1 // SPDX-License-Identifier: GPL-2.0-only 1 // SPDX-License-Identifier: GPL-2.0-only
2 /* 2 /*
3 * kernel/workqueue.c - generic async executio 3 * kernel/workqueue.c - generic async execution with shared worker pool
4 * 4 *
5 * Copyright (C) 2002 Ingo Molnar 5 * Copyright (C) 2002 Ingo Molnar
6 * 6 *
7 * Derived from the taskqueue/keventd code b 7 * Derived from the taskqueue/keventd code by:
8 * David Woodhouse <dwmw2@infradead.org> 8 * David Woodhouse <dwmw2@infradead.org>
9 * Andrew Morton 9 * Andrew Morton
10 * Kai Petzke <wpp@marie.physik.tu-berlin. 10 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
11 * Theodore Ts'o <tytso@mit.edu> 11 * Theodore Ts'o <tytso@mit.edu>
12 * 12 *
13 * Made to use alloc_percpu by Christoph Lamet 13 * Made to use alloc_percpu by Christoph Lameter.
14 * 14 *
15 * Copyright (C) 2010 SUSE Linux Pro 15 * Copyright (C) 2010 SUSE Linux Products GmbH
16 * Copyright (C) 2010 Tejun Heo <tj@ 16 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
17 * 17 *
18 * This is the generic async execution mechani 18 * This is the generic async execution mechanism. Work items as are
19 * executed in process context. The worker po 19 * executed in process context. The worker pool is shared and
20 * automatically managed. There are two worke 20 * automatically managed. There are two worker pools for each CPU (one for
21 * normal work items and the other for high pr 21 * normal work items and the other for high priority ones) and some extra
22 * pools for workqueues which are not bound to 22 * pools for workqueues which are not bound to any specific CPU - the
23 * number of these backing pools is dynamic. 23 * number of these backing pools is dynamic.
24 * 24 *
25 * Please read Documentation/core-api/workqueu 25 * Please read Documentation/core-api/workqueue.rst for details.
26 */ 26 */
27 27
28 #include <linux/export.h> 28 #include <linux/export.h>
29 #include <linux/kernel.h> 29 #include <linux/kernel.h>
30 #include <linux/sched.h> 30 #include <linux/sched.h>
31 #include <linux/init.h> 31 #include <linux/init.h>
32 #include <linux/interrupt.h> <<
33 #include <linux/signal.h> 32 #include <linux/signal.h>
34 #include <linux/completion.h> 33 #include <linux/completion.h>
35 #include <linux/workqueue.h> 34 #include <linux/workqueue.h>
36 #include <linux/slab.h> 35 #include <linux/slab.h>
37 #include <linux/cpu.h> 36 #include <linux/cpu.h>
38 #include <linux/notifier.h> 37 #include <linux/notifier.h>
39 #include <linux/kthread.h> 38 #include <linux/kthread.h>
40 #include <linux/hardirq.h> 39 #include <linux/hardirq.h>
41 #include <linux/mempolicy.h> 40 #include <linux/mempolicy.h>
42 #include <linux/freezer.h> 41 #include <linux/freezer.h>
43 #include <linux/debug_locks.h> 42 #include <linux/debug_locks.h>
44 #include <linux/lockdep.h> 43 #include <linux/lockdep.h>
45 #include <linux/idr.h> 44 #include <linux/idr.h>
46 #include <linux/jhash.h> 45 #include <linux/jhash.h>
47 #include <linux/hashtable.h> 46 #include <linux/hashtable.h>
48 #include <linux/rculist.h> 47 #include <linux/rculist.h>
49 #include <linux/nodemask.h> 48 #include <linux/nodemask.h>
50 #include <linux/moduleparam.h> 49 #include <linux/moduleparam.h>
51 #include <linux/uaccess.h> 50 #include <linux/uaccess.h>
52 #include <linux/sched/isolation.h> 51 #include <linux/sched/isolation.h>
53 #include <linux/sched/debug.h> 52 #include <linux/sched/debug.h>
54 #include <linux/nmi.h> 53 #include <linux/nmi.h>
55 #include <linux/kvm_para.h> 54 #include <linux/kvm_para.h>
56 #include <linux/delay.h> 55 #include <linux/delay.h>
57 #include <linux/irq_work.h> <<
58 56
59 #include "workqueue_internal.h" 57 #include "workqueue_internal.h"
60 58
61 enum worker_pool_flags { !! 59 enum {
62 /* 60 /*
63 * worker_pool flags 61 * worker_pool flags
64 * 62 *
65 * A bound pool is either associated o 63 * A bound pool is either associated or disassociated with its CPU.
66 * While associated (!DISASSOCIATED), 64 * While associated (!DISASSOCIATED), all workers are bound to the
67 * CPU and none has %WORKER_UNBOUND se 65 * CPU and none has %WORKER_UNBOUND set and concurrency management
68 * is in effect. 66 * is in effect.
69 * 67 *
70 * While DISASSOCIATED, the cpu may be 68 * While DISASSOCIATED, the cpu may be offline and all workers have
71 * %WORKER_UNBOUND set and concurrency 69 * %WORKER_UNBOUND set and concurrency management disabled, and may
72 * be executing on any CPU. The pool 70 * be executing on any CPU. The pool behaves as an unbound one.
73 * 71 *
74 * Note that DISASSOCIATED should be f 72 * Note that DISASSOCIATED should be flipped only while holding
75 * wq_pool_attach_mutex to avoid chang 73 * wq_pool_attach_mutex to avoid changing binding state while
76 * worker_attach_to_pool() is in progr 74 * worker_attach_to_pool() is in progress.
77 * <<
78 * As there can only be one concurrent <<
79 * BH pool is per-CPU and always DISAS <<
80 */ 75 */
81 POOL_BH = 1 << 0, !! 76 POOL_MANAGER_ACTIVE = 1 << 0, /* being managed */
82 POOL_MANAGER_ACTIVE = 1 << 1, <<
83 POOL_DISASSOCIATED = 1 << 2, 77 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
84 POOL_BH_DRAINING = 1 << 3, <<
85 }; <<
86 78
87 enum worker_flags { <<
88 /* worker flags */ 79 /* worker flags */
89 WORKER_DIE = 1 << 1, 80 WORKER_DIE = 1 << 1, /* die die die */
90 WORKER_IDLE = 1 << 2, 81 WORKER_IDLE = 1 << 2, /* is idle */
91 WORKER_PREP = 1 << 3, 82 WORKER_PREP = 1 << 3, /* preparing to run works */
92 WORKER_CPU_INTENSIVE = 1 << 6, 83 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
93 WORKER_UNBOUND = 1 << 7, 84 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
94 WORKER_REBOUND = 1 << 8, 85 WORKER_REBOUND = 1 << 8, /* worker was rebound */
95 86
96 WORKER_NOT_RUNNING = WORKER_PREP 87 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
97 WORKER_UNBOU 88 WORKER_UNBOUND | WORKER_REBOUND,
98 }; <<
99 <<
100 enum work_cancel_flags { <<
101 WORK_CANCEL_DELAYED = 1 << 0, <<
102 WORK_CANCEL_DISABLE = 1 << 1, <<
103 }; <<
104 89
105 enum wq_internal_consts { <<
106 NR_STD_WORKER_POOLS = 2, 90 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
107 91
108 UNBOUND_POOL_HASH_ORDER = 6, 92 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
109 BUSY_WORKER_HASH_ORDER = 6, 93 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
110 94
111 MAX_IDLE_WORKERS_RATIO = 4, 95 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
112 IDLE_WORKER_TIMEOUT = 300 * HZ, 96 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
113 97
114 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 98 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
115 99 /* call for help after 10ms
116 100 (min two ticks) */
117 MAYDAY_INTERVAL = HZ / 10, 101 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
118 CREATE_COOLDOWN = HZ, 102 CREATE_COOLDOWN = HZ, /* time to breath after fail */
119 103
120 /* 104 /*
121 * Rescue workers are used only on eme 105 * Rescue workers are used only on emergencies and shared by
122 * all cpus. Give MIN_NICE. 106 * all cpus. Give MIN_NICE.
123 */ 107 */
124 RESCUER_NICE_LEVEL = MIN_NICE, 108 RESCUER_NICE_LEVEL = MIN_NICE,
125 HIGHPRI_NICE_LEVEL = MIN_NICE, 109 HIGHPRI_NICE_LEVEL = MIN_NICE,
126 110
127 WQ_NAME_LEN = 32, !! 111 WQ_NAME_LEN = 24,
128 WORKER_ID_LEN = 10 + WQ_NAME <<
129 }; 112 };
130 113
131 /* 114 /*
132 * We don't want to trap softirq for too long. <<
133 * MAX_SOFTIRQ_RESTART in kernel/softirq.c. Th <<
134 * msecs_to_jiffies() can't be an initializer. <<
135 */ <<
136 #define BH_WORKER_JIFFIES msecs_to_jiffi <<
137 #define BH_WORKER_RESTARTS 10 <<
138 <<
139 /* <<
140 * Structure fields follow one of the followin 115 * Structure fields follow one of the following exclusion rules.
141 * 116 *
142 * I: Modifiable by initialization/destruction 117 * I: Modifiable by initialization/destruction paths and read-only for
143 * everyone else. 118 * everyone else.
144 * 119 *
145 * P: Preemption protected. Disabling preempt 120 * P: Preemption protected. Disabling preemption is enough and should
146 * only be modified and accessed from the l 121 * only be modified and accessed from the local cpu.
147 * 122 *
148 * L: pool->lock protected. Access with pool- 123 * L: pool->lock protected. Access with pool->lock held.
149 * 124 *
150 * LN: pool->lock and wq_node_nr_active->lock <<
151 * reads. <<
152 * <<
153 * K: Only modified by worker while holding po 125 * K: Only modified by worker while holding pool->lock. Can be safely read by
154 * self, while holding pool->lock or from I 126 * self, while holding pool->lock or from IRQ context if %current is the
155 * kworker. 127 * kworker.
156 * 128 *
157 * S: Only modified by worker self. 129 * S: Only modified by worker self.
158 * 130 *
159 * A: wq_pool_attach_mutex protected. 131 * A: wq_pool_attach_mutex protected.
160 * 132 *
161 * PL: wq_pool_mutex protected. 133 * PL: wq_pool_mutex protected.
162 * 134 *
163 * PR: wq_pool_mutex protected for writes. RC 135 * PR: wq_pool_mutex protected for writes. RCU protected for reads.
164 * 136 *
165 * PW: wq_pool_mutex and wq->mutex protected f 137 * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads.
166 * 138 *
167 * PWR: wq_pool_mutex and wq->mutex protected 139 * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or
168 * RCU for reads. 140 * RCU for reads.
169 * 141 *
170 * WQ: wq->mutex protected. 142 * WQ: wq->mutex protected.
171 * 143 *
172 * WR: wq->mutex protected for writes. RCU pr 144 * WR: wq->mutex protected for writes. RCU protected for reads.
173 * 145 *
174 * WO: wq->mutex protected for writes. Updated <<
175 * with READ_ONCE() without locking. <<
176 * <<
177 * MD: wq_mayday_lock protected. 146 * MD: wq_mayday_lock protected.
178 * 147 *
179 * WD: Used internally by the watchdog. 148 * WD: Used internally by the watchdog.
180 */ 149 */
181 150
182 /* struct worker is defined in workqueue_inter 151 /* struct worker is defined in workqueue_internal.h */
183 152
184 struct worker_pool { 153 struct worker_pool {
185 raw_spinlock_t lock; 154 raw_spinlock_t lock; /* the pool lock */
186 int cpu; 155 int cpu; /* I: the associated cpu */
187 int node; 156 int node; /* I: the associated node ID */
188 int id; 157 int id; /* I: pool ID */
189 unsigned int flags; 158 unsigned int flags; /* L: flags */
190 159
191 unsigned long watchdog_ts; 160 unsigned long watchdog_ts; /* L: watchdog timestamp */
192 bool cpu_stall; 161 bool cpu_stall; /* WD: stalled cpu bound pool */
193 162
194 /* 163 /*
195 * The counter is incremented in a pro 164 * The counter is incremented in a process context on the associated CPU
196 * w/ preemption disabled, and decreme 165 * w/ preemption disabled, and decremented or reset in the same context
197 * but w/ pool->lock held. The readers 166 * but w/ pool->lock held. The readers grab pool->lock and are
198 * guaranteed to see if the counter re 167 * guaranteed to see if the counter reached zero.
199 */ 168 */
200 int nr_running; 169 int nr_running;
201 170
202 struct list_head worklist; 171 struct list_head worklist; /* L: list of pending works */
203 172
204 int nr_workers; 173 int nr_workers; /* L: total number of workers */
205 int nr_idle; 174 int nr_idle; /* L: currently idle workers */
206 175
207 struct list_head idle_list; 176 struct list_head idle_list; /* L: list of idle workers */
208 struct timer_list idle_timer; 177 struct timer_list idle_timer; /* L: worker idle timeout */
209 struct work_struct idle_cull_work 178 struct work_struct idle_cull_work; /* L: worker idle cleanup */
210 179
211 struct timer_list mayday_timer; 180 struct timer_list mayday_timer; /* L: SOS timer for workers */
212 181
213 /* a workers is either on busy_hash or 182 /* a workers is either on busy_hash or idle_list, or the manager */
214 DECLARE_HASHTABLE(busy_hash, BUSY_WORK 183 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
215 184 /* L: hash of busy workers */
216 185
217 struct worker *manager; 186 struct worker *manager; /* L: purely informational */
218 struct list_head workers; 187 struct list_head workers; /* A: attached workers */
>> 188 struct list_head dying_workers; /* A: workers about to die */
>> 189 struct completion *detach_completion; /* all workers detached */
219 190
220 struct ida worker_ida; 191 struct ida worker_ida; /* worker IDs for task name */
221 192
222 struct workqueue_attrs *attrs; 193 struct workqueue_attrs *attrs; /* I: worker attributes */
223 struct hlist_node hash_node; 194 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
224 int refcnt; 195 int refcnt; /* PL: refcnt for unbound pools */
225 196
226 /* 197 /*
227 * Destruction of pool is RCU protecte 198 * Destruction of pool is RCU protected to allow dereferences
228 * from get_work_pool(). 199 * from get_work_pool().
229 */ 200 */
230 struct rcu_head rcu; 201 struct rcu_head rcu;
231 }; 202 };
232 203
233 /* 204 /*
234 * Per-pool_workqueue statistics. These can be 205 * Per-pool_workqueue statistics. These can be monitored using
235 * tools/workqueue/wq_monitor.py. 206 * tools/workqueue/wq_monitor.py.
236 */ 207 */
237 enum pool_workqueue_stats { 208 enum pool_workqueue_stats {
238 PWQ_STAT_STARTED, /* work items 209 PWQ_STAT_STARTED, /* work items started execution */
239 PWQ_STAT_COMPLETED, /* work items 210 PWQ_STAT_COMPLETED, /* work items completed execution */
240 PWQ_STAT_CPU_TIME, /* total CPU t 211 PWQ_STAT_CPU_TIME, /* total CPU time consumed */
241 PWQ_STAT_CPU_INTENSIVE, /* wq_cpu_inte 212 PWQ_STAT_CPU_INTENSIVE, /* wq_cpu_intensive_thresh_us violations */
242 PWQ_STAT_CM_WAKEUP, /* concurrency 213 PWQ_STAT_CM_WAKEUP, /* concurrency-management worker wakeups */
243 PWQ_STAT_REPATRIATED, /* unbound wor 214 PWQ_STAT_REPATRIATED, /* unbound workers brought back into scope */
244 PWQ_STAT_MAYDAY, /* maydays to 215 PWQ_STAT_MAYDAY, /* maydays to rescuer */
245 PWQ_STAT_RESCUED, /* linked work 216 PWQ_STAT_RESCUED, /* linked work items executed by rescuer */
246 217
247 PWQ_NR_STATS, 218 PWQ_NR_STATS,
248 }; 219 };
249 220
250 /* 221 /*
251 * The per-pool workqueue. While queued, bits !! 222 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
252 * of work_struct->data are used for flags and 223 * of work_struct->data are used for flags and the remaining high bits
253 * point to the pwq; thus, pwqs need to be ali 224 * point to the pwq; thus, pwqs need to be aligned at two's power of the
254 * number of flag bits. 225 * number of flag bits.
255 */ 226 */
256 struct pool_workqueue { 227 struct pool_workqueue {
257 struct worker_pool *pool; 228 struct worker_pool *pool; /* I: the associated pool */
258 struct workqueue_struct *wq; 229 struct workqueue_struct *wq; /* I: the owning workqueue */
259 int work_color; 230 int work_color; /* L: current color */
260 int flush_color; 231 int flush_color; /* L: flushing color */
261 int refcnt; 232 int refcnt; /* L: reference count */
262 int nr_in_flight[W 233 int nr_in_flight[WORK_NR_COLORS];
263 234 /* L: nr of in_flight works */
264 bool plugged; <<
265 235
266 /* 236 /*
267 * nr_active management and WORK_STRUC 237 * nr_active management and WORK_STRUCT_INACTIVE:
268 * 238 *
269 * When pwq->nr_active >= max_active, 239 * When pwq->nr_active >= max_active, new work item is queued to
270 * pwq->inactive_works instead of pool 240 * pwq->inactive_works instead of pool->worklist and marked with
271 * WORK_STRUCT_INACTIVE. 241 * WORK_STRUCT_INACTIVE.
272 * 242 *
273 * All work items marked with WORK_STR !! 243 * All work items marked with WORK_STRUCT_INACTIVE do not participate
274 * nr_active and all work items in pwq !! 244 * in pwq->nr_active and all work items in pwq->inactive_works are
275 * WORK_STRUCT_INACTIVE. But not all W !! 245 * marked with WORK_STRUCT_INACTIVE. But not all WORK_STRUCT_INACTIVE
276 * in pwq->inactive_works. Some of the !! 246 * work items are in pwq->inactive_works. Some of them are ready to
277 * pool->worklist or worker->scheduled !! 247 * run in pool->worklist or worker->scheduled. Those work itmes are
278 * wq_barrier which is used for flush_ !! 248 * only struct wq_barrier which is used for flush_work() and should
279 * in nr_active. For non-barrier work !! 249 * not participate in pwq->nr_active. For non-barrier work item, it
280 * WORK_STRUCT_INACTIVE iff it is in p !! 250 * is marked with WORK_STRUCT_INACTIVE iff it is in pwq->inactive_works.
281 */ 251 */
282 int nr_active; 252 int nr_active; /* L: nr of active works */
>> 253 int max_active; /* L: max active works */
283 struct list_head inactive_works 254 struct list_head inactive_works; /* L: inactive works */
284 struct list_head pending_node; <<
285 struct list_head pwqs_node; 255 struct list_head pwqs_node; /* WR: node on wq->pwqs */
286 struct list_head mayday_node; 256 struct list_head mayday_node; /* MD: node on wq->maydays */
287 257
288 u64 stats[PWQ_NR_S 258 u64 stats[PWQ_NR_STATS];
289 259
290 /* 260 /*
291 * Release of unbound pwq is punted to 261 * Release of unbound pwq is punted to a kthread_worker. See put_pwq()
292 * and pwq_release_workfn() for detail 262 * and pwq_release_workfn() for details. pool_workqueue itself is also
293 * RCU protected so that the first pwq 263 * RCU protected so that the first pwq can be determined without
294 * grabbing wq->mutex. 264 * grabbing wq->mutex.
295 */ 265 */
296 struct kthread_work release_work; 266 struct kthread_work release_work;
297 struct rcu_head rcu; 267 struct rcu_head rcu;
298 } __aligned(1 << WORK_STRUCT_PWQ_SHIFT); !! 268 } __aligned(1 << WORK_STRUCT_FLAG_BITS);
299 269
300 /* 270 /*
301 * Structure used to wait for workqueue flush. 271 * Structure used to wait for workqueue flush.
302 */ 272 */
303 struct wq_flusher { 273 struct wq_flusher {
304 struct list_head list; 274 struct list_head list; /* WQ: list of flushers */
305 int flush_color; 275 int flush_color; /* WQ: flush color waiting for */
306 struct completion done; 276 struct completion done; /* flush completion */
307 }; 277 };
308 278
309 struct wq_device; 279 struct wq_device;
310 280
311 /* 281 /*
312 * Unlike in a per-cpu workqueue where max_act <<
313 * on each CPU, in an unbound workqueue, max_a <<
314 * As sharing a single nr_active across multip <<
315 * the counting and enforcement is per NUMA no <<
316 * <<
317 * The following struct is used to enforce per <<
318 * to start executing a work item, it should i <<
319 * tryinc_node_nr_active(). If acquisition fai <<
320 * ->max, the pwq is queued on ->pending_pwqs. <<
321 * and decrement ->nr, node_activate_pending_p <<
322 * round-robin order. <<
323 */ <<
324 struct wq_node_nr_active { <<
325 int max; <<
326 atomic_t nr; <<
327 raw_spinlock_t lock; <<
328 struct list_head pending_pwqs; <<
329 }; <<
330 <<
331 /* <<
332 * The externally visible workqueue. It relay 282 * The externally visible workqueue. It relays the issued work items to
333 * the appropriate worker_pool through its poo 283 * the appropriate worker_pool through its pool_workqueues.
334 */ 284 */
335 struct workqueue_struct { 285 struct workqueue_struct {
336 struct list_head pwqs; 286 struct list_head pwqs; /* WR: all pwqs of this wq */
337 struct list_head list; 287 struct list_head list; /* PR: list of all workqueues */
338 288
339 struct mutex mutex; 289 struct mutex mutex; /* protects this wq */
340 int work_color; 290 int work_color; /* WQ: current work color */
341 int flush_color; 291 int flush_color; /* WQ: current flush color */
342 atomic_t nr_pwqs_to_flu 292 atomic_t nr_pwqs_to_flush; /* flush in progress */
343 struct wq_flusher *first_flusher 293 struct wq_flusher *first_flusher; /* WQ: first flusher */
344 struct list_head flusher_queue; 294 struct list_head flusher_queue; /* WQ: flush waiters */
345 struct list_head flusher_overfl 295 struct list_head flusher_overflow; /* WQ: flush overflow list */
346 296
347 struct list_head maydays; 297 struct list_head maydays; /* MD: pwqs requesting rescue */
348 struct worker *rescuer; 298 struct worker *rescuer; /* MD: rescue worker */
349 299
350 int nr_drainers; 300 int nr_drainers; /* WQ: drain in progress */
351 !! 301 int saved_max_active; /* WQ: saved pwq max_active */
352 /* See alloc_workqueue() function comm <<
353 int max_active; <<
354 int min_active; <<
355 int saved_max_acti <<
356 int saved_min_acti <<
357 302
358 struct workqueue_attrs *unbound_attrs 303 struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */
359 struct pool_workqueue __rcu *dfl_pwq; !! 304 struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */
360 305
361 #ifdef CONFIG_SYSFS 306 #ifdef CONFIG_SYSFS
362 struct wq_device *wq_dev; 307 struct wq_device *wq_dev; /* I: for sysfs interface */
363 #endif 308 #endif
364 #ifdef CONFIG_LOCKDEP 309 #ifdef CONFIG_LOCKDEP
365 char *lock_name; 310 char *lock_name;
366 struct lock_class_key key; 311 struct lock_class_key key;
367 struct lockdep_map lockdep_map; 312 struct lockdep_map lockdep_map;
368 #endif 313 #endif
369 char name[WQ_NAME_L 314 char name[WQ_NAME_LEN]; /* I: workqueue name */
370 315
371 /* 316 /*
372 * Destruction of workqueue_struct is 317 * Destruction of workqueue_struct is RCU protected to allow walking
373 * the workqueues list without grabbin 318 * the workqueues list without grabbing wq_pool_mutex.
374 * This is used to dump all workqueues 319 * This is used to dump all workqueues from sysrq.
375 */ 320 */
376 struct rcu_head rcu; 321 struct rcu_head rcu;
377 322
378 /* hot fields used during command issu 323 /* hot fields used during command issue, aligned to cacheline */
379 unsigned int flags ____cach 324 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
380 struct pool_workqueue __rcu * __percpu !! 325 struct pool_workqueue __percpu __rcu **cpu_pwq; /* I: per-cpu pwqs */
381 struct wq_node_nr_active *node_nr_acti <<
382 }; 326 };
383 327
>> 328 static struct kmem_cache *pwq_cache;
>> 329
384 /* 330 /*
385 * Each pod type describes how CPUs should be 331 * Each pod type describes how CPUs should be grouped for unbound workqueues.
386 * See the comment above workqueue_attrs->affn 332 * See the comment above workqueue_attrs->affn_scope.
387 */ 333 */
388 struct wq_pod_type { 334 struct wq_pod_type {
389 int nr_pods; 335 int nr_pods; /* number of pods */
390 cpumask_var_t *pod_cpus; 336 cpumask_var_t *pod_cpus; /* pod -> cpus */
391 int *pod_node; 337 int *pod_node; /* pod -> node */
392 int *cpu_pod; 338 int *cpu_pod; /* cpu -> pod */
393 }; 339 };
394 340
395 struct work_offq_data { !! 341 static struct wq_pod_type wq_pod_types[WQ_AFFN_NR_TYPES];
396 u32 pool_id; !! 342 static enum wq_affn_scope wq_affn_dfl = WQ_AFFN_CACHE;
397 u32 disable; <<
398 u32 flags; <<
399 }; <<
400 343
401 static const char *wq_affn_names[WQ_AFFN_NR_TY 344 static const char *wq_affn_names[WQ_AFFN_NR_TYPES] = {
402 [WQ_AFFN_DFL] = "default", !! 345 [WQ_AFFN_DFL] = "default",
403 [WQ_AFFN_CPU] = "cpu", !! 346 [WQ_AFFN_CPU] = "cpu",
404 [WQ_AFFN_SMT] = "smt", !! 347 [WQ_AFFN_SMT] = "smt",
405 [WQ_AFFN_CACHE] = "cache", !! 348 [WQ_AFFN_CACHE] = "cache",
406 [WQ_AFFN_NUMA] = "numa", !! 349 [WQ_AFFN_NUMA] = "numa",
407 [WQ_AFFN_SYSTEM] = "system", !! 350 [WQ_AFFN_SYSTEM] = "system",
408 }; 351 };
409 352
410 /* 353 /*
411 * Per-cpu work items which run for longer tha 354 * Per-cpu work items which run for longer than the following threshold are
412 * automatically considered CPU intensive and 355 * automatically considered CPU intensive and excluded from concurrency
413 * management to prevent them from noticeably 356 * management to prevent them from noticeably delaying other per-cpu work items.
414 * ULONG_MAX indicates that the user hasn't ov 357 * ULONG_MAX indicates that the user hasn't overridden it with a boot parameter.
415 * The actual value is initialized in wq_cpu_i 358 * The actual value is initialized in wq_cpu_intensive_thresh_init().
416 */ 359 */
417 static unsigned long wq_cpu_intensive_thresh_u 360 static unsigned long wq_cpu_intensive_thresh_us = ULONG_MAX;
418 module_param_named(cpu_intensive_thresh_us, wq 361 module_param_named(cpu_intensive_thresh_us, wq_cpu_intensive_thresh_us, ulong, 0644);
419 #ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT <<
420 static unsigned int wq_cpu_intensive_warning_t <<
421 module_param_named(cpu_intensive_warning_thres <<
422 #endif <<
423 362
424 /* see the comment above the definition of WQ_ 363 /* see the comment above the definition of WQ_POWER_EFFICIENT */
425 static bool wq_power_efficient = IS_ENABLED(CO 364 static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
426 module_param_named(power_efficient, wq_power_e 365 module_param_named(power_efficient, wq_power_efficient, bool, 0444);
427 366
428 static bool wq_online; /* can 367 static bool wq_online; /* can kworkers be created yet? */
429 static bool wq_topo_initialized __read_mostly <<
430 <<
431 static struct kmem_cache *pwq_cache; <<
432 <<
433 static struct wq_pod_type wq_pod_types[WQ_AFFN <<
434 static enum wq_affn_scope wq_affn_dfl = WQ_AFF <<
435 368
436 /* buf for wq_update_unbound_pod_attrs(), prot 369 /* buf for wq_update_unbound_pod_attrs(), protected by CPU hotplug exclusion */
437 static struct workqueue_attrs *unbound_wq_upda !! 370 static struct workqueue_attrs *wq_update_pod_attrs_buf;
438 371
439 static DEFINE_MUTEX(wq_pool_mutex); /* pro 372 static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
440 static DEFINE_MUTEX(wq_pool_attach_mutex); /* 373 static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */
441 static DEFINE_RAW_SPINLOCK(wq_mayday_lock); 374 static DEFINE_RAW_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
442 /* wait for manager to go away */ 375 /* wait for manager to go away */
443 static struct rcuwait manager_wait = __RCUWAIT 376 static struct rcuwait manager_wait = __RCUWAIT_INITIALIZER(manager_wait);
444 377
445 static LIST_HEAD(workqueues); /* PR: 378 static LIST_HEAD(workqueues); /* PR: list of all workqueues */
446 static bool workqueue_freezing; /* PL: 379 static bool workqueue_freezing; /* PL: have wqs started freezing? */
447 380
448 /* PL: mirror the cpu_online_mask excluding th <<
449 static cpumask_var_t wq_online_cpumask; <<
450 <<
451 /* PL&A: allowable cpus for unbound wqs and wo 381 /* PL&A: allowable cpus for unbound wqs and work items */
452 static cpumask_var_t wq_unbound_cpumask; 382 static cpumask_var_t wq_unbound_cpumask;
453 383
454 /* PL: user requested unbound cpumask via sysf <<
455 static cpumask_var_t wq_requested_unbound_cpum <<
456 <<
457 /* PL: isolated cpumask to be excluded from un <<
458 static cpumask_var_t wq_isolated_cpumask; <<
459 <<
460 /* for further constrain wq_unbound_cpumask by 384 /* for further constrain wq_unbound_cpumask by cmdline parameter*/
461 static struct cpumask wq_cmdline_cpumask __ini 385 static struct cpumask wq_cmdline_cpumask __initdata;
462 386
463 /* CPU where unbound work was last round robin 387 /* CPU where unbound work was last round robin scheduled from this CPU */
464 static DEFINE_PER_CPU(int, wq_rr_cpu_last); 388 static DEFINE_PER_CPU(int, wq_rr_cpu_last);
465 389
466 /* 390 /*
467 * Local execution of unbound work items is no 391 * Local execution of unbound work items is no longer guaranteed. The
468 * following always forces round-robin CPU sel 392 * following always forces round-robin CPU selection on unbound work items
469 * to uncover usages which depend on it. 393 * to uncover usages which depend on it.
470 */ 394 */
471 #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU 395 #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
472 static bool wq_debug_force_rr_cpu = true; 396 static bool wq_debug_force_rr_cpu = true;
473 #else 397 #else
474 static bool wq_debug_force_rr_cpu = false; 398 static bool wq_debug_force_rr_cpu = false;
475 #endif 399 #endif
476 module_param_named(debug_force_rr_cpu, wq_debu 400 module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
477 401
478 /* to raise softirq for the BH worker pools on <<
479 static DEFINE_PER_CPU_SHARED_ALIGNED(struct ir <<
480 bh_pool_i <<
481 <<
482 /* the BH worker pools */ <<
483 static DEFINE_PER_CPU_SHARED_ALIGNED(struct wo <<
484 bh_worker <<
485 <<
486 /* the per-cpu worker pools */ 402 /* the per-cpu worker pools */
487 static DEFINE_PER_CPU_SHARED_ALIGNED(struct wo !! 403 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools);
488 cpu_worke <<
489 404
490 static DEFINE_IDR(worker_pool_idr); /* PR: 405 static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
491 406
492 /* PL: hash of all unbound pools keyed by pool 407 /* PL: hash of all unbound pools keyed by pool->attrs */
493 static DEFINE_HASHTABLE(unbound_pool_hash, UNB 408 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
494 409
495 /* I: attributes used when instantiating stand 410 /* I: attributes used when instantiating standard unbound pools on demand */
496 static struct workqueue_attrs *unbound_std_wq_ 411 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
497 412
498 /* I: attributes used when instantiating order 413 /* I: attributes used when instantiating ordered pools on demand */
499 static struct workqueue_attrs *ordered_wq_attr 414 static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
500 415
501 /* 416 /*
502 * I: kthread_worker to release pwq's. pwq rel 417 * I: kthread_worker to release pwq's. pwq release needs to be bounced to a
503 * process context while holding a pool lock. 418 * process context while holding a pool lock. Bounce to a dedicated kthread
504 * worker to avoid A-A deadlocks. 419 * worker to avoid A-A deadlocks.
505 */ 420 */
506 static struct kthread_worker *pwq_release_work !! 421 static struct kthread_worker *pwq_release_worker;
507 422
508 struct workqueue_struct *system_wq __ro_after_ !! 423 struct workqueue_struct *system_wq __read_mostly;
509 EXPORT_SYMBOL(system_wq); 424 EXPORT_SYMBOL(system_wq);
510 struct workqueue_struct *system_highpri_wq __r !! 425 struct workqueue_struct *system_highpri_wq __read_mostly;
511 EXPORT_SYMBOL_GPL(system_highpri_wq); 426 EXPORT_SYMBOL_GPL(system_highpri_wq);
512 struct workqueue_struct *system_long_wq __ro_a !! 427 struct workqueue_struct *system_long_wq __read_mostly;
513 EXPORT_SYMBOL_GPL(system_long_wq); 428 EXPORT_SYMBOL_GPL(system_long_wq);
514 struct workqueue_struct *system_unbound_wq __r !! 429 struct workqueue_struct *system_unbound_wq __read_mostly;
515 EXPORT_SYMBOL_GPL(system_unbound_wq); 430 EXPORT_SYMBOL_GPL(system_unbound_wq);
516 struct workqueue_struct *system_freezable_wq _ !! 431 struct workqueue_struct *system_freezable_wq __read_mostly;
517 EXPORT_SYMBOL_GPL(system_freezable_wq); 432 EXPORT_SYMBOL_GPL(system_freezable_wq);
518 struct workqueue_struct *system_power_efficien !! 433 struct workqueue_struct *system_power_efficient_wq __read_mostly;
519 EXPORT_SYMBOL_GPL(system_power_efficient_wq); 434 EXPORT_SYMBOL_GPL(system_power_efficient_wq);
520 struct workqueue_struct *system_freezable_powe !! 435 struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
521 EXPORT_SYMBOL_GPL(system_freezable_power_effic 436 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
522 struct workqueue_struct *system_bh_wq; <<
523 EXPORT_SYMBOL_GPL(system_bh_wq); <<
524 struct workqueue_struct *system_bh_highpri_wq; <<
525 EXPORT_SYMBOL_GPL(system_bh_highpri_wq); <<
526 437
527 static int worker_thread(void *__worker); 438 static int worker_thread(void *__worker);
528 static void workqueue_sysfs_unregister(struct 439 static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
529 static void show_pwq(struct pool_workqueue *pw 440 static void show_pwq(struct pool_workqueue *pwq);
530 static void show_one_worker_pool(struct worker 441 static void show_one_worker_pool(struct worker_pool *pool);
531 442
532 #define CREATE_TRACE_POINTS 443 #define CREATE_TRACE_POINTS
533 #include <trace/events/workqueue.h> 444 #include <trace/events/workqueue.h>
534 445
535 #define assert_rcu_or_pool_mutex() 446 #define assert_rcu_or_pool_mutex() \
536 RCU_LOCKDEP_WARN(!rcu_read_lock_any_he !! 447 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
537 !lockdep_is_held(&wq_ 448 !lockdep_is_held(&wq_pool_mutex), \
538 "RCU or wq_pool_mutex 449 "RCU or wq_pool_mutex should be held")
539 450
540 #define assert_rcu_or_wq_mutex_or_pool_mutex(w 451 #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
541 RCU_LOCKDEP_WARN(!rcu_read_lock_any_he !! 452 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
542 !lockdep_is_held(&wq- 453 !lockdep_is_held(&wq->mutex) && \
543 !lockdep_is_held(&wq_ 454 !lockdep_is_held(&wq_pool_mutex), \
544 "RCU, wq->mutex or wq 455 "RCU, wq->mutex or wq_pool_mutex should be held")
545 456
546 #define for_each_bh_worker_pool(pool, cpu) <<
547 for ((pool) = &per_cpu(bh_worker_pools <<
548 (pool) < &per_cpu(bh_worker_pools <<
549 (pool)++) <<
550 <<
551 #define for_each_cpu_worker_pool(pool, cpu) 457 #define for_each_cpu_worker_pool(pool, cpu) \
552 for ((pool) = &per_cpu(cpu_worker_pool 458 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
553 (pool) < &per_cpu(cpu_worker_pool 459 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
554 (pool)++) 460 (pool)++)
555 461
556 /** 462 /**
557 * for_each_pool - iterate through all worker_ 463 * for_each_pool - iterate through all worker_pools in the system
558 * @pool: iteration cursor 464 * @pool: iteration cursor
559 * @pi: integer used for iteration 465 * @pi: integer used for iteration
560 * 466 *
561 * This must be called either with wq_pool_mut 467 * This must be called either with wq_pool_mutex held or RCU read
562 * locked. If the pool needs to be used beyon 468 * locked. If the pool needs to be used beyond the locking in effect, the
563 * caller is responsible for guaranteeing that 469 * caller is responsible for guaranteeing that the pool stays online.
564 * 470 *
565 * The if/else clause exists only for the lock 471 * The if/else clause exists only for the lockdep assertion and can be
566 * ignored. 472 * ignored.
567 */ 473 */
568 #define for_each_pool(pool, pi) 474 #define for_each_pool(pool, pi) \
569 idr_for_each_entry(&worker_pool_idr, p 475 idr_for_each_entry(&worker_pool_idr, pool, pi) \
570 if (({ assert_rcu_or_pool_mute 476 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
571 else 477 else
572 478
573 /** 479 /**
574 * for_each_pool_worker - iterate through all 480 * for_each_pool_worker - iterate through all workers of a worker_pool
575 * @worker: iteration cursor 481 * @worker: iteration cursor
576 * @pool: worker_pool to iterate workers of 482 * @pool: worker_pool to iterate workers of
577 * 483 *
578 * This must be called with wq_pool_attach_mut 484 * This must be called with wq_pool_attach_mutex.
579 * 485 *
580 * The if/else clause exists only for the lock 486 * The if/else clause exists only for the lockdep assertion and can be
581 * ignored. 487 * ignored.
582 */ 488 */
583 #define for_each_pool_worker(worker, pool) 489 #define for_each_pool_worker(worker, pool) \
584 list_for_each_entry((worker), &(pool)- 490 list_for_each_entry((worker), &(pool)->workers, node) \
585 if (({ lockdep_assert_held(&wq 491 if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \
586 else 492 else
587 493
588 /** 494 /**
589 * for_each_pwq - iterate through all pool_wor 495 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
590 * @pwq: iteration cursor 496 * @pwq: iteration cursor
591 * @wq: the target workqueue 497 * @wq: the target workqueue
592 * 498 *
593 * This must be called either with wq->mutex h 499 * This must be called either with wq->mutex held or RCU read locked.
594 * If the pwq needs to be used beyond the lock 500 * If the pwq needs to be used beyond the locking in effect, the caller is
595 * responsible for guaranteeing that the pwq s 501 * responsible for guaranteeing that the pwq stays online.
596 * 502 *
597 * The if/else clause exists only for the lock 503 * The if/else clause exists only for the lockdep assertion and can be
598 * ignored. 504 * ignored.
599 */ 505 */
600 #define for_each_pwq(pwq, wq) 506 #define for_each_pwq(pwq, wq) \
601 list_for_each_entry_rcu((pwq), &(wq)-> 507 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \
602 lockdep_is_he 508 lockdep_is_held(&(wq->mutex)))
603 509
604 #ifdef CONFIG_DEBUG_OBJECTS_WORK 510 #ifdef CONFIG_DEBUG_OBJECTS_WORK
605 511
606 static const struct debug_obj_descr work_debug 512 static const struct debug_obj_descr work_debug_descr;
607 513
608 static void *work_debug_hint(void *addr) 514 static void *work_debug_hint(void *addr)
609 { 515 {
610 return ((struct work_struct *) addr)-> 516 return ((struct work_struct *) addr)->func;
611 } 517 }
612 518
613 static bool work_is_static_object(void *addr) 519 static bool work_is_static_object(void *addr)
614 { 520 {
615 struct work_struct *work = addr; 521 struct work_struct *work = addr;
616 522
617 return test_bit(WORK_STRUCT_STATIC_BIT 523 return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work));
618 } 524 }
619 525
620 /* 526 /*
621 * fixup_init is called when: 527 * fixup_init is called when:
622 * - an active object is initialized 528 * - an active object is initialized
623 */ 529 */
624 static bool work_fixup_init(void *addr, enum d 530 static bool work_fixup_init(void *addr, enum debug_obj_state state)
625 { 531 {
626 struct work_struct *work = addr; 532 struct work_struct *work = addr;
627 533
628 switch (state) { 534 switch (state) {
629 case ODEBUG_STATE_ACTIVE: 535 case ODEBUG_STATE_ACTIVE:
630 cancel_work_sync(work); 536 cancel_work_sync(work);
631 debug_object_init(work, &work_ 537 debug_object_init(work, &work_debug_descr);
632 return true; 538 return true;
633 default: 539 default:
634 return false; 540 return false;
635 } 541 }
636 } 542 }
637 543
638 /* 544 /*
639 * fixup_free is called when: 545 * fixup_free is called when:
640 * - an active object is freed 546 * - an active object is freed
641 */ 547 */
642 static bool work_fixup_free(void *addr, enum d 548 static bool work_fixup_free(void *addr, enum debug_obj_state state)
643 { 549 {
644 struct work_struct *work = addr; 550 struct work_struct *work = addr;
645 551
646 switch (state) { 552 switch (state) {
647 case ODEBUG_STATE_ACTIVE: 553 case ODEBUG_STATE_ACTIVE:
648 cancel_work_sync(work); 554 cancel_work_sync(work);
649 debug_object_free(work, &work_ 555 debug_object_free(work, &work_debug_descr);
650 return true; 556 return true;
651 default: 557 default:
652 return false; 558 return false;
653 } 559 }
654 } 560 }
655 561
656 static const struct debug_obj_descr work_debug 562 static const struct debug_obj_descr work_debug_descr = {
657 .name = "work_struct", 563 .name = "work_struct",
658 .debug_hint = work_debug_hint, 564 .debug_hint = work_debug_hint,
659 .is_static_object = work_is_static_obj 565 .is_static_object = work_is_static_object,
660 .fixup_init = work_fixup_init, 566 .fixup_init = work_fixup_init,
661 .fixup_free = work_fixup_free, 567 .fixup_free = work_fixup_free,
662 }; 568 };
663 569
664 static inline void debug_work_activate(struct 570 static inline void debug_work_activate(struct work_struct *work)
665 { 571 {
666 debug_object_activate(work, &work_debu 572 debug_object_activate(work, &work_debug_descr);
667 } 573 }
668 574
669 static inline void debug_work_deactivate(struc 575 static inline void debug_work_deactivate(struct work_struct *work)
670 { 576 {
671 debug_object_deactivate(work, &work_de 577 debug_object_deactivate(work, &work_debug_descr);
672 } 578 }
673 579
674 void __init_work(struct work_struct *work, int 580 void __init_work(struct work_struct *work, int onstack)
675 { 581 {
676 if (onstack) 582 if (onstack)
677 debug_object_init_on_stack(wor 583 debug_object_init_on_stack(work, &work_debug_descr);
678 else 584 else
679 debug_object_init(work, &work_ 585 debug_object_init(work, &work_debug_descr);
680 } 586 }
681 EXPORT_SYMBOL_GPL(__init_work); 587 EXPORT_SYMBOL_GPL(__init_work);
682 588
683 void destroy_work_on_stack(struct work_struct 589 void destroy_work_on_stack(struct work_struct *work)
684 { 590 {
685 debug_object_free(work, &work_debug_de 591 debug_object_free(work, &work_debug_descr);
686 } 592 }
687 EXPORT_SYMBOL_GPL(destroy_work_on_stack); 593 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
688 594
689 void destroy_delayed_work_on_stack(struct dela 595 void destroy_delayed_work_on_stack(struct delayed_work *work)
690 { 596 {
691 destroy_timer_on_stack(&work->timer); 597 destroy_timer_on_stack(&work->timer);
692 debug_object_free(&work->work, &work_d 598 debug_object_free(&work->work, &work_debug_descr);
693 } 599 }
694 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stac 600 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
695 601
696 #else 602 #else
697 static inline void debug_work_activate(struct 603 static inline void debug_work_activate(struct work_struct *work) { }
698 static inline void debug_work_deactivate(struc 604 static inline void debug_work_deactivate(struct work_struct *work) { }
699 #endif 605 #endif
700 606
701 /** 607 /**
702 * worker_pool_assign_id - allocate ID and ass 608 * worker_pool_assign_id - allocate ID and assign it to @pool
703 * @pool: the pool pointer of interest 609 * @pool: the pool pointer of interest
704 * 610 *
705 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) 611 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
706 * successfully, -errno on failure. 612 * successfully, -errno on failure.
707 */ 613 */
708 static int worker_pool_assign_id(struct worker 614 static int worker_pool_assign_id(struct worker_pool *pool)
709 { 615 {
710 int ret; 616 int ret;
711 617
712 lockdep_assert_held(&wq_pool_mutex); 618 lockdep_assert_held(&wq_pool_mutex);
713 619
714 ret = idr_alloc(&worker_pool_idr, pool 620 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
715 GFP_KERNEL); 621 GFP_KERNEL);
716 if (ret >= 0) { 622 if (ret >= 0) {
717 pool->id = ret; 623 pool->id = ret;
718 return 0; 624 return 0;
719 } 625 }
720 return ret; 626 return ret;
721 } 627 }
722 628
723 static struct pool_workqueue __rcu ** <<
724 unbound_pwq_slot(struct workqueue_struct *wq, <<
725 { <<
726 if (cpu >= 0) <<
727 return per_cpu_ptr(wq->cpu_pwq, <<
728 else <<
729 return &wq->dfl_pwq; <<
730 } <<
731 <<
732 /* @cpu < 0 for dfl_pwq */ <<
733 static struct pool_workqueue *unbound_pwq(stru <<
734 { <<
735 return rcu_dereference_check(*unbound_ <<
736 lockdep_i <<
737 lockdep_i <<
738 } <<
739 <<
740 /** <<
741 * unbound_effective_cpumask - effective cpuma <<
742 * @wq: workqueue of interest <<
743 * <<
744 * @wq->unbound_attrs->cpumask contains the cp <<
745 * is masked with wq_unbound_cpumask to determ <<
746 * default pwq is always mapped to the pool wi <<
747 */ <<
748 static struct cpumask *unbound_effective_cpuma <<
749 { <<
750 return unbound_pwq(wq, -1)->pool->attr <<
751 } <<
752 <<
753 static unsigned int work_color_to_flags(int co 629 static unsigned int work_color_to_flags(int color)
754 { 630 {
755 return color << WORK_STRUCT_COLOR_SHIF 631 return color << WORK_STRUCT_COLOR_SHIFT;
756 } 632 }
757 633
758 static int get_work_color(unsigned long work_d 634 static int get_work_color(unsigned long work_data)
759 { 635 {
760 return (work_data >> WORK_STRUCT_COLOR 636 return (work_data >> WORK_STRUCT_COLOR_SHIFT) &
761 ((1 << WORK_STRUCT_COLOR_BITS) 637 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
762 } 638 }
763 639
764 static int work_next_color(int color) 640 static int work_next_color(int color)
765 { 641 {
766 return (color + 1) % WORK_NR_COLORS; 642 return (color + 1) % WORK_NR_COLORS;
767 } 643 }
768 644
769 static unsigned long pool_offq_flags(struct wo <<
770 { <<
771 return (pool->flags & POOL_BH) ? WORK_ <<
772 } <<
773 <<
774 /* 645 /*
775 * While queued, %WORK_STRUCT_PWQ is set and n 646 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
776 * contain the pointer to the queued pwq. Onc 647 * contain the pointer to the queued pwq. Once execution starts, the flag
777 * is cleared and the high bits contain OFFQ f 648 * is cleared and the high bits contain OFFQ flags and pool ID.
778 * 649 *
779 * set_work_pwq(), set_work_pool_and_clear_pen !! 650 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
780 * can be used to set the pwq, pool or clear w !! 651 * and clear_work_data() can be used to set the pwq, pool or clear
781 * only be called while the work is owned - ie !! 652 * work->data. These functions should only be called while the work is
>> 653 * owned - ie. while the PENDING bit is set.
782 * 654 *
783 * get_work_pool() and get_work_pwq() can be u 655 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
784 * corresponding to a work. Pool is available 656 * corresponding to a work. Pool is available once the work has been
785 * queued anywhere after initialization until 657 * queued anywhere after initialization until it is sync canceled. pwq is
786 * available only while the work item is queue 658 * available only while the work item is queued.
>> 659 *
>> 660 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
>> 661 * canceled. While being canceled, a work item may have its PENDING set
>> 662 * but stay off timer and worklist for arbitrarily long and nobody should
>> 663 * try to steal the PENDING bit.
787 */ 664 */
788 static inline void set_work_data(struct work_s !! 665 static inline void set_work_data(struct work_struct *work, unsigned long data,
>> 666 unsigned long flags)
789 { 667 {
790 WARN_ON_ONCE(!work_pending(work)); 668 WARN_ON_ONCE(!work_pending(work));
791 atomic_long_set(&work->data, data | wo !! 669 atomic_long_set(&work->data, data | flags | work_static(work));
792 } 670 }
793 671
794 static void set_work_pwq(struct work_struct *w 672 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
795 unsigned long flags) !! 673 unsigned long extra_flags)
796 { 674 {
797 set_work_data(work, (unsigned long)pwq !! 675 set_work_data(work, (unsigned long)pwq,
798 WORK_STRUCT_PWQ | flags) !! 676 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
799 } 677 }
800 678
801 static void set_work_pool_and_keep_pending(str 679 static void set_work_pool_and_keep_pending(struct work_struct *work,
802 int !! 680 int pool_id)
803 { 681 {
804 set_work_data(work, ((unsigned long)po !! 682 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
805 WORK_STRUCT_PENDING | fl !! 683 WORK_STRUCT_PENDING);
806 } 684 }
807 685
808 static void set_work_pool_and_clear_pending(st 686 static void set_work_pool_and_clear_pending(struct work_struct *work,
809 in !! 687 int pool_id)
810 { 688 {
811 /* 689 /*
812 * The following wmb is paired with th 690 * The following wmb is paired with the implied mb in
813 * test_and_set_bit(PENDING) and ensur 691 * test_and_set_bit(PENDING) and ensures all updates to @work made
814 * here are visible to and precede any 692 * here are visible to and precede any updates by the next PENDING
815 * owner. 693 * owner.
816 */ 694 */
817 smp_wmb(); 695 smp_wmb();
818 set_work_data(work, ((unsigned long)po !! 696 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
819 flags); <<
820 /* 697 /*
821 * The following mb guarantees that pr 698 * The following mb guarantees that previous clear of a PENDING bit
822 * will not be reordered with any spec 699 * will not be reordered with any speculative LOADS or STORES from
823 * work->current_func, which is execut 700 * work->current_func, which is executed afterwards. This possible
824 * reordering can lead to a missed exe 701 * reordering can lead to a missed execution on attempt to queue
825 * the same @work. E.g. consider this 702 * the same @work. E.g. consider this case:
826 * 703 *
827 * CPU#0 CPU 704 * CPU#0 CPU#1
828 * ---------------------------- --- 705 * ---------------------------- --------------------------------
829 * 706 *
830 * 1 STORE event_indicated 707 * 1 STORE event_indicated
831 * 2 queue_work_on() { 708 * 2 queue_work_on() {
832 * 3 test_and_set_bit(PENDING) 709 * 3 test_and_set_bit(PENDING)
833 * 4 } set 710 * 4 } set_..._and_clear_pending() {
834 * 5 s 711 * 5 set_work_data() # clear bit
835 * 6 s 712 * 6 smp_mb()
836 * 7 wor 713 * 7 work->current_func() {
837 * 8 714 * 8 LOAD event_indicated
838 * } 715 * }
839 * 716 *
840 * Without an explicit full barrier sp 717 * Without an explicit full barrier speculative LOAD on line 8 can
841 * be executed before CPU#0 does STORE 718 * be executed before CPU#0 does STORE on line 1. If that happens,
842 * CPU#0 observes the PENDING bit is s 719 * CPU#0 observes the PENDING bit is still set and new execution of
843 * a @work is not queued in a hope, th 720 * a @work is not queued in a hope, that CPU#1 will eventually
844 * finish the queued @work. Meanwhile 721 * finish the queued @work. Meanwhile CPU#1 does not see
845 * event_indicated is set, because spe 722 * event_indicated is set, because speculative LOAD was executed
846 * before actual STORE. 723 * before actual STORE.
847 */ 724 */
848 smp_mb(); 725 smp_mb();
849 } 726 }
850 727
>> 728 static void clear_work_data(struct work_struct *work)
>> 729 {
>> 730 smp_wmb(); /* see set_work_pool_and_clear_pending() */
>> 731 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
>> 732 }
>> 733
851 static inline struct pool_workqueue *work_stru 734 static inline struct pool_workqueue *work_struct_pwq(unsigned long data)
852 { 735 {
853 return (struct pool_workqueue *)(data !! 736 return (struct pool_workqueue *)(data & WORK_STRUCT_WQ_DATA_MASK);
854 } 737 }
855 738
856 static struct pool_workqueue *get_work_pwq(str 739 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
857 { 740 {
858 unsigned long data = atomic_long_read( 741 unsigned long data = atomic_long_read(&work->data);
859 742
860 if (data & WORK_STRUCT_PWQ) 743 if (data & WORK_STRUCT_PWQ)
861 return work_struct_pwq(data); 744 return work_struct_pwq(data);
862 else 745 else
863 return NULL; 746 return NULL;
864 } 747 }
865 748
866 /** 749 /**
867 * get_work_pool - return the worker_pool a gi 750 * get_work_pool - return the worker_pool a given work was associated with
868 * @work: the work item of interest 751 * @work: the work item of interest
869 * 752 *
870 * Pools are created and destroyed under wq_po 753 * Pools are created and destroyed under wq_pool_mutex, and allows read
871 * access under RCU read lock. As such, this 754 * access under RCU read lock. As such, this function should be
872 * called under wq_pool_mutex or inside of a r 755 * called under wq_pool_mutex or inside of a rcu_read_lock() region.
873 * 756 *
874 * All fields of the returned pool are accessi 757 * All fields of the returned pool are accessible as long as the above
875 * mentioned locking is in effect. If the ret 758 * mentioned locking is in effect. If the returned pool needs to be used
876 * beyond the critical section, the caller is 759 * beyond the critical section, the caller is responsible for ensuring the
877 * returned pool is and stays online. 760 * returned pool is and stays online.
878 * 761 *
879 * Return: The worker_pool @work was last asso 762 * Return: The worker_pool @work was last associated with. %NULL if none.
880 */ 763 */
881 static struct worker_pool *get_work_pool(struc 764 static struct worker_pool *get_work_pool(struct work_struct *work)
882 { 765 {
883 unsigned long data = atomic_long_read( 766 unsigned long data = atomic_long_read(&work->data);
884 int pool_id; 767 int pool_id;
885 768
886 assert_rcu_or_pool_mutex(); 769 assert_rcu_or_pool_mutex();
887 770
888 if (data & WORK_STRUCT_PWQ) 771 if (data & WORK_STRUCT_PWQ)
889 return work_struct_pwq(data)-> 772 return work_struct_pwq(data)->pool;
890 773
891 pool_id = data >> WORK_OFFQ_POOL_SHIFT 774 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
892 if (pool_id == WORK_OFFQ_POOL_NONE) 775 if (pool_id == WORK_OFFQ_POOL_NONE)
893 return NULL; 776 return NULL;
894 777
895 return idr_find(&worker_pool_idr, pool 778 return idr_find(&worker_pool_idr, pool_id);
896 } 779 }
897 780
898 static unsigned long shift_and_mask(unsigned l !! 781 /**
>> 782 * get_work_pool_id - return the worker pool ID a given work is associated with
>> 783 * @work: the work item of interest
>> 784 *
>> 785 * Return: The worker_pool ID @work was last associated with.
>> 786 * %WORK_OFFQ_POOL_NONE if none.
>> 787 */
>> 788 static int get_work_pool_id(struct work_struct *work)
899 { 789 {
900 return (v >> shift) & ((1U << bits) - !! 790 unsigned long data = atomic_long_read(&work->data);
>> 791
>> 792 if (data & WORK_STRUCT_PWQ)
>> 793 return work_struct_pwq(data)->pool->id;
>> 794
>> 795 return data >> WORK_OFFQ_POOL_SHIFT;
901 } 796 }
902 797
903 static void work_offqd_unpack(struct work_offq !! 798 static void mark_work_canceling(struct work_struct *work)
904 { 799 {
905 WARN_ON_ONCE(data & WORK_STRUCT_PWQ); !! 800 unsigned long pool_id = get_work_pool_id(work);
906 801
907 offqd->pool_id = shift_and_mask(data, !! 802 pool_id <<= WORK_OFFQ_POOL_SHIFT;
908 WORK_O !! 803 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
909 offqd->disable = shift_and_mask(data, <<
910 WORK_O <<
911 offqd->flags = data & WORK_OFFQ_FLAG_M <<
912 } 804 }
913 805
914 static unsigned long work_offqd_pack_flags(str !! 806 static bool work_is_canceling(struct work_struct *work)
915 { 807 {
916 return ((unsigned long)offqd->disable !! 808 unsigned long data = atomic_long_read(&work->data);
917 ((unsigned long)offqd->flags); !! 809
>> 810 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
918 } 811 }
919 812
920 /* 813 /*
921 * Policy functions. These define the policie 814 * Policy functions. These define the policies on how the global worker
922 * pools are managed. Unless noted otherwise, 815 * pools are managed. Unless noted otherwise, these functions assume that
923 * they're being called with pool->lock held. 816 * they're being called with pool->lock held.
924 */ 817 */
925 818
926 /* 819 /*
927 * Need to wake up a worker? Called from anyt 820 * Need to wake up a worker? Called from anything but currently
928 * running workers. 821 * running workers.
929 * 822 *
930 * Note that, because unbound workers never co 823 * Note that, because unbound workers never contribute to nr_running, this
931 * function will always return %true for unbou 824 * function will always return %true for unbound pools as long as the
932 * worklist isn't empty. 825 * worklist isn't empty.
933 */ 826 */
934 static bool need_more_worker(struct worker_poo 827 static bool need_more_worker(struct worker_pool *pool)
935 { 828 {
936 return !list_empty(&pool->worklist) && 829 return !list_empty(&pool->worklist) && !pool->nr_running;
937 } 830 }
938 831
939 /* Can I start working? Called from busy but 832 /* Can I start working? Called from busy but !running workers. */
940 static bool may_start_working(struct worker_po 833 static bool may_start_working(struct worker_pool *pool)
941 { 834 {
942 return pool->nr_idle; 835 return pool->nr_idle;
943 } 836 }
944 837
945 /* Do I need to keep working? Called from cur 838 /* Do I need to keep working? Called from currently running workers. */
946 static bool keep_working(struct worker_pool *p 839 static bool keep_working(struct worker_pool *pool)
947 { 840 {
948 return !list_empty(&pool->worklist) && 841 return !list_empty(&pool->worklist) && (pool->nr_running <= 1);
949 } 842 }
950 843
951 /* Do we need a new worker? Called from manag 844 /* Do we need a new worker? Called from manager. */
952 static bool need_to_create_worker(struct worke 845 static bool need_to_create_worker(struct worker_pool *pool)
953 { 846 {
954 return need_more_worker(pool) && !may_ 847 return need_more_worker(pool) && !may_start_working(pool);
955 } 848 }
956 849
957 /* Do we have too many workers and should some 850 /* Do we have too many workers and should some go away? */
958 static bool too_many_workers(struct worker_poo 851 static bool too_many_workers(struct worker_pool *pool)
959 { 852 {
960 bool managing = pool->flags & POOL_MAN 853 bool managing = pool->flags & POOL_MANAGER_ACTIVE;
961 int nr_idle = pool->nr_idle + managing 854 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
962 int nr_busy = pool->nr_workers - nr_id 855 int nr_busy = pool->nr_workers - nr_idle;
963 856
964 return nr_idle > 2 && (nr_idle - 2) * 857 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
965 } 858 }
966 859
967 /** 860 /**
968 * worker_set_flags - set worker flags and adj 861 * worker_set_flags - set worker flags and adjust nr_running accordingly
969 * @worker: self 862 * @worker: self
970 * @flags: flags to set 863 * @flags: flags to set
971 * 864 *
972 * Set @flags in @worker->flags and adjust nr_ 865 * Set @flags in @worker->flags and adjust nr_running accordingly.
973 */ 866 */
974 static inline void worker_set_flags(struct wor 867 static inline void worker_set_flags(struct worker *worker, unsigned int flags)
975 { 868 {
976 struct worker_pool *pool = worker->poo 869 struct worker_pool *pool = worker->pool;
977 870
978 lockdep_assert_held(&pool->lock); 871 lockdep_assert_held(&pool->lock);
979 872
980 /* If transitioning into NOT_RUNNING, 873 /* If transitioning into NOT_RUNNING, adjust nr_running. */
981 if ((flags & WORKER_NOT_RUNNING) && 874 if ((flags & WORKER_NOT_RUNNING) &&
982 !(worker->flags & WORKER_NOT_RUNNI 875 !(worker->flags & WORKER_NOT_RUNNING)) {
983 pool->nr_running--; 876 pool->nr_running--;
984 } 877 }
985 878
986 worker->flags |= flags; 879 worker->flags |= flags;
987 } 880 }
988 881
989 /** 882 /**
990 * worker_clr_flags - clear worker flags and a 883 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
991 * @worker: self 884 * @worker: self
992 * @flags: flags to clear 885 * @flags: flags to clear
993 * 886 *
994 * Clear @flags in @worker->flags and adjust n 887 * Clear @flags in @worker->flags and adjust nr_running accordingly.
995 */ 888 */
996 static inline void worker_clr_flags(struct wor 889 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
997 { 890 {
998 struct worker_pool *pool = worker->poo 891 struct worker_pool *pool = worker->pool;
999 unsigned int oflags = worker->flags; 892 unsigned int oflags = worker->flags;
1000 893
1001 lockdep_assert_held(&pool->lock); 894 lockdep_assert_held(&pool->lock);
1002 895
1003 worker->flags &= ~flags; 896 worker->flags &= ~flags;
1004 897
1005 /* 898 /*
1006 * If transitioning out of NOT_RUNNIN 899 * If transitioning out of NOT_RUNNING, increment nr_running. Note
1007 * that the nested NOT_RUNNING is not 900 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
1008 * of multiple flags, not a single fl 901 * of multiple flags, not a single flag.
1009 */ 902 */
1010 if ((flags & WORKER_NOT_RUNNING) && ( 903 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
1011 if (!(worker->flags & WORKER_ 904 if (!(worker->flags & WORKER_NOT_RUNNING))
1012 pool->nr_running++; 905 pool->nr_running++;
1013 } 906 }
1014 907
1015 /* Return the first idle worker. Called with 908 /* Return the first idle worker. Called with pool->lock held. */
1016 static struct worker *first_idle_worker(struc 909 static struct worker *first_idle_worker(struct worker_pool *pool)
1017 { 910 {
1018 if (unlikely(list_empty(&pool->idle_l 911 if (unlikely(list_empty(&pool->idle_list)))
1019 return NULL; 912 return NULL;
1020 913
1021 return list_first_entry(&pool->idle_l 914 return list_first_entry(&pool->idle_list, struct worker, entry);
1022 } 915 }
1023 916
1024 /** 917 /**
1025 * worker_enter_idle - enter idle state 918 * worker_enter_idle - enter idle state
1026 * @worker: worker which is entering idle sta 919 * @worker: worker which is entering idle state
1027 * 920 *
1028 * @worker is entering idle state. Update st 921 * @worker is entering idle state. Update stats and idle timer if
1029 * necessary. 922 * necessary.
1030 * 923 *
1031 * LOCKING: 924 * LOCKING:
1032 * raw_spin_lock_irq(pool->lock). 925 * raw_spin_lock_irq(pool->lock).
1033 */ 926 */
1034 static void worker_enter_idle(struct worker * 927 static void worker_enter_idle(struct worker *worker)
1035 { 928 {
1036 struct worker_pool *pool = worker->po 929 struct worker_pool *pool = worker->pool;
1037 930
1038 if (WARN_ON_ONCE(worker->flags & WORK 931 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1039 WARN_ON_ONCE(!list_empty(&worker- 932 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1040 (worker->hentry.next 933 (worker->hentry.next || worker->hentry.pprev)))
1041 return; 934 return;
1042 935
1043 /* can't use worker_set_flags(), also 936 /* can't use worker_set_flags(), also called from create_worker() */
1044 worker->flags |= WORKER_IDLE; 937 worker->flags |= WORKER_IDLE;
1045 pool->nr_idle++; 938 pool->nr_idle++;
1046 worker->last_active = jiffies; 939 worker->last_active = jiffies;
1047 940
1048 /* idle_list is LIFO */ 941 /* idle_list is LIFO */
1049 list_add(&worker->entry, &pool->idle_ 942 list_add(&worker->entry, &pool->idle_list);
1050 943
1051 if (too_many_workers(pool) && !timer_ 944 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1052 mod_timer(&pool->idle_timer, 945 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1053 946
1054 /* Sanity check nr_running. */ 947 /* Sanity check nr_running. */
1055 WARN_ON_ONCE(pool->nr_workers == pool 948 WARN_ON_ONCE(pool->nr_workers == pool->nr_idle && pool->nr_running);
1056 } 949 }
1057 950
1058 /** 951 /**
1059 * worker_leave_idle - leave idle state 952 * worker_leave_idle - leave idle state
1060 * @worker: worker which is leaving idle stat 953 * @worker: worker which is leaving idle state
1061 * 954 *
1062 * @worker is leaving idle state. Update sta 955 * @worker is leaving idle state. Update stats.
1063 * 956 *
1064 * LOCKING: 957 * LOCKING:
1065 * raw_spin_lock_irq(pool->lock). 958 * raw_spin_lock_irq(pool->lock).
1066 */ 959 */
1067 static void worker_leave_idle(struct worker * 960 static void worker_leave_idle(struct worker *worker)
1068 { 961 {
1069 struct worker_pool *pool = worker->po 962 struct worker_pool *pool = worker->pool;
1070 963
1071 if (WARN_ON_ONCE(!(worker->flags & WO 964 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1072 return; 965 return;
1073 worker_clr_flags(worker, WORKER_IDLE) 966 worker_clr_flags(worker, WORKER_IDLE);
1074 pool->nr_idle--; 967 pool->nr_idle--;
1075 list_del_init(&worker->entry); 968 list_del_init(&worker->entry);
1076 } 969 }
1077 970
1078 /** 971 /**
1079 * find_worker_executing_work - find worker w 972 * find_worker_executing_work - find worker which is executing a work
1080 * @pool: pool of interest 973 * @pool: pool of interest
1081 * @work: work to find worker for 974 * @work: work to find worker for
1082 * 975 *
1083 * Find a worker which is executing @work on 976 * Find a worker which is executing @work on @pool by searching
1084 * @pool->busy_hash which is keyed by the add 977 * @pool->busy_hash which is keyed by the address of @work. For a worker
1085 * to match, its current execution should mat 978 * to match, its current execution should match the address of @work and
1086 * its work function. This is to avoid unwan 979 * its work function. This is to avoid unwanted dependency between
1087 * unrelated work executions through a work i 980 * unrelated work executions through a work item being recycled while still
1088 * being executed. 981 * being executed.
1089 * 982 *
1090 * This is a bit tricky. A work item may be 983 * This is a bit tricky. A work item may be freed once its execution
1091 * starts and nothing prevents the freed area 984 * starts and nothing prevents the freed area from being recycled for
1092 * another work item. If the same work item 985 * another work item. If the same work item address ends up being reused
1093 * before the original execution finishes, wo 986 * before the original execution finishes, workqueue will identify the
1094 * recycled work item as currently executing 987 * recycled work item as currently executing and make it wait until the
1095 * current execution finishes, introducing an 988 * current execution finishes, introducing an unwanted dependency.
1096 * 989 *
1097 * This function checks the work item address 990 * This function checks the work item address and work function to avoid
1098 * false positives. Note that this isn't com 991 * false positives. Note that this isn't complete as one may construct a
1099 * work function which can introduce dependen 992 * work function which can introduce dependency onto itself through a
1100 * recycled work item. Well, if somebody wan 993 * recycled work item. Well, if somebody wants to shoot oneself in the
1101 * foot that badly, there's only so much we c 994 * foot that badly, there's only so much we can do, and if such deadlock
1102 * actually occurs, it should be easy to loca 995 * actually occurs, it should be easy to locate the culprit work function.
1103 * 996 *
1104 * CONTEXT: 997 * CONTEXT:
1105 * raw_spin_lock_irq(pool->lock). 998 * raw_spin_lock_irq(pool->lock).
1106 * 999 *
1107 * Return: 1000 * Return:
1108 * Pointer to worker which is executing @work 1001 * Pointer to worker which is executing @work if found, %NULL
1109 * otherwise. 1002 * otherwise.
1110 */ 1003 */
1111 static struct worker *find_worker_executing_w 1004 static struct worker *find_worker_executing_work(struct worker_pool *pool,
1112 1005 struct work_struct *work)
1113 { 1006 {
1114 struct worker *worker; 1007 struct worker *worker;
1115 1008
1116 hash_for_each_possible(pool->busy_has 1009 hash_for_each_possible(pool->busy_hash, worker, hentry,
1117 (unsigned long 1010 (unsigned long)work)
1118 if (worker->current_work == w 1011 if (worker->current_work == work &&
1119 worker->current_func == w 1012 worker->current_func == work->func)
1120 return worker; 1013 return worker;
1121 1014
1122 return NULL; 1015 return NULL;
1123 } 1016 }
1124 1017
1125 /** 1018 /**
1126 * move_linked_works - move linked works to a 1019 * move_linked_works - move linked works to a list
1127 * @work: start of series of works to be sche 1020 * @work: start of series of works to be scheduled
1128 * @head: target list to append @work to 1021 * @head: target list to append @work to
1129 * @nextp: out parameter for nested worklist 1022 * @nextp: out parameter for nested worklist walking
1130 * 1023 *
1131 * Schedule linked works starting from @work 1024 * Schedule linked works starting from @work to @head. Work series to be
1132 * scheduled starts at @work and includes any 1025 * scheduled starts at @work and includes any consecutive work with
1133 * WORK_STRUCT_LINKED set in its predecessor. 1026 * WORK_STRUCT_LINKED set in its predecessor. See assign_work() for details on
1134 * @nextp. 1027 * @nextp.
1135 * 1028 *
1136 * CONTEXT: 1029 * CONTEXT:
1137 * raw_spin_lock_irq(pool->lock). 1030 * raw_spin_lock_irq(pool->lock).
1138 */ 1031 */
1139 static void move_linked_works(struct work_str 1032 static void move_linked_works(struct work_struct *work, struct list_head *head,
1140 struct work_str 1033 struct work_struct **nextp)
1141 { 1034 {
1142 struct work_struct *n; 1035 struct work_struct *n;
1143 1036
1144 /* 1037 /*
1145 * Linked worklist will always end be 1038 * Linked worklist will always end before the end of the list,
1146 * use NULL for list head. 1039 * use NULL for list head.
1147 */ 1040 */
1148 list_for_each_entry_safe_from(work, n 1041 list_for_each_entry_safe_from(work, n, NULL, entry) {
1149 list_move_tail(&work->entry, 1042 list_move_tail(&work->entry, head);
1150 if (!(*work_data_bits(work) & 1043 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1151 break; 1044 break;
1152 } 1045 }
1153 1046
1154 /* 1047 /*
1155 * If we're already inside safe list 1048 * If we're already inside safe list traversal and have moved
1156 * multiple works to the scheduled qu 1049 * multiple works to the scheduled queue, the next position
1157 * needs to be updated. 1050 * needs to be updated.
1158 */ 1051 */
1159 if (nextp) 1052 if (nextp)
1160 *nextp = n; 1053 *nextp = n;
1161 } 1054 }
1162 1055
1163 /** 1056 /**
1164 * assign_work - assign a work item and its l 1057 * assign_work - assign a work item and its linked work items to a worker
1165 * @work: work to assign 1058 * @work: work to assign
1166 * @worker: worker to assign to 1059 * @worker: worker to assign to
1167 * @nextp: out parameter for nested worklist 1060 * @nextp: out parameter for nested worklist walking
1168 * 1061 *
1169 * Assign @work and its linked work items to 1062 * Assign @work and its linked work items to @worker. If @work is already being
1170 * executed by another worker in the same poo 1063 * executed by another worker in the same pool, it'll be punted there.
1171 * 1064 *
1172 * If @nextp is not NULL, it's updated to poi 1065 * If @nextp is not NULL, it's updated to point to the next work of the last
1173 * scheduled work. This allows assign_work() 1066 * scheduled work. This allows assign_work() to be nested inside
1174 * list_for_each_entry_safe(). 1067 * list_for_each_entry_safe().
1175 * 1068 *
1176 * Returns %true if @work was successfully as 1069 * Returns %true if @work was successfully assigned to @worker. %false if @work
1177 * was punted to another worker already execu 1070 * was punted to another worker already executing it.
1178 */ 1071 */
1179 static bool assign_work(struct work_struct *w 1072 static bool assign_work(struct work_struct *work, struct worker *worker,
1180 struct work_struct ** 1073 struct work_struct **nextp)
1181 { 1074 {
1182 struct worker_pool *pool = worker->po 1075 struct worker_pool *pool = worker->pool;
1183 struct worker *collision; 1076 struct worker *collision;
1184 1077
1185 lockdep_assert_held(&pool->lock); 1078 lockdep_assert_held(&pool->lock);
1186 1079
1187 /* 1080 /*
1188 * A single work shouldn't be execute 1081 * A single work shouldn't be executed concurrently by multiple workers.
1189 * __queue_work() ensures that @work 1082 * __queue_work() ensures that @work doesn't jump to a different pool
1190 * while still running in the previou 1083 * while still running in the previous pool. Here, we should ensure that
1191 * @work is not executed concurrently 1084 * @work is not executed concurrently by multiple workers from the same
1192 * pool. Check whether anyone is alre 1085 * pool. Check whether anyone is already processing the work. If so,
1193 * defer the work to the currently ex 1086 * defer the work to the currently executing one.
1194 */ 1087 */
1195 collision = find_worker_executing_wor 1088 collision = find_worker_executing_work(pool, work);
1196 if (unlikely(collision)) { 1089 if (unlikely(collision)) {
1197 move_linked_works(work, &coll 1090 move_linked_works(work, &collision->scheduled, nextp);
1198 return false; 1091 return false;
1199 } 1092 }
1200 1093
1201 move_linked_works(work, &worker->sche 1094 move_linked_works(work, &worker->scheduled, nextp);
1202 return true; 1095 return true;
1203 } 1096 }
1204 1097
1205 static struct irq_work *bh_pool_irq_work(stru <<
1206 { <<
1207 int high = pool->attrs->nice == HIGHP <<
1208 <<
1209 return &per_cpu(bh_pool_irq_works, po <<
1210 } <<
1211 <<
1212 static void kick_bh_pool(struct worker_pool * <<
1213 { <<
1214 #ifdef CONFIG_SMP <<
1215 /* see drain_dead_softirq_workfn() fo <<
1216 if (unlikely(pool->cpu != smp_process <<
1217 !(pool->flags & POOL_BH_ <<
1218 irq_work_queue_on(bh_pool_irq <<
1219 return; <<
1220 } <<
1221 #endif <<
1222 if (pool->attrs->nice == HIGHPRI_NICE <<
1223 raise_softirq_irqoff(HI_SOFTI <<
1224 else <<
1225 raise_softirq_irqoff(TASKLET_ <<
1226 } <<
1227 <<
1228 /** 1098 /**
1229 * kick_pool - wake up an idle worker if nece 1099 * kick_pool - wake up an idle worker if necessary
1230 * @pool: pool to kick 1100 * @pool: pool to kick
1231 * 1101 *
1232 * @pool may have pending work items. Wake up 1102 * @pool may have pending work items. Wake up worker if necessary. Returns
1233 * whether a worker was woken up. 1103 * whether a worker was woken up.
1234 */ 1104 */
1235 static bool kick_pool(struct worker_pool *poo 1105 static bool kick_pool(struct worker_pool *pool)
1236 { 1106 {
1237 struct worker *worker = first_idle_wo 1107 struct worker *worker = first_idle_worker(pool);
1238 struct task_struct *p; 1108 struct task_struct *p;
1239 1109
1240 lockdep_assert_held(&pool->lock); 1110 lockdep_assert_held(&pool->lock);
1241 1111
1242 if (!need_more_worker(pool) || !worke 1112 if (!need_more_worker(pool) || !worker)
1243 return false; 1113 return false;
1244 1114
1245 if (pool->flags & POOL_BH) { <<
1246 kick_bh_pool(pool); <<
1247 return true; <<
1248 } <<
1249 <<
1250 p = worker->task; 1115 p = worker->task;
1251 1116
1252 #ifdef CONFIG_SMP 1117 #ifdef CONFIG_SMP
1253 /* 1118 /*
1254 * Idle @worker is about to execute @ 1119 * Idle @worker is about to execute @work and waking up provides an
1255 * opportunity to migrate @worker at 1120 * opportunity to migrate @worker at a lower cost by setting the task's
1256 * wake_cpu field. Let's see if we wa 1121 * wake_cpu field. Let's see if we want to move @worker to improve
1257 * execution locality. 1122 * execution locality.
1258 * 1123 *
1259 * We're waking the worker that went 1124 * We're waking the worker that went idle the latest and there's some
1260 * chance that @worker is marked idle 1125 * chance that @worker is marked idle but hasn't gone off CPU yet. If
1261 * so, setting the wake_cpu won't do 1126 * so, setting the wake_cpu won't do anything. As this is a best-effort
1262 * optimization and the race window i 1127 * optimization and the race window is narrow, let's leave as-is for
1263 * now. If this becomes pronounced, w 1128 * now. If this becomes pronounced, we can skip over workers which are
1264 * still on cpu when picking an idle 1129 * still on cpu when picking an idle worker.
1265 * 1130 *
1266 * If @pool has non-strict affinity, 1131 * If @pool has non-strict affinity, @worker might have ended up outside
1267 * its affinity scope. Repatriate. 1132 * its affinity scope. Repatriate.
1268 */ 1133 */
1269 if (!pool->attrs->affn_strict && 1134 if (!pool->attrs->affn_strict &&
1270 !cpumask_test_cpu(p->wake_cpu, po 1135 !cpumask_test_cpu(p->wake_cpu, pool->attrs->__pod_cpumask)) {
1271 struct work_struct *work = li 1136 struct work_struct *work = list_first_entry(&pool->worklist,
1272 1137 struct work_struct, entry);
1273 int wake_cpu = cpumask_any_an 1138 int wake_cpu = cpumask_any_and_distribute(pool->attrs->__pod_cpumask,
1274 1139 cpu_online_mask);
1275 if (wake_cpu < nr_cpu_ids) { 1140 if (wake_cpu < nr_cpu_ids) {
1276 p->wake_cpu = wake_cp 1141 p->wake_cpu = wake_cpu;
1277 get_work_pwq(work)->s 1142 get_work_pwq(work)->stats[PWQ_STAT_REPATRIATED]++;
1278 } 1143 }
1279 } 1144 }
1280 #endif 1145 #endif
1281 wake_up_process(p); 1146 wake_up_process(p);
1282 return true; 1147 return true;
1283 } 1148 }
1284 1149
1285 #ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT 1150 #ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT
1286 1151
1287 /* 1152 /*
1288 * Concurrency-managed per-cpu work items tha 1153 * Concurrency-managed per-cpu work items that hog CPU for longer than
1289 * wq_cpu_intensive_thresh_us trigger the aut 1154 * wq_cpu_intensive_thresh_us trigger the automatic CPU_INTENSIVE mechanism,
1290 * which prevents them from stalling other co 1155 * which prevents them from stalling other concurrency-managed work items. If a
1291 * work function keeps triggering this mechan 1156 * work function keeps triggering this mechanism, it's likely that the work item
1292 * should be using an unbound workqueue inste 1157 * should be using an unbound workqueue instead.
1293 * 1158 *
1294 * wq_cpu_intensive_report() tracks work func 1159 * wq_cpu_intensive_report() tracks work functions which trigger such conditions
1295 * and report them so that they can be examin 1160 * and report them so that they can be examined and converted to use unbound
1296 * workqueues as appropriate. To avoid floodi 1161 * workqueues as appropriate. To avoid flooding the console, each violating work
1297 * function is tracked and reported with expo 1162 * function is tracked and reported with exponential backoff.
1298 */ 1163 */
1299 #define WCI_MAX_ENTS 128 1164 #define WCI_MAX_ENTS 128
1300 1165
1301 struct wci_ent { 1166 struct wci_ent {
1302 work_func_t func; 1167 work_func_t func;
1303 atomic64_t cnt; 1168 atomic64_t cnt;
1304 struct hlist_node hash_node; 1169 struct hlist_node hash_node;
1305 }; 1170 };
1306 1171
1307 static struct wci_ent wci_ents[WCI_MAX_ENTS]; 1172 static struct wci_ent wci_ents[WCI_MAX_ENTS];
1308 static int wci_nr_ents; 1173 static int wci_nr_ents;
1309 static DEFINE_RAW_SPINLOCK(wci_lock); 1174 static DEFINE_RAW_SPINLOCK(wci_lock);
1310 static DEFINE_HASHTABLE(wci_hash, ilog2(WCI_M 1175 static DEFINE_HASHTABLE(wci_hash, ilog2(WCI_MAX_ENTS));
1311 1176
1312 static struct wci_ent *wci_find_ent(work_func 1177 static struct wci_ent *wci_find_ent(work_func_t func)
1313 { 1178 {
1314 struct wci_ent *ent; 1179 struct wci_ent *ent;
1315 1180
1316 hash_for_each_possible_rcu(wci_hash, 1181 hash_for_each_possible_rcu(wci_hash, ent, hash_node,
1317 (unsigned 1182 (unsigned long)func) {
1318 if (ent->func == func) 1183 if (ent->func == func)
1319 return ent; 1184 return ent;
1320 } 1185 }
1321 return NULL; 1186 return NULL;
1322 } 1187 }
1323 1188
1324 static void wq_cpu_intensive_report(work_func 1189 static void wq_cpu_intensive_report(work_func_t func)
1325 { 1190 {
1326 struct wci_ent *ent; 1191 struct wci_ent *ent;
1327 1192
1328 restart: 1193 restart:
1329 ent = wci_find_ent(func); 1194 ent = wci_find_ent(func);
1330 if (ent) { 1195 if (ent) {
1331 u64 cnt; 1196 u64 cnt;
1332 1197
1333 /* 1198 /*
1334 * Start reporting from the w !! 1199 * Start reporting from the fourth time and back off
1335 * exponentially. 1200 * exponentially.
1336 */ 1201 */
1337 cnt = atomic64_inc_return_rel 1202 cnt = atomic64_inc_return_relaxed(&ent->cnt);
1338 if (wq_cpu_intensive_warning_ !! 1203 if (cnt >= 4 && is_power_of_2(cnt))
1339 cnt >= wq_cpu_intensive_w <<
1340 is_power_of_2(cnt + 1 - w <<
1341 printk_deferred(KERN_ 1204 printk_deferred(KERN_WARNING "workqueue: %ps hogged CPU for >%luus %llu times, consider switching to WQ_UNBOUND\n",
1342 ent-> 1205 ent->func, wq_cpu_intensive_thresh_us,
1343 atomi 1206 atomic64_read(&ent->cnt));
1344 return; 1207 return;
1345 } 1208 }
1346 1209
1347 /* 1210 /*
1348 * @func is a new violation. Allocate 1211 * @func is a new violation. Allocate a new entry for it. If wcn_ents[]
1349 * is exhausted, something went reall 1212 * is exhausted, something went really wrong and we probably made enough
1350 * noise already. 1213 * noise already.
1351 */ 1214 */
1352 if (wci_nr_ents >= WCI_MAX_ENTS) 1215 if (wci_nr_ents >= WCI_MAX_ENTS)
1353 return; 1216 return;
1354 1217
1355 raw_spin_lock(&wci_lock); 1218 raw_spin_lock(&wci_lock);
1356 1219
1357 if (wci_nr_ents >= WCI_MAX_ENTS) { 1220 if (wci_nr_ents >= WCI_MAX_ENTS) {
1358 raw_spin_unlock(&wci_lock); 1221 raw_spin_unlock(&wci_lock);
1359 return; 1222 return;
1360 } 1223 }
1361 1224
1362 if (wci_find_ent(func)) { 1225 if (wci_find_ent(func)) {
1363 raw_spin_unlock(&wci_lock); 1226 raw_spin_unlock(&wci_lock);
1364 goto restart; 1227 goto restart;
1365 } 1228 }
1366 1229
1367 ent = &wci_ents[wci_nr_ents++]; 1230 ent = &wci_ents[wci_nr_ents++];
1368 ent->func = func; 1231 ent->func = func;
1369 atomic64_set(&ent->cnt, 0); !! 1232 atomic64_set(&ent->cnt, 1);
1370 hash_add_rcu(wci_hash, &ent->hash_nod 1233 hash_add_rcu(wci_hash, &ent->hash_node, (unsigned long)func);
1371 1234
1372 raw_spin_unlock(&wci_lock); 1235 raw_spin_unlock(&wci_lock);
1373 <<
1374 goto restart; <<
1375 } 1236 }
1376 1237
1377 #else /* CONFIG_WQ_CPU_INTENSIVE_REPORT */ 1238 #else /* CONFIG_WQ_CPU_INTENSIVE_REPORT */
1378 static void wq_cpu_intensive_report(work_func 1239 static void wq_cpu_intensive_report(work_func_t func) {}
1379 #endif /* CONFIG_WQ_CPU_INTENSIVE_REPORT */ 1240 #endif /* CONFIG_WQ_CPU_INTENSIVE_REPORT */
1380 1241
1381 /** 1242 /**
1382 * wq_worker_running - a worker is running ag 1243 * wq_worker_running - a worker is running again
1383 * @task: task waking up 1244 * @task: task waking up
1384 * 1245 *
1385 * This function is called when a worker retu 1246 * This function is called when a worker returns from schedule()
1386 */ 1247 */
1387 void wq_worker_running(struct task_struct *ta 1248 void wq_worker_running(struct task_struct *task)
1388 { 1249 {
1389 struct worker *worker = kthread_data( 1250 struct worker *worker = kthread_data(task);
1390 1251
1391 if (!READ_ONCE(worker->sleeping)) 1252 if (!READ_ONCE(worker->sleeping))
1392 return; 1253 return;
1393 1254
1394 /* 1255 /*
1395 * If preempted by unbind_workers() b 1256 * If preempted by unbind_workers() between the WORKER_NOT_RUNNING check
1396 * and the nr_running increment below 1257 * and the nr_running increment below, we may ruin the nr_running reset
1397 * and leave with an unexpected pool- 1258 * and leave with an unexpected pool->nr_running == 1 on the newly unbound
1398 * pool. Protect against such race. 1259 * pool. Protect against such race.
1399 */ 1260 */
1400 preempt_disable(); 1261 preempt_disable();
1401 if (!(worker->flags & WORKER_NOT_RUNN 1262 if (!(worker->flags & WORKER_NOT_RUNNING))
1402 worker->pool->nr_running++; 1263 worker->pool->nr_running++;
1403 preempt_enable(); 1264 preempt_enable();
1404 1265
1405 /* 1266 /*
1406 * CPU intensive auto-detection cares 1267 * CPU intensive auto-detection cares about how long a work item hogged
1407 * CPU without sleeping. Reset the st 1268 * CPU without sleeping. Reset the starting timestamp on wakeup.
1408 */ 1269 */
1409 worker->current_at = worker->task->se 1270 worker->current_at = worker->task->se.sum_exec_runtime;
1410 1271
1411 WRITE_ONCE(worker->sleeping, 0); 1272 WRITE_ONCE(worker->sleeping, 0);
1412 } 1273 }
1413 1274
1414 /** 1275 /**
1415 * wq_worker_sleeping - a worker is going to 1276 * wq_worker_sleeping - a worker is going to sleep
1416 * @task: task going to sleep 1277 * @task: task going to sleep
1417 * 1278 *
1418 * This function is called from schedule() wh 1279 * This function is called from schedule() when a busy worker is
1419 * going to sleep. 1280 * going to sleep.
1420 */ 1281 */
1421 void wq_worker_sleeping(struct task_struct *t 1282 void wq_worker_sleeping(struct task_struct *task)
1422 { 1283 {
1423 struct worker *worker = kthread_data( 1284 struct worker *worker = kthread_data(task);
1424 struct worker_pool *pool; 1285 struct worker_pool *pool;
1425 1286
1426 /* 1287 /*
1427 * Rescuers, which may not have all t 1288 * Rescuers, which may not have all the fields set up like normal
1428 * workers, also reach here, let's no 1289 * workers, also reach here, let's not access anything before
1429 * checking NOT_RUNNING. 1290 * checking NOT_RUNNING.
1430 */ 1291 */
1431 if (worker->flags & WORKER_NOT_RUNNIN 1292 if (worker->flags & WORKER_NOT_RUNNING)
1432 return; 1293 return;
1433 1294
1434 pool = worker->pool; 1295 pool = worker->pool;
1435 1296
1436 /* Return if preempted before wq_work 1297 /* Return if preempted before wq_worker_running() was reached */
1437 if (READ_ONCE(worker->sleeping)) 1298 if (READ_ONCE(worker->sleeping))
1438 return; 1299 return;
1439 1300
1440 WRITE_ONCE(worker->sleeping, 1); 1301 WRITE_ONCE(worker->sleeping, 1);
1441 raw_spin_lock_irq(&pool->lock); 1302 raw_spin_lock_irq(&pool->lock);
1442 1303
1443 /* 1304 /*
1444 * Recheck in case unbind_workers() p 1305 * Recheck in case unbind_workers() preempted us. We don't
1445 * want to decrement nr_running after 1306 * want to decrement nr_running after the worker is unbound
1446 * and nr_running has been reset. 1307 * and nr_running has been reset.
1447 */ 1308 */
1448 if (worker->flags & WORKER_NOT_RUNNIN 1309 if (worker->flags & WORKER_NOT_RUNNING) {
1449 raw_spin_unlock_irq(&pool->lo 1310 raw_spin_unlock_irq(&pool->lock);
1450 return; 1311 return;
1451 } 1312 }
1452 1313
1453 pool->nr_running--; 1314 pool->nr_running--;
1454 if (kick_pool(pool)) 1315 if (kick_pool(pool))
1455 worker->current_pwq->stats[PW 1316 worker->current_pwq->stats[PWQ_STAT_CM_WAKEUP]++;
1456 1317
1457 raw_spin_unlock_irq(&pool->lock); 1318 raw_spin_unlock_irq(&pool->lock);
1458 } 1319 }
1459 1320
1460 /** 1321 /**
1461 * wq_worker_tick - a scheduler tick occurred 1322 * wq_worker_tick - a scheduler tick occurred while a kworker is running
1462 * @task: task currently running 1323 * @task: task currently running
1463 * 1324 *
1464 * Called from sched_tick(). We're in the IRQ !! 1325 * Called from scheduler_tick(). We're in the IRQ context and the current
1465 * worker's fields which follow the 'K' locki 1326 * worker's fields which follow the 'K' locking rule can be accessed safely.
1466 */ 1327 */
1467 void wq_worker_tick(struct task_struct *task) 1328 void wq_worker_tick(struct task_struct *task)
1468 { 1329 {
1469 struct worker *worker = kthread_data( 1330 struct worker *worker = kthread_data(task);
1470 struct pool_workqueue *pwq = worker-> 1331 struct pool_workqueue *pwq = worker->current_pwq;
1471 struct worker_pool *pool = worker->po 1332 struct worker_pool *pool = worker->pool;
1472 1333
1473 if (!pwq) 1334 if (!pwq)
1474 return; 1335 return;
1475 1336
1476 pwq->stats[PWQ_STAT_CPU_TIME] += TICK 1337 pwq->stats[PWQ_STAT_CPU_TIME] += TICK_USEC;
1477 1338
1478 if (!wq_cpu_intensive_thresh_us) 1339 if (!wq_cpu_intensive_thresh_us)
1479 return; 1340 return;
1480 1341
1481 /* 1342 /*
1482 * If the current worker is concurren 1343 * If the current worker is concurrency managed and hogged the CPU for
1483 * longer than wq_cpu_intensive_thres 1344 * longer than wq_cpu_intensive_thresh_us, it's automatically marked
1484 * CPU_INTENSIVE to avoid stalling ot 1345 * CPU_INTENSIVE to avoid stalling other concurrency-managed work items.
1485 * 1346 *
1486 * Set @worker->sleeping means that @ 1347 * Set @worker->sleeping means that @worker is in the process of
1487 * switching out voluntarily and won' 1348 * switching out voluntarily and won't be contributing to
1488 * @pool->nr_running until it wakes u 1349 * @pool->nr_running until it wakes up. As wq_worker_sleeping() also
1489 * decrements ->nr_running, setting C 1350 * decrements ->nr_running, setting CPU_INTENSIVE here can lead to
1490 * double decrements. The task is rel 1351 * double decrements. The task is releasing the CPU anyway. Let's skip.
1491 * We probably want to make this pret 1352 * We probably want to make this prettier in the future.
1492 */ 1353 */
1493 if ((worker->flags & WORKER_NOT_RUNNI 1354 if ((worker->flags & WORKER_NOT_RUNNING) || READ_ONCE(worker->sleeping) ||
1494 worker->task->se.sum_exec_runtime 1355 worker->task->se.sum_exec_runtime - worker->current_at <
1495 wq_cpu_intensive_thresh_us * NSEC 1356 wq_cpu_intensive_thresh_us * NSEC_PER_USEC)
1496 return; 1357 return;
1497 1358
1498 raw_spin_lock(&pool->lock); 1359 raw_spin_lock(&pool->lock);
1499 1360
1500 worker_set_flags(worker, WORKER_CPU_I 1361 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
1501 wq_cpu_intensive_report(worker->curre 1362 wq_cpu_intensive_report(worker->current_func);
1502 pwq->stats[PWQ_STAT_CPU_INTENSIVE]++; 1363 pwq->stats[PWQ_STAT_CPU_INTENSIVE]++;
1503 1364
1504 if (kick_pool(pool)) 1365 if (kick_pool(pool))
1505 pwq->stats[PWQ_STAT_CM_WAKEUP 1366 pwq->stats[PWQ_STAT_CM_WAKEUP]++;
1506 1367
1507 raw_spin_unlock(&pool->lock); 1368 raw_spin_unlock(&pool->lock);
1508 } 1369 }
1509 1370
1510 /** 1371 /**
1511 * wq_worker_last_func - retrieve worker's la 1372 * wq_worker_last_func - retrieve worker's last work function
1512 * @task: Task to retrieve last work function 1373 * @task: Task to retrieve last work function of.
1513 * 1374 *
1514 * Determine the last function a worker execu 1375 * Determine the last function a worker executed. This is called from
1515 * the scheduler to get a worker's last known 1376 * the scheduler to get a worker's last known identity.
1516 * 1377 *
1517 * CONTEXT: 1378 * CONTEXT:
1518 * raw_spin_lock_irq(rq->lock) 1379 * raw_spin_lock_irq(rq->lock)
1519 * 1380 *
1520 * This function is called during schedule() 1381 * This function is called during schedule() when a kworker is going
1521 * to sleep. It's used by psi to identify agg 1382 * to sleep. It's used by psi to identify aggregation workers during
1522 * dequeuing, to allow periodic aggregation t 1383 * dequeuing, to allow periodic aggregation to shut-off when that
1523 * worker is the last task in the system or c 1384 * worker is the last task in the system or cgroup to go to sleep.
1524 * 1385 *
1525 * As this function doesn't involve any workq 1386 * As this function doesn't involve any workqueue-related locking, it
1526 * only returns stable values when called fro 1387 * only returns stable values when called from inside the scheduler's
1527 * queuing and dequeuing paths, when @task, w 1388 * queuing and dequeuing paths, when @task, which must be a kworker,
1528 * is guaranteed to not be processing any wor 1389 * is guaranteed to not be processing any works.
1529 * 1390 *
1530 * Return: 1391 * Return:
1531 * The last work function %current executed a 1392 * The last work function %current executed as a worker, NULL if it
1532 * hasn't executed any work yet. 1393 * hasn't executed any work yet.
1533 */ 1394 */
1534 work_func_t wq_worker_last_func(struct task_s 1395 work_func_t wq_worker_last_func(struct task_struct *task)
1535 { 1396 {
1536 struct worker *worker = kthread_data( 1397 struct worker *worker = kthread_data(task);
1537 1398
1538 return worker->last_func; 1399 return worker->last_func;
1539 } 1400 }
1540 1401
1541 /** 1402 /**
1542 * wq_node_nr_active - Determine wq_node_nr_a <<
1543 * @wq: workqueue of interest <<
1544 * @node: NUMA node, can be %NUMA_NO_NODE <<
1545 * <<
1546 * Determine wq_node_nr_active to use for @wq <<
1547 * <<
1548 * - %NULL for per-cpu workqueues as they don <<
1549 * <<
1550 * - node_nr_active[nr_node_ids] if @node is <<
1551 * <<
1552 * - Otherwise, node_nr_active[@node]. <<
1553 */ <<
1554 static struct wq_node_nr_active *wq_node_nr_a <<
1555 <<
1556 { <<
1557 if (!(wq->flags & WQ_UNBOUND)) <<
1558 return NULL; <<
1559 <<
1560 if (node == NUMA_NO_NODE) <<
1561 node = nr_node_ids; <<
1562 <<
1563 return wq->node_nr_active[node]; <<
1564 } <<
1565 <<
1566 /** <<
1567 * wq_update_node_max_active - Update per-nod <<
1568 * @wq: workqueue to update <<
1569 * @off_cpu: CPU that's going down, -1 if a C <<
1570 * <<
1571 * Update @wq->node_nr_active[]->max. @wq mus <<
1572 * distributed among nodes according to the p <<
1573 * cpus. The result is always between @wq->mi <<
1574 */ <<
1575 static void wq_update_node_max_active(struct <<
1576 { <<
1577 struct cpumask *effective = unbound_e <<
1578 int min_active = READ_ONCE(wq->min_ac <<
1579 int max_active = READ_ONCE(wq->max_ac <<
1580 int total_cpus, node; <<
1581 <<
1582 lockdep_assert_held(&wq->mutex); <<
1583 <<
1584 if (!wq_topo_initialized) <<
1585 return; <<
1586 <<
1587 if (off_cpu >= 0 && !cpumask_test_cpu <<
1588 off_cpu = -1; <<
1589 <<
1590 total_cpus = cpumask_weight_and(effec <<
1591 if (off_cpu >= 0) <<
1592 total_cpus--; <<
1593 <<
1594 /* If all CPUs of the wq get offline, <<
1595 if (unlikely(!total_cpus)) { <<
1596 for_each_node(node) <<
1597 wq_node_nr_active(wq, <<
1598 <<
1599 wq_node_nr_active(wq, NUMA_NO <<
1600 return; <<
1601 } <<
1602 <<
1603 for_each_node(node) { <<
1604 int node_cpus; <<
1605 <<
1606 node_cpus = cpumask_weight_an <<
1607 if (off_cpu >= 0 && cpu_to_no <<
1608 node_cpus--; <<
1609 <<
1610 wq_node_nr_active(wq, node)-> <<
1611 clamp(DIV_ROUND_UP(ma <<
1612 min_active, max <<
1613 } <<
1614 <<
1615 wq_node_nr_active(wq, NUMA_NO_NODE)-> <<
1616 } <<
1617 <<
1618 /** <<
1619 * get_pwq - get an extra reference on the sp 1403 * get_pwq - get an extra reference on the specified pool_workqueue
1620 * @pwq: pool_workqueue to get 1404 * @pwq: pool_workqueue to get
1621 * 1405 *
1622 * Obtain an extra reference on @pwq. The ca 1406 * Obtain an extra reference on @pwq. The caller should guarantee that
1623 * @pwq has positive refcnt and be holding th 1407 * @pwq has positive refcnt and be holding the matching pool->lock.
1624 */ 1408 */
1625 static void get_pwq(struct pool_workqueue *pw 1409 static void get_pwq(struct pool_workqueue *pwq)
1626 { 1410 {
1627 lockdep_assert_held(&pwq->pool->lock) 1411 lockdep_assert_held(&pwq->pool->lock);
1628 WARN_ON_ONCE(pwq->refcnt <= 0); 1412 WARN_ON_ONCE(pwq->refcnt <= 0);
1629 pwq->refcnt++; 1413 pwq->refcnt++;
1630 } 1414 }
1631 1415
1632 /** 1416 /**
1633 * put_pwq - put a pool_workqueue reference 1417 * put_pwq - put a pool_workqueue reference
1634 * @pwq: pool_workqueue to put 1418 * @pwq: pool_workqueue to put
1635 * 1419 *
1636 * Drop a reference of @pwq. If its refcnt r 1420 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1637 * destruction. The caller should be holding 1421 * destruction. The caller should be holding the matching pool->lock.
1638 */ 1422 */
1639 static void put_pwq(struct pool_workqueue *pw 1423 static void put_pwq(struct pool_workqueue *pwq)
1640 { 1424 {
1641 lockdep_assert_held(&pwq->pool->lock) 1425 lockdep_assert_held(&pwq->pool->lock);
1642 if (likely(--pwq->refcnt)) 1426 if (likely(--pwq->refcnt))
1643 return; 1427 return;
1644 /* 1428 /*
1645 * @pwq can't be released under pool- 1429 * @pwq can't be released under pool->lock, bounce to a dedicated
1646 * kthread_worker to avoid A-A deadlo 1430 * kthread_worker to avoid A-A deadlocks.
1647 */ 1431 */
1648 kthread_queue_work(pwq_release_worker 1432 kthread_queue_work(pwq_release_worker, &pwq->release_work);
1649 } 1433 }
1650 1434
1651 /** 1435 /**
1652 * put_pwq_unlocked - put_pwq() with surround 1436 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1653 * @pwq: pool_workqueue to put (can be %NULL) 1437 * @pwq: pool_workqueue to put (can be %NULL)
1654 * 1438 *
1655 * put_pwq() with locking. This function als 1439 * put_pwq() with locking. This function also allows %NULL @pwq.
1656 */ 1440 */
1657 static void put_pwq_unlocked(struct pool_work 1441 static void put_pwq_unlocked(struct pool_workqueue *pwq)
1658 { 1442 {
1659 if (pwq) { 1443 if (pwq) {
1660 /* 1444 /*
1661 * As both pwqs and pools are 1445 * As both pwqs and pools are RCU protected, the
1662 * following lock operations 1446 * following lock operations are safe.
1663 */ 1447 */
1664 raw_spin_lock_irq(&pwq->pool- 1448 raw_spin_lock_irq(&pwq->pool->lock);
1665 put_pwq(pwq); 1449 put_pwq(pwq);
1666 raw_spin_unlock_irq(&pwq->poo 1450 raw_spin_unlock_irq(&pwq->pool->lock);
1667 } 1451 }
1668 } 1452 }
1669 1453
1670 static bool pwq_is_empty(struct pool_workqueu !! 1454 static void pwq_activate_inactive_work(struct work_struct *work)
1671 { <<
1672 return !pwq->nr_active && list_empty( <<
1673 } <<
1674 <<
1675 static void __pwq_activate_work(struct pool_w <<
1676 struct work_s <<
1677 { 1455 {
1678 unsigned long *wdb = work_data_bits(w !! 1456 struct pool_workqueue *pwq = get_work_pwq(work);
1679 1457
1680 WARN_ON_ONCE(!(*wdb & WORK_STRUCT_INA <<
1681 trace_workqueue_activate_work(work); 1458 trace_workqueue_activate_work(work);
1682 if (list_empty(&pwq->pool->worklist)) 1459 if (list_empty(&pwq->pool->worklist))
1683 pwq->pool->watchdog_ts = jiff 1460 pwq->pool->watchdog_ts = jiffies;
1684 move_linked_works(work, &pwq->pool->w 1461 move_linked_works(work, &pwq->pool->worklist, NULL);
1685 __clear_bit(WORK_STRUCT_INACTIVE_BIT, !! 1462 __clear_bit(WORK_STRUCT_INACTIVE_BIT, work_data_bits(work));
>> 1463 pwq->nr_active++;
1686 } 1464 }
1687 1465
1688 static bool tryinc_node_nr_active(struct wq_n !! 1466 static void pwq_activate_first_inactive(struct pool_workqueue *pwq)
1689 { 1467 {
1690 int max = READ_ONCE(nna->max); !! 1468 struct work_struct *work = list_first_entry(&pwq->inactive_works,
1691 !! 1469 struct work_struct, entry);
1692 while (true) { <<
1693 int old, tmp; <<
1694 <<
1695 old = atomic_read(&nna->nr); <<
1696 if (old >= max) <<
1697 return false; <<
1698 tmp = atomic_cmpxchg_relaxed( <<
1699 if (tmp == old) <<
1700 return true; <<
1701 } <<
1702 } <<
1703 <<
1704 /** <<
1705 * pwq_tryinc_nr_active - Try to increment nr <<
1706 * @pwq: pool_workqueue of interest <<
1707 * @fill: max_active may have increased, try <<
1708 * <<
1709 * Try to increment nr_active for @pwq. Retur <<
1710 * successfully obtained. %false otherwise. <<
1711 */ <<
1712 static bool pwq_tryinc_nr_active(struct pool_ <<
1713 { <<
1714 struct workqueue_struct *wq = pwq->wq <<
1715 struct worker_pool *pool = pwq->pool; <<
1716 struct wq_node_nr_active *nna = wq_no <<
1717 bool obtained = false; <<
1718 <<
1719 lockdep_assert_held(&pool->lock); <<
1720 <<
1721 if (!nna) { <<
1722 /* BH or per-cpu workqueue, p <<
1723 obtained = pwq->nr_active < R <<
1724 goto out; <<
1725 } <<
1726 <<
1727 if (unlikely(pwq->plugged)) <<
1728 return false; <<
1729 <<
1730 /* <<
1731 * Unbound workqueue uses per-node sh <<
1732 * already waiting on $nna, pwq_dec_n <<
1733 * concurrency level. Don't jump the <<
1734 * <<
1735 * We need to ignore the pending test <<
1736 * pwq_dec_nr_active() can only maint <<
1737 * increase it. This is indicated by <<
1738 */ <<
1739 if (!list_empty(&pwq->pending_node) & <<
1740 goto out; <<
1741 <<
1742 obtained = tryinc_node_nr_active(nna) <<
1743 if (obtained) <<
1744 goto out; <<
1745 <<
1746 /* <<
1747 * Lockless acquisition failed. Lock, <<
1748 * and try again. The smp_mb() is pai <<
1749 * of atomic_dec_return() in pwq_dec_ <<
1750 * we see the decremented $nna->nr or <<
1751 * $nna->pending_pwqs. <<
1752 */ <<
1753 raw_spin_lock(&nna->lock); <<
1754 <<
1755 if (list_empty(&pwq->pending_node)) <<
1756 list_add_tail(&pwq->pending_n <<
1757 else if (likely(!fill)) <<
1758 goto out_unlock; <<
1759 <<
1760 smp_mb(); <<
1761 <<
1762 obtained = tryinc_node_nr_active(nna) <<
1763 <<
1764 /* <<
1765 * If @fill, @pwq might have already <<
1766 * pending in cold paths doesn't affe <<
1767 */ <<
1768 if (obtained && likely(!fill)) <<
1769 list_del_init(&pwq->pending_n <<
1770 <<
1771 out_unlock: <<
1772 raw_spin_unlock(&nna->lock); <<
1773 out: <<
1774 if (obtained) <<
1775 pwq->nr_active++; <<
1776 return obtained; <<
1777 } <<
1778 <<
1779 /** <<
1780 * pwq_activate_first_inactive - Activate the <<
1781 * @pwq: pool_workqueue of interest <<
1782 * @fill: max_active may have increased, try <<
1783 * <<
1784 * Activate the first inactive work item of @ <<
1785 * max_active limit. <<
1786 * <<
1787 * Returns %true if an inactive work item has <<
1788 * inactive work item is found or max_active <<
1789 */ <<
1790 static bool pwq_activate_first_inactive(struc <<
1791 { <<
1792 struct work_struct *work = <<
1793 list_first_entry_or_null(&pwq <<
1794 stru <<
1795 <<
1796 if (work && pwq_tryinc_nr_active(pwq, <<
1797 __pwq_activate_work(pwq, work <<
1798 return true; <<
1799 } else { <<
1800 return false; <<
1801 } <<
1802 } <<
1803 <<
1804 /** <<
1805 * unplug_oldest_pwq - unplug the oldest pool <<
1806 * @wq: workqueue_struct where its oldest pwq <<
1807 * <<
1808 * This function should only be called for or <<
1809 * oldest pwq is unplugged, the others are pl <<
1810 * ensure proper work item ordering:: <<
1811 * <<
1812 * dfl_pwq --------------+ [P] - plugg <<
1813 * | <<
1814 * v <<
1815 * pwqs -> A -> B [P] -> C [P] (newest) <<
1816 * | | | <<
1817 * 1 3 5 <<
1818 * | | | <<
1819 * 2 4 6 <<
1820 * <<
1821 * When the oldest pwq is drained and removed <<
1822 * to unplug the next oldest one to start its <<
1823 * pwq's are linked into wq->pwqs with the ol <<
1824 * the list is the oldest. <<
1825 */ <<
1826 static void unplug_oldest_pwq(struct workqueu <<
1827 { <<
1828 struct pool_workqueue *pwq; <<
1829 <<
1830 lockdep_assert_held(&wq->mutex); <<
1831 <<
1832 /* Caller should make sure that pwqs <<
1833 pwq = list_first_entry_or_null(&wq->p <<
1834 pwqs_n <<
1835 raw_spin_lock_irq(&pwq->pool->lock); <<
1836 if (pwq->plugged) { <<
1837 pwq->plugged = false; <<
1838 if (pwq_activate_first_inacti <<
1839 kick_pool(pwq->pool); <<
1840 } <<
1841 raw_spin_unlock_irq(&pwq->pool->lock) <<
1842 } <<
1843 <<
1844 /** <<
1845 * node_activate_pending_pwq - Activate a pen <<
1846 * @nna: wq_node_nr_active to activate a pend <<
1847 * @caller_pool: worker_pool the caller is lo <<
1848 * <<
1849 * Activate a pwq in @nna->pending_pwqs. Call <<
1850 * @caller_pool may be unlocked and relocked <<
1851 */ <<
1852 static void node_activate_pending_pwq(struct <<
1853 struct <<
1854 { <<
1855 struct worker_pool *locked_pool = cal <<
1856 struct pool_workqueue *pwq; <<
1857 struct work_struct *work; <<
1858 <<
1859 lockdep_assert_held(&caller_pool->loc <<
1860 <<
1861 raw_spin_lock(&nna->lock); <<
1862 retry: <<
1863 pwq = list_first_entry_or_null(&nna-> <<
1864 struct <<
1865 if (!pwq) <<
1866 goto out_unlock; <<
1867 <<
1868 /* <<
1869 * If @pwq is for a different pool th <<
1870 * @pwq->pool->lock. Let's trylock fi <<
1871 * / lock dance. For that, we also ne <<
1872 * nested inside pool locks. <<
1873 */ <<
1874 if (pwq->pool != locked_pool) { <<
1875 raw_spin_unlock(&locked_pool- <<
1876 locked_pool = pwq->pool; <<
1877 if (!raw_spin_trylock(&locked <<
1878 raw_spin_unlock(&nna- <<
1879 raw_spin_lock(&locked <<
1880 raw_spin_lock(&nna->l <<
1881 goto retry; <<
1882 } <<
1883 } <<
1884 <<
1885 /* <<
1886 * $pwq may not have any inactive wor <<
1887 * Drop it from pending_pwqs and see <<
1888 */ <<
1889 work = list_first_entry_or_null(&pwq- <<
1890 struc <<
1891 if (!work) { <<
1892 list_del_init(&pwq->pending_n <<
1893 goto retry; <<
1894 } <<
1895 <<
1896 /* <<
1897 * Acquire an nr_active count and act <<
1898 * $pwq still has inactive work items <<
1899 * pending_pwqs so that we round-robi <<
1900 * inactive work items are not activa <<
1901 * given that there has never been an <<
1902 */ <<
1903 if (likely(tryinc_node_nr_active(nna) <<
1904 pwq->nr_active++; <<
1905 __pwq_activate_work(pwq, work <<
1906 <<
1907 if (list_empty(&pwq->inactive <<
1908 list_del_init(&pwq->p <<
1909 else <<
1910 list_move_tail(&pwq-> <<
1911 <<
1912 /* if activating a foreign po <<
1913 if (pwq->pool != caller_pool) <<
1914 kick_pool(pwq->pool); <<
1915 } <<
1916 <<
1917 out_unlock: <<
1918 raw_spin_unlock(&nna->lock); <<
1919 if (locked_pool != caller_pool) { <<
1920 raw_spin_unlock(&locked_pool- <<
1921 raw_spin_lock(&caller_pool->l <<
1922 } <<
1923 } <<
1924 <<
1925 /** <<
1926 * pwq_dec_nr_active - Retire an active count <<
1927 * @pwq: pool_workqueue of interest <<
1928 * <<
1929 * Decrement @pwq's nr_active and try to acti <<
1930 * For unbound workqueues, this function may <<
1931 */ <<
1932 static void pwq_dec_nr_active(struct pool_wor <<
1933 { <<
1934 struct worker_pool *pool = pwq->pool; <<
1935 struct wq_node_nr_active *nna = wq_no <<
1936 <<
1937 lockdep_assert_held(&pool->lock); <<
1938 <<
1939 /* <<
1940 * @pwq->nr_active should be decremen <<
1941 * workqueues. <<
1942 */ <<
1943 pwq->nr_active--; <<
1944 <<
1945 /* <<
1946 * For a percpu workqueue, it's simpl <<
1947 * inactive work item on @pwq itself. <<
1948 */ <<
1949 if (!nna) { <<
1950 pwq_activate_first_inactive(p <<
1951 return; <<
1952 } <<
1953 <<
1954 /* <<
1955 * If @pwq is for an unbound workqueu <<
1956 * multiple pwqs and pools may be sha <<
1957 * pwq needs to wait for an nr_active <<
1958 * $nna->pending_pwqs. The following <<
1959 * memory barrier is paired with smp_ <<
1960 * guarantee that either we see non-e <<
1961 * decremented $nna->nr. <<
1962 * <<
1963 * $nna->max may change as CPUs come <<
1964 * max_active gets updated. However, <<
1965 * larger than @pwq->wq->min_active w <<
1966 * This maintains the forward progres <<
1967 */ <<
1968 if (atomic_dec_return(&nna->nr) >= RE <<
1969 return; <<
1970 1470
1971 if (!list_empty(&nna->pending_pwqs)) !! 1471 pwq_activate_inactive_work(work);
1972 node_activate_pending_pwq(nna <<
1973 } 1472 }
1974 1473
1975 /** 1474 /**
1976 * pwq_dec_nr_in_flight - decrement pwq's nr_ 1475 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1977 * @pwq: pwq of interest 1476 * @pwq: pwq of interest
1978 * @work_data: work_data of work which left t 1477 * @work_data: work_data of work which left the queue
1979 * 1478 *
1980 * A work either has completed or is removed 1479 * A work either has completed or is removed from pending queue,
1981 * decrement nr_in_flight of its pwq and hand 1480 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1982 * 1481 *
1983 * NOTE: <<
1984 * For unbound workqueues, this function may <<
1985 * and thus should be called after all other <<
1986 * work item is complete. <<
1987 * <<
1988 * CONTEXT: 1482 * CONTEXT:
1989 * raw_spin_lock_irq(pool->lock). 1483 * raw_spin_lock_irq(pool->lock).
1990 */ 1484 */
1991 static void pwq_dec_nr_in_flight(struct pool_ 1485 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, unsigned long work_data)
1992 { 1486 {
1993 int color = get_work_color(work_data) 1487 int color = get_work_color(work_data);
1994 1488
1995 if (!(work_data & WORK_STRUCT_INACTIV !! 1489 if (!(work_data & WORK_STRUCT_INACTIVE)) {
1996 pwq_dec_nr_active(pwq); !! 1490 pwq->nr_active--;
>> 1491 if (!list_empty(&pwq->inactive_works)) {
>> 1492 /* one down, submit an inactive one */
>> 1493 if (pwq->nr_active < pwq->max_active)
>> 1494 pwq_activate_first_inactive(pwq);
>> 1495 }
>> 1496 }
1997 1497
1998 pwq->nr_in_flight[color]--; 1498 pwq->nr_in_flight[color]--;
1999 1499
2000 /* is flush in progress and are we at 1500 /* is flush in progress and are we at the flushing tip? */
2001 if (likely(pwq->flush_color != color) 1501 if (likely(pwq->flush_color != color))
2002 goto out_put; 1502 goto out_put;
2003 1503
2004 /* are there still in-flight works? * 1504 /* are there still in-flight works? */
2005 if (pwq->nr_in_flight[color]) 1505 if (pwq->nr_in_flight[color])
2006 goto out_put; 1506 goto out_put;
2007 1507
2008 /* this pwq is done, clear flush_colo 1508 /* this pwq is done, clear flush_color */
2009 pwq->flush_color = -1; 1509 pwq->flush_color = -1;
2010 1510
2011 /* 1511 /*
2012 * If this was the last pwq, wake up 1512 * If this was the last pwq, wake up the first flusher. It
2013 * will handle the rest. 1513 * will handle the rest.
2014 */ 1514 */
2015 if (atomic_dec_and_test(&pwq->wq->nr_ 1515 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
2016 complete(&pwq->wq->first_flus 1516 complete(&pwq->wq->first_flusher->done);
2017 out_put: 1517 out_put:
2018 put_pwq(pwq); 1518 put_pwq(pwq);
2019 } 1519 }
2020 1520
2021 /** 1521 /**
2022 * try_to_grab_pending - steal work item from 1522 * try_to_grab_pending - steal work item from worklist and disable irq
2023 * @work: work item to steal 1523 * @work: work item to steal
2024 * @cflags: %WORK_CANCEL_ flags !! 1524 * @is_dwork: @work is a delayed_work
2025 * @irq_flags: place to store irq state !! 1525 * @flags: place to store irq state
2026 * 1526 *
2027 * Try to grab PENDING bit of @work. This fu 1527 * Try to grab PENDING bit of @work. This function can handle @work in any
2028 * stable state - idle, on timer or on workli 1528 * stable state - idle, on timer or on worklist.
2029 * 1529 *
2030 * Return: 1530 * Return:
2031 * 1531 *
2032 * ======== ============================= 1532 * ======== ================================================================
2033 * 1 if @work was pending and we s 1533 * 1 if @work was pending and we successfully stole PENDING
2034 * 0 if @work was idle and we clai 1534 * 0 if @work was idle and we claimed PENDING
2035 * -EAGAIN if PENDING couldn't be grabbe 1535 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
>> 1536 * -ENOENT if someone else is canceling @work, this state may persist
>> 1537 * for arbitrarily long
2036 * ======== ============================= 1538 * ======== ================================================================
2037 * 1539 *
2038 * Note: 1540 * Note:
2039 * On >= 0 return, the caller owns @work's PE 1541 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
2040 * interrupted while holding PENDING and @wor 1542 * interrupted while holding PENDING and @work off queue, irq must be
2041 * disabled on entry. This, combined with de 1543 * disabled on entry. This, combined with delayed_work->timer being
2042 * irqsafe, ensures that we return -EAGAIN fo 1544 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
2043 * 1545 *
2044 * On successful return, >= 0, irq is disable 1546 * On successful return, >= 0, irq is disabled and the caller is
2045 * responsible for releasing it using local_i !! 1547 * responsible for releasing it using local_irq_restore(*@flags).
2046 * 1548 *
2047 * This function is safe to call from any con 1549 * This function is safe to call from any context including IRQ handler.
2048 */ 1550 */
2049 static int try_to_grab_pending(struct work_st !! 1551 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
2050 unsigned long !! 1552 unsigned long *flags)
2051 { 1553 {
2052 struct worker_pool *pool; 1554 struct worker_pool *pool;
2053 struct pool_workqueue *pwq; 1555 struct pool_workqueue *pwq;
2054 1556
2055 local_irq_save(*irq_flags); !! 1557 local_irq_save(*flags);
2056 1558
2057 /* try to steal the timer if it exist 1559 /* try to steal the timer if it exists */
2058 if (cflags & WORK_CANCEL_DELAYED) { !! 1560 if (is_dwork) {
2059 struct delayed_work *dwork = 1561 struct delayed_work *dwork = to_delayed_work(work);
2060 1562
2061 /* 1563 /*
2062 * dwork->timer is irqsafe. 1564 * dwork->timer is irqsafe. If del_timer() fails, it's
2063 * guaranteed that the timer 1565 * guaranteed that the timer is not queued anywhere and not
2064 * running on the local CPU. 1566 * running on the local CPU.
2065 */ 1567 */
2066 if (likely(del_timer(&dwork-> 1568 if (likely(del_timer(&dwork->timer)))
2067 return 1; 1569 return 1;
2068 } 1570 }
2069 1571
2070 /* try to claim PENDING the normal wa 1572 /* try to claim PENDING the normal way */
2071 if (!test_and_set_bit(WORK_STRUCT_PEN 1573 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
2072 return 0; 1574 return 0;
2073 1575
2074 rcu_read_lock(); 1576 rcu_read_lock();
2075 /* 1577 /*
2076 * The queueing is in progress, or it 1578 * The queueing is in progress, or it is already queued. Try to
2077 * steal it from ->worklist without c 1579 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
2078 */ 1580 */
2079 pool = get_work_pool(work); 1581 pool = get_work_pool(work);
2080 if (!pool) 1582 if (!pool)
2081 goto fail; 1583 goto fail;
2082 1584
2083 raw_spin_lock(&pool->lock); 1585 raw_spin_lock(&pool->lock);
2084 /* 1586 /*
2085 * work->data is guaranteed to point 1587 * work->data is guaranteed to point to pwq only while the work
2086 * item is queued on pwq->wq, and bot 1588 * item is queued on pwq->wq, and both updating work->data to point
2087 * to pwq on queueing and to pool on 1589 * to pwq on queueing and to pool on dequeueing are done under
2088 * pwq->pool->lock. This in turn gua 1590 * pwq->pool->lock. This in turn guarantees that, if work->data
2089 * points to pwq which is associated 1591 * points to pwq which is associated with a locked pool, the work
2090 * item is currently queued on that p 1592 * item is currently queued on that pool.
2091 */ 1593 */
2092 pwq = get_work_pwq(work); 1594 pwq = get_work_pwq(work);
2093 if (pwq && pwq->pool == pool) { 1595 if (pwq && pwq->pool == pool) {
2094 unsigned long work_data = *wo <<
2095 <<
2096 debug_work_deactivate(work); 1596 debug_work_deactivate(work);
2097 1597
2098 /* 1598 /*
2099 * A cancelable inactive work 1599 * A cancelable inactive work item must be in the
2100 * pwq->inactive_works since 1600 * pwq->inactive_works since a queued barrier can't be
2101 * canceled (see the comments 1601 * canceled (see the comments in insert_wq_barrier()).
2102 * 1602 *
2103 * An inactive work item cann !! 1603 * An inactive work item cannot be grabbed directly because
2104 * it might have linked barri 1604 * it might have linked barrier work items which, if left
2105 * on the inactive_works list 1605 * on the inactive_works list, will confuse pwq->nr_active
2106 * management later on and ca !! 1606 * management later on and cause stall. Make sure the work
2107 * barrier work items to the !! 1607 * item is activated before grabbing.
2108 * item. Also keep WORK_STRUC <<
2109 * it doesn't participate in <<
2110 * pwq_dec_nr_in_flight(). <<
2111 */ 1608 */
2112 if (work_data & WORK_STRUCT_I !! 1609 if (*work_data_bits(work) & WORK_STRUCT_INACTIVE)
2113 move_linked_works(wor !! 1610 pwq_activate_inactive_work(work);
2114 1611
2115 list_del_init(&work->entry); 1612 list_del_init(&work->entry);
>> 1613 pwq_dec_nr_in_flight(pwq, *work_data_bits(work));
2116 1614
2117 /* !! 1615 /* work->data points to pwq iff queued, point to pool */
2118 * work->data points to pwq i !! 1616 set_work_pool_and_keep_pending(work, pool->id);
2119 * this destroys work->data n <<
2120 */ <<
2121 set_work_pool_and_keep_pendin <<
2122 <<
2123 <<
2124 /* must be the last step, see <<
2125 pwq_dec_nr_in_flight(pwq, wor <<
2126 1617
2127 raw_spin_unlock(&pool->lock); 1618 raw_spin_unlock(&pool->lock);
2128 rcu_read_unlock(); 1619 rcu_read_unlock();
2129 return 1; 1620 return 1;
2130 } 1621 }
2131 raw_spin_unlock(&pool->lock); 1622 raw_spin_unlock(&pool->lock);
2132 fail: 1623 fail:
2133 rcu_read_unlock(); 1624 rcu_read_unlock();
2134 local_irq_restore(*irq_flags); !! 1625 local_irq_restore(*flags);
>> 1626 if (work_is_canceling(work))
>> 1627 return -ENOENT;
>> 1628 cpu_relax();
2135 return -EAGAIN; 1629 return -EAGAIN;
2136 } 1630 }
2137 1631
2138 /** 1632 /**
2139 * work_grab_pending - steal work item from w <<
2140 * @work: work item to steal <<
2141 * @cflags: %WORK_CANCEL_ flags <<
2142 * @irq_flags: place to store IRQ state <<
2143 * <<
2144 * Grab PENDING bit of @work. @work can be in <<
2145 * or on worklist. <<
2146 * <<
2147 * Can be called from any context. IRQ is dis <<
2148 * stored in *@irq_flags. The caller is respo <<
2149 * local_irq_restore(). <<
2150 * <<
2151 * Returns %true if @work was pending. %false <<
2152 */ <<
2153 static bool work_grab_pending(struct work_str <<
2154 unsigned long * <<
2155 { <<
2156 int ret; <<
2157 <<
2158 while (true) { <<
2159 ret = try_to_grab_pending(wor <<
2160 if (ret >= 0) <<
2161 return ret; <<
2162 cpu_relax(); <<
2163 } <<
2164 } <<
2165 <<
2166 /** <<
2167 * insert_work - insert a work into a pool 1633 * insert_work - insert a work into a pool
2168 * @pwq: pwq @work belongs to 1634 * @pwq: pwq @work belongs to
2169 * @work: work to insert 1635 * @work: work to insert
2170 * @head: insertion point 1636 * @head: insertion point
2171 * @extra_flags: extra WORK_STRUCT_* flags to 1637 * @extra_flags: extra WORK_STRUCT_* flags to set
2172 * 1638 *
2173 * Insert @work which belongs to @pwq after @ 1639 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
2174 * work_struct flags. 1640 * work_struct flags.
2175 * 1641 *
2176 * CONTEXT: 1642 * CONTEXT:
2177 * raw_spin_lock_irq(pool->lock). 1643 * raw_spin_lock_irq(pool->lock).
2178 */ 1644 */
2179 static void insert_work(struct pool_workqueue 1645 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
2180 struct list_head *hea 1646 struct list_head *head, unsigned int extra_flags)
2181 { 1647 {
2182 debug_work_activate(work); 1648 debug_work_activate(work);
2183 1649
2184 /* record the work call stack in orde 1650 /* record the work call stack in order to print it in KASAN reports */
2185 kasan_record_aux_stack_noalloc(work); 1651 kasan_record_aux_stack_noalloc(work);
2186 1652
2187 /* we own @work, set data and link */ 1653 /* we own @work, set data and link */
2188 set_work_pwq(work, pwq, extra_flags); 1654 set_work_pwq(work, pwq, extra_flags);
2189 list_add_tail(&work->entry, head); 1655 list_add_tail(&work->entry, head);
2190 get_pwq(pwq); 1656 get_pwq(pwq);
2191 } 1657 }
2192 1658
2193 /* 1659 /*
2194 * Test whether @work is being queued from an 1660 * Test whether @work is being queued from another work executing on the
2195 * same workqueue. 1661 * same workqueue.
2196 */ 1662 */
2197 static bool is_chained_work(struct workqueue_ 1663 static bool is_chained_work(struct workqueue_struct *wq)
2198 { 1664 {
2199 struct worker *worker; 1665 struct worker *worker;
2200 1666
2201 worker = current_wq_worker(); 1667 worker = current_wq_worker();
2202 /* 1668 /*
2203 * Return %true iff I'm a worker exec 1669 * Return %true iff I'm a worker executing a work item on @wq. If
2204 * I'm @worker, it's safe to derefere 1670 * I'm @worker, it's safe to dereference it without locking.
2205 */ 1671 */
2206 return worker && worker->current_pwq- 1672 return worker && worker->current_pwq->wq == wq;
2207 } 1673 }
2208 1674
2209 /* 1675 /*
2210 * When queueing an unbound work item to a wq 1676 * When queueing an unbound work item to a wq, prefer local CPU if allowed
2211 * by wq_unbound_cpumask. Otherwise, round r 1677 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
2212 * avoid perturbing sensitive tasks. 1678 * avoid perturbing sensitive tasks.
2213 */ 1679 */
2214 static int wq_select_unbound_cpu(int cpu) 1680 static int wq_select_unbound_cpu(int cpu)
2215 { 1681 {
2216 int new_cpu; 1682 int new_cpu;
2217 1683
2218 if (likely(!wq_debug_force_rr_cpu)) { 1684 if (likely(!wq_debug_force_rr_cpu)) {
2219 if (cpumask_test_cpu(cpu, wq_ 1685 if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
2220 return cpu; 1686 return cpu;
2221 } else { 1687 } else {
2222 pr_warn_once("workqueue: roun 1688 pr_warn_once("workqueue: round-robin CPU selection forced, expect performance impact\n");
2223 } 1689 }
2224 1690
2225 new_cpu = __this_cpu_read(wq_rr_cpu_l 1691 new_cpu = __this_cpu_read(wq_rr_cpu_last);
2226 new_cpu = cpumask_next_and(new_cpu, w 1692 new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
2227 if (unlikely(new_cpu >= nr_cpu_ids)) 1693 if (unlikely(new_cpu >= nr_cpu_ids)) {
2228 new_cpu = cpumask_first_and(w 1694 new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
2229 if (unlikely(new_cpu >= nr_cp 1695 if (unlikely(new_cpu >= nr_cpu_ids))
2230 return cpu; 1696 return cpu;
2231 } 1697 }
2232 __this_cpu_write(wq_rr_cpu_last, new_ 1698 __this_cpu_write(wq_rr_cpu_last, new_cpu);
2233 1699
2234 return new_cpu; 1700 return new_cpu;
2235 } 1701 }
2236 1702
2237 static void __queue_work(int cpu, struct work 1703 static void __queue_work(int cpu, struct workqueue_struct *wq,
2238 struct work_struct * 1704 struct work_struct *work)
2239 { 1705 {
2240 struct pool_workqueue *pwq; 1706 struct pool_workqueue *pwq;
2241 struct worker_pool *last_pool, *pool; 1707 struct worker_pool *last_pool, *pool;
2242 unsigned int work_flags; 1708 unsigned int work_flags;
2243 unsigned int req_cpu = cpu; 1709 unsigned int req_cpu = cpu;
2244 1710
2245 /* 1711 /*
2246 * While a work item is PENDING && of 1712 * While a work item is PENDING && off queue, a task trying to
2247 * steal the PENDING will busy-loop w 1713 * steal the PENDING will busy-loop waiting for it to either get
2248 * queued or lose PENDING. Grabbing 1714 * queued or lose PENDING. Grabbing PENDING and queueing should
2249 * happen with IRQ disabled. 1715 * happen with IRQ disabled.
2250 */ 1716 */
2251 lockdep_assert_irqs_disabled(); 1717 lockdep_assert_irqs_disabled();
2252 1718
>> 1719
2253 /* 1720 /*
2254 * For a draining wq, only works from 1721 * For a draining wq, only works from the same workqueue are
2255 * allowed. The __WQ_DESTROYING helps 1722 * allowed. The __WQ_DESTROYING helps to spot the issue that
2256 * queues a new work item to a wq aft 1723 * queues a new work item to a wq after destroy_workqueue(wq).
2257 */ 1724 */
2258 if (unlikely(wq->flags & (__WQ_DESTRO 1725 if (unlikely(wq->flags & (__WQ_DESTROYING | __WQ_DRAINING) &&
2259 WARN_ON_ONCE(!is_chained 1726 WARN_ON_ONCE(!is_chained_work(wq))))
2260 return; 1727 return;
2261 rcu_read_lock(); 1728 rcu_read_lock();
2262 retry: 1729 retry:
2263 /* pwq which will be used unless @wor 1730 /* pwq which will be used unless @work is executing elsewhere */
2264 if (req_cpu == WORK_CPU_UNBOUND) { 1731 if (req_cpu == WORK_CPU_UNBOUND) {
2265 if (wq->flags & WQ_UNBOUND) 1732 if (wq->flags & WQ_UNBOUND)
2266 cpu = wq_select_unbou 1733 cpu = wq_select_unbound_cpu(raw_smp_processor_id());
2267 else 1734 else
2268 cpu = raw_smp_process 1735 cpu = raw_smp_processor_id();
2269 } 1736 }
2270 1737
2271 pwq = rcu_dereference(*per_cpu_ptr(wq 1738 pwq = rcu_dereference(*per_cpu_ptr(wq->cpu_pwq, cpu));
2272 pool = pwq->pool; 1739 pool = pwq->pool;
2273 1740
2274 /* 1741 /*
2275 * If @work was previously on a diffe 1742 * If @work was previously on a different pool, it might still be
2276 * running there, in which case the w 1743 * running there, in which case the work needs to be queued on that
2277 * pool to guarantee non-reentrancy. 1744 * pool to guarantee non-reentrancy.
2278 * <<
2279 * For ordered workqueue, work items <<
2280 * for accurate order management. Gu <<
2281 * non-reentrancy. See the comments <<
2282 */ 1745 */
2283 last_pool = get_work_pool(work); 1746 last_pool = get_work_pool(work);
2284 if (last_pool && last_pool != pool && !! 1747 if (last_pool && last_pool != pool) {
2285 struct worker *worker; 1748 struct worker *worker;
2286 1749
2287 raw_spin_lock(&last_pool->loc 1750 raw_spin_lock(&last_pool->lock);
2288 1751
2289 worker = find_worker_executin 1752 worker = find_worker_executing_work(last_pool, work);
2290 1753
2291 if (worker && worker->current 1754 if (worker && worker->current_pwq->wq == wq) {
2292 pwq = worker->current 1755 pwq = worker->current_pwq;
2293 pool = pwq->pool; 1756 pool = pwq->pool;
2294 WARN_ON_ONCE(pool != 1757 WARN_ON_ONCE(pool != last_pool);
2295 } else { 1758 } else {
2296 /* meh... not running 1759 /* meh... not running there, queue here */
2297 raw_spin_unlock(&last 1760 raw_spin_unlock(&last_pool->lock);
2298 raw_spin_lock(&pool-> 1761 raw_spin_lock(&pool->lock);
2299 } 1762 }
2300 } else { 1763 } else {
2301 raw_spin_lock(&pool->lock); 1764 raw_spin_lock(&pool->lock);
2302 } 1765 }
2303 1766
2304 /* 1767 /*
2305 * pwq is determined and locked. For 1768 * pwq is determined and locked. For unbound pools, we could have raced
2306 * with pwq release and it could alre 1769 * with pwq release and it could already be dead. If its refcnt is zero,
2307 * repeat pwq selection. Note that un 1770 * repeat pwq selection. Note that unbound pwqs never die without
2308 * another pwq replacing it in cpu_pw 1771 * another pwq replacing it in cpu_pwq or while work items are executing
2309 * on it, so the retrying is guarante 1772 * on it, so the retrying is guaranteed to make forward-progress.
2310 */ 1773 */
2311 if (unlikely(!pwq->refcnt)) { 1774 if (unlikely(!pwq->refcnt)) {
2312 if (wq->flags & WQ_UNBOUND) { 1775 if (wq->flags & WQ_UNBOUND) {
2313 raw_spin_unlock(&pool 1776 raw_spin_unlock(&pool->lock);
2314 cpu_relax(); 1777 cpu_relax();
2315 goto retry; 1778 goto retry;
2316 } 1779 }
2317 /* oops */ 1780 /* oops */
2318 WARN_ONCE(true, "workqueue: p 1781 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
2319 wq->name, cpu); 1782 wq->name, cpu);
2320 } 1783 }
2321 1784
2322 /* pwq determined, queue */ 1785 /* pwq determined, queue */
2323 trace_workqueue_queue_work(req_cpu, p 1786 trace_workqueue_queue_work(req_cpu, pwq, work);
2324 1787
2325 if (WARN_ON(!list_empty(&work->entry) 1788 if (WARN_ON(!list_empty(&work->entry)))
2326 goto out; 1789 goto out;
2327 1790
2328 pwq->nr_in_flight[pwq->work_color]++; 1791 pwq->nr_in_flight[pwq->work_color]++;
2329 work_flags = work_color_to_flags(pwq- 1792 work_flags = work_color_to_flags(pwq->work_color);
2330 1793
2331 /* !! 1794 if (likely(pwq->nr_active < pwq->max_active)) {
2332 * Limit the number of concurrently a <<
2333 * @work must also queue behind exist <<
2334 * ordering when max_active changes. <<
2335 */ <<
2336 if (list_empty(&pwq->inactive_works) <<
2337 if (list_empty(&pool->worklis 1795 if (list_empty(&pool->worklist))
2338 pool->watchdog_ts = j 1796 pool->watchdog_ts = jiffies;
2339 1797
2340 trace_workqueue_activate_work 1798 trace_workqueue_activate_work(work);
>> 1799 pwq->nr_active++;
2341 insert_work(pwq, work, &pool- 1800 insert_work(pwq, work, &pool->worklist, work_flags);
2342 kick_pool(pool); 1801 kick_pool(pool);
2343 } else { 1802 } else {
2344 work_flags |= WORK_STRUCT_INA 1803 work_flags |= WORK_STRUCT_INACTIVE;
2345 insert_work(pwq, work, &pwq-> 1804 insert_work(pwq, work, &pwq->inactive_works, work_flags);
2346 } 1805 }
2347 1806
2348 out: 1807 out:
2349 raw_spin_unlock(&pool->lock); 1808 raw_spin_unlock(&pool->lock);
2350 rcu_read_unlock(); 1809 rcu_read_unlock();
2351 } 1810 }
2352 1811
2353 static bool clear_pending_if_disabled(struct <<
2354 { <<
2355 unsigned long data = *work_data_bits( <<
2356 struct work_offq_data offqd; <<
2357 <<
2358 if (likely((data & WORK_STRUCT_PWQ) | <<
2359 !(data & WORK_OFFQ_DISABLE <<
2360 return false; <<
2361 <<
2362 work_offqd_unpack(&offqd, data); <<
2363 set_work_pool_and_clear_pending(work, <<
2364 work_ <<
2365 return true; <<
2366 } <<
2367 <<
2368 /** 1812 /**
2369 * queue_work_on - queue work on specific cpu 1813 * queue_work_on - queue work on specific cpu
2370 * @cpu: CPU number to execute work on 1814 * @cpu: CPU number to execute work on
2371 * @wq: workqueue to use 1815 * @wq: workqueue to use
2372 * @work: work to queue 1816 * @work: work to queue
2373 * 1817 *
2374 * We queue the work to a specific CPU, the c 1818 * We queue the work to a specific CPU, the caller must ensure it
2375 * can't go away. Callers that fail to ensur 1819 * can't go away. Callers that fail to ensure that the specified
2376 * CPU cannot go away will execute on a rando 1820 * CPU cannot go away will execute on a randomly chosen CPU.
2377 * But note well that callers specifying a CP 1821 * But note well that callers specifying a CPU that never has been
2378 * online will get a splat. 1822 * online will get a splat.
2379 * 1823 *
2380 * Return: %false if @work was already on a q 1824 * Return: %false if @work was already on a queue, %true otherwise.
2381 */ 1825 */
2382 bool queue_work_on(int cpu, struct workqueue_ 1826 bool queue_work_on(int cpu, struct workqueue_struct *wq,
2383 struct work_struct *work) 1827 struct work_struct *work)
2384 { 1828 {
2385 bool ret = false; 1829 bool ret = false;
2386 unsigned long irq_flags; !! 1830 unsigned long flags;
2387 1831
2388 local_irq_save(irq_flags); !! 1832 local_irq_save(flags);
2389 1833
2390 if (!test_and_set_bit(WORK_STRUCT_PEN !! 1834 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
2391 !clear_pending_if_disabled(work)) <<
2392 __queue_work(cpu, wq, work); 1835 __queue_work(cpu, wq, work);
2393 ret = true; 1836 ret = true;
2394 } 1837 }
2395 1838
2396 local_irq_restore(irq_flags); !! 1839 local_irq_restore(flags);
2397 return ret; 1840 return ret;
2398 } 1841 }
2399 EXPORT_SYMBOL(queue_work_on); 1842 EXPORT_SYMBOL(queue_work_on);
2400 1843
2401 /** 1844 /**
2402 * select_numa_node_cpu - Select a CPU based 1845 * select_numa_node_cpu - Select a CPU based on NUMA node
2403 * @node: NUMA node ID that we want to select 1846 * @node: NUMA node ID that we want to select a CPU from
2404 * 1847 *
2405 * This function will attempt to find a "rand 1848 * This function will attempt to find a "random" cpu available on a given
2406 * node. If there are no CPUs available on th 1849 * node. If there are no CPUs available on the given node it will return
2407 * WORK_CPU_UNBOUND indicating that we should 1850 * WORK_CPU_UNBOUND indicating that we should just schedule to any
2408 * available CPU if we need to schedule this 1851 * available CPU if we need to schedule this work.
2409 */ 1852 */
2410 static int select_numa_node_cpu(int node) 1853 static int select_numa_node_cpu(int node)
2411 { 1854 {
2412 int cpu; 1855 int cpu;
2413 1856
2414 /* Delay binding to CPU if node is no 1857 /* Delay binding to CPU if node is not valid or online */
2415 if (node < 0 || node >= MAX_NUMNODES 1858 if (node < 0 || node >= MAX_NUMNODES || !node_online(node))
2416 return WORK_CPU_UNBOUND; 1859 return WORK_CPU_UNBOUND;
2417 1860
2418 /* Use local node/cpu if we are alrea 1861 /* Use local node/cpu if we are already there */
2419 cpu = raw_smp_processor_id(); 1862 cpu = raw_smp_processor_id();
2420 if (node == cpu_to_node(cpu)) 1863 if (node == cpu_to_node(cpu))
2421 return cpu; 1864 return cpu;
2422 1865
2423 /* Use "random" otherwise know as "fi 1866 /* Use "random" otherwise know as "first" online CPU of node */
2424 cpu = cpumask_any_and(cpumask_of_node 1867 cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask);
2425 1868
2426 /* If CPU is valid return that, other 1869 /* If CPU is valid return that, otherwise just defer */
2427 return cpu < nr_cpu_ids ? cpu : WORK_ 1870 return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND;
2428 } 1871 }
2429 1872
2430 /** 1873 /**
2431 * queue_work_node - queue work on a "random" 1874 * queue_work_node - queue work on a "random" cpu for a given NUMA node
2432 * @node: NUMA node that we are targeting the 1875 * @node: NUMA node that we are targeting the work for
2433 * @wq: workqueue to use 1876 * @wq: workqueue to use
2434 * @work: work to queue 1877 * @work: work to queue
2435 * 1878 *
2436 * We queue the work to a "random" CPU within 1879 * We queue the work to a "random" CPU within a given NUMA node. The basic
2437 * idea here is to provide a way to somehow a 1880 * idea here is to provide a way to somehow associate work with a given
2438 * NUMA node. 1881 * NUMA node.
2439 * 1882 *
2440 * This function will only make a best effort 1883 * This function will only make a best effort attempt at getting this onto
2441 * the right NUMA node. If no node is request 1884 * the right NUMA node. If no node is requested or the requested node is
2442 * offline then we just fall back to standard 1885 * offline then we just fall back to standard queue_work behavior.
2443 * 1886 *
2444 * Currently the "random" CPU ends up being t 1887 * Currently the "random" CPU ends up being the first available CPU in the
2445 * intersection of cpu_online_mask and the cp 1888 * intersection of cpu_online_mask and the cpumask of the node, unless we
2446 * are running on the node. In that case we j 1889 * are running on the node. In that case we just use the current CPU.
2447 * 1890 *
2448 * Return: %false if @work was already on a q 1891 * Return: %false if @work was already on a queue, %true otherwise.
2449 */ 1892 */
2450 bool queue_work_node(int node, struct workque 1893 bool queue_work_node(int node, struct workqueue_struct *wq,
2451 struct work_struct *work 1894 struct work_struct *work)
2452 { 1895 {
2453 unsigned long irq_flags; !! 1896 unsigned long flags;
2454 bool ret = false; 1897 bool ret = false;
2455 1898
2456 /* 1899 /*
2457 * This current implementation is spe 1900 * This current implementation is specific to unbound workqueues.
2458 * Specifically we only return the fi 1901 * Specifically we only return the first available CPU for a given
2459 * node instead of cycling through in 1902 * node instead of cycling through individual CPUs within the node.
2460 * 1903 *
2461 * If this is used with a per-cpu wor 1904 * If this is used with a per-cpu workqueue then the logic in
2462 * workqueue_select_cpu_near would ne 1905 * workqueue_select_cpu_near would need to be updated to allow for
2463 * some round robin type logic. 1906 * some round robin type logic.
2464 */ 1907 */
2465 WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND 1908 WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND));
2466 1909
2467 local_irq_save(irq_flags); !! 1910 local_irq_save(flags);
2468 1911
2469 if (!test_and_set_bit(WORK_STRUCT_PEN !! 1912 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
2470 !clear_pending_if_disabled(work)) <<
2471 int cpu = select_numa_node_cp 1913 int cpu = select_numa_node_cpu(node);
2472 1914
2473 __queue_work(cpu, wq, work); 1915 __queue_work(cpu, wq, work);
2474 ret = true; 1916 ret = true;
2475 } 1917 }
2476 1918
2477 local_irq_restore(irq_flags); !! 1919 local_irq_restore(flags);
2478 return ret; 1920 return ret;
2479 } 1921 }
2480 EXPORT_SYMBOL_GPL(queue_work_node); 1922 EXPORT_SYMBOL_GPL(queue_work_node);
2481 1923
2482 void delayed_work_timer_fn(struct timer_list 1924 void delayed_work_timer_fn(struct timer_list *t)
2483 { 1925 {
2484 struct delayed_work *dwork = from_tim 1926 struct delayed_work *dwork = from_timer(dwork, t, timer);
2485 1927
2486 /* should have been called from irqsa 1928 /* should have been called from irqsafe timer with irq already off */
2487 __queue_work(dwork->cpu, dwork->wq, & 1929 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2488 } 1930 }
2489 EXPORT_SYMBOL(delayed_work_timer_fn); 1931 EXPORT_SYMBOL(delayed_work_timer_fn);
2490 1932
2491 static void __queue_delayed_work(int cpu, str 1933 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
2492 struct delaye 1934 struct delayed_work *dwork, unsigned long delay)
2493 { 1935 {
2494 struct timer_list *timer = &dwork->ti 1936 struct timer_list *timer = &dwork->timer;
2495 struct work_struct *work = &dwork->wo 1937 struct work_struct *work = &dwork->work;
2496 1938
2497 WARN_ON_ONCE(!wq); 1939 WARN_ON_ONCE(!wq);
2498 WARN_ON_ONCE(timer->function != delay 1940 WARN_ON_ONCE(timer->function != delayed_work_timer_fn);
2499 WARN_ON_ONCE(timer_pending(timer)); 1941 WARN_ON_ONCE(timer_pending(timer));
2500 WARN_ON_ONCE(!list_empty(&work->entry 1942 WARN_ON_ONCE(!list_empty(&work->entry));
2501 1943
2502 /* 1944 /*
2503 * If @delay is 0, queue @dwork->work 1945 * If @delay is 0, queue @dwork->work immediately. This is for
2504 * both optimization and correctness. 1946 * both optimization and correctness. The earliest @timer can
2505 * expire is on the closest next tick 1947 * expire is on the closest next tick and delayed_work users depend
2506 * on that there's no such delay when 1948 * on that there's no such delay when @delay is 0.
2507 */ 1949 */
2508 if (!delay) { 1950 if (!delay) {
2509 __queue_work(cpu, wq, &dwork- 1951 __queue_work(cpu, wq, &dwork->work);
2510 return; 1952 return;
2511 } 1953 }
2512 1954
2513 dwork->wq = wq; 1955 dwork->wq = wq;
2514 dwork->cpu = cpu; 1956 dwork->cpu = cpu;
2515 timer->expires = jiffies + delay; 1957 timer->expires = jiffies + delay;
2516 1958
2517 if (housekeeping_enabled(HK_TYPE_TIME !! 1959 if (unlikely(cpu != WORK_CPU_UNBOUND))
2518 /* If the current cpu is a ho <<
2519 cpu = smp_processor_id(); <<
2520 if (!housekeeping_test_cpu(cp <<
2521 cpu = housekeeping_an <<
2522 add_timer_on(timer, cpu); 1960 add_timer_on(timer, cpu);
2523 } else { !! 1961 else
2524 if (likely(cpu == WORK_CPU_UN !! 1962 add_timer(timer);
2525 add_timer_global(time <<
2526 else <<
2527 add_timer_on(timer, c <<
2528 } <<
2529 } 1963 }
2530 1964
2531 /** 1965 /**
2532 * queue_delayed_work_on - queue work on spec 1966 * queue_delayed_work_on - queue work on specific CPU after delay
2533 * @cpu: CPU number to execute work on 1967 * @cpu: CPU number to execute work on
2534 * @wq: workqueue to use 1968 * @wq: workqueue to use
2535 * @dwork: work to queue 1969 * @dwork: work to queue
2536 * @delay: number of jiffies to wait before q 1970 * @delay: number of jiffies to wait before queueing
2537 * 1971 *
2538 * Return: %false if @work was already on a q 1972 * Return: %false if @work was already on a queue, %true otherwise. If
2539 * @delay is zero and @dwork is idle, it will 1973 * @delay is zero and @dwork is idle, it will be scheduled for immediate
2540 * execution. 1974 * execution.
2541 */ 1975 */
2542 bool queue_delayed_work_on(int cpu, struct wo 1976 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
2543 struct delayed_wor 1977 struct delayed_work *dwork, unsigned long delay)
2544 { 1978 {
2545 struct work_struct *work = &dwork->wo 1979 struct work_struct *work = &dwork->work;
2546 bool ret = false; 1980 bool ret = false;
2547 unsigned long irq_flags; !! 1981 unsigned long flags;
2548 1982
2549 /* read the comment in __queue_work() 1983 /* read the comment in __queue_work() */
2550 local_irq_save(irq_flags); !! 1984 local_irq_save(flags);
2551 1985
2552 if (!test_and_set_bit(WORK_STRUCT_PEN !! 1986 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
2553 !clear_pending_if_disabled(work)) <<
2554 __queue_delayed_work(cpu, wq, 1987 __queue_delayed_work(cpu, wq, dwork, delay);
2555 ret = true; 1988 ret = true;
2556 } 1989 }
2557 1990
2558 local_irq_restore(irq_flags); !! 1991 local_irq_restore(flags);
2559 return ret; 1992 return ret;
2560 } 1993 }
2561 EXPORT_SYMBOL(queue_delayed_work_on); 1994 EXPORT_SYMBOL(queue_delayed_work_on);
2562 1995
2563 /** 1996 /**
2564 * mod_delayed_work_on - modify delay of or q 1997 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
2565 * @cpu: CPU number to execute work on 1998 * @cpu: CPU number to execute work on
2566 * @wq: workqueue to use 1999 * @wq: workqueue to use
2567 * @dwork: work to queue 2000 * @dwork: work to queue
2568 * @delay: number of jiffies to wait before q 2001 * @delay: number of jiffies to wait before queueing
2569 * 2002 *
2570 * If @dwork is idle, equivalent to queue_del 2003 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
2571 * modify @dwork's timer so that it expires a 2004 * modify @dwork's timer so that it expires after @delay. If @delay is
2572 * zero, @work is guaranteed to be scheduled 2005 * zero, @work is guaranteed to be scheduled immediately regardless of its
2573 * current state. 2006 * current state.
2574 * 2007 *
2575 * Return: %false if @dwork was idle and queu 2008 * Return: %false if @dwork was idle and queued, %true if @dwork was
2576 * pending and its timer was modified. 2009 * pending and its timer was modified.
2577 * 2010 *
2578 * This function is safe to call from any con 2011 * This function is safe to call from any context including IRQ handler.
2579 * See try_to_grab_pending() for details. 2012 * See try_to_grab_pending() for details.
2580 */ 2013 */
2581 bool mod_delayed_work_on(int cpu, struct work 2014 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
2582 struct delayed_work 2015 struct delayed_work *dwork, unsigned long delay)
2583 { 2016 {
2584 unsigned long irq_flags; !! 2017 unsigned long flags;
2585 bool ret; !! 2018 int ret;
2586 2019
2587 ret = work_grab_pending(&dwork->work, !! 2020 do {
>> 2021 ret = try_to_grab_pending(&dwork->work, true, &flags);
>> 2022 } while (unlikely(ret == -EAGAIN));
2588 2023
2589 if (!clear_pending_if_disabled(&dwork !! 2024 if (likely(ret >= 0)) {
2590 __queue_delayed_work(cpu, wq, 2025 __queue_delayed_work(cpu, wq, dwork, delay);
>> 2026 local_irq_restore(flags);
>> 2027 }
2591 2028
2592 local_irq_restore(irq_flags); !! 2029 /* -ENOENT from try_to_grab_pending() becomes %true */
2593 return ret; 2030 return ret;
2594 } 2031 }
2595 EXPORT_SYMBOL_GPL(mod_delayed_work_on); 2032 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
2596 2033
2597 static void rcu_work_rcufn(struct rcu_head *r 2034 static void rcu_work_rcufn(struct rcu_head *rcu)
2598 { 2035 {
2599 struct rcu_work *rwork = container_of 2036 struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu);
2600 2037
2601 /* read the comment in __queue_work() 2038 /* read the comment in __queue_work() */
2602 local_irq_disable(); 2039 local_irq_disable();
2603 __queue_work(WORK_CPU_UNBOUND, rwork- 2040 __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work);
2604 local_irq_enable(); 2041 local_irq_enable();
2605 } 2042 }
2606 2043
2607 /** 2044 /**
2608 * queue_rcu_work - queue work after a RCU gr 2045 * queue_rcu_work - queue work after a RCU grace period
2609 * @wq: workqueue to use 2046 * @wq: workqueue to use
2610 * @rwork: work to queue 2047 * @rwork: work to queue
2611 * 2048 *
2612 * Return: %false if @rwork was already pendi 2049 * Return: %false if @rwork was already pending, %true otherwise. Note
2613 * that a full RCU grace period is guaranteed 2050 * that a full RCU grace period is guaranteed only after a %true return.
2614 * While @rwork is guaranteed to be executed 2051 * While @rwork is guaranteed to be executed after a %false return, the
2615 * execution may happen before a full RCU gra 2052 * execution may happen before a full RCU grace period has passed.
2616 */ 2053 */
2617 bool queue_rcu_work(struct workqueue_struct * 2054 bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork)
2618 { 2055 {
2619 struct work_struct *work = &rwork->wo 2056 struct work_struct *work = &rwork->work;
2620 2057
2621 /* !! 2058 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
2622 * rcu_work can't be canceled or disa <<
2623 * inside @rwork and disabled the inn <<
2624 */ <<
2625 if (!test_and_set_bit(WORK_STRUCT_PEN <<
2626 !WARN_ON_ONCE(clear_pending_if_di <<
2627 rwork->wq = wq; 2059 rwork->wq = wq;
2628 call_rcu_hurry(&rwork->rcu, r 2060 call_rcu_hurry(&rwork->rcu, rcu_work_rcufn);
2629 return true; 2061 return true;
2630 } 2062 }
2631 2063
2632 return false; 2064 return false;
2633 } 2065 }
2634 EXPORT_SYMBOL(queue_rcu_work); 2066 EXPORT_SYMBOL(queue_rcu_work);
2635 2067
2636 static struct worker *alloc_worker(int node) 2068 static struct worker *alloc_worker(int node)
2637 { 2069 {
2638 struct worker *worker; 2070 struct worker *worker;
2639 2071
2640 worker = kzalloc_node(sizeof(*worker) 2072 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
2641 if (worker) { 2073 if (worker) {
2642 INIT_LIST_HEAD(&worker->entry 2074 INIT_LIST_HEAD(&worker->entry);
2643 INIT_LIST_HEAD(&worker->sched 2075 INIT_LIST_HEAD(&worker->scheduled);
2644 INIT_LIST_HEAD(&worker->node) 2076 INIT_LIST_HEAD(&worker->node);
2645 /* on creation a worker is in 2077 /* on creation a worker is in !idle && prep state */
2646 worker->flags = WORKER_PREP; 2078 worker->flags = WORKER_PREP;
2647 } 2079 }
2648 return worker; 2080 return worker;
2649 } 2081 }
2650 2082
2651 static cpumask_t *pool_allowed_cpus(struct wo 2083 static cpumask_t *pool_allowed_cpus(struct worker_pool *pool)
2652 { 2084 {
2653 if (pool->cpu < 0 && pool->attrs->aff 2085 if (pool->cpu < 0 && pool->attrs->affn_strict)
2654 return pool->attrs->__pod_cpu 2086 return pool->attrs->__pod_cpumask;
2655 else 2087 else
2656 return pool->attrs->cpumask; 2088 return pool->attrs->cpumask;
2657 } 2089 }
2658 2090
2659 /** 2091 /**
2660 * worker_attach_to_pool() - attach a worker 2092 * worker_attach_to_pool() - attach a worker to a pool
2661 * @worker: worker to be attached 2093 * @worker: worker to be attached
2662 * @pool: the target pool 2094 * @pool: the target pool
2663 * 2095 *
2664 * Attach @worker to @pool. Once attached, t 2096 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
2665 * cpu-binding of @worker are kept coordinate 2097 * cpu-binding of @worker are kept coordinated with the pool across
2666 * cpu-[un]hotplugs. 2098 * cpu-[un]hotplugs.
2667 */ 2099 */
2668 static void worker_attach_to_pool(struct work 2100 static void worker_attach_to_pool(struct worker *worker,
2669 struct work !! 2101 struct worker_pool *pool)
2670 { 2102 {
2671 mutex_lock(&wq_pool_attach_mutex); 2103 mutex_lock(&wq_pool_attach_mutex);
2672 2104
2673 /* 2105 /*
2674 * The wq_pool_attach_mutex ensures % !! 2106 * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains
2675 * across this function. See the comm !! 2107 * stable across this function. See the comments above the flag
2676 * details. BH workers are, while per !! 2108 * definition for details.
2677 */ 2109 */
2678 if (pool->flags & POOL_DISASSOCIATED) !! 2110 if (pool->flags & POOL_DISASSOCIATED)
2679 worker->flags |= WORKER_UNBOU 2111 worker->flags |= WORKER_UNBOUND;
2680 } else { !! 2112 else
2681 WARN_ON_ONCE(pool->flags & PO <<
2682 kthread_set_per_cpu(worker->t 2113 kthread_set_per_cpu(worker->task, pool->cpu);
2683 } <<
2684 2114
2685 if (worker->rescue_wq) 2115 if (worker->rescue_wq)
2686 set_cpus_allowed_ptr(worker-> 2116 set_cpus_allowed_ptr(worker->task, pool_allowed_cpus(pool));
2687 2117
2688 list_add_tail(&worker->node, &pool->w 2118 list_add_tail(&worker->node, &pool->workers);
2689 worker->pool = pool; 2119 worker->pool = pool;
2690 2120
2691 mutex_unlock(&wq_pool_attach_mutex); 2121 mutex_unlock(&wq_pool_attach_mutex);
2692 } 2122 }
2693 2123
2694 static void unbind_worker(struct worker *work <<
2695 { <<
2696 lockdep_assert_held(&wq_pool_attach_m <<
2697 <<
2698 kthread_set_per_cpu(worker->task, -1) <<
2699 if (cpumask_intersects(wq_unbound_cpu <<
2700 WARN_ON_ONCE(set_cpus_allowed <<
2701 else <<
2702 WARN_ON_ONCE(set_cpus_allowed <<
2703 } <<
2704 <<
2705 <<
2706 static void detach_worker(struct worker *work <<
2707 { <<
2708 lockdep_assert_held(&wq_pool_attach_m <<
2709 <<
2710 unbind_worker(worker); <<
2711 list_del(&worker->node); <<
2712 } <<
2713 <<
2714 /** 2124 /**
2715 * worker_detach_from_pool() - detach a worke 2125 * worker_detach_from_pool() - detach a worker from its pool
2716 * @worker: worker which is attached to its p 2126 * @worker: worker which is attached to its pool
2717 * 2127 *
2718 * Undo the attaching which had been done in 2128 * Undo the attaching which had been done in worker_attach_to_pool(). The
2719 * caller worker shouldn't access to the pool 2129 * caller worker shouldn't access to the pool after detached except it has
2720 * other reference to the pool. 2130 * other reference to the pool.
2721 */ 2131 */
2722 static void worker_detach_from_pool(struct wo 2132 static void worker_detach_from_pool(struct worker *worker)
2723 { 2133 {
2724 struct worker_pool *pool = worker->po 2134 struct worker_pool *pool = worker->pool;
2725 !! 2135 struct completion *detach_completion = NULL;
2726 /* there is one permanent BH worker p <<
2727 WARN_ON_ONCE(pool->flags & POOL_BH); <<
2728 2136
2729 mutex_lock(&wq_pool_attach_mutex); 2137 mutex_lock(&wq_pool_attach_mutex);
2730 detach_worker(worker); !! 2138
>> 2139 kthread_set_per_cpu(worker->task, -1);
>> 2140 list_del(&worker->node);
2731 worker->pool = NULL; 2141 worker->pool = NULL;
>> 2142
>> 2143 if (list_empty(&pool->workers) && list_empty(&pool->dying_workers))
>> 2144 detach_completion = pool->detach_completion;
2732 mutex_unlock(&wq_pool_attach_mutex); 2145 mutex_unlock(&wq_pool_attach_mutex);
2733 2146
2734 /* clear leftover flags without pool- 2147 /* clear leftover flags without pool->lock after it is detached */
2735 worker->flags &= ~(WORKER_UNBOUND | W 2148 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
2736 } <<
2737 <<
2738 static int format_worker_id(char *buf, size_t <<
2739 struct worker_poo <<
2740 { <<
2741 if (worker->rescue_wq) <<
2742 return scnprintf(buf, size, " <<
2743 worker->resc <<
2744 2149
2745 if (pool) { !! 2150 if (detach_completion)
2746 if (pool->cpu >= 0) !! 2151 complete(detach_completion);
2747 return scnprintf(buf, <<
2748 pool <<
2749 pool <<
2750 else <<
2751 return scnprintf(buf, <<
2752 pool <<
2753 } else { <<
2754 return scnprintf(buf, size, " <<
2755 } <<
2756 } 2152 }
2757 2153
2758 /** 2154 /**
2759 * create_worker - create a new workqueue wor 2155 * create_worker - create a new workqueue worker
2760 * @pool: pool the new worker will belong to 2156 * @pool: pool the new worker will belong to
2761 * 2157 *
2762 * Create and start a new worker which is att 2158 * Create and start a new worker which is attached to @pool.
2763 * 2159 *
2764 * CONTEXT: 2160 * CONTEXT:
2765 * Might sleep. Does GFP_KERNEL allocations. 2161 * Might sleep. Does GFP_KERNEL allocations.
2766 * 2162 *
2767 * Return: 2163 * Return:
2768 * Pointer to the newly created worker. 2164 * Pointer to the newly created worker.
2769 */ 2165 */
2770 static struct worker *create_worker(struct wo 2166 static struct worker *create_worker(struct worker_pool *pool)
2771 { 2167 {
2772 struct worker *worker; 2168 struct worker *worker;
2773 int id; 2169 int id;
>> 2170 char id_buf[23];
2774 2171
2775 /* ID is needed to determine kthread 2172 /* ID is needed to determine kthread name */
2776 id = ida_alloc(&pool->worker_ida, GFP 2173 id = ida_alloc(&pool->worker_ida, GFP_KERNEL);
2777 if (id < 0) { 2174 if (id < 0) {
2778 pr_err_once("workqueue: Faile 2175 pr_err_once("workqueue: Failed to allocate a worker ID: %pe\n",
2779 ERR_PTR(id)); 2176 ERR_PTR(id));
2780 return NULL; 2177 return NULL;
2781 } 2178 }
2782 2179
2783 worker = alloc_worker(pool->node); 2180 worker = alloc_worker(pool->node);
2784 if (!worker) { 2181 if (!worker) {
2785 pr_err_once("workqueue: Faile 2182 pr_err_once("workqueue: Failed to allocate a worker\n");
2786 goto fail; 2183 goto fail;
2787 } 2184 }
2788 2185
2789 worker->id = id; 2186 worker->id = id;
2790 2187
2791 if (!(pool->flags & POOL_BH)) { !! 2188 if (pool->cpu >= 0)
2792 char id_buf[WORKER_ID_LEN]; !! 2189 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
>> 2190 pool->attrs->nice < 0 ? "H" : "");
>> 2191 else
>> 2192 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
2793 2193
2794 format_worker_id(id_buf, size !! 2194 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
2795 worker->task = kthread_create !! 2195 "kworker/%s", id_buf);
2796 !! 2196 if (IS_ERR(worker->task)) {
2797 if (IS_ERR(worker->task)) { !! 2197 if (PTR_ERR(worker->task) == -EINTR) {
2798 if (PTR_ERR(worker->t !! 2198 pr_err("workqueue: Interrupted when creating a worker thread \"kworker/%s\"\n",
2799 pr_err("workq !! 2199 id_buf);
2800 id_buf !! 2200 } else {
2801 } else { !! 2201 pr_err_once("workqueue: Failed to create a worker thread: %pe",
2802 pr_err_once(" !! 2202 worker->task);
2803 w <<
2804 } <<
2805 goto fail; <<
2806 } 2203 }
2807 !! 2204 goto fail;
2808 set_user_nice(worker->task, p <<
2809 kthread_bind_mask(worker->tas <<
2810 } 2205 }
2811 2206
>> 2207 set_user_nice(worker->task, pool->attrs->nice);
>> 2208 kthread_bind_mask(worker->task, pool_allowed_cpus(pool));
>> 2209
2812 /* successful, attach the worker to t 2210 /* successful, attach the worker to the pool */
2813 worker_attach_to_pool(worker, pool); 2211 worker_attach_to_pool(worker, pool);
2814 2212
2815 /* start the newly created worker */ 2213 /* start the newly created worker */
2816 raw_spin_lock_irq(&pool->lock); 2214 raw_spin_lock_irq(&pool->lock);
2817 2215
2818 worker->pool->nr_workers++; 2216 worker->pool->nr_workers++;
2819 worker_enter_idle(worker); 2217 worker_enter_idle(worker);
>> 2218 kick_pool(pool);
2820 2219
2821 /* 2220 /*
2822 * @worker is waiting on a completion 2221 * @worker is waiting on a completion in kthread() and will trigger hung
2823 * check if not woken up soon. As kic !! 2222 * check if not woken up soon. As kick_pool() might not have waken it
2824 * wake it up explicitly. !! 2223 * up, wake it up explicitly once more.
2825 */ 2224 */
2826 if (worker->task) !! 2225 wake_up_process(worker->task);
2827 wake_up_process(worker->task) <<
2828 2226
2829 raw_spin_unlock_irq(&pool->lock); 2227 raw_spin_unlock_irq(&pool->lock);
2830 2228
2831 return worker; 2229 return worker;
2832 2230
2833 fail: 2231 fail:
2834 ida_free(&pool->worker_ida, id); 2232 ida_free(&pool->worker_ida, id);
2835 kfree(worker); 2233 kfree(worker);
2836 return NULL; 2234 return NULL;
2837 } 2235 }
2838 2236
2839 static void detach_dying_workers(struct list_ !! 2237 static void unbind_worker(struct worker *worker)
2840 { 2238 {
2841 struct worker *worker; !! 2239 lockdep_assert_held(&wq_pool_attach_mutex);
2842 2240
2843 list_for_each_entry(worker, cull_list !! 2241 kthread_set_per_cpu(worker->task, -1);
2844 detach_worker(worker); !! 2242 if (cpumask_intersects(wq_unbound_cpumask, cpu_active_mask))
>> 2243 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, wq_unbound_cpumask) < 0);
>> 2244 else
>> 2245 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0);
2845 } 2246 }
2846 2247
2847 static void reap_dying_workers(struct list_he !! 2248 static void wake_dying_workers(struct list_head *cull_list)
2848 { 2249 {
2849 struct worker *worker, *tmp; 2250 struct worker *worker, *tmp;
2850 2251
2851 list_for_each_entry_safe(worker, tmp, 2252 list_for_each_entry_safe(worker, tmp, cull_list, entry) {
2852 list_del_init(&worker->entry) 2253 list_del_init(&worker->entry);
2853 kthread_stop_put(worker->task !! 2254 unbind_worker(worker);
2854 kfree(worker); !! 2255 /*
>> 2256 * If the worker was somehow already running, then it had to be
>> 2257 * in pool->idle_list when set_worker_dying() happened or we
>> 2258 * wouldn't have gotten here.
>> 2259 *
>> 2260 * Thus, the worker must either have observed the WORKER_DIE
>> 2261 * flag, or have set its state to TASK_IDLE. Either way, the
>> 2262 * below will be observed by the worker and is safe to do
>> 2263 * outside of pool->lock.
>> 2264 */
>> 2265 wake_up_process(worker->task);
2855 } 2266 }
2856 } 2267 }
2857 2268
2858 /** 2269 /**
2859 * set_worker_dying - Tag a worker for destru 2270 * set_worker_dying - Tag a worker for destruction
2860 * @worker: worker to be destroyed 2271 * @worker: worker to be destroyed
2861 * @list: transfer worker away from its pool- 2272 * @list: transfer worker away from its pool->idle_list and into list
2862 * 2273 *
2863 * Tag @worker for destruction and adjust @po 2274 * Tag @worker for destruction and adjust @pool stats accordingly. The worker
2864 * should be idle. 2275 * should be idle.
2865 * 2276 *
2866 * CONTEXT: 2277 * CONTEXT:
2867 * raw_spin_lock_irq(pool->lock). 2278 * raw_spin_lock_irq(pool->lock).
2868 */ 2279 */
2869 static void set_worker_dying(struct worker *w 2280 static void set_worker_dying(struct worker *worker, struct list_head *list)
2870 { 2281 {
2871 struct worker_pool *pool = worker->po 2282 struct worker_pool *pool = worker->pool;
2872 2283
2873 lockdep_assert_held(&pool->lock); 2284 lockdep_assert_held(&pool->lock);
2874 lockdep_assert_held(&wq_pool_attach_m 2285 lockdep_assert_held(&wq_pool_attach_mutex);
2875 2286
2876 /* sanity check frenzy */ 2287 /* sanity check frenzy */
2877 if (WARN_ON(worker->current_work) || 2288 if (WARN_ON(worker->current_work) ||
2878 WARN_ON(!list_empty(&worker->sche 2289 WARN_ON(!list_empty(&worker->scheduled)) ||
2879 WARN_ON(!(worker->flags & WORKER_ 2290 WARN_ON(!(worker->flags & WORKER_IDLE)))
2880 return; 2291 return;
2881 2292
2882 pool->nr_workers--; 2293 pool->nr_workers--;
2883 pool->nr_idle--; 2294 pool->nr_idle--;
2884 2295
2885 worker->flags |= WORKER_DIE; 2296 worker->flags |= WORKER_DIE;
2886 2297
2887 list_move(&worker->entry, list); 2298 list_move(&worker->entry, list);
2888 !! 2299 list_move(&worker->node, &pool->dying_workers);
2889 /* get an extra task struct reference <<
2890 get_task_struct(worker->task); <<
2891 } 2300 }
2892 2301
2893 /** 2302 /**
2894 * idle_worker_timeout - check if some idle w 2303 * idle_worker_timeout - check if some idle workers can now be deleted.
2895 * @t: The pool's idle_timer that just expire 2304 * @t: The pool's idle_timer that just expired
2896 * 2305 *
2897 * The timer is armed in worker_enter_idle(). 2306 * The timer is armed in worker_enter_idle(). Note that it isn't disarmed in
2898 * worker_leave_idle(), as a worker flicking 2307 * worker_leave_idle(), as a worker flicking between idle and active while its
2899 * pool is at the too_many_workers() tipping 2308 * pool is at the too_many_workers() tipping point would cause too much timer
2900 * housekeeping overhead. Since IDLE_WORKER_T 2309 * housekeeping overhead. Since IDLE_WORKER_TIMEOUT is long enough, we just let
2901 * it expire and re-evaluate things from ther 2310 * it expire and re-evaluate things from there.
2902 */ 2311 */
2903 static void idle_worker_timeout(struct timer_ 2312 static void idle_worker_timeout(struct timer_list *t)
2904 { 2313 {
2905 struct worker_pool *pool = from_timer 2314 struct worker_pool *pool = from_timer(pool, t, idle_timer);
2906 bool do_cull = false; 2315 bool do_cull = false;
2907 2316
2908 if (work_pending(&pool->idle_cull_wor 2317 if (work_pending(&pool->idle_cull_work))
2909 return; 2318 return;
2910 2319
2911 raw_spin_lock_irq(&pool->lock); 2320 raw_spin_lock_irq(&pool->lock);
2912 2321
2913 if (too_many_workers(pool)) { 2322 if (too_many_workers(pool)) {
2914 struct worker *worker; 2323 struct worker *worker;
2915 unsigned long expires; 2324 unsigned long expires;
2916 2325
2917 /* idle_list is kept in LIFO 2326 /* idle_list is kept in LIFO order, check the last one */
2918 worker = list_last_entry(&poo !! 2327 worker = list_entry(pool->idle_list.prev, struct worker, entry);
2919 expires = worker->last_active 2328 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2920 do_cull = !time_before(jiffie 2329 do_cull = !time_before(jiffies, expires);
2921 2330
2922 if (!do_cull) 2331 if (!do_cull)
2923 mod_timer(&pool->idle 2332 mod_timer(&pool->idle_timer, expires);
2924 } 2333 }
2925 raw_spin_unlock_irq(&pool->lock); 2334 raw_spin_unlock_irq(&pool->lock);
2926 2335
2927 if (do_cull) 2336 if (do_cull)
2928 queue_work(system_unbound_wq, 2337 queue_work(system_unbound_wq, &pool->idle_cull_work);
2929 } 2338 }
2930 2339
2931 /** 2340 /**
2932 * idle_cull_fn - cull workers that have been 2341 * idle_cull_fn - cull workers that have been idle for too long.
2933 * @work: the pool's work for handling these 2342 * @work: the pool's work for handling these idle workers
2934 * 2343 *
2935 * This goes through a pool's idle workers an 2344 * This goes through a pool's idle workers and gets rid of those that have been
2936 * idle for at least IDLE_WORKER_TIMEOUT seco 2345 * idle for at least IDLE_WORKER_TIMEOUT seconds.
2937 * 2346 *
2938 * We don't want to disturb isolated CPUs bec 2347 * We don't want to disturb isolated CPUs because of a pcpu kworker being
2939 * culled, so this also resets worker affinit 2348 * culled, so this also resets worker affinity. This requires a sleepable
2940 * context, hence the split between timer cal 2349 * context, hence the split between timer callback and work item.
2941 */ 2350 */
2942 static void idle_cull_fn(struct work_struct * 2351 static void idle_cull_fn(struct work_struct *work)
2943 { 2352 {
2944 struct worker_pool *pool = container_ 2353 struct worker_pool *pool = container_of(work, struct worker_pool, idle_cull_work);
2945 LIST_HEAD(cull_list); 2354 LIST_HEAD(cull_list);
2946 2355
2947 /* 2356 /*
2948 * Grabbing wq_pool_attach_mutex here 2357 * Grabbing wq_pool_attach_mutex here ensures an already-running worker
2949 * cannot proceed beyong set_pf_worke !! 2358 * cannot proceed beyong worker_detach_from_pool() in its self-destruct
2950 * This is required as a previously-p !! 2359 * path. This is required as a previously-preempted worker could run after
2951 * set_worker_dying() has happened bu !! 2360 * set_worker_dying() has happened but before wake_dying_workers() did.
2952 */ 2361 */
2953 mutex_lock(&wq_pool_attach_mutex); 2362 mutex_lock(&wq_pool_attach_mutex);
2954 raw_spin_lock_irq(&pool->lock); 2363 raw_spin_lock_irq(&pool->lock);
2955 2364
2956 while (too_many_workers(pool)) { 2365 while (too_many_workers(pool)) {
2957 struct worker *worker; 2366 struct worker *worker;
2958 unsigned long expires; 2367 unsigned long expires;
2959 2368
2960 worker = list_last_entry(&poo !! 2369 worker = list_entry(pool->idle_list.prev, struct worker, entry);
2961 expires = worker->last_active 2370 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2962 2371
2963 if (time_before(jiffies, expi 2372 if (time_before(jiffies, expires)) {
2964 mod_timer(&pool->idle 2373 mod_timer(&pool->idle_timer, expires);
2965 break; 2374 break;
2966 } 2375 }
2967 2376
2968 set_worker_dying(worker, &cul 2377 set_worker_dying(worker, &cull_list);
2969 } 2378 }
2970 2379
2971 raw_spin_unlock_irq(&pool->lock); 2380 raw_spin_unlock_irq(&pool->lock);
2972 detach_dying_workers(&cull_list); !! 2381 wake_dying_workers(&cull_list);
2973 mutex_unlock(&wq_pool_attach_mutex); 2382 mutex_unlock(&wq_pool_attach_mutex);
2974 <<
2975 reap_dying_workers(&cull_list); <<
2976 } 2383 }
2977 2384
2978 static void send_mayday(struct work_struct *w 2385 static void send_mayday(struct work_struct *work)
2979 { 2386 {
2980 struct pool_workqueue *pwq = get_work 2387 struct pool_workqueue *pwq = get_work_pwq(work);
2981 struct workqueue_struct *wq = pwq->wq 2388 struct workqueue_struct *wq = pwq->wq;
2982 2389
2983 lockdep_assert_held(&wq_mayday_lock); 2390 lockdep_assert_held(&wq_mayday_lock);
2984 2391
2985 if (!wq->rescuer) 2392 if (!wq->rescuer)
2986 return; 2393 return;
2987 2394
2988 /* mayday mayday mayday */ 2395 /* mayday mayday mayday */
2989 if (list_empty(&pwq->mayday_node)) { 2396 if (list_empty(&pwq->mayday_node)) {
2990 /* 2397 /*
2991 * If @pwq is for an unbound 2398 * If @pwq is for an unbound wq, its base ref may be put at
2992 * any time due to an attribu 2399 * any time due to an attribute change. Pin @pwq until the
2993 * rescuer is done with it. 2400 * rescuer is done with it.
2994 */ 2401 */
2995 get_pwq(pwq); 2402 get_pwq(pwq);
2996 list_add_tail(&pwq->mayday_no 2403 list_add_tail(&pwq->mayday_node, &wq->maydays);
2997 wake_up_process(wq->rescuer-> 2404 wake_up_process(wq->rescuer->task);
2998 pwq->stats[PWQ_STAT_MAYDAY]++ 2405 pwq->stats[PWQ_STAT_MAYDAY]++;
2999 } 2406 }
3000 } 2407 }
3001 2408
3002 static void pool_mayday_timeout(struct timer_ 2409 static void pool_mayday_timeout(struct timer_list *t)
3003 { 2410 {
3004 struct worker_pool *pool = from_timer 2411 struct worker_pool *pool = from_timer(pool, t, mayday_timer);
3005 struct work_struct *work; 2412 struct work_struct *work;
3006 2413
3007 raw_spin_lock_irq(&pool->lock); 2414 raw_spin_lock_irq(&pool->lock);
3008 raw_spin_lock(&wq_mayday_lock); 2415 raw_spin_lock(&wq_mayday_lock); /* for wq->maydays */
3009 2416
3010 if (need_to_create_worker(pool)) { 2417 if (need_to_create_worker(pool)) {
3011 /* 2418 /*
3012 * We've been trying to creat 2419 * We've been trying to create a new worker but
3013 * haven't been successful. 2420 * haven't been successful. We might be hitting an
3014 * allocation deadlock. Send 2421 * allocation deadlock. Send distress signals to
3015 * rescuers. 2422 * rescuers.
3016 */ 2423 */
3017 list_for_each_entry(work, &po 2424 list_for_each_entry(work, &pool->worklist, entry)
3018 send_mayday(work); 2425 send_mayday(work);
3019 } 2426 }
3020 2427
3021 raw_spin_unlock(&wq_mayday_lock); 2428 raw_spin_unlock(&wq_mayday_lock);
3022 raw_spin_unlock_irq(&pool->lock); 2429 raw_spin_unlock_irq(&pool->lock);
3023 2430
3024 mod_timer(&pool->mayday_timer, jiffie 2431 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
3025 } 2432 }
3026 2433
3027 /** 2434 /**
3028 * maybe_create_worker - create a new worker 2435 * maybe_create_worker - create a new worker if necessary
3029 * @pool: pool to create a new worker for 2436 * @pool: pool to create a new worker for
3030 * 2437 *
3031 * Create a new worker for @pool if necessary 2438 * Create a new worker for @pool if necessary. @pool is guaranteed to
3032 * have at least one idle worker on return fr 2439 * have at least one idle worker on return from this function. If
3033 * creating a new worker takes longer than MA 2440 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
3034 * sent to all rescuers with works scheduled 2441 * sent to all rescuers with works scheduled on @pool to resolve
3035 * possible allocation deadlock. 2442 * possible allocation deadlock.
3036 * 2443 *
3037 * On return, need_to_create_worker() is guar 2444 * On return, need_to_create_worker() is guaranteed to be %false and
3038 * may_start_working() %true. 2445 * may_start_working() %true.
3039 * 2446 *
3040 * LOCKING: 2447 * LOCKING:
3041 * raw_spin_lock_irq(pool->lock) which may be 2448 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
3042 * multiple times. Does GFP_KERNEL allocatio 2449 * multiple times. Does GFP_KERNEL allocations. Called only from
3043 * manager. 2450 * manager.
3044 */ 2451 */
3045 static void maybe_create_worker(struct worker 2452 static void maybe_create_worker(struct worker_pool *pool)
3046 __releases(&pool->lock) 2453 __releases(&pool->lock)
3047 __acquires(&pool->lock) 2454 __acquires(&pool->lock)
3048 { 2455 {
3049 restart: 2456 restart:
3050 raw_spin_unlock_irq(&pool->lock); 2457 raw_spin_unlock_irq(&pool->lock);
3051 2458
3052 /* if we don't make progress in MAYDA 2459 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
3053 mod_timer(&pool->mayday_timer, jiffie 2460 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
3054 2461
3055 while (true) { 2462 while (true) {
3056 if (create_worker(pool) || !n 2463 if (create_worker(pool) || !need_to_create_worker(pool))
3057 break; 2464 break;
3058 2465
3059 schedule_timeout_interruptibl 2466 schedule_timeout_interruptible(CREATE_COOLDOWN);
3060 2467
3061 if (!need_to_create_worker(po 2468 if (!need_to_create_worker(pool))
3062 break; 2469 break;
3063 } 2470 }
3064 2471
3065 del_timer_sync(&pool->mayday_timer); 2472 del_timer_sync(&pool->mayday_timer);
3066 raw_spin_lock_irq(&pool->lock); 2473 raw_spin_lock_irq(&pool->lock);
3067 /* 2474 /*
3068 * This is necessary even after a new 2475 * This is necessary even after a new worker was just successfully
3069 * created as @pool->lock was dropped 2476 * created as @pool->lock was dropped and the new worker might have
3070 * already become busy. 2477 * already become busy.
3071 */ 2478 */
3072 if (need_to_create_worker(pool)) 2479 if (need_to_create_worker(pool))
3073 goto restart; 2480 goto restart;
3074 } 2481 }
3075 2482
3076 /** 2483 /**
3077 * manage_workers - manage worker pool 2484 * manage_workers - manage worker pool
3078 * @worker: self 2485 * @worker: self
3079 * 2486 *
3080 * Assume the manager role and manage the wor 2487 * Assume the manager role and manage the worker pool @worker belongs
3081 * to. At any given time, there can be only 2488 * to. At any given time, there can be only zero or one manager per
3082 * pool. The exclusion is handled automatica 2489 * pool. The exclusion is handled automatically by this function.
3083 * 2490 *
3084 * The caller can safely start processing wor 2491 * The caller can safely start processing works on false return. On
3085 * true return, it's guaranteed that need_to_ 2492 * true return, it's guaranteed that need_to_create_worker() is false
3086 * and may_start_working() is true. 2493 * and may_start_working() is true.
3087 * 2494 *
3088 * CONTEXT: 2495 * CONTEXT:
3089 * raw_spin_lock_irq(pool->lock) which may be 2496 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
3090 * multiple times. Does GFP_KERNEL allocatio 2497 * multiple times. Does GFP_KERNEL allocations.
3091 * 2498 *
3092 * Return: 2499 * Return:
3093 * %false if the pool doesn't need management 2500 * %false if the pool doesn't need management and the caller can safely
3094 * start processing works, %true if managemen 2501 * start processing works, %true if management function was performed and
3095 * the conditions that the caller verified be 2502 * the conditions that the caller verified before calling the function may
3096 * no longer be true. 2503 * no longer be true.
3097 */ 2504 */
3098 static bool manage_workers(struct worker *wor 2505 static bool manage_workers(struct worker *worker)
3099 { 2506 {
3100 struct worker_pool *pool = worker->po 2507 struct worker_pool *pool = worker->pool;
3101 2508
3102 if (pool->flags & POOL_MANAGER_ACTIVE 2509 if (pool->flags & POOL_MANAGER_ACTIVE)
3103 return false; 2510 return false;
3104 2511
3105 pool->flags |= POOL_MANAGER_ACTIVE; 2512 pool->flags |= POOL_MANAGER_ACTIVE;
3106 pool->manager = worker; 2513 pool->manager = worker;
3107 2514
3108 maybe_create_worker(pool); 2515 maybe_create_worker(pool);
3109 2516
3110 pool->manager = NULL; 2517 pool->manager = NULL;
3111 pool->flags &= ~POOL_MANAGER_ACTIVE; 2518 pool->flags &= ~POOL_MANAGER_ACTIVE;
3112 rcuwait_wake_up(&manager_wait); 2519 rcuwait_wake_up(&manager_wait);
3113 return true; 2520 return true;
3114 } 2521 }
3115 2522
3116 /** 2523 /**
3117 * process_one_work - process single work 2524 * process_one_work - process single work
3118 * @worker: self 2525 * @worker: self
3119 * @work: work to process 2526 * @work: work to process
3120 * 2527 *
3121 * Process @work. This function contains all 2528 * Process @work. This function contains all the logics necessary to
3122 * process a single work including synchroniz 2529 * process a single work including synchronization against and
3123 * interaction with other workers on the same 2530 * interaction with other workers on the same cpu, queueing and
3124 * flushing. As long as context requirement 2531 * flushing. As long as context requirement is met, any worker can
3125 * call this function to process a work. 2532 * call this function to process a work.
3126 * 2533 *
3127 * CONTEXT: 2534 * CONTEXT:
3128 * raw_spin_lock_irq(pool->lock) which is rel 2535 * raw_spin_lock_irq(pool->lock) which is released and regrabbed.
3129 */ 2536 */
3130 static void process_one_work(struct worker *w 2537 static void process_one_work(struct worker *worker, struct work_struct *work)
3131 __releases(&pool->lock) 2538 __releases(&pool->lock)
3132 __acquires(&pool->lock) 2539 __acquires(&pool->lock)
3133 { 2540 {
3134 struct pool_workqueue *pwq = get_work 2541 struct pool_workqueue *pwq = get_work_pwq(work);
3135 struct worker_pool *pool = worker->po 2542 struct worker_pool *pool = worker->pool;
3136 unsigned long work_data; 2543 unsigned long work_data;
3137 int lockdep_start_depth, rcu_start_de <<
3138 bool bh_draining = pool->flags & POOL <<
3139 #ifdef CONFIG_LOCKDEP 2544 #ifdef CONFIG_LOCKDEP
3140 /* 2545 /*
3141 * It is permissible to free the stru 2546 * It is permissible to free the struct work_struct from
3142 * inside the function that is called 2547 * inside the function that is called from it, this we need to
3143 * take into account for lockdep too. 2548 * take into account for lockdep too. To avoid bogus "held
3144 * lock freed" warnings as well as pr 2549 * lock freed" warnings as well as problems when looking into
3145 * work->lockdep_map, make a copy and 2550 * work->lockdep_map, make a copy and use that here.
3146 */ 2551 */
3147 struct lockdep_map lockdep_map; 2552 struct lockdep_map lockdep_map;
3148 2553
3149 lockdep_copy_map(&lockdep_map, &work- 2554 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
3150 #endif 2555 #endif
3151 /* ensure we're on the correct CPU */ 2556 /* ensure we're on the correct CPU */
3152 WARN_ON_ONCE(!(pool->flags & POOL_DIS 2557 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
3153 raw_smp_processor_id() ! 2558 raw_smp_processor_id() != pool->cpu);
3154 2559
3155 /* claim and dequeue */ 2560 /* claim and dequeue */
3156 debug_work_deactivate(work); 2561 debug_work_deactivate(work);
3157 hash_add(pool->busy_hash, &worker->he 2562 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
3158 worker->current_work = work; 2563 worker->current_work = work;
3159 worker->current_func = work->func; 2564 worker->current_func = work->func;
3160 worker->current_pwq = pwq; 2565 worker->current_pwq = pwq;
3161 if (worker->task) !! 2566 worker->current_at = worker->task->se.sum_exec_runtime;
3162 worker->current_at = worker-> <<
3163 work_data = *work_data_bits(work); 2567 work_data = *work_data_bits(work);
3164 worker->current_color = get_work_colo 2568 worker->current_color = get_work_color(work_data);
3165 2569
3166 /* 2570 /*
3167 * Record wq name for cmdline and deb 2571 * Record wq name for cmdline and debug reporting, may get
3168 * overridden through set_worker_desc 2572 * overridden through set_worker_desc().
3169 */ 2573 */
3170 strscpy(worker->desc, pwq->wq->name, 2574 strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN);
3171 2575
3172 list_del_init(&work->entry); 2576 list_del_init(&work->entry);
3173 2577
3174 /* 2578 /*
3175 * CPU intensive works don't particip 2579 * CPU intensive works don't participate in concurrency management.
3176 * They're the scheduler's responsibi 2580 * They're the scheduler's responsibility. This takes @worker out
3177 * of concurrency management and the 2581 * of concurrency management and the next code block will chain
3178 * execution of the pending work item 2582 * execution of the pending work items.
3179 */ 2583 */
3180 if (unlikely(pwq->wq->flags & WQ_CPU_ 2584 if (unlikely(pwq->wq->flags & WQ_CPU_INTENSIVE))
3181 worker_set_flags(worker, WORK 2585 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
3182 2586
3183 /* 2587 /*
3184 * Kick @pool if necessary. It's alwa 2588 * Kick @pool if necessary. It's always noop for per-cpu worker pools
3185 * since nr_running would always be > 2589 * since nr_running would always be >= 1 at this point. This is used to
3186 * chain execution of the pending wor 2590 * chain execution of the pending work items for WORKER_NOT_RUNNING
3187 * workers such as the UNBOUND and CP 2591 * workers such as the UNBOUND and CPU_INTENSIVE ones.
3188 */ 2592 */
3189 kick_pool(pool); 2593 kick_pool(pool);
3190 2594
3191 /* 2595 /*
3192 * Record the last pool and clear PEN 2596 * Record the last pool and clear PENDING which should be the last
3193 * update to @work. Also, do this in 2597 * update to @work. Also, do this inside @pool->lock so that
3194 * PENDING and queued state changes h 2598 * PENDING and queued state changes happen together while IRQ is
3195 * disabled. 2599 * disabled.
3196 */ 2600 */
3197 set_work_pool_and_clear_pending(work, !! 2601 set_work_pool_and_clear_pending(work, pool->id);
3198 2602
3199 pwq->stats[PWQ_STAT_STARTED]++; 2603 pwq->stats[PWQ_STAT_STARTED]++;
3200 raw_spin_unlock_irq(&pool->lock); 2604 raw_spin_unlock_irq(&pool->lock);
3201 2605
3202 rcu_start_depth = rcu_preempt_depth() !! 2606 lock_map_acquire(&pwq->wq->lockdep_map);
3203 lockdep_start_depth = lockdep_depth(c <<
3204 /* see drain_dead_softirq_workfn() */ <<
3205 if (!bh_draining) <<
3206 lock_map_acquire(&pwq->wq->lo <<
3207 lock_map_acquire(&lockdep_map); 2607 lock_map_acquire(&lockdep_map);
3208 /* 2608 /*
3209 * Strictly speaking we should mark t 2609 * Strictly speaking we should mark the invariant state without holding
3210 * any locks, that is, before these t 2610 * any locks, that is, before these two lock_map_acquire()'s.
3211 * 2611 *
3212 * However, that would result in: 2612 * However, that would result in:
3213 * 2613 *
3214 * A(W1) 2614 * A(W1)
3215 * WFC(C) 2615 * WFC(C)
3216 * A(W1) 2616 * A(W1)
3217 * C(C) 2617 * C(C)
3218 * 2618 *
3219 * Which would create W1->C->W1 depen 2619 * Which would create W1->C->W1 dependencies, even though there is no
3220 * actual deadlock possible. There ar 2620 * actual deadlock possible. There are two solutions, using a
3221 * read-recursive acquire on the work 2621 * read-recursive acquire on the work(queue) 'locks', but this will then
3222 * hit the lockdep limitation on recu 2622 * hit the lockdep limitation on recursive locks, or simply discard
3223 * these locks. 2623 * these locks.
3224 * 2624 *
3225 * AFAICT there is no possible deadlo 2625 * AFAICT there is no possible deadlock scenario between the
3226 * flush_work() and complete() primit 2626 * flush_work() and complete() primitives (except for single-threaded
3227 * workqueues), so hiding them isn't 2627 * workqueues), so hiding them isn't a problem.
3228 */ 2628 */
3229 lockdep_invariant_state(true); 2629 lockdep_invariant_state(true);
3230 trace_workqueue_execute_start(work); 2630 trace_workqueue_execute_start(work);
3231 worker->current_func(work); 2631 worker->current_func(work);
3232 /* 2632 /*
3233 * While we must be careful to not us 2633 * While we must be careful to not use "work" after this, the trace
3234 * point will only record its address 2634 * point will only record its address.
3235 */ 2635 */
3236 trace_workqueue_execute_end(work, wor 2636 trace_workqueue_execute_end(work, worker->current_func);
3237 pwq->stats[PWQ_STAT_COMPLETED]++; 2637 pwq->stats[PWQ_STAT_COMPLETED]++;
3238 lock_map_release(&lockdep_map); 2638 lock_map_release(&lockdep_map);
3239 if (!bh_draining) !! 2639 lock_map_release(&pwq->wq->lockdep_map);
3240 lock_map_release(&pwq->wq->lo <<
3241 2640
3242 if (unlikely((worker->task && in_atom !! 2641 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
3243 lockdep_depth(current) ! !! 2642 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
3244 rcu_preempt_depth() != r !! 2643 " last function: %ps\n",
3245 pr_err("BUG: workqueue leaked !! 2644 current->comm, preempt_count(), task_pid_nr(current),
3246 " preempt=0x%08x l <<
3247 current->comm, task_pi <<
3248 lockdep_start_depth, l <<
3249 rcu_start_depth, rcu_p <<
3250 worker->current_func); 2645 worker->current_func);
3251 debug_show_held_locks(current 2646 debug_show_held_locks(current);
3252 dump_stack(); 2647 dump_stack();
3253 } 2648 }
3254 2649
3255 /* 2650 /*
3256 * The following prevents a kworker f 2651 * The following prevents a kworker from hogging CPU on !PREEMPTION
3257 * kernels, where a requeueing work i 2652 * kernels, where a requeueing work item waiting for something to
3258 * happen could deadlock with stop_ma 2653 * happen could deadlock with stop_machine as such work item could
3259 * indefinitely requeue itself while 2654 * indefinitely requeue itself while all other CPUs are trapped in
3260 * stop_machine. At the same time, re 2655 * stop_machine. At the same time, report a quiescent RCU state so
3261 * the same condition doesn't freeze 2656 * the same condition doesn't freeze RCU.
3262 */ 2657 */
3263 if (worker->task) !! 2658 cond_resched();
3264 cond_resched(); <<
3265 2659
3266 raw_spin_lock_irq(&pool->lock); 2660 raw_spin_lock_irq(&pool->lock);
3267 2661
3268 /* 2662 /*
3269 * In addition to %WQ_CPU_INTENSIVE, 2663 * In addition to %WQ_CPU_INTENSIVE, @worker may also have been marked
3270 * CPU intensive by wq_worker_tick() 2664 * CPU intensive by wq_worker_tick() if @work hogged CPU longer than
3271 * wq_cpu_intensive_thresh_us. Clear 2665 * wq_cpu_intensive_thresh_us. Clear it.
3272 */ 2666 */
3273 worker_clr_flags(worker, WORKER_CPU_I 2667 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
3274 2668
3275 /* tag the worker for identification 2669 /* tag the worker for identification in schedule() */
3276 worker->last_func = worker->current_f 2670 worker->last_func = worker->current_func;
3277 2671
3278 /* we're done with it, release */ 2672 /* we're done with it, release */
3279 hash_del(&worker->hentry); 2673 hash_del(&worker->hentry);
3280 worker->current_work = NULL; 2674 worker->current_work = NULL;
3281 worker->current_func = NULL; 2675 worker->current_func = NULL;
3282 worker->current_pwq = NULL; 2676 worker->current_pwq = NULL;
3283 worker->current_color = INT_MAX; 2677 worker->current_color = INT_MAX;
3284 <<
3285 /* must be the last step, see the fun <<
3286 pwq_dec_nr_in_flight(pwq, work_data); 2678 pwq_dec_nr_in_flight(pwq, work_data);
3287 } 2679 }
3288 2680
3289 /** 2681 /**
3290 * process_scheduled_works - process schedule 2682 * process_scheduled_works - process scheduled works
3291 * @worker: self 2683 * @worker: self
3292 * 2684 *
3293 * Process all scheduled works. Please note 2685 * Process all scheduled works. Please note that the scheduled list
3294 * may change while processing a work, so thi 2686 * may change while processing a work, so this function repeatedly
3295 * fetches a work from the top and executes i 2687 * fetches a work from the top and executes it.
3296 * 2688 *
3297 * CONTEXT: 2689 * CONTEXT:
3298 * raw_spin_lock_irq(pool->lock) which may be 2690 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
3299 * multiple times. 2691 * multiple times.
3300 */ 2692 */
3301 static void process_scheduled_works(struct wo 2693 static void process_scheduled_works(struct worker *worker)
3302 { 2694 {
3303 struct work_struct *work; 2695 struct work_struct *work;
3304 bool first = true; 2696 bool first = true;
3305 2697
3306 while ((work = list_first_entry_or_nu 2698 while ((work = list_first_entry_or_null(&worker->scheduled,
3307 2699 struct work_struct, entry))) {
3308 if (first) { 2700 if (first) {
3309 worker->pool->watchdo 2701 worker->pool->watchdog_ts = jiffies;
3310 first = false; 2702 first = false;
3311 } 2703 }
3312 process_one_work(worker, work 2704 process_one_work(worker, work);
3313 } 2705 }
3314 } 2706 }
3315 2707
3316 static void set_pf_worker(bool val) 2708 static void set_pf_worker(bool val)
3317 { 2709 {
3318 mutex_lock(&wq_pool_attach_mutex); 2710 mutex_lock(&wq_pool_attach_mutex);
3319 if (val) 2711 if (val)
3320 current->flags |= PF_WQ_WORKE 2712 current->flags |= PF_WQ_WORKER;
3321 else 2713 else
3322 current->flags &= ~PF_WQ_WORK 2714 current->flags &= ~PF_WQ_WORKER;
3323 mutex_unlock(&wq_pool_attach_mutex); 2715 mutex_unlock(&wq_pool_attach_mutex);
3324 } 2716 }
3325 2717
3326 /** 2718 /**
3327 * worker_thread - the worker thread function 2719 * worker_thread - the worker thread function
3328 * @__worker: self 2720 * @__worker: self
3329 * 2721 *
3330 * The worker thread function. All workers b 2722 * The worker thread function. All workers belong to a worker_pool -
3331 * either a per-cpu one or dynamic unbound on 2723 * either a per-cpu one or dynamic unbound one. These workers process all
3332 * work items regardless of their specific ta 2724 * work items regardless of their specific target workqueue. The only
3333 * exception is work items which belong to wo 2725 * exception is work items which belong to workqueues with a rescuer which
3334 * will be explained in rescuer_thread(). 2726 * will be explained in rescuer_thread().
3335 * 2727 *
3336 * Return: 0 2728 * Return: 0
3337 */ 2729 */
3338 static int worker_thread(void *__worker) 2730 static int worker_thread(void *__worker)
3339 { 2731 {
3340 struct worker *worker = __worker; 2732 struct worker *worker = __worker;
3341 struct worker_pool *pool = worker->po 2733 struct worker_pool *pool = worker->pool;
3342 2734
3343 /* tell the scheduler that this is a 2735 /* tell the scheduler that this is a workqueue worker */
3344 set_pf_worker(true); 2736 set_pf_worker(true);
3345 woke_up: 2737 woke_up:
3346 raw_spin_lock_irq(&pool->lock); 2738 raw_spin_lock_irq(&pool->lock);
3347 2739
3348 /* am I supposed to die? */ 2740 /* am I supposed to die? */
3349 if (unlikely(worker->flags & WORKER_D 2741 if (unlikely(worker->flags & WORKER_DIE)) {
3350 raw_spin_unlock_irq(&pool->lo 2742 raw_spin_unlock_irq(&pool->lock);
3351 set_pf_worker(false); 2743 set_pf_worker(false);
3352 /* !! 2744
3353 * The worker is dead and PF_ !! 2745 set_task_comm(worker->task, "kworker/dying");
3354 * shouldn't be accessed, res <<
3355 */ <<
3356 worker->pool = NULL; <<
3357 ida_free(&pool->worker_ida, w 2746 ida_free(&pool->worker_ida, worker->id);
>> 2747 worker_detach_from_pool(worker);
>> 2748 WARN_ON_ONCE(!list_empty(&worker->entry));
>> 2749 kfree(worker);
3358 return 0; 2750 return 0;
3359 } 2751 }
3360 2752
3361 worker_leave_idle(worker); 2753 worker_leave_idle(worker);
3362 recheck: 2754 recheck:
3363 /* no more worker necessary? */ 2755 /* no more worker necessary? */
3364 if (!need_more_worker(pool)) 2756 if (!need_more_worker(pool))
3365 goto sleep; 2757 goto sleep;
3366 2758
3367 /* do we need to manage? */ 2759 /* do we need to manage? */
3368 if (unlikely(!may_start_working(pool) 2760 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
3369 goto recheck; 2761 goto recheck;
3370 2762
3371 /* 2763 /*
3372 * ->scheduled list can only be fille 2764 * ->scheduled list can only be filled while a worker is
3373 * preparing to process a work or act 2765 * preparing to process a work or actually processing it.
3374 * Make sure nobody diddled with it w 2766 * Make sure nobody diddled with it while I was sleeping.
3375 */ 2767 */
3376 WARN_ON_ONCE(!list_empty(&worker->sch 2768 WARN_ON_ONCE(!list_empty(&worker->scheduled));
3377 2769
3378 /* 2770 /*
3379 * Finish PREP stage. We're guarante 2771 * Finish PREP stage. We're guaranteed to have at least one idle
3380 * worker or that someone else has al 2772 * worker or that someone else has already assumed the manager
3381 * role. This is where @worker start 2773 * role. This is where @worker starts participating in concurrency
3382 * management if applicable and concu 2774 * management if applicable and concurrency management is restored
3383 * after being rebound. See rebind_w 2775 * after being rebound. See rebind_workers() for details.
3384 */ 2776 */
3385 worker_clr_flags(worker, WORKER_PREP 2777 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
3386 2778
3387 do { 2779 do {
3388 struct work_struct *work = 2780 struct work_struct *work =
3389 list_first_entry(&poo 2781 list_first_entry(&pool->worklist,
3390 stru 2782 struct work_struct, entry);
3391 2783
3392 if (assign_work(work, worker, 2784 if (assign_work(work, worker, NULL))
3393 process_scheduled_wor 2785 process_scheduled_works(worker);
3394 } while (keep_working(pool)); 2786 } while (keep_working(pool));
3395 2787
3396 worker_set_flags(worker, WORKER_PREP) 2788 worker_set_flags(worker, WORKER_PREP);
3397 sleep: 2789 sleep:
3398 /* 2790 /*
3399 * pool->lock is held and there's no 2791 * pool->lock is held and there's no work to process and no need to
3400 * manage, sleep. Workers are woken 2792 * manage, sleep. Workers are woken up only while holding
3401 * pool->lock or from local cpu, so s 2793 * pool->lock or from local cpu, so setting the current state
3402 * before releasing pool->lock is eno 2794 * before releasing pool->lock is enough to prevent losing any
3403 * event. 2795 * event.
3404 */ 2796 */
3405 worker_enter_idle(worker); 2797 worker_enter_idle(worker);
3406 __set_current_state(TASK_IDLE); 2798 __set_current_state(TASK_IDLE);
3407 raw_spin_unlock_irq(&pool->lock); 2799 raw_spin_unlock_irq(&pool->lock);
3408 schedule(); 2800 schedule();
3409 goto woke_up; 2801 goto woke_up;
3410 } 2802 }
3411 2803
3412 /** 2804 /**
3413 * rescuer_thread - the rescuer thread functi 2805 * rescuer_thread - the rescuer thread function
3414 * @__rescuer: self 2806 * @__rescuer: self
3415 * 2807 *
3416 * Workqueue rescuer thread function. There' 2808 * Workqueue rescuer thread function. There's one rescuer for each
3417 * workqueue which has WQ_MEM_RECLAIM set. 2809 * workqueue which has WQ_MEM_RECLAIM set.
3418 * 2810 *
3419 * Regular work processing on a pool may bloc 2811 * Regular work processing on a pool may block trying to create a new
3420 * worker which uses GFP_KERNEL allocation wh 2812 * worker which uses GFP_KERNEL allocation which has slight chance of
3421 * developing into deadlock if some works cur 2813 * developing into deadlock if some works currently on the same queue
3422 * need to be processed to satisfy the GFP_KE 2814 * need to be processed to satisfy the GFP_KERNEL allocation. This is
3423 * the problem rescuer solves. 2815 * the problem rescuer solves.
3424 * 2816 *
3425 * When such condition is possible, the pool 2817 * When such condition is possible, the pool summons rescuers of all
3426 * workqueues which have works queued on the 2818 * workqueues which have works queued on the pool and let them process
3427 * those works so that forward progress can b 2819 * those works so that forward progress can be guaranteed.
3428 * 2820 *
3429 * This should happen rarely. 2821 * This should happen rarely.
3430 * 2822 *
3431 * Return: 0 2823 * Return: 0
3432 */ 2824 */
3433 static int rescuer_thread(void *__rescuer) 2825 static int rescuer_thread(void *__rescuer)
3434 { 2826 {
3435 struct worker *rescuer = __rescuer; 2827 struct worker *rescuer = __rescuer;
3436 struct workqueue_struct *wq = rescuer 2828 struct workqueue_struct *wq = rescuer->rescue_wq;
3437 bool should_stop; 2829 bool should_stop;
3438 2830
3439 set_user_nice(current, RESCUER_NICE_L 2831 set_user_nice(current, RESCUER_NICE_LEVEL);
3440 2832
3441 /* 2833 /*
3442 * Mark rescuer as worker too. As WO 2834 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
3443 * doesn't participate in concurrency 2835 * doesn't participate in concurrency management.
3444 */ 2836 */
3445 set_pf_worker(true); 2837 set_pf_worker(true);
3446 repeat: 2838 repeat:
3447 set_current_state(TASK_IDLE); 2839 set_current_state(TASK_IDLE);
3448 2840
3449 /* 2841 /*
3450 * By the time the rescuer is request 2842 * By the time the rescuer is requested to stop, the workqueue
3451 * shouldn't have any work pending, b 2843 * shouldn't have any work pending, but @wq->maydays may still have
3452 * pwq(s) queued. This can happen by 2844 * pwq(s) queued. This can happen by non-rescuer workers consuming
3453 * all the work items before the resc 2845 * all the work items before the rescuer got to them. Go through
3454 * @wq->maydays processing before act 2846 * @wq->maydays processing before acting on should_stop so that the
3455 * list is always empty on exit. 2847 * list is always empty on exit.
3456 */ 2848 */
3457 should_stop = kthread_should_stop(); 2849 should_stop = kthread_should_stop();
3458 2850
3459 /* see whether any pwq is asking for 2851 /* see whether any pwq is asking for help */
3460 raw_spin_lock_irq(&wq_mayday_lock); 2852 raw_spin_lock_irq(&wq_mayday_lock);
3461 2853
3462 while (!list_empty(&wq->maydays)) { 2854 while (!list_empty(&wq->maydays)) {
3463 struct pool_workqueue *pwq = 2855 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
3464 struc 2856 struct pool_workqueue, mayday_node);
3465 struct worker_pool *pool = pw 2857 struct worker_pool *pool = pwq->pool;
3466 struct work_struct *work, *n; 2858 struct work_struct *work, *n;
3467 2859
3468 __set_current_state(TASK_RUNN 2860 __set_current_state(TASK_RUNNING);
3469 list_del_init(&pwq->mayday_no 2861 list_del_init(&pwq->mayday_node);
3470 2862
3471 raw_spin_unlock_irq(&wq_mayda 2863 raw_spin_unlock_irq(&wq_mayday_lock);
3472 2864
3473 worker_attach_to_pool(rescuer 2865 worker_attach_to_pool(rescuer, pool);
3474 2866
3475 raw_spin_lock_irq(&pool->lock 2867 raw_spin_lock_irq(&pool->lock);
3476 2868
3477 /* 2869 /*
3478 * Slurp in all works issued 2870 * Slurp in all works issued via this workqueue and
3479 * process'em. 2871 * process'em.
3480 */ 2872 */
3481 WARN_ON_ONCE(!list_empty(&res 2873 WARN_ON_ONCE(!list_empty(&rescuer->scheduled));
3482 list_for_each_entry_safe(work 2874 list_for_each_entry_safe(work, n, &pool->worklist, entry) {
3483 if (get_work_pwq(work 2875 if (get_work_pwq(work) == pwq &&
3484 assign_work(work, 2876 assign_work(work, rescuer, &n))
3485 pwq->stats[PW 2877 pwq->stats[PWQ_STAT_RESCUED]++;
3486 } 2878 }
3487 2879
3488 if (!list_empty(&rescuer->sch 2880 if (!list_empty(&rescuer->scheduled)) {
3489 process_scheduled_wor 2881 process_scheduled_works(rescuer);
3490 2882
3491 /* 2883 /*
3492 * The above executio 2884 * The above execution of rescued work items could
3493 * have created more 2885 * have created more to rescue through
3494 * pwq_activate_first 2886 * pwq_activate_first_inactive() or chained
3495 * queueing. Let's p 2887 * queueing. Let's put @pwq back on mayday list so
3496 * that such back-to- 2888 * that such back-to-back work items, which may be
3497 * being used to reli 2889 * being used to relieve memory pressure, don't
3498 * incur MAYDAY_INTER 2890 * incur MAYDAY_INTERVAL delay inbetween.
3499 */ 2891 */
3500 if (pwq->nr_active && 2892 if (pwq->nr_active && need_to_create_worker(pool)) {
3501 raw_spin_lock 2893 raw_spin_lock(&wq_mayday_lock);
3502 /* 2894 /*
3503 * Queue iff 2895 * Queue iff we aren't racing destruction
3504 * and somebo 2896 * and somebody else hasn't queued it already.
3505 */ 2897 */
3506 if (wq->rescu 2898 if (wq->rescuer && list_empty(&pwq->mayday_node)) {
3507 get_p 2899 get_pwq(pwq);
3508 list_ 2900 list_add_tail(&pwq->mayday_node, &wq->maydays);
3509 } 2901 }
3510 raw_spin_unlo 2902 raw_spin_unlock(&wq_mayday_lock);
3511 } 2903 }
3512 } 2904 }
3513 2905
3514 /* 2906 /*
3515 * Put the reference grabbed 2907 * Put the reference grabbed by send_mayday(). @pool won't
3516 * go away while we're still 2908 * go away while we're still attached to it.
3517 */ 2909 */
3518 put_pwq(pwq); 2910 put_pwq(pwq);
3519 2911
3520 /* 2912 /*
3521 * Leave this pool. Notify re 2913 * Leave this pool. Notify regular workers; otherwise, we end up
3522 * with 0 concurrency and sta 2914 * with 0 concurrency and stalling the execution.
3523 */ 2915 */
3524 kick_pool(pool); 2916 kick_pool(pool);
3525 2917
3526 raw_spin_unlock_irq(&pool->lo 2918 raw_spin_unlock_irq(&pool->lock);
3527 2919
3528 worker_detach_from_pool(rescu 2920 worker_detach_from_pool(rescuer);
3529 2921
3530 raw_spin_lock_irq(&wq_mayday_ 2922 raw_spin_lock_irq(&wq_mayday_lock);
3531 } 2923 }
3532 2924
3533 raw_spin_unlock_irq(&wq_mayday_lock); 2925 raw_spin_unlock_irq(&wq_mayday_lock);
3534 2926
3535 if (should_stop) { 2927 if (should_stop) {
3536 __set_current_state(TASK_RUNN 2928 __set_current_state(TASK_RUNNING);
3537 set_pf_worker(false); 2929 set_pf_worker(false);
3538 return 0; 2930 return 0;
3539 } 2931 }
3540 2932
3541 /* rescuers should never participate 2933 /* rescuers should never participate in concurrency management */
3542 WARN_ON_ONCE(!(rescuer->flags & WORKE 2934 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
3543 schedule(); 2935 schedule();
3544 goto repeat; 2936 goto repeat;
3545 } 2937 }
3546 2938
3547 static void bh_worker(struct worker *worker) <<
3548 { <<
3549 struct worker_pool *pool = worker->po <<
3550 int nr_restarts = BH_WORKER_RESTARTS; <<
3551 unsigned long end = jiffies + BH_WORK <<
3552 <<
3553 raw_spin_lock_irq(&pool->lock); <<
3554 worker_leave_idle(worker); <<
3555 <<
3556 /* <<
3557 * This function follows the structur <<
3558 * explanations on each step. <<
3559 */ <<
3560 if (!need_more_worker(pool)) <<
3561 goto done; <<
3562 <<
3563 WARN_ON_ONCE(!list_empty(&worker->sch <<
3564 worker_clr_flags(worker, WORKER_PREP <<
3565 <<
3566 do { <<
3567 struct work_struct *work = <<
3568 list_first_entry(&poo <<
3569 stru <<
3570 <<
3571 if (assign_work(work, worker, <<
3572 process_scheduled_wor <<
3573 } while (keep_working(pool) && <<
3574 --nr_restarts && time_before <<
3575 <<
3576 worker_set_flags(worker, WORKER_PREP) <<
3577 done: <<
3578 worker_enter_idle(worker); <<
3579 kick_pool(pool); <<
3580 raw_spin_unlock_irq(&pool->lock); <<
3581 } <<
3582 <<
3583 /* <<
3584 * TODO: Convert all tasklet users to workque <<
3585 * <<
3586 * This is currently called from tasklet[_hi] <<
3587 * whenever there are tasklets to run. Let's <<
3588 * queued. Once conversion from tasklet is co <<
3589 * can be dropped. <<
3590 * <<
3591 * After full conversion, we'll add worker->s <<
3592 * softirq action and obtain the worker point <<
3593 */ <<
3594 void workqueue_softirq_action(bool highpri) <<
3595 { <<
3596 struct worker_pool *pool = <<
3597 &per_cpu(bh_worker_pools, smp <<
3598 if (need_more_worker(pool)) <<
3599 bh_worker(list_first_entry(&p <<
3600 } <<
3601 <<
3602 struct wq_drain_dead_softirq_work { <<
3603 struct work_struct work; <<
3604 struct worker_pool *pool; <<
3605 struct completion done; <<
3606 }; <<
3607 <<
3608 static void drain_dead_softirq_workfn(struct <<
3609 { <<
3610 struct wq_drain_dead_softirq_work *de <<
3611 container_of(work, struct wq_ <<
3612 struct worker_pool *pool = dead_work- <<
3613 bool repeat; <<
3614 <<
3615 /* <<
3616 * @pool's CPU is dead and we want to <<
3617 * items from this BH work item which <<
3618 * its CPU is dead, @pool can't be ki <<
3619 * will be nested, a lockdep annotati <<
3620 * @pool with %POOL_BH_DRAINING for t <<
3621 */ <<
3622 raw_spin_lock_irq(&pool->lock); <<
3623 pool->flags |= POOL_BH_DRAINING; <<
3624 raw_spin_unlock_irq(&pool->lock); <<
3625 <<
3626 bh_worker(list_first_entry(&pool->wor <<
3627 <<
3628 raw_spin_lock_irq(&pool->lock); <<
3629 pool->flags &= ~POOL_BH_DRAINING; <<
3630 repeat = need_more_worker(pool); <<
3631 raw_spin_unlock_irq(&pool->lock); <<
3632 <<
3633 /* <<
3634 * bh_worker() might hit consecutive <<
3635 * still are pending work items, resc <<
3636 * don't hog this CPU's BH. <<
3637 */ <<
3638 if (repeat) { <<
3639 if (pool->attrs->nice == HIGH <<
3640 queue_work(system_bh_ <<
3641 else <<
3642 queue_work(system_bh_ <<
3643 } else { <<
3644 complete(&dead_work->done); <<
3645 } <<
3646 } <<
3647 <<
3648 /* <<
3649 * @cpu is dead. Drain the remaining BH work <<
3650 * possible to allocate dead_work per CPU and <<
3651 * have to worry about draining overlapping w <<
3652 * nesting (one CPU's dead_work queued on ano <<
3653 * on). Let's keep it simple and drain them s <<
3654 * items which shouldn't be requeued on the s <<
3655 */ <<
3656 void workqueue_softirq_dead(unsigned int cpu) <<
3657 { <<
3658 int i; <<
3659 <<
3660 for (i = 0; i < NR_STD_WORKER_POOLS; <<
3661 struct worker_pool *pool = &p <<
3662 struct wq_drain_dead_softirq_ <<
3663 <<
3664 if (!need_more_worker(pool)) <<
3665 continue; <<
3666 <<
3667 INIT_WORK_ONSTACK(&dead_work. <<
3668 dead_work.pool = pool; <<
3669 init_completion(&dead_work.do <<
3670 <<
3671 if (pool->attrs->nice == HIGH <<
3672 queue_work(system_bh_ <<
3673 else <<
3674 queue_work(system_bh_ <<
3675 <<
3676 wait_for_completion(&dead_wor <<
3677 destroy_work_on_stack(&dead_w <<
3678 } <<
3679 } <<
3680 <<
3681 /** 2939 /**
3682 * check_flush_dependency - check for flush d 2940 * check_flush_dependency - check for flush dependency sanity
3683 * @target_wq: workqueue being flushed 2941 * @target_wq: workqueue being flushed
3684 * @target_work: work item being flushed (NUL 2942 * @target_work: work item being flushed (NULL for workqueue flushes)
3685 * 2943 *
3686 * %current is trying to flush the whole @tar 2944 * %current is trying to flush the whole @target_wq or @target_work on it.
3687 * If @target_wq doesn't have %WQ_MEM_RECLAIM 2945 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
3688 * reclaiming memory or running on a workqueu 2946 * reclaiming memory or running on a workqueue which doesn't have
3689 * %WQ_MEM_RECLAIM as that can break forward- 2947 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
3690 * a deadlock. 2948 * a deadlock.
3691 */ 2949 */
3692 static void check_flush_dependency(struct wor 2950 static void check_flush_dependency(struct workqueue_struct *target_wq,
3693 struct wor 2951 struct work_struct *target_work)
3694 { 2952 {
3695 work_func_t target_func = target_work 2953 work_func_t target_func = target_work ? target_work->func : NULL;
3696 struct worker *worker; 2954 struct worker *worker;
3697 2955
3698 if (target_wq->flags & WQ_MEM_RECLAIM 2956 if (target_wq->flags & WQ_MEM_RECLAIM)
3699 return; 2957 return;
3700 2958
3701 worker = current_wq_worker(); 2959 worker = current_wq_worker();
3702 2960
3703 WARN_ONCE(current->flags & PF_MEMALLO 2961 WARN_ONCE(current->flags & PF_MEMALLOC,
3704 "workqueue: PF_MEMALLOC tas 2962 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps",
3705 current->pid, current->comm 2963 current->pid, current->comm, target_wq->name, target_func);
3706 WARN_ONCE(worker && ((worker->current 2964 WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
3707 (WQ_MEM_RECLAIM 2965 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
3708 "workqueue: WQ_MEM_RECLAIM 2966 "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps",
3709 worker->current_pwq->wq->na 2967 worker->current_pwq->wq->name, worker->current_func,
3710 target_wq->name, target_fun 2968 target_wq->name, target_func);
3711 } 2969 }
3712 2970
3713 struct wq_barrier { 2971 struct wq_barrier {
3714 struct work_struct work; 2972 struct work_struct work;
3715 struct completion done; 2973 struct completion done;
3716 struct task_struct *task; /* pu 2974 struct task_struct *task; /* purely informational */
3717 }; 2975 };
3718 2976
3719 static void wq_barrier_func(struct work_struc 2977 static void wq_barrier_func(struct work_struct *work)
3720 { 2978 {
3721 struct wq_barrier *barr = container_o 2979 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
3722 complete(&barr->done); 2980 complete(&barr->done);
3723 } 2981 }
3724 2982
3725 /** 2983 /**
3726 * insert_wq_barrier - insert a barrier work 2984 * insert_wq_barrier - insert a barrier work
3727 * @pwq: pwq to insert barrier into 2985 * @pwq: pwq to insert barrier into
3728 * @barr: wq_barrier to insert 2986 * @barr: wq_barrier to insert
3729 * @target: target work to attach @barr to 2987 * @target: target work to attach @barr to
3730 * @worker: worker currently executing @targe 2988 * @worker: worker currently executing @target, NULL if @target is not executing
3731 * 2989 *
3732 * @barr is linked to @target such that @barr 2990 * @barr is linked to @target such that @barr is completed only after
3733 * @target finishes execution. Please note t 2991 * @target finishes execution. Please note that the ordering
3734 * guarantee is observed only with respect to 2992 * guarantee is observed only with respect to @target and on the local
3735 * cpu. 2993 * cpu.
3736 * 2994 *
3737 * Currently, a queued barrier can't be cance 2995 * Currently, a queued barrier can't be canceled. This is because
3738 * try_to_grab_pending() can't determine whet 2996 * try_to_grab_pending() can't determine whether the work to be
3739 * grabbed is at the head of the queue and th 2997 * grabbed is at the head of the queue and thus can't clear LINKED
3740 * flag of the previous work while there must 2998 * flag of the previous work while there must be a valid next work
3741 * after a work with LINKED flag set. 2999 * after a work with LINKED flag set.
3742 * 3000 *
3743 * Note that when @worker is non-NULL, @targe 3001 * Note that when @worker is non-NULL, @target may be modified
3744 * underneath us, so we can't reliably determ 3002 * underneath us, so we can't reliably determine pwq from @target.
3745 * 3003 *
3746 * CONTEXT: 3004 * CONTEXT:
3747 * raw_spin_lock_irq(pool->lock). 3005 * raw_spin_lock_irq(pool->lock).
3748 */ 3006 */
3749 static void insert_wq_barrier(struct pool_wor 3007 static void insert_wq_barrier(struct pool_workqueue *pwq,
3750 struct wq_barri 3008 struct wq_barrier *barr,
3751 struct work_str 3009 struct work_struct *target, struct worker *worker)
3752 { 3010 {
3753 static __maybe_unused struct lock_cla <<
3754 unsigned int work_flags = 0; 3011 unsigned int work_flags = 0;
3755 unsigned int work_color; 3012 unsigned int work_color;
3756 struct list_head *head; 3013 struct list_head *head;
3757 3014
3758 /* 3015 /*
3759 * debugobject calls are safe here ev 3016 * debugobject calls are safe here even with pool->lock locked
3760 * as we know for sure that this will 3017 * as we know for sure that this will not trigger any of the
3761 * checks and call back into the fixu 3018 * checks and call back into the fixup functions where we
3762 * might deadlock. 3019 * might deadlock.
3763 * <<
3764 * BH and threaded workqueues need se <<
3765 * spuriously triggering "inconsisten <<
3766 * usage". <<
3767 */ 3020 */
3768 INIT_WORK_ONSTACK_KEY(&barr->work, wq !! 3021 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
3769 (pwq->wq->flags <<
3770 __set_bit(WORK_STRUCT_PENDING_BIT, wo 3022 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
3771 3023
3772 init_completion_map(&barr->done, &tar 3024 init_completion_map(&barr->done, &target->lockdep_map);
3773 3025
3774 barr->task = current; 3026 barr->task = current;
3775 3027
3776 /* The barrier work item does not par !! 3028 /* The barrier work item does not participate in pwq->nr_active. */
3777 work_flags |= WORK_STRUCT_INACTIVE; 3029 work_flags |= WORK_STRUCT_INACTIVE;
3778 3030
3779 /* 3031 /*
3780 * If @target is currently being exec 3032 * If @target is currently being executed, schedule the
3781 * barrier to the worker; otherwise, 3033 * barrier to the worker; otherwise, put it after @target.
3782 */ 3034 */
3783 if (worker) { 3035 if (worker) {
3784 head = worker->scheduled.next 3036 head = worker->scheduled.next;
3785 work_color = worker->current_ 3037 work_color = worker->current_color;
3786 } else { 3038 } else {
3787 unsigned long *bits = work_da 3039 unsigned long *bits = work_data_bits(target);
3788 3040
3789 head = target->entry.next; 3041 head = target->entry.next;
3790 /* there can already be other 3042 /* there can already be other linked works, inherit and set */
3791 work_flags |= *bits & WORK_ST 3043 work_flags |= *bits & WORK_STRUCT_LINKED;
3792 work_color = get_work_color(* 3044 work_color = get_work_color(*bits);
3793 __set_bit(WORK_STRUCT_LINKED_ 3045 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
3794 } 3046 }
3795 3047
3796 pwq->nr_in_flight[work_color]++; 3048 pwq->nr_in_flight[work_color]++;
3797 work_flags |= work_color_to_flags(wor 3049 work_flags |= work_color_to_flags(work_color);
3798 3050
3799 insert_work(pwq, &barr->work, head, w 3051 insert_work(pwq, &barr->work, head, work_flags);
3800 } 3052 }
3801 3053
3802 /** 3054 /**
3803 * flush_workqueue_prep_pwqs - prepare pwqs f 3055 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
3804 * @wq: workqueue being flushed 3056 * @wq: workqueue being flushed
3805 * @flush_color: new flush color, < 0 for no- 3057 * @flush_color: new flush color, < 0 for no-op
3806 * @work_color: new work color, < 0 for no-op 3058 * @work_color: new work color, < 0 for no-op
3807 * 3059 *
3808 * Prepare pwqs for workqueue flushing. 3060 * Prepare pwqs for workqueue flushing.
3809 * 3061 *
3810 * If @flush_color is non-negative, flush_col 3062 * If @flush_color is non-negative, flush_color on all pwqs should be
3811 * -1. If no pwq has in-flight commands at t 3063 * -1. If no pwq has in-flight commands at the specified color, all
3812 * pwq->flush_color's stay at -1 and %false i 3064 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
3813 * has in flight commands, its pwq->flush_col 3065 * has in flight commands, its pwq->flush_color is set to
3814 * @flush_color, @wq->nr_pwqs_to_flush is upd 3066 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
3815 * wakeup logic is armed and %true is returne 3067 * wakeup logic is armed and %true is returned.
3816 * 3068 *
3817 * The caller should have initialized @wq->fi 3069 * The caller should have initialized @wq->first_flusher prior to
3818 * calling this function with non-negative @f 3070 * calling this function with non-negative @flush_color. If
3819 * @flush_color is negative, no flush color u 3071 * @flush_color is negative, no flush color update is done and %false
3820 * is returned. 3072 * is returned.
3821 * 3073 *
3822 * If @work_color is non-negative, all pwqs s 3074 * If @work_color is non-negative, all pwqs should have the same
3823 * work_color which is previous to @work_colo 3075 * work_color which is previous to @work_color and all will be
3824 * advanced to @work_color. 3076 * advanced to @work_color.
3825 * 3077 *
3826 * CONTEXT: 3078 * CONTEXT:
3827 * mutex_lock(wq->mutex). 3079 * mutex_lock(wq->mutex).
3828 * 3080 *
3829 * Return: 3081 * Return:
3830 * %true if @flush_color >= 0 and there's som 3082 * %true if @flush_color >= 0 and there's something to flush. %false
3831 * otherwise. 3083 * otherwise.
3832 */ 3084 */
3833 static bool flush_workqueue_prep_pwqs(struct 3085 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
3834 int flu 3086 int flush_color, int work_color)
3835 { 3087 {
3836 bool wait = false; 3088 bool wait = false;
3837 struct pool_workqueue *pwq; 3089 struct pool_workqueue *pwq;
3838 3090
3839 if (flush_color >= 0) { 3091 if (flush_color >= 0) {
3840 WARN_ON_ONCE(atomic_read(&wq- 3092 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
3841 atomic_set(&wq->nr_pwqs_to_fl 3093 atomic_set(&wq->nr_pwqs_to_flush, 1);
3842 } 3094 }
3843 3095
3844 for_each_pwq(pwq, wq) { 3096 for_each_pwq(pwq, wq) {
3845 struct worker_pool *pool = pw 3097 struct worker_pool *pool = pwq->pool;
3846 3098
3847 raw_spin_lock_irq(&pool->lock 3099 raw_spin_lock_irq(&pool->lock);
3848 3100
3849 if (flush_color >= 0) { 3101 if (flush_color >= 0) {
3850 WARN_ON_ONCE(pwq->flu 3102 WARN_ON_ONCE(pwq->flush_color != -1);
3851 3103
3852 if (pwq->nr_in_flight 3104 if (pwq->nr_in_flight[flush_color]) {
3853 pwq->flush_co 3105 pwq->flush_color = flush_color;
3854 atomic_inc(&w 3106 atomic_inc(&wq->nr_pwqs_to_flush);
3855 wait = true; 3107 wait = true;
3856 } 3108 }
3857 } 3109 }
3858 3110
3859 if (work_color >= 0) { 3111 if (work_color >= 0) {
3860 WARN_ON_ONCE(work_col 3112 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
3861 pwq->work_color = wor 3113 pwq->work_color = work_color;
3862 } 3114 }
3863 3115
3864 raw_spin_unlock_irq(&pool->lo 3116 raw_spin_unlock_irq(&pool->lock);
3865 } 3117 }
3866 3118
3867 if (flush_color >= 0 && atomic_dec_an 3119 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
3868 complete(&wq->first_flusher-> 3120 complete(&wq->first_flusher->done);
3869 3121
3870 return wait; 3122 return wait;
3871 } 3123 }
3872 3124
3873 static void touch_wq_lockdep_map(struct workq <<
3874 { <<
3875 #ifdef CONFIG_LOCKDEP <<
3876 if (wq->flags & WQ_BH) <<
3877 local_bh_disable(); <<
3878 <<
3879 lock_map_acquire(&wq->lockdep_map); <<
3880 lock_map_release(&wq->lockdep_map); <<
3881 <<
3882 if (wq->flags & WQ_BH) <<
3883 local_bh_enable(); <<
3884 #endif <<
3885 } <<
3886 <<
3887 static void touch_work_lockdep_map(struct wor <<
3888 struct wor <<
3889 { <<
3890 #ifdef CONFIG_LOCKDEP <<
3891 if (wq->flags & WQ_BH) <<
3892 local_bh_disable(); <<
3893 <<
3894 lock_map_acquire(&work->lockdep_map); <<
3895 lock_map_release(&work->lockdep_map); <<
3896 <<
3897 if (wq->flags & WQ_BH) <<
3898 local_bh_enable(); <<
3899 #endif <<
3900 } <<
3901 <<
3902 /** 3125 /**
3903 * __flush_workqueue - ensure that any schedu 3126 * __flush_workqueue - ensure that any scheduled work has run to completion.
3904 * @wq: workqueue to flush 3127 * @wq: workqueue to flush
3905 * 3128 *
3906 * This function sleeps until all work items 3129 * This function sleeps until all work items which were queued on entry
3907 * have finished execution, but it is not liv 3130 * have finished execution, but it is not livelocked by new incoming ones.
3908 */ 3131 */
3909 void __flush_workqueue(struct workqueue_struc 3132 void __flush_workqueue(struct workqueue_struct *wq)
3910 { 3133 {
3911 struct wq_flusher this_flusher = { 3134 struct wq_flusher this_flusher = {
3912 .list = LIST_HEAD_INIT(this_f 3135 .list = LIST_HEAD_INIT(this_flusher.list),
3913 .flush_color = -1, 3136 .flush_color = -1,
3914 .done = COMPLETION_INITIALIZE 3137 .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map),
3915 }; 3138 };
3916 int next_color; 3139 int next_color;
3917 3140
3918 if (WARN_ON(!wq_online)) 3141 if (WARN_ON(!wq_online))
3919 return; 3142 return;
3920 3143
3921 touch_wq_lockdep_map(wq); !! 3144 lock_map_acquire(&wq->lockdep_map);
>> 3145 lock_map_release(&wq->lockdep_map);
3922 3146
3923 mutex_lock(&wq->mutex); 3147 mutex_lock(&wq->mutex);
3924 3148
3925 /* 3149 /*
3926 * Start-to-wait phase 3150 * Start-to-wait phase
3927 */ 3151 */
3928 next_color = work_next_color(wq->work 3152 next_color = work_next_color(wq->work_color);
3929 3153
3930 if (next_color != wq->flush_color) { 3154 if (next_color != wq->flush_color) {
3931 /* 3155 /*
3932 * Color space is not full. 3156 * Color space is not full. The current work_color
3933 * becomes our flush_color an 3157 * becomes our flush_color and work_color is advanced
3934 * by one. 3158 * by one.
3935 */ 3159 */
3936 WARN_ON_ONCE(!list_empty(&wq- 3160 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
3937 this_flusher.flush_color = wq 3161 this_flusher.flush_color = wq->work_color;
3938 wq->work_color = next_color; 3162 wq->work_color = next_color;
3939 3163
3940 if (!wq->first_flusher) { 3164 if (!wq->first_flusher) {
3941 /* no flush in progre 3165 /* no flush in progress, become the first flusher */
3942 WARN_ON_ONCE(wq->flus 3166 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
3943 3167
3944 wq->first_flusher = & 3168 wq->first_flusher = &this_flusher;
3945 3169
3946 if (!flush_workqueue_ 3170 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
3947 3171 wq->work_color)) {
3948 /* nothing to 3172 /* nothing to flush, done */
3949 wq->flush_col 3173 wq->flush_color = next_color;
3950 wq->first_flu 3174 wq->first_flusher = NULL;
3951 goto out_unlo 3175 goto out_unlock;
3952 } 3176 }
3953 } else { 3177 } else {
3954 /* wait in queue */ 3178 /* wait in queue */
3955 WARN_ON_ONCE(wq->flus 3179 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
3956 list_add_tail(&this_f 3180 list_add_tail(&this_flusher.list, &wq->flusher_queue);
3957 flush_workqueue_prep_ 3181 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
3958 } 3182 }
3959 } else { 3183 } else {
3960 /* 3184 /*
3961 * Oops, color space is full, 3185 * Oops, color space is full, wait on overflow queue.
3962 * The next flush completion 3186 * The next flush completion will assign us
3963 * flush_color and transfer t 3187 * flush_color and transfer to flusher_queue.
3964 */ 3188 */
3965 list_add_tail(&this_flusher.l 3189 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
3966 } 3190 }
3967 3191
3968 check_flush_dependency(wq, NULL); 3192 check_flush_dependency(wq, NULL);
3969 3193
3970 mutex_unlock(&wq->mutex); 3194 mutex_unlock(&wq->mutex);
3971 3195
3972 wait_for_completion(&this_flusher.don 3196 wait_for_completion(&this_flusher.done);
3973 3197
3974 /* 3198 /*
3975 * Wake-up-and-cascade phase 3199 * Wake-up-and-cascade phase
3976 * 3200 *
3977 * First flushers are responsible for 3201 * First flushers are responsible for cascading flushes and
3978 * handling overflow. Non-first flus 3202 * handling overflow. Non-first flushers can simply return.
3979 */ 3203 */
3980 if (READ_ONCE(wq->first_flusher) != & 3204 if (READ_ONCE(wq->first_flusher) != &this_flusher)
3981 return; 3205 return;
3982 3206
3983 mutex_lock(&wq->mutex); 3207 mutex_lock(&wq->mutex);
3984 3208
3985 /* we might have raced, check again w 3209 /* we might have raced, check again with mutex held */
3986 if (wq->first_flusher != &this_flushe 3210 if (wq->first_flusher != &this_flusher)
3987 goto out_unlock; 3211 goto out_unlock;
3988 3212
3989 WRITE_ONCE(wq->first_flusher, NULL); 3213 WRITE_ONCE(wq->first_flusher, NULL);
3990 3214
3991 WARN_ON_ONCE(!list_empty(&this_flushe 3215 WARN_ON_ONCE(!list_empty(&this_flusher.list));
3992 WARN_ON_ONCE(wq->flush_color != this_ 3216 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
3993 3217
3994 while (true) { 3218 while (true) {
3995 struct wq_flusher *next, *tmp 3219 struct wq_flusher *next, *tmp;
3996 3220
3997 /* complete all the flushers 3221 /* complete all the flushers sharing the current flush color */
3998 list_for_each_entry_safe(next 3222 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
3999 if (next->flush_color 3223 if (next->flush_color != wq->flush_color)
4000 break; 3224 break;
4001 list_del_init(&next-> 3225 list_del_init(&next->list);
4002 complete(&next->done) 3226 complete(&next->done);
4003 } 3227 }
4004 3228
4005 WARN_ON_ONCE(!list_empty(&wq- 3229 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
4006 wq->flush_color 3230 wq->flush_color != work_next_color(wq->work_color));
4007 3231
4008 /* this flush_color is finish 3232 /* this flush_color is finished, advance by one */
4009 wq->flush_color = work_next_c 3233 wq->flush_color = work_next_color(wq->flush_color);
4010 3234
4011 /* one color has been freed, 3235 /* one color has been freed, handle overflow queue */
4012 if (!list_empty(&wq->flusher_ 3236 if (!list_empty(&wq->flusher_overflow)) {
4013 /* 3237 /*
4014 * Assign the same co 3238 * Assign the same color to all overflowed
4015 * flushers, advance 3239 * flushers, advance work_color and append to
4016 * flusher_queue. Th 3240 * flusher_queue. This is the start-to-wait
4017 * phase for these ov 3241 * phase for these overflowed flushers.
4018 */ 3242 */
4019 list_for_each_entry(t 3243 list_for_each_entry(tmp, &wq->flusher_overflow, list)
4020 tmp->flush_co 3244 tmp->flush_color = wq->work_color;
4021 3245
4022 wq->work_color = work 3246 wq->work_color = work_next_color(wq->work_color);
4023 3247
4024 list_splice_tail_init 3248 list_splice_tail_init(&wq->flusher_overflow,
4025 3249 &wq->flusher_queue);
4026 flush_workqueue_prep_ 3250 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
4027 } 3251 }
4028 3252
4029 if (list_empty(&wq->flusher_q 3253 if (list_empty(&wq->flusher_queue)) {
4030 WARN_ON_ONCE(wq->flus 3254 WARN_ON_ONCE(wq->flush_color != wq->work_color);
4031 break; 3255 break;
4032 } 3256 }
4033 3257
4034 /* 3258 /*
4035 * Need to flush more colors. 3259 * Need to flush more colors. Make the next flusher
4036 * the new first flusher and 3260 * the new first flusher and arm pwqs.
4037 */ 3261 */
4038 WARN_ON_ONCE(wq->flush_color 3262 WARN_ON_ONCE(wq->flush_color == wq->work_color);
4039 WARN_ON_ONCE(wq->flush_color 3263 WARN_ON_ONCE(wq->flush_color != next->flush_color);
4040 3264
4041 list_del_init(&next->list); 3265 list_del_init(&next->list);
4042 wq->first_flusher = next; 3266 wq->first_flusher = next;
4043 3267
4044 if (flush_workqueue_prep_pwqs 3268 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
4045 break; 3269 break;
4046 3270
4047 /* 3271 /*
4048 * Meh... this color is alrea 3272 * Meh... this color is already done, clear first
4049 * flusher and repeat cascadi 3273 * flusher and repeat cascading.
4050 */ 3274 */
4051 wq->first_flusher = NULL; 3275 wq->first_flusher = NULL;
4052 } 3276 }
4053 3277
4054 out_unlock: 3278 out_unlock:
4055 mutex_unlock(&wq->mutex); 3279 mutex_unlock(&wq->mutex);
4056 } 3280 }
4057 EXPORT_SYMBOL(__flush_workqueue); 3281 EXPORT_SYMBOL(__flush_workqueue);
4058 3282
4059 /** 3283 /**
4060 * drain_workqueue - drain a workqueue 3284 * drain_workqueue - drain a workqueue
4061 * @wq: workqueue to drain 3285 * @wq: workqueue to drain
4062 * 3286 *
4063 * Wait until the workqueue becomes empty. W 3287 * Wait until the workqueue becomes empty. While draining is in progress,
4064 * only chain queueing is allowed. IOW, only 3288 * only chain queueing is allowed. IOW, only currently pending or running
4065 * work items on @wq can queue further work i 3289 * work items on @wq can queue further work items on it. @wq is flushed
4066 * repeatedly until it becomes empty. The nu 3290 * repeatedly until it becomes empty. The number of flushing is determined
4067 * by the depth of chaining and should be rel 3291 * by the depth of chaining and should be relatively short. Whine if it
4068 * takes too long. 3292 * takes too long.
4069 */ 3293 */
4070 void drain_workqueue(struct workqueue_struct 3294 void drain_workqueue(struct workqueue_struct *wq)
4071 { 3295 {
4072 unsigned int flush_cnt = 0; 3296 unsigned int flush_cnt = 0;
4073 struct pool_workqueue *pwq; 3297 struct pool_workqueue *pwq;
4074 3298
4075 /* 3299 /*
4076 * __queue_work() needs to test wheth 3300 * __queue_work() needs to test whether there are drainers, is much
4077 * hotter than drain_workqueue() and 3301 * hotter than drain_workqueue() and already looks at @wq->flags.
4078 * Use __WQ_DRAINING so that queue do 3302 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
4079 */ 3303 */
4080 mutex_lock(&wq->mutex); 3304 mutex_lock(&wq->mutex);
4081 if (!wq->nr_drainers++) 3305 if (!wq->nr_drainers++)
4082 wq->flags |= __WQ_DRAINING; 3306 wq->flags |= __WQ_DRAINING;
4083 mutex_unlock(&wq->mutex); 3307 mutex_unlock(&wq->mutex);
4084 reflush: 3308 reflush:
4085 __flush_workqueue(wq); 3309 __flush_workqueue(wq);
4086 3310
4087 mutex_lock(&wq->mutex); 3311 mutex_lock(&wq->mutex);
4088 3312
4089 for_each_pwq(pwq, wq) { 3313 for_each_pwq(pwq, wq) {
4090 bool drained; 3314 bool drained;
4091 3315
4092 raw_spin_lock_irq(&pwq->pool- 3316 raw_spin_lock_irq(&pwq->pool->lock);
4093 drained = pwq_is_empty(pwq); !! 3317 drained = !pwq->nr_active && list_empty(&pwq->inactive_works);
4094 raw_spin_unlock_irq(&pwq->poo 3318 raw_spin_unlock_irq(&pwq->pool->lock);
4095 3319
4096 if (drained) 3320 if (drained)
4097 continue; 3321 continue;
4098 3322
4099 if (++flush_cnt == 10 || 3323 if (++flush_cnt == 10 ||
4100 (flush_cnt % 100 == 0 && 3324 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
4101 pr_warn("workqueue %s 3325 pr_warn("workqueue %s: %s() isn't complete after %u tries\n",
4102 wq->name, __f 3326 wq->name, __func__, flush_cnt);
4103 3327
4104 mutex_unlock(&wq->mutex); 3328 mutex_unlock(&wq->mutex);
4105 goto reflush; 3329 goto reflush;
4106 } 3330 }
4107 3331
4108 if (!--wq->nr_drainers) 3332 if (!--wq->nr_drainers)
4109 wq->flags &= ~__WQ_DRAINING; 3333 wq->flags &= ~__WQ_DRAINING;
4110 mutex_unlock(&wq->mutex); 3334 mutex_unlock(&wq->mutex);
4111 } 3335 }
4112 EXPORT_SYMBOL_GPL(drain_workqueue); 3336 EXPORT_SYMBOL_GPL(drain_workqueue);
4113 3337
4114 static bool start_flush_work(struct work_stru 3338 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
4115 bool from_cancel 3339 bool from_cancel)
4116 { 3340 {
4117 struct worker *worker = NULL; 3341 struct worker *worker = NULL;
4118 struct worker_pool *pool; 3342 struct worker_pool *pool;
4119 struct pool_workqueue *pwq; 3343 struct pool_workqueue *pwq;
4120 struct workqueue_struct *wq; !! 3344
>> 3345 might_sleep();
4121 3346
4122 rcu_read_lock(); 3347 rcu_read_lock();
4123 pool = get_work_pool(work); 3348 pool = get_work_pool(work);
4124 if (!pool) { 3349 if (!pool) {
4125 rcu_read_unlock(); 3350 rcu_read_unlock();
4126 return false; 3351 return false;
4127 } 3352 }
4128 3353
4129 raw_spin_lock_irq(&pool->lock); 3354 raw_spin_lock_irq(&pool->lock);
4130 /* see the comment in try_to_grab_pen 3355 /* see the comment in try_to_grab_pending() with the same code */
4131 pwq = get_work_pwq(work); 3356 pwq = get_work_pwq(work);
4132 if (pwq) { 3357 if (pwq) {
4133 if (unlikely(pwq->pool != poo 3358 if (unlikely(pwq->pool != pool))
4134 goto already_gone; 3359 goto already_gone;
4135 } else { 3360 } else {
4136 worker = find_worker_executin 3361 worker = find_worker_executing_work(pool, work);
4137 if (!worker) 3362 if (!worker)
4138 goto already_gone; 3363 goto already_gone;
4139 pwq = worker->current_pwq; 3364 pwq = worker->current_pwq;
4140 } 3365 }
4141 3366
4142 wq = pwq->wq; !! 3367 check_flush_dependency(pwq->wq, work);
4143 check_flush_dependency(wq, work); <<
4144 3368
4145 insert_wq_barrier(pwq, barr, work, wo 3369 insert_wq_barrier(pwq, barr, work, worker);
4146 raw_spin_unlock_irq(&pool->lock); 3370 raw_spin_unlock_irq(&pool->lock);
4147 3371
4148 touch_work_lockdep_map(work, wq); <<
4149 <<
4150 /* 3372 /*
4151 * Force a lock recursion deadlock wh 3373 * Force a lock recursion deadlock when using flush_work() inside a
4152 * single-threaded or rescuer equippe 3374 * single-threaded or rescuer equipped workqueue.
4153 * 3375 *
4154 * For single threaded workqueues the 3376 * For single threaded workqueues the deadlock happens when the work
4155 * is after the work issuing the flus 3377 * is after the work issuing the flush_work(). For rescuer equipped
4156 * workqueues the deadlock happens wh 3378 * workqueues the deadlock happens when the rescuer stalls, blocking
4157 * forward progress. 3379 * forward progress.
4158 */ 3380 */
4159 if (!from_cancel && (wq->saved_max_ac !! 3381 if (!from_cancel &&
4160 touch_wq_lockdep_map(wq); !! 3382 (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) {
4161 !! 3383 lock_map_acquire(&pwq->wq->lockdep_map);
>> 3384 lock_map_release(&pwq->wq->lockdep_map);
>> 3385 }
4162 rcu_read_unlock(); 3386 rcu_read_unlock();
4163 return true; 3387 return true;
4164 already_gone: 3388 already_gone:
4165 raw_spin_unlock_irq(&pool->lock); 3389 raw_spin_unlock_irq(&pool->lock);
4166 rcu_read_unlock(); 3390 rcu_read_unlock();
4167 return false; 3391 return false;
4168 } 3392 }
4169 3393
4170 static bool __flush_work(struct work_struct * 3394 static bool __flush_work(struct work_struct *work, bool from_cancel)
4171 { 3395 {
4172 struct wq_barrier barr; 3396 struct wq_barrier barr;
4173 3397
4174 if (WARN_ON(!wq_online)) 3398 if (WARN_ON(!wq_online))
4175 return false; 3399 return false;
4176 3400
4177 if (WARN_ON(!work->func)) 3401 if (WARN_ON(!work->func))
4178 return false; 3402 return false;
4179 3403
4180 if (!start_flush_work(work, &barr, fr !! 3404 lock_map_acquire(&work->lockdep_map);
4181 return false; !! 3405 lock_map_release(&work->lockdep_map);
4182 <<
4183 /* <<
4184 * start_flush_work() returned %true. <<
4185 * that @work must have been executin <<
4186 * can't currently be queued. Its dat <<
4187 * was queued on a BH workqueue, we a <<
4188 * BH context and thus can be busy-wa <<
4189 */ <<
4190 if (from_cancel) { <<
4191 unsigned long data = *work_da <<
4192 3406
4193 if (!WARN_ON_ONCE(data & WORK !! 3407 if (start_flush_work(work, &barr, from_cancel)) {
4194 (data & WORK_OFFQ_BH)) { !! 3408 wait_for_completion(&barr.done);
4195 /* !! 3409 destroy_work_on_stack(&barr.work);
4196 * On RT, prevent a l !! 3410 return true;
4197 * soft interrupt pro !! 3411 } else {
4198 * running by keeping !! 3412 return false;
4199 * runs on a differen <<
4200 * than doing the BH <<
4201 * This is copied fro <<
4202 * kernel/softirq.c:: <<
4203 */ <<
4204 while (!try_wait_for_ <<
4205 if (IS_ENABLE <<
4206 local <<
4207 local <<
4208 } else { <<
4209 cpu_r <<
4210 } <<
4211 } <<
4212 goto out_destroy; <<
4213 } <<
4214 } 3413 }
4215 <<
4216 wait_for_completion(&barr.done); <<
4217 <<
4218 out_destroy: <<
4219 destroy_work_on_stack(&barr.work); <<
4220 return true; <<
4221 } 3414 }
4222 3415
4223 /** 3416 /**
4224 * flush_work - wait for a work to finish exe 3417 * flush_work - wait for a work to finish executing the last queueing instance
4225 * @work: the work to flush 3418 * @work: the work to flush
4226 * 3419 *
4227 * Wait until @work has finished execution. 3420 * Wait until @work has finished execution. @work is guaranteed to be idle
4228 * on return if it hasn't been requeued since 3421 * on return if it hasn't been requeued since flush started.
4229 * 3422 *
4230 * Return: 3423 * Return:
4231 * %true if flush_work() waited for the work 3424 * %true if flush_work() waited for the work to finish execution,
4232 * %false if it was already idle. 3425 * %false if it was already idle.
4233 */ 3426 */
4234 bool flush_work(struct work_struct *work) 3427 bool flush_work(struct work_struct *work)
4235 { 3428 {
4236 might_sleep(); <<
4237 return __flush_work(work, false); 3429 return __flush_work(work, false);
4238 } 3430 }
4239 EXPORT_SYMBOL_GPL(flush_work); 3431 EXPORT_SYMBOL_GPL(flush_work);
4240 3432
>> 3433 struct cwt_wait {
>> 3434 wait_queue_entry_t wait;
>> 3435 struct work_struct *work;
>> 3436 };
>> 3437
>> 3438 static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
>> 3439 {
>> 3440 struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
>> 3441
>> 3442 if (cwait->work != key)
>> 3443 return 0;
>> 3444 return autoremove_wake_function(wait, mode, sync, key);
>> 3445 }
>> 3446
>> 3447 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
>> 3448 {
>> 3449 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
>> 3450 unsigned long flags;
>> 3451 int ret;
>> 3452
>> 3453 do {
>> 3454 ret = try_to_grab_pending(work, is_dwork, &flags);
>> 3455 /*
>> 3456 * If someone else is already canceling, wait for it to
>> 3457 * finish. flush_work() doesn't work for PREEMPT_NONE
>> 3458 * because we may get scheduled between @work's completion
>> 3459 * and the other canceling task resuming and clearing
>> 3460 * CANCELING - flush_work() will return false immediately
>> 3461 * as @work is no longer busy, try_to_grab_pending() will
>> 3462 * return -ENOENT as @work is still being canceled and the
>> 3463 * other canceling task won't be able to clear CANCELING as
>> 3464 * we're hogging the CPU.
>> 3465 *
>> 3466 * Let's wait for completion using a waitqueue. As this
>> 3467 * may lead to the thundering herd problem, use a custom
>> 3468 * wake function which matches @work along with exclusive
>> 3469 * wait and wakeup.
>> 3470 */
>> 3471 if (unlikely(ret == -ENOENT)) {
>> 3472 struct cwt_wait cwait;
>> 3473
>> 3474 init_wait(&cwait.wait);
>> 3475 cwait.wait.func = cwt_wakefn;
>> 3476 cwait.work = work;
>> 3477
>> 3478 prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
>> 3479 TASK_UNINTERRUPTIBLE);
>> 3480 if (work_is_canceling(work))
>> 3481 schedule();
>> 3482 finish_wait(&cancel_waitq, &cwait.wait);
>> 3483 }
>> 3484 } while (unlikely(ret < 0));
>> 3485
>> 3486 /* tell other tasks trying to grab @work to back off */
>> 3487 mark_work_canceling(work);
>> 3488 local_irq_restore(flags);
>> 3489
>> 3490 /*
>> 3491 * This allows canceling during early boot. We know that @work
>> 3492 * isn't executing.
>> 3493 */
>> 3494 if (wq_online)
>> 3495 __flush_work(work, true);
>> 3496
>> 3497 clear_work_data(work);
>> 3498
>> 3499 /*
>> 3500 * Paired with prepare_to_wait() above so that either
>> 3501 * waitqueue_active() is visible here or !work_is_canceling() is
>> 3502 * visible there.
>> 3503 */
>> 3504 smp_mb();
>> 3505 if (waitqueue_active(&cancel_waitq))
>> 3506 __wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
>> 3507
>> 3508 return ret;
>> 3509 }
>> 3510
>> 3511 /**
>> 3512 * cancel_work_sync - cancel a work and wait for it to finish
>> 3513 * @work: the work to cancel
>> 3514 *
>> 3515 * Cancel @work and wait for its execution to finish. This function
>> 3516 * can be used even if the work re-queues itself or migrates to
>> 3517 * another workqueue. On return from this function, @work is
>> 3518 * guaranteed to be not pending or executing on any CPU.
>> 3519 *
>> 3520 * cancel_work_sync(&delayed_work->work) must not be used for
>> 3521 * delayed_work's. Use cancel_delayed_work_sync() instead.
>> 3522 *
>> 3523 * The caller must ensure that the workqueue on which @work was last
>> 3524 * queued can't be destroyed before this function returns.
>> 3525 *
>> 3526 * Return:
>> 3527 * %true if @work was pending, %false otherwise.
>> 3528 */
>> 3529 bool cancel_work_sync(struct work_struct *work)
>> 3530 {
>> 3531 return __cancel_work_timer(work, false);
>> 3532 }
>> 3533 EXPORT_SYMBOL_GPL(cancel_work_sync);
>> 3534
4241 /** 3535 /**
4242 * flush_delayed_work - wait for a dwork to f 3536 * flush_delayed_work - wait for a dwork to finish executing the last queueing
4243 * @dwork: the delayed work to flush 3537 * @dwork: the delayed work to flush
4244 * 3538 *
4245 * Delayed timer is cancelled and the pending 3539 * Delayed timer is cancelled and the pending work is queued for
4246 * immediate execution. Like flush_work(), t 3540 * immediate execution. Like flush_work(), this function only
4247 * considers the last queueing instance of @d 3541 * considers the last queueing instance of @dwork.
4248 * 3542 *
4249 * Return: 3543 * Return:
4250 * %true if flush_work() waited for the work 3544 * %true if flush_work() waited for the work to finish execution,
4251 * %false if it was already idle. 3545 * %false if it was already idle.
4252 */ 3546 */
4253 bool flush_delayed_work(struct delayed_work * 3547 bool flush_delayed_work(struct delayed_work *dwork)
4254 { 3548 {
4255 local_irq_disable(); 3549 local_irq_disable();
4256 if (del_timer_sync(&dwork->timer)) 3550 if (del_timer_sync(&dwork->timer))
4257 __queue_work(dwork->cpu, dwor 3551 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
4258 local_irq_enable(); 3552 local_irq_enable();
4259 return flush_work(&dwork->work); 3553 return flush_work(&dwork->work);
4260 } 3554 }
4261 EXPORT_SYMBOL(flush_delayed_work); 3555 EXPORT_SYMBOL(flush_delayed_work);
4262 3556
4263 /** 3557 /**
4264 * flush_rcu_work - wait for a rwork to finis 3558 * flush_rcu_work - wait for a rwork to finish executing the last queueing
4265 * @rwork: the rcu work to flush 3559 * @rwork: the rcu work to flush
4266 * 3560 *
4267 * Return: 3561 * Return:
4268 * %true if flush_rcu_work() waited for the w 3562 * %true if flush_rcu_work() waited for the work to finish execution,
4269 * %false if it was already idle. 3563 * %false if it was already idle.
4270 */ 3564 */
4271 bool flush_rcu_work(struct rcu_work *rwork) 3565 bool flush_rcu_work(struct rcu_work *rwork)
4272 { 3566 {
4273 if (test_bit(WORK_STRUCT_PENDING_BIT, 3567 if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) {
4274 rcu_barrier(); 3568 rcu_barrier();
4275 flush_work(&rwork->work); 3569 flush_work(&rwork->work);
4276 return true; 3570 return true;
4277 } else { 3571 } else {
4278 return flush_work(&rwork->wor 3572 return flush_work(&rwork->work);
4279 } 3573 }
4280 } 3574 }
4281 EXPORT_SYMBOL(flush_rcu_work); 3575 EXPORT_SYMBOL(flush_rcu_work);
4282 3576
4283 static void work_offqd_disable(struct work_of !! 3577 static bool __cancel_work(struct work_struct *work, bool is_dwork)
4284 { <<
4285 const unsigned long max = (1lu << WOR <<
4286 <<
4287 if (likely(offqd->disable < max)) <<
4288 offqd->disable++; <<
4289 else <<
4290 WARN_ONCE(true, "workqueue: w <<
4291 } <<
4292 <<
4293 static void work_offqd_enable(struct work_off <<
4294 { 3578 {
4295 if (likely(offqd->disable > 0)) !! 3579 unsigned long flags;
4296 offqd->disable--; <<
4297 else <<
4298 WARN_ONCE(true, "workqueue: w <<
4299 } <<
4300 <<
4301 static bool __cancel_work(struct work_struct <<
4302 { <<
4303 struct work_offq_data offqd; <<
4304 unsigned long irq_flags; <<
4305 int ret; 3580 int ret;
4306 3581
4307 ret = work_grab_pending(work, cflags, !! 3582 do {
4308 !! 3583 ret = try_to_grab_pending(work, is_dwork, &flags);
4309 work_offqd_unpack(&offqd, *work_data_ !! 3584 } while (unlikely(ret == -EAGAIN));
4310 <<
4311 if (cflags & WORK_CANCEL_DISABLE) <<
4312 work_offqd_disable(&offqd); <<
4313 <<
4314 set_work_pool_and_clear_pending(work, <<
4315 work_ <<
4316 local_irq_restore(irq_flags); <<
4317 return ret; <<
4318 } <<
4319 <<
4320 static bool __cancel_work_sync(struct work_st <<
4321 { <<
4322 bool ret; <<
4323 <<
4324 ret = __cancel_work(work, cflags | WO <<
4325 <<
4326 if (*work_data_bits(work) & WORK_OFFQ <<
4327 WARN_ON_ONCE(in_hardirq()); <<
4328 else <<
4329 might_sleep(); <<
4330 <<
4331 /* <<
4332 * Skip __flush_work() during early b <<
4333 * executing. This allows canceling d <<
4334 */ <<
4335 if (wq_online) <<
4336 __flush_work(work, true); <<
4337 3585
4338 if (!(cflags & WORK_CANCEL_DISABLE)) !! 3586 if (unlikely(ret < 0))
4339 enable_work(work); !! 3587 return false;
4340 3588
>> 3589 set_work_pool_and_clear_pending(work, get_work_pool_id(work));
>> 3590 local_irq_restore(flags);
4341 return ret; 3591 return ret;
4342 } 3592 }
4343 3593
4344 /* 3594 /*
4345 * See cancel_delayed_work() 3595 * See cancel_delayed_work()
4346 */ 3596 */
4347 bool cancel_work(struct work_struct *work) 3597 bool cancel_work(struct work_struct *work)
4348 { 3598 {
4349 return __cancel_work(work, 0); !! 3599 return __cancel_work(work, false);
4350 } 3600 }
4351 EXPORT_SYMBOL(cancel_work); 3601 EXPORT_SYMBOL(cancel_work);
4352 3602
4353 /** 3603 /**
4354 * cancel_work_sync - cancel a work and wait <<
4355 * @work: the work to cancel <<
4356 * <<
4357 * Cancel @work and wait for its execution to <<
4358 * even if the work re-queues itself or migra <<
4359 * from this function, @work is guaranteed to <<
4360 * CPU as long as there aren't racing enqueue <<
4361 * <<
4362 * cancel_work_sync(&delayed_work->work) must <<
4363 * Use cancel_delayed_work_sync() instead. <<
4364 * <<
4365 * Must be called from a sleepable context if <<
4366 * workqueue. Can also be called from non-har <<
4367 * if @work was last queued on a BH workqueue <<
4368 * <<
4369 * Returns %true if @work was pending, %false <<
4370 */ <<
4371 bool cancel_work_sync(struct work_struct *wor <<
4372 { <<
4373 return __cancel_work_sync(work, 0); <<
4374 } <<
4375 EXPORT_SYMBOL_GPL(cancel_work_sync); <<
4376 <<
4377 /** <<
4378 * cancel_delayed_work - cancel a delayed wor 3604 * cancel_delayed_work - cancel a delayed work
4379 * @dwork: delayed_work to cancel 3605 * @dwork: delayed_work to cancel
4380 * 3606 *
4381 * Kill off a pending delayed_work. 3607 * Kill off a pending delayed_work.
4382 * 3608 *
4383 * Return: %true if @dwork was pending and ca 3609 * Return: %true if @dwork was pending and canceled; %false if it wasn't
4384 * pending. 3610 * pending.
4385 * 3611 *
4386 * Note: 3612 * Note:
4387 * The work callback function may still be ru 3613 * The work callback function may still be running on return, unless
4388 * it returns %true and the work doesn't re-a 3614 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
4389 * use cancel_delayed_work_sync() to wait on 3615 * use cancel_delayed_work_sync() to wait on it.
4390 * 3616 *
4391 * This function is safe to call from any con 3617 * This function is safe to call from any context including IRQ handler.
4392 */ 3618 */
4393 bool cancel_delayed_work(struct delayed_work 3619 bool cancel_delayed_work(struct delayed_work *dwork)
4394 { 3620 {
4395 return __cancel_work(&dwork->work, WO !! 3621 return __cancel_work(&dwork->work, true);
4396 } 3622 }
4397 EXPORT_SYMBOL(cancel_delayed_work); 3623 EXPORT_SYMBOL(cancel_delayed_work);
4398 3624
4399 /** 3625 /**
4400 * cancel_delayed_work_sync - cancel a delaye 3626 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
4401 * @dwork: the delayed work cancel 3627 * @dwork: the delayed work cancel
4402 * 3628 *
4403 * This is cancel_work_sync() for delayed wor 3629 * This is cancel_work_sync() for delayed works.
4404 * 3630 *
4405 * Return: 3631 * Return:
4406 * %true if @dwork was pending, %false otherw 3632 * %true if @dwork was pending, %false otherwise.
4407 */ 3633 */
4408 bool cancel_delayed_work_sync(struct delayed_ 3634 bool cancel_delayed_work_sync(struct delayed_work *dwork)
4409 { 3635 {
4410 return __cancel_work_sync(&dwork->wor !! 3636 return __cancel_work_timer(&dwork->work, true);
4411 } 3637 }
4412 EXPORT_SYMBOL(cancel_delayed_work_sync); 3638 EXPORT_SYMBOL(cancel_delayed_work_sync);
4413 3639
4414 /** 3640 /**
4415 * disable_work - Disable and cancel a work i <<
4416 * @work: work item to disable <<
4417 * <<
4418 * Disable @work by incrementing its disable <<
4419 * pending. As long as the disable count is n <<
4420 * will fail and return %false. The maximum s <<
4421 * power of %WORK_OFFQ_DISABLE_BITS, currentl <<
4422 * <<
4423 * Can be called from any context. Returns %t <<
4424 * otherwise. <<
4425 */ <<
4426 bool disable_work(struct work_struct *work) <<
4427 { <<
4428 return __cancel_work(work, WORK_CANCE <<
4429 } <<
4430 EXPORT_SYMBOL_GPL(disable_work); <<
4431 <<
4432 /** <<
4433 * disable_work_sync - Disable, cancel and dr <<
4434 * @work: work item to disable <<
4435 * <<
4436 * Similar to disable_work() but also wait fo <<
4437 * executing. <<
4438 * <<
4439 * Must be called from a sleepable context if <<
4440 * workqueue. Can also be called from non-har <<
4441 * if @work was last queued on a BH workqueue <<
4442 * <<
4443 * Returns %true if @work was pending, %false <<
4444 */ <<
4445 bool disable_work_sync(struct work_struct *wo <<
4446 { <<
4447 return __cancel_work_sync(work, WORK_ <<
4448 } <<
4449 EXPORT_SYMBOL_GPL(disable_work_sync); <<
4450 <<
4451 /** <<
4452 * enable_work - Enable a work item <<
4453 * @work: work item to enable <<
4454 * <<
4455 * Undo disable_work[_sync]() by decrementing <<
4456 * only be queued if its disable count is 0. <<
4457 * <<
4458 * Can be called from any context. Returns %t <<
4459 * Otherwise, %false. <<
4460 */ <<
4461 bool enable_work(struct work_struct *work) <<
4462 { <<
4463 struct work_offq_data offqd; <<
4464 unsigned long irq_flags; <<
4465 <<
4466 work_grab_pending(work, 0, &irq_flags <<
4467 <<
4468 work_offqd_unpack(&offqd, *work_data_ <<
4469 work_offqd_enable(&offqd); <<
4470 set_work_pool_and_clear_pending(work, <<
4471 work_ <<
4472 local_irq_restore(irq_flags); <<
4473 <<
4474 return !offqd.disable; <<
4475 } <<
4476 EXPORT_SYMBOL_GPL(enable_work); <<
4477 <<
4478 /** <<
4479 * disable_delayed_work - Disable and cancel <<
4480 * @dwork: delayed work item to disable <<
4481 * <<
4482 * disable_work() for delayed work items. <<
4483 */ <<
4484 bool disable_delayed_work(struct delayed_work <<
4485 { <<
4486 return __cancel_work(&dwork->work, <<
4487 WORK_CANCEL_DELA <<
4488 } <<
4489 EXPORT_SYMBOL_GPL(disable_delayed_work); <<
4490 <<
4491 /** <<
4492 * disable_delayed_work_sync - Disable, cance <<
4493 * @dwork: delayed work item to disable <<
4494 * <<
4495 * disable_work_sync() for delayed work items <<
4496 */ <<
4497 bool disable_delayed_work_sync(struct delayed <<
4498 { <<
4499 return __cancel_work_sync(&dwork->wor <<
4500 WORK_CANCEL <<
4501 } <<
4502 EXPORT_SYMBOL_GPL(disable_delayed_work_sync); <<
4503 <<
4504 /** <<
4505 * enable_delayed_work - Enable a delayed wor <<
4506 * @dwork: delayed work item to enable <<
4507 * <<
4508 * enable_work() for delayed work items. <<
4509 */ <<
4510 bool enable_delayed_work(struct delayed_work <<
4511 { <<
4512 return enable_work(&dwork->work); <<
4513 } <<
4514 EXPORT_SYMBOL_GPL(enable_delayed_work); <<
4515 <<
4516 /** <<
4517 * schedule_on_each_cpu - execute a function 3641 * schedule_on_each_cpu - execute a function synchronously on each online CPU
4518 * @func: the function to call 3642 * @func: the function to call
4519 * 3643 *
4520 * schedule_on_each_cpu() executes @func on e 3644 * schedule_on_each_cpu() executes @func on each online CPU using the
4521 * system workqueue and blocks until all CPUs 3645 * system workqueue and blocks until all CPUs have completed.
4522 * schedule_on_each_cpu() is very slow. 3646 * schedule_on_each_cpu() is very slow.
4523 * 3647 *
4524 * Return: 3648 * Return:
4525 * 0 on success, -errno on failure. 3649 * 0 on success, -errno on failure.
4526 */ 3650 */
4527 int schedule_on_each_cpu(work_func_t func) 3651 int schedule_on_each_cpu(work_func_t func)
4528 { 3652 {
4529 int cpu; 3653 int cpu;
4530 struct work_struct __percpu *works; 3654 struct work_struct __percpu *works;
4531 3655
4532 works = alloc_percpu(struct work_stru 3656 works = alloc_percpu(struct work_struct);
4533 if (!works) 3657 if (!works)
4534 return -ENOMEM; 3658 return -ENOMEM;
4535 3659
4536 cpus_read_lock(); 3660 cpus_read_lock();
4537 3661
4538 for_each_online_cpu(cpu) { 3662 for_each_online_cpu(cpu) {
4539 struct work_struct *work = pe 3663 struct work_struct *work = per_cpu_ptr(works, cpu);
4540 3664
4541 INIT_WORK(work, func); 3665 INIT_WORK(work, func);
4542 schedule_work_on(cpu, work); 3666 schedule_work_on(cpu, work);
4543 } 3667 }
4544 3668
4545 for_each_online_cpu(cpu) 3669 for_each_online_cpu(cpu)
4546 flush_work(per_cpu_ptr(works, 3670 flush_work(per_cpu_ptr(works, cpu));
4547 3671
4548 cpus_read_unlock(); 3672 cpus_read_unlock();
4549 free_percpu(works); 3673 free_percpu(works);
4550 return 0; 3674 return 0;
4551 } 3675 }
4552 3676
4553 /** 3677 /**
4554 * execute_in_process_context - reliably exec 3678 * execute_in_process_context - reliably execute the routine with user context
4555 * @fn: the function to execute 3679 * @fn: the function to execute
4556 * @ew: guaranteed storage for the ex 3680 * @ew: guaranteed storage for the execute work structure (must
4557 * be available when the work ex 3681 * be available when the work executes)
4558 * 3682 *
4559 * Executes the function immediately if proce 3683 * Executes the function immediately if process context is available,
4560 * otherwise schedules the function for delay 3684 * otherwise schedules the function for delayed execution.
4561 * 3685 *
4562 * Return: 0 - function was executed 3686 * Return: 0 - function was executed
4563 * 1 - function was scheduled fo 3687 * 1 - function was scheduled for execution
4564 */ 3688 */
4565 int execute_in_process_context(work_func_t fn 3689 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
4566 { 3690 {
4567 if (!in_interrupt()) { 3691 if (!in_interrupt()) {
4568 fn(&ew->work); 3692 fn(&ew->work);
4569 return 0; 3693 return 0;
4570 } 3694 }
4571 3695
4572 INIT_WORK(&ew->work, fn); 3696 INIT_WORK(&ew->work, fn);
4573 schedule_work(&ew->work); 3697 schedule_work(&ew->work);
4574 3698
4575 return 1; 3699 return 1;
4576 } 3700 }
4577 EXPORT_SYMBOL_GPL(execute_in_process_context) 3701 EXPORT_SYMBOL_GPL(execute_in_process_context);
4578 3702
4579 /** 3703 /**
4580 * free_workqueue_attrs - free a workqueue_at 3704 * free_workqueue_attrs - free a workqueue_attrs
4581 * @attrs: workqueue_attrs to free 3705 * @attrs: workqueue_attrs to free
4582 * 3706 *
4583 * Undo alloc_workqueue_attrs(). 3707 * Undo alloc_workqueue_attrs().
4584 */ 3708 */
4585 void free_workqueue_attrs(struct workqueue_at 3709 void free_workqueue_attrs(struct workqueue_attrs *attrs)
4586 { 3710 {
4587 if (attrs) { 3711 if (attrs) {
4588 free_cpumask_var(attrs->cpuma 3712 free_cpumask_var(attrs->cpumask);
4589 free_cpumask_var(attrs->__pod 3713 free_cpumask_var(attrs->__pod_cpumask);
4590 kfree(attrs); 3714 kfree(attrs);
4591 } 3715 }
4592 } 3716 }
4593 3717
4594 /** 3718 /**
4595 * alloc_workqueue_attrs - allocate a workque 3719 * alloc_workqueue_attrs - allocate a workqueue_attrs
4596 * 3720 *
4597 * Allocate a new workqueue_attrs, initialize 3721 * Allocate a new workqueue_attrs, initialize with default settings and
4598 * return it. 3722 * return it.
4599 * 3723 *
4600 * Return: The allocated new workqueue_attr o 3724 * Return: The allocated new workqueue_attr on success. %NULL on failure.
4601 */ 3725 */
4602 struct workqueue_attrs *alloc_workqueue_attrs 3726 struct workqueue_attrs *alloc_workqueue_attrs(void)
4603 { 3727 {
4604 struct workqueue_attrs *attrs; 3728 struct workqueue_attrs *attrs;
4605 3729
4606 attrs = kzalloc(sizeof(*attrs), GFP_K 3730 attrs = kzalloc(sizeof(*attrs), GFP_KERNEL);
4607 if (!attrs) 3731 if (!attrs)
4608 goto fail; 3732 goto fail;
4609 if (!alloc_cpumask_var(&attrs->cpumas 3733 if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL))
4610 goto fail; 3734 goto fail;
4611 if (!alloc_cpumask_var(&attrs->__pod_ 3735 if (!alloc_cpumask_var(&attrs->__pod_cpumask, GFP_KERNEL))
4612 goto fail; 3736 goto fail;
4613 3737
4614 cpumask_copy(attrs->cpumask, cpu_poss 3738 cpumask_copy(attrs->cpumask, cpu_possible_mask);
4615 attrs->affn_scope = WQ_AFFN_DFL; 3739 attrs->affn_scope = WQ_AFFN_DFL;
4616 return attrs; 3740 return attrs;
4617 fail: 3741 fail:
4618 free_workqueue_attrs(attrs); 3742 free_workqueue_attrs(attrs);
4619 return NULL; 3743 return NULL;
4620 } 3744 }
4621 3745
4622 static void copy_workqueue_attrs(struct workq 3746 static void copy_workqueue_attrs(struct workqueue_attrs *to,
4623 const struct 3747 const struct workqueue_attrs *from)
4624 { 3748 {
4625 to->nice = from->nice; 3749 to->nice = from->nice;
4626 cpumask_copy(to->cpumask, from->cpuma 3750 cpumask_copy(to->cpumask, from->cpumask);
4627 cpumask_copy(to->__pod_cpumask, from- 3751 cpumask_copy(to->__pod_cpumask, from->__pod_cpumask);
4628 to->affn_strict = from->affn_strict; 3752 to->affn_strict = from->affn_strict;
4629 3753
4630 /* 3754 /*
4631 * Unlike hash and equality test, cop 3755 * Unlike hash and equality test, copying shouldn't ignore wq-only
4632 * fields as copying is used for both 3756 * fields as copying is used for both pool and wq attrs. Instead,
4633 * get_unbound_pool() explicitly clea 3757 * get_unbound_pool() explicitly clears the fields.
4634 */ 3758 */
4635 to->affn_scope = from->affn_scope; 3759 to->affn_scope = from->affn_scope;
4636 to->ordered = from->ordered; 3760 to->ordered = from->ordered;
4637 } 3761 }
4638 3762
4639 /* 3763 /*
4640 * Some attrs fields are workqueue-only. Clea 3764 * Some attrs fields are workqueue-only. Clear them for worker_pool's. See the
4641 * comments in 'struct workqueue_attrs' defin 3765 * comments in 'struct workqueue_attrs' definition.
4642 */ 3766 */
4643 static void wqattrs_clear_for_pool(struct wor 3767 static void wqattrs_clear_for_pool(struct workqueue_attrs *attrs)
4644 { 3768 {
4645 attrs->affn_scope = WQ_AFFN_NR_TYPES; 3769 attrs->affn_scope = WQ_AFFN_NR_TYPES;
4646 attrs->ordered = false; 3770 attrs->ordered = false;
4647 if (attrs->affn_strict) <<
4648 cpumask_copy(attrs->cpumask, <<
4649 } 3771 }
4650 3772
4651 /* hash value of the content of @attr */ 3773 /* hash value of the content of @attr */
4652 static u32 wqattrs_hash(const struct workqueu 3774 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
4653 { 3775 {
4654 u32 hash = 0; 3776 u32 hash = 0;
4655 3777
4656 hash = jhash_1word(attrs->nice, hash) 3778 hash = jhash_1word(attrs->nice, hash);
4657 hash = jhash_1word(attrs->affn_strict !! 3779 hash = jhash(cpumask_bits(attrs->cpumask),
>> 3780 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
4658 hash = jhash(cpumask_bits(attrs->__po 3781 hash = jhash(cpumask_bits(attrs->__pod_cpumask),
4659 BITS_TO_LONGS(nr_cpumask 3782 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
4660 if (!attrs->affn_strict) !! 3783 hash = jhash_1word(attrs->affn_strict, hash);
4661 hash = jhash(cpumask_bits(att <<
4662 BITS_TO_LONGS(nr <<
4663 return hash; 3784 return hash;
4664 } 3785 }
4665 3786
4666 /* content equality test */ 3787 /* content equality test */
4667 static bool wqattrs_equal(const struct workqu 3788 static bool wqattrs_equal(const struct workqueue_attrs *a,
4668 const struct workqu 3789 const struct workqueue_attrs *b)
4669 { 3790 {
4670 if (a->nice != b->nice) 3791 if (a->nice != b->nice)
4671 return false; 3792 return false;
4672 if (a->affn_strict != b->affn_strict) !! 3793 if (!cpumask_equal(a->cpumask, b->cpumask))
4673 return false; 3794 return false;
4674 if (!cpumask_equal(a->__pod_cpumask, 3795 if (!cpumask_equal(a->__pod_cpumask, b->__pod_cpumask))
4675 return false; 3796 return false;
4676 if (!a->affn_strict && !cpumask_equal !! 3797 if (a->affn_strict != b->affn_strict)
4677 return false; 3798 return false;
4678 return true; 3799 return true;
4679 } 3800 }
4680 3801
4681 /* Update @attrs with actually available CPUs 3802 /* Update @attrs with actually available CPUs */
4682 static void wqattrs_actualize_cpumask(struct 3803 static void wqattrs_actualize_cpumask(struct workqueue_attrs *attrs,
4683 const c 3804 const cpumask_t *unbound_cpumask)
4684 { 3805 {
4685 /* 3806 /*
4686 * Calculate the effective CPU mask o 3807 * Calculate the effective CPU mask of @attrs given @unbound_cpumask. If
4687 * @attrs->cpumask doesn't overlap wi 3808 * @attrs->cpumask doesn't overlap with @unbound_cpumask, we fallback to
4688 * @unbound_cpumask. 3809 * @unbound_cpumask.
4689 */ 3810 */
4690 cpumask_and(attrs->cpumask, attrs->cp 3811 cpumask_and(attrs->cpumask, attrs->cpumask, unbound_cpumask);
4691 if (unlikely(cpumask_empty(attrs->cpu 3812 if (unlikely(cpumask_empty(attrs->cpumask)))
4692 cpumask_copy(attrs->cpumask, 3813 cpumask_copy(attrs->cpumask, unbound_cpumask);
4693 } 3814 }
4694 3815
4695 /* find wq_pod_type to use for @attrs */ 3816 /* find wq_pod_type to use for @attrs */
4696 static const struct wq_pod_type * 3817 static const struct wq_pod_type *
4697 wqattrs_pod_type(const struct workqueue_attrs 3818 wqattrs_pod_type(const struct workqueue_attrs *attrs)
4698 { 3819 {
4699 enum wq_affn_scope scope; 3820 enum wq_affn_scope scope;
4700 struct wq_pod_type *pt; 3821 struct wq_pod_type *pt;
4701 3822
4702 /* to synchronize access to wq_affn_d 3823 /* to synchronize access to wq_affn_dfl */
4703 lockdep_assert_held(&wq_pool_mutex); 3824 lockdep_assert_held(&wq_pool_mutex);
4704 3825
4705 if (attrs->affn_scope == WQ_AFFN_DFL) 3826 if (attrs->affn_scope == WQ_AFFN_DFL)
4706 scope = wq_affn_dfl; 3827 scope = wq_affn_dfl;
4707 else 3828 else
4708 scope = attrs->affn_scope; 3829 scope = attrs->affn_scope;
4709 3830
4710 pt = &wq_pod_types[scope]; 3831 pt = &wq_pod_types[scope];
4711 3832
4712 if (!WARN_ON_ONCE(attrs->affn_scope = 3833 if (!WARN_ON_ONCE(attrs->affn_scope == WQ_AFFN_NR_TYPES) &&
4713 likely(pt->nr_pods)) 3834 likely(pt->nr_pods))
4714 return pt; 3835 return pt;
4715 3836
4716 /* 3837 /*
4717 * Before workqueue_init_topology(), 3838 * Before workqueue_init_topology(), only SYSTEM is available which is
4718 * initialized in workqueue_init_earl 3839 * initialized in workqueue_init_early().
4719 */ 3840 */
4720 pt = &wq_pod_types[WQ_AFFN_SYSTEM]; 3841 pt = &wq_pod_types[WQ_AFFN_SYSTEM];
4721 BUG_ON(!pt->nr_pods); 3842 BUG_ON(!pt->nr_pods);
4722 return pt; 3843 return pt;
4723 } 3844 }
4724 3845
4725 /** 3846 /**
4726 * init_worker_pool - initialize a newly zall 3847 * init_worker_pool - initialize a newly zalloc'd worker_pool
4727 * @pool: worker_pool to initialize 3848 * @pool: worker_pool to initialize
4728 * 3849 *
4729 * Initialize a newly zalloc'd @pool. It als 3850 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
4730 * 3851 *
4731 * Return: 0 on success, -errno on failure. 3852 * Return: 0 on success, -errno on failure. Even on failure, all fields
4732 * inside @pool proper are initialized and pu 3853 * inside @pool proper are initialized and put_unbound_pool() can be called
4733 * on @pool safely to release it. 3854 * on @pool safely to release it.
4734 */ 3855 */
4735 static int init_worker_pool(struct worker_poo 3856 static int init_worker_pool(struct worker_pool *pool)
4736 { 3857 {
4737 raw_spin_lock_init(&pool->lock); 3858 raw_spin_lock_init(&pool->lock);
4738 pool->id = -1; 3859 pool->id = -1;
4739 pool->cpu = -1; 3860 pool->cpu = -1;
4740 pool->node = NUMA_NO_NODE; 3861 pool->node = NUMA_NO_NODE;
4741 pool->flags |= POOL_DISASSOCIATED; 3862 pool->flags |= POOL_DISASSOCIATED;
4742 pool->watchdog_ts = jiffies; 3863 pool->watchdog_ts = jiffies;
4743 INIT_LIST_HEAD(&pool->worklist); 3864 INIT_LIST_HEAD(&pool->worklist);
4744 INIT_LIST_HEAD(&pool->idle_list); 3865 INIT_LIST_HEAD(&pool->idle_list);
4745 hash_init(pool->busy_hash); 3866 hash_init(pool->busy_hash);
4746 3867
4747 timer_setup(&pool->idle_timer, idle_w 3868 timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE);
4748 INIT_WORK(&pool->idle_cull_work, idle 3869 INIT_WORK(&pool->idle_cull_work, idle_cull_fn);
4749 3870
4750 timer_setup(&pool->mayday_timer, pool 3871 timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0);
4751 3872
4752 INIT_LIST_HEAD(&pool->workers); 3873 INIT_LIST_HEAD(&pool->workers);
>> 3874 INIT_LIST_HEAD(&pool->dying_workers);
4753 3875
4754 ida_init(&pool->worker_ida); 3876 ida_init(&pool->worker_ida);
4755 INIT_HLIST_NODE(&pool->hash_node); 3877 INIT_HLIST_NODE(&pool->hash_node);
4756 pool->refcnt = 1; 3878 pool->refcnt = 1;
4757 3879
4758 /* shouldn't fail above this point */ 3880 /* shouldn't fail above this point */
4759 pool->attrs = alloc_workqueue_attrs() 3881 pool->attrs = alloc_workqueue_attrs();
4760 if (!pool->attrs) 3882 if (!pool->attrs)
4761 return -ENOMEM; 3883 return -ENOMEM;
4762 3884
4763 wqattrs_clear_for_pool(pool->attrs); 3885 wqattrs_clear_for_pool(pool->attrs);
4764 3886
4765 return 0; 3887 return 0;
4766 } 3888 }
4767 3889
4768 #ifdef CONFIG_LOCKDEP 3890 #ifdef CONFIG_LOCKDEP
4769 static void wq_init_lockdep(struct workqueue_ 3891 static void wq_init_lockdep(struct workqueue_struct *wq)
4770 { 3892 {
4771 char *lock_name; 3893 char *lock_name;
4772 3894
4773 lockdep_register_key(&wq->key); 3895 lockdep_register_key(&wq->key);
4774 lock_name = kasprintf(GFP_KERNEL, "%s 3896 lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name);
4775 if (!lock_name) 3897 if (!lock_name)
4776 lock_name = wq->name; 3898 lock_name = wq->name;
4777 3899
4778 wq->lock_name = lock_name; 3900 wq->lock_name = lock_name;
4779 lockdep_init_map(&wq->lockdep_map, lo 3901 lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0);
4780 } 3902 }
4781 3903
4782 static void wq_unregister_lockdep(struct work 3904 static void wq_unregister_lockdep(struct workqueue_struct *wq)
4783 { 3905 {
4784 lockdep_unregister_key(&wq->key); 3906 lockdep_unregister_key(&wq->key);
4785 } 3907 }
4786 3908
4787 static void wq_free_lockdep(struct workqueue_ 3909 static void wq_free_lockdep(struct workqueue_struct *wq)
4788 { 3910 {
4789 if (wq->lock_name != wq->name) 3911 if (wq->lock_name != wq->name)
4790 kfree(wq->lock_name); 3912 kfree(wq->lock_name);
4791 } 3913 }
4792 #else 3914 #else
4793 static void wq_init_lockdep(struct workqueue_ 3915 static void wq_init_lockdep(struct workqueue_struct *wq)
4794 { 3916 {
4795 } 3917 }
4796 3918
4797 static void wq_unregister_lockdep(struct work 3919 static void wq_unregister_lockdep(struct workqueue_struct *wq)
4798 { 3920 {
4799 } 3921 }
4800 3922
4801 static void wq_free_lockdep(struct workqueue_ 3923 static void wq_free_lockdep(struct workqueue_struct *wq)
4802 { 3924 {
4803 } 3925 }
4804 #endif 3926 #endif
4805 3927
4806 static void free_node_nr_active(struct wq_nod <<
4807 { <<
4808 int node; <<
4809 <<
4810 for_each_node(node) { <<
4811 kfree(nna_ar[node]); <<
4812 nna_ar[node] = NULL; <<
4813 } <<
4814 <<
4815 kfree(nna_ar[nr_node_ids]); <<
4816 nna_ar[nr_node_ids] = NULL; <<
4817 } <<
4818 <<
4819 static void init_node_nr_active(struct wq_nod <<
4820 { <<
4821 nna->max = WQ_DFL_MIN_ACTIVE; <<
4822 atomic_set(&nna->nr, 0); <<
4823 raw_spin_lock_init(&nna->lock); <<
4824 INIT_LIST_HEAD(&nna->pending_pwqs); <<
4825 } <<
4826 <<
4827 /* <<
4828 * Each node's nr_active counter will be acce <<
4829 * should be allocated in the node. <<
4830 */ <<
4831 static int alloc_node_nr_active(struct wq_nod <<
4832 { <<
4833 struct wq_node_nr_active *nna; <<
4834 int node; <<
4835 <<
4836 for_each_node(node) { <<
4837 nna = kzalloc_node(sizeof(*nn <<
4838 if (!nna) <<
4839 goto err_free; <<
4840 init_node_nr_active(nna); <<
4841 nna_ar[node] = nna; <<
4842 } <<
4843 <<
4844 /* [nr_node_ids] is used as the fallb <<
4845 nna = kzalloc_node(sizeof(*nna), GFP_ <<
4846 if (!nna) <<
4847 goto err_free; <<
4848 init_node_nr_active(nna); <<
4849 nna_ar[nr_node_ids] = nna; <<
4850 <<
4851 return 0; <<
4852 <<
4853 err_free: <<
4854 free_node_nr_active(nna_ar); <<
4855 return -ENOMEM; <<
4856 } <<
4857 <<
4858 static void rcu_free_wq(struct rcu_head *rcu) 3928 static void rcu_free_wq(struct rcu_head *rcu)
4859 { 3929 {
4860 struct workqueue_struct *wq = 3930 struct workqueue_struct *wq =
4861 container_of(rcu, struct work 3931 container_of(rcu, struct workqueue_struct, rcu);
4862 3932
4863 if (wq->flags & WQ_UNBOUND) <<
4864 free_node_nr_active(wq->node_ <<
4865 <<
4866 wq_free_lockdep(wq); 3933 wq_free_lockdep(wq);
4867 free_percpu(wq->cpu_pwq); 3934 free_percpu(wq->cpu_pwq);
4868 free_workqueue_attrs(wq->unbound_attr 3935 free_workqueue_attrs(wq->unbound_attrs);
4869 kfree(wq); 3936 kfree(wq);
4870 } 3937 }
4871 3938
4872 static void rcu_free_pool(struct rcu_head *rc 3939 static void rcu_free_pool(struct rcu_head *rcu)
4873 { 3940 {
4874 struct worker_pool *pool = container_ 3941 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
4875 3942
4876 ida_destroy(&pool->worker_ida); 3943 ida_destroy(&pool->worker_ida);
4877 free_workqueue_attrs(pool->attrs); 3944 free_workqueue_attrs(pool->attrs);
4878 kfree(pool); 3945 kfree(pool);
4879 } 3946 }
4880 3947
4881 /** 3948 /**
4882 * put_unbound_pool - put a worker_pool 3949 * put_unbound_pool - put a worker_pool
4883 * @pool: worker_pool to put 3950 * @pool: worker_pool to put
4884 * 3951 *
4885 * Put @pool. If its refcnt reaches zero, it 3952 * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU
4886 * safe manner. get_unbound_pool() calls thi 3953 * safe manner. get_unbound_pool() calls this function on its failure path
4887 * and this function should be able to releas 3954 * and this function should be able to release pools which went through,
4888 * successfully or not, init_worker_pool(). 3955 * successfully or not, init_worker_pool().
4889 * 3956 *
4890 * Should be called with wq_pool_mutex held. 3957 * Should be called with wq_pool_mutex held.
4891 */ 3958 */
4892 static void put_unbound_pool(struct worker_po 3959 static void put_unbound_pool(struct worker_pool *pool)
4893 { 3960 {
>> 3961 DECLARE_COMPLETION_ONSTACK(detach_completion);
4894 struct worker *worker; 3962 struct worker *worker;
4895 LIST_HEAD(cull_list); 3963 LIST_HEAD(cull_list);
4896 3964
4897 lockdep_assert_held(&wq_pool_mutex); 3965 lockdep_assert_held(&wq_pool_mutex);
4898 3966
4899 if (--pool->refcnt) 3967 if (--pool->refcnt)
4900 return; 3968 return;
4901 3969
4902 /* sanity checks */ 3970 /* sanity checks */
4903 if (WARN_ON(!(pool->cpu < 0)) || 3971 if (WARN_ON(!(pool->cpu < 0)) ||
4904 WARN_ON(!list_empty(&pool->workli 3972 WARN_ON(!list_empty(&pool->worklist)))
4905 return; 3973 return;
4906 3974
4907 /* release id and unhash */ 3975 /* release id and unhash */
4908 if (pool->id >= 0) 3976 if (pool->id >= 0)
4909 idr_remove(&worker_pool_idr, 3977 idr_remove(&worker_pool_idr, pool->id);
4910 hash_del(&pool->hash_node); 3978 hash_del(&pool->hash_node);
4911 3979
4912 /* 3980 /*
4913 * Become the manager and destroy all 3981 * Become the manager and destroy all workers. This prevents
4914 * @pool's workers from blocking on a 3982 * @pool's workers from blocking on attach_mutex. We're the last
4915 * manager and @pool gets freed with 3983 * manager and @pool gets freed with the flag set.
4916 * 3984 *
4917 * Having a concurrent manager is qui 3985 * Having a concurrent manager is quite unlikely to happen as we can
4918 * only get here with 3986 * only get here with
4919 * pwq->refcnt == pool->refcnt == 0 3987 * pwq->refcnt == pool->refcnt == 0
4920 * which implies no work queued to th 3988 * which implies no work queued to the pool, which implies no worker can
4921 * become the manager. However a work 3989 * become the manager. However a worker could have taken the role of
4922 * manager before the refcnts dropped 3990 * manager before the refcnts dropped to 0, since maybe_create_worker()
4923 * drops pool->lock 3991 * drops pool->lock
4924 */ 3992 */
4925 while (true) { 3993 while (true) {
4926 rcuwait_wait_event(&manager_w 3994 rcuwait_wait_event(&manager_wait,
4927 !(pool->fl 3995 !(pool->flags & POOL_MANAGER_ACTIVE),
4928 TASK_UNINT 3996 TASK_UNINTERRUPTIBLE);
4929 3997
4930 mutex_lock(&wq_pool_attach_mu 3998 mutex_lock(&wq_pool_attach_mutex);
4931 raw_spin_lock_irq(&pool->lock 3999 raw_spin_lock_irq(&pool->lock);
4932 if (!(pool->flags & POOL_MANA 4000 if (!(pool->flags & POOL_MANAGER_ACTIVE)) {
4933 pool->flags |= POOL_M 4001 pool->flags |= POOL_MANAGER_ACTIVE;
4934 break; 4002 break;
4935 } 4003 }
4936 raw_spin_unlock_irq(&pool->lo 4004 raw_spin_unlock_irq(&pool->lock);
4937 mutex_unlock(&wq_pool_attach_ 4005 mutex_unlock(&wq_pool_attach_mutex);
4938 } 4006 }
4939 4007
4940 while ((worker = first_idle_worker(po 4008 while ((worker = first_idle_worker(pool)))
4941 set_worker_dying(worker, &cul 4009 set_worker_dying(worker, &cull_list);
4942 WARN_ON(pool->nr_workers || pool->nr_ 4010 WARN_ON(pool->nr_workers || pool->nr_idle);
4943 raw_spin_unlock_irq(&pool->lock); 4011 raw_spin_unlock_irq(&pool->lock);
4944 4012
4945 detach_dying_workers(&cull_list); !! 4013 wake_dying_workers(&cull_list);
4946 4014
>> 4015 if (!list_empty(&pool->workers) || !list_empty(&pool->dying_workers))
>> 4016 pool->detach_completion = &detach_completion;
4947 mutex_unlock(&wq_pool_attach_mutex); 4017 mutex_unlock(&wq_pool_attach_mutex);
4948 4018
4949 reap_dying_workers(&cull_list); !! 4019 if (pool->detach_completion)
>> 4020 wait_for_completion(pool->detach_completion);
4950 4021
4951 /* shut down the timers */ 4022 /* shut down the timers */
4952 del_timer_sync(&pool->idle_timer); 4023 del_timer_sync(&pool->idle_timer);
4953 cancel_work_sync(&pool->idle_cull_wor 4024 cancel_work_sync(&pool->idle_cull_work);
4954 del_timer_sync(&pool->mayday_timer); 4025 del_timer_sync(&pool->mayday_timer);
4955 4026
4956 /* RCU protected to allow dereference 4027 /* RCU protected to allow dereferences from get_work_pool() */
4957 call_rcu(&pool->rcu, rcu_free_pool); 4028 call_rcu(&pool->rcu, rcu_free_pool);
4958 } 4029 }
4959 4030
4960 /** 4031 /**
4961 * get_unbound_pool - get a worker_pool with 4032 * get_unbound_pool - get a worker_pool with the specified attributes
4962 * @attrs: the attributes of the worker_pool 4033 * @attrs: the attributes of the worker_pool to get
4963 * 4034 *
4964 * Obtain a worker_pool which has the same at 4035 * Obtain a worker_pool which has the same attributes as @attrs, bump the
4965 * reference count and return it. If there a 4036 * reference count and return it. If there already is a matching
4966 * worker_pool, it will be used; otherwise, t 4037 * worker_pool, it will be used; otherwise, this function attempts to
4967 * create a new one. 4038 * create a new one.
4968 * 4039 *
4969 * Should be called with wq_pool_mutex held. 4040 * Should be called with wq_pool_mutex held.
4970 * 4041 *
4971 * Return: On success, a worker_pool with the 4042 * Return: On success, a worker_pool with the same attributes as @attrs.
4972 * On failure, %NULL. 4043 * On failure, %NULL.
4973 */ 4044 */
4974 static struct worker_pool *get_unbound_pool(c 4045 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
4975 { 4046 {
4976 struct wq_pod_type *pt = &wq_pod_type 4047 struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_NUMA];
4977 u32 hash = wqattrs_hash(attrs); 4048 u32 hash = wqattrs_hash(attrs);
4978 struct worker_pool *pool; 4049 struct worker_pool *pool;
4979 int pod, node = NUMA_NO_NODE; 4050 int pod, node = NUMA_NO_NODE;
4980 4051
4981 lockdep_assert_held(&wq_pool_mutex); 4052 lockdep_assert_held(&wq_pool_mutex);
4982 4053
4983 /* do we already have a matching pool 4054 /* do we already have a matching pool? */
4984 hash_for_each_possible(unbound_pool_h 4055 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
4985 if (wqattrs_equal(pool->attrs 4056 if (wqattrs_equal(pool->attrs, attrs)) {
4986 pool->refcnt++; 4057 pool->refcnt++;
4987 return pool; 4058 return pool;
4988 } 4059 }
4989 } 4060 }
4990 4061
4991 /* If __pod_cpumask is contained insi 4062 /* If __pod_cpumask is contained inside a NUMA pod, that's our node */
4992 for (pod = 0; pod < pt->nr_pods; pod+ 4063 for (pod = 0; pod < pt->nr_pods; pod++) {
4993 if (cpumask_subset(attrs->__p 4064 if (cpumask_subset(attrs->__pod_cpumask, pt->pod_cpus[pod])) {
4994 node = pt->pod_node[p 4065 node = pt->pod_node[pod];
4995 break; 4066 break;
4996 } 4067 }
4997 } 4068 }
4998 4069
4999 /* nope, create a new one */ 4070 /* nope, create a new one */
5000 pool = kzalloc_node(sizeof(*pool), GF 4071 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, node);
5001 if (!pool || init_worker_pool(pool) < 4072 if (!pool || init_worker_pool(pool) < 0)
5002 goto fail; 4073 goto fail;
5003 4074
5004 pool->node = node; 4075 pool->node = node;
5005 copy_workqueue_attrs(pool->attrs, att 4076 copy_workqueue_attrs(pool->attrs, attrs);
5006 wqattrs_clear_for_pool(pool->attrs); 4077 wqattrs_clear_for_pool(pool->attrs);
5007 4078
5008 if (worker_pool_assign_id(pool) < 0) 4079 if (worker_pool_assign_id(pool) < 0)
5009 goto fail; 4080 goto fail;
5010 4081
5011 /* create and start the initial worke 4082 /* create and start the initial worker */
5012 if (wq_online && !create_worker(pool) 4083 if (wq_online && !create_worker(pool))
5013 goto fail; 4084 goto fail;
5014 4085
5015 /* install */ 4086 /* install */
5016 hash_add(unbound_pool_hash, &pool->ha 4087 hash_add(unbound_pool_hash, &pool->hash_node, hash);
5017 4088
5018 return pool; 4089 return pool;
5019 fail: 4090 fail:
5020 if (pool) 4091 if (pool)
5021 put_unbound_pool(pool); 4092 put_unbound_pool(pool);
5022 return NULL; 4093 return NULL;
5023 } 4094 }
5024 4095
>> 4096 static void rcu_free_pwq(struct rcu_head *rcu)
>> 4097 {
>> 4098 kmem_cache_free(pwq_cache,
>> 4099 container_of(rcu, struct pool_workqueue, rcu));
>> 4100 }
>> 4101
5025 /* 4102 /*
5026 * Scheduled on pwq_release_worker by put_pwq 4103 * Scheduled on pwq_release_worker by put_pwq() when an unbound pwq hits zero
5027 * refcnt and needs to be destroyed. 4104 * refcnt and needs to be destroyed.
5028 */ 4105 */
5029 static void pwq_release_workfn(struct kthread 4106 static void pwq_release_workfn(struct kthread_work *work)
5030 { 4107 {
5031 struct pool_workqueue *pwq = containe 4108 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
5032 4109 release_work);
5033 struct workqueue_struct *wq = pwq->wq 4110 struct workqueue_struct *wq = pwq->wq;
5034 struct worker_pool *pool = pwq->pool; 4111 struct worker_pool *pool = pwq->pool;
5035 bool is_last = false; 4112 bool is_last = false;
5036 4113
5037 /* 4114 /*
5038 * When @pwq is not linked, it doesn' 4115 * When @pwq is not linked, it doesn't hold any reference to the
5039 * @wq, and @wq is invalid to access. 4116 * @wq, and @wq is invalid to access.
5040 */ 4117 */
5041 if (!list_empty(&pwq->pwqs_node)) { 4118 if (!list_empty(&pwq->pwqs_node)) {
5042 mutex_lock(&wq->mutex); 4119 mutex_lock(&wq->mutex);
5043 list_del_rcu(&pwq->pwqs_node) 4120 list_del_rcu(&pwq->pwqs_node);
5044 is_last = list_empty(&wq->pwq 4121 is_last = list_empty(&wq->pwqs);
5045 <<
5046 /* <<
5047 * For ordered workqueue with <<
5048 */ <<
5049 if (!is_last && (wq->flags & <<
5050 unplug_oldest_pwq(wq) <<
5051 <<
5052 mutex_unlock(&wq->mutex); 4122 mutex_unlock(&wq->mutex);
5053 } 4123 }
5054 4124
5055 if (wq->flags & WQ_UNBOUND) { 4125 if (wq->flags & WQ_UNBOUND) {
5056 mutex_lock(&wq_pool_mutex); 4126 mutex_lock(&wq_pool_mutex);
5057 put_unbound_pool(pool); 4127 put_unbound_pool(pool);
5058 mutex_unlock(&wq_pool_mutex); 4128 mutex_unlock(&wq_pool_mutex);
5059 } 4129 }
5060 4130
5061 if (!list_empty(&pwq->pending_node)) !! 4131 call_rcu(&pwq->rcu, rcu_free_pwq);
5062 struct wq_node_nr_active *nna <<
5063 wq_node_nr_active(pwq <<
5064 <<
5065 raw_spin_lock_irq(&nna->lock) <<
5066 list_del_init(&pwq->pending_n <<
5067 raw_spin_unlock_irq(&nna->loc <<
5068 } <<
5069 <<
5070 kfree_rcu(pwq, rcu); <<
5071 4132
5072 /* 4133 /*
5073 * If we're the last pwq going away, 4134 * If we're the last pwq going away, @wq is already dead and no one
5074 * is gonna access it anymore. Sched 4135 * is gonna access it anymore. Schedule RCU free.
5075 */ 4136 */
5076 if (is_last) { 4137 if (is_last) {
5077 wq_unregister_lockdep(wq); 4138 wq_unregister_lockdep(wq);
5078 call_rcu(&wq->rcu, rcu_free_w 4139 call_rcu(&wq->rcu, rcu_free_wq);
5079 } 4140 }
5080 } 4141 }
5081 4142
>> 4143 /**
>> 4144 * pwq_adjust_max_active - update a pwq's max_active to the current setting
>> 4145 * @pwq: target pool_workqueue
>> 4146 *
>> 4147 * If @pwq isn't freezing, set @pwq->max_active to the associated
>> 4148 * workqueue's saved_max_active and activate inactive work items
>> 4149 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
>> 4150 */
>> 4151 static void pwq_adjust_max_active(struct pool_workqueue *pwq)
>> 4152 {
>> 4153 struct workqueue_struct *wq = pwq->wq;
>> 4154 bool freezable = wq->flags & WQ_FREEZABLE;
>> 4155 unsigned long flags;
>> 4156
>> 4157 /* for @wq->saved_max_active */
>> 4158 lockdep_assert_held(&wq->mutex);
>> 4159
>> 4160 /* fast exit for non-freezable wqs */
>> 4161 if (!freezable && pwq->max_active == wq->saved_max_active)
>> 4162 return;
>> 4163
>> 4164 /* this function can be called during early boot w/ irq disabled */
>> 4165 raw_spin_lock_irqsave(&pwq->pool->lock, flags);
>> 4166
>> 4167 /*
>> 4168 * During [un]freezing, the caller is responsible for ensuring that
>> 4169 * this function is called at least once after @workqueue_freezing
>> 4170 * is updated and visible.
>> 4171 */
>> 4172 if (!freezable || !workqueue_freezing) {
>> 4173 pwq->max_active = wq->saved_max_active;
>> 4174
>> 4175 while (!list_empty(&pwq->inactive_works) &&
>> 4176 pwq->nr_active < pwq->max_active)
>> 4177 pwq_activate_first_inactive(pwq);
>> 4178
>> 4179 kick_pool(pwq->pool);
>> 4180 } else {
>> 4181 pwq->max_active = 0;
>> 4182 }
>> 4183
>> 4184 raw_spin_unlock_irqrestore(&pwq->pool->lock, flags);
>> 4185 }
>> 4186
5082 /* initialize newly allocated @pwq which is a 4187 /* initialize newly allocated @pwq which is associated with @wq and @pool */
5083 static void init_pwq(struct pool_workqueue *p 4188 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
5084 struct worker_pool *pool 4189 struct worker_pool *pool)
5085 { 4190 {
5086 BUG_ON((unsigned long)pwq & ~WORK_STR !! 4191 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
5087 4192
5088 memset(pwq, 0, sizeof(*pwq)); 4193 memset(pwq, 0, sizeof(*pwq));
5089 4194
5090 pwq->pool = pool; 4195 pwq->pool = pool;
5091 pwq->wq = wq; 4196 pwq->wq = wq;
5092 pwq->flush_color = -1; 4197 pwq->flush_color = -1;
5093 pwq->refcnt = 1; 4198 pwq->refcnt = 1;
5094 INIT_LIST_HEAD(&pwq->inactive_works); 4199 INIT_LIST_HEAD(&pwq->inactive_works);
5095 INIT_LIST_HEAD(&pwq->pending_node); <<
5096 INIT_LIST_HEAD(&pwq->pwqs_node); 4200 INIT_LIST_HEAD(&pwq->pwqs_node);
5097 INIT_LIST_HEAD(&pwq->mayday_node); 4201 INIT_LIST_HEAD(&pwq->mayday_node);
5098 kthread_init_work(&pwq->release_work, 4202 kthread_init_work(&pwq->release_work, pwq_release_workfn);
5099 } 4203 }
5100 4204
5101 /* sync @pwq with the current state of its as 4205 /* sync @pwq with the current state of its associated wq and link it */
5102 static void link_pwq(struct pool_workqueue *p 4206 static void link_pwq(struct pool_workqueue *pwq)
5103 { 4207 {
5104 struct workqueue_struct *wq = pwq->wq 4208 struct workqueue_struct *wq = pwq->wq;
5105 4209
5106 lockdep_assert_held(&wq->mutex); 4210 lockdep_assert_held(&wq->mutex);
5107 4211
5108 /* may be called multiple times, igno 4212 /* may be called multiple times, ignore if already linked */
5109 if (!list_empty(&pwq->pwqs_node)) 4213 if (!list_empty(&pwq->pwqs_node))
5110 return; 4214 return;
5111 4215
5112 /* set the matching work_color */ 4216 /* set the matching work_color */
5113 pwq->work_color = wq->work_color; 4217 pwq->work_color = wq->work_color;
5114 4218
>> 4219 /* sync max_active to the current setting */
>> 4220 pwq_adjust_max_active(pwq);
>> 4221
5115 /* link in @pwq */ 4222 /* link in @pwq */
5116 list_add_tail_rcu(&pwq->pwqs_node, &w !! 4223 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
5117 } 4224 }
5118 4225
5119 /* obtain a pool matching @attr and create a 4226 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
5120 static struct pool_workqueue *alloc_unbound_p 4227 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
5121 const 4228 const struct workqueue_attrs *attrs)
5122 { 4229 {
5123 struct worker_pool *pool; 4230 struct worker_pool *pool;
5124 struct pool_workqueue *pwq; 4231 struct pool_workqueue *pwq;
5125 4232
5126 lockdep_assert_held(&wq_pool_mutex); 4233 lockdep_assert_held(&wq_pool_mutex);
5127 4234
5128 pool = get_unbound_pool(attrs); 4235 pool = get_unbound_pool(attrs);
5129 if (!pool) 4236 if (!pool)
5130 return NULL; 4237 return NULL;
5131 4238
5132 pwq = kmem_cache_alloc_node(pwq_cache 4239 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
5133 if (!pwq) { 4240 if (!pwq) {
5134 put_unbound_pool(pool); 4241 put_unbound_pool(pool);
5135 return NULL; 4242 return NULL;
5136 } 4243 }
5137 4244
5138 init_pwq(pwq, wq, pool); 4245 init_pwq(pwq, wq, pool);
5139 return pwq; 4246 return pwq;
5140 } 4247 }
5141 4248
5142 static void apply_wqattrs_lock(void) <<
5143 { <<
5144 mutex_lock(&wq_pool_mutex); <<
5145 } <<
5146 <<
5147 static void apply_wqattrs_unlock(void) <<
5148 { <<
5149 mutex_unlock(&wq_pool_mutex); <<
5150 } <<
5151 <<
5152 /** 4249 /**
5153 * wq_calc_pod_cpumask - calculate a wq_attrs 4250 * wq_calc_pod_cpumask - calculate a wq_attrs' cpumask for a pod
5154 * @attrs: the wq_attrs of the default pwq of 4251 * @attrs: the wq_attrs of the default pwq of the target workqueue
5155 * @cpu: the target CPU 4252 * @cpu: the target CPU
>> 4253 * @cpu_going_down: if >= 0, the CPU to consider as offline
5156 * 4254 *
5157 * Calculate the cpumask a workqueue with @at !! 4255 * Calculate the cpumask a workqueue with @attrs should use on @pod. If
>> 4256 * @cpu_going_down is >= 0, that cpu is considered offline during calculation.
5158 * The result is stored in @attrs->__pod_cpum 4257 * The result is stored in @attrs->__pod_cpumask.
5159 * 4258 *
5160 * If pod affinity is not enabled, @attrs->cp 4259 * If pod affinity is not enabled, @attrs->cpumask is always used. If enabled
5161 * and @pod has online CPUs requested by @att 4260 * and @pod has online CPUs requested by @attrs, the returned cpumask is the
5162 * intersection of the possible CPUs of @pod 4261 * intersection of the possible CPUs of @pod and @attrs->cpumask.
5163 * 4262 *
5164 * The caller is responsible for ensuring tha 4263 * The caller is responsible for ensuring that the cpumask of @pod stays stable.
5165 */ 4264 */
5166 static void wq_calc_pod_cpumask(struct workqu !! 4265 static void wq_calc_pod_cpumask(struct workqueue_attrs *attrs, int cpu,
>> 4266 int cpu_going_down)
5167 { 4267 {
5168 const struct wq_pod_type *pt = wqattr 4268 const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
5169 int pod = pt->cpu_pod[cpu]; 4269 int pod = pt->cpu_pod[cpu];
5170 4270
5171 /* calculate possible CPUs in @pod th <<
5172 cpumask_and(attrs->__pod_cpumask, pt- <<
5173 /* does @pod have any online CPUs @at 4271 /* does @pod have any online CPUs @attrs wants? */
5174 if (!cpumask_intersects(attrs->__pod_ !! 4272 cpumask_and(attrs->__pod_cpumask, pt->pod_cpus[pod], attrs->cpumask);
>> 4273 cpumask_and(attrs->__pod_cpumask, attrs->__pod_cpumask, cpu_online_mask);
>> 4274 if (cpu_going_down >= 0)
>> 4275 cpumask_clear_cpu(cpu_going_down, attrs->__pod_cpumask);
>> 4276
>> 4277 if (cpumask_empty(attrs->__pod_cpumask)) {
5175 cpumask_copy(attrs->__pod_cpu 4278 cpumask_copy(attrs->__pod_cpumask, attrs->cpumask);
5176 return; 4279 return;
5177 } 4280 }
>> 4281
>> 4282 /* yeap, return possible CPUs in @pod that @attrs wants */
>> 4283 cpumask_and(attrs->__pod_cpumask, attrs->cpumask, pt->pod_cpus[pod]);
>> 4284
>> 4285 if (cpumask_empty(attrs->__pod_cpumask))
>> 4286 pr_warn_once("WARNING: workqueue cpumask: online intersect > "
>> 4287 "possible intersect\n");
5178 } 4288 }
5179 4289
5180 /* install @pwq into @wq and return the old p !! 4290 /* install @pwq into @wq's cpu_pwq and return the old pwq */
5181 static struct pool_workqueue *install_unbound 4291 static struct pool_workqueue *install_unbound_pwq(struct workqueue_struct *wq,
5182 int c 4292 int cpu, struct pool_workqueue *pwq)
5183 { 4293 {
5184 struct pool_workqueue __rcu **slot = <<
5185 struct pool_workqueue *old_pwq; 4294 struct pool_workqueue *old_pwq;
5186 4295
5187 lockdep_assert_held(&wq_pool_mutex); 4296 lockdep_assert_held(&wq_pool_mutex);
5188 lockdep_assert_held(&wq->mutex); 4297 lockdep_assert_held(&wq->mutex);
5189 4298
5190 /* link_pwq() can handle duplicate ca 4299 /* link_pwq() can handle duplicate calls */
5191 link_pwq(pwq); 4300 link_pwq(pwq);
5192 4301
5193 old_pwq = rcu_access_pointer(*slot); !! 4302 old_pwq = rcu_access_pointer(*per_cpu_ptr(wq->cpu_pwq, cpu));
5194 rcu_assign_pointer(*slot, pwq); !! 4303 rcu_assign_pointer(*per_cpu_ptr(wq->cpu_pwq, cpu), pwq);
5195 return old_pwq; 4304 return old_pwq;
5196 } 4305 }
5197 4306
5198 /* context to store the prepared attrs & pwqs 4307 /* context to store the prepared attrs & pwqs before applying */
5199 struct apply_wqattrs_ctx { 4308 struct apply_wqattrs_ctx {
5200 struct workqueue_struct *wq; 4309 struct workqueue_struct *wq; /* target workqueue */
5201 struct workqueue_attrs *attrs; 4310 struct workqueue_attrs *attrs; /* attrs to apply */
5202 struct list_head list; 4311 struct list_head list; /* queued for batching commit */
5203 struct pool_workqueue *dfl_pwq; 4312 struct pool_workqueue *dfl_pwq;
5204 struct pool_workqueue *pwq_tbl[]; 4313 struct pool_workqueue *pwq_tbl[];
5205 }; 4314 };
5206 4315
5207 /* free the resources after success or abort 4316 /* free the resources after success or abort */
5208 static void apply_wqattrs_cleanup(struct appl 4317 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
5209 { 4318 {
5210 if (ctx) { 4319 if (ctx) {
5211 int cpu; 4320 int cpu;
5212 4321
5213 for_each_possible_cpu(cpu) 4322 for_each_possible_cpu(cpu)
5214 put_pwq_unlocked(ctx- 4323 put_pwq_unlocked(ctx->pwq_tbl[cpu]);
5215 put_pwq_unlocked(ctx->dfl_pwq 4324 put_pwq_unlocked(ctx->dfl_pwq);
5216 4325
5217 free_workqueue_attrs(ctx->att 4326 free_workqueue_attrs(ctx->attrs);
5218 4327
5219 kfree(ctx); 4328 kfree(ctx);
5220 } 4329 }
5221 } 4330 }
5222 4331
5223 /* allocate the attrs and pwqs for later inst 4332 /* allocate the attrs and pwqs for later installation */
5224 static struct apply_wqattrs_ctx * 4333 static struct apply_wqattrs_ctx *
5225 apply_wqattrs_prepare(struct workqueue_struct 4334 apply_wqattrs_prepare(struct workqueue_struct *wq,
5226 const struct workqueue_ 4335 const struct workqueue_attrs *attrs,
5227 const cpumask_var_t unb 4336 const cpumask_var_t unbound_cpumask)
5228 { 4337 {
5229 struct apply_wqattrs_ctx *ctx; 4338 struct apply_wqattrs_ctx *ctx;
5230 struct workqueue_attrs *new_attrs; 4339 struct workqueue_attrs *new_attrs;
5231 int cpu; 4340 int cpu;
5232 4341
5233 lockdep_assert_held(&wq_pool_mutex); 4342 lockdep_assert_held(&wq_pool_mutex);
5234 4343
5235 if (WARN_ON(attrs->affn_scope < 0 || 4344 if (WARN_ON(attrs->affn_scope < 0 ||
5236 attrs->affn_scope >= WQ_A 4345 attrs->affn_scope >= WQ_AFFN_NR_TYPES))
5237 return ERR_PTR(-EINVAL); 4346 return ERR_PTR(-EINVAL);
5238 4347
5239 ctx = kzalloc(struct_size(ctx, pwq_tb 4348 ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_cpu_ids), GFP_KERNEL);
5240 4349
5241 new_attrs = alloc_workqueue_attrs(); 4350 new_attrs = alloc_workqueue_attrs();
5242 if (!ctx || !new_attrs) 4351 if (!ctx || !new_attrs)
5243 goto out_free; 4352 goto out_free;
5244 4353
5245 /* 4354 /*
5246 * If something goes wrong during CPU 4355 * If something goes wrong during CPU up/down, we'll fall back to
5247 * the default pwq covering whole @at 4356 * the default pwq covering whole @attrs->cpumask. Always create
5248 * it even if we don't use it immedia 4357 * it even if we don't use it immediately.
5249 */ 4358 */
5250 copy_workqueue_attrs(new_attrs, attrs 4359 copy_workqueue_attrs(new_attrs, attrs);
5251 wqattrs_actualize_cpumask(new_attrs, 4360 wqattrs_actualize_cpumask(new_attrs, unbound_cpumask);
5252 cpumask_copy(new_attrs->__pod_cpumask 4361 cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask);
5253 ctx->dfl_pwq = alloc_unbound_pwq(wq, 4362 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
5254 if (!ctx->dfl_pwq) 4363 if (!ctx->dfl_pwq)
5255 goto out_free; 4364 goto out_free;
5256 4365
5257 for_each_possible_cpu(cpu) { 4366 for_each_possible_cpu(cpu) {
5258 if (new_attrs->ordered) { 4367 if (new_attrs->ordered) {
5259 ctx->dfl_pwq->refcnt+ 4368 ctx->dfl_pwq->refcnt++;
5260 ctx->pwq_tbl[cpu] = c 4369 ctx->pwq_tbl[cpu] = ctx->dfl_pwq;
5261 } else { 4370 } else {
5262 wq_calc_pod_cpumask(n !! 4371 wq_calc_pod_cpumask(new_attrs, cpu, -1);
5263 ctx->pwq_tbl[cpu] = a 4372 ctx->pwq_tbl[cpu] = alloc_unbound_pwq(wq, new_attrs);
5264 if (!ctx->pwq_tbl[cpu 4373 if (!ctx->pwq_tbl[cpu])
5265 goto out_free 4374 goto out_free;
5266 } 4375 }
5267 } 4376 }
5268 4377
5269 /* save the user configured attrs and 4378 /* save the user configured attrs and sanitize it. */
5270 copy_workqueue_attrs(new_attrs, attrs 4379 copy_workqueue_attrs(new_attrs, attrs);
5271 cpumask_and(new_attrs->cpumask, new_a 4380 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
5272 cpumask_copy(new_attrs->__pod_cpumask 4381 cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask);
5273 ctx->attrs = new_attrs; 4382 ctx->attrs = new_attrs;
5274 4383
5275 /* <<
5276 * For initialized ordered workqueues <<
5277 * (dfl_pwq). Set the plugged flag of <<
5278 * of newly queued work items until e <<
5279 * the old pwq's have completed. <<
5280 */ <<
5281 if ((wq->flags & __WQ_ORDERED) && !li <<
5282 ctx->dfl_pwq->plugged = true; <<
5283 <<
5284 ctx->wq = wq; 4384 ctx->wq = wq;
5285 return ctx; 4385 return ctx;
5286 4386
5287 out_free: 4387 out_free:
5288 free_workqueue_attrs(new_attrs); 4388 free_workqueue_attrs(new_attrs);
5289 apply_wqattrs_cleanup(ctx); 4389 apply_wqattrs_cleanup(ctx);
5290 return ERR_PTR(-ENOMEM); 4390 return ERR_PTR(-ENOMEM);
5291 } 4391 }
5292 4392
5293 /* set attrs and install prepared pwqs, @ctx 4393 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
5294 static void apply_wqattrs_commit(struct apply 4394 static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
5295 { 4395 {
5296 int cpu; 4396 int cpu;
5297 4397
5298 /* all pwqs have been created success 4398 /* all pwqs have been created successfully, let's install'em */
5299 mutex_lock(&ctx->wq->mutex); 4399 mutex_lock(&ctx->wq->mutex);
5300 4400
5301 copy_workqueue_attrs(ctx->wq->unbound 4401 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
5302 4402
5303 /* save the previous pwqs and install !! 4403 /* save the previous pwq and install the new one */
5304 for_each_possible_cpu(cpu) 4404 for_each_possible_cpu(cpu)
5305 ctx->pwq_tbl[cpu] = install_u 4405 ctx->pwq_tbl[cpu] = install_unbound_pwq(ctx->wq, cpu,
5306 4406 ctx->pwq_tbl[cpu]);
5307 ctx->dfl_pwq = install_unbound_pwq(ct <<
5308 4407
5309 /* update node_nr_active->max */ !! 4408 /* @dfl_pwq might not have been used, ensure it's linked */
5310 wq_update_node_max_active(ctx->wq, -1 !! 4409 link_pwq(ctx->dfl_pwq);
5311 !! 4410 swap(ctx->wq->dfl_pwq, ctx->dfl_pwq);
5312 /* rescuer needs to respect wq cpumas <<
5313 if (ctx->wq->rescuer) <<
5314 set_cpus_allowed_ptr(ctx->wq- <<
5315 unbound_ <<
5316 4411
5317 mutex_unlock(&ctx->wq->mutex); 4412 mutex_unlock(&ctx->wq->mutex);
5318 } 4413 }
5319 4414
>> 4415 static void apply_wqattrs_lock(void)
>> 4416 {
>> 4417 /* CPUs should stay stable across pwq creations and installations */
>> 4418 cpus_read_lock();
>> 4419 mutex_lock(&wq_pool_mutex);
>> 4420 }
>> 4421
>> 4422 static void apply_wqattrs_unlock(void)
>> 4423 {
>> 4424 mutex_unlock(&wq_pool_mutex);
>> 4425 cpus_read_unlock();
>> 4426 }
>> 4427
5320 static int apply_workqueue_attrs_locked(struc 4428 static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
5321 const 4429 const struct workqueue_attrs *attrs)
5322 { 4430 {
5323 struct apply_wqattrs_ctx *ctx; 4431 struct apply_wqattrs_ctx *ctx;
5324 4432
5325 /* only unbound workqueues can change 4433 /* only unbound workqueues can change attributes */
5326 if (WARN_ON(!(wq->flags & WQ_UNBOUND) 4434 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
5327 return -EINVAL; 4435 return -EINVAL;
5328 4436
>> 4437 /* creating multiple pwqs breaks ordering guarantee */
>> 4438 if (!list_empty(&wq->pwqs)) {
>> 4439 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
>> 4440 return -EINVAL;
>> 4441
>> 4442 wq->flags &= ~__WQ_ORDERED;
>> 4443 }
>> 4444
5329 ctx = apply_wqattrs_prepare(wq, attrs 4445 ctx = apply_wqattrs_prepare(wq, attrs, wq_unbound_cpumask);
5330 if (IS_ERR(ctx)) 4446 if (IS_ERR(ctx))
5331 return PTR_ERR(ctx); 4447 return PTR_ERR(ctx);
5332 4448
5333 /* the ctx has been prepared successf 4449 /* the ctx has been prepared successfully, let's commit it */
5334 apply_wqattrs_commit(ctx); 4450 apply_wqattrs_commit(ctx);
5335 apply_wqattrs_cleanup(ctx); 4451 apply_wqattrs_cleanup(ctx);
5336 4452
5337 return 0; 4453 return 0;
5338 } 4454 }
5339 4455
5340 /** 4456 /**
5341 * apply_workqueue_attrs - apply new workqueu 4457 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
5342 * @wq: the target workqueue 4458 * @wq: the target workqueue
5343 * @attrs: the workqueue_attrs to apply, allo 4459 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
5344 * 4460 *
5345 * Apply @attrs to an unbound workqueue @wq. 4461 * Apply @attrs to an unbound workqueue @wq. Unless disabled, this function maps
5346 * a separate pwq to each CPU pod with possib 4462 * a separate pwq to each CPU pod with possibles CPUs in @attrs->cpumask so that
5347 * work items are affine to the pod it was is 4463 * work items are affine to the pod it was issued on. Older pwqs are released as
5348 * in-flight work items finish. Note that a w 4464 * in-flight work items finish. Note that a work item which repeatedly requeues
5349 * itself back-to-back will stay on its curre 4465 * itself back-to-back will stay on its current pwq.
5350 * 4466 *
5351 * Performs GFP_KERNEL allocations. 4467 * Performs GFP_KERNEL allocations.
5352 * 4468 *
>> 4469 * Assumes caller has CPU hotplug read exclusion, i.e. cpus_read_lock().
>> 4470 *
5353 * Return: 0 on success and -errno on failure 4471 * Return: 0 on success and -errno on failure.
5354 */ 4472 */
5355 int apply_workqueue_attrs(struct workqueue_st 4473 int apply_workqueue_attrs(struct workqueue_struct *wq,
5356 const struct workqu 4474 const struct workqueue_attrs *attrs)
5357 { 4475 {
5358 int ret; 4476 int ret;
5359 4477
>> 4478 lockdep_assert_cpus_held();
>> 4479
5360 mutex_lock(&wq_pool_mutex); 4480 mutex_lock(&wq_pool_mutex);
5361 ret = apply_workqueue_attrs_locked(wq 4481 ret = apply_workqueue_attrs_locked(wq, attrs);
5362 mutex_unlock(&wq_pool_mutex); 4482 mutex_unlock(&wq_pool_mutex);
5363 4483
5364 return ret; 4484 return ret;
5365 } 4485 }
5366 4486
5367 /** 4487 /**
5368 * unbound_wq_update_pwq - update a pwq slot !! 4488 * wq_update_pod - update pod affinity of a wq for CPU hot[un]plug
5369 * @wq: the target workqueue 4489 * @wq: the target workqueue
5370 * @cpu: the CPU to update the pwq slot for !! 4490 * @cpu: the CPU to update pool association for
>> 4491 * @hotplug_cpu: the CPU coming up or going down
>> 4492 * @online: whether @cpu is coming up or going down
5371 * 4493 *
5372 * This function is to be called from %CPU_DO 4494 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
5373 * %CPU_DOWN_FAILED. @cpu is in the same pod !! 4495 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update pod affinity of
>> 4496 * @wq accordingly.
5374 * 4497 *
5375 * 4498 *
5376 * If pod affinity can't be adjusted due to m 4499 * If pod affinity can't be adjusted due to memory allocation failure, it falls
5377 * back to @wq->dfl_pwq which may not be opti 4500 * back to @wq->dfl_pwq which may not be optimal but is always correct.
5378 * 4501 *
5379 * Note that when the last allowed CPU of a p 4502 * Note that when the last allowed CPU of a pod goes offline for a workqueue
5380 * with a cpumask spanning multiple pods, the 4503 * with a cpumask spanning multiple pods, the workers which were already
5381 * executing the work items for the workqueue 4504 * executing the work items for the workqueue will lose their CPU affinity and
5382 * may execute on any CPU. This is similar to 4505 * may execute on any CPU. This is similar to how per-cpu workqueues behave on
5383 * CPU_DOWN. If a workqueue user wants strict 4506 * CPU_DOWN. If a workqueue user wants strict affinity, it's the user's
5384 * responsibility to flush the work item from 4507 * responsibility to flush the work item from CPU_DOWN_PREPARE.
5385 */ 4508 */
5386 static void unbound_wq_update_pwq(struct work !! 4509 static void wq_update_pod(struct workqueue_struct *wq, int cpu,
>> 4510 int hotplug_cpu, bool online)
5387 { 4511 {
>> 4512 int off_cpu = online ? -1 : hotplug_cpu;
5388 struct pool_workqueue *old_pwq = NULL 4513 struct pool_workqueue *old_pwq = NULL, *pwq;
5389 struct workqueue_attrs *target_attrs; 4514 struct workqueue_attrs *target_attrs;
5390 4515
5391 lockdep_assert_held(&wq_pool_mutex); 4516 lockdep_assert_held(&wq_pool_mutex);
5392 4517
5393 if (!(wq->flags & WQ_UNBOUND) || wq-> 4518 if (!(wq->flags & WQ_UNBOUND) || wq->unbound_attrs->ordered)
5394 return; 4519 return;
5395 4520
5396 /* 4521 /*
5397 * We don't wanna alloc/free wq_attrs 4522 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
5398 * Let's use a preallocated one. The 4523 * Let's use a preallocated one. The following buf is protected by
5399 * CPU hotplug exclusion. 4524 * CPU hotplug exclusion.
5400 */ 4525 */
5401 target_attrs = unbound_wq_update_pwq_ !! 4526 target_attrs = wq_update_pod_attrs_buf;
5402 4527
5403 copy_workqueue_attrs(target_attrs, wq 4528 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
5404 wqattrs_actualize_cpumask(target_attr 4529 wqattrs_actualize_cpumask(target_attrs, wq_unbound_cpumask);
5405 4530
5406 /* nothing to do if the target cpumas 4531 /* nothing to do if the target cpumask matches the current pwq */
5407 wq_calc_pod_cpumask(target_attrs, cpu !! 4532 wq_calc_pod_cpumask(target_attrs, cpu, off_cpu);
5408 if (wqattrs_equal(target_attrs, unbou !! 4533 pwq = rcu_dereference_protected(*per_cpu_ptr(wq->cpu_pwq, cpu),
>> 4534 lockdep_is_held(&wq_pool_mutex));
>> 4535 if (wqattrs_equal(target_attrs, pwq->pool->attrs))
5409 return; 4536 return;
5410 4537
5411 /* create a new pwq */ 4538 /* create a new pwq */
5412 pwq = alloc_unbound_pwq(wq, target_at 4539 pwq = alloc_unbound_pwq(wq, target_attrs);
5413 if (!pwq) { 4540 if (!pwq) {
5414 pr_warn("workqueue: allocatio 4541 pr_warn("workqueue: allocation failed while updating CPU pod affinity of \"%s\"\n",
5415 wq->name); 4542 wq->name);
5416 goto use_dfl_pwq; 4543 goto use_dfl_pwq;
5417 } 4544 }
5418 4545
5419 /* Install the new pwq. */ 4546 /* Install the new pwq. */
5420 mutex_lock(&wq->mutex); 4547 mutex_lock(&wq->mutex);
5421 old_pwq = install_unbound_pwq(wq, cpu 4548 old_pwq = install_unbound_pwq(wq, cpu, pwq);
5422 goto out_unlock; 4549 goto out_unlock;
5423 4550
5424 use_dfl_pwq: 4551 use_dfl_pwq:
5425 mutex_lock(&wq->mutex); 4552 mutex_lock(&wq->mutex);
5426 pwq = unbound_pwq(wq, -1); !! 4553 raw_spin_lock_irq(&wq->dfl_pwq->pool->lock);
5427 raw_spin_lock_irq(&pwq->pool->lock); !! 4554 get_pwq(wq->dfl_pwq);
5428 get_pwq(pwq); !! 4555 raw_spin_unlock_irq(&wq->dfl_pwq->pool->lock);
5429 raw_spin_unlock_irq(&pwq->pool->lock) !! 4556 old_pwq = install_unbound_pwq(wq, cpu, wq->dfl_pwq);
5430 old_pwq = install_unbound_pwq(wq, cpu <<
5431 out_unlock: 4557 out_unlock:
5432 mutex_unlock(&wq->mutex); 4558 mutex_unlock(&wq->mutex);
5433 put_pwq_unlocked(old_pwq); 4559 put_pwq_unlocked(old_pwq);
5434 } 4560 }
5435 4561
5436 static int alloc_and_link_pwqs(struct workque 4562 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
5437 { 4563 {
5438 bool highpri = wq->flags & WQ_HIGHPRI 4564 bool highpri = wq->flags & WQ_HIGHPRI;
5439 int cpu, ret; 4565 int cpu, ret;
5440 4566
5441 lockdep_assert_held(&wq_pool_mutex); <<
5442 <<
5443 wq->cpu_pwq = alloc_percpu(struct poo 4567 wq->cpu_pwq = alloc_percpu(struct pool_workqueue *);
5444 if (!wq->cpu_pwq) 4568 if (!wq->cpu_pwq)
5445 goto enomem; 4569 goto enomem;
5446 4570
5447 if (!(wq->flags & WQ_UNBOUND)) { 4571 if (!(wq->flags & WQ_UNBOUND)) {
5448 struct worker_pool __percpu * <<
5449 <<
5450 if (wq->flags & WQ_BH) <<
5451 pools = bh_worker_poo <<
5452 else <<
5453 pools = cpu_worker_po <<
5454 <<
5455 for_each_possible_cpu(cpu) { 4572 for_each_possible_cpu(cpu) {
5456 struct pool_workqueue !! 4573 struct pool_workqueue **pwq_p =
5457 struct worker_pool *p !! 4574 per_cpu_ptr(wq->cpu_pwq, cpu);
5458 !! 4575 struct worker_pool *pool =
5459 pool = &(per_cpu_ptr( !! 4576 &(per_cpu_ptr(cpu_worker_pools, cpu)[highpri]);
5460 pwq_p = per_cpu_ptr(w <<
5461 4577
5462 *pwq_p = kmem_cache_a 4578 *pwq_p = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL,
5463 4579 pool->node);
5464 if (!*pwq_p) 4580 if (!*pwq_p)
5465 goto enomem; 4581 goto enomem;
5466 4582
5467 init_pwq(*pwq_p, wq, 4583 init_pwq(*pwq_p, wq, pool);
5468 4584
5469 mutex_lock(&wq->mutex 4585 mutex_lock(&wq->mutex);
5470 link_pwq(*pwq_p); 4586 link_pwq(*pwq_p);
5471 mutex_unlock(&wq->mut 4587 mutex_unlock(&wq->mutex);
5472 } 4588 }
5473 return 0; 4589 return 0;
5474 } 4590 }
5475 4591
>> 4592 cpus_read_lock();
5476 if (wq->flags & __WQ_ORDERED) { 4593 if (wq->flags & __WQ_ORDERED) {
5477 struct pool_workqueue *dfl_pw !! 4594 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
5478 <<
5479 ret = apply_workqueue_attrs_l <<
5480 /* there should only be singl 4595 /* there should only be single pwq for ordering guarantee */
5481 dfl_pwq = rcu_access_pointer( !! 4596 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
5482 WARN(!ret && (wq->pwqs.next ! !! 4597 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
5483 wq->pwqs.prev ! <<
5484 "ordering guarantee brok 4598 "ordering guarantee broken for workqueue %s\n", wq->name);
5485 } else { 4599 } else {
5486 ret = apply_workqueue_attrs_l !! 4600 ret = apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
5487 } 4601 }
>> 4602 cpus_read_unlock();
>> 4603
>> 4604 /* for unbound pwq, flush the pwq_release_worker ensures that the
>> 4605 * pwq_release_workfn() completes before calling kfree(wq).
>> 4606 */
>> 4607 if (ret)
>> 4608 kthread_flush_worker(pwq_release_worker);
5488 4609
5489 return ret; 4610 return ret;
5490 4611
5491 enomem: 4612 enomem:
5492 if (wq->cpu_pwq) { 4613 if (wq->cpu_pwq) {
5493 for_each_possible_cpu(cpu) { 4614 for_each_possible_cpu(cpu) {
5494 struct pool_workqueue 4615 struct pool_workqueue *pwq = *per_cpu_ptr(wq->cpu_pwq, cpu);
5495 4616
5496 if (pwq) 4617 if (pwq)
5497 kmem_cache_fr 4618 kmem_cache_free(pwq_cache, pwq);
5498 } 4619 }
5499 free_percpu(wq->cpu_pwq); 4620 free_percpu(wq->cpu_pwq);
5500 wq->cpu_pwq = NULL; 4621 wq->cpu_pwq = NULL;
5501 } 4622 }
5502 return -ENOMEM; 4623 return -ENOMEM;
5503 } 4624 }
5504 4625
5505 static int wq_clamp_max_active(int max_active 4626 static int wq_clamp_max_active(int max_active, unsigned int flags,
5506 const char *na 4627 const char *name)
5507 { 4628 {
5508 if (max_active < 1 || max_active > WQ 4629 if (max_active < 1 || max_active > WQ_MAX_ACTIVE)
5509 pr_warn("workqueue: max_activ 4630 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
5510 max_active, name, 1, 4631 max_active, name, 1, WQ_MAX_ACTIVE);
5511 4632
5512 return clamp_val(max_active, 1, WQ_MA 4633 return clamp_val(max_active, 1, WQ_MAX_ACTIVE);
5513 } 4634 }
5514 4635
5515 /* 4636 /*
5516 * Workqueues which may be used during memory 4637 * Workqueues which may be used during memory reclaim should have a rescuer
5517 * to guarantee forward progress. 4638 * to guarantee forward progress.
5518 */ 4639 */
5519 static int init_rescuer(struct workqueue_stru 4640 static int init_rescuer(struct workqueue_struct *wq)
5520 { 4641 {
5521 struct worker *rescuer; 4642 struct worker *rescuer;
5522 char id_buf[WORKER_ID_LEN]; <<
5523 int ret; 4643 int ret;
5524 4644
5525 lockdep_assert_held(&wq_pool_mutex); <<
5526 <<
5527 if (!(wq->flags & WQ_MEM_RECLAIM)) 4645 if (!(wq->flags & WQ_MEM_RECLAIM))
5528 return 0; 4646 return 0;
5529 4647
5530 rescuer = alloc_worker(NUMA_NO_NODE); 4648 rescuer = alloc_worker(NUMA_NO_NODE);
5531 if (!rescuer) { 4649 if (!rescuer) {
5532 pr_err("workqueue: Failed to 4650 pr_err("workqueue: Failed to allocate a rescuer for wq \"%s\"\n",
5533 wq->name); 4651 wq->name);
5534 return -ENOMEM; 4652 return -ENOMEM;
5535 } 4653 }
5536 4654
5537 rescuer->rescue_wq = wq; 4655 rescuer->rescue_wq = wq;
5538 format_worker_id(id_buf, sizeof(id_bu !! 4656 rescuer->task = kthread_create(rescuer_thread, rescuer, "kworker/R-%s", wq->name);
5539 <<
5540 rescuer->task = kthread_create(rescue <<
5541 if (IS_ERR(rescuer->task)) { 4657 if (IS_ERR(rescuer->task)) {
5542 ret = PTR_ERR(rescuer->task); 4658 ret = PTR_ERR(rescuer->task);
5543 pr_err("workqueue: Failed to 4659 pr_err("workqueue: Failed to create a rescuer kthread for wq \"%s\": %pe",
5544 wq->name, ERR_PTR(ret) 4660 wq->name, ERR_PTR(ret));
5545 kfree(rescuer); 4661 kfree(rescuer);
5546 return ret; 4662 return ret;
5547 } 4663 }
5548 4664
5549 wq->rescuer = rescuer; 4665 wq->rescuer = rescuer;
5550 if (wq->flags & WQ_UNBOUND) !! 4666 kthread_bind_mask(rescuer->task, cpu_possible_mask);
5551 kthread_bind_mask(rescuer->ta <<
5552 else <<
5553 kthread_bind_mask(rescuer->ta <<
5554 wake_up_process(rescuer->task); 4667 wake_up_process(rescuer->task);
5555 4668
5556 return 0; 4669 return 0;
5557 } 4670 }
5558 4671
5559 /** <<
5560 * wq_adjust_max_active - update a wq's max_a <<
5561 * @wq: target workqueue <<
5562 * <<
5563 * If @wq isn't freezing, set @wq->max_active <<
5564 * activate inactive work items accordingly. <<
5565 * @wq->max_active to zero. <<
5566 */ <<
5567 static void wq_adjust_max_active(struct workq <<
5568 { <<
5569 bool activated; <<
5570 int new_max, new_min; <<
5571 <<
5572 lockdep_assert_held(&wq->mutex); <<
5573 <<
5574 if ((wq->flags & WQ_FREEZABLE) && wor <<
5575 new_max = 0; <<
5576 new_min = 0; <<
5577 } else { <<
5578 new_max = wq->saved_max_activ <<
5579 new_min = wq->saved_min_activ <<
5580 } <<
5581 <<
5582 if (wq->max_active == new_max && wq-> <<
5583 return; <<
5584 <<
5585 /* <<
5586 * Update @wq->max/min_active and the <<
5587 * active work items are allowed. Thi <<
5588 * because new work items are always <<
5589 * work items if there are any. <<
5590 */ <<
5591 WRITE_ONCE(wq->max_active, new_max); <<
5592 WRITE_ONCE(wq->min_active, new_min); <<
5593 <<
5594 if (wq->flags & WQ_UNBOUND) <<
5595 wq_update_node_max_active(wq, <<
5596 <<
5597 if (new_max == 0) <<
5598 return; <<
5599 <<
5600 /* <<
5601 * Round-robin through pwq's activati <<
5602 * until max_active is filled. <<
5603 */ <<
5604 do { <<
5605 struct pool_workqueue *pwq; <<
5606 <<
5607 activated = false; <<
5608 for_each_pwq(pwq, wq) { <<
5609 unsigned long irq_fla <<
5610 <<
5611 /* can be called duri <<
5612 raw_spin_lock_irqsave <<
5613 if (pwq_activate_firs <<
5614 activated = t <<
5615 kick_pool(pwq <<
5616 } <<
5617 raw_spin_unlock_irqre <<
5618 } <<
5619 } while (activated); <<
5620 } <<
5621 <<
5622 __printf(1, 4) 4672 __printf(1, 4)
5623 struct workqueue_struct *alloc_workqueue(cons 4673 struct workqueue_struct *alloc_workqueue(const char *fmt,
5624 unsi 4674 unsigned int flags,
5625 int 4675 int max_active, ...)
5626 { 4676 {
5627 va_list args; 4677 va_list args;
5628 struct workqueue_struct *wq; 4678 struct workqueue_struct *wq;
5629 size_t wq_size; !! 4679 struct pool_workqueue *pwq;
5630 int name_len; <<
5631 4680
5632 if (flags & WQ_BH) { !! 4681 /*
5633 if (WARN_ON_ONCE(flags & ~__W !! 4682 * Unbound && max_active == 1 used to imply ordered, which is no longer
5634 return NULL; !! 4683 * the case on many machines due to per-pod pools. While
5635 if (WARN_ON_ONCE(max_active)) !! 4684 * alloc_ordered_workqueue() is the right way to create an ordered
5636 return NULL; !! 4685 * workqueue, keep the previous behavior to avoid subtle breakages.
5637 } !! 4686 */
>> 4687 if ((flags & WQ_UNBOUND) && max_active == 1)
>> 4688 flags |= __WQ_ORDERED;
5638 4689
5639 /* see the comment above the definiti 4690 /* see the comment above the definition of WQ_POWER_EFFICIENT */
5640 if ((flags & WQ_POWER_EFFICIENT) && w 4691 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
5641 flags |= WQ_UNBOUND; 4692 flags |= WQ_UNBOUND;
5642 4693
5643 /* allocate wq and format name */ 4694 /* allocate wq and format name */
5644 if (flags & WQ_UNBOUND) !! 4695 wq = kzalloc(sizeof(*wq), GFP_KERNEL);
5645 wq_size = struct_size(wq, nod <<
5646 else <<
5647 wq_size = sizeof(*wq); <<
5648 <<
5649 wq = kzalloc(wq_size, GFP_KERNEL); <<
5650 if (!wq) 4696 if (!wq)
5651 return NULL; 4697 return NULL;
5652 4698
5653 if (flags & WQ_UNBOUND) { 4699 if (flags & WQ_UNBOUND) {
5654 wq->unbound_attrs = alloc_wor 4700 wq->unbound_attrs = alloc_workqueue_attrs();
5655 if (!wq->unbound_attrs) 4701 if (!wq->unbound_attrs)
5656 goto err_free_wq; 4702 goto err_free_wq;
5657 } 4703 }
5658 4704
5659 va_start(args, max_active); 4705 va_start(args, max_active);
5660 name_len = vsnprintf(wq->name, sizeof !! 4706 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
5661 va_end(args); 4707 va_end(args);
5662 4708
5663 if (name_len >= WQ_NAME_LEN) !! 4709 max_active = max_active ?: WQ_DFL_ACTIVE;
5664 pr_warn_once("workqueue: name !! 4710 max_active = wq_clamp_max_active(max_active, flags, wq->name);
5665 wq->name); <<
5666 <<
5667 if (flags & WQ_BH) { <<
5668 /* <<
5669 * BH workqueues always share <<
5670 * and don't impose any max_a <<
5671 */ <<
5672 max_active = INT_MAX; <<
5673 } else { <<
5674 max_active = max_active ?: WQ <<
5675 max_active = wq_clamp_max_act <<
5676 } <<
5677 4711
5678 /* init wq */ 4712 /* init wq */
5679 wq->flags = flags; 4713 wq->flags = flags;
5680 wq->max_active = max_active; !! 4714 wq->saved_max_active = max_active;
5681 wq->min_active = min(max_active, WQ_D <<
5682 wq->saved_max_active = wq->max_active <<
5683 wq->saved_min_active = wq->min_active <<
5684 mutex_init(&wq->mutex); 4715 mutex_init(&wq->mutex);
5685 atomic_set(&wq->nr_pwqs_to_flush, 0); 4716 atomic_set(&wq->nr_pwqs_to_flush, 0);
5686 INIT_LIST_HEAD(&wq->pwqs); 4717 INIT_LIST_HEAD(&wq->pwqs);
5687 INIT_LIST_HEAD(&wq->flusher_queue); 4718 INIT_LIST_HEAD(&wq->flusher_queue);
5688 INIT_LIST_HEAD(&wq->flusher_overflow) 4719 INIT_LIST_HEAD(&wq->flusher_overflow);
5689 INIT_LIST_HEAD(&wq->maydays); 4720 INIT_LIST_HEAD(&wq->maydays);
5690 4721
5691 wq_init_lockdep(wq); 4722 wq_init_lockdep(wq);
5692 INIT_LIST_HEAD(&wq->list); 4723 INIT_LIST_HEAD(&wq->list);
5693 4724
5694 if (flags & WQ_UNBOUND) { !! 4725 if (alloc_and_link_pwqs(wq) < 0)
5695 if (alloc_node_nr_active(wq-> !! 4726 goto err_unreg_lockdep;
5696 goto err_unreg_lockde !! 4727
5697 } !! 4728 if (wq_online && init_rescuer(wq) < 0)
>> 4729 goto err_destroy;
>> 4730
>> 4731 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
>> 4732 goto err_destroy;
5698 4733
5699 /* 4734 /*
5700 * wq_pool_mutex protects the workque !! 4735 * wq_pool_mutex protects global freeze state and workqueues list.
5701 * and the global freeze state. !! 4736 * Grab it, adjust max_active and add the new @wq to workqueues
>> 4737 * list.
5702 */ 4738 */
5703 apply_wqattrs_lock(); !! 4739 mutex_lock(&wq_pool_mutex);
5704 <<
5705 if (alloc_and_link_pwqs(wq) < 0) <<
5706 goto err_unlock_free_node_nr_ <<
5707 4740
5708 mutex_lock(&wq->mutex); 4741 mutex_lock(&wq->mutex);
5709 wq_adjust_max_active(wq); !! 4742 for_each_pwq(pwq, wq)
>> 4743 pwq_adjust_max_active(pwq);
5710 mutex_unlock(&wq->mutex); 4744 mutex_unlock(&wq->mutex);
5711 4745
5712 list_add_tail_rcu(&wq->list, &workque 4746 list_add_tail_rcu(&wq->list, &workqueues);
5713 4747
5714 if (wq_online && init_rescuer(wq) < 0 !! 4748 mutex_unlock(&wq_pool_mutex);
5715 goto err_unlock_destroy; <<
5716 <<
5717 apply_wqattrs_unlock(); <<
5718 <<
5719 if ((wq->flags & WQ_SYSFS) && workque <<
5720 goto err_destroy; <<
5721 4749
5722 return wq; 4750 return wq;
5723 4751
5724 err_unlock_free_node_nr_active: <<
5725 apply_wqattrs_unlock(); <<
5726 /* <<
5727 * Failed alloc_and_link_pwqs() may l <<
5728 * flushing the pwq_release_worker en <<
5729 * completes before calling kfree(wq) <<
5730 */ <<
5731 if (wq->flags & WQ_UNBOUND) { <<
5732 kthread_flush_worker(pwq_rele <<
5733 free_node_nr_active(wq->node_ <<
5734 } <<
5735 err_unreg_lockdep: 4752 err_unreg_lockdep:
5736 wq_unregister_lockdep(wq); 4753 wq_unregister_lockdep(wq);
5737 wq_free_lockdep(wq); 4754 wq_free_lockdep(wq);
5738 err_free_wq: 4755 err_free_wq:
5739 free_workqueue_attrs(wq->unbound_attr 4756 free_workqueue_attrs(wq->unbound_attrs);
5740 kfree(wq); 4757 kfree(wq);
5741 return NULL; 4758 return NULL;
5742 err_unlock_destroy: <<
5743 apply_wqattrs_unlock(); <<
5744 err_destroy: 4759 err_destroy:
5745 destroy_workqueue(wq); 4760 destroy_workqueue(wq);
5746 return NULL; 4761 return NULL;
5747 } 4762 }
5748 EXPORT_SYMBOL_GPL(alloc_workqueue); 4763 EXPORT_SYMBOL_GPL(alloc_workqueue);
5749 4764
5750 static bool pwq_busy(struct pool_workqueue *p 4765 static bool pwq_busy(struct pool_workqueue *pwq)
5751 { 4766 {
5752 int i; 4767 int i;
5753 4768
5754 for (i = 0; i < WORK_NR_COLORS; i++) 4769 for (i = 0; i < WORK_NR_COLORS; i++)
5755 if (pwq->nr_in_flight[i]) 4770 if (pwq->nr_in_flight[i])
5756 return true; 4771 return true;
5757 4772
5758 if ((pwq != rcu_access_pointer(pwq->w !! 4773 if ((pwq != pwq->wq->dfl_pwq) && (pwq->refcnt > 1))
5759 return true; 4774 return true;
5760 if (!pwq_is_empty(pwq)) !! 4775 if (pwq->nr_active || !list_empty(&pwq->inactive_works))
5761 return true; 4776 return true;
5762 4777
5763 return false; 4778 return false;
5764 } 4779 }
5765 4780
5766 /** 4781 /**
5767 * destroy_workqueue - safely terminate a wor 4782 * destroy_workqueue - safely terminate a workqueue
5768 * @wq: target workqueue 4783 * @wq: target workqueue
5769 * 4784 *
5770 * Safely destroy a workqueue. All work curre 4785 * Safely destroy a workqueue. All work currently pending will be done first.
5771 */ 4786 */
5772 void destroy_workqueue(struct workqueue_struc 4787 void destroy_workqueue(struct workqueue_struct *wq)
5773 { 4788 {
5774 struct pool_workqueue *pwq; 4789 struct pool_workqueue *pwq;
5775 int cpu; 4790 int cpu;
5776 4791
5777 /* 4792 /*
5778 * Remove it from sysfs first so that 4793 * Remove it from sysfs first so that sanity check failure doesn't
5779 * lead to sysfs name conflicts. 4794 * lead to sysfs name conflicts.
5780 */ 4795 */
5781 workqueue_sysfs_unregister(wq); 4796 workqueue_sysfs_unregister(wq);
5782 4797
5783 /* mark the workqueue destruction is 4798 /* mark the workqueue destruction is in progress */
5784 mutex_lock(&wq->mutex); 4799 mutex_lock(&wq->mutex);
5785 wq->flags |= __WQ_DESTROYING; 4800 wq->flags |= __WQ_DESTROYING;
5786 mutex_unlock(&wq->mutex); 4801 mutex_unlock(&wq->mutex);
5787 4802
5788 /* drain it before proceeding with de 4803 /* drain it before proceeding with destruction */
5789 drain_workqueue(wq); 4804 drain_workqueue(wq);
5790 4805
5791 /* kill rescuer, if sanity checks fai 4806 /* kill rescuer, if sanity checks fail, leave it w/o rescuer */
5792 if (wq->rescuer) { 4807 if (wq->rescuer) {
5793 struct worker *rescuer = wq-> 4808 struct worker *rescuer = wq->rescuer;
5794 4809
5795 /* this prevents new queueing 4810 /* this prevents new queueing */
5796 raw_spin_lock_irq(&wq_mayday_ 4811 raw_spin_lock_irq(&wq_mayday_lock);
5797 wq->rescuer = NULL; 4812 wq->rescuer = NULL;
5798 raw_spin_unlock_irq(&wq_mayda 4813 raw_spin_unlock_irq(&wq_mayday_lock);
5799 4814
5800 /* rescuer will empty maydays 4815 /* rescuer will empty maydays list before exiting */
5801 kthread_stop(rescuer->task); 4816 kthread_stop(rescuer->task);
5802 kfree(rescuer); 4817 kfree(rescuer);
5803 } 4818 }
5804 4819
5805 /* 4820 /*
5806 * Sanity checks - grab all the locks 4821 * Sanity checks - grab all the locks so that we wait for all
5807 * in-flight operations which may do 4822 * in-flight operations which may do put_pwq().
5808 */ 4823 */
5809 mutex_lock(&wq_pool_mutex); 4824 mutex_lock(&wq_pool_mutex);
5810 mutex_lock(&wq->mutex); 4825 mutex_lock(&wq->mutex);
5811 for_each_pwq(pwq, wq) { 4826 for_each_pwq(pwq, wq) {
5812 raw_spin_lock_irq(&pwq->pool- 4827 raw_spin_lock_irq(&pwq->pool->lock);
5813 if (WARN_ON(pwq_busy(pwq))) { 4828 if (WARN_ON(pwq_busy(pwq))) {
5814 pr_warn("%s: %s has t 4829 pr_warn("%s: %s has the following busy pwq\n",
5815 __func__, wq- 4830 __func__, wq->name);
5816 show_pwq(pwq); 4831 show_pwq(pwq);
5817 raw_spin_unlock_irq(& 4832 raw_spin_unlock_irq(&pwq->pool->lock);
5818 mutex_unlock(&wq->mut 4833 mutex_unlock(&wq->mutex);
5819 mutex_unlock(&wq_pool 4834 mutex_unlock(&wq_pool_mutex);
5820 show_one_workqueue(wq 4835 show_one_workqueue(wq);
5821 return; 4836 return;
5822 } 4837 }
5823 raw_spin_unlock_irq(&pwq->poo 4838 raw_spin_unlock_irq(&pwq->pool->lock);
5824 } 4839 }
5825 mutex_unlock(&wq->mutex); 4840 mutex_unlock(&wq->mutex);
5826 4841
5827 /* 4842 /*
5828 * wq list is used to freeze wq, remo 4843 * wq list is used to freeze wq, remove from list after
5829 * flushing is complete in case freez 4844 * flushing is complete in case freeze races us.
5830 */ 4845 */
5831 list_del_rcu(&wq->list); 4846 list_del_rcu(&wq->list);
5832 mutex_unlock(&wq_pool_mutex); 4847 mutex_unlock(&wq_pool_mutex);
5833 4848
5834 /* 4849 /*
5835 * We're the sole accessor of @wq. Di 4850 * We're the sole accessor of @wq. Directly access cpu_pwq and dfl_pwq
5836 * to put the base refs. @wq will be 4851 * to put the base refs. @wq will be auto-destroyed from the last
5837 * pwq_put. RCU read lock prevents @w 4852 * pwq_put. RCU read lock prevents @wq from going away from under us.
5838 */ 4853 */
5839 rcu_read_lock(); 4854 rcu_read_lock();
5840 4855
5841 for_each_possible_cpu(cpu) { 4856 for_each_possible_cpu(cpu) {
5842 put_pwq_unlocked(unbound_pwq( !! 4857 pwq = rcu_access_pointer(*per_cpu_ptr(wq->cpu_pwq, cpu));
5843 RCU_INIT_POINTER(*unbound_pwq !! 4858 RCU_INIT_POINTER(*per_cpu_ptr(wq->cpu_pwq, cpu), NULL);
>> 4859 put_pwq_unlocked(pwq);
5844 } 4860 }
5845 4861
5846 put_pwq_unlocked(unbound_pwq(wq, -1)) !! 4862 put_pwq_unlocked(wq->dfl_pwq);
5847 RCU_INIT_POINTER(*unbound_pwq_slot(wq !! 4863 wq->dfl_pwq = NULL;
5848 4864
5849 rcu_read_unlock(); 4865 rcu_read_unlock();
5850 } 4866 }
5851 EXPORT_SYMBOL_GPL(destroy_workqueue); 4867 EXPORT_SYMBOL_GPL(destroy_workqueue);
5852 4868
5853 /** 4869 /**
5854 * workqueue_set_max_active - adjust max_acti 4870 * workqueue_set_max_active - adjust max_active of a workqueue
5855 * @wq: target workqueue 4871 * @wq: target workqueue
5856 * @max_active: new max_active value. 4872 * @max_active: new max_active value.
5857 * 4873 *
5858 * Set max_active of @wq to @max_active. See !! 4874 * Set max_active of @wq to @max_active.
5859 * comment. <<
5860 * 4875 *
5861 * CONTEXT: 4876 * CONTEXT:
5862 * Don't call from IRQ context. 4877 * Don't call from IRQ context.
5863 */ 4878 */
5864 void workqueue_set_max_active(struct workqueu 4879 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
5865 { 4880 {
5866 /* max_active doesn't mean anything f !! 4881 struct pool_workqueue *pwq;
5867 if (WARN_ON(wq->flags & WQ_BH)) !! 4882
5868 return; <<
5869 /* disallow meddling with max_active 4883 /* disallow meddling with max_active for ordered workqueues */
5870 if (WARN_ON(wq->flags & __WQ_ORDERED) !! 4884 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
5871 return; 4885 return;
5872 4886
5873 max_active = wq_clamp_max_active(max_ 4887 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
5874 4888
5875 mutex_lock(&wq->mutex); 4889 mutex_lock(&wq->mutex);
5876 4890
>> 4891 wq->flags &= ~__WQ_ORDERED;
5877 wq->saved_max_active = max_active; 4892 wq->saved_max_active = max_active;
5878 if (wq->flags & WQ_UNBOUND) <<
5879 wq->saved_min_active = min(wq <<
5880 4893
5881 wq_adjust_max_active(wq); !! 4894 for_each_pwq(pwq, wq)
>> 4895 pwq_adjust_max_active(pwq);
5882 4896
5883 mutex_unlock(&wq->mutex); 4897 mutex_unlock(&wq->mutex);
5884 } 4898 }
5885 EXPORT_SYMBOL_GPL(workqueue_set_max_active); 4899 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
5886 4900
5887 /** 4901 /**
5888 * workqueue_set_min_active - adjust min_acti <<
5889 * @wq: target unbound workqueue <<
5890 * @min_active: new min_active value <<
5891 * <<
5892 * Set min_active of an unbound workqueue. Un <<
5893 * unbound workqueue is not guaranteed to be <<
5894 * interdependent work items. Instead, an unb <<
5895 * able to process min_active number of inter <<
5896 * %WQ_DFL_MIN_ACTIVE by default. <<
5897 * <<
5898 * Use this function to adjust the min_active <<
5899 * max_active. <<
5900 */ <<
5901 void workqueue_set_min_active(struct workqueu <<
5902 { <<
5903 /* min_active is only meaningful for <<
5904 if (WARN_ON((wq->flags & (WQ_BH | WQ_ <<
5905 WQ_UNBOUND)) <<
5906 return; <<
5907 <<
5908 mutex_lock(&wq->mutex); <<
5909 wq->saved_min_active = clamp(min_acti <<
5910 wq_adjust_max_active(wq); <<
5911 mutex_unlock(&wq->mutex); <<
5912 } <<
5913 <<
5914 /** <<
5915 * current_work - retrieve %current task's wo 4902 * current_work - retrieve %current task's work struct
5916 * 4903 *
5917 * Determine if %current task is a workqueue 4904 * Determine if %current task is a workqueue worker and what it's working on.
5918 * Useful to find out the context that the %c 4905 * Useful to find out the context that the %current task is running in.
5919 * 4906 *
5920 * Return: work struct if %current task is a 4907 * Return: work struct if %current task is a workqueue worker, %NULL otherwise.
5921 */ 4908 */
5922 struct work_struct *current_work(void) 4909 struct work_struct *current_work(void)
5923 { 4910 {
5924 struct worker *worker = current_wq_wo 4911 struct worker *worker = current_wq_worker();
5925 4912
5926 return worker ? worker->current_work 4913 return worker ? worker->current_work : NULL;
5927 } 4914 }
5928 EXPORT_SYMBOL(current_work); 4915 EXPORT_SYMBOL(current_work);
5929 4916
5930 /** 4917 /**
5931 * current_is_workqueue_rescuer - is %current 4918 * current_is_workqueue_rescuer - is %current workqueue rescuer?
5932 * 4919 *
5933 * Determine whether %current is a workqueue 4920 * Determine whether %current is a workqueue rescuer. Can be used from
5934 * work functions to determine whether it's b 4921 * work functions to determine whether it's being run off the rescuer task.
5935 * 4922 *
5936 * Return: %true if %current is a workqueue r 4923 * Return: %true if %current is a workqueue rescuer. %false otherwise.
5937 */ 4924 */
5938 bool current_is_workqueue_rescuer(void) 4925 bool current_is_workqueue_rescuer(void)
5939 { 4926 {
5940 struct worker *worker = current_wq_wo 4927 struct worker *worker = current_wq_worker();
5941 4928
5942 return worker && worker->rescue_wq; 4929 return worker && worker->rescue_wq;
5943 } 4930 }
5944 4931
5945 /** 4932 /**
5946 * workqueue_congested - test whether a workq 4933 * workqueue_congested - test whether a workqueue is congested
5947 * @cpu: CPU in question 4934 * @cpu: CPU in question
5948 * @wq: target workqueue 4935 * @wq: target workqueue
5949 * 4936 *
5950 * Test whether @wq's cpu workqueue for @cpu 4937 * Test whether @wq's cpu workqueue for @cpu is congested. There is
5951 * no synchronization around this function an 4938 * no synchronization around this function and the test result is
5952 * unreliable and only useful as advisory hin 4939 * unreliable and only useful as advisory hints or for debugging.
5953 * 4940 *
5954 * If @cpu is WORK_CPU_UNBOUND, the test is p 4941 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
5955 * 4942 *
5956 * With the exception of ordered workqueues, 4943 * With the exception of ordered workqueues, all workqueues have per-cpu
5957 * pool_workqueues, each with its own congest 4944 * pool_workqueues, each with its own congested state. A workqueue being
5958 * congested on one CPU doesn't mean that the 4945 * congested on one CPU doesn't mean that the workqueue is contested on any
5959 * other CPUs. 4946 * other CPUs.
5960 * 4947 *
5961 * Return: 4948 * Return:
5962 * %true if congested, %false otherwise. 4949 * %true if congested, %false otherwise.
5963 */ 4950 */
5964 bool workqueue_congested(int cpu, struct work 4951 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
5965 { 4952 {
5966 struct pool_workqueue *pwq; 4953 struct pool_workqueue *pwq;
5967 bool ret; 4954 bool ret;
5968 4955
5969 rcu_read_lock(); 4956 rcu_read_lock();
5970 preempt_disable(); 4957 preempt_disable();
5971 4958
5972 if (cpu == WORK_CPU_UNBOUND) 4959 if (cpu == WORK_CPU_UNBOUND)
5973 cpu = smp_processor_id(); 4960 cpu = smp_processor_id();
5974 4961
5975 pwq = *per_cpu_ptr(wq->cpu_pwq, cpu); 4962 pwq = *per_cpu_ptr(wq->cpu_pwq, cpu);
5976 ret = !list_empty(&pwq->inactive_work 4963 ret = !list_empty(&pwq->inactive_works);
5977 4964
5978 preempt_enable(); 4965 preempt_enable();
5979 rcu_read_unlock(); 4966 rcu_read_unlock();
5980 4967
5981 return ret; 4968 return ret;
5982 } 4969 }
5983 EXPORT_SYMBOL_GPL(workqueue_congested); 4970 EXPORT_SYMBOL_GPL(workqueue_congested);
5984 4971
5985 /** 4972 /**
5986 * work_busy - test whether a work is current 4973 * work_busy - test whether a work is currently pending or running
5987 * @work: the work to be tested 4974 * @work: the work to be tested
5988 * 4975 *
5989 * Test whether @work is currently pending or 4976 * Test whether @work is currently pending or running. There is no
5990 * synchronization around this function and t 4977 * synchronization around this function and the test result is
5991 * unreliable and only useful as advisory hin 4978 * unreliable and only useful as advisory hints or for debugging.
5992 * 4979 *
5993 * Return: 4980 * Return:
5994 * OR'd bitmask of WORK_BUSY_* bits. 4981 * OR'd bitmask of WORK_BUSY_* bits.
5995 */ 4982 */
5996 unsigned int work_busy(struct work_struct *wo 4983 unsigned int work_busy(struct work_struct *work)
5997 { 4984 {
5998 struct worker_pool *pool; 4985 struct worker_pool *pool;
5999 unsigned long irq_flags; !! 4986 unsigned long flags;
6000 unsigned int ret = 0; 4987 unsigned int ret = 0;
6001 4988
6002 if (work_pending(work)) 4989 if (work_pending(work))
6003 ret |= WORK_BUSY_PENDING; 4990 ret |= WORK_BUSY_PENDING;
6004 4991
6005 rcu_read_lock(); 4992 rcu_read_lock();
6006 pool = get_work_pool(work); 4993 pool = get_work_pool(work);
6007 if (pool) { 4994 if (pool) {
6008 raw_spin_lock_irqsave(&pool-> !! 4995 raw_spin_lock_irqsave(&pool->lock, flags);
6009 if (find_worker_executing_wor 4996 if (find_worker_executing_work(pool, work))
6010 ret |= WORK_BUSY_RUNN 4997 ret |= WORK_BUSY_RUNNING;
6011 raw_spin_unlock_irqrestore(&p !! 4998 raw_spin_unlock_irqrestore(&pool->lock, flags);
6012 } 4999 }
6013 rcu_read_unlock(); 5000 rcu_read_unlock();
6014 5001
6015 return ret; 5002 return ret;
6016 } 5003 }
6017 EXPORT_SYMBOL_GPL(work_busy); 5004 EXPORT_SYMBOL_GPL(work_busy);
6018 5005
6019 /** 5006 /**
6020 * set_worker_desc - set description for the 5007 * set_worker_desc - set description for the current work item
6021 * @fmt: printf-style format string 5008 * @fmt: printf-style format string
6022 * @...: arguments for the format string 5009 * @...: arguments for the format string
6023 * 5010 *
6024 * This function can be called by a running w 5011 * This function can be called by a running work function to describe what
6025 * the work item is about. If the worker tas 5012 * the work item is about. If the worker task gets dumped, this
6026 * information will be printed out together t 5013 * information will be printed out together to help debugging. The
6027 * description can be at most WORKER_DESC_LEN 5014 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
6028 */ 5015 */
6029 void set_worker_desc(const char *fmt, ...) 5016 void set_worker_desc(const char *fmt, ...)
6030 { 5017 {
6031 struct worker *worker = current_wq_wo 5018 struct worker *worker = current_wq_worker();
6032 va_list args; 5019 va_list args;
6033 5020
6034 if (worker) { 5021 if (worker) {
6035 va_start(args, fmt); 5022 va_start(args, fmt);
6036 vsnprintf(worker->desc, sizeo 5023 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
6037 va_end(args); 5024 va_end(args);
6038 } 5025 }
6039 } 5026 }
6040 EXPORT_SYMBOL_GPL(set_worker_desc); 5027 EXPORT_SYMBOL_GPL(set_worker_desc);
6041 5028
6042 /** 5029 /**
6043 * print_worker_info - print out worker infor 5030 * print_worker_info - print out worker information and description
6044 * @log_lvl: the log level to use when printi 5031 * @log_lvl: the log level to use when printing
6045 * @task: target task 5032 * @task: target task
6046 * 5033 *
6047 * If @task is a worker and currently executi 5034 * If @task is a worker and currently executing a work item, print out the
6048 * name of the workqueue being serviced and w 5035 * name of the workqueue being serviced and worker description set with
6049 * set_worker_desc() by the currently executi 5036 * set_worker_desc() by the currently executing work item.
6050 * 5037 *
6051 * This function can be safely called on any 5038 * This function can be safely called on any task as long as the
6052 * task_struct itself is accessible. While s 5039 * task_struct itself is accessible. While safe, this function isn't
6053 * synchronized and may print out mixups or g 5040 * synchronized and may print out mixups or garbages of limited length.
6054 */ 5041 */
6055 void print_worker_info(const char *log_lvl, s 5042 void print_worker_info(const char *log_lvl, struct task_struct *task)
6056 { 5043 {
6057 work_func_t *fn = NULL; 5044 work_func_t *fn = NULL;
6058 char name[WQ_NAME_LEN] = { }; 5045 char name[WQ_NAME_LEN] = { };
6059 char desc[WORKER_DESC_LEN] = { }; 5046 char desc[WORKER_DESC_LEN] = { };
6060 struct pool_workqueue *pwq = NULL; 5047 struct pool_workqueue *pwq = NULL;
6061 struct workqueue_struct *wq = NULL; 5048 struct workqueue_struct *wq = NULL;
6062 struct worker *worker; 5049 struct worker *worker;
6063 5050
6064 if (!(task->flags & PF_WQ_WORKER)) 5051 if (!(task->flags & PF_WQ_WORKER))
6065 return; 5052 return;
6066 5053
6067 /* 5054 /*
6068 * This function is called without an 5055 * This function is called without any synchronization and @task
6069 * could be in any state. Be careful 5056 * could be in any state. Be careful with dereferences.
6070 */ 5057 */
6071 worker = kthread_probe_data(task); 5058 worker = kthread_probe_data(task);
6072 5059
6073 /* 5060 /*
6074 * Carefully copy the associated work 5061 * Carefully copy the associated workqueue's workfn, name and desc.
6075 * Keep the original last '\0' in cas 5062 * Keep the original last '\0' in case the original is garbage.
6076 */ 5063 */
6077 copy_from_kernel_nofault(&fn, &worker 5064 copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn));
6078 copy_from_kernel_nofault(&pwq, &worke 5065 copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq));
6079 copy_from_kernel_nofault(&wq, &pwq->w 5066 copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq));
6080 copy_from_kernel_nofault(name, wq->na 5067 copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1);
6081 copy_from_kernel_nofault(desc, worker 5068 copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1);
6082 5069
6083 if (fn || name[0] || desc[0]) { 5070 if (fn || name[0] || desc[0]) {
6084 printk("%sWorkqueue: %s %ps", 5071 printk("%sWorkqueue: %s %ps", log_lvl, name, fn);
6085 if (strcmp(name, desc)) 5072 if (strcmp(name, desc))
6086 pr_cont(" (%s)", desc 5073 pr_cont(" (%s)", desc);
6087 pr_cont("\n"); 5074 pr_cont("\n");
6088 } 5075 }
6089 } 5076 }
6090 5077
6091 static void pr_cont_pool_info(struct worker_p 5078 static void pr_cont_pool_info(struct worker_pool *pool)
6092 { 5079 {
6093 pr_cont(" cpus=%*pbl", nr_cpumask_bit 5080 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
6094 if (pool->node != NUMA_NO_NODE) 5081 if (pool->node != NUMA_NO_NODE)
6095 pr_cont(" node=%d", pool->nod 5082 pr_cont(" node=%d", pool->node);
6096 pr_cont(" flags=0x%x", pool->flags); !! 5083 pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
6097 if (pool->flags & POOL_BH) <<
6098 pr_cont(" bh%s", <<
6099 pool->attrs->nice == <<
6100 else <<
6101 pr_cont(" nice=%d", pool->att <<
6102 } <<
6103 <<
6104 static void pr_cont_worker_id(struct worker * <<
6105 { <<
6106 struct worker_pool *pool = worker->po <<
6107 <<
6108 if (pool->flags & WQ_BH) <<
6109 pr_cont("bh%s", <<
6110 pool->attrs->nice == <<
6111 else <<
6112 pr_cont("%d%s", task_pid_nr(w <<
6113 worker->rescue_wq ? " <<
6114 } 5084 }
6115 5085
6116 struct pr_cont_work_struct { 5086 struct pr_cont_work_struct {
6117 bool comma; 5087 bool comma;
6118 work_func_t func; 5088 work_func_t func;
6119 long ctr; 5089 long ctr;
6120 }; 5090 };
6121 5091
6122 static void pr_cont_work_flush(bool comma, wo 5092 static void pr_cont_work_flush(bool comma, work_func_t func, struct pr_cont_work_struct *pcwsp)
6123 { 5093 {
6124 if (!pcwsp->ctr) 5094 if (!pcwsp->ctr)
6125 goto out_record; 5095 goto out_record;
6126 if (func == pcwsp->func) { 5096 if (func == pcwsp->func) {
6127 pcwsp->ctr++; 5097 pcwsp->ctr++;
6128 return; 5098 return;
6129 } 5099 }
6130 if (pcwsp->ctr == 1) 5100 if (pcwsp->ctr == 1)
6131 pr_cont("%s %ps", pcwsp->comm 5101 pr_cont("%s %ps", pcwsp->comma ? "," : "", pcwsp->func);
6132 else 5102 else
6133 pr_cont("%s %ld*%ps", pcwsp-> 5103 pr_cont("%s %ld*%ps", pcwsp->comma ? "," : "", pcwsp->ctr, pcwsp->func);
6134 pcwsp->ctr = 0; 5104 pcwsp->ctr = 0;
6135 out_record: 5105 out_record:
6136 if ((long)func == -1L) 5106 if ((long)func == -1L)
6137 return; 5107 return;
6138 pcwsp->comma = comma; 5108 pcwsp->comma = comma;
6139 pcwsp->func = func; 5109 pcwsp->func = func;
6140 pcwsp->ctr = 1; 5110 pcwsp->ctr = 1;
6141 } 5111 }
6142 5112
6143 static void pr_cont_work(bool comma, struct w 5113 static void pr_cont_work(bool comma, struct work_struct *work, struct pr_cont_work_struct *pcwsp)
6144 { 5114 {
6145 if (work->func == wq_barrier_func) { 5115 if (work->func == wq_barrier_func) {
6146 struct wq_barrier *barr; 5116 struct wq_barrier *barr;
6147 5117
6148 barr = container_of(work, str 5118 barr = container_of(work, struct wq_barrier, work);
6149 5119
6150 pr_cont_work_flush(comma, (wo 5120 pr_cont_work_flush(comma, (work_func_t)-1, pcwsp);
6151 pr_cont("%s BAR(%d)", comma ? 5121 pr_cont("%s BAR(%d)", comma ? "," : "",
6152 task_pid_nr(barr->tas 5122 task_pid_nr(barr->task));
6153 } else { 5123 } else {
6154 if (!comma) 5124 if (!comma)
6155 pr_cont_work_flush(co 5125 pr_cont_work_flush(comma, (work_func_t)-1, pcwsp);
6156 pr_cont_work_flush(comma, wor 5126 pr_cont_work_flush(comma, work->func, pcwsp);
6157 } 5127 }
6158 } 5128 }
6159 5129
6160 static void show_pwq(struct pool_workqueue *p 5130 static void show_pwq(struct pool_workqueue *pwq)
6161 { 5131 {
6162 struct pr_cont_work_struct pcws = { . 5132 struct pr_cont_work_struct pcws = { .ctr = 0, };
6163 struct worker_pool *pool = pwq->pool; 5133 struct worker_pool *pool = pwq->pool;
6164 struct work_struct *work; 5134 struct work_struct *work;
6165 struct worker *worker; 5135 struct worker *worker;
6166 bool has_in_flight = false, has_pendi 5136 bool has_in_flight = false, has_pending = false;
6167 int bkt; 5137 int bkt;
6168 5138
6169 pr_info(" pwq %d:", pool->id); 5139 pr_info(" pwq %d:", pool->id);
6170 pr_cont_pool_info(pool); 5140 pr_cont_pool_info(pool);
6171 5141
6172 pr_cont(" active=%d refcnt=%d%s\n", !! 5142 pr_cont(" active=%d/%d refcnt=%d%s\n",
6173 pwq->nr_active, pwq->refcnt, !! 5143 pwq->nr_active, pwq->max_active, pwq->refcnt,
6174 !list_empty(&pwq->mayday_node 5144 !list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
6175 5145
6176 hash_for_each(pool->busy_hash, bkt, w 5146 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
6177 if (worker->current_pwq == pw 5147 if (worker->current_pwq == pwq) {
6178 has_in_flight = true; 5148 has_in_flight = true;
6179 break; 5149 break;
6180 } 5150 }
6181 } 5151 }
6182 if (has_in_flight) { 5152 if (has_in_flight) {
6183 bool comma = false; 5153 bool comma = false;
6184 5154
6185 pr_info(" in-flight:"); 5155 pr_info(" in-flight:");
6186 hash_for_each(pool->busy_hash 5156 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
6187 if (worker->current_p 5157 if (worker->current_pwq != pwq)
6188 continue; 5158 continue;
6189 5159
6190 pr_cont(" %s", comma !! 5160 pr_cont("%s %d%s:%ps", comma ? "," : "",
6191 pr_cont_worker_id(wor !! 5161 task_pid_nr(worker->task),
6192 pr_cont(":%ps", worke !! 5162 worker->rescue_wq ? "(RESCUER)" : "",
>> 5163 worker->current_func);
6193 list_for_each_entry(w 5164 list_for_each_entry(work, &worker->scheduled, entry)
6194 pr_cont_work( 5165 pr_cont_work(false, work, &pcws);
6195 pr_cont_work_flush(co 5166 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
6196 comma = true; 5167 comma = true;
6197 } 5168 }
6198 pr_cont("\n"); 5169 pr_cont("\n");
6199 } 5170 }
6200 5171
6201 list_for_each_entry(work, &pool->work 5172 list_for_each_entry(work, &pool->worklist, entry) {
6202 if (get_work_pwq(work) == pwq 5173 if (get_work_pwq(work) == pwq) {
6203 has_pending = true; 5174 has_pending = true;
6204 break; 5175 break;
6205 } 5176 }
6206 } 5177 }
6207 if (has_pending) { 5178 if (has_pending) {
6208 bool comma = false; 5179 bool comma = false;
6209 5180
6210 pr_info(" pending:"); 5181 pr_info(" pending:");
6211 list_for_each_entry(work, &po 5182 list_for_each_entry(work, &pool->worklist, entry) {
6212 if (get_work_pwq(work 5183 if (get_work_pwq(work) != pwq)
6213 continue; 5184 continue;
6214 5185
6215 pr_cont_work(comma, w 5186 pr_cont_work(comma, work, &pcws);
6216 comma = !(*work_data_ 5187 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
6217 } 5188 }
6218 pr_cont_work_flush(comma, (wo 5189 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
6219 pr_cont("\n"); 5190 pr_cont("\n");
6220 } 5191 }
6221 5192
6222 if (!list_empty(&pwq->inactive_works) 5193 if (!list_empty(&pwq->inactive_works)) {
6223 bool comma = false; 5194 bool comma = false;
6224 5195
6225 pr_info(" inactive:"); 5196 pr_info(" inactive:");
6226 list_for_each_entry(work, &pw 5197 list_for_each_entry(work, &pwq->inactive_works, entry) {
6227 pr_cont_work(comma, w 5198 pr_cont_work(comma, work, &pcws);
6228 comma = !(*work_data_ 5199 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
6229 } 5200 }
6230 pr_cont_work_flush(comma, (wo 5201 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
6231 pr_cont("\n"); 5202 pr_cont("\n");
6232 } 5203 }
6233 } 5204 }
6234 5205
6235 /** 5206 /**
6236 * show_one_workqueue - dump state of specifi 5207 * show_one_workqueue - dump state of specified workqueue
6237 * @wq: workqueue whose state will be printed 5208 * @wq: workqueue whose state will be printed
6238 */ 5209 */
6239 void show_one_workqueue(struct workqueue_stru 5210 void show_one_workqueue(struct workqueue_struct *wq)
6240 { 5211 {
6241 struct pool_workqueue *pwq; 5212 struct pool_workqueue *pwq;
6242 bool idle = true; 5213 bool idle = true;
6243 unsigned long irq_flags; !! 5214 unsigned long flags;
6244 5215
6245 for_each_pwq(pwq, wq) { 5216 for_each_pwq(pwq, wq) {
6246 if (!pwq_is_empty(pwq)) { !! 5217 if (pwq->nr_active || !list_empty(&pwq->inactive_works)) {
6247 idle = false; 5218 idle = false;
6248 break; 5219 break;
6249 } 5220 }
6250 } 5221 }
6251 if (idle) /* Nothing to print for idl 5222 if (idle) /* Nothing to print for idle workqueue */
6252 return; 5223 return;
6253 5224
6254 pr_info("workqueue %s: flags=0x%x\n", 5225 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
6255 5226
6256 for_each_pwq(pwq, wq) { 5227 for_each_pwq(pwq, wq) {
6257 raw_spin_lock_irqsave(&pwq->p !! 5228 raw_spin_lock_irqsave(&pwq->pool->lock, flags);
6258 if (!pwq_is_empty(pwq)) { !! 5229 if (pwq->nr_active || !list_empty(&pwq->inactive_works)) {
6259 /* 5230 /*
6260 * Defer printing to 5231 * Defer printing to avoid deadlocks in console
6261 * drivers that queue 5232 * drivers that queue work while holding locks
6262 * also taken in thei 5233 * also taken in their write paths.
6263 */ 5234 */
6264 printk_deferred_enter 5235 printk_deferred_enter();
6265 show_pwq(pwq); 5236 show_pwq(pwq);
6266 printk_deferred_exit( 5237 printk_deferred_exit();
6267 } 5238 }
6268 raw_spin_unlock_irqrestore(&p !! 5239 raw_spin_unlock_irqrestore(&pwq->pool->lock, flags);
6269 /* 5240 /*
6270 * We could be printing a lot 5241 * We could be printing a lot from atomic context, e.g.
6271 * sysrq-t -> show_all_workqu 5242 * sysrq-t -> show_all_workqueues(). Avoid triggering
6272 * hard lockup. 5243 * hard lockup.
6273 */ 5244 */
6274 touch_nmi_watchdog(); 5245 touch_nmi_watchdog();
6275 } 5246 }
6276 5247
6277 } 5248 }
6278 5249
6279 /** 5250 /**
6280 * show_one_worker_pool - dump state of speci 5251 * show_one_worker_pool - dump state of specified worker pool
6281 * @pool: worker pool whose state will be pri 5252 * @pool: worker pool whose state will be printed
6282 */ 5253 */
6283 static void show_one_worker_pool(struct worke 5254 static void show_one_worker_pool(struct worker_pool *pool)
6284 { 5255 {
6285 struct worker *worker; 5256 struct worker *worker;
6286 bool first = true; 5257 bool first = true;
6287 unsigned long irq_flags; !! 5258 unsigned long flags;
6288 unsigned long hung = 0; 5259 unsigned long hung = 0;
6289 5260
6290 raw_spin_lock_irqsave(&pool->lock, ir !! 5261 raw_spin_lock_irqsave(&pool->lock, flags);
6291 if (pool->nr_workers == pool->nr_idle 5262 if (pool->nr_workers == pool->nr_idle)
6292 goto next_pool; 5263 goto next_pool;
6293 5264
6294 /* How long the first pending work is 5265 /* How long the first pending work is waiting for a worker. */
6295 if (!list_empty(&pool->worklist)) 5266 if (!list_empty(&pool->worklist))
6296 hung = jiffies_to_msecs(jiffi 5267 hung = jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000;
6297 5268
6298 /* 5269 /*
6299 * Defer printing to avoid deadlocks 5270 * Defer printing to avoid deadlocks in console drivers that
6300 * queue work while holding locks als 5271 * queue work while holding locks also taken in their write
6301 * paths. 5272 * paths.
6302 */ 5273 */
6303 printk_deferred_enter(); 5274 printk_deferred_enter();
6304 pr_info("pool %d:", pool->id); 5275 pr_info("pool %d:", pool->id);
6305 pr_cont_pool_info(pool); 5276 pr_cont_pool_info(pool);
6306 pr_cont(" hung=%lus workers=%d", hung 5277 pr_cont(" hung=%lus workers=%d", hung, pool->nr_workers);
6307 if (pool->manager) 5278 if (pool->manager)
6308 pr_cont(" manager: %d", 5279 pr_cont(" manager: %d",
6309 task_pid_nr(pool->man 5280 task_pid_nr(pool->manager->task));
6310 list_for_each_entry(worker, &pool->id 5281 list_for_each_entry(worker, &pool->idle_list, entry) {
6311 pr_cont(" %s", first ? "idle: !! 5282 pr_cont(" %s%d", first ? "idle: " : "",
6312 pr_cont_worker_id(worker); !! 5283 task_pid_nr(worker->task));
6313 first = false; 5284 first = false;
6314 } 5285 }
6315 pr_cont("\n"); 5286 pr_cont("\n");
6316 printk_deferred_exit(); 5287 printk_deferred_exit();
6317 next_pool: 5288 next_pool:
6318 raw_spin_unlock_irqrestore(&pool->loc !! 5289 raw_spin_unlock_irqrestore(&pool->lock, flags);
6319 /* 5290 /*
6320 * We could be printing a lot from at 5291 * We could be printing a lot from atomic context, e.g.
6321 * sysrq-t -> show_all_workqueues(). 5292 * sysrq-t -> show_all_workqueues(). Avoid triggering
6322 * hard lockup. 5293 * hard lockup.
6323 */ 5294 */
6324 touch_nmi_watchdog(); 5295 touch_nmi_watchdog();
6325 5296
6326 } 5297 }
6327 5298
6328 /** 5299 /**
6329 * show_all_workqueues - dump workqueue state 5300 * show_all_workqueues - dump workqueue state
6330 * 5301 *
6331 * Called from a sysrq handler and prints out 5302 * Called from a sysrq handler and prints out all busy workqueues and pools.
6332 */ 5303 */
6333 void show_all_workqueues(void) 5304 void show_all_workqueues(void)
6334 { 5305 {
6335 struct workqueue_struct *wq; 5306 struct workqueue_struct *wq;
6336 struct worker_pool *pool; 5307 struct worker_pool *pool;
6337 int pi; 5308 int pi;
6338 5309
6339 rcu_read_lock(); 5310 rcu_read_lock();
6340 5311
6341 pr_info("Showing busy workqueues and 5312 pr_info("Showing busy workqueues and worker pools:\n");
6342 5313
6343 list_for_each_entry_rcu(wq, &workqueu 5314 list_for_each_entry_rcu(wq, &workqueues, list)
6344 show_one_workqueue(wq); 5315 show_one_workqueue(wq);
6345 5316
6346 for_each_pool(pool, pi) 5317 for_each_pool(pool, pi)
6347 show_one_worker_pool(pool); 5318 show_one_worker_pool(pool);
6348 5319
6349 rcu_read_unlock(); 5320 rcu_read_unlock();
6350 } 5321 }
6351 5322
6352 /** 5323 /**
6353 * show_freezable_workqueues - dump freezable 5324 * show_freezable_workqueues - dump freezable workqueue state
6354 * 5325 *
6355 * Called from try_to_freeze_tasks() and prin 5326 * Called from try_to_freeze_tasks() and prints out all freezable workqueues
6356 * still busy. 5327 * still busy.
6357 */ 5328 */
6358 void show_freezable_workqueues(void) 5329 void show_freezable_workqueues(void)
6359 { 5330 {
6360 struct workqueue_struct *wq; 5331 struct workqueue_struct *wq;
6361 5332
6362 rcu_read_lock(); 5333 rcu_read_lock();
6363 5334
6364 pr_info("Showing freezable workqueues 5335 pr_info("Showing freezable workqueues that are still busy:\n");
6365 5336
6366 list_for_each_entry_rcu(wq, &workqueu 5337 list_for_each_entry_rcu(wq, &workqueues, list) {
6367 if (!(wq->flags & WQ_FREEZABL 5338 if (!(wq->flags & WQ_FREEZABLE))
6368 continue; 5339 continue;
6369 show_one_workqueue(wq); 5340 show_one_workqueue(wq);
6370 } 5341 }
6371 5342
6372 rcu_read_unlock(); 5343 rcu_read_unlock();
6373 } 5344 }
6374 5345
6375 /* used to show worker information through /p 5346 /* used to show worker information through /proc/PID/{comm,stat,status} */
6376 void wq_worker_comm(char *buf, size_t size, s 5347 void wq_worker_comm(char *buf, size_t size, struct task_struct *task)
6377 { 5348 {
>> 5349 int off;
>> 5350
>> 5351 /* always show the actual comm */
>> 5352 off = strscpy(buf, task->comm, size);
>> 5353 if (off < 0)
>> 5354 return;
>> 5355
6378 /* stabilize PF_WQ_WORKER and worker 5356 /* stabilize PF_WQ_WORKER and worker pool association */
6379 mutex_lock(&wq_pool_attach_mutex); 5357 mutex_lock(&wq_pool_attach_mutex);
6380 5358
6381 if (task->flags & PF_WQ_WORKER) { 5359 if (task->flags & PF_WQ_WORKER) {
6382 struct worker *worker = kthre 5360 struct worker *worker = kthread_data(task);
6383 struct worker_pool *pool = wo 5361 struct worker_pool *pool = worker->pool;
6384 int off; <<
6385 <<
6386 off = format_worker_id(buf, s <<
6387 5362
6388 if (pool) { 5363 if (pool) {
6389 raw_spin_lock_irq(&po 5364 raw_spin_lock_irq(&pool->lock);
6390 /* 5365 /*
6391 * ->desc tracks info 5366 * ->desc tracks information (wq name or
6392 * set_worker_desc()) 5367 * set_worker_desc()) for the latest execution. If
6393 * current, prepend ' 5368 * current, prepend '+', otherwise '-'.
6394 */ 5369 */
6395 if (worker->desc[0] ! 5370 if (worker->desc[0] != '\0') {
6396 if (worker->c 5371 if (worker->current_work)
6397 scnpr 5372 scnprintf(buf + off, size - off, "+%s",
6398 5373 worker->desc);
6399 else 5374 else
6400 scnpr 5375 scnprintf(buf + off, size - off, "-%s",
6401 5376 worker->desc);
6402 } 5377 }
6403 raw_spin_unlock_irq(& 5378 raw_spin_unlock_irq(&pool->lock);
6404 } 5379 }
6405 } else { <<
6406 strscpy(buf, task->comm, size <<
6407 } 5380 }
6408 5381
6409 mutex_unlock(&wq_pool_attach_mutex); 5382 mutex_unlock(&wq_pool_attach_mutex);
6410 } 5383 }
6411 5384
6412 #ifdef CONFIG_SMP 5385 #ifdef CONFIG_SMP
6413 5386
6414 /* 5387 /*
6415 * CPU hotplug. 5388 * CPU hotplug.
6416 * 5389 *
6417 * There are two challenges in supporting CPU 5390 * There are two challenges in supporting CPU hotplug. Firstly, there
6418 * are a lot of assumptions on strong associa 5391 * are a lot of assumptions on strong associations among work, pwq and
6419 * pool which make migrating pending and sche 5392 * pool which make migrating pending and scheduled works very
6420 * difficult to implement without impacting h 5393 * difficult to implement without impacting hot paths. Secondly,
6421 * worker pools serve mix of short, long and 5394 * worker pools serve mix of short, long and very long running works making
6422 * blocked draining impractical. 5395 * blocked draining impractical.
6423 * 5396 *
6424 * This is solved by allowing the pools to be 5397 * This is solved by allowing the pools to be disassociated from the CPU
6425 * running as an unbound one and allowing it 5398 * running as an unbound one and allowing it to be reattached later if the
6426 * cpu comes back online. 5399 * cpu comes back online.
6427 */ 5400 */
6428 5401
6429 static void unbind_workers(int cpu) 5402 static void unbind_workers(int cpu)
6430 { 5403 {
6431 struct worker_pool *pool; 5404 struct worker_pool *pool;
6432 struct worker *worker; 5405 struct worker *worker;
6433 5406
6434 for_each_cpu_worker_pool(pool, cpu) { 5407 for_each_cpu_worker_pool(pool, cpu) {
6435 mutex_lock(&wq_pool_attach_mu 5408 mutex_lock(&wq_pool_attach_mutex);
6436 raw_spin_lock_irq(&pool->lock 5409 raw_spin_lock_irq(&pool->lock);
6437 5410
6438 /* 5411 /*
6439 * We've blocked all attach/d 5412 * We've blocked all attach/detach operations. Make all workers
6440 * unbound and set DISASSOCIA 5413 * unbound and set DISASSOCIATED. Before this, all workers
6441 * must be on the cpu. After 5414 * must be on the cpu. After this, they may become diasporas.
6442 * And the preemption disable 5415 * And the preemption disabled section in their sched callbacks
6443 * are guaranteed to see WORK 5416 * are guaranteed to see WORKER_UNBOUND since the code here
6444 * is on the same cpu. 5417 * is on the same cpu.
6445 */ 5418 */
6446 for_each_pool_worker(worker, 5419 for_each_pool_worker(worker, pool)
6447 worker->flags |= WORK 5420 worker->flags |= WORKER_UNBOUND;
6448 5421
6449 pool->flags |= POOL_DISASSOCI 5422 pool->flags |= POOL_DISASSOCIATED;
6450 5423
6451 /* 5424 /*
6452 * The handling of nr_running 5425 * The handling of nr_running in sched callbacks are disabled
6453 * now. Zap nr_running. Aft 5426 * now. Zap nr_running. After this, nr_running stays zero and
6454 * need_more_worker() and kee 5427 * need_more_worker() and keep_working() are always true as
6455 * long as the worklist is no 5428 * long as the worklist is not empty. This pool now behaves as
6456 * an unbound (in terms of co 5429 * an unbound (in terms of concurrency management) pool which
6457 * are served by workers tied 5430 * are served by workers tied to the pool.
6458 */ 5431 */
6459 pool->nr_running = 0; 5432 pool->nr_running = 0;
6460 5433
6461 /* 5434 /*
6462 * With concurrency managemen 5435 * With concurrency management just turned off, a busy
6463 * worker blocking could lead 5436 * worker blocking could lead to lengthy stalls. Kick off
6464 * unbound chain execution of 5437 * unbound chain execution of currently pending work items.
6465 */ 5438 */
6466 kick_pool(pool); 5439 kick_pool(pool);
6467 5440
6468 raw_spin_unlock_irq(&pool->lo 5441 raw_spin_unlock_irq(&pool->lock);
6469 5442
6470 for_each_pool_worker(worker, 5443 for_each_pool_worker(worker, pool)
6471 unbind_worker(worker) 5444 unbind_worker(worker);
6472 5445
6473 mutex_unlock(&wq_pool_attach_ 5446 mutex_unlock(&wq_pool_attach_mutex);
6474 } 5447 }
6475 } 5448 }
6476 5449
6477 /** 5450 /**
6478 * rebind_workers - rebind all workers of a p 5451 * rebind_workers - rebind all workers of a pool to the associated CPU
6479 * @pool: pool of interest 5452 * @pool: pool of interest
6480 * 5453 *
6481 * @pool->cpu is coming online. Rebind all w 5454 * @pool->cpu is coming online. Rebind all workers to the CPU.
6482 */ 5455 */
6483 static void rebind_workers(struct worker_pool 5456 static void rebind_workers(struct worker_pool *pool)
6484 { 5457 {
6485 struct worker *worker; 5458 struct worker *worker;
6486 5459
6487 lockdep_assert_held(&wq_pool_attach_m 5460 lockdep_assert_held(&wq_pool_attach_mutex);
6488 5461
6489 /* 5462 /*
6490 * Restore CPU affinity of all worker 5463 * Restore CPU affinity of all workers. As all idle workers should
6491 * be on the run-queue of the associa 5464 * be on the run-queue of the associated CPU before any local
6492 * wake-ups for concurrency managemen 5465 * wake-ups for concurrency management happen, restore CPU affinity
6493 * of all workers first and then clea 5466 * of all workers first and then clear UNBOUND. As we're called
6494 * from CPU_ONLINE, the following sho 5467 * from CPU_ONLINE, the following shouldn't fail.
6495 */ 5468 */
6496 for_each_pool_worker(worker, pool) { 5469 for_each_pool_worker(worker, pool) {
6497 kthread_set_per_cpu(worker->t 5470 kthread_set_per_cpu(worker->task, pool->cpu);
6498 WARN_ON_ONCE(set_cpus_allowed 5471 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
6499 5472 pool_allowed_cpus(pool)) < 0);
6500 } 5473 }
6501 5474
6502 raw_spin_lock_irq(&pool->lock); 5475 raw_spin_lock_irq(&pool->lock);
6503 5476
6504 pool->flags &= ~POOL_DISASSOCIATED; 5477 pool->flags &= ~POOL_DISASSOCIATED;
6505 5478
6506 for_each_pool_worker(worker, pool) { 5479 for_each_pool_worker(worker, pool) {
6507 unsigned int worker_flags = w 5480 unsigned int worker_flags = worker->flags;
6508 5481
6509 /* 5482 /*
6510 * We want to clear UNBOUND b 5483 * We want to clear UNBOUND but can't directly call
6511 * worker_clr_flags() or adju 5484 * worker_clr_flags() or adjust nr_running. Atomically
6512 * replace UNBOUND with anoth 5485 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
6513 * @worker will clear REBOUND 5486 * @worker will clear REBOUND using worker_clr_flags() when
6514 * it initiates the next exec 5487 * it initiates the next execution cycle thus restoring
6515 * concurrency management. N 5488 * concurrency management. Note that when or whether
6516 * @worker clears REBOUND doe 5489 * @worker clears REBOUND doesn't affect correctness.
6517 * 5490 *
6518 * WRITE_ONCE() is necessary 5491 * WRITE_ONCE() is necessary because @worker->flags may be
6519 * tested without holding any 5492 * tested without holding any lock in
6520 * wq_worker_running(). With 5493 * wq_worker_running(). Without it, NOT_RUNNING test may
6521 * fail incorrectly leading t 5494 * fail incorrectly leading to premature concurrency
6522 * management operations. 5495 * management operations.
6523 */ 5496 */
6524 WARN_ON_ONCE(!(worker_flags & 5497 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
6525 worker_flags |= WORKER_REBOUN 5498 worker_flags |= WORKER_REBOUND;
6526 worker_flags &= ~WORKER_UNBOU 5499 worker_flags &= ~WORKER_UNBOUND;
6527 WRITE_ONCE(worker->flags, wor 5500 WRITE_ONCE(worker->flags, worker_flags);
6528 } 5501 }
6529 5502
6530 raw_spin_unlock_irq(&pool->lock); 5503 raw_spin_unlock_irq(&pool->lock);
6531 } 5504 }
6532 5505
6533 /** 5506 /**
6534 * restore_unbound_workers_cpumask - restore 5507 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
6535 * @pool: unbound pool of interest 5508 * @pool: unbound pool of interest
6536 * @cpu: the CPU which is coming up 5509 * @cpu: the CPU which is coming up
6537 * 5510 *
6538 * An unbound pool may end up with a cpumask 5511 * An unbound pool may end up with a cpumask which doesn't have any online
6539 * CPUs. When a worker of such pool get sche 5512 * CPUs. When a worker of such pool get scheduled, the scheduler resets
6540 * its cpus_allowed. If @cpu is in @pool's c 5513 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
6541 * online CPU before, cpus_allowed of all its 5514 * online CPU before, cpus_allowed of all its workers should be restored.
6542 */ 5515 */
6543 static void restore_unbound_workers_cpumask(s 5516 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
6544 { 5517 {
6545 static cpumask_t cpumask; 5518 static cpumask_t cpumask;
6546 struct worker *worker; 5519 struct worker *worker;
6547 5520
6548 lockdep_assert_held(&wq_pool_attach_m 5521 lockdep_assert_held(&wq_pool_attach_mutex);
6549 5522
6550 /* is @cpu allowed for @pool? */ 5523 /* is @cpu allowed for @pool? */
6551 if (!cpumask_test_cpu(cpu, pool->attr 5524 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
6552 return; 5525 return;
6553 5526
6554 cpumask_and(&cpumask, pool->attrs->cp 5527 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
6555 5528
6556 /* as we're called from CPU_ONLINE, t 5529 /* as we're called from CPU_ONLINE, the following shouldn't fail */
6557 for_each_pool_worker(worker, pool) 5530 for_each_pool_worker(worker, pool)
6558 WARN_ON_ONCE(set_cpus_allowed 5531 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0);
6559 } 5532 }
6560 5533
6561 int workqueue_prepare_cpu(unsigned int cpu) 5534 int workqueue_prepare_cpu(unsigned int cpu)
6562 { 5535 {
6563 struct worker_pool *pool; 5536 struct worker_pool *pool;
6564 5537
6565 for_each_cpu_worker_pool(pool, cpu) { 5538 for_each_cpu_worker_pool(pool, cpu) {
6566 if (pool->nr_workers) 5539 if (pool->nr_workers)
6567 continue; 5540 continue;
6568 if (!create_worker(pool)) 5541 if (!create_worker(pool))
6569 return -ENOMEM; 5542 return -ENOMEM;
6570 } 5543 }
6571 return 0; 5544 return 0;
6572 } 5545 }
6573 5546
6574 int workqueue_online_cpu(unsigned int cpu) 5547 int workqueue_online_cpu(unsigned int cpu)
6575 { 5548 {
6576 struct worker_pool *pool; 5549 struct worker_pool *pool;
6577 struct workqueue_struct *wq; 5550 struct workqueue_struct *wq;
6578 int pi; 5551 int pi;
6579 5552
6580 mutex_lock(&wq_pool_mutex); 5553 mutex_lock(&wq_pool_mutex);
6581 5554
6582 cpumask_set_cpu(cpu, wq_online_cpumas <<
6583 <<
6584 for_each_pool(pool, pi) { 5555 for_each_pool(pool, pi) {
6585 /* BH pools aren't affected b <<
6586 if (pool->flags & POOL_BH) <<
6587 continue; <<
6588 <<
6589 mutex_lock(&wq_pool_attach_mu 5556 mutex_lock(&wq_pool_attach_mutex);
>> 5557
6590 if (pool->cpu == cpu) 5558 if (pool->cpu == cpu)
6591 rebind_workers(pool); 5559 rebind_workers(pool);
6592 else if (pool->cpu < 0) 5560 else if (pool->cpu < 0)
6593 restore_unbound_worke 5561 restore_unbound_workers_cpumask(pool, cpu);
>> 5562
6594 mutex_unlock(&wq_pool_attach_ 5563 mutex_unlock(&wq_pool_attach_mutex);
6595 } 5564 }
6596 5565
6597 /* update pod affinity of unbound wor 5566 /* update pod affinity of unbound workqueues */
6598 list_for_each_entry(wq, &workqueues, 5567 list_for_each_entry(wq, &workqueues, list) {
6599 struct workqueue_attrs *attrs 5568 struct workqueue_attrs *attrs = wq->unbound_attrs;
6600 5569
6601 if (attrs) { 5570 if (attrs) {
6602 const struct wq_pod_t 5571 const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
6603 int tcpu; 5572 int tcpu;
6604 5573
6605 for_each_cpu(tcpu, pt 5574 for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]])
6606 unbound_wq_up !! 5575 wq_update_pod(wq, tcpu, cpu, true);
6607 <<
6608 mutex_lock(&wq->mutex <<
6609 wq_update_node_max_ac <<
6610 mutex_unlock(&wq->mut <<
6611 } 5576 }
6612 } 5577 }
6613 5578
6614 mutex_unlock(&wq_pool_mutex); 5579 mutex_unlock(&wq_pool_mutex);
6615 return 0; 5580 return 0;
6616 } 5581 }
6617 5582
6618 int workqueue_offline_cpu(unsigned int cpu) 5583 int workqueue_offline_cpu(unsigned int cpu)
6619 { 5584 {
6620 struct workqueue_struct *wq; 5585 struct workqueue_struct *wq;
6621 5586
6622 /* unbinding per-cpu workers should h 5587 /* unbinding per-cpu workers should happen on the local CPU */
6623 if (WARN_ON(cpu != smp_processor_id() 5588 if (WARN_ON(cpu != smp_processor_id()))
6624 return -1; 5589 return -1;
6625 5590
6626 unbind_workers(cpu); 5591 unbind_workers(cpu);
6627 5592
6628 /* update pod affinity of unbound wor 5593 /* update pod affinity of unbound workqueues */
6629 mutex_lock(&wq_pool_mutex); 5594 mutex_lock(&wq_pool_mutex);
6630 <<
6631 cpumask_clear_cpu(cpu, wq_online_cpum <<
6632 <<
6633 list_for_each_entry(wq, &workqueues, 5595 list_for_each_entry(wq, &workqueues, list) {
6634 struct workqueue_attrs *attrs 5596 struct workqueue_attrs *attrs = wq->unbound_attrs;
6635 5597
6636 if (attrs) { 5598 if (attrs) {
6637 const struct wq_pod_t 5599 const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
6638 int tcpu; 5600 int tcpu;
6639 5601
6640 for_each_cpu(tcpu, pt 5602 for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]])
6641 unbound_wq_up !! 5603 wq_update_pod(wq, tcpu, cpu, false);
6642 <<
6643 mutex_lock(&wq->mutex <<
6644 wq_update_node_max_ac <<
6645 mutex_unlock(&wq->mut <<
6646 } 5604 }
6647 } 5605 }
6648 mutex_unlock(&wq_pool_mutex); 5606 mutex_unlock(&wq_pool_mutex);
6649 5607
6650 return 0; 5608 return 0;
6651 } 5609 }
6652 5610
6653 struct work_for_cpu { 5611 struct work_for_cpu {
6654 struct work_struct work; 5612 struct work_struct work;
6655 long (*fn)(void *); 5613 long (*fn)(void *);
6656 void *arg; 5614 void *arg;
6657 long ret; 5615 long ret;
6658 }; 5616 };
6659 5617
6660 static void work_for_cpu_fn(struct work_struc 5618 static void work_for_cpu_fn(struct work_struct *work)
6661 { 5619 {
6662 struct work_for_cpu *wfc = container_ 5620 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
6663 5621
6664 wfc->ret = wfc->fn(wfc->arg); 5622 wfc->ret = wfc->fn(wfc->arg);
6665 } 5623 }
6666 5624
6667 /** 5625 /**
6668 * work_on_cpu_key - run a function in thread 5626 * work_on_cpu_key - run a function in thread context on a particular cpu
6669 * @cpu: the cpu to run on 5627 * @cpu: the cpu to run on
6670 * @fn: the function to run 5628 * @fn: the function to run
6671 * @arg: the function arg 5629 * @arg: the function arg
6672 * @key: The lock class key for lock debuggin 5630 * @key: The lock class key for lock debugging purposes
6673 * 5631 *
6674 * It is up to the caller to ensure that the 5632 * It is up to the caller to ensure that the cpu doesn't go offline.
6675 * The caller must not hold any locks which w 5633 * The caller must not hold any locks which would prevent @fn from completing.
6676 * 5634 *
6677 * Return: The value @fn returns. 5635 * Return: The value @fn returns.
6678 */ 5636 */
6679 long work_on_cpu_key(int cpu, long (*fn)(void 5637 long work_on_cpu_key(int cpu, long (*fn)(void *),
6680 void *arg, struct lock_c 5638 void *arg, struct lock_class_key *key)
6681 { 5639 {
6682 struct work_for_cpu wfc = { .fn = fn, 5640 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
6683 5641
6684 INIT_WORK_ONSTACK_KEY(&wfc.work, work 5642 INIT_WORK_ONSTACK_KEY(&wfc.work, work_for_cpu_fn, key);
6685 schedule_work_on(cpu, &wfc.work); 5643 schedule_work_on(cpu, &wfc.work);
6686 flush_work(&wfc.work); 5644 flush_work(&wfc.work);
6687 destroy_work_on_stack(&wfc.work); 5645 destroy_work_on_stack(&wfc.work);
6688 return wfc.ret; 5646 return wfc.ret;
6689 } 5647 }
6690 EXPORT_SYMBOL_GPL(work_on_cpu_key); 5648 EXPORT_SYMBOL_GPL(work_on_cpu_key);
6691 5649
6692 /** 5650 /**
6693 * work_on_cpu_safe_key - run a function in t 5651 * work_on_cpu_safe_key - run a function in thread context on a particular cpu
6694 * @cpu: the cpu to run on 5652 * @cpu: the cpu to run on
6695 * @fn: the function to run 5653 * @fn: the function to run
6696 * @arg: the function argument 5654 * @arg: the function argument
6697 * @key: The lock class key for lock debuggin 5655 * @key: The lock class key for lock debugging purposes
6698 * 5656 *
6699 * Disables CPU hotplug and calls work_on_cpu 5657 * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold
6700 * any locks which would prevent @fn from com 5658 * any locks which would prevent @fn from completing.
6701 * 5659 *
6702 * Return: The value @fn returns. 5660 * Return: The value @fn returns.
6703 */ 5661 */
6704 long work_on_cpu_safe_key(int cpu, long (*fn) 5662 long work_on_cpu_safe_key(int cpu, long (*fn)(void *),
6705 void *arg, struct l 5663 void *arg, struct lock_class_key *key)
6706 { 5664 {
6707 long ret = -ENODEV; 5665 long ret = -ENODEV;
6708 5666
6709 cpus_read_lock(); 5667 cpus_read_lock();
6710 if (cpu_online(cpu)) 5668 if (cpu_online(cpu))
6711 ret = work_on_cpu_key(cpu, fn 5669 ret = work_on_cpu_key(cpu, fn, arg, key);
6712 cpus_read_unlock(); 5670 cpus_read_unlock();
6713 return ret; 5671 return ret;
6714 } 5672 }
6715 EXPORT_SYMBOL_GPL(work_on_cpu_safe_key); 5673 EXPORT_SYMBOL_GPL(work_on_cpu_safe_key);
6716 #endif /* CONFIG_SMP */ 5674 #endif /* CONFIG_SMP */
6717 5675
6718 #ifdef CONFIG_FREEZER 5676 #ifdef CONFIG_FREEZER
6719 5677
6720 /** 5678 /**
6721 * freeze_workqueues_begin - begin freezing w 5679 * freeze_workqueues_begin - begin freezing workqueues
6722 * 5680 *
6723 * Start freezing workqueues. After this fun 5681 * Start freezing workqueues. After this function returns, all freezable
6724 * workqueues will queue new works to their i 5682 * workqueues will queue new works to their inactive_works list instead of
6725 * pool->worklist. 5683 * pool->worklist.
6726 * 5684 *
6727 * CONTEXT: 5685 * CONTEXT:
6728 * Grabs and releases wq_pool_mutex, wq->mute 5686 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
6729 */ 5687 */
6730 void freeze_workqueues_begin(void) 5688 void freeze_workqueues_begin(void)
6731 { 5689 {
6732 struct workqueue_struct *wq; 5690 struct workqueue_struct *wq;
>> 5691 struct pool_workqueue *pwq;
6733 5692
6734 mutex_lock(&wq_pool_mutex); 5693 mutex_lock(&wq_pool_mutex);
6735 5694
6736 WARN_ON_ONCE(workqueue_freezing); 5695 WARN_ON_ONCE(workqueue_freezing);
6737 workqueue_freezing = true; 5696 workqueue_freezing = true;
6738 5697
6739 list_for_each_entry(wq, &workqueues, 5698 list_for_each_entry(wq, &workqueues, list) {
6740 mutex_lock(&wq->mutex); 5699 mutex_lock(&wq->mutex);
6741 wq_adjust_max_active(wq); !! 5700 for_each_pwq(pwq, wq)
>> 5701 pwq_adjust_max_active(pwq);
6742 mutex_unlock(&wq->mutex); 5702 mutex_unlock(&wq->mutex);
6743 } 5703 }
6744 5704
6745 mutex_unlock(&wq_pool_mutex); 5705 mutex_unlock(&wq_pool_mutex);
6746 } 5706 }
6747 5707
6748 /** 5708 /**
6749 * freeze_workqueues_busy - are freezable wor 5709 * freeze_workqueues_busy - are freezable workqueues still busy?
6750 * 5710 *
6751 * Check whether freezing is complete. This 5711 * Check whether freezing is complete. This function must be called
6752 * between freeze_workqueues_begin() and thaw 5712 * between freeze_workqueues_begin() and thaw_workqueues().
6753 * 5713 *
6754 * CONTEXT: 5714 * CONTEXT:
6755 * Grabs and releases wq_pool_mutex. 5715 * Grabs and releases wq_pool_mutex.
6756 * 5716 *
6757 * Return: 5717 * Return:
6758 * %true if some freezable workqueues are sti 5718 * %true if some freezable workqueues are still busy. %false if freezing
6759 * is complete. 5719 * is complete.
6760 */ 5720 */
6761 bool freeze_workqueues_busy(void) 5721 bool freeze_workqueues_busy(void)
6762 { 5722 {
6763 bool busy = false; 5723 bool busy = false;
6764 struct workqueue_struct *wq; 5724 struct workqueue_struct *wq;
6765 struct pool_workqueue *pwq; 5725 struct pool_workqueue *pwq;
6766 5726
6767 mutex_lock(&wq_pool_mutex); 5727 mutex_lock(&wq_pool_mutex);
6768 5728
6769 WARN_ON_ONCE(!workqueue_freezing); 5729 WARN_ON_ONCE(!workqueue_freezing);
6770 5730
6771 list_for_each_entry(wq, &workqueues, 5731 list_for_each_entry(wq, &workqueues, list) {
6772 if (!(wq->flags & WQ_FREEZABL 5732 if (!(wq->flags & WQ_FREEZABLE))
6773 continue; 5733 continue;
6774 /* 5734 /*
6775 * nr_active is monotonically 5735 * nr_active is monotonically decreasing. It's safe
6776 * to peek without lock. 5736 * to peek without lock.
6777 */ 5737 */
6778 rcu_read_lock(); 5738 rcu_read_lock();
6779 for_each_pwq(pwq, wq) { 5739 for_each_pwq(pwq, wq) {
6780 WARN_ON_ONCE(pwq->nr_ 5740 WARN_ON_ONCE(pwq->nr_active < 0);
6781 if (pwq->nr_active) { 5741 if (pwq->nr_active) {
6782 busy = true; 5742 busy = true;
6783 rcu_read_unlo 5743 rcu_read_unlock();
6784 goto out_unlo 5744 goto out_unlock;
6785 } 5745 }
6786 } 5746 }
6787 rcu_read_unlock(); 5747 rcu_read_unlock();
6788 } 5748 }
6789 out_unlock: 5749 out_unlock:
6790 mutex_unlock(&wq_pool_mutex); 5750 mutex_unlock(&wq_pool_mutex);
6791 return busy; 5751 return busy;
6792 } 5752 }
6793 5753
6794 /** 5754 /**
6795 * thaw_workqueues - thaw workqueues 5755 * thaw_workqueues - thaw workqueues
6796 * 5756 *
6797 * Thaw workqueues. Normal queueing is resto 5757 * Thaw workqueues. Normal queueing is restored and all collected
6798 * frozen works are transferred to their resp 5758 * frozen works are transferred to their respective pool worklists.
6799 * 5759 *
6800 * CONTEXT: 5760 * CONTEXT:
6801 * Grabs and releases wq_pool_mutex, wq->mute 5761 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
6802 */ 5762 */
6803 void thaw_workqueues(void) 5763 void thaw_workqueues(void)
6804 { 5764 {
6805 struct workqueue_struct *wq; 5765 struct workqueue_struct *wq;
>> 5766 struct pool_workqueue *pwq;
6806 5767
6807 mutex_lock(&wq_pool_mutex); 5768 mutex_lock(&wq_pool_mutex);
6808 5769
6809 if (!workqueue_freezing) 5770 if (!workqueue_freezing)
6810 goto out_unlock; 5771 goto out_unlock;
6811 5772
6812 workqueue_freezing = false; 5773 workqueue_freezing = false;
6813 5774
6814 /* restore max_active and repopulate 5775 /* restore max_active and repopulate worklist */
6815 list_for_each_entry(wq, &workqueues, 5776 list_for_each_entry(wq, &workqueues, list) {
6816 mutex_lock(&wq->mutex); 5777 mutex_lock(&wq->mutex);
6817 wq_adjust_max_active(wq); !! 5778 for_each_pwq(pwq, wq)
>> 5779 pwq_adjust_max_active(pwq);
6818 mutex_unlock(&wq->mutex); 5780 mutex_unlock(&wq->mutex);
6819 } 5781 }
6820 5782
6821 out_unlock: 5783 out_unlock:
6822 mutex_unlock(&wq_pool_mutex); 5784 mutex_unlock(&wq_pool_mutex);
6823 } 5785 }
6824 #endif /* CONFIG_FREEZER */ 5786 #endif /* CONFIG_FREEZER */
6825 5787
6826 static int workqueue_apply_unbound_cpumask(co 5788 static int workqueue_apply_unbound_cpumask(const cpumask_var_t unbound_cpumask)
6827 { 5789 {
6828 LIST_HEAD(ctxs); 5790 LIST_HEAD(ctxs);
6829 int ret = 0; 5791 int ret = 0;
6830 struct workqueue_struct *wq; 5792 struct workqueue_struct *wq;
6831 struct apply_wqattrs_ctx *ctx, *n; 5793 struct apply_wqattrs_ctx *ctx, *n;
6832 5794
6833 lockdep_assert_held(&wq_pool_mutex); 5795 lockdep_assert_held(&wq_pool_mutex);
6834 5796
6835 list_for_each_entry(wq, &workqueues, 5797 list_for_each_entry(wq, &workqueues, list) {
6836 if (!(wq->flags & WQ_UNBOUND) !! 5798 if (!(wq->flags & WQ_UNBOUND))
>> 5799 continue;
>> 5800 /* creating multiple pwqs breaks ordering guarantee */
>> 5801 if (wq->flags & __WQ_ORDERED)
6837 continue; 5802 continue;
6838 5803
6839 ctx = apply_wqattrs_prepare(w 5804 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs, unbound_cpumask);
6840 if (IS_ERR(ctx)) { 5805 if (IS_ERR(ctx)) {
6841 ret = PTR_ERR(ctx); 5806 ret = PTR_ERR(ctx);
6842 break; 5807 break;
6843 } 5808 }
6844 5809
6845 list_add_tail(&ctx->list, &ct 5810 list_add_tail(&ctx->list, &ctxs);
6846 } 5811 }
6847 5812
6848 list_for_each_entry_safe(ctx, n, &ctx 5813 list_for_each_entry_safe(ctx, n, &ctxs, list) {
6849 if (!ret) 5814 if (!ret)
6850 apply_wqattrs_commit( 5815 apply_wqattrs_commit(ctx);
6851 apply_wqattrs_cleanup(ctx); 5816 apply_wqattrs_cleanup(ctx);
6852 } 5817 }
6853 5818
6854 if (!ret) { 5819 if (!ret) {
6855 mutex_lock(&wq_pool_attach_mu 5820 mutex_lock(&wq_pool_attach_mutex);
6856 cpumask_copy(wq_unbound_cpuma 5821 cpumask_copy(wq_unbound_cpumask, unbound_cpumask);
6857 mutex_unlock(&wq_pool_attach_ 5822 mutex_unlock(&wq_pool_attach_mutex);
6858 } 5823 }
6859 return ret; 5824 return ret;
6860 } 5825 }
6861 5826
6862 /** 5827 /**
6863 * workqueue_unbound_exclude_cpumask - Exclud !! 5828 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
6864 * @exclude_cpumask: the cpumask to be exclud !! 5829 * @cpumask: the cpumask to set
6865 * 5830 *
6866 * This function can be called from cpuset co !! 5831 * The low-level workqueues cpumask is a global cpumask that limits
6867 * CPUs that should be excluded from wq_unbou !! 5832 * the affinity of all unbound workqueues. This function check the @cpumask
>> 5833 * and apply it to all unbound workqueues and updates all pwqs of them.
>> 5834 *
>> 5835 * Return: 0 - Success
>> 5836 * -EINVAL - Invalid @cpumask
>> 5837 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
6868 */ 5838 */
6869 int workqueue_unbound_exclude_cpumask(cpumask !! 5839 int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
6870 { 5840 {
6871 cpumask_var_t cpumask; !! 5841 int ret = -EINVAL;
6872 int ret = 0; <<
6873 <<
6874 if (!zalloc_cpumask_var(&cpumask, GFP <<
6875 return -ENOMEM; <<
6876 <<
6877 mutex_lock(&wq_pool_mutex); <<
6878 5842
6879 /* 5843 /*
6880 * If the operation fails, it will fa !! 5844 * Not excluding isolated cpus on purpose.
6881 * wq_requested_unbound_cpumask which !! 5845 * If the user wishes to include them, we allow that.
6882 * (HK_TYPE_WQ ∩ HK_TYPE_DOMAIN) ho !! 5846 */
6883 * by any subsequent write to workque !! 5847 cpumask_and(cpumask, cpumask, cpu_possible_mask);
6884 */ !! 5848 if (!cpumask_empty(cpumask)) {
6885 if (!cpumask_andnot(cpumask, wq_reque !! 5849 apply_wqattrs_lock();
6886 cpumask_copy(cpumask, wq_requ !! 5850 if (cpumask_equal(cpumask, wq_unbound_cpumask)) {
6887 if (!cpumask_equal(cpumask, wq_unboun !! 5851 ret = 0;
>> 5852 goto out_unlock;
>> 5853 }
>> 5854
6888 ret = workqueue_apply_unbound 5855 ret = workqueue_apply_unbound_cpumask(cpumask);
6889 5856
6890 /* Save the current isolated cpumask !! 5857 out_unlock:
6891 if (!ret) !! 5858 apply_wqattrs_unlock();
6892 cpumask_copy(wq_isolated_cpum !! 5859 }
6893 5860
6894 mutex_unlock(&wq_pool_mutex); <<
6895 free_cpumask_var(cpumask); <<
6896 return ret; 5861 return ret;
6897 } 5862 }
6898 5863
6899 static int parse_affn_scope(const char *val) 5864 static int parse_affn_scope(const char *val)
6900 { 5865 {
6901 int i; 5866 int i;
6902 5867
6903 for (i = 0; i < ARRAY_SIZE(wq_affn_na 5868 for (i = 0; i < ARRAY_SIZE(wq_affn_names); i++) {
6904 if (!strncasecmp(val, wq_affn 5869 if (!strncasecmp(val, wq_affn_names[i], strlen(wq_affn_names[i])))
6905 return i; 5870 return i;
6906 } 5871 }
6907 return -EINVAL; 5872 return -EINVAL;
6908 } 5873 }
6909 5874
6910 static int wq_affn_dfl_set(const char *val, c 5875 static int wq_affn_dfl_set(const char *val, const struct kernel_param *kp)
6911 { 5876 {
6912 struct workqueue_struct *wq; 5877 struct workqueue_struct *wq;
6913 int affn, cpu; 5878 int affn, cpu;
6914 5879
6915 affn = parse_affn_scope(val); 5880 affn = parse_affn_scope(val);
6916 if (affn < 0) 5881 if (affn < 0)
6917 return affn; 5882 return affn;
6918 if (affn == WQ_AFFN_DFL) 5883 if (affn == WQ_AFFN_DFL)
6919 return -EINVAL; 5884 return -EINVAL;
6920 5885
6921 cpus_read_lock(); 5886 cpus_read_lock();
6922 mutex_lock(&wq_pool_mutex); 5887 mutex_lock(&wq_pool_mutex);
6923 5888
6924 wq_affn_dfl = affn; 5889 wq_affn_dfl = affn;
6925 5890
6926 list_for_each_entry(wq, &workqueues, 5891 list_for_each_entry(wq, &workqueues, list) {
6927 for_each_online_cpu(cpu) !! 5892 for_each_online_cpu(cpu) {
6928 unbound_wq_update_pwq !! 5893 wq_update_pod(wq, cpu, cpu, true);
>> 5894 }
6929 } 5895 }
6930 5896
6931 mutex_unlock(&wq_pool_mutex); 5897 mutex_unlock(&wq_pool_mutex);
6932 cpus_read_unlock(); 5898 cpus_read_unlock();
6933 5899
6934 return 0; 5900 return 0;
6935 } 5901 }
6936 5902
6937 static int wq_affn_dfl_get(char *buffer, cons 5903 static int wq_affn_dfl_get(char *buffer, const struct kernel_param *kp)
6938 { 5904 {
6939 return scnprintf(buffer, PAGE_SIZE, " 5905 return scnprintf(buffer, PAGE_SIZE, "%s\n", wq_affn_names[wq_affn_dfl]);
6940 } 5906 }
6941 5907
6942 static const struct kernel_param_ops wq_affn_ 5908 static const struct kernel_param_ops wq_affn_dfl_ops = {
6943 .set = wq_affn_dfl_set, 5909 .set = wq_affn_dfl_set,
6944 .get = wq_affn_dfl_get, 5910 .get = wq_affn_dfl_get,
6945 }; 5911 };
6946 5912
6947 module_param_cb(default_affinity_scope, &wq_a 5913 module_param_cb(default_affinity_scope, &wq_affn_dfl_ops, NULL, 0644);
6948 5914
6949 #ifdef CONFIG_SYSFS 5915 #ifdef CONFIG_SYSFS
6950 /* 5916 /*
6951 * Workqueues with WQ_SYSFS flag set is visib 5917 * Workqueues with WQ_SYSFS flag set is visible to userland via
6952 * /sys/bus/workqueue/devices/WQ_NAME. All v 5918 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
6953 * following attributes. 5919 * following attributes.
6954 * 5920 *
6955 * per_cpu RO bool : whether the 5921 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
6956 * max_active RW int : maximum num 5922 * max_active RW int : maximum number of in-flight work items
6957 * 5923 *
6958 * Unbound workqueues have the following extr 5924 * Unbound workqueues have the following extra attributes.
6959 * 5925 *
6960 * nice RW int : nice value 5926 * nice RW int : nice value of the workers
6961 * cpumask RW mask : bitmask of 5927 * cpumask RW mask : bitmask of allowed CPUs for the workers
6962 * affinity_scope RW str : worker CPU 5928 * affinity_scope RW str : worker CPU affinity scope (cache, numa, none)
6963 * affinity_strict RW bool : worker CPU 5929 * affinity_strict RW bool : worker CPU affinity is strict
6964 */ 5930 */
6965 struct wq_device { 5931 struct wq_device {
6966 struct workqueue_struct *wq; 5932 struct workqueue_struct *wq;
6967 struct device dev; 5933 struct device dev;
6968 }; 5934 };
6969 5935
6970 static struct workqueue_struct *dev_to_wq(str 5936 static struct workqueue_struct *dev_to_wq(struct device *dev)
6971 { 5937 {
6972 struct wq_device *wq_dev = container_ 5938 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
6973 5939
6974 return wq_dev->wq; 5940 return wq_dev->wq;
6975 } 5941 }
6976 5942
6977 static ssize_t per_cpu_show(struct device *de 5943 static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
6978 char *buf) 5944 char *buf)
6979 { 5945 {
6980 struct workqueue_struct *wq = dev_to_ 5946 struct workqueue_struct *wq = dev_to_wq(dev);
6981 5947
6982 return scnprintf(buf, PAGE_SIZE, "%d\ 5948 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
6983 } 5949 }
6984 static DEVICE_ATTR_RO(per_cpu); 5950 static DEVICE_ATTR_RO(per_cpu);
6985 5951
6986 static ssize_t max_active_show(struct device 5952 static ssize_t max_active_show(struct device *dev,
6987 struct device_ 5953 struct device_attribute *attr, char *buf)
6988 { 5954 {
6989 struct workqueue_struct *wq = dev_to_ 5955 struct workqueue_struct *wq = dev_to_wq(dev);
6990 5956
6991 return scnprintf(buf, PAGE_SIZE, "%d\ 5957 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
6992 } 5958 }
6993 5959
6994 static ssize_t max_active_store(struct device 5960 static ssize_t max_active_store(struct device *dev,
6995 struct device 5961 struct device_attribute *attr, const char *buf,
6996 size_t count) 5962 size_t count)
6997 { 5963 {
6998 struct workqueue_struct *wq = dev_to_ 5964 struct workqueue_struct *wq = dev_to_wq(dev);
6999 int val; 5965 int val;
7000 5966
7001 if (sscanf(buf, "%d", &val) != 1 || v 5967 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
7002 return -EINVAL; 5968 return -EINVAL;
7003 5969
7004 workqueue_set_max_active(wq, val); 5970 workqueue_set_max_active(wq, val);
7005 return count; 5971 return count;
7006 } 5972 }
7007 static DEVICE_ATTR_RW(max_active); 5973 static DEVICE_ATTR_RW(max_active);
7008 5974
7009 static struct attribute *wq_sysfs_attrs[] = { 5975 static struct attribute *wq_sysfs_attrs[] = {
7010 &dev_attr_per_cpu.attr, 5976 &dev_attr_per_cpu.attr,
7011 &dev_attr_max_active.attr, 5977 &dev_attr_max_active.attr,
7012 NULL, 5978 NULL,
7013 }; 5979 };
7014 ATTRIBUTE_GROUPS(wq_sysfs); 5980 ATTRIBUTE_GROUPS(wq_sysfs);
7015 5981
7016 static ssize_t wq_nice_show(struct device *de 5982 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
7017 char *buf) 5983 char *buf)
7018 { 5984 {
7019 struct workqueue_struct *wq = dev_to_ 5985 struct workqueue_struct *wq = dev_to_wq(dev);
7020 int written; 5986 int written;
7021 5987
7022 mutex_lock(&wq->mutex); 5988 mutex_lock(&wq->mutex);
7023 written = scnprintf(buf, PAGE_SIZE, " 5989 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
7024 mutex_unlock(&wq->mutex); 5990 mutex_unlock(&wq->mutex);
7025 5991
7026 return written; 5992 return written;
7027 } 5993 }
7028 5994
7029 /* prepare workqueue_attrs for sysfs store op 5995 /* prepare workqueue_attrs for sysfs store operations */
7030 static struct workqueue_attrs *wq_sysfs_prep_ 5996 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
7031 { 5997 {
7032 struct workqueue_attrs *attrs; 5998 struct workqueue_attrs *attrs;
7033 5999
7034 lockdep_assert_held(&wq_pool_mutex); 6000 lockdep_assert_held(&wq_pool_mutex);
7035 6001
7036 attrs = alloc_workqueue_attrs(); 6002 attrs = alloc_workqueue_attrs();
7037 if (!attrs) 6003 if (!attrs)
7038 return NULL; 6004 return NULL;
7039 6005
7040 copy_workqueue_attrs(attrs, wq->unbou 6006 copy_workqueue_attrs(attrs, wq->unbound_attrs);
7041 return attrs; 6007 return attrs;
7042 } 6008 }
7043 6009
7044 static ssize_t wq_nice_store(struct device *d 6010 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
7045 const char *buf, 6011 const char *buf, size_t count)
7046 { 6012 {
7047 struct workqueue_struct *wq = dev_to_ 6013 struct workqueue_struct *wq = dev_to_wq(dev);
7048 struct workqueue_attrs *attrs; 6014 struct workqueue_attrs *attrs;
7049 int ret = -ENOMEM; 6015 int ret = -ENOMEM;
7050 6016
7051 apply_wqattrs_lock(); 6017 apply_wqattrs_lock();
7052 6018
7053 attrs = wq_sysfs_prep_attrs(wq); 6019 attrs = wq_sysfs_prep_attrs(wq);
7054 if (!attrs) 6020 if (!attrs)
7055 goto out_unlock; 6021 goto out_unlock;
7056 6022
7057 if (sscanf(buf, "%d", &attrs->nice) = 6023 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
7058 attrs->nice >= MIN_NICE && attrs- 6024 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
7059 ret = apply_workqueue_attrs_l 6025 ret = apply_workqueue_attrs_locked(wq, attrs);
7060 else 6026 else
7061 ret = -EINVAL; 6027 ret = -EINVAL;
7062 6028
7063 out_unlock: 6029 out_unlock:
7064 apply_wqattrs_unlock(); 6030 apply_wqattrs_unlock();
7065 free_workqueue_attrs(attrs); 6031 free_workqueue_attrs(attrs);
7066 return ret ?: count; 6032 return ret ?: count;
7067 } 6033 }
7068 6034
7069 static ssize_t wq_cpumask_show(struct device 6035 static ssize_t wq_cpumask_show(struct device *dev,
7070 struct device_ 6036 struct device_attribute *attr, char *buf)
7071 { 6037 {
7072 struct workqueue_struct *wq = dev_to_ 6038 struct workqueue_struct *wq = dev_to_wq(dev);
7073 int written; 6039 int written;
7074 6040
7075 mutex_lock(&wq->mutex); 6041 mutex_lock(&wq->mutex);
7076 written = scnprintf(buf, PAGE_SIZE, " 6042 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
7077 cpumask_pr_args(w 6043 cpumask_pr_args(wq->unbound_attrs->cpumask));
7078 mutex_unlock(&wq->mutex); 6044 mutex_unlock(&wq->mutex);
7079 return written; 6045 return written;
7080 } 6046 }
7081 6047
7082 static ssize_t wq_cpumask_store(struct device 6048 static ssize_t wq_cpumask_store(struct device *dev,
7083 struct device 6049 struct device_attribute *attr,
7084 const char *b 6050 const char *buf, size_t count)
7085 { 6051 {
7086 struct workqueue_struct *wq = dev_to_ 6052 struct workqueue_struct *wq = dev_to_wq(dev);
7087 struct workqueue_attrs *attrs; 6053 struct workqueue_attrs *attrs;
7088 int ret = -ENOMEM; 6054 int ret = -ENOMEM;
7089 6055
7090 apply_wqattrs_lock(); 6056 apply_wqattrs_lock();
7091 6057
7092 attrs = wq_sysfs_prep_attrs(wq); 6058 attrs = wq_sysfs_prep_attrs(wq);
7093 if (!attrs) 6059 if (!attrs)
7094 goto out_unlock; 6060 goto out_unlock;
7095 6061
7096 ret = cpumask_parse(buf, attrs->cpuma 6062 ret = cpumask_parse(buf, attrs->cpumask);
7097 if (!ret) 6063 if (!ret)
7098 ret = apply_workqueue_attrs_l 6064 ret = apply_workqueue_attrs_locked(wq, attrs);
7099 6065
7100 out_unlock: 6066 out_unlock:
7101 apply_wqattrs_unlock(); 6067 apply_wqattrs_unlock();
7102 free_workqueue_attrs(attrs); 6068 free_workqueue_attrs(attrs);
7103 return ret ?: count; 6069 return ret ?: count;
7104 } 6070 }
7105 6071
7106 static ssize_t wq_affn_scope_show(struct devi 6072 static ssize_t wq_affn_scope_show(struct device *dev,
7107 struct devi 6073 struct device_attribute *attr, char *buf)
7108 { 6074 {
7109 struct workqueue_struct *wq = dev_to_ 6075 struct workqueue_struct *wq = dev_to_wq(dev);
7110 int written; 6076 int written;
7111 6077
7112 mutex_lock(&wq->mutex); 6078 mutex_lock(&wq->mutex);
7113 if (wq->unbound_attrs->affn_scope == 6079 if (wq->unbound_attrs->affn_scope == WQ_AFFN_DFL)
7114 written = scnprintf(buf, PAGE 6080 written = scnprintf(buf, PAGE_SIZE, "%s (%s)\n",
7115 wq_affn_n 6081 wq_affn_names[WQ_AFFN_DFL],
7116 wq_affn_n 6082 wq_affn_names[wq_affn_dfl]);
7117 else 6083 else
7118 written = scnprintf(buf, PAGE 6084 written = scnprintf(buf, PAGE_SIZE, "%s\n",
7119 wq_affn_n 6085 wq_affn_names[wq->unbound_attrs->affn_scope]);
7120 mutex_unlock(&wq->mutex); 6086 mutex_unlock(&wq->mutex);
7121 6087
7122 return written; 6088 return written;
7123 } 6089 }
7124 6090
7125 static ssize_t wq_affn_scope_store(struct dev 6091 static ssize_t wq_affn_scope_store(struct device *dev,
7126 struct dev 6092 struct device_attribute *attr,
7127 const char 6093 const char *buf, size_t count)
7128 { 6094 {
7129 struct workqueue_struct *wq = dev_to_ 6095 struct workqueue_struct *wq = dev_to_wq(dev);
7130 struct workqueue_attrs *attrs; 6096 struct workqueue_attrs *attrs;
7131 int affn, ret = -ENOMEM; 6097 int affn, ret = -ENOMEM;
7132 6098
7133 affn = parse_affn_scope(buf); 6099 affn = parse_affn_scope(buf);
7134 if (affn < 0) 6100 if (affn < 0)
7135 return affn; 6101 return affn;
7136 6102
7137 apply_wqattrs_lock(); 6103 apply_wqattrs_lock();
7138 attrs = wq_sysfs_prep_attrs(wq); 6104 attrs = wq_sysfs_prep_attrs(wq);
7139 if (attrs) { 6105 if (attrs) {
7140 attrs->affn_scope = affn; 6106 attrs->affn_scope = affn;
7141 ret = apply_workqueue_attrs_l 6107 ret = apply_workqueue_attrs_locked(wq, attrs);
7142 } 6108 }
7143 apply_wqattrs_unlock(); 6109 apply_wqattrs_unlock();
7144 free_workqueue_attrs(attrs); 6110 free_workqueue_attrs(attrs);
7145 return ret ?: count; 6111 return ret ?: count;
7146 } 6112 }
7147 6113
7148 static ssize_t wq_affinity_strict_show(struct 6114 static ssize_t wq_affinity_strict_show(struct device *dev,
7149 struct 6115 struct device_attribute *attr, char *buf)
7150 { 6116 {
7151 struct workqueue_struct *wq = dev_to_ 6117 struct workqueue_struct *wq = dev_to_wq(dev);
7152 6118
7153 return scnprintf(buf, PAGE_SIZE, "%d\ 6119 return scnprintf(buf, PAGE_SIZE, "%d\n",
7154 wq->unbound_attrs->a 6120 wq->unbound_attrs->affn_strict);
7155 } 6121 }
7156 6122
7157 static ssize_t wq_affinity_strict_store(struc 6123 static ssize_t wq_affinity_strict_store(struct device *dev,
7158 struc 6124 struct device_attribute *attr,
7159 const 6125 const char *buf, size_t count)
7160 { 6126 {
7161 struct workqueue_struct *wq = dev_to_ 6127 struct workqueue_struct *wq = dev_to_wq(dev);
7162 struct workqueue_attrs *attrs; 6128 struct workqueue_attrs *attrs;
7163 int v, ret = -ENOMEM; 6129 int v, ret = -ENOMEM;
7164 6130
7165 if (sscanf(buf, "%d", &v) != 1) 6131 if (sscanf(buf, "%d", &v) != 1)
7166 return -EINVAL; 6132 return -EINVAL;
7167 6133
7168 apply_wqattrs_lock(); 6134 apply_wqattrs_lock();
7169 attrs = wq_sysfs_prep_attrs(wq); 6135 attrs = wq_sysfs_prep_attrs(wq);
7170 if (attrs) { 6136 if (attrs) {
7171 attrs->affn_strict = (bool)v; 6137 attrs->affn_strict = (bool)v;
7172 ret = apply_workqueue_attrs_l 6138 ret = apply_workqueue_attrs_locked(wq, attrs);
7173 } 6139 }
7174 apply_wqattrs_unlock(); 6140 apply_wqattrs_unlock();
7175 free_workqueue_attrs(attrs); 6141 free_workqueue_attrs(attrs);
7176 return ret ?: count; 6142 return ret ?: count;
7177 } 6143 }
7178 6144
7179 static struct device_attribute wq_sysfs_unbou 6145 static struct device_attribute wq_sysfs_unbound_attrs[] = {
7180 __ATTR(nice, 0644, wq_nice_show, wq_n 6146 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
7181 __ATTR(cpumask, 0644, wq_cpumask_show 6147 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
7182 __ATTR(affinity_scope, 0644, wq_affn_ 6148 __ATTR(affinity_scope, 0644, wq_affn_scope_show, wq_affn_scope_store),
7183 __ATTR(affinity_strict, 0644, wq_affi 6149 __ATTR(affinity_strict, 0644, wq_affinity_strict_show, wq_affinity_strict_store),
7184 __ATTR_NULL, 6150 __ATTR_NULL,
7185 }; 6151 };
7186 6152
7187 static const struct bus_type wq_subsys = { !! 6153 static struct bus_type wq_subsys = {
7188 .name = "wo 6154 .name = "workqueue",
7189 .dev_groups = wq_ 6155 .dev_groups = wq_sysfs_groups,
7190 }; 6156 };
7191 6157
7192 /** !! 6158 static ssize_t wq_unbound_cpumask_show(struct device *dev,
7193 * workqueue_set_unbound_cpumask - Set the l !! 6159 struct device_attribute *attr, char *buf)
7194 * @cpumask: the cpumask to set <<
7195 * <<
7196 * The low-level workqueues cpumask is a glo <<
7197 * the affinity of all unbound workqueues. <<
7198 * and apply it to all unbound workqueues an <<
7199 * <<
7200 * Return: 0 - Success <<
7201 * -EINVAL - Invalid @cpumask <<
7202 * -ENOMEM - Failed to allocate <<
7203 */ <<
7204 static int workqueue_set_unbound_cpumask(cpum <<
7205 { <<
7206 int ret = -EINVAL; <<
7207 <<
7208 /* <<
7209 * Not excluding isolated cpus on pur <<
7210 * If the user wishes to include them <<
7211 */ <<
7212 cpumask_and(cpumask, cpumask, cpu_pos <<
7213 if (!cpumask_empty(cpumask)) { <<
7214 ret = 0; <<
7215 apply_wqattrs_lock(); <<
7216 if (!cpumask_equal(cpumask, w <<
7217 ret = workqueue_apply <<
7218 if (!ret) <<
7219 cpumask_copy(wq_reque <<
7220 apply_wqattrs_unlock(); <<
7221 } <<
7222 <<
7223 return ret; <<
7224 } <<
7225 <<
7226 static ssize_t __wq_cpumask_show(struct devic <<
7227 struct device_attribute *attr <<
7228 { 6160 {
7229 int written; 6161 int written;
7230 6162
7231 mutex_lock(&wq_pool_mutex); 6163 mutex_lock(&wq_pool_mutex);
7232 written = scnprintf(buf, PAGE_SIZE, " !! 6164 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
>> 6165 cpumask_pr_args(wq_unbound_cpumask));
7233 mutex_unlock(&wq_pool_mutex); 6166 mutex_unlock(&wq_pool_mutex);
7234 6167
7235 return written; 6168 return written;
7236 } 6169 }
7237 6170
7238 static ssize_t cpumask_requested_show(struct !! 6171 static ssize_t wq_unbound_cpumask_store(struct device *dev,
7239 struct device_attribute *attr <<
7240 { <<
7241 return __wq_cpumask_show(dev, attr, b <<
7242 } <<
7243 static DEVICE_ATTR_RO(cpumask_requested); <<
7244 <<
7245 static ssize_t cpumask_isolated_show(struct d <<
7246 struct device_attribute *attr <<
7247 { <<
7248 return __wq_cpumask_show(dev, attr, b <<
7249 } <<
7250 static DEVICE_ATTR_RO(cpumask_isolated); <<
7251 <<
7252 static ssize_t cpumask_show(struct device *de <<
7253 struct device_attribute *attr <<
7254 { <<
7255 return __wq_cpumask_show(dev, attr, b <<
7256 } <<
7257 <<
7258 static ssize_t cpumask_store(struct device *d <<
7259 struct device_attribute *attr 6172 struct device_attribute *attr, const char *buf, size_t count)
7260 { 6173 {
7261 cpumask_var_t cpumask; 6174 cpumask_var_t cpumask;
7262 int ret; 6175 int ret;
7263 6176
7264 if (!zalloc_cpumask_var(&cpumask, GFP 6177 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
7265 return -ENOMEM; 6178 return -ENOMEM;
7266 6179
7267 ret = cpumask_parse(buf, cpumask); 6180 ret = cpumask_parse(buf, cpumask);
7268 if (!ret) 6181 if (!ret)
7269 ret = workqueue_set_unbound_c 6182 ret = workqueue_set_unbound_cpumask(cpumask);
7270 6183
7271 free_cpumask_var(cpumask); 6184 free_cpumask_var(cpumask);
7272 return ret ? ret : count; 6185 return ret ? ret : count;
7273 } 6186 }
7274 static DEVICE_ATTR_RW(cpumask); <<
7275 6187
7276 static struct attribute *wq_sysfs_cpumask_att !! 6188 static struct device_attribute wq_sysfs_cpumask_attr =
7277 &dev_attr_cpumask.attr, !! 6189 __ATTR(cpumask, 0644, wq_unbound_cpumask_show,
7278 &dev_attr_cpumask_requested.attr, !! 6190 wq_unbound_cpumask_store);
7279 &dev_attr_cpumask_isolated.attr, <<
7280 NULL, <<
7281 }; <<
7282 ATTRIBUTE_GROUPS(wq_sysfs_cpumask); <<
7283 6191
7284 static int __init wq_sysfs_init(void) 6192 static int __init wq_sysfs_init(void)
7285 { 6193 {
7286 return subsys_virtual_register(&wq_su !! 6194 struct device *dev_root;
>> 6195 int err;
>> 6196
>> 6197 err = subsys_virtual_register(&wq_subsys, NULL);
>> 6198 if (err)
>> 6199 return err;
>> 6200
>> 6201 dev_root = bus_get_dev_root(&wq_subsys);
>> 6202 if (dev_root) {
>> 6203 err = device_create_file(dev_root, &wq_sysfs_cpumask_attr);
>> 6204 put_device(dev_root);
>> 6205 }
>> 6206 return err;
7287 } 6207 }
7288 core_initcall(wq_sysfs_init); 6208 core_initcall(wq_sysfs_init);
7289 6209
7290 static void wq_device_release(struct device * 6210 static void wq_device_release(struct device *dev)
7291 { 6211 {
7292 struct wq_device *wq_dev = container_ 6212 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
7293 6213
7294 kfree(wq_dev); 6214 kfree(wq_dev);
7295 } 6215 }
7296 6216
7297 /** 6217 /**
7298 * workqueue_sysfs_register - make a workqueu 6218 * workqueue_sysfs_register - make a workqueue visible in sysfs
7299 * @wq: the workqueue to register 6219 * @wq: the workqueue to register
7300 * 6220 *
7301 * Expose @wq in sysfs under /sys/bus/workque 6221 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
7302 * alloc_workqueue*() automatically calls thi 6222 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
7303 * which is the preferred method. 6223 * which is the preferred method.
7304 * 6224 *
7305 * Workqueue user should use this function di 6225 * Workqueue user should use this function directly iff it wants to apply
7306 * workqueue_attrs before making the workqueu 6226 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
7307 * apply_workqueue_attrs() may race against u 6227 * apply_workqueue_attrs() may race against userland updating the
7308 * attributes. 6228 * attributes.
7309 * 6229 *
7310 * Return: 0 on success, -errno on failure. 6230 * Return: 0 on success, -errno on failure.
7311 */ 6231 */
7312 int workqueue_sysfs_register(struct workqueue 6232 int workqueue_sysfs_register(struct workqueue_struct *wq)
7313 { 6233 {
7314 struct wq_device *wq_dev; 6234 struct wq_device *wq_dev;
7315 int ret; 6235 int ret;
7316 6236
7317 /* 6237 /*
7318 * Adjusting max_active breaks orderi !! 6238 * Adjusting max_active or creating new pwqs by applying
7319 * ordered workqueues. !! 6239 * attributes breaks ordering guarantee. Disallow exposing ordered
>> 6240 * workqueues.
7320 */ 6241 */
7321 if (WARN_ON(wq->flags & __WQ_ORDERED) !! 6242 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
7322 return -EINVAL; 6243 return -EINVAL;
7323 6244
7324 wq->wq_dev = wq_dev = kzalloc(sizeof( 6245 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
7325 if (!wq_dev) 6246 if (!wq_dev)
7326 return -ENOMEM; 6247 return -ENOMEM;
7327 6248
7328 wq_dev->wq = wq; 6249 wq_dev->wq = wq;
7329 wq_dev->dev.bus = &wq_subsys; 6250 wq_dev->dev.bus = &wq_subsys;
7330 wq_dev->dev.release = wq_device_relea 6251 wq_dev->dev.release = wq_device_release;
7331 dev_set_name(&wq_dev->dev, "%s", wq-> 6252 dev_set_name(&wq_dev->dev, "%s", wq->name);
7332 6253
7333 /* 6254 /*
7334 * unbound_attrs are created separate 6255 * unbound_attrs are created separately. Suppress uevent until
7335 * everything is ready. 6256 * everything is ready.
7336 */ 6257 */
7337 dev_set_uevent_suppress(&wq_dev->dev, 6258 dev_set_uevent_suppress(&wq_dev->dev, true);
7338 6259
7339 ret = device_register(&wq_dev->dev); 6260 ret = device_register(&wq_dev->dev);
7340 if (ret) { 6261 if (ret) {
7341 put_device(&wq_dev->dev); 6262 put_device(&wq_dev->dev);
7342 wq->wq_dev = NULL; 6263 wq->wq_dev = NULL;
7343 return ret; 6264 return ret;
7344 } 6265 }
7345 6266
7346 if (wq->flags & WQ_UNBOUND) { 6267 if (wq->flags & WQ_UNBOUND) {
7347 struct device_attribute *attr 6268 struct device_attribute *attr;
7348 6269
7349 for (attr = wq_sysfs_unbound_ 6270 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
7350 ret = device_create_f 6271 ret = device_create_file(&wq_dev->dev, attr);
7351 if (ret) { 6272 if (ret) {
7352 device_unregi 6273 device_unregister(&wq_dev->dev);
7353 wq->wq_dev = 6274 wq->wq_dev = NULL;
7354 return ret; 6275 return ret;
7355 } 6276 }
7356 } 6277 }
7357 } 6278 }
7358 6279
7359 dev_set_uevent_suppress(&wq_dev->dev, 6280 dev_set_uevent_suppress(&wq_dev->dev, false);
7360 kobject_uevent(&wq_dev->dev.kobj, KOB 6281 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
7361 return 0; 6282 return 0;
7362 } 6283 }
7363 6284
7364 /** 6285 /**
7365 * workqueue_sysfs_unregister - undo workqueu 6286 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
7366 * @wq: the workqueue to unregister 6287 * @wq: the workqueue to unregister
7367 * 6288 *
7368 * If @wq is registered to sysfs by workqueue 6289 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
7369 */ 6290 */
7370 static void workqueue_sysfs_unregister(struct 6291 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
7371 { 6292 {
7372 struct wq_device *wq_dev = wq->wq_dev 6293 struct wq_device *wq_dev = wq->wq_dev;
7373 6294
7374 if (!wq->wq_dev) 6295 if (!wq->wq_dev)
7375 return; 6296 return;
7376 6297
7377 wq->wq_dev = NULL; 6298 wq->wq_dev = NULL;
7378 device_unregister(&wq_dev->dev); 6299 device_unregister(&wq_dev->dev);
7379 } 6300 }
7380 #else /* CONFIG_SYSFS */ 6301 #else /* CONFIG_SYSFS */
7381 static void workqueue_sysfs_unregister(struct 6302 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
7382 #endif /* CONFIG_SYSFS */ 6303 #endif /* CONFIG_SYSFS */
7383 6304
7384 /* 6305 /*
7385 * Workqueue watchdog. 6306 * Workqueue watchdog.
7386 * 6307 *
7387 * Stall may be caused by various bugs - miss 6308 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
7388 * flush dependency, a concurrency managed wo 6309 * flush dependency, a concurrency managed work item which stays RUNNING
7389 * indefinitely. Workqueue stalls can be ver 6310 * indefinitely. Workqueue stalls can be very difficult to debug as the
7390 * usual warning mechanisms don't trigger and 6311 * usual warning mechanisms don't trigger and internal workqueue state is
7391 * largely opaque. 6312 * largely opaque.
7392 * 6313 *
7393 * Workqueue watchdog monitors all worker poo 6314 * Workqueue watchdog monitors all worker pools periodically and dumps
7394 * state if some pools failed to make forward 6315 * state if some pools failed to make forward progress for a while where
7395 * forward progress is defined as the first i 6316 * forward progress is defined as the first item on ->worklist changing.
7396 * 6317 *
7397 * This mechanism is controlled through the k 6318 * This mechanism is controlled through the kernel parameter
7398 * "workqueue.watchdog_thresh" which can be u 6319 * "workqueue.watchdog_thresh" which can be updated at runtime through the
7399 * corresponding sysfs parameter file. 6320 * corresponding sysfs parameter file.
7400 */ 6321 */
7401 #ifdef CONFIG_WQ_WATCHDOG 6322 #ifdef CONFIG_WQ_WATCHDOG
7402 6323
7403 static unsigned long wq_watchdog_thresh = 30; 6324 static unsigned long wq_watchdog_thresh = 30;
7404 static struct timer_list wq_watchdog_timer; 6325 static struct timer_list wq_watchdog_timer;
7405 6326
7406 static unsigned long wq_watchdog_touched = IN 6327 static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
7407 static DEFINE_PER_CPU(unsigned long, wq_watch 6328 static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
7408 6329
7409 /* 6330 /*
7410 * Show workers that might prevent the proces 6331 * Show workers that might prevent the processing of pending work items.
7411 * The only candidates are CPU-bound workers 6332 * The only candidates are CPU-bound workers in the running state.
7412 * Pending work items should be handled by an 6333 * Pending work items should be handled by another idle worker
7413 * in all other situations. 6334 * in all other situations.
7414 */ 6335 */
7415 static void show_cpu_pool_hog(struct worker_p 6336 static void show_cpu_pool_hog(struct worker_pool *pool)
7416 { 6337 {
7417 struct worker *worker; 6338 struct worker *worker;
7418 unsigned long irq_flags; !! 6339 unsigned long flags;
7419 int bkt; 6340 int bkt;
7420 6341
7421 raw_spin_lock_irqsave(&pool->lock, ir !! 6342 raw_spin_lock_irqsave(&pool->lock, flags);
7422 6343
7423 hash_for_each(pool->busy_hash, bkt, w 6344 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
7424 if (task_is_running(worker->t 6345 if (task_is_running(worker->task)) {
7425 /* 6346 /*
7426 * Defer printing to 6347 * Defer printing to avoid deadlocks in console
7427 * drivers that queue 6348 * drivers that queue work while holding locks
7428 * also taken in thei 6349 * also taken in their write paths.
7429 */ 6350 */
7430 printk_deferred_enter 6351 printk_deferred_enter();
7431 6352
7432 pr_info("pool %d:\n", 6353 pr_info("pool %d:\n", pool->id);
7433 sched_show_task(worke 6354 sched_show_task(worker->task);
7434 6355
7435 printk_deferred_exit( 6356 printk_deferred_exit();
7436 } 6357 }
7437 } 6358 }
7438 6359
7439 raw_spin_unlock_irqrestore(&pool->loc !! 6360 raw_spin_unlock_irqrestore(&pool->lock, flags);
7440 } 6361 }
7441 6362
7442 static void show_cpu_pools_hogs(void) 6363 static void show_cpu_pools_hogs(void)
7443 { 6364 {
7444 struct worker_pool *pool; 6365 struct worker_pool *pool;
7445 int pi; 6366 int pi;
7446 6367
7447 pr_info("Showing backtraces of runnin 6368 pr_info("Showing backtraces of running workers in stalled CPU-bound worker pools:\n");
7448 6369
7449 rcu_read_lock(); 6370 rcu_read_lock();
7450 6371
7451 for_each_pool(pool, pi) { 6372 for_each_pool(pool, pi) {
7452 if (pool->cpu_stall) 6373 if (pool->cpu_stall)
7453 show_cpu_pool_hog(poo 6374 show_cpu_pool_hog(pool);
7454 6375
7455 } 6376 }
7456 6377
7457 rcu_read_unlock(); 6378 rcu_read_unlock();
7458 } 6379 }
7459 6380
7460 static void wq_watchdog_reset_touched(void) 6381 static void wq_watchdog_reset_touched(void)
7461 { 6382 {
7462 int cpu; 6383 int cpu;
7463 6384
7464 wq_watchdog_touched = jiffies; 6385 wq_watchdog_touched = jiffies;
7465 for_each_possible_cpu(cpu) 6386 for_each_possible_cpu(cpu)
7466 per_cpu(wq_watchdog_touched_c 6387 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
7467 } 6388 }
7468 6389
7469 static void wq_watchdog_timer_fn(struct timer 6390 static void wq_watchdog_timer_fn(struct timer_list *unused)
7470 { 6391 {
7471 unsigned long thresh = READ_ONCE(wq_w 6392 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
7472 bool lockup_detected = false; 6393 bool lockup_detected = false;
7473 bool cpu_pool_stall = false; 6394 bool cpu_pool_stall = false;
7474 unsigned long now = jiffies; 6395 unsigned long now = jiffies;
7475 struct worker_pool *pool; 6396 struct worker_pool *pool;
7476 int pi; 6397 int pi;
7477 6398
7478 if (!thresh) 6399 if (!thresh)
7479 return; 6400 return;
7480 6401
7481 rcu_read_lock(); 6402 rcu_read_lock();
7482 6403
7483 for_each_pool(pool, pi) { 6404 for_each_pool(pool, pi) {
7484 unsigned long pool_ts, touche 6405 unsigned long pool_ts, touched, ts;
7485 6406
7486 pool->cpu_stall = false; 6407 pool->cpu_stall = false;
7487 if (list_empty(&pool->worklis 6408 if (list_empty(&pool->worklist))
7488 continue; 6409 continue;
7489 6410
7490 /* 6411 /*
7491 * If a virtual machine is st 6412 * If a virtual machine is stopped by the host it can look to
7492 * the watchdog like a stall. 6413 * the watchdog like a stall.
7493 */ 6414 */
7494 kvm_check_and_clear_guest_pau 6415 kvm_check_and_clear_guest_paused();
7495 6416
7496 /* get the latest of pool and 6417 /* get the latest of pool and touched timestamps */
7497 if (pool->cpu >= 0) 6418 if (pool->cpu >= 0)
7498 touched = READ_ONCE(p 6419 touched = READ_ONCE(per_cpu(wq_watchdog_touched_cpu, pool->cpu));
7499 else 6420 else
7500 touched = READ_ONCE(w 6421 touched = READ_ONCE(wq_watchdog_touched);
7501 pool_ts = READ_ONCE(pool->wat 6422 pool_ts = READ_ONCE(pool->watchdog_ts);
7502 6423
7503 if (time_after(pool_ts, touch 6424 if (time_after(pool_ts, touched))
7504 ts = pool_ts; 6425 ts = pool_ts;
7505 else 6426 else
7506 ts = touched; 6427 ts = touched;
7507 6428
7508 /* did we stall? */ 6429 /* did we stall? */
7509 if (time_after(now, ts + thre 6430 if (time_after(now, ts + thresh)) {
7510 lockup_detected = tru 6431 lockup_detected = true;
7511 if (pool->cpu >= 0 && !! 6432 if (pool->cpu >= 0) {
7512 pool->cpu_sta 6433 pool->cpu_stall = true;
7513 cpu_pool_stal 6434 cpu_pool_stall = true;
7514 } 6435 }
7515 pr_emerg("BUG: workqu 6436 pr_emerg("BUG: workqueue lockup - pool");
7516 pr_cont_pool_info(poo 6437 pr_cont_pool_info(pool);
7517 pr_cont(" stuck for % 6438 pr_cont(" stuck for %us!\n",
7518 jiffies_to_ms 6439 jiffies_to_msecs(now - pool_ts) / 1000);
7519 } 6440 }
7520 6441
7521 6442
7522 } 6443 }
7523 6444
7524 rcu_read_unlock(); 6445 rcu_read_unlock();
7525 6446
7526 if (lockup_detected) 6447 if (lockup_detected)
7527 show_all_workqueues(); 6448 show_all_workqueues();
7528 6449
7529 if (cpu_pool_stall) 6450 if (cpu_pool_stall)
7530 show_cpu_pools_hogs(); 6451 show_cpu_pools_hogs();
7531 6452
7532 wq_watchdog_reset_touched(); 6453 wq_watchdog_reset_touched();
7533 mod_timer(&wq_watchdog_timer, jiffies 6454 mod_timer(&wq_watchdog_timer, jiffies + thresh);
7534 } 6455 }
7535 6456
7536 notrace void wq_watchdog_touch(int cpu) 6457 notrace void wq_watchdog_touch(int cpu)
7537 { 6458 {
7538 unsigned long thresh = READ_ONCE(wq_w 6459 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
7539 unsigned long touch_ts = READ_ONCE(wq 6460 unsigned long touch_ts = READ_ONCE(wq_watchdog_touched);
7540 unsigned long now = jiffies; 6461 unsigned long now = jiffies;
7541 6462
7542 if (cpu >= 0) 6463 if (cpu >= 0)
7543 per_cpu(wq_watchdog_touched_c 6464 per_cpu(wq_watchdog_touched_cpu, cpu) = now;
7544 else 6465 else
7545 WARN_ONCE(1, "%s should be ca 6466 WARN_ONCE(1, "%s should be called with valid CPU", __func__);
7546 6467
7547 /* Don't unnecessarily store to globa 6468 /* Don't unnecessarily store to global cacheline */
7548 if (time_after(now, touch_ts + thresh 6469 if (time_after(now, touch_ts + thresh / 4))
7549 WRITE_ONCE(wq_watchdog_touche 6470 WRITE_ONCE(wq_watchdog_touched, jiffies);
7550 } 6471 }
7551 6472
7552 static void wq_watchdog_set_thresh(unsigned l 6473 static void wq_watchdog_set_thresh(unsigned long thresh)
7553 { 6474 {
7554 wq_watchdog_thresh = 0; 6475 wq_watchdog_thresh = 0;
7555 del_timer_sync(&wq_watchdog_timer); 6476 del_timer_sync(&wq_watchdog_timer);
7556 6477
7557 if (thresh) { 6478 if (thresh) {
7558 wq_watchdog_thresh = thresh; 6479 wq_watchdog_thresh = thresh;
7559 wq_watchdog_reset_touched(); 6480 wq_watchdog_reset_touched();
7560 mod_timer(&wq_watchdog_timer, 6481 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
7561 } 6482 }
7562 } 6483 }
7563 6484
7564 static int wq_watchdog_param_set_thresh(const 6485 static int wq_watchdog_param_set_thresh(const char *val,
7565 const 6486 const struct kernel_param *kp)
7566 { 6487 {
7567 unsigned long thresh; 6488 unsigned long thresh;
7568 int ret; 6489 int ret;
7569 6490
7570 ret = kstrtoul(val, 0, &thresh); 6491 ret = kstrtoul(val, 0, &thresh);
7571 if (ret) 6492 if (ret)
7572 return ret; 6493 return ret;
7573 6494
7574 if (system_wq) 6495 if (system_wq)
7575 wq_watchdog_set_thresh(thresh 6496 wq_watchdog_set_thresh(thresh);
7576 else 6497 else
7577 wq_watchdog_thresh = thresh; 6498 wq_watchdog_thresh = thresh;
7578 6499
7579 return 0; 6500 return 0;
7580 } 6501 }
7581 6502
7582 static const struct kernel_param_ops wq_watch 6503 static const struct kernel_param_ops wq_watchdog_thresh_ops = {
7583 .set = wq_watchdog_param_set_thres 6504 .set = wq_watchdog_param_set_thresh,
7584 .get = param_get_ulong, 6505 .get = param_get_ulong,
7585 }; 6506 };
7586 6507
7587 module_param_cb(watchdog_thresh, &wq_watchdog 6508 module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
7588 0644); 6509 0644);
7589 6510
7590 static void wq_watchdog_init(void) 6511 static void wq_watchdog_init(void)
7591 { 6512 {
7592 timer_setup(&wq_watchdog_timer, wq_wa 6513 timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE);
7593 wq_watchdog_set_thresh(wq_watchdog_th 6514 wq_watchdog_set_thresh(wq_watchdog_thresh);
7594 } 6515 }
7595 6516
7596 #else /* CONFIG_WQ_WATCHDOG */ 6517 #else /* CONFIG_WQ_WATCHDOG */
7597 6518
7598 static inline void wq_watchdog_init(void) { } 6519 static inline void wq_watchdog_init(void) { }
7599 6520
7600 #endif /* CONFIG_WQ_WATCHDOG */ 6521 #endif /* CONFIG_WQ_WATCHDOG */
7601 6522
7602 static void bh_pool_kick_normal(struct irq_wo <<
7603 { <<
7604 raise_softirq_irqoff(TASKLET_SOFTIRQ) <<
7605 } <<
7606 <<
7607 static void bh_pool_kick_highpri(struct irq_w <<
7608 { <<
7609 raise_softirq_irqoff(HI_SOFTIRQ); <<
7610 } <<
7611 <<
7612 static void __init restrict_unbound_cpumask(c 6523 static void __init restrict_unbound_cpumask(const char *name, const struct cpumask *mask)
7613 { 6524 {
7614 if (!cpumask_intersects(wq_unbound_cp 6525 if (!cpumask_intersects(wq_unbound_cpumask, mask)) {
7615 pr_warn("workqueue: Restricti 6526 pr_warn("workqueue: Restricting unbound_cpumask (%*pb) with %s (%*pb) leaves no CPU, ignoring\n",
7616 cpumask_pr_args(wq_un 6527 cpumask_pr_args(wq_unbound_cpumask), name, cpumask_pr_args(mask));
7617 return; 6528 return;
7618 } 6529 }
7619 6530
7620 cpumask_and(wq_unbound_cpumask, wq_un 6531 cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, mask);
7621 } 6532 }
7622 6533
7623 static void __init init_cpu_worker_pool(struc <<
7624 { <<
7625 BUG_ON(init_worker_pool(pool)); <<
7626 pool->cpu = cpu; <<
7627 cpumask_copy(pool->attrs->cpumask, cp <<
7628 cpumask_copy(pool->attrs->__pod_cpuma <<
7629 pool->attrs->nice = nice; <<
7630 pool->attrs->affn_strict = true; <<
7631 pool->node = cpu_to_node(cpu); <<
7632 <<
7633 /* alloc pool ID */ <<
7634 mutex_lock(&wq_pool_mutex); <<
7635 BUG_ON(worker_pool_assign_id(pool)); <<
7636 mutex_unlock(&wq_pool_mutex); <<
7637 } <<
7638 <<
7639 /** 6534 /**
7640 * workqueue_init_early - early init for work 6535 * workqueue_init_early - early init for workqueue subsystem
7641 * 6536 *
7642 * This is the first step of three-staged wor 6537 * This is the first step of three-staged workqueue subsystem initialization and
7643 * invoked as soon as the bare basics - memor 6538 * invoked as soon as the bare basics - memory allocation, cpumasks and idr are
7644 * up. It sets up all the data structures and 6539 * up. It sets up all the data structures and system workqueues and allows early
7645 * boot code to create workqueues and queue/c 6540 * boot code to create workqueues and queue/cancel work items. Actual work item
7646 * execution starts only after kthreads can b 6541 * execution starts only after kthreads can be created and scheduled right
7647 * before early initcalls. 6542 * before early initcalls.
7648 */ 6543 */
7649 void __init workqueue_init_early(void) 6544 void __init workqueue_init_early(void)
7650 { 6545 {
7651 struct wq_pod_type *pt = &wq_pod_type 6546 struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_SYSTEM];
7652 int std_nice[NR_STD_WORKER_POOLS] = { 6547 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
7653 void (*irq_work_fns[2])(struct irq_wo <<
7654 <<
7655 int i, cpu; 6548 int i, cpu;
7656 6549
7657 BUILD_BUG_ON(__alignof__(struct pool_ 6550 BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
7658 6551
7659 BUG_ON(!alloc_cpumask_var(&wq_online_ <<
7660 BUG_ON(!alloc_cpumask_var(&wq_unbound 6552 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
7661 BUG_ON(!alloc_cpumask_var(&wq_request <<
7662 BUG_ON(!zalloc_cpumask_var(&wq_isolat <<
7663 <<
7664 cpumask_copy(wq_online_cpumask, cpu_o <<
7665 cpumask_copy(wq_unbound_cpumask, cpu_ 6553 cpumask_copy(wq_unbound_cpumask, cpu_possible_mask);
7666 restrict_unbound_cpumask("HK_TYPE_WQ" 6554 restrict_unbound_cpumask("HK_TYPE_WQ", housekeeping_cpumask(HK_TYPE_WQ));
7667 restrict_unbound_cpumask("HK_TYPE_DOM 6555 restrict_unbound_cpumask("HK_TYPE_DOMAIN", housekeeping_cpumask(HK_TYPE_DOMAIN));
7668 if (!cpumask_empty(&wq_cmdline_cpumas 6556 if (!cpumask_empty(&wq_cmdline_cpumask))
7669 restrict_unbound_cpumask("wor 6557 restrict_unbound_cpumask("workqueue.unbound_cpus", &wq_cmdline_cpumask);
7670 6558
7671 cpumask_copy(wq_requested_unbound_cpu <<
7672 <<
7673 pwq_cache = KMEM_CACHE(pool_workqueue 6559 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
7674 6560
7675 unbound_wq_update_pwq_attrs_buf = all !! 6561 wq_update_pod_attrs_buf = alloc_workqueue_attrs();
7676 BUG_ON(!unbound_wq_update_pwq_attrs_b !! 6562 BUG_ON(!wq_update_pod_attrs_buf);
7677 <<
7678 /* <<
7679 * If nohz_full is enabled, set power <<
7680 * This allows workqueue items to be <<
7681 */ <<
7682 if (housekeeping_enabled(HK_TYPE_TICK <<
7683 wq_power_efficient = true; <<
7684 6563
7685 /* initialize WQ_AFFN_SYSTEM pods */ 6564 /* initialize WQ_AFFN_SYSTEM pods */
7686 pt->pod_cpus = kcalloc(1, sizeof(pt-> 6565 pt->pod_cpus = kcalloc(1, sizeof(pt->pod_cpus[0]), GFP_KERNEL);
7687 pt->pod_node = kcalloc(1, sizeof(pt-> 6566 pt->pod_node = kcalloc(1, sizeof(pt->pod_node[0]), GFP_KERNEL);
7688 pt->cpu_pod = kcalloc(nr_cpu_ids, siz 6567 pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL);
7689 BUG_ON(!pt->pod_cpus || !pt->pod_node 6568 BUG_ON(!pt->pod_cpus || !pt->pod_node || !pt->cpu_pod);
7690 6569
7691 BUG_ON(!zalloc_cpumask_var_node(&pt-> 6570 BUG_ON(!zalloc_cpumask_var_node(&pt->pod_cpus[0], GFP_KERNEL, NUMA_NO_NODE));
7692 6571
7693 pt->nr_pods = 1; 6572 pt->nr_pods = 1;
7694 cpumask_copy(pt->pod_cpus[0], cpu_pos 6573 cpumask_copy(pt->pod_cpus[0], cpu_possible_mask);
7695 pt->pod_node[0] = NUMA_NO_NODE; 6574 pt->pod_node[0] = NUMA_NO_NODE;
7696 pt->cpu_pod[0] = 0; 6575 pt->cpu_pod[0] = 0;
7697 6576
7698 /* initialize BH and CPU pools */ !! 6577 /* initialize CPU pools */
7699 for_each_possible_cpu(cpu) { 6578 for_each_possible_cpu(cpu) {
7700 struct worker_pool *pool; 6579 struct worker_pool *pool;
7701 6580
7702 i = 0; 6581 i = 0;
7703 for_each_bh_worker_pool(pool, !! 6582 for_each_cpu_worker_pool(pool, cpu) {
7704 init_cpu_worker_pool( !! 6583 BUG_ON(init_worker_pool(pool));
7705 pool->flags |= POOL_B !! 6584 pool->cpu = cpu;
7706 init_irq_work(bh_pool !! 6585 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
7707 i++; !! 6586 cpumask_copy(pool->attrs->__pod_cpumask, cpumask_of(cpu));
7708 } !! 6587 pool->attrs->nice = std_nice[i++];
>> 6588 pool->attrs->affn_strict = true;
>> 6589 pool->node = cpu_to_node(cpu);
7709 6590
7710 i = 0; !! 6591 /* alloc pool ID */
7711 for_each_cpu_worker_pool(pool !! 6592 mutex_lock(&wq_pool_mutex);
7712 init_cpu_worker_pool( !! 6593 BUG_ON(worker_pool_assign_id(pool));
>> 6594 mutex_unlock(&wq_pool_mutex);
>> 6595 }
7713 } 6596 }
7714 6597
7715 /* create default unbound and ordered 6598 /* create default unbound and ordered wq attrs */
7716 for (i = 0; i < NR_STD_WORKER_POOLS; 6599 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
7717 struct workqueue_attrs *attrs 6600 struct workqueue_attrs *attrs;
7718 6601
7719 BUG_ON(!(attrs = alloc_workqu 6602 BUG_ON(!(attrs = alloc_workqueue_attrs()));
7720 attrs->nice = std_nice[i]; 6603 attrs->nice = std_nice[i];
7721 unbound_std_wq_attrs[i] = att 6604 unbound_std_wq_attrs[i] = attrs;
7722 6605
7723 /* 6606 /*
7724 * An ordered wq should have 6607 * An ordered wq should have only one pwq as ordering is
7725 * guaranteed by max_active w 6608 * guaranteed by max_active which is enforced by pwqs.
7726 */ 6609 */
7727 BUG_ON(!(attrs = alloc_workqu 6610 BUG_ON(!(attrs = alloc_workqueue_attrs()));
7728 attrs->nice = std_nice[i]; 6611 attrs->nice = std_nice[i];
7729 attrs->ordered = true; 6612 attrs->ordered = true;
7730 ordered_wq_attrs[i] = attrs; 6613 ordered_wq_attrs[i] = attrs;
7731 } 6614 }
7732 6615
7733 system_wq = alloc_workqueue("events", 6616 system_wq = alloc_workqueue("events", 0, 0);
7734 system_highpri_wq = alloc_workqueue(" 6617 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
7735 system_long_wq = alloc_workqueue("eve 6618 system_long_wq = alloc_workqueue("events_long", 0, 0);
7736 system_unbound_wq = alloc_workqueue(" 6619 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
7737 W 6620 WQ_MAX_ACTIVE);
7738 system_freezable_wq = alloc_workqueue 6621 system_freezable_wq = alloc_workqueue("events_freezable",
7739 6622 WQ_FREEZABLE, 0);
7740 system_power_efficient_wq = alloc_wor 6623 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
7741 6624 WQ_POWER_EFFICIENT, 0);
7742 system_freezable_power_efficient_wq = !! 6625 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
7743 6626 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
7744 6627 0);
7745 system_bh_wq = alloc_workqueue("event <<
7746 system_bh_highpri_wq = alloc_workqueu <<
7747 <<
7748 BUG_ON(!system_wq || !system_highpri_ 6628 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
7749 !system_unbound_wq || !system_ 6629 !system_unbound_wq || !system_freezable_wq ||
7750 !system_power_efficient_wq || 6630 !system_power_efficient_wq ||
7751 !system_freezable_power_effici !! 6631 !system_freezable_power_efficient_wq);
7752 !system_bh_wq || !system_bh_hi <<
7753 } 6632 }
7754 6633
7755 static void __init wq_cpu_intensive_thresh_in 6634 static void __init wq_cpu_intensive_thresh_init(void)
7756 { 6635 {
7757 unsigned long thresh; 6636 unsigned long thresh;
7758 unsigned long bogo; 6637 unsigned long bogo;
7759 6638
7760 pwq_release_worker = kthread_create_w 6639 pwq_release_worker = kthread_create_worker(0, "pool_workqueue_release");
7761 BUG_ON(IS_ERR(pwq_release_worker)); 6640 BUG_ON(IS_ERR(pwq_release_worker));
7762 6641
7763 /* if the user set it to a specific v 6642 /* if the user set it to a specific value, keep it */
7764 if (wq_cpu_intensive_thresh_us != ULO 6643 if (wq_cpu_intensive_thresh_us != ULONG_MAX)
7765 return; 6644 return;
7766 6645
7767 /* 6646 /*
7768 * The default of 10ms is derived fro 6647 * The default of 10ms is derived from the fact that most modern (as of
7769 * 2023) processors can do a lot in 1 6648 * 2023) processors can do a lot in 10ms and that it's just below what
7770 * most consider human-perceivable. H 6649 * most consider human-perceivable. However, the kernel also runs on a
7771 * lot slower CPUs including microcon 6650 * lot slower CPUs including microcontrollers where the threshold is way
7772 * too low. 6651 * too low.
7773 * 6652 *
7774 * Let's scale up the threshold upto 6653 * Let's scale up the threshold upto 1 second if BogoMips is below 4000.
7775 * This is by no means accurate but i 6654 * This is by no means accurate but it doesn't have to be. The mechanism
7776 * is still useful even when the thre 6655 * is still useful even when the threshold is fully scaled up. Also, as
7777 * the reports would usually be appli 6656 * the reports would usually be applicable to everyone, some machines
7778 * operating on longer thresholds won 6657 * operating on longer thresholds won't significantly diminish their
7779 * usefulness. 6658 * usefulness.
7780 */ 6659 */
7781 thresh = 10 * USEC_PER_MSEC; 6660 thresh = 10 * USEC_PER_MSEC;
7782 6661
7783 /* see init/calibrate.c for lpj -> Bo 6662 /* see init/calibrate.c for lpj -> BogoMIPS calculation */
7784 bogo = max_t(unsigned long, loops_per 6663 bogo = max_t(unsigned long, loops_per_jiffy / 500000 * HZ, 1);
7785 if (bogo < 4000) 6664 if (bogo < 4000)
7786 thresh = min_t(unsigned long, 6665 thresh = min_t(unsigned long, thresh * 4000 / bogo, USEC_PER_SEC);
7787 6666
7788 pr_debug("wq_cpu_intensive_thresh: lp 6667 pr_debug("wq_cpu_intensive_thresh: lpj=%lu BogoMIPS=%lu thresh_us=%lu\n",
7789 loops_per_jiffy, bogo, thres 6668 loops_per_jiffy, bogo, thresh);
7790 6669
7791 wq_cpu_intensive_thresh_us = thresh; 6670 wq_cpu_intensive_thresh_us = thresh;
7792 } 6671 }
7793 6672
7794 /** 6673 /**
7795 * workqueue_init - bring workqueue subsystem 6674 * workqueue_init - bring workqueue subsystem fully online
7796 * 6675 *
7797 * This is the second step of three-staged wo 6676 * This is the second step of three-staged workqueue subsystem initialization
7798 * and invoked as soon as kthreads can be cre 6677 * and invoked as soon as kthreads can be created and scheduled. Workqueues have
7799 * been created and work items queued on them 6678 * been created and work items queued on them, but there are no kworkers
7800 * executing the work items yet. Populate the 6679 * executing the work items yet. Populate the worker pools with the initial
7801 * workers and enable future kworker creation 6680 * workers and enable future kworker creations.
7802 */ 6681 */
7803 void __init workqueue_init(void) 6682 void __init workqueue_init(void)
7804 { 6683 {
7805 struct workqueue_struct *wq; 6684 struct workqueue_struct *wq;
7806 struct worker_pool *pool; 6685 struct worker_pool *pool;
7807 int cpu, bkt; 6686 int cpu, bkt;
7808 6687
7809 wq_cpu_intensive_thresh_init(); 6688 wq_cpu_intensive_thresh_init();
7810 6689
7811 mutex_lock(&wq_pool_mutex); 6690 mutex_lock(&wq_pool_mutex);
7812 6691
7813 /* 6692 /*
7814 * Per-cpu pools created earlier coul 6693 * Per-cpu pools created earlier could be missing node hint. Fix them
7815 * up. Also, create a rescuer for wor 6694 * up. Also, create a rescuer for workqueues that requested it.
7816 */ 6695 */
7817 for_each_possible_cpu(cpu) { 6696 for_each_possible_cpu(cpu) {
7818 for_each_bh_worker_pool(pool, !! 6697 for_each_cpu_worker_pool(pool, cpu) {
7819 pool->node = cpu_to_n <<
7820 for_each_cpu_worker_pool(pool <<
7821 pool->node = cpu_to_n 6698 pool->node = cpu_to_node(cpu);
>> 6699 }
7822 } 6700 }
7823 6701
7824 list_for_each_entry(wq, &workqueues, 6702 list_for_each_entry(wq, &workqueues, list) {
7825 WARN(init_rescuer(wq), 6703 WARN(init_rescuer(wq),
7826 "workqueue: failed to cr 6704 "workqueue: failed to create early rescuer for %s",
7827 wq->name); 6705 wq->name);
7828 } 6706 }
7829 6707
7830 mutex_unlock(&wq_pool_mutex); 6708 mutex_unlock(&wq_pool_mutex);
7831 6709
7832 /* !! 6710 /* create the initial workers */
7833 * Create the initial workers. A BH p <<
7834 * represents the shared BH execution <<
7835 * affected by hotplug events. Create <<
7836 * possible CPUs here. <<
7837 */ <<
7838 for_each_possible_cpu(cpu) <<
7839 for_each_bh_worker_pool(pool, <<
7840 BUG_ON(!create_worker <<
7841 <<
7842 for_each_online_cpu(cpu) { 6711 for_each_online_cpu(cpu) {
7843 for_each_cpu_worker_pool(pool 6712 for_each_cpu_worker_pool(pool, cpu) {
7844 pool->flags &= ~POOL_ 6713 pool->flags &= ~POOL_DISASSOCIATED;
7845 BUG_ON(!create_worker 6714 BUG_ON(!create_worker(pool));
7846 } 6715 }
7847 } 6716 }
7848 6717
7849 hash_for_each(unbound_pool_hash, bkt, 6718 hash_for_each(unbound_pool_hash, bkt, pool, hash_node)
7850 BUG_ON(!create_worker(pool)); 6719 BUG_ON(!create_worker(pool));
7851 6720
7852 wq_online = true; 6721 wq_online = true;
7853 wq_watchdog_init(); 6722 wq_watchdog_init();
7854 } 6723 }
7855 6724
7856 /* 6725 /*
7857 * Initialize @pt by first initializing @pt-> 6726 * Initialize @pt by first initializing @pt->cpu_pod[] with pod IDs according to
7858 * @cpu_shares_pod(). Each subset of CPUs tha 6727 * @cpu_shares_pod(). Each subset of CPUs that share a pod is assigned a unique
7859 * and consecutive pod ID. The rest of @pt is 6728 * and consecutive pod ID. The rest of @pt is initialized accordingly.
7860 */ 6729 */
7861 static void __init init_pod_type(struct wq_po 6730 static void __init init_pod_type(struct wq_pod_type *pt,
7862 bool (*cpus_ 6731 bool (*cpus_share_pod)(int, int))
7863 { 6732 {
7864 int cur, pre, cpu, pod; 6733 int cur, pre, cpu, pod;
7865 6734
7866 pt->nr_pods = 0; 6735 pt->nr_pods = 0;
7867 6736
7868 /* init @pt->cpu_pod[] according to @ 6737 /* init @pt->cpu_pod[] according to @cpus_share_pod() */
7869 pt->cpu_pod = kcalloc(nr_cpu_ids, siz 6738 pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL);
7870 BUG_ON(!pt->cpu_pod); 6739 BUG_ON(!pt->cpu_pod);
7871 6740
7872 for_each_possible_cpu(cur) { 6741 for_each_possible_cpu(cur) {
7873 for_each_possible_cpu(pre) { 6742 for_each_possible_cpu(pre) {
7874 if (pre >= cur) { 6743 if (pre >= cur) {
7875 pt->cpu_pod[c 6744 pt->cpu_pod[cur] = pt->nr_pods++;
7876 break; 6745 break;
7877 } 6746 }
7878 if (cpus_share_pod(cu 6747 if (cpus_share_pod(cur, pre)) {
7879 pt->cpu_pod[c 6748 pt->cpu_pod[cur] = pt->cpu_pod[pre];
7880 break; 6749 break;
7881 } 6750 }
7882 } 6751 }
7883 } 6752 }
7884 6753
7885 /* init the rest to match @pt->cpu_po 6754 /* init the rest to match @pt->cpu_pod[] */
7886 pt->pod_cpus = kcalloc(pt->nr_pods, s 6755 pt->pod_cpus = kcalloc(pt->nr_pods, sizeof(pt->pod_cpus[0]), GFP_KERNEL);
7887 pt->pod_node = kcalloc(pt->nr_pods, s 6756 pt->pod_node = kcalloc(pt->nr_pods, sizeof(pt->pod_node[0]), GFP_KERNEL);
7888 BUG_ON(!pt->pod_cpus || !pt->pod_node 6757 BUG_ON(!pt->pod_cpus || !pt->pod_node);
7889 6758
7890 for (pod = 0; pod < pt->nr_pods; pod+ 6759 for (pod = 0; pod < pt->nr_pods; pod++)
7891 BUG_ON(!zalloc_cpumask_var(&p 6760 BUG_ON(!zalloc_cpumask_var(&pt->pod_cpus[pod], GFP_KERNEL));
7892 6761
7893 for_each_possible_cpu(cpu) { 6762 for_each_possible_cpu(cpu) {
7894 cpumask_set_cpu(cpu, pt->pod_ 6763 cpumask_set_cpu(cpu, pt->pod_cpus[pt->cpu_pod[cpu]]);
7895 pt->pod_node[pt->cpu_pod[cpu] 6764 pt->pod_node[pt->cpu_pod[cpu]] = cpu_to_node(cpu);
7896 } 6765 }
7897 } 6766 }
7898 6767
7899 static bool __init cpus_dont_share(int cpu0, 6768 static bool __init cpus_dont_share(int cpu0, int cpu1)
7900 { 6769 {
7901 return false; 6770 return false;
7902 } 6771 }
7903 6772
7904 static bool __init cpus_share_smt(int cpu0, i 6773 static bool __init cpus_share_smt(int cpu0, int cpu1)
7905 { 6774 {
7906 #ifdef CONFIG_SCHED_SMT 6775 #ifdef CONFIG_SCHED_SMT
7907 return cpumask_test_cpu(cpu0, cpu_smt 6776 return cpumask_test_cpu(cpu0, cpu_smt_mask(cpu1));
7908 #else 6777 #else
7909 return false; 6778 return false;
7910 #endif 6779 #endif
7911 } 6780 }
7912 6781
7913 static bool __init cpus_share_numa(int cpu0, 6782 static bool __init cpus_share_numa(int cpu0, int cpu1)
7914 { 6783 {
7915 return cpu_to_node(cpu0) == cpu_to_no 6784 return cpu_to_node(cpu0) == cpu_to_node(cpu1);
7916 } 6785 }
7917 6786
7918 /** 6787 /**
7919 * workqueue_init_topology - initialize CPU p 6788 * workqueue_init_topology - initialize CPU pods for unbound workqueues
7920 * 6789 *
7921 * This is the third step of three-staged wor !! 6790 * This is the third step of there-staged workqueue subsystem initialization and
7922 * invoked after SMP and topology information 6791 * invoked after SMP and topology information are fully initialized. It
7923 * initializes the unbound CPU pods according 6792 * initializes the unbound CPU pods accordingly.
7924 */ 6793 */
7925 void __init workqueue_init_topology(void) 6794 void __init workqueue_init_topology(void)
7926 { 6795 {
7927 struct workqueue_struct *wq; 6796 struct workqueue_struct *wq;
7928 int cpu; 6797 int cpu;
7929 6798
7930 init_pod_type(&wq_pod_types[WQ_AFFN_C 6799 init_pod_type(&wq_pod_types[WQ_AFFN_CPU], cpus_dont_share);
7931 init_pod_type(&wq_pod_types[WQ_AFFN_S 6800 init_pod_type(&wq_pod_types[WQ_AFFN_SMT], cpus_share_smt);
7932 init_pod_type(&wq_pod_types[WQ_AFFN_C 6801 init_pod_type(&wq_pod_types[WQ_AFFN_CACHE], cpus_share_cache);
7933 init_pod_type(&wq_pod_types[WQ_AFFN_N 6802 init_pod_type(&wq_pod_types[WQ_AFFN_NUMA], cpus_share_numa);
7934 6803
7935 wq_topo_initialized = true; <<
7936 <<
7937 mutex_lock(&wq_pool_mutex); 6804 mutex_lock(&wq_pool_mutex);
7938 6805
7939 /* 6806 /*
7940 * Workqueues allocated earlier would 6807 * Workqueues allocated earlier would have all CPUs sharing the default
7941 * worker pool. Explicitly call unbou !! 6808 * worker pool. Explicitly call wq_update_pod() on all workqueue and CPU
7942 * and CPU combinations to apply per- !! 6809 * combinations to apply per-pod sharing.
7943 */ 6810 */
7944 list_for_each_entry(wq, &workqueues, 6811 list_for_each_entry(wq, &workqueues, list) {
7945 for_each_online_cpu(cpu) !! 6812 for_each_online_cpu(cpu) {
7946 unbound_wq_update_pwq !! 6813 wq_update_pod(wq, cpu, cpu, true);
7947 if (wq->flags & WQ_UNBOUND) { <<
7948 mutex_lock(&wq->mutex <<
7949 wq_update_node_max_ac <<
7950 mutex_unlock(&wq->mut <<
7951 } 6814 }
7952 } 6815 }
7953 6816
7954 mutex_unlock(&wq_pool_mutex); 6817 mutex_unlock(&wq_pool_mutex);
7955 } 6818 }
7956 6819
7957 void __warn_flushing_systemwide_wq(void) 6820 void __warn_flushing_systemwide_wq(void)
7958 { 6821 {
7959 pr_warn("WARNING: Flushing system-wid 6822 pr_warn("WARNING: Flushing system-wide workqueues will be prohibited in near future.\n");
7960 dump_stack(); 6823 dump_stack();
7961 } 6824 }
7962 EXPORT_SYMBOL(__warn_flushing_systemwide_wq); 6825 EXPORT_SYMBOL(__warn_flushing_systemwide_wq);
7963 6826
7964 static int __init workqueue_unbound_cpus_setu 6827 static int __init workqueue_unbound_cpus_setup(char *str)
7965 { 6828 {
7966 if (cpulist_parse(str, &wq_cmdline_cp 6829 if (cpulist_parse(str, &wq_cmdline_cpumask) < 0) {
7967 cpumask_clear(&wq_cmdline_cpu 6830 cpumask_clear(&wq_cmdline_cpumask);
7968 pr_warn("workqueue.unbound_cp 6831 pr_warn("workqueue.unbound_cpus: incorrect CPU range, using default\n");
7969 } 6832 }
7970 6833
7971 return 1; 6834 return 1;
7972 } 6835 }
7973 __setup("workqueue.unbound_cpus=", workqueue_ 6836 __setup("workqueue.unbound_cpus=", workqueue_unbound_cpus_setup);
7974 6837
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.