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> <<
54 #include <linux/nmi.h> 52 #include <linux/nmi.h>
55 #include <linux/kvm_para.h> <<
56 #include <linux/delay.h> <<
57 #include <linux/irq_work.h> <<
58 53
59 #include "workqueue_internal.h" 54 #include "workqueue_internal.h"
60 55
61 enum worker_pool_flags { !! 56 enum {
62 /* 57 /*
63 * worker_pool flags 58 * worker_pool flags
64 * 59 *
65 * A bound pool is either associated o 60 * A bound pool is either associated or disassociated with its CPU.
66 * While associated (!DISASSOCIATED), 61 * While associated (!DISASSOCIATED), all workers are bound to the
67 * CPU and none has %WORKER_UNBOUND se 62 * CPU and none has %WORKER_UNBOUND set and concurrency management
68 * is in effect. 63 * is in effect.
69 * 64 *
70 * While DISASSOCIATED, the cpu may be 65 * While DISASSOCIATED, the cpu may be offline and all workers have
71 * %WORKER_UNBOUND set and concurrency 66 * %WORKER_UNBOUND set and concurrency management disabled, and may
72 * be executing on any CPU. The pool 67 * be executing on any CPU. The pool behaves as an unbound one.
73 * 68 *
74 * Note that DISASSOCIATED should be f 69 * Note that DISASSOCIATED should be flipped only while holding
75 * wq_pool_attach_mutex to avoid chang 70 * wq_pool_attach_mutex to avoid changing binding state while
76 * worker_attach_to_pool() is in progr 71 * 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 */ 72 */
81 POOL_BH = 1 << 0, !! 73 POOL_MANAGER_ACTIVE = 1 << 0, /* being managed */
82 POOL_MANAGER_ACTIVE = 1 << 1, <<
83 POOL_DISASSOCIATED = 1 << 2, 74 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
84 POOL_BH_DRAINING = 1 << 3, <<
85 }; <<
86 75
87 enum worker_flags { <<
88 /* worker flags */ 76 /* worker flags */
89 WORKER_DIE = 1 << 1, 77 WORKER_DIE = 1 << 1, /* die die die */
90 WORKER_IDLE = 1 << 2, 78 WORKER_IDLE = 1 << 2, /* is idle */
91 WORKER_PREP = 1 << 3, 79 WORKER_PREP = 1 << 3, /* preparing to run works */
92 WORKER_CPU_INTENSIVE = 1 << 6, 80 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
93 WORKER_UNBOUND = 1 << 7, 81 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
94 WORKER_REBOUND = 1 << 8, 82 WORKER_REBOUND = 1 << 8, /* worker was rebound */
95 83
96 WORKER_NOT_RUNNING = WORKER_PREP 84 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
97 WORKER_UNBOU 85 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 86
105 enum wq_internal_consts { <<
106 NR_STD_WORKER_POOLS = 2, 87 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
107 88
108 UNBOUND_POOL_HASH_ORDER = 6, 89 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
109 BUSY_WORKER_HASH_ORDER = 6, 90 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
110 91
111 MAX_IDLE_WORKERS_RATIO = 4, 92 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
112 IDLE_WORKER_TIMEOUT = 300 * HZ, 93 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
113 94
114 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 95 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
115 96 /* call for help after 10ms
116 97 (min two ticks) */
117 MAYDAY_INTERVAL = HZ / 10, 98 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
118 CREATE_COOLDOWN = HZ, 99 CREATE_COOLDOWN = HZ, /* time to breath after fail */
119 100
120 /* 101 /*
121 * Rescue workers are used only on eme 102 * Rescue workers are used only on emergencies and shared by
122 * all cpus. Give MIN_NICE. 103 * all cpus. Give MIN_NICE.
123 */ 104 */
124 RESCUER_NICE_LEVEL = MIN_NICE, 105 RESCUER_NICE_LEVEL = MIN_NICE,
125 HIGHPRI_NICE_LEVEL = MIN_NICE, 106 HIGHPRI_NICE_LEVEL = MIN_NICE,
126 107
127 WQ_NAME_LEN = 32, !! 108 WQ_NAME_LEN = 24,
128 WORKER_ID_LEN = 10 + WQ_NAME <<
129 }; 109 };
130 110
131 /* 111 /*
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 112 * Structure fields follow one of the following exclusion rules.
141 * 113 *
142 * I: Modifiable by initialization/destruction 114 * I: Modifiable by initialization/destruction paths and read-only for
143 * everyone else. 115 * everyone else.
144 * 116 *
145 * P: Preemption protected. Disabling preempt 117 * P: Preemption protected. Disabling preemption is enough and should
146 * only be modified and accessed from the l 118 * only be modified and accessed from the local cpu.
147 * 119 *
148 * L: pool->lock protected. Access with pool- 120 * L: pool->lock protected. Access with pool->lock held.
149 * 121 *
150 * LN: pool->lock and wq_node_nr_active->lock !! 122 * X: During normal operation, modification requires pool->lock and should
151 * reads. !! 123 * be done only from local cpu. Either disabling preemption on local
152 * !! 124 * cpu or grabbing pool->lock is enough for read access. If
153 * K: Only modified by worker while holding po !! 125 * POOL_DISASSOCIATED is set, it's identical to L.
154 * self, while holding pool->lock or from I <<
155 * kworker. <<
156 * <<
157 * S: Only modified by worker self. <<
158 * 126 *
159 * A: wq_pool_attach_mutex protected. 127 * A: wq_pool_attach_mutex protected.
160 * 128 *
161 * PL: wq_pool_mutex protected. 129 * PL: wq_pool_mutex protected.
162 * 130 *
163 * PR: wq_pool_mutex protected for writes. RC 131 * PR: wq_pool_mutex protected for writes. RCU protected for reads.
164 * 132 *
165 * PW: wq_pool_mutex and wq->mutex protected f 133 * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads.
166 * 134 *
167 * PWR: wq_pool_mutex and wq->mutex protected 135 * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or
168 * RCU for reads. 136 * RCU for reads.
169 * 137 *
170 * WQ: wq->mutex protected. 138 * WQ: wq->mutex protected.
171 * 139 *
172 * WR: wq->mutex protected for writes. RCU pr 140 * WR: wq->mutex protected for writes. RCU protected for reads.
173 * 141 *
174 * WO: wq->mutex protected for writes. Updated <<
175 * with READ_ONCE() without locking. <<
176 * <<
177 * MD: wq_mayday_lock protected. 142 * MD: wq_mayday_lock protected.
178 * <<
179 * WD: Used internally by the watchdog. <<
180 */ 143 */
181 144
182 /* struct worker is defined in workqueue_inter 145 /* struct worker is defined in workqueue_internal.h */
183 146
184 struct worker_pool { 147 struct worker_pool {
185 raw_spinlock_t lock; 148 raw_spinlock_t lock; /* the pool lock */
186 int cpu; 149 int cpu; /* I: the associated cpu */
187 int node; 150 int node; /* I: the associated node ID */
188 int id; 151 int id; /* I: pool ID */
189 unsigned int flags; !! 152 unsigned int flags; /* X: flags */
190 153
191 unsigned long watchdog_ts; 154 unsigned long watchdog_ts; /* L: watchdog timestamp */
192 bool cpu_stall; <<
193 <<
194 /* <<
195 * The counter is incremented in a pro <<
196 * w/ preemption disabled, and decreme <<
197 * but w/ pool->lock held. The readers <<
198 * guaranteed to see if the counter re <<
199 */ <<
200 int nr_running; <<
201 155
202 struct list_head worklist; 156 struct list_head worklist; /* L: list of pending works */
203 157
204 int nr_workers; 158 int nr_workers; /* L: total number of workers */
205 int nr_idle; 159 int nr_idle; /* L: currently idle workers */
206 160
207 struct list_head idle_list; !! 161 struct list_head idle_list; /* X: list of idle workers */
208 struct timer_list idle_timer; 162 struct timer_list idle_timer; /* L: worker idle timeout */
209 struct work_struct idle_cull_work !! 163 struct timer_list mayday_timer; /* L: SOS timer for workers */
210 <<
211 struct timer_list mayday_timer; <<
212 164
213 /* a workers is either on busy_hash or 165 /* a workers is either on busy_hash or idle_list, or the manager */
214 DECLARE_HASHTABLE(busy_hash, BUSY_WORK 166 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
215 167 /* L: hash of busy workers */
216 168
217 struct worker *manager; 169 struct worker *manager; /* L: purely informational */
218 struct list_head workers; 170 struct list_head workers; /* A: attached workers */
>> 171 struct completion *detach_completion; /* all workers detached */
219 172
220 struct ida worker_ida; 173 struct ida worker_ida; /* worker IDs for task name */
221 174
222 struct workqueue_attrs *attrs; 175 struct workqueue_attrs *attrs; /* I: worker attributes */
223 struct hlist_node hash_node; 176 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
224 int refcnt; 177 int refcnt; /* PL: refcnt for unbound pools */
225 178
226 /* 179 /*
>> 180 * The current concurrency level. As it's likely to be accessed
>> 181 * from other CPUs during try_to_wake_up(), put it in a separate
>> 182 * cacheline.
>> 183 */
>> 184 atomic_t nr_running ____cacheline_aligned_in_smp;
>> 185
>> 186 /*
227 * Destruction of pool is RCU protecte 187 * Destruction of pool is RCU protected to allow dereferences
228 * from get_work_pool(). 188 * from get_work_pool().
229 */ 189 */
230 struct rcu_head rcu; 190 struct rcu_head rcu;
231 }; !! 191 } ____cacheline_aligned_in_smp;
232 192
233 /* 193 /*
234 * Per-pool_workqueue statistics. These can be !! 194 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
235 * tools/workqueue/wq_monitor.py. <<
236 */ <<
237 enum pool_workqueue_stats { <<
238 PWQ_STAT_STARTED, /* work items <<
239 PWQ_STAT_COMPLETED, /* work items <<
240 PWQ_STAT_CPU_TIME, /* total CPU t <<
241 PWQ_STAT_CPU_INTENSIVE, /* wq_cpu_inte <<
242 PWQ_STAT_CM_WAKEUP, /* concurrency <<
243 PWQ_STAT_REPATRIATED, /* unbound wor <<
244 PWQ_STAT_MAYDAY, /* maydays to <<
245 PWQ_STAT_RESCUED, /* linked work <<
246 <<
247 PWQ_NR_STATS, <<
248 }; <<
249 <<
250 /* <<
251 * The per-pool workqueue. While queued, bits <<
252 * of work_struct->data are used for flags and 195 * of work_struct->data are used for flags and the remaining high bits
253 * point to the pwq; thus, pwqs need to be ali 196 * point to the pwq; thus, pwqs need to be aligned at two's power of the
254 * number of flag bits. 197 * number of flag bits.
255 */ 198 */
256 struct pool_workqueue { 199 struct pool_workqueue {
257 struct worker_pool *pool; 200 struct worker_pool *pool; /* I: the associated pool */
258 struct workqueue_struct *wq; 201 struct workqueue_struct *wq; /* I: the owning workqueue */
259 int work_color; 202 int work_color; /* L: current color */
260 int flush_color; 203 int flush_color; /* L: flushing color */
261 int refcnt; 204 int refcnt; /* L: reference count */
262 int nr_in_flight[W 205 int nr_in_flight[WORK_NR_COLORS];
263 206 /* L: nr of in_flight works */
264 bool plugged; <<
265 <<
266 /* <<
267 * nr_active management and WORK_STRUC <<
268 * <<
269 * When pwq->nr_active >= max_active, <<
270 * pwq->inactive_works instead of pool <<
271 * WORK_STRUCT_INACTIVE. <<
272 * <<
273 * All work items marked with WORK_STR <<
274 * nr_active and all work items in pwq <<
275 * WORK_STRUCT_INACTIVE. But not all W <<
276 * in pwq->inactive_works. Some of the <<
277 * pool->worklist or worker->scheduled <<
278 * wq_barrier which is used for flush_ <<
279 * in nr_active. For non-barrier work <<
280 * WORK_STRUCT_INACTIVE iff it is in p <<
281 */ <<
282 int nr_active; 207 int nr_active; /* L: nr of active works */
283 struct list_head inactive_works !! 208 int max_active; /* L: max active works */
284 struct list_head pending_node; !! 209 struct list_head delayed_works; /* L: delayed works */
285 struct list_head pwqs_node; 210 struct list_head pwqs_node; /* WR: node on wq->pwqs */
286 struct list_head mayday_node; 211 struct list_head mayday_node; /* MD: node on wq->maydays */
287 212
288 u64 stats[PWQ_NR_S <<
289 <<
290 /* 213 /*
291 * Release of unbound pwq is punted to !! 214 * Release of unbound pwq is punted to system_wq. See put_pwq()
292 * and pwq_release_workfn() for detail !! 215 * and pwq_unbound_release_workfn() for details. pool_workqueue
293 * RCU protected so that the first pwq !! 216 * itself is also RCU protected so that the first pwq can be
294 * grabbing wq->mutex. !! 217 * determined without grabbing wq->mutex.
295 */ 218 */
296 struct kthread_work release_work; !! 219 struct work_struct unbound_release_work;
297 struct rcu_head rcu; 220 struct rcu_head rcu;
298 } __aligned(1 << WORK_STRUCT_PWQ_SHIFT); !! 221 } __aligned(1 << WORK_STRUCT_FLAG_BITS);
299 222
300 /* 223 /*
301 * Structure used to wait for workqueue flush. 224 * Structure used to wait for workqueue flush.
302 */ 225 */
303 struct wq_flusher { 226 struct wq_flusher {
304 struct list_head list; 227 struct list_head list; /* WQ: list of flushers */
305 int flush_color; 228 int flush_color; /* WQ: flush color waiting for */
306 struct completion done; 229 struct completion done; /* flush completion */
307 }; 230 };
308 231
309 struct wq_device; 232 struct wq_device;
310 233
311 /* 234 /*
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 235 * The externally visible workqueue. It relays the issued work items to
333 * the appropriate worker_pool through its poo 236 * the appropriate worker_pool through its pool_workqueues.
334 */ 237 */
335 struct workqueue_struct { 238 struct workqueue_struct {
336 struct list_head pwqs; 239 struct list_head pwqs; /* WR: all pwqs of this wq */
337 struct list_head list; 240 struct list_head list; /* PR: list of all workqueues */
338 241
339 struct mutex mutex; 242 struct mutex mutex; /* protects this wq */
340 int work_color; 243 int work_color; /* WQ: current work color */
341 int flush_color; 244 int flush_color; /* WQ: current flush color */
342 atomic_t nr_pwqs_to_flu 245 atomic_t nr_pwqs_to_flush; /* flush in progress */
343 struct wq_flusher *first_flusher 246 struct wq_flusher *first_flusher; /* WQ: first flusher */
344 struct list_head flusher_queue; 247 struct list_head flusher_queue; /* WQ: flush waiters */
345 struct list_head flusher_overfl 248 struct list_head flusher_overflow; /* WQ: flush overflow list */
346 249
347 struct list_head maydays; 250 struct list_head maydays; /* MD: pwqs requesting rescue */
348 struct worker *rescuer; 251 struct worker *rescuer; /* MD: rescue worker */
349 252
350 int nr_drainers; 253 int nr_drainers; /* WQ: drain in progress */
351 !! 254 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 <<
357 255
358 struct workqueue_attrs *unbound_attrs 256 struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */
359 struct pool_workqueue __rcu *dfl_pwq; !! 257 struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */
360 258
361 #ifdef CONFIG_SYSFS 259 #ifdef CONFIG_SYSFS
362 struct wq_device *wq_dev; 260 struct wq_device *wq_dev; /* I: for sysfs interface */
363 #endif 261 #endif
364 #ifdef CONFIG_LOCKDEP 262 #ifdef CONFIG_LOCKDEP
365 char *lock_name; 263 char *lock_name;
366 struct lock_class_key key; 264 struct lock_class_key key;
367 struct lockdep_map lockdep_map; 265 struct lockdep_map lockdep_map;
368 #endif 266 #endif
369 char name[WQ_NAME_L 267 char name[WQ_NAME_LEN]; /* I: workqueue name */
370 268
371 /* 269 /*
372 * Destruction of workqueue_struct is 270 * Destruction of workqueue_struct is RCU protected to allow walking
373 * the workqueues list without grabbin 271 * the workqueues list without grabbing wq_pool_mutex.
374 * This is used to dump all workqueues 272 * This is used to dump all workqueues from sysrq.
375 */ 273 */
376 struct rcu_head rcu; 274 struct rcu_head rcu;
377 275
378 /* hot fields used during command issu 276 /* hot fields used during command issue, aligned to cacheline */
379 unsigned int flags ____cach 277 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
380 struct pool_workqueue __rcu * __percpu !! 278 struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
381 struct wq_node_nr_active *node_nr_acti !! 279 struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */
382 }; 280 };
383 281
384 /* !! 282 static struct kmem_cache *pwq_cache;
385 * Each pod type describes how CPUs should be <<
386 * See the comment above workqueue_attrs->affn <<
387 */ <<
388 struct wq_pod_type { <<
389 int nr_pods; <<
390 cpumask_var_t *pod_cpus; <<
391 int *pod_node; <<
392 int *cpu_pod; <<
393 }; <<
394 283
395 struct work_offq_data { !! 284 static cpumask_var_t *wq_numa_possible_cpumask;
396 u32 pool_id; !! 285 /* possible CPUs of each node */
397 u32 disable; <<
398 u32 flags; <<
399 }; <<
400 286
401 static const char *wq_affn_names[WQ_AFFN_NR_TY !! 287 static bool wq_disable_numa;
402 [WQ_AFFN_DFL] = "default", !! 288 module_param_named(disable_numa, wq_disable_numa, bool, 0444);
403 [WQ_AFFN_CPU] = "cpu", <<
404 [WQ_AFFN_SMT] = "smt", <<
405 [WQ_AFFN_CACHE] = "cache", <<
406 [WQ_AFFN_NUMA] = "numa", <<
407 [WQ_AFFN_SYSTEM] = "system", <<
408 }; <<
409 <<
410 /* <<
411 * Per-cpu work items which run for longer tha <<
412 * automatically considered CPU intensive and <<
413 * management to prevent them from noticeably <<
414 * ULONG_MAX indicates that the user hasn't ov <<
415 * The actual value is initialized in wq_cpu_i <<
416 */ <<
417 static unsigned long wq_cpu_intensive_thresh_u <<
418 module_param_named(cpu_intensive_thresh_us, wq <<
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 <<
423 289
424 /* see the comment above the definition of WQ_ 290 /* see the comment above the definition of WQ_POWER_EFFICIENT */
425 static bool wq_power_efficient = IS_ENABLED(CO 291 static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
426 module_param_named(power_efficient, wq_power_e 292 module_param_named(power_efficient, wq_power_efficient, bool, 0444);
427 293
428 static bool wq_online; /* can 294 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; <<
432 295
433 static struct wq_pod_type wq_pod_types[WQ_AFFN !! 296 static bool wq_numa_enabled; /* unbound NUMA affinity enabled */
434 static enum wq_affn_scope wq_affn_dfl = WQ_AFF <<
435 297
436 /* buf for wq_update_unbound_pod_attrs(), prot !! 298 /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
437 static struct workqueue_attrs *unbound_wq_upda !! 299 static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
438 300
439 static DEFINE_MUTEX(wq_pool_mutex); /* pro 301 static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
440 static DEFINE_MUTEX(wq_pool_attach_mutex); /* 302 static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */
441 static DEFINE_RAW_SPINLOCK(wq_mayday_lock); 303 static DEFINE_RAW_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
442 /* wait for manager to go away */ 304 /* wait for manager to go away */
443 static struct rcuwait manager_wait = __RCUWAIT 305 static struct rcuwait manager_wait = __RCUWAIT_INITIALIZER(manager_wait);
444 306
445 static LIST_HEAD(workqueues); /* PR: 307 static LIST_HEAD(workqueues); /* PR: list of all workqueues */
446 static bool workqueue_freezing; /* PL: 308 static bool workqueue_freezing; /* PL: have wqs started freezing? */
447 309
448 /* PL: mirror the cpu_online_mask excluding th !! 310 /* PL: allowable cpus for unbound wqs and work items */
449 static cpumask_var_t wq_online_cpumask; <<
450 <<
451 /* PL&A: allowable cpus for unbound wqs and wo <<
452 static cpumask_var_t wq_unbound_cpumask; 311 static cpumask_var_t wq_unbound_cpumask;
453 312
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 <<
461 static struct cpumask wq_cmdline_cpumask __ini <<
462 <<
463 /* CPU where unbound work was last round robin 313 /* CPU where unbound work was last round robin scheduled from this CPU */
464 static DEFINE_PER_CPU(int, wq_rr_cpu_last); 314 static DEFINE_PER_CPU(int, wq_rr_cpu_last);
465 315
466 /* 316 /*
467 * Local execution of unbound work items is no 317 * Local execution of unbound work items is no longer guaranteed. The
468 * following always forces round-robin CPU sel 318 * following always forces round-robin CPU selection on unbound work items
469 * to uncover usages which depend on it. 319 * to uncover usages which depend on it.
470 */ 320 */
471 #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU 321 #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
472 static bool wq_debug_force_rr_cpu = true; 322 static bool wq_debug_force_rr_cpu = true;
473 #else 323 #else
474 static bool wq_debug_force_rr_cpu = false; 324 static bool wq_debug_force_rr_cpu = false;
475 #endif 325 #endif
476 module_param_named(debug_force_rr_cpu, wq_debu 326 module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
477 327
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 */ 328 /* the per-cpu worker pools */
487 static DEFINE_PER_CPU_SHARED_ALIGNED(struct wo !! 329 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools);
488 cpu_worke <<
489 330
490 static DEFINE_IDR(worker_pool_idr); /* PR: 331 static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
491 332
492 /* PL: hash of all unbound pools keyed by pool 333 /* PL: hash of all unbound pools keyed by pool->attrs */
493 static DEFINE_HASHTABLE(unbound_pool_hash, UNB 334 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
494 335
495 /* I: attributes used when instantiating stand 336 /* I: attributes used when instantiating standard unbound pools on demand */
496 static struct workqueue_attrs *unbound_std_wq_ 337 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
497 338
498 /* I: attributes used when instantiating order 339 /* I: attributes used when instantiating ordered pools on demand */
499 static struct workqueue_attrs *ordered_wq_attr 340 static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
500 341
501 /* !! 342 struct workqueue_struct *system_wq __read_mostly;
502 * I: kthread_worker to release pwq's. pwq rel <<
503 * process context while holding a pool lock. <<
504 * worker to avoid A-A deadlocks. <<
505 */ <<
506 static struct kthread_worker *pwq_release_work <<
507 <<
508 struct workqueue_struct *system_wq __ro_after_ <<
509 EXPORT_SYMBOL(system_wq); 343 EXPORT_SYMBOL(system_wq);
510 struct workqueue_struct *system_highpri_wq __r !! 344 struct workqueue_struct *system_highpri_wq __read_mostly;
511 EXPORT_SYMBOL_GPL(system_highpri_wq); 345 EXPORT_SYMBOL_GPL(system_highpri_wq);
512 struct workqueue_struct *system_long_wq __ro_a !! 346 struct workqueue_struct *system_long_wq __read_mostly;
513 EXPORT_SYMBOL_GPL(system_long_wq); 347 EXPORT_SYMBOL_GPL(system_long_wq);
514 struct workqueue_struct *system_unbound_wq __r !! 348 struct workqueue_struct *system_unbound_wq __read_mostly;
515 EXPORT_SYMBOL_GPL(system_unbound_wq); 349 EXPORT_SYMBOL_GPL(system_unbound_wq);
516 struct workqueue_struct *system_freezable_wq _ !! 350 struct workqueue_struct *system_freezable_wq __read_mostly;
517 EXPORT_SYMBOL_GPL(system_freezable_wq); 351 EXPORT_SYMBOL_GPL(system_freezable_wq);
518 struct workqueue_struct *system_power_efficien !! 352 struct workqueue_struct *system_power_efficient_wq __read_mostly;
519 EXPORT_SYMBOL_GPL(system_power_efficient_wq); 353 EXPORT_SYMBOL_GPL(system_power_efficient_wq);
520 struct workqueue_struct *system_freezable_powe !! 354 struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
521 EXPORT_SYMBOL_GPL(system_freezable_power_effic 355 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); <<
526 356
527 static int worker_thread(void *__worker); 357 static int worker_thread(void *__worker);
528 static void workqueue_sysfs_unregister(struct 358 static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
529 static void show_pwq(struct pool_workqueue *pw 359 static void show_pwq(struct pool_workqueue *pwq);
530 static void show_one_worker_pool(struct worker <<
531 360
532 #define CREATE_TRACE_POINTS 361 #define CREATE_TRACE_POINTS
533 #include <trace/events/workqueue.h> 362 #include <trace/events/workqueue.h>
534 363
535 #define assert_rcu_or_pool_mutex() 364 #define assert_rcu_or_pool_mutex() \
536 RCU_LOCKDEP_WARN(!rcu_read_lock_any_he !! 365 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
537 !lockdep_is_held(&wq_ 366 !lockdep_is_held(&wq_pool_mutex), \
538 "RCU or wq_pool_mutex 367 "RCU or wq_pool_mutex should be held")
539 368
540 #define assert_rcu_or_wq_mutex_or_pool_mutex(w 369 #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
541 RCU_LOCKDEP_WARN(!rcu_read_lock_any_he !! 370 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
542 !lockdep_is_held(&wq- 371 !lockdep_is_held(&wq->mutex) && \
543 !lockdep_is_held(&wq_ 372 !lockdep_is_held(&wq_pool_mutex), \
544 "RCU, wq->mutex or wq 373 "RCU, wq->mutex or wq_pool_mutex should be held")
545 374
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) 375 #define for_each_cpu_worker_pool(pool, cpu) \
552 for ((pool) = &per_cpu(cpu_worker_pool 376 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
553 (pool) < &per_cpu(cpu_worker_pool 377 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
554 (pool)++) 378 (pool)++)
555 379
556 /** 380 /**
557 * for_each_pool - iterate through all worker_ 381 * for_each_pool - iterate through all worker_pools in the system
558 * @pool: iteration cursor 382 * @pool: iteration cursor
559 * @pi: integer used for iteration 383 * @pi: integer used for iteration
560 * 384 *
561 * This must be called either with wq_pool_mut 385 * This must be called either with wq_pool_mutex held or RCU read
562 * locked. If the pool needs to be used beyon 386 * locked. If the pool needs to be used beyond the locking in effect, the
563 * caller is responsible for guaranteeing that 387 * caller is responsible for guaranteeing that the pool stays online.
564 * 388 *
565 * The if/else clause exists only for the lock 389 * The if/else clause exists only for the lockdep assertion and can be
566 * ignored. 390 * ignored.
567 */ 391 */
568 #define for_each_pool(pool, pi) 392 #define for_each_pool(pool, pi) \
569 idr_for_each_entry(&worker_pool_idr, p 393 idr_for_each_entry(&worker_pool_idr, pool, pi) \
570 if (({ assert_rcu_or_pool_mute 394 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
571 else 395 else
572 396
573 /** 397 /**
574 * for_each_pool_worker - iterate through all 398 * for_each_pool_worker - iterate through all workers of a worker_pool
575 * @worker: iteration cursor 399 * @worker: iteration cursor
576 * @pool: worker_pool to iterate workers of 400 * @pool: worker_pool to iterate workers of
577 * 401 *
578 * This must be called with wq_pool_attach_mut 402 * This must be called with wq_pool_attach_mutex.
579 * 403 *
580 * The if/else clause exists only for the lock 404 * The if/else clause exists only for the lockdep assertion and can be
581 * ignored. 405 * ignored.
582 */ 406 */
583 #define for_each_pool_worker(worker, pool) 407 #define for_each_pool_worker(worker, pool) \
584 list_for_each_entry((worker), &(pool)- 408 list_for_each_entry((worker), &(pool)->workers, node) \
585 if (({ lockdep_assert_held(&wq 409 if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \
586 else 410 else
587 411
588 /** 412 /**
589 * for_each_pwq - iterate through all pool_wor 413 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
590 * @pwq: iteration cursor 414 * @pwq: iteration cursor
591 * @wq: the target workqueue 415 * @wq: the target workqueue
592 * 416 *
593 * This must be called either with wq->mutex h 417 * This must be called either with wq->mutex held or RCU read locked.
594 * If the pwq needs to be used beyond the lock 418 * If the pwq needs to be used beyond the locking in effect, the caller is
595 * responsible for guaranteeing that the pwq s 419 * responsible for guaranteeing that the pwq stays online.
596 * 420 *
597 * The if/else clause exists only for the lock 421 * The if/else clause exists only for the lockdep assertion and can be
598 * ignored. 422 * ignored.
599 */ 423 */
600 #define for_each_pwq(pwq, wq) 424 #define for_each_pwq(pwq, wq) \
601 list_for_each_entry_rcu((pwq), &(wq)-> 425 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \
602 lockdep_is_he 426 lockdep_is_held(&(wq->mutex)))
603 427
604 #ifdef CONFIG_DEBUG_OBJECTS_WORK 428 #ifdef CONFIG_DEBUG_OBJECTS_WORK
605 429
606 static const struct debug_obj_descr work_debug 430 static const struct debug_obj_descr work_debug_descr;
607 431
608 static void *work_debug_hint(void *addr) 432 static void *work_debug_hint(void *addr)
609 { 433 {
610 return ((struct work_struct *) addr)-> 434 return ((struct work_struct *) addr)->func;
611 } 435 }
612 436
613 static bool work_is_static_object(void *addr) 437 static bool work_is_static_object(void *addr)
614 { 438 {
615 struct work_struct *work = addr; 439 struct work_struct *work = addr;
616 440
617 return test_bit(WORK_STRUCT_STATIC_BIT 441 return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work));
618 } 442 }
619 443
620 /* 444 /*
621 * fixup_init is called when: 445 * fixup_init is called when:
622 * - an active object is initialized 446 * - an active object is initialized
623 */ 447 */
624 static bool work_fixup_init(void *addr, enum d 448 static bool work_fixup_init(void *addr, enum debug_obj_state state)
625 { 449 {
626 struct work_struct *work = addr; 450 struct work_struct *work = addr;
627 451
628 switch (state) { 452 switch (state) {
629 case ODEBUG_STATE_ACTIVE: 453 case ODEBUG_STATE_ACTIVE:
630 cancel_work_sync(work); 454 cancel_work_sync(work);
631 debug_object_init(work, &work_ 455 debug_object_init(work, &work_debug_descr);
632 return true; 456 return true;
633 default: 457 default:
634 return false; 458 return false;
635 } 459 }
636 } 460 }
637 461
638 /* 462 /*
639 * fixup_free is called when: 463 * fixup_free is called when:
640 * - an active object is freed 464 * - an active object is freed
641 */ 465 */
642 static bool work_fixup_free(void *addr, enum d 466 static bool work_fixup_free(void *addr, enum debug_obj_state state)
643 { 467 {
644 struct work_struct *work = addr; 468 struct work_struct *work = addr;
645 469
646 switch (state) { 470 switch (state) {
647 case ODEBUG_STATE_ACTIVE: 471 case ODEBUG_STATE_ACTIVE:
648 cancel_work_sync(work); 472 cancel_work_sync(work);
649 debug_object_free(work, &work_ 473 debug_object_free(work, &work_debug_descr);
650 return true; 474 return true;
651 default: 475 default:
652 return false; 476 return false;
653 } 477 }
654 } 478 }
655 479
656 static const struct debug_obj_descr work_debug 480 static const struct debug_obj_descr work_debug_descr = {
657 .name = "work_struct", 481 .name = "work_struct",
658 .debug_hint = work_debug_hint, 482 .debug_hint = work_debug_hint,
659 .is_static_object = work_is_static_obj 483 .is_static_object = work_is_static_object,
660 .fixup_init = work_fixup_init, 484 .fixup_init = work_fixup_init,
661 .fixup_free = work_fixup_free, 485 .fixup_free = work_fixup_free,
662 }; 486 };
663 487
664 static inline void debug_work_activate(struct 488 static inline void debug_work_activate(struct work_struct *work)
665 { 489 {
666 debug_object_activate(work, &work_debu 490 debug_object_activate(work, &work_debug_descr);
667 } 491 }
668 492
669 static inline void debug_work_deactivate(struc 493 static inline void debug_work_deactivate(struct work_struct *work)
670 { 494 {
671 debug_object_deactivate(work, &work_de 495 debug_object_deactivate(work, &work_debug_descr);
672 } 496 }
673 497
674 void __init_work(struct work_struct *work, int 498 void __init_work(struct work_struct *work, int onstack)
675 { 499 {
676 if (onstack) 500 if (onstack)
677 debug_object_init_on_stack(wor 501 debug_object_init_on_stack(work, &work_debug_descr);
678 else 502 else
679 debug_object_init(work, &work_ 503 debug_object_init(work, &work_debug_descr);
680 } 504 }
681 EXPORT_SYMBOL_GPL(__init_work); 505 EXPORT_SYMBOL_GPL(__init_work);
682 506
683 void destroy_work_on_stack(struct work_struct 507 void destroy_work_on_stack(struct work_struct *work)
684 { 508 {
685 debug_object_free(work, &work_debug_de 509 debug_object_free(work, &work_debug_descr);
686 } 510 }
687 EXPORT_SYMBOL_GPL(destroy_work_on_stack); 511 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
688 512
689 void destroy_delayed_work_on_stack(struct dela 513 void destroy_delayed_work_on_stack(struct delayed_work *work)
690 { 514 {
691 destroy_timer_on_stack(&work->timer); 515 destroy_timer_on_stack(&work->timer);
692 debug_object_free(&work->work, &work_d 516 debug_object_free(&work->work, &work_debug_descr);
693 } 517 }
694 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stac 518 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
695 519
696 #else 520 #else
697 static inline void debug_work_activate(struct 521 static inline void debug_work_activate(struct work_struct *work) { }
698 static inline void debug_work_deactivate(struc 522 static inline void debug_work_deactivate(struct work_struct *work) { }
699 #endif 523 #endif
700 524
701 /** 525 /**
702 * worker_pool_assign_id - allocate ID and ass !! 526 * worker_pool_assign_id - allocate ID and assing it to @pool
703 * @pool: the pool pointer of interest 527 * @pool: the pool pointer of interest
704 * 528 *
705 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) 529 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
706 * successfully, -errno on failure. 530 * successfully, -errno on failure.
707 */ 531 */
708 static int worker_pool_assign_id(struct worker 532 static int worker_pool_assign_id(struct worker_pool *pool)
709 { 533 {
710 int ret; 534 int ret;
711 535
712 lockdep_assert_held(&wq_pool_mutex); 536 lockdep_assert_held(&wq_pool_mutex);
713 537
714 ret = idr_alloc(&worker_pool_idr, pool 538 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
715 GFP_KERNEL); 539 GFP_KERNEL);
716 if (ret >= 0) { 540 if (ret >= 0) {
717 pool->id = ret; 541 pool->id = ret;
718 return 0; 542 return 0;
719 } 543 }
720 return ret; 544 return ret;
721 } 545 }
722 546
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 /** 547 /**
741 * unbound_effective_cpumask - effective cpuma !! 548 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
742 * @wq: workqueue of interest !! 549 * @wq: the target workqueue
>> 550 * @node: the node ID
>> 551 *
>> 552 * This must be called with any of wq_pool_mutex, wq->mutex or RCU
>> 553 * read locked.
>> 554 * If the pwq needs to be used beyond the locking in effect, the caller is
>> 555 * responsible for guaranteeing that the pwq stays online.
743 * 556 *
744 * @wq->unbound_attrs->cpumask contains the cp !! 557 * Return: The unbound pool_workqueue for @node.
745 * is masked with wq_unbound_cpumask to determ <<
746 * default pwq is always mapped to the pool wi <<
747 */ 558 */
748 static struct cpumask *unbound_effective_cpuma !! 559 static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
>> 560 int node)
749 { 561 {
750 return unbound_pwq(wq, -1)->pool->attr !! 562 assert_rcu_or_wq_mutex_or_pool_mutex(wq);
>> 563
>> 564 /*
>> 565 * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a
>> 566 * delayed item is pending. The plan is to keep CPU -> NODE
>> 567 * mapping valid and stable across CPU on/offlines. Once that
>> 568 * happens, this workaround can be removed.
>> 569 */
>> 570 if (unlikely(node == NUMA_NO_NODE))
>> 571 return wq->dfl_pwq;
>> 572
>> 573 return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
751 } 574 }
752 575
753 static unsigned int work_color_to_flags(int co 576 static unsigned int work_color_to_flags(int color)
754 { 577 {
755 return color << WORK_STRUCT_COLOR_SHIF 578 return color << WORK_STRUCT_COLOR_SHIFT;
756 } 579 }
757 580
758 static int get_work_color(unsigned long work_d !! 581 static int get_work_color(struct work_struct *work)
759 { 582 {
760 return (work_data >> WORK_STRUCT_COLOR !! 583 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
761 ((1 << WORK_STRUCT_COLOR_BITS) 584 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
762 } 585 }
763 586
764 static int work_next_color(int color) 587 static int work_next_color(int color)
765 { 588 {
766 return (color + 1) % WORK_NR_COLORS; 589 return (color + 1) % WORK_NR_COLORS;
767 } 590 }
768 591
769 static unsigned long pool_offq_flags(struct wo <<
770 { <<
771 return (pool->flags & POOL_BH) ? WORK_ <<
772 } <<
773 <<
774 /* 592 /*
775 * While queued, %WORK_STRUCT_PWQ is set and n 593 * 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 594 * contain the pointer to the queued pwq. Once execution starts, the flag
777 * is cleared and the high bits contain OFFQ f 595 * is cleared and the high bits contain OFFQ flags and pool ID.
778 * 596 *
779 * set_work_pwq(), set_work_pool_and_clear_pen !! 597 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
780 * can be used to set the pwq, pool or clear w !! 598 * and clear_work_data() can be used to set the pwq, pool or clear
781 * only be called while the work is owned - ie !! 599 * work->data. These functions should only be called while the work is
>> 600 * owned - ie. while the PENDING bit is set.
782 * 601 *
783 * get_work_pool() and get_work_pwq() can be u 602 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
784 * corresponding to a work. Pool is available 603 * corresponding to a work. Pool is available once the work has been
785 * queued anywhere after initialization until 604 * queued anywhere after initialization until it is sync canceled. pwq is
786 * available only while the work item is queue 605 * available only while the work item is queued.
>> 606 *
>> 607 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
>> 608 * canceled. While being canceled, a work item may have its PENDING set
>> 609 * but stay off timer and worklist for arbitrarily long and nobody should
>> 610 * try to steal the PENDING bit.
787 */ 611 */
788 static inline void set_work_data(struct work_s !! 612 static inline void set_work_data(struct work_struct *work, unsigned long data,
>> 613 unsigned long flags)
789 { 614 {
790 WARN_ON_ONCE(!work_pending(work)); 615 WARN_ON_ONCE(!work_pending(work));
791 atomic_long_set(&work->data, data | wo !! 616 atomic_long_set(&work->data, data | flags | work_static(work));
792 } 617 }
793 618
794 static void set_work_pwq(struct work_struct *w 619 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
795 unsigned long flags) !! 620 unsigned long extra_flags)
796 { 621 {
797 set_work_data(work, (unsigned long)pwq !! 622 set_work_data(work, (unsigned long)pwq,
798 WORK_STRUCT_PWQ | flags) !! 623 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
799 } 624 }
800 625
801 static void set_work_pool_and_keep_pending(str 626 static void set_work_pool_and_keep_pending(struct work_struct *work,
802 int !! 627 int pool_id)
803 { 628 {
804 set_work_data(work, ((unsigned long)po !! 629 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
805 WORK_STRUCT_PENDING | fl !! 630 WORK_STRUCT_PENDING);
806 } 631 }
807 632
808 static void set_work_pool_and_clear_pending(st 633 static void set_work_pool_and_clear_pending(struct work_struct *work,
809 in !! 634 int pool_id)
810 { 635 {
811 /* 636 /*
812 * The following wmb is paired with th 637 * The following wmb is paired with the implied mb in
813 * test_and_set_bit(PENDING) and ensur 638 * test_and_set_bit(PENDING) and ensures all updates to @work made
814 * here are visible to and precede any 639 * here are visible to and precede any updates by the next PENDING
815 * owner. 640 * owner.
816 */ 641 */
817 smp_wmb(); 642 smp_wmb();
818 set_work_data(work, ((unsigned long)po !! 643 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
819 flags); <<
820 /* 644 /*
821 * The following mb guarantees that pr 645 * The following mb guarantees that previous clear of a PENDING bit
822 * will not be reordered with any spec 646 * will not be reordered with any speculative LOADS or STORES from
823 * work->current_func, which is execut 647 * work->current_func, which is executed afterwards. This possible
824 * reordering can lead to a missed exe 648 * reordering can lead to a missed execution on attempt to queue
825 * the same @work. E.g. consider this 649 * the same @work. E.g. consider this case:
826 * 650 *
827 * CPU#0 CPU 651 * CPU#0 CPU#1
828 * ---------------------------- --- 652 * ---------------------------- --------------------------------
829 * 653 *
830 * 1 STORE event_indicated 654 * 1 STORE event_indicated
831 * 2 queue_work_on() { 655 * 2 queue_work_on() {
832 * 3 test_and_set_bit(PENDING) 656 * 3 test_and_set_bit(PENDING)
833 * 4 } set 657 * 4 } set_..._and_clear_pending() {
834 * 5 s 658 * 5 set_work_data() # clear bit
835 * 6 s 659 * 6 smp_mb()
836 * 7 wor 660 * 7 work->current_func() {
837 * 8 661 * 8 LOAD event_indicated
838 * } 662 * }
839 * 663 *
840 * Without an explicit full barrier sp 664 * Without an explicit full barrier speculative LOAD on line 8 can
841 * be executed before CPU#0 does STORE 665 * be executed before CPU#0 does STORE on line 1. If that happens,
842 * CPU#0 observes the PENDING bit is s 666 * CPU#0 observes the PENDING bit is still set and new execution of
843 * a @work is not queued in a hope, th 667 * a @work is not queued in a hope, that CPU#1 will eventually
844 * finish the queued @work. Meanwhile 668 * finish the queued @work. Meanwhile CPU#1 does not see
845 * event_indicated is set, because spe 669 * event_indicated is set, because speculative LOAD was executed
846 * before actual STORE. 670 * before actual STORE.
847 */ 671 */
848 smp_mb(); 672 smp_mb();
849 } 673 }
850 674
851 static inline struct pool_workqueue *work_stru !! 675 static void clear_work_data(struct work_struct *work)
852 { 676 {
853 return (struct pool_workqueue *)(data !! 677 smp_wmb(); /* see set_work_pool_and_clear_pending() */
>> 678 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
854 } 679 }
855 680
856 static struct pool_workqueue *get_work_pwq(str 681 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
857 { 682 {
858 unsigned long data = atomic_long_read( 683 unsigned long data = atomic_long_read(&work->data);
859 684
860 if (data & WORK_STRUCT_PWQ) 685 if (data & WORK_STRUCT_PWQ)
861 return work_struct_pwq(data); !! 686 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
862 else 687 else
863 return NULL; 688 return NULL;
864 } 689 }
865 690
866 /** 691 /**
867 * get_work_pool - return the worker_pool a gi 692 * get_work_pool - return the worker_pool a given work was associated with
868 * @work: the work item of interest 693 * @work: the work item of interest
869 * 694 *
870 * Pools are created and destroyed under wq_po 695 * Pools are created and destroyed under wq_pool_mutex, and allows read
871 * access under RCU read lock. As such, this 696 * access under RCU read lock. As such, this function should be
872 * called under wq_pool_mutex or inside of a r 697 * called under wq_pool_mutex or inside of a rcu_read_lock() region.
873 * 698 *
874 * All fields of the returned pool are accessi 699 * All fields of the returned pool are accessible as long as the above
875 * mentioned locking is in effect. If the ret 700 * mentioned locking is in effect. If the returned pool needs to be used
876 * beyond the critical section, the caller is 701 * beyond the critical section, the caller is responsible for ensuring the
877 * returned pool is and stays online. 702 * returned pool is and stays online.
878 * 703 *
879 * Return: The worker_pool @work was last asso 704 * Return: The worker_pool @work was last associated with. %NULL if none.
880 */ 705 */
881 static struct worker_pool *get_work_pool(struc 706 static struct worker_pool *get_work_pool(struct work_struct *work)
882 { 707 {
883 unsigned long data = atomic_long_read( 708 unsigned long data = atomic_long_read(&work->data);
884 int pool_id; 709 int pool_id;
885 710
886 assert_rcu_or_pool_mutex(); 711 assert_rcu_or_pool_mutex();
887 712
888 if (data & WORK_STRUCT_PWQ) 713 if (data & WORK_STRUCT_PWQ)
889 return work_struct_pwq(data)-> !! 714 return ((struct pool_workqueue *)
>> 715 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
890 716
891 pool_id = data >> WORK_OFFQ_POOL_SHIFT 717 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
892 if (pool_id == WORK_OFFQ_POOL_NONE) 718 if (pool_id == WORK_OFFQ_POOL_NONE)
893 return NULL; 719 return NULL;
894 720
895 return idr_find(&worker_pool_idr, pool 721 return idr_find(&worker_pool_idr, pool_id);
896 } 722 }
897 723
898 static unsigned long shift_and_mask(unsigned l !! 724 /**
>> 725 * get_work_pool_id - return the worker pool ID a given work is associated with
>> 726 * @work: the work item of interest
>> 727 *
>> 728 * Return: The worker_pool ID @work was last associated with.
>> 729 * %WORK_OFFQ_POOL_NONE if none.
>> 730 */
>> 731 static int get_work_pool_id(struct work_struct *work)
899 { 732 {
900 return (v >> shift) & ((1U << bits) - !! 733 unsigned long data = atomic_long_read(&work->data);
>> 734
>> 735 if (data & WORK_STRUCT_PWQ)
>> 736 return ((struct pool_workqueue *)
>> 737 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
>> 738
>> 739 return data >> WORK_OFFQ_POOL_SHIFT;
901 } 740 }
902 741
903 static void work_offqd_unpack(struct work_offq !! 742 static void mark_work_canceling(struct work_struct *work)
904 { 743 {
905 WARN_ON_ONCE(data & WORK_STRUCT_PWQ); !! 744 unsigned long pool_id = get_work_pool_id(work);
906 745
907 offqd->pool_id = shift_and_mask(data, !! 746 pool_id <<= WORK_OFFQ_POOL_SHIFT;
908 WORK_O !! 747 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 } 748 }
913 749
914 static unsigned long work_offqd_pack_flags(str !! 750 static bool work_is_canceling(struct work_struct *work)
915 { 751 {
916 return ((unsigned long)offqd->disable !! 752 unsigned long data = atomic_long_read(&work->data);
917 ((unsigned long)offqd->flags); !! 753
>> 754 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
918 } 755 }
919 756
920 /* 757 /*
921 * Policy functions. These define the policie 758 * Policy functions. These define the policies on how the global worker
922 * pools are managed. Unless noted otherwise, 759 * pools are managed. Unless noted otherwise, these functions assume that
923 * they're being called with pool->lock held. 760 * they're being called with pool->lock held.
924 */ 761 */
925 762
>> 763 static bool __need_more_worker(struct worker_pool *pool)
>> 764 {
>> 765 return !atomic_read(&pool->nr_running);
>> 766 }
>> 767
926 /* 768 /*
927 * Need to wake up a worker? Called from anyt 769 * Need to wake up a worker? Called from anything but currently
928 * running workers. 770 * running workers.
929 * 771 *
930 * Note that, because unbound workers never co 772 * Note that, because unbound workers never contribute to nr_running, this
931 * function will always return %true for unbou 773 * function will always return %true for unbound pools as long as the
932 * worklist isn't empty. 774 * worklist isn't empty.
933 */ 775 */
934 static bool need_more_worker(struct worker_poo 776 static bool need_more_worker(struct worker_pool *pool)
935 { 777 {
936 return !list_empty(&pool->worklist) && !! 778 return !list_empty(&pool->worklist) && __need_more_worker(pool);
937 } 779 }
938 780
939 /* Can I start working? Called from busy but 781 /* Can I start working? Called from busy but !running workers. */
940 static bool may_start_working(struct worker_po 782 static bool may_start_working(struct worker_pool *pool)
941 { 783 {
942 return pool->nr_idle; 784 return pool->nr_idle;
943 } 785 }
944 786
945 /* Do I need to keep working? Called from cur 787 /* Do I need to keep working? Called from currently running workers. */
946 static bool keep_working(struct worker_pool *p 788 static bool keep_working(struct worker_pool *pool)
947 { 789 {
948 return !list_empty(&pool->worklist) && !! 790 return !list_empty(&pool->worklist) &&
>> 791 atomic_read(&pool->nr_running) <= 1;
949 } 792 }
950 793
951 /* Do we need a new worker? Called from manag 794 /* Do we need a new worker? Called from manager. */
952 static bool need_to_create_worker(struct worke 795 static bool need_to_create_worker(struct worker_pool *pool)
953 { 796 {
954 return need_more_worker(pool) && !may_ 797 return need_more_worker(pool) && !may_start_working(pool);
955 } 798 }
956 799
957 /* Do we have too many workers and should some 800 /* Do we have too many workers and should some go away? */
958 static bool too_many_workers(struct worker_poo 801 static bool too_many_workers(struct worker_pool *pool)
959 { 802 {
960 bool managing = pool->flags & POOL_MAN 803 bool managing = pool->flags & POOL_MANAGER_ACTIVE;
961 int nr_idle = pool->nr_idle + managing 804 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
962 int nr_busy = pool->nr_workers - nr_id 805 int nr_busy = pool->nr_workers - nr_idle;
963 806
964 return nr_idle > 2 && (nr_idle - 2) * 807 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
965 } 808 }
966 809
>> 810 /*
>> 811 * Wake up functions.
>> 812 */
>> 813
>> 814 /* Return the first idle worker. Safe with preemption disabled */
>> 815 static struct worker *first_idle_worker(struct worker_pool *pool)
>> 816 {
>> 817 if (unlikely(list_empty(&pool->idle_list)))
>> 818 return NULL;
>> 819
>> 820 return list_first_entry(&pool->idle_list, struct worker, entry);
>> 821 }
>> 822
>> 823 /**
>> 824 * wake_up_worker - wake up an idle worker
>> 825 * @pool: worker pool to wake worker from
>> 826 *
>> 827 * Wake up the first idle worker of @pool.
>> 828 *
>> 829 * CONTEXT:
>> 830 * raw_spin_lock_irq(pool->lock).
>> 831 */
>> 832 static void wake_up_worker(struct worker_pool *pool)
>> 833 {
>> 834 struct worker *worker = first_idle_worker(pool);
>> 835
>> 836 if (likely(worker))
>> 837 wake_up_process(worker->task);
>> 838 }
>> 839
>> 840 /**
>> 841 * wq_worker_running - a worker is running again
>> 842 * @task: task waking up
>> 843 *
>> 844 * This function is called when a worker returns from schedule()
>> 845 */
>> 846 void wq_worker_running(struct task_struct *task)
>> 847 {
>> 848 struct worker *worker = kthread_data(task);
>> 849
>> 850 if (!worker->sleeping)
>> 851 return;
>> 852 if (!(worker->flags & WORKER_NOT_RUNNING))
>> 853 atomic_inc(&worker->pool->nr_running);
>> 854 worker->sleeping = 0;
>> 855 }
>> 856
>> 857 /**
>> 858 * wq_worker_sleeping - a worker is going to sleep
>> 859 * @task: task going to sleep
>> 860 *
>> 861 * This function is called from schedule() when a busy worker is
>> 862 * going to sleep. Preemption needs to be disabled to protect ->sleeping
>> 863 * assignment.
>> 864 */
>> 865 void wq_worker_sleeping(struct task_struct *task)
>> 866 {
>> 867 struct worker *next, *worker = kthread_data(task);
>> 868 struct worker_pool *pool;
>> 869
>> 870 /*
>> 871 * Rescuers, which may not have all the fields set up like normal
>> 872 * workers, also reach here, let's not access anything before
>> 873 * checking NOT_RUNNING.
>> 874 */
>> 875 if (worker->flags & WORKER_NOT_RUNNING)
>> 876 return;
>> 877
>> 878 pool = worker->pool;
>> 879
>> 880 /* Return if preempted before wq_worker_running() was reached */
>> 881 if (worker->sleeping)
>> 882 return;
>> 883
>> 884 worker->sleeping = 1;
>> 885 raw_spin_lock_irq(&pool->lock);
>> 886
>> 887 /*
>> 888 * The counterpart of the following dec_and_test, implied mb,
>> 889 * worklist not empty test sequence is in insert_work().
>> 890 * Please read comment there.
>> 891 *
>> 892 * NOT_RUNNING is clear. This means that we're bound to and
>> 893 * running on the local cpu w/ rq lock held and preemption
>> 894 * disabled, which in turn means that none else could be
>> 895 * manipulating idle_list, so dereferencing idle_list without pool
>> 896 * lock is safe.
>> 897 */
>> 898 if (atomic_dec_and_test(&pool->nr_running) &&
>> 899 !list_empty(&pool->worklist)) {
>> 900 next = first_idle_worker(pool);
>> 901 if (next)
>> 902 wake_up_process(next->task);
>> 903 }
>> 904 raw_spin_unlock_irq(&pool->lock);
>> 905 }
>> 906
>> 907 /**
>> 908 * wq_worker_last_func - retrieve worker's last work function
>> 909 * @task: Task to retrieve last work function of.
>> 910 *
>> 911 * Determine the last function a worker executed. This is called from
>> 912 * the scheduler to get a worker's last known identity.
>> 913 *
>> 914 * CONTEXT:
>> 915 * raw_spin_lock_irq(rq->lock)
>> 916 *
>> 917 * This function is called during schedule() when a kworker is going
>> 918 * to sleep. It's used by psi to identify aggregation workers during
>> 919 * dequeuing, to allow periodic aggregation to shut-off when that
>> 920 * worker is the last task in the system or cgroup to go to sleep.
>> 921 *
>> 922 * As this function doesn't involve any workqueue-related locking, it
>> 923 * only returns stable values when called from inside the scheduler's
>> 924 * queuing and dequeuing paths, when @task, which must be a kworker,
>> 925 * is guaranteed to not be processing any works.
>> 926 *
>> 927 * Return:
>> 928 * The last work function %current executed as a worker, NULL if it
>> 929 * hasn't executed any work yet.
>> 930 */
>> 931 work_func_t wq_worker_last_func(struct task_struct *task)
>> 932 {
>> 933 struct worker *worker = kthread_data(task);
>> 934
>> 935 return worker->last_func;
>> 936 }
>> 937
967 /** 938 /**
968 * worker_set_flags - set worker flags and adj 939 * worker_set_flags - set worker flags and adjust nr_running accordingly
969 * @worker: self 940 * @worker: self
970 * @flags: flags to set 941 * @flags: flags to set
971 * 942 *
972 * Set @flags in @worker->flags and adjust nr_ 943 * Set @flags in @worker->flags and adjust nr_running accordingly.
>> 944 *
>> 945 * CONTEXT:
>> 946 * raw_spin_lock_irq(pool->lock)
973 */ 947 */
974 static inline void worker_set_flags(struct wor 948 static inline void worker_set_flags(struct worker *worker, unsigned int flags)
975 { 949 {
976 struct worker_pool *pool = worker->poo 950 struct worker_pool *pool = worker->pool;
977 951
978 lockdep_assert_held(&pool->lock); !! 952 WARN_ON_ONCE(worker->task != current);
979 953
980 /* If transitioning into NOT_RUNNING, 954 /* If transitioning into NOT_RUNNING, adjust nr_running. */
981 if ((flags & WORKER_NOT_RUNNING) && 955 if ((flags & WORKER_NOT_RUNNING) &&
982 !(worker->flags & WORKER_NOT_RUNNI 956 !(worker->flags & WORKER_NOT_RUNNING)) {
983 pool->nr_running--; !! 957 atomic_dec(&pool->nr_running);
984 } 958 }
985 959
986 worker->flags |= flags; 960 worker->flags |= flags;
987 } 961 }
988 962
989 /** 963 /**
990 * worker_clr_flags - clear worker flags and a 964 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
991 * @worker: self 965 * @worker: self
992 * @flags: flags to clear 966 * @flags: flags to clear
993 * 967 *
994 * Clear @flags in @worker->flags and adjust n 968 * Clear @flags in @worker->flags and adjust nr_running accordingly.
>> 969 *
>> 970 * CONTEXT:
>> 971 * raw_spin_lock_irq(pool->lock)
995 */ 972 */
996 static inline void worker_clr_flags(struct wor 973 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
997 { 974 {
998 struct worker_pool *pool = worker->poo 975 struct worker_pool *pool = worker->pool;
999 unsigned int oflags = worker->flags; 976 unsigned int oflags = worker->flags;
1000 977
1001 lockdep_assert_held(&pool->lock); !! 978 WARN_ON_ONCE(worker->task != current);
1002 979
1003 worker->flags &= ~flags; 980 worker->flags &= ~flags;
1004 981
1005 /* 982 /*
1006 * If transitioning out of NOT_RUNNIN 983 * If transitioning out of NOT_RUNNING, increment nr_running. Note
1007 * that the nested NOT_RUNNING is not 984 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
1008 * of multiple flags, not a single fl 985 * of multiple flags, not a single flag.
1009 */ 986 */
1010 if ((flags & WORKER_NOT_RUNNING) && ( 987 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
1011 if (!(worker->flags & WORKER_ 988 if (!(worker->flags & WORKER_NOT_RUNNING))
1012 pool->nr_running++; !! 989 atomic_inc(&pool->nr_running);
1013 } <<
1014 <<
1015 /* Return the first idle worker. Called with <<
1016 static struct worker *first_idle_worker(struc <<
1017 { <<
1018 if (unlikely(list_empty(&pool->idle_l <<
1019 return NULL; <<
1020 <<
1021 return list_first_entry(&pool->idle_l <<
1022 } <<
1023 <<
1024 /** <<
1025 * worker_enter_idle - enter idle state <<
1026 * @worker: worker which is entering idle sta <<
1027 * <<
1028 * @worker is entering idle state. Update st <<
1029 * necessary. <<
1030 * <<
1031 * LOCKING: <<
1032 * raw_spin_lock_irq(pool->lock). <<
1033 */ <<
1034 static void worker_enter_idle(struct worker * <<
1035 { <<
1036 struct worker_pool *pool = worker->po <<
1037 <<
1038 if (WARN_ON_ONCE(worker->flags & WORK <<
1039 WARN_ON_ONCE(!list_empty(&worker- <<
1040 (worker->hentry.next <<
1041 return; <<
1042 <<
1043 /* can't use worker_set_flags(), also <<
1044 worker->flags |= WORKER_IDLE; <<
1045 pool->nr_idle++; <<
1046 worker->last_active = jiffies; <<
1047 <<
1048 /* idle_list is LIFO */ <<
1049 list_add(&worker->entry, &pool->idle_ <<
1050 <<
1051 if (too_many_workers(pool) && !timer_ <<
1052 mod_timer(&pool->idle_timer, <<
1053 <<
1054 /* Sanity check nr_running. */ <<
1055 WARN_ON_ONCE(pool->nr_workers == pool <<
1056 } <<
1057 <<
1058 /** <<
1059 * worker_leave_idle - leave idle state <<
1060 * @worker: worker which is leaving idle stat <<
1061 * <<
1062 * @worker is leaving idle state. Update sta <<
1063 * <<
1064 * LOCKING: <<
1065 * raw_spin_lock_irq(pool->lock). <<
1066 */ <<
1067 static void worker_leave_idle(struct worker * <<
1068 { <<
1069 struct worker_pool *pool = worker->po <<
1070 <<
1071 if (WARN_ON_ONCE(!(worker->flags & WO <<
1072 return; <<
1073 worker_clr_flags(worker, WORKER_IDLE) <<
1074 pool->nr_idle--; <<
1075 list_del_init(&worker->entry); <<
1076 } 990 }
1077 991
1078 /** 992 /**
1079 * find_worker_executing_work - find worker w 993 * find_worker_executing_work - find worker which is executing a work
1080 * @pool: pool of interest 994 * @pool: pool of interest
1081 * @work: work to find worker for 995 * @work: work to find worker for
1082 * 996 *
1083 * Find a worker which is executing @work on 997 * Find a worker which is executing @work on @pool by searching
1084 * @pool->busy_hash which is keyed by the add 998 * @pool->busy_hash which is keyed by the address of @work. For a worker
1085 * to match, its current execution should mat 999 * to match, its current execution should match the address of @work and
1086 * its work function. This is to avoid unwan 1000 * its work function. This is to avoid unwanted dependency between
1087 * unrelated work executions through a work i 1001 * unrelated work executions through a work item being recycled while still
1088 * being executed. 1002 * being executed.
1089 * 1003 *
1090 * This is a bit tricky. A work item may be 1004 * This is a bit tricky. A work item may be freed once its execution
1091 * starts and nothing prevents the freed area 1005 * starts and nothing prevents the freed area from being recycled for
1092 * another work item. If the same work item 1006 * another work item. If the same work item address ends up being reused
1093 * before the original execution finishes, wo 1007 * before the original execution finishes, workqueue will identify the
1094 * recycled work item as currently executing 1008 * recycled work item as currently executing and make it wait until the
1095 * current execution finishes, introducing an 1009 * current execution finishes, introducing an unwanted dependency.
1096 * 1010 *
1097 * This function checks the work item address 1011 * This function checks the work item address and work function to avoid
1098 * false positives. Note that this isn't com 1012 * false positives. Note that this isn't complete as one may construct a
1099 * work function which can introduce dependen 1013 * work function which can introduce dependency onto itself through a
1100 * recycled work item. Well, if somebody wan 1014 * recycled work item. Well, if somebody wants to shoot oneself in the
1101 * foot that badly, there's only so much we c 1015 * foot that badly, there's only so much we can do, and if such deadlock
1102 * actually occurs, it should be easy to loca 1016 * actually occurs, it should be easy to locate the culprit work function.
1103 * 1017 *
1104 * CONTEXT: 1018 * CONTEXT:
1105 * raw_spin_lock_irq(pool->lock). 1019 * raw_spin_lock_irq(pool->lock).
1106 * 1020 *
1107 * Return: 1021 * Return:
1108 * Pointer to worker which is executing @work 1022 * Pointer to worker which is executing @work if found, %NULL
1109 * otherwise. 1023 * otherwise.
1110 */ 1024 */
1111 static struct worker *find_worker_executing_w 1025 static struct worker *find_worker_executing_work(struct worker_pool *pool,
1112 1026 struct work_struct *work)
1113 { 1027 {
1114 struct worker *worker; 1028 struct worker *worker;
1115 1029
1116 hash_for_each_possible(pool->busy_has 1030 hash_for_each_possible(pool->busy_hash, worker, hentry,
1117 (unsigned long 1031 (unsigned long)work)
1118 if (worker->current_work == w 1032 if (worker->current_work == work &&
1119 worker->current_func == w 1033 worker->current_func == work->func)
1120 return worker; 1034 return worker;
1121 1035
1122 return NULL; 1036 return NULL;
1123 } 1037 }
1124 1038
1125 /** 1039 /**
1126 * move_linked_works - move linked works to a 1040 * move_linked_works - move linked works to a list
1127 * @work: start of series of works to be sche 1041 * @work: start of series of works to be scheduled
1128 * @head: target list to append @work to 1042 * @head: target list to append @work to
1129 * @nextp: out parameter for nested worklist 1043 * @nextp: out parameter for nested worklist walking
1130 * 1044 *
1131 * Schedule linked works starting from @work !! 1045 * Schedule linked works starting from @work to @head. Work series to
1132 * scheduled starts at @work and includes any !! 1046 * be scheduled starts at @work and includes any consecutive work with
1133 * WORK_STRUCT_LINKED set in its predecessor. !! 1047 * WORK_STRUCT_LINKED set in its predecessor.
1134 * @nextp. !! 1048 *
>> 1049 * If @nextp is not NULL, it's updated to point to the next work of
>> 1050 * the last scheduled work. This allows move_linked_works() to be
>> 1051 * nested inside outer list_for_each_entry_safe().
1135 * 1052 *
1136 * CONTEXT: 1053 * CONTEXT:
1137 * raw_spin_lock_irq(pool->lock). 1054 * raw_spin_lock_irq(pool->lock).
1138 */ 1055 */
1139 static void move_linked_works(struct work_str 1056 static void move_linked_works(struct work_struct *work, struct list_head *head,
1140 struct work_str 1057 struct work_struct **nextp)
1141 { 1058 {
1142 struct work_struct *n; 1059 struct work_struct *n;
1143 1060
1144 /* 1061 /*
1145 * Linked worklist will always end be 1062 * Linked worklist will always end before the end of the list,
1146 * use NULL for list head. 1063 * use NULL for list head.
1147 */ 1064 */
1148 list_for_each_entry_safe_from(work, n 1065 list_for_each_entry_safe_from(work, n, NULL, entry) {
1149 list_move_tail(&work->entry, 1066 list_move_tail(&work->entry, head);
1150 if (!(*work_data_bits(work) & 1067 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1151 break; 1068 break;
1152 } 1069 }
1153 1070
1154 /* 1071 /*
1155 * If we're already inside safe list 1072 * If we're already inside safe list traversal and have moved
1156 * multiple works to the scheduled qu 1073 * multiple works to the scheduled queue, the next position
1157 * needs to be updated. 1074 * needs to be updated.
1158 */ 1075 */
1159 if (nextp) 1076 if (nextp)
1160 *nextp = n; 1077 *nextp = n;
1161 } 1078 }
1162 1079
1163 /** 1080 /**
1164 * assign_work - assign a work item and its l <<
1165 * @work: work to assign <<
1166 * @worker: worker to assign to <<
1167 * @nextp: out parameter for nested worklist <<
1168 * <<
1169 * Assign @work and its linked work items to <<
1170 * executed by another worker in the same poo <<
1171 * <<
1172 * If @nextp is not NULL, it's updated to poi <<
1173 * scheduled work. This allows assign_work() <<
1174 * list_for_each_entry_safe(). <<
1175 * <<
1176 * Returns %true if @work was successfully as <<
1177 * was punted to another worker already execu <<
1178 */ <<
1179 static bool assign_work(struct work_struct *w <<
1180 struct work_struct ** <<
1181 { <<
1182 struct worker_pool *pool = worker->po <<
1183 struct worker *collision; <<
1184 <<
1185 lockdep_assert_held(&pool->lock); <<
1186 <<
1187 /* <<
1188 * A single work shouldn't be execute <<
1189 * __queue_work() ensures that @work <<
1190 * while still running in the previou <<
1191 * @work is not executed concurrently <<
1192 * pool. Check whether anyone is alre <<
1193 * defer the work to the currently ex <<
1194 */ <<
1195 collision = find_worker_executing_wor <<
1196 if (unlikely(collision)) { <<
1197 move_linked_works(work, &coll <<
1198 return false; <<
1199 } <<
1200 <<
1201 move_linked_works(work, &worker->sche <<
1202 return true; <<
1203 } <<
1204 <<
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 /** <<
1229 * kick_pool - wake up an idle worker if nece <<
1230 * @pool: pool to kick <<
1231 * <<
1232 * @pool may have pending work items. Wake up <<
1233 * whether a worker was woken up. <<
1234 */ <<
1235 static bool kick_pool(struct worker_pool *poo <<
1236 { <<
1237 struct worker *worker = first_idle_wo <<
1238 struct task_struct *p; <<
1239 <<
1240 lockdep_assert_held(&pool->lock); <<
1241 <<
1242 if (!need_more_worker(pool) || !worke <<
1243 return false; <<
1244 <<
1245 if (pool->flags & POOL_BH) { <<
1246 kick_bh_pool(pool); <<
1247 return true; <<
1248 } <<
1249 <<
1250 p = worker->task; <<
1251 <<
1252 #ifdef CONFIG_SMP <<
1253 /* <<
1254 * Idle @worker is about to execute @ <<
1255 * opportunity to migrate @worker at <<
1256 * wake_cpu field. Let's see if we wa <<
1257 * execution locality. <<
1258 * <<
1259 * We're waking the worker that went <<
1260 * chance that @worker is marked idle <<
1261 * so, setting the wake_cpu won't do <<
1262 * optimization and the race window i <<
1263 * now. If this becomes pronounced, w <<
1264 * still on cpu when picking an idle <<
1265 * <<
1266 * If @pool has non-strict affinity, <<
1267 * its affinity scope. Repatriate. <<
1268 */ <<
1269 if (!pool->attrs->affn_strict && <<
1270 !cpumask_test_cpu(p->wake_cpu, po <<
1271 struct work_struct *work = li <<
1272 <<
1273 int wake_cpu = cpumask_any_an <<
1274 <<
1275 if (wake_cpu < nr_cpu_ids) { <<
1276 p->wake_cpu = wake_cp <<
1277 get_work_pwq(work)->s <<
1278 } <<
1279 } <<
1280 #endif <<
1281 wake_up_process(p); <<
1282 return true; <<
1283 } <<
1284 <<
1285 #ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT <<
1286 <<
1287 /* <<
1288 * Concurrency-managed per-cpu work items tha <<
1289 * wq_cpu_intensive_thresh_us trigger the aut <<
1290 * which prevents them from stalling other co <<
1291 * work function keeps triggering this mechan <<
1292 * should be using an unbound workqueue inste <<
1293 * <<
1294 * wq_cpu_intensive_report() tracks work func <<
1295 * and report them so that they can be examin <<
1296 * workqueues as appropriate. To avoid floodi <<
1297 * function is tracked and reported with expo <<
1298 */ <<
1299 #define WCI_MAX_ENTS 128 <<
1300 <<
1301 struct wci_ent { <<
1302 work_func_t func; <<
1303 atomic64_t cnt; <<
1304 struct hlist_node hash_node; <<
1305 }; <<
1306 <<
1307 static struct wci_ent wci_ents[WCI_MAX_ENTS]; <<
1308 static int wci_nr_ents; <<
1309 static DEFINE_RAW_SPINLOCK(wci_lock); <<
1310 static DEFINE_HASHTABLE(wci_hash, ilog2(WCI_M <<
1311 <<
1312 static struct wci_ent *wci_find_ent(work_func <<
1313 { <<
1314 struct wci_ent *ent; <<
1315 <<
1316 hash_for_each_possible_rcu(wci_hash, <<
1317 (unsigned <<
1318 if (ent->func == func) <<
1319 return ent; <<
1320 } <<
1321 return NULL; <<
1322 } <<
1323 <<
1324 static void wq_cpu_intensive_report(work_func <<
1325 { <<
1326 struct wci_ent *ent; <<
1327 <<
1328 restart: <<
1329 ent = wci_find_ent(func); <<
1330 if (ent) { <<
1331 u64 cnt; <<
1332 <<
1333 /* <<
1334 * Start reporting from the w <<
1335 * exponentially. <<
1336 */ <<
1337 cnt = atomic64_inc_return_rel <<
1338 if (wq_cpu_intensive_warning_ <<
1339 cnt >= wq_cpu_intensive_w <<
1340 is_power_of_2(cnt + 1 - w <<
1341 printk_deferred(KERN_ <<
1342 ent-> <<
1343 atomi <<
1344 return; <<
1345 } <<
1346 <<
1347 /* <<
1348 * @func is a new violation. Allocate <<
1349 * is exhausted, something went reall <<
1350 * noise already. <<
1351 */ <<
1352 if (wci_nr_ents >= WCI_MAX_ENTS) <<
1353 return; <<
1354 <<
1355 raw_spin_lock(&wci_lock); <<
1356 <<
1357 if (wci_nr_ents >= WCI_MAX_ENTS) { <<
1358 raw_spin_unlock(&wci_lock); <<
1359 return; <<
1360 } <<
1361 <<
1362 if (wci_find_ent(func)) { <<
1363 raw_spin_unlock(&wci_lock); <<
1364 goto restart; <<
1365 } <<
1366 <<
1367 ent = &wci_ents[wci_nr_ents++]; <<
1368 ent->func = func; <<
1369 atomic64_set(&ent->cnt, 0); <<
1370 hash_add_rcu(wci_hash, &ent->hash_nod <<
1371 <<
1372 raw_spin_unlock(&wci_lock); <<
1373 <<
1374 goto restart; <<
1375 } <<
1376 <<
1377 #else /* CONFIG_WQ_CPU_INTENSIVE_REPORT */ <<
1378 static void wq_cpu_intensive_report(work_func <<
1379 #endif /* CONFIG_WQ_CPU_INTENSIVE_REPORT */ <<
1380 <<
1381 /** <<
1382 * wq_worker_running - a worker is running ag <<
1383 * @task: task waking up <<
1384 * <<
1385 * This function is called when a worker retu <<
1386 */ <<
1387 void wq_worker_running(struct task_struct *ta <<
1388 { <<
1389 struct worker *worker = kthread_data( <<
1390 <<
1391 if (!READ_ONCE(worker->sleeping)) <<
1392 return; <<
1393 <<
1394 /* <<
1395 * If preempted by unbind_workers() b <<
1396 * and the nr_running increment below <<
1397 * and leave with an unexpected pool- <<
1398 * pool. Protect against such race. <<
1399 */ <<
1400 preempt_disable(); <<
1401 if (!(worker->flags & WORKER_NOT_RUNN <<
1402 worker->pool->nr_running++; <<
1403 preempt_enable(); <<
1404 <<
1405 /* <<
1406 * CPU intensive auto-detection cares <<
1407 * CPU without sleeping. Reset the st <<
1408 */ <<
1409 worker->current_at = worker->task->se <<
1410 <<
1411 WRITE_ONCE(worker->sleeping, 0); <<
1412 } <<
1413 <<
1414 /** <<
1415 * wq_worker_sleeping - a worker is going to <<
1416 * @task: task going to sleep <<
1417 * <<
1418 * This function is called from schedule() wh <<
1419 * going to sleep. <<
1420 */ <<
1421 void wq_worker_sleeping(struct task_struct *t <<
1422 { <<
1423 struct worker *worker = kthread_data( <<
1424 struct worker_pool *pool; <<
1425 <<
1426 /* <<
1427 * Rescuers, which may not have all t <<
1428 * workers, also reach here, let's no <<
1429 * checking NOT_RUNNING. <<
1430 */ <<
1431 if (worker->flags & WORKER_NOT_RUNNIN <<
1432 return; <<
1433 <<
1434 pool = worker->pool; <<
1435 <<
1436 /* Return if preempted before wq_work <<
1437 if (READ_ONCE(worker->sleeping)) <<
1438 return; <<
1439 <<
1440 WRITE_ONCE(worker->sleeping, 1); <<
1441 raw_spin_lock_irq(&pool->lock); <<
1442 <<
1443 /* <<
1444 * Recheck in case unbind_workers() p <<
1445 * want to decrement nr_running after <<
1446 * and nr_running has been reset. <<
1447 */ <<
1448 if (worker->flags & WORKER_NOT_RUNNIN <<
1449 raw_spin_unlock_irq(&pool->lo <<
1450 return; <<
1451 } <<
1452 <<
1453 pool->nr_running--; <<
1454 if (kick_pool(pool)) <<
1455 worker->current_pwq->stats[PW <<
1456 <<
1457 raw_spin_unlock_irq(&pool->lock); <<
1458 } <<
1459 <<
1460 /** <<
1461 * wq_worker_tick - a scheduler tick occurred <<
1462 * @task: task currently running <<
1463 * <<
1464 * Called from sched_tick(). We're in the IRQ <<
1465 * worker's fields which follow the 'K' locki <<
1466 */ <<
1467 void wq_worker_tick(struct task_struct *task) <<
1468 { <<
1469 struct worker *worker = kthread_data( <<
1470 struct pool_workqueue *pwq = worker-> <<
1471 struct worker_pool *pool = worker->po <<
1472 <<
1473 if (!pwq) <<
1474 return; <<
1475 <<
1476 pwq->stats[PWQ_STAT_CPU_TIME] += TICK <<
1477 <<
1478 if (!wq_cpu_intensive_thresh_us) <<
1479 return; <<
1480 <<
1481 /* <<
1482 * If the current worker is concurren <<
1483 * longer than wq_cpu_intensive_thres <<
1484 * CPU_INTENSIVE to avoid stalling ot <<
1485 * <<
1486 * Set @worker->sleeping means that @ <<
1487 * switching out voluntarily and won' <<
1488 * @pool->nr_running until it wakes u <<
1489 * decrements ->nr_running, setting C <<
1490 * double decrements. The task is rel <<
1491 * We probably want to make this pret <<
1492 */ <<
1493 if ((worker->flags & WORKER_NOT_RUNNI <<
1494 worker->task->se.sum_exec_runtime <<
1495 wq_cpu_intensive_thresh_us * NSEC <<
1496 return; <<
1497 <<
1498 raw_spin_lock(&pool->lock); <<
1499 <<
1500 worker_set_flags(worker, WORKER_CPU_I <<
1501 wq_cpu_intensive_report(worker->curre <<
1502 pwq->stats[PWQ_STAT_CPU_INTENSIVE]++; <<
1503 <<
1504 if (kick_pool(pool)) <<
1505 pwq->stats[PWQ_STAT_CM_WAKEUP <<
1506 <<
1507 raw_spin_unlock(&pool->lock); <<
1508 } <<
1509 <<
1510 /** <<
1511 * wq_worker_last_func - retrieve worker's la <<
1512 * @task: Task to retrieve last work function <<
1513 * <<
1514 * Determine the last function a worker execu <<
1515 * the scheduler to get a worker's last known <<
1516 * <<
1517 * CONTEXT: <<
1518 * raw_spin_lock_irq(rq->lock) <<
1519 * <<
1520 * This function is called during schedule() <<
1521 * to sleep. It's used by psi to identify agg <<
1522 * dequeuing, to allow periodic aggregation t <<
1523 * worker is the last task in the system or c <<
1524 * <<
1525 * As this function doesn't involve any workq <<
1526 * only returns stable values when called fro <<
1527 * queuing and dequeuing paths, when @task, w <<
1528 * is guaranteed to not be processing any wor <<
1529 * <<
1530 * Return: <<
1531 * The last work function %current executed a <<
1532 * hasn't executed any work yet. <<
1533 */ <<
1534 work_func_t wq_worker_last_func(struct task_s <<
1535 { <<
1536 struct worker *worker = kthread_data( <<
1537 <<
1538 return worker->last_func; <<
1539 } <<
1540 <<
1541 /** <<
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 1081 * get_pwq - get an extra reference on the specified pool_workqueue
1620 * @pwq: pool_workqueue to get 1082 * @pwq: pool_workqueue to get
1621 * 1083 *
1622 * Obtain an extra reference on @pwq. The ca 1084 * Obtain an extra reference on @pwq. The caller should guarantee that
1623 * @pwq has positive refcnt and be holding th 1085 * @pwq has positive refcnt and be holding the matching pool->lock.
1624 */ 1086 */
1625 static void get_pwq(struct pool_workqueue *pw 1087 static void get_pwq(struct pool_workqueue *pwq)
1626 { 1088 {
1627 lockdep_assert_held(&pwq->pool->lock) 1089 lockdep_assert_held(&pwq->pool->lock);
1628 WARN_ON_ONCE(pwq->refcnt <= 0); 1090 WARN_ON_ONCE(pwq->refcnt <= 0);
1629 pwq->refcnt++; 1091 pwq->refcnt++;
1630 } 1092 }
1631 1093
1632 /** 1094 /**
1633 * put_pwq - put a pool_workqueue reference 1095 * put_pwq - put a pool_workqueue reference
1634 * @pwq: pool_workqueue to put 1096 * @pwq: pool_workqueue to put
1635 * 1097 *
1636 * Drop a reference of @pwq. If its refcnt r 1098 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1637 * destruction. The caller should be holding 1099 * destruction. The caller should be holding the matching pool->lock.
1638 */ 1100 */
1639 static void put_pwq(struct pool_workqueue *pw 1101 static void put_pwq(struct pool_workqueue *pwq)
1640 { 1102 {
1641 lockdep_assert_held(&pwq->pool->lock) 1103 lockdep_assert_held(&pwq->pool->lock);
1642 if (likely(--pwq->refcnt)) 1104 if (likely(--pwq->refcnt))
1643 return; 1105 return;
>> 1106 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
>> 1107 return;
1644 /* 1108 /*
1645 * @pwq can't be released under pool- !! 1109 * @pwq can't be released under pool->lock, bounce to
1646 * kthread_worker to avoid A-A deadlo !! 1110 * pwq_unbound_release_workfn(). This never recurses on the same
>> 1111 * pool->lock as this path is taken only for unbound workqueues and
>> 1112 * the release work item is scheduled on a per-cpu workqueue. To
>> 1113 * avoid lockdep warning, unbound pool->locks are given lockdep
>> 1114 * subclass of 1 in get_unbound_pool().
1647 */ 1115 */
1648 kthread_queue_work(pwq_release_worker !! 1116 schedule_work(&pwq->unbound_release_work);
1649 } 1117 }
1650 1118
1651 /** 1119 /**
1652 * put_pwq_unlocked - put_pwq() with surround 1120 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1653 * @pwq: pool_workqueue to put (can be %NULL) 1121 * @pwq: pool_workqueue to put (can be %NULL)
1654 * 1122 *
1655 * put_pwq() with locking. This function als 1123 * put_pwq() with locking. This function also allows %NULL @pwq.
1656 */ 1124 */
1657 static void put_pwq_unlocked(struct pool_work 1125 static void put_pwq_unlocked(struct pool_workqueue *pwq)
1658 { 1126 {
1659 if (pwq) { 1127 if (pwq) {
1660 /* 1128 /*
1661 * As both pwqs and pools are 1129 * As both pwqs and pools are RCU protected, the
1662 * following lock operations 1130 * following lock operations are safe.
1663 */ 1131 */
1664 raw_spin_lock_irq(&pwq->pool- 1132 raw_spin_lock_irq(&pwq->pool->lock);
1665 put_pwq(pwq); 1133 put_pwq(pwq);
1666 raw_spin_unlock_irq(&pwq->poo 1134 raw_spin_unlock_irq(&pwq->pool->lock);
1667 } 1135 }
1668 } 1136 }
1669 1137
1670 static bool pwq_is_empty(struct pool_workqueu !! 1138 static void pwq_activate_delayed_work(struct work_struct *work)
1671 { 1139 {
1672 return !pwq->nr_active && list_empty( !! 1140 struct pool_workqueue *pwq = get_work_pwq(work);
1673 } <<
1674 <<
1675 static void __pwq_activate_work(struct pool_w <<
1676 struct work_s <<
1677 { <<
1678 unsigned long *wdb = work_data_bits(w <<
1679 1141
1680 WARN_ON_ONCE(!(*wdb & WORK_STRUCT_INA <<
1681 trace_workqueue_activate_work(work); 1142 trace_workqueue_activate_work(work);
1682 if (list_empty(&pwq->pool->worklist)) 1143 if (list_empty(&pwq->pool->worklist))
1683 pwq->pool->watchdog_ts = jiff 1144 pwq->pool->watchdog_ts = jiffies;
1684 move_linked_works(work, &pwq->pool->w 1145 move_linked_works(work, &pwq->pool->worklist, NULL);
1685 __clear_bit(WORK_STRUCT_INACTIVE_BIT, !! 1146 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
>> 1147 pwq->nr_active++;
1686 } 1148 }
1687 1149
1688 static bool tryinc_node_nr_active(struct wq_n !! 1150 static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1689 { 1151 {
1690 int max = READ_ONCE(nna->max); !! 1152 struct work_struct *work = list_first_entry(&pwq->delayed_works,
1691 !! 1153 struct work_struct, entry);
1692 while (true) { <<
1693 int old, tmp; <<
1694 1154
1695 old = atomic_read(&nna->nr); !! 1155 pwq_activate_delayed_work(work);
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 <<
1971 if (!list_empty(&nna->pending_pwqs)) <<
1972 node_activate_pending_pwq(nna <<
1973 } 1156 }
1974 1157
1975 /** 1158 /**
1976 * pwq_dec_nr_in_flight - decrement pwq's nr_ 1159 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1977 * @pwq: pwq of interest 1160 * @pwq: pwq of interest
1978 * @work_data: work_data of work which left t !! 1161 * @color: color of work which left the queue
1979 * 1162 *
1980 * A work either has completed or is removed 1163 * A work either has completed or is removed from pending queue,
1981 * decrement nr_in_flight of its pwq and hand 1164 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1982 * 1165 *
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: 1166 * CONTEXT:
1989 * raw_spin_lock_irq(pool->lock). 1167 * raw_spin_lock_irq(pool->lock).
1990 */ 1168 */
1991 static void pwq_dec_nr_in_flight(struct pool_ !! 1169 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1992 { 1170 {
1993 int color = get_work_color(work_data) !! 1171 /* uncolored work items don't participate in flushing or nr_active */
1994 !! 1172 if (color == WORK_NO_COLOR)
1995 if (!(work_data & WORK_STRUCT_INACTIV !! 1173 goto out_put;
1996 pwq_dec_nr_active(pwq); <<
1997 1174
1998 pwq->nr_in_flight[color]--; 1175 pwq->nr_in_flight[color]--;
1999 1176
>> 1177 pwq->nr_active--;
>> 1178 if (!list_empty(&pwq->delayed_works)) {
>> 1179 /* one down, submit a delayed one */
>> 1180 if (pwq->nr_active < pwq->max_active)
>> 1181 pwq_activate_first_delayed(pwq);
>> 1182 }
>> 1183
2000 /* is flush in progress and are we at 1184 /* is flush in progress and are we at the flushing tip? */
2001 if (likely(pwq->flush_color != color) 1185 if (likely(pwq->flush_color != color))
2002 goto out_put; 1186 goto out_put;
2003 1187
2004 /* are there still in-flight works? * 1188 /* are there still in-flight works? */
2005 if (pwq->nr_in_flight[color]) 1189 if (pwq->nr_in_flight[color])
2006 goto out_put; 1190 goto out_put;
2007 1191
2008 /* this pwq is done, clear flush_colo 1192 /* this pwq is done, clear flush_color */
2009 pwq->flush_color = -1; 1193 pwq->flush_color = -1;
2010 1194
2011 /* 1195 /*
2012 * If this was the last pwq, wake up 1196 * If this was the last pwq, wake up the first flusher. It
2013 * will handle the rest. 1197 * will handle the rest.
2014 */ 1198 */
2015 if (atomic_dec_and_test(&pwq->wq->nr_ 1199 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
2016 complete(&pwq->wq->first_flus 1200 complete(&pwq->wq->first_flusher->done);
2017 out_put: 1201 out_put:
2018 put_pwq(pwq); 1202 put_pwq(pwq);
2019 } 1203 }
2020 1204
2021 /** 1205 /**
2022 * try_to_grab_pending - steal work item from 1206 * try_to_grab_pending - steal work item from worklist and disable irq
2023 * @work: work item to steal 1207 * @work: work item to steal
2024 * @cflags: %WORK_CANCEL_ flags !! 1208 * @is_dwork: @work is a delayed_work
2025 * @irq_flags: place to store irq state !! 1209 * @flags: place to store irq state
2026 * 1210 *
2027 * Try to grab PENDING bit of @work. This fu 1211 * Try to grab PENDING bit of @work. This function can handle @work in any
2028 * stable state - idle, on timer or on workli 1212 * stable state - idle, on timer or on worklist.
2029 * 1213 *
2030 * Return: 1214 * Return:
2031 * 1215 *
2032 * ======== ============================= 1216 * ======== ================================================================
2033 * 1 if @work was pending and we s 1217 * 1 if @work was pending and we successfully stole PENDING
2034 * 0 if @work was idle and we clai 1218 * 0 if @work was idle and we claimed PENDING
2035 * -EAGAIN if PENDING couldn't be grabbe 1219 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
>> 1220 * -ENOENT if someone else is canceling @work, this state may persist
>> 1221 * for arbitrarily long
2036 * ======== ============================= 1222 * ======== ================================================================
2037 * 1223 *
2038 * Note: 1224 * Note:
2039 * On >= 0 return, the caller owns @work's PE 1225 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
2040 * interrupted while holding PENDING and @wor 1226 * interrupted while holding PENDING and @work off queue, irq must be
2041 * disabled on entry. This, combined with de 1227 * disabled on entry. This, combined with delayed_work->timer being
2042 * irqsafe, ensures that we return -EAGAIN fo 1228 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
2043 * 1229 *
2044 * On successful return, >= 0, irq is disable 1230 * On successful return, >= 0, irq is disabled and the caller is
2045 * responsible for releasing it using local_i !! 1231 * responsible for releasing it using local_irq_restore(*@flags).
2046 * 1232 *
2047 * This function is safe to call from any con 1233 * This function is safe to call from any context including IRQ handler.
2048 */ 1234 */
2049 static int try_to_grab_pending(struct work_st !! 1235 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
2050 unsigned long !! 1236 unsigned long *flags)
2051 { 1237 {
2052 struct worker_pool *pool; 1238 struct worker_pool *pool;
2053 struct pool_workqueue *pwq; 1239 struct pool_workqueue *pwq;
2054 1240
2055 local_irq_save(*irq_flags); !! 1241 local_irq_save(*flags);
2056 1242
2057 /* try to steal the timer if it exist 1243 /* try to steal the timer if it exists */
2058 if (cflags & WORK_CANCEL_DELAYED) { !! 1244 if (is_dwork) {
2059 struct delayed_work *dwork = 1245 struct delayed_work *dwork = to_delayed_work(work);
2060 1246
2061 /* 1247 /*
2062 * dwork->timer is irqsafe. 1248 * dwork->timer is irqsafe. If del_timer() fails, it's
2063 * guaranteed that the timer 1249 * guaranteed that the timer is not queued anywhere and not
2064 * running on the local CPU. 1250 * running on the local CPU.
2065 */ 1251 */
2066 if (likely(del_timer(&dwork-> 1252 if (likely(del_timer(&dwork->timer)))
2067 return 1; 1253 return 1;
2068 } 1254 }
2069 1255
2070 /* try to claim PENDING the normal wa 1256 /* try to claim PENDING the normal way */
2071 if (!test_and_set_bit(WORK_STRUCT_PEN 1257 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
2072 return 0; 1258 return 0;
2073 1259
2074 rcu_read_lock(); 1260 rcu_read_lock();
2075 /* 1261 /*
2076 * The queueing is in progress, or it 1262 * The queueing is in progress, or it is already queued. Try to
2077 * steal it from ->worklist without c 1263 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
2078 */ 1264 */
2079 pool = get_work_pool(work); 1265 pool = get_work_pool(work);
2080 if (!pool) 1266 if (!pool)
2081 goto fail; 1267 goto fail;
2082 1268
2083 raw_spin_lock(&pool->lock); 1269 raw_spin_lock(&pool->lock);
2084 /* 1270 /*
2085 * work->data is guaranteed to point 1271 * work->data is guaranteed to point to pwq only while the work
2086 * item is queued on pwq->wq, and bot 1272 * item is queued on pwq->wq, and both updating work->data to point
2087 * to pwq on queueing and to pool on 1273 * to pwq on queueing and to pool on dequeueing are done under
2088 * pwq->pool->lock. This in turn gua 1274 * pwq->pool->lock. This in turn guarantees that, if work->data
2089 * points to pwq which is associated 1275 * points to pwq which is associated with a locked pool, the work
2090 * item is currently queued on that p 1276 * item is currently queued on that pool.
2091 */ 1277 */
2092 pwq = get_work_pwq(work); 1278 pwq = get_work_pwq(work);
2093 if (pwq && pwq->pool == pool) { 1279 if (pwq && pwq->pool == pool) {
2094 unsigned long work_data = *wo <<
2095 <<
2096 debug_work_deactivate(work); 1280 debug_work_deactivate(work);
2097 1281
2098 /* 1282 /*
2099 * A cancelable inactive work !! 1283 * A delayed work item cannot be grabbed directly because
2100 * pwq->inactive_works since !! 1284 * it might have linked NO_COLOR work items which, if left
2101 * canceled (see the comments !! 1285 * on the delayed_list, will confuse pwq->nr_active
2102 * !! 1286 * management later on and cause stall. Make sure the work
2103 * An inactive work item cann !! 1287 * item is activated before grabbing.
2104 * it might have linked barri <<
2105 * on the inactive_works list <<
2106 * management later on and ca <<
2107 * barrier work items to the <<
2108 * item. Also keep WORK_STRUC <<
2109 * it doesn't participate in <<
2110 * pwq_dec_nr_in_flight(). <<
2111 */ 1288 */
2112 if (work_data & WORK_STRUCT_I !! 1289 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
2113 move_linked_works(wor !! 1290 pwq_activate_delayed_work(work);
2114 1291
2115 list_del_init(&work->entry); 1292 list_del_init(&work->entry);
>> 1293 pwq_dec_nr_in_flight(pwq, get_work_color(work));
2116 1294
2117 /* !! 1295 /* work->data points to pwq iff queued, point to pool */
2118 * work->data points to pwq i !! 1296 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 1297
2127 raw_spin_unlock(&pool->lock); 1298 raw_spin_unlock(&pool->lock);
2128 rcu_read_unlock(); 1299 rcu_read_unlock();
2129 return 1; 1300 return 1;
2130 } 1301 }
2131 raw_spin_unlock(&pool->lock); 1302 raw_spin_unlock(&pool->lock);
2132 fail: 1303 fail:
2133 rcu_read_unlock(); 1304 rcu_read_unlock();
2134 local_irq_restore(*irq_flags); !! 1305 local_irq_restore(*flags);
>> 1306 if (work_is_canceling(work))
>> 1307 return -ENOENT;
>> 1308 cpu_relax();
2135 return -EAGAIN; 1309 return -EAGAIN;
2136 } 1310 }
2137 1311
2138 /** 1312 /**
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 1313 * insert_work - insert a work into a pool
2168 * @pwq: pwq @work belongs to 1314 * @pwq: pwq @work belongs to
2169 * @work: work to insert 1315 * @work: work to insert
2170 * @head: insertion point 1316 * @head: insertion point
2171 * @extra_flags: extra WORK_STRUCT_* flags to 1317 * @extra_flags: extra WORK_STRUCT_* flags to set
2172 * 1318 *
2173 * Insert @work which belongs to @pwq after @ 1319 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
2174 * work_struct flags. 1320 * work_struct flags.
2175 * 1321 *
2176 * CONTEXT: 1322 * CONTEXT:
2177 * raw_spin_lock_irq(pool->lock). 1323 * raw_spin_lock_irq(pool->lock).
2178 */ 1324 */
2179 static void insert_work(struct pool_workqueue 1325 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
2180 struct list_head *hea 1326 struct list_head *head, unsigned int extra_flags)
2181 { 1327 {
2182 debug_work_activate(work); !! 1328 struct worker_pool *pool = pwq->pool;
2183 1329
2184 /* record the work call stack in orde 1330 /* record the work call stack in order to print it in KASAN reports */
2185 kasan_record_aux_stack_noalloc(work); !! 1331 kasan_record_aux_stack(work);
2186 1332
2187 /* we own @work, set data and link */ 1333 /* we own @work, set data and link */
2188 set_work_pwq(work, pwq, extra_flags); 1334 set_work_pwq(work, pwq, extra_flags);
2189 list_add_tail(&work->entry, head); 1335 list_add_tail(&work->entry, head);
2190 get_pwq(pwq); 1336 get_pwq(pwq);
>> 1337
>> 1338 /*
>> 1339 * Ensure either wq_worker_sleeping() sees the above
>> 1340 * list_add_tail() or we see zero nr_running to avoid workers lying
>> 1341 * around lazily while there are works to be processed.
>> 1342 */
>> 1343 smp_mb();
>> 1344
>> 1345 if (__need_more_worker(pool))
>> 1346 wake_up_worker(pool);
2191 } 1347 }
2192 1348
2193 /* 1349 /*
2194 * Test whether @work is being queued from an 1350 * Test whether @work is being queued from another work executing on the
2195 * same workqueue. 1351 * same workqueue.
2196 */ 1352 */
2197 static bool is_chained_work(struct workqueue_ 1353 static bool is_chained_work(struct workqueue_struct *wq)
2198 { 1354 {
2199 struct worker *worker; 1355 struct worker *worker;
2200 1356
2201 worker = current_wq_worker(); 1357 worker = current_wq_worker();
2202 /* 1358 /*
2203 * Return %true iff I'm a worker exec 1359 * Return %true iff I'm a worker executing a work item on @wq. If
2204 * I'm @worker, it's safe to derefere 1360 * I'm @worker, it's safe to dereference it without locking.
2205 */ 1361 */
2206 return worker && worker->current_pwq- 1362 return worker && worker->current_pwq->wq == wq;
2207 } 1363 }
2208 1364
2209 /* 1365 /*
2210 * When queueing an unbound work item to a wq 1366 * When queueing an unbound work item to a wq, prefer local CPU if allowed
2211 * by wq_unbound_cpumask. Otherwise, round r 1367 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
2212 * avoid perturbing sensitive tasks. 1368 * avoid perturbing sensitive tasks.
2213 */ 1369 */
2214 static int wq_select_unbound_cpu(int cpu) 1370 static int wq_select_unbound_cpu(int cpu)
2215 { 1371 {
>> 1372 static bool printed_dbg_warning;
2216 int new_cpu; 1373 int new_cpu;
2217 1374
2218 if (likely(!wq_debug_force_rr_cpu)) { 1375 if (likely(!wq_debug_force_rr_cpu)) {
2219 if (cpumask_test_cpu(cpu, wq_ 1376 if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
2220 return cpu; 1377 return cpu;
2221 } else { !! 1378 } else if (!printed_dbg_warning) {
2222 pr_warn_once("workqueue: roun !! 1379 pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n");
>> 1380 printed_dbg_warning = true;
2223 } 1381 }
2224 1382
>> 1383 if (cpumask_empty(wq_unbound_cpumask))
>> 1384 return cpu;
>> 1385
2225 new_cpu = __this_cpu_read(wq_rr_cpu_l 1386 new_cpu = __this_cpu_read(wq_rr_cpu_last);
2226 new_cpu = cpumask_next_and(new_cpu, w 1387 new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
2227 if (unlikely(new_cpu >= nr_cpu_ids)) 1388 if (unlikely(new_cpu >= nr_cpu_ids)) {
2228 new_cpu = cpumask_first_and(w 1389 new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
2229 if (unlikely(new_cpu >= nr_cp 1390 if (unlikely(new_cpu >= nr_cpu_ids))
2230 return cpu; 1391 return cpu;
2231 } 1392 }
2232 __this_cpu_write(wq_rr_cpu_last, new_ 1393 __this_cpu_write(wq_rr_cpu_last, new_cpu);
2233 1394
2234 return new_cpu; 1395 return new_cpu;
2235 } 1396 }
2236 1397
2237 static void __queue_work(int cpu, struct work 1398 static void __queue_work(int cpu, struct workqueue_struct *wq,
2238 struct work_struct * 1399 struct work_struct *work)
2239 { 1400 {
2240 struct pool_workqueue *pwq; 1401 struct pool_workqueue *pwq;
2241 struct worker_pool *last_pool, *pool; !! 1402 struct worker_pool *last_pool;
>> 1403 struct list_head *worklist;
2242 unsigned int work_flags; 1404 unsigned int work_flags;
2243 unsigned int req_cpu = cpu; 1405 unsigned int req_cpu = cpu;
2244 1406
2245 /* 1407 /*
2246 * While a work item is PENDING && of 1408 * While a work item is PENDING && off queue, a task trying to
2247 * steal the PENDING will busy-loop w 1409 * steal the PENDING will busy-loop waiting for it to either get
2248 * queued or lose PENDING. Grabbing 1410 * queued or lose PENDING. Grabbing PENDING and queueing should
2249 * happen with IRQ disabled. 1411 * happen with IRQ disabled.
2250 */ 1412 */
2251 lockdep_assert_irqs_disabled(); 1413 lockdep_assert_irqs_disabled();
2252 1414
2253 /* !! 1415
2254 * For a draining wq, only works from !! 1416 /* if draining, only works from the same workqueue are allowed */
2255 * allowed. The __WQ_DESTROYING helps !! 1417 if (unlikely(wq->flags & __WQ_DRAINING) &&
2256 * queues a new work item to a wq aft !! 1418 WARN_ON_ONCE(!is_chained_work(wq)))
2257 */ <<
2258 if (unlikely(wq->flags & (__WQ_DESTRO <<
2259 WARN_ON_ONCE(!is_chained <<
2260 return; 1419 return;
2261 rcu_read_lock(); 1420 rcu_read_lock();
2262 retry: 1421 retry:
2263 /* pwq which will be used unless @wor 1422 /* pwq which will be used unless @work is executing elsewhere */
2264 if (req_cpu == WORK_CPU_UNBOUND) { !! 1423 if (wq->flags & WQ_UNBOUND) {
2265 if (wq->flags & WQ_UNBOUND) !! 1424 if (req_cpu == WORK_CPU_UNBOUND)
2266 cpu = wq_select_unbou 1425 cpu = wq_select_unbound_cpu(raw_smp_processor_id());
2267 else !! 1426 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
>> 1427 } else {
>> 1428 if (req_cpu == WORK_CPU_UNBOUND)
2268 cpu = raw_smp_process 1429 cpu = raw_smp_processor_id();
>> 1430 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
2269 } 1431 }
2270 1432
2271 pwq = rcu_dereference(*per_cpu_ptr(wq <<
2272 pool = pwq->pool; <<
2273 <<
2274 /* 1433 /*
2275 * If @work was previously on a diffe 1434 * If @work was previously on a different pool, it might still be
2276 * running there, in which case the w 1435 * running there, in which case the work needs to be queued on that
2277 * pool to guarantee non-reentrancy. 1436 * 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 */ 1437 */
2283 last_pool = get_work_pool(work); 1438 last_pool = get_work_pool(work);
2284 if (last_pool && last_pool != pool && !! 1439 if (last_pool && last_pool != pwq->pool) {
2285 struct worker *worker; 1440 struct worker *worker;
2286 1441
2287 raw_spin_lock(&last_pool->loc 1442 raw_spin_lock(&last_pool->lock);
2288 1443
2289 worker = find_worker_executin 1444 worker = find_worker_executing_work(last_pool, work);
2290 1445
2291 if (worker && worker->current 1446 if (worker && worker->current_pwq->wq == wq) {
2292 pwq = worker->current 1447 pwq = worker->current_pwq;
2293 pool = pwq->pool; <<
2294 WARN_ON_ONCE(pool != <<
2295 } else { 1448 } else {
2296 /* meh... not running 1449 /* meh... not running there, queue here */
2297 raw_spin_unlock(&last 1450 raw_spin_unlock(&last_pool->lock);
2298 raw_spin_lock(&pool-> !! 1451 raw_spin_lock(&pwq->pool->lock);
2299 } 1452 }
2300 } else { 1453 } else {
2301 raw_spin_lock(&pool->lock); !! 1454 raw_spin_lock(&pwq->pool->lock);
2302 } 1455 }
2303 1456
2304 /* 1457 /*
2305 * pwq is determined and locked. For !! 1458 * pwq is determined and locked. For unbound pools, we could have
2306 * with pwq release and it could alre !! 1459 * raced with pwq release and it could already be dead. If its
2307 * repeat pwq selection. Note that un !! 1460 * refcnt is zero, repeat pwq selection. Note that pwqs never die
2308 * another pwq replacing it in cpu_pw !! 1461 * without another pwq replacing it in the numa_pwq_tbl or while
2309 * on it, so the retrying is guarante !! 1462 * work items are executing on it, so the retrying is guaranteed to
>> 1463 * make forward-progress.
2310 */ 1464 */
2311 if (unlikely(!pwq->refcnt)) { 1465 if (unlikely(!pwq->refcnt)) {
2312 if (wq->flags & WQ_UNBOUND) { 1466 if (wq->flags & WQ_UNBOUND) {
2313 raw_spin_unlock(&pool !! 1467 raw_spin_unlock(&pwq->pool->lock);
2314 cpu_relax(); 1468 cpu_relax();
2315 goto retry; 1469 goto retry;
2316 } 1470 }
2317 /* oops */ 1471 /* oops */
2318 WARN_ONCE(true, "workqueue: p 1472 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
2319 wq->name, cpu); 1473 wq->name, cpu);
2320 } 1474 }
2321 1475
2322 /* pwq determined, queue */ 1476 /* pwq determined, queue */
2323 trace_workqueue_queue_work(req_cpu, p 1477 trace_workqueue_queue_work(req_cpu, pwq, work);
2324 1478
2325 if (WARN_ON(!list_empty(&work->entry) 1479 if (WARN_ON(!list_empty(&work->entry)))
2326 goto out; 1480 goto out;
2327 1481
2328 pwq->nr_in_flight[pwq->work_color]++; 1482 pwq->nr_in_flight[pwq->work_color]++;
2329 work_flags = work_color_to_flags(pwq- 1483 work_flags = work_color_to_flags(pwq->work_color);
2330 1484
2331 /* !! 1485 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 <<
2338 pool->watchdog_ts = j <<
2339 <<
2340 trace_workqueue_activate_work 1486 trace_workqueue_activate_work(work);
2341 insert_work(pwq, work, &pool- !! 1487 pwq->nr_active++;
2342 kick_pool(pool); !! 1488 worklist = &pwq->pool->worklist;
>> 1489 if (list_empty(worklist))
>> 1490 pwq->pool->watchdog_ts = jiffies;
2343 } else { 1491 } else {
2344 work_flags |= WORK_STRUCT_INA !! 1492 work_flags |= WORK_STRUCT_DELAYED;
2345 insert_work(pwq, work, &pwq-> !! 1493 worklist = &pwq->delayed_works;
2346 } 1494 }
2347 1495
>> 1496 debug_work_activate(work);
>> 1497 insert_work(pwq, work, worklist, work_flags);
>> 1498
2348 out: 1499 out:
2349 raw_spin_unlock(&pool->lock); !! 1500 raw_spin_unlock(&pwq->pool->lock);
2350 rcu_read_unlock(); 1501 rcu_read_unlock();
2351 } 1502 }
2352 1503
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 /** 1504 /**
2369 * queue_work_on - queue work on specific cpu 1505 * queue_work_on - queue work on specific cpu
2370 * @cpu: CPU number to execute work on 1506 * @cpu: CPU number to execute work on
2371 * @wq: workqueue to use 1507 * @wq: workqueue to use
2372 * @work: work to queue 1508 * @work: work to queue
2373 * 1509 *
2374 * We queue the work to a specific CPU, the c 1510 * We queue the work to a specific CPU, the caller must ensure it
2375 * can't go away. Callers that fail to ensur !! 1511 * can't go away.
2376 * CPU cannot go away will execute on a rando <<
2377 * But note well that callers specifying a CP <<
2378 * online will get a splat. <<
2379 * 1512 *
2380 * Return: %false if @work was already on a q 1513 * Return: %false if @work was already on a queue, %true otherwise.
2381 */ 1514 */
2382 bool queue_work_on(int cpu, struct workqueue_ 1515 bool queue_work_on(int cpu, struct workqueue_struct *wq,
2383 struct work_struct *work) 1516 struct work_struct *work)
2384 { 1517 {
2385 bool ret = false; 1518 bool ret = false;
2386 unsigned long irq_flags; !! 1519 unsigned long flags;
2387 1520
2388 local_irq_save(irq_flags); !! 1521 local_irq_save(flags);
2389 1522
2390 if (!test_and_set_bit(WORK_STRUCT_PEN !! 1523 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); 1524 __queue_work(cpu, wq, work);
2393 ret = true; 1525 ret = true;
2394 } 1526 }
2395 1527
2396 local_irq_restore(irq_flags); !! 1528 local_irq_restore(flags);
2397 return ret; 1529 return ret;
2398 } 1530 }
2399 EXPORT_SYMBOL(queue_work_on); 1531 EXPORT_SYMBOL(queue_work_on);
2400 1532
2401 /** 1533 /**
2402 * select_numa_node_cpu - Select a CPU based !! 1534 * workqueue_select_cpu_near - Select a CPU based on NUMA node
2403 * @node: NUMA node ID that we want to select 1535 * @node: NUMA node ID that we want to select a CPU from
2404 * 1536 *
2405 * This function will attempt to find a "rand 1537 * This function will attempt to find a "random" cpu available on a given
2406 * node. If there are no CPUs available on th 1538 * node. If there are no CPUs available on the given node it will return
2407 * WORK_CPU_UNBOUND indicating that we should 1539 * WORK_CPU_UNBOUND indicating that we should just schedule to any
2408 * available CPU if we need to schedule this 1540 * available CPU if we need to schedule this work.
2409 */ 1541 */
2410 static int select_numa_node_cpu(int node) !! 1542 static int workqueue_select_cpu_near(int node)
2411 { 1543 {
2412 int cpu; 1544 int cpu;
2413 1545
>> 1546 /* No point in doing this if NUMA isn't enabled for workqueues */
>> 1547 if (!wq_numa_enabled)
>> 1548 return WORK_CPU_UNBOUND;
>> 1549
2414 /* Delay binding to CPU if node is no 1550 /* Delay binding to CPU if node is not valid or online */
2415 if (node < 0 || node >= MAX_NUMNODES 1551 if (node < 0 || node >= MAX_NUMNODES || !node_online(node))
2416 return WORK_CPU_UNBOUND; 1552 return WORK_CPU_UNBOUND;
2417 1553
2418 /* Use local node/cpu if we are alrea 1554 /* Use local node/cpu if we are already there */
2419 cpu = raw_smp_processor_id(); 1555 cpu = raw_smp_processor_id();
2420 if (node == cpu_to_node(cpu)) 1556 if (node == cpu_to_node(cpu))
2421 return cpu; 1557 return cpu;
2422 1558
2423 /* Use "random" otherwise know as "fi 1559 /* Use "random" otherwise know as "first" online CPU of node */
2424 cpu = cpumask_any_and(cpumask_of_node 1560 cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask);
2425 1561
2426 /* If CPU is valid return that, other 1562 /* If CPU is valid return that, otherwise just defer */
2427 return cpu < nr_cpu_ids ? cpu : WORK_ 1563 return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND;
2428 } 1564 }
2429 1565
2430 /** 1566 /**
2431 * queue_work_node - queue work on a "random" 1567 * queue_work_node - queue work on a "random" cpu for a given NUMA node
2432 * @node: NUMA node that we are targeting the 1568 * @node: NUMA node that we are targeting the work for
2433 * @wq: workqueue to use 1569 * @wq: workqueue to use
2434 * @work: work to queue 1570 * @work: work to queue
2435 * 1571 *
2436 * We queue the work to a "random" CPU within 1572 * 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 1573 * idea here is to provide a way to somehow associate work with a given
2438 * NUMA node. 1574 * NUMA node.
2439 * 1575 *
2440 * This function will only make a best effort 1576 * This function will only make a best effort attempt at getting this onto
2441 * the right NUMA node. If no node is request 1577 * the right NUMA node. If no node is requested or the requested node is
2442 * offline then we just fall back to standard 1578 * offline then we just fall back to standard queue_work behavior.
2443 * 1579 *
2444 * Currently the "random" CPU ends up being t 1580 * Currently the "random" CPU ends up being the first available CPU in the
2445 * intersection of cpu_online_mask and the cp 1581 * intersection of cpu_online_mask and the cpumask of the node, unless we
2446 * are running on the node. In that case we j 1582 * are running on the node. In that case we just use the current CPU.
2447 * 1583 *
2448 * Return: %false if @work was already on a q 1584 * Return: %false if @work was already on a queue, %true otherwise.
2449 */ 1585 */
2450 bool queue_work_node(int node, struct workque 1586 bool queue_work_node(int node, struct workqueue_struct *wq,
2451 struct work_struct *work 1587 struct work_struct *work)
2452 { 1588 {
2453 unsigned long irq_flags; !! 1589 unsigned long flags;
2454 bool ret = false; 1590 bool ret = false;
2455 1591
2456 /* 1592 /*
2457 * This current implementation is spe 1593 * This current implementation is specific to unbound workqueues.
2458 * Specifically we only return the fi 1594 * Specifically we only return the first available CPU for a given
2459 * node instead of cycling through in 1595 * node instead of cycling through individual CPUs within the node.
2460 * 1596 *
2461 * If this is used with a per-cpu wor 1597 * If this is used with a per-cpu workqueue then the logic in
2462 * workqueue_select_cpu_near would ne 1598 * workqueue_select_cpu_near would need to be updated to allow for
2463 * some round robin type logic. 1599 * some round robin type logic.
2464 */ 1600 */
2465 WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND 1601 WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND));
2466 1602
2467 local_irq_save(irq_flags); !! 1603 local_irq_save(flags);
2468 1604
2469 if (!test_and_set_bit(WORK_STRUCT_PEN !! 1605 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
2470 !clear_pending_if_disabled(work)) !! 1606 int cpu = workqueue_select_cpu_near(node);
2471 int cpu = select_numa_node_cp <<
2472 1607
2473 __queue_work(cpu, wq, work); 1608 __queue_work(cpu, wq, work);
2474 ret = true; 1609 ret = true;
2475 } 1610 }
2476 1611
2477 local_irq_restore(irq_flags); !! 1612 local_irq_restore(flags);
2478 return ret; 1613 return ret;
2479 } 1614 }
2480 EXPORT_SYMBOL_GPL(queue_work_node); 1615 EXPORT_SYMBOL_GPL(queue_work_node);
2481 1616
2482 void delayed_work_timer_fn(struct timer_list 1617 void delayed_work_timer_fn(struct timer_list *t)
2483 { 1618 {
2484 struct delayed_work *dwork = from_tim 1619 struct delayed_work *dwork = from_timer(dwork, t, timer);
2485 1620
2486 /* should have been called from irqsa 1621 /* should have been called from irqsafe timer with irq already off */
2487 __queue_work(dwork->cpu, dwork->wq, & 1622 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2488 } 1623 }
2489 EXPORT_SYMBOL(delayed_work_timer_fn); 1624 EXPORT_SYMBOL(delayed_work_timer_fn);
2490 1625
2491 static void __queue_delayed_work(int cpu, str 1626 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
2492 struct delaye 1627 struct delayed_work *dwork, unsigned long delay)
2493 { 1628 {
2494 struct timer_list *timer = &dwork->ti 1629 struct timer_list *timer = &dwork->timer;
2495 struct work_struct *work = &dwork->wo 1630 struct work_struct *work = &dwork->work;
2496 1631
2497 WARN_ON_ONCE(!wq); 1632 WARN_ON_ONCE(!wq);
2498 WARN_ON_ONCE(timer->function != delay 1633 WARN_ON_ONCE(timer->function != delayed_work_timer_fn);
2499 WARN_ON_ONCE(timer_pending(timer)); 1634 WARN_ON_ONCE(timer_pending(timer));
2500 WARN_ON_ONCE(!list_empty(&work->entry 1635 WARN_ON_ONCE(!list_empty(&work->entry));
2501 1636
2502 /* 1637 /*
2503 * If @delay is 0, queue @dwork->work 1638 * If @delay is 0, queue @dwork->work immediately. This is for
2504 * both optimization and correctness. 1639 * both optimization and correctness. The earliest @timer can
2505 * expire is on the closest next tick 1640 * expire is on the closest next tick and delayed_work users depend
2506 * on that there's no such delay when 1641 * on that there's no such delay when @delay is 0.
2507 */ 1642 */
2508 if (!delay) { 1643 if (!delay) {
2509 __queue_work(cpu, wq, &dwork- 1644 __queue_work(cpu, wq, &dwork->work);
2510 return; 1645 return;
2511 } 1646 }
2512 1647
2513 dwork->wq = wq; 1648 dwork->wq = wq;
2514 dwork->cpu = cpu; 1649 dwork->cpu = cpu;
2515 timer->expires = jiffies + delay; 1650 timer->expires = jiffies + delay;
2516 1651
2517 if (housekeeping_enabled(HK_TYPE_TIME !! 1652 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); 1653 add_timer_on(timer, cpu);
2523 } else { !! 1654 else
2524 if (likely(cpu == WORK_CPU_UN !! 1655 add_timer(timer);
2525 add_timer_global(time <<
2526 else <<
2527 add_timer_on(timer, c <<
2528 } <<
2529 } 1656 }
2530 1657
2531 /** 1658 /**
2532 * queue_delayed_work_on - queue work on spec 1659 * queue_delayed_work_on - queue work on specific CPU after delay
2533 * @cpu: CPU number to execute work on 1660 * @cpu: CPU number to execute work on
2534 * @wq: workqueue to use 1661 * @wq: workqueue to use
2535 * @dwork: work to queue 1662 * @dwork: work to queue
2536 * @delay: number of jiffies to wait before q 1663 * @delay: number of jiffies to wait before queueing
2537 * 1664 *
2538 * Return: %false if @work was already on a q 1665 * Return: %false if @work was already on a queue, %true otherwise. If
2539 * @delay is zero and @dwork is idle, it will 1666 * @delay is zero and @dwork is idle, it will be scheduled for immediate
2540 * execution. 1667 * execution.
2541 */ 1668 */
2542 bool queue_delayed_work_on(int cpu, struct wo 1669 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
2543 struct delayed_wor 1670 struct delayed_work *dwork, unsigned long delay)
2544 { 1671 {
2545 struct work_struct *work = &dwork->wo 1672 struct work_struct *work = &dwork->work;
2546 bool ret = false; 1673 bool ret = false;
2547 unsigned long irq_flags; !! 1674 unsigned long flags;
2548 1675
2549 /* read the comment in __queue_work() 1676 /* read the comment in __queue_work() */
2550 local_irq_save(irq_flags); !! 1677 local_irq_save(flags);
2551 1678
2552 if (!test_and_set_bit(WORK_STRUCT_PEN !! 1679 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, 1680 __queue_delayed_work(cpu, wq, dwork, delay);
2555 ret = true; 1681 ret = true;
2556 } 1682 }
2557 1683
2558 local_irq_restore(irq_flags); !! 1684 local_irq_restore(flags);
2559 return ret; 1685 return ret;
2560 } 1686 }
2561 EXPORT_SYMBOL(queue_delayed_work_on); 1687 EXPORT_SYMBOL(queue_delayed_work_on);
2562 1688
2563 /** 1689 /**
2564 * mod_delayed_work_on - modify delay of or q 1690 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
2565 * @cpu: CPU number to execute work on 1691 * @cpu: CPU number to execute work on
2566 * @wq: workqueue to use 1692 * @wq: workqueue to use
2567 * @dwork: work to queue 1693 * @dwork: work to queue
2568 * @delay: number of jiffies to wait before q 1694 * @delay: number of jiffies to wait before queueing
2569 * 1695 *
2570 * If @dwork is idle, equivalent to queue_del 1696 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
2571 * modify @dwork's timer so that it expires a 1697 * modify @dwork's timer so that it expires after @delay. If @delay is
2572 * zero, @work is guaranteed to be scheduled 1698 * zero, @work is guaranteed to be scheduled immediately regardless of its
2573 * current state. 1699 * current state.
2574 * 1700 *
2575 * Return: %false if @dwork was idle and queu 1701 * Return: %false if @dwork was idle and queued, %true if @dwork was
2576 * pending and its timer was modified. 1702 * pending and its timer was modified.
2577 * 1703 *
2578 * This function is safe to call from any con 1704 * This function is safe to call from any context including IRQ handler.
2579 * See try_to_grab_pending() for details. 1705 * See try_to_grab_pending() for details.
2580 */ 1706 */
2581 bool mod_delayed_work_on(int cpu, struct work 1707 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
2582 struct delayed_work 1708 struct delayed_work *dwork, unsigned long delay)
2583 { 1709 {
2584 unsigned long irq_flags; !! 1710 unsigned long flags;
2585 bool ret; !! 1711 int ret;
2586 1712
2587 ret = work_grab_pending(&dwork->work, !! 1713 do {
>> 1714 ret = try_to_grab_pending(&dwork->work, true, &flags);
>> 1715 } while (unlikely(ret == -EAGAIN));
2588 1716
2589 if (!clear_pending_if_disabled(&dwork !! 1717 if (likely(ret >= 0)) {
2590 __queue_delayed_work(cpu, wq, 1718 __queue_delayed_work(cpu, wq, dwork, delay);
>> 1719 local_irq_restore(flags);
>> 1720 }
2591 1721
2592 local_irq_restore(irq_flags); !! 1722 /* -ENOENT from try_to_grab_pending() becomes %true */
2593 return ret; 1723 return ret;
2594 } 1724 }
2595 EXPORT_SYMBOL_GPL(mod_delayed_work_on); 1725 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
2596 1726
2597 static void rcu_work_rcufn(struct rcu_head *r 1727 static void rcu_work_rcufn(struct rcu_head *rcu)
2598 { 1728 {
2599 struct rcu_work *rwork = container_of 1729 struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu);
2600 1730
2601 /* read the comment in __queue_work() 1731 /* read the comment in __queue_work() */
2602 local_irq_disable(); 1732 local_irq_disable();
2603 __queue_work(WORK_CPU_UNBOUND, rwork- 1733 __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work);
2604 local_irq_enable(); 1734 local_irq_enable();
2605 } 1735 }
2606 1736
2607 /** 1737 /**
2608 * queue_rcu_work - queue work after a RCU gr 1738 * queue_rcu_work - queue work after a RCU grace period
2609 * @wq: workqueue to use 1739 * @wq: workqueue to use
2610 * @rwork: work to queue 1740 * @rwork: work to queue
2611 * 1741 *
2612 * Return: %false if @rwork was already pendi 1742 * Return: %false if @rwork was already pending, %true otherwise. Note
2613 * that a full RCU grace period is guaranteed 1743 * that a full RCU grace period is guaranteed only after a %true return.
2614 * While @rwork is guaranteed to be executed 1744 * While @rwork is guaranteed to be executed after a %false return, the
2615 * execution may happen before a full RCU gra 1745 * execution may happen before a full RCU grace period has passed.
2616 */ 1746 */
2617 bool queue_rcu_work(struct workqueue_struct * 1747 bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork)
2618 { 1748 {
2619 struct work_struct *work = &rwork->wo 1749 struct work_struct *work = &rwork->work;
2620 1750
2621 /* !! 1751 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; 1752 rwork->wq = wq;
2628 call_rcu_hurry(&rwork->rcu, r !! 1753 call_rcu(&rwork->rcu, rcu_work_rcufn);
2629 return true; 1754 return true;
2630 } 1755 }
2631 1756
2632 return false; 1757 return false;
2633 } 1758 }
2634 EXPORT_SYMBOL(queue_rcu_work); 1759 EXPORT_SYMBOL(queue_rcu_work);
2635 1760
>> 1761 /**
>> 1762 * worker_enter_idle - enter idle state
>> 1763 * @worker: worker which is entering idle state
>> 1764 *
>> 1765 * @worker is entering idle state. Update stats and idle timer if
>> 1766 * necessary.
>> 1767 *
>> 1768 * LOCKING:
>> 1769 * raw_spin_lock_irq(pool->lock).
>> 1770 */
>> 1771 static void worker_enter_idle(struct worker *worker)
>> 1772 {
>> 1773 struct worker_pool *pool = worker->pool;
>> 1774
>> 1775 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
>> 1776 WARN_ON_ONCE(!list_empty(&worker->entry) &&
>> 1777 (worker->hentry.next || worker->hentry.pprev)))
>> 1778 return;
>> 1779
>> 1780 /* can't use worker_set_flags(), also called from create_worker() */
>> 1781 worker->flags |= WORKER_IDLE;
>> 1782 pool->nr_idle++;
>> 1783 worker->last_active = jiffies;
>> 1784
>> 1785 /* idle_list is LIFO */
>> 1786 list_add(&worker->entry, &pool->idle_list);
>> 1787
>> 1788 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
>> 1789 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
>> 1790
>> 1791 /*
>> 1792 * Sanity check nr_running. Because unbind_workers() releases
>> 1793 * pool->lock between setting %WORKER_UNBOUND and zapping
>> 1794 * nr_running, the warning may trigger spuriously. Check iff
>> 1795 * unbind is not in progress.
>> 1796 */
>> 1797 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
>> 1798 pool->nr_workers == pool->nr_idle &&
>> 1799 atomic_read(&pool->nr_running));
>> 1800 }
>> 1801
>> 1802 /**
>> 1803 * worker_leave_idle - leave idle state
>> 1804 * @worker: worker which is leaving idle state
>> 1805 *
>> 1806 * @worker is leaving idle state. Update stats.
>> 1807 *
>> 1808 * LOCKING:
>> 1809 * raw_spin_lock_irq(pool->lock).
>> 1810 */
>> 1811 static void worker_leave_idle(struct worker *worker)
>> 1812 {
>> 1813 struct worker_pool *pool = worker->pool;
>> 1814
>> 1815 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
>> 1816 return;
>> 1817 worker_clr_flags(worker, WORKER_IDLE);
>> 1818 pool->nr_idle--;
>> 1819 list_del_init(&worker->entry);
>> 1820 }
>> 1821
2636 static struct worker *alloc_worker(int node) 1822 static struct worker *alloc_worker(int node)
2637 { 1823 {
2638 struct worker *worker; 1824 struct worker *worker;
2639 1825
2640 worker = kzalloc_node(sizeof(*worker) 1826 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
2641 if (worker) { 1827 if (worker) {
2642 INIT_LIST_HEAD(&worker->entry 1828 INIT_LIST_HEAD(&worker->entry);
2643 INIT_LIST_HEAD(&worker->sched 1829 INIT_LIST_HEAD(&worker->scheduled);
2644 INIT_LIST_HEAD(&worker->node) 1830 INIT_LIST_HEAD(&worker->node);
2645 /* on creation a worker is in 1831 /* on creation a worker is in !idle && prep state */
2646 worker->flags = WORKER_PREP; 1832 worker->flags = WORKER_PREP;
2647 } 1833 }
2648 return worker; 1834 return worker;
2649 } 1835 }
2650 1836
2651 static cpumask_t *pool_allowed_cpus(struct wo <<
2652 { <<
2653 if (pool->cpu < 0 && pool->attrs->aff <<
2654 return pool->attrs->__pod_cpu <<
2655 else <<
2656 return pool->attrs->cpumask; <<
2657 } <<
2658 <<
2659 /** 1837 /**
2660 * worker_attach_to_pool() - attach a worker 1838 * worker_attach_to_pool() - attach a worker to a pool
2661 * @worker: worker to be attached 1839 * @worker: worker to be attached
2662 * @pool: the target pool 1840 * @pool: the target pool
2663 * 1841 *
2664 * Attach @worker to @pool. Once attached, t 1842 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
2665 * cpu-binding of @worker are kept coordinate 1843 * cpu-binding of @worker are kept coordinated with the pool across
2666 * cpu-[un]hotplugs. 1844 * cpu-[un]hotplugs.
2667 */ 1845 */
2668 static void worker_attach_to_pool(struct work 1846 static void worker_attach_to_pool(struct worker *worker,
2669 struct work !! 1847 struct worker_pool *pool)
2670 { 1848 {
2671 mutex_lock(&wq_pool_attach_mutex); 1849 mutex_lock(&wq_pool_attach_mutex);
2672 1850
2673 /* 1851 /*
2674 * The wq_pool_attach_mutex ensures % !! 1852 * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains
2675 * across this function. See the comm !! 1853 * stable across this function. See the comments above the flag
2676 * details. BH workers are, while per !! 1854 * definition for details.
2677 */ 1855 */
2678 if (pool->flags & POOL_DISASSOCIATED) !! 1856 if (pool->flags & POOL_DISASSOCIATED)
2679 worker->flags |= WORKER_UNBOU 1857 worker->flags |= WORKER_UNBOUND;
2680 } else { !! 1858 else
2681 WARN_ON_ONCE(pool->flags & PO <<
2682 kthread_set_per_cpu(worker->t 1859 kthread_set_per_cpu(worker->task, pool->cpu);
2683 } <<
2684 1860
2685 if (worker->rescue_wq) 1861 if (worker->rescue_wq)
2686 set_cpus_allowed_ptr(worker-> !! 1862 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
2687 1863
2688 list_add_tail(&worker->node, &pool->w 1864 list_add_tail(&worker->node, &pool->workers);
2689 worker->pool = pool; 1865 worker->pool = pool;
2690 1866
2691 mutex_unlock(&wq_pool_attach_mutex); 1867 mutex_unlock(&wq_pool_attach_mutex);
2692 } 1868 }
2693 1869
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 /** 1870 /**
2715 * worker_detach_from_pool() - detach a worke 1871 * worker_detach_from_pool() - detach a worker from its pool
2716 * @worker: worker which is attached to its p 1872 * @worker: worker which is attached to its pool
2717 * 1873 *
2718 * Undo the attaching which had been done in 1874 * Undo the attaching which had been done in worker_attach_to_pool(). The
2719 * caller worker shouldn't access to the pool 1875 * caller worker shouldn't access to the pool after detached except it has
2720 * other reference to the pool. 1876 * other reference to the pool.
2721 */ 1877 */
2722 static void worker_detach_from_pool(struct wo 1878 static void worker_detach_from_pool(struct worker *worker)
2723 { 1879 {
2724 struct worker_pool *pool = worker->po 1880 struct worker_pool *pool = worker->pool;
2725 !! 1881 struct completion *detach_completion = NULL;
2726 /* there is one permanent BH worker p <<
2727 WARN_ON_ONCE(pool->flags & POOL_BH); <<
2728 1882
2729 mutex_lock(&wq_pool_attach_mutex); 1883 mutex_lock(&wq_pool_attach_mutex);
2730 detach_worker(worker); !! 1884
>> 1885 kthread_set_per_cpu(worker->task, -1);
>> 1886 list_del(&worker->node);
2731 worker->pool = NULL; 1887 worker->pool = NULL;
>> 1888
>> 1889 if (list_empty(&pool->workers))
>> 1890 detach_completion = pool->detach_completion;
2732 mutex_unlock(&wq_pool_attach_mutex); 1891 mutex_unlock(&wq_pool_attach_mutex);
2733 1892
2734 /* clear leftover flags without pool- 1893 /* clear leftover flags without pool->lock after it is detached */
2735 worker->flags &= ~(WORKER_UNBOUND | W 1894 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 1895
2745 if (pool) { !! 1896 if (detach_completion)
2746 if (pool->cpu >= 0) !! 1897 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 } 1898 }
2757 1899
2758 /** 1900 /**
2759 * create_worker - create a new workqueue wor 1901 * create_worker - create a new workqueue worker
2760 * @pool: pool the new worker will belong to 1902 * @pool: pool the new worker will belong to
2761 * 1903 *
2762 * Create and start a new worker which is att 1904 * Create and start a new worker which is attached to @pool.
2763 * 1905 *
2764 * CONTEXT: 1906 * CONTEXT:
2765 * Might sleep. Does GFP_KERNEL allocations. 1907 * Might sleep. Does GFP_KERNEL allocations.
2766 * 1908 *
2767 * Return: 1909 * Return:
2768 * Pointer to the newly created worker. 1910 * Pointer to the newly created worker.
2769 */ 1911 */
2770 static struct worker *create_worker(struct wo 1912 static struct worker *create_worker(struct worker_pool *pool)
2771 { 1913 {
2772 struct worker *worker; !! 1914 struct worker *worker = NULL;
2773 int id; !! 1915 int id = -1;
>> 1916 char id_buf[16];
2774 1917
2775 /* ID is needed to determine kthread 1918 /* ID is needed to determine kthread name */
2776 id = ida_alloc(&pool->worker_ida, GFP !! 1919 id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL);
2777 if (id < 0) { !! 1920 if (id < 0)
2778 pr_err_once("workqueue: Faile !! 1921 goto fail;
2779 ERR_PTR(id)); <<
2780 return NULL; <<
2781 } <<
2782 1922
2783 worker = alloc_worker(pool->node); 1923 worker = alloc_worker(pool->node);
2784 if (!worker) { !! 1924 if (!worker)
2785 pr_err_once("workqueue: Faile <<
2786 goto fail; 1925 goto fail;
2787 } <<
2788 1926
2789 worker->id = id; 1927 worker->id = id;
2790 1928
2791 if (!(pool->flags & POOL_BH)) { !! 1929 if (pool->cpu >= 0)
2792 char id_buf[WORKER_ID_LEN]; !! 1930 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
>> 1931 pool->attrs->nice < 0 ? "H" : "");
>> 1932 else
>> 1933 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
2793 1934
2794 format_worker_id(id_buf, size !! 1935 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
2795 worker->task = kthread_create !! 1936 "kworker/%s", id_buf);
2796 !! 1937 if (IS_ERR(worker->task))
2797 if (IS_ERR(worker->task)) { !! 1938 goto fail;
2798 if (PTR_ERR(worker->t <<
2799 pr_err("workq <<
2800 id_buf <<
2801 } else { <<
2802 pr_err_once(" <<
2803 w <<
2804 } <<
2805 goto fail; <<
2806 } <<
2807 1939
2808 set_user_nice(worker->task, p !! 1940 set_user_nice(worker->task, pool->attrs->nice);
2809 kthread_bind_mask(worker->tas !! 1941 kthread_bind_mask(worker->task, pool->attrs->cpumask);
2810 } <<
2811 1942
2812 /* successful, attach the worker to t 1943 /* successful, attach the worker to the pool */
2813 worker_attach_to_pool(worker, pool); 1944 worker_attach_to_pool(worker, pool);
2814 1945
2815 /* start the newly created worker */ 1946 /* start the newly created worker */
2816 raw_spin_lock_irq(&pool->lock); 1947 raw_spin_lock_irq(&pool->lock);
2817 <<
2818 worker->pool->nr_workers++; 1948 worker->pool->nr_workers++;
2819 worker_enter_idle(worker); 1949 worker_enter_idle(worker);
2820 !! 1950 wake_up_process(worker->task);
2821 /* <<
2822 * @worker is waiting on a completion <<
2823 * check if not woken up soon. As kic <<
2824 * wake it up explicitly. <<
2825 */ <<
2826 if (worker->task) <<
2827 wake_up_process(worker->task) <<
2828 <<
2829 raw_spin_unlock_irq(&pool->lock); 1951 raw_spin_unlock_irq(&pool->lock);
2830 1952
2831 return worker; 1953 return worker;
2832 1954
2833 fail: 1955 fail:
2834 ida_free(&pool->worker_ida, id); !! 1956 if (id >= 0)
>> 1957 ida_simple_remove(&pool->worker_ida, id);
2835 kfree(worker); 1958 kfree(worker);
2836 return NULL; 1959 return NULL;
2837 } 1960 }
2838 1961
2839 static void detach_dying_workers(struct list_ <<
2840 { <<
2841 struct worker *worker; <<
2842 <<
2843 list_for_each_entry(worker, cull_list <<
2844 detach_worker(worker); <<
2845 } <<
2846 <<
2847 static void reap_dying_workers(struct list_he <<
2848 { <<
2849 struct worker *worker, *tmp; <<
2850 <<
2851 list_for_each_entry_safe(worker, tmp, <<
2852 list_del_init(&worker->entry) <<
2853 kthread_stop_put(worker->task <<
2854 kfree(worker); <<
2855 } <<
2856 } <<
2857 <<
2858 /** 1962 /**
2859 * set_worker_dying - Tag a worker for destru !! 1963 * destroy_worker - destroy a workqueue worker
2860 * @worker: worker to be destroyed 1964 * @worker: worker to be destroyed
2861 * @list: transfer worker away from its pool- <<
2862 * 1965 *
2863 * Tag @worker for destruction and adjust @po !! 1966 * Destroy @worker and adjust @pool stats accordingly. The worker should
2864 * should be idle. !! 1967 * be idle.
2865 * 1968 *
2866 * CONTEXT: 1969 * CONTEXT:
2867 * raw_spin_lock_irq(pool->lock). 1970 * raw_spin_lock_irq(pool->lock).
2868 */ 1971 */
2869 static void set_worker_dying(struct worker *w !! 1972 static void destroy_worker(struct worker *worker)
2870 { 1973 {
2871 struct worker_pool *pool = worker->po 1974 struct worker_pool *pool = worker->pool;
2872 1975
2873 lockdep_assert_held(&pool->lock); 1976 lockdep_assert_held(&pool->lock);
2874 lockdep_assert_held(&wq_pool_attach_m <<
2875 1977
2876 /* sanity check frenzy */ 1978 /* sanity check frenzy */
2877 if (WARN_ON(worker->current_work) || 1979 if (WARN_ON(worker->current_work) ||
2878 WARN_ON(!list_empty(&worker->sche 1980 WARN_ON(!list_empty(&worker->scheduled)) ||
2879 WARN_ON(!(worker->flags & WORKER_ 1981 WARN_ON(!(worker->flags & WORKER_IDLE)))
2880 return; 1982 return;
2881 1983
2882 pool->nr_workers--; 1984 pool->nr_workers--;
2883 pool->nr_idle--; 1985 pool->nr_idle--;
2884 1986
>> 1987 list_del_init(&worker->entry);
2885 worker->flags |= WORKER_DIE; 1988 worker->flags |= WORKER_DIE;
2886 !! 1989 wake_up_process(worker->task);
2887 list_move(&worker->entry, list); <<
2888 <<
2889 /* get an extra task struct reference <<
2890 get_task_struct(worker->task); <<
2891 } 1990 }
2892 1991
2893 /** <<
2894 * idle_worker_timeout - check if some idle w <<
2895 * @t: The pool's idle_timer that just expire <<
2896 * <<
2897 * The timer is armed in worker_enter_idle(). <<
2898 * worker_leave_idle(), as a worker flicking <<
2899 * pool is at the too_many_workers() tipping <<
2900 * housekeeping overhead. Since IDLE_WORKER_T <<
2901 * it expire and re-evaluate things from ther <<
2902 */ <<
2903 static void idle_worker_timeout(struct timer_ 1992 static void idle_worker_timeout(struct timer_list *t)
2904 { 1993 {
2905 struct worker_pool *pool = from_timer 1994 struct worker_pool *pool = from_timer(pool, t, idle_timer);
2906 bool do_cull = false; <<
2907 <<
2908 if (work_pending(&pool->idle_cull_wor <<
2909 return; <<
2910 <<
2911 raw_spin_lock_irq(&pool->lock); <<
2912 <<
2913 if (too_many_workers(pool)) { <<
2914 struct worker *worker; <<
2915 unsigned long expires; <<
2916 <<
2917 /* idle_list is kept in LIFO <<
2918 worker = list_last_entry(&poo <<
2919 expires = worker->last_active <<
2920 do_cull = !time_before(jiffie <<
2921 <<
2922 if (!do_cull) <<
2923 mod_timer(&pool->idle <<
2924 } <<
2925 raw_spin_unlock_irq(&pool->lock); <<
2926 <<
2927 if (do_cull) <<
2928 queue_work(system_unbound_wq, <<
2929 } <<
2930 1995
2931 /** <<
2932 * idle_cull_fn - cull workers that have been <<
2933 * @work: the pool's work for handling these <<
2934 * <<
2935 * This goes through a pool's idle workers an <<
2936 * idle for at least IDLE_WORKER_TIMEOUT seco <<
2937 * <<
2938 * We don't want to disturb isolated CPUs bec <<
2939 * culled, so this also resets worker affinit <<
2940 * context, hence the split between timer cal <<
2941 */ <<
2942 static void idle_cull_fn(struct work_struct * <<
2943 { <<
2944 struct worker_pool *pool = container_ <<
2945 LIST_HEAD(cull_list); <<
2946 <<
2947 /* <<
2948 * Grabbing wq_pool_attach_mutex here <<
2949 * cannot proceed beyong set_pf_worke <<
2950 * This is required as a previously-p <<
2951 * set_worker_dying() has happened bu <<
2952 */ <<
2953 mutex_lock(&wq_pool_attach_mutex); <<
2954 raw_spin_lock_irq(&pool->lock); 1996 raw_spin_lock_irq(&pool->lock);
2955 1997
2956 while (too_many_workers(pool)) { 1998 while (too_many_workers(pool)) {
2957 struct worker *worker; 1999 struct worker *worker;
2958 unsigned long expires; 2000 unsigned long expires;
2959 2001
2960 worker = list_last_entry(&poo !! 2002 /* idle_list is kept in LIFO order, check the last one */
>> 2003 worker = list_entry(pool->idle_list.prev, struct worker, entry);
2961 expires = worker->last_active 2004 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2962 2005
2963 if (time_before(jiffies, expi 2006 if (time_before(jiffies, expires)) {
2964 mod_timer(&pool->idle 2007 mod_timer(&pool->idle_timer, expires);
2965 break; 2008 break;
2966 } 2009 }
2967 2010
2968 set_worker_dying(worker, &cul !! 2011 destroy_worker(worker);
2969 } 2012 }
2970 2013
2971 raw_spin_unlock_irq(&pool->lock); 2014 raw_spin_unlock_irq(&pool->lock);
2972 detach_dying_workers(&cull_list); <<
2973 mutex_unlock(&wq_pool_attach_mutex); <<
2974 <<
2975 reap_dying_workers(&cull_list); <<
2976 } 2015 }
2977 2016
2978 static void send_mayday(struct work_struct *w 2017 static void send_mayday(struct work_struct *work)
2979 { 2018 {
2980 struct pool_workqueue *pwq = get_work 2019 struct pool_workqueue *pwq = get_work_pwq(work);
2981 struct workqueue_struct *wq = pwq->wq 2020 struct workqueue_struct *wq = pwq->wq;
2982 2021
2983 lockdep_assert_held(&wq_mayday_lock); 2022 lockdep_assert_held(&wq_mayday_lock);
2984 2023
2985 if (!wq->rescuer) 2024 if (!wq->rescuer)
2986 return; 2025 return;
2987 2026
2988 /* mayday mayday mayday */ 2027 /* mayday mayday mayday */
2989 if (list_empty(&pwq->mayday_node)) { 2028 if (list_empty(&pwq->mayday_node)) {
2990 /* 2029 /*
2991 * If @pwq is for an unbound 2030 * If @pwq is for an unbound wq, its base ref may be put at
2992 * any time due to an attribu 2031 * any time due to an attribute change. Pin @pwq until the
2993 * rescuer is done with it. 2032 * rescuer is done with it.
2994 */ 2033 */
2995 get_pwq(pwq); 2034 get_pwq(pwq);
2996 list_add_tail(&pwq->mayday_no 2035 list_add_tail(&pwq->mayday_node, &wq->maydays);
2997 wake_up_process(wq->rescuer-> 2036 wake_up_process(wq->rescuer->task);
2998 pwq->stats[PWQ_STAT_MAYDAY]++ <<
2999 } 2037 }
3000 } 2038 }
3001 2039
3002 static void pool_mayday_timeout(struct timer_ 2040 static void pool_mayday_timeout(struct timer_list *t)
3003 { 2041 {
3004 struct worker_pool *pool = from_timer 2042 struct worker_pool *pool = from_timer(pool, t, mayday_timer);
3005 struct work_struct *work; 2043 struct work_struct *work;
3006 2044
3007 raw_spin_lock_irq(&pool->lock); 2045 raw_spin_lock_irq(&pool->lock);
3008 raw_spin_lock(&wq_mayday_lock); 2046 raw_spin_lock(&wq_mayday_lock); /* for wq->maydays */
3009 2047
3010 if (need_to_create_worker(pool)) { 2048 if (need_to_create_worker(pool)) {
3011 /* 2049 /*
3012 * We've been trying to creat 2050 * We've been trying to create a new worker but
3013 * haven't been successful. 2051 * haven't been successful. We might be hitting an
3014 * allocation deadlock. Send 2052 * allocation deadlock. Send distress signals to
3015 * rescuers. 2053 * rescuers.
3016 */ 2054 */
3017 list_for_each_entry(work, &po 2055 list_for_each_entry(work, &pool->worklist, entry)
3018 send_mayday(work); 2056 send_mayday(work);
3019 } 2057 }
3020 2058
3021 raw_spin_unlock(&wq_mayday_lock); 2059 raw_spin_unlock(&wq_mayday_lock);
3022 raw_spin_unlock_irq(&pool->lock); 2060 raw_spin_unlock_irq(&pool->lock);
3023 2061
3024 mod_timer(&pool->mayday_timer, jiffie 2062 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
3025 } 2063 }
3026 2064
3027 /** 2065 /**
3028 * maybe_create_worker - create a new worker 2066 * maybe_create_worker - create a new worker if necessary
3029 * @pool: pool to create a new worker for 2067 * @pool: pool to create a new worker for
3030 * 2068 *
3031 * Create a new worker for @pool if necessary 2069 * Create a new worker for @pool if necessary. @pool is guaranteed to
3032 * have at least one idle worker on return fr 2070 * have at least one idle worker on return from this function. If
3033 * creating a new worker takes longer than MA 2071 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
3034 * sent to all rescuers with works scheduled 2072 * sent to all rescuers with works scheduled on @pool to resolve
3035 * possible allocation deadlock. 2073 * possible allocation deadlock.
3036 * 2074 *
3037 * On return, need_to_create_worker() is guar 2075 * On return, need_to_create_worker() is guaranteed to be %false and
3038 * may_start_working() %true. 2076 * may_start_working() %true.
3039 * 2077 *
3040 * LOCKING: 2078 * LOCKING:
3041 * raw_spin_lock_irq(pool->lock) which may be 2079 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
3042 * multiple times. Does GFP_KERNEL allocatio 2080 * multiple times. Does GFP_KERNEL allocations. Called only from
3043 * manager. 2081 * manager.
3044 */ 2082 */
3045 static void maybe_create_worker(struct worker 2083 static void maybe_create_worker(struct worker_pool *pool)
3046 __releases(&pool->lock) 2084 __releases(&pool->lock)
3047 __acquires(&pool->lock) 2085 __acquires(&pool->lock)
3048 { 2086 {
3049 restart: 2087 restart:
3050 raw_spin_unlock_irq(&pool->lock); 2088 raw_spin_unlock_irq(&pool->lock);
3051 2089
3052 /* if we don't make progress in MAYDA 2090 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
3053 mod_timer(&pool->mayday_timer, jiffie 2091 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
3054 2092
3055 while (true) { 2093 while (true) {
3056 if (create_worker(pool) || !n 2094 if (create_worker(pool) || !need_to_create_worker(pool))
3057 break; 2095 break;
3058 2096
3059 schedule_timeout_interruptibl 2097 schedule_timeout_interruptible(CREATE_COOLDOWN);
3060 2098
3061 if (!need_to_create_worker(po 2099 if (!need_to_create_worker(pool))
3062 break; 2100 break;
3063 } 2101 }
3064 2102
3065 del_timer_sync(&pool->mayday_timer); 2103 del_timer_sync(&pool->mayday_timer);
3066 raw_spin_lock_irq(&pool->lock); 2104 raw_spin_lock_irq(&pool->lock);
3067 /* 2105 /*
3068 * This is necessary even after a new 2106 * This is necessary even after a new worker was just successfully
3069 * created as @pool->lock was dropped 2107 * created as @pool->lock was dropped and the new worker might have
3070 * already become busy. 2108 * already become busy.
3071 */ 2109 */
3072 if (need_to_create_worker(pool)) 2110 if (need_to_create_worker(pool))
3073 goto restart; 2111 goto restart;
3074 } 2112 }
3075 2113
3076 /** 2114 /**
3077 * manage_workers - manage worker pool 2115 * manage_workers - manage worker pool
3078 * @worker: self 2116 * @worker: self
3079 * 2117 *
3080 * Assume the manager role and manage the wor 2118 * Assume the manager role and manage the worker pool @worker belongs
3081 * to. At any given time, there can be only 2119 * to. At any given time, there can be only zero or one manager per
3082 * pool. The exclusion is handled automatica 2120 * pool. The exclusion is handled automatically by this function.
3083 * 2121 *
3084 * The caller can safely start processing wor 2122 * The caller can safely start processing works on false return. On
3085 * true return, it's guaranteed that need_to_ 2123 * true return, it's guaranteed that need_to_create_worker() is false
3086 * and may_start_working() is true. 2124 * and may_start_working() is true.
3087 * 2125 *
3088 * CONTEXT: 2126 * CONTEXT:
3089 * raw_spin_lock_irq(pool->lock) which may be 2127 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
3090 * multiple times. Does GFP_KERNEL allocatio 2128 * multiple times. Does GFP_KERNEL allocations.
3091 * 2129 *
3092 * Return: 2130 * Return:
3093 * %false if the pool doesn't need management 2131 * %false if the pool doesn't need management and the caller can safely
3094 * start processing works, %true if managemen 2132 * start processing works, %true if management function was performed and
3095 * the conditions that the caller verified be 2133 * the conditions that the caller verified before calling the function may
3096 * no longer be true. 2134 * no longer be true.
3097 */ 2135 */
3098 static bool manage_workers(struct worker *wor 2136 static bool manage_workers(struct worker *worker)
3099 { 2137 {
3100 struct worker_pool *pool = worker->po 2138 struct worker_pool *pool = worker->pool;
3101 2139
3102 if (pool->flags & POOL_MANAGER_ACTIVE 2140 if (pool->flags & POOL_MANAGER_ACTIVE)
3103 return false; 2141 return false;
3104 2142
3105 pool->flags |= POOL_MANAGER_ACTIVE; 2143 pool->flags |= POOL_MANAGER_ACTIVE;
3106 pool->manager = worker; 2144 pool->manager = worker;
3107 2145
3108 maybe_create_worker(pool); 2146 maybe_create_worker(pool);
3109 2147
3110 pool->manager = NULL; 2148 pool->manager = NULL;
3111 pool->flags &= ~POOL_MANAGER_ACTIVE; 2149 pool->flags &= ~POOL_MANAGER_ACTIVE;
3112 rcuwait_wake_up(&manager_wait); 2150 rcuwait_wake_up(&manager_wait);
3113 return true; 2151 return true;
3114 } 2152 }
3115 2153
3116 /** 2154 /**
3117 * process_one_work - process single work 2155 * process_one_work - process single work
3118 * @worker: self 2156 * @worker: self
3119 * @work: work to process 2157 * @work: work to process
3120 * 2158 *
3121 * Process @work. This function contains all 2159 * Process @work. This function contains all the logics necessary to
3122 * process a single work including synchroniz 2160 * process a single work including synchronization against and
3123 * interaction with other workers on the same 2161 * interaction with other workers on the same cpu, queueing and
3124 * flushing. As long as context requirement 2162 * flushing. As long as context requirement is met, any worker can
3125 * call this function to process a work. 2163 * call this function to process a work.
3126 * 2164 *
3127 * CONTEXT: 2165 * CONTEXT:
3128 * raw_spin_lock_irq(pool->lock) which is rel 2166 * raw_spin_lock_irq(pool->lock) which is released and regrabbed.
3129 */ 2167 */
3130 static void process_one_work(struct worker *w 2168 static void process_one_work(struct worker *worker, struct work_struct *work)
3131 __releases(&pool->lock) 2169 __releases(&pool->lock)
3132 __acquires(&pool->lock) 2170 __acquires(&pool->lock)
3133 { 2171 {
3134 struct pool_workqueue *pwq = get_work 2172 struct pool_workqueue *pwq = get_work_pwq(work);
3135 struct worker_pool *pool = worker->po 2173 struct worker_pool *pool = worker->pool;
3136 unsigned long work_data; !! 2174 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
3137 int lockdep_start_depth, rcu_start_de !! 2175 int work_color;
3138 bool bh_draining = pool->flags & POOL !! 2176 struct worker *collision;
3139 #ifdef CONFIG_LOCKDEP 2177 #ifdef CONFIG_LOCKDEP
3140 /* 2178 /*
3141 * It is permissible to free the stru 2179 * It is permissible to free the struct work_struct from
3142 * inside the function that is called 2180 * inside the function that is called from it, this we need to
3143 * take into account for lockdep too. 2181 * take into account for lockdep too. To avoid bogus "held
3144 * lock freed" warnings as well as pr 2182 * lock freed" warnings as well as problems when looking into
3145 * work->lockdep_map, make a copy and 2183 * work->lockdep_map, make a copy and use that here.
3146 */ 2184 */
3147 struct lockdep_map lockdep_map; 2185 struct lockdep_map lockdep_map;
3148 2186
3149 lockdep_copy_map(&lockdep_map, &work- 2187 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
3150 #endif 2188 #endif
3151 /* ensure we're on the correct CPU */ 2189 /* ensure we're on the correct CPU */
3152 WARN_ON_ONCE(!(pool->flags & POOL_DIS 2190 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
3153 raw_smp_processor_id() ! 2191 raw_smp_processor_id() != pool->cpu);
3154 2192
>> 2193 /*
>> 2194 * A single work shouldn't be executed concurrently by
>> 2195 * multiple workers on a single cpu. Check whether anyone is
>> 2196 * already processing the work. If so, defer the work to the
>> 2197 * currently executing one.
>> 2198 */
>> 2199 collision = find_worker_executing_work(pool, work);
>> 2200 if (unlikely(collision)) {
>> 2201 move_linked_works(work, &collision->scheduled, NULL);
>> 2202 return;
>> 2203 }
>> 2204
3155 /* claim and dequeue */ 2205 /* claim and dequeue */
3156 debug_work_deactivate(work); 2206 debug_work_deactivate(work);
3157 hash_add(pool->busy_hash, &worker->he 2207 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
3158 worker->current_work = work; 2208 worker->current_work = work;
3159 worker->current_func = work->func; 2209 worker->current_func = work->func;
3160 worker->current_pwq = pwq; 2210 worker->current_pwq = pwq;
3161 if (worker->task) !! 2211 work_color = get_work_color(work);
3162 worker->current_at = worker-> <<
3163 work_data = *work_data_bits(work); <<
3164 worker->current_color = get_work_colo <<
3165 2212
3166 /* 2213 /*
3167 * Record wq name for cmdline and deb 2214 * Record wq name for cmdline and debug reporting, may get
3168 * overridden through set_worker_desc 2215 * overridden through set_worker_desc().
3169 */ 2216 */
3170 strscpy(worker->desc, pwq->wq->name, 2217 strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN);
3171 2218
3172 list_del_init(&work->entry); 2219 list_del_init(&work->entry);
3173 2220
3174 /* 2221 /*
3175 * CPU intensive works don't particip 2222 * CPU intensive works don't participate in concurrency management.
3176 * They're the scheduler's responsibi 2223 * They're the scheduler's responsibility. This takes @worker out
3177 * of concurrency management and the 2224 * of concurrency management and the next code block will chain
3178 * execution of the pending work item 2225 * execution of the pending work items.
3179 */ 2226 */
3180 if (unlikely(pwq->wq->flags & WQ_CPU_ !! 2227 if (unlikely(cpu_intensive))
3181 worker_set_flags(worker, WORK 2228 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
3182 2229
3183 /* 2230 /*
3184 * Kick @pool if necessary. It's alwa !! 2231 * Wake up another worker if necessary. The condition is always
3185 * since nr_running would always be > !! 2232 * false for normal per-cpu workers since nr_running would always
3186 * chain execution of the pending wor !! 2233 * be >= 1 at this point. This is used to chain execution of the
3187 * workers such as the UNBOUND and CP !! 2234 * pending work items for WORKER_NOT_RUNNING workers such as the
>> 2235 * UNBOUND and CPU_INTENSIVE ones.
3188 */ 2236 */
3189 kick_pool(pool); !! 2237 if (need_more_worker(pool))
>> 2238 wake_up_worker(pool);
3190 2239
3191 /* 2240 /*
3192 * Record the last pool and clear PEN 2241 * Record the last pool and clear PENDING which should be the last
3193 * update to @work. Also, do this in 2242 * update to @work. Also, do this inside @pool->lock so that
3194 * PENDING and queued state changes h 2243 * PENDING and queued state changes happen together while IRQ is
3195 * disabled. 2244 * disabled.
3196 */ 2245 */
3197 set_work_pool_and_clear_pending(work, !! 2246 set_work_pool_and_clear_pending(work, pool->id);
3198 2247
3199 pwq->stats[PWQ_STAT_STARTED]++; <<
3200 raw_spin_unlock_irq(&pool->lock); 2248 raw_spin_unlock_irq(&pool->lock);
3201 2249
3202 rcu_start_depth = rcu_preempt_depth() !! 2250 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); 2251 lock_map_acquire(&lockdep_map);
3208 /* 2252 /*
3209 * Strictly speaking we should mark t 2253 * Strictly speaking we should mark the invariant state without holding
3210 * any locks, that is, before these t 2254 * any locks, that is, before these two lock_map_acquire()'s.
3211 * 2255 *
3212 * However, that would result in: 2256 * However, that would result in:
3213 * 2257 *
3214 * A(W1) 2258 * A(W1)
3215 * WFC(C) 2259 * WFC(C)
3216 * A(W1) 2260 * A(W1)
3217 * C(C) 2261 * C(C)
3218 * 2262 *
3219 * Which would create W1->C->W1 depen 2263 * Which would create W1->C->W1 dependencies, even though there is no
3220 * actual deadlock possible. There ar 2264 * actual deadlock possible. There are two solutions, using a
3221 * read-recursive acquire on the work 2265 * read-recursive acquire on the work(queue) 'locks', but this will then
3222 * hit the lockdep limitation on recu 2266 * hit the lockdep limitation on recursive locks, or simply discard
3223 * these locks. 2267 * these locks.
3224 * 2268 *
3225 * AFAICT there is no possible deadlo 2269 * AFAICT there is no possible deadlock scenario between the
3226 * flush_work() and complete() primit 2270 * flush_work() and complete() primitives (except for single-threaded
3227 * workqueues), so hiding them isn't 2271 * workqueues), so hiding them isn't a problem.
3228 */ 2272 */
3229 lockdep_invariant_state(true); 2273 lockdep_invariant_state(true);
3230 trace_workqueue_execute_start(work); 2274 trace_workqueue_execute_start(work);
3231 worker->current_func(work); 2275 worker->current_func(work);
3232 /* 2276 /*
3233 * While we must be careful to not us 2277 * While we must be careful to not use "work" after this, the trace
3234 * point will only record its address 2278 * point will only record its address.
3235 */ 2279 */
3236 trace_workqueue_execute_end(work, wor 2280 trace_workqueue_execute_end(work, worker->current_func);
3237 pwq->stats[PWQ_STAT_COMPLETED]++; <<
3238 lock_map_release(&lockdep_map); 2281 lock_map_release(&lockdep_map);
3239 if (!bh_draining) !! 2282 lock_map_release(&pwq->wq->lockdep_map);
3240 lock_map_release(&pwq->wq->lo <<
3241 2283
3242 if (unlikely((worker->task && in_atom !! 2284 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
3243 lockdep_depth(current) ! !! 2285 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
3244 rcu_preempt_depth() != r !! 2286 " last function: %ps\n",
3245 pr_err("BUG: workqueue leaked !! 2287 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); 2288 worker->current_func);
3251 debug_show_held_locks(current 2289 debug_show_held_locks(current);
3252 dump_stack(); 2290 dump_stack();
3253 } 2291 }
3254 2292
3255 /* 2293 /*
3256 * The following prevents a kworker f 2294 * The following prevents a kworker from hogging CPU on !PREEMPTION
3257 * kernels, where a requeueing work i 2295 * kernels, where a requeueing work item waiting for something to
3258 * happen could deadlock with stop_ma 2296 * happen could deadlock with stop_machine as such work item could
3259 * indefinitely requeue itself while 2297 * indefinitely requeue itself while all other CPUs are trapped in
3260 * stop_machine. At the same time, re 2298 * stop_machine. At the same time, report a quiescent RCU state so
3261 * the same condition doesn't freeze 2299 * the same condition doesn't freeze RCU.
3262 */ 2300 */
3263 if (worker->task) !! 2301 cond_resched();
3264 cond_resched(); <<
3265 2302
3266 raw_spin_lock_irq(&pool->lock); 2303 raw_spin_lock_irq(&pool->lock);
3267 2304
3268 /* !! 2305 /* clear cpu intensive status */
3269 * In addition to %WQ_CPU_INTENSIVE, !! 2306 if (unlikely(cpu_intensive))
3270 * CPU intensive by wq_worker_tick() !! 2307 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
3271 * wq_cpu_intensive_thresh_us. Clear <<
3272 */ <<
3273 worker_clr_flags(worker, WORKER_CPU_I <<
3274 2308
3275 /* tag the worker for identification 2309 /* tag the worker for identification in schedule() */
3276 worker->last_func = worker->current_f 2310 worker->last_func = worker->current_func;
3277 2311
3278 /* we're done with it, release */ 2312 /* we're done with it, release */
3279 hash_del(&worker->hentry); 2313 hash_del(&worker->hentry);
3280 worker->current_work = NULL; 2314 worker->current_work = NULL;
3281 worker->current_func = NULL; 2315 worker->current_func = NULL;
3282 worker->current_pwq = NULL; 2316 worker->current_pwq = NULL;
3283 worker->current_color = INT_MAX; !! 2317 pwq_dec_nr_in_flight(pwq, work_color);
3284 <<
3285 /* must be the last step, see the fun <<
3286 pwq_dec_nr_in_flight(pwq, work_data); <<
3287 } 2318 }
3288 2319
3289 /** 2320 /**
3290 * process_scheduled_works - process schedule 2321 * process_scheduled_works - process scheduled works
3291 * @worker: self 2322 * @worker: self
3292 * 2323 *
3293 * Process all scheduled works. Please note 2324 * Process all scheduled works. Please note that the scheduled list
3294 * may change while processing a work, so thi 2325 * may change while processing a work, so this function repeatedly
3295 * fetches a work from the top and executes i 2326 * fetches a work from the top and executes it.
3296 * 2327 *
3297 * CONTEXT: 2328 * CONTEXT:
3298 * raw_spin_lock_irq(pool->lock) which may be 2329 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
3299 * multiple times. 2330 * multiple times.
3300 */ 2331 */
3301 static void process_scheduled_works(struct wo 2332 static void process_scheduled_works(struct worker *worker)
3302 { 2333 {
3303 struct work_struct *work; !! 2334 while (!list_empty(&worker->scheduled)) {
3304 bool first = true; !! 2335 struct work_struct *work = list_first_entry(&worker->scheduled,
3305 !! 2336 struct work_struct, entry);
3306 while ((work = list_first_entry_or_nu <<
3307 <<
3308 if (first) { <<
3309 worker->pool->watchdo <<
3310 first = false; <<
3311 } <<
3312 process_one_work(worker, work 2337 process_one_work(worker, work);
3313 } 2338 }
3314 } 2339 }
3315 2340
3316 static void set_pf_worker(bool val) 2341 static void set_pf_worker(bool val)
3317 { 2342 {
3318 mutex_lock(&wq_pool_attach_mutex); 2343 mutex_lock(&wq_pool_attach_mutex);
3319 if (val) 2344 if (val)
3320 current->flags |= PF_WQ_WORKE 2345 current->flags |= PF_WQ_WORKER;
3321 else 2346 else
3322 current->flags &= ~PF_WQ_WORK 2347 current->flags &= ~PF_WQ_WORKER;
3323 mutex_unlock(&wq_pool_attach_mutex); 2348 mutex_unlock(&wq_pool_attach_mutex);
3324 } 2349 }
3325 2350
3326 /** 2351 /**
3327 * worker_thread - the worker thread function 2352 * worker_thread - the worker thread function
3328 * @__worker: self 2353 * @__worker: self
3329 * 2354 *
3330 * The worker thread function. All workers b 2355 * The worker thread function. All workers belong to a worker_pool -
3331 * either a per-cpu one or dynamic unbound on 2356 * either a per-cpu one or dynamic unbound one. These workers process all
3332 * work items regardless of their specific ta 2357 * work items regardless of their specific target workqueue. The only
3333 * exception is work items which belong to wo 2358 * exception is work items which belong to workqueues with a rescuer which
3334 * will be explained in rescuer_thread(). 2359 * will be explained in rescuer_thread().
3335 * 2360 *
3336 * Return: 0 2361 * Return: 0
3337 */ 2362 */
3338 static int worker_thread(void *__worker) 2363 static int worker_thread(void *__worker)
3339 { 2364 {
3340 struct worker *worker = __worker; 2365 struct worker *worker = __worker;
3341 struct worker_pool *pool = worker->po 2366 struct worker_pool *pool = worker->pool;
3342 2367
3343 /* tell the scheduler that this is a 2368 /* tell the scheduler that this is a workqueue worker */
3344 set_pf_worker(true); 2369 set_pf_worker(true);
3345 woke_up: 2370 woke_up:
3346 raw_spin_lock_irq(&pool->lock); 2371 raw_spin_lock_irq(&pool->lock);
3347 2372
3348 /* am I supposed to die? */ 2373 /* am I supposed to die? */
3349 if (unlikely(worker->flags & WORKER_D 2374 if (unlikely(worker->flags & WORKER_DIE)) {
3350 raw_spin_unlock_irq(&pool->lo 2375 raw_spin_unlock_irq(&pool->lock);
>> 2376 WARN_ON_ONCE(!list_empty(&worker->entry));
3351 set_pf_worker(false); 2377 set_pf_worker(false);
3352 /* !! 2378
3353 * The worker is dead and PF_ !! 2379 set_task_comm(worker->task, "kworker/dying");
3354 * shouldn't be accessed, res !! 2380 ida_simple_remove(&pool->worker_ida, worker->id);
3355 */ !! 2381 worker_detach_from_pool(worker);
3356 worker->pool = NULL; !! 2382 kfree(worker);
3357 ida_free(&pool->worker_ida, w <<
3358 return 0; 2383 return 0;
3359 } 2384 }
3360 2385
3361 worker_leave_idle(worker); 2386 worker_leave_idle(worker);
3362 recheck: 2387 recheck:
3363 /* no more worker necessary? */ 2388 /* no more worker necessary? */
3364 if (!need_more_worker(pool)) 2389 if (!need_more_worker(pool))
3365 goto sleep; 2390 goto sleep;
3366 2391
3367 /* do we need to manage? */ 2392 /* do we need to manage? */
3368 if (unlikely(!may_start_working(pool) 2393 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
3369 goto recheck; 2394 goto recheck;
3370 2395
3371 /* 2396 /*
3372 * ->scheduled list can only be fille 2397 * ->scheduled list can only be filled while a worker is
3373 * preparing to process a work or act 2398 * preparing to process a work or actually processing it.
3374 * Make sure nobody diddled with it w 2399 * Make sure nobody diddled with it while I was sleeping.
3375 */ 2400 */
3376 WARN_ON_ONCE(!list_empty(&worker->sch 2401 WARN_ON_ONCE(!list_empty(&worker->scheduled));
3377 2402
3378 /* 2403 /*
3379 * Finish PREP stage. We're guarante 2404 * Finish PREP stage. We're guaranteed to have at least one idle
3380 * worker or that someone else has al 2405 * worker or that someone else has already assumed the manager
3381 * role. This is where @worker start 2406 * role. This is where @worker starts participating in concurrency
3382 * management if applicable and concu 2407 * management if applicable and concurrency management is restored
3383 * after being rebound. See rebind_w 2408 * after being rebound. See rebind_workers() for details.
3384 */ 2409 */
3385 worker_clr_flags(worker, WORKER_PREP 2410 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
3386 2411
3387 do { 2412 do {
3388 struct work_struct *work = 2413 struct work_struct *work =
3389 list_first_entry(&poo 2414 list_first_entry(&pool->worklist,
3390 stru 2415 struct work_struct, entry);
3391 2416
3392 if (assign_work(work, worker, !! 2417 pool->watchdog_ts = jiffies;
>> 2418
>> 2419 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
>> 2420 /* optimization path, not strictly necessary */
>> 2421 process_one_work(worker, work);
>> 2422 if (unlikely(!list_empty(&worker->scheduled)))
>> 2423 process_scheduled_works(worker);
>> 2424 } else {
>> 2425 move_linked_works(work, &worker->scheduled, NULL);
3393 process_scheduled_wor 2426 process_scheduled_works(worker);
>> 2427 }
3394 } while (keep_working(pool)); 2428 } while (keep_working(pool));
3395 2429
3396 worker_set_flags(worker, WORKER_PREP) 2430 worker_set_flags(worker, WORKER_PREP);
3397 sleep: 2431 sleep:
3398 /* 2432 /*
3399 * pool->lock is held and there's no 2433 * pool->lock is held and there's no work to process and no need to
3400 * manage, sleep. Workers are woken 2434 * manage, sleep. Workers are woken up only while holding
3401 * pool->lock or from local cpu, so s 2435 * pool->lock or from local cpu, so setting the current state
3402 * before releasing pool->lock is eno 2436 * before releasing pool->lock is enough to prevent losing any
3403 * event. 2437 * event.
3404 */ 2438 */
3405 worker_enter_idle(worker); 2439 worker_enter_idle(worker);
3406 __set_current_state(TASK_IDLE); 2440 __set_current_state(TASK_IDLE);
3407 raw_spin_unlock_irq(&pool->lock); 2441 raw_spin_unlock_irq(&pool->lock);
3408 schedule(); 2442 schedule();
3409 goto woke_up; 2443 goto woke_up;
3410 } 2444 }
3411 2445
3412 /** 2446 /**
3413 * rescuer_thread - the rescuer thread functi 2447 * rescuer_thread - the rescuer thread function
3414 * @__rescuer: self 2448 * @__rescuer: self
3415 * 2449 *
3416 * Workqueue rescuer thread function. There' 2450 * Workqueue rescuer thread function. There's one rescuer for each
3417 * workqueue which has WQ_MEM_RECLAIM set. 2451 * workqueue which has WQ_MEM_RECLAIM set.
3418 * 2452 *
3419 * Regular work processing on a pool may bloc 2453 * Regular work processing on a pool may block trying to create a new
3420 * worker which uses GFP_KERNEL allocation wh 2454 * worker which uses GFP_KERNEL allocation which has slight chance of
3421 * developing into deadlock if some works cur 2455 * developing into deadlock if some works currently on the same queue
3422 * need to be processed to satisfy the GFP_KE 2456 * need to be processed to satisfy the GFP_KERNEL allocation. This is
3423 * the problem rescuer solves. 2457 * the problem rescuer solves.
3424 * 2458 *
3425 * When such condition is possible, the pool 2459 * When such condition is possible, the pool summons rescuers of all
3426 * workqueues which have works queued on the 2460 * workqueues which have works queued on the pool and let them process
3427 * those works so that forward progress can b 2461 * those works so that forward progress can be guaranteed.
3428 * 2462 *
3429 * This should happen rarely. 2463 * This should happen rarely.
3430 * 2464 *
3431 * Return: 0 2465 * Return: 0
3432 */ 2466 */
3433 static int rescuer_thread(void *__rescuer) 2467 static int rescuer_thread(void *__rescuer)
3434 { 2468 {
3435 struct worker *rescuer = __rescuer; 2469 struct worker *rescuer = __rescuer;
3436 struct workqueue_struct *wq = rescuer 2470 struct workqueue_struct *wq = rescuer->rescue_wq;
>> 2471 struct list_head *scheduled = &rescuer->scheduled;
3437 bool should_stop; 2472 bool should_stop;
3438 2473
3439 set_user_nice(current, RESCUER_NICE_L 2474 set_user_nice(current, RESCUER_NICE_LEVEL);
3440 2475
3441 /* 2476 /*
3442 * Mark rescuer as worker too. As WO 2477 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
3443 * doesn't participate in concurrency 2478 * doesn't participate in concurrency management.
3444 */ 2479 */
3445 set_pf_worker(true); 2480 set_pf_worker(true);
3446 repeat: 2481 repeat:
3447 set_current_state(TASK_IDLE); 2482 set_current_state(TASK_IDLE);
3448 2483
3449 /* 2484 /*
3450 * By the time the rescuer is request 2485 * By the time the rescuer is requested to stop, the workqueue
3451 * shouldn't have any work pending, b 2486 * shouldn't have any work pending, but @wq->maydays may still have
3452 * pwq(s) queued. This can happen by 2487 * pwq(s) queued. This can happen by non-rescuer workers consuming
3453 * all the work items before the resc 2488 * all the work items before the rescuer got to them. Go through
3454 * @wq->maydays processing before act 2489 * @wq->maydays processing before acting on should_stop so that the
3455 * list is always empty on exit. 2490 * list is always empty on exit.
3456 */ 2491 */
3457 should_stop = kthread_should_stop(); 2492 should_stop = kthread_should_stop();
3458 2493
3459 /* see whether any pwq is asking for 2494 /* see whether any pwq is asking for help */
3460 raw_spin_lock_irq(&wq_mayday_lock); 2495 raw_spin_lock_irq(&wq_mayday_lock);
3461 2496
3462 while (!list_empty(&wq->maydays)) { 2497 while (!list_empty(&wq->maydays)) {
3463 struct pool_workqueue *pwq = 2498 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
3464 struc 2499 struct pool_workqueue, mayday_node);
3465 struct worker_pool *pool = pw 2500 struct worker_pool *pool = pwq->pool;
3466 struct work_struct *work, *n; 2501 struct work_struct *work, *n;
>> 2502 bool first = true;
3467 2503
3468 __set_current_state(TASK_RUNN 2504 __set_current_state(TASK_RUNNING);
3469 list_del_init(&pwq->mayday_no 2505 list_del_init(&pwq->mayday_node);
3470 2506
3471 raw_spin_unlock_irq(&wq_mayda 2507 raw_spin_unlock_irq(&wq_mayday_lock);
3472 2508
3473 worker_attach_to_pool(rescuer 2509 worker_attach_to_pool(rescuer, pool);
3474 2510
3475 raw_spin_lock_irq(&pool->lock 2511 raw_spin_lock_irq(&pool->lock);
3476 2512
3477 /* 2513 /*
3478 * Slurp in all works issued 2514 * Slurp in all works issued via this workqueue and
3479 * process'em. 2515 * process'em.
3480 */ 2516 */
3481 WARN_ON_ONCE(!list_empty(&res !! 2517 WARN_ON_ONCE(!list_empty(scheduled));
3482 list_for_each_entry_safe(work 2518 list_for_each_entry_safe(work, n, &pool->worklist, entry) {
3483 if (get_work_pwq(work !! 2519 if (get_work_pwq(work) == pwq) {
3484 assign_work(work, !! 2520 if (first)
3485 pwq->stats[PW !! 2521 pool->watchdog_ts = jiffies;
>> 2522 move_linked_works(work, scheduled, &n);
>> 2523 }
>> 2524 first = false;
3486 } 2525 }
3487 2526
3488 if (!list_empty(&rescuer->sch !! 2527 if (!list_empty(scheduled)) {
3489 process_scheduled_wor 2528 process_scheduled_works(rescuer);
3490 2529
3491 /* 2530 /*
3492 * The above executio 2531 * The above execution of rescued work items could
3493 * have created more 2532 * have created more to rescue through
3494 * pwq_activate_first !! 2533 * pwq_activate_first_delayed() or chained
3495 * queueing. Let's p 2534 * queueing. Let's put @pwq back on mayday list so
3496 * that such back-to- 2535 * that such back-to-back work items, which may be
3497 * being used to reli 2536 * being used to relieve memory pressure, don't
3498 * incur MAYDAY_INTER 2537 * incur MAYDAY_INTERVAL delay inbetween.
3499 */ 2538 */
3500 if (pwq->nr_active && 2539 if (pwq->nr_active && need_to_create_worker(pool)) {
3501 raw_spin_lock 2540 raw_spin_lock(&wq_mayday_lock);
3502 /* 2541 /*
3503 * Queue iff 2542 * Queue iff we aren't racing destruction
3504 * and somebo 2543 * and somebody else hasn't queued it already.
3505 */ 2544 */
3506 if (wq->rescu 2545 if (wq->rescuer && list_empty(&pwq->mayday_node)) {
3507 get_p 2546 get_pwq(pwq);
3508 list_ 2547 list_add_tail(&pwq->mayday_node, &wq->maydays);
3509 } 2548 }
3510 raw_spin_unlo 2549 raw_spin_unlock(&wq_mayday_lock);
3511 } 2550 }
3512 } 2551 }
3513 2552
3514 /* 2553 /*
3515 * Put the reference grabbed 2554 * Put the reference grabbed by send_mayday(). @pool won't
3516 * go away while we're still 2555 * go away while we're still attached to it.
3517 */ 2556 */
3518 put_pwq(pwq); 2557 put_pwq(pwq);
3519 2558
3520 /* 2559 /*
3521 * Leave this pool. Notify re !! 2560 * Leave this pool. If need_more_worker() is %true, notify a
3522 * with 0 concurrency and sta !! 2561 * regular worker; otherwise, we end up with 0 concurrency
>> 2562 * and stalling the execution.
3523 */ 2563 */
3524 kick_pool(pool); !! 2564 if (need_more_worker(pool))
>> 2565 wake_up_worker(pool);
3525 2566
3526 raw_spin_unlock_irq(&pool->lo 2567 raw_spin_unlock_irq(&pool->lock);
3527 2568
3528 worker_detach_from_pool(rescu 2569 worker_detach_from_pool(rescuer);
3529 2570
3530 raw_spin_lock_irq(&wq_mayday_ 2571 raw_spin_lock_irq(&wq_mayday_lock);
3531 } 2572 }
3532 2573
3533 raw_spin_unlock_irq(&wq_mayday_lock); 2574 raw_spin_unlock_irq(&wq_mayday_lock);
3534 2575
3535 if (should_stop) { 2576 if (should_stop) {
3536 __set_current_state(TASK_RUNN 2577 __set_current_state(TASK_RUNNING);
3537 set_pf_worker(false); 2578 set_pf_worker(false);
3538 return 0; 2579 return 0;
3539 } 2580 }
3540 2581
3541 /* rescuers should never participate 2582 /* rescuers should never participate in concurrency management */
3542 WARN_ON_ONCE(!(rescuer->flags & WORKE 2583 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
3543 schedule(); 2584 schedule();
3544 goto repeat; 2585 goto repeat;
3545 } 2586 }
3546 2587
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 /** 2588 /**
3682 * check_flush_dependency - check for flush d 2589 * check_flush_dependency - check for flush dependency sanity
3683 * @target_wq: workqueue being flushed 2590 * @target_wq: workqueue being flushed
3684 * @target_work: work item being flushed (NUL 2591 * @target_work: work item being flushed (NULL for workqueue flushes)
3685 * 2592 *
3686 * %current is trying to flush the whole @tar 2593 * %current is trying to flush the whole @target_wq or @target_work on it.
3687 * If @target_wq doesn't have %WQ_MEM_RECLAIM 2594 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
3688 * reclaiming memory or running on a workqueu 2595 * reclaiming memory or running on a workqueue which doesn't have
3689 * %WQ_MEM_RECLAIM as that can break forward- 2596 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
3690 * a deadlock. 2597 * a deadlock.
3691 */ 2598 */
3692 static void check_flush_dependency(struct wor 2599 static void check_flush_dependency(struct workqueue_struct *target_wq,
3693 struct wor 2600 struct work_struct *target_work)
3694 { 2601 {
3695 work_func_t target_func = target_work 2602 work_func_t target_func = target_work ? target_work->func : NULL;
3696 struct worker *worker; 2603 struct worker *worker;
3697 2604
3698 if (target_wq->flags & WQ_MEM_RECLAIM 2605 if (target_wq->flags & WQ_MEM_RECLAIM)
3699 return; 2606 return;
3700 2607
3701 worker = current_wq_worker(); 2608 worker = current_wq_worker();
3702 2609
3703 WARN_ONCE(current->flags & PF_MEMALLO 2610 WARN_ONCE(current->flags & PF_MEMALLOC,
3704 "workqueue: PF_MEMALLOC tas 2611 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps",
3705 current->pid, current->comm 2612 current->pid, current->comm, target_wq->name, target_func);
3706 WARN_ONCE(worker && ((worker->current 2613 WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
3707 (WQ_MEM_RECLAIM 2614 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
3708 "workqueue: WQ_MEM_RECLAIM 2615 "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps",
3709 worker->current_pwq->wq->na 2616 worker->current_pwq->wq->name, worker->current_func,
3710 target_wq->name, target_fun 2617 target_wq->name, target_func);
3711 } 2618 }
3712 2619
3713 struct wq_barrier { 2620 struct wq_barrier {
3714 struct work_struct work; 2621 struct work_struct work;
3715 struct completion done; 2622 struct completion done;
3716 struct task_struct *task; /* pu 2623 struct task_struct *task; /* purely informational */
3717 }; 2624 };
3718 2625
3719 static void wq_barrier_func(struct work_struc 2626 static void wq_barrier_func(struct work_struct *work)
3720 { 2627 {
3721 struct wq_barrier *barr = container_o 2628 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
3722 complete(&barr->done); 2629 complete(&barr->done);
3723 } 2630 }
3724 2631
3725 /** 2632 /**
3726 * insert_wq_barrier - insert a barrier work 2633 * insert_wq_barrier - insert a barrier work
3727 * @pwq: pwq to insert barrier into 2634 * @pwq: pwq to insert barrier into
3728 * @barr: wq_barrier to insert 2635 * @barr: wq_barrier to insert
3729 * @target: target work to attach @barr to 2636 * @target: target work to attach @barr to
3730 * @worker: worker currently executing @targe 2637 * @worker: worker currently executing @target, NULL if @target is not executing
3731 * 2638 *
3732 * @barr is linked to @target such that @barr 2639 * @barr is linked to @target such that @barr is completed only after
3733 * @target finishes execution. Please note t 2640 * @target finishes execution. Please note that the ordering
3734 * guarantee is observed only with respect to 2641 * guarantee is observed only with respect to @target and on the local
3735 * cpu. 2642 * cpu.
3736 * 2643 *
3737 * Currently, a queued barrier can't be cance 2644 * Currently, a queued barrier can't be canceled. This is because
3738 * try_to_grab_pending() can't determine whet 2645 * try_to_grab_pending() can't determine whether the work to be
3739 * grabbed is at the head of the queue and th 2646 * grabbed is at the head of the queue and thus can't clear LINKED
3740 * flag of the previous work while there must 2647 * flag of the previous work while there must be a valid next work
3741 * after a work with LINKED flag set. 2648 * after a work with LINKED flag set.
3742 * 2649 *
3743 * Note that when @worker is non-NULL, @targe 2650 * Note that when @worker is non-NULL, @target may be modified
3744 * underneath us, so we can't reliably determ 2651 * underneath us, so we can't reliably determine pwq from @target.
3745 * 2652 *
3746 * CONTEXT: 2653 * CONTEXT:
3747 * raw_spin_lock_irq(pool->lock). 2654 * raw_spin_lock_irq(pool->lock).
3748 */ 2655 */
3749 static void insert_wq_barrier(struct pool_wor 2656 static void insert_wq_barrier(struct pool_workqueue *pwq,
3750 struct wq_barri 2657 struct wq_barrier *barr,
3751 struct work_str 2658 struct work_struct *target, struct worker *worker)
3752 { 2659 {
3753 static __maybe_unused struct lock_cla <<
3754 unsigned int work_flags = 0; <<
3755 unsigned int work_color; <<
3756 struct list_head *head; 2660 struct list_head *head;
>> 2661 unsigned int linked = 0;
3757 2662
3758 /* 2663 /*
3759 * debugobject calls are safe here ev 2664 * debugobject calls are safe here even with pool->lock locked
3760 * as we know for sure that this will 2665 * as we know for sure that this will not trigger any of the
3761 * checks and call back into the fixu 2666 * checks and call back into the fixup functions where we
3762 * might deadlock. 2667 * might deadlock.
3763 * <<
3764 * BH and threaded workqueues need se <<
3765 * spuriously triggering "inconsisten <<
3766 * usage". <<
3767 */ 2668 */
3768 INIT_WORK_ONSTACK_KEY(&barr->work, wq !! 2669 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
3769 (pwq->wq->flags <<
3770 __set_bit(WORK_STRUCT_PENDING_BIT, wo 2670 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
3771 2671
3772 init_completion_map(&barr->done, &tar 2672 init_completion_map(&barr->done, &target->lockdep_map);
3773 2673
3774 barr->task = current; 2674 barr->task = current;
3775 2675
3776 /* The barrier work item does not par <<
3777 work_flags |= WORK_STRUCT_INACTIVE; <<
3778 <<
3779 /* 2676 /*
3780 * If @target is currently being exec 2677 * If @target is currently being executed, schedule the
3781 * barrier to the worker; otherwise, 2678 * barrier to the worker; otherwise, put it after @target.
3782 */ 2679 */
3783 if (worker) { !! 2680 if (worker)
3784 head = worker->scheduled.next 2681 head = worker->scheduled.next;
3785 work_color = worker->current_ !! 2682 else {
3786 } else { <<
3787 unsigned long *bits = work_da 2683 unsigned long *bits = work_data_bits(target);
3788 2684
3789 head = target->entry.next; 2685 head = target->entry.next;
3790 /* there can already be other 2686 /* there can already be other linked works, inherit and set */
3791 work_flags |= *bits & WORK_ST !! 2687 linked = *bits & WORK_STRUCT_LINKED;
3792 work_color = get_work_color(* <<
3793 __set_bit(WORK_STRUCT_LINKED_ 2688 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
3794 } 2689 }
3795 2690
3796 pwq->nr_in_flight[work_color]++; !! 2691 debug_work_activate(&barr->work);
3797 work_flags |= work_color_to_flags(wor !! 2692 insert_work(pwq, &barr->work, head,
3798 !! 2693 work_color_to_flags(WORK_NO_COLOR) | linked);
3799 insert_work(pwq, &barr->work, head, w <<
3800 } 2694 }
3801 2695
3802 /** 2696 /**
3803 * flush_workqueue_prep_pwqs - prepare pwqs f 2697 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
3804 * @wq: workqueue being flushed 2698 * @wq: workqueue being flushed
3805 * @flush_color: new flush color, < 0 for no- 2699 * @flush_color: new flush color, < 0 for no-op
3806 * @work_color: new work color, < 0 for no-op 2700 * @work_color: new work color, < 0 for no-op
3807 * 2701 *
3808 * Prepare pwqs for workqueue flushing. 2702 * Prepare pwqs for workqueue flushing.
3809 * 2703 *
3810 * If @flush_color is non-negative, flush_col 2704 * If @flush_color is non-negative, flush_color on all pwqs should be
3811 * -1. If no pwq has in-flight commands at t 2705 * -1. If no pwq has in-flight commands at the specified color, all
3812 * pwq->flush_color's stay at -1 and %false i 2706 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
3813 * has in flight commands, its pwq->flush_col 2707 * has in flight commands, its pwq->flush_color is set to
3814 * @flush_color, @wq->nr_pwqs_to_flush is upd 2708 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
3815 * wakeup logic is armed and %true is returne 2709 * wakeup logic is armed and %true is returned.
3816 * 2710 *
3817 * The caller should have initialized @wq->fi 2711 * The caller should have initialized @wq->first_flusher prior to
3818 * calling this function with non-negative @f 2712 * calling this function with non-negative @flush_color. If
3819 * @flush_color is negative, no flush color u 2713 * @flush_color is negative, no flush color update is done and %false
3820 * is returned. 2714 * is returned.
3821 * 2715 *
3822 * If @work_color is non-negative, all pwqs s 2716 * If @work_color is non-negative, all pwqs should have the same
3823 * work_color which is previous to @work_colo 2717 * work_color which is previous to @work_color and all will be
3824 * advanced to @work_color. 2718 * advanced to @work_color.
3825 * 2719 *
3826 * CONTEXT: 2720 * CONTEXT:
3827 * mutex_lock(wq->mutex). 2721 * mutex_lock(wq->mutex).
3828 * 2722 *
3829 * Return: 2723 * Return:
3830 * %true if @flush_color >= 0 and there's som 2724 * %true if @flush_color >= 0 and there's something to flush. %false
3831 * otherwise. 2725 * otherwise.
3832 */ 2726 */
3833 static bool flush_workqueue_prep_pwqs(struct 2727 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
3834 int flu 2728 int flush_color, int work_color)
3835 { 2729 {
3836 bool wait = false; 2730 bool wait = false;
3837 struct pool_workqueue *pwq; 2731 struct pool_workqueue *pwq;
3838 2732
3839 if (flush_color >= 0) { 2733 if (flush_color >= 0) {
3840 WARN_ON_ONCE(atomic_read(&wq- 2734 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
3841 atomic_set(&wq->nr_pwqs_to_fl 2735 atomic_set(&wq->nr_pwqs_to_flush, 1);
3842 } 2736 }
3843 2737
3844 for_each_pwq(pwq, wq) { 2738 for_each_pwq(pwq, wq) {
3845 struct worker_pool *pool = pw 2739 struct worker_pool *pool = pwq->pool;
3846 2740
3847 raw_spin_lock_irq(&pool->lock 2741 raw_spin_lock_irq(&pool->lock);
3848 2742
3849 if (flush_color >= 0) { 2743 if (flush_color >= 0) {
3850 WARN_ON_ONCE(pwq->flu 2744 WARN_ON_ONCE(pwq->flush_color != -1);
3851 2745
3852 if (pwq->nr_in_flight 2746 if (pwq->nr_in_flight[flush_color]) {
3853 pwq->flush_co 2747 pwq->flush_color = flush_color;
3854 atomic_inc(&w 2748 atomic_inc(&wq->nr_pwqs_to_flush);
3855 wait = true; 2749 wait = true;
3856 } 2750 }
3857 } 2751 }
3858 2752
3859 if (work_color >= 0) { 2753 if (work_color >= 0) {
3860 WARN_ON_ONCE(work_col 2754 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
3861 pwq->work_color = wor 2755 pwq->work_color = work_color;
3862 } 2756 }
3863 2757
3864 raw_spin_unlock_irq(&pool->lo 2758 raw_spin_unlock_irq(&pool->lock);
3865 } 2759 }
3866 2760
3867 if (flush_color >= 0 && atomic_dec_an 2761 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
3868 complete(&wq->first_flusher-> 2762 complete(&wq->first_flusher->done);
3869 2763
3870 return wait; 2764 return wait;
3871 } 2765 }
3872 2766
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 /** 2767 /**
3903 * __flush_workqueue - ensure that any schedu !! 2768 * flush_workqueue - ensure that any scheduled work has run to completion.
3904 * @wq: workqueue to flush 2769 * @wq: workqueue to flush
3905 * 2770 *
3906 * This function sleeps until all work items 2771 * This function sleeps until all work items which were queued on entry
3907 * have finished execution, but it is not liv 2772 * have finished execution, but it is not livelocked by new incoming ones.
3908 */ 2773 */
3909 void __flush_workqueue(struct workqueue_struc !! 2774 void flush_workqueue(struct workqueue_struct *wq)
3910 { 2775 {
3911 struct wq_flusher this_flusher = { 2776 struct wq_flusher this_flusher = {
3912 .list = LIST_HEAD_INIT(this_f 2777 .list = LIST_HEAD_INIT(this_flusher.list),
3913 .flush_color = -1, 2778 .flush_color = -1,
3914 .done = COMPLETION_INITIALIZE 2779 .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map),
3915 }; 2780 };
3916 int next_color; 2781 int next_color;
3917 2782
3918 if (WARN_ON(!wq_online)) 2783 if (WARN_ON(!wq_online))
3919 return; 2784 return;
3920 2785
3921 touch_wq_lockdep_map(wq); !! 2786 lock_map_acquire(&wq->lockdep_map);
>> 2787 lock_map_release(&wq->lockdep_map);
3922 2788
3923 mutex_lock(&wq->mutex); 2789 mutex_lock(&wq->mutex);
3924 2790
3925 /* 2791 /*
3926 * Start-to-wait phase 2792 * Start-to-wait phase
3927 */ 2793 */
3928 next_color = work_next_color(wq->work 2794 next_color = work_next_color(wq->work_color);
3929 2795
3930 if (next_color != wq->flush_color) { 2796 if (next_color != wq->flush_color) {
3931 /* 2797 /*
3932 * Color space is not full. 2798 * Color space is not full. The current work_color
3933 * becomes our flush_color an 2799 * becomes our flush_color and work_color is advanced
3934 * by one. 2800 * by one.
3935 */ 2801 */
3936 WARN_ON_ONCE(!list_empty(&wq- 2802 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
3937 this_flusher.flush_color = wq 2803 this_flusher.flush_color = wq->work_color;
3938 wq->work_color = next_color; 2804 wq->work_color = next_color;
3939 2805
3940 if (!wq->first_flusher) { 2806 if (!wq->first_flusher) {
3941 /* no flush in progre 2807 /* no flush in progress, become the first flusher */
3942 WARN_ON_ONCE(wq->flus 2808 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
3943 2809
3944 wq->first_flusher = & 2810 wq->first_flusher = &this_flusher;
3945 2811
3946 if (!flush_workqueue_ 2812 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
3947 2813 wq->work_color)) {
3948 /* nothing to 2814 /* nothing to flush, done */
3949 wq->flush_col 2815 wq->flush_color = next_color;
3950 wq->first_flu 2816 wq->first_flusher = NULL;
3951 goto out_unlo 2817 goto out_unlock;
3952 } 2818 }
3953 } else { 2819 } else {
3954 /* wait in queue */ 2820 /* wait in queue */
3955 WARN_ON_ONCE(wq->flus 2821 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
3956 list_add_tail(&this_f 2822 list_add_tail(&this_flusher.list, &wq->flusher_queue);
3957 flush_workqueue_prep_ 2823 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
3958 } 2824 }
3959 } else { 2825 } else {
3960 /* 2826 /*
3961 * Oops, color space is full, 2827 * Oops, color space is full, wait on overflow queue.
3962 * The next flush completion 2828 * The next flush completion will assign us
3963 * flush_color and transfer t 2829 * flush_color and transfer to flusher_queue.
3964 */ 2830 */
3965 list_add_tail(&this_flusher.l 2831 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
3966 } 2832 }
3967 2833
3968 check_flush_dependency(wq, NULL); 2834 check_flush_dependency(wq, NULL);
3969 2835
3970 mutex_unlock(&wq->mutex); 2836 mutex_unlock(&wq->mutex);
3971 2837
3972 wait_for_completion(&this_flusher.don 2838 wait_for_completion(&this_flusher.done);
3973 2839
3974 /* 2840 /*
3975 * Wake-up-and-cascade phase 2841 * Wake-up-and-cascade phase
3976 * 2842 *
3977 * First flushers are responsible for 2843 * First flushers are responsible for cascading flushes and
3978 * handling overflow. Non-first flus 2844 * handling overflow. Non-first flushers can simply return.
3979 */ 2845 */
3980 if (READ_ONCE(wq->first_flusher) != & 2846 if (READ_ONCE(wq->first_flusher) != &this_flusher)
3981 return; 2847 return;
3982 2848
3983 mutex_lock(&wq->mutex); 2849 mutex_lock(&wq->mutex);
3984 2850
3985 /* we might have raced, check again w 2851 /* we might have raced, check again with mutex held */
3986 if (wq->first_flusher != &this_flushe 2852 if (wq->first_flusher != &this_flusher)
3987 goto out_unlock; 2853 goto out_unlock;
3988 2854
3989 WRITE_ONCE(wq->first_flusher, NULL); 2855 WRITE_ONCE(wq->first_flusher, NULL);
3990 2856
3991 WARN_ON_ONCE(!list_empty(&this_flushe 2857 WARN_ON_ONCE(!list_empty(&this_flusher.list));
3992 WARN_ON_ONCE(wq->flush_color != this_ 2858 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
3993 2859
3994 while (true) { 2860 while (true) {
3995 struct wq_flusher *next, *tmp 2861 struct wq_flusher *next, *tmp;
3996 2862
3997 /* complete all the flushers 2863 /* complete all the flushers sharing the current flush color */
3998 list_for_each_entry_safe(next 2864 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
3999 if (next->flush_color 2865 if (next->flush_color != wq->flush_color)
4000 break; 2866 break;
4001 list_del_init(&next-> 2867 list_del_init(&next->list);
4002 complete(&next->done) 2868 complete(&next->done);
4003 } 2869 }
4004 2870
4005 WARN_ON_ONCE(!list_empty(&wq- 2871 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
4006 wq->flush_color 2872 wq->flush_color != work_next_color(wq->work_color));
4007 2873
4008 /* this flush_color is finish 2874 /* this flush_color is finished, advance by one */
4009 wq->flush_color = work_next_c 2875 wq->flush_color = work_next_color(wq->flush_color);
4010 2876
4011 /* one color has been freed, 2877 /* one color has been freed, handle overflow queue */
4012 if (!list_empty(&wq->flusher_ 2878 if (!list_empty(&wq->flusher_overflow)) {
4013 /* 2879 /*
4014 * Assign the same co 2880 * Assign the same color to all overflowed
4015 * flushers, advance 2881 * flushers, advance work_color and append to
4016 * flusher_queue. Th 2882 * flusher_queue. This is the start-to-wait
4017 * phase for these ov 2883 * phase for these overflowed flushers.
4018 */ 2884 */
4019 list_for_each_entry(t 2885 list_for_each_entry(tmp, &wq->flusher_overflow, list)
4020 tmp->flush_co 2886 tmp->flush_color = wq->work_color;
4021 2887
4022 wq->work_color = work 2888 wq->work_color = work_next_color(wq->work_color);
4023 2889
4024 list_splice_tail_init 2890 list_splice_tail_init(&wq->flusher_overflow,
4025 2891 &wq->flusher_queue);
4026 flush_workqueue_prep_ 2892 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
4027 } 2893 }
4028 2894
4029 if (list_empty(&wq->flusher_q 2895 if (list_empty(&wq->flusher_queue)) {
4030 WARN_ON_ONCE(wq->flus 2896 WARN_ON_ONCE(wq->flush_color != wq->work_color);
4031 break; 2897 break;
4032 } 2898 }
4033 2899
4034 /* 2900 /*
4035 * Need to flush more colors. 2901 * Need to flush more colors. Make the next flusher
4036 * the new first flusher and 2902 * the new first flusher and arm pwqs.
4037 */ 2903 */
4038 WARN_ON_ONCE(wq->flush_color 2904 WARN_ON_ONCE(wq->flush_color == wq->work_color);
4039 WARN_ON_ONCE(wq->flush_color 2905 WARN_ON_ONCE(wq->flush_color != next->flush_color);
4040 2906
4041 list_del_init(&next->list); 2907 list_del_init(&next->list);
4042 wq->first_flusher = next; 2908 wq->first_flusher = next;
4043 2909
4044 if (flush_workqueue_prep_pwqs 2910 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
4045 break; 2911 break;
4046 2912
4047 /* 2913 /*
4048 * Meh... this color is alrea 2914 * Meh... this color is already done, clear first
4049 * flusher and repeat cascadi 2915 * flusher and repeat cascading.
4050 */ 2916 */
4051 wq->first_flusher = NULL; 2917 wq->first_flusher = NULL;
4052 } 2918 }
4053 2919
4054 out_unlock: 2920 out_unlock:
4055 mutex_unlock(&wq->mutex); 2921 mutex_unlock(&wq->mutex);
4056 } 2922 }
4057 EXPORT_SYMBOL(__flush_workqueue); !! 2923 EXPORT_SYMBOL(flush_workqueue);
4058 2924
4059 /** 2925 /**
4060 * drain_workqueue - drain a workqueue 2926 * drain_workqueue - drain a workqueue
4061 * @wq: workqueue to drain 2927 * @wq: workqueue to drain
4062 * 2928 *
4063 * Wait until the workqueue becomes empty. W 2929 * Wait until the workqueue becomes empty. While draining is in progress,
4064 * only chain queueing is allowed. IOW, only 2930 * only chain queueing is allowed. IOW, only currently pending or running
4065 * work items on @wq can queue further work i 2931 * work items on @wq can queue further work items on it. @wq is flushed
4066 * repeatedly until it becomes empty. The nu 2932 * repeatedly until it becomes empty. The number of flushing is determined
4067 * by the depth of chaining and should be rel 2933 * by the depth of chaining and should be relatively short. Whine if it
4068 * takes too long. 2934 * takes too long.
4069 */ 2935 */
4070 void drain_workqueue(struct workqueue_struct 2936 void drain_workqueue(struct workqueue_struct *wq)
4071 { 2937 {
4072 unsigned int flush_cnt = 0; 2938 unsigned int flush_cnt = 0;
4073 struct pool_workqueue *pwq; 2939 struct pool_workqueue *pwq;
4074 2940
4075 /* 2941 /*
4076 * __queue_work() needs to test wheth 2942 * __queue_work() needs to test whether there are drainers, is much
4077 * hotter than drain_workqueue() and 2943 * hotter than drain_workqueue() and already looks at @wq->flags.
4078 * Use __WQ_DRAINING so that queue do 2944 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
4079 */ 2945 */
4080 mutex_lock(&wq->mutex); 2946 mutex_lock(&wq->mutex);
4081 if (!wq->nr_drainers++) 2947 if (!wq->nr_drainers++)
4082 wq->flags |= __WQ_DRAINING; 2948 wq->flags |= __WQ_DRAINING;
4083 mutex_unlock(&wq->mutex); 2949 mutex_unlock(&wq->mutex);
4084 reflush: 2950 reflush:
4085 __flush_workqueue(wq); !! 2951 flush_workqueue(wq);
4086 2952
4087 mutex_lock(&wq->mutex); 2953 mutex_lock(&wq->mutex);
4088 2954
4089 for_each_pwq(pwq, wq) { 2955 for_each_pwq(pwq, wq) {
4090 bool drained; 2956 bool drained;
4091 2957
4092 raw_spin_lock_irq(&pwq->pool- 2958 raw_spin_lock_irq(&pwq->pool->lock);
4093 drained = pwq_is_empty(pwq); !! 2959 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
4094 raw_spin_unlock_irq(&pwq->poo 2960 raw_spin_unlock_irq(&pwq->pool->lock);
4095 2961
4096 if (drained) 2962 if (drained)
4097 continue; 2963 continue;
4098 2964
4099 if (++flush_cnt == 10 || 2965 if (++flush_cnt == 10 ||
4100 (flush_cnt % 100 == 0 && 2966 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
4101 pr_warn("workqueue %s !! 2967 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
4102 wq->name, __f !! 2968 wq->name, flush_cnt);
4103 2969
4104 mutex_unlock(&wq->mutex); 2970 mutex_unlock(&wq->mutex);
4105 goto reflush; 2971 goto reflush;
4106 } 2972 }
4107 2973
4108 if (!--wq->nr_drainers) 2974 if (!--wq->nr_drainers)
4109 wq->flags &= ~__WQ_DRAINING; 2975 wq->flags &= ~__WQ_DRAINING;
4110 mutex_unlock(&wq->mutex); 2976 mutex_unlock(&wq->mutex);
4111 } 2977 }
4112 EXPORT_SYMBOL_GPL(drain_workqueue); 2978 EXPORT_SYMBOL_GPL(drain_workqueue);
4113 2979
4114 static bool start_flush_work(struct work_stru 2980 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
4115 bool from_cancel 2981 bool from_cancel)
4116 { 2982 {
4117 struct worker *worker = NULL; 2983 struct worker *worker = NULL;
4118 struct worker_pool *pool; 2984 struct worker_pool *pool;
4119 struct pool_workqueue *pwq; 2985 struct pool_workqueue *pwq;
4120 struct workqueue_struct *wq; !! 2986
>> 2987 might_sleep();
4121 2988
4122 rcu_read_lock(); 2989 rcu_read_lock();
4123 pool = get_work_pool(work); 2990 pool = get_work_pool(work);
4124 if (!pool) { 2991 if (!pool) {
4125 rcu_read_unlock(); 2992 rcu_read_unlock();
4126 return false; 2993 return false;
4127 } 2994 }
4128 2995
4129 raw_spin_lock_irq(&pool->lock); 2996 raw_spin_lock_irq(&pool->lock);
4130 /* see the comment in try_to_grab_pen 2997 /* see the comment in try_to_grab_pending() with the same code */
4131 pwq = get_work_pwq(work); 2998 pwq = get_work_pwq(work);
4132 if (pwq) { 2999 if (pwq) {
4133 if (unlikely(pwq->pool != poo 3000 if (unlikely(pwq->pool != pool))
4134 goto already_gone; 3001 goto already_gone;
4135 } else { 3002 } else {
4136 worker = find_worker_executin 3003 worker = find_worker_executing_work(pool, work);
4137 if (!worker) 3004 if (!worker)
4138 goto already_gone; 3005 goto already_gone;
4139 pwq = worker->current_pwq; 3006 pwq = worker->current_pwq;
4140 } 3007 }
4141 3008
4142 wq = pwq->wq; !! 3009 check_flush_dependency(pwq->wq, work);
4143 check_flush_dependency(wq, work); <<
4144 3010
4145 insert_wq_barrier(pwq, barr, work, wo 3011 insert_wq_barrier(pwq, barr, work, worker);
4146 raw_spin_unlock_irq(&pool->lock); 3012 raw_spin_unlock_irq(&pool->lock);
4147 3013
4148 touch_work_lockdep_map(work, wq); <<
4149 <<
4150 /* 3014 /*
4151 * Force a lock recursion deadlock wh 3015 * Force a lock recursion deadlock when using flush_work() inside a
4152 * single-threaded or rescuer equippe 3016 * single-threaded or rescuer equipped workqueue.
4153 * 3017 *
4154 * For single threaded workqueues the 3018 * For single threaded workqueues the deadlock happens when the work
4155 * is after the work issuing the flus 3019 * is after the work issuing the flush_work(). For rescuer equipped
4156 * workqueues the deadlock happens wh 3020 * workqueues the deadlock happens when the rescuer stalls, blocking
4157 * forward progress. 3021 * forward progress.
4158 */ 3022 */
4159 if (!from_cancel && (wq->saved_max_ac !! 3023 if (!from_cancel &&
4160 touch_wq_lockdep_map(wq); !! 3024 (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) {
4161 !! 3025 lock_map_acquire(&pwq->wq->lockdep_map);
>> 3026 lock_map_release(&pwq->wq->lockdep_map);
>> 3027 }
4162 rcu_read_unlock(); 3028 rcu_read_unlock();
4163 return true; 3029 return true;
4164 already_gone: 3030 already_gone:
4165 raw_spin_unlock_irq(&pool->lock); 3031 raw_spin_unlock_irq(&pool->lock);
4166 rcu_read_unlock(); 3032 rcu_read_unlock();
4167 return false; 3033 return false;
4168 } 3034 }
4169 3035
4170 static bool __flush_work(struct work_struct * 3036 static bool __flush_work(struct work_struct *work, bool from_cancel)
4171 { 3037 {
4172 struct wq_barrier barr; 3038 struct wq_barrier barr;
4173 3039
4174 if (WARN_ON(!wq_online)) 3040 if (WARN_ON(!wq_online))
4175 return false; 3041 return false;
4176 3042
4177 if (WARN_ON(!work->func)) 3043 if (WARN_ON(!work->func))
4178 return false; 3044 return false;
4179 3045
4180 if (!start_flush_work(work, &barr, fr !! 3046 if (!from_cancel) {
4181 return false; !! 3047 lock_map_acquire(&work->lockdep_map);
4182 !! 3048 lock_map_release(&work->lockdep_map);
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 <<
4193 if (!WARN_ON_ONCE(data & WORK <<
4194 (data & WORK_OFFQ_BH)) { <<
4195 /* <<
4196 * On RT, prevent a l <<
4197 * soft interrupt pro <<
4198 * running by keeping <<
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 } 3049 }
4215 3050
4216 wait_for_completion(&barr.done); !! 3051 if (start_flush_work(work, &barr, from_cancel)) {
4217 !! 3052 wait_for_completion(&barr.done);
4218 out_destroy: !! 3053 destroy_work_on_stack(&barr.work);
4219 destroy_work_on_stack(&barr.work); !! 3054 return true;
4220 return true; !! 3055 } else {
>> 3056 return false;
>> 3057 }
4221 } 3058 }
4222 3059
4223 /** 3060 /**
4224 * flush_work - wait for a work to finish exe 3061 * flush_work - wait for a work to finish executing the last queueing instance
4225 * @work: the work to flush 3062 * @work: the work to flush
4226 * 3063 *
4227 * Wait until @work has finished execution. 3064 * Wait until @work has finished execution. @work is guaranteed to be idle
4228 * on return if it hasn't been requeued since 3065 * on return if it hasn't been requeued since flush started.
4229 * 3066 *
4230 * Return: 3067 * Return:
4231 * %true if flush_work() waited for the work 3068 * %true if flush_work() waited for the work to finish execution,
4232 * %false if it was already idle. 3069 * %false if it was already idle.
4233 */ 3070 */
4234 bool flush_work(struct work_struct *work) 3071 bool flush_work(struct work_struct *work)
4235 { 3072 {
4236 might_sleep(); <<
4237 return __flush_work(work, false); 3073 return __flush_work(work, false);
4238 } 3074 }
4239 EXPORT_SYMBOL_GPL(flush_work); 3075 EXPORT_SYMBOL_GPL(flush_work);
4240 3076
>> 3077 struct cwt_wait {
>> 3078 wait_queue_entry_t wait;
>> 3079 struct work_struct *work;
>> 3080 };
>> 3081
>> 3082 static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
>> 3083 {
>> 3084 struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
>> 3085
>> 3086 if (cwait->work != key)
>> 3087 return 0;
>> 3088 return autoremove_wake_function(wait, mode, sync, key);
>> 3089 }
>> 3090
>> 3091 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
>> 3092 {
>> 3093 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
>> 3094 unsigned long flags;
>> 3095 int ret;
>> 3096
>> 3097 do {
>> 3098 ret = try_to_grab_pending(work, is_dwork, &flags);
>> 3099 /*
>> 3100 * If someone else is already canceling, wait for it to
>> 3101 * finish. flush_work() doesn't work for PREEMPT_NONE
>> 3102 * because we may get scheduled between @work's completion
>> 3103 * and the other canceling task resuming and clearing
>> 3104 * CANCELING - flush_work() will return false immediately
>> 3105 * as @work is no longer busy, try_to_grab_pending() will
>> 3106 * return -ENOENT as @work is still being canceled and the
>> 3107 * other canceling task won't be able to clear CANCELING as
>> 3108 * we're hogging the CPU.
>> 3109 *
>> 3110 * Let's wait for completion using a waitqueue. As this
>> 3111 * may lead to the thundering herd problem, use a custom
>> 3112 * wake function which matches @work along with exclusive
>> 3113 * wait and wakeup.
>> 3114 */
>> 3115 if (unlikely(ret == -ENOENT)) {
>> 3116 struct cwt_wait cwait;
>> 3117
>> 3118 init_wait(&cwait.wait);
>> 3119 cwait.wait.func = cwt_wakefn;
>> 3120 cwait.work = work;
>> 3121
>> 3122 prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
>> 3123 TASK_UNINTERRUPTIBLE);
>> 3124 if (work_is_canceling(work))
>> 3125 schedule();
>> 3126 finish_wait(&cancel_waitq, &cwait.wait);
>> 3127 }
>> 3128 } while (unlikely(ret < 0));
>> 3129
>> 3130 /* tell other tasks trying to grab @work to back off */
>> 3131 mark_work_canceling(work);
>> 3132 local_irq_restore(flags);
>> 3133
>> 3134 /*
>> 3135 * This allows canceling during early boot. We know that @work
>> 3136 * isn't executing.
>> 3137 */
>> 3138 if (wq_online)
>> 3139 __flush_work(work, true);
>> 3140
>> 3141 clear_work_data(work);
>> 3142
>> 3143 /*
>> 3144 * Paired with prepare_to_wait() above so that either
>> 3145 * waitqueue_active() is visible here or !work_is_canceling() is
>> 3146 * visible there.
>> 3147 */
>> 3148 smp_mb();
>> 3149 if (waitqueue_active(&cancel_waitq))
>> 3150 __wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
>> 3151
>> 3152 return ret;
>> 3153 }
>> 3154
>> 3155 /**
>> 3156 * cancel_work_sync - cancel a work and wait for it to finish
>> 3157 * @work: the work to cancel
>> 3158 *
>> 3159 * Cancel @work and wait for its execution to finish. This function
>> 3160 * can be used even if the work re-queues itself or migrates to
>> 3161 * another workqueue. On return from this function, @work is
>> 3162 * guaranteed to be not pending or executing on any CPU.
>> 3163 *
>> 3164 * cancel_work_sync(&delayed_work->work) must not be used for
>> 3165 * delayed_work's. Use cancel_delayed_work_sync() instead.
>> 3166 *
>> 3167 * The caller must ensure that the workqueue on which @work was last
>> 3168 * queued can't be destroyed before this function returns.
>> 3169 *
>> 3170 * Return:
>> 3171 * %true if @work was pending, %false otherwise.
>> 3172 */
>> 3173 bool cancel_work_sync(struct work_struct *work)
>> 3174 {
>> 3175 return __cancel_work_timer(work, false);
>> 3176 }
>> 3177 EXPORT_SYMBOL_GPL(cancel_work_sync);
>> 3178
4241 /** 3179 /**
4242 * flush_delayed_work - wait for a dwork to f 3180 * flush_delayed_work - wait for a dwork to finish executing the last queueing
4243 * @dwork: the delayed work to flush 3181 * @dwork: the delayed work to flush
4244 * 3182 *
4245 * Delayed timer is cancelled and the pending 3183 * Delayed timer is cancelled and the pending work is queued for
4246 * immediate execution. Like flush_work(), t 3184 * immediate execution. Like flush_work(), this function only
4247 * considers the last queueing instance of @d 3185 * considers the last queueing instance of @dwork.
4248 * 3186 *
4249 * Return: 3187 * Return:
4250 * %true if flush_work() waited for the work 3188 * %true if flush_work() waited for the work to finish execution,
4251 * %false if it was already idle. 3189 * %false if it was already idle.
4252 */ 3190 */
4253 bool flush_delayed_work(struct delayed_work * 3191 bool flush_delayed_work(struct delayed_work *dwork)
4254 { 3192 {
4255 local_irq_disable(); 3193 local_irq_disable();
4256 if (del_timer_sync(&dwork->timer)) 3194 if (del_timer_sync(&dwork->timer))
4257 __queue_work(dwork->cpu, dwor 3195 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
4258 local_irq_enable(); 3196 local_irq_enable();
4259 return flush_work(&dwork->work); 3197 return flush_work(&dwork->work);
4260 } 3198 }
4261 EXPORT_SYMBOL(flush_delayed_work); 3199 EXPORT_SYMBOL(flush_delayed_work);
4262 3200
4263 /** 3201 /**
4264 * flush_rcu_work - wait for a rwork to finis 3202 * flush_rcu_work - wait for a rwork to finish executing the last queueing
4265 * @rwork: the rcu work to flush 3203 * @rwork: the rcu work to flush
4266 * 3204 *
4267 * Return: 3205 * Return:
4268 * %true if flush_rcu_work() waited for the w 3206 * %true if flush_rcu_work() waited for the work to finish execution,
4269 * %false if it was already idle. 3207 * %false if it was already idle.
4270 */ 3208 */
4271 bool flush_rcu_work(struct rcu_work *rwork) 3209 bool flush_rcu_work(struct rcu_work *rwork)
4272 { 3210 {
4273 if (test_bit(WORK_STRUCT_PENDING_BIT, 3211 if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) {
4274 rcu_barrier(); 3212 rcu_barrier();
4275 flush_work(&rwork->work); 3213 flush_work(&rwork->work);
4276 return true; 3214 return true;
4277 } else { 3215 } else {
4278 return flush_work(&rwork->wor 3216 return flush_work(&rwork->work);
4279 } 3217 }
4280 } 3218 }
4281 EXPORT_SYMBOL(flush_rcu_work); 3219 EXPORT_SYMBOL(flush_rcu_work);
4282 3220
4283 static void work_offqd_disable(struct work_of !! 3221 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 { 3222 {
4295 if (likely(offqd->disable > 0)) !! 3223 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; 3224 int ret;
4306 3225
4307 ret = work_grab_pending(work, cflags, !! 3226 do {
4308 !! 3227 ret = try_to_grab_pending(work, is_dwork, &flags);
4309 work_offqd_unpack(&offqd, *work_data_ !! 3228 } 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 3229
4338 if (!(cflags & WORK_CANCEL_DISABLE)) !! 3230 if (unlikely(ret < 0))
4339 enable_work(work); !! 3231 return false;
4340 3232
>> 3233 set_work_pool_and_clear_pending(work, get_work_pool_id(work));
>> 3234 local_irq_restore(flags);
4341 return ret; 3235 return ret;
4342 } 3236 }
4343 3237
4344 /* <<
4345 * See cancel_delayed_work() <<
4346 */ <<
4347 bool cancel_work(struct work_struct *work) <<
4348 { <<
4349 return __cancel_work(work, 0); <<
4350 } <<
4351 EXPORT_SYMBOL(cancel_work); <<
4352 <<
4353 /** <<
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 /** 3238 /**
4378 * cancel_delayed_work - cancel a delayed wor 3239 * cancel_delayed_work - cancel a delayed work
4379 * @dwork: delayed_work to cancel 3240 * @dwork: delayed_work to cancel
4380 * 3241 *
4381 * Kill off a pending delayed_work. 3242 * Kill off a pending delayed_work.
4382 * 3243 *
4383 * Return: %true if @dwork was pending and ca 3244 * Return: %true if @dwork was pending and canceled; %false if it wasn't
4384 * pending. 3245 * pending.
4385 * 3246 *
4386 * Note: 3247 * Note:
4387 * The work callback function may still be ru 3248 * The work callback function may still be running on return, unless
4388 * it returns %true and the work doesn't re-a 3249 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
4389 * use cancel_delayed_work_sync() to wait on 3250 * use cancel_delayed_work_sync() to wait on it.
4390 * 3251 *
4391 * This function is safe to call from any con 3252 * This function is safe to call from any context including IRQ handler.
4392 */ 3253 */
4393 bool cancel_delayed_work(struct delayed_work 3254 bool cancel_delayed_work(struct delayed_work *dwork)
4394 { 3255 {
4395 return __cancel_work(&dwork->work, WO !! 3256 return __cancel_work(&dwork->work, true);
4396 } 3257 }
4397 EXPORT_SYMBOL(cancel_delayed_work); 3258 EXPORT_SYMBOL(cancel_delayed_work);
4398 3259
4399 /** 3260 /**
4400 * cancel_delayed_work_sync - cancel a delaye 3261 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
4401 * @dwork: the delayed work cancel 3262 * @dwork: the delayed work cancel
4402 * 3263 *
4403 * This is cancel_work_sync() for delayed wor 3264 * This is cancel_work_sync() for delayed works.
4404 * 3265 *
4405 * Return: 3266 * Return:
4406 * %true if @dwork was pending, %false otherw 3267 * %true if @dwork was pending, %false otherwise.
4407 */ 3268 */
4408 bool cancel_delayed_work_sync(struct delayed_ 3269 bool cancel_delayed_work_sync(struct delayed_work *dwork)
4409 { 3270 {
4410 return __cancel_work_sync(&dwork->wor !! 3271 return __cancel_work_timer(&dwork->work, true);
4411 } 3272 }
4412 EXPORT_SYMBOL(cancel_delayed_work_sync); 3273 EXPORT_SYMBOL(cancel_delayed_work_sync);
4413 3274
4414 /** 3275 /**
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 3276 * schedule_on_each_cpu - execute a function synchronously on each online CPU
4518 * @func: the function to call 3277 * @func: the function to call
4519 * 3278 *
4520 * schedule_on_each_cpu() executes @func on e 3279 * schedule_on_each_cpu() executes @func on each online CPU using the
4521 * system workqueue and blocks until all CPUs 3280 * system workqueue and blocks until all CPUs have completed.
4522 * schedule_on_each_cpu() is very slow. 3281 * schedule_on_each_cpu() is very slow.
4523 * 3282 *
4524 * Return: 3283 * Return:
4525 * 0 on success, -errno on failure. 3284 * 0 on success, -errno on failure.
4526 */ 3285 */
4527 int schedule_on_each_cpu(work_func_t func) 3286 int schedule_on_each_cpu(work_func_t func)
4528 { 3287 {
4529 int cpu; 3288 int cpu;
4530 struct work_struct __percpu *works; 3289 struct work_struct __percpu *works;
4531 3290
4532 works = alloc_percpu(struct work_stru 3291 works = alloc_percpu(struct work_struct);
4533 if (!works) 3292 if (!works)
4534 return -ENOMEM; 3293 return -ENOMEM;
4535 3294
4536 cpus_read_lock(); !! 3295 get_online_cpus();
4537 3296
4538 for_each_online_cpu(cpu) { 3297 for_each_online_cpu(cpu) {
4539 struct work_struct *work = pe 3298 struct work_struct *work = per_cpu_ptr(works, cpu);
4540 3299
4541 INIT_WORK(work, func); 3300 INIT_WORK(work, func);
4542 schedule_work_on(cpu, work); 3301 schedule_work_on(cpu, work);
4543 } 3302 }
4544 3303
4545 for_each_online_cpu(cpu) 3304 for_each_online_cpu(cpu)
4546 flush_work(per_cpu_ptr(works, 3305 flush_work(per_cpu_ptr(works, cpu));
4547 3306
4548 cpus_read_unlock(); !! 3307 put_online_cpus();
4549 free_percpu(works); 3308 free_percpu(works);
4550 return 0; 3309 return 0;
4551 } 3310 }
4552 3311
4553 /** 3312 /**
4554 * execute_in_process_context - reliably exec 3313 * execute_in_process_context - reliably execute the routine with user context
4555 * @fn: the function to execute 3314 * @fn: the function to execute
4556 * @ew: guaranteed storage for the ex 3315 * @ew: guaranteed storage for the execute work structure (must
4557 * be available when the work ex 3316 * be available when the work executes)
4558 * 3317 *
4559 * Executes the function immediately if proce 3318 * Executes the function immediately if process context is available,
4560 * otherwise schedules the function for delay 3319 * otherwise schedules the function for delayed execution.
4561 * 3320 *
4562 * Return: 0 - function was executed 3321 * Return: 0 - function was executed
4563 * 1 - function was scheduled fo 3322 * 1 - function was scheduled for execution
4564 */ 3323 */
4565 int execute_in_process_context(work_func_t fn 3324 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
4566 { 3325 {
4567 if (!in_interrupt()) { 3326 if (!in_interrupt()) {
4568 fn(&ew->work); 3327 fn(&ew->work);
4569 return 0; 3328 return 0;
4570 } 3329 }
4571 3330
4572 INIT_WORK(&ew->work, fn); 3331 INIT_WORK(&ew->work, fn);
4573 schedule_work(&ew->work); 3332 schedule_work(&ew->work);
4574 3333
4575 return 1; 3334 return 1;
4576 } 3335 }
4577 EXPORT_SYMBOL_GPL(execute_in_process_context) 3336 EXPORT_SYMBOL_GPL(execute_in_process_context);
4578 3337
4579 /** 3338 /**
4580 * free_workqueue_attrs - free a workqueue_at 3339 * free_workqueue_attrs - free a workqueue_attrs
4581 * @attrs: workqueue_attrs to free 3340 * @attrs: workqueue_attrs to free
4582 * 3341 *
4583 * Undo alloc_workqueue_attrs(). 3342 * Undo alloc_workqueue_attrs().
4584 */ 3343 */
4585 void free_workqueue_attrs(struct workqueue_at 3344 void free_workqueue_attrs(struct workqueue_attrs *attrs)
4586 { 3345 {
4587 if (attrs) { 3346 if (attrs) {
4588 free_cpumask_var(attrs->cpuma 3347 free_cpumask_var(attrs->cpumask);
4589 free_cpumask_var(attrs->__pod <<
4590 kfree(attrs); 3348 kfree(attrs);
4591 } 3349 }
4592 } 3350 }
4593 3351
4594 /** 3352 /**
4595 * alloc_workqueue_attrs - allocate a workque 3353 * alloc_workqueue_attrs - allocate a workqueue_attrs
4596 * 3354 *
4597 * Allocate a new workqueue_attrs, initialize 3355 * Allocate a new workqueue_attrs, initialize with default settings and
4598 * return it. 3356 * return it.
4599 * 3357 *
4600 * Return: The allocated new workqueue_attr o 3358 * Return: The allocated new workqueue_attr on success. %NULL on failure.
4601 */ 3359 */
4602 struct workqueue_attrs *alloc_workqueue_attrs 3360 struct workqueue_attrs *alloc_workqueue_attrs(void)
4603 { 3361 {
4604 struct workqueue_attrs *attrs; 3362 struct workqueue_attrs *attrs;
4605 3363
4606 attrs = kzalloc(sizeof(*attrs), GFP_K 3364 attrs = kzalloc(sizeof(*attrs), GFP_KERNEL);
4607 if (!attrs) 3365 if (!attrs)
4608 goto fail; 3366 goto fail;
4609 if (!alloc_cpumask_var(&attrs->cpumas 3367 if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL))
4610 goto fail; 3368 goto fail;
4611 if (!alloc_cpumask_var(&attrs->__pod_ <<
4612 goto fail; <<
4613 3369
4614 cpumask_copy(attrs->cpumask, cpu_poss 3370 cpumask_copy(attrs->cpumask, cpu_possible_mask);
4615 attrs->affn_scope = WQ_AFFN_DFL; <<
4616 return attrs; 3371 return attrs;
4617 fail: 3372 fail:
4618 free_workqueue_attrs(attrs); 3373 free_workqueue_attrs(attrs);
4619 return NULL; 3374 return NULL;
4620 } 3375 }
4621 3376
4622 static void copy_workqueue_attrs(struct workq 3377 static void copy_workqueue_attrs(struct workqueue_attrs *to,
4623 const struct 3378 const struct workqueue_attrs *from)
4624 { 3379 {
4625 to->nice = from->nice; 3380 to->nice = from->nice;
4626 cpumask_copy(to->cpumask, from->cpuma 3381 cpumask_copy(to->cpumask, from->cpumask);
4627 cpumask_copy(to->__pod_cpumask, from- <<
4628 to->affn_strict = from->affn_strict; <<
4629 <<
4630 /* 3382 /*
4631 * Unlike hash and equality test, cop !! 3383 * Unlike hash and equality test, this function doesn't ignore
4632 * fields as copying is used for both !! 3384 * ->no_numa as it is used for both pool and wq attrs. Instead,
4633 * get_unbound_pool() explicitly clea !! 3385 * get_unbound_pool() explicitly clears ->no_numa after copying.
4634 */ 3386 */
4635 to->affn_scope = from->affn_scope; !! 3387 to->no_numa = from->no_numa;
4636 to->ordered = from->ordered; <<
4637 } <<
4638 <<
4639 /* <<
4640 * Some attrs fields are workqueue-only. Clea <<
4641 * comments in 'struct workqueue_attrs' defin <<
4642 */ <<
4643 static void wqattrs_clear_for_pool(struct wor <<
4644 { <<
4645 attrs->affn_scope = WQ_AFFN_NR_TYPES; <<
4646 attrs->ordered = false; <<
4647 if (attrs->affn_strict) <<
4648 cpumask_copy(attrs->cpumask, <<
4649 } 3388 }
4650 3389
4651 /* hash value of the content of @attr */ 3390 /* hash value of the content of @attr */
4652 static u32 wqattrs_hash(const struct workqueu 3391 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
4653 { 3392 {
4654 u32 hash = 0; 3393 u32 hash = 0;
4655 3394
4656 hash = jhash_1word(attrs->nice, hash) 3395 hash = jhash_1word(attrs->nice, hash);
4657 hash = jhash_1word(attrs->affn_strict !! 3396 hash = jhash(cpumask_bits(attrs->cpumask),
4658 hash = jhash(cpumask_bits(attrs->__po <<
4659 BITS_TO_LONGS(nr_cpumask 3397 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
4660 if (!attrs->affn_strict) <<
4661 hash = jhash(cpumask_bits(att <<
4662 BITS_TO_LONGS(nr <<
4663 return hash; 3398 return hash;
4664 } 3399 }
4665 3400
4666 /* content equality test */ 3401 /* content equality test */
4667 static bool wqattrs_equal(const struct workqu 3402 static bool wqattrs_equal(const struct workqueue_attrs *a,
4668 const struct workqu 3403 const struct workqueue_attrs *b)
4669 { 3404 {
4670 if (a->nice != b->nice) 3405 if (a->nice != b->nice)
4671 return false; 3406 return false;
4672 if (a->affn_strict != b->affn_strict) !! 3407 if (!cpumask_equal(a->cpumask, b->cpumask))
4673 return false; <<
4674 if (!cpumask_equal(a->__pod_cpumask, <<
4675 return false; <<
4676 if (!a->affn_strict && !cpumask_equal <<
4677 return false; 3408 return false;
4678 return true; 3409 return true;
4679 } 3410 }
4680 3411
4681 /* Update @attrs with actually available CPUs <<
4682 static void wqattrs_actualize_cpumask(struct <<
4683 const c <<
4684 { <<
4685 /* <<
4686 * Calculate the effective CPU mask o <<
4687 * @attrs->cpumask doesn't overlap wi <<
4688 * @unbound_cpumask. <<
4689 */ <<
4690 cpumask_and(attrs->cpumask, attrs->cp <<
4691 if (unlikely(cpumask_empty(attrs->cpu <<
4692 cpumask_copy(attrs->cpumask, <<
4693 } <<
4694 <<
4695 /* find wq_pod_type to use for @attrs */ <<
4696 static const struct wq_pod_type * <<
4697 wqattrs_pod_type(const struct workqueue_attrs <<
4698 { <<
4699 enum wq_affn_scope scope; <<
4700 struct wq_pod_type *pt; <<
4701 <<
4702 /* to synchronize access to wq_affn_d <<
4703 lockdep_assert_held(&wq_pool_mutex); <<
4704 <<
4705 if (attrs->affn_scope == WQ_AFFN_DFL) <<
4706 scope = wq_affn_dfl; <<
4707 else <<
4708 scope = attrs->affn_scope; <<
4709 <<
4710 pt = &wq_pod_types[scope]; <<
4711 <<
4712 if (!WARN_ON_ONCE(attrs->affn_scope = <<
4713 likely(pt->nr_pods)) <<
4714 return pt; <<
4715 <<
4716 /* <<
4717 * Before workqueue_init_topology(), <<
4718 * initialized in workqueue_init_earl <<
4719 */ <<
4720 pt = &wq_pod_types[WQ_AFFN_SYSTEM]; <<
4721 BUG_ON(!pt->nr_pods); <<
4722 return pt; <<
4723 } <<
4724 <<
4725 /** 3412 /**
4726 * init_worker_pool - initialize a newly zall 3413 * init_worker_pool - initialize a newly zalloc'd worker_pool
4727 * @pool: worker_pool to initialize 3414 * @pool: worker_pool to initialize
4728 * 3415 *
4729 * Initialize a newly zalloc'd @pool. It als 3416 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
4730 * 3417 *
4731 * Return: 0 on success, -errno on failure. 3418 * Return: 0 on success, -errno on failure. Even on failure, all fields
4732 * inside @pool proper are initialized and pu 3419 * inside @pool proper are initialized and put_unbound_pool() can be called
4733 * on @pool safely to release it. 3420 * on @pool safely to release it.
4734 */ 3421 */
4735 static int init_worker_pool(struct worker_poo 3422 static int init_worker_pool(struct worker_pool *pool)
4736 { 3423 {
4737 raw_spin_lock_init(&pool->lock); 3424 raw_spin_lock_init(&pool->lock);
4738 pool->id = -1; 3425 pool->id = -1;
4739 pool->cpu = -1; 3426 pool->cpu = -1;
4740 pool->node = NUMA_NO_NODE; 3427 pool->node = NUMA_NO_NODE;
4741 pool->flags |= POOL_DISASSOCIATED; 3428 pool->flags |= POOL_DISASSOCIATED;
4742 pool->watchdog_ts = jiffies; 3429 pool->watchdog_ts = jiffies;
4743 INIT_LIST_HEAD(&pool->worklist); 3430 INIT_LIST_HEAD(&pool->worklist);
4744 INIT_LIST_HEAD(&pool->idle_list); 3431 INIT_LIST_HEAD(&pool->idle_list);
4745 hash_init(pool->busy_hash); 3432 hash_init(pool->busy_hash);
4746 3433
4747 timer_setup(&pool->idle_timer, idle_w 3434 timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE);
4748 INIT_WORK(&pool->idle_cull_work, idle <<
4749 3435
4750 timer_setup(&pool->mayday_timer, pool 3436 timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0);
4751 3437
4752 INIT_LIST_HEAD(&pool->workers); 3438 INIT_LIST_HEAD(&pool->workers);
4753 3439
4754 ida_init(&pool->worker_ida); 3440 ida_init(&pool->worker_ida);
4755 INIT_HLIST_NODE(&pool->hash_node); 3441 INIT_HLIST_NODE(&pool->hash_node);
4756 pool->refcnt = 1; 3442 pool->refcnt = 1;
4757 3443
4758 /* shouldn't fail above this point */ 3444 /* shouldn't fail above this point */
4759 pool->attrs = alloc_workqueue_attrs() 3445 pool->attrs = alloc_workqueue_attrs();
4760 if (!pool->attrs) 3446 if (!pool->attrs)
4761 return -ENOMEM; 3447 return -ENOMEM;
4762 <<
4763 wqattrs_clear_for_pool(pool->attrs); <<
4764 <<
4765 return 0; 3448 return 0;
4766 } 3449 }
4767 3450
4768 #ifdef CONFIG_LOCKDEP 3451 #ifdef CONFIG_LOCKDEP
4769 static void wq_init_lockdep(struct workqueue_ 3452 static void wq_init_lockdep(struct workqueue_struct *wq)
4770 { 3453 {
4771 char *lock_name; 3454 char *lock_name;
4772 3455
4773 lockdep_register_key(&wq->key); 3456 lockdep_register_key(&wq->key);
4774 lock_name = kasprintf(GFP_KERNEL, "%s 3457 lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name);
4775 if (!lock_name) 3458 if (!lock_name)
4776 lock_name = wq->name; 3459 lock_name = wq->name;
4777 3460
4778 wq->lock_name = lock_name; 3461 wq->lock_name = lock_name;
4779 lockdep_init_map(&wq->lockdep_map, lo 3462 lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0);
4780 } 3463 }
4781 3464
4782 static void wq_unregister_lockdep(struct work 3465 static void wq_unregister_lockdep(struct workqueue_struct *wq)
4783 { 3466 {
4784 lockdep_unregister_key(&wq->key); 3467 lockdep_unregister_key(&wq->key);
4785 } 3468 }
4786 3469
4787 static void wq_free_lockdep(struct workqueue_ 3470 static void wq_free_lockdep(struct workqueue_struct *wq)
4788 { 3471 {
4789 if (wq->lock_name != wq->name) 3472 if (wq->lock_name != wq->name)
4790 kfree(wq->lock_name); 3473 kfree(wq->lock_name);
4791 } 3474 }
4792 #else 3475 #else
4793 static void wq_init_lockdep(struct workqueue_ 3476 static void wq_init_lockdep(struct workqueue_struct *wq)
4794 { 3477 {
4795 } 3478 }
4796 3479
4797 static void wq_unregister_lockdep(struct work 3480 static void wq_unregister_lockdep(struct workqueue_struct *wq)
4798 { 3481 {
4799 } 3482 }
4800 3483
4801 static void wq_free_lockdep(struct workqueue_ 3484 static void wq_free_lockdep(struct workqueue_struct *wq)
4802 { 3485 {
4803 } 3486 }
4804 #endif 3487 #endif
4805 3488
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) 3489 static void rcu_free_wq(struct rcu_head *rcu)
4859 { 3490 {
4860 struct workqueue_struct *wq = 3491 struct workqueue_struct *wq =
4861 container_of(rcu, struct work 3492 container_of(rcu, struct workqueue_struct, rcu);
4862 3493
4863 if (wq->flags & WQ_UNBOUND) <<
4864 free_node_nr_active(wq->node_ <<
4865 <<
4866 wq_free_lockdep(wq); 3494 wq_free_lockdep(wq);
4867 free_percpu(wq->cpu_pwq); !! 3495
4868 free_workqueue_attrs(wq->unbound_attr !! 3496 if (!(wq->flags & WQ_UNBOUND))
>> 3497 free_percpu(wq->cpu_pwqs);
>> 3498 else
>> 3499 free_workqueue_attrs(wq->unbound_attrs);
>> 3500
4869 kfree(wq); 3501 kfree(wq);
4870 } 3502 }
4871 3503
4872 static void rcu_free_pool(struct rcu_head *rc 3504 static void rcu_free_pool(struct rcu_head *rcu)
4873 { 3505 {
4874 struct worker_pool *pool = container_ 3506 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
4875 3507
4876 ida_destroy(&pool->worker_ida); 3508 ida_destroy(&pool->worker_ida);
4877 free_workqueue_attrs(pool->attrs); 3509 free_workqueue_attrs(pool->attrs);
4878 kfree(pool); 3510 kfree(pool);
4879 } 3511 }
4880 3512
>> 3513 /* This returns with the lock held on success (pool manager is inactive). */
>> 3514 static bool wq_manager_inactive(struct worker_pool *pool)
>> 3515 {
>> 3516 raw_spin_lock_irq(&pool->lock);
>> 3517
>> 3518 if (pool->flags & POOL_MANAGER_ACTIVE) {
>> 3519 raw_spin_unlock_irq(&pool->lock);
>> 3520 return false;
>> 3521 }
>> 3522 return true;
>> 3523 }
>> 3524
4881 /** 3525 /**
4882 * put_unbound_pool - put a worker_pool 3526 * put_unbound_pool - put a worker_pool
4883 * @pool: worker_pool to put 3527 * @pool: worker_pool to put
4884 * 3528 *
4885 * Put @pool. If its refcnt reaches zero, it 3529 * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU
4886 * safe manner. get_unbound_pool() calls thi 3530 * safe manner. get_unbound_pool() calls this function on its failure path
4887 * and this function should be able to releas 3531 * and this function should be able to release pools which went through,
4888 * successfully or not, init_worker_pool(). 3532 * successfully or not, init_worker_pool().
4889 * 3533 *
4890 * Should be called with wq_pool_mutex held. 3534 * Should be called with wq_pool_mutex held.
4891 */ 3535 */
4892 static void put_unbound_pool(struct worker_po 3536 static void put_unbound_pool(struct worker_pool *pool)
4893 { 3537 {
>> 3538 DECLARE_COMPLETION_ONSTACK(detach_completion);
4894 struct worker *worker; 3539 struct worker *worker;
4895 LIST_HEAD(cull_list); <<
4896 3540
4897 lockdep_assert_held(&wq_pool_mutex); 3541 lockdep_assert_held(&wq_pool_mutex);
4898 3542
4899 if (--pool->refcnt) 3543 if (--pool->refcnt)
4900 return; 3544 return;
4901 3545
4902 /* sanity checks */ 3546 /* sanity checks */
4903 if (WARN_ON(!(pool->cpu < 0)) || 3547 if (WARN_ON(!(pool->cpu < 0)) ||
4904 WARN_ON(!list_empty(&pool->workli 3548 WARN_ON(!list_empty(&pool->worklist)))
4905 return; 3549 return;
4906 3550
4907 /* release id and unhash */ 3551 /* release id and unhash */
4908 if (pool->id >= 0) 3552 if (pool->id >= 0)
4909 idr_remove(&worker_pool_idr, 3553 idr_remove(&worker_pool_idr, pool->id);
4910 hash_del(&pool->hash_node); 3554 hash_del(&pool->hash_node);
4911 3555
4912 /* 3556 /*
4913 * Become the manager and destroy all 3557 * Become the manager and destroy all workers. This prevents
4914 * @pool's workers from blocking on a 3558 * @pool's workers from blocking on attach_mutex. We're the last
4915 * manager and @pool gets freed with 3559 * manager and @pool gets freed with the flag set.
4916 * !! 3560 * Because of how wq_manager_inactive() works, we will hold the
4917 * Having a concurrent manager is qui !! 3561 * spinlock after a successful wait.
4918 * only get here with <<
4919 * pwq->refcnt == pool->refcnt == 0 <<
4920 * which implies no work queued to th <<
4921 * become the manager. However a work <<
4922 * manager before the refcnts dropped <<
4923 * drops pool->lock <<
4924 */ 3562 */
4925 while (true) { !! 3563 rcuwait_wait_event(&manager_wait, wq_manager_inactive(pool),
4926 rcuwait_wait_event(&manager_w !! 3564 TASK_UNINTERRUPTIBLE);
4927 !(pool->fl !! 3565 pool->flags |= POOL_MANAGER_ACTIVE;
4928 TASK_UNINT <<
4929 <<
4930 mutex_lock(&wq_pool_attach_mu <<
4931 raw_spin_lock_irq(&pool->lock <<
4932 if (!(pool->flags & POOL_MANA <<
4933 pool->flags |= POOL_M <<
4934 break; <<
4935 } <<
4936 raw_spin_unlock_irq(&pool->lo <<
4937 mutex_unlock(&wq_pool_attach_ <<
4938 } <<
4939 3566
4940 while ((worker = first_idle_worker(po 3567 while ((worker = first_idle_worker(pool)))
4941 set_worker_dying(worker, &cul !! 3568 destroy_worker(worker);
4942 WARN_ON(pool->nr_workers || pool->nr_ 3569 WARN_ON(pool->nr_workers || pool->nr_idle);
4943 raw_spin_unlock_irq(&pool->lock); 3570 raw_spin_unlock_irq(&pool->lock);
4944 3571
4945 detach_dying_workers(&cull_list); !! 3572 mutex_lock(&wq_pool_attach_mutex);
4946 !! 3573 if (!list_empty(&pool->workers))
>> 3574 pool->detach_completion = &detach_completion;
4947 mutex_unlock(&wq_pool_attach_mutex); 3575 mutex_unlock(&wq_pool_attach_mutex);
4948 3576
4949 reap_dying_workers(&cull_list); !! 3577 if (pool->detach_completion)
>> 3578 wait_for_completion(pool->detach_completion);
4950 3579
4951 /* shut down the timers */ 3580 /* shut down the timers */
4952 del_timer_sync(&pool->idle_timer); 3581 del_timer_sync(&pool->idle_timer);
4953 cancel_work_sync(&pool->idle_cull_wor <<
4954 del_timer_sync(&pool->mayday_timer); 3582 del_timer_sync(&pool->mayday_timer);
4955 3583
4956 /* RCU protected to allow dereference 3584 /* RCU protected to allow dereferences from get_work_pool() */
4957 call_rcu(&pool->rcu, rcu_free_pool); 3585 call_rcu(&pool->rcu, rcu_free_pool);
4958 } 3586 }
4959 3587
4960 /** 3588 /**
4961 * get_unbound_pool - get a worker_pool with 3589 * get_unbound_pool - get a worker_pool with the specified attributes
4962 * @attrs: the attributes of the worker_pool 3590 * @attrs: the attributes of the worker_pool to get
4963 * 3591 *
4964 * Obtain a worker_pool which has the same at 3592 * Obtain a worker_pool which has the same attributes as @attrs, bump the
4965 * reference count and return it. If there a 3593 * reference count and return it. If there already is a matching
4966 * worker_pool, it will be used; otherwise, t 3594 * worker_pool, it will be used; otherwise, this function attempts to
4967 * create a new one. 3595 * create a new one.
4968 * 3596 *
4969 * Should be called with wq_pool_mutex held. 3597 * Should be called with wq_pool_mutex held.
4970 * 3598 *
4971 * Return: On success, a worker_pool with the 3599 * Return: On success, a worker_pool with the same attributes as @attrs.
4972 * On failure, %NULL. 3600 * On failure, %NULL.
4973 */ 3601 */
4974 static struct worker_pool *get_unbound_pool(c 3602 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
4975 { 3603 {
4976 struct wq_pod_type *pt = &wq_pod_type <<
4977 u32 hash = wqattrs_hash(attrs); 3604 u32 hash = wqattrs_hash(attrs);
4978 struct worker_pool *pool; 3605 struct worker_pool *pool;
4979 int pod, node = NUMA_NO_NODE; !! 3606 int node;
>> 3607 int target_node = NUMA_NO_NODE;
4980 3608
4981 lockdep_assert_held(&wq_pool_mutex); 3609 lockdep_assert_held(&wq_pool_mutex);
4982 3610
4983 /* do we already have a matching pool 3611 /* do we already have a matching pool? */
4984 hash_for_each_possible(unbound_pool_h 3612 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
4985 if (wqattrs_equal(pool->attrs 3613 if (wqattrs_equal(pool->attrs, attrs)) {
4986 pool->refcnt++; 3614 pool->refcnt++;
4987 return pool; 3615 return pool;
4988 } 3616 }
4989 } 3617 }
4990 3618
4991 /* If __pod_cpumask is contained insi !! 3619 /* if cpumask is contained inside a NUMA node, we belong to that node */
4992 for (pod = 0; pod < pt->nr_pods; pod+ !! 3620 if (wq_numa_enabled) {
4993 if (cpumask_subset(attrs->__p !! 3621 for_each_node(node) {
4994 node = pt->pod_node[p !! 3622 if (cpumask_subset(attrs->cpumask,
4995 break; !! 3623 wq_numa_possible_cpumask[node])) {
>> 3624 target_node = node;
>> 3625 break;
>> 3626 }
4996 } 3627 }
4997 } 3628 }
4998 3629
4999 /* nope, create a new one */ 3630 /* nope, create a new one */
5000 pool = kzalloc_node(sizeof(*pool), GF !! 3631 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node);
5001 if (!pool || init_worker_pool(pool) < 3632 if (!pool || init_worker_pool(pool) < 0)
5002 goto fail; 3633 goto fail;
5003 3634
5004 pool->node = node; !! 3635 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
5005 copy_workqueue_attrs(pool->attrs, att 3636 copy_workqueue_attrs(pool->attrs, attrs);
5006 wqattrs_clear_for_pool(pool->attrs); !! 3637 pool->node = target_node;
>> 3638
>> 3639 /*
>> 3640 * no_numa isn't a worker_pool attribute, always clear it. See
>> 3641 * 'struct workqueue_attrs' comments for detail.
>> 3642 */
>> 3643 pool->attrs->no_numa = false;
5007 3644
5008 if (worker_pool_assign_id(pool) < 0) 3645 if (worker_pool_assign_id(pool) < 0)
5009 goto fail; 3646 goto fail;
5010 3647
5011 /* create and start the initial worke 3648 /* create and start the initial worker */
5012 if (wq_online && !create_worker(pool) 3649 if (wq_online && !create_worker(pool))
5013 goto fail; 3650 goto fail;
5014 3651
5015 /* install */ 3652 /* install */
5016 hash_add(unbound_pool_hash, &pool->ha 3653 hash_add(unbound_pool_hash, &pool->hash_node, hash);
5017 3654
5018 return pool; 3655 return pool;
5019 fail: 3656 fail:
5020 if (pool) 3657 if (pool)
5021 put_unbound_pool(pool); 3658 put_unbound_pool(pool);
5022 return NULL; 3659 return NULL;
5023 } 3660 }
5024 3661
>> 3662 static void rcu_free_pwq(struct rcu_head *rcu)
>> 3663 {
>> 3664 kmem_cache_free(pwq_cache,
>> 3665 container_of(rcu, struct pool_workqueue, rcu));
>> 3666 }
>> 3667
5025 /* 3668 /*
5026 * Scheduled on pwq_release_worker by put_pwq !! 3669 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
5027 * refcnt and needs to be destroyed. !! 3670 * and needs to be destroyed.
5028 */ 3671 */
5029 static void pwq_release_workfn(struct kthread !! 3672 static void pwq_unbound_release_workfn(struct work_struct *work)
5030 { 3673 {
5031 struct pool_workqueue *pwq = containe 3674 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
5032 !! 3675 unbound_release_work);
5033 struct workqueue_struct *wq = pwq->wq 3676 struct workqueue_struct *wq = pwq->wq;
5034 struct worker_pool *pool = pwq->pool; 3677 struct worker_pool *pool = pwq->pool;
5035 bool is_last = false; !! 3678 bool is_last;
5036 3679
5037 /* !! 3680 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
5038 * When @pwq is not linked, it doesn' !! 3681 return;
5039 * @wq, and @wq is invalid to access. <<
5040 */ <<
5041 if (!list_empty(&pwq->pwqs_node)) { <<
5042 mutex_lock(&wq->mutex); <<
5043 list_del_rcu(&pwq->pwqs_node) <<
5044 is_last = list_empty(&wq->pwq <<
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); <<
5053 } <<
5054 <<
5055 if (wq->flags & WQ_UNBOUND) { <<
5056 mutex_lock(&wq_pool_mutex); <<
5057 put_unbound_pool(pool); <<
5058 mutex_unlock(&wq_pool_mutex); <<
5059 } <<
5060 3682
5061 if (!list_empty(&pwq->pending_node)) !! 3683 mutex_lock(&wq->mutex);
5062 struct wq_node_nr_active *nna !! 3684 list_del_rcu(&pwq->pwqs_node);
5063 wq_node_nr_active(pwq !! 3685 is_last = list_empty(&wq->pwqs);
>> 3686 mutex_unlock(&wq->mutex);
5064 3687
5065 raw_spin_lock_irq(&nna->lock) !! 3688 mutex_lock(&wq_pool_mutex);
5066 list_del_init(&pwq->pending_n !! 3689 put_unbound_pool(pool);
5067 raw_spin_unlock_irq(&nna->loc !! 3690 mutex_unlock(&wq_pool_mutex);
5068 } <<
5069 3691
5070 kfree_rcu(pwq, rcu); !! 3692 call_rcu(&pwq->rcu, rcu_free_pwq);
5071 3693
5072 /* 3694 /*
5073 * If we're the last pwq going away, 3695 * If we're the last pwq going away, @wq is already dead and no one
5074 * is gonna access it anymore. Sched 3696 * is gonna access it anymore. Schedule RCU free.
5075 */ 3697 */
5076 if (is_last) { 3698 if (is_last) {
5077 wq_unregister_lockdep(wq); 3699 wq_unregister_lockdep(wq);
5078 call_rcu(&wq->rcu, rcu_free_w 3700 call_rcu(&wq->rcu, rcu_free_wq);
5079 } 3701 }
5080 } 3702 }
5081 3703
5082 /* initialize newly allocated @pwq which is a !! 3704 /**
>> 3705 * pwq_adjust_max_active - update a pwq's max_active to the current setting
>> 3706 * @pwq: target pool_workqueue
>> 3707 *
>> 3708 * If @pwq isn't freezing, set @pwq->max_active to the associated
>> 3709 * workqueue's saved_max_active and activate delayed work items
>> 3710 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
>> 3711 */
>> 3712 static void pwq_adjust_max_active(struct pool_workqueue *pwq)
>> 3713 {
>> 3714 struct workqueue_struct *wq = pwq->wq;
>> 3715 bool freezable = wq->flags & WQ_FREEZABLE;
>> 3716 unsigned long flags;
>> 3717
>> 3718 /* for @wq->saved_max_active */
>> 3719 lockdep_assert_held(&wq->mutex);
>> 3720
>> 3721 /* fast exit for non-freezable wqs */
>> 3722 if (!freezable && pwq->max_active == wq->saved_max_active)
>> 3723 return;
>> 3724
>> 3725 /* this function can be called during early boot w/ irq disabled */
>> 3726 raw_spin_lock_irqsave(&pwq->pool->lock, flags);
>> 3727
>> 3728 /*
>> 3729 * During [un]freezing, the caller is responsible for ensuring that
>> 3730 * this function is called at least once after @workqueue_freezing
>> 3731 * is updated and visible.
>> 3732 */
>> 3733 if (!freezable || !workqueue_freezing) {
>> 3734 bool kick = false;
>> 3735
>> 3736 pwq->max_active = wq->saved_max_active;
>> 3737
>> 3738 while (!list_empty(&pwq->delayed_works) &&
>> 3739 pwq->nr_active < pwq->max_active) {
>> 3740 pwq_activate_first_delayed(pwq);
>> 3741 kick = true;
>> 3742 }
>> 3743
>> 3744 /*
>> 3745 * Need to kick a worker after thawed or an unbound wq's
>> 3746 * max_active is bumped. In realtime scenarios, always kicking a
>> 3747 * worker will cause interference on the isolated cpu cores, so
>> 3748 * let's kick iff work items were activated.
>> 3749 */
>> 3750 if (kick)
>> 3751 wake_up_worker(pwq->pool);
>> 3752 } else {
>> 3753 pwq->max_active = 0;
>> 3754 }
>> 3755
>> 3756 raw_spin_unlock_irqrestore(&pwq->pool->lock, flags);
>> 3757 }
>> 3758
>> 3759 /* initialize newly alloced @pwq which is associated with @wq and @pool */
5083 static void init_pwq(struct pool_workqueue *p 3760 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
5084 struct worker_pool *pool 3761 struct worker_pool *pool)
5085 { 3762 {
5086 BUG_ON((unsigned long)pwq & ~WORK_STR !! 3763 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
5087 3764
5088 memset(pwq, 0, sizeof(*pwq)); 3765 memset(pwq, 0, sizeof(*pwq));
5089 3766
5090 pwq->pool = pool; 3767 pwq->pool = pool;
5091 pwq->wq = wq; 3768 pwq->wq = wq;
5092 pwq->flush_color = -1; 3769 pwq->flush_color = -1;
5093 pwq->refcnt = 1; 3770 pwq->refcnt = 1;
5094 INIT_LIST_HEAD(&pwq->inactive_works); !! 3771 INIT_LIST_HEAD(&pwq->delayed_works);
5095 INIT_LIST_HEAD(&pwq->pending_node); <<
5096 INIT_LIST_HEAD(&pwq->pwqs_node); 3772 INIT_LIST_HEAD(&pwq->pwqs_node);
5097 INIT_LIST_HEAD(&pwq->mayday_node); 3773 INIT_LIST_HEAD(&pwq->mayday_node);
5098 kthread_init_work(&pwq->release_work, !! 3774 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
5099 } 3775 }
5100 3776
5101 /* sync @pwq with the current state of its as 3777 /* sync @pwq with the current state of its associated wq and link it */
5102 static void link_pwq(struct pool_workqueue *p 3778 static void link_pwq(struct pool_workqueue *pwq)
5103 { 3779 {
5104 struct workqueue_struct *wq = pwq->wq 3780 struct workqueue_struct *wq = pwq->wq;
5105 3781
5106 lockdep_assert_held(&wq->mutex); 3782 lockdep_assert_held(&wq->mutex);
5107 3783
5108 /* may be called multiple times, igno 3784 /* may be called multiple times, ignore if already linked */
5109 if (!list_empty(&pwq->pwqs_node)) 3785 if (!list_empty(&pwq->pwqs_node))
5110 return; 3786 return;
5111 3787
5112 /* set the matching work_color */ 3788 /* set the matching work_color */
5113 pwq->work_color = wq->work_color; 3789 pwq->work_color = wq->work_color;
5114 3790
>> 3791 /* sync max_active to the current setting */
>> 3792 pwq_adjust_max_active(pwq);
>> 3793
5115 /* link in @pwq */ 3794 /* link in @pwq */
5116 list_add_tail_rcu(&pwq->pwqs_node, &w !! 3795 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
5117 } 3796 }
5118 3797
5119 /* obtain a pool matching @attr and create a 3798 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
5120 static struct pool_workqueue *alloc_unbound_p 3799 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
5121 const 3800 const struct workqueue_attrs *attrs)
5122 { 3801 {
5123 struct worker_pool *pool; 3802 struct worker_pool *pool;
5124 struct pool_workqueue *pwq; 3803 struct pool_workqueue *pwq;
5125 3804
5126 lockdep_assert_held(&wq_pool_mutex); 3805 lockdep_assert_held(&wq_pool_mutex);
5127 3806
5128 pool = get_unbound_pool(attrs); 3807 pool = get_unbound_pool(attrs);
5129 if (!pool) 3808 if (!pool)
5130 return NULL; 3809 return NULL;
5131 3810
5132 pwq = kmem_cache_alloc_node(pwq_cache 3811 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
5133 if (!pwq) { 3812 if (!pwq) {
5134 put_unbound_pool(pool); 3813 put_unbound_pool(pool);
5135 return NULL; 3814 return NULL;
5136 } 3815 }
5137 3816
5138 init_pwq(pwq, wq, pool); 3817 init_pwq(pwq, wq, pool);
5139 return pwq; 3818 return pwq;
5140 } 3819 }
5141 3820
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 /** 3821 /**
5153 * wq_calc_pod_cpumask - calculate a wq_attrs !! 3822 * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
5154 * @attrs: the wq_attrs of the default pwq of 3823 * @attrs: the wq_attrs of the default pwq of the target workqueue
5155 * @cpu: the target CPU !! 3824 * @node: the target NUMA node
>> 3825 * @cpu_going_down: if >= 0, the CPU to consider as offline
>> 3826 * @cpumask: outarg, the resulting cpumask
>> 3827 *
>> 3828 * Calculate the cpumask a workqueue with @attrs should use on @node. If
>> 3829 * @cpu_going_down is >= 0, that cpu is considered offline during
>> 3830 * calculation. The result is stored in @cpumask.
5156 * 3831 *
5157 * Calculate the cpumask a workqueue with @at !! 3832 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
5158 * The result is stored in @attrs->__pod_cpum !! 3833 * enabled and @node has online CPUs requested by @attrs, the returned
>> 3834 * cpumask is the intersection of the possible CPUs of @node and
>> 3835 * @attrs->cpumask.
5159 * 3836 *
5160 * If pod affinity is not enabled, @attrs->cp !! 3837 * The caller is responsible for ensuring that the cpumask of @node stays
5161 * and @pod has online CPUs requested by @att !! 3838 * stable.
5162 * intersection of the possible CPUs of @pod <<
5163 * 3839 *
5164 * The caller is responsible for ensuring tha !! 3840 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
>> 3841 * %false if equal.
5165 */ 3842 */
5166 static void wq_calc_pod_cpumask(struct workqu !! 3843 static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
>> 3844 int cpu_going_down, cpumask_t *cpumask)
5167 { 3845 {
5168 const struct wq_pod_type *pt = wqattr !! 3846 if (!wq_numa_enabled || attrs->no_numa)
5169 int pod = pt->cpu_pod[cpu]; !! 3847 goto use_dfl;
5170 3848
5171 /* calculate possible CPUs in @pod th !! 3849 /* does @node have any online CPUs @attrs wants? */
5172 cpumask_and(attrs->__pod_cpumask, pt- !! 3850 cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
5173 /* does @pod have any online CPUs @at !! 3851 if (cpu_going_down >= 0)
5174 if (!cpumask_intersects(attrs->__pod_ !! 3852 cpumask_clear_cpu(cpu_going_down, cpumask);
5175 cpumask_copy(attrs->__pod_cpu !! 3853
5176 return; !! 3854 if (cpumask_empty(cpumask))
>> 3855 goto use_dfl;
>> 3856
>> 3857 /* yeap, return possible CPUs in @node that @attrs wants */
>> 3858 cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
>> 3859
>> 3860 if (cpumask_empty(cpumask)) {
>> 3861 pr_warn_once("WARNING: workqueue cpumask: online intersect > "
>> 3862 "possible intersect\n");
>> 3863 return false;
5177 } 3864 }
>> 3865
>> 3866 return !cpumask_equal(cpumask, attrs->cpumask);
>> 3867
>> 3868 use_dfl:
>> 3869 cpumask_copy(cpumask, attrs->cpumask);
>> 3870 return false;
5178 } 3871 }
5179 3872
5180 /* install @pwq into @wq and return the old p !! 3873 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
5181 static struct pool_workqueue *install_unbound !! 3874 static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
5182 int c !! 3875 int node,
>> 3876 struct pool_workqueue *pwq)
5183 { 3877 {
5184 struct pool_workqueue __rcu **slot = <<
5185 struct pool_workqueue *old_pwq; 3878 struct pool_workqueue *old_pwq;
5186 3879
5187 lockdep_assert_held(&wq_pool_mutex); 3880 lockdep_assert_held(&wq_pool_mutex);
5188 lockdep_assert_held(&wq->mutex); 3881 lockdep_assert_held(&wq->mutex);
5189 3882
5190 /* link_pwq() can handle duplicate ca 3883 /* link_pwq() can handle duplicate calls */
5191 link_pwq(pwq); 3884 link_pwq(pwq);
5192 3885
5193 old_pwq = rcu_access_pointer(*slot); !! 3886 old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
5194 rcu_assign_pointer(*slot, pwq); !! 3887 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
5195 return old_pwq; 3888 return old_pwq;
5196 } 3889 }
5197 3890
5198 /* context to store the prepared attrs & pwqs 3891 /* context to store the prepared attrs & pwqs before applying */
5199 struct apply_wqattrs_ctx { 3892 struct apply_wqattrs_ctx {
5200 struct workqueue_struct *wq; 3893 struct workqueue_struct *wq; /* target workqueue */
5201 struct workqueue_attrs *attrs; 3894 struct workqueue_attrs *attrs; /* attrs to apply */
5202 struct list_head list; 3895 struct list_head list; /* queued for batching commit */
5203 struct pool_workqueue *dfl_pwq; 3896 struct pool_workqueue *dfl_pwq;
5204 struct pool_workqueue *pwq_tbl[]; 3897 struct pool_workqueue *pwq_tbl[];
5205 }; 3898 };
5206 3899
5207 /* free the resources after success or abort 3900 /* free the resources after success or abort */
5208 static void apply_wqattrs_cleanup(struct appl 3901 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
5209 { 3902 {
5210 if (ctx) { 3903 if (ctx) {
5211 int cpu; !! 3904 int node;
5212 3905
5213 for_each_possible_cpu(cpu) !! 3906 for_each_node(node)
5214 put_pwq_unlocked(ctx- !! 3907 put_pwq_unlocked(ctx->pwq_tbl[node]);
5215 put_pwq_unlocked(ctx->dfl_pwq 3908 put_pwq_unlocked(ctx->dfl_pwq);
5216 3909
5217 free_workqueue_attrs(ctx->att 3910 free_workqueue_attrs(ctx->attrs);
5218 3911
5219 kfree(ctx); 3912 kfree(ctx);
5220 } 3913 }
5221 } 3914 }
5222 3915
5223 /* allocate the attrs and pwqs for later inst 3916 /* allocate the attrs and pwqs for later installation */
5224 static struct apply_wqattrs_ctx * 3917 static struct apply_wqattrs_ctx *
5225 apply_wqattrs_prepare(struct workqueue_struct 3918 apply_wqattrs_prepare(struct workqueue_struct *wq,
5226 const struct workqueue_ !! 3919 const struct workqueue_attrs *attrs)
5227 const cpumask_var_t unb <<
5228 { 3920 {
5229 struct apply_wqattrs_ctx *ctx; 3921 struct apply_wqattrs_ctx *ctx;
5230 struct workqueue_attrs *new_attrs; !! 3922 struct workqueue_attrs *new_attrs, *tmp_attrs;
5231 int cpu; !! 3923 int node;
5232 3924
5233 lockdep_assert_held(&wq_pool_mutex); 3925 lockdep_assert_held(&wq_pool_mutex);
5234 3926
5235 if (WARN_ON(attrs->affn_scope < 0 || !! 3927 ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_node_ids), GFP_KERNEL);
5236 attrs->affn_scope >= WQ_A <<
5237 return ERR_PTR(-EINVAL); <<
5238 <<
5239 ctx = kzalloc(struct_size(ctx, pwq_tb <<
5240 3928
5241 new_attrs = alloc_workqueue_attrs(); 3929 new_attrs = alloc_workqueue_attrs();
5242 if (!ctx || !new_attrs) !! 3930 tmp_attrs = alloc_workqueue_attrs();
>> 3931 if (!ctx || !new_attrs || !tmp_attrs)
5243 goto out_free; 3932 goto out_free;
5244 3933
5245 /* 3934 /*
>> 3935 * Calculate the attrs of the default pwq.
>> 3936 * If the user configured cpumask doesn't overlap with the
>> 3937 * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
>> 3938 */
>> 3939 copy_workqueue_attrs(new_attrs, attrs);
>> 3940 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask);
>> 3941 if (unlikely(cpumask_empty(new_attrs->cpumask)))
>> 3942 cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask);
>> 3943
>> 3944 /*
>> 3945 * We may create multiple pwqs with differing cpumasks. Make a
>> 3946 * copy of @new_attrs which will be modified and used to obtain
>> 3947 * pools.
>> 3948 */
>> 3949 copy_workqueue_attrs(tmp_attrs, new_attrs);
>> 3950
>> 3951 /*
5246 * If something goes wrong during CPU 3952 * If something goes wrong during CPU up/down, we'll fall back to
5247 * the default pwq covering whole @at 3953 * the default pwq covering whole @attrs->cpumask. Always create
5248 * it even if we don't use it immedia 3954 * it even if we don't use it immediately.
5249 */ 3955 */
5250 copy_workqueue_attrs(new_attrs, attrs <<
5251 wqattrs_actualize_cpumask(new_attrs, <<
5252 cpumask_copy(new_attrs->__pod_cpumask <<
5253 ctx->dfl_pwq = alloc_unbound_pwq(wq, 3956 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
5254 if (!ctx->dfl_pwq) 3957 if (!ctx->dfl_pwq)
5255 goto out_free; 3958 goto out_free;
5256 3959
5257 for_each_possible_cpu(cpu) { !! 3960 for_each_node(node) {
5258 if (new_attrs->ordered) { !! 3961 if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) {
5259 ctx->dfl_pwq->refcnt+ !! 3962 ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
5260 ctx->pwq_tbl[cpu] = c !! 3963 if (!ctx->pwq_tbl[node])
5261 } else { <<
5262 wq_calc_pod_cpumask(n <<
5263 ctx->pwq_tbl[cpu] = a <<
5264 if (!ctx->pwq_tbl[cpu <<
5265 goto out_free 3964 goto out_free;
>> 3965 } else {
>> 3966 ctx->dfl_pwq->refcnt++;
>> 3967 ctx->pwq_tbl[node] = ctx->dfl_pwq;
5266 } 3968 }
5267 } 3969 }
5268 3970
5269 /* save the user configured attrs and 3971 /* save the user configured attrs and sanitize it. */
5270 copy_workqueue_attrs(new_attrs, attrs 3972 copy_workqueue_attrs(new_attrs, attrs);
5271 cpumask_and(new_attrs->cpumask, new_a 3973 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
5272 cpumask_copy(new_attrs->__pod_cpumask <<
5273 ctx->attrs = new_attrs; 3974 ctx->attrs = new_attrs;
5274 3975
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; 3976 ctx->wq = wq;
>> 3977 free_workqueue_attrs(tmp_attrs);
5285 return ctx; 3978 return ctx;
5286 3979
5287 out_free: 3980 out_free:
>> 3981 free_workqueue_attrs(tmp_attrs);
5288 free_workqueue_attrs(new_attrs); 3982 free_workqueue_attrs(new_attrs);
5289 apply_wqattrs_cleanup(ctx); 3983 apply_wqattrs_cleanup(ctx);
5290 return ERR_PTR(-ENOMEM); !! 3984 return NULL;
5291 } 3985 }
5292 3986
5293 /* set attrs and install prepared pwqs, @ctx 3987 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
5294 static void apply_wqattrs_commit(struct apply 3988 static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
5295 { 3989 {
5296 int cpu; !! 3990 int node;
5297 3991
5298 /* all pwqs have been created success 3992 /* all pwqs have been created successfully, let's install'em */
5299 mutex_lock(&ctx->wq->mutex); 3993 mutex_lock(&ctx->wq->mutex);
5300 3994
5301 copy_workqueue_attrs(ctx->wq->unbound 3995 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
5302 3996
5303 /* save the previous pwqs and install !! 3997 /* save the previous pwq and install the new one */
5304 for_each_possible_cpu(cpu) !! 3998 for_each_node(node)
5305 ctx->pwq_tbl[cpu] = install_u !! 3999 ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node,
5306 !! 4000 ctx->pwq_tbl[node]);
5307 ctx->dfl_pwq = install_unbound_pwq(ct !! 4001
5308 !! 4002 /* @dfl_pwq might not have been used, ensure it's linked */
5309 /* update node_nr_active->max */ !! 4003 link_pwq(ctx->dfl_pwq);
5310 wq_update_node_max_active(ctx->wq, -1 !! 4004 swap(ctx->wq->dfl_pwq, ctx->dfl_pwq);
5311 <<
5312 /* rescuer needs to respect wq cpumas <<
5313 if (ctx->wq->rescuer) <<
5314 set_cpus_allowed_ptr(ctx->wq- <<
5315 unbound_ <<
5316 4005
5317 mutex_unlock(&ctx->wq->mutex); 4006 mutex_unlock(&ctx->wq->mutex);
5318 } 4007 }
5319 4008
>> 4009 static void apply_wqattrs_lock(void)
>> 4010 {
>> 4011 /* CPUs should stay stable across pwq creations and installations */
>> 4012 get_online_cpus();
>> 4013 mutex_lock(&wq_pool_mutex);
>> 4014 }
>> 4015
>> 4016 static void apply_wqattrs_unlock(void)
>> 4017 {
>> 4018 mutex_unlock(&wq_pool_mutex);
>> 4019 put_online_cpus();
>> 4020 }
>> 4021
5320 static int apply_workqueue_attrs_locked(struc 4022 static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
5321 const 4023 const struct workqueue_attrs *attrs)
5322 { 4024 {
5323 struct apply_wqattrs_ctx *ctx; 4025 struct apply_wqattrs_ctx *ctx;
5324 4026
5325 /* only unbound workqueues can change 4027 /* only unbound workqueues can change attributes */
5326 if (WARN_ON(!(wq->flags & WQ_UNBOUND) 4028 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
5327 return -EINVAL; 4029 return -EINVAL;
5328 4030
5329 ctx = apply_wqattrs_prepare(wq, attrs !! 4031 /* creating multiple pwqs breaks ordering guarantee */
5330 if (IS_ERR(ctx)) !! 4032 if (!list_empty(&wq->pwqs)) {
5331 return PTR_ERR(ctx); !! 4033 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
>> 4034 return -EINVAL;
>> 4035
>> 4036 wq->flags &= ~__WQ_ORDERED;
>> 4037 }
>> 4038
>> 4039 ctx = apply_wqattrs_prepare(wq, attrs);
>> 4040 if (!ctx)
>> 4041 return -ENOMEM;
5332 4042
5333 /* the ctx has been prepared successf 4043 /* the ctx has been prepared successfully, let's commit it */
5334 apply_wqattrs_commit(ctx); 4044 apply_wqattrs_commit(ctx);
5335 apply_wqattrs_cleanup(ctx); 4045 apply_wqattrs_cleanup(ctx);
5336 4046
5337 return 0; 4047 return 0;
5338 } 4048 }
5339 4049
5340 /** 4050 /**
5341 * apply_workqueue_attrs - apply new workqueu 4051 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
5342 * @wq: the target workqueue 4052 * @wq: the target workqueue
5343 * @attrs: the workqueue_attrs to apply, allo 4053 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
5344 * 4054 *
5345 * Apply @attrs to an unbound workqueue @wq. !! 4055 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
5346 * a separate pwq to each CPU pod with possib !! 4056 * machines, this function maps a separate pwq to each NUMA node with
5347 * work items are affine to the pod it was is !! 4057 * possibles CPUs in @attrs->cpumask so that work items are affine to the
5348 * in-flight work items finish. Note that a w !! 4058 * NUMA node it was issued on. Older pwqs are released as in-flight work
5349 * itself back-to-back will stay on its curre !! 4059 * items finish. Note that a work item which repeatedly requeues itself
>> 4060 * back-to-back will stay on its current pwq.
5350 * 4061 *
5351 * Performs GFP_KERNEL allocations. 4062 * Performs GFP_KERNEL allocations.
5352 * 4063 *
>> 4064 * Assumes caller has CPU hotplug read exclusion, i.e. get_online_cpus().
>> 4065 *
5353 * Return: 0 on success and -errno on failure 4066 * Return: 0 on success and -errno on failure.
5354 */ 4067 */
5355 int apply_workqueue_attrs(struct workqueue_st 4068 int apply_workqueue_attrs(struct workqueue_struct *wq,
5356 const struct workqu 4069 const struct workqueue_attrs *attrs)
5357 { 4070 {
5358 int ret; 4071 int ret;
5359 4072
>> 4073 lockdep_assert_cpus_held();
>> 4074
5360 mutex_lock(&wq_pool_mutex); 4075 mutex_lock(&wq_pool_mutex);
5361 ret = apply_workqueue_attrs_locked(wq 4076 ret = apply_workqueue_attrs_locked(wq, attrs);
5362 mutex_unlock(&wq_pool_mutex); 4077 mutex_unlock(&wq_pool_mutex);
5363 4078
5364 return ret; 4079 return ret;
5365 } 4080 }
5366 4081
5367 /** 4082 /**
5368 * unbound_wq_update_pwq - update a pwq slot !! 4083 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
5369 * @wq: the target workqueue 4084 * @wq: the target workqueue
5370 * @cpu: the CPU to update the pwq slot for !! 4085 * @cpu: the CPU coming up or going down
>> 4086 * @online: whether @cpu is coming up or going down
5371 * 4087 *
5372 * This function is to be called from %CPU_DO 4088 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
5373 * %CPU_DOWN_FAILED. @cpu is in the same pod !! 4089 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
5374 * !! 4090 * @wq accordingly.
5375 * 4091 *
5376 * If pod affinity can't be adjusted due to m !! 4092 * If NUMA affinity can't be adjusted due to memory allocation failure, it
5377 * back to @wq->dfl_pwq which may not be opti !! 4093 * falls back to @wq->dfl_pwq which may not be optimal but is always
5378 * !! 4094 * correct.
5379 * Note that when the last allowed CPU of a p !! 4095 *
5380 * with a cpumask spanning multiple pods, the !! 4096 * Note that when the last allowed CPU of a NUMA node goes offline for a
5381 * executing the work items for the workqueue !! 4097 * workqueue with a cpumask spanning multiple nodes, the workers which were
5382 * may execute on any CPU. This is similar to !! 4098 * already executing the work items for the workqueue will lose their CPU
5383 * CPU_DOWN. If a workqueue user wants strict !! 4099 * affinity and may execute on any CPU. This is similar to how per-cpu
5384 * responsibility to flush the work item from !! 4100 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
>> 4101 * affinity, it's the user's responsibility to flush the work item from
>> 4102 * CPU_DOWN_PREPARE.
5385 */ 4103 */
5386 static void unbound_wq_update_pwq(struct work !! 4104 static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
>> 4105 bool online)
5387 { 4106 {
>> 4107 int node = cpu_to_node(cpu);
>> 4108 int cpu_off = online ? -1 : cpu;
5388 struct pool_workqueue *old_pwq = NULL 4109 struct pool_workqueue *old_pwq = NULL, *pwq;
5389 struct workqueue_attrs *target_attrs; 4110 struct workqueue_attrs *target_attrs;
>> 4111 cpumask_t *cpumask;
5390 4112
5391 lockdep_assert_held(&wq_pool_mutex); 4113 lockdep_assert_held(&wq_pool_mutex);
5392 4114
5393 if (!(wq->flags & WQ_UNBOUND) || wq-> !! 4115 if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) ||
>> 4116 wq->unbound_attrs->no_numa)
5394 return; 4117 return;
5395 4118
5396 /* 4119 /*
5397 * We don't wanna alloc/free wq_attrs 4120 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
5398 * Let's use a preallocated one. The 4121 * Let's use a preallocated one. The following buf is protected by
5399 * CPU hotplug exclusion. 4122 * CPU hotplug exclusion.
5400 */ 4123 */
5401 target_attrs = unbound_wq_update_pwq_ !! 4124 target_attrs = wq_update_unbound_numa_attrs_buf;
>> 4125 cpumask = target_attrs->cpumask;
5402 4126
5403 copy_workqueue_attrs(target_attrs, wq 4127 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
5404 wqattrs_actualize_cpumask(target_attr !! 4128 pwq = unbound_pwq_by_node(wq, node);
5405 4129
5406 /* nothing to do if the target cpumas !! 4130 /*
5407 wq_calc_pod_cpumask(target_attrs, cpu !! 4131 * Let's determine what needs to be done. If the target cpumask is
5408 if (wqattrs_equal(target_attrs, unbou !! 4132 * different from the default pwq's, we need to compare it to @pwq's
5409 return; !! 4133 * and create a new one if they don't match. If the target cpumask
>> 4134 * equals the default pwq's, the default pwq should be used.
>> 4135 */
>> 4136 if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) {
>> 4137 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
>> 4138 return;
>> 4139 } else {
>> 4140 goto use_dfl_pwq;
>> 4141 }
5410 4142
5411 /* create a new pwq */ 4143 /* create a new pwq */
5412 pwq = alloc_unbound_pwq(wq, target_at 4144 pwq = alloc_unbound_pwq(wq, target_attrs);
5413 if (!pwq) { 4145 if (!pwq) {
5414 pr_warn("workqueue: allocatio !! 4146 pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
5415 wq->name); 4147 wq->name);
5416 goto use_dfl_pwq; 4148 goto use_dfl_pwq;
5417 } 4149 }
5418 4150
5419 /* Install the new pwq. */ 4151 /* Install the new pwq. */
5420 mutex_lock(&wq->mutex); 4152 mutex_lock(&wq->mutex);
5421 old_pwq = install_unbound_pwq(wq, cpu !! 4153 old_pwq = numa_pwq_tbl_install(wq, node, pwq);
5422 goto out_unlock; 4154 goto out_unlock;
5423 4155
5424 use_dfl_pwq: 4156 use_dfl_pwq:
5425 mutex_lock(&wq->mutex); 4157 mutex_lock(&wq->mutex);
5426 pwq = unbound_pwq(wq, -1); !! 4158 raw_spin_lock_irq(&wq->dfl_pwq->pool->lock);
5427 raw_spin_lock_irq(&pwq->pool->lock); !! 4159 get_pwq(wq->dfl_pwq);
5428 get_pwq(pwq); !! 4160 raw_spin_unlock_irq(&wq->dfl_pwq->pool->lock);
5429 raw_spin_unlock_irq(&pwq->pool->lock) !! 4161 old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
5430 old_pwq = install_unbound_pwq(wq, cpu <<
5431 out_unlock: 4162 out_unlock:
5432 mutex_unlock(&wq->mutex); 4163 mutex_unlock(&wq->mutex);
5433 put_pwq_unlocked(old_pwq); 4164 put_pwq_unlocked(old_pwq);
5434 } 4165 }
5435 4166
5436 static int alloc_and_link_pwqs(struct workque 4167 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
5437 { 4168 {
5438 bool highpri = wq->flags & WQ_HIGHPRI 4169 bool highpri = wq->flags & WQ_HIGHPRI;
5439 int cpu, ret; 4170 int cpu, ret;
5440 4171
5441 lockdep_assert_held(&wq_pool_mutex); <<
5442 <<
5443 wq->cpu_pwq = alloc_percpu(struct poo <<
5444 if (!wq->cpu_pwq) <<
5445 goto enomem; <<
5446 <<
5447 if (!(wq->flags & WQ_UNBOUND)) { 4172 if (!(wq->flags & WQ_UNBOUND)) {
5448 struct worker_pool __percpu * !! 4173 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
5449 !! 4174 if (!wq->cpu_pwqs)
5450 if (wq->flags & WQ_BH) !! 4175 return -ENOMEM;
5451 pools = bh_worker_poo <<
5452 else <<
5453 pools = cpu_worker_po <<
5454 4176
5455 for_each_possible_cpu(cpu) { 4177 for_each_possible_cpu(cpu) {
5456 struct pool_workqueue !! 4178 struct pool_workqueue *pwq =
5457 struct worker_pool *p !! 4179 per_cpu_ptr(wq->cpu_pwqs, cpu);
>> 4180 struct worker_pool *cpu_pools =
>> 4181 per_cpu(cpu_worker_pools, cpu);
5458 4182
5459 pool = &(per_cpu_ptr( !! 4183 init_pwq(pwq, wq, &cpu_pools[highpri]);
5460 pwq_p = per_cpu_ptr(w <<
5461 <<
5462 *pwq_p = kmem_cache_a <<
5463 <<
5464 if (!*pwq_p) <<
5465 goto enomem; <<
5466 <<
5467 init_pwq(*pwq_p, wq, <<
5468 4184
5469 mutex_lock(&wq->mutex 4185 mutex_lock(&wq->mutex);
5470 link_pwq(*pwq_p); !! 4186 link_pwq(pwq);
5471 mutex_unlock(&wq->mut 4187 mutex_unlock(&wq->mutex);
5472 } 4188 }
5473 return 0; 4189 return 0;
5474 } 4190 }
5475 4191
>> 4192 get_online_cpus();
5476 if (wq->flags & __WQ_ORDERED) { 4193 if (wq->flags & __WQ_ORDERED) {
5477 struct pool_workqueue *dfl_pw !! 4194 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
5478 <<
5479 ret = apply_workqueue_attrs_l <<
5480 /* there should only be singl 4195 /* there should only be single pwq for ordering guarantee */
5481 dfl_pwq = rcu_access_pointer( !! 4196 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
5482 WARN(!ret && (wq->pwqs.next ! !! 4197 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
5483 wq->pwqs.prev ! <<
5484 "ordering guarantee brok 4198 "ordering guarantee broken for workqueue %s\n", wq->name);
5485 } else { 4199 } else {
5486 ret = apply_workqueue_attrs_l !! 4200 ret = apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
5487 } 4201 }
>> 4202 put_online_cpus();
5488 4203
5489 return ret; 4204 return ret;
5490 <<
5491 enomem: <<
5492 if (wq->cpu_pwq) { <<
5493 for_each_possible_cpu(cpu) { <<
5494 struct pool_workqueue <<
5495 <<
5496 if (pwq) <<
5497 kmem_cache_fr <<
5498 } <<
5499 free_percpu(wq->cpu_pwq); <<
5500 wq->cpu_pwq = NULL; <<
5501 } <<
5502 return -ENOMEM; <<
5503 } 4205 }
5504 4206
5505 static int wq_clamp_max_active(int max_active 4207 static int wq_clamp_max_active(int max_active, unsigned int flags,
5506 const char *na 4208 const char *name)
5507 { 4209 {
5508 if (max_active < 1 || max_active > WQ !! 4210 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
>> 4211
>> 4212 if (max_active < 1 || max_active > lim)
5509 pr_warn("workqueue: max_activ 4213 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
5510 max_active, name, 1, !! 4214 max_active, name, 1, lim);
5511 4215
5512 return clamp_val(max_active, 1, WQ_MA !! 4216 return clamp_val(max_active, 1, lim);
5513 } 4217 }
5514 4218
5515 /* 4219 /*
5516 * Workqueues which may be used during memory 4220 * Workqueues which may be used during memory reclaim should have a rescuer
5517 * to guarantee forward progress. 4221 * to guarantee forward progress.
5518 */ 4222 */
5519 static int init_rescuer(struct workqueue_stru 4223 static int init_rescuer(struct workqueue_struct *wq)
5520 { 4224 {
5521 struct worker *rescuer; 4225 struct worker *rescuer;
5522 char id_buf[WORKER_ID_LEN]; <<
5523 int ret; 4226 int ret;
5524 4227
5525 lockdep_assert_held(&wq_pool_mutex); <<
5526 <<
5527 if (!(wq->flags & WQ_MEM_RECLAIM)) 4228 if (!(wq->flags & WQ_MEM_RECLAIM))
5528 return 0; 4229 return 0;
5529 4230
5530 rescuer = alloc_worker(NUMA_NO_NODE); 4231 rescuer = alloc_worker(NUMA_NO_NODE);
5531 if (!rescuer) { !! 4232 if (!rescuer)
5532 pr_err("workqueue: Failed to <<
5533 wq->name); <<
5534 return -ENOMEM; 4233 return -ENOMEM;
5535 } <<
5536 4234
5537 rescuer->rescue_wq = wq; 4235 rescuer->rescue_wq = wq;
5538 format_worker_id(id_buf, sizeof(id_bu !! 4236 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", wq->name);
5539 <<
5540 rescuer->task = kthread_create(rescue <<
5541 if (IS_ERR(rescuer->task)) { 4237 if (IS_ERR(rescuer->task)) {
5542 ret = PTR_ERR(rescuer->task); 4238 ret = PTR_ERR(rescuer->task);
5543 pr_err("workqueue: Failed to <<
5544 wq->name, ERR_PTR(ret) <<
5545 kfree(rescuer); 4239 kfree(rescuer);
5546 return ret; 4240 return ret;
5547 } 4241 }
5548 4242
5549 wq->rescuer = rescuer; 4243 wq->rescuer = rescuer;
5550 if (wq->flags & WQ_UNBOUND) !! 4244 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); 4245 wake_up_process(rescuer->task);
5555 4246
5556 return 0; 4247 return 0;
5557 } 4248 }
5558 4249
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) 4250 __printf(1, 4)
5623 struct workqueue_struct *alloc_workqueue(cons 4251 struct workqueue_struct *alloc_workqueue(const char *fmt,
5624 unsi 4252 unsigned int flags,
5625 int 4253 int max_active, ...)
5626 { 4254 {
>> 4255 size_t tbl_size = 0;
5627 va_list args; 4256 va_list args;
5628 struct workqueue_struct *wq; 4257 struct workqueue_struct *wq;
5629 size_t wq_size; !! 4258 struct pool_workqueue *pwq;
5630 int name_len; <<
5631 4259
5632 if (flags & WQ_BH) { !! 4260 /*
5633 if (WARN_ON_ONCE(flags & ~__W !! 4261 * Unbound && max_active == 1 used to imply ordered, which is no
5634 return NULL; !! 4262 * longer the case on NUMA machines due to per-node pools. While
5635 if (WARN_ON_ONCE(max_active)) !! 4263 * alloc_ordered_workqueue() is the right way to create an ordered
5636 return NULL; !! 4264 * workqueue, keep the previous behavior to avoid subtle breakages
5637 } !! 4265 * on NUMA.
>> 4266 */
>> 4267 if ((flags & WQ_UNBOUND) && max_active == 1)
>> 4268 flags |= __WQ_ORDERED;
5638 4269
5639 /* see the comment above the definiti 4270 /* see the comment above the definition of WQ_POWER_EFFICIENT */
5640 if ((flags & WQ_POWER_EFFICIENT) && w 4271 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
5641 flags |= WQ_UNBOUND; 4272 flags |= WQ_UNBOUND;
5642 4273
5643 /* allocate wq and format name */ 4274 /* allocate wq and format name */
5644 if (flags & WQ_UNBOUND) 4275 if (flags & WQ_UNBOUND)
5645 wq_size = struct_size(wq, nod !! 4276 tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]);
5646 else <<
5647 wq_size = sizeof(*wq); <<
5648 4277
5649 wq = kzalloc(wq_size, GFP_KERNEL); !! 4278 wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
5650 if (!wq) 4279 if (!wq)
5651 return NULL; 4280 return NULL;
5652 4281
5653 if (flags & WQ_UNBOUND) { 4282 if (flags & WQ_UNBOUND) {
5654 wq->unbound_attrs = alloc_wor 4283 wq->unbound_attrs = alloc_workqueue_attrs();
5655 if (!wq->unbound_attrs) 4284 if (!wq->unbound_attrs)
5656 goto err_free_wq; 4285 goto err_free_wq;
5657 } 4286 }
5658 4287
5659 va_start(args, max_active); 4288 va_start(args, max_active);
5660 name_len = vsnprintf(wq->name, sizeof !! 4289 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
5661 va_end(args); 4290 va_end(args);
5662 4291
5663 if (name_len >= WQ_NAME_LEN) !! 4292 max_active = max_active ?: WQ_DFL_ACTIVE;
5664 pr_warn_once("workqueue: name !! 4293 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 4294
5678 /* init wq */ 4295 /* init wq */
5679 wq->flags = flags; 4296 wq->flags = flags;
5680 wq->max_active = max_active; !! 4297 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); 4298 mutex_init(&wq->mutex);
5685 atomic_set(&wq->nr_pwqs_to_flush, 0); 4299 atomic_set(&wq->nr_pwqs_to_flush, 0);
5686 INIT_LIST_HEAD(&wq->pwqs); 4300 INIT_LIST_HEAD(&wq->pwqs);
5687 INIT_LIST_HEAD(&wq->flusher_queue); 4301 INIT_LIST_HEAD(&wq->flusher_queue);
5688 INIT_LIST_HEAD(&wq->flusher_overflow) 4302 INIT_LIST_HEAD(&wq->flusher_overflow);
5689 INIT_LIST_HEAD(&wq->maydays); 4303 INIT_LIST_HEAD(&wq->maydays);
5690 4304
5691 wq_init_lockdep(wq); 4305 wq_init_lockdep(wq);
5692 INIT_LIST_HEAD(&wq->list); 4306 INIT_LIST_HEAD(&wq->list);
5693 4307
5694 if (flags & WQ_UNBOUND) { !! 4308 if (alloc_and_link_pwqs(wq) < 0)
5695 if (alloc_node_nr_active(wq-> !! 4309 goto err_unreg_lockdep;
5696 goto err_unreg_lockde !! 4310
5697 } !! 4311 if (wq_online && init_rescuer(wq) < 0)
>> 4312 goto err_destroy;
>> 4313
>> 4314 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
>> 4315 goto err_destroy;
5698 4316
5699 /* 4317 /*
5700 * wq_pool_mutex protects the workque !! 4318 * wq_pool_mutex protects global freeze state and workqueues list.
5701 * and the global freeze state. !! 4319 * Grab it, adjust max_active and add the new @wq to workqueues
>> 4320 * list.
5702 */ 4321 */
5703 apply_wqattrs_lock(); !! 4322 mutex_lock(&wq_pool_mutex);
5704 <<
5705 if (alloc_and_link_pwqs(wq) < 0) <<
5706 goto err_unlock_free_node_nr_ <<
5707 4323
5708 mutex_lock(&wq->mutex); 4324 mutex_lock(&wq->mutex);
5709 wq_adjust_max_active(wq); !! 4325 for_each_pwq(pwq, wq)
>> 4326 pwq_adjust_max_active(pwq);
5710 mutex_unlock(&wq->mutex); 4327 mutex_unlock(&wq->mutex);
5711 4328
5712 list_add_tail_rcu(&wq->list, &workque 4329 list_add_tail_rcu(&wq->list, &workqueues);
5713 4330
5714 if (wq_online && init_rescuer(wq) < 0 !! 4331 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 4332
5722 return wq; 4333 return wq;
5723 4334
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: 4335 err_unreg_lockdep:
5736 wq_unregister_lockdep(wq); 4336 wq_unregister_lockdep(wq);
5737 wq_free_lockdep(wq); 4337 wq_free_lockdep(wq);
5738 err_free_wq: 4338 err_free_wq:
5739 free_workqueue_attrs(wq->unbound_attr 4339 free_workqueue_attrs(wq->unbound_attrs);
5740 kfree(wq); 4340 kfree(wq);
5741 return NULL; 4341 return NULL;
5742 err_unlock_destroy: <<
5743 apply_wqattrs_unlock(); <<
5744 err_destroy: 4342 err_destroy:
5745 destroy_workqueue(wq); 4343 destroy_workqueue(wq);
5746 return NULL; 4344 return NULL;
5747 } 4345 }
5748 EXPORT_SYMBOL_GPL(alloc_workqueue); 4346 EXPORT_SYMBOL_GPL(alloc_workqueue);
5749 4347
5750 static bool pwq_busy(struct pool_workqueue *p 4348 static bool pwq_busy(struct pool_workqueue *pwq)
5751 { 4349 {
5752 int i; 4350 int i;
5753 4351
5754 for (i = 0; i < WORK_NR_COLORS; i++) 4352 for (i = 0; i < WORK_NR_COLORS; i++)
5755 if (pwq->nr_in_flight[i]) 4353 if (pwq->nr_in_flight[i])
5756 return true; 4354 return true;
5757 4355
5758 if ((pwq != rcu_access_pointer(pwq->w !! 4356 if ((pwq != pwq->wq->dfl_pwq) && (pwq->refcnt > 1))
5759 return true; 4357 return true;
5760 if (!pwq_is_empty(pwq)) !! 4358 if (pwq->nr_active || !list_empty(&pwq->delayed_works))
5761 return true; 4359 return true;
5762 4360
5763 return false; 4361 return false;
5764 } 4362 }
5765 4363
5766 /** 4364 /**
5767 * destroy_workqueue - safely terminate a wor 4365 * destroy_workqueue - safely terminate a workqueue
5768 * @wq: target workqueue 4366 * @wq: target workqueue
5769 * 4367 *
5770 * Safely destroy a workqueue. All work curre 4368 * Safely destroy a workqueue. All work currently pending will be done first.
5771 */ 4369 */
5772 void destroy_workqueue(struct workqueue_struc 4370 void destroy_workqueue(struct workqueue_struct *wq)
5773 { 4371 {
5774 struct pool_workqueue *pwq; 4372 struct pool_workqueue *pwq;
5775 int cpu; !! 4373 int node;
5776 4374
5777 /* 4375 /*
5778 * Remove it from sysfs first so that 4376 * Remove it from sysfs first so that sanity check failure doesn't
5779 * lead to sysfs name conflicts. 4377 * lead to sysfs name conflicts.
5780 */ 4378 */
5781 workqueue_sysfs_unregister(wq); 4379 workqueue_sysfs_unregister(wq);
5782 4380
5783 /* mark the workqueue destruction is <<
5784 mutex_lock(&wq->mutex); <<
5785 wq->flags |= __WQ_DESTROYING; <<
5786 mutex_unlock(&wq->mutex); <<
5787 <<
5788 /* drain it before proceeding with de 4381 /* drain it before proceeding with destruction */
5789 drain_workqueue(wq); 4382 drain_workqueue(wq);
5790 4383
5791 /* kill rescuer, if sanity checks fai 4384 /* kill rescuer, if sanity checks fail, leave it w/o rescuer */
5792 if (wq->rescuer) { 4385 if (wq->rescuer) {
5793 struct worker *rescuer = wq-> 4386 struct worker *rescuer = wq->rescuer;
5794 4387
5795 /* this prevents new queueing 4388 /* this prevents new queueing */
5796 raw_spin_lock_irq(&wq_mayday_ 4389 raw_spin_lock_irq(&wq_mayday_lock);
5797 wq->rescuer = NULL; 4390 wq->rescuer = NULL;
5798 raw_spin_unlock_irq(&wq_mayda 4391 raw_spin_unlock_irq(&wq_mayday_lock);
5799 4392
5800 /* rescuer will empty maydays 4393 /* rescuer will empty maydays list before exiting */
5801 kthread_stop(rescuer->task); 4394 kthread_stop(rescuer->task);
5802 kfree(rescuer); 4395 kfree(rescuer);
5803 } 4396 }
5804 4397
5805 /* 4398 /*
5806 * Sanity checks - grab all the locks 4399 * Sanity checks - grab all the locks so that we wait for all
5807 * in-flight operations which may do 4400 * in-flight operations which may do put_pwq().
5808 */ 4401 */
5809 mutex_lock(&wq_pool_mutex); 4402 mutex_lock(&wq_pool_mutex);
5810 mutex_lock(&wq->mutex); 4403 mutex_lock(&wq->mutex);
5811 for_each_pwq(pwq, wq) { 4404 for_each_pwq(pwq, wq) {
5812 raw_spin_lock_irq(&pwq->pool- 4405 raw_spin_lock_irq(&pwq->pool->lock);
5813 if (WARN_ON(pwq_busy(pwq))) { 4406 if (WARN_ON(pwq_busy(pwq))) {
5814 pr_warn("%s: %s has t 4407 pr_warn("%s: %s has the following busy pwq\n",
5815 __func__, wq- 4408 __func__, wq->name);
5816 show_pwq(pwq); 4409 show_pwq(pwq);
5817 raw_spin_unlock_irq(& 4410 raw_spin_unlock_irq(&pwq->pool->lock);
5818 mutex_unlock(&wq->mut 4411 mutex_unlock(&wq->mutex);
5819 mutex_unlock(&wq_pool 4412 mutex_unlock(&wq_pool_mutex);
5820 show_one_workqueue(wq !! 4413 show_workqueue_state();
5821 return; 4414 return;
5822 } 4415 }
5823 raw_spin_unlock_irq(&pwq->poo 4416 raw_spin_unlock_irq(&pwq->pool->lock);
5824 } 4417 }
5825 mutex_unlock(&wq->mutex); 4418 mutex_unlock(&wq->mutex);
5826 4419
5827 /* 4420 /*
5828 * wq list is used to freeze wq, remo 4421 * wq list is used to freeze wq, remove from list after
5829 * flushing is complete in case freez 4422 * flushing is complete in case freeze races us.
5830 */ 4423 */
5831 list_del_rcu(&wq->list); 4424 list_del_rcu(&wq->list);
5832 mutex_unlock(&wq_pool_mutex); 4425 mutex_unlock(&wq_pool_mutex);
5833 4426
5834 /* !! 4427 if (!(wq->flags & WQ_UNBOUND)) {
5835 * We're the sole accessor of @wq. Di !! 4428 wq_unregister_lockdep(wq);
5836 * to put the base refs. @wq will be !! 4429 /*
5837 * pwq_put. RCU read lock prevents @w !! 4430 * The base ref is never dropped on per-cpu pwqs. Directly
5838 */ !! 4431 * schedule RCU free.
5839 rcu_read_lock(); !! 4432 */
>> 4433 call_rcu(&wq->rcu, rcu_free_wq);
>> 4434 } else {
>> 4435 /*
>> 4436 * We're the sole accessor of @wq at this point. Directly
>> 4437 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
>> 4438 * @wq will be freed when the last pwq is released.
>> 4439 */
>> 4440 for_each_node(node) {
>> 4441 pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
>> 4442 RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
>> 4443 put_pwq_unlocked(pwq);
>> 4444 }
5840 4445
5841 for_each_possible_cpu(cpu) { !! 4446 /*
5842 put_pwq_unlocked(unbound_pwq( !! 4447 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
5843 RCU_INIT_POINTER(*unbound_pwq !! 4448 * put. Don't access it afterwards.
>> 4449 */
>> 4450 pwq = wq->dfl_pwq;
>> 4451 wq->dfl_pwq = NULL;
>> 4452 put_pwq_unlocked(pwq);
5844 } 4453 }
5845 <<
5846 put_pwq_unlocked(unbound_pwq(wq, -1)) <<
5847 RCU_INIT_POINTER(*unbound_pwq_slot(wq <<
5848 <<
5849 rcu_read_unlock(); <<
5850 } 4454 }
5851 EXPORT_SYMBOL_GPL(destroy_workqueue); 4455 EXPORT_SYMBOL_GPL(destroy_workqueue);
5852 4456
5853 /** 4457 /**
5854 * workqueue_set_max_active - adjust max_acti 4458 * workqueue_set_max_active - adjust max_active of a workqueue
5855 * @wq: target workqueue 4459 * @wq: target workqueue
5856 * @max_active: new max_active value. 4460 * @max_active: new max_active value.
5857 * 4461 *
5858 * Set max_active of @wq to @max_active. See !! 4462 * Set max_active of @wq to @max_active.
5859 * comment. <<
5860 * 4463 *
5861 * CONTEXT: 4464 * CONTEXT:
5862 * Don't call from IRQ context. 4465 * Don't call from IRQ context.
5863 */ 4466 */
5864 void workqueue_set_max_active(struct workqueu 4467 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
5865 { 4468 {
5866 /* max_active doesn't mean anything f !! 4469 struct pool_workqueue *pwq;
5867 if (WARN_ON(wq->flags & WQ_BH)) !! 4470
5868 return; <<
5869 /* disallow meddling with max_active 4471 /* disallow meddling with max_active for ordered workqueues */
5870 if (WARN_ON(wq->flags & __WQ_ORDERED) !! 4472 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
5871 return; 4473 return;
5872 4474
5873 max_active = wq_clamp_max_active(max_ 4475 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
5874 4476
5875 mutex_lock(&wq->mutex); 4477 mutex_lock(&wq->mutex);
5876 4478
>> 4479 wq->flags &= ~__WQ_ORDERED;
5877 wq->saved_max_active = max_active; 4480 wq->saved_max_active = max_active;
5878 if (wq->flags & WQ_UNBOUND) <<
5879 wq->saved_min_active = min(wq <<
5880 4481
5881 wq_adjust_max_active(wq); !! 4482 for_each_pwq(pwq, wq)
>> 4483 pwq_adjust_max_active(pwq);
5882 4484
5883 mutex_unlock(&wq->mutex); 4485 mutex_unlock(&wq->mutex);
5884 } 4486 }
5885 EXPORT_SYMBOL_GPL(workqueue_set_max_active); 4487 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
5886 4488
5887 /** 4489 /**
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 4490 * current_work - retrieve %current task's work struct
5916 * 4491 *
5917 * Determine if %current task is a workqueue 4492 * Determine if %current task is a workqueue worker and what it's working on.
5918 * Useful to find out the context that the %c 4493 * Useful to find out the context that the %current task is running in.
5919 * 4494 *
5920 * Return: work struct if %current task is a 4495 * Return: work struct if %current task is a workqueue worker, %NULL otherwise.
5921 */ 4496 */
5922 struct work_struct *current_work(void) 4497 struct work_struct *current_work(void)
5923 { 4498 {
5924 struct worker *worker = current_wq_wo 4499 struct worker *worker = current_wq_worker();
5925 4500
5926 return worker ? worker->current_work 4501 return worker ? worker->current_work : NULL;
5927 } 4502 }
5928 EXPORT_SYMBOL(current_work); 4503 EXPORT_SYMBOL(current_work);
5929 4504
5930 /** 4505 /**
5931 * current_is_workqueue_rescuer - is %current 4506 * current_is_workqueue_rescuer - is %current workqueue rescuer?
5932 * 4507 *
5933 * Determine whether %current is a workqueue 4508 * Determine whether %current is a workqueue rescuer. Can be used from
5934 * work functions to determine whether it's b 4509 * work functions to determine whether it's being run off the rescuer task.
5935 * 4510 *
5936 * Return: %true if %current is a workqueue r 4511 * Return: %true if %current is a workqueue rescuer. %false otherwise.
5937 */ 4512 */
5938 bool current_is_workqueue_rescuer(void) 4513 bool current_is_workqueue_rescuer(void)
5939 { 4514 {
5940 struct worker *worker = current_wq_wo 4515 struct worker *worker = current_wq_worker();
5941 4516
5942 return worker && worker->rescue_wq; 4517 return worker && worker->rescue_wq;
5943 } 4518 }
5944 4519
5945 /** 4520 /**
5946 * workqueue_congested - test whether a workq 4521 * workqueue_congested - test whether a workqueue is congested
5947 * @cpu: CPU in question 4522 * @cpu: CPU in question
5948 * @wq: target workqueue 4523 * @wq: target workqueue
5949 * 4524 *
5950 * Test whether @wq's cpu workqueue for @cpu 4525 * Test whether @wq's cpu workqueue for @cpu is congested. There is
5951 * no synchronization around this function an 4526 * no synchronization around this function and the test result is
5952 * unreliable and only useful as advisory hin 4527 * unreliable and only useful as advisory hints or for debugging.
5953 * 4528 *
5954 * If @cpu is WORK_CPU_UNBOUND, the test is p 4529 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
5955 * !! 4530 * Note that both per-cpu and unbound workqueues may be associated with
5956 * With the exception of ordered workqueues, !! 4531 * multiple pool_workqueues which have separate congested states. A
5957 * pool_workqueues, each with its own congest !! 4532 * workqueue being congested on one CPU doesn't mean the workqueue is also
5958 * congested on one CPU doesn't mean that the !! 4533 * contested on other CPUs / NUMA nodes.
5959 * other CPUs. <<
5960 * 4534 *
5961 * Return: 4535 * Return:
5962 * %true if congested, %false otherwise. 4536 * %true if congested, %false otherwise.
5963 */ 4537 */
5964 bool workqueue_congested(int cpu, struct work 4538 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
5965 { 4539 {
5966 struct pool_workqueue *pwq; 4540 struct pool_workqueue *pwq;
5967 bool ret; 4541 bool ret;
5968 4542
5969 rcu_read_lock(); 4543 rcu_read_lock();
5970 preempt_disable(); 4544 preempt_disable();
5971 4545
5972 if (cpu == WORK_CPU_UNBOUND) 4546 if (cpu == WORK_CPU_UNBOUND)
5973 cpu = smp_processor_id(); 4547 cpu = smp_processor_id();
5974 4548
5975 pwq = *per_cpu_ptr(wq->cpu_pwq, cpu); !! 4549 if (!(wq->flags & WQ_UNBOUND))
5976 ret = !list_empty(&pwq->inactive_work !! 4550 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
>> 4551 else
>> 4552 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
5977 4553
>> 4554 ret = !list_empty(&pwq->delayed_works);
5978 preempt_enable(); 4555 preempt_enable();
5979 rcu_read_unlock(); 4556 rcu_read_unlock();
5980 4557
5981 return ret; 4558 return ret;
5982 } 4559 }
5983 EXPORT_SYMBOL_GPL(workqueue_congested); 4560 EXPORT_SYMBOL_GPL(workqueue_congested);
5984 4561
5985 /** 4562 /**
5986 * work_busy - test whether a work is current 4563 * work_busy - test whether a work is currently pending or running
5987 * @work: the work to be tested 4564 * @work: the work to be tested
5988 * 4565 *
5989 * Test whether @work is currently pending or 4566 * Test whether @work is currently pending or running. There is no
5990 * synchronization around this function and t 4567 * synchronization around this function and the test result is
5991 * unreliable and only useful as advisory hin 4568 * unreliable and only useful as advisory hints or for debugging.
5992 * 4569 *
5993 * Return: 4570 * Return:
5994 * OR'd bitmask of WORK_BUSY_* bits. 4571 * OR'd bitmask of WORK_BUSY_* bits.
5995 */ 4572 */
5996 unsigned int work_busy(struct work_struct *wo 4573 unsigned int work_busy(struct work_struct *work)
5997 { 4574 {
5998 struct worker_pool *pool; 4575 struct worker_pool *pool;
5999 unsigned long irq_flags; !! 4576 unsigned long flags;
6000 unsigned int ret = 0; 4577 unsigned int ret = 0;
6001 4578
6002 if (work_pending(work)) 4579 if (work_pending(work))
6003 ret |= WORK_BUSY_PENDING; 4580 ret |= WORK_BUSY_PENDING;
6004 4581
6005 rcu_read_lock(); 4582 rcu_read_lock();
6006 pool = get_work_pool(work); 4583 pool = get_work_pool(work);
6007 if (pool) { 4584 if (pool) {
6008 raw_spin_lock_irqsave(&pool-> !! 4585 raw_spin_lock_irqsave(&pool->lock, flags);
6009 if (find_worker_executing_wor 4586 if (find_worker_executing_work(pool, work))
6010 ret |= WORK_BUSY_RUNN 4587 ret |= WORK_BUSY_RUNNING;
6011 raw_spin_unlock_irqrestore(&p !! 4588 raw_spin_unlock_irqrestore(&pool->lock, flags);
6012 } 4589 }
6013 rcu_read_unlock(); 4590 rcu_read_unlock();
6014 4591
6015 return ret; 4592 return ret;
6016 } 4593 }
6017 EXPORT_SYMBOL_GPL(work_busy); 4594 EXPORT_SYMBOL_GPL(work_busy);
6018 4595
6019 /** 4596 /**
6020 * set_worker_desc - set description for the 4597 * set_worker_desc - set description for the current work item
6021 * @fmt: printf-style format string 4598 * @fmt: printf-style format string
6022 * @...: arguments for the format string 4599 * @...: arguments for the format string
6023 * 4600 *
6024 * This function can be called by a running w 4601 * This function can be called by a running work function to describe what
6025 * the work item is about. If the worker tas 4602 * the work item is about. If the worker task gets dumped, this
6026 * information will be printed out together t 4603 * information will be printed out together to help debugging. The
6027 * description can be at most WORKER_DESC_LEN 4604 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
6028 */ 4605 */
6029 void set_worker_desc(const char *fmt, ...) 4606 void set_worker_desc(const char *fmt, ...)
6030 { 4607 {
6031 struct worker *worker = current_wq_wo 4608 struct worker *worker = current_wq_worker();
6032 va_list args; 4609 va_list args;
6033 4610
6034 if (worker) { 4611 if (worker) {
6035 va_start(args, fmt); 4612 va_start(args, fmt);
6036 vsnprintf(worker->desc, sizeo 4613 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
6037 va_end(args); 4614 va_end(args);
6038 } 4615 }
6039 } 4616 }
6040 EXPORT_SYMBOL_GPL(set_worker_desc); 4617 EXPORT_SYMBOL_GPL(set_worker_desc);
6041 4618
6042 /** 4619 /**
6043 * print_worker_info - print out worker infor 4620 * print_worker_info - print out worker information and description
6044 * @log_lvl: the log level to use when printi 4621 * @log_lvl: the log level to use when printing
6045 * @task: target task 4622 * @task: target task
6046 * 4623 *
6047 * If @task is a worker and currently executi 4624 * If @task is a worker and currently executing a work item, print out the
6048 * name of the workqueue being serviced and w 4625 * name of the workqueue being serviced and worker description set with
6049 * set_worker_desc() by the currently executi 4626 * set_worker_desc() by the currently executing work item.
6050 * 4627 *
6051 * This function can be safely called on any 4628 * This function can be safely called on any task as long as the
6052 * task_struct itself is accessible. While s 4629 * task_struct itself is accessible. While safe, this function isn't
6053 * synchronized and may print out mixups or g 4630 * synchronized and may print out mixups or garbages of limited length.
6054 */ 4631 */
6055 void print_worker_info(const char *log_lvl, s 4632 void print_worker_info(const char *log_lvl, struct task_struct *task)
6056 { 4633 {
6057 work_func_t *fn = NULL; 4634 work_func_t *fn = NULL;
6058 char name[WQ_NAME_LEN] = { }; 4635 char name[WQ_NAME_LEN] = { };
6059 char desc[WORKER_DESC_LEN] = { }; 4636 char desc[WORKER_DESC_LEN] = { };
6060 struct pool_workqueue *pwq = NULL; 4637 struct pool_workqueue *pwq = NULL;
6061 struct workqueue_struct *wq = NULL; 4638 struct workqueue_struct *wq = NULL;
6062 struct worker *worker; 4639 struct worker *worker;
6063 4640
6064 if (!(task->flags & PF_WQ_WORKER)) 4641 if (!(task->flags & PF_WQ_WORKER))
6065 return; 4642 return;
6066 4643
6067 /* 4644 /*
6068 * This function is called without an 4645 * This function is called without any synchronization and @task
6069 * could be in any state. Be careful 4646 * could be in any state. Be careful with dereferences.
6070 */ 4647 */
6071 worker = kthread_probe_data(task); 4648 worker = kthread_probe_data(task);
6072 4649
6073 /* 4650 /*
6074 * Carefully copy the associated work 4651 * Carefully copy the associated workqueue's workfn, name and desc.
6075 * Keep the original last '\0' in cas 4652 * Keep the original last '\0' in case the original is garbage.
6076 */ 4653 */
6077 copy_from_kernel_nofault(&fn, &worker 4654 copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn));
6078 copy_from_kernel_nofault(&pwq, &worke 4655 copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq));
6079 copy_from_kernel_nofault(&wq, &pwq->w 4656 copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq));
6080 copy_from_kernel_nofault(name, wq->na 4657 copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1);
6081 copy_from_kernel_nofault(desc, worker 4658 copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1);
6082 4659
6083 if (fn || name[0] || desc[0]) { 4660 if (fn || name[0] || desc[0]) {
6084 printk("%sWorkqueue: %s %ps", 4661 printk("%sWorkqueue: %s %ps", log_lvl, name, fn);
6085 if (strcmp(name, desc)) 4662 if (strcmp(name, desc))
6086 pr_cont(" (%s)", desc 4663 pr_cont(" (%s)", desc);
6087 pr_cont("\n"); 4664 pr_cont("\n");
6088 } 4665 }
6089 } 4666 }
6090 4667
6091 static void pr_cont_pool_info(struct worker_p 4668 static void pr_cont_pool_info(struct worker_pool *pool)
6092 { 4669 {
6093 pr_cont(" cpus=%*pbl", nr_cpumask_bit 4670 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
6094 if (pool->node != NUMA_NO_NODE) 4671 if (pool->node != NUMA_NO_NODE)
6095 pr_cont(" node=%d", pool->nod 4672 pr_cont(" node=%d", pool->node);
6096 pr_cont(" flags=0x%x", pool->flags); !! 4673 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 } 4674 }
6115 4675
6116 struct pr_cont_work_struct { !! 4676 static void pr_cont_work(bool comma, struct work_struct *work)
6117 bool comma; <<
6118 work_func_t func; <<
6119 long ctr; <<
6120 }; <<
6121 <<
6122 static void pr_cont_work_flush(bool comma, wo <<
6123 { <<
6124 if (!pcwsp->ctr) <<
6125 goto out_record; <<
6126 if (func == pcwsp->func) { <<
6127 pcwsp->ctr++; <<
6128 return; <<
6129 } <<
6130 if (pcwsp->ctr == 1) <<
6131 pr_cont("%s %ps", pcwsp->comm <<
6132 else <<
6133 pr_cont("%s %ld*%ps", pcwsp-> <<
6134 pcwsp->ctr = 0; <<
6135 out_record: <<
6136 if ((long)func == -1L) <<
6137 return; <<
6138 pcwsp->comma = comma; <<
6139 pcwsp->func = func; <<
6140 pcwsp->ctr = 1; <<
6141 } <<
6142 <<
6143 static void pr_cont_work(bool comma, struct w <<
6144 { 4677 {
6145 if (work->func == wq_barrier_func) { 4678 if (work->func == wq_barrier_func) {
6146 struct wq_barrier *barr; 4679 struct wq_barrier *barr;
6147 4680
6148 barr = container_of(work, str 4681 barr = container_of(work, struct wq_barrier, work);
6149 4682
6150 pr_cont_work_flush(comma, (wo <<
6151 pr_cont("%s BAR(%d)", comma ? 4683 pr_cont("%s BAR(%d)", comma ? "," : "",
6152 task_pid_nr(barr->tas 4684 task_pid_nr(barr->task));
6153 } else { 4685 } else {
6154 if (!comma) !! 4686 pr_cont("%s %ps", comma ? "," : "", work->func);
6155 pr_cont_work_flush(co <<
6156 pr_cont_work_flush(comma, wor <<
6157 } 4687 }
6158 } 4688 }
6159 4689
6160 static void show_pwq(struct pool_workqueue *p 4690 static void show_pwq(struct pool_workqueue *pwq)
6161 { 4691 {
6162 struct pr_cont_work_struct pcws = { . <<
6163 struct worker_pool *pool = pwq->pool; 4692 struct worker_pool *pool = pwq->pool;
6164 struct work_struct *work; 4693 struct work_struct *work;
6165 struct worker *worker; 4694 struct worker *worker;
6166 bool has_in_flight = false, has_pendi 4695 bool has_in_flight = false, has_pending = false;
6167 int bkt; 4696 int bkt;
6168 4697
6169 pr_info(" pwq %d:", pool->id); 4698 pr_info(" pwq %d:", pool->id);
6170 pr_cont_pool_info(pool); 4699 pr_cont_pool_info(pool);
6171 4700
6172 pr_cont(" active=%d refcnt=%d%s\n", !! 4701 pr_cont(" active=%d/%d refcnt=%d%s\n",
6173 pwq->nr_active, pwq->refcnt, !! 4702 pwq->nr_active, pwq->max_active, pwq->refcnt,
6174 !list_empty(&pwq->mayday_node 4703 !list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
6175 4704
6176 hash_for_each(pool->busy_hash, bkt, w 4705 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
6177 if (worker->current_pwq == pw 4706 if (worker->current_pwq == pwq) {
6178 has_in_flight = true; 4707 has_in_flight = true;
6179 break; 4708 break;
6180 } 4709 }
6181 } 4710 }
6182 if (has_in_flight) { 4711 if (has_in_flight) {
6183 bool comma = false; 4712 bool comma = false;
6184 4713
6185 pr_info(" in-flight:"); 4714 pr_info(" in-flight:");
6186 hash_for_each(pool->busy_hash 4715 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
6187 if (worker->current_p 4716 if (worker->current_pwq != pwq)
6188 continue; 4717 continue;
6189 4718
6190 pr_cont(" %s", comma !! 4719 pr_cont("%s %d%s:%ps", comma ? "," : "",
6191 pr_cont_worker_id(wor !! 4720 task_pid_nr(worker->task),
6192 pr_cont(":%ps", worke !! 4721 worker->rescue_wq ? "(RESCUER)" : "",
>> 4722 worker->current_func);
6193 list_for_each_entry(w 4723 list_for_each_entry(work, &worker->scheduled, entry)
6194 pr_cont_work( !! 4724 pr_cont_work(false, work);
6195 pr_cont_work_flush(co <<
6196 comma = true; 4725 comma = true;
6197 } 4726 }
6198 pr_cont("\n"); 4727 pr_cont("\n");
6199 } 4728 }
6200 4729
6201 list_for_each_entry(work, &pool->work 4730 list_for_each_entry(work, &pool->worklist, entry) {
6202 if (get_work_pwq(work) == pwq 4731 if (get_work_pwq(work) == pwq) {
6203 has_pending = true; 4732 has_pending = true;
6204 break; 4733 break;
6205 } 4734 }
6206 } 4735 }
6207 if (has_pending) { 4736 if (has_pending) {
6208 bool comma = false; 4737 bool comma = false;
6209 4738
6210 pr_info(" pending:"); 4739 pr_info(" pending:");
6211 list_for_each_entry(work, &po 4740 list_for_each_entry(work, &pool->worklist, entry) {
6212 if (get_work_pwq(work 4741 if (get_work_pwq(work) != pwq)
6213 continue; 4742 continue;
6214 4743
6215 pr_cont_work(comma, w !! 4744 pr_cont_work(comma, work);
6216 comma = !(*work_data_ 4745 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
6217 } 4746 }
6218 pr_cont_work_flush(comma, (wo <<
6219 pr_cont("\n"); 4747 pr_cont("\n");
6220 } 4748 }
6221 4749
6222 if (!list_empty(&pwq->inactive_works) !! 4750 if (!list_empty(&pwq->delayed_works)) {
6223 bool comma = false; 4751 bool comma = false;
6224 4752
6225 pr_info(" inactive:"); !! 4753 pr_info(" delayed:");
6226 list_for_each_entry(work, &pw !! 4754 list_for_each_entry(work, &pwq->delayed_works, entry) {
6227 pr_cont_work(comma, w !! 4755 pr_cont_work(comma, work);
6228 comma = !(*work_data_ 4756 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
6229 } 4757 }
6230 pr_cont_work_flush(comma, (wo <<
6231 pr_cont("\n"); 4758 pr_cont("\n");
6232 } 4759 }
6233 } 4760 }
6234 4761
6235 /** 4762 /**
6236 * show_one_workqueue - dump state of specifi !! 4763 * show_workqueue_state - dump workqueue state
6237 * @wq: workqueue whose state will be printed <<
6238 */ <<
6239 void show_one_workqueue(struct workqueue_stru <<
6240 { <<
6241 struct pool_workqueue *pwq; <<
6242 bool idle = true; <<
6243 unsigned long irq_flags; <<
6244 <<
6245 for_each_pwq(pwq, wq) { <<
6246 if (!pwq_is_empty(pwq)) { <<
6247 idle = false; <<
6248 break; <<
6249 } <<
6250 } <<
6251 if (idle) /* Nothing to print for idl <<
6252 return; <<
6253 <<
6254 pr_info("workqueue %s: flags=0x%x\n", <<
6255 <<
6256 for_each_pwq(pwq, wq) { <<
6257 raw_spin_lock_irqsave(&pwq->p <<
6258 if (!pwq_is_empty(pwq)) { <<
6259 /* <<
6260 * Defer printing to <<
6261 * drivers that queue <<
6262 * also taken in thei <<
6263 */ <<
6264 printk_deferred_enter <<
6265 show_pwq(pwq); <<
6266 printk_deferred_exit( <<
6267 } <<
6268 raw_spin_unlock_irqrestore(&p <<
6269 /* <<
6270 * We could be printing a lot <<
6271 * sysrq-t -> show_all_workqu <<
6272 * hard lockup. <<
6273 */ <<
6274 touch_nmi_watchdog(); <<
6275 } <<
6276 <<
6277 } <<
6278 <<
6279 /** <<
6280 * show_one_worker_pool - dump state of speci <<
6281 * @pool: worker pool whose state will be pri <<
6282 */ <<
6283 static void show_one_worker_pool(struct worke <<
6284 { <<
6285 struct worker *worker; <<
6286 bool first = true; <<
6287 unsigned long irq_flags; <<
6288 unsigned long hung = 0; <<
6289 <<
6290 raw_spin_lock_irqsave(&pool->lock, ir <<
6291 if (pool->nr_workers == pool->nr_idle <<
6292 goto next_pool; <<
6293 <<
6294 /* How long the first pending work is <<
6295 if (!list_empty(&pool->worklist)) <<
6296 hung = jiffies_to_msecs(jiffi <<
6297 <<
6298 /* <<
6299 * Defer printing to avoid deadlocks <<
6300 * queue work while holding locks als <<
6301 * paths. <<
6302 */ <<
6303 printk_deferred_enter(); <<
6304 pr_info("pool %d:", pool->id); <<
6305 pr_cont_pool_info(pool); <<
6306 pr_cont(" hung=%lus workers=%d", hung <<
6307 if (pool->manager) <<
6308 pr_cont(" manager: %d", <<
6309 task_pid_nr(pool->man <<
6310 list_for_each_entry(worker, &pool->id <<
6311 pr_cont(" %s", first ? "idle: <<
6312 pr_cont_worker_id(worker); <<
6313 first = false; <<
6314 } <<
6315 pr_cont("\n"); <<
6316 printk_deferred_exit(); <<
6317 next_pool: <<
6318 raw_spin_unlock_irqrestore(&pool->loc <<
6319 /* <<
6320 * We could be printing a lot from at <<
6321 * sysrq-t -> show_all_workqueues(). <<
6322 * hard lockup. <<
6323 */ <<
6324 touch_nmi_watchdog(); <<
6325 <<
6326 } <<
6327 <<
6328 /** <<
6329 * show_all_workqueues - dump workqueue state <<
6330 * 4764 *
6331 * Called from a sysrq handler and prints out !! 4765 * Called from a sysrq handler or try_to_freeze_tasks() and prints out
>> 4766 * all busy workqueues and pools.
6332 */ 4767 */
6333 void show_all_workqueues(void) !! 4768 void show_workqueue_state(void)
6334 { 4769 {
6335 struct workqueue_struct *wq; 4770 struct workqueue_struct *wq;
6336 struct worker_pool *pool; 4771 struct worker_pool *pool;
>> 4772 unsigned long flags;
6337 int pi; 4773 int pi;
6338 4774
6339 rcu_read_lock(); 4775 rcu_read_lock();
6340 4776
6341 pr_info("Showing busy workqueues and 4777 pr_info("Showing busy workqueues and worker pools:\n");
6342 4778
6343 list_for_each_entry_rcu(wq, &workqueu !! 4779 list_for_each_entry_rcu(wq, &workqueues, list) {
6344 show_one_workqueue(wq); !! 4780 struct pool_workqueue *pwq;
6345 !! 4781 bool idle = true;
6346 for_each_pool(pool, pi) <<
6347 show_one_worker_pool(pool); <<
6348 4782
6349 rcu_read_unlock(); !! 4783 for_each_pwq(pwq, wq) {
6350 } !! 4784 if (pwq->nr_active || !list_empty(&pwq->delayed_works)) {
>> 4785 idle = false;
>> 4786 break;
>> 4787 }
>> 4788 }
>> 4789 if (idle)
>> 4790 continue;
6351 4791
6352 /** !! 4792 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
6353 * show_freezable_workqueues - dump freezable <<
6354 * <<
6355 * Called from try_to_freeze_tasks() and prin <<
6356 * still busy. <<
6357 */ <<
6358 void show_freezable_workqueues(void) <<
6359 { <<
6360 struct workqueue_struct *wq; <<
6361 4793
6362 rcu_read_lock(); !! 4794 for_each_pwq(pwq, wq) {
>> 4795 raw_spin_lock_irqsave(&pwq->pool->lock, flags);
>> 4796 if (pwq->nr_active || !list_empty(&pwq->delayed_works))
>> 4797 show_pwq(pwq);
>> 4798 raw_spin_unlock_irqrestore(&pwq->pool->lock, flags);
>> 4799 /*
>> 4800 * We could be printing a lot from atomic context, e.g.
>> 4801 * sysrq-t -> show_workqueue_state(). Avoid triggering
>> 4802 * hard lockup.
>> 4803 */
>> 4804 touch_nmi_watchdog();
>> 4805 }
>> 4806 }
6363 4807
6364 pr_info("Showing freezable workqueues !! 4808 for_each_pool(pool, pi) {
>> 4809 struct worker *worker;
>> 4810 bool first = true;
6365 4811
6366 list_for_each_entry_rcu(wq, &workqueu !! 4812 raw_spin_lock_irqsave(&pool->lock, flags);
6367 if (!(wq->flags & WQ_FREEZABL !! 4813 if (pool->nr_workers == pool->nr_idle)
6368 continue; !! 4814 goto next_pool;
6369 show_one_workqueue(wq); !! 4815
>> 4816 pr_info("pool %d:", pool->id);
>> 4817 pr_cont_pool_info(pool);
>> 4818 pr_cont(" hung=%us workers=%d",
>> 4819 jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000,
>> 4820 pool->nr_workers);
>> 4821 if (pool->manager)
>> 4822 pr_cont(" manager: %d",
>> 4823 task_pid_nr(pool->manager->task));
>> 4824 list_for_each_entry(worker, &pool->idle_list, entry) {
>> 4825 pr_cont(" %s%d", first ? "idle: " : "",
>> 4826 task_pid_nr(worker->task));
>> 4827 first = false;
>> 4828 }
>> 4829 pr_cont("\n");
>> 4830 next_pool:
>> 4831 raw_spin_unlock_irqrestore(&pool->lock, flags);
>> 4832 /*
>> 4833 * We could be printing a lot from atomic context, e.g.
>> 4834 * sysrq-t -> show_workqueue_state(). Avoid triggering
>> 4835 * hard lockup.
>> 4836 */
>> 4837 touch_nmi_watchdog();
6370 } 4838 }
6371 4839
6372 rcu_read_unlock(); 4840 rcu_read_unlock();
6373 } 4841 }
6374 4842
6375 /* used to show worker information through /p 4843 /* used to show worker information through /proc/PID/{comm,stat,status} */
6376 void wq_worker_comm(char *buf, size_t size, s 4844 void wq_worker_comm(char *buf, size_t size, struct task_struct *task)
6377 { 4845 {
>> 4846 int off;
>> 4847
>> 4848 /* always show the actual comm */
>> 4849 off = strscpy(buf, task->comm, size);
>> 4850 if (off < 0)
>> 4851 return;
>> 4852
6378 /* stabilize PF_WQ_WORKER and worker 4853 /* stabilize PF_WQ_WORKER and worker pool association */
6379 mutex_lock(&wq_pool_attach_mutex); 4854 mutex_lock(&wq_pool_attach_mutex);
6380 4855
6381 if (task->flags & PF_WQ_WORKER) { 4856 if (task->flags & PF_WQ_WORKER) {
6382 struct worker *worker = kthre 4857 struct worker *worker = kthread_data(task);
6383 struct worker_pool *pool = wo 4858 struct worker_pool *pool = worker->pool;
6384 int off; <<
6385 <<
6386 off = format_worker_id(buf, s <<
6387 4859
6388 if (pool) { 4860 if (pool) {
6389 raw_spin_lock_irq(&po 4861 raw_spin_lock_irq(&pool->lock);
6390 /* 4862 /*
6391 * ->desc tracks info 4863 * ->desc tracks information (wq name or
6392 * set_worker_desc()) 4864 * set_worker_desc()) for the latest execution. If
6393 * current, prepend ' 4865 * current, prepend '+', otherwise '-'.
6394 */ 4866 */
6395 if (worker->desc[0] ! 4867 if (worker->desc[0] != '\0') {
6396 if (worker->c 4868 if (worker->current_work)
6397 scnpr 4869 scnprintf(buf + off, size - off, "+%s",
6398 4870 worker->desc);
6399 else 4871 else
6400 scnpr 4872 scnprintf(buf + off, size - off, "-%s",
6401 4873 worker->desc);
6402 } 4874 }
6403 raw_spin_unlock_irq(& 4875 raw_spin_unlock_irq(&pool->lock);
6404 } 4876 }
6405 } else { <<
6406 strscpy(buf, task->comm, size <<
6407 } 4877 }
6408 4878
6409 mutex_unlock(&wq_pool_attach_mutex); 4879 mutex_unlock(&wq_pool_attach_mutex);
6410 } 4880 }
6411 4881
6412 #ifdef CONFIG_SMP 4882 #ifdef CONFIG_SMP
6413 4883
6414 /* 4884 /*
6415 * CPU hotplug. 4885 * CPU hotplug.
6416 * 4886 *
6417 * There are two challenges in supporting CPU 4887 * There are two challenges in supporting CPU hotplug. Firstly, there
6418 * are a lot of assumptions on strong associa 4888 * are a lot of assumptions on strong associations among work, pwq and
6419 * pool which make migrating pending and sche 4889 * pool which make migrating pending and scheduled works very
6420 * difficult to implement without impacting h 4890 * difficult to implement without impacting hot paths. Secondly,
6421 * worker pools serve mix of short, long and 4891 * worker pools serve mix of short, long and very long running works making
6422 * blocked draining impractical. 4892 * blocked draining impractical.
6423 * 4893 *
6424 * This is solved by allowing the pools to be 4894 * This is solved by allowing the pools to be disassociated from the CPU
6425 * running as an unbound one and allowing it 4895 * running as an unbound one and allowing it to be reattached later if the
6426 * cpu comes back online. 4896 * cpu comes back online.
6427 */ 4897 */
6428 4898
6429 static void unbind_workers(int cpu) 4899 static void unbind_workers(int cpu)
6430 { 4900 {
6431 struct worker_pool *pool; 4901 struct worker_pool *pool;
6432 struct worker *worker; 4902 struct worker *worker;
6433 4903
6434 for_each_cpu_worker_pool(pool, cpu) { 4904 for_each_cpu_worker_pool(pool, cpu) {
6435 mutex_lock(&wq_pool_attach_mu 4905 mutex_lock(&wq_pool_attach_mutex);
6436 raw_spin_lock_irq(&pool->lock 4906 raw_spin_lock_irq(&pool->lock);
6437 4907
6438 /* 4908 /*
6439 * We've blocked all attach/d 4909 * We've blocked all attach/detach operations. Make all workers
6440 * unbound and set DISASSOCIA 4910 * unbound and set DISASSOCIATED. Before this, all workers
6441 * must be on the cpu. After !! 4911 * except for the ones which are still executing works from
6442 * And the preemption disable !! 4912 * before the last CPU down must be on the cpu. After
6443 * are guaranteed to see WORK !! 4913 * this, they may become diasporas.
6444 * is on the same cpu. <<
6445 */ 4914 */
6446 for_each_pool_worker(worker, 4915 for_each_pool_worker(worker, pool)
6447 worker->flags |= WORK 4916 worker->flags |= WORKER_UNBOUND;
6448 4917
6449 pool->flags |= POOL_DISASSOCI 4918 pool->flags |= POOL_DISASSOCIATED;
6450 4919
>> 4920 raw_spin_unlock_irq(&pool->lock);
>> 4921
>> 4922 for_each_pool_worker(worker, pool) {
>> 4923 kthread_set_per_cpu(worker->task, -1);
>> 4924 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0);
>> 4925 }
>> 4926
>> 4927 mutex_unlock(&wq_pool_attach_mutex);
>> 4928
6451 /* 4929 /*
6452 * The handling of nr_running !! 4930 * Call schedule() so that we cross rq->lock and thus can
6453 * now. Zap nr_running. Aft !! 4931 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
6454 * need_more_worker() and kee !! 4932 * This is necessary as scheduler callbacks may be invoked
6455 * long as the worklist is no !! 4933 * from other cpus.
6456 * an unbound (in terms of co !! 4934 */
>> 4935 schedule();
>> 4936
>> 4937 /*
>> 4938 * Sched callbacks are disabled now. Zap nr_running.
>> 4939 * After this, nr_running stays zero and need_more_worker()
>> 4940 * and keep_working() are always true as long as the
>> 4941 * worklist is not empty. This pool now behaves as an
>> 4942 * unbound (in terms of concurrency management) pool which
6457 * are served by workers tied 4943 * are served by workers tied to the pool.
6458 */ 4944 */
6459 pool->nr_running = 0; !! 4945 atomic_set(&pool->nr_running, 0);
6460 4946
6461 /* 4947 /*
6462 * With concurrency managemen 4948 * With concurrency management just turned off, a busy
6463 * worker blocking could lead 4949 * worker blocking could lead to lengthy stalls. Kick off
6464 * unbound chain execution of 4950 * unbound chain execution of currently pending work items.
6465 */ 4951 */
6466 kick_pool(pool); !! 4952 raw_spin_lock_irq(&pool->lock);
6467 !! 4953 wake_up_worker(pool);
6468 raw_spin_unlock_irq(&pool->lo 4954 raw_spin_unlock_irq(&pool->lock);
6469 <<
6470 for_each_pool_worker(worker, <<
6471 unbind_worker(worker) <<
6472 <<
6473 mutex_unlock(&wq_pool_attach_ <<
6474 } 4955 }
6475 } 4956 }
6476 4957
6477 /** 4958 /**
6478 * rebind_workers - rebind all workers of a p 4959 * rebind_workers - rebind all workers of a pool to the associated CPU
6479 * @pool: pool of interest 4960 * @pool: pool of interest
6480 * 4961 *
6481 * @pool->cpu is coming online. Rebind all w 4962 * @pool->cpu is coming online. Rebind all workers to the CPU.
6482 */ 4963 */
6483 static void rebind_workers(struct worker_pool 4964 static void rebind_workers(struct worker_pool *pool)
6484 { 4965 {
6485 struct worker *worker; 4966 struct worker *worker;
6486 4967
6487 lockdep_assert_held(&wq_pool_attach_m 4968 lockdep_assert_held(&wq_pool_attach_mutex);
6488 4969
6489 /* 4970 /*
6490 * Restore CPU affinity of all worker 4971 * Restore CPU affinity of all workers. As all idle workers should
6491 * be on the run-queue of the associa 4972 * be on the run-queue of the associated CPU before any local
6492 * wake-ups for concurrency managemen 4973 * wake-ups for concurrency management happen, restore CPU affinity
6493 * of all workers first and then clea 4974 * of all workers first and then clear UNBOUND. As we're called
6494 * from CPU_ONLINE, the following sho 4975 * from CPU_ONLINE, the following shouldn't fail.
6495 */ 4976 */
6496 for_each_pool_worker(worker, pool) { 4977 for_each_pool_worker(worker, pool) {
6497 kthread_set_per_cpu(worker->t 4978 kthread_set_per_cpu(worker->task, pool->cpu);
6498 WARN_ON_ONCE(set_cpus_allowed 4979 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
6499 !! 4980 pool->attrs->cpumask) < 0);
6500 } 4981 }
6501 4982
6502 raw_spin_lock_irq(&pool->lock); 4983 raw_spin_lock_irq(&pool->lock);
6503 4984
6504 pool->flags &= ~POOL_DISASSOCIATED; 4985 pool->flags &= ~POOL_DISASSOCIATED;
6505 4986
6506 for_each_pool_worker(worker, pool) { 4987 for_each_pool_worker(worker, pool) {
6507 unsigned int worker_flags = w 4988 unsigned int worker_flags = worker->flags;
6508 4989
6509 /* 4990 /*
>> 4991 * A bound idle worker should actually be on the runqueue
>> 4992 * of the associated CPU for local wake-ups targeting it to
>> 4993 * work. Kick all idle workers so that they migrate to the
>> 4994 * associated CPU. Doing this in the same loop as
>> 4995 * replacing UNBOUND with REBOUND is safe as no worker will
>> 4996 * be bound before @pool->lock is released.
>> 4997 */
>> 4998 if (worker_flags & WORKER_IDLE)
>> 4999 wake_up_process(worker->task);
>> 5000
>> 5001 /*
6510 * We want to clear UNBOUND b 5002 * We want to clear UNBOUND but can't directly call
6511 * worker_clr_flags() or adju 5003 * worker_clr_flags() or adjust nr_running. Atomically
6512 * replace UNBOUND with anoth 5004 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
6513 * @worker will clear REBOUND 5005 * @worker will clear REBOUND using worker_clr_flags() when
6514 * it initiates the next exec 5006 * it initiates the next execution cycle thus restoring
6515 * concurrency management. N 5007 * concurrency management. Note that when or whether
6516 * @worker clears REBOUND doe 5008 * @worker clears REBOUND doesn't affect correctness.
6517 * 5009 *
6518 * WRITE_ONCE() is necessary 5010 * WRITE_ONCE() is necessary because @worker->flags may be
6519 * tested without holding any 5011 * tested without holding any lock in
6520 * wq_worker_running(). With 5012 * wq_worker_running(). Without it, NOT_RUNNING test may
6521 * fail incorrectly leading t 5013 * fail incorrectly leading to premature concurrency
6522 * management operations. 5014 * management operations.
6523 */ 5015 */
6524 WARN_ON_ONCE(!(worker_flags & 5016 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
6525 worker_flags |= WORKER_REBOUN 5017 worker_flags |= WORKER_REBOUND;
6526 worker_flags &= ~WORKER_UNBOU 5018 worker_flags &= ~WORKER_UNBOUND;
6527 WRITE_ONCE(worker->flags, wor 5019 WRITE_ONCE(worker->flags, worker_flags);
6528 } 5020 }
6529 5021
6530 raw_spin_unlock_irq(&pool->lock); 5022 raw_spin_unlock_irq(&pool->lock);
6531 } 5023 }
6532 5024
6533 /** 5025 /**
6534 * restore_unbound_workers_cpumask - restore 5026 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
6535 * @pool: unbound pool of interest 5027 * @pool: unbound pool of interest
6536 * @cpu: the CPU which is coming up 5028 * @cpu: the CPU which is coming up
6537 * 5029 *
6538 * An unbound pool may end up with a cpumask 5030 * 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 5031 * CPUs. When a worker of such pool get scheduled, the scheduler resets
6540 * its cpus_allowed. If @cpu is in @pool's c 5032 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
6541 * online CPU before, cpus_allowed of all its 5033 * online CPU before, cpus_allowed of all its workers should be restored.
6542 */ 5034 */
6543 static void restore_unbound_workers_cpumask(s 5035 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
6544 { 5036 {
6545 static cpumask_t cpumask; 5037 static cpumask_t cpumask;
6546 struct worker *worker; 5038 struct worker *worker;
6547 5039
6548 lockdep_assert_held(&wq_pool_attach_m 5040 lockdep_assert_held(&wq_pool_attach_mutex);
6549 5041
6550 /* is @cpu allowed for @pool? */ 5042 /* is @cpu allowed for @pool? */
6551 if (!cpumask_test_cpu(cpu, pool->attr 5043 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
6552 return; 5044 return;
6553 5045
6554 cpumask_and(&cpumask, pool->attrs->cp 5046 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
6555 5047
6556 /* as we're called from CPU_ONLINE, t 5048 /* as we're called from CPU_ONLINE, the following shouldn't fail */
6557 for_each_pool_worker(worker, pool) 5049 for_each_pool_worker(worker, pool)
6558 WARN_ON_ONCE(set_cpus_allowed 5050 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0);
6559 } 5051 }
6560 5052
6561 int workqueue_prepare_cpu(unsigned int cpu) 5053 int workqueue_prepare_cpu(unsigned int cpu)
6562 { 5054 {
6563 struct worker_pool *pool; 5055 struct worker_pool *pool;
6564 5056
6565 for_each_cpu_worker_pool(pool, cpu) { 5057 for_each_cpu_worker_pool(pool, cpu) {
6566 if (pool->nr_workers) 5058 if (pool->nr_workers)
6567 continue; 5059 continue;
6568 if (!create_worker(pool)) 5060 if (!create_worker(pool))
6569 return -ENOMEM; 5061 return -ENOMEM;
6570 } 5062 }
6571 return 0; 5063 return 0;
6572 } 5064 }
6573 5065
6574 int workqueue_online_cpu(unsigned int cpu) 5066 int workqueue_online_cpu(unsigned int cpu)
6575 { 5067 {
6576 struct worker_pool *pool; 5068 struct worker_pool *pool;
6577 struct workqueue_struct *wq; 5069 struct workqueue_struct *wq;
6578 int pi; 5070 int pi;
6579 5071
6580 mutex_lock(&wq_pool_mutex); 5072 mutex_lock(&wq_pool_mutex);
6581 5073
6582 cpumask_set_cpu(cpu, wq_online_cpumas <<
6583 <<
6584 for_each_pool(pool, pi) { 5074 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 5075 mutex_lock(&wq_pool_attach_mutex);
>> 5076
6590 if (pool->cpu == cpu) 5077 if (pool->cpu == cpu)
6591 rebind_workers(pool); 5078 rebind_workers(pool);
6592 else if (pool->cpu < 0) 5079 else if (pool->cpu < 0)
6593 restore_unbound_worke 5080 restore_unbound_workers_cpumask(pool, cpu);
>> 5081
6594 mutex_unlock(&wq_pool_attach_ 5082 mutex_unlock(&wq_pool_attach_mutex);
6595 } 5083 }
6596 5084
6597 /* update pod affinity of unbound wor !! 5085 /* update NUMA affinity of unbound workqueues */
6598 list_for_each_entry(wq, &workqueues, !! 5086 list_for_each_entry(wq, &workqueues, list)
6599 struct workqueue_attrs *attrs !! 5087 wq_update_unbound_numa(wq, cpu, true);
6600 <<
6601 if (attrs) { <<
6602 const struct wq_pod_t <<
6603 int tcpu; <<
6604 <<
6605 for_each_cpu(tcpu, pt <<
6606 unbound_wq_up <<
6607 <<
6608 mutex_lock(&wq->mutex <<
6609 wq_update_node_max_ac <<
6610 mutex_unlock(&wq->mut <<
6611 } <<
6612 } <<
6613 5088
6614 mutex_unlock(&wq_pool_mutex); 5089 mutex_unlock(&wq_pool_mutex);
6615 return 0; 5090 return 0;
6616 } 5091 }
6617 5092
6618 int workqueue_offline_cpu(unsigned int cpu) 5093 int workqueue_offline_cpu(unsigned int cpu)
6619 { 5094 {
6620 struct workqueue_struct *wq; 5095 struct workqueue_struct *wq;
6621 5096
6622 /* unbinding per-cpu workers should h 5097 /* unbinding per-cpu workers should happen on the local CPU */
6623 if (WARN_ON(cpu != smp_processor_id() 5098 if (WARN_ON(cpu != smp_processor_id()))
6624 return -1; 5099 return -1;
6625 5100
6626 unbind_workers(cpu); 5101 unbind_workers(cpu);
6627 5102
6628 /* update pod affinity of unbound wor !! 5103 /* update NUMA affinity of unbound workqueues */
6629 mutex_lock(&wq_pool_mutex); 5104 mutex_lock(&wq_pool_mutex);
6630 !! 5105 list_for_each_entry(wq, &workqueues, list)
6631 cpumask_clear_cpu(cpu, wq_online_cpum !! 5106 wq_update_unbound_numa(wq, cpu, false);
6632 <<
6633 list_for_each_entry(wq, &workqueues, <<
6634 struct workqueue_attrs *attrs <<
6635 <<
6636 if (attrs) { <<
6637 const struct wq_pod_t <<
6638 int tcpu; <<
6639 <<
6640 for_each_cpu(tcpu, pt <<
6641 unbound_wq_up <<
6642 <<
6643 mutex_lock(&wq->mutex <<
6644 wq_update_node_max_ac <<
6645 mutex_unlock(&wq->mut <<
6646 } <<
6647 } <<
6648 mutex_unlock(&wq_pool_mutex); 5107 mutex_unlock(&wq_pool_mutex);
6649 5108
6650 return 0; 5109 return 0;
6651 } 5110 }
6652 5111
6653 struct work_for_cpu { 5112 struct work_for_cpu {
6654 struct work_struct work; 5113 struct work_struct work;
6655 long (*fn)(void *); 5114 long (*fn)(void *);
6656 void *arg; 5115 void *arg;
6657 long ret; 5116 long ret;
6658 }; 5117 };
6659 5118
6660 static void work_for_cpu_fn(struct work_struc 5119 static void work_for_cpu_fn(struct work_struct *work)
6661 { 5120 {
6662 struct work_for_cpu *wfc = container_ 5121 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
6663 5122
6664 wfc->ret = wfc->fn(wfc->arg); 5123 wfc->ret = wfc->fn(wfc->arg);
6665 } 5124 }
6666 5125
6667 /** 5126 /**
6668 * work_on_cpu_key - run a function in thread !! 5127 * work_on_cpu - run a function in thread context on a particular cpu
6669 * @cpu: the cpu to run on 5128 * @cpu: the cpu to run on
6670 * @fn: the function to run 5129 * @fn: the function to run
6671 * @arg: the function arg 5130 * @arg: the function arg
6672 * @key: The lock class key for lock debuggin <<
6673 * 5131 *
6674 * It is up to the caller to ensure that the 5132 * 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 5133 * The caller must not hold any locks which would prevent @fn from completing.
6676 * 5134 *
6677 * Return: The value @fn returns. 5135 * Return: The value @fn returns.
6678 */ 5136 */
6679 long work_on_cpu_key(int cpu, long (*fn)(void !! 5137 long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
6680 void *arg, struct lock_c <<
6681 { 5138 {
6682 struct work_for_cpu wfc = { .fn = fn, 5139 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
6683 5140
6684 INIT_WORK_ONSTACK_KEY(&wfc.work, work !! 5141 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
6685 schedule_work_on(cpu, &wfc.work); 5142 schedule_work_on(cpu, &wfc.work);
6686 flush_work(&wfc.work); 5143 flush_work(&wfc.work);
6687 destroy_work_on_stack(&wfc.work); 5144 destroy_work_on_stack(&wfc.work);
6688 return wfc.ret; 5145 return wfc.ret;
6689 } 5146 }
6690 EXPORT_SYMBOL_GPL(work_on_cpu_key); !! 5147 EXPORT_SYMBOL_GPL(work_on_cpu);
6691 5148
6692 /** 5149 /**
6693 * work_on_cpu_safe_key - run a function in t !! 5150 * work_on_cpu_safe - run a function in thread context on a particular cpu
6694 * @cpu: the cpu to run on 5151 * @cpu: the cpu to run on
6695 * @fn: the function to run 5152 * @fn: the function to run
6696 * @arg: the function argument 5153 * @arg: the function argument
6697 * @key: The lock class key for lock debuggin <<
6698 * 5154 *
6699 * Disables CPU hotplug and calls work_on_cpu 5155 * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold
6700 * any locks which would prevent @fn from com 5156 * any locks which would prevent @fn from completing.
6701 * 5157 *
6702 * Return: The value @fn returns. 5158 * Return: The value @fn returns.
6703 */ 5159 */
6704 long work_on_cpu_safe_key(int cpu, long (*fn) !! 5160 long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg)
6705 void *arg, struct l <<
6706 { 5161 {
6707 long ret = -ENODEV; 5162 long ret = -ENODEV;
6708 5163
6709 cpus_read_lock(); !! 5164 get_online_cpus();
6710 if (cpu_online(cpu)) 5165 if (cpu_online(cpu))
6711 ret = work_on_cpu_key(cpu, fn !! 5166 ret = work_on_cpu(cpu, fn, arg);
6712 cpus_read_unlock(); !! 5167 put_online_cpus();
6713 return ret; 5168 return ret;
6714 } 5169 }
6715 EXPORT_SYMBOL_GPL(work_on_cpu_safe_key); !! 5170 EXPORT_SYMBOL_GPL(work_on_cpu_safe);
6716 #endif /* CONFIG_SMP */ 5171 #endif /* CONFIG_SMP */
6717 5172
6718 #ifdef CONFIG_FREEZER 5173 #ifdef CONFIG_FREEZER
6719 5174
6720 /** 5175 /**
6721 * freeze_workqueues_begin - begin freezing w 5176 * freeze_workqueues_begin - begin freezing workqueues
6722 * 5177 *
6723 * Start freezing workqueues. After this fun 5178 * Start freezing workqueues. After this function returns, all freezable
6724 * workqueues will queue new works to their i !! 5179 * workqueues will queue new works to their delayed_works list instead of
6725 * pool->worklist. 5180 * pool->worklist.
6726 * 5181 *
6727 * CONTEXT: 5182 * CONTEXT:
6728 * Grabs and releases wq_pool_mutex, wq->mute 5183 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
6729 */ 5184 */
6730 void freeze_workqueues_begin(void) 5185 void freeze_workqueues_begin(void)
6731 { 5186 {
6732 struct workqueue_struct *wq; 5187 struct workqueue_struct *wq;
>> 5188 struct pool_workqueue *pwq;
6733 5189
6734 mutex_lock(&wq_pool_mutex); 5190 mutex_lock(&wq_pool_mutex);
6735 5191
6736 WARN_ON_ONCE(workqueue_freezing); 5192 WARN_ON_ONCE(workqueue_freezing);
6737 workqueue_freezing = true; 5193 workqueue_freezing = true;
6738 5194
6739 list_for_each_entry(wq, &workqueues, 5195 list_for_each_entry(wq, &workqueues, list) {
6740 mutex_lock(&wq->mutex); 5196 mutex_lock(&wq->mutex);
6741 wq_adjust_max_active(wq); !! 5197 for_each_pwq(pwq, wq)
>> 5198 pwq_adjust_max_active(pwq);
6742 mutex_unlock(&wq->mutex); 5199 mutex_unlock(&wq->mutex);
6743 } 5200 }
6744 5201
6745 mutex_unlock(&wq_pool_mutex); 5202 mutex_unlock(&wq_pool_mutex);
6746 } 5203 }
6747 5204
6748 /** 5205 /**
6749 * freeze_workqueues_busy - are freezable wor 5206 * freeze_workqueues_busy - are freezable workqueues still busy?
6750 * 5207 *
6751 * Check whether freezing is complete. This 5208 * Check whether freezing is complete. This function must be called
6752 * between freeze_workqueues_begin() and thaw 5209 * between freeze_workqueues_begin() and thaw_workqueues().
6753 * 5210 *
6754 * CONTEXT: 5211 * CONTEXT:
6755 * Grabs and releases wq_pool_mutex. 5212 * Grabs and releases wq_pool_mutex.
6756 * 5213 *
6757 * Return: 5214 * Return:
6758 * %true if some freezable workqueues are sti 5215 * %true if some freezable workqueues are still busy. %false if freezing
6759 * is complete. 5216 * is complete.
6760 */ 5217 */
6761 bool freeze_workqueues_busy(void) 5218 bool freeze_workqueues_busy(void)
6762 { 5219 {
6763 bool busy = false; 5220 bool busy = false;
6764 struct workqueue_struct *wq; 5221 struct workqueue_struct *wq;
6765 struct pool_workqueue *pwq; 5222 struct pool_workqueue *pwq;
6766 5223
6767 mutex_lock(&wq_pool_mutex); 5224 mutex_lock(&wq_pool_mutex);
6768 5225
6769 WARN_ON_ONCE(!workqueue_freezing); 5226 WARN_ON_ONCE(!workqueue_freezing);
6770 5227
6771 list_for_each_entry(wq, &workqueues, 5228 list_for_each_entry(wq, &workqueues, list) {
6772 if (!(wq->flags & WQ_FREEZABL 5229 if (!(wq->flags & WQ_FREEZABLE))
6773 continue; 5230 continue;
6774 /* 5231 /*
6775 * nr_active is monotonically 5232 * nr_active is monotonically decreasing. It's safe
6776 * to peek without lock. 5233 * to peek without lock.
6777 */ 5234 */
6778 rcu_read_lock(); 5235 rcu_read_lock();
6779 for_each_pwq(pwq, wq) { 5236 for_each_pwq(pwq, wq) {
6780 WARN_ON_ONCE(pwq->nr_ 5237 WARN_ON_ONCE(pwq->nr_active < 0);
6781 if (pwq->nr_active) { 5238 if (pwq->nr_active) {
6782 busy = true; 5239 busy = true;
6783 rcu_read_unlo 5240 rcu_read_unlock();
6784 goto out_unlo 5241 goto out_unlock;
6785 } 5242 }
6786 } 5243 }
6787 rcu_read_unlock(); 5244 rcu_read_unlock();
6788 } 5245 }
6789 out_unlock: 5246 out_unlock:
6790 mutex_unlock(&wq_pool_mutex); 5247 mutex_unlock(&wq_pool_mutex);
6791 return busy; 5248 return busy;
6792 } 5249 }
6793 5250
6794 /** 5251 /**
6795 * thaw_workqueues - thaw workqueues 5252 * thaw_workqueues - thaw workqueues
6796 * 5253 *
6797 * Thaw workqueues. Normal queueing is resto 5254 * Thaw workqueues. Normal queueing is restored and all collected
6798 * frozen works are transferred to their resp 5255 * frozen works are transferred to their respective pool worklists.
6799 * 5256 *
6800 * CONTEXT: 5257 * CONTEXT:
6801 * Grabs and releases wq_pool_mutex, wq->mute 5258 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
6802 */ 5259 */
6803 void thaw_workqueues(void) 5260 void thaw_workqueues(void)
6804 { 5261 {
6805 struct workqueue_struct *wq; 5262 struct workqueue_struct *wq;
>> 5263 struct pool_workqueue *pwq;
6806 5264
6807 mutex_lock(&wq_pool_mutex); 5265 mutex_lock(&wq_pool_mutex);
6808 5266
6809 if (!workqueue_freezing) 5267 if (!workqueue_freezing)
6810 goto out_unlock; 5268 goto out_unlock;
6811 5269
6812 workqueue_freezing = false; 5270 workqueue_freezing = false;
6813 5271
6814 /* restore max_active and repopulate 5272 /* restore max_active and repopulate worklist */
6815 list_for_each_entry(wq, &workqueues, 5273 list_for_each_entry(wq, &workqueues, list) {
6816 mutex_lock(&wq->mutex); 5274 mutex_lock(&wq->mutex);
6817 wq_adjust_max_active(wq); !! 5275 for_each_pwq(pwq, wq)
>> 5276 pwq_adjust_max_active(pwq);
6818 mutex_unlock(&wq->mutex); 5277 mutex_unlock(&wq->mutex);
6819 } 5278 }
6820 5279
6821 out_unlock: 5280 out_unlock:
6822 mutex_unlock(&wq_pool_mutex); 5281 mutex_unlock(&wq_pool_mutex);
6823 } 5282 }
6824 #endif /* CONFIG_FREEZER */ 5283 #endif /* CONFIG_FREEZER */
6825 5284
6826 static int workqueue_apply_unbound_cpumask(co !! 5285 static int workqueue_apply_unbound_cpumask(void)
6827 { 5286 {
6828 LIST_HEAD(ctxs); 5287 LIST_HEAD(ctxs);
6829 int ret = 0; 5288 int ret = 0;
6830 struct workqueue_struct *wq; 5289 struct workqueue_struct *wq;
6831 struct apply_wqattrs_ctx *ctx, *n; 5290 struct apply_wqattrs_ctx *ctx, *n;
6832 5291
6833 lockdep_assert_held(&wq_pool_mutex); 5292 lockdep_assert_held(&wq_pool_mutex);
6834 5293
6835 list_for_each_entry(wq, &workqueues, 5294 list_for_each_entry(wq, &workqueues, list) {
6836 if (!(wq->flags & WQ_UNBOUND) !! 5295 if (!(wq->flags & WQ_UNBOUND))
>> 5296 continue;
>> 5297 /* creating multiple pwqs breaks ordering guarantee */
>> 5298 if (wq->flags & __WQ_ORDERED)
6837 continue; 5299 continue;
6838 5300
6839 ctx = apply_wqattrs_prepare(w !! 5301 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs);
6840 if (IS_ERR(ctx)) { !! 5302 if (!ctx) {
6841 ret = PTR_ERR(ctx); !! 5303 ret = -ENOMEM;
6842 break; 5304 break;
6843 } 5305 }
6844 5306
6845 list_add_tail(&ctx->list, &ct 5307 list_add_tail(&ctx->list, &ctxs);
6846 } 5308 }
6847 5309
6848 list_for_each_entry_safe(ctx, n, &ctx 5310 list_for_each_entry_safe(ctx, n, &ctxs, list) {
6849 if (!ret) 5311 if (!ret)
6850 apply_wqattrs_commit( 5312 apply_wqattrs_commit(ctx);
6851 apply_wqattrs_cleanup(ctx); 5313 apply_wqattrs_cleanup(ctx);
6852 } 5314 }
6853 5315
6854 if (!ret) { <<
6855 mutex_lock(&wq_pool_attach_mu <<
6856 cpumask_copy(wq_unbound_cpuma <<
6857 mutex_unlock(&wq_pool_attach_ <<
6858 } <<
6859 return ret; 5316 return ret;
6860 } 5317 }
6861 5318
6862 /** 5319 /**
6863 * workqueue_unbound_exclude_cpumask - Exclud !! 5320 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
6864 * @exclude_cpumask: the cpumask to be exclud !! 5321 * @cpumask: the cpumask to set
6865 * 5322 *
6866 * This function can be called from cpuset co !! 5323 * The low-level workqueues cpumask is a global cpumask that limits
6867 * CPUs that should be excluded from wq_unbou !! 5324 * the affinity of all unbound workqueues. This function check the @cpumask
>> 5325 * and apply it to all unbound workqueues and updates all pwqs of them.
>> 5326 *
>> 5327 * Retun: 0 - Success
>> 5328 * -EINVAL - Invalid @cpumask
>> 5329 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
6868 */ 5330 */
6869 int workqueue_unbound_exclude_cpumask(cpumask !! 5331 int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
6870 { 5332 {
6871 cpumask_var_t cpumask; !! 5333 int ret = -EINVAL;
6872 int ret = 0; !! 5334 cpumask_var_t saved_cpumask;
6873 5335
6874 if (!zalloc_cpumask_var(&cpumask, GFP !! 5336 if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL))
6875 return -ENOMEM; 5337 return -ENOMEM;
6876 5338
6877 mutex_lock(&wq_pool_mutex); <<
6878 <<
6879 /* 5339 /*
6880 * If the operation fails, it will fa !! 5340 * Not excluding isolated cpus on purpose.
6881 * wq_requested_unbound_cpumask which !! 5341 * If the user wishes to include them, we allow that.
6882 * (HK_TYPE_WQ ∩ HK_TYPE_DOMAIN) ho !! 5342 */
6883 * by any subsequent write to workque !! 5343 cpumask_and(cpumask, cpumask, cpu_possible_mask);
6884 */ !! 5344 if (!cpumask_empty(cpumask)) {
6885 if (!cpumask_andnot(cpumask, wq_reque !! 5345 apply_wqattrs_lock();
6886 cpumask_copy(cpumask, wq_requ <<
6887 if (!cpumask_equal(cpumask, wq_unboun <<
6888 ret = workqueue_apply_unbound <<
6889 <<
6890 /* Save the current isolated cpumask <<
6891 if (!ret) <<
6892 cpumask_copy(wq_isolated_cpum <<
6893 <<
6894 mutex_unlock(&wq_pool_mutex); <<
6895 free_cpumask_var(cpumask); <<
6896 return ret; <<
6897 } <<
6898 <<
6899 static int parse_affn_scope(const char *val) <<
6900 { <<
6901 int i; <<
6902 <<
6903 for (i = 0; i < ARRAY_SIZE(wq_affn_na <<
6904 if (!strncasecmp(val, wq_affn <<
6905 return i; <<
6906 } <<
6907 return -EINVAL; <<
6908 } <<
6909 <<
6910 static int wq_affn_dfl_set(const char *val, c <<
6911 { <<
6912 struct workqueue_struct *wq; <<
6913 int affn, cpu; <<
6914 <<
6915 affn = parse_affn_scope(val); <<
6916 if (affn < 0) <<
6917 return affn; <<
6918 if (affn == WQ_AFFN_DFL) <<
6919 return -EINVAL; <<
6920 5346
6921 cpus_read_lock(); !! 5347 /* save the old wq_unbound_cpumask. */
6922 mutex_lock(&wq_pool_mutex); !! 5348 cpumask_copy(saved_cpumask, wq_unbound_cpumask);
6923 5349
6924 wq_affn_dfl = affn; !! 5350 /* update wq_unbound_cpumask at first and apply it to wqs. */
>> 5351 cpumask_copy(wq_unbound_cpumask, cpumask);
>> 5352 ret = workqueue_apply_unbound_cpumask();
>> 5353
>> 5354 /* restore the wq_unbound_cpumask when failed. */
>> 5355 if (ret < 0)
>> 5356 cpumask_copy(wq_unbound_cpumask, saved_cpumask);
6925 5357
6926 list_for_each_entry(wq, &workqueues, !! 5358 apply_wqattrs_unlock();
6927 for_each_online_cpu(cpu) <<
6928 unbound_wq_update_pwq <<
6929 } 5359 }
6930 5360
6931 mutex_unlock(&wq_pool_mutex); !! 5361 free_cpumask_var(saved_cpumask);
6932 cpus_read_unlock(); !! 5362 return ret;
6933 <<
6934 return 0; <<
6935 } <<
6936 <<
6937 static int wq_affn_dfl_get(char *buffer, cons <<
6938 { <<
6939 return scnprintf(buffer, PAGE_SIZE, " <<
6940 } 5363 }
6941 5364
6942 static const struct kernel_param_ops wq_affn_ <<
6943 .set = wq_affn_dfl_set, <<
6944 .get = wq_affn_dfl_get, <<
6945 }; <<
6946 <<
6947 module_param_cb(default_affinity_scope, &wq_a <<
6948 <<
6949 #ifdef CONFIG_SYSFS 5365 #ifdef CONFIG_SYSFS
6950 /* 5366 /*
6951 * Workqueues with WQ_SYSFS flag set is visib 5367 * Workqueues with WQ_SYSFS flag set is visible to userland via
6952 * /sys/bus/workqueue/devices/WQ_NAME. All v 5368 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
6953 * following attributes. 5369 * following attributes.
6954 * 5370 *
6955 * per_cpu RO bool : whether the !! 5371 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
6956 * max_active RW int : maximum num !! 5372 * max_active RW int : maximum number of in-flight work items
6957 * 5373 *
6958 * Unbound workqueues have the following extr 5374 * Unbound workqueues have the following extra attributes.
6959 * 5375 *
6960 * nice RW int : nice value !! 5376 * pool_ids RO int : the associated pool IDs for each node
6961 * cpumask RW mask : bitmask of !! 5377 * nice RW int : nice value of the workers
6962 * affinity_scope RW str : worker CPU !! 5378 * cpumask RW mask : bitmask of allowed CPUs for the workers
6963 * affinity_strict RW bool : worker CPU !! 5379 * numa RW bool : whether enable NUMA affinity
6964 */ 5380 */
6965 struct wq_device { 5381 struct wq_device {
6966 struct workqueue_struct *wq; 5382 struct workqueue_struct *wq;
6967 struct device dev; 5383 struct device dev;
6968 }; 5384 };
6969 5385
6970 static struct workqueue_struct *dev_to_wq(str 5386 static struct workqueue_struct *dev_to_wq(struct device *dev)
6971 { 5387 {
6972 struct wq_device *wq_dev = container_ 5388 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
6973 5389
6974 return wq_dev->wq; 5390 return wq_dev->wq;
6975 } 5391 }
6976 5392
6977 static ssize_t per_cpu_show(struct device *de 5393 static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
6978 char *buf) 5394 char *buf)
6979 { 5395 {
6980 struct workqueue_struct *wq = dev_to_ 5396 struct workqueue_struct *wq = dev_to_wq(dev);
6981 5397
6982 return scnprintf(buf, PAGE_SIZE, "%d\ 5398 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
6983 } 5399 }
6984 static DEVICE_ATTR_RO(per_cpu); 5400 static DEVICE_ATTR_RO(per_cpu);
6985 5401
6986 static ssize_t max_active_show(struct device 5402 static ssize_t max_active_show(struct device *dev,
6987 struct device_ 5403 struct device_attribute *attr, char *buf)
6988 { 5404 {
6989 struct workqueue_struct *wq = dev_to_ 5405 struct workqueue_struct *wq = dev_to_wq(dev);
6990 5406
6991 return scnprintf(buf, PAGE_SIZE, "%d\ 5407 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
6992 } 5408 }
6993 5409
6994 static ssize_t max_active_store(struct device 5410 static ssize_t max_active_store(struct device *dev,
6995 struct device 5411 struct device_attribute *attr, const char *buf,
6996 size_t count) 5412 size_t count)
6997 { 5413 {
6998 struct workqueue_struct *wq = dev_to_ 5414 struct workqueue_struct *wq = dev_to_wq(dev);
6999 int val; 5415 int val;
7000 5416
7001 if (sscanf(buf, "%d", &val) != 1 || v 5417 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
7002 return -EINVAL; 5418 return -EINVAL;
7003 5419
7004 workqueue_set_max_active(wq, val); 5420 workqueue_set_max_active(wq, val);
7005 return count; 5421 return count;
7006 } 5422 }
7007 static DEVICE_ATTR_RW(max_active); 5423 static DEVICE_ATTR_RW(max_active);
7008 5424
7009 static struct attribute *wq_sysfs_attrs[] = { 5425 static struct attribute *wq_sysfs_attrs[] = {
7010 &dev_attr_per_cpu.attr, 5426 &dev_attr_per_cpu.attr,
7011 &dev_attr_max_active.attr, 5427 &dev_attr_max_active.attr,
7012 NULL, 5428 NULL,
7013 }; 5429 };
7014 ATTRIBUTE_GROUPS(wq_sysfs); 5430 ATTRIBUTE_GROUPS(wq_sysfs);
7015 5431
>> 5432 static ssize_t wq_pool_ids_show(struct device *dev,
>> 5433 struct device_attribute *attr, char *buf)
>> 5434 {
>> 5435 struct workqueue_struct *wq = dev_to_wq(dev);
>> 5436 const char *delim = "";
>> 5437 int node, written = 0;
>> 5438
>> 5439 get_online_cpus();
>> 5440 rcu_read_lock();
>> 5441 for_each_node(node) {
>> 5442 written += scnprintf(buf + written, PAGE_SIZE - written,
>> 5443 "%s%d:%d", delim, node,
>> 5444 unbound_pwq_by_node(wq, node)->pool->id);
>> 5445 delim = " ";
>> 5446 }
>> 5447 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
>> 5448 rcu_read_unlock();
>> 5449 put_online_cpus();
>> 5450
>> 5451 return written;
>> 5452 }
>> 5453
7016 static ssize_t wq_nice_show(struct device *de 5454 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
7017 char *buf) 5455 char *buf)
7018 { 5456 {
7019 struct workqueue_struct *wq = dev_to_ 5457 struct workqueue_struct *wq = dev_to_wq(dev);
7020 int written; 5458 int written;
7021 5459
7022 mutex_lock(&wq->mutex); 5460 mutex_lock(&wq->mutex);
7023 written = scnprintf(buf, PAGE_SIZE, " 5461 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
7024 mutex_unlock(&wq->mutex); 5462 mutex_unlock(&wq->mutex);
7025 5463
7026 return written; 5464 return written;
7027 } 5465 }
7028 5466
7029 /* prepare workqueue_attrs for sysfs store op 5467 /* prepare workqueue_attrs for sysfs store operations */
7030 static struct workqueue_attrs *wq_sysfs_prep_ 5468 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
7031 { 5469 {
7032 struct workqueue_attrs *attrs; 5470 struct workqueue_attrs *attrs;
7033 5471
7034 lockdep_assert_held(&wq_pool_mutex); 5472 lockdep_assert_held(&wq_pool_mutex);
7035 5473
7036 attrs = alloc_workqueue_attrs(); 5474 attrs = alloc_workqueue_attrs();
7037 if (!attrs) 5475 if (!attrs)
7038 return NULL; 5476 return NULL;
7039 5477
7040 copy_workqueue_attrs(attrs, wq->unbou 5478 copy_workqueue_attrs(attrs, wq->unbound_attrs);
7041 return attrs; 5479 return attrs;
7042 } 5480 }
7043 5481
7044 static ssize_t wq_nice_store(struct device *d 5482 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
7045 const char *buf, 5483 const char *buf, size_t count)
7046 { 5484 {
7047 struct workqueue_struct *wq = dev_to_ 5485 struct workqueue_struct *wq = dev_to_wq(dev);
7048 struct workqueue_attrs *attrs; 5486 struct workqueue_attrs *attrs;
7049 int ret = -ENOMEM; 5487 int ret = -ENOMEM;
7050 5488
7051 apply_wqattrs_lock(); 5489 apply_wqattrs_lock();
7052 5490
7053 attrs = wq_sysfs_prep_attrs(wq); 5491 attrs = wq_sysfs_prep_attrs(wq);
7054 if (!attrs) 5492 if (!attrs)
7055 goto out_unlock; 5493 goto out_unlock;
7056 5494
7057 if (sscanf(buf, "%d", &attrs->nice) = 5495 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
7058 attrs->nice >= MIN_NICE && attrs- 5496 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
7059 ret = apply_workqueue_attrs_l 5497 ret = apply_workqueue_attrs_locked(wq, attrs);
7060 else 5498 else
7061 ret = -EINVAL; 5499 ret = -EINVAL;
7062 5500
7063 out_unlock: 5501 out_unlock:
7064 apply_wqattrs_unlock(); 5502 apply_wqattrs_unlock();
7065 free_workqueue_attrs(attrs); 5503 free_workqueue_attrs(attrs);
7066 return ret ?: count; 5504 return ret ?: count;
7067 } 5505 }
7068 5506
7069 static ssize_t wq_cpumask_show(struct device 5507 static ssize_t wq_cpumask_show(struct device *dev,
7070 struct device_ 5508 struct device_attribute *attr, char *buf)
7071 { 5509 {
7072 struct workqueue_struct *wq = dev_to_ 5510 struct workqueue_struct *wq = dev_to_wq(dev);
7073 int written; 5511 int written;
7074 5512
7075 mutex_lock(&wq->mutex); 5513 mutex_lock(&wq->mutex);
7076 written = scnprintf(buf, PAGE_SIZE, " 5514 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
7077 cpumask_pr_args(w 5515 cpumask_pr_args(wq->unbound_attrs->cpumask));
7078 mutex_unlock(&wq->mutex); 5516 mutex_unlock(&wq->mutex);
7079 return written; 5517 return written;
7080 } 5518 }
7081 5519
7082 static ssize_t wq_cpumask_store(struct device 5520 static ssize_t wq_cpumask_store(struct device *dev,
7083 struct device 5521 struct device_attribute *attr,
7084 const char *b 5522 const char *buf, size_t count)
7085 { 5523 {
7086 struct workqueue_struct *wq = dev_to_ 5524 struct workqueue_struct *wq = dev_to_wq(dev);
7087 struct workqueue_attrs *attrs; 5525 struct workqueue_attrs *attrs;
7088 int ret = -ENOMEM; 5526 int ret = -ENOMEM;
7089 5527
7090 apply_wqattrs_lock(); 5528 apply_wqattrs_lock();
7091 5529
7092 attrs = wq_sysfs_prep_attrs(wq); 5530 attrs = wq_sysfs_prep_attrs(wq);
7093 if (!attrs) 5531 if (!attrs)
7094 goto out_unlock; 5532 goto out_unlock;
7095 5533
7096 ret = cpumask_parse(buf, attrs->cpuma 5534 ret = cpumask_parse(buf, attrs->cpumask);
7097 if (!ret) 5535 if (!ret)
7098 ret = apply_workqueue_attrs_l 5536 ret = apply_workqueue_attrs_locked(wq, attrs);
7099 5537
7100 out_unlock: 5538 out_unlock:
7101 apply_wqattrs_unlock(); 5539 apply_wqattrs_unlock();
7102 free_workqueue_attrs(attrs); 5540 free_workqueue_attrs(attrs);
7103 return ret ?: count; 5541 return ret ?: count;
7104 } 5542 }
7105 5543
7106 static ssize_t wq_affn_scope_show(struct devi !! 5544 static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
7107 struct devi !! 5545 char *buf)
7108 { 5546 {
7109 struct workqueue_struct *wq = dev_to_ 5547 struct workqueue_struct *wq = dev_to_wq(dev);
7110 int written; 5548 int written;
7111 5549
7112 mutex_lock(&wq->mutex); 5550 mutex_lock(&wq->mutex);
7113 if (wq->unbound_attrs->affn_scope == !! 5551 written = scnprintf(buf, PAGE_SIZE, "%d\n",
7114 written = scnprintf(buf, PAGE !! 5552 !wq->unbound_attrs->no_numa);
7115 wq_affn_n <<
7116 wq_affn_n <<
7117 else <<
7118 written = scnprintf(buf, PAGE <<
7119 wq_affn_n <<
7120 mutex_unlock(&wq->mutex); 5553 mutex_unlock(&wq->mutex);
7121 5554
7122 return written; 5555 return written;
7123 } 5556 }
7124 5557
7125 static ssize_t wq_affn_scope_store(struct dev !! 5558 static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
7126 struct dev !! 5559 const char *buf, size_t count)
7127 const char <<
7128 { <<
7129 struct workqueue_struct *wq = dev_to_ <<
7130 struct workqueue_attrs *attrs; <<
7131 int affn, ret = -ENOMEM; <<
7132 <<
7133 affn = parse_affn_scope(buf); <<
7134 if (affn < 0) <<
7135 return affn; <<
7136 <<
7137 apply_wqattrs_lock(); <<
7138 attrs = wq_sysfs_prep_attrs(wq); <<
7139 if (attrs) { <<
7140 attrs->affn_scope = affn; <<
7141 ret = apply_workqueue_attrs_l <<
7142 } <<
7143 apply_wqattrs_unlock(); <<
7144 free_workqueue_attrs(attrs); <<
7145 return ret ?: count; <<
7146 } <<
7147 <<
7148 static ssize_t wq_affinity_strict_show(struct <<
7149 struct <<
7150 { <<
7151 struct workqueue_struct *wq = dev_to_ <<
7152 <<
7153 return scnprintf(buf, PAGE_SIZE, "%d\ <<
7154 wq->unbound_attrs->a <<
7155 } <<
7156 <<
7157 static ssize_t wq_affinity_strict_store(struc <<
7158 struc <<
7159 const <<
7160 { 5560 {
7161 struct workqueue_struct *wq = dev_to_ 5561 struct workqueue_struct *wq = dev_to_wq(dev);
7162 struct workqueue_attrs *attrs; 5562 struct workqueue_attrs *attrs;
7163 int v, ret = -ENOMEM; 5563 int v, ret = -ENOMEM;
7164 5564
7165 if (sscanf(buf, "%d", &v) != 1) <<
7166 return -EINVAL; <<
7167 <<
7168 apply_wqattrs_lock(); 5565 apply_wqattrs_lock();
>> 5566
7169 attrs = wq_sysfs_prep_attrs(wq); 5567 attrs = wq_sysfs_prep_attrs(wq);
7170 if (attrs) { !! 5568 if (!attrs)
7171 attrs->affn_strict = (bool)v; !! 5569 goto out_unlock;
>> 5570
>> 5571 ret = -EINVAL;
>> 5572 if (sscanf(buf, "%d", &v) == 1) {
>> 5573 attrs->no_numa = !v;
7172 ret = apply_workqueue_attrs_l 5574 ret = apply_workqueue_attrs_locked(wq, attrs);
7173 } 5575 }
>> 5576
>> 5577 out_unlock:
7174 apply_wqattrs_unlock(); 5578 apply_wqattrs_unlock();
7175 free_workqueue_attrs(attrs); 5579 free_workqueue_attrs(attrs);
7176 return ret ?: count; 5580 return ret ?: count;
7177 } 5581 }
7178 5582
7179 static struct device_attribute wq_sysfs_unbou 5583 static struct device_attribute wq_sysfs_unbound_attrs[] = {
>> 5584 __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
7180 __ATTR(nice, 0644, wq_nice_show, wq_n 5585 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
7181 __ATTR(cpumask, 0644, wq_cpumask_show 5586 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
7182 __ATTR(affinity_scope, 0644, wq_affn_ !! 5587 __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
7183 __ATTR(affinity_strict, 0644, wq_affi <<
7184 __ATTR_NULL, 5588 __ATTR_NULL,
7185 }; 5589 };
7186 5590
7187 static const struct bus_type wq_subsys = { !! 5591 static struct bus_type wq_subsys = {
7188 .name = "wo 5592 .name = "workqueue",
7189 .dev_groups = wq_ 5593 .dev_groups = wq_sysfs_groups,
7190 }; 5594 };
7191 5595
7192 /** !! 5596 static ssize_t wq_unbound_cpumask_show(struct device *dev,
7193 * workqueue_set_unbound_cpumask - Set the l !! 5597 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 { 5598 {
7229 int written; 5599 int written;
7230 5600
7231 mutex_lock(&wq_pool_mutex); 5601 mutex_lock(&wq_pool_mutex);
7232 written = scnprintf(buf, PAGE_SIZE, " !! 5602 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
>> 5603 cpumask_pr_args(wq_unbound_cpumask));
7233 mutex_unlock(&wq_pool_mutex); 5604 mutex_unlock(&wq_pool_mutex);
7234 5605
7235 return written; 5606 return written;
7236 } 5607 }
7237 5608
7238 static ssize_t cpumask_requested_show(struct !! 5609 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 5610 struct device_attribute *attr, const char *buf, size_t count)
7260 { 5611 {
7261 cpumask_var_t cpumask; 5612 cpumask_var_t cpumask;
7262 int ret; 5613 int ret;
7263 5614
7264 if (!zalloc_cpumask_var(&cpumask, GFP 5615 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
7265 return -ENOMEM; 5616 return -ENOMEM;
7266 5617
7267 ret = cpumask_parse(buf, cpumask); 5618 ret = cpumask_parse(buf, cpumask);
7268 if (!ret) 5619 if (!ret)
7269 ret = workqueue_set_unbound_c 5620 ret = workqueue_set_unbound_cpumask(cpumask);
7270 5621
7271 free_cpumask_var(cpumask); 5622 free_cpumask_var(cpumask);
7272 return ret ? ret : count; 5623 return ret ? ret : count;
7273 } 5624 }
7274 static DEVICE_ATTR_RW(cpumask); <<
7275 5625
7276 static struct attribute *wq_sysfs_cpumask_att !! 5626 static struct device_attribute wq_sysfs_cpumask_attr =
7277 &dev_attr_cpumask.attr, !! 5627 __ATTR(cpumask, 0644, wq_unbound_cpumask_show,
7278 &dev_attr_cpumask_requested.attr, !! 5628 wq_unbound_cpumask_store);
7279 &dev_attr_cpumask_isolated.attr, <<
7280 NULL, <<
7281 }; <<
7282 ATTRIBUTE_GROUPS(wq_sysfs_cpumask); <<
7283 5629
7284 static int __init wq_sysfs_init(void) 5630 static int __init wq_sysfs_init(void)
7285 { 5631 {
7286 return subsys_virtual_register(&wq_su !! 5632 int err;
>> 5633
>> 5634 err = subsys_virtual_register(&wq_subsys, NULL);
>> 5635 if (err)
>> 5636 return err;
>> 5637
>> 5638 return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr);
7287 } 5639 }
7288 core_initcall(wq_sysfs_init); 5640 core_initcall(wq_sysfs_init);
7289 5641
7290 static void wq_device_release(struct device * 5642 static void wq_device_release(struct device *dev)
7291 { 5643 {
7292 struct wq_device *wq_dev = container_ 5644 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
7293 5645
7294 kfree(wq_dev); 5646 kfree(wq_dev);
7295 } 5647 }
7296 5648
7297 /** 5649 /**
7298 * workqueue_sysfs_register - make a workqueu 5650 * workqueue_sysfs_register - make a workqueue visible in sysfs
7299 * @wq: the workqueue to register 5651 * @wq: the workqueue to register
7300 * 5652 *
7301 * Expose @wq in sysfs under /sys/bus/workque 5653 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
7302 * alloc_workqueue*() automatically calls thi 5654 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
7303 * which is the preferred method. 5655 * which is the preferred method.
7304 * 5656 *
7305 * Workqueue user should use this function di 5657 * Workqueue user should use this function directly iff it wants to apply
7306 * workqueue_attrs before making the workqueu 5658 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
7307 * apply_workqueue_attrs() may race against u 5659 * apply_workqueue_attrs() may race against userland updating the
7308 * attributes. 5660 * attributes.
7309 * 5661 *
7310 * Return: 0 on success, -errno on failure. 5662 * Return: 0 on success, -errno on failure.
7311 */ 5663 */
7312 int workqueue_sysfs_register(struct workqueue 5664 int workqueue_sysfs_register(struct workqueue_struct *wq)
7313 { 5665 {
7314 struct wq_device *wq_dev; 5666 struct wq_device *wq_dev;
7315 int ret; 5667 int ret;
7316 5668
7317 /* 5669 /*
7318 * Adjusting max_active breaks orderi !! 5670 * Adjusting max_active or creating new pwqs by applying
7319 * ordered workqueues. !! 5671 * attributes breaks ordering guarantee. Disallow exposing ordered
>> 5672 * workqueues.
7320 */ 5673 */
7321 if (WARN_ON(wq->flags & __WQ_ORDERED) !! 5674 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
7322 return -EINVAL; 5675 return -EINVAL;
7323 5676
7324 wq->wq_dev = wq_dev = kzalloc(sizeof( 5677 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
7325 if (!wq_dev) 5678 if (!wq_dev)
7326 return -ENOMEM; 5679 return -ENOMEM;
7327 5680
7328 wq_dev->wq = wq; 5681 wq_dev->wq = wq;
7329 wq_dev->dev.bus = &wq_subsys; 5682 wq_dev->dev.bus = &wq_subsys;
7330 wq_dev->dev.release = wq_device_relea 5683 wq_dev->dev.release = wq_device_release;
7331 dev_set_name(&wq_dev->dev, "%s", wq-> 5684 dev_set_name(&wq_dev->dev, "%s", wq->name);
7332 5685
7333 /* 5686 /*
7334 * unbound_attrs are created separate 5687 * unbound_attrs are created separately. Suppress uevent until
7335 * everything is ready. 5688 * everything is ready.
7336 */ 5689 */
7337 dev_set_uevent_suppress(&wq_dev->dev, 5690 dev_set_uevent_suppress(&wq_dev->dev, true);
7338 5691
7339 ret = device_register(&wq_dev->dev); 5692 ret = device_register(&wq_dev->dev);
7340 if (ret) { 5693 if (ret) {
7341 put_device(&wq_dev->dev); 5694 put_device(&wq_dev->dev);
7342 wq->wq_dev = NULL; 5695 wq->wq_dev = NULL;
7343 return ret; 5696 return ret;
7344 } 5697 }
7345 5698
7346 if (wq->flags & WQ_UNBOUND) { 5699 if (wq->flags & WQ_UNBOUND) {
7347 struct device_attribute *attr 5700 struct device_attribute *attr;
7348 5701
7349 for (attr = wq_sysfs_unbound_ 5702 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
7350 ret = device_create_f 5703 ret = device_create_file(&wq_dev->dev, attr);
7351 if (ret) { 5704 if (ret) {
7352 device_unregi 5705 device_unregister(&wq_dev->dev);
7353 wq->wq_dev = 5706 wq->wq_dev = NULL;
7354 return ret; 5707 return ret;
7355 } 5708 }
7356 } 5709 }
7357 } 5710 }
7358 5711
7359 dev_set_uevent_suppress(&wq_dev->dev, 5712 dev_set_uevent_suppress(&wq_dev->dev, false);
7360 kobject_uevent(&wq_dev->dev.kobj, KOB 5713 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
7361 return 0; 5714 return 0;
7362 } 5715 }
7363 5716
7364 /** 5717 /**
7365 * workqueue_sysfs_unregister - undo workqueu 5718 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
7366 * @wq: the workqueue to unregister 5719 * @wq: the workqueue to unregister
7367 * 5720 *
7368 * If @wq is registered to sysfs by workqueue 5721 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
7369 */ 5722 */
7370 static void workqueue_sysfs_unregister(struct 5723 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
7371 { 5724 {
7372 struct wq_device *wq_dev = wq->wq_dev 5725 struct wq_device *wq_dev = wq->wq_dev;
7373 5726
7374 if (!wq->wq_dev) 5727 if (!wq->wq_dev)
7375 return; 5728 return;
7376 5729
7377 wq->wq_dev = NULL; 5730 wq->wq_dev = NULL;
7378 device_unregister(&wq_dev->dev); 5731 device_unregister(&wq_dev->dev);
7379 } 5732 }
7380 #else /* CONFIG_SYSFS */ 5733 #else /* CONFIG_SYSFS */
7381 static void workqueue_sysfs_unregister(struct 5734 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
7382 #endif /* CONFIG_SYSFS */ 5735 #endif /* CONFIG_SYSFS */
7383 5736
7384 /* 5737 /*
7385 * Workqueue watchdog. 5738 * Workqueue watchdog.
7386 * 5739 *
7387 * Stall may be caused by various bugs - miss 5740 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
7388 * flush dependency, a concurrency managed wo 5741 * flush dependency, a concurrency managed work item which stays RUNNING
7389 * indefinitely. Workqueue stalls can be ver 5742 * indefinitely. Workqueue stalls can be very difficult to debug as the
7390 * usual warning mechanisms don't trigger and 5743 * usual warning mechanisms don't trigger and internal workqueue state is
7391 * largely opaque. 5744 * largely opaque.
7392 * 5745 *
7393 * Workqueue watchdog monitors all worker poo 5746 * Workqueue watchdog monitors all worker pools periodically and dumps
7394 * state if some pools failed to make forward 5747 * state if some pools failed to make forward progress for a while where
7395 * forward progress is defined as the first i 5748 * forward progress is defined as the first item on ->worklist changing.
7396 * 5749 *
7397 * This mechanism is controlled through the k 5750 * This mechanism is controlled through the kernel parameter
7398 * "workqueue.watchdog_thresh" which can be u 5751 * "workqueue.watchdog_thresh" which can be updated at runtime through the
7399 * corresponding sysfs parameter file. 5752 * corresponding sysfs parameter file.
7400 */ 5753 */
7401 #ifdef CONFIG_WQ_WATCHDOG 5754 #ifdef CONFIG_WQ_WATCHDOG
7402 5755
7403 static unsigned long wq_watchdog_thresh = 30; 5756 static unsigned long wq_watchdog_thresh = 30;
7404 static struct timer_list wq_watchdog_timer; 5757 static struct timer_list wq_watchdog_timer;
7405 5758
7406 static unsigned long wq_watchdog_touched = IN 5759 static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
7407 static DEFINE_PER_CPU(unsigned long, wq_watch 5760 static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
7408 5761
7409 /* <<
7410 * Show workers that might prevent the proces <<
7411 * The only candidates are CPU-bound workers <<
7412 * Pending work items should be handled by an <<
7413 * in all other situations. <<
7414 */ <<
7415 static void show_cpu_pool_hog(struct worker_p <<
7416 { <<
7417 struct worker *worker; <<
7418 unsigned long irq_flags; <<
7419 int bkt; <<
7420 <<
7421 raw_spin_lock_irqsave(&pool->lock, ir <<
7422 <<
7423 hash_for_each(pool->busy_hash, bkt, w <<
7424 if (task_is_running(worker->t <<
7425 /* <<
7426 * Defer printing to <<
7427 * drivers that queue <<
7428 * also taken in thei <<
7429 */ <<
7430 printk_deferred_enter <<
7431 <<
7432 pr_info("pool %d:\n", <<
7433 sched_show_task(worke <<
7434 <<
7435 printk_deferred_exit( <<
7436 } <<
7437 } <<
7438 <<
7439 raw_spin_unlock_irqrestore(&pool->loc <<
7440 } <<
7441 <<
7442 static void show_cpu_pools_hogs(void) <<
7443 { <<
7444 struct worker_pool *pool; <<
7445 int pi; <<
7446 <<
7447 pr_info("Showing backtraces of runnin <<
7448 <<
7449 rcu_read_lock(); <<
7450 <<
7451 for_each_pool(pool, pi) { <<
7452 if (pool->cpu_stall) <<
7453 show_cpu_pool_hog(poo <<
7454 <<
7455 } <<
7456 <<
7457 rcu_read_unlock(); <<
7458 } <<
7459 <<
7460 static void wq_watchdog_reset_touched(void) 5762 static void wq_watchdog_reset_touched(void)
7461 { 5763 {
7462 int cpu; 5764 int cpu;
7463 5765
7464 wq_watchdog_touched = jiffies; 5766 wq_watchdog_touched = jiffies;
7465 for_each_possible_cpu(cpu) 5767 for_each_possible_cpu(cpu)
7466 per_cpu(wq_watchdog_touched_c 5768 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
7467 } 5769 }
7468 5770
7469 static void wq_watchdog_timer_fn(struct timer 5771 static void wq_watchdog_timer_fn(struct timer_list *unused)
7470 { 5772 {
7471 unsigned long thresh = READ_ONCE(wq_w 5773 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
7472 bool lockup_detected = false; 5774 bool lockup_detected = false;
7473 bool cpu_pool_stall = false; <<
7474 unsigned long now = jiffies; <<
7475 struct worker_pool *pool; 5775 struct worker_pool *pool;
7476 int pi; 5776 int pi;
7477 5777
7478 if (!thresh) 5778 if (!thresh)
7479 return; 5779 return;
7480 5780
7481 rcu_read_lock(); 5781 rcu_read_lock();
7482 5782
7483 for_each_pool(pool, pi) { 5783 for_each_pool(pool, pi) {
7484 unsigned long pool_ts, touche 5784 unsigned long pool_ts, touched, ts;
7485 5785
7486 pool->cpu_stall = false; <<
7487 if (list_empty(&pool->worklis 5786 if (list_empty(&pool->worklist))
7488 continue; 5787 continue;
7489 5788
7490 /* <<
7491 * If a virtual machine is st <<
7492 * the watchdog like a stall. <<
7493 */ <<
7494 kvm_check_and_clear_guest_pau <<
7495 <<
7496 /* get the latest of pool and 5789 /* get the latest of pool and touched timestamps */
7497 if (pool->cpu >= 0) <<
7498 touched = READ_ONCE(p <<
7499 else <<
7500 touched = READ_ONCE(w <<
7501 pool_ts = READ_ONCE(pool->wat 5790 pool_ts = READ_ONCE(pool->watchdog_ts);
>> 5791 touched = READ_ONCE(wq_watchdog_touched);
7502 5792
7503 if (time_after(pool_ts, touch 5793 if (time_after(pool_ts, touched))
7504 ts = pool_ts; 5794 ts = pool_ts;
7505 else 5795 else
7506 ts = touched; 5796 ts = touched;
7507 5797
>> 5798 if (pool->cpu >= 0) {
>> 5799 unsigned long cpu_touched =
>> 5800 READ_ONCE(per_cpu(wq_watchdog_touched_cpu,
>> 5801 pool->cpu));
>> 5802 if (time_after(cpu_touched, ts))
>> 5803 ts = cpu_touched;
>> 5804 }
>> 5805
7508 /* did we stall? */ 5806 /* did we stall? */
7509 if (time_after(now, ts + thre !! 5807 if (time_after(jiffies, ts + thresh)) {
7510 lockup_detected = tru 5808 lockup_detected = true;
7511 if (pool->cpu >= 0 && <<
7512 pool->cpu_sta <<
7513 cpu_pool_stal <<
7514 } <<
7515 pr_emerg("BUG: workqu 5809 pr_emerg("BUG: workqueue lockup - pool");
7516 pr_cont_pool_info(poo 5810 pr_cont_pool_info(pool);
7517 pr_cont(" stuck for % 5811 pr_cont(" stuck for %us!\n",
7518 jiffies_to_ms !! 5812 jiffies_to_msecs(jiffies - pool_ts) / 1000);
7519 } 5813 }
7520 <<
7521 <<
7522 } 5814 }
7523 5815
7524 rcu_read_unlock(); 5816 rcu_read_unlock();
7525 5817
7526 if (lockup_detected) 5818 if (lockup_detected)
7527 show_all_workqueues(); !! 5819 show_workqueue_state();
7528 <<
7529 if (cpu_pool_stall) <<
7530 show_cpu_pools_hogs(); <<
7531 5820
7532 wq_watchdog_reset_touched(); 5821 wq_watchdog_reset_touched();
7533 mod_timer(&wq_watchdog_timer, jiffies 5822 mod_timer(&wq_watchdog_timer, jiffies + thresh);
7534 } 5823 }
7535 5824
7536 notrace void wq_watchdog_touch(int cpu) 5825 notrace void wq_watchdog_touch(int cpu)
7537 { 5826 {
7538 unsigned long thresh = READ_ONCE(wq_w <<
7539 unsigned long touch_ts = READ_ONCE(wq <<
7540 unsigned long now = jiffies; <<
7541 <<
7542 if (cpu >= 0) 5827 if (cpu >= 0)
7543 per_cpu(wq_watchdog_touched_c !! 5828 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
7544 else 5829 else
7545 WARN_ONCE(1, "%s should be ca !! 5830 wq_watchdog_touched = jiffies;
7546 <<
7547 /* Don't unnecessarily store to globa <<
7548 if (time_after(now, touch_ts + thresh <<
7549 WRITE_ONCE(wq_watchdog_touche <<
7550 } 5831 }
7551 5832
7552 static void wq_watchdog_set_thresh(unsigned l 5833 static void wq_watchdog_set_thresh(unsigned long thresh)
7553 { 5834 {
7554 wq_watchdog_thresh = 0; 5835 wq_watchdog_thresh = 0;
7555 del_timer_sync(&wq_watchdog_timer); 5836 del_timer_sync(&wq_watchdog_timer);
7556 5837
7557 if (thresh) { 5838 if (thresh) {
7558 wq_watchdog_thresh = thresh; 5839 wq_watchdog_thresh = thresh;
7559 wq_watchdog_reset_touched(); 5840 wq_watchdog_reset_touched();
7560 mod_timer(&wq_watchdog_timer, 5841 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
7561 } 5842 }
7562 } 5843 }
7563 5844
7564 static int wq_watchdog_param_set_thresh(const 5845 static int wq_watchdog_param_set_thresh(const char *val,
7565 const 5846 const struct kernel_param *kp)
7566 { 5847 {
7567 unsigned long thresh; 5848 unsigned long thresh;
7568 int ret; 5849 int ret;
7569 5850
7570 ret = kstrtoul(val, 0, &thresh); 5851 ret = kstrtoul(val, 0, &thresh);
7571 if (ret) 5852 if (ret)
7572 return ret; 5853 return ret;
7573 5854
7574 if (system_wq) 5855 if (system_wq)
7575 wq_watchdog_set_thresh(thresh 5856 wq_watchdog_set_thresh(thresh);
7576 else 5857 else
7577 wq_watchdog_thresh = thresh; 5858 wq_watchdog_thresh = thresh;
7578 5859
7579 return 0; 5860 return 0;
7580 } 5861 }
7581 5862
7582 static const struct kernel_param_ops wq_watch 5863 static const struct kernel_param_ops wq_watchdog_thresh_ops = {
7583 .set = wq_watchdog_param_set_thres 5864 .set = wq_watchdog_param_set_thresh,
7584 .get = param_get_ulong, 5865 .get = param_get_ulong,
7585 }; 5866 };
7586 5867
7587 module_param_cb(watchdog_thresh, &wq_watchdog 5868 module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
7588 0644); 5869 0644);
7589 5870
7590 static void wq_watchdog_init(void) 5871 static void wq_watchdog_init(void)
7591 { 5872 {
7592 timer_setup(&wq_watchdog_timer, wq_wa 5873 timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE);
7593 wq_watchdog_set_thresh(wq_watchdog_th 5874 wq_watchdog_set_thresh(wq_watchdog_thresh);
7594 } 5875 }
7595 5876
7596 #else /* CONFIG_WQ_WATCHDOG */ 5877 #else /* CONFIG_WQ_WATCHDOG */
7597 5878
7598 static inline void wq_watchdog_init(void) { } 5879 static inline void wq_watchdog_init(void) { }
7599 5880
7600 #endif /* CONFIG_WQ_WATCHDOG */ 5881 #endif /* CONFIG_WQ_WATCHDOG */
7601 5882
7602 static void bh_pool_kick_normal(struct irq_wo !! 5883 static void __init wq_numa_init(void)
7603 { 5884 {
7604 raise_softirq_irqoff(TASKLET_SOFTIRQ) !! 5885 cpumask_var_t *tbl;
7605 } !! 5886 int node, cpu;
7606 5887
7607 static void bh_pool_kick_highpri(struct irq_w !! 5888 if (num_possible_nodes() <= 1)
7608 { !! 5889 return;
7609 raise_softirq_irqoff(HI_SOFTIRQ); <<
7610 } <<
7611 5890
7612 static void __init restrict_unbound_cpumask(c !! 5891 if (wq_disable_numa) {
7613 { !! 5892 pr_info("workqueue: NUMA affinity support disabled\n");
7614 if (!cpumask_intersects(wq_unbound_cp <<
7615 pr_warn("workqueue: Restricti <<
7616 cpumask_pr_args(wq_un <<
7617 return; 5893 return;
7618 } 5894 }
7619 5895
7620 cpumask_and(wq_unbound_cpumask, wq_un !! 5896 wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs();
7621 } !! 5897 BUG_ON(!wq_update_unbound_numa_attrs_buf);
7622 5898
7623 static void __init init_cpu_worker_pool(struc !! 5899 /*
7624 { !! 5900 * We want masks of possible CPUs of each node which isn't readily
7625 BUG_ON(init_worker_pool(pool)); !! 5901 * available. Build one from cpu_to_node() which should have been
7626 pool->cpu = cpu; !! 5902 * fully initialized by now.
7627 cpumask_copy(pool->attrs->cpumask, cp !! 5903 */
7628 cpumask_copy(pool->attrs->__pod_cpuma !! 5904 tbl = kcalloc(nr_node_ids, sizeof(tbl[0]), GFP_KERNEL);
7629 pool->attrs->nice = nice; !! 5905 BUG_ON(!tbl);
7630 pool->attrs->affn_strict = true; <<
7631 pool->node = cpu_to_node(cpu); <<
7632 5906
7633 /* alloc pool ID */ !! 5907 for_each_node(node)
7634 mutex_lock(&wq_pool_mutex); !! 5908 BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
7635 BUG_ON(worker_pool_assign_id(pool)); !! 5909 node_online(node) ? node : NUMA_NO_NODE));
7636 mutex_unlock(&wq_pool_mutex); !! 5910
>> 5911 for_each_possible_cpu(cpu) {
>> 5912 node = cpu_to_node(cpu);
>> 5913 if (WARN_ON(node == NUMA_NO_NODE)) {
>> 5914 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
>> 5915 /* happens iff arch is bonkers, let's just proceed */
>> 5916 return;
>> 5917 }
>> 5918 cpumask_set_cpu(cpu, tbl[node]);
>> 5919 }
>> 5920
>> 5921 wq_numa_possible_cpumask = tbl;
>> 5922 wq_numa_enabled = true;
7637 } 5923 }
7638 5924
7639 /** 5925 /**
7640 * workqueue_init_early - early init for work 5926 * workqueue_init_early - early init for workqueue subsystem
7641 * 5927 *
7642 * This is the first step of three-staged wor !! 5928 * This is the first half of two-staged workqueue subsystem initialization
7643 * invoked as soon as the bare basics - memor !! 5929 * and invoked as soon as the bare basics - memory allocation, cpumasks and
7644 * up. It sets up all the data structures and !! 5930 * idr are up. It sets up all the data structures and system workqueues
7645 * boot code to create workqueues and queue/c !! 5931 * and allows early boot code to create workqueues and queue/cancel work
7646 * execution starts only after kthreads can b !! 5932 * items. Actual work item execution starts only after kthreads can be
7647 * before early initcalls. !! 5933 * created and scheduled right before early initcalls.
7648 */ 5934 */
7649 void __init workqueue_init_early(void) 5935 void __init workqueue_init_early(void)
7650 { 5936 {
7651 struct wq_pod_type *pt = &wq_pod_type <<
7652 int std_nice[NR_STD_WORKER_POOLS] = { 5937 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
7653 void (*irq_work_fns[2])(struct irq_wo !! 5938 int hk_flags = HK_FLAG_DOMAIN | HK_FLAG_WQ;
7654 <<
7655 int i, cpu; 5939 int i, cpu;
7656 5940
7657 BUILD_BUG_ON(__alignof__(struct pool_ 5941 BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
7658 5942
7659 BUG_ON(!alloc_cpumask_var(&wq_online_ <<
7660 BUG_ON(!alloc_cpumask_var(&wq_unbound 5943 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
7661 BUG_ON(!alloc_cpumask_var(&wq_request !! 5944 cpumask_copy(wq_unbound_cpumask, housekeeping_cpumask(hk_flags));
7662 BUG_ON(!zalloc_cpumask_var(&wq_isolat <<
7663 <<
7664 cpumask_copy(wq_online_cpumask, cpu_o <<
7665 cpumask_copy(wq_unbound_cpumask, cpu_ <<
7666 restrict_unbound_cpumask("HK_TYPE_WQ" <<
7667 restrict_unbound_cpumask("HK_TYPE_DOM <<
7668 if (!cpumask_empty(&wq_cmdline_cpumas <<
7669 restrict_unbound_cpumask("wor <<
7670 <<
7671 cpumask_copy(wq_requested_unbound_cpu <<
7672 5945
7673 pwq_cache = KMEM_CACHE(pool_workqueue 5946 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
7674 5947
7675 unbound_wq_update_pwq_attrs_buf = all !! 5948 /* initialize CPU pools */
7676 BUG_ON(!unbound_wq_update_pwq_attrs_b <<
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 <<
7685 /* initialize WQ_AFFN_SYSTEM pods */ <<
7686 pt->pod_cpus = kcalloc(1, sizeof(pt-> <<
7687 pt->pod_node = kcalloc(1, sizeof(pt-> <<
7688 pt->cpu_pod = kcalloc(nr_cpu_ids, siz <<
7689 BUG_ON(!pt->pod_cpus || !pt->pod_node <<
7690 <<
7691 BUG_ON(!zalloc_cpumask_var_node(&pt-> <<
7692 <<
7693 pt->nr_pods = 1; <<
7694 cpumask_copy(pt->pod_cpus[0], cpu_pos <<
7695 pt->pod_node[0] = NUMA_NO_NODE; <<
7696 pt->cpu_pod[0] = 0; <<
7697 <<
7698 /* initialize BH and CPU pools */ <<
7699 for_each_possible_cpu(cpu) { 5949 for_each_possible_cpu(cpu) {
7700 struct worker_pool *pool; 5950 struct worker_pool *pool;
7701 5951
7702 i = 0; 5952 i = 0;
7703 for_each_bh_worker_pool(pool, !! 5953 for_each_cpu_worker_pool(pool, cpu) {
7704 init_cpu_worker_pool( !! 5954 BUG_ON(init_worker_pool(pool));
7705 pool->flags |= POOL_B !! 5955 pool->cpu = cpu;
7706 init_irq_work(bh_pool !! 5956 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
7707 i++; !! 5957 pool->attrs->nice = std_nice[i++];
7708 } !! 5958 pool->node = cpu_to_node(cpu);
7709 5959
7710 i = 0; !! 5960 /* alloc pool ID */
7711 for_each_cpu_worker_pool(pool !! 5961 mutex_lock(&wq_pool_mutex);
7712 init_cpu_worker_pool( !! 5962 BUG_ON(worker_pool_assign_id(pool));
>> 5963 mutex_unlock(&wq_pool_mutex);
>> 5964 }
7713 } 5965 }
7714 5966
7715 /* create default unbound and ordered 5967 /* create default unbound and ordered wq attrs */
7716 for (i = 0; i < NR_STD_WORKER_POOLS; 5968 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
7717 struct workqueue_attrs *attrs 5969 struct workqueue_attrs *attrs;
7718 5970
7719 BUG_ON(!(attrs = alloc_workqu 5971 BUG_ON(!(attrs = alloc_workqueue_attrs()));
7720 attrs->nice = std_nice[i]; 5972 attrs->nice = std_nice[i];
7721 unbound_std_wq_attrs[i] = att 5973 unbound_std_wq_attrs[i] = attrs;
7722 5974
7723 /* 5975 /*
7724 * An ordered wq should have 5976 * An ordered wq should have only one pwq as ordering is
7725 * guaranteed by max_active w 5977 * guaranteed by max_active which is enforced by pwqs.
>> 5978 * Turn off NUMA so that dfl_pwq is used for all nodes.
7726 */ 5979 */
7727 BUG_ON(!(attrs = alloc_workqu 5980 BUG_ON(!(attrs = alloc_workqueue_attrs()));
7728 attrs->nice = std_nice[i]; 5981 attrs->nice = std_nice[i];
7729 attrs->ordered = true; !! 5982 attrs->no_numa = true;
7730 ordered_wq_attrs[i] = attrs; 5983 ordered_wq_attrs[i] = attrs;
7731 } 5984 }
7732 5985
7733 system_wq = alloc_workqueue("events", 5986 system_wq = alloc_workqueue("events", 0, 0);
7734 system_highpri_wq = alloc_workqueue(" 5987 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
7735 system_long_wq = alloc_workqueue("eve 5988 system_long_wq = alloc_workqueue("events_long", 0, 0);
7736 system_unbound_wq = alloc_workqueue(" 5989 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
7737 W !! 5990 WQ_UNBOUND_MAX_ACTIVE);
7738 system_freezable_wq = alloc_workqueue 5991 system_freezable_wq = alloc_workqueue("events_freezable",
7739 5992 WQ_FREEZABLE, 0);
7740 system_power_efficient_wq = alloc_wor 5993 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
7741 5994 WQ_POWER_EFFICIENT, 0);
7742 system_freezable_power_efficient_wq = !! 5995 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
7743 5996 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
7744 5997 0);
7745 system_bh_wq = alloc_workqueue("event <<
7746 system_bh_highpri_wq = alloc_workqueu <<
7747 <<
7748 BUG_ON(!system_wq || !system_highpri_ 5998 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
7749 !system_unbound_wq || !system_ 5999 !system_unbound_wq || !system_freezable_wq ||
7750 !system_power_efficient_wq || 6000 !system_power_efficient_wq ||
7751 !system_freezable_power_effici !! 6001 !system_freezable_power_efficient_wq);
7752 !system_bh_wq || !system_bh_hi <<
7753 } <<
7754 <<
7755 static void __init wq_cpu_intensive_thresh_in <<
7756 { <<
7757 unsigned long thresh; <<
7758 unsigned long bogo; <<
7759 <<
7760 pwq_release_worker = kthread_create_w <<
7761 BUG_ON(IS_ERR(pwq_release_worker)); <<
7762 <<
7763 /* if the user set it to a specific v <<
7764 if (wq_cpu_intensive_thresh_us != ULO <<
7765 return; <<
7766 <<
7767 /* <<
7768 * The default of 10ms is derived fro <<
7769 * 2023) processors can do a lot in 1 <<
7770 * most consider human-perceivable. H <<
7771 * lot slower CPUs including microcon <<
7772 * too low. <<
7773 * <<
7774 * Let's scale up the threshold upto <<
7775 * This is by no means accurate but i <<
7776 * is still useful even when the thre <<
7777 * the reports would usually be appli <<
7778 * operating on longer thresholds won <<
7779 * usefulness. <<
7780 */ <<
7781 thresh = 10 * USEC_PER_MSEC; <<
7782 <<
7783 /* see init/calibrate.c for lpj -> Bo <<
7784 bogo = max_t(unsigned long, loops_per <<
7785 if (bogo < 4000) <<
7786 thresh = min_t(unsigned long, <<
7787 <<
7788 pr_debug("wq_cpu_intensive_thresh: lp <<
7789 loops_per_jiffy, bogo, thres <<
7790 <<
7791 wq_cpu_intensive_thresh_us = thresh; <<
7792 } 6002 }
7793 6003
7794 /** 6004 /**
7795 * workqueue_init - bring workqueue subsystem 6005 * workqueue_init - bring workqueue subsystem fully online
7796 * 6006 *
7797 * This is the second step of three-staged wo !! 6007 * This is the latter half of two-staged workqueue subsystem initialization
7798 * and invoked as soon as kthreads can be cre !! 6008 * and invoked as soon as kthreads can be created and scheduled.
7799 * been created and work items queued on them !! 6009 * Workqueues have been created and work items queued on them, but there
7800 * executing the work items yet. Populate the !! 6010 * are no kworkers executing the work items yet. Populate the worker pools
7801 * workers and enable future kworker creation !! 6011 * with the initial workers and enable future kworker creations.
7802 */ 6012 */
7803 void __init workqueue_init(void) 6013 void __init workqueue_init(void)
7804 { 6014 {
7805 struct workqueue_struct *wq; 6015 struct workqueue_struct *wq;
7806 struct worker_pool *pool; 6016 struct worker_pool *pool;
7807 int cpu, bkt; 6017 int cpu, bkt;
7808 6018
7809 wq_cpu_intensive_thresh_init(); !! 6019 /*
>> 6020 * It'd be simpler to initialize NUMA in workqueue_init_early() but
>> 6021 * CPU to node mapping may not be available that early on some
>> 6022 * archs such as power and arm64. As per-cpu pools created
>> 6023 * previously could be missing node hint and unbound pools NUMA
>> 6024 * affinity, fix them up.
>> 6025 *
>> 6026 * Also, while iterating workqueues, create rescuers if requested.
>> 6027 */
>> 6028 wq_numa_init();
7810 6029
7811 mutex_lock(&wq_pool_mutex); 6030 mutex_lock(&wq_pool_mutex);
7812 6031
7813 /* <<
7814 * Per-cpu pools created earlier coul <<
7815 * up. Also, create a rescuer for wor <<
7816 */ <<
7817 for_each_possible_cpu(cpu) { 6032 for_each_possible_cpu(cpu) {
7818 for_each_bh_worker_pool(pool, !! 6033 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 6034 pool->node = cpu_to_node(cpu);
>> 6035 }
7822 } 6036 }
7823 6037
7824 list_for_each_entry(wq, &workqueues, 6038 list_for_each_entry(wq, &workqueues, list) {
>> 6039 wq_update_unbound_numa(wq, smp_processor_id(), true);
7825 WARN(init_rescuer(wq), 6040 WARN(init_rescuer(wq),
7826 "workqueue: failed to cr 6041 "workqueue: failed to create early rescuer for %s",
7827 wq->name); 6042 wq->name);
7828 } 6043 }
7829 6044
7830 mutex_unlock(&wq_pool_mutex); 6045 mutex_unlock(&wq_pool_mutex);
7831 6046
7832 /* !! 6047 /* 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) { 6048 for_each_online_cpu(cpu) {
7843 for_each_cpu_worker_pool(pool 6049 for_each_cpu_worker_pool(pool, cpu) {
7844 pool->flags &= ~POOL_ 6050 pool->flags &= ~POOL_DISASSOCIATED;
7845 BUG_ON(!create_worker 6051 BUG_ON(!create_worker(pool));
7846 } 6052 }
7847 } 6053 }
7848 6054
7849 hash_for_each(unbound_pool_hash, bkt, 6055 hash_for_each(unbound_pool_hash, bkt, pool, hash_node)
7850 BUG_ON(!create_worker(pool)); 6056 BUG_ON(!create_worker(pool));
7851 6057
7852 wq_online = true; 6058 wq_online = true;
7853 wq_watchdog_init(); 6059 wq_watchdog_init();
7854 } 6060 }
7855 <<
7856 /* <<
7857 * Initialize @pt by first initializing @pt-> <<
7858 * @cpu_shares_pod(). Each subset of CPUs tha <<
7859 * and consecutive pod ID. The rest of @pt is <<
7860 */ <<
7861 static void __init init_pod_type(struct wq_po <<
7862 bool (*cpus_ <<
7863 { <<
7864 int cur, pre, cpu, pod; <<
7865 <<
7866 pt->nr_pods = 0; <<
7867 <<
7868 /* init @pt->cpu_pod[] according to @ <<
7869 pt->cpu_pod = kcalloc(nr_cpu_ids, siz <<
7870 BUG_ON(!pt->cpu_pod); <<
7871 <<
7872 for_each_possible_cpu(cur) { <<
7873 for_each_possible_cpu(pre) { <<
7874 if (pre >= cur) { <<
7875 pt->cpu_pod[c <<
7876 break; <<
7877 } <<
7878 if (cpus_share_pod(cu <<
7879 pt->cpu_pod[c <<
7880 break; <<
7881 } <<
7882 } <<
7883 } <<
7884 <<
7885 /* init the rest to match @pt->cpu_po <<
7886 pt->pod_cpus = kcalloc(pt->nr_pods, s <<
7887 pt->pod_node = kcalloc(pt->nr_pods, s <<
7888 BUG_ON(!pt->pod_cpus || !pt->pod_node <<
7889 <<
7890 for (pod = 0; pod < pt->nr_pods; pod+ <<
7891 BUG_ON(!zalloc_cpumask_var(&p <<
7892 <<
7893 for_each_possible_cpu(cpu) { <<
7894 cpumask_set_cpu(cpu, pt->pod_ <<
7895 pt->pod_node[pt->cpu_pod[cpu] <<
7896 } <<
7897 } <<
7898 <<
7899 static bool __init cpus_dont_share(int cpu0, <<
7900 { <<
7901 return false; <<
7902 } <<
7903 <<
7904 static bool __init cpus_share_smt(int cpu0, i <<
7905 { <<
7906 #ifdef CONFIG_SCHED_SMT <<
7907 return cpumask_test_cpu(cpu0, cpu_smt <<
7908 #else <<
7909 return false; <<
7910 #endif <<
7911 } <<
7912 <<
7913 static bool __init cpus_share_numa(int cpu0, <<
7914 { <<
7915 return cpu_to_node(cpu0) == cpu_to_no <<
7916 } <<
7917 <<
7918 /** <<
7919 * workqueue_init_topology - initialize CPU p <<
7920 * <<
7921 * This is the third step of three-staged wor <<
7922 * invoked after SMP and topology information <<
7923 * initializes the unbound CPU pods according <<
7924 */ <<
7925 void __init workqueue_init_topology(void) <<
7926 { <<
7927 struct workqueue_struct *wq; <<
7928 int cpu; <<
7929 <<
7930 init_pod_type(&wq_pod_types[WQ_AFFN_C <<
7931 init_pod_type(&wq_pod_types[WQ_AFFN_S <<
7932 init_pod_type(&wq_pod_types[WQ_AFFN_C <<
7933 init_pod_type(&wq_pod_types[WQ_AFFN_N <<
7934 <<
7935 wq_topo_initialized = true; <<
7936 <<
7937 mutex_lock(&wq_pool_mutex); <<
7938 <<
7939 /* <<
7940 * Workqueues allocated earlier would <<
7941 * worker pool. Explicitly call unbou <<
7942 * and CPU combinations to apply per- <<
7943 */ <<
7944 list_for_each_entry(wq, &workqueues, <<
7945 for_each_online_cpu(cpu) <<
7946 unbound_wq_update_pwq <<
7947 if (wq->flags & WQ_UNBOUND) { <<
7948 mutex_lock(&wq->mutex <<
7949 wq_update_node_max_ac <<
7950 mutex_unlock(&wq->mut <<
7951 } <<
7952 } <<
7953 <<
7954 mutex_unlock(&wq_pool_mutex); <<
7955 } <<
7956 <<
7957 void __warn_flushing_systemwide_wq(void) <<
7958 { <<
7959 pr_warn("WARNING: Flushing system-wid <<
7960 dump_stack(); <<
7961 } <<
7962 EXPORT_SYMBOL(__warn_flushing_systemwide_wq); <<
7963 <<
7964 static int __init workqueue_unbound_cpus_setu <<
7965 { <<
7966 if (cpulist_parse(str, &wq_cmdline_cp <<
7967 cpumask_clear(&wq_cmdline_cpu <<
7968 pr_warn("workqueue.unbound_cp <<
7969 } <<
7970 <<
7971 return 1; <<
7972 } <<
7973 __setup("workqueue.unbound_cpus=", workqueue_ <<
7974 6061
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.