1 // SPDX-License-Identifier: GPL-2.0-only <<
2 /* 1 /*
3 * kernel/workqueue.c - generic async executio 2 * kernel/workqueue.c - generic async execution with shared worker pool
4 * 3 *
5 * Copyright (C) 2002 Ingo Molnar 4 * Copyright (C) 2002 Ingo Molnar
6 * 5 *
7 * Derived from the taskqueue/keventd code b 6 * Derived from the taskqueue/keventd code by:
8 * David Woodhouse <dwmw2@infradead.org> 7 * David Woodhouse <dwmw2@infradead.org>
9 * Andrew Morton 8 * Andrew Morton
10 * Kai Petzke <wpp@marie.physik.tu-berlin. 9 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
11 * Theodore Ts'o <tytso@mit.edu> 10 * Theodore Ts'o <tytso@mit.edu>
12 * 11 *
13 * Made to use alloc_percpu by Christoph Lamet 12 * Made to use alloc_percpu by Christoph Lameter.
14 * 13 *
15 * Copyright (C) 2010 SUSE Linux Pro 14 * Copyright (C) 2010 SUSE Linux Products GmbH
16 * Copyright (C) 2010 Tejun Heo <tj@ 15 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
17 * 16 *
18 * This is the generic async execution mechani 17 * This is the generic async execution mechanism. Work items as are
19 * executed in process context. The worker po 18 * executed in process context. The worker pool is shared and
20 * automatically managed. There are two worke 19 * automatically managed. There are two worker pools for each CPU (one for
21 * normal work items and the other for high pr 20 * normal work items and the other for high priority ones) and some extra
22 * pools for workqueues which are not bound to 21 * pools for workqueues which are not bound to any specific CPU - the
23 * number of these backing pools is dynamic. 22 * number of these backing pools is dynamic.
24 * 23 *
25 * Please read Documentation/core-api/workqueu 24 * Please read Documentation/core-api/workqueue.rst for details.
26 */ 25 */
27 26
28 #include <linux/export.h> 27 #include <linux/export.h>
29 #include <linux/kernel.h> 28 #include <linux/kernel.h>
30 #include <linux/sched.h> 29 #include <linux/sched.h>
31 #include <linux/init.h> 30 #include <linux/init.h>
32 #include <linux/interrupt.h> <<
33 #include <linux/signal.h> 31 #include <linux/signal.h>
34 #include <linux/completion.h> 32 #include <linux/completion.h>
35 #include <linux/workqueue.h> 33 #include <linux/workqueue.h>
36 #include <linux/slab.h> 34 #include <linux/slab.h>
37 #include <linux/cpu.h> 35 #include <linux/cpu.h>
38 #include <linux/notifier.h> 36 #include <linux/notifier.h>
39 #include <linux/kthread.h> 37 #include <linux/kthread.h>
40 #include <linux/hardirq.h> 38 #include <linux/hardirq.h>
41 #include <linux/mempolicy.h> 39 #include <linux/mempolicy.h>
42 #include <linux/freezer.h> 40 #include <linux/freezer.h>
43 #include <linux/debug_locks.h> 41 #include <linux/debug_locks.h>
44 #include <linux/lockdep.h> 42 #include <linux/lockdep.h>
45 #include <linux/idr.h> 43 #include <linux/idr.h>
46 #include <linux/jhash.h> 44 #include <linux/jhash.h>
47 #include <linux/hashtable.h> 45 #include <linux/hashtable.h>
48 #include <linux/rculist.h> 46 #include <linux/rculist.h>
49 #include <linux/nodemask.h> 47 #include <linux/nodemask.h>
50 #include <linux/moduleparam.h> 48 #include <linux/moduleparam.h>
51 #include <linux/uaccess.h> 49 #include <linux/uaccess.h>
52 #include <linux/sched/isolation.h> 50 #include <linux/sched/isolation.h>
53 #include <linux/sched/debug.h> <<
54 #include <linux/nmi.h> 51 #include <linux/nmi.h>
55 #include <linux/kvm_para.h> <<
56 #include <linux/delay.h> <<
57 #include <linux/irq_work.h> <<
58 52
59 #include "workqueue_internal.h" 53 #include "workqueue_internal.h"
60 54
61 enum worker_pool_flags { !! 55 enum {
62 /* 56 /*
63 * worker_pool flags 57 * worker_pool flags
64 * 58 *
65 * A bound pool is either associated o 59 * A bound pool is either associated or disassociated with its CPU.
66 * While associated (!DISASSOCIATED), 60 * While associated (!DISASSOCIATED), all workers are bound to the
67 * CPU and none has %WORKER_UNBOUND se 61 * CPU and none has %WORKER_UNBOUND set and concurrency management
68 * is in effect. 62 * is in effect.
69 * 63 *
70 * While DISASSOCIATED, the cpu may be 64 * While DISASSOCIATED, the cpu may be offline and all workers have
71 * %WORKER_UNBOUND set and concurrency 65 * %WORKER_UNBOUND set and concurrency management disabled, and may
72 * be executing on any CPU. The pool 66 * be executing on any CPU. The pool behaves as an unbound one.
73 * 67 *
74 * Note that DISASSOCIATED should be f 68 * Note that DISASSOCIATED should be flipped only while holding
75 * wq_pool_attach_mutex to avoid chang !! 69 * attach_mutex to avoid changing binding state while
76 * worker_attach_to_pool() is in progr 70 * 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 */ 71 */
81 POOL_BH = 1 << 0, !! 72 POOL_MANAGER_ACTIVE = 1 << 0, /* being managed */
82 POOL_MANAGER_ACTIVE = 1 << 1, <<
83 POOL_DISASSOCIATED = 1 << 2, 73 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
84 POOL_BH_DRAINING = 1 << 3, <<
85 }; <<
86 74
87 enum worker_flags { <<
88 /* worker flags */ 75 /* worker flags */
89 WORKER_DIE = 1 << 1, 76 WORKER_DIE = 1 << 1, /* die die die */
90 WORKER_IDLE = 1 << 2, 77 WORKER_IDLE = 1 << 2, /* is idle */
91 WORKER_PREP = 1 << 3, 78 WORKER_PREP = 1 << 3, /* preparing to run works */
92 WORKER_CPU_INTENSIVE = 1 << 6, 79 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
93 WORKER_UNBOUND = 1 << 7, 80 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
94 WORKER_REBOUND = 1 << 8, 81 WORKER_REBOUND = 1 << 8, /* worker was rebound */
95 82
96 WORKER_NOT_RUNNING = WORKER_PREP 83 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
97 WORKER_UNBOU 84 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 85
105 enum wq_internal_consts { <<
106 NR_STD_WORKER_POOLS = 2, 86 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
107 87
108 UNBOUND_POOL_HASH_ORDER = 6, 88 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
109 BUSY_WORKER_HASH_ORDER = 6, 89 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
110 90
111 MAX_IDLE_WORKERS_RATIO = 4, 91 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
112 IDLE_WORKER_TIMEOUT = 300 * HZ, 92 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
113 93
114 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 94 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
115 95 /* call for help after 10ms
116 96 (min two ticks) */
117 MAYDAY_INTERVAL = HZ / 10, 97 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
118 CREATE_COOLDOWN = HZ, 98 CREATE_COOLDOWN = HZ, /* time to breath after fail */
119 99
120 /* 100 /*
121 * Rescue workers are used only on eme 101 * Rescue workers are used only on emergencies and shared by
122 * all cpus. Give MIN_NICE. 102 * all cpus. Give MIN_NICE.
123 */ 103 */
124 RESCUER_NICE_LEVEL = MIN_NICE, 104 RESCUER_NICE_LEVEL = MIN_NICE,
125 HIGHPRI_NICE_LEVEL = MIN_NICE, 105 HIGHPRI_NICE_LEVEL = MIN_NICE,
126 106
127 WQ_NAME_LEN = 32, !! 107 WQ_NAME_LEN = 24,
128 WORKER_ID_LEN = 10 + WQ_NAME <<
129 }; 108 };
130 109
131 /* 110 /*
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 111 * Structure fields follow one of the following exclusion rules.
141 * 112 *
142 * I: Modifiable by initialization/destruction 113 * I: Modifiable by initialization/destruction paths and read-only for
143 * everyone else. 114 * everyone else.
144 * 115 *
145 * P: Preemption protected. Disabling preempt 116 * P: Preemption protected. Disabling preemption is enough and should
146 * only be modified and accessed from the l 117 * only be modified and accessed from the local cpu.
147 * 118 *
148 * L: pool->lock protected. Access with pool- 119 * L: pool->lock protected. Access with pool->lock held.
149 * 120 *
150 * LN: pool->lock and wq_node_nr_active->lock !! 121 * X: During normal operation, modification requires pool->lock and should
151 * reads. !! 122 * be done only from local cpu. Either disabling preemption on local
>> 123 * cpu or grabbing pool->lock is enough for read access. If
>> 124 * POOL_DISASSOCIATED is set, it's identical to L.
152 * 125 *
153 * K: Only modified by worker while holding po !! 126 * A: pool->attach_mutex protected.
154 * self, while holding pool->lock or from I <<
155 * kworker. <<
156 * <<
157 * S: Only modified by worker self. <<
158 * <<
159 * A: wq_pool_attach_mutex protected. <<
160 * 127 *
161 * PL: wq_pool_mutex protected. 128 * PL: wq_pool_mutex protected.
162 * 129 *
163 * PR: wq_pool_mutex protected for writes. RC !! 130 * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads.
164 * 131 *
165 * PW: wq_pool_mutex and wq->mutex protected f 132 * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads.
166 * 133 *
167 * PWR: wq_pool_mutex and wq->mutex protected 134 * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or
168 * RCU for reads. !! 135 * sched-RCU for reads.
169 * 136 *
170 * WQ: wq->mutex protected. 137 * WQ: wq->mutex protected.
171 * 138 *
172 * WR: wq->mutex protected for writes. RCU pr !! 139 * WR: wq->mutex protected for writes. Sched-RCU protected for reads.
173 * <<
174 * WO: wq->mutex protected for writes. Updated <<
175 * with READ_ONCE() without locking. <<
176 * 140 *
177 * MD: wq_mayday_lock protected. 141 * MD: wq_mayday_lock protected.
178 * <<
179 * WD: Used internally by the watchdog. <<
180 */ 142 */
181 143
182 /* struct worker is defined in workqueue_inter 144 /* struct worker is defined in workqueue_internal.h */
183 145
184 struct worker_pool { 146 struct worker_pool {
185 raw_spinlock_t lock; !! 147 spinlock_t lock; /* the pool lock */
186 int cpu; 148 int cpu; /* I: the associated cpu */
187 int node; 149 int node; /* I: the associated node ID */
188 int id; 150 int id; /* I: pool ID */
189 unsigned int flags; !! 151 unsigned int flags; /* X: flags */
190 152
191 unsigned long watchdog_ts; 153 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; <<
201 154
202 struct list_head worklist; 155 struct list_head worklist; /* L: list of pending works */
203 <<
204 int nr_workers; 156 int nr_workers; /* L: total number of workers */
205 int nr_idle; <<
206 157
207 struct list_head idle_list; !! 158 /* nr_idle includes the ones off idle_list for rebinding */
208 struct timer_list idle_timer; !! 159 int nr_idle; /* L: currently idle ones */
209 struct work_struct idle_cull_work <<
210 160
211 struct timer_list mayday_timer; !! 161 struct list_head idle_list; /* X: list of idle workers */
>> 162 struct timer_list idle_timer; /* L: worker idle timeout */
>> 163 struct timer_list mayday_timer; /* L: SOS timer for workers */
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
>> 169 /* see manage_workers() for details on the two manager mutexes */
217 struct worker *manager; 170 struct worker *manager; /* L: purely informational */
>> 171 struct mutex attach_mutex; /* attach/detach exclusion */
218 struct list_head workers; 172 struct list_head workers; /* A: attached workers */
>> 173 struct completion *detach_completion; /* all workers detached */
219 174
220 struct ida worker_ida; 175 struct ida worker_ida; /* worker IDs for task name */
221 176
222 struct workqueue_attrs *attrs; 177 struct workqueue_attrs *attrs; /* I: worker attributes */
223 struct hlist_node hash_node; 178 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
224 int refcnt; 179 int refcnt; /* PL: refcnt for unbound pools */
225 180
226 /* 181 /*
227 * Destruction of pool is RCU protecte !! 182 * The current concurrency level. As it's likely to be accessed
>> 183 * from other CPUs during try_to_wake_up(), put it in a separate
>> 184 * cacheline.
>> 185 */
>> 186 atomic_t nr_running ____cacheline_aligned_in_smp;
>> 187
>> 188 /*
>> 189 * Destruction of pool is sched-RCU protected to allow dereferences
228 * from get_work_pool(). 190 * from get_work_pool().
229 */ 191 */
230 struct rcu_head rcu; 192 struct rcu_head rcu;
231 }; !! 193 } ____cacheline_aligned_in_smp;
232 194
233 /* 195 /*
234 * Per-pool_workqueue statistics. These can be !! 196 * 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 197 * of work_struct->data are used for flags and the remaining high bits
253 * point to the pwq; thus, pwqs need to be ali 198 * point to the pwq; thus, pwqs need to be aligned at two's power of the
254 * number of flag bits. 199 * number of flag bits.
255 */ 200 */
256 struct pool_workqueue { 201 struct pool_workqueue {
257 struct worker_pool *pool; 202 struct worker_pool *pool; /* I: the associated pool */
258 struct workqueue_struct *wq; 203 struct workqueue_struct *wq; /* I: the owning workqueue */
259 int work_color; 204 int work_color; /* L: current color */
260 int flush_color; 205 int flush_color; /* L: flushing color */
261 int refcnt; 206 int refcnt; /* L: reference count */
262 int nr_in_flight[W 207 int nr_in_flight[WORK_NR_COLORS];
263 208 /* 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; 209 int nr_active; /* L: nr of active works */
283 struct list_head inactive_works !! 210 int max_active; /* L: max active works */
284 struct list_head pending_node; !! 211 struct list_head delayed_works; /* L: delayed works */
285 struct list_head pwqs_node; 212 struct list_head pwqs_node; /* WR: node on wq->pwqs */
286 struct list_head mayday_node; 213 struct list_head mayday_node; /* MD: node on wq->maydays */
287 214
288 u64 stats[PWQ_NR_S <<
289 <<
290 /* 215 /*
291 * Release of unbound pwq is punted to !! 216 * Release of unbound pwq is punted to system_wq. See put_pwq()
292 * and pwq_release_workfn() for detail !! 217 * and pwq_unbound_release_workfn() for details. pool_workqueue
293 * RCU protected so that the first pwq !! 218 * itself is also sched-RCU protected so that the first pwq can be
294 * grabbing wq->mutex. !! 219 * determined without grabbing wq->mutex.
295 */ 220 */
296 struct kthread_work release_work; !! 221 struct work_struct unbound_release_work;
297 struct rcu_head rcu; 222 struct rcu_head rcu;
298 } __aligned(1 << WORK_STRUCT_PWQ_SHIFT); !! 223 } __aligned(1 << WORK_STRUCT_FLAG_BITS);
299 224
300 /* 225 /*
301 * Structure used to wait for workqueue flush. 226 * Structure used to wait for workqueue flush.
302 */ 227 */
303 struct wq_flusher { 228 struct wq_flusher {
304 struct list_head list; 229 struct list_head list; /* WQ: list of flushers */
305 int flush_color; 230 int flush_color; /* WQ: flush color waiting for */
306 struct completion done; 231 struct completion done; /* flush completion */
307 }; 232 };
308 233
309 struct wq_device; 234 struct wq_device;
310 235
311 /* 236 /*
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 237 * The externally visible workqueue. It relays the issued work items to
333 * the appropriate worker_pool through its poo 238 * the appropriate worker_pool through its pool_workqueues.
334 */ 239 */
335 struct workqueue_struct { 240 struct workqueue_struct {
336 struct list_head pwqs; 241 struct list_head pwqs; /* WR: all pwqs of this wq */
337 struct list_head list; 242 struct list_head list; /* PR: list of all workqueues */
338 243
339 struct mutex mutex; 244 struct mutex mutex; /* protects this wq */
340 int work_color; 245 int work_color; /* WQ: current work color */
341 int flush_color; 246 int flush_color; /* WQ: current flush color */
342 atomic_t nr_pwqs_to_flu 247 atomic_t nr_pwqs_to_flush; /* flush in progress */
343 struct wq_flusher *first_flusher 248 struct wq_flusher *first_flusher; /* WQ: first flusher */
344 struct list_head flusher_queue; 249 struct list_head flusher_queue; /* WQ: flush waiters */
345 struct list_head flusher_overfl 250 struct list_head flusher_overflow; /* WQ: flush overflow list */
346 251
347 struct list_head maydays; 252 struct list_head maydays; /* MD: pwqs requesting rescue */
348 struct worker *rescuer; !! 253 struct worker *rescuer; /* I: rescue worker */
349 254
350 int nr_drainers; 255 int nr_drainers; /* WQ: drain in progress */
351 !! 256 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 <<
357 257
358 struct workqueue_attrs *unbound_attrs 258 struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */
359 struct pool_workqueue __rcu *dfl_pwq; !! 259 struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */
360 260
361 #ifdef CONFIG_SYSFS 261 #ifdef CONFIG_SYSFS
362 struct wq_device *wq_dev; 262 struct wq_device *wq_dev; /* I: for sysfs interface */
363 #endif 263 #endif
364 #ifdef CONFIG_LOCKDEP 264 #ifdef CONFIG_LOCKDEP
365 char *lock_name; <<
366 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 sched-RCU protected to allow
373 * the workqueues list without grabbin !! 271 * walking 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 <<
395 struct work_offq_data { <<
396 u32 pool_id; <<
397 u32 disable; <<
398 u32 flags; <<
399 }; <<
400 283
401 static const char *wq_affn_names[WQ_AFFN_NR_TY !! 284 static cpumask_var_t *wq_numa_possible_cpumask;
402 [WQ_AFFN_DFL] = "default", !! 285 /* possible CPUs of each node */
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 }; <<
409 286
410 /* !! 287 static bool wq_disable_numa;
411 * Per-cpu work items which run for longer tha !! 288 module_param_named(disable_numa, wq_disable_numa, bool, 0444);
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_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
441 static DEFINE_RAW_SPINLOCK(wq_mayday_lock); !! 303 static DECLARE_WAIT_QUEUE_HEAD(wq_manager_wait); /* wait for manager to go away */
442 /* wait for manager to go away */ <<
443 static struct rcuwait manager_wait = __RCUWAIT <<
444 304
445 static LIST_HEAD(workqueues); /* PR: 305 static LIST_HEAD(workqueues); /* PR: list of all workqueues */
446 static bool workqueue_freezing; /* PL: 306 static bool workqueue_freezing; /* PL: have wqs started freezing? */
447 307
448 /* PL: mirror the cpu_online_mask excluding th !! 308 /* 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; 309 static cpumask_var_t wq_unbound_cpumask;
453 310
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 311 /* CPU where unbound work was last round robin scheduled from this CPU */
464 static DEFINE_PER_CPU(int, wq_rr_cpu_last); 312 static DEFINE_PER_CPU(int, wq_rr_cpu_last);
465 313
466 /* 314 /*
467 * Local execution of unbound work items is no 315 * Local execution of unbound work items is no longer guaranteed. The
468 * following always forces round-robin CPU sel 316 * following always forces round-robin CPU selection on unbound work items
469 * to uncover usages which depend on it. 317 * to uncover usages which depend on it.
470 */ 318 */
471 #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU 319 #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
472 static bool wq_debug_force_rr_cpu = true; 320 static bool wq_debug_force_rr_cpu = true;
473 #else 321 #else
474 static bool wq_debug_force_rr_cpu = false; 322 static bool wq_debug_force_rr_cpu = false;
475 #endif 323 #endif
476 module_param_named(debug_force_rr_cpu, wq_debu 324 module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
477 325
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 */ 326 /* the per-cpu worker pools */
487 static DEFINE_PER_CPU_SHARED_ALIGNED(struct wo !! 327 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools);
488 cpu_worke <<
489 328
490 static DEFINE_IDR(worker_pool_idr); /* PR: 329 static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
491 330
492 /* PL: hash of all unbound pools keyed by pool 331 /* PL: hash of all unbound pools keyed by pool->attrs */
493 static DEFINE_HASHTABLE(unbound_pool_hash, UNB 332 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
494 333
495 /* I: attributes used when instantiating stand 334 /* I: attributes used when instantiating standard unbound pools on demand */
496 static struct workqueue_attrs *unbound_std_wq_ 335 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
497 336
498 /* I: attributes used when instantiating order 337 /* I: attributes used when instantiating ordered pools on demand */
499 static struct workqueue_attrs *ordered_wq_attr 338 static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
500 339
501 /* !! 340 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); 341 EXPORT_SYMBOL(system_wq);
510 struct workqueue_struct *system_highpri_wq __r !! 342 struct workqueue_struct *system_highpri_wq __read_mostly;
511 EXPORT_SYMBOL_GPL(system_highpri_wq); 343 EXPORT_SYMBOL_GPL(system_highpri_wq);
512 struct workqueue_struct *system_long_wq __ro_a !! 344 struct workqueue_struct *system_long_wq __read_mostly;
513 EXPORT_SYMBOL_GPL(system_long_wq); 345 EXPORT_SYMBOL_GPL(system_long_wq);
514 struct workqueue_struct *system_unbound_wq __r !! 346 struct workqueue_struct *system_unbound_wq __read_mostly;
515 EXPORT_SYMBOL_GPL(system_unbound_wq); 347 EXPORT_SYMBOL_GPL(system_unbound_wq);
516 struct workqueue_struct *system_freezable_wq _ !! 348 struct workqueue_struct *system_freezable_wq __read_mostly;
517 EXPORT_SYMBOL_GPL(system_freezable_wq); 349 EXPORT_SYMBOL_GPL(system_freezable_wq);
518 struct workqueue_struct *system_power_efficien !! 350 struct workqueue_struct *system_power_efficient_wq __read_mostly;
519 EXPORT_SYMBOL_GPL(system_power_efficient_wq); 351 EXPORT_SYMBOL_GPL(system_power_efficient_wq);
520 struct workqueue_struct *system_freezable_powe !! 352 struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
521 EXPORT_SYMBOL_GPL(system_freezable_power_effic 353 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); <<
526 354
527 static int worker_thread(void *__worker); 355 static int worker_thread(void *__worker);
528 static void workqueue_sysfs_unregister(struct 356 static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
529 static void show_pwq(struct pool_workqueue *pw <<
530 static void show_one_worker_pool(struct worker <<
531 357
532 #define CREATE_TRACE_POINTS 358 #define CREATE_TRACE_POINTS
533 #include <trace/events/workqueue.h> 359 #include <trace/events/workqueue.h>
534 360
535 #define assert_rcu_or_pool_mutex() 361 #define assert_rcu_or_pool_mutex() \
536 RCU_LOCKDEP_WARN(!rcu_read_lock_any_he !! 362 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
537 !lockdep_is_held(&wq_ 363 !lockdep_is_held(&wq_pool_mutex), \
538 "RCU or wq_pool_mutex !! 364 "sched RCU or wq_pool_mutex should be held")
>> 365
>> 366 #define assert_rcu_or_wq_mutex(wq) \
>> 367 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
>> 368 !lockdep_is_held(&wq->mutex), \
>> 369 "sched RCU or wq->mutex should be held")
539 370
540 #define assert_rcu_or_wq_mutex_or_pool_mutex(w 371 #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
541 RCU_LOCKDEP_WARN(!rcu_read_lock_any_he !! 372 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
542 !lockdep_is_held(&wq- 373 !lockdep_is_held(&wq->mutex) && \
543 !lockdep_is_held(&wq_ 374 !lockdep_is_held(&wq_pool_mutex), \
544 "RCU, wq->mutex or wq !! 375 "sched RCU, wq->mutex or wq_pool_mutex should be held")
545 <<
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)++) <<
550 376
551 #define for_each_cpu_worker_pool(pool, cpu) 377 #define for_each_cpu_worker_pool(pool, cpu) \
552 for ((pool) = &per_cpu(cpu_worker_pool 378 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
553 (pool) < &per_cpu(cpu_worker_pool 379 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
554 (pool)++) 380 (pool)++)
555 381
556 /** 382 /**
557 * for_each_pool - iterate through all worker_ 383 * for_each_pool - iterate through all worker_pools in the system
558 * @pool: iteration cursor 384 * @pool: iteration cursor
559 * @pi: integer used for iteration 385 * @pi: integer used for iteration
560 * 386 *
561 * This must be called either with wq_pool_mut !! 387 * This must be called either with wq_pool_mutex held or sched RCU read
562 * locked. If the pool needs to be used beyon 388 * locked. If the pool needs to be used beyond the locking in effect, the
563 * caller is responsible for guaranteeing that 389 * caller is responsible for guaranteeing that the pool stays online.
564 * 390 *
565 * The if/else clause exists only for the lock 391 * The if/else clause exists only for the lockdep assertion and can be
566 * ignored. 392 * ignored.
567 */ 393 */
568 #define for_each_pool(pool, pi) 394 #define for_each_pool(pool, pi) \
569 idr_for_each_entry(&worker_pool_idr, p 395 idr_for_each_entry(&worker_pool_idr, pool, pi) \
570 if (({ assert_rcu_or_pool_mute 396 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
571 else 397 else
572 398
573 /** 399 /**
574 * for_each_pool_worker - iterate through all 400 * for_each_pool_worker - iterate through all workers of a worker_pool
575 * @worker: iteration cursor 401 * @worker: iteration cursor
576 * @pool: worker_pool to iterate workers of 402 * @pool: worker_pool to iterate workers of
577 * 403 *
578 * This must be called with wq_pool_attach_mut !! 404 * This must be called with @pool->attach_mutex.
579 * 405 *
580 * The if/else clause exists only for the lock 406 * The if/else clause exists only for the lockdep assertion and can be
581 * ignored. 407 * ignored.
582 */ 408 */
583 #define for_each_pool_worker(worker, pool) 409 #define for_each_pool_worker(worker, pool) \
584 list_for_each_entry((worker), &(pool)- 410 list_for_each_entry((worker), &(pool)->workers, node) \
585 if (({ lockdep_assert_held(&wq !! 411 if (({ lockdep_assert_held(&pool->attach_mutex); false; })) { } \
586 else 412 else
587 413
588 /** 414 /**
589 * for_each_pwq - iterate through all pool_wor 415 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
590 * @pwq: iteration cursor 416 * @pwq: iteration cursor
591 * @wq: the target workqueue 417 * @wq: the target workqueue
592 * 418 *
593 * This must be called either with wq->mutex h !! 419 * This must be called either with wq->mutex held or sched RCU read locked.
594 * If the pwq needs to be used beyond the lock 420 * If the pwq needs to be used beyond the locking in effect, the caller is
595 * responsible for guaranteeing that the pwq s 421 * responsible for guaranteeing that the pwq stays online.
596 * 422 *
597 * The if/else clause exists only for the lock 423 * The if/else clause exists only for the lockdep assertion and can be
598 * ignored. 424 * ignored.
599 */ 425 */
600 #define for_each_pwq(pwq, wq) 426 #define for_each_pwq(pwq, wq) \
601 list_for_each_entry_rcu((pwq), &(wq)-> !! 427 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
602 lockdep_is_he !! 428 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
>> 429 else
603 430
604 #ifdef CONFIG_DEBUG_OBJECTS_WORK 431 #ifdef CONFIG_DEBUG_OBJECTS_WORK
605 432
606 static const struct debug_obj_descr work_debug !! 433 static struct debug_obj_descr work_debug_descr;
607 434
608 static void *work_debug_hint(void *addr) 435 static void *work_debug_hint(void *addr)
609 { 436 {
610 return ((struct work_struct *) addr)-> 437 return ((struct work_struct *) addr)->func;
611 } 438 }
612 439
613 static bool work_is_static_object(void *addr) 440 static bool work_is_static_object(void *addr)
614 { 441 {
615 struct work_struct *work = addr; 442 struct work_struct *work = addr;
616 443
617 return test_bit(WORK_STRUCT_STATIC_BIT 444 return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work));
618 } 445 }
619 446
620 /* 447 /*
621 * fixup_init is called when: 448 * fixup_init is called when:
622 * - an active object is initialized 449 * - an active object is initialized
623 */ 450 */
624 static bool work_fixup_init(void *addr, enum d 451 static bool work_fixup_init(void *addr, enum debug_obj_state state)
625 { 452 {
626 struct work_struct *work = addr; 453 struct work_struct *work = addr;
627 454
628 switch (state) { 455 switch (state) {
629 case ODEBUG_STATE_ACTIVE: 456 case ODEBUG_STATE_ACTIVE:
630 cancel_work_sync(work); 457 cancel_work_sync(work);
631 debug_object_init(work, &work_ 458 debug_object_init(work, &work_debug_descr);
632 return true; 459 return true;
633 default: 460 default:
634 return false; 461 return false;
635 } 462 }
636 } 463 }
637 464
638 /* 465 /*
639 * fixup_free is called when: 466 * fixup_free is called when:
640 * - an active object is freed 467 * - an active object is freed
641 */ 468 */
642 static bool work_fixup_free(void *addr, enum d 469 static bool work_fixup_free(void *addr, enum debug_obj_state state)
643 { 470 {
644 struct work_struct *work = addr; 471 struct work_struct *work = addr;
645 472
646 switch (state) { 473 switch (state) {
647 case ODEBUG_STATE_ACTIVE: 474 case ODEBUG_STATE_ACTIVE:
648 cancel_work_sync(work); 475 cancel_work_sync(work);
649 debug_object_free(work, &work_ 476 debug_object_free(work, &work_debug_descr);
650 return true; 477 return true;
651 default: 478 default:
652 return false; 479 return false;
653 } 480 }
654 } 481 }
655 482
656 static const struct debug_obj_descr work_debug !! 483 static struct debug_obj_descr work_debug_descr = {
657 .name = "work_struct", 484 .name = "work_struct",
658 .debug_hint = work_debug_hint, 485 .debug_hint = work_debug_hint,
659 .is_static_object = work_is_static_obj 486 .is_static_object = work_is_static_object,
660 .fixup_init = work_fixup_init, 487 .fixup_init = work_fixup_init,
661 .fixup_free = work_fixup_free, 488 .fixup_free = work_fixup_free,
662 }; 489 };
663 490
664 static inline void debug_work_activate(struct 491 static inline void debug_work_activate(struct work_struct *work)
665 { 492 {
666 debug_object_activate(work, &work_debu 493 debug_object_activate(work, &work_debug_descr);
667 } 494 }
668 495
669 static inline void debug_work_deactivate(struc 496 static inline void debug_work_deactivate(struct work_struct *work)
670 { 497 {
671 debug_object_deactivate(work, &work_de 498 debug_object_deactivate(work, &work_debug_descr);
672 } 499 }
673 500
674 void __init_work(struct work_struct *work, int 501 void __init_work(struct work_struct *work, int onstack)
675 { 502 {
676 if (onstack) 503 if (onstack)
677 debug_object_init_on_stack(wor 504 debug_object_init_on_stack(work, &work_debug_descr);
678 else 505 else
679 debug_object_init(work, &work_ 506 debug_object_init(work, &work_debug_descr);
680 } 507 }
681 EXPORT_SYMBOL_GPL(__init_work); 508 EXPORT_SYMBOL_GPL(__init_work);
682 509
683 void destroy_work_on_stack(struct work_struct 510 void destroy_work_on_stack(struct work_struct *work)
684 { 511 {
685 debug_object_free(work, &work_debug_de 512 debug_object_free(work, &work_debug_descr);
686 } 513 }
687 EXPORT_SYMBOL_GPL(destroy_work_on_stack); 514 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
688 515
689 void destroy_delayed_work_on_stack(struct dela 516 void destroy_delayed_work_on_stack(struct delayed_work *work)
690 { 517 {
691 destroy_timer_on_stack(&work->timer); 518 destroy_timer_on_stack(&work->timer);
692 debug_object_free(&work->work, &work_d 519 debug_object_free(&work->work, &work_debug_descr);
693 } 520 }
694 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stac 521 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
695 522
696 #else 523 #else
697 static inline void debug_work_activate(struct 524 static inline void debug_work_activate(struct work_struct *work) { }
698 static inline void debug_work_deactivate(struc 525 static inline void debug_work_deactivate(struct work_struct *work) { }
699 #endif 526 #endif
700 527
701 /** 528 /**
702 * worker_pool_assign_id - allocate ID and ass !! 529 * worker_pool_assign_id - allocate ID and assing it to @pool
703 * @pool: the pool pointer of interest 530 * @pool: the pool pointer of interest
704 * 531 *
705 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) 532 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
706 * successfully, -errno on failure. 533 * successfully, -errno on failure.
707 */ 534 */
708 static int worker_pool_assign_id(struct worker 535 static int worker_pool_assign_id(struct worker_pool *pool)
709 { 536 {
710 int ret; 537 int ret;
711 538
712 lockdep_assert_held(&wq_pool_mutex); 539 lockdep_assert_held(&wq_pool_mutex);
713 540
714 ret = idr_alloc(&worker_pool_idr, pool 541 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
715 GFP_KERNEL); 542 GFP_KERNEL);
716 if (ret >= 0) { 543 if (ret >= 0) {
717 pool->id = ret; 544 pool->id = ret;
718 return 0; 545 return 0;
719 } 546 }
720 return ret; 547 return ret;
721 } 548 }
722 549
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 /** 550 /**
741 * unbound_effective_cpumask - effective cpuma !! 551 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
742 * @wq: workqueue of interest !! 552 * @wq: the target workqueue
>> 553 * @node: the node ID
743 * 554 *
744 * @wq->unbound_attrs->cpumask contains the cp !! 555 * This must be called with any of wq_pool_mutex, wq->mutex or sched RCU
745 * is masked with wq_unbound_cpumask to determ !! 556 * read locked.
746 * default pwq is always mapped to the pool wi !! 557 * If the pwq needs to be used beyond the locking in effect, the caller is
>> 558 * responsible for guaranteeing that the pwq stays online.
>> 559 *
>> 560 * Return: The unbound pool_workqueue for @node.
747 */ 561 */
748 static struct cpumask *unbound_effective_cpuma !! 562 static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
>> 563 int node)
749 { 564 {
750 return unbound_pwq(wq, -1)->pool->attr !! 565 assert_rcu_or_wq_mutex_or_pool_mutex(wq);
>> 566
>> 567 /*
>> 568 * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a
>> 569 * delayed item is pending. The plan is to keep CPU -> NODE
>> 570 * mapping valid and stable across CPU on/offlines. Once that
>> 571 * happens, this workaround can be removed.
>> 572 */
>> 573 if (unlikely(node == NUMA_NO_NODE))
>> 574 return wq->dfl_pwq;
>> 575
>> 576 return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
751 } 577 }
752 578
753 static unsigned int work_color_to_flags(int co 579 static unsigned int work_color_to_flags(int color)
754 { 580 {
755 return color << WORK_STRUCT_COLOR_SHIF 581 return color << WORK_STRUCT_COLOR_SHIFT;
756 } 582 }
757 583
758 static int get_work_color(unsigned long work_d !! 584 static int get_work_color(struct work_struct *work)
759 { 585 {
760 return (work_data >> WORK_STRUCT_COLOR !! 586 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
761 ((1 << WORK_STRUCT_COLOR_BITS) 587 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
762 } 588 }
763 589
764 static int work_next_color(int color) 590 static int work_next_color(int color)
765 { 591 {
766 return (color + 1) % WORK_NR_COLORS; 592 return (color + 1) % WORK_NR_COLORS;
767 } 593 }
768 594
769 static unsigned long pool_offq_flags(struct wo <<
770 { <<
771 return (pool->flags & POOL_BH) ? WORK_ <<
772 } <<
773 <<
774 /* 595 /*
775 * While queued, %WORK_STRUCT_PWQ is set and n 596 * 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 597 * contain the pointer to the queued pwq. Once execution starts, the flag
777 * is cleared and the high bits contain OFFQ f 598 * is cleared and the high bits contain OFFQ flags and pool ID.
778 * 599 *
779 * set_work_pwq(), set_work_pool_and_clear_pen !! 600 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
780 * can be used to set the pwq, pool or clear w !! 601 * and clear_work_data() can be used to set the pwq, pool or clear
781 * only be called while the work is owned - ie !! 602 * work->data. These functions should only be called while the work is
>> 603 * owned - ie. while the PENDING bit is set.
782 * 604 *
783 * get_work_pool() and get_work_pwq() can be u 605 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
784 * corresponding to a work. Pool is available 606 * corresponding to a work. Pool is available once the work has been
785 * queued anywhere after initialization until 607 * queued anywhere after initialization until it is sync canceled. pwq is
786 * available only while the work item is queue 608 * available only while the work item is queued.
>> 609 *
>> 610 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
>> 611 * canceled. While being canceled, a work item may have its PENDING set
>> 612 * but stay off timer and worklist for arbitrarily long and nobody should
>> 613 * try to steal the PENDING bit.
787 */ 614 */
788 static inline void set_work_data(struct work_s !! 615 static inline void set_work_data(struct work_struct *work, unsigned long data,
>> 616 unsigned long flags)
789 { 617 {
790 WARN_ON_ONCE(!work_pending(work)); 618 WARN_ON_ONCE(!work_pending(work));
791 atomic_long_set(&work->data, data | wo !! 619 atomic_long_set(&work->data, data | flags | work_static(work));
792 } 620 }
793 621
794 static void set_work_pwq(struct work_struct *w 622 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
795 unsigned long flags) !! 623 unsigned long extra_flags)
796 { 624 {
797 set_work_data(work, (unsigned long)pwq !! 625 set_work_data(work, (unsigned long)pwq,
798 WORK_STRUCT_PWQ | flags) !! 626 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
799 } 627 }
800 628
801 static void set_work_pool_and_keep_pending(str 629 static void set_work_pool_and_keep_pending(struct work_struct *work,
802 int !! 630 int pool_id)
803 { 631 {
804 set_work_data(work, ((unsigned long)po !! 632 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
805 WORK_STRUCT_PENDING | fl !! 633 WORK_STRUCT_PENDING);
806 } 634 }
807 635
808 static void set_work_pool_and_clear_pending(st 636 static void set_work_pool_and_clear_pending(struct work_struct *work,
809 in !! 637 int pool_id)
810 { 638 {
811 /* 639 /*
812 * The following wmb is paired with th 640 * The following wmb is paired with the implied mb in
813 * test_and_set_bit(PENDING) and ensur 641 * test_and_set_bit(PENDING) and ensures all updates to @work made
814 * here are visible to and precede any 642 * here are visible to and precede any updates by the next PENDING
815 * owner. 643 * owner.
816 */ 644 */
817 smp_wmb(); 645 smp_wmb();
818 set_work_data(work, ((unsigned long)po !! 646 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
819 flags); <<
820 /* 647 /*
821 * The following mb guarantees that pr 648 * The following mb guarantees that previous clear of a PENDING bit
822 * will not be reordered with any spec 649 * will not be reordered with any speculative LOADS or STORES from
823 * work->current_func, which is execut 650 * work->current_func, which is executed afterwards. This possible
824 * reordering can lead to a missed exe !! 651 * reordering can lead to a missed execution on attempt to qeueue
825 * the same @work. E.g. consider this 652 * the same @work. E.g. consider this case:
826 * 653 *
827 * CPU#0 CPU 654 * CPU#0 CPU#1
828 * ---------------------------- --- 655 * ---------------------------- --------------------------------
829 * 656 *
830 * 1 STORE event_indicated 657 * 1 STORE event_indicated
831 * 2 queue_work_on() { 658 * 2 queue_work_on() {
832 * 3 test_and_set_bit(PENDING) 659 * 3 test_and_set_bit(PENDING)
833 * 4 } set 660 * 4 } set_..._and_clear_pending() {
834 * 5 s 661 * 5 set_work_data() # clear bit
835 * 6 s 662 * 6 smp_mb()
836 * 7 wor 663 * 7 work->current_func() {
837 * 8 664 * 8 LOAD event_indicated
838 * } 665 * }
839 * 666 *
840 * Without an explicit full barrier sp 667 * Without an explicit full barrier speculative LOAD on line 8 can
841 * be executed before CPU#0 does STORE 668 * be executed before CPU#0 does STORE on line 1. If that happens,
842 * CPU#0 observes the PENDING bit is s 669 * CPU#0 observes the PENDING bit is still set and new execution of
843 * a @work is not queued in a hope, th 670 * a @work is not queued in a hope, that CPU#1 will eventually
844 * finish the queued @work. Meanwhile 671 * finish the queued @work. Meanwhile CPU#1 does not see
845 * event_indicated is set, because spe 672 * event_indicated is set, because speculative LOAD was executed
846 * before actual STORE. 673 * before actual STORE.
847 */ 674 */
848 smp_mb(); 675 smp_mb();
849 } 676 }
850 677
851 static inline struct pool_workqueue *work_stru !! 678 static void clear_work_data(struct work_struct *work)
852 { 679 {
853 return (struct pool_workqueue *)(data !! 680 smp_wmb(); /* see set_work_pool_and_clear_pending() */
>> 681 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
854 } 682 }
855 683
856 static struct pool_workqueue *get_work_pwq(str 684 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
857 { 685 {
858 unsigned long data = atomic_long_read( 686 unsigned long data = atomic_long_read(&work->data);
859 687
860 if (data & WORK_STRUCT_PWQ) 688 if (data & WORK_STRUCT_PWQ)
861 return work_struct_pwq(data); !! 689 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
862 else 690 else
863 return NULL; 691 return NULL;
864 } 692 }
865 693
866 /** 694 /**
867 * get_work_pool - return the worker_pool a gi 695 * get_work_pool - return the worker_pool a given work was associated with
868 * @work: the work item of interest 696 * @work: the work item of interest
869 * 697 *
870 * Pools are created and destroyed under wq_po 698 * Pools are created and destroyed under wq_pool_mutex, and allows read
871 * access under RCU read lock. As such, this !! 699 * access under sched-RCU read lock. As such, this function should be
872 * called under wq_pool_mutex or inside of a r !! 700 * called under wq_pool_mutex or with preemption disabled.
873 * 701 *
874 * All fields of the returned pool are accessi 702 * All fields of the returned pool are accessible as long as the above
875 * mentioned locking is in effect. If the ret 703 * mentioned locking is in effect. If the returned pool needs to be used
876 * beyond the critical section, the caller is 704 * beyond the critical section, the caller is responsible for ensuring the
877 * returned pool is and stays online. 705 * returned pool is and stays online.
878 * 706 *
879 * Return: The worker_pool @work was last asso 707 * Return: The worker_pool @work was last associated with. %NULL if none.
880 */ 708 */
881 static struct worker_pool *get_work_pool(struc 709 static struct worker_pool *get_work_pool(struct work_struct *work)
882 { 710 {
883 unsigned long data = atomic_long_read( 711 unsigned long data = atomic_long_read(&work->data);
884 int pool_id; 712 int pool_id;
885 713
886 assert_rcu_or_pool_mutex(); 714 assert_rcu_or_pool_mutex();
887 715
888 if (data & WORK_STRUCT_PWQ) 716 if (data & WORK_STRUCT_PWQ)
889 return work_struct_pwq(data)-> !! 717 return ((struct pool_workqueue *)
>> 718 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
890 719
891 pool_id = data >> WORK_OFFQ_POOL_SHIFT 720 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
892 if (pool_id == WORK_OFFQ_POOL_NONE) 721 if (pool_id == WORK_OFFQ_POOL_NONE)
893 return NULL; 722 return NULL;
894 723
895 return idr_find(&worker_pool_idr, pool 724 return idr_find(&worker_pool_idr, pool_id);
896 } 725 }
897 726
898 static unsigned long shift_and_mask(unsigned l !! 727 /**
>> 728 * get_work_pool_id - return the worker pool ID a given work is associated with
>> 729 * @work: the work item of interest
>> 730 *
>> 731 * Return: The worker_pool ID @work was last associated with.
>> 732 * %WORK_OFFQ_POOL_NONE if none.
>> 733 */
>> 734 static int get_work_pool_id(struct work_struct *work)
899 { 735 {
900 return (v >> shift) & ((1U << bits) - !! 736 unsigned long data = atomic_long_read(&work->data);
>> 737
>> 738 if (data & WORK_STRUCT_PWQ)
>> 739 return ((struct pool_workqueue *)
>> 740 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
>> 741
>> 742 return data >> WORK_OFFQ_POOL_SHIFT;
901 } 743 }
902 744
903 static void work_offqd_unpack(struct work_offq !! 745 static void mark_work_canceling(struct work_struct *work)
904 { 746 {
905 WARN_ON_ONCE(data & WORK_STRUCT_PWQ); !! 747 unsigned long pool_id = get_work_pool_id(work);
906 748
907 offqd->pool_id = shift_and_mask(data, !! 749 pool_id <<= WORK_OFFQ_POOL_SHIFT;
908 WORK_O !! 750 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 } 751 }
913 752
914 static unsigned long work_offqd_pack_flags(str !! 753 static bool work_is_canceling(struct work_struct *work)
915 { 754 {
916 return ((unsigned long)offqd->disable !! 755 unsigned long data = atomic_long_read(&work->data);
917 ((unsigned long)offqd->flags); !! 756
>> 757 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
918 } 758 }
919 759
920 /* 760 /*
921 * Policy functions. These define the policie 761 * Policy functions. These define the policies on how the global worker
922 * pools are managed. Unless noted otherwise, 762 * pools are managed. Unless noted otherwise, these functions assume that
923 * they're being called with pool->lock held. 763 * they're being called with pool->lock held.
924 */ 764 */
925 765
>> 766 static bool __need_more_worker(struct worker_pool *pool)
>> 767 {
>> 768 return !atomic_read(&pool->nr_running);
>> 769 }
>> 770
926 /* 771 /*
927 * Need to wake up a worker? Called from anyt 772 * Need to wake up a worker? Called from anything but currently
928 * running workers. 773 * running workers.
929 * 774 *
930 * Note that, because unbound workers never co 775 * Note that, because unbound workers never contribute to nr_running, this
931 * function will always return %true for unbou 776 * function will always return %true for unbound pools as long as the
932 * worklist isn't empty. 777 * worklist isn't empty.
933 */ 778 */
934 static bool need_more_worker(struct worker_poo 779 static bool need_more_worker(struct worker_pool *pool)
935 { 780 {
936 return !list_empty(&pool->worklist) && !! 781 return !list_empty(&pool->worklist) && __need_more_worker(pool);
937 } 782 }
938 783
939 /* Can I start working? Called from busy but 784 /* Can I start working? Called from busy but !running workers. */
940 static bool may_start_working(struct worker_po 785 static bool may_start_working(struct worker_pool *pool)
941 { 786 {
942 return pool->nr_idle; 787 return pool->nr_idle;
943 } 788 }
944 789
945 /* Do I need to keep working? Called from cur 790 /* Do I need to keep working? Called from currently running workers. */
946 static bool keep_working(struct worker_pool *p 791 static bool keep_working(struct worker_pool *pool)
947 { 792 {
948 return !list_empty(&pool->worklist) && !! 793 return !list_empty(&pool->worklist) &&
>> 794 atomic_read(&pool->nr_running) <= 1;
949 } 795 }
950 796
951 /* Do we need a new worker? Called from manag 797 /* Do we need a new worker? Called from manager. */
952 static bool need_to_create_worker(struct worke 798 static bool need_to_create_worker(struct worker_pool *pool)
953 { 799 {
954 return need_more_worker(pool) && !may_ 800 return need_more_worker(pool) && !may_start_working(pool);
955 } 801 }
956 802
957 /* Do we have too many workers and should some 803 /* Do we have too many workers and should some go away? */
958 static bool too_many_workers(struct worker_poo 804 static bool too_many_workers(struct worker_pool *pool)
959 { 805 {
960 bool managing = pool->flags & POOL_MAN 806 bool managing = pool->flags & POOL_MANAGER_ACTIVE;
961 int nr_idle = pool->nr_idle + managing 807 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
962 int nr_busy = pool->nr_workers - nr_id 808 int nr_busy = pool->nr_workers - nr_idle;
963 809
964 return nr_idle > 2 && (nr_idle - 2) * 810 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
965 } 811 }
966 812
>> 813 /*
>> 814 * Wake up functions.
>> 815 */
>> 816
>> 817 /* Return the first idle worker. Safe with preemption disabled */
>> 818 static struct worker *first_idle_worker(struct worker_pool *pool)
>> 819 {
>> 820 if (unlikely(list_empty(&pool->idle_list)))
>> 821 return NULL;
>> 822
>> 823 return list_first_entry(&pool->idle_list, struct worker, entry);
>> 824 }
>> 825
>> 826 /**
>> 827 * wake_up_worker - wake up an idle worker
>> 828 * @pool: worker pool to wake worker from
>> 829 *
>> 830 * Wake up the first idle worker of @pool.
>> 831 *
>> 832 * CONTEXT:
>> 833 * spin_lock_irq(pool->lock).
>> 834 */
>> 835 static void wake_up_worker(struct worker_pool *pool)
>> 836 {
>> 837 struct worker *worker = first_idle_worker(pool);
>> 838
>> 839 if (likely(worker))
>> 840 wake_up_process(worker->task);
>> 841 }
>> 842
>> 843 /**
>> 844 * wq_worker_waking_up - a worker is waking up
>> 845 * @task: task waking up
>> 846 * @cpu: CPU @task is waking up to
>> 847 *
>> 848 * This function is called during try_to_wake_up() when a worker is
>> 849 * being awoken.
>> 850 *
>> 851 * CONTEXT:
>> 852 * spin_lock_irq(rq->lock)
>> 853 */
>> 854 void wq_worker_waking_up(struct task_struct *task, int cpu)
>> 855 {
>> 856 struct worker *worker = kthread_data(task);
>> 857
>> 858 if (!(worker->flags & WORKER_NOT_RUNNING)) {
>> 859 WARN_ON_ONCE(worker->pool->cpu != cpu);
>> 860 atomic_inc(&worker->pool->nr_running);
>> 861 }
>> 862 }
>> 863
>> 864 /**
>> 865 * wq_worker_sleeping - a worker is going to sleep
>> 866 * @task: task going to sleep
>> 867 *
>> 868 * This function is called during schedule() when a busy worker is
>> 869 * going to sleep. Worker on the same cpu can be woken up by
>> 870 * returning pointer to its task.
>> 871 *
>> 872 * CONTEXT:
>> 873 * spin_lock_irq(rq->lock)
>> 874 *
>> 875 * Return:
>> 876 * Worker task on @cpu to wake up, %NULL if none.
>> 877 */
>> 878 struct task_struct *wq_worker_sleeping(struct task_struct *task)
>> 879 {
>> 880 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
>> 881 struct worker_pool *pool;
>> 882
>> 883 /*
>> 884 * Rescuers, which may not have all the fields set up like normal
>> 885 * workers, also reach here, let's not access anything before
>> 886 * checking NOT_RUNNING.
>> 887 */
>> 888 if (worker->flags & WORKER_NOT_RUNNING)
>> 889 return NULL;
>> 890
>> 891 pool = worker->pool;
>> 892
>> 893 /* this can only happen on the local cpu */
>> 894 if (WARN_ON_ONCE(pool->cpu != raw_smp_processor_id()))
>> 895 return NULL;
>> 896
>> 897 /*
>> 898 * The counterpart of the following dec_and_test, implied mb,
>> 899 * worklist not empty test sequence is in insert_work().
>> 900 * Please read comment there.
>> 901 *
>> 902 * NOT_RUNNING is clear. This means that we're bound to and
>> 903 * running on the local cpu w/ rq lock held and preemption
>> 904 * disabled, which in turn means that none else could be
>> 905 * manipulating idle_list, so dereferencing idle_list without pool
>> 906 * lock is safe.
>> 907 */
>> 908 if (atomic_dec_and_test(&pool->nr_running) &&
>> 909 !list_empty(&pool->worklist))
>> 910 to_wakeup = first_idle_worker(pool);
>> 911 return to_wakeup ? to_wakeup->task : NULL;
>> 912 }
>> 913
967 /** 914 /**
968 * worker_set_flags - set worker flags and adj 915 * worker_set_flags - set worker flags and adjust nr_running accordingly
969 * @worker: self 916 * @worker: self
970 * @flags: flags to set 917 * @flags: flags to set
971 * 918 *
972 * Set @flags in @worker->flags and adjust nr_ 919 * Set @flags in @worker->flags and adjust nr_running accordingly.
>> 920 *
>> 921 * CONTEXT:
>> 922 * spin_lock_irq(pool->lock)
973 */ 923 */
974 static inline void worker_set_flags(struct wor 924 static inline void worker_set_flags(struct worker *worker, unsigned int flags)
975 { 925 {
976 struct worker_pool *pool = worker->poo 926 struct worker_pool *pool = worker->pool;
977 927
978 lockdep_assert_held(&pool->lock); !! 928 WARN_ON_ONCE(worker->task != current);
979 929
980 /* If transitioning into NOT_RUNNING, 930 /* If transitioning into NOT_RUNNING, adjust nr_running. */
981 if ((flags & WORKER_NOT_RUNNING) && 931 if ((flags & WORKER_NOT_RUNNING) &&
982 !(worker->flags & WORKER_NOT_RUNNI 932 !(worker->flags & WORKER_NOT_RUNNING)) {
983 pool->nr_running--; !! 933 atomic_dec(&pool->nr_running);
984 } 934 }
985 935
986 worker->flags |= flags; 936 worker->flags |= flags;
987 } 937 }
988 938
989 /** 939 /**
990 * worker_clr_flags - clear worker flags and a 940 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
991 * @worker: self 941 * @worker: self
992 * @flags: flags to clear 942 * @flags: flags to clear
993 * 943 *
994 * Clear @flags in @worker->flags and adjust n 944 * Clear @flags in @worker->flags and adjust nr_running accordingly.
>> 945 *
>> 946 * CONTEXT:
>> 947 * spin_lock_irq(pool->lock)
995 */ 948 */
996 static inline void worker_clr_flags(struct wor 949 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
997 { 950 {
998 struct worker_pool *pool = worker->poo 951 struct worker_pool *pool = worker->pool;
999 unsigned int oflags = worker->flags; 952 unsigned int oflags = worker->flags;
1000 953
1001 lockdep_assert_held(&pool->lock); !! 954 WARN_ON_ONCE(worker->task != current);
1002 955
1003 worker->flags &= ~flags; 956 worker->flags &= ~flags;
1004 957
1005 /* 958 /*
1006 * If transitioning out of NOT_RUNNIN 959 * If transitioning out of NOT_RUNNING, increment nr_running. Note
1007 * that the nested NOT_RUNNING is not 960 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
1008 * of multiple flags, not a single fl 961 * of multiple flags, not a single flag.
1009 */ 962 */
1010 if ((flags & WORKER_NOT_RUNNING) && ( 963 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
1011 if (!(worker->flags & WORKER_ 964 if (!(worker->flags & WORKER_NOT_RUNNING))
1012 pool->nr_running++; !! 965 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 } 966 }
1077 967
1078 /** 968 /**
1079 * find_worker_executing_work - find worker w 969 * find_worker_executing_work - find worker which is executing a work
1080 * @pool: pool of interest 970 * @pool: pool of interest
1081 * @work: work to find worker for 971 * @work: work to find worker for
1082 * 972 *
1083 * Find a worker which is executing @work on 973 * Find a worker which is executing @work on @pool by searching
1084 * @pool->busy_hash which is keyed by the add 974 * @pool->busy_hash which is keyed by the address of @work. For a worker
1085 * to match, its current execution should mat 975 * to match, its current execution should match the address of @work and
1086 * its work function. This is to avoid unwan 976 * its work function. This is to avoid unwanted dependency between
1087 * unrelated work executions through a work i 977 * unrelated work executions through a work item being recycled while still
1088 * being executed. 978 * being executed.
1089 * 979 *
1090 * This is a bit tricky. A work item may be 980 * This is a bit tricky. A work item may be freed once its execution
1091 * starts and nothing prevents the freed area 981 * starts and nothing prevents the freed area from being recycled for
1092 * another work item. If the same work item 982 * another work item. If the same work item address ends up being reused
1093 * before the original execution finishes, wo 983 * before the original execution finishes, workqueue will identify the
1094 * recycled work item as currently executing 984 * recycled work item as currently executing and make it wait until the
1095 * current execution finishes, introducing an 985 * current execution finishes, introducing an unwanted dependency.
1096 * 986 *
1097 * This function checks the work item address 987 * This function checks the work item address and work function to avoid
1098 * false positives. Note that this isn't com 988 * false positives. Note that this isn't complete as one may construct a
1099 * work function which can introduce dependen 989 * work function which can introduce dependency onto itself through a
1100 * recycled work item. Well, if somebody wan 990 * recycled work item. Well, if somebody wants to shoot oneself in the
1101 * foot that badly, there's only so much we c 991 * foot that badly, there's only so much we can do, and if such deadlock
1102 * actually occurs, it should be easy to loca 992 * actually occurs, it should be easy to locate the culprit work function.
1103 * 993 *
1104 * CONTEXT: 994 * CONTEXT:
1105 * raw_spin_lock_irq(pool->lock). !! 995 * spin_lock_irq(pool->lock).
1106 * 996 *
1107 * Return: 997 * Return:
1108 * Pointer to worker which is executing @work 998 * Pointer to worker which is executing @work if found, %NULL
1109 * otherwise. 999 * otherwise.
1110 */ 1000 */
1111 static struct worker *find_worker_executing_w 1001 static struct worker *find_worker_executing_work(struct worker_pool *pool,
1112 1002 struct work_struct *work)
1113 { 1003 {
1114 struct worker *worker; 1004 struct worker *worker;
1115 1005
1116 hash_for_each_possible(pool->busy_has 1006 hash_for_each_possible(pool->busy_hash, worker, hentry,
1117 (unsigned long 1007 (unsigned long)work)
1118 if (worker->current_work == w 1008 if (worker->current_work == work &&
1119 worker->current_func == w 1009 worker->current_func == work->func)
1120 return worker; 1010 return worker;
1121 1011
1122 return NULL; 1012 return NULL;
1123 } 1013 }
1124 1014
1125 /** 1015 /**
1126 * move_linked_works - move linked works to a 1016 * move_linked_works - move linked works to a list
1127 * @work: start of series of works to be sche 1017 * @work: start of series of works to be scheduled
1128 * @head: target list to append @work to 1018 * @head: target list to append @work to
1129 * @nextp: out parameter for nested worklist 1019 * @nextp: out parameter for nested worklist walking
1130 * 1020 *
1131 * Schedule linked works starting from @work !! 1021 * Schedule linked works starting from @work to @head. Work series to
1132 * scheduled starts at @work and includes any !! 1022 * be scheduled starts at @work and includes any consecutive work with
1133 * WORK_STRUCT_LINKED set in its predecessor. !! 1023 * WORK_STRUCT_LINKED set in its predecessor.
1134 * @nextp. !! 1024 *
>> 1025 * If @nextp is not NULL, it's updated to point to the next work of
>> 1026 * the last scheduled work. This allows move_linked_works() to be
>> 1027 * nested inside outer list_for_each_entry_safe().
1135 * 1028 *
1136 * CONTEXT: 1029 * CONTEXT:
1137 * raw_spin_lock_irq(pool->lock). !! 1030 * spin_lock_irq(pool->lock).
1138 */ 1031 */
1139 static void move_linked_works(struct work_str 1032 static void move_linked_works(struct work_struct *work, struct list_head *head,
1140 struct work_str 1033 struct work_struct **nextp)
1141 { 1034 {
1142 struct work_struct *n; 1035 struct work_struct *n;
1143 1036
1144 /* 1037 /*
1145 * Linked worklist will always end be 1038 * Linked worklist will always end before the end of the list,
1146 * use NULL for list head. 1039 * use NULL for list head.
1147 */ 1040 */
1148 list_for_each_entry_safe_from(work, n 1041 list_for_each_entry_safe_from(work, n, NULL, entry) {
1149 list_move_tail(&work->entry, 1042 list_move_tail(&work->entry, head);
1150 if (!(*work_data_bits(work) & 1043 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1151 break; 1044 break;
1152 } 1045 }
1153 1046
1154 /* 1047 /*
1155 * If we're already inside safe list 1048 * If we're already inside safe list traversal and have moved
1156 * multiple works to the scheduled qu 1049 * multiple works to the scheduled queue, the next position
1157 * needs to be updated. 1050 * needs to be updated.
1158 */ 1051 */
1159 if (nextp) 1052 if (nextp)
1160 *nextp = n; 1053 *nextp = n;
1161 } 1054 }
1162 1055
1163 /** 1056 /**
1164 * assign_work - assign a work item and its l <<
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 1057 * get_pwq - get an extra reference on the specified pool_workqueue
1620 * @pwq: pool_workqueue to get 1058 * @pwq: pool_workqueue to get
1621 * 1059 *
1622 * Obtain an extra reference on @pwq. The ca 1060 * Obtain an extra reference on @pwq. The caller should guarantee that
1623 * @pwq has positive refcnt and be holding th 1061 * @pwq has positive refcnt and be holding the matching pool->lock.
1624 */ 1062 */
1625 static void get_pwq(struct pool_workqueue *pw 1063 static void get_pwq(struct pool_workqueue *pwq)
1626 { 1064 {
1627 lockdep_assert_held(&pwq->pool->lock) 1065 lockdep_assert_held(&pwq->pool->lock);
1628 WARN_ON_ONCE(pwq->refcnt <= 0); 1066 WARN_ON_ONCE(pwq->refcnt <= 0);
1629 pwq->refcnt++; 1067 pwq->refcnt++;
1630 } 1068 }
1631 1069
1632 /** 1070 /**
1633 * put_pwq - put a pool_workqueue reference 1071 * put_pwq - put a pool_workqueue reference
1634 * @pwq: pool_workqueue to put 1072 * @pwq: pool_workqueue to put
1635 * 1073 *
1636 * Drop a reference of @pwq. If its refcnt r 1074 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1637 * destruction. The caller should be holding 1075 * destruction. The caller should be holding the matching pool->lock.
1638 */ 1076 */
1639 static void put_pwq(struct pool_workqueue *pw 1077 static void put_pwq(struct pool_workqueue *pwq)
1640 { 1078 {
1641 lockdep_assert_held(&pwq->pool->lock) 1079 lockdep_assert_held(&pwq->pool->lock);
1642 if (likely(--pwq->refcnt)) 1080 if (likely(--pwq->refcnt))
1643 return; 1081 return;
>> 1082 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
>> 1083 return;
1644 /* 1084 /*
1645 * @pwq can't be released under pool- !! 1085 * @pwq can't be released under pool->lock, bounce to
1646 * kthread_worker to avoid A-A deadlo !! 1086 * pwq_unbound_release_workfn(). This never recurses on the same
>> 1087 * pool->lock as this path is taken only for unbound workqueues and
>> 1088 * the release work item is scheduled on a per-cpu workqueue. To
>> 1089 * avoid lockdep warning, unbound pool->locks are given lockdep
>> 1090 * subclass of 1 in get_unbound_pool().
1647 */ 1091 */
1648 kthread_queue_work(pwq_release_worker !! 1092 schedule_work(&pwq->unbound_release_work);
1649 } 1093 }
1650 1094
1651 /** 1095 /**
1652 * put_pwq_unlocked - put_pwq() with surround 1096 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1653 * @pwq: pool_workqueue to put (can be %NULL) 1097 * @pwq: pool_workqueue to put (can be %NULL)
1654 * 1098 *
1655 * put_pwq() with locking. This function als 1099 * put_pwq() with locking. This function also allows %NULL @pwq.
1656 */ 1100 */
1657 static void put_pwq_unlocked(struct pool_work 1101 static void put_pwq_unlocked(struct pool_workqueue *pwq)
1658 { 1102 {
1659 if (pwq) { 1103 if (pwq) {
1660 /* 1104 /*
1661 * As both pwqs and pools are !! 1105 * As both pwqs and pools are sched-RCU protected, the
1662 * following lock operations 1106 * following lock operations are safe.
1663 */ 1107 */
1664 raw_spin_lock_irq(&pwq->pool- !! 1108 spin_lock_irq(&pwq->pool->lock);
1665 put_pwq(pwq); 1109 put_pwq(pwq);
1666 raw_spin_unlock_irq(&pwq->poo !! 1110 spin_unlock_irq(&pwq->pool->lock);
1667 } 1111 }
1668 } 1112 }
1669 1113
1670 static bool pwq_is_empty(struct pool_workqueu !! 1114 static void pwq_activate_delayed_work(struct work_struct *work)
1671 { <<
1672 return !pwq->nr_active && list_empty( <<
1673 } <<
1674 <<
1675 static void __pwq_activate_work(struct pool_w <<
1676 struct work_s <<
1677 { 1115 {
1678 unsigned long *wdb = work_data_bits(w !! 1116 struct pool_workqueue *pwq = get_work_pwq(work);
1679 1117
1680 WARN_ON_ONCE(!(*wdb & WORK_STRUCT_INA <<
1681 trace_workqueue_activate_work(work); 1118 trace_workqueue_activate_work(work);
1682 if (list_empty(&pwq->pool->worklist)) 1119 if (list_empty(&pwq->pool->worklist))
1683 pwq->pool->watchdog_ts = jiff 1120 pwq->pool->watchdog_ts = jiffies;
1684 move_linked_works(work, &pwq->pool->w 1121 move_linked_works(work, &pwq->pool->worklist, NULL);
1685 __clear_bit(WORK_STRUCT_INACTIVE_BIT, !! 1122 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
>> 1123 pwq->nr_active++;
1686 } 1124 }
1687 1125
1688 static bool tryinc_node_nr_active(struct wq_n !! 1126 static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1689 { 1127 {
1690 int max = READ_ONCE(nna->max); !! 1128 struct work_struct *work = list_first_entry(&pwq->delayed_works,
1691 !! 1129 struct work_struct, entry);
1692 while (true) { <<
1693 int old, tmp; <<
1694 1130
1695 old = atomic_read(&nna->nr); !! 1131 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 } 1132 }
1974 1133
1975 /** 1134 /**
1976 * pwq_dec_nr_in_flight - decrement pwq's nr_ 1135 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1977 * @pwq: pwq of interest 1136 * @pwq: pwq of interest
1978 * @work_data: work_data of work which left t !! 1137 * @color: color of work which left the queue
1979 * 1138 *
1980 * A work either has completed or is removed 1139 * A work either has completed or is removed from pending queue,
1981 * decrement nr_in_flight of its pwq and hand 1140 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1982 * 1141 *
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: 1142 * CONTEXT:
1989 * raw_spin_lock_irq(pool->lock). !! 1143 * spin_lock_irq(pool->lock).
1990 */ 1144 */
1991 static void pwq_dec_nr_in_flight(struct pool_ !! 1145 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1992 { 1146 {
1993 int color = get_work_color(work_data) !! 1147 /* uncolored work items don't participate in flushing or nr_active */
1994 !! 1148 if (color == WORK_NO_COLOR)
1995 if (!(work_data & WORK_STRUCT_INACTIV !! 1149 goto out_put;
1996 pwq_dec_nr_active(pwq); <<
1997 1150
1998 pwq->nr_in_flight[color]--; 1151 pwq->nr_in_flight[color]--;
1999 1152
>> 1153 pwq->nr_active--;
>> 1154 if (!list_empty(&pwq->delayed_works)) {
>> 1155 /* one down, submit a delayed one */
>> 1156 if (pwq->nr_active < pwq->max_active)
>> 1157 pwq_activate_first_delayed(pwq);
>> 1158 }
>> 1159
2000 /* is flush in progress and are we at 1160 /* is flush in progress and are we at the flushing tip? */
2001 if (likely(pwq->flush_color != color) 1161 if (likely(pwq->flush_color != color))
2002 goto out_put; 1162 goto out_put;
2003 1163
2004 /* are there still in-flight works? * 1164 /* are there still in-flight works? */
2005 if (pwq->nr_in_flight[color]) 1165 if (pwq->nr_in_flight[color])
2006 goto out_put; 1166 goto out_put;
2007 1167
2008 /* this pwq is done, clear flush_colo 1168 /* this pwq is done, clear flush_color */
2009 pwq->flush_color = -1; 1169 pwq->flush_color = -1;
2010 1170
2011 /* 1171 /*
2012 * If this was the last pwq, wake up 1172 * If this was the last pwq, wake up the first flusher. It
2013 * will handle the rest. 1173 * will handle the rest.
2014 */ 1174 */
2015 if (atomic_dec_and_test(&pwq->wq->nr_ 1175 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
2016 complete(&pwq->wq->first_flus 1176 complete(&pwq->wq->first_flusher->done);
2017 out_put: 1177 out_put:
2018 put_pwq(pwq); 1178 put_pwq(pwq);
2019 } 1179 }
2020 1180
2021 /** 1181 /**
2022 * try_to_grab_pending - steal work item from 1182 * try_to_grab_pending - steal work item from worklist and disable irq
2023 * @work: work item to steal 1183 * @work: work item to steal
2024 * @cflags: %WORK_CANCEL_ flags !! 1184 * @is_dwork: @work is a delayed_work
2025 * @irq_flags: place to store irq state !! 1185 * @flags: place to store irq state
2026 * 1186 *
2027 * Try to grab PENDING bit of @work. This fu 1187 * Try to grab PENDING bit of @work. This function can handle @work in any
2028 * stable state - idle, on timer or on workli 1188 * stable state - idle, on timer or on worklist.
2029 * 1189 *
2030 * Return: 1190 * Return:
2031 * <<
2032 * ======== ============================= <<
2033 * 1 if @work was pending and we s 1191 * 1 if @work was pending and we successfully stole PENDING
2034 * 0 if @work was idle and we clai 1192 * 0 if @work was idle and we claimed PENDING
2035 * -EAGAIN if PENDING couldn't be grabbe 1193 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
2036 * ======== ============================= !! 1194 * -ENOENT if someone else is canceling @work, this state may persist
>> 1195 * for arbitrarily long
2037 * 1196 *
2038 * Note: 1197 * Note:
2039 * On >= 0 return, the caller owns @work's PE 1198 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
2040 * interrupted while holding PENDING and @wor 1199 * interrupted while holding PENDING and @work off queue, irq must be
2041 * disabled on entry. This, combined with de 1200 * disabled on entry. This, combined with delayed_work->timer being
2042 * irqsafe, ensures that we return -EAGAIN fo 1201 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
2043 * 1202 *
2044 * On successful return, >= 0, irq is disable 1203 * On successful return, >= 0, irq is disabled and the caller is
2045 * responsible for releasing it using local_i !! 1204 * responsible for releasing it using local_irq_restore(*@flags).
2046 * 1205 *
2047 * This function is safe to call from any con 1206 * This function is safe to call from any context including IRQ handler.
2048 */ 1207 */
2049 static int try_to_grab_pending(struct work_st !! 1208 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
2050 unsigned long !! 1209 unsigned long *flags)
2051 { 1210 {
2052 struct worker_pool *pool; 1211 struct worker_pool *pool;
2053 struct pool_workqueue *pwq; 1212 struct pool_workqueue *pwq;
2054 1213
2055 local_irq_save(*irq_flags); !! 1214 local_irq_save(*flags);
2056 1215
2057 /* try to steal the timer if it exist 1216 /* try to steal the timer if it exists */
2058 if (cflags & WORK_CANCEL_DELAYED) { !! 1217 if (is_dwork) {
2059 struct delayed_work *dwork = 1218 struct delayed_work *dwork = to_delayed_work(work);
2060 1219
2061 /* 1220 /*
2062 * dwork->timer is irqsafe. 1221 * dwork->timer is irqsafe. If del_timer() fails, it's
2063 * guaranteed that the timer 1222 * guaranteed that the timer is not queued anywhere and not
2064 * running on the local CPU. 1223 * running on the local CPU.
2065 */ 1224 */
2066 if (likely(del_timer(&dwork-> 1225 if (likely(del_timer(&dwork->timer)))
2067 return 1; 1226 return 1;
2068 } 1227 }
2069 1228
2070 /* try to claim PENDING the normal wa 1229 /* try to claim PENDING the normal way */
2071 if (!test_and_set_bit(WORK_STRUCT_PEN 1230 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
2072 return 0; 1231 return 0;
2073 1232
2074 rcu_read_lock(); <<
2075 /* 1233 /*
2076 * The queueing is in progress, or it 1234 * The queueing is in progress, or it is already queued. Try to
2077 * steal it from ->worklist without c 1235 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
2078 */ 1236 */
2079 pool = get_work_pool(work); 1237 pool = get_work_pool(work);
2080 if (!pool) 1238 if (!pool)
2081 goto fail; 1239 goto fail;
2082 1240
2083 raw_spin_lock(&pool->lock); !! 1241 spin_lock(&pool->lock);
2084 /* 1242 /*
2085 * work->data is guaranteed to point 1243 * work->data is guaranteed to point to pwq only while the work
2086 * item is queued on pwq->wq, and bot 1244 * item is queued on pwq->wq, and both updating work->data to point
2087 * to pwq on queueing and to pool on 1245 * to pwq on queueing and to pool on dequeueing are done under
2088 * pwq->pool->lock. This in turn gua 1246 * pwq->pool->lock. This in turn guarantees that, if work->data
2089 * points to pwq which is associated 1247 * points to pwq which is associated with a locked pool, the work
2090 * item is currently queued on that p 1248 * item is currently queued on that pool.
2091 */ 1249 */
2092 pwq = get_work_pwq(work); 1250 pwq = get_work_pwq(work);
2093 if (pwq && pwq->pool == pool) { 1251 if (pwq && pwq->pool == pool) {
2094 unsigned long work_data = *wo <<
2095 <<
2096 debug_work_deactivate(work); 1252 debug_work_deactivate(work);
2097 1253
2098 /* 1254 /*
2099 * A cancelable inactive work !! 1255 * A delayed work item cannot be grabbed directly because
2100 * pwq->inactive_works since !! 1256 * it might have linked NO_COLOR work items which, if left
2101 * canceled (see the comments !! 1257 * on the delayed_list, will confuse pwq->nr_active
2102 * !! 1258 * management later on and cause stall. Make sure the work
2103 * An inactive work item cann !! 1259 * 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 */ 1260 */
2112 if (work_data & WORK_STRUCT_I !! 1261 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
2113 move_linked_works(wor !! 1262 pwq_activate_delayed_work(work);
2114 1263
2115 list_del_init(&work->entry); 1264 list_del_init(&work->entry);
>> 1265 pwq_dec_nr_in_flight(pwq, get_work_color(work));
2116 1266
2117 /* !! 1267 /* work->data points to pwq iff queued, point to pool */
2118 * work->data points to pwq i !! 1268 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 1269
2127 raw_spin_unlock(&pool->lock); !! 1270 spin_unlock(&pool->lock);
2128 rcu_read_unlock(); <<
2129 return 1; 1271 return 1;
2130 } 1272 }
2131 raw_spin_unlock(&pool->lock); !! 1273 spin_unlock(&pool->lock);
2132 fail: 1274 fail:
2133 rcu_read_unlock(); !! 1275 local_irq_restore(*flags);
2134 local_irq_restore(*irq_flags); !! 1276 if (work_is_canceling(work))
>> 1277 return -ENOENT;
>> 1278 cpu_relax();
2135 return -EAGAIN; 1279 return -EAGAIN;
2136 } 1280 }
2137 1281
2138 /** 1282 /**
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 1283 * insert_work - insert a work into a pool
2168 * @pwq: pwq @work belongs to 1284 * @pwq: pwq @work belongs to
2169 * @work: work to insert 1285 * @work: work to insert
2170 * @head: insertion point 1286 * @head: insertion point
2171 * @extra_flags: extra WORK_STRUCT_* flags to 1287 * @extra_flags: extra WORK_STRUCT_* flags to set
2172 * 1288 *
2173 * Insert @work which belongs to @pwq after @ 1289 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
2174 * work_struct flags. 1290 * work_struct flags.
2175 * 1291 *
2176 * CONTEXT: 1292 * CONTEXT:
2177 * raw_spin_lock_irq(pool->lock). !! 1293 * spin_lock_irq(pool->lock).
2178 */ 1294 */
2179 static void insert_work(struct pool_workqueue 1295 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
2180 struct list_head *hea 1296 struct list_head *head, unsigned int extra_flags)
2181 { 1297 {
2182 debug_work_activate(work); !! 1298 struct worker_pool *pool = pwq->pool;
2183 <<
2184 /* record the work call stack in orde <<
2185 kasan_record_aux_stack_noalloc(work); <<
2186 1299
2187 /* we own @work, set data and link */ 1300 /* we own @work, set data and link */
2188 set_work_pwq(work, pwq, extra_flags); 1301 set_work_pwq(work, pwq, extra_flags);
2189 list_add_tail(&work->entry, head); 1302 list_add_tail(&work->entry, head);
2190 get_pwq(pwq); 1303 get_pwq(pwq);
>> 1304
>> 1305 /*
>> 1306 * Ensure either wq_worker_sleeping() sees the above
>> 1307 * list_add_tail() or we see zero nr_running to avoid workers lying
>> 1308 * around lazily while there are works to be processed.
>> 1309 */
>> 1310 smp_mb();
>> 1311
>> 1312 if (__need_more_worker(pool))
>> 1313 wake_up_worker(pool);
2191 } 1314 }
2192 1315
2193 /* 1316 /*
2194 * Test whether @work is being queued from an 1317 * Test whether @work is being queued from another work executing on the
2195 * same workqueue. 1318 * same workqueue.
2196 */ 1319 */
2197 static bool is_chained_work(struct workqueue_ 1320 static bool is_chained_work(struct workqueue_struct *wq)
2198 { 1321 {
2199 struct worker *worker; 1322 struct worker *worker;
2200 1323
2201 worker = current_wq_worker(); 1324 worker = current_wq_worker();
2202 /* 1325 /*
2203 * Return %true iff I'm a worker exec !! 1326 * Return %true iff I'm a worker execuing a work item on @wq. If
2204 * I'm @worker, it's safe to derefere 1327 * I'm @worker, it's safe to dereference it without locking.
2205 */ 1328 */
2206 return worker && worker->current_pwq- 1329 return worker && worker->current_pwq->wq == wq;
2207 } 1330 }
2208 1331
2209 /* 1332 /*
2210 * When queueing an unbound work item to a wq 1333 * When queueing an unbound work item to a wq, prefer local CPU if allowed
2211 * by wq_unbound_cpumask. Otherwise, round r 1334 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
2212 * avoid perturbing sensitive tasks. 1335 * avoid perturbing sensitive tasks.
2213 */ 1336 */
2214 static int wq_select_unbound_cpu(int cpu) 1337 static int wq_select_unbound_cpu(int cpu)
2215 { 1338 {
>> 1339 static bool printed_dbg_warning;
2216 int new_cpu; 1340 int new_cpu;
2217 1341
2218 if (likely(!wq_debug_force_rr_cpu)) { 1342 if (likely(!wq_debug_force_rr_cpu)) {
2219 if (cpumask_test_cpu(cpu, wq_ 1343 if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
2220 return cpu; 1344 return cpu;
2221 } else { !! 1345 } else if (!printed_dbg_warning) {
2222 pr_warn_once("workqueue: roun !! 1346 pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n");
>> 1347 printed_dbg_warning = true;
2223 } 1348 }
2224 1349
>> 1350 if (cpumask_empty(wq_unbound_cpumask))
>> 1351 return cpu;
>> 1352
2225 new_cpu = __this_cpu_read(wq_rr_cpu_l 1353 new_cpu = __this_cpu_read(wq_rr_cpu_last);
2226 new_cpu = cpumask_next_and(new_cpu, w 1354 new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
2227 if (unlikely(new_cpu >= nr_cpu_ids)) 1355 if (unlikely(new_cpu >= nr_cpu_ids)) {
2228 new_cpu = cpumask_first_and(w 1356 new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
2229 if (unlikely(new_cpu >= nr_cp 1357 if (unlikely(new_cpu >= nr_cpu_ids))
2230 return cpu; 1358 return cpu;
2231 } 1359 }
2232 __this_cpu_write(wq_rr_cpu_last, new_ 1360 __this_cpu_write(wq_rr_cpu_last, new_cpu);
2233 1361
2234 return new_cpu; 1362 return new_cpu;
2235 } 1363 }
2236 1364
2237 static void __queue_work(int cpu, struct work 1365 static void __queue_work(int cpu, struct workqueue_struct *wq,
2238 struct work_struct * 1366 struct work_struct *work)
2239 { 1367 {
2240 struct pool_workqueue *pwq; 1368 struct pool_workqueue *pwq;
2241 struct worker_pool *last_pool, *pool; !! 1369 struct worker_pool *last_pool;
>> 1370 struct list_head *worklist;
2242 unsigned int work_flags; 1371 unsigned int work_flags;
2243 unsigned int req_cpu = cpu; 1372 unsigned int req_cpu = cpu;
2244 1373
2245 /* 1374 /*
2246 * While a work item is PENDING && of 1375 * While a work item is PENDING && off queue, a task trying to
2247 * steal the PENDING will busy-loop w 1376 * steal the PENDING will busy-loop waiting for it to either get
2248 * queued or lose PENDING. Grabbing 1377 * queued or lose PENDING. Grabbing PENDING and queueing should
2249 * happen with IRQ disabled. 1378 * happen with IRQ disabled.
2250 */ 1379 */
2251 lockdep_assert_irqs_disabled(); 1380 lockdep_assert_irqs_disabled();
2252 1381
2253 /* !! 1382 debug_work_activate(work);
2254 * For a draining wq, only works from !! 1383
2255 * allowed. The __WQ_DESTROYING helps !! 1384 /* if draining, only works from the same workqueue are allowed */
2256 * queues a new work item to a wq aft !! 1385 if (unlikely(wq->flags & __WQ_DRAINING) &&
2257 */ !! 1386 WARN_ON_ONCE(!is_chained_work(wq)))
2258 if (unlikely(wq->flags & (__WQ_DESTRO <<
2259 WARN_ON_ONCE(!is_chained <<
2260 return; 1387 return;
2261 rcu_read_lock(); <<
2262 retry: 1388 retry:
2263 /* pwq which will be used unless @wor !! 1389 if (req_cpu == WORK_CPU_UNBOUND)
2264 if (req_cpu == WORK_CPU_UNBOUND) { !! 1390 cpu = wq_select_unbound_cpu(raw_smp_processor_id());
2265 if (wq->flags & WQ_UNBOUND) <<
2266 cpu = wq_select_unbou <<
2267 else <<
2268 cpu = raw_smp_process <<
2269 } <<
2270 1391
2271 pwq = rcu_dereference(*per_cpu_ptr(wq !! 1392 /* pwq which will be used unless @work is executing elsewhere */
2272 pool = pwq->pool; !! 1393 if (!(wq->flags & WQ_UNBOUND))
>> 1394 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
>> 1395 else
>> 1396 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
2273 1397
2274 /* 1398 /*
2275 * If @work was previously on a diffe 1399 * If @work was previously on a different pool, it might still be
2276 * running there, in which case the w 1400 * running there, in which case the work needs to be queued on that
2277 * pool to guarantee non-reentrancy. 1401 * 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 */ 1402 */
2283 last_pool = get_work_pool(work); 1403 last_pool = get_work_pool(work);
2284 if (last_pool && last_pool != pool && !! 1404 if (last_pool && last_pool != pwq->pool) {
2285 struct worker *worker; 1405 struct worker *worker;
2286 1406
2287 raw_spin_lock(&last_pool->loc !! 1407 spin_lock(&last_pool->lock);
2288 1408
2289 worker = find_worker_executin 1409 worker = find_worker_executing_work(last_pool, work);
2290 1410
2291 if (worker && worker->current 1411 if (worker && worker->current_pwq->wq == wq) {
2292 pwq = worker->current 1412 pwq = worker->current_pwq;
2293 pool = pwq->pool; <<
2294 WARN_ON_ONCE(pool != <<
2295 } else { 1413 } else {
2296 /* meh... not running 1414 /* meh... not running there, queue here */
2297 raw_spin_unlock(&last !! 1415 spin_unlock(&last_pool->lock);
2298 raw_spin_lock(&pool-> !! 1416 spin_lock(&pwq->pool->lock);
2299 } 1417 }
2300 } else { 1418 } else {
2301 raw_spin_lock(&pool->lock); !! 1419 spin_lock(&pwq->pool->lock);
2302 } 1420 }
2303 1421
2304 /* 1422 /*
2305 * pwq is determined and locked. For !! 1423 * pwq is determined and locked. For unbound pools, we could have
2306 * with pwq release and it could alre !! 1424 * raced with pwq release and it could already be dead. If its
2307 * repeat pwq selection. Note that un !! 1425 * refcnt is zero, repeat pwq selection. Note that pwqs never die
2308 * another pwq replacing it in cpu_pw !! 1426 * without another pwq replacing it in the numa_pwq_tbl or while
2309 * on it, so the retrying is guarante !! 1427 * work items are executing on it, so the retrying is guaranteed to
>> 1428 * make forward-progress.
2310 */ 1429 */
2311 if (unlikely(!pwq->refcnt)) { 1430 if (unlikely(!pwq->refcnt)) {
2312 if (wq->flags & WQ_UNBOUND) { 1431 if (wq->flags & WQ_UNBOUND) {
2313 raw_spin_unlock(&pool !! 1432 spin_unlock(&pwq->pool->lock);
2314 cpu_relax(); 1433 cpu_relax();
2315 goto retry; 1434 goto retry;
2316 } 1435 }
2317 /* oops */ 1436 /* oops */
2318 WARN_ONCE(true, "workqueue: p 1437 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
2319 wq->name, cpu); 1438 wq->name, cpu);
2320 } 1439 }
2321 1440
2322 /* pwq determined, queue */ 1441 /* pwq determined, queue */
2323 trace_workqueue_queue_work(req_cpu, p 1442 trace_workqueue_queue_work(req_cpu, pwq, work);
2324 1443
2325 if (WARN_ON(!list_empty(&work->entry) !! 1444 if (WARN_ON(!list_empty(&work->entry))) {
2326 goto out; !! 1445 spin_unlock(&pwq->pool->lock);
>> 1446 return;
>> 1447 }
2327 1448
2328 pwq->nr_in_flight[pwq->work_color]++; 1449 pwq->nr_in_flight[pwq->work_color]++;
2329 work_flags = work_color_to_flags(pwq- 1450 work_flags = work_color_to_flags(pwq->work_color);
2330 1451
2331 /* !! 1452 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 1453 trace_workqueue_activate_work(work);
2341 insert_work(pwq, work, &pool- !! 1454 pwq->nr_active++;
2342 kick_pool(pool); !! 1455 worklist = &pwq->pool->worklist;
>> 1456 if (list_empty(worklist))
>> 1457 pwq->pool->watchdog_ts = jiffies;
2343 } else { 1458 } else {
2344 work_flags |= WORK_STRUCT_INA !! 1459 work_flags |= WORK_STRUCT_DELAYED;
2345 insert_work(pwq, work, &pwq-> !! 1460 worklist = &pwq->delayed_works;
2346 } 1461 }
2347 1462
2348 out: !! 1463 insert_work(pwq, work, worklist, work_flags);
2349 raw_spin_unlock(&pool->lock); <<
2350 rcu_read_unlock(); <<
2351 } <<
2352 <<
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 1464
2362 work_offqd_unpack(&offqd, data); !! 1465 spin_unlock(&pwq->pool->lock);
2363 set_work_pool_and_clear_pending(work, <<
2364 work_ <<
2365 return true; <<
2366 } 1466 }
2367 1467
2368 /** 1468 /**
2369 * queue_work_on - queue work on specific cpu 1469 * queue_work_on - queue work on specific cpu
2370 * @cpu: CPU number to execute work on 1470 * @cpu: CPU number to execute work on
2371 * @wq: workqueue to use 1471 * @wq: workqueue to use
2372 * @work: work to queue 1472 * @work: work to queue
2373 * 1473 *
2374 * We queue the work to a specific CPU, the c 1474 * We queue the work to a specific CPU, the caller must ensure it
2375 * can't go away. Callers that fail to ensur !! 1475 * 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 * 1476 *
2380 * Return: %false if @work was already on a q 1477 * Return: %false if @work was already on a queue, %true otherwise.
2381 */ 1478 */
2382 bool queue_work_on(int cpu, struct workqueue_ 1479 bool queue_work_on(int cpu, struct workqueue_struct *wq,
2383 struct work_struct *work) 1480 struct work_struct *work)
2384 { 1481 {
2385 bool ret = false; 1482 bool ret = false;
2386 unsigned long irq_flags; !! 1483 unsigned long flags;
2387 1484
2388 local_irq_save(irq_flags); !! 1485 local_irq_save(flags);
2389 1486
2390 if (!test_and_set_bit(WORK_STRUCT_PEN !! 1487 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); 1488 __queue_work(cpu, wq, work);
2393 ret = true; 1489 ret = true;
2394 } 1490 }
2395 1491
2396 local_irq_restore(irq_flags); !! 1492 local_irq_restore(flags);
2397 return ret; 1493 return ret;
2398 } 1494 }
2399 EXPORT_SYMBOL(queue_work_on); 1495 EXPORT_SYMBOL(queue_work_on);
2400 1496
2401 /** <<
2402 * select_numa_node_cpu - Select a CPU based <<
2403 * @node: NUMA node ID that we want to select <<
2404 * <<
2405 * This function will attempt to find a "rand <<
2406 * node. If there are no CPUs available on th <<
2407 * WORK_CPU_UNBOUND indicating that we should <<
2408 * available CPU if we need to schedule this <<
2409 */ <<
2410 static int select_numa_node_cpu(int node) <<
2411 { <<
2412 int cpu; <<
2413 <<
2414 /* Delay binding to CPU if node is no <<
2415 if (node < 0 || node >= MAX_NUMNODES <<
2416 return WORK_CPU_UNBOUND; <<
2417 <<
2418 /* Use local node/cpu if we are alrea <<
2419 cpu = raw_smp_processor_id(); <<
2420 if (node == cpu_to_node(cpu)) <<
2421 return cpu; <<
2422 <<
2423 /* Use "random" otherwise know as "fi <<
2424 cpu = cpumask_any_and(cpumask_of_node <<
2425 <<
2426 /* If CPU is valid return that, other <<
2427 return cpu < nr_cpu_ids ? cpu : WORK_ <<
2428 } <<
2429 <<
2430 /** <<
2431 * queue_work_node - queue work on a "random" <<
2432 * @node: NUMA node that we are targeting the <<
2433 * @wq: workqueue to use <<
2434 * @work: work to queue <<
2435 * <<
2436 * We queue the work to a "random" CPU within <<
2437 * idea here is to provide a way to somehow a <<
2438 * NUMA node. <<
2439 * <<
2440 * This function will only make a best effort <<
2441 * the right NUMA node. If no node is request <<
2442 * offline then we just fall back to standard <<
2443 * <<
2444 * Currently the "random" CPU ends up being t <<
2445 * intersection of cpu_online_mask and the cp <<
2446 * are running on the node. In that case we j <<
2447 * <<
2448 * Return: %false if @work was already on a q <<
2449 */ <<
2450 bool queue_work_node(int node, struct workque <<
2451 struct work_struct *work <<
2452 { <<
2453 unsigned long irq_flags; <<
2454 bool ret = false; <<
2455 <<
2456 /* <<
2457 * This current implementation is spe <<
2458 * Specifically we only return the fi <<
2459 * node instead of cycling through in <<
2460 * <<
2461 * If this is used with a per-cpu wor <<
2462 * workqueue_select_cpu_near would ne <<
2463 * some round robin type logic. <<
2464 */ <<
2465 WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND <<
2466 <<
2467 local_irq_save(irq_flags); <<
2468 <<
2469 if (!test_and_set_bit(WORK_STRUCT_PEN <<
2470 !clear_pending_if_disabled(work)) <<
2471 int cpu = select_numa_node_cp <<
2472 <<
2473 __queue_work(cpu, wq, work); <<
2474 ret = true; <<
2475 } <<
2476 <<
2477 local_irq_restore(irq_flags); <<
2478 return ret; <<
2479 } <<
2480 EXPORT_SYMBOL_GPL(queue_work_node); <<
2481 <<
2482 void delayed_work_timer_fn(struct timer_list 1497 void delayed_work_timer_fn(struct timer_list *t)
2483 { 1498 {
2484 struct delayed_work *dwork = from_tim 1499 struct delayed_work *dwork = from_timer(dwork, t, timer);
2485 1500
2486 /* should have been called from irqsa 1501 /* should have been called from irqsafe timer with irq already off */
2487 __queue_work(dwork->cpu, dwork->wq, & 1502 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2488 } 1503 }
2489 EXPORT_SYMBOL(delayed_work_timer_fn); 1504 EXPORT_SYMBOL(delayed_work_timer_fn);
2490 1505
2491 static void __queue_delayed_work(int cpu, str 1506 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
2492 struct delaye 1507 struct delayed_work *dwork, unsigned long delay)
2493 { 1508 {
2494 struct timer_list *timer = &dwork->ti 1509 struct timer_list *timer = &dwork->timer;
2495 struct work_struct *work = &dwork->wo 1510 struct work_struct *work = &dwork->work;
2496 1511
2497 WARN_ON_ONCE(!wq); 1512 WARN_ON_ONCE(!wq);
2498 WARN_ON_ONCE(timer->function != delay 1513 WARN_ON_ONCE(timer->function != delayed_work_timer_fn);
2499 WARN_ON_ONCE(timer_pending(timer)); 1514 WARN_ON_ONCE(timer_pending(timer));
2500 WARN_ON_ONCE(!list_empty(&work->entry 1515 WARN_ON_ONCE(!list_empty(&work->entry));
2501 1516
2502 /* 1517 /*
2503 * If @delay is 0, queue @dwork->work 1518 * If @delay is 0, queue @dwork->work immediately. This is for
2504 * both optimization and correctness. 1519 * both optimization and correctness. The earliest @timer can
2505 * expire is on the closest next tick 1520 * expire is on the closest next tick and delayed_work users depend
2506 * on that there's no such delay when 1521 * on that there's no such delay when @delay is 0.
2507 */ 1522 */
2508 if (!delay) { 1523 if (!delay) {
2509 __queue_work(cpu, wq, &dwork- 1524 __queue_work(cpu, wq, &dwork->work);
2510 return; 1525 return;
2511 } 1526 }
2512 1527
2513 dwork->wq = wq; 1528 dwork->wq = wq;
2514 dwork->cpu = cpu; 1529 dwork->cpu = cpu;
2515 timer->expires = jiffies + delay; 1530 timer->expires = jiffies + delay;
2516 1531
2517 if (housekeeping_enabled(HK_TYPE_TIME !! 1532 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); 1533 add_timer_on(timer, cpu);
2523 } else { !! 1534 else
2524 if (likely(cpu == WORK_CPU_UN !! 1535 add_timer(timer);
2525 add_timer_global(time <<
2526 else <<
2527 add_timer_on(timer, c <<
2528 } <<
2529 } 1536 }
2530 1537
2531 /** 1538 /**
2532 * queue_delayed_work_on - queue work on spec 1539 * queue_delayed_work_on - queue work on specific CPU after delay
2533 * @cpu: CPU number to execute work on 1540 * @cpu: CPU number to execute work on
2534 * @wq: workqueue to use 1541 * @wq: workqueue to use
2535 * @dwork: work to queue 1542 * @dwork: work to queue
2536 * @delay: number of jiffies to wait before q 1543 * @delay: number of jiffies to wait before queueing
2537 * 1544 *
2538 * Return: %false if @work was already on a q 1545 * Return: %false if @work was already on a queue, %true otherwise. If
2539 * @delay is zero and @dwork is idle, it will 1546 * @delay is zero and @dwork is idle, it will be scheduled for immediate
2540 * execution. 1547 * execution.
2541 */ 1548 */
2542 bool queue_delayed_work_on(int cpu, struct wo 1549 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
2543 struct delayed_wor 1550 struct delayed_work *dwork, unsigned long delay)
2544 { 1551 {
2545 struct work_struct *work = &dwork->wo 1552 struct work_struct *work = &dwork->work;
2546 bool ret = false; 1553 bool ret = false;
2547 unsigned long irq_flags; !! 1554 unsigned long flags;
2548 1555
2549 /* read the comment in __queue_work() 1556 /* read the comment in __queue_work() */
2550 local_irq_save(irq_flags); !! 1557 local_irq_save(flags);
2551 1558
2552 if (!test_and_set_bit(WORK_STRUCT_PEN !! 1559 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, 1560 __queue_delayed_work(cpu, wq, dwork, delay);
2555 ret = true; 1561 ret = true;
2556 } 1562 }
2557 1563
2558 local_irq_restore(irq_flags); !! 1564 local_irq_restore(flags);
2559 return ret; 1565 return ret;
2560 } 1566 }
2561 EXPORT_SYMBOL(queue_delayed_work_on); 1567 EXPORT_SYMBOL(queue_delayed_work_on);
2562 1568
2563 /** 1569 /**
2564 * mod_delayed_work_on - modify delay of or q 1570 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
2565 * @cpu: CPU number to execute work on 1571 * @cpu: CPU number to execute work on
2566 * @wq: workqueue to use 1572 * @wq: workqueue to use
2567 * @dwork: work to queue 1573 * @dwork: work to queue
2568 * @delay: number of jiffies to wait before q 1574 * @delay: number of jiffies to wait before queueing
2569 * 1575 *
2570 * If @dwork is idle, equivalent to queue_del 1576 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
2571 * modify @dwork's timer so that it expires a 1577 * modify @dwork's timer so that it expires after @delay. If @delay is
2572 * zero, @work is guaranteed to be scheduled 1578 * zero, @work is guaranteed to be scheduled immediately regardless of its
2573 * current state. 1579 * current state.
2574 * 1580 *
2575 * Return: %false if @dwork was idle and queu 1581 * Return: %false if @dwork was idle and queued, %true if @dwork was
2576 * pending and its timer was modified. 1582 * pending and its timer was modified.
2577 * 1583 *
2578 * This function is safe to call from any con 1584 * This function is safe to call from any context including IRQ handler.
2579 * See try_to_grab_pending() for details. 1585 * See try_to_grab_pending() for details.
2580 */ 1586 */
2581 bool mod_delayed_work_on(int cpu, struct work 1587 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
2582 struct delayed_work 1588 struct delayed_work *dwork, unsigned long delay)
2583 { 1589 {
2584 unsigned long irq_flags; !! 1590 unsigned long flags;
2585 bool ret; !! 1591 int ret;
2586 1592
2587 ret = work_grab_pending(&dwork->work, !! 1593 do {
>> 1594 ret = try_to_grab_pending(&dwork->work, true, &flags);
>> 1595 } while (unlikely(ret == -EAGAIN));
2588 1596
2589 if (!clear_pending_if_disabled(&dwork !! 1597 if (likely(ret >= 0)) {
2590 __queue_delayed_work(cpu, wq, 1598 __queue_delayed_work(cpu, wq, dwork, delay);
>> 1599 local_irq_restore(flags);
>> 1600 }
2591 1601
2592 local_irq_restore(irq_flags); !! 1602 /* -ENOENT from try_to_grab_pending() becomes %true */
2593 return ret; 1603 return ret;
2594 } 1604 }
2595 EXPORT_SYMBOL_GPL(mod_delayed_work_on); 1605 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
2596 1606
2597 static void rcu_work_rcufn(struct rcu_head *r !! 1607 /**
>> 1608 * worker_enter_idle - enter idle state
>> 1609 * @worker: worker which is entering idle state
>> 1610 *
>> 1611 * @worker is entering idle state. Update stats and idle timer if
>> 1612 * necessary.
>> 1613 *
>> 1614 * LOCKING:
>> 1615 * spin_lock_irq(pool->lock).
>> 1616 */
>> 1617 static void worker_enter_idle(struct worker *worker)
2598 { 1618 {
2599 struct rcu_work *rwork = container_of !! 1619 struct worker_pool *pool = worker->pool;
2600 1620
2601 /* read the comment in __queue_work() !! 1621 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
2602 local_irq_disable(); !! 1622 WARN_ON_ONCE(!list_empty(&worker->entry) &&
2603 __queue_work(WORK_CPU_UNBOUND, rwork- !! 1623 (worker->hentry.next || worker->hentry.pprev)))
2604 local_irq_enable(); !! 1624 return;
>> 1625
>> 1626 /* can't use worker_set_flags(), also called from create_worker() */
>> 1627 worker->flags |= WORKER_IDLE;
>> 1628 pool->nr_idle++;
>> 1629 worker->last_active = jiffies;
>> 1630
>> 1631 /* idle_list is LIFO */
>> 1632 list_add(&worker->entry, &pool->idle_list);
>> 1633
>> 1634 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
>> 1635 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
>> 1636
>> 1637 /*
>> 1638 * Sanity check nr_running. Because unbind_workers() releases
>> 1639 * pool->lock between setting %WORKER_UNBOUND and zapping
>> 1640 * nr_running, the warning may trigger spuriously. Check iff
>> 1641 * unbind is not in progress.
>> 1642 */
>> 1643 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
>> 1644 pool->nr_workers == pool->nr_idle &&
>> 1645 atomic_read(&pool->nr_running));
2605 } 1646 }
2606 1647
2607 /** 1648 /**
2608 * queue_rcu_work - queue work after a RCU gr !! 1649 * worker_leave_idle - leave idle state
2609 * @wq: workqueue to use !! 1650 * @worker: worker which is leaving idle state
2610 * @rwork: work to queue !! 1651 *
>> 1652 * @worker is leaving idle state. Update stats.
2611 * 1653 *
2612 * Return: %false if @rwork was already pendi !! 1654 * LOCKING:
2613 * that a full RCU grace period is guaranteed !! 1655 * spin_lock_irq(pool->lock).
2614 * While @rwork is guaranteed to be executed <<
2615 * execution may happen before a full RCU gra <<
2616 */ 1656 */
2617 bool queue_rcu_work(struct workqueue_struct * !! 1657 static void worker_leave_idle(struct worker *worker)
2618 { 1658 {
2619 struct work_struct *work = &rwork->wo !! 1659 struct worker_pool *pool = worker->pool;
2620 <<
2621 /* <<
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; <<
2628 call_rcu_hurry(&rwork->rcu, r <<
2629 return true; <<
2630 } <<
2631 1660
2632 return false; !! 1661 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
>> 1662 return;
>> 1663 worker_clr_flags(worker, WORKER_IDLE);
>> 1664 pool->nr_idle--;
>> 1665 list_del_init(&worker->entry);
2633 } 1666 }
2634 EXPORT_SYMBOL(queue_rcu_work); <<
2635 1667
2636 static struct worker *alloc_worker(int node) 1668 static struct worker *alloc_worker(int node)
2637 { 1669 {
2638 struct worker *worker; 1670 struct worker *worker;
2639 1671
2640 worker = kzalloc_node(sizeof(*worker) 1672 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
2641 if (worker) { 1673 if (worker) {
2642 INIT_LIST_HEAD(&worker->entry 1674 INIT_LIST_HEAD(&worker->entry);
2643 INIT_LIST_HEAD(&worker->sched 1675 INIT_LIST_HEAD(&worker->scheduled);
2644 INIT_LIST_HEAD(&worker->node) 1676 INIT_LIST_HEAD(&worker->node);
2645 /* on creation a worker is in 1677 /* on creation a worker is in !idle && prep state */
2646 worker->flags = WORKER_PREP; 1678 worker->flags = WORKER_PREP;
2647 } 1679 }
2648 return worker; 1680 return worker;
2649 } 1681 }
2650 1682
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 /** 1683 /**
2660 * worker_attach_to_pool() - attach a worker 1684 * worker_attach_to_pool() - attach a worker to a pool
2661 * @worker: worker to be attached 1685 * @worker: worker to be attached
2662 * @pool: the target pool 1686 * @pool: the target pool
2663 * 1687 *
2664 * Attach @worker to @pool. Once attached, t 1688 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
2665 * cpu-binding of @worker are kept coordinate 1689 * cpu-binding of @worker are kept coordinated with the pool across
2666 * cpu-[un]hotplugs. 1690 * cpu-[un]hotplugs.
2667 */ 1691 */
2668 static void worker_attach_to_pool(struct work 1692 static void worker_attach_to_pool(struct worker *worker,
2669 struct work !! 1693 struct worker_pool *pool)
2670 { 1694 {
2671 mutex_lock(&wq_pool_attach_mutex); !! 1695 mutex_lock(&pool->attach_mutex);
2672 1696
2673 /* 1697 /*
2674 * The wq_pool_attach_mutex ensures % !! 1698 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
2675 * across this function. See the comm !! 1699 * online CPUs. It'll be re-applied when any of the CPUs come up.
2676 * details. BH workers are, while per <<
2677 */ 1700 */
2678 if (pool->flags & POOL_DISASSOCIATED) !! 1701 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
2679 worker->flags |= WORKER_UNBOU <<
2680 } else { <<
2681 WARN_ON_ONCE(pool->flags & PO <<
2682 kthread_set_per_cpu(worker->t <<
2683 } <<
2684 1702
2685 if (worker->rescue_wq) !! 1703 /*
2686 set_cpus_allowed_ptr(worker-> !! 1704 * The pool->attach_mutex ensures %POOL_DISASSOCIATED remains
>> 1705 * stable across this function. See the comments above the
>> 1706 * flag definition for details.
>> 1707 */
>> 1708 if (pool->flags & POOL_DISASSOCIATED)
>> 1709 worker->flags |= WORKER_UNBOUND;
2687 1710
2688 list_add_tail(&worker->node, &pool->w 1711 list_add_tail(&worker->node, &pool->workers);
2689 worker->pool = pool; <<
2690 <<
2691 mutex_unlock(&wq_pool_attach_mutex); <<
2692 } <<
2693 <<
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 1712
2706 static void detach_worker(struct worker *work !! 1713 mutex_unlock(&pool->attach_mutex);
2707 { <<
2708 lockdep_assert_held(&wq_pool_attach_m <<
2709 <<
2710 unbind_worker(worker); <<
2711 list_del(&worker->node); <<
2712 } 1714 }
2713 1715
2714 /** 1716 /**
2715 * worker_detach_from_pool() - detach a worke 1717 * worker_detach_from_pool() - detach a worker from its pool
2716 * @worker: worker which is attached to its p 1718 * @worker: worker which is attached to its pool
>> 1719 * @pool: the pool @worker is attached to
2717 * 1720 *
2718 * Undo the attaching which had been done in 1721 * Undo the attaching which had been done in worker_attach_to_pool(). The
2719 * caller worker shouldn't access to the pool 1722 * caller worker shouldn't access to the pool after detached except it has
2720 * other reference to the pool. 1723 * other reference to the pool.
2721 */ 1724 */
2722 static void worker_detach_from_pool(struct wo !! 1725 static void worker_detach_from_pool(struct worker *worker,
>> 1726 struct worker_pool *pool)
2723 { 1727 {
2724 struct worker_pool *pool = worker->po !! 1728 struct completion *detach_completion = NULL;
2725 1729
2726 /* there is one permanent BH worker p !! 1730 mutex_lock(&pool->attach_mutex);
2727 WARN_ON_ONCE(pool->flags & POOL_BH); !! 1731 list_del(&worker->node);
2728 !! 1732 if (list_empty(&pool->workers))
2729 mutex_lock(&wq_pool_attach_mutex); !! 1733 detach_completion = pool->detach_completion;
2730 detach_worker(worker); !! 1734 mutex_unlock(&pool->attach_mutex);
2731 worker->pool = NULL; <<
2732 mutex_unlock(&wq_pool_attach_mutex); <<
2733 1735
2734 /* clear leftover flags without pool- 1736 /* clear leftover flags without pool->lock after it is detached */
2735 worker->flags &= ~(WORKER_UNBOUND | W 1737 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
2736 } <<
2737 1738
2738 static int format_worker_id(char *buf, size_t !! 1739 if (detach_completion)
2739 struct worker_poo !! 1740 complete(detach_completion);
2740 { <<
2741 if (worker->rescue_wq) <<
2742 return scnprintf(buf, size, " <<
2743 worker->resc <<
2744 <<
2745 if (pool) { <<
2746 if (pool->cpu >= 0) <<
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 } 1741 }
2757 1742
2758 /** 1743 /**
2759 * create_worker - create a new workqueue wor 1744 * create_worker - create a new workqueue worker
2760 * @pool: pool the new worker will belong to 1745 * @pool: pool the new worker will belong to
2761 * 1746 *
2762 * Create and start a new worker which is att 1747 * Create and start a new worker which is attached to @pool.
2763 * 1748 *
2764 * CONTEXT: 1749 * CONTEXT:
2765 * Might sleep. Does GFP_KERNEL allocations. 1750 * Might sleep. Does GFP_KERNEL allocations.
2766 * 1751 *
2767 * Return: 1752 * Return:
2768 * Pointer to the newly created worker. 1753 * Pointer to the newly created worker.
2769 */ 1754 */
2770 static struct worker *create_worker(struct wo 1755 static struct worker *create_worker(struct worker_pool *pool)
2771 { 1756 {
2772 struct worker *worker; !! 1757 struct worker *worker = NULL;
2773 int id; !! 1758 int id = -1;
>> 1759 char id_buf[16];
2774 1760
2775 /* ID is needed to determine kthread 1761 /* ID is needed to determine kthread name */
2776 id = ida_alloc(&pool->worker_ida, GFP !! 1762 id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL);
2777 if (id < 0) { !! 1763 if (id < 0)
2778 pr_err_once("workqueue: Faile !! 1764 goto fail;
2779 ERR_PTR(id)); <<
2780 return NULL; <<
2781 } <<
2782 1765
2783 worker = alloc_worker(pool->node); 1766 worker = alloc_worker(pool->node);
2784 if (!worker) { !! 1767 if (!worker)
2785 pr_err_once("workqueue: Faile <<
2786 goto fail; 1768 goto fail;
2787 } <<
2788 1769
>> 1770 worker->pool = pool;
2789 worker->id = id; 1771 worker->id = id;
2790 1772
2791 if (!(pool->flags & POOL_BH)) { !! 1773 if (pool->cpu >= 0)
2792 char id_buf[WORKER_ID_LEN]; !! 1774 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
>> 1775 pool->attrs->nice < 0 ? "H" : "");
>> 1776 else
>> 1777 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
2793 1778
2794 format_worker_id(id_buf, size !! 1779 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
2795 worker->task = kthread_create !! 1780 "kworker/%s", id_buf);
2796 !! 1781 if (IS_ERR(worker->task))
2797 if (IS_ERR(worker->task)) { !! 1782 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 1783
2808 set_user_nice(worker->task, p !! 1784 set_user_nice(worker->task, pool->attrs->nice);
2809 kthread_bind_mask(worker->tas !! 1785 kthread_bind_mask(worker->task, pool->attrs->cpumask);
2810 } <<
2811 1786
2812 /* successful, attach the worker to t 1787 /* successful, attach the worker to the pool */
2813 worker_attach_to_pool(worker, pool); 1788 worker_attach_to_pool(worker, pool);
2814 1789
2815 /* start the newly created worker */ 1790 /* start the newly created worker */
2816 raw_spin_lock_irq(&pool->lock); !! 1791 spin_lock_irq(&pool->lock);
2817 <<
2818 worker->pool->nr_workers++; 1792 worker->pool->nr_workers++;
2819 worker_enter_idle(worker); 1793 worker_enter_idle(worker);
2820 !! 1794 wake_up_process(worker->task);
2821 /* !! 1795 spin_unlock_irq(&pool->lock);
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); <<
2830 1796
2831 return worker; 1797 return worker;
2832 1798
2833 fail: 1799 fail:
2834 ida_free(&pool->worker_ida, id); !! 1800 if (id >= 0)
>> 1801 ida_simple_remove(&pool->worker_ida, id);
2835 kfree(worker); 1802 kfree(worker);
2836 return NULL; 1803 return NULL;
2837 } 1804 }
2838 1805
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 /** 1806 /**
2859 * set_worker_dying - Tag a worker for destru !! 1807 * destroy_worker - destroy a workqueue worker
2860 * @worker: worker to be destroyed 1808 * @worker: worker to be destroyed
2861 * @list: transfer worker away from its pool- <<
2862 * 1809 *
2863 * Tag @worker for destruction and adjust @po !! 1810 * Destroy @worker and adjust @pool stats accordingly. The worker should
2864 * should be idle. !! 1811 * be idle.
2865 * 1812 *
2866 * CONTEXT: 1813 * CONTEXT:
2867 * raw_spin_lock_irq(pool->lock). !! 1814 * spin_lock_irq(pool->lock).
2868 */ 1815 */
2869 static void set_worker_dying(struct worker *w !! 1816 static void destroy_worker(struct worker *worker)
2870 { 1817 {
2871 struct worker_pool *pool = worker->po 1818 struct worker_pool *pool = worker->pool;
2872 1819
2873 lockdep_assert_held(&pool->lock); 1820 lockdep_assert_held(&pool->lock);
2874 lockdep_assert_held(&wq_pool_attach_m <<
2875 1821
2876 /* sanity check frenzy */ 1822 /* sanity check frenzy */
2877 if (WARN_ON(worker->current_work) || 1823 if (WARN_ON(worker->current_work) ||
2878 WARN_ON(!list_empty(&worker->sche 1824 WARN_ON(!list_empty(&worker->scheduled)) ||
2879 WARN_ON(!(worker->flags & WORKER_ 1825 WARN_ON(!(worker->flags & WORKER_IDLE)))
2880 return; 1826 return;
2881 1827
2882 pool->nr_workers--; 1828 pool->nr_workers--;
2883 pool->nr_idle--; 1829 pool->nr_idle--;
2884 1830
>> 1831 list_del_init(&worker->entry);
2885 worker->flags |= WORKER_DIE; 1832 worker->flags |= WORKER_DIE;
2886 !! 1833 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 } 1834 }
2892 1835
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_ 1836 static void idle_worker_timeout(struct timer_list *t)
2904 { 1837 {
2905 struct worker_pool *pool = from_timer 1838 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 1839
2927 if (do_cull) !! 1840 spin_lock_irq(&pool->lock);
2928 queue_work(system_unbound_wq, <<
2929 } <<
2930 <<
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); <<
2955 1841
2956 while (too_many_workers(pool)) { 1842 while (too_many_workers(pool)) {
2957 struct worker *worker; 1843 struct worker *worker;
2958 unsigned long expires; 1844 unsigned long expires;
2959 1845
2960 worker = list_last_entry(&poo !! 1846 /* idle_list is kept in LIFO order, check the last one */
>> 1847 worker = list_entry(pool->idle_list.prev, struct worker, entry);
2961 expires = worker->last_active 1848 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2962 1849
2963 if (time_before(jiffies, expi 1850 if (time_before(jiffies, expires)) {
2964 mod_timer(&pool->idle 1851 mod_timer(&pool->idle_timer, expires);
2965 break; 1852 break;
2966 } 1853 }
2967 1854
2968 set_worker_dying(worker, &cul !! 1855 destroy_worker(worker);
2969 } 1856 }
2970 1857
2971 raw_spin_unlock_irq(&pool->lock); !! 1858 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 } 1859 }
2977 1860
2978 static void send_mayday(struct work_struct *w 1861 static void send_mayday(struct work_struct *work)
2979 { 1862 {
2980 struct pool_workqueue *pwq = get_work 1863 struct pool_workqueue *pwq = get_work_pwq(work);
2981 struct workqueue_struct *wq = pwq->wq 1864 struct workqueue_struct *wq = pwq->wq;
2982 1865
2983 lockdep_assert_held(&wq_mayday_lock); 1866 lockdep_assert_held(&wq_mayday_lock);
2984 1867
2985 if (!wq->rescuer) 1868 if (!wq->rescuer)
2986 return; 1869 return;
2987 1870
2988 /* mayday mayday mayday */ 1871 /* mayday mayday mayday */
2989 if (list_empty(&pwq->mayday_node)) { 1872 if (list_empty(&pwq->mayday_node)) {
2990 /* 1873 /*
2991 * If @pwq is for an unbound 1874 * If @pwq is for an unbound wq, its base ref may be put at
2992 * any time due to an attribu 1875 * any time due to an attribute change. Pin @pwq until the
2993 * rescuer is done with it. 1876 * rescuer is done with it.
2994 */ 1877 */
2995 get_pwq(pwq); 1878 get_pwq(pwq);
2996 list_add_tail(&pwq->mayday_no 1879 list_add_tail(&pwq->mayday_node, &wq->maydays);
2997 wake_up_process(wq->rescuer-> 1880 wake_up_process(wq->rescuer->task);
2998 pwq->stats[PWQ_STAT_MAYDAY]++ <<
2999 } 1881 }
3000 } 1882 }
3001 1883
3002 static void pool_mayday_timeout(struct timer_ 1884 static void pool_mayday_timeout(struct timer_list *t)
3003 { 1885 {
3004 struct worker_pool *pool = from_timer 1886 struct worker_pool *pool = from_timer(pool, t, mayday_timer);
3005 struct work_struct *work; 1887 struct work_struct *work;
3006 1888
3007 raw_spin_lock_irq(&pool->lock); !! 1889 spin_lock_irq(&pool->lock);
3008 raw_spin_lock(&wq_mayday_lock); !! 1890 spin_lock(&wq_mayday_lock); /* for wq->maydays */
3009 1891
3010 if (need_to_create_worker(pool)) { 1892 if (need_to_create_worker(pool)) {
3011 /* 1893 /*
3012 * We've been trying to creat 1894 * We've been trying to create a new worker but
3013 * haven't been successful. 1895 * haven't been successful. We might be hitting an
3014 * allocation deadlock. Send 1896 * allocation deadlock. Send distress signals to
3015 * rescuers. 1897 * rescuers.
3016 */ 1898 */
3017 list_for_each_entry(work, &po 1899 list_for_each_entry(work, &pool->worklist, entry)
3018 send_mayday(work); 1900 send_mayday(work);
3019 } 1901 }
3020 1902
3021 raw_spin_unlock(&wq_mayday_lock); !! 1903 spin_unlock(&wq_mayday_lock);
3022 raw_spin_unlock_irq(&pool->lock); !! 1904 spin_unlock_irq(&pool->lock);
3023 1905
3024 mod_timer(&pool->mayday_timer, jiffie 1906 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
3025 } 1907 }
3026 1908
3027 /** 1909 /**
3028 * maybe_create_worker - create a new worker 1910 * maybe_create_worker - create a new worker if necessary
3029 * @pool: pool to create a new worker for 1911 * @pool: pool to create a new worker for
3030 * 1912 *
3031 * Create a new worker for @pool if necessary 1913 * Create a new worker for @pool if necessary. @pool is guaranteed to
3032 * have at least one idle worker on return fr 1914 * have at least one idle worker on return from this function. If
3033 * creating a new worker takes longer than MA 1915 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
3034 * sent to all rescuers with works scheduled 1916 * sent to all rescuers with works scheduled on @pool to resolve
3035 * possible allocation deadlock. 1917 * possible allocation deadlock.
3036 * 1918 *
3037 * On return, need_to_create_worker() is guar 1919 * On return, need_to_create_worker() is guaranteed to be %false and
3038 * may_start_working() %true. 1920 * may_start_working() %true.
3039 * 1921 *
3040 * LOCKING: 1922 * LOCKING:
3041 * raw_spin_lock_irq(pool->lock) which may be !! 1923 * spin_lock_irq(pool->lock) which may be released and regrabbed
3042 * multiple times. Does GFP_KERNEL allocatio 1924 * multiple times. Does GFP_KERNEL allocations. Called only from
3043 * manager. 1925 * manager.
3044 */ 1926 */
3045 static void maybe_create_worker(struct worker 1927 static void maybe_create_worker(struct worker_pool *pool)
3046 __releases(&pool->lock) 1928 __releases(&pool->lock)
3047 __acquires(&pool->lock) 1929 __acquires(&pool->lock)
3048 { 1930 {
3049 restart: 1931 restart:
3050 raw_spin_unlock_irq(&pool->lock); !! 1932 spin_unlock_irq(&pool->lock);
3051 1933
3052 /* if we don't make progress in MAYDA 1934 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
3053 mod_timer(&pool->mayday_timer, jiffie 1935 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
3054 1936
3055 while (true) { 1937 while (true) {
3056 if (create_worker(pool) || !n 1938 if (create_worker(pool) || !need_to_create_worker(pool))
3057 break; 1939 break;
3058 1940
3059 schedule_timeout_interruptibl 1941 schedule_timeout_interruptible(CREATE_COOLDOWN);
3060 1942
3061 if (!need_to_create_worker(po 1943 if (!need_to_create_worker(pool))
3062 break; 1944 break;
3063 } 1945 }
3064 1946
3065 del_timer_sync(&pool->mayday_timer); 1947 del_timer_sync(&pool->mayday_timer);
3066 raw_spin_lock_irq(&pool->lock); !! 1948 spin_lock_irq(&pool->lock);
3067 /* 1949 /*
3068 * This is necessary even after a new 1950 * This is necessary even after a new worker was just successfully
3069 * created as @pool->lock was dropped 1951 * created as @pool->lock was dropped and the new worker might have
3070 * already become busy. 1952 * already become busy.
3071 */ 1953 */
3072 if (need_to_create_worker(pool)) 1954 if (need_to_create_worker(pool))
3073 goto restart; 1955 goto restart;
3074 } 1956 }
3075 1957
3076 /** 1958 /**
3077 * manage_workers - manage worker pool 1959 * manage_workers - manage worker pool
3078 * @worker: self 1960 * @worker: self
3079 * 1961 *
3080 * Assume the manager role and manage the wor 1962 * Assume the manager role and manage the worker pool @worker belongs
3081 * to. At any given time, there can be only 1963 * to. At any given time, there can be only zero or one manager per
3082 * pool. The exclusion is handled automatica 1964 * pool. The exclusion is handled automatically by this function.
3083 * 1965 *
3084 * The caller can safely start processing wor 1966 * The caller can safely start processing works on false return. On
3085 * true return, it's guaranteed that need_to_ 1967 * true return, it's guaranteed that need_to_create_worker() is false
3086 * and may_start_working() is true. 1968 * and may_start_working() is true.
3087 * 1969 *
3088 * CONTEXT: 1970 * CONTEXT:
3089 * raw_spin_lock_irq(pool->lock) which may be !! 1971 * spin_lock_irq(pool->lock) which may be released and regrabbed
3090 * multiple times. Does GFP_KERNEL allocatio 1972 * multiple times. Does GFP_KERNEL allocations.
3091 * 1973 *
3092 * Return: 1974 * Return:
3093 * %false if the pool doesn't need management 1975 * %false if the pool doesn't need management and the caller can safely
3094 * start processing works, %true if managemen 1976 * start processing works, %true if management function was performed and
3095 * the conditions that the caller verified be 1977 * the conditions that the caller verified before calling the function may
3096 * no longer be true. 1978 * no longer be true.
3097 */ 1979 */
3098 static bool manage_workers(struct worker *wor 1980 static bool manage_workers(struct worker *worker)
3099 { 1981 {
3100 struct worker_pool *pool = worker->po 1982 struct worker_pool *pool = worker->pool;
3101 1983
3102 if (pool->flags & POOL_MANAGER_ACTIVE 1984 if (pool->flags & POOL_MANAGER_ACTIVE)
3103 return false; 1985 return false;
3104 1986
3105 pool->flags |= POOL_MANAGER_ACTIVE; 1987 pool->flags |= POOL_MANAGER_ACTIVE;
3106 pool->manager = worker; 1988 pool->manager = worker;
3107 1989
3108 maybe_create_worker(pool); 1990 maybe_create_worker(pool);
3109 1991
3110 pool->manager = NULL; 1992 pool->manager = NULL;
3111 pool->flags &= ~POOL_MANAGER_ACTIVE; 1993 pool->flags &= ~POOL_MANAGER_ACTIVE;
3112 rcuwait_wake_up(&manager_wait); !! 1994 wake_up(&wq_manager_wait);
3113 return true; 1995 return true;
3114 } 1996 }
3115 1997
3116 /** 1998 /**
3117 * process_one_work - process single work 1999 * process_one_work - process single work
3118 * @worker: self 2000 * @worker: self
3119 * @work: work to process 2001 * @work: work to process
3120 * 2002 *
3121 * Process @work. This function contains all 2003 * Process @work. This function contains all the logics necessary to
3122 * process a single work including synchroniz 2004 * process a single work including synchronization against and
3123 * interaction with other workers on the same 2005 * interaction with other workers on the same cpu, queueing and
3124 * flushing. As long as context requirement 2006 * flushing. As long as context requirement is met, any worker can
3125 * call this function to process a work. 2007 * call this function to process a work.
3126 * 2008 *
3127 * CONTEXT: 2009 * CONTEXT:
3128 * raw_spin_lock_irq(pool->lock) which is rel !! 2010 * spin_lock_irq(pool->lock) which is released and regrabbed.
3129 */ 2011 */
3130 static void process_one_work(struct worker *w 2012 static void process_one_work(struct worker *worker, struct work_struct *work)
3131 __releases(&pool->lock) 2013 __releases(&pool->lock)
3132 __acquires(&pool->lock) 2014 __acquires(&pool->lock)
3133 { 2015 {
3134 struct pool_workqueue *pwq = get_work 2016 struct pool_workqueue *pwq = get_work_pwq(work);
3135 struct worker_pool *pool = worker->po 2017 struct worker_pool *pool = worker->pool;
3136 unsigned long work_data; !! 2018 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
3137 int lockdep_start_depth, rcu_start_de !! 2019 int work_color;
3138 bool bh_draining = pool->flags & POOL !! 2020 struct worker *collision;
3139 #ifdef CONFIG_LOCKDEP 2021 #ifdef CONFIG_LOCKDEP
3140 /* 2022 /*
3141 * It is permissible to free the stru 2023 * It is permissible to free the struct work_struct from
3142 * inside the function that is called 2024 * inside the function that is called from it, this we need to
3143 * take into account for lockdep too. 2025 * take into account for lockdep too. To avoid bogus "held
3144 * lock freed" warnings as well as pr 2026 * lock freed" warnings as well as problems when looking into
3145 * work->lockdep_map, make a copy and 2027 * work->lockdep_map, make a copy and use that here.
3146 */ 2028 */
3147 struct lockdep_map lockdep_map; 2029 struct lockdep_map lockdep_map;
3148 2030
3149 lockdep_copy_map(&lockdep_map, &work- 2031 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
3150 #endif 2032 #endif
3151 /* ensure we're on the correct CPU */ 2033 /* ensure we're on the correct CPU */
3152 WARN_ON_ONCE(!(pool->flags & POOL_DIS 2034 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
3153 raw_smp_processor_id() ! 2035 raw_smp_processor_id() != pool->cpu);
3154 2036
>> 2037 /*
>> 2038 * A single work shouldn't be executed concurrently by
>> 2039 * multiple workers on a single cpu. Check whether anyone is
>> 2040 * already processing the work. If so, defer the work to the
>> 2041 * currently executing one.
>> 2042 */
>> 2043 collision = find_worker_executing_work(pool, work);
>> 2044 if (unlikely(collision)) {
>> 2045 move_linked_works(work, &collision->scheduled, NULL);
>> 2046 return;
>> 2047 }
>> 2048
3155 /* claim and dequeue */ 2049 /* claim and dequeue */
3156 debug_work_deactivate(work); 2050 debug_work_deactivate(work);
3157 hash_add(pool->busy_hash, &worker->he 2051 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
3158 worker->current_work = work; 2052 worker->current_work = work;
3159 worker->current_func = work->func; 2053 worker->current_func = work->func;
3160 worker->current_pwq = pwq; 2054 worker->current_pwq = pwq;
3161 if (worker->task) !! 2055 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 <<
3166 /* <<
3167 * Record wq name for cmdline and deb <<
3168 * overridden through set_worker_desc <<
3169 */ <<
3170 strscpy(worker->desc, pwq->wq->name, <<
3171 2056
3172 list_del_init(&work->entry); 2057 list_del_init(&work->entry);
3173 2058
3174 /* 2059 /*
3175 * CPU intensive works don't particip 2060 * CPU intensive works don't participate in concurrency management.
3176 * They're the scheduler's responsibi 2061 * They're the scheduler's responsibility. This takes @worker out
3177 * of concurrency management and the 2062 * of concurrency management and the next code block will chain
3178 * execution of the pending work item 2063 * execution of the pending work items.
3179 */ 2064 */
3180 if (unlikely(pwq->wq->flags & WQ_CPU_ !! 2065 if (unlikely(cpu_intensive))
3181 worker_set_flags(worker, WORK 2066 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
3182 2067
3183 /* 2068 /*
3184 * Kick @pool if necessary. It's alwa !! 2069 * Wake up another worker if necessary. The condition is always
3185 * since nr_running would always be > !! 2070 * false for normal per-cpu workers since nr_running would always
3186 * chain execution of the pending wor !! 2071 * be >= 1 at this point. This is used to chain execution of the
3187 * workers such as the UNBOUND and CP !! 2072 * pending work items for WORKER_NOT_RUNNING workers such as the
>> 2073 * UNBOUND and CPU_INTENSIVE ones.
3188 */ 2074 */
3189 kick_pool(pool); !! 2075 if (need_more_worker(pool))
>> 2076 wake_up_worker(pool);
3190 2077
3191 /* 2078 /*
3192 * Record the last pool and clear PEN 2079 * Record the last pool and clear PENDING which should be the last
3193 * update to @work. Also, do this in 2080 * update to @work. Also, do this inside @pool->lock so that
3194 * PENDING and queued state changes h 2081 * PENDING and queued state changes happen together while IRQ is
3195 * disabled. 2082 * disabled.
3196 */ 2083 */
3197 set_work_pool_and_clear_pending(work, !! 2084 set_work_pool_and_clear_pending(work, pool->id);
3198 2085
3199 pwq->stats[PWQ_STAT_STARTED]++; !! 2086 spin_unlock_irq(&pool->lock);
3200 raw_spin_unlock_irq(&pool->lock); <<
3201 2087
3202 rcu_start_depth = rcu_preempt_depth() !! 2088 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); 2089 lock_map_acquire(&lockdep_map);
3208 /* 2090 /*
3209 * Strictly speaking we should mark t 2091 * Strictly speaking we should mark the invariant state without holding
3210 * any locks, that is, before these t 2092 * any locks, that is, before these two lock_map_acquire()'s.
3211 * 2093 *
3212 * However, that would result in: 2094 * However, that would result in:
3213 * 2095 *
3214 * A(W1) 2096 * A(W1)
3215 * WFC(C) 2097 * WFC(C)
3216 * A(W1) 2098 * A(W1)
3217 * C(C) 2099 * C(C)
3218 * 2100 *
3219 * Which would create W1->C->W1 depen 2101 * Which would create W1->C->W1 dependencies, even though there is no
3220 * actual deadlock possible. There ar 2102 * actual deadlock possible. There are two solutions, using a
3221 * read-recursive acquire on the work 2103 * read-recursive acquire on the work(queue) 'locks', but this will then
3222 * hit the lockdep limitation on recu 2104 * hit the lockdep limitation on recursive locks, or simply discard
3223 * these locks. 2105 * these locks.
3224 * 2106 *
3225 * AFAICT there is no possible deadlo 2107 * AFAICT there is no possible deadlock scenario between the
3226 * flush_work() and complete() primit 2108 * flush_work() and complete() primitives (except for single-threaded
3227 * workqueues), so hiding them isn't 2109 * workqueues), so hiding them isn't a problem.
3228 */ 2110 */
3229 lockdep_invariant_state(true); 2111 lockdep_invariant_state(true);
3230 trace_workqueue_execute_start(work); 2112 trace_workqueue_execute_start(work);
3231 worker->current_func(work); 2113 worker->current_func(work);
3232 /* 2114 /*
3233 * While we must be careful to not us 2115 * While we must be careful to not use "work" after this, the trace
3234 * point will only record its address 2116 * point will only record its address.
3235 */ 2117 */
3236 trace_workqueue_execute_end(work, wor !! 2118 trace_workqueue_execute_end(work);
3237 pwq->stats[PWQ_STAT_COMPLETED]++; <<
3238 lock_map_release(&lockdep_map); 2119 lock_map_release(&lockdep_map);
3239 if (!bh_draining) !! 2120 lock_map_release(&pwq->wq->lockdep_map);
3240 lock_map_release(&pwq->wq->lo <<
3241 2121
3242 if (unlikely((worker->task && in_atom !! 2122 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
3243 lockdep_depth(current) ! !! 2123 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
3244 rcu_preempt_depth() != r !! 2124 " last function: %pf\n",
3245 pr_err("BUG: workqueue leaked !! 2125 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); 2126 worker->current_func);
3251 debug_show_held_locks(current 2127 debug_show_held_locks(current);
3252 dump_stack(); 2128 dump_stack();
3253 } 2129 }
3254 2130
3255 /* 2131 /*
3256 * The following prevents a kworker f !! 2132 * The following prevents a kworker from hogging CPU on !PREEMPT
3257 * kernels, where a requeueing work i 2133 * kernels, where a requeueing work item waiting for something to
3258 * happen could deadlock with stop_ma 2134 * happen could deadlock with stop_machine as such work item could
3259 * indefinitely requeue itself while 2135 * indefinitely requeue itself while all other CPUs are trapped in
3260 * stop_machine. At the same time, re 2136 * stop_machine. At the same time, report a quiescent RCU state so
3261 * the same condition doesn't freeze 2137 * the same condition doesn't freeze RCU.
3262 */ 2138 */
3263 if (worker->task) !! 2139 cond_resched();
3264 cond_resched(); <<
3265 2140
3266 raw_spin_lock_irq(&pool->lock); !! 2141 spin_lock_irq(&pool->lock);
3267 2142
3268 /* !! 2143 /* clear cpu intensive status */
3269 * In addition to %WQ_CPU_INTENSIVE, !! 2144 if (unlikely(cpu_intensive))
3270 * CPU intensive by wq_worker_tick() !! 2145 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
3271 * wq_cpu_intensive_thresh_us. Clear <<
3272 */ <<
3273 worker_clr_flags(worker, WORKER_CPU_I <<
3274 <<
3275 /* tag the worker for identification <<
3276 worker->last_func = worker->current_f <<
3277 2146
3278 /* we're done with it, release */ 2147 /* we're done with it, release */
3279 hash_del(&worker->hentry); 2148 hash_del(&worker->hentry);
3280 worker->current_work = NULL; 2149 worker->current_work = NULL;
3281 worker->current_func = NULL; 2150 worker->current_func = NULL;
3282 worker->current_pwq = NULL; 2151 worker->current_pwq = NULL;
3283 worker->current_color = INT_MAX; !! 2152 worker->desc_valid = false;
3284 !! 2153 pwq_dec_nr_in_flight(pwq, work_color);
3285 /* must be the last step, see the fun <<
3286 pwq_dec_nr_in_flight(pwq, work_data); <<
3287 } 2154 }
3288 2155
3289 /** 2156 /**
3290 * process_scheduled_works - process schedule 2157 * process_scheduled_works - process scheduled works
3291 * @worker: self 2158 * @worker: self
3292 * 2159 *
3293 * Process all scheduled works. Please note 2160 * Process all scheduled works. Please note that the scheduled list
3294 * may change while processing a work, so thi 2161 * may change while processing a work, so this function repeatedly
3295 * fetches a work from the top and executes i 2162 * fetches a work from the top and executes it.
3296 * 2163 *
3297 * CONTEXT: 2164 * CONTEXT:
3298 * raw_spin_lock_irq(pool->lock) which may be !! 2165 * spin_lock_irq(pool->lock) which may be released and regrabbed
3299 * multiple times. 2166 * multiple times.
3300 */ 2167 */
3301 static void process_scheduled_works(struct wo 2168 static void process_scheduled_works(struct worker *worker)
3302 { 2169 {
3303 struct work_struct *work; !! 2170 while (!list_empty(&worker->scheduled)) {
3304 bool first = true; !! 2171 struct work_struct *work = list_first_entry(&worker->scheduled,
3305 !! 2172 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 2173 process_one_work(worker, work);
3313 } 2174 }
3314 } 2175 }
3315 2176
3316 static void set_pf_worker(bool val) <<
3317 { <<
3318 mutex_lock(&wq_pool_attach_mutex); <<
3319 if (val) <<
3320 current->flags |= PF_WQ_WORKE <<
3321 else <<
3322 current->flags &= ~PF_WQ_WORK <<
3323 mutex_unlock(&wq_pool_attach_mutex); <<
3324 } <<
3325 <<
3326 /** 2177 /**
3327 * worker_thread - the worker thread function 2178 * worker_thread - the worker thread function
3328 * @__worker: self 2179 * @__worker: self
3329 * 2180 *
3330 * The worker thread function. All workers b 2181 * The worker thread function. All workers belong to a worker_pool -
3331 * either a per-cpu one or dynamic unbound on 2182 * either a per-cpu one or dynamic unbound one. These workers process all
3332 * work items regardless of their specific ta 2183 * work items regardless of their specific target workqueue. The only
3333 * exception is work items which belong to wo 2184 * exception is work items which belong to workqueues with a rescuer which
3334 * will be explained in rescuer_thread(). 2185 * will be explained in rescuer_thread().
3335 * 2186 *
3336 * Return: 0 2187 * Return: 0
3337 */ 2188 */
3338 static int worker_thread(void *__worker) 2189 static int worker_thread(void *__worker)
3339 { 2190 {
3340 struct worker *worker = __worker; 2191 struct worker *worker = __worker;
3341 struct worker_pool *pool = worker->po 2192 struct worker_pool *pool = worker->pool;
3342 2193
3343 /* tell the scheduler that this is a 2194 /* tell the scheduler that this is a workqueue worker */
3344 set_pf_worker(true); !! 2195 worker->task->flags |= PF_WQ_WORKER;
3345 woke_up: 2196 woke_up:
3346 raw_spin_lock_irq(&pool->lock); !! 2197 spin_lock_irq(&pool->lock);
3347 2198
3348 /* am I supposed to die? */ 2199 /* am I supposed to die? */
3349 if (unlikely(worker->flags & WORKER_D 2200 if (unlikely(worker->flags & WORKER_DIE)) {
3350 raw_spin_unlock_irq(&pool->lo !! 2201 spin_unlock_irq(&pool->lock);
3351 set_pf_worker(false); !! 2202 WARN_ON_ONCE(!list_empty(&worker->entry));
3352 /* !! 2203 worker->task->flags &= ~PF_WQ_WORKER;
3353 * The worker is dead and PF_ !! 2204
3354 * shouldn't be accessed, res !! 2205 set_task_comm(worker->task, "kworker/dying");
3355 */ !! 2206 ida_simple_remove(&pool->worker_ida, worker->id);
3356 worker->pool = NULL; !! 2207 worker_detach_from_pool(worker, pool);
3357 ida_free(&pool->worker_ida, w !! 2208 kfree(worker);
3358 return 0; 2209 return 0;
3359 } 2210 }
3360 2211
3361 worker_leave_idle(worker); 2212 worker_leave_idle(worker);
3362 recheck: 2213 recheck:
3363 /* no more worker necessary? */ 2214 /* no more worker necessary? */
3364 if (!need_more_worker(pool)) 2215 if (!need_more_worker(pool))
3365 goto sleep; 2216 goto sleep;
3366 2217
3367 /* do we need to manage? */ 2218 /* do we need to manage? */
3368 if (unlikely(!may_start_working(pool) 2219 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
3369 goto recheck; 2220 goto recheck;
3370 2221
3371 /* 2222 /*
3372 * ->scheduled list can only be fille 2223 * ->scheduled list can only be filled while a worker is
3373 * preparing to process a work or act 2224 * preparing to process a work or actually processing it.
3374 * Make sure nobody diddled with it w 2225 * Make sure nobody diddled with it while I was sleeping.
3375 */ 2226 */
3376 WARN_ON_ONCE(!list_empty(&worker->sch 2227 WARN_ON_ONCE(!list_empty(&worker->scheduled));
3377 2228
3378 /* 2229 /*
3379 * Finish PREP stage. We're guarante 2230 * Finish PREP stage. We're guaranteed to have at least one idle
3380 * worker or that someone else has al 2231 * worker or that someone else has already assumed the manager
3381 * role. This is where @worker start 2232 * role. This is where @worker starts participating in concurrency
3382 * management if applicable and concu 2233 * management if applicable and concurrency management is restored
3383 * after being rebound. See rebind_w 2234 * after being rebound. See rebind_workers() for details.
3384 */ 2235 */
3385 worker_clr_flags(worker, WORKER_PREP 2236 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
3386 2237
3387 do { 2238 do {
3388 struct work_struct *work = 2239 struct work_struct *work =
3389 list_first_entry(&poo 2240 list_first_entry(&pool->worklist,
3390 stru 2241 struct work_struct, entry);
3391 2242
3392 if (assign_work(work, worker, !! 2243 pool->watchdog_ts = jiffies;
>> 2244
>> 2245 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
>> 2246 /* optimization path, not strictly necessary */
>> 2247 process_one_work(worker, work);
>> 2248 if (unlikely(!list_empty(&worker->scheduled)))
>> 2249 process_scheduled_works(worker);
>> 2250 } else {
>> 2251 move_linked_works(work, &worker->scheduled, NULL);
3393 process_scheduled_wor 2252 process_scheduled_works(worker);
>> 2253 }
3394 } while (keep_working(pool)); 2254 } while (keep_working(pool));
3395 2255
3396 worker_set_flags(worker, WORKER_PREP) 2256 worker_set_flags(worker, WORKER_PREP);
3397 sleep: 2257 sleep:
3398 /* 2258 /*
3399 * pool->lock is held and there's no 2259 * pool->lock is held and there's no work to process and no need to
3400 * manage, sleep. Workers are woken 2260 * manage, sleep. Workers are woken up only while holding
3401 * pool->lock or from local cpu, so s 2261 * pool->lock or from local cpu, so setting the current state
3402 * before releasing pool->lock is eno 2262 * before releasing pool->lock is enough to prevent losing any
3403 * event. 2263 * event.
3404 */ 2264 */
3405 worker_enter_idle(worker); 2265 worker_enter_idle(worker);
3406 __set_current_state(TASK_IDLE); 2266 __set_current_state(TASK_IDLE);
3407 raw_spin_unlock_irq(&pool->lock); !! 2267 spin_unlock_irq(&pool->lock);
3408 schedule(); 2268 schedule();
3409 goto woke_up; 2269 goto woke_up;
3410 } 2270 }
3411 2271
3412 /** 2272 /**
3413 * rescuer_thread - the rescuer thread functi 2273 * rescuer_thread - the rescuer thread function
3414 * @__rescuer: self 2274 * @__rescuer: self
3415 * 2275 *
3416 * Workqueue rescuer thread function. There' 2276 * Workqueue rescuer thread function. There's one rescuer for each
3417 * workqueue which has WQ_MEM_RECLAIM set. 2277 * workqueue which has WQ_MEM_RECLAIM set.
3418 * 2278 *
3419 * Regular work processing on a pool may bloc 2279 * Regular work processing on a pool may block trying to create a new
3420 * worker which uses GFP_KERNEL allocation wh 2280 * worker which uses GFP_KERNEL allocation which has slight chance of
3421 * developing into deadlock if some works cur 2281 * developing into deadlock if some works currently on the same queue
3422 * need to be processed to satisfy the GFP_KE 2282 * need to be processed to satisfy the GFP_KERNEL allocation. This is
3423 * the problem rescuer solves. 2283 * the problem rescuer solves.
3424 * 2284 *
3425 * When such condition is possible, the pool 2285 * When such condition is possible, the pool summons rescuers of all
3426 * workqueues which have works queued on the 2286 * workqueues which have works queued on the pool and let them process
3427 * those works so that forward progress can b 2287 * those works so that forward progress can be guaranteed.
3428 * 2288 *
3429 * This should happen rarely. 2289 * This should happen rarely.
3430 * 2290 *
3431 * Return: 0 2291 * Return: 0
3432 */ 2292 */
3433 static int rescuer_thread(void *__rescuer) 2293 static int rescuer_thread(void *__rescuer)
3434 { 2294 {
3435 struct worker *rescuer = __rescuer; 2295 struct worker *rescuer = __rescuer;
3436 struct workqueue_struct *wq = rescuer 2296 struct workqueue_struct *wq = rescuer->rescue_wq;
>> 2297 struct list_head *scheduled = &rescuer->scheduled;
3437 bool should_stop; 2298 bool should_stop;
3438 2299
3439 set_user_nice(current, RESCUER_NICE_L 2300 set_user_nice(current, RESCUER_NICE_LEVEL);
3440 2301
3441 /* 2302 /*
3442 * Mark rescuer as worker too. As WO 2303 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
3443 * doesn't participate in concurrency 2304 * doesn't participate in concurrency management.
3444 */ 2305 */
3445 set_pf_worker(true); !! 2306 rescuer->task->flags |= PF_WQ_WORKER;
3446 repeat: 2307 repeat:
3447 set_current_state(TASK_IDLE); 2308 set_current_state(TASK_IDLE);
3448 2309
3449 /* 2310 /*
3450 * By the time the rescuer is request 2311 * By the time the rescuer is requested to stop, the workqueue
3451 * shouldn't have any work pending, b 2312 * shouldn't have any work pending, but @wq->maydays may still have
3452 * pwq(s) queued. This can happen by 2313 * pwq(s) queued. This can happen by non-rescuer workers consuming
3453 * all the work items before the resc 2314 * all the work items before the rescuer got to them. Go through
3454 * @wq->maydays processing before act 2315 * @wq->maydays processing before acting on should_stop so that the
3455 * list is always empty on exit. 2316 * list is always empty on exit.
3456 */ 2317 */
3457 should_stop = kthread_should_stop(); 2318 should_stop = kthread_should_stop();
3458 2319
3459 /* see whether any pwq is asking for 2320 /* see whether any pwq is asking for help */
3460 raw_spin_lock_irq(&wq_mayday_lock); !! 2321 spin_lock_irq(&wq_mayday_lock);
3461 2322
3462 while (!list_empty(&wq->maydays)) { 2323 while (!list_empty(&wq->maydays)) {
3463 struct pool_workqueue *pwq = 2324 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
3464 struc 2325 struct pool_workqueue, mayday_node);
3465 struct worker_pool *pool = pw 2326 struct worker_pool *pool = pwq->pool;
3466 struct work_struct *work, *n; 2327 struct work_struct *work, *n;
>> 2328 bool first = true;
3467 2329
3468 __set_current_state(TASK_RUNN 2330 __set_current_state(TASK_RUNNING);
3469 list_del_init(&pwq->mayday_no 2331 list_del_init(&pwq->mayday_node);
3470 2332
3471 raw_spin_unlock_irq(&wq_mayda !! 2333 spin_unlock_irq(&wq_mayday_lock);
3472 2334
3473 worker_attach_to_pool(rescuer 2335 worker_attach_to_pool(rescuer, pool);
3474 2336
3475 raw_spin_lock_irq(&pool->lock !! 2337 spin_lock_irq(&pool->lock);
>> 2338 rescuer->pool = pool;
3476 2339
3477 /* 2340 /*
3478 * Slurp in all works issued 2341 * Slurp in all works issued via this workqueue and
3479 * process'em. 2342 * process'em.
3480 */ 2343 */
3481 WARN_ON_ONCE(!list_empty(&res !! 2344 WARN_ON_ONCE(!list_empty(scheduled));
3482 list_for_each_entry_safe(work 2345 list_for_each_entry_safe(work, n, &pool->worklist, entry) {
3483 if (get_work_pwq(work !! 2346 if (get_work_pwq(work) == pwq) {
3484 assign_work(work, !! 2347 if (first)
3485 pwq->stats[PW !! 2348 pool->watchdog_ts = jiffies;
>> 2349 move_linked_works(work, scheduled, &n);
>> 2350 }
>> 2351 first = false;
3486 } 2352 }
3487 2353
3488 if (!list_empty(&rescuer->sch !! 2354 if (!list_empty(scheduled)) {
3489 process_scheduled_wor 2355 process_scheduled_works(rescuer);
3490 2356
3491 /* 2357 /*
3492 * The above executio 2358 * The above execution of rescued work items could
3493 * have created more 2359 * have created more to rescue through
3494 * pwq_activate_first !! 2360 * pwq_activate_first_delayed() or chained
3495 * queueing. Let's p 2361 * queueing. Let's put @pwq back on mayday list so
3496 * that such back-to- 2362 * that such back-to-back work items, which may be
3497 * being used to reli 2363 * being used to relieve memory pressure, don't
3498 * incur MAYDAY_INTER 2364 * incur MAYDAY_INTERVAL delay inbetween.
3499 */ 2365 */
3500 if (pwq->nr_active && !! 2366 if (need_to_create_worker(pool)) {
3501 raw_spin_lock !! 2367 spin_lock(&wq_mayday_lock);
3502 /* !! 2368 get_pwq(pwq);
3503 * Queue iff !! 2369 list_move_tail(&pwq->mayday_node, &wq->maydays);
3504 * and somebo !! 2370 spin_unlock(&wq_mayday_lock);
3505 */ <<
3506 if (wq->rescu <<
3507 get_p <<
3508 list_ <<
3509 } <<
3510 raw_spin_unlo <<
3511 } 2371 }
3512 } 2372 }
3513 2373
3514 /* 2374 /*
3515 * Put the reference grabbed 2375 * Put the reference grabbed by send_mayday(). @pool won't
3516 * go away while we're still 2376 * go away while we're still attached to it.
3517 */ 2377 */
3518 put_pwq(pwq); 2378 put_pwq(pwq);
3519 2379
3520 /* 2380 /*
3521 * Leave this pool. Notify re !! 2381 * Leave this pool. If need_more_worker() is %true, notify a
3522 * with 0 concurrency and sta !! 2382 * regular worker; otherwise, we end up with 0 concurrency
>> 2383 * and stalling the execution.
3523 */ 2384 */
3524 kick_pool(pool); !! 2385 if (need_more_worker(pool))
>> 2386 wake_up_worker(pool);
3525 2387
3526 raw_spin_unlock_irq(&pool->lo !! 2388 rescuer->pool = NULL;
>> 2389 spin_unlock_irq(&pool->lock);
3527 2390
3528 worker_detach_from_pool(rescu !! 2391 worker_detach_from_pool(rescuer, pool);
3529 2392
3530 raw_spin_lock_irq(&wq_mayday_ !! 2393 spin_lock_irq(&wq_mayday_lock);
3531 } 2394 }
3532 2395
3533 raw_spin_unlock_irq(&wq_mayday_lock); !! 2396 spin_unlock_irq(&wq_mayday_lock);
3534 2397
3535 if (should_stop) { 2398 if (should_stop) {
3536 __set_current_state(TASK_RUNN 2399 __set_current_state(TASK_RUNNING);
3537 set_pf_worker(false); !! 2400 rescuer->task->flags &= ~PF_WQ_WORKER;
3538 return 0; 2401 return 0;
3539 } 2402 }
3540 2403
3541 /* rescuers should never participate 2404 /* rescuers should never participate in concurrency management */
3542 WARN_ON_ONCE(!(rescuer->flags & WORKE 2405 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
3543 schedule(); 2406 schedule();
3544 goto repeat; 2407 goto repeat;
3545 } 2408 }
3546 2409
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 /** 2410 /**
3682 * check_flush_dependency - check for flush d 2411 * check_flush_dependency - check for flush dependency sanity
3683 * @target_wq: workqueue being flushed 2412 * @target_wq: workqueue being flushed
3684 * @target_work: work item being flushed (NUL 2413 * @target_work: work item being flushed (NULL for workqueue flushes)
3685 * 2414 *
3686 * %current is trying to flush the whole @tar 2415 * %current is trying to flush the whole @target_wq or @target_work on it.
3687 * If @target_wq doesn't have %WQ_MEM_RECLAIM 2416 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
3688 * reclaiming memory or running on a workqueu 2417 * reclaiming memory or running on a workqueue which doesn't have
3689 * %WQ_MEM_RECLAIM as that can break forward- 2418 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
3690 * a deadlock. 2419 * a deadlock.
3691 */ 2420 */
3692 static void check_flush_dependency(struct wor 2421 static void check_flush_dependency(struct workqueue_struct *target_wq,
3693 struct wor 2422 struct work_struct *target_work)
3694 { 2423 {
3695 work_func_t target_func = target_work 2424 work_func_t target_func = target_work ? target_work->func : NULL;
3696 struct worker *worker; 2425 struct worker *worker;
3697 2426
3698 if (target_wq->flags & WQ_MEM_RECLAIM 2427 if (target_wq->flags & WQ_MEM_RECLAIM)
3699 return; 2428 return;
3700 2429
3701 worker = current_wq_worker(); 2430 worker = current_wq_worker();
3702 2431
3703 WARN_ONCE(current->flags & PF_MEMALLO 2432 WARN_ONCE(current->flags & PF_MEMALLOC,
3704 "workqueue: PF_MEMALLOC tas !! 2433 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%pf",
3705 current->pid, current->comm 2434 current->pid, current->comm, target_wq->name, target_func);
3706 WARN_ONCE(worker && ((worker->current 2435 WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
3707 (WQ_MEM_RECLAIM 2436 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
3708 "workqueue: WQ_MEM_RECLAIM !! 2437 "workqueue: WQ_MEM_RECLAIM %s:%pf is flushing !WQ_MEM_RECLAIM %s:%pf",
3709 worker->current_pwq->wq->na 2438 worker->current_pwq->wq->name, worker->current_func,
3710 target_wq->name, target_fun 2439 target_wq->name, target_func);
3711 } 2440 }
3712 2441
3713 struct wq_barrier { 2442 struct wq_barrier {
3714 struct work_struct work; 2443 struct work_struct work;
3715 struct completion done; 2444 struct completion done;
3716 struct task_struct *task; /* pu 2445 struct task_struct *task; /* purely informational */
3717 }; 2446 };
3718 2447
3719 static void wq_barrier_func(struct work_struc 2448 static void wq_barrier_func(struct work_struct *work)
3720 { 2449 {
3721 struct wq_barrier *barr = container_o 2450 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
3722 complete(&barr->done); 2451 complete(&barr->done);
3723 } 2452 }
3724 2453
3725 /** 2454 /**
3726 * insert_wq_barrier - insert a barrier work 2455 * insert_wq_barrier - insert a barrier work
3727 * @pwq: pwq to insert barrier into 2456 * @pwq: pwq to insert barrier into
3728 * @barr: wq_barrier to insert 2457 * @barr: wq_barrier to insert
3729 * @target: target work to attach @barr to 2458 * @target: target work to attach @barr to
3730 * @worker: worker currently executing @targe 2459 * @worker: worker currently executing @target, NULL if @target is not executing
3731 * 2460 *
3732 * @barr is linked to @target such that @barr 2461 * @barr is linked to @target such that @barr is completed only after
3733 * @target finishes execution. Please note t 2462 * @target finishes execution. Please note that the ordering
3734 * guarantee is observed only with respect to 2463 * guarantee is observed only with respect to @target and on the local
3735 * cpu. 2464 * cpu.
3736 * 2465 *
3737 * Currently, a queued barrier can't be cance 2466 * Currently, a queued barrier can't be canceled. This is because
3738 * try_to_grab_pending() can't determine whet 2467 * try_to_grab_pending() can't determine whether the work to be
3739 * grabbed is at the head of the queue and th 2468 * grabbed is at the head of the queue and thus can't clear LINKED
3740 * flag of the previous work while there must 2469 * flag of the previous work while there must be a valid next work
3741 * after a work with LINKED flag set. 2470 * after a work with LINKED flag set.
3742 * 2471 *
3743 * Note that when @worker is non-NULL, @targe 2472 * Note that when @worker is non-NULL, @target may be modified
3744 * underneath us, so we can't reliably determ 2473 * underneath us, so we can't reliably determine pwq from @target.
3745 * 2474 *
3746 * CONTEXT: 2475 * CONTEXT:
3747 * raw_spin_lock_irq(pool->lock). !! 2476 * spin_lock_irq(pool->lock).
3748 */ 2477 */
3749 static void insert_wq_barrier(struct pool_wor 2478 static void insert_wq_barrier(struct pool_workqueue *pwq,
3750 struct wq_barri 2479 struct wq_barrier *barr,
3751 struct work_str 2480 struct work_struct *target, struct worker *worker)
3752 { 2481 {
3753 static __maybe_unused struct lock_cla <<
3754 unsigned int work_flags = 0; <<
3755 unsigned int work_color; <<
3756 struct list_head *head; 2482 struct list_head *head;
>> 2483 unsigned int linked = 0;
3757 2484
3758 /* 2485 /*
3759 * debugobject calls are safe here ev 2486 * debugobject calls are safe here even with pool->lock locked
3760 * as we know for sure that this will 2487 * as we know for sure that this will not trigger any of the
3761 * checks and call back into the fixu 2488 * checks and call back into the fixup functions where we
3762 * might deadlock. 2489 * might deadlock.
3763 * <<
3764 * BH and threaded workqueues need se <<
3765 * spuriously triggering "inconsisten <<
3766 * usage". <<
3767 */ 2490 */
3768 INIT_WORK_ONSTACK_KEY(&barr->work, wq !! 2491 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
3769 (pwq->wq->flags <<
3770 __set_bit(WORK_STRUCT_PENDING_BIT, wo 2492 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
3771 2493
3772 init_completion_map(&barr->done, &tar 2494 init_completion_map(&barr->done, &target->lockdep_map);
3773 2495
3774 barr->task = current; 2496 barr->task = current;
3775 2497
3776 /* The barrier work item does not par <<
3777 work_flags |= WORK_STRUCT_INACTIVE; <<
3778 <<
3779 /* 2498 /*
3780 * If @target is currently being exec 2499 * If @target is currently being executed, schedule the
3781 * barrier to the worker; otherwise, 2500 * barrier to the worker; otherwise, put it after @target.
3782 */ 2501 */
3783 if (worker) { !! 2502 if (worker)
3784 head = worker->scheduled.next 2503 head = worker->scheduled.next;
3785 work_color = worker->current_ !! 2504 else {
3786 } else { <<
3787 unsigned long *bits = work_da 2505 unsigned long *bits = work_data_bits(target);
3788 2506
3789 head = target->entry.next; 2507 head = target->entry.next;
3790 /* there can already be other 2508 /* there can already be other linked works, inherit and set */
3791 work_flags |= *bits & WORK_ST !! 2509 linked = *bits & WORK_STRUCT_LINKED;
3792 work_color = get_work_color(* <<
3793 __set_bit(WORK_STRUCT_LINKED_ 2510 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
3794 } 2511 }
3795 2512
3796 pwq->nr_in_flight[work_color]++; !! 2513 debug_work_activate(&barr->work);
3797 work_flags |= work_color_to_flags(wor !! 2514 insert_work(pwq, &barr->work, head,
3798 !! 2515 work_color_to_flags(WORK_NO_COLOR) | linked);
3799 insert_work(pwq, &barr->work, head, w <<
3800 } 2516 }
3801 2517
3802 /** 2518 /**
3803 * flush_workqueue_prep_pwqs - prepare pwqs f 2519 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
3804 * @wq: workqueue being flushed 2520 * @wq: workqueue being flushed
3805 * @flush_color: new flush color, < 0 for no- 2521 * @flush_color: new flush color, < 0 for no-op
3806 * @work_color: new work color, < 0 for no-op 2522 * @work_color: new work color, < 0 for no-op
3807 * 2523 *
3808 * Prepare pwqs for workqueue flushing. 2524 * Prepare pwqs for workqueue flushing.
3809 * 2525 *
3810 * If @flush_color is non-negative, flush_col 2526 * If @flush_color is non-negative, flush_color on all pwqs should be
3811 * -1. If no pwq has in-flight commands at t 2527 * -1. If no pwq has in-flight commands at the specified color, all
3812 * pwq->flush_color's stay at -1 and %false i 2528 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
3813 * has in flight commands, its pwq->flush_col 2529 * has in flight commands, its pwq->flush_color is set to
3814 * @flush_color, @wq->nr_pwqs_to_flush is upd 2530 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
3815 * wakeup logic is armed and %true is returne 2531 * wakeup logic is armed and %true is returned.
3816 * 2532 *
3817 * The caller should have initialized @wq->fi 2533 * The caller should have initialized @wq->first_flusher prior to
3818 * calling this function with non-negative @f 2534 * calling this function with non-negative @flush_color. If
3819 * @flush_color is negative, no flush color u 2535 * @flush_color is negative, no flush color update is done and %false
3820 * is returned. 2536 * is returned.
3821 * 2537 *
3822 * If @work_color is non-negative, all pwqs s 2538 * If @work_color is non-negative, all pwqs should have the same
3823 * work_color which is previous to @work_colo 2539 * work_color which is previous to @work_color and all will be
3824 * advanced to @work_color. 2540 * advanced to @work_color.
3825 * 2541 *
3826 * CONTEXT: 2542 * CONTEXT:
3827 * mutex_lock(wq->mutex). 2543 * mutex_lock(wq->mutex).
3828 * 2544 *
3829 * Return: 2545 * Return:
3830 * %true if @flush_color >= 0 and there's som 2546 * %true if @flush_color >= 0 and there's something to flush. %false
3831 * otherwise. 2547 * otherwise.
3832 */ 2548 */
3833 static bool flush_workqueue_prep_pwqs(struct 2549 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
3834 int flu 2550 int flush_color, int work_color)
3835 { 2551 {
3836 bool wait = false; 2552 bool wait = false;
3837 struct pool_workqueue *pwq; 2553 struct pool_workqueue *pwq;
3838 2554
3839 if (flush_color >= 0) { 2555 if (flush_color >= 0) {
3840 WARN_ON_ONCE(atomic_read(&wq- 2556 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
3841 atomic_set(&wq->nr_pwqs_to_fl 2557 atomic_set(&wq->nr_pwqs_to_flush, 1);
3842 } 2558 }
3843 2559
3844 for_each_pwq(pwq, wq) { 2560 for_each_pwq(pwq, wq) {
3845 struct worker_pool *pool = pw 2561 struct worker_pool *pool = pwq->pool;
3846 2562
3847 raw_spin_lock_irq(&pool->lock !! 2563 spin_lock_irq(&pool->lock);
3848 2564
3849 if (flush_color >= 0) { 2565 if (flush_color >= 0) {
3850 WARN_ON_ONCE(pwq->flu 2566 WARN_ON_ONCE(pwq->flush_color != -1);
3851 2567
3852 if (pwq->nr_in_flight 2568 if (pwq->nr_in_flight[flush_color]) {
3853 pwq->flush_co 2569 pwq->flush_color = flush_color;
3854 atomic_inc(&w 2570 atomic_inc(&wq->nr_pwqs_to_flush);
3855 wait = true; 2571 wait = true;
3856 } 2572 }
3857 } 2573 }
3858 2574
3859 if (work_color >= 0) { 2575 if (work_color >= 0) {
3860 WARN_ON_ONCE(work_col 2576 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
3861 pwq->work_color = wor 2577 pwq->work_color = work_color;
3862 } 2578 }
3863 2579
3864 raw_spin_unlock_irq(&pool->lo !! 2580 spin_unlock_irq(&pool->lock);
3865 } 2581 }
3866 2582
3867 if (flush_color >= 0 && atomic_dec_an 2583 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
3868 complete(&wq->first_flusher-> 2584 complete(&wq->first_flusher->done);
3869 2585
3870 return wait; 2586 return wait;
3871 } 2587 }
3872 2588
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 /** 2589 /**
3903 * __flush_workqueue - ensure that any schedu !! 2590 * flush_workqueue - ensure that any scheduled work has run to completion.
3904 * @wq: workqueue to flush 2591 * @wq: workqueue to flush
3905 * 2592 *
3906 * This function sleeps until all work items 2593 * This function sleeps until all work items which were queued on entry
3907 * have finished execution, but it is not liv 2594 * have finished execution, but it is not livelocked by new incoming ones.
3908 */ 2595 */
3909 void __flush_workqueue(struct workqueue_struc !! 2596 void flush_workqueue(struct workqueue_struct *wq)
3910 { 2597 {
3911 struct wq_flusher this_flusher = { 2598 struct wq_flusher this_flusher = {
3912 .list = LIST_HEAD_INIT(this_f 2599 .list = LIST_HEAD_INIT(this_flusher.list),
3913 .flush_color = -1, 2600 .flush_color = -1,
3914 .done = COMPLETION_INITIALIZE 2601 .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map),
3915 }; 2602 };
3916 int next_color; 2603 int next_color;
3917 2604
3918 if (WARN_ON(!wq_online)) 2605 if (WARN_ON(!wq_online))
3919 return; 2606 return;
3920 2607
3921 touch_wq_lockdep_map(wq); <<
3922 <<
3923 mutex_lock(&wq->mutex); 2608 mutex_lock(&wq->mutex);
3924 2609
3925 /* 2610 /*
3926 * Start-to-wait phase 2611 * Start-to-wait phase
3927 */ 2612 */
3928 next_color = work_next_color(wq->work 2613 next_color = work_next_color(wq->work_color);
3929 2614
3930 if (next_color != wq->flush_color) { 2615 if (next_color != wq->flush_color) {
3931 /* 2616 /*
3932 * Color space is not full. 2617 * Color space is not full. The current work_color
3933 * becomes our flush_color an 2618 * becomes our flush_color and work_color is advanced
3934 * by one. 2619 * by one.
3935 */ 2620 */
3936 WARN_ON_ONCE(!list_empty(&wq- 2621 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
3937 this_flusher.flush_color = wq 2622 this_flusher.flush_color = wq->work_color;
3938 wq->work_color = next_color; 2623 wq->work_color = next_color;
3939 2624
3940 if (!wq->first_flusher) { 2625 if (!wq->first_flusher) {
3941 /* no flush in progre 2626 /* no flush in progress, become the first flusher */
3942 WARN_ON_ONCE(wq->flus 2627 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
3943 2628
3944 wq->first_flusher = & 2629 wq->first_flusher = &this_flusher;
3945 2630
3946 if (!flush_workqueue_ 2631 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
3947 2632 wq->work_color)) {
3948 /* nothing to 2633 /* nothing to flush, done */
3949 wq->flush_col 2634 wq->flush_color = next_color;
3950 wq->first_flu 2635 wq->first_flusher = NULL;
3951 goto out_unlo 2636 goto out_unlock;
3952 } 2637 }
3953 } else { 2638 } else {
3954 /* wait in queue */ 2639 /* wait in queue */
3955 WARN_ON_ONCE(wq->flus 2640 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
3956 list_add_tail(&this_f 2641 list_add_tail(&this_flusher.list, &wq->flusher_queue);
3957 flush_workqueue_prep_ 2642 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
3958 } 2643 }
3959 } else { 2644 } else {
3960 /* 2645 /*
3961 * Oops, color space is full, 2646 * Oops, color space is full, wait on overflow queue.
3962 * The next flush completion 2647 * The next flush completion will assign us
3963 * flush_color and transfer t 2648 * flush_color and transfer to flusher_queue.
3964 */ 2649 */
3965 list_add_tail(&this_flusher.l 2650 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
3966 } 2651 }
3967 2652
3968 check_flush_dependency(wq, NULL); 2653 check_flush_dependency(wq, NULL);
3969 2654
3970 mutex_unlock(&wq->mutex); 2655 mutex_unlock(&wq->mutex);
3971 2656
3972 wait_for_completion(&this_flusher.don 2657 wait_for_completion(&this_flusher.done);
3973 2658
3974 /* 2659 /*
3975 * Wake-up-and-cascade phase 2660 * Wake-up-and-cascade phase
3976 * 2661 *
3977 * First flushers are responsible for 2662 * First flushers are responsible for cascading flushes and
3978 * handling overflow. Non-first flus 2663 * handling overflow. Non-first flushers can simply return.
3979 */ 2664 */
3980 if (READ_ONCE(wq->first_flusher) != & !! 2665 if (wq->first_flusher != &this_flusher)
3981 return; 2666 return;
3982 2667
3983 mutex_lock(&wq->mutex); 2668 mutex_lock(&wq->mutex);
3984 2669
3985 /* we might have raced, check again w 2670 /* we might have raced, check again with mutex held */
3986 if (wq->first_flusher != &this_flushe 2671 if (wq->first_flusher != &this_flusher)
3987 goto out_unlock; 2672 goto out_unlock;
3988 2673
3989 WRITE_ONCE(wq->first_flusher, NULL); !! 2674 wq->first_flusher = NULL;
3990 2675
3991 WARN_ON_ONCE(!list_empty(&this_flushe 2676 WARN_ON_ONCE(!list_empty(&this_flusher.list));
3992 WARN_ON_ONCE(wq->flush_color != this_ 2677 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
3993 2678
3994 while (true) { 2679 while (true) {
3995 struct wq_flusher *next, *tmp 2680 struct wq_flusher *next, *tmp;
3996 2681
3997 /* complete all the flushers 2682 /* complete all the flushers sharing the current flush color */
3998 list_for_each_entry_safe(next 2683 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
3999 if (next->flush_color 2684 if (next->flush_color != wq->flush_color)
4000 break; 2685 break;
4001 list_del_init(&next-> 2686 list_del_init(&next->list);
4002 complete(&next->done) 2687 complete(&next->done);
4003 } 2688 }
4004 2689
4005 WARN_ON_ONCE(!list_empty(&wq- 2690 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
4006 wq->flush_color 2691 wq->flush_color != work_next_color(wq->work_color));
4007 2692
4008 /* this flush_color is finish 2693 /* this flush_color is finished, advance by one */
4009 wq->flush_color = work_next_c 2694 wq->flush_color = work_next_color(wq->flush_color);
4010 2695
4011 /* one color has been freed, 2696 /* one color has been freed, handle overflow queue */
4012 if (!list_empty(&wq->flusher_ 2697 if (!list_empty(&wq->flusher_overflow)) {
4013 /* 2698 /*
4014 * Assign the same co 2699 * Assign the same color to all overflowed
4015 * flushers, advance 2700 * flushers, advance work_color and append to
4016 * flusher_queue. Th 2701 * flusher_queue. This is the start-to-wait
4017 * phase for these ov 2702 * phase for these overflowed flushers.
4018 */ 2703 */
4019 list_for_each_entry(t 2704 list_for_each_entry(tmp, &wq->flusher_overflow, list)
4020 tmp->flush_co 2705 tmp->flush_color = wq->work_color;
4021 2706
4022 wq->work_color = work 2707 wq->work_color = work_next_color(wq->work_color);
4023 2708
4024 list_splice_tail_init 2709 list_splice_tail_init(&wq->flusher_overflow,
4025 2710 &wq->flusher_queue);
4026 flush_workqueue_prep_ 2711 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
4027 } 2712 }
4028 2713
4029 if (list_empty(&wq->flusher_q 2714 if (list_empty(&wq->flusher_queue)) {
4030 WARN_ON_ONCE(wq->flus 2715 WARN_ON_ONCE(wq->flush_color != wq->work_color);
4031 break; 2716 break;
4032 } 2717 }
4033 2718
4034 /* 2719 /*
4035 * Need to flush more colors. 2720 * Need to flush more colors. Make the next flusher
4036 * the new first flusher and 2721 * the new first flusher and arm pwqs.
4037 */ 2722 */
4038 WARN_ON_ONCE(wq->flush_color 2723 WARN_ON_ONCE(wq->flush_color == wq->work_color);
4039 WARN_ON_ONCE(wq->flush_color 2724 WARN_ON_ONCE(wq->flush_color != next->flush_color);
4040 2725
4041 list_del_init(&next->list); 2726 list_del_init(&next->list);
4042 wq->first_flusher = next; 2727 wq->first_flusher = next;
4043 2728
4044 if (flush_workqueue_prep_pwqs 2729 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
4045 break; 2730 break;
4046 2731
4047 /* 2732 /*
4048 * Meh... this color is alrea 2733 * Meh... this color is already done, clear first
4049 * flusher and repeat cascadi 2734 * flusher and repeat cascading.
4050 */ 2735 */
4051 wq->first_flusher = NULL; 2736 wq->first_flusher = NULL;
4052 } 2737 }
4053 2738
4054 out_unlock: 2739 out_unlock:
4055 mutex_unlock(&wq->mutex); 2740 mutex_unlock(&wq->mutex);
4056 } 2741 }
4057 EXPORT_SYMBOL(__flush_workqueue); !! 2742 EXPORT_SYMBOL(flush_workqueue);
4058 2743
4059 /** 2744 /**
4060 * drain_workqueue - drain a workqueue 2745 * drain_workqueue - drain a workqueue
4061 * @wq: workqueue to drain 2746 * @wq: workqueue to drain
4062 * 2747 *
4063 * Wait until the workqueue becomes empty. W 2748 * Wait until the workqueue becomes empty. While draining is in progress,
4064 * only chain queueing is allowed. IOW, only 2749 * only chain queueing is allowed. IOW, only currently pending or running
4065 * work items on @wq can queue further work i 2750 * work items on @wq can queue further work items on it. @wq is flushed
4066 * repeatedly until it becomes empty. The nu 2751 * repeatedly until it becomes empty. The number of flushing is determined
4067 * by the depth of chaining and should be rel 2752 * by the depth of chaining and should be relatively short. Whine if it
4068 * takes too long. 2753 * takes too long.
4069 */ 2754 */
4070 void drain_workqueue(struct workqueue_struct 2755 void drain_workqueue(struct workqueue_struct *wq)
4071 { 2756 {
4072 unsigned int flush_cnt = 0; 2757 unsigned int flush_cnt = 0;
4073 struct pool_workqueue *pwq; 2758 struct pool_workqueue *pwq;
4074 2759
4075 /* 2760 /*
4076 * __queue_work() needs to test wheth 2761 * __queue_work() needs to test whether there are drainers, is much
4077 * hotter than drain_workqueue() and 2762 * hotter than drain_workqueue() and already looks at @wq->flags.
4078 * Use __WQ_DRAINING so that queue do 2763 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
4079 */ 2764 */
4080 mutex_lock(&wq->mutex); 2765 mutex_lock(&wq->mutex);
4081 if (!wq->nr_drainers++) 2766 if (!wq->nr_drainers++)
4082 wq->flags |= __WQ_DRAINING; 2767 wq->flags |= __WQ_DRAINING;
4083 mutex_unlock(&wq->mutex); 2768 mutex_unlock(&wq->mutex);
4084 reflush: 2769 reflush:
4085 __flush_workqueue(wq); !! 2770 flush_workqueue(wq);
4086 2771
4087 mutex_lock(&wq->mutex); 2772 mutex_lock(&wq->mutex);
4088 2773
4089 for_each_pwq(pwq, wq) { 2774 for_each_pwq(pwq, wq) {
4090 bool drained; 2775 bool drained;
4091 2776
4092 raw_spin_lock_irq(&pwq->pool- !! 2777 spin_lock_irq(&pwq->pool->lock);
4093 drained = pwq_is_empty(pwq); !! 2778 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
4094 raw_spin_unlock_irq(&pwq->poo !! 2779 spin_unlock_irq(&pwq->pool->lock);
4095 2780
4096 if (drained) 2781 if (drained)
4097 continue; 2782 continue;
4098 2783
4099 if (++flush_cnt == 10 || 2784 if (++flush_cnt == 10 ||
4100 (flush_cnt % 100 == 0 && 2785 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
4101 pr_warn("workqueue %s !! 2786 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
4102 wq->name, __f !! 2787 wq->name, flush_cnt);
4103 2788
4104 mutex_unlock(&wq->mutex); 2789 mutex_unlock(&wq->mutex);
4105 goto reflush; 2790 goto reflush;
4106 } 2791 }
4107 2792
4108 if (!--wq->nr_drainers) 2793 if (!--wq->nr_drainers)
4109 wq->flags &= ~__WQ_DRAINING; 2794 wq->flags &= ~__WQ_DRAINING;
4110 mutex_unlock(&wq->mutex); 2795 mutex_unlock(&wq->mutex);
4111 } 2796 }
4112 EXPORT_SYMBOL_GPL(drain_workqueue); 2797 EXPORT_SYMBOL_GPL(drain_workqueue);
4113 2798
4114 static bool start_flush_work(struct work_stru !! 2799 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
4115 bool from_cancel <<
4116 { 2800 {
4117 struct worker *worker = NULL; 2801 struct worker *worker = NULL;
4118 struct worker_pool *pool; 2802 struct worker_pool *pool;
4119 struct pool_workqueue *pwq; 2803 struct pool_workqueue *pwq;
4120 struct workqueue_struct *wq; <<
4121 2804
4122 rcu_read_lock(); !! 2805 might_sleep();
>> 2806
>> 2807 local_irq_disable();
4123 pool = get_work_pool(work); 2808 pool = get_work_pool(work);
4124 if (!pool) { 2809 if (!pool) {
4125 rcu_read_unlock(); !! 2810 local_irq_enable();
4126 return false; 2811 return false;
4127 } 2812 }
4128 2813
4129 raw_spin_lock_irq(&pool->lock); !! 2814 spin_lock(&pool->lock);
4130 /* see the comment in try_to_grab_pen 2815 /* see the comment in try_to_grab_pending() with the same code */
4131 pwq = get_work_pwq(work); 2816 pwq = get_work_pwq(work);
4132 if (pwq) { 2817 if (pwq) {
4133 if (unlikely(pwq->pool != poo 2818 if (unlikely(pwq->pool != pool))
4134 goto already_gone; 2819 goto already_gone;
4135 } else { 2820 } else {
4136 worker = find_worker_executin 2821 worker = find_worker_executing_work(pool, work);
4137 if (!worker) 2822 if (!worker)
4138 goto already_gone; 2823 goto already_gone;
4139 pwq = worker->current_pwq; 2824 pwq = worker->current_pwq;
4140 } 2825 }
4141 2826
4142 wq = pwq->wq; !! 2827 check_flush_dependency(pwq->wq, work);
4143 check_flush_dependency(wq, work); <<
4144 2828
4145 insert_wq_barrier(pwq, barr, work, wo 2829 insert_wq_barrier(pwq, barr, work, worker);
4146 raw_spin_unlock_irq(&pool->lock); !! 2830 spin_unlock_irq(&pool->lock);
4147 <<
4148 touch_work_lockdep_map(work, wq); <<
4149 2831
4150 /* 2832 /*
4151 * Force a lock recursion deadlock wh 2833 * Force a lock recursion deadlock when using flush_work() inside a
4152 * single-threaded or rescuer equippe 2834 * single-threaded or rescuer equipped workqueue.
4153 * 2835 *
4154 * For single threaded workqueues the 2836 * For single threaded workqueues the deadlock happens when the work
4155 * is after the work issuing the flus 2837 * is after the work issuing the flush_work(). For rescuer equipped
4156 * workqueues the deadlock happens wh 2838 * workqueues the deadlock happens when the rescuer stalls, blocking
4157 * forward progress. 2839 * forward progress.
4158 */ 2840 */
4159 if (!from_cancel && (wq->saved_max_ac !! 2841 if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer) {
4160 touch_wq_lockdep_map(wq); !! 2842 lock_map_acquire(&pwq->wq->lockdep_map);
>> 2843 lock_map_release(&pwq->wq->lockdep_map);
>> 2844 }
4161 2845
4162 rcu_read_unlock(); <<
4163 return true; 2846 return true;
4164 already_gone: 2847 already_gone:
4165 raw_spin_unlock_irq(&pool->lock); !! 2848 spin_unlock_irq(&pool->lock);
4166 rcu_read_unlock(); <<
4167 return false; 2849 return false;
4168 } 2850 }
4169 2851
4170 static bool __flush_work(struct work_struct * !! 2852 /**
>> 2853 * flush_work - wait for a work to finish executing the last queueing instance
>> 2854 * @work: the work to flush
>> 2855 *
>> 2856 * Wait until @work has finished execution. @work is guaranteed to be idle
>> 2857 * on return if it hasn't been requeued since flush started.
>> 2858 *
>> 2859 * Return:
>> 2860 * %true if flush_work() waited for the work to finish execution,
>> 2861 * %false if it was already idle.
>> 2862 */
>> 2863 bool flush_work(struct work_struct *work)
4171 { 2864 {
4172 struct wq_barrier barr; 2865 struct wq_barrier barr;
4173 2866
4174 if (WARN_ON(!wq_online)) 2867 if (WARN_ON(!wq_online))
4175 return false; 2868 return false;
4176 2869
4177 if (WARN_ON(!work->func)) !! 2870 if (start_flush_work(work, &barr)) {
>> 2871 wait_for_completion(&barr.done);
>> 2872 destroy_work_on_stack(&barr.work);
>> 2873 return true;
>> 2874 } else {
4178 return false; 2875 return false;
>> 2876 }
>> 2877 }
>> 2878 EXPORT_SYMBOL_GPL(flush_work);
4179 2879
4180 if (!start_flush_work(work, &barr, fr !! 2880 struct cwt_wait {
4181 return false; !! 2881 wait_queue_entry_t wait;
>> 2882 struct work_struct *work;
>> 2883 };
>> 2884
>> 2885 static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
>> 2886 {
>> 2887 struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
>> 2888
>> 2889 if (cwait->work != key)
>> 2890 return 0;
>> 2891 return autoremove_wake_function(wait, mode, sync, key);
>> 2892 }
>> 2893
>> 2894 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
>> 2895 {
>> 2896 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
>> 2897 unsigned long flags;
>> 2898 int ret;
>> 2899
>> 2900 do {
>> 2901 ret = try_to_grab_pending(work, is_dwork, &flags);
>> 2902 /*
>> 2903 * If someone else is already canceling, wait for it to
>> 2904 * finish. flush_work() doesn't work for PREEMPT_NONE
>> 2905 * because we may get scheduled between @work's completion
>> 2906 * and the other canceling task resuming and clearing
>> 2907 * CANCELING - flush_work() will return false immediately
>> 2908 * as @work is no longer busy, try_to_grab_pending() will
>> 2909 * return -ENOENT as @work is still being canceled and the
>> 2910 * other canceling task won't be able to clear CANCELING as
>> 2911 * we're hogging the CPU.
>> 2912 *
>> 2913 * Let's wait for completion using a waitqueue. As this
>> 2914 * may lead to the thundering herd problem, use a custom
>> 2915 * wake function which matches @work along with exclusive
>> 2916 * wait and wakeup.
>> 2917 */
>> 2918 if (unlikely(ret == -ENOENT)) {
>> 2919 struct cwt_wait cwait;
>> 2920
>> 2921 init_wait(&cwait.wait);
>> 2922 cwait.wait.func = cwt_wakefn;
>> 2923 cwait.work = work;
>> 2924
>> 2925 prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
>> 2926 TASK_UNINTERRUPTIBLE);
>> 2927 if (work_is_canceling(work))
>> 2928 schedule();
>> 2929 finish_wait(&cancel_waitq, &cwait.wait);
>> 2930 }
>> 2931 } while (unlikely(ret < 0));
>> 2932
>> 2933 /* tell other tasks trying to grab @work to back off */
>> 2934 mark_work_canceling(work);
>> 2935 local_irq_restore(flags);
4182 2936
4183 /* 2937 /*
4184 * start_flush_work() returned %true. !! 2938 * This allows canceling during early boot. We know that @work
4185 * that @work must have been executin !! 2939 * isn't executing.
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 */ 2940 */
4190 if (from_cancel) { !! 2941 if (wq_online)
4191 unsigned long data = *work_da !! 2942 flush_work(work);
4192 2943
4193 if (!WARN_ON_ONCE(data & WORK !! 2944 clear_work_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 } <<
4215 2945
4216 wait_for_completion(&barr.done); !! 2946 /*
>> 2947 * Paired with prepare_to_wait() above so that either
>> 2948 * waitqueue_active() is visible here or !work_is_canceling() is
>> 2949 * visible there.
>> 2950 */
>> 2951 smp_mb();
>> 2952 if (waitqueue_active(&cancel_waitq))
>> 2953 __wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
4217 2954
4218 out_destroy: !! 2955 return ret;
4219 destroy_work_on_stack(&barr.work); <<
4220 return true; <<
4221 } 2956 }
4222 2957
4223 /** 2958 /**
4224 * flush_work - wait for a work to finish exe !! 2959 * cancel_work_sync - cancel a work and wait for it to finish
4225 * @work: the work to flush !! 2960 * @work: the work to cancel
4226 * 2961 *
4227 * Wait until @work has finished execution. !! 2962 * Cancel @work and wait for its execution to finish. This function
4228 * on return if it hasn't been requeued since !! 2963 * can be used even if the work re-queues itself or migrates to
>> 2964 * another workqueue. On return from this function, @work is
>> 2965 * guaranteed to be not pending or executing on any CPU.
>> 2966 *
>> 2967 * cancel_work_sync(&delayed_work->work) must not be used for
>> 2968 * delayed_work's. Use cancel_delayed_work_sync() instead.
>> 2969 *
>> 2970 * The caller must ensure that the workqueue on which @work was last
>> 2971 * queued can't be destroyed before this function returns.
4229 * 2972 *
4230 * Return: 2973 * Return:
4231 * %true if flush_work() waited for the work !! 2974 * %true if @work was pending, %false otherwise.
4232 * %false if it was already idle. <<
4233 */ 2975 */
4234 bool flush_work(struct work_struct *work) !! 2976 bool cancel_work_sync(struct work_struct *work)
4235 { 2977 {
4236 might_sleep(); !! 2978 return __cancel_work_timer(work, false);
4237 return __flush_work(work, false); <<
4238 } 2979 }
4239 EXPORT_SYMBOL_GPL(flush_work); !! 2980 EXPORT_SYMBOL_GPL(cancel_work_sync);
4240 2981
4241 /** 2982 /**
4242 * flush_delayed_work - wait for a dwork to f 2983 * flush_delayed_work - wait for a dwork to finish executing the last queueing
4243 * @dwork: the delayed work to flush 2984 * @dwork: the delayed work to flush
4244 * 2985 *
4245 * Delayed timer is cancelled and the pending 2986 * Delayed timer is cancelled and the pending work is queued for
4246 * immediate execution. Like flush_work(), t 2987 * immediate execution. Like flush_work(), this function only
4247 * considers the last queueing instance of @d 2988 * considers the last queueing instance of @dwork.
4248 * 2989 *
4249 * Return: 2990 * Return:
4250 * %true if flush_work() waited for the work 2991 * %true if flush_work() waited for the work to finish execution,
4251 * %false if it was already idle. 2992 * %false if it was already idle.
4252 */ 2993 */
4253 bool flush_delayed_work(struct delayed_work * 2994 bool flush_delayed_work(struct delayed_work *dwork)
4254 { 2995 {
4255 local_irq_disable(); 2996 local_irq_disable();
4256 if (del_timer_sync(&dwork->timer)) 2997 if (del_timer_sync(&dwork->timer))
4257 __queue_work(dwork->cpu, dwor 2998 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
4258 local_irq_enable(); 2999 local_irq_enable();
4259 return flush_work(&dwork->work); 3000 return flush_work(&dwork->work);
4260 } 3001 }
4261 EXPORT_SYMBOL(flush_delayed_work); 3002 EXPORT_SYMBOL(flush_delayed_work);
4262 3003
4263 /** !! 3004 static bool __cancel_work(struct work_struct *work, bool is_dwork)
4264 * flush_rcu_work - wait for a rwork to finis <<
4265 * @rwork: the rcu work to flush <<
4266 * <<
4267 * Return: <<
4268 * %true if flush_rcu_work() waited for the w <<
4269 * %false if it was already idle. <<
4270 */ <<
4271 bool flush_rcu_work(struct rcu_work *rwork) <<
4272 { <<
4273 if (test_bit(WORK_STRUCT_PENDING_BIT, <<
4274 rcu_barrier(); <<
4275 flush_work(&rwork->work); <<
4276 return true; <<
4277 } else { <<
4278 return flush_work(&rwork->wor <<
4279 } <<
4280 } <<
4281 EXPORT_SYMBOL(flush_rcu_work); <<
4282 <<
4283 static void work_offqd_disable(struct work_of <<
4284 { 3005 {
4285 const unsigned long max = (1lu << WOR !! 3006 unsigned long flags;
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 { <<
4295 if (likely(offqd->disable > 0)) <<
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; 3007 int ret;
4306 3008
4307 ret = work_grab_pending(work, cflags, !! 3009 do {
4308 !! 3010 ret = try_to_grab_pending(work, is_dwork, &flags);
4309 work_offqd_unpack(&offqd, *work_data_ !! 3011 } 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 3012
4338 if (!(cflags & WORK_CANCEL_DISABLE)) !! 3013 if (unlikely(ret < 0))
4339 enable_work(work); !! 3014 return false;
4340 3015
>> 3016 set_work_pool_and_clear_pending(work, get_work_pool_id(work));
>> 3017 local_irq_restore(flags);
4341 return ret; 3018 return ret;
4342 } 3019 }
4343 3020
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 /** 3021 /**
4378 * cancel_delayed_work - cancel a delayed wor 3022 * cancel_delayed_work - cancel a delayed work
4379 * @dwork: delayed_work to cancel 3023 * @dwork: delayed_work to cancel
4380 * 3024 *
4381 * Kill off a pending delayed_work. 3025 * Kill off a pending delayed_work.
4382 * 3026 *
4383 * Return: %true if @dwork was pending and ca 3027 * Return: %true if @dwork was pending and canceled; %false if it wasn't
4384 * pending. 3028 * pending.
4385 * 3029 *
4386 * Note: 3030 * Note:
4387 * The work callback function may still be ru 3031 * The work callback function may still be running on return, unless
4388 * it returns %true and the work doesn't re-a 3032 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
4389 * use cancel_delayed_work_sync() to wait on 3033 * use cancel_delayed_work_sync() to wait on it.
4390 * 3034 *
4391 * This function is safe to call from any con 3035 * This function is safe to call from any context including IRQ handler.
4392 */ 3036 */
4393 bool cancel_delayed_work(struct delayed_work 3037 bool cancel_delayed_work(struct delayed_work *dwork)
4394 { 3038 {
4395 return __cancel_work(&dwork->work, WO !! 3039 return __cancel_work(&dwork->work, true);
4396 } 3040 }
4397 EXPORT_SYMBOL(cancel_delayed_work); 3041 EXPORT_SYMBOL(cancel_delayed_work);
4398 3042
4399 /** 3043 /**
4400 * cancel_delayed_work_sync - cancel a delaye 3044 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
4401 * @dwork: the delayed work cancel 3045 * @dwork: the delayed work cancel
4402 * 3046 *
4403 * This is cancel_work_sync() for delayed wor 3047 * This is cancel_work_sync() for delayed works.
4404 * 3048 *
4405 * Return: 3049 * Return:
4406 * %true if @dwork was pending, %false otherw 3050 * %true if @dwork was pending, %false otherwise.
4407 */ 3051 */
4408 bool cancel_delayed_work_sync(struct delayed_ 3052 bool cancel_delayed_work_sync(struct delayed_work *dwork)
4409 { 3053 {
4410 return __cancel_work_sync(&dwork->wor !! 3054 return __cancel_work_timer(&dwork->work, true);
4411 } 3055 }
4412 EXPORT_SYMBOL(cancel_delayed_work_sync); 3056 EXPORT_SYMBOL(cancel_delayed_work_sync);
4413 3057
4414 /** 3058 /**
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 3059 * schedule_on_each_cpu - execute a function synchronously on each online CPU
4518 * @func: the function to call 3060 * @func: the function to call
4519 * 3061 *
4520 * schedule_on_each_cpu() executes @func on e 3062 * schedule_on_each_cpu() executes @func on each online CPU using the
4521 * system workqueue and blocks until all CPUs 3063 * system workqueue and blocks until all CPUs have completed.
4522 * schedule_on_each_cpu() is very slow. 3064 * schedule_on_each_cpu() is very slow.
4523 * 3065 *
4524 * Return: 3066 * Return:
4525 * 0 on success, -errno on failure. 3067 * 0 on success, -errno on failure.
4526 */ 3068 */
4527 int schedule_on_each_cpu(work_func_t func) 3069 int schedule_on_each_cpu(work_func_t func)
4528 { 3070 {
4529 int cpu; 3071 int cpu;
4530 struct work_struct __percpu *works; 3072 struct work_struct __percpu *works;
4531 3073
4532 works = alloc_percpu(struct work_stru 3074 works = alloc_percpu(struct work_struct);
4533 if (!works) 3075 if (!works)
4534 return -ENOMEM; 3076 return -ENOMEM;
4535 3077
4536 cpus_read_lock(); !! 3078 get_online_cpus();
4537 3079
4538 for_each_online_cpu(cpu) { 3080 for_each_online_cpu(cpu) {
4539 struct work_struct *work = pe 3081 struct work_struct *work = per_cpu_ptr(works, cpu);
4540 3082
4541 INIT_WORK(work, func); 3083 INIT_WORK(work, func);
4542 schedule_work_on(cpu, work); 3084 schedule_work_on(cpu, work);
4543 } 3085 }
4544 3086
4545 for_each_online_cpu(cpu) 3087 for_each_online_cpu(cpu)
4546 flush_work(per_cpu_ptr(works, 3088 flush_work(per_cpu_ptr(works, cpu));
4547 3089
4548 cpus_read_unlock(); !! 3090 put_online_cpus();
4549 free_percpu(works); 3091 free_percpu(works);
4550 return 0; 3092 return 0;
4551 } 3093 }
4552 3094
4553 /** 3095 /**
4554 * execute_in_process_context - reliably exec 3096 * execute_in_process_context - reliably execute the routine with user context
4555 * @fn: the function to execute 3097 * @fn: the function to execute
4556 * @ew: guaranteed storage for the ex 3098 * @ew: guaranteed storage for the execute work structure (must
4557 * be available when the work ex 3099 * be available when the work executes)
4558 * 3100 *
4559 * Executes the function immediately if proce 3101 * Executes the function immediately if process context is available,
4560 * otherwise schedules the function for delay 3102 * otherwise schedules the function for delayed execution.
4561 * 3103 *
4562 * Return: 0 - function was executed 3104 * Return: 0 - function was executed
4563 * 1 - function was scheduled fo 3105 * 1 - function was scheduled for execution
4564 */ 3106 */
4565 int execute_in_process_context(work_func_t fn 3107 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
4566 { 3108 {
4567 if (!in_interrupt()) { 3109 if (!in_interrupt()) {
4568 fn(&ew->work); 3110 fn(&ew->work);
4569 return 0; 3111 return 0;
4570 } 3112 }
4571 3113
4572 INIT_WORK(&ew->work, fn); 3114 INIT_WORK(&ew->work, fn);
4573 schedule_work(&ew->work); 3115 schedule_work(&ew->work);
4574 3116
4575 return 1; 3117 return 1;
4576 } 3118 }
4577 EXPORT_SYMBOL_GPL(execute_in_process_context) 3119 EXPORT_SYMBOL_GPL(execute_in_process_context);
4578 3120
4579 /** 3121 /**
4580 * free_workqueue_attrs - free a workqueue_at 3122 * free_workqueue_attrs - free a workqueue_attrs
4581 * @attrs: workqueue_attrs to free 3123 * @attrs: workqueue_attrs to free
4582 * 3124 *
4583 * Undo alloc_workqueue_attrs(). 3125 * Undo alloc_workqueue_attrs().
4584 */ 3126 */
4585 void free_workqueue_attrs(struct workqueue_at 3127 void free_workqueue_attrs(struct workqueue_attrs *attrs)
4586 { 3128 {
4587 if (attrs) { 3129 if (attrs) {
4588 free_cpumask_var(attrs->cpuma 3130 free_cpumask_var(attrs->cpumask);
4589 free_cpumask_var(attrs->__pod <<
4590 kfree(attrs); 3131 kfree(attrs);
4591 } 3132 }
4592 } 3133 }
4593 3134
4594 /** 3135 /**
4595 * alloc_workqueue_attrs - allocate a workque 3136 * alloc_workqueue_attrs - allocate a workqueue_attrs
>> 3137 * @gfp_mask: allocation mask to use
4596 * 3138 *
4597 * Allocate a new workqueue_attrs, initialize 3139 * Allocate a new workqueue_attrs, initialize with default settings and
4598 * return it. 3140 * return it.
4599 * 3141 *
4600 * Return: The allocated new workqueue_attr o 3142 * Return: The allocated new workqueue_attr on success. %NULL on failure.
4601 */ 3143 */
4602 struct workqueue_attrs *alloc_workqueue_attrs !! 3144 struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
4603 { 3145 {
4604 struct workqueue_attrs *attrs; 3146 struct workqueue_attrs *attrs;
4605 3147
4606 attrs = kzalloc(sizeof(*attrs), GFP_K !! 3148 attrs = kzalloc(sizeof(*attrs), gfp_mask);
4607 if (!attrs) 3149 if (!attrs)
4608 goto fail; 3150 goto fail;
4609 if (!alloc_cpumask_var(&attrs->cpumas !! 3151 if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
4610 goto fail; <<
4611 if (!alloc_cpumask_var(&attrs->__pod_ <<
4612 goto fail; 3152 goto fail;
4613 3153
4614 cpumask_copy(attrs->cpumask, cpu_poss 3154 cpumask_copy(attrs->cpumask, cpu_possible_mask);
4615 attrs->affn_scope = WQ_AFFN_DFL; <<
4616 return attrs; 3155 return attrs;
4617 fail: 3156 fail:
4618 free_workqueue_attrs(attrs); 3157 free_workqueue_attrs(attrs);
4619 return NULL; 3158 return NULL;
4620 } 3159 }
4621 3160
4622 static void copy_workqueue_attrs(struct workq 3161 static void copy_workqueue_attrs(struct workqueue_attrs *to,
4623 const struct 3162 const struct workqueue_attrs *from)
4624 { 3163 {
4625 to->nice = from->nice; 3164 to->nice = from->nice;
4626 cpumask_copy(to->cpumask, from->cpuma 3165 cpumask_copy(to->cpumask, from->cpumask);
4627 cpumask_copy(to->__pod_cpumask, from- <<
4628 to->affn_strict = from->affn_strict; <<
4629 <<
4630 /* 3166 /*
4631 * Unlike hash and equality test, cop !! 3167 * Unlike hash and equality test, this function doesn't ignore
4632 * fields as copying is used for both !! 3168 * ->no_numa as it is used for both pool and wq attrs. Instead,
4633 * get_unbound_pool() explicitly clea !! 3169 * get_unbound_pool() explicitly clears ->no_numa after copying.
4634 */ 3170 */
4635 to->affn_scope = from->affn_scope; !! 3171 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 } 3172 }
4650 3173
4651 /* hash value of the content of @attr */ 3174 /* hash value of the content of @attr */
4652 static u32 wqattrs_hash(const struct workqueu 3175 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
4653 { 3176 {
4654 u32 hash = 0; 3177 u32 hash = 0;
4655 3178
4656 hash = jhash_1word(attrs->nice, hash) 3179 hash = jhash_1word(attrs->nice, hash);
4657 hash = jhash_1word(attrs->affn_strict !! 3180 hash = jhash(cpumask_bits(attrs->cpumask),
4658 hash = jhash(cpumask_bits(attrs->__po <<
4659 BITS_TO_LONGS(nr_cpumask 3181 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; 3182 return hash;
4664 } 3183 }
4665 3184
4666 /* content equality test */ 3185 /* content equality test */
4667 static bool wqattrs_equal(const struct workqu 3186 static bool wqattrs_equal(const struct workqueue_attrs *a,
4668 const struct workqu 3187 const struct workqueue_attrs *b)
4669 { 3188 {
4670 if (a->nice != b->nice) 3189 if (a->nice != b->nice)
4671 return false; 3190 return false;
4672 if (a->affn_strict != b->affn_strict) !! 3191 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; 3192 return false;
4678 return true; 3193 return true;
4679 } 3194 }
4680 3195
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 /** 3196 /**
4726 * init_worker_pool - initialize a newly zall 3197 * init_worker_pool - initialize a newly zalloc'd worker_pool
4727 * @pool: worker_pool to initialize 3198 * @pool: worker_pool to initialize
4728 * 3199 *
4729 * Initialize a newly zalloc'd @pool. It als 3200 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
4730 * 3201 *
4731 * Return: 0 on success, -errno on failure. 3202 * Return: 0 on success, -errno on failure. Even on failure, all fields
4732 * inside @pool proper are initialized and pu 3203 * inside @pool proper are initialized and put_unbound_pool() can be called
4733 * on @pool safely to release it. 3204 * on @pool safely to release it.
4734 */ 3205 */
4735 static int init_worker_pool(struct worker_poo 3206 static int init_worker_pool(struct worker_pool *pool)
4736 { 3207 {
4737 raw_spin_lock_init(&pool->lock); !! 3208 spin_lock_init(&pool->lock);
4738 pool->id = -1; 3209 pool->id = -1;
4739 pool->cpu = -1; 3210 pool->cpu = -1;
4740 pool->node = NUMA_NO_NODE; 3211 pool->node = NUMA_NO_NODE;
4741 pool->flags |= POOL_DISASSOCIATED; 3212 pool->flags |= POOL_DISASSOCIATED;
4742 pool->watchdog_ts = jiffies; 3213 pool->watchdog_ts = jiffies;
4743 INIT_LIST_HEAD(&pool->worklist); 3214 INIT_LIST_HEAD(&pool->worklist);
4744 INIT_LIST_HEAD(&pool->idle_list); 3215 INIT_LIST_HEAD(&pool->idle_list);
4745 hash_init(pool->busy_hash); 3216 hash_init(pool->busy_hash);
4746 3217
4747 timer_setup(&pool->idle_timer, idle_w 3218 timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE);
4748 INIT_WORK(&pool->idle_cull_work, idle <<
4749 3219
4750 timer_setup(&pool->mayday_timer, pool 3220 timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0);
4751 3221
>> 3222 mutex_init(&pool->attach_mutex);
4752 INIT_LIST_HEAD(&pool->workers); 3223 INIT_LIST_HEAD(&pool->workers);
4753 3224
4754 ida_init(&pool->worker_ida); 3225 ida_init(&pool->worker_ida);
4755 INIT_HLIST_NODE(&pool->hash_node); 3226 INIT_HLIST_NODE(&pool->hash_node);
4756 pool->refcnt = 1; 3227 pool->refcnt = 1;
4757 3228
4758 /* shouldn't fail above this point */ 3229 /* shouldn't fail above this point */
4759 pool->attrs = alloc_workqueue_attrs() !! 3230 pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
4760 if (!pool->attrs) 3231 if (!pool->attrs)
4761 return -ENOMEM; 3232 return -ENOMEM;
4762 <<
4763 wqattrs_clear_for_pool(pool->attrs); <<
4764 <<
4765 return 0; <<
4766 } <<
4767 <<
4768 #ifdef CONFIG_LOCKDEP <<
4769 static void wq_init_lockdep(struct workqueue_ <<
4770 { <<
4771 char *lock_name; <<
4772 <<
4773 lockdep_register_key(&wq->key); <<
4774 lock_name = kasprintf(GFP_KERNEL, "%s <<
4775 if (!lock_name) <<
4776 lock_name = wq->name; <<
4777 <<
4778 wq->lock_name = lock_name; <<
4779 lockdep_init_map(&wq->lockdep_map, lo <<
4780 } <<
4781 <<
4782 static void wq_unregister_lockdep(struct work <<
4783 { <<
4784 lockdep_unregister_key(&wq->key); <<
4785 } <<
4786 <<
4787 static void wq_free_lockdep(struct workqueue_ <<
4788 { <<
4789 if (wq->lock_name != wq->name) <<
4790 kfree(wq->lock_name); <<
4791 } <<
4792 #else <<
4793 static void wq_init_lockdep(struct workqueue_ <<
4794 { <<
4795 } <<
4796 <<
4797 static void wq_unregister_lockdep(struct work <<
4798 { <<
4799 } <<
4800 <<
4801 static void wq_free_lockdep(struct workqueue_ <<
4802 { <<
4803 } <<
4804 #endif <<
4805 <<
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; 3233 return 0;
4852 <<
4853 err_free: <<
4854 free_node_nr_active(nna_ar); <<
4855 return -ENOMEM; <<
4856 } 3234 }
4857 3235
4858 static void rcu_free_wq(struct rcu_head *rcu) 3236 static void rcu_free_wq(struct rcu_head *rcu)
4859 { 3237 {
4860 struct workqueue_struct *wq = 3238 struct workqueue_struct *wq =
4861 container_of(rcu, struct work 3239 container_of(rcu, struct workqueue_struct, rcu);
4862 3240
4863 if (wq->flags & WQ_UNBOUND) !! 3241 if (!(wq->flags & WQ_UNBOUND))
4864 free_node_nr_active(wq->node_ !! 3242 free_percpu(wq->cpu_pwqs);
>> 3243 else
>> 3244 free_workqueue_attrs(wq->unbound_attrs);
4865 3245
4866 wq_free_lockdep(wq); !! 3246 kfree(wq->rescuer);
4867 free_percpu(wq->cpu_pwq); <<
4868 free_workqueue_attrs(wq->unbound_attr <<
4869 kfree(wq); 3247 kfree(wq);
4870 } 3248 }
4871 3249
4872 static void rcu_free_pool(struct rcu_head *rc 3250 static void rcu_free_pool(struct rcu_head *rcu)
4873 { 3251 {
4874 struct worker_pool *pool = container_ 3252 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
4875 3253
4876 ida_destroy(&pool->worker_ida); 3254 ida_destroy(&pool->worker_ida);
4877 free_workqueue_attrs(pool->attrs); 3255 free_workqueue_attrs(pool->attrs);
4878 kfree(pool); 3256 kfree(pool);
4879 } 3257 }
4880 3258
4881 /** 3259 /**
4882 * put_unbound_pool - put a worker_pool 3260 * put_unbound_pool - put a worker_pool
4883 * @pool: worker_pool to put 3261 * @pool: worker_pool to put
4884 * 3262 *
4885 * Put @pool. If its refcnt reaches zero, it !! 3263 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
4886 * safe manner. get_unbound_pool() calls thi 3264 * safe manner. get_unbound_pool() calls this function on its failure path
4887 * and this function should be able to releas 3265 * and this function should be able to release pools which went through,
4888 * successfully or not, init_worker_pool(). 3266 * successfully or not, init_worker_pool().
4889 * 3267 *
4890 * Should be called with wq_pool_mutex held. 3268 * Should be called with wq_pool_mutex held.
4891 */ 3269 */
4892 static void put_unbound_pool(struct worker_po 3270 static void put_unbound_pool(struct worker_pool *pool)
4893 { 3271 {
>> 3272 DECLARE_COMPLETION_ONSTACK(detach_completion);
4894 struct worker *worker; 3273 struct worker *worker;
4895 LIST_HEAD(cull_list); <<
4896 3274
4897 lockdep_assert_held(&wq_pool_mutex); 3275 lockdep_assert_held(&wq_pool_mutex);
4898 3276
4899 if (--pool->refcnt) 3277 if (--pool->refcnt)
4900 return; 3278 return;
4901 3279
4902 /* sanity checks */ 3280 /* sanity checks */
4903 if (WARN_ON(!(pool->cpu < 0)) || 3281 if (WARN_ON(!(pool->cpu < 0)) ||
4904 WARN_ON(!list_empty(&pool->workli 3282 WARN_ON(!list_empty(&pool->worklist)))
4905 return; 3283 return;
4906 3284
4907 /* release id and unhash */ 3285 /* release id and unhash */
4908 if (pool->id >= 0) 3286 if (pool->id >= 0)
4909 idr_remove(&worker_pool_idr, 3287 idr_remove(&worker_pool_idr, pool->id);
4910 hash_del(&pool->hash_node); 3288 hash_del(&pool->hash_node);
4911 3289
4912 /* 3290 /*
4913 * Become the manager and destroy all 3291 * Become the manager and destroy all workers. This prevents
4914 * @pool's workers from blocking on a 3292 * @pool's workers from blocking on attach_mutex. We're the last
4915 * manager and @pool gets freed with 3293 * manager and @pool gets freed with the flag set.
4916 * <<
4917 * Having a concurrent manager is qui <<
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 */ 3294 */
4925 while (true) { !! 3295 spin_lock_irq(&pool->lock);
4926 rcuwait_wait_event(&manager_w !! 3296 wait_event_lock_irq(wq_manager_wait,
4927 !(pool->fl !! 3297 !(pool->flags & POOL_MANAGER_ACTIVE), pool->lock);
4928 TASK_UNINT !! 3298 pool->flags |= POOL_MANAGER_ACTIVE;
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 3299
4940 while ((worker = first_idle_worker(po 3300 while ((worker = first_idle_worker(pool)))
4941 set_worker_dying(worker, &cul !! 3301 destroy_worker(worker);
4942 WARN_ON(pool->nr_workers || pool->nr_ 3302 WARN_ON(pool->nr_workers || pool->nr_idle);
4943 raw_spin_unlock_irq(&pool->lock); !! 3303 spin_unlock_irq(&pool->lock);
4944 <<
4945 detach_dying_workers(&cull_list); <<
4946 3304
4947 mutex_unlock(&wq_pool_attach_mutex); !! 3305 mutex_lock(&pool->attach_mutex);
>> 3306 if (!list_empty(&pool->workers))
>> 3307 pool->detach_completion = &detach_completion;
>> 3308 mutex_unlock(&pool->attach_mutex);
4948 3309
4949 reap_dying_workers(&cull_list); !! 3310 if (pool->detach_completion)
>> 3311 wait_for_completion(pool->detach_completion);
4950 3312
4951 /* shut down the timers */ 3313 /* shut down the timers */
4952 del_timer_sync(&pool->idle_timer); 3314 del_timer_sync(&pool->idle_timer);
4953 cancel_work_sync(&pool->idle_cull_wor <<
4954 del_timer_sync(&pool->mayday_timer); 3315 del_timer_sync(&pool->mayday_timer);
4955 3316
4956 /* RCU protected to allow dereference !! 3317 /* sched-RCU protected to allow dereferences from get_work_pool() */
4957 call_rcu(&pool->rcu, rcu_free_pool); !! 3318 call_rcu_sched(&pool->rcu, rcu_free_pool);
4958 } 3319 }
4959 3320
4960 /** 3321 /**
4961 * get_unbound_pool - get a worker_pool with 3322 * get_unbound_pool - get a worker_pool with the specified attributes
4962 * @attrs: the attributes of the worker_pool 3323 * @attrs: the attributes of the worker_pool to get
4963 * 3324 *
4964 * Obtain a worker_pool which has the same at 3325 * Obtain a worker_pool which has the same attributes as @attrs, bump the
4965 * reference count and return it. If there a 3326 * reference count and return it. If there already is a matching
4966 * worker_pool, it will be used; otherwise, t 3327 * worker_pool, it will be used; otherwise, this function attempts to
4967 * create a new one. 3328 * create a new one.
4968 * 3329 *
4969 * Should be called with wq_pool_mutex held. 3330 * Should be called with wq_pool_mutex held.
4970 * 3331 *
4971 * Return: On success, a worker_pool with the 3332 * Return: On success, a worker_pool with the same attributes as @attrs.
4972 * On failure, %NULL. 3333 * On failure, %NULL.
4973 */ 3334 */
4974 static struct worker_pool *get_unbound_pool(c 3335 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
4975 { 3336 {
4976 struct wq_pod_type *pt = &wq_pod_type <<
4977 u32 hash = wqattrs_hash(attrs); 3337 u32 hash = wqattrs_hash(attrs);
4978 struct worker_pool *pool; 3338 struct worker_pool *pool;
4979 int pod, node = NUMA_NO_NODE; !! 3339 int node;
>> 3340 int target_node = NUMA_NO_NODE;
4980 3341
4981 lockdep_assert_held(&wq_pool_mutex); 3342 lockdep_assert_held(&wq_pool_mutex);
4982 3343
4983 /* do we already have a matching pool 3344 /* do we already have a matching pool? */
4984 hash_for_each_possible(unbound_pool_h 3345 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
4985 if (wqattrs_equal(pool->attrs 3346 if (wqattrs_equal(pool->attrs, attrs)) {
4986 pool->refcnt++; 3347 pool->refcnt++;
4987 return pool; 3348 return pool;
4988 } 3349 }
4989 } 3350 }
4990 3351
4991 /* If __pod_cpumask is contained insi !! 3352 /* if cpumask is contained inside a NUMA node, we belong to that node */
4992 for (pod = 0; pod < pt->nr_pods; pod+ !! 3353 if (wq_numa_enabled) {
4993 if (cpumask_subset(attrs->__p !! 3354 for_each_node(node) {
4994 node = pt->pod_node[p !! 3355 if (cpumask_subset(attrs->cpumask,
4995 break; !! 3356 wq_numa_possible_cpumask[node])) {
>> 3357 target_node = node;
>> 3358 break;
>> 3359 }
4996 } 3360 }
4997 } 3361 }
4998 3362
4999 /* nope, create a new one */ 3363 /* nope, create a new one */
5000 pool = kzalloc_node(sizeof(*pool), GF !! 3364 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node);
5001 if (!pool || init_worker_pool(pool) < 3365 if (!pool || init_worker_pool(pool) < 0)
5002 goto fail; 3366 goto fail;
5003 3367
5004 pool->node = node; !! 3368 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
5005 copy_workqueue_attrs(pool->attrs, att 3369 copy_workqueue_attrs(pool->attrs, attrs);
5006 wqattrs_clear_for_pool(pool->attrs); !! 3370 pool->node = target_node;
>> 3371
>> 3372 /*
>> 3373 * no_numa isn't a worker_pool attribute, always clear it. See
>> 3374 * 'struct workqueue_attrs' comments for detail.
>> 3375 */
>> 3376 pool->attrs->no_numa = false;
5007 3377
5008 if (worker_pool_assign_id(pool) < 0) 3378 if (worker_pool_assign_id(pool) < 0)
5009 goto fail; 3379 goto fail;
5010 3380
5011 /* create and start the initial worke 3381 /* create and start the initial worker */
5012 if (wq_online && !create_worker(pool) 3382 if (wq_online && !create_worker(pool))
5013 goto fail; 3383 goto fail;
5014 3384
5015 /* install */ 3385 /* install */
5016 hash_add(unbound_pool_hash, &pool->ha 3386 hash_add(unbound_pool_hash, &pool->hash_node, hash);
5017 3387
5018 return pool; 3388 return pool;
5019 fail: 3389 fail:
5020 if (pool) 3390 if (pool)
5021 put_unbound_pool(pool); 3391 put_unbound_pool(pool);
5022 return NULL; 3392 return NULL;
5023 } 3393 }
5024 3394
>> 3395 static void rcu_free_pwq(struct rcu_head *rcu)
>> 3396 {
>> 3397 kmem_cache_free(pwq_cache,
>> 3398 container_of(rcu, struct pool_workqueue, rcu));
>> 3399 }
>> 3400
5025 /* 3401 /*
5026 * Scheduled on pwq_release_worker by put_pwq !! 3402 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
5027 * refcnt and needs to be destroyed. !! 3403 * and needs to be destroyed.
5028 */ 3404 */
5029 static void pwq_release_workfn(struct kthread !! 3405 static void pwq_unbound_release_workfn(struct work_struct *work)
5030 { 3406 {
5031 struct pool_workqueue *pwq = containe 3407 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
5032 !! 3408 unbound_release_work);
5033 struct workqueue_struct *wq = pwq->wq 3409 struct workqueue_struct *wq = pwq->wq;
5034 struct worker_pool *pool = pwq->pool; 3410 struct worker_pool *pool = pwq->pool;
5035 bool is_last = false; !! 3411 bool is_last;
5036 3412
5037 /* !! 3413 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
5038 * When @pwq is not linked, it doesn' !! 3414 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 3415
5046 /* !! 3416 mutex_lock(&wq->mutex);
5047 * For ordered workqueue with !! 3417 list_del_rcu(&pwq->pwqs_node);
5048 */ !! 3418 is_last = list_empty(&wq->pwqs);
5049 if (!is_last && (wq->flags & !! 3419 mutex_unlock(&wq->mutex);
5050 unplug_oldest_pwq(wq) <<
5051 3420
5052 mutex_unlock(&wq->mutex); !! 3421 mutex_lock(&wq_pool_mutex);
5053 } !! 3422 put_unbound_pool(pool);
>> 3423 mutex_unlock(&wq_pool_mutex);
5054 3424
5055 if (wq->flags & WQ_UNBOUND) { !! 3425 call_rcu_sched(&pwq->rcu, rcu_free_pwq);
5056 mutex_lock(&wq_pool_mutex); !! 3426
5057 put_unbound_pool(pool); !! 3427 /*
5058 mutex_unlock(&wq_pool_mutex); !! 3428 * If we're the last pwq going away, @wq is already dead and no one
5059 } !! 3429 * is gonna access it anymore. Schedule RCU free.
>> 3430 */
>> 3431 if (is_last)
>> 3432 call_rcu_sched(&wq->rcu, rcu_free_wq);
>> 3433 }
5060 3434
5061 if (!list_empty(&pwq->pending_node)) !! 3435 /**
5062 struct wq_node_nr_active *nna !! 3436 * pwq_adjust_max_active - update a pwq's max_active to the current setting
5063 wq_node_nr_active(pwq !! 3437 * @pwq: target pool_workqueue
>> 3438 *
>> 3439 * If @pwq isn't freezing, set @pwq->max_active to the associated
>> 3440 * workqueue's saved_max_active and activate delayed work items
>> 3441 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
>> 3442 */
>> 3443 static void pwq_adjust_max_active(struct pool_workqueue *pwq)
>> 3444 {
>> 3445 struct workqueue_struct *wq = pwq->wq;
>> 3446 bool freezable = wq->flags & WQ_FREEZABLE;
>> 3447 unsigned long flags;
5064 3448
5065 raw_spin_lock_irq(&nna->lock) !! 3449 /* for @wq->saved_max_active */
5066 list_del_init(&pwq->pending_n !! 3450 lockdep_assert_held(&wq->mutex);
5067 raw_spin_unlock_irq(&nna->loc !! 3451
5068 } !! 3452 /* fast exit for non-freezable wqs */
>> 3453 if (!freezable && pwq->max_active == wq->saved_max_active)
>> 3454 return;
5069 3455
5070 kfree_rcu(pwq, rcu); !! 3456 /* this function can be called during early boot w/ irq disabled */
>> 3457 spin_lock_irqsave(&pwq->pool->lock, flags);
5071 3458
5072 /* 3459 /*
5073 * If we're the last pwq going away, !! 3460 * During [un]freezing, the caller is responsible for ensuring that
5074 * is gonna access it anymore. Sched !! 3461 * this function is called at least once after @workqueue_freezing
>> 3462 * is updated and visible.
5075 */ 3463 */
5076 if (is_last) { !! 3464 if (!freezable || !workqueue_freezing) {
5077 wq_unregister_lockdep(wq); !! 3465 pwq->max_active = wq->saved_max_active;
5078 call_rcu(&wq->rcu, rcu_free_w !! 3466
>> 3467 while (!list_empty(&pwq->delayed_works) &&
>> 3468 pwq->nr_active < pwq->max_active)
>> 3469 pwq_activate_first_delayed(pwq);
>> 3470
>> 3471 /*
>> 3472 * Need to kick a worker after thawed or an unbound wq's
>> 3473 * max_active is bumped. It's a slow path. Do it always.
>> 3474 */
>> 3475 wake_up_worker(pwq->pool);
>> 3476 } else {
>> 3477 pwq->max_active = 0;
5079 } 3478 }
>> 3479
>> 3480 spin_unlock_irqrestore(&pwq->pool->lock, flags);
5080 } 3481 }
5081 3482
5082 /* initialize newly allocated @pwq which is a !! 3483 /* initialize newly alloced @pwq which is associated with @wq and @pool */
5083 static void init_pwq(struct pool_workqueue *p 3484 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
5084 struct worker_pool *pool 3485 struct worker_pool *pool)
5085 { 3486 {
5086 BUG_ON((unsigned long)pwq & ~WORK_STR !! 3487 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
5087 3488
5088 memset(pwq, 0, sizeof(*pwq)); 3489 memset(pwq, 0, sizeof(*pwq));
5089 3490
5090 pwq->pool = pool; 3491 pwq->pool = pool;
5091 pwq->wq = wq; 3492 pwq->wq = wq;
5092 pwq->flush_color = -1; 3493 pwq->flush_color = -1;
5093 pwq->refcnt = 1; 3494 pwq->refcnt = 1;
5094 INIT_LIST_HEAD(&pwq->inactive_works); !! 3495 INIT_LIST_HEAD(&pwq->delayed_works);
5095 INIT_LIST_HEAD(&pwq->pending_node); <<
5096 INIT_LIST_HEAD(&pwq->pwqs_node); 3496 INIT_LIST_HEAD(&pwq->pwqs_node);
5097 INIT_LIST_HEAD(&pwq->mayday_node); 3497 INIT_LIST_HEAD(&pwq->mayday_node);
5098 kthread_init_work(&pwq->release_work, !! 3498 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
5099 } 3499 }
5100 3500
5101 /* sync @pwq with the current state of its as 3501 /* sync @pwq with the current state of its associated wq and link it */
5102 static void link_pwq(struct pool_workqueue *p 3502 static void link_pwq(struct pool_workqueue *pwq)
5103 { 3503 {
5104 struct workqueue_struct *wq = pwq->wq 3504 struct workqueue_struct *wq = pwq->wq;
5105 3505
5106 lockdep_assert_held(&wq->mutex); 3506 lockdep_assert_held(&wq->mutex);
5107 3507
5108 /* may be called multiple times, igno 3508 /* may be called multiple times, ignore if already linked */
5109 if (!list_empty(&pwq->pwqs_node)) 3509 if (!list_empty(&pwq->pwqs_node))
5110 return; 3510 return;
5111 3511
5112 /* set the matching work_color */ 3512 /* set the matching work_color */
5113 pwq->work_color = wq->work_color; 3513 pwq->work_color = wq->work_color;
5114 3514
>> 3515 /* sync max_active to the current setting */
>> 3516 pwq_adjust_max_active(pwq);
>> 3517
5115 /* link in @pwq */ 3518 /* link in @pwq */
5116 list_add_tail_rcu(&pwq->pwqs_node, &w !! 3519 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
5117 } 3520 }
5118 3521
5119 /* obtain a pool matching @attr and create a 3522 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
5120 static struct pool_workqueue *alloc_unbound_p 3523 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
5121 const 3524 const struct workqueue_attrs *attrs)
5122 { 3525 {
5123 struct worker_pool *pool; 3526 struct worker_pool *pool;
5124 struct pool_workqueue *pwq; 3527 struct pool_workqueue *pwq;
5125 3528
5126 lockdep_assert_held(&wq_pool_mutex); 3529 lockdep_assert_held(&wq_pool_mutex);
5127 3530
5128 pool = get_unbound_pool(attrs); 3531 pool = get_unbound_pool(attrs);
5129 if (!pool) 3532 if (!pool)
5130 return NULL; 3533 return NULL;
5131 3534
5132 pwq = kmem_cache_alloc_node(pwq_cache 3535 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
5133 if (!pwq) { 3536 if (!pwq) {
5134 put_unbound_pool(pool); 3537 put_unbound_pool(pool);
5135 return NULL; 3538 return NULL;
5136 } 3539 }
5137 3540
5138 init_pwq(pwq, wq, pool); 3541 init_pwq(pwq, wq, pool);
5139 return pwq; 3542 return pwq;
5140 } 3543 }
5141 3544
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 /** 3545 /**
5153 * wq_calc_pod_cpumask - calculate a wq_attrs !! 3546 * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
5154 * @attrs: the wq_attrs of the default pwq of 3547 * @attrs: the wq_attrs of the default pwq of the target workqueue
5155 * @cpu: the target CPU !! 3548 * @node: the target NUMA node
>> 3549 * @cpu_going_down: if >= 0, the CPU to consider as offline
>> 3550 * @cpumask: outarg, the resulting cpumask
5156 * 3551 *
5157 * Calculate the cpumask a workqueue with @at !! 3552 * Calculate the cpumask a workqueue with @attrs should use on @node. If
5158 * The result is stored in @attrs->__pod_cpum !! 3553 * @cpu_going_down is >= 0, that cpu is considered offline during
>> 3554 * calculation. The result is stored in @cpumask.
5159 * 3555 *
5160 * If pod affinity is not enabled, @attrs->cp !! 3556 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
5161 * and @pod has online CPUs requested by @att !! 3557 * enabled and @node has online CPUs requested by @attrs, the returned
5162 * intersection of the possible CPUs of @pod !! 3558 * cpumask is the intersection of the possible CPUs of @node and
>> 3559 * @attrs->cpumask.
5163 * 3560 *
5164 * The caller is responsible for ensuring tha !! 3561 * The caller is responsible for ensuring that the cpumask of @node stays
>> 3562 * stable.
>> 3563 *
>> 3564 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
>> 3565 * %false if equal.
5165 */ 3566 */
5166 static void wq_calc_pod_cpumask(struct workqu !! 3567 static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
>> 3568 int cpu_going_down, cpumask_t *cpumask)
5167 { 3569 {
5168 const struct wq_pod_type *pt = wqattr !! 3570 if (!wq_numa_enabled || attrs->no_numa)
5169 int pod = pt->cpu_pod[cpu]; !! 3571 goto use_dfl;
5170 3572
5171 /* calculate possible CPUs in @pod th !! 3573 /* does @node have any online CPUs @attrs wants? */
5172 cpumask_and(attrs->__pod_cpumask, pt- !! 3574 cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
5173 /* does @pod have any online CPUs @at !! 3575 if (cpu_going_down >= 0)
5174 if (!cpumask_intersects(attrs->__pod_ !! 3576 cpumask_clear_cpu(cpu_going_down, cpumask);
5175 cpumask_copy(attrs->__pod_cpu !! 3577
5176 return; !! 3578 if (cpumask_empty(cpumask))
>> 3579 goto use_dfl;
>> 3580
>> 3581 /* yeap, return possible CPUs in @node that @attrs wants */
>> 3582 cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
>> 3583
>> 3584 if (cpumask_empty(cpumask)) {
>> 3585 pr_warn_once("WARNING: workqueue cpumask: online intersect > "
>> 3586 "possible intersect\n");
>> 3587 return false;
5177 } 3588 }
>> 3589
>> 3590 return !cpumask_equal(cpumask, attrs->cpumask);
>> 3591
>> 3592 use_dfl:
>> 3593 cpumask_copy(cpumask, attrs->cpumask);
>> 3594 return false;
5178 } 3595 }
5179 3596
5180 /* install @pwq into @wq and return the old p !! 3597 /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
5181 static struct pool_workqueue *install_unbound !! 3598 static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
5182 int c !! 3599 int node,
>> 3600 struct pool_workqueue *pwq)
5183 { 3601 {
5184 struct pool_workqueue __rcu **slot = <<
5185 struct pool_workqueue *old_pwq; 3602 struct pool_workqueue *old_pwq;
5186 3603
5187 lockdep_assert_held(&wq_pool_mutex); 3604 lockdep_assert_held(&wq_pool_mutex);
5188 lockdep_assert_held(&wq->mutex); 3605 lockdep_assert_held(&wq->mutex);
5189 3606
5190 /* link_pwq() can handle duplicate ca 3607 /* link_pwq() can handle duplicate calls */
5191 link_pwq(pwq); 3608 link_pwq(pwq);
5192 3609
5193 old_pwq = rcu_access_pointer(*slot); !! 3610 old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
5194 rcu_assign_pointer(*slot, pwq); !! 3611 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
5195 return old_pwq; 3612 return old_pwq;
5196 } 3613 }
5197 3614
5198 /* context to store the prepared attrs & pwqs 3615 /* context to store the prepared attrs & pwqs before applying */
5199 struct apply_wqattrs_ctx { 3616 struct apply_wqattrs_ctx {
5200 struct workqueue_struct *wq; 3617 struct workqueue_struct *wq; /* target workqueue */
5201 struct workqueue_attrs *attrs; 3618 struct workqueue_attrs *attrs; /* attrs to apply */
5202 struct list_head list; 3619 struct list_head list; /* queued for batching commit */
5203 struct pool_workqueue *dfl_pwq; 3620 struct pool_workqueue *dfl_pwq;
5204 struct pool_workqueue *pwq_tbl[]; 3621 struct pool_workqueue *pwq_tbl[];
5205 }; 3622 };
5206 3623
5207 /* free the resources after success or abort 3624 /* free the resources after success or abort */
5208 static void apply_wqattrs_cleanup(struct appl 3625 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
5209 { 3626 {
5210 if (ctx) { 3627 if (ctx) {
5211 int cpu; !! 3628 int node;
5212 3629
5213 for_each_possible_cpu(cpu) !! 3630 for_each_node(node)
5214 put_pwq_unlocked(ctx- !! 3631 put_pwq_unlocked(ctx->pwq_tbl[node]);
5215 put_pwq_unlocked(ctx->dfl_pwq 3632 put_pwq_unlocked(ctx->dfl_pwq);
5216 3633
5217 free_workqueue_attrs(ctx->att 3634 free_workqueue_attrs(ctx->attrs);
5218 3635
5219 kfree(ctx); 3636 kfree(ctx);
5220 } 3637 }
5221 } 3638 }
5222 3639
5223 /* allocate the attrs and pwqs for later inst 3640 /* allocate the attrs and pwqs for later installation */
5224 static struct apply_wqattrs_ctx * 3641 static struct apply_wqattrs_ctx *
5225 apply_wqattrs_prepare(struct workqueue_struct 3642 apply_wqattrs_prepare(struct workqueue_struct *wq,
5226 const struct workqueue_ !! 3643 const struct workqueue_attrs *attrs)
5227 const cpumask_var_t unb <<
5228 { 3644 {
5229 struct apply_wqattrs_ctx *ctx; 3645 struct apply_wqattrs_ctx *ctx;
5230 struct workqueue_attrs *new_attrs; !! 3646 struct workqueue_attrs *new_attrs, *tmp_attrs;
5231 int cpu; !! 3647 int node;
5232 3648
5233 lockdep_assert_held(&wq_pool_mutex); 3649 lockdep_assert_held(&wq_pool_mutex);
5234 3650
5235 if (WARN_ON(attrs->affn_scope < 0 || !! 3651 ctx = kzalloc(sizeof(*ctx) + nr_node_ids * sizeof(ctx->pwq_tbl[0]),
5236 attrs->affn_scope >= WQ_A !! 3652 GFP_KERNEL);
5237 return ERR_PTR(-EINVAL); <<
5238 3653
5239 ctx = kzalloc(struct_size(ctx, pwq_tb !! 3654 new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
5240 !! 3655 tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL);
5241 new_attrs = alloc_workqueue_attrs(); !! 3656 if (!ctx || !new_attrs || !tmp_attrs)
5242 if (!ctx || !new_attrs) <<
5243 goto out_free; 3657 goto out_free;
5244 3658
5245 /* 3659 /*
>> 3660 * Calculate the attrs of the default pwq.
>> 3661 * If the user configured cpumask doesn't overlap with the
>> 3662 * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
>> 3663 */
>> 3664 copy_workqueue_attrs(new_attrs, attrs);
>> 3665 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask);
>> 3666 if (unlikely(cpumask_empty(new_attrs->cpumask)))
>> 3667 cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask);
>> 3668
>> 3669 /*
>> 3670 * We may create multiple pwqs with differing cpumasks. Make a
>> 3671 * copy of @new_attrs which will be modified and used to obtain
>> 3672 * pools.
>> 3673 */
>> 3674 copy_workqueue_attrs(tmp_attrs, new_attrs);
>> 3675
>> 3676 /*
5246 * If something goes wrong during CPU 3677 * If something goes wrong during CPU up/down, we'll fall back to
5247 * the default pwq covering whole @at 3678 * the default pwq covering whole @attrs->cpumask. Always create
5248 * it even if we don't use it immedia 3679 * it even if we don't use it immediately.
5249 */ 3680 */
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, 3681 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
5254 if (!ctx->dfl_pwq) 3682 if (!ctx->dfl_pwq)
5255 goto out_free; 3683 goto out_free;
5256 3684
5257 for_each_possible_cpu(cpu) { !! 3685 for_each_node(node) {
5258 if (new_attrs->ordered) { !! 3686 if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) {
5259 ctx->dfl_pwq->refcnt+ !! 3687 ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
5260 ctx->pwq_tbl[cpu] = c !! 3688 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 3689 goto out_free;
>> 3690 } else {
>> 3691 ctx->dfl_pwq->refcnt++;
>> 3692 ctx->pwq_tbl[node] = ctx->dfl_pwq;
5266 } 3693 }
5267 } 3694 }
5268 3695
5269 /* save the user configured attrs and 3696 /* save the user configured attrs and sanitize it. */
5270 copy_workqueue_attrs(new_attrs, attrs 3697 copy_workqueue_attrs(new_attrs, attrs);
5271 cpumask_and(new_attrs->cpumask, new_a 3698 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
5272 cpumask_copy(new_attrs->__pod_cpumask <<
5273 ctx->attrs = new_attrs; 3699 ctx->attrs = new_attrs;
5274 3700
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; 3701 ctx->wq = wq;
>> 3702 free_workqueue_attrs(tmp_attrs);
5285 return ctx; 3703 return ctx;
5286 3704
5287 out_free: 3705 out_free:
>> 3706 free_workqueue_attrs(tmp_attrs);
5288 free_workqueue_attrs(new_attrs); 3707 free_workqueue_attrs(new_attrs);
5289 apply_wqattrs_cleanup(ctx); 3708 apply_wqattrs_cleanup(ctx);
5290 return ERR_PTR(-ENOMEM); !! 3709 return NULL;
5291 } 3710 }
5292 3711
5293 /* set attrs and install prepared pwqs, @ctx 3712 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
5294 static void apply_wqattrs_commit(struct apply 3713 static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
5295 { 3714 {
5296 int cpu; !! 3715 int node;
5297 3716
5298 /* all pwqs have been created success 3717 /* all pwqs have been created successfully, let's install'em */
5299 mutex_lock(&ctx->wq->mutex); 3718 mutex_lock(&ctx->wq->mutex);
5300 3719
5301 copy_workqueue_attrs(ctx->wq->unbound 3720 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
5302 3721
5303 /* save the previous pwqs and install !! 3722 /* save the previous pwq and install the new one */
5304 for_each_possible_cpu(cpu) !! 3723 for_each_node(node)
5305 ctx->pwq_tbl[cpu] = install_u !! 3724 ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node,
5306 !! 3725 ctx->pwq_tbl[node]);
5307 ctx->dfl_pwq = install_unbound_pwq(ct !! 3726
5308 !! 3727 /* @dfl_pwq might not have been used, ensure it's linked */
5309 /* update node_nr_active->max */ !! 3728 link_pwq(ctx->dfl_pwq);
5310 wq_update_node_max_active(ctx->wq, -1 !! 3729 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 3730
5317 mutex_unlock(&ctx->wq->mutex); 3731 mutex_unlock(&ctx->wq->mutex);
5318 } 3732 }
5319 3733
>> 3734 static void apply_wqattrs_lock(void)
>> 3735 {
>> 3736 /* CPUs should stay stable across pwq creations and installations */
>> 3737 get_online_cpus();
>> 3738 mutex_lock(&wq_pool_mutex);
>> 3739 }
>> 3740
>> 3741 static void apply_wqattrs_unlock(void)
>> 3742 {
>> 3743 mutex_unlock(&wq_pool_mutex);
>> 3744 put_online_cpus();
>> 3745 }
>> 3746
5320 static int apply_workqueue_attrs_locked(struc 3747 static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
5321 const 3748 const struct workqueue_attrs *attrs)
5322 { 3749 {
5323 struct apply_wqattrs_ctx *ctx; 3750 struct apply_wqattrs_ctx *ctx;
5324 3751
5325 /* only unbound workqueues can change 3752 /* only unbound workqueues can change attributes */
5326 if (WARN_ON(!(wq->flags & WQ_UNBOUND) 3753 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
5327 return -EINVAL; 3754 return -EINVAL;
5328 3755
5329 ctx = apply_wqattrs_prepare(wq, attrs !! 3756 /* creating multiple pwqs breaks ordering guarantee */
5330 if (IS_ERR(ctx)) !! 3757 if (!list_empty(&wq->pwqs)) {
5331 return PTR_ERR(ctx); !! 3758 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
>> 3759 return -EINVAL;
>> 3760
>> 3761 wq->flags &= ~__WQ_ORDERED;
>> 3762 }
>> 3763
>> 3764 ctx = apply_wqattrs_prepare(wq, attrs);
>> 3765 if (!ctx)
>> 3766 return -ENOMEM;
5332 3767
5333 /* the ctx has been prepared successf 3768 /* the ctx has been prepared successfully, let's commit it */
5334 apply_wqattrs_commit(ctx); 3769 apply_wqattrs_commit(ctx);
5335 apply_wqattrs_cleanup(ctx); 3770 apply_wqattrs_cleanup(ctx);
5336 3771
5337 return 0; 3772 return 0;
5338 } 3773 }
5339 3774
5340 /** 3775 /**
5341 * apply_workqueue_attrs - apply new workqueu 3776 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
5342 * @wq: the target workqueue 3777 * @wq: the target workqueue
5343 * @attrs: the workqueue_attrs to apply, allo 3778 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
5344 * 3779 *
5345 * Apply @attrs to an unbound workqueue @wq. !! 3780 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
5346 * a separate pwq to each CPU pod with possib !! 3781 * machines, this function maps a separate pwq to each NUMA node with
5347 * work items are affine to the pod it was is !! 3782 * possibles CPUs in @attrs->cpumask so that work items are affine to the
5348 * in-flight work items finish. Note that a w !! 3783 * NUMA node it was issued on. Older pwqs are released as in-flight work
5349 * itself back-to-back will stay on its curre !! 3784 * items finish. Note that a work item which repeatedly requeues itself
>> 3785 * back-to-back will stay on its current pwq.
5350 * 3786 *
5351 * Performs GFP_KERNEL allocations. 3787 * Performs GFP_KERNEL allocations.
5352 * 3788 *
5353 * Return: 0 on success and -errno on failure 3789 * Return: 0 on success and -errno on failure.
5354 */ 3790 */
5355 int apply_workqueue_attrs(struct workqueue_st 3791 int apply_workqueue_attrs(struct workqueue_struct *wq,
5356 const struct workqu 3792 const struct workqueue_attrs *attrs)
5357 { 3793 {
5358 int ret; 3794 int ret;
5359 3795
5360 mutex_lock(&wq_pool_mutex); !! 3796 apply_wqattrs_lock();
5361 ret = apply_workqueue_attrs_locked(wq 3797 ret = apply_workqueue_attrs_locked(wq, attrs);
5362 mutex_unlock(&wq_pool_mutex); !! 3798 apply_wqattrs_unlock();
5363 3799
5364 return ret; 3800 return ret;
5365 } 3801 }
>> 3802 EXPORT_SYMBOL_GPL(apply_workqueue_attrs);
5366 3803
5367 /** 3804 /**
5368 * unbound_wq_update_pwq - update a pwq slot !! 3805 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
5369 * @wq: the target workqueue 3806 * @wq: the target workqueue
5370 * @cpu: the CPU to update the pwq slot for !! 3807 * @cpu: the CPU coming up or going down
>> 3808 * @online: whether @cpu is coming up or going down
5371 * 3809 *
5372 * This function is to be called from %CPU_DO 3810 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
5373 * %CPU_DOWN_FAILED. @cpu is in the same pod !! 3811 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
5374 * !! 3812 * @wq accordingly.
5375 * 3813 *
5376 * If pod affinity can't be adjusted due to m !! 3814 * If NUMA affinity can't be adjusted due to memory allocation failure, it
5377 * back to @wq->dfl_pwq which may not be opti !! 3815 * falls back to @wq->dfl_pwq which may not be optimal but is always
5378 * !! 3816 * correct.
5379 * Note that when the last allowed CPU of a p !! 3817 *
5380 * with a cpumask spanning multiple pods, the !! 3818 * Note that when the last allowed CPU of a NUMA node goes offline for a
5381 * executing the work items for the workqueue !! 3819 * workqueue with a cpumask spanning multiple nodes, the workers which were
5382 * may execute on any CPU. This is similar to !! 3820 * already executing the work items for the workqueue will lose their CPU
5383 * CPU_DOWN. If a workqueue user wants strict !! 3821 * affinity and may execute on any CPU. This is similar to how per-cpu
5384 * responsibility to flush the work item from !! 3822 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
>> 3823 * affinity, it's the user's responsibility to flush the work item from
>> 3824 * CPU_DOWN_PREPARE.
5385 */ 3825 */
5386 static void unbound_wq_update_pwq(struct work !! 3826 static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
>> 3827 bool online)
5387 { 3828 {
>> 3829 int node = cpu_to_node(cpu);
>> 3830 int cpu_off = online ? -1 : cpu;
5388 struct pool_workqueue *old_pwq = NULL 3831 struct pool_workqueue *old_pwq = NULL, *pwq;
5389 struct workqueue_attrs *target_attrs; 3832 struct workqueue_attrs *target_attrs;
>> 3833 cpumask_t *cpumask;
5390 3834
5391 lockdep_assert_held(&wq_pool_mutex); 3835 lockdep_assert_held(&wq_pool_mutex);
5392 3836
5393 if (!(wq->flags & WQ_UNBOUND) || wq-> !! 3837 if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) ||
>> 3838 wq->unbound_attrs->no_numa)
5394 return; 3839 return;
5395 3840
5396 /* 3841 /*
5397 * We don't wanna alloc/free wq_attrs 3842 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
5398 * Let's use a preallocated one. The 3843 * Let's use a preallocated one. The following buf is protected by
5399 * CPU hotplug exclusion. 3844 * CPU hotplug exclusion.
5400 */ 3845 */
5401 target_attrs = unbound_wq_update_pwq_ !! 3846 target_attrs = wq_update_unbound_numa_attrs_buf;
>> 3847 cpumask = target_attrs->cpumask;
5402 3848
5403 copy_workqueue_attrs(target_attrs, wq 3849 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
5404 wqattrs_actualize_cpumask(target_attr !! 3850 pwq = unbound_pwq_by_node(wq, node);
5405 3851
5406 /* nothing to do if the target cpumas !! 3852 /*
5407 wq_calc_pod_cpumask(target_attrs, cpu !! 3853 * Let's determine what needs to be done. If the target cpumask is
5408 if (wqattrs_equal(target_attrs, unbou !! 3854 * different from the default pwq's, we need to compare it to @pwq's
5409 return; !! 3855 * and create a new one if they don't match. If the target cpumask
>> 3856 * equals the default pwq's, the default pwq should be used.
>> 3857 */
>> 3858 if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) {
>> 3859 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
>> 3860 return;
>> 3861 } else {
>> 3862 goto use_dfl_pwq;
>> 3863 }
5410 3864
5411 /* create a new pwq */ 3865 /* create a new pwq */
5412 pwq = alloc_unbound_pwq(wq, target_at 3866 pwq = alloc_unbound_pwq(wq, target_attrs);
5413 if (!pwq) { 3867 if (!pwq) {
5414 pr_warn("workqueue: allocatio !! 3868 pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
5415 wq->name); 3869 wq->name);
5416 goto use_dfl_pwq; 3870 goto use_dfl_pwq;
5417 } 3871 }
5418 3872
5419 /* Install the new pwq. */ 3873 /* Install the new pwq. */
5420 mutex_lock(&wq->mutex); 3874 mutex_lock(&wq->mutex);
5421 old_pwq = install_unbound_pwq(wq, cpu !! 3875 old_pwq = numa_pwq_tbl_install(wq, node, pwq);
5422 goto out_unlock; 3876 goto out_unlock;
5423 3877
5424 use_dfl_pwq: 3878 use_dfl_pwq:
5425 mutex_lock(&wq->mutex); 3879 mutex_lock(&wq->mutex);
5426 pwq = unbound_pwq(wq, -1); !! 3880 spin_lock_irq(&wq->dfl_pwq->pool->lock);
5427 raw_spin_lock_irq(&pwq->pool->lock); !! 3881 get_pwq(wq->dfl_pwq);
5428 get_pwq(pwq); !! 3882 spin_unlock_irq(&wq->dfl_pwq->pool->lock);
5429 raw_spin_unlock_irq(&pwq->pool->lock) !! 3883 old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
5430 old_pwq = install_unbound_pwq(wq, cpu <<
5431 out_unlock: 3884 out_unlock:
5432 mutex_unlock(&wq->mutex); 3885 mutex_unlock(&wq->mutex);
5433 put_pwq_unlocked(old_pwq); 3886 put_pwq_unlocked(old_pwq);
5434 } 3887 }
5435 3888
5436 static int alloc_and_link_pwqs(struct workque 3889 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
5437 { 3890 {
5438 bool highpri = wq->flags & WQ_HIGHPRI 3891 bool highpri = wq->flags & WQ_HIGHPRI;
5439 int cpu, ret; 3892 int cpu, ret;
5440 3893
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)) { 3894 if (!(wq->flags & WQ_UNBOUND)) {
5448 struct worker_pool __percpu * !! 3895 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
5449 !! 3896 if (!wq->cpu_pwqs)
5450 if (wq->flags & WQ_BH) !! 3897 return -ENOMEM;
5451 pools = bh_worker_poo <<
5452 else <<
5453 pools = cpu_worker_po <<
5454 3898
5455 for_each_possible_cpu(cpu) { 3899 for_each_possible_cpu(cpu) {
5456 struct pool_workqueue !! 3900 struct pool_workqueue *pwq =
5457 struct worker_pool *p !! 3901 per_cpu_ptr(wq->cpu_pwqs, cpu);
5458 !! 3902 struct worker_pool *cpu_pools =
5459 pool = &(per_cpu_ptr( !! 3903 per_cpu(cpu_worker_pools, cpu);
5460 pwq_p = per_cpu_ptr(w <<
5461 3904
5462 *pwq_p = kmem_cache_a !! 3905 init_pwq(pwq, wq, &cpu_pools[highpri]);
5463 <<
5464 if (!*pwq_p) <<
5465 goto enomem; <<
5466 <<
5467 init_pwq(*pwq_p, wq, <<
5468 3906
5469 mutex_lock(&wq->mutex 3907 mutex_lock(&wq->mutex);
5470 link_pwq(*pwq_p); !! 3908 link_pwq(pwq);
5471 mutex_unlock(&wq->mut 3909 mutex_unlock(&wq->mutex);
5472 } 3910 }
5473 return 0; 3911 return 0;
5474 } !! 3912 } else if (wq->flags & __WQ_ORDERED) {
5475 !! 3913 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
5476 if (wq->flags & __WQ_ORDERED) { <<
5477 struct pool_workqueue *dfl_pw <<
5478 <<
5479 ret = apply_workqueue_attrs_l <<
5480 /* there should only be singl 3914 /* there should only be single pwq for ordering guarantee */
5481 dfl_pwq = rcu_access_pointer( !! 3915 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
5482 WARN(!ret && (wq->pwqs.next ! !! 3916 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
5483 wq->pwqs.prev ! <<
5484 "ordering guarantee brok 3917 "ordering guarantee broken for workqueue %s\n", wq->name);
>> 3918 return ret;
5485 } else { 3919 } else {
5486 ret = apply_workqueue_attrs_l !! 3920 return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
5487 } 3921 }
5488 <<
5489 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 } 3922 }
5504 3923
5505 static int wq_clamp_max_active(int max_active 3924 static int wq_clamp_max_active(int max_active, unsigned int flags,
5506 const char *na 3925 const char *name)
5507 { 3926 {
5508 if (max_active < 1 || max_active > WQ !! 3927 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
>> 3928
>> 3929 if (max_active < 1 || max_active > lim)
5509 pr_warn("workqueue: max_activ 3930 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
5510 max_active, name, 1, !! 3931 max_active, name, 1, lim);
5511 3932
5512 return clamp_val(max_active, 1, WQ_MA !! 3933 return clamp_val(max_active, 1, lim);
5513 } 3934 }
5514 3935
5515 /* 3936 /*
5516 * Workqueues which may be used during memory 3937 * Workqueues which may be used during memory reclaim should have a rescuer
5517 * to guarantee forward progress. 3938 * to guarantee forward progress.
5518 */ 3939 */
5519 static int init_rescuer(struct workqueue_stru 3940 static int init_rescuer(struct workqueue_struct *wq)
5520 { 3941 {
5521 struct worker *rescuer; 3942 struct worker *rescuer;
5522 char id_buf[WORKER_ID_LEN]; <<
5523 int ret; 3943 int ret;
5524 3944
5525 lockdep_assert_held(&wq_pool_mutex); <<
5526 <<
5527 if (!(wq->flags & WQ_MEM_RECLAIM)) 3945 if (!(wq->flags & WQ_MEM_RECLAIM))
5528 return 0; 3946 return 0;
5529 3947
5530 rescuer = alloc_worker(NUMA_NO_NODE); 3948 rescuer = alloc_worker(NUMA_NO_NODE);
5531 if (!rescuer) { !! 3949 if (!rescuer)
5532 pr_err("workqueue: Failed to <<
5533 wq->name); <<
5534 return -ENOMEM; 3950 return -ENOMEM;
5535 } <<
5536 3951
5537 rescuer->rescue_wq = wq; 3952 rescuer->rescue_wq = wq;
5538 format_worker_id(id_buf, sizeof(id_bu !! 3953 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", wq->name);
5539 !! 3954 ret = PTR_ERR_OR_ZERO(rescuer->task);
5540 rescuer->task = kthread_create(rescue !! 3955 if (ret) {
5541 if (IS_ERR(rescuer->task)) { <<
5542 ret = PTR_ERR(rescuer->task); <<
5543 pr_err("workqueue: Failed to <<
5544 wq->name, ERR_PTR(ret) <<
5545 kfree(rescuer); 3956 kfree(rescuer);
5546 return ret; 3957 return ret;
5547 } 3958 }
5548 3959
5549 wq->rescuer = rescuer; 3960 wq->rescuer = rescuer;
5550 if (wq->flags & WQ_UNBOUND) !! 3961 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); 3962 wake_up_process(rescuer->task);
5555 3963
5556 return 0; 3964 return 0;
5557 } 3965 }
5558 3966
5559 /** !! 3967 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
5560 * wq_adjust_max_active - update a wq's max_a !! 3968 unsigned int flags,
5561 * @wq: target workqueue !! 3969 int max_active,
5562 * !! 3970 struct lock_class_key *key,
5563 * If @wq isn't freezing, set @wq->max_active !! 3971 const char *lock_name, ...)
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) <<
5623 struct workqueue_struct *alloc_workqueue(cons <<
5624 unsi <<
5625 int <<
5626 { 3972 {
>> 3973 size_t tbl_size = 0;
5627 va_list args; 3974 va_list args;
5628 struct workqueue_struct *wq; 3975 struct workqueue_struct *wq;
5629 size_t wq_size; !! 3976 struct pool_workqueue *pwq;
5630 int name_len; <<
5631 3977
5632 if (flags & WQ_BH) { !! 3978 /*
5633 if (WARN_ON_ONCE(flags & ~__W !! 3979 * Unbound && max_active == 1 used to imply ordered, which is no
5634 return NULL; !! 3980 * longer the case on NUMA machines due to per-node pools. While
5635 if (WARN_ON_ONCE(max_active)) !! 3981 * alloc_ordered_workqueue() is the right way to create an ordered
5636 return NULL; !! 3982 * workqueue, keep the previous behavior to avoid subtle breakages
5637 } !! 3983 * on NUMA.
>> 3984 */
>> 3985 if ((flags & WQ_UNBOUND) && max_active == 1)
>> 3986 flags |= __WQ_ORDERED;
5638 3987
5639 /* see the comment above the definiti 3988 /* see the comment above the definition of WQ_POWER_EFFICIENT */
5640 if ((flags & WQ_POWER_EFFICIENT) && w 3989 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
5641 flags |= WQ_UNBOUND; 3990 flags |= WQ_UNBOUND;
5642 3991
5643 /* allocate wq and format name */ 3992 /* allocate wq and format name */
5644 if (flags & WQ_UNBOUND) 3993 if (flags & WQ_UNBOUND)
5645 wq_size = struct_size(wq, nod !! 3994 tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]);
5646 else <<
5647 wq_size = sizeof(*wq); <<
5648 3995
5649 wq = kzalloc(wq_size, GFP_KERNEL); !! 3996 wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
5650 if (!wq) 3997 if (!wq)
5651 return NULL; 3998 return NULL;
5652 3999
5653 if (flags & WQ_UNBOUND) { 4000 if (flags & WQ_UNBOUND) {
5654 wq->unbound_attrs = alloc_wor !! 4001 wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
5655 if (!wq->unbound_attrs) 4002 if (!wq->unbound_attrs)
5656 goto err_free_wq; 4003 goto err_free_wq;
5657 } 4004 }
5658 4005
5659 va_start(args, max_active); !! 4006 va_start(args, lock_name);
5660 name_len = vsnprintf(wq->name, sizeof !! 4007 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
5661 va_end(args); 4008 va_end(args);
5662 4009
5663 if (name_len >= WQ_NAME_LEN) !! 4010 max_active = max_active ?: WQ_DFL_ACTIVE;
5664 pr_warn_once("workqueue: name !! 4011 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 4012
5678 /* init wq */ 4013 /* init wq */
5679 wq->flags = flags; 4014 wq->flags = flags;
5680 wq->max_active = max_active; !! 4015 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); 4016 mutex_init(&wq->mutex);
5685 atomic_set(&wq->nr_pwqs_to_flush, 0); 4017 atomic_set(&wq->nr_pwqs_to_flush, 0);
5686 INIT_LIST_HEAD(&wq->pwqs); 4018 INIT_LIST_HEAD(&wq->pwqs);
5687 INIT_LIST_HEAD(&wq->flusher_queue); 4019 INIT_LIST_HEAD(&wq->flusher_queue);
5688 INIT_LIST_HEAD(&wq->flusher_overflow) 4020 INIT_LIST_HEAD(&wq->flusher_overflow);
5689 INIT_LIST_HEAD(&wq->maydays); 4021 INIT_LIST_HEAD(&wq->maydays);
5690 4022
5691 wq_init_lockdep(wq); !! 4023 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
5692 INIT_LIST_HEAD(&wq->list); 4024 INIT_LIST_HEAD(&wq->list);
5693 4025
5694 if (flags & WQ_UNBOUND) { !! 4026 if (alloc_and_link_pwqs(wq) < 0)
5695 if (alloc_node_nr_active(wq-> !! 4027 goto err_free_wq;
5696 goto err_unreg_lockde !! 4028
5697 } !! 4029 if (wq_online && init_rescuer(wq) < 0)
>> 4030 goto err_destroy;
>> 4031
>> 4032 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
>> 4033 goto err_destroy;
5698 4034
5699 /* 4035 /*
5700 * wq_pool_mutex protects the workque !! 4036 * wq_pool_mutex protects global freeze state and workqueues list.
5701 * and the global freeze state. !! 4037 * Grab it, adjust max_active and add the new @wq to workqueues
>> 4038 * list.
5702 */ 4039 */
5703 apply_wqattrs_lock(); !! 4040 mutex_lock(&wq_pool_mutex);
5704 <<
5705 if (alloc_and_link_pwqs(wq) < 0) <<
5706 goto err_unlock_free_node_nr_ <<
5707 4041
5708 mutex_lock(&wq->mutex); 4042 mutex_lock(&wq->mutex);
5709 wq_adjust_max_active(wq); !! 4043 for_each_pwq(pwq, wq)
>> 4044 pwq_adjust_max_active(pwq);
5710 mutex_unlock(&wq->mutex); 4045 mutex_unlock(&wq->mutex);
5711 4046
5712 list_add_tail_rcu(&wq->list, &workque 4047 list_add_tail_rcu(&wq->list, &workqueues);
5713 4048
5714 if (wq_online && init_rescuer(wq) < 0 !! 4049 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 4050
5722 return wq; 4051 return wq;
5723 4052
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: <<
5736 wq_unregister_lockdep(wq); <<
5737 wq_free_lockdep(wq); <<
5738 err_free_wq: 4053 err_free_wq:
5739 free_workqueue_attrs(wq->unbound_attr 4054 free_workqueue_attrs(wq->unbound_attrs);
5740 kfree(wq); 4055 kfree(wq);
5741 return NULL; 4056 return NULL;
5742 err_unlock_destroy: <<
5743 apply_wqattrs_unlock(); <<
5744 err_destroy: 4057 err_destroy:
5745 destroy_workqueue(wq); 4058 destroy_workqueue(wq);
5746 return NULL; 4059 return NULL;
5747 } 4060 }
5748 EXPORT_SYMBOL_GPL(alloc_workqueue); !! 4061 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
5749 <<
5750 static bool pwq_busy(struct pool_workqueue *p <<
5751 { <<
5752 int i; <<
5753 <<
5754 for (i = 0; i < WORK_NR_COLORS; i++) <<
5755 if (pwq->nr_in_flight[i]) <<
5756 return true; <<
5757 <<
5758 if ((pwq != rcu_access_pointer(pwq->w <<
5759 return true; <<
5760 if (!pwq_is_empty(pwq)) <<
5761 return true; <<
5762 <<
5763 return false; <<
5764 } <<
5765 4062
5766 /** 4063 /**
5767 * destroy_workqueue - safely terminate a wor 4064 * destroy_workqueue - safely terminate a workqueue
5768 * @wq: target workqueue 4065 * @wq: target workqueue
5769 * 4066 *
5770 * Safely destroy a workqueue. All work curre 4067 * Safely destroy a workqueue. All work currently pending will be done first.
5771 */ 4068 */
5772 void destroy_workqueue(struct workqueue_struc 4069 void destroy_workqueue(struct workqueue_struct *wq)
5773 { 4070 {
5774 struct pool_workqueue *pwq; 4071 struct pool_workqueue *pwq;
5775 int cpu; !! 4072 int node;
5776 <<
5777 /* <<
5778 * Remove it from sysfs first so that <<
5779 * lead to sysfs name conflicts. <<
5780 */ <<
5781 workqueue_sysfs_unregister(wq); <<
5782 <<
5783 /* mark the workqueue destruction is <<
5784 mutex_lock(&wq->mutex); <<
5785 wq->flags |= __WQ_DESTROYING; <<
5786 mutex_unlock(&wq->mutex); <<
5787 4073
5788 /* drain it before proceeding with de 4074 /* drain it before proceeding with destruction */
5789 drain_workqueue(wq); 4075 drain_workqueue(wq);
5790 4076
5791 /* kill rescuer, if sanity checks fai !! 4077 /* sanity checks */
5792 if (wq->rescuer) { <<
5793 struct worker *rescuer = wq-> <<
5794 <<
5795 /* this prevents new queueing <<
5796 raw_spin_lock_irq(&wq_mayday_ <<
5797 wq->rescuer = NULL; <<
5798 raw_spin_unlock_irq(&wq_mayda <<
5799 <<
5800 /* rescuer will empty maydays <<
5801 kthread_stop(rescuer->task); <<
5802 kfree(rescuer); <<
5803 } <<
5804 <<
5805 /* <<
5806 * Sanity checks - grab all the locks <<
5807 * in-flight operations which may do <<
5808 */ <<
5809 mutex_lock(&wq_pool_mutex); <<
5810 mutex_lock(&wq->mutex); 4078 mutex_lock(&wq->mutex);
5811 for_each_pwq(pwq, wq) { 4079 for_each_pwq(pwq, wq) {
5812 raw_spin_lock_irq(&pwq->pool- !! 4080 int i;
5813 if (WARN_ON(pwq_busy(pwq))) { !! 4081
5814 pr_warn("%s: %s has t !! 4082 for (i = 0; i < WORK_NR_COLORS; i++) {
5815 __func__, wq- !! 4083 if (WARN_ON(pwq->nr_in_flight[i])) {
5816 show_pwq(pwq); !! 4084 mutex_unlock(&wq->mutex);
5817 raw_spin_unlock_irq(& !! 4085 show_workqueue_state();
>> 4086 return;
>> 4087 }
>> 4088 }
>> 4089
>> 4090 if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) ||
>> 4091 WARN_ON(pwq->nr_active) ||
>> 4092 WARN_ON(!list_empty(&pwq->delayed_works))) {
5818 mutex_unlock(&wq->mut 4093 mutex_unlock(&wq->mutex);
5819 mutex_unlock(&wq_pool !! 4094 show_workqueue_state();
5820 show_one_workqueue(wq <<
5821 return; 4095 return;
5822 } 4096 }
5823 raw_spin_unlock_irq(&pwq->poo <<
5824 } 4097 }
5825 mutex_unlock(&wq->mutex); 4098 mutex_unlock(&wq->mutex);
5826 4099
5827 /* 4100 /*
5828 * wq list is used to freeze wq, remo 4101 * wq list is used to freeze wq, remove from list after
5829 * flushing is complete in case freez 4102 * flushing is complete in case freeze races us.
5830 */ 4103 */
>> 4104 mutex_lock(&wq_pool_mutex);
5831 list_del_rcu(&wq->list); 4105 list_del_rcu(&wq->list);
5832 mutex_unlock(&wq_pool_mutex); 4106 mutex_unlock(&wq_pool_mutex);
5833 4107
5834 /* !! 4108 workqueue_sysfs_unregister(wq);
5835 * We're the sole accessor of @wq. Di <<
5836 * to put the base refs. @wq will be <<
5837 * pwq_put. RCU read lock prevents @w <<
5838 */ <<
5839 rcu_read_lock(); <<
5840 4109
5841 for_each_possible_cpu(cpu) { !! 4110 if (wq->rescuer)
5842 put_pwq_unlocked(unbound_pwq( !! 4111 kthread_stop(wq->rescuer->task);
5843 RCU_INIT_POINTER(*unbound_pwq <<
5844 } <<
5845 4112
5846 put_pwq_unlocked(unbound_pwq(wq, -1)) !! 4113 if (!(wq->flags & WQ_UNBOUND)) {
5847 RCU_INIT_POINTER(*unbound_pwq_slot(wq !! 4114 /*
>> 4115 * The base ref is never dropped on per-cpu pwqs. Directly
>> 4116 * schedule RCU free.
>> 4117 */
>> 4118 call_rcu_sched(&wq->rcu, rcu_free_wq);
>> 4119 } else {
>> 4120 /*
>> 4121 * We're the sole accessor of @wq at this point. Directly
>> 4122 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
>> 4123 * @wq will be freed when the last pwq is released.
>> 4124 */
>> 4125 for_each_node(node) {
>> 4126 pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
>> 4127 RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
>> 4128 put_pwq_unlocked(pwq);
>> 4129 }
5848 4130
5849 rcu_read_unlock(); !! 4131 /*
>> 4132 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
>> 4133 * put. Don't access it afterwards.
>> 4134 */
>> 4135 pwq = wq->dfl_pwq;
>> 4136 wq->dfl_pwq = NULL;
>> 4137 put_pwq_unlocked(pwq);
>> 4138 }
5850 } 4139 }
5851 EXPORT_SYMBOL_GPL(destroy_workqueue); 4140 EXPORT_SYMBOL_GPL(destroy_workqueue);
5852 4141
5853 /** 4142 /**
5854 * workqueue_set_max_active - adjust max_acti 4143 * workqueue_set_max_active - adjust max_active of a workqueue
5855 * @wq: target workqueue 4144 * @wq: target workqueue
5856 * @max_active: new max_active value. 4145 * @max_active: new max_active value.
5857 * 4146 *
5858 * Set max_active of @wq to @max_active. See !! 4147 * Set max_active of @wq to @max_active.
5859 * comment. <<
5860 * 4148 *
5861 * CONTEXT: 4149 * CONTEXT:
5862 * Don't call from IRQ context. 4150 * Don't call from IRQ context.
5863 */ 4151 */
5864 void workqueue_set_max_active(struct workqueu 4152 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
5865 { 4153 {
5866 /* max_active doesn't mean anything f !! 4154 struct pool_workqueue *pwq;
5867 if (WARN_ON(wq->flags & WQ_BH)) !! 4155
5868 return; <<
5869 /* disallow meddling with max_active 4156 /* disallow meddling with max_active for ordered workqueues */
5870 if (WARN_ON(wq->flags & __WQ_ORDERED) !! 4157 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
5871 return; 4158 return;
5872 4159
5873 max_active = wq_clamp_max_active(max_ 4160 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
5874 4161
5875 mutex_lock(&wq->mutex); 4162 mutex_lock(&wq->mutex);
5876 4163
>> 4164 wq->flags &= ~__WQ_ORDERED;
5877 wq->saved_max_active = max_active; 4165 wq->saved_max_active = max_active;
5878 if (wq->flags & WQ_UNBOUND) <<
5879 wq->saved_min_active = min(wq <<
5880 4166
5881 wq_adjust_max_active(wq); !! 4167 for_each_pwq(pwq, wq)
>> 4168 pwq_adjust_max_active(pwq);
5882 4169
5883 mutex_unlock(&wq->mutex); 4170 mutex_unlock(&wq->mutex);
5884 } 4171 }
5885 EXPORT_SYMBOL_GPL(workqueue_set_max_active); 4172 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
5886 4173
5887 /** 4174 /**
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 4175 * current_work - retrieve %current task's work struct
5916 * 4176 *
5917 * Determine if %current task is a workqueue 4177 * Determine if %current task is a workqueue worker and what it's working on.
5918 * Useful to find out the context that the %c 4178 * Useful to find out the context that the %current task is running in.
5919 * 4179 *
5920 * Return: work struct if %current task is a 4180 * Return: work struct if %current task is a workqueue worker, %NULL otherwise.
5921 */ 4181 */
5922 struct work_struct *current_work(void) 4182 struct work_struct *current_work(void)
5923 { 4183 {
5924 struct worker *worker = current_wq_wo 4184 struct worker *worker = current_wq_worker();
5925 4185
5926 return worker ? worker->current_work 4186 return worker ? worker->current_work : NULL;
5927 } 4187 }
5928 EXPORT_SYMBOL(current_work); 4188 EXPORT_SYMBOL(current_work);
5929 4189
5930 /** 4190 /**
5931 * current_is_workqueue_rescuer - is %current 4191 * current_is_workqueue_rescuer - is %current workqueue rescuer?
5932 * 4192 *
5933 * Determine whether %current is a workqueue 4193 * Determine whether %current is a workqueue rescuer. Can be used from
5934 * work functions to determine whether it's b 4194 * work functions to determine whether it's being run off the rescuer task.
5935 * 4195 *
5936 * Return: %true if %current is a workqueue r 4196 * Return: %true if %current is a workqueue rescuer. %false otherwise.
5937 */ 4197 */
5938 bool current_is_workqueue_rescuer(void) 4198 bool current_is_workqueue_rescuer(void)
5939 { 4199 {
5940 struct worker *worker = current_wq_wo 4200 struct worker *worker = current_wq_worker();
5941 4201
5942 return worker && worker->rescue_wq; 4202 return worker && worker->rescue_wq;
5943 } 4203 }
5944 4204
5945 /** 4205 /**
5946 * workqueue_congested - test whether a workq 4206 * workqueue_congested - test whether a workqueue is congested
5947 * @cpu: CPU in question 4207 * @cpu: CPU in question
5948 * @wq: target workqueue 4208 * @wq: target workqueue
5949 * 4209 *
5950 * Test whether @wq's cpu workqueue for @cpu 4210 * Test whether @wq's cpu workqueue for @cpu is congested. There is
5951 * no synchronization around this function an 4211 * no synchronization around this function and the test result is
5952 * unreliable and only useful as advisory hin 4212 * unreliable and only useful as advisory hints or for debugging.
5953 * 4213 *
5954 * If @cpu is WORK_CPU_UNBOUND, the test is p 4214 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
5955 * !! 4215 * Note that both per-cpu and unbound workqueues may be associated with
5956 * With the exception of ordered workqueues, !! 4216 * multiple pool_workqueues which have separate congested states. A
5957 * pool_workqueues, each with its own congest !! 4217 * workqueue being congested on one CPU doesn't mean the workqueue is also
5958 * congested on one CPU doesn't mean that the !! 4218 * contested on other CPUs / NUMA nodes.
5959 * other CPUs. <<
5960 * 4219 *
5961 * Return: 4220 * Return:
5962 * %true if congested, %false otherwise. 4221 * %true if congested, %false otherwise.
5963 */ 4222 */
5964 bool workqueue_congested(int cpu, struct work 4223 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
5965 { 4224 {
5966 struct pool_workqueue *pwq; 4225 struct pool_workqueue *pwq;
5967 bool ret; 4226 bool ret;
5968 4227
5969 rcu_read_lock(); !! 4228 rcu_read_lock_sched();
5970 preempt_disable(); <<
5971 4229
5972 if (cpu == WORK_CPU_UNBOUND) 4230 if (cpu == WORK_CPU_UNBOUND)
5973 cpu = smp_processor_id(); 4231 cpu = smp_processor_id();
5974 4232
5975 pwq = *per_cpu_ptr(wq->cpu_pwq, cpu); !! 4233 if (!(wq->flags & WQ_UNBOUND))
5976 ret = !list_empty(&pwq->inactive_work !! 4234 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
>> 4235 else
>> 4236 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
5977 4237
5978 preempt_enable(); !! 4238 ret = !list_empty(&pwq->delayed_works);
5979 rcu_read_unlock(); !! 4239 rcu_read_unlock_sched();
5980 4240
5981 return ret; 4241 return ret;
5982 } 4242 }
5983 EXPORT_SYMBOL_GPL(workqueue_congested); 4243 EXPORT_SYMBOL_GPL(workqueue_congested);
5984 4244
5985 /** 4245 /**
5986 * work_busy - test whether a work is current 4246 * work_busy - test whether a work is currently pending or running
5987 * @work: the work to be tested 4247 * @work: the work to be tested
5988 * 4248 *
5989 * Test whether @work is currently pending or 4249 * Test whether @work is currently pending or running. There is no
5990 * synchronization around this function and t 4250 * synchronization around this function and the test result is
5991 * unreliable and only useful as advisory hin 4251 * unreliable and only useful as advisory hints or for debugging.
5992 * 4252 *
5993 * Return: 4253 * Return:
5994 * OR'd bitmask of WORK_BUSY_* bits. 4254 * OR'd bitmask of WORK_BUSY_* bits.
5995 */ 4255 */
5996 unsigned int work_busy(struct work_struct *wo 4256 unsigned int work_busy(struct work_struct *work)
5997 { 4257 {
5998 struct worker_pool *pool; 4258 struct worker_pool *pool;
5999 unsigned long irq_flags; !! 4259 unsigned long flags;
6000 unsigned int ret = 0; 4260 unsigned int ret = 0;
6001 4261
6002 if (work_pending(work)) 4262 if (work_pending(work))
6003 ret |= WORK_BUSY_PENDING; 4263 ret |= WORK_BUSY_PENDING;
6004 4264
6005 rcu_read_lock(); !! 4265 local_irq_save(flags);
6006 pool = get_work_pool(work); 4266 pool = get_work_pool(work);
6007 if (pool) { 4267 if (pool) {
6008 raw_spin_lock_irqsave(&pool-> !! 4268 spin_lock(&pool->lock);
6009 if (find_worker_executing_wor 4269 if (find_worker_executing_work(pool, work))
6010 ret |= WORK_BUSY_RUNN 4270 ret |= WORK_BUSY_RUNNING;
6011 raw_spin_unlock_irqrestore(&p !! 4271 spin_unlock(&pool->lock);
6012 } 4272 }
6013 rcu_read_unlock(); !! 4273 local_irq_restore(flags);
6014 4274
6015 return ret; 4275 return ret;
6016 } 4276 }
6017 EXPORT_SYMBOL_GPL(work_busy); 4277 EXPORT_SYMBOL_GPL(work_busy);
6018 4278
6019 /** 4279 /**
6020 * set_worker_desc - set description for the 4280 * set_worker_desc - set description for the current work item
6021 * @fmt: printf-style format string 4281 * @fmt: printf-style format string
6022 * @...: arguments for the format string 4282 * @...: arguments for the format string
6023 * 4283 *
6024 * This function can be called by a running w 4284 * This function can be called by a running work function to describe what
6025 * the work item is about. If the worker tas 4285 * the work item is about. If the worker task gets dumped, this
6026 * information will be printed out together t 4286 * information will be printed out together to help debugging. The
6027 * description can be at most WORKER_DESC_LEN 4287 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
6028 */ 4288 */
6029 void set_worker_desc(const char *fmt, ...) 4289 void set_worker_desc(const char *fmt, ...)
6030 { 4290 {
6031 struct worker *worker = current_wq_wo 4291 struct worker *worker = current_wq_worker();
6032 va_list args; 4292 va_list args;
6033 4293
6034 if (worker) { 4294 if (worker) {
6035 va_start(args, fmt); 4295 va_start(args, fmt);
6036 vsnprintf(worker->desc, sizeo 4296 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
6037 va_end(args); 4297 va_end(args);
>> 4298 worker->desc_valid = true;
6038 } 4299 }
6039 } 4300 }
6040 EXPORT_SYMBOL_GPL(set_worker_desc); <<
6041 4301
6042 /** 4302 /**
6043 * print_worker_info - print out worker infor 4303 * print_worker_info - print out worker information and description
6044 * @log_lvl: the log level to use when printi 4304 * @log_lvl: the log level to use when printing
6045 * @task: target task 4305 * @task: target task
6046 * 4306 *
6047 * If @task is a worker and currently executi 4307 * If @task is a worker and currently executing a work item, print out the
6048 * name of the workqueue being serviced and w 4308 * name of the workqueue being serviced and worker description set with
6049 * set_worker_desc() by the currently executi 4309 * set_worker_desc() by the currently executing work item.
6050 * 4310 *
6051 * This function can be safely called on any 4311 * This function can be safely called on any task as long as the
6052 * task_struct itself is accessible. While s 4312 * task_struct itself is accessible. While safe, this function isn't
6053 * synchronized and may print out mixups or g 4313 * synchronized and may print out mixups or garbages of limited length.
6054 */ 4314 */
6055 void print_worker_info(const char *log_lvl, s 4315 void print_worker_info(const char *log_lvl, struct task_struct *task)
6056 { 4316 {
6057 work_func_t *fn = NULL; 4317 work_func_t *fn = NULL;
6058 char name[WQ_NAME_LEN] = { }; 4318 char name[WQ_NAME_LEN] = { };
6059 char desc[WORKER_DESC_LEN] = { }; 4319 char desc[WORKER_DESC_LEN] = { };
6060 struct pool_workqueue *pwq = NULL; 4320 struct pool_workqueue *pwq = NULL;
6061 struct workqueue_struct *wq = NULL; 4321 struct workqueue_struct *wq = NULL;
>> 4322 bool desc_valid = false;
6062 struct worker *worker; 4323 struct worker *worker;
6063 4324
6064 if (!(task->flags & PF_WQ_WORKER)) 4325 if (!(task->flags & PF_WQ_WORKER))
6065 return; 4326 return;
6066 4327
6067 /* 4328 /*
6068 * This function is called without an 4329 * This function is called without any synchronization and @task
6069 * could be in any state. Be careful 4330 * could be in any state. Be careful with dereferences.
6070 */ 4331 */
6071 worker = kthread_probe_data(task); 4332 worker = kthread_probe_data(task);
6072 4333
6073 /* 4334 /*
6074 * Carefully copy the associated work !! 4335 * Carefully copy the associated workqueue's workfn and name. Keep
6075 * Keep the original last '\0' in cas !! 4336 * the original last '\0' in case the original contains garbage.
6076 */ 4337 */
6077 copy_from_kernel_nofault(&fn, &worker !! 4338 probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
6078 copy_from_kernel_nofault(&pwq, &worke !! 4339 probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
6079 copy_from_kernel_nofault(&wq, &pwq->w !! 4340 probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
6080 copy_from_kernel_nofault(name, wq->na !! 4341 probe_kernel_read(name, wq->name, sizeof(name) - 1);
6081 copy_from_kernel_nofault(desc, worker !! 4342
>> 4343 /* copy worker description */
>> 4344 probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid));
>> 4345 if (desc_valid)
>> 4346 probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
6082 4347
6083 if (fn || name[0] || desc[0]) { 4348 if (fn || name[0] || desc[0]) {
6084 printk("%sWorkqueue: %s %ps", !! 4349 printk("%sWorkqueue: %s %pf", log_lvl, name, fn);
6085 if (strcmp(name, desc)) !! 4350 if (desc[0])
6086 pr_cont(" (%s)", desc 4351 pr_cont(" (%s)", desc);
6087 pr_cont("\n"); 4352 pr_cont("\n");
6088 } 4353 }
6089 } 4354 }
6090 4355
6091 static void pr_cont_pool_info(struct worker_p 4356 static void pr_cont_pool_info(struct worker_pool *pool)
6092 { 4357 {
6093 pr_cont(" cpus=%*pbl", nr_cpumask_bit 4358 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
6094 if (pool->node != NUMA_NO_NODE) 4359 if (pool->node != NUMA_NO_NODE)
6095 pr_cont(" node=%d", pool->nod 4360 pr_cont(" node=%d", pool->node);
6096 pr_cont(" flags=0x%x", pool->flags); !! 4361 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 } 4362 }
6103 4363
6104 static void pr_cont_worker_id(struct worker * !! 4364 static void pr_cont_work(bool comma, struct work_struct *work)
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 } <<
6115 <<
6116 struct pr_cont_work_struct { <<
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 { 4365 {
6145 if (work->func == wq_barrier_func) { 4366 if (work->func == wq_barrier_func) {
6146 struct wq_barrier *barr; 4367 struct wq_barrier *barr;
6147 4368
6148 barr = container_of(work, str 4369 barr = container_of(work, struct wq_barrier, work);
6149 4370
6150 pr_cont_work_flush(comma, (wo <<
6151 pr_cont("%s BAR(%d)", comma ? 4371 pr_cont("%s BAR(%d)", comma ? "," : "",
6152 task_pid_nr(barr->tas 4372 task_pid_nr(barr->task));
6153 } else { 4373 } else {
6154 if (!comma) !! 4374 pr_cont("%s %pf", comma ? "," : "", work->func);
6155 pr_cont_work_flush(co <<
6156 pr_cont_work_flush(comma, wor <<
6157 } 4375 }
6158 } 4376 }
6159 4377
6160 static void show_pwq(struct pool_workqueue *p 4378 static void show_pwq(struct pool_workqueue *pwq)
6161 { 4379 {
6162 struct pr_cont_work_struct pcws = { . <<
6163 struct worker_pool *pool = pwq->pool; 4380 struct worker_pool *pool = pwq->pool;
6164 struct work_struct *work; 4381 struct work_struct *work;
6165 struct worker *worker; 4382 struct worker *worker;
6166 bool has_in_flight = false, has_pendi 4383 bool has_in_flight = false, has_pending = false;
6167 int bkt; 4384 int bkt;
6168 4385
6169 pr_info(" pwq %d:", pool->id); 4386 pr_info(" pwq %d:", pool->id);
6170 pr_cont_pool_info(pool); 4387 pr_cont_pool_info(pool);
6171 4388
6172 pr_cont(" active=%d refcnt=%d%s\n", !! 4389 pr_cont(" active=%d/%d%s\n", pwq->nr_active, pwq->max_active,
6173 pwq->nr_active, pwq->refcnt, <<
6174 !list_empty(&pwq->mayday_node 4390 !list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
6175 4391
6176 hash_for_each(pool->busy_hash, bkt, w 4392 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
6177 if (worker->current_pwq == pw 4393 if (worker->current_pwq == pwq) {
6178 has_in_flight = true; 4394 has_in_flight = true;
6179 break; 4395 break;
6180 } 4396 }
6181 } 4397 }
6182 if (has_in_flight) { 4398 if (has_in_flight) {
6183 bool comma = false; 4399 bool comma = false;
6184 4400
6185 pr_info(" in-flight:"); 4401 pr_info(" in-flight:");
6186 hash_for_each(pool->busy_hash 4402 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
6187 if (worker->current_p 4403 if (worker->current_pwq != pwq)
6188 continue; 4404 continue;
6189 4405
6190 pr_cont(" %s", comma !! 4406 pr_cont("%s %d%s:%pf", comma ? "," : "",
6191 pr_cont_worker_id(wor !! 4407 task_pid_nr(worker->task),
6192 pr_cont(":%ps", worke !! 4408 worker == pwq->wq->rescuer ? "(RESCUER)" : "",
>> 4409 worker->current_func);
6193 list_for_each_entry(w 4410 list_for_each_entry(work, &worker->scheduled, entry)
6194 pr_cont_work( !! 4411 pr_cont_work(false, work);
6195 pr_cont_work_flush(co <<
6196 comma = true; 4412 comma = true;
6197 } 4413 }
6198 pr_cont("\n"); 4414 pr_cont("\n");
6199 } 4415 }
6200 4416
6201 list_for_each_entry(work, &pool->work 4417 list_for_each_entry(work, &pool->worklist, entry) {
6202 if (get_work_pwq(work) == pwq 4418 if (get_work_pwq(work) == pwq) {
6203 has_pending = true; 4419 has_pending = true;
6204 break; 4420 break;
6205 } 4421 }
6206 } 4422 }
6207 if (has_pending) { 4423 if (has_pending) {
6208 bool comma = false; 4424 bool comma = false;
6209 4425
6210 pr_info(" pending:"); 4426 pr_info(" pending:");
6211 list_for_each_entry(work, &po 4427 list_for_each_entry(work, &pool->worklist, entry) {
6212 if (get_work_pwq(work 4428 if (get_work_pwq(work) != pwq)
6213 continue; 4429 continue;
6214 4430
6215 pr_cont_work(comma, w !! 4431 pr_cont_work(comma, work);
6216 comma = !(*work_data_ 4432 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
6217 } 4433 }
6218 pr_cont_work_flush(comma, (wo <<
6219 pr_cont("\n"); 4434 pr_cont("\n");
6220 } 4435 }
6221 4436
6222 if (!list_empty(&pwq->inactive_works) !! 4437 if (!list_empty(&pwq->delayed_works)) {
6223 bool comma = false; 4438 bool comma = false;
6224 4439
6225 pr_info(" inactive:"); !! 4440 pr_info(" delayed:");
6226 list_for_each_entry(work, &pw !! 4441 list_for_each_entry(work, &pwq->delayed_works, entry) {
6227 pr_cont_work(comma, w !! 4442 pr_cont_work(comma, work);
6228 comma = !(*work_data_ 4443 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
6229 } 4444 }
6230 pr_cont_work_flush(comma, (wo <<
6231 pr_cont("\n"); 4445 pr_cont("\n");
6232 } 4446 }
6233 } 4447 }
6234 4448
6235 /** 4449 /**
6236 * show_one_workqueue - dump state of specifi !! 4450 * 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 * 4451 *
6331 * Called from a sysrq handler and prints out !! 4452 * Called from a sysrq handler or try_to_freeze_tasks() and prints out
>> 4453 * all busy workqueues and pools.
6332 */ 4454 */
6333 void show_all_workqueues(void) !! 4455 void show_workqueue_state(void)
6334 { 4456 {
6335 struct workqueue_struct *wq; 4457 struct workqueue_struct *wq;
6336 struct worker_pool *pool; 4458 struct worker_pool *pool;
>> 4459 unsigned long flags;
6337 int pi; 4460 int pi;
6338 4461
6339 rcu_read_lock(); !! 4462 rcu_read_lock_sched();
6340 4463
6341 pr_info("Showing busy workqueues and 4464 pr_info("Showing busy workqueues and worker pools:\n");
6342 4465
6343 list_for_each_entry_rcu(wq, &workqueu <<
6344 show_one_workqueue(wq); <<
6345 <<
6346 for_each_pool(pool, pi) <<
6347 show_one_worker_pool(pool); <<
6348 <<
6349 rcu_read_unlock(); <<
6350 } <<
6351 <<
6352 /** <<
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 <<
6362 rcu_read_lock(); <<
6363 <<
6364 pr_info("Showing freezable workqueues <<
6365 <<
6366 list_for_each_entry_rcu(wq, &workqueu 4466 list_for_each_entry_rcu(wq, &workqueues, list) {
6367 if (!(wq->flags & WQ_FREEZABL !! 4467 struct pool_workqueue *pwq;
6368 continue; !! 4468 bool idle = true;
6369 show_one_workqueue(wq); <<
6370 } <<
6371 <<
6372 rcu_read_unlock(); <<
6373 } <<
6374 <<
6375 /* used to show worker information through /p <<
6376 void wq_worker_comm(char *buf, size_t size, s <<
6377 { <<
6378 /* stabilize PF_WQ_WORKER and worker <<
6379 mutex_lock(&wq_pool_attach_mutex); <<
6380 4469
6381 if (task->flags & PF_WQ_WORKER) { !! 4470 for_each_pwq(pwq, wq) {
6382 struct worker *worker = kthre !! 4471 if (pwq->nr_active || !list_empty(&pwq->delayed_works)) {
6383 struct worker_pool *pool = wo !! 4472 idle = false;
6384 int off; !! 4473 break;
>> 4474 }
>> 4475 }
>> 4476 if (idle)
>> 4477 continue;
6385 4478
6386 off = format_worker_id(buf, s !! 4479 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
6387 4480
6388 if (pool) { !! 4481 for_each_pwq(pwq, wq) {
6389 raw_spin_lock_irq(&po !! 4482 spin_lock_irqsave(&pwq->pool->lock, flags);
>> 4483 if (pwq->nr_active || !list_empty(&pwq->delayed_works))
>> 4484 show_pwq(pwq);
>> 4485 spin_unlock_irqrestore(&pwq->pool->lock, flags);
6390 /* 4486 /*
6391 * ->desc tracks info !! 4487 * We could be printing a lot from atomic context, e.g.
6392 * set_worker_desc()) !! 4488 * sysrq-t -> show_workqueue_state(). Avoid triggering
6393 * current, prepend ' !! 4489 * hard lockup.
6394 */ 4490 */
6395 if (worker->desc[0] ! !! 4491 touch_nmi_watchdog();
6396 if (worker->c <<
6397 scnpr <<
6398 <<
6399 else <<
6400 scnpr <<
6401 <<
6402 } <<
6403 raw_spin_unlock_irq(& <<
6404 } 4492 }
6405 } else { <<
6406 strscpy(buf, task->comm, size <<
6407 } 4493 }
6408 4494
6409 mutex_unlock(&wq_pool_attach_mutex); !! 4495 for_each_pool(pool, pi) {
6410 } !! 4496 struct worker *worker;
>> 4497 bool first = true;
6411 4498
6412 #ifdef CONFIG_SMP !! 4499 spin_lock_irqsave(&pool->lock, flags);
>> 4500 if (pool->nr_workers == pool->nr_idle)
>> 4501 goto next_pool;
>> 4502
>> 4503 pr_info("pool %d:", pool->id);
>> 4504 pr_cont_pool_info(pool);
>> 4505 pr_cont(" hung=%us workers=%d",
>> 4506 jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000,
>> 4507 pool->nr_workers);
>> 4508 if (pool->manager)
>> 4509 pr_cont(" manager: %d",
>> 4510 task_pid_nr(pool->manager->task));
>> 4511 list_for_each_entry(worker, &pool->idle_list, entry) {
>> 4512 pr_cont(" %s%d", first ? "idle: " : "",
>> 4513 task_pid_nr(worker->task));
>> 4514 first = false;
>> 4515 }
>> 4516 pr_cont("\n");
>> 4517 next_pool:
>> 4518 spin_unlock_irqrestore(&pool->lock, flags);
>> 4519 /*
>> 4520 * We could be printing a lot from atomic context, e.g.
>> 4521 * sysrq-t -> show_workqueue_state(). Avoid triggering
>> 4522 * hard lockup.
>> 4523 */
>> 4524 touch_nmi_watchdog();
>> 4525 }
>> 4526
>> 4527 rcu_read_unlock_sched();
>> 4528 }
6413 4529
6414 /* 4530 /*
6415 * CPU hotplug. 4531 * CPU hotplug.
6416 * 4532 *
6417 * There are two challenges in supporting CPU 4533 * There are two challenges in supporting CPU hotplug. Firstly, there
6418 * are a lot of assumptions on strong associa 4534 * are a lot of assumptions on strong associations among work, pwq and
6419 * pool which make migrating pending and sche 4535 * pool which make migrating pending and scheduled works very
6420 * difficult to implement without impacting h 4536 * difficult to implement without impacting hot paths. Secondly,
6421 * worker pools serve mix of short, long and 4537 * worker pools serve mix of short, long and very long running works making
6422 * blocked draining impractical. 4538 * blocked draining impractical.
6423 * 4539 *
6424 * This is solved by allowing the pools to be 4540 * This is solved by allowing the pools to be disassociated from the CPU
6425 * running as an unbound one and allowing it 4541 * running as an unbound one and allowing it to be reattached later if the
6426 * cpu comes back online. 4542 * cpu comes back online.
6427 */ 4543 */
6428 4544
6429 static void unbind_workers(int cpu) 4545 static void unbind_workers(int cpu)
6430 { 4546 {
6431 struct worker_pool *pool; 4547 struct worker_pool *pool;
6432 struct worker *worker; 4548 struct worker *worker;
6433 4549
6434 for_each_cpu_worker_pool(pool, cpu) { 4550 for_each_cpu_worker_pool(pool, cpu) {
6435 mutex_lock(&wq_pool_attach_mu !! 4551 mutex_lock(&pool->attach_mutex);
6436 raw_spin_lock_irq(&pool->lock !! 4552 spin_lock_irq(&pool->lock);
6437 4553
6438 /* 4554 /*
6439 * We've blocked all attach/d 4555 * We've blocked all attach/detach operations. Make all workers
6440 * unbound and set DISASSOCIA 4556 * unbound and set DISASSOCIATED. Before this, all workers
6441 * must be on the cpu. After !! 4557 * except for the ones which are still executing works from
6442 * And the preemption disable !! 4558 * before the last CPU down must be on the cpu. After
6443 * are guaranteed to see WORK !! 4559 * this, they may become diasporas.
6444 * is on the same cpu. <<
6445 */ 4560 */
6446 for_each_pool_worker(worker, 4561 for_each_pool_worker(worker, pool)
6447 worker->flags |= WORK 4562 worker->flags |= WORKER_UNBOUND;
6448 4563
6449 pool->flags |= POOL_DISASSOCI 4564 pool->flags |= POOL_DISASSOCIATED;
6450 4565
>> 4566 spin_unlock_irq(&pool->lock);
>> 4567 mutex_unlock(&pool->attach_mutex);
>> 4568
6451 /* 4569 /*
6452 * The handling of nr_running !! 4570 * Call schedule() so that we cross rq->lock and thus can
6453 * now. Zap nr_running. Aft !! 4571 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
6454 * need_more_worker() and kee !! 4572 * This is necessary as scheduler callbacks may be invoked
6455 * long as the worklist is no !! 4573 * from other cpus.
6456 * an unbound (in terms of co !! 4574 */
>> 4575 schedule();
>> 4576
>> 4577 /*
>> 4578 * Sched callbacks are disabled now. Zap nr_running.
>> 4579 * After this, nr_running stays zero and need_more_worker()
>> 4580 * and keep_working() are always true as long as the
>> 4581 * worklist is not empty. This pool now behaves as an
>> 4582 * unbound (in terms of concurrency management) pool which
6457 * are served by workers tied 4583 * are served by workers tied to the pool.
6458 */ 4584 */
6459 pool->nr_running = 0; !! 4585 atomic_set(&pool->nr_running, 0);
6460 4586
6461 /* 4587 /*
6462 * With concurrency managemen 4588 * With concurrency management just turned off, a busy
6463 * worker blocking could lead 4589 * worker blocking could lead to lengthy stalls. Kick off
6464 * unbound chain execution of 4590 * unbound chain execution of currently pending work items.
6465 */ 4591 */
6466 kick_pool(pool); !! 4592 spin_lock_irq(&pool->lock);
6467 !! 4593 wake_up_worker(pool);
6468 raw_spin_unlock_irq(&pool->lo !! 4594 spin_unlock_irq(&pool->lock);
6469 <<
6470 for_each_pool_worker(worker, <<
6471 unbind_worker(worker) <<
6472 <<
6473 mutex_unlock(&wq_pool_attach_ <<
6474 } 4595 }
6475 } 4596 }
6476 4597
6477 /** 4598 /**
6478 * rebind_workers - rebind all workers of a p 4599 * rebind_workers - rebind all workers of a pool to the associated CPU
6479 * @pool: pool of interest 4600 * @pool: pool of interest
6480 * 4601 *
6481 * @pool->cpu is coming online. Rebind all w 4602 * @pool->cpu is coming online. Rebind all workers to the CPU.
6482 */ 4603 */
6483 static void rebind_workers(struct worker_pool 4604 static void rebind_workers(struct worker_pool *pool)
6484 { 4605 {
6485 struct worker *worker; 4606 struct worker *worker;
6486 4607
6487 lockdep_assert_held(&wq_pool_attach_m !! 4608 lockdep_assert_held(&pool->attach_mutex);
6488 4609
6489 /* 4610 /*
6490 * Restore CPU affinity of all worker 4611 * Restore CPU affinity of all workers. As all idle workers should
6491 * be on the run-queue of the associa 4612 * be on the run-queue of the associated CPU before any local
6492 * wake-ups for concurrency managemen 4613 * wake-ups for concurrency management happen, restore CPU affinity
6493 * of all workers first and then clea 4614 * of all workers first and then clear UNBOUND. As we're called
6494 * from CPU_ONLINE, the following sho 4615 * from CPU_ONLINE, the following shouldn't fail.
6495 */ 4616 */
6496 for_each_pool_worker(worker, pool) { !! 4617 for_each_pool_worker(worker, pool)
6497 kthread_set_per_cpu(worker->t <<
6498 WARN_ON_ONCE(set_cpus_allowed 4618 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
6499 !! 4619 pool->attrs->cpumask) < 0);
6500 } <<
6501 4620
6502 raw_spin_lock_irq(&pool->lock); !! 4621 spin_lock_irq(&pool->lock);
6503 4622
6504 pool->flags &= ~POOL_DISASSOCIATED; 4623 pool->flags &= ~POOL_DISASSOCIATED;
6505 4624
6506 for_each_pool_worker(worker, pool) { 4625 for_each_pool_worker(worker, pool) {
6507 unsigned int worker_flags = w 4626 unsigned int worker_flags = worker->flags;
6508 4627
6509 /* 4628 /*
>> 4629 * A bound idle worker should actually be on the runqueue
>> 4630 * of the associated CPU for local wake-ups targeting it to
>> 4631 * work. Kick all idle workers so that they migrate to the
>> 4632 * associated CPU. Doing this in the same loop as
>> 4633 * replacing UNBOUND with REBOUND is safe as no worker will
>> 4634 * be bound before @pool->lock is released.
>> 4635 */
>> 4636 if (worker_flags & WORKER_IDLE)
>> 4637 wake_up_process(worker->task);
>> 4638
>> 4639 /*
6510 * We want to clear UNBOUND b 4640 * We want to clear UNBOUND but can't directly call
6511 * worker_clr_flags() or adju 4641 * worker_clr_flags() or adjust nr_running. Atomically
6512 * replace UNBOUND with anoth 4642 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
6513 * @worker will clear REBOUND 4643 * @worker will clear REBOUND using worker_clr_flags() when
6514 * it initiates the next exec 4644 * it initiates the next execution cycle thus restoring
6515 * concurrency management. N 4645 * concurrency management. Note that when or whether
6516 * @worker clears REBOUND doe 4646 * @worker clears REBOUND doesn't affect correctness.
6517 * 4647 *
6518 * WRITE_ONCE() is necessary 4648 * WRITE_ONCE() is necessary because @worker->flags may be
6519 * tested without holding any 4649 * tested without holding any lock in
6520 * wq_worker_running(). With !! 4650 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
6521 * fail incorrectly leading t 4651 * fail incorrectly leading to premature concurrency
6522 * management operations. 4652 * management operations.
6523 */ 4653 */
6524 WARN_ON_ONCE(!(worker_flags & 4654 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
6525 worker_flags |= WORKER_REBOUN 4655 worker_flags |= WORKER_REBOUND;
6526 worker_flags &= ~WORKER_UNBOU 4656 worker_flags &= ~WORKER_UNBOUND;
6527 WRITE_ONCE(worker->flags, wor 4657 WRITE_ONCE(worker->flags, worker_flags);
6528 } 4658 }
6529 4659
6530 raw_spin_unlock_irq(&pool->lock); !! 4660 spin_unlock_irq(&pool->lock);
6531 } 4661 }
6532 4662
6533 /** 4663 /**
6534 * restore_unbound_workers_cpumask - restore 4664 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
6535 * @pool: unbound pool of interest 4665 * @pool: unbound pool of interest
6536 * @cpu: the CPU which is coming up 4666 * @cpu: the CPU which is coming up
6537 * 4667 *
6538 * An unbound pool may end up with a cpumask 4668 * 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 4669 * CPUs. When a worker of such pool get scheduled, the scheduler resets
6540 * its cpus_allowed. If @cpu is in @pool's c 4670 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
6541 * online CPU before, cpus_allowed of all its 4671 * online CPU before, cpus_allowed of all its workers should be restored.
6542 */ 4672 */
6543 static void restore_unbound_workers_cpumask(s 4673 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
6544 { 4674 {
6545 static cpumask_t cpumask; 4675 static cpumask_t cpumask;
6546 struct worker *worker; 4676 struct worker *worker;
6547 4677
6548 lockdep_assert_held(&wq_pool_attach_m !! 4678 lockdep_assert_held(&pool->attach_mutex);
6549 4679
6550 /* is @cpu allowed for @pool? */ 4680 /* is @cpu allowed for @pool? */
6551 if (!cpumask_test_cpu(cpu, pool->attr 4681 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
6552 return; 4682 return;
6553 4683
6554 cpumask_and(&cpumask, pool->attrs->cp 4684 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
6555 4685
6556 /* as we're called from CPU_ONLINE, t 4686 /* as we're called from CPU_ONLINE, the following shouldn't fail */
6557 for_each_pool_worker(worker, pool) 4687 for_each_pool_worker(worker, pool)
6558 WARN_ON_ONCE(set_cpus_allowed 4688 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0);
6559 } 4689 }
6560 4690
6561 int workqueue_prepare_cpu(unsigned int cpu) 4691 int workqueue_prepare_cpu(unsigned int cpu)
6562 { 4692 {
6563 struct worker_pool *pool; 4693 struct worker_pool *pool;
6564 4694
6565 for_each_cpu_worker_pool(pool, cpu) { 4695 for_each_cpu_worker_pool(pool, cpu) {
6566 if (pool->nr_workers) 4696 if (pool->nr_workers)
6567 continue; 4697 continue;
6568 if (!create_worker(pool)) 4698 if (!create_worker(pool))
6569 return -ENOMEM; 4699 return -ENOMEM;
6570 } 4700 }
6571 return 0; 4701 return 0;
6572 } 4702 }
6573 4703
6574 int workqueue_online_cpu(unsigned int cpu) 4704 int workqueue_online_cpu(unsigned int cpu)
6575 { 4705 {
6576 struct worker_pool *pool; 4706 struct worker_pool *pool;
6577 struct workqueue_struct *wq; 4707 struct workqueue_struct *wq;
6578 int pi; 4708 int pi;
6579 4709
6580 mutex_lock(&wq_pool_mutex); 4710 mutex_lock(&wq_pool_mutex);
6581 4711
6582 cpumask_set_cpu(cpu, wq_online_cpumas <<
6583 <<
6584 for_each_pool(pool, pi) { 4712 for_each_pool(pool, pi) {
6585 /* BH pools aren't affected b !! 4713 mutex_lock(&pool->attach_mutex);
6586 if (pool->flags & POOL_BH) <<
6587 continue; <<
6588 4714
6589 mutex_lock(&wq_pool_attach_mu <<
6590 if (pool->cpu == cpu) 4715 if (pool->cpu == cpu)
6591 rebind_workers(pool); 4716 rebind_workers(pool);
6592 else if (pool->cpu < 0) 4717 else if (pool->cpu < 0)
6593 restore_unbound_worke 4718 restore_unbound_workers_cpumask(pool, cpu);
6594 mutex_unlock(&wq_pool_attach_ <<
6595 } <<
6596 <<
6597 /* update pod affinity of unbound wor <<
6598 list_for_each_entry(wq, &workqueues, <<
6599 struct workqueue_attrs *attrs <<
6600 <<
6601 if (attrs) { <<
6602 const struct wq_pod_t <<
6603 int tcpu; <<
6604 4719
6605 for_each_cpu(tcpu, pt !! 4720 mutex_unlock(&pool->attach_mutex);
6606 unbound_wq_up <<
6607 <<
6608 mutex_lock(&wq->mutex <<
6609 wq_update_node_max_ac <<
6610 mutex_unlock(&wq->mut <<
6611 } <<
6612 } 4721 }
6613 4722
>> 4723 /* update NUMA affinity of unbound workqueues */
>> 4724 list_for_each_entry(wq, &workqueues, list)
>> 4725 wq_update_unbound_numa(wq, cpu, true);
>> 4726
6614 mutex_unlock(&wq_pool_mutex); 4727 mutex_unlock(&wq_pool_mutex);
6615 return 0; 4728 return 0;
6616 } 4729 }
6617 4730
6618 int workqueue_offline_cpu(unsigned int cpu) 4731 int workqueue_offline_cpu(unsigned int cpu)
6619 { 4732 {
6620 struct workqueue_struct *wq; 4733 struct workqueue_struct *wq;
6621 4734
6622 /* unbinding per-cpu workers should h 4735 /* unbinding per-cpu workers should happen on the local CPU */
6623 if (WARN_ON(cpu != smp_processor_id() 4736 if (WARN_ON(cpu != smp_processor_id()))
6624 return -1; 4737 return -1;
6625 4738
6626 unbind_workers(cpu); 4739 unbind_workers(cpu);
6627 4740
6628 /* update pod affinity of unbound wor !! 4741 /* update NUMA affinity of unbound workqueues */
6629 mutex_lock(&wq_pool_mutex); 4742 mutex_lock(&wq_pool_mutex);
6630 !! 4743 list_for_each_entry(wq, &workqueues, list)
6631 cpumask_clear_cpu(cpu, wq_online_cpum !! 4744 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); 4745 mutex_unlock(&wq_pool_mutex);
6649 4746
6650 return 0; 4747 return 0;
6651 } 4748 }
6652 4749
>> 4750 #ifdef CONFIG_SMP
>> 4751
6653 struct work_for_cpu { 4752 struct work_for_cpu {
6654 struct work_struct work; 4753 struct work_struct work;
6655 long (*fn)(void *); 4754 long (*fn)(void *);
6656 void *arg; 4755 void *arg;
6657 long ret; 4756 long ret;
6658 }; 4757 };
6659 4758
6660 static void work_for_cpu_fn(struct work_struc 4759 static void work_for_cpu_fn(struct work_struct *work)
6661 { 4760 {
6662 struct work_for_cpu *wfc = container_ 4761 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
6663 4762
6664 wfc->ret = wfc->fn(wfc->arg); 4763 wfc->ret = wfc->fn(wfc->arg);
6665 } 4764 }
6666 4765
6667 /** 4766 /**
6668 * work_on_cpu_key - run a function in thread !! 4767 * work_on_cpu - run a function in thread context on a particular cpu
6669 * @cpu: the cpu to run on 4768 * @cpu: the cpu to run on
6670 * @fn: the function to run 4769 * @fn: the function to run
6671 * @arg: the function arg 4770 * @arg: the function arg
6672 * @key: The lock class key for lock debuggin <<
6673 * 4771 *
6674 * It is up to the caller to ensure that the 4772 * 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 4773 * The caller must not hold any locks which would prevent @fn from completing.
6676 * 4774 *
6677 * Return: The value @fn returns. 4775 * Return: The value @fn returns.
6678 */ 4776 */
6679 long work_on_cpu_key(int cpu, long (*fn)(void !! 4777 long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
6680 void *arg, struct lock_c <<
6681 { 4778 {
6682 struct work_for_cpu wfc = { .fn = fn, 4779 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
6683 4780
6684 INIT_WORK_ONSTACK_KEY(&wfc.work, work !! 4781 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
6685 schedule_work_on(cpu, &wfc.work); 4782 schedule_work_on(cpu, &wfc.work);
6686 flush_work(&wfc.work); 4783 flush_work(&wfc.work);
6687 destroy_work_on_stack(&wfc.work); 4784 destroy_work_on_stack(&wfc.work);
6688 return wfc.ret; 4785 return wfc.ret;
6689 } 4786 }
6690 EXPORT_SYMBOL_GPL(work_on_cpu_key); !! 4787 EXPORT_SYMBOL_GPL(work_on_cpu);
6691 4788
6692 /** 4789 /**
6693 * work_on_cpu_safe_key - run a function in t !! 4790 * work_on_cpu_safe - run a function in thread context on a particular cpu
6694 * @cpu: the cpu to run on 4791 * @cpu: the cpu to run on
6695 * @fn: the function to run 4792 * @fn: the function to run
6696 * @arg: the function argument 4793 * @arg: the function argument
6697 * @key: The lock class key for lock debuggin <<
6698 * 4794 *
6699 * Disables CPU hotplug and calls work_on_cpu 4795 * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold
6700 * any locks which would prevent @fn from com 4796 * any locks which would prevent @fn from completing.
6701 * 4797 *
6702 * Return: The value @fn returns. 4798 * Return: The value @fn returns.
6703 */ 4799 */
6704 long work_on_cpu_safe_key(int cpu, long (*fn) !! 4800 long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg)
6705 void *arg, struct l <<
6706 { 4801 {
6707 long ret = -ENODEV; 4802 long ret = -ENODEV;
6708 4803
6709 cpus_read_lock(); !! 4804 get_online_cpus();
6710 if (cpu_online(cpu)) 4805 if (cpu_online(cpu))
6711 ret = work_on_cpu_key(cpu, fn !! 4806 ret = work_on_cpu(cpu, fn, arg);
6712 cpus_read_unlock(); !! 4807 put_online_cpus();
6713 return ret; 4808 return ret;
6714 } 4809 }
6715 EXPORT_SYMBOL_GPL(work_on_cpu_safe_key); !! 4810 EXPORT_SYMBOL_GPL(work_on_cpu_safe);
6716 #endif /* CONFIG_SMP */ 4811 #endif /* CONFIG_SMP */
6717 4812
6718 #ifdef CONFIG_FREEZER 4813 #ifdef CONFIG_FREEZER
6719 4814
6720 /** 4815 /**
6721 * freeze_workqueues_begin - begin freezing w 4816 * freeze_workqueues_begin - begin freezing workqueues
6722 * 4817 *
6723 * Start freezing workqueues. After this fun 4818 * Start freezing workqueues. After this function returns, all freezable
6724 * workqueues will queue new works to their i !! 4819 * workqueues will queue new works to their delayed_works list instead of
6725 * pool->worklist. 4820 * pool->worklist.
6726 * 4821 *
6727 * CONTEXT: 4822 * CONTEXT:
6728 * Grabs and releases wq_pool_mutex, wq->mute 4823 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
6729 */ 4824 */
6730 void freeze_workqueues_begin(void) 4825 void freeze_workqueues_begin(void)
6731 { 4826 {
6732 struct workqueue_struct *wq; 4827 struct workqueue_struct *wq;
>> 4828 struct pool_workqueue *pwq;
6733 4829
6734 mutex_lock(&wq_pool_mutex); 4830 mutex_lock(&wq_pool_mutex);
6735 4831
6736 WARN_ON_ONCE(workqueue_freezing); 4832 WARN_ON_ONCE(workqueue_freezing);
6737 workqueue_freezing = true; 4833 workqueue_freezing = true;
6738 4834
6739 list_for_each_entry(wq, &workqueues, 4835 list_for_each_entry(wq, &workqueues, list) {
6740 mutex_lock(&wq->mutex); 4836 mutex_lock(&wq->mutex);
6741 wq_adjust_max_active(wq); !! 4837 for_each_pwq(pwq, wq)
>> 4838 pwq_adjust_max_active(pwq);
6742 mutex_unlock(&wq->mutex); 4839 mutex_unlock(&wq->mutex);
6743 } 4840 }
6744 4841
6745 mutex_unlock(&wq_pool_mutex); 4842 mutex_unlock(&wq_pool_mutex);
6746 } 4843 }
6747 4844
6748 /** 4845 /**
6749 * freeze_workqueues_busy - are freezable wor 4846 * freeze_workqueues_busy - are freezable workqueues still busy?
6750 * 4847 *
6751 * Check whether freezing is complete. This 4848 * Check whether freezing is complete. This function must be called
6752 * between freeze_workqueues_begin() and thaw 4849 * between freeze_workqueues_begin() and thaw_workqueues().
6753 * 4850 *
6754 * CONTEXT: 4851 * CONTEXT:
6755 * Grabs and releases wq_pool_mutex. 4852 * Grabs and releases wq_pool_mutex.
6756 * 4853 *
6757 * Return: 4854 * Return:
6758 * %true if some freezable workqueues are sti 4855 * %true if some freezable workqueues are still busy. %false if freezing
6759 * is complete. 4856 * is complete.
6760 */ 4857 */
6761 bool freeze_workqueues_busy(void) 4858 bool freeze_workqueues_busy(void)
6762 { 4859 {
6763 bool busy = false; 4860 bool busy = false;
6764 struct workqueue_struct *wq; 4861 struct workqueue_struct *wq;
6765 struct pool_workqueue *pwq; 4862 struct pool_workqueue *pwq;
6766 4863
6767 mutex_lock(&wq_pool_mutex); 4864 mutex_lock(&wq_pool_mutex);
6768 4865
6769 WARN_ON_ONCE(!workqueue_freezing); 4866 WARN_ON_ONCE(!workqueue_freezing);
6770 4867
6771 list_for_each_entry(wq, &workqueues, 4868 list_for_each_entry(wq, &workqueues, list) {
6772 if (!(wq->flags & WQ_FREEZABL 4869 if (!(wq->flags & WQ_FREEZABLE))
6773 continue; 4870 continue;
6774 /* 4871 /*
6775 * nr_active is monotonically 4872 * nr_active is monotonically decreasing. It's safe
6776 * to peek without lock. 4873 * to peek without lock.
6777 */ 4874 */
6778 rcu_read_lock(); !! 4875 rcu_read_lock_sched();
6779 for_each_pwq(pwq, wq) { 4876 for_each_pwq(pwq, wq) {
6780 WARN_ON_ONCE(pwq->nr_ 4877 WARN_ON_ONCE(pwq->nr_active < 0);
6781 if (pwq->nr_active) { 4878 if (pwq->nr_active) {
6782 busy = true; 4879 busy = true;
6783 rcu_read_unlo !! 4880 rcu_read_unlock_sched();
6784 goto out_unlo 4881 goto out_unlock;
6785 } 4882 }
6786 } 4883 }
6787 rcu_read_unlock(); !! 4884 rcu_read_unlock_sched();
6788 } 4885 }
6789 out_unlock: 4886 out_unlock:
6790 mutex_unlock(&wq_pool_mutex); 4887 mutex_unlock(&wq_pool_mutex);
6791 return busy; 4888 return busy;
6792 } 4889 }
6793 4890
6794 /** 4891 /**
6795 * thaw_workqueues - thaw workqueues 4892 * thaw_workqueues - thaw workqueues
6796 * 4893 *
6797 * Thaw workqueues. Normal queueing is resto 4894 * Thaw workqueues. Normal queueing is restored and all collected
6798 * frozen works are transferred to their resp 4895 * frozen works are transferred to their respective pool worklists.
6799 * 4896 *
6800 * CONTEXT: 4897 * CONTEXT:
6801 * Grabs and releases wq_pool_mutex, wq->mute 4898 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
6802 */ 4899 */
6803 void thaw_workqueues(void) 4900 void thaw_workqueues(void)
6804 { 4901 {
6805 struct workqueue_struct *wq; 4902 struct workqueue_struct *wq;
>> 4903 struct pool_workqueue *pwq;
6806 4904
6807 mutex_lock(&wq_pool_mutex); 4905 mutex_lock(&wq_pool_mutex);
6808 4906
6809 if (!workqueue_freezing) 4907 if (!workqueue_freezing)
6810 goto out_unlock; 4908 goto out_unlock;
6811 4909
6812 workqueue_freezing = false; 4910 workqueue_freezing = false;
6813 4911
6814 /* restore max_active and repopulate 4912 /* restore max_active and repopulate worklist */
6815 list_for_each_entry(wq, &workqueues, 4913 list_for_each_entry(wq, &workqueues, list) {
6816 mutex_lock(&wq->mutex); 4914 mutex_lock(&wq->mutex);
6817 wq_adjust_max_active(wq); !! 4915 for_each_pwq(pwq, wq)
>> 4916 pwq_adjust_max_active(pwq);
6818 mutex_unlock(&wq->mutex); 4917 mutex_unlock(&wq->mutex);
6819 } 4918 }
6820 4919
6821 out_unlock: 4920 out_unlock:
6822 mutex_unlock(&wq_pool_mutex); 4921 mutex_unlock(&wq_pool_mutex);
6823 } 4922 }
6824 #endif /* CONFIG_FREEZER */ 4923 #endif /* CONFIG_FREEZER */
6825 4924
6826 static int workqueue_apply_unbound_cpumask(co !! 4925 static int workqueue_apply_unbound_cpumask(void)
6827 { 4926 {
6828 LIST_HEAD(ctxs); 4927 LIST_HEAD(ctxs);
6829 int ret = 0; 4928 int ret = 0;
6830 struct workqueue_struct *wq; 4929 struct workqueue_struct *wq;
6831 struct apply_wqattrs_ctx *ctx, *n; 4930 struct apply_wqattrs_ctx *ctx, *n;
6832 4931
6833 lockdep_assert_held(&wq_pool_mutex); 4932 lockdep_assert_held(&wq_pool_mutex);
6834 4933
6835 list_for_each_entry(wq, &workqueues, 4934 list_for_each_entry(wq, &workqueues, list) {
6836 if (!(wq->flags & WQ_UNBOUND) !! 4935 if (!(wq->flags & WQ_UNBOUND))
>> 4936 continue;
>> 4937 /* creating multiple pwqs breaks ordering guarantee */
>> 4938 if (wq->flags & __WQ_ORDERED)
6837 continue; 4939 continue;
6838 4940
6839 ctx = apply_wqattrs_prepare(w !! 4941 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs);
6840 if (IS_ERR(ctx)) { !! 4942 if (!ctx) {
6841 ret = PTR_ERR(ctx); !! 4943 ret = -ENOMEM;
6842 break; 4944 break;
6843 } 4945 }
6844 4946
6845 list_add_tail(&ctx->list, &ct 4947 list_add_tail(&ctx->list, &ctxs);
6846 } 4948 }
6847 4949
6848 list_for_each_entry_safe(ctx, n, &ctx 4950 list_for_each_entry_safe(ctx, n, &ctxs, list) {
6849 if (!ret) 4951 if (!ret)
6850 apply_wqattrs_commit( 4952 apply_wqattrs_commit(ctx);
6851 apply_wqattrs_cleanup(ctx); 4953 apply_wqattrs_cleanup(ctx);
6852 } 4954 }
6853 4955
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; 4956 return ret;
6860 } 4957 }
6861 4958
6862 /** 4959 /**
6863 * workqueue_unbound_exclude_cpumask - Exclud !! 4960 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
6864 * @exclude_cpumask: the cpumask to be exclud !! 4961 * @cpumask: the cpumask to set
>> 4962 *
>> 4963 * The low-level workqueues cpumask is a global cpumask that limits
>> 4964 * the affinity of all unbound workqueues. This function check the @cpumask
>> 4965 * and apply it to all unbound workqueues and updates all pwqs of them.
6865 * 4966 *
6866 * This function can be called from cpuset co !! 4967 * Retun: 0 - Success
6867 * CPUs that should be excluded from wq_unbou !! 4968 * -EINVAL - Invalid @cpumask
>> 4969 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
6868 */ 4970 */
6869 int workqueue_unbound_exclude_cpumask(cpumask !! 4971 int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
6870 { 4972 {
6871 cpumask_var_t cpumask; !! 4973 int ret = -EINVAL;
6872 int ret = 0; !! 4974 cpumask_var_t saved_cpumask;
6873 4975
6874 if (!zalloc_cpumask_var(&cpumask, GFP !! 4976 if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL))
6875 return -ENOMEM; 4977 return -ENOMEM;
6876 4978
6877 mutex_lock(&wq_pool_mutex); <<
6878 <<
6879 /* 4979 /*
6880 * If the operation fails, it will fa !! 4980 * Not excluding isolated cpus on purpose.
6881 * wq_requested_unbound_cpumask which !! 4981 * If the user wishes to include them, we allow that.
6882 * (HK_TYPE_WQ ∩ HK_TYPE_DOMAIN) ho !! 4982 */
6883 * by any subsequent write to workque !! 4983 cpumask_and(cpumask, cpumask, cpu_possible_mask);
6884 */ !! 4984 if (!cpumask_empty(cpumask)) {
6885 if (!cpumask_andnot(cpumask, wq_reque !! 4985 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 4986
6921 cpus_read_lock(); !! 4987 /* save the old wq_unbound_cpumask. */
6922 mutex_lock(&wq_pool_mutex); !! 4988 cpumask_copy(saved_cpumask, wq_unbound_cpumask);
6923 4989
6924 wq_affn_dfl = affn; !! 4990 /* update wq_unbound_cpumask at first and apply it to wqs. */
>> 4991 cpumask_copy(wq_unbound_cpumask, cpumask);
>> 4992 ret = workqueue_apply_unbound_cpumask();
>> 4993
>> 4994 /* restore the wq_unbound_cpumask when failed. */
>> 4995 if (ret < 0)
>> 4996 cpumask_copy(wq_unbound_cpumask, saved_cpumask);
6925 4997
6926 list_for_each_entry(wq, &workqueues, !! 4998 apply_wqattrs_unlock();
6927 for_each_online_cpu(cpu) <<
6928 unbound_wq_update_pwq <<
6929 } 4999 }
6930 5000
6931 mutex_unlock(&wq_pool_mutex); !! 5001 free_cpumask_var(saved_cpumask);
6932 cpus_read_unlock(); !! 5002 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 } 5003 }
6941 5004
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 5005 #ifdef CONFIG_SYSFS
6950 /* 5006 /*
6951 * Workqueues with WQ_SYSFS flag set is visib 5007 * Workqueues with WQ_SYSFS flag set is visible to userland via
6952 * /sys/bus/workqueue/devices/WQ_NAME. All v 5008 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
6953 * following attributes. 5009 * following attributes.
6954 * 5010 *
6955 * per_cpu RO bool : whether the !! 5011 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
6956 * max_active RW int : maximum num !! 5012 * max_active RW int : maximum number of in-flight work items
6957 * 5013 *
6958 * Unbound workqueues have the following extr 5014 * Unbound workqueues have the following extra attributes.
6959 * 5015 *
6960 * nice RW int : nice value !! 5016 * pool_ids RO int : the associated pool IDs for each node
6961 * cpumask RW mask : bitmask of !! 5017 * nice RW int : nice value of the workers
6962 * affinity_scope RW str : worker CPU !! 5018 * cpumask RW mask : bitmask of allowed CPUs for the workers
6963 * affinity_strict RW bool : worker CPU !! 5019 * numa RW bool : whether enable NUMA affinity
6964 */ 5020 */
6965 struct wq_device { 5021 struct wq_device {
6966 struct workqueue_struct *wq; 5022 struct workqueue_struct *wq;
6967 struct device dev; 5023 struct device dev;
6968 }; 5024 };
6969 5025
6970 static struct workqueue_struct *dev_to_wq(str 5026 static struct workqueue_struct *dev_to_wq(struct device *dev)
6971 { 5027 {
6972 struct wq_device *wq_dev = container_ 5028 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
6973 5029
6974 return wq_dev->wq; 5030 return wq_dev->wq;
6975 } 5031 }
6976 5032
6977 static ssize_t per_cpu_show(struct device *de 5033 static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
6978 char *buf) 5034 char *buf)
6979 { 5035 {
6980 struct workqueue_struct *wq = dev_to_ 5036 struct workqueue_struct *wq = dev_to_wq(dev);
6981 5037
6982 return scnprintf(buf, PAGE_SIZE, "%d\ 5038 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
6983 } 5039 }
6984 static DEVICE_ATTR_RO(per_cpu); 5040 static DEVICE_ATTR_RO(per_cpu);
6985 5041
6986 static ssize_t max_active_show(struct device 5042 static ssize_t max_active_show(struct device *dev,
6987 struct device_ 5043 struct device_attribute *attr, char *buf)
6988 { 5044 {
6989 struct workqueue_struct *wq = dev_to_ 5045 struct workqueue_struct *wq = dev_to_wq(dev);
6990 5046
6991 return scnprintf(buf, PAGE_SIZE, "%d\ 5047 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
6992 } 5048 }
6993 5049
6994 static ssize_t max_active_store(struct device 5050 static ssize_t max_active_store(struct device *dev,
6995 struct device 5051 struct device_attribute *attr, const char *buf,
6996 size_t count) 5052 size_t count)
6997 { 5053 {
6998 struct workqueue_struct *wq = dev_to_ 5054 struct workqueue_struct *wq = dev_to_wq(dev);
6999 int val; 5055 int val;
7000 5056
7001 if (sscanf(buf, "%d", &val) != 1 || v 5057 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
7002 return -EINVAL; 5058 return -EINVAL;
7003 5059
7004 workqueue_set_max_active(wq, val); 5060 workqueue_set_max_active(wq, val);
7005 return count; 5061 return count;
7006 } 5062 }
7007 static DEVICE_ATTR_RW(max_active); 5063 static DEVICE_ATTR_RW(max_active);
7008 5064
7009 static struct attribute *wq_sysfs_attrs[] = { 5065 static struct attribute *wq_sysfs_attrs[] = {
7010 &dev_attr_per_cpu.attr, 5066 &dev_attr_per_cpu.attr,
7011 &dev_attr_max_active.attr, 5067 &dev_attr_max_active.attr,
7012 NULL, 5068 NULL,
7013 }; 5069 };
7014 ATTRIBUTE_GROUPS(wq_sysfs); 5070 ATTRIBUTE_GROUPS(wq_sysfs);
7015 5071
>> 5072 static ssize_t wq_pool_ids_show(struct device *dev,
>> 5073 struct device_attribute *attr, char *buf)
>> 5074 {
>> 5075 struct workqueue_struct *wq = dev_to_wq(dev);
>> 5076 const char *delim = "";
>> 5077 int node, written = 0;
>> 5078
>> 5079 rcu_read_lock_sched();
>> 5080 for_each_node(node) {
>> 5081 written += scnprintf(buf + written, PAGE_SIZE - written,
>> 5082 "%s%d:%d", delim, node,
>> 5083 unbound_pwq_by_node(wq, node)->pool->id);
>> 5084 delim = " ";
>> 5085 }
>> 5086 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
>> 5087 rcu_read_unlock_sched();
>> 5088
>> 5089 return written;
>> 5090 }
>> 5091
7016 static ssize_t wq_nice_show(struct device *de 5092 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
7017 char *buf) 5093 char *buf)
7018 { 5094 {
7019 struct workqueue_struct *wq = dev_to_ 5095 struct workqueue_struct *wq = dev_to_wq(dev);
7020 int written; 5096 int written;
7021 5097
7022 mutex_lock(&wq->mutex); 5098 mutex_lock(&wq->mutex);
7023 written = scnprintf(buf, PAGE_SIZE, " 5099 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
7024 mutex_unlock(&wq->mutex); 5100 mutex_unlock(&wq->mutex);
7025 5101
7026 return written; 5102 return written;
7027 } 5103 }
7028 5104
7029 /* prepare workqueue_attrs for sysfs store op 5105 /* prepare workqueue_attrs for sysfs store operations */
7030 static struct workqueue_attrs *wq_sysfs_prep_ 5106 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
7031 { 5107 {
7032 struct workqueue_attrs *attrs; 5108 struct workqueue_attrs *attrs;
7033 5109
7034 lockdep_assert_held(&wq_pool_mutex); 5110 lockdep_assert_held(&wq_pool_mutex);
7035 5111
7036 attrs = alloc_workqueue_attrs(); !! 5112 attrs = alloc_workqueue_attrs(GFP_KERNEL);
7037 if (!attrs) 5113 if (!attrs)
7038 return NULL; 5114 return NULL;
7039 5115
7040 copy_workqueue_attrs(attrs, wq->unbou 5116 copy_workqueue_attrs(attrs, wq->unbound_attrs);
7041 return attrs; 5117 return attrs;
7042 } 5118 }
7043 5119
7044 static ssize_t wq_nice_store(struct device *d 5120 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
7045 const char *buf, 5121 const char *buf, size_t count)
7046 { 5122 {
7047 struct workqueue_struct *wq = dev_to_ 5123 struct workqueue_struct *wq = dev_to_wq(dev);
7048 struct workqueue_attrs *attrs; 5124 struct workqueue_attrs *attrs;
7049 int ret = -ENOMEM; 5125 int ret = -ENOMEM;
7050 5126
7051 apply_wqattrs_lock(); 5127 apply_wqattrs_lock();
7052 5128
7053 attrs = wq_sysfs_prep_attrs(wq); 5129 attrs = wq_sysfs_prep_attrs(wq);
7054 if (!attrs) 5130 if (!attrs)
7055 goto out_unlock; 5131 goto out_unlock;
7056 5132
7057 if (sscanf(buf, "%d", &attrs->nice) = 5133 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
7058 attrs->nice >= MIN_NICE && attrs- 5134 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
7059 ret = apply_workqueue_attrs_l 5135 ret = apply_workqueue_attrs_locked(wq, attrs);
7060 else 5136 else
7061 ret = -EINVAL; 5137 ret = -EINVAL;
7062 5138
7063 out_unlock: 5139 out_unlock:
7064 apply_wqattrs_unlock(); 5140 apply_wqattrs_unlock();
7065 free_workqueue_attrs(attrs); 5141 free_workqueue_attrs(attrs);
7066 return ret ?: count; 5142 return ret ?: count;
7067 } 5143 }
7068 5144
7069 static ssize_t wq_cpumask_show(struct device 5145 static ssize_t wq_cpumask_show(struct device *dev,
7070 struct device_ 5146 struct device_attribute *attr, char *buf)
7071 { 5147 {
7072 struct workqueue_struct *wq = dev_to_ 5148 struct workqueue_struct *wq = dev_to_wq(dev);
7073 int written; 5149 int written;
7074 5150
7075 mutex_lock(&wq->mutex); 5151 mutex_lock(&wq->mutex);
7076 written = scnprintf(buf, PAGE_SIZE, " 5152 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
7077 cpumask_pr_args(w 5153 cpumask_pr_args(wq->unbound_attrs->cpumask));
7078 mutex_unlock(&wq->mutex); 5154 mutex_unlock(&wq->mutex);
7079 return written; 5155 return written;
7080 } 5156 }
7081 5157
7082 static ssize_t wq_cpumask_store(struct device 5158 static ssize_t wq_cpumask_store(struct device *dev,
7083 struct device 5159 struct device_attribute *attr,
7084 const char *b 5160 const char *buf, size_t count)
7085 { 5161 {
7086 struct workqueue_struct *wq = dev_to_ 5162 struct workqueue_struct *wq = dev_to_wq(dev);
7087 struct workqueue_attrs *attrs; 5163 struct workqueue_attrs *attrs;
7088 int ret = -ENOMEM; 5164 int ret = -ENOMEM;
7089 5165
7090 apply_wqattrs_lock(); 5166 apply_wqattrs_lock();
7091 5167
7092 attrs = wq_sysfs_prep_attrs(wq); 5168 attrs = wq_sysfs_prep_attrs(wq);
7093 if (!attrs) 5169 if (!attrs)
7094 goto out_unlock; 5170 goto out_unlock;
7095 5171
7096 ret = cpumask_parse(buf, attrs->cpuma 5172 ret = cpumask_parse(buf, attrs->cpumask);
7097 if (!ret) 5173 if (!ret)
7098 ret = apply_workqueue_attrs_l 5174 ret = apply_workqueue_attrs_locked(wq, attrs);
7099 5175
7100 out_unlock: 5176 out_unlock:
7101 apply_wqattrs_unlock(); 5177 apply_wqattrs_unlock();
7102 free_workqueue_attrs(attrs); 5178 free_workqueue_attrs(attrs);
7103 return ret ?: count; 5179 return ret ?: count;
7104 } 5180 }
7105 5181
7106 static ssize_t wq_affn_scope_show(struct devi !! 5182 static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
7107 struct devi !! 5183 char *buf)
7108 { 5184 {
7109 struct workqueue_struct *wq = dev_to_ 5185 struct workqueue_struct *wq = dev_to_wq(dev);
7110 int written; 5186 int written;
7111 5187
7112 mutex_lock(&wq->mutex); 5188 mutex_lock(&wq->mutex);
7113 if (wq->unbound_attrs->affn_scope == !! 5189 written = scnprintf(buf, PAGE_SIZE, "%d\n",
7114 written = scnprintf(buf, PAGE !! 5190 !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); 5191 mutex_unlock(&wq->mutex);
7121 5192
7122 return written; 5193 return written;
7123 } 5194 }
7124 5195
7125 static ssize_t wq_affn_scope_store(struct dev !! 5196 static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
7126 struct dev !! 5197 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 { 5198 {
7161 struct workqueue_struct *wq = dev_to_ 5199 struct workqueue_struct *wq = dev_to_wq(dev);
7162 struct workqueue_attrs *attrs; 5200 struct workqueue_attrs *attrs;
7163 int v, ret = -ENOMEM; 5201 int v, ret = -ENOMEM;
7164 5202
7165 if (sscanf(buf, "%d", &v) != 1) <<
7166 return -EINVAL; <<
7167 <<
7168 apply_wqattrs_lock(); 5203 apply_wqattrs_lock();
>> 5204
7169 attrs = wq_sysfs_prep_attrs(wq); 5205 attrs = wq_sysfs_prep_attrs(wq);
7170 if (attrs) { !! 5206 if (!attrs)
7171 attrs->affn_strict = (bool)v; !! 5207 goto out_unlock;
>> 5208
>> 5209 ret = -EINVAL;
>> 5210 if (sscanf(buf, "%d", &v) == 1) {
>> 5211 attrs->no_numa = !v;
7172 ret = apply_workqueue_attrs_l 5212 ret = apply_workqueue_attrs_locked(wq, attrs);
7173 } 5213 }
>> 5214
>> 5215 out_unlock:
7174 apply_wqattrs_unlock(); 5216 apply_wqattrs_unlock();
7175 free_workqueue_attrs(attrs); 5217 free_workqueue_attrs(attrs);
7176 return ret ?: count; 5218 return ret ?: count;
7177 } 5219 }
7178 5220
7179 static struct device_attribute wq_sysfs_unbou 5221 static struct device_attribute wq_sysfs_unbound_attrs[] = {
>> 5222 __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
7180 __ATTR(nice, 0644, wq_nice_show, wq_n 5223 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
7181 __ATTR(cpumask, 0644, wq_cpumask_show 5224 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
7182 __ATTR(affinity_scope, 0644, wq_affn_ !! 5225 __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
7183 __ATTR(affinity_strict, 0644, wq_affi <<
7184 __ATTR_NULL, 5226 __ATTR_NULL,
7185 }; 5227 };
7186 5228
7187 static const struct bus_type wq_subsys = { !! 5229 static struct bus_type wq_subsys = {
7188 .name = "wo 5230 .name = "workqueue",
7189 .dev_groups = wq_ 5231 .dev_groups = wq_sysfs_groups,
7190 }; 5232 };
7191 5233
7192 /** !! 5234 static ssize_t wq_unbound_cpumask_show(struct device *dev,
7193 * workqueue_set_unbound_cpumask - Set the l !! 5235 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 { 5236 {
7229 int written; 5237 int written;
7230 5238
7231 mutex_lock(&wq_pool_mutex); 5239 mutex_lock(&wq_pool_mutex);
7232 written = scnprintf(buf, PAGE_SIZE, " !! 5240 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
>> 5241 cpumask_pr_args(wq_unbound_cpumask));
7233 mutex_unlock(&wq_pool_mutex); 5242 mutex_unlock(&wq_pool_mutex);
7234 5243
7235 return written; 5244 return written;
7236 } 5245 }
7237 5246
7238 static ssize_t cpumask_requested_show(struct !! 5247 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 5248 struct device_attribute *attr, const char *buf, size_t count)
7260 { 5249 {
7261 cpumask_var_t cpumask; 5250 cpumask_var_t cpumask;
7262 int ret; 5251 int ret;
7263 5252
7264 if (!zalloc_cpumask_var(&cpumask, GFP 5253 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
7265 return -ENOMEM; 5254 return -ENOMEM;
7266 5255
7267 ret = cpumask_parse(buf, cpumask); 5256 ret = cpumask_parse(buf, cpumask);
7268 if (!ret) 5257 if (!ret)
7269 ret = workqueue_set_unbound_c 5258 ret = workqueue_set_unbound_cpumask(cpumask);
7270 5259
7271 free_cpumask_var(cpumask); 5260 free_cpumask_var(cpumask);
7272 return ret ? ret : count; 5261 return ret ? ret : count;
7273 } 5262 }
7274 static DEVICE_ATTR_RW(cpumask); <<
7275 5263
7276 static struct attribute *wq_sysfs_cpumask_att !! 5264 static struct device_attribute wq_sysfs_cpumask_attr =
7277 &dev_attr_cpumask.attr, !! 5265 __ATTR(cpumask, 0644, wq_unbound_cpumask_show,
7278 &dev_attr_cpumask_requested.attr, !! 5266 wq_unbound_cpumask_store);
7279 &dev_attr_cpumask_isolated.attr, <<
7280 NULL, <<
7281 }; <<
7282 ATTRIBUTE_GROUPS(wq_sysfs_cpumask); <<
7283 5267
7284 static int __init wq_sysfs_init(void) 5268 static int __init wq_sysfs_init(void)
7285 { 5269 {
7286 return subsys_virtual_register(&wq_su !! 5270 int err;
>> 5271
>> 5272 err = subsys_virtual_register(&wq_subsys, NULL);
>> 5273 if (err)
>> 5274 return err;
>> 5275
>> 5276 return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr);
7287 } 5277 }
7288 core_initcall(wq_sysfs_init); 5278 core_initcall(wq_sysfs_init);
7289 5279
7290 static void wq_device_release(struct device * 5280 static void wq_device_release(struct device *dev)
7291 { 5281 {
7292 struct wq_device *wq_dev = container_ 5282 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
7293 5283
7294 kfree(wq_dev); 5284 kfree(wq_dev);
7295 } 5285 }
7296 5286
7297 /** 5287 /**
7298 * workqueue_sysfs_register - make a workqueu 5288 * workqueue_sysfs_register - make a workqueue visible in sysfs
7299 * @wq: the workqueue to register 5289 * @wq: the workqueue to register
7300 * 5290 *
7301 * Expose @wq in sysfs under /sys/bus/workque 5291 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
7302 * alloc_workqueue*() automatically calls thi 5292 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
7303 * which is the preferred method. 5293 * which is the preferred method.
7304 * 5294 *
7305 * Workqueue user should use this function di 5295 * Workqueue user should use this function directly iff it wants to apply
7306 * workqueue_attrs before making the workqueu 5296 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
7307 * apply_workqueue_attrs() may race against u 5297 * apply_workqueue_attrs() may race against userland updating the
7308 * attributes. 5298 * attributes.
7309 * 5299 *
7310 * Return: 0 on success, -errno on failure. 5300 * Return: 0 on success, -errno on failure.
7311 */ 5301 */
7312 int workqueue_sysfs_register(struct workqueue 5302 int workqueue_sysfs_register(struct workqueue_struct *wq)
7313 { 5303 {
7314 struct wq_device *wq_dev; 5304 struct wq_device *wq_dev;
7315 int ret; 5305 int ret;
7316 5306
7317 /* 5307 /*
7318 * Adjusting max_active breaks orderi !! 5308 * Adjusting max_active or creating new pwqs by applying
7319 * ordered workqueues. !! 5309 * attributes breaks ordering guarantee. Disallow exposing ordered
>> 5310 * workqueues.
7320 */ 5311 */
7321 if (WARN_ON(wq->flags & __WQ_ORDERED) !! 5312 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
7322 return -EINVAL; 5313 return -EINVAL;
7323 5314
7324 wq->wq_dev = wq_dev = kzalloc(sizeof( 5315 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
7325 if (!wq_dev) 5316 if (!wq_dev)
7326 return -ENOMEM; 5317 return -ENOMEM;
7327 5318
7328 wq_dev->wq = wq; 5319 wq_dev->wq = wq;
7329 wq_dev->dev.bus = &wq_subsys; 5320 wq_dev->dev.bus = &wq_subsys;
7330 wq_dev->dev.release = wq_device_relea 5321 wq_dev->dev.release = wq_device_release;
7331 dev_set_name(&wq_dev->dev, "%s", wq-> 5322 dev_set_name(&wq_dev->dev, "%s", wq->name);
7332 5323
7333 /* 5324 /*
7334 * unbound_attrs are created separate 5325 * unbound_attrs are created separately. Suppress uevent until
7335 * everything is ready. 5326 * everything is ready.
7336 */ 5327 */
7337 dev_set_uevent_suppress(&wq_dev->dev, 5328 dev_set_uevent_suppress(&wq_dev->dev, true);
7338 5329
7339 ret = device_register(&wq_dev->dev); 5330 ret = device_register(&wq_dev->dev);
7340 if (ret) { 5331 if (ret) {
7341 put_device(&wq_dev->dev); 5332 put_device(&wq_dev->dev);
7342 wq->wq_dev = NULL; 5333 wq->wq_dev = NULL;
7343 return ret; 5334 return ret;
7344 } 5335 }
7345 5336
7346 if (wq->flags & WQ_UNBOUND) { 5337 if (wq->flags & WQ_UNBOUND) {
7347 struct device_attribute *attr 5338 struct device_attribute *attr;
7348 5339
7349 for (attr = wq_sysfs_unbound_ 5340 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
7350 ret = device_create_f 5341 ret = device_create_file(&wq_dev->dev, attr);
7351 if (ret) { 5342 if (ret) {
7352 device_unregi 5343 device_unregister(&wq_dev->dev);
7353 wq->wq_dev = 5344 wq->wq_dev = NULL;
7354 return ret; 5345 return ret;
7355 } 5346 }
7356 } 5347 }
7357 } 5348 }
7358 5349
7359 dev_set_uevent_suppress(&wq_dev->dev, 5350 dev_set_uevent_suppress(&wq_dev->dev, false);
7360 kobject_uevent(&wq_dev->dev.kobj, KOB 5351 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
7361 return 0; 5352 return 0;
7362 } 5353 }
7363 5354
7364 /** 5355 /**
7365 * workqueue_sysfs_unregister - undo workqueu 5356 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
7366 * @wq: the workqueue to unregister 5357 * @wq: the workqueue to unregister
7367 * 5358 *
7368 * If @wq is registered to sysfs by workqueue 5359 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
7369 */ 5360 */
7370 static void workqueue_sysfs_unregister(struct 5361 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
7371 { 5362 {
7372 struct wq_device *wq_dev = wq->wq_dev 5363 struct wq_device *wq_dev = wq->wq_dev;
7373 5364
7374 if (!wq->wq_dev) 5365 if (!wq->wq_dev)
7375 return; 5366 return;
7376 5367
7377 wq->wq_dev = NULL; 5368 wq->wq_dev = NULL;
7378 device_unregister(&wq_dev->dev); 5369 device_unregister(&wq_dev->dev);
7379 } 5370 }
7380 #else /* CONFIG_SYSFS */ 5371 #else /* CONFIG_SYSFS */
7381 static void workqueue_sysfs_unregister(struct 5372 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
7382 #endif /* CONFIG_SYSFS */ 5373 #endif /* CONFIG_SYSFS */
7383 5374
7384 /* 5375 /*
7385 * Workqueue watchdog. 5376 * Workqueue watchdog.
7386 * 5377 *
7387 * Stall may be caused by various bugs - miss 5378 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
7388 * flush dependency, a concurrency managed wo 5379 * flush dependency, a concurrency managed work item which stays RUNNING
7389 * indefinitely. Workqueue stalls can be ver 5380 * indefinitely. Workqueue stalls can be very difficult to debug as the
7390 * usual warning mechanisms don't trigger and 5381 * usual warning mechanisms don't trigger and internal workqueue state is
7391 * largely opaque. 5382 * largely opaque.
7392 * 5383 *
7393 * Workqueue watchdog monitors all worker poo 5384 * Workqueue watchdog monitors all worker pools periodically and dumps
7394 * state if some pools failed to make forward 5385 * state if some pools failed to make forward progress for a while where
7395 * forward progress is defined as the first i 5386 * forward progress is defined as the first item on ->worklist changing.
7396 * 5387 *
7397 * This mechanism is controlled through the k 5388 * This mechanism is controlled through the kernel parameter
7398 * "workqueue.watchdog_thresh" which can be u 5389 * "workqueue.watchdog_thresh" which can be updated at runtime through the
7399 * corresponding sysfs parameter file. 5390 * corresponding sysfs parameter file.
7400 */ 5391 */
7401 #ifdef CONFIG_WQ_WATCHDOG 5392 #ifdef CONFIG_WQ_WATCHDOG
7402 5393
7403 static unsigned long wq_watchdog_thresh = 30; 5394 static unsigned long wq_watchdog_thresh = 30;
7404 static struct timer_list wq_watchdog_timer; 5395 static struct timer_list wq_watchdog_timer;
7405 5396
7406 static unsigned long wq_watchdog_touched = IN 5397 static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
7407 static DEFINE_PER_CPU(unsigned long, wq_watch 5398 static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
7408 5399
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) 5400 static void wq_watchdog_reset_touched(void)
7461 { 5401 {
7462 int cpu; 5402 int cpu;
7463 5403
7464 wq_watchdog_touched = jiffies; 5404 wq_watchdog_touched = jiffies;
7465 for_each_possible_cpu(cpu) 5405 for_each_possible_cpu(cpu)
7466 per_cpu(wq_watchdog_touched_c 5406 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
7467 } 5407 }
7468 5408
7469 static void wq_watchdog_timer_fn(struct timer 5409 static void wq_watchdog_timer_fn(struct timer_list *unused)
7470 { 5410 {
7471 unsigned long thresh = READ_ONCE(wq_w 5411 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
7472 bool lockup_detected = false; 5412 bool lockup_detected = false;
7473 bool cpu_pool_stall = false; <<
7474 unsigned long now = jiffies; <<
7475 struct worker_pool *pool; 5413 struct worker_pool *pool;
7476 int pi; 5414 int pi;
7477 5415
7478 if (!thresh) 5416 if (!thresh)
7479 return; 5417 return;
7480 5418
7481 rcu_read_lock(); 5419 rcu_read_lock();
7482 5420
7483 for_each_pool(pool, pi) { 5421 for_each_pool(pool, pi) {
7484 unsigned long pool_ts, touche 5422 unsigned long pool_ts, touched, ts;
7485 5423
7486 pool->cpu_stall = false; <<
7487 if (list_empty(&pool->worklis 5424 if (list_empty(&pool->worklist))
7488 continue; 5425 continue;
7489 5426
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 5427 /* 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 5428 pool_ts = READ_ONCE(pool->watchdog_ts);
>> 5429 touched = READ_ONCE(wq_watchdog_touched);
7502 5430
7503 if (time_after(pool_ts, touch 5431 if (time_after(pool_ts, touched))
7504 ts = pool_ts; 5432 ts = pool_ts;
7505 else 5433 else
7506 ts = touched; 5434 ts = touched;
7507 5435
>> 5436 if (pool->cpu >= 0) {
>> 5437 unsigned long cpu_touched =
>> 5438 READ_ONCE(per_cpu(wq_watchdog_touched_cpu,
>> 5439 pool->cpu));
>> 5440 if (time_after(cpu_touched, ts))
>> 5441 ts = cpu_touched;
>> 5442 }
>> 5443
7508 /* did we stall? */ 5444 /* did we stall? */
7509 if (time_after(now, ts + thre !! 5445 if (time_after(jiffies, ts + thresh)) {
7510 lockup_detected = tru 5446 lockup_detected = true;
7511 if (pool->cpu >= 0 && <<
7512 pool->cpu_sta <<
7513 cpu_pool_stal <<
7514 } <<
7515 pr_emerg("BUG: workqu 5447 pr_emerg("BUG: workqueue lockup - pool");
7516 pr_cont_pool_info(poo 5448 pr_cont_pool_info(pool);
7517 pr_cont(" stuck for % 5449 pr_cont(" stuck for %us!\n",
7518 jiffies_to_ms !! 5450 jiffies_to_msecs(jiffies - pool_ts) / 1000);
7519 } 5451 }
7520 <<
7521 <<
7522 } 5452 }
7523 5453
7524 rcu_read_unlock(); 5454 rcu_read_unlock();
7525 5455
7526 if (lockup_detected) 5456 if (lockup_detected)
7527 show_all_workqueues(); !! 5457 show_workqueue_state();
7528 <<
7529 if (cpu_pool_stall) <<
7530 show_cpu_pools_hogs(); <<
7531 5458
7532 wq_watchdog_reset_touched(); 5459 wq_watchdog_reset_touched();
7533 mod_timer(&wq_watchdog_timer, jiffies 5460 mod_timer(&wq_watchdog_timer, jiffies + thresh);
7534 } 5461 }
7535 5462
7536 notrace void wq_watchdog_touch(int cpu) !! 5463 void wq_watchdog_touch(int cpu)
7537 { 5464 {
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) 5465 if (cpu >= 0)
7543 per_cpu(wq_watchdog_touched_c !! 5466 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
7544 else 5467 else
7545 WARN_ONCE(1, "%s should be ca !! 5468 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 } 5469 }
7551 5470
7552 static void wq_watchdog_set_thresh(unsigned l 5471 static void wq_watchdog_set_thresh(unsigned long thresh)
7553 { 5472 {
7554 wq_watchdog_thresh = 0; 5473 wq_watchdog_thresh = 0;
7555 del_timer_sync(&wq_watchdog_timer); 5474 del_timer_sync(&wq_watchdog_timer);
7556 5475
7557 if (thresh) { 5476 if (thresh) {
7558 wq_watchdog_thresh = thresh; 5477 wq_watchdog_thresh = thresh;
7559 wq_watchdog_reset_touched(); 5478 wq_watchdog_reset_touched();
7560 mod_timer(&wq_watchdog_timer, 5479 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
7561 } 5480 }
7562 } 5481 }
7563 5482
7564 static int wq_watchdog_param_set_thresh(const 5483 static int wq_watchdog_param_set_thresh(const char *val,
7565 const 5484 const struct kernel_param *kp)
7566 { 5485 {
7567 unsigned long thresh; 5486 unsigned long thresh;
7568 int ret; 5487 int ret;
7569 5488
7570 ret = kstrtoul(val, 0, &thresh); 5489 ret = kstrtoul(val, 0, &thresh);
7571 if (ret) 5490 if (ret)
7572 return ret; 5491 return ret;
7573 5492
7574 if (system_wq) 5493 if (system_wq)
7575 wq_watchdog_set_thresh(thresh 5494 wq_watchdog_set_thresh(thresh);
7576 else 5495 else
7577 wq_watchdog_thresh = thresh; 5496 wq_watchdog_thresh = thresh;
7578 5497
7579 return 0; 5498 return 0;
7580 } 5499 }
7581 5500
7582 static const struct kernel_param_ops wq_watch 5501 static const struct kernel_param_ops wq_watchdog_thresh_ops = {
7583 .set = wq_watchdog_param_set_thres 5502 .set = wq_watchdog_param_set_thresh,
7584 .get = param_get_ulong, 5503 .get = param_get_ulong,
7585 }; 5504 };
7586 5505
7587 module_param_cb(watchdog_thresh, &wq_watchdog 5506 module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
7588 0644); 5507 0644);
7589 5508
7590 static void wq_watchdog_init(void) 5509 static void wq_watchdog_init(void)
7591 { 5510 {
7592 timer_setup(&wq_watchdog_timer, wq_wa 5511 timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE);
7593 wq_watchdog_set_thresh(wq_watchdog_th 5512 wq_watchdog_set_thresh(wq_watchdog_thresh);
7594 } 5513 }
7595 5514
7596 #else /* CONFIG_WQ_WATCHDOG */ 5515 #else /* CONFIG_WQ_WATCHDOG */
7597 5516
7598 static inline void wq_watchdog_init(void) { } 5517 static inline void wq_watchdog_init(void) { }
7599 5518
7600 #endif /* CONFIG_WQ_WATCHDOG */ 5519 #endif /* CONFIG_WQ_WATCHDOG */
7601 5520
7602 static void bh_pool_kick_normal(struct irq_wo !! 5521 static void __init wq_numa_init(void)
7603 { 5522 {
7604 raise_softirq_irqoff(TASKLET_SOFTIRQ) !! 5523 cpumask_var_t *tbl;
7605 } !! 5524 int node, cpu;
7606 5525
7607 static void bh_pool_kick_highpri(struct irq_w !! 5526 if (num_possible_nodes() <= 1)
7608 { !! 5527 return;
7609 raise_softirq_irqoff(HI_SOFTIRQ); <<
7610 } <<
7611 5528
7612 static void __init restrict_unbound_cpumask(c !! 5529 if (wq_disable_numa) {
7613 { !! 5530 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; 5531 return;
7618 } 5532 }
7619 5533
7620 cpumask_and(wq_unbound_cpumask, wq_un !! 5534 wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL);
7621 } !! 5535 BUG_ON(!wq_update_unbound_numa_attrs_buf);
7622 5536
7623 static void __init init_cpu_worker_pool(struc !! 5537 /*
7624 { !! 5538 * We want masks of possible CPUs of each node which isn't readily
7625 BUG_ON(init_worker_pool(pool)); !! 5539 * available. Build one from cpu_to_node() which should have been
7626 pool->cpu = cpu; !! 5540 * fully initialized by now.
7627 cpumask_copy(pool->attrs->cpumask, cp !! 5541 */
7628 cpumask_copy(pool->attrs->__pod_cpuma !! 5542 tbl = kzalloc(nr_node_ids * sizeof(tbl[0]), GFP_KERNEL);
7629 pool->attrs->nice = nice; !! 5543 BUG_ON(!tbl);
7630 pool->attrs->affn_strict = true; <<
7631 pool->node = cpu_to_node(cpu); <<
7632 5544
7633 /* alloc pool ID */ !! 5545 for_each_node(node)
7634 mutex_lock(&wq_pool_mutex); !! 5546 BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
7635 BUG_ON(worker_pool_assign_id(pool)); !! 5547 node_online(node) ? node : NUMA_NO_NODE));
7636 mutex_unlock(&wq_pool_mutex); !! 5548
>> 5549 for_each_possible_cpu(cpu) {
>> 5550 node = cpu_to_node(cpu);
>> 5551 if (WARN_ON(node == NUMA_NO_NODE)) {
>> 5552 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
>> 5553 /* happens iff arch is bonkers, let's just proceed */
>> 5554 return;
>> 5555 }
>> 5556 cpumask_set_cpu(cpu, tbl[node]);
>> 5557 }
>> 5558
>> 5559 wq_numa_possible_cpumask = tbl;
>> 5560 wq_numa_enabled = true;
7637 } 5561 }
7638 5562
7639 /** 5563 /**
7640 * workqueue_init_early - early init for work 5564 * workqueue_init_early - early init for workqueue subsystem
7641 * 5565 *
7642 * This is the first step of three-staged wor !! 5566 * This is the first half of two-staged workqueue subsystem initialization
7643 * invoked as soon as the bare basics - memor !! 5567 * and invoked as soon as the bare basics - memory allocation, cpumasks and
7644 * up. It sets up all the data structures and !! 5568 * idr are up. It sets up all the data structures and system workqueues
7645 * boot code to create workqueues and queue/c !! 5569 * and allows early boot code to create workqueues and queue/cancel work
7646 * execution starts only after kthreads can b !! 5570 * items. Actual work item execution starts only after kthreads can be
7647 * before early initcalls. !! 5571 * created and scheduled right before early initcalls.
7648 */ 5572 */
7649 void __init workqueue_init_early(void) !! 5573 int __init workqueue_init_early(void)
7650 { 5574 {
7651 struct wq_pod_type *pt = &wq_pod_type <<
7652 int std_nice[NR_STD_WORKER_POOLS] = { 5575 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
7653 void (*irq_work_fns[2])(struct irq_wo <<
7654 <<
7655 int i, cpu; 5576 int i, cpu;
7656 5577
7657 BUILD_BUG_ON(__alignof__(struct pool_ !! 5578 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
7658 5579
7659 BUG_ON(!alloc_cpumask_var(&wq_online_ <<
7660 BUG_ON(!alloc_cpumask_var(&wq_unbound 5580 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
7661 BUG_ON(!alloc_cpumask_var(&wq_request !! 5581 cpumask_copy(wq_unbound_cpumask, housekeeping_cpumask(HK_FLAG_DOMAIN));
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 5582
7673 pwq_cache = KMEM_CACHE(pool_workqueue 5583 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
7674 5584
7675 unbound_wq_update_pwq_attrs_buf = all !! 5585 /* 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) { 5586 for_each_possible_cpu(cpu) {
7700 struct worker_pool *pool; 5587 struct worker_pool *pool;
7701 5588
7702 i = 0; 5589 i = 0;
7703 for_each_bh_worker_pool(pool, !! 5590 for_each_cpu_worker_pool(pool, cpu) {
7704 init_cpu_worker_pool( !! 5591 BUG_ON(init_worker_pool(pool));
7705 pool->flags |= POOL_B !! 5592 pool->cpu = cpu;
7706 init_irq_work(bh_pool !! 5593 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
7707 i++; !! 5594 pool->attrs->nice = std_nice[i++];
7708 } !! 5595 pool->node = cpu_to_node(cpu);
7709 5596
7710 i = 0; !! 5597 /* alloc pool ID */
7711 for_each_cpu_worker_pool(pool !! 5598 mutex_lock(&wq_pool_mutex);
7712 init_cpu_worker_pool( !! 5599 BUG_ON(worker_pool_assign_id(pool));
>> 5600 mutex_unlock(&wq_pool_mutex);
>> 5601 }
7713 } 5602 }
7714 5603
7715 /* create default unbound and ordered 5604 /* create default unbound and ordered wq attrs */
7716 for (i = 0; i < NR_STD_WORKER_POOLS; 5605 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
7717 struct workqueue_attrs *attrs 5606 struct workqueue_attrs *attrs;
7718 5607
7719 BUG_ON(!(attrs = alloc_workqu !! 5608 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
7720 attrs->nice = std_nice[i]; 5609 attrs->nice = std_nice[i];
7721 unbound_std_wq_attrs[i] = att 5610 unbound_std_wq_attrs[i] = attrs;
7722 5611
7723 /* 5612 /*
7724 * An ordered wq should have 5613 * An ordered wq should have only one pwq as ordering is
7725 * guaranteed by max_active w 5614 * guaranteed by max_active which is enforced by pwqs.
>> 5615 * Turn off NUMA so that dfl_pwq is used for all nodes.
7726 */ 5616 */
7727 BUG_ON(!(attrs = alloc_workqu !! 5617 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
7728 attrs->nice = std_nice[i]; 5618 attrs->nice = std_nice[i];
7729 attrs->ordered = true; !! 5619 attrs->no_numa = true;
7730 ordered_wq_attrs[i] = attrs; 5620 ordered_wq_attrs[i] = attrs;
7731 } 5621 }
7732 5622
7733 system_wq = alloc_workqueue("events", 5623 system_wq = alloc_workqueue("events", 0, 0);
7734 system_highpri_wq = alloc_workqueue(" 5624 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
7735 system_long_wq = alloc_workqueue("eve 5625 system_long_wq = alloc_workqueue("events_long", 0, 0);
7736 system_unbound_wq = alloc_workqueue(" 5626 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
7737 W !! 5627 WQ_UNBOUND_MAX_ACTIVE);
7738 system_freezable_wq = alloc_workqueue 5628 system_freezable_wq = alloc_workqueue("events_freezable",
7739 5629 WQ_FREEZABLE, 0);
7740 system_power_efficient_wq = alloc_wor 5630 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
7741 5631 WQ_POWER_EFFICIENT, 0);
7742 system_freezable_power_efficient_wq = !! 5632 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
7743 5633 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
7744 5634 0);
7745 system_bh_wq = alloc_workqueue("event <<
7746 system_bh_highpri_wq = alloc_workqueu <<
7747 <<
7748 BUG_ON(!system_wq || !system_highpri_ 5635 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
7749 !system_unbound_wq || !system_ 5636 !system_unbound_wq || !system_freezable_wq ||
7750 !system_power_efficient_wq || 5637 !system_power_efficient_wq ||
7751 !system_freezable_power_effici !! 5638 !system_freezable_power_efficient_wq);
7752 !system_bh_wq || !system_bh_hi <<
7753 } <<
7754 5639
7755 static void __init wq_cpu_intensive_thresh_in !! 5640 return 0;
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 } 5641 }
7793 5642
7794 /** 5643 /**
7795 * workqueue_init - bring workqueue subsystem 5644 * workqueue_init - bring workqueue subsystem fully online
7796 * 5645 *
7797 * This is the second step of three-staged wo !! 5646 * This is the latter half of two-staged workqueue subsystem initialization
7798 * and invoked as soon as kthreads can be cre !! 5647 * and invoked as soon as kthreads can be created and scheduled.
7799 * been created and work items queued on them !! 5648 * Workqueues have been created and work items queued on them, but there
7800 * executing the work items yet. Populate the !! 5649 * are no kworkers executing the work items yet. Populate the worker pools
7801 * workers and enable future kworker creation !! 5650 * with the initial workers and enable future kworker creations.
7802 */ 5651 */
7803 void __init workqueue_init(void) !! 5652 int __init workqueue_init(void)
7804 { 5653 {
7805 struct workqueue_struct *wq; 5654 struct workqueue_struct *wq;
7806 struct worker_pool *pool; 5655 struct worker_pool *pool;
7807 int cpu, bkt; 5656 int cpu, bkt;
7808 5657
7809 wq_cpu_intensive_thresh_init(); !! 5658 /*
>> 5659 * It'd be simpler to initialize NUMA in workqueue_init_early() but
>> 5660 * CPU to node mapping may not be available that early on some
>> 5661 * archs such as power and arm64. As per-cpu pools created
>> 5662 * previously could be missing node hint and unbound pools NUMA
>> 5663 * affinity, fix them up.
>> 5664 *
>> 5665 * Also, while iterating workqueues, create rescuers if requested.
>> 5666 */
>> 5667 wq_numa_init();
7810 5668
7811 mutex_lock(&wq_pool_mutex); 5669 mutex_lock(&wq_pool_mutex);
7812 5670
7813 /* <<
7814 * Per-cpu pools created earlier coul <<
7815 * up. Also, create a rescuer for wor <<
7816 */ <<
7817 for_each_possible_cpu(cpu) { 5671 for_each_possible_cpu(cpu) {
7818 for_each_bh_worker_pool(pool, !! 5672 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 5673 pool->node = cpu_to_node(cpu);
>> 5674 }
7822 } 5675 }
7823 5676
7824 list_for_each_entry(wq, &workqueues, 5677 list_for_each_entry(wq, &workqueues, list) {
>> 5678 wq_update_unbound_numa(wq, smp_processor_id(), true);
7825 WARN(init_rescuer(wq), 5679 WARN(init_rescuer(wq),
7826 "workqueue: failed to cr 5680 "workqueue: failed to create early rescuer for %s",
7827 wq->name); 5681 wq->name);
7828 } 5682 }
7829 5683
7830 mutex_unlock(&wq_pool_mutex); 5684 mutex_unlock(&wq_pool_mutex);
7831 5685
7832 /* !! 5686 /* 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) { 5687 for_each_online_cpu(cpu) {
7843 for_each_cpu_worker_pool(pool 5688 for_each_cpu_worker_pool(pool, cpu) {
7844 pool->flags &= ~POOL_ 5689 pool->flags &= ~POOL_DISASSOCIATED;
7845 BUG_ON(!create_worker 5690 BUG_ON(!create_worker(pool));
7846 } 5691 }
7847 } 5692 }
7848 5693
7849 hash_for_each(unbound_pool_hash, bkt, 5694 hash_for_each(unbound_pool_hash, bkt, pool, hash_node)
7850 BUG_ON(!create_worker(pool)); 5695 BUG_ON(!create_worker(pool));
7851 5696
7852 wq_online = true; 5697 wq_online = true;
7853 wq_watchdog_init(); 5698 wq_watchdog_init();
7854 } <<
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 5699
7866 pt->nr_pods = 0; !! 5700 return 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 } 5701 }
7973 __setup("workqueue.unbound_cpus=", workqueue_ <<
7974 5702
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.