Commit | Line | Data |
---|---|---|
457c8996 | 1 | // SPDX-License-Identifier: GPL-2.0-only |
1da177e4 | 2 | /* |
c54fce6e | 3 | * kernel/workqueue.c - generic async execution with shared worker pool |
1da177e4 | 4 | * |
c54fce6e | 5 | * Copyright (C) 2002 Ingo Molnar |
1da177e4 | 6 | * |
c54fce6e TH |
7 | * Derived from the taskqueue/keventd code by: |
8 | * David Woodhouse <dwmw2@infradead.org> | |
9 | * Andrew Morton | |
10 | * Kai Petzke <wpp@marie.physik.tu-berlin.de> | |
11 | * Theodore Ts'o <tytso@mit.edu> | |
1da177e4 | 12 | * |
c54fce6e | 13 | * Made to use alloc_percpu by Christoph Lameter. |
1da177e4 | 14 | * |
c54fce6e TH |
15 | * Copyright (C) 2010 SUSE Linux Products GmbH |
16 | * Copyright (C) 2010 Tejun Heo <tj@kernel.org> | |
89ada679 | 17 | * |
c54fce6e TH |
18 | * This is the generic async execution mechanism. Work items as are |
19 | * executed in process context. The worker pool is shared and | |
b11895c4 L |
20 | * automatically managed. There are two worker pools for each CPU (one for |
21 | * normal work items and the other for high priority ones) and some extra | |
22 | * pools for workqueues which are not bound to any specific CPU - the | |
23 | * number of these backing pools is dynamic. | |
c54fce6e | 24 | * |
9a261491 | 25 | * Please read Documentation/core-api/workqueue.rst for details. |
1da177e4 LT |
26 | */ |
27 | ||
9984de1a | 28 | #include <linux/export.h> |
1da177e4 LT |
29 | #include <linux/kernel.h> |
30 | #include <linux/sched.h> | |
31 | #include <linux/init.h> | |
32 | #include <linux/signal.h> | |
33 | #include <linux/completion.h> | |
34 | #include <linux/workqueue.h> | |
35 | #include <linux/slab.h> | |
36 | #include <linux/cpu.h> | |
37 | #include <linux/notifier.h> | |
38 | #include <linux/kthread.h> | |
1fa44eca | 39 | #include <linux/hardirq.h> |
46934023 | 40 | #include <linux/mempolicy.h> |
341a5958 | 41 | #include <linux/freezer.h> |
d5abe669 | 42 | #include <linux/debug_locks.h> |
4e6045f1 | 43 | #include <linux/lockdep.h> |
c34056a3 | 44 | #include <linux/idr.h> |
29c91e99 | 45 | #include <linux/jhash.h> |
42f8570f | 46 | #include <linux/hashtable.h> |
76af4d93 | 47 | #include <linux/rculist.h> |
bce90380 | 48 | #include <linux/nodemask.h> |
4c16bd32 | 49 | #include <linux/moduleparam.h> |
3d1cb205 | 50 | #include <linux/uaccess.h> |
c98a9805 | 51 | #include <linux/sched/isolation.h> |
cd2440d6 | 52 | #include <linux/sched/debug.h> |
62635ea8 | 53 | #include <linux/nmi.h> |
940d71c6 | 54 | #include <linux/kvm_para.h> |
aa6fde93 | 55 | #include <linux/delay.h> |
e22bee78 | 56 | |
ea138446 | 57 | #include "workqueue_internal.h" |
1da177e4 | 58 | |
e563d0a7 | 59 | enum worker_pool_flags { |
24647570 TH |
60 | /* |
61 | * worker_pool flags | |
bc2ae0f5 | 62 | * |
24647570 | 63 | * A bound pool is either associated or disassociated with its CPU. |
bc2ae0f5 TH |
64 | * While associated (!DISASSOCIATED), all workers are bound to the |
65 | * CPU and none has %WORKER_UNBOUND set and concurrency management | |
66 | * is in effect. | |
67 | * | |
68 | * While DISASSOCIATED, the cpu may be offline and all workers have | |
69 | * %WORKER_UNBOUND set and concurrency management disabled, and may | |
24647570 | 70 | * be executing on any CPU. The pool behaves as an unbound one. |
bc2ae0f5 | 71 | * |
bc3a1afc | 72 | * Note that DISASSOCIATED should be flipped only while holding |
1258fae7 | 73 | * wq_pool_attach_mutex to avoid changing binding state while |
4736cbf7 | 74 | * worker_attach_to_pool() is in progress. |
bc2ae0f5 | 75 | */ |
692b4825 | 76 | POOL_MANAGER_ACTIVE = 1 << 0, /* being managed */ |
24647570 | 77 | POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */ |
e563d0a7 | 78 | }; |
db7bccf4 | 79 | |
e563d0a7 | 80 | enum worker_flags { |
c8e55f36 | 81 | /* worker flags */ |
c8e55f36 TH |
82 | WORKER_DIE = 1 << 1, /* die die die */ |
83 | WORKER_IDLE = 1 << 2, /* is idle */ | |
e22bee78 | 84 | WORKER_PREP = 1 << 3, /* preparing to run works */ |
fb0e7beb | 85 | WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */ |
f3421797 | 86 | WORKER_UNBOUND = 1 << 7, /* worker is unbound */ |
a9ab775b | 87 | WORKER_REBOUND = 1 << 8, /* worker was rebound */ |
e22bee78 | 88 | |
a9ab775b TH |
89 | WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE | |
90 | WORKER_UNBOUND | WORKER_REBOUND, | |
e563d0a7 | 91 | }; |
db7bccf4 | 92 | |
e563d0a7 | 93 | enum wq_internal_consts { |
e34cdddb | 94 | NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */ |
4ce62e9e | 95 | |
29c91e99 | 96 | UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */ |
c8e55f36 | 97 | BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */ |
db7bccf4 | 98 | |
e22bee78 TH |
99 | MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */ |
100 | IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */ | |
101 | ||
3233cdbd TH |
102 | MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2, |
103 | /* call for help after 10ms | |
104 | (min two ticks) */ | |
e22bee78 TH |
105 | MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */ |
106 | CREATE_COOLDOWN = HZ, /* time to breath after fail */ | |
e22bee78 TH |
107 | |
108 | /* | |
109 | * Rescue workers are used only on emergencies and shared by | |
8698a745 | 110 | * all cpus. Give MIN_NICE. |
e22bee78 | 111 | */ |
8698a745 DY |
112 | RESCUER_NICE_LEVEL = MIN_NICE, |
113 | HIGHPRI_NICE_LEVEL = MIN_NICE, | |
ecf6881f | 114 | |
31c89007 | 115 | WQ_NAME_LEN = 32, |
c8e55f36 | 116 | }; |
1da177e4 LT |
117 | |
118 | /* | |
4690c4ab TH |
119 | * Structure fields follow one of the following exclusion rules. |
120 | * | |
e41e704b TH |
121 | * I: Modifiable by initialization/destruction paths and read-only for |
122 | * everyone else. | |
4690c4ab | 123 | * |
e22bee78 TH |
124 | * P: Preemption protected. Disabling preemption is enough and should |
125 | * only be modified and accessed from the local cpu. | |
126 | * | |
d565ed63 | 127 | * L: pool->lock protected. Access with pool->lock held. |
4690c4ab | 128 | * |
5797b1c1 TH |
129 | * LN: pool->lock and wq_node_nr_active->lock protected for writes. Either for |
130 | * reads. | |
131 | * | |
bdf8b9bf TH |
132 | * K: Only modified by worker while holding pool->lock. Can be safely read by |
133 | * self, while holding pool->lock or from IRQ context if %current is the | |
134 | * kworker. | |
135 | * | |
136 | * S: Only modified by worker self. | |
137 | * | |
1258fae7 | 138 | * A: wq_pool_attach_mutex protected. |
822d8405 | 139 | * |
68e13a67 | 140 | * PL: wq_pool_mutex protected. |
5bcab335 | 141 | * |
24acfb71 | 142 | * PR: wq_pool_mutex protected for writes. RCU protected for reads. |
76af4d93 | 143 | * |
5b95e1af LJ |
144 | * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads. |
145 | * | |
146 | * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or | |
24acfb71 | 147 | * RCU for reads. |
5b95e1af | 148 | * |
3c25a55d LJ |
149 | * WQ: wq->mutex protected. |
150 | * | |
24acfb71 | 151 | * WR: wq->mutex protected for writes. RCU protected for reads. |
2e109a28 | 152 | * |
a045a272 TH |
153 | * WO: wq->mutex protected for writes. Updated with WRITE_ONCE() and can be read |
154 | * with READ_ONCE() without locking. | |
155 | * | |
2e109a28 | 156 | * MD: wq_mayday_lock protected. |
cd2440d6 PM |
157 | * |
158 | * WD: Used internally by the watchdog. | |
1da177e4 | 159 | */ |
1da177e4 | 160 | |
2eaebdb3 | 161 | /* struct worker is defined in workqueue_internal.h */ |
c34056a3 | 162 | |
bd7bdd43 | 163 | struct worker_pool { |
a9b8a985 | 164 | raw_spinlock_t lock; /* the pool lock */ |
d84ff051 | 165 | int cpu; /* I: the associated cpu */ |
f3f90ad4 | 166 | int node; /* I: the associated node ID */ |
9daf9e67 | 167 | int id; /* I: pool ID */ |
bc8b50c2 | 168 | unsigned int flags; /* L: flags */ |
bd7bdd43 | 169 | |
82607adc | 170 | unsigned long watchdog_ts; /* L: watchdog timestamp */ |
cd2440d6 | 171 | bool cpu_stall; /* WD: stalled cpu bound pool */ |
82607adc | 172 | |
bc35f7ef LJ |
173 | /* |
174 | * The counter is incremented in a process context on the associated CPU | |
175 | * w/ preemption disabled, and decremented or reset in the same context | |
176 | * but w/ pool->lock held. The readers grab pool->lock and are | |
177 | * guaranteed to see if the counter reached zero. | |
178 | */ | |
179 | int nr_running; | |
84f91c62 | 180 | |
bd7bdd43 | 181 | struct list_head worklist; /* L: list of pending works */ |
ea1abd61 | 182 | |
5826cc8f LJ |
183 | int nr_workers; /* L: total number of workers */ |
184 | int nr_idle; /* L: currently idle workers */ | |
bd7bdd43 | 185 | |
2c1f1a91 | 186 | struct list_head idle_list; /* L: list of idle workers */ |
bd7bdd43 | 187 | struct timer_list idle_timer; /* L: worker idle timeout */ |
3f959aa3 VS |
188 | struct work_struct idle_cull_work; /* L: worker idle cleanup */ |
189 | ||
190 | struct timer_list mayday_timer; /* L: SOS timer for workers */ | |
bd7bdd43 | 191 | |
c5aa87bb | 192 | /* a workers is either on busy_hash or idle_list, or the manager */ |
c9e7cf27 TH |
193 | DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER); |
194 | /* L: hash of busy workers */ | |
195 | ||
2607d7a6 | 196 | struct worker *manager; /* L: purely informational */ |
92f9c5c4 | 197 | struct list_head workers; /* A: attached workers */ |
e02b9312 | 198 | struct list_head dying_workers; /* A: workers about to die */ |
60f5a4bc | 199 | struct completion *detach_completion; /* all workers detached */ |
e19e397a | 200 | |
7cda9aae | 201 | struct ida worker_ida; /* worker IDs for task name */ |
e19e397a | 202 | |
7a4e344c | 203 | struct workqueue_attrs *attrs; /* I: worker attributes */ |
68e13a67 LJ |
204 | struct hlist_node hash_node; /* PL: unbound_pool_hash node */ |
205 | int refcnt; /* PL: refcnt for unbound pools */ | |
7a4e344c | 206 | |
29c91e99 | 207 | /* |
24acfb71 | 208 | * Destruction of pool is RCU protected to allow dereferences |
29c91e99 TH |
209 | * from get_work_pool(). |
210 | */ | |
211 | struct rcu_head rcu; | |
84f91c62 | 212 | }; |
8b03ae3c | 213 | |
725e8ec5 TH |
214 | /* |
215 | * Per-pool_workqueue statistics. These can be monitored using | |
216 | * tools/workqueue/wq_monitor.py. | |
217 | */ | |
218 | enum pool_workqueue_stats { | |
219 | PWQ_STAT_STARTED, /* work items started execution */ | |
220 | PWQ_STAT_COMPLETED, /* work items completed execution */ | |
8a1dd1e5 | 221 | PWQ_STAT_CPU_TIME, /* total CPU time consumed */ |
616db877 | 222 | PWQ_STAT_CPU_INTENSIVE, /* wq_cpu_intensive_thresh_us violations */ |
725e8ec5 | 223 | PWQ_STAT_CM_WAKEUP, /* concurrency-management worker wakeups */ |
8639eceb | 224 | PWQ_STAT_REPATRIATED, /* unbound workers brought back into scope */ |
725e8ec5 TH |
225 | PWQ_STAT_MAYDAY, /* maydays to rescuer */ |
226 | PWQ_STAT_RESCUED, /* linked work items executed by rescuer */ | |
227 | ||
228 | PWQ_NR_STATS, | |
229 | }; | |
230 | ||
1da177e4 | 231 | /* |
112202d9 TH |
232 | * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS |
233 | * of work_struct->data are used for flags and the remaining high bits | |
234 | * point to the pwq; thus, pwqs need to be aligned at two's power of the | |
235 | * number of flag bits. | |
1da177e4 | 236 | */ |
112202d9 | 237 | struct pool_workqueue { |
bd7bdd43 | 238 | struct worker_pool *pool; /* I: the associated pool */ |
4690c4ab | 239 | struct workqueue_struct *wq; /* I: the owning workqueue */ |
73f53c4a TH |
240 | int work_color; /* L: current color */ |
241 | int flush_color; /* L: flushing color */ | |
8864b4e5 | 242 | int refcnt; /* L: reference count */ |
73f53c4a TH |
243 | int nr_in_flight[WORK_NR_COLORS]; |
244 | /* L: nr of in_flight works */ | |
018f3a13 LJ |
245 | |
246 | /* | |
247 | * nr_active management and WORK_STRUCT_INACTIVE: | |
248 | * | |
249 | * When pwq->nr_active >= max_active, new work item is queued to | |
250 | * pwq->inactive_works instead of pool->worklist and marked with | |
251 | * WORK_STRUCT_INACTIVE. | |
252 | * | |
5797b1c1 TH |
253 | * All work items marked with WORK_STRUCT_INACTIVE do not participate in |
254 | * nr_active and all work items in pwq->inactive_works are marked with | |
255 | * WORK_STRUCT_INACTIVE. But not all WORK_STRUCT_INACTIVE work items are | |
256 | * in pwq->inactive_works. Some of them are ready to run in | |
257 | * pool->worklist or worker->scheduled. Those work itmes are only struct | |
258 | * wq_barrier which is used for flush_work() and should not participate | |
259 | * in nr_active. For non-barrier work item, it is marked with | |
260 | * WORK_STRUCT_INACTIVE iff it is in pwq->inactive_works. | |
018f3a13 | 261 | */ |
1e19ffc6 | 262 | int nr_active; /* L: nr of active works */ |
f97a4a1a | 263 | struct list_head inactive_works; /* L: inactive works */ |
5797b1c1 | 264 | struct list_head pending_node; /* LN: node on wq_node_nr_active->pending_pwqs */ |
3c25a55d | 265 | struct list_head pwqs_node; /* WR: node on wq->pwqs */ |
2e109a28 | 266 | struct list_head mayday_node; /* MD: node on wq->maydays */ |
8864b4e5 | 267 | |
725e8ec5 TH |
268 | u64 stats[PWQ_NR_STATS]; |
269 | ||
8864b4e5 | 270 | /* |
967b494e | 271 | * Release of unbound pwq is punted to a kthread_worker. See put_pwq() |
687a9aa5 TH |
272 | * and pwq_release_workfn() for details. pool_workqueue itself is also |
273 | * RCU protected so that the first pwq can be determined without | |
967b494e | 274 | * grabbing wq->mutex. |
8864b4e5 | 275 | */ |
687a9aa5 | 276 | struct kthread_work release_work; |
8864b4e5 | 277 | struct rcu_head rcu; |
e904e6c2 | 278 | } __aligned(1 << WORK_STRUCT_FLAG_BITS); |
1da177e4 | 279 | |
73f53c4a TH |
280 | /* |
281 | * Structure used to wait for workqueue flush. | |
282 | */ | |
283 | struct wq_flusher { | |
3c25a55d LJ |
284 | struct list_head list; /* WQ: list of flushers */ |
285 | int flush_color; /* WQ: flush color waiting for */ | |
73f53c4a TH |
286 | struct completion done; /* flush completion */ |
287 | }; | |
288 | ||
226223ab TH |
289 | struct wq_device; |
290 | ||
91ccc6e7 TH |
291 | /* |
292 | * Unlike in a per-cpu workqueue where max_active limits its concurrency level | |
293 | * on each CPU, in an unbound workqueue, max_active applies to the whole system. | |
294 | * As sharing a single nr_active across multiple sockets can be very expensive, | |
295 | * the counting and enforcement is per NUMA node. | |
5797b1c1 TH |
296 | * |
297 | * The following struct is used to enforce per-node max_active. When a pwq wants | |
298 | * to start executing a work item, it should increment ->nr using | |
299 | * tryinc_node_nr_active(). If acquisition fails due to ->nr already being over | |
300 | * ->max, the pwq is queued on ->pending_pwqs. As in-flight work items finish | |
301 | * and decrement ->nr, node_activate_pending_pwq() activates the pending pwqs in | |
302 | * round-robin order. | |
91ccc6e7 TH |
303 | */ |
304 | struct wq_node_nr_active { | |
5797b1c1 TH |
305 | int max; /* per-node max_active */ |
306 | atomic_t nr; /* per-node nr_active */ | |
307 | raw_spinlock_t lock; /* nests inside pool locks */ | |
308 | struct list_head pending_pwqs; /* LN: pwqs with inactive works */ | |
91ccc6e7 TH |
309 | }; |
310 | ||
1da177e4 | 311 | /* |
c5aa87bb TH |
312 | * The externally visible workqueue. It relays the issued work items to |
313 | * the appropriate worker_pool through its pool_workqueues. | |
1da177e4 LT |
314 | */ |
315 | struct workqueue_struct { | |
3c25a55d | 316 | struct list_head pwqs; /* WR: all pwqs of this wq */ |
e2dca7ad | 317 | struct list_head list; /* PR: list of all workqueues */ |
73f53c4a | 318 | |
3c25a55d LJ |
319 | struct mutex mutex; /* protects this wq */ |
320 | int work_color; /* WQ: current work color */ | |
321 | int flush_color; /* WQ: current flush color */ | |
112202d9 | 322 | atomic_t nr_pwqs_to_flush; /* flush in progress */ |
3c25a55d LJ |
323 | struct wq_flusher *first_flusher; /* WQ: first flusher */ |
324 | struct list_head flusher_queue; /* WQ: flush waiters */ | |
325 | struct list_head flusher_overflow; /* WQ: flush overflow list */ | |
73f53c4a | 326 | |
2e109a28 | 327 | struct list_head maydays; /* MD: pwqs requesting rescue */ |
30ae2fc0 | 328 | struct worker *rescuer; /* MD: rescue worker */ |
e22bee78 | 329 | |
87fc741e | 330 | int nr_drainers; /* WQ: drain in progress */ |
5797b1c1 TH |
331 | |
332 | /* See alloc_workqueue() function comment for info on min/max_active */ | |
a045a272 | 333 | int max_active; /* WO: max active works */ |
5797b1c1 | 334 | int min_active; /* WO: min active works */ |
a045a272 | 335 | int saved_max_active; /* WQ: saved max_active */ |
5797b1c1 | 336 | int saved_min_active; /* WQ: saved min_active */ |
226223ab | 337 | |
5b95e1af | 338 | struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */ |
9f66cff2 | 339 | struct pool_workqueue __rcu *dfl_pwq; /* PW: only for unbound wqs */ |
6029a918 | 340 | |
226223ab TH |
341 | #ifdef CONFIG_SYSFS |
342 | struct wq_device *wq_dev; /* I: for sysfs interface */ | |
343 | #endif | |
4e6045f1 | 344 | #ifdef CONFIG_LOCKDEP |
669de8bd BVA |
345 | char *lock_name; |
346 | struct lock_class_key key; | |
4690c4ab | 347 | struct lockdep_map lockdep_map; |
4e6045f1 | 348 | #endif |
ecf6881f | 349 | char name[WQ_NAME_LEN]; /* I: workqueue name */ |
2728fd2f | 350 | |
e2dca7ad | 351 | /* |
24acfb71 TG |
352 | * Destruction of workqueue_struct is RCU protected to allow walking |
353 | * the workqueues list without grabbing wq_pool_mutex. | |
e2dca7ad TH |
354 | * This is used to dump all workqueues from sysrq. |
355 | */ | |
356 | struct rcu_head rcu; | |
357 | ||
2728fd2f TH |
358 | /* hot fields used during command issue, aligned to cacheline */ |
359 | unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */ | |
636b927e | 360 | struct pool_workqueue __percpu __rcu **cpu_pwq; /* I: per-cpu pwqs */ |
91ccc6e7 | 361 | struct wq_node_nr_active *node_nr_active[]; /* I: per-node nr_active */ |
1da177e4 LT |
362 | }; |
363 | ||
e904e6c2 TH |
364 | static struct kmem_cache *pwq_cache; |
365 | ||
84193c07 TH |
366 | /* |
367 | * Each pod type describes how CPUs should be grouped for unbound workqueues. | |
368 | * See the comment above workqueue_attrs->affn_scope. | |
369 | */ | |
370 | struct wq_pod_type { | |
371 | int nr_pods; /* number of pods */ | |
372 | cpumask_var_t *pod_cpus; /* pod -> cpus */ | |
373 | int *pod_node; /* pod -> node */ | |
374 | int *cpu_pod; /* cpu -> pod */ | |
375 | }; | |
376 | ||
377 | static struct wq_pod_type wq_pod_types[WQ_AFFN_NR_TYPES]; | |
523a301e | 378 | static enum wq_affn_scope wq_affn_dfl = WQ_AFFN_CACHE; |
63c5484e TH |
379 | |
380 | static const char *wq_affn_names[WQ_AFFN_NR_TYPES] = { | |
523a301e | 381 | [WQ_AFFN_DFL] = "default", |
63c5484e TH |
382 | [WQ_AFFN_CPU] = "cpu", |
383 | [WQ_AFFN_SMT] = "smt", | |
384 | [WQ_AFFN_CACHE] = "cache", | |
385 | [WQ_AFFN_NUMA] = "numa", | |
386 | [WQ_AFFN_SYSTEM] = "system", | |
387 | }; | |
bce90380 | 388 | |
616db877 TH |
389 | /* |
390 | * Per-cpu work items which run for longer than the following threshold are | |
391 | * automatically considered CPU intensive and excluded from concurrency | |
392 | * management to prevent them from noticeably delaying other per-cpu work items. | |
aa6fde93 TH |
393 | * ULONG_MAX indicates that the user hasn't overridden it with a boot parameter. |
394 | * The actual value is initialized in wq_cpu_intensive_thresh_init(). | |
616db877 | 395 | */ |
aa6fde93 | 396 | static unsigned long wq_cpu_intensive_thresh_us = ULONG_MAX; |
616db877 TH |
397 | module_param_named(cpu_intensive_thresh_us, wq_cpu_intensive_thresh_us, ulong, 0644); |
398 | ||
cee22a15 | 399 | /* see the comment above the definition of WQ_POWER_EFFICIENT */ |
552f530c | 400 | static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT); |
cee22a15 VK |
401 | module_param_named(power_efficient, wq_power_efficient, bool, 0444); |
402 | ||
863b710b | 403 | static bool wq_online; /* can kworkers be created yet? */ |
3347fa09 | 404 | |
fef59c9c TH |
405 | /* buf for wq_update_unbound_pod_attrs(), protected by CPU hotplug exclusion */ |
406 | static struct workqueue_attrs *wq_update_pod_attrs_buf; | |
4c16bd32 | 407 | |
68e13a67 | 408 | static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ |
1258fae7 | 409 | static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */ |
a9b8a985 | 410 | static DEFINE_RAW_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ |
d8bb65ab SAS |
411 | /* wait for manager to go away */ |
412 | static struct rcuwait manager_wait = __RCUWAIT_INITIALIZER(manager_wait); | |
5bcab335 | 413 | |
e2dca7ad | 414 | static LIST_HEAD(workqueues); /* PR: list of all workqueues */ |
68e13a67 | 415 | static bool workqueue_freezing; /* PL: have wqs started freezing? */ |
7d19c5ce | 416 | |
99c621ef | 417 | /* PL&A: allowable cpus for unbound wqs and work items */ |
ef557180 MG |
418 | static cpumask_var_t wq_unbound_cpumask; |
419 | ||
fe28f631 WL |
420 | /* PL: user requested unbound cpumask via sysfs */ |
421 | static cpumask_var_t wq_requested_unbound_cpumask; | |
422 | ||
423 | /* PL: isolated cpumask to be excluded from unbound cpumask */ | |
424 | static cpumask_var_t wq_isolated_cpumask; | |
425 | ||
ace3c549 | 426 | /* for further constrain wq_unbound_cpumask by cmdline parameter*/ |
427 | static struct cpumask wq_cmdline_cpumask __initdata; | |
428 | ||
ef557180 MG |
429 | /* CPU where unbound work was last round robin scheduled from this CPU */ |
430 | static DEFINE_PER_CPU(int, wq_rr_cpu_last); | |
b05a7928 | 431 | |
f303fccb TH |
432 | /* |
433 | * Local execution of unbound work items is no longer guaranteed. The | |
434 | * following always forces round-robin CPU selection on unbound work items | |
435 | * to uncover usages which depend on it. | |
436 | */ | |
437 | #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU | |
438 | static bool wq_debug_force_rr_cpu = true; | |
439 | #else | |
440 | static bool wq_debug_force_rr_cpu = false; | |
441 | #endif | |
442 | module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644); | |
443 | ||
7d19c5ce | 444 | /* the per-cpu worker pools */ |
25528213 | 445 | static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools); |
7d19c5ce | 446 | |
68e13a67 | 447 | static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ |
7d19c5ce | 448 | |
68e13a67 | 449 | /* PL: hash of all unbound pools keyed by pool->attrs */ |
29c91e99 TH |
450 | static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); |
451 | ||
c5aa87bb | 452 | /* I: attributes used when instantiating standard unbound pools on demand */ |
29c91e99 TH |
453 | static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; |
454 | ||
8a2b7538 TH |
455 | /* I: attributes used when instantiating ordered pools on demand */ |
456 | static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS]; | |
457 | ||
967b494e TH |
458 | /* |
459 | * I: kthread_worker to release pwq's. pwq release needs to be bounced to a | |
460 | * process context while holding a pool lock. Bounce to a dedicated kthread | |
461 | * worker to avoid A-A deadlocks. | |
462 | */ | |
68279f9c | 463 | static struct kthread_worker *pwq_release_worker __ro_after_init; |
967b494e | 464 | |
68279f9c | 465 | struct workqueue_struct *system_wq __ro_after_init; |
ad7b1f84 | 466 | EXPORT_SYMBOL(system_wq); |
68279f9c | 467 | struct workqueue_struct *system_highpri_wq __ro_after_init; |
1aabe902 | 468 | EXPORT_SYMBOL_GPL(system_highpri_wq); |
68279f9c | 469 | struct workqueue_struct *system_long_wq __ro_after_init; |
d320c038 | 470 | EXPORT_SYMBOL_GPL(system_long_wq); |
68279f9c | 471 | struct workqueue_struct *system_unbound_wq __ro_after_init; |
f3421797 | 472 | EXPORT_SYMBOL_GPL(system_unbound_wq); |
68279f9c | 473 | struct workqueue_struct *system_freezable_wq __ro_after_init; |
24d51add | 474 | EXPORT_SYMBOL_GPL(system_freezable_wq); |
68279f9c | 475 | struct workqueue_struct *system_power_efficient_wq __ro_after_init; |
0668106c | 476 | EXPORT_SYMBOL_GPL(system_power_efficient_wq); |
68279f9c | 477 | struct workqueue_struct *system_freezable_power_efficient_wq __ro_after_init; |
0668106c | 478 | EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq); |
d320c038 | 479 | |
7d19c5ce | 480 | static int worker_thread(void *__worker); |
6ba94429 | 481 | static void workqueue_sysfs_unregister(struct workqueue_struct *wq); |
c29eb853 | 482 | static void show_pwq(struct pool_workqueue *pwq); |
55df0933 | 483 | static void show_one_worker_pool(struct worker_pool *pool); |
7d19c5ce | 484 | |
97bd2347 TH |
485 | #define CREATE_TRACE_POINTS |
486 | #include <trace/events/workqueue.h> | |
487 | ||
68e13a67 | 488 | #define assert_rcu_or_pool_mutex() \ |
24acfb71 | 489 | RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \ |
f78f5b90 | 490 | !lockdep_is_held(&wq_pool_mutex), \ |
24acfb71 | 491 | "RCU or wq_pool_mutex should be held") |
5bcab335 | 492 | |
5b95e1af | 493 | #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \ |
24acfb71 | 494 | RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \ |
f78f5b90 PM |
495 | !lockdep_is_held(&wq->mutex) && \ |
496 | !lockdep_is_held(&wq_pool_mutex), \ | |
24acfb71 | 497 | "RCU, wq->mutex or wq_pool_mutex should be held") |
5b95e1af | 498 | |
f02ae73a TH |
499 | #define for_each_cpu_worker_pool(pool, cpu) \ |
500 | for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ | |
501 | (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ | |
7a62c2c8 | 502 | (pool)++) |
4ce62e9e | 503 | |
17116969 TH |
504 | /** |
505 | * for_each_pool - iterate through all worker_pools in the system | |
506 | * @pool: iteration cursor | |
611c92a0 | 507 | * @pi: integer used for iteration |
fa1b54e6 | 508 | * |
24acfb71 | 509 | * This must be called either with wq_pool_mutex held or RCU read |
68e13a67 LJ |
510 | * locked. If the pool needs to be used beyond the locking in effect, the |
511 | * caller is responsible for guaranteeing that the pool stays online. | |
fa1b54e6 TH |
512 | * |
513 | * The if/else clause exists only for the lockdep assertion and can be | |
514 | * ignored. | |
17116969 | 515 | */ |
611c92a0 TH |
516 | #define for_each_pool(pool, pi) \ |
517 | idr_for_each_entry(&worker_pool_idr, pool, pi) \ | |
68e13a67 | 518 | if (({ assert_rcu_or_pool_mutex(); false; })) { } \ |
fa1b54e6 | 519 | else |
17116969 | 520 | |
822d8405 TH |
521 | /** |
522 | * for_each_pool_worker - iterate through all workers of a worker_pool | |
523 | * @worker: iteration cursor | |
822d8405 TH |
524 | * @pool: worker_pool to iterate workers of |
525 | * | |
1258fae7 | 526 | * This must be called with wq_pool_attach_mutex. |
822d8405 TH |
527 | * |
528 | * The if/else clause exists only for the lockdep assertion and can be | |
529 | * ignored. | |
530 | */ | |
da028469 LJ |
531 | #define for_each_pool_worker(worker, pool) \ |
532 | list_for_each_entry((worker), &(pool)->workers, node) \ | |
1258fae7 | 533 | if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \ |
822d8405 TH |
534 | else |
535 | ||
49e3cf44 TH |
536 | /** |
537 | * for_each_pwq - iterate through all pool_workqueues of the specified workqueue | |
538 | * @pwq: iteration cursor | |
539 | * @wq: the target workqueue | |
76af4d93 | 540 | * |
24acfb71 | 541 | * This must be called either with wq->mutex held or RCU read locked. |
794b18bc TH |
542 | * If the pwq needs to be used beyond the locking in effect, the caller is |
543 | * responsible for guaranteeing that the pwq stays online. | |
76af4d93 TH |
544 | * |
545 | * The if/else clause exists only for the lockdep assertion and can be | |
546 | * ignored. | |
49e3cf44 TH |
547 | */ |
548 | #define for_each_pwq(pwq, wq) \ | |
49e9d1a9 | 549 | list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \ |
5a644662 | 550 | lockdep_is_held(&(wq->mutex))) |
f3421797 | 551 | |
dc186ad7 TG |
552 | #ifdef CONFIG_DEBUG_OBJECTS_WORK |
553 | ||
f9e62f31 | 554 | static const struct debug_obj_descr work_debug_descr; |
dc186ad7 | 555 | |
99777288 SG |
556 | static void *work_debug_hint(void *addr) |
557 | { | |
558 | return ((struct work_struct *) addr)->func; | |
559 | } | |
560 | ||
b9fdac7f DC |
561 | static bool work_is_static_object(void *addr) |
562 | { | |
563 | struct work_struct *work = addr; | |
564 | ||
565 | return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work)); | |
566 | } | |
567 | ||
dc186ad7 TG |
568 | /* |
569 | * fixup_init is called when: | |
570 | * - an active object is initialized | |
571 | */ | |
02a982a6 | 572 | static bool work_fixup_init(void *addr, enum debug_obj_state state) |
dc186ad7 TG |
573 | { |
574 | struct work_struct *work = addr; | |
575 | ||
576 | switch (state) { | |
577 | case ODEBUG_STATE_ACTIVE: | |
578 | cancel_work_sync(work); | |
579 | debug_object_init(work, &work_debug_descr); | |
02a982a6 | 580 | return true; |
dc186ad7 | 581 | default: |
02a982a6 | 582 | return false; |
dc186ad7 TG |
583 | } |
584 | } | |
585 | ||
dc186ad7 TG |
586 | /* |
587 | * fixup_free is called when: | |
588 | * - an active object is freed | |
589 | */ | |
02a982a6 | 590 | static bool work_fixup_free(void *addr, enum debug_obj_state state) |
dc186ad7 TG |
591 | { |
592 | struct work_struct *work = addr; | |
593 | ||
594 | switch (state) { | |
595 | case ODEBUG_STATE_ACTIVE: | |
596 | cancel_work_sync(work); | |
597 | debug_object_free(work, &work_debug_descr); | |
02a982a6 | 598 | return true; |
dc186ad7 | 599 | default: |
02a982a6 | 600 | return false; |
dc186ad7 TG |
601 | } |
602 | } | |
603 | ||
f9e62f31 | 604 | static const struct debug_obj_descr work_debug_descr = { |
dc186ad7 | 605 | .name = "work_struct", |
99777288 | 606 | .debug_hint = work_debug_hint, |
b9fdac7f | 607 | .is_static_object = work_is_static_object, |
dc186ad7 | 608 | .fixup_init = work_fixup_init, |
dc186ad7 TG |
609 | .fixup_free = work_fixup_free, |
610 | }; | |
611 | ||
612 | static inline void debug_work_activate(struct work_struct *work) | |
613 | { | |
614 | debug_object_activate(work, &work_debug_descr); | |
615 | } | |
616 | ||
617 | static inline void debug_work_deactivate(struct work_struct *work) | |
618 | { | |
619 | debug_object_deactivate(work, &work_debug_descr); | |
620 | } | |
621 | ||
622 | void __init_work(struct work_struct *work, int onstack) | |
623 | { | |
624 | if (onstack) | |
625 | debug_object_init_on_stack(work, &work_debug_descr); | |
626 | else | |
627 | debug_object_init(work, &work_debug_descr); | |
628 | } | |
629 | EXPORT_SYMBOL_GPL(__init_work); | |
630 | ||
631 | void destroy_work_on_stack(struct work_struct *work) | |
632 | { | |
633 | debug_object_free(work, &work_debug_descr); | |
634 | } | |
635 | EXPORT_SYMBOL_GPL(destroy_work_on_stack); | |
636 | ||
ea2e64f2 TG |
637 | void destroy_delayed_work_on_stack(struct delayed_work *work) |
638 | { | |
639 | destroy_timer_on_stack(&work->timer); | |
640 | debug_object_free(&work->work, &work_debug_descr); | |
641 | } | |
642 | EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack); | |
643 | ||
dc186ad7 TG |
644 | #else |
645 | static inline void debug_work_activate(struct work_struct *work) { } | |
646 | static inline void debug_work_deactivate(struct work_struct *work) { } | |
647 | #endif | |
648 | ||
4e8b22bd | 649 | /** |
67dc8325 | 650 | * worker_pool_assign_id - allocate ID and assign it to @pool |
4e8b22bd LB |
651 | * @pool: the pool pointer of interest |
652 | * | |
653 | * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned | |
654 | * successfully, -errno on failure. | |
655 | */ | |
9daf9e67 TH |
656 | static int worker_pool_assign_id(struct worker_pool *pool) |
657 | { | |
658 | int ret; | |
659 | ||
68e13a67 | 660 | lockdep_assert_held(&wq_pool_mutex); |
5bcab335 | 661 | |
4e8b22bd LB |
662 | ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE, |
663 | GFP_KERNEL); | |
229641a6 | 664 | if (ret >= 0) { |
e68035fb | 665 | pool->id = ret; |
229641a6 TH |
666 | return 0; |
667 | } | |
fa1b54e6 | 668 | return ret; |
7c3eed5c TH |
669 | } |
670 | ||
9f66cff2 TH |
671 | static struct pool_workqueue __rcu ** |
672 | unbound_pwq_slot(struct workqueue_struct *wq, int cpu) | |
673 | { | |
674 | if (cpu >= 0) | |
675 | return per_cpu_ptr(wq->cpu_pwq, cpu); | |
676 | else | |
677 | return &wq->dfl_pwq; | |
678 | } | |
679 | ||
680 | /* @cpu < 0 for dfl_pwq */ | |
681 | static struct pool_workqueue *unbound_pwq(struct workqueue_struct *wq, int cpu) | |
682 | { | |
683 | return rcu_dereference_check(*unbound_pwq_slot(wq, cpu), | |
684 | lockdep_is_held(&wq_pool_mutex) || | |
685 | lockdep_is_held(&wq->mutex)); | |
686 | } | |
687 | ||
5797b1c1 TH |
688 | /** |
689 | * unbound_effective_cpumask - effective cpumask of an unbound workqueue | |
690 | * @wq: workqueue of interest | |
691 | * | |
692 | * @wq->unbound_attrs->cpumask contains the cpumask requested by the user which | |
693 | * is masked with wq_unbound_cpumask to determine the effective cpumask. The | |
694 | * default pwq is always mapped to the pool with the current effective cpumask. | |
695 | */ | |
696 | static struct cpumask *unbound_effective_cpumask(struct workqueue_struct *wq) | |
697 | { | |
698 | return unbound_pwq(wq, -1)->pool->attrs->__pod_cpumask; | |
699 | } | |
700 | ||
73f53c4a TH |
701 | static unsigned int work_color_to_flags(int color) |
702 | { | |
703 | return color << WORK_STRUCT_COLOR_SHIFT; | |
704 | } | |
705 | ||
c4560c2c | 706 | static int get_work_color(unsigned long work_data) |
73f53c4a | 707 | { |
c4560c2c | 708 | return (work_data >> WORK_STRUCT_COLOR_SHIFT) & |
73f53c4a TH |
709 | ((1 << WORK_STRUCT_COLOR_BITS) - 1); |
710 | } | |
711 | ||
712 | static int work_next_color(int color) | |
713 | { | |
714 | return (color + 1) % WORK_NR_COLORS; | |
715 | } | |
1da177e4 | 716 | |
14441960 | 717 | /* |
112202d9 TH |
718 | * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data |
719 | * contain the pointer to the queued pwq. Once execution starts, the flag | |
7c3eed5c | 720 | * is cleared and the high bits contain OFFQ flags and pool ID. |
7a22ad75 | 721 | * |
112202d9 TH |
722 | * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() |
723 | * and clear_work_data() can be used to set the pwq, pool or clear | |
bbb68dfa TH |
724 | * work->data. These functions should only be called while the work is |
725 | * owned - ie. while the PENDING bit is set. | |
7a22ad75 | 726 | * |
112202d9 | 727 | * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq |
7c3eed5c | 728 | * corresponding to a work. Pool is available once the work has been |
112202d9 | 729 | * queued anywhere after initialization until it is sync canceled. pwq is |
7c3eed5c | 730 | * available only while the work item is queued. |
7a22ad75 | 731 | * |
bbb68dfa TH |
732 | * %WORK_OFFQ_CANCELING is used to mark a work item which is being |
733 | * canceled. While being canceled, a work item may have its PENDING set | |
734 | * but stay off timer and worklist for arbitrarily long and nobody should | |
735 | * try to steal the PENDING bit. | |
14441960 | 736 | */ |
7a22ad75 TH |
737 | static inline void set_work_data(struct work_struct *work, unsigned long data, |
738 | unsigned long flags) | |
365970a1 | 739 | { |
6183c009 | 740 | WARN_ON_ONCE(!work_pending(work)); |
7a22ad75 TH |
741 | atomic_long_set(&work->data, data | flags | work_static(work)); |
742 | } | |
365970a1 | 743 | |
112202d9 | 744 | static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, |
7a22ad75 TH |
745 | unsigned long extra_flags) |
746 | { | |
112202d9 TH |
747 | set_work_data(work, (unsigned long)pwq, |
748 | WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); | |
365970a1 DH |
749 | } |
750 | ||
4468a00f LJ |
751 | static void set_work_pool_and_keep_pending(struct work_struct *work, |
752 | int pool_id) | |
753 | { | |
754 | set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, | |
755 | WORK_STRUCT_PENDING); | |
756 | } | |
757 | ||
7c3eed5c TH |
758 | static void set_work_pool_and_clear_pending(struct work_struct *work, |
759 | int pool_id) | |
7a22ad75 | 760 | { |
23657bb1 TH |
761 | /* |
762 | * The following wmb is paired with the implied mb in | |
763 | * test_and_set_bit(PENDING) and ensures all updates to @work made | |
764 | * here are visible to and precede any updates by the next PENDING | |
765 | * owner. | |
766 | */ | |
767 | smp_wmb(); | |
7c3eed5c | 768 | set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); |
346c09f8 RP |
769 | /* |
770 | * The following mb guarantees that previous clear of a PENDING bit | |
771 | * will not be reordered with any speculative LOADS or STORES from | |
772 | * work->current_func, which is executed afterwards. This possible | |
8bdc6201 | 773 | * reordering can lead to a missed execution on attempt to queue |
346c09f8 RP |
774 | * the same @work. E.g. consider this case: |
775 | * | |
776 | * CPU#0 CPU#1 | |
777 | * ---------------------------- -------------------------------- | |
778 | * | |
779 | * 1 STORE event_indicated | |
780 | * 2 queue_work_on() { | |
781 | * 3 test_and_set_bit(PENDING) | |
782 | * 4 } set_..._and_clear_pending() { | |
783 | * 5 set_work_data() # clear bit | |
784 | * 6 smp_mb() | |
785 | * 7 work->current_func() { | |
786 | * 8 LOAD event_indicated | |
787 | * } | |
788 | * | |
789 | * Without an explicit full barrier speculative LOAD on line 8 can | |
790 | * be executed before CPU#0 does STORE on line 1. If that happens, | |
791 | * CPU#0 observes the PENDING bit is still set and new execution of | |
792 | * a @work is not queued in a hope, that CPU#1 will eventually | |
793 | * finish the queued @work. Meanwhile CPU#1 does not see | |
794 | * event_indicated is set, because speculative LOAD was executed | |
795 | * before actual STORE. | |
796 | */ | |
797 | smp_mb(); | |
7a22ad75 | 798 | } |
f756d5e2 | 799 | |
7a22ad75 | 800 | static void clear_work_data(struct work_struct *work) |
1da177e4 | 801 | { |
7c3eed5c TH |
802 | smp_wmb(); /* see set_work_pool_and_clear_pending() */ |
803 | set_work_data(work, WORK_STRUCT_NO_POOL, 0); | |
1da177e4 LT |
804 | } |
805 | ||
afa4bb77 LT |
806 | static inline struct pool_workqueue *work_struct_pwq(unsigned long data) |
807 | { | |
808 | return (struct pool_workqueue *)(data & WORK_STRUCT_WQ_DATA_MASK); | |
809 | } | |
810 | ||
112202d9 | 811 | static struct pool_workqueue *get_work_pwq(struct work_struct *work) |
b1f4ec17 | 812 | { |
e120153d | 813 | unsigned long data = atomic_long_read(&work->data); |
7a22ad75 | 814 | |
112202d9 | 815 | if (data & WORK_STRUCT_PWQ) |
afa4bb77 | 816 | return work_struct_pwq(data); |
e120153d TH |
817 | else |
818 | return NULL; | |
4d707b9f ON |
819 | } |
820 | ||
7c3eed5c TH |
821 | /** |
822 | * get_work_pool - return the worker_pool a given work was associated with | |
823 | * @work: the work item of interest | |
824 | * | |
68e13a67 | 825 | * Pools are created and destroyed under wq_pool_mutex, and allows read |
24acfb71 TG |
826 | * access under RCU read lock. As such, this function should be |
827 | * called under wq_pool_mutex or inside of a rcu_read_lock() region. | |
fa1b54e6 TH |
828 | * |
829 | * All fields of the returned pool are accessible as long as the above | |
830 | * mentioned locking is in effect. If the returned pool needs to be used | |
831 | * beyond the critical section, the caller is responsible for ensuring the | |
832 | * returned pool is and stays online. | |
d185af30 YB |
833 | * |
834 | * Return: The worker_pool @work was last associated with. %NULL if none. | |
7c3eed5c TH |
835 | */ |
836 | static struct worker_pool *get_work_pool(struct work_struct *work) | |
365970a1 | 837 | { |
e120153d | 838 | unsigned long data = atomic_long_read(&work->data); |
7c3eed5c | 839 | int pool_id; |
7a22ad75 | 840 | |
68e13a67 | 841 | assert_rcu_or_pool_mutex(); |
fa1b54e6 | 842 | |
112202d9 | 843 | if (data & WORK_STRUCT_PWQ) |
afa4bb77 | 844 | return work_struct_pwq(data)->pool; |
7a22ad75 | 845 | |
7c3eed5c TH |
846 | pool_id = data >> WORK_OFFQ_POOL_SHIFT; |
847 | if (pool_id == WORK_OFFQ_POOL_NONE) | |
7a22ad75 TH |
848 | return NULL; |
849 | ||
fa1b54e6 | 850 | return idr_find(&worker_pool_idr, pool_id); |
7c3eed5c TH |
851 | } |
852 | ||
853 | /** | |
854 | * get_work_pool_id - return the worker pool ID a given work is associated with | |
855 | * @work: the work item of interest | |
856 | * | |
d185af30 | 857 | * Return: The worker_pool ID @work was last associated with. |
7c3eed5c TH |
858 | * %WORK_OFFQ_POOL_NONE if none. |
859 | */ | |
860 | static int get_work_pool_id(struct work_struct *work) | |
861 | { | |
54d5b7d0 LJ |
862 | unsigned long data = atomic_long_read(&work->data); |
863 | ||
112202d9 | 864 | if (data & WORK_STRUCT_PWQ) |
afa4bb77 | 865 | return work_struct_pwq(data)->pool->id; |
7c3eed5c | 866 | |
54d5b7d0 | 867 | return data >> WORK_OFFQ_POOL_SHIFT; |
7c3eed5c TH |
868 | } |
869 | ||
bbb68dfa TH |
870 | static void mark_work_canceling(struct work_struct *work) |
871 | { | |
7c3eed5c | 872 | unsigned long pool_id = get_work_pool_id(work); |
bbb68dfa | 873 | |
7c3eed5c TH |
874 | pool_id <<= WORK_OFFQ_POOL_SHIFT; |
875 | set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); | |
bbb68dfa TH |
876 | } |
877 | ||
878 | static bool work_is_canceling(struct work_struct *work) | |
879 | { | |
880 | unsigned long data = atomic_long_read(&work->data); | |
881 | ||
112202d9 | 882 | return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); |
bbb68dfa TH |
883 | } |
884 | ||
e22bee78 | 885 | /* |
3270476a TH |
886 | * Policy functions. These define the policies on how the global worker |
887 | * pools are managed. Unless noted otherwise, these functions assume that | |
d565ed63 | 888 | * they're being called with pool->lock held. |
e22bee78 TH |
889 | */ |
890 | ||
4594bf15 | 891 | /* |
e22bee78 TH |
892 | * Need to wake up a worker? Called from anything but currently |
893 | * running workers. | |
974271c4 TH |
894 | * |
895 | * Note that, because unbound workers never contribute to nr_running, this | |
706026c2 | 896 | * function will always return %true for unbound pools as long as the |
974271c4 | 897 | * worklist isn't empty. |
4594bf15 | 898 | */ |
63d95a91 | 899 | static bool need_more_worker(struct worker_pool *pool) |
365970a1 | 900 | { |
0219a352 | 901 | return !list_empty(&pool->worklist) && !pool->nr_running; |
e22bee78 | 902 | } |
4594bf15 | 903 | |
e22bee78 | 904 | /* Can I start working? Called from busy but !running workers. */ |
63d95a91 | 905 | static bool may_start_working(struct worker_pool *pool) |
e22bee78 | 906 | { |
63d95a91 | 907 | return pool->nr_idle; |
e22bee78 TH |
908 | } |
909 | ||
910 | /* Do I need to keep working? Called from currently running workers. */ | |
63d95a91 | 911 | static bool keep_working(struct worker_pool *pool) |
e22bee78 | 912 | { |
bc35f7ef | 913 | return !list_empty(&pool->worklist) && (pool->nr_running <= 1); |
e22bee78 TH |
914 | } |
915 | ||
916 | /* Do we need a new worker? Called from manager. */ | |
63d95a91 | 917 | static bool need_to_create_worker(struct worker_pool *pool) |
e22bee78 | 918 | { |
63d95a91 | 919 | return need_more_worker(pool) && !may_start_working(pool); |
e22bee78 | 920 | } |
365970a1 | 921 | |
e22bee78 | 922 | /* Do we have too many workers and should some go away? */ |
63d95a91 | 923 | static bool too_many_workers(struct worker_pool *pool) |
e22bee78 | 924 | { |
692b4825 | 925 | bool managing = pool->flags & POOL_MANAGER_ACTIVE; |
63d95a91 TH |
926 | int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ |
927 | int nr_busy = pool->nr_workers - nr_idle; | |
e22bee78 TH |
928 | |
929 | return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; | |
365970a1 DH |
930 | } |
931 | ||
c54d5046 TH |
932 | /** |
933 | * worker_set_flags - set worker flags and adjust nr_running accordingly | |
934 | * @worker: self | |
935 | * @flags: flags to set | |
936 | * | |
937 | * Set @flags in @worker->flags and adjust nr_running accordingly. | |
c54d5046 TH |
938 | */ |
939 | static inline void worker_set_flags(struct worker *worker, unsigned int flags) | |
940 | { | |
941 | struct worker_pool *pool = worker->pool; | |
942 | ||
bc8b50c2 | 943 | lockdep_assert_held(&pool->lock); |
c54d5046 TH |
944 | |
945 | /* If transitioning into NOT_RUNNING, adjust nr_running. */ | |
946 | if ((flags & WORKER_NOT_RUNNING) && | |
947 | !(worker->flags & WORKER_NOT_RUNNING)) { | |
948 | pool->nr_running--; | |
949 | } | |
950 | ||
951 | worker->flags |= flags; | |
952 | } | |
953 | ||
954 | /** | |
955 | * worker_clr_flags - clear worker flags and adjust nr_running accordingly | |
956 | * @worker: self | |
957 | * @flags: flags to clear | |
958 | * | |
959 | * Clear @flags in @worker->flags and adjust nr_running accordingly. | |
c54d5046 TH |
960 | */ |
961 | static inline void worker_clr_flags(struct worker *worker, unsigned int flags) | |
962 | { | |
963 | struct worker_pool *pool = worker->pool; | |
964 | unsigned int oflags = worker->flags; | |
965 | ||
bc8b50c2 | 966 | lockdep_assert_held(&pool->lock); |
c54d5046 TH |
967 | |
968 | worker->flags &= ~flags; | |
969 | ||
970 | /* | |
971 | * If transitioning out of NOT_RUNNING, increment nr_running. Note | |
972 | * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask | |
973 | * of multiple flags, not a single flag. | |
974 | */ | |
975 | if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) | |
976 | if (!(worker->flags & WORKER_NOT_RUNNING)) | |
977 | pool->nr_running++; | |
978 | } | |
979 | ||
797e8345 TH |
980 | /* Return the first idle worker. Called with pool->lock held. */ |
981 | static struct worker *first_idle_worker(struct worker_pool *pool) | |
982 | { | |
983 | if (unlikely(list_empty(&pool->idle_list))) | |
984 | return NULL; | |
985 | ||
986 | return list_first_entry(&pool->idle_list, struct worker, entry); | |
987 | } | |
988 | ||
989 | /** | |
990 | * worker_enter_idle - enter idle state | |
991 | * @worker: worker which is entering idle state | |
992 | * | |
993 | * @worker is entering idle state. Update stats and idle timer if | |
994 | * necessary. | |
995 | * | |
996 | * LOCKING: | |
997 | * raw_spin_lock_irq(pool->lock). | |
998 | */ | |
999 | static void worker_enter_idle(struct worker *worker) | |
1000 | { | |
1001 | struct worker_pool *pool = worker->pool; | |
1002 | ||
1003 | if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || | |
1004 | WARN_ON_ONCE(!list_empty(&worker->entry) && | |
1005 | (worker->hentry.next || worker->hentry.pprev))) | |
1006 | return; | |
1007 | ||
1008 | /* can't use worker_set_flags(), also called from create_worker() */ | |
1009 | worker->flags |= WORKER_IDLE; | |
1010 | pool->nr_idle++; | |
1011 | worker->last_active = jiffies; | |
1012 | ||
1013 | /* idle_list is LIFO */ | |
1014 | list_add(&worker->entry, &pool->idle_list); | |
1015 | ||
1016 | if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) | |
1017 | mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); | |
1018 | ||
1019 | /* Sanity check nr_running. */ | |
1020 | WARN_ON_ONCE(pool->nr_workers == pool->nr_idle && pool->nr_running); | |
1021 | } | |
1022 | ||
1023 | /** | |
1024 | * worker_leave_idle - leave idle state | |
1025 | * @worker: worker which is leaving idle state | |
1026 | * | |
1027 | * @worker is leaving idle state. Update stats. | |
1028 | * | |
1029 | * LOCKING: | |
1030 | * raw_spin_lock_irq(pool->lock). | |
1031 | */ | |
1032 | static void worker_leave_idle(struct worker *worker) | |
1033 | { | |
1034 | struct worker_pool *pool = worker->pool; | |
1035 | ||
1036 | if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) | |
1037 | return; | |
1038 | worker_clr_flags(worker, WORKER_IDLE); | |
1039 | pool->nr_idle--; | |
1040 | list_del_init(&worker->entry); | |
1041 | } | |
1042 | ||
1043 | /** | |
1044 | * find_worker_executing_work - find worker which is executing a work | |
1045 | * @pool: pool of interest | |
1046 | * @work: work to find worker for | |
1047 | * | |
1048 | * Find a worker which is executing @work on @pool by searching | |
1049 | * @pool->busy_hash which is keyed by the address of @work. For a worker | |
1050 | * to match, its current execution should match the address of @work and | |
1051 | * its work function. This is to avoid unwanted dependency between | |
1052 | * unrelated work executions through a work item being recycled while still | |
1053 | * being executed. | |
1054 | * | |
1055 | * This is a bit tricky. A work item may be freed once its execution | |
1056 | * starts and nothing prevents the freed area from being recycled for | |
1057 | * another work item. If the same work item address ends up being reused | |
1058 | * before the original execution finishes, workqueue will identify the | |
1059 | * recycled work item as currently executing and make it wait until the | |
1060 | * current execution finishes, introducing an unwanted dependency. | |
1061 | * | |
1062 | * This function checks the work item address and work function to avoid | |
1063 | * false positives. Note that this isn't complete as one may construct a | |
1064 | * work function which can introduce dependency onto itself through a | |
1065 | * recycled work item. Well, if somebody wants to shoot oneself in the | |
1066 | * foot that badly, there's only so much we can do, and if such deadlock | |
1067 | * actually occurs, it should be easy to locate the culprit work function. | |
1068 | * | |
1069 | * CONTEXT: | |
1070 | * raw_spin_lock_irq(pool->lock). | |
1071 | * | |
1072 | * Return: | |
1073 | * Pointer to worker which is executing @work if found, %NULL | |
1074 | * otherwise. | |
1075 | */ | |
1076 | static struct worker *find_worker_executing_work(struct worker_pool *pool, | |
1077 | struct work_struct *work) | |
1078 | { | |
1079 | struct worker *worker; | |
1080 | ||
1081 | hash_for_each_possible(pool->busy_hash, worker, hentry, | |
1082 | (unsigned long)work) | |
1083 | if (worker->current_work == work && | |
1084 | worker->current_func == work->func) | |
1085 | return worker; | |
1086 | ||
1087 | return NULL; | |
1088 | } | |
1089 | ||
1090 | /** | |
1091 | * move_linked_works - move linked works to a list | |
1092 | * @work: start of series of works to be scheduled | |
1093 | * @head: target list to append @work to | |
1094 | * @nextp: out parameter for nested worklist walking | |
1095 | * | |
873eaca6 TH |
1096 | * Schedule linked works starting from @work to @head. Work series to be |
1097 | * scheduled starts at @work and includes any consecutive work with | |
1098 | * WORK_STRUCT_LINKED set in its predecessor. See assign_work() for details on | |
1099 | * @nextp. | |
797e8345 TH |
1100 | * |
1101 | * CONTEXT: | |
1102 | * raw_spin_lock_irq(pool->lock). | |
1103 | */ | |
1104 | static void move_linked_works(struct work_struct *work, struct list_head *head, | |
1105 | struct work_struct **nextp) | |
1106 | { | |
1107 | struct work_struct *n; | |
1108 | ||
1109 | /* | |
1110 | * Linked worklist will always end before the end of the list, | |
1111 | * use NULL for list head. | |
1112 | */ | |
1113 | list_for_each_entry_safe_from(work, n, NULL, entry) { | |
1114 | list_move_tail(&work->entry, head); | |
1115 | if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) | |
1116 | break; | |
1117 | } | |
1118 | ||
1119 | /* | |
1120 | * If we're already inside safe list traversal and have moved | |
1121 | * multiple works to the scheduled queue, the next position | |
1122 | * needs to be updated. | |
1123 | */ | |
1124 | if (nextp) | |
1125 | *nextp = n; | |
1126 | } | |
1127 | ||
873eaca6 TH |
1128 | /** |
1129 | * assign_work - assign a work item and its linked work items to a worker | |
1130 | * @work: work to assign | |
1131 | * @worker: worker to assign to | |
1132 | * @nextp: out parameter for nested worklist walking | |
1133 | * | |
1134 | * Assign @work and its linked work items to @worker. If @work is already being | |
1135 | * executed by another worker in the same pool, it'll be punted there. | |
1136 | * | |
1137 | * If @nextp is not NULL, it's updated to point to the next work of the last | |
1138 | * scheduled work. This allows assign_work() to be nested inside | |
1139 | * list_for_each_entry_safe(). | |
1140 | * | |
1141 | * Returns %true if @work was successfully assigned to @worker. %false if @work | |
1142 | * was punted to another worker already executing it. | |
1143 | */ | |
1144 | static bool assign_work(struct work_struct *work, struct worker *worker, | |
1145 | struct work_struct **nextp) | |
1146 | { | |
1147 | struct worker_pool *pool = worker->pool; | |
1148 | struct worker *collision; | |
1149 | ||
1150 | lockdep_assert_held(&pool->lock); | |
1151 | ||
1152 | /* | |
1153 | * A single work shouldn't be executed concurrently by multiple workers. | |
1154 | * __queue_work() ensures that @work doesn't jump to a different pool | |
1155 | * while still running in the previous pool. Here, we should ensure that | |
1156 | * @work is not executed concurrently by multiple workers from the same | |
1157 | * pool. Check whether anyone is already processing the work. If so, | |
1158 | * defer the work to the currently executing one. | |
1159 | */ | |
1160 | collision = find_worker_executing_work(pool, work); | |
1161 | if (unlikely(collision)) { | |
1162 | move_linked_works(work, &collision->scheduled, nextp); | |
1163 | return false; | |
1164 | } | |
1165 | ||
1166 | move_linked_works(work, &worker->scheduled, nextp); | |
1167 | return true; | |
1168 | } | |
1169 | ||
797e8345 | 1170 | /** |
0219a352 TH |
1171 | * kick_pool - wake up an idle worker if necessary |
1172 | * @pool: pool to kick | |
797e8345 | 1173 | * |
0219a352 TH |
1174 | * @pool may have pending work items. Wake up worker if necessary. Returns |
1175 | * whether a worker was woken up. | |
797e8345 | 1176 | */ |
0219a352 | 1177 | static bool kick_pool(struct worker_pool *pool) |
797e8345 TH |
1178 | { |
1179 | struct worker *worker = first_idle_worker(pool); | |
8639eceb | 1180 | struct task_struct *p; |
797e8345 | 1181 | |
0219a352 TH |
1182 | lockdep_assert_held(&pool->lock); |
1183 | ||
1184 | if (!need_more_worker(pool) || !worker) | |
1185 | return false; | |
1186 | ||
8639eceb TH |
1187 | p = worker->task; |
1188 | ||
1189 | #ifdef CONFIG_SMP | |
1190 | /* | |
1191 | * Idle @worker is about to execute @work and waking up provides an | |
1192 | * opportunity to migrate @worker at a lower cost by setting the task's | |
1193 | * wake_cpu field. Let's see if we want to move @worker to improve | |
1194 | * execution locality. | |
1195 | * | |
1196 | * We're waking the worker that went idle the latest and there's some | |
1197 | * chance that @worker is marked idle but hasn't gone off CPU yet. If | |
1198 | * so, setting the wake_cpu won't do anything. As this is a best-effort | |
1199 | * optimization and the race window is narrow, let's leave as-is for | |
1200 | * now. If this becomes pronounced, we can skip over workers which are | |
1201 | * still on cpu when picking an idle worker. | |
1202 | * | |
1203 | * If @pool has non-strict affinity, @worker might have ended up outside | |
1204 | * its affinity scope. Repatriate. | |
1205 | */ | |
1206 | if (!pool->attrs->affn_strict && | |
1207 | !cpumask_test_cpu(p->wake_cpu, pool->attrs->__pod_cpumask)) { | |
1208 | struct work_struct *work = list_first_entry(&pool->worklist, | |
1209 | struct work_struct, entry); | |
1210 | p->wake_cpu = cpumask_any_distribute(pool->attrs->__pod_cpumask); | |
1211 | get_work_pwq(work)->stats[PWQ_STAT_REPATRIATED]++; | |
1212 | } | |
1213 | #endif | |
1214 | wake_up_process(p); | |
0219a352 | 1215 | return true; |
797e8345 TH |
1216 | } |
1217 | ||
63638450 TH |
1218 | #ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT |
1219 | ||
1220 | /* | |
1221 | * Concurrency-managed per-cpu work items that hog CPU for longer than | |
1222 | * wq_cpu_intensive_thresh_us trigger the automatic CPU_INTENSIVE mechanism, | |
1223 | * which prevents them from stalling other concurrency-managed work items. If a | |
1224 | * work function keeps triggering this mechanism, it's likely that the work item | |
1225 | * should be using an unbound workqueue instead. | |
1226 | * | |
1227 | * wq_cpu_intensive_report() tracks work functions which trigger such conditions | |
1228 | * and report them so that they can be examined and converted to use unbound | |
1229 | * workqueues as appropriate. To avoid flooding the console, each violating work | |
1230 | * function is tracked and reported with exponential backoff. | |
1231 | */ | |
1232 | #define WCI_MAX_ENTS 128 | |
1233 | ||
1234 | struct wci_ent { | |
1235 | work_func_t func; | |
1236 | atomic64_t cnt; | |
1237 | struct hlist_node hash_node; | |
1238 | }; | |
1239 | ||
1240 | static struct wci_ent wci_ents[WCI_MAX_ENTS]; | |
1241 | static int wci_nr_ents; | |
1242 | static DEFINE_RAW_SPINLOCK(wci_lock); | |
1243 | static DEFINE_HASHTABLE(wci_hash, ilog2(WCI_MAX_ENTS)); | |
1244 | ||
1245 | static struct wci_ent *wci_find_ent(work_func_t func) | |
1246 | { | |
1247 | struct wci_ent *ent; | |
1248 | ||
1249 | hash_for_each_possible_rcu(wci_hash, ent, hash_node, | |
1250 | (unsigned long)func) { | |
1251 | if (ent->func == func) | |
1252 | return ent; | |
1253 | } | |
1254 | return NULL; | |
1255 | } | |
1256 | ||
1257 | static void wq_cpu_intensive_report(work_func_t func) | |
1258 | { | |
1259 | struct wci_ent *ent; | |
1260 | ||
1261 | restart: | |
1262 | ent = wci_find_ent(func); | |
1263 | if (ent) { | |
1264 | u64 cnt; | |
1265 | ||
1266 | /* | |
1267 | * Start reporting from the fourth time and back off | |
1268 | * exponentially. | |
1269 | */ | |
1270 | cnt = atomic64_inc_return_relaxed(&ent->cnt); | |
1271 | if (cnt >= 4 && is_power_of_2(cnt)) | |
1272 | printk_deferred(KERN_WARNING "workqueue: %ps hogged CPU for >%luus %llu times, consider switching to WQ_UNBOUND\n", | |
1273 | ent->func, wq_cpu_intensive_thresh_us, | |
1274 | atomic64_read(&ent->cnt)); | |
1275 | return; | |
1276 | } | |
1277 | ||
1278 | /* | |
1279 | * @func is a new violation. Allocate a new entry for it. If wcn_ents[] | |
1280 | * is exhausted, something went really wrong and we probably made enough | |
1281 | * noise already. | |
1282 | */ | |
1283 | if (wci_nr_ents >= WCI_MAX_ENTS) | |
1284 | return; | |
1285 | ||
1286 | raw_spin_lock(&wci_lock); | |
1287 | ||
1288 | if (wci_nr_ents >= WCI_MAX_ENTS) { | |
1289 | raw_spin_unlock(&wci_lock); | |
1290 | return; | |
1291 | } | |
1292 | ||
1293 | if (wci_find_ent(func)) { | |
1294 | raw_spin_unlock(&wci_lock); | |
1295 | goto restart; | |
1296 | } | |
1297 | ||
1298 | ent = &wci_ents[wci_nr_ents++]; | |
1299 | ent->func = func; | |
1300 | atomic64_set(&ent->cnt, 1); | |
1301 | hash_add_rcu(wci_hash, &ent->hash_node, (unsigned long)func); | |
1302 | ||
1303 | raw_spin_unlock(&wci_lock); | |
1304 | } | |
1305 | ||
1306 | #else /* CONFIG_WQ_CPU_INTENSIVE_REPORT */ | |
1307 | static void wq_cpu_intensive_report(work_func_t func) {} | |
1308 | #endif /* CONFIG_WQ_CPU_INTENSIVE_REPORT */ | |
1309 | ||
d302f017 | 1310 | /** |
6d25be57 | 1311 | * wq_worker_running - a worker is running again |
e22bee78 | 1312 | * @task: task waking up |
e22bee78 | 1313 | * |
6d25be57 | 1314 | * This function is called when a worker returns from schedule() |
e22bee78 | 1315 | */ |
6d25be57 | 1316 | void wq_worker_running(struct task_struct *task) |
e22bee78 TH |
1317 | { |
1318 | struct worker *worker = kthread_data(task); | |
1319 | ||
c8f6219b | 1320 | if (!READ_ONCE(worker->sleeping)) |
6d25be57 | 1321 | return; |
07edfece FW |
1322 | |
1323 | /* | |
1324 | * If preempted by unbind_workers() between the WORKER_NOT_RUNNING check | |
1325 | * and the nr_running increment below, we may ruin the nr_running reset | |
1326 | * and leave with an unexpected pool->nr_running == 1 on the newly unbound | |
1327 | * pool. Protect against such race. | |
1328 | */ | |
1329 | preempt_disable(); | |
6d25be57 | 1330 | if (!(worker->flags & WORKER_NOT_RUNNING)) |
bc35f7ef | 1331 | worker->pool->nr_running++; |
07edfece | 1332 | preempt_enable(); |
616db877 TH |
1333 | |
1334 | /* | |
1335 | * CPU intensive auto-detection cares about how long a work item hogged | |
1336 | * CPU without sleeping. Reset the starting timestamp on wakeup. | |
1337 | */ | |
1338 | worker->current_at = worker->task->se.sum_exec_runtime; | |
1339 | ||
c8f6219b | 1340 | WRITE_ONCE(worker->sleeping, 0); |
e22bee78 TH |
1341 | } |
1342 | ||
1343 | /** | |
1344 | * wq_worker_sleeping - a worker is going to sleep | |
1345 | * @task: task going to sleep | |
e22bee78 | 1346 | * |
6d25be57 | 1347 | * This function is called from schedule() when a busy worker is |
ccf45156 | 1348 | * going to sleep. |
e22bee78 | 1349 | */ |
6d25be57 | 1350 | void wq_worker_sleeping(struct task_struct *task) |
e22bee78 | 1351 | { |
cc5bff38 | 1352 | struct worker *worker = kthread_data(task); |
111c225a | 1353 | struct worker_pool *pool; |
e22bee78 | 1354 | |
111c225a TH |
1355 | /* |
1356 | * Rescuers, which may not have all the fields set up like normal | |
1357 | * workers, also reach here, let's not access anything before | |
1358 | * checking NOT_RUNNING. | |
1359 | */ | |
2d64672e | 1360 | if (worker->flags & WORKER_NOT_RUNNING) |
6d25be57 | 1361 | return; |
e22bee78 | 1362 | |
111c225a | 1363 | pool = worker->pool; |
111c225a | 1364 | |
62849a96 | 1365 | /* Return if preempted before wq_worker_running() was reached */ |
c8f6219b | 1366 | if (READ_ONCE(worker->sleeping)) |
6d25be57 TG |
1367 | return; |
1368 | ||
c8f6219b | 1369 | WRITE_ONCE(worker->sleeping, 1); |
a9b8a985 | 1370 | raw_spin_lock_irq(&pool->lock); |
e22bee78 | 1371 | |
45c753f5 FW |
1372 | /* |
1373 | * Recheck in case unbind_workers() preempted us. We don't | |
1374 | * want to decrement nr_running after the worker is unbound | |
1375 | * and nr_running has been reset. | |
1376 | */ | |
1377 | if (worker->flags & WORKER_NOT_RUNNING) { | |
1378 | raw_spin_unlock_irq(&pool->lock); | |
1379 | return; | |
1380 | } | |
1381 | ||
bc35f7ef | 1382 | pool->nr_running--; |
0219a352 | 1383 | if (kick_pool(pool)) |
725e8ec5 | 1384 | worker->current_pwq->stats[PWQ_STAT_CM_WAKEUP]++; |
0219a352 | 1385 | |
a9b8a985 | 1386 | raw_spin_unlock_irq(&pool->lock); |
e22bee78 TH |
1387 | } |
1388 | ||
616db877 TH |
1389 | /** |
1390 | * wq_worker_tick - a scheduler tick occurred while a kworker is running | |
1391 | * @task: task currently running | |
1392 | * | |
1393 | * Called from scheduler_tick(). We're in the IRQ context and the current | |
1394 | * worker's fields which follow the 'K' locking rule can be accessed safely. | |
1395 | */ | |
1396 | void wq_worker_tick(struct task_struct *task) | |
1397 | { | |
1398 | struct worker *worker = kthread_data(task); | |
1399 | struct pool_workqueue *pwq = worker->current_pwq; | |
1400 | struct worker_pool *pool = worker->pool; | |
1401 | ||
1402 | if (!pwq) | |
1403 | return; | |
1404 | ||
8a1dd1e5 TH |
1405 | pwq->stats[PWQ_STAT_CPU_TIME] += TICK_USEC; |
1406 | ||
18c8ae81 Z |
1407 | if (!wq_cpu_intensive_thresh_us) |
1408 | return; | |
1409 | ||
616db877 TH |
1410 | /* |
1411 | * If the current worker is concurrency managed and hogged the CPU for | |
1412 | * longer than wq_cpu_intensive_thresh_us, it's automatically marked | |
1413 | * CPU_INTENSIVE to avoid stalling other concurrency-managed work items. | |
c8f6219b Z |
1414 | * |
1415 | * Set @worker->sleeping means that @worker is in the process of | |
1416 | * switching out voluntarily and won't be contributing to | |
1417 | * @pool->nr_running until it wakes up. As wq_worker_sleeping() also | |
1418 | * decrements ->nr_running, setting CPU_INTENSIVE here can lead to | |
1419 | * double decrements. The task is releasing the CPU anyway. Let's skip. | |
1420 | * We probably want to make this prettier in the future. | |
616db877 | 1421 | */ |
c8f6219b | 1422 | if ((worker->flags & WORKER_NOT_RUNNING) || READ_ONCE(worker->sleeping) || |
616db877 TH |
1423 | worker->task->se.sum_exec_runtime - worker->current_at < |
1424 | wq_cpu_intensive_thresh_us * NSEC_PER_USEC) | |
1425 | return; | |
1426 | ||
1427 | raw_spin_lock(&pool->lock); | |
1428 | ||
1429 | worker_set_flags(worker, WORKER_CPU_INTENSIVE); | |
63638450 | 1430 | wq_cpu_intensive_report(worker->current_func); |
616db877 TH |
1431 | pwq->stats[PWQ_STAT_CPU_INTENSIVE]++; |
1432 | ||
0219a352 | 1433 | if (kick_pool(pool)) |
616db877 | 1434 | pwq->stats[PWQ_STAT_CM_WAKEUP]++; |
616db877 TH |
1435 | |
1436 | raw_spin_unlock(&pool->lock); | |
1437 | } | |
1438 | ||
1b69ac6b JW |
1439 | /** |
1440 | * wq_worker_last_func - retrieve worker's last work function | |
8194fe94 | 1441 | * @task: Task to retrieve last work function of. |
1b69ac6b JW |
1442 | * |
1443 | * Determine the last function a worker executed. This is called from | |
1444 | * the scheduler to get a worker's last known identity. | |
1445 | * | |
1446 | * CONTEXT: | |
a9b8a985 | 1447 | * raw_spin_lock_irq(rq->lock) |
1b69ac6b | 1448 | * |
4b047002 JW |
1449 | * This function is called during schedule() when a kworker is going |
1450 | * to sleep. It's used by psi to identify aggregation workers during | |
1451 | * dequeuing, to allow periodic aggregation to shut-off when that | |
1452 | * worker is the last task in the system or cgroup to go to sleep. | |
1453 | * | |
1454 | * As this function doesn't involve any workqueue-related locking, it | |
1455 | * only returns stable values when called from inside the scheduler's | |
1456 | * queuing and dequeuing paths, when @task, which must be a kworker, | |
1457 | * is guaranteed to not be processing any works. | |
1458 | * | |
1b69ac6b JW |
1459 | * Return: |
1460 | * The last work function %current executed as a worker, NULL if it | |
1461 | * hasn't executed any work yet. | |
1462 | */ | |
1463 | work_func_t wq_worker_last_func(struct task_struct *task) | |
1464 | { | |
1465 | struct worker *worker = kthread_data(task); | |
1466 | ||
1467 | return worker->last_func; | |
1468 | } | |
1469 | ||
91ccc6e7 TH |
1470 | /** |
1471 | * wq_node_nr_active - Determine wq_node_nr_active to use | |
1472 | * @wq: workqueue of interest | |
1473 | * @node: NUMA node, can be %NUMA_NO_NODE | |
1474 | * | |
1475 | * Determine wq_node_nr_active to use for @wq on @node. Returns: | |
1476 | * | |
1477 | * - %NULL for per-cpu workqueues as they don't need to use shared nr_active. | |
1478 | * | |
1479 | * - node_nr_active[nr_node_ids] if @node is %NUMA_NO_NODE. | |
1480 | * | |
1481 | * - Otherwise, node_nr_active[@node]. | |
1482 | */ | |
1483 | static struct wq_node_nr_active *wq_node_nr_active(struct workqueue_struct *wq, | |
1484 | int node) | |
1485 | { | |
1486 | if (!(wq->flags & WQ_UNBOUND)) | |
1487 | return NULL; | |
1488 | ||
1489 | if (node == NUMA_NO_NODE) | |
1490 | node = nr_node_ids; | |
1491 | ||
1492 | return wq->node_nr_active[node]; | |
1493 | } | |
1494 | ||
5797b1c1 TH |
1495 | /** |
1496 | * wq_update_node_max_active - Update per-node max_actives to use | |
1497 | * @wq: workqueue to update | |
1498 | * @off_cpu: CPU that's going down, -1 if a CPU is not going down | |
1499 | * | |
1500 | * Update @wq->node_nr_active[]->max. @wq must be unbound. max_active is | |
1501 | * distributed among nodes according to the proportions of numbers of online | |
1502 | * cpus. The result is always between @wq->min_active and max_active. | |
1503 | */ | |
1504 | static void wq_update_node_max_active(struct workqueue_struct *wq, int off_cpu) | |
1505 | { | |
1506 | struct cpumask *effective = unbound_effective_cpumask(wq); | |
1507 | int min_active = READ_ONCE(wq->min_active); | |
1508 | int max_active = READ_ONCE(wq->max_active); | |
1509 | int total_cpus, node; | |
1510 | ||
1511 | lockdep_assert_held(&wq->mutex); | |
1512 | ||
1513 | if (!cpumask_test_cpu(off_cpu, effective)) | |
1514 | off_cpu = -1; | |
1515 | ||
1516 | total_cpus = cpumask_weight_and(effective, cpu_online_mask); | |
1517 | if (off_cpu >= 0) | |
1518 | total_cpus--; | |
1519 | ||
1520 | for_each_node(node) { | |
1521 | int node_cpus; | |
1522 | ||
1523 | node_cpus = cpumask_weight_and(effective, cpumask_of_node(node)); | |
1524 | if (off_cpu >= 0 && cpu_to_node(off_cpu) == node) | |
1525 | node_cpus--; | |
1526 | ||
1527 | wq_node_nr_active(wq, node)->max = | |
1528 | clamp(DIV_ROUND_UP(max_active * node_cpus, total_cpus), | |
1529 | min_active, max_active); | |
1530 | } | |
1531 | ||
1532 | wq_node_nr_active(wq, NUMA_NO_NODE)->max = min_active; | |
1533 | } | |
1534 | ||
8864b4e5 TH |
1535 | /** |
1536 | * get_pwq - get an extra reference on the specified pool_workqueue | |
1537 | * @pwq: pool_workqueue to get | |
1538 | * | |
1539 | * Obtain an extra reference on @pwq. The caller should guarantee that | |
1540 | * @pwq has positive refcnt and be holding the matching pool->lock. | |
1541 | */ | |
1542 | static void get_pwq(struct pool_workqueue *pwq) | |
1543 | { | |
1544 | lockdep_assert_held(&pwq->pool->lock); | |
1545 | WARN_ON_ONCE(pwq->refcnt <= 0); | |
1546 | pwq->refcnt++; | |
1547 | } | |
1548 | ||
1549 | /** | |
1550 | * put_pwq - put a pool_workqueue reference | |
1551 | * @pwq: pool_workqueue to put | |
1552 | * | |
1553 | * Drop a reference of @pwq. If its refcnt reaches zero, schedule its | |
1554 | * destruction. The caller should be holding the matching pool->lock. | |
1555 | */ | |
1556 | static void put_pwq(struct pool_workqueue *pwq) | |
1557 | { | |
1558 | lockdep_assert_held(&pwq->pool->lock); | |
1559 | if (likely(--pwq->refcnt)) | |
1560 | return; | |
8864b4e5 | 1561 | /* |
967b494e TH |
1562 | * @pwq can't be released under pool->lock, bounce to a dedicated |
1563 | * kthread_worker to avoid A-A deadlocks. | |
8864b4e5 | 1564 | */ |
687a9aa5 | 1565 | kthread_queue_work(pwq_release_worker, &pwq->release_work); |
8864b4e5 TH |
1566 | } |
1567 | ||
dce90d47 TH |
1568 | /** |
1569 | * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock | |
1570 | * @pwq: pool_workqueue to put (can be %NULL) | |
1571 | * | |
1572 | * put_pwq() with locking. This function also allows %NULL @pwq. | |
1573 | */ | |
1574 | static void put_pwq_unlocked(struct pool_workqueue *pwq) | |
1575 | { | |
1576 | if (pwq) { | |
1577 | /* | |
24acfb71 | 1578 | * As both pwqs and pools are RCU protected, the |
dce90d47 TH |
1579 | * following lock operations are safe. |
1580 | */ | |
a9b8a985 | 1581 | raw_spin_lock_irq(&pwq->pool->lock); |
dce90d47 | 1582 | put_pwq(pwq); |
a9b8a985 | 1583 | raw_spin_unlock_irq(&pwq->pool->lock); |
dce90d47 TH |
1584 | } |
1585 | } | |
1586 | ||
afa87ce8 TH |
1587 | static bool pwq_is_empty(struct pool_workqueue *pwq) |
1588 | { | |
1589 | return !pwq->nr_active && list_empty(&pwq->inactive_works); | |
1590 | } | |
1591 | ||
4c638030 TH |
1592 | static void __pwq_activate_work(struct pool_workqueue *pwq, |
1593 | struct work_struct *work) | |
bf4ede01 | 1594 | { |
1c270b79 TH |
1595 | unsigned long *wdb = work_data_bits(work); |
1596 | ||
1597 | WARN_ON_ONCE(!(*wdb & WORK_STRUCT_INACTIVE)); | |
bf4ede01 | 1598 | trace_workqueue_activate_work(work); |
82607adc TH |
1599 | if (list_empty(&pwq->pool->worklist)) |
1600 | pwq->pool->watchdog_ts = jiffies; | |
112202d9 | 1601 | move_linked_works(work, &pwq->pool->worklist, NULL); |
1c270b79 | 1602 | __clear_bit(WORK_STRUCT_INACTIVE_BIT, wdb); |
4c638030 TH |
1603 | } |
1604 | ||
1605 | /** | |
1606 | * pwq_activate_work - Activate a work item if inactive | |
1607 | * @pwq: pool_workqueue @work belongs to | |
1608 | * @work: work item to activate | |
1609 | * | |
1610 | * Returns %true if activated. %false if already active. | |
1611 | */ | |
1612 | static bool pwq_activate_work(struct pool_workqueue *pwq, | |
1613 | struct work_struct *work) | |
1614 | { | |
1615 | struct worker_pool *pool = pwq->pool; | |
91ccc6e7 | 1616 | struct wq_node_nr_active *nna; |
4c638030 TH |
1617 | |
1618 | lockdep_assert_held(&pool->lock); | |
1619 | ||
1620 | if (!(*work_data_bits(work) & WORK_STRUCT_INACTIVE)) | |
1621 | return false; | |
1622 | ||
91ccc6e7 TH |
1623 | nna = wq_node_nr_active(pwq->wq, pool->node); |
1624 | if (nna) | |
1625 | atomic_inc(&nna->nr); | |
1626 | ||
112202d9 | 1627 | pwq->nr_active++; |
4c638030 TH |
1628 | __pwq_activate_work(pwq, work); |
1629 | return true; | |
bf4ede01 TH |
1630 | } |
1631 | ||
5797b1c1 TH |
1632 | static bool tryinc_node_nr_active(struct wq_node_nr_active *nna) |
1633 | { | |
1634 | int max = READ_ONCE(nna->max); | |
1635 | ||
1636 | while (true) { | |
1637 | int old, tmp; | |
1638 | ||
1639 | old = atomic_read(&nna->nr); | |
1640 | if (old >= max) | |
1641 | return false; | |
1642 | tmp = atomic_cmpxchg_relaxed(&nna->nr, old, old + 1); | |
1643 | if (tmp == old) | |
1644 | return true; | |
1645 | } | |
1646 | } | |
1647 | ||
1c270b79 TH |
1648 | /** |
1649 | * pwq_tryinc_nr_active - Try to increment nr_active for a pwq | |
1650 | * @pwq: pool_workqueue of interest | |
5797b1c1 | 1651 | * @fill: max_active may have increased, try to increase concurrency level |
1c270b79 TH |
1652 | * |
1653 | * Try to increment nr_active for @pwq. Returns %true if an nr_active count is | |
1654 | * successfully obtained. %false otherwise. | |
1655 | */ | |
5797b1c1 | 1656 | static bool pwq_tryinc_nr_active(struct pool_workqueue *pwq, bool fill) |
1c270b79 TH |
1657 | { |
1658 | struct workqueue_struct *wq = pwq->wq; | |
1659 | struct worker_pool *pool = pwq->pool; | |
91ccc6e7 | 1660 | struct wq_node_nr_active *nna = wq_node_nr_active(wq, pool->node); |
5797b1c1 | 1661 | bool obtained = false; |
1c270b79 TH |
1662 | |
1663 | lockdep_assert_held(&pool->lock); | |
1664 | ||
5797b1c1 TH |
1665 | if (!nna) { |
1666 | /* per-cpu workqueue, pwq->nr_active is sufficient */ | |
1667 | obtained = pwq->nr_active < READ_ONCE(wq->max_active); | |
1668 | goto out; | |
1669 | } | |
1670 | ||
1671 | /* | |
1672 | * Unbound workqueue uses per-node shared nr_active $nna. If @pwq is | |
1673 | * already waiting on $nna, pwq_dec_nr_active() will maintain the | |
1674 | * concurrency level. Don't jump the line. | |
1675 | * | |
1676 | * We need to ignore the pending test after max_active has increased as | |
1677 | * pwq_dec_nr_active() can only maintain the concurrency level but not | |
1678 | * increase it. This is indicated by @fill. | |
1679 | */ | |
1680 | if (!list_empty(&pwq->pending_node) && likely(!fill)) | |
1681 | goto out; | |
1682 | ||
1683 | obtained = tryinc_node_nr_active(nna); | |
1684 | if (obtained) | |
1685 | goto out; | |
1686 | ||
1687 | /* | |
1688 | * Lockless acquisition failed. Lock, add ourself to $nna->pending_pwqs | |
1689 | * and try again. The smp_mb() is paired with the implied memory barrier | |
1690 | * of atomic_dec_return() in pwq_dec_nr_active() to ensure that either | |
1691 | * we see the decremented $nna->nr or they see non-empty | |
1692 | * $nna->pending_pwqs. | |
1693 | */ | |
1694 | raw_spin_lock(&nna->lock); | |
1695 | ||
1696 | if (list_empty(&pwq->pending_node)) | |
1697 | list_add_tail(&pwq->pending_node, &nna->pending_pwqs); | |
1698 | else if (likely(!fill)) | |
1699 | goto out_unlock; | |
1700 | ||
1701 | smp_mb(); | |
1702 | ||
1703 | obtained = tryinc_node_nr_active(nna); | |
1c270b79 | 1704 | |
5797b1c1 TH |
1705 | /* |
1706 | * If @fill, @pwq might have already been pending. Being spuriously | |
1707 | * pending in cold paths doesn't affect anything. Let's leave it be. | |
1708 | */ | |
1709 | if (obtained && likely(!fill)) | |
1710 | list_del_init(&pwq->pending_node); | |
1711 | ||
1712 | out_unlock: | |
1713 | raw_spin_unlock(&nna->lock); | |
1714 | out: | |
1715 | if (obtained) | |
1c270b79 TH |
1716 | pwq->nr_active++; |
1717 | return obtained; | |
1718 | } | |
1719 | ||
1720 | /** | |
1721 | * pwq_activate_first_inactive - Activate the first inactive work item on a pwq | |
1722 | * @pwq: pool_workqueue of interest | |
5797b1c1 | 1723 | * @fill: max_active may have increased, try to increase concurrency level |
1c270b79 TH |
1724 | * |
1725 | * Activate the first inactive work item of @pwq if available and allowed by | |
1726 | * max_active limit. | |
1727 | * | |
1728 | * Returns %true if an inactive work item has been activated. %false if no | |
1729 | * inactive work item is found or max_active limit is reached. | |
1730 | */ | |
5797b1c1 | 1731 | static bool pwq_activate_first_inactive(struct pool_workqueue *pwq, bool fill) |
1c270b79 TH |
1732 | { |
1733 | struct work_struct *work = | |
1734 | list_first_entry_or_null(&pwq->inactive_works, | |
1735 | struct work_struct, entry); | |
1736 | ||
5797b1c1 | 1737 | if (work && pwq_tryinc_nr_active(pwq, fill)) { |
1c270b79 TH |
1738 | __pwq_activate_work(pwq, work); |
1739 | return true; | |
1740 | } else { | |
1741 | return false; | |
1742 | } | |
1743 | } | |
1744 | ||
5797b1c1 TH |
1745 | /** |
1746 | * node_activate_pending_pwq - Activate a pending pwq on a wq_node_nr_active | |
1747 | * @nna: wq_node_nr_active to activate a pending pwq for | |
1748 | * @caller_pool: worker_pool the caller is locking | |
1749 | * | |
1750 | * Activate a pwq in @nna->pending_pwqs. Called with @caller_pool locked. | |
1751 | * @caller_pool may be unlocked and relocked to lock other worker_pools. | |
1752 | */ | |
1753 | static void node_activate_pending_pwq(struct wq_node_nr_active *nna, | |
1754 | struct worker_pool *caller_pool) | |
1755 | { | |
1756 | struct worker_pool *locked_pool = caller_pool; | |
1757 | struct pool_workqueue *pwq; | |
1758 | struct work_struct *work; | |
1759 | ||
1760 | lockdep_assert_held(&caller_pool->lock); | |
1761 | ||
1762 | raw_spin_lock(&nna->lock); | |
1763 | retry: | |
1764 | pwq = list_first_entry_or_null(&nna->pending_pwqs, | |
1765 | struct pool_workqueue, pending_node); | |
1766 | if (!pwq) | |
1767 | goto out_unlock; | |
1768 | ||
1769 | /* | |
1770 | * If @pwq is for a different pool than @locked_pool, we need to lock | |
1771 | * @pwq->pool->lock. Let's trylock first. If unsuccessful, do the unlock | |
1772 | * / lock dance. For that, we also need to release @nna->lock as it's | |
1773 | * nested inside pool locks. | |
1774 | */ | |
1775 | if (pwq->pool != locked_pool) { | |
1776 | raw_spin_unlock(&locked_pool->lock); | |
1777 | locked_pool = pwq->pool; | |
1778 | if (!raw_spin_trylock(&locked_pool->lock)) { | |
1779 | raw_spin_unlock(&nna->lock); | |
1780 | raw_spin_lock(&locked_pool->lock); | |
1781 | raw_spin_lock(&nna->lock); | |
1782 | goto retry; | |
1783 | } | |
1784 | } | |
1785 | ||
1786 | /* | |
1787 | * $pwq may not have any inactive work items due to e.g. cancellations. | |
1788 | * Drop it from pending_pwqs and see if there's another one. | |
1789 | */ | |
1790 | work = list_first_entry_or_null(&pwq->inactive_works, | |
1791 | struct work_struct, entry); | |
1792 | if (!work) { | |
1793 | list_del_init(&pwq->pending_node); | |
1794 | goto retry; | |
1795 | } | |
1796 | ||
1797 | /* | |
1798 | * Acquire an nr_active count and activate the inactive work item. If | |
1799 | * $pwq still has inactive work items, rotate it to the end of the | |
1800 | * pending_pwqs so that we round-robin through them. This means that | |
1801 | * inactive work items are not activated in queueing order which is fine | |
1802 | * given that there has never been any ordering across different pwqs. | |
1803 | */ | |
1804 | if (likely(tryinc_node_nr_active(nna))) { | |
1805 | pwq->nr_active++; | |
1806 | __pwq_activate_work(pwq, work); | |
1807 | ||
1808 | if (list_empty(&pwq->inactive_works)) | |
1809 | list_del_init(&pwq->pending_node); | |
1810 | else | |
1811 | list_move_tail(&pwq->pending_node, &nna->pending_pwqs); | |
1812 | ||
1813 | /* if activating a foreign pool, make sure it's running */ | |
1814 | if (pwq->pool != caller_pool) | |
1815 | kick_pool(pwq->pool); | |
1816 | } | |
1817 | ||
1818 | out_unlock: | |
1819 | raw_spin_unlock(&nna->lock); | |
1820 | if (locked_pool != caller_pool) { | |
1821 | raw_spin_unlock(&locked_pool->lock); | |
1822 | raw_spin_lock(&caller_pool->lock); | |
1823 | } | |
1824 | } | |
1825 | ||
1c270b79 TH |
1826 | /** |
1827 | * pwq_dec_nr_active - Retire an active count | |
1828 | * @pwq: pool_workqueue of interest | |
1829 | * | |
1830 | * Decrement @pwq's nr_active and try to activate the first inactive work item. | |
5797b1c1 | 1831 | * For unbound workqueues, this function may temporarily drop @pwq->pool->lock. |
1c270b79 TH |
1832 | */ |
1833 | static void pwq_dec_nr_active(struct pool_workqueue *pwq) | |
3aa62497 | 1834 | { |
1c270b79 | 1835 | struct worker_pool *pool = pwq->pool; |
91ccc6e7 | 1836 | struct wq_node_nr_active *nna = wq_node_nr_active(pwq->wq, pool->node); |
3aa62497 | 1837 | |
1c270b79 TH |
1838 | lockdep_assert_held(&pool->lock); |
1839 | ||
91ccc6e7 TH |
1840 | /* |
1841 | * @pwq->nr_active should be decremented for both percpu and unbound | |
1842 | * workqueues. | |
1843 | */ | |
1c270b79 | 1844 | pwq->nr_active--; |
91ccc6e7 TH |
1845 | |
1846 | /* | |
1847 | * For a percpu workqueue, it's simple. Just need to kick the first | |
1848 | * inactive work item on @pwq itself. | |
1849 | */ | |
1850 | if (!nna) { | |
5797b1c1 | 1851 | pwq_activate_first_inactive(pwq, false); |
91ccc6e7 TH |
1852 | return; |
1853 | } | |
1854 | ||
5797b1c1 TH |
1855 | /* |
1856 | * If @pwq is for an unbound workqueue, it's more complicated because | |
1857 | * multiple pwqs and pools may be sharing the nr_active count. When a | |
1858 | * pwq needs to wait for an nr_active count, it puts itself on | |
1859 | * $nna->pending_pwqs. The following atomic_dec_return()'s implied | |
1860 | * memory barrier is paired with smp_mb() in pwq_tryinc_nr_active() to | |
1861 | * guarantee that either we see non-empty pending_pwqs or they see | |
1862 | * decremented $nna->nr. | |
1863 | * | |
1864 | * $nna->max may change as CPUs come online/offline and @pwq->wq's | |
1865 | * max_active gets updated. However, it is guaranteed to be equal to or | |
1866 | * larger than @pwq->wq->min_active which is above zero unless freezing. | |
1867 | * This maintains the forward progress guarantee. | |
1868 | */ | |
1869 | if (atomic_dec_return(&nna->nr) >= READ_ONCE(nna->max)) | |
1870 | return; | |
1871 | ||
1872 | if (!list_empty(&nna->pending_pwqs)) | |
1873 | node_activate_pending_pwq(nna, pool); | |
3aa62497 LJ |
1874 | } |
1875 | ||
bf4ede01 | 1876 | /** |
112202d9 TH |
1877 | * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight |
1878 | * @pwq: pwq of interest | |
c4560c2c | 1879 | * @work_data: work_data of work which left the queue |
bf4ede01 TH |
1880 | * |
1881 | * A work either has completed or is removed from pending queue, | |
112202d9 | 1882 | * decrement nr_in_flight of its pwq and handle workqueue flushing. |
bf4ede01 | 1883 | * |
dd6c3c54 TH |
1884 | * NOTE: |
1885 | * For unbound workqueues, this function may temporarily drop @pwq->pool->lock | |
1886 | * and thus should be called after all other state updates for the in-flight | |
1887 | * work item is complete. | |
1888 | * | |
bf4ede01 | 1889 | * CONTEXT: |
a9b8a985 | 1890 | * raw_spin_lock_irq(pool->lock). |
bf4ede01 | 1891 | */ |
c4560c2c | 1892 | static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, unsigned long work_data) |
bf4ede01 | 1893 | { |
c4560c2c LJ |
1894 | int color = get_work_color(work_data); |
1895 | ||
1c270b79 TH |
1896 | if (!(work_data & WORK_STRUCT_INACTIVE)) |
1897 | pwq_dec_nr_active(pwq); | |
018f3a13 | 1898 | |
112202d9 | 1899 | pwq->nr_in_flight[color]--; |
bf4ede01 | 1900 | |
bf4ede01 | 1901 | /* is flush in progress and are we at the flushing tip? */ |
112202d9 | 1902 | if (likely(pwq->flush_color != color)) |
8864b4e5 | 1903 | goto out_put; |
bf4ede01 TH |
1904 | |
1905 | /* are there still in-flight works? */ | |
112202d9 | 1906 | if (pwq->nr_in_flight[color]) |
8864b4e5 | 1907 | goto out_put; |
bf4ede01 | 1908 | |
112202d9 TH |
1909 | /* this pwq is done, clear flush_color */ |
1910 | pwq->flush_color = -1; | |
bf4ede01 TH |
1911 | |
1912 | /* | |
112202d9 | 1913 | * If this was the last pwq, wake up the first flusher. It |
bf4ede01 TH |
1914 | * will handle the rest. |
1915 | */ | |
112202d9 TH |
1916 | if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) |
1917 | complete(&pwq->wq->first_flusher->done); | |
8864b4e5 TH |
1918 | out_put: |
1919 | put_pwq(pwq); | |
bf4ede01 TH |
1920 | } |
1921 | ||
36e227d2 | 1922 | /** |
bbb68dfa | 1923 | * try_to_grab_pending - steal work item from worklist and disable irq |
36e227d2 TH |
1924 | * @work: work item to steal |
1925 | * @is_dwork: @work is a delayed_work | |
bbb68dfa | 1926 | * @flags: place to store irq state |
36e227d2 TH |
1927 | * |
1928 | * Try to grab PENDING bit of @work. This function can handle @work in any | |
d185af30 | 1929 | * stable state - idle, on timer or on worklist. |
36e227d2 | 1930 | * |
d185af30 | 1931 | * Return: |
3eb6b31b MCC |
1932 | * |
1933 | * ======== ================================================================ | |
36e227d2 TH |
1934 | * 1 if @work was pending and we successfully stole PENDING |
1935 | * 0 if @work was idle and we claimed PENDING | |
1936 | * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry | |
bbb68dfa TH |
1937 | * -ENOENT if someone else is canceling @work, this state may persist |
1938 | * for arbitrarily long | |
3eb6b31b | 1939 | * ======== ================================================================ |
36e227d2 | 1940 | * |
d185af30 | 1941 | * Note: |
bbb68dfa | 1942 | * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting |
e0aecdd8 TH |
1943 | * interrupted while holding PENDING and @work off queue, irq must be |
1944 | * disabled on entry. This, combined with delayed_work->timer being | |
1945 | * irqsafe, ensures that we return -EAGAIN for finite short period of time. | |
bbb68dfa TH |
1946 | * |
1947 | * On successful return, >= 0, irq is disabled and the caller is | |
1948 | * responsible for releasing it using local_irq_restore(*@flags). | |
1949 | * | |
e0aecdd8 | 1950 | * This function is safe to call from any context including IRQ handler. |
bf4ede01 | 1951 | */ |
bbb68dfa TH |
1952 | static int try_to_grab_pending(struct work_struct *work, bool is_dwork, |
1953 | unsigned long *flags) | |
bf4ede01 | 1954 | { |
d565ed63 | 1955 | struct worker_pool *pool; |
112202d9 | 1956 | struct pool_workqueue *pwq; |
bf4ede01 | 1957 | |
bbb68dfa TH |
1958 | local_irq_save(*flags); |
1959 | ||
36e227d2 TH |
1960 | /* try to steal the timer if it exists */ |
1961 | if (is_dwork) { | |
1962 | struct delayed_work *dwork = to_delayed_work(work); | |
1963 | ||
e0aecdd8 TH |
1964 | /* |
1965 | * dwork->timer is irqsafe. If del_timer() fails, it's | |
1966 | * guaranteed that the timer is not queued anywhere and not | |
1967 | * running on the local CPU. | |
1968 | */ | |
36e227d2 TH |
1969 | if (likely(del_timer(&dwork->timer))) |
1970 | return 1; | |
1971 | } | |
1972 | ||
1973 | /* try to claim PENDING the normal way */ | |
bf4ede01 TH |
1974 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) |
1975 | return 0; | |
1976 | ||
24acfb71 | 1977 | rcu_read_lock(); |
bf4ede01 TH |
1978 | /* |
1979 | * The queueing is in progress, or it is already queued. Try to | |
1980 | * steal it from ->worklist without clearing WORK_STRUCT_PENDING. | |
1981 | */ | |
d565ed63 TH |
1982 | pool = get_work_pool(work); |
1983 | if (!pool) | |
bbb68dfa | 1984 | goto fail; |
bf4ede01 | 1985 | |
a9b8a985 | 1986 | raw_spin_lock(&pool->lock); |
0b3dae68 | 1987 | /* |
112202d9 TH |
1988 | * work->data is guaranteed to point to pwq only while the work |
1989 | * item is queued on pwq->wq, and both updating work->data to point | |
1990 | * to pwq on queueing and to pool on dequeueing are done under | |
1991 | * pwq->pool->lock. This in turn guarantees that, if work->data | |
1992 | * points to pwq which is associated with a locked pool, the work | |
0b3dae68 LJ |
1993 | * item is currently queued on that pool. |
1994 | */ | |
112202d9 TH |
1995 | pwq = get_work_pwq(work); |
1996 | if (pwq && pwq->pool == pool) { | |
16062836 TH |
1997 | debug_work_deactivate(work); |
1998 | ||
1999 | /* | |
018f3a13 LJ |
2000 | * A cancelable inactive work item must be in the |
2001 | * pwq->inactive_works since a queued barrier can't be | |
2002 | * canceled (see the comments in insert_wq_barrier()). | |
2003 | * | |
f97a4a1a | 2004 | * An inactive work item cannot be grabbed directly because |
d812796e | 2005 | * it might have linked barrier work items which, if left |
f97a4a1a | 2006 | * on the inactive_works list, will confuse pwq->nr_active |
16062836 TH |
2007 | * management later on and cause stall. Make sure the work |
2008 | * item is activated before grabbing. | |
2009 | */ | |
4c638030 | 2010 | pwq_activate_work(pwq, work); |
16062836 TH |
2011 | |
2012 | list_del_init(&work->entry); | |
16062836 | 2013 | |
112202d9 | 2014 | /* work->data points to pwq iff queued, point to pool */ |
16062836 TH |
2015 | set_work_pool_and_keep_pending(work, pool->id); |
2016 | ||
dd6c3c54 TH |
2017 | /* must be the last step, see the function comment */ |
2018 | pwq_dec_nr_in_flight(pwq, *work_data_bits(work)); | |
2019 | ||
a9b8a985 | 2020 | raw_spin_unlock(&pool->lock); |
24acfb71 | 2021 | rcu_read_unlock(); |
16062836 | 2022 | return 1; |
bf4ede01 | 2023 | } |
a9b8a985 | 2024 | raw_spin_unlock(&pool->lock); |
bbb68dfa | 2025 | fail: |
24acfb71 | 2026 | rcu_read_unlock(); |
bbb68dfa TH |
2027 | local_irq_restore(*flags); |
2028 | if (work_is_canceling(work)) | |
2029 | return -ENOENT; | |
2030 | cpu_relax(); | |
36e227d2 | 2031 | return -EAGAIN; |
bf4ede01 TH |
2032 | } |
2033 | ||
4690c4ab | 2034 | /** |
706026c2 | 2035 | * insert_work - insert a work into a pool |
112202d9 | 2036 | * @pwq: pwq @work belongs to |
4690c4ab TH |
2037 | * @work: work to insert |
2038 | * @head: insertion point | |
2039 | * @extra_flags: extra WORK_STRUCT_* flags to set | |
2040 | * | |
112202d9 | 2041 | * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to |
706026c2 | 2042 | * work_struct flags. |
4690c4ab TH |
2043 | * |
2044 | * CONTEXT: | |
a9b8a985 | 2045 | * raw_spin_lock_irq(pool->lock). |
4690c4ab | 2046 | */ |
112202d9 TH |
2047 | static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, |
2048 | struct list_head *head, unsigned int extra_flags) | |
b89deed3 | 2049 | { |
fe089f87 | 2050 | debug_work_activate(work); |
e22bee78 | 2051 | |
e89a85d6 | 2052 | /* record the work call stack in order to print it in KASAN reports */ |
f70da745 | 2053 | kasan_record_aux_stack_noalloc(work); |
e89a85d6 | 2054 | |
4690c4ab | 2055 | /* we own @work, set data and link */ |
112202d9 | 2056 | set_work_pwq(work, pwq, extra_flags); |
1a4d9b0a | 2057 | list_add_tail(&work->entry, head); |
8864b4e5 | 2058 | get_pwq(pwq); |
b89deed3 ON |
2059 | } |
2060 | ||
c8efcc25 TH |
2061 | /* |
2062 | * Test whether @work is being queued from another work executing on the | |
8d03ecfe | 2063 | * same workqueue. |
c8efcc25 TH |
2064 | */ |
2065 | static bool is_chained_work(struct workqueue_struct *wq) | |
2066 | { | |
8d03ecfe TH |
2067 | struct worker *worker; |
2068 | ||
2069 | worker = current_wq_worker(); | |
2070 | /* | |
bf393fd4 | 2071 | * Return %true iff I'm a worker executing a work item on @wq. If |
8d03ecfe TH |
2072 | * I'm @worker, it's safe to dereference it without locking. |
2073 | */ | |
112202d9 | 2074 | return worker && worker->current_pwq->wq == wq; |
c8efcc25 TH |
2075 | } |
2076 | ||
ef557180 MG |
2077 | /* |
2078 | * When queueing an unbound work item to a wq, prefer local CPU if allowed | |
2079 | * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to | |
2080 | * avoid perturbing sensitive tasks. | |
2081 | */ | |
2082 | static int wq_select_unbound_cpu(int cpu) | |
2083 | { | |
2084 | int new_cpu; | |
2085 | ||
f303fccb TH |
2086 | if (likely(!wq_debug_force_rr_cpu)) { |
2087 | if (cpumask_test_cpu(cpu, wq_unbound_cpumask)) | |
2088 | return cpu; | |
a8ec5880 AF |
2089 | } else { |
2090 | pr_warn_once("workqueue: round-robin CPU selection forced, expect performance impact\n"); | |
f303fccb TH |
2091 | } |
2092 | ||
ef557180 MG |
2093 | new_cpu = __this_cpu_read(wq_rr_cpu_last); |
2094 | new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask); | |
2095 | if (unlikely(new_cpu >= nr_cpu_ids)) { | |
2096 | new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask); | |
2097 | if (unlikely(new_cpu >= nr_cpu_ids)) | |
2098 | return cpu; | |
2099 | } | |
2100 | __this_cpu_write(wq_rr_cpu_last, new_cpu); | |
2101 | ||
2102 | return new_cpu; | |
2103 | } | |
2104 | ||
d84ff051 | 2105 | static void __queue_work(int cpu, struct workqueue_struct *wq, |
1da177e4 LT |
2106 | struct work_struct *work) |
2107 | { | |
112202d9 | 2108 | struct pool_workqueue *pwq; |
fe089f87 | 2109 | struct worker_pool *last_pool, *pool; |
8a2e8e5d | 2110 | unsigned int work_flags; |
b75cac93 | 2111 | unsigned int req_cpu = cpu; |
8930caba TH |
2112 | |
2113 | /* | |
2114 | * While a work item is PENDING && off queue, a task trying to | |
2115 | * steal the PENDING will busy-loop waiting for it to either get | |
2116 | * queued or lose PENDING. Grabbing PENDING and queueing should | |
2117 | * happen with IRQ disabled. | |
2118 | */ | |
8e8eb730 | 2119 | lockdep_assert_irqs_disabled(); |
1da177e4 | 2120 | |
1e19ffc6 | 2121 | |
33e3f0a3 RC |
2122 | /* |
2123 | * For a draining wq, only works from the same workqueue are | |
2124 | * allowed. The __WQ_DESTROYING helps to spot the issue that | |
2125 | * queues a new work item to a wq after destroy_workqueue(wq). | |
2126 | */ | |
2127 | if (unlikely(wq->flags & (__WQ_DESTROYING | __WQ_DRAINING) && | |
2128 | WARN_ON_ONCE(!is_chained_work(wq)))) | |
e41e704b | 2129 | return; |
24acfb71 | 2130 | rcu_read_lock(); |
9e8cd2f5 | 2131 | retry: |
c9178087 | 2132 | /* pwq which will be used unless @work is executing elsewhere */ |
636b927e TH |
2133 | if (req_cpu == WORK_CPU_UNBOUND) { |
2134 | if (wq->flags & WQ_UNBOUND) | |
aa202f1f | 2135 | cpu = wq_select_unbound_cpu(raw_smp_processor_id()); |
636b927e | 2136 | else |
aa202f1f | 2137 | cpu = raw_smp_processor_id(); |
aa202f1f | 2138 | } |
dbf2576e | 2139 | |
636b927e | 2140 | pwq = rcu_dereference(*per_cpu_ptr(wq->cpu_pwq, cpu)); |
fe089f87 TH |
2141 | pool = pwq->pool; |
2142 | ||
c9178087 TH |
2143 | /* |
2144 | * If @work was previously on a different pool, it might still be | |
2145 | * running there, in which case the work needs to be queued on that | |
2146 | * pool to guarantee non-reentrancy. | |
2147 | */ | |
2148 | last_pool = get_work_pool(work); | |
fe089f87 | 2149 | if (last_pool && last_pool != pool) { |
c9178087 | 2150 | struct worker *worker; |
18aa9eff | 2151 | |
a9b8a985 | 2152 | raw_spin_lock(&last_pool->lock); |
18aa9eff | 2153 | |
c9178087 | 2154 | worker = find_worker_executing_work(last_pool, work); |
18aa9eff | 2155 | |
c9178087 TH |
2156 | if (worker && worker->current_pwq->wq == wq) { |
2157 | pwq = worker->current_pwq; | |
fe089f87 TH |
2158 | pool = pwq->pool; |
2159 | WARN_ON_ONCE(pool != last_pool); | |
8930caba | 2160 | } else { |
c9178087 | 2161 | /* meh... not running there, queue here */ |
a9b8a985 | 2162 | raw_spin_unlock(&last_pool->lock); |
fe089f87 | 2163 | raw_spin_lock(&pool->lock); |
8930caba | 2164 | } |
f3421797 | 2165 | } else { |
fe089f87 | 2166 | raw_spin_lock(&pool->lock); |
502ca9d8 TH |
2167 | } |
2168 | ||
9e8cd2f5 | 2169 | /* |
636b927e TH |
2170 | * pwq is determined and locked. For unbound pools, we could have raced |
2171 | * with pwq release and it could already be dead. If its refcnt is zero, | |
2172 | * repeat pwq selection. Note that unbound pwqs never die without | |
2173 | * another pwq replacing it in cpu_pwq or while work items are executing | |
2174 | * on it, so the retrying is guaranteed to make forward-progress. | |
9e8cd2f5 TH |
2175 | */ |
2176 | if (unlikely(!pwq->refcnt)) { | |
2177 | if (wq->flags & WQ_UNBOUND) { | |
fe089f87 | 2178 | raw_spin_unlock(&pool->lock); |
9e8cd2f5 TH |
2179 | cpu_relax(); |
2180 | goto retry; | |
2181 | } | |
2182 | /* oops */ | |
2183 | WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", | |
2184 | wq->name, cpu); | |
2185 | } | |
2186 | ||
112202d9 TH |
2187 | /* pwq determined, queue */ |
2188 | trace_workqueue_queue_work(req_cpu, pwq, work); | |
502ca9d8 | 2189 | |
24acfb71 TG |
2190 | if (WARN_ON(!list_empty(&work->entry))) |
2191 | goto out; | |
1e19ffc6 | 2192 | |
112202d9 TH |
2193 | pwq->nr_in_flight[pwq->work_color]++; |
2194 | work_flags = work_color_to_flags(pwq->work_color); | |
1e19ffc6 | 2195 | |
a045a272 TH |
2196 | /* |
2197 | * Limit the number of concurrently active work items to max_active. | |
2198 | * @work must also queue behind existing inactive work items to maintain | |
2199 | * ordering when max_active changes. See wq_adjust_max_active(). | |
2200 | */ | |
5797b1c1 | 2201 | if (list_empty(&pwq->inactive_works) && pwq_tryinc_nr_active(pwq, false)) { |
fe089f87 TH |
2202 | if (list_empty(&pool->worklist)) |
2203 | pool->watchdog_ts = jiffies; | |
2204 | ||
cdadf009 | 2205 | trace_workqueue_activate_work(work); |
fe089f87 | 2206 | insert_work(pwq, work, &pool->worklist, work_flags); |
0219a352 | 2207 | kick_pool(pool); |
8a2e8e5d | 2208 | } else { |
f97a4a1a | 2209 | work_flags |= WORK_STRUCT_INACTIVE; |
fe089f87 | 2210 | insert_work(pwq, work, &pwq->inactive_works, work_flags); |
8a2e8e5d | 2211 | } |
1e19ffc6 | 2212 | |
24acfb71 | 2213 | out: |
fe089f87 | 2214 | raw_spin_unlock(&pool->lock); |
24acfb71 | 2215 | rcu_read_unlock(); |
1da177e4 LT |
2216 | } |
2217 | ||
0fcb78c2 | 2218 | /** |
c1a220e7 ZR |
2219 | * queue_work_on - queue work on specific cpu |
2220 | * @cpu: CPU number to execute work on | |
0fcb78c2 REB |
2221 | * @wq: workqueue to use |
2222 | * @work: work to queue | |
2223 | * | |
c1a220e7 | 2224 | * We queue the work to a specific CPU, the caller must ensure it |
443378f0 PM |
2225 | * can't go away. Callers that fail to ensure that the specified |
2226 | * CPU cannot go away will execute on a randomly chosen CPU. | |
854f5cc5 PM |
2227 | * But note well that callers specifying a CPU that never has been |
2228 | * online will get a splat. | |
d185af30 YB |
2229 | * |
2230 | * Return: %false if @work was already on a queue, %true otherwise. | |
1da177e4 | 2231 | */ |
d4283e93 TH |
2232 | bool queue_work_on(int cpu, struct workqueue_struct *wq, |
2233 | struct work_struct *work) | |
1da177e4 | 2234 | { |
d4283e93 | 2235 | bool ret = false; |
8930caba | 2236 | unsigned long flags; |
ef1ca236 | 2237 | |
8930caba | 2238 | local_irq_save(flags); |
c1a220e7 | 2239 | |
22df02bb | 2240 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { |
4690c4ab | 2241 | __queue_work(cpu, wq, work); |
d4283e93 | 2242 | ret = true; |
c1a220e7 | 2243 | } |
ef1ca236 | 2244 | |
8930caba | 2245 | local_irq_restore(flags); |
1da177e4 LT |
2246 | return ret; |
2247 | } | |
ad7b1f84 | 2248 | EXPORT_SYMBOL(queue_work_on); |
1da177e4 | 2249 | |
8204e0c1 | 2250 | /** |
fef59c9c | 2251 | * select_numa_node_cpu - Select a CPU based on NUMA node |
8204e0c1 AD |
2252 | * @node: NUMA node ID that we want to select a CPU from |
2253 | * | |
2254 | * This function will attempt to find a "random" cpu available on a given | |
2255 | * node. If there are no CPUs available on the given node it will return | |
2256 | * WORK_CPU_UNBOUND indicating that we should just schedule to any | |
2257 | * available CPU if we need to schedule this work. | |
2258 | */ | |
fef59c9c | 2259 | static int select_numa_node_cpu(int node) |
8204e0c1 AD |
2260 | { |
2261 | int cpu; | |
2262 | ||
8204e0c1 AD |
2263 | /* Delay binding to CPU if node is not valid or online */ |
2264 | if (node < 0 || node >= MAX_NUMNODES || !node_online(node)) | |
2265 | return WORK_CPU_UNBOUND; | |
2266 | ||
2267 | /* Use local node/cpu if we are already there */ | |
2268 | cpu = raw_smp_processor_id(); | |
2269 | if (node == cpu_to_node(cpu)) | |
2270 | return cpu; | |
2271 | ||
2272 | /* Use "random" otherwise know as "first" online CPU of node */ | |
2273 | cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask); | |
2274 | ||
2275 | /* If CPU is valid return that, otherwise just defer */ | |
2276 | return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND; | |
2277 | } | |
2278 | ||
2279 | /** | |
2280 | * queue_work_node - queue work on a "random" cpu for a given NUMA node | |
2281 | * @node: NUMA node that we are targeting the work for | |
2282 | * @wq: workqueue to use | |
2283 | * @work: work to queue | |
2284 | * | |
2285 | * We queue the work to a "random" CPU within a given NUMA node. The basic | |
2286 | * idea here is to provide a way to somehow associate work with a given | |
2287 | * NUMA node. | |
2288 | * | |
2289 | * This function will only make a best effort attempt at getting this onto | |
2290 | * the right NUMA node. If no node is requested or the requested node is | |
2291 | * offline then we just fall back to standard queue_work behavior. | |
2292 | * | |
2293 | * Currently the "random" CPU ends up being the first available CPU in the | |
2294 | * intersection of cpu_online_mask and the cpumask of the node, unless we | |
2295 | * are running on the node. In that case we just use the current CPU. | |
2296 | * | |
2297 | * Return: %false if @work was already on a queue, %true otherwise. | |
2298 | */ | |
2299 | bool queue_work_node(int node, struct workqueue_struct *wq, | |
2300 | struct work_struct *work) | |
2301 | { | |
2302 | unsigned long flags; | |
2303 | bool ret = false; | |
2304 | ||
2305 | /* | |
2306 | * This current implementation is specific to unbound workqueues. | |
2307 | * Specifically we only return the first available CPU for a given | |
2308 | * node instead of cycling through individual CPUs within the node. | |
2309 | * | |
2310 | * If this is used with a per-cpu workqueue then the logic in | |
2311 | * workqueue_select_cpu_near would need to be updated to allow for | |
2312 | * some round robin type logic. | |
2313 | */ | |
2314 | WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)); | |
2315 | ||
2316 | local_irq_save(flags); | |
2317 | ||
2318 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { | |
fef59c9c | 2319 | int cpu = select_numa_node_cpu(node); |
8204e0c1 AD |
2320 | |
2321 | __queue_work(cpu, wq, work); | |
2322 | ret = true; | |
2323 | } | |
2324 | ||
2325 | local_irq_restore(flags); | |
2326 | return ret; | |
2327 | } | |
2328 | EXPORT_SYMBOL_GPL(queue_work_node); | |
2329 | ||
8c20feb6 | 2330 | void delayed_work_timer_fn(struct timer_list *t) |
1da177e4 | 2331 | { |
8c20feb6 | 2332 | struct delayed_work *dwork = from_timer(dwork, t, timer); |
1da177e4 | 2333 | |
e0aecdd8 | 2334 | /* should have been called from irqsafe timer with irq already off */ |
60c057bc | 2335 | __queue_work(dwork->cpu, dwork->wq, &dwork->work); |
1da177e4 | 2336 | } |
1438ade5 | 2337 | EXPORT_SYMBOL(delayed_work_timer_fn); |
1da177e4 | 2338 | |
7beb2edf TH |
2339 | static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, |
2340 | struct delayed_work *dwork, unsigned long delay) | |
1da177e4 | 2341 | { |
7beb2edf TH |
2342 | struct timer_list *timer = &dwork->timer; |
2343 | struct work_struct *work = &dwork->work; | |
7beb2edf | 2344 | |
637fdbae | 2345 | WARN_ON_ONCE(!wq); |
4b243563 | 2346 | WARN_ON_ONCE(timer->function != delayed_work_timer_fn); |
fc4b514f TH |
2347 | WARN_ON_ONCE(timer_pending(timer)); |
2348 | WARN_ON_ONCE(!list_empty(&work->entry)); | |
7beb2edf | 2349 | |
8852aac2 TH |
2350 | /* |
2351 | * If @delay is 0, queue @dwork->work immediately. This is for | |
2352 | * both optimization and correctness. The earliest @timer can | |
2353 | * expire is on the closest next tick and delayed_work users depend | |
2354 | * on that there's no such delay when @delay is 0. | |
2355 | */ | |
2356 | if (!delay) { | |
2357 | __queue_work(cpu, wq, &dwork->work); | |
2358 | return; | |
2359 | } | |
2360 | ||
60c057bc | 2361 | dwork->wq = wq; |
1265057f | 2362 | dwork->cpu = cpu; |
7beb2edf TH |
2363 | timer->expires = jiffies + delay; |
2364 | ||
041bd12e TH |
2365 | if (unlikely(cpu != WORK_CPU_UNBOUND)) |
2366 | add_timer_on(timer, cpu); | |
2367 | else | |
2368 | add_timer(timer); | |
1da177e4 LT |
2369 | } |
2370 | ||
0fcb78c2 REB |
2371 | /** |
2372 | * queue_delayed_work_on - queue work on specific CPU after delay | |
2373 | * @cpu: CPU number to execute work on | |
2374 | * @wq: workqueue to use | |
af9997e4 | 2375 | * @dwork: work to queue |
0fcb78c2 REB |
2376 | * @delay: number of jiffies to wait before queueing |
2377 | * | |
d185af30 | 2378 | * Return: %false if @work was already on a queue, %true otherwise. If |
715f1300 TH |
2379 | * @delay is zero and @dwork is idle, it will be scheduled for immediate |
2380 | * execution. | |
0fcb78c2 | 2381 | */ |
d4283e93 TH |
2382 | bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, |
2383 | struct delayed_work *dwork, unsigned long delay) | |
7a6bc1cd | 2384 | { |
52bad64d | 2385 | struct work_struct *work = &dwork->work; |
d4283e93 | 2386 | bool ret = false; |
8930caba | 2387 | unsigned long flags; |
7a6bc1cd | 2388 | |
8930caba TH |
2389 | /* read the comment in __queue_work() */ |
2390 | local_irq_save(flags); | |
7a6bc1cd | 2391 | |
22df02bb | 2392 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { |
7beb2edf | 2393 | __queue_delayed_work(cpu, wq, dwork, delay); |
d4283e93 | 2394 | ret = true; |
7a6bc1cd | 2395 | } |
8a3e77cc | 2396 | |
8930caba | 2397 | local_irq_restore(flags); |
7a6bc1cd VP |
2398 | return ret; |
2399 | } | |
ad7b1f84 | 2400 | EXPORT_SYMBOL(queue_delayed_work_on); |
c7fc77f7 | 2401 | |
8376fe22 TH |
2402 | /** |
2403 | * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU | |
2404 | * @cpu: CPU number to execute work on | |
2405 | * @wq: workqueue to use | |
2406 | * @dwork: work to queue | |
2407 | * @delay: number of jiffies to wait before queueing | |
2408 | * | |
2409 | * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, | |
2410 | * modify @dwork's timer so that it expires after @delay. If @delay is | |
2411 | * zero, @work is guaranteed to be scheduled immediately regardless of its | |
2412 | * current state. | |
2413 | * | |
d185af30 | 2414 | * Return: %false if @dwork was idle and queued, %true if @dwork was |
8376fe22 TH |
2415 | * pending and its timer was modified. |
2416 | * | |
e0aecdd8 | 2417 | * This function is safe to call from any context including IRQ handler. |
8376fe22 TH |
2418 | * See try_to_grab_pending() for details. |
2419 | */ | |
2420 | bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, | |
2421 | struct delayed_work *dwork, unsigned long delay) | |
2422 | { | |
2423 | unsigned long flags; | |
2424 | int ret; | |
c7fc77f7 | 2425 | |
8376fe22 TH |
2426 | do { |
2427 | ret = try_to_grab_pending(&dwork->work, true, &flags); | |
2428 | } while (unlikely(ret == -EAGAIN)); | |
63bc0362 | 2429 | |
8376fe22 TH |
2430 | if (likely(ret >= 0)) { |
2431 | __queue_delayed_work(cpu, wq, dwork, delay); | |
2432 | local_irq_restore(flags); | |
7a6bc1cd | 2433 | } |
8376fe22 TH |
2434 | |
2435 | /* -ENOENT from try_to_grab_pending() becomes %true */ | |
7a6bc1cd VP |
2436 | return ret; |
2437 | } | |
8376fe22 TH |
2438 | EXPORT_SYMBOL_GPL(mod_delayed_work_on); |
2439 | ||
05f0fe6b TH |
2440 | static void rcu_work_rcufn(struct rcu_head *rcu) |
2441 | { | |
2442 | struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu); | |
2443 | ||
2444 | /* read the comment in __queue_work() */ | |
2445 | local_irq_disable(); | |
2446 | __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work); | |
2447 | local_irq_enable(); | |
2448 | } | |
2449 | ||
2450 | /** | |
2451 | * queue_rcu_work - queue work after a RCU grace period | |
2452 | * @wq: workqueue to use | |
2453 | * @rwork: work to queue | |
2454 | * | |
2455 | * Return: %false if @rwork was already pending, %true otherwise. Note | |
2456 | * that a full RCU grace period is guaranteed only after a %true return. | |
bf393fd4 | 2457 | * While @rwork is guaranteed to be executed after a %false return, the |
05f0fe6b TH |
2458 | * execution may happen before a full RCU grace period has passed. |
2459 | */ | |
2460 | bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork) | |
2461 | { | |
2462 | struct work_struct *work = &rwork->work; | |
2463 | ||
2464 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { | |
2465 | rwork->wq = wq; | |
a7e30c0e | 2466 | call_rcu_hurry(&rwork->rcu, rcu_work_rcufn); |
05f0fe6b TH |
2467 | return true; |
2468 | } | |
2469 | ||
2470 | return false; | |
2471 | } | |
2472 | EXPORT_SYMBOL(queue_rcu_work); | |
2473 | ||
f7537df5 | 2474 | static struct worker *alloc_worker(int node) |
c34056a3 TH |
2475 | { |
2476 | struct worker *worker; | |
2477 | ||
f7537df5 | 2478 | worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node); |
c8e55f36 TH |
2479 | if (worker) { |
2480 | INIT_LIST_HEAD(&worker->entry); | |
affee4b2 | 2481 | INIT_LIST_HEAD(&worker->scheduled); |
da028469 | 2482 | INIT_LIST_HEAD(&worker->node); |
e22bee78 TH |
2483 | /* on creation a worker is in !idle && prep state */ |
2484 | worker->flags = WORKER_PREP; | |
c8e55f36 | 2485 | } |
c34056a3 TH |
2486 | return worker; |
2487 | } | |
2488 | ||
9546b29e TH |
2489 | static cpumask_t *pool_allowed_cpus(struct worker_pool *pool) |
2490 | { | |
8639eceb TH |
2491 | if (pool->cpu < 0 && pool->attrs->affn_strict) |
2492 | return pool->attrs->__pod_cpumask; | |
2493 | else | |
2494 | return pool->attrs->cpumask; | |
9546b29e TH |
2495 | } |
2496 | ||
4736cbf7 LJ |
2497 | /** |
2498 | * worker_attach_to_pool() - attach a worker to a pool | |
2499 | * @worker: worker to be attached | |
2500 | * @pool: the target pool | |
2501 | * | |
2502 | * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and | |
2503 | * cpu-binding of @worker are kept coordinated with the pool across | |
2504 | * cpu-[un]hotplugs. | |
2505 | */ | |
2506 | static void worker_attach_to_pool(struct worker *worker, | |
2507 | struct worker_pool *pool) | |
2508 | { | |
1258fae7 | 2509 | mutex_lock(&wq_pool_attach_mutex); |
4736cbf7 | 2510 | |
4736cbf7 | 2511 | /* |
1258fae7 TH |
2512 | * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains |
2513 | * stable across this function. See the comments above the flag | |
2514 | * definition for details. | |
4736cbf7 LJ |
2515 | */ |
2516 | if (pool->flags & POOL_DISASSOCIATED) | |
2517 | worker->flags |= WORKER_UNBOUND; | |
5c25b5ff PZ |
2518 | else |
2519 | kthread_set_per_cpu(worker->task, pool->cpu); | |
4736cbf7 | 2520 | |
640f17c8 | 2521 | if (worker->rescue_wq) |
9546b29e | 2522 | set_cpus_allowed_ptr(worker->task, pool_allowed_cpus(pool)); |
640f17c8 | 2523 | |
4736cbf7 | 2524 | list_add_tail(&worker->node, &pool->workers); |
a2d812a2 | 2525 | worker->pool = pool; |
4736cbf7 | 2526 | |
1258fae7 | 2527 | mutex_unlock(&wq_pool_attach_mutex); |
4736cbf7 LJ |
2528 | } |
2529 | ||
60f5a4bc LJ |
2530 | /** |
2531 | * worker_detach_from_pool() - detach a worker from its pool | |
2532 | * @worker: worker which is attached to its pool | |
60f5a4bc | 2533 | * |
4736cbf7 LJ |
2534 | * Undo the attaching which had been done in worker_attach_to_pool(). The |
2535 | * caller worker shouldn't access to the pool after detached except it has | |
2536 | * other reference to the pool. | |
60f5a4bc | 2537 | */ |
a2d812a2 | 2538 | static void worker_detach_from_pool(struct worker *worker) |
60f5a4bc | 2539 | { |
a2d812a2 | 2540 | struct worker_pool *pool = worker->pool; |
60f5a4bc LJ |
2541 | struct completion *detach_completion = NULL; |
2542 | ||
1258fae7 | 2543 | mutex_lock(&wq_pool_attach_mutex); |
a2d812a2 | 2544 | |
5c25b5ff | 2545 | kthread_set_per_cpu(worker->task, -1); |
da028469 | 2546 | list_del(&worker->node); |
a2d812a2 TH |
2547 | worker->pool = NULL; |
2548 | ||
e02b9312 | 2549 | if (list_empty(&pool->workers) && list_empty(&pool->dying_workers)) |
60f5a4bc | 2550 | detach_completion = pool->detach_completion; |
1258fae7 | 2551 | mutex_unlock(&wq_pool_attach_mutex); |
60f5a4bc | 2552 | |
b62c0751 LJ |
2553 | /* clear leftover flags without pool->lock after it is detached */ |
2554 | worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND); | |
2555 | ||
60f5a4bc LJ |
2556 | if (detach_completion) |
2557 | complete(detach_completion); | |
2558 | } | |
2559 | ||
c34056a3 TH |
2560 | /** |
2561 | * create_worker - create a new workqueue worker | |
63d95a91 | 2562 | * @pool: pool the new worker will belong to |
c34056a3 | 2563 | * |
051e1850 | 2564 | * Create and start a new worker which is attached to @pool. |
c34056a3 TH |
2565 | * |
2566 | * CONTEXT: | |
2567 | * Might sleep. Does GFP_KERNEL allocations. | |
2568 | * | |
d185af30 | 2569 | * Return: |
c34056a3 TH |
2570 | * Pointer to the newly created worker. |
2571 | */ | |
bc2ae0f5 | 2572 | static struct worker *create_worker(struct worker_pool *pool) |
c34056a3 | 2573 | { |
e441b56f ZL |
2574 | struct worker *worker; |
2575 | int id; | |
5d9c7a1e | 2576 | char id_buf[23]; |
c34056a3 | 2577 | |
7cda9aae | 2578 | /* ID is needed to determine kthread name */ |
e441b56f | 2579 | id = ida_alloc(&pool->worker_ida, GFP_KERNEL); |
3f0ea0b8 PM |
2580 | if (id < 0) { |
2581 | pr_err_once("workqueue: Failed to allocate a worker ID: %pe\n", | |
2582 | ERR_PTR(id)); | |
e441b56f | 2583 | return NULL; |
3f0ea0b8 | 2584 | } |
c34056a3 | 2585 | |
f7537df5 | 2586 | worker = alloc_worker(pool->node); |
3f0ea0b8 PM |
2587 | if (!worker) { |
2588 | pr_err_once("workqueue: Failed to allocate a worker\n"); | |
c34056a3 | 2589 | goto fail; |
3f0ea0b8 | 2590 | } |
c34056a3 | 2591 | |
c34056a3 TH |
2592 | worker->id = id; |
2593 | ||
29c91e99 | 2594 | if (pool->cpu >= 0) |
e3c916a4 TH |
2595 | snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, |
2596 | pool->attrs->nice < 0 ? "H" : ""); | |
f3421797 | 2597 | else |
e3c916a4 TH |
2598 | snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); |
2599 | ||
f3f90ad4 | 2600 | worker->task = kthread_create_on_node(worker_thread, worker, pool->node, |
e3c916a4 | 2601 | "kworker/%s", id_buf); |
3f0ea0b8 | 2602 | if (IS_ERR(worker->task)) { |
60f54038 PM |
2603 | if (PTR_ERR(worker->task) == -EINTR) { |
2604 | pr_err("workqueue: Interrupted when creating a worker thread \"kworker/%s\"\n", | |
2605 | id_buf); | |
2606 | } else { | |
2607 | pr_err_once("workqueue: Failed to create a worker thread: %pe", | |
2608 | worker->task); | |
2609 | } | |
c34056a3 | 2610 | goto fail; |
3f0ea0b8 | 2611 | } |
c34056a3 | 2612 | |
91151228 | 2613 | set_user_nice(worker->task, pool->attrs->nice); |
9546b29e | 2614 | kthread_bind_mask(worker->task, pool_allowed_cpus(pool)); |
91151228 | 2615 | |
da028469 | 2616 | /* successful, attach the worker to the pool */ |
4736cbf7 | 2617 | worker_attach_to_pool(worker, pool); |
822d8405 | 2618 | |
051e1850 | 2619 | /* start the newly created worker */ |
a9b8a985 | 2620 | raw_spin_lock_irq(&pool->lock); |
0219a352 | 2621 | |
051e1850 LJ |
2622 | worker->pool->nr_workers++; |
2623 | worker_enter_idle(worker); | |
0219a352 TH |
2624 | |
2625 | /* | |
2626 | * @worker is waiting on a completion in kthread() and will trigger hung | |
6a229b0e TH |
2627 | * check if not woken up soon. As kick_pool() is noop if @pool is empty, |
2628 | * wake it up explicitly. | |
0219a352 | 2629 | */ |
051e1850 | 2630 | wake_up_process(worker->task); |
0219a352 | 2631 | |
a9b8a985 | 2632 | raw_spin_unlock_irq(&pool->lock); |
051e1850 | 2633 | |
c34056a3 | 2634 | return worker; |
822d8405 | 2635 | |
c34056a3 | 2636 | fail: |
e441b56f | 2637 | ida_free(&pool->worker_ida, id); |
c34056a3 TH |
2638 | kfree(worker); |
2639 | return NULL; | |
2640 | } | |
2641 | ||
793777bc VS |
2642 | static void unbind_worker(struct worker *worker) |
2643 | { | |
2644 | lockdep_assert_held(&wq_pool_attach_mutex); | |
2645 | ||
2646 | kthread_set_per_cpu(worker->task, -1); | |
2647 | if (cpumask_intersects(wq_unbound_cpumask, cpu_active_mask)) | |
2648 | WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, wq_unbound_cpumask) < 0); | |
2649 | else | |
2650 | WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0); | |
2651 | } | |
2652 | ||
e02b9312 VS |
2653 | static void wake_dying_workers(struct list_head *cull_list) |
2654 | { | |
2655 | struct worker *worker, *tmp; | |
2656 | ||
2657 | list_for_each_entry_safe(worker, tmp, cull_list, entry) { | |
2658 | list_del_init(&worker->entry); | |
2659 | unbind_worker(worker); | |
2660 | /* | |
2661 | * If the worker was somehow already running, then it had to be | |
2662 | * in pool->idle_list when set_worker_dying() happened or we | |
2663 | * wouldn't have gotten here. | |
2664 | * | |
2665 | * Thus, the worker must either have observed the WORKER_DIE | |
2666 | * flag, or have set its state to TASK_IDLE. Either way, the | |
2667 | * below will be observed by the worker and is safe to do | |
2668 | * outside of pool->lock. | |
2669 | */ | |
2670 | wake_up_process(worker->task); | |
2671 | } | |
2672 | } | |
2673 | ||
c34056a3 | 2674 | /** |
e02b9312 | 2675 | * set_worker_dying - Tag a worker for destruction |
c34056a3 | 2676 | * @worker: worker to be destroyed |
e02b9312 | 2677 | * @list: transfer worker away from its pool->idle_list and into list |
c34056a3 | 2678 | * |
e02b9312 VS |
2679 | * Tag @worker for destruction and adjust @pool stats accordingly. The worker |
2680 | * should be idle. | |
c8e55f36 TH |
2681 | * |
2682 | * CONTEXT: | |
a9b8a985 | 2683 | * raw_spin_lock_irq(pool->lock). |
c34056a3 | 2684 | */ |
e02b9312 | 2685 | static void set_worker_dying(struct worker *worker, struct list_head *list) |
c34056a3 | 2686 | { |
bd7bdd43 | 2687 | struct worker_pool *pool = worker->pool; |
c34056a3 | 2688 | |
cd549687 | 2689 | lockdep_assert_held(&pool->lock); |
e02b9312 | 2690 | lockdep_assert_held(&wq_pool_attach_mutex); |
cd549687 | 2691 | |
c34056a3 | 2692 | /* sanity check frenzy */ |
6183c009 | 2693 | if (WARN_ON(worker->current_work) || |
73eb7fe7 LJ |
2694 | WARN_ON(!list_empty(&worker->scheduled)) || |
2695 | WARN_ON(!(worker->flags & WORKER_IDLE))) | |
6183c009 | 2696 | return; |
c34056a3 | 2697 | |
73eb7fe7 LJ |
2698 | pool->nr_workers--; |
2699 | pool->nr_idle--; | |
5bdfff96 | 2700 | |
cb444766 | 2701 | worker->flags |= WORKER_DIE; |
e02b9312 VS |
2702 | |
2703 | list_move(&worker->entry, list); | |
2704 | list_move(&worker->node, &pool->dying_workers); | |
c34056a3 TH |
2705 | } |
2706 | ||
3f959aa3 VS |
2707 | /** |
2708 | * idle_worker_timeout - check if some idle workers can now be deleted. | |
2709 | * @t: The pool's idle_timer that just expired | |
2710 | * | |
2711 | * The timer is armed in worker_enter_idle(). Note that it isn't disarmed in | |
2712 | * worker_leave_idle(), as a worker flicking between idle and active while its | |
2713 | * pool is at the too_many_workers() tipping point would cause too much timer | |
2714 | * housekeeping overhead. Since IDLE_WORKER_TIMEOUT is long enough, we just let | |
2715 | * it expire and re-evaluate things from there. | |
2716 | */ | |
32a6c723 | 2717 | static void idle_worker_timeout(struct timer_list *t) |
e22bee78 | 2718 | { |
32a6c723 | 2719 | struct worker_pool *pool = from_timer(pool, t, idle_timer); |
3f959aa3 VS |
2720 | bool do_cull = false; |
2721 | ||
2722 | if (work_pending(&pool->idle_cull_work)) | |
2723 | return; | |
e22bee78 | 2724 | |
a9b8a985 | 2725 | raw_spin_lock_irq(&pool->lock); |
e22bee78 | 2726 | |
3f959aa3 | 2727 | if (too_many_workers(pool)) { |
e22bee78 TH |
2728 | struct worker *worker; |
2729 | unsigned long expires; | |
2730 | ||
2731 | /* idle_list is kept in LIFO order, check the last one */ | |
3f959aa3 VS |
2732 | worker = list_entry(pool->idle_list.prev, struct worker, entry); |
2733 | expires = worker->last_active + IDLE_WORKER_TIMEOUT; | |
2734 | do_cull = !time_before(jiffies, expires); | |
2735 | ||
2736 | if (!do_cull) | |
2737 | mod_timer(&pool->idle_timer, expires); | |
2738 | } | |
2739 | raw_spin_unlock_irq(&pool->lock); | |
2740 | ||
2741 | if (do_cull) | |
2742 | queue_work(system_unbound_wq, &pool->idle_cull_work); | |
2743 | } | |
2744 | ||
2745 | /** | |
2746 | * idle_cull_fn - cull workers that have been idle for too long. | |
2747 | * @work: the pool's work for handling these idle workers | |
2748 | * | |
2749 | * This goes through a pool's idle workers and gets rid of those that have been | |
2750 | * idle for at least IDLE_WORKER_TIMEOUT seconds. | |
e02b9312 VS |
2751 | * |
2752 | * We don't want to disturb isolated CPUs because of a pcpu kworker being | |
2753 | * culled, so this also resets worker affinity. This requires a sleepable | |
2754 | * context, hence the split between timer callback and work item. | |
3f959aa3 VS |
2755 | */ |
2756 | static void idle_cull_fn(struct work_struct *work) | |
2757 | { | |
2758 | struct worker_pool *pool = container_of(work, struct worker_pool, idle_cull_work); | |
9680540c | 2759 | LIST_HEAD(cull_list); |
3f959aa3 | 2760 | |
e02b9312 VS |
2761 | /* |
2762 | * Grabbing wq_pool_attach_mutex here ensures an already-running worker | |
2763 | * cannot proceed beyong worker_detach_from_pool() in its self-destruct | |
2764 | * path. This is required as a previously-preempted worker could run after | |
2765 | * set_worker_dying() has happened but before wake_dying_workers() did. | |
2766 | */ | |
2767 | mutex_lock(&wq_pool_attach_mutex); | |
3f959aa3 VS |
2768 | raw_spin_lock_irq(&pool->lock); |
2769 | ||
2770 | while (too_many_workers(pool)) { | |
2771 | struct worker *worker; | |
2772 | unsigned long expires; | |
2773 | ||
63d95a91 | 2774 | worker = list_entry(pool->idle_list.prev, struct worker, entry); |
e22bee78 TH |
2775 | expires = worker->last_active + IDLE_WORKER_TIMEOUT; |
2776 | ||
3347fc9f | 2777 | if (time_before(jiffies, expires)) { |
63d95a91 | 2778 | mod_timer(&pool->idle_timer, expires); |
3347fc9f | 2779 | break; |
d5abe669 | 2780 | } |
3347fc9f | 2781 | |
e02b9312 | 2782 | set_worker_dying(worker, &cull_list); |
e22bee78 TH |
2783 | } |
2784 | ||
a9b8a985 | 2785 | raw_spin_unlock_irq(&pool->lock); |
e02b9312 VS |
2786 | wake_dying_workers(&cull_list); |
2787 | mutex_unlock(&wq_pool_attach_mutex); | |
e22bee78 | 2788 | } |
d5abe669 | 2789 | |
493a1724 | 2790 | static void send_mayday(struct work_struct *work) |
e22bee78 | 2791 | { |
112202d9 TH |
2792 | struct pool_workqueue *pwq = get_work_pwq(work); |
2793 | struct workqueue_struct *wq = pwq->wq; | |
493a1724 | 2794 | |
2e109a28 | 2795 | lockdep_assert_held(&wq_mayday_lock); |
e22bee78 | 2796 | |
493008a8 | 2797 | if (!wq->rescuer) |
493a1724 | 2798 | return; |
e22bee78 TH |
2799 | |
2800 | /* mayday mayday mayday */ | |
493a1724 | 2801 | if (list_empty(&pwq->mayday_node)) { |
77668c8b LJ |
2802 | /* |
2803 | * If @pwq is for an unbound wq, its base ref may be put at | |
2804 | * any time due to an attribute change. Pin @pwq until the | |
2805 | * rescuer is done with it. | |
2806 | */ | |
2807 | get_pwq(pwq); | |
493a1724 | 2808 | list_add_tail(&pwq->mayday_node, &wq->maydays); |
e22bee78 | 2809 | wake_up_process(wq->rescuer->task); |
725e8ec5 | 2810 | pwq->stats[PWQ_STAT_MAYDAY]++; |
493a1724 | 2811 | } |
e22bee78 TH |
2812 | } |
2813 | ||
32a6c723 | 2814 | static void pool_mayday_timeout(struct timer_list *t) |
e22bee78 | 2815 | { |
32a6c723 | 2816 | struct worker_pool *pool = from_timer(pool, t, mayday_timer); |
e22bee78 TH |
2817 | struct work_struct *work; |
2818 | ||
a9b8a985 SAS |
2819 | raw_spin_lock_irq(&pool->lock); |
2820 | raw_spin_lock(&wq_mayday_lock); /* for wq->maydays */ | |
e22bee78 | 2821 | |
63d95a91 | 2822 | if (need_to_create_worker(pool)) { |
e22bee78 TH |
2823 | /* |
2824 | * We've been trying to create a new worker but | |
2825 | * haven't been successful. We might be hitting an | |
2826 | * allocation deadlock. Send distress signals to | |
2827 | * rescuers. | |
2828 | */ | |
63d95a91 | 2829 | list_for_each_entry(work, &pool->worklist, entry) |
e22bee78 | 2830 | send_mayday(work); |
1da177e4 | 2831 | } |
e22bee78 | 2832 | |
a9b8a985 SAS |
2833 | raw_spin_unlock(&wq_mayday_lock); |
2834 | raw_spin_unlock_irq(&pool->lock); | |
e22bee78 | 2835 | |
63d95a91 | 2836 | mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); |
1da177e4 LT |
2837 | } |
2838 | ||
e22bee78 TH |
2839 | /** |
2840 | * maybe_create_worker - create a new worker if necessary | |
63d95a91 | 2841 | * @pool: pool to create a new worker for |
e22bee78 | 2842 | * |
63d95a91 | 2843 | * Create a new worker for @pool if necessary. @pool is guaranteed to |
e22bee78 TH |
2844 | * have at least one idle worker on return from this function. If |
2845 | * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is | |
63d95a91 | 2846 | * sent to all rescuers with works scheduled on @pool to resolve |
e22bee78 TH |
2847 | * possible allocation deadlock. |
2848 | * | |
c5aa87bb TH |
2849 | * On return, need_to_create_worker() is guaranteed to be %false and |
2850 | * may_start_working() %true. | |
e22bee78 TH |
2851 | * |
2852 | * LOCKING: | |
a9b8a985 | 2853 | * raw_spin_lock_irq(pool->lock) which may be released and regrabbed |
e22bee78 TH |
2854 | * multiple times. Does GFP_KERNEL allocations. Called only from |
2855 | * manager. | |
e22bee78 | 2856 | */ |
29187a9e | 2857 | static void maybe_create_worker(struct worker_pool *pool) |
d565ed63 TH |
2858 | __releases(&pool->lock) |
2859 | __acquires(&pool->lock) | |
1da177e4 | 2860 | { |
e22bee78 | 2861 | restart: |
a9b8a985 | 2862 | raw_spin_unlock_irq(&pool->lock); |
9f9c2364 | 2863 | |
e22bee78 | 2864 | /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ |
63d95a91 | 2865 | mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); |
e22bee78 TH |
2866 | |
2867 | while (true) { | |
051e1850 | 2868 | if (create_worker(pool) || !need_to_create_worker(pool)) |
e22bee78 | 2869 | break; |
1da177e4 | 2870 | |
e212f361 | 2871 | schedule_timeout_interruptible(CREATE_COOLDOWN); |
9f9c2364 | 2872 | |
63d95a91 | 2873 | if (!need_to_create_worker(pool)) |
e22bee78 TH |
2874 | break; |
2875 | } | |
2876 | ||
63d95a91 | 2877 | del_timer_sync(&pool->mayday_timer); |
a9b8a985 | 2878 | raw_spin_lock_irq(&pool->lock); |
051e1850 LJ |
2879 | /* |
2880 | * This is necessary even after a new worker was just successfully | |
2881 | * created as @pool->lock was dropped and the new worker might have | |
2882 | * already become busy. | |
2883 | */ | |
63d95a91 | 2884 | if (need_to_create_worker(pool)) |
e22bee78 | 2885 | goto restart; |
e22bee78 TH |
2886 | } |
2887 | ||
73f53c4a | 2888 | /** |
e22bee78 TH |
2889 | * manage_workers - manage worker pool |
2890 | * @worker: self | |
73f53c4a | 2891 | * |
706026c2 | 2892 | * Assume the manager role and manage the worker pool @worker belongs |
e22bee78 | 2893 | * to. At any given time, there can be only zero or one manager per |
706026c2 | 2894 | * pool. The exclusion is handled automatically by this function. |
e22bee78 TH |
2895 | * |
2896 | * The caller can safely start processing works on false return. On | |
2897 | * true return, it's guaranteed that need_to_create_worker() is false | |
2898 | * and may_start_working() is true. | |
73f53c4a TH |
2899 | * |
2900 | * CONTEXT: | |
a9b8a985 | 2901 | * raw_spin_lock_irq(pool->lock) which may be released and regrabbed |
e22bee78 TH |
2902 | * multiple times. Does GFP_KERNEL allocations. |
2903 | * | |
d185af30 | 2904 | * Return: |
29187a9e TH |
2905 | * %false if the pool doesn't need management and the caller can safely |
2906 | * start processing works, %true if management function was performed and | |
2907 | * the conditions that the caller verified before calling the function may | |
2908 | * no longer be true. | |
73f53c4a | 2909 | */ |
e22bee78 | 2910 | static bool manage_workers(struct worker *worker) |
73f53c4a | 2911 | { |
63d95a91 | 2912 | struct worker_pool *pool = worker->pool; |
73f53c4a | 2913 | |
692b4825 | 2914 | if (pool->flags & POOL_MANAGER_ACTIVE) |
29187a9e | 2915 | return false; |
692b4825 TH |
2916 | |
2917 | pool->flags |= POOL_MANAGER_ACTIVE; | |
2607d7a6 | 2918 | pool->manager = worker; |
1e19ffc6 | 2919 | |
29187a9e | 2920 | maybe_create_worker(pool); |
e22bee78 | 2921 | |
2607d7a6 | 2922 | pool->manager = NULL; |
692b4825 | 2923 | pool->flags &= ~POOL_MANAGER_ACTIVE; |
d8bb65ab | 2924 | rcuwait_wake_up(&manager_wait); |
29187a9e | 2925 | return true; |
73f53c4a TH |
2926 | } |
2927 | ||
a62428c0 TH |
2928 | /** |
2929 | * process_one_work - process single work | |
c34056a3 | 2930 | * @worker: self |
a62428c0 TH |
2931 | * @work: work to process |
2932 | * | |
2933 | * Process @work. This function contains all the logics necessary to | |
2934 | * process a single work including synchronization against and | |
2935 | * interaction with other workers on the same cpu, queueing and | |
2936 | * flushing. As long as context requirement is met, any worker can | |
2937 | * call this function to process a work. | |
2938 | * | |
2939 | * CONTEXT: | |
a9b8a985 | 2940 | * raw_spin_lock_irq(pool->lock) which is released and regrabbed. |
a62428c0 | 2941 | */ |
c34056a3 | 2942 | static void process_one_work(struct worker *worker, struct work_struct *work) |
d565ed63 TH |
2943 | __releases(&pool->lock) |
2944 | __acquires(&pool->lock) | |
a62428c0 | 2945 | { |
112202d9 | 2946 | struct pool_workqueue *pwq = get_work_pwq(work); |
bd7bdd43 | 2947 | struct worker_pool *pool = worker->pool; |
c4560c2c | 2948 | unsigned long work_data; |
a62428c0 TH |
2949 | #ifdef CONFIG_LOCKDEP |
2950 | /* | |
2951 | * It is permissible to free the struct work_struct from | |
2952 | * inside the function that is called from it, this we need to | |
2953 | * take into account for lockdep too. To avoid bogus "held | |
2954 | * lock freed" warnings as well as problems when looking into | |
2955 | * work->lockdep_map, make a copy and use that here. | |
2956 | */ | |
4d82a1de PZ |
2957 | struct lockdep_map lockdep_map; |
2958 | ||
2959 | lockdep_copy_map(&lockdep_map, &work->lockdep_map); | |
a62428c0 | 2960 | #endif |
807407c0 | 2961 | /* ensure we're on the correct CPU */ |
85327af6 | 2962 | WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && |
ec22ca5e | 2963 | raw_smp_processor_id() != pool->cpu); |
25511a47 | 2964 | |
8930caba | 2965 | /* claim and dequeue */ |
a62428c0 | 2966 | debug_work_deactivate(work); |
c9e7cf27 | 2967 | hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); |
c34056a3 | 2968 | worker->current_work = work; |
a2c1c57b | 2969 | worker->current_func = work->func; |
112202d9 | 2970 | worker->current_pwq = pwq; |
616db877 | 2971 | worker->current_at = worker->task->se.sum_exec_runtime; |
c4560c2c | 2972 | work_data = *work_data_bits(work); |
d812796e | 2973 | worker->current_color = get_work_color(work_data); |
7a22ad75 | 2974 | |
8bf89593 TH |
2975 | /* |
2976 | * Record wq name for cmdline and debug reporting, may get | |
2977 | * overridden through set_worker_desc(). | |
2978 | */ | |
2979 | strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN); | |
2980 | ||
a62428c0 TH |
2981 | list_del_init(&work->entry); |
2982 | ||
fb0e7beb | 2983 | /* |
228f1d00 LJ |
2984 | * CPU intensive works don't participate in concurrency management. |
2985 | * They're the scheduler's responsibility. This takes @worker out | |
2986 | * of concurrency management and the next code block will chain | |
2987 | * execution of the pending work items. | |
fb0e7beb | 2988 | */ |
616db877 | 2989 | if (unlikely(pwq->wq->flags & WQ_CPU_INTENSIVE)) |
228f1d00 | 2990 | worker_set_flags(worker, WORKER_CPU_INTENSIVE); |
fb0e7beb | 2991 | |
974271c4 | 2992 | /* |
0219a352 TH |
2993 | * Kick @pool if necessary. It's always noop for per-cpu worker pools |
2994 | * since nr_running would always be >= 1 at this point. This is used to | |
2995 | * chain execution of the pending work items for WORKER_NOT_RUNNING | |
2996 | * workers such as the UNBOUND and CPU_INTENSIVE ones. | |
974271c4 | 2997 | */ |
0219a352 | 2998 | kick_pool(pool); |
974271c4 | 2999 | |
8930caba | 3000 | /* |
7c3eed5c | 3001 | * Record the last pool and clear PENDING which should be the last |
d565ed63 | 3002 | * update to @work. Also, do this inside @pool->lock so that |
23657bb1 TH |
3003 | * PENDING and queued state changes happen together while IRQ is |
3004 | * disabled. | |
8930caba | 3005 | */ |
7c3eed5c | 3006 | set_work_pool_and_clear_pending(work, pool->id); |
a62428c0 | 3007 | |
fe48ba7d | 3008 | pwq->stats[PWQ_STAT_STARTED]++; |
a9b8a985 | 3009 | raw_spin_unlock_irq(&pool->lock); |
a62428c0 | 3010 | |
a1d14934 | 3011 | lock_map_acquire(&pwq->wq->lockdep_map); |
a62428c0 | 3012 | lock_map_acquire(&lockdep_map); |
e6f3faa7 | 3013 | /* |
f52be570 PZ |
3014 | * Strictly speaking we should mark the invariant state without holding |
3015 | * any locks, that is, before these two lock_map_acquire()'s. | |
e6f3faa7 PZ |
3016 | * |
3017 | * However, that would result in: | |
3018 | * | |
3019 | * A(W1) | |
3020 | * WFC(C) | |
3021 | * A(W1) | |
3022 | * C(C) | |
3023 | * | |
3024 | * Which would create W1->C->W1 dependencies, even though there is no | |
3025 | * actual deadlock possible. There are two solutions, using a | |
3026 | * read-recursive acquire on the work(queue) 'locks', but this will then | |
f52be570 | 3027 | * hit the lockdep limitation on recursive locks, or simply discard |
e6f3faa7 PZ |
3028 | * these locks. |
3029 | * | |
3030 | * AFAICT there is no possible deadlock scenario between the | |
3031 | * flush_work() and complete() primitives (except for single-threaded | |
3032 | * workqueues), so hiding them isn't a problem. | |
3033 | */ | |
f52be570 | 3034 | lockdep_invariant_state(true); |
e36c886a | 3035 | trace_workqueue_execute_start(work); |
a2c1c57b | 3036 | worker->current_func(work); |
e36c886a AV |
3037 | /* |
3038 | * While we must be careful to not use "work" after this, the trace | |
3039 | * point will only record its address. | |
3040 | */ | |
1c5da0ec | 3041 | trace_workqueue_execute_end(work, worker->current_func); |
725e8ec5 | 3042 | pwq->stats[PWQ_STAT_COMPLETED]++; |
a62428c0 | 3043 | lock_map_release(&lockdep_map); |
112202d9 | 3044 | lock_map_release(&pwq->wq->lockdep_map); |
a62428c0 | 3045 | |
1a65a6d1 XY |
3046 | if (unlikely(in_atomic() || lockdep_depth(current) > 0 || |
3047 | rcu_preempt_depth() > 0)) { | |
3048 | pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d/%d\n" | |
d75f773c | 3049 | " last function: %ps\n", |
1a65a6d1 XY |
3050 | current->comm, preempt_count(), rcu_preempt_depth(), |
3051 | task_pid_nr(current), worker->current_func); | |
a62428c0 TH |
3052 | debug_show_held_locks(current); |
3053 | dump_stack(); | |
3054 | } | |
3055 | ||
b22ce278 | 3056 | /* |
025f50f3 | 3057 | * The following prevents a kworker from hogging CPU on !PREEMPTION |
b22ce278 TH |
3058 | * kernels, where a requeueing work item waiting for something to |
3059 | * happen could deadlock with stop_machine as such work item could | |
3060 | * indefinitely requeue itself while all other CPUs are trapped in | |
789cbbec JL |
3061 | * stop_machine. At the same time, report a quiescent RCU state so |
3062 | * the same condition doesn't freeze RCU. | |
b22ce278 | 3063 | */ |
a7e6425e | 3064 | cond_resched(); |
b22ce278 | 3065 | |
a9b8a985 | 3066 | raw_spin_lock_irq(&pool->lock); |
a62428c0 | 3067 | |
616db877 TH |
3068 | /* |
3069 | * In addition to %WQ_CPU_INTENSIVE, @worker may also have been marked | |
3070 | * CPU intensive by wq_worker_tick() if @work hogged CPU longer than | |
3071 | * wq_cpu_intensive_thresh_us. Clear it. | |
3072 | */ | |
3073 | worker_clr_flags(worker, WORKER_CPU_INTENSIVE); | |
fb0e7beb | 3074 | |
1b69ac6b JW |
3075 | /* tag the worker for identification in schedule() */ |
3076 | worker->last_func = worker->current_func; | |
3077 | ||
a62428c0 | 3078 | /* we're done with it, release */ |
42f8570f | 3079 | hash_del(&worker->hentry); |
c34056a3 | 3080 | worker->current_work = NULL; |
a2c1c57b | 3081 | worker->current_func = NULL; |
112202d9 | 3082 | worker->current_pwq = NULL; |
d812796e | 3083 | worker->current_color = INT_MAX; |
dd6c3c54 TH |
3084 | |
3085 | /* must be the last step, see the function comment */ | |
c4560c2c | 3086 | pwq_dec_nr_in_flight(pwq, work_data); |
a62428c0 TH |
3087 | } |
3088 | ||
affee4b2 TH |
3089 | /** |
3090 | * process_scheduled_works - process scheduled works | |
3091 | * @worker: self | |
3092 | * | |
3093 | * Process all scheduled works. Please note that the scheduled list | |
3094 | * may change while processing a work, so this function repeatedly | |
3095 | * fetches a work from the top and executes it. | |
3096 | * | |
3097 | * CONTEXT: | |
a9b8a985 | 3098 | * raw_spin_lock_irq(pool->lock) which may be released and regrabbed |
affee4b2 TH |
3099 | * multiple times. |
3100 | */ | |
3101 | static void process_scheduled_works(struct worker *worker) | |
1da177e4 | 3102 | { |
c0ab017d TH |
3103 | struct work_struct *work; |
3104 | bool first = true; | |
3105 | ||
3106 | while ((work = list_first_entry_or_null(&worker->scheduled, | |
3107 | struct work_struct, entry))) { | |
3108 | if (first) { | |
3109 | worker->pool->watchdog_ts = jiffies; | |
3110 | first = false; | |
3111 | } | |
c34056a3 | 3112 | process_one_work(worker, work); |
1da177e4 | 3113 | } |
1da177e4 LT |
3114 | } |
3115 | ||
197f6acc TH |
3116 | static void set_pf_worker(bool val) |
3117 | { | |
3118 | mutex_lock(&wq_pool_attach_mutex); | |
3119 | if (val) | |
3120 | current->flags |= PF_WQ_WORKER; | |
3121 | else | |
3122 | current->flags &= ~PF_WQ_WORKER; | |
3123 | mutex_unlock(&wq_pool_attach_mutex); | |
3124 | } | |
3125 | ||
4690c4ab TH |
3126 | /** |
3127 | * worker_thread - the worker thread function | |
c34056a3 | 3128 | * @__worker: self |
4690c4ab | 3129 | * |
c5aa87bb TH |
3130 | * The worker thread function. All workers belong to a worker_pool - |
3131 | * either a per-cpu one or dynamic unbound one. These workers process all | |
3132 | * work items regardless of their specific target workqueue. The only | |
3133 | * exception is work items which belong to workqueues with a rescuer which | |
3134 | * will be explained in rescuer_thread(). | |
d185af30 YB |
3135 | * |
3136 | * Return: 0 | |
4690c4ab | 3137 | */ |
c34056a3 | 3138 | static int worker_thread(void *__worker) |
1da177e4 | 3139 | { |
c34056a3 | 3140 | struct worker *worker = __worker; |
bd7bdd43 | 3141 | struct worker_pool *pool = worker->pool; |
1da177e4 | 3142 | |
e22bee78 | 3143 | /* tell the scheduler that this is a workqueue worker */ |
197f6acc | 3144 | set_pf_worker(true); |
c8e55f36 | 3145 | woke_up: |
a9b8a985 | 3146 | raw_spin_lock_irq(&pool->lock); |
1da177e4 | 3147 | |
a9ab775b TH |
3148 | /* am I supposed to die? */ |
3149 | if (unlikely(worker->flags & WORKER_DIE)) { | |
a9b8a985 | 3150 | raw_spin_unlock_irq(&pool->lock); |
197f6acc | 3151 | set_pf_worker(false); |
60f5a4bc LJ |
3152 | |
3153 | set_task_comm(worker->task, "kworker/dying"); | |
e441b56f | 3154 | ida_free(&pool->worker_ida, worker->id); |
a2d812a2 | 3155 | worker_detach_from_pool(worker); |
e02b9312 | 3156 | WARN_ON_ONCE(!list_empty(&worker->entry)); |
60f5a4bc | 3157 | kfree(worker); |
a9ab775b | 3158 | return 0; |
c8e55f36 | 3159 | } |
affee4b2 | 3160 | |
c8e55f36 | 3161 | worker_leave_idle(worker); |
db7bccf4 | 3162 | recheck: |
e22bee78 | 3163 | /* no more worker necessary? */ |
63d95a91 | 3164 | if (!need_more_worker(pool)) |
e22bee78 TH |
3165 | goto sleep; |
3166 | ||
3167 | /* do we need to manage? */ | |
63d95a91 | 3168 | if (unlikely(!may_start_working(pool)) && manage_workers(worker)) |
e22bee78 TH |
3169 | goto recheck; |
3170 | ||
c8e55f36 TH |
3171 | /* |
3172 | * ->scheduled list can only be filled while a worker is | |
3173 | * preparing to process a work or actually processing it. | |
3174 | * Make sure nobody diddled with it while I was sleeping. | |
3175 | */ | |
6183c009 | 3176 | WARN_ON_ONCE(!list_empty(&worker->scheduled)); |
c8e55f36 | 3177 | |
e22bee78 | 3178 | /* |
a9ab775b TH |
3179 | * Finish PREP stage. We're guaranteed to have at least one idle |
3180 | * worker or that someone else has already assumed the manager | |
3181 | * role. This is where @worker starts participating in concurrency | |
3182 | * management if applicable and concurrency management is restored | |
3183 | * after being rebound. See rebind_workers() for details. | |
e22bee78 | 3184 | */ |
a9ab775b | 3185 | worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); |
e22bee78 TH |
3186 | |
3187 | do { | |
c8e55f36 | 3188 | struct work_struct *work = |
bd7bdd43 | 3189 | list_first_entry(&pool->worklist, |
c8e55f36 TH |
3190 | struct work_struct, entry); |
3191 | ||
873eaca6 TH |
3192 | if (assign_work(work, worker, NULL)) |
3193 | process_scheduled_works(worker); | |
63d95a91 | 3194 | } while (keep_working(pool)); |
e22bee78 | 3195 | |
228f1d00 | 3196 | worker_set_flags(worker, WORKER_PREP); |
d313dd85 | 3197 | sleep: |
c8e55f36 | 3198 | /* |
d565ed63 TH |
3199 | * pool->lock is held and there's no work to process and no need to |
3200 | * manage, sleep. Workers are woken up only while holding | |
3201 | * pool->lock or from local cpu, so setting the current state | |
3202 | * before releasing pool->lock is enough to prevent losing any | |
3203 | * event. | |
c8e55f36 TH |
3204 | */ |
3205 | worker_enter_idle(worker); | |
c5a94a61 | 3206 | __set_current_state(TASK_IDLE); |
a9b8a985 | 3207 | raw_spin_unlock_irq(&pool->lock); |
c8e55f36 TH |
3208 | schedule(); |
3209 | goto woke_up; | |
1da177e4 LT |
3210 | } |
3211 | ||
e22bee78 TH |
3212 | /** |
3213 | * rescuer_thread - the rescuer thread function | |
111c225a | 3214 | * @__rescuer: self |
e22bee78 TH |
3215 | * |
3216 | * Workqueue rescuer thread function. There's one rescuer for each | |
493008a8 | 3217 | * workqueue which has WQ_MEM_RECLAIM set. |
e22bee78 | 3218 | * |
706026c2 | 3219 | * Regular work processing on a pool may block trying to create a new |
e22bee78 TH |
3220 | * worker which uses GFP_KERNEL allocation which has slight chance of |
3221 | * developing into deadlock if some works currently on the same queue | |
3222 | * need to be processed to satisfy the GFP_KERNEL allocation. This is | |
3223 | * the problem rescuer solves. | |
3224 | * | |
706026c2 TH |
3225 | * When such condition is possible, the pool summons rescuers of all |
3226 | * workqueues which have works queued on the pool and let them process | |
e22bee78 TH |
3227 | * those works so that forward progress can be guaranteed. |
3228 | * | |
3229 | * This should happen rarely. | |
d185af30 YB |
3230 | * |
3231 | * Return: 0 | |
e22bee78 | 3232 | */ |
111c225a | 3233 | static int rescuer_thread(void *__rescuer) |
e22bee78 | 3234 | { |
111c225a TH |
3235 | struct worker *rescuer = __rescuer; |
3236 | struct workqueue_struct *wq = rescuer->rescue_wq; | |
4d595b86 | 3237 | bool should_stop; |
e22bee78 TH |
3238 | |
3239 | set_user_nice(current, RESCUER_NICE_LEVEL); | |
111c225a TH |
3240 | |
3241 | /* | |
3242 | * Mark rescuer as worker too. As WORKER_PREP is never cleared, it | |
3243 | * doesn't participate in concurrency management. | |
3244 | */ | |
197f6acc | 3245 | set_pf_worker(true); |
e22bee78 | 3246 | repeat: |
c5a94a61 | 3247 | set_current_state(TASK_IDLE); |
e22bee78 | 3248 | |
4d595b86 LJ |
3249 | /* |
3250 | * By the time the rescuer is requested to stop, the workqueue | |
3251 | * shouldn't have any work pending, but @wq->maydays may still have | |
3252 | * pwq(s) queued. This can happen by non-rescuer workers consuming | |
3253 | * all the work items before the rescuer got to them. Go through | |
3254 | * @wq->maydays processing before acting on should_stop so that the | |
3255 | * list is always empty on exit. | |
3256 | */ | |
3257 | should_stop = kthread_should_stop(); | |
e22bee78 | 3258 | |
493a1724 | 3259 | /* see whether any pwq is asking for help */ |
a9b8a985 | 3260 | raw_spin_lock_irq(&wq_mayday_lock); |
493a1724 TH |
3261 | |
3262 | while (!list_empty(&wq->maydays)) { | |
3263 | struct pool_workqueue *pwq = list_first_entry(&wq->maydays, | |
3264 | struct pool_workqueue, mayday_node); | |
112202d9 | 3265 | struct worker_pool *pool = pwq->pool; |
e22bee78 TH |
3266 | struct work_struct *work, *n; |
3267 | ||
3268 | __set_current_state(TASK_RUNNING); | |
493a1724 TH |
3269 | list_del_init(&pwq->mayday_node); |
3270 | ||
a9b8a985 | 3271 | raw_spin_unlock_irq(&wq_mayday_lock); |
e22bee78 | 3272 | |
51697d39 LJ |
3273 | worker_attach_to_pool(rescuer, pool); |
3274 | ||
a9b8a985 | 3275 | raw_spin_lock_irq(&pool->lock); |
e22bee78 TH |
3276 | |
3277 | /* | |
3278 | * Slurp in all works issued via this workqueue and | |
3279 | * process'em. | |
3280 | */ | |
873eaca6 | 3281 | WARN_ON_ONCE(!list_empty(&rescuer->scheduled)); |
82607adc | 3282 | list_for_each_entry_safe(work, n, &pool->worklist, entry) { |
873eaca6 TH |
3283 | if (get_work_pwq(work) == pwq && |
3284 | assign_work(work, rescuer, &n)) | |
725e8ec5 | 3285 | pwq->stats[PWQ_STAT_RESCUED]++; |
82607adc | 3286 | } |
e22bee78 | 3287 | |
873eaca6 | 3288 | if (!list_empty(&rescuer->scheduled)) { |
008847f6 N |
3289 | process_scheduled_works(rescuer); |
3290 | ||
3291 | /* | |
3292 | * The above execution of rescued work items could | |
3293 | * have created more to rescue through | |
f97a4a1a | 3294 | * pwq_activate_first_inactive() or chained |
008847f6 N |
3295 | * queueing. Let's put @pwq back on mayday list so |
3296 | * that such back-to-back work items, which may be | |
3297 | * being used to relieve memory pressure, don't | |
3298 | * incur MAYDAY_INTERVAL delay inbetween. | |
3299 | */ | |
4f3f4cf3 | 3300 | if (pwq->nr_active && need_to_create_worker(pool)) { |
a9b8a985 | 3301 | raw_spin_lock(&wq_mayday_lock); |
e66b39af TH |
3302 | /* |
3303 | * Queue iff we aren't racing destruction | |
3304 | * and somebody else hasn't queued it already. | |
3305 | */ | |
3306 | if (wq->rescuer && list_empty(&pwq->mayday_node)) { | |
3307 | get_pwq(pwq); | |
3308 | list_add_tail(&pwq->mayday_node, &wq->maydays); | |
3309 | } | |
a9b8a985 | 3310 | raw_spin_unlock(&wq_mayday_lock); |
008847f6 N |
3311 | } |
3312 | } | |
7576958a | 3313 | |
77668c8b LJ |
3314 | /* |
3315 | * Put the reference grabbed by send_mayday(). @pool won't | |
13b1d625 | 3316 | * go away while we're still attached to it. |
77668c8b LJ |
3317 | */ |
3318 | put_pwq(pwq); | |
3319 | ||
7576958a | 3320 | /* |
0219a352 TH |
3321 | * Leave this pool. Notify regular workers; otherwise, we end up |
3322 | * with 0 concurrency and stalling the execution. | |
7576958a | 3323 | */ |
0219a352 | 3324 | kick_pool(pool); |
7576958a | 3325 | |
a9b8a985 | 3326 | raw_spin_unlock_irq(&pool->lock); |
13b1d625 | 3327 | |
a2d812a2 | 3328 | worker_detach_from_pool(rescuer); |
13b1d625 | 3329 | |
a9b8a985 | 3330 | raw_spin_lock_irq(&wq_mayday_lock); |
e22bee78 TH |
3331 | } |
3332 | ||
a9b8a985 | 3333 | raw_spin_unlock_irq(&wq_mayday_lock); |
493a1724 | 3334 | |
4d595b86 LJ |
3335 | if (should_stop) { |
3336 | __set_current_state(TASK_RUNNING); | |
197f6acc | 3337 | set_pf_worker(false); |
4d595b86 LJ |
3338 | return 0; |
3339 | } | |
3340 | ||
111c225a TH |
3341 | /* rescuers should never participate in concurrency management */ |
3342 | WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); | |
e22bee78 TH |
3343 | schedule(); |
3344 | goto repeat; | |
1da177e4 LT |
3345 | } |
3346 | ||
fca839c0 TH |
3347 | /** |
3348 | * check_flush_dependency - check for flush dependency sanity | |
3349 | * @target_wq: workqueue being flushed | |
3350 | * @target_work: work item being flushed (NULL for workqueue flushes) | |
3351 | * | |
3352 | * %current is trying to flush the whole @target_wq or @target_work on it. | |
3353 | * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not | |
3354 | * reclaiming memory or running on a workqueue which doesn't have | |
3355 | * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to | |
3356 | * a deadlock. | |
3357 | */ | |
3358 | static void check_flush_dependency(struct workqueue_struct *target_wq, | |
3359 | struct work_struct *target_work) | |
3360 | { | |
3361 | work_func_t target_func = target_work ? target_work->func : NULL; | |
3362 | struct worker *worker; | |
3363 | ||
3364 | if (target_wq->flags & WQ_MEM_RECLAIM) | |
3365 | return; | |
3366 | ||
3367 | worker = current_wq_worker(); | |
3368 | ||
3369 | WARN_ONCE(current->flags & PF_MEMALLOC, | |
d75f773c | 3370 | "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps", |
fca839c0 | 3371 | current->pid, current->comm, target_wq->name, target_func); |
23d11a58 TH |
3372 | WARN_ONCE(worker && ((worker->current_pwq->wq->flags & |
3373 | (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM), | |
d75f773c | 3374 | "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps", |
fca839c0 TH |
3375 | worker->current_pwq->wq->name, worker->current_func, |
3376 | target_wq->name, target_func); | |
3377 | } | |
3378 | ||
fc2e4d70 ON |
3379 | struct wq_barrier { |
3380 | struct work_struct work; | |
3381 | struct completion done; | |
2607d7a6 | 3382 | struct task_struct *task; /* purely informational */ |
fc2e4d70 ON |
3383 | }; |
3384 | ||
3385 | static void wq_barrier_func(struct work_struct *work) | |
3386 | { | |
3387 | struct wq_barrier *barr = container_of(work, struct wq_barrier, work); | |
3388 | complete(&barr->done); | |
3389 | } | |
3390 | ||
4690c4ab TH |
3391 | /** |
3392 | * insert_wq_barrier - insert a barrier work | |
112202d9 | 3393 | * @pwq: pwq to insert barrier into |
4690c4ab | 3394 | * @barr: wq_barrier to insert |
affee4b2 TH |
3395 | * @target: target work to attach @barr to |
3396 | * @worker: worker currently executing @target, NULL if @target is not executing | |
4690c4ab | 3397 | * |
affee4b2 TH |
3398 | * @barr is linked to @target such that @barr is completed only after |
3399 | * @target finishes execution. Please note that the ordering | |
3400 | * guarantee is observed only with respect to @target and on the local | |
3401 | * cpu. | |
3402 | * | |
3403 | * Currently, a queued barrier can't be canceled. This is because | |
3404 | * try_to_grab_pending() can't determine whether the work to be | |
3405 | * grabbed is at the head of the queue and thus can't clear LINKED | |
3406 | * flag of the previous work while there must be a valid next work | |
3407 | * after a work with LINKED flag set. | |
3408 | * | |
3409 | * Note that when @worker is non-NULL, @target may be modified | |
112202d9 | 3410 | * underneath us, so we can't reliably determine pwq from @target. |
4690c4ab TH |
3411 | * |
3412 | * CONTEXT: | |
a9b8a985 | 3413 | * raw_spin_lock_irq(pool->lock). |
4690c4ab | 3414 | */ |
112202d9 | 3415 | static void insert_wq_barrier(struct pool_workqueue *pwq, |
affee4b2 TH |
3416 | struct wq_barrier *barr, |
3417 | struct work_struct *target, struct worker *worker) | |
fc2e4d70 | 3418 | { |
d812796e LJ |
3419 | unsigned int work_flags = 0; |
3420 | unsigned int work_color; | |
affee4b2 | 3421 | struct list_head *head; |
affee4b2 | 3422 | |
dc186ad7 | 3423 | /* |
d565ed63 | 3424 | * debugobject calls are safe here even with pool->lock locked |
dc186ad7 TG |
3425 | * as we know for sure that this will not trigger any of the |
3426 | * checks and call back into the fixup functions where we | |
3427 | * might deadlock. | |
3428 | */ | |
ca1cab37 | 3429 | INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); |
22df02bb | 3430 | __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); |
52fa5bc5 | 3431 | |
fd1a5b04 BP |
3432 | init_completion_map(&barr->done, &target->lockdep_map); |
3433 | ||
2607d7a6 | 3434 | barr->task = current; |
83c22520 | 3435 | |
5797b1c1 | 3436 | /* The barrier work item does not participate in nr_active. */ |
018f3a13 LJ |
3437 | work_flags |= WORK_STRUCT_INACTIVE; |
3438 | ||
affee4b2 TH |
3439 | /* |
3440 | * If @target is currently being executed, schedule the | |
3441 | * barrier to the worker; otherwise, put it after @target. | |
3442 | */ | |
d812796e | 3443 | if (worker) { |
affee4b2 | 3444 | head = worker->scheduled.next; |
d812796e LJ |
3445 | work_color = worker->current_color; |
3446 | } else { | |
affee4b2 TH |
3447 | unsigned long *bits = work_data_bits(target); |
3448 | ||
3449 | head = target->entry.next; | |
3450 | /* there can already be other linked works, inherit and set */ | |
d21cece0 | 3451 | work_flags |= *bits & WORK_STRUCT_LINKED; |
d812796e | 3452 | work_color = get_work_color(*bits); |
affee4b2 TH |
3453 | __set_bit(WORK_STRUCT_LINKED_BIT, bits); |
3454 | } | |
3455 | ||
d812796e LJ |
3456 | pwq->nr_in_flight[work_color]++; |
3457 | work_flags |= work_color_to_flags(work_color); | |
3458 | ||
d21cece0 | 3459 | insert_work(pwq, &barr->work, head, work_flags); |
fc2e4d70 ON |
3460 | } |
3461 | ||
73f53c4a | 3462 | /** |
112202d9 | 3463 | * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing |
73f53c4a TH |
3464 | * @wq: workqueue being flushed |
3465 | * @flush_color: new flush color, < 0 for no-op | |
3466 | * @work_color: new work color, < 0 for no-op | |
3467 | * | |
112202d9 | 3468 | * Prepare pwqs for workqueue flushing. |
73f53c4a | 3469 | * |
112202d9 TH |
3470 | * If @flush_color is non-negative, flush_color on all pwqs should be |
3471 | * -1. If no pwq has in-flight commands at the specified color, all | |
3472 | * pwq->flush_color's stay at -1 and %false is returned. If any pwq | |
3473 | * has in flight commands, its pwq->flush_color is set to | |
3474 | * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq | |
73f53c4a TH |
3475 | * wakeup logic is armed and %true is returned. |
3476 | * | |
3477 | * The caller should have initialized @wq->first_flusher prior to | |
3478 | * calling this function with non-negative @flush_color. If | |
3479 | * @flush_color is negative, no flush color update is done and %false | |
3480 | * is returned. | |
3481 | * | |
112202d9 | 3482 | * If @work_color is non-negative, all pwqs should have the same |
73f53c4a TH |
3483 | * work_color which is previous to @work_color and all will be |
3484 | * advanced to @work_color. | |
3485 | * | |
3486 | * CONTEXT: | |
3c25a55d | 3487 | * mutex_lock(wq->mutex). |
73f53c4a | 3488 | * |
d185af30 | 3489 | * Return: |
73f53c4a TH |
3490 | * %true if @flush_color >= 0 and there's something to flush. %false |
3491 | * otherwise. | |
3492 | */ | |
112202d9 | 3493 | static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, |
73f53c4a | 3494 | int flush_color, int work_color) |
1da177e4 | 3495 | { |
73f53c4a | 3496 | bool wait = false; |
49e3cf44 | 3497 | struct pool_workqueue *pwq; |
1da177e4 | 3498 | |
73f53c4a | 3499 | if (flush_color >= 0) { |
6183c009 | 3500 | WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); |
112202d9 | 3501 | atomic_set(&wq->nr_pwqs_to_flush, 1); |
1da177e4 | 3502 | } |
2355b70f | 3503 | |
49e3cf44 | 3504 | for_each_pwq(pwq, wq) { |
112202d9 | 3505 | struct worker_pool *pool = pwq->pool; |
fc2e4d70 | 3506 | |
a9b8a985 | 3507 | raw_spin_lock_irq(&pool->lock); |
83c22520 | 3508 | |
73f53c4a | 3509 | if (flush_color >= 0) { |
6183c009 | 3510 | WARN_ON_ONCE(pwq->flush_color != -1); |
fc2e4d70 | 3511 | |
112202d9 TH |
3512 | if (pwq->nr_in_flight[flush_color]) { |
3513 | pwq->flush_color = flush_color; | |
3514 | atomic_inc(&wq->nr_pwqs_to_flush); | |
73f53c4a TH |
3515 | wait = true; |
3516 | } | |
3517 | } | |
1da177e4 | 3518 | |
73f53c4a | 3519 | if (work_color >= 0) { |
6183c009 | 3520 | WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); |
112202d9 | 3521 | pwq->work_color = work_color; |
73f53c4a | 3522 | } |
1da177e4 | 3523 | |
a9b8a985 | 3524 | raw_spin_unlock_irq(&pool->lock); |
1da177e4 | 3525 | } |
2355b70f | 3526 | |
112202d9 | 3527 | if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) |
73f53c4a | 3528 | complete(&wq->first_flusher->done); |
14441960 | 3529 | |
73f53c4a | 3530 | return wait; |
1da177e4 LT |
3531 | } |
3532 | ||
0fcb78c2 | 3533 | /** |
c4f135d6 | 3534 | * __flush_workqueue - ensure that any scheduled work has run to completion. |
0fcb78c2 | 3535 | * @wq: workqueue to flush |
1da177e4 | 3536 | * |
c5aa87bb TH |
3537 | * This function sleeps until all work items which were queued on entry |
3538 | * have finished execution, but it is not livelocked by new incoming ones. | |
1da177e4 | 3539 | */ |
c4f135d6 | 3540 | void __flush_workqueue(struct workqueue_struct *wq) |
1da177e4 | 3541 | { |
73f53c4a TH |
3542 | struct wq_flusher this_flusher = { |
3543 | .list = LIST_HEAD_INIT(this_flusher.list), | |
3544 | .flush_color = -1, | |
fd1a5b04 | 3545 | .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map), |
73f53c4a TH |
3546 | }; |
3547 | int next_color; | |
1da177e4 | 3548 | |
3347fa09 TH |
3549 | if (WARN_ON(!wq_online)) |
3550 | return; | |
3551 | ||
87915adc JB |
3552 | lock_map_acquire(&wq->lockdep_map); |
3553 | lock_map_release(&wq->lockdep_map); | |
3554 | ||
3c25a55d | 3555 | mutex_lock(&wq->mutex); |
73f53c4a TH |
3556 | |
3557 | /* | |
3558 | * Start-to-wait phase | |
3559 | */ | |
3560 | next_color = work_next_color(wq->work_color); | |
3561 | ||
3562 | if (next_color != wq->flush_color) { | |
3563 | /* | |
3564 | * Color space is not full. The current work_color | |
3565 | * becomes our flush_color and work_color is advanced | |
3566 | * by one. | |
3567 | */ | |
6183c009 | 3568 | WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); |
73f53c4a TH |
3569 | this_flusher.flush_color = wq->work_color; |
3570 | wq->work_color = next_color; | |
3571 | ||
3572 | if (!wq->first_flusher) { | |
3573 | /* no flush in progress, become the first flusher */ | |
6183c009 | 3574 | WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); |
73f53c4a TH |
3575 | |
3576 | wq->first_flusher = &this_flusher; | |
3577 | ||
112202d9 | 3578 | if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, |
73f53c4a TH |
3579 | wq->work_color)) { |
3580 | /* nothing to flush, done */ | |
3581 | wq->flush_color = next_color; | |
3582 | wq->first_flusher = NULL; | |
3583 | goto out_unlock; | |
3584 | } | |
3585 | } else { | |
3586 | /* wait in queue */ | |
6183c009 | 3587 | WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); |
73f53c4a | 3588 | list_add_tail(&this_flusher.list, &wq->flusher_queue); |
112202d9 | 3589 | flush_workqueue_prep_pwqs(wq, -1, wq->work_color); |
73f53c4a TH |
3590 | } |
3591 | } else { | |
3592 | /* | |
3593 | * Oops, color space is full, wait on overflow queue. | |
3594 | * The next flush completion will assign us | |
3595 | * flush_color and transfer to flusher_queue. | |
3596 | */ | |
3597 | list_add_tail(&this_flusher.list, &wq->flusher_overflow); | |
3598 | } | |
3599 | ||
fca839c0 TH |
3600 | check_flush_dependency(wq, NULL); |
3601 | ||
3c25a55d | 3602 | mutex_unlock(&wq->mutex); |
73f53c4a TH |
3603 | |
3604 | wait_for_completion(&this_flusher.done); | |
3605 | ||
3606 | /* | |
3607 | * Wake-up-and-cascade phase | |
3608 | * | |
3609 | * First flushers are responsible for cascading flushes and | |
3610 | * handling overflow. Non-first flushers can simply return. | |
3611 | */ | |
00d5d15b | 3612 | if (READ_ONCE(wq->first_flusher) != &this_flusher) |
73f53c4a TH |
3613 | return; |
3614 | ||
3c25a55d | 3615 | mutex_lock(&wq->mutex); |
73f53c4a | 3616 | |
4ce48b37 TH |
3617 | /* we might have raced, check again with mutex held */ |
3618 | if (wq->first_flusher != &this_flusher) | |
3619 | goto out_unlock; | |
3620 | ||
00d5d15b | 3621 | WRITE_ONCE(wq->first_flusher, NULL); |
73f53c4a | 3622 | |
6183c009 TH |
3623 | WARN_ON_ONCE(!list_empty(&this_flusher.list)); |
3624 | WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); | |
73f53c4a TH |
3625 | |
3626 | while (true) { | |
3627 | struct wq_flusher *next, *tmp; | |
3628 | ||
3629 | /* complete all the flushers sharing the current flush color */ | |
3630 | list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { | |
3631 | if (next->flush_color != wq->flush_color) | |
3632 | break; | |
3633 | list_del_init(&next->list); | |
3634 | complete(&next->done); | |
3635 | } | |
3636 | ||
6183c009 TH |
3637 | WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && |
3638 | wq->flush_color != work_next_color(wq->work_color)); | |
73f53c4a TH |
3639 | |
3640 | /* this flush_color is finished, advance by one */ | |
3641 | wq->flush_color = work_next_color(wq->flush_color); | |
3642 | ||
3643 | /* one color has been freed, handle overflow queue */ | |
3644 | if (!list_empty(&wq->flusher_overflow)) { | |
3645 | /* | |
3646 | * Assign the same color to all overflowed | |
3647 | * flushers, advance work_color and append to | |
3648 | * flusher_queue. This is the start-to-wait | |
3649 | * phase for these overflowed flushers. | |
3650 | */ | |
3651 | list_for_each_entry(tmp, &wq->flusher_overflow, list) | |
3652 | tmp->flush_color = wq->work_color; | |
3653 | ||
3654 | wq->work_color = work_next_color(wq->work_color); | |
3655 | ||
3656 | list_splice_tail_init(&wq->flusher_overflow, | |
3657 | &wq->flusher_queue); | |
112202d9 | 3658 | flush_workqueue_prep_pwqs(wq, -1, wq->work_color); |
73f53c4a TH |
3659 | } |
3660 | ||
3661 | if (list_empty(&wq->flusher_queue)) { | |
6183c009 | 3662 | WARN_ON_ONCE(wq->flush_color != wq->work_color); |
73f53c4a TH |
3663 | break; |
3664 | } | |
3665 | ||
3666 | /* | |
3667 | * Need to flush more colors. Make the next flusher | |
112202d9 | 3668 | * the new first flusher and arm pwqs. |
73f53c4a | 3669 | */ |
6183c009 TH |
3670 | WARN_ON_ONCE(wq->flush_color == wq->work_color); |
3671 | WARN_ON_ONCE(wq->flush_color != next->flush_color); | |
73f53c4a TH |
3672 | |
3673 | list_del_init(&next->list); | |
3674 | wq->first_flusher = next; | |
3675 | ||
112202d9 | 3676 | if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) |
73f53c4a TH |
3677 | break; |
3678 | ||
3679 | /* | |
3680 | * Meh... this color is already done, clear first | |
3681 | * flusher and repeat cascading. | |
3682 | */ | |
3683 | wq->first_flusher = NULL; | |
3684 | } | |
3685 | ||
3686 | out_unlock: | |
3c25a55d | 3687 | mutex_unlock(&wq->mutex); |
1da177e4 | 3688 | } |
c4f135d6 | 3689 | EXPORT_SYMBOL(__flush_workqueue); |
1da177e4 | 3690 | |
9c5a2ba7 TH |
3691 | /** |
3692 | * drain_workqueue - drain a workqueue | |
3693 | * @wq: workqueue to drain | |
3694 | * | |
3695 | * Wait until the workqueue becomes empty. While draining is in progress, | |
3696 | * only chain queueing is allowed. IOW, only currently pending or running | |
3697 | * work items on @wq can queue further work items on it. @wq is flushed | |
b749b1b6 | 3698 | * repeatedly until it becomes empty. The number of flushing is determined |
9c5a2ba7 TH |
3699 | * by the depth of chaining and should be relatively short. Whine if it |
3700 | * takes too long. | |
3701 | */ | |
3702 | void drain_workqueue(struct workqueue_struct *wq) | |
3703 | { | |
3704 | unsigned int flush_cnt = 0; | |
49e3cf44 | 3705 | struct pool_workqueue *pwq; |
9c5a2ba7 TH |
3706 | |
3707 | /* | |
3708 | * __queue_work() needs to test whether there are drainers, is much | |
3709 | * hotter than drain_workqueue() and already looks at @wq->flags. | |
618b01eb | 3710 | * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. |
9c5a2ba7 | 3711 | */ |
87fc741e | 3712 | mutex_lock(&wq->mutex); |
9c5a2ba7 | 3713 | if (!wq->nr_drainers++) |
618b01eb | 3714 | wq->flags |= __WQ_DRAINING; |
87fc741e | 3715 | mutex_unlock(&wq->mutex); |
9c5a2ba7 | 3716 | reflush: |
c4f135d6 | 3717 | __flush_workqueue(wq); |
9c5a2ba7 | 3718 | |
b09f4fd3 | 3719 | mutex_lock(&wq->mutex); |
76af4d93 | 3720 | |
49e3cf44 | 3721 | for_each_pwq(pwq, wq) { |
fa2563e4 | 3722 | bool drained; |
9c5a2ba7 | 3723 | |
a9b8a985 | 3724 | raw_spin_lock_irq(&pwq->pool->lock); |
afa87ce8 | 3725 | drained = pwq_is_empty(pwq); |
a9b8a985 | 3726 | raw_spin_unlock_irq(&pwq->pool->lock); |
fa2563e4 TT |
3727 | |
3728 | if (drained) | |
9c5a2ba7 TH |
3729 | continue; |
3730 | ||
3731 | if (++flush_cnt == 10 || | |
3732 | (flush_cnt % 100 == 0 && flush_cnt <= 1000)) | |
e9ad2eb3 SZ |
3733 | pr_warn("workqueue %s: %s() isn't complete after %u tries\n", |
3734 | wq->name, __func__, flush_cnt); | |
76af4d93 | 3735 | |
b09f4fd3 | 3736 | mutex_unlock(&wq->mutex); |
9c5a2ba7 TH |
3737 | goto reflush; |
3738 | } | |
3739 | ||
9c5a2ba7 | 3740 | if (!--wq->nr_drainers) |
618b01eb | 3741 | wq->flags &= ~__WQ_DRAINING; |
87fc741e | 3742 | mutex_unlock(&wq->mutex); |
9c5a2ba7 TH |
3743 | } |
3744 | EXPORT_SYMBOL_GPL(drain_workqueue); | |
3745 | ||
d6e89786 JB |
3746 | static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr, |
3747 | bool from_cancel) | |
db700897 | 3748 | { |
affee4b2 | 3749 | struct worker *worker = NULL; |
c9e7cf27 | 3750 | struct worker_pool *pool; |
112202d9 | 3751 | struct pool_workqueue *pwq; |
db700897 ON |
3752 | |
3753 | might_sleep(); | |
fa1b54e6 | 3754 | |
24acfb71 | 3755 | rcu_read_lock(); |
c9e7cf27 | 3756 | pool = get_work_pool(work); |
fa1b54e6 | 3757 | if (!pool) { |
24acfb71 | 3758 | rcu_read_unlock(); |
baf59022 | 3759 | return false; |
fa1b54e6 | 3760 | } |
db700897 | 3761 | |
a9b8a985 | 3762 | raw_spin_lock_irq(&pool->lock); |
0b3dae68 | 3763 | /* see the comment in try_to_grab_pending() with the same code */ |
112202d9 TH |
3764 | pwq = get_work_pwq(work); |
3765 | if (pwq) { | |
3766 | if (unlikely(pwq->pool != pool)) | |
4690c4ab | 3767 | goto already_gone; |
606a5020 | 3768 | } else { |
c9e7cf27 | 3769 | worker = find_worker_executing_work(pool, work); |
affee4b2 | 3770 | if (!worker) |
4690c4ab | 3771 | goto already_gone; |
112202d9 | 3772 | pwq = worker->current_pwq; |
606a5020 | 3773 | } |
db700897 | 3774 | |
fca839c0 TH |
3775 | check_flush_dependency(pwq->wq, work); |
3776 | ||
112202d9 | 3777 | insert_wq_barrier(pwq, barr, work, worker); |
a9b8a985 | 3778 | raw_spin_unlock_irq(&pool->lock); |
7a22ad75 | 3779 | |
e159489b | 3780 | /* |
a1d14934 PZ |
3781 | * Force a lock recursion deadlock when using flush_work() inside a |
3782 | * single-threaded or rescuer equipped workqueue. | |
3783 | * | |
3784 | * For single threaded workqueues the deadlock happens when the work | |
3785 | * is after the work issuing the flush_work(). For rescuer equipped | |
3786 | * workqueues the deadlock happens when the rescuer stalls, blocking | |
3787 | * forward progress. | |
e159489b | 3788 | */ |
d6e89786 JB |
3789 | if (!from_cancel && |
3790 | (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) { | |
112202d9 | 3791 | lock_map_acquire(&pwq->wq->lockdep_map); |
a1d14934 PZ |
3792 | lock_map_release(&pwq->wq->lockdep_map); |
3793 | } | |
24acfb71 | 3794 | rcu_read_unlock(); |
401a8d04 | 3795 | return true; |
4690c4ab | 3796 | already_gone: |
a9b8a985 | 3797 | raw_spin_unlock_irq(&pool->lock); |
24acfb71 | 3798 | rcu_read_unlock(); |
401a8d04 | 3799 | return false; |
db700897 | 3800 | } |
baf59022 | 3801 | |
d6e89786 JB |
3802 | static bool __flush_work(struct work_struct *work, bool from_cancel) |
3803 | { | |
3804 | struct wq_barrier barr; | |
3805 | ||
3806 | if (WARN_ON(!wq_online)) | |
3807 | return false; | |
3808 | ||
4d43d395 TH |
3809 | if (WARN_ON(!work->func)) |
3810 | return false; | |
3811 | ||
c0feea59 TH |
3812 | lock_map_acquire(&work->lockdep_map); |
3813 | lock_map_release(&work->lockdep_map); | |
87915adc | 3814 | |
d6e89786 JB |
3815 | if (start_flush_work(work, &barr, from_cancel)) { |
3816 | wait_for_completion(&barr.done); | |
3817 | destroy_work_on_stack(&barr.work); | |
3818 | return true; | |
3819 | } else { | |
3820 | return false; | |
3821 | } | |
3822 | } | |
3823 | ||
baf59022 TH |
3824 | /** |
3825 | * flush_work - wait for a work to finish executing the last queueing instance | |
3826 | * @work: the work to flush | |
3827 | * | |
606a5020 TH |
3828 | * Wait until @work has finished execution. @work is guaranteed to be idle |
3829 | * on return if it hasn't been requeued since flush started. | |
baf59022 | 3830 | * |
d185af30 | 3831 | * Return: |
baf59022 TH |
3832 | * %true if flush_work() waited for the work to finish execution, |
3833 | * %false if it was already idle. | |
3834 | */ | |
3835 | bool flush_work(struct work_struct *work) | |
3836 | { | |
d6e89786 | 3837 | return __flush_work(work, false); |
6e84d644 | 3838 | } |
606a5020 | 3839 | EXPORT_SYMBOL_GPL(flush_work); |
6e84d644 | 3840 | |
8603e1b3 | 3841 | struct cwt_wait { |
ac6424b9 | 3842 | wait_queue_entry_t wait; |
8603e1b3 TH |
3843 | struct work_struct *work; |
3844 | }; | |
3845 | ||
ac6424b9 | 3846 | static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key) |
8603e1b3 TH |
3847 | { |
3848 | struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait); | |
3849 | ||
3850 | if (cwait->work != key) | |
3851 | return 0; | |
3852 | return autoremove_wake_function(wait, mode, sync, key); | |
3853 | } | |
3854 | ||
36e227d2 | 3855 | static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) |
1f1f642e | 3856 | { |
8603e1b3 | 3857 | static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq); |
bbb68dfa | 3858 | unsigned long flags; |
1f1f642e ON |
3859 | int ret; |
3860 | ||
3861 | do { | |
bbb68dfa TH |
3862 | ret = try_to_grab_pending(work, is_dwork, &flags); |
3863 | /* | |
8603e1b3 TH |
3864 | * If someone else is already canceling, wait for it to |
3865 | * finish. flush_work() doesn't work for PREEMPT_NONE | |
3866 | * because we may get scheduled between @work's completion | |
3867 | * and the other canceling task resuming and clearing | |
3868 | * CANCELING - flush_work() will return false immediately | |
3869 | * as @work is no longer busy, try_to_grab_pending() will | |
3870 | * return -ENOENT as @work is still being canceled and the | |
3871 | * other canceling task won't be able to clear CANCELING as | |
3872 | * we're hogging the CPU. | |
3873 | * | |
3874 | * Let's wait for completion using a waitqueue. As this | |
3875 | * may lead to the thundering herd problem, use a custom | |
3876 | * wake function which matches @work along with exclusive | |
3877 | * wait and wakeup. | |
bbb68dfa | 3878 | */ |
8603e1b3 TH |
3879 | if (unlikely(ret == -ENOENT)) { |
3880 | struct cwt_wait cwait; | |
3881 | ||
3882 | init_wait(&cwait.wait); | |
3883 | cwait.wait.func = cwt_wakefn; | |
3884 | cwait.work = work; | |
3885 | ||
3886 | prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait, | |
3887 | TASK_UNINTERRUPTIBLE); | |
3888 | if (work_is_canceling(work)) | |
3889 | schedule(); | |
3890 | finish_wait(&cancel_waitq, &cwait.wait); | |
3891 | } | |
1f1f642e ON |
3892 | } while (unlikely(ret < 0)); |
3893 | ||
bbb68dfa TH |
3894 | /* tell other tasks trying to grab @work to back off */ |
3895 | mark_work_canceling(work); | |
3896 | local_irq_restore(flags); | |
3897 | ||
3347fa09 TH |
3898 | /* |
3899 | * This allows canceling during early boot. We know that @work | |
3900 | * isn't executing. | |
3901 | */ | |
3902 | if (wq_online) | |
d6e89786 | 3903 | __flush_work(work, true); |
3347fa09 | 3904 | |
7a22ad75 | 3905 | clear_work_data(work); |
8603e1b3 TH |
3906 | |
3907 | /* | |
3908 | * Paired with prepare_to_wait() above so that either | |
3909 | * waitqueue_active() is visible here or !work_is_canceling() is | |
3910 | * visible there. | |
3911 | */ | |
3912 | smp_mb(); | |
3913 | if (waitqueue_active(&cancel_waitq)) | |
3914 | __wake_up(&cancel_waitq, TASK_NORMAL, 1, work); | |
3915 | ||
1f1f642e ON |
3916 | return ret; |
3917 | } | |
3918 | ||
6e84d644 | 3919 | /** |
401a8d04 TH |
3920 | * cancel_work_sync - cancel a work and wait for it to finish |
3921 | * @work: the work to cancel | |
6e84d644 | 3922 | * |
401a8d04 TH |
3923 | * Cancel @work and wait for its execution to finish. This function |
3924 | * can be used even if the work re-queues itself or migrates to | |
3925 | * another workqueue. On return from this function, @work is | |
3926 | * guaranteed to be not pending or executing on any CPU. | |
1f1f642e | 3927 | * |
401a8d04 TH |
3928 | * cancel_work_sync(&delayed_work->work) must not be used for |
3929 | * delayed_work's. Use cancel_delayed_work_sync() instead. | |
6e84d644 | 3930 | * |
401a8d04 | 3931 | * The caller must ensure that the workqueue on which @work was last |
6e84d644 | 3932 | * queued can't be destroyed before this function returns. |
401a8d04 | 3933 | * |
d185af30 | 3934 | * Return: |
401a8d04 | 3935 | * %true if @work was pending, %false otherwise. |
6e84d644 | 3936 | */ |
401a8d04 | 3937 | bool cancel_work_sync(struct work_struct *work) |
6e84d644 | 3938 | { |
36e227d2 | 3939 | return __cancel_work_timer(work, false); |
b89deed3 | 3940 | } |
28e53bdd | 3941 | EXPORT_SYMBOL_GPL(cancel_work_sync); |
b89deed3 | 3942 | |
6e84d644 | 3943 | /** |
401a8d04 TH |
3944 | * flush_delayed_work - wait for a dwork to finish executing the last queueing |
3945 | * @dwork: the delayed work to flush | |
6e84d644 | 3946 | * |
401a8d04 TH |
3947 | * Delayed timer is cancelled and the pending work is queued for |
3948 | * immediate execution. Like flush_work(), this function only | |
3949 | * considers the last queueing instance of @dwork. | |
1f1f642e | 3950 | * |
d185af30 | 3951 | * Return: |
401a8d04 TH |
3952 | * %true if flush_work() waited for the work to finish execution, |
3953 | * %false if it was already idle. | |
6e84d644 | 3954 | */ |
401a8d04 TH |
3955 | bool flush_delayed_work(struct delayed_work *dwork) |
3956 | { | |
8930caba | 3957 | local_irq_disable(); |
401a8d04 | 3958 | if (del_timer_sync(&dwork->timer)) |
60c057bc | 3959 | __queue_work(dwork->cpu, dwork->wq, &dwork->work); |
8930caba | 3960 | local_irq_enable(); |
401a8d04 TH |
3961 | return flush_work(&dwork->work); |
3962 | } | |
3963 | EXPORT_SYMBOL(flush_delayed_work); | |
3964 | ||
05f0fe6b TH |
3965 | /** |
3966 | * flush_rcu_work - wait for a rwork to finish executing the last queueing | |
3967 | * @rwork: the rcu work to flush | |
3968 | * | |
3969 | * Return: | |
3970 | * %true if flush_rcu_work() waited for the work to finish execution, | |
3971 | * %false if it was already idle. | |
3972 | */ | |
3973 | bool flush_rcu_work(struct rcu_work *rwork) | |
3974 | { | |
3975 | if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) { | |
3976 | rcu_barrier(); | |
3977 | flush_work(&rwork->work); | |
3978 | return true; | |
3979 | } else { | |
3980 | return flush_work(&rwork->work); | |
3981 | } | |
3982 | } | |
3983 | EXPORT_SYMBOL(flush_rcu_work); | |
3984 | ||
f72b8792 JA |
3985 | static bool __cancel_work(struct work_struct *work, bool is_dwork) |
3986 | { | |
3987 | unsigned long flags; | |
3988 | int ret; | |
3989 | ||
3990 | do { | |
3991 | ret = try_to_grab_pending(work, is_dwork, &flags); | |
3992 | } while (unlikely(ret == -EAGAIN)); | |
3993 | ||
3994 | if (unlikely(ret < 0)) | |
3995 | return false; | |
3996 | ||
3997 | set_work_pool_and_clear_pending(work, get_work_pool_id(work)); | |
3998 | local_irq_restore(flags); | |
3999 | return ret; | |
4000 | } | |
4001 | ||
73b4b532 AG |
4002 | /* |
4003 | * See cancel_delayed_work() | |
4004 | */ | |
4005 | bool cancel_work(struct work_struct *work) | |
4006 | { | |
4007 | return __cancel_work(work, false); | |
4008 | } | |
4009 | EXPORT_SYMBOL(cancel_work); | |
4010 | ||
09383498 | 4011 | /** |
57b30ae7 TH |
4012 | * cancel_delayed_work - cancel a delayed work |
4013 | * @dwork: delayed_work to cancel | |
09383498 | 4014 | * |
d185af30 YB |
4015 | * Kill off a pending delayed_work. |
4016 | * | |
4017 | * Return: %true if @dwork was pending and canceled; %false if it wasn't | |
4018 | * pending. | |
4019 | * | |
4020 | * Note: | |
4021 | * The work callback function may still be running on return, unless | |
4022 | * it returns %true and the work doesn't re-arm itself. Explicitly flush or | |
4023 | * use cancel_delayed_work_sync() to wait on it. | |
09383498 | 4024 | * |
57b30ae7 | 4025 | * This function is safe to call from any context including IRQ handler. |
09383498 | 4026 | */ |
57b30ae7 | 4027 | bool cancel_delayed_work(struct delayed_work *dwork) |
09383498 | 4028 | { |
f72b8792 | 4029 | return __cancel_work(&dwork->work, true); |
09383498 | 4030 | } |
57b30ae7 | 4031 | EXPORT_SYMBOL(cancel_delayed_work); |
09383498 | 4032 | |
401a8d04 TH |
4033 | /** |
4034 | * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish | |
4035 | * @dwork: the delayed work cancel | |
4036 | * | |
4037 | * This is cancel_work_sync() for delayed works. | |
4038 | * | |
d185af30 | 4039 | * Return: |
401a8d04 TH |
4040 | * %true if @dwork was pending, %false otherwise. |
4041 | */ | |
4042 | bool cancel_delayed_work_sync(struct delayed_work *dwork) | |
6e84d644 | 4043 | { |
36e227d2 | 4044 | return __cancel_work_timer(&dwork->work, true); |
6e84d644 | 4045 | } |
f5a421a4 | 4046 | EXPORT_SYMBOL(cancel_delayed_work_sync); |
1da177e4 | 4047 | |
b6136773 | 4048 | /** |
31ddd871 | 4049 | * schedule_on_each_cpu - execute a function synchronously on each online CPU |
b6136773 | 4050 | * @func: the function to call |
b6136773 | 4051 | * |
31ddd871 TH |
4052 | * schedule_on_each_cpu() executes @func on each online CPU using the |
4053 | * system workqueue and blocks until all CPUs have completed. | |
b6136773 | 4054 | * schedule_on_each_cpu() is very slow. |
31ddd871 | 4055 | * |
d185af30 | 4056 | * Return: |
31ddd871 | 4057 | * 0 on success, -errno on failure. |
b6136773 | 4058 | */ |
65f27f38 | 4059 | int schedule_on_each_cpu(work_func_t func) |
15316ba8 CL |
4060 | { |
4061 | int cpu; | |
38f51568 | 4062 | struct work_struct __percpu *works; |
15316ba8 | 4063 | |
b6136773 AM |
4064 | works = alloc_percpu(struct work_struct); |
4065 | if (!works) | |
15316ba8 | 4066 | return -ENOMEM; |
b6136773 | 4067 | |
ffd8bea8 | 4068 | cpus_read_lock(); |
93981800 | 4069 | |
15316ba8 | 4070 | for_each_online_cpu(cpu) { |
9bfb1839 IM |
4071 | struct work_struct *work = per_cpu_ptr(works, cpu); |
4072 | ||
4073 | INIT_WORK(work, func); | |
b71ab8c2 | 4074 | schedule_work_on(cpu, work); |
65a64464 | 4075 | } |
93981800 TH |
4076 | |
4077 | for_each_online_cpu(cpu) | |
4078 | flush_work(per_cpu_ptr(works, cpu)); | |
4079 | ||
ffd8bea8 | 4080 | cpus_read_unlock(); |
b6136773 | 4081 | free_percpu(works); |
15316ba8 CL |
4082 | return 0; |
4083 | } | |
4084 | ||
1fa44eca JB |
4085 | /** |
4086 | * execute_in_process_context - reliably execute the routine with user context | |
4087 | * @fn: the function to execute | |
1fa44eca JB |
4088 | * @ew: guaranteed storage for the execute work structure (must |
4089 | * be available when the work executes) | |
4090 | * | |
4091 | * Executes the function immediately if process context is available, | |
4092 | * otherwise schedules the function for delayed execution. | |
4093 | * | |
d185af30 | 4094 | * Return: 0 - function was executed |
1fa44eca JB |
4095 | * 1 - function was scheduled for execution |
4096 | */ | |
65f27f38 | 4097 | int execute_in_process_context(work_func_t fn, struct execute_work *ew) |
1fa44eca JB |
4098 | { |
4099 | if (!in_interrupt()) { | |
65f27f38 | 4100 | fn(&ew->work); |
1fa44eca JB |
4101 | return 0; |
4102 | } | |
4103 | ||
65f27f38 | 4104 | INIT_WORK(&ew->work, fn); |
1fa44eca JB |
4105 | schedule_work(&ew->work); |
4106 | ||
4107 | return 1; | |
4108 | } | |
4109 | EXPORT_SYMBOL_GPL(execute_in_process_context); | |
4110 | ||
6ba94429 FW |
4111 | /** |
4112 | * free_workqueue_attrs - free a workqueue_attrs | |
4113 | * @attrs: workqueue_attrs to free | |
226223ab | 4114 | * |
6ba94429 | 4115 | * Undo alloc_workqueue_attrs(). |
226223ab | 4116 | */ |
513c98d0 | 4117 | void free_workqueue_attrs(struct workqueue_attrs *attrs) |
226223ab | 4118 | { |
6ba94429 FW |
4119 | if (attrs) { |
4120 | free_cpumask_var(attrs->cpumask); | |
9546b29e | 4121 | free_cpumask_var(attrs->__pod_cpumask); |
6ba94429 FW |
4122 | kfree(attrs); |
4123 | } | |
226223ab TH |
4124 | } |
4125 | ||
6ba94429 FW |
4126 | /** |
4127 | * alloc_workqueue_attrs - allocate a workqueue_attrs | |
6ba94429 FW |
4128 | * |
4129 | * Allocate a new workqueue_attrs, initialize with default settings and | |
4130 | * return it. | |
4131 | * | |
4132 | * Return: The allocated new workqueue_attr on success. %NULL on failure. | |
4133 | */ | |
513c98d0 | 4134 | struct workqueue_attrs *alloc_workqueue_attrs(void) |
226223ab | 4135 | { |
6ba94429 | 4136 | struct workqueue_attrs *attrs; |
226223ab | 4137 | |
be69d00d | 4138 | attrs = kzalloc(sizeof(*attrs), GFP_KERNEL); |
6ba94429 FW |
4139 | if (!attrs) |
4140 | goto fail; | |
be69d00d | 4141 | if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL)) |
6ba94429 | 4142 | goto fail; |
9546b29e TH |
4143 | if (!alloc_cpumask_var(&attrs->__pod_cpumask, GFP_KERNEL)) |
4144 | goto fail; | |
6ba94429 FW |
4145 | |
4146 | cpumask_copy(attrs->cpumask, cpu_possible_mask); | |
523a301e | 4147 | attrs->affn_scope = WQ_AFFN_DFL; |
6ba94429 FW |
4148 | return attrs; |
4149 | fail: | |
4150 | free_workqueue_attrs(attrs); | |
4151 | return NULL; | |
226223ab TH |
4152 | } |
4153 | ||
6ba94429 FW |
4154 | static void copy_workqueue_attrs(struct workqueue_attrs *to, |
4155 | const struct workqueue_attrs *from) | |
226223ab | 4156 | { |
6ba94429 FW |
4157 | to->nice = from->nice; |
4158 | cpumask_copy(to->cpumask, from->cpumask); | |
9546b29e | 4159 | cpumask_copy(to->__pod_cpumask, from->__pod_cpumask); |
8639eceb | 4160 | to->affn_strict = from->affn_strict; |
84193c07 | 4161 | |
6ba94429 | 4162 | /* |
84193c07 TH |
4163 | * Unlike hash and equality test, copying shouldn't ignore wq-only |
4164 | * fields as copying is used for both pool and wq attrs. Instead, | |
4165 | * get_unbound_pool() explicitly clears the fields. | |
6ba94429 | 4166 | */ |
84193c07 | 4167 | to->affn_scope = from->affn_scope; |
af73f5c9 | 4168 | to->ordered = from->ordered; |
226223ab TH |
4169 | } |
4170 | ||
5de7a03c TH |
4171 | /* |
4172 | * Some attrs fields are workqueue-only. Clear them for worker_pool's. See the | |
4173 | * comments in 'struct workqueue_attrs' definition. | |
4174 | */ | |
4175 | static void wqattrs_clear_for_pool(struct workqueue_attrs *attrs) | |
4176 | { | |
84193c07 | 4177 | attrs->affn_scope = WQ_AFFN_NR_TYPES; |
5de7a03c TH |
4178 | attrs->ordered = false; |
4179 | } | |
4180 | ||
6ba94429 FW |
4181 | /* hash value of the content of @attr */ |
4182 | static u32 wqattrs_hash(const struct workqueue_attrs *attrs) | |
226223ab | 4183 | { |
6ba94429 | 4184 | u32 hash = 0; |
226223ab | 4185 | |
6ba94429 FW |
4186 | hash = jhash_1word(attrs->nice, hash); |
4187 | hash = jhash(cpumask_bits(attrs->cpumask), | |
4188 | BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); | |
9546b29e TH |
4189 | hash = jhash(cpumask_bits(attrs->__pod_cpumask), |
4190 | BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); | |
8639eceb | 4191 | hash = jhash_1word(attrs->affn_strict, hash); |
6ba94429 | 4192 | return hash; |
226223ab | 4193 | } |
226223ab | 4194 | |
6ba94429 FW |
4195 | /* content equality test */ |
4196 | static bool wqattrs_equal(const struct workqueue_attrs *a, | |
4197 | const struct workqueue_attrs *b) | |
226223ab | 4198 | { |
6ba94429 FW |
4199 | if (a->nice != b->nice) |
4200 | return false; | |
4201 | if (!cpumask_equal(a->cpumask, b->cpumask)) | |
4202 | return false; | |
9546b29e TH |
4203 | if (!cpumask_equal(a->__pod_cpumask, b->__pod_cpumask)) |
4204 | return false; | |
8639eceb TH |
4205 | if (a->affn_strict != b->affn_strict) |
4206 | return false; | |
6ba94429 | 4207 | return true; |
226223ab TH |
4208 | } |
4209 | ||
0f36ee24 TH |
4210 | /* Update @attrs with actually available CPUs */ |
4211 | static void wqattrs_actualize_cpumask(struct workqueue_attrs *attrs, | |
4212 | const cpumask_t *unbound_cpumask) | |
4213 | { | |
4214 | /* | |
4215 | * Calculate the effective CPU mask of @attrs given @unbound_cpumask. If | |
4216 | * @attrs->cpumask doesn't overlap with @unbound_cpumask, we fallback to | |
4217 | * @unbound_cpumask. | |
4218 | */ | |
4219 | cpumask_and(attrs->cpumask, attrs->cpumask, unbound_cpumask); | |
4220 | if (unlikely(cpumask_empty(attrs->cpumask))) | |
4221 | cpumask_copy(attrs->cpumask, unbound_cpumask); | |
4222 | } | |
4223 | ||
84193c07 TH |
4224 | /* find wq_pod_type to use for @attrs */ |
4225 | static const struct wq_pod_type * | |
4226 | wqattrs_pod_type(const struct workqueue_attrs *attrs) | |
4227 | { | |
523a301e TH |
4228 | enum wq_affn_scope scope; |
4229 | struct wq_pod_type *pt; | |
4230 | ||
4231 | /* to synchronize access to wq_affn_dfl */ | |
4232 | lockdep_assert_held(&wq_pool_mutex); | |
4233 | ||
4234 | if (attrs->affn_scope == WQ_AFFN_DFL) | |
4235 | scope = wq_affn_dfl; | |
4236 | else | |
4237 | scope = attrs->affn_scope; | |
4238 | ||
4239 | pt = &wq_pod_types[scope]; | |
84193c07 TH |
4240 | |
4241 | if (!WARN_ON_ONCE(attrs->affn_scope == WQ_AFFN_NR_TYPES) && | |
4242 | likely(pt->nr_pods)) | |
4243 | return pt; | |
4244 | ||
4245 | /* | |
4246 | * Before workqueue_init_topology(), only SYSTEM is available which is | |
4247 | * initialized in workqueue_init_early(). | |
4248 | */ | |
4249 | pt = &wq_pod_types[WQ_AFFN_SYSTEM]; | |
4250 | BUG_ON(!pt->nr_pods); | |
4251 | return pt; | |
4252 | } | |
4253 | ||
6ba94429 FW |
4254 | /** |
4255 | * init_worker_pool - initialize a newly zalloc'd worker_pool | |
4256 | * @pool: worker_pool to initialize | |
4257 | * | |
402dd89d | 4258 | * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs. |
6ba94429 FW |
4259 | * |
4260 | * Return: 0 on success, -errno on failure. Even on failure, all fields | |
4261 | * inside @pool proper are initialized and put_unbound_pool() can be called | |
4262 | * on @pool safely to release it. | |
4263 | */ | |
4264 | static int init_worker_pool(struct worker_pool *pool) | |
226223ab | 4265 | { |
a9b8a985 | 4266 | raw_spin_lock_init(&pool->lock); |
6ba94429 FW |
4267 | pool->id = -1; |
4268 | pool->cpu = -1; | |
4269 | pool->node = NUMA_NO_NODE; | |
4270 | pool->flags |= POOL_DISASSOCIATED; | |
82607adc | 4271 | pool->watchdog_ts = jiffies; |
6ba94429 FW |
4272 | INIT_LIST_HEAD(&pool->worklist); |
4273 | INIT_LIST_HEAD(&pool->idle_list); | |
4274 | hash_init(pool->busy_hash); | |
226223ab | 4275 | |
32a6c723 | 4276 | timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE); |
3f959aa3 | 4277 | INIT_WORK(&pool->idle_cull_work, idle_cull_fn); |
226223ab | 4278 | |
32a6c723 | 4279 | timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0); |
226223ab | 4280 | |
6ba94429 | 4281 | INIT_LIST_HEAD(&pool->workers); |
e02b9312 | 4282 | INIT_LIST_HEAD(&pool->dying_workers); |
226223ab | 4283 | |
6ba94429 FW |
4284 | ida_init(&pool->worker_ida); |
4285 | INIT_HLIST_NODE(&pool->hash_node); | |
4286 | pool->refcnt = 1; | |
226223ab | 4287 | |
6ba94429 | 4288 | /* shouldn't fail above this point */ |
be69d00d | 4289 | pool->attrs = alloc_workqueue_attrs(); |
6ba94429 FW |
4290 | if (!pool->attrs) |
4291 | return -ENOMEM; | |
5de7a03c TH |
4292 | |
4293 | wqattrs_clear_for_pool(pool->attrs); | |
4294 | ||
6ba94429 | 4295 | return 0; |
226223ab TH |
4296 | } |
4297 | ||
669de8bd BVA |
4298 | #ifdef CONFIG_LOCKDEP |
4299 | static void wq_init_lockdep(struct workqueue_struct *wq) | |
4300 | { | |
4301 | char *lock_name; | |
4302 | ||
4303 | lockdep_register_key(&wq->key); | |
4304 | lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name); | |
4305 | if (!lock_name) | |
4306 | lock_name = wq->name; | |
69a106c0 QC |
4307 | |
4308 | wq->lock_name = lock_name; | |
669de8bd BVA |
4309 | lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0); |
4310 | } | |
4311 | ||
4312 | static void wq_unregister_lockdep(struct workqueue_struct *wq) | |
4313 | { | |
4314 | lockdep_unregister_key(&wq->key); | |
4315 | } | |
4316 | ||
4317 | static void wq_free_lockdep(struct workqueue_struct *wq) | |
4318 | { | |
4319 | if (wq->lock_name != wq->name) | |
4320 | kfree(wq->lock_name); | |
4321 | } | |
4322 | #else | |
4323 | static void wq_init_lockdep(struct workqueue_struct *wq) | |
4324 | { | |
4325 | } | |
4326 | ||
4327 | static void wq_unregister_lockdep(struct workqueue_struct *wq) | |
4328 | { | |
4329 | } | |
4330 | ||
4331 | static void wq_free_lockdep(struct workqueue_struct *wq) | |
4332 | { | |
4333 | } | |
4334 | #endif | |
4335 | ||
91ccc6e7 TH |
4336 | static void free_node_nr_active(struct wq_node_nr_active **nna_ar) |
4337 | { | |
4338 | int node; | |
4339 | ||
4340 | for_each_node(node) { | |
4341 | kfree(nna_ar[node]); | |
4342 | nna_ar[node] = NULL; | |
4343 | } | |
4344 | ||
4345 | kfree(nna_ar[nr_node_ids]); | |
4346 | nna_ar[nr_node_ids] = NULL; | |
4347 | } | |
4348 | ||
4349 | static void init_node_nr_active(struct wq_node_nr_active *nna) | |
4350 | { | |
4351 | atomic_set(&nna->nr, 0); | |
5797b1c1 TH |
4352 | raw_spin_lock_init(&nna->lock); |
4353 | INIT_LIST_HEAD(&nna->pending_pwqs); | |
91ccc6e7 TH |
4354 | } |
4355 | ||
4356 | /* | |
4357 | * Each node's nr_active counter will be accessed mostly from its own node and | |
4358 | * should be allocated in the node. | |
4359 | */ | |
4360 | static int alloc_node_nr_active(struct wq_node_nr_active **nna_ar) | |
4361 | { | |
4362 | struct wq_node_nr_active *nna; | |
4363 | int node; | |
4364 | ||
4365 | for_each_node(node) { | |
4366 | nna = kzalloc_node(sizeof(*nna), GFP_KERNEL, node); | |
4367 | if (!nna) | |
4368 | goto err_free; | |
4369 | init_node_nr_active(nna); | |
4370 | nna_ar[node] = nna; | |
4371 | } | |
4372 | ||
4373 | /* [nr_node_ids] is used as the fallback */ | |
4374 | nna = kzalloc_node(sizeof(*nna), GFP_KERNEL, NUMA_NO_NODE); | |
4375 | if (!nna) | |
4376 | goto err_free; | |
4377 | init_node_nr_active(nna); | |
4378 | nna_ar[nr_node_ids] = nna; | |
4379 | ||
4380 | return 0; | |
4381 | ||
4382 | err_free: | |
4383 | free_node_nr_active(nna_ar); | |
4384 | return -ENOMEM; | |
4385 | } | |
4386 | ||
6ba94429 | 4387 | static void rcu_free_wq(struct rcu_head *rcu) |
226223ab | 4388 | { |
6ba94429 FW |
4389 | struct workqueue_struct *wq = |
4390 | container_of(rcu, struct workqueue_struct, rcu); | |
226223ab | 4391 | |
91ccc6e7 TH |
4392 | if (wq->flags & WQ_UNBOUND) |
4393 | free_node_nr_active(wq->node_nr_active); | |
4394 | ||
669de8bd | 4395 | wq_free_lockdep(wq); |
636b927e TH |
4396 | free_percpu(wq->cpu_pwq); |
4397 | free_workqueue_attrs(wq->unbound_attrs); | |
6ba94429 | 4398 | kfree(wq); |
226223ab TH |
4399 | } |
4400 | ||
6ba94429 | 4401 | static void rcu_free_pool(struct rcu_head *rcu) |
226223ab | 4402 | { |
6ba94429 | 4403 | struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); |
226223ab | 4404 | |
6ba94429 FW |
4405 | ida_destroy(&pool->worker_ida); |
4406 | free_workqueue_attrs(pool->attrs); | |
4407 | kfree(pool); | |
226223ab TH |
4408 | } |
4409 | ||
6ba94429 FW |
4410 | /** |
4411 | * put_unbound_pool - put a worker_pool | |
4412 | * @pool: worker_pool to put | |
4413 | * | |
24acfb71 | 4414 | * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU |
6ba94429 FW |
4415 | * safe manner. get_unbound_pool() calls this function on its failure path |
4416 | * and this function should be able to release pools which went through, | |
4417 | * successfully or not, init_worker_pool(). | |
4418 | * | |
4419 | * Should be called with wq_pool_mutex held. | |
4420 | */ | |
4421 | static void put_unbound_pool(struct worker_pool *pool) | |
226223ab | 4422 | { |
6ba94429 FW |
4423 | DECLARE_COMPLETION_ONSTACK(detach_completion); |
4424 | struct worker *worker; | |
9680540c | 4425 | LIST_HEAD(cull_list); |
e02b9312 | 4426 | |
6ba94429 | 4427 | lockdep_assert_held(&wq_pool_mutex); |
226223ab | 4428 | |
6ba94429 FW |
4429 | if (--pool->refcnt) |
4430 | return; | |
226223ab | 4431 | |
6ba94429 FW |
4432 | /* sanity checks */ |
4433 | if (WARN_ON(!(pool->cpu < 0)) || | |
4434 | WARN_ON(!list_empty(&pool->worklist))) | |
4435 | return; | |
226223ab | 4436 | |
6ba94429 FW |
4437 | /* release id and unhash */ |
4438 | if (pool->id >= 0) | |
4439 | idr_remove(&worker_pool_idr, pool->id); | |
4440 | hash_del(&pool->hash_node); | |
d55262c4 | 4441 | |
6ba94429 | 4442 | /* |
692b4825 TH |
4443 | * Become the manager and destroy all workers. This prevents |
4444 | * @pool's workers from blocking on attach_mutex. We're the last | |
4445 | * manager and @pool gets freed with the flag set. | |
9ab03be4 VS |
4446 | * |
4447 | * Having a concurrent manager is quite unlikely to happen as we can | |
4448 | * only get here with | |
4449 | * pwq->refcnt == pool->refcnt == 0 | |
4450 | * which implies no work queued to the pool, which implies no worker can | |
4451 | * become the manager. However a worker could have taken the role of | |
4452 | * manager before the refcnts dropped to 0, since maybe_create_worker() | |
4453 | * drops pool->lock | |
6ba94429 | 4454 | */ |
9ab03be4 VS |
4455 | while (true) { |
4456 | rcuwait_wait_event(&manager_wait, | |
4457 | !(pool->flags & POOL_MANAGER_ACTIVE), | |
4458 | TASK_UNINTERRUPTIBLE); | |
e02b9312 VS |
4459 | |
4460 | mutex_lock(&wq_pool_attach_mutex); | |
9ab03be4 VS |
4461 | raw_spin_lock_irq(&pool->lock); |
4462 | if (!(pool->flags & POOL_MANAGER_ACTIVE)) { | |
4463 | pool->flags |= POOL_MANAGER_ACTIVE; | |
4464 | break; | |
4465 | } | |
4466 | raw_spin_unlock_irq(&pool->lock); | |
e02b9312 | 4467 | mutex_unlock(&wq_pool_attach_mutex); |
9ab03be4 | 4468 | } |
692b4825 | 4469 | |
6ba94429 | 4470 | while ((worker = first_idle_worker(pool))) |
e02b9312 | 4471 | set_worker_dying(worker, &cull_list); |
6ba94429 | 4472 | WARN_ON(pool->nr_workers || pool->nr_idle); |
a9b8a985 | 4473 | raw_spin_unlock_irq(&pool->lock); |
d55262c4 | 4474 | |
e02b9312 VS |
4475 | wake_dying_workers(&cull_list); |
4476 | ||
4477 | if (!list_empty(&pool->workers) || !list_empty(&pool->dying_workers)) | |
6ba94429 | 4478 | pool->detach_completion = &detach_completion; |
1258fae7 | 4479 | mutex_unlock(&wq_pool_attach_mutex); |
226223ab | 4480 | |
6ba94429 FW |
4481 | if (pool->detach_completion) |
4482 | wait_for_completion(pool->detach_completion); | |
226223ab | 4483 | |
6ba94429 FW |
4484 | /* shut down the timers */ |
4485 | del_timer_sync(&pool->idle_timer); | |
3f959aa3 | 4486 | cancel_work_sync(&pool->idle_cull_work); |
6ba94429 | 4487 | del_timer_sync(&pool->mayday_timer); |
226223ab | 4488 | |
24acfb71 | 4489 | /* RCU protected to allow dereferences from get_work_pool() */ |
25b00775 | 4490 | call_rcu(&pool->rcu, rcu_free_pool); |
226223ab TH |
4491 | } |
4492 | ||
4493 | /** | |
6ba94429 FW |
4494 | * get_unbound_pool - get a worker_pool with the specified attributes |
4495 | * @attrs: the attributes of the worker_pool to get | |
226223ab | 4496 | * |
6ba94429 FW |
4497 | * Obtain a worker_pool which has the same attributes as @attrs, bump the |
4498 | * reference count and return it. If there already is a matching | |
4499 | * worker_pool, it will be used; otherwise, this function attempts to | |
4500 | * create a new one. | |
226223ab | 4501 | * |
6ba94429 | 4502 | * Should be called with wq_pool_mutex held. |
226223ab | 4503 | * |
6ba94429 FW |
4504 | * Return: On success, a worker_pool with the same attributes as @attrs. |
4505 | * On failure, %NULL. | |
226223ab | 4506 | */ |
6ba94429 | 4507 | static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) |
226223ab | 4508 | { |
84193c07 | 4509 | struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_NUMA]; |
6ba94429 FW |
4510 | u32 hash = wqattrs_hash(attrs); |
4511 | struct worker_pool *pool; | |
84193c07 | 4512 | int pod, node = NUMA_NO_NODE; |
226223ab | 4513 | |
6ba94429 | 4514 | lockdep_assert_held(&wq_pool_mutex); |
226223ab | 4515 | |
6ba94429 FW |
4516 | /* do we already have a matching pool? */ |
4517 | hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { | |
4518 | if (wqattrs_equal(pool->attrs, attrs)) { | |
4519 | pool->refcnt++; | |
4520 | return pool; | |
4521 | } | |
4522 | } | |
226223ab | 4523 | |
9546b29e | 4524 | /* If __pod_cpumask is contained inside a NUMA pod, that's our node */ |
84193c07 | 4525 | for (pod = 0; pod < pt->nr_pods; pod++) { |
9546b29e | 4526 | if (cpumask_subset(attrs->__pod_cpumask, pt->pod_cpus[pod])) { |
84193c07 TH |
4527 | node = pt->pod_node[pod]; |
4528 | break; | |
e2273584 XP |
4529 | } |
4530 | } | |
4531 | ||
6ba94429 | 4532 | /* nope, create a new one */ |
84193c07 | 4533 | pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, node); |
6ba94429 FW |
4534 | if (!pool || init_worker_pool(pool) < 0) |
4535 | goto fail; | |
4536 | ||
84193c07 | 4537 | pool->node = node; |
5de7a03c TH |
4538 | copy_workqueue_attrs(pool->attrs, attrs); |
4539 | wqattrs_clear_for_pool(pool->attrs); | |
226223ab | 4540 | |
6ba94429 FW |
4541 | if (worker_pool_assign_id(pool) < 0) |
4542 | goto fail; | |
226223ab | 4543 | |
6ba94429 | 4544 | /* create and start the initial worker */ |
3347fa09 | 4545 | if (wq_online && !create_worker(pool)) |
6ba94429 | 4546 | goto fail; |
226223ab | 4547 | |
6ba94429 FW |
4548 | /* install */ |
4549 | hash_add(unbound_pool_hash, &pool->hash_node, hash); | |
226223ab | 4550 | |
6ba94429 FW |
4551 | return pool; |
4552 | fail: | |
4553 | if (pool) | |
4554 | put_unbound_pool(pool); | |
4555 | return NULL; | |
226223ab | 4556 | } |
226223ab | 4557 | |
6ba94429 | 4558 | static void rcu_free_pwq(struct rcu_head *rcu) |
7a4e344c | 4559 | { |
6ba94429 FW |
4560 | kmem_cache_free(pwq_cache, |
4561 | container_of(rcu, struct pool_workqueue, rcu)); | |
7a4e344c TH |
4562 | } |
4563 | ||
6ba94429 | 4564 | /* |
967b494e TH |
4565 | * Scheduled on pwq_release_worker by put_pwq() when an unbound pwq hits zero |
4566 | * refcnt and needs to be destroyed. | |
7a4e344c | 4567 | */ |
687a9aa5 | 4568 | static void pwq_release_workfn(struct kthread_work *work) |
7a4e344c | 4569 | { |
6ba94429 | 4570 | struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, |
687a9aa5 | 4571 | release_work); |
6ba94429 FW |
4572 | struct workqueue_struct *wq = pwq->wq; |
4573 | struct worker_pool *pool = pwq->pool; | |
b42b0bdd | 4574 | bool is_last = false; |
7a4e344c | 4575 | |
b42b0bdd | 4576 | /* |
687a9aa5 | 4577 | * When @pwq is not linked, it doesn't hold any reference to the |
b42b0bdd YY |
4578 | * @wq, and @wq is invalid to access. |
4579 | */ | |
4580 | if (!list_empty(&pwq->pwqs_node)) { | |
b42b0bdd YY |
4581 | mutex_lock(&wq->mutex); |
4582 | list_del_rcu(&pwq->pwqs_node); | |
4583 | is_last = list_empty(&wq->pwqs); | |
4584 | mutex_unlock(&wq->mutex); | |
4585 | } | |
6ba94429 | 4586 | |
687a9aa5 TH |
4587 | if (wq->flags & WQ_UNBOUND) { |
4588 | mutex_lock(&wq_pool_mutex); | |
4589 | put_unbound_pool(pool); | |
4590 | mutex_unlock(&wq_pool_mutex); | |
4591 | } | |
6ba94429 | 4592 | |
5797b1c1 TH |
4593 | if (!list_empty(&pwq->pending_node)) { |
4594 | struct wq_node_nr_active *nna = | |
4595 | wq_node_nr_active(pwq->wq, pwq->pool->node); | |
4596 | ||
4597 | raw_spin_lock_irq(&nna->lock); | |
4598 | list_del_init(&pwq->pending_node); | |
4599 | raw_spin_unlock_irq(&nna->lock); | |
4600 | } | |
4601 | ||
25b00775 | 4602 | call_rcu(&pwq->rcu, rcu_free_pwq); |
7a4e344c | 4603 | |
2865a8fb | 4604 | /* |
6ba94429 FW |
4605 | * If we're the last pwq going away, @wq is already dead and no one |
4606 | * is gonna access it anymore. Schedule RCU free. | |
2865a8fb | 4607 | */ |
669de8bd BVA |
4608 | if (is_last) { |
4609 | wq_unregister_lockdep(wq); | |
25b00775 | 4610 | call_rcu(&wq->rcu, rcu_free_wq); |
669de8bd | 4611 | } |
29c91e99 TH |
4612 | } |
4613 | ||
67dc8325 | 4614 | /* initialize newly allocated @pwq which is associated with @wq and @pool */ |
6ba94429 FW |
4615 | static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, |
4616 | struct worker_pool *pool) | |
29c91e99 | 4617 | { |
6ba94429 | 4618 | BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); |
29c91e99 | 4619 | |
6ba94429 FW |
4620 | memset(pwq, 0, sizeof(*pwq)); |
4621 | ||
4622 | pwq->pool = pool; | |
4623 | pwq->wq = wq; | |
4624 | pwq->flush_color = -1; | |
4625 | pwq->refcnt = 1; | |
f97a4a1a | 4626 | INIT_LIST_HEAD(&pwq->inactive_works); |
5797b1c1 | 4627 | INIT_LIST_HEAD(&pwq->pending_node); |
6ba94429 FW |
4628 | INIT_LIST_HEAD(&pwq->pwqs_node); |
4629 | INIT_LIST_HEAD(&pwq->mayday_node); | |
687a9aa5 | 4630 | kthread_init_work(&pwq->release_work, pwq_release_workfn); |
29c91e99 TH |
4631 | } |
4632 | ||
6ba94429 FW |
4633 | /* sync @pwq with the current state of its associated wq and link it */ |
4634 | static void link_pwq(struct pool_workqueue *pwq) | |
29c91e99 | 4635 | { |
6ba94429 | 4636 | struct workqueue_struct *wq = pwq->wq; |
29c91e99 | 4637 | |
6ba94429 | 4638 | lockdep_assert_held(&wq->mutex); |
a892cacc | 4639 | |
6ba94429 FW |
4640 | /* may be called multiple times, ignore if already linked */ |
4641 | if (!list_empty(&pwq->pwqs_node)) | |
29c91e99 | 4642 | return; |
29c91e99 | 4643 | |
6ba94429 FW |
4644 | /* set the matching work_color */ |
4645 | pwq->work_color = wq->work_color; | |
29c91e99 | 4646 | |
6ba94429 FW |
4647 | /* link in @pwq */ |
4648 | list_add_rcu(&pwq->pwqs_node, &wq->pwqs); | |
4649 | } | |
29c91e99 | 4650 | |
6ba94429 FW |
4651 | /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ |
4652 | static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, | |
4653 | const struct workqueue_attrs *attrs) | |
4654 | { | |
4655 | struct worker_pool *pool; | |
4656 | struct pool_workqueue *pwq; | |
60f5a4bc | 4657 | |
6ba94429 | 4658 | lockdep_assert_held(&wq_pool_mutex); |
60f5a4bc | 4659 | |
6ba94429 FW |
4660 | pool = get_unbound_pool(attrs); |
4661 | if (!pool) | |
4662 | return NULL; | |
60f5a4bc | 4663 | |
6ba94429 FW |
4664 | pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); |
4665 | if (!pwq) { | |
4666 | put_unbound_pool(pool); | |
4667 | return NULL; | |
4668 | } | |
29c91e99 | 4669 | |
6ba94429 FW |
4670 | init_pwq(pwq, wq, pool); |
4671 | return pwq; | |
4672 | } | |
29c91e99 | 4673 | |
29c91e99 | 4674 | /** |
fef59c9c | 4675 | * wq_calc_pod_cpumask - calculate a wq_attrs' cpumask for a pod |
042f7df1 | 4676 | * @attrs: the wq_attrs of the default pwq of the target workqueue |
84193c07 | 4677 | * @cpu: the target CPU |
6ba94429 | 4678 | * @cpu_going_down: if >= 0, the CPU to consider as offline |
29c91e99 | 4679 | * |
fef59c9c TH |
4680 | * Calculate the cpumask a workqueue with @attrs should use on @pod. If |
4681 | * @cpu_going_down is >= 0, that cpu is considered offline during calculation. | |
9546b29e | 4682 | * The result is stored in @attrs->__pod_cpumask. |
a892cacc | 4683 | * |
fef59c9c TH |
4684 | * If pod affinity is not enabled, @attrs->cpumask is always used. If enabled |
4685 | * and @pod has online CPUs requested by @attrs, the returned cpumask is the | |
4686 | * intersection of the possible CPUs of @pod and @attrs->cpumask. | |
d185af30 | 4687 | * |
fef59c9c | 4688 | * The caller is responsible for ensuring that the cpumask of @pod stays stable. |
29c91e99 | 4689 | */ |
9546b29e TH |
4690 | static void wq_calc_pod_cpumask(struct workqueue_attrs *attrs, int cpu, |
4691 | int cpu_going_down) | |
29c91e99 | 4692 | { |
84193c07 TH |
4693 | const struct wq_pod_type *pt = wqattrs_pod_type(attrs); |
4694 | int pod = pt->cpu_pod[cpu]; | |
29c91e99 | 4695 | |
fef59c9c | 4696 | /* does @pod have any online CPUs @attrs wants? */ |
9546b29e TH |
4697 | cpumask_and(attrs->__pod_cpumask, pt->pod_cpus[pod], attrs->cpumask); |
4698 | cpumask_and(attrs->__pod_cpumask, attrs->__pod_cpumask, cpu_online_mask); | |
6ba94429 | 4699 | if (cpu_going_down >= 0) |
9546b29e | 4700 | cpumask_clear_cpu(cpu_going_down, attrs->__pod_cpumask); |
29c91e99 | 4701 | |
9546b29e TH |
4702 | if (cpumask_empty(attrs->__pod_cpumask)) { |
4703 | cpumask_copy(attrs->__pod_cpumask, attrs->cpumask); | |
84193c07 TH |
4704 | return; |
4705 | } | |
4c16bd32 | 4706 | |
fef59c9c | 4707 | /* yeap, return possible CPUs in @pod that @attrs wants */ |
9546b29e | 4708 | cpumask_and(attrs->__pod_cpumask, attrs->cpumask, pt->pod_cpus[pod]); |
1ad0f0a7 | 4709 | |
9546b29e | 4710 | if (cpumask_empty(attrs->__pod_cpumask)) |
1ad0f0a7 MB |
4711 | pr_warn_once("WARNING: workqueue cpumask: online intersect > " |
4712 | "possible intersect\n"); | |
4c16bd32 TH |
4713 | } |
4714 | ||
9f66cff2 | 4715 | /* install @pwq into @wq and return the old pwq, @cpu < 0 for dfl_pwq */ |
636b927e TH |
4716 | static struct pool_workqueue *install_unbound_pwq(struct workqueue_struct *wq, |
4717 | int cpu, struct pool_workqueue *pwq) | |
1befcf30 | 4718 | { |
9f66cff2 | 4719 | struct pool_workqueue __rcu **slot = unbound_pwq_slot(wq, cpu); |
1befcf30 TH |
4720 | struct pool_workqueue *old_pwq; |
4721 | ||
5b95e1af | 4722 | lockdep_assert_held(&wq_pool_mutex); |
1befcf30 TH |
4723 | lockdep_assert_held(&wq->mutex); |
4724 | ||
4725 | /* link_pwq() can handle duplicate calls */ | |
4726 | link_pwq(pwq); | |
4727 | ||
9f66cff2 TH |
4728 | old_pwq = rcu_access_pointer(*slot); |
4729 | rcu_assign_pointer(*slot, pwq); | |
1befcf30 TH |
4730 | return old_pwq; |
4731 | } | |
4732 | ||
2d5f0764 LJ |
4733 | /* context to store the prepared attrs & pwqs before applying */ |
4734 | struct apply_wqattrs_ctx { | |
4735 | struct workqueue_struct *wq; /* target workqueue */ | |
4736 | struct workqueue_attrs *attrs; /* attrs to apply */ | |
042f7df1 | 4737 | struct list_head list; /* queued for batching commit */ |
2d5f0764 LJ |
4738 | struct pool_workqueue *dfl_pwq; |
4739 | struct pool_workqueue *pwq_tbl[]; | |
4740 | }; | |
4741 | ||
4742 | /* free the resources after success or abort */ | |
4743 | static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx) | |
4744 | { | |
4745 | if (ctx) { | |
636b927e | 4746 | int cpu; |
2d5f0764 | 4747 | |
636b927e TH |
4748 | for_each_possible_cpu(cpu) |
4749 | put_pwq_unlocked(ctx->pwq_tbl[cpu]); | |
2d5f0764 LJ |
4750 | put_pwq_unlocked(ctx->dfl_pwq); |
4751 | ||
4752 | free_workqueue_attrs(ctx->attrs); | |
4753 | ||
4754 | kfree(ctx); | |
4755 | } | |
4756 | } | |
4757 | ||
4758 | /* allocate the attrs and pwqs for later installation */ | |
4759 | static struct apply_wqattrs_ctx * | |
4760 | apply_wqattrs_prepare(struct workqueue_struct *wq, | |
99c621ef LJ |
4761 | const struct workqueue_attrs *attrs, |
4762 | const cpumask_var_t unbound_cpumask) | |
9e8cd2f5 | 4763 | { |
2d5f0764 | 4764 | struct apply_wqattrs_ctx *ctx; |
9546b29e | 4765 | struct workqueue_attrs *new_attrs; |
636b927e | 4766 | int cpu; |
9e8cd2f5 | 4767 | |
2d5f0764 | 4768 | lockdep_assert_held(&wq_pool_mutex); |
9e8cd2f5 | 4769 | |
84193c07 TH |
4770 | if (WARN_ON(attrs->affn_scope < 0 || |
4771 | attrs->affn_scope >= WQ_AFFN_NR_TYPES)) | |
4772 | return ERR_PTR(-EINVAL); | |
4773 | ||
636b927e | 4774 | ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_cpu_ids), GFP_KERNEL); |
8719dcea | 4775 | |
be69d00d | 4776 | new_attrs = alloc_workqueue_attrs(); |
9546b29e | 4777 | if (!ctx || !new_attrs) |
2d5f0764 | 4778 | goto out_free; |
13e2e556 | 4779 | |
4c16bd32 TH |
4780 | /* |
4781 | * If something goes wrong during CPU up/down, we'll fall back to | |
4782 | * the default pwq covering whole @attrs->cpumask. Always create | |
4783 | * it even if we don't use it immediately. | |
4784 | */ | |
0f36ee24 TH |
4785 | copy_workqueue_attrs(new_attrs, attrs); |
4786 | wqattrs_actualize_cpumask(new_attrs, unbound_cpumask); | |
9546b29e | 4787 | cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask); |
2d5f0764 LJ |
4788 | ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs); |
4789 | if (!ctx->dfl_pwq) | |
4790 | goto out_free; | |
4c16bd32 | 4791 | |
636b927e | 4792 | for_each_possible_cpu(cpu) { |
af73f5c9 | 4793 | if (new_attrs->ordered) { |
2d5f0764 | 4794 | ctx->dfl_pwq->refcnt++; |
636b927e TH |
4795 | ctx->pwq_tbl[cpu] = ctx->dfl_pwq; |
4796 | } else { | |
9546b29e TH |
4797 | wq_calc_pod_cpumask(new_attrs, cpu, -1); |
4798 | ctx->pwq_tbl[cpu] = alloc_unbound_pwq(wq, new_attrs); | |
636b927e TH |
4799 | if (!ctx->pwq_tbl[cpu]) |
4800 | goto out_free; | |
4c16bd32 TH |
4801 | } |
4802 | } | |
4803 | ||
042f7df1 LJ |
4804 | /* save the user configured attrs and sanitize it. */ |
4805 | copy_workqueue_attrs(new_attrs, attrs); | |
4806 | cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); | |
9546b29e | 4807 | cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask); |
2d5f0764 | 4808 | ctx->attrs = new_attrs; |
042f7df1 | 4809 | |
2d5f0764 | 4810 | ctx->wq = wq; |
2d5f0764 LJ |
4811 | return ctx; |
4812 | ||
4813 | out_free: | |
2d5f0764 LJ |
4814 | free_workqueue_attrs(new_attrs); |
4815 | apply_wqattrs_cleanup(ctx); | |
84193c07 | 4816 | return ERR_PTR(-ENOMEM); |
2d5f0764 LJ |
4817 | } |
4818 | ||
4819 | /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */ | |
4820 | static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx) | |
4821 | { | |
636b927e | 4822 | int cpu; |
9e8cd2f5 | 4823 | |
4c16bd32 | 4824 | /* all pwqs have been created successfully, let's install'em */ |
2d5f0764 | 4825 | mutex_lock(&ctx->wq->mutex); |
a892cacc | 4826 | |
2d5f0764 | 4827 | copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs); |
4c16bd32 | 4828 | |
9f66cff2 | 4829 | /* save the previous pwqs and install the new ones */ |
636b927e TH |
4830 | for_each_possible_cpu(cpu) |
4831 | ctx->pwq_tbl[cpu] = install_unbound_pwq(ctx->wq, cpu, | |
4832 | ctx->pwq_tbl[cpu]); | |
9f66cff2 | 4833 | ctx->dfl_pwq = install_unbound_pwq(ctx->wq, -1, ctx->dfl_pwq); |
f147f29e | 4834 | |
5797b1c1 TH |
4835 | /* update node_nr_active->max */ |
4836 | wq_update_node_max_active(ctx->wq, -1); | |
4837 | ||
2d5f0764 LJ |
4838 | mutex_unlock(&ctx->wq->mutex); |
4839 | } | |
9e8cd2f5 | 4840 | |
a0111cf6 LJ |
4841 | static int apply_workqueue_attrs_locked(struct workqueue_struct *wq, |
4842 | const struct workqueue_attrs *attrs) | |
2d5f0764 LJ |
4843 | { |
4844 | struct apply_wqattrs_ctx *ctx; | |
4c16bd32 | 4845 | |
2d5f0764 LJ |
4846 | /* only unbound workqueues can change attributes */ |
4847 | if (WARN_ON(!(wq->flags & WQ_UNBOUND))) | |
4848 | return -EINVAL; | |
13e2e556 | 4849 | |
2d5f0764 | 4850 | /* creating multiple pwqs breaks ordering guarantee */ |
0a94efb5 TH |
4851 | if (!list_empty(&wq->pwqs)) { |
4852 | if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) | |
4853 | return -EINVAL; | |
4854 | ||
4855 | wq->flags &= ~__WQ_ORDERED; | |
4856 | } | |
2d5f0764 | 4857 | |
99c621ef | 4858 | ctx = apply_wqattrs_prepare(wq, attrs, wq_unbound_cpumask); |
84193c07 TH |
4859 | if (IS_ERR(ctx)) |
4860 | return PTR_ERR(ctx); | |
2d5f0764 LJ |
4861 | |
4862 | /* the ctx has been prepared successfully, let's commit it */ | |
6201171e | 4863 | apply_wqattrs_commit(ctx); |
2d5f0764 LJ |
4864 | apply_wqattrs_cleanup(ctx); |
4865 | ||
6201171e | 4866 | return 0; |
9e8cd2f5 TH |
4867 | } |
4868 | ||
a0111cf6 LJ |
4869 | /** |
4870 | * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue | |
4871 | * @wq: the target workqueue | |
4872 | * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() | |
4873 | * | |
fef59c9c TH |
4874 | * Apply @attrs to an unbound workqueue @wq. Unless disabled, this function maps |
4875 | * a separate pwq to each CPU pod with possibles CPUs in @attrs->cpumask so that | |
4876 | * work items are affine to the pod it was issued on. Older pwqs are released as | |
4877 | * in-flight work items finish. Note that a work item which repeatedly requeues | |
4878 | * itself back-to-back will stay on its current pwq. | |
a0111cf6 LJ |
4879 | * |
4880 | * Performs GFP_KERNEL allocations. | |
4881 | * | |
ffd8bea8 | 4882 | * Assumes caller has CPU hotplug read exclusion, i.e. cpus_read_lock(). |
509b3204 | 4883 | * |
a0111cf6 LJ |
4884 | * Return: 0 on success and -errno on failure. |
4885 | */ | |
513c98d0 | 4886 | int apply_workqueue_attrs(struct workqueue_struct *wq, |
a0111cf6 LJ |
4887 | const struct workqueue_attrs *attrs) |
4888 | { | |
4889 | int ret; | |
4890 | ||
509b3204 DJ |
4891 | lockdep_assert_cpus_held(); |
4892 | ||
4893 | mutex_lock(&wq_pool_mutex); | |
a0111cf6 | 4894 | ret = apply_workqueue_attrs_locked(wq, attrs); |
509b3204 | 4895 | mutex_unlock(&wq_pool_mutex); |
a0111cf6 LJ |
4896 | |
4897 | return ret; | |
4898 | } | |
4899 | ||
4c16bd32 | 4900 | /** |
fef59c9c | 4901 | * wq_update_pod - update pod affinity of a wq for CPU hot[un]plug |
4c16bd32 | 4902 | * @wq: the target workqueue |
4cbfd3de TH |
4903 | * @cpu: the CPU to update pool association for |
4904 | * @hotplug_cpu: the CPU coming up or going down | |
4c16bd32 TH |
4905 | * @online: whether @cpu is coming up or going down |
4906 | * | |
4907 | * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and | |
fef59c9c | 4908 | * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update pod affinity of |
4c16bd32 TH |
4909 | * @wq accordingly. |
4910 | * | |
fef59c9c TH |
4911 | * |
4912 | * If pod affinity can't be adjusted due to memory allocation failure, it falls | |
4913 | * back to @wq->dfl_pwq which may not be optimal but is always correct. | |
4914 | * | |
4915 | * Note that when the last allowed CPU of a pod goes offline for a workqueue | |
4916 | * with a cpumask spanning multiple pods, the workers which were already | |
4917 | * executing the work items for the workqueue will lose their CPU affinity and | |
4918 | * may execute on any CPU. This is similar to how per-cpu workqueues behave on | |
4919 | * CPU_DOWN. If a workqueue user wants strict affinity, it's the user's | |
4920 | * responsibility to flush the work item from CPU_DOWN_PREPARE. | |
4c16bd32 | 4921 | */ |
fef59c9c TH |
4922 | static void wq_update_pod(struct workqueue_struct *wq, int cpu, |
4923 | int hotplug_cpu, bool online) | |
4c16bd32 | 4924 | { |
4cbfd3de | 4925 | int off_cpu = online ? -1 : hotplug_cpu; |
4c16bd32 TH |
4926 | struct pool_workqueue *old_pwq = NULL, *pwq; |
4927 | struct workqueue_attrs *target_attrs; | |
4c16bd32 TH |
4928 | |
4929 | lockdep_assert_held(&wq_pool_mutex); | |
4930 | ||
84193c07 | 4931 | if (!(wq->flags & WQ_UNBOUND) || wq->unbound_attrs->ordered) |
4c16bd32 TH |
4932 | return; |
4933 | ||
4934 | /* | |
4935 | * We don't wanna alloc/free wq_attrs for each wq for each CPU. | |
4936 | * Let's use a preallocated one. The following buf is protected by | |
4937 | * CPU hotplug exclusion. | |
4938 | */ | |
fef59c9c | 4939 | target_attrs = wq_update_pod_attrs_buf; |
4c16bd32 | 4940 | |
4c16bd32 | 4941 | copy_workqueue_attrs(target_attrs, wq->unbound_attrs); |
0f36ee24 | 4942 | wqattrs_actualize_cpumask(target_attrs, wq_unbound_cpumask); |
4c16bd32 | 4943 | |
636b927e | 4944 | /* nothing to do if the target cpumask matches the current pwq */ |
9546b29e | 4945 | wq_calc_pod_cpumask(target_attrs, cpu, off_cpu); |
9f66cff2 | 4946 | if (wqattrs_equal(target_attrs, unbound_pwq(wq, cpu)->pool->attrs)) |
636b927e | 4947 | return; |
4c16bd32 | 4948 | |
4c16bd32 TH |
4949 | /* create a new pwq */ |
4950 | pwq = alloc_unbound_pwq(wq, target_attrs); | |
4951 | if (!pwq) { | |
fef59c9c | 4952 | pr_warn("workqueue: allocation failed while updating CPU pod affinity of \"%s\"\n", |
2d916033 | 4953 | wq->name); |
77f300b1 | 4954 | goto use_dfl_pwq; |
4c16bd32 TH |
4955 | } |
4956 | ||
f7142ed4 | 4957 | /* Install the new pwq. */ |
4c16bd32 | 4958 | mutex_lock(&wq->mutex); |
636b927e | 4959 | old_pwq = install_unbound_pwq(wq, cpu, pwq); |
4c16bd32 TH |
4960 | goto out_unlock; |
4961 | ||
4962 | use_dfl_pwq: | |
f7142ed4 | 4963 | mutex_lock(&wq->mutex); |
9f66cff2 TH |
4964 | pwq = unbound_pwq(wq, -1); |
4965 | raw_spin_lock_irq(&pwq->pool->lock); | |
4966 | get_pwq(pwq); | |
4967 | raw_spin_unlock_irq(&pwq->pool->lock); | |
4968 | old_pwq = install_unbound_pwq(wq, cpu, pwq); | |
4c16bd32 TH |
4969 | out_unlock: |
4970 | mutex_unlock(&wq->mutex); | |
4971 | put_pwq_unlocked(old_pwq); | |
4972 | } | |
4973 | ||
30cdf249 | 4974 | static int alloc_and_link_pwqs(struct workqueue_struct *wq) |
0f900049 | 4975 | { |
49e3cf44 | 4976 | bool highpri = wq->flags & WQ_HIGHPRI; |
8a2b7538 | 4977 | int cpu, ret; |
30cdf249 | 4978 | |
636b927e TH |
4979 | wq->cpu_pwq = alloc_percpu(struct pool_workqueue *); |
4980 | if (!wq->cpu_pwq) | |
4981 | goto enomem; | |
30cdf249 | 4982 | |
636b927e | 4983 | if (!(wq->flags & WQ_UNBOUND)) { |
30cdf249 | 4984 | for_each_possible_cpu(cpu) { |
687a9aa5 | 4985 | struct pool_workqueue **pwq_p = |
ee1ceef7 | 4986 | per_cpu_ptr(wq->cpu_pwq, cpu); |
687a9aa5 TH |
4987 | struct worker_pool *pool = |
4988 | &(per_cpu_ptr(cpu_worker_pools, cpu)[highpri]); | |
4989 | ||
4990 | *pwq_p = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, | |
4991 | pool->node); | |
4992 | if (!*pwq_p) | |
4993 | goto enomem; | |
f3421797 | 4994 | |
687a9aa5 | 4995 | init_pwq(*pwq_p, wq, pool); |
f147f29e TH |
4996 | |
4997 | mutex_lock(&wq->mutex); | |
687a9aa5 | 4998 | link_pwq(*pwq_p); |
f147f29e | 4999 | mutex_unlock(&wq->mutex); |
30cdf249 | 5000 | } |
9e8cd2f5 | 5001 | return 0; |
509b3204 DJ |
5002 | } |
5003 | ||
ffd8bea8 | 5004 | cpus_read_lock(); |
509b3204 | 5005 | if (wq->flags & __WQ_ORDERED) { |
9f66cff2 TH |
5006 | struct pool_workqueue *dfl_pwq; |
5007 | ||
8a2b7538 TH |
5008 | ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]); |
5009 | /* there should only be single pwq for ordering guarantee */ | |
9f66cff2 TH |
5010 | dfl_pwq = rcu_access_pointer(wq->dfl_pwq); |
5011 | WARN(!ret && (wq->pwqs.next != &dfl_pwq->pwqs_node || | |
5012 | wq->pwqs.prev != &dfl_pwq->pwqs_node), | |
8a2b7538 | 5013 | "ordering guarantee broken for workqueue %s\n", wq->name); |
30cdf249 | 5014 | } else { |
509b3204 | 5015 | ret = apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); |
30cdf249 | 5016 | } |
ffd8bea8 | 5017 | cpus_read_unlock(); |
509b3204 | 5018 | |
64344553 Z |
5019 | /* for unbound pwq, flush the pwq_release_worker ensures that the |
5020 | * pwq_release_workfn() completes before calling kfree(wq). | |
5021 | */ | |
5022 | if (ret) | |
5023 | kthread_flush_worker(pwq_release_worker); | |
5024 | ||
509b3204 | 5025 | return ret; |
687a9aa5 TH |
5026 | |
5027 | enomem: | |
5028 | if (wq->cpu_pwq) { | |
7b42f401 Z |
5029 | for_each_possible_cpu(cpu) { |
5030 | struct pool_workqueue *pwq = *per_cpu_ptr(wq->cpu_pwq, cpu); | |
5031 | ||
5032 | if (pwq) | |
5033 | kmem_cache_free(pwq_cache, pwq); | |
5034 | } | |
687a9aa5 TH |
5035 | free_percpu(wq->cpu_pwq); |
5036 | wq->cpu_pwq = NULL; | |
5037 | } | |
5038 | return -ENOMEM; | |
0f900049 TH |
5039 | } |
5040 | ||
f3421797 TH |
5041 | static int wq_clamp_max_active(int max_active, unsigned int flags, |
5042 | const char *name) | |
b71ab8c2 | 5043 | { |
636b927e | 5044 | if (max_active < 1 || max_active > WQ_MAX_ACTIVE) |
044c782c | 5045 | pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", |
636b927e | 5046 | max_active, name, 1, WQ_MAX_ACTIVE); |
b71ab8c2 | 5047 | |
636b927e | 5048 | return clamp_val(max_active, 1, WQ_MAX_ACTIVE); |
b71ab8c2 TH |
5049 | } |
5050 | ||
983c7515 TH |
5051 | /* |
5052 | * Workqueues which may be used during memory reclaim should have a rescuer | |
5053 | * to guarantee forward progress. | |
5054 | */ | |
5055 | static int init_rescuer(struct workqueue_struct *wq) | |
5056 | { | |
5057 | struct worker *rescuer; | |
b92b36ea | 5058 | int ret; |
983c7515 TH |
5059 | |
5060 | if (!(wq->flags & WQ_MEM_RECLAIM)) | |
5061 | return 0; | |
5062 | ||
5063 | rescuer = alloc_worker(NUMA_NO_NODE); | |
4c0736a7 PM |
5064 | if (!rescuer) { |
5065 | pr_err("workqueue: Failed to allocate a rescuer for wq \"%s\"\n", | |
5066 | wq->name); | |
983c7515 | 5067 | return -ENOMEM; |
4c0736a7 | 5068 | } |
983c7515 TH |
5069 | |
5070 | rescuer->rescue_wq = wq; | |
b6a46f72 | 5071 | rescuer->task = kthread_create(rescuer_thread, rescuer, "kworker/R-%s", wq->name); |
f187b697 | 5072 | if (IS_ERR(rescuer->task)) { |
b92b36ea | 5073 | ret = PTR_ERR(rescuer->task); |
4c0736a7 PM |
5074 | pr_err("workqueue: Failed to create a rescuer kthread for wq \"%s\": %pe", |
5075 | wq->name, ERR_PTR(ret)); | |
983c7515 | 5076 | kfree(rescuer); |
b92b36ea | 5077 | return ret; |
983c7515 TH |
5078 | } |
5079 | ||
5080 | wq->rescuer = rescuer; | |
85f0ab43 JL |
5081 | if (wq->flags & WQ_UNBOUND) |
5082 | kthread_bind_mask(rescuer->task, wq->unbound_attrs->cpumask); | |
5083 | else | |
5084 | kthread_bind_mask(rescuer->task, cpu_possible_mask); | |
983c7515 TH |
5085 | wake_up_process(rescuer->task); |
5086 | ||
5087 | return 0; | |
5088 | } | |
5089 | ||
a045a272 TH |
5090 | /** |
5091 | * wq_adjust_max_active - update a wq's max_active to the current setting | |
5092 | * @wq: target workqueue | |
5093 | * | |
5094 | * If @wq isn't freezing, set @wq->max_active to the saved_max_active and | |
5095 | * activate inactive work items accordingly. If @wq is freezing, clear | |
5096 | * @wq->max_active to zero. | |
5097 | */ | |
5098 | static void wq_adjust_max_active(struct workqueue_struct *wq) | |
5099 | { | |
c5404d4e | 5100 | bool activated; |
5797b1c1 | 5101 | int new_max, new_min; |
a045a272 TH |
5102 | |
5103 | lockdep_assert_held(&wq->mutex); | |
5104 | ||
5105 | if ((wq->flags & WQ_FREEZABLE) && workqueue_freezing) { | |
5797b1c1 TH |
5106 | new_max = 0; |
5107 | new_min = 0; | |
5108 | } else { | |
5109 | new_max = wq->saved_max_active; | |
5110 | new_min = wq->saved_min_active; | |
a045a272 TH |
5111 | } |
5112 | ||
5797b1c1 | 5113 | if (wq->max_active == new_max && wq->min_active == new_min) |
a045a272 TH |
5114 | return; |
5115 | ||
5116 | /* | |
5797b1c1 | 5117 | * Update @wq->max/min_active and then kick inactive work items if more |
a045a272 TH |
5118 | * active work items are allowed. This doesn't break work item ordering |
5119 | * because new work items are always queued behind existing inactive | |
5120 | * work items if there are any. | |
5121 | */ | |
5797b1c1 TH |
5122 | WRITE_ONCE(wq->max_active, new_max); |
5123 | WRITE_ONCE(wq->min_active, new_min); | |
5124 | ||
5125 | if (wq->flags & WQ_UNBOUND) | |
5126 | wq_update_node_max_active(wq, -1); | |
5127 | ||
5128 | if (new_max == 0) | |
5129 | return; | |
a045a272 | 5130 | |
c5404d4e TH |
5131 | /* |
5132 | * Round-robin through pwq's activating the first inactive work item | |
5133 | * until max_active is filled. | |
5134 | */ | |
5135 | do { | |
5136 | struct pool_workqueue *pwq; | |
a045a272 | 5137 | |
c5404d4e TH |
5138 | activated = false; |
5139 | for_each_pwq(pwq, wq) { | |
5140 | unsigned long flags; | |
a045a272 | 5141 | |
c5404d4e TH |
5142 | /* can be called during early boot w/ irq disabled */ |
5143 | raw_spin_lock_irqsave(&pwq->pool->lock, flags); | |
5797b1c1 | 5144 | if (pwq_activate_first_inactive(pwq, true)) { |
c5404d4e TH |
5145 | activated = true; |
5146 | kick_pool(pwq->pool); | |
5147 | } | |
5148 | raw_spin_unlock_irqrestore(&pwq->pool->lock, flags); | |
5149 | } | |
5150 | } while (activated); | |
a045a272 TH |
5151 | } |
5152 | ||
a2775bbc | 5153 | __printf(1, 4) |
669de8bd BVA |
5154 | struct workqueue_struct *alloc_workqueue(const char *fmt, |
5155 | unsigned int flags, | |
5156 | int max_active, ...) | |
1da177e4 | 5157 | { |
ecf6881f | 5158 | va_list args; |
1da177e4 | 5159 | struct workqueue_struct *wq; |
91ccc6e7 TH |
5160 | size_t wq_size; |
5161 | int name_len; | |
b196be89 | 5162 | |
5c0338c6 | 5163 | /* |
fef59c9c TH |
5164 | * Unbound && max_active == 1 used to imply ordered, which is no longer |
5165 | * the case on many machines due to per-pod pools. While | |
5c0338c6 | 5166 | * alloc_ordered_workqueue() is the right way to create an ordered |
fef59c9c | 5167 | * workqueue, keep the previous behavior to avoid subtle breakages. |
5c0338c6 TH |
5168 | */ |
5169 | if ((flags & WQ_UNBOUND) && max_active == 1) | |
5170 | flags |= __WQ_ORDERED; | |
5171 | ||
cee22a15 VK |
5172 | /* see the comment above the definition of WQ_POWER_EFFICIENT */ |
5173 | if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) | |
5174 | flags |= WQ_UNBOUND; | |
5175 | ||
ecf6881f | 5176 | /* allocate wq and format name */ |
91ccc6e7 TH |
5177 | if (flags & WQ_UNBOUND) |
5178 | wq_size = struct_size(wq, node_nr_active, nr_node_ids + 1); | |
5179 | else | |
5180 | wq_size = sizeof(*wq); | |
5181 | ||
5182 | wq = kzalloc(wq_size, GFP_KERNEL); | |
b196be89 | 5183 | if (!wq) |
d2c1d404 | 5184 | return NULL; |
b196be89 | 5185 | |
6029a918 | 5186 | if (flags & WQ_UNBOUND) { |
be69d00d | 5187 | wq->unbound_attrs = alloc_workqueue_attrs(); |
6029a918 TH |
5188 | if (!wq->unbound_attrs) |
5189 | goto err_free_wq; | |
5190 | } | |
5191 | ||
669de8bd | 5192 | va_start(args, max_active); |
91ccc6e7 | 5193 | name_len = vsnprintf(wq->name, sizeof(wq->name), fmt, args); |
b196be89 | 5194 | va_end(args); |
1da177e4 | 5195 | |
91ccc6e7 TH |
5196 | if (name_len >= WQ_NAME_LEN) |
5197 | pr_warn_once("workqueue: name exceeds WQ_NAME_LEN. Truncating to: %s\n", | |
5198 | wq->name); | |
31c89007 | 5199 | |
d320c038 | 5200 | max_active = max_active ?: WQ_DFL_ACTIVE; |
b196be89 | 5201 | max_active = wq_clamp_max_active(max_active, flags, wq->name); |
3af24433 | 5202 | |
b196be89 | 5203 | /* init wq */ |
97e37d7b | 5204 | wq->flags = flags; |
a045a272 | 5205 | wq->max_active = max_active; |
5797b1c1 TH |
5206 | wq->min_active = min(max_active, WQ_DFL_MIN_ACTIVE); |
5207 | wq->saved_max_active = wq->max_active; | |
5208 | wq->saved_min_active = wq->min_active; | |
3c25a55d | 5209 | mutex_init(&wq->mutex); |
112202d9 | 5210 | atomic_set(&wq->nr_pwqs_to_flush, 0); |
30cdf249 | 5211 | INIT_LIST_HEAD(&wq->pwqs); |
73f53c4a TH |
5212 | INIT_LIST_HEAD(&wq->flusher_queue); |
5213 | INIT_LIST_HEAD(&wq->flusher_overflow); | |
493a1724 | 5214 | INIT_LIST_HEAD(&wq->maydays); |
502ca9d8 | 5215 | |
669de8bd | 5216 | wq_init_lockdep(wq); |
cce1a165 | 5217 | INIT_LIST_HEAD(&wq->list); |
3af24433 | 5218 | |
91ccc6e7 TH |
5219 | if (flags & WQ_UNBOUND) { |
5220 | if (alloc_node_nr_active(wq->node_nr_active) < 0) | |
5221 | goto err_unreg_lockdep; | |
5222 | } | |
5223 | ||
30cdf249 | 5224 | if (alloc_and_link_pwqs(wq) < 0) |
91ccc6e7 | 5225 | goto err_free_node_nr_active; |
1537663f | 5226 | |
40c17f75 | 5227 | if (wq_online && init_rescuer(wq) < 0) |
983c7515 | 5228 | goto err_destroy; |
3af24433 | 5229 | |
226223ab TH |
5230 | if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) |
5231 | goto err_destroy; | |
5232 | ||
a0a1a5fd | 5233 | /* |
68e13a67 LJ |
5234 | * wq_pool_mutex protects global freeze state and workqueues list. |
5235 | * Grab it, adjust max_active and add the new @wq to workqueues | |
5236 | * list. | |
a0a1a5fd | 5237 | */ |
68e13a67 | 5238 | mutex_lock(&wq_pool_mutex); |
a0a1a5fd | 5239 | |
a357fc03 | 5240 | mutex_lock(&wq->mutex); |
a045a272 | 5241 | wq_adjust_max_active(wq); |
a357fc03 | 5242 | mutex_unlock(&wq->mutex); |
a0a1a5fd | 5243 | |
e2dca7ad | 5244 | list_add_tail_rcu(&wq->list, &workqueues); |
a0a1a5fd | 5245 | |
68e13a67 | 5246 | mutex_unlock(&wq_pool_mutex); |
1537663f | 5247 | |
3af24433 | 5248 | return wq; |
d2c1d404 | 5249 | |
91ccc6e7 TH |
5250 | err_free_node_nr_active: |
5251 | if (wq->flags & WQ_UNBOUND) | |
5252 | free_node_nr_active(wq->node_nr_active); | |
82efcab3 | 5253 | err_unreg_lockdep: |
009bb421 BVA |
5254 | wq_unregister_lockdep(wq); |
5255 | wq_free_lockdep(wq); | |
82efcab3 | 5256 | err_free_wq: |
6029a918 | 5257 | free_workqueue_attrs(wq->unbound_attrs); |
d2c1d404 TH |
5258 | kfree(wq); |
5259 | return NULL; | |
5260 | err_destroy: | |
5261 | destroy_workqueue(wq); | |
4690c4ab | 5262 | return NULL; |
3af24433 | 5263 | } |
669de8bd | 5264 | EXPORT_SYMBOL_GPL(alloc_workqueue); |
1da177e4 | 5265 | |
c29eb853 TH |
5266 | static bool pwq_busy(struct pool_workqueue *pwq) |
5267 | { | |
5268 | int i; | |
5269 | ||
5270 | for (i = 0; i < WORK_NR_COLORS; i++) | |
5271 | if (pwq->nr_in_flight[i]) | |
5272 | return true; | |
5273 | ||
9f66cff2 | 5274 | if ((pwq != rcu_access_pointer(pwq->wq->dfl_pwq)) && (pwq->refcnt > 1)) |
c29eb853 | 5275 | return true; |
afa87ce8 | 5276 | if (!pwq_is_empty(pwq)) |
c29eb853 TH |
5277 | return true; |
5278 | ||
5279 | return false; | |
5280 | } | |
5281 | ||
3af24433 ON |
5282 | /** |
5283 | * destroy_workqueue - safely terminate a workqueue | |
5284 | * @wq: target workqueue | |
5285 | * | |
5286 | * Safely destroy a workqueue. All work currently pending will be done first. | |
5287 | */ | |
5288 | void destroy_workqueue(struct workqueue_struct *wq) | |
5289 | { | |
49e3cf44 | 5290 | struct pool_workqueue *pwq; |
636b927e | 5291 | int cpu; |
3af24433 | 5292 | |
def98c84 TH |
5293 | /* |
5294 | * Remove it from sysfs first so that sanity check failure doesn't | |
5295 | * lead to sysfs name conflicts. | |
5296 | */ | |
5297 | workqueue_sysfs_unregister(wq); | |
5298 | ||
33e3f0a3 RC |
5299 | /* mark the workqueue destruction is in progress */ |
5300 | mutex_lock(&wq->mutex); | |
5301 | wq->flags |= __WQ_DESTROYING; | |
5302 | mutex_unlock(&wq->mutex); | |
5303 | ||
9c5a2ba7 TH |
5304 | /* drain it before proceeding with destruction */ |
5305 | drain_workqueue(wq); | |
c8efcc25 | 5306 | |
def98c84 TH |
5307 | /* kill rescuer, if sanity checks fail, leave it w/o rescuer */ |
5308 | if (wq->rescuer) { | |
5309 | struct worker *rescuer = wq->rescuer; | |
5310 | ||
5311 | /* this prevents new queueing */ | |
a9b8a985 | 5312 | raw_spin_lock_irq(&wq_mayday_lock); |
def98c84 | 5313 | wq->rescuer = NULL; |
a9b8a985 | 5314 | raw_spin_unlock_irq(&wq_mayday_lock); |
def98c84 TH |
5315 | |
5316 | /* rescuer will empty maydays list before exiting */ | |
5317 | kthread_stop(rescuer->task); | |
8efe1223 | 5318 | kfree(rescuer); |
def98c84 TH |
5319 | } |
5320 | ||
c29eb853 TH |
5321 | /* |
5322 | * Sanity checks - grab all the locks so that we wait for all | |
5323 | * in-flight operations which may do put_pwq(). | |
5324 | */ | |
5325 | mutex_lock(&wq_pool_mutex); | |
b09f4fd3 | 5326 | mutex_lock(&wq->mutex); |
49e3cf44 | 5327 | for_each_pwq(pwq, wq) { |
a9b8a985 | 5328 | raw_spin_lock_irq(&pwq->pool->lock); |
c29eb853 | 5329 | if (WARN_ON(pwq_busy(pwq))) { |
1d9a6159 KW |
5330 | pr_warn("%s: %s has the following busy pwq\n", |
5331 | __func__, wq->name); | |
c29eb853 | 5332 | show_pwq(pwq); |
a9b8a985 | 5333 | raw_spin_unlock_irq(&pwq->pool->lock); |
b09f4fd3 | 5334 | mutex_unlock(&wq->mutex); |
c29eb853 | 5335 | mutex_unlock(&wq_pool_mutex); |
55df0933 | 5336 | show_one_workqueue(wq); |
6183c009 | 5337 | return; |
76af4d93 | 5338 | } |
a9b8a985 | 5339 | raw_spin_unlock_irq(&pwq->pool->lock); |
6183c009 | 5340 | } |
b09f4fd3 | 5341 | mutex_unlock(&wq->mutex); |
6183c009 | 5342 | |
a0a1a5fd TH |
5343 | /* |
5344 | * wq list is used to freeze wq, remove from list after | |
5345 | * flushing is complete in case freeze races us. | |
5346 | */ | |
e2dca7ad | 5347 | list_del_rcu(&wq->list); |
68e13a67 | 5348 | mutex_unlock(&wq_pool_mutex); |
3af24433 | 5349 | |
636b927e TH |
5350 | /* |
5351 | * We're the sole accessor of @wq. Directly access cpu_pwq and dfl_pwq | |
5352 | * to put the base refs. @wq will be auto-destroyed from the last | |
5353 | * pwq_put. RCU read lock prevents @wq from going away from under us. | |
5354 | */ | |
5355 | rcu_read_lock(); | |
4c16bd32 | 5356 | |
636b927e | 5357 | for_each_possible_cpu(cpu) { |
9f66cff2 TH |
5358 | put_pwq_unlocked(unbound_pwq(wq, cpu)); |
5359 | RCU_INIT_POINTER(*unbound_pwq_slot(wq, cpu), NULL); | |
29c91e99 | 5360 | } |
636b927e | 5361 | |
9f66cff2 TH |
5362 | put_pwq_unlocked(unbound_pwq(wq, -1)); |
5363 | RCU_INIT_POINTER(*unbound_pwq_slot(wq, -1), NULL); | |
636b927e TH |
5364 | |
5365 | rcu_read_unlock(); | |
3af24433 ON |
5366 | } |
5367 | EXPORT_SYMBOL_GPL(destroy_workqueue); | |
5368 | ||
dcd989cb TH |
5369 | /** |
5370 | * workqueue_set_max_active - adjust max_active of a workqueue | |
5371 | * @wq: target workqueue | |
5372 | * @max_active: new max_active value. | |
5373 | * | |
5797b1c1 TH |
5374 | * Set max_active of @wq to @max_active. See the alloc_workqueue() function |
5375 | * comment. | |
dcd989cb TH |
5376 | * |
5377 | * CONTEXT: | |
5378 | * Don't call from IRQ context. | |
5379 | */ | |
5380 | void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) | |
5381 | { | |
8719dcea | 5382 | /* disallow meddling with max_active for ordered workqueues */ |
0a94efb5 | 5383 | if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) |
8719dcea TH |
5384 | return; |
5385 | ||
f3421797 | 5386 | max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); |
dcd989cb | 5387 | |
a357fc03 | 5388 | mutex_lock(&wq->mutex); |
dcd989cb | 5389 | |
0a94efb5 | 5390 | wq->flags &= ~__WQ_ORDERED; |
dcd989cb | 5391 | wq->saved_max_active = max_active; |
5797b1c1 TH |
5392 | if (wq->flags & WQ_UNBOUND) |
5393 | wq->saved_min_active = min(wq->saved_min_active, max_active); | |
5394 | ||
a045a272 | 5395 | wq_adjust_max_active(wq); |
93981800 | 5396 | |
a357fc03 | 5397 | mutex_unlock(&wq->mutex); |
15316ba8 | 5398 | } |
dcd989cb | 5399 | EXPORT_SYMBOL_GPL(workqueue_set_max_active); |
15316ba8 | 5400 | |
27d4ee03 LW |
5401 | /** |
5402 | * current_work - retrieve %current task's work struct | |
5403 | * | |
5404 | * Determine if %current task is a workqueue worker and what it's working on. | |
5405 | * Useful to find out the context that the %current task is running in. | |
5406 | * | |
5407 | * Return: work struct if %current task is a workqueue worker, %NULL otherwise. | |
5408 | */ | |
5409 | struct work_struct *current_work(void) | |
5410 | { | |
5411 | struct worker *worker = current_wq_worker(); | |
5412 | ||
5413 | return worker ? worker->current_work : NULL; | |
5414 | } | |
5415 | EXPORT_SYMBOL(current_work); | |
5416 | ||
e6267616 TH |
5417 | /** |
5418 | * current_is_workqueue_rescuer - is %current workqueue rescuer? | |
5419 | * | |
5420 | * Determine whether %current is a workqueue rescuer. Can be used from | |
5421 | * work functions to determine whether it's being run off the rescuer task. | |
d185af30 YB |
5422 | * |
5423 | * Return: %true if %current is a workqueue rescuer. %false otherwise. | |
e6267616 TH |
5424 | */ |
5425 | bool current_is_workqueue_rescuer(void) | |
5426 | { | |
5427 | struct worker *worker = current_wq_worker(); | |
5428 | ||
6a092dfd | 5429 | return worker && worker->rescue_wq; |
e6267616 TH |
5430 | } |
5431 | ||
eef6a7d5 | 5432 | /** |
dcd989cb TH |
5433 | * workqueue_congested - test whether a workqueue is congested |
5434 | * @cpu: CPU in question | |
5435 | * @wq: target workqueue | |
eef6a7d5 | 5436 | * |
dcd989cb TH |
5437 | * Test whether @wq's cpu workqueue for @cpu is congested. There is |
5438 | * no synchronization around this function and the test result is | |
5439 | * unreliable and only useful as advisory hints or for debugging. | |
eef6a7d5 | 5440 | * |
d3251859 | 5441 | * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. |
636b927e TH |
5442 | * |
5443 | * With the exception of ordered workqueues, all workqueues have per-cpu | |
5444 | * pool_workqueues, each with its own congested state. A workqueue being | |
5445 | * congested on one CPU doesn't mean that the workqueue is contested on any | |
5446 | * other CPUs. | |
d3251859 | 5447 | * |
d185af30 | 5448 | * Return: |
dcd989cb | 5449 | * %true if congested, %false otherwise. |
eef6a7d5 | 5450 | */ |
d84ff051 | 5451 | bool workqueue_congested(int cpu, struct workqueue_struct *wq) |
1da177e4 | 5452 | { |
7fb98ea7 | 5453 | struct pool_workqueue *pwq; |
76af4d93 TH |
5454 | bool ret; |
5455 | ||
24acfb71 TG |
5456 | rcu_read_lock(); |
5457 | preempt_disable(); | |
7fb98ea7 | 5458 | |
d3251859 TH |
5459 | if (cpu == WORK_CPU_UNBOUND) |
5460 | cpu = smp_processor_id(); | |
5461 | ||
636b927e | 5462 | pwq = *per_cpu_ptr(wq->cpu_pwq, cpu); |
f97a4a1a | 5463 | ret = !list_empty(&pwq->inactive_works); |
636b927e | 5464 | |
24acfb71 TG |
5465 | preempt_enable(); |
5466 | rcu_read_unlock(); | |
76af4d93 TH |
5467 | |
5468 | return ret; | |
1da177e4 | 5469 | } |
dcd989cb | 5470 | EXPORT_SYMBOL_GPL(workqueue_congested); |
1da177e4 | 5471 | |
dcd989cb TH |
5472 | /** |
5473 | * work_busy - test whether a work is currently pending or running | |
5474 | * @work: the work to be tested | |
5475 | * | |
5476 | * Test whether @work is currently pending or running. There is no | |
5477 | * synchronization around this function and the test result is | |
5478 | * unreliable and only useful as advisory hints or for debugging. | |
dcd989cb | 5479 | * |
d185af30 | 5480 | * Return: |
dcd989cb TH |
5481 | * OR'd bitmask of WORK_BUSY_* bits. |
5482 | */ | |
5483 | unsigned int work_busy(struct work_struct *work) | |
1da177e4 | 5484 | { |
fa1b54e6 | 5485 | struct worker_pool *pool; |
dcd989cb TH |
5486 | unsigned long flags; |
5487 | unsigned int ret = 0; | |
1da177e4 | 5488 | |
dcd989cb TH |
5489 | if (work_pending(work)) |
5490 | ret |= WORK_BUSY_PENDING; | |
1da177e4 | 5491 | |
24acfb71 | 5492 | rcu_read_lock(); |
fa1b54e6 | 5493 | pool = get_work_pool(work); |
038366c5 | 5494 | if (pool) { |
a9b8a985 | 5495 | raw_spin_lock_irqsave(&pool->lock, flags); |
038366c5 LJ |
5496 | if (find_worker_executing_work(pool, work)) |
5497 | ret |= WORK_BUSY_RUNNING; | |
a9b8a985 | 5498 | raw_spin_unlock_irqrestore(&pool->lock, flags); |
038366c5 | 5499 | } |
24acfb71 | 5500 | rcu_read_unlock(); |
1da177e4 | 5501 | |
dcd989cb | 5502 | return ret; |
1da177e4 | 5503 | } |
dcd989cb | 5504 | EXPORT_SYMBOL_GPL(work_busy); |
1da177e4 | 5505 | |
3d1cb205 TH |
5506 | /** |
5507 | * set_worker_desc - set description for the current work item | |
5508 | * @fmt: printf-style format string | |
5509 | * @...: arguments for the format string | |
5510 | * | |
5511 | * This function can be called by a running work function to describe what | |
5512 | * the work item is about. If the worker task gets dumped, this | |
5513 | * information will be printed out together to help debugging. The | |
5514 | * description can be at most WORKER_DESC_LEN including the trailing '\0'. | |
5515 | */ | |
5516 | void set_worker_desc(const char *fmt, ...) | |
5517 | { | |
5518 | struct worker *worker = current_wq_worker(); | |
5519 | va_list args; | |
5520 | ||
5521 | if (worker) { | |
5522 | va_start(args, fmt); | |
5523 | vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); | |
5524 | va_end(args); | |
3d1cb205 TH |
5525 | } |
5526 | } | |
5c750d58 | 5527 | EXPORT_SYMBOL_GPL(set_worker_desc); |
3d1cb205 TH |
5528 | |
5529 | /** | |
5530 | * print_worker_info - print out worker information and description | |
5531 | * @log_lvl: the log level to use when printing | |
5532 | * @task: target task | |
5533 | * | |
5534 | * If @task is a worker and currently executing a work item, print out the | |
5535 | * name of the workqueue being serviced and worker description set with | |
5536 | * set_worker_desc() by the currently executing work item. | |
5537 | * | |
5538 | * This function can be safely called on any task as long as the | |
5539 | * task_struct itself is accessible. While safe, this function isn't | |
5540 | * synchronized and may print out mixups or garbages of limited length. | |
5541 | */ | |
5542 | void print_worker_info(const char *log_lvl, struct task_struct *task) | |
5543 | { | |
5544 | work_func_t *fn = NULL; | |
5545 | char name[WQ_NAME_LEN] = { }; | |
5546 | char desc[WORKER_DESC_LEN] = { }; | |
5547 | struct pool_workqueue *pwq = NULL; | |
5548 | struct workqueue_struct *wq = NULL; | |
3d1cb205 TH |
5549 | struct worker *worker; |
5550 | ||
5551 | if (!(task->flags & PF_WQ_WORKER)) | |
5552 | return; | |
5553 | ||
5554 | /* | |
5555 | * This function is called without any synchronization and @task | |
5556 | * could be in any state. Be careful with dereferences. | |
5557 | */ | |
e700591a | 5558 | worker = kthread_probe_data(task); |
3d1cb205 TH |
5559 | |
5560 | /* | |
8bf89593 TH |
5561 | * Carefully copy the associated workqueue's workfn, name and desc. |
5562 | * Keep the original last '\0' in case the original is garbage. | |
3d1cb205 | 5563 | */ |
fe557319 CH |
5564 | copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn)); |
5565 | copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq)); | |
5566 | copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq)); | |
5567 | copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1); | |
5568 | copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1); | |
3d1cb205 TH |
5569 | |
5570 | if (fn || name[0] || desc[0]) { | |
d75f773c | 5571 | printk("%sWorkqueue: %s %ps", log_lvl, name, fn); |
8bf89593 | 5572 | if (strcmp(name, desc)) |
3d1cb205 TH |
5573 | pr_cont(" (%s)", desc); |
5574 | pr_cont("\n"); | |
5575 | } | |
5576 | } | |
5577 | ||
3494fc30 TH |
5578 | static void pr_cont_pool_info(struct worker_pool *pool) |
5579 | { | |
5580 | pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask); | |
5581 | if (pool->node != NUMA_NO_NODE) | |
5582 | pr_cont(" node=%d", pool->node); | |
5583 | pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice); | |
5584 | } | |
5585 | ||
c76feb0d PM |
5586 | struct pr_cont_work_struct { |
5587 | bool comma; | |
5588 | work_func_t func; | |
5589 | long ctr; | |
5590 | }; | |
5591 | ||
5592 | static void pr_cont_work_flush(bool comma, work_func_t func, struct pr_cont_work_struct *pcwsp) | |
5593 | { | |
5594 | if (!pcwsp->ctr) | |
5595 | goto out_record; | |
5596 | if (func == pcwsp->func) { | |
5597 | pcwsp->ctr++; | |
5598 | return; | |
5599 | } | |
5600 | if (pcwsp->ctr == 1) | |
5601 | pr_cont("%s %ps", pcwsp->comma ? "," : "", pcwsp->func); | |
5602 | else | |
5603 | pr_cont("%s %ld*%ps", pcwsp->comma ? "," : "", pcwsp->ctr, pcwsp->func); | |
5604 | pcwsp->ctr = 0; | |
5605 | out_record: | |
5606 | if ((long)func == -1L) | |
5607 | return; | |
5608 | pcwsp->comma = comma; | |
5609 | pcwsp->func = func; | |
5610 | pcwsp->ctr = 1; | |
5611 | } | |
5612 | ||
5613 | static void pr_cont_work(bool comma, struct work_struct *work, struct pr_cont_work_struct *pcwsp) | |
3494fc30 TH |
5614 | { |
5615 | if (work->func == wq_barrier_func) { | |
5616 | struct wq_barrier *barr; | |
5617 | ||
5618 | barr = container_of(work, struct wq_barrier, work); | |
5619 | ||
c76feb0d | 5620 | pr_cont_work_flush(comma, (work_func_t)-1, pcwsp); |
3494fc30 TH |
5621 | pr_cont("%s BAR(%d)", comma ? "," : "", |
5622 | task_pid_nr(barr->task)); | |
5623 | } else { | |
c76feb0d PM |
5624 | if (!comma) |
5625 | pr_cont_work_flush(comma, (work_func_t)-1, pcwsp); | |
5626 | pr_cont_work_flush(comma, work->func, pcwsp); | |
3494fc30 TH |
5627 | } |
5628 | } | |
5629 | ||
5630 | static void show_pwq(struct pool_workqueue *pwq) | |
5631 | { | |
c76feb0d | 5632 | struct pr_cont_work_struct pcws = { .ctr = 0, }; |
3494fc30 TH |
5633 | struct worker_pool *pool = pwq->pool; |
5634 | struct work_struct *work; | |
5635 | struct worker *worker; | |
5636 | bool has_in_flight = false, has_pending = false; | |
5637 | int bkt; | |
5638 | ||
5639 | pr_info(" pwq %d:", pool->id); | |
5640 | pr_cont_pool_info(pool); | |
5641 | ||
a045a272 TH |
5642 | pr_cont(" active=%d refcnt=%d%s\n", |
5643 | pwq->nr_active, pwq->refcnt, | |
3494fc30 TH |
5644 | !list_empty(&pwq->mayday_node) ? " MAYDAY" : ""); |
5645 | ||
5646 | hash_for_each(pool->busy_hash, bkt, worker, hentry) { | |
5647 | if (worker->current_pwq == pwq) { | |
5648 | has_in_flight = true; | |
5649 | break; | |
5650 | } | |
5651 | } | |
5652 | if (has_in_flight) { | |
5653 | bool comma = false; | |
5654 | ||
5655 | pr_info(" in-flight:"); | |
5656 | hash_for_each(pool->busy_hash, bkt, worker, hentry) { | |
5657 | if (worker->current_pwq != pwq) | |
5658 | continue; | |
5659 | ||
d75f773c | 5660 | pr_cont("%s %d%s:%ps", comma ? "," : "", |
3494fc30 | 5661 | task_pid_nr(worker->task), |
30ae2fc0 | 5662 | worker->rescue_wq ? "(RESCUER)" : "", |
3494fc30 TH |
5663 | worker->current_func); |
5664 | list_for_each_entry(work, &worker->scheduled, entry) | |
c76feb0d PM |
5665 | pr_cont_work(false, work, &pcws); |
5666 | pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); | |
3494fc30 TH |
5667 | comma = true; |
5668 | } | |
5669 | pr_cont("\n"); | |
5670 | } | |
5671 | ||
5672 | list_for_each_entry(work, &pool->worklist, entry) { | |
5673 | if (get_work_pwq(work) == pwq) { | |
5674 | has_pending = true; | |
5675 | break; | |
5676 | } | |
5677 | } | |
5678 | if (has_pending) { | |
5679 | bool comma = false; | |
5680 | ||
5681 | pr_info(" pending:"); | |
5682 | list_for_each_entry(work, &pool->worklist, entry) { | |
5683 | if (get_work_pwq(work) != pwq) | |
5684 | continue; | |
5685 | ||
c76feb0d | 5686 | pr_cont_work(comma, work, &pcws); |
3494fc30 TH |
5687 | comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); |
5688 | } | |
c76feb0d | 5689 | pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); |
3494fc30 TH |
5690 | pr_cont("\n"); |
5691 | } | |
5692 | ||
f97a4a1a | 5693 | if (!list_empty(&pwq->inactive_works)) { |
3494fc30 TH |
5694 | bool comma = false; |
5695 | ||
f97a4a1a LJ |
5696 | pr_info(" inactive:"); |
5697 | list_for_each_entry(work, &pwq->inactive_works, entry) { | |
c76feb0d | 5698 | pr_cont_work(comma, work, &pcws); |
3494fc30 TH |
5699 | comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); |
5700 | } | |
c76feb0d | 5701 | pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); |
3494fc30 TH |
5702 | pr_cont("\n"); |
5703 | } | |
5704 | } | |
5705 | ||
5706 | /** | |
55df0933 IK |
5707 | * show_one_workqueue - dump state of specified workqueue |
5708 | * @wq: workqueue whose state will be printed | |
3494fc30 | 5709 | */ |
55df0933 | 5710 | void show_one_workqueue(struct workqueue_struct *wq) |
3494fc30 | 5711 | { |
55df0933 IK |
5712 | struct pool_workqueue *pwq; |
5713 | bool idle = true; | |
3494fc30 | 5714 | unsigned long flags; |
3494fc30 | 5715 | |
55df0933 | 5716 | for_each_pwq(pwq, wq) { |
afa87ce8 | 5717 | if (!pwq_is_empty(pwq)) { |
55df0933 IK |
5718 | idle = false; |
5719 | break; | |
3494fc30 | 5720 | } |
55df0933 IK |
5721 | } |
5722 | if (idle) /* Nothing to print for idle workqueue */ | |
5723 | return; | |
3494fc30 | 5724 | |
55df0933 | 5725 | pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags); |
3494fc30 | 5726 | |
55df0933 IK |
5727 | for_each_pwq(pwq, wq) { |
5728 | raw_spin_lock_irqsave(&pwq->pool->lock, flags); | |
afa87ce8 | 5729 | if (!pwq_is_empty(pwq)) { |
62635ea8 | 5730 | /* |
55df0933 IK |
5731 | * Defer printing to avoid deadlocks in console |
5732 | * drivers that queue work while holding locks | |
5733 | * also taken in their write paths. | |
62635ea8 | 5734 | */ |
55df0933 IK |
5735 | printk_deferred_enter(); |
5736 | show_pwq(pwq); | |
5737 | printk_deferred_exit(); | |
3494fc30 | 5738 | } |
55df0933 | 5739 | raw_spin_unlock_irqrestore(&pwq->pool->lock, flags); |
62635ea8 SS |
5740 | /* |
5741 | * We could be printing a lot from atomic context, e.g. | |
55df0933 | 5742 | * sysrq-t -> show_all_workqueues(). Avoid triggering |
62635ea8 SS |
5743 | * hard lockup. |
5744 | */ | |
5745 | touch_nmi_watchdog(); | |
3494fc30 TH |
5746 | } |
5747 | ||
55df0933 IK |
5748 | } |
5749 | ||
5750 | /** | |
5751 | * show_one_worker_pool - dump state of specified worker pool | |
5752 | * @pool: worker pool whose state will be printed | |
5753 | */ | |
5754 | static void show_one_worker_pool(struct worker_pool *pool) | |
5755 | { | |
5756 | struct worker *worker; | |
5757 | bool first = true; | |
5758 | unsigned long flags; | |
335a42eb | 5759 | unsigned long hung = 0; |
55df0933 IK |
5760 | |
5761 | raw_spin_lock_irqsave(&pool->lock, flags); | |
5762 | if (pool->nr_workers == pool->nr_idle) | |
5763 | goto next_pool; | |
335a42eb PM |
5764 | |
5765 | /* How long the first pending work is waiting for a worker. */ | |
5766 | if (!list_empty(&pool->worklist)) | |
5767 | hung = jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000; | |
5768 | ||
55df0933 IK |
5769 | /* |
5770 | * Defer printing to avoid deadlocks in console drivers that | |
5771 | * queue work while holding locks also taken in their write | |
5772 | * paths. | |
5773 | */ | |
5774 | printk_deferred_enter(); | |
5775 | pr_info("pool %d:", pool->id); | |
5776 | pr_cont_pool_info(pool); | |
335a42eb | 5777 | pr_cont(" hung=%lus workers=%d", hung, pool->nr_workers); |
55df0933 IK |
5778 | if (pool->manager) |
5779 | pr_cont(" manager: %d", | |
5780 | task_pid_nr(pool->manager->task)); | |
5781 | list_for_each_entry(worker, &pool->idle_list, entry) { | |
5782 | pr_cont(" %s%d", first ? "idle: " : "", | |
5783 | task_pid_nr(worker->task)); | |
5784 | first = false; | |
5785 | } | |
5786 | pr_cont("\n"); | |
5787 | printk_deferred_exit(); | |
5788 | next_pool: | |
5789 | raw_spin_unlock_irqrestore(&pool->lock, flags); | |
5790 | /* | |
5791 | * We could be printing a lot from atomic context, e.g. | |
5792 | * sysrq-t -> show_all_workqueues(). Avoid triggering | |
5793 | * hard lockup. | |
5794 | */ | |
5795 | touch_nmi_watchdog(); | |
5796 | ||
5797 | } | |
5798 | ||
5799 | /** | |
5800 | * show_all_workqueues - dump workqueue state | |
5801 | * | |
704bc669 | 5802 | * Called from a sysrq handler and prints out all busy workqueues and pools. |
55df0933 IK |
5803 | */ |
5804 | void show_all_workqueues(void) | |
5805 | { | |
5806 | struct workqueue_struct *wq; | |
5807 | struct worker_pool *pool; | |
5808 | int pi; | |
5809 | ||
5810 | rcu_read_lock(); | |
5811 | ||
5812 | pr_info("Showing busy workqueues and worker pools:\n"); | |
5813 | ||
5814 | list_for_each_entry_rcu(wq, &workqueues, list) | |
5815 | show_one_workqueue(wq); | |
5816 | ||
5817 | for_each_pool(pool, pi) | |
5818 | show_one_worker_pool(pool); | |
5819 | ||
24acfb71 | 5820 | rcu_read_unlock(); |
3494fc30 TH |
5821 | } |
5822 | ||
704bc669 JL |
5823 | /** |
5824 | * show_freezable_workqueues - dump freezable workqueue state | |
5825 | * | |
5826 | * Called from try_to_freeze_tasks() and prints out all freezable workqueues | |
5827 | * still busy. | |
5828 | */ | |
5829 | void show_freezable_workqueues(void) | |
5830 | { | |
5831 | struct workqueue_struct *wq; | |
5832 | ||
5833 | rcu_read_lock(); | |
5834 | ||
5835 | pr_info("Showing freezable workqueues that are still busy:\n"); | |
5836 | ||
5837 | list_for_each_entry_rcu(wq, &workqueues, list) { | |
5838 | if (!(wq->flags & WQ_FREEZABLE)) | |
5839 | continue; | |
5840 | show_one_workqueue(wq); | |
5841 | } | |
5842 | ||
5843 | rcu_read_unlock(); | |
5844 | } | |
5845 | ||
6b59808b TH |
5846 | /* used to show worker information through /proc/PID/{comm,stat,status} */ |
5847 | void wq_worker_comm(char *buf, size_t size, struct task_struct *task) | |
5848 | { | |
6b59808b TH |
5849 | int off; |
5850 | ||
5851 | /* always show the actual comm */ | |
5852 | off = strscpy(buf, task->comm, size); | |
5853 | if (off < 0) | |
5854 | return; | |
5855 | ||
197f6acc | 5856 | /* stabilize PF_WQ_WORKER and worker pool association */ |
6b59808b TH |
5857 | mutex_lock(&wq_pool_attach_mutex); |
5858 | ||
197f6acc TH |
5859 | if (task->flags & PF_WQ_WORKER) { |
5860 | struct worker *worker = kthread_data(task); | |
5861 | struct worker_pool *pool = worker->pool; | |
6b59808b | 5862 | |
197f6acc | 5863 | if (pool) { |
a9b8a985 | 5864 | raw_spin_lock_irq(&pool->lock); |
197f6acc TH |
5865 | /* |
5866 | * ->desc tracks information (wq name or | |
5867 | * set_worker_desc()) for the latest execution. If | |
5868 | * current, prepend '+', otherwise '-'. | |
5869 | */ | |
5870 | if (worker->desc[0] != '\0') { | |
5871 | if (worker->current_work) | |
5872 | scnprintf(buf + off, size - off, "+%s", | |
5873 | worker->desc); | |
5874 | else | |
5875 | scnprintf(buf + off, size - off, "-%s", | |
5876 | worker->desc); | |
5877 | } | |
a9b8a985 | 5878 | raw_spin_unlock_irq(&pool->lock); |
6b59808b | 5879 | } |
6b59808b TH |
5880 | } |
5881 | ||
5882 | mutex_unlock(&wq_pool_attach_mutex); | |
5883 | } | |
5884 | ||
66448bc2 MM |
5885 | #ifdef CONFIG_SMP |
5886 | ||
db7bccf4 TH |
5887 | /* |
5888 | * CPU hotplug. | |
5889 | * | |
e22bee78 | 5890 | * There are two challenges in supporting CPU hotplug. Firstly, there |
112202d9 | 5891 | * are a lot of assumptions on strong associations among work, pwq and |
706026c2 | 5892 | * pool which make migrating pending and scheduled works very |
e22bee78 | 5893 | * difficult to implement without impacting hot paths. Secondly, |
94cf58bb | 5894 | * worker pools serve mix of short, long and very long running works making |
e22bee78 TH |
5895 | * blocked draining impractical. |
5896 | * | |
24647570 | 5897 | * This is solved by allowing the pools to be disassociated from the CPU |
628c78e7 TH |
5898 | * running as an unbound one and allowing it to be reattached later if the |
5899 | * cpu comes back online. | |
db7bccf4 | 5900 | */ |
1da177e4 | 5901 | |
e8b3f8db | 5902 | static void unbind_workers(int cpu) |
3af24433 | 5903 | { |
4ce62e9e | 5904 | struct worker_pool *pool; |
db7bccf4 | 5905 | struct worker *worker; |
3af24433 | 5906 | |
f02ae73a | 5907 | for_each_cpu_worker_pool(pool, cpu) { |
1258fae7 | 5908 | mutex_lock(&wq_pool_attach_mutex); |
a9b8a985 | 5909 | raw_spin_lock_irq(&pool->lock); |
3af24433 | 5910 | |
94cf58bb | 5911 | /* |
92f9c5c4 | 5912 | * We've blocked all attach/detach operations. Make all workers |
94cf58bb | 5913 | * unbound and set DISASSOCIATED. Before this, all workers |
11b45b0b | 5914 | * must be on the cpu. After this, they may become diasporas. |
b4ac9384 LJ |
5915 | * And the preemption disabled section in their sched callbacks |
5916 | * are guaranteed to see WORKER_UNBOUND since the code here | |
5917 | * is on the same cpu. | |
94cf58bb | 5918 | */ |
da028469 | 5919 | for_each_pool_worker(worker, pool) |
c9e7cf27 | 5920 | worker->flags |= WORKER_UNBOUND; |
06ba38a9 | 5921 | |
24647570 | 5922 | pool->flags |= POOL_DISASSOCIATED; |
f2d5a0ee | 5923 | |
eb283428 | 5924 | /* |
989442d7 LJ |
5925 | * The handling of nr_running in sched callbacks are disabled |
5926 | * now. Zap nr_running. After this, nr_running stays zero and | |
5927 | * need_more_worker() and keep_working() are always true as | |
5928 | * long as the worklist is not empty. This pool now behaves as | |
5929 | * an unbound (in terms of concurrency management) pool which | |
eb283428 LJ |
5930 | * are served by workers tied to the pool. |
5931 | */ | |
bc35f7ef | 5932 | pool->nr_running = 0; |
eb283428 LJ |
5933 | |
5934 | /* | |
5935 | * With concurrency management just turned off, a busy | |
5936 | * worker blocking could lead to lengthy stalls. Kick off | |
5937 | * unbound chain execution of currently pending work items. | |
5938 | */ | |
0219a352 | 5939 | kick_pool(pool); |
989442d7 | 5940 | |
a9b8a985 | 5941 | raw_spin_unlock_irq(&pool->lock); |
989442d7 | 5942 | |
793777bc VS |
5943 | for_each_pool_worker(worker, pool) |
5944 | unbind_worker(worker); | |
989442d7 LJ |
5945 | |
5946 | mutex_unlock(&wq_pool_attach_mutex); | |
eb283428 | 5947 | } |
3af24433 | 5948 | } |
3af24433 | 5949 | |
bd7c089e TH |
5950 | /** |
5951 | * rebind_workers - rebind all workers of a pool to the associated CPU | |
5952 | * @pool: pool of interest | |
5953 | * | |
a9ab775b | 5954 | * @pool->cpu is coming online. Rebind all workers to the CPU. |
bd7c089e TH |
5955 | */ |
5956 | static void rebind_workers(struct worker_pool *pool) | |
5957 | { | |
a9ab775b | 5958 | struct worker *worker; |
bd7c089e | 5959 | |
1258fae7 | 5960 | lockdep_assert_held(&wq_pool_attach_mutex); |
bd7c089e | 5961 | |
a9ab775b TH |
5962 | /* |
5963 | * Restore CPU affinity of all workers. As all idle workers should | |
5964 | * be on the run-queue of the associated CPU before any local | |
402dd89d | 5965 | * wake-ups for concurrency management happen, restore CPU affinity |
a9ab775b TH |
5966 | * of all workers first and then clear UNBOUND. As we're called |
5967 | * from CPU_ONLINE, the following shouldn't fail. | |
5968 | */ | |
c63a2e52 VS |
5969 | for_each_pool_worker(worker, pool) { |
5970 | kthread_set_per_cpu(worker->task, pool->cpu); | |
5971 | WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, | |
9546b29e | 5972 | pool_allowed_cpus(pool)) < 0); |
c63a2e52 | 5973 | } |
bd7c089e | 5974 | |
a9b8a985 | 5975 | raw_spin_lock_irq(&pool->lock); |
f7c17d26 | 5976 | |
3de5e884 | 5977 | pool->flags &= ~POOL_DISASSOCIATED; |
bd7c089e | 5978 | |
da028469 | 5979 | for_each_pool_worker(worker, pool) { |
a9ab775b | 5980 | unsigned int worker_flags = worker->flags; |
bd7c089e | 5981 | |
a9ab775b TH |
5982 | /* |
5983 | * We want to clear UNBOUND but can't directly call | |
5984 | * worker_clr_flags() or adjust nr_running. Atomically | |
5985 | * replace UNBOUND with another NOT_RUNNING flag REBOUND. | |
5986 | * @worker will clear REBOUND using worker_clr_flags() when | |
5987 | * it initiates the next execution cycle thus restoring | |
5988 | * concurrency management. Note that when or whether | |
5989 | * @worker clears REBOUND doesn't affect correctness. | |
5990 | * | |
c95491ed | 5991 | * WRITE_ONCE() is necessary because @worker->flags may be |
a9ab775b | 5992 | * tested without holding any lock in |
6d25be57 | 5993 | * wq_worker_running(). Without it, NOT_RUNNING test may |
a9ab775b TH |
5994 | * fail incorrectly leading to premature concurrency |
5995 | * management operations. | |
5996 | */ | |
5997 | WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); | |
5998 | worker_flags |= WORKER_REBOUND; | |
5999 | worker_flags &= ~WORKER_UNBOUND; | |
c95491ed | 6000 | WRITE_ONCE(worker->flags, worker_flags); |
bd7c089e | 6001 | } |
a9ab775b | 6002 | |
a9b8a985 | 6003 | raw_spin_unlock_irq(&pool->lock); |
bd7c089e TH |
6004 | } |
6005 | ||
7dbc725e TH |
6006 | /** |
6007 | * restore_unbound_workers_cpumask - restore cpumask of unbound workers | |
6008 | * @pool: unbound pool of interest | |
6009 | * @cpu: the CPU which is coming up | |
6010 | * | |
6011 | * An unbound pool may end up with a cpumask which doesn't have any online | |
6012 | * CPUs. When a worker of such pool get scheduled, the scheduler resets | |
6013 | * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any | |
6014 | * online CPU before, cpus_allowed of all its workers should be restored. | |
6015 | */ | |
6016 | static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) | |
6017 | { | |
6018 | static cpumask_t cpumask; | |
6019 | struct worker *worker; | |
7dbc725e | 6020 | |
1258fae7 | 6021 | lockdep_assert_held(&wq_pool_attach_mutex); |
7dbc725e TH |
6022 | |
6023 | /* is @cpu allowed for @pool? */ | |
6024 | if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) | |
6025 | return; | |
6026 | ||
7dbc725e | 6027 | cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); |
7dbc725e TH |
6028 | |
6029 | /* as we're called from CPU_ONLINE, the following shouldn't fail */ | |
da028469 | 6030 | for_each_pool_worker(worker, pool) |
d945b5e9 | 6031 | WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0); |
7dbc725e TH |
6032 | } |
6033 | ||
7ee681b2 TG |
6034 | int workqueue_prepare_cpu(unsigned int cpu) |
6035 | { | |
6036 | struct worker_pool *pool; | |
6037 | ||
6038 | for_each_cpu_worker_pool(pool, cpu) { | |
6039 | if (pool->nr_workers) | |
6040 | continue; | |
6041 | if (!create_worker(pool)) | |
6042 | return -ENOMEM; | |
6043 | } | |
6044 | return 0; | |
6045 | } | |
6046 | ||
6047 | int workqueue_online_cpu(unsigned int cpu) | |
3af24433 | 6048 | { |
4ce62e9e | 6049 | struct worker_pool *pool; |
4c16bd32 | 6050 | struct workqueue_struct *wq; |
7dbc725e | 6051 | int pi; |
3ce63377 | 6052 | |
7ee681b2 | 6053 | mutex_lock(&wq_pool_mutex); |
7dbc725e | 6054 | |
7ee681b2 | 6055 | for_each_pool(pool, pi) { |
1258fae7 | 6056 | mutex_lock(&wq_pool_attach_mutex); |
94cf58bb | 6057 | |
7ee681b2 TG |
6058 | if (pool->cpu == cpu) |
6059 | rebind_workers(pool); | |
6060 | else if (pool->cpu < 0) | |
6061 | restore_unbound_workers_cpumask(pool, cpu); | |
94cf58bb | 6062 | |
1258fae7 | 6063 | mutex_unlock(&wq_pool_attach_mutex); |
7ee681b2 | 6064 | } |
6ba94429 | 6065 | |
fef59c9c | 6066 | /* update pod affinity of unbound workqueues */ |
4cbfd3de | 6067 | list_for_each_entry(wq, &workqueues, list) { |
84193c07 TH |
6068 | struct workqueue_attrs *attrs = wq->unbound_attrs; |
6069 | ||
6070 | if (attrs) { | |
6071 | const struct wq_pod_type *pt = wqattrs_pod_type(attrs); | |
6072 | int tcpu; | |
4cbfd3de | 6073 | |
84193c07 | 6074 | for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]]) |
fef59c9c | 6075 | wq_update_pod(wq, tcpu, cpu, true); |
5797b1c1 TH |
6076 | |
6077 | mutex_lock(&wq->mutex); | |
6078 | wq_update_node_max_active(wq, -1); | |
6079 | mutex_unlock(&wq->mutex); | |
4cbfd3de TH |
6080 | } |
6081 | } | |
6ba94429 | 6082 | |
7ee681b2 TG |
6083 | mutex_unlock(&wq_pool_mutex); |
6084 | return 0; | |
6ba94429 FW |
6085 | } |
6086 | ||
7ee681b2 | 6087 | int workqueue_offline_cpu(unsigned int cpu) |
6ba94429 | 6088 | { |
6ba94429 FW |
6089 | struct workqueue_struct *wq; |
6090 | ||
7ee681b2 | 6091 | /* unbinding per-cpu workers should happen on the local CPU */ |
e8b3f8db LJ |
6092 | if (WARN_ON(cpu != smp_processor_id())) |
6093 | return -1; | |
6094 | ||
6095 | unbind_workers(cpu); | |
7ee681b2 | 6096 | |
fef59c9c | 6097 | /* update pod affinity of unbound workqueues */ |
7ee681b2 | 6098 | mutex_lock(&wq_pool_mutex); |
4cbfd3de | 6099 | list_for_each_entry(wq, &workqueues, list) { |
84193c07 TH |
6100 | struct workqueue_attrs *attrs = wq->unbound_attrs; |
6101 | ||
6102 | if (attrs) { | |
6103 | const struct wq_pod_type *pt = wqattrs_pod_type(attrs); | |
6104 | int tcpu; | |
4cbfd3de | 6105 | |
84193c07 | 6106 | for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]]) |
fef59c9c | 6107 | wq_update_pod(wq, tcpu, cpu, false); |
5797b1c1 TH |
6108 | |
6109 | mutex_lock(&wq->mutex); | |
6110 | wq_update_node_max_active(wq, cpu); | |
6111 | mutex_unlock(&wq->mutex); | |
4cbfd3de TH |
6112 | } |
6113 | } | |
7ee681b2 TG |
6114 | mutex_unlock(&wq_pool_mutex); |
6115 | ||
7ee681b2 | 6116 | return 0; |
6ba94429 FW |
6117 | } |
6118 | ||
6ba94429 FW |
6119 | struct work_for_cpu { |
6120 | struct work_struct work; | |
6121 | long (*fn)(void *); | |
6122 | void *arg; | |
6123 | long ret; | |
6124 | }; | |
6125 | ||
6126 | static void work_for_cpu_fn(struct work_struct *work) | |
6127 | { | |
6128 | struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); | |
6129 | ||
6130 | wfc->ret = wfc->fn(wfc->arg); | |
6131 | } | |
6132 | ||
6133 | /** | |
265f3ed0 | 6134 | * work_on_cpu_key - run a function in thread context on a particular cpu |
6ba94429 FW |
6135 | * @cpu: the cpu to run on |
6136 | * @fn: the function to run | |
6137 | * @arg: the function arg | |
265f3ed0 | 6138 | * @key: The lock class key for lock debugging purposes |
6ba94429 FW |
6139 | * |
6140 | * It is up to the caller to ensure that the cpu doesn't go offline. | |
6141 | * The caller must not hold any locks which would prevent @fn from completing. | |
6142 | * | |
6143 | * Return: The value @fn returns. | |
6144 | */ | |
265f3ed0 FW |
6145 | long work_on_cpu_key(int cpu, long (*fn)(void *), |
6146 | void *arg, struct lock_class_key *key) | |
6ba94429 FW |
6147 | { |
6148 | struct work_for_cpu wfc = { .fn = fn, .arg = arg }; | |
6149 | ||
265f3ed0 | 6150 | INIT_WORK_ONSTACK_KEY(&wfc.work, work_for_cpu_fn, key); |
6ba94429 FW |
6151 | schedule_work_on(cpu, &wfc.work); |
6152 | flush_work(&wfc.work); | |
6153 | destroy_work_on_stack(&wfc.work); | |
6154 | return wfc.ret; | |
6155 | } | |
265f3ed0 | 6156 | EXPORT_SYMBOL_GPL(work_on_cpu_key); |
0e8d6a93 TG |
6157 | |
6158 | /** | |
265f3ed0 | 6159 | * work_on_cpu_safe_key - run a function in thread context on a particular cpu |
0e8d6a93 TG |
6160 | * @cpu: the cpu to run on |
6161 | * @fn: the function to run | |
6162 | * @arg: the function argument | |
265f3ed0 | 6163 | * @key: The lock class key for lock debugging purposes |
0e8d6a93 TG |
6164 | * |
6165 | * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold | |
6166 | * any locks which would prevent @fn from completing. | |
6167 | * | |
6168 | * Return: The value @fn returns. | |
6169 | */ | |
265f3ed0 FW |
6170 | long work_on_cpu_safe_key(int cpu, long (*fn)(void *), |
6171 | void *arg, struct lock_class_key *key) | |
0e8d6a93 TG |
6172 | { |
6173 | long ret = -ENODEV; | |
6174 | ||
ffd8bea8 | 6175 | cpus_read_lock(); |
0e8d6a93 | 6176 | if (cpu_online(cpu)) |
265f3ed0 | 6177 | ret = work_on_cpu_key(cpu, fn, arg, key); |
ffd8bea8 | 6178 | cpus_read_unlock(); |
0e8d6a93 TG |
6179 | return ret; |
6180 | } | |
265f3ed0 | 6181 | EXPORT_SYMBOL_GPL(work_on_cpu_safe_key); |
6ba94429 FW |
6182 | #endif /* CONFIG_SMP */ |
6183 | ||
6184 | #ifdef CONFIG_FREEZER | |
6185 | ||
6186 | /** | |
6187 | * freeze_workqueues_begin - begin freezing workqueues | |
6188 | * | |
6189 | * Start freezing workqueues. After this function returns, all freezable | |
f97a4a1a | 6190 | * workqueues will queue new works to their inactive_works list instead of |
6ba94429 FW |
6191 | * pool->worklist. |
6192 | * | |
6193 | * CONTEXT: | |
6194 | * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. | |
6195 | */ | |
6196 | void freeze_workqueues_begin(void) | |
6197 | { | |
6198 | struct workqueue_struct *wq; | |
6ba94429 FW |
6199 | |
6200 | mutex_lock(&wq_pool_mutex); | |
6201 | ||
6202 | WARN_ON_ONCE(workqueue_freezing); | |
6203 | workqueue_freezing = true; | |
6204 | ||
6205 | list_for_each_entry(wq, &workqueues, list) { | |
6206 | mutex_lock(&wq->mutex); | |
a045a272 | 6207 | wq_adjust_max_active(wq); |
6ba94429 FW |
6208 | mutex_unlock(&wq->mutex); |
6209 | } | |
6210 | ||
6211 | mutex_unlock(&wq_pool_mutex); | |
6212 | } | |
6213 | ||
6214 | /** | |
6215 | * freeze_workqueues_busy - are freezable workqueues still busy? | |
6216 | * | |
6217 | * Check whether freezing is complete. This function must be called | |
6218 | * between freeze_workqueues_begin() and thaw_workqueues(). | |
6219 | * | |
6220 | * CONTEXT: | |
6221 | * Grabs and releases wq_pool_mutex. | |
6222 | * | |
6223 | * Return: | |
6224 | * %true if some freezable workqueues are still busy. %false if freezing | |
6225 | * is complete. | |
6226 | */ | |
6227 | bool freeze_workqueues_busy(void) | |
6228 | { | |
6229 | bool busy = false; | |
6230 | struct workqueue_struct *wq; | |
6231 | struct pool_workqueue *pwq; | |
6232 | ||
6233 | mutex_lock(&wq_pool_mutex); | |
6234 | ||
6235 | WARN_ON_ONCE(!workqueue_freezing); | |
6236 | ||
6237 | list_for_each_entry(wq, &workqueues, list) { | |
6238 | if (!(wq->flags & WQ_FREEZABLE)) | |
6239 | continue; | |
6240 | /* | |
6241 | * nr_active is monotonically decreasing. It's safe | |
6242 | * to peek without lock. | |
6243 | */ | |
24acfb71 | 6244 | rcu_read_lock(); |
6ba94429 FW |
6245 | for_each_pwq(pwq, wq) { |
6246 | WARN_ON_ONCE(pwq->nr_active < 0); | |
6247 | if (pwq->nr_active) { | |
6248 | busy = true; | |
24acfb71 | 6249 | rcu_read_unlock(); |
6ba94429 FW |
6250 | goto out_unlock; |
6251 | } | |
6252 | } | |
24acfb71 | 6253 | rcu_read_unlock(); |
6ba94429 FW |
6254 | } |
6255 | out_unlock: | |
6256 | mutex_unlock(&wq_pool_mutex); | |
6257 | return busy; | |
6258 | } | |
6259 | ||
6260 | /** | |
6261 | * thaw_workqueues - thaw workqueues | |
6262 | * | |
6263 | * Thaw workqueues. Normal queueing is restored and all collected | |
6264 | * frozen works are transferred to their respective pool worklists. | |
6265 | * | |
6266 | * CONTEXT: | |
6267 | * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. | |
6268 | */ | |
6269 | void thaw_workqueues(void) | |
6270 | { | |
6271 | struct workqueue_struct *wq; | |
6ba94429 FW |
6272 | |
6273 | mutex_lock(&wq_pool_mutex); | |
6274 | ||
6275 | if (!workqueue_freezing) | |
6276 | goto out_unlock; | |
6277 | ||
6278 | workqueue_freezing = false; | |
6279 | ||
6280 | /* restore max_active and repopulate worklist */ | |
6281 | list_for_each_entry(wq, &workqueues, list) { | |
6282 | mutex_lock(&wq->mutex); | |
a045a272 | 6283 | wq_adjust_max_active(wq); |
6ba94429 FW |
6284 | mutex_unlock(&wq->mutex); |
6285 | } | |
6286 | ||
6287 | out_unlock: | |
6288 | mutex_unlock(&wq_pool_mutex); | |
6289 | } | |
6290 | #endif /* CONFIG_FREEZER */ | |
6291 | ||
99c621ef | 6292 | static int workqueue_apply_unbound_cpumask(const cpumask_var_t unbound_cpumask) |
042f7df1 LJ |
6293 | { |
6294 | LIST_HEAD(ctxs); | |
6295 | int ret = 0; | |
6296 | struct workqueue_struct *wq; | |
6297 | struct apply_wqattrs_ctx *ctx, *n; | |
6298 | ||
6299 | lockdep_assert_held(&wq_pool_mutex); | |
6300 | ||
6301 | list_for_each_entry(wq, &workqueues, list) { | |
6302 | if (!(wq->flags & WQ_UNBOUND)) | |
6303 | continue; | |
ca10d851 | 6304 | |
042f7df1 | 6305 | /* creating multiple pwqs breaks ordering guarantee */ |
ca10d851 WL |
6306 | if (!list_empty(&wq->pwqs)) { |
6307 | if (wq->flags & __WQ_ORDERED_EXPLICIT) | |
6308 | continue; | |
6309 | wq->flags &= ~__WQ_ORDERED; | |
6310 | } | |
042f7df1 | 6311 | |
99c621ef | 6312 | ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs, unbound_cpumask); |
84193c07 TH |
6313 | if (IS_ERR(ctx)) { |
6314 | ret = PTR_ERR(ctx); | |
042f7df1 LJ |
6315 | break; |
6316 | } | |
6317 | ||
6318 | list_add_tail(&ctx->list, &ctxs); | |
6319 | } | |
6320 | ||
6321 | list_for_each_entry_safe(ctx, n, &ctxs, list) { | |
6322 | if (!ret) | |
6323 | apply_wqattrs_commit(ctx); | |
6324 | apply_wqattrs_cleanup(ctx); | |
6325 | } | |
6326 | ||
99c621ef LJ |
6327 | if (!ret) { |
6328 | mutex_lock(&wq_pool_attach_mutex); | |
6329 | cpumask_copy(wq_unbound_cpumask, unbound_cpumask); | |
6330 | mutex_unlock(&wq_pool_attach_mutex); | |
6331 | } | |
042f7df1 LJ |
6332 | return ret; |
6333 | } | |
6334 | ||
fe28f631 WL |
6335 | /** |
6336 | * workqueue_unbound_exclude_cpumask - Exclude given CPUs from unbound cpumask | |
6337 | * @exclude_cpumask: the cpumask to be excluded from wq_unbound_cpumask | |
6338 | * | |
6339 | * This function can be called from cpuset code to provide a set of isolated | |
6340 | * CPUs that should be excluded from wq_unbound_cpumask. The caller must hold | |
6341 | * either cpus_read_lock or cpus_write_lock. | |
6342 | */ | |
6343 | int workqueue_unbound_exclude_cpumask(cpumask_var_t exclude_cpumask) | |
6344 | { | |
6345 | cpumask_var_t cpumask; | |
6346 | int ret = 0; | |
6347 | ||
6348 | if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) | |
6349 | return -ENOMEM; | |
6350 | ||
6351 | lockdep_assert_cpus_held(); | |
6352 | mutex_lock(&wq_pool_mutex); | |
6353 | ||
6354 | /* Save the current isolated cpumask & export it via sysfs */ | |
6355 | cpumask_copy(wq_isolated_cpumask, exclude_cpumask); | |
6356 | ||
6357 | /* | |
6358 | * If the operation fails, it will fall back to | |
6359 | * wq_requested_unbound_cpumask which is initially set to | |
6360 | * (HK_TYPE_WQ ∩ HK_TYPE_DOMAIN) house keeping mask and rewritten | |
6361 | * by any subsequent write to workqueue/cpumask sysfs file. | |
6362 | */ | |
6363 | if (!cpumask_andnot(cpumask, wq_requested_unbound_cpumask, exclude_cpumask)) | |
6364 | cpumask_copy(cpumask, wq_requested_unbound_cpumask); | |
6365 | if (!cpumask_equal(cpumask, wq_unbound_cpumask)) | |
6366 | ret = workqueue_apply_unbound_cpumask(cpumask); | |
6367 | ||
6368 | mutex_unlock(&wq_pool_mutex); | |
6369 | free_cpumask_var(cpumask); | |
6370 | return ret; | |
6371 | } | |
6372 | ||
63c5484e TH |
6373 | static int parse_affn_scope(const char *val) |
6374 | { | |
6375 | int i; | |
6376 | ||
6377 | for (i = 0; i < ARRAY_SIZE(wq_affn_names); i++) { | |
6378 | if (!strncasecmp(val, wq_affn_names[i], strlen(wq_affn_names[i]))) | |
6379 | return i; | |
6380 | } | |
6381 | return -EINVAL; | |
6382 | } | |
6383 | ||
6384 | static int wq_affn_dfl_set(const char *val, const struct kernel_param *kp) | |
6385 | { | |
523a301e TH |
6386 | struct workqueue_struct *wq; |
6387 | int affn, cpu; | |
63c5484e TH |
6388 | |
6389 | affn = parse_affn_scope(val); | |
6390 | if (affn < 0) | |
6391 | return affn; | |
523a301e TH |
6392 | if (affn == WQ_AFFN_DFL) |
6393 | return -EINVAL; | |
6394 | ||
6395 | cpus_read_lock(); | |
6396 | mutex_lock(&wq_pool_mutex); | |
63c5484e TH |
6397 | |
6398 | wq_affn_dfl = affn; | |
523a301e TH |
6399 | |
6400 | list_for_each_entry(wq, &workqueues, list) { | |
6401 | for_each_online_cpu(cpu) { | |
6402 | wq_update_pod(wq, cpu, cpu, true); | |
6403 | } | |
6404 | } | |
6405 | ||
6406 | mutex_unlock(&wq_pool_mutex); | |
6407 | cpus_read_unlock(); | |
6408 | ||
63c5484e TH |
6409 | return 0; |
6410 | } | |
6411 | ||
6412 | static int wq_affn_dfl_get(char *buffer, const struct kernel_param *kp) | |
6413 | { | |
6414 | return scnprintf(buffer, PAGE_SIZE, "%s\n", wq_affn_names[wq_affn_dfl]); | |
6415 | } | |
6416 | ||
6417 | static const struct kernel_param_ops wq_affn_dfl_ops = { | |
6418 | .set = wq_affn_dfl_set, | |
6419 | .get = wq_affn_dfl_get, | |
6420 | }; | |
6421 | ||
6422 | module_param_cb(default_affinity_scope, &wq_affn_dfl_ops, NULL, 0644); | |
6423 | ||
6ba94429 FW |
6424 | #ifdef CONFIG_SYSFS |
6425 | /* | |
6426 | * Workqueues with WQ_SYSFS flag set is visible to userland via | |
6427 | * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the | |
6428 | * following attributes. | |
6429 | * | |
63c5484e TH |
6430 | * per_cpu RO bool : whether the workqueue is per-cpu or unbound |
6431 | * max_active RW int : maximum number of in-flight work items | |
6ba94429 FW |
6432 | * |
6433 | * Unbound workqueues have the following extra attributes. | |
6434 | * | |
63c5484e TH |
6435 | * nice RW int : nice value of the workers |
6436 | * cpumask RW mask : bitmask of allowed CPUs for the workers | |
6437 | * affinity_scope RW str : worker CPU affinity scope (cache, numa, none) | |
8639eceb | 6438 | * affinity_strict RW bool : worker CPU affinity is strict |
6ba94429 FW |
6439 | */ |
6440 | struct wq_device { | |
6441 | struct workqueue_struct *wq; | |
6442 | struct device dev; | |
6443 | }; | |
6444 | ||
6445 | static struct workqueue_struct *dev_to_wq(struct device *dev) | |
6446 | { | |
6447 | struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); | |
6448 | ||
6449 | return wq_dev->wq; | |
6450 | } | |
6451 | ||
6452 | static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, | |
6453 | char *buf) | |
6454 | { | |
6455 | struct workqueue_struct *wq = dev_to_wq(dev); | |
6456 | ||
6457 | return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); | |
6458 | } | |
6459 | static DEVICE_ATTR_RO(per_cpu); | |
6460 | ||
6461 | static ssize_t max_active_show(struct device *dev, | |
6462 | struct device_attribute *attr, char *buf) | |
6463 | { | |
6464 | struct workqueue_struct *wq = dev_to_wq(dev); | |
6465 | ||
6466 | return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); | |
6467 | } | |
6468 | ||
6469 | static ssize_t max_active_store(struct device *dev, | |
6470 | struct device_attribute *attr, const char *buf, | |
6471 | size_t count) | |
6472 | { | |
6473 | struct workqueue_struct *wq = dev_to_wq(dev); | |
6474 | int val; | |
6475 | ||
6476 | if (sscanf(buf, "%d", &val) != 1 || val <= 0) | |
6477 | return -EINVAL; | |
6478 | ||
6479 | workqueue_set_max_active(wq, val); | |
6480 | return count; | |
6481 | } | |
6482 | static DEVICE_ATTR_RW(max_active); | |
6483 | ||
6484 | static struct attribute *wq_sysfs_attrs[] = { | |
6485 | &dev_attr_per_cpu.attr, | |
6486 | &dev_attr_max_active.attr, | |
6487 | NULL, | |
6488 | }; | |
6489 | ATTRIBUTE_GROUPS(wq_sysfs); | |
6490 | ||
49277a5b WL |
6491 | static void apply_wqattrs_lock(void) |
6492 | { | |
6493 | /* CPUs should stay stable across pwq creations and installations */ | |
6494 | cpus_read_lock(); | |
6495 | mutex_lock(&wq_pool_mutex); | |
6496 | } | |
6497 | ||
6498 | static void apply_wqattrs_unlock(void) | |
6499 | { | |
6500 | mutex_unlock(&wq_pool_mutex); | |
6501 | cpus_read_unlock(); | |
6502 | } | |
6503 | ||
6ba94429 FW |
6504 | static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, |
6505 | char *buf) | |
6506 | { | |
6507 | struct workqueue_struct *wq = dev_to_wq(dev); | |
6508 | int written; | |
6509 | ||
6510 | mutex_lock(&wq->mutex); | |
6511 | written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); | |
6512 | mutex_unlock(&wq->mutex); | |
6513 | ||
6514 | return written; | |
6515 | } | |
6516 | ||
6517 | /* prepare workqueue_attrs for sysfs store operations */ | |
6518 | static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) | |
6519 | { | |
6520 | struct workqueue_attrs *attrs; | |
6521 | ||
899a94fe LJ |
6522 | lockdep_assert_held(&wq_pool_mutex); |
6523 | ||
be69d00d | 6524 | attrs = alloc_workqueue_attrs(); |
6ba94429 FW |
6525 | if (!attrs) |
6526 | return NULL; | |
6527 | ||
6ba94429 | 6528 | copy_workqueue_attrs(attrs, wq->unbound_attrs); |
6ba94429 FW |
6529 | return attrs; |
6530 | } | |
6531 | ||
6532 | static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, | |
6533 | const char *buf, size_t count) | |
6534 | { | |
6535 | struct workqueue_struct *wq = dev_to_wq(dev); | |
6536 | struct workqueue_attrs *attrs; | |
d4d3e257 LJ |
6537 | int ret = -ENOMEM; |
6538 | ||
6539 | apply_wqattrs_lock(); | |
6ba94429 FW |
6540 | |
6541 | attrs = wq_sysfs_prep_attrs(wq); | |
6542 | if (!attrs) | |
d4d3e257 | 6543 | goto out_unlock; |
6ba94429 FW |
6544 | |
6545 | if (sscanf(buf, "%d", &attrs->nice) == 1 && | |
6546 | attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) | |
d4d3e257 | 6547 | ret = apply_workqueue_attrs_locked(wq, attrs); |
6ba94429 FW |
6548 | else |
6549 | ret = -EINVAL; | |
6550 | ||
d4d3e257 LJ |
6551 | out_unlock: |
6552 | apply_wqattrs_unlock(); | |
6ba94429 FW |
6553 | free_workqueue_attrs(attrs); |
6554 | return ret ?: count; | |
6555 | } | |
6556 | ||
6557 | static ssize_t wq_cpumask_show(struct device *dev, | |
6558 | struct device_attribute *attr, char *buf) | |
6559 | { | |
6560 | struct workqueue_struct *wq = dev_to_wq(dev); | |
6561 | int written; | |
6562 | ||
6563 | mutex_lock(&wq->mutex); | |
6564 | written = scnprintf(buf, PAGE_SIZE, "%*pb\n", | |
6565 | cpumask_pr_args(wq->unbound_attrs->cpumask)); | |
6566 | mutex_unlock(&wq->mutex); | |
6567 | return written; | |
6568 | } | |
6569 | ||
6570 | static ssize_t wq_cpumask_store(struct device *dev, | |
6571 | struct device_attribute *attr, | |
6572 | const char *buf, size_t count) | |
6573 | { | |
6574 | struct workqueue_struct *wq = dev_to_wq(dev); | |
6575 | struct workqueue_attrs *attrs; | |
d4d3e257 LJ |
6576 | int ret = -ENOMEM; |
6577 | ||
6578 | apply_wqattrs_lock(); | |
6ba94429 FW |
6579 | |
6580 | attrs = wq_sysfs_prep_attrs(wq); | |
6581 | if (!attrs) | |
d4d3e257 | 6582 | goto out_unlock; |
6ba94429 FW |
6583 | |
6584 | ret = cpumask_parse(buf, attrs->cpumask); | |
6585 | if (!ret) | |
d4d3e257 | 6586 | ret = apply_workqueue_attrs_locked(wq, attrs); |
6ba94429 | 6587 | |
d4d3e257 LJ |
6588 | out_unlock: |
6589 | apply_wqattrs_unlock(); | |
6ba94429 FW |
6590 | free_workqueue_attrs(attrs); |
6591 | return ret ?: count; | |
6592 | } | |
6593 | ||
63c5484e TH |
6594 | static ssize_t wq_affn_scope_show(struct device *dev, |
6595 | struct device_attribute *attr, char *buf) | |
6596 | { | |
6597 | struct workqueue_struct *wq = dev_to_wq(dev); | |
6598 | int written; | |
6599 | ||
6600 | mutex_lock(&wq->mutex); | |
523a301e TH |
6601 | if (wq->unbound_attrs->affn_scope == WQ_AFFN_DFL) |
6602 | written = scnprintf(buf, PAGE_SIZE, "%s (%s)\n", | |
6603 | wq_affn_names[WQ_AFFN_DFL], | |
6604 | wq_affn_names[wq_affn_dfl]); | |
6605 | else | |
6606 | written = scnprintf(buf, PAGE_SIZE, "%s\n", | |
6607 | wq_affn_names[wq->unbound_attrs->affn_scope]); | |
63c5484e TH |
6608 | mutex_unlock(&wq->mutex); |
6609 | ||
6610 | return written; | |
6611 | } | |
6612 | ||
6613 | static ssize_t wq_affn_scope_store(struct device *dev, | |
6614 | struct device_attribute *attr, | |
6615 | const char *buf, size_t count) | |
6616 | { | |
6617 | struct workqueue_struct *wq = dev_to_wq(dev); | |
6618 | struct workqueue_attrs *attrs; | |
6619 | int affn, ret = -ENOMEM; | |
6620 | ||
6621 | affn = parse_affn_scope(buf); | |
6622 | if (affn < 0) | |
6623 | return affn; | |
6624 | ||
6625 | apply_wqattrs_lock(); | |
6626 | attrs = wq_sysfs_prep_attrs(wq); | |
6627 | if (attrs) { | |
6628 | attrs->affn_scope = affn; | |
6629 | ret = apply_workqueue_attrs_locked(wq, attrs); | |
6630 | } | |
6631 | apply_wqattrs_unlock(); | |
6632 | free_workqueue_attrs(attrs); | |
6633 | return ret ?: count; | |
6634 | } | |
6635 | ||
8639eceb TH |
6636 | static ssize_t wq_affinity_strict_show(struct device *dev, |
6637 | struct device_attribute *attr, char *buf) | |
6638 | { | |
6639 | struct workqueue_struct *wq = dev_to_wq(dev); | |
6640 | ||
6641 | return scnprintf(buf, PAGE_SIZE, "%d\n", | |
6642 | wq->unbound_attrs->affn_strict); | |
6643 | } | |
6644 | ||
6645 | static ssize_t wq_affinity_strict_store(struct device *dev, | |
6646 | struct device_attribute *attr, | |
6647 | const char *buf, size_t count) | |
6648 | { | |
6649 | struct workqueue_struct *wq = dev_to_wq(dev); | |
6650 | struct workqueue_attrs *attrs; | |
6651 | int v, ret = -ENOMEM; | |
6652 | ||
6653 | if (sscanf(buf, "%d", &v) != 1) | |
6654 | return -EINVAL; | |
6655 | ||
6656 | apply_wqattrs_lock(); | |
6657 | attrs = wq_sysfs_prep_attrs(wq); | |
6658 | if (attrs) { | |
6659 | attrs->affn_strict = (bool)v; | |
6660 | ret = apply_workqueue_attrs_locked(wq, attrs); | |
6661 | } | |
6662 | apply_wqattrs_unlock(); | |
6663 | free_workqueue_attrs(attrs); | |
6664 | return ret ?: count; | |
6665 | } | |
6666 | ||
6ba94429 | 6667 | static struct device_attribute wq_sysfs_unbound_attrs[] = { |
6ba94429 FW |
6668 | __ATTR(nice, 0644, wq_nice_show, wq_nice_store), |
6669 | __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), | |
63c5484e | 6670 | __ATTR(affinity_scope, 0644, wq_affn_scope_show, wq_affn_scope_store), |
8639eceb | 6671 | __ATTR(affinity_strict, 0644, wq_affinity_strict_show, wq_affinity_strict_store), |
6ba94429 FW |
6672 | __ATTR_NULL, |
6673 | }; | |
8ccad40d | 6674 | |
6ba94429 FW |
6675 | static struct bus_type wq_subsys = { |
6676 | .name = "workqueue", | |
6677 | .dev_groups = wq_sysfs_groups, | |
2d3854a3 RR |
6678 | }; |
6679 | ||
49277a5b WL |
6680 | /** |
6681 | * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask | |
6682 | * @cpumask: the cpumask to set | |
6683 | * | |
6684 | * The low-level workqueues cpumask is a global cpumask that limits | |
6685 | * the affinity of all unbound workqueues. This function check the @cpumask | |
6686 | * and apply it to all unbound workqueues and updates all pwqs of them. | |
6687 | * | |
6688 | * Return: 0 - Success | |
6689 | * -EINVAL - Invalid @cpumask | |
6690 | * -ENOMEM - Failed to allocate memory for attrs or pwqs. | |
6691 | */ | |
6692 | static int workqueue_set_unbound_cpumask(cpumask_var_t cpumask) | |
6693 | { | |
6694 | int ret = -EINVAL; | |
6695 | ||
6696 | /* | |
6697 | * Not excluding isolated cpus on purpose. | |
6698 | * If the user wishes to include them, we allow that. | |
6699 | */ | |
6700 | cpumask_and(cpumask, cpumask, cpu_possible_mask); | |
6701 | if (!cpumask_empty(cpumask)) { | |
6702 | apply_wqattrs_lock(); | |
6703 | cpumask_copy(wq_requested_unbound_cpumask, cpumask); | |
6704 | if (cpumask_equal(cpumask, wq_unbound_cpumask)) { | |
6705 | ret = 0; | |
6706 | goto out_unlock; | |
6707 | } | |
6708 | ||
6709 | ret = workqueue_apply_unbound_cpumask(cpumask); | |
6710 | ||
6711 | out_unlock: | |
6712 | apply_wqattrs_unlock(); | |
6713 | } | |
6714 | ||
6715 | return ret; | |
6716 | } | |
6717 | ||
fe28f631 WL |
6718 | static ssize_t __wq_cpumask_show(struct device *dev, |
6719 | struct device_attribute *attr, char *buf, cpumask_var_t mask) | |
b05a7928 FW |
6720 | { |
6721 | int written; | |
6722 | ||
042f7df1 | 6723 | mutex_lock(&wq_pool_mutex); |
fe28f631 | 6724 | written = scnprintf(buf, PAGE_SIZE, "%*pb\n", cpumask_pr_args(mask)); |
042f7df1 | 6725 | mutex_unlock(&wq_pool_mutex); |
b05a7928 FW |
6726 | |
6727 | return written; | |
6728 | } | |
6729 | ||
fe28f631 WL |
6730 | static ssize_t wq_unbound_cpumask_show(struct device *dev, |
6731 | struct device_attribute *attr, char *buf) | |
6732 | { | |
6733 | return __wq_cpumask_show(dev, attr, buf, wq_unbound_cpumask); | |
6734 | } | |
6735 | ||
6736 | static ssize_t wq_requested_cpumask_show(struct device *dev, | |
6737 | struct device_attribute *attr, char *buf) | |
6738 | { | |
6739 | return __wq_cpumask_show(dev, attr, buf, wq_requested_unbound_cpumask); | |
6740 | } | |
6741 | ||
6742 | static ssize_t wq_isolated_cpumask_show(struct device *dev, | |
6743 | struct device_attribute *attr, char *buf) | |
6744 | { | |
6745 | return __wq_cpumask_show(dev, attr, buf, wq_isolated_cpumask); | |
6746 | } | |
6747 | ||
042f7df1 LJ |
6748 | static ssize_t wq_unbound_cpumask_store(struct device *dev, |
6749 | struct device_attribute *attr, const char *buf, size_t count) | |
6750 | { | |
6751 | cpumask_var_t cpumask; | |
6752 | int ret; | |
6753 | ||
6754 | if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) | |
6755 | return -ENOMEM; | |
6756 | ||
6757 | ret = cpumask_parse(buf, cpumask); | |
6758 | if (!ret) | |
6759 | ret = workqueue_set_unbound_cpumask(cpumask); | |
6760 | ||
6761 | free_cpumask_var(cpumask); | |
6762 | return ret ? ret : count; | |
6763 | } | |
6764 | ||
fe28f631 | 6765 | static struct device_attribute wq_sysfs_cpumask_attrs[] = { |
042f7df1 | 6766 | __ATTR(cpumask, 0644, wq_unbound_cpumask_show, |
fe28f631 WL |
6767 | wq_unbound_cpumask_store), |
6768 | __ATTR(cpumask_requested, 0444, wq_requested_cpumask_show, NULL), | |
6769 | __ATTR(cpumask_isolated, 0444, wq_isolated_cpumask_show, NULL), | |
6770 | __ATTR_NULL, | |
6771 | }; | |
b05a7928 | 6772 | |
6ba94429 | 6773 | static int __init wq_sysfs_init(void) |
2d3854a3 | 6774 | { |
686f6697 | 6775 | struct device *dev_root; |
b05a7928 FW |
6776 | int err; |
6777 | ||
6778 | err = subsys_virtual_register(&wq_subsys, NULL); | |
6779 | if (err) | |
6780 | return err; | |
6781 | ||
686f6697 GKH |
6782 | dev_root = bus_get_dev_root(&wq_subsys); |
6783 | if (dev_root) { | |
fe28f631 WL |
6784 | struct device_attribute *attr; |
6785 | ||
6786 | for (attr = wq_sysfs_cpumask_attrs; attr->attr.name; attr++) { | |
6787 | err = device_create_file(dev_root, attr); | |
6788 | if (err) | |
6789 | break; | |
6790 | } | |
686f6697 GKH |
6791 | put_device(dev_root); |
6792 | } | |
6793 | return err; | |
2d3854a3 | 6794 | } |
6ba94429 | 6795 | core_initcall(wq_sysfs_init); |
2d3854a3 | 6796 | |
6ba94429 | 6797 | static void wq_device_release(struct device *dev) |
2d3854a3 | 6798 | { |
6ba94429 | 6799 | struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); |
6b44003e | 6800 | |
6ba94429 | 6801 | kfree(wq_dev); |
2d3854a3 | 6802 | } |
a0a1a5fd TH |
6803 | |
6804 | /** | |
6ba94429 FW |
6805 | * workqueue_sysfs_register - make a workqueue visible in sysfs |
6806 | * @wq: the workqueue to register | |
a0a1a5fd | 6807 | * |
6ba94429 FW |
6808 | * Expose @wq in sysfs under /sys/bus/workqueue/devices. |
6809 | * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set | |
6810 | * which is the preferred method. | |
a0a1a5fd | 6811 | * |
6ba94429 FW |
6812 | * Workqueue user should use this function directly iff it wants to apply |
6813 | * workqueue_attrs before making the workqueue visible in sysfs; otherwise, | |
6814 | * apply_workqueue_attrs() may race against userland updating the | |
6815 | * attributes. | |
6816 | * | |
6817 | * Return: 0 on success, -errno on failure. | |
a0a1a5fd | 6818 | */ |
6ba94429 | 6819 | int workqueue_sysfs_register(struct workqueue_struct *wq) |
a0a1a5fd | 6820 | { |
6ba94429 FW |
6821 | struct wq_device *wq_dev; |
6822 | int ret; | |
a0a1a5fd | 6823 | |
6ba94429 | 6824 | /* |
402dd89d | 6825 | * Adjusting max_active or creating new pwqs by applying |
6ba94429 FW |
6826 | * attributes breaks ordering guarantee. Disallow exposing ordered |
6827 | * workqueues. | |
6828 | */ | |
0a94efb5 | 6829 | if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) |
6ba94429 | 6830 | return -EINVAL; |
a0a1a5fd | 6831 | |
6ba94429 FW |
6832 | wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); |
6833 | if (!wq_dev) | |
6834 | return -ENOMEM; | |
5bcab335 | 6835 | |
6ba94429 FW |
6836 | wq_dev->wq = wq; |
6837 | wq_dev->dev.bus = &wq_subsys; | |
6ba94429 | 6838 | wq_dev->dev.release = wq_device_release; |
23217b44 | 6839 | dev_set_name(&wq_dev->dev, "%s", wq->name); |
a0a1a5fd | 6840 | |
6ba94429 FW |
6841 | /* |
6842 | * unbound_attrs are created separately. Suppress uevent until | |
6843 | * everything is ready. | |
6844 | */ | |
6845 | dev_set_uevent_suppress(&wq_dev->dev, true); | |
a0a1a5fd | 6846 | |
6ba94429 FW |
6847 | ret = device_register(&wq_dev->dev); |
6848 | if (ret) { | |
537f4146 | 6849 | put_device(&wq_dev->dev); |
6ba94429 FW |
6850 | wq->wq_dev = NULL; |
6851 | return ret; | |
6852 | } | |
a0a1a5fd | 6853 | |
6ba94429 FW |
6854 | if (wq->flags & WQ_UNBOUND) { |
6855 | struct device_attribute *attr; | |
a0a1a5fd | 6856 | |
6ba94429 FW |
6857 | for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { |
6858 | ret = device_create_file(&wq_dev->dev, attr); | |
6859 | if (ret) { | |
6860 | device_unregister(&wq_dev->dev); | |
6861 | wq->wq_dev = NULL; | |
6862 | return ret; | |
a0a1a5fd TH |
6863 | } |
6864 | } | |
6865 | } | |
6ba94429 FW |
6866 | |
6867 | dev_set_uevent_suppress(&wq_dev->dev, false); | |
6868 | kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); | |
6869 | return 0; | |
a0a1a5fd TH |
6870 | } |
6871 | ||
6872 | /** | |
6ba94429 FW |
6873 | * workqueue_sysfs_unregister - undo workqueue_sysfs_register() |
6874 | * @wq: the workqueue to unregister | |
a0a1a5fd | 6875 | * |
6ba94429 | 6876 | * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. |
a0a1a5fd | 6877 | */ |
6ba94429 | 6878 | static void workqueue_sysfs_unregister(struct workqueue_struct *wq) |
a0a1a5fd | 6879 | { |
6ba94429 | 6880 | struct wq_device *wq_dev = wq->wq_dev; |
8b03ae3c | 6881 | |
6ba94429 FW |
6882 | if (!wq->wq_dev) |
6883 | return; | |
a0a1a5fd | 6884 | |
6ba94429 FW |
6885 | wq->wq_dev = NULL; |
6886 | device_unregister(&wq_dev->dev); | |
a0a1a5fd | 6887 | } |
6ba94429 FW |
6888 | #else /* CONFIG_SYSFS */ |
6889 | static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } | |
6890 | #endif /* CONFIG_SYSFS */ | |
a0a1a5fd | 6891 | |
82607adc TH |
6892 | /* |
6893 | * Workqueue watchdog. | |
6894 | * | |
6895 | * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal | |
6896 | * flush dependency, a concurrency managed work item which stays RUNNING | |
6897 | * indefinitely. Workqueue stalls can be very difficult to debug as the | |
6898 | * usual warning mechanisms don't trigger and internal workqueue state is | |
6899 | * largely opaque. | |
6900 | * | |
6901 | * Workqueue watchdog monitors all worker pools periodically and dumps | |
6902 | * state if some pools failed to make forward progress for a while where | |
6903 | * forward progress is defined as the first item on ->worklist changing. | |
6904 | * | |
6905 | * This mechanism is controlled through the kernel parameter | |
6906 | * "workqueue.watchdog_thresh" which can be updated at runtime through the | |
6907 | * corresponding sysfs parameter file. | |
6908 | */ | |
6909 | #ifdef CONFIG_WQ_WATCHDOG | |
6910 | ||
82607adc | 6911 | static unsigned long wq_watchdog_thresh = 30; |
5cd79d6a | 6912 | static struct timer_list wq_watchdog_timer; |
82607adc TH |
6913 | |
6914 | static unsigned long wq_watchdog_touched = INITIAL_JIFFIES; | |
6915 | static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES; | |
6916 | ||
cd2440d6 PM |
6917 | /* |
6918 | * Show workers that might prevent the processing of pending work items. | |
6919 | * The only candidates are CPU-bound workers in the running state. | |
6920 | * Pending work items should be handled by another idle worker | |
6921 | * in all other situations. | |
6922 | */ | |
6923 | static void show_cpu_pool_hog(struct worker_pool *pool) | |
6924 | { | |
6925 | struct worker *worker; | |
6926 | unsigned long flags; | |
6927 | int bkt; | |
6928 | ||
6929 | raw_spin_lock_irqsave(&pool->lock, flags); | |
6930 | ||
6931 | hash_for_each(pool->busy_hash, bkt, worker, hentry) { | |
6932 | if (task_is_running(worker->task)) { | |
6933 | /* | |
6934 | * Defer printing to avoid deadlocks in console | |
6935 | * drivers that queue work while holding locks | |
6936 | * also taken in their write paths. | |
6937 | */ | |
6938 | printk_deferred_enter(); | |
6939 | ||
6940 | pr_info("pool %d:\n", pool->id); | |
6941 | sched_show_task(worker->task); | |
6942 | ||
6943 | printk_deferred_exit(); | |
6944 | } | |
6945 | } | |
6946 | ||
6947 | raw_spin_unlock_irqrestore(&pool->lock, flags); | |
6948 | } | |
6949 | ||
6950 | static void show_cpu_pools_hogs(void) | |
6951 | { | |
6952 | struct worker_pool *pool; | |
6953 | int pi; | |
6954 | ||
6955 | pr_info("Showing backtraces of running workers in stalled CPU-bound worker pools:\n"); | |
6956 | ||
6957 | rcu_read_lock(); | |
6958 | ||
6959 | for_each_pool(pool, pi) { | |
6960 | if (pool->cpu_stall) | |
6961 | show_cpu_pool_hog(pool); | |
6962 | ||
6963 | } | |
6964 | ||
6965 | rcu_read_unlock(); | |
6966 | } | |
6967 | ||
82607adc TH |
6968 | static void wq_watchdog_reset_touched(void) |
6969 | { | |
6970 | int cpu; | |
6971 | ||
6972 | wq_watchdog_touched = jiffies; | |
6973 | for_each_possible_cpu(cpu) | |
6974 | per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; | |
6975 | } | |
6976 | ||
5cd79d6a | 6977 | static void wq_watchdog_timer_fn(struct timer_list *unused) |
82607adc TH |
6978 | { |
6979 | unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ; | |
6980 | bool lockup_detected = false; | |
cd2440d6 | 6981 | bool cpu_pool_stall = false; |
940d71c6 | 6982 | unsigned long now = jiffies; |
82607adc TH |
6983 | struct worker_pool *pool; |
6984 | int pi; | |
6985 | ||
6986 | if (!thresh) | |
6987 | return; | |
6988 | ||
6989 | rcu_read_lock(); | |
6990 | ||
6991 | for_each_pool(pool, pi) { | |
6992 | unsigned long pool_ts, touched, ts; | |
6993 | ||
cd2440d6 | 6994 | pool->cpu_stall = false; |
82607adc TH |
6995 | if (list_empty(&pool->worklist)) |
6996 | continue; | |
6997 | ||
940d71c6 SS |
6998 | /* |
6999 | * If a virtual machine is stopped by the host it can look to | |
7000 | * the watchdog like a stall. | |
7001 | */ | |
7002 | kvm_check_and_clear_guest_paused(); | |
7003 | ||
82607adc | 7004 | /* get the latest of pool and touched timestamps */ |
89e28ce6 WQ |
7005 | if (pool->cpu >= 0) |
7006 | touched = READ_ONCE(per_cpu(wq_watchdog_touched_cpu, pool->cpu)); | |
7007 | else | |
7008 | touched = READ_ONCE(wq_watchdog_touched); | |
82607adc | 7009 | pool_ts = READ_ONCE(pool->watchdog_ts); |
82607adc TH |
7010 | |
7011 | if (time_after(pool_ts, touched)) | |
7012 | ts = pool_ts; | |
7013 | else | |
7014 | ts = touched; | |
7015 | ||
82607adc | 7016 | /* did we stall? */ |
940d71c6 | 7017 | if (time_after(now, ts + thresh)) { |
82607adc | 7018 | lockup_detected = true; |
cd2440d6 PM |
7019 | if (pool->cpu >= 0) { |
7020 | pool->cpu_stall = true; | |
7021 | cpu_pool_stall = true; | |
7022 | } | |
82607adc TH |
7023 | pr_emerg("BUG: workqueue lockup - pool"); |
7024 | pr_cont_pool_info(pool); | |
7025 | pr_cont(" stuck for %us!\n", | |
940d71c6 | 7026 | jiffies_to_msecs(now - pool_ts) / 1000); |
82607adc | 7027 | } |
cd2440d6 PM |
7028 | |
7029 | ||
82607adc TH |
7030 | } |
7031 | ||
7032 | rcu_read_unlock(); | |
7033 | ||
7034 | if (lockup_detected) | |
55df0933 | 7035 | show_all_workqueues(); |
82607adc | 7036 | |
cd2440d6 PM |
7037 | if (cpu_pool_stall) |
7038 | show_cpu_pools_hogs(); | |
7039 | ||
82607adc TH |
7040 | wq_watchdog_reset_touched(); |
7041 | mod_timer(&wq_watchdog_timer, jiffies + thresh); | |
7042 | } | |
7043 | ||
cb9d7fd5 | 7044 | notrace void wq_watchdog_touch(int cpu) |
82607adc TH |
7045 | { |
7046 | if (cpu >= 0) | |
7047 | per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; | |
89e28ce6 WQ |
7048 | |
7049 | wq_watchdog_touched = jiffies; | |
82607adc TH |
7050 | } |
7051 | ||
7052 | static void wq_watchdog_set_thresh(unsigned long thresh) | |
7053 | { | |
7054 | wq_watchdog_thresh = 0; | |
7055 | del_timer_sync(&wq_watchdog_timer); | |
7056 | ||
7057 | if (thresh) { | |
7058 | wq_watchdog_thresh = thresh; | |
7059 | wq_watchdog_reset_touched(); | |
7060 | mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ); | |
7061 | } | |
7062 | } | |
7063 | ||
7064 | static int wq_watchdog_param_set_thresh(const char *val, | |
7065 | const struct kernel_param *kp) | |
7066 | { | |
7067 | unsigned long thresh; | |
7068 | int ret; | |
7069 | ||
7070 | ret = kstrtoul(val, 0, &thresh); | |
7071 | if (ret) | |
7072 | return ret; | |
7073 | ||
7074 | if (system_wq) | |
7075 | wq_watchdog_set_thresh(thresh); | |
7076 | else | |
7077 | wq_watchdog_thresh = thresh; | |
7078 | ||
7079 | return 0; | |
7080 | } | |
7081 | ||
7082 | static const struct kernel_param_ops wq_watchdog_thresh_ops = { | |
7083 | .set = wq_watchdog_param_set_thresh, | |
7084 | .get = param_get_ulong, | |
7085 | }; | |
7086 | ||
7087 | module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh, | |
7088 | 0644); | |
7089 | ||
7090 | static void wq_watchdog_init(void) | |
7091 | { | |
5cd79d6a | 7092 | timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE); |
82607adc TH |
7093 | wq_watchdog_set_thresh(wq_watchdog_thresh); |
7094 | } | |
7095 | ||
7096 | #else /* CONFIG_WQ_WATCHDOG */ | |
7097 | ||
7098 | static inline void wq_watchdog_init(void) { } | |
7099 | ||
7100 | #endif /* CONFIG_WQ_WATCHDOG */ | |
7101 | ||
4a6c5607 TH |
7102 | static void __init restrict_unbound_cpumask(const char *name, const struct cpumask *mask) |
7103 | { | |
7104 | if (!cpumask_intersects(wq_unbound_cpumask, mask)) { | |
7105 | pr_warn("workqueue: Restricting unbound_cpumask (%*pb) with %s (%*pb) leaves no CPU, ignoring\n", | |
7106 | cpumask_pr_args(wq_unbound_cpumask), name, cpumask_pr_args(mask)); | |
7107 | return; | |
7108 | } | |
7109 | ||
7110 | cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, mask); | |
7111 | } | |
7112 | ||
3347fa09 TH |
7113 | /** |
7114 | * workqueue_init_early - early init for workqueue subsystem | |
7115 | * | |
2930155b TH |
7116 | * This is the first step of three-staged workqueue subsystem initialization and |
7117 | * invoked as soon as the bare basics - memory allocation, cpumasks and idr are | |
7118 | * up. It sets up all the data structures and system workqueues and allows early | |
7119 | * boot code to create workqueues and queue/cancel work items. Actual work item | |
7120 | * execution starts only after kthreads can be created and scheduled right | |
7121 | * before early initcalls. | |
3347fa09 | 7122 | */ |
2333e829 | 7123 | void __init workqueue_init_early(void) |
1da177e4 | 7124 | { |
84193c07 | 7125 | struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_SYSTEM]; |
7a4e344c TH |
7126 | int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; |
7127 | int i, cpu; | |
c34056a3 | 7128 | |
10cdb157 | 7129 | BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); |
e904e6c2 | 7130 | |
b05a7928 | 7131 | BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL)); |
fe28f631 WL |
7132 | BUG_ON(!alloc_cpumask_var(&wq_requested_unbound_cpumask, GFP_KERNEL)); |
7133 | BUG_ON(!zalloc_cpumask_var(&wq_isolated_cpumask, GFP_KERNEL)); | |
b05a7928 | 7134 | |
4a6c5607 TH |
7135 | cpumask_copy(wq_unbound_cpumask, cpu_possible_mask); |
7136 | restrict_unbound_cpumask("HK_TYPE_WQ", housekeeping_cpumask(HK_TYPE_WQ)); | |
7137 | restrict_unbound_cpumask("HK_TYPE_DOMAIN", housekeeping_cpumask(HK_TYPE_DOMAIN)); | |
ace3c549 | 7138 | if (!cpumask_empty(&wq_cmdline_cpumask)) |
4a6c5607 | 7139 | restrict_unbound_cpumask("workqueue.unbound_cpus", &wq_cmdline_cpumask); |
ace3c549 | 7140 | |
fe28f631 | 7141 | cpumask_copy(wq_requested_unbound_cpumask, wq_unbound_cpumask); |
ace3c549 | 7142 | |
e904e6c2 TH |
7143 | pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); |
7144 | ||
2930155b TH |
7145 | wq_update_pod_attrs_buf = alloc_workqueue_attrs(); |
7146 | BUG_ON(!wq_update_pod_attrs_buf); | |
7147 | ||
7bd20b6b MT |
7148 | /* |
7149 | * If nohz_full is enabled, set power efficient workqueue as unbound. | |
7150 | * This allows workqueue items to be moved to HK CPUs. | |
7151 | */ | |
7152 | if (housekeeping_enabled(HK_TYPE_TICK)) | |
7153 | wq_power_efficient = true; | |
7154 | ||
84193c07 TH |
7155 | /* initialize WQ_AFFN_SYSTEM pods */ |
7156 | pt->pod_cpus = kcalloc(1, sizeof(pt->pod_cpus[0]), GFP_KERNEL); | |
7157 | pt->pod_node = kcalloc(1, sizeof(pt->pod_node[0]), GFP_KERNEL); | |
7158 | pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL); | |
7159 | BUG_ON(!pt->pod_cpus || !pt->pod_node || !pt->cpu_pod); | |
7160 | ||
7161 | BUG_ON(!zalloc_cpumask_var_node(&pt->pod_cpus[0], GFP_KERNEL, NUMA_NO_NODE)); | |
7162 | ||
84193c07 TH |
7163 | pt->nr_pods = 1; |
7164 | cpumask_copy(pt->pod_cpus[0], cpu_possible_mask); | |
7165 | pt->pod_node[0] = NUMA_NO_NODE; | |
7166 | pt->cpu_pod[0] = 0; | |
7167 | ||
706026c2 | 7168 | /* initialize CPU pools */ |
29c91e99 | 7169 | for_each_possible_cpu(cpu) { |
4ce62e9e | 7170 | struct worker_pool *pool; |
8b03ae3c | 7171 | |
7a4e344c | 7172 | i = 0; |
f02ae73a | 7173 | for_each_cpu_worker_pool(pool, cpu) { |
7a4e344c | 7174 | BUG_ON(init_worker_pool(pool)); |
ec22ca5e | 7175 | pool->cpu = cpu; |
29c91e99 | 7176 | cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); |
9546b29e | 7177 | cpumask_copy(pool->attrs->__pod_cpumask, cpumask_of(cpu)); |
7a4e344c | 7178 | pool->attrs->nice = std_nice[i++]; |
8639eceb | 7179 | pool->attrs->affn_strict = true; |
f3f90ad4 | 7180 | pool->node = cpu_to_node(cpu); |
7a4e344c | 7181 | |
9daf9e67 | 7182 | /* alloc pool ID */ |
68e13a67 | 7183 | mutex_lock(&wq_pool_mutex); |
9daf9e67 | 7184 | BUG_ON(worker_pool_assign_id(pool)); |
68e13a67 | 7185 | mutex_unlock(&wq_pool_mutex); |
4ce62e9e | 7186 | } |
8b03ae3c TH |
7187 | } |
7188 | ||
8a2b7538 | 7189 | /* create default unbound and ordered wq attrs */ |
29c91e99 TH |
7190 | for (i = 0; i < NR_STD_WORKER_POOLS; i++) { |
7191 | struct workqueue_attrs *attrs; | |
7192 | ||
be69d00d | 7193 | BUG_ON(!(attrs = alloc_workqueue_attrs())); |
29c91e99 | 7194 | attrs->nice = std_nice[i]; |
29c91e99 | 7195 | unbound_std_wq_attrs[i] = attrs; |
8a2b7538 TH |
7196 | |
7197 | /* | |
7198 | * An ordered wq should have only one pwq as ordering is | |
7199 | * guaranteed by max_active which is enforced by pwqs. | |
8a2b7538 | 7200 | */ |
be69d00d | 7201 | BUG_ON(!(attrs = alloc_workqueue_attrs())); |
8a2b7538 | 7202 | attrs->nice = std_nice[i]; |
af73f5c9 | 7203 | attrs->ordered = true; |
8a2b7538 | 7204 | ordered_wq_attrs[i] = attrs; |
29c91e99 TH |
7205 | } |
7206 | ||
d320c038 | 7207 | system_wq = alloc_workqueue("events", 0, 0); |
1aabe902 | 7208 | system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); |
d320c038 | 7209 | system_long_wq = alloc_workqueue("events_long", 0, 0); |
f3421797 | 7210 | system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, |
636b927e | 7211 | WQ_MAX_ACTIVE); |
24d51add TH |
7212 | system_freezable_wq = alloc_workqueue("events_freezable", |
7213 | WQ_FREEZABLE, 0); | |
0668106c VK |
7214 | system_power_efficient_wq = alloc_workqueue("events_power_efficient", |
7215 | WQ_POWER_EFFICIENT, 0); | |
8318d6a6 | 7216 | system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_pwr_efficient", |
0668106c VK |
7217 | WQ_FREEZABLE | WQ_POWER_EFFICIENT, |
7218 | 0); | |
1aabe902 | 7219 | BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || |
0668106c VK |
7220 | !system_unbound_wq || !system_freezable_wq || |
7221 | !system_power_efficient_wq || | |
7222 | !system_freezable_power_efficient_wq); | |
3347fa09 TH |
7223 | } |
7224 | ||
aa6fde93 TH |
7225 | static void __init wq_cpu_intensive_thresh_init(void) |
7226 | { | |
7227 | unsigned long thresh; | |
7228 | unsigned long bogo; | |
7229 | ||
dd64c873 Z |
7230 | pwq_release_worker = kthread_create_worker(0, "pool_workqueue_release"); |
7231 | BUG_ON(IS_ERR(pwq_release_worker)); | |
7232 | ||
aa6fde93 TH |
7233 | /* if the user set it to a specific value, keep it */ |
7234 | if (wq_cpu_intensive_thresh_us != ULONG_MAX) | |
7235 | return; | |
7236 | ||
7237 | /* | |
7238 | * The default of 10ms is derived from the fact that most modern (as of | |
7239 | * 2023) processors can do a lot in 10ms and that it's just below what | |
7240 | * most consider human-perceivable. However, the kernel also runs on a | |
7241 | * lot slower CPUs including microcontrollers where the threshold is way | |
7242 | * too low. | |
7243 | * | |
7244 | * Let's scale up the threshold upto 1 second if BogoMips is below 4000. | |
7245 | * This is by no means accurate but it doesn't have to be. The mechanism | |
7246 | * is still useful even when the threshold is fully scaled up. Also, as | |
7247 | * the reports would usually be applicable to everyone, some machines | |
7248 | * operating on longer thresholds won't significantly diminish their | |
7249 | * usefulness. | |
7250 | */ | |
7251 | thresh = 10 * USEC_PER_MSEC; | |
7252 | ||
7253 | /* see init/calibrate.c for lpj -> BogoMIPS calculation */ | |
7254 | bogo = max_t(unsigned long, loops_per_jiffy / 500000 * HZ, 1); | |
7255 | if (bogo < 4000) | |
7256 | thresh = min_t(unsigned long, thresh * 4000 / bogo, USEC_PER_SEC); | |
7257 | ||
7258 | pr_debug("wq_cpu_intensive_thresh: lpj=%lu BogoMIPS=%lu thresh_us=%lu\n", | |
7259 | loops_per_jiffy, bogo, thresh); | |
7260 | ||
7261 | wq_cpu_intensive_thresh_us = thresh; | |
7262 | } | |
7263 | ||
3347fa09 TH |
7264 | /** |
7265 | * workqueue_init - bring workqueue subsystem fully online | |
7266 | * | |
2930155b TH |
7267 | * This is the second step of three-staged workqueue subsystem initialization |
7268 | * and invoked as soon as kthreads can be created and scheduled. Workqueues have | |
7269 | * been created and work items queued on them, but there are no kworkers | |
7270 | * executing the work items yet. Populate the worker pools with the initial | |
7271 | * workers and enable future kworker creations. | |
3347fa09 | 7272 | */ |
2333e829 | 7273 | void __init workqueue_init(void) |
3347fa09 | 7274 | { |
2186d9f9 | 7275 | struct workqueue_struct *wq; |
3347fa09 TH |
7276 | struct worker_pool *pool; |
7277 | int cpu, bkt; | |
7278 | ||
aa6fde93 TH |
7279 | wq_cpu_intensive_thresh_init(); |
7280 | ||
2186d9f9 TH |
7281 | mutex_lock(&wq_pool_mutex); |
7282 | ||
2930155b TH |
7283 | /* |
7284 | * Per-cpu pools created earlier could be missing node hint. Fix them | |
7285 | * up. Also, create a rescuer for workqueues that requested it. | |
7286 | */ | |
2186d9f9 TH |
7287 | for_each_possible_cpu(cpu) { |
7288 | for_each_cpu_worker_pool(pool, cpu) { | |
7289 | pool->node = cpu_to_node(cpu); | |
7290 | } | |
7291 | } | |
7292 | ||
40c17f75 | 7293 | list_for_each_entry(wq, &workqueues, list) { |
40c17f75 TH |
7294 | WARN(init_rescuer(wq), |
7295 | "workqueue: failed to create early rescuer for %s", | |
7296 | wq->name); | |
7297 | } | |
2186d9f9 TH |
7298 | |
7299 | mutex_unlock(&wq_pool_mutex); | |
7300 | ||
3347fa09 TH |
7301 | /* create the initial workers */ |
7302 | for_each_online_cpu(cpu) { | |
7303 | for_each_cpu_worker_pool(pool, cpu) { | |
7304 | pool->flags &= ~POOL_DISASSOCIATED; | |
7305 | BUG_ON(!create_worker(pool)); | |
7306 | } | |
7307 | } | |
7308 | ||
7309 | hash_for_each(unbound_pool_hash, bkt, pool, hash_node) | |
7310 | BUG_ON(!create_worker(pool)); | |
7311 | ||
7312 | wq_online = true; | |
82607adc | 7313 | wq_watchdog_init(); |
1da177e4 | 7314 | } |
c4f135d6 | 7315 | |
025e1684 TH |
7316 | /* |
7317 | * Initialize @pt by first initializing @pt->cpu_pod[] with pod IDs according to | |
7318 | * @cpu_shares_pod(). Each subset of CPUs that share a pod is assigned a unique | |
7319 | * and consecutive pod ID. The rest of @pt is initialized accordingly. | |
7320 | */ | |
7321 | static void __init init_pod_type(struct wq_pod_type *pt, | |
7322 | bool (*cpus_share_pod)(int, int)) | |
7323 | { | |
7324 | int cur, pre, cpu, pod; | |
7325 | ||
7326 | pt->nr_pods = 0; | |
7327 | ||
7328 | /* init @pt->cpu_pod[] according to @cpus_share_pod() */ | |
7329 | pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL); | |
7330 | BUG_ON(!pt->cpu_pod); | |
7331 | ||
7332 | for_each_possible_cpu(cur) { | |
7333 | for_each_possible_cpu(pre) { | |
7334 | if (pre >= cur) { | |
7335 | pt->cpu_pod[cur] = pt->nr_pods++; | |
7336 | break; | |
7337 | } | |
7338 | if (cpus_share_pod(cur, pre)) { | |
7339 | pt->cpu_pod[cur] = pt->cpu_pod[pre]; | |
7340 | break; | |
7341 | } | |
7342 | } | |
7343 | } | |
7344 | ||
7345 | /* init the rest to match @pt->cpu_pod[] */ | |
7346 | pt->pod_cpus = kcalloc(pt->nr_pods, sizeof(pt->pod_cpus[0]), GFP_KERNEL); | |
7347 | pt->pod_node = kcalloc(pt->nr_pods, sizeof(pt->pod_node[0]), GFP_KERNEL); | |
7348 | BUG_ON(!pt->pod_cpus || !pt->pod_node); | |
7349 | ||
7350 | for (pod = 0; pod < pt->nr_pods; pod++) | |
7351 | BUG_ON(!zalloc_cpumask_var(&pt->pod_cpus[pod], GFP_KERNEL)); | |
7352 | ||
7353 | for_each_possible_cpu(cpu) { | |
7354 | cpumask_set_cpu(cpu, pt->pod_cpus[pt->cpu_pod[cpu]]); | |
7355 | pt->pod_node[pt->cpu_pod[cpu]] = cpu_to_node(cpu); | |
7356 | } | |
7357 | } | |
7358 | ||
63c5484e TH |
7359 | static bool __init cpus_dont_share(int cpu0, int cpu1) |
7360 | { | |
7361 | return false; | |
7362 | } | |
7363 | ||
7364 | static bool __init cpus_share_smt(int cpu0, int cpu1) | |
7365 | { | |
7366 | #ifdef CONFIG_SCHED_SMT | |
7367 | return cpumask_test_cpu(cpu0, cpu_smt_mask(cpu1)); | |
7368 | #else | |
7369 | return false; | |
7370 | #endif | |
7371 | } | |
7372 | ||
025e1684 TH |
7373 | static bool __init cpus_share_numa(int cpu0, int cpu1) |
7374 | { | |
7375 | return cpu_to_node(cpu0) == cpu_to_node(cpu1); | |
7376 | } | |
7377 | ||
2930155b TH |
7378 | /** |
7379 | * workqueue_init_topology - initialize CPU pods for unbound workqueues | |
7380 | * | |
7381 | * This is the third step of there-staged workqueue subsystem initialization and | |
7382 | * invoked after SMP and topology information are fully initialized. It | |
7383 | * initializes the unbound CPU pods accordingly. | |
7384 | */ | |
7385 | void __init workqueue_init_topology(void) | |
a86feae6 | 7386 | { |
2930155b | 7387 | struct workqueue_struct *wq; |
025e1684 | 7388 | int cpu; |
a86feae6 | 7389 | |
63c5484e TH |
7390 | init_pod_type(&wq_pod_types[WQ_AFFN_CPU], cpus_dont_share); |
7391 | init_pod_type(&wq_pod_types[WQ_AFFN_SMT], cpus_share_smt); | |
7392 | init_pod_type(&wq_pod_types[WQ_AFFN_CACHE], cpus_share_cache); | |
025e1684 | 7393 | init_pod_type(&wq_pod_types[WQ_AFFN_NUMA], cpus_share_numa); |
a86feae6 | 7394 | |
2930155b | 7395 | mutex_lock(&wq_pool_mutex); |
a86feae6 | 7396 | |
2930155b TH |
7397 | /* |
7398 | * Workqueues allocated earlier would have all CPUs sharing the default | |
7399 | * worker pool. Explicitly call wq_update_pod() on all workqueue and CPU | |
7400 | * combinations to apply per-pod sharing. | |
7401 | */ | |
7402 | list_for_each_entry(wq, &workqueues, list) { | |
5797b1c1 | 7403 | for_each_online_cpu(cpu) |
2930155b | 7404 | wq_update_pod(wq, cpu, cpu, true); |
5797b1c1 TH |
7405 | if (wq->flags & WQ_UNBOUND) { |
7406 | mutex_lock(&wq->mutex); | |
7407 | wq_update_node_max_active(wq, -1); | |
7408 | mutex_unlock(&wq->mutex); | |
2930155b TH |
7409 | } |
7410 | } | |
7411 | ||
7412 | mutex_unlock(&wq_pool_mutex); | |
a86feae6 TH |
7413 | } |
7414 | ||
20bdedaf TH |
7415 | void __warn_flushing_systemwide_wq(void) |
7416 | { | |
7417 | pr_warn("WARNING: Flushing system-wide workqueues will be prohibited in near future.\n"); | |
7418 | dump_stack(); | |
7419 | } | |
c4f135d6 | 7420 | EXPORT_SYMBOL(__warn_flushing_systemwide_wq); |
ace3c549 | 7421 | |
7422 | static int __init workqueue_unbound_cpus_setup(char *str) | |
7423 | { | |
7424 | if (cpulist_parse(str, &wq_cmdline_cpumask) < 0) { | |
7425 | cpumask_clear(&wq_cmdline_cpumask); | |
7426 | pr_warn("workqueue.unbound_cpus: incorrect CPU range, using default\n"); | |
7427 | } | |
7428 | ||
7429 | return 1; | |
7430 | } | |
7431 | __setup("workqueue.unbound_cpus=", workqueue_unbound_cpus_setup); |