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 | |
c5f8cd6c TH |
389 | static bool wq_topo_initialized __read_mostly = false; |
390 | ||
616db877 TH |
391 | /* |
392 | * Per-cpu work items which run for longer than the following threshold are | |
393 | * automatically considered CPU intensive and excluded from concurrency | |
394 | * management to prevent them from noticeably delaying other per-cpu work items. | |
aa6fde93 TH |
395 | * ULONG_MAX indicates that the user hasn't overridden it with a boot parameter. |
396 | * The actual value is initialized in wq_cpu_intensive_thresh_init(). | |
616db877 | 397 | */ |
aa6fde93 | 398 | static unsigned long wq_cpu_intensive_thresh_us = ULONG_MAX; |
616db877 TH |
399 | module_param_named(cpu_intensive_thresh_us, wq_cpu_intensive_thresh_us, ulong, 0644); |
400 | ||
cee22a15 | 401 | /* see the comment above the definition of WQ_POWER_EFFICIENT */ |
552f530c | 402 | static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT); |
cee22a15 VK |
403 | module_param_named(power_efficient, wq_power_efficient, bool, 0444); |
404 | ||
863b710b | 405 | static bool wq_online; /* can kworkers be created yet? */ |
3347fa09 | 406 | |
fef59c9c TH |
407 | /* buf for wq_update_unbound_pod_attrs(), protected by CPU hotplug exclusion */ |
408 | static struct workqueue_attrs *wq_update_pod_attrs_buf; | |
4c16bd32 | 409 | |
68e13a67 | 410 | static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ |
1258fae7 | 411 | static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */ |
a9b8a985 | 412 | static DEFINE_RAW_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ |
d8bb65ab SAS |
413 | /* wait for manager to go away */ |
414 | static struct rcuwait manager_wait = __RCUWAIT_INITIALIZER(manager_wait); | |
5bcab335 | 415 | |
e2dca7ad | 416 | static LIST_HEAD(workqueues); /* PR: list of all workqueues */ |
68e13a67 | 417 | static bool workqueue_freezing; /* PL: have wqs started freezing? */ |
7d19c5ce | 418 | |
99c621ef | 419 | /* PL&A: allowable cpus for unbound wqs and work items */ |
ef557180 MG |
420 | static cpumask_var_t wq_unbound_cpumask; |
421 | ||
fe28f631 WL |
422 | /* PL: user requested unbound cpumask via sysfs */ |
423 | static cpumask_var_t wq_requested_unbound_cpumask; | |
424 | ||
425 | /* PL: isolated cpumask to be excluded from unbound cpumask */ | |
426 | static cpumask_var_t wq_isolated_cpumask; | |
427 | ||
ace3c549 | 428 | /* for further constrain wq_unbound_cpumask by cmdline parameter*/ |
429 | static struct cpumask wq_cmdline_cpumask __initdata; | |
430 | ||
ef557180 MG |
431 | /* CPU where unbound work was last round robin scheduled from this CPU */ |
432 | static DEFINE_PER_CPU(int, wq_rr_cpu_last); | |
b05a7928 | 433 | |
f303fccb TH |
434 | /* |
435 | * Local execution of unbound work items is no longer guaranteed. The | |
436 | * following always forces round-robin CPU selection on unbound work items | |
437 | * to uncover usages which depend on it. | |
438 | */ | |
439 | #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU | |
440 | static bool wq_debug_force_rr_cpu = true; | |
441 | #else | |
442 | static bool wq_debug_force_rr_cpu = false; | |
443 | #endif | |
444 | module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644); | |
445 | ||
7d19c5ce | 446 | /* the per-cpu worker pools */ |
25528213 | 447 | static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools); |
7d19c5ce | 448 | |
68e13a67 | 449 | static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ |
7d19c5ce | 450 | |
68e13a67 | 451 | /* PL: hash of all unbound pools keyed by pool->attrs */ |
29c91e99 TH |
452 | static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); |
453 | ||
c5aa87bb | 454 | /* I: attributes used when instantiating standard unbound pools on demand */ |
29c91e99 TH |
455 | static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; |
456 | ||
8a2b7538 TH |
457 | /* I: attributes used when instantiating ordered pools on demand */ |
458 | static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS]; | |
459 | ||
967b494e TH |
460 | /* |
461 | * I: kthread_worker to release pwq's. pwq release needs to be bounced to a | |
462 | * process context while holding a pool lock. Bounce to a dedicated kthread | |
463 | * worker to avoid A-A deadlocks. | |
464 | */ | |
68279f9c | 465 | static struct kthread_worker *pwq_release_worker __ro_after_init; |
967b494e | 466 | |
68279f9c | 467 | struct workqueue_struct *system_wq __ro_after_init; |
ad7b1f84 | 468 | EXPORT_SYMBOL(system_wq); |
68279f9c | 469 | struct workqueue_struct *system_highpri_wq __ro_after_init; |
1aabe902 | 470 | EXPORT_SYMBOL_GPL(system_highpri_wq); |
68279f9c | 471 | struct workqueue_struct *system_long_wq __ro_after_init; |
d320c038 | 472 | EXPORT_SYMBOL_GPL(system_long_wq); |
68279f9c | 473 | struct workqueue_struct *system_unbound_wq __ro_after_init; |
f3421797 | 474 | EXPORT_SYMBOL_GPL(system_unbound_wq); |
68279f9c | 475 | struct workqueue_struct *system_freezable_wq __ro_after_init; |
24d51add | 476 | EXPORT_SYMBOL_GPL(system_freezable_wq); |
68279f9c | 477 | struct workqueue_struct *system_power_efficient_wq __ro_after_init; |
0668106c | 478 | EXPORT_SYMBOL_GPL(system_power_efficient_wq); |
68279f9c | 479 | struct workqueue_struct *system_freezable_power_efficient_wq __ro_after_init; |
0668106c | 480 | EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq); |
d320c038 | 481 | |
7d19c5ce | 482 | static int worker_thread(void *__worker); |
6ba94429 | 483 | static void workqueue_sysfs_unregister(struct workqueue_struct *wq); |
c29eb853 | 484 | static void show_pwq(struct pool_workqueue *pwq); |
55df0933 | 485 | static void show_one_worker_pool(struct worker_pool *pool); |
7d19c5ce | 486 | |
97bd2347 TH |
487 | #define CREATE_TRACE_POINTS |
488 | #include <trace/events/workqueue.h> | |
489 | ||
68e13a67 | 490 | #define assert_rcu_or_pool_mutex() \ |
24acfb71 | 491 | RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \ |
f78f5b90 | 492 | !lockdep_is_held(&wq_pool_mutex), \ |
24acfb71 | 493 | "RCU or wq_pool_mutex should be held") |
5bcab335 | 494 | |
5b95e1af | 495 | #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \ |
24acfb71 | 496 | RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \ |
f78f5b90 PM |
497 | !lockdep_is_held(&wq->mutex) && \ |
498 | !lockdep_is_held(&wq_pool_mutex), \ | |
24acfb71 | 499 | "RCU, wq->mutex or wq_pool_mutex should be held") |
5b95e1af | 500 | |
f02ae73a TH |
501 | #define for_each_cpu_worker_pool(pool, cpu) \ |
502 | for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ | |
503 | (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ | |
7a62c2c8 | 504 | (pool)++) |
4ce62e9e | 505 | |
17116969 TH |
506 | /** |
507 | * for_each_pool - iterate through all worker_pools in the system | |
508 | * @pool: iteration cursor | |
611c92a0 | 509 | * @pi: integer used for iteration |
fa1b54e6 | 510 | * |
24acfb71 | 511 | * This must be called either with wq_pool_mutex held or RCU read |
68e13a67 LJ |
512 | * locked. If the pool needs to be used beyond the locking in effect, the |
513 | * caller is responsible for guaranteeing that the pool stays online. | |
fa1b54e6 TH |
514 | * |
515 | * The if/else clause exists only for the lockdep assertion and can be | |
516 | * ignored. | |
17116969 | 517 | */ |
611c92a0 TH |
518 | #define for_each_pool(pool, pi) \ |
519 | idr_for_each_entry(&worker_pool_idr, pool, pi) \ | |
68e13a67 | 520 | if (({ assert_rcu_or_pool_mutex(); false; })) { } \ |
fa1b54e6 | 521 | else |
17116969 | 522 | |
822d8405 TH |
523 | /** |
524 | * for_each_pool_worker - iterate through all workers of a worker_pool | |
525 | * @worker: iteration cursor | |
822d8405 TH |
526 | * @pool: worker_pool to iterate workers of |
527 | * | |
1258fae7 | 528 | * This must be called with wq_pool_attach_mutex. |
822d8405 TH |
529 | * |
530 | * The if/else clause exists only for the lockdep assertion and can be | |
531 | * ignored. | |
532 | */ | |
da028469 LJ |
533 | #define for_each_pool_worker(worker, pool) \ |
534 | list_for_each_entry((worker), &(pool)->workers, node) \ | |
1258fae7 | 535 | if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \ |
822d8405 TH |
536 | else |
537 | ||
49e3cf44 TH |
538 | /** |
539 | * for_each_pwq - iterate through all pool_workqueues of the specified workqueue | |
540 | * @pwq: iteration cursor | |
541 | * @wq: the target workqueue | |
76af4d93 | 542 | * |
24acfb71 | 543 | * This must be called either with wq->mutex held or RCU read locked. |
794b18bc TH |
544 | * If the pwq needs to be used beyond the locking in effect, the caller is |
545 | * responsible for guaranteeing that the pwq stays online. | |
76af4d93 TH |
546 | * |
547 | * The if/else clause exists only for the lockdep assertion and can be | |
548 | * ignored. | |
49e3cf44 TH |
549 | */ |
550 | #define for_each_pwq(pwq, wq) \ | |
49e9d1a9 | 551 | list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \ |
5a644662 | 552 | lockdep_is_held(&(wq->mutex))) |
f3421797 | 553 | |
dc186ad7 TG |
554 | #ifdef CONFIG_DEBUG_OBJECTS_WORK |
555 | ||
f9e62f31 | 556 | static const struct debug_obj_descr work_debug_descr; |
dc186ad7 | 557 | |
99777288 SG |
558 | static void *work_debug_hint(void *addr) |
559 | { | |
560 | return ((struct work_struct *) addr)->func; | |
561 | } | |
562 | ||
b9fdac7f DC |
563 | static bool work_is_static_object(void *addr) |
564 | { | |
565 | struct work_struct *work = addr; | |
566 | ||
567 | return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work)); | |
568 | } | |
569 | ||
dc186ad7 TG |
570 | /* |
571 | * fixup_init is called when: | |
572 | * - an active object is initialized | |
573 | */ | |
02a982a6 | 574 | static bool work_fixup_init(void *addr, enum debug_obj_state state) |
dc186ad7 TG |
575 | { |
576 | struct work_struct *work = addr; | |
577 | ||
578 | switch (state) { | |
579 | case ODEBUG_STATE_ACTIVE: | |
580 | cancel_work_sync(work); | |
581 | debug_object_init(work, &work_debug_descr); | |
02a982a6 | 582 | return true; |
dc186ad7 | 583 | default: |
02a982a6 | 584 | return false; |
dc186ad7 TG |
585 | } |
586 | } | |
587 | ||
dc186ad7 TG |
588 | /* |
589 | * fixup_free is called when: | |
590 | * - an active object is freed | |
591 | */ | |
02a982a6 | 592 | static bool work_fixup_free(void *addr, enum debug_obj_state state) |
dc186ad7 TG |
593 | { |
594 | struct work_struct *work = addr; | |
595 | ||
596 | switch (state) { | |
597 | case ODEBUG_STATE_ACTIVE: | |
598 | cancel_work_sync(work); | |
599 | debug_object_free(work, &work_debug_descr); | |
02a982a6 | 600 | return true; |
dc186ad7 | 601 | default: |
02a982a6 | 602 | return false; |
dc186ad7 TG |
603 | } |
604 | } | |
605 | ||
f9e62f31 | 606 | static const struct debug_obj_descr work_debug_descr = { |
dc186ad7 | 607 | .name = "work_struct", |
99777288 | 608 | .debug_hint = work_debug_hint, |
b9fdac7f | 609 | .is_static_object = work_is_static_object, |
dc186ad7 | 610 | .fixup_init = work_fixup_init, |
dc186ad7 TG |
611 | .fixup_free = work_fixup_free, |
612 | }; | |
613 | ||
614 | static inline void debug_work_activate(struct work_struct *work) | |
615 | { | |
616 | debug_object_activate(work, &work_debug_descr); | |
617 | } | |
618 | ||
619 | static inline void debug_work_deactivate(struct work_struct *work) | |
620 | { | |
621 | debug_object_deactivate(work, &work_debug_descr); | |
622 | } | |
623 | ||
624 | void __init_work(struct work_struct *work, int onstack) | |
625 | { | |
626 | if (onstack) | |
627 | debug_object_init_on_stack(work, &work_debug_descr); | |
628 | else | |
629 | debug_object_init(work, &work_debug_descr); | |
630 | } | |
631 | EXPORT_SYMBOL_GPL(__init_work); | |
632 | ||
633 | void destroy_work_on_stack(struct work_struct *work) | |
634 | { | |
635 | debug_object_free(work, &work_debug_descr); | |
636 | } | |
637 | EXPORT_SYMBOL_GPL(destroy_work_on_stack); | |
638 | ||
ea2e64f2 TG |
639 | void destroy_delayed_work_on_stack(struct delayed_work *work) |
640 | { | |
641 | destroy_timer_on_stack(&work->timer); | |
642 | debug_object_free(&work->work, &work_debug_descr); | |
643 | } | |
644 | EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack); | |
645 | ||
dc186ad7 TG |
646 | #else |
647 | static inline void debug_work_activate(struct work_struct *work) { } | |
648 | static inline void debug_work_deactivate(struct work_struct *work) { } | |
649 | #endif | |
650 | ||
4e8b22bd | 651 | /** |
67dc8325 | 652 | * worker_pool_assign_id - allocate ID and assign it to @pool |
4e8b22bd LB |
653 | * @pool: the pool pointer of interest |
654 | * | |
655 | * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned | |
656 | * successfully, -errno on failure. | |
657 | */ | |
9daf9e67 TH |
658 | static int worker_pool_assign_id(struct worker_pool *pool) |
659 | { | |
660 | int ret; | |
661 | ||
68e13a67 | 662 | lockdep_assert_held(&wq_pool_mutex); |
5bcab335 | 663 | |
4e8b22bd LB |
664 | ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE, |
665 | GFP_KERNEL); | |
229641a6 | 666 | if (ret >= 0) { |
e68035fb | 667 | pool->id = ret; |
229641a6 TH |
668 | return 0; |
669 | } | |
fa1b54e6 | 670 | return ret; |
7c3eed5c TH |
671 | } |
672 | ||
9f66cff2 TH |
673 | static struct pool_workqueue __rcu ** |
674 | unbound_pwq_slot(struct workqueue_struct *wq, int cpu) | |
675 | { | |
676 | if (cpu >= 0) | |
677 | return per_cpu_ptr(wq->cpu_pwq, cpu); | |
678 | else | |
679 | return &wq->dfl_pwq; | |
680 | } | |
681 | ||
682 | /* @cpu < 0 for dfl_pwq */ | |
683 | static struct pool_workqueue *unbound_pwq(struct workqueue_struct *wq, int cpu) | |
684 | { | |
685 | return rcu_dereference_check(*unbound_pwq_slot(wq, cpu), | |
686 | lockdep_is_held(&wq_pool_mutex) || | |
687 | lockdep_is_held(&wq->mutex)); | |
688 | } | |
689 | ||
5797b1c1 TH |
690 | /** |
691 | * unbound_effective_cpumask - effective cpumask of an unbound workqueue | |
692 | * @wq: workqueue of interest | |
693 | * | |
694 | * @wq->unbound_attrs->cpumask contains the cpumask requested by the user which | |
695 | * is masked with wq_unbound_cpumask to determine the effective cpumask. The | |
696 | * default pwq is always mapped to the pool with the current effective cpumask. | |
697 | */ | |
698 | static struct cpumask *unbound_effective_cpumask(struct workqueue_struct *wq) | |
699 | { | |
700 | return unbound_pwq(wq, -1)->pool->attrs->__pod_cpumask; | |
701 | } | |
702 | ||
73f53c4a TH |
703 | static unsigned int work_color_to_flags(int color) |
704 | { | |
705 | return color << WORK_STRUCT_COLOR_SHIFT; | |
706 | } | |
707 | ||
c4560c2c | 708 | static int get_work_color(unsigned long work_data) |
73f53c4a | 709 | { |
c4560c2c | 710 | return (work_data >> WORK_STRUCT_COLOR_SHIFT) & |
73f53c4a TH |
711 | ((1 << WORK_STRUCT_COLOR_BITS) - 1); |
712 | } | |
713 | ||
714 | static int work_next_color(int color) | |
715 | { | |
716 | return (color + 1) % WORK_NR_COLORS; | |
717 | } | |
1da177e4 | 718 | |
14441960 | 719 | /* |
112202d9 TH |
720 | * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data |
721 | * contain the pointer to the queued pwq. Once execution starts, the flag | |
7c3eed5c | 722 | * is cleared and the high bits contain OFFQ flags and pool ID. |
7a22ad75 | 723 | * |
112202d9 TH |
724 | * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() |
725 | * and clear_work_data() can be used to set the pwq, pool or clear | |
bbb68dfa TH |
726 | * work->data. These functions should only be called while the work is |
727 | * owned - ie. while the PENDING bit is set. | |
7a22ad75 | 728 | * |
112202d9 | 729 | * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq |
7c3eed5c | 730 | * corresponding to a work. Pool is available once the work has been |
112202d9 | 731 | * queued anywhere after initialization until it is sync canceled. pwq is |
7c3eed5c | 732 | * available only while the work item is queued. |
7a22ad75 | 733 | * |
bbb68dfa TH |
734 | * %WORK_OFFQ_CANCELING is used to mark a work item which is being |
735 | * canceled. While being canceled, a work item may have its PENDING set | |
736 | * but stay off timer and worklist for arbitrarily long and nobody should | |
737 | * try to steal the PENDING bit. | |
14441960 | 738 | */ |
7a22ad75 TH |
739 | static inline void set_work_data(struct work_struct *work, unsigned long data, |
740 | unsigned long flags) | |
365970a1 | 741 | { |
6183c009 | 742 | WARN_ON_ONCE(!work_pending(work)); |
7a22ad75 TH |
743 | atomic_long_set(&work->data, data | flags | work_static(work)); |
744 | } | |
365970a1 | 745 | |
112202d9 | 746 | static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, |
7a22ad75 TH |
747 | unsigned long extra_flags) |
748 | { | |
112202d9 TH |
749 | set_work_data(work, (unsigned long)pwq, |
750 | WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); | |
365970a1 DH |
751 | } |
752 | ||
4468a00f LJ |
753 | static void set_work_pool_and_keep_pending(struct work_struct *work, |
754 | int pool_id) | |
755 | { | |
756 | set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, | |
757 | WORK_STRUCT_PENDING); | |
758 | } | |
759 | ||
7c3eed5c TH |
760 | static void set_work_pool_and_clear_pending(struct work_struct *work, |
761 | int pool_id) | |
7a22ad75 | 762 | { |
23657bb1 TH |
763 | /* |
764 | * The following wmb is paired with the implied mb in | |
765 | * test_and_set_bit(PENDING) and ensures all updates to @work made | |
766 | * here are visible to and precede any updates by the next PENDING | |
767 | * owner. | |
768 | */ | |
769 | smp_wmb(); | |
7c3eed5c | 770 | set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); |
346c09f8 RP |
771 | /* |
772 | * The following mb guarantees that previous clear of a PENDING bit | |
773 | * will not be reordered with any speculative LOADS or STORES from | |
774 | * work->current_func, which is executed afterwards. This possible | |
8bdc6201 | 775 | * reordering can lead to a missed execution on attempt to queue |
346c09f8 RP |
776 | * the same @work. E.g. consider this case: |
777 | * | |
778 | * CPU#0 CPU#1 | |
779 | * ---------------------------- -------------------------------- | |
780 | * | |
781 | * 1 STORE event_indicated | |
782 | * 2 queue_work_on() { | |
783 | * 3 test_and_set_bit(PENDING) | |
784 | * 4 } set_..._and_clear_pending() { | |
785 | * 5 set_work_data() # clear bit | |
786 | * 6 smp_mb() | |
787 | * 7 work->current_func() { | |
788 | * 8 LOAD event_indicated | |
789 | * } | |
790 | * | |
791 | * Without an explicit full barrier speculative LOAD on line 8 can | |
792 | * be executed before CPU#0 does STORE on line 1. If that happens, | |
793 | * CPU#0 observes the PENDING bit is still set and new execution of | |
794 | * a @work is not queued in a hope, that CPU#1 will eventually | |
795 | * finish the queued @work. Meanwhile CPU#1 does not see | |
796 | * event_indicated is set, because speculative LOAD was executed | |
797 | * before actual STORE. | |
798 | */ | |
799 | smp_mb(); | |
7a22ad75 | 800 | } |
f756d5e2 | 801 | |
7a22ad75 | 802 | static void clear_work_data(struct work_struct *work) |
1da177e4 | 803 | { |
7c3eed5c TH |
804 | smp_wmb(); /* see set_work_pool_and_clear_pending() */ |
805 | set_work_data(work, WORK_STRUCT_NO_POOL, 0); | |
1da177e4 LT |
806 | } |
807 | ||
afa4bb77 LT |
808 | static inline struct pool_workqueue *work_struct_pwq(unsigned long data) |
809 | { | |
810 | return (struct pool_workqueue *)(data & WORK_STRUCT_WQ_DATA_MASK); | |
811 | } | |
812 | ||
112202d9 | 813 | static struct pool_workqueue *get_work_pwq(struct work_struct *work) |
b1f4ec17 | 814 | { |
e120153d | 815 | unsigned long data = atomic_long_read(&work->data); |
7a22ad75 | 816 | |
112202d9 | 817 | if (data & WORK_STRUCT_PWQ) |
afa4bb77 | 818 | return work_struct_pwq(data); |
e120153d TH |
819 | else |
820 | return NULL; | |
4d707b9f ON |
821 | } |
822 | ||
7c3eed5c TH |
823 | /** |
824 | * get_work_pool - return the worker_pool a given work was associated with | |
825 | * @work: the work item of interest | |
826 | * | |
68e13a67 | 827 | * Pools are created and destroyed under wq_pool_mutex, and allows read |
24acfb71 TG |
828 | * access under RCU read lock. As such, this function should be |
829 | * called under wq_pool_mutex or inside of a rcu_read_lock() region. | |
fa1b54e6 TH |
830 | * |
831 | * All fields of the returned pool are accessible as long as the above | |
832 | * mentioned locking is in effect. If the returned pool needs to be used | |
833 | * beyond the critical section, the caller is responsible for ensuring the | |
834 | * returned pool is and stays online. | |
d185af30 YB |
835 | * |
836 | * Return: The worker_pool @work was last associated with. %NULL if none. | |
7c3eed5c TH |
837 | */ |
838 | static struct worker_pool *get_work_pool(struct work_struct *work) | |
365970a1 | 839 | { |
e120153d | 840 | unsigned long data = atomic_long_read(&work->data); |
7c3eed5c | 841 | int pool_id; |
7a22ad75 | 842 | |
68e13a67 | 843 | assert_rcu_or_pool_mutex(); |
fa1b54e6 | 844 | |
112202d9 | 845 | if (data & WORK_STRUCT_PWQ) |
afa4bb77 | 846 | return work_struct_pwq(data)->pool; |
7a22ad75 | 847 | |
7c3eed5c TH |
848 | pool_id = data >> WORK_OFFQ_POOL_SHIFT; |
849 | if (pool_id == WORK_OFFQ_POOL_NONE) | |
7a22ad75 TH |
850 | return NULL; |
851 | ||
fa1b54e6 | 852 | return idr_find(&worker_pool_idr, pool_id); |
7c3eed5c TH |
853 | } |
854 | ||
855 | /** | |
856 | * get_work_pool_id - return the worker pool ID a given work is associated with | |
857 | * @work: the work item of interest | |
858 | * | |
d185af30 | 859 | * Return: The worker_pool ID @work was last associated with. |
7c3eed5c TH |
860 | * %WORK_OFFQ_POOL_NONE if none. |
861 | */ | |
862 | static int get_work_pool_id(struct work_struct *work) | |
863 | { | |
54d5b7d0 LJ |
864 | unsigned long data = atomic_long_read(&work->data); |
865 | ||
112202d9 | 866 | if (data & WORK_STRUCT_PWQ) |
afa4bb77 | 867 | return work_struct_pwq(data)->pool->id; |
7c3eed5c | 868 | |
54d5b7d0 | 869 | return data >> WORK_OFFQ_POOL_SHIFT; |
7c3eed5c TH |
870 | } |
871 | ||
bbb68dfa TH |
872 | static void mark_work_canceling(struct work_struct *work) |
873 | { | |
7c3eed5c | 874 | unsigned long pool_id = get_work_pool_id(work); |
bbb68dfa | 875 | |
7c3eed5c TH |
876 | pool_id <<= WORK_OFFQ_POOL_SHIFT; |
877 | set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); | |
bbb68dfa TH |
878 | } |
879 | ||
880 | static bool work_is_canceling(struct work_struct *work) | |
881 | { | |
882 | unsigned long data = atomic_long_read(&work->data); | |
883 | ||
112202d9 | 884 | return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); |
bbb68dfa TH |
885 | } |
886 | ||
e22bee78 | 887 | /* |
3270476a TH |
888 | * Policy functions. These define the policies on how the global worker |
889 | * pools are managed. Unless noted otherwise, these functions assume that | |
d565ed63 | 890 | * they're being called with pool->lock held. |
e22bee78 TH |
891 | */ |
892 | ||
4594bf15 | 893 | /* |
e22bee78 TH |
894 | * Need to wake up a worker? Called from anything but currently |
895 | * running workers. | |
974271c4 TH |
896 | * |
897 | * Note that, because unbound workers never contribute to nr_running, this | |
706026c2 | 898 | * function will always return %true for unbound pools as long as the |
974271c4 | 899 | * worklist isn't empty. |
4594bf15 | 900 | */ |
63d95a91 | 901 | static bool need_more_worker(struct worker_pool *pool) |
365970a1 | 902 | { |
0219a352 | 903 | return !list_empty(&pool->worklist) && !pool->nr_running; |
e22bee78 | 904 | } |
4594bf15 | 905 | |
e22bee78 | 906 | /* Can I start working? Called from busy but !running workers. */ |
63d95a91 | 907 | static bool may_start_working(struct worker_pool *pool) |
e22bee78 | 908 | { |
63d95a91 | 909 | return pool->nr_idle; |
e22bee78 TH |
910 | } |
911 | ||
912 | /* Do I need to keep working? Called from currently running workers. */ | |
63d95a91 | 913 | static bool keep_working(struct worker_pool *pool) |
e22bee78 | 914 | { |
bc35f7ef | 915 | return !list_empty(&pool->worklist) && (pool->nr_running <= 1); |
e22bee78 TH |
916 | } |
917 | ||
918 | /* Do we need a new worker? Called from manager. */ | |
63d95a91 | 919 | static bool need_to_create_worker(struct worker_pool *pool) |
e22bee78 | 920 | { |
63d95a91 | 921 | return need_more_worker(pool) && !may_start_working(pool); |
e22bee78 | 922 | } |
365970a1 | 923 | |
e22bee78 | 924 | /* Do we have too many workers and should some go away? */ |
63d95a91 | 925 | static bool too_many_workers(struct worker_pool *pool) |
e22bee78 | 926 | { |
692b4825 | 927 | bool managing = pool->flags & POOL_MANAGER_ACTIVE; |
63d95a91 TH |
928 | int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ |
929 | int nr_busy = pool->nr_workers - nr_idle; | |
e22bee78 TH |
930 | |
931 | return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; | |
365970a1 DH |
932 | } |
933 | ||
c54d5046 TH |
934 | /** |
935 | * worker_set_flags - set worker flags and adjust nr_running accordingly | |
936 | * @worker: self | |
937 | * @flags: flags to set | |
938 | * | |
939 | * Set @flags in @worker->flags and adjust nr_running accordingly. | |
c54d5046 TH |
940 | */ |
941 | static inline void worker_set_flags(struct worker *worker, unsigned int flags) | |
942 | { | |
943 | struct worker_pool *pool = worker->pool; | |
944 | ||
bc8b50c2 | 945 | lockdep_assert_held(&pool->lock); |
c54d5046 TH |
946 | |
947 | /* If transitioning into NOT_RUNNING, adjust nr_running. */ | |
948 | if ((flags & WORKER_NOT_RUNNING) && | |
949 | !(worker->flags & WORKER_NOT_RUNNING)) { | |
950 | pool->nr_running--; | |
951 | } | |
952 | ||
953 | worker->flags |= flags; | |
954 | } | |
955 | ||
956 | /** | |
957 | * worker_clr_flags - clear worker flags and adjust nr_running accordingly | |
958 | * @worker: self | |
959 | * @flags: flags to clear | |
960 | * | |
961 | * Clear @flags in @worker->flags and adjust nr_running accordingly. | |
c54d5046 TH |
962 | */ |
963 | static inline void worker_clr_flags(struct worker *worker, unsigned int flags) | |
964 | { | |
965 | struct worker_pool *pool = worker->pool; | |
966 | unsigned int oflags = worker->flags; | |
967 | ||
bc8b50c2 | 968 | lockdep_assert_held(&pool->lock); |
c54d5046 TH |
969 | |
970 | worker->flags &= ~flags; | |
971 | ||
972 | /* | |
973 | * If transitioning out of NOT_RUNNING, increment nr_running. Note | |
974 | * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask | |
975 | * of multiple flags, not a single flag. | |
976 | */ | |
977 | if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) | |
978 | if (!(worker->flags & WORKER_NOT_RUNNING)) | |
979 | pool->nr_running++; | |
980 | } | |
981 | ||
797e8345 TH |
982 | /* Return the first idle worker. Called with pool->lock held. */ |
983 | static struct worker *first_idle_worker(struct worker_pool *pool) | |
984 | { | |
985 | if (unlikely(list_empty(&pool->idle_list))) | |
986 | return NULL; | |
987 | ||
988 | return list_first_entry(&pool->idle_list, struct worker, entry); | |
989 | } | |
990 | ||
991 | /** | |
992 | * worker_enter_idle - enter idle state | |
993 | * @worker: worker which is entering idle state | |
994 | * | |
995 | * @worker is entering idle state. Update stats and idle timer if | |
996 | * necessary. | |
997 | * | |
998 | * LOCKING: | |
999 | * raw_spin_lock_irq(pool->lock). | |
1000 | */ | |
1001 | static void worker_enter_idle(struct worker *worker) | |
1002 | { | |
1003 | struct worker_pool *pool = worker->pool; | |
1004 | ||
1005 | if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || | |
1006 | WARN_ON_ONCE(!list_empty(&worker->entry) && | |
1007 | (worker->hentry.next || worker->hentry.pprev))) | |
1008 | return; | |
1009 | ||
1010 | /* can't use worker_set_flags(), also called from create_worker() */ | |
1011 | worker->flags |= WORKER_IDLE; | |
1012 | pool->nr_idle++; | |
1013 | worker->last_active = jiffies; | |
1014 | ||
1015 | /* idle_list is LIFO */ | |
1016 | list_add(&worker->entry, &pool->idle_list); | |
1017 | ||
1018 | if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) | |
1019 | mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); | |
1020 | ||
1021 | /* Sanity check nr_running. */ | |
1022 | WARN_ON_ONCE(pool->nr_workers == pool->nr_idle && pool->nr_running); | |
1023 | } | |
1024 | ||
1025 | /** | |
1026 | * worker_leave_idle - leave idle state | |
1027 | * @worker: worker which is leaving idle state | |
1028 | * | |
1029 | * @worker is leaving idle state. Update stats. | |
1030 | * | |
1031 | * LOCKING: | |
1032 | * raw_spin_lock_irq(pool->lock). | |
1033 | */ | |
1034 | static void worker_leave_idle(struct worker *worker) | |
1035 | { | |
1036 | struct worker_pool *pool = worker->pool; | |
1037 | ||
1038 | if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) | |
1039 | return; | |
1040 | worker_clr_flags(worker, WORKER_IDLE); | |
1041 | pool->nr_idle--; | |
1042 | list_del_init(&worker->entry); | |
1043 | } | |
1044 | ||
1045 | /** | |
1046 | * find_worker_executing_work - find worker which is executing a work | |
1047 | * @pool: pool of interest | |
1048 | * @work: work to find worker for | |
1049 | * | |
1050 | * Find a worker which is executing @work on @pool by searching | |
1051 | * @pool->busy_hash which is keyed by the address of @work. For a worker | |
1052 | * to match, its current execution should match the address of @work and | |
1053 | * its work function. This is to avoid unwanted dependency between | |
1054 | * unrelated work executions through a work item being recycled while still | |
1055 | * being executed. | |
1056 | * | |
1057 | * This is a bit tricky. A work item may be freed once its execution | |
1058 | * starts and nothing prevents the freed area from being recycled for | |
1059 | * another work item. If the same work item address ends up being reused | |
1060 | * before the original execution finishes, workqueue will identify the | |
1061 | * recycled work item as currently executing and make it wait until the | |
1062 | * current execution finishes, introducing an unwanted dependency. | |
1063 | * | |
1064 | * This function checks the work item address and work function to avoid | |
1065 | * false positives. Note that this isn't complete as one may construct a | |
1066 | * work function which can introduce dependency onto itself through a | |
1067 | * recycled work item. Well, if somebody wants to shoot oneself in the | |
1068 | * foot that badly, there's only so much we can do, and if such deadlock | |
1069 | * actually occurs, it should be easy to locate the culprit work function. | |
1070 | * | |
1071 | * CONTEXT: | |
1072 | * raw_spin_lock_irq(pool->lock). | |
1073 | * | |
1074 | * Return: | |
1075 | * Pointer to worker which is executing @work if found, %NULL | |
1076 | * otherwise. | |
1077 | */ | |
1078 | static struct worker *find_worker_executing_work(struct worker_pool *pool, | |
1079 | struct work_struct *work) | |
1080 | { | |
1081 | struct worker *worker; | |
1082 | ||
1083 | hash_for_each_possible(pool->busy_hash, worker, hentry, | |
1084 | (unsigned long)work) | |
1085 | if (worker->current_work == work && | |
1086 | worker->current_func == work->func) | |
1087 | return worker; | |
1088 | ||
1089 | return NULL; | |
1090 | } | |
1091 | ||
1092 | /** | |
1093 | * move_linked_works - move linked works to a list | |
1094 | * @work: start of series of works to be scheduled | |
1095 | * @head: target list to append @work to | |
1096 | * @nextp: out parameter for nested worklist walking | |
1097 | * | |
873eaca6 TH |
1098 | * Schedule linked works starting from @work to @head. Work series to be |
1099 | * scheduled starts at @work and includes any consecutive work with | |
1100 | * WORK_STRUCT_LINKED set in its predecessor. See assign_work() for details on | |
1101 | * @nextp. | |
797e8345 TH |
1102 | * |
1103 | * CONTEXT: | |
1104 | * raw_spin_lock_irq(pool->lock). | |
1105 | */ | |
1106 | static void move_linked_works(struct work_struct *work, struct list_head *head, | |
1107 | struct work_struct **nextp) | |
1108 | { | |
1109 | struct work_struct *n; | |
1110 | ||
1111 | /* | |
1112 | * Linked worklist will always end before the end of the list, | |
1113 | * use NULL for list head. | |
1114 | */ | |
1115 | list_for_each_entry_safe_from(work, n, NULL, entry) { | |
1116 | list_move_tail(&work->entry, head); | |
1117 | if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) | |
1118 | break; | |
1119 | } | |
1120 | ||
1121 | /* | |
1122 | * If we're already inside safe list traversal and have moved | |
1123 | * multiple works to the scheduled queue, the next position | |
1124 | * needs to be updated. | |
1125 | */ | |
1126 | if (nextp) | |
1127 | *nextp = n; | |
1128 | } | |
1129 | ||
873eaca6 TH |
1130 | /** |
1131 | * assign_work - assign a work item and its linked work items to a worker | |
1132 | * @work: work to assign | |
1133 | * @worker: worker to assign to | |
1134 | * @nextp: out parameter for nested worklist walking | |
1135 | * | |
1136 | * Assign @work and its linked work items to @worker. If @work is already being | |
1137 | * executed by another worker in the same pool, it'll be punted there. | |
1138 | * | |
1139 | * If @nextp is not NULL, it's updated to point to the next work of the last | |
1140 | * scheduled work. This allows assign_work() to be nested inside | |
1141 | * list_for_each_entry_safe(). | |
1142 | * | |
1143 | * Returns %true if @work was successfully assigned to @worker. %false if @work | |
1144 | * was punted to another worker already executing it. | |
1145 | */ | |
1146 | static bool assign_work(struct work_struct *work, struct worker *worker, | |
1147 | struct work_struct **nextp) | |
1148 | { | |
1149 | struct worker_pool *pool = worker->pool; | |
1150 | struct worker *collision; | |
1151 | ||
1152 | lockdep_assert_held(&pool->lock); | |
1153 | ||
1154 | /* | |
1155 | * A single work shouldn't be executed concurrently by multiple workers. | |
1156 | * __queue_work() ensures that @work doesn't jump to a different pool | |
1157 | * while still running in the previous pool. Here, we should ensure that | |
1158 | * @work is not executed concurrently by multiple workers from the same | |
1159 | * pool. Check whether anyone is already processing the work. If so, | |
1160 | * defer the work to the currently executing one. | |
1161 | */ | |
1162 | collision = find_worker_executing_work(pool, work); | |
1163 | if (unlikely(collision)) { | |
1164 | move_linked_works(work, &collision->scheduled, nextp); | |
1165 | return false; | |
1166 | } | |
1167 | ||
1168 | move_linked_works(work, &worker->scheduled, nextp); | |
1169 | return true; | |
1170 | } | |
1171 | ||
797e8345 | 1172 | /** |
0219a352 TH |
1173 | * kick_pool - wake up an idle worker if necessary |
1174 | * @pool: pool to kick | |
797e8345 | 1175 | * |
0219a352 TH |
1176 | * @pool may have pending work items. Wake up worker if necessary. Returns |
1177 | * whether a worker was woken up. | |
797e8345 | 1178 | */ |
0219a352 | 1179 | static bool kick_pool(struct worker_pool *pool) |
797e8345 TH |
1180 | { |
1181 | struct worker *worker = first_idle_worker(pool); | |
8639eceb | 1182 | struct task_struct *p; |
797e8345 | 1183 | |
0219a352 TH |
1184 | lockdep_assert_held(&pool->lock); |
1185 | ||
1186 | if (!need_more_worker(pool) || !worker) | |
1187 | return false; | |
1188 | ||
8639eceb TH |
1189 | p = worker->task; |
1190 | ||
1191 | #ifdef CONFIG_SMP | |
1192 | /* | |
1193 | * Idle @worker is about to execute @work and waking up provides an | |
1194 | * opportunity to migrate @worker at a lower cost by setting the task's | |
1195 | * wake_cpu field. Let's see if we want to move @worker to improve | |
1196 | * execution locality. | |
1197 | * | |
1198 | * We're waking the worker that went idle the latest and there's some | |
1199 | * chance that @worker is marked idle but hasn't gone off CPU yet. If | |
1200 | * so, setting the wake_cpu won't do anything. As this is a best-effort | |
1201 | * optimization and the race window is narrow, let's leave as-is for | |
1202 | * now. If this becomes pronounced, we can skip over workers which are | |
1203 | * still on cpu when picking an idle worker. | |
1204 | * | |
1205 | * If @pool has non-strict affinity, @worker might have ended up outside | |
1206 | * its affinity scope. Repatriate. | |
1207 | */ | |
1208 | if (!pool->attrs->affn_strict && | |
1209 | !cpumask_test_cpu(p->wake_cpu, pool->attrs->__pod_cpumask)) { | |
1210 | struct work_struct *work = list_first_entry(&pool->worklist, | |
1211 | struct work_struct, entry); | |
1212 | p->wake_cpu = cpumask_any_distribute(pool->attrs->__pod_cpumask); | |
1213 | get_work_pwq(work)->stats[PWQ_STAT_REPATRIATED]++; | |
1214 | } | |
1215 | #endif | |
1216 | wake_up_process(p); | |
0219a352 | 1217 | return true; |
797e8345 TH |
1218 | } |
1219 | ||
63638450 TH |
1220 | #ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT |
1221 | ||
1222 | /* | |
1223 | * Concurrency-managed per-cpu work items that hog CPU for longer than | |
1224 | * wq_cpu_intensive_thresh_us trigger the automatic CPU_INTENSIVE mechanism, | |
1225 | * which prevents them from stalling other concurrency-managed work items. If a | |
1226 | * work function keeps triggering this mechanism, it's likely that the work item | |
1227 | * should be using an unbound workqueue instead. | |
1228 | * | |
1229 | * wq_cpu_intensive_report() tracks work functions which trigger such conditions | |
1230 | * and report them so that they can be examined and converted to use unbound | |
1231 | * workqueues as appropriate. To avoid flooding the console, each violating work | |
1232 | * function is tracked and reported with exponential backoff. | |
1233 | */ | |
1234 | #define WCI_MAX_ENTS 128 | |
1235 | ||
1236 | struct wci_ent { | |
1237 | work_func_t func; | |
1238 | atomic64_t cnt; | |
1239 | struct hlist_node hash_node; | |
1240 | }; | |
1241 | ||
1242 | static struct wci_ent wci_ents[WCI_MAX_ENTS]; | |
1243 | static int wci_nr_ents; | |
1244 | static DEFINE_RAW_SPINLOCK(wci_lock); | |
1245 | static DEFINE_HASHTABLE(wci_hash, ilog2(WCI_MAX_ENTS)); | |
1246 | ||
1247 | static struct wci_ent *wci_find_ent(work_func_t func) | |
1248 | { | |
1249 | struct wci_ent *ent; | |
1250 | ||
1251 | hash_for_each_possible_rcu(wci_hash, ent, hash_node, | |
1252 | (unsigned long)func) { | |
1253 | if (ent->func == func) | |
1254 | return ent; | |
1255 | } | |
1256 | return NULL; | |
1257 | } | |
1258 | ||
1259 | static void wq_cpu_intensive_report(work_func_t func) | |
1260 | { | |
1261 | struct wci_ent *ent; | |
1262 | ||
1263 | restart: | |
1264 | ent = wci_find_ent(func); | |
1265 | if (ent) { | |
1266 | u64 cnt; | |
1267 | ||
1268 | /* | |
1269 | * Start reporting from the fourth time and back off | |
1270 | * exponentially. | |
1271 | */ | |
1272 | cnt = atomic64_inc_return_relaxed(&ent->cnt); | |
1273 | if (cnt >= 4 && is_power_of_2(cnt)) | |
1274 | printk_deferred(KERN_WARNING "workqueue: %ps hogged CPU for >%luus %llu times, consider switching to WQ_UNBOUND\n", | |
1275 | ent->func, wq_cpu_intensive_thresh_us, | |
1276 | atomic64_read(&ent->cnt)); | |
1277 | return; | |
1278 | } | |
1279 | ||
1280 | /* | |
1281 | * @func is a new violation. Allocate a new entry for it. If wcn_ents[] | |
1282 | * is exhausted, something went really wrong and we probably made enough | |
1283 | * noise already. | |
1284 | */ | |
1285 | if (wci_nr_ents >= WCI_MAX_ENTS) | |
1286 | return; | |
1287 | ||
1288 | raw_spin_lock(&wci_lock); | |
1289 | ||
1290 | if (wci_nr_ents >= WCI_MAX_ENTS) { | |
1291 | raw_spin_unlock(&wci_lock); | |
1292 | return; | |
1293 | } | |
1294 | ||
1295 | if (wci_find_ent(func)) { | |
1296 | raw_spin_unlock(&wci_lock); | |
1297 | goto restart; | |
1298 | } | |
1299 | ||
1300 | ent = &wci_ents[wci_nr_ents++]; | |
1301 | ent->func = func; | |
1302 | atomic64_set(&ent->cnt, 1); | |
1303 | hash_add_rcu(wci_hash, &ent->hash_node, (unsigned long)func); | |
1304 | ||
1305 | raw_spin_unlock(&wci_lock); | |
1306 | } | |
1307 | ||
1308 | #else /* CONFIG_WQ_CPU_INTENSIVE_REPORT */ | |
1309 | static void wq_cpu_intensive_report(work_func_t func) {} | |
1310 | #endif /* CONFIG_WQ_CPU_INTENSIVE_REPORT */ | |
1311 | ||
d302f017 | 1312 | /** |
6d25be57 | 1313 | * wq_worker_running - a worker is running again |
e22bee78 | 1314 | * @task: task waking up |
e22bee78 | 1315 | * |
6d25be57 | 1316 | * This function is called when a worker returns from schedule() |
e22bee78 | 1317 | */ |
6d25be57 | 1318 | void wq_worker_running(struct task_struct *task) |
e22bee78 TH |
1319 | { |
1320 | struct worker *worker = kthread_data(task); | |
1321 | ||
c8f6219b | 1322 | if (!READ_ONCE(worker->sleeping)) |
6d25be57 | 1323 | return; |
07edfece FW |
1324 | |
1325 | /* | |
1326 | * If preempted by unbind_workers() between the WORKER_NOT_RUNNING check | |
1327 | * and the nr_running increment below, we may ruin the nr_running reset | |
1328 | * and leave with an unexpected pool->nr_running == 1 on the newly unbound | |
1329 | * pool. Protect against such race. | |
1330 | */ | |
1331 | preempt_disable(); | |
6d25be57 | 1332 | if (!(worker->flags & WORKER_NOT_RUNNING)) |
bc35f7ef | 1333 | worker->pool->nr_running++; |
07edfece | 1334 | preempt_enable(); |
616db877 TH |
1335 | |
1336 | /* | |
1337 | * CPU intensive auto-detection cares about how long a work item hogged | |
1338 | * CPU without sleeping. Reset the starting timestamp on wakeup. | |
1339 | */ | |
1340 | worker->current_at = worker->task->se.sum_exec_runtime; | |
1341 | ||
c8f6219b | 1342 | WRITE_ONCE(worker->sleeping, 0); |
e22bee78 TH |
1343 | } |
1344 | ||
1345 | /** | |
1346 | * wq_worker_sleeping - a worker is going to sleep | |
1347 | * @task: task going to sleep | |
e22bee78 | 1348 | * |
6d25be57 | 1349 | * This function is called from schedule() when a busy worker is |
ccf45156 | 1350 | * going to sleep. |
e22bee78 | 1351 | */ |
6d25be57 | 1352 | void wq_worker_sleeping(struct task_struct *task) |
e22bee78 | 1353 | { |
cc5bff38 | 1354 | struct worker *worker = kthread_data(task); |
111c225a | 1355 | struct worker_pool *pool; |
e22bee78 | 1356 | |
111c225a TH |
1357 | /* |
1358 | * Rescuers, which may not have all the fields set up like normal | |
1359 | * workers, also reach here, let's not access anything before | |
1360 | * checking NOT_RUNNING. | |
1361 | */ | |
2d64672e | 1362 | if (worker->flags & WORKER_NOT_RUNNING) |
6d25be57 | 1363 | return; |
e22bee78 | 1364 | |
111c225a | 1365 | pool = worker->pool; |
111c225a | 1366 | |
62849a96 | 1367 | /* Return if preempted before wq_worker_running() was reached */ |
c8f6219b | 1368 | if (READ_ONCE(worker->sleeping)) |
6d25be57 TG |
1369 | return; |
1370 | ||
c8f6219b | 1371 | WRITE_ONCE(worker->sleeping, 1); |
a9b8a985 | 1372 | raw_spin_lock_irq(&pool->lock); |
e22bee78 | 1373 | |
45c753f5 FW |
1374 | /* |
1375 | * Recheck in case unbind_workers() preempted us. We don't | |
1376 | * want to decrement nr_running after the worker is unbound | |
1377 | * and nr_running has been reset. | |
1378 | */ | |
1379 | if (worker->flags & WORKER_NOT_RUNNING) { | |
1380 | raw_spin_unlock_irq(&pool->lock); | |
1381 | return; | |
1382 | } | |
1383 | ||
bc35f7ef | 1384 | pool->nr_running--; |
0219a352 | 1385 | if (kick_pool(pool)) |
725e8ec5 | 1386 | worker->current_pwq->stats[PWQ_STAT_CM_WAKEUP]++; |
0219a352 | 1387 | |
a9b8a985 | 1388 | raw_spin_unlock_irq(&pool->lock); |
e22bee78 TH |
1389 | } |
1390 | ||
616db877 TH |
1391 | /** |
1392 | * wq_worker_tick - a scheduler tick occurred while a kworker is running | |
1393 | * @task: task currently running | |
1394 | * | |
1395 | * Called from scheduler_tick(). We're in the IRQ context and the current | |
1396 | * worker's fields which follow the 'K' locking rule can be accessed safely. | |
1397 | */ | |
1398 | void wq_worker_tick(struct task_struct *task) | |
1399 | { | |
1400 | struct worker *worker = kthread_data(task); | |
1401 | struct pool_workqueue *pwq = worker->current_pwq; | |
1402 | struct worker_pool *pool = worker->pool; | |
1403 | ||
1404 | if (!pwq) | |
1405 | return; | |
1406 | ||
8a1dd1e5 TH |
1407 | pwq->stats[PWQ_STAT_CPU_TIME] += TICK_USEC; |
1408 | ||
18c8ae81 Z |
1409 | if (!wq_cpu_intensive_thresh_us) |
1410 | return; | |
1411 | ||
616db877 TH |
1412 | /* |
1413 | * If the current worker is concurrency managed and hogged the CPU for | |
1414 | * longer than wq_cpu_intensive_thresh_us, it's automatically marked | |
1415 | * CPU_INTENSIVE to avoid stalling other concurrency-managed work items. | |
c8f6219b Z |
1416 | * |
1417 | * Set @worker->sleeping means that @worker is in the process of | |
1418 | * switching out voluntarily and won't be contributing to | |
1419 | * @pool->nr_running until it wakes up. As wq_worker_sleeping() also | |
1420 | * decrements ->nr_running, setting CPU_INTENSIVE here can lead to | |
1421 | * double decrements. The task is releasing the CPU anyway. Let's skip. | |
1422 | * We probably want to make this prettier in the future. | |
616db877 | 1423 | */ |
c8f6219b | 1424 | if ((worker->flags & WORKER_NOT_RUNNING) || READ_ONCE(worker->sleeping) || |
616db877 TH |
1425 | worker->task->se.sum_exec_runtime - worker->current_at < |
1426 | wq_cpu_intensive_thresh_us * NSEC_PER_USEC) | |
1427 | return; | |
1428 | ||
1429 | raw_spin_lock(&pool->lock); | |
1430 | ||
1431 | worker_set_flags(worker, WORKER_CPU_INTENSIVE); | |
63638450 | 1432 | wq_cpu_intensive_report(worker->current_func); |
616db877 TH |
1433 | pwq->stats[PWQ_STAT_CPU_INTENSIVE]++; |
1434 | ||
0219a352 | 1435 | if (kick_pool(pool)) |
616db877 | 1436 | pwq->stats[PWQ_STAT_CM_WAKEUP]++; |
616db877 TH |
1437 | |
1438 | raw_spin_unlock(&pool->lock); | |
1439 | } | |
1440 | ||
1b69ac6b JW |
1441 | /** |
1442 | * wq_worker_last_func - retrieve worker's last work function | |
8194fe94 | 1443 | * @task: Task to retrieve last work function of. |
1b69ac6b JW |
1444 | * |
1445 | * Determine the last function a worker executed. This is called from | |
1446 | * the scheduler to get a worker's last known identity. | |
1447 | * | |
1448 | * CONTEXT: | |
a9b8a985 | 1449 | * raw_spin_lock_irq(rq->lock) |
1b69ac6b | 1450 | * |
4b047002 JW |
1451 | * This function is called during schedule() when a kworker is going |
1452 | * to sleep. It's used by psi to identify aggregation workers during | |
1453 | * dequeuing, to allow periodic aggregation to shut-off when that | |
1454 | * worker is the last task in the system or cgroup to go to sleep. | |
1455 | * | |
1456 | * As this function doesn't involve any workqueue-related locking, it | |
1457 | * only returns stable values when called from inside the scheduler's | |
1458 | * queuing and dequeuing paths, when @task, which must be a kworker, | |
1459 | * is guaranteed to not be processing any works. | |
1460 | * | |
1b69ac6b JW |
1461 | * Return: |
1462 | * The last work function %current executed as a worker, NULL if it | |
1463 | * hasn't executed any work yet. | |
1464 | */ | |
1465 | work_func_t wq_worker_last_func(struct task_struct *task) | |
1466 | { | |
1467 | struct worker *worker = kthread_data(task); | |
1468 | ||
1469 | return worker->last_func; | |
1470 | } | |
1471 | ||
91ccc6e7 TH |
1472 | /** |
1473 | * wq_node_nr_active - Determine wq_node_nr_active to use | |
1474 | * @wq: workqueue of interest | |
1475 | * @node: NUMA node, can be %NUMA_NO_NODE | |
1476 | * | |
1477 | * Determine wq_node_nr_active to use for @wq on @node. Returns: | |
1478 | * | |
1479 | * - %NULL for per-cpu workqueues as they don't need to use shared nr_active. | |
1480 | * | |
1481 | * - node_nr_active[nr_node_ids] if @node is %NUMA_NO_NODE. | |
1482 | * | |
1483 | * - Otherwise, node_nr_active[@node]. | |
1484 | */ | |
1485 | static struct wq_node_nr_active *wq_node_nr_active(struct workqueue_struct *wq, | |
1486 | int node) | |
1487 | { | |
1488 | if (!(wq->flags & WQ_UNBOUND)) | |
1489 | return NULL; | |
1490 | ||
1491 | if (node == NUMA_NO_NODE) | |
1492 | node = nr_node_ids; | |
1493 | ||
1494 | return wq->node_nr_active[node]; | |
1495 | } | |
1496 | ||
5797b1c1 TH |
1497 | /** |
1498 | * wq_update_node_max_active - Update per-node max_actives to use | |
1499 | * @wq: workqueue to update | |
1500 | * @off_cpu: CPU that's going down, -1 if a CPU is not going down | |
1501 | * | |
1502 | * Update @wq->node_nr_active[]->max. @wq must be unbound. max_active is | |
1503 | * distributed among nodes according to the proportions of numbers of online | |
1504 | * cpus. The result is always between @wq->min_active and max_active. | |
1505 | */ | |
1506 | static void wq_update_node_max_active(struct workqueue_struct *wq, int off_cpu) | |
1507 | { | |
1508 | struct cpumask *effective = unbound_effective_cpumask(wq); | |
1509 | int min_active = READ_ONCE(wq->min_active); | |
1510 | int max_active = READ_ONCE(wq->max_active); | |
1511 | int total_cpus, node; | |
1512 | ||
1513 | lockdep_assert_held(&wq->mutex); | |
1514 | ||
c5f8cd6c TH |
1515 | if (!wq_topo_initialized) |
1516 | return; | |
1517 | ||
15930da4 | 1518 | if (off_cpu >= 0 && !cpumask_test_cpu(off_cpu, effective)) |
5797b1c1 TH |
1519 | off_cpu = -1; |
1520 | ||
1521 | total_cpus = cpumask_weight_and(effective, cpu_online_mask); | |
1522 | if (off_cpu >= 0) | |
1523 | total_cpus--; | |
1524 | ||
1525 | for_each_node(node) { | |
1526 | int node_cpus; | |
1527 | ||
1528 | node_cpus = cpumask_weight_and(effective, cpumask_of_node(node)); | |
1529 | if (off_cpu >= 0 && cpu_to_node(off_cpu) == node) | |
1530 | node_cpus--; | |
1531 | ||
1532 | wq_node_nr_active(wq, node)->max = | |
1533 | clamp(DIV_ROUND_UP(max_active * node_cpus, total_cpus), | |
1534 | min_active, max_active); | |
1535 | } | |
1536 | ||
1537 | wq_node_nr_active(wq, NUMA_NO_NODE)->max = min_active; | |
1538 | } | |
1539 | ||
8864b4e5 TH |
1540 | /** |
1541 | * get_pwq - get an extra reference on the specified pool_workqueue | |
1542 | * @pwq: pool_workqueue to get | |
1543 | * | |
1544 | * Obtain an extra reference on @pwq. The caller should guarantee that | |
1545 | * @pwq has positive refcnt and be holding the matching pool->lock. | |
1546 | */ | |
1547 | static void get_pwq(struct pool_workqueue *pwq) | |
1548 | { | |
1549 | lockdep_assert_held(&pwq->pool->lock); | |
1550 | WARN_ON_ONCE(pwq->refcnt <= 0); | |
1551 | pwq->refcnt++; | |
1552 | } | |
1553 | ||
1554 | /** | |
1555 | * put_pwq - put a pool_workqueue reference | |
1556 | * @pwq: pool_workqueue to put | |
1557 | * | |
1558 | * Drop a reference of @pwq. If its refcnt reaches zero, schedule its | |
1559 | * destruction. The caller should be holding the matching pool->lock. | |
1560 | */ | |
1561 | static void put_pwq(struct pool_workqueue *pwq) | |
1562 | { | |
1563 | lockdep_assert_held(&pwq->pool->lock); | |
1564 | if (likely(--pwq->refcnt)) | |
1565 | return; | |
8864b4e5 | 1566 | /* |
967b494e TH |
1567 | * @pwq can't be released under pool->lock, bounce to a dedicated |
1568 | * kthread_worker to avoid A-A deadlocks. | |
8864b4e5 | 1569 | */ |
687a9aa5 | 1570 | kthread_queue_work(pwq_release_worker, &pwq->release_work); |
8864b4e5 TH |
1571 | } |
1572 | ||
dce90d47 TH |
1573 | /** |
1574 | * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock | |
1575 | * @pwq: pool_workqueue to put (can be %NULL) | |
1576 | * | |
1577 | * put_pwq() with locking. This function also allows %NULL @pwq. | |
1578 | */ | |
1579 | static void put_pwq_unlocked(struct pool_workqueue *pwq) | |
1580 | { | |
1581 | if (pwq) { | |
1582 | /* | |
24acfb71 | 1583 | * As both pwqs and pools are RCU protected, the |
dce90d47 TH |
1584 | * following lock operations are safe. |
1585 | */ | |
a9b8a985 | 1586 | raw_spin_lock_irq(&pwq->pool->lock); |
dce90d47 | 1587 | put_pwq(pwq); |
a9b8a985 | 1588 | raw_spin_unlock_irq(&pwq->pool->lock); |
dce90d47 TH |
1589 | } |
1590 | } | |
1591 | ||
afa87ce8 TH |
1592 | static bool pwq_is_empty(struct pool_workqueue *pwq) |
1593 | { | |
1594 | return !pwq->nr_active && list_empty(&pwq->inactive_works); | |
1595 | } | |
1596 | ||
4c638030 TH |
1597 | static void __pwq_activate_work(struct pool_workqueue *pwq, |
1598 | struct work_struct *work) | |
bf4ede01 | 1599 | { |
1c270b79 TH |
1600 | unsigned long *wdb = work_data_bits(work); |
1601 | ||
1602 | WARN_ON_ONCE(!(*wdb & WORK_STRUCT_INACTIVE)); | |
bf4ede01 | 1603 | trace_workqueue_activate_work(work); |
82607adc TH |
1604 | if (list_empty(&pwq->pool->worklist)) |
1605 | pwq->pool->watchdog_ts = jiffies; | |
112202d9 | 1606 | move_linked_works(work, &pwq->pool->worklist, NULL); |
1c270b79 | 1607 | __clear_bit(WORK_STRUCT_INACTIVE_BIT, wdb); |
4c638030 TH |
1608 | } |
1609 | ||
1610 | /** | |
1611 | * pwq_activate_work - Activate a work item if inactive | |
1612 | * @pwq: pool_workqueue @work belongs to | |
1613 | * @work: work item to activate | |
1614 | * | |
1615 | * Returns %true if activated. %false if already active. | |
1616 | */ | |
1617 | static bool pwq_activate_work(struct pool_workqueue *pwq, | |
1618 | struct work_struct *work) | |
1619 | { | |
1620 | struct worker_pool *pool = pwq->pool; | |
91ccc6e7 | 1621 | struct wq_node_nr_active *nna; |
4c638030 TH |
1622 | |
1623 | lockdep_assert_held(&pool->lock); | |
1624 | ||
1625 | if (!(*work_data_bits(work) & WORK_STRUCT_INACTIVE)) | |
1626 | return false; | |
1627 | ||
91ccc6e7 TH |
1628 | nna = wq_node_nr_active(pwq->wq, pool->node); |
1629 | if (nna) | |
1630 | atomic_inc(&nna->nr); | |
1631 | ||
112202d9 | 1632 | pwq->nr_active++; |
4c638030 TH |
1633 | __pwq_activate_work(pwq, work); |
1634 | return true; | |
bf4ede01 TH |
1635 | } |
1636 | ||
5797b1c1 TH |
1637 | static bool tryinc_node_nr_active(struct wq_node_nr_active *nna) |
1638 | { | |
1639 | int max = READ_ONCE(nna->max); | |
1640 | ||
1641 | while (true) { | |
1642 | int old, tmp; | |
1643 | ||
1644 | old = atomic_read(&nna->nr); | |
1645 | if (old >= max) | |
1646 | return false; | |
1647 | tmp = atomic_cmpxchg_relaxed(&nna->nr, old, old + 1); | |
1648 | if (tmp == old) | |
1649 | return true; | |
1650 | } | |
1651 | } | |
1652 | ||
1c270b79 TH |
1653 | /** |
1654 | * pwq_tryinc_nr_active - Try to increment nr_active for a pwq | |
1655 | * @pwq: pool_workqueue of interest | |
5797b1c1 | 1656 | * @fill: max_active may have increased, try to increase concurrency level |
1c270b79 TH |
1657 | * |
1658 | * Try to increment nr_active for @pwq. Returns %true if an nr_active count is | |
1659 | * successfully obtained. %false otherwise. | |
1660 | */ | |
5797b1c1 | 1661 | static bool pwq_tryinc_nr_active(struct pool_workqueue *pwq, bool fill) |
1c270b79 TH |
1662 | { |
1663 | struct workqueue_struct *wq = pwq->wq; | |
1664 | struct worker_pool *pool = pwq->pool; | |
91ccc6e7 | 1665 | struct wq_node_nr_active *nna = wq_node_nr_active(wq, pool->node); |
5797b1c1 | 1666 | bool obtained = false; |
1c270b79 TH |
1667 | |
1668 | lockdep_assert_held(&pool->lock); | |
1669 | ||
5797b1c1 TH |
1670 | if (!nna) { |
1671 | /* per-cpu workqueue, pwq->nr_active is sufficient */ | |
1672 | obtained = pwq->nr_active < READ_ONCE(wq->max_active); | |
1673 | goto out; | |
1674 | } | |
1675 | ||
1676 | /* | |
1677 | * Unbound workqueue uses per-node shared nr_active $nna. If @pwq is | |
1678 | * already waiting on $nna, pwq_dec_nr_active() will maintain the | |
1679 | * concurrency level. Don't jump the line. | |
1680 | * | |
1681 | * We need to ignore the pending test after max_active has increased as | |
1682 | * pwq_dec_nr_active() can only maintain the concurrency level but not | |
1683 | * increase it. This is indicated by @fill. | |
1684 | */ | |
1685 | if (!list_empty(&pwq->pending_node) && likely(!fill)) | |
1686 | goto out; | |
1687 | ||
1688 | obtained = tryinc_node_nr_active(nna); | |
1689 | if (obtained) | |
1690 | goto out; | |
1691 | ||
1692 | /* | |
1693 | * Lockless acquisition failed. Lock, add ourself to $nna->pending_pwqs | |
1694 | * and try again. The smp_mb() is paired with the implied memory barrier | |
1695 | * of atomic_dec_return() in pwq_dec_nr_active() to ensure that either | |
1696 | * we see the decremented $nna->nr or they see non-empty | |
1697 | * $nna->pending_pwqs. | |
1698 | */ | |
1699 | raw_spin_lock(&nna->lock); | |
1700 | ||
1701 | if (list_empty(&pwq->pending_node)) | |
1702 | list_add_tail(&pwq->pending_node, &nna->pending_pwqs); | |
1703 | else if (likely(!fill)) | |
1704 | goto out_unlock; | |
1705 | ||
1706 | smp_mb(); | |
1707 | ||
1708 | obtained = tryinc_node_nr_active(nna); | |
1c270b79 | 1709 | |
5797b1c1 TH |
1710 | /* |
1711 | * If @fill, @pwq might have already been pending. Being spuriously | |
1712 | * pending in cold paths doesn't affect anything. Let's leave it be. | |
1713 | */ | |
1714 | if (obtained && likely(!fill)) | |
1715 | list_del_init(&pwq->pending_node); | |
1716 | ||
1717 | out_unlock: | |
1718 | raw_spin_unlock(&nna->lock); | |
1719 | out: | |
1720 | if (obtained) | |
1c270b79 TH |
1721 | pwq->nr_active++; |
1722 | return obtained; | |
1723 | } | |
1724 | ||
1725 | /** | |
1726 | * pwq_activate_first_inactive - Activate the first inactive work item on a pwq | |
1727 | * @pwq: pool_workqueue of interest | |
5797b1c1 | 1728 | * @fill: max_active may have increased, try to increase concurrency level |
1c270b79 TH |
1729 | * |
1730 | * Activate the first inactive work item of @pwq if available and allowed by | |
1731 | * max_active limit. | |
1732 | * | |
1733 | * Returns %true if an inactive work item has been activated. %false if no | |
1734 | * inactive work item is found or max_active limit is reached. | |
1735 | */ | |
5797b1c1 | 1736 | static bool pwq_activate_first_inactive(struct pool_workqueue *pwq, bool fill) |
1c270b79 TH |
1737 | { |
1738 | struct work_struct *work = | |
1739 | list_first_entry_or_null(&pwq->inactive_works, | |
1740 | struct work_struct, entry); | |
1741 | ||
5797b1c1 | 1742 | if (work && pwq_tryinc_nr_active(pwq, fill)) { |
1c270b79 TH |
1743 | __pwq_activate_work(pwq, work); |
1744 | return true; | |
1745 | } else { | |
1746 | return false; | |
1747 | } | |
1748 | } | |
1749 | ||
5797b1c1 TH |
1750 | /** |
1751 | * node_activate_pending_pwq - Activate a pending pwq on a wq_node_nr_active | |
1752 | * @nna: wq_node_nr_active to activate a pending pwq for | |
1753 | * @caller_pool: worker_pool the caller is locking | |
1754 | * | |
1755 | * Activate a pwq in @nna->pending_pwqs. Called with @caller_pool locked. | |
1756 | * @caller_pool may be unlocked and relocked to lock other worker_pools. | |
1757 | */ | |
1758 | static void node_activate_pending_pwq(struct wq_node_nr_active *nna, | |
1759 | struct worker_pool *caller_pool) | |
1760 | { | |
1761 | struct worker_pool *locked_pool = caller_pool; | |
1762 | struct pool_workqueue *pwq; | |
1763 | struct work_struct *work; | |
1764 | ||
1765 | lockdep_assert_held(&caller_pool->lock); | |
1766 | ||
1767 | raw_spin_lock(&nna->lock); | |
1768 | retry: | |
1769 | pwq = list_first_entry_or_null(&nna->pending_pwqs, | |
1770 | struct pool_workqueue, pending_node); | |
1771 | if (!pwq) | |
1772 | goto out_unlock; | |
1773 | ||
1774 | /* | |
1775 | * If @pwq is for a different pool than @locked_pool, we need to lock | |
1776 | * @pwq->pool->lock. Let's trylock first. If unsuccessful, do the unlock | |
1777 | * / lock dance. For that, we also need to release @nna->lock as it's | |
1778 | * nested inside pool locks. | |
1779 | */ | |
1780 | if (pwq->pool != locked_pool) { | |
1781 | raw_spin_unlock(&locked_pool->lock); | |
1782 | locked_pool = pwq->pool; | |
1783 | if (!raw_spin_trylock(&locked_pool->lock)) { | |
1784 | raw_spin_unlock(&nna->lock); | |
1785 | raw_spin_lock(&locked_pool->lock); | |
1786 | raw_spin_lock(&nna->lock); | |
1787 | goto retry; | |
1788 | } | |
1789 | } | |
1790 | ||
1791 | /* | |
1792 | * $pwq may not have any inactive work items due to e.g. cancellations. | |
1793 | * Drop it from pending_pwqs and see if there's another one. | |
1794 | */ | |
1795 | work = list_first_entry_or_null(&pwq->inactive_works, | |
1796 | struct work_struct, entry); | |
1797 | if (!work) { | |
1798 | list_del_init(&pwq->pending_node); | |
1799 | goto retry; | |
1800 | } | |
1801 | ||
1802 | /* | |
1803 | * Acquire an nr_active count and activate the inactive work item. If | |
1804 | * $pwq still has inactive work items, rotate it to the end of the | |
1805 | * pending_pwqs so that we round-robin through them. This means that | |
1806 | * inactive work items are not activated in queueing order which is fine | |
1807 | * given that there has never been any ordering across different pwqs. | |
1808 | */ | |
1809 | if (likely(tryinc_node_nr_active(nna))) { | |
1810 | pwq->nr_active++; | |
1811 | __pwq_activate_work(pwq, work); | |
1812 | ||
1813 | if (list_empty(&pwq->inactive_works)) | |
1814 | list_del_init(&pwq->pending_node); | |
1815 | else | |
1816 | list_move_tail(&pwq->pending_node, &nna->pending_pwqs); | |
1817 | ||
1818 | /* if activating a foreign pool, make sure it's running */ | |
1819 | if (pwq->pool != caller_pool) | |
1820 | kick_pool(pwq->pool); | |
1821 | } | |
1822 | ||
1823 | out_unlock: | |
1824 | raw_spin_unlock(&nna->lock); | |
1825 | if (locked_pool != caller_pool) { | |
1826 | raw_spin_unlock(&locked_pool->lock); | |
1827 | raw_spin_lock(&caller_pool->lock); | |
1828 | } | |
1829 | } | |
1830 | ||
1c270b79 TH |
1831 | /** |
1832 | * pwq_dec_nr_active - Retire an active count | |
1833 | * @pwq: pool_workqueue of interest | |
1834 | * | |
1835 | * Decrement @pwq's nr_active and try to activate the first inactive work item. | |
5797b1c1 | 1836 | * For unbound workqueues, this function may temporarily drop @pwq->pool->lock. |
1c270b79 TH |
1837 | */ |
1838 | static void pwq_dec_nr_active(struct pool_workqueue *pwq) | |
3aa62497 | 1839 | { |
1c270b79 | 1840 | struct worker_pool *pool = pwq->pool; |
91ccc6e7 | 1841 | struct wq_node_nr_active *nna = wq_node_nr_active(pwq->wq, pool->node); |
3aa62497 | 1842 | |
1c270b79 TH |
1843 | lockdep_assert_held(&pool->lock); |
1844 | ||
91ccc6e7 TH |
1845 | /* |
1846 | * @pwq->nr_active should be decremented for both percpu and unbound | |
1847 | * workqueues. | |
1848 | */ | |
1c270b79 | 1849 | pwq->nr_active--; |
91ccc6e7 TH |
1850 | |
1851 | /* | |
1852 | * For a percpu workqueue, it's simple. Just need to kick the first | |
1853 | * inactive work item on @pwq itself. | |
1854 | */ | |
1855 | if (!nna) { | |
5797b1c1 | 1856 | pwq_activate_first_inactive(pwq, false); |
91ccc6e7 TH |
1857 | return; |
1858 | } | |
1859 | ||
5797b1c1 TH |
1860 | /* |
1861 | * If @pwq is for an unbound workqueue, it's more complicated because | |
1862 | * multiple pwqs and pools may be sharing the nr_active count. When a | |
1863 | * pwq needs to wait for an nr_active count, it puts itself on | |
1864 | * $nna->pending_pwqs. The following atomic_dec_return()'s implied | |
1865 | * memory barrier is paired with smp_mb() in pwq_tryinc_nr_active() to | |
1866 | * guarantee that either we see non-empty pending_pwqs or they see | |
1867 | * decremented $nna->nr. | |
1868 | * | |
1869 | * $nna->max may change as CPUs come online/offline and @pwq->wq's | |
1870 | * max_active gets updated. However, it is guaranteed to be equal to or | |
1871 | * larger than @pwq->wq->min_active which is above zero unless freezing. | |
1872 | * This maintains the forward progress guarantee. | |
1873 | */ | |
1874 | if (atomic_dec_return(&nna->nr) >= READ_ONCE(nna->max)) | |
1875 | return; | |
1876 | ||
1877 | if (!list_empty(&nna->pending_pwqs)) | |
1878 | node_activate_pending_pwq(nna, pool); | |
3aa62497 LJ |
1879 | } |
1880 | ||
bf4ede01 | 1881 | /** |
112202d9 TH |
1882 | * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight |
1883 | * @pwq: pwq of interest | |
c4560c2c | 1884 | * @work_data: work_data of work which left the queue |
bf4ede01 TH |
1885 | * |
1886 | * A work either has completed or is removed from pending queue, | |
112202d9 | 1887 | * decrement nr_in_flight of its pwq and handle workqueue flushing. |
bf4ede01 | 1888 | * |
dd6c3c54 TH |
1889 | * NOTE: |
1890 | * For unbound workqueues, this function may temporarily drop @pwq->pool->lock | |
1891 | * and thus should be called after all other state updates for the in-flight | |
1892 | * work item is complete. | |
1893 | * | |
bf4ede01 | 1894 | * CONTEXT: |
a9b8a985 | 1895 | * raw_spin_lock_irq(pool->lock). |
bf4ede01 | 1896 | */ |
c4560c2c | 1897 | static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, unsigned long work_data) |
bf4ede01 | 1898 | { |
c4560c2c LJ |
1899 | int color = get_work_color(work_data); |
1900 | ||
1c270b79 TH |
1901 | if (!(work_data & WORK_STRUCT_INACTIVE)) |
1902 | pwq_dec_nr_active(pwq); | |
018f3a13 | 1903 | |
112202d9 | 1904 | pwq->nr_in_flight[color]--; |
bf4ede01 | 1905 | |
bf4ede01 | 1906 | /* is flush in progress and are we at the flushing tip? */ |
112202d9 | 1907 | if (likely(pwq->flush_color != color)) |
8864b4e5 | 1908 | goto out_put; |
bf4ede01 TH |
1909 | |
1910 | /* are there still in-flight works? */ | |
112202d9 | 1911 | if (pwq->nr_in_flight[color]) |
8864b4e5 | 1912 | goto out_put; |
bf4ede01 | 1913 | |
112202d9 TH |
1914 | /* this pwq is done, clear flush_color */ |
1915 | pwq->flush_color = -1; | |
bf4ede01 TH |
1916 | |
1917 | /* | |
112202d9 | 1918 | * If this was the last pwq, wake up the first flusher. It |
bf4ede01 TH |
1919 | * will handle the rest. |
1920 | */ | |
112202d9 TH |
1921 | if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) |
1922 | complete(&pwq->wq->first_flusher->done); | |
8864b4e5 TH |
1923 | out_put: |
1924 | put_pwq(pwq); | |
bf4ede01 TH |
1925 | } |
1926 | ||
36e227d2 | 1927 | /** |
bbb68dfa | 1928 | * try_to_grab_pending - steal work item from worklist and disable irq |
36e227d2 TH |
1929 | * @work: work item to steal |
1930 | * @is_dwork: @work is a delayed_work | |
bbb68dfa | 1931 | * @flags: place to store irq state |
36e227d2 TH |
1932 | * |
1933 | * Try to grab PENDING bit of @work. This function can handle @work in any | |
d185af30 | 1934 | * stable state - idle, on timer or on worklist. |
36e227d2 | 1935 | * |
d185af30 | 1936 | * Return: |
3eb6b31b MCC |
1937 | * |
1938 | * ======== ================================================================ | |
36e227d2 TH |
1939 | * 1 if @work was pending and we successfully stole PENDING |
1940 | * 0 if @work was idle and we claimed PENDING | |
1941 | * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry | |
bbb68dfa TH |
1942 | * -ENOENT if someone else is canceling @work, this state may persist |
1943 | * for arbitrarily long | |
3eb6b31b | 1944 | * ======== ================================================================ |
36e227d2 | 1945 | * |
d185af30 | 1946 | * Note: |
bbb68dfa | 1947 | * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting |
e0aecdd8 TH |
1948 | * interrupted while holding PENDING and @work off queue, irq must be |
1949 | * disabled on entry. This, combined with delayed_work->timer being | |
1950 | * irqsafe, ensures that we return -EAGAIN for finite short period of time. | |
bbb68dfa TH |
1951 | * |
1952 | * On successful return, >= 0, irq is disabled and the caller is | |
1953 | * responsible for releasing it using local_irq_restore(*@flags). | |
1954 | * | |
e0aecdd8 | 1955 | * This function is safe to call from any context including IRQ handler. |
bf4ede01 | 1956 | */ |
bbb68dfa TH |
1957 | static int try_to_grab_pending(struct work_struct *work, bool is_dwork, |
1958 | unsigned long *flags) | |
bf4ede01 | 1959 | { |
d565ed63 | 1960 | struct worker_pool *pool; |
112202d9 | 1961 | struct pool_workqueue *pwq; |
bf4ede01 | 1962 | |
bbb68dfa TH |
1963 | local_irq_save(*flags); |
1964 | ||
36e227d2 TH |
1965 | /* try to steal the timer if it exists */ |
1966 | if (is_dwork) { | |
1967 | struct delayed_work *dwork = to_delayed_work(work); | |
1968 | ||
e0aecdd8 TH |
1969 | /* |
1970 | * dwork->timer is irqsafe. If del_timer() fails, it's | |
1971 | * guaranteed that the timer is not queued anywhere and not | |
1972 | * running on the local CPU. | |
1973 | */ | |
36e227d2 TH |
1974 | if (likely(del_timer(&dwork->timer))) |
1975 | return 1; | |
1976 | } | |
1977 | ||
1978 | /* try to claim PENDING the normal way */ | |
bf4ede01 TH |
1979 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) |
1980 | return 0; | |
1981 | ||
24acfb71 | 1982 | rcu_read_lock(); |
bf4ede01 TH |
1983 | /* |
1984 | * The queueing is in progress, or it is already queued. Try to | |
1985 | * steal it from ->worklist without clearing WORK_STRUCT_PENDING. | |
1986 | */ | |
d565ed63 TH |
1987 | pool = get_work_pool(work); |
1988 | if (!pool) | |
bbb68dfa | 1989 | goto fail; |
bf4ede01 | 1990 | |
a9b8a985 | 1991 | raw_spin_lock(&pool->lock); |
0b3dae68 | 1992 | /* |
112202d9 TH |
1993 | * work->data is guaranteed to point to pwq only while the work |
1994 | * item is queued on pwq->wq, and both updating work->data to point | |
1995 | * to pwq on queueing and to pool on dequeueing are done under | |
1996 | * pwq->pool->lock. This in turn guarantees that, if work->data | |
1997 | * points to pwq which is associated with a locked pool, the work | |
0b3dae68 LJ |
1998 | * item is currently queued on that pool. |
1999 | */ | |
112202d9 TH |
2000 | pwq = get_work_pwq(work); |
2001 | if (pwq && pwq->pool == pool) { | |
c70e1779 TH |
2002 | unsigned long work_data; |
2003 | ||
16062836 TH |
2004 | debug_work_deactivate(work); |
2005 | ||
2006 | /* | |
018f3a13 LJ |
2007 | * A cancelable inactive work item must be in the |
2008 | * pwq->inactive_works since a queued barrier can't be | |
2009 | * canceled (see the comments in insert_wq_barrier()). | |
2010 | * | |
f97a4a1a | 2011 | * An inactive work item cannot be grabbed directly because |
d812796e | 2012 | * it might have linked barrier work items which, if left |
f97a4a1a | 2013 | * on the inactive_works list, will confuse pwq->nr_active |
16062836 TH |
2014 | * management later on and cause stall. Make sure the work |
2015 | * item is activated before grabbing. | |
2016 | */ | |
4c638030 | 2017 | pwq_activate_work(pwq, work); |
16062836 TH |
2018 | |
2019 | list_del_init(&work->entry); | |
16062836 | 2020 | |
c70e1779 TH |
2021 | /* |
2022 | * work->data points to pwq iff queued. Let's point to pool. As | |
2023 | * this destroys work->data needed by the next step, stash it. | |
2024 | */ | |
2025 | work_data = *work_data_bits(work); | |
16062836 TH |
2026 | set_work_pool_and_keep_pending(work, pool->id); |
2027 | ||
dd6c3c54 | 2028 | /* must be the last step, see the function comment */ |
c70e1779 | 2029 | pwq_dec_nr_in_flight(pwq, work_data); |
dd6c3c54 | 2030 | |
a9b8a985 | 2031 | raw_spin_unlock(&pool->lock); |
24acfb71 | 2032 | rcu_read_unlock(); |
16062836 | 2033 | return 1; |
bf4ede01 | 2034 | } |
a9b8a985 | 2035 | raw_spin_unlock(&pool->lock); |
bbb68dfa | 2036 | fail: |
24acfb71 | 2037 | rcu_read_unlock(); |
bbb68dfa TH |
2038 | local_irq_restore(*flags); |
2039 | if (work_is_canceling(work)) | |
2040 | return -ENOENT; | |
2041 | cpu_relax(); | |
36e227d2 | 2042 | return -EAGAIN; |
bf4ede01 TH |
2043 | } |
2044 | ||
4690c4ab | 2045 | /** |
706026c2 | 2046 | * insert_work - insert a work into a pool |
112202d9 | 2047 | * @pwq: pwq @work belongs to |
4690c4ab TH |
2048 | * @work: work to insert |
2049 | * @head: insertion point | |
2050 | * @extra_flags: extra WORK_STRUCT_* flags to set | |
2051 | * | |
112202d9 | 2052 | * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to |
706026c2 | 2053 | * work_struct flags. |
4690c4ab TH |
2054 | * |
2055 | * CONTEXT: | |
a9b8a985 | 2056 | * raw_spin_lock_irq(pool->lock). |
4690c4ab | 2057 | */ |
112202d9 TH |
2058 | static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, |
2059 | struct list_head *head, unsigned int extra_flags) | |
b89deed3 | 2060 | { |
fe089f87 | 2061 | debug_work_activate(work); |
e22bee78 | 2062 | |
e89a85d6 | 2063 | /* record the work call stack in order to print it in KASAN reports */ |
f70da745 | 2064 | kasan_record_aux_stack_noalloc(work); |
e89a85d6 | 2065 | |
4690c4ab | 2066 | /* we own @work, set data and link */ |
112202d9 | 2067 | set_work_pwq(work, pwq, extra_flags); |
1a4d9b0a | 2068 | list_add_tail(&work->entry, head); |
8864b4e5 | 2069 | get_pwq(pwq); |
b89deed3 ON |
2070 | } |
2071 | ||
c8efcc25 TH |
2072 | /* |
2073 | * Test whether @work is being queued from another work executing on the | |
8d03ecfe | 2074 | * same workqueue. |
c8efcc25 TH |
2075 | */ |
2076 | static bool is_chained_work(struct workqueue_struct *wq) | |
2077 | { | |
8d03ecfe TH |
2078 | struct worker *worker; |
2079 | ||
2080 | worker = current_wq_worker(); | |
2081 | /* | |
bf393fd4 | 2082 | * Return %true iff I'm a worker executing a work item on @wq. If |
8d03ecfe TH |
2083 | * I'm @worker, it's safe to dereference it without locking. |
2084 | */ | |
112202d9 | 2085 | return worker && worker->current_pwq->wq == wq; |
c8efcc25 TH |
2086 | } |
2087 | ||
ef557180 MG |
2088 | /* |
2089 | * When queueing an unbound work item to a wq, prefer local CPU if allowed | |
2090 | * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to | |
2091 | * avoid perturbing sensitive tasks. | |
2092 | */ | |
2093 | static int wq_select_unbound_cpu(int cpu) | |
2094 | { | |
2095 | int new_cpu; | |
2096 | ||
f303fccb TH |
2097 | if (likely(!wq_debug_force_rr_cpu)) { |
2098 | if (cpumask_test_cpu(cpu, wq_unbound_cpumask)) | |
2099 | return cpu; | |
a8ec5880 AF |
2100 | } else { |
2101 | pr_warn_once("workqueue: round-robin CPU selection forced, expect performance impact\n"); | |
f303fccb TH |
2102 | } |
2103 | ||
ef557180 MG |
2104 | new_cpu = __this_cpu_read(wq_rr_cpu_last); |
2105 | new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask); | |
2106 | if (unlikely(new_cpu >= nr_cpu_ids)) { | |
2107 | new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask); | |
2108 | if (unlikely(new_cpu >= nr_cpu_ids)) | |
2109 | return cpu; | |
2110 | } | |
2111 | __this_cpu_write(wq_rr_cpu_last, new_cpu); | |
2112 | ||
2113 | return new_cpu; | |
2114 | } | |
2115 | ||
d84ff051 | 2116 | static void __queue_work(int cpu, struct workqueue_struct *wq, |
1da177e4 LT |
2117 | struct work_struct *work) |
2118 | { | |
112202d9 | 2119 | struct pool_workqueue *pwq; |
fe089f87 | 2120 | struct worker_pool *last_pool, *pool; |
8a2e8e5d | 2121 | unsigned int work_flags; |
b75cac93 | 2122 | unsigned int req_cpu = cpu; |
8930caba TH |
2123 | |
2124 | /* | |
2125 | * While a work item is PENDING && off queue, a task trying to | |
2126 | * steal the PENDING will busy-loop waiting for it to either get | |
2127 | * queued or lose PENDING. Grabbing PENDING and queueing should | |
2128 | * happen with IRQ disabled. | |
2129 | */ | |
8e8eb730 | 2130 | lockdep_assert_irqs_disabled(); |
1da177e4 | 2131 | |
1e19ffc6 | 2132 | |
33e3f0a3 RC |
2133 | /* |
2134 | * For a draining wq, only works from the same workqueue are | |
2135 | * allowed. The __WQ_DESTROYING helps to spot the issue that | |
2136 | * queues a new work item to a wq after destroy_workqueue(wq). | |
2137 | */ | |
2138 | if (unlikely(wq->flags & (__WQ_DESTROYING | __WQ_DRAINING) && | |
2139 | WARN_ON_ONCE(!is_chained_work(wq)))) | |
e41e704b | 2140 | return; |
24acfb71 | 2141 | rcu_read_lock(); |
9e8cd2f5 | 2142 | retry: |
c9178087 | 2143 | /* pwq which will be used unless @work is executing elsewhere */ |
636b927e TH |
2144 | if (req_cpu == WORK_CPU_UNBOUND) { |
2145 | if (wq->flags & WQ_UNBOUND) | |
aa202f1f | 2146 | cpu = wq_select_unbound_cpu(raw_smp_processor_id()); |
636b927e | 2147 | else |
aa202f1f | 2148 | cpu = raw_smp_processor_id(); |
aa202f1f | 2149 | } |
dbf2576e | 2150 | |
636b927e | 2151 | pwq = rcu_dereference(*per_cpu_ptr(wq->cpu_pwq, cpu)); |
fe089f87 TH |
2152 | pool = pwq->pool; |
2153 | ||
c9178087 TH |
2154 | /* |
2155 | * If @work was previously on a different pool, it might still be | |
2156 | * running there, in which case the work needs to be queued on that | |
2157 | * pool to guarantee non-reentrancy. | |
2158 | */ | |
2159 | last_pool = get_work_pool(work); | |
fe089f87 | 2160 | if (last_pool && last_pool != pool) { |
c9178087 | 2161 | struct worker *worker; |
18aa9eff | 2162 | |
a9b8a985 | 2163 | raw_spin_lock(&last_pool->lock); |
18aa9eff | 2164 | |
c9178087 | 2165 | worker = find_worker_executing_work(last_pool, work); |
18aa9eff | 2166 | |
c9178087 TH |
2167 | if (worker && worker->current_pwq->wq == wq) { |
2168 | pwq = worker->current_pwq; | |
fe089f87 TH |
2169 | pool = pwq->pool; |
2170 | WARN_ON_ONCE(pool != last_pool); | |
8930caba | 2171 | } else { |
c9178087 | 2172 | /* meh... not running there, queue here */ |
a9b8a985 | 2173 | raw_spin_unlock(&last_pool->lock); |
fe089f87 | 2174 | raw_spin_lock(&pool->lock); |
8930caba | 2175 | } |
f3421797 | 2176 | } else { |
fe089f87 | 2177 | raw_spin_lock(&pool->lock); |
502ca9d8 TH |
2178 | } |
2179 | ||
9e8cd2f5 | 2180 | /* |
636b927e TH |
2181 | * pwq is determined and locked. For unbound pools, we could have raced |
2182 | * with pwq release and it could already be dead. If its refcnt is zero, | |
2183 | * repeat pwq selection. Note that unbound pwqs never die without | |
2184 | * another pwq replacing it in cpu_pwq or while work items are executing | |
2185 | * on it, so the retrying is guaranteed to make forward-progress. | |
9e8cd2f5 TH |
2186 | */ |
2187 | if (unlikely(!pwq->refcnt)) { | |
2188 | if (wq->flags & WQ_UNBOUND) { | |
fe089f87 | 2189 | raw_spin_unlock(&pool->lock); |
9e8cd2f5 TH |
2190 | cpu_relax(); |
2191 | goto retry; | |
2192 | } | |
2193 | /* oops */ | |
2194 | WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", | |
2195 | wq->name, cpu); | |
2196 | } | |
2197 | ||
112202d9 TH |
2198 | /* pwq determined, queue */ |
2199 | trace_workqueue_queue_work(req_cpu, pwq, work); | |
502ca9d8 | 2200 | |
24acfb71 TG |
2201 | if (WARN_ON(!list_empty(&work->entry))) |
2202 | goto out; | |
1e19ffc6 | 2203 | |
112202d9 TH |
2204 | pwq->nr_in_flight[pwq->work_color]++; |
2205 | work_flags = work_color_to_flags(pwq->work_color); | |
1e19ffc6 | 2206 | |
a045a272 TH |
2207 | /* |
2208 | * Limit the number of concurrently active work items to max_active. | |
2209 | * @work must also queue behind existing inactive work items to maintain | |
2210 | * ordering when max_active changes. See wq_adjust_max_active(). | |
2211 | */ | |
5797b1c1 | 2212 | if (list_empty(&pwq->inactive_works) && pwq_tryinc_nr_active(pwq, false)) { |
fe089f87 TH |
2213 | if (list_empty(&pool->worklist)) |
2214 | pool->watchdog_ts = jiffies; | |
2215 | ||
cdadf009 | 2216 | trace_workqueue_activate_work(work); |
fe089f87 | 2217 | insert_work(pwq, work, &pool->worklist, work_flags); |
0219a352 | 2218 | kick_pool(pool); |
8a2e8e5d | 2219 | } else { |
f97a4a1a | 2220 | work_flags |= WORK_STRUCT_INACTIVE; |
fe089f87 | 2221 | insert_work(pwq, work, &pwq->inactive_works, work_flags); |
8a2e8e5d | 2222 | } |
1e19ffc6 | 2223 | |
24acfb71 | 2224 | out: |
fe089f87 | 2225 | raw_spin_unlock(&pool->lock); |
24acfb71 | 2226 | rcu_read_unlock(); |
1da177e4 LT |
2227 | } |
2228 | ||
0fcb78c2 | 2229 | /** |
c1a220e7 ZR |
2230 | * queue_work_on - queue work on specific cpu |
2231 | * @cpu: CPU number to execute work on | |
0fcb78c2 REB |
2232 | * @wq: workqueue to use |
2233 | * @work: work to queue | |
2234 | * | |
c1a220e7 | 2235 | * We queue the work to a specific CPU, the caller must ensure it |
443378f0 PM |
2236 | * can't go away. Callers that fail to ensure that the specified |
2237 | * CPU cannot go away will execute on a randomly chosen CPU. | |
854f5cc5 PM |
2238 | * But note well that callers specifying a CPU that never has been |
2239 | * online will get a splat. | |
d185af30 YB |
2240 | * |
2241 | * Return: %false if @work was already on a queue, %true otherwise. | |
1da177e4 | 2242 | */ |
d4283e93 TH |
2243 | bool queue_work_on(int cpu, struct workqueue_struct *wq, |
2244 | struct work_struct *work) | |
1da177e4 | 2245 | { |
d4283e93 | 2246 | bool ret = false; |
8930caba | 2247 | unsigned long flags; |
ef1ca236 | 2248 | |
8930caba | 2249 | local_irq_save(flags); |
c1a220e7 | 2250 | |
22df02bb | 2251 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { |
4690c4ab | 2252 | __queue_work(cpu, wq, work); |
d4283e93 | 2253 | ret = true; |
c1a220e7 | 2254 | } |
ef1ca236 | 2255 | |
8930caba | 2256 | local_irq_restore(flags); |
1da177e4 LT |
2257 | return ret; |
2258 | } | |
ad7b1f84 | 2259 | EXPORT_SYMBOL(queue_work_on); |
1da177e4 | 2260 | |
8204e0c1 | 2261 | /** |
fef59c9c | 2262 | * select_numa_node_cpu - Select a CPU based on NUMA node |
8204e0c1 AD |
2263 | * @node: NUMA node ID that we want to select a CPU from |
2264 | * | |
2265 | * This function will attempt to find a "random" cpu available on a given | |
2266 | * node. If there are no CPUs available on the given node it will return | |
2267 | * WORK_CPU_UNBOUND indicating that we should just schedule to any | |
2268 | * available CPU if we need to schedule this work. | |
2269 | */ | |
fef59c9c | 2270 | static int select_numa_node_cpu(int node) |
8204e0c1 AD |
2271 | { |
2272 | int cpu; | |
2273 | ||
8204e0c1 AD |
2274 | /* Delay binding to CPU if node is not valid or online */ |
2275 | if (node < 0 || node >= MAX_NUMNODES || !node_online(node)) | |
2276 | return WORK_CPU_UNBOUND; | |
2277 | ||
2278 | /* Use local node/cpu if we are already there */ | |
2279 | cpu = raw_smp_processor_id(); | |
2280 | if (node == cpu_to_node(cpu)) | |
2281 | return cpu; | |
2282 | ||
2283 | /* Use "random" otherwise know as "first" online CPU of node */ | |
2284 | cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask); | |
2285 | ||
2286 | /* If CPU is valid return that, otherwise just defer */ | |
2287 | return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND; | |
2288 | } | |
2289 | ||
2290 | /** | |
2291 | * queue_work_node - queue work on a "random" cpu for a given NUMA node | |
2292 | * @node: NUMA node that we are targeting the work for | |
2293 | * @wq: workqueue to use | |
2294 | * @work: work to queue | |
2295 | * | |
2296 | * We queue the work to a "random" CPU within a given NUMA node. The basic | |
2297 | * idea here is to provide a way to somehow associate work with a given | |
2298 | * NUMA node. | |
2299 | * | |
2300 | * This function will only make a best effort attempt at getting this onto | |
2301 | * the right NUMA node. If no node is requested or the requested node is | |
2302 | * offline then we just fall back to standard queue_work behavior. | |
2303 | * | |
2304 | * Currently the "random" CPU ends up being the first available CPU in the | |
2305 | * intersection of cpu_online_mask and the cpumask of the node, unless we | |
2306 | * are running on the node. In that case we just use the current CPU. | |
2307 | * | |
2308 | * Return: %false if @work was already on a queue, %true otherwise. | |
2309 | */ | |
2310 | bool queue_work_node(int node, struct workqueue_struct *wq, | |
2311 | struct work_struct *work) | |
2312 | { | |
2313 | unsigned long flags; | |
2314 | bool ret = false; | |
2315 | ||
2316 | /* | |
2317 | * This current implementation is specific to unbound workqueues. | |
2318 | * Specifically we only return the first available CPU for a given | |
2319 | * node instead of cycling through individual CPUs within the node. | |
2320 | * | |
2321 | * If this is used with a per-cpu workqueue then the logic in | |
2322 | * workqueue_select_cpu_near would need to be updated to allow for | |
2323 | * some round robin type logic. | |
2324 | */ | |
2325 | WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)); | |
2326 | ||
2327 | local_irq_save(flags); | |
2328 | ||
2329 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { | |
fef59c9c | 2330 | int cpu = select_numa_node_cpu(node); |
8204e0c1 AD |
2331 | |
2332 | __queue_work(cpu, wq, work); | |
2333 | ret = true; | |
2334 | } | |
2335 | ||
2336 | local_irq_restore(flags); | |
2337 | return ret; | |
2338 | } | |
2339 | EXPORT_SYMBOL_GPL(queue_work_node); | |
2340 | ||
8c20feb6 | 2341 | void delayed_work_timer_fn(struct timer_list *t) |
1da177e4 | 2342 | { |
8c20feb6 | 2343 | struct delayed_work *dwork = from_timer(dwork, t, timer); |
1da177e4 | 2344 | |
e0aecdd8 | 2345 | /* should have been called from irqsafe timer with irq already off */ |
60c057bc | 2346 | __queue_work(dwork->cpu, dwork->wq, &dwork->work); |
1da177e4 | 2347 | } |
1438ade5 | 2348 | EXPORT_SYMBOL(delayed_work_timer_fn); |
1da177e4 | 2349 | |
7beb2edf TH |
2350 | static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, |
2351 | struct delayed_work *dwork, unsigned long delay) | |
1da177e4 | 2352 | { |
7beb2edf TH |
2353 | struct timer_list *timer = &dwork->timer; |
2354 | struct work_struct *work = &dwork->work; | |
7beb2edf | 2355 | |
637fdbae | 2356 | WARN_ON_ONCE(!wq); |
4b243563 | 2357 | WARN_ON_ONCE(timer->function != delayed_work_timer_fn); |
fc4b514f TH |
2358 | WARN_ON_ONCE(timer_pending(timer)); |
2359 | WARN_ON_ONCE(!list_empty(&work->entry)); | |
7beb2edf | 2360 | |
8852aac2 TH |
2361 | /* |
2362 | * If @delay is 0, queue @dwork->work immediately. This is for | |
2363 | * both optimization and correctness. The earliest @timer can | |
2364 | * expire is on the closest next tick and delayed_work users depend | |
2365 | * on that there's no such delay when @delay is 0. | |
2366 | */ | |
2367 | if (!delay) { | |
2368 | __queue_work(cpu, wq, &dwork->work); | |
2369 | return; | |
2370 | } | |
2371 | ||
60c057bc | 2372 | dwork->wq = wq; |
1265057f | 2373 | dwork->cpu = cpu; |
7beb2edf TH |
2374 | timer->expires = jiffies + delay; |
2375 | ||
aae17ebb LB |
2376 | if (housekeeping_enabled(HK_TYPE_TIMER)) { |
2377 | /* If the current cpu is a housekeeping cpu, use it. */ | |
2378 | cpu = smp_processor_id(); | |
2379 | if (!housekeeping_test_cpu(cpu, HK_TYPE_TIMER)) | |
2380 | cpu = housekeeping_any_cpu(HK_TYPE_TIMER); | |
041bd12e | 2381 | add_timer_on(timer, cpu); |
aae17ebb LB |
2382 | } else { |
2383 | if (likely(cpu == WORK_CPU_UNBOUND)) | |
2384 | add_timer(timer); | |
2385 | else | |
2386 | add_timer_on(timer, cpu); | |
2387 | } | |
1da177e4 LT |
2388 | } |
2389 | ||
0fcb78c2 REB |
2390 | /** |
2391 | * queue_delayed_work_on - queue work on specific CPU after delay | |
2392 | * @cpu: CPU number to execute work on | |
2393 | * @wq: workqueue to use | |
af9997e4 | 2394 | * @dwork: work to queue |
0fcb78c2 REB |
2395 | * @delay: number of jiffies to wait before queueing |
2396 | * | |
d185af30 | 2397 | * Return: %false if @work was already on a queue, %true otherwise. If |
715f1300 TH |
2398 | * @delay is zero and @dwork is idle, it will be scheduled for immediate |
2399 | * execution. | |
0fcb78c2 | 2400 | */ |
d4283e93 TH |
2401 | bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, |
2402 | struct delayed_work *dwork, unsigned long delay) | |
7a6bc1cd | 2403 | { |
52bad64d | 2404 | struct work_struct *work = &dwork->work; |
d4283e93 | 2405 | bool ret = false; |
8930caba | 2406 | unsigned long flags; |
7a6bc1cd | 2407 | |
8930caba TH |
2408 | /* read the comment in __queue_work() */ |
2409 | local_irq_save(flags); | |
7a6bc1cd | 2410 | |
22df02bb | 2411 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { |
7beb2edf | 2412 | __queue_delayed_work(cpu, wq, dwork, delay); |
d4283e93 | 2413 | ret = true; |
7a6bc1cd | 2414 | } |
8a3e77cc | 2415 | |
8930caba | 2416 | local_irq_restore(flags); |
7a6bc1cd VP |
2417 | return ret; |
2418 | } | |
ad7b1f84 | 2419 | EXPORT_SYMBOL(queue_delayed_work_on); |
c7fc77f7 | 2420 | |
8376fe22 TH |
2421 | /** |
2422 | * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU | |
2423 | * @cpu: CPU number to execute work on | |
2424 | * @wq: workqueue to use | |
2425 | * @dwork: work to queue | |
2426 | * @delay: number of jiffies to wait before queueing | |
2427 | * | |
2428 | * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, | |
2429 | * modify @dwork's timer so that it expires after @delay. If @delay is | |
2430 | * zero, @work is guaranteed to be scheduled immediately regardless of its | |
2431 | * current state. | |
2432 | * | |
d185af30 | 2433 | * Return: %false if @dwork was idle and queued, %true if @dwork was |
8376fe22 TH |
2434 | * pending and its timer was modified. |
2435 | * | |
e0aecdd8 | 2436 | * This function is safe to call from any context including IRQ handler. |
8376fe22 TH |
2437 | * See try_to_grab_pending() for details. |
2438 | */ | |
2439 | bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, | |
2440 | struct delayed_work *dwork, unsigned long delay) | |
2441 | { | |
2442 | unsigned long flags; | |
2443 | int ret; | |
c7fc77f7 | 2444 | |
8376fe22 TH |
2445 | do { |
2446 | ret = try_to_grab_pending(&dwork->work, true, &flags); | |
2447 | } while (unlikely(ret == -EAGAIN)); | |
63bc0362 | 2448 | |
8376fe22 TH |
2449 | if (likely(ret >= 0)) { |
2450 | __queue_delayed_work(cpu, wq, dwork, delay); | |
2451 | local_irq_restore(flags); | |
7a6bc1cd | 2452 | } |
8376fe22 TH |
2453 | |
2454 | /* -ENOENT from try_to_grab_pending() becomes %true */ | |
7a6bc1cd VP |
2455 | return ret; |
2456 | } | |
8376fe22 TH |
2457 | EXPORT_SYMBOL_GPL(mod_delayed_work_on); |
2458 | ||
05f0fe6b TH |
2459 | static void rcu_work_rcufn(struct rcu_head *rcu) |
2460 | { | |
2461 | struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu); | |
2462 | ||
2463 | /* read the comment in __queue_work() */ | |
2464 | local_irq_disable(); | |
2465 | __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work); | |
2466 | local_irq_enable(); | |
2467 | } | |
2468 | ||
2469 | /** | |
2470 | * queue_rcu_work - queue work after a RCU grace period | |
2471 | * @wq: workqueue to use | |
2472 | * @rwork: work to queue | |
2473 | * | |
2474 | * Return: %false if @rwork was already pending, %true otherwise. Note | |
2475 | * that a full RCU grace period is guaranteed only after a %true return. | |
bf393fd4 | 2476 | * While @rwork is guaranteed to be executed after a %false return, the |
05f0fe6b TH |
2477 | * execution may happen before a full RCU grace period has passed. |
2478 | */ | |
2479 | bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork) | |
2480 | { | |
2481 | struct work_struct *work = &rwork->work; | |
2482 | ||
2483 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { | |
2484 | rwork->wq = wq; | |
a7e30c0e | 2485 | call_rcu_hurry(&rwork->rcu, rcu_work_rcufn); |
05f0fe6b TH |
2486 | return true; |
2487 | } | |
2488 | ||
2489 | return false; | |
2490 | } | |
2491 | EXPORT_SYMBOL(queue_rcu_work); | |
2492 | ||
f7537df5 | 2493 | static struct worker *alloc_worker(int node) |
c34056a3 TH |
2494 | { |
2495 | struct worker *worker; | |
2496 | ||
f7537df5 | 2497 | worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node); |
c8e55f36 TH |
2498 | if (worker) { |
2499 | INIT_LIST_HEAD(&worker->entry); | |
affee4b2 | 2500 | INIT_LIST_HEAD(&worker->scheduled); |
da028469 | 2501 | INIT_LIST_HEAD(&worker->node); |
e22bee78 TH |
2502 | /* on creation a worker is in !idle && prep state */ |
2503 | worker->flags = WORKER_PREP; | |
c8e55f36 | 2504 | } |
c34056a3 TH |
2505 | return worker; |
2506 | } | |
2507 | ||
9546b29e TH |
2508 | static cpumask_t *pool_allowed_cpus(struct worker_pool *pool) |
2509 | { | |
8639eceb TH |
2510 | if (pool->cpu < 0 && pool->attrs->affn_strict) |
2511 | return pool->attrs->__pod_cpumask; | |
2512 | else | |
2513 | return pool->attrs->cpumask; | |
9546b29e TH |
2514 | } |
2515 | ||
4736cbf7 LJ |
2516 | /** |
2517 | * worker_attach_to_pool() - attach a worker to a pool | |
2518 | * @worker: worker to be attached | |
2519 | * @pool: the target pool | |
2520 | * | |
2521 | * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and | |
2522 | * cpu-binding of @worker are kept coordinated with the pool across | |
2523 | * cpu-[un]hotplugs. | |
2524 | */ | |
2525 | static void worker_attach_to_pool(struct worker *worker, | |
2526 | struct worker_pool *pool) | |
2527 | { | |
1258fae7 | 2528 | mutex_lock(&wq_pool_attach_mutex); |
4736cbf7 | 2529 | |
4736cbf7 | 2530 | /* |
1258fae7 TH |
2531 | * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains |
2532 | * stable across this function. See the comments above the flag | |
2533 | * definition for details. | |
4736cbf7 LJ |
2534 | */ |
2535 | if (pool->flags & POOL_DISASSOCIATED) | |
2536 | worker->flags |= WORKER_UNBOUND; | |
5c25b5ff PZ |
2537 | else |
2538 | kthread_set_per_cpu(worker->task, pool->cpu); | |
4736cbf7 | 2539 | |
640f17c8 | 2540 | if (worker->rescue_wq) |
9546b29e | 2541 | set_cpus_allowed_ptr(worker->task, pool_allowed_cpus(pool)); |
640f17c8 | 2542 | |
4736cbf7 | 2543 | list_add_tail(&worker->node, &pool->workers); |
a2d812a2 | 2544 | worker->pool = pool; |
4736cbf7 | 2545 | |
1258fae7 | 2546 | mutex_unlock(&wq_pool_attach_mutex); |
4736cbf7 LJ |
2547 | } |
2548 | ||
60f5a4bc LJ |
2549 | /** |
2550 | * worker_detach_from_pool() - detach a worker from its pool | |
2551 | * @worker: worker which is attached to its pool | |
60f5a4bc | 2552 | * |
4736cbf7 LJ |
2553 | * Undo the attaching which had been done in worker_attach_to_pool(). The |
2554 | * caller worker shouldn't access to the pool after detached except it has | |
2555 | * other reference to the pool. | |
60f5a4bc | 2556 | */ |
a2d812a2 | 2557 | static void worker_detach_from_pool(struct worker *worker) |
60f5a4bc | 2558 | { |
a2d812a2 | 2559 | struct worker_pool *pool = worker->pool; |
60f5a4bc LJ |
2560 | struct completion *detach_completion = NULL; |
2561 | ||
1258fae7 | 2562 | mutex_lock(&wq_pool_attach_mutex); |
a2d812a2 | 2563 | |
5c25b5ff | 2564 | kthread_set_per_cpu(worker->task, -1); |
da028469 | 2565 | list_del(&worker->node); |
a2d812a2 TH |
2566 | worker->pool = NULL; |
2567 | ||
e02b9312 | 2568 | if (list_empty(&pool->workers) && list_empty(&pool->dying_workers)) |
60f5a4bc | 2569 | detach_completion = pool->detach_completion; |
1258fae7 | 2570 | mutex_unlock(&wq_pool_attach_mutex); |
60f5a4bc | 2571 | |
b62c0751 LJ |
2572 | /* clear leftover flags without pool->lock after it is detached */ |
2573 | worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND); | |
2574 | ||
60f5a4bc LJ |
2575 | if (detach_completion) |
2576 | complete(detach_completion); | |
2577 | } | |
2578 | ||
c34056a3 TH |
2579 | /** |
2580 | * create_worker - create a new workqueue worker | |
63d95a91 | 2581 | * @pool: pool the new worker will belong to |
c34056a3 | 2582 | * |
051e1850 | 2583 | * Create and start a new worker which is attached to @pool. |
c34056a3 TH |
2584 | * |
2585 | * CONTEXT: | |
2586 | * Might sleep. Does GFP_KERNEL allocations. | |
2587 | * | |
d185af30 | 2588 | * Return: |
c34056a3 TH |
2589 | * Pointer to the newly created worker. |
2590 | */ | |
bc2ae0f5 | 2591 | static struct worker *create_worker(struct worker_pool *pool) |
c34056a3 | 2592 | { |
e441b56f ZL |
2593 | struct worker *worker; |
2594 | int id; | |
5d9c7a1e | 2595 | char id_buf[23]; |
c34056a3 | 2596 | |
7cda9aae | 2597 | /* ID is needed to determine kthread name */ |
e441b56f | 2598 | id = ida_alloc(&pool->worker_ida, GFP_KERNEL); |
3f0ea0b8 PM |
2599 | if (id < 0) { |
2600 | pr_err_once("workqueue: Failed to allocate a worker ID: %pe\n", | |
2601 | ERR_PTR(id)); | |
e441b56f | 2602 | return NULL; |
3f0ea0b8 | 2603 | } |
c34056a3 | 2604 | |
f7537df5 | 2605 | worker = alloc_worker(pool->node); |
3f0ea0b8 PM |
2606 | if (!worker) { |
2607 | pr_err_once("workqueue: Failed to allocate a worker\n"); | |
c34056a3 | 2608 | goto fail; |
3f0ea0b8 | 2609 | } |
c34056a3 | 2610 | |
c34056a3 TH |
2611 | worker->id = id; |
2612 | ||
29c91e99 | 2613 | if (pool->cpu >= 0) |
e3c916a4 TH |
2614 | snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, |
2615 | pool->attrs->nice < 0 ? "H" : ""); | |
f3421797 | 2616 | else |
e3c916a4 TH |
2617 | snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); |
2618 | ||
f3f90ad4 | 2619 | worker->task = kthread_create_on_node(worker_thread, worker, pool->node, |
e3c916a4 | 2620 | "kworker/%s", id_buf); |
3f0ea0b8 | 2621 | if (IS_ERR(worker->task)) { |
60f54038 PM |
2622 | if (PTR_ERR(worker->task) == -EINTR) { |
2623 | pr_err("workqueue: Interrupted when creating a worker thread \"kworker/%s\"\n", | |
2624 | id_buf); | |
2625 | } else { | |
2626 | pr_err_once("workqueue: Failed to create a worker thread: %pe", | |
2627 | worker->task); | |
2628 | } | |
c34056a3 | 2629 | goto fail; |
3f0ea0b8 | 2630 | } |
c34056a3 | 2631 | |
91151228 | 2632 | set_user_nice(worker->task, pool->attrs->nice); |
9546b29e | 2633 | kthread_bind_mask(worker->task, pool_allowed_cpus(pool)); |
91151228 | 2634 | |
da028469 | 2635 | /* successful, attach the worker to the pool */ |
4736cbf7 | 2636 | worker_attach_to_pool(worker, pool); |
822d8405 | 2637 | |
051e1850 | 2638 | /* start the newly created worker */ |
a9b8a985 | 2639 | raw_spin_lock_irq(&pool->lock); |
0219a352 | 2640 | |
051e1850 LJ |
2641 | worker->pool->nr_workers++; |
2642 | worker_enter_idle(worker); | |
0219a352 TH |
2643 | |
2644 | /* | |
2645 | * @worker is waiting on a completion in kthread() and will trigger hung | |
6a229b0e TH |
2646 | * check if not woken up soon. As kick_pool() is noop if @pool is empty, |
2647 | * wake it up explicitly. | |
0219a352 | 2648 | */ |
051e1850 | 2649 | wake_up_process(worker->task); |
0219a352 | 2650 | |
a9b8a985 | 2651 | raw_spin_unlock_irq(&pool->lock); |
051e1850 | 2652 | |
c34056a3 | 2653 | return worker; |
822d8405 | 2654 | |
c34056a3 | 2655 | fail: |
e441b56f | 2656 | ida_free(&pool->worker_ida, id); |
c34056a3 TH |
2657 | kfree(worker); |
2658 | return NULL; | |
2659 | } | |
2660 | ||
793777bc VS |
2661 | static void unbind_worker(struct worker *worker) |
2662 | { | |
2663 | lockdep_assert_held(&wq_pool_attach_mutex); | |
2664 | ||
2665 | kthread_set_per_cpu(worker->task, -1); | |
2666 | if (cpumask_intersects(wq_unbound_cpumask, cpu_active_mask)) | |
2667 | WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, wq_unbound_cpumask) < 0); | |
2668 | else | |
2669 | WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0); | |
2670 | } | |
2671 | ||
e02b9312 VS |
2672 | static void wake_dying_workers(struct list_head *cull_list) |
2673 | { | |
2674 | struct worker *worker, *tmp; | |
2675 | ||
2676 | list_for_each_entry_safe(worker, tmp, cull_list, entry) { | |
2677 | list_del_init(&worker->entry); | |
2678 | unbind_worker(worker); | |
2679 | /* | |
2680 | * If the worker was somehow already running, then it had to be | |
2681 | * in pool->idle_list when set_worker_dying() happened or we | |
2682 | * wouldn't have gotten here. | |
2683 | * | |
2684 | * Thus, the worker must either have observed the WORKER_DIE | |
2685 | * flag, or have set its state to TASK_IDLE. Either way, the | |
2686 | * below will be observed by the worker and is safe to do | |
2687 | * outside of pool->lock. | |
2688 | */ | |
2689 | wake_up_process(worker->task); | |
2690 | } | |
2691 | } | |
2692 | ||
c34056a3 | 2693 | /** |
e02b9312 | 2694 | * set_worker_dying - Tag a worker for destruction |
c34056a3 | 2695 | * @worker: worker to be destroyed |
e02b9312 | 2696 | * @list: transfer worker away from its pool->idle_list and into list |
c34056a3 | 2697 | * |
e02b9312 VS |
2698 | * Tag @worker for destruction and adjust @pool stats accordingly. The worker |
2699 | * should be idle. | |
c8e55f36 TH |
2700 | * |
2701 | * CONTEXT: | |
a9b8a985 | 2702 | * raw_spin_lock_irq(pool->lock). |
c34056a3 | 2703 | */ |
e02b9312 | 2704 | static void set_worker_dying(struct worker *worker, struct list_head *list) |
c34056a3 | 2705 | { |
bd7bdd43 | 2706 | struct worker_pool *pool = worker->pool; |
c34056a3 | 2707 | |
cd549687 | 2708 | lockdep_assert_held(&pool->lock); |
e02b9312 | 2709 | lockdep_assert_held(&wq_pool_attach_mutex); |
cd549687 | 2710 | |
c34056a3 | 2711 | /* sanity check frenzy */ |
6183c009 | 2712 | if (WARN_ON(worker->current_work) || |
73eb7fe7 LJ |
2713 | WARN_ON(!list_empty(&worker->scheduled)) || |
2714 | WARN_ON(!(worker->flags & WORKER_IDLE))) | |
6183c009 | 2715 | return; |
c34056a3 | 2716 | |
73eb7fe7 LJ |
2717 | pool->nr_workers--; |
2718 | pool->nr_idle--; | |
5bdfff96 | 2719 | |
cb444766 | 2720 | worker->flags |= WORKER_DIE; |
e02b9312 VS |
2721 | |
2722 | list_move(&worker->entry, list); | |
2723 | list_move(&worker->node, &pool->dying_workers); | |
c34056a3 TH |
2724 | } |
2725 | ||
3f959aa3 VS |
2726 | /** |
2727 | * idle_worker_timeout - check if some idle workers can now be deleted. | |
2728 | * @t: The pool's idle_timer that just expired | |
2729 | * | |
2730 | * The timer is armed in worker_enter_idle(). Note that it isn't disarmed in | |
2731 | * worker_leave_idle(), as a worker flicking between idle and active while its | |
2732 | * pool is at the too_many_workers() tipping point would cause too much timer | |
2733 | * housekeeping overhead. Since IDLE_WORKER_TIMEOUT is long enough, we just let | |
2734 | * it expire and re-evaluate things from there. | |
2735 | */ | |
32a6c723 | 2736 | static void idle_worker_timeout(struct timer_list *t) |
e22bee78 | 2737 | { |
32a6c723 | 2738 | struct worker_pool *pool = from_timer(pool, t, idle_timer); |
3f959aa3 VS |
2739 | bool do_cull = false; |
2740 | ||
2741 | if (work_pending(&pool->idle_cull_work)) | |
2742 | return; | |
e22bee78 | 2743 | |
a9b8a985 | 2744 | raw_spin_lock_irq(&pool->lock); |
e22bee78 | 2745 | |
3f959aa3 | 2746 | if (too_many_workers(pool)) { |
e22bee78 TH |
2747 | struct worker *worker; |
2748 | unsigned long expires; | |
2749 | ||
2750 | /* idle_list is kept in LIFO order, check the last one */ | |
3f959aa3 VS |
2751 | worker = list_entry(pool->idle_list.prev, struct worker, entry); |
2752 | expires = worker->last_active + IDLE_WORKER_TIMEOUT; | |
2753 | do_cull = !time_before(jiffies, expires); | |
2754 | ||
2755 | if (!do_cull) | |
2756 | mod_timer(&pool->idle_timer, expires); | |
2757 | } | |
2758 | raw_spin_unlock_irq(&pool->lock); | |
2759 | ||
2760 | if (do_cull) | |
2761 | queue_work(system_unbound_wq, &pool->idle_cull_work); | |
2762 | } | |
2763 | ||
2764 | /** | |
2765 | * idle_cull_fn - cull workers that have been idle for too long. | |
2766 | * @work: the pool's work for handling these idle workers | |
2767 | * | |
2768 | * This goes through a pool's idle workers and gets rid of those that have been | |
2769 | * idle for at least IDLE_WORKER_TIMEOUT seconds. | |
e02b9312 VS |
2770 | * |
2771 | * We don't want to disturb isolated CPUs because of a pcpu kworker being | |
2772 | * culled, so this also resets worker affinity. This requires a sleepable | |
2773 | * context, hence the split between timer callback and work item. | |
3f959aa3 VS |
2774 | */ |
2775 | static void idle_cull_fn(struct work_struct *work) | |
2776 | { | |
2777 | struct worker_pool *pool = container_of(work, struct worker_pool, idle_cull_work); | |
9680540c | 2778 | LIST_HEAD(cull_list); |
3f959aa3 | 2779 | |
e02b9312 VS |
2780 | /* |
2781 | * Grabbing wq_pool_attach_mutex here ensures an already-running worker | |
2782 | * cannot proceed beyong worker_detach_from_pool() in its self-destruct | |
2783 | * path. This is required as a previously-preempted worker could run after | |
2784 | * set_worker_dying() has happened but before wake_dying_workers() did. | |
2785 | */ | |
2786 | mutex_lock(&wq_pool_attach_mutex); | |
3f959aa3 VS |
2787 | raw_spin_lock_irq(&pool->lock); |
2788 | ||
2789 | while (too_many_workers(pool)) { | |
2790 | struct worker *worker; | |
2791 | unsigned long expires; | |
2792 | ||
63d95a91 | 2793 | worker = list_entry(pool->idle_list.prev, struct worker, entry); |
e22bee78 TH |
2794 | expires = worker->last_active + IDLE_WORKER_TIMEOUT; |
2795 | ||
3347fc9f | 2796 | if (time_before(jiffies, expires)) { |
63d95a91 | 2797 | mod_timer(&pool->idle_timer, expires); |
3347fc9f | 2798 | break; |
d5abe669 | 2799 | } |
3347fc9f | 2800 | |
e02b9312 | 2801 | set_worker_dying(worker, &cull_list); |
e22bee78 TH |
2802 | } |
2803 | ||
a9b8a985 | 2804 | raw_spin_unlock_irq(&pool->lock); |
e02b9312 VS |
2805 | wake_dying_workers(&cull_list); |
2806 | mutex_unlock(&wq_pool_attach_mutex); | |
e22bee78 | 2807 | } |
d5abe669 | 2808 | |
493a1724 | 2809 | static void send_mayday(struct work_struct *work) |
e22bee78 | 2810 | { |
112202d9 TH |
2811 | struct pool_workqueue *pwq = get_work_pwq(work); |
2812 | struct workqueue_struct *wq = pwq->wq; | |
493a1724 | 2813 | |
2e109a28 | 2814 | lockdep_assert_held(&wq_mayday_lock); |
e22bee78 | 2815 | |
493008a8 | 2816 | if (!wq->rescuer) |
493a1724 | 2817 | return; |
e22bee78 TH |
2818 | |
2819 | /* mayday mayday mayday */ | |
493a1724 | 2820 | if (list_empty(&pwq->mayday_node)) { |
77668c8b LJ |
2821 | /* |
2822 | * If @pwq is for an unbound wq, its base ref may be put at | |
2823 | * any time due to an attribute change. Pin @pwq until the | |
2824 | * rescuer is done with it. | |
2825 | */ | |
2826 | get_pwq(pwq); | |
493a1724 | 2827 | list_add_tail(&pwq->mayday_node, &wq->maydays); |
e22bee78 | 2828 | wake_up_process(wq->rescuer->task); |
725e8ec5 | 2829 | pwq->stats[PWQ_STAT_MAYDAY]++; |
493a1724 | 2830 | } |
e22bee78 TH |
2831 | } |
2832 | ||
32a6c723 | 2833 | static void pool_mayday_timeout(struct timer_list *t) |
e22bee78 | 2834 | { |
32a6c723 | 2835 | struct worker_pool *pool = from_timer(pool, t, mayday_timer); |
e22bee78 TH |
2836 | struct work_struct *work; |
2837 | ||
a9b8a985 SAS |
2838 | raw_spin_lock_irq(&pool->lock); |
2839 | raw_spin_lock(&wq_mayday_lock); /* for wq->maydays */ | |
e22bee78 | 2840 | |
63d95a91 | 2841 | if (need_to_create_worker(pool)) { |
e22bee78 TH |
2842 | /* |
2843 | * We've been trying to create a new worker but | |
2844 | * haven't been successful. We might be hitting an | |
2845 | * allocation deadlock. Send distress signals to | |
2846 | * rescuers. | |
2847 | */ | |
63d95a91 | 2848 | list_for_each_entry(work, &pool->worklist, entry) |
e22bee78 | 2849 | send_mayday(work); |
1da177e4 | 2850 | } |
e22bee78 | 2851 | |
a9b8a985 SAS |
2852 | raw_spin_unlock(&wq_mayday_lock); |
2853 | raw_spin_unlock_irq(&pool->lock); | |
e22bee78 | 2854 | |
63d95a91 | 2855 | mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); |
1da177e4 LT |
2856 | } |
2857 | ||
e22bee78 TH |
2858 | /** |
2859 | * maybe_create_worker - create a new worker if necessary | |
63d95a91 | 2860 | * @pool: pool to create a new worker for |
e22bee78 | 2861 | * |
63d95a91 | 2862 | * Create a new worker for @pool if necessary. @pool is guaranteed to |
e22bee78 TH |
2863 | * have at least one idle worker on return from this function. If |
2864 | * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is | |
63d95a91 | 2865 | * sent to all rescuers with works scheduled on @pool to resolve |
e22bee78 TH |
2866 | * possible allocation deadlock. |
2867 | * | |
c5aa87bb TH |
2868 | * On return, need_to_create_worker() is guaranteed to be %false and |
2869 | * may_start_working() %true. | |
e22bee78 TH |
2870 | * |
2871 | * LOCKING: | |
a9b8a985 | 2872 | * raw_spin_lock_irq(pool->lock) which may be released and regrabbed |
e22bee78 TH |
2873 | * multiple times. Does GFP_KERNEL allocations. Called only from |
2874 | * manager. | |
e22bee78 | 2875 | */ |
29187a9e | 2876 | static void maybe_create_worker(struct worker_pool *pool) |
d565ed63 TH |
2877 | __releases(&pool->lock) |
2878 | __acquires(&pool->lock) | |
1da177e4 | 2879 | { |
e22bee78 | 2880 | restart: |
a9b8a985 | 2881 | raw_spin_unlock_irq(&pool->lock); |
9f9c2364 | 2882 | |
e22bee78 | 2883 | /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ |
63d95a91 | 2884 | mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); |
e22bee78 TH |
2885 | |
2886 | while (true) { | |
051e1850 | 2887 | if (create_worker(pool) || !need_to_create_worker(pool)) |
e22bee78 | 2888 | break; |
1da177e4 | 2889 | |
e212f361 | 2890 | schedule_timeout_interruptible(CREATE_COOLDOWN); |
9f9c2364 | 2891 | |
63d95a91 | 2892 | if (!need_to_create_worker(pool)) |
e22bee78 TH |
2893 | break; |
2894 | } | |
2895 | ||
63d95a91 | 2896 | del_timer_sync(&pool->mayday_timer); |
a9b8a985 | 2897 | raw_spin_lock_irq(&pool->lock); |
051e1850 LJ |
2898 | /* |
2899 | * This is necessary even after a new worker was just successfully | |
2900 | * created as @pool->lock was dropped and the new worker might have | |
2901 | * already become busy. | |
2902 | */ | |
63d95a91 | 2903 | if (need_to_create_worker(pool)) |
e22bee78 | 2904 | goto restart; |
e22bee78 TH |
2905 | } |
2906 | ||
73f53c4a | 2907 | /** |
e22bee78 TH |
2908 | * manage_workers - manage worker pool |
2909 | * @worker: self | |
73f53c4a | 2910 | * |
706026c2 | 2911 | * Assume the manager role and manage the worker pool @worker belongs |
e22bee78 | 2912 | * to. At any given time, there can be only zero or one manager per |
706026c2 | 2913 | * pool. The exclusion is handled automatically by this function. |
e22bee78 TH |
2914 | * |
2915 | * The caller can safely start processing works on false return. On | |
2916 | * true return, it's guaranteed that need_to_create_worker() is false | |
2917 | * and may_start_working() is true. | |
73f53c4a TH |
2918 | * |
2919 | * CONTEXT: | |
a9b8a985 | 2920 | * raw_spin_lock_irq(pool->lock) which may be released and regrabbed |
e22bee78 TH |
2921 | * multiple times. Does GFP_KERNEL allocations. |
2922 | * | |
d185af30 | 2923 | * Return: |
29187a9e TH |
2924 | * %false if the pool doesn't need management and the caller can safely |
2925 | * start processing works, %true if management function was performed and | |
2926 | * the conditions that the caller verified before calling the function may | |
2927 | * no longer be true. | |
73f53c4a | 2928 | */ |
e22bee78 | 2929 | static bool manage_workers(struct worker *worker) |
73f53c4a | 2930 | { |
63d95a91 | 2931 | struct worker_pool *pool = worker->pool; |
73f53c4a | 2932 | |
692b4825 | 2933 | if (pool->flags & POOL_MANAGER_ACTIVE) |
29187a9e | 2934 | return false; |
692b4825 TH |
2935 | |
2936 | pool->flags |= POOL_MANAGER_ACTIVE; | |
2607d7a6 | 2937 | pool->manager = worker; |
1e19ffc6 | 2938 | |
29187a9e | 2939 | maybe_create_worker(pool); |
e22bee78 | 2940 | |
2607d7a6 | 2941 | pool->manager = NULL; |
692b4825 | 2942 | pool->flags &= ~POOL_MANAGER_ACTIVE; |
d8bb65ab | 2943 | rcuwait_wake_up(&manager_wait); |
29187a9e | 2944 | return true; |
73f53c4a TH |
2945 | } |
2946 | ||
a62428c0 TH |
2947 | /** |
2948 | * process_one_work - process single work | |
c34056a3 | 2949 | * @worker: self |
a62428c0 TH |
2950 | * @work: work to process |
2951 | * | |
2952 | * Process @work. This function contains all the logics necessary to | |
2953 | * process a single work including synchronization against and | |
2954 | * interaction with other workers on the same cpu, queueing and | |
2955 | * flushing. As long as context requirement is met, any worker can | |
2956 | * call this function to process a work. | |
2957 | * | |
2958 | * CONTEXT: | |
a9b8a985 | 2959 | * raw_spin_lock_irq(pool->lock) which is released and regrabbed. |
a62428c0 | 2960 | */ |
c34056a3 | 2961 | static void process_one_work(struct worker *worker, struct work_struct *work) |
d565ed63 TH |
2962 | __releases(&pool->lock) |
2963 | __acquires(&pool->lock) | |
a62428c0 | 2964 | { |
112202d9 | 2965 | struct pool_workqueue *pwq = get_work_pwq(work); |
bd7bdd43 | 2966 | struct worker_pool *pool = worker->pool; |
c4560c2c | 2967 | unsigned long work_data; |
a62428c0 TH |
2968 | #ifdef CONFIG_LOCKDEP |
2969 | /* | |
2970 | * It is permissible to free the struct work_struct from | |
2971 | * inside the function that is called from it, this we need to | |
2972 | * take into account for lockdep too. To avoid bogus "held | |
2973 | * lock freed" warnings as well as problems when looking into | |
2974 | * work->lockdep_map, make a copy and use that here. | |
2975 | */ | |
4d82a1de PZ |
2976 | struct lockdep_map lockdep_map; |
2977 | ||
2978 | lockdep_copy_map(&lockdep_map, &work->lockdep_map); | |
a62428c0 | 2979 | #endif |
807407c0 | 2980 | /* ensure we're on the correct CPU */ |
85327af6 | 2981 | WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && |
ec22ca5e | 2982 | raw_smp_processor_id() != pool->cpu); |
25511a47 | 2983 | |
8930caba | 2984 | /* claim and dequeue */ |
a62428c0 | 2985 | debug_work_deactivate(work); |
c9e7cf27 | 2986 | hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); |
c34056a3 | 2987 | worker->current_work = work; |
a2c1c57b | 2988 | worker->current_func = work->func; |
112202d9 | 2989 | worker->current_pwq = pwq; |
616db877 | 2990 | worker->current_at = worker->task->se.sum_exec_runtime; |
c4560c2c | 2991 | work_data = *work_data_bits(work); |
d812796e | 2992 | worker->current_color = get_work_color(work_data); |
7a22ad75 | 2993 | |
8bf89593 TH |
2994 | /* |
2995 | * Record wq name for cmdline and debug reporting, may get | |
2996 | * overridden through set_worker_desc(). | |
2997 | */ | |
2998 | strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN); | |
2999 | ||
a62428c0 TH |
3000 | list_del_init(&work->entry); |
3001 | ||
fb0e7beb | 3002 | /* |
228f1d00 LJ |
3003 | * CPU intensive works don't participate in concurrency management. |
3004 | * They're the scheduler's responsibility. This takes @worker out | |
3005 | * of concurrency management and the next code block will chain | |
3006 | * execution of the pending work items. | |
fb0e7beb | 3007 | */ |
616db877 | 3008 | if (unlikely(pwq->wq->flags & WQ_CPU_INTENSIVE)) |
228f1d00 | 3009 | worker_set_flags(worker, WORKER_CPU_INTENSIVE); |
fb0e7beb | 3010 | |
974271c4 | 3011 | /* |
0219a352 TH |
3012 | * Kick @pool if necessary. It's always noop for per-cpu worker pools |
3013 | * since nr_running would always be >= 1 at this point. This is used to | |
3014 | * chain execution of the pending work items for WORKER_NOT_RUNNING | |
3015 | * workers such as the UNBOUND and CPU_INTENSIVE ones. | |
974271c4 | 3016 | */ |
0219a352 | 3017 | kick_pool(pool); |
974271c4 | 3018 | |
8930caba | 3019 | /* |
7c3eed5c | 3020 | * Record the last pool and clear PENDING which should be the last |
d565ed63 | 3021 | * update to @work. Also, do this inside @pool->lock so that |
23657bb1 TH |
3022 | * PENDING and queued state changes happen together while IRQ is |
3023 | * disabled. | |
8930caba | 3024 | */ |
7c3eed5c | 3025 | set_work_pool_and_clear_pending(work, pool->id); |
a62428c0 | 3026 | |
fe48ba7d | 3027 | pwq->stats[PWQ_STAT_STARTED]++; |
a9b8a985 | 3028 | raw_spin_unlock_irq(&pool->lock); |
a62428c0 | 3029 | |
a1d14934 | 3030 | lock_map_acquire(&pwq->wq->lockdep_map); |
a62428c0 | 3031 | lock_map_acquire(&lockdep_map); |
e6f3faa7 | 3032 | /* |
f52be570 PZ |
3033 | * Strictly speaking we should mark the invariant state without holding |
3034 | * any locks, that is, before these two lock_map_acquire()'s. | |
e6f3faa7 PZ |
3035 | * |
3036 | * However, that would result in: | |
3037 | * | |
3038 | * A(W1) | |
3039 | * WFC(C) | |
3040 | * A(W1) | |
3041 | * C(C) | |
3042 | * | |
3043 | * Which would create W1->C->W1 dependencies, even though there is no | |
3044 | * actual deadlock possible. There are two solutions, using a | |
3045 | * read-recursive acquire on the work(queue) 'locks', but this will then | |
f52be570 | 3046 | * hit the lockdep limitation on recursive locks, or simply discard |
e6f3faa7 PZ |
3047 | * these locks. |
3048 | * | |
3049 | * AFAICT there is no possible deadlock scenario between the | |
3050 | * flush_work() and complete() primitives (except for single-threaded | |
3051 | * workqueues), so hiding them isn't a problem. | |
3052 | */ | |
f52be570 | 3053 | lockdep_invariant_state(true); |
e36c886a | 3054 | trace_workqueue_execute_start(work); |
a2c1c57b | 3055 | worker->current_func(work); |
e36c886a AV |
3056 | /* |
3057 | * While we must be careful to not use "work" after this, the trace | |
3058 | * point will only record its address. | |
3059 | */ | |
1c5da0ec | 3060 | trace_workqueue_execute_end(work, worker->current_func); |
725e8ec5 | 3061 | pwq->stats[PWQ_STAT_COMPLETED]++; |
a62428c0 | 3062 | lock_map_release(&lockdep_map); |
112202d9 | 3063 | lock_map_release(&pwq->wq->lockdep_map); |
a62428c0 | 3064 | |
1a65a6d1 XY |
3065 | if (unlikely(in_atomic() || lockdep_depth(current) > 0 || |
3066 | rcu_preempt_depth() > 0)) { | |
3067 | pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d/%d\n" | |
d75f773c | 3068 | " last function: %ps\n", |
1a65a6d1 XY |
3069 | current->comm, preempt_count(), rcu_preempt_depth(), |
3070 | task_pid_nr(current), worker->current_func); | |
a62428c0 TH |
3071 | debug_show_held_locks(current); |
3072 | dump_stack(); | |
3073 | } | |
3074 | ||
b22ce278 | 3075 | /* |
025f50f3 | 3076 | * The following prevents a kworker from hogging CPU on !PREEMPTION |
b22ce278 TH |
3077 | * kernels, where a requeueing work item waiting for something to |
3078 | * happen could deadlock with stop_machine as such work item could | |
3079 | * indefinitely requeue itself while all other CPUs are trapped in | |
789cbbec JL |
3080 | * stop_machine. At the same time, report a quiescent RCU state so |
3081 | * the same condition doesn't freeze RCU. | |
b22ce278 | 3082 | */ |
a7e6425e | 3083 | cond_resched(); |
b22ce278 | 3084 | |
a9b8a985 | 3085 | raw_spin_lock_irq(&pool->lock); |
a62428c0 | 3086 | |
616db877 TH |
3087 | /* |
3088 | * In addition to %WQ_CPU_INTENSIVE, @worker may also have been marked | |
3089 | * CPU intensive by wq_worker_tick() if @work hogged CPU longer than | |
3090 | * wq_cpu_intensive_thresh_us. Clear it. | |
3091 | */ | |
3092 | worker_clr_flags(worker, WORKER_CPU_INTENSIVE); | |
fb0e7beb | 3093 | |
1b69ac6b JW |
3094 | /* tag the worker for identification in schedule() */ |
3095 | worker->last_func = worker->current_func; | |
3096 | ||
a62428c0 | 3097 | /* we're done with it, release */ |
42f8570f | 3098 | hash_del(&worker->hentry); |
c34056a3 | 3099 | worker->current_work = NULL; |
a2c1c57b | 3100 | worker->current_func = NULL; |
112202d9 | 3101 | worker->current_pwq = NULL; |
d812796e | 3102 | worker->current_color = INT_MAX; |
dd6c3c54 TH |
3103 | |
3104 | /* must be the last step, see the function comment */ | |
c4560c2c | 3105 | pwq_dec_nr_in_flight(pwq, work_data); |
a62428c0 TH |
3106 | } |
3107 | ||
affee4b2 TH |
3108 | /** |
3109 | * process_scheduled_works - process scheduled works | |
3110 | * @worker: self | |
3111 | * | |
3112 | * Process all scheduled works. Please note that the scheduled list | |
3113 | * may change while processing a work, so this function repeatedly | |
3114 | * fetches a work from the top and executes it. | |
3115 | * | |
3116 | * CONTEXT: | |
a9b8a985 | 3117 | * raw_spin_lock_irq(pool->lock) which may be released and regrabbed |
affee4b2 TH |
3118 | * multiple times. |
3119 | */ | |
3120 | static void process_scheduled_works(struct worker *worker) | |
1da177e4 | 3121 | { |
c0ab017d TH |
3122 | struct work_struct *work; |
3123 | bool first = true; | |
3124 | ||
3125 | while ((work = list_first_entry_or_null(&worker->scheduled, | |
3126 | struct work_struct, entry))) { | |
3127 | if (first) { | |
3128 | worker->pool->watchdog_ts = jiffies; | |
3129 | first = false; | |
3130 | } | |
c34056a3 | 3131 | process_one_work(worker, work); |
1da177e4 | 3132 | } |
1da177e4 LT |
3133 | } |
3134 | ||
197f6acc TH |
3135 | static void set_pf_worker(bool val) |
3136 | { | |
3137 | mutex_lock(&wq_pool_attach_mutex); | |
3138 | if (val) | |
3139 | current->flags |= PF_WQ_WORKER; | |
3140 | else | |
3141 | current->flags &= ~PF_WQ_WORKER; | |
3142 | mutex_unlock(&wq_pool_attach_mutex); | |
3143 | } | |
3144 | ||
4690c4ab TH |
3145 | /** |
3146 | * worker_thread - the worker thread function | |
c34056a3 | 3147 | * @__worker: self |
4690c4ab | 3148 | * |
c5aa87bb TH |
3149 | * The worker thread function. All workers belong to a worker_pool - |
3150 | * either a per-cpu one or dynamic unbound one. These workers process all | |
3151 | * work items regardless of their specific target workqueue. The only | |
3152 | * exception is work items which belong to workqueues with a rescuer which | |
3153 | * will be explained in rescuer_thread(). | |
d185af30 YB |
3154 | * |
3155 | * Return: 0 | |
4690c4ab | 3156 | */ |
c34056a3 | 3157 | static int worker_thread(void *__worker) |
1da177e4 | 3158 | { |
c34056a3 | 3159 | struct worker *worker = __worker; |
bd7bdd43 | 3160 | struct worker_pool *pool = worker->pool; |
1da177e4 | 3161 | |
e22bee78 | 3162 | /* tell the scheduler that this is a workqueue worker */ |
197f6acc | 3163 | set_pf_worker(true); |
c8e55f36 | 3164 | woke_up: |
a9b8a985 | 3165 | raw_spin_lock_irq(&pool->lock); |
1da177e4 | 3166 | |
a9ab775b TH |
3167 | /* am I supposed to die? */ |
3168 | if (unlikely(worker->flags & WORKER_DIE)) { | |
a9b8a985 | 3169 | raw_spin_unlock_irq(&pool->lock); |
197f6acc | 3170 | set_pf_worker(false); |
60f5a4bc LJ |
3171 | |
3172 | set_task_comm(worker->task, "kworker/dying"); | |
e441b56f | 3173 | ida_free(&pool->worker_ida, worker->id); |
a2d812a2 | 3174 | worker_detach_from_pool(worker); |
e02b9312 | 3175 | WARN_ON_ONCE(!list_empty(&worker->entry)); |
60f5a4bc | 3176 | kfree(worker); |
a9ab775b | 3177 | return 0; |
c8e55f36 | 3178 | } |
affee4b2 | 3179 | |
c8e55f36 | 3180 | worker_leave_idle(worker); |
db7bccf4 | 3181 | recheck: |
e22bee78 | 3182 | /* no more worker necessary? */ |
63d95a91 | 3183 | if (!need_more_worker(pool)) |
e22bee78 TH |
3184 | goto sleep; |
3185 | ||
3186 | /* do we need to manage? */ | |
63d95a91 | 3187 | if (unlikely(!may_start_working(pool)) && manage_workers(worker)) |
e22bee78 TH |
3188 | goto recheck; |
3189 | ||
c8e55f36 TH |
3190 | /* |
3191 | * ->scheduled list can only be filled while a worker is | |
3192 | * preparing to process a work or actually processing it. | |
3193 | * Make sure nobody diddled with it while I was sleeping. | |
3194 | */ | |
6183c009 | 3195 | WARN_ON_ONCE(!list_empty(&worker->scheduled)); |
c8e55f36 | 3196 | |
e22bee78 | 3197 | /* |
a9ab775b TH |
3198 | * Finish PREP stage. We're guaranteed to have at least one idle |
3199 | * worker or that someone else has already assumed the manager | |
3200 | * role. This is where @worker starts participating in concurrency | |
3201 | * management if applicable and concurrency management is restored | |
3202 | * after being rebound. See rebind_workers() for details. | |
e22bee78 | 3203 | */ |
a9ab775b | 3204 | worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); |
e22bee78 TH |
3205 | |
3206 | do { | |
c8e55f36 | 3207 | struct work_struct *work = |
bd7bdd43 | 3208 | list_first_entry(&pool->worklist, |
c8e55f36 TH |
3209 | struct work_struct, entry); |
3210 | ||
873eaca6 TH |
3211 | if (assign_work(work, worker, NULL)) |
3212 | process_scheduled_works(worker); | |
63d95a91 | 3213 | } while (keep_working(pool)); |
e22bee78 | 3214 | |
228f1d00 | 3215 | worker_set_flags(worker, WORKER_PREP); |
d313dd85 | 3216 | sleep: |
c8e55f36 | 3217 | /* |
d565ed63 TH |
3218 | * pool->lock is held and there's no work to process and no need to |
3219 | * manage, sleep. Workers are woken up only while holding | |
3220 | * pool->lock or from local cpu, so setting the current state | |
3221 | * before releasing pool->lock is enough to prevent losing any | |
3222 | * event. | |
c8e55f36 TH |
3223 | */ |
3224 | worker_enter_idle(worker); | |
c5a94a61 | 3225 | __set_current_state(TASK_IDLE); |
a9b8a985 | 3226 | raw_spin_unlock_irq(&pool->lock); |
c8e55f36 TH |
3227 | schedule(); |
3228 | goto woke_up; | |
1da177e4 LT |
3229 | } |
3230 | ||
e22bee78 TH |
3231 | /** |
3232 | * rescuer_thread - the rescuer thread function | |
111c225a | 3233 | * @__rescuer: self |
e22bee78 TH |
3234 | * |
3235 | * Workqueue rescuer thread function. There's one rescuer for each | |
493008a8 | 3236 | * workqueue which has WQ_MEM_RECLAIM set. |
e22bee78 | 3237 | * |
706026c2 | 3238 | * Regular work processing on a pool may block trying to create a new |
e22bee78 TH |
3239 | * worker which uses GFP_KERNEL allocation which has slight chance of |
3240 | * developing into deadlock if some works currently on the same queue | |
3241 | * need to be processed to satisfy the GFP_KERNEL allocation. This is | |
3242 | * the problem rescuer solves. | |
3243 | * | |
706026c2 TH |
3244 | * When such condition is possible, the pool summons rescuers of all |
3245 | * workqueues which have works queued on the pool and let them process | |
e22bee78 TH |
3246 | * those works so that forward progress can be guaranteed. |
3247 | * | |
3248 | * This should happen rarely. | |
d185af30 YB |
3249 | * |
3250 | * Return: 0 | |
e22bee78 | 3251 | */ |
111c225a | 3252 | static int rescuer_thread(void *__rescuer) |
e22bee78 | 3253 | { |
111c225a TH |
3254 | struct worker *rescuer = __rescuer; |
3255 | struct workqueue_struct *wq = rescuer->rescue_wq; | |
4d595b86 | 3256 | bool should_stop; |
e22bee78 TH |
3257 | |
3258 | set_user_nice(current, RESCUER_NICE_LEVEL); | |
111c225a TH |
3259 | |
3260 | /* | |
3261 | * Mark rescuer as worker too. As WORKER_PREP is never cleared, it | |
3262 | * doesn't participate in concurrency management. | |
3263 | */ | |
197f6acc | 3264 | set_pf_worker(true); |
e22bee78 | 3265 | repeat: |
c5a94a61 | 3266 | set_current_state(TASK_IDLE); |
e22bee78 | 3267 | |
4d595b86 LJ |
3268 | /* |
3269 | * By the time the rescuer is requested to stop, the workqueue | |
3270 | * shouldn't have any work pending, but @wq->maydays may still have | |
3271 | * pwq(s) queued. This can happen by non-rescuer workers consuming | |
3272 | * all the work items before the rescuer got to them. Go through | |
3273 | * @wq->maydays processing before acting on should_stop so that the | |
3274 | * list is always empty on exit. | |
3275 | */ | |
3276 | should_stop = kthread_should_stop(); | |
e22bee78 | 3277 | |
493a1724 | 3278 | /* see whether any pwq is asking for help */ |
a9b8a985 | 3279 | raw_spin_lock_irq(&wq_mayday_lock); |
493a1724 TH |
3280 | |
3281 | while (!list_empty(&wq->maydays)) { | |
3282 | struct pool_workqueue *pwq = list_first_entry(&wq->maydays, | |
3283 | struct pool_workqueue, mayday_node); | |
112202d9 | 3284 | struct worker_pool *pool = pwq->pool; |
e22bee78 TH |
3285 | struct work_struct *work, *n; |
3286 | ||
3287 | __set_current_state(TASK_RUNNING); | |
493a1724 TH |
3288 | list_del_init(&pwq->mayday_node); |
3289 | ||
a9b8a985 | 3290 | raw_spin_unlock_irq(&wq_mayday_lock); |
e22bee78 | 3291 | |
51697d39 LJ |
3292 | worker_attach_to_pool(rescuer, pool); |
3293 | ||
a9b8a985 | 3294 | raw_spin_lock_irq(&pool->lock); |
e22bee78 TH |
3295 | |
3296 | /* | |
3297 | * Slurp in all works issued via this workqueue and | |
3298 | * process'em. | |
3299 | */ | |
873eaca6 | 3300 | WARN_ON_ONCE(!list_empty(&rescuer->scheduled)); |
82607adc | 3301 | list_for_each_entry_safe(work, n, &pool->worklist, entry) { |
873eaca6 TH |
3302 | if (get_work_pwq(work) == pwq && |
3303 | assign_work(work, rescuer, &n)) | |
725e8ec5 | 3304 | pwq->stats[PWQ_STAT_RESCUED]++; |
82607adc | 3305 | } |
e22bee78 | 3306 | |
873eaca6 | 3307 | if (!list_empty(&rescuer->scheduled)) { |
008847f6 N |
3308 | process_scheduled_works(rescuer); |
3309 | ||
3310 | /* | |
3311 | * The above execution of rescued work items could | |
3312 | * have created more to rescue through | |
f97a4a1a | 3313 | * pwq_activate_first_inactive() or chained |
008847f6 N |
3314 | * queueing. Let's put @pwq back on mayday list so |
3315 | * that such back-to-back work items, which may be | |
3316 | * being used to relieve memory pressure, don't | |
3317 | * incur MAYDAY_INTERVAL delay inbetween. | |
3318 | */ | |
4f3f4cf3 | 3319 | if (pwq->nr_active && need_to_create_worker(pool)) { |
a9b8a985 | 3320 | raw_spin_lock(&wq_mayday_lock); |
e66b39af TH |
3321 | /* |
3322 | * Queue iff we aren't racing destruction | |
3323 | * and somebody else hasn't queued it already. | |
3324 | */ | |
3325 | if (wq->rescuer && list_empty(&pwq->mayday_node)) { | |
3326 | get_pwq(pwq); | |
3327 | list_add_tail(&pwq->mayday_node, &wq->maydays); | |
3328 | } | |
a9b8a985 | 3329 | raw_spin_unlock(&wq_mayday_lock); |
008847f6 N |
3330 | } |
3331 | } | |
7576958a | 3332 | |
77668c8b LJ |
3333 | /* |
3334 | * Put the reference grabbed by send_mayday(). @pool won't | |
13b1d625 | 3335 | * go away while we're still attached to it. |
77668c8b LJ |
3336 | */ |
3337 | put_pwq(pwq); | |
3338 | ||
7576958a | 3339 | /* |
0219a352 TH |
3340 | * Leave this pool. Notify regular workers; otherwise, we end up |
3341 | * with 0 concurrency and stalling the execution. | |
7576958a | 3342 | */ |
0219a352 | 3343 | kick_pool(pool); |
7576958a | 3344 | |
a9b8a985 | 3345 | raw_spin_unlock_irq(&pool->lock); |
13b1d625 | 3346 | |
a2d812a2 | 3347 | worker_detach_from_pool(rescuer); |
13b1d625 | 3348 | |
a9b8a985 | 3349 | raw_spin_lock_irq(&wq_mayday_lock); |
e22bee78 TH |
3350 | } |
3351 | ||
a9b8a985 | 3352 | raw_spin_unlock_irq(&wq_mayday_lock); |
493a1724 | 3353 | |
4d595b86 LJ |
3354 | if (should_stop) { |
3355 | __set_current_state(TASK_RUNNING); | |
197f6acc | 3356 | set_pf_worker(false); |
4d595b86 LJ |
3357 | return 0; |
3358 | } | |
3359 | ||
111c225a TH |
3360 | /* rescuers should never participate in concurrency management */ |
3361 | WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); | |
e22bee78 TH |
3362 | schedule(); |
3363 | goto repeat; | |
1da177e4 LT |
3364 | } |
3365 | ||
fca839c0 TH |
3366 | /** |
3367 | * check_flush_dependency - check for flush dependency sanity | |
3368 | * @target_wq: workqueue being flushed | |
3369 | * @target_work: work item being flushed (NULL for workqueue flushes) | |
3370 | * | |
3371 | * %current is trying to flush the whole @target_wq or @target_work on it. | |
3372 | * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not | |
3373 | * reclaiming memory or running on a workqueue which doesn't have | |
3374 | * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to | |
3375 | * a deadlock. | |
3376 | */ | |
3377 | static void check_flush_dependency(struct workqueue_struct *target_wq, | |
3378 | struct work_struct *target_work) | |
3379 | { | |
3380 | work_func_t target_func = target_work ? target_work->func : NULL; | |
3381 | struct worker *worker; | |
3382 | ||
3383 | if (target_wq->flags & WQ_MEM_RECLAIM) | |
3384 | return; | |
3385 | ||
3386 | worker = current_wq_worker(); | |
3387 | ||
3388 | WARN_ONCE(current->flags & PF_MEMALLOC, | |
d75f773c | 3389 | "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps", |
fca839c0 | 3390 | current->pid, current->comm, target_wq->name, target_func); |
23d11a58 TH |
3391 | WARN_ONCE(worker && ((worker->current_pwq->wq->flags & |
3392 | (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM), | |
d75f773c | 3393 | "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps", |
fca839c0 TH |
3394 | worker->current_pwq->wq->name, worker->current_func, |
3395 | target_wq->name, target_func); | |
3396 | } | |
3397 | ||
fc2e4d70 ON |
3398 | struct wq_barrier { |
3399 | struct work_struct work; | |
3400 | struct completion done; | |
2607d7a6 | 3401 | struct task_struct *task; /* purely informational */ |
fc2e4d70 ON |
3402 | }; |
3403 | ||
3404 | static void wq_barrier_func(struct work_struct *work) | |
3405 | { | |
3406 | struct wq_barrier *barr = container_of(work, struct wq_barrier, work); | |
3407 | complete(&barr->done); | |
3408 | } | |
3409 | ||
4690c4ab TH |
3410 | /** |
3411 | * insert_wq_barrier - insert a barrier work | |
112202d9 | 3412 | * @pwq: pwq to insert barrier into |
4690c4ab | 3413 | * @barr: wq_barrier to insert |
affee4b2 TH |
3414 | * @target: target work to attach @barr to |
3415 | * @worker: worker currently executing @target, NULL if @target is not executing | |
4690c4ab | 3416 | * |
affee4b2 TH |
3417 | * @barr is linked to @target such that @barr is completed only after |
3418 | * @target finishes execution. Please note that the ordering | |
3419 | * guarantee is observed only with respect to @target and on the local | |
3420 | * cpu. | |
3421 | * | |
3422 | * Currently, a queued barrier can't be canceled. This is because | |
3423 | * try_to_grab_pending() can't determine whether the work to be | |
3424 | * grabbed is at the head of the queue and thus can't clear LINKED | |
3425 | * flag of the previous work while there must be a valid next work | |
3426 | * after a work with LINKED flag set. | |
3427 | * | |
3428 | * Note that when @worker is non-NULL, @target may be modified | |
112202d9 | 3429 | * underneath us, so we can't reliably determine pwq from @target. |
4690c4ab TH |
3430 | * |
3431 | * CONTEXT: | |
a9b8a985 | 3432 | * raw_spin_lock_irq(pool->lock). |
4690c4ab | 3433 | */ |
112202d9 | 3434 | static void insert_wq_barrier(struct pool_workqueue *pwq, |
affee4b2 TH |
3435 | struct wq_barrier *barr, |
3436 | struct work_struct *target, struct worker *worker) | |
fc2e4d70 | 3437 | { |
d812796e LJ |
3438 | unsigned int work_flags = 0; |
3439 | unsigned int work_color; | |
affee4b2 | 3440 | struct list_head *head; |
affee4b2 | 3441 | |
dc186ad7 | 3442 | /* |
d565ed63 | 3443 | * debugobject calls are safe here even with pool->lock locked |
dc186ad7 TG |
3444 | * as we know for sure that this will not trigger any of the |
3445 | * checks and call back into the fixup functions where we | |
3446 | * might deadlock. | |
3447 | */ | |
ca1cab37 | 3448 | INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); |
22df02bb | 3449 | __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); |
52fa5bc5 | 3450 | |
fd1a5b04 BP |
3451 | init_completion_map(&barr->done, &target->lockdep_map); |
3452 | ||
2607d7a6 | 3453 | barr->task = current; |
83c22520 | 3454 | |
5797b1c1 | 3455 | /* The barrier work item does not participate in nr_active. */ |
018f3a13 LJ |
3456 | work_flags |= WORK_STRUCT_INACTIVE; |
3457 | ||
affee4b2 TH |
3458 | /* |
3459 | * If @target is currently being executed, schedule the | |
3460 | * barrier to the worker; otherwise, put it after @target. | |
3461 | */ | |
d812796e | 3462 | if (worker) { |
affee4b2 | 3463 | head = worker->scheduled.next; |
d812796e LJ |
3464 | work_color = worker->current_color; |
3465 | } else { | |
affee4b2 TH |
3466 | unsigned long *bits = work_data_bits(target); |
3467 | ||
3468 | head = target->entry.next; | |
3469 | /* there can already be other linked works, inherit and set */ | |
d21cece0 | 3470 | work_flags |= *bits & WORK_STRUCT_LINKED; |
d812796e | 3471 | work_color = get_work_color(*bits); |
affee4b2 TH |
3472 | __set_bit(WORK_STRUCT_LINKED_BIT, bits); |
3473 | } | |
3474 | ||
d812796e LJ |
3475 | pwq->nr_in_flight[work_color]++; |
3476 | work_flags |= work_color_to_flags(work_color); | |
3477 | ||
d21cece0 | 3478 | insert_work(pwq, &barr->work, head, work_flags); |
fc2e4d70 ON |
3479 | } |
3480 | ||
73f53c4a | 3481 | /** |
112202d9 | 3482 | * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing |
73f53c4a TH |
3483 | * @wq: workqueue being flushed |
3484 | * @flush_color: new flush color, < 0 for no-op | |
3485 | * @work_color: new work color, < 0 for no-op | |
3486 | * | |
112202d9 | 3487 | * Prepare pwqs for workqueue flushing. |
73f53c4a | 3488 | * |
112202d9 TH |
3489 | * If @flush_color is non-negative, flush_color on all pwqs should be |
3490 | * -1. If no pwq has in-flight commands at the specified color, all | |
3491 | * pwq->flush_color's stay at -1 and %false is returned. If any pwq | |
3492 | * has in flight commands, its pwq->flush_color is set to | |
3493 | * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq | |
73f53c4a TH |
3494 | * wakeup logic is armed and %true is returned. |
3495 | * | |
3496 | * The caller should have initialized @wq->first_flusher prior to | |
3497 | * calling this function with non-negative @flush_color. If | |
3498 | * @flush_color is negative, no flush color update is done and %false | |
3499 | * is returned. | |
3500 | * | |
112202d9 | 3501 | * If @work_color is non-negative, all pwqs should have the same |
73f53c4a TH |
3502 | * work_color which is previous to @work_color and all will be |
3503 | * advanced to @work_color. | |
3504 | * | |
3505 | * CONTEXT: | |
3c25a55d | 3506 | * mutex_lock(wq->mutex). |
73f53c4a | 3507 | * |
d185af30 | 3508 | * Return: |
73f53c4a TH |
3509 | * %true if @flush_color >= 0 and there's something to flush. %false |
3510 | * otherwise. | |
3511 | */ | |
112202d9 | 3512 | static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, |
73f53c4a | 3513 | int flush_color, int work_color) |
1da177e4 | 3514 | { |
73f53c4a | 3515 | bool wait = false; |
49e3cf44 | 3516 | struct pool_workqueue *pwq; |
1da177e4 | 3517 | |
73f53c4a | 3518 | if (flush_color >= 0) { |
6183c009 | 3519 | WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); |
112202d9 | 3520 | atomic_set(&wq->nr_pwqs_to_flush, 1); |
1da177e4 | 3521 | } |
2355b70f | 3522 | |
49e3cf44 | 3523 | for_each_pwq(pwq, wq) { |
112202d9 | 3524 | struct worker_pool *pool = pwq->pool; |
fc2e4d70 | 3525 | |
a9b8a985 | 3526 | raw_spin_lock_irq(&pool->lock); |
83c22520 | 3527 | |
73f53c4a | 3528 | if (flush_color >= 0) { |
6183c009 | 3529 | WARN_ON_ONCE(pwq->flush_color != -1); |
fc2e4d70 | 3530 | |
112202d9 TH |
3531 | if (pwq->nr_in_flight[flush_color]) { |
3532 | pwq->flush_color = flush_color; | |
3533 | atomic_inc(&wq->nr_pwqs_to_flush); | |
73f53c4a TH |
3534 | wait = true; |
3535 | } | |
3536 | } | |
1da177e4 | 3537 | |
73f53c4a | 3538 | if (work_color >= 0) { |
6183c009 | 3539 | WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); |
112202d9 | 3540 | pwq->work_color = work_color; |
73f53c4a | 3541 | } |
1da177e4 | 3542 | |
a9b8a985 | 3543 | raw_spin_unlock_irq(&pool->lock); |
1da177e4 | 3544 | } |
2355b70f | 3545 | |
112202d9 | 3546 | if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) |
73f53c4a | 3547 | complete(&wq->first_flusher->done); |
14441960 | 3548 | |
73f53c4a | 3549 | return wait; |
1da177e4 LT |
3550 | } |
3551 | ||
0fcb78c2 | 3552 | /** |
c4f135d6 | 3553 | * __flush_workqueue - ensure that any scheduled work has run to completion. |
0fcb78c2 | 3554 | * @wq: workqueue to flush |
1da177e4 | 3555 | * |
c5aa87bb TH |
3556 | * This function sleeps until all work items which were queued on entry |
3557 | * have finished execution, but it is not livelocked by new incoming ones. | |
1da177e4 | 3558 | */ |
c4f135d6 | 3559 | void __flush_workqueue(struct workqueue_struct *wq) |
1da177e4 | 3560 | { |
73f53c4a TH |
3561 | struct wq_flusher this_flusher = { |
3562 | .list = LIST_HEAD_INIT(this_flusher.list), | |
3563 | .flush_color = -1, | |
fd1a5b04 | 3564 | .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map), |
73f53c4a TH |
3565 | }; |
3566 | int next_color; | |
1da177e4 | 3567 | |
3347fa09 TH |
3568 | if (WARN_ON(!wq_online)) |
3569 | return; | |
3570 | ||
87915adc JB |
3571 | lock_map_acquire(&wq->lockdep_map); |
3572 | lock_map_release(&wq->lockdep_map); | |
3573 | ||
3c25a55d | 3574 | mutex_lock(&wq->mutex); |
73f53c4a TH |
3575 | |
3576 | /* | |
3577 | * Start-to-wait phase | |
3578 | */ | |
3579 | next_color = work_next_color(wq->work_color); | |
3580 | ||
3581 | if (next_color != wq->flush_color) { | |
3582 | /* | |
3583 | * Color space is not full. The current work_color | |
3584 | * becomes our flush_color and work_color is advanced | |
3585 | * by one. | |
3586 | */ | |
6183c009 | 3587 | WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); |
73f53c4a TH |
3588 | this_flusher.flush_color = wq->work_color; |
3589 | wq->work_color = next_color; | |
3590 | ||
3591 | if (!wq->first_flusher) { | |
3592 | /* no flush in progress, become the first flusher */ | |
6183c009 | 3593 | WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); |
73f53c4a TH |
3594 | |
3595 | wq->first_flusher = &this_flusher; | |
3596 | ||
112202d9 | 3597 | if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, |
73f53c4a TH |
3598 | wq->work_color)) { |
3599 | /* nothing to flush, done */ | |
3600 | wq->flush_color = next_color; | |
3601 | wq->first_flusher = NULL; | |
3602 | goto out_unlock; | |
3603 | } | |
3604 | } else { | |
3605 | /* wait in queue */ | |
6183c009 | 3606 | WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); |
73f53c4a | 3607 | list_add_tail(&this_flusher.list, &wq->flusher_queue); |
112202d9 | 3608 | flush_workqueue_prep_pwqs(wq, -1, wq->work_color); |
73f53c4a TH |
3609 | } |
3610 | } else { | |
3611 | /* | |
3612 | * Oops, color space is full, wait on overflow queue. | |
3613 | * The next flush completion will assign us | |
3614 | * flush_color and transfer to flusher_queue. | |
3615 | */ | |
3616 | list_add_tail(&this_flusher.list, &wq->flusher_overflow); | |
3617 | } | |
3618 | ||
fca839c0 TH |
3619 | check_flush_dependency(wq, NULL); |
3620 | ||
3c25a55d | 3621 | mutex_unlock(&wq->mutex); |
73f53c4a TH |
3622 | |
3623 | wait_for_completion(&this_flusher.done); | |
3624 | ||
3625 | /* | |
3626 | * Wake-up-and-cascade phase | |
3627 | * | |
3628 | * First flushers are responsible for cascading flushes and | |
3629 | * handling overflow. Non-first flushers can simply return. | |
3630 | */ | |
00d5d15b | 3631 | if (READ_ONCE(wq->first_flusher) != &this_flusher) |
73f53c4a TH |
3632 | return; |
3633 | ||
3c25a55d | 3634 | mutex_lock(&wq->mutex); |
73f53c4a | 3635 | |
4ce48b37 TH |
3636 | /* we might have raced, check again with mutex held */ |
3637 | if (wq->first_flusher != &this_flusher) | |
3638 | goto out_unlock; | |
3639 | ||
00d5d15b | 3640 | WRITE_ONCE(wq->first_flusher, NULL); |
73f53c4a | 3641 | |
6183c009 TH |
3642 | WARN_ON_ONCE(!list_empty(&this_flusher.list)); |
3643 | WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); | |
73f53c4a TH |
3644 | |
3645 | while (true) { | |
3646 | struct wq_flusher *next, *tmp; | |
3647 | ||
3648 | /* complete all the flushers sharing the current flush color */ | |
3649 | list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { | |
3650 | if (next->flush_color != wq->flush_color) | |
3651 | break; | |
3652 | list_del_init(&next->list); | |
3653 | complete(&next->done); | |
3654 | } | |
3655 | ||
6183c009 TH |
3656 | WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && |
3657 | wq->flush_color != work_next_color(wq->work_color)); | |
73f53c4a TH |
3658 | |
3659 | /* this flush_color is finished, advance by one */ | |
3660 | wq->flush_color = work_next_color(wq->flush_color); | |
3661 | ||
3662 | /* one color has been freed, handle overflow queue */ | |
3663 | if (!list_empty(&wq->flusher_overflow)) { | |
3664 | /* | |
3665 | * Assign the same color to all overflowed | |
3666 | * flushers, advance work_color and append to | |
3667 | * flusher_queue. This is the start-to-wait | |
3668 | * phase for these overflowed flushers. | |
3669 | */ | |
3670 | list_for_each_entry(tmp, &wq->flusher_overflow, list) | |
3671 | tmp->flush_color = wq->work_color; | |
3672 | ||
3673 | wq->work_color = work_next_color(wq->work_color); | |
3674 | ||
3675 | list_splice_tail_init(&wq->flusher_overflow, | |
3676 | &wq->flusher_queue); | |
112202d9 | 3677 | flush_workqueue_prep_pwqs(wq, -1, wq->work_color); |
73f53c4a TH |
3678 | } |
3679 | ||
3680 | if (list_empty(&wq->flusher_queue)) { | |
6183c009 | 3681 | WARN_ON_ONCE(wq->flush_color != wq->work_color); |
73f53c4a TH |
3682 | break; |
3683 | } | |
3684 | ||
3685 | /* | |
3686 | * Need to flush more colors. Make the next flusher | |
112202d9 | 3687 | * the new first flusher and arm pwqs. |
73f53c4a | 3688 | */ |
6183c009 TH |
3689 | WARN_ON_ONCE(wq->flush_color == wq->work_color); |
3690 | WARN_ON_ONCE(wq->flush_color != next->flush_color); | |
73f53c4a TH |
3691 | |
3692 | list_del_init(&next->list); | |
3693 | wq->first_flusher = next; | |
3694 | ||
112202d9 | 3695 | if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) |
73f53c4a TH |
3696 | break; |
3697 | ||
3698 | /* | |
3699 | * Meh... this color is already done, clear first | |
3700 | * flusher and repeat cascading. | |
3701 | */ | |
3702 | wq->first_flusher = NULL; | |
3703 | } | |
3704 | ||
3705 | out_unlock: | |
3c25a55d | 3706 | mutex_unlock(&wq->mutex); |
1da177e4 | 3707 | } |
c4f135d6 | 3708 | EXPORT_SYMBOL(__flush_workqueue); |
1da177e4 | 3709 | |
9c5a2ba7 TH |
3710 | /** |
3711 | * drain_workqueue - drain a workqueue | |
3712 | * @wq: workqueue to drain | |
3713 | * | |
3714 | * Wait until the workqueue becomes empty. While draining is in progress, | |
3715 | * only chain queueing is allowed. IOW, only currently pending or running | |
3716 | * work items on @wq can queue further work items on it. @wq is flushed | |
b749b1b6 | 3717 | * repeatedly until it becomes empty. The number of flushing is determined |
9c5a2ba7 TH |
3718 | * by the depth of chaining and should be relatively short. Whine if it |
3719 | * takes too long. | |
3720 | */ | |
3721 | void drain_workqueue(struct workqueue_struct *wq) | |
3722 | { | |
3723 | unsigned int flush_cnt = 0; | |
49e3cf44 | 3724 | struct pool_workqueue *pwq; |
9c5a2ba7 TH |
3725 | |
3726 | /* | |
3727 | * __queue_work() needs to test whether there are drainers, is much | |
3728 | * hotter than drain_workqueue() and already looks at @wq->flags. | |
618b01eb | 3729 | * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. |
9c5a2ba7 | 3730 | */ |
87fc741e | 3731 | mutex_lock(&wq->mutex); |
9c5a2ba7 | 3732 | if (!wq->nr_drainers++) |
618b01eb | 3733 | wq->flags |= __WQ_DRAINING; |
87fc741e | 3734 | mutex_unlock(&wq->mutex); |
9c5a2ba7 | 3735 | reflush: |
c4f135d6 | 3736 | __flush_workqueue(wq); |
9c5a2ba7 | 3737 | |
b09f4fd3 | 3738 | mutex_lock(&wq->mutex); |
76af4d93 | 3739 | |
49e3cf44 | 3740 | for_each_pwq(pwq, wq) { |
fa2563e4 | 3741 | bool drained; |
9c5a2ba7 | 3742 | |
a9b8a985 | 3743 | raw_spin_lock_irq(&pwq->pool->lock); |
afa87ce8 | 3744 | drained = pwq_is_empty(pwq); |
a9b8a985 | 3745 | raw_spin_unlock_irq(&pwq->pool->lock); |
fa2563e4 TT |
3746 | |
3747 | if (drained) | |
9c5a2ba7 TH |
3748 | continue; |
3749 | ||
3750 | if (++flush_cnt == 10 || | |
3751 | (flush_cnt % 100 == 0 && flush_cnt <= 1000)) | |
e9ad2eb3 SZ |
3752 | pr_warn("workqueue %s: %s() isn't complete after %u tries\n", |
3753 | wq->name, __func__, flush_cnt); | |
76af4d93 | 3754 | |
b09f4fd3 | 3755 | mutex_unlock(&wq->mutex); |
9c5a2ba7 TH |
3756 | goto reflush; |
3757 | } | |
3758 | ||
9c5a2ba7 | 3759 | if (!--wq->nr_drainers) |
618b01eb | 3760 | wq->flags &= ~__WQ_DRAINING; |
87fc741e | 3761 | mutex_unlock(&wq->mutex); |
9c5a2ba7 TH |
3762 | } |
3763 | EXPORT_SYMBOL_GPL(drain_workqueue); | |
3764 | ||
d6e89786 JB |
3765 | static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr, |
3766 | bool from_cancel) | |
db700897 | 3767 | { |
affee4b2 | 3768 | struct worker *worker = NULL; |
c9e7cf27 | 3769 | struct worker_pool *pool; |
112202d9 | 3770 | struct pool_workqueue *pwq; |
db700897 ON |
3771 | |
3772 | might_sleep(); | |
fa1b54e6 | 3773 | |
24acfb71 | 3774 | rcu_read_lock(); |
c9e7cf27 | 3775 | pool = get_work_pool(work); |
fa1b54e6 | 3776 | if (!pool) { |
24acfb71 | 3777 | rcu_read_unlock(); |
baf59022 | 3778 | return false; |
fa1b54e6 | 3779 | } |
db700897 | 3780 | |
a9b8a985 | 3781 | raw_spin_lock_irq(&pool->lock); |
0b3dae68 | 3782 | /* see the comment in try_to_grab_pending() with the same code */ |
112202d9 TH |
3783 | pwq = get_work_pwq(work); |
3784 | if (pwq) { | |
3785 | if (unlikely(pwq->pool != pool)) | |
4690c4ab | 3786 | goto already_gone; |
606a5020 | 3787 | } else { |
c9e7cf27 | 3788 | worker = find_worker_executing_work(pool, work); |
affee4b2 | 3789 | if (!worker) |
4690c4ab | 3790 | goto already_gone; |
112202d9 | 3791 | pwq = worker->current_pwq; |
606a5020 | 3792 | } |
db700897 | 3793 | |
fca839c0 TH |
3794 | check_flush_dependency(pwq->wq, work); |
3795 | ||
112202d9 | 3796 | insert_wq_barrier(pwq, barr, work, worker); |
a9b8a985 | 3797 | raw_spin_unlock_irq(&pool->lock); |
7a22ad75 | 3798 | |
e159489b | 3799 | /* |
a1d14934 PZ |
3800 | * Force a lock recursion deadlock when using flush_work() inside a |
3801 | * single-threaded or rescuer equipped workqueue. | |
3802 | * | |
3803 | * For single threaded workqueues the deadlock happens when the work | |
3804 | * is after the work issuing the flush_work(). For rescuer equipped | |
3805 | * workqueues the deadlock happens when the rescuer stalls, blocking | |
3806 | * forward progress. | |
e159489b | 3807 | */ |
d6e89786 JB |
3808 | if (!from_cancel && |
3809 | (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) { | |
112202d9 | 3810 | lock_map_acquire(&pwq->wq->lockdep_map); |
a1d14934 PZ |
3811 | lock_map_release(&pwq->wq->lockdep_map); |
3812 | } | |
24acfb71 | 3813 | rcu_read_unlock(); |
401a8d04 | 3814 | return true; |
4690c4ab | 3815 | already_gone: |
a9b8a985 | 3816 | raw_spin_unlock_irq(&pool->lock); |
24acfb71 | 3817 | rcu_read_unlock(); |
401a8d04 | 3818 | return false; |
db700897 | 3819 | } |
baf59022 | 3820 | |
d6e89786 JB |
3821 | static bool __flush_work(struct work_struct *work, bool from_cancel) |
3822 | { | |
3823 | struct wq_barrier barr; | |
3824 | ||
3825 | if (WARN_ON(!wq_online)) | |
3826 | return false; | |
3827 | ||
4d43d395 TH |
3828 | if (WARN_ON(!work->func)) |
3829 | return false; | |
3830 | ||
c0feea59 TH |
3831 | lock_map_acquire(&work->lockdep_map); |
3832 | lock_map_release(&work->lockdep_map); | |
87915adc | 3833 | |
d6e89786 JB |
3834 | if (start_flush_work(work, &barr, from_cancel)) { |
3835 | wait_for_completion(&barr.done); | |
3836 | destroy_work_on_stack(&barr.work); | |
3837 | return true; | |
3838 | } else { | |
3839 | return false; | |
3840 | } | |
3841 | } | |
3842 | ||
baf59022 TH |
3843 | /** |
3844 | * flush_work - wait for a work to finish executing the last queueing instance | |
3845 | * @work: the work to flush | |
3846 | * | |
606a5020 TH |
3847 | * Wait until @work has finished execution. @work is guaranteed to be idle |
3848 | * on return if it hasn't been requeued since flush started. | |
baf59022 | 3849 | * |
d185af30 | 3850 | * Return: |
baf59022 TH |
3851 | * %true if flush_work() waited for the work to finish execution, |
3852 | * %false if it was already idle. | |
3853 | */ | |
3854 | bool flush_work(struct work_struct *work) | |
3855 | { | |
d6e89786 | 3856 | return __flush_work(work, false); |
6e84d644 | 3857 | } |
606a5020 | 3858 | EXPORT_SYMBOL_GPL(flush_work); |
6e84d644 | 3859 | |
8603e1b3 | 3860 | struct cwt_wait { |
ac6424b9 | 3861 | wait_queue_entry_t wait; |
8603e1b3 TH |
3862 | struct work_struct *work; |
3863 | }; | |
3864 | ||
ac6424b9 | 3865 | static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key) |
8603e1b3 TH |
3866 | { |
3867 | struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait); | |
3868 | ||
3869 | if (cwait->work != key) | |
3870 | return 0; | |
3871 | return autoremove_wake_function(wait, mode, sync, key); | |
3872 | } | |
3873 | ||
36e227d2 | 3874 | static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) |
1f1f642e | 3875 | { |
8603e1b3 | 3876 | static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq); |
bbb68dfa | 3877 | unsigned long flags; |
1f1f642e ON |
3878 | int ret; |
3879 | ||
3880 | do { | |
bbb68dfa TH |
3881 | ret = try_to_grab_pending(work, is_dwork, &flags); |
3882 | /* | |
8603e1b3 TH |
3883 | * If someone else is already canceling, wait for it to |
3884 | * finish. flush_work() doesn't work for PREEMPT_NONE | |
3885 | * because we may get scheduled between @work's completion | |
3886 | * and the other canceling task resuming and clearing | |
3887 | * CANCELING - flush_work() will return false immediately | |
3888 | * as @work is no longer busy, try_to_grab_pending() will | |
3889 | * return -ENOENT as @work is still being canceled and the | |
3890 | * other canceling task won't be able to clear CANCELING as | |
3891 | * we're hogging the CPU. | |
3892 | * | |
3893 | * Let's wait for completion using a waitqueue. As this | |
3894 | * may lead to the thundering herd problem, use a custom | |
3895 | * wake function which matches @work along with exclusive | |
3896 | * wait and wakeup. | |
bbb68dfa | 3897 | */ |
8603e1b3 TH |
3898 | if (unlikely(ret == -ENOENT)) { |
3899 | struct cwt_wait cwait; | |
3900 | ||
3901 | init_wait(&cwait.wait); | |
3902 | cwait.wait.func = cwt_wakefn; | |
3903 | cwait.work = work; | |
3904 | ||
3905 | prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait, | |
3906 | TASK_UNINTERRUPTIBLE); | |
3907 | if (work_is_canceling(work)) | |
3908 | schedule(); | |
3909 | finish_wait(&cancel_waitq, &cwait.wait); | |
3910 | } | |
1f1f642e ON |
3911 | } while (unlikely(ret < 0)); |
3912 | ||
bbb68dfa TH |
3913 | /* tell other tasks trying to grab @work to back off */ |
3914 | mark_work_canceling(work); | |
3915 | local_irq_restore(flags); | |
3916 | ||
3347fa09 TH |
3917 | /* |
3918 | * This allows canceling during early boot. We know that @work | |
3919 | * isn't executing. | |
3920 | */ | |
3921 | if (wq_online) | |
d6e89786 | 3922 | __flush_work(work, true); |
3347fa09 | 3923 | |
7a22ad75 | 3924 | clear_work_data(work); |
8603e1b3 TH |
3925 | |
3926 | /* | |
3927 | * Paired with prepare_to_wait() above so that either | |
3928 | * waitqueue_active() is visible here or !work_is_canceling() is | |
3929 | * visible there. | |
3930 | */ | |
3931 | smp_mb(); | |
3932 | if (waitqueue_active(&cancel_waitq)) | |
3933 | __wake_up(&cancel_waitq, TASK_NORMAL, 1, work); | |
3934 | ||
1f1f642e ON |
3935 | return ret; |
3936 | } | |
3937 | ||
6e84d644 | 3938 | /** |
401a8d04 TH |
3939 | * cancel_work_sync - cancel a work and wait for it to finish |
3940 | * @work: the work to cancel | |
6e84d644 | 3941 | * |
401a8d04 TH |
3942 | * Cancel @work and wait for its execution to finish. This function |
3943 | * can be used even if the work re-queues itself or migrates to | |
3944 | * another workqueue. On return from this function, @work is | |
3945 | * guaranteed to be not pending or executing on any CPU. | |
1f1f642e | 3946 | * |
401a8d04 TH |
3947 | * cancel_work_sync(&delayed_work->work) must not be used for |
3948 | * delayed_work's. Use cancel_delayed_work_sync() instead. | |
6e84d644 | 3949 | * |
401a8d04 | 3950 | * The caller must ensure that the workqueue on which @work was last |
6e84d644 | 3951 | * queued can't be destroyed before this function returns. |
401a8d04 | 3952 | * |
d185af30 | 3953 | * Return: |
401a8d04 | 3954 | * %true if @work was pending, %false otherwise. |
6e84d644 | 3955 | */ |
401a8d04 | 3956 | bool cancel_work_sync(struct work_struct *work) |
6e84d644 | 3957 | { |
36e227d2 | 3958 | return __cancel_work_timer(work, false); |
b89deed3 | 3959 | } |
28e53bdd | 3960 | EXPORT_SYMBOL_GPL(cancel_work_sync); |
b89deed3 | 3961 | |
6e84d644 | 3962 | /** |
401a8d04 TH |
3963 | * flush_delayed_work - wait for a dwork to finish executing the last queueing |
3964 | * @dwork: the delayed work to flush | |
6e84d644 | 3965 | * |
401a8d04 TH |
3966 | * Delayed timer is cancelled and the pending work is queued for |
3967 | * immediate execution. Like flush_work(), this function only | |
3968 | * considers the last queueing instance of @dwork. | |
1f1f642e | 3969 | * |
d185af30 | 3970 | * Return: |
401a8d04 TH |
3971 | * %true if flush_work() waited for the work to finish execution, |
3972 | * %false if it was already idle. | |
6e84d644 | 3973 | */ |
401a8d04 TH |
3974 | bool flush_delayed_work(struct delayed_work *dwork) |
3975 | { | |
8930caba | 3976 | local_irq_disable(); |
401a8d04 | 3977 | if (del_timer_sync(&dwork->timer)) |
60c057bc | 3978 | __queue_work(dwork->cpu, dwork->wq, &dwork->work); |
8930caba | 3979 | local_irq_enable(); |
401a8d04 TH |
3980 | return flush_work(&dwork->work); |
3981 | } | |
3982 | EXPORT_SYMBOL(flush_delayed_work); | |
3983 | ||
05f0fe6b TH |
3984 | /** |
3985 | * flush_rcu_work - wait for a rwork to finish executing the last queueing | |
3986 | * @rwork: the rcu work to flush | |
3987 | * | |
3988 | * Return: | |
3989 | * %true if flush_rcu_work() waited for the work to finish execution, | |
3990 | * %false if it was already idle. | |
3991 | */ | |
3992 | bool flush_rcu_work(struct rcu_work *rwork) | |
3993 | { | |
3994 | if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) { | |
3995 | rcu_barrier(); | |
3996 | flush_work(&rwork->work); | |
3997 | return true; | |
3998 | } else { | |
3999 | return flush_work(&rwork->work); | |
4000 | } | |
4001 | } | |
4002 | EXPORT_SYMBOL(flush_rcu_work); | |
4003 | ||
f72b8792 JA |
4004 | static bool __cancel_work(struct work_struct *work, bool is_dwork) |
4005 | { | |
4006 | unsigned long flags; | |
4007 | int ret; | |
4008 | ||
4009 | do { | |
4010 | ret = try_to_grab_pending(work, is_dwork, &flags); | |
4011 | } while (unlikely(ret == -EAGAIN)); | |
4012 | ||
4013 | if (unlikely(ret < 0)) | |
4014 | return false; | |
4015 | ||
4016 | set_work_pool_and_clear_pending(work, get_work_pool_id(work)); | |
4017 | local_irq_restore(flags); | |
4018 | return ret; | |
4019 | } | |
4020 | ||
73b4b532 AG |
4021 | /* |
4022 | * See cancel_delayed_work() | |
4023 | */ | |
4024 | bool cancel_work(struct work_struct *work) | |
4025 | { | |
4026 | return __cancel_work(work, false); | |
4027 | } | |
4028 | EXPORT_SYMBOL(cancel_work); | |
4029 | ||
09383498 | 4030 | /** |
57b30ae7 TH |
4031 | * cancel_delayed_work - cancel a delayed work |
4032 | * @dwork: delayed_work to cancel | |
09383498 | 4033 | * |
d185af30 YB |
4034 | * Kill off a pending delayed_work. |
4035 | * | |
4036 | * Return: %true if @dwork was pending and canceled; %false if it wasn't | |
4037 | * pending. | |
4038 | * | |
4039 | * Note: | |
4040 | * The work callback function may still be running on return, unless | |
4041 | * it returns %true and the work doesn't re-arm itself. Explicitly flush or | |
4042 | * use cancel_delayed_work_sync() to wait on it. | |
09383498 | 4043 | * |
57b30ae7 | 4044 | * This function is safe to call from any context including IRQ handler. |
09383498 | 4045 | */ |
57b30ae7 | 4046 | bool cancel_delayed_work(struct delayed_work *dwork) |
09383498 | 4047 | { |
f72b8792 | 4048 | return __cancel_work(&dwork->work, true); |
09383498 | 4049 | } |
57b30ae7 | 4050 | EXPORT_SYMBOL(cancel_delayed_work); |
09383498 | 4051 | |
401a8d04 TH |
4052 | /** |
4053 | * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish | |
4054 | * @dwork: the delayed work cancel | |
4055 | * | |
4056 | * This is cancel_work_sync() for delayed works. | |
4057 | * | |
d185af30 | 4058 | * Return: |
401a8d04 TH |
4059 | * %true if @dwork was pending, %false otherwise. |
4060 | */ | |
4061 | bool cancel_delayed_work_sync(struct delayed_work *dwork) | |
6e84d644 | 4062 | { |
36e227d2 | 4063 | return __cancel_work_timer(&dwork->work, true); |
6e84d644 | 4064 | } |
f5a421a4 | 4065 | EXPORT_SYMBOL(cancel_delayed_work_sync); |
1da177e4 | 4066 | |
b6136773 | 4067 | /** |
31ddd871 | 4068 | * schedule_on_each_cpu - execute a function synchronously on each online CPU |
b6136773 | 4069 | * @func: the function to call |
b6136773 | 4070 | * |
31ddd871 TH |
4071 | * schedule_on_each_cpu() executes @func on each online CPU using the |
4072 | * system workqueue and blocks until all CPUs have completed. | |
b6136773 | 4073 | * schedule_on_each_cpu() is very slow. |
31ddd871 | 4074 | * |
d185af30 | 4075 | * Return: |
31ddd871 | 4076 | * 0 on success, -errno on failure. |
b6136773 | 4077 | */ |
65f27f38 | 4078 | int schedule_on_each_cpu(work_func_t func) |
15316ba8 CL |
4079 | { |
4080 | int cpu; | |
38f51568 | 4081 | struct work_struct __percpu *works; |
15316ba8 | 4082 | |
b6136773 AM |
4083 | works = alloc_percpu(struct work_struct); |
4084 | if (!works) | |
15316ba8 | 4085 | return -ENOMEM; |
b6136773 | 4086 | |
ffd8bea8 | 4087 | cpus_read_lock(); |
93981800 | 4088 | |
15316ba8 | 4089 | for_each_online_cpu(cpu) { |
9bfb1839 IM |
4090 | struct work_struct *work = per_cpu_ptr(works, cpu); |
4091 | ||
4092 | INIT_WORK(work, func); | |
b71ab8c2 | 4093 | schedule_work_on(cpu, work); |
65a64464 | 4094 | } |
93981800 TH |
4095 | |
4096 | for_each_online_cpu(cpu) | |
4097 | flush_work(per_cpu_ptr(works, cpu)); | |
4098 | ||
ffd8bea8 | 4099 | cpus_read_unlock(); |
b6136773 | 4100 | free_percpu(works); |
15316ba8 CL |
4101 | return 0; |
4102 | } | |
4103 | ||
1fa44eca JB |
4104 | /** |
4105 | * execute_in_process_context - reliably execute the routine with user context | |
4106 | * @fn: the function to execute | |
1fa44eca JB |
4107 | * @ew: guaranteed storage for the execute work structure (must |
4108 | * be available when the work executes) | |
4109 | * | |
4110 | * Executes the function immediately if process context is available, | |
4111 | * otherwise schedules the function for delayed execution. | |
4112 | * | |
d185af30 | 4113 | * Return: 0 - function was executed |
1fa44eca JB |
4114 | * 1 - function was scheduled for execution |
4115 | */ | |
65f27f38 | 4116 | int execute_in_process_context(work_func_t fn, struct execute_work *ew) |
1fa44eca JB |
4117 | { |
4118 | if (!in_interrupt()) { | |
65f27f38 | 4119 | fn(&ew->work); |
1fa44eca JB |
4120 | return 0; |
4121 | } | |
4122 | ||
65f27f38 | 4123 | INIT_WORK(&ew->work, fn); |
1fa44eca JB |
4124 | schedule_work(&ew->work); |
4125 | ||
4126 | return 1; | |
4127 | } | |
4128 | EXPORT_SYMBOL_GPL(execute_in_process_context); | |
4129 | ||
6ba94429 FW |
4130 | /** |
4131 | * free_workqueue_attrs - free a workqueue_attrs | |
4132 | * @attrs: workqueue_attrs to free | |
226223ab | 4133 | * |
6ba94429 | 4134 | * Undo alloc_workqueue_attrs(). |
226223ab | 4135 | */ |
513c98d0 | 4136 | void free_workqueue_attrs(struct workqueue_attrs *attrs) |
226223ab | 4137 | { |
6ba94429 FW |
4138 | if (attrs) { |
4139 | free_cpumask_var(attrs->cpumask); | |
9546b29e | 4140 | free_cpumask_var(attrs->__pod_cpumask); |
6ba94429 FW |
4141 | kfree(attrs); |
4142 | } | |
226223ab TH |
4143 | } |
4144 | ||
6ba94429 FW |
4145 | /** |
4146 | * alloc_workqueue_attrs - allocate a workqueue_attrs | |
6ba94429 FW |
4147 | * |
4148 | * Allocate a new workqueue_attrs, initialize with default settings and | |
4149 | * return it. | |
4150 | * | |
4151 | * Return: The allocated new workqueue_attr on success. %NULL on failure. | |
4152 | */ | |
513c98d0 | 4153 | struct workqueue_attrs *alloc_workqueue_attrs(void) |
226223ab | 4154 | { |
6ba94429 | 4155 | struct workqueue_attrs *attrs; |
226223ab | 4156 | |
be69d00d | 4157 | attrs = kzalloc(sizeof(*attrs), GFP_KERNEL); |
6ba94429 FW |
4158 | if (!attrs) |
4159 | goto fail; | |
be69d00d | 4160 | if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL)) |
6ba94429 | 4161 | goto fail; |
9546b29e TH |
4162 | if (!alloc_cpumask_var(&attrs->__pod_cpumask, GFP_KERNEL)) |
4163 | goto fail; | |
6ba94429 FW |
4164 | |
4165 | cpumask_copy(attrs->cpumask, cpu_possible_mask); | |
523a301e | 4166 | attrs->affn_scope = WQ_AFFN_DFL; |
6ba94429 FW |
4167 | return attrs; |
4168 | fail: | |
4169 | free_workqueue_attrs(attrs); | |
4170 | return NULL; | |
226223ab TH |
4171 | } |
4172 | ||
6ba94429 FW |
4173 | static void copy_workqueue_attrs(struct workqueue_attrs *to, |
4174 | const struct workqueue_attrs *from) | |
226223ab | 4175 | { |
6ba94429 FW |
4176 | to->nice = from->nice; |
4177 | cpumask_copy(to->cpumask, from->cpumask); | |
9546b29e | 4178 | cpumask_copy(to->__pod_cpumask, from->__pod_cpumask); |
8639eceb | 4179 | to->affn_strict = from->affn_strict; |
84193c07 | 4180 | |
6ba94429 | 4181 | /* |
84193c07 TH |
4182 | * Unlike hash and equality test, copying shouldn't ignore wq-only |
4183 | * fields as copying is used for both pool and wq attrs. Instead, | |
4184 | * get_unbound_pool() explicitly clears the fields. | |
6ba94429 | 4185 | */ |
84193c07 | 4186 | to->affn_scope = from->affn_scope; |
af73f5c9 | 4187 | to->ordered = from->ordered; |
226223ab TH |
4188 | } |
4189 | ||
5de7a03c TH |
4190 | /* |
4191 | * Some attrs fields are workqueue-only. Clear them for worker_pool's. See the | |
4192 | * comments in 'struct workqueue_attrs' definition. | |
4193 | */ | |
4194 | static void wqattrs_clear_for_pool(struct workqueue_attrs *attrs) | |
4195 | { | |
84193c07 | 4196 | attrs->affn_scope = WQ_AFFN_NR_TYPES; |
5de7a03c TH |
4197 | attrs->ordered = false; |
4198 | } | |
4199 | ||
6ba94429 FW |
4200 | /* hash value of the content of @attr */ |
4201 | static u32 wqattrs_hash(const struct workqueue_attrs *attrs) | |
226223ab | 4202 | { |
6ba94429 | 4203 | u32 hash = 0; |
226223ab | 4204 | |
6ba94429 FW |
4205 | hash = jhash_1word(attrs->nice, hash); |
4206 | hash = jhash(cpumask_bits(attrs->cpumask), | |
4207 | BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); | |
9546b29e TH |
4208 | hash = jhash(cpumask_bits(attrs->__pod_cpumask), |
4209 | BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); | |
8639eceb | 4210 | hash = jhash_1word(attrs->affn_strict, hash); |
6ba94429 | 4211 | return hash; |
226223ab | 4212 | } |
226223ab | 4213 | |
6ba94429 FW |
4214 | /* content equality test */ |
4215 | static bool wqattrs_equal(const struct workqueue_attrs *a, | |
4216 | const struct workqueue_attrs *b) | |
226223ab | 4217 | { |
6ba94429 FW |
4218 | if (a->nice != b->nice) |
4219 | return false; | |
4220 | if (!cpumask_equal(a->cpumask, b->cpumask)) | |
4221 | return false; | |
9546b29e TH |
4222 | if (!cpumask_equal(a->__pod_cpumask, b->__pod_cpumask)) |
4223 | return false; | |
8639eceb TH |
4224 | if (a->affn_strict != b->affn_strict) |
4225 | return false; | |
6ba94429 | 4226 | return true; |
226223ab TH |
4227 | } |
4228 | ||
0f36ee24 TH |
4229 | /* Update @attrs with actually available CPUs */ |
4230 | static void wqattrs_actualize_cpumask(struct workqueue_attrs *attrs, | |
4231 | const cpumask_t *unbound_cpumask) | |
4232 | { | |
4233 | /* | |
4234 | * Calculate the effective CPU mask of @attrs given @unbound_cpumask. If | |
4235 | * @attrs->cpumask doesn't overlap with @unbound_cpumask, we fallback to | |
4236 | * @unbound_cpumask. | |
4237 | */ | |
4238 | cpumask_and(attrs->cpumask, attrs->cpumask, unbound_cpumask); | |
4239 | if (unlikely(cpumask_empty(attrs->cpumask))) | |
4240 | cpumask_copy(attrs->cpumask, unbound_cpumask); | |
4241 | } | |
4242 | ||
84193c07 TH |
4243 | /* find wq_pod_type to use for @attrs */ |
4244 | static const struct wq_pod_type * | |
4245 | wqattrs_pod_type(const struct workqueue_attrs *attrs) | |
4246 | { | |
523a301e TH |
4247 | enum wq_affn_scope scope; |
4248 | struct wq_pod_type *pt; | |
4249 | ||
4250 | /* to synchronize access to wq_affn_dfl */ | |
4251 | lockdep_assert_held(&wq_pool_mutex); | |
4252 | ||
4253 | if (attrs->affn_scope == WQ_AFFN_DFL) | |
4254 | scope = wq_affn_dfl; | |
4255 | else | |
4256 | scope = attrs->affn_scope; | |
4257 | ||
4258 | pt = &wq_pod_types[scope]; | |
84193c07 TH |
4259 | |
4260 | if (!WARN_ON_ONCE(attrs->affn_scope == WQ_AFFN_NR_TYPES) && | |
4261 | likely(pt->nr_pods)) | |
4262 | return pt; | |
4263 | ||
4264 | /* | |
4265 | * Before workqueue_init_topology(), only SYSTEM is available which is | |
4266 | * initialized in workqueue_init_early(). | |
4267 | */ | |
4268 | pt = &wq_pod_types[WQ_AFFN_SYSTEM]; | |
4269 | BUG_ON(!pt->nr_pods); | |
4270 | return pt; | |
4271 | } | |
4272 | ||
6ba94429 FW |
4273 | /** |
4274 | * init_worker_pool - initialize a newly zalloc'd worker_pool | |
4275 | * @pool: worker_pool to initialize | |
4276 | * | |
402dd89d | 4277 | * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs. |
6ba94429 FW |
4278 | * |
4279 | * Return: 0 on success, -errno on failure. Even on failure, all fields | |
4280 | * inside @pool proper are initialized and put_unbound_pool() can be called | |
4281 | * on @pool safely to release it. | |
4282 | */ | |
4283 | static int init_worker_pool(struct worker_pool *pool) | |
226223ab | 4284 | { |
a9b8a985 | 4285 | raw_spin_lock_init(&pool->lock); |
6ba94429 FW |
4286 | pool->id = -1; |
4287 | pool->cpu = -1; | |
4288 | pool->node = NUMA_NO_NODE; | |
4289 | pool->flags |= POOL_DISASSOCIATED; | |
82607adc | 4290 | pool->watchdog_ts = jiffies; |
6ba94429 FW |
4291 | INIT_LIST_HEAD(&pool->worklist); |
4292 | INIT_LIST_HEAD(&pool->idle_list); | |
4293 | hash_init(pool->busy_hash); | |
226223ab | 4294 | |
32a6c723 | 4295 | timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE); |
3f959aa3 | 4296 | INIT_WORK(&pool->idle_cull_work, idle_cull_fn); |
226223ab | 4297 | |
32a6c723 | 4298 | timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0); |
226223ab | 4299 | |
6ba94429 | 4300 | INIT_LIST_HEAD(&pool->workers); |
e02b9312 | 4301 | INIT_LIST_HEAD(&pool->dying_workers); |
226223ab | 4302 | |
6ba94429 FW |
4303 | ida_init(&pool->worker_ida); |
4304 | INIT_HLIST_NODE(&pool->hash_node); | |
4305 | pool->refcnt = 1; | |
226223ab | 4306 | |
6ba94429 | 4307 | /* shouldn't fail above this point */ |
be69d00d | 4308 | pool->attrs = alloc_workqueue_attrs(); |
6ba94429 FW |
4309 | if (!pool->attrs) |
4310 | return -ENOMEM; | |
5de7a03c TH |
4311 | |
4312 | wqattrs_clear_for_pool(pool->attrs); | |
4313 | ||
6ba94429 | 4314 | return 0; |
226223ab TH |
4315 | } |
4316 | ||
669de8bd BVA |
4317 | #ifdef CONFIG_LOCKDEP |
4318 | static void wq_init_lockdep(struct workqueue_struct *wq) | |
4319 | { | |
4320 | char *lock_name; | |
4321 | ||
4322 | lockdep_register_key(&wq->key); | |
4323 | lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name); | |
4324 | if (!lock_name) | |
4325 | lock_name = wq->name; | |
69a106c0 QC |
4326 | |
4327 | wq->lock_name = lock_name; | |
669de8bd BVA |
4328 | lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0); |
4329 | } | |
4330 | ||
4331 | static void wq_unregister_lockdep(struct workqueue_struct *wq) | |
4332 | { | |
4333 | lockdep_unregister_key(&wq->key); | |
4334 | } | |
4335 | ||
4336 | static void wq_free_lockdep(struct workqueue_struct *wq) | |
4337 | { | |
4338 | if (wq->lock_name != wq->name) | |
4339 | kfree(wq->lock_name); | |
4340 | } | |
4341 | #else | |
4342 | static void wq_init_lockdep(struct workqueue_struct *wq) | |
4343 | { | |
4344 | } | |
4345 | ||
4346 | static void wq_unregister_lockdep(struct workqueue_struct *wq) | |
4347 | { | |
4348 | } | |
4349 | ||
4350 | static void wq_free_lockdep(struct workqueue_struct *wq) | |
4351 | { | |
4352 | } | |
4353 | #endif | |
4354 | ||
91ccc6e7 TH |
4355 | static void free_node_nr_active(struct wq_node_nr_active **nna_ar) |
4356 | { | |
4357 | int node; | |
4358 | ||
4359 | for_each_node(node) { | |
4360 | kfree(nna_ar[node]); | |
4361 | nna_ar[node] = NULL; | |
4362 | } | |
4363 | ||
4364 | kfree(nna_ar[nr_node_ids]); | |
4365 | nna_ar[nr_node_ids] = NULL; | |
4366 | } | |
4367 | ||
4368 | static void init_node_nr_active(struct wq_node_nr_active *nna) | |
4369 | { | |
c5f8cd6c | 4370 | nna->max = WQ_DFL_MIN_ACTIVE; |
91ccc6e7 | 4371 | atomic_set(&nna->nr, 0); |
5797b1c1 TH |
4372 | raw_spin_lock_init(&nna->lock); |
4373 | INIT_LIST_HEAD(&nna->pending_pwqs); | |
91ccc6e7 TH |
4374 | } |
4375 | ||
4376 | /* | |
4377 | * Each node's nr_active counter will be accessed mostly from its own node and | |
4378 | * should be allocated in the node. | |
4379 | */ | |
4380 | static int alloc_node_nr_active(struct wq_node_nr_active **nna_ar) | |
4381 | { | |
4382 | struct wq_node_nr_active *nna; | |
4383 | int node; | |
4384 | ||
4385 | for_each_node(node) { | |
4386 | nna = kzalloc_node(sizeof(*nna), GFP_KERNEL, node); | |
4387 | if (!nna) | |
4388 | goto err_free; | |
4389 | init_node_nr_active(nna); | |
4390 | nna_ar[node] = nna; | |
4391 | } | |
4392 | ||
4393 | /* [nr_node_ids] is used as the fallback */ | |
4394 | nna = kzalloc_node(sizeof(*nna), GFP_KERNEL, NUMA_NO_NODE); | |
4395 | if (!nna) | |
4396 | goto err_free; | |
4397 | init_node_nr_active(nna); | |
4398 | nna_ar[nr_node_ids] = nna; | |
4399 | ||
4400 | return 0; | |
4401 | ||
4402 | err_free: | |
4403 | free_node_nr_active(nna_ar); | |
4404 | return -ENOMEM; | |
4405 | } | |
4406 | ||
6ba94429 | 4407 | static void rcu_free_wq(struct rcu_head *rcu) |
226223ab | 4408 | { |
6ba94429 FW |
4409 | struct workqueue_struct *wq = |
4410 | container_of(rcu, struct workqueue_struct, rcu); | |
226223ab | 4411 | |
91ccc6e7 TH |
4412 | if (wq->flags & WQ_UNBOUND) |
4413 | free_node_nr_active(wq->node_nr_active); | |
4414 | ||
669de8bd | 4415 | wq_free_lockdep(wq); |
636b927e TH |
4416 | free_percpu(wq->cpu_pwq); |
4417 | free_workqueue_attrs(wq->unbound_attrs); | |
6ba94429 | 4418 | kfree(wq); |
226223ab TH |
4419 | } |
4420 | ||
6ba94429 | 4421 | static void rcu_free_pool(struct rcu_head *rcu) |
226223ab | 4422 | { |
6ba94429 | 4423 | struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); |
226223ab | 4424 | |
6ba94429 FW |
4425 | ida_destroy(&pool->worker_ida); |
4426 | free_workqueue_attrs(pool->attrs); | |
4427 | kfree(pool); | |
226223ab TH |
4428 | } |
4429 | ||
6ba94429 FW |
4430 | /** |
4431 | * put_unbound_pool - put a worker_pool | |
4432 | * @pool: worker_pool to put | |
4433 | * | |
24acfb71 | 4434 | * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU |
6ba94429 FW |
4435 | * safe manner. get_unbound_pool() calls this function on its failure path |
4436 | * and this function should be able to release pools which went through, | |
4437 | * successfully or not, init_worker_pool(). | |
4438 | * | |
4439 | * Should be called with wq_pool_mutex held. | |
4440 | */ | |
4441 | static void put_unbound_pool(struct worker_pool *pool) | |
226223ab | 4442 | { |
6ba94429 FW |
4443 | DECLARE_COMPLETION_ONSTACK(detach_completion); |
4444 | struct worker *worker; | |
9680540c | 4445 | LIST_HEAD(cull_list); |
e02b9312 | 4446 | |
6ba94429 | 4447 | lockdep_assert_held(&wq_pool_mutex); |
226223ab | 4448 | |
6ba94429 FW |
4449 | if (--pool->refcnt) |
4450 | return; | |
226223ab | 4451 | |
6ba94429 FW |
4452 | /* sanity checks */ |
4453 | if (WARN_ON(!(pool->cpu < 0)) || | |
4454 | WARN_ON(!list_empty(&pool->worklist))) | |
4455 | return; | |
226223ab | 4456 | |
6ba94429 FW |
4457 | /* release id and unhash */ |
4458 | if (pool->id >= 0) | |
4459 | idr_remove(&worker_pool_idr, pool->id); | |
4460 | hash_del(&pool->hash_node); | |
d55262c4 | 4461 | |
6ba94429 | 4462 | /* |
692b4825 TH |
4463 | * Become the manager and destroy all workers. This prevents |
4464 | * @pool's workers from blocking on attach_mutex. We're the last | |
4465 | * manager and @pool gets freed with the flag set. | |
9ab03be4 VS |
4466 | * |
4467 | * Having a concurrent manager is quite unlikely to happen as we can | |
4468 | * only get here with | |
4469 | * pwq->refcnt == pool->refcnt == 0 | |
4470 | * which implies no work queued to the pool, which implies no worker can | |
4471 | * become the manager. However a worker could have taken the role of | |
4472 | * manager before the refcnts dropped to 0, since maybe_create_worker() | |
4473 | * drops pool->lock | |
6ba94429 | 4474 | */ |
9ab03be4 VS |
4475 | while (true) { |
4476 | rcuwait_wait_event(&manager_wait, | |
4477 | !(pool->flags & POOL_MANAGER_ACTIVE), | |
4478 | TASK_UNINTERRUPTIBLE); | |
e02b9312 VS |
4479 | |
4480 | mutex_lock(&wq_pool_attach_mutex); | |
9ab03be4 VS |
4481 | raw_spin_lock_irq(&pool->lock); |
4482 | if (!(pool->flags & POOL_MANAGER_ACTIVE)) { | |
4483 | pool->flags |= POOL_MANAGER_ACTIVE; | |
4484 | break; | |
4485 | } | |
4486 | raw_spin_unlock_irq(&pool->lock); | |
e02b9312 | 4487 | mutex_unlock(&wq_pool_attach_mutex); |
9ab03be4 | 4488 | } |
692b4825 | 4489 | |
6ba94429 | 4490 | while ((worker = first_idle_worker(pool))) |
e02b9312 | 4491 | set_worker_dying(worker, &cull_list); |
6ba94429 | 4492 | WARN_ON(pool->nr_workers || pool->nr_idle); |
a9b8a985 | 4493 | raw_spin_unlock_irq(&pool->lock); |
d55262c4 | 4494 | |
e02b9312 VS |
4495 | wake_dying_workers(&cull_list); |
4496 | ||
4497 | if (!list_empty(&pool->workers) || !list_empty(&pool->dying_workers)) | |
6ba94429 | 4498 | pool->detach_completion = &detach_completion; |
1258fae7 | 4499 | mutex_unlock(&wq_pool_attach_mutex); |
226223ab | 4500 | |
6ba94429 FW |
4501 | if (pool->detach_completion) |
4502 | wait_for_completion(pool->detach_completion); | |
226223ab | 4503 | |
6ba94429 FW |
4504 | /* shut down the timers */ |
4505 | del_timer_sync(&pool->idle_timer); | |
3f959aa3 | 4506 | cancel_work_sync(&pool->idle_cull_work); |
6ba94429 | 4507 | del_timer_sync(&pool->mayday_timer); |
226223ab | 4508 | |
24acfb71 | 4509 | /* RCU protected to allow dereferences from get_work_pool() */ |
25b00775 | 4510 | call_rcu(&pool->rcu, rcu_free_pool); |
226223ab TH |
4511 | } |
4512 | ||
4513 | /** | |
6ba94429 FW |
4514 | * get_unbound_pool - get a worker_pool with the specified attributes |
4515 | * @attrs: the attributes of the worker_pool to get | |
226223ab | 4516 | * |
6ba94429 FW |
4517 | * Obtain a worker_pool which has the same attributes as @attrs, bump the |
4518 | * reference count and return it. If there already is a matching | |
4519 | * worker_pool, it will be used; otherwise, this function attempts to | |
4520 | * create a new one. | |
226223ab | 4521 | * |
6ba94429 | 4522 | * Should be called with wq_pool_mutex held. |
226223ab | 4523 | * |
6ba94429 FW |
4524 | * Return: On success, a worker_pool with the same attributes as @attrs. |
4525 | * On failure, %NULL. | |
226223ab | 4526 | */ |
6ba94429 | 4527 | static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) |
226223ab | 4528 | { |
84193c07 | 4529 | struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_NUMA]; |
6ba94429 FW |
4530 | u32 hash = wqattrs_hash(attrs); |
4531 | struct worker_pool *pool; | |
84193c07 | 4532 | int pod, node = NUMA_NO_NODE; |
226223ab | 4533 | |
6ba94429 | 4534 | lockdep_assert_held(&wq_pool_mutex); |
226223ab | 4535 | |
6ba94429 FW |
4536 | /* do we already have a matching pool? */ |
4537 | hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { | |
4538 | if (wqattrs_equal(pool->attrs, attrs)) { | |
4539 | pool->refcnt++; | |
4540 | return pool; | |
4541 | } | |
4542 | } | |
226223ab | 4543 | |
9546b29e | 4544 | /* If __pod_cpumask is contained inside a NUMA pod, that's our node */ |
84193c07 | 4545 | for (pod = 0; pod < pt->nr_pods; pod++) { |
9546b29e | 4546 | if (cpumask_subset(attrs->__pod_cpumask, pt->pod_cpus[pod])) { |
84193c07 TH |
4547 | node = pt->pod_node[pod]; |
4548 | break; | |
e2273584 XP |
4549 | } |
4550 | } | |
4551 | ||
6ba94429 | 4552 | /* nope, create a new one */ |
84193c07 | 4553 | pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, node); |
6ba94429 FW |
4554 | if (!pool || init_worker_pool(pool) < 0) |
4555 | goto fail; | |
4556 | ||
84193c07 | 4557 | pool->node = node; |
5de7a03c TH |
4558 | copy_workqueue_attrs(pool->attrs, attrs); |
4559 | wqattrs_clear_for_pool(pool->attrs); | |
226223ab | 4560 | |
6ba94429 FW |
4561 | if (worker_pool_assign_id(pool) < 0) |
4562 | goto fail; | |
226223ab | 4563 | |
6ba94429 | 4564 | /* create and start the initial worker */ |
3347fa09 | 4565 | if (wq_online && !create_worker(pool)) |
6ba94429 | 4566 | goto fail; |
226223ab | 4567 | |
6ba94429 FW |
4568 | /* install */ |
4569 | hash_add(unbound_pool_hash, &pool->hash_node, hash); | |
226223ab | 4570 | |
6ba94429 FW |
4571 | return pool; |
4572 | fail: | |
4573 | if (pool) | |
4574 | put_unbound_pool(pool); | |
4575 | return NULL; | |
226223ab | 4576 | } |
226223ab | 4577 | |
6ba94429 | 4578 | static void rcu_free_pwq(struct rcu_head *rcu) |
7a4e344c | 4579 | { |
6ba94429 FW |
4580 | kmem_cache_free(pwq_cache, |
4581 | container_of(rcu, struct pool_workqueue, rcu)); | |
7a4e344c TH |
4582 | } |
4583 | ||
6ba94429 | 4584 | /* |
967b494e TH |
4585 | * Scheduled on pwq_release_worker by put_pwq() when an unbound pwq hits zero |
4586 | * refcnt and needs to be destroyed. | |
7a4e344c | 4587 | */ |
687a9aa5 | 4588 | static void pwq_release_workfn(struct kthread_work *work) |
7a4e344c | 4589 | { |
6ba94429 | 4590 | struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, |
687a9aa5 | 4591 | release_work); |
6ba94429 FW |
4592 | struct workqueue_struct *wq = pwq->wq; |
4593 | struct worker_pool *pool = pwq->pool; | |
b42b0bdd | 4594 | bool is_last = false; |
7a4e344c | 4595 | |
b42b0bdd | 4596 | /* |
687a9aa5 | 4597 | * When @pwq is not linked, it doesn't hold any reference to the |
b42b0bdd YY |
4598 | * @wq, and @wq is invalid to access. |
4599 | */ | |
4600 | if (!list_empty(&pwq->pwqs_node)) { | |
b42b0bdd YY |
4601 | mutex_lock(&wq->mutex); |
4602 | list_del_rcu(&pwq->pwqs_node); | |
4603 | is_last = list_empty(&wq->pwqs); | |
4604 | mutex_unlock(&wq->mutex); | |
4605 | } | |
6ba94429 | 4606 | |
687a9aa5 TH |
4607 | if (wq->flags & WQ_UNBOUND) { |
4608 | mutex_lock(&wq_pool_mutex); | |
4609 | put_unbound_pool(pool); | |
4610 | mutex_unlock(&wq_pool_mutex); | |
4611 | } | |
6ba94429 | 4612 | |
5797b1c1 TH |
4613 | if (!list_empty(&pwq->pending_node)) { |
4614 | struct wq_node_nr_active *nna = | |
4615 | wq_node_nr_active(pwq->wq, pwq->pool->node); | |
4616 | ||
4617 | raw_spin_lock_irq(&nna->lock); | |
4618 | list_del_init(&pwq->pending_node); | |
4619 | raw_spin_unlock_irq(&nna->lock); | |
4620 | } | |
4621 | ||
25b00775 | 4622 | call_rcu(&pwq->rcu, rcu_free_pwq); |
7a4e344c | 4623 | |
2865a8fb | 4624 | /* |
6ba94429 FW |
4625 | * If we're the last pwq going away, @wq is already dead and no one |
4626 | * is gonna access it anymore. Schedule RCU free. | |
2865a8fb | 4627 | */ |
669de8bd BVA |
4628 | if (is_last) { |
4629 | wq_unregister_lockdep(wq); | |
25b00775 | 4630 | call_rcu(&wq->rcu, rcu_free_wq); |
669de8bd | 4631 | } |
29c91e99 TH |
4632 | } |
4633 | ||
67dc8325 | 4634 | /* initialize newly allocated @pwq which is associated with @wq and @pool */ |
6ba94429 FW |
4635 | static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, |
4636 | struct worker_pool *pool) | |
29c91e99 | 4637 | { |
6ba94429 | 4638 | BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); |
29c91e99 | 4639 | |
6ba94429 FW |
4640 | memset(pwq, 0, sizeof(*pwq)); |
4641 | ||
4642 | pwq->pool = pool; | |
4643 | pwq->wq = wq; | |
4644 | pwq->flush_color = -1; | |
4645 | pwq->refcnt = 1; | |
f97a4a1a | 4646 | INIT_LIST_HEAD(&pwq->inactive_works); |
5797b1c1 | 4647 | INIT_LIST_HEAD(&pwq->pending_node); |
6ba94429 FW |
4648 | INIT_LIST_HEAD(&pwq->pwqs_node); |
4649 | INIT_LIST_HEAD(&pwq->mayday_node); | |
687a9aa5 | 4650 | kthread_init_work(&pwq->release_work, pwq_release_workfn); |
29c91e99 TH |
4651 | } |
4652 | ||
6ba94429 FW |
4653 | /* sync @pwq with the current state of its associated wq and link it */ |
4654 | static void link_pwq(struct pool_workqueue *pwq) | |
29c91e99 | 4655 | { |
6ba94429 | 4656 | struct workqueue_struct *wq = pwq->wq; |
29c91e99 | 4657 | |
6ba94429 | 4658 | lockdep_assert_held(&wq->mutex); |
a892cacc | 4659 | |
6ba94429 FW |
4660 | /* may be called multiple times, ignore if already linked */ |
4661 | if (!list_empty(&pwq->pwqs_node)) | |
29c91e99 | 4662 | return; |
29c91e99 | 4663 | |
6ba94429 FW |
4664 | /* set the matching work_color */ |
4665 | pwq->work_color = wq->work_color; | |
29c91e99 | 4666 | |
6ba94429 FW |
4667 | /* link in @pwq */ |
4668 | list_add_rcu(&pwq->pwqs_node, &wq->pwqs); | |
4669 | } | |
29c91e99 | 4670 | |
6ba94429 FW |
4671 | /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ |
4672 | static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, | |
4673 | const struct workqueue_attrs *attrs) | |
4674 | { | |
4675 | struct worker_pool *pool; | |
4676 | struct pool_workqueue *pwq; | |
60f5a4bc | 4677 | |
6ba94429 | 4678 | lockdep_assert_held(&wq_pool_mutex); |
60f5a4bc | 4679 | |
6ba94429 FW |
4680 | pool = get_unbound_pool(attrs); |
4681 | if (!pool) | |
4682 | return NULL; | |
60f5a4bc | 4683 | |
6ba94429 FW |
4684 | pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); |
4685 | if (!pwq) { | |
4686 | put_unbound_pool(pool); | |
4687 | return NULL; | |
4688 | } | |
29c91e99 | 4689 | |
6ba94429 FW |
4690 | init_pwq(pwq, wq, pool); |
4691 | return pwq; | |
4692 | } | |
29c91e99 | 4693 | |
29c91e99 | 4694 | /** |
fef59c9c | 4695 | * wq_calc_pod_cpumask - calculate a wq_attrs' cpumask for a pod |
042f7df1 | 4696 | * @attrs: the wq_attrs of the default pwq of the target workqueue |
84193c07 | 4697 | * @cpu: the target CPU |
6ba94429 | 4698 | * @cpu_going_down: if >= 0, the CPU to consider as offline |
29c91e99 | 4699 | * |
fef59c9c TH |
4700 | * Calculate the cpumask a workqueue with @attrs should use on @pod. If |
4701 | * @cpu_going_down is >= 0, that cpu is considered offline during calculation. | |
9546b29e | 4702 | * The result is stored in @attrs->__pod_cpumask. |
a892cacc | 4703 | * |
fef59c9c TH |
4704 | * If pod affinity is not enabled, @attrs->cpumask is always used. If enabled |
4705 | * and @pod has online CPUs requested by @attrs, the returned cpumask is the | |
4706 | * intersection of the possible CPUs of @pod and @attrs->cpumask. | |
d185af30 | 4707 | * |
fef59c9c | 4708 | * The caller is responsible for ensuring that the cpumask of @pod stays stable. |
29c91e99 | 4709 | */ |
9546b29e TH |
4710 | static void wq_calc_pod_cpumask(struct workqueue_attrs *attrs, int cpu, |
4711 | int cpu_going_down) | |
29c91e99 | 4712 | { |
84193c07 TH |
4713 | const struct wq_pod_type *pt = wqattrs_pod_type(attrs); |
4714 | int pod = pt->cpu_pod[cpu]; | |
29c91e99 | 4715 | |
fef59c9c | 4716 | /* does @pod have any online CPUs @attrs wants? */ |
9546b29e TH |
4717 | cpumask_and(attrs->__pod_cpumask, pt->pod_cpus[pod], attrs->cpumask); |
4718 | cpumask_and(attrs->__pod_cpumask, attrs->__pod_cpumask, cpu_online_mask); | |
6ba94429 | 4719 | if (cpu_going_down >= 0) |
9546b29e | 4720 | cpumask_clear_cpu(cpu_going_down, attrs->__pod_cpumask); |
29c91e99 | 4721 | |
9546b29e TH |
4722 | if (cpumask_empty(attrs->__pod_cpumask)) { |
4723 | cpumask_copy(attrs->__pod_cpumask, attrs->cpumask); | |
84193c07 TH |
4724 | return; |
4725 | } | |
4c16bd32 | 4726 | |
fef59c9c | 4727 | /* yeap, return possible CPUs in @pod that @attrs wants */ |
9546b29e | 4728 | cpumask_and(attrs->__pod_cpumask, attrs->cpumask, pt->pod_cpus[pod]); |
1ad0f0a7 | 4729 | |
9546b29e | 4730 | if (cpumask_empty(attrs->__pod_cpumask)) |
1ad0f0a7 MB |
4731 | pr_warn_once("WARNING: workqueue cpumask: online intersect > " |
4732 | "possible intersect\n"); | |
4c16bd32 TH |
4733 | } |
4734 | ||
9f66cff2 | 4735 | /* install @pwq into @wq and return the old pwq, @cpu < 0 for dfl_pwq */ |
636b927e TH |
4736 | static struct pool_workqueue *install_unbound_pwq(struct workqueue_struct *wq, |
4737 | int cpu, struct pool_workqueue *pwq) | |
1befcf30 | 4738 | { |
9f66cff2 | 4739 | struct pool_workqueue __rcu **slot = unbound_pwq_slot(wq, cpu); |
1befcf30 TH |
4740 | struct pool_workqueue *old_pwq; |
4741 | ||
5b95e1af | 4742 | lockdep_assert_held(&wq_pool_mutex); |
1befcf30 TH |
4743 | lockdep_assert_held(&wq->mutex); |
4744 | ||
4745 | /* link_pwq() can handle duplicate calls */ | |
4746 | link_pwq(pwq); | |
4747 | ||
9f66cff2 TH |
4748 | old_pwq = rcu_access_pointer(*slot); |
4749 | rcu_assign_pointer(*slot, pwq); | |
1befcf30 TH |
4750 | return old_pwq; |
4751 | } | |
4752 | ||
2d5f0764 LJ |
4753 | /* context to store the prepared attrs & pwqs before applying */ |
4754 | struct apply_wqattrs_ctx { | |
4755 | struct workqueue_struct *wq; /* target workqueue */ | |
4756 | struct workqueue_attrs *attrs; /* attrs to apply */ | |
042f7df1 | 4757 | struct list_head list; /* queued for batching commit */ |
2d5f0764 LJ |
4758 | struct pool_workqueue *dfl_pwq; |
4759 | struct pool_workqueue *pwq_tbl[]; | |
4760 | }; | |
4761 | ||
4762 | /* free the resources after success or abort */ | |
4763 | static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx) | |
4764 | { | |
4765 | if (ctx) { | |
636b927e | 4766 | int cpu; |
2d5f0764 | 4767 | |
636b927e TH |
4768 | for_each_possible_cpu(cpu) |
4769 | put_pwq_unlocked(ctx->pwq_tbl[cpu]); | |
2d5f0764 LJ |
4770 | put_pwq_unlocked(ctx->dfl_pwq); |
4771 | ||
4772 | free_workqueue_attrs(ctx->attrs); | |
4773 | ||
4774 | kfree(ctx); | |
4775 | } | |
4776 | } | |
4777 | ||
4778 | /* allocate the attrs and pwqs for later installation */ | |
4779 | static struct apply_wqattrs_ctx * | |
4780 | apply_wqattrs_prepare(struct workqueue_struct *wq, | |
99c621ef LJ |
4781 | const struct workqueue_attrs *attrs, |
4782 | const cpumask_var_t unbound_cpumask) | |
9e8cd2f5 | 4783 | { |
2d5f0764 | 4784 | struct apply_wqattrs_ctx *ctx; |
9546b29e | 4785 | struct workqueue_attrs *new_attrs; |
636b927e | 4786 | int cpu; |
9e8cd2f5 | 4787 | |
2d5f0764 | 4788 | lockdep_assert_held(&wq_pool_mutex); |
9e8cd2f5 | 4789 | |
84193c07 TH |
4790 | if (WARN_ON(attrs->affn_scope < 0 || |
4791 | attrs->affn_scope >= WQ_AFFN_NR_TYPES)) | |
4792 | return ERR_PTR(-EINVAL); | |
4793 | ||
636b927e | 4794 | ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_cpu_ids), GFP_KERNEL); |
8719dcea | 4795 | |
be69d00d | 4796 | new_attrs = alloc_workqueue_attrs(); |
9546b29e | 4797 | if (!ctx || !new_attrs) |
2d5f0764 | 4798 | goto out_free; |
13e2e556 | 4799 | |
4c16bd32 TH |
4800 | /* |
4801 | * If something goes wrong during CPU up/down, we'll fall back to | |
4802 | * the default pwq covering whole @attrs->cpumask. Always create | |
4803 | * it even if we don't use it immediately. | |
4804 | */ | |
0f36ee24 TH |
4805 | copy_workqueue_attrs(new_attrs, attrs); |
4806 | wqattrs_actualize_cpumask(new_attrs, unbound_cpumask); | |
9546b29e | 4807 | cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask); |
2d5f0764 LJ |
4808 | ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs); |
4809 | if (!ctx->dfl_pwq) | |
4810 | goto out_free; | |
4c16bd32 | 4811 | |
636b927e | 4812 | for_each_possible_cpu(cpu) { |
af73f5c9 | 4813 | if (new_attrs->ordered) { |
2d5f0764 | 4814 | ctx->dfl_pwq->refcnt++; |
636b927e TH |
4815 | ctx->pwq_tbl[cpu] = ctx->dfl_pwq; |
4816 | } else { | |
9546b29e TH |
4817 | wq_calc_pod_cpumask(new_attrs, cpu, -1); |
4818 | ctx->pwq_tbl[cpu] = alloc_unbound_pwq(wq, new_attrs); | |
636b927e TH |
4819 | if (!ctx->pwq_tbl[cpu]) |
4820 | goto out_free; | |
4c16bd32 TH |
4821 | } |
4822 | } | |
4823 | ||
042f7df1 LJ |
4824 | /* save the user configured attrs and sanitize it. */ |
4825 | copy_workqueue_attrs(new_attrs, attrs); | |
4826 | cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); | |
9546b29e | 4827 | cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask); |
2d5f0764 | 4828 | ctx->attrs = new_attrs; |
042f7df1 | 4829 | |
2d5f0764 | 4830 | ctx->wq = wq; |
2d5f0764 LJ |
4831 | return ctx; |
4832 | ||
4833 | out_free: | |
2d5f0764 LJ |
4834 | free_workqueue_attrs(new_attrs); |
4835 | apply_wqattrs_cleanup(ctx); | |
84193c07 | 4836 | return ERR_PTR(-ENOMEM); |
2d5f0764 LJ |
4837 | } |
4838 | ||
4839 | /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */ | |
4840 | static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx) | |
4841 | { | |
636b927e | 4842 | int cpu; |
9e8cd2f5 | 4843 | |
4c16bd32 | 4844 | /* all pwqs have been created successfully, let's install'em */ |
2d5f0764 | 4845 | mutex_lock(&ctx->wq->mutex); |
a892cacc | 4846 | |
2d5f0764 | 4847 | copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs); |
4c16bd32 | 4848 | |
9f66cff2 | 4849 | /* save the previous pwqs and install the new ones */ |
636b927e TH |
4850 | for_each_possible_cpu(cpu) |
4851 | ctx->pwq_tbl[cpu] = install_unbound_pwq(ctx->wq, cpu, | |
4852 | ctx->pwq_tbl[cpu]); | |
9f66cff2 | 4853 | ctx->dfl_pwq = install_unbound_pwq(ctx->wq, -1, ctx->dfl_pwq); |
f147f29e | 4854 | |
5797b1c1 TH |
4855 | /* update node_nr_active->max */ |
4856 | wq_update_node_max_active(ctx->wq, -1); | |
4857 | ||
2d5f0764 LJ |
4858 | mutex_unlock(&ctx->wq->mutex); |
4859 | } | |
9e8cd2f5 | 4860 | |
a0111cf6 LJ |
4861 | static int apply_workqueue_attrs_locked(struct workqueue_struct *wq, |
4862 | const struct workqueue_attrs *attrs) | |
2d5f0764 LJ |
4863 | { |
4864 | struct apply_wqattrs_ctx *ctx; | |
4c16bd32 | 4865 | |
2d5f0764 LJ |
4866 | /* only unbound workqueues can change attributes */ |
4867 | if (WARN_ON(!(wq->flags & WQ_UNBOUND))) | |
4868 | return -EINVAL; | |
13e2e556 | 4869 | |
2d5f0764 | 4870 | /* creating multiple pwqs breaks ordering guarantee */ |
0a94efb5 TH |
4871 | if (!list_empty(&wq->pwqs)) { |
4872 | if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) | |
4873 | return -EINVAL; | |
4874 | ||
4875 | wq->flags &= ~__WQ_ORDERED; | |
4876 | } | |
2d5f0764 | 4877 | |
99c621ef | 4878 | ctx = apply_wqattrs_prepare(wq, attrs, wq_unbound_cpumask); |
84193c07 TH |
4879 | if (IS_ERR(ctx)) |
4880 | return PTR_ERR(ctx); | |
2d5f0764 LJ |
4881 | |
4882 | /* the ctx has been prepared successfully, let's commit it */ | |
6201171e | 4883 | apply_wqattrs_commit(ctx); |
2d5f0764 LJ |
4884 | apply_wqattrs_cleanup(ctx); |
4885 | ||
6201171e | 4886 | return 0; |
9e8cd2f5 TH |
4887 | } |
4888 | ||
a0111cf6 LJ |
4889 | /** |
4890 | * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue | |
4891 | * @wq: the target workqueue | |
4892 | * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() | |
4893 | * | |
fef59c9c TH |
4894 | * Apply @attrs to an unbound workqueue @wq. Unless disabled, this function maps |
4895 | * a separate pwq to each CPU pod with possibles CPUs in @attrs->cpumask so that | |
4896 | * work items are affine to the pod it was issued on. Older pwqs are released as | |
4897 | * in-flight work items finish. Note that a work item which repeatedly requeues | |
4898 | * itself back-to-back will stay on its current pwq. | |
a0111cf6 LJ |
4899 | * |
4900 | * Performs GFP_KERNEL allocations. | |
4901 | * | |
ffd8bea8 | 4902 | * Assumes caller has CPU hotplug read exclusion, i.e. cpus_read_lock(). |
509b3204 | 4903 | * |
a0111cf6 LJ |
4904 | * Return: 0 on success and -errno on failure. |
4905 | */ | |
513c98d0 | 4906 | int apply_workqueue_attrs(struct workqueue_struct *wq, |
a0111cf6 LJ |
4907 | const struct workqueue_attrs *attrs) |
4908 | { | |
4909 | int ret; | |
4910 | ||
509b3204 DJ |
4911 | lockdep_assert_cpus_held(); |
4912 | ||
4913 | mutex_lock(&wq_pool_mutex); | |
a0111cf6 | 4914 | ret = apply_workqueue_attrs_locked(wq, attrs); |
509b3204 | 4915 | mutex_unlock(&wq_pool_mutex); |
a0111cf6 LJ |
4916 | |
4917 | return ret; | |
4918 | } | |
4919 | ||
4c16bd32 | 4920 | /** |
fef59c9c | 4921 | * wq_update_pod - update pod affinity of a wq for CPU hot[un]plug |
4c16bd32 | 4922 | * @wq: the target workqueue |
4cbfd3de TH |
4923 | * @cpu: the CPU to update pool association for |
4924 | * @hotplug_cpu: the CPU coming up or going down | |
4c16bd32 TH |
4925 | * @online: whether @cpu is coming up or going down |
4926 | * | |
4927 | * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and | |
fef59c9c | 4928 | * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update pod affinity of |
4c16bd32 TH |
4929 | * @wq accordingly. |
4930 | * | |
fef59c9c TH |
4931 | * |
4932 | * If pod affinity can't be adjusted due to memory allocation failure, it falls | |
4933 | * back to @wq->dfl_pwq which may not be optimal but is always correct. | |
4934 | * | |
4935 | * Note that when the last allowed CPU of a pod goes offline for a workqueue | |
4936 | * with a cpumask spanning multiple pods, the workers which were already | |
4937 | * executing the work items for the workqueue will lose their CPU affinity and | |
4938 | * may execute on any CPU. This is similar to how per-cpu workqueues behave on | |
4939 | * CPU_DOWN. If a workqueue user wants strict affinity, it's the user's | |
4940 | * responsibility to flush the work item from CPU_DOWN_PREPARE. | |
4c16bd32 | 4941 | */ |
fef59c9c TH |
4942 | static void wq_update_pod(struct workqueue_struct *wq, int cpu, |
4943 | int hotplug_cpu, bool online) | |
4c16bd32 | 4944 | { |
4cbfd3de | 4945 | int off_cpu = online ? -1 : hotplug_cpu; |
4c16bd32 TH |
4946 | struct pool_workqueue *old_pwq = NULL, *pwq; |
4947 | struct workqueue_attrs *target_attrs; | |
4c16bd32 TH |
4948 | |
4949 | lockdep_assert_held(&wq_pool_mutex); | |
4950 | ||
84193c07 | 4951 | if (!(wq->flags & WQ_UNBOUND) || wq->unbound_attrs->ordered) |
4c16bd32 TH |
4952 | return; |
4953 | ||
4954 | /* | |
4955 | * We don't wanna alloc/free wq_attrs for each wq for each CPU. | |
4956 | * Let's use a preallocated one. The following buf is protected by | |
4957 | * CPU hotplug exclusion. | |
4958 | */ | |
fef59c9c | 4959 | target_attrs = wq_update_pod_attrs_buf; |
4c16bd32 | 4960 | |
4c16bd32 | 4961 | copy_workqueue_attrs(target_attrs, wq->unbound_attrs); |
0f36ee24 | 4962 | wqattrs_actualize_cpumask(target_attrs, wq_unbound_cpumask); |
4c16bd32 | 4963 | |
636b927e | 4964 | /* nothing to do if the target cpumask matches the current pwq */ |
9546b29e | 4965 | wq_calc_pod_cpumask(target_attrs, cpu, off_cpu); |
9f66cff2 | 4966 | if (wqattrs_equal(target_attrs, unbound_pwq(wq, cpu)->pool->attrs)) |
636b927e | 4967 | return; |
4c16bd32 | 4968 | |
4c16bd32 TH |
4969 | /* create a new pwq */ |
4970 | pwq = alloc_unbound_pwq(wq, target_attrs); | |
4971 | if (!pwq) { | |
fef59c9c | 4972 | pr_warn("workqueue: allocation failed while updating CPU pod affinity of \"%s\"\n", |
2d916033 | 4973 | wq->name); |
77f300b1 | 4974 | goto use_dfl_pwq; |
4c16bd32 TH |
4975 | } |
4976 | ||
f7142ed4 | 4977 | /* Install the new pwq. */ |
4c16bd32 | 4978 | mutex_lock(&wq->mutex); |
636b927e | 4979 | old_pwq = install_unbound_pwq(wq, cpu, pwq); |
4c16bd32 TH |
4980 | goto out_unlock; |
4981 | ||
4982 | use_dfl_pwq: | |
f7142ed4 | 4983 | mutex_lock(&wq->mutex); |
9f66cff2 TH |
4984 | pwq = unbound_pwq(wq, -1); |
4985 | raw_spin_lock_irq(&pwq->pool->lock); | |
4986 | get_pwq(pwq); | |
4987 | raw_spin_unlock_irq(&pwq->pool->lock); | |
4988 | old_pwq = install_unbound_pwq(wq, cpu, pwq); | |
4c16bd32 TH |
4989 | out_unlock: |
4990 | mutex_unlock(&wq->mutex); | |
4991 | put_pwq_unlocked(old_pwq); | |
4992 | } | |
4993 | ||
30cdf249 | 4994 | static int alloc_and_link_pwqs(struct workqueue_struct *wq) |
0f900049 | 4995 | { |
49e3cf44 | 4996 | bool highpri = wq->flags & WQ_HIGHPRI; |
8a2b7538 | 4997 | int cpu, ret; |
30cdf249 | 4998 | |
636b927e TH |
4999 | wq->cpu_pwq = alloc_percpu(struct pool_workqueue *); |
5000 | if (!wq->cpu_pwq) | |
5001 | goto enomem; | |
30cdf249 | 5002 | |
636b927e | 5003 | if (!(wq->flags & WQ_UNBOUND)) { |
30cdf249 | 5004 | for_each_possible_cpu(cpu) { |
687a9aa5 | 5005 | struct pool_workqueue **pwq_p = |
ee1ceef7 | 5006 | per_cpu_ptr(wq->cpu_pwq, cpu); |
687a9aa5 TH |
5007 | struct worker_pool *pool = |
5008 | &(per_cpu_ptr(cpu_worker_pools, cpu)[highpri]); | |
5009 | ||
5010 | *pwq_p = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, | |
5011 | pool->node); | |
5012 | if (!*pwq_p) | |
5013 | goto enomem; | |
f3421797 | 5014 | |
687a9aa5 | 5015 | init_pwq(*pwq_p, wq, pool); |
f147f29e TH |
5016 | |
5017 | mutex_lock(&wq->mutex); | |
687a9aa5 | 5018 | link_pwq(*pwq_p); |
f147f29e | 5019 | mutex_unlock(&wq->mutex); |
30cdf249 | 5020 | } |
9e8cd2f5 | 5021 | return 0; |
509b3204 DJ |
5022 | } |
5023 | ||
ffd8bea8 | 5024 | cpus_read_lock(); |
509b3204 | 5025 | if (wq->flags & __WQ_ORDERED) { |
9f66cff2 TH |
5026 | struct pool_workqueue *dfl_pwq; |
5027 | ||
8a2b7538 TH |
5028 | ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]); |
5029 | /* there should only be single pwq for ordering guarantee */ | |
9f66cff2 TH |
5030 | dfl_pwq = rcu_access_pointer(wq->dfl_pwq); |
5031 | WARN(!ret && (wq->pwqs.next != &dfl_pwq->pwqs_node || | |
5032 | wq->pwqs.prev != &dfl_pwq->pwqs_node), | |
8a2b7538 | 5033 | "ordering guarantee broken for workqueue %s\n", wq->name); |
30cdf249 | 5034 | } else { |
509b3204 | 5035 | ret = apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); |
30cdf249 | 5036 | } |
ffd8bea8 | 5037 | cpus_read_unlock(); |
509b3204 | 5038 | |
64344553 Z |
5039 | /* for unbound pwq, flush the pwq_release_worker ensures that the |
5040 | * pwq_release_workfn() completes before calling kfree(wq). | |
5041 | */ | |
5042 | if (ret) | |
5043 | kthread_flush_worker(pwq_release_worker); | |
5044 | ||
509b3204 | 5045 | return ret; |
687a9aa5 TH |
5046 | |
5047 | enomem: | |
5048 | if (wq->cpu_pwq) { | |
7b42f401 Z |
5049 | for_each_possible_cpu(cpu) { |
5050 | struct pool_workqueue *pwq = *per_cpu_ptr(wq->cpu_pwq, cpu); | |
5051 | ||
5052 | if (pwq) | |
5053 | kmem_cache_free(pwq_cache, pwq); | |
5054 | } | |
687a9aa5 TH |
5055 | free_percpu(wq->cpu_pwq); |
5056 | wq->cpu_pwq = NULL; | |
5057 | } | |
5058 | return -ENOMEM; | |
0f900049 TH |
5059 | } |
5060 | ||
f3421797 TH |
5061 | static int wq_clamp_max_active(int max_active, unsigned int flags, |
5062 | const char *name) | |
b71ab8c2 | 5063 | { |
636b927e | 5064 | if (max_active < 1 || max_active > WQ_MAX_ACTIVE) |
044c782c | 5065 | pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", |
636b927e | 5066 | max_active, name, 1, WQ_MAX_ACTIVE); |
b71ab8c2 | 5067 | |
636b927e | 5068 | return clamp_val(max_active, 1, WQ_MAX_ACTIVE); |
b71ab8c2 TH |
5069 | } |
5070 | ||
983c7515 TH |
5071 | /* |
5072 | * Workqueues which may be used during memory reclaim should have a rescuer | |
5073 | * to guarantee forward progress. | |
5074 | */ | |
5075 | static int init_rescuer(struct workqueue_struct *wq) | |
5076 | { | |
5077 | struct worker *rescuer; | |
b92b36ea | 5078 | int ret; |
983c7515 TH |
5079 | |
5080 | if (!(wq->flags & WQ_MEM_RECLAIM)) | |
5081 | return 0; | |
5082 | ||
5083 | rescuer = alloc_worker(NUMA_NO_NODE); | |
4c0736a7 PM |
5084 | if (!rescuer) { |
5085 | pr_err("workqueue: Failed to allocate a rescuer for wq \"%s\"\n", | |
5086 | wq->name); | |
983c7515 | 5087 | return -ENOMEM; |
4c0736a7 | 5088 | } |
983c7515 TH |
5089 | |
5090 | rescuer->rescue_wq = wq; | |
b6a46f72 | 5091 | rescuer->task = kthread_create(rescuer_thread, rescuer, "kworker/R-%s", wq->name); |
f187b697 | 5092 | if (IS_ERR(rescuer->task)) { |
b92b36ea | 5093 | ret = PTR_ERR(rescuer->task); |
4c0736a7 PM |
5094 | pr_err("workqueue: Failed to create a rescuer kthread for wq \"%s\": %pe", |
5095 | wq->name, ERR_PTR(ret)); | |
983c7515 | 5096 | kfree(rescuer); |
b92b36ea | 5097 | return ret; |
983c7515 TH |
5098 | } |
5099 | ||
5100 | wq->rescuer = rescuer; | |
85f0ab43 JL |
5101 | if (wq->flags & WQ_UNBOUND) |
5102 | kthread_bind_mask(rescuer->task, wq->unbound_attrs->cpumask); | |
5103 | else | |
5104 | kthread_bind_mask(rescuer->task, cpu_possible_mask); | |
983c7515 TH |
5105 | wake_up_process(rescuer->task); |
5106 | ||
5107 | return 0; | |
5108 | } | |
5109 | ||
a045a272 TH |
5110 | /** |
5111 | * wq_adjust_max_active - update a wq's max_active to the current setting | |
5112 | * @wq: target workqueue | |
5113 | * | |
5114 | * If @wq isn't freezing, set @wq->max_active to the saved_max_active and | |
5115 | * activate inactive work items accordingly. If @wq is freezing, clear | |
5116 | * @wq->max_active to zero. | |
5117 | */ | |
5118 | static void wq_adjust_max_active(struct workqueue_struct *wq) | |
5119 | { | |
c5404d4e | 5120 | bool activated; |
5797b1c1 | 5121 | int new_max, new_min; |
a045a272 TH |
5122 | |
5123 | lockdep_assert_held(&wq->mutex); | |
5124 | ||
5125 | if ((wq->flags & WQ_FREEZABLE) && workqueue_freezing) { | |
5797b1c1 TH |
5126 | new_max = 0; |
5127 | new_min = 0; | |
5128 | } else { | |
5129 | new_max = wq->saved_max_active; | |
5130 | new_min = wq->saved_min_active; | |
a045a272 TH |
5131 | } |
5132 | ||
5797b1c1 | 5133 | if (wq->max_active == new_max && wq->min_active == new_min) |
a045a272 TH |
5134 | return; |
5135 | ||
5136 | /* | |
5797b1c1 | 5137 | * Update @wq->max/min_active and then kick inactive work items if more |
a045a272 TH |
5138 | * active work items are allowed. This doesn't break work item ordering |
5139 | * because new work items are always queued behind existing inactive | |
5140 | * work items if there are any. | |
5141 | */ | |
5797b1c1 TH |
5142 | WRITE_ONCE(wq->max_active, new_max); |
5143 | WRITE_ONCE(wq->min_active, new_min); | |
5144 | ||
5145 | if (wq->flags & WQ_UNBOUND) | |
5146 | wq_update_node_max_active(wq, -1); | |
5147 | ||
5148 | if (new_max == 0) | |
5149 | return; | |
a045a272 | 5150 | |
c5404d4e TH |
5151 | /* |
5152 | * Round-robin through pwq's activating the first inactive work item | |
5153 | * until max_active is filled. | |
5154 | */ | |
5155 | do { | |
5156 | struct pool_workqueue *pwq; | |
a045a272 | 5157 | |
c5404d4e TH |
5158 | activated = false; |
5159 | for_each_pwq(pwq, wq) { | |
5160 | unsigned long flags; | |
a045a272 | 5161 | |
c5404d4e TH |
5162 | /* can be called during early boot w/ irq disabled */ |
5163 | raw_spin_lock_irqsave(&pwq->pool->lock, flags); | |
5797b1c1 | 5164 | if (pwq_activate_first_inactive(pwq, true)) { |
c5404d4e TH |
5165 | activated = true; |
5166 | kick_pool(pwq->pool); | |
5167 | } | |
5168 | raw_spin_unlock_irqrestore(&pwq->pool->lock, flags); | |
5169 | } | |
5170 | } while (activated); | |
a045a272 TH |
5171 | } |
5172 | ||
a2775bbc | 5173 | __printf(1, 4) |
669de8bd BVA |
5174 | struct workqueue_struct *alloc_workqueue(const char *fmt, |
5175 | unsigned int flags, | |
5176 | int max_active, ...) | |
1da177e4 | 5177 | { |
ecf6881f | 5178 | va_list args; |
1da177e4 | 5179 | struct workqueue_struct *wq; |
91ccc6e7 TH |
5180 | size_t wq_size; |
5181 | int name_len; | |
b196be89 | 5182 | |
5c0338c6 | 5183 | /* |
fef59c9c TH |
5184 | * Unbound && max_active == 1 used to imply ordered, which is no longer |
5185 | * the case on many machines due to per-pod pools. While | |
5c0338c6 | 5186 | * alloc_ordered_workqueue() is the right way to create an ordered |
fef59c9c | 5187 | * workqueue, keep the previous behavior to avoid subtle breakages. |
5c0338c6 TH |
5188 | */ |
5189 | if ((flags & WQ_UNBOUND) && max_active == 1) | |
5190 | flags |= __WQ_ORDERED; | |
5191 | ||
cee22a15 VK |
5192 | /* see the comment above the definition of WQ_POWER_EFFICIENT */ |
5193 | if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) | |
5194 | flags |= WQ_UNBOUND; | |
5195 | ||
ecf6881f | 5196 | /* allocate wq and format name */ |
91ccc6e7 TH |
5197 | if (flags & WQ_UNBOUND) |
5198 | wq_size = struct_size(wq, node_nr_active, nr_node_ids + 1); | |
5199 | else | |
5200 | wq_size = sizeof(*wq); | |
5201 | ||
5202 | wq = kzalloc(wq_size, GFP_KERNEL); | |
b196be89 | 5203 | if (!wq) |
d2c1d404 | 5204 | return NULL; |
b196be89 | 5205 | |
6029a918 | 5206 | if (flags & WQ_UNBOUND) { |
be69d00d | 5207 | wq->unbound_attrs = alloc_workqueue_attrs(); |
6029a918 TH |
5208 | if (!wq->unbound_attrs) |
5209 | goto err_free_wq; | |
5210 | } | |
5211 | ||
669de8bd | 5212 | va_start(args, max_active); |
91ccc6e7 | 5213 | name_len = vsnprintf(wq->name, sizeof(wq->name), fmt, args); |
b196be89 | 5214 | va_end(args); |
1da177e4 | 5215 | |
91ccc6e7 TH |
5216 | if (name_len >= WQ_NAME_LEN) |
5217 | pr_warn_once("workqueue: name exceeds WQ_NAME_LEN. Truncating to: %s\n", | |
5218 | wq->name); | |
31c89007 | 5219 | |
d320c038 | 5220 | max_active = max_active ?: WQ_DFL_ACTIVE; |
b196be89 | 5221 | max_active = wq_clamp_max_active(max_active, flags, wq->name); |
3af24433 | 5222 | |
b196be89 | 5223 | /* init wq */ |
97e37d7b | 5224 | wq->flags = flags; |
a045a272 | 5225 | wq->max_active = max_active; |
5797b1c1 TH |
5226 | wq->min_active = min(max_active, WQ_DFL_MIN_ACTIVE); |
5227 | wq->saved_max_active = wq->max_active; | |
5228 | wq->saved_min_active = wq->min_active; | |
3c25a55d | 5229 | mutex_init(&wq->mutex); |
112202d9 | 5230 | atomic_set(&wq->nr_pwqs_to_flush, 0); |
30cdf249 | 5231 | INIT_LIST_HEAD(&wq->pwqs); |
73f53c4a TH |
5232 | INIT_LIST_HEAD(&wq->flusher_queue); |
5233 | INIT_LIST_HEAD(&wq->flusher_overflow); | |
493a1724 | 5234 | INIT_LIST_HEAD(&wq->maydays); |
502ca9d8 | 5235 | |
669de8bd | 5236 | wq_init_lockdep(wq); |
cce1a165 | 5237 | INIT_LIST_HEAD(&wq->list); |
3af24433 | 5238 | |
91ccc6e7 TH |
5239 | if (flags & WQ_UNBOUND) { |
5240 | if (alloc_node_nr_active(wq->node_nr_active) < 0) | |
5241 | goto err_unreg_lockdep; | |
5242 | } | |
5243 | ||
30cdf249 | 5244 | if (alloc_and_link_pwqs(wq) < 0) |
91ccc6e7 | 5245 | goto err_free_node_nr_active; |
1537663f | 5246 | |
40c17f75 | 5247 | if (wq_online && init_rescuer(wq) < 0) |
983c7515 | 5248 | goto err_destroy; |
3af24433 | 5249 | |
226223ab TH |
5250 | if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) |
5251 | goto err_destroy; | |
5252 | ||
a0a1a5fd | 5253 | /* |
68e13a67 LJ |
5254 | * wq_pool_mutex protects global freeze state and workqueues list. |
5255 | * Grab it, adjust max_active and add the new @wq to workqueues | |
5256 | * list. | |
a0a1a5fd | 5257 | */ |
68e13a67 | 5258 | mutex_lock(&wq_pool_mutex); |
a0a1a5fd | 5259 | |
a357fc03 | 5260 | mutex_lock(&wq->mutex); |
a045a272 | 5261 | wq_adjust_max_active(wq); |
a357fc03 | 5262 | mutex_unlock(&wq->mutex); |
a0a1a5fd | 5263 | |
e2dca7ad | 5264 | list_add_tail_rcu(&wq->list, &workqueues); |
a0a1a5fd | 5265 | |
68e13a67 | 5266 | mutex_unlock(&wq_pool_mutex); |
1537663f | 5267 | |
3af24433 | 5268 | return wq; |
d2c1d404 | 5269 | |
91ccc6e7 TH |
5270 | err_free_node_nr_active: |
5271 | if (wq->flags & WQ_UNBOUND) | |
5272 | free_node_nr_active(wq->node_nr_active); | |
82efcab3 | 5273 | err_unreg_lockdep: |
009bb421 BVA |
5274 | wq_unregister_lockdep(wq); |
5275 | wq_free_lockdep(wq); | |
82efcab3 | 5276 | err_free_wq: |
6029a918 | 5277 | free_workqueue_attrs(wq->unbound_attrs); |
d2c1d404 TH |
5278 | kfree(wq); |
5279 | return NULL; | |
5280 | err_destroy: | |
5281 | destroy_workqueue(wq); | |
4690c4ab | 5282 | return NULL; |
3af24433 | 5283 | } |
669de8bd | 5284 | EXPORT_SYMBOL_GPL(alloc_workqueue); |
1da177e4 | 5285 | |
c29eb853 TH |
5286 | static bool pwq_busy(struct pool_workqueue *pwq) |
5287 | { | |
5288 | int i; | |
5289 | ||
5290 | for (i = 0; i < WORK_NR_COLORS; i++) | |
5291 | if (pwq->nr_in_flight[i]) | |
5292 | return true; | |
5293 | ||
9f66cff2 | 5294 | if ((pwq != rcu_access_pointer(pwq->wq->dfl_pwq)) && (pwq->refcnt > 1)) |
c29eb853 | 5295 | return true; |
afa87ce8 | 5296 | if (!pwq_is_empty(pwq)) |
c29eb853 TH |
5297 | return true; |
5298 | ||
5299 | return false; | |
5300 | } | |
5301 | ||
3af24433 ON |
5302 | /** |
5303 | * destroy_workqueue - safely terminate a workqueue | |
5304 | * @wq: target workqueue | |
5305 | * | |
5306 | * Safely destroy a workqueue. All work currently pending will be done first. | |
5307 | */ | |
5308 | void destroy_workqueue(struct workqueue_struct *wq) | |
5309 | { | |
49e3cf44 | 5310 | struct pool_workqueue *pwq; |
636b927e | 5311 | int cpu; |
3af24433 | 5312 | |
def98c84 TH |
5313 | /* |
5314 | * Remove it from sysfs first so that sanity check failure doesn't | |
5315 | * lead to sysfs name conflicts. | |
5316 | */ | |
5317 | workqueue_sysfs_unregister(wq); | |
5318 | ||
33e3f0a3 RC |
5319 | /* mark the workqueue destruction is in progress */ |
5320 | mutex_lock(&wq->mutex); | |
5321 | wq->flags |= __WQ_DESTROYING; | |
5322 | mutex_unlock(&wq->mutex); | |
5323 | ||
9c5a2ba7 TH |
5324 | /* drain it before proceeding with destruction */ |
5325 | drain_workqueue(wq); | |
c8efcc25 | 5326 | |
def98c84 TH |
5327 | /* kill rescuer, if sanity checks fail, leave it w/o rescuer */ |
5328 | if (wq->rescuer) { | |
5329 | struct worker *rescuer = wq->rescuer; | |
5330 | ||
5331 | /* this prevents new queueing */ | |
a9b8a985 | 5332 | raw_spin_lock_irq(&wq_mayday_lock); |
def98c84 | 5333 | wq->rescuer = NULL; |
a9b8a985 | 5334 | raw_spin_unlock_irq(&wq_mayday_lock); |
def98c84 TH |
5335 | |
5336 | /* rescuer will empty maydays list before exiting */ | |
5337 | kthread_stop(rescuer->task); | |
8efe1223 | 5338 | kfree(rescuer); |
def98c84 TH |
5339 | } |
5340 | ||
c29eb853 TH |
5341 | /* |
5342 | * Sanity checks - grab all the locks so that we wait for all | |
5343 | * in-flight operations which may do put_pwq(). | |
5344 | */ | |
5345 | mutex_lock(&wq_pool_mutex); | |
b09f4fd3 | 5346 | mutex_lock(&wq->mutex); |
49e3cf44 | 5347 | for_each_pwq(pwq, wq) { |
a9b8a985 | 5348 | raw_spin_lock_irq(&pwq->pool->lock); |
c29eb853 | 5349 | if (WARN_ON(pwq_busy(pwq))) { |
1d9a6159 KW |
5350 | pr_warn("%s: %s has the following busy pwq\n", |
5351 | __func__, wq->name); | |
c29eb853 | 5352 | show_pwq(pwq); |
a9b8a985 | 5353 | raw_spin_unlock_irq(&pwq->pool->lock); |
b09f4fd3 | 5354 | mutex_unlock(&wq->mutex); |
c29eb853 | 5355 | mutex_unlock(&wq_pool_mutex); |
55df0933 | 5356 | show_one_workqueue(wq); |
6183c009 | 5357 | return; |
76af4d93 | 5358 | } |
a9b8a985 | 5359 | raw_spin_unlock_irq(&pwq->pool->lock); |
6183c009 | 5360 | } |
b09f4fd3 | 5361 | mutex_unlock(&wq->mutex); |
6183c009 | 5362 | |
a0a1a5fd TH |
5363 | /* |
5364 | * wq list is used to freeze wq, remove from list after | |
5365 | * flushing is complete in case freeze races us. | |
5366 | */ | |
e2dca7ad | 5367 | list_del_rcu(&wq->list); |
68e13a67 | 5368 | mutex_unlock(&wq_pool_mutex); |
3af24433 | 5369 | |
636b927e TH |
5370 | /* |
5371 | * We're the sole accessor of @wq. Directly access cpu_pwq and dfl_pwq | |
5372 | * to put the base refs. @wq will be auto-destroyed from the last | |
5373 | * pwq_put. RCU read lock prevents @wq from going away from under us. | |
5374 | */ | |
5375 | rcu_read_lock(); | |
4c16bd32 | 5376 | |
636b927e | 5377 | for_each_possible_cpu(cpu) { |
9f66cff2 TH |
5378 | put_pwq_unlocked(unbound_pwq(wq, cpu)); |
5379 | RCU_INIT_POINTER(*unbound_pwq_slot(wq, cpu), NULL); | |
29c91e99 | 5380 | } |
636b927e | 5381 | |
9f66cff2 TH |
5382 | put_pwq_unlocked(unbound_pwq(wq, -1)); |
5383 | RCU_INIT_POINTER(*unbound_pwq_slot(wq, -1), NULL); | |
636b927e TH |
5384 | |
5385 | rcu_read_unlock(); | |
3af24433 ON |
5386 | } |
5387 | EXPORT_SYMBOL_GPL(destroy_workqueue); | |
5388 | ||
dcd989cb TH |
5389 | /** |
5390 | * workqueue_set_max_active - adjust max_active of a workqueue | |
5391 | * @wq: target workqueue | |
5392 | * @max_active: new max_active value. | |
5393 | * | |
5797b1c1 TH |
5394 | * Set max_active of @wq to @max_active. See the alloc_workqueue() function |
5395 | * comment. | |
dcd989cb TH |
5396 | * |
5397 | * CONTEXT: | |
5398 | * Don't call from IRQ context. | |
5399 | */ | |
5400 | void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) | |
5401 | { | |
8719dcea | 5402 | /* disallow meddling with max_active for ordered workqueues */ |
0a94efb5 | 5403 | if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) |
8719dcea TH |
5404 | return; |
5405 | ||
f3421797 | 5406 | max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); |
dcd989cb | 5407 | |
a357fc03 | 5408 | mutex_lock(&wq->mutex); |
dcd989cb | 5409 | |
0a94efb5 | 5410 | wq->flags &= ~__WQ_ORDERED; |
dcd989cb | 5411 | wq->saved_max_active = max_active; |
5797b1c1 TH |
5412 | if (wq->flags & WQ_UNBOUND) |
5413 | wq->saved_min_active = min(wq->saved_min_active, max_active); | |
5414 | ||
a045a272 | 5415 | wq_adjust_max_active(wq); |
93981800 | 5416 | |
a357fc03 | 5417 | mutex_unlock(&wq->mutex); |
15316ba8 | 5418 | } |
dcd989cb | 5419 | EXPORT_SYMBOL_GPL(workqueue_set_max_active); |
15316ba8 | 5420 | |
27d4ee03 LW |
5421 | /** |
5422 | * current_work - retrieve %current task's work struct | |
5423 | * | |
5424 | * Determine if %current task is a workqueue worker and what it's working on. | |
5425 | * Useful to find out the context that the %current task is running in. | |
5426 | * | |
5427 | * Return: work struct if %current task is a workqueue worker, %NULL otherwise. | |
5428 | */ | |
5429 | struct work_struct *current_work(void) | |
5430 | { | |
5431 | struct worker *worker = current_wq_worker(); | |
5432 | ||
5433 | return worker ? worker->current_work : NULL; | |
5434 | } | |
5435 | EXPORT_SYMBOL(current_work); | |
5436 | ||
e6267616 TH |
5437 | /** |
5438 | * current_is_workqueue_rescuer - is %current workqueue rescuer? | |
5439 | * | |
5440 | * Determine whether %current is a workqueue rescuer. Can be used from | |
5441 | * work functions to determine whether it's being run off the rescuer task. | |
d185af30 YB |
5442 | * |
5443 | * Return: %true if %current is a workqueue rescuer. %false otherwise. | |
e6267616 TH |
5444 | */ |
5445 | bool current_is_workqueue_rescuer(void) | |
5446 | { | |
5447 | struct worker *worker = current_wq_worker(); | |
5448 | ||
6a092dfd | 5449 | return worker && worker->rescue_wq; |
e6267616 TH |
5450 | } |
5451 | ||
eef6a7d5 | 5452 | /** |
dcd989cb TH |
5453 | * workqueue_congested - test whether a workqueue is congested |
5454 | * @cpu: CPU in question | |
5455 | * @wq: target workqueue | |
eef6a7d5 | 5456 | * |
dcd989cb TH |
5457 | * Test whether @wq's cpu workqueue for @cpu is congested. There is |
5458 | * no synchronization around this function and the test result is | |
5459 | * unreliable and only useful as advisory hints or for debugging. | |
eef6a7d5 | 5460 | * |
d3251859 | 5461 | * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. |
636b927e TH |
5462 | * |
5463 | * With the exception of ordered workqueues, all workqueues have per-cpu | |
5464 | * pool_workqueues, each with its own congested state. A workqueue being | |
5465 | * congested on one CPU doesn't mean that the workqueue is contested on any | |
5466 | * other CPUs. | |
d3251859 | 5467 | * |
d185af30 | 5468 | * Return: |
dcd989cb | 5469 | * %true if congested, %false otherwise. |
eef6a7d5 | 5470 | */ |
d84ff051 | 5471 | bool workqueue_congested(int cpu, struct workqueue_struct *wq) |
1da177e4 | 5472 | { |
7fb98ea7 | 5473 | struct pool_workqueue *pwq; |
76af4d93 TH |
5474 | bool ret; |
5475 | ||
24acfb71 TG |
5476 | rcu_read_lock(); |
5477 | preempt_disable(); | |
7fb98ea7 | 5478 | |
d3251859 TH |
5479 | if (cpu == WORK_CPU_UNBOUND) |
5480 | cpu = smp_processor_id(); | |
5481 | ||
636b927e | 5482 | pwq = *per_cpu_ptr(wq->cpu_pwq, cpu); |
f97a4a1a | 5483 | ret = !list_empty(&pwq->inactive_works); |
636b927e | 5484 | |
24acfb71 TG |
5485 | preempt_enable(); |
5486 | rcu_read_unlock(); | |
76af4d93 TH |
5487 | |
5488 | return ret; | |
1da177e4 | 5489 | } |
dcd989cb | 5490 | EXPORT_SYMBOL_GPL(workqueue_congested); |
1da177e4 | 5491 | |
dcd989cb TH |
5492 | /** |
5493 | * work_busy - test whether a work is currently pending or running | |
5494 | * @work: the work to be tested | |
5495 | * | |
5496 | * Test whether @work is currently pending or running. There is no | |
5497 | * synchronization around this function and the test result is | |
5498 | * unreliable and only useful as advisory hints or for debugging. | |
dcd989cb | 5499 | * |
d185af30 | 5500 | * Return: |
dcd989cb TH |
5501 | * OR'd bitmask of WORK_BUSY_* bits. |
5502 | */ | |
5503 | unsigned int work_busy(struct work_struct *work) | |
1da177e4 | 5504 | { |
fa1b54e6 | 5505 | struct worker_pool *pool; |
dcd989cb TH |
5506 | unsigned long flags; |
5507 | unsigned int ret = 0; | |
1da177e4 | 5508 | |
dcd989cb TH |
5509 | if (work_pending(work)) |
5510 | ret |= WORK_BUSY_PENDING; | |
1da177e4 | 5511 | |
24acfb71 | 5512 | rcu_read_lock(); |
fa1b54e6 | 5513 | pool = get_work_pool(work); |
038366c5 | 5514 | if (pool) { |
a9b8a985 | 5515 | raw_spin_lock_irqsave(&pool->lock, flags); |
038366c5 LJ |
5516 | if (find_worker_executing_work(pool, work)) |
5517 | ret |= WORK_BUSY_RUNNING; | |
a9b8a985 | 5518 | raw_spin_unlock_irqrestore(&pool->lock, flags); |
038366c5 | 5519 | } |
24acfb71 | 5520 | rcu_read_unlock(); |
1da177e4 | 5521 | |
dcd989cb | 5522 | return ret; |
1da177e4 | 5523 | } |
dcd989cb | 5524 | EXPORT_SYMBOL_GPL(work_busy); |
1da177e4 | 5525 | |
3d1cb205 TH |
5526 | /** |
5527 | * set_worker_desc - set description for the current work item | |
5528 | * @fmt: printf-style format string | |
5529 | * @...: arguments for the format string | |
5530 | * | |
5531 | * This function can be called by a running work function to describe what | |
5532 | * the work item is about. If the worker task gets dumped, this | |
5533 | * information will be printed out together to help debugging. The | |
5534 | * description can be at most WORKER_DESC_LEN including the trailing '\0'. | |
5535 | */ | |
5536 | void set_worker_desc(const char *fmt, ...) | |
5537 | { | |
5538 | struct worker *worker = current_wq_worker(); | |
5539 | va_list args; | |
5540 | ||
5541 | if (worker) { | |
5542 | va_start(args, fmt); | |
5543 | vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); | |
5544 | va_end(args); | |
3d1cb205 TH |
5545 | } |
5546 | } | |
5c750d58 | 5547 | EXPORT_SYMBOL_GPL(set_worker_desc); |
3d1cb205 TH |
5548 | |
5549 | /** | |
5550 | * print_worker_info - print out worker information and description | |
5551 | * @log_lvl: the log level to use when printing | |
5552 | * @task: target task | |
5553 | * | |
5554 | * If @task is a worker and currently executing a work item, print out the | |
5555 | * name of the workqueue being serviced and worker description set with | |
5556 | * set_worker_desc() by the currently executing work item. | |
5557 | * | |
5558 | * This function can be safely called on any task as long as the | |
5559 | * task_struct itself is accessible. While safe, this function isn't | |
5560 | * synchronized and may print out mixups or garbages of limited length. | |
5561 | */ | |
5562 | void print_worker_info(const char *log_lvl, struct task_struct *task) | |
5563 | { | |
5564 | work_func_t *fn = NULL; | |
5565 | char name[WQ_NAME_LEN] = { }; | |
5566 | char desc[WORKER_DESC_LEN] = { }; | |
5567 | struct pool_workqueue *pwq = NULL; | |
5568 | struct workqueue_struct *wq = NULL; | |
3d1cb205 TH |
5569 | struct worker *worker; |
5570 | ||
5571 | if (!(task->flags & PF_WQ_WORKER)) | |
5572 | return; | |
5573 | ||
5574 | /* | |
5575 | * This function is called without any synchronization and @task | |
5576 | * could be in any state. Be careful with dereferences. | |
5577 | */ | |
e700591a | 5578 | worker = kthread_probe_data(task); |
3d1cb205 TH |
5579 | |
5580 | /* | |
8bf89593 TH |
5581 | * Carefully copy the associated workqueue's workfn, name and desc. |
5582 | * Keep the original last '\0' in case the original is garbage. | |
3d1cb205 | 5583 | */ |
fe557319 CH |
5584 | copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn)); |
5585 | copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq)); | |
5586 | copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq)); | |
5587 | copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1); | |
5588 | copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1); | |
3d1cb205 TH |
5589 | |
5590 | if (fn || name[0] || desc[0]) { | |
d75f773c | 5591 | printk("%sWorkqueue: %s %ps", log_lvl, name, fn); |
8bf89593 | 5592 | if (strcmp(name, desc)) |
3d1cb205 TH |
5593 | pr_cont(" (%s)", desc); |
5594 | pr_cont("\n"); | |
5595 | } | |
5596 | } | |
5597 | ||
3494fc30 TH |
5598 | static void pr_cont_pool_info(struct worker_pool *pool) |
5599 | { | |
5600 | pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask); | |
5601 | if (pool->node != NUMA_NO_NODE) | |
5602 | pr_cont(" node=%d", pool->node); | |
5603 | pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice); | |
5604 | } | |
5605 | ||
c76feb0d PM |
5606 | struct pr_cont_work_struct { |
5607 | bool comma; | |
5608 | work_func_t func; | |
5609 | long ctr; | |
5610 | }; | |
5611 | ||
5612 | static void pr_cont_work_flush(bool comma, work_func_t func, struct pr_cont_work_struct *pcwsp) | |
5613 | { | |
5614 | if (!pcwsp->ctr) | |
5615 | goto out_record; | |
5616 | if (func == pcwsp->func) { | |
5617 | pcwsp->ctr++; | |
5618 | return; | |
5619 | } | |
5620 | if (pcwsp->ctr == 1) | |
5621 | pr_cont("%s %ps", pcwsp->comma ? "," : "", pcwsp->func); | |
5622 | else | |
5623 | pr_cont("%s %ld*%ps", pcwsp->comma ? "," : "", pcwsp->ctr, pcwsp->func); | |
5624 | pcwsp->ctr = 0; | |
5625 | out_record: | |
5626 | if ((long)func == -1L) | |
5627 | return; | |
5628 | pcwsp->comma = comma; | |
5629 | pcwsp->func = func; | |
5630 | pcwsp->ctr = 1; | |
5631 | } | |
5632 | ||
5633 | static void pr_cont_work(bool comma, struct work_struct *work, struct pr_cont_work_struct *pcwsp) | |
3494fc30 TH |
5634 | { |
5635 | if (work->func == wq_barrier_func) { | |
5636 | struct wq_barrier *barr; | |
5637 | ||
5638 | barr = container_of(work, struct wq_barrier, work); | |
5639 | ||
c76feb0d | 5640 | pr_cont_work_flush(comma, (work_func_t)-1, pcwsp); |
3494fc30 TH |
5641 | pr_cont("%s BAR(%d)", comma ? "," : "", |
5642 | task_pid_nr(barr->task)); | |
5643 | } else { | |
c76feb0d PM |
5644 | if (!comma) |
5645 | pr_cont_work_flush(comma, (work_func_t)-1, pcwsp); | |
5646 | pr_cont_work_flush(comma, work->func, pcwsp); | |
3494fc30 TH |
5647 | } |
5648 | } | |
5649 | ||
5650 | static void show_pwq(struct pool_workqueue *pwq) | |
5651 | { | |
c76feb0d | 5652 | struct pr_cont_work_struct pcws = { .ctr = 0, }; |
3494fc30 TH |
5653 | struct worker_pool *pool = pwq->pool; |
5654 | struct work_struct *work; | |
5655 | struct worker *worker; | |
5656 | bool has_in_flight = false, has_pending = false; | |
5657 | int bkt; | |
5658 | ||
5659 | pr_info(" pwq %d:", pool->id); | |
5660 | pr_cont_pool_info(pool); | |
5661 | ||
a045a272 TH |
5662 | pr_cont(" active=%d refcnt=%d%s\n", |
5663 | pwq->nr_active, pwq->refcnt, | |
3494fc30 TH |
5664 | !list_empty(&pwq->mayday_node) ? " MAYDAY" : ""); |
5665 | ||
5666 | hash_for_each(pool->busy_hash, bkt, worker, hentry) { | |
5667 | if (worker->current_pwq == pwq) { | |
5668 | has_in_flight = true; | |
5669 | break; | |
5670 | } | |
5671 | } | |
5672 | if (has_in_flight) { | |
5673 | bool comma = false; | |
5674 | ||
5675 | pr_info(" in-flight:"); | |
5676 | hash_for_each(pool->busy_hash, bkt, worker, hentry) { | |
5677 | if (worker->current_pwq != pwq) | |
5678 | continue; | |
5679 | ||
d75f773c | 5680 | pr_cont("%s %d%s:%ps", comma ? "," : "", |
3494fc30 | 5681 | task_pid_nr(worker->task), |
30ae2fc0 | 5682 | worker->rescue_wq ? "(RESCUER)" : "", |
3494fc30 TH |
5683 | worker->current_func); |
5684 | list_for_each_entry(work, &worker->scheduled, entry) | |
c76feb0d PM |
5685 | pr_cont_work(false, work, &pcws); |
5686 | pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); | |
3494fc30 TH |
5687 | comma = true; |
5688 | } | |
5689 | pr_cont("\n"); | |
5690 | } | |
5691 | ||
5692 | list_for_each_entry(work, &pool->worklist, entry) { | |
5693 | if (get_work_pwq(work) == pwq) { | |
5694 | has_pending = true; | |
5695 | break; | |
5696 | } | |
5697 | } | |
5698 | if (has_pending) { | |
5699 | bool comma = false; | |
5700 | ||
5701 | pr_info(" pending:"); | |
5702 | list_for_each_entry(work, &pool->worklist, entry) { | |
5703 | if (get_work_pwq(work) != pwq) | |
5704 | continue; | |
5705 | ||
c76feb0d | 5706 | pr_cont_work(comma, work, &pcws); |
3494fc30 TH |
5707 | comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); |
5708 | } | |
c76feb0d | 5709 | pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); |
3494fc30 TH |
5710 | pr_cont("\n"); |
5711 | } | |
5712 | ||
f97a4a1a | 5713 | if (!list_empty(&pwq->inactive_works)) { |
3494fc30 TH |
5714 | bool comma = false; |
5715 | ||
f97a4a1a LJ |
5716 | pr_info(" inactive:"); |
5717 | list_for_each_entry(work, &pwq->inactive_works, entry) { | |
c76feb0d | 5718 | pr_cont_work(comma, work, &pcws); |
3494fc30 TH |
5719 | comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); |
5720 | } | |
c76feb0d | 5721 | pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); |
3494fc30 TH |
5722 | pr_cont("\n"); |
5723 | } | |
5724 | } | |
5725 | ||
5726 | /** | |
55df0933 IK |
5727 | * show_one_workqueue - dump state of specified workqueue |
5728 | * @wq: workqueue whose state will be printed | |
3494fc30 | 5729 | */ |
55df0933 | 5730 | void show_one_workqueue(struct workqueue_struct *wq) |
3494fc30 | 5731 | { |
55df0933 IK |
5732 | struct pool_workqueue *pwq; |
5733 | bool idle = true; | |
3494fc30 | 5734 | unsigned long flags; |
3494fc30 | 5735 | |
55df0933 | 5736 | for_each_pwq(pwq, wq) { |
afa87ce8 | 5737 | if (!pwq_is_empty(pwq)) { |
55df0933 IK |
5738 | idle = false; |
5739 | break; | |
3494fc30 | 5740 | } |
55df0933 IK |
5741 | } |
5742 | if (idle) /* Nothing to print for idle workqueue */ | |
5743 | return; | |
3494fc30 | 5744 | |
55df0933 | 5745 | pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags); |
3494fc30 | 5746 | |
55df0933 IK |
5747 | for_each_pwq(pwq, wq) { |
5748 | raw_spin_lock_irqsave(&pwq->pool->lock, flags); | |
afa87ce8 | 5749 | if (!pwq_is_empty(pwq)) { |
62635ea8 | 5750 | /* |
55df0933 IK |
5751 | * Defer printing to avoid deadlocks in console |
5752 | * drivers that queue work while holding locks | |
5753 | * also taken in their write paths. | |
62635ea8 | 5754 | */ |
55df0933 IK |
5755 | printk_deferred_enter(); |
5756 | show_pwq(pwq); | |
5757 | printk_deferred_exit(); | |
3494fc30 | 5758 | } |
55df0933 | 5759 | raw_spin_unlock_irqrestore(&pwq->pool->lock, flags); |
62635ea8 SS |
5760 | /* |
5761 | * We could be printing a lot from atomic context, e.g. | |
55df0933 | 5762 | * sysrq-t -> show_all_workqueues(). Avoid triggering |
62635ea8 SS |
5763 | * hard lockup. |
5764 | */ | |
5765 | touch_nmi_watchdog(); | |
3494fc30 TH |
5766 | } |
5767 | ||
55df0933 IK |
5768 | } |
5769 | ||
5770 | /** | |
5771 | * show_one_worker_pool - dump state of specified worker pool | |
5772 | * @pool: worker pool whose state will be printed | |
5773 | */ | |
5774 | static void show_one_worker_pool(struct worker_pool *pool) | |
5775 | { | |
5776 | struct worker *worker; | |
5777 | bool first = true; | |
5778 | unsigned long flags; | |
335a42eb | 5779 | unsigned long hung = 0; |
55df0933 IK |
5780 | |
5781 | raw_spin_lock_irqsave(&pool->lock, flags); | |
5782 | if (pool->nr_workers == pool->nr_idle) | |
5783 | goto next_pool; | |
335a42eb PM |
5784 | |
5785 | /* How long the first pending work is waiting for a worker. */ | |
5786 | if (!list_empty(&pool->worklist)) | |
5787 | hung = jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000; | |
5788 | ||
55df0933 IK |
5789 | /* |
5790 | * Defer printing to avoid deadlocks in console drivers that | |
5791 | * queue work while holding locks also taken in their write | |
5792 | * paths. | |
5793 | */ | |
5794 | printk_deferred_enter(); | |
5795 | pr_info("pool %d:", pool->id); | |
5796 | pr_cont_pool_info(pool); | |
335a42eb | 5797 | pr_cont(" hung=%lus workers=%d", hung, pool->nr_workers); |
55df0933 IK |
5798 | if (pool->manager) |
5799 | pr_cont(" manager: %d", | |
5800 | task_pid_nr(pool->manager->task)); | |
5801 | list_for_each_entry(worker, &pool->idle_list, entry) { | |
5802 | pr_cont(" %s%d", first ? "idle: " : "", | |
5803 | task_pid_nr(worker->task)); | |
5804 | first = false; | |
5805 | } | |
5806 | pr_cont("\n"); | |
5807 | printk_deferred_exit(); | |
5808 | next_pool: | |
5809 | raw_spin_unlock_irqrestore(&pool->lock, flags); | |
5810 | /* | |
5811 | * We could be printing a lot from atomic context, e.g. | |
5812 | * sysrq-t -> show_all_workqueues(). Avoid triggering | |
5813 | * hard lockup. | |
5814 | */ | |
5815 | touch_nmi_watchdog(); | |
5816 | ||
5817 | } | |
5818 | ||
5819 | /** | |
5820 | * show_all_workqueues - dump workqueue state | |
5821 | * | |
704bc669 | 5822 | * Called from a sysrq handler and prints out all busy workqueues and pools. |
55df0933 IK |
5823 | */ |
5824 | void show_all_workqueues(void) | |
5825 | { | |
5826 | struct workqueue_struct *wq; | |
5827 | struct worker_pool *pool; | |
5828 | int pi; | |
5829 | ||
5830 | rcu_read_lock(); | |
5831 | ||
5832 | pr_info("Showing busy workqueues and worker pools:\n"); | |
5833 | ||
5834 | list_for_each_entry_rcu(wq, &workqueues, list) | |
5835 | show_one_workqueue(wq); | |
5836 | ||
5837 | for_each_pool(pool, pi) | |
5838 | show_one_worker_pool(pool); | |
5839 | ||
24acfb71 | 5840 | rcu_read_unlock(); |
3494fc30 TH |
5841 | } |
5842 | ||
704bc669 JL |
5843 | /** |
5844 | * show_freezable_workqueues - dump freezable workqueue state | |
5845 | * | |
5846 | * Called from try_to_freeze_tasks() and prints out all freezable workqueues | |
5847 | * still busy. | |
5848 | */ | |
5849 | void show_freezable_workqueues(void) | |
5850 | { | |
5851 | struct workqueue_struct *wq; | |
5852 | ||
5853 | rcu_read_lock(); | |
5854 | ||
5855 | pr_info("Showing freezable workqueues that are still busy:\n"); | |
5856 | ||
5857 | list_for_each_entry_rcu(wq, &workqueues, list) { | |
5858 | if (!(wq->flags & WQ_FREEZABLE)) | |
5859 | continue; | |
5860 | show_one_workqueue(wq); | |
5861 | } | |
5862 | ||
5863 | rcu_read_unlock(); | |
5864 | } | |
5865 | ||
6b59808b TH |
5866 | /* used to show worker information through /proc/PID/{comm,stat,status} */ |
5867 | void wq_worker_comm(char *buf, size_t size, struct task_struct *task) | |
5868 | { | |
6b59808b TH |
5869 | int off; |
5870 | ||
5871 | /* always show the actual comm */ | |
5872 | off = strscpy(buf, task->comm, size); | |
5873 | if (off < 0) | |
5874 | return; | |
5875 | ||
197f6acc | 5876 | /* stabilize PF_WQ_WORKER and worker pool association */ |
6b59808b TH |
5877 | mutex_lock(&wq_pool_attach_mutex); |
5878 | ||
197f6acc TH |
5879 | if (task->flags & PF_WQ_WORKER) { |
5880 | struct worker *worker = kthread_data(task); | |
5881 | struct worker_pool *pool = worker->pool; | |
6b59808b | 5882 | |
197f6acc | 5883 | if (pool) { |
a9b8a985 | 5884 | raw_spin_lock_irq(&pool->lock); |
197f6acc TH |
5885 | /* |
5886 | * ->desc tracks information (wq name or | |
5887 | * set_worker_desc()) for the latest execution. If | |
5888 | * current, prepend '+', otherwise '-'. | |
5889 | */ | |
5890 | if (worker->desc[0] != '\0') { | |
5891 | if (worker->current_work) | |
5892 | scnprintf(buf + off, size - off, "+%s", | |
5893 | worker->desc); | |
5894 | else | |
5895 | scnprintf(buf + off, size - off, "-%s", | |
5896 | worker->desc); | |
5897 | } | |
a9b8a985 | 5898 | raw_spin_unlock_irq(&pool->lock); |
6b59808b | 5899 | } |
6b59808b TH |
5900 | } |
5901 | ||
5902 | mutex_unlock(&wq_pool_attach_mutex); | |
5903 | } | |
5904 | ||
66448bc2 MM |
5905 | #ifdef CONFIG_SMP |
5906 | ||
db7bccf4 TH |
5907 | /* |
5908 | * CPU hotplug. | |
5909 | * | |
e22bee78 | 5910 | * There are two challenges in supporting CPU hotplug. Firstly, there |
112202d9 | 5911 | * are a lot of assumptions on strong associations among work, pwq and |
706026c2 | 5912 | * pool which make migrating pending and scheduled works very |
e22bee78 | 5913 | * difficult to implement without impacting hot paths. Secondly, |
94cf58bb | 5914 | * worker pools serve mix of short, long and very long running works making |
e22bee78 TH |
5915 | * blocked draining impractical. |
5916 | * | |
24647570 | 5917 | * This is solved by allowing the pools to be disassociated from the CPU |
628c78e7 TH |
5918 | * running as an unbound one and allowing it to be reattached later if the |
5919 | * cpu comes back online. | |
db7bccf4 | 5920 | */ |
1da177e4 | 5921 | |
e8b3f8db | 5922 | static void unbind_workers(int cpu) |
3af24433 | 5923 | { |
4ce62e9e | 5924 | struct worker_pool *pool; |
db7bccf4 | 5925 | struct worker *worker; |
3af24433 | 5926 | |
f02ae73a | 5927 | for_each_cpu_worker_pool(pool, cpu) { |
1258fae7 | 5928 | mutex_lock(&wq_pool_attach_mutex); |
a9b8a985 | 5929 | raw_spin_lock_irq(&pool->lock); |
3af24433 | 5930 | |
94cf58bb | 5931 | /* |
92f9c5c4 | 5932 | * We've blocked all attach/detach operations. Make all workers |
94cf58bb | 5933 | * unbound and set DISASSOCIATED. Before this, all workers |
11b45b0b | 5934 | * must be on the cpu. After this, they may become diasporas. |
b4ac9384 LJ |
5935 | * And the preemption disabled section in their sched callbacks |
5936 | * are guaranteed to see WORKER_UNBOUND since the code here | |
5937 | * is on the same cpu. | |
94cf58bb | 5938 | */ |
da028469 | 5939 | for_each_pool_worker(worker, pool) |
c9e7cf27 | 5940 | worker->flags |= WORKER_UNBOUND; |
06ba38a9 | 5941 | |
24647570 | 5942 | pool->flags |= POOL_DISASSOCIATED; |
f2d5a0ee | 5943 | |
eb283428 | 5944 | /* |
989442d7 LJ |
5945 | * The handling of nr_running in sched callbacks are disabled |
5946 | * now. Zap nr_running. After this, nr_running stays zero and | |
5947 | * need_more_worker() and keep_working() are always true as | |
5948 | * long as the worklist is not empty. This pool now behaves as | |
5949 | * an unbound (in terms of concurrency management) pool which | |
eb283428 LJ |
5950 | * are served by workers tied to the pool. |
5951 | */ | |
bc35f7ef | 5952 | pool->nr_running = 0; |
eb283428 LJ |
5953 | |
5954 | /* | |
5955 | * With concurrency management just turned off, a busy | |
5956 | * worker blocking could lead to lengthy stalls. Kick off | |
5957 | * unbound chain execution of currently pending work items. | |
5958 | */ | |
0219a352 | 5959 | kick_pool(pool); |
989442d7 | 5960 | |
a9b8a985 | 5961 | raw_spin_unlock_irq(&pool->lock); |
989442d7 | 5962 | |
793777bc VS |
5963 | for_each_pool_worker(worker, pool) |
5964 | unbind_worker(worker); | |
989442d7 LJ |
5965 | |
5966 | mutex_unlock(&wq_pool_attach_mutex); | |
eb283428 | 5967 | } |
3af24433 | 5968 | } |
3af24433 | 5969 | |
bd7c089e TH |
5970 | /** |
5971 | * rebind_workers - rebind all workers of a pool to the associated CPU | |
5972 | * @pool: pool of interest | |
5973 | * | |
a9ab775b | 5974 | * @pool->cpu is coming online. Rebind all workers to the CPU. |
bd7c089e TH |
5975 | */ |
5976 | static void rebind_workers(struct worker_pool *pool) | |
5977 | { | |
a9ab775b | 5978 | struct worker *worker; |
bd7c089e | 5979 | |
1258fae7 | 5980 | lockdep_assert_held(&wq_pool_attach_mutex); |
bd7c089e | 5981 | |
a9ab775b TH |
5982 | /* |
5983 | * Restore CPU affinity of all workers. As all idle workers should | |
5984 | * be on the run-queue of the associated CPU before any local | |
402dd89d | 5985 | * wake-ups for concurrency management happen, restore CPU affinity |
a9ab775b TH |
5986 | * of all workers first and then clear UNBOUND. As we're called |
5987 | * from CPU_ONLINE, the following shouldn't fail. | |
5988 | */ | |
c63a2e52 VS |
5989 | for_each_pool_worker(worker, pool) { |
5990 | kthread_set_per_cpu(worker->task, pool->cpu); | |
5991 | WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, | |
9546b29e | 5992 | pool_allowed_cpus(pool)) < 0); |
c63a2e52 | 5993 | } |
bd7c089e | 5994 | |
a9b8a985 | 5995 | raw_spin_lock_irq(&pool->lock); |
f7c17d26 | 5996 | |
3de5e884 | 5997 | pool->flags &= ~POOL_DISASSOCIATED; |
bd7c089e | 5998 | |
da028469 | 5999 | for_each_pool_worker(worker, pool) { |
a9ab775b | 6000 | unsigned int worker_flags = worker->flags; |
bd7c089e | 6001 | |
a9ab775b TH |
6002 | /* |
6003 | * We want to clear UNBOUND but can't directly call | |
6004 | * worker_clr_flags() or adjust nr_running. Atomically | |
6005 | * replace UNBOUND with another NOT_RUNNING flag REBOUND. | |
6006 | * @worker will clear REBOUND using worker_clr_flags() when | |
6007 | * it initiates the next execution cycle thus restoring | |
6008 | * concurrency management. Note that when or whether | |
6009 | * @worker clears REBOUND doesn't affect correctness. | |
6010 | * | |
c95491ed | 6011 | * WRITE_ONCE() is necessary because @worker->flags may be |
a9ab775b | 6012 | * tested without holding any lock in |
6d25be57 | 6013 | * wq_worker_running(). Without it, NOT_RUNNING test may |
a9ab775b TH |
6014 | * fail incorrectly leading to premature concurrency |
6015 | * management operations. | |
6016 | */ | |
6017 | WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); | |
6018 | worker_flags |= WORKER_REBOUND; | |
6019 | worker_flags &= ~WORKER_UNBOUND; | |
c95491ed | 6020 | WRITE_ONCE(worker->flags, worker_flags); |
bd7c089e | 6021 | } |
a9ab775b | 6022 | |
a9b8a985 | 6023 | raw_spin_unlock_irq(&pool->lock); |
bd7c089e TH |
6024 | } |
6025 | ||
7dbc725e TH |
6026 | /** |
6027 | * restore_unbound_workers_cpumask - restore cpumask of unbound workers | |
6028 | * @pool: unbound pool of interest | |
6029 | * @cpu: the CPU which is coming up | |
6030 | * | |
6031 | * An unbound pool may end up with a cpumask which doesn't have any online | |
6032 | * CPUs. When a worker of such pool get scheduled, the scheduler resets | |
6033 | * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any | |
6034 | * online CPU before, cpus_allowed of all its workers should be restored. | |
6035 | */ | |
6036 | static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) | |
6037 | { | |
6038 | static cpumask_t cpumask; | |
6039 | struct worker *worker; | |
7dbc725e | 6040 | |
1258fae7 | 6041 | lockdep_assert_held(&wq_pool_attach_mutex); |
7dbc725e TH |
6042 | |
6043 | /* is @cpu allowed for @pool? */ | |
6044 | if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) | |
6045 | return; | |
6046 | ||
7dbc725e | 6047 | cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); |
7dbc725e TH |
6048 | |
6049 | /* as we're called from CPU_ONLINE, the following shouldn't fail */ | |
da028469 | 6050 | for_each_pool_worker(worker, pool) |
d945b5e9 | 6051 | WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0); |
7dbc725e TH |
6052 | } |
6053 | ||
7ee681b2 TG |
6054 | int workqueue_prepare_cpu(unsigned int cpu) |
6055 | { | |
6056 | struct worker_pool *pool; | |
6057 | ||
6058 | for_each_cpu_worker_pool(pool, cpu) { | |
6059 | if (pool->nr_workers) | |
6060 | continue; | |
6061 | if (!create_worker(pool)) | |
6062 | return -ENOMEM; | |
6063 | } | |
6064 | return 0; | |
6065 | } | |
6066 | ||
6067 | int workqueue_online_cpu(unsigned int cpu) | |
3af24433 | 6068 | { |
4ce62e9e | 6069 | struct worker_pool *pool; |
4c16bd32 | 6070 | struct workqueue_struct *wq; |
7dbc725e | 6071 | int pi; |
3ce63377 | 6072 | |
7ee681b2 | 6073 | mutex_lock(&wq_pool_mutex); |
7dbc725e | 6074 | |
7ee681b2 | 6075 | for_each_pool(pool, pi) { |
1258fae7 | 6076 | mutex_lock(&wq_pool_attach_mutex); |
94cf58bb | 6077 | |
7ee681b2 TG |
6078 | if (pool->cpu == cpu) |
6079 | rebind_workers(pool); | |
6080 | else if (pool->cpu < 0) | |
6081 | restore_unbound_workers_cpumask(pool, cpu); | |
94cf58bb | 6082 | |
1258fae7 | 6083 | mutex_unlock(&wq_pool_attach_mutex); |
7ee681b2 | 6084 | } |
6ba94429 | 6085 | |
fef59c9c | 6086 | /* update pod affinity of unbound workqueues */ |
4cbfd3de | 6087 | list_for_each_entry(wq, &workqueues, list) { |
84193c07 TH |
6088 | struct workqueue_attrs *attrs = wq->unbound_attrs; |
6089 | ||
6090 | if (attrs) { | |
6091 | const struct wq_pod_type *pt = wqattrs_pod_type(attrs); | |
6092 | int tcpu; | |
4cbfd3de | 6093 | |
84193c07 | 6094 | for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]]) |
fef59c9c | 6095 | wq_update_pod(wq, tcpu, cpu, true); |
5797b1c1 TH |
6096 | |
6097 | mutex_lock(&wq->mutex); | |
6098 | wq_update_node_max_active(wq, -1); | |
6099 | mutex_unlock(&wq->mutex); | |
4cbfd3de TH |
6100 | } |
6101 | } | |
6ba94429 | 6102 | |
7ee681b2 TG |
6103 | mutex_unlock(&wq_pool_mutex); |
6104 | return 0; | |
6ba94429 FW |
6105 | } |
6106 | ||
7ee681b2 | 6107 | int workqueue_offline_cpu(unsigned int cpu) |
6ba94429 | 6108 | { |
6ba94429 FW |
6109 | struct workqueue_struct *wq; |
6110 | ||
7ee681b2 | 6111 | /* unbinding per-cpu workers should happen on the local CPU */ |
e8b3f8db LJ |
6112 | if (WARN_ON(cpu != smp_processor_id())) |
6113 | return -1; | |
6114 | ||
6115 | unbind_workers(cpu); | |
7ee681b2 | 6116 | |
fef59c9c | 6117 | /* update pod affinity of unbound workqueues */ |
7ee681b2 | 6118 | mutex_lock(&wq_pool_mutex); |
4cbfd3de | 6119 | list_for_each_entry(wq, &workqueues, list) { |
84193c07 TH |
6120 | struct workqueue_attrs *attrs = wq->unbound_attrs; |
6121 | ||
6122 | if (attrs) { | |
6123 | const struct wq_pod_type *pt = wqattrs_pod_type(attrs); | |
6124 | int tcpu; | |
4cbfd3de | 6125 | |
84193c07 | 6126 | for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]]) |
fef59c9c | 6127 | wq_update_pod(wq, tcpu, cpu, false); |
5797b1c1 TH |
6128 | |
6129 | mutex_lock(&wq->mutex); | |
6130 | wq_update_node_max_active(wq, cpu); | |
6131 | mutex_unlock(&wq->mutex); | |
4cbfd3de TH |
6132 | } |
6133 | } | |
7ee681b2 TG |
6134 | mutex_unlock(&wq_pool_mutex); |
6135 | ||
7ee681b2 | 6136 | return 0; |
6ba94429 FW |
6137 | } |
6138 | ||
6ba94429 FW |
6139 | struct work_for_cpu { |
6140 | struct work_struct work; | |
6141 | long (*fn)(void *); | |
6142 | void *arg; | |
6143 | long ret; | |
6144 | }; | |
6145 | ||
6146 | static void work_for_cpu_fn(struct work_struct *work) | |
6147 | { | |
6148 | struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); | |
6149 | ||
6150 | wfc->ret = wfc->fn(wfc->arg); | |
6151 | } | |
6152 | ||
6153 | /** | |
265f3ed0 | 6154 | * work_on_cpu_key - run a function in thread context on a particular cpu |
6ba94429 FW |
6155 | * @cpu: the cpu to run on |
6156 | * @fn: the function to run | |
6157 | * @arg: the function arg | |
265f3ed0 | 6158 | * @key: The lock class key for lock debugging purposes |
6ba94429 FW |
6159 | * |
6160 | * It is up to the caller to ensure that the cpu doesn't go offline. | |
6161 | * The caller must not hold any locks which would prevent @fn from completing. | |
6162 | * | |
6163 | * Return: The value @fn returns. | |
6164 | */ | |
265f3ed0 FW |
6165 | long work_on_cpu_key(int cpu, long (*fn)(void *), |
6166 | void *arg, struct lock_class_key *key) | |
6ba94429 FW |
6167 | { |
6168 | struct work_for_cpu wfc = { .fn = fn, .arg = arg }; | |
6169 | ||
265f3ed0 | 6170 | INIT_WORK_ONSTACK_KEY(&wfc.work, work_for_cpu_fn, key); |
6ba94429 FW |
6171 | schedule_work_on(cpu, &wfc.work); |
6172 | flush_work(&wfc.work); | |
6173 | destroy_work_on_stack(&wfc.work); | |
6174 | return wfc.ret; | |
6175 | } | |
265f3ed0 | 6176 | EXPORT_SYMBOL_GPL(work_on_cpu_key); |
0e8d6a93 TG |
6177 | |
6178 | /** | |
265f3ed0 | 6179 | * work_on_cpu_safe_key - run a function in thread context on a particular cpu |
0e8d6a93 TG |
6180 | * @cpu: the cpu to run on |
6181 | * @fn: the function to run | |
6182 | * @arg: the function argument | |
265f3ed0 | 6183 | * @key: The lock class key for lock debugging purposes |
0e8d6a93 TG |
6184 | * |
6185 | * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold | |
6186 | * any locks which would prevent @fn from completing. | |
6187 | * | |
6188 | * Return: The value @fn returns. | |
6189 | */ | |
265f3ed0 FW |
6190 | long work_on_cpu_safe_key(int cpu, long (*fn)(void *), |
6191 | void *arg, struct lock_class_key *key) | |
0e8d6a93 TG |
6192 | { |
6193 | long ret = -ENODEV; | |
6194 | ||
ffd8bea8 | 6195 | cpus_read_lock(); |
0e8d6a93 | 6196 | if (cpu_online(cpu)) |
265f3ed0 | 6197 | ret = work_on_cpu_key(cpu, fn, arg, key); |
ffd8bea8 | 6198 | cpus_read_unlock(); |
0e8d6a93 TG |
6199 | return ret; |
6200 | } | |
265f3ed0 | 6201 | EXPORT_SYMBOL_GPL(work_on_cpu_safe_key); |
6ba94429 FW |
6202 | #endif /* CONFIG_SMP */ |
6203 | ||
6204 | #ifdef CONFIG_FREEZER | |
6205 | ||
6206 | /** | |
6207 | * freeze_workqueues_begin - begin freezing workqueues | |
6208 | * | |
6209 | * Start freezing workqueues. After this function returns, all freezable | |
f97a4a1a | 6210 | * workqueues will queue new works to their inactive_works list instead of |
6ba94429 FW |
6211 | * pool->worklist. |
6212 | * | |
6213 | * CONTEXT: | |
6214 | * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. | |
6215 | */ | |
6216 | void freeze_workqueues_begin(void) | |
6217 | { | |
6218 | struct workqueue_struct *wq; | |
6ba94429 FW |
6219 | |
6220 | mutex_lock(&wq_pool_mutex); | |
6221 | ||
6222 | WARN_ON_ONCE(workqueue_freezing); | |
6223 | workqueue_freezing = true; | |
6224 | ||
6225 | list_for_each_entry(wq, &workqueues, list) { | |
6226 | mutex_lock(&wq->mutex); | |
a045a272 | 6227 | wq_adjust_max_active(wq); |
6ba94429 FW |
6228 | mutex_unlock(&wq->mutex); |
6229 | } | |
6230 | ||
6231 | mutex_unlock(&wq_pool_mutex); | |
6232 | } | |
6233 | ||
6234 | /** | |
6235 | * freeze_workqueues_busy - are freezable workqueues still busy? | |
6236 | * | |
6237 | * Check whether freezing is complete. This function must be called | |
6238 | * between freeze_workqueues_begin() and thaw_workqueues(). | |
6239 | * | |
6240 | * CONTEXT: | |
6241 | * Grabs and releases wq_pool_mutex. | |
6242 | * | |
6243 | * Return: | |
6244 | * %true if some freezable workqueues are still busy. %false if freezing | |
6245 | * is complete. | |
6246 | */ | |
6247 | bool freeze_workqueues_busy(void) | |
6248 | { | |
6249 | bool busy = false; | |
6250 | struct workqueue_struct *wq; | |
6251 | struct pool_workqueue *pwq; | |
6252 | ||
6253 | mutex_lock(&wq_pool_mutex); | |
6254 | ||
6255 | WARN_ON_ONCE(!workqueue_freezing); | |
6256 | ||
6257 | list_for_each_entry(wq, &workqueues, list) { | |
6258 | if (!(wq->flags & WQ_FREEZABLE)) | |
6259 | continue; | |
6260 | /* | |
6261 | * nr_active is monotonically decreasing. It's safe | |
6262 | * to peek without lock. | |
6263 | */ | |
24acfb71 | 6264 | rcu_read_lock(); |
6ba94429 FW |
6265 | for_each_pwq(pwq, wq) { |
6266 | WARN_ON_ONCE(pwq->nr_active < 0); | |
6267 | if (pwq->nr_active) { | |
6268 | busy = true; | |
24acfb71 | 6269 | rcu_read_unlock(); |
6ba94429 FW |
6270 | goto out_unlock; |
6271 | } | |
6272 | } | |
24acfb71 | 6273 | rcu_read_unlock(); |
6ba94429 FW |
6274 | } |
6275 | out_unlock: | |
6276 | mutex_unlock(&wq_pool_mutex); | |
6277 | return busy; | |
6278 | } | |
6279 | ||
6280 | /** | |
6281 | * thaw_workqueues - thaw workqueues | |
6282 | * | |
6283 | * Thaw workqueues. Normal queueing is restored and all collected | |
6284 | * frozen works are transferred to their respective pool worklists. | |
6285 | * | |
6286 | * CONTEXT: | |
6287 | * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. | |
6288 | */ | |
6289 | void thaw_workqueues(void) | |
6290 | { | |
6291 | struct workqueue_struct *wq; | |
6ba94429 FW |
6292 | |
6293 | mutex_lock(&wq_pool_mutex); | |
6294 | ||
6295 | if (!workqueue_freezing) | |
6296 | goto out_unlock; | |
6297 | ||
6298 | workqueue_freezing = false; | |
6299 | ||
6300 | /* restore max_active and repopulate worklist */ | |
6301 | list_for_each_entry(wq, &workqueues, list) { | |
6302 | mutex_lock(&wq->mutex); | |
a045a272 | 6303 | wq_adjust_max_active(wq); |
6ba94429 FW |
6304 | mutex_unlock(&wq->mutex); |
6305 | } | |
6306 | ||
6307 | out_unlock: | |
6308 | mutex_unlock(&wq_pool_mutex); | |
6309 | } | |
6310 | #endif /* CONFIG_FREEZER */ | |
6311 | ||
99c621ef | 6312 | static int workqueue_apply_unbound_cpumask(const cpumask_var_t unbound_cpumask) |
042f7df1 LJ |
6313 | { |
6314 | LIST_HEAD(ctxs); | |
6315 | int ret = 0; | |
6316 | struct workqueue_struct *wq; | |
6317 | struct apply_wqattrs_ctx *ctx, *n; | |
6318 | ||
6319 | lockdep_assert_held(&wq_pool_mutex); | |
6320 | ||
6321 | list_for_each_entry(wq, &workqueues, list) { | |
6322 | if (!(wq->flags & WQ_UNBOUND)) | |
6323 | continue; | |
ca10d851 | 6324 | |
042f7df1 | 6325 | /* creating multiple pwqs breaks ordering guarantee */ |
ca10d851 WL |
6326 | if (!list_empty(&wq->pwqs)) { |
6327 | if (wq->flags & __WQ_ORDERED_EXPLICIT) | |
6328 | continue; | |
6329 | wq->flags &= ~__WQ_ORDERED; | |
6330 | } | |
042f7df1 | 6331 | |
99c621ef | 6332 | ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs, unbound_cpumask); |
84193c07 TH |
6333 | if (IS_ERR(ctx)) { |
6334 | ret = PTR_ERR(ctx); | |
042f7df1 LJ |
6335 | break; |
6336 | } | |
6337 | ||
6338 | list_add_tail(&ctx->list, &ctxs); | |
6339 | } | |
6340 | ||
6341 | list_for_each_entry_safe(ctx, n, &ctxs, list) { | |
6342 | if (!ret) | |
6343 | apply_wqattrs_commit(ctx); | |
6344 | apply_wqattrs_cleanup(ctx); | |
6345 | } | |
6346 | ||
99c621ef LJ |
6347 | if (!ret) { |
6348 | mutex_lock(&wq_pool_attach_mutex); | |
6349 | cpumask_copy(wq_unbound_cpumask, unbound_cpumask); | |
6350 | mutex_unlock(&wq_pool_attach_mutex); | |
6351 | } | |
042f7df1 LJ |
6352 | return ret; |
6353 | } | |
6354 | ||
fe28f631 WL |
6355 | /** |
6356 | * workqueue_unbound_exclude_cpumask - Exclude given CPUs from unbound cpumask | |
6357 | * @exclude_cpumask: the cpumask to be excluded from wq_unbound_cpumask | |
6358 | * | |
6359 | * This function can be called from cpuset code to provide a set of isolated | |
6360 | * CPUs that should be excluded from wq_unbound_cpumask. The caller must hold | |
6361 | * either cpus_read_lock or cpus_write_lock. | |
6362 | */ | |
6363 | int workqueue_unbound_exclude_cpumask(cpumask_var_t exclude_cpumask) | |
6364 | { | |
6365 | cpumask_var_t cpumask; | |
6366 | int ret = 0; | |
6367 | ||
6368 | if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) | |
6369 | return -ENOMEM; | |
6370 | ||
6371 | lockdep_assert_cpus_held(); | |
6372 | mutex_lock(&wq_pool_mutex); | |
6373 | ||
6374 | /* Save the current isolated cpumask & export it via sysfs */ | |
6375 | cpumask_copy(wq_isolated_cpumask, exclude_cpumask); | |
6376 | ||
6377 | /* | |
6378 | * If the operation fails, it will fall back to | |
6379 | * wq_requested_unbound_cpumask which is initially set to | |
6380 | * (HK_TYPE_WQ ∩ HK_TYPE_DOMAIN) house keeping mask and rewritten | |
6381 | * by any subsequent write to workqueue/cpumask sysfs file. | |
6382 | */ | |
6383 | if (!cpumask_andnot(cpumask, wq_requested_unbound_cpumask, exclude_cpumask)) | |
6384 | cpumask_copy(cpumask, wq_requested_unbound_cpumask); | |
6385 | if (!cpumask_equal(cpumask, wq_unbound_cpumask)) | |
6386 | ret = workqueue_apply_unbound_cpumask(cpumask); | |
6387 | ||
6388 | mutex_unlock(&wq_pool_mutex); | |
6389 | free_cpumask_var(cpumask); | |
6390 | return ret; | |
6391 | } | |
6392 | ||
63c5484e TH |
6393 | static int parse_affn_scope(const char *val) |
6394 | { | |
6395 | int i; | |
6396 | ||
6397 | for (i = 0; i < ARRAY_SIZE(wq_affn_names); i++) { | |
6398 | if (!strncasecmp(val, wq_affn_names[i], strlen(wq_affn_names[i]))) | |
6399 | return i; | |
6400 | } | |
6401 | return -EINVAL; | |
6402 | } | |
6403 | ||
6404 | static int wq_affn_dfl_set(const char *val, const struct kernel_param *kp) | |
6405 | { | |
523a301e TH |
6406 | struct workqueue_struct *wq; |
6407 | int affn, cpu; | |
63c5484e TH |
6408 | |
6409 | affn = parse_affn_scope(val); | |
6410 | if (affn < 0) | |
6411 | return affn; | |
523a301e TH |
6412 | if (affn == WQ_AFFN_DFL) |
6413 | return -EINVAL; | |
6414 | ||
6415 | cpus_read_lock(); | |
6416 | mutex_lock(&wq_pool_mutex); | |
63c5484e TH |
6417 | |
6418 | wq_affn_dfl = affn; | |
523a301e TH |
6419 | |
6420 | list_for_each_entry(wq, &workqueues, list) { | |
6421 | for_each_online_cpu(cpu) { | |
6422 | wq_update_pod(wq, cpu, cpu, true); | |
6423 | } | |
6424 | } | |
6425 | ||
6426 | mutex_unlock(&wq_pool_mutex); | |
6427 | cpus_read_unlock(); | |
6428 | ||
63c5484e TH |
6429 | return 0; |
6430 | } | |
6431 | ||
6432 | static int wq_affn_dfl_get(char *buffer, const struct kernel_param *kp) | |
6433 | { | |
6434 | return scnprintf(buffer, PAGE_SIZE, "%s\n", wq_affn_names[wq_affn_dfl]); | |
6435 | } | |
6436 | ||
6437 | static const struct kernel_param_ops wq_affn_dfl_ops = { | |
6438 | .set = wq_affn_dfl_set, | |
6439 | .get = wq_affn_dfl_get, | |
6440 | }; | |
6441 | ||
6442 | module_param_cb(default_affinity_scope, &wq_affn_dfl_ops, NULL, 0644); | |
6443 | ||
6ba94429 FW |
6444 | #ifdef CONFIG_SYSFS |
6445 | /* | |
6446 | * Workqueues with WQ_SYSFS flag set is visible to userland via | |
6447 | * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the | |
6448 | * following attributes. | |
6449 | * | |
63c5484e TH |
6450 | * per_cpu RO bool : whether the workqueue is per-cpu or unbound |
6451 | * max_active RW int : maximum number of in-flight work items | |
6ba94429 FW |
6452 | * |
6453 | * Unbound workqueues have the following extra attributes. | |
6454 | * | |
63c5484e TH |
6455 | * nice RW int : nice value of the workers |
6456 | * cpumask RW mask : bitmask of allowed CPUs for the workers | |
6457 | * affinity_scope RW str : worker CPU affinity scope (cache, numa, none) | |
8639eceb | 6458 | * affinity_strict RW bool : worker CPU affinity is strict |
6ba94429 FW |
6459 | */ |
6460 | struct wq_device { | |
6461 | struct workqueue_struct *wq; | |
6462 | struct device dev; | |
6463 | }; | |
6464 | ||
6465 | static struct workqueue_struct *dev_to_wq(struct device *dev) | |
6466 | { | |
6467 | struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); | |
6468 | ||
6469 | return wq_dev->wq; | |
6470 | } | |
6471 | ||
6472 | static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, | |
6473 | char *buf) | |
6474 | { | |
6475 | struct workqueue_struct *wq = dev_to_wq(dev); | |
6476 | ||
6477 | return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); | |
6478 | } | |
6479 | static DEVICE_ATTR_RO(per_cpu); | |
6480 | ||
6481 | static ssize_t max_active_show(struct device *dev, | |
6482 | struct device_attribute *attr, char *buf) | |
6483 | { | |
6484 | struct workqueue_struct *wq = dev_to_wq(dev); | |
6485 | ||
6486 | return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); | |
6487 | } | |
6488 | ||
6489 | static ssize_t max_active_store(struct device *dev, | |
6490 | struct device_attribute *attr, const char *buf, | |
6491 | size_t count) | |
6492 | { | |
6493 | struct workqueue_struct *wq = dev_to_wq(dev); | |
6494 | int val; | |
6495 | ||
6496 | if (sscanf(buf, "%d", &val) != 1 || val <= 0) | |
6497 | return -EINVAL; | |
6498 | ||
6499 | workqueue_set_max_active(wq, val); | |
6500 | return count; | |
6501 | } | |
6502 | static DEVICE_ATTR_RW(max_active); | |
6503 | ||
6504 | static struct attribute *wq_sysfs_attrs[] = { | |
6505 | &dev_attr_per_cpu.attr, | |
6506 | &dev_attr_max_active.attr, | |
6507 | NULL, | |
6508 | }; | |
6509 | ATTRIBUTE_GROUPS(wq_sysfs); | |
6510 | ||
49277a5b WL |
6511 | static void apply_wqattrs_lock(void) |
6512 | { | |
6513 | /* CPUs should stay stable across pwq creations and installations */ | |
6514 | cpus_read_lock(); | |
6515 | mutex_lock(&wq_pool_mutex); | |
6516 | } | |
6517 | ||
6518 | static void apply_wqattrs_unlock(void) | |
6519 | { | |
6520 | mutex_unlock(&wq_pool_mutex); | |
6521 | cpus_read_unlock(); | |
6522 | } | |
6523 | ||
6ba94429 FW |
6524 | static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, |
6525 | char *buf) | |
6526 | { | |
6527 | struct workqueue_struct *wq = dev_to_wq(dev); | |
6528 | int written; | |
6529 | ||
6530 | mutex_lock(&wq->mutex); | |
6531 | written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); | |
6532 | mutex_unlock(&wq->mutex); | |
6533 | ||
6534 | return written; | |
6535 | } | |
6536 | ||
6537 | /* prepare workqueue_attrs for sysfs store operations */ | |
6538 | static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) | |
6539 | { | |
6540 | struct workqueue_attrs *attrs; | |
6541 | ||
899a94fe LJ |
6542 | lockdep_assert_held(&wq_pool_mutex); |
6543 | ||
be69d00d | 6544 | attrs = alloc_workqueue_attrs(); |
6ba94429 FW |
6545 | if (!attrs) |
6546 | return NULL; | |
6547 | ||
6ba94429 | 6548 | copy_workqueue_attrs(attrs, wq->unbound_attrs); |
6ba94429 FW |
6549 | return attrs; |
6550 | } | |
6551 | ||
6552 | static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, | |
6553 | const char *buf, size_t count) | |
6554 | { | |
6555 | struct workqueue_struct *wq = dev_to_wq(dev); | |
6556 | struct workqueue_attrs *attrs; | |
d4d3e257 LJ |
6557 | int ret = -ENOMEM; |
6558 | ||
6559 | apply_wqattrs_lock(); | |
6ba94429 FW |
6560 | |
6561 | attrs = wq_sysfs_prep_attrs(wq); | |
6562 | if (!attrs) | |
d4d3e257 | 6563 | goto out_unlock; |
6ba94429 FW |
6564 | |
6565 | if (sscanf(buf, "%d", &attrs->nice) == 1 && | |
6566 | attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) | |
d4d3e257 | 6567 | ret = apply_workqueue_attrs_locked(wq, attrs); |
6ba94429 FW |
6568 | else |
6569 | ret = -EINVAL; | |
6570 | ||
d4d3e257 LJ |
6571 | out_unlock: |
6572 | apply_wqattrs_unlock(); | |
6ba94429 FW |
6573 | free_workqueue_attrs(attrs); |
6574 | return ret ?: count; | |
6575 | } | |
6576 | ||
6577 | static ssize_t wq_cpumask_show(struct device *dev, | |
6578 | struct device_attribute *attr, char *buf) | |
6579 | { | |
6580 | struct workqueue_struct *wq = dev_to_wq(dev); | |
6581 | int written; | |
6582 | ||
6583 | mutex_lock(&wq->mutex); | |
6584 | written = scnprintf(buf, PAGE_SIZE, "%*pb\n", | |
6585 | cpumask_pr_args(wq->unbound_attrs->cpumask)); | |
6586 | mutex_unlock(&wq->mutex); | |
6587 | return written; | |
6588 | } | |
6589 | ||
6590 | static ssize_t wq_cpumask_store(struct device *dev, | |
6591 | struct device_attribute *attr, | |
6592 | const char *buf, size_t count) | |
6593 | { | |
6594 | struct workqueue_struct *wq = dev_to_wq(dev); | |
6595 | struct workqueue_attrs *attrs; | |
d4d3e257 LJ |
6596 | int ret = -ENOMEM; |
6597 | ||
6598 | apply_wqattrs_lock(); | |
6ba94429 FW |
6599 | |
6600 | attrs = wq_sysfs_prep_attrs(wq); | |
6601 | if (!attrs) | |
d4d3e257 | 6602 | goto out_unlock; |
6ba94429 FW |
6603 | |
6604 | ret = cpumask_parse(buf, attrs->cpumask); | |
6605 | if (!ret) | |
d4d3e257 | 6606 | ret = apply_workqueue_attrs_locked(wq, attrs); |
6ba94429 | 6607 | |
d4d3e257 LJ |
6608 | out_unlock: |
6609 | apply_wqattrs_unlock(); | |
6ba94429 FW |
6610 | free_workqueue_attrs(attrs); |
6611 | return ret ?: count; | |
6612 | } | |
6613 | ||
63c5484e TH |
6614 | static ssize_t wq_affn_scope_show(struct device *dev, |
6615 | struct device_attribute *attr, char *buf) | |
6616 | { | |
6617 | struct workqueue_struct *wq = dev_to_wq(dev); | |
6618 | int written; | |
6619 | ||
6620 | mutex_lock(&wq->mutex); | |
523a301e TH |
6621 | if (wq->unbound_attrs->affn_scope == WQ_AFFN_DFL) |
6622 | written = scnprintf(buf, PAGE_SIZE, "%s (%s)\n", | |
6623 | wq_affn_names[WQ_AFFN_DFL], | |
6624 | wq_affn_names[wq_affn_dfl]); | |
6625 | else | |
6626 | written = scnprintf(buf, PAGE_SIZE, "%s\n", | |
6627 | wq_affn_names[wq->unbound_attrs->affn_scope]); | |
63c5484e TH |
6628 | mutex_unlock(&wq->mutex); |
6629 | ||
6630 | return written; | |
6631 | } | |
6632 | ||
6633 | static ssize_t wq_affn_scope_store(struct device *dev, | |
6634 | struct device_attribute *attr, | |
6635 | const char *buf, size_t count) | |
6636 | { | |
6637 | struct workqueue_struct *wq = dev_to_wq(dev); | |
6638 | struct workqueue_attrs *attrs; | |
6639 | int affn, ret = -ENOMEM; | |
6640 | ||
6641 | affn = parse_affn_scope(buf); | |
6642 | if (affn < 0) | |
6643 | return affn; | |
6644 | ||
6645 | apply_wqattrs_lock(); | |
6646 | attrs = wq_sysfs_prep_attrs(wq); | |
6647 | if (attrs) { | |
6648 | attrs->affn_scope = affn; | |
6649 | ret = apply_workqueue_attrs_locked(wq, attrs); | |
6650 | } | |
6651 | apply_wqattrs_unlock(); | |
6652 | free_workqueue_attrs(attrs); | |
6653 | return ret ?: count; | |
6654 | } | |
6655 | ||
8639eceb TH |
6656 | static ssize_t wq_affinity_strict_show(struct device *dev, |
6657 | struct device_attribute *attr, char *buf) | |
6658 | { | |
6659 | struct workqueue_struct *wq = dev_to_wq(dev); | |
6660 | ||
6661 | return scnprintf(buf, PAGE_SIZE, "%d\n", | |
6662 | wq->unbound_attrs->affn_strict); | |
6663 | } | |
6664 | ||
6665 | static ssize_t wq_affinity_strict_store(struct device *dev, | |
6666 | struct device_attribute *attr, | |
6667 | const char *buf, size_t count) | |
6668 | { | |
6669 | struct workqueue_struct *wq = dev_to_wq(dev); | |
6670 | struct workqueue_attrs *attrs; | |
6671 | int v, ret = -ENOMEM; | |
6672 | ||
6673 | if (sscanf(buf, "%d", &v) != 1) | |
6674 | return -EINVAL; | |
6675 | ||
6676 | apply_wqattrs_lock(); | |
6677 | attrs = wq_sysfs_prep_attrs(wq); | |
6678 | if (attrs) { | |
6679 | attrs->affn_strict = (bool)v; | |
6680 | ret = apply_workqueue_attrs_locked(wq, attrs); | |
6681 | } | |
6682 | apply_wqattrs_unlock(); | |
6683 | free_workqueue_attrs(attrs); | |
6684 | return ret ?: count; | |
6685 | } | |
6686 | ||
6ba94429 | 6687 | static struct device_attribute wq_sysfs_unbound_attrs[] = { |
6ba94429 FW |
6688 | __ATTR(nice, 0644, wq_nice_show, wq_nice_store), |
6689 | __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), | |
63c5484e | 6690 | __ATTR(affinity_scope, 0644, wq_affn_scope_show, wq_affn_scope_store), |
8639eceb | 6691 | __ATTR(affinity_strict, 0644, wq_affinity_strict_show, wq_affinity_strict_store), |
6ba94429 FW |
6692 | __ATTR_NULL, |
6693 | }; | |
8ccad40d | 6694 | |
d412ace1 | 6695 | static const struct bus_type wq_subsys = { |
6ba94429 FW |
6696 | .name = "workqueue", |
6697 | .dev_groups = wq_sysfs_groups, | |
2d3854a3 RR |
6698 | }; |
6699 | ||
49277a5b WL |
6700 | /** |
6701 | * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask | |
6702 | * @cpumask: the cpumask to set | |
6703 | * | |
6704 | * The low-level workqueues cpumask is a global cpumask that limits | |
6705 | * the affinity of all unbound workqueues. This function check the @cpumask | |
6706 | * and apply it to all unbound workqueues and updates all pwqs of them. | |
6707 | * | |
6708 | * Return: 0 - Success | |
6709 | * -EINVAL - Invalid @cpumask | |
6710 | * -ENOMEM - Failed to allocate memory for attrs or pwqs. | |
6711 | */ | |
6712 | static int workqueue_set_unbound_cpumask(cpumask_var_t cpumask) | |
6713 | { | |
6714 | int ret = -EINVAL; | |
6715 | ||
6716 | /* | |
6717 | * Not excluding isolated cpus on purpose. | |
6718 | * If the user wishes to include them, we allow that. | |
6719 | */ | |
6720 | cpumask_and(cpumask, cpumask, cpu_possible_mask); | |
6721 | if (!cpumask_empty(cpumask)) { | |
6722 | apply_wqattrs_lock(); | |
6723 | cpumask_copy(wq_requested_unbound_cpumask, cpumask); | |
6724 | if (cpumask_equal(cpumask, wq_unbound_cpumask)) { | |
6725 | ret = 0; | |
6726 | goto out_unlock; | |
6727 | } | |
6728 | ||
6729 | ret = workqueue_apply_unbound_cpumask(cpumask); | |
6730 | ||
6731 | out_unlock: | |
6732 | apply_wqattrs_unlock(); | |
6733 | } | |
6734 | ||
6735 | return ret; | |
6736 | } | |
6737 | ||
fe28f631 WL |
6738 | static ssize_t __wq_cpumask_show(struct device *dev, |
6739 | struct device_attribute *attr, char *buf, cpumask_var_t mask) | |
b05a7928 FW |
6740 | { |
6741 | int written; | |
6742 | ||
042f7df1 | 6743 | mutex_lock(&wq_pool_mutex); |
fe28f631 | 6744 | written = scnprintf(buf, PAGE_SIZE, "%*pb\n", cpumask_pr_args(mask)); |
042f7df1 | 6745 | mutex_unlock(&wq_pool_mutex); |
b05a7928 FW |
6746 | |
6747 | return written; | |
6748 | } | |
6749 | ||
fe28f631 WL |
6750 | static ssize_t wq_unbound_cpumask_show(struct device *dev, |
6751 | struct device_attribute *attr, char *buf) | |
6752 | { | |
6753 | return __wq_cpumask_show(dev, attr, buf, wq_unbound_cpumask); | |
6754 | } | |
6755 | ||
6756 | static ssize_t wq_requested_cpumask_show(struct device *dev, | |
6757 | struct device_attribute *attr, char *buf) | |
6758 | { | |
6759 | return __wq_cpumask_show(dev, attr, buf, wq_requested_unbound_cpumask); | |
6760 | } | |
6761 | ||
6762 | static ssize_t wq_isolated_cpumask_show(struct device *dev, | |
6763 | struct device_attribute *attr, char *buf) | |
6764 | { | |
6765 | return __wq_cpumask_show(dev, attr, buf, wq_isolated_cpumask); | |
6766 | } | |
6767 | ||
042f7df1 LJ |
6768 | static ssize_t wq_unbound_cpumask_store(struct device *dev, |
6769 | struct device_attribute *attr, const char *buf, size_t count) | |
6770 | { | |
6771 | cpumask_var_t cpumask; | |
6772 | int ret; | |
6773 | ||
6774 | if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) | |
6775 | return -ENOMEM; | |
6776 | ||
6777 | ret = cpumask_parse(buf, cpumask); | |
6778 | if (!ret) | |
6779 | ret = workqueue_set_unbound_cpumask(cpumask); | |
6780 | ||
6781 | free_cpumask_var(cpumask); | |
6782 | return ret ? ret : count; | |
6783 | } | |
6784 | ||
fe28f631 | 6785 | static struct device_attribute wq_sysfs_cpumask_attrs[] = { |
042f7df1 | 6786 | __ATTR(cpumask, 0644, wq_unbound_cpumask_show, |
fe28f631 WL |
6787 | wq_unbound_cpumask_store), |
6788 | __ATTR(cpumask_requested, 0444, wq_requested_cpumask_show, NULL), | |
6789 | __ATTR(cpumask_isolated, 0444, wq_isolated_cpumask_show, NULL), | |
6790 | __ATTR_NULL, | |
6791 | }; | |
b05a7928 | 6792 | |
6ba94429 | 6793 | static int __init wq_sysfs_init(void) |
2d3854a3 | 6794 | { |
686f6697 | 6795 | struct device *dev_root; |
b05a7928 FW |
6796 | int err; |
6797 | ||
6798 | err = subsys_virtual_register(&wq_subsys, NULL); | |
6799 | if (err) | |
6800 | return err; | |
6801 | ||
686f6697 GKH |
6802 | dev_root = bus_get_dev_root(&wq_subsys); |
6803 | if (dev_root) { | |
fe28f631 WL |
6804 | struct device_attribute *attr; |
6805 | ||
6806 | for (attr = wq_sysfs_cpumask_attrs; attr->attr.name; attr++) { | |
6807 | err = device_create_file(dev_root, attr); | |
6808 | if (err) | |
6809 | break; | |
6810 | } | |
686f6697 GKH |
6811 | put_device(dev_root); |
6812 | } | |
6813 | return err; | |
2d3854a3 | 6814 | } |
6ba94429 | 6815 | core_initcall(wq_sysfs_init); |
2d3854a3 | 6816 | |
6ba94429 | 6817 | static void wq_device_release(struct device *dev) |
2d3854a3 | 6818 | { |
6ba94429 | 6819 | struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); |
6b44003e | 6820 | |
6ba94429 | 6821 | kfree(wq_dev); |
2d3854a3 | 6822 | } |
a0a1a5fd TH |
6823 | |
6824 | /** | |
6ba94429 FW |
6825 | * workqueue_sysfs_register - make a workqueue visible in sysfs |
6826 | * @wq: the workqueue to register | |
a0a1a5fd | 6827 | * |
6ba94429 FW |
6828 | * Expose @wq in sysfs under /sys/bus/workqueue/devices. |
6829 | * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set | |
6830 | * which is the preferred method. | |
a0a1a5fd | 6831 | * |
6ba94429 FW |
6832 | * Workqueue user should use this function directly iff it wants to apply |
6833 | * workqueue_attrs before making the workqueue visible in sysfs; otherwise, | |
6834 | * apply_workqueue_attrs() may race against userland updating the | |
6835 | * attributes. | |
6836 | * | |
6837 | * Return: 0 on success, -errno on failure. | |
a0a1a5fd | 6838 | */ |
6ba94429 | 6839 | int workqueue_sysfs_register(struct workqueue_struct *wq) |
a0a1a5fd | 6840 | { |
6ba94429 FW |
6841 | struct wq_device *wq_dev; |
6842 | int ret; | |
a0a1a5fd | 6843 | |
6ba94429 | 6844 | /* |
402dd89d | 6845 | * Adjusting max_active or creating new pwqs by applying |
6ba94429 FW |
6846 | * attributes breaks ordering guarantee. Disallow exposing ordered |
6847 | * workqueues. | |
6848 | */ | |
0a94efb5 | 6849 | if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) |
6ba94429 | 6850 | return -EINVAL; |
a0a1a5fd | 6851 | |
6ba94429 FW |
6852 | wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); |
6853 | if (!wq_dev) | |
6854 | return -ENOMEM; | |
5bcab335 | 6855 | |
6ba94429 FW |
6856 | wq_dev->wq = wq; |
6857 | wq_dev->dev.bus = &wq_subsys; | |
6ba94429 | 6858 | wq_dev->dev.release = wq_device_release; |
23217b44 | 6859 | dev_set_name(&wq_dev->dev, "%s", wq->name); |
a0a1a5fd | 6860 | |
6ba94429 FW |
6861 | /* |
6862 | * unbound_attrs are created separately. Suppress uevent until | |
6863 | * everything is ready. | |
6864 | */ | |
6865 | dev_set_uevent_suppress(&wq_dev->dev, true); | |
a0a1a5fd | 6866 | |
6ba94429 FW |
6867 | ret = device_register(&wq_dev->dev); |
6868 | if (ret) { | |
537f4146 | 6869 | put_device(&wq_dev->dev); |
6ba94429 FW |
6870 | wq->wq_dev = NULL; |
6871 | return ret; | |
6872 | } | |
a0a1a5fd | 6873 | |
6ba94429 FW |
6874 | if (wq->flags & WQ_UNBOUND) { |
6875 | struct device_attribute *attr; | |
a0a1a5fd | 6876 | |
6ba94429 FW |
6877 | for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { |
6878 | ret = device_create_file(&wq_dev->dev, attr); | |
6879 | if (ret) { | |
6880 | device_unregister(&wq_dev->dev); | |
6881 | wq->wq_dev = NULL; | |
6882 | return ret; | |
a0a1a5fd TH |
6883 | } |
6884 | } | |
6885 | } | |
6ba94429 FW |
6886 | |
6887 | dev_set_uevent_suppress(&wq_dev->dev, false); | |
6888 | kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); | |
6889 | return 0; | |
a0a1a5fd TH |
6890 | } |
6891 | ||
6892 | /** | |
6ba94429 FW |
6893 | * workqueue_sysfs_unregister - undo workqueue_sysfs_register() |
6894 | * @wq: the workqueue to unregister | |
a0a1a5fd | 6895 | * |
6ba94429 | 6896 | * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. |
a0a1a5fd | 6897 | */ |
6ba94429 | 6898 | static void workqueue_sysfs_unregister(struct workqueue_struct *wq) |
a0a1a5fd | 6899 | { |
6ba94429 | 6900 | struct wq_device *wq_dev = wq->wq_dev; |
8b03ae3c | 6901 | |
6ba94429 FW |
6902 | if (!wq->wq_dev) |
6903 | return; | |
a0a1a5fd | 6904 | |
6ba94429 FW |
6905 | wq->wq_dev = NULL; |
6906 | device_unregister(&wq_dev->dev); | |
a0a1a5fd | 6907 | } |
6ba94429 FW |
6908 | #else /* CONFIG_SYSFS */ |
6909 | static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } | |
6910 | #endif /* CONFIG_SYSFS */ | |
a0a1a5fd | 6911 | |
82607adc TH |
6912 | /* |
6913 | * Workqueue watchdog. | |
6914 | * | |
6915 | * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal | |
6916 | * flush dependency, a concurrency managed work item which stays RUNNING | |
6917 | * indefinitely. Workqueue stalls can be very difficult to debug as the | |
6918 | * usual warning mechanisms don't trigger and internal workqueue state is | |
6919 | * largely opaque. | |
6920 | * | |
6921 | * Workqueue watchdog monitors all worker pools periodically and dumps | |
6922 | * state if some pools failed to make forward progress for a while where | |
6923 | * forward progress is defined as the first item on ->worklist changing. | |
6924 | * | |
6925 | * This mechanism is controlled through the kernel parameter | |
6926 | * "workqueue.watchdog_thresh" which can be updated at runtime through the | |
6927 | * corresponding sysfs parameter file. | |
6928 | */ | |
6929 | #ifdef CONFIG_WQ_WATCHDOG | |
6930 | ||
82607adc | 6931 | static unsigned long wq_watchdog_thresh = 30; |
5cd79d6a | 6932 | static struct timer_list wq_watchdog_timer; |
82607adc TH |
6933 | |
6934 | static unsigned long wq_watchdog_touched = INITIAL_JIFFIES; | |
6935 | static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES; | |
6936 | ||
cd2440d6 PM |
6937 | /* |
6938 | * Show workers that might prevent the processing of pending work items. | |
6939 | * The only candidates are CPU-bound workers in the running state. | |
6940 | * Pending work items should be handled by another idle worker | |
6941 | * in all other situations. | |
6942 | */ | |
6943 | static void show_cpu_pool_hog(struct worker_pool *pool) | |
6944 | { | |
6945 | struct worker *worker; | |
6946 | unsigned long flags; | |
6947 | int bkt; | |
6948 | ||
6949 | raw_spin_lock_irqsave(&pool->lock, flags); | |
6950 | ||
6951 | hash_for_each(pool->busy_hash, bkt, worker, hentry) { | |
6952 | if (task_is_running(worker->task)) { | |
6953 | /* | |
6954 | * Defer printing to avoid deadlocks in console | |
6955 | * drivers that queue work while holding locks | |
6956 | * also taken in their write paths. | |
6957 | */ | |
6958 | printk_deferred_enter(); | |
6959 | ||
6960 | pr_info("pool %d:\n", pool->id); | |
6961 | sched_show_task(worker->task); | |
6962 | ||
6963 | printk_deferred_exit(); | |
6964 | } | |
6965 | } | |
6966 | ||
6967 | raw_spin_unlock_irqrestore(&pool->lock, flags); | |
6968 | } | |
6969 | ||
6970 | static void show_cpu_pools_hogs(void) | |
6971 | { | |
6972 | struct worker_pool *pool; | |
6973 | int pi; | |
6974 | ||
6975 | pr_info("Showing backtraces of running workers in stalled CPU-bound worker pools:\n"); | |
6976 | ||
6977 | rcu_read_lock(); | |
6978 | ||
6979 | for_each_pool(pool, pi) { | |
6980 | if (pool->cpu_stall) | |
6981 | show_cpu_pool_hog(pool); | |
6982 | ||
6983 | } | |
6984 | ||
6985 | rcu_read_unlock(); | |
6986 | } | |
6987 | ||
82607adc TH |
6988 | static void wq_watchdog_reset_touched(void) |
6989 | { | |
6990 | int cpu; | |
6991 | ||
6992 | wq_watchdog_touched = jiffies; | |
6993 | for_each_possible_cpu(cpu) | |
6994 | per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; | |
6995 | } | |
6996 | ||
5cd79d6a | 6997 | static void wq_watchdog_timer_fn(struct timer_list *unused) |
82607adc TH |
6998 | { |
6999 | unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ; | |
7000 | bool lockup_detected = false; | |
cd2440d6 | 7001 | bool cpu_pool_stall = false; |
940d71c6 | 7002 | unsigned long now = jiffies; |
82607adc TH |
7003 | struct worker_pool *pool; |
7004 | int pi; | |
7005 | ||
7006 | if (!thresh) | |
7007 | return; | |
7008 | ||
7009 | rcu_read_lock(); | |
7010 | ||
7011 | for_each_pool(pool, pi) { | |
7012 | unsigned long pool_ts, touched, ts; | |
7013 | ||
cd2440d6 | 7014 | pool->cpu_stall = false; |
82607adc TH |
7015 | if (list_empty(&pool->worklist)) |
7016 | continue; | |
7017 | ||
940d71c6 SS |
7018 | /* |
7019 | * If a virtual machine is stopped by the host it can look to | |
7020 | * the watchdog like a stall. | |
7021 | */ | |
7022 | kvm_check_and_clear_guest_paused(); | |
7023 | ||
82607adc | 7024 | /* get the latest of pool and touched timestamps */ |
89e28ce6 WQ |
7025 | if (pool->cpu >= 0) |
7026 | touched = READ_ONCE(per_cpu(wq_watchdog_touched_cpu, pool->cpu)); | |
7027 | else | |
7028 | touched = READ_ONCE(wq_watchdog_touched); | |
82607adc | 7029 | pool_ts = READ_ONCE(pool->watchdog_ts); |
82607adc TH |
7030 | |
7031 | if (time_after(pool_ts, touched)) | |
7032 | ts = pool_ts; | |
7033 | else | |
7034 | ts = touched; | |
7035 | ||
82607adc | 7036 | /* did we stall? */ |
940d71c6 | 7037 | if (time_after(now, ts + thresh)) { |
82607adc | 7038 | lockup_detected = true; |
cd2440d6 PM |
7039 | if (pool->cpu >= 0) { |
7040 | pool->cpu_stall = true; | |
7041 | cpu_pool_stall = true; | |
7042 | } | |
82607adc TH |
7043 | pr_emerg("BUG: workqueue lockup - pool"); |
7044 | pr_cont_pool_info(pool); | |
7045 | pr_cont(" stuck for %us!\n", | |
940d71c6 | 7046 | jiffies_to_msecs(now - pool_ts) / 1000); |
82607adc | 7047 | } |
cd2440d6 PM |
7048 | |
7049 | ||
82607adc TH |
7050 | } |
7051 | ||
7052 | rcu_read_unlock(); | |
7053 | ||
7054 | if (lockup_detected) | |
55df0933 | 7055 | show_all_workqueues(); |
82607adc | 7056 | |
cd2440d6 PM |
7057 | if (cpu_pool_stall) |
7058 | show_cpu_pools_hogs(); | |
7059 | ||
82607adc TH |
7060 | wq_watchdog_reset_touched(); |
7061 | mod_timer(&wq_watchdog_timer, jiffies + thresh); | |
7062 | } | |
7063 | ||
cb9d7fd5 | 7064 | notrace void wq_watchdog_touch(int cpu) |
82607adc TH |
7065 | { |
7066 | if (cpu >= 0) | |
7067 | per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; | |
89e28ce6 WQ |
7068 | |
7069 | wq_watchdog_touched = jiffies; | |
82607adc TH |
7070 | } |
7071 | ||
7072 | static void wq_watchdog_set_thresh(unsigned long thresh) | |
7073 | { | |
7074 | wq_watchdog_thresh = 0; | |
7075 | del_timer_sync(&wq_watchdog_timer); | |
7076 | ||
7077 | if (thresh) { | |
7078 | wq_watchdog_thresh = thresh; | |
7079 | wq_watchdog_reset_touched(); | |
7080 | mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ); | |
7081 | } | |
7082 | } | |
7083 | ||
7084 | static int wq_watchdog_param_set_thresh(const char *val, | |
7085 | const struct kernel_param *kp) | |
7086 | { | |
7087 | unsigned long thresh; | |
7088 | int ret; | |
7089 | ||
7090 | ret = kstrtoul(val, 0, &thresh); | |
7091 | if (ret) | |
7092 | return ret; | |
7093 | ||
7094 | if (system_wq) | |
7095 | wq_watchdog_set_thresh(thresh); | |
7096 | else | |
7097 | wq_watchdog_thresh = thresh; | |
7098 | ||
7099 | return 0; | |
7100 | } | |
7101 | ||
7102 | static const struct kernel_param_ops wq_watchdog_thresh_ops = { | |
7103 | .set = wq_watchdog_param_set_thresh, | |
7104 | .get = param_get_ulong, | |
7105 | }; | |
7106 | ||
7107 | module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh, | |
7108 | 0644); | |
7109 | ||
7110 | static void wq_watchdog_init(void) | |
7111 | { | |
5cd79d6a | 7112 | timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE); |
82607adc TH |
7113 | wq_watchdog_set_thresh(wq_watchdog_thresh); |
7114 | } | |
7115 | ||
7116 | #else /* CONFIG_WQ_WATCHDOG */ | |
7117 | ||
7118 | static inline void wq_watchdog_init(void) { } | |
7119 | ||
7120 | #endif /* CONFIG_WQ_WATCHDOG */ | |
7121 | ||
4a6c5607 TH |
7122 | static void __init restrict_unbound_cpumask(const char *name, const struct cpumask *mask) |
7123 | { | |
7124 | if (!cpumask_intersects(wq_unbound_cpumask, mask)) { | |
7125 | pr_warn("workqueue: Restricting unbound_cpumask (%*pb) with %s (%*pb) leaves no CPU, ignoring\n", | |
7126 | cpumask_pr_args(wq_unbound_cpumask), name, cpumask_pr_args(mask)); | |
7127 | return; | |
7128 | } | |
7129 | ||
7130 | cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, mask); | |
7131 | } | |
7132 | ||
3347fa09 TH |
7133 | /** |
7134 | * workqueue_init_early - early init for workqueue subsystem | |
7135 | * | |
2930155b TH |
7136 | * This is the first step of three-staged workqueue subsystem initialization and |
7137 | * invoked as soon as the bare basics - memory allocation, cpumasks and idr are | |
7138 | * up. It sets up all the data structures and system workqueues and allows early | |
7139 | * boot code to create workqueues and queue/cancel work items. Actual work item | |
7140 | * execution starts only after kthreads can be created and scheduled right | |
7141 | * before early initcalls. | |
3347fa09 | 7142 | */ |
2333e829 | 7143 | void __init workqueue_init_early(void) |
1da177e4 | 7144 | { |
84193c07 | 7145 | struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_SYSTEM]; |
7a4e344c TH |
7146 | int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; |
7147 | int i, cpu; | |
c34056a3 | 7148 | |
10cdb157 | 7149 | BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); |
e904e6c2 | 7150 | |
b05a7928 | 7151 | BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL)); |
fe28f631 WL |
7152 | BUG_ON(!alloc_cpumask_var(&wq_requested_unbound_cpumask, GFP_KERNEL)); |
7153 | BUG_ON(!zalloc_cpumask_var(&wq_isolated_cpumask, GFP_KERNEL)); | |
b05a7928 | 7154 | |
4a6c5607 TH |
7155 | cpumask_copy(wq_unbound_cpumask, cpu_possible_mask); |
7156 | restrict_unbound_cpumask("HK_TYPE_WQ", housekeeping_cpumask(HK_TYPE_WQ)); | |
7157 | restrict_unbound_cpumask("HK_TYPE_DOMAIN", housekeeping_cpumask(HK_TYPE_DOMAIN)); | |
ace3c549 | 7158 | if (!cpumask_empty(&wq_cmdline_cpumask)) |
4a6c5607 | 7159 | restrict_unbound_cpumask("workqueue.unbound_cpus", &wq_cmdline_cpumask); |
ace3c549 | 7160 | |
fe28f631 | 7161 | cpumask_copy(wq_requested_unbound_cpumask, wq_unbound_cpumask); |
ace3c549 | 7162 | |
e904e6c2 TH |
7163 | pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); |
7164 | ||
2930155b TH |
7165 | wq_update_pod_attrs_buf = alloc_workqueue_attrs(); |
7166 | BUG_ON(!wq_update_pod_attrs_buf); | |
7167 | ||
7bd20b6b MT |
7168 | /* |
7169 | * If nohz_full is enabled, set power efficient workqueue as unbound. | |
7170 | * This allows workqueue items to be moved to HK CPUs. | |
7171 | */ | |
7172 | if (housekeeping_enabled(HK_TYPE_TICK)) | |
7173 | wq_power_efficient = true; | |
7174 | ||
84193c07 TH |
7175 | /* initialize WQ_AFFN_SYSTEM pods */ |
7176 | pt->pod_cpus = kcalloc(1, sizeof(pt->pod_cpus[0]), GFP_KERNEL); | |
7177 | pt->pod_node = kcalloc(1, sizeof(pt->pod_node[0]), GFP_KERNEL); | |
7178 | pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL); | |
7179 | BUG_ON(!pt->pod_cpus || !pt->pod_node || !pt->cpu_pod); | |
7180 | ||
7181 | BUG_ON(!zalloc_cpumask_var_node(&pt->pod_cpus[0], GFP_KERNEL, NUMA_NO_NODE)); | |
7182 | ||
84193c07 TH |
7183 | pt->nr_pods = 1; |
7184 | cpumask_copy(pt->pod_cpus[0], cpu_possible_mask); | |
7185 | pt->pod_node[0] = NUMA_NO_NODE; | |
7186 | pt->cpu_pod[0] = 0; | |
7187 | ||
706026c2 | 7188 | /* initialize CPU pools */ |
29c91e99 | 7189 | for_each_possible_cpu(cpu) { |
4ce62e9e | 7190 | struct worker_pool *pool; |
8b03ae3c | 7191 | |
7a4e344c | 7192 | i = 0; |
f02ae73a | 7193 | for_each_cpu_worker_pool(pool, cpu) { |
7a4e344c | 7194 | BUG_ON(init_worker_pool(pool)); |
ec22ca5e | 7195 | pool->cpu = cpu; |
29c91e99 | 7196 | cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); |
9546b29e | 7197 | cpumask_copy(pool->attrs->__pod_cpumask, cpumask_of(cpu)); |
7a4e344c | 7198 | pool->attrs->nice = std_nice[i++]; |
8639eceb | 7199 | pool->attrs->affn_strict = true; |
f3f90ad4 | 7200 | pool->node = cpu_to_node(cpu); |
7a4e344c | 7201 | |
9daf9e67 | 7202 | /* alloc pool ID */ |
68e13a67 | 7203 | mutex_lock(&wq_pool_mutex); |
9daf9e67 | 7204 | BUG_ON(worker_pool_assign_id(pool)); |
68e13a67 | 7205 | mutex_unlock(&wq_pool_mutex); |
4ce62e9e | 7206 | } |
8b03ae3c TH |
7207 | } |
7208 | ||
8a2b7538 | 7209 | /* create default unbound and ordered wq attrs */ |
29c91e99 TH |
7210 | for (i = 0; i < NR_STD_WORKER_POOLS; i++) { |
7211 | struct workqueue_attrs *attrs; | |
7212 | ||
be69d00d | 7213 | BUG_ON(!(attrs = alloc_workqueue_attrs())); |
29c91e99 | 7214 | attrs->nice = std_nice[i]; |
29c91e99 | 7215 | unbound_std_wq_attrs[i] = attrs; |
8a2b7538 TH |
7216 | |
7217 | /* | |
7218 | * An ordered wq should have only one pwq as ordering is | |
7219 | * guaranteed by max_active which is enforced by pwqs. | |
8a2b7538 | 7220 | */ |
be69d00d | 7221 | BUG_ON(!(attrs = alloc_workqueue_attrs())); |
8a2b7538 | 7222 | attrs->nice = std_nice[i]; |
af73f5c9 | 7223 | attrs->ordered = true; |
8a2b7538 | 7224 | ordered_wq_attrs[i] = attrs; |
29c91e99 TH |
7225 | } |
7226 | ||
d320c038 | 7227 | system_wq = alloc_workqueue("events", 0, 0); |
1aabe902 | 7228 | system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); |
d320c038 | 7229 | system_long_wq = alloc_workqueue("events_long", 0, 0); |
f3421797 | 7230 | system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, |
636b927e | 7231 | WQ_MAX_ACTIVE); |
24d51add TH |
7232 | system_freezable_wq = alloc_workqueue("events_freezable", |
7233 | WQ_FREEZABLE, 0); | |
0668106c VK |
7234 | system_power_efficient_wq = alloc_workqueue("events_power_efficient", |
7235 | WQ_POWER_EFFICIENT, 0); | |
8318d6a6 | 7236 | system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_pwr_efficient", |
0668106c VK |
7237 | WQ_FREEZABLE | WQ_POWER_EFFICIENT, |
7238 | 0); | |
1aabe902 | 7239 | BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || |
0668106c VK |
7240 | !system_unbound_wq || !system_freezable_wq || |
7241 | !system_power_efficient_wq || | |
7242 | !system_freezable_power_efficient_wq); | |
3347fa09 TH |
7243 | } |
7244 | ||
aa6fde93 TH |
7245 | static void __init wq_cpu_intensive_thresh_init(void) |
7246 | { | |
7247 | unsigned long thresh; | |
7248 | unsigned long bogo; | |
7249 | ||
dd64c873 Z |
7250 | pwq_release_worker = kthread_create_worker(0, "pool_workqueue_release"); |
7251 | BUG_ON(IS_ERR(pwq_release_worker)); | |
7252 | ||
aa6fde93 TH |
7253 | /* if the user set it to a specific value, keep it */ |
7254 | if (wq_cpu_intensive_thresh_us != ULONG_MAX) | |
7255 | return; | |
7256 | ||
7257 | /* | |
7258 | * The default of 10ms is derived from the fact that most modern (as of | |
7259 | * 2023) processors can do a lot in 10ms and that it's just below what | |
7260 | * most consider human-perceivable. However, the kernel also runs on a | |
7261 | * lot slower CPUs including microcontrollers where the threshold is way | |
7262 | * too low. | |
7263 | * | |
7264 | * Let's scale up the threshold upto 1 second if BogoMips is below 4000. | |
7265 | * This is by no means accurate but it doesn't have to be. The mechanism | |
7266 | * is still useful even when the threshold is fully scaled up. Also, as | |
7267 | * the reports would usually be applicable to everyone, some machines | |
7268 | * operating on longer thresholds won't significantly diminish their | |
7269 | * usefulness. | |
7270 | */ | |
7271 | thresh = 10 * USEC_PER_MSEC; | |
7272 | ||
7273 | /* see init/calibrate.c for lpj -> BogoMIPS calculation */ | |
7274 | bogo = max_t(unsigned long, loops_per_jiffy / 500000 * HZ, 1); | |
7275 | if (bogo < 4000) | |
7276 | thresh = min_t(unsigned long, thresh * 4000 / bogo, USEC_PER_SEC); | |
7277 | ||
7278 | pr_debug("wq_cpu_intensive_thresh: lpj=%lu BogoMIPS=%lu thresh_us=%lu\n", | |
7279 | loops_per_jiffy, bogo, thresh); | |
7280 | ||
7281 | wq_cpu_intensive_thresh_us = thresh; | |
7282 | } | |
7283 | ||
3347fa09 TH |
7284 | /** |
7285 | * workqueue_init - bring workqueue subsystem fully online | |
7286 | * | |
2930155b TH |
7287 | * This is the second step of three-staged workqueue subsystem initialization |
7288 | * and invoked as soon as kthreads can be created and scheduled. Workqueues have | |
7289 | * been created and work items queued on them, but there are no kworkers | |
7290 | * executing the work items yet. Populate the worker pools with the initial | |
7291 | * workers and enable future kworker creations. | |
3347fa09 | 7292 | */ |
2333e829 | 7293 | void __init workqueue_init(void) |
3347fa09 | 7294 | { |
2186d9f9 | 7295 | struct workqueue_struct *wq; |
3347fa09 TH |
7296 | struct worker_pool *pool; |
7297 | int cpu, bkt; | |
7298 | ||
aa6fde93 TH |
7299 | wq_cpu_intensive_thresh_init(); |
7300 | ||
2186d9f9 TH |
7301 | mutex_lock(&wq_pool_mutex); |
7302 | ||
2930155b TH |
7303 | /* |
7304 | * Per-cpu pools created earlier could be missing node hint. Fix them | |
7305 | * up. Also, create a rescuer for workqueues that requested it. | |
7306 | */ | |
2186d9f9 TH |
7307 | for_each_possible_cpu(cpu) { |
7308 | for_each_cpu_worker_pool(pool, cpu) { | |
7309 | pool->node = cpu_to_node(cpu); | |
7310 | } | |
7311 | } | |
7312 | ||
40c17f75 | 7313 | list_for_each_entry(wq, &workqueues, list) { |
40c17f75 TH |
7314 | WARN(init_rescuer(wq), |
7315 | "workqueue: failed to create early rescuer for %s", | |
7316 | wq->name); | |
7317 | } | |
2186d9f9 TH |
7318 | |
7319 | mutex_unlock(&wq_pool_mutex); | |
7320 | ||
3347fa09 TH |
7321 | /* create the initial workers */ |
7322 | for_each_online_cpu(cpu) { | |
7323 | for_each_cpu_worker_pool(pool, cpu) { | |
7324 | pool->flags &= ~POOL_DISASSOCIATED; | |
7325 | BUG_ON(!create_worker(pool)); | |
7326 | } | |
7327 | } | |
7328 | ||
7329 | hash_for_each(unbound_pool_hash, bkt, pool, hash_node) | |
7330 | BUG_ON(!create_worker(pool)); | |
7331 | ||
7332 | wq_online = true; | |
82607adc | 7333 | wq_watchdog_init(); |
1da177e4 | 7334 | } |
c4f135d6 | 7335 | |
025e1684 TH |
7336 | /* |
7337 | * Initialize @pt by first initializing @pt->cpu_pod[] with pod IDs according to | |
7338 | * @cpu_shares_pod(). Each subset of CPUs that share a pod is assigned a unique | |
7339 | * and consecutive pod ID. The rest of @pt is initialized accordingly. | |
7340 | */ | |
7341 | static void __init init_pod_type(struct wq_pod_type *pt, | |
7342 | bool (*cpus_share_pod)(int, int)) | |
7343 | { | |
7344 | int cur, pre, cpu, pod; | |
7345 | ||
7346 | pt->nr_pods = 0; | |
7347 | ||
7348 | /* init @pt->cpu_pod[] according to @cpus_share_pod() */ | |
7349 | pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL); | |
7350 | BUG_ON(!pt->cpu_pod); | |
7351 | ||
7352 | for_each_possible_cpu(cur) { | |
7353 | for_each_possible_cpu(pre) { | |
7354 | if (pre >= cur) { | |
7355 | pt->cpu_pod[cur] = pt->nr_pods++; | |
7356 | break; | |
7357 | } | |
7358 | if (cpus_share_pod(cur, pre)) { | |
7359 | pt->cpu_pod[cur] = pt->cpu_pod[pre]; | |
7360 | break; | |
7361 | } | |
7362 | } | |
7363 | } | |
7364 | ||
7365 | /* init the rest to match @pt->cpu_pod[] */ | |
7366 | pt->pod_cpus = kcalloc(pt->nr_pods, sizeof(pt->pod_cpus[0]), GFP_KERNEL); | |
7367 | pt->pod_node = kcalloc(pt->nr_pods, sizeof(pt->pod_node[0]), GFP_KERNEL); | |
7368 | BUG_ON(!pt->pod_cpus || !pt->pod_node); | |
7369 | ||
7370 | for (pod = 0; pod < pt->nr_pods; pod++) | |
7371 | BUG_ON(!zalloc_cpumask_var(&pt->pod_cpus[pod], GFP_KERNEL)); | |
7372 | ||
7373 | for_each_possible_cpu(cpu) { | |
7374 | cpumask_set_cpu(cpu, pt->pod_cpus[pt->cpu_pod[cpu]]); | |
7375 | pt->pod_node[pt->cpu_pod[cpu]] = cpu_to_node(cpu); | |
7376 | } | |
7377 | } | |
7378 | ||
63c5484e TH |
7379 | static bool __init cpus_dont_share(int cpu0, int cpu1) |
7380 | { | |
7381 | return false; | |
7382 | } | |
7383 | ||
7384 | static bool __init cpus_share_smt(int cpu0, int cpu1) | |
7385 | { | |
7386 | #ifdef CONFIG_SCHED_SMT | |
7387 | return cpumask_test_cpu(cpu0, cpu_smt_mask(cpu1)); | |
7388 | #else | |
7389 | return false; | |
7390 | #endif | |
7391 | } | |
7392 | ||
025e1684 TH |
7393 | static bool __init cpus_share_numa(int cpu0, int cpu1) |
7394 | { | |
7395 | return cpu_to_node(cpu0) == cpu_to_node(cpu1); | |
7396 | } | |
7397 | ||
2930155b TH |
7398 | /** |
7399 | * workqueue_init_topology - initialize CPU pods for unbound workqueues | |
7400 | * | |
7401 | * This is the third step of there-staged workqueue subsystem initialization and | |
7402 | * invoked after SMP and topology information are fully initialized. It | |
7403 | * initializes the unbound CPU pods accordingly. | |
7404 | */ | |
7405 | void __init workqueue_init_topology(void) | |
a86feae6 | 7406 | { |
2930155b | 7407 | struct workqueue_struct *wq; |
025e1684 | 7408 | int cpu; |
a86feae6 | 7409 | |
63c5484e TH |
7410 | init_pod_type(&wq_pod_types[WQ_AFFN_CPU], cpus_dont_share); |
7411 | init_pod_type(&wq_pod_types[WQ_AFFN_SMT], cpus_share_smt); | |
7412 | init_pod_type(&wq_pod_types[WQ_AFFN_CACHE], cpus_share_cache); | |
025e1684 | 7413 | init_pod_type(&wq_pod_types[WQ_AFFN_NUMA], cpus_share_numa); |
a86feae6 | 7414 | |
c5f8cd6c TH |
7415 | wq_topo_initialized = true; |
7416 | ||
2930155b | 7417 | mutex_lock(&wq_pool_mutex); |
a86feae6 | 7418 | |
2930155b TH |
7419 | /* |
7420 | * Workqueues allocated earlier would have all CPUs sharing the default | |
7421 | * worker pool. Explicitly call wq_update_pod() on all workqueue and CPU | |
7422 | * combinations to apply per-pod sharing. | |
7423 | */ | |
7424 | list_for_each_entry(wq, &workqueues, list) { | |
5797b1c1 | 7425 | for_each_online_cpu(cpu) |
2930155b | 7426 | wq_update_pod(wq, cpu, cpu, true); |
5797b1c1 TH |
7427 | if (wq->flags & WQ_UNBOUND) { |
7428 | mutex_lock(&wq->mutex); | |
7429 | wq_update_node_max_active(wq, -1); | |
7430 | mutex_unlock(&wq->mutex); | |
2930155b TH |
7431 | } |
7432 | } | |
7433 | ||
7434 | mutex_unlock(&wq_pool_mutex); | |
a86feae6 TH |
7435 | } |
7436 | ||
20bdedaf TH |
7437 | void __warn_flushing_systemwide_wq(void) |
7438 | { | |
7439 | pr_warn("WARNING: Flushing system-wide workqueues will be prohibited in near future.\n"); | |
7440 | dump_stack(); | |
7441 | } | |
c4f135d6 | 7442 | EXPORT_SYMBOL(__warn_flushing_systemwide_wq); |
ace3c549 | 7443 | |
7444 | static int __init workqueue_unbound_cpus_setup(char *str) | |
7445 | { | |
7446 | if (cpulist_parse(str, &wq_cmdline_cpumask) < 0) { | |
7447 | cpumask_clear(&wq_cmdline_cpumask); | |
7448 | pr_warn("workqueue.unbound_cpus: incorrect CPU range, using default\n"); | |
7449 | } | |
7450 | ||
7451 | return 1; | |
7452 | } | |
7453 | __setup("workqueue.unbound_cpus=", workqueue_unbound_cpus_setup); |