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> | |
4cb1ef64 | 32 | #include <linux/interrupt.h> |
1da177e4 LT |
33 | #include <linux/signal.h> |
34 | #include <linux/completion.h> | |
35 | #include <linux/workqueue.h> | |
36 | #include <linux/slab.h> | |
37 | #include <linux/cpu.h> | |
38 | #include <linux/notifier.h> | |
39 | #include <linux/kthread.h> | |
1fa44eca | 40 | #include <linux/hardirq.h> |
46934023 | 41 | #include <linux/mempolicy.h> |
341a5958 | 42 | #include <linux/freezer.h> |
d5abe669 | 43 | #include <linux/debug_locks.h> |
4e6045f1 | 44 | #include <linux/lockdep.h> |
c34056a3 | 45 | #include <linux/idr.h> |
29c91e99 | 46 | #include <linux/jhash.h> |
42f8570f | 47 | #include <linux/hashtable.h> |
76af4d93 | 48 | #include <linux/rculist.h> |
bce90380 | 49 | #include <linux/nodemask.h> |
4c16bd32 | 50 | #include <linux/moduleparam.h> |
3d1cb205 | 51 | #include <linux/uaccess.h> |
c98a9805 | 52 | #include <linux/sched/isolation.h> |
cd2440d6 | 53 | #include <linux/sched/debug.h> |
62635ea8 | 54 | #include <linux/nmi.h> |
940d71c6 | 55 | #include <linux/kvm_para.h> |
aa6fde93 | 56 | #include <linux/delay.h> |
2f34d733 | 57 | #include <linux/irq_work.h> |
e22bee78 | 58 | |
ea138446 | 59 | #include "workqueue_internal.h" |
1da177e4 | 60 | |
e563d0a7 | 61 | enum worker_pool_flags { |
24647570 TH |
62 | /* |
63 | * worker_pool flags | |
bc2ae0f5 | 64 | * |
24647570 | 65 | * A bound pool is either associated or disassociated with its CPU. |
bc2ae0f5 TH |
66 | * While associated (!DISASSOCIATED), all workers are bound to the |
67 | * CPU and none has %WORKER_UNBOUND set and concurrency management | |
68 | * is in effect. | |
69 | * | |
70 | * While DISASSOCIATED, the cpu may be offline and all workers have | |
71 | * %WORKER_UNBOUND set and concurrency management disabled, and may | |
24647570 | 72 | * be executing on any CPU. The pool behaves as an unbound one. |
bc2ae0f5 | 73 | * |
bc3a1afc | 74 | * Note that DISASSOCIATED should be flipped only while holding |
1258fae7 | 75 | * wq_pool_attach_mutex to avoid changing binding state while |
4736cbf7 | 76 | * worker_attach_to_pool() is in progress. |
4cb1ef64 TH |
77 | * |
78 | * As there can only be one concurrent BH execution context per CPU, a | |
79 | * BH pool is per-CPU and always DISASSOCIATED. | |
bc2ae0f5 | 80 | */ |
4cb1ef64 TH |
81 | POOL_BH = 1 << 0, /* is a BH pool */ |
82 | POOL_MANAGER_ACTIVE = 1 << 1, /* being managed */ | |
24647570 | 83 | POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */ |
1acd92d9 | 84 | POOL_BH_DRAINING = 1 << 3, /* draining after CPU offline */ |
e563d0a7 | 85 | }; |
db7bccf4 | 86 | |
e563d0a7 | 87 | enum worker_flags { |
c8e55f36 | 88 | /* worker flags */ |
c8e55f36 TH |
89 | WORKER_DIE = 1 << 1, /* die die die */ |
90 | WORKER_IDLE = 1 << 2, /* is idle */ | |
e22bee78 | 91 | WORKER_PREP = 1 << 3, /* preparing to run works */ |
fb0e7beb | 92 | WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */ |
f3421797 | 93 | WORKER_UNBOUND = 1 << 7, /* worker is unbound */ |
a9ab775b | 94 | WORKER_REBOUND = 1 << 8, /* worker was rebound */ |
e22bee78 | 95 | |
a9ab775b TH |
96 | WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE | |
97 | WORKER_UNBOUND | WORKER_REBOUND, | |
e563d0a7 | 98 | }; |
db7bccf4 | 99 | |
c5f5b942 TH |
100 | enum work_cancel_flags { |
101 | WORK_CANCEL_DELAYED = 1 << 0, /* canceling a delayed_work */ | |
86898fa6 | 102 | WORK_CANCEL_DISABLE = 1 << 1, /* canceling to disable */ |
c5f5b942 TH |
103 | }; |
104 | ||
e563d0a7 | 105 | enum wq_internal_consts { |
e34cdddb | 106 | NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */ |
4ce62e9e | 107 | |
29c91e99 | 108 | UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */ |
c8e55f36 | 109 | BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */ |
db7bccf4 | 110 | |
e22bee78 TH |
111 | MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */ |
112 | IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */ | |
113 | ||
3233cdbd TH |
114 | MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2, |
115 | /* call for help after 10ms | |
116 | (min two ticks) */ | |
e22bee78 TH |
117 | MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */ |
118 | CREATE_COOLDOWN = HZ, /* time to breath after fail */ | |
e22bee78 TH |
119 | |
120 | /* | |
121 | * Rescue workers are used only on emergencies and shared by | |
8698a745 | 122 | * all cpus. Give MIN_NICE. |
e22bee78 | 123 | */ |
8698a745 DY |
124 | RESCUER_NICE_LEVEL = MIN_NICE, |
125 | HIGHPRI_NICE_LEVEL = MIN_NICE, | |
ecf6881f | 126 | |
31c89007 | 127 | WQ_NAME_LEN = 32, |
2a1b02bc | 128 | WORKER_ID_LEN = 10 + WQ_NAME_LEN, /* "kworker/R-" + WQ_NAME_LEN */ |
c8e55f36 | 129 | }; |
1da177e4 | 130 | |
4cb1ef64 TH |
131 | /* |
132 | * We don't want to trap softirq for too long. See MAX_SOFTIRQ_TIME and | |
133 | * MAX_SOFTIRQ_RESTART in kernel/softirq.c. These are macros because | |
134 | * msecs_to_jiffies() can't be an initializer. | |
135 | */ | |
136 | #define BH_WORKER_JIFFIES msecs_to_jiffies(2) | |
137 | #define BH_WORKER_RESTARTS 10 | |
138 | ||
1da177e4 | 139 | /* |
4690c4ab TH |
140 | * Structure fields follow one of the following exclusion rules. |
141 | * | |
e41e704b TH |
142 | * I: Modifiable by initialization/destruction paths and read-only for |
143 | * everyone else. | |
4690c4ab | 144 | * |
e22bee78 TH |
145 | * P: Preemption protected. Disabling preemption is enough and should |
146 | * only be modified and accessed from the local cpu. | |
147 | * | |
d565ed63 | 148 | * L: pool->lock protected. Access with pool->lock held. |
4690c4ab | 149 | * |
5797b1c1 TH |
150 | * LN: pool->lock and wq_node_nr_active->lock protected for writes. Either for |
151 | * reads. | |
152 | * | |
bdf8b9bf TH |
153 | * K: Only modified by worker while holding pool->lock. Can be safely read by |
154 | * self, while holding pool->lock or from IRQ context if %current is the | |
155 | * kworker. | |
156 | * | |
157 | * S: Only modified by worker self. | |
158 | * | |
1258fae7 | 159 | * A: wq_pool_attach_mutex protected. |
822d8405 | 160 | * |
68e13a67 | 161 | * PL: wq_pool_mutex protected. |
5bcab335 | 162 | * |
24acfb71 | 163 | * PR: wq_pool_mutex protected for writes. RCU protected for reads. |
76af4d93 | 164 | * |
5b95e1af LJ |
165 | * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads. |
166 | * | |
167 | * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or | |
24acfb71 | 168 | * RCU for reads. |
5b95e1af | 169 | * |
3c25a55d LJ |
170 | * WQ: wq->mutex protected. |
171 | * | |
24acfb71 | 172 | * WR: wq->mutex protected for writes. RCU protected for reads. |
2e109a28 | 173 | * |
a045a272 TH |
174 | * WO: wq->mutex protected for writes. Updated with WRITE_ONCE() and can be read |
175 | * with READ_ONCE() without locking. | |
176 | * | |
2e109a28 | 177 | * MD: wq_mayday_lock protected. |
cd2440d6 PM |
178 | * |
179 | * WD: Used internally by the watchdog. | |
1da177e4 | 180 | */ |
1da177e4 | 181 | |
2eaebdb3 | 182 | /* struct worker is defined in workqueue_internal.h */ |
c34056a3 | 183 | |
bd7bdd43 | 184 | struct worker_pool { |
a9b8a985 | 185 | raw_spinlock_t lock; /* the pool lock */ |
d84ff051 | 186 | int cpu; /* I: the associated cpu */ |
f3f90ad4 | 187 | int node; /* I: the associated node ID */ |
9daf9e67 | 188 | int id; /* I: pool ID */ |
bc8b50c2 | 189 | unsigned int flags; /* L: flags */ |
bd7bdd43 | 190 | |
82607adc | 191 | unsigned long watchdog_ts; /* L: watchdog timestamp */ |
cd2440d6 | 192 | bool cpu_stall; /* WD: stalled cpu bound pool */ |
82607adc | 193 | |
bc35f7ef LJ |
194 | /* |
195 | * The counter is incremented in a process context on the associated CPU | |
196 | * w/ preemption disabled, and decremented or reset in the same context | |
197 | * but w/ pool->lock held. The readers grab pool->lock and are | |
198 | * guaranteed to see if the counter reached zero. | |
199 | */ | |
200 | int nr_running; | |
84f91c62 | 201 | |
bd7bdd43 | 202 | struct list_head worklist; /* L: list of pending works */ |
ea1abd61 | 203 | |
5826cc8f LJ |
204 | int nr_workers; /* L: total number of workers */ |
205 | int nr_idle; /* L: currently idle workers */ | |
bd7bdd43 | 206 | |
2c1f1a91 | 207 | struct list_head idle_list; /* L: list of idle workers */ |
bd7bdd43 | 208 | struct timer_list idle_timer; /* L: worker idle timeout */ |
3f959aa3 VS |
209 | struct work_struct idle_cull_work; /* L: worker idle cleanup */ |
210 | ||
211 | struct timer_list mayday_timer; /* L: SOS timer for workers */ | |
bd7bdd43 | 212 | |
c5aa87bb | 213 | /* a workers is either on busy_hash or idle_list, or the manager */ |
c9e7cf27 TH |
214 | DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER); |
215 | /* L: hash of busy workers */ | |
216 | ||
2607d7a6 | 217 | struct worker *manager; /* L: purely informational */ |
92f9c5c4 | 218 | struct list_head workers; /* A: attached workers */ |
e19e397a | 219 | |
7cda9aae | 220 | struct ida worker_ida; /* worker IDs for task name */ |
e19e397a | 221 | |
7a4e344c | 222 | struct workqueue_attrs *attrs; /* I: worker attributes */ |
68e13a67 LJ |
223 | struct hlist_node hash_node; /* PL: unbound_pool_hash node */ |
224 | int refcnt; /* PL: refcnt for unbound pools */ | |
7a4e344c | 225 | |
29c91e99 | 226 | /* |
24acfb71 | 227 | * Destruction of pool is RCU protected to allow dereferences |
29c91e99 TH |
228 | * from get_work_pool(). |
229 | */ | |
230 | struct rcu_head rcu; | |
84f91c62 | 231 | }; |
8b03ae3c | 232 | |
725e8ec5 TH |
233 | /* |
234 | * Per-pool_workqueue statistics. These can be monitored using | |
235 | * tools/workqueue/wq_monitor.py. | |
236 | */ | |
237 | enum pool_workqueue_stats { | |
238 | PWQ_STAT_STARTED, /* work items started execution */ | |
239 | PWQ_STAT_COMPLETED, /* work items completed execution */ | |
8a1dd1e5 | 240 | PWQ_STAT_CPU_TIME, /* total CPU time consumed */ |
616db877 | 241 | PWQ_STAT_CPU_INTENSIVE, /* wq_cpu_intensive_thresh_us violations */ |
725e8ec5 | 242 | PWQ_STAT_CM_WAKEUP, /* concurrency-management worker wakeups */ |
8639eceb | 243 | PWQ_STAT_REPATRIATED, /* unbound workers brought back into scope */ |
725e8ec5 TH |
244 | PWQ_STAT_MAYDAY, /* maydays to rescuer */ |
245 | PWQ_STAT_RESCUED, /* linked work items executed by rescuer */ | |
246 | ||
247 | PWQ_NR_STATS, | |
248 | }; | |
249 | ||
1da177e4 | 250 | /* |
e9a8e01f | 251 | * The per-pool workqueue. While queued, bits below WORK_PWQ_SHIFT |
112202d9 TH |
252 | * of work_struct->data are used for flags and the remaining high bits |
253 | * point to the pwq; thus, pwqs need to be aligned at two's power of the | |
254 | * number of flag bits. | |
1da177e4 | 255 | */ |
112202d9 | 256 | struct pool_workqueue { |
bd7bdd43 | 257 | struct worker_pool *pool; /* I: the associated pool */ |
4690c4ab | 258 | struct workqueue_struct *wq; /* I: the owning workqueue */ |
73f53c4a TH |
259 | int work_color; /* L: current color */ |
260 | int flush_color; /* L: flushing color */ | |
8864b4e5 | 261 | int refcnt; /* L: reference count */ |
73f53c4a TH |
262 | int nr_in_flight[WORK_NR_COLORS]; |
263 | /* L: nr of in_flight works */ | |
4c065dbc | 264 | bool plugged; /* L: execution suspended */ |
018f3a13 LJ |
265 | |
266 | /* | |
267 | * nr_active management and WORK_STRUCT_INACTIVE: | |
268 | * | |
269 | * When pwq->nr_active >= max_active, new work item is queued to | |
270 | * pwq->inactive_works instead of pool->worklist and marked with | |
271 | * WORK_STRUCT_INACTIVE. | |
272 | * | |
5797b1c1 TH |
273 | * All work items marked with WORK_STRUCT_INACTIVE do not participate in |
274 | * nr_active and all work items in pwq->inactive_works are marked with | |
275 | * WORK_STRUCT_INACTIVE. But not all WORK_STRUCT_INACTIVE work items are | |
276 | * in pwq->inactive_works. Some of them are ready to run in | |
277 | * pool->worklist or worker->scheduled. Those work itmes are only struct | |
278 | * wq_barrier which is used for flush_work() and should not participate | |
279 | * in nr_active. For non-barrier work item, it is marked with | |
280 | * WORK_STRUCT_INACTIVE iff it is in pwq->inactive_works. | |
018f3a13 | 281 | */ |
1e19ffc6 | 282 | int nr_active; /* L: nr of active works */ |
f97a4a1a | 283 | struct list_head inactive_works; /* L: inactive works */ |
5797b1c1 | 284 | struct list_head pending_node; /* LN: node on wq_node_nr_active->pending_pwqs */ |
3c25a55d | 285 | struct list_head pwqs_node; /* WR: node on wq->pwqs */ |
2e109a28 | 286 | struct list_head mayday_node; /* MD: node on wq->maydays */ |
8864b4e5 | 287 | |
725e8ec5 TH |
288 | u64 stats[PWQ_NR_STATS]; |
289 | ||
8864b4e5 | 290 | /* |
967b494e | 291 | * Release of unbound pwq is punted to a kthread_worker. See put_pwq() |
687a9aa5 TH |
292 | * and pwq_release_workfn() for details. pool_workqueue itself is also |
293 | * RCU protected so that the first pwq can be determined without | |
967b494e | 294 | * grabbing wq->mutex. |
8864b4e5 | 295 | */ |
687a9aa5 | 296 | struct kthread_work release_work; |
8864b4e5 | 297 | struct rcu_head rcu; |
e9a8e01f | 298 | } __aligned(1 << WORK_STRUCT_PWQ_SHIFT); |
1da177e4 | 299 | |
73f53c4a TH |
300 | /* |
301 | * Structure used to wait for workqueue flush. | |
302 | */ | |
303 | struct wq_flusher { | |
3c25a55d LJ |
304 | struct list_head list; /* WQ: list of flushers */ |
305 | int flush_color; /* WQ: flush color waiting for */ | |
73f53c4a TH |
306 | struct completion done; /* flush completion */ |
307 | }; | |
308 | ||
226223ab TH |
309 | struct wq_device; |
310 | ||
91ccc6e7 TH |
311 | /* |
312 | * Unlike in a per-cpu workqueue where max_active limits its concurrency level | |
313 | * on each CPU, in an unbound workqueue, max_active applies to the whole system. | |
314 | * As sharing a single nr_active across multiple sockets can be very expensive, | |
315 | * the counting and enforcement is per NUMA node. | |
5797b1c1 TH |
316 | * |
317 | * The following struct is used to enforce per-node max_active. When a pwq wants | |
318 | * to start executing a work item, it should increment ->nr using | |
319 | * tryinc_node_nr_active(). If acquisition fails due to ->nr already being over | |
320 | * ->max, the pwq is queued on ->pending_pwqs. As in-flight work items finish | |
321 | * and decrement ->nr, node_activate_pending_pwq() activates the pending pwqs in | |
322 | * round-robin order. | |
91ccc6e7 TH |
323 | */ |
324 | struct wq_node_nr_active { | |
5797b1c1 TH |
325 | int max; /* per-node max_active */ |
326 | atomic_t nr; /* per-node nr_active */ | |
327 | raw_spinlock_t lock; /* nests inside pool locks */ | |
328 | struct list_head pending_pwqs; /* LN: pwqs with inactive works */ | |
91ccc6e7 TH |
329 | }; |
330 | ||
1da177e4 | 331 | /* |
c5aa87bb TH |
332 | * The externally visible workqueue. It relays the issued work items to |
333 | * the appropriate worker_pool through its pool_workqueues. | |
1da177e4 LT |
334 | */ |
335 | struct workqueue_struct { | |
3c25a55d | 336 | struct list_head pwqs; /* WR: all pwqs of this wq */ |
e2dca7ad | 337 | struct list_head list; /* PR: list of all workqueues */ |
73f53c4a | 338 | |
3c25a55d LJ |
339 | struct mutex mutex; /* protects this wq */ |
340 | int work_color; /* WQ: current work color */ | |
341 | int flush_color; /* WQ: current flush color */ | |
112202d9 | 342 | atomic_t nr_pwqs_to_flush; /* flush in progress */ |
3c25a55d LJ |
343 | struct wq_flusher *first_flusher; /* WQ: first flusher */ |
344 | struct list_head flusher_queue; /* WQ: flush waiters */ | |
345 | struct list_head flusher_overflow; /* WQ: flush overflow list */ | |
73f53c4a | 346 | |
2e109a28 | 347 | struct list_head maydays; /* MD: pwqs requesting rescue */ |
30ae2fc0 | 348 | struct worker *rescuer; /* MD: rescue worker */ |
e22bee78 | 349 | |
87fc741e | 350 | int nr_drainers; /* WQ: drain in progress */ |
5797b1c1 TH |
351 | |
352 | /* See alloc_workqueue() function comment for info on min/max_active */ | |
a045a272 | 353 | int max_active; /* WO: max active works */ |
5797b1c1 | 354 | int min_active; /* WO: min active works */ |
a045a272 | 355 | int saved_max_active; /* WQ: saved max_active */ |
5797b1c1 | 356 | int saved_min_active; /* WQ: saved min_active */ |
226223ab | 357 | |
5b95e1af | 358 | struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */ |
9f66cff2 | 359 | struct pool_workqueue __rcu *dfl_pwq; /* PW: only for unbound wqs */ |
6029a918 | 360 | |
226223ab TH |
361 | #ifdef CONFIG_SYSFS |
362 | struct wq_device *wq_dev; /* I: for sysfs interface */ | |
363 | #endif | |
4e6045f1 | 364 | #ifdef CONFIG_LOCKDEP |
669de8bd BVA |
365 | char *lock_name; |
366 | struct lock_class_key key; | |
4f022f43 MB |
367 | struct lockdep_map __lockdep_map; |
368 | struct lockdep_map *lockdep_map; | |
4e6045f1 | 369 | #endif |
ecf6881f | 370 | char name[WQ_NAME_LEN]; /* I: workqueue name */ |
2728fd2f | 371 | |
e2dca7ad | 372 | /* |
24acfb71 TG |
373 | * Destruction of workqueue_struct is RCU protected to allow walking |
374 | * the workqueues list without grabbing wq_pool_mutex. | |
e2dca7ad TH |
375 | * This is used to dump all workqueues from sysrq. |
376 | */ | |
377 | struct rcu_head rcu; | |
378 | ||
2728fd2f TH |
379 | /* hot fields used during command issue, aligned to cacheline */ |
380 | unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */ | |
c4c8f369 | 381 | struct pool_workqueue __rcu * __percpu *cpu_pwq; /* I: per-cpu pwqs */ |
91ccc6e7 | 382 | struct wq_node_nr_active *node_nr_active[]; /* I: per-node nr_active */ |
1da177e4 LT |
383 | }; |
384 | ||
84193c07 TH |
385 | /* |
386 | * Each pod type describes how CPUs should be grouped for unbound workqueues. | |
387 | * See the comment above workqueue_attrs->affn_scope. | |
388 | */ | |
389 | struct wq_pod_type { | |
390 | int nr_pods; /* number of pods */ | |
391 | cpumask_var_t *pod_cpus; /* pod -> cpus */ | |
392 | int *pod_node; /* pod -> node */ | |
393 | int *cpu_pod; /* cpu -> pod */ | |
394 | }; | |
395 | ||
1211f3b2 TH |
396 | struct work_offq_data { |
397 | u32 pool_id; | |
86898fa6 | 398 | u32 disable; |
1211f3b2 TH |
399 | u32 flags; |
400 | }; | |
401 | ||
63c5484e | 402 | static const char *wq_affn_names[WQ_AFFN_NR_TYPES] = { |
c7a40c49 TH |
403 | [WQ_AFFN_DFL] = "default", |
404 | [WQ_AFFN_CPU] = "cpu", | |
405 | [WQ_AFFN_SMT] = "smt", | |
406 | [WQ_AFFN_CACHE] = "cache", | |
407 | [WQ_AFFN_NUMA] = "numa", | |
408 | [WQ_AFFN_SYSTEM] = "system", | |
63c5484e | 409 | }; |
bce90380 | 410 | |
616db877 TH |
411 | /* |
412 | * Per-cpu work items which run for longer than the following threshold are | |
413 | * automatically considered CPU intensive and excluded from concurrency | |
414 | * management to prevent them from noticeably delaying other per-cpu work items. | |
aa6fde93 TH |
415 | * ULONG_MAX indicates that the user hasn't overridden it with a boot parameter. |
416 | * The actual value is initialized in wq_cpu_intensive_thresh_init(). | |
616db877 | 417 | */ |
aa6fde93 | 418 | static unsigned long wq_cpu_intensive_thresh_us = ULONG_MAX; |
616db877 | 419 | module_param_named(cpu_intensive_thresh_us, wq_cpu_intensive_thresh_us, ulong, 0644); |
ccdec921 XY |
420 | #ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT |
421 | static unsigned int wq_cpu_intensive_warning_thresh = 4; | |
422 | module_param_named(cpu_intensive_warning_thresh, wq_cpu_intensive_warning_thresh, uint, 0644); | |
423 | #endif | |
616db877 | 424 | |
cee22a15 | 425 | /* see the comment above the definition of WQ_POWER_EFFICIENT */ |
552f530c | 426 | static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT); |
cee22a15 VK |
427 | module_param_named(power_efficient, wq_power_efficient, bool, 0444); |
428 | ||
863b710b | 429 | static bool wq_online; /* can kworkers be created yet? */ |
c7a40c49 TH |
430 | static bool wq_topo_initialized __read_mostly = false; |
431 | ||
432 | static struct kmem_cache *pwq_cache; | |
433 | ||
434 | static struct wq_pod_type wq_pod_types[WQ_AFFN_NR_TYPES]; | |
435 | static enum wq_affn_scope wq_affn_dfl = WQ_AFFN_CACHE; | |
3347fa09 | 436 | |
fef59c9c | 437 | /* buf for wq_update_unbound_pod_attrs(), protected by CPU hotplug exclusion */ |
b2b1f933 | 438 | static struct workqueue_attrs *unbound_wq_update_pwq_attrs_buf; |
4c16bd32 | 439 | |
68e13a67 | 440 | static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ |
1258fae7 | 441 | static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */ |
a9b8a985 | 442 | static DEFINE_RAW_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ |
d8bb65ab SAS |
443 | /* wait for manager to go away */ |
444 | static struct rcuwait manager_wait = __RCUWAIT_INITIALIZER(manager_wait); | |
5bcab335 | 445 | |
e2dca7ad | 446 | static LIST_HEAD(workqueues); /* PR: list of all workqueues */ |
68e13a67 | 447 | static bool workqueue_freezing; /* PL: have wqs started freezing? */ |
7d19c5ce | 448 | |
8d84baf7 LJ |
449 | /* PL: mirror the cpu_online_mask excluding the CPU in the midst of hotplugging */ |
450 | static cpumask_var_t wq_online_cpumask; | |
451 | ||
99c621ef | 452 | /* PL&A: allowable cpus for unbound wqs and work items */ |
ef557180 MG |
453 | static cpumask_var_t wq_unbound_cpumask; |
454 | ||
fe28f631 WL |
455 | /* PL: user requested unbound cpumask via sysfs */ |
456 | static cpumask_var_t wq_requested_unbound_cpumask; | |
457 | ||
458 | /* PL: isolated cpumask to be excluded from unbound cpumask */ | |
459 | static cpumask_var_t wq_isolated_cpumask; | |
460 | ||
ace3c549 | 461 | /* for further constrain wq_unbound_cpumask by cmdline parameter*/ |
462 | static struct cpumask wq_cmdline_cpumask __initdata; | |
463 | ||
ef557180 MG |
464 | /* CPU where unbound work was last round robin scheduled from this CPU */ |
465 | static DEFINE_PER_CPU(int, wq_rr_cpu_last); | |
b05a7928 | 466 | |
f303fccb TH |
467 | /* |
468 | * Local execution of unbound work items is no longer guaranteed. The | |
469 | * following always forces round-robin CPU selection on unbound work items | |
470 | * to uncover usages which depend on it. | |
471 | */ | |
472 | #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU | |
473 | static bool wq_debug_force_rr_cpu = true; | |
474 | #else | |
475 | static bool wq_debug_force_rr_cpu = false; | |
476 | #endif | |
477 | module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644); | |
478 | ||
2f34d733 | 479 | /* to raise softirq for the BH worker pools on other CPUs */ |
b4722b85 | 480 | static DEFINE_PER_CPU_SHARED_ALIGNED(struct irq_work [NR_STD_WORKER_POOLS], bh_pool_irq_works); |
2f34d733 | 481 | |
4cb1ef64 | 482 | /* the BH worker pools */ |
b4722b85 | 483 | static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], bh_worker_pools); |
4cb1ef64 | 484 | |
7d19c5ce | 485 | /* the per-cpu worker pools */ |
b4722b85 | 486 | static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools); |
7d19c5ce | 487 | |
68e13a67 | 488 | static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ |
7d19c5ce | 489 | |
68e13a67 | 490 | /* PL: hash of all unbound pools keyed by pool->attrs */ |
29c91e99 TH |
491 | static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); |
492 | ||
c5aa87bb | 493 | /* I: attributes used when instantiating standard unbound pools on demand */ |
29c91e99 TH |
494 | static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; |
495 | ||
8a2b7538 TH |
496 | /* I: attributes used when instantiating ordered pools on demand */ |
497 | static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS]; | |
498 | ||
967b494e TH |
499 | /* |
500 | * I: kthread_worker to release pwq's. pwq release needs to be bounced to a | |
501 | * process context while holding a pool lock. Bounce to a dedicated kthread | |
502 | * worker to avoid A-A deadlocks. | |
503 | */ | |
68279f9c | 504 | static struct kthread_worker *pwq_release_worker __ro_after_init; |
967b494e | 505 | |
68279f9c | 506 | struct workqueue_struct *system_wq __ro_after_init; |
ad7b1f84 | 507 | EXPORT_SYMBOL(system_wq); |
68279f9c | 508 | struct workqueue_struct *system_highpri_wq __ro_after_init; |
1aabe902 | 509 | EXPORT_SYMBOL_GPL(system_highpri_wq); |
68279f9c | 510 | struct workqueue_struct *system_long_wq __ro_after_init; |
d320c038 | 511 | EXPORT_SYMBOL_GPL(system_long_wq); |
68279f9c | 512 | struct workqueue_struct *system_unbound_wq __ro_after_init; |
f3421797 | 513 | EXPORT_SYMBOL_GPL(system_unbound_wq); |
68279f9c | 514 | struct workqueue_struct *system_freezable_wq __ro_after_init; |
24d51add | 515 | EXPORT_SYMBOL_GPL(system_freezable_wq); |
68279f9c | 516 | struct workqueue_struct *system_power_efficient_wq __ro_after_init; |
0668106c | 517 | EXPORT_SYMBOL_GPL(system_power_efficient_wq); |
68279f9c | 518 | struct workqueue_struct *system_freezable_power_efficient_wq __ro_after_init; |
0668106c | 519 | EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq); |
4cb1ef64 TH |
520 | struct workqueue_struct *system_bh_wq; |
521 | EXPORT_SYMBOL_GPL(system_bh_wq); | |
522 | struct workqueue_struct *system_bh_highpri_wq; | |
523 | EXPORT_SYMBOL_GPL(system_bh_highpri_wq); | |
d320c038 | 524 | |
7d19c5ce | 525 | static int worker_thread(void *__worker); |
6ba94429 | 526 | static void workqueue_sysfs_unregister(struct workqueue_struct *wq); |
c29eb853 | 527 | static void show_pwq(struct pool_workqueue *pwq); |
55df0933 | 528 | static void show_one_worker_pool(struct worker_pool *pool); |
7d19c5ce | 529 | |
97bd2347 TH |
530 | #define CREATE_TRACE_POINTS |
531 | #include <trace/events/workqueue.h> | |
532 | ||
68e13a67 | 533 | #define assert_rcu_or_pool_mutex() \ |
d355001f | 534 | RCU_LOCKDEP_WARN(!rcu_read_lock_any_held() && \ |
f78f5b90 | 535 | !lockdep_is_held(&wq_pool_mutex), \ |
24acfb71 | 536 | "RCU or wq_pool_mutex should be held") |
5bcab335 | 537 | |
5b95e1af | 538 | #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \ |
d355001f | 539 | RCU_LOCKDEP_WARN(!rcu_read_lock_any_held() && \ |
f78f5b90 PM |
540 | !lockdep_is_held(&wq->mutex) && \ |
541 | !lockdep_is_held(&wq_pool_mutex), \ | |
24acfb71 | 542 | "RCU, wq->mutex or wq_pool_mutex should be held") |
5b95e1af | 543 | |
4cb1ef64 TH |
544 | #define for_each_bh_worker_pool(pool, cpu) \ |
545 | for ((pool) = &per_cpu(bh_worker_pools, cpu)[0]; \ | |
546 | (pool) < &per_cpu(bh_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ | |
547 | (pool)++) | |
548 | ||
f02ae73a TH |
549 | #define for_each_cpu_worker_pool(pool, cpu) \ |
550 | for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ | |
551 | (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ | |
7a62c2c8 | 552 | (pool)++) |
4ce62e9e | 553 | |
17116969 TH |
554 | /** |
555 | * for_each_pool - iterate through all worker_pools in the system | |
556 | * @pool: iteration cursor | |
611c92a0 | 557 | * @pi: integer used for iteration |
fa1b54e6 | 558 | * |
24acfb71 | 559 | * This must be called either with wq_pool_mutex held or RCU read |
68e13a67 LJ |
560 | * locked. If the pool needs to be used beyond the locking in effect, the |
561 | * caller is responsible for guaranteeing that the pool stays online. | |
fa1b54e6 TH |
562 | * |
563 | * The if/else clause exists only for the lockdep assertion and can be | |
564 | * ignored. | |
17116969 | 565 | */ |
611c92a0 TH |
566 | #define for_each_pool(pool, pi) \ |
567 | idr_for_each_entry(&worker_pool_idr, pool, pi) \ | |
68e13a67 | 568 | if (({ assert_rcu_or_pool_mutex(); false; })) { } \ |
fa1b54e6 | 569 | else |
17116969 | 570 | |
822d8405 TH |
571 | /** |
572 | * for_each_pool_worker - iterate through all workers of a worker_pool | |
573 | * @worker: iteration cursor | |
822d8405 TH |
574 | * @pool: worker_pool to iterate workers of |
575 | * | |
1258fae7 | 576 | * This must be called with wq_pool_attach_mutex. |
822d8405 TH |
577 | * |
578 | * The if/else clause exists only for the lockdep assertion and can be | |
579 | * ignored. | |
580 | */ | |
da028469 LJ |
581 | #define for_each_pool_worker(worker, pool) \ |
582 | list_for_each_entry((worker), &(pool)->workers, node) \ | |
1258fae7 | 583 | if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \ |
822d8405 TH |
584 | else |
585 | ||
49e3cf44 TH |
586 | /** |
587 | * for_each_pwq - iterate through all pool_workqueues of the specified workqueue | |
588 | * @pwq: iteration cursor | |
589 | * @wq: the target workqueue | |
76af4d93 | 590 | * |
24acfb71 | 591 | * This must be called either with wq->mutex held or RCU read locked. |
794b18bc TH |
592 | * If the pwq needs to be used beyond the locking in effect, the caller is |
593 | * responsible for guaranteeing that the pwq stays online. | |
76af4d93 TH |
594 | * |
595 | * The if/else clause exists only for the lockdep assertion and can be | |
596 | * ignored. | |
49e3cf44 TH |
597 | */ |
598 | #define for_each_pwq(pwq, wq) \ | |
49e9d1a9 | 599 | list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \ |
5a644662 | 600 | lockdep_is_held(&(wq->mutex))) |
f3421797 | 601 | |
dc186ad7 TG |
602 | #ifdef CONFIG_DEBUG_OBJECTS_WORK |
603 | ||
f9e62f31 | 604 | static const struct debug_obj_descr work_debug_descr; |
dc186ad7 | 605 | |
99777288 SG |
606 | static void *work_debug_hint(void *addr) |
607 | { | |
608 | return ((struct work_struct *) addr)->func; | |
609 | } | |
610 | ||
b9fdac7f DC |
611 | static bool work_is_static_object(void *addr) |
612 | { | |
613 | struct work_struct *work = addr; | |
614 | ||
615 | return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work)); | |
616 | } | |
617 | ||
dc186ad7 TG |
618 | /* |
619 | * fixup_init is called when: | |
620 | * - an active object is initialized | |
621 | */ | |
02a982a6 | 622 | static bool work_fixup_init(void *addr, enum debug_obj_state state) |
dc186ad7 TG |
623 | { |
624 | struct work_struct *work = addr; | |
625 | ||
626 | switch (state) { | |
627 | case ODEBUG_STATE_ACTIVE: | |
628 | cancel_work_sync(work); | |
629 | debug_object_init(work, &work_debug_descr); | |
02a982a6 | 630 | return true; |
dc186ad7 | 631 | default: |
02a982a6 | 632 | return false; |
dc186ad7 TG |
633 | } |
634 | } | |
635 | ||
dc186ad7 TG |
636 | /* |
637 | * fixup_free is called when: | |
638 | * - an active object is freed | |
639 | */ | |
02a982a6 | 640 | static bool work_fixup_free(void *addr, enum debug_obj_state state) |
dc186ad7 TG |
641 | { |
642 | struct work_struct *work = addr; | |
643 | ||
644 | switch (state) { | |
645 | case ODEBUG_STATE_ACTIVE: | |
646 | cancel_work_sync(work); | |
647 | debug_object_free(work, &work_debug_descr); | |
02a982a6 | 648 | return true; |
dc186ad7 | 649 | default: |
02a982a6 | 650 | return false; |
dc186ad7 TG |
651 | } |
652 | } | |
653 | ||
f9e62f31 | 654 | static const struct debug_obj_descr work_debug_descr = { |
dc186ad7 | 655 | .name = "work_struct", |
99777288 | 656 | .debug_hint = work_debug_hint, |
b9fdac7f | 657 | .is_static_object = work_is_static_object, |
dc186ad7 | 658 | .fixup_init = work_fixup_init, |
dc186ad7 TG |
659 | .fixup_free = work_fixup_free, |
660 | }; | |
661 | ||
662 | static inline void debug_work_activate(struct work_struct *work) | |
663 | { | |
664 | debug_object_activate(work, &work_debug_descr); | |
665 | } | |
666 | ||
667 | static inline void debug_work_deactivate(struct work_struct *work) | |
668 | { | |
669 | debug_object_deactivate(work, &work_debug_descr); | |
670 | } | |
671 | ||
672 | void __init_work(struct work_struct *work, int onstack) | |
673 | { | |
674 | if (onstack) | |
675 | debug_object_init_on_stack(work, &work_debug_descr); | |
676 | else | |
677 | debug_object_init(work, &work_debug_descr); | |
678 | } | |
679 | EXPORT_SYMBOL_GPL(__init_work); | |
680 | ||
681 | void destroy_work_on_stack(struct work_struct *work) | |
682 | { | |
683 | debug_object_free(work, &work_debug_descr); | |
684 | } | |
685 | EXPORT_SYMBOL_GPL(destroy_work_on_stack); | |
686 | ||
ea2e64f2 TG |
687 | void destroy_delayed_work_on_stack(struct delayed_work *work) |
688 | { | |
689 | destroy_timer_on_stack(&work->timer); | |
690 | debug_object_free(&work->work, &work_debug_descr); | |
691 | } | |
692 | EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack); | |
693 | ||
dc186ad7 TG |
694 | #else |
695 | static inline void debug_work_activate(struct work_struct *work) { } | |
696 | static inline void debug_work_deactivate(struct work_struct *work) { } | |
697 | #endif | |
698 | ||
4e8b22bd | 699 | /** |
67dc8325 | 700 | * worker_pool_assign_id - allocate ID and assign it to @pool |
4e8b22bd LB |
701 | * @pool: the pool pointer of interest |
702 | * | |
703 | * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned | |
704 | * successfully, -errno on failure. | |
705 | */ | |
9daf9e67 TH |
706 | static int worker_pool_assign_id(struct worker_pool *pool) |
707 | { | |
708 | int ret; | |
709 | ||
68e13a67 | 710 | lockdep_assert_held(&wq_pool_mutex); |
5bcab335 | 711 | |
4e8b22bd LB |
712 | ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE, |
713 | GFP_KERNEL); | |
229641a6 | 714 | if (ret >= 0) { |
e68035fb | 715 | pool->id = ret; |
229641a6 TH |
716 | return 0; |
717 | } | |
fa1b54e6 | 718 | return ret; |
7c3eed5c TH |
719 | } |
720 | ||
9f66cff2 TH |
721 | static struct pool_workqueue __rcu ** |
722 | unbound_pwq_slot(struct workqueue_struct *wq, int cpu) | |
723 | { | |
724 | if (cpu >= 0) | |
725 | return per_cpu_ptr(wq->cpu_pwq, cpu); | |
726 | else | |
727 | return &wq->dfl_pwq; | |
728 | } | |
729 | ||
730 | /* @cpu < 0 for dfl_pwq */ | |
731 | static struct pool_workqueue *unbound_pwq(struct workqueue_struct *wq, int cpu) | |
732 | { | |
733 | return rcu_dereference_check(*unbound_pwq_slot(wq, cpu), | |
734 | lockdep_is_held(&wq_pool_mutex) || | |
735 | lockdep_is_held(&wq->mutex)); | |
736 | } | |
737 | ||
5797b1c1 TH |
738 | /** |
739 | * unbound_effective_cpumask - effective cpumask of an unbound workqueue | |
740 | * @wq: workqueue of interest | |
741 | * | |
742 | * @wq->unbound_attrs->cpumask contains the cpumask requested by the user which | |
743 | * is masked with wq_unbound_cpumask to determine the effective cpumask. The | |
744 | * default pwq is always mapped to the pool with the current effective cpumask. | |
745 | */ | |
746 | static struct cpumask *unbound_effective_cpumask(struct workqueue_struct *wq) | |
747 | { | |
748 | return unbound_pwq(wq, -1)->pool->attrs->__pod_cpumask; | |
749 | } | |
750 | ||
73f53c4a TH |
751 | static unsigned int work_color_to_flags(int color) |
752 | { | |
753 | return color << WORK_STRUCT_COLOR_SHIFT; | |
754 | } | |
755 | ||
c4560c2c | 756 | static int get_work_color(unsigned long work_data) |
73f53c4a | 757 | { |
c4560c2c | 758 | return (work_data >> WORK_STRUCT_COLOR_SHIFT) & |
73f53c4a TH |
759 | ((1 << WORK_STRUCT_COLOR_BITS) - 1); |
760 | } | |
761 | ||
762 | static int work_next_color(int color) | |
763 | { | |
764 | return (color + 1) % WORK_NR_COLORS; | |
765 | } | |
1da177e4 | 766 | |
456a78ee TH |
767 | static unsigned long pool_offq_flags(struct worker_pool *pool) |
768 | { | |
769 | return (pool->flags & POOL_BH) ? WORK_OFFQ_BH : 0; | |
770 | } | |
771 | ||
14441960 | 772 | /* |
112202d9 TH |
773 | * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data |
774 | * contain the pointer to the queued pwq. Once execution starts, the flag | |
7c3eed5c | 775 | * is cleared and the high bits contain OFFQ flags and pool ID. |
7a22ad75 | 776 | * |
afe928c1 TH |
777 | * set_work_pwq(), set_work_pool_and_clear_pending() and mark_work_canceling() |
778 | * can be used to set the pwq, pool or clear work->data. These functions should | |
779 | * only be called while the work is owned - ie. while the PENDING bit is set. | |
7a22ad75 | 780 | * |
112202d9 | 781 | * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq |
7c3eed5c | 782 | * corresponding to a work. Pool is available once the work has been |
112202d9 | 783 | * queued anywhere after initialization until it is sync canceled. pwq is |
7c3eed5c | 784 | * available only while the work item is queued. |
14441960 | 785 | */ |
bccdc1fa | 786 | static inline void set_work_data(struct work_struct *work, unsigned long data) |
365970a1 | 787 | { |
6183c009 | 788 | WARN_ON_ONCE(!work_pending(work)); |
bccdc1fa | 789 | atomic_long_set(&work->data, data | work_static(work)); |
7a22ad75 | 790 | } |
365970a1 | 791 | |
112202d9 | 792 | static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, |
bccdc1fa | 793 | unsigned long flags) |
7a22ad75 | 794 | { |
bccdc1fa TH |
795 | set_work_data(work, (unsigned long)pwq | WORK_STRUCT_PENDING | |
796 | WORK_STRUCT_PWQ | flags); | |
365970a1 DH |
797 | } |
798 | ||
4468a00f | 799 | static void set_work_pool_and_keep_pending(struct work_struct *work, |
bccdc1fa | 800 | int pool_id, unsigned long flags) |
4468a00f | 801 | { |
bccdc1fa TH |
802 | set_work_data(work, ((unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT) | |
803 | WORK_STRUCT_PENDING | flags); | |
4468a00f LJ |
804 | } |
805 | ||
7c3eed5c | 806 | static void set_work_pool_and_clear_pending(struct work_struct *work, |
bccdc1fa | 807 | int pool_id, unsigned long flags) |
7a22ad75 | 808 | { |
23657bb1 TH |
809 | /* |
810 | * The following wmb is paired with the implied mb in | |
811 | * test_and_set_bit(PENDING) and ensures all updates to @work made | |
812 | * here are visible to and precede any updates by the next PENDING | |
813 | * owner. | |
814 | */ | |
815 | smp_wmb(); | |
bccdc1fa TH |
816 | set_work_data(work, ((unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT) | |
817 | flags); | |
346c09f8 RP |
818 | /* |
819 | * The following mb guarantees that previous clear of a PENDING bit | |
820 | * will not be reordered with any speculative LOADS or STORES from | |
821 | * work->current_func, which is executed afterwards. This possible | |
8bdc6201 | 822 | * reordering can lead to a missed execution on attempt to queue |
346c09f8 RP |
823 | * the same @work. E.g. consider this case: |
824 | * | |
825 | * CPU#0 CPU#1 | |
826 | * ---------------------------- -------------------------------- | |
827 | * | |
828 | * 1 STORE event_indicated | |
829 | * 2 queue_work_on() { | |
830 | * 3 test_and_set_bit(PENDING) | |
831 | * 4 } set_..._and_clear_pending() { | |
832 | * 5 set_work_data() # clear bit | |
833 | * 6 smp_mb() | |
834 | * 7 work->current_func() { | |
835 | * 8 LOAD event_indicated | |
836 | * } | |
837 | * | |
838 | * Without an explicit full barrier speculative LOAD on line 8 can | |
839 | * be executed before CPU#0 does STORE on line 1. If that happens, | |
840 | * CPU#0 observes the PENDING bit is still set and new execution of | |
841 | * a @work is not queued in a hope, that CPU#1 will eventually | |
842 | * finish the queued @work. Meanwhile CPU#1 does not see | |
843 | * event_indicated is set, because speculative LOAD was executed | |
844 | * before actual STORE. | |
845 | */ | |
846 | smp_mb(); | |
7a22ad75 | 847 | } |
f756d5e2 | 848 | |
afa4bb77 LT |
849 | static inline struct pool_workqueue *work_struct_pwq(unsigned long data) |
850 | { | |
e9a8e01f | 851 | return (struct pool_workqueue *)(data & WORK_STRUCT_PWQ_MASK); |
afa4bb77 LT |
852 | } |
853 | ||
112202d9 | 854 | static struct pool_workqueue *get_work_pwq(struct work_struct *work) |
b1f4ec17 | 855 | { |
e120153d | 856 | unsigned long data = atomic_long_read(&work->data); |
7a22ad75 | 857 | |
112202d9 | 858 | if (data & WORK_STRUCT_PWQ) |
afa4bb77 | 859 | return work_struct_pwq(data); |
e120153d TH |
860 | else |
861 | return NULL; | |
4d707b9f ON |
862 | } |
863 | ||
7c3eed5c TH |
864 | /** |
865 | * get_work_pool - return the worker_pool a given work was associated with | |
866 | * @work: the work item of interest | |
867 | * | |
68e13a67 | 868 | * Pools are created and destroyed under wq_pool_mutex, and allows read |
24acfb71 TG |
869 | * access under RCU read lock. As such, this function should be |
870 | * called under wq_pool_mutex or inside of a rcu_read_lock() region. | |
fa1b54e6 TH |
871 | * |
872 | * All fields of the returned pool are accessible as long as the above | |
873 | * mentioned locking is in effect. If the returned pool needs to be used | |
874 | * beyond the critical section, the caller is responsible for ensuring the | |
875 | * returned pool is and stays online. | |
d185af30 YB |
876 | * |
877 | * Return: The worker_pool @work was last associated with. %NULL if none. | |
7c3eed5c TH |
878 | */ |
879 | static struct worker_pool *get_work_pool(struct work_struct *work) | |
365970a1 | 880 | { |
e120153d | 881 | unsigned long data = atomic_long_read(&work->data); |
7c3eed5c | 882 | int pool_id; |
7a22ad75 | 883 | |
68e13a67 | 884 | assert_rcu_or_pool_mutex(); |
fa1b54e6 | 885 | |
112202d9 | 886 | if (data & WORK_STRUCT_PWQ) |
afa4bb77 | 887 | return work_struct_pwq(data)->pool; |
7a22ad75 | 888 | |
7c3eed5c TH |
889 | pool_id = data >> WORK_OFFQ_POOL_SHIFT; |
890 | if (pool_id == WORK_OFFQ_POOL_NONE) | |
7a22ad75 TH |
891 | return NULL; |
892 | ||
fa1b54e6 | 893 | return idr_find(&worker_pool_idr, pool_id); |
7c3eed5c TH |
894 | } |
895 | ||
1211f3b2 | 896 | static unsigned long shift_and_mask(unsigned long v, u32 shift, u32 bits) |
7c3eed5c | 897 | { |
38f7e145 | 898 | return (v >> shift) & ((1U << bits) - 1); |
7c3eed5c TH |
899 | } |
900 | ||
1211f3b2 | 901 | static void work_offqd_unpack(struct work_offq_data *offqd, unsigned long data) |
bbb68dfa | 902 | { |
1211f3b2 | 903 | WARN_ON_ONCE(data & WORK_STRUCT_PWQ); |
bbb68dfa | 904 | |
1211f3b2 TH |
905 | offqd->pool_id = shift_and_mask(data, WORK_OFFQ_POOL_SHIFT, |
906 | WORK_OFFQ_POOL_BITS); | |
86898fa6 TH |
907 | offqd->disable = shift_and_mask(data, WORK_OFFQ_DISABLE_SHIFT, |
908 | WORK_OFFQ_DISABLE_BITS); | |
1211f3b2 | 909 | offqd->flags = data & WORK_OFFQ_FLAG_MASK; |
bbb68dfa TH |
910 | } |
911 | ||
1211f3b2 | 912 | static unsigned long work_offqd_pack_flags(struct work_offq_data *offqd) |
bbb68dfa | 913 | { |
86898fa6 TH |
914 | return ((unsigned long)offqd->disable << WORK_OFFQ_DISABLE_SHIFT) | |
915 | ((unsigned long)offqd->flags); | |
bbb68dfa TH |
916 | } |
917 | ||
e22bee78 | 918 | /* |
3270476a TH |
919 | * Policy functions. These define the policies on how the global worker |
920 | * pools are managed. Unless noted otherwise, these functions assume that | |
d565ed63 | 921 | * they're being called with pool->lock held. |
e22bee78 TH |
922 | */ |
923 | ||
4594bf15 | 924 | /* |
e22bee78 TH |
925 | * Need to wake up a worker? Called from anything but currently |
926 | * running workers. | |
974271c4 TH |
927 | * |
928 | * Note that, because unbound workers never contribute to nr_running, this | |
706026c2 | 929 | * function will always return %true for unbound pools as long as the |
974271c4 | 930 | * worklist isn't empty. |
4594bf15 | 931 | */ |
63d95a91 | 932 | static bool need_more_worker(struct worker_pool *pool) |
365970a1 | 933 | { |
0219a352 | 934 | return !list_empty(&pool->worklist) && !pool->nr_running; |
e22bee78 | 935 | } |
4594bf15 | 936 | |
e22bee78 | 937 | /* Can I start working? Called from busy but !running workers. */ |
63d95a91 | 938 | static bool may_start_working(struct worker_pool *pool) |
e22bee78 | 939 | { |
63d95a91 | 940 | return pool->nr_idle; |
e22bee78 TH |
941 | } |
942 | ||
943 | /* Do I need to keep working? Called from currently running workers. */ | |
63d95a91 | 944 | static bool keep_working(struct worker_pool *pool) |
e22bee78 | 945 | { |
bc35f7ef | 946 | return !list_empty(&pool->worklist) && (pool->nr_running <= 1); |
e22bee78 TH |
947 | } |
948 | ||
949 | /* Do we need a new worker? Called from manager. */ | |
63d95a91 | 950 | static bool need_to_create_worker(struct worker_pool *pool) |
e22bee78 | 951 | { |
63d95a91 | 952 | return need_more_worker(pool) && !may_start_working(pool); |
e22bee78 | 953 | } |
365970a1 | 954 | |
e22bee78 | 955 | /* Do we have too many workers and should some go away? */ |
63d95a91 | 956 | static bool too_many_workers(struct worker_pool *pool) |
e22bee78 | 957 | { |
692b4825 | 958 | bool managing = pool->flags & POOL_MANAGER_ACTIVE; |
63d95a91 TH |
959 | int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ |
960 | int nr_busy = pool->nr_workers - nr_idle; | |
e22bee78 TH |
961 | |
962 | return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; | |
365970a1 DH |
963 | } |
964 | ||
c54d5046 TH |
965 | /** |
966 | * worker_set_flags - set worker flags and adjust nr_running accordingly | |
967 | * @worker: self | |
968 | * @flags: flags to set | |
969 | * | |
970 | * Set @flags in @worker->flags and adjust nr_running accordingly. | |
c54d5046 TH |
971 | */ |
972 | static inline void worker_set_flags(struct worker *worker, unsigned int flags) | |
973 | { | |
974 | struct worker_pool *pool = worker->pool; | |
975 | ||
bc8b50c2 | 976 | lockdep_assert_held(&pool->lock); |
c54d5046 TH |
977 | |
978 | /* If transitioning into NOT_RUNNING, adjust nr_running. */ | |
979 | if ((flags & WORKER_NOT_RUNNING) && | |
980 | !(worker->flags & WORKER_NOT_RUNNING)) { | |
981 | pool->nr_running--; | |
982 | } | |
983 | ||
984 | worker->flags |= flags; | |
985 | } | |
986 | ||
987 | /** | |
988 | * worker_clr_flags - clear worker flags and adjust nr_running accordingly | |
989 | * @worker: self | |
990 | * @flags: flags to clear | |
991 | * | |
992 | * Clear @flags in @worker->flags and adjust nr_running accordingly. | |
c54d5046 TH |
993 | */ |
994 | static inline void worker_clr_flags(struct worker *worker, unsigned int flags) | |
995 | { | |
996 | struct worker_pool *pool = worker->pool; | |
997 | unsigned int oflags = worker->flags; | |
998 | ||
bc8b50c2 | 999 | lockdep_assert_held(&pool->lock); |
c54d5046 TH |
1000 | |
1001 | worker->flags &= ~flags; | |
1002 | ||
1003 | /* | |
1004 | * If transitioning out of NOT_RUNNING, increment nr_running. Note | |
1005 | * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask | |
1006 | * of multiple flags, not a single flag. | |
1007 | */ | |
1008 | if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) | |
1009 | if (!(worker->flags & WORKER_NOT_RUNNING)) | |
1010 | pool->nr_running++; | |
1011 | } | |
1012 | ||
797e8345 TH |
1013 | /* Return the first idle worker. Called with pool->lock held. */ |
1014 | static struct worker *first_idle_worker(struct worker_pool *pool) | |
1015 | { | |
1016 | if (unlikely(list_empty(&pool->idle_list))) | |
1017 | return NULL; | |
1018 | ||
1019 | return list_first_entry(&pool->idle_list, struct worker, entry); | |
1020 | } | |
1021 | ||
1022 | /** | |
1023 | * worker_enter_idle - enter idle state | |
1024 | * @worker: worker which is entering idle state | |
1025 | * | |
1026 | * @worker is entering idle state. Update stats and idle timer if | |
1027 | * necessary. | |
1028 | * | |
1029 | * LOCKING: | |
1030 | * raw_spin_lock_irq(pool->lock). | |
1031 | */ | |
1032 | static void worker_enter_idle(struct worker *worker) | |
1033 | { | |
1034 | struct worker_pool *pool = worker->pool; | |
1035 | ||
1036 | if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || | |
1037 | WARN_ON_ONCE(!list_empty(&worker->entry) && | |
1038 | (worker->hentry.next || worker->hentry.pprev))) | |
1039 | return; | |
1040 | ||
1041 | /* can't use worker_set_flags(), also called from create_worker() */ | |
1042 | worker->flags |= WORKER_IDLE; | |
1043 | pool->nr_idle++; | |
1044 | worker->last_active = jiffies; | |
1045 | ||
1046 | /* idle_list is LIFO */ | |
1047 | list_add(&worker->entry, &pool->idle_list); | |
1048 | ||
1049 | if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) | |
1050 | mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); | |
1051 | ||
1052 | /* Sanity check nr_running. */ | |
1053 | WARN_ON_ONCE(pool->nr_workers == pool->nr_idle && pool->nr_running); | |
1054 | } | |
1055 | ||
1056 | /** | |
1057 | * worker_leave_idle - leave idle state | |
1058 | * @worker: worker which is leaving idle state | |
1059 | * | |
1060 | * @worker is leaving idle state. Update stats. | |
1061 | * | |
1062 | * LOCKING: | |
1063 | * raw_spin_lock_irq(pool->lock). | |
1064 | */ | |
1065 | static void worker_leave_idle(struct worker *worker) | |
1066 | { | |
1067 | struct worker_pool *pool = worker->pool; | |
1068 | ||
1069 | if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) | |
1070 | return; | |
1071 | worker_clr_flags(worker, WORKER_IDLE); | |
1072 | pool->nr_idle--; | |
1073 | list_del_init(&worker->entry); | |
1074 | } | |
1075 | ||
1076 | /** | |
1077 | * find_worker_executing_work - find worker which is executing a work | |
1078 | * @pool: pool of interest | |
1079 | * @work: work to find worker for | |
1080 | * | |
1081 | * Find a worker which is executing @work on @pool by searching | |
1082 | * @pool->busy_hash which is keyed by the address of @work. For a worker | |
1083 | * to match, its current execution should match the address of @work and | |
1084 | * its work function. This is to avoid unwanted dependency between | |
1085 | * unrelated work executions through a work item being recycled while still | |
1086 | * being executed. | |
1087 | * | |
1088 | * This is a bit tricky. A work item may be freed once its execution | |
1089 | * starts and nothing prevents the freed area from being recycled for | |
1090 | * another work item. If the same work item address ends up being reused | |
1091 | * before the original execution finishes, workqueue will identify the | |
1092 | * recycled work item as currently executing and make it wait until the | |
1093 | * current execution finishes, introducing an unwanted dependency. | |
1094 | * | |
1095 | * This function checks the work item address and work function to avoid | |
1096 | * false positives. Note that this isn't complete as one may construct a | |
1097 | * work function which can introduce dependency onto itself through a | |
1098 | * recycled work item. Well, if somebody wants to shoot oneself in the | |
1099 | * foot that badly, there's only so much we can do, and if such deadlock | |
1100 | * actually occurs, it should be easy to locate the culprit work function. | |
1101 | * | |
1102 | * CONTEXT: | |
1103 | * raw_spin_lock_irq(pool->lock). | |
1104 | * | |
1105 | * Return: | |
1106 | * Pointer to worker which is executing @work if found, %NULL | |
1107 | * otherwise. | |
1108 | */ | |
1109 | static struct worker *find_worker_executing_work(struct worker_pool *pool, | |
1110 | struct work_struct *work) | |
1111 | { | |
1112 | struct worker *worker; | |
1113 | ||
1114 | hash_for_each_possible(pool->busy_hash, worker, hentry, | |
1115 | (unsigned long)work) | |
1116 | if (worker->current_work == work && | |
1117 | worker->current_func == work->func) | |
1118 | return worker; | |
1119 | ||
1120 | return NULL; | |
1121 | } | |
1122 | ||
1123 | /** | |
1124 | * move_linked_works - move linked works to a list | |
1125 | * @work: start of series of works to be scheduled | |
1126 | * @head: target list to append @work to | |
1127 | * @nextp: out parameter for nested worklist walking | |
1128 | * | |
873eaca6 TH |
1129 | * Schedule linked works starting from @work to @head. Work series to be |
1130 | * scheduled starts at @work and includes any consecutive work with | |
1131 | * WORK_STRUCT_LINKED set in its predecessor. See assign_work() for details on | |
1132 | * @nextp. | |
797e8345 TH |
1133 | * |
1134 | * CONTEXT: | |
1135 | * raw_spin_lock_irq(pool->lock). | |
1136 | */ | |
1137 | static void move_linked_works(struct work_struct *work, struct list_head *head, | |
1138 | struct work_struct **nextp) | |
1139 | { | |
1140 | struct work_struct *n; | |
1141 | ||
1142 | /* | |
1143 | * Linked worklist will always end before the end of the list, | |
1144 | * use NULL for list head. | |
1145 | */ | |
1146 | list_for_each_entry_safe_from(work, n, NULL, entry) { | |
1147 | list_move_tail(&work->entry, head); | |
1148 | if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) | |
1149 | break; | |
1150 | } | |
1151 | ||
1152 | /* | |
1153 | * If we're already inside safe list traversal and have moved | |
1154 | * multiple works to the scheduled queue, the next position | |
1155 | * needs to be updated. | |
1156 | */ | |
1157 | if (nextp) | |
1158 | *nextp = n; | |
1159 | } | |
1160 | ||
873eaca6 TH |
1161 | /** |
1162 | * assign_work - assign a work item and its linked work items to a worker | |
1163 | * @work: work to assign | |
1164 | * @worker: worker to assign to | |
1165 | * @nextp: out parameter for nested worklist walking | |
1166 | * | |
1167 | * Assign @work and its linked work items to @worker. If @work is already being | |
1168 | * executed by another worker in the same pool, it'll be punted there. | |
1169 | * | |
1170 | * If @nextp is not NULL, it's updated to point to the next work of the last | |
1171 | * scheduled work. This allows assign_work() to be nested inside | |
1172 | * list_for_each_entry_safe(). | |
1173 | * | |
1174 | * Returns %true if @work was successfully assigned to @worker. %false if @work | |
1175 | * was punted to another worker already executing it. | |
1176 | */ | |
1177 | static bool assign_work(struct work_struct *work, struct worker *worker, | |
1178 | struct work_struct **nextp) | |
1179 | { | |
1180 | struct worker_pool *pool = worker->pool; | |
1181 | struct worker *collision; | |
1182 | ||
1183 | lockdep_assert_held(&pool->lock); | |
1184 | ||
1185 | /* | |
1186 | * A single work shouldn't be executed concurrently by multiple workers. | |
1187 | * __queue_work() ensures that @work doesn't jump to a different pool | |
1188 | * while still running in the previous pool. Here, we should ensure that | |
1189 | * @work is not executed concurrently by multiple workers from the same | |
1190 | * pool. Check whether anyone is already processing the work. If so, | |
1191 | * defer the work to the currently executing one. | |
1192 | */ | |
1193 | collision = find_worker_executing_work(pool, work); | |
1194 | if (unlikely(collision)) { | |
1195 | move_linked_works(work, &collision->scheduled, nextp); | |
1196 | return false; | |
1197 | } | |
1198 | ||
1199 | move_linked_works(work, &worker->scheduled, nextp); | |
1200 | return true; | |
1201 | } | |
1202 | ||
2f34d733 TH |
1203 | static struct irq_work *bh_pool_irq_work(struct worker_pool *pool) |
1204 | { | |
1205 | int high = pool->attrs->nice == HIGHPRI_NICE_LEVEL ? 1 : 0; | |
1206 | ||
1207 | return &per_cpu(bh_pool_irq_works, pool->cpu)[high]; | |
1208 | } | |
1209 | ||
fd0a68a2 TH |
1210 | static void kick_bh_pool(struct worker_pool *pool) |
1211 | { | |
1212 | #ifdef CONFIG_SMP | |
1acd92d9 TH |
1213 | /* see drain_dead_softirq_workfn() for BH_DRAINING */ |
1214 | if (unlikely(pool->cpu != smp_processor_id() && | |
1215 | !(pool->flags & POOL_BH_DRAINING))) { | |
fd0a68a2 TH |
1216 | irq_work_queue_on(bh_pool_irq_work(pool), pool->cpu); |
1217 | return; | |
1218 | } | |
1219 | #endif | |
1220 | if (pool->attrs->nice == HIGHPRI_NICE_LEVEL) | |
1221 | raise_softirq_irqoff(HI_SOFTIRQ); | |
1222 | else | |
1223 | raise_softirq_irqoff(TASKLET_SOFTIRQ); | |
1224 | } | |
1225 | ||
797e8345 | 1226 | /** |
0219a352 TH |
1227 | * kick_pool - wake up an idle worker if necessary |
1228 | * @pool: pool to kick | |
797e8345 | 1229 | * |
0219a352 TH |
1230 | * @pool may have pending work items. Wake up worker if necessary. Returns |
1231 | * whether a worker was woken up. | |
797e8345 | 1232 | */ |
0219a352 | 1233 | static bool kick_pool(struct worker_pool *pool) |
797e8345 TH |
1234 | { |
1235 | struct worker *worker = first_idle_worker(pool); | |
8639eceb | 1236 | struct task_struct *p; |
797e8345 | 1237 | |
0219a352 TH |
1238 | lockdep_assert_held(&pool->lock); |
1239 | ||
1240 | if (!need_more_worker(pool) || !worker) | |
1241 | return false; | |
1242 | ||
4cb1ef64 | 1243 | if (pool->flags & POOL_BH) { |
fd0a68a2 | 1244 | kick_bh_pool(pool); |
4cb1ef64 TH |
1245 | return true; |
1246 | } | |
1247 | ||
8639eceb TH |
1248 | p = worker->task; |
1249 | ||
1250 | #ifdef CONFIG_SMP | |
1251 | /* | |
1252 | * Idle @worker is about to execute @work and waking up provides an | |
1253 | * opportunity to migrate @worker at a lower cost by setting the task's | |
1254 | * wake_cpu field. Let's see if we want to move @worker to improve | |
1255 | * execution locality. | |
1256 | * | |
1257 | * We're waking the worker that went idle the latest and there's some | |
1258 | * chance that @worker is marked idle but hasn't gone off CPU yet. If | |
1259 | * so, setting the wake_cpu won't do anything. As this is a best-effort | |
1260 | * optimization and the race window is narrow, let's leave as-is for | |
1261 | * now. If this becomes pronounced, we can skip over workers which are | |
1262 | * still on cpu when picking an idle worker. | |
1263 | * | |
1264 | * If @pool has non-strict affinity, @worker might have ended up outside | |
1265 | * its affinity scope. Repatriate. | |
1266 | */ | |
1267 | if (!pool->attrs->affn_strict && | |
1268 | !cpumask_test_cpu(p->wake_cpu, pool->attrs->__pod_cpumask)) { | |
1269 | struct work_struct *work = list_first_entry(&pool->worklist, | |
1270 | struct work_struct, entry); | |
57a01eaf SS |
1271 | int wake_cpu = cpumask_any_and_distribute(pool->attrs->__pod_cpumask, |
1272 | cpu_online_mask); | |
1273 | if (wake_cpu < nr_cpu_ids) { | |
1274 | p->wake_cpu = wake_cpu; | |
1275 | get_work_pwq(work)->stats[PWQ_STAT_REPATRIATED]++; | |
1276 | } | |
8639eceb TH |
1277 | } |
1278 | #endif | |
1279 | wake_up_process(p); | |
0219a352 | 1280 | return true; |
797e8345 TH |
1281 | } |
1282 | ||
63638450 TH |
1283 | #ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT |
1284 | ||
1285 | /* | |
1286 | * Concurrency-managed per-cpu work items that hog CPU for longer than | |
1287 | * wq_cpu_intensive_thresh_us trigger the automatic CPU_INTENSIVE mechanism, | |
1288 | * which prevents them from stalling other concurrency-managed work items. If a | |
1289 | * work function keeps triggering this mechanism, it's likely that the work item | |
1290 | * should be using an unbound workqueue instead. | |
1291 | * | |
1292 | * wq_cpu_intensive_report() tracks work functions which trigger such conditions | |
1293 | * and report them so that they can be examined and converted to use unbound | |
1294 | * workqueues as appropriate. To avoid flooding the console, each violating work | |
1295 | * function is tracked and reported with exponential backoff. | |
1296 | */ | |
1297 | #define WCI_MAX_ENTS 128 | |
1298 | ||
1299 | struct wci_ent { | |
1300 | work_func_t func; | |
1301 | atomic64_t cnt; | |
1302 | struct hlist_node hash_node; | |
1303 | }; | |
1304 | ||
1305 | static struct wci_ent wci_ents[WCI_MAX_ENTS]; | |
1306 | static int wci_nr_ents; | |
1307 | static DEFINE_RAW_SPINLOCK(wci_lock); | |
1308 | static DEFINE_HASHTABLE(wci_hash, ilog2(WCI_MAX_ENTS)); | |
1309 | ||
1310 | static struct wci_ent *wci_find_ent(work_func_t func) | |
1311 | { | |
1312 | struct wci_ent *ent; | |
1313 | ||
1314 | hash_for_each_possible_rcu(wci_hash, ent, hash_node, | |
1315 | (unsigned long)func) { | |
1316 | if (ent->func == func) | |
1317 | return ent; | |
1318 | } | |
1319 | return NULL; | |
1320 | } | |
1321 | ||
1322 | static void wq_cpu_intensive_report(work_func_t func) | |
1323 | { | |
1324 | struct wci_ent *ent; | |
1325 | ||
1326 | restart: | |
1327 | ent = wci_find_ent(func); | |
1328 | if (ent) { | |
1329 | u64 cnt; | |
1330 | ||
1331 | /* | |
ccdec921 | 1332 | * Start reporting from the warning_thresh and back off |
63638450 TH |
1333 | * exponentially. |
1334 | */ | |
1335 | cnt = atomic64_inc_return_relaxed(&ent->cnt); | |
ccdec921 XY |
1336 | if (wq_cpu_intensive_warning_thresh && |
1337 | cnt >= wq_cpu_intensive_warning_thresh && | |
1338 | is_power_of_2(cnt + 1 - wq_cpu_intensive_warning_thresh)) | |
63638450 TH |
1339 | printk_deferred(KERN_WARNING "workqueue: %ps hogged CPU for >%luus %llu times, consider switching to WQ_UNBOUND\n", |
1340 | ent->func, wq_cpu_intensive_thresh_us, | |
1341 | atomic64_read(&ent->cnt)); | |
1342 | return; | |
1343 | } | |
1344 | ||
1345 | /* | |
1346 | * @func is a new violation. Allocate a new entry for it. If wcn_ents[] | |
1347 | * is exhausted, something went really wrong and we probably made enough | |
1348 | * noise already. | |
1349 | */ | |
1350 | if (wci_nr_ents >= WCI_MAX_ENTS) | |
1351 | return; | |
1352 | ||
1353 | raw_spin_lock(&wci_lock); | |
1354 | ||
1355 | if (wci_nr_ents >= WCI_MAX_ENTS) { | |
1356 | raw_spin_unlock(&wci_lock); | |
1357 | return; | |
1358 | } | |
1359 | ||
1360 | if (wci_find_ent(func)) { | |
1361 | raw_spin_unlock(&wci_lock); | |
1362 | goto restart; | |
1363 | } | |
1364 | ||
1365 | ent = &wci_ents[wci_nr_ents++]; | |
1366 | ent->func = func; | |
ccdec921 | 1367 | atomic64_set(&ent->cnt, 0); |
63638450 TH |
1368 | hash_add_rcu(wci_hash, &ent->hash_node, (unsigned long)func); |
1369 | ||
1370 | raw_spin_unlock(&wci_lock); | |
ccdec921 XY |
1371 | |
1372 | goto restart; | |
63638450 TH |
1373 | } |
1374 | ||
1375 | #else /* CONFIG_WQ_CPU_INTENSIVE_REPORT */ | |
1376 | static void wq_cpu_intensive_report(work_func_t func) {} | |
1377 | #endif /* CONFIG_WQ_CPU_INTENSIVE_REPORT */ | |
1378 | ||
d302f017 | 1379 | /** |
6d25be57 | 1380 | * wq_worker_running - a worker is running again |
e22bee78 | 1381 | * @task: task waking up |
e22bee78 | 1382 | * |
6d25be57 | 1383 | * This function is called when a worker returns from schedule() |
e22bee78 | 1384 | */ |
6d25be57 | 1385 | void wq_worker_running(struct task_struct *task) |
e22bee78 TH |
1386 | { |
1387 | struct worker *worker = kthread_data(task); | |
1388 | ||
c8f6219b | 1389 | if (!READ_ONCE(worker->sleeping)) |
6d25be57 | 1390 | return; |
07edfece FW |
1391 | |
1392 | /* | |
1393 | * If preempted by unbind_workers() between the WORKER_NOT_RUNNING check | |
1394 | * and the nr_running increment below, we may ruin the nr_running reset | |
1395 | * and leave with an unexpected pool->nr_running == 1 on the newly unbound | |
1396 | * pool. Protect against such race. | |
1397 | */ | |
1398 | preempt_disable(); | |
6d25be57 | 1399 | if (!(worker->flags & WORKER_NOT_RUNNING)) |
bc35f7ef | 1400 | worker->pool->nr_running++; |
07edfece | 1401 | preempt_enable(); |
616db877 TH |
1402 | |
1403 | /* | |
1404 | * CPU intensive auto-detection cares about how long a work item hogged | |
1405 | * CPU without sleeping. Reset the starting timestamp on wakeup. | |
1406 | */ | |
1407 | worker->current_at = worker->task->se.sum_exec_runtime; | |
1408 | ||
c8f6219b | 1409 | WRITE_ONCE(worker->sleeping, 0); |
e22bee78 TH |
1410 | } |
1411 | ||
1412 | /** | |
1413 | * wq_worker_sleeping - a worker is going to sleep | |
1414 | * @task: task going to sleep | |
e22bee78 | 1415 | * |
6d25be57 | 1416 | * This function is called from schedule() when a busy worker is |
ccf45156 | 1417 | * going to sleep. |
e22bee78 | 1418 | */ |
6d25be57 | 1419 | void wq_worker_sleeping(struct task_struct *task) |
e22bee78 | 1420 | { |
cc5bff38 | 1421 | struct worker *worker = kthread_data(task); |
111c225a | 1422 | struct worker_pool *pool; |
e22bee78 | 1423 | |
111c225a TH |
1424 | /* |
1425 | * Rescuers, which may not have all the fields set up like normal | |
1426 | * workers, also reach here, let's not access anything before | |
1427 | * checking NOT_RUNNING. | |
1428 | */ | |
2d64672e | 1429 | if (worker->flags & WORKER_NOT_RUNNING) |
6d25be57 | 1430 | return; |
e22bee78 | 1431 | |
111c225a | 1432 | pool = worker->pool; |
111c225a | 1433 | |
62849a96 | 1434 | /* Return if preempted before wq_worker_running() was reached */ |
c8f6219b | 1435 | if (READ_ONCE(worker->sleeping)) |
6d25be57 TG |
1436 | return; |
1437 | ||
c8f6219b | 1438 | WRITE_ONCE(worker->sleeping, 1); |
a9b8a985 | 1439 | raw_spin_lock_irq(&pool->lock); |
e22bee78 | 1440 | |
45c753f5 FW |
1441 | /* |
1442 | * Recheck in case unbind_workers() preempted us. We don't | |
1443 | * want to decrement nr_running after the worker is unbound | |
1444 | * and nr_running has been reset. | |
1445 | */ | |
1446 | if (worker->flags & WORKER_NOT_RUNNING) { | |
1447 | raw_spin_unlock_irq(&pool->lock); | |
1448 | return; | |
1449 | } | |
1450 | ||
bc35f7ef | 1451 | pool->nr_running--; |
0219a352 | 1452 | if (kick_pool(pool)) |
725e8ec5 | 1453 | worker->current_pwq->stats[PWQ_STAT_CM_WAKEUP]++; |
0219a352 | 1454 | |
a9b8a985 | 1455 | raw_spin_unlock_irq(&pool->lock); |
e22bee78 TH |
1456 | } |
1457 | ||
616db877 TH |
1458 | /** |
1459 | * wq_worker_tick - a scheduler tick occurred while a kworker is running | |
1460 | * @task: task currently running | |
1461 | * | |
86dd6c04 | 1462 | * Called from sched_tick(). We're in the IRQ context and the current |
616db877 TH |
1463 | * worker's fields which follow the 'K' locking rule can be accessed safely. |
1464 | */ | |
1465 | void wq_worker_tick(struct task_struct *task) | |
1466 | { | |
1467 | struct worker *worker = kthread_data(task); | |
1468 | struct pool_workqueue *pwq = worker->current_pwq; | |
1469 | struct worker_pool *pool = worker->pool; | |
1470 | ||
1471 | if (!pwq) | |
1472 | return; | |
1473 | ||
8a1dd1e5 TH |
1474 | pwq->stats[PWQ_STAT_CPU_TIME] += TICK_USEC; |
1475 | ||
18c8ae81 Z |
1476 | if (!wq_cpu_intensive_thresh_us) |
1477 | return; | |
1478 | ||
616db877 TH |
1479 | /* |
1480 | * If the current worker is concurrency managed and hogged the CPU for | |
1481 | * longer than wq_cpu_intensive_thresh_us, it's automatically marked | |
1482 | * CPU_INTENSIVE to avoid stalling other concurrency-managed work items. | |
c8f6219b Z |
1483 | * |
1484 | * Set @worker->sleeping means that @worker is in the process of | |
1485 | * switching out voluntarily and won't be contributing to | |
1486 | * @pool->nr_running until it wakes up. As wq_worker_sleeping() also | |
1487 | * decrements ->nr_running, setting CPU_INTENSIVE here can lead to | |
1488 | * double decrements. The task is releasing the CPU anyway. Let's skip. | |
1489 | * We probably want to make this prettier in the future. | |
616db877 | 1490 | */ |
c8f6219b | 1491 | if ((worker->flags & WORKER_NOT_RUNNING) || READ_ONCE(worker->sleeping) || |
616db877 TH |
1492 | worker->task->se.sum_exec_runtime - worker->current_at < |
1493 | wq_cpu_intensive_thresh_us * NSEC_PER_USEC) | |
1494 | return; | |
1495 | ||
1496 | raw_spin_lock(&pool->lock); | |
1497 | ||
1498 | worker_set_flags(worker, WORKER_CPU_INTENSIVE); | |
63638450 | 1499 | wq_cpu_intensive_report(worker->current_func); |
616db877 TH |
1500 | pwq->stats[PWQ_STAT_CPU_INTENSIVE]++; |
1501 | ||
0219a352 | 1502 | if (kick_pool(pool)) |
616db877 | 1503 | pwq->stats[PWQ_STAT_CM_WAKEUP]++; |
616db877 TH |
1504 | |
1505 | raw_spin_unlock(&pool->lock); | |
1506 | } | |
1507 | ||
1b69ac6b JW |
1508 | /** |
1509 | * wq_worker_last_func - retrieve worker's last work function | |
8194fe94 | 1510 | * @task: Task to retrieve last work function of. |
1b69ac6b JW |
1511 | * |
1512 | * Determine the last function a worker executed. This is called from | |
1513 | * the scheduler to get a worker's last known identity. | |
1514 | * | |
1515 | * CONTEXT: | |
a9b8a985 | 1516 | * raw_spin_lock_irq(rq->lock) |
1b69ac6b | 1517 | * |
4b047002 JW |
1518 | * This function is called during schedule() when a kworker is going |
1519 | * to sleep. It's used by psi to identify aggregation workers during | |
1520 | * dequeuing, to allow periodic aggregation to shut-off when that | |
1521 | * worker is the last task in the system or cgroup to go to sleep. | |
1522 | * | |
1523 | * As this function doesn't involve any workqueue-related locking, it | |
1524 | * only returns stable values when called from inside the scheduler's | |
1525 | * queuing and dequeuing paths, when @task, which must be a kworker, | |
1526 | * is guaranteed to not be processing any works. | |
1527 | * | |
1b69ac6b JW |
1528 | * Return: |
1529 | * The last work function %current executed as a worker, NULL if it | |
1530 | * hasn't executed any work yet. | |
1531 | */ | |
1532 | work_func_t wq_worker_last_func(struct task_struct *task) | |
1533 | { | |
1534 | struct worker *worker = kthread_data(task); | |
1535 | ||
1536 | return worker->last_func; | |
1537 | } | |
1538 | ||
91ccc6e7 TH |
1539 | /** |
1540 | * wq_node_nr_active - Determine wq_node_nr_active to use | |
1541 | * @wq: workqueue of interest | |
1542 | * @node: NUMA node, can be %NUMA_NO_NODE | |
1543 | * | |
1544 | * Determine wq_node_nr_active to use for @wq on @node. Returns: | |
1545 | * | |
1546 | * - %NULL for per-cpu workqueues as they don't need to use shared nr_active. | |
1547 | * | |
1548 | * - node_nr_active[nr_node_ids] if @node is %NUMA_NO_NODE. | |
1549 | * | |
1550 | * - Otherwise, node_nr_active[@node]. | |
1551 | */ | |
1552 | static struct wq_node_nr_active *wq_node_nr_active(struct workqueue_struct *wq, | |
1553 | int node) | |
1554 | { | |
1555 | if (!(wq->flags & WQ_UNBOUND)) | |
1556 | return NULL; | |
1557 | ||
1558 | if (node == NUMA_NO_NODE) | |
1559 | node = nr_node_ids; | |
1560 | ||
1561 | return wq->node_nr_active[node]; | |
1562 | } | |
1563 | ||
5797b1c1 TH |
1564 | /** |
1565 | * wq_update_node_max_active - Update per-node max_actives to use | |
1566 | * @wq: workqueue to update | |
1567 | * @off_cpu: CPU that's going down, -1 if a CPU is not going down | |
1568 | * | |
1569 | * Update @wq->node_nr_active[]->max. @wq must be unbound. max_active is | |
1570 | * distributed among nodes according to the proportions of numbers of online | |
1571 | * cpus. The result is always between @wq->min_active and max_active. | |
1572 | */ | |
1573 | static void wq_update_node_max_active(struct workqueue_struct *wq, int off_cpu) | |
1574 | { | |
1575 | struct cpumask *effective = unbound_effective_cpumask(wq); | |
1576 | int min_active = READ_ONCE(wq->min_active); | |
1577 | int max_active = READ_ONCE(wq->max_active); | |
1578 | int total_cpus, node; | |
1579 | ||
1580 | lockdep_assert_held(&wq->mutex); | |
1581 | ||
c5f8cd6c TH |
1582 | if (!wq_topo_initialized) |
1583 | return; | |
1584 | ||
15930da4 | 1585 | if (off_cpu >= 0 && !cpumask_test_cpu(off_cpu, effective)) |
5797b1c1 TH |
1586 | off_cpu = -1; |
1587 | ||
1588 | total_cpus = cpumask_weight_and(effective, cpu_online_mask); | |
1589 | if (off_cpu >= 0) | |
1590 | total_cpus--; | |
1591 | ||
91f09870 LJ |
1592 | /* If all CPUs of the wq get offline, use the default values */ |
1593 | if (unlikely(!total_cpus)) { | |
1594 | for_each_node(node) | |
1595 | wq_node_nr_active(wq, node)->max = min_active; | |
1596 | ||
1597 | wq_node_nr_active(wq, NUMA_NO_NODE)->max = max_active; | |
1598 | return; | |
1599 | } | |
1600 | ||
5797b1c1 TH |
1601 | for_each_node(node) { |
1602 | int node_cpus; | |
1603 | ||
1604 | node_cpus = cpumask_weight_and(effective, cpumask_of_node(node)); | |
1605 | if (off_cpu >= 0 && cpu_to_node(off_cpu) == node) | |
1606 | node_cpus--; | |
1607 | ||
1608 | wq_node_nr_active(wq, node)->max = | |
1609 | clamp(DIV_ROUND_UP(max_active * node_cpus, total_cpus), | |
1610 | min_active, max_active); | |
1611 | } | |
1612 | ||
d40f9202 | 1613 | wq_node_nr_active(wq, NUMA_NO_NODE)->max = max_active; |
5797b1c1 TH |
1614 | } |
1615 | ||
8864b4e5 TH |
1616 | /** |
1617 | * get_pwq - get an extra reference on the specified pool_workqueue | |
1618 | * @pwq: pool_workqueue to get | |
1619 | * | |
1620 | * Obtain an extra reference on @pwq. The caller should guarantee that | |
1621 | * @pwq has positive refcnt and be holding the matching pool->lock. | |
1622 | */ | |
1623 | static void get_pwq(struct pool_workqueue *pwq) | |
1624 | { | |
1625 | lockdep_assert_held(&pwq->pool->lock); | |
1626 | WARN_ON_ONCE(pwq->refcnt <= 0); | |
1627 | pwq->refcnt++; | |
1628 | } | |
1629 | ||
1630 | /** | |
1631 | * put_pwq - put a pool_workqueue reference | |
1632 | * @pwq: pool_workqueue to put | |
1633 | * | |
1634 | * Drop a reference of @pwq. If its refcnt reaches zero, schedule its | |
1635 | * destruction. The caller should be holding the matching pool->lock. | |
1636 | */ | |
1637 | static void put_pwq(struct pool_workqueue *pwq) | |
1638 | { | |
1639 | lockdep_assert_held(&pwq->pool->lock); | |
1640 | if (likely(--pwq->refcnt)) | |
1641 | return; | |
8864b4e5 | 1642 | /* |
967b494e TH |
1643 | * @pwq can't be released under pool->lock, bounce to a dedicated |
1644 | * kthread_worker to avoid A-A deadlocks. | |
8864b4e5 | 1645 | */ |
687a9aa5 | 1646 | kthread_queue_work(pwq_release_worker, &pwq->release_work); |
8864b4e5 TH |
1647 | } |
1648 | ||
dce90d47 TH |
1649 | /** |
1650 | * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock | |
1651 | * @pwq: pool_workqueue to put (can be %NULL) | |
1652 | * | |
1653 | * put_pwq() with locking. This function also allows %NULL @pwq. | |
1654 | */ | |
1655 | static void put_pwq_unlocked(struct pool_workqueue *pwq) | |
1656 | { | |
1657 | if (pwq) { | |
1658 | /* | |
24acfb71 | 1659 | * As both pwqs and pools are RCU protected, the |
dce90d47 TH |
1660 | * following lock operations are safe. |
1661 | */ | |
a9b8a985 | 1662 | raw_spin_lock_irq(&pwq->pool->lock); |
dce90d47 | 1663 | put_pwq(pwq); |
a9b8a985 | 1664 | raw_spin_unlock_irq(&pwq->pool->lock); |
dce90d47 TH |
1665 | } |
1666 | } | |
1667 | ||
afa87ce8 TH |
1668 | static bool pwq_is_empty(struct pool_workqueue *pwq) |
1669 | { | |
1670 | return !pwq->nr_active && list_empty(&pwq->inactive_works); | |
1671 | } | |
1672 | ||
4c638030 TH |
1673 | static void __pwq_activate_work(struct pool_workqueue *pwq, |
1674 | struct work_struct *work) | |
bf4ede01 | 1675 | { |
1c270b79 TH |
1676 | unsigned long *wdb = work_data_bits(work); |
1677 | ||
1678 | WARN_ON_ONCE(!(*wdb & WORK_STRUCT_INACTIVE)); | |
bf4ede01 | 1679 | trace_workqueue_activate_work(work); |
82607adc TH |
1680 | if (list_empty(&pwq->pool->worklist)) |
1681 | pwq->pool->watchdog_ts = jiffies; | |
112202d9 | 1682 | move_linked_works(work, &pwq->pool->worklist, NULL); |
1c270b79 | 1683 | __clear_bit(WORK_STRUCT_INACTIVE_BIT, wdb); |
4c638030 TH |
1684 | } |
1685 | ||
5797b1c1 TH |
1686 | static bool tryinc_node_nr_active(struct wq_node_nr_active *nna) |
1687 | { | |
1688 | int max = READ_ONCE(nna->max); | |
1689 | ||
1690 | while (true) { | |
1691 | int old, tmp; | |
1692 | ||
1693 | old = atomic_read(&nna->nr); | |
1694 | if (old >= max) | |
1695 | return false; | |
1696 | tmp = atomic_cmpxchg_relaxed(&nna->nr, old, old + 1); | |
1697 | if (tmp == old) | |
1698 | return true; | |
1699 | } | |
1700 | } | |
1701 | ||
1c270b79 TH |
1702 | /** |
1703 | * pwq_tryinc_nr_active - Try to increment nr_active for a pwq | |
1704 | * @pwq: pool_workqueue of interest | |
5797b1c1 | 1705 | * @fill: max_active may have increased, try to increase concurrency level |
1c270b79 TH |
1706 | * |
1707 | * Try to increment nr_active for @pwq. Returns %true if an nr_active count is | |
1708 | * successfully obtained. %false otherwise. | |
1709 | */ | |
5797b1c1 | 1710 | static bool pwq_tryinc_nr_active(struct pool_workqueue *pwq, bool fill) |
1c270b79 TH |
1711 | { |
1712 | struct workqueue_struct *wq = pwq->wq; | |
1713 | struct worker_pool *pool = pwq->pool; | |
91ccc6e7 | 1714 | struct wq_node_nr_active *nna = wq_node_nr_active(wq, pool->node); |
5797b1c1 | 1715 | bool obtained = false; |
1c270b79 TH |
1716 | |
1717 | lockdep_assert_held(&pool->lock); | |
1718 | ||
5797b1c1 | 1719 | if (!nna) { |
4cb1ef64 | 1720 | /* BH or per-cpu workqueue, pwq->nr_active is sufficient */ |
5797b1c1 TH |
1721 | obtained = pwq->nr_active < READ_ONCE(wq->max_active); |
1722 | goto out; | |
1723 | } | |
1724 | ||
4c065dbc WL |
1725 | if (unlikely(pwq->plugged)) |
1726 | return false; | |
1727 | ||
5797b1c1 TH |
1728 | /* |
1729 | * Unbound workqueue uses per-node shared nr_active $nna. If @pwq is | |
1730 | * already waiting on $nna, pwq_dec_nr_active() will maintain the | |
1731 | * concurrency level. Don't jump the line. | |
1732 | * | |
1733 | * We need to ignore the pending test after max_active has increased as | |
1734 | * pwq_dec_nr_active() can only maintain the concurrency level but not | |
1735 | * increase it. This is indicated by @fill. | |
1736 | */ | |
1737 | if (!list_empty(&pwq->pending_node) && likely(!fill)) | |
1738 | goto out; | |
1739 | ||
1740 | obtained = tryinc_node_nr_active(nna); | |
1741 | if (obtained) | |
1742 | goto out; | |
1743 | ||
1744 | /* | |
1745 | * Lockless acquisition failed. Lock, add ourself to $nna->pending_pwqs | |
1746 | * and try again. The smp_mb() is paired with the implied memory barrier | |
1747 | * of atomic_dec_return() in pwq_dec_nr_active() to ensure that either | |
1748 | * we see the decremented $nna->nr or they see non-empty | |
1749 | * $nna->pending_pwqs. | |
1750 | */ | |
1751 | raw_spin_lock(&nna->lock); | |
1752 | ||
1753 | if (list_empty(&pwq->pending_node)) | |
1754 | list_add_tail(&pwq->pending_node, &nna->pending_pwqs); | |
1755 | else if (likely(!fill)) | |
1756 | goto out_unlock; | |
1757 | ||
1758 | smp_mb(); | |
1759 | ||
1760 | obtained = tryinc_node_nr_active(nna); | |
1c270b79 | 1761 | |
5797b1c1 TH |
1762 | /* |
1763 | * If @fill, @pwq might have already been pending. Being spuriously | |
1764 | * pending in cold paths doesn't affect anything. Let's leave it be. | |
1765 | */ | |
1766 | if (obtained && likely(!fill)) | |
1767 | list_del_init(&pwq->pending_node); | |
1768 | ||
1769 | out_unlock: | |
1770 | raw_spin_unlock(&nna->lock); | |
1771 | out: | |
1772 | if (obtained) | |
1c270b79 TH |
1773 | pwq->nr_active++; |
1774 | return obtained; | |
1775 | } | |
1776 | ||
1777 | /** | |
1778 | * pwq_activate_first_inactive - Activate the first inactive work item on a pwq | |
1779 | * @pwq: pool_workqueue of interest | |
5797b1c1 | 1780 | * @fill: max_active may have increased, try to increase concurrency level |
1c270b79 TH |
1781 | * |
1782 | * Activate the first inactive work item of @pwq if available and allowed by | |
1783 | * max_active limit. | |
1784 | * | |
1785 | * Returns %true if an inactive work item has been activated. %false if no | |
1786 | * inactive work item is found or max_active limit is reached. | |
1787 | */ | |
5797b1c1 | 1788 | static bool pwq_activate_first_inactive(struct pool_workqueue *pwq, bool fill) |
1c270b79 TH |
1789 | { |
1790 | struct work_struct *work = | |
1791 | list_first_entry_or_null(&pwq->inactive_works, | |
1792 | struct work_struct, entry); | |
1793 | ||
5797b1c1 | 1794 | if (work && pwq_tryinc_nr_active(pwq, fill)) { |
1c270b79 TH |
1795 | __pwq_activate_work(pwq, work); |
1796 | return true; | |
1797 | } else { | |
1798 | return false; | |
1799 | } | |
1800 | } | |
1801 | ||
4c065dbc | 1802 | /** |
516d3dc9 WL |
1803 | * unplug_oldest_pwq - unplug the oldest pool_workqueue |
1804 | * @wq: workqueue_struct where its oldest pwq is to be unplugged | |
4c065dbc | 1805 | * |
516d3dc9 WL |
1806 | * This function should only be called for ordered workqueues where only the |
1807 | * oldest pwq is unplugged, the others are plugged to suspend execution to | |
1808 | * ensure proper work item ordering:: | |
4c065dbc WL |
1809 | * |
1810 | * dfl_pwq --------------+ [P] - plugged | |
1811 | * | | |
1812 | * v | |
1813 | * pwqs -> A -> B [P] -> C [P] (newest) | |
1814 | * | | | | |
1815 | * 1 3 5 | |
1816 | * | | | | |
1817 | * 2 4 6 | |
516d3dc9 WL |
1818 | * |
1819 | * When the oldest pwq is drained and removed, this function should be called | |
1820 | * to unplug the next oldest one to start its work item execution. Note that | |
1821 | * pwq's are linked into wq->pwqs with the oldest first, so the first one in | |
1822 | * the list is the oldest. | |
4c065dbc WL |
1823 | */ |
1824 | static void unplug_oldest_pwq(struct workqueue_struct *wq) | |
1825 | { | |
1826 | struct pool_workqueue *pwq; | |
1827 | ||
1828 | lockdep_assert_held(&wq->mutex); | |
1829 | ||
1830 | /* Caller should make sure that pwqs isn't empty before calling */ | |
1831 | pwq = list_first_entry_or_null(&wq->pwqs, struct pool_workqueue, | |
1832 | pwqs_node); | |
1833 | raw_spin_lock_irq(&pwq->pool->lock); | |
1834 | if (pwq->plugged) { | |
1835 | pwq->plugged = false; | |
1836 | if (pwq_activate_first_inactive(pwq, true)) | |
1837 | kick_pool(pwq->pool); | |
1838 | } | |
1839 | raw_spin_unlock_irq(&pwq->pool->lock); | |
1840 | } | |
1841 | ||
5797b1c1 TH |
1842 | /** |
1843 | * node_activate_pending_pwq - Activate a pending pwq on a wq_node_nr_active | |
1844 | * @nna: wq_node_nr_active to activate a pending pwq for | |
1845 | * @caller_pool: worker_pool the caller is locking | |
1846 | * | |
1847 | * Activate a pwq in @nna->pending_pwqs. Called with @caller_pool locked. | |
1848 | * @caller_pool may be unlocked and relocked to lock other worker_pools. | |
1849 | */ | |
1850 | static void node_activate_pending_pwq(struct wq_node_nr_active *nna, | |
1851 | struct worker_pool *caller_pool) | |
1852 | { | |
1853 | struct worker_pool *locked_pool = caller_pool; | |
1854 | struct pool_workqueue *pwq; | |
1855 | struct work_struct *work; | |
1856 | ||
1857 | lockdep_assert_held(&caller_pool->lock); | |
1858 | ||
1859 | raw_spin_lock(&nna->lock); | |
1860 | retry: | |
1861 | pwq = list_first_entry_or_null(&nna->pending_pwqs, | |
1862 | struct pool_workqueue, pending_node); | |
1863 | if (!pwq) | |
1864 | goto out_unlock; | |
1865 | ||
1866 | /* | |
1867 | * If @pwq is for a different pool than @locked_pool, we need to lock | |
1868 | * @pwq->pool->lock. Let's trylock first. If unsuccessful, do the unlock | |
1869 | * / lock dance. For that, we also need to release @nna->lock as it's | |
1870 | * nested inside pool locks. | |
1871 | */ | |
1872 | if (pwq->pool != locked_pool) { | |
1873 | raw_spin_unlock(&locked_pool->lock); | |
1874 | locked_pool = pwq->pool; | |
1875 | if (!raw_spin_trylock(&locked_pool->lock)) { | |
1876 | raw_spin_unlock(&nna->lock); | |
1877 | raw_spin_lock(&locked_pool->lock); | |
1878 | raw_spin_lock(&nna->lock); | |
1879 | goto retry; | |
1880 | } | |
1881 | } | |
1882 | ||
1883 | /* | |
1884 | * $pwq may not have any inactive work items due to e.g. cancellations. | |
1885 | * Drop it from pending_pwqs and see if there's another one. | |
1886 | */ | |
1887 | work = list_first_entry_or_null(&pwq->inactive_works, | |
1888 | struct work_struct, entry); | |
1889 | if (!work) { | |
1890 | list_del_init(&pwq->pending_node); | |
1891 | goto retry; | |
1892 | } | |
1893 | ||
1894 | /* | |
1895 | * Acquire an nr_active count and activate the inactive work item. If | |
1896 | * $pwq still has inactive work items, rotate it to the end of the | |
1897 | * pending_pwqs so that we round-robin through them. This means that | |
1898 | * inactive work items are not activated in queueing order which is fine | |
1899 | * given that there has never been any ordering across different pwqs. | |
1900 | */ | |
1901 | if (likely(tryinc_node_nr_active(nna))) { | |
1902 | pwq->nr_active++; | |
1903 | __pwq_activate_work(pwq, work); | |
1904 | ||
1905 | if (list_empty(&pwq->inactive_works)) | |
1906 | list_del_init(&pwq->pending_node); | |
1907 | else | |
1908 | list_move_tail(&pwq->pending_node, &nna->pending_pwqs); | |
1909 | ||
1910 | /* if activating a foreign pool, make sure it's running */ | |
1911 | if (pwq->pool != caller_pool) | |
1912 | kick_pool(pwq->pool); | |
1913 | } | |
1914 | ||
1915 | out_unlock: | |
1916 | raw_spin_unlock(&nna->lock); | |
1917 | if (locked_pool != caller_pool) { | |
1918 | raw_spin_unlock(&locked_pool->lock); | |
1919 | raw_spin_lock(&caller_pool->lock); | |
1920 | } | |
1921 | } | |
1922 | ||
1c270b79 TH |
1923 | /** |
1924 | * pwq_dec_nr_active - Retire an active count | |
1925 | * @pwq: pool_workqueue of interest | |
1926 | * | |
1927 | * Decrement @pwq's nr_active and try to activate the first inactive work item. | |
5797b1c1 | 1928 | * For unbound workqueues, this function may temporarily drop @pwq->pool->lock. |
1c270b79 TH |
1929 | */ |
1930 | static void pwq_dec_nr_active(struct pool_workqueue *pwq) | |
3aa62497 | 1931 | { |
1c270b79 | 1932 | struct worker_pool *pool = pwq->pool; |
91ccc6e7 | 1933 | struct wq_node_nr_active *nna = wq_node_nr_active(pwq->wq, pool->node); |
3aa62497 | 1934 | |
1c270b79 TH |
1935 | lockdep_assert_held(&pool->lock); |
1936 | ||
91ccc6e7 TH |
1937 | /* |
1938 | * @pwq->nr_active should be decremented for both percpu and unbound | |
1939 | * workqueues. | |
1940 | */ | |
1c270b79 | 1941 | pwq->nr_active--; |
91ccc6e7 TH |
1942 | |
1943 | /* | |
1944 | * For a percpu workqueue, it's simple. Just need to kick the first | |
1945 | * inactive work item on @pwq itself. | |
1946 | */ | |
1947 | if (!nna) { | |
5797b1c1 | 1948 | pwq_activate_first_inactive(pwq, false); |
91ccc6e7 TH |
1949 | return; |
1950 | } | |
1951 | ||
5797b1c1 TH |
1952 | /* |
1953 | * If @pwq is for an unbound workqueue, it's more complicated because | |
1954 | * multiple pwqs and pools may be sharing the nr_active count. When a | |
1955 | * pwq needs to wait for an nr_active count, it puts itself on | |
1956 | * $nna->pending_pwqs. The following atomic_dec_return()'s implied | |
1957 | * memory barrier is paired with smp_mb() in pwq_tryinc_nr_active() to | |
1958 | * guarantee that either we see non-empty pending_pwqs or they see | |
1959 | * decremented $nna->nr. | |
1960 | * | |
1961 | * $nna->max may change as CPUs come online/offline and @pwq->wq's | |
1962 | * max_active gets updated. However, it is guaranteed to be equal to or | |
1963 | * larger than @pwq->wq->min_active which is above zero unless freezing. | |
1964 | * This maintains the forward progress guarantee. | |
1965 | */ | |
1966 | if (atomic_dec_return(&nna->nr) >= READ_ONCE(nna->max)) | |
1967 | return; | |
1968 | ||
1969 | if (!list_empty(&nna->pending_pwqs)) | |
1970 | node_activate_pending_pwq(nna, pool); | |
3aa62497 LJ |
1971 | } |
1972 | ||
bf4ede01 | 1973 | /** |
112202d9 TH |
1974 | * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight |
1975 | * @pwq: pwq of interest | |
c4560c2c | 1976 | * @work_data: work_data of work which left the queue |
bf4ede01 TH |
1977 | * |
1978 | * A work either has completed or is removed from pending queue, | |
112202d9 | 1979 | * decrement nr_in_flight of its pwq and handle workqueue flushing. |
bf4ede01 | 1980 | * |
dd6c3c54 TH |
1981 | * NOTE: |
1982 | * For unbound workqueues, this function may temporarily drop @pwq->pool->lock | |
1983 | * and thus should be called after all other state updates for the in-flight | |
1984 | * work item is complete. | |
1985 | * | |
bf4ede01 | 1986 | * CONTEXT: |
a9b8a985 | 1987 | * raw_spin_lock_irq(pool->lock). |
bf4ede01 | 1988 | */ |
c4560c2c | 1989 | static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, unsigned long work_data) |
bf4ede01 | 1990 | { |
c4560c2c LJ |
1991 | int color = get_work_color(work_data); |
1992 | ||
1c270b79 TH |
1993 | if (!(work_data & WORK_STRUCT_INACTIVE)) |
1994 | pwq_dec_nr_active(pwq); | |
018f3a13 | 1995 | |
112202d9 | 1996 | pwq->nr_in_flight[color]--; |
bf4ede01 | 1997 | |
bf4ede01 | 1998 | /* is flush in progress and are we at the flushing tip? */ |
112202d9 | 1999 | if (likely(pwq->flush_color != color)) |
8864b4e5 | 2000 | goto out_put; |
bf4ede01 TH |
2001 | |
2002 | /* are there still in-flight works? */ | |
112202d9 | 2003 | if (pwq->nr_in_flight[color]) |
8864b4e5 | 2004 | goto out_put; |
bf4ede01 | 2005 | |
112202d9 TH |
2006 | /* this pwq is done, clear flush_color */ |
2007 | pwq->flush_color = -1; | |
bf4ede01 TH |
2008 | |
2009 | /* | |
112202d9 | 2010 | * If this was the last pwq, wake up the first flusher. It |
bf4ede01 TH |
2011 | * will handle the rest. |
2012 | */ | |
112202d9 TH |
2013 | if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) |
2014 | complete(&pwq->wq->first_flusher->done); | |
8864b4e5 TH |
2015 | out_put: |
2016 | put_pwq(pwq); | |
bf4ede01 TH |
2017 | } |
2018 | ||
36e227d2 | 2019 | /** |
bbb68dfa | 2020 | * try_to_grab_pending - steal work item from worklist and disable irq |
36e227d2 | 2021 | * @work: work item to steal |
c5f5b942 | 2022 | * @cflags: %WORK_CANCEL_ flags |
c26e2f2e | 2023 | * @irq_flags: place to store irq state |
36e227d2 TH |
2024 | * |
2025 | * Try to grab PENDING bit of @work. This function can handle @work in any | |
d185af30 | 2026 | * stable state - idle, on timer or on worklist. |
36e227d2 | 2027 | * |
d185af30 | 2028 | * Return: |
3eb6b31b MCC |
2029 | * |
2030 | * ======== ================================================================ | |
36e227d2 TH |
2031 | * 1 if @work was pending and we successfully stole PENDING |
2032 | * 0 if @work was idle and we claimed PENDING | |
2033 | * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry | |
3eb6b31b | 2034 | * ======== ================================================================ |
36e227d2 | 2035 | * |
d185af30 | 2036 | * Note: |
bbb68dfa | 2037 | * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting |
e0aecdd8 TH |
2038 | * interrupted while holding PENDING and @work off queue, irq must be |
2039 | * disabled on entry. This, combined with delayed_work->timer being | |
2040 | * irqsafe, ensures that we return -EAGAIN for finite short period of time. | |
bbb68dfa TH |
2041 | * |
2042 | * On successful return, >= 0, irq is disabled and the caller is | |
c26e2f2e | 2043 | * responsible for releasing it using local_irq_restore(*@irq_flags). |
bbb68dfa | 2044 | * |
e0aecdd8 | 2045 | * This function is safe to call from any context including IRQ handler. |
bf4ede01 | 2046 | */ |
c5f5b942 | 2047 | static int try_to_grab_pending(struct work_struct *work, u32 cflags, |
c26e2f2e | 2048 | unsigned long *irq_flags) |
bf4ede01 | 2049 | { |
d565ed63 | 2050 | struct worker_pool *pool; |
112202d9 | 2051 | struct pool_workqueue *pwq; |
bf4ede01 | 2052 | |
c26e2f2e | 2053 | local_irq_save(*irq_flags); |
bbb68dfa | 2054 | |
36e227d2 | 2055 | /* try to steal the timer if it exists */ |
c5f5b942 | 2056 | if (cflags & WORK_CANCEL_DELAYED) { |
36e227d2 TH |
2057 | struct delayed_work *dwork = to_delayed_work(work); |
2058 | ||
e0aecdd8 TH |
2059 | /* |
2060 | * dwork->timer is irqsafe. If del_timer() fails, it's | |
2061 | * guaranteed that the timer is not queued anywhere and not | |
2062 | * running on the local CPU. | |
2063 | */ | |
36e227d2 TH |
2064 | if (likely(del_timer(&dwork->timer))) |
2065 | return 1; | |
2066 | } | |
2067 | ||
2068 | /* try to claim PENDING the normal way */ | |
bf4ede01 TH |
2069 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) |
2070 | return 0; | |
2071 | ||
24acfb71 | 2072 | rcu_read_lock(); |
bf4ede01 TH |
2073 | /* |
2074 | * The queueing is in progress, or it is already queued. Try to | |
2075 | * steal it from ->worklist without clearing WORK_STRUCT_PENDING. | |
2076 | */ | |
d565ed63 TH |
2077 | pool = get_work_pool(work); |
2078 | if (!pool) | |
bbb68dfa | 2079 | goto fail; |
bf4ede01 | 2080 | |
a9b8a985 | 2081 | raw_spin_lock(&pool->lock); |
0b3dae68 | 2082 | /* |
112202d9 TH |
2083 | * work->data is guaranteed to point to pwq only while the work |
2084 | * item is queued on pwq->wq, and both updating work->data to point | |
2085 | * to pwq on queueing and to pool on dequeueing are done under | |
2086 | * pwq->pool->lock. This in turn guarantees that, if work->data | |
2087 | * points to pwq which is associated with a locked pool, the work | |
0b3dae68 LJ |
2088 | * item is currently queued on that pool. |
2089 | */ | |
112202d9 TH |
2090 | pwq = get_work_pwq(work); |
2091 | if (pwq && pwq->pool == pool) { | |
b56c7207 | 2092 | unsigned long work_data = *work_data_bits(work); |
c70e1779 | 2093 | |
16062836 TH |
2094 | debug_work_deactivate(work); |
2095 | ||
2096 | /* | |
018f3a13 LJ |
2097 | * A cancelable inactive work item must be in the |
2098 | * pwq->inactive_works since a queued barrier can't be | |
2099 | * canceled (see the comments in insert_wq_barrier()). | |
2100 | * | |
b56c7207 | 2101 | * An inactive work item cannot be deleted directly because |
d812796e | 2102 | * it might have linked barrier work items which, if left |
f97a4a1a | 2103 | * on the inactive_works list, will confuse pwq->nr_active |
b56c7207 LJ |
2104 | * management later on and cause stall. Move the linked |
2105 | * barrier work items to the worklist when deleting the grabbed | |
2106 | * item. Also keep WORK_STRUCT_INACTIVE in work_data, so that | |
2107 | * it doesn't participate in nr_active management in later | |
2108 | * pwq_dec_nr_in_flight(). | |
16062836 | 2109 | */ |
b56c7207 LJ |
2110 | if (work_data & WORK_STRUCT_INACTIVE) |
2111 | move_linked_works(work, &pwq->pool->worklist, NULL); | |
16062836 TH |
2112 | |
2113 | list_del_init(&work->entry); | |
16062836 | 2114 | |
c70e1779 TH |
2115 | /* |
2116 | * work->data points to pwq iff queued. Let's point to pool. As | |
2117 | * this destroys work->data needed by the next step, stash it. | |
2118 | */ | |
456a78ee TH |
2119 | set_work_pool_and_keep_pending(work, pool->id, |
2120 | pool_offq_flags(pool)); | |
16062836 | 2121 | |
dd6c3c54 | 2122 | /* must be the last step, see the function comment */ |
c70e1779 | 2123 | pwq_dec_nr_in_flight(pwq, work_data); |
dd6c3c54 | 2124 | |
a9b8a985 | 2125 | raw_spin_unlock(&pool->lock); |
24acfb71 | 2126 | rcu_read_unlock(); |
16062836 | 2127 | return 1; |
bf4ede01 | 2128 | } |
a9b8a985 | 2129 | raw_spin_unlock(&pool->lock); |
bbb68dfa | 2130 | fail: |
24acfb71 | 2131 | rcu_read_unlock(); |
c26e2f2e | 2132 | local_irq_restore(*irq_flags); |
36e227d2 | 2133 | return -EAGAIN; |
bf4ede01 TH |
2134 | } |
2135 | ||
978b8409 TH |
2136 | /** |
2137 | * work_grab_pending - steal work item from worklist and disable irq | |
2138 | * @work: work item to steal | |
2139 | * @cflags: %WORK_CANCEL_ flags | |
2140 | * @irq_flags: place to store IRQ state | |
2141 | * | |
2142 | * Grab PENDING bit of @work. @work can be in any stable state - idle, on timer | |
2143 | * or on worklist. | |
2144 | * | |
f09b10b6 | 2145 | * Can be called from any context. IRQ is disabled on return with IRQ state |
978b8409 TH |
2146 | * stored in *@irq_flags. The caller is responsible for re-enabling it using |
2147 | * local_irq_restore(). | |
2148 | * | |
2149 | * Returns %true if @work was pending. %false if idle. | |
2150 | */ | |
2151 | static bool work_grab_pending(struct work_struct *work, u32 cflags, | |
2152 | unsigned long *irq_flags) | |
2153 | { | |
978b8409 TH |
2154 | int ret; |
2155 | ||
f09b10b6 TH |
2156 | while (true) { |
2157 | ret = try_to_grab_pending(work, cflags, irq_flags); | |
2158 | if (ret >= 0) | |
2159 | return ret; | |
2160 | cpu_relax(); | |
2161 | } | |
978b8409 TH |
2162 | } |
2163 | ||
4690c4ab | 2164 | /** |
706026c2 | 2165 | * insert_work - insert a work into a pool |
112202d9 | 2166 | * @pwq: pwq @work belongs to |
4690c4ab TH |
2167 | * @work: work to insert |
2168 | * @head: insertion point | |
2169 | * @extra_flags: extra WORK_STRUCT_* flags to set | |
2170 | * | |
112202d9 | 2171 | * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to |
706026c2 | 2172 | * work_struct flags. |
4690c4ab TH |
2173 | * |
2174 | * CONTEXT: | |
a9b8a985 | 2175 | * raw_spin_lock_irq(pool->lock). |
4690c4ab | 2176 | */ |
112202d9 TH |
2177 | static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, |
2178 | struct list_head *head, unsigned int extra_flags) | |
b89deed3 | 2179 | { |
fe089f87 | 2180 | debug_work_activate(work); |
e22bee78 | 2181 | |
e89a85d6 | 2182 | /* record the work call stack in order to print it in KASAN reports */ |
f70da745 | 2183 | kasan_record_aux_stack_noalloc(work); |
e89a85d6 | 2184 | |
4690c4ab | 2185 | /* we own @work, set data and link */ |
112202d9 | 2186 | set_work_pwq(work, pwq, extra_flags); |
1a4d9b0a | 2187 | list_add_tail(&work->entry, head); |
8864b4e5 | 2188 | get_pwq(pwq); |
b89deed3 ON |
2189 | } |
2190 | ||
c8efcc25 TH |
2191 | /* |
2192 | * Test whether @work is being queued from another work executing on the | |
8d03ecfe | 2193 | * same workqueue. |
c8efcc25 TH |
2194 | */ |
2195 | static bool is_chained_work(struct workqueue_struct *wq) | |
2196 | { | |
8d03ecfe TH |
2197 | struct worker *worker; |
2198 | ||
2199 | worker = current_wq_worker(); | |
2200 | /* | |
bf393fd4 | 2201 | * Return %true iff I'm a worker executing a work item on @wq. If |
8d03ecfe TH |
2202 | * I'm @worker, it's safe to dereference it without locking. |
2203 | */ | |
112202d9 | 2204 | return worker && worker->current_pwq->wq == wq; |
c8efcc25 TH |
2205 | } |
2206 | ||
ef557180 MG |
2207 | /* |
2208 | * When queueing an unbound work item to a wq, prefer local CPU if allowed | |
2209 | * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to | |
2210 | * avoid perturbing sensitive tasks. | |
2211 | */ | |
2212 | static int wq_select_unbound_cpu(int cpu) | |
2213 | { | |
2214 | int new_cpu; | |
2215 | ||
f303fccb TH |
2216 | if (likely(!wq_debug_force_rr_cpu)) { |
2217 | if (cpumask_test_cpu(cpu, wq_unbound_cpumask)) | |
2218 | return cpu; | |
a8ec5880 AF |
2219 | } else { |
2220 | pr_warn_once("workqueue: round-robin CPU selection forced, expect performance impact\n"); | |
f303fccb TH |
2221 | } |
2222 | ||
ef557180 MG |
2223 | new_cpu = __this_cpu_read(wq_rr_cpu_last); |
2224 | new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask); | |
2225 | if (unlikely(new_cpu >= nr_cpu_ids)) { | |
2226 | new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask); | |
2227 | if (unlikely(new_cpu >= nr_cpu_ids)) | |
2228 | return cpu; | |
2229 | } | |
2230 | __this_cpu_write(wq_rr_cpu_last, new_cpu); | |
2231 | ||
2232 | return new_cpu; | |
2233 | } | |
2234 | ||
d84ff051 | 2235 | static void __queue_work(int cpu, struct workqueue_struct *wq, |
1da177e4 LT |
2236 | struct work_struct *work) |
2237 | { | |
112202d9 | 2238 | struct pool_workqueue *pwq; |
fe089f87 | 2239 | struct worker_pool *last_pool, *pool; |
8a2e8e5d | 2240 | unsigned int work_flags; |
b75cac93 | 2241 | unsigned int req_cpu = cpu; |
8930caba TH |
2242 | |
2243 | /* | |
2244 | * While a work item is PENDING && off queue, a task trying to | |
2245 | * steal the PENDING will busy-loop waiting for it to either get | |
2246 | * queued or lose PENDING. Grabbing PENDING and queueing should | |
2247 | * happen with IRQ disabled. | |
2248 | */ | |
8e8eb730 | 2249 | lockdep_assert_irqs_disabled(); |
1da177e4 | 2250 | |
33e3f0a3 RC |
2251 | /* |
2252 | * For a draining wq, only works from the same workqueue are | |
2253 | * allowed. The __WQ_DESTROYING helps to spot the issue that | |
2254 | * queues a new work item to a wq after destroy_workqueue(wq). | |
2255 | */ | |
2256 | if (unlikely(wq->flags & (__WQ_DESTROYING | __WQ_DRAINING) && | |
2257 | WARN_ON_ONCE(!is_chained_work(wq)))) | |
e41e704b | 2258 | return; |
24acfb71 | 2259 | rcu_read_lock(); |
9e8cd2f5 | 2260 | retry: |
c9178087 | 2261 | /* pwq which will be used unless @work is executing elsewhere */ |
636b927e TH |
2262 | if (req_cpu == WORK_CPU_UNBOUND) { |
2263 | if (wq->flags & WQ_UNBOUND) | |
aa202f1f | 2264 | cpu = wq_select_unbound_cpu(raw_smp_processor_id()); |
636b927e | 2265 | else |
aa202f1f | 2266 | cpu = raw_smp_processor_id(); |
aa202f1f | 2267 | } |
dbf2576e | 2268 | |
636b927e | 2269 | pwq = rcu_dereference(*per_cpu_ptr(wq->cpu_pwq, cpu)); |
fe089f87 TH |
2270 | pool = pwq->pool; |
2271 | ||
c9178087 TH |
2272 | /* |
2273 | * If @work was previously on a different pool, it might still be | |
2274 | * running there, in which case the work needs to be queued on that | |
2275 | * pool to guarantee non-reentrancy. | |
58629d48 LJ |
2276 | * |
2277 | * For ordered workqueue, work items must be queued on the newest pwq | |
2278 | * for accurate order management. Guaranteed order also guarantees | |
2279 | * non-reentrancy. See the comments above unplug_oldest_pwq(). | |
c9178087 TH |
2280 | */ |
2281 | last_pool = get_work_pool(work); | |
58629d48 | 2282 | if (last_pool && last_pool != pool && !(wq->flags & __WQ_ORDERED)) { |
c9178087 | 2283 | struct worker *worker; |
18aa9eff | 2284 | |
a9b8a985 | 2285 | raw_spin_lock(&last_pool->lock); |
18aa9eff | 2286 | |
c9178087 | 2287 | worker = find_worker_executing_work(last_pool, work); |
18aa9eff | 2288 | |
c9178087 TH |
2289 | if (worker && worker->current_pwq->wq == wq) { |
2290 | pwq = worker->current_pwq; | |
fe089f87 TH |
2291 | pool = pwq->pool; |
2292 | WARN_ON_ONCE(pool != last_pool); | |
8930caba | 2293 | } else { |
c9178087 | 2294 | /* meh... not running there, queue here */ |
a9b8a985 | 2295 | raw_spin_unlock(&last_pool->lock); |
fe089f87 | 2296 | raw_spin_lock(&pool->lock); |
8930caba | 2297 | } |
f3421797 | 2298 | } else { |
fe089f87 | 2299 | raw_spin_lock(&pool->lock); |
502ca9d8 TH |
2300 | } |
2301 | ||
9e8cd2f5 | 2302 | /* |
636b927e TH |
2303 | * pwq is determined and locked. For unbound pools, we could have raced |
2304 | * with pwq release and it could already be dead. If its refcnt is zero, | |
2305 | * repeat pwq selection. Note that unbound pwqs never die without | |
2306 | * another pwq replacing it in cpu_pwq or while work items are executing | |
2307 | * on it, so the retrying is guaranteed to make forward-progress. | |
9e8cd2f5 TH |
2308 | */ |
2309 | if (unlikely(!pwq->refcnt)) { | |
2310 | if (wq->flags & WQ_UNBOUND) { | |
fe089f87 | 2311 | raw_spin_unlock(&pool->lock); |
9e8cd2f5 TH |
2312 | cpu_relax(); |
2313 | goto retry; | |
2314 | } | |
2315 | /* oops */ | |
2316 | WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", | |
2317 | wq->name, cpu); | |
2318 | } | |
2319 | ||
112202d9 TH |
2320 | /* pwq determined, queue */ |
2321 | trace_workqueue_queue_work(req_cpu, pwq, work); | |
502ca9d8 | 2322 | |
24acfb71 TG |
2323 | if (WARN_ON(!list_empty(&work->entry))) |
2324 | goto out; | |
1e19ffc6 | 2325 | |
112202d9 TH |
2326 | pwq->nr_in_flight[pwq->work_color]++; |
2327 | work_flags = work_color_to_flags(pwq->work_color); | |
1e19ffc6 | 2328 | |
a045a272 TH |
2329 | /* |
2330 | * Limit the number of concurrently active work items to max_active. | |
2331 | * @work must also queue behind existing inactive work items to maintain | |
2332 | * ordering when max_active changes. See wq_adjust_max_active(). | |
2333 | */ | |
5797b1c1 | 2334 | if (list_empty(&pwq->inactive_works) && pwq_tryinc_nr_active(pwq, false)) { |
fe089f87 TH |
2335 | if (list_empty(&pool->worklist)) |
2336 | pool->watchdog_ts = jiffies; | |
2337 | ||
cdadf009 | 2338 | trace_workqueue_activate_work(work); |
fe089f87 | 2339 | insert_work(pwq, work, &pool->worklist, work_flags); |
0219a352 | 2340 | kick_pool(pool); |
8a2e8e5d | 2341 | } else { |
f97a4a1a | 2342 | work_flags |= WORK_STRUCT_INACTIVE; |
fe089f87 | 2343 | insert_work(pwq, work, &pwq->inactive_works, work_flags); |
8a2e8e5d | 2344 | } |
1e19ffc6 | 2345 | |
24acfb71 | 2346 | out: |
fe089f87 | 2347 | raw_spin_unlock(&pool->lock); |
24acfb71 | 2348 | rcu_read_unlock(); |
1da177e4 LT |
2349 | } |
2350 | ||
86898fa6 TH |
2351 | static bool clear_pending_if_disabled(struct work_struct *work) |
2352 | { | |
2353 | unsigned long data = *work_data_bits(work); | |
2354 | struct work_offq_data offqd; | |
2355 | ||
2356 | if (likely((data & WORK_STRUCT_PWQ) || | |
2357 | !(data & WORK_OFFQ_DISABLE_MASK))) | |
2358 | return false; | |
2359 | ||
2360 | work_offqd_unpack(&offqd, data); | |
2361 | set_work_pool_and_clear_pending(work, offqd.pool_id, | |
2362 | work_offqd_pack_flags(&offqd)); | |
2363 | return true; | |
2364 | } | |
2365 | ||
0fcb78c2 | 2366 | /** |
c1a220e7 ZR |
2367 | * queue_work_on - queue work on specific cpu |
2368 | * @cpu: CPU number to execute work on | |
0fcb78c2 REB |
2369 | * @wq: workqueue to use |
2370 | * @work: work to queue | |
2371 | * | |
c1a220e7 | 2372 | * We queue the work to a specific CPU, the caller must ensure it |
443378f0 PM |
2373 | * can't go away. Callers that fail to ensure that the specified |
2374 | * CPU cannot go away will execute on a randomly chosen CPU. | |
854f5cc5 PM |
2375 | * But note well that callers specifying a CPU that never has been |
2376 | * online will get a splat. | |
d185af30 YB |
2377 | * |
2378 | * Return: %false if @work was already on a queue, %true otherwise. | |
1da177e4 | 2379 | */ |
d4283e93 TH |
2380 | bool queue_work_on(int cpu, struct workqueue_struct *wq, |
2381 | struct work_struct *work) | |
1da177e4 | 2382 | { |
d4283e93 | 2383 | bool ret = false; |
c26e2f2e | 2384 | unsigned long irq_flags; |
ef1ca236 | 2385 | |
c26e2f2e | 2386 | local_irq_save(irq_flags); |
c1a220e7 | 2387 | |
86898fa6 TH |
2388 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)) && |
2389 | !clear_pending_if_disabled(work)) { | |
4690c4ab | 2390 | __queue_work(cpu, wq, work); |
d4283e93 | 2391 | ret = true; |
c1a220e7 | 2392 | } |
ef1ca236 | 2393 | |
c26e2f2e | 2394 | local_irq_restore(irq_flags); |
1da177e4 LT |
2395 | return ret; |
2396 | } | |
ad7b1f84 | 2397 | EXPORT_SYMBOL(queue_work_on); |
1da177e4 | 2398 | |
8204e0c1 | 2399 | /** |
fef59c9c | 2400 | * select_numa_node_cpu - Select a CPU based on NUMA node |
8204e0c1 AD |
2401 | * @node: NUMA node ID that we want to select a CPU from |
2402 | * | |
2403 | * This function will attempt to find a "random" cpu available on a given | |
2404 | * node. If there are no CPUs available on the given node it will return | |
2405 | * WORK_CPU_UNBOUND indicating that we should just schedule to any | |
2406 | * available CPU if we need to schedule this work. | |
2407 | */ | |
fef59c9c | 2408 | static int select_numa_node_cpu(int node) |
8204e0c1 AD |
2409 | { |
2410 | int cpu; | |
2411 | ||
8204e0c1 AD |
2412 | /* Delay binding to CPU if node is not valid or online */ |
2413 | if (node < 0 || node >= MAX_NUMNODES || !node_online(node)) | |
2414 | return WORK_CPU_UNBOUND; | |
2415 | ||
2416 | /* Use local node/cpu if we are already there */ | |
2417 | cpu = raw_smp_processor_id(); | |
2418 | if (node == cpu_to_node(cpu)) | |
2419 | return cpu; | |
2420 | ||
2421 | /* Use "random" otherwise know as "first" online CPU of node */ | |
2422 | cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask); | |
2423 | ||
2424 | /* If CPU is valid return that, otherwise just defer */ | |
2425 | return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND; | |
2426 | } | |
2427 | ||
2428 | /** | |
2429 | * queue_work_node - queue work on a "random" cpu for a given NUMA node | |
2430 | * @node: NUMA node that we are targeting the work for | |
2431 | * @wq: workqueue to use | |
2432 | * @work: work to queue | |
2433 | * | |
2434 | * We queue the work to a "random" CPU within a given NUMA node. The basic | |
2435 | * idea here is to provide a way to somehow associate work with a given | |
2436 | * NUMA node. | |
2437 | * | |
2438 | * This function will only make a best effort attempt at getting this onto | |
2439 | * the right NUMA node. If no node is requested or the requested node is | |
2440 | * offline then we just fall back to standard queue_work behavior. | |
2441 | * | |
2442 | * Currently the "random" CPU ends up being the first available CPU in the | |
2443 | * intersection of cpu_online_mask and the cpumask of the node, unless we | |
2444 | * are running on the node. In that case we just use the current CPU. | |
2445 | * | |
2446 | * Return: %false if @work was already on a queue, %true otherwise. | |
2447 | */ | |
2448 | bool queue_work_node(int node, struct workqueue_struct *wq, | |
2449 | struct work_struct *work) | |
2450 | { | |
c26e2f2e | 2451 | unsigned long irq_flags; |
8204e0c1 AD |
2452 | bool ret = false; |
2453 | ||
2454 | /* | |
2455 | * This current implementation is specific to unbound workqueues. | |
2456 | * Specifically we only return the first available CPU for a given | |
2457 | * node instead of cycling through individual CPUs within the node. | |
2458 | * | |
2459 | * If this is used with a per-cpu workqueue then the logic in | |
2460 | * workqueue_select_cpu_near would need to be updated to allow for | |
2461 | * some round robin type logic. | |
2462 | */ | |
2463 | WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)); | |
2464 | ||
c26e2f2e | 2465 | local_irq_save(irq_flags); |
8204e0c1 | 2466 | |
86898fa6 TH |
2467 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)) && |
2468 | !clear_pending_if_disabled(work)) { | |
fef59c9c | 2469 | int cpu = select_numa_node_cpu(node); |
8204e0c1 AD |
2470 | |
2471 | __queue_work(cpu, wq, work); | |
2472 | ret = true; | |
2473 | } | |
2474 | ||
c26e2f2e | 2475 | local_irq_restore(irq_flags); |
8204e0c1 AD |
2476 | return ret; |
2477 | } | |
2478 | EXPORT_SYMBOL_GPL(queue_work_node); | |
2479 | ||
8c20feb6 | 2480 | void delayed_work_timer_fn(struct timer_list *t) |
1da177e4 | 2481 | { |
8c20feb6 | 2482 | struct delayed_work *dwork = from_timer(dwork, t, timer); |
1da177e4 | 2483 | |
e0aecdd8 | 2484 | /* should have been called from irqsafe timer with irq already off */ |
60c057bc | 2485 | __queue_work(dwork->cpu, dwork->wq, &dwork->work); |
1da177e4 | 2486 | } |
1438ade5 | 2487 | EXPORT_SYMBOL(delayed_work_timer_fn); |
1da177e4 | 2488 | |
7beb2edf TH |
2489 | static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, |
2490 | struct delayed_work *dwork, unsigned long delay) | |
1da177e4 | 2491 | { |
7beb2edf TH |
2492 | struct timer_list *timer = &dwork->timer; |
2493 | struct work_struct *work = &dwork->work; | |
7beb2edf | 2494 | |
637fdbae | 2495 | WARN_ON_ONCE(!wq); |
4b243563 | 2496 | WARN_ON_ONCE(timer->function != delayed_work_timer_fn); |
fc4b514f TH |
2497 | WARN_ON_ONCE(timer_pending(timer)); |
2498 | WARN_ON_ONCE(!list_empty(&work->entry)); | |
7beb2edf | 2499 | |
8852aac2 TH |
2500 | /* |
2501 | * If @delay is 0, queue @dwork->work immediately. This is for | |
2502 | * both optimization and correctness. The earliest @timer can | |
2503 | * expire is on the closest next tick and delayed_work users depend | |
2504 | * on that there's no such delay when @delay is 0. | |
2505 | */ | |
2506 | if (!delay) { | |
2507 | __queue_work(cpu, wq, &dwork->work); | |
2508 | return; | |
2509 | } | |
2510 | ||
60c057bc | 2511 | dwork->wq = wq; |
1265057f | 2512 | dwork->cpu = cpu; |
7beb2edf TH |
2513 | timer->expires = jiffies + delay; |
2514 | ||
aae17ebb LB |
2515 | if (housekeeping_enabled(HK_TYPE_TIMER)) { |
2516 | /* If the current cpu is a housekeeping cpu, use it. */ | |
2517 | cpu = smp_processor_id(); | |
2518 | if (!housekeeping_test_cpu(cpu, HK_TYPE_TIMER)) | |
2519 | cpu = housekeeping_any_cpu(HK_TYPE_TIMER); | |
041bd12e | 2520 | add_timer_on(timer, cpu); |
aae17ebb LB |
2521 | } else { |
2522 | if (likely(cpu == WORK_CPU_UNBOUND)) | |
d08c407f | 2523 | add_timer_global(timer); |
aae17ebb LB |
2524 | else |
2525 | add_timer_on(timer, cpu); | |
2526 | } | |
1da177e4 LT |
2527 | } |
2528 | ||
0fcb78c2 REB |
2529 | /** |
2530 | * queue_delayed_work_on - queue work on specific CPU after delay | |
2531 | * @cpu: CPU number to execute work on | |
2532 | * @wq: workqueue to use | |
af9997e4 | 2533 | * @dwork: work to queue |
0fcb78c2 REB |
2534 | * @delay: number of jiffies to wait before queueing |
2535 | * | |
d185af30 | 2536 | * Return: %false if @work was already on a queue, %true otherwise. If |
715f1300 TH |
2537 | * @delay is zero and @dwork is idle, it will be scheduled for immediate |
2538 | * execution. | |
0fcb78c2 | 2539 | */ |
d4283e93 TH |
2540 | bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, |
2541 | struct delayed_work *dwork, unsigned long delay) | |
7a6bc1cd | 2542 | { |
52bad64d | 2543 | struct work_struct *work = &dwork->work; |
d4283e93 | 2544 | bool ret = false; |
c26e2f2e | 2545 | unsigned long irq_flags; |
7a6bc1cd | 2546 | |
8930caba | 2547 | /* read the comment in __queue_work() */ |
c26e2f2e | 2548 | local_irq_save(irq_flags); |
7a6bc1cd | 2549 | |
86898fa6 TH |
2550 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)) && |
2551 | !clear_pending_if_disabled(work)) { | |
7beb2edf | 2552 | __queue_delayed_work(cpu, wq, dwork, delay); |
d4283e93 | 2553 | ret = true; |
7a6bc1cd | 2554 | } |
8a3e77cc | 2555 | |
c26e2f2e | 2556 | local_irq_restore(irq_flags); |
7a6bc1cd VP |
2557 | return ret; |
2558 | } | |
ad7b1f84 | 2559 | EXPORT_SYMBOL(queue_delayed_work_on); |
c7fc77f7 | 2560 | |
8376fe22 TH |
2561 | /** |
2562 | * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU | |
2563 | * @cpu: CPU number to execute work on | |
2564 | * @wq: workqueue to use | |
2565 | * @dwork: work to queue | |
2566 | * @delay: number of jiffies to wait before queueing | |
2567 | * | |
2568 | * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, | |
2569 | * modify @dwork's timer so that it expires after @delay. If @delay is | |
2570 | * zero, @work is guaranteed to be scheduled immediately regardless of its | |
2571 | * current state. | |
2572 | * | |
d185af30 | 2573 | * Return: %false if @dwork was idle and queued, %true if @dwork was |
8376fe22 TH |
2574 | * pending and its timer was modified. |
2575 | * | |
e0aecdd8 | 2576 | * This function is safe to call from any context including IRQ handler. |
8376fe22 TH |
2577 | * See try_to_grab_pending() for details. |
2578 | */ | |
2579 | bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, | |
2580 | struct delayed_work *dwork, unsigned long delay) | |
2581 | { | |
c26e2f2e | 2582 | unsigned long irq_flags; |
f09b10b6 | 2583 | bool ret; |
c7fc77f7 | 2584 | |
f09b10b6 | 2585 | ret = work_grab_pending(&dwork->work, WORK_CANCEL_DELAYED, &irq_flags); |
63bc0362 | 2586 | |
f09b10b6 | 2587 | if (!clear_pending_if_disabled(&dwork->work)) |
8376fe22 | 2588 | __queue_delayed_work(cpu, wq, dwork, delay); |
8376fe22 | 2589 | |
f09b10b6 | 2590 | local_irq_restore(irq_flags); |
7a6bc1cd VP |
2591 | return ret; |
2592 | } | |
8376fe22 TH |
2593 | EXPORT_SYMBOL_GPL(mod_delayed_work_on); |
2594 | ||
05f0fe6b TH |
2595 | static void rcu_work_rcufn(struct rcu_head *rcu) |
2596 | { | |
2597 | struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu); | |
2598 | ||
2599 | /* read the comment in __queue_work() */ | |
2600 | local_irq_disable(); | |
2601 | __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work); | |
2602 | local_irq_enable(); | |
2603 | } | |
2604 | ||
2605 | /** | |
2606 | * queue_rcu_work - queue work after a RCU grace period | |
2607 | * @wq: workqueue to use | |
2608 | * @rwork: work to queue | |
2609 | * | |
2610 | * Return: %false if @rwork was already pending, %true otherwise. Note | |
2611 | * that a full RCU grace period is guaranteed only after a %true return. | |
bf393fd4 | 2612 | * While @rwork is guaranteed to be executed after a %false return, the |
05f0fe6b TH |
2613 | * execution may happen before a full RCU grace period has passed. |
2614 | */ | |
2615 | bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork) | |
2616 | { | |
2617 | struct work_struct *work = &rwork->work; | |
2618 | ||
86898fa6 TH |
2619 | /* |
2620 | * rcu_work can't be canceled or disabled. Warn if the user reached | |
2621 | * inside @rwork and disabled the inner work. | |
2622 | */ | |
2623 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)) && | |
2624 | !WARN_ON_ONCE(clear_pending_if_disabled(work))) { | |
05f0fe6b | 2625 | rwork->wq = wq; |
a7e30c0e | 2626 | call_rcu_hurry(&rwork->rcu, rcu_work_rcufn); |
05f0fe6b TH |
2627 | return true; |
2628 | } | |
2629 | ||
2630 | return false; | |
2631 | } | |
2632 | EXPORT_SYMBOL(queue_rcu_work); | |
2633 | ||
f7537df5 | 2634 | static struct worker *alloc_worker(int node) |
c34056a3 TH |
2635 | { |
2636 | struct worker *worker; | |
2637 | ||
f7537df5 | 2638 | worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node); |
c8e55f36 TH |
2639 | if (worker) { |
2640 | INIT_LIST_HEAD(&worker->entry); | |
affee4b2 | 2641 | INIT_LIST_HEAD(&worker->scheduled); |
da028469 | 2642 | INIT_LIST_HEAD(&worker->node); |
e22bee78 TH |
2643 | /* on creation a worker is in !idle && prep state */ |
2644 | worker->flags = WORKER_PREP; | |
c8e55f36 | 2645 | } |
c34056a3 TH |
2646 | return worker; |
2647 | } | |
2648 | ||
9546b29e TH |
2649 | static cpumask_t *pool_allowed_cpus(struct worker_pool *pool) |
2650 | { | |
8639eceb TH |
2651 | if (pool->cpu < 0 && pool->attrs->affn_strict) |
2652 | return pool->attrs->__pod_cpumask; | |
2653 | else | |
2654 | return pool->attrs->cpumask; | |
9546b29e TH |
2655 | } |
2656 | ||
4736cbf7 LJ |
2657 | /** |
2658 | * worker_attach_to_pool() - attach a worker to a pool | |
2659 | * @worker: worker to be attached | |
2660 | * @pool: the target pool | |
2661 | * | |
2662 | * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and | |
2663 | * cpu-binding of @worker are kept coordinated with the pool across | |
2664 | * cpu-[un]hotplugs. | |
2665 | */ | |
2666 | static void worker_attach_to_pool(struct worker *worker, | |
4cb1ef64 | 2667 | struct worker_pool *pool) |
4736cbf7 | 2668 | { |
1258fae7 | 2669 | mutex_lock(&wq_pool_attach_mutex); |
4736cbf7 | 2670 | |
4736cbf7 | 2671 | /* |
4cb1ef64 TH |
2672 | * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains stable |
2673 | * across this function. See the comments above the flag definition for | |
2674 | * details. BH workers are, while per-CPU, always DISASSOCIATED. | |
4736cbf7 | 2675 | */ |
4cb1ef64 | 2676 | if (pool->flags & POOL_DISASSOCIATED) { |
4736cbf7 | 2677 | worker->flags |= WORKER_UNBOUND; |
4cb1ef64 TH |
2678 | } else { |
2679 | WARN_ON_ONCE(pool->flags & POOL_BH); | |
5c25b5ff | 2680 | kthread_set_per_cpu(worker->task, pool->cpu); |
4cb1ef64 | 2681 | } |
4736cbf7 | 2682 | |
640f17c8 | 2683 | if (worker->rescue_wq) |
9546b29e | 2684 | set_cpus_allowed_ptr(worker->task, pool_allowed_cpus(pool)); |
640f17c8 | 2685 | |
4736cbf7 | 2686 | list_add_tail(&worker->node, &pool->workers); |
a2d812a2 | 2687 | worker->pool = pool; |
4736cbf7 | 2688 | |
1258fae7 | 2689 | mutex_unlock(&wq_pool_attach_mutex); |
4736cbf7 LJ |
2690 | } |
2691 | ||
f45b1c3c LJ |
2692 | static void unbind_worker(struct worker *worker) |
2693 | { | |
2694 | lockdep_assert_held(&wq_pool_attach_mutex); | |
2695 | ||
2696 | kthread_set_per_cpu(worker->task, -1); | |
2697 | if (cpumask_intersects(wq_unbound_cpumask, cpu_active_mask)) | |
2698 | WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, wq_unbound_cpumask) < 0); | |
2699 | else | |
2700 | WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0); | |
2701 | } | |
2702 | ||
f4b7b53c LJ |
2703 | |
2704 | static void detach_worker(struct worker *worker) | |
2705 | { | |
2706 | lockdep_assert_held(&wq_pool_attach_mutex); | |
2707 | ||
2708 | unbind_worker(worker); | |
2709 | list_del(&worker->node); | |
f4b7b53c LJ |
2710 | } |
2711 | ||
60f5a4bc LJ |
2712 | /** |
2713 | * worker_detach_from_pool() - detach a worker from its pool | |
2714 | * @worker: worker which is attached to its pool | |
60f5a4bc | 2715 | * |
4736cbf7 LJ |
2716 | * Undo the attaching which had been done in worker_attach_to_pool(). The |
2717 | * caller worker shouldn't access to the pool after detached except it has | |
2718 | * other reference to the pool. | |
60f5a4bc | 2719 | */ |
a2d812a2 | 2720 | static void worker_detach_from_pool(struct worker *worker) |
60f5a4bc | 2721 | { |
a2d812a2 | 2722 | struct worker_pool *pool = worker->pool; |
60f5a4bc | 2723 | |
4cb1ef64 TH |
2724 | /* there is one permanent BH worker per CPU which should never detach */ |
2725 | WARN_ON_ONCE(pool->flags & POOL_BH); | |
2726 | ||
1258fae7 | 2727 | mutex_lock(&wq_pool_attach_mutex); |
f4b7b53c | 2728 | detach_worker(worker); |
73613840 | 2729 | worker->pool = NULL; |
1258fae7 | 2730 | mutex_unlock(&wq_pool_attach_mutex); |
60f5a4bc | 2731 | |
b62c0751 LJ |
2732 | /* clear leftover flags without pool->lock after it is detached */ |
2733 | worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND); | |
60f5a4bc LJ |
2734 | } |
2735 | ||
2a1b02bc TH |
2736 | static int format_worker_id(char *buf, size_t size, struct worker *worker, |
2737 | struct worker_pool *pool) | |
2738 | { | |
2739 | if (worker->rescue_wq) | |
2740 | return scnprintf(buf, size, "kworker/R-%s", | |
2741 | worker->rescue_wq->name); | |
2742 | ||
2743 | if (pool) { | |
2744 | if (pool->cpu >= 0) | |
2745 | return scnprintf(buf, size, "kworker/%d:%d%s", | |
2746 | pool->cpu, worker->id, | |
2747 | pool->attrs->nice < 0 ? "H" : ""); | |
2748 | else | |
2749 | return scnprintf(buf, size, "kworker/u%d:%d", | |
2750 | pool->id, worker->id); | |
2751 | } else { | |
2752 | return scnprintf(buf, size, "kworker/dying"); | |
2753 | } | |
2754 | } | |
2755 | ||
c34056a3 TH |
2756 | /** |
2757 | * create_worker - create a new workqueue worker | |
63d95a91 | 2758 | * @pool: pool the new worker will belong to |
c34056a3 | 2759 | * |
051e1850 | 2760 | * Create and start a new worker which is attached to @pool. |
c34056a3 TH |
2761 | * |
2762 | * CONTEXT: | |
2763 | * Might sleep. Does GFP_KERNEL allocations. | |
2764 | * | |
d185af30 | 2765 | * Return: |
c34056a3 TH |
2766 | * Pointer to the newly created worker. |
2767 | */ | |
bc2ae0f5 | 2768 | static struct worker *create_worker(struct worker_pool *pool) |
c34056a3 | 2769 | { |
e441b56f ZL |
2770 | struct worker *worker; |
2771 | int id; | |
c34056a3 | 2772 | |
7cda9aae | 2773 | /* ID is needed to determine kthread name */ |
e441b56f | 2774 | id = ida_alloc(&pool->worker_ida, GFP_KERNEL); |
3f0ea0b8 PM |
2775 | if (id < 0) { |
2776 | pr_err_once("workqueue: Failed to allocate a worker ID: %pe\n", | |
2777 | ERR_PTR(id)); | |
e441b56f | 2778 | return NULL; |
3f0ea0b8 | 2779 | } |
c34056a3 | 2780 | |
f7537df5 | 2781 | worker = alloc_worker(pool->node); |
3f0ea0b8 PM |
2782 | if (!worker) { |
2783 | pr_err_once("workqueue: Failed to allocate a worker\n"); | |
c34056a3 | 2784 | goto fail; |
3f0ea0b8 | 2785 | } |
c34056a3 | 2786 | |
c34056a3 TH |
2787 | worker->id = id; |
2788 | ||
4cb1ef64 | 2789 | if (!(pool->flags & POOL_BH)) { |
2a1b02bc | 2790 | char id_buf[WORKER_ID_LEN]; |
4cb1ef64 | 2791 | |
2a1b02bc | 2792 | format_worker_id(id_buf, sizeof(id_buf), worker, pool); |
4cb1ef64 | 2793 | worker->task = kthread_create_on_node(worker_thread, worker, |
2a1b02bc | 2794 | pool->node, "%s", id_buf); |
4cb1ef64 TH |
2795 | if (IS_ERR(worker->task)) { |
2796 | if (PTR_ERR(worker->task) == -EINTR) { | |
2a1b02bc | 2797 | pr_err("workqueue: Interrupted when creating a worker thread \"%s\"\n", |
4cb1ef64 TH |
2798 | id_buf); |
2799 | } else { | |
2800 | pr_err_once("workqueue: Failed to create a worker thread: %pe", | |
2801 | worker->task); | |
2802 | } | |
2803 | goto fail; | |
60f54038 | 2804 | } |
c34056a3 | 2805 | |
4cb1ef64 TH |
2806 | set_user_nice(worker->task, pool->attrs->nice); |
2807 | kthread_bind_mask(worker->task, pool_allowed_cpus(pool)); | |
2808 | } | |
91151228 | 2809 | |
da028469 | 2810 | /* successful, attach the worker to the pool */ |
4736cbf7 | 2811 | worker_attach_to_pool(worker, pool); |
822d8405 | 2812 | |
051e1850 | 2813 | /* start the newly created worker */ |
a9b8a985 | 2814 | raw_spin_lock_irq(&pool->lock); |
0219a352 | 2815 | |
051e1850 LJ |
2816 | worker->pool->nr_workers++; |
2817 | worker_enter_idle(worker); | |
0219a352 TH |
2818 | |
2819 | /* | |
2820 | * @worker is waiting on a completion in kthread() and will trigger hung | |
6a229b0e TH |
2821 | * check if not woken up soon. As kick_pool() is noop if @pool is empty, |
2822 | * wake it up explicitly. | |
0219a352 | 2823 | */ |
4cb1ef64 TH |
2824 | if (worker->task) |
2825 | wake_up_process(worker->task); | |
0219a352 | 2826 | |
a9b8a985 | 2827 | raw_spin_unlock_irq(&pool->lock); |
051e1850 | 2828 | |
c34056a3 | 2829 | return worker; |
822d8405 | 2830 | |
c34056a3 | 2831 | fail: |
e441b56f | 2832 | ida_free(&pool->worker_ida, id); |
c34056a3 TH |
2833 | kfree(worker); |
2834 | return NULL; | |
2835 | } | |
2836 | ||
f4b7b53c | 2837 | static void detach_dying_workers(struct list_head *cull_list) |
793777bc | 2838 | { |
68f83057 | 2839 | struct worker *worker; |
793777bc | 2840 | |
f4b7b53c LJ |
2841 | list_for_each_entry(worker, cull_list, entry) |
2842 | detach_worker(worker); | |
793777bc VS |
2843 | } |
2844 | ||
68f83057 | 2845 | static void reap_dying_workers(struct list_head *cull_list) |
e02b9312 VS |
2846 | { |
2847 | struct worker *worker, *tmp; | |
2848 | ||
2849 | list_for_each_entry_safe(worker, tmp, cull_list, entry) { | |
2850 | list_del_init(&worker->entry); | |
68f83057 LJ |
2851 | kthread_stop_put(worker->task); |
2852 | kfree(worker); | |
e02b9312 VS |
2853 | } |
2854 | } | |
2855 | ||
c34056a3 | 2856 | /** |
e02b9312 | 2857 | * set_worker_dying - Tag a worker for destruction |
c34056a3 | 2858 | * @worker: worker to be destroyed |
e02b9312 | 2859 | * @list: transfer worker away from its pool->idle_list and into list |
c34056a3 | 2860 | * |
e02b9312 VS |
2861 | * Tag @worker for destruction and adjust @pool stats accordingly. The worker |
2862 | * should be idle. | |
c8e55f36 TH |
2863 | * |
2864 | * CONTEXT: | |
a9b8a985 | 2865 | * raw_spin_lock_irq(pool->lock). |
c34056a3 | 2866 | */ |
e02b9312 | 2867 | static void set_worker_dying(struct worker *worker, struct list_head *list) |
c34056a3 | 2868 | { |
bd7bdd43 | 2869 | struct worker_pool *pool = worker->pool; |
c34056a3 | 2870 | |
cd549687 | 2871 | lockdep_assert_held(&pool->lock); |
e02b9312 | 2872 | lockdep_assert_held(&wq_pool_attach_mutex); |
cd549687 | 2873 | |
c34056a3 | 2874 | /* sanity check frenzy */ |
6183c009 | 2875 | if (WARN_ON(worker->current_work) || |
73eb7fe7 LJ |
2876 | WARN_ON(!list_empty(&worker->scheduled)) || |
2877 | WARN_ON(!(worker->flags & WORKER_IDLE))) | |
6183c009 | 2878 | return; |
c34056a3 | 2879 | |
73eb7fe7 LJ |
2880 | pool->nr_workers--; |
2881 | pool->nr_idle--; | |
5bdfff96 | 2882 | |
cb444766 | 2883 | worker->flags |= WORKER_DIE; |
e02b9312 VS |
2884 | |
2885 | list_move(&worker->entry, list); | |
68f83057 LJ |
2886 | |
2887 | /* get an extra task struct reference for later kthread_stop_put() */ | |
2888 | get_task_struct(worker->task); | |
c34056a3 TH |
2889 | } |
2890 | ||
3f959aa3 VS |
2891 | /** |
2892 | * idle_worker_timeout - check if some idle workers can now be deleted. | |
2893 | * @t: The pool's idle_timer that just expired | |
2894 | * | |
2895 | * The timer is armed in worker_enter_idle(). Note that it isn't disarmed in | |
2896 | * worker_leave_idle(), as a worker flicking between idle and active while its | |
2897 | * pool is at the too_many_workers() tipping point would cause too much timer | |
2898 | * housekeeping overhead. Since IDLE_WORKER_TIMEOUT is long enough, we just let | |
2899 | * it expire and re-evaluate things from there. | |
2900 | */ | |
32a6c723 | 2901 | static void idle_worker_timeout(struct timer_list *t) |
e22bee78 | 2902 | { |
32a6c723 | 2903 | struct worker_pool *pool = from_timer(pool, t, idle_timer); |
3f959aa3 VS |
2904 | bool do_cull = false; |
2905 | ||
2906 | if (work_pending(&pool->idle_cull_work)) | |
2907 | return; | |
e22bee78 | 2908 | |
a9b8a985 | 2909 | raw_spin_lock_irq(&pool->lock); |
e22bee78 | 2910 | |
3f959aa3 | 2911 | if (too_many_workers(pool)) { |
e22bee78 TH |
2912 | struct worker *worker; |
2913 | unsigned long expires; | |
2914 | ||
2915 | /* idle_list is kept in LIFO order, check the last one */ | |
d70f5d57 | 2916 | worker = list_last_entry(&pool->idle_list, struct worker, entry); |
3f959aa3 VS |
2917 | expires = worker->last_active + IDLE_WORKER_TIMEOUT; |
2918 | do_cull = !time_before(jiffies, expires); | |
2919 | ||
2920 | if (!do_cull) | |
2921 | mod_timer(&pool->idle_timer, expires); | |
2922 | } | |
2923 | raw_spin_unlock_irq(&pool->lock); | |
2924 | ||
2925 | if (do_cull) | |
2926 | queue_work(system_unbound_wq, &pool->idle_cull_work); | |
2927 | } | |
2928 | ||
2929 | /** | |
2930 | * idle_cull_fn - cull workers that have been idle for too long. | |
2931 | * @work: the pool's work for handling these idle workers | |
2932 | * | |
2933 | * This goes through a pool's idle workers and gets rid of those that have been | |
2934 | * idle for at least IDLE_WORKER_TIMEOUT seconds. | |
e02b9312 VS |
2935 | * |
2936 | * We don't want to disturb isolated CPUs because of a pcpu kworker being | |
2937 | * culled, so this also resets worker affinity. This requires a sleepable | |
2938 | * context, hence the split between timer callback and work item. | |
3f959aa3 VS |
2939 | */ |
2940 | static void idle_cull_fn(struct work_struct *work) | |
2941 | { | |
2942 | struct worker_pool *pool = container_of(work, struct worker_pool, idle_cull_work); | |
9680540c | 2943 | LIST_HEAD(cull_list); |
3f959aa3 | 2944 | |
e02b9312 VS |
2945 | /* |
2946 | * Grabbing wq_pool_attach_mutex here ensures an already-running worker | |
f4b7b53c LJ |
2947 | * cannot proceed beyong set_pf_worker() in its self-destruct path. |
2948 | * This is required as a previously-preempted worker could run after | |
2949 | * set_worker_dying() has happened but before detach_dying_workers() did. | |
e02b9312 VS |
2950 | */ |
2951 | mutex_lock(&wq_pool_attach_mutex); | |
3f959aa3 VS |
2952 | raw_spin_lock_irq(&pool->lock); |
2953 | ||
2954 | while (too_many_workers(pool)) { | |
2955 | struct worker *worker; | |
2956 | unsigned long expires; | |
2957 | ||
d70f5d57 | 2958 | worker = list_last_entry(&pool->idle_list, struct worker, entry); |
e22bee78 TH |
2959 | expires = worker->last_active + IDLE_WORKER_TIMEOUT; |
2960 | ||
3347fc9f | 2961 | if (time_before(jiffies, expires)) { |
63d95a91 | 2962 | mod_timer(&pool->idle_timer, expires); |
3347fc9f | 2963 | break; |
d5abe669 | 2964 | } |
3347fc9f | 2965 | |
e02b9312 | 2966 | set_worker_dying(worker, &cull_list); |
e22bee78 TH |
2967 | } |
2968 | ||
a9b8a985 | 2969 | raw_spin_unlock_irq(&pool->lock); |
f4b7b53c | 2970 | detach_dying_workers(&cull_list); |
e02b9312 | 2971 | mutex_unlock(&wq_pool_attach_mutex); |
68f83057 LJ |
2972 | |
2973 | reap_dying_workers(&cull_list); | |
e22bee78 | 2974 | } |
d5abe669 | 2975 | |
493a1724 | 2976 | static void send_mayday(struct work_struct *work) |
e22bee78 | 2977 | { |
112202d9 TH |
2978 | struct pool_workqueue *pwq = get_work_pwq(work); |
2979 | struct workqueue_struct *wq = pwq->wq; | |
493a1724 | 2980 | |
2e109a28 | 2981 | lockdep_assert_held(&wq_mayday_lock); |
e22bee78 | 2982 | |
493008a8 | 2983 | if (!wq->rescuer) |
493a1724 | 2984 | return; |
e22bee78 TH |
2985 | |
2986 | /* mayday mayday mayday */ | |
493a1724 | 2987 | if (list_empty(&pwq->mayday_node)) { |
77668c8b LJ |
2988 | /* |
2989 | * If @pwq is for an unbound wq, its base ref may be put at | |
2990 | * any time due to an attribute change. Pin @pwq until the | |
2991 | * rescuer is done with it. | |
2992 | */ | |
2993 | get_pwq(pwq); | |
493a1724 | 2994 | list_add_tail(&pwq->mayday_node, &wq->maydays); |
e22bee78 | 2995 | wake_up_process(wq->rescuer->task); |
725e8ec5 | 2996 | pwq->stats[PWQ_STAT_MAYDAY]++; |
493a1724 | 2997 | } |
e22bee78 TH |
2998 | } |
2999 | ||
32a6c723 | 3000 | static void pool_mayday_timeout(struct timer_list *t) |
e22bee78 | 3001 | { |
32a6c723 | 3002 | struct worker_pool *pool = from_timer(pool, t, mayday_timer); |
e22bee78 TH |
3003 | struct work_struct *work; |
3004 | ||
a9b8a985 SAS |
3005 | raw_spin_lock_irq(&pool->lock); |
3006 | raw_spin_lock(&wq_mayday_lock); /* for wq->maydays */ | |
e22bee78 | 3007 | |
63d95a91 | 3008 | if (need_to_create_worker(pool)) { |
e22bee78 TH |
3009 | /* |
3010 | * We've been trying to create a new worker but | |
3011 | * haven't been successful. We might be hitting an | |
3012 | * allocation deadlock. Send distress signals to | |
3013 | * rescuers. | |
3014 | */ | |
63d95a91 | 3015 | list_for_each_entry(work, &pool->worklist, entry) |
e22bee78 | 3016 | send_mayday(work); |
1da177e4 | 3017 | } |
e22bee78 | 3018 | |
a9b8a985 SAS |
3019 | raw_spin_unlock(&wq_mayday_lock); |
3020 | raw_spin_unlock_irq(&pool->lock); | |
e22bee78 | 3021 | |
63d95a91 | 3022 | mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); |
1da177e4 LT |
3023 | } |
3024 | ||
e22bee78 TH |
3025 | /** |
3026 | * maybe_create_worker - create a new worker if necessary | |
63d95a91 | 3027 | * @pool: pool to create a new worker for |
e22bee78 | 3028 | * |
63d95a91 | 3029 | * Create a new worker for @pool if necessary. @pool is guaranteed to |
e22bee78 TH |
3030 | * have at least one idle worker on return from this function. If |
3031 | * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is | |
63d95a91 | 3032 | * sent to all rescuers with works scheduled on @pool to resolve |
e22bee78 TH |
3033 | * possible allocation deadlock. |
3034 | * | |
c5aa87bb TH |
3035 | * On return, need_to_create_worker() is guaranteed to be %false and |
3036 | * may_start_working() %true. | |
e22bee78 TH |
3037 | * |
3038 | * LOCKING: | |
a9b8a985 | 3039 | * raw_spin_lock_irq(pool->lock) which may be released and regrabbed |
e22bee78 TH |
3040 | * multiple times. Does GFP_KERNEL allocations. Called only from |
3041 | * manager. | |
e22bee78 | 3042 | */ |
29187a9e | 3043 | static void maybe_create_worker(struct worker_pool *pool) |
d565ed63 TH |
3044 | __releases(&pool->lock) |
3045 | __acquires(&pool->lock) | |
1da177e4 | 3046 | { |
e22bee78 | 3047 | restart: |
a9b8a985 | 3048 | raw_spin_unlock_irq(&pool->lock); |
9f9c2364 | 3049 | |
e22bee78 | 3050 | /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ |
63d95a91 | 3051 | mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); |
e22bee78 TH |
3052 | |
3053 | while (true) { | |
051e1850 | 3054 | if (create_worker(pool) || !need_to_create_worker(pool)) |
e22bee78 | 3055 | break; |
1da177e4 | 3056 | |
e212f361 | 3057 | schedule_timeout_interruptible(CREATE_COOLDOWN); |
9f9c2364 | 3058 | |
63d95a91 | 3059 | if (!need_to_create_worker(pool)) |
e22bee78 TH |
3060 | break; |
3061 | } | |
3062 | ||
63d95a91 | 3063 | del_timer_sync(&pool->mayday_timer); |
a9b8a985 | 3064 | raw_spin_lock_irq(&pool->lock); |
051e1850 LJ |
3065 | /* |
3066 | * This is necessary even after a new worker was just successfully | |
3067 | * created as @pool->lock was dropped and the new worker might have | |
3068 | * already become busy. | |
3069 | */ | |
63d95a91 | 3070 | if (need_to_create_worker(pool)) |
e22bee78 | 3071 | goto restart; |
e22bee78 TH |
3072 | } |
3073 | ||
73f53c4a | 3074 | /** |
e22bee78 TH |
3075 | * manage_workers - manage worker pool |
3076 | * @worker: self | |
73f53c4a | 3077 | * |
706026c2 | 3078 | * Assume the manager role and manage the worker pool @worker belongs |
e22bee78 | 3079 | * to. At any given time, there can be only zero or one manager per |
706026c2 | 3080 | * pool. The exclusion is handled automatically by this function. |
e22bee78 TH |
3081 | * |
3082 | * The caller can safely start processing works on false return. On | |
3083 | * true return, it's guaranteed that need_to_create_worker() is false | |
3084 | * and may_start_working() is true. | |
73f53c4a TH |
3085 | * |
3086 | * CONTEXT: | |
a9b8a985 | 3087 | * raw_spin_lock_irq(pool->lock) which may be released and regrabbed |
e22bee78 TH |
3088 | * multiple times. Does GFP_KERNEL allocations. |
3089 | * | |
d185af30 | 3090 | * Return: |
29187a9e TH |
3091 | * %false if the pool doesn't need management and the caller can safely |
3092 | * start processing works, %true if management function was performed and | |
3093 | * the conditions that the caller verified before calling the function may | |
3094 | * no longer be true. | |
73f53c4a | 3095 | */ |
e22bee78 | 3096 | static bool manage_workers(struct worker *worker) |
73f53c4a | 3097 | { |
63d95a91 | 3098 | struct worker_pool *pool = worker->pool; |
73f53c4a | 3099 | |
692b4825 | 3100 | if (pool->flags & POOL_MANAGER_ACTIVE) |
29187a9e | 3101 | return false; |
692b4825 TH |
3102 | |
3103 | pool->flags |= POOL_MANAGER_ACTIVE; | |
2607d7a6 | 3104 | pool->manager = worker; |
1e19ffc6 | 3105 | |
29187a9e | 3106 | maybe_create_worker(pool); |
e22bee78 | 3107 | |
2607d7a6 | 3108 | pool->manager = NULL; |
692b4825 | 3109 | pool->flags &= ~POOL_MANAGER_ACTIVE; |
d8bb65ab | 3110 | rcuwait_wake_up(&manager_wait); |
29187a9e | 3111 | return true; |
73f53c4a TH |
3112 | } |
3113 | ||
a62428c0 TH |
3114 | /** |
3115 | * process_one_work - process single work | |
c34056a3 | 3116 | * @worker: self |
a62428c0 TH |
3117 | * @work: work to process |
3118 | * | |
3119 | * Process @work. This function contains all the logics necessary to | |
3120 | * process a single work including synchronization against and | |
3121 | * interaction with other workers on the same cpu, queueing and | |
3122 | * flushing. As long as context requirement is met, any worker can | |
3123 | * call this function to process a work. | |
3124 | * | |
3125 | * CONTEXT: | |
a9b8a985 | 3126 | * raw_spin_lock_irq(pool->lock) which is released and regrabbed. |
a62428c0 | 3127 | */ |
c34056a3 | 3128 | static void process_one_work(struct worker *worker, struct work_struct *work) |
d565ed63 TH |
3129 | __releases(&pool->lock) |
3130 | __acquires(&pool->lock) | |
a62428c0 | 3131 | { |
112202d9 | 3132 | struct pool_workqueue *pwq = get_work_pwq(work); |
bd7bdd43 | 3133 | struct worker_pool *pool = worker->pool; |
c4560c2c | 3134 | unsigned long work_data; |
c35aea39 | 3135 | int lockdep_start_depth, rcu_start_depth; |
1acd92d9 | 3136 | bool bh_draining = pool->flags & POOL_BH_DRAINING; |
a62428c0 TH |
3137 | #ifdef CONFIG_LOCKDEP |
3138 | /* | |
3139 | * It is permissible to free the struct work_struct from | |
3140 | * inside the function that is called from it, this we need to | |
3141 | * take into account for lockdep too. To avoid bogus "held | |
3142 | * lock freed" warnings as well as problems when looking into | |
3143 | * work->lockdep_map, make a copy and use that here. | |
3144 | */ | |
4d82a1de PZ |
3145 | struct lockdep_map lockdep_map; |
3146 | ||
3147 | lockdep_copy_map(&lockdep_map, &work->lockdep_map); | |
a62428c0 | 3148 | #endif |
807407c0 | 3149 | /* ensure we're on the correct CPU */ |
85327af6 | 3150 | WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && |
ec22ca5e | 3151 | raw_smp_processor_id() != pool->cpu); |
25511a47 | 3152 | |
8930caba | 3153 | /* claim and dequeue */ |
a62428c0 | 3154 | debug_work_deactivate(work); |
c9e7cf27 | 3155 | hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); |
c34056a3 | 3156 | worker->current_work = work; |
a2c1c57b | 3157 | worker->current_func = work->func; |
112202d9 | 3158 | worker->current_pwq = pwq; |
4cb1ef64 TH |
3159 | if (worker->task) |
3160 | worker->current_at = worker->task->se.sum_exec_runtime; | |
c4560c2c | 3161 | work_data = *work_data_bits(work); |
d812796e | 3162 | worker->current_color = get_work_color(work_data); |
7a22ad75 | 3163 | |
8bf89593 TH |
3164 | /* |
3165 | * Record wq name for cmdline and debug reporting, may get | |
3166 | * overridden through set_worker_desc(). | |
3167 | */ | |
3168 | strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN); | |
3169 | ||
a62428c0 TH |
3170 | list_del_init(&work->entry); |
3171 | ||
fb0e7beb | 3172 | /* |
228f1d00 LJ |
3173 | * CPU intensive works don't participate in concurrency management. |
3174 | * They're the scheduler's responsibility. This takes @worker out | |
3175 | * of concurrency management and the next code block will chain | |
3176 | * execution of the pending work items. | |
fb0e7beb | 3177 | */ |
616db877 | 3178 | if (unlikely(pwq->wq->flags & WQ_CPU_INTENSIVE)) |
228f1d00 | 3179 | worker_set_flags(worker, WORKER_CPU_INTENSIVE); |
fb0e7beb | 3180 | |
974271c4 | 3181 | /* |
0219a352 TH |
3182 | * Kick @pool if necessary. It's always noop for per-cpu worker pools |
3183 | * since nr_running would always be >= 1 at this point. This is used to | |
3184 | * chain execution of the pending work items for WORKER_NOT_RUNNING | |
3185 | * workers such as the UNBOUND and CPU_INTENSIVE ones. | |
974271c4 | 3186 | */ |
0219a352 | 3187 | kick_pool(pool); |
974271c4 | 3188 | |
8930caba | 3189 | /* |
7c3eed5c | 3190 | * Record the last pool and clear PENDING which should be the last |
d565ed63 | 3191 | * update to @work. Also, do this inside @pool->lock so that |
23657bb1 TH |
3192 | * PENDING and queued state changes happen together while IRQ is |
3193 | * disabled. | |
8930caba | 3194 | */ |
456a78ee | 3195 | set_work_pool_and_clear_pending(work, pool->id, pool_offq_flags(pool)); |
a62428c0 | 3196 | |
fe48ba7d | 3197 | pwq->stats[PWQ_STAT_STARTED]++; |
a9b8a985 | 3198 | raw_spin_unlock_irq(&pool->lock); |
a62428c0 | 3199 | |
c35aea39 TH |
3200 | rcu_start_depth = rcu_preempt_depth(); |
3201 | lockdep_start_depth = lockdep_depth(current); | |
1acd92d9 TH |
3202 | /* see drain_dead_softirq_workfn() */ |
3203 | if (!bh_draining) | |
4f022f43 | 3204 | lock_map_acquire(pwq->wq->lockdep_map); |
a62428c0 | 3205 | lock_map_acquire(&lockdep_map); |
e6f3faa7 | 3206 | /* |
f52be570 PZ |
3207 | * Strictly speaking we should mark the invariant state without holding |
3208 | * any locks, that is, before these two lock_map_acquire()'s. | |
e6f3faa7 PZ |
3209 | * |
3210 | * However, that would result in: | |
3211 | * | |
3212 | * A(W1) | |
3213 | * WFC(C) | |
3214 | * A(W1) | |
3215 | * C(C) | |
3216 | * | |
3217 | * Which would create W1->C->W1 dependencies, even though there is no | |
3218 | * actual deadlock possible. There are two solutions, using a | |
3219 | * read-recursive acquire on the work(queue) 'locks', but this will then | |
f52be570 | 3220 | * hit the lockdep limitation on recursive locks, or simply discard |
e6f3faa7 PZ |
3221 | * these locks. |
3222 | * | |
3223 | * AFAICT there is no possible deadlock scenario between the | |
3224 | * flush_work() and complete() primitives (except for single-threaded | |
3225 | * workqueues), so hiding them isn't a problem. | |
3226 | */ | |
f52be570 | 3227 | lockdep_invariant_state(true); |
e36c886a | 3228 | trace_workqueue_execute_start(work); |
a2c1c57b | 3229 | worker->current_func(work); |
e36c886a AV |
3230 | /* |
3231 | * While we must be careful to not use "work" after this, the trace | |
3232 | * point will only record its address. | |
3233 | */ | |
1c5da0ec | 3234 | trace_workqueue_execute_end(work, worker->current_func); |
725e8ec5 | 3235 | pwq->stats[PWQ_STAT_COMPLETED]++; |
a62428c0 | 3236 | lock_map_release(&lockdep_map); |
1acd92d9 | 3237 | if (!bh_draining) |
4f022f43 | 3238 | lock_map_release(pwq->wq->lockdep_map); |
a62428c0 | 3239 | |
c35aea39 TH |
3240 | if (unlikely((worker->task && in_atomic()) || |
3241 | lockdep_depth(current) != lockdep_start_depth || | |
3242 | rcu_preempt_depth() != rcu_start_depth)) { | |
3243 | pr_err("BUG: workqueue leaked atomic, lock or RCU: %s[%d]\n" | |
3244 | " preempt=0x%08x lock=%d->%d RCU=%d->%d workfn=%ps\n", | |
3245 | current->comm, task_pid_nr(current), preempt_count(), | |
3246 | lockdep_start_depth, lockdep_depth(current), | |
3247 | rcu_start_depth, rcu_preempt_depth(), | |
3248 | worker->current_func); | |
a62428c0 TH |
3249 | debug_show_held_locks(current); |
3250 | dump_stack(); | |
3251 | } | |
3252 | ||
b22ce278 | 3253 | /* |
025f50f3 | 3254 | * The following prevents a kworker from hogging CPU on !PREEMPTION |
b22ce278 TH |
3255 | * kernels, where a requeueing work item waiting for something to |
3256 | * happen could deadlock with stop_machine as such work item could | |
3257 | * indefinitely requeue itself while all other CPUs are trapped in | |
789cbbec JL |
3258 | * stop_machine. At the same time, report a quiescent RCU state so |
3259 | * the same condition doesn't freeze RCU. | |
b22ce278 | 3260 | */ |
4cb1ef64 TH |
3261 | if (worker->task) |
3262 | cond_resched(); | |
b22ce278 | 3263 | |
a9b8a985 | 3264 | raw_spin_lock_irq(&pool->lock); |
a62428c0 | 3265 | |
616db877 TH |
3266 | /* |
3267 | * In addition to %WQ_CPU_INTENSIVE, @worker may also have been marked | |
3268 | * CPU intensive by wq_worker_tick() if @work hogged CPU longer than | |
3269 | * wq_cpu_intensive_thresh_us. Clear it. | |
3270 | */ | |
3271 | worker_clr_flags(worker, WORKER_CPU_INTENSIVE); | |
fb0e7beb | 3272 | |
1b69ac6b JW |
3273 | /* tag the worker for identification in schedule() */ |
3274 | worker->last_func = worker->current_func; | |
3275 | ||
a62428c0 | 3276 | /* we're done with it, release */ |
42f8570f | 3277 | hash_del(&worker->hentry); |
c34056a3 | 3278 | worker->current_work = NULL; |
a2c1c57b | 3279 | worker->current_func = NULL; |
112202d9 | 3280 | worker->current_pwq = NULL; |
d812796e | 3281 | worker->current_color = INT_MAX; |
dd6c3c54 TH |
3282 | |
3283 | /* must be the last step, see the function comment */ | |
c4560c2c | 3284 | pwq_dec_nr_in_flight(pwq, work_data); |
a62428c0 TH |
3285 | } |
3286 | ||
affee4b2 TH |
3287 | /** |
3288 | * process_scheduled_works - process scheduled works | |
3289 | * @worker: self | |
3290 | * | |
3291 | * Process all scheduled works. Please note that the scheduled list | |
3292 | * may change while processing a work, so this function repeatedly | |
3293 | * fetches a work from the top and executes it. | |
3294 | * | |
3295 | * CONTEXT: | |
a9b8a985 | 3296 | * raw_spin_lock_irq(pool->lock) which may be released and regrabbed |
affee4b2 TH |
3297 | * multiple times. |
3298 | */ | |
3299 | static void process_scheduled_works(struct worker *worker) | |
1da177e4 | 3300 | { |
c0ab017d TH |
3301 | struct work_struct *work; |
3302 | bool first = true; | |
3303 | ||
3304 | while ((work = list_first_entry_or_null(&worker->scheduled, | |
3305 | struct work_struct, entry))) { | |
3306 | if (first) { | |
3307 | worker->pool->watchdog_ts = jiffies; | |
3308 | first = false; | |
3309 | } | |
c34056a3 | 3310 | process_one_work(worker, work); |
1da177e4 | 3311 | } |
1da177e4 LT |
3312 | } |
3313 | ||
197f6acc TH |
3314 | static void set_pf_worker(bool val) |
3315 | { | |
3316 | mutex_lock(&wq_pool_attach_mutex); | |
3317 | if (val) | |
3318 | current->flags |= PF_WQ_WORKER; | |
3319 | else | |
3320 | current->flags &= ~PF_WQ_WORKER; | |
3321 | mutex_unlock(&wq_pool_attach_mutex); | |
3322 | } | |
3323 | ||
4690c4ab TH |
3324 | /** |
3325 | * worker_thread - the worker thread function | |
c34056a3 | 3326 | * @__worker: self |
4690c4ab | 3327 | * |
c5aa87bb TH |
3328 | * The worker thread function. All workers belong to a worker_pool - |
3329 | * either a per-cpu one or dynamic unbound one. These workers process all | |
3330 | * work items regardless of their specific target workqueue. The only | |
3331 | * exception is work items which belong to workqueues with a rescuer which | |
3332 | * will be explained in rescuer_thread(). | |
d185af30 YB |
3333 | * |
3334 | * Return: 0 | |
4690c4ab | 3335 | */ |
c34056a3 | 3336 | static int worker_thread(void *__worker) |
1da177e4 | 3337 | { |
c34056a3 | 3338 | struct worker *worker = __worker; |
bd7bdd43 | 3339 | struct worker_pool *pool = worker->pool; |
1da177e4 | 3340 | |
e22bee78 | 3341 | /* tell the scheduler that this is a workqueue worker */ |
197f6acc | 3342 | set_pf_worker(true); |
c8e55f36 | 3343 | woke_up: |
a9b8a985 | 3344 | raw_spin_lock_irq(&pool->lock); |
1da177e4 | 3345 | |
a9ab775b TH |
3346 | /* am I supposed to die? */ |
3347 | if (unlikely(worker->flags & WORKER_DIE)) { | |
a9b8a985 | 3348 | raw_spin_unlock_irq(&pool->lock); |
197f6acc | 3349 | set_pf_worker(false); |
73613840 LJ |
3350 | /* |
3351 | * The worker is dead and PF_WQ_WORKER is cleared, worker->pool | |
3352 | * shouldn't be accessed, reset it to NULL in case otherwise. | |
3353 | */ | |
3354 | worker->pool = NULL; | |
e441b56f | 3355 | ida_free(&pool->worker_ida, worker->id); |
a9ab775b | 3356 | return 0; |
c8e55f36 | 3357 | } |
affee4b2 | 3358 | |
c8e55f36 | 3359 | worker_leave_idle(worker); |
db7bccf4 | 3360 | recheck: |
e22bee78 | 3361 | /* no more worker necessary? */ |
63d95a91 | 3362 | if (!need_more_worker(pool)) |
e22bee78 TH |
3363 | goto sleep; |
3364 | ||
3365 | /* do we need to manage? */ | |
63d95a91 | 3366 | if (unlikely(!may_start_working(pool)) && manage_workers(worker)) |
e22bee78 TH |
3367 | goto recheck; |
3368 | ||
c8e55f36 TH |
3369 | /* |
3370 | * ->scheduled list can only be filled while a worker is | |
3371 | * preparing to process a work or actually processing it. | |
3372 | * Make sure nobody diddled with it while I was sleeping. | |
3373 | */ | |
6183c009 | 3374 | WARN_ON_ONCE(!list_empty(&worker->scheduled)); |
c8e55f36 | 3375 | |
e22bee78 | 3376 | /* |
a9ab775b TH |
3377 | * Finish PREP stage. We're guaranteed to have at least one idle |
3378 | * worker or that someone else has already assumed the manager | |
3379 | * role. This is where @worker starts participating in concurrency | |
3380 | * management if applicable and concurrency management is restored | |
3381 | * after being rebound. See rebind_workers() for details. | |
e22bee78 | 3382 | */ |
a9ab775b | 3383 | worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); |
e22bee78 TH |
3384 | |
3385 | do { | |
c8e55f36 | 3386 | struct work_struct *work = |
bd7bdd43 | 3387 | list_first_entry(&pool->worklist, |
c8e55f36 TH |
3388 | struct work_struct, entry); |
3389 | ||
873eaca6 TH |
3390 | if (assign_work(work, worker, NULL)) |
3391 | process_scheduled_works(worker); | |
63d95a91 | 3392 | } while (keep_working(pool)); |
e22bee78 | 3393 | |
228f1d00 | 3394 | worker_set_flags(worker, WORKER_PREP); |
d313dd85 | 3395 | sleep: |
c8e55f36 | 3396 | /* |
d565ed63 TH |
3397 | * pool->lock is held and there's no work to process and no need to |
3398 | * manage, sleep. Workers are woken up only while holding | |
3399 | * pool->lock or from local cpu, so setting the current state | |
3400 | * before releasing pool->lock is enough to prevent losing any | |
3401 | * event. | |
c8e55f36 TH |
3402 | */ |
3403 | worker_enter_idle(worker); | |
c5a94a61 | 3404 | __set_current_state(TASK_IDLE); |
a9b8a985 | 3405 | raw_spin_unlock_irq(&pool->lock); |
c8e55f36 TH |
3406 | schedule(); |
3407 | goto woke_up; | |
1da177e4 LT |
3408 | } |
3409 | ||
e22bee78 TH |
3410 | /** |
3411 | * rescuer_thread - the rescuer thread function | |
111c225a | 3412 | * @__rescuer: self |
e22bee78 TH |
3413 | * |
3414 | * Workqueue rescuer thread function. There's one rescuer for each | |
493008a8 | 3415 | * workqueue which has WQ_MEM_RECLAIM set. |
e22bee78 | 3416 | * |
706026c2 | 3417 | * Regular work processing on a pool may block trying to create a new |
e22bee78 TH |
3418 | * worker which uses GFP_KERNEL allocation which has slight chance of |
3419 | * developing into deadlock if some works currently on the same queue | |
3420 | * need to be processed to satisfy the GFP_KERNEL allocation. This is | |
3421 | * the problem rescuer solves. | |
3422 | * | |
706026c2 TH |
3423 | * When such condition is possible, the pool summons rescuers of all |
3424 | * workqueues which have works queued on the pool and let them process | |
e22bee78 TH |
3425 | * those works so that forward progress can be guaranteed. |
3426 | * | |
3427 | * This should happen rarely. | |
d185af30 YB |
3428 | * |
3429 | * Return: 0 | |
e22bee78 | 3430 | */ |
111c225a | 3431 | static int rescuer_thread(void *__rescuer) |
e22bee78 | 3432 | { |
111c225a TH |
3433 | struct worker *rescuer = __rescuer; |
3434 | struct workqueue_struct *wq = rescuer->rescue_wq; | |
4d595b86 | 3435 | bool should_stop; |
e22bee78 TH |
3436 | |
3437 | set_user_nice(current, RESCUER_NICE_LEVEL); | |
111c225a TH |
3438 | |
3439 | /* | |
3440 | * Mark rescuer as worker too. As WORKER_PREP is never cleared, it | |
3441 | * doesn't participate in concurrency management. | |
3442 | */ | |
197f6acc | 3443 | set_pf_worker(true); |
e22bee78 | 3444 | repeat: |
c5a94a61 | 3445 | set_current_state(TASK_IDLE); |
e22bee78 | 3446 | |
4d595b86 LJ |
3447 | /* |
3448 | * By the time the rescuer is requested to stop, the workqueue | |
3449 | * shouldn't have any work pending, but @wq->maydays may still have | |
3450 | * pwq(s) queued. This can happen by non-rescuer workers consuming | |
3451 | * all the work items before the rescuer got to them. Go through | |
3452 | * @wq->maydays processing before acting on should_stop so that the | |
3453 | * list is always empty on exit. | |
3454 | */ | |
3455 | should_stop = kthread_should_stop(); | |
e22bee78 | 3456 | |
493a1724 | 3457 | /* see whether any pwq is asking for help */ |
a9b8a985 | 3458 | raw_spin_lock_irq(&wq_mayday_lock); |
493a1724 TH |
3459 | |
3460 | while (!list_empty(&wq->maydays)) { | |
3461 | struct pool_workqueue *pwq = list_first_entry(&wq->maydays, | |
3462 | struct pool_workqueue, mayday_node); | |
112202d9 | 3463 | struct worker_pool *pool = pwq->pool; |
e22bee78 TH |
3464 | struct work_struct *work, *n; |
3465 | ||
3466 | __set_current_state(TASK_RUNNING); | |
493a1724 TH |
3467 | list_del_init(&pwq->mayday_node); |
3468 | ||
a9b8a985 | 3469 | raw_spin_unlock_irq(&wq_mayday_lock); |
e22bee78 | 3470 | |
51697d39 LJ |
3471 | worker_attach_to_pool(rescuer, pool); |
3472 | ||
a9b8a985 | 3473 | raw_spin_lock_irq(&pool->lock); |
e22bee78 TH |
3474 | |
3475 | /* | |
3476 | * Slurp in all works issued via this workqueue and | |
3477 | * process'em. | |
3478 | */ | |
873eaca6 | 3479 | WARN_ON_ONCE(!list_empty(&rescuer->scheduled)); |
82607adc | 3480 | list_for_each_entry_safe(work, n, &pool->worklist, entry) { |
873eaca6 TH |
3481 | if (get_work_pwq(work) == pwq && |
3482 | assign_work(work, rescuer, &n)) | |
725e8ec5 | 3483 | pwq->stats[PWQ_STAT_RESCUED]++; |
82607adc | 3484 | } |
e22bee78 | 3485 | |
873eaca6 | 3486 | if (!list_empty(&rescuer->scheduled)) { |
008847f6 N |
3487 | process_scheduled_works(rescuer); |
3488 | ||
3489 | /* | |
3490 | * The above execution of rescued work items could | |
3491 | * have created more to rescue through | |
f97a4a1a | 3492 | * pwq_activate_first_inactive() or chained |
008847f6 N |
3493 | * queueing. Let's put @pwq back on mayday list so |
3494 | * that such back-to-back work items, which may be | |
3495 | * being used to relieve memory pressure, don't | |
3496 | * incur MAYDAY_INTERVAL delay inbetween. | |
3497 | */ | |
4f3f4cf3 | 3498 | if (pwq->nr_active && need_to_create_worker(pool)) { |
a9b8a985 | 3499 | raw_spin_lock(&wq_mayday_lock); |
e66b39af TH |
3500 | /* |
3501 | * Queue iff we aren't racing destruction | |
3502 | * and somebody else hasn't queued it already. | |
3503 | */ | |
3504 | if (wq->rescuer && list_empty(&pwq->mayday_node)) { | |
3505 | get_pwq(pwq); | |
3506 | list_add_tail(&pwq->mayday_node, &wq->maydays); | |
3507 | } | |
a9b8a985 | 3508 | raw_spin_unlock(&wq_mayday_lock); |
008847f6 N |
3509 | } |
3510 | } | |
7576958a | 3511 | |
77668c8b LJ |
3512 | /* |
3513 | * Put the reference grabbed by send_mayday(). @pool won't | |
13b1d625 | 3514 | * go away while we're still attached to it. |
77668c8b LJ |
3515 | */ |
3516 | put_pwq(pwq); | |
3517 | ||
7576958a | 3518 | /* |
0219a352 TH |
3519 | * Leave this pool. Notify regular workers; otherwise, we end up |
3520 | * with 0 concurrency and stalling the execution. | |
7576958a | 3521 | */ |
0219a352 | 3522 | kick_pool(pool); |
7576958a | 3523 | |
a9b8a985 | 3524 | raw_spin_unlock_irq(&pool->lock); |
13b1d625 | 3525 | |
a2d812a2 | 3526 | worker_detach_from_pool(rescuer); |
13b1d625 | 3527 | |
a9b8a985 | 3528 | raw_spin_lock_irq(&wq_mayday_lock); |
e22bee78 TH |
3529 | } |
3530 | ||
a9b8a985 | 3531 | raw_spin_unlock_irq(&wq_mayday_lock); |
493a1724 | 3532 | |
4d595b86 LJ |
3533 | if (should_stop) { |
3534 | __set_current_state(TASK_RUNNING); | |
197f6acc | 3535 | set_pf_worker(false); |
4d595b86 LJ |
3536 | return 0; |
3537 | } | |
3538 | ||
111c225a TH |
3539 | /* rescuers should never participate in concurrency management */ |
3540 | WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); | |
e22bee78 TH |
3541 | schedule(); |
3542 | goto repeat; | |
1da177e4 LT |
3543 | } |
3544 | ||
4cb1ef64 TH |
3545 | static void bh_worker(struct worker *worker) |
3546 | { | |
3547 | struct worker_pool *pool = worker->pool; | |
3548 | int nr_restarts = BH_WORKER_RESTARTS; | |
3549 | unsigned long end = jiffies + BH_WORKER_JIFFIES; | |
3550 | ||
3551 | raw_spin_lock_irq(&pool->lock); | |
3552 | worker_leave_idle(worker); | |
3553 | ||
3554 | /* | |
3555 | * This function follows the structure of worker_thread(). See there for | |
3556 | * explanations on each step. | |
3557 | */ | |
3558 | if (!need_more_worker(pool)) | |
3559 | goto done; | |
3560 | ||
3561 | WARN_ON_ONCE(!list_empty(&worker->scheduled)); | |
3562 | worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); | |
3563 | ||
3564 | do { | |
3565 | struct work_struct *work = | |
3566 | list_first_entry(&pool->worklist, | |
3567 | struct work_struct, entry); | |
3568 | ||
3569 | if (assign_work(work, worker, NULL)) | |
3570 | process_scheduled_works(worker); | |
3571 | } while (keep_working(pool) && | |
3572 | --nr_restarts && time_before(jiffies, end)); | |
3573 | ||
3574 | worker_set_flags(worker, WORKER_PREP); | |
3575 | done: | |
3576 | worker_enter_idle(worker); | |
3577 | kick_pool(pool); | |
3578 | raw_spin_unlock_irq(&pool->lock); | |
3579 | } | |
3580 | ||
3581 | /* | |
3582 | * TODO: Convert all tasklet users to workqueue and use softirq directly. | |
3583 | * | |
3584 | * This is currently called from tasklet[_hi]action() and thus is also called | |
3585 | * whenever there are tasklets to run. Let's do an early exit if there's nothing | |
3586 | * queued. Once conversion from tasklet is complete, the need_more_worker() test | |
3587 | * can be dropped. | |
3588 | * | |
3589 | * After full conversion, we'll add worker->softirq_action, directly use the | |
3590 | * softirq action and obtain the worker pointer from the softirq_action pointer. | |
3591 | */ | |
3592 | void workqueue_softirq_action(bool highpri) | |
3593 | { | |
3594 | struct worker_pool *pool = | |
3595 | &per_cpu(bh_worker_pools, smp_processor_id())[highpri]; | |
3596 | if (need_more_worker(pool)) | |
3597 | bh_worker(list_first_entry(&pool->workers, struct worker, node)); | |
3598 | } | |
3599 | ||
1acd92d9 TH |
3600 | struct wq_drain_dead_softirq_work { |
3601 | struct work_struct work; | |
3602 | struct worker_pool *pool; | |
3603 | struct completion done; | |
3604 | }; | |
3605 | ||
3606 | static void drain_dead_softirq_workfn(struct work_struct *work) | |
3607 | { | |
3608 | struct wq_drain_dead_softirq_work *dead_work = | |
3609 | container_of(work, struct wq_drain_dead_softirq_work, work); | |
3610 | struct worker_pool *pool = dead_work->pool; | |
3611 | bool repeat; | |
3612 | ||
3613 | /* | |
3614 | * @pool's CPU is dead and we want to execute its still pending work | |
3615 | * items from this BH work item which is running on a different CPU. As | |
3616 | * its CPU is dead, @pool can't be kicked and, as work execution path | |
3617 | * will be nested, a lockdep annotation needs to be suppressed. Mark | |
3618 | * @pool with %POOL_BH_DRAINING for the special treatments. | |
3619 | */ | |
3620 | raw_spin_lock_irq(&pool->lock); | |
3621 | pool->flags |= POOL_BH_DRAINING; | |
3622 | raw_spin_unlock_irq(&pool->lock); | |
3623 | ||
3624 | bh_worker(list_first_entry(&pool->workers, struct worker, node)); | |
3625 | ||
3626 | raw_spin_lock_irq(&pool->lock); | |
3627 | pool->flags &= ~POOL_BH_DRAINING; | |
3628 | repeat = need_more_worker(pool); | |
3629 | raw_spin_unlock_irq(&pool->lock); | |
3630 | ||
3631 | /* | |
3632 | * bh_worker() might hit consecutive execution limit and bail. If there | |
3633 | * still are pending work items, reschedule self and return so that we | |
3634 | * don't hog this CPU's BH. | |
3635 | */ | |
3636 | if (repeat) { | |
3637 | if (pool->attrs->nice == HIGHPRI_NICE_LEVEL) | |
3638 | queue_work(system_bh_highpri_wq, work); | |
3639 | else | |
3640 | queue_work(system_bh_wq, work); | |
3641 | } else { | |
3642 | complete(&dead_work->done); | |
3643 | } | |
3644 | } | |
3645 | ||
3646 | /* | |
3647 | * @cpu is dead. Drain the remaining BH work items on the current CPU. It's | |
3648 | * possible to allocate dead_work per CPU and avoid flushing. However, then we | |
3649 | * have to worry about draining overlapping with CPU coming back online or | |
3650 | * nesting (one CPU's dead_work queued on another CPU which is also dead and so | |
3651 | * on). Let's keep it simple and drain them synchronously. These are BH work | |
3652 | * items which shouldn't be requeued on the same pool. Shouldn't take long. | |
3653 | */ | |
3654 | void workqueue_softirq_dead(unsigned int cpu) | |
3655 | { | |
3656 | int i; | |
3657 | ||
3658 | for (i = 0; i < NR_STD_WORKER_POOLS; i++) { | |
3659 | struct worker_pool *pool = &per_cpu(bh_worker_pools, cpu)[i]; | |
3660 | struct wq_drain_dead_softirq_work dead_work; | |
3661 | ||
3662 | if (!need_more_worker(pool)) | |
3663 | continue; | |
3664 | ||
e7cc3be6 | 3665 | INIT_WORK_ONSTACK(&dead_work.work, drain_dead_softirq_workfn); |
1acd92d9 TH |
3666 | dead_work.pool = pool; |
3667 | init_completion(&dead_work.done); | |
3668 | ||
3669 | if (pool->attrs->nice == HIGHPRI_NICE_LEVEL) | |
3670 | queue_work(system_bh_highpri_wq, &dead_work.work); | |
3671 | else | |
3672 | queue_work(system_bh_wq, &dead_work.work); | |
3673 | ||
3674 | wait_for_completion(&dead_work.done); | |
31103f40 | 3675 | destroy_work_on_stack(&dead_work.work); |
1acd92d9 TH |
3676 | } |
3677 | } | |
3678 | ||
fca839c0 TH |
3679 | /** |
3680 | * check_flush_dependency - check for flush dependency sanity | |
3681 | * @target_wq: workqueue being flushed | |
3682 | * @target_work: work item being flushed (NULL for workqueue flushes) | |
3683 | * | |
3684 | * %current is trying to flush the whole @target_wq or @target_work on it. | |
3685 | * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not | |
3686 | * reclaiming memory or running on a workqueue which doesn't have | |
3687 | * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to | |
3688 | * a deadlock. | |
3689 | */ | |
3690 | static void check_flush_dependency(struct workqueue_struct *target_wq, | |
3691 | struct work_struct *target_work) | |
3692 | { | |
3693 | work_func_t target_func = target_work ? target_work->func : NULL; | |
3694 | struct worker *worker; | |
3695 | ||
3696 | if (target_wq->flags & WQ_MEM_RECLAIM) | |
3697 | return; | |
3698 | ||
3699 | worker = current_wq_worker(); | |
3700 | ||
3701 | WARN_ONCE(current->flags & PF_MEMALLOC, | |
d75f773c | 3702 | "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps", |
fca839c0 | 3703 | current->pid, current->comm, target_wq->name, target_func); |
23d11a58 TH |
3704 | WARN_ONCE(worker && ((worker->current_pwq->wq->flags & |
3705 | (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM), | |
d75f773c | 3706 | "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps", |
fca839c0 TH |
3707 | worker->current_pwq->wq->name, worker->current_func, |
3708 | target_wq->name, target_func); | |
3709 | } | |
3710 | ||
fc2e4d70 ON |
3711 | struct wq_barrier { |
3712 | struct work_struct work; | |
3713 | struct completion done; | |
2607d7a6 | 3714 | struct task_struct *task; /* purely informational */ |
fc2e4d70 ON |
3715 | }; |
3716 | ||
3717 | static void wq_barrier_func(struct work_struct *work) | |
3718 | { | |
3719 | struct wq_barrier *barr = container_of(work, struct wq_barrier, work); | |
3720 | complete(&barr->done); | |
3721 | } | |
3722 | ||
4690c4ab TH |
3723 | /** |
3724 | * insert_wq_barrier - insert a barrier work | |
112202d9 | 3725 | * @pwq: pwq to insert barrier into |
4690c4ab | 3726 | * @barr: wq_barrier to insert |
affee4b2 TH |
3727 | * @target: target work to attach @barr to |
3728 | * @worker: worker currently executing @target, NULL if @target is not executing | |
4690c4ab | 3729 | * |
affee4b2 TH |
3730 | * @barr is linked to @target such that @barr is completed only after |
3731 | * @target finishes execution. Please note that the ordering | |
3732 | * guarantee is observed only with respect to @target and on the local | |
3733 | * cpu. | |
3734 | * | |
3735 | * Currently, a queued barrier can't be canceled. This is because | |
3736 | * try_to_grab_pending() can't determine whether the work to be | |
3737 | * grabbed is at the head of the queue and thus can't clear LINKED | |
3738 | * flag of the previous work while there must be a valid next work | |
3739 | * after a work with LINKED flag set. | |
3740 | * | |
3741 | * Note that when @worker is non-NULL, @target may be modified | |
112202d9 | 3742 | * underneath us, so we can't reliably determine pwq from @target. |
4690c4ab TH |
3743 | * |
3744 | * CONTEXT: | |
a9b8a985 | 3745 | * raw_spin_lock_irq(pool->lock). |
4690c4ab | 3746 | */ |
112202d9 | 3747 | static void insert_wq_barrier(struct pool_workqueue *pwq, |
affee4b2 TH |
3748 | struct wq_barrier *barr, |
3749 | struct work_struct *target, struct worker *worker) | |
fc2e4d70 | 3750 | { |
4cb1ef64 | 3751 | static __maybe_unused struct lock_class_key bh_key, thr_key; |
d812796e LJ |
3752 | unsigned int work_flags = 0; |
3753 | unsigned int work_color; | |
affee4b2 | 3754 | struct list_head *head; |
affee4b2 | 3755 | |
dc186ad7 | 3756 | /* |
d565ed63 | 3757 | * debugobject calls are safe here even with pool->lock locked |
dc186ad7 TG |
3758 | * as we know for sure that this will not trigger any of the |
3759 | * checks and call back into the fixup functions where we | |
3760 | * might deadlock. | |
4cb1ef64 TH |
3761 | * |
3762 | * BH and threaded workqueues need separate lockdep keys to avoid | |
3763 | * spuriously triggering "inconsistent {SOFTIRQ-ON-W} -> {IN-SOFTIRQ-W} | |
3764 | * usage". | |
dc186ad7 | 3765 | */ |
4cb1ef64 TH |
3766 | INIT_WORK_ONSTACK_KEY(&barr->work, wq_barrier_func, |
3767 | (pwq->wq->flags & WQ_BH) ? &bh_key : &thr_key); | |
22df02bb | 3768 | __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); |
52fa5bc5 | 3769 | |
fd1a5b04 BP |
3770 | init_completion_map(&barr->done, &target->lockdep_map); |
3771 | ||
2607d7a6 | 3772 | barr->task = current; |
83c22520 | 3773 | |
5797b1c1 | 3774 | /* The barrier work item does not participate in nr_active. */ |
018f3a13 LJ |
3775 | work_flags |= WORK_STRUCT_INACTIVE; |
3776 | ||
affee4b2 TH |
3777 | /* |
3778 | * If @target is currently being executed, schedule the | |
3779 | * barrier to the worker; otherwise, put it after @target. | |
3780 | */ | |
d812796e | 3781 | if (worker) { |
affee4b2 | 3782 | head = worker->scheduled.next; |
d812796e LJ |
3783 | work_color = worker->current_color; |
3784 | } else { | |
affee4b2 TH |
3785 | unsigned long *bits = work_data_bits(target); |
3786 | ||
3787 | head = target->entry.next; | |
3788 | /* there can already be other linked works, inherit and set */ | |
d21cece0 | 3789 | work_flags |= *bits & WORK_STRUCT_LINKED; |
d812796e | 3790 | work_color = get_work_color(*bits); |
affee4b2 TH |
3791 | __set_bit(WORK_STRUCT_LINKED_BIT, bits); |
3792 | } | |
3793 | ||
d812796e LJ |
3794 | pwq->nr_in_flight[work_color]++; |
3795 | work_flags |= work_color_to_flags(work_color); | |
3796 | ||
d21cece0 | 3797 | insert_work(pwq, &barr->work, head, work_flags); |
fc2e4d70 ON |
3798 | } |
3799 | ||
73f53c4a | 3800 | /** |
112202d9 | 3801 | * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing |
73f53c4a TH |
3802 | * @wq: workqueue being flushed |
3803 | * @flush_color: new flush color, < 0 for no-op | |
3804 | * @work_color: new work color, < 0 for no-op | |
3805 | * | |
112202d9 | 3806 | * Prepare pwqs for workqueue flushing. |
73f53c4a | 3807 | * |
112202d9 TH |
3808 | * If @flush_color is non-negative, flush_color on all pwqs should be |
3809 | * -1. If no pwq has in-flight commands at the specified color, all | |
3810 | * pwq->flush_color's stay at -1 and %false is returned. If any pwq | |
3811 | * has in flight commands, its pwq->flush_color is set to | |
3812 | * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq | |
73f53c4a TH |
3813 | * wakeup logic is armed and %true is returned. |
3814 | * | |
3815 | * The caller should have initialized @wq->first_flusher prior to | |
3816 | * calling this function with non-negative @flush_color. If | |
3817 | * @flush_color is negative, no flush color update is done and %false | |
3818 | * is returned. | |
3819 | * | |
112202d9 | 3820 | * If @work_color is non-negative, all pwqs should have the same |
73f53c4a TH |
3821 | * work_color which is previous to @work_color and all will be |
3822 | * advanced to @work_color. | |
3823 | * | |
3824 | * CONTEXT: | |
3c25a55d | 3825 | * mutex_lock(wq->mutex). |
73f53c4a | 3826 | * |
d185af30 | 3827 | * Return: |
73f53c4a TH |
3828 | * %true if @flush_color >= 0 and there's something to flush. %false |
3829 | * otherwise. | |
3830 | */ | |
112202d9 | 3831 | static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, |
73f53c4a | 3832 | int flush_color, int work_color) |
1da177e4 | 3833 | { |
73f53c4a | 3834 | bool wait = false; |
49e3cf44 | 3835 | struct pool_workqueue *pwq; |
1da177e4 | 3836 | |
73f53c4a | 3837 | if (flush_color >= 0) { |
6183c009 | 3838 | WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); |
112202d9 | 3839 | atomic_set(&wq->nr_pwqs_to_flush, 1); |
1da177e4 | 3840 | } |
2355b70f | 3841 | |
49e3cf44 | 3842 | for_each_pwq(pwq, wq) { |
112202d9 | 3843 | struct worker_pool *pool = pwq->pool; |
fc2e4d70 | 3844 | |
a9b8a985 | 3845 | raw_spin_lock_irq(&pool->lock); |
83c22520 | 3846 | |
73f53c4a | 3847 | if (flush_color >= 0) { |
6183c009 | 3848 | WARN_ON_ONCE(pwq->flush_color != -1); |
fc2e4d70 | 3849 | |
112202d9 TH |
3850 | if (pwq->nr_in_flight[flush_color]) { |
3851 | pwq->flush_color = flush_color; | |
3852 | atomic_inc(&wq->nr_pwqs_to_flush); | |
73f53c4a TH |
3853 | wait = true; |
3854 | } | |
3855 | } | |
1da177e4 | 3856 | |
73f53c4a | 3857 | if (work_color >= 0) { |
6183c009 | 3858 | WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); |
112202d9 | 3859 | pwq->work_color = work_color; |
73f53c4a | 3860 | } |
1da177e4 | 3861 | |
a9b8a985 | 3862 | raw_spin_unlock_irq(&pool->lock); |
1da177e4 | 3863 | } |
2355b70f | 3864 | |
112202d9 | 3865 | if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) |
73f53c4a | 3866 | complete(&wq->first_flusher->done); |
14441960 | 3867 | |
73f53c4a | 3868 | return wait; |
1da177e4 LT |
3869 | } |
3870 | ||
c35aea39 TH |
3871 | static void touch_wq_lockdep_map(struct workqueue_struct *wq) |
3872 | { | |
4cb1ef64 | 3873 | #ifdef CONFIG_LOCKDEP |
84c425be SS |
3874 | if (unlikely(!wq->lockdep_map)) |
3875 | return; | |
3876 | ||
4cb1ef64 TH |
3877 | if (wq->flags & WQ_BH) |
3878 | local_bh_disable(); | |
3879 | ||
4f022f43 MB |
3880 | lock_map_acquire(wq->lockdep_map); |
3881 | lock_map_release(wq->lockdep_map); | |
4cb1ef64 TH |
3882 | |
3883 | if (wq->flags & WQ_BH) | |
3884 | local_bh_enable(); | |
3885 | #endif | |
c35aea39 TH |
3886 | } |
3887 | ||
3888 | static void touch_work_lockdep_map(struct work_struct *work, | |
3889 | struct workqueue_struct *wq) | |
3890 | { | |
4cb1ef64 TH |
3891 | #ifdef CONFIG_LOCKDEP |
3892 | if (wq->flags & WQ_BH) | |
3893 | local_bh_disable(); | |
3894 | ||
c35aea39 TH |
3895 | lock_map_acquire(&work->lockdep_map); |
3896 | lock_map_release(&work->lockdep_map); | |
4cb1ef64 TH |
3897 | |
3898 | if (wq->flags & WQ_BH) | |
3899 | local_bh_enable(); | |
3900 | #endif | |
c35aea39 TH |
3901 | } |
3902 | ||
0fcb78c2 | 3903 | /** |
c4f135d6 | 3904 | * __flush_workqueue - ensure that any scheduled work has run to completion. |
0fcb78c2 | 3905 | * @wq: workqueue to flush |
1da177e4 | 3906 | * |
c5aa87bb TH |
3907 | * This function sleeps until all work items which were queued on entry |
3908 | * have finished execution, but it is not livelocked by new incoming ones. | |
1da177e4 | 3909 | */ |
c4f135d6 | 3910 | void __flush_workqueue(struct workqueue_struct *wq) |
1da177e4 | 3911 | { |
73f53c4a TH |
3912 | struct wq_flusher this_flusher = { |
3913 | .list = LIST_HEAD_INIT(this_flusher.list), | |
3914 | .flush_color = -1, | |
4f022f43 | 3915 | .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, (*wq->lockdep_map)), |
73f53c4a TH |
3916 | }; |
3917 | int next_color; | |
1da177e4 | 3918 | |
3347fa09 TH |
3919 | if (WARN_ON(!wq_online)) |
3920 | return; | |
3921 | ||
c35aea39 | 3922 | touch_wq_lockdep_map(wq); |
87915adc | 3923 | |
3c25a55d | 3924 | mutex_lock(&wq->mutex); |
73f53c4a TH |
3925 | |
3926 | /* | |
3927 | * Start-to-wait phase | |
3928 | */ | |
3929 | next_color = work_next_color(wq->work_color); | |
3930 | ||
3931 | if (next_color != wq->flush_color) { | |
3932 | /* | |
3933 | * Color space is not full. The current work_color | |
3934 | * becomes our flush_color and work_color is advanced | |
3935 | * by one. | |
3936 | */ | |
6183c009 | 3937 | WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); |
73f53c4a TH |
3938 | this_flusher.flush_color = wq->work_color; |
3939 | wq->work_color = next_color; | |
3940 | ||
3941 | if (!wq->first_flusher) { | |
3942 | /* no flush in progress, become the first flusher */ | |
6183c009 | 3943 | WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); |
73f53c4a TH |
3944 | |
3945 | wq->first_flusher = &this_flusher; | |
3946 | ||
112202d9 | 3947 | if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, |
73f53c4a TH |
3948 | wq->work_color)) { |
3949 | /* nothing to flush, done */ | |
3950 | wq->flush_color = next_color; | |
3951 | wq->first_flusher = NULL; | |
3952 | goto out_unlock; | |
3953 | } | |
3954 | } else { | |
3955 | /* wait in queue */ | |
6183c009 | 3956 | WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); |
73f53c4a | 3957 | list_add_tail(&this_flusher.list, &wq->flusher_queue); |
112202d9 | 3958 | flush_workqueue_prep_pwqs(wq, -1, wq->work_color); |
73f53c4a TH |
3959 | } |
3960 | } else { | |
3961 | /* | |
3962 | * Oops, color space is full, wait on overflow queue. | |
3963 | * The next flush completion will assign us | |
3964 | * flush_color and transfer to flusher_queue. | |
3965 | */ | |
3966 | list_add_tail(&this_flusher.list, &wq->flusher_overflow); | |
3967 | } | |
3968 | ||
fca839c0 TH |
3969 | check_flush_dependency(wq, NULL); |
3970 | ||
3c25a55d | 3971 | mutex_unlock(&wq->mutex); |
73f53c4a TH |
3972 | |
3973 | wait_for_completion(&this_flusher.done); | |
3974 | ||
3975 | /* | |
3976 | * Wake-up-and-cascade phase | |
3977 | * | |
3978 | * First flushers are responsible for cascading flushes and | |
3979 | * handling overflow. Non-first flushers can simply return. | |
3980 | */ | |
00d5d15b | 3981 | if (READ_ONCE(wq->first_flusher) != &this_flusher) |
73f53c4a TH |
3982 | return; |
3983 | ||
3c25a55d | 3984 | mutex_lock(&wq->mutex); |
73f53c4a | 3985 | |
4ce48b37 TH |
3986 | /* we might have raced, check again with mutex held */ |
3987 | if (wq->first_flusher != &this_flusher) | |
3988 | goto out_unlock; | |
3989 | ||
00d5d15b | 3990 | WRITE_ONCE(wq->first_flusher, NULL); |
73f53c4a | 3991 | |
6183c009 TH |
3992 | WARN_ON_ONCE(!list_empty(&this_flusher.list)); |
3993 | WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); | |
73f53c4a TH |
3994 | |
3995 | while (true) { | |
3996 | struct wq_flusher *next, *tmp; | |
3997 | ||
3998 | /* complete all the flushers sharing the current flush color */ | |
3999 | list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { | |
4000 | if (next->flush_color != wq->flush_color) | |
4001 | break; | |
4002 | list_del_init(&next->list); | |
4003 | complete(&next->done); | |
4004 | } | |
4005 | ||
6183c009 TH |
4006 | WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && |
4007 | wq->flush_color != work_next_color(wq->work_color)); | |
73f53c4a TH |
4008 | |
4009 | /* this flush_color is finished, advance by one */ | |
4010 | wq->flush_color = work_next_color(wq->flush_color); | |
4011 | ||
4012 | /* one color has been freed, handle overflow queue */ | |
4013 | if (!list_empty(&wq->flusher_overflow)) { | |
4014 | /* | |
4015 | * Assign the same color to all overflowed | |
4016 | * flushers, advance work_color and append to | |
4017 | * flusher_queue. This is the start-to-wait | |
4018 | * phase for these overflowed flushers. | |
4019 | */ | |
4020 | list_for_each_entry(tmp, &wq->flusher_overflow, list) | |
4021 | tmp->flush_color = wq->work_color; | |
4022 | ||
4023 | wq->work_color = work_next_color(wq->work_color); | |
4024 | ||
4025 | list_splice_tail_init(&wq->flusher_overflow, | |
4026 | &wq->flusher_queue); | |
112202d9 | 4027 | flush_workqueue_prep_pwqs(wq, -1, wq->work_color); |
73f53c4a TH |
4028 | } |
4029 | ||
4030 | if (list_empty(&wq->flusher_queue)) { | |
6183c009 | 4031 | WARN_ON_ONCE(wq->flush_color != wq->work_color); |
73f53c4a TH |
4032 | break; |
4033 | } | |
4034 | ||
4035 | /* | |
4036 | * Need to flush more colors. Make the next flusher | |
112202d9 | 4037 | * the new first flusher and arm pwqs. |
73f53c4a | 4038 | */ |
6183c009 TH |
4039 | WARN_ON_ONCE(wq->flush_color == wq->work_color); |
4040 | WARN_ON_ONCE(wq->flush_color != next->flush_color); | |
73f53c4a TH |
4041 | |
4042 | list_del_init(&next->list); | |
4043 | wq->first_flusher = next; | |
4044 | ||
112202d9 | 4045 | if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) |
73f53c4a TH |
4046 | break; |
4047 | ||
4048 | /* | |
4049 | * Meh... this color is already done, clear first | |
4050 | * flusher and repeat cascading. | |
4051 | */ | |
4052 | wq->first_flusher = NULL; | |
4053 | } | |
4054 | ||
4055 | out_unlock: | |
3c25a55d | 4056 | mutex_unlock(&wq->mutex); |
1da177e4 | 4057 | } |
c4f135d6 | 4058 | EXPORT_SYMBOL(__flush_workqueue); |
1da177e4 | 4059 | |
9c5a2ba7 TH |
4060 | /** |
4061 | * drain_workqueue - drain a workqueue | |
4062 | * @wq: workqueue to drain | |
4063 | * | |
4064 | * Wait until the workqueue becomes empty. While draining is in progress, | |
4065 | * only chain queueing is allowed. IOW, only currently pending or running | |
4066 | * work items on @wq can queue further work items on it. @wq is flushed | |
b749b1b6 | 4067 | * repeatedly until it becomes empty. The number of flushing is determined |
9c5a2ba7 TH |
4068 | * by the depth of chaining and should be relatively short. Whine if it |
4069 | * takes too long. | |
4070 | */ | |
4071 | void drain_workqueue(struct workqueue_struct *wq) | |
4072 | { | |
4073 | unsigned int flush_cnt = 0; | |
49e3cf44 | 4074 | struct pool_workqueue *pwq; |
9c5a2ba7 TH |
4075 | |
4076 | /* | |
4077 | * __queue_work() needs to test whether there are drainers, is much | |
4078 | * hotter than drain_workqueue() and already looks at @wq->flags. | |
618b01eb | 4079 | * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. |
9c5a2ba7 | 4080 | */ |
87fc741e | 4081 | mutex_lock(&wq->mutex); |
9c5a2ba7 | 4082 | if (!wq->nr_drainers++) |
618b01eb | 4083 | wq->flags |= __WQ_DRAINING; |
87fc741e | 4084 | mutex_unlock(&wq->mutex); |
9c5a2ba7 | 4085 | reflush: |
c4f135d6 | 4086 | __flush_workqueue(wq); |
9c5a2ba7 | 4087 | |
b09f4fd3 | 4088 | mutex_lock(&wq->mutex); |
76af4d93 | 4089 | |
49e3cf44 | 4090 | for_each_pwq(pwq, wq) { |
fa2563e4 | 4091 | bool drained; |
9c5a2ba7 | 4092 | |
a9b8a985 | 4093 | raw_spin_lock_irq(&pwq->pool->lock); |
afa87ce8 | 4094 | drained = pwq_is_empty(pwq); |
a9b8a985 | 4095 | raw_spin_unlock_irq(&pwq->pool->lock); |
fa2563e4 TT |
4096 | |
4097 | if (drained) | |
9c5a2ba7 TH |
4098 | continue; |
4099 | ||
4100 | if (++flush_cnt == 10 || | |
4101 | (flush_cnt % 100 == 0 && flush_cnt <= 1000)) | |
e9ad2eb3 SZ |
4102 | pr_warn("workqueue %s: %s() isn't complete after %u tries\n", |
4103 | wq->name, __func__, flush_cnt); | |
76af4d93 | 4104 | |
b09f4fd3 | 4105 | mutex_unlock(&wq->mutex); |
9c5a2ba7 TH |
4106 | goto reflush; |
4107 | } | |
4108 | ||
9c5a2ba7 | 4109 | if (!--wq->nr_drainers) |
618b01eb | 4110 | wq->flags &= ~__WQ_DRAINING; |
87fc741e | 4111 | mutex_unlock(&wq->mutex); |
9c5a2ba7 TH |
4112 | } |
4113 | EXPORT_SYMBOL_GPL(drain_workqueue); | |
4114 | ||
d6e89786 JB |
4115 | static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr, |
4116 | bool from_cancel) | |
db700897 | 4117 | { |
affee4b2 | 4118 | struct worker *worker = NULL; |
c9e7cf27 | 4119 | struct worker_pool *pool; |
112202d9 | 4120 | struct pool_workqueue *pwq; |
c35aea39 | 4121 | struct workqueue_struct *wq; |
db700897 | 4122 | |
24acfb71 | 4123 | rcu_read_lock(); |
c9e7cf27 | 4124 | pool = get_work_pool(work); |
fa1b54e6 | 4125 | if (!pool) { |
24acfb71 | 4126 | rcu_read_unlock(); |
baf59022 | 4127 | return false; |
fa1b54e6 | 4128 | } |
db700897 | 4129 | |
a9b8a985 | 4130 | raw_spin_lock_irq(&pool->lock); |
0b3dae68 | 4131 | /* see the comment in try_to_grab_pending() with the same code */ |
112202d9 TH |
4132 | pwq = get_work_pwq(work); |
4133 | if (pwq) { | |
4134 | if (unlikely(pwq->pool != pool)) | |
4690c4ab | 4135 | goto already_gone; |
606a5020 | 4136 | } else { |
c9e7cf27 | 4137 | worker = find_worker_executing_work(pool, work); |
affee4b2 | 4138 | if (!worker) |
4690c4ab | 4139 | goto already_gone; |
112202d9 | 4140 | pwq = worker->current_pwq; |
606a5020 | 4141 | } |
db700897 | 4142 | |
c35aea39 TH |
4143 | wq = pwq->wq; |
4144 | check_flush_dependency(wq, work); | |
fca839c0 | 4145 | |
112202d9 | 4146 | insert_wq_barrier(pwq, barr, work, worker); |
a9b8a985 | 4147 | raw_spin_unlock_irq(&pool->lock); |
7a22ad75 | 4148 | |
c35aea39 TH |
4149 | touch_work_lockdep_map(work, wq); |
4150 | ||
e159489b | 4151 | /* |
a1d14934 PZ |
4152 | * Force a lock recursion deadlock when using flush_work() inside a |
4153 | * single-threaded or rescuer equipped workqueue. | |
4154 | * | |
4155 | * For single threaded workqueues the deadlock happens when the work | |
4156 | * is after the work issuing the flush_work(). For rescuer equipped | |
4157 | * workqueues the deadlock happens when the rescuer stalls, blocking | |
4158 | * forward progress. | |
e159489b | 4159 | */ |
c35aea39 TH |
4160 | if (!from_cancel && (wq->saved_max_active == 1 || wq->rescuer)) |
4161 | touch_wq_lockdep_map(wq); | |
4162 | ||
24acfb71 | 4163 | rcu_read_unlock(); |
401a8d04 | 4164 | return true; |
4690c4ab | 4165 | already_gone: |
a9b8a985 | 4166 | raw_spin_unlock_irq(&pool->lock); |
24acfb71 | 4167 | rcu_read_unlock(); |
401a8d04 | 4168 | return false; |
db700897 | 4169 | } |
baf59022 | 4170 | |
d6e89786 JB |
4171 | static bool __flush_work(struct work_struct *work, bool from_cancel) |
4172 | { | |
4173 | struct wq_barrier barr; | |
4174 | ||
4175 | if (WARN_ON(!wq_online)) | |
4176 | return false; | |
4177 | ||
4d43d395 TH |
4178 | if (WARN_ON(!work->func)) |
4179 | return false; | |
4180 | ||
134874e2 | 4181 | if (!start_flush_work(work, &barr, from_cancel)) |
d6e89786 | 4182 | return false; |
134874e2 TH |
4183 | |
4184 | /* | |
4185 | * start_flush_work() returned %true. If @from_cancel is set, we know | |
4186 | * that @work must have been executing during start_flush_work() and | |
4187 | * can't currently be queued. Its data must contain OFFQ bits. If @work | |
4188 | * was queued on a BH workqueue, we also know that it was running in the | |
4189 | * BH context and thus can be busy-waited. | |
4190 | */ | |
8bc35475 TH |
4191 | if (from_cancel) { |
4192 | unsigned long data = *work_data_bits(work); | |
4193 | ||
4194 | if (!WARN_ON_ONCE(data & WORK_STRUCT_PWQ) && | |
4195 | (data & WORK_OFFQ_BH)) { | |
4196 | /* | |
4197 | * On RT, prevent a live lock when %current preempted | |
4198 | * soft interrupt processing or prevents ksoftirqd from | |
4199 | * running by keeping flipping BH. If the BH work item | |
4200 | * runs on a different CPU then this has no effect other | |
4201 | * than doing the BH disable/enable dance for nothing. | |
4202 | * This is copied from | |
4203 | * kernel/softirq.c::tasklet_unlock_spin_wait(). | |
4204 | */ | |
4205 | while (!try_wait_for_completion(&barr.done)) { | |
4206 | if (IS_ENABLED(CONFIG_PREEMPT_RT)) { | |
4207 | local_bh_disable(); | |
4208 | local_bh_enable(); | |
4209 | } else { | |
4210 | cpu_relax(); | |
4211 | } | |
134874e2 | 4212 | } |
8bc35475 | 4213 | goto out_destroy; |
134874e2 | 4214 | } |
d6e89786 | 4215 | } |
134874e2 | 4216 | |
8bc35475 TH |
4217 | wait_for_completion(&barr.done); |
4218 | ||
4219 | out_destroy: | |
134874e2 TH |
4220 | destroy_work_on_stack(&barr.work); |
4221 | return true; | |
d6e89786 JB |
4222 | } |
4223 | ||
baf59022 TH |
4224 | /** |
4225 | * flush_work - wait for a work to finish executing the last queueing instance | |
4226 | * @work: the work to flush | |
4227 | * | |
606a5020 TH |
4228 | * Wait until @work has finished execution. @work is guaranteed to be idle |
4229 | * on return if it hasn't been requeued since flush started. | |
baf59022 | 4230 | * |
d185af30 | 4231 | * Return: |
baf59022 TH |
4232 | * %true if flush_work() waited for the work to finish execution, |
4233 | * %false if it was already idle. | |
4234 | */ | |
4235 | bool flush_work(struct work_struct *work) | |
4236 | { | |
134874e2 | 4237 | might_sleep(); |
d6e89786 | 4238 | return __flush_work(work, false); |
6e84d644 | 4239 | } |
606a5020 | 4240 | EXPORT_SYMBOL_GPL(flush_work); |
6e84d644 | 4241 | |
cdc6e4b3 TH |
4242 | /** |
4243 | * flush_delayed_work - wait for a dwork to finish executing the last queueing | |
4244 | * @dwork: the delayed work to flush | |
4245 | * | |
4246 | * Delayed timer is cancelled and the pending work is queued for | |
4247 | * immediate execution. Like flush_work(), this function only | |
4248 | * considers the last queueing instance of @dwork. | |
4249 | * | |
4250 | * Return: | |
4251 | * %true if flush_work() waited for the work to finish execution, | |
4252 | * %false if it was already idle. | |
4253 | */ | |
4254 | bool flush_delayed_work(struct delayed_work *dwork) | |
4255 | { | |
4256 | local_irq_disable(); | |
4257 | if (del_timer_sync(&dwork->timer)) | |
4258 | __queue_work(dwork->cpu, dwork->wq, &dwork->work); | |
4259 | local_irq_enable(); | |
4260 | return flush_work(&dwork->work); | |
4261 | } | |
4262 | EXPORT_SYMBOL(flush_delayed_work); | |
4263 | ||
4264 | /** | |
4265 | * flush_rcu_work - wait for a rwork to finish executing the last queueing | |
4266 | * @rwork: the rcu work to flush | |
4267 | * | |
4268 | * Return: | |
4269 | * %true if flush_rcu_work() waited for the work to finish execution, | |
4270 | * %false if it was already idle. | |
4271 | */ | |
4272 | bool flush_rcu_work(struct rcu_work *rwork) | |
4273 | { | |
4274 | if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) { | |
4275 | rcu_barrier(); | |
4276 | flush_work(&rwork->work); | |
4277 | return true; | |
4278 | } else { | |
4279 | return flush_work(&rwork->work); | |
4280 | } | |
4281 | } | |
4282 | EXPORT_SYMBOL(flush_rcu_work); | |
4283 | ||
86898fa6 TH |
4284 | static void work_offqd_disable(struct work_offq_data *offqd) |
4285 | { | |
4286 | const unsigned long max = (1lu << WORK_OFFQ_DISABLE_BITS) - 1; | |
4287 | ||
4288 | if (likely(offqd->disable < max)) | |
4289 | offqd->disable++; | |
4290 | else | |
4291 | WARN_ONCE(true, "workqueue: work disable count overflowed\n"); | |
4292 | } | |
4293 | ||
4294 | static void work_offqd_enable(struct work_offq_data *offqd) | |
4295 | { | |
4296 | if (likely(offqd->disable > 0)) | |
4297 | offqd->disable--; | |
4298 | else | |
4299 | WARN_ONCE(true, "workqueue: work disable count underflowed\n"); | |
4300 | } | |
4301 | ||
c5f5b942 | 4302 | static bool __cancel_work(struct work_struct *work, u32 cflags) |
cdc6e4b3 | 4303 | { |
1211f3b2 | 4304 | struct work_offq_data offqd; |
c26e2f2e | 4305 | unsigned long irq_flags; |
cdc6e4b3 TH |
4306 | int ret; |
4307 | ||
f09b10b6 | 4308 | ret = work_grab_pending(work, cflags, &irq_flags); |
cdc6e4b3 | 4309 | |
1211f3b2 | 4310 | work_offqd_unpack(&offqd, *work_data_bits(work)); |
cdc6e4b3 | 4311 | |
86898fa6 TH |
4312 | if (cflags & WORK_CANCEL_DISABLE) |
4313 | work_offqd_disable(&offqd); | |
4314 | ||
1211f3b2 TH |
4315 | set_work_pool_and_clear_pending(work, offqd.pool_id, |
4316 | work_offqd_pack_flags(&offqd)); | |
c26e2f2e | 4317 | local_irq_restore(irq_flags); |
cdc6e4b3 TH |
4318 | return ret; |
4319 | } | |
4320 | ||
c5f5b942 | 4321 | static bool __cancel_work_sync(struct work_struct *work, u32 cflags) |
1f1f642e | 4322 | { |
978b8409 | 4323 | bool ret; |
1f1f642e | 4324 | |
f09b10b6 | 4325 | ret = __cancel_work(work, cflags | WORK_CANCEL_DISABLE); |
bbb68dfa | 4326 | |
134874e2 TH |
4327 | if (*work_data_bits(work) & WORK_OFFQ_BH) |
4328 | WARN_ON_ONCE(in_hardirq()); | |
4329 | else | |
4330 | might_sleep(); | |
bbb68dfa | 4331 | |
3347fa09 | 4332 | /* |
c7a40c49 TH |
4333 | * Skip __flush_work() during early boot when we know that @work isn't |
4334 | * executing. This allows canceling during early boot. | |
3347fa09 TH |
4335 | */ |
4336 | if (wq_online) | |
d6e89786 | 4337 | __flush_work(work, true); |
3347fa09 | 4338 | |
f09b10b6 TH |
4339 | if (!(cflags & WORK_CANCEL_DISABLE)) |
4340 | enable_work(work); | |
8603e1b3 | 4341 | |
1f1f642e ON |
4342 | return ret; |
4343 | } | |
4344 | ||
cdc6e4b3 TH |
4345 | /* |
4346 | * See cancel_delayed_work() | |
4347 | */ | |
4348 | bool cancel_work(struct work_struct *work) | |
4349 | { | |
c5f5b942 | 4350 | return __cancel_work(work, 0); |
cdc6e4b3 TH |
4351 | } |
4352 | EXPORT_SYMBOL(cancel_work); | |
4353 | ||
6e84d644 | 4354 | /** |
401a8d04 TH |
4355 | * cancel_work_sync - cancel a work and wait for it to finish |
4356 | * @work: the work to cancel | |
6e84d644 | 4357 | * |
134874e2 TH |
4358 | * Cancel @work and wait for its execution to finish. This function can be used |
4359 | * even if the work re-queues itself or migrates to another workqueue. On return | |
4360 | * from this function, @work is guaranteed to be not pending or executing on any | |
4361 | * CPU as long as there aren't racing enqueues. | |
1f1f642e | 4362 | * |
134874e2 TH |
4363 | * cancel_work_sync(&delayed_work->work) must not be used for delayed_work's. |
4364 | * Use cancel_delayed_work_sync() instead. | |
6e84d644 | 4365 | * |
134874e2 TH |
4366 | * Must be called from a sleepable context if @work was last queued on a non-BH |
4367 | * workqueue. Can also be called from non-hardirq atomic contexts including BH | |
4368 | * if @work was last queued on a BH workqueue. | |
401a8d04 | 4369 | * |
134874e2 | 4370 | * Returns %true if @work was pending, %false otherwise. |
6e84d644 | 4371 | */ |
401a8d04 | 4372 | bool cancel_work_sync(struct work_struct *work) |
6e84d644 | 4373 | { |
c5f5b942 | 4374 | return __cancel_work_sync(work, 0); |
b89deed3 | 4375 | } |
28e53bdd | 4376 | EXPORT_SYMBOL_GPL(cancel_work_sync); |
b89deed3 | 4377 | |
09383498 | 4378 | /** |
57b30ae7 TH |
4379 | * cancel_delayed_work - cancel a delayed work |
4380 | * @dwork: delayed_work to cancel | |
09383498 | 4381 | * |
d185af30 YB |
4382 | * Kill off a pending delayed_work. |
4383 | * | |
4384 | * Return: %true if @dwork was pending and canceled; %false if it wasn't | |
4385 | * pending. | |
4386 | * | |
4387 | * Note: | |
4388 | * The work callback function may still be running on return, unless | |
4389 | * it returns %true and the work doesn't re-arm itself. Explicitly flush or | |
4390 | * use cancel_delayed_work_sync() to wait on it. | |
09383498 | 4391 | * |
57b30ae7 | 4392 | * This function is safe to call from any context including IRQ handler. |
09383498 | 4393 | */ |
57b30ae7 | 4394 | bool cancel_delayed_work(struct delayed_work *dwork) |
09383498 | 4395 | { |
c5f5b942 | 4396 | return __cancel_work(&dwork->work, WORK_CANCEL_DELAYED); |
09383498 | 4397 | } |
57b30ae7 | 4398 | EXPORT_SYMBOL(cancel_delayed_work); |
09383498 | 4399 | |
401a8d04 TH |
4400 | /** |
4401 | * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish | |
4402 | * @dwork: the delayed work cancel | |
4403 | * | |
4404 | * This is cancel_work_sync() for delayed works. | |
4405 | * | |
d185af30 | 4406 | * Return: |
401a8d04 TH |
4407 | * %true if @dwork was pending, %false otherwise. |
4408 | */ | |
4409 | bool cancel_delayed_work_sync(struct delayed_work *dwork) | |
6e84d644 | 4410 | { |
c5f5b942 | 4411 | return __cancel_work_sync(&dwork->work, WORK_CANCEL_DELAYED); |
6e84d644 | 4412 | } |
f5a421a4 | 4413 | EXPORT_SYMBOL(cancel_delayed_work_sync); |
1da177e4 | 4414 | |
86898fa6 TH |
4415 | /** |
4416 | * disable_work - Disable and cancel a work item | |
4417 | * @work: work item to disable | |
4418 | * | |
4419 | * Disable @work by incrementing its disable count and cancel it if currently | |
4420 | * pending. As long as the disable count is non-zero, any attempt to queue @work | |
4421 | * will fail and return %false. The maximum supported disable depth is 2 to the | |
4422 | * power of %WORK_OFFQ_DISABLE_BITS, currently 65536. | |
4423 | * | |
f09b10b6 TH |
4424 | * Can be called from any context. Returns %true if @work was pending, %false |
4425 | * otherwise. | |
86898fa6 TH |
4426 | */ |
4427 | bool disable_work(struct work_struct *work) | |
4428 | { | |
4429 | return __cancel_work(work, WORK_CANCEL_DISABLE); | |
4430 | } | |
4431 | EXPORT_SYMBOL_GPL(disable_work); | |
4432 | ||
4433 | /** | |
4434 | * disable_work_sync - Disable, cancel and drain a work item | |
4435 | * @work: work item to disable | |
4436 | * | |
4437 | * Similar to disable_work() but also wait for @work to finish if currently | |
4438 | * executing. | |
4439 | * | |
134874e2 TH |
4440 | * Must be called from a sleepable context if @work was last queued on a non-BH |
4441 | * workqueue. Can also be called from non-hardirq atomic contexts including BH | |
4442 | * if @work was last queued on a BH workqueue. | |
4443 | * | |
4444 | * Returns %true if @work was pending, %false otherwise. | |
86898fa6 TH |
4445 | */ |
4446 | bool disable_work_sync(struct work_struct *work) | |
4447 | { | |
4448 | return __cancel_work_sync(work, WORK_CANCEL_DISABLE); | |
4449 | } | |
4450 | EXPORT_SYMBOL_GPL(disable_work_sync); | |
4451 | ||
4452 | /** | |
4453 | * enable_work - Enable a work item | |
4454 | * @work: work item to enable | |
4455 | * | |
4456 | * Undo disable_work[_sync]() by decrementing @work's disable count. @work can | |
4457 | * only be queued if its disable count is 0. | |
4458 | * | |
f09b10b6 TH |
4459 | * Can be called from any context. Returns %true if the disable count reached 0. |
4460 | * Otherwise, %false. | |
86898fa6 TH |
4461 | */ |
4462 | bool enable_work(struct work_struct *work) | |
4463 | { | |
4464 | struct work_offq_data offqd; | |
4465 | unsigned long irq_flags; | |
4466 | ||
4467 | work_grab_pending(work, 0, &irq_flags); | |
4468 | ||
4469 | work_offqd_unpack(&offqd, *work_data_bits(work)); | |
4470 | work_offqd_enable(&offqd); | |
4471 | set_work_pool_and_clear_pending(work, offqd.pool_id, | |
4472 | work_offqd_pack_flags(&offqd)); | |
4473 | local_irq_restore(irq_flags); | |
4474 | ||
4475 | return !offqd.disable; | |
4476 | } | |
4477 | EXPORT_SYMBOL_GPL(enable_work); | |
4478 | ||
4479 | /** | |
4480 | * disable_delayed_work - Disable and cancel a delayed work item | |
4481 | * @dwork: delayed work item to disable | |
4482 | * | |
4483 | * disable_work() for delayed work items. | |
4484 | */ | |
4485 | bool disable_delayed_work(struct delayed_work *dwork) | |
4486 | { | |
4487 | return __cancel_work(&dwork->work, | |
4488 | WORK_CANCEL_DELAYED | WORK_CANCEL_DISABLE); | |
4489 | } | |
4490 | EXPORT_SYMBOL_GPL(disable_delayed_work); | |
4491 | ||
4492 | /** | |
4493 | * disable_delayed_work_sync - Disable, cancel and drain a delayed work item | |
4494 | * @dwork: delayed work item to disable | |
4495 | * | |
4496 | * disable_work_sync() for delayed work items. | |
4497 | */ | |
4498 | bool disable_delayed_work_sync(struct delayed_work *dwork) | |
4499 | { | |
4500 | return __cancel_work_sync(&dwork->work, | |
4501 | WORK_CANCEL_DELAYED | WORK_CANCEL_DISABLE); | |
4502 | } | |
4503 | EXPORT_SYMBOL_GPL(disable_delayed_work_sync); | |
4504 | ||
4505 | /** | |
4506 | * enable_delayed_work - Enable a delayed work item | |
4507 | * @dwork: delayed work item to enable | |
4508 | * | |
4509 | * enable_work() for delayed work items. | |
4510 | */ | |
4511 | bool enable_delayed_work(struct delayed_work *dwork) | |
4512 | { | |
4513 | return enable_work(&dwork->work); | |
4514 | } | |
4515 | EXPORT_SYMBOL_GPL(enable_delayed_work); | |
4516 | ||
b6136773 | 4517 | /** |
31ddd871 | 4518 | * schedule_on_each_cpu - execute a function synchronously on each online CPU |
b6136773 | 4519 | * @func: the function to call |
b6136773 | 4520 | * |
31ddd871 TH |
4521 | * schedule_on_each_cpu() executes @func on each online CPU using the |
4522 | * system workqueue and blocks until all CPUs have completed. | |
b6136773 | 4523 | * schedule_on_each_cpu() is very slow. |
31ddd871 | 4524 | * |
d185af30 | 4525 | * Return: |
31ddd871 | 4526 | * 0 on success, -errno on failure. |
b6136773 | 4527 | */ |
65f27f38 | 4528 | int schedule_on_each_cpu(work_func_t func) |
15316ba8 CL |
4529 | { |
4530 | int cpu; | |
38f51568 | 4531 | struct work_struct __percpu *works; |
15316ba8 | 4532 | |
b6136773 AM |
4533 | works = alloc_percpu(struct work_struct); |
4534 | if (!works) | |
15316ba8 | 4535 | return -ENOMEM; |
b6136773 | 4536 | |
ffd8bea8 | 4537 | cpus_read_lock(); |
93981800 | 4538 | |
15316ba8 | 4539 | for_each_online_cpu(cpu) { |
9bfb1839 IM |
4540 | struct work_struct *work = per_cpu_ptr(works, cpu); |
4541 | ||
4542 | INIT_WORK(work, func); | |
b71ab8c2 | 4543 | schedule_work_on(cpu, work); |
65a64464 | 4544 | } |
93981800 TH |
4545 | |
4546 | for_each_online_cpu(cpu) | |
4547 | flush_work(per_cpu_ptr(works, cpu)); | |
4548 | ||
ffd8bea8 | 4549 | cpus_read_unlock(); |
b6136773 | 4550 | free_percpu(works); |
15316ba8 CL |
4551 | return 0; |
4552 | } | |
4553 | ||
1fa44eca JB |
4554 | /** |
4555 | * execute_in_process_context - reliably execute the routine with user context | |
4556 | * @fn: the function to execute | |
1fa44eca JB |
4557 | * @ew: guaranteed storage for the execute work structure (must |
4558 | * be available when the work executes) | |
4559 | * | |
4560 | * Executes the function immediately if process context is available, | |
4561 | * otherwise schedules the function for delayed execution. | |
4562 | * | |
d185af30 | 4563 | * Return: 0 - function was executed |
1fa44eca JB |
4564 | * 1 - function was scheduled for execution |
4565 | */ | |
65f27f38 | 4566 | int execute_in_process_context(work_func_t fn, struct execute_work *ew) |
1fa44eca JB |
4567 | { |
4568 | if (!in_interrupt()) { | |
65f27f38 | 4569 | fn(&ew->work); |
1fa44eca JB |
4570 | return 0; |
4571 | } | |
4572 | ||
65f27f38 | 4573 | INIT_WORK(&ew->work, fn); |
1fa44eca JB |
4574 | schedule_work(&ew->work); |
4575 | ||
4576 | return 1; | |
4577 | } | |
4578 | EXPORT_SYMBOL_GPL(execute_in_process_context); | |
4579 | ||
6ba94429 FW |
4580 | /** |
4581 | * free_workqueue_attrs - free a workqueue_attrs | |
4582 | * @attrs: workqueue_attrs to free | |
226223ab | 4583 | * |
6ba94429 | 4584 | * Undo alloc_workqueue_attrs(). |
226223ab | 4585 | */ |
513c98d0 | 4586 | void free_workqueue_attrs(struct workqueue_attrs *attrs) |
226223ab | 4587 | { |
6ba94429 FW |
4588 | if (attrs) { |
4589 | free_cpumask_var(attrs->cpumask); | |
9546b29e | 4590 | free_cpumask_var(attrs->__pod_cpumask); |
6ba94429 FW |
4591 | kfree(attrs); |
4592 | } | |
226223ab TH |
4593 | } |
4594 | ||
6ba94429 FW |
4595 | /** |
4596 | * alloc_workqueue_attrs - allocate a workqueue_attrs | |
6ba94429 FW |
4597 | * |
4598 | * Allocate a new workqueue_attrs, initialize with default settings and | |
4599 | * return it. | |
4600 | * | |
4601 | * Return: The allocated new workqueue_attr on success. %NULL on failure. | |
4602 | */ | |
513c98d0 | 4603 | struct workqueue_attrs *alloc_workqueue_attrs(void) |
226223ab | 4604 | { |
6ba94429 | 4605 | struct workqueue_attrs *attrs; |
226223ab | 4606 | |
be69d00d | 4607 | attrs = kzalloc(sizeof(*attrs), GFP_KERNEL); |
6ba94429 FW |
4608 | if (!attrs) |
4609 | goto fail; | |
be69d00d | 4610 | if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL)) |
6ba94429 | 4611 | goto fail; |
9546b29e TH |
4612 | if (!alloc_cpumask_var(&attrs->__pod_cpumask, GFP_KERNEL)) |
4613 | goto fail; | |
6ba94429 FW |
4614 | |
4615 | cpumask_copy(attrs->cpumask, cpu_possible_mask); | |
523a301e | 4616 | attrs->affn_scope = WQ_AFFN_DFL; |
6ba94429 FW |
4617 | return attrs; |
4618 | fail: | |
4619 | free_workqueue_attrs(attrs); | |
4620 | return NULL; | |
226223ab TH |
4621 | } |
4622 | ||
6ba94429 FW |
4623 | static void copy_workqueue_attrs(struct workqueue_attrs *to, |
4624 | const struct workqueue_attrs *from) | |
226223ab | 4625 | { |
6ba94429 FW |
4626 | to->nice = from->nice; |
4627 | cpumask_copy(to->cpumask, from->cpumask); | |
9546b29e | 4628 | cpumask_copy(to->__pod_cpumask, from->__pod_cpumask); |
8639eceb | 4629 | to->affn_strict = from->affn_strict; |
84193c07 | 4630 | |
6ba94429 | 4631 | /* |
84193c07 TH |
4632 | * Unlike hash and equality test, copying shouldn't ignore wq-only |
4633 | * fields as copying is used for both pool and wq attrs. Instead, | |
4634 | * get_unbound_pool() explicitly clears the fields. | |
6ba94429 | 4635 | */ |
84193c07 | 4636 | to->affn_scope = from->affn_scope; |
af73f5c9 | 4637 | to->ordered = from->ordered; |
226223ab TH |
4638 | } |
4639 | ||
5de7a03c TH |
4640 | /* |
4641 | * Some attrs fields are workqueue-only. Clear them for worker_pool's. See the | |
4642 | * comments in 'struct workqueue_attrs' definition. | |
4643 | */ | |
4644 | static void wqattrs_clear_for_pool(struct workqueue_attrs *attrs) | |
4645 | { | |
84193c07 | 4646 | attrs->affn_scope = WQ_AFFN_NR_TYPES; |
5de7a03c | 4647 | attrs->ordered = false; |
ae1296a7 LJ |
4648 | if (attrs->affn_strict) |
4649 | cpumask_copy(attrs->cpumask, cpu_possible_mask); | |
5de7a03c TH |
4650 | } |
4651 | ||
6ba94429 FW |
4652 | /* hash value of the content of @attr */ |
4653 | static u32 wqattrs_hash(const struct workqueue_attrs *attrs) | |
226223ab | 4654 | { |
6ba94429 | 4655 | u32 hash = 0; |
226223ab | 4656 | |
6ba94429 | 4657 | hash = jhash_1word(attrs->nice, hash); |
ae1296a7 | 4658 | hash = jhash_1word(attrs->affn_strict, hash); |
9546b29e TH |
4659 | hash = jhash(cpumask_bits(attrs->__pod_cpumask), |
4660 | BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); | |
ae1296a7 LJ |
4661 | if (!attrs->affn_strict) |
4662 | hash = jhash(cpumask_bits(attrs->cpumask), | |
4663 | BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); | |
6ba94429 | 4664 | return hash; |
226223ab | 4665 | } |
226223ab | 4666 | |
6ba94429 FW |
4667 | /* content equality test */ |
4668 | static bool wqattrs_equal(const struct workqueue_attrs *a, | |
4669 | const struct workqueue_attrs *b) | |
226223ab | 4670 | { |
6ba94429 FW |
4671 | if (a->nice != b->nice) |
4672 | return false; | |
ae1296a7 | 4673 | if (a->affn_strict != b->affn_strict) |
6ba94429 | 4674 | return false; |
9546b29e TH |
4675 | if (!cpumask_equal(a->__pod_cpumask, b->__pod_cpumask)) |
4676 | return false; | |
ae1296a7 | 4677 | if (!a->affn_strict && !cpumask_equal(a->cpumask, b->cpumask)) |
8639eceb | 4678 | return false; |
6ba94429 | 4679 | return true; |
226223ab TH |
4680 | } |
4681 | ||
0f36ee24 TH |
4682 | /* Update @attrs with actually available CPUs */ |
4683 | static void wqattrs_actualize_cpumask(struct workqueue_attrs *attrs, | |
4684 | const cpumask_t *unbound_cpumask) | |
4685 | { | |
4686 | /* | |
4687 | * Calculate the effective CPU mask of @attrs given @unbound_cpumask. If | |
4688 | * @attrs->cpumask doesn't overlap with @unbound_cpumask, we fallback to | |
4689 | * @unbound_cpumask. | |
4690 | */ | |
4691 | cpumask_and(attrs->cpumask, attrs->cpumask, unbound_cpumask); | |
4692 | if (unlikely(cpumask_empty(attrs->cpumask))) | |
4693 | cpumask_copy(attrs->cpumask, unbound_cpumask); | |
4694 | } | |
4695 | ||
84193c07 TH |
4696 | /* find wq_pod_type to use for @attrs */ |
4697 | static const struct wq_pod_type * | |
4698 | wqattrs_pod_type(const struct workqueue_attrs *attrs) | |
4699 | { | |
523a301e TH |
4700 | enum wq_affn_scope scope; |
4701 | struct wq_pod_type *pt; | |
4702 | ||
4703 | /* to synchronize access to wq_affn_dfl */ | |
4704 | lockdep_assert_held(&wq_pool_mutex); | |
4705 | ||
4706 | if (attrs->affn_scope == WQ_AFFN_DFL) | |
4707 | scope = wq_affn_dfl; | |
4708 | else | |
4709 | scope = attrs->affn_scope; | |
4710 | ||
4711 | pt = &wq_pod_types[scope]; | |
84193c07 TH |
4712 | |
4713 | if (!WARN_ON_ONCE(attrs->affn_scope == WQ_AFFN_NR_TYPES) && | |
4714 | likely(pt->nr_pods)) | |
4715 | return pt; | |
4716 | ||
4717 | /* | |
4718 | * Before workqueue_init_topology(), only SYSTEM is available which is | |
4719 | * initialized in workqueue_init_early(). | |
4720 | */ | |
4721 | pt = &wq_pod_types[WQ_AFFN_SYSTEM]; | |
4722 | BUG_ON(!pt->nr_pods); | |
4723 | return pt; | |
4724 | } | |
4725 | ||
6ba94429 FW |
4726 | /** |
4727 | * init_worker_pool - initialize a newly zalloc'd worker_pool | |
4728 | * @pool: worker_pool to initialize | |
4729 | * | |
402dd89d | 4730 | * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs. |
6ba94429 FW |
4731 | * |
4732 | * Return: 0 on success, -errno on failure. Even on failure, all fields | |
4733 | * inside @pool proper are initialized and put_unbound_pool() can be called | |
4734 | * on @pool safely to release it. | |
4735 | */ | |
4736 | static int init_worker_pool(struct worker_pool *pool) | |
226223ab | 4737 | { |
a9b8a985 | 4738 | raw_spin_lock_init(&pool->lock); |
6ba94429 FW |
4739 | pool->id = -1; |
4740 | pool->cpu = -1; | |
4741 | pool->node = NUMA_NO_NODE; | |
4742 | pool->flags |= POOL_DISASSOCIATED; | |
82607adc | 4743 | pool->watchdog_ts = jiffies; |
6ba94429 FW |
4744 | INIT_LIST_HEAD(&pool->worklist); |
4745 | INIT_LIST_HEAD(&pool->idle_list); | |
4746 | hash_init(pool->busy_hash); | |
226223ab | 4747 | |
32a6c723 | 4748 | timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE); |
3f959aa3 | 4749 | INIT_WORK(&pool->idle_cull_work, idle_cull_fn); |
226223ab | 4750 | |
32a6c723 | 4751 | timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0); |
226223ab | 4752 | |
6ba94429 | 4753 | INIT_LIST_HEAD(&pool->workers); |
226223ab | 4754 | |
6ba94429 FW |
4755 | ida_init(&pool->worker_ida); |
4756 | INIT_HLIST_NODE(&pool->hash_node); | |
4757 | pool->refcnt = 1; | |
226223ab | 4758 | |
6ba94429 | 4759 | /* shouldn't fail above this point */ |
be69d00d | 4760 | pool->attrs = alloc_workqueue_attrs(); |
6ba94429 FW |
4761 | if (!pool->attrs) |
4762 | return -ENOMEM; | |
5de7a03c TH |
4763 | |
4764 | wqattrs_clear_for_pool(pool->attrs); | |
4765 | ||
6ba94429 | 4766 | return 0; |
226223ab TH |
4767 | } |
4768 | ||
669de8bd BVA |
4769 | #ifdef CONFIG_LOCKDEP |
4770 | static void wq_init_lockdep(struct workqueue_struct *wq) | |
4771 | { | |
4772 | char *lock_name; | |
4773 | ||
4774 | lockdep_register_key(&wq->key); | |
4775 | lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name); | |
4776 | if (!lock_name) | |
4777 | lock_name = wq->name; | |
69a106c0 QC |
4778 | |
4779 | wq->lock_name = lock_name; | |
4f022f43 MB |
4780 | wq->lockdep_map = &wq->__lockdep_map; |
4781 | lockdep_init_map(wq->lockdep_map, lock_name, &wq->key, 0); | |
669de8bd BVA |
4782 | } |
4783 | ||
4784 | static void wq_unregister_lockdep(struct workqueue_struct *wq) | |
4785 | { | |
ec0a7d44 MB |
4786 | if (wq->lockdep_map != &wq->__lockdep_map) |
4787 | return; | |
4788 | ||
669de8bd BVA |
4789 | lockdep_unregister_key(&wq->key); |
4790 | } | |
4791 | ||
4792 | static void wq_free_lockdep(struct workqueue_struct *wq) | |
4793 | { | |
ec0a7d44 MB |
4794 | if (wq->lockdep_map != &wq->__lockdep_map) |
4795 | return; | |
4796 | ||
669de8bd BVA |
4797 | if (wq->lock_name != wq->name) |
4798 | kfree(wq->lock_name); | |
4799 | } | |
4800 | #else | |
4801 | static void wq_init_lockdep(struct workqueue_struct *wq) | |
4802 | { | |
4803 | } | |
4804 | ||
4805 | static void wq_unregister_lockdep(struct workqueue_struct *wq) | |
4806 | { | |
4807 | } | |
4808 | ||
4809 | static void wq_free_lockdep(struct workqueue_struct *wq) | |
4810 | { | |
4811 | } | |
4812 | #endif | |
4813 | ||
91ccc6e7 TH |
4814 | static void free_node_nr_active(struct wq_node_nr_active **nna_ar) |
4815 | { | |
4816 | int node; | |
4817 | ||
4818 | for_each_node(node) { | |
4819 | kfree(nna_ar[node]); | |
4820 | nna_ar[node] = NULL; | |
4821 | } | |
4822 | ||
4823 | kfree(nna_ar[nr_node_ids]); | |
4824 | nna_ar[nr_node_ids] = NULL; | |
4825 | } | |
4826 | ||
4827 | static void init_node_nr_active(struct wq_node_nr_active *nna) | |
4828 | { | |
c5f8cd6c | 4829 | nna->max = WQ_DFL_MIN_ACTIVE; |
91ccc6e7 | 4830 | atomic_set(&nna->nr, 0); |
5797b1c1 TH |
4831 | raw_spin_lock_init(&nna->lock); |
4832 | INIT_LIST_HEAD(&nna->pending_pwqs); | |
91ccc6e7 TH |
4833 | } |
4834 | ||
4835 | /* | |
4836 | * Each node's nr_active counter will be accessed mostly from its own node and | |
4837 | * should be allocated in the node. | |
4838 | */ | |
4839 | static int alloc_node_nr_active(struct wq_node_nr_active **nna_ar) | |
4840 | { | |
4841 | struct wq_node_nr_active *nna; | |
4842 | int node; | |
4843 | ||
4844 | for_each_node(node) { | |
4845 | nna = kzalloc_node(sizeof(*nna), GFP_KERNEL, node); | |
4846 | if (!nna) | |
4847 | goto err_free; | |
4848 | init_node_nr_active(nna); | |
4849 | nna_ar[node] = nna; | |
4850 | } | |
4851 | ||
4852 | /* [nr_node_ids] is used as the fallback */ | |
4853 | nna = kzalloc_node(sizeof(*nna), GFP_KERNEL, NUMA_NO_NODE); | |
4854 | if (!nna) | |
4855 | goto err_free; | |
4856 | init_node_nr_active(nna); | |
4857 | nna_ar[nr_node_ids] = nna; | |
4858 | ||
4859 | return 0; | |
4860 | ||
4861 | err_free: | |
4862 | free_node_nr_active(nna_ar); | |
4863 | return -ENOMEM; | |
4864 | } | |
4865 | ||
6ba94429 | 4866 | static void rcu_free_wq(struct rcu_head *rcu) |
226223ab | 4867 | { |
6ba94429 FW |
4868 | struct workqueue_struct *wq = |
4869 | container_of(rcu, struct workqueue_struct, rcu); | |
226223ab | 4870 | |
91ccc6e7 TH |
4871 | if (wq->flags & WQ_UNBOUND) |
4872 | free_node_nr_active(wq->node_nr_active); | |
4873 | ||
669de8bd | 4874 | wq_free_lockdep(wq); |
636b927e TH |
4875 | free_percpu(wq->cpu_pwq); |
4876 | free_workqueue_attrs(wq->unbound_attrs); | |
6ba94429 | 4877 | kfree(wq); |
226223ab TH |
4878 | } |
4879 | ||
6ba94429 | 4880 | static void rcu_free_pool(struct rcu_head *rcu) |
226223ab | 4881 | { |
6ba94429 | 4882 | struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); |
226223ab | 4883 | |
6ba94429 FW |
4884 | ida_destroy(&pool->worker_ida); |
4885 | free_workqueue_attrs(pool->attrs); | |
4886 | kfree(pool); | |
226223ab TH |
4887 | } |
4888 | ||
6ba94429 FW |
4889 | /** |
4890 | * put_unbound_pool - put a worker_pool | |
4891 | * @pool: worker_pool to put | |
4892 | * | |
24acfb71 | 4893 | * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU |
6ba94429 FW |
4894 | * safe manner. get_unbound_pool() calls this function on its failure path |
4895 | * and this function should be able to release pools which went through, | |
4896 | * successfully or not, init_worker_pool(). | |
4897 | * | |
4898 | * Should be called with wq_pool_mutex held. | |
4899 | */ | |
4900 | static void put_unbound_pool(struct worker_pool *pool) | |
226223ab | 4901 | { |
6ba94429 | 4902 | struct worker *worker; |
9680540c | 4903 | LIST_HEAD(cull_list); |
e02b9312 | 4904 | |
6ba94429 | 4905 | lockdep_assert_held(&wq_pool_mutex); |
226223ab | 4906 | |
6ba94429 FW |
4907 | if (--pool->refcnt) |
4908 | return; | |
226223ab | 4909 | |
6ba94429 FW |
4910 | /* sanity checks */ |
4911 | if (WARN_ON(!(pool->cpu < 0)) || | |
4912 | WARN_ON(!list_empty(&pool->worklist))) | |
4913 | return; | |
226223ab | 4914 | |
6ba94429 FW |
4915 | /* release id and unhash */ |
4916 | if (pool->id >= 0) | |
4917 | idr_remove(&worker_pool_idr, pool->id); | |
4918 | hash_del(&pool->hash_node); | |
d55262c4 | 4919 | |
6ba94429 | 4920 | /* |
692b4825 TH |
4921 | * Become the manager and destroy all workers. This prevents |
4922 | * @pool's workers from blocking on attach_mutex. We're the last | |
4923 | * manager and @pool gets freed with the flag set. | |
9ab03be4 VS |
4924 | * |
4925 | * Having a concurrent manager is quite unlikely to happen as we can | |
4926 | * only get here with | |
4927 | * pwq->refcnt == pool->refcnt == 0 | |
4928 | * which implies no work queued to the pool, which implies no worker can | |
4929 | * become the manager. However a worker could have taken the role of | |
4930 | * manager before the refcnts dropped to 0, since maybe_create_worker() | |
4931 | * drops pool->lock | |
6ba94429 | 4932 | */ |
9ab03be4 VS |
4933 | while (true) { |
4934 | rcuwait_wait_event(&manager_wait, | |
4935 | !(pool->flags & POOL_MANAGER_ACTIVE), | |
4936 | TASK_UNINTERRUPTIBLE); | |
e02b9312 VS |
4937 | |
4938 | mutex_lock(&wq_pool_attach_mutex); | |
9ab03be4 VS |
4939 | raw_spin_lock_irq(&pool->lock); |
4940 | if (!(pool->flags & POOL_MANAGER_ACTIVE)) { | |
4941 | pool->flags |= POOL_MANAGER_ACTIVE; | |
4942 | break; | |
4943 | } | |
4944 | raw_spin_unlock_irq(&pool->lock); | |
e02b9312 | 4945 | mutex_unlock(&wq_pool_attach_mutex); |
9ab03be4 | 4946 | } |
692b4825 | 4947 | |
6ba94429 | 4948 | while ((worker = first_idle_worker(pool))) |
e02b9312 | 4949 | set_worker_dying(worker, &cull_list); |
6ba94429 | 4950 | WARN_ON(pool->nr_workers || pool->nr_idle); |
a9b8a985 | 4951 | raw_spin_unlock_irq(&pool->lock); |
d55262c4 | 4952 | |
f4b7b53c | 4953 | detach_dying_workers(&cull_list); |
e02b9312 | 4954 | |
1258fae7 | 4955 | mutex_unlock(&wq_pool_attach_mutex); |
226223ab | 4956 | |
68f83057 | 4957 | reap_dying_workers(&cull_list); |
226223ab | 4958 | |
6ba94429 FW |
4959 | /* shut down the timers */ |
4960 | del_timer_sync(&pool->idle_timer); | |
3f959aa3 | 4961 | cancel_work_sync(&pool->idle_cull_work); |
6ba94429 | 4962 | del_timer_sync(&pool->mayday_timer); |
226223ab | 4963 | |
24acfb71 | 4964 | /* RCU protected to allow dereferences from get_work_pool() */ |
25b00775 | 4965 | call_rcu(&pool->rcu, rcu_free_pool); |
226223ab TH |
4966 | } |
4967 | ||
4968 | /** | |
6ba94429 FW |
4969 | * get_unbound_pool - get a worker_pool with the specified attributes |
4970 | * @attrs: the attributes of the worker_pool to get | |
226223ab | 4971 | * |
6ba94429 FW |
4972 | * Obtain a worker_pool which has the same attributes as @attrs, bump the |
4973 | * reference count and return it. If there already is a matching | |
4974 | * worker_pool, it will be used; otherwise, this function attempts to | |
4975 | * create a new one. | |
226223ab | 4976 | * |
6ba94429 | 4977 | * Should be called with wq_pool_mutex held. |
226223ab | 4978 | * |
6ba94429 FW |
4979 | * Return: On success, a worker_pool with the same attributes as @attrs. |
4980 | * On failure, %NULL. | |
226223ab | 4981 | */ |
6ba94429 | 4982 | static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) |
226223ab | 4983 | { |
84193c07 | 4984 | struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_NUMA]; |
6ba94429 FW |
4985 | u32 hash = wqattrs_hash(attrs); |
4986 | struct worker_pool *pool; | |
84193c07 | 4987 | int pod, node = NUMA_NO_NODE; |
226223ab | 4988 | |
6ba94429 | 4989 | lockdep_assert_held(&wq_pool_mutex); |
226223ab | 4990 | |
6ba94429 FW |
4991 | /* do we already have a matching pool? */ |
4992 | hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { | |
4993 | if (wqattrs_equal(pool->attrs, attrs)) { | |
4994 | pool->refcnt++; | |
4995 | return pool; | |
4996 | } | |
4997 | } | |
226223ab | 4998 | |
9546b29e | 4999 | /* If __pod_cpumask is contained inside a NUMA pod, that's our node */ |
84193c07 | 5000 | for (pod = 0; pod < pt->nr_pods; pod++) { |
9546b29e | 5001 | if (cpumask_subset(attrs->__pod_cpumask, pt->pod_cpus[pod])) { |
84193c07 TH |
5002 | node = pt->pod_node[pod]; |
5003 | break; | |
e2273584 XP |
5004 | } |
5005 | } | |
5006 | ||
6ba94429 | 5007 | /* nope, create a new one */ |
84193c07 | 5008 | pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, node); |
6ba94429 FW |
5009 | if (!pool || init_worker_pool(pool) < 0) |
5010 | goto fail; | |
5011 | ||
84193c07 | 5012 | pool->node = node; |
5de7a03c TH |
5013 | copy_workqueue_attrs(pool->attrs, attrs); |
5014 | wqattrs_clear_for_pool(pool->attrs); | |
226223ab | 5015 | |
6ba94429 FW |
5016 | if (worker_pool_assign_id(pool) < 0) |
5017 | goto fail; | |
226223ab | 5018 | |
6ba94429 | 5019 | /* create and start the initial worker */ |
3347fa09 | 5020 | if (wq_online && !create_worker(pool)) |
6ba94429 | 5021 | goto fail; |
226223ab | 5022 | |
6ba94429 FW |
5023 | /* install */ |
5024 | hash_add(unbound_pool_hash, &pool->hash_node, hash); | |
226223ab | 5025 | |
6ba94429 FW |
5026 | return pool; |
5027 | fail: | |
5028 | if (pool) | |
5029 | put_unbound_pool(pool); | |
5030 | return NULL; | |
226223ab | 5031 | } |
226223ab | 5032 | |
6ba94429 | 5033 | /* |
967b494e TH |
5034 | * Scheduled on pwq_release_worker by put_pwq() when an unbound pwq hits zero |
5035 | * refcnt and needs to be destroyed. | |
7a4e344c | 5036 | */ |
687a9aa5 | 5037 | static void pwq_release_workfn(struct kthread_work *work) |
7a4e344c | 5038 | { |
6ba94429 | 5039 | struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, |
687a9aa5 | 5040 | release_work); |
6ba94429 FW |
5041 | struct workqueue_struct *wq = pwq->wq; |
5042 | struct worker_pool *pool = pwq->pool; | |
b42b0bdd | 5043 | bool is_last = false; |
7a4e344c | 5044 | |
b42b0bdd | 5045 | /* |
687a9aa5 | 5046 | * When @pwq is not linked, it doesn't hold any reference to the |
b42b0bdd YY |
5047 | * @wq, and @wq is invalid to access. |
5048 | */ | |
5049 | if (!list_empty(&pwq->pwqs_node)) { | |
b42b0bdd YY |
5050 | mutex_lock(&wq->mutex); |
5051 | list_del_rcu(&pwq->pwqs_node); | |
5052 | is_last = list_empty(&wq->pwqs); | |
4c065dbc WL |
5053 | |
5054 | /* | |
5055 | * For ordered workqueue with a plugged dfl_pwq, restart it now. | |
5056 | */ | |
5057 | if (!is_last && (wq->flags & __WQ_ORDERED)) | |
5058 | unplug_oldest_pwq(wq); | |
5059 | ||
b42b0bdd YY |
5060 | mutex_unlock(&wq->mutex); |
5061 | } | |
6ba94429 | 5062 | |
687a9aa5 TH |
5063 | if (wq->flags & WQ_UNBOUND) { |
5064 | mutex_lock(&wq_pool_mutex); | |
5065 | put_unbound_pool(pool); | |
5066 | mutex_unlock(&wq_pool_mutex); | |
5067 | } | |
6ba94429 | 5068 | |
5797b1c1 TH |
5069 | if (!list_empty(&pwq->pending_node)) { |
5070 | struct wq_node_nr_active *nna = | |
5071 | wq_node_nr_active(pwq->wq, pwq->pool->node); | |
5072 | ||
5073 | raw_spin_lock_irq(&nna->lock); | |
5074 | list_del_init(&pwq->pending_node); | |
5075 | raw_spin_unlock_irq(&nna->lock); | |
5076 | } | |
5077 | ||
37c2277f | 5078 | kfree_rcu(pwq, rcu); |
7a4e344c | 5079 | |
2865a8fb | 5080 | /* |
6ba94429 FW |
5081 | * If we're the last pwq going away, @wq is already dead and no one |
5082 | * is gonna access it anymore. Schedule RCU free. | |
2865a8fb | 5083 | */ |
669de8bd BVA |
5084 | if (is_last) { |
5085 | wq_unregister_lockdep(wq); | |
25b00775 | 5086 | call_rcu(&wq->rcu, rcu_free_wq); |
669de8bd | 5087 | } |
29c91e99 TH |
5088 | } |
5089 | ||
67dc8325 | 5090 | /* initialize newly allocated @pwq which is associated with @wq and @pool */ |
6ba94429 FW |
5091 | static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, |
5092 | struct worker_pool *pool) | |
29c91e99 | 5093 | { |
e9a8e01f | 5094 | BUG_ON((unsigned long)pwq & ~WORK_STRUCT_PWQ_MASK); |
29c91e99 | 5095 | |
6ba94429 FW |
5096 | memset(pwq, 0, sizeof(*pwq)); |
5097 | ||
5098 | pwq->pool = pool; | |
5099 | pwq->wq = wq; | |
5100 | pwq->flush_color = -1; | |
5101 | pwq->refcnt = 1; | |
f97a4a1a | 5102 | INIT_LIST_HEAD(&pwq->inactive_works); |
5797b1c1 | 5103 | INIT_LIST_HEAD(&pwq->pending_node); |
6ba94429 FW |
5104 | INIT_LIST_HEAD(&pwq->pwqs_node); |
5105 | INIT_LIST_HEAD(&pwq->mayday_node); | |
687a9aa5 | 5106 | kthread_init_work(&pwq->release_work, pwq_release_workfn); |
29c91e99 TH |
5107 | } |
5108 | ||
6ba94429 FW |
5109 | /* sync @pwq with the current state of its associated wq and link it */ |
5110 | static void link_pwq(struct pool_workqueue *pwq) | |
29c91e99 | 5111 | { |
6ba94429 | 5112 | struct workqueue_struct *wq = pwq->wq; |
29c91e99 | 5113 | |
6ba94429 | 5114 | lockdep_assert_held(&wq->mutex); |
a892cacc | 5115 | |
6ba94429 FW |
5116 | /* may be called multiple times, ignore if already linked */ |
5117 | if (!list_empty(&pwq->pwqs_node)) | |
29c91e99 | 5118 | return; |
29c91e99 | 5119 | |
6ba94429 FW |
5120 | /* set the matching work_color */ |
5121 | pwq->work_color = wq->work_color; | |
29c91e99 | 5122 | |
6ba94429 | 5123 | /* link in @pwq */ |
26fb7e3d | 5124 | list_add_tail_rcu(&pwq->pwqs_node, &wq->pwqs); |
6ba94429 | 5125 | } |
29c91e99 | 5126 | |
6ba94429 FW |
5127 | /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ |
5128 | static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, | |
5129 | const struct workqueue_attrs *attrs) | |
5130 | { | |
5131 | struct worker_pool *pool; | |
5132 | struct pool_workqueue *pwq; | |
60f5a4bc | 5133 | |
6ba94429 | 5134 | lockdep_assert_held(&wq_pool_mutex); |
60f5a4bc | 5135 | |
6ba94429 FW |
5136 | pool = get_unbound_pool(attrs); |
5137 | if (!pool) | |
5138 | return NULL; | |
60f5a4bc | 5139 | |
6ba94429 FW |
5140 | pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); |
5141 | if (!pwq) { | |
5142 | put_unbound_pool(pool); | |
5143 | return NULL; | |
5144 | } | |
29c91e99 | 5145 | |
6ba94429 FW |
5146 | init_pwq(pwq, wq, pool); |
5147 | return pwq; | |
5148 | } | |
29c91e99 | 5149 | |
1726a171 LJ |
5150 | static void apply_wqattrs_lock(void) |
5151 | { | |
1726a171 LJ |
5152 | mutex_lock(&wq_pool_mutex); |
5153 | } | |
5154 | ||
5155 | static void apply_wqattrs_unlock(void) | |
5156 | { | |
5157 | mutex_unlock(&wq_pool_mutex); | |
1726a171 LJ |
5158 | } |
5159 | ||
29c91e99 | 5160 | /** |
fef59c9c | 5161 | * wq_calc_pod_cpumask - calculate a wq_attrs' cpumask for a pod |
042f7df1 | 5162 | * @attrs: the wq_attrs of the default pwq of the target workqueue |
84193c07 | 5163 | * @cpu: the target CPU |
29c91e99 | 5164 | * |
88a41b18 | 5165 | * Calculate the cpumask a workqueue with @attrs should use on @pod. |
9546b29e | 5166 | * The result is stored in @attrs->__pod_cpumask. |
a892cacc | 5167 | * |
fef59c9c TH |
5168 | * If pod affinity is not enabled, @attrs->cpumask is always used. If enabled |
5169 | * and @pod has online CPUs requested by @attrs, the returned cpumask is the | |
5170 | * intersection of the possible CPUs of @pod and @attrs->cpumask. | |
d185af30 | 5171 | * |
fef59c9c | 5172 | * The caller is responsible for ensuring that the cpumask of @pod stays stable. |
29c91e99 | 5173 | */ |
88a41b18 | 5174 | static void wq_calc_pod_cpumask(struct workqueue_attrs *attrs, int cpu) |
29c91e99 | 5175 | { |
84193c07 TH |
5176 | const struct wq_pod_type *pt = wqattrs_pod_type(attrs); |
5177 | int pod = pt->cpu_pod[cpu]; | |
29c91e99 | 5178 | |
fbb3d4c1 | 5179 | /* calculate possible CPUs in @pod that @attrs wants */ |
9546b29e | 5180 | cpumask_and(attrs->__pod_cpumask, pt->pod_cpus[pod], attrs->cpumask); |
fbb3d4c1 LJ |
5181 | /* does @pod have any online CPUs @attrs wants? */ |
5182 | if (!cpumask_intersects(attrs->__pod_cpumask, wq_online_cpumask)) { | |
9546b29e | 5183 | cpumask_copy(attrs->__pod_cpumask, attrs->cpumask); |
84193c07 TH |
5184 | return; |
5185 | } | |
4c16bd32 TH |
5186 | } |
5187 | ||
9f66cff2 | 5188 | /* install @pwq into @wq and return the old pwq, @cpu < 0 for dfl_pwq */ |
636b927e TH |
5189 | static struct pool_workqueue *install_unbound_pwq(struct workqueue_struct *wq, |
5190 | int cpu, struct pool_workqueue *pwq) | |
1befcf30 | 5191 | { |
9f66cff2 | 5192 | struct pool_workqueue __rcu **slot = unbound_pwq_slot(wq, cpu); |
1befcf30 TH |
5193 | struct pool_workqueue *old_pwq; |
5194 | ||
5b95e1af | 5195 | lockdep_assert_held(&wq_pool_mutex); |
1befcf30 TH |
5196 | lockdep_assert_held(&wq->mutex); |
5197 | ||
5198 | /* link_pwq() can handle duplicate calls */ | |
5199 | link_pwq(pwq); | |
5200 | ||
9f66cff2 TH |
5201 | old_pwq = rcu_access_pointer(*slot); |
5202 | rcu_assign_pointer(*slot, pwq); | |
1befcf30 TH |
5203 | return old_pwq; |
5204 | } | |
5205 | ||
2d5f0764 LJ |
5206 | /* context to store the prepared attrs & pwqs before applying */ |
5207 | struct apply_wqattrs_ctx { | |
5208 | struct workqueue_struct *wq; /* target workqueue */ | |
5209 | struct workqueue_attrs *attrs; /* attrs to apply */ | |
042f7df1 | 5210 | struct list_head list; /* queued for batching commit */ |
2d5f0764 LJ |
5211 | struct pool_workqueue *dfl_pwq; |
5212 | struct pool_workqueue *pwq_tbl[]; | |
5213 | }; | |
5214 | ||
5215 | /* free the resources after success or abort */ | |
5216 | static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx) | |
5217 | { | |
5218 | if (ctx) { | |
636b927e | 5219 | int cpu; |
2d5f0764 | 5220 | |
636b927e TH |
5221 | for_each_possible_cpu(cpu) |
5222 | put_pwq_unlocked(ctx->pwq_tbl[cpu]); | |
2d5f0764 LJ |
5223 | put_pwq_unlocked(ctx->dfl_pwq); |
5224 | ||
5225 | free_workqueue_attrs(ctx->attrs); | |
5226 | ||
5227 | kfree(ctx); | |
5228 | } | |
5229 | } | |
5230 | ||
5231 | /* allocate the attrs and pwqs for later installation */ | |
5232 | static struct apply_wqattrs_ctx * | |
5233 | apply_wqattrs_prepare(struct workqueue_struct *wq, | |
99c621ef LJ |
5234 | const struct workqueue_attrs *attrs, |
5235 | const cpumask_var_t unbound_cpumask) | |
9e8cd2f5 | 5236 | { |
2d5f0764 | 5237 | struct apply_wqattrs_ctx *ctx; |
9546b29e | 5238 | struct workqueue_attrs *new_attrs; |
636b927e | 5239 | int cpu; |
9e8cd2f5 | 5240 | |
2d5f0764 | 5241 | lockdep_assert_held(&wq_pool_mutex); |
9e8cd2f5 | 5242 | |
84193c07 TH |
5243 | if (WARN_ON(attrs->affn_scope < 0 || |
5244 | attrs->affn_scope >= WQ_AFFN_NR_TYPES)) | |
5245 | return ERR_PTR(-EINVAL); | |
5246 | ||
636b927e | 5247 | ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_cpu_ids), GFP_KERNEL); |
8719dcea | 5248 | |
be69d00d | 5249 | new_attrs = alloc_workqueue_attrs(); |
9546b29e | 5250 | if (!ctx || !new_attrs) |
2d5f0764 | 5251 | goto out_free; |
13e2e556 | 5252 | |
4c16bd32 TH |
5253 | /* |
5254 | * If something goes wrong during CPU up/down, we'll fall back to | |
5255 | * the default pwq covering whole @attrs->cpumask. Always create | |
5256 | * it even if we don't use it immediately. | |
5257 | */ | |
0f36ee24 TH |
5258 | copy_workqueue_attrs(new_attrs, attrs); |
5259 | wqattrs_actualize_cpumask(new_attrs, unbound_cpumask); | |
9546b29e | 5260 | cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask); |
2d5f0764 LJ |
5261 | ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs); |
5262 | if (!ctx->dfl_pwq) | |
5263 | goto out_free; | |
4c16bd32 | 5264 | |
636b927e | 5265 | for_each_possible_cpu(cpu) { |
af73f5c9 | 5266 | if (new_attrs->ordered) { |
2d5f0764 | 5267 | ctx->dfl_pwq->refcnt++; |
636b927e TH |
5268 | ctx->pwq_tbl[cpu] = ctx->dfl_pwq; |
5269 | } else { | |
88a41b18 | 5270 | wq_calc_pod_cpumask(new_attrs, cpu); |
9546b29e | 5271 | ctx->pwq_tbl[cpu] = alloc_unbound_pwq(wq, new_attrs); |
636b927e TH |
5272 | if (!ctx->pwq_tbl[cpu]) |
5273 | goto out_free; | |
4c16bd32 TH |
5274 | } |
5275 | } | |
5276 | ||
042f7df1 LJ |
5277 | /* save the user configured attrs and sanitize it. */ |
5278 | copy_workqueue_attrs(new_attrs, attrs); | |
5279 | cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); | |
9546b29e | 5280 | cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask); |
2d5f0764 | 5281 | ctx->attrs = new_attrs; |
042f7df1 | 5282 | |
4c065dbc WL |
5283 | /* |
5284 | * For initialized ordered workqueues, there should only be one pwq | |
5285 | * (dfl_pwq). Set the plugged flag of ctx->dfl_pwq to suspend execution | |
5286 | * of newly queued work items until execution of older work items in | |
5287 | * the old pwq's have completed. | |
5288 | */ | |
5289 | if ((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs)) | |
5290 | ctx->dfl_pwq->plugged = true; | |
5291 | ||
2d5f0764 | 5292 | ctx->wq = wq; |
2d5f0764 LJ |
5293 | return ctx; |
5294 | ||
5295 | out_free: | |
2d5f0764 LJ |
5296 | free_workqueue_attrs(new_attrs); |
5297 | apply_wqattrs_cleanup(ctx); | |
84193c07 | 5298 | return ERR_PTR(-ENOMEM); |
2d5f0764 LJ |
5299 | } |
5300 | ||
5301 | /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */ | |
5302 | static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx) | |
5303 | { | |
636b927e | 5304 | int cpu; |
9e8cd2f5 | 5305 | |
4c16bd32 | 5306 | /* all pwqs have been created successfully, let's install'em */ |
2d5f0764 | 5307 | mutex_lock(&ctx->wq->mutex); |
a892cacc | 5308 | |
2d5f0764 | 5309 | copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs); |
4c16bd32 | 5310 | |
9f66cff2 | 5311 | /* save the previous pwqs and install the new ones */ |
636b927e TH |
5312 | for_each_possible_cpu(cpu) |
5313 | ctx->pwq_tbl[cpu] = install_unbound_pwq(ctx->wq, cpu, | |
5314 | ctx->pwq_tbl[cpu]); | |
9f66cff2 | 5315 | ctx->dfl_pwq = install_unbound_pwq(ctx->wq, -1, ctx->dfl_pwq); |
f147f29e | 5316 | |
5797b1c1 TH |
5317 | /* update node_nr_active->max */ |
5318 | wq_update_node_max_active(ctx->wq, -1); | |
5319 | ||
d64f2fa0 JL |
5320 | /* rescuer needs to respect wq cpumask changes */ |
5321 | if (ctx->wq->rescuer) | |
5322 | set_cpus_allowed_ptr(ctx->wq->rescuer->task, | |
5323 | unbound_effective_cpumask(ctx->wq)); | |
5324 | ||
2d5f0764 LJ |
5325 | mutex_unlock(&ctx->wq->mutex); |
5326 | } | |
9e8cd2f5 | 5327 | |
a0111cf6 LJ |
5328 | static int apply_workqueue_attrs_locked(struct workqueue_struct *wq, |
5329 | const struct workqueue_attrs *attrs) | |
2d5f0764 LJ |
5330 | { |
5331 | struct apply_wqattrs_ctx *ctx; | |
4c16bd32 | 5332 | |
2d5f0764 LJ |
5333 | /* only unbound workqueues can change attributes */ |
5334 | if (WARN_ON(!(wq->flags & WQ_UNBOUND))) | |
5335 | return -EINVAL; | |
13e2e556 | 5336 | |
99c621ef | 5337 | ctx = apply_wqattrs_prepare(wq, attrs, wq_unbound_cpumask); |
84193c07 TH |
5338 | if (IS_ERR(ctx)) |
5339 | return PTR_ERR(ctx); | |
2d5f0764 LJ |
5340 | |
5341 | /* the ctx has been prepared successfully, let's commit it */ | |
6201171e | 5342 | apply_wqattrs_commit(ctx); |
2d5f0764 LJ |
5343 | apply_wqattrs_cleanup(ctx); |
5344 | ||
6201171e | 5345 | return 0; |
9e8cd2f5 TH |
5346 | } |
5347 | ||
a0111cf6 LJ |
5348 | /** |
5349 | * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue | |
5350 | * @wq: the target workqueue | |
5351 | * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() | |
5352 | * | |
fef59c9c TH |
5353 | * Apply @attrs to an unbound workqueue @wq. Unless disabled, this function maps |
5354 | * a separate pwq to each CPU pod with possibles CPUs in @attrs->cpumask so that | |
5355 | * work items are affine to the pod it was issued on. Older pwqs are released as | |
5356 | * in-flight work items finish. Note that a work item which repeatedly requeues | |
5357 | * itself back-to-back will stay on its current pwq. | |
a0111cf6 LJ |
5358 | * |
5359 | * Performs GFP_KERNEL allocations. | |
5360 | * | |
5361 | * Return: 0 on success and -errno on failure. | |
5362 | */ | |
513c98d0 | 5363 | int apply_workqueue_attrs(struct workqueue_struct *wq, |
a0111cf6 LJ |
5364 | const struct workqueue_attrs *attrs) |
5365 | { | |
5366 | int ret; | |
5367 | ||
509b3204 | 5368 | mutex_lock(&wq_pool_mutex); |
a0111cf6 | 5369 | ret = apply_workqueue_attrs_locked(wq, attrs); |
509b3204 | 5370 | mutex_unlock(&wq_pool_mutex); |
a0111cf6 LJ |
5371 | |
5372 | return ret; | |
5373 | } | |
5374 | ||
4c16bd32 | 5375 | /** |
b2b1f933 | 5376 | * unbound_wq_update_pwq - update a pwq slot for CPU hot[un]plug |
4c16bd32 | 5377 | * @wq: the target workqueue |
b2b1f933 | 5378 | * @cpu: the CPU to update the pwq slot for |
4c16bd32 TH |
5379 | * |
5380 | * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and | |
b2b1f933 | 5381 | * %CPU_DOWN_FAILED. @cpu is in the same pod of the CPU being hot[un]plugged. |
4c16bd32 | 5382 | * |
fef59c9c TH |
5383 | * |
5384 | * If pod affinity can't be adjusted due to memory allocation failure, it falls | |
5385 | * back to @wq->dfl_pwq which may not be optimal but is always correct. | |
5386 | * | |
5387 | * Note that when the last allowed CPU of a pod goes offline for a workqueue | |
5388 | * with a cpumask spanning multiple pods, the workers which were already | |
5389 | * executing the work items for the workqueue will lose their CPU affinity and | |
5390 | * may execute on any CPU. This is similar to how per-cpu workqueues behave on | |
5391 | * CPU_DOWN. If a workqueue user wants strict affinity, it's the user's | |
5392 | * responsibility to flush the work item from CPU_DOWN_PREPARE. | |
4c16bd32 | 5393 | */ |
b2b1f933 | 5394 | static void unbound_wq_update_pwq(struct workqueue_struct *wq, int cpu) |
4c16bd32 | 5395 | { |
4c16bd32 TH |
5396 | struct pool_workqueue *old_pwq = NULL, *pwq; |
5397 | struct workqueue_attrs *target_attrs; | |
4c16bd32 TH |
5398 | |
5399 | lockdep_assert_held(&wq_pool_mutex); | |
5400 | ||
84193c07 | 5401 | if (!(wq->flags & WQ_UNBOUND) || wq->unbound_attrs->ordered) |
4c16bd32 TH |
5402 | return; |
5403 | ||
5404 | /* | |
5405 | * We don't wanna alloc/free wq_attrs for each wq for each CPU. | |
5406 | * Let's use a preallocated one. The following buf is protected by | |
5407 | * CPU hotplug exclusion. | |
5408 | */ | |
b2b1f933 | 5409 | target_attrs = unbound_wq_update_pwq_attrs_buf; |
4c16bd32 | 5410 | |
4c16bd32 | 5411 | copy_workqueue_attrs(target_attrs, wq->unbound_attrs); |
0f36ee24 | 5412 | wqattrs_actualize_cpumask(target_attrs, wq_unbound_cpumask); |
4c16bd32 | 5413 | |
636b927e | 5414 | /* nothing to do if the target cpumask matches the current pwq */ |
88a41b18 | 5415 | wq_calc_pod_cpumask(target_attrs, cpu); |
9f66cff2 | 5416 | if (wqattrs_equal(target_attrs, unbound_pwq(wq, cpu)->pool->attrs)) |
636b927e | 5417 | return; |
4c16bd32 | 5418 | |
4c16bd32 TH |
5419 | /* create a new pwq */ |
5420 | pwq = alloc_unbound_pwq(wq, target_attrs); | |
5421 | if (!pwq) { | |
fef59c9c | 5422 | pr_warn("workqueue: allocation failed while updating CPU pod affinity of \"%s\"\n", |
2d916033 | 5423 | wq->name); |
77f300b1 | 5424 | goto use_dfl_pwq; |
4c16bd32 TH |
5425 | } |
5426 | ||
f7142ed4 | 5427 | /* Install the new pwq. */ |
4c16bd32 | 5428 | mutex_lock(&wq->mutex); |
636b927e | 5429 | old_pwq = install_unbound_pwq(wq, cpu, pwq); |
4c16bd32 TH |
5430 | goto out_unlock; |
5431 | ||
5432 | use_dfl_pwq: | |
f7142ed4 | 5433 | mutex_lock(&wq->mutex); |
9f66cff2 TH |
5434 | pwq = unbound_pwq(wq, -1); |
5435 | raw_spin_lock_irq(&pwq->pool->lock); | |
5436 | get_pwq(pwq); | |
5437 | raw_spin_unlock_irq(&pwq->pool->lock); | |
5438 | old_pwq = install_unbound_pwq(wq, cpu, pwq); | |
4c16bd32 TH |
5439 | out_unlock: |
5440 | mutex_unlock(&wq->mutex); | |
5441 | put_pwq_unlocked(old_pwq); | |
5442 | } | |
5443 | ||
30cdf249 | 5444 | static int alloc_and_link_pwqs(struct workqueue_struct *wq) |
0f900049 | 5445 | { |
49e3cf44 | 5446 | bool highpri = wq->flags & WQ_HIGHPRI; |
8a2b7538 | 5447 | int cpu, ret; |
30cdf249 | 5448 | |
1726a171 LJ |
5449 | lockdep_assert_held(&wq_pool_mutex); |
5450 | ||
636b927e TH |
5451 | wq->cpu_pwq = alloc_percpu(struct pool_workqueue *); |
5452 | if (!wq->cpu_pwq) | |
5453 | goto enomem; | |
30cdf249 | 5454 | |
636b927e | 5455 | if (!(wq->flags & WQ_UNBOUND)) { |
7ccc2151 WH |
5456 | struct worker_pool __percpu *pools; |
5457 | ||
5458 | if (wq->flags & WQ_BH) | |
5459 | pools = bh_worker_pools; | |
5460 | else | |
5461 | pools = cpu_worker_pools; | |
5462 | ||
30cdf249 | 5463 | for_each_possible_cpu(cpu) { |
4cb1ef64 | 5464 | struct pool_workqueue **pwq_p; |
4cb1ef64 TH |
5465 | struct worker_pool *pool; |
5466 | ||
4cb1ef64 TH |
5467 | pool = &(per_cpu_ptr(pools, cpu)[highpri]); |
5468 | pwq_p = per_cpu_ptr(wq->cpu_pwq, cpu); | |
687a9aa5 TH |
5469 | |
5470 | *pwq_p = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, | |
5471 | pool->node); | |
5472 | if (!*pwq_p) | |
5473 | goto enomem; | |
f3421797 | 5474 | |
687a9aa5 | 5475 | init_pwq(*pwq_p, wq, pool); |
f147f29e TH |
5476 | |
5477 | mutex_lock(&wq->mutex); | |
687a9aa5 | 5478 | link_pwq(*pwq_p); |
f147f29e | 5479 | mutex_unlock(&wq->mutex); |
30cdf249 | 5480 | } |
9e8cd2f5 | 5481 | return 0; |
509b3204 DJ |
5482 | } |
5483 | ||
509b3204 | 5484 | if (wq->flags & __WQ_ORDERED) { |
9f66cff2 TH |
5485 | struct pool_workqueue *dfl_pwq; |
5486 | ||
1726a171 | 5487 | ret = apply_workqueue_attrs_locked(wq, ordered_wq_attrs[highpri]); |
8a2b7538 | 5488 | /* there should only be single pwq for ordering guarantee */ |
9f66cff2 TH |
5489 | dfl_pwq = rcu_access_pointer(wq->dfl_pwq); |
5490 | WARN(!ret && (wq->pwqs.next != &dfl_pwq->pwqs_node || | |
5491 | wq->pwqs.prev != &dfl_pwq->pwqs_node), | |
8a2b7538 | 5492 | "ordering guarantee broken for workqueue %s\n", wq->name); |
30cdf249 | 5493 | } else { |
1726a171 | 5494 | ret = apply_workqueue_attrs_locked(wq, unbound_std_wq_attrs[highpri]); |
30cdf249 | 5495 | } |
64344553 | 5496 | |
509b3204 | 5497 | return ret; |
687a9aa5 TH |
5498 | |
5499 | enomem: | |
5500 | if (wq->cpu_pwq) { | |
7b42f401 Z |
5501 | for_each_possible_cpu(cpu) { |
5502 | struct pool_workqueue *pwq = *per_cpu_ptr(wq->cpu_pwq, cpu); | |
5503 | ||
5504 | if (pwq) | |
5505 | kmem_cache_free(pwq_cache, pwq); | |
5506 | } | |
687a9aa5 TH |
5507 | free_percpu(wq->cpu_pwq); |
5508 | wq->cpu_pwq = NULL; | |
5509 | } | |
5510 | return -ENOMEM; | |
0f900049 TH |
5511 | } |
5512 | ||
f3421797 TH |
5513 | static int wq_clamp_max_active(int max_active, unsigned int flags, |
5514 | const char *name) | |
b71ab8c2 | 5515 | { |
636b927e | 5516 | if (max_active < 1 || max_active > WQ_MAX_ACTIVE) |
044c782c | 5517 | pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", |
636b927e | 5518 | max_active, name, 1, WQ_MAX_ACTIVE); |
b71ab8c2 | 5519 | |
636b927e | 5520 | return clamp_val(max_active, 1, WQ_MAX_ACTIVE); |
b71ab8c2 TH |
5521 | } |
5522 | ||
983c7515 TH |
5523 | /* |
5524 | * Workqueues which may be used during memory reclaim should have a rescuer | |
5525 | * to guarantee forward progress. | |
5526 | */ | |
5527 | static int init_rescuer(struct workqueue_struct *wq) | |
5528 | { | |
5529 | struct worker *rescuer; | |
2a1b02bc | 5530 | char id_buf[WORKER_ID_LEN]; |
b92b36ea | 5531 | int ret; |
983c7515 | 5532 | |
449b31ad LJ |
5533 | lockdep_assert_held(&wq_pool_mutex); |
5534 | ||
983c7515 TH |
5535 | if (!(wq->flags & WQ_MEM_RECLAIM)) |
5536 | return 0; | |
5537 | ||
5538 | rescuer = alloc_worker(NUMA_NO_NODE); | |
4c0736a7 PM |
5539 | if (!rescuer) { |
5540 | pr_err("workqueue: Failed to allocate a rescuer for wq \"%s\"\n", | |
5541 | wq->name); | |
983c7515 | 5542 | return -ENOMEM; |
4c0736a7 | 5543 | } |
983c7515 TH |
5544 | |
5545 | rescuer->rescue_wq = wq; | |
2a1b02bc TH |
5546 | format_worker_id(id_buf, sizeof(id_buf), rescuer, NULL); |
5547 | ||
5548 | rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", id_buf); | |
f187b697 | 5549 | if (IS_ERR(rescuer->task)) { |
b92b36ea | 5550 | ret = PTR_ERR(rescuer->task); |
4c0736a7 PM |
5551 | pr_err("workqueue: Failed to create a rescuer kthread for wq \"%s\": %pe", |
5552 | wq->name, ERR_PTR(ret)); | |
983c7515 | 5553 | kfree(rescuer); |
b92b36ea | 5554 | return ret; |
983c7515 TH |
5555 | } |
5556 | ||
5557 | wq->rescuer = rescuer; | |
85f0ab43 | 5558 | if (wq->flags & WQ_UNBOUND) |
449b31ad | 5559 | kthread_bind_mask(rescuer->task, unbound_effective_cpumask(wq)); |
85f0ab43 JL |
5560 | else |
5561 | kthread_bind_mask(rescuer->task, cpu_possible_mask); | |
983c7515 TH |
5562 | wake_up_process(rescuer->task); |
5563 | ||
5564 | return 0; | |
5565 | } | |
5566 | ||
a045a272 TH |
5567 | /** |
5568 | * wq_adjust_max_active - update a wq's max_active to the current setting | |
5569 | * @wq: target workqueue | |
5570 | * | |
5571 | * If @wq isn't freezing, set @wq->max_active to the saved_max_active and | |
5572 | * activate inactive work items accordingly. If @wq is freezing, clear | |
5573 | * @wq->max_active to zero. | |
5574 | */ | |
5575 | static void wq_adjust_max_active(struct workqueue_struct *wq) | |
5576 | { | |
c5404d4e | 5577 | bool activated; |
5797b1c1 | 5578 | int new_max, new_min; |
a045a272 TH |
5579 | |
5580 | lockdep_assert_held(&wq->mutex); | |
5581 | ||
5582 | if ((wq->flags & WQ_FREEZABLE) && workqueue_freezing) { | |
5797b1c1 TH |
5583 | new_max = 0; |
5584 | new_min = 0; | |
5585 | } else { | |
5586 | new_max = wq->saved_max_active; | |
5587 | new_min = wq->saved_min_active; | |
a045a272 TH |
5588 | } |
5589 | ||
5797b1c1 | 5590 | if (wq->max_active == new_max && wq->min_active == new_min) |
a045a272 TH |
5591 | return; |
5592 | ||
5593 | /* | |
5797b1c1 | 5594 | * Update @wq->max/min_active and then kick inactive work items if more |
a045a272 TH |
5595 | * active work items are allowed. This doesn't break work item ordering |
5596 | * because new work items are always queued behind existing inactive | |
5597 | * work items if there are any. | |
5598 | */ | |
5797b1c1 TH |
5599 | WRITE_ONCE(wq->max_active, new_max); |
5600 | WRITE_ONCE(wq->min_active, new_min); | |
5601 | ||
5602 | if (wq->flags & WQ_UNBOUND) | |
5603 | wq_update_node_max_active(wq, -1); | |
5604 | ||
5605 | if (new_max == 0) | |
5606 | return; | |
a045a272 | 5607 | |
c5404d4e TH |
5608 | /* |
5609 | * Round-robin through pwq's activating the first inactive work item | |
5610 | * until max_active is filled. | |
5611 | */ | |
5612 | do { | |
5613 | struct pool_workqueue *pwq; | |
a045a272 | 5614 | |
c5404d4e TH |
5615 | activated = false; |
5616 | for_each_pwq(pwq, wq) { | |
c26e2f2e | 5617 | unsigned long irq_flags; |
a045a272 | 5618 | |
c5404d4e | 5619 | /* can be called during early boot w/ irq disabled */ |
c26e2f2e | 5620 | raw_spin_lock_irqsave(&pwq->pool->lock, irq_flags); |
5797b1c1 | 5621 | if (pwq_activate_first_inactive(pwq, true)) { |
c5404d4e TH |
5622 | activated = true; |
5623 | kick_pool(pwq->pool); | |
5624 | } | |
c26e2f2e | 5625 | raw_spin_unlock_irqrestore(&pwq->pool->lock, irq_flags); |
c5404d4e TH |
5626 | } |
5627 | } while (activated); | |
a045a272 TH |
5628 | } |
5629 | ||
b188c57a MB |
5630 | static struct workqueue_struct *__alloc_workqueue(const char *fmt, |
5631 | unsigned int flags, | |
9b59a85a | 5632 | int max_active, va_list args) |
1da177e4 | 5633 | { |
1da177e4 | 5634 | struct workqueue_struct *wq; |
91ccc6e7 TH |
5635 | size_t wq_size; |
5636 | int name_len; | |
b196be89 | 5637 | |
4cb1ef64 TH |
5638 | if (flags & WQ_BH) { |
5639 | if (WARN_ON_ONCE(flags & ~__WQ_BH_ALLOWS)) | |
5640 | return NULL; | |
5641 | if (WARN_ON_ONCE(max_active)) | |
5642 | return NULL; | |
5643 | } | |
5644 | ||
cee22a15 VK |
5645 | /* see the comment above the definition of WQ_POWER_EFFICIENT */ |
5646 | if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) | |
5647 | flags |= WQ_UNBOUND; | |
5648 | ||
ecf6881f | 5649 | /* allocate wq and format name */ |
91ccc6e7 TH |
5650 | if (flags & WQ_UNBOUND) |
5651 | wq_size = struct_size(wq, node_nr_active, nr_node_ids + 1); | |
5652 | else | |
5653 | wq_size = sizeof(*wq); | |
5654 | ||
5655 | wq = kzalloc(wq_size, GFP_KERNEL); | |
b196be89 | 5656 | if (!wq) |
d2c1d404 | 5657 | return NULL; |
b196be89 | 5658 | |
6029a918 | 5659 | if (flags & WQ_UNBOUND) { |
be69d00d | 5660 | wq->unbound_attrs = alloc_workqueue_attrs(); |
6029a918 TH |
5661 | if (!wq->unbound_attrs) |
5662 | goto err_free_wq; | |
5663 | } | |
5664 | ||
91ccc6e7 | 5665 | name_len = vsnprintf(wq->name, sizeof(wq->name), fmt, args); |
1da177e4 | 5666 | |
91ccc6e7 TH |
5667 | if (name_len >= WQ_NAME_LEN) |
5668 | pr_warn_once("workqueue: name exceeds WQ_NAME_LEN. Truncating to: %s\n", | |
5669 | wq->name); | |
31c89007 | 5670 | |
4cb1ef64 TH |
5671 | if (flags & WQ_BH) { |
5672 | /* | |
5673 | * BH workqueues always share a single execution context per CPU | |
5674 | * and don't impose any max_active limit. | |
5675 | */ | |
5676 | max_active = INT_MAX; | |
5677 | } else { | |
5678 | max_active = max_active ?: WQ_DFL_ACTIVE; | |
5679 | max_active = wq_clamp_max_active(max_active, flags, wq->name); | |
5680 | } | |
3af24433 | 5681 | |
b196be89 | 5682 | /* init wq */ |
97e37d7b | 5683 | wq->flags = flags; |
a045a272 | 5684 | wq->max_active = max_active; |
5797b1c1 TH |
5685 | wq->min_active = min(max_active, WQ_DFL_MIN_ACTIVE); |
5686 | wq->saved_max_active = wq->max_active; | |
5687 | wq->saved_min_active = wq->min_active; | |
3c25a55d | 5688 | mutex_init(&wq->mutex); |
112202d9 | 5689 | atomic_set(&wq->nr_pwqs_to_flush, 0); |
30cdf249 | 5690 | INIT_LIST_HEAD(&wq->pwqs); |
73f53c4a TH |
5691 | INIT_LIST_HEAD(&wq->flusher_queue); |
5692 | INIT_LIST_HEAD(&wq->flusher_overflow); | |
493a1724 | 5693 | INIT_LIST_HEAD(&wq->maydays); |
502ca9d8 | 5694 | |
cce1a165 | 5695 | INIT_LIST_HEAD(&wq->list); |
3af24433 | 5696 | |
91ccc6e7 TH |
5697 | if (flags & WQ_UNBOUND) { |
5698 | if (alloc_node_nr_active(wq->node_nr_active) < 0) | |
b188c57a | 5699 | goto err_free_wq; |
91ccc6e7 TH |
5700 | } |
5701 | ||
a0a1a5fd | 5702 | /* |
1726a171 | 5703 | * wq_pool_mutex protects the workqueues list, allocations of PWQs, |
aa868475 | 5704 | * and the global freeze state. |
a0a1a5fd | 5705 | */ |
1726a171 LJ |
5706 | apply_wqattrs_lock(); |
5707 | ||
5708 | if (alloc_and_link_pwqs(wq) < 0) | |
5709 | goto err_unlock_free_node_nr_active; | |
a0a1a5fd | 5710 | |
a357fc03 | 5711 | mutex_lock(&wq->mutex); |
a045a272 | 5712 | wq_adjust_max_active(wq); |
a357fc03 | 5713 | mutex_unlock(&wq->mutex); |
a0a1a5fd | 5714 | |
e2dca7ad | 5715 | list_add_tail_rcu(&wq->list, &workqueues); |
a0a1a5fd | 5716 | |
c5178e6c | 5717 | if (wq_online && init_rescuer(wq) < 0) |
449b31ad LJ |
5718 | goto err_unlock_destroy; |
5719 | ||
5720 | apply_wqattrs_unlock(); | |
c5178e6c | 5721 | |
c3138f38 LJ |
5722 | if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) |
5723 | goto err_destroy; | |
1537663f | 5724 | |
3af24433 | 5725 | return wq; |
d2c1d404 | 5726 | |
1726a171 LJ |
5727 | err_unlock_free_node_nr_active: |
5728 | apply_wqattrs_unlock(); | |
4e9a3738 LJ |
5729 | /* |
5730 | * Failed alloc_and_link_pwqs() may leave pending pwq->release_work, | |
5731 | * flushing the pwq_release_worker ensures that the pwq_release_workfn() | |
5732 | * completes before calling kfree(wq). | |
5733 | */ | |
5734 | if (wq->flags & WQ_UNBOUND) { | |
5735 | kthread_flush_worker(pwq_release_worker); | |
91ccc6e7 | 5736 | free_node_nr_active(wq->node_nr_active); |
4e9a3738 | 5737 | } |
82efcab3 | 5738 | err_free_wq: |
6029a918 | 5739 | free_workqueue_attrs(wq->unbound_attrs); |
d2c1d404 TH |
5740 | kfree(wq); |
5741 | return NULL; | |
449b31ad LJ |
5742 | err_unlock_destroy: |
5743 | apply_wqattrs_unlock(); | |
d2c1d404 TH |
5744 | err_destroy: |
5745 | destroy_workqueue(wq); | |
4690c4ab | 5746 | return NULL; |
3af24433 | 5747 | } |
b188c57a MB |
5748 | |
5749 | __printf(1, 4) | |
5750 | struct workqueue_struct *alloc_workqueue(const char *fmt, | |
5751 | unsigned int flags, | |
5752 | int max_active, ...) | |
5753 | { | |
5754 | struct workqueue_struct *wq; | |
5755 | va_list args; | |
5756 | ||
5757 | va_start(args, max_active); | |
5758 | wq = __alloc_workqueue(fmt, flags, max_active, args); | |
5759 | va_end(args); | |
5760 | if (!wq) | |
5761 | return NULL; | |
5762 | ||
5763 | wq_init_lockdep(wq); | |
5764 | ||
5765 | return wq; | |
5766 | } | |
669de8bd | 5767 | EXPORT_SYMBOL_GPL(alloc_workqueue); |
1da177e4 | 5768 | |
ec0a7d44 MB |
5769 | #ifdef CONFIG_LOCKDEP |
5770 | __printf(1, 5) | |
5771 | struct workqueue_struct * | |
5772 | alloc_workqueue_lockdep_map(const char *fmt, unsigned int flags, | |
5773 | int max_active, struct lockdep_map *lockdep_map, ...) | |
5774 | { | |
5775 | struct workqueue_struct *wq; | |
5776 | va_list args; | |
5777 | ||
5778 | va_start(args, lockdep_map); | |
5779 | wq = __alloc_workqueue(fmt, flags, max_active, args); | |
5780 | va_end(args); | |
5781 | if (!wq) | |
5782 | return NULL; | |
5783 | ||
5784 | wq->lockdep_map = lockdep_map; | |
5785 | ||
5786 | return wq; | |
5787 | } | |
5788 | EXPORT_SYMBOL_GPL(alloc_workqueue_lockdep_map); | |
5789 | #endif | |
5790 | ||
c29eb853 TH |
5791 | static bool pwq_busy(struct pool_workqueue *pwq) |
5792 | { | |
5793 | int i; | |
5794 | ||
5795 | for (i = 0; i < WORK_NR_COLORS; i++) | |
5796 | if (pwq->nr_in_flight[i]) | |
5797 | return true; | |
5798 | ||
9f66cff2 | 5799 | if ((pwq != rcu_access_pointer(pwq->wq->dfl_pwq)) && (pwq->refcnt > 1)) |
c29eb853 | 5800 | return true; |
afa87ce8 | 5801 | if (!pwq_is_empty(pwq)) |
c29eb853 TH |
5802 | return true; |
5803 | ||
5804 | return false; | |
5805 | } | |
5806 | ||
3af24433 ON |
5807 | /** |
5808 | * destroy_workqueue - safely terminate a workqueue | |
5809 | * @wq: target workqueue | |
5810 | * | |
5811 | * Safely destroy a workqueue. All work currently pending will be done first. | |
5812 | */ | |
5813 | void destroy_workqueue(struct workqueue_struct *wq) | |
5814 | { | |
49e3cf44 | 5815 | struct pool_workqueue *pwq; |
636b927e | 5816 | int cpu; |
3af24433 | 5817 | |
def98c84 TH |
5818 | /* |
5819 | * Remove it from sysfs first so that sanity check failure doesn't | |
5820 | * lead to sysfs name conflicts. | |
5821 | */ | |
5822 | workqueue_sysfs_unregister(wq); | |
5823 | ||
33e3f0a3 RC |
5824 | /* mark the workqueue destruction is in progress */ |
5825 | mutex_lock(&wq->mutex); | |
5826 | wq->flags |= __WQ_DESTROYING; | |
5827 | mutex_unlock(&wq->mutex); | |
5828 | ||
9c5a2ba7 TH |
5829 | /* drain it before proceeding with destruction */ |
5830 | drain_workqueue(wq); | |
c8efcc25 | 5831 | |
def98c84 TH |
5832 | /* kill rescuer, if sanity checks fail, leave it w/o rescuer */ |
5833 | if (wq->rescuer) { | |
5834 | struct worker *rescuer = wq->rescuer; | |
5835 | ||
5836 | /* this prevents new queueing */ | |
a9b8a985 | 5837 | raw_spin_lock_irq(&wq_mayday_lock); |
def98c84 | 5838 | wq->rescuer = NULL; |
a9b8a985 | 5839 | raw_spin_unlock_irq(&wq_mayday_lock); |
def98c84 TH |
5840 | |
5841 | /* rescuer will empty maydays list before exiting */ | |
5842 | kthread_stop(rescuer->task); | |
8efe1223 | 5843 | kfree(rescuer); |
def98c84 TH |
5844 | } |
5845 | ||
c29eb853 TH |
5846 | /* |
5847 | * Sanity checks - grab all the locks so that we wait for all | |
5848 | * in-flight operations which may do put_pwq(). | |
5849 | */ | |
5850 | mutex_lock(&wq_pool_mutex); | |
b09f4fd3 | 5851 | mutex_lock(&wq->mutex); |
49e3cf44 | 5852 | for_each_pwq(pwq, wq) { |
a9b8a985 | 5853 | raw_spin_lock_irq(&pwq->pool->lock); |
c29eb853 | 5854 | if (WARN_ON(pwq_busy(pwq))) { |
1d9a6159 KW |
5855 | pr_warn("%s: %s has the following busy pwq\n", |
5856 | __func__, wq->name); | |
c29eb853 | 5857 | show_pwq(pwq); |
a9b8a985 | 5858 | raw_spin_unlock_irq(&pwq->pool->lock); |
b09f4fd3 | 5859 | mutex_unlock(&wq->mutex); |
c29eb853 | 5860 | mutex_unlock(&wq_pool_mutex); |
55df0933 | 5861 | show_one_workqueue(wq); |
6183c009 | 5862 | return; |
76af4d93 | 5863 | } |
a9b8a985 | 5864 | raw_spin_unlock_irq(&pwq->pool->lock); |
6183c009 | 5865 | } |
b09f4fd3 | 5866 | mutex_unlock(&wq->mutex); |
6183c009 | 5867 | |
a0a1a5fd TH |
5868 | /* |
5869 | * wq list is used to freeze wq, remove from list after | |
5870 | * flushing is complete in case freeze races us. | |
5871 | */ | |
e2dca7ad | 5872 | list_del_rcu(&wq->list); |
68e13a67 | 5873 | mutex_unlock(&wq_pool_mutex); |
3af24433 | 5874 | |
636b927e TH |
5875 | /* |
5876 | * We're the sole accessor of @wq. Directly access cpu_pwq and dfl_pwq | |
5877 | * to put the base refs. @wq will be auto-destroyed from the last | |
5878 | * pwq_put. RCU read lock prevents @wq from going away from under us. | |
5879 | */ | |
5880 | rcu_read_lock(); | |
4c16bd32 | 5881 | |
636b927e | 5882 | for_each_possible_cpu(cpu) { |
9f66cff2 TH |
5883 | put_pwq_unlocked(unbound_pwq(wq, cpu)); |
5884 | RCU_INIT_POINTER(*unbound_pwq_slot(wq, cpu), NULL); | |
29c91e99 | 5885 | } |
636b927e | 5886 | |
9f66cff2 TH |
5887 | put_pwq_unlocked(unbound_pwq(wq, -1)); |
5888 | RCU_INIT_POINTER(*unbound_pwq_slot(wq, -1), NULL); | |
636b927e TH |
5889 | |
5890 | rcu_read_unlock(); | |
3af24433 ON |
5891 | } |
5892 | EXPORT_SYMBOL_GPL(destroy_workqueue); | |
5893 | ||
dcd989cb TH |
5894 | /** |
5895 | * workqueue_set_max_active - adjust max_active of a workqueue | |
5896 | * @wq: target workqueue | |
5897 | * @max_active: new max_active value. | |
5898 | * | |
5797b1c1 TH |
5899 | * Set max_active of @wq to @max_active. See the alloc_workqueue() function |
5900 | * comment. | |
dcd989cb TH |
5901 | * |
5902 | * CONTEXT: | |
5903 | * Don't call from IRQ context. | |
5904 | */ | |
5905 | void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) | |
5906 | { | |
4cb1ef64 TH |
5907 | /* max_active doesn't mean anything for BH workqueues */ |
5908 | if (WARN_ON(wq->flags & WQ_BH)) | |
5909 | return; | |
8719dcea | 5910 | /* disallow meddling with max_active for ordered workqueues */ |
3bc1e711 | 5911 | if (WARN_ON(wq->flags & __WQ_ORDERED)) |
8719dcea TH |
5912 | return; |
5913 | ||
f3421797 | 5914 | max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); |
dcd989cb | 5915 | |
a357fc03 | 5916 | mutex_lock(&wq->mutex); |
dcd989cb TH |
5917 | |
5918 | wq->saved_max_active = max_active; | |
5797b1c1 TH |
5919 | if (wq->flags & WQ_UNBOUND) |
5920 | wq->saved_min_active = min(wq->saved_min_active, max_active); | |
5921 | ||
a045a272 | 5922 | wq_adjust_max_active(wq); |
93981800 | 5923 | |
a357fc03 | 5924 | mutex_unlock(&wq->mutex); |
15316ba8 | 5925 | } |
dcd989cb | 5926 | EXPORT_SYMBOL_GPL(workqueue_set_max_active); |
15316ba8 | 5927 | |
8f172181 TH |
5928 | /** |
5929 | * workqueue_set_min_active - adjust min_active of an unbound workqueue | |
5930 | * @wq: target unbound workqueue | |
5931 | * @min_active: new min_active value | |
5932 | * | |
5933 | * Set min_active of an unbound workqueue. Unlike other types of workqueues, an | |
5934 | * unbound workqueue is not guaranteed to be able to process max_active | |
5935 | * interdependent work items. Instead, an unbound workqueue is guaranteed to be | |
5936 | * able to process min_active number of interdependent work items which is | |
5937 | * %WQ_DFL_MIN_ACTIVE by default. | |
5938 | * | |
5939 | * Use this function to adjust the min_active value between 0 and the current | |
5940 | * max_active. | |
5941 | */ | |
5942 | void workqueue_set_min_active(struct workqueue_struct *wq, int min_active) | |
5943 | { | |
5944 | /* min_active is only meaningful for non-ordered unbound workqueues */ | |
5945 | if (WARN_ON((wq->flags & (WQ_BH | WQ_UNBOUND | __WQ_ORDERED)) != | |
5946 | WQ_UNBOUND)) | |
5947 | return; | |
5948 | ||
5949 | mutex_lock(&wq->mutex); | |
5950 | wq->saved_min_active = clamp(min_active, 0, wq->saved_max_active); | |
5951 | wq_adjust_max_active(wq); | |
5952 | mutex_unlock(&wq->mutex); | |
5953 | } | |
5954 | ||
27d4ee03 LW |
5955 | /** |
5956 | * current_work - retrieve %current task's work struct | |
5957 | * | |
5958 | * Determine if %current task is a workqueue worker and what it's working on. | |
5959 | * Useful to find out the context that the %current task is running in. | |
5960 | * | |
5961 | * Return: work struct if %current task is a workqueue worker, %NULL otherwise. | |
5962 | */ | |
5963 | struct work_struct *current_work(void) | |
5964 | { | |
5965 | struct worker *worker = current_wq_worker(); | |
5966 | ||
5967 | return worker ? worker->current_work : NULL; | |
5968 | } | |
5969 | EXPORT_SYMBOL(current_work); | |
5970 | ||
e6267616 TH |
5971 | /** |
5972 | * current_is_workqueue_rescuer - is %current workqueue rescuer? | |
5973 | * | |
5974 | * Determine whether %current is a workqueue rescuer. Can be used from | |
5975 | * work functions to determine whether it's being run off the rescuer task. | |
d185af30 YB |
5976 | * |
5977 | * Return: %true if %current is a workqueue rescuer. %false otherwise. | |
e6267616 TH |
5978 | */ |
5979 | bool current_is_workqueue_rescuer(void) | |
5980 | { | |
5981 | struct worker *worker = current_wq_worker(); | |
5982 | ||
6a092dfd | 5983 | return worker && worker->rescue_wq; |
e6267616 TH |
5984 | } |
5985 | ||
eef6a7d5 | 5986 | /** |
dcd989cb TH |
5987 | * workqueue_congested - test whether a workqueue is congested |
5988 | * @cpu: CPU in question | |
5989 | * @wq: target workqueue | |
eef6a7d5 | 5990 | * |
dcd989cb TH |
5991 | * Test whether @wq's cpu workqueue for @cpu is congested. There is |
5992 | * no synchronization around this function and the test result is | |
5993 | * unreliable and only useful as advisory hints or for debugging. | |
eef6a7d5 | 5994 | * |
d3251859 | 5995 | * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. |
636b927e TH |
5996 | * |
5997 | * With the exception of ordered workqueues, all workqueues have per-cpu | |
5998 | * pool_workqueues, each with its own congested state. A workqueue being | |
5999 | * congested on one CPU doesn't mean that the workqueue is contested on any | |
6000 | * other CPUs. | |
d3251859 | 6001 | * |
d185af30 | 6002 | * Return: |
dcd989cb | 6003 | * %true if congested, %false otherwise. |
eef6a7d5 | 6004 | */ |
d84ff051 | 6005 | bool workqueue_congested(int cpu, struct workqueue_struct *wq) |
1da177e4 | 6006 | { |
7fb98ea7 | 6007 | struct pool_workqueue *pwq; |
76af4d93 TH |
6008 | bool ret; |
6009 | ||
24acfb71 TG |
6010 | rcu_read_lock(); |
6011 | preempt_disable(); | |
7fb98ea7 | 6012 | |
d3251859 TH |
6013 | if (cpu == WORK_CPU_UNBOUND) |
6014 | cpu = smp_processor_id(); | |
6015 | ||
636b927e | 6016 | pwq = *per_cpu_ptr(wq->cpu_pwq, cpu); |
f97a4a1a | 6017 | ret = !list_empty(&pwq->inactive_works); |
636b927e | 6018 | |
24acfb71 TG |
6019 | preempt_enable(); |
6020 | rcu_read_unlock(); | |
76af4d93 TH |
6021 | |
6022 | return ret; | |
1da177e4 | 6023 | } |
dcd989cb | 6024 | EXPORT_SYMBOL_GPL(workqueue_congested); |
1da177e4 | 6025 | |
dcd989cb TH |
6026 | /** |
6027 | * work_busy - test whether a work is currently pending or running | |
6028 | * @work: the work to be tested | |
6029 | * | |
6030 | * Test whether @work is currently pending or running. There is no | |
6031 | * synchronization around this function and the test result is | |
6032 | * unreliable and only useful as advisory hints or for debugging. | |
dcd989cb | 6033 | * |
d185af30 | 6034 | * Return: |
dcd989cb TH |
6035 | * OR'd bitmask of WORK_BUSY_* bits. |
6036 | */ | |
6037 | unsigned int work_busy(struct work_struct *work) | |
1da177e4 | 6038 | { |
fa1b54e6 | 6039 | struct worker_pool *pool; |
c26e2f2e | 6040 | unsigned long irq_flags; |
dcd989cb | 6041 | unsigned int ret = 0; |
1da177e4 | 6042 | |
dcd989cb TH |
6043 | if (work_pending(work)) |
6044 | ret |= WORK_BUSY_PENDING; | |
1da177e4 | 6045 | |
24acfb71 | 6046 | rcu_read_lock(); |
fa1b54e6 | 6047 | pool = get_work_pool(work); |
038366c5 | 6048 | if (pool) { |
c26e2f2e | 6049 | raw_spin_lock_irqsave(&pool->lock, irq_flags); |
038366c5 LJ |
6050 | if (find_worker_executing_work(pool, work)) |
6051 | ret |= WORK_BUSY_RUNNING; | |
c26e2f2e | 6052 | raw_spin_unlock_irqrestore(&pool->lock, irq_flags); |
038366c5 | 6053 | } |
24acfb71 | 6054 | rcu_read_unlock(); |
1da177e4 | 6055 | |
dcd989cb | 6056 | return ret; |
1da177e4 | 6057 | } |
dcd989cb | 6058 | EXPORT_SYMBOL_GPL(work_busy); |
1da177e4 | 6059 | |
3d1cb205 TH |
6060 | /** |
6061 | * set_worker_desc - set description for the current work item | |
6062 | * @fmt: printf-style format string | |
6063 | * @...: arguments for the format string | |
6064 | * | |
6065 | * This function can be called by a running work function to describe what | |
6066 | * the work item is about. If the worker task gets dumped, this | |
6067 | * information will be printed out together to help debugging. The | |
6068 | * description can be at most WORKER_DESC_LEN including the trailing '\0'. | |
6069 | */ | |
6070 | void set_worker_desc(const char *fmt, ...) | |
6071 | { | |
6072 | struct worker *worker = current_wq_worker(); | |
6073 | va_list args; | |
6074 | ||
6075 | if (worker) { | |
6076 | va_start(args, fmt); | |
6077 | vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); | |
6078 | va_end(args); | |
3d1cb205 TH |
6079 | } |
6080 | } | |
5c750d58 | 6081 | EXPORT_SYMBOL_GPL(set_worker_desc); |
3d1cb205 TH |
6082 | |
6083 | /** | |
6084 | * print_worker_info - print out worker information and description | |
6085 | * @log_lvl: the log level to use when printing | |
6086 | * @task: target task | |
6087 | * | |
6088 | * If @task is a worker and currently executing a work item, print out the | |
6089 | * name of the workqueue being serviced and worker description set with | |
6090 | * set_worker_desc() by the currently executing work item. | |
6091 | * | |
6092 | * This function can be safely called on any task as long as the | |
6093 | * task_struct itself is accessible. While safe, this function isn't | |
6094 | * synchronized and may print out mixups or garbages of limited length. | |
6095 | */ | |
6096 | void print_worker_info(const char *log_lvl, struct task_struct *task) | |
6097 | { | |
6098 | work_func_t *fn = NULL; | |
6099 | char name[WQ_NAME_LEN] = { }; | |
6100 | char desc[WORKER_DESC_LEN] = { }; | |
6101 | struct pool_workqueue *pwq = NULL; | |
6102 | struct workqueue_struct *wq = NULL; | |
3d1cb205 TH |
6103 | struct worker *worker; |
6104 | ||
6105 | if (!(task->flags & PF_WQ_WORKER)) | |
6106 | return; | |
6107 | ||
6108 | /* | |
6109 | * This function is called without any synchronization and @task | |
6110 | * could be in any state. Be careful with dereferences. | |
6111 | */ | |
e700591a | 6112 | worker = kthread_probe_data(task); |
3d1cb205 TH |
6113 | |
6114 | /* | |
8bf89593 TH |
6115 | * Carefully copy the associated workqueue's workfn, name and desc. |
6116 | * Keep the original last '\0' in case the original is garbage. | |
3d1cb205 | 6117 | */ |
fe557319 CH |
6118 | copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn)); |
6119 | copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq)); | |
6120 | copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq)); | |
6121 | copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1); | |
6122 | copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1); | |
3d1cb205 TH |
6123 | |
6124 | if (fn || name[0] || desc[0]) { | |
d75f773c | 6125 | printk("%sWorkqueue: %s %ps", log_lvl, name, fn); |
8bf89593 | 6126 | if (strcmp(name, desc)) |
3d1cb205 TH |
6127 | pr_cont(" (%s)", desc); |
6128 | pr_cont("\n"); | |
6129 | } | |
6130 | } | |
6131 | ||
3494fc30 TH |
6132 | static void pr_cont_pool_info(struct worker_pool *pool) |
6133 | { | |
6134 | pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask); | |
6135 | if (pool->node != NUMA_NO_NODE) | |
6136 | pr_cont(" node=%d", pool->node); | |
4cb1ef64 TH |
6137 | pr_cont(" flags=0x%x", pool->flags); |
6138 | if (pool->flags & POOL_BH) | |
6139 | pr_cont(" bh%s", | |
6140 | pool->attrs->nice == HIGHPRI_NICE_LEVEL ? "-hi" : ""); | |
6141 | else | |
6142 | pr_cont(" nice=%d", pool->attrs->nice); | |
6143 | } | |
6144 | ||
6145 | static void pr_cont_worker_id(struct worker *worker) | |
6146 | { | |
6147 | struct worker_pool *pool = worker->pool; | |
6148 | ||
6149 | if (pool->flags & WQ_BH) | |
6150 | pr_cont("bh%s", | |
6151 | pool->attrs->nice == HIGHPRI_NICE_LEVEL ? "-hi" : ""); | |
6152 | else | |
6153 | pr_cont("%d%s", task_pid_nr(worker->task), | |
6154 | worker->rescue_wq ? "(RESCUER)" : ""); | |
3494fc30 TH |
6155 | } |
6156 | ||
c76feb0d PM |
6157 | struct pr_cont_work_struct { |
6158 | bool comma; | |
6159 | work_func_t func; | |
6160 | long ctr; | |
6161 | }; | |
6162 | ||
6163 | static void pr_cont_work_flush(bool comma, work_func_t func, struct pr_cont_work_struct *pcwsp) | |
6164 | { | |
6165 | if (!pcwsp->ctr) | |
6166 | goto out_record; | |
6167 | if (func == pcwsp->func) { | |
6168 | pcwsp->ctr++; | |
6169 | return; | |
6170 | } | |
6171 | if (pcwsp->ctr == 1) | |
6172 | pr_cont("%s %ps", pcwsp->comma ? "," : "", pcwsp->func); | |
6173 | else | |
6174 | pr_cont("%s %ld*%ps", pcwsp->comma ? "," : "", pcwsp->ctr, pcwsp->func); | |
6175 | pcwsp->ctr = 0; | |
6176 | out_record: | |
6177 | if ((long)func == -1L) | |
6178 | return; | |
6179 | pcwsp->comma = comma; | |
6180 | pcwsp->func = func; | |
6181 | pcwsp->ctr = 1; | |
6182 | } | |
6183 | ||
6184 | static void pr_cont_work(bool comma, struct work_struct *work, struct pr_cont_work_struct *pcwsp) | |
3494fc30 TH |
6185 | { |
6186 | if (work->func == wq_barrier_func) { | |
6187 | struct wq_barrier *barr; | |
6188 | ||
6189 | barr = container_of(work, struct wq_barrier, work); | |
6190 | ||
c76feb0d | 6191 | pr_cont_work_flush(comma, (work_func_t)-1, pcwsp); |
3494fc30 TH |
6192 | pr_cont("%s BAR(%d)", comma ? "," : "", |
6193 | task_pid_nr(barr->task)); | |
6194 | } else { | |
c76feb0d PM |
6195 | if (!comma) |
6196 | pr_cont_work_flush(comma, (work_func_t)-1, pcwsp); | |
6197 | pr_cont_work_flush(comma, work->func, pcwsp); | |
3494fc30 TH |
6198 | } |
6199 | } | |
6200 | ||
6201 | static void show_pwq(struct pool_workqueue *pwq) | |
6202 | { | |
c76feb0d | 6203 | struct pr_cont_work_struct pcws = { .ctr = 0, }; |
3494fc30 TH |
6204 | struct worker_pool *pool = pwq->pool; |
6205 | struct work_struct *work; | |
6206 | struct worker *worker; | |
6207 | bool has_in_flight = false, has_pending = false; | |
6208 | int bkt; | |
6209 | ||
6210 | pr_info(" pwq %d:", pool->id); | |
6211 | pr_cont_pool_info(pool); | |
6212 | ||
a045a272 TH |
6213 | pr_cont(" active=%d refcnt=%d%s\n", |
6214 | pwq->nr_active, pwq->refcnt, | |
3494fc30 TH |
6215 | !list_empty(&pwq->mayday_node) ? " MAYDAY" : ""); |
6216 | ||
6217 | hash_for_each(pool->busy_hash, bkt, worker, hentry) { | |
6218 | if (worker->current_pwq == pwq) { | |
6219 | has_in_flight = true; | |
6220 | break; | |
6221 | } | |
6222 | } | |
6223 | if (has_in_flight) { | |
6224 | bool comma = false; | |
6225 | ||
6226 | pr_info(" in-flight:"); | |
6227 | hash_for_each(pool->busy_hash, bkt, worker, hentry) { | |
6228 | if (worker->current_pwq != pwq) | |
6229 | continue; | |
6230 | ||
4cb1ef64 TH |
6231 | pr_cont(" %s", comma ? "," : ""); |
6232 | pr_cont_worker_id(worker); | |
6233 | pr_cont(":%ps", worker->current_func); | |
3494fc30 | 6234 | list_for_each_entry(work, &worker->scheduled, entry) |
c76feb0d PM |
6235 | pr_cont_work(false, work, &pcws); |
6236 | pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); | |
3494fc30 TH |
6237 | comma = true; |
6238 | } | |
6239 | pr_cont("\n"); | |
6240 | } | |
6241 | ||
6242 | list_for_each_entry(work, &pool->worklist, entry) { | |
6243 | if (get_work_pwq(work) == pwq) { | |
6244 | has_pending = true; | |
6245 | break; | |
6246 | } | |
6247 | } | |
6248 | if (has_pending) { | |
6249 | bool comma = false; | |
6250 | ||
6251 | pr_info(" pending:"); | |
6252 | list_for_each_entry(work, &pool->worklist, entry) { | |
6253 | if (get_work_pwq(work) != pwq) | |
6254 | continue; | |
6255 | ||
c76feb0d | 6256 | pr_cont_work(comma, work, &pcws); |
3494fc30 TH |
6257 | comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); |
6258 | } | |
c76feb0d | 6259 | pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); |
3494fc30 TH |
6260 | pr_cont("\n"); |
6261 | } | |
6262 | ||
f97a4a1a | 6263 | if (!list_empty(&pwq->inactive_works)) { |
3494fc30 TH |
6264 | bool comma = false; |
6265 | ||
f97a4a1a LJ |
6266 | pr_info(" inactive:"); |
6267 | list_for_each_entry(work, &pwq->inactive_works, entry) { | |
c76feb0d | 6268 | pr_cont_work(comma, work, &pcws); |
3494fc30 TH |
6269 | comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); |
6270 | } | |
c76feb0d | 6271 | pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); |
3494fc30 TH |
6272 | pr_cont("\n"); |
6273 | } | |
6274 | } | |
6275 | ||
6276 | /** | |
55df0933 IK |
6277 | * show_one_workqueue - dump state of specified workqueue |
6278 | * @wq: workqueue whose state will be printed | |
3494fc30 | 6279 | */ |
55df0933 | 6280 | void show_one_workqueue(struct workqueue_struct *wq) |
3494fc30 | 6281 | { |
55df0933 IK |
6282 | struct pool_workqueue *pwq; |
6283 | bool idle = true; | |
c26e2f2e | 6284 | unsigned long irq_flags; |
3494fc30 | 6285 | |
55df0933 | 6286 | for_each_pwq(pwq, wq) { |
afa87ce8 | 6287 | if (!pwq_is_empty(pwq)) { |
55df0933 IK |
6288 | idle = false; |
6289 | break; | |
3494fc30 | 6290 | } |
55df0933 IK |
6291 | } |
6292 | if (idle) /* Nothing to print for idle workqueue */ | |
6293 | return; | |
3494fc30 | 6294 | |
55df0933 | 6295 | pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags); |
3494fc30 | 6296 | |
55df0933 | 6297 | for_each_pwq(pwq, wq) { |
c26e2f2e | 6298 | raw_spin_lock_irqsave(&pwq->pool->lock, irq_flags); |
afa87ce8 | 6299 | if (!pwq_is_empty(pwq)) { |
62635ea8 | 6300 | /* |
55df0933 IK |
6301 | * Defer printing to avoid deadlocks in console |
6302 | * drivers that queue work while holding locks | |
6303 | * also taken in their write paths. | |
62635ea8 | 6304 | */ |
55df0933 IK |
6305 | printk_deferred_enter(); |
6306 | show_pwq(pwq); | |
6307 | printk_deferred_exit(); | |
3494fc30 | 6308 | } |
c26e2f2e | 6309 | raw_spin_unlock_irqrestore(&pwq->pool->lock, irq_flags); |
62635ea8 SS |
6310 | /* |
6311 | * We could be printing a lot from atomic context, e.g. | |
55df0933 | 6312 | * sysrq-t -> show_all_workqueues(). Avoid triggering |
62635ea8 SS |
6313 | * hard lockup. |
6314 | */ | |
6315 | touch_nmi_watchdog(); | |
3494fc30 TH |
6316 | } |
6317 | ||
55df0933 IK |
6318 | } |
6319 | ||
6320 | /** | |
6321 | * show_one_worker_pool - dump state of specified worker pool | |
6322 | * @pool: worker pool whose state will be printed | |
6323 | */ | |
6324 | static void show_one_worker_pool(struct worker_pool *pool) | |
6325 | { | |
6326 | struct worker *worker; | |
6327 | bool first = true; | |
c26e2f2e | 6328 | unsigned long irq_flags; |
335a42eb | 6329 | unsigned long hung = 0; |
55df0933 | 6330 | |
c26e2f2e | 6331 | raw_spin_lock_irqsave(&pool->lock, irq_flags); |
55df0933 IK |
6332 | if (pool->nr_workers == pool->nr_idle) |
6333 | goto next_pool; | |
335a42eb PM |
6334 | |
6335 | /* How long the first pending work is waiting for a worker. */ | |
6336 | if (!list_empty(&pool->worklist)) | |
6337 | hung = jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000; | |
6338 | ||
55df0933 IK |
6339 | /* |
6340 | * Defer printing to avoid deadlocks in console drivers that | |
6341 | * queue work while holding locks also taken in their write | |
6342 | * paths. | |
6343 | */ | |
6344 | printk_deferred_enter(); | |
6345 | pr_info("pool %d:", pool->id); | |
6346 | pr_cont_pool_info(pool); | |
335a42eb | 6347 | pr_cont(" hung=%lus workers=%d", hung, pool->nr_workers); |
55df0933 IK |
6348 | if (pool->manager) |
6349 | pr_cont(" manager: %d", | |
6350 | task_pid_nr(pool->manager->task)); | |
6351 | list_for_each_entry(worker, &pool->idle_list, entry) { | |
4cb1ef64 TH |
6352 | pr_cont(" %s", first ? "idle: " : ""); |
6353 | pr_cont_worker_id(worker); | |
55df0933 IK |
6354 | first = false; |
6355 | } | |
6356 | pr_cont("\n"); | |
6357 | printk_deferred_exit(); | |
6358 | next_pool: | |
c26e2f2e | 6359 | raw_spin_unlock_irqrestore(&pool->lock, irq_flags); |
55df0933 IK |
6360 | /* |
6361 | * We could be printing a lot from atomic context, e.g. | |
6362 | * sysrq-t -> show_all_workqueues(). Avoid triggering | |
6363 | * hard lockup. | |
6364 | */ | |
6365 | touch_nmi_watchdog(); | |
6366 | ||
6367 | } | |
6368 | ||
6369 | /** | |
6370 | * show_all_workqueues - dump workqueue state | |
6371 | * | |
704bc669 | 6372 | * Called from a sysrq handler and prints out all busy workqueues and pools. |
55df0933 IK |
6373 | */ |
6374 | void show_all_workqueues(void) | |
6375 | { | |
6376 | struct workqueue_struct *wq; | |
6377 | struct worker_pool *pool; | |
6378 | int pi; | |
6379 | ||
6380 | rcu_read_lock(); | |
6381 | ||
6382 | pr_info("Showing busy workqueues and worker pools:\n"); | |
6383 | ||
6384 | list_for_each_entry_rcu(wq, &workqueues, list) | |
6385 | show_one_workqueue(wq); | |
6386 | ||
6387 | for_each_pool(pool, pi) | |
6388 | show_one_worker_pool(pool); | |
6389 | ||
24acfb71 | 6390 | rcu_read_unlock(); |
3494fc30 TH |
6391 | } |
6392 | ||
704bc669 JL |
6393 | /** |
6394 | * show_freezable_workqueues - dump freezable workqueue state | |
6395 | * | |
6396 | * Called from try_to_freeze_tasks() and prints out all freezable workqueues | |
6397 | * still busy. | |
6398 | */ | |
6399 | void show_freezable_workqueues(void) | |
6400 | { | |
6401 | struct workqueue_struct *wq; | |
6402 | ||
6403 | rcu_read_lock(); | |
6404 | ||
6405 | pr_info("Showing freezable workqueues that are still busy:\n"); | |
6406 | ||
6407 | list_for_each_entry_rcu(wq, &workqueues, list) { | |
6408 | if (!(wq->flags & WQ_FREEZABLE)) | |
6409 | continue; | |
6410 | show_one_workqueue(wq); | |
6411 | } | |
6412 | ||
6413 | rcu_read_unlock(); | |
6414 | } | |
6415 | ||
6b59808b TH |
6416 | /* used to show worker information through /proc/PID/{comm,stat,status} */ |
6417 | void wq_worker_comm(char *buf, size_t size, struct task_struct *task) | |
6418 | { | |
197f6acc | 6419 | /* stabilize PF_WQ_WORKER and worker pool association */ |
6b59808b TH |
6420 | mutex_lock(&wq_pool_attach_mutex); |
6421 | ||
197f6acc TH |
6422 | if (task->flags & PF_WQ_WORKER) { |
6423 | struct worker *worker = kthread_data(task); | |
6424 | struct worker_pool *pool = worker->pool; | |
2a1b02bc TH |
6425 | int off; |
6426 | ||
6427 | off = format_worker_id(buf, size, worker, pool); | |
6b59808b | 6428 | |
197f6acc | 6429 | if (pool) { |
a9b8a985 | 6430 | raw_spin_lock_irq(&pool->lock); |
197f6acc TH |
6431 | /* |
6432 | * ->desc tracks information (wq name or | |
6433 | * set_worker_desc()) for the latest execution. If | |
6434 | * current, prepend '+', otherwise '-'. | |
6435 | */ | |
6436 | if (worker->desc[0] != '\0') { | |
6437 | if (worker->current_work) | |
6438 | scnprintf(buf + off, size - off, "+%s", | |
6439 | worker->desc); | |
6440 | else | |
6441 | scnprintf(buf + off, size - off, "-%s", | |
6442 | worker->desc); | |
6443 | } | |
a9b8a985 | 6444 | raw_spin_unlock_irq(&pool->lock); |
6b59808b | 6445 | } |
2a1b02bc TH |
6446 | } else { |
6447 | strscpy(buf, task->comm, size); | |
6b59808b TH |
6448 | } |
6449 | ||
6450 | mutex_unlock(&wq_pool_attach_mutex); | |
6451 | } | |
6452 | ||
66448bc2 MM |
6453 | #ifdef CONFIG_SMP |
6454 | ||
db7bccf4 TH |
6455 | /* |
6456 | * CPU hotplug. | |
6457 | * | |
e22bee78 | 6458 | * There are two challenges in supporting CPU hotplug. Firstly, there |
112202d9 | 6459 | * are a lot of assumptions on strong associations among work, pwq and |
706026c2 | 6460 | * pool which make migrating pending and scheduled works very |
e22bee78 | 6461 | * difficult to implement without impacting hot paths. Secondly, |
94cf58bb | 6462 | * worker pools serve mix of short, long and very long running works making |
e22bee78 TH |
6463 | * blocked draining impractical. |
6464 | * | |
24647570 | 6465 | * This is solved by allowing the pools to be disassociated from the CPU |
628c78e7 TH |
6466 | * running as an unbound one and allowing it to be reattached later if the |
6467 | * cpu comes back online. | |
db7bccf4 | 6468 | */ |
1da177e4 | 6469 | |
e8b3f8db | 6470 | static void unbind_workers(int cpu) |
3af24433 | 6471 | { |
4ce62e9e | 6472 | struct worker_pool *pool; |
db7bccf4 | 6473 | struct worker *worker; |
3af24433 | 6474 | |
f02ae73a | 6475 | for_each_cpu_worker_pool(pool, cpu) { |
1258fae7 | 6476 | mutex_lock(&wq_pool_attach_mutex); |
a9b8a985 | 6477 | raw_spin_lock_irq(&pool->lock); |
3af24433 | 6478 | |
94cf58bb | 6479 | /* |
92f9c5c4 | 6480 | * We've blocked all attach/detach operations. Make all workers |
94cf58bb | 6481 | * unbound and set DISASSOCIATED. Before this, all workers |
11b45b0b | 6482 | * must be on the cpu. After this, they may become diasporas. |
b4ac9384 LJ |
6483 | * And the preemption disabled section in their sched callbacks |
6484 | * are guaranteed to see WORKER_UNBOUND since the code here | |
6485 | * is on the same cpu. | |
94cf58bb | 6486 | */ |
da028469 | 6487 | for_each_pool_worker(worker, pool) |
c9e7cf27 | 6488 | worker->flags |= WORKER_UNBOUND; |
06ba38a9 | 6489 | |
24647570 | 6490 | pool->flags |= POOL_DISASSOCIATED; |
f2d5a0ee | 6491 | |
eb283428 | 6492 | /* |
989442d7 LJ |
6493 | * The handling of nr_running in sched callbacks are disabled |
6494 | * now. Zap nr_running. After this, nr_running stays zero and | |
6495 | * need_more_worker() and keep_working() are always true as | |
6496 | * long as the worklist is not empty. This pool now behaves as | |
6497 | * an unbound (in terms of concurrency management) pool which | |
eb283428 LJ |
6498 | * are served by workers tied to the pool. |
6499 | */ | |
bc35f7ef | 6500 | pool->nr_running = 0; |
eb283428 LJ |
6501 | |
6502 | /* | |
6503 | * With concurrency management just turned off, a busy | |
6504 | * worker blocking could lead to lengthy stalls. Kick off | |
6505 | * unbound chain execution of currently pending work items. | |
6506 | */ | |
0219a352 | 6507 | kick_pool(pool); |
989442d7 | 6508 | |
a9b8a985 | 6509 | raw_spin_unlock_irq(&pool->lock); |
989442d7 | 6510 | |
793777bc VS |
6511 | for_each_pool_worker(worker, pool) |
6512 | unbind_worker(worker); | |
989442d7 LJ |
6513 | |
6514 | mutex_unlock(&wq_pool_attach_mutex); | |
eb283428 | 6515 | } |
3af24433 | 6516 | } |
3af24433 | 6517 | |
bd7c089e TH |
6518 | /** |
6519 | * rebind_workers - rebind all workers of a pool to the associated CPU | |
6520 | * @pool: pool of interest | |
6521 | * | |
a9ab775b | 6522 | * @pool->cpu is coming online. Rebind all workers to the CPU. |
bd7c089e TH |
6523 | */ |
6524 | static void rebind_workers(struct worker_pool *pool) | |
6525 | { | |
a9ab775b | 6526 | struct worker *worker; |
bd7c089e | 6527 | |
1258fae7 | 6528 | lockdep_assert_held(&wq_pool_attach_mutex); |
bd7c089e | 6529 | |
a9ab775b TH |
6530 | /* |
6531 | * Restore CPU affinity of all workers. As all idle workers should | |
6532 | * be on the run-queue of the associated CPU before any local | |
402dd89d | 6533 | * wake-ups for concurrency management happen, restore CPU affinity |
a9ab775b TH |
6534 | * of all workers first and then clear UNBOUND. As we're called |
6535 | * from CPU_ONLINE, the following shouldn't fail. | |
6536 | */ | |
c63a2e52 VS |
6537 | for_each_pool_worker(worker, pool) { |
6538 | kthread_set_per_cpu(worker->task, pool->cpu); | |
6539 | WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, | |
9546b29e | 6540 | pool_allowed_cpus(pool)) < 0); |
c63a2e52 | 6541 | } |
bd7c089e | 6542 | |
a9b8a985 | 6543 | raw_spin_lock_irq(&pool->lock); |
f7c17d26 | 6544 | |
3de5e884 | 6545 | pool->flags &= ~POOL_DISASSOCIATED; |
bd7c089e | 6546 | |
da028469 | 6547 | for_each_pool_worker(worker, pool) { |
a9ab775b | 6548 | unsigned int worker_flags = worker->flags; |
bd7c089e | 6549 | |
a9ab775b TH |
6550 | /* |
6551 | * We want to clear UNBOUND but can't directly call | |
6552 | * worker_clr_flags() or adjust nr_running. Atomically | |
6553 | * replace UNBOUND with another NOT_RUNNING flag REBOUND. | |
6554 | * @worker will clear REBOUND using worker_clr_flags() when | |
6555 | * it initiates the next execution cycle thus restoring | |
6556 | * concurrency management. Note that when or whether | |
6557 | * @worker clears REBOUND doesn't affect correctness. | |
6558 | * | |
c95491ed | 6559 | * WRITE_ONCE() is necessary because @worker->flags may be |
a9ab775b | 6560 | * tested without holding any lock in |
6d25be57 | 6561 | * wq_worker_running(). Without it, NOT_RUNNING test may |
a9ab775b TH |
6562 | * fail incorrectly leading to premature concurrency |
6563 | * management operations. | |
6564 | */ | |
6565 | WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); | |
6566 | worker_flags |= WORKER_REBOUND; | |
6567 | worker_flags &= ~WORKER_UNBOUND; | |
c95491ed | 6568 | WRITE_ONCE(worker->flags, worker_flags); |
bd7c089e | 6569 | } |
a9ab775b | 6570 | |
a9b8a985 | 6571 | raw_spin_unlock_irq(&pool->lock); |
bd7c089e TH |
6572 | } |
6573 | ||
7dbc725e TH |
6574 | /** |
6575 | * restore_unbound_workers_cpumask - restore cpumask of unbound workers | |
6576 | * @pool: unbound pool of interest | |
6577 | * @cpu: the CPU which is coming up | |
6578 | * | |
6579 | * An unbound pool may end up with a cpumask which doesn't have any online | |
6580 | * CPUs. When a worker of such pool get scheduled, the scheduler resets | |
6581 | * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any | |
6582 | * online CPU before, cpus_allowed of all its workers should be restored. | |
6583 | */ | |
6584 | static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) | |
6585 | { | |
6586 | static cpumask_t cpumask; | |
6587 | struct worker *worker; | |
7dbc725e | 6588 | |
1258fae7 | 6589 | lockdep_assert_held(&wq_pool_attach_mutex); |
7dbc725e TH |
6590 | |
6591 | /* is @cpu allowed for @pool? */ | |
6592 | if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) | |
6593 | return; | |
6594 | ||
7dbc725e | 6595 | cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); |
7dbc725e TH |
6596 | |
6597 | /* as we're called from CPU_ONLINE, the following shouldn't fail */ | |
da028469 | 6598 | for_each_pool_worker(worker, pool) |
d945b5e9 | 6599 | WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0); |
7dbc725e TH |
6600 | } |
6601 | ||
7ee681b2 TG |
6602 | int workqueue_prepare_cpu(unsigned int cpu) |
6603 | { | |
6604 | struct worker_pool *pool; | |
6605 | ||
6606 | for_each_cpu_worker_pool(pool, cpu) { | |
6607 | if (pool->nr_workers) | |
6608 | continue; | |
6609 | if (!create_worker(pool)) | |
6610 | return -ENOMEM; | |
6611 | } | |
6612 | return 0; | |
6613 | } | |
6614 | ||
6615 | int workqueue_online_cpu(unsigned int cpu) | |
3af24433 | 6616 | { |
4ce62e9e | 6617 | struct worker_pool *pool; |
4c16bd32 | 6618 | struct workqueue_struct *wq; |
7dbc725e | 6619 | int pi; |
3ce63377 | 6620 | |
7ee681b2 | 6621 | mutex_lock(&wq_pool_mutex); |
7dbc725e | 6622 | |
8d84baf7 LJ |
6623 | cpumask_set_cpu(cpu, wq_online_cpumask); |
6624 | ||
7ee681b2 | 6625 | for_each_pool(pool, pi) { |
4cb1ef64 TH |
6626 | /* BH pools aren't affected by hotplug */ |
6627 | if (pool->flags & POOL_BH) | |
6628 | continue; | |
94cf58bb | 6629 | |
4cb1ef64 | 6630 | mutex_lock(&wq_pool_attach_mutex); |
7ee681b2 TG |
6631 | if (pool->cpu == cpu) |
6632 | rebind_workers(pool); | |
6633 | else if (pool->cpu < 0) | |
6634 | restore_unbound_workers_cpumask(pool, cpu); | |
1258fae7 | 6635 | mutex_unlock(&wq_pool_attach_mutex); |
7ee681b2 | 6636 | } |
6ba94429 | 6637 | |
fef59c9c | 6638 | /* update pod affinity of unbound workqueues */ |
4cbfd3de | 6639 | list_for_each_entry(wq, &workqueues, list) { |
84193c07 TH |
6640 | struct workqueue_attrs *attrs = wq->unbound_attrs; |
6641 | ||
6642 | if (attrs) { | |
6643 | const struct wq_pod_type *pt = wqattrs_pod_type(attrs); | |
6644 | int tcpu; | |
4cbfd3de | 6645 | |
84193c07 | 6646 | for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]]) |
b2b1f933 | 6647 | unbound_wq_update_pwq(wq, tcpu); |
5797b1c1 TH |
6648 | |
6649 | mutex_lock(&wq->mutex); | |
6650 | wq_update_node_max_active(wq, -1); | |
6651 | mutex_unlock(&wq->mutex); | |
4cbfd3de TH |
6652 | } |
6653 | } | |
6ba94429 | 6654 | |
7ee681b2 TG |
6655 | mutex_unlock(&wq_pool_mutex); |
6656 | return 0; | |
6ba94429 FW |
6657 | } |
6658 | ||
7ee681b2 | 6659 | int workqueue_offline_cpu(unsigned int cpu) |
6ba94429 | 6660 | { |
6ba94429 FW |
6661 | struct workqueue_struct *wq; |
6662 | ||
7ee681b2 | 6663 | /* unbinding per-cpu workers should happen on the local CPU */ |
e8b3f8db LJ |
6664 | if (WARN_ON(cpu != smp_processor_id())) |
6665 | return -1; | |
6666 | ||
6667 | unbind_workers(cpu); | |
7ee681b2 | 6668 | |
fef59c9c | 6669 | /* update pod affinity of unbound workqueues */ |
7ee681b2 | 6670 | mutex_lock(&wq_pool_mutex); |
8d84baf7 LJ |
6671 | |
6672 | cpumask_clear_cpu(cpu, wq_online_cpumask); | |
6673 | ||
4cbfd3de | 6674 | list_for_each_entry(wq, &workqueues, list) { |
84193c07 TH |
6675 | struct workqueue_attrs *attrs = wq->unbound_attrs; |
6676 | ||
6677 | if (attrs) { | |
6678 | const struct wq_pod_type *pt = wqattrs_pod_type(attrs); | |
6679 | int tcpu; | |
4cbfd3de | 6680 | |
84193c07 | 6681 | for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]]) |
b2b1f933 | 6682 | unbound_wq_update_pwq(wq, tcpu); |
5797b1c1 TH |
6683 | |
6684 | mutex_lock(&wq->mutex); | |
6685 | wq_update_node_max_active(wq, cpu); | |
6686 | mutex_unlock(&wq->mutex); | |
4cbfd3de TH |
6687 | } |
6688 | } | |
7ee681b2 TG |
6689 | mutex_unlock(&wq_pool_mutex); |
6690 | ||
7ee681b2 | 6691 | return 0; |
6ba94429 FW |
6692 | } |
6693 | ||
6ba94429 FW |
6694 | struct work_for_cpu { |
6695 | struct work_struct work; | |
6696 | long (*fn)(void *); | |
6697 | void *arg; | |
6698 | long ret; | |
6699 | }; | |
6700 | ||
6701 | static void work_for_cpu_fn(struct work_struct *work) | |
6702 | { | |
6703 | struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); | |
6704 | ||
6705 | wfc->ret = wfc->fn(wfc->arg); | |
6706 | } | |
6707 | ||
6708 | /** | |
265f3ed0 | 6709 | * work_on_cpu_key - run a function in thread context on a particular cpu |
6ba94429 FW |
6710 | * @cpu: the cpu to run on |
6711 | * @fn: the function to run | |
6712 | * @arg: the function arg | |
265f3ed0 | 6713 | * @key: The lock class key for lock debugging purposes |
6ba94429 FW |
6714 | * |
6715 | * It is up to the caller to ensure that the cpu doesn't go offline. | |
6716 | * The caller must not hold any locks which would prevent @fn from completing. | |
6717 | * | |
6718 | * Return: The value @fn returns. | |
6719 | */ | |
265f3ed0 FW |
6720 | long work_on_cpu_key(int cpu, long (*fn)(void *), |
6721 | void *arg, struct lock_class_key *key) | |
6ba94429 FW |
6722 | { |
6723 | struct work_for_cpu wfc = { .fn = fn, .arg = arg }; | |
6724 | ||
265f3ed0 | 6725 | INIT_WORK_ONSTACK_KEY(&wfc.work, work_for_cpu_fn, key); |
6ba94429 FW |
6726 | schedule_work_on(cpu, &wfc.work); |
6727 | flush_work(&wfc.work); | |
6728 | destroy_work_on_stack(&wfc.work); | |
6729 | return wfc.ret; | |
6730 | } | |
265f3ed0 | 6731 | EXPORT_SYMBOL_GPL(work_on_cpu_key); |
0e8d6a93 TG |
6732 | |
6733 | /** | |
265f3ed0 | 6734 | * work_on_cpu_safe_key - run a function in thread context on a particular cpu |
0e8d6a93 TG |
6735 | * @cpu: the cpu to run on |
6736 | * @fn: the function to run | |
6737 | * @arg: the function argument | |
265f3ed0 | 6738 | * @key: The lock class key for lock debugging purposes |
0e8d6a93 TG |
6739 | * |
6740 | * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold | |
6741 | * any locks which would prevent @fn from completing. | |
6742 | * | |
6743 | * Return: The value @fn returns. | |
6744 | */ | |
265f3ed0 FW |
6745 | long work_on_cpu_safe_key(int cpu, long (*fn)(void *), |
6746 | void *arg, struct lock_class_key *key) | |
0e8d6a93 TG |
6747 | { |
6748 | long ret = -ENODEV; | |
6749 | ||
ffd8bea8 | 6750 | cpus_read_lock(); |
0e8d6a93 | 6751 | if (cpu_online(cpu)) |
265f3ed0 | 6752 | ret = work_on_cpu_key(cpu, fn, arg, key); |
ffd8bea8 | 6753 | cpus_read_unlock(); |
0e8d6a93 TG |
6754 | return ret; |
6755 | } | |
265f3ed0 | 6756 | EXPORT_SYMBOL_GPL(work_on_cpu_safe_key); |
6ba94429 FW |
6757 | #endif /* CONFIG_SMP */ |
6758 | ||
6759 | #ifdef CONFIG_FREEZER | |
6760 | ||
6761 | /** | |
6762 | * freeze_workqueues_begin - begin freezing workqueues | |
6763 | * | |
6764 | * Start freezing workqueues. After this function returns, all freezable | |
f97a4a1a | 6765 | * workqueues will queue new works to their inactive_works list instead of |
6ba94429 FW |
6766 | * pool->worklist. |
6767 | * | |
6768 | * CONTEXT: | |
6769 | * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. | |
6770 | */ | |
6771 | void freeze_workqueues_begin(void) | |
6772 | { | |
6773 | struct workqueue_struct *wq; | |
6ba94429 FW |
6774 | |
6775 | mutex_lock(&wq_pool_mutex); | |
6776 | ||
6777 | WARN_ON_ONCE(workqueue_freezing); | |
6778 | workqueue_freezing = true; | |
6779 | ||
6780 | list_for_each_entry(wq, &workqueues, list) { | |
6781 | mutex_lock(&wq->mutex); | |
a045a272 | 6782 | wq_adjust_max_active(wq); |
6ba94429 FW |
6783 | mutex_unlock(&wq->mutex); |
6784 | } | |
6785 | ||
6786 | mutex_unlock(&wq_pool_mutex); | |
6787 | } | |
6788 | ||
6789 | /** | |
6790 | * freeze_workqueues_busy - are freezable workqueues still busy? | |
6791 | * | |
6792 | * Check whether freezing is complete. This function must be called | |
6793 | * between freeze_workqueues_begin() and thaw_workqueues(). | |
6794 | * | |
6795 | * CONTEXT: | |
6796 | * Grabs and releases wq_pool_mutex. | |
6797 | * | |
6798 | * Return: | |
6799 | * %true if some freezable workqueues are still busy. %false if freezing | |
6800 | * is complete. | |
6801 | */ | |
6802 | bool freeze_workqueues_busy(void) | |
6803 | { | |
6804 | bool busy = false; | |
6805 | struct workqueue_struct *wq; | |
6806 | struct pool_workqueue *pwq; | |
6807 | ||
6808 | mutex_lock(&wq_pool_mutex); | |
6809 | ||
6810 | WARN_ON_ONCE(!workqueue_freezing); | |
6811 | ||
6812 | list_for_each_entry(wq, &workqueues, list) { | |
6813 | if (!(wq->flags & WQ_FREEZABLE)) | |
6814 | continue; | |
6815 | /* | |
6816 | * nr_active is monotonically decreasing. It's safe | |
6817 | * to peek without lock. | |
6818 | */ | |
24acfb71 | 6819 | rcu_read_lock(); |
6ba94429 FW |
6820 | for_each_pwq(pwq, wq) { |
6821 | WARN_ON_ONCE(pwq->nr_active < 0); | |
6822 | if (pwq->nr_active) { | |
6823 | busy = true; | |
24acfb71 | 6824 | rcu_read_unlock(); |
6ba94429 FW |
6825 | goto out_unlock; |
6826 | } | |
6827 | } | |
24acfb71 | 6828 | rcu_read_unlock(); |
6ba94429 FW |
6829 | } |
6830 | out_unlock: | |
6831 | mutex_unlock(&wq_pool_mutex); | |
6832 | return busy; | |
6833 | } | |
6834 | ||
6835 | /** | |
6836 | * thaw_workqueues - thaw workqueues | |
6837 | * | |
6838 | * Thaw workqueues. Normal queueing is restored and all collected | |
6839 | * frozen works are transferred to their respective pool worklists. | |
6840 | * | |
6841 | * CONTEXT: | |
6842 | * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. | |
6843 | */ | |
6844 | void thaw_workqueues(void) | |
6845 | { | |
6846 | struct workqueue_struct *wq; | |
6ba94429 FW |
6847 | |
6848 | mutex_lock(&wq_pool_mutex); | |
6849 | ||
6850 | if (!workqueue_freezing) | |
6851 | goto out_unlock; | |
6852 | ||
6853 | workqueue_freezing = false; | |
6854 | ||
6855 | /* restore max_active and repopulate worklist */ | |
6856 | list_for_each_entry(wq, &workqueues, list) { | |
6857 | mutex_lock(&wq->mutex); | |
a045a272 | 6858 | wq_adjust_max_active(wq); |
6ba94429 FW |
6859 | mutex_unlock(&wq->mutex); |
6860 | } | |
6861 | ||
6862 | out_unlock: | |
6863 | mutex_unlock(&wq_pool_mutex); | |
6864 | } | |
6865 | #endif /* CONFIG_FREEZER */ | |
6866 | ||
99c621ef | 6867 | static int workqueue_apply_unbound_cpumask(const cpumask_var_t unbound_cpumask) |
042f7df1 LJ |
6868 | { |
6869 | LIST_HEAD(ctxs); | |
6870 | int ret = 0; | |
6871 | struct workqueue_struct *wq; | |
6872 | struct apply_wqattrs_ctx *ctx, *n; | |
6873 | ||
6874 | lockdep_assert_held(&wq_pool_mutex); | |
6875 | ||
6876 | list_for_each_entry(wq, &workqueues, list) { | |
8eb17dc1 | 6877 | if (!(wq->flags & WQ_UNBOUND) || (wq->flags & __WQ_DESTROYING)) |
042f7df1 | 6878 | continue; |
042f7df1 | 6879 | |
99c621ef | 6880 | ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs, unbound_cpumask); |
84193c07 TH |
6881 | if (IS_ERR(ctx)) { |
6882 | ret = PTR_ERR(ctx); | |
042f7df1 LJ |
6883 | break; |
6884 | } | |
6885 | ||
6886 | list_add_tail(&ctx->list, &ctxs); | |
6887 | } | |
6888 | ||
6889 | list_for_each_entry_safe(ctx, n, &ctxs, list) { | |
6890 | if (!ret) | |
6891 | apply_wqattrs_commit(ctx); | |
6892 | apply_wqattrs_cleanup(ctx); | |
6893 | } | |
6894 | ||
99c621ef LJ |
6895 | if (!ret) { |
6896 | mutex_lock(&wq_pool_attach_mutex); | |
6897 | cpumask_copy(wq_unbound_cpumask, unbound_cpumask); | |
6898 | mutex_unlock(&wq_pool_attach_mutex); | |
6899 | } | |
042f7df1 LJ |
6900 | return ret; |
6901 | } | |
6902 | ||
fe28f631 WL |
6903 | /** |
6904 | * workqueue_unbound_exclude_cpumask - Exclude given CPUs from unbound cpumask | |
6905 | * @exclude_cpumask: the cpumask to be excluded from wq_unbound_cpumask | |
6906 | * | |
6907 | * This function can be called from cpuset code to provide a set of isolated | |
aa868475 | 6908 | * CPUs that should be excluded from wq_unbound_cpumask. |
fe28f631 WL |
6909 | */ |
6910 | int workqueue_unbound_exclude_cpumask(cpumask_var_t exclude_cpumask) | |
6911 | { | |
6912 | cpumask_var_t cpumask; | |
6913 | int ret = 0; | |
6914 | ||
6915 | if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) | |
6916 | return -ENOMEM; | |
6917 | ||
fe28f631 WL |
6918 | mutex_lock(&wq_pool_mutex); |
6919 | ||
fe28f631 WL |
6920 | /* |
6921 | * If the operation fails, it will fall back to | |
6922 | * wq_requested_unbound_cpumask which is initially set to | |
6923 | * (HK_TYPE_WQ ∩ HK_TYPE_DOMAIN) house keeping mask and rewritten | |
6924 | * by any subsequent write to workqueue/cpumask sysfs file. | |
6925 | */ | |
6926 | if (!cpumask_andnot(cpumask, wq_requested_unbound_cpumask, exclude_cpumask)) | |
6927 | cpumask_copy(cpumask, wq_requested_unbound_cpumask); | |
6928 | if (!cpumask_equal(cpumask, wq_unbound_cpumask)) | |
6929 | ret = workqueue_apply_unbound_cpumask(cpumask); | |
6930 | ||
84165883 LJ |
6931 | /* Save the current isolated cpumask & export it via sysfs */ |
6932 | if (!ret) | |
6933 | cpumask_copy(wq_isolated_cpumask, exclude_cpumask); | |
6934 | ||
fe28f631 WL |
6935 | mutex_unlock(&wq_pool_mutex); |
6936 | free_cpumask_var(cpumask); | |
6937 | return ret; | |
6938 | } | |
6939 | ||
63c5484e TH |
6940 | static int parse_affn_scope(const char *val) |
6941 | { | |
6942 | int i; | |
6943 | ||
6944 | for (i = 0; i < ARRAY_SIZE(wq_affn_names); i++) { | |
6945 | if (!strncasecmp(val, wq_affn_names[i], strlen(wq_affn_names[i]))) | |
6946 | return i; | |
6947 | } | |
6948 | return -EINVAL; | |
6949 | } | |
6950 | ||
6951 | static int wq_affn_dfl_set(const char *val, const struct kernel_param *kp) | |
6952 | { | |
523a301e TH |
6953 | struct workqueue_struct *wq; |
6954 | int affn, cpu; | |
63c5484e TH |
6955 | |
6956 | affn = parse_affn_scope(val); | |
6957 | if (affn < 0) | |
6958 | return affn; | |
523a301e TH |
6959 | if (affn == WQ_AFFN_DFL) |
6960 | return -EINVAL; | |
6961 | ||
6962 | cpus_read_lock(); | |
6963 | mutex_lock(&wq_pool_mutex); | |
63c5484e TH |
6964 | |
6965 | wq_affn_dfl = affn; | |
523a301e TH |
6966 | |
6967 | list_for_each_entry(wq, &workqueues, list) { | |
b2b1f933 LJ |
6968 | for_each_online_cpu(cpu) |
6969 | unbound_wq_update_pwq(wq, cpu); | |
523a301e TH |
6970 | } |
6971 | ||
6972 | mutex_unlock(&wq_pool_mutex); | |
6973 | cpus_read_unlock(); | |
6974 | ||
63c5484e TH |
6975 | return 0; |
6976 | } | |
6977 | ||
6978 | static int wq_affn_dfl_get(char *buffer, const struct kernel_param *kp) | |
6979 | { | |
6980 | return scnprintf(buffer, PAGE_SIZE, "%s\n", wq_affn_names[wq_affn_dfl]); | |
6981 | } | |
6982 | ||
6983 | static const struct kernel_param_ops wq_affn_dfl_ops = { | |
6984 | .set = wq_affn_dfl_set, | |
6985 | .get = wq_affn_dfl_get, | |
6986 | }; | |
6987 | ||
6988 | module_param_cb(default_affinity_scope, &wq_affn_dfl_ops, NULL, 0644); | |
6989 | ||
6ba94429 FW |
6990 | #ifdef CONFIG_SYSFS |
6991 | /* | |
6992 | * Workqueues with WQ_SYSFS flag set is visible to userland via | |
6993 | * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the | |
6994 | * following attributes. | |
6995 | * | |
63c5484e TH |
6996 | * per_cpu RO bool : whether the workqueue is per-cpu or unbound |
6997 | * max_active RW int : maximum number of in-flight work items | |
6ba94429 FW |
6998 | * |
6999 | * Unbound workqueues have the following extra attributes. | |
7000 | * | |
63c5484e TH |
7001 | * nice RW int : nice value of the workers |
7002 | * cpumask RW mask : bitmask of allowed CPUs for the workers | |
7003 | * affinity_scope RW str : worker CPU affinity scope (cache, numa, none) | |
8639eceb | 7004 | * affinity_strict RW bool : worker CPU affinity is strict |
6ba94429 FW |
7005 | */ |
7006 | struct wq_device { | |
7007 | struct workqueue_struct *wq; | |
7008 | struct device dev; | |
7009 | }; | |
7010 | ||
7011 | static struct workqueue_struct *dev_to_wq(struct device *dev) | |
7012 | { | |
7013 | struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); | |
7014 | ||
7015 | return wq_dev->wq; | |
7016 | } | |
7017 | ||
7018 | static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, | |
7019 | char *buf) | |
7020 | { | |
7021 | struct workqueue_struct *wq = dev_to_wq(dev); | |
7022 | ||
7023 | return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); | |
7024 | } | |
7025 | static DEVICE_ATTR_RO(per_cpu); | |
7026 | ||
7027 | static ssize_t max_active_show(struct device *dev, | |
7028 | struct device_attribute *attr, char *buf) | |
7029 | { | |
7030 | struct workqueue_struct *wq = dev_to_wq(dev); | |
7031 | ||
7032 | return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); | |
7033 | } | |
7034 | ||
7035 | static ssize_t max_active_store(struct device *dev, | |
7036 | struct device_attribute *attr, const char *buf, | |
7037 | size_t count) | |
7038 | { | |
7039 | struct workqueue_struct *wq = dev_to_wq(dev); | |
7040 | int val; | |
7041 | ||
7042 | if (sscanf(buf, "%d", &val) != 1 || val <= 0) | |
7043 | return -EINVAL; | |
7044 | ||
7045 | workqueue_set_max_active(wq, val); | |
7046 | return count; | |
7047 | } | |
7048 | static DEVICE_ATTR_RW(max_active); | |
7049 | ||
7050 | static struct attribute *wq_sysfs_attrs[] = { | |
7051 | &dev_attr_per_cpu.attr, | |
7052 | &dev_attr_max_active.attr, | |
7053 | NULL, | |
7054 | }; | |
7055 | ATTRIBUTE_GROUPS(wq_sysfs); | |
7056 | ||
6ba94429 FW |
7057 | static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, |
7058 | char *buf) | |
7059 | { | |
7060 | struct workqueue_struct *wq = dev_to_wq(dev); | |
7061 | int written; | |
7062 | ||
7063 | mutex_lock(&wq->mutex); | |
7064 | written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); | |
7065 | mutex_unlock(&wq->mutex); | |
7066 | ||
7067 | return written; | |
7068 | } | |
7069 | ||
7070 | /* prepare workqueue_attrs for sysfs store operations */ | |
7071 | static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) | |
7072 | { | |
7073 | struct workqueue_attrs *attrs; | |
7074 | ||
899a94fe LJ |
7075 | lockdep_assert_held(&wq_pool_mutex); |
7076 | ||
be69d00d | 7077 | attrs = alloc_workqueue_attrs(); |
6ba94429 FW |
7078 | if (!attrs) |
7079 | return NULL; | |
7080 | ||
6ba94429 | 7081 | copy_workqueue_attrs(attrs, wq->unbound_attrs); |
6ba94429 FW |
7082 | return attrs; |
7083 | } | |
7084 | ||
7085 | static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, | |
7086 | const char *buf, size_t count) | |
7087 | { | |
7088 | struct workqueue_struct *wq = dev_to_wq(dev); | |
7089 | struct workqueue_attrs *attrs; | |
d4d3e257 LJ |
7090 | int ret = -ENOMEM; |
7091 | ||
7092 | apply_wqattrs_lock(); | |
6ba94429 FW |
7093 | |
7094 | attrs = wq_sysfs_prep_attrs(wq); | |
7095 | if (!attrs) | |
d4d3e257 | 7096 | goto out_unlock; |
6ba94429 FW |
7097 | |
7098 | if (sscanf(buf, "%d", &attrs->nice) == 1 && | |
7099 | attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) | |
d4d3e257 | 7100 | ret = apply_workqueue_attrs_locked(wq, attrs); |
6ba94429 FW |
7101 | else |
7102 | ret = -EINVAL; | |
7103 | ||
d4d3e257 LJ |
7104 | out_unlock: |
7105 | apply_wqattrs_unlock(); | |
6ba94429 FW |
7106 | free_workqueue_attrs(attrs); |
7107 | return ret ?: count; | |
7108 | } | |
7109 | ||
7110 | static ssize_t wq_cpumask_show(struct device *dev, | |
7111 | struct device_attribute *attr, char *buf) | |
7112 | { | |
7113 | struct workqueue_struct *wq = dev_to_wq(dev); | |
7114 | int written; | |
7115 | ||
7116 | mutex_lock(&wq->mutex); | |
7117 | written = scnprintf(buf, PAGE_SIZE, "%*pb\n", | |
7118 | cpumask_pr_args(wq->unbound_attrs->cpumask)); | |
7119 | mutex_unlock(&wq->mutex); | |
7120 | return written; | |
7121 | } | |
7122 | ||
7123 | static ssize_t wq_cpumask_store(struct device *dev, | |
7124 | struct device_attribute *attr, | |
7125 | const char *buf, size_t count) | |
7126 | { | |
7127 | struct workqueue_struct *wq = dev_to_wq(dev); | |
7128 | struct workqueue_attrs *attrs; | |
d4d3e257 LJ |
7129 | int ret = -ENOMEM; |
7130 | ||
7131 | apply_wqattrs_lock(); | |
6ba94429 FW |
7132 | |
7133 | attrs = wq_sysfs_prep_attrs(wq); | |
7134 | if (!attrs) | |
d4d3e257 | 7135 | goto out_unlock; |
6ba94429 FW |
7136 | |
7137 | ret = cpumask_parse(buf, attrs->cpumask); | |
7138 | if (!ret) | |
d4d3e257 | 7139 | ret = apply_workqueue_attrs_locked(wq, attrs); |
6ba94429 | 7140 | |
d4d3e257 LJ |
7141 | out_unlock: |
7142 | apply_wqattrs_unlock(); | |
6ba94429 FW |
7143 | free_workqueue_attrs(attrs); |
7144 | return ret ?: count; | |
7145 | } | |
7146 | ||
63c5484e TH |
7147 | static ssize_t wq_affn_scope_show(struct device *dev, |
7148 | struct device_attribute *attr, char *buf) | |
7149 | { | |
7150 | struct workqueue_struct *wq = dev_to_wq(dev); | |
7151 | int written; | |
7152 | ||
7153 | mutex_lock(&wq->mutex); | |
523a301e TH |
7154 | if (wq->unbound_attrs->affn_scope == WQ_AFFN_DFL) |
7155 | written = scnprintf(buf, PAGE_SIZE, "%s (%s)\n", | |
7156 | wq_affn_names[WQ_AFFN_DFL], | |
7157 | wq_affn_names[wq_affn_dfl]); | |
7158 | else | |
7159 | written = scnprintf(buf, PAGE_SIZE, "%s\n", | |
7160 | wq_affn_names[wq->unbound_attrs->affn_scope]); | |
63c5484e TH |
7161 | mutex_unlock(&wq->mutex); |
7162 | ||
7163 | return written; | |
7164 | } | |
7165 | ||
7166 | static ssize_t wq_affn_scope_store(struct device *dev, | |
7167 | struct device_attribute *attr, | |
7168 | const char *buf, size_t count) | |
7169 | { | |
7170 | struct workqueue_struct *wq = dev_to_wq(dev); | |
7171 | struct workqueue_attrs *attrs; | |
7172 | int affn, ret = -ENOMEM; | |
7173 | ||
7174 | affn = parse_affn_scope(buf); | |
7175 | if (affn < 0) | |
7176 | return affn; | |
7177 | ||
7178 | apply_wqattrs_lock(); | |
7179 | attrs = wq_sysfs_prep_attrs(wq); | |
7180 | if (attrs) { | |
7181 | attrs->affn_scope = affn; | |
7182 | ret = apply_workqueue_attrs_locked(wq, attrs); | |
7183 | } | |
7184 | apply_wqattrs_unlock(); | |
7185 | free_workqueue_attrs(attrs); | |
7186 | return ret ?: count; | |
7187 | } | |
7188 | ||
8639eceb TH |
7189 | static ssize_t wq_affinity_strict_show(struct device *dev, |
7190 | struct device_attribute *attr, char *buf) | |
7191 | { | |
7192 | struct workqueue_struct *wq = dev_to_wq(dev); | |
7193 | ||
7194 | return scnprintf(buf, PAGE_SIZE, "%d\n", | |
7195 | wq->unbound_attrs->affn_strict); | |
7196 | } | |
7197 | ||
7198 | static ssize_t wq_affinity_strict_store(struct device *dev, | |
7199 | struct device_attribute *attr, | |
7200 | const char *buf, size_t count) | |
7201 | { | |
7202 | struct workqueue_struct *wq = dev_to_wq(dev); | |
7203 | struct workqueue_attrs *attrs; | |
7204 | int v, ret = -ENOMEM; | |
7205 | ||
7206 | if (sscanf(buf, "%d", &v) != 1) | |
7207 | return -EINVAL; | |
7208 | ||
7209 | apply_wqattrs_lock(); | |
7210 | attrs = wq_sysfs_prep_attrs(wq); | |
7211 | if (attrs) { | |
7212 | attrs->affn_strict = (bool)v; | |
7213 | ret = apply_workqueue_attrs_locked(wq, attrs); | |
7214 | } | |
7215 | apply_wqattrs_unlock(); | |
7216 | free_workqueue_attrs(attrs); | |
7217 | return ret ?: count; | |
7218 | } | |
7219 | ||
6ba94429 | 7220 | static struct device_attribute wq_sysfs_unbound_attrs[] = { |
6ba94429 FW |
7221 | __ATTR(nice, 0644, wq_nice_show, wq_nice_store), |
7222 | __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), | |
63c5484e | 7223 | __ATTR(affinity_scope, 0644, wq_affn_scope_show, wq_affn_scope_store), |
8639eceb | 7224 | __ATTR(affinity_strict, 0644, wq_affinity_strict_show, wq_affinity_strict_store), |
6ba94429 FW |
7225 | __ATTR_NULL, |
7226 | }; | |
8ccad40d | 7227 | |
5df9197e | 7228 | static const struct bus_type wq_subsys = { |
6ba94429 FW |
7229 | .name = "workqueue", |
7230 | .dev_groups = wq_sysfs_groups, | |
2d3854a3 RR |
7231 | }; |
7232 | ||
49277a5b WL |
7233 | /** |
7234 | * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask | |
7235 | * @cpumask: the cpumask to set | |
7236 | * | |
7237 | * The low-level workqueues cpumask is a global cpumask that limits | |
7238 | * the affinity of all unbound workqueues. This function check the @cpumask | |
7239 | * and apply it to all unbound workqueues and updates all pwqs of them. | |
7240 | * | |
7241 | * Return: 0 - Success | |
7242 | * -EINVAL - Invalid @cpumask | |
7243 | * -ENOMEM - Failed to allocate memory for attrs or pwqs. | |
7244 | */ | |
7245 | static int workqueue_set_unbound_cpumask(cpumask_var_t cpumask) | |
7246 | { | |
7247 | int ret = -EINVAL; | |
7248 | ||
7249 | /* | |
7250 | * Not excluding isolated cpus on purpose. | |
7251 | * If the user wishes to include them, we allow that. | |
7252 | */ | |
7253 | cpumask_and(cpumask, cpumask, cpu_possible_mask); | |
7254 | if (!cpumask_empty(cpumask)) { | |
b3d20916 | 7255 | ret = 0; |
49277a5b | 7256 | apply_wqattrs_lock(); |
b3d20916 LJ |
7257 | if (!cpumask_equal(cpumask, wq_unbound_cpumask)) |
7258 | ret = workqueue_apply_unbound_cpumask(cpumask); | |
84165883 LJ |
7259 | if (!ret) |
7260 | cpumask_copy(wq_requested_unbound_cpumask, cpumask); | |
49277a5b WL |
7261 | apply_wqattrs_unlock(); |
7262 | } | |
7263 | ||
7264 | return ret; | |
7265 | } | |
7266 | ||
fe28f631 WL |
7267 | static ssize_t __wq_cpumask_show(struct device *dev, |
7268 | struct device_attribute *attr, char *buf, cpumask_var_t mask) | |
b05a7928 FW |
7269 | { |
7270 | int written; | |
7271 | ||
042f7df1 | 7272 | mutex_lock(&wq_pool_mutex); |
fe28f631 | 7273 | written = scnprintf(buf, PAGE_SIZE, "%*pb\n", cpumask_pr_args(mask)); |
042f7df1 | 7274 | mutex_unlock(&wq_pool_mutex); |
b05a7928 FW |
7275 | |
7276 | return written; | |
7277 | } | |
7278 | ||
79202591 | 7279 | static ssize_t cpumask_requested_show(struct device *dev, |
fe28f631 WL |
7280 | struct device_attribute *attr, char *buf) |
7281 | { | |
79202591 | 7282 | return __wq_cpumask_show(dev, attr, buf, wq_requested_unbound_cpumask); |
fe28f631 | 7283 | } |
79202591 | 7284 | static DEVICE_ATTR_RO(cpumask_requested); |
fe28f631 | 7285 | |
79202591 | 7286 | static ssize_t cpumask_isolated_show(struct device *dev, |
fe28f631 WL |
7287 | struct device_attribute *attr, char *buf) |
7288 | { | |
79202591 | 7289 | return __wq_cpumask_show(dev, attr, buf, wq_isolated_cpumask); |
fe28f631 | 7290 | } |
79202591 | 7291 | static DEVICE_ATTR_RO(cpumask_isolated); |
fe28f631 | 7292 | |
79202591 | 7293 | static ssize_t cpumask_show(struct device *dev, |
fe28f631 WL |
7294 | struct device_attribute *attr, char *buf) |
7295 | { | |
79202591 | 7296 | return __wq_cpumask_show(dev, attr, buf, wq_unbound_cpumask); |
fe28f631 WL |
7297 | } |
7298 | ||
79202591 | 7299 | static ssize_t cpumask_store(struct device *dev, |
042f7df1 LJ |
7300 | struct device_attribute *attr, const char *buf, size_t count) |
7301 | { | |
7302 | cpumask_var_t cpumask; | |
7303 | int ret; | |
7304 | ||
7305 | if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) | |
7306 | return -ENOMEM; | |
7307 | ||
7308 | ret = cpumask_parse(buf, cpumask); | |
7309 | if (!ret) | |
7310 | ret = workqueue_set_unbound_cpumask(cpumask); | |
7311 | ||
7312 | free_cpumask_var(cpumask); | |
7313 | return ret ? ret : count; | |
7314 | } | |
79202591 | 7315 | static DEVICE_ATTR_RW(cpumask); |
042f7df1 | 7316 | |
79202591 DW |
7317 | static struct attribute *wq_sysfs_cpumask_attrs[] = { |
7318 | &dev_attr_cpumask.attr, | |
7319 | &dev_attr_cpumask_requested.attr, | |
7320 | &dev_attr_cpumask_isolated.attr, | |
7321 | NULL, | |
fe28f631 | 7322 | }; |
79202591 | 7323 | ATTRIBUTE_GROUPS(wq_sysfs_cpumask); |
b05a7928 | 7324 | |
6ba94429 | 7325 | static int __init wq_sysfs_init(void) |
2d3854a3 | 7326 | { |
79202591 | 7327 | return subsys_virtual_register(&wq_subsys, wq_sysfs_cpumask_groups); |
2d3854a3 | 7328 | } |
6ba94429 | 7329 | core_initcall(wq_sysfs_init); |
2d3854a3 | 7330 | |
6ba94429 | 7331 | static void wq_device_release(struct device *dev) |
2d3854a3 | 7332 | { |
6ba94429 | 7333 | struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); |
6b44003e | 7334 | |
6ba94429 | 7335 | kfree(wq_dev); |
2d3854a3 | 7336 | } |
a0a1a5fd TH |
7337 | |
7338 | /** | |
6ba94429 FW |
7339 | * workqueue_sysfs_register - make a workqueue visible in sysfs |
7340 | * @wq: the workqueue to register | |
a0a1a5fd | 7341 | * |
6ba94429 FW |
7342 | * Expose @wq in sysfs under /sys/bus/workqueue/devices. |
7343 | * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set | |
7344 | * which is the preferred method. | |
a0a1a5fd | 7345 | * |
6ba94429 FW |
7346 | * Workqueue user should use this function directly iff it wants to apply |
7347 | * workqueue_attrs before making the workqueue visible in sysfs; otherwise, | |
7348 | * apply_workqueue_attrs() may race against userland updating the | |
7349 | * attributes. | |
7350 | * | |
7351 | * Return: 0 on success, -errno on failure. | |
a0a1a5fd | 7352 | */ |
6ba94429 | 7353 | int workqueue_sysfs_register(struct workqueue_struct *wq) |
a0a1a5fd | 7354 | { |
6ba94429 FW |
7355 | struct wq_device *wq_dev; |
7356 | int ret; | |
a0a1a5fd | 7357 | |
6ba94429 | 7358 | /* |
4c065dbc WL |
7359 | * Adjusting max_active breaks ordering guarantee. Disallow exposing |
7360 | * ordered workqueues. | |
6ba94429 | 7361 | */ |
3bc1e711 | 7362 | if (WARN_ON(wq->flags & __WQ_ORDERED)) |
6ba94429 | 7363 | return -EINVAL; |
a0a1a5fd | 7364 | |
6ba94429 FW |
7365 | wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); |
7366 | if (!wq_dev) | |
7367 | return -ENOMEM; | |
5bcab335 | 7368 | |
6ba94429 FW |
7369 | wq_dev->wq = wq; |
7370 | wq_dev->dev.bus = &wq_subsys; | |
6ba94429 | 7371 | wq_dev->dev.release = wq_device_release; |
23217b44 | 7372 | dev_set_name(&wq_dev->dev, "%s", wq->name); |
a0a1a5fd | 7373 | |
6ba94429 FW |
7374 | /* |
7375 | * unbound_attrs are created separately. Suppress uevent until | |
7376 | * everything is ready. | |
7377 | */ | |
7378 | dev_set_uevent_suppress(&wq_dev->dev, true); | |
a0a1a5fd | 7379 | |
6ba94429 FW |
7380 | ret = device_register(&wq_dev->dev); |
7381 | if (ret) { | |
537f4146 | 7382 | put_device(&wq_dev->dev); |
6ba94429 FW |
7383 | wq->wq_dev = NULL; |
7384 | return ret; | |
7385 | } | |
a0a1a5fd | 7386 | |
6ba94429 FW |
7387 | if (wq->flags & WQ_UNBOUND) { |
7388 | struct device_attribute *attr; | |
a0a1a5fd | 7389 | |
6ba94429 FW |
7390 | for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { |
7391 | ret = device_create_file(&wq_dev->dev, attr); | |
7392 | if (ret) { | |
7393 | device_unregister(&wq_dev->dev); | |
7394 | wq->wq_dev = NULL; | |
7395 | return ret; | |
a0a1a5fd TH |
7396 | } |
7397 | } | |
7398 | } | |
6ba94429 FW |
7399 | |
7400 | dev_set_uevent_suppress(&wq_dev->dev, false); | |
7401 | kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); | |
7402 | return 0; | |
a0a1a5fd TH |
7403 | } |
7404 | ||
7405 | /** | |
6ba94429 FW |
7406 | * workqueue_sysfs_unregister - undo workqueue_sysfs_register() |
7407 | * @wq: the workqueue to unregister | |
a0a1a5fd | 7408 | * |
6ba94429 | 7409 | * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. |
a0a1a5fd | 7410 | */ |
6ba94429 | 7411 | static void workqueue_sysfs_unregister(struct workqueue_struct *wq) |
a0a1a5fd | 7412 | { |
6ba94429 | 7413 | struct wq_device *wq_dev = wq->wq_dev; |
8b03ae3c | 7414 | |
6ba94429 FW |
7415 | if (!wq->wq_dev) |
7416 | return; | |
a0a1a5fd | 7417 | |
6ba94429 FW |
7418 | wq->wq_dev = NULL; |
7419 | device_unregister(&wq_dev->dev); | |
a0a1a5fd | 7420 | } |
6ba94429 FW |
7421 | #else /* CONFIG_SYSFS */ |
7422 | static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } | |
7423 | #endif /* CONFIG_SYSFS */ | |
a0a1a5fd | 7424 | |
82607adc TH |
7425 | /* |
7426 | * Workqueue watchdog. | |
7427 | * | |
7428 | * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal | |
7429 | * flush dependency, a concurrency managed work item which stays RUNNING | |
7430 | * indefinitely. Workqueue stalls can be very difficult to debug as the | |
7431 | * usual warning mechanisms don't trigger and internal workqueue state is | |
7432 | * largely opaque. | |
7433 | * | |
7434 | * Workqueue watchdog monitors all worker pools periodically and dumps | |
7435 | * state if some pools failed to make forward progress for a while where | |
7436 | * forward progress is defined as the first item on ->worklist changing. | |
7437 | * | |
7438 | * This mechanism is controlled through the kernel parameter | |
7439 | * "workqueue.watchdog_thresh" which can be updated at runtime through the | |
7440 | * corresponding sysfs parameter file. | |
7441 | */ | |
7442 | #ifdef CONFIG_WQ_WATCHDOG | |
7443 | ||
82607adc | 7444 | static unsigned long wq_watchdog_thresh = 30; |
5cd79d6a | 7445 | static struct timer_list wq_watchdog_timer; |
82607adc TH |
7446 | |
7447 | static unsigned long wq_watchdog_touched = INITIAL_JIFFIES; | |
7448 | static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES; | |
7449 | ||
073107b3 SK |
7450 | static unsigned int wq_panic_on_stall; |
7451 | module_param_named(panic_on_stall, wq_panic_on_stall, uint, 0644); | |
7452 | ||
cd2440d6 PM |
7453 | /* |
7454 | * Show workers that might prevent the processing of pending work items. | |
7455 | * The only candidates are CPU-bound workers in the running state. | |
7456 | * Pending work items should be handled by another idle worker | |
7457 | * in all other situations. | |
7458 | */ | |
7459 | static void show_cpu_pool_hog(struct worker_pool *pool) | |
7460 | { | |
7461 | struct worker *worker; | |
c26e2f2e | 7462 | unsigned long irq_flags; |
cd2440d6 PM |
7463 | int bkt; |
7464 | ||
c26e2f2e | 7465 | raw_spin_lock_irqsave(&pool->lock, irq_flags); |
cd2440d6 PM |
7466 | |
7467 | hash_for_each(pool->busy_hash, bkt, worker, hentry) { | |
7468 | if (task_is_running(worker->task)) { | |
7469 | /* | |
7470 | * Defer printing to avoid deadlocks in console | |
7471 | * drivers that queue work while holding locks | |
7472 | * also taken in their write paths. | |
7473 | */ | |
7474 | printk_deferred_enter(); | |
7475 | ||
7476 | pr_info("pool %d:\n", pool->id); | |
7477 | sched_show_task(worker->task); | |
7478 | ||
7479 | printk_deferred_exit(); | |
7480 | } | |
7481 | } | |
7482 | ||
c26e2f2e | 7483 | raw_spin_unlock_irqrestore(&pool->lock, irq_flags); |
cd2440d6 PM |
7484 | } |
7485 | ||
7486 | static void show_cpu_pools_hogs(void) | |
7487 | { | |
7488 | struct worker_pool *pool; | |
7489 | int pi; | |
7490 | ||
7491 | pr_info("Showing backtraces of running workers in stalled CPU-bound worker pools:\n"); | |
7492 | ||
7493 | rcu_read_lock(); | |
7494 | ||
7495 | for_each_pool(pool, pi) { | |
7496 | if (pool->cpu_stall) | |
7497 | show_cpu_pool_hog(pool); | |
7498 | ||
7499 | } | |
7500 | ||
7501 | rcu_read_unlock(); | |
7502 | } | |
7503 | ||
073107b3 SK |
7504 | static void panic_on_wq_watchdog(void) |
7505 | { | |
7506 | static unsigned int wq_stall; | |
7507 | ||
7508 | if (wq_panic_on_stall) { | |
7509 | wq_stall++; | |
7510 | BUG_ON(wq_stall >= wq_panic_on_stall); | |
7511 | } | |
7512 | } | |
7513 | ||
82607adc TH |
7514 | static void wq_watchdog_reset_touched(void) |
7515 | { | |
7516 | int cpu; | |
7517 | ||
7518 | wq_watchdog_touched = jiffies; | |
7519 | for_each_possible_cpu(cpu) | |
7520 | per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; | |
7521 | } | |
7522 | ||
5cd79d6a | 7523 | static void wq_watchdog_timer_fn(struct timer_list *unused) |
82607adc TH |
7524 | { |
7525 | unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ; | |
7526 | bool lockup_detected = false; | |
cd2440d6 | 7527 | bool cpu_pool_stall = false; |
940d71c6 | 7528 | unsigned long now = jiffies; |
82607adc TH |
7529 | struct worker_pool *pool; |
7530 | int pi; | |
7531 | ||
7532 | if (!thresh) | |
7533 | return; | |
7534 | ||
7535 | rcu_read_lock(); | |
7536 | ||
7537 | for_each_pool(pool, pi) { | |
7538 | unsigned long pool_ts, touched, ts; | |
7539 | ||
cd2440d6 | 7540 | pool->cpu_stall = false; |
82607adc TH |
7541 | if (list_empty(&pool->worklist)) |
7542 | continue; | |
7543 | ||
940d71c6 SS |
7544 | /* |
7545 | * If a virtual machine is stopped by the host it can look to | |
7546 | * the watchdog like a stall. | |
7547 | */ | |
7548 | kvm_check_and_clear_guest_paused(); | |
7549 | ||
82607adc | 7550 | /* get the latest of pool and touched timestamps */ |
89e28ce6 WQ |
7551 | if (pool->cpu >= 0) |
7552 | touched = READ_ONCE(per_cpu(wq_watchdog_touched_cpu, pool->cpu)); | |
7553 | else | |
7554 | touched = READ_ONCE(wq_watchdog_touched); | |
82607adc | 7555 | pool_ts = READ_ONCE(pool->watchdog_ts); |
82607adc TH |
7556 | |
7557 | if (time_after(pool_ts, touched)) | |
7558 | ts = pool_ts; | |
7559 | else | |
7560 | ts = touched; | |
7561 | ||
82607adc | 7562 | /* did we stall? */ |
940d71c6 | 7563 | if (time_after(now, ts + thresh)) { |
82607adc | 7564 | lockup_detected = true; |
4cb1ef64 | 7565 | if (pool->cpu >= 0 && !(pool->flags & POOL_BH)) { |
cd2440d6 PM |
7566 | pool->cpu_stall = true; |
7567 | cpu_pool_stall = true; | |
7568 | } | |
82607adc TH |
7569 | pr_emerg("BUG: workqueue lockup - pool"); |
7570 | pr_cont_pool_info(pool); | |
7571 | pr_cont(" stuck for %us!\n", | |
940d71c6 | 7572 | jiffies_to_msecs(now - pool_ts) / 1000); |
82607adc | 7573 | } |
cd2440d6 PM |
7574 | |
7575 | ||
82607adc TH |
7576 | } |
7577 | ||
7578 | rcu_read_unlock(); | |
7579 | ||
7580 | if (lockup_detected) | |
55df0933 | 7581 | show_all_workqueues(); |
82607adc | 7582 | |
cd2440d6 PM |
7583 | if (cpu_pool_stall) |
7584 | show_cpu_pools_hogs(); | |
7585 | ||
073107b3 SK |
7586 | if (lockup_detected) |
7587 | panic_on_wq_watchdog(); | |
7588 | ||
82607adc TH |
7589 | wq_watchdog_reset_touched(); |
7590 | mod_timer(&wq_watchdog_timer, jiffies + thresh); | |
7591 | } | |
7592 | ||
cb9d7fd5 | 7593 | notrace void wq_watchdog_touch(int cpu) |
82607adc | 7594 | { |
98f887f8 NP |
7595 | unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ; |
7596 | unsigned long touch_ts = READ_ONCE(wq_watchdog_touched); | |
7597 | unsigned long now = jiffies; | |
7598 | ||
82607adc | 7599 | if (cpu >= 0) |
98f887f8 | 7600 | per_cpu(wq_watchdog_touched_cpu, cpu) = now; |
18e24deb NP |
7601 | else |
7602 | WARN_ONCE(1, "%s should be called with valid CPU", __func__); | |
89e28ce6 | 7603 | |
98f887f8 NP |
7604 | /* Don't unnecessarily store to global cacheline */ |
7605 | if (time_after(now, touch_ts + thresh / 4)) | |
7606 | WRITE_ONCE(wq_watchdog_touched, jiffies); | |
82607adc TH |
7607 | } |
7608 | ||
7609 | static void wq_watchdog_set_thresh(unsigned long thresh) | |
7610 | { | |
7611 | wq_watchdog_thresh = 0; | |
7612 | del_timer_sync(&wq_watchdog_timer); | |
7613 | ||
7614 | if (thresh) { | |
7615 | wq_watchdog_thresh = thresh; | |
7616 | wq_watchdog_reset_touched(); | |
7617 | mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ); | |
7618 | } | |
7619 | } | |
7620 | ||
7621 | static int wq_watchdog_param_set_thresh(const char *val, | |
7622 | const struct kernel_param *kp) | |
7623 | { | |
7624 | unsigned long thresh; | |
7625 | int ret; | |
7626 | ||
7627 | ret = kstrtoul(val, 0, &thresh); | |
7628 | if (ret) | |
7629 | return ret; | |
7630 | ||
7631 | if (system_wq) | |
7632 | wq_watchdog_set_thresh(thresh); | |
7633 | else | |
7634 | wq_watchdog_thresh = thresh; | |
7635 | ||
7636 | return 0; | |
7637 | } | |
7638 | ||
7639 | static const struct kernel_param_ops wq_watchdog_thresh_ops = { | |
7640 | .set = wq_watchdog_param_set_thresh, | |
7641 | .get = param_get_ulong, | |
7642 | }; | |
7643 | ||
7644 | module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh, | |
7645 | 0644); | |
7646 | ||
7647 | static void wq_watchdog_init(void) | |
7648 | { | |
5cd79d6a | 7649 | timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE); |
82607adc TH |
7650 | wq_watchdog_set_thresh(wq_watchdog_thresh); |
7651 | } | |
7652 | ||
7653 | #else /* CONFIG_WQ_WATCHDOG */ | |
7654 | ||
7655 | static inline void wq_watchdog_init(void) { } | |
7656 | ||
7657 | #endif /* CONFIG_WQ_WATCHDOG */ | |
7658 | ||
2f34d733 TH |
7659 | static void bh_pool_kick_normal(struct irq_work *irq_work) |
7660 | { | |
7661 | raise_softirq_irqoff(TASKLET_SOFTIRQ); | |
7662 | } | |
7663 | ||
7664 | static void bh_pool_kick_highpri(struct irq_work *irq_work) | |
7665 | { | |
7666 | raise_softirq_irqoff(HI_SOFTIRQ); | |
7667 | } | |
7668 | ||
4a6c5607 TH |
7669 | static void __init restrict_unbound_cpumask(const char *name, const struct cpumask *mask) |
7670 | { | |
7671 | if (!cpumask_intersects(wq_unbound_cpumask, mask)) { | |
7672 | pr_warn("workqueue: Restricting unbound_cpumask (%*pb) with %s (%*pb) leaves no CPU, ignoring\n", | |
7673 | cpumask_pr_args(wq_unbound_cpumask), name, cpumask_pr_args(mask)); | |
7674 | return; | |
7675 | } | |
7676 | ||
7677 | cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, mask); | |
7678 | } | |
7679 | ||
2fcdb1b4 TH |
7680 | static void __init init_cpu_worker_pool(struct worker_pool *pool, int cpu, int nice) |
7681 | { | |
7682 | BUG_ON(init_worker_pool(pool)); | |
7683 | pool->cpu = cpu; | |
7684 | cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); | |
7685 | cpumask_copy(pool->attrs->__pod_cpumask, cpumask_of(cpu)); | |
7686 | pool->attrs->nice = nice; | |
7687 | pool->attrs->affn_strict = true; | |
7688 | pool->node = cpu_to_node(cpu); | |
7689 | ||
7690 | /* alloc pool ID */ | |
7691 | mutex_lock(&wq_pool_mutex); | |
7692 | BUG_ON(worker_pool_assign_id(pool)); | |
7693 | mutex_unlock(&wq_pool_mutex); | |
7694 | } | |
7695 | ||
3347fa09 TH |
7696 | /** |
7697 | * workqueue_init_early - early init for workqueue subsystem | |
7698 | * | |
2930155b TH |
7699 | * This is the first step of three-staged workqueue subsystem initialization and |
7700 | * invoked as soon as the bare basics - memory allocation, cpumasks and idr are | |
7701 | * up. It sets up all the data structures and system workqueues and allows early | |
7702 | * boot code to create workqueues and queue/cancel work items. Actual work item | |
7703 | * execution starts only after kthreads can be created and scheduled right | |
7704 | * before early initcalls. | |
3347fa09 | 7705 | */ |
2333e829 | 7706 | void __init workqueue_init_early(void) |
1da177e4 | 7707 | { |
84193c07 | 7708 | struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_SYSTEM]; |
7a4e344c | 7709 | int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; |
2f34d733 TH |
7710 | void (*irq_work_fns[2])(struct irq_work *) = { bh_pool_kick_normal, |
7711 | bh_pool_kick_highpri }; | |
7a4e344c | 7712 | int i, cpu; |
c34056a3 | 7713 | |
10cdb157 | 7714 | BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); |
e904e6c2 | 7715 | |
8d84baf7 | 7716 | BUG_ON(!alloc_cpumask_var(&wq_online_cpumask, GFP_KERNEL)); |
b05a7928 | 7717 | BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL)); |
fe28f631 WL |
7718 | BUG_ON(!alloc_cpumask_var(&wq_requested_unbound_cpumask, GFP_KERNEL)); |
7719 | BUG_ON(!zalloc_cpumask_var(&wq_isolated_cpumask, GFP_KERNEL)); | |
b05a7928 | 7720 | |
8d84baf7 | 7721 | cpumask_copy(wq_online_cpumask, cpu_online_mask); |
4a6c5607 TH |
7722 | cpumask_copy(wq_unbound_cpumask, cpu_possible_mask); |
7723 | restrict_unbound_cpumask("HK_TYPE_WQ", housekeeping_cpumask(HK_TYPE_WQ)); | |
7724 | restrict_unbound_cpumask("HK_TYPE_DOMAIN", housekeeping_cpumask(HK_TYPE_DOMAIN)); | |
ace3c549 | 7725 | if (!cpumask_empty(&wq_cmdline_cpumask)) |
4a6c5607 | 7726 | restrict_unbound_cpumask("workqueue.unbound_cpus", &wq_cmdline_cpumask); |
ace3c549 | 7727 | |
fe28f631 | 7728 | cpumask_copy(wq_requested_unbound_cpumask, wq_unbound_cpumask); |
ace3c549 | 7729 | |
e904e6c2 TH |
7730 | pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); |
7731 | ||
b2b1f933 LJ |
7732 | unbound_wq_update_pwq_attrs_buf = alloc_workqueue_attrs(); |
7733 | BUG_ON(!unbound_wq_update_pwq_attrs_buf); | |
2930155b | 7734 | |
7bd20b6b MT |
7735 | /* |
7736 | * If nohz_full is enabled, set power efficient workqueue as unbound. | |
7737 | * This allows workqueue items to be moved to HK CPUs. | |
7738 | */ | |
7739 | if (housekeeping_enabled(HK_TYPE_TICK)) | |
7740 | wq_power_efficient = true; | |
7741 | ||
84193c07 TH |
7742 | /* initialize WQ_AFFN_SYSTEM pods */ |
7743 | pt->pod_cpus = kcalloc(1, sizeof(pt->pod_cpus[0]), GFP_KERNEL); | |
7744 | pt->pod_node = kcalloc(1, sizeof(pt->pod_node[0]), GFP_KERNEL); | |
7745 | pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL); | |
7746 | BUG_ON(!pt->pod_cpus || !pt->pod_node || !pt->cpu_pod); | |
7747 | ||
7748 | BUG_ON(!zalloc_cpumask_var_node(&pt->pod_cpus[0], GFP_KERNEL, NUMA_NO_NODE)); | |
7749 | ||
84193c07 TH |
7750 | pt->nr_pods = 1; |
7751 | cpumask_copy(pt->pod_cpus[0], cpu_possible_mask); | |
7752 | pt->pod_node[0] = NUMA_NO_NODE; | |
7753 | pt->cpu_pod[0] = 0; | |
7754 | ||
4cb1ef64 | 7755 | /* initialize BH and CPU pools */ |
29c91e99 | 7756 | for_each_possible_cpu(cpu) { |
4ce62e9e | 7757 | struct worker_pool *pool; |
8b03ae3c | 7758 | |
4cb1ef64 TH |
7759 | i = 0; |
7760 | for_each_bh_worker_pool(pool, cpu) { | |
2f34d733 | 7761 | init_cpu_worker_pool(pool, cpu, std_nice[i]); |
4cb1ef64 | 7762 | pool->flags |= POOL_BH; |
2f34d733 TH |
7763 | init_irq_work(bh_pool_irq_work(pool), irq_work_fns[i]); |
7764 | i++; | |
4cb1ef64 TH |
7765 | } |
7766 | ||
7a4e344c | 7767 | i = 0; |
2fcdb1b4 TH |
7768 | for_each_cpu_worker_pool(pool, cpu) |
7769 | init_cpu_worker_pool(pool, cpu, std_nice[i++]); | |
8b03ae3c TH |
7770 | } |
7771 | ||
8a2b7538 | 7772 | /* create default unbound and ordered wq attrs */ |
29c91e99 TH |
7773 | for (i = 0; i < NR_STD_WORKER_POOLS; i++) { |
7774 | struct workqueue_attrs *attrs; | |
7775 | ||
be69d00d | 7776 | BUG_ON(!(attrs = alloc_workqueue_attrs())); |
29c91e99 | 7777 | attrs->nice = std_nice[i]; |
29c91e99 | 7778 | unbound_std_wq_attrs[i] = attrs; |
8a2b7538 TH |
7779 | |
7780 | /* | |
7781 | * An ordered wq should have only one pwq as ordering is | |
7782 | * guaranteed by max_active which is enforced by pwqs. | |
8a2b7538 | 7783 | */ |
be69d00d | 7784 | BUG_ON(!(attrs = alloc_workqueue_attrs())); |
8a2b7538 | 7785 | attrs->nice = std_nice[i]; |
af73f5c9 | 7786 | attrs->ordered = true; |
8a2b7538 | 7787 | ordered_wq_attrs[i] = attrs; |
29c91e99 TH |
7788 | } |
7789 | ||
d320c038 | 7790 | system_wq = alloc_workqueue("events", 0, 0); |
1aabe902 | 7791 | system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); |
d320c038 | 7792 | system_long_wq = alloc_workqueue("events_long", 0, 0); |
f3421797 | 7793 | system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, |
636b927e | 7794 | WQ_MAX_ACTIVE); |
24d51add TH |
7795 | system_freezable_wq = alloc_workqueue("events_freezable", |
7796 | WQ_FREEZABLE, 0); | |
0668106c VK |
7797 | system_power_efficient_wq = alloc_workqueue("events_power_efficient", |
7798 | WQ_POWER_EFFICIENT, 0); | |
8318d6a6 | 7799 | system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_pwr_efficient", |
0668106c VK |
7800 | WQ_FREEZABLE | WQ_POWER_EFFICIENT, |
7801 | 0); | |
4cb1ef64 TH |
7802 | system_bh_wq = alloc_workqueue("events_bh", WQ_BH, 0); |
7803 | system_bh_highpri_wq = alloc_workqueue("events_bh_highpri", | |
7804 | WQ_BH | WQ_HIGHPRI, 0); | |
1aabe902 | 7805 | BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || |
0668106c VK |
7806 | !system_unbound_wq || !system_freezable_wq || |
7807 | !system_power_efficient_wq || | |
4cb1ef64 TH |
7808 | !system_freezable_power_efficient_wq || |
7809 | !system_bh_wq || !system_bh_highpri_wq); | |
3347fa09 TH |
7810 | } |
7811 | ||
aa6fde93 TH |
7812 | static void __init wq_cpu_intensive_thresh_init(void) |
7813 | { | |
7814 | unsigned long thresh; | |
7815 | unsigned long bogo; | |
7816 | ||
dd64c873 Z |
7817 | pwq_release_worker = kthread_create_worker(0, "pool_workqueue_release"); |
7818 | BUG_ON(IS_ERR(pwq_release_worker)); | |
7819 | ||
aa6fde93 TH |
7820 | /* if the user set it to a specific value, keep it */ |
7821 | if (wq_cpu_intensive_thresh_us != ULONG_MAX) | |
7822 | return; | |
7823 | ||
7824 | /* | |
7825 | * The default of 10ms is derived from the fact that most modern (as of | |
7826 | * 2023) processors can do a lot in 10ms and that it's just below what | |
7827 | * most consider human-perceivable. However, the kernel also runs on a | |
7828 | * lot slower CPUs including microcontrollers where the threshold is way | |
7829 | * too low. | |
7830 | * | |
7831 | * Let's scale up the threshold upto 1 second if BogoMips is below 4000. | |
7832 | * This is by no means accurate but it doesn't have to be. The mechanism | |
7833 | * is still useful even when the threshold is fully scaled up. Also, as | |
7834 | * the reports would usually be applicable to everyone, some machines | |
7835 | * operating on longer thresholds won't significantly diminish their | |
7836 | * usefulness. | |
7837 | */ | |
7838 | thresh = 10 * USEC_PER_MSEC; | |
7839 | ||
7840 | /* see init/calibrate.c for lpj -> BogoMIPS calculation */ | |
7841 | bogo = max_t(unsigned long, loops_per_jiffy / 500000 * HZ, 1); | |
7842 | if (bogo < 4000) | |
7843 | thresh = min_t(unsigned long, thresh * 4000 / bogo, USEC_PER_SEC); | |
7844 | ||
7845 | pr_debug("wq_cpu_intensive_thresh: lpj=%lu BogoMIPS=%lu thresh_us=%lu\n", | |
7846 | loops_per_jiffy, bogo, thresh); | |
7847 | ||
7848 | wq_cpu_intensive_thresh_us = thresh; | |
7849 | } | |
7850 | ||
3347fa09 TH |
7851 | /** |
7852 | * workqueue_init - bring workqueue subsystem fully online | |
7853 | * | |
2930155b TH |
7854 | * This is the second step of three-staged workqueue subsystem initialization |
7855 | * and invoked as soon as kthreads can be created and scheduled. Workqueues have | |
7856 | * been created and work items queued on them, but there are no kworkers | |
7857 | * executing the work items yet. Populate the worker pools with the initial | |
7858 | * workers and enable future kworker creations. | |
3347fa09 | 7859 | */ |
2333e829 | 7860 | void __init workqueue_init(void) |
3347fa09 | 7861 | { |
2186d9f9 | 7862 | struct workqueue_struct *wq; |
3347fa09 TH |
7863 | struct worker_pool *pool; |
7864 | int cpu, bkt; | |
7865 | ||
aa6fde93 TH |
7866 | wq_cpu_intensive_thresh_init(); |
7867 | ||
2186d9f9 TH |
7868 | mutex_lock(&wq_pool_mutex); |
7869 | ||
2930155b TH |
7870 | /* |
7871 | * Per-cpu pools created earlier could be missing node hint. Fix them | |
7872 | * up. Also, create a rescuer for workqueues that requested it. | |
7873 | */ | |
2186d9f9 | 7874 | for_each_possible_cpu(cpu) { |
4cb1ef64 TH |
7875 | for_each_bh_worker_pool(pool, cpu) |
7876 | pool->node = cpu_to_node(cpu); | |
7877 | for_each_cpu_worker_pool(pool, cpu) | |
2186d9f9 | 7878 | pool->node = cpu_to_node(cpu); |
2186d9f9 TH |
7879 | } |
7880 | ||
40c17f75 | 7881 | list_for_each_entry(wq, &workqueues, list) { |
40c17f75 TH |
7882 | WARN(init_rescuer(wq), |
7883 | "workqueue: failed to create early rescuer for %s", | |
7884 | wq->name); | |
7885 | } | |
2186d9f9 TH |
7886 | |
7887 | mutex_unlock(&wq_pool_mutex); | |
7888 | ||
4cb1ef64 TH |
7889 | /* |
7890 | * Create the initial workers. A BH pool has one pseudo worker that | |
7891 | * represents the shared BH execution context and thus doesn't get | |
7892 | * affected by hotplug events. Create the BH pseudo workers for all | |
7893 | * possible CPUs here. | |
7894 | */ | |
7895 | for_each_possible_cpu(cpu) | |
7896 | for_each_bh_worker_pool(pool, cpu) | |
7897 | BUG_ON(!create_worker(pool)); | |
7898 | ||
3347fa09 TH |
7899 | for_each_online_cpu(cpu) { |
7900 | for_each_cpu_worker_pool(pool, cpu) { | |
7901 | pool->flags &= ~POOL_DISASSOCIATED; | |
7902 | BUG_ON(!create_worker(pool)); | |
7903 | } | |
7904 | } | |
7905 | ||
7906 | hash_for_each(unbound_pool_hash, bkt, pool, hash_node) | |
7907 | BUG_ON(!create_worker(pool)); | |
7908 | ||
7909 | wq_online = true; | |
82607adc | 7910 | wq_watchdog_init(); |
1da177e4 | 7911 | } |
c4f135d6 | 7912 | |
025e1684 TH |
7913 | /* |
7914 | * Initialize @pt by first initializing @pt->cpu_pod[] with pod IDs according to | |
7915 | * @cpu_shares_pod(). Each subset of CPUs that share a pod is assigned a unique | |
7916 | * and consecutive pod ID. The rest of @pt is initialized accordingly. | |
7917 | */ | |
7918 | static void __init init_pod_type(struct wq_pod_type *pt, | |
7919 | bool (*cpus_share_pod)(int, int)) | |
7920 | { | |
7921 | int cur, pre, cpu, pod; | |
7922 | ||
7923 | pt->nr_pods = 0; | |
7924 | ||
7925 | /* init @pt->cpu_pod[] according to @cpus_share_pod() */ | |
7926 | pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL); | |
7927 | BUG_ON(!pt->cpu_pod); | |
7928 | ||
7929 | for_each_possible_cpu(cur) { | |
7930 | for_each_possible_cpu(pre) { | |
7931 | if (pre >= cur) { | |
7932 | pt->cpu_pod[cur] = pt->nr_pods++; | |
7933 | break; | |
7934 | } | |
7935 | if (cpus_share_pod(cur, pre)) { | |
7936 | pt->cpu_pod[cur] = pt->cpu_pod[pre]; | |
7937 | break; | |
7938 | } | |
7939 | } | |
7940 | } | |
7941 | ||
7942 | /* init the rest to match @pt->cpu_pod[] */ | |
7943 | pt->pod_cpus = kcalloc(pt->nr_pods, sizeof(pt->pod_cpus[0]), GFP_KERNEL); | |
7944 | pt->pod_node = kcalloc(pt->nr_pods, sizeof(pt->pod_node[0]), GFP_KERNEL); | |
7945 | BUG_ON(!pt->pod_cpus || !pt->pod_node); | |
7946 | ||
7947 | for (pod = 0; pod < pt->nr_pods; pod++) | |
7948 | BUG_ON(!zalloc_cpumask_var(&pt->pod_cpus[pod], GFP_KERNEL)); | |
7949 | ||
7950 | for_each_possible_cpu(cpu) { | |
7951 | cpumask_set_cpu(cpu, pt->pod_cpus[pt->cpu_pod[cpu]]); | |
7952 | pt->pod_node[pt->cpu_pod[cpu]] = cpu_to_node(cpu); | |
7953 | } | |
7954 | } | |
7955 | ||
63c5484e TH |
7956 | static bool __init cpus_dont_share(int cpu0, int cpu1) |
7957 | { | |
7958 | return false; | |
7959 | } | |
7960 | ||
7961 | static bool __init cpus_share_smt(int cpu0, int cpu1) | |
7962 | { | |
7963 | #ifdef CONFIG_SCHED_SMT | |
7964 | return cpumask_test_cpu(cpu0, cpu_smt_mask(cpu1)); | |
7965 | #else | |
7966 | return false; | |
7967 | #endif | |
7968 | } | |
7969 | ||
025e1684 TH |
7970 | static bool __init cpus_share_numa(int cpu0, int cpu1) |
7971 | { | |
7972 | return cpu_to_node(cpu0) == cpu_to_node(cpu1); | |
7973 | } | |
7974 | ||
2930155b TH |
7975 | /** |
7976 | * workqueue_init_topology - initialize CPU pods for unbound workqueues | |
7977 | * | |
96068b60 | 7978 | * This is the third step of three-staged workqueue subsystem initialization and |
2930155b TH |
7979 | * invoked after SMP and topology information are fully initialized. It |
7980 | * initializes the unbound CPU pods accordingly. | |
7981 | */ | |
7982 | void __init workqueue_init_topology(void) | |
a86feae6 | 7983 | { |
2930155b | 7984 | struct workqueue_struct *wq; |
025e1684 | 7985 | int cpu; |
a86feae6 | 7986 | |
63c5484e TH |
7987 | init_pod_type(&wq_pod_types[WQ_AFFN_CPU], cpus_dont_share); |
7988 | init_pod_type(&wq_pod_types[WQ_AFFN_SMT], cpus_share_smt); | |
7989 | init_pod_type(&wq_pod_types[WQ_AFFN_CACHE], cpus_share_cache); | |
025e1684 | 7990 | init_pod_type(&wq_pod_types[WQ_AFFN_NUMA], cpus_share_numa); |
a86feae6 | 7991 | |
c5f8cd6c TH |
7992 | wq_topo_initialized = true; |
7993 | ||
2930155b | 7994 | mutex_lock(&wq_pool_mutex); |
a86feae6 | 7995 | |
2930155b TH |
7996 | /* |
7997 | * Workqueues allocated earlier would have all CPUs sharing the default | |
b2b1f933 LJ |
7998 | * worker pool. Explicitly call unbound_wq_update_pwq() on all workqueue |
7999 | * and CPU combinations to apply per-pod sharing. | |
2930155b TH |
8000 | */ |
8001 | list_for_each_entry(wq, &workqueues, list) { | |
5797b1c1 | 8002 | for_each_online_cpu(cpu) |
b2b1f933 | 8003 | unbound_wq_update_pwq(wq, cpu); |
5797b1c1 TH |
8004 | if (wq->flags & WQ_UNBOUND) { |
8005 | mutex_lock(&wq->mutex); | |
8006 | wq_update_node_max_active(wq, -1); | |
8007 | mutex_unlock(&wq->mutex); | |
2930155b TH |
8008 | } |
8009 | } | |
8010 | ||
8011 | mutex_unlock(&wq_pool_mutex); | |
a86feae6 TH |
8012 | } |
8013 | ||
20bdedaf TH |
8014 | void __warn_flushing_systemwide_wq(void) |
8015 | { | |
8016 | pr_warn("WARNING: Flushing system-wide workqueues will be prohibited in near future.\n"); | |
8017 | dump_stack(); | |
8018 | } | |
c4f135d6 | 8019 | EXPORT_SYMBOL(__warn_flushing_systemwide_wq); |
ace3c549 | 8020 | |
8021 | static int __init workqueue_unbound_cpus_setup(char *str) | |
8022 | { | |
8023 | if (cpulist_parse(str, &wq_cmdline_cpumask) < 0) { | |
8024 | cpumask_clear(&wq_cmdline_cpumask); | |
8025 | pr_warn("workqueue.unbound_cpus: incorrect CPU range, using default\n"); | |
8026 | } | |
8027 | ||
8028 | return 1; | |
8029 | } | |
8030 | __setup("workqueue.unbound_cpus=", workqueue_unbound_cpus_setup); |