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, |
c8e55f36 | 128 | }; |
1da177e4 | 129 | |
4cb1ef64 TH |
130 | /* |
131 | * We don't want to trap softirq for too long. See MAX_SOFTIRQ_TIME and | |
132 | * MAX_SOFTIRQ_RESTART in kernel/softirq.c. These are macros because | |
133 | * msecs_to_jiffies() can't be an initializer. | |
134 | */ | |
135 | #define BH_WORKER_JIFFIES msecs_to_jiffies(2) | |
136 | #define BH_WORKER_RESTARTS 10 | |
137 | ||
1da177e4 | 138 | /* |
4690c4ab TH |
139 | * Structure fields follow one of the following exclusion rules. |
140 | * | |
e41e704b TH |
141 | * I: Modifiable by initialization/destruction paths and read-only for |
142 | * everyone else. | |
4690c4ab | 143 | * |
e22bee78 TH |
144 | * P: Preemption protected. Disabling preemption is enough and should |
145 | * only be modified and accessed from the local cpu. | |
146 | * | |
d565ed63 | 147 | * L: pool->lock protected. Access with pool->lock held. |
4690c4ab | 148 | * |
5797b1c1 TH |
149 | * LN: pool->lock and wq_node_nr_active->lock protected for writes. Either for |
150 | * reads. | |
151 | * | |
bdf8b9bf TH |
152 | * K: Only modified by worker while holding pool->lock. Can be safely read by |
153 | * self, while holding pool->lock or from IRQ context if %current is the | |
154 | * kworker. | |
155 | * | |
156 | * S: Only modified by worker self. | |
157 | * | |
1258fae7 | 158 | * A: wq_pool_attach_mutex protected. |
822d8405 | 159 | * |
68e13a67 | 160 | * PL: wq_pool_mutex protected. |
5bcab335 | 161 | * |
24acfb71 | 162 | * PR: wq_pool_mutex protected for writes. RCU protected for reads. |
76af4d93 | 163 | * |
5b95e1af LJ |
164 | * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads. |
165 | * | |
166 | * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or | |
24acfb71 | 167 | * RCU for reads. |
5b95e1af | 168 | * |
3c25a55d LJ |
169 | * WQ: wq->mutex protected. |
170 | * | |
24acfb71 | 171 | * WR: wq->mutex protected for writes. RCU protected for reads. |
2e109a28 | 172 | * |
a045a272 TH |
173 | * WO: wq->mutex protected for writes. Updated with WRITE_ONCE() and can be read |
174 | * with READ_ONCE() without locking. | |
175 | * | |
2e109a28 | 176 | * MD: wq_mayday_lock protected. |
cd2440d6 PM |
177 | * |
178 | * WD: Used internally by the watchdog. | |
1da177e4 | 179 | */ |
1da177e4 | 180 | |
2eaebdb3 | 181 | /* struct worker is defined in workqueue_internal.h */ |
c34056a3 | 182 | |
bd7bdd43 | 183 | struct worker_pool { |
a9b8a985 | 184 | raw_spinlock_t lock; /* the pool lock */ |
d84ff051 | 185 | int cpu; /* I: the associated cpu */ |
f3f90ad4 | 186 | int node; /* I: the associated node ID */ |
9daf9e67 | 187 | int id; /* I: pool ID */ |
bc8b50c2 | 188 | unsigned int flags; /* L: flags */ |
bd7bdd43 | 189 | |
82607adc | 190 | unsigned long watchdog_ts; /* L: watchdog timestamp */ |
cd2440d6 | 191 | bool cpu_stall; /* WD: stalled cpu bound pool */ |
82607adc | 192 | |
bc35f7ef LJ |
193 | /* |
194 | * The counter is incremented in a process context on the associated CPU | |
195 | * w/ preemption disabled, and decremented or reset in the same context | |
196 | * but w/ pool->lock held. The readers grab pool->lock and are | |
197 | * guaranteed to see if the counter reached zero. | |
198 | */ | |
199 | int nr_running; | |
84f91c62 | 200 | |
bd7bdd43 | 201 | struct list_head worklist; /* L: list of pending works */ |
ea1abd61 | 202 | |
5826cc8f LJ |
203 | int nr_workers; /* L: total number of workers */ |
204 | int nr_idle; /* L: currently idle workers */ | |
bd7bdd43 | 205 | |
2c1f1a91 | 206 | struct list_head idle_list; /* L: list of idle workers */ |
bd7bdd43 | 207 | struct timer_list idle_timer; /* L: worker idle timeout */ |
3f959aa3 VS |
208 | struct work_struct idle_cull_work; /* L: worker idle cleanup */ |
209 | ||
210 | struct timer_list mayday_timer; /* L: SOS timer for workers */ | |
bd7bdd43 | 211 | |
c5aa87bb | 212 | /* a workers is either on busy_hash or idle_list, or the manager */ |
c9e7cf27 TH |
213 | DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER); |
214 | /* L: hash of busy workers */ | |
215 | ||
2607d7a6 | 216 | struct worker *manager; /* L: purely informational */ |
92f9c5c4 | 217 | struct list_head workers; /* A: attached workers */ |
e02b9312 | 218 | struct list_head dying_workers; /* A: workers about to die */ |
60f5a4bc | 219 | struct completion *detach_completion; /* all workers detached */ |
e19e397a | 220 | |
7cda9aae | 221 | struct ida worker_ida; /* worker IDs for task name */ |
e19e397a | 222 | |
7a4e344c | 223 | struct workqueue_attrs *attrs; /* I: worker attributes */ |
68e13a67 LJ |
224 | struct hlist_node hash_node; /* PL: unbound_pool_hash node */ |
225 | int refcnt; /* PL: refcnt for unbound pools */ | |
7a4e344c | 226 | |
29c91e99 | 227 | /* |
24acfb71 | 228 | * Destruction of pool is RCU protected to allow dereferences |
29c91e99 TH |
229 | * from get_work_pool(). |
230 | */ | |
231 | struct rcu_head rcu; | |
84f91c62 | 232 | }; |
8b03ae3c | 233 | |
725e8ec5 TH |
234 | /* |
235 | * Per-pool_workqueue statistics. These can be monitored using | |
236 | * tools/workqueue/wq_monitor.py. | |
237 | */ | |
238 | enum pool_workqueue_stats { | |
239 | PWQ_STAT_STARTED, /* work items started execution */ | |
240 | PWQ_STAT_COMPLETED, /* work items completed execution */ | |
8a1dd1e5 | 241 | PWQ_STAT_CPU_TIME, /* total CPU time consumed */ |
616db877 | 242 | PWQ_STAT_CPU_INTENSIVE, /* wq_cpu_intensive_thresh_us violations */ |
725e8ec5 | 243 | PWQ_STAT_CM_WAKEUP, /* concurrency-management worker wakeups */ |
8639eceb | 244 | PWQ_STAT_REPATRIATED, /* unbound workers brought back into scope */ |
725e8ec5 TH |
245 | PWQ_STAT_MAYDAY, /* maydays to rescuer */ |
246 | PWQ_STAT_RESCUED, /* linked work items executed by rescuer */ | |
247 | ||
248 | PWQ_NR_STATS, | |
249 | }; | |
250 | ||
1da177e4 | 251 | /* |
e9a8e01f | 252 | * The per-pool workqueue. While queued, bits below WORK_PWQ_SHIFT |
112202d9 TH |
253 | * of work_struct->data are used for flags and the remaining high bits |
254 | * point to the pwq; thus, pwqs need to be aligned at two's power of the | |
255 | * number of flag bits. | |
1da177e4 | 256 | */ |
112202d9 | 257 | struct pool_workqueue { |
bd7bdd43 | 258 | struct worker_pool *pool; /* I: the associated pool */ |
4690c4ab | 259 | struct workqueue_struct *wq; /* I: the owning workqueue */ |
73f53c4a TH |
260 | int work_color; /* L: current color */ |
261 | int flush_color; /* L: flushing color */ | |
8864b4e5 | 262 | int refcnt; /* L: reference count */ |
73f53c4a TH |
263 | int nr_in_flight[WORK_NR_COLORS]; |
264 | /* L: nr of in_flight works */ | |
4c065dbc | 265 | bool plugged; /* L: execution suspended */ |
018f3a13 LJ |
266 | |
267 | /* | |
268 | * nr_active management and WORK_STRUCT_INACTIVE: | |
269 | * | |
270 | * When pwq->nr_active >= max_active, new work item is queued to | |
271 | * pwq->inactive_works instead of pool->worklist and marked with | |
272 | * WORK_STRUCT_INACTIVE. | |
273 | * | |
5797b1c1 TH |
274 | * All work items marked with WORK_STRUCT_INACTIVE do not participate in |
275 | * nr_active and all work items in pwq->inactive_works are marked with | |
276 | * WORK_STRUCT_INACTIVE. But not all WORK_STRUCT_INACTIVE work items are | |
277 | * in pwq->inactive_works. Some of them are ready to run in | |
278 | * pool->worklist or worker->scheduled. Those work itmes are only struct | |
279 | * wq_barrier which is used for flush_work() and should not participate | |
280 | * in nr_active. For non-barrier work item, it is marked with | |
281 | * WORK_STRUCT_INACTIVE iff it is in pwq->inactive_works. | |
018f3a13 | 282 | */ |
1e19ffc6 | 283 | int nr_active; /* L: nr of active works */ |
f97a4a1a | 284 | struct list_head inactive_works; /* L: inactive works */ |
5797b1c1 | 285 | struct list_head pending_node; /* LN: node on wq_node_nr_active->pending_pwqs */ |
3c25a55d | 286 | struct list_head pwqs_node; /* WR: node on wq->pwqs */ |
2e109a28 | 287 | struct list_head mayday_node; /* MD: node on wq->maydays */ |
8864b4e5 | 288 | |
725e8ec5 TH |
289 | u64 stats[PWQ_NR_STATS]; |
290 | ||
8864b4e5 | 291 | /* |
967b494e | 292 | * Release of unbound pwq is punted to a kthread_worker. See put_pwq() |
687a9aa5 TH |
293 | * and pwq_release_workfn() for details. pool_workqueue itself is also |
294 | * RCU protected so that the first pwq can be determined without | |
967b494e | 295 | * grabbing wq->mutex. |
8864b4e5 | 296 | */ |
687a9aa5 | 297 | struct kthread_work release_work; |
8864b4e5 | 298 | struct rcu_head rcu; |
e9a8e01f | 299 | } __aligned(1 << WORK_STRUCT_PWQ_SHIFT); |
1da177e4 | 300 | |
73f53c4a TH |
301 | /* |
302 | * Structure used to wait for workqueue flush. | |
303 | */ | |
304 | struct wq_flusher { | |
3c25a55d LJ |
305 | struct list_head list; /* WQ: list of flushers */ |
306 | int flush_color; /* WQ: flush color waiting for */ | |
73f53c4a TH |
307 | struct completion done; /* flush completion */ |
308 | }; | |
309 | ||
226223ab TH |
310 | struct wq_device; |
311 | ||
91ccc6e7 TH |
312 | /* |
313 | * Unlike in a per-cpu workqueue where max_active limits its concurrency level | |
314 | * on each CPU, in an unbound workqueue, max_active applies to the whole system. | |
315 | * As sharing a single nr_active across multiple sockets can be very expensive, | |
316 | * the counting and enforcement is per NUMA node. | |
5797b1c1 TH |
317 | * |
318 | * The following struct is used to enforce per-node max_active. When a pwq wants | |
319 | * to start executing a work item, it should increment ->nr using | |
320 | * tryinc_node_nr_active(). If acquisition fails due to ->nr already being over | |
321 | * ->max, the pwq is queued on ->pending_pwqs. As in-flight work items finish | |
322 | * and decrement ->nr, node_activate_pending_pwq() activates the pending pwqs in | |
323 | * round-robin order. | |
91ccc6e7 TH |
324 | */ |
325 | struct wq_node_nr_active { | |
5797b1c1 TH |
326 | int max; /* per-node max_active */ |
327 | atomic_t nr; /* per-node nr_active */ | |
328 | raw_spinlock_t lock; /* nests inside pool locks */ | |
329 | struct list_head pending_pwqs; /* LN: pwqs with inactive works */ | |
91ccc6e7 TH |
330 | }; |
331 | ||
1da177e4 | 332 | /* |
c5aa87bb TH |
333 | * The externally visible workqueue. It relays the issued work items to |
334 | * the appropriate worker_pool through its pool_workqueues. | |
1da177e4 LT |
335 | */ |
336 | struct workqueue_struct { | |
3c25a55d | 337 | struct list_head pwqs; /* WR: all pwqs of this wq */ |
e2dca7ad | 338 | struct list_head list; /* PR: list of all workqueues */ |
73f53c4a | 339 | |
3c25a55d LJ |
340 | struct mutex mutex; /* protects this wq */ |
341 | int work_color; /* WQ: current work color */ | |
342 | int flush_color; /* WQ: current flush color */ | |
112202d9 | 343 | atomic_t nr_pwqs_to_flush; /* flush in progress */ |
3c25a55d LJ |
344 | struct wq_flusher *first_flusher; /* WQ: first flusher */ |
345 | struct list_head flusher_queue; /* WQ: flush waiters */ | |
346 | struct list_head flusher_overflow; /* WQ: flush overflow list */ | |
73f53c4a | 347 | |
2e109a28 | 348 | struct list_head maydays; /* MD: pwqs requesting rescue */ |
30ae2fc0 | 349 | struct worker *rescuer; /* MD: rescue worker */ |
e22bee78 | 350 | |
87fc741e | 351 | int nr_drainers; /* WQ: drain in progress */ |
5797b1c1 TH |
352 | |
353 | /* See alloc_workqueue() function comment for info on min/max_active */ | |
a045a272 | 354 | int max_active; /* WO: max active works */ |
5797b1c1 | 355 | int min_active; /* WO: min active works */ |
a045a272 | 356 | int saved_max_active; /* WQ: saved max_active */ |
5797b1c1 | 357 | int saved_min_active; /* WQ: saved min_active */ |
226223ab | 358 | |
5b95e1af | 359 | struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */ |
9f66cff2 | 360 | struct pool_workqueue __rcu *dfl_pwq; /* PW: only for unbound wqs */ |
6029a918 | 361 | |
226223ab TH |
362 | #ifdef CONFIG_SYSFS |
363 | struct wq_device *wq_dev; /* I: for sysfs interface */ | |
364 | #endif | |
4e6045f1 | 365 | #ifdef CONFIG_LOCKDEP |
669de8bd BVA |
366 | char *lock_name; |
367 | struct lock_class_key key; | |
4690c4ab | 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 */ | |
636b927e | 381 | struct pool_workqueue __percpu __rcu **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 TH |
437 | /* buf for wq_update_unbound_pod_attrs(), protected by CPU hotplug exclusion */ |
438 | static struct workqueue_attrs *wq_update_pod_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 | |
99c621ef | 449 | /* PL&A: allowable cpus for unbound wqs and work items */ |
ef557180 MG |
450 | static cpumask_var_t wq_unbound_cpumask; |
451 | ||
fe28f631 WL |
452 | /* PL: user requested unbound cpumask via sysfs */ |
453 | static cpumask_var_t wq_requested_unbound_cpumask; | |
454 | ||
455 | /* PL: isolated cpumask to be excluded from unbound cpumask */ | |
456 | static cpumask_var_t wq_isolated_cpumask; | |
457 | ||
ace3c549 | 458 | /* for further constrain wq_unbound_cpumask by cmdline parameter*/ |
459 | static struct cpumask wq_cmdline_cpumask __initdata; | |
460 | ||
ef557180 MG |
461 | /* CPU where unbound work was last round robin scheduled from this CPU */ |
462 | static DEFINE_PER_CPU(int, wq_rr_cpu_last); | |
b05a7928 | 463 | |
f303fccb TH |
464 | /* |
465 | * Local execution of unbound work items is no longer guaranteed. The | |
466 | * following always forces round-robin CPU selection on unbound work items | |
467 | * to uncover usages which depend on it. | |
468 | */ | |
469 | #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU | |
470 | static bool wq_debug_force_rr_cpu = true; | |
471 | #else | |
472 | static bool wq_debug_force_rr_cpu = false; | |
473 | #endif | |
474 | module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644); | |
475 | ||
2f34d733 TH |
476 | /* to raise softirq for the BH worker pools on other CPUs */ |
477 | static DEFINE_PER_CPU_SHARED_ALIGNED(struct irq_work [NR_STD_WORKER_POOLS], | |
478 | bh_pool_irq_works); | |
479 | ||
4cb1ef64 TH |
480 | /* the BH worker pools */ |
481 | static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], | |
482 | bh_worker_pools); | |
483 | ||
7d19c5ce | 484 | /* the per-cpu worker pools */ |
4cb1ef64 TH |
485 | static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], |
486 | 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 | { |
1211f3b2 TH |
898 | return (v >> shift) & ((1 << bits) - 1); |
899 | } | |
54d5b7d0 | 900 | |
1211f3b2 TH |
901 | static void work_offqd_unpack(struct work_offq_data *offqd, unsigned long data) |
902 | { | |
903 | WARN_ON_ONCE(data & WORK_STRUCT_PWQ); | |
7c3eed5c | 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; |
7c3eed5c 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); | |
1271 | p->wake_cpu = cpumask_any_distribute(pool->attrs->__pod_cpumask); | |
1272 | get_work_pwq(work)->stats[PWQ_STAT_REPATRIATED]++; | |
1273 | } | |
1274 | #endif | |
1275 | wake_up_process(p); | |
0219a352 | 1276 | return true; |
797e8345 TH |
1277 | } |
1278 | ||
63638450 TH |
1279 | #ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT |
1280 | ||
1281 | /* | |
1282 | * Concurrency-managed per-cpu work items that hog CPU for longer than | |
1283 | * wq_cpu_intensive_thresh_us trigger the automatic CPU_INTENSIVE mechanism, | |
1284 | * which prevents them from stalling other concurrency-managed work items. If a | |
1285 | * work function keeps triggering this mechanism, it's likely that the work item | |
1286 | * should be using an unbound workqueue instead. | |
1287 | * | |
1288 | * wq_cpu_intensive_report() tracks work functions which trigger such conditions | |
1289 | * and report them so that they can be examined and converted to use unbound | |
1290 | * workqueues as appropriate. To avoid flooding the console, each violating work | |
1291 | * function is tracked and reported with exponential backoff. | |
1292 | */ | |
1293 | #define WCI_MAX_ENTS 128 | |
1294 | ||
1295 | struct wci_ent { | |
1296 | work_func_t func; | |
1297 | atomic64_t cnt; | |
1298 | struct hlist_node hash_node; | |
1299 | }; | |
1300 | ||
1301 | static struct wci_ent wci_ents[WCI_MAX_ENTS]; | |
1302 | static int wci_nr_ents; | |
1303 | static DEFINE_RAW_SPINLOCK(wci_lock); | |
1304 | static DEFINE_HASHTABLE(wci_hash, ilog2(WCI_MAX_ENTS)); | |
1305 | ||
1306 | static struct wci_ent *wci_find_ent(work_func_t func) | |
1307 | { | |
1308 | struct wci_ent *ent; | |
1309 | ||
1310 | hash_for_each_possible_rcu(wci_hash, ent, hash_node, | |
1311 | (unsigned long)func) { | |
1312 | if (ent->func == func) | |
1313 | return ent; | |
1314 | } | |
1315 | return NULL; | |
1316 | } | |
1317 | ||
1318 | static void wq_cpu_intensive_report(work_func_t func) | |
1319 | { | |
1320 | struct wci_ent *ent; | |
1321 | ||
1322 | restart: | |
1323 | ent = wci_find_ent(func); | |
1324 | if (ent) { | |
1325 | u64 cnt; | |
1326 | ||
1327 | /* | |
ccdec921 | 1328 | * Start reporting from the warning_thresh and back off |
63638450 TH |
1329 | * exponentially. |
1330 | */ | |
1331 | cnt = atomic64_inc_return_relaxed(&ent->cnt); | |
ccdec921 XY |
1332 | if (wq_cpu_intensive_warning_thresh && |
1333 | cnt >= wq_cpu_intensive_warning_thresh && | |
1334 | is_power_of_2(cnt + 1 - wq_cpu_intensive_warning_thresh)) | |
63638450 TH |
1335 | printk_deferred(KERN_WARNING "workqueue: %ps hogged CPU for >%luus %llu times, consider switching to WQ_UNBOUND\n", |
1336 | ent->func, wq_cpu_intensive_thresh_us, | |
1337 | atomic64_read(&ent->cnt)); | |
1338 | return; | |
1339 | } | |
1340 | ||
1341 | /* | |
1342 | * @func is a new violation. Allocate a new entry for it. If wcn_ents[] | |
1343 | * is exhausted, something went really wrong and we probably made enough | |
1344 | * noise already. | |
1345 | */ | |
1346 | if (wci_nr_ents >= WCI_MAX_ENTS) | |
1347 | return; | |
1348 | ||
1349 | raw_spin_lock(&wci_lock); | |
1350 | ||
1351 | if (wci_nr_ents >= WCI_MAX_ENTS) { | |
1352 | raw_spin_unlock(&wci_lock); | |
1353 | return; | |
1354 | } | |
1355 | ||
1356 | if (wci_find_ent(func)) { | |
1357 | raw_spin_unlock(&wci_lock); | |
1358 | goto restart; | |
1359 | } | |
1360 | ||
1361 | ent = &wci_ents[wci_nr_ents++]; | |
1362 | ent->func = func; | |
ccdec921 | 1363 | atomic64_set(&ent->cnt, 0); |
63638450 TH |
1364 | hash_add_rcu(wci_hash, &ent->hash_node, (unsigned long)func); |
1365 | ||
1366 | raw_spin_unlock(&wci_lock); | |
ccdec921 XY |
1367 | |
1368 | goto restart; | |
63638450 TH |
1369 | } |
1370 | ||
1371 | #else /* CONFIG_WQ_CPU_INTENSIVE_REPORT */ | |
1372 | static void wq_cpu_intensive_report(work_func_t func) {} | |
1373 | #endif /* CONFIG_WQ_CPU_INTENSIVE_REPORT */ | |
1374 | ||
d302f017 | 1375 | /** |
6d25be57 | 1376 | * wq_worker_running - a worker is running again |
e22bee78 | 1377 | * @task: task waking up |
e22bee78 | 1378 | * |
6d25be57 | 1379 | * This function is called when a worker returns from schedule() |
e22bee78 | 1380 | */ |
6d25be57 | 1381 | void wq_worker_running(struct task_struct *task) |
e22bee78 TH |
1382 | { |
1383 | struct worker *worker = kthread_data(task); | |
1384 | ||
c8f6219b | 1385 | if (!READ_ONCE(worker->sleeping)) |
6d25be57 | 1386 | return; |
07edfece FW |
1387 | |
1388 | /* | |
1389 | * If preempted by unbind_workers() between the WORKER_NOT_RUNNING check | |
1390 | * and the nr_running increment below, we may ruin the nr_running reset | |
1391 | * and leave with an unexpected pool->nr_running == 1 on the newly unbound | |
1392 | * pool. Protect against such race. | |
1393 | */ | |
1394 | preempt_disable(); | |
6d25be57 | 1395 | if (!(worker->flags & WORKER_NOT_RUNNING)) |
bc35f7ef | 1396 | worker->pool->nr_running++; |
07edfece | 1397 | preempt_enable(); |
616db877 TH |
1398 | |
1399 | /* | |
1400 | * CPU intensive auto-detection cares about how long a work item hogged | |
1401 | * CPU without sleeping. Reset the starting timestamp on wakeup. | |
1402 | */ | |
1403 | worker->current_at = worker->task->se.sum_exec_runtime; | |
1404 | ||
c8f6219b | 1405 | WRITE_ONCE(worker->sleeping, 0); |
e22bee78 TH |
1406 | } |
1407 | ||
1408 | /** | |
1409 | * wq_worker_sleeping - a worker is going to sleep | |
1410 | * @task: task going to sleep | |
e22bee78 | 1411 | * |
6d25be57 | 1412 | * This function is called from schedule() when a busy worker is |
ccf45156 | 1413 | * going to sleep. |
e22bee78 | 1414 | */ |
6d25be57 | 1415 | void wq_worker_sleeping(struct task_struct *task) |
e22bee78 | 1416 | { |
cc5bff38 | 1417 | struct worker *worker = kthread_data(task); |
111c225a | 1418 | struct worker_pool *pool; |
e22bee78 | 1419 | |
111c225a TH |
1420 | /* |
1421 | * Rescuers, which may not have all the fields set up like normal | |
1422 | * workers, also reach here, let's not access anything before | |
1423 | * checking NOT_RUNNING. | |
1424 | */ | |
2d64672e | 1425 | if (worker->flags & WORKER_NOT_RUNNING) |
6d25be57 | 1426 | return; |
e22bee78 | 1427 | |
111c225a | 1428 | pool = worker->pool; |
111c225a | 1429 | |
62849a96 | 1430 | /* Return if preempted before wq_worker_running() was reached */ |
c8f6219b | 1431 | if (READ_ONCE(worker->sleeping)) |
6d25be57 TG |
1432 | return; |
1433 | ||
c8f6219b | 1434 | WRITE_ONCE(worker->sleeping, 1); |
a9b8a985 | 1435 | raw_spin_lock_irq(&pool->lock); |
e22bee78 | 1436 | |
45c753f5 FW |
1437 | /* |
1438 | * Recheck in case unbind_workers() preempted us. We don't | |
1439 | * want to decrement nr_running after the worker is unbound | |
1440 | * and nr_running has been reset. | |
1441 | */ | |
1442 | if (worker->flags & WORKER_NOT_RUNNING) { | |
1443 | raw_spin_unlock_irq(&pool->lock); | |
1444 | return; | |
1445 | } | |
1446 | ||
bc35f7ef | 1447 | pool->nr_running--; |
0219a352 | 1448 | if (kick_pool(pool)) |
725e8ec5 | 1449 | worker->current_pwq->stats[PWQ_STAT_CM_WAKEUP]++; |
0219a352 | 1450 | |
a9b8a985 | 1451 | raw_spin_unlock_irq(&pool->lock); |
e22bee78 TH |
1452 | } |
1453 | ||
616db877 TH |
1454 | /** |
1455 | * wq_worker_tick - a scheduler tick occurred while a kworker is running | |
1456 | * @task: task currently running | |
1457 | * | |
1458 | * Called from scheduler_tick(). We're in the IRQ context and the current | |
1459 | * worker's fields which follow the 'K' locking rule can be accessed safely. | |
1460 | */ | |
1461 | void wq_worker_tick(struct task_struct *task) | |
1462 | { | |
1463 | struct worker *worker = kthread_data(task); | |
1464 | struct pool_workqueue *pwq = worker->current_pwq; | |
1465 | struct worker_pool *pool = worker->pool; | |
1466 | ||
1467 | if (!pwq) | |
1468 | return; | |
1469 | ||
8a1dd1e5 TH |
1470 | pwq->stats[PWQ_STAT_CPU_TIME] += TICK_USEC; |
1471 | ||
18c8ae81 Z |
1472 | if (!wq_cpu_intensive_thresh_us) |
1473 | return; | |
1474 | ||
616db877 TH |
1475 | /* |
1476 | * If the current worker is concurrency managed and hogged the CPU for | |
1477 | * longer than wq_cpu_intensive_thresh_us, it's automatically marked | |
1478 | * CPU_INTENSIVE to avoid stalling other concurrency-managed work items. | |
c8f6219b Z |
1479 | * |
1480 | * Set @worker->sleeping means that @worker is in the process of | |
1481 | * switching out voluntarily and won't be contributing to | |
1482 | * @pool->nr_running until it wakes up. As wq_worker_sleeping() also | |
1483 | * decrements ->nr_running, setting CPU_INTENSIVE here can lead to | |
1484 | * double decrements. The task is releasing the CPU anyway. Let's skip. | |
1485 | * We probably want to make this prettier in the future. | |
616db877 | 1486 | */ |
c8f6219b | 1487 | if ((worker->flags & WORKER_NOT_RUNNING) || READ_ONCE(worker->sleeping) || |
616db877 TH |
1488 | worker->task->se.sum_exec_runtime - worker->current_at < |
1489 | wq_cpu_intensive_thresh_us * NSEC_PER_USEC) | |
1490 | return; | |
1491 | ||
1492 | raw_spin_lock(&pool->lock); | |
1493 | ||
1494 | worker_set_flags(worker, WORKER_CPU_INTENSIVE); | |
63638450 | 1495 | wq_cpu_intensive_report(worker->current_func); |
616db877 TH |
1496 | pwq->stats[PWQ_STAT_CPU_INTENSIVE]++; |
1497 | ||
0219a352 | 1498 | if (kick_pool(pool)) |
616db877 | 1499 | pwq->stats[PWQ_STAT_CM_WAKEUP]++; |
616db877 TH |
1500 | |
1501 | raw_spin_unlock(&pool->lock); | |
1502 | } | |
1503 | ||
1b69ac6b JW |
1504 | /** |
1505 | * wq_worker_last_func - retrieve worker's last work function | |
8194fe94 | 1506 | * @task: Task to retrieve last work function of. |
1b69ac6b JW |
1507 | * |
1508 | * Determine the last function a worker executed. This is called from | |
1509 | * the scheduler to get a worker's last known identity. | |
1510 | * | |
1511 | * CONTEXT: | |
a9b8a985 | 1512 | * raw_spin_lock_irq(rq->lock) |
1b69ac6b | 1513 | * |
4b047002 JW |
1514 | * This function is called during schedule() when a kworker is going |
1515 | * to sleep. It's used by psi to identify aggregation workers during | |
1516 | * dequeuing, to allow periodic aggregation to shut-off when that | |
1517 | * worker is the last task in the system or cgroup to go to sleep. | |
1518 | * | |
1519 | * As this function doesn't involve any workqueue-related locking, it | |
1520 | * only returns stable values when called from inside the scheduler's | |
1521 | * queuing and dequeuing paths, when @task, which must be a kworker, | |
1522 | * is guaranteed to not be processing any works. | |
1523 | * | |
1b69ac6b JW |
1524 | * Return: |
1525 | * The last work function %current executed as a worker, NULL if it | |
1526 | * hasn't executed any work yet. | |
1527 | */ | |
1528 | work_func_t wq_worker_last_func(struct task_struct *task) | |
1529 | { | |
1530 | struct worker *worker = kthread_data(task); | |
1531 | ||
1532 | return worker->last_func; | |
1533 | } | |
1534 | ||
91ccc6e7 TH |
1535 | /** |
1536 | * wq_node_nr_active - Determine wq_node_nr_active to use | |
1537 | * @wq: workqueue of interest | |
1538 | * @node: NUMA node, can be %NUMA_NO_NODE | |
1539 | * | |
1540 | * Determine wq_node_nr_active to use for @wq on @node. Returns: | |
1541 | * | |
1542 | * - %NULL for per-cpu workqueues as they don't need to use shared nr_active. | |
1543 | * | |
1544 | * - node_nr_active[nr_node_ids] if @node is %NUMA_NO_NODE. | |
1545 | * | |
1546 | * - Otherwise, node_nr_active[@node]. | |
1547 | */ | |
1548 | static struct wq_node_nr_active *wq_node_nr_active(struct workqueue_struct *wq, | |
1549 | int node) | |
1550 | { | |
1551 | if (!(wq->flags & WQ_UNBOUND)) | |
1552 | return NULL; | |
1553 | ||
1554 | if (node == NUMA_NO_NODE) | |
1555 | node = nr_node_ids; | |
1556 | ||
1557 | return wq->node_nr_active[node]; | |
1558 | } | |
1559 | ||
5797b1c1 TH |
1560 | /** |
1561 | * wq_update_node_max_active - Update per-node max_actives to use | |
1562 | * @wq: workqueue to update | |
1563 | * @off_cpu: CPU that's going down, -1 if a CPU is not going down | |
1564 | * | |
1565 | * Update @wq->node_nr_active[]->max. @wq must be unbound. max_active is | |
1566 | * distributed among nodes according to the proportions of numbers of online | |
1567 | * cpus. The result is always between @wq->min_active and max_active. | |
1568 | */ | |
1569 | static void wq_update_node_max_active(struct workqueue_struct *wq, int off_cpu) | |
1570 | { | |
1571 | struct cpumask *effective = unbound_effective_cpumask(wq); | |
1572 | int min_active = READ_ONCE(wq->min_active); | |
1573 | int max_active = READ_ONCE(wq->max_active); | |
1574 | int total_cpus, node; | |
1575 | ||
1576 | lockdep_assert_held(&wq->mutex); | |
1577 | ||
c5f8cd6c TH |
1578 | if (!wq_topo_initialized) |
1579 | return; | |
1580 | ||
15930da4 | 1581 | if (off_cpu >= 0 && !cpumask_test_cpu(off_cpu, effective)) |
5797b1c1 TH |
1582 | off_cpu = -1; |
1583 | ||
1584 | total_cpus = cpumask_weight_and(effective, cpu_online_mask); | |
1585 | if (off_cpu >= 0) | |
1586 | total_cpus--; | |
1587 | ||
1588 | for_each_node(node) { | |
1589 | int node_cpus; | |
1590 | ||
1591 | node_cpus = cpumask_weight_and(effective, cpumask_of_node(node)); | |
1592 | if (off_cpu >= 0 && cpu_to_node(off_cpu) == node) | |
1593 | node_cpus--; | |
1594 | ||
1595 | wq_node_nr_active(wq, node)->max = | |
1596 | clamp(DIV_ROUND_UP(max_active * node_cpus, total_cpus), | |
1597 | min_active, max_active); | |
1598 | } | |
1599 | ||
1600 | wq_node_nr_active(wq, NUMA_NO_NODE)->max = min_active; | |
1601 | } | |
1602 | ||
8864b4e5 TH |
1603 | /** |
1604 | * get_pwq - get an extra reference on the specified pool_workqueue | |
1605 | * @pwq: pool_workqueue to get | |
1606 | * | |
1607 | * Obtain an extra reference on @pwq. The caller should guarantee that | |
1608 | * @pwq has positive refcnt and be holding the matching pool->lock. | |
1609 | */ | |
1610 | static void get_pwq(struct pool_workqueue *pwq) | |
1611 | { | |
1612 | lockdep_assert_held(&pwq->pool->lock); | |
1613 | WARN_ON_ONCE(pwq->refcnt <= 0); | |
1614 | pwq->refcnt++; | |
1615 | } | |
1616 | ||
1617 | /** | |
1618 | * put_pwq - put a pool_workqueue reference | |
1619 | * @pwq: pool_workqueue to put | |
1620 | * | |
1621 | * Drop a reference of @pwq. If its refcnt reaches zero, schedule its | |
1622 | * destruction. The caller should be holding the matching pool->lock. | |
1623 | */ | |
1624 | static void put_pwq(struct pool_workqueue *pwq) | |
1625 | { | |
1626 | lockdep_assert_held(&pwq->pool->lock); | |
1627 | if (likely(--pwq->refcnt)) | |
1628 | return; | |
8864b4e5 | 1629 | /* |
967b494e TH |
1630 | * @pwq can't be released under pool->lock, bounce to a dedicated |
1631 | * kthread_worker to avoid A-A deadlocks. | |
8864b4e5 | 1632 | */ |
687a9aa5 | 1633 | kthread_queue_work(pwq_release_worker, &pwq->release_work); |
8864b4e5 TH |
1634 | } |
1635 | ||
dce90d47 TH |
1636 | /** |
1637 | * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock | |
1638 | * @pwq: pool_workqueue to put (can be %NULL) | |
1639 | * | |
1640 | * put_pwq() with locking. This function also allows %NULL @pwq. | |
1641 | */ | |
1642 | static void put_pwq_unlocked(struct pool_workqueue *pwq) | |
1643 | { | |
1644 | if (pwq) { | |
1645 | /* | |
24acfb71 | 1646 | * As both pwqs and pools are RCU protected, the |
dce90d47 TH |
1647 | * following lock operations are safe. |
1648 | */ | |
a9b8a985 | 1649 | raw_spin_lock_irq(&pwq->pool->lock); |
dce90d47 | 1650 | put_pwq(pwq); |
a9b8a985 | 1651 | raw_spin_unlock_irq(&pwq->pool->lock); |
dce90d47 TH |
1652 | } |
1653 | } | |
1654 | ||
afa87ce8 TH |
1655 | static bool pwq_is_empty(struct pool_workqueue *pwq) |
1656 | { | |
1657 | return !pwq->nr_active && list_empty(&pwq->inactive_works); | |
1658 | } | |
1659 | ||
4c638030 TH |
1660 | static void __pwq_activate_work(struct pool_workqueue *pwq, |
1661 | struct work_struct *work) | |
bf4ede01 | 1662 | { |
1c270b79 TH |
1663 | unsigned long *wdb = work_data_bits(work); |
1664 | ||
1665 | WARN_ON_ONCE(!(*wdb & WORK_STRUCT_INACTIVE)); | |
bf4ede01 | 1666 | trace_workqueue_activate_work(work); |
82607adc TH |
1667 | if (list_empty(&pwq->pool->worklist)) |
1668 | pwq->pool->watchdog_ts = jiffies; | |
112202d9 | 1669 | move_linked_works(work, &pwq->pool->worklist, NULL); |
1c270b79 | 1670 | __clear_bit(WORK_STRUCT_INACTIVE_BIT, wdb); |
4c638030 TH |
1671 | } |
1672 | ||
1673 | /** | |
1674 | * pwq_activate_work - Activate a work item if inactive | |
1675 | * @pwq: pool_workqueue @work belongs to | |
1676 | * @work: work item to activate | |
1677 | * | |
1678 | * Returns %true if activated. %false if already active. | |
1679 | */ | |
1680 | static bool pwq_activate_work(struct pool_workqueue *pwq, | |
1681 | struct work_struct *work) | |
1682 | { | |
1683 | struct worker_pool *pool = pwq->pool; | |
91ccc6e7 | 1684 | struct wq_node_nr_active *nna; |
4c638030 TH |
1685 | |
1686 | lockdep_assert_held(&pool->lock); | |
1687 | ||
1688 | if (!(*work_data_bits(work) & WORK_STRUCT_INACTIVE)) | |
1689 | return false; | |
1690 | ||
91ccc6e7 TH |
1691 | nna = wq_node_nr_active(pwq->wq, pool->node); |
1692 | if (nna) | |
1693 | atomic_inc(&nna->nr); | |
1694 | ||
112202d9 | 1695 | pwq->nr_active++; |
4c638030 TH |
1696 | __pwq_activate_work(pwq, work); |
1697 | return true; | |
bf4ede01 TH |
1698 | } |
1699 | ||
5797b1c1 TH |
1700 | static bool tryinc_node_nr_active(struct wq_node_nr_active *nna) |
1701 | { | |
1702 | int max = READ_ONCE(nna->max); | |
1703 | ||
1704 | while (true) { | |
1705 | int old, tmp; | |
1706 | ||
1707 | old = atomic_read(&nna->nr); | |
1708 | if (old >= max) | |
1709 | return false; | |
1710 | tmp = atomic_cmpxchg_relaxed(&nna->nr, old, old + 1); | |
1711 | if (tmp == old) | |
1712 | return true; | |
1713 | } | |
1714 | } | |
1715 | ||
1c270b79 TH |
1716 | /** |
1717 | * pwq_tryinc_nr_active - Try to increment nr_active for a pwq | |
1718 | * @pwq: pool_workqueue of interest | |
5797b1c1 | 1719 | * @fill: max_active may have increased, try to increase concurrency level |
1c270b79 TH |
1720 | * |
1721 | * Try to increment nr_active for @pwq. Returns %true if an nr_active count is | |
1722 | * successfully obtained. %false otherwise. | |
1723 | */ | |
5797b1c1 | 1724 | static bool pwq_tryinc_nr_active(struct pool_workqueue *pwq, bool fill) |
1c270b79 TH |
1725 | { |
1726 | struct workqueue_struct *wq = pwq->wq; | |
1727 | struct worker_pool *pool = pwq->pool; | |
91ccc6e7 | 1728 | struct wq_node_nr_active *nna = wq_node_nr_active(wq, pool->node); |
5797b1c1 | 1729 | bool obtained = false; |
1c270b79 TH |
1730 | |
1731 | lockdep_assert_held(&pool->lock); | |
1732 | ||
5797b1c1 | 1733 | if (!nna) { |
4cb1ef64 | 1734 | /* BH or per-cpu workqueue, pwq->nr_active is sufficient */ |
5797b1c1 TH |
1735 | obtained = pwq->nr_active < READ_ONCE(wq->max_active); |
1736 | goto out; | |
1737 | } | |
1738 | ||
4c065dbc WL |
1739 | if (unlikely(pwq->plugged)) |
1740 | return false; | |
1741 | ||
5797b1c1 TH |
1742 | /* |
1743 | * Unbound workqueue uses per-node shared nr_active $nna. If @pwq is | |
1744 | * already waiting on $nna, pwq_dec_nr_active() will maintain the | |
1745 | * concurrency level. Don't jump the line. | |
1746 | * | |
1747 | * We need to ignore the pending test after max_active has increased as | |
1748 | * pwq_dec_nr_active() can only maintain the concurrency level but not | |
1749 | * increase it. This is indicated by @fill. | |
1750 | */ | |
1751 | if (!list_empty(&pwq->pending_node) && likely(!fill)) | |
1752 | goto out; | |
1753 | ||
1754 | obtained = tryinc_node_nr_active(nna); | |
1755 | if (obtained) | |
1756 | goto out; | |
1757 | ||
1758 | /* | |
1759 | * Lockless acquisition failed. Lock, add ourself to $nna->pending_pwqs | |
1760 | * and try again. The smp_mb() is paired with the implied memory barrier | |
1761 | * of atomic_dec_return() in pwq_dec_nr_active() to ensure that either | |
1762 | * we see the decremented $nna->nr or they see non-empty | |
1763 | * $nna->pending_pwqs. | |
1764 | */ | |
1765 | raw_spin_lock(&nna->lock); | |
1766 | ||
1767 | if (list_empty(&pwq->pending_node)) | |
1768 | list_add_tail(&pwq->pending_node, &nna->pending_pwqs); | |
1769 | else if (likely(!fill)) | |
1770 | goto out_unlock; | |
1771 | ||
1772 | smp_mb(); | |
1773 | ||
1774 | obtained = tryinc_node_nr_active(nna); | |
1c270b79 | 1775 | |
5797b1c1 TH |
1776 | /* |
1777 | * If @fill, @pwq might have already been pending. Being spuriously | |
1778 | * pending in cold paths doesn't affect anything. Let's leave it be. | |
1779 | */ | |
1780 | if (obtained && likely(!fill)) | |
1781 | list_del_init(&pwq->pending_node); | |
1782 | ||
1783 | out_unlock: | |
1784 | raw_spin_unlock(&nna->lock); | |
1785 | out: | |
1786 | if (obtained) | |
1c270b79 TH |
1787 | pwq->nr_active++; |
1788 | return obtained; | |
1789 | } | |
1790 | ||
1791 | /** | |
1792 | * pwq_activate_first_inactive - Activate the first inactive work item on a pwq | |
1793 | * @pwq: pool_workqueue of interest | |
5797b1c1 | 1794 | * @fill: max_active may have increased, try to increase concurrency level |
1c270b79 TH |
1795 | * |
1796 | * Activate the first inactive work item of @pwq if available and allowed by | |
1797 | * max_active limit. | |
1798 | * | |
1799 | * Returns %true if an inactive work item has been activated. %false if no | |
1800 | * inactive work item is found or max_active limit is reached. | |
1801 | */ | |
5797b1c1 | 1802 | static bool pwq_activate_first_inactive(struct pool_workqueue *pwq, bool fill) |
1c270b79 TH |
1803 | { |
1804 | struct work_struct *work = | |
1805 | list_first_entry_or_null(&pwq->inactive_works, | |
1806 | struct work_struct, entry); | |
1807 | ||
5797b1c1 | 1808 | if (work && pwq_tryinc_nr_active(pwq, fill)) { |
1c270b79 TH |
1809 | __pwq_activate_work(pwq, work); |
1810 | return true; | |
1811 | } else { | |
1812 | return false; | |
1813 | } | |
1814 | } | |
1815 | ||
4c065dbc | 1816 | /** |
516d3dc9 WL |
1817 | * unplug_oldest_pwq - unplug the oldest pool_workqueue |
1818 | * @wq: workqueue_struct where its oldest pwq is to be unplugged | |
4c065dbc | 1819 | * |
516d3dc9 WL |
1820 | * This function should only be called for ordered workqueues where only the |
1821 | * oldest pwq is unplugged, the others are plugged to suspend execution to | |
1822 | * ensure proper work item ordering:: | |
4c065dbc WL |
1823 | * |
1824 | * dfl_pwq --------------+ [P] - plugged | |
1825 | * | | |
1826 | * v | |
1827 | * pwqs -> A -> B [P] -> C [P] (newest) | |
1828 | * | | | | |
1829 | * 1 3 5 | |
1830 | * | | | | |
1831 | * 2 4 6 | |
516d3dc9 WL |
1832 | * |
1833 | * When the oldest pwq is drained and removed, this function should be called | |
1834 | * to unplug the next oldest one to start its work item execution. Note that | |
1835 | * pwq's are linked into wq->pwqs with the oldest first, so the first one in | |
1836 | * the list is the oldest. | |
4c065dbc WL |
1837 | */ |
1838 | static void unplug_oldest_pwq(struct workqueue_struct *wq) | |
1839 | { | |
1840 | struct pool_workqueue *pwq; | |
1841 | ||
1842 | lockdep_assert_held(&wq->mutex); | |
1843 | ||
1844 | /* Caller should make sure that pwqs isn't empty before calling */ | |
1845 | pwq = list_first_entry_or_null(&wq->pwqs, struct pool_workqueue, | |
1846 | pwqs_node); | |
1847 | raw_spin_lock_irq(&pwq->pool->lock); | |
1848 | if (pwq->plugged) { | |
1849 | pwq->plugged = false; | |
1850 | if (pwq_activate_first_inactive(pwq, true)) | |
1851 | kick_pool(pwq->pool); | |
1852 | } | |
1853 | raw_spin_unlock_irq(&pwq->pool->lock); | |
1854 | } | |
1855 | ||
5797b1c1 TH |
1856 | /** |
1857 | * node_activate_pending_pwq - Activate a pending pwq on a wq_node_nr_active | |
1858 | * @nna: wq_node_nr_active to activate a pending pwq for | |
1859 | * @caller_pool: worker_pool the caller is locking | |
1860 | * | |
1861 | * Activate a pwq in @nna->pending_pwqs. Called with @caller_pool locked. | |
1862 | * @caller_pool may be unlocked and relocked to lock other worker_pools. | |
1863 | */ | |
1864 | static void node_activate_pending_pwq(struct wq_node_nr_active *nna, | |
1865 | struct worker_pool *caller_pool) | |
1866 | { | |
1867 | struct worker_pool *locked_pool = caller_pool; | |
1868 | struct pool_workqueue *pwq; | |
1869 | struct work_struct *work; | |
1870 | ||
1871 | lockdep_assert_held(&caller_pool->lock); | |
1872 | ||
1873 | raw_spin_lock(&nna->lock); | |
1874 | retry: | |
1875 | pwq = list_first_entry_or_null(&nna->pending_pwqs, | |
1876 | struct pool_workqueue, pending_node); | |
1877 | if (!pwq) | |
1878 | goto out_unlock; | |
1879 | ||
1880 | /* | |
1881 | * If @pwq is for a different pool than @locked_pool, we need to lock | |
1882 | * @pwq->pool->lock. Let's trylock first. If unsuccessful, do the unlock | |
1883 | * / lock dance. For that, we also need to release @nna->lock as it's | |
1884 | * nested inside pool locks. | |
1885 | */ | |
1886 | if (pwq->pool != locked_pool) { | |
1887 | raw_spin_unlock(&locked_pool->lock); | |
1888 | locked_pool = pwq->pool; | |
1889 | if (!raw_spin_trylock(&locked_pool->lock)) { | |
1890 | raw_spin_unlock(&nna->lock); | |
1891 | raw_spin_lock(&locked_pool->lock); | |
1892 | raw_spin_lock(&nna->lock); | |
1893 | goto retry; | |
1894 | } | |
1895 | } | |
1896 | ||
1897 | /* | |
1898 | * $pwq may not have any inactive work items due to e.g. cancellations. | |
1899 | * Drop it from pending_pwqs and see if there's another one. | |
1900 | */ | |
1901 | work = list_first_entry_or_null(&pwq->inactive_works, | |
1902 | struct work_struct, entry); | |
1903 | if (!work) { | |
1904 | list_del_init(&pwq->pending_node); | |
1905 | goto retry; | |
1906 | } | |
1907 | ||
1908 | /* | |
1909 | * Acquire an nr_active count and activate the inactive work item. If | |
1910 | * $pwq still has inactive work items, rotate it to the end of the | |
1911 | * pending_pwqs so that we round-robin through them. This means that | |
1912 | * inactive work items are not activated in queueing order which is fine | |
1913 | * given that there has never been any ordering across different pwqs. | |
1914 | */ | |
1915 | if (likely(tryinc_node_nr_active(nna))) { | |
1916 | pwq->nr_active++; | |
1917 | __pwq_activate_work(pwq, work); | |
1918 | ||
1919 | if (list_empty(&pwq->inactive_works)) | |
1920 | list_del_init(&pwq->pending_node); | |
1921 | else | |
1922 | list_move_tail(&pwq->pending_node, &nna->pending_pwqs); | |
1923 | ||
1924 | /* if activating a foreign pool, make sure it's running */ | |
1925 | if (pwq->pool != caller_pool) | |
1926 | kick_pool(pwq->pool); | |
1927 | } | |
1928 | ||
1929 | out_unlock: | |
1930 | raw_spin_unlock(&nna->lock); | |
1931 | if (locked_pool != caller_pool) { | |
1932 | raw_spin_unlock(&locked_pool->lock); | |
1933 | raw_spin_lock(&caller_pool->lock); | |
1934 | } | |
1935 | } | |
1936 | ||
1c270b79 TH |
1937 | /** |
1938 | * pwq_dec_nr_active - Retire an active count | |
1939 | * @pwq: pool_workqueue of interest | |
1940 | * | |
1941 | * Decrement @pwq's nr_active and try to activate the first inactive work item. | |
5797b1c1 | 1942 | * For unbound workqueues, this function may temporarily drop @pwq->pool->lock. |
1c270b79 TH |
1943 | */ |
1944 | static void pwq_dec_nr_active(struct pool_workqueue *pwq) | |
3aa62497 | 1945 | { |
1c270b79 | 1946 | struct worker_pool *pool = pwq->pool; |
91ccc6e7 | 1947 | struct wq_node_nr_active *nna = wq_node_nr_active(pwq->wq, pool->node); |
3aa62497 | 1948 | |
1c270b79 TH |
1949 | lockdep_assert_held(&pool->lock); |
1950 | ||
91ccc6e7 TH |
1951 | /* |
1952 | * @pwq->nr_active should be decremented for both percpu and unbound | |
1953 | * workqueues. | |
1954 | */ | |
1c270b79 | 1955 | pwq->nr_active--; |
91ccc6e7 TH |
1956 | |
1957 | /* | |
1958 | * For a percpu workqueue, it's simple. Just need to kick the first | |
1959 | * inactive work item on @pwq itself. | |
1960 | */ | |
1961 | if (!nna) { | |
5797b1c1 | 1962 | pwq_activate_first_inactive(pwq, false); |
91ccc6e7 TH |
1963 | return; |
1964 | } | |
1965 | ||
5797b1c1 TH |
1966 | /* |
1967 | * If @pwq is for an unbound workqueue, it's more complicated because | |
1968 | * multiple pwqs and pools may be sharing the nr_active count. When a | |
1969 | * pwq needs to wait for an nr_active count, it puts itself on | |
1970 | * $nna->pending_pwqs. The following atomic_dec_return()'s implied | |
1971 | * memory barrier is paired with smp_mb() in pwq_tryinc_nr_active() to | |
1972 | * guarantee that either we see non-empty pending_pwqs or they see | |
1973 | * decremented $nna->nr. | |
1974 | * | |
1975 | * $nna->max may change as CPUs come online/offline and @pwq->wq's | |
1976 | * max_active gets updated. However, it is guaranteed to be equal to or | |
1977 | * larger than @pwq->wq->min_active which is above zero unless freezing. | |
1978 | * This maintains the forward progress guarantee. | |
1979 | */ | |
1980 | if (atomic_dec_return(&nna->nr) >= READ_ONCE(nna->max)) | |
1981 | return; | |
1982 | ||
1983 | if (!list_empty(&nna->pending_pwqs)) | |
1984 | node_activate_pending_pwq(nna, pool); | |
3aa62497 LJ |
1985 | } |
1986 | ||
bf4ede01 | 1987 | /** |
112202d9 TH |
1988 | * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight |
1989 | * @pwq: pwq of interest | |
c4560c2c | 1990 | * @work_data: work_data of work which left the queue |
bf4ede01 TH |
1991 | * |
1992 | * A work either has completed or is removed from pending queue, | |
112202d9 | 1993 | * decrement nr_in_flight of its pwq and handle workqueue flushing. |
bf4ede01 | 1994 | * |
dd6c3c54 TH |
1995 | * NOTE: |
1996 | * For unbound workqueues, this function may temporarily drop @pwq->pool->lock | |
1997 | * and thus should be called after all other state updates for the in-flight | |
1998 | * work item is complete. | |
1999 | * | |
bf4ede01 | 2000 | * CONTEXT: |
a9b8a985 | 2001 | * raw_spin_lock_irq(pool->lock). |
bf4ede01 | 2002 | */ |
c4560c2c | 2003 | static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, unsigned long work_data) |
bf4ede01 | 2004 | { |
c4560c2c LJ |
2005 | int color = get_work_color(work_data); |
2006 | ||
1c270b79 TH |
2007 | if (!(work_data & WORK_STRUCT_INACTIVE)) |
2008 | pwq_dec_nr_active(pwq); | |
018f3a13 | 2009 | |
112202d9 | 2010 | pwq->nr_in_flight[color]--; |
bf4ede01 | 2011 | |
bf4ede01 | 2012 | /* is flush in progress and are we at the flushing tip? */ |
112202d9 | 2013 | if (likely(pwq->flush_color != color)) |
8864b4e5 | 2014 | goto out_put; |
bf4ede01 TH |
2015 | |
2016 | /* are there still in-flight works? */ | |
112202d9 | 2017 | if (pwq->nr_in_flight[color]) |
8864b4e5 | 2018 | goto out_put; |
bf4ede01 | 2019 | |
112202d9 TH |
2020 | /* this pwq is done, clear flush_color */ |
2021 | pwq->flush_color = -1; | |
bf4ede01 TH |
2022 | |
2023 | /* | |
112202d9 | 2024 | * If this was the last pwq, wake up the first flusher. It |
bf4ede01 TH |
2025 | * will handle the rest. |
2026 | */ | |
112202d9 TH |
2027 | if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) |
2028 | complete(&pwq->wq->first_flusher->done); | |
8864b4e5 TH |
2029 | out_put: |
2030 | put_pwq(pwq); | |
bf4ede01 TH |
2031 | } |
2032 | ||
36e227d2 | 2033 | /** |
bbb68dfa | 2034 | * try_to_grab_pending - steal work item from worklist and disable irq |
36e227d2 | 2035 | * @work: work item to steal |
c5f5b942 | 2036 | * @cflags: %WORK_CANCEL_ flags |
c26e2f2e | 2037 | * @irq_flags: place to store irq state |
36e227d2 TH |
2038 | * |
2039 | * Try to grab PENDING bit of @work. This function can handle @work in any | |
d185af30 | 2040 | * stable state - idle, on timer or on worklist. |
36e227d2 | 2041 | * |
d185af30 | 2042 | * Return: |
3eb6b31b MCC |
2043 | * |
2044 | * ======== ================================================================ | |
36e227d2 TH |
2045 | * 1 if @work was pending and we successfully stole PENDING |
2046 | * 0 if @work was idle and we claimed PENDING | |
2047 | * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry | |
3eb6b31b | 2048 | * ======== ================================================================ |
36e227d2 | 2049 | * |
d185af30 | 2050 | * Note: |
bbb68dfa | 2051 | * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting |
e0aecdd8 TH |
2052 | * interrupted while holding PENDING and @work off queue, irq must be |
2053 | * disabled on entry. This, combined with delayed_work->timer being | |
2054 | * irqsafe, ensures that we return -EAGAIN for finite short period of time. | |
bbb68dfa TH |
2055 | * |
2056 | * On successful return, >= 0, irq is disabled and the caller is | |
c26e2f2e | 2057 | * responsible for releasing it using local_irq_restore(*@irq_flags). |
bbb68dfa | 2058 | * |
e0aecdd8 | 2059 | * This function is safe to call from any context including IRQ handler. |
bf4ede01 | 2060 | */ |
c5f5b942 | 2061 | static int try_to_grab_pending(struct work_struct *work, u32 cflags, |
c26e2f2e | 2062 | unsigned long *irq_flags) |
bf4ede01 | 2063 | { |
d565ed63 | 2064 | struct worker_pool *pool; |
112202d9 | 2065 | struct pool_workqueue *pwq; |
bf4ede01 | 2066 | |
c26e2f2e | 2067 | local_irq_save(*irq_flags); |
bbb68dfa | 2068 | |
36e227d2 | 2069 | /* try to steal the timer if it exists */ |
c5f5b942 | 2070 | if (cflags & WORK_CANCEL_DELAYED) { |
36e227d2 TH |
2071 | struct delayed_work *dwork = to_delayed_work(work); |
2072 | ||
e0aecdd8 TH |
2073 | /* |
2074 | * dwork->timer is irqsafe. If del_timer() fails, it's | |
2075 | * guaranteed that the timer is not queued anywhere and not | |
2076 | * running on the local CPU. | |
2077 | */ | |
36e227d2 TH |
2078 | if (likely(del_timer(&dwork->timer))) |
2079 | return 1; | |
2080 | } | |
2081 | ||
2082 | /* try to claim PENDING the normal way */ | |
bf4ede01 TH |
2083 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) |
2084 | return 0; | |
2085 | ||
24acfb71 | 2086 | rcu_read_lock(); |
bf4ede01 TH |
2087 | /* |
2088 | * The queueing is in progress, or it is already queued. Try to | |
2089 | * steal it from ->worklist without clearing WORK_STRUCT_PENDING. | |
2090 | */ | |
d565ed63 TH |
2091 | pool = get_work_pool(work); |
2092 | if (!pool) | |
bbb68dfa | 2093 | goto fail; |
bf4ede01 | 2094 | |
a9b8a985 | 2095 | raw_spin_lock(&pool->lock); |
0b3dae68 | 2096 | /* |
112202d9 TH |
2097 | * work->data is guaranteed to point to pwq only while the work |
2098 | * item is queued on pwq->wq, and both updating work->data to point | |
2099 | * to pwq on queueing and to pool on dequeueing are done under | |
2100 | * pwq->pool->lock. This in turn guarantees that, if work->data | |
2101 | * points to pwq which is associated with a locked pool, the work | |
0b3dae68 LJ |
2102 | * item is currently queued on that pool. |
2103 | */ | |
112202d9 TH |
2104 | pwq = get_work_pwq(work); |
2105 | if (pwq && pwq->pool == pool) { | |
c70e1779 TH |
2106 | unsigned long work_data; |
2107 | ||
16062836 TH |
2108 | debug_work_deactivate(work); |
2109 | ||
2110 | /* | |
018f3a13 LJ |
2111 | * A cancelable inactive work item must be in the |
2112 | * pwq->inactive_works since a queued barrier can't be | |
2113 | * canceled (see the comments in insert_wq_barrier()). | |
2114 | * | |
f97a4a1a | 2115 | * An inactive work item cannot be grabbed directly because |
d812796e | 2116 | * it might have linked barrier work items which, if left |
f97a4a1a | 2117 | * on the inactive_works list, will confuse pwq->nr_active |
16062836 TH |
2118 | * management later on and cause stall. Make sure the work |
2119 | * item is activated before grabbing. | |
2120 | */ | |
4c638030 | 2121 | pwq_activate_work(pwq, work); |
16062836 TH |
2122 | |
2123 | list_del_init(&work->entry); | |
16062836 | 2124 | |
c70e1779 TH |
2125 | /* |
2126 | * work->data points to pwq iff queued. Let's point to pool. As | |
2127 | * this destroys work->data needed by the next step, stash it. | |
2128 | */ | |
2129 | work_data = *work_data_bits(work); | |
456a78ee TH |
2130 | set_work_pool_and_keep_pending(work, pool->id, |
2131 | pool_offq_flags(pool)); | |
16062836 | 2132 | |
dd6c3c54 | 2133 | /* must be the last step, see the function comment */ |
c70e1779 | 2134 | pwq_dec_nr_in_flight(pwq, work_data); |
dd6c3c54 | 2135 | |
a9b8a985 | 2136 | raw_spin_unlock(&pool->lock); |
24acfb71 | 2137 | rcu_read_unlock(); |
16062836 | 2138 | return 1; |
bf4ede01 | 2139 | } |
a9b8a985 | 2140 | raw_spin_unlock(&pool->lock); |
bbb68dfa | 2141 | fail: |
24acfb71 | 2142 | rcu_read_unlock(); |
c26e2f2e | 2143 | local_irq_restore(*irq_flags); |
36e227d2 | 2144 | return -EAGAIN; |
bf4ede01 TH |
2145 | } |
2146 | ||
978b8409 TH |
2147 | /** |
2148 | * work_grab_pending - steal work item from worklist and disable irq | |
2149 | * @work: work item to steal | |
2150 | * @cflags: %WORK_CANCEL_ flags | |
2151 | * @irq_flags: place to store IRQ state | |
2152 | * | |
2153 | * Grab PENDING bit of @work. @work can be in any stable state - idle, on timer | |
2154 | * or on worklist. | |
2155 | * | |
f09b10b6 | 2156 | * Can be called from any context. IRQ is disabled on return with IRQ state |
978b8409 TH |
2157 | * stored in *@irq_flags. The caller is responsible for re-enabling it using |
2158 | * local_irq_restore(). | |
2159 | * | |
2160 | * Returns %true if @work was pending. %false if idle. | |
2161 | */ | |
2162 | static bool work_grab_pending(struct work_struct *work, u32 cflags, | |
2163 | unsigned long *irq_flags) | |
2164 | { | |
978b8409 TH |
2165 | int ret; |
2166 | ||
f09b10b6 TH |
2167 | while (true) { |
2168 | ret = try_to_grab_pending(work, cflags, irq_flags); | |
2169 | if (ret >= 0) | |
2170 | return ret; | |
2171 | cpu_relax(); | |
2172 | } | |
978b8409 TH |
2173 | } |
2174 | ||
4690c4ab | 2175 | /** |
706026c2 | 2176 | * insert_work - insert a work into a pool |
112202d9 | 2177 | * @pwq: pwq @work belongs to |
4690c4ab TH |
2178 | * @work: work to insert |
2179 | * @head: insertion point | |
2180 | * @extra_flags: extra WORK_STRUCT_* flags to set | |
2181 | * | |
112202d9 | 2182 | * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to |
706026c2 | 2183 | * work_struct flags. |
4690c4ab TH |
2184 | * |
2185 | * CONTEXT: | |
a9b8a985 | 2186 | * raw_spin_lock_irq(pool->lock). |
4690c4ab | 2187 | */ |
112202d9 TH |
2188 | static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, |
2189 | struct list_head *head, unsigned int extra_flags) | |
b89deed3 | 2190 | { |
fe089f87 | 2191 | debug_work_activate(work); |
e22bee78 | 2192 | |
e89a85d6 | 2193 | /* record the work call stack in order to print it in KASAN reports */ |
f70da745 | 2194 | kasan_record_aux_stack_noalloc(work); |
e89a85d6 | 2195 | |
4690c4ab | 2196 | /* we own @work, set data and link */ |
112202d9 | 2197 | set_work_pwq(work, pwq, extra_flags); |
1a4d9b0a | 2198 | list_add_tail(&work->entry, head); |
8864b4e5 | 2199 | get_pwq(pwq); |
b89deed3 ON |
2200 | } |
2201 | ||
c8efcc25 TH |
2202 | /* |
2203 | * Test whether @work is being queued from another work executing on the | |
8d03ecfe | 2204 | * same workqueue. |
c8efcc25 TH |
2205 | */ |
2206 | static bool is_chained_work(struct workqueue_struct *wq) | |
2207 | { | |
8d03ecfe TH |
2208 | struct worker *worker; |
2209 | ||
2210 | worker = current_wq_worker(); | |
2211 | /* | |
bf393fd4 | 2212 | * Return %true iff I'm a worker executing a work item on @wq. If |
8d03ecfe TH |
2213 | * I'm @worker, it's safe to dereference it without locking. |
2214 | */ | |
112202d9 | 2215 | return worker && worker->current_pwq->wq == wq; |
c8efcc25 TH |
2216 | } |
2217 | ||
ef557180 MG |
2218 | /* |
2219 | * When queueing an unbound work item to a wq, prefer local CPU if allowed | |
2220 | * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to | |
2221 | * avoid perturbing sensitive tasks. | |
2222 | */ | |
2223 | static int wq_select_unbound_cpu(int cpu) | |
2224 | { | |
2225 | int new_cpu; | |
2226 | ||
f303fccb TH |
2227 | if (likely(!wq_debug_force_rr_cpu)) { |
2228 | if (cpumask_test_cpu(cpu, wq_unbound_cpumask)) | |
2229 | return cpu; | |
a8ec5880 AF |
2230 | } else { |
2231 | pr_warn_once("workqueue: round-robin CPU selection forced, expect performance impact\n"); | |
f303fccb TH |
2232 | } |
2233 | ||
ef557180 MG |
2234 | new_cpu = __this_cpu_read(wq_rr_cpu_last); |
2235 | new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask); | |
2236 | if (unlikely(new_cpu >= nr_cpu_ids)) { | |
2237 | new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask); | |
2238 | if (unlikely(new_cpu >= nr_cpu_ids)) | |
2239 | return cpu; | |
2240 | } | |
2241 | __this_cpu_write(wq_rr_cpu_last, new_cpu); | |
2242 | ||
2243 | return new_cpu; | |
2244 | } | |
2245 | ||
d84ff051 | 2246 | static void __queue_work(int cpu, struct workqueue_struct *wq, |
1da177e4 LT |
2247 | struct work_struct *work) |
2248 | { | |
112202d9 | 2249 | struct pool_workqueue *pwq; |
fe089f87 | 2250 | struct worker_pool *last_pool, *pool; |
8a2e8e5d | 2251 | unsigned int work_flags; |
b75cac93 | 2252 | unsigned int req_cpu = cpu; |
8930caba TH |
2253 | |
2254 | /* | |
2255 | * While a work item is PENDING && off queue, a task trying to | |
2256 | * steal the PENDING will busy-loop waiting for it to either get | |
2257 | * queued or lose PENDING. Grabbing PENDING and queueing should | |
2258 | * happen with IRQ disabled. | |
2259 | */ | |
8e8eb730 | 2260 | lockdep_assert_irqs_disabled(); |
1da177e4 | 2261 | |
33e3f0a3 RC |
2262 | /* |
2263 | * For a draining wq, only works from the same workqueue are | |
2264 | * allowed. The __WQ_DESTROYING helps to spot the issue that | |
2265 | * queues a new work item to a wq after destroy_workqueue(wq). | |
2266 | */ | |
2267 | if (unlikely(wq->flags & (__WQ_DESTROYING | __WQ_DRAINING) && | |
2268 | WARN_ON_ONCE(!is_chained_work(wq)))) | |
e41e704b | 2269 | return; |
24acfb71 | 2270 | rcu_read_lock(); |
9e8cd2f5 | 2271 | retry: |
c9178087 | 2272 | /* pwq which will be used unless @work is executing elsewhere */ |
636b927e TH |
2273 | if (req_cpu == WORK_CPU_UNBOUND) { |
2274 | if (wq->flags & WQ_UNBOUND) | |
aa202f1f | 2275 | cpu = wq_select_unbound_cpu(raw_smp_processor_id()); |
636b927e | 2276 | else |
aa202f1f | 2277 | cpu = raw_smp_processor_id(); |
aa202f1f | 2278 | } |
dbf2576e | 2279 | |
636b927e | 2280 | pwq = rcu_dereference(*per_cpu_ptr(wq->cpu_pwq, cpu)); |
fe089f87 TH |
2281 | pool = pwq->pool; |
2282 | ||
c9178087 TH |
2283 | /* |
2284 | * If @work was previously on a different pool, it might still be | |
2285 | * running there, in which case the work needs to be queued on that | |
2286 | * pool to guarantee non-reentrancy. | |
2287 | */ | |
2288 | last_pool = get_work_pool(work); | |
fe089f87 | 2289 | if (last_pool && last_pool != pool) { |
c9178087 | 2290 | struct worker *worker; |
18aa9eff | 2291 | |
a9b8a985 | 2292 | raw_spin_lock(&last_pool->lock); |
18aa9eff | 2293 | |
c9178087 | 2294 | worker = find_worker_executing_work(last_pool, work); |
18aa9eff | 2295 | |
c9178087 TH |
2296 | if (worker && worker->current_pwq->wq == wq) { |
2297 | pwq = worker->current_pwq; | |
fe089f87 TH |
2298 | pool = pwq->pool; |
2299 | WARN_ON_ONCE(pool != last_pool); | |
8930caba | 2300 | } else { |
c9178087 | 2301 | /* meh... not running there, queue here */ |
a9b8a985 | 2302 | raw_spin_unlock(&last_pool->lock); |
fe089f87 | 2303 | raw_spin_lock(&pool->lock); |
8930caba | 2304 | } |
f3421797 | 2305 | } else { |
fe089f87 | 2306 | raw_spin_lock(&pool->lock); |
502ca9d8 TH |
2307 | } |
2308 | ||
9e8cd2f5 | 2309 | /* |
636b927e TH |
2310 | * pwq is determined and locked. For unbound pools, we could have raced |
2311 | * with pwq release and it could already be dead. If its refcnt is zero, | |
2312 | * repeat pwq selection. Note that unbound pwqs never die without | |
2313 | * another pwq replacing it in cpu_pwq or while work items are executing | |
2314 | * on it, so the retrying is guaranteed to make forward-progress. | |
9e8cd2f5 TH |
2315 | */ |
2316 | if (unlikely(!pwq->refcnt)) { | |
2317 | if (wq->flags & WQ_UNBOUND) { | |
fe089f87 | 2318 | raw_spin_unlock(&pool->lock); |
9e8cd2f5 TH |
2319 | cpu_relax(); |
2320 | goto retry; | |
2321 | } | |
2322 | /* oops */ | |
2323 | WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", | |
2324 | wq->name, cpu); | |
2325 | } | |
2326 | ||
112202d9 TH |
2327 | /* pwq determined, queue */ |
2328 | trace_workqueue_queue_work(req_cpu, pwq, work); | |
502ca9d8 | 2329 | |
24acfb71 TG |
2330 | if (WARN_ON(!list_empty(&work->entry))) |
2331 | goto out; | |
1e19ffc6 | 2332 | |
112202d9 TH |
2333 | pwq->nr_in_flight[pwq->work_color]++; |
2334 | work_flags = work_color_to_flags(pwq->work_color); | |
1e19ffc6 | 2335 | |
a045a272 TH |
2336 | /* |
2337 | * Limit the number of concurrently active work items to max_active. | |
2338 | * @work must also queue behind existing inactive work items to maintain | |
2339 | * ordering when max_active changes. See wq_adjust_max_active(). | |
2340 | */ | |
5797b1c1 | 2341 | if (list_empty(&pwq->inactive_works) && pwq_tryinc_nr_active(pwq, false)) { |
fe089f87 TH |
2342 | if (list_empty(&pool->worklist)) |
2343 | pool->watchdog_ts = jiffies; | |
2344 | ||
cdadf009 | 2345 | trace_workqueue_activate_work(work); |
fe089f87 | 2346 | insert_work(pwq, work, &pool->worklist, work_flags); |
0219a352 | 2347 | kick_pool(pool); |
8a2e8e5d | 2348 | } else { |
f97a4a1a | 2349 | work_flags |= WORK_STRUCT_INACTIVE; |
fe089f87 | 2350 | insert_work(pwq, work, &pwq->inactive_works, work_flags); |
8a2e8e5d | 2351 | } |
1e19ffc6 | 2352 | |
24acfb71 | 2353 | out: |
fe089f87 | 2354 | raw_spin_unlock(&pool->lock); |
24acfb71 | 2355 | rcu_read_unlock(); |
1da177e4 LT |
2356 | } |
2357 | ||
86898fa6 TH |
2358 | static bool clear_pending_if_disabled(struct work_struct *work) |
2359 | { | |
2360 | unsigned long data = *work_data_bits(work); | |
2361 | struct work_offq_data offqd; | |
2362 | ||
2363 | if (likely((data & WORK_STRUCT_PWQ) || | |
2364 | !(data & WORK_OFFQ_DISABLE_MASK))) | |
2365 | return false; | |
2366 | ||
2367 | work_offqd_unpack(&offqd, data); | |
2368 | set_work_pool_and_clear_pending(work, offqd.pool_id, | |
2369 | work_offqd_pack_flags(&offqd)); | |
2370 | return true; | |
2371 | } | |
2372 | ||
0fcb78c2 | 2373 | /** |
c1a220e7 ZR |
2374 | * queue_work_on - queue work on specific cpu |
2375 | * @cpu: CPU number to execute work on | |
0fcb78c2 REB |
2376 | * @wq: workqueue to use |
2377 | * @work: work to queue | |
2378 | * | |
c1a220e7 | 2379 | * We queue the work to a specific CPU, the caller must ensure it |
443378f0 PM |
2380 | * can't go away. Callers that fail to ensure that the specified |
2381 | * CPU cannot go away will execute on a randomly chosen CPU. | |
854f5cc5 PM |
2382 | * But note well that callers specifying a CPU that never has been |
2383 | * online will get a splat. | |
d185af30 YB |
2384 | * |
2385 | * Return: %false if @work was already on a queue, %true otherwise. | |
1da177e4 | 2386 | */ |
d4283e93 TH |
2387 | bool queue_work_on(int cpu, struct workqueue_struct *wq, |
2388 | struct work_struct *work) | |
1da177e4 | 2389 | { |
d4283e93 | 2390 | bool ret = false; |
c26e2f2e | 2391 | unsigned long irq_flags; |
ef1ca236 | 2392 | |
c26e2f2e | 2393 | local_irq_save(irq_flags); |
c1a220e7 | 2394 | |
86898fa6 TH |
2395 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)) && |
2396 | !clear_pending_if_disabled(work)) { | |
4690c4ab | 2397 | __queue_work(cpu, wq, work); |
d4283e93 | 2398 | ret = true; |
c1a220e7 | 2399 | } |
ef1ca236 | 2400 | |
c26e2f2e | 2401 | local_irq_restore(irq_flags); |
1da177e4 LT |
2402 | return ret; |
2403 | } | |
ad7b1f84 | 2404 | EXPORT_SYMBOL(queue_work_on); |
1da177e4 | 2405 | |
8204e0c1 | 2406 | /** |
fef59c9c | 2407 | * select_numa_node_cpu - Select a CPU based on NUMA node |
8204e0c1 AD |
2408 | * @node: NUMA node ID that we want to select a CPU from |
2409 | * | |
2410 | * This function will attempt to find a "random" cpu available on a given | |
2411 | * node. If there are no CPUs available on the given node it will return | |
2412 | * WORK_CPU_UNBOUND indicating that we should just schedule to any | |
2413 | * available CPU if we need to schedule this work. | |
2414 | */ | |
fef59c9c | 2415 | static int select_numa_node_cpu(int node) |
8204e0c1 AD |
2416 | { |
2417 | int cpu; | |
2418 | ||
8204e0c1 AD |
2419 | /* Delay binding to CPU if node is not valid or online */ |
2420 | if (node < 0 || node >= MAX_NUMNODES || !node_online(node)) | |
2421 | return WORK_CPU_UNBOUND; | |
2422 | ||
2423 | /* Use local node/cpu if we are already there */ | |
2424 | cpu = raw_smp_processor_id(); | |
2425 | if (node == cpu_to_node(cpu)) | |
2426 | return cpu; | |
2427 | ||
2428 | /* Use "random" otherwise know as "first" online CPU of node */ | |
2429 | cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask); | |
2430 | ||
2431 | /* If CPU is valid return that, otherwise just defer */ | |
2432 | return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND; | |
2433 | } | |
2434 | ||
2435 | /** | |
2436 | * queue_work_node - queue work on a "random" cpu for a given NUMA node | |
2437 | * @node: NUMA node that we are targeting the work for | |
2438 | * @wq: workqueue to use | |
2439 | * @work: work to queue | |
2440 | * | |
2441 | * We queue the work to a "random" CPU within a given NUMA node. The basic | |
2442 | * idea here is to provide a way to somehow associate work with a given | |
2443 | * NUMA node. | |
2444 | * | |
2445 | * This function will only make a best effort attempt at getting this onto | |
2446 | * the right NUMA node. If no node is requested or the requested node is | |
2447 | * offline then we just fall back to standard queue_work behavior. | |
2448 | * | |
2449 | * Currently the "random" CPU ends up being the first available CPU in the | |
2450 | * intersection of cpu_online_mask and the cpumask of the node, unless we | |
2451 | * are running on the node. In that case we just use the current CPU. | |
2452 | * | |
2453 | * Return: %false if @work was already on a queue, %true otherwise. | |
2454 | */ | |
2455 | bool queue_work_node(int node, struct workqueue_struct *wq, | |
2456 | struct work_struct *work) | |
2457 | { | |
c26e2f2e | 2458 | unsigned long irq_flags; |
8204e0c1 AD |
2459 | bool ret = false; |
2460 | ||
2461 | /* | |
2462 | * This current implementation is specific to unbound workqueues. | |
2463 | * Specifically we only return the first available CPU for a given | |
2464 | * node instead of cycling through individual CPUs within the node. | |
2465 | * | |
2466 | * If this is used with a per-cpu workqueue then the logic in | |
2467 | * workqueue_select_cpu_near would need to be updated to allow for | |
2468 | * some round robin type logic. | |
2469 | */ | |
2470 | WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)); | |
2471 | ||
c26e2f2e | 2472 | local_irq_save(irq_flags); |
8204e0c1 | 2473 | |
86898fa6 TH |
2474 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)) && |
2475 | !clear_pending_if_disabled(work)) { | |
fef59c9c | 2476 | int cpu = select_numa_node_cpu(node); |
8204e0c1 AD |
2477 | |
2478 | __queue_work(cpu, wq, work); | |
2479 | ret = true; | |
2480 | } | |
2481 | ||
c26e2f2e | 2482 | local_irq_restore(irq_flags); |
8204e0c1 AD |
2483 | return ret; |
2484 | } | |
2485 | EXPORT_SYMBOL_GPL(queue_work_node); | |
2486 | ||
8c20feb6 | 2487 | void delayed_work_timer_fn(struct timer_list *t) |
1da177e4 | 2488 | { |
8c20feb6 | 2489 | struct delayed_work *dwork = from_timer(dwork, t, timer); |
1da177e4 | 2490 | |
e0aecdd8 | 2491 | /* should have been called from irqsafe timer with irq already off */ |
60c057bc | 2492 | __queue_work(dwork->cpu, dwork->wq, &dwork->work); |
1da177e4 | 2493 | } |
1438ade5 | 2494 | EXPORT_SYMBOL(delayed_work_timer_fn); |
1da177e4 | 2495 | |
7beb2edf TH |
2496 | static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, |
2497 | struct delayed_work *dwork, unsigned long delay) | |
1da177e4 | 2498 | { |
7beb2edf TH |
2499 | struct timer_list *timer = &dwork->timer; |
2500 | struct work_struct *work = &dwork->work; | |
7beb2edf | 2501 | |
637fdbae | 2502 | WARN_ON_ONCE(!wq); |
4b243563 | 2503 | WARN_ON_ONCE(timer->function != delayed_work_timer_fn); |
fc4b514f TH |
2504 | WARN_ON_ONCE(timer_pending(timer)); |
2505 | WARN_ON_ONCE(!list_empty(&work->entry)); | |
7beb2edf | 2506 | |
8852aac2 TH |
2507 | /* |
2508 | * If @delay is 0, queue @dwork->work immediately. This is for | |
2509 | * both optimization and correctness. The earliest @timer can | |
2510 | * expire is on the closest next tick and delayed_work users depend | |
2511 | * on that there's no such delay when @delay is 0. | |
2512 | */ | |
2513 | if (!delay) { | |
2514 | __queue_work(cpu, wq, &dwork->work); | |
2515 | return; | |
2516 | } | |
2517 | ||
60c057bc | 2518 | dwork->wq = wq; |
1265057f | 2519 | dwork->cpu = cpu; |
7beb2edf TH |
2520 | timer->expires = jiffies + delay; |
2521 | ||
aae17ebb LB |
2522 | if (housekeeping_enabled(HK_TYPE_TIMER)) { |
2523 | /* If the current cpu is a housekeeping cpu, use it. */ | |
2524 | cpu = smp_processor_id(); | |
2525 | if (!housekeeping_test_cpu(cpu, HK_TYPE_TIMER)) | |
2526 | cpu = housekeeping_any_cpu(HK_TYPE_TIMER); | |
041bd12e | 2527 | add_timer_on(timer, cpu); |
aae17ebb LB |
2528 | } else { |
2529 | if (likely(cpu == WORK_CPU_UNBOUND)) | |
d08c407f | 2530 | add_timer_global(timer); |
aae17ebb LB |
2531 | else |
2532 | add_timer_on(timer, cpu); | |
2533 | } | |
1da177e4 LT |
2534 | } |
2535 | ||
0fcb78c2 REB |
2536 | /** |
2537 | * queue_delayed_work_on - queue work on specific CPU after delay | |
2538 | * @cpu: CPU number to execute work on | |
2539 | * @wq: workqueue to use | |
af9997e4 | 2540 | * @dwork: work to queue |
0fcb78c2 REB |
2541 | * @delay: number of jiffies to wait before queueing |
2542 | * | |
d185af30 | 2543 | * Return: %false if @work was already on a queue, %true otherwise. If |
715f1300 TH |
2544 | * @delay is zero and @dwork is idle, it will be scheduled for immediate |
2545 | * execution. | |
0fcb78c2 | 2546 | */ |
d4283e93 TH |
2547 | bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, |
2548 | struct delayed_work *dwork, unsigned long delay) | |
7a6bc1cd | 2549 | { |
52bad64d | 2550 | struct work_struct *work = &dwork->work; |
d4283e93 | 2551 | bool ret = false; |
c26e2f2e | 2552 | unsigned long irq_flags; |
7a6bc1cd | 2553 | |
8930caba | 2554 | /* read the comment in __queue_work() */ |
c26e2f2e | 2555 | local_irq_save(irq_flags); |
7a6bc1cd | 2556 | |
86898fa6 TH |
2557 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)) && |
2558 | !clear_pending_if_disabled(work)) { | |
7beb2edf | 2559 | __queue_delayed_work(cpu, wq, dwork, delay); |
d4283e93 | 2560 | ret = true; |
7a6bc1cd | 2561 | } |
8a3e77cc | 2562 | |
c26e2f2e | 2563 | local_irq_restore(irq_flags); |
7a6bc1cd VP |
2564 | return ret; |
2565 | } | |
ad7b1f84 | 2566 | EXPORT_SYMBOL(queue_delayed_work_on); |
c7fc77f7 | 2567 | |
8376fe22 TH |
2568 | /** |
2569 | * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU | |
2570 | * @cpu: CPU number to execute work on | |
2571 | * @wq: workqueue to use | |
2572 | * @dwork: work to queue | |
2573 | * @delay: number of jiffies to wait before queueing | |
2574 | * | |
2575 | * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, | |
2576 | * modify @dwork's timer so that it expires after @delay. If @delay is | |
2577 | * zero, @work is guaranteed to be scheduled immediately regardless of its | |
2578 | * current state. | |
2579 | * | |
d185af30 | 2580 | * Return: %false if @dwork was idle and queued, %true if @dwork was |
8376fe22 TH |
2581 | * pending and its timer was modified. |
2582 | * | |
e0aecdd8 | 2583 | * This function is safe to call from any context including IRQ handler. |
8376fe22 TH |
2584 | * See try_to_grab_pending() for details. |
2585 | */ | |
2586 | bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, | |
2587 | struct delayed_work *dwork, unsigned long delay) | |
2588 | { | |
c26e2f2e | 2589 | unsigned long irq_flags; |
f09b10b6 | 2590 | bool ret; |
c7fc77f7 | 2591 | |
f09b10b6 | 2592 | ret = work_grab_pending(&dwork->work, WORK_CANCEL_DELAYED, &irq_flags); |
63bc0362 | 2593 | |
f09b10b6 | 2594 | if (!clear_pending_if_disabled(&dwork->work)) |
8376fe22 | 2595 | __queue_delayed_work(cpu, wq, dwork, delay); |
8376fe22 | 2596 | |
f09b10b6 | 2597 | local_irq_restore(irq_flags); |
7a6bc1cd VP |
2598 | return ret; |
2599 | } | |
8376fe22 TH |
2600 | EXPORT_SYMBOL_GPL(mod_delayed_work_on); |
2601 | ||
05f0fe6b TH |
2602 | static void rcu_work_rcufn(struct rcu_head *rcu) |
2603 | { | |
2604 | struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu); | |
2605 | ||
2606 | /* read the comment in __queue_work() */ | |
2607 | local_irq_disable(); | |
2608 | __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work); | |
2609 | local_irq_enable(); | |
2610 | } | |
2611 | ||
2612 | /** | |
2613 | * queue_rcu_work - queue work after a RCU grace period | |
2614 | * @wq: workqueue to use | |
2615 | * @rwork: work to queue | |
2616 | * | |
2617 | * Return: %false if @rwork was already pending, %true otherwise. Note | |
2618 | * that a full RCU grace period is guaranteed only after a %true return. | |
bf393fd4 | 2619 | * While @rwork is guaranteed to be executed after a %false return, the |
05f0fe6b TH |
2620 | * execution may happen before a full RCU grace period has passed. |
2621 | */ | |
2622 | bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork) | |
2623 | { | |
2624 | struct work_struct *work = &rwork->work; | |
2625 | ||
86898fa6 TH |
2626 | /* |
2627 | * rcu_work can't be canceled or disabled. Warn if the user reached | |
2628 | * inside @rwork and disabled the inner work. | |
2629 | */ | |
2630 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)) && | |
2631 | !WARN_ON_ONCE(clear_pending_if_disabled(work))) { | |
05f0fe6b | 2632 | rwork->wq = wq; |
a7e30c0e | 2633 | call_rcu_hurry(&rwork->rcu, rcu_work_rcufn); |
05f0fe6b TH |
2634 | return true; |
2635 | } | |
2636 | ||
2637 | return false; | |
2638 | } | |
2639 | EXPORT_SYMBOL(queue_rcu_work); | |
2640 | ||
f7537df5 | 2641 | static struct worker *alloc_worker(int node) |
c34056a3 TH |
2642 | { |
2643 | struct worker *worker; | |
2644 | ||
f7537df5 | 2645 | worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node); |
c8e55f36 TH |
2646 | if (worker) { |
2647 | INIT_LIST_HEAD(&worker->entry); | |
affee4b2 | 2648 | INIT_LIST_HEAD(&worker->scheduled); |
da028469 | 2649 | INIT_LIST_HEAD(&worker->node); |
e22bee78 TH |
2650 | /* on creation a worker is in !idle && prep state */ |
2651 | worker->flags = WORKER_PREP; | |
c8e55f36 | 2652 | } |
c34056a3 TH |
2653 | return worker; |
2654 | } | |
2655 | ||
9546b29e TH |
2656 | static cpumask_t *pool_allowed_cpus(struct worker_pool *pool) |
2657 | { | |
8639eceb TH |
2658 | if (pool->cpu < 0 && pool->attrs->affn_strict) |
2659 | return pool->attrs->__pod_cpumask; | |
2660 | else | |
2661 | return pool->attrs->cpumask; | |
9546b29e TH |
2662 | } |
2663 | ||
4736cbf7 LJ |
2664 | /** |
2665 | * worker_attach_to_pool() - attach a worker to a pool | |
2666 | * @worker: worker to be attached | |
2667 | * @pool: the target pool | |
2668 | * | |
2669 | * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and | |
2670 | * cpu-binding of @worker are kept coordinated with the pool across | |
2671 | * cpu-[un]hotplugs. | |
2672 | */ | |
2673 | static void worker_attach_to_pool(struct worker *worker, | |
4cb1ef64 | 2674 | struct worker_pool *pool) |
4736cbf7 | 2675 | { |
1258fae7 | 2676 | mutex_lock(&wq_pool_attach_mutex); |
4736cbf7 | 2677 | |
4736cbf7 | 2678 | /* |
4cb1ef64 TH |
2679 | * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains stable |
2680 | * across this function. See the comments above the flag definition for | |
2681 | * details. BH workers are, while per-CPU, always DISASSOCIATED. | |
4736cbf7 | 2682 | */ |
4cb1ef64 | 2683 | if (pool->flags & POOL_DISASSOCIATED) { |
4736cbf7 | 2684 | worker->flags |= WORKER_UNBOUND; |
4cb1ef64 TH |
2685 | } else { |
2686 | WARN_ON_ONCE(pool->flags & POOL_BH); | |
5c25b5ff | 2687 | kthread_set_per_cpu(worker->task, pool->cpu); |
4cb1ef64 | 2688 | } |
4736cbf7 | 2689 | |
640f17c8 | 2690 | if (worker->rescue_wq) |
9546b29e | 2691 | set_cpus_allowed_ptr(worker->task, pool_allowed_cpus(pool)); |
640f17c8 | 2692 | |
4736cbf7 | 2693 | list_add_tail(&worker->node, &pool->workers); |
a2d812a2 | 2694 | worker->pool = pool; |
4736cbf7 | 2695 | |
1258fae7 | 2696 | mutex_unlock(&wq_pool_attach_mutex); |
4736cbf7 LJ |
2697 | } |
2698 | ||
60f5a4bc LJ |
2699 | /** |
2700 | * worker_detach_from_pool() - detach a worker from its pool | |
2701 | * @worker: worker which is attached to its pool | |
60f5a4bc | 2702 | * |
4736cbf7 LJ |
2703 | * Undo the attaching which had been done in worker_attach_to_pool(). The |
2704 | * caller worker shouldn't access to the pool after detached except it has | |
2705 | * other reference to the pool. | |
60f5a4bc | 2706 | */ |
a2d812a2 | 2707 | static void worker_detach_from_pool(struct worker *worker) |
60f5a4bc | 2708 | { |
a2d812a2 | 2709 | struct worker_pool *pool = worker->pool; |
60f5a4bc LJ |
2710 | struct completion *detach_completion = NULL; |
2711 | ||
4cb1ef64 TH |
2712 | /* there is one permanent BH worker per CPU which should never detach */ |
2713 | WARN_ON_ONCE(pool->flags & POOL_BH); | |
2714 | ||
1258fae7 | 2715 | mutex_lock(&wq_pool_attach_mutex); |
a2d812a2 | 2716 | |
5c25b5ff | 2717 | kthread_set_per_cpu(worker->task, -1); |
da028469 | 2718 | list_del(&worker->node); |
a2d812a2 TH |
2719 | worker->pool = NULL; |
2720 | ||
e02b9312 | 2721 | if (list_empty(&pool->workers) && list_empty(&pool->dying_workers)) |
60f5a4bc | 2722 | detach_completion = pool->detach_completion; |
1258fae7 | 2723 | mutex_unlock(&wq_pool_attach_mutex); |
60f5a4bc | 2724 | |
b62c0751 LJ |
2725 | /* clear leftover flags without pool->lock after it is detached */ |
2726 | worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND); | |
2727 | ||
60f5a4bc LJ |
2728 | if (detach_completion) |
2729 | complete(detach_completion); | |
2730 | } | |
2731 | ||
c34056a3 TH |
2732 | /** |
2733 | * create_worker - create a new workqueue worker | |
63d95a91 | 2734 | * @pool: pool the new worker will belong to |
c34056a3 | 2735 | * |
051e1850 | 2736 | * Create and start a new worker which is attached to @pool. |
c34056a3 TH |
2737 | * |
2738 | * CONTEXT: | |
2739 | * Might sleep. Does GFP_KERNEL allocations. | |
2740 | * | |
d185af30 | 2741 | * Return: |
c34056a3 TH |
2742 | * Pointer to the newly created worker. |
2743 | */ | |
bc2ae0f5 | 2744 | static struct worker *create_worker(struct worker_pool *pool) |
c34056a3 | 2745 | { |
e441b56f ZL |
2746 | struct worker *worker; |
2747 | int id; | |
5d9c7a1e | 2748 | char id_buf[23]; |
c34056a3 | 2749 | |
7cda9aae | 2750 | /* ID is needed to determine kthread name */ |
e441b56f | 2751 | id = ida_alloc(&pool->worker_ida, GFP_KERNEL); |
3f0ea0b8 PM |
2752 | if (id < 0) { |
2753 | pr_err_once("workqueue: Failed to allocate a worker ID: %pe\n", | |
2754 | ERR_PTR(id)); | |
e441b56f | 2755 | return NULL; |
3f0ea0b8 | 2756 | } |
c34056a3 | 2757 | |
f7537df5 | 2758 | worker = alloc_worker(pool->node); |
3f0ea0b8 PM |
2759 | if (!worker) { |
2760 | pr_err_once("workqueue: Failed to allocate a worker\n"); | |
c34056a3 | 2761 | goto fail; |
3f0ea0b8 | 2762 | } |
c34056a3 | 2763 | |
c34056a3 TH |
2764 | worker->id = id; |
2765 | ||
4cb1ef64 TH |
2766 | if (!(pool->flags & POOL_BH)) { |
2767 | if (pool->cpu >= 0) | |
2768 | snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, | |
2769 | pool->attrs->nice < 0 ? "H" : ""); | |
2770 | else | |
2771 | snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); | |
2772 | ||
2773 | worker->task = kthread_create_on_node(worker_thread, worker, | |
2774 | pool->node, "kworker/%s", id_buf); | |
2775 | if (IS_ERR(worker->task)) { | |
2776 | if (PTR_ERR(worker->task) == -EINTR) { | |
2777 | pr_err("workqueue: Interrupted when creating a worker thread \"kworker/%s\"\n", | |
2778 | id_buf); | |
2779 | } else { | |
2780 | pr_err_once("workqueue: Failed to create a worker thread: %pe", | |
2781 | worker->task); | |
2782 | } | |
2783 | goto fail; | |
60f54038 | 2784 | } |
c34056a3 | 2785 | |
4cb1ef64 TH |
2786 | set_user_nice(worker->task, pool->attrs->nice); |
2787 | kthread_bind_mask(worker->task, pool_allowed_cpus(pool)); | |
2788 | } | |
91151228 | 2789 | |
da028469 | 2790 | /* successful, attach the worker to the pool */ |
4736cbf7 | 2791 | worker_attach_to_pool(worker, pool); |
822d8405 | 2792 | |
051e1850 | 2793 | /* start the newly created worker */ |
a9b8a985 | 2794 | raw_spin_lock_irq(&pool->lock); |
0219a352 | 2795 | |
051e1850 LJ |
2796 | worker->pool->nr_workers++; |
2797 | worker_enter_idle(worker); | |
0219a352 TH |
2798 | |
2799 | /* | |
2800 | * @worker is waiting on a completion in kthread() and will trigger hung | |
6a229b0e TH |
2801 | * check if not woken up soon. As kick_pool() is noop if @pool is empty, |
2802 | * wake it up explicitly. | |
0219a352 | 2803 | */ |
4cb1ef64 TH |
2804 | if (worker->task) |
2805 | wake_up_process(worker->task); | |
0219a352 | 2806 | |
a9b8a985 | 2807 | raw_spin_unlock_irq(&pool->lock); |
051e1850 | 2808 | |
c34056a3 | 2809 | return worker; |
822d8405 | 2810 | |
c34056a3 | 2811 | fail: |
e441b56f | 2812 | ida_free(&pool->worker_ida, id); |
c34056a3 TH |
2813 | kfree(worker); |
2814 | return NULL; | |
2815 | } | |
2816 | ||
793777bc VS |
2817 | static void unbind_worker(struct worker *worker) |
2818 | { | |
2819 | lockdep_assert_held(&wq_pool_attach_mutex); | |
2820 | ||
2821 | kthread_set_per_cpu(worker->task, -1); | |
2822 | if (cpumask_intersects(wq_unbound_cpumask, cpu_active_mask)) | |
2823 | WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, wq_unbound_cpumask) < 0); | |
2824 | else | |
2825 | WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0); | |
2826 | } | |
2827 | ||
e02b9312 VS |
2828 | static void wake_dying_workers(struct list_head *cull_list) |
2829 | { | |
2830 | struct worker *worker, *tmp; | |
2831 | ||
2832 | list_for_each_entry_safe(worker, tmp, cull_list, entry) { | |
2833 | list_del_init(&worker->entry); | |
2834 | unbind_worker(worker); | |
2835 | /* | |
2836 | * If the worker was somehow already running, then it had to be | |
2837 | * in pool->idle_list when set_worker_dying() happened or we | |
2838 | * wouldn't have gotten here. | |
2839 | * | |
2840 | * Thus, the worker must either have observed the WORKER_DIE | |
2841 | * flag, or have set its state to TASK_IDLE. Either way, the | |
2842 | * below will be observed by the worker and is safe to do | |
2843 | * outside of pool->lock. | |
2844 | */ | |
2845 | wake_up_process(worker->task); | |
2846 | } | |
2847 | } | |
2848 | ||
c34056a3 | 2849 | /** |
e02b9312 | 2850 | * set_worker_dying - Tag a worker for destruction |
c34056a3 | 2851 | * @worker: worker to be destroyed |
e02b9312 | 2852 | * @list: transfer worker away from its pool->idle_list and into list |
c34056a3 | 2853 | * |
e02b9312 VS |
2854 | * Tag @worker for destruction and adjust @pool stats accordingly. The worker |
2855 | * should be idle. | |
c8e55f36 TH |
2856 | * |
2857 | * CONTEXT: | |
a9b8a985 | 2858 | * raw_spin_lock_irq(pool->lock). |
c34056a3 | 2859 | */ |
e02b9312 | 2860 | static void set_worker_dying(struct worker *worker, struct list_head *list) |
c34056a3 | 2861 | { |
bd7bdd43 | 2862 | struct worker_pool *pool = worker->pool; |
c34056a3 | 2863 | |
cd549687 | 2864 | lockdep_assert_held(&pool->lock); |
e02b9312 | 2865 | lockdep_assert_held(&wq_pool_attach_mutex); |
cd549687 | 2866 | |
c34056a3 | 2867 | /* sanity check frenzy */ |
6183c009 | 2868 | if (WARN_ON(worker->current_work) || |
73eb7fe7 LJ |
2869 | WARN_ON(!list_empty(&worker->scheduled)) || |
2870 | WARN_ON(!(worker->flags & WORKER_IDLE))) | |
6183c009 | 2871 | return; |
c34056a3 | 2872 | |
73eb7fe7 LJ |
2873 | pool->nr_workers--; |
2874 | pool->nr_idle--; | |
5bdfff96 | 2875 | |
cb444766 | 2876 | worker->flags |= WORKER_DIE; |
e02b9312 VS |
2877 | |
2878 | list_move(&worker->entry, list); | |
2879 | list_move(&worker->node, &pool->dying_workers); | |
c34056a3 TH |
2880 | } |
2881 | ||
3f959aa3 VS |
2882 | /** |
2883 | * idle_worker_timeout - check if some idle workers can now be deleted. | |
2884 | * @t: The pool's idle_timer that just expired | |
2885 | * | |
2886 | * The timer is armed in worker_enter_idle(). Note that it isn't disarmed in | |
2887 | * worker_leave_idle(), as a worker flicking between idle and active while its | |
2888 | * pool is at the too_many_workers() tipping point would cause too much timer | |
2889 | * housekeeping overhead. Since IDLE_WORKER_TIMEOUT is long enough, we just let | |
2890 | * it expire and re-evaluate things from there. | |
2891 | */ | |
32a6c723 | 2892 | static void idle_worker_timeout(struct timer_list *t) |
e22bee78 | 2893 | { |
32a6c723 | 2894 | struct worker_pool *pool = from_timer(pool, t, idle_timer); |
3f959aa3 VS |
2895 | bool do_cull = false; |
2896 | ||
2897 | if (work_pending(&pool->idle_cull_work)) | |
2898 | return; | |
e22bee78 | 2899 | |
a9b8a985 | 2900 | raw_spin_lock_irq(&pool->lock); |
e22bee78 | 2901 | |
3f959aa3 | 2902 | if (too_many_workers(pool)) { |
e22bee78 TH |
2903 | struct worker *worker; |
2904 | unsigned long expires; | |
2905 | ||
2906 | /* idle_list is kept in LIFO order, check the last one */ | |
d70f5d57 | 2907 | worker = list_last_entry(&pool->idle_list, struct worker, entry); |
3f959aa3 VS |
2908 | expires = worker->last_active + IDLE_WORKER_TIMEOUT; |
2909 | do_cull = !time_before(jiffies, expires); | |
2910 | ||
2911 | if (!do_cull) | |
2912 | mod_timer(&pool->idle_timer, expires); | |
2913 | } | |
2914 | raw_spin_unlock_irq(&pool->lock); | |
2915 | ||
2916 | if (do_cull) | |
2917 | queue_work(system_unbound_wq, &pool->idle_cull_work); | |
2918 | } | |
2919 | ||
2920 | /** | |
2921 | * idle_cull_fn - cull workers that have been idle for too long. | |
2922 | * @work: the pool's work for handling these idle workers | |
2923 | * | |
2924 | * This goes through a pool's idle workers and gets rid of those that have been | |
2925 | * idle for at least IDLE_WORKER_TIMEOUT seconds. | |
e02b9312 VS |
2926 | * |
2927 | * We don't want to disturb isolated CPUs because of a pcpu kworker being | |
2928 | * culled, so this also resets worker affinity. This requires a sleepable | |
2929 | * context, hence the split between timer callback and work item. | |
3f959aa3 VS |
2930 | */ |
2931 | static void idle_cull_fn(struct work_struct *work) | |
2932 | { | |
2933 | struct worker_pool *pool = container_of(work, struct worker_pool, idle_cull_work); | |
9680540c | 2934 | LIST_HEAD(cull_list); |
3f959aa3 | 2935 | |
e02b9312 VS |
2936 | /* |
2937 | * Grabbing wq_pool_attach_mutex here ensures an already-running worker | |
2938 | * cannot proceed beyong worker_detach_from_pool() in its self-destruct | |
2939 | * path. This is required as a previously-preempted worker could run after | |
2940 | * set_worker_dying() has happened but before wake_dying_workers() did. | |
2941 | */ | |
2942 | mutex_lock(&wq_pool_attach_mutex); | |
3f959aa3 VS |
2943 | raw_spin_lock_irq(&pool->lock); |
2944 | ||
2945 | while (too_many_workers(pool)) { | |
2946 | struct worker *worker; | |
2947 | unsigned long expires; | |
2948 | ||
d70f5d57 | 2949 | worker = list_last_entry(&pool->idle_list, struct worker, entry); |
e22bee78 TH |
2950 | expires = worker->last_active + IDLE_WORKER_TIMEOUT; |
2951 | ||
3347fc9f | 2952 | if (time_before(jiffies, expires)) { |
63d95a91 | 2953 | mod_timer(&pool->idle_timer, expires); |
3347fc9f | 2954 | break; |
d5abe669 | 2955 | } |
3347fc9f | 2956 | |
e02b9312 | 2957 | set_worker_dying(worker, &cull_list); |
e22bee78 TH |
2958 | } |
2959 | ||
a9b8a985 | 2960 | raw_spin_unlock_irq(&pool->lock); |
e02b9312 VS |
2961 | wake_dying_workers(&cull_list); |
2962 | mutex_unlock(&wq_pool_attach_mutex); | |
e22bee78 | 2963 | } |
d5abe669 | 2964 | |
493a1724 | 2965 | static void send_mayday(struct work_struct *work) |
e22bee78 | 2966 | { |
112202d9 TH |
2967 | struct pool_workqueue *pwq = get_work_pwq(work); |
2968 | struct workqueue_struct *wq = pwq->wq; | |
493a1724 | 2969 | |
2e109a28 | 2970 | lockdep_assert_held(&wq_mayday_lock); |
e22bee78 | 2971 | |
493008a8 | 2972 | if (!wq->rescuer) |
493a1724 | 2973 | return; |
e22bee78 TH |
2974 | |
2975 | /* mayday mayday mayday */ | |
493a1724 | 2976 | if (list_empty(&pwq->mayday_node)) { |
77668c8b LJ |
2977 | /* |
2978 | * If @pwq is for an unbound wq, its base ref may be put at | |
2979 | * any time due to an attribute change. Pin @pwq until the | |
2980 | * rescuer is done with it. | |
2981 | */ | |
2982 | get_pwq(pwq); | |
493a1724 | 2983 | list_add_tail(&pwq->mayday_node, &wq->maydays); |
e22bee78 | 2984 | wake_up_process(wq->rescuer->task); |
725e8ec5 | 2985 | pwq->stats[PWQ_STAT_MAYDAY]++; |
493a1724 | 2986 | } |
e22bee78 TH |
2987 | } |
2988 | ||
32a6c723 | 2989 | static void pool_mayday_timeout(struct timer_list *t) |
e22bee78 | 2990 | { |
32a6c723 | 2991 | struct worker_pool *pool = from_timer(pool, t, mayday_timer); |
e22bee78 TH |
2992 | struct work_struct *work; |
2993 | ||
a9b8a985 SAS |
2994 | raw_spin_lock_irq(&pool->lock); |
2995 | raw_spin_lock(&wq_mayday_lock); /* for wq->maydays */ | |
e22bee78 | 2996 | |
63d95a91 | 2997 | if (need_to_create_worker(pool)) { |
e22bee78 TH |
2998 | /* |
2999 | * We've been trying to create a new worker but | |
3000 | * haven't been successful. We might be hitting an | |
3001 | * allocation deadlock. Send distress signals to | |
3002 | * rescuers. | |
3003 | */ | |
63d95a91 | 3004 | list_for_each_entry(work, &pool->worklist, entry) |
e22bee78 | 3005 | send_mayday(work); |
1da177e4 | 3006 | } |
e22bee78 | 3007 | |
a9b8a985 SAS |
3008 | raw_spin_unlock(&wq_mayday_lock); |
3009 | raw_spin_unlock_irq(&pool->lock); | |
e22bee78 | 3010 | |
63d95a91 | 3011 | mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); |
1da177e4 LT |
3012 | } |
3013 | ||
e22bee78 TH |
3014 | /** |
3015 | * maybe_create_worker - create a new worker if necessary | |
63d95a91 | 3016 | * @pool: pool to create a new worker for |
e22bee78 | 3017 | * |
63d95a91 | 3018 | * Create a new worker for @pool if necessary. @pool is guaranteed to |
e22bee78 TH |
3019 | * have at least one idle worker on return from this function. If |
3020 | * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is | |
63d95a91 | 3021 | * sent to all rescuers with works scheduled on @pool to resolve |
e22bee78 TH |
3022 | * possible allocation deadlock. |
3023 | * | |
c5aa87bb TH |
3024 | * On return, need_to_create_worker() is guaranteed to be %false and |
3025 | * may_start_working() %true. | |
e22bee78 TH |
3026 | * |
3027 | * LOCKING: | |
a9b8a985 | 3028 | * raw_spin_lock_irq(pool->lock) which may be released and regrabbed |
e22bee78 TH |
3029 | * multiple times. Does GFP_KERNEL allocations. Called only from |
3030 | * manager. | |
e22bee78 | 3031 | */ |
29187a9e | 3032 | static void maybe_create_worker(struct worker_pool *pool) |
d565ed63 TH |
3033 | __releases(&pool->lock) |
3034 | __acquires(&pool->lock) | |
1da177e4 | 3035 | { |
e22bee78 | 3036 | restart: |
a9b8a985 | 3037 | raw_spin_unlock_irq(&pool->lock); |
9f9c2364 | 3038 | |
e22bee78 | 3039 | /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ |
63d95a91 | 3040 | mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); |
e22bee78 TH |
3041 | |
3042 | while (true) { | |
051e1850 | 3043 | if (create_worker(pool) || !need_to_create_worker(pool)) |
e22bee78 | 3044 | break; |
1da177e4 | 3045 | |
e212f361 | 3046 | schedule_timeout_interruptible(CREATE_COOLDOWN); |
9f9c2364 | 3047 | |
63d95a91 | 3048 | if (!need_to_create_worker(pool)) |
e22bee78 TH |
3049 | break; |
3050 | } | |
3051 | ||
63d95a91 | 3052 | del_timer_sync(&pool->mayday_timer); |
a9b8a985 | 3053 | raw_spin_lock_irq(&pool->lock); |
051e1850 LJ |
3054 | /* |
3055 | * This is necessary even after a new worker was just successfully | |
3056 | * created as @pool->lock was dropped and the new worker might have | |
3057 | * already become busy. | |
3058 | */ | |
63d95a91 | 3059 | if (need_to_create_worker(pool)) |
e22bee78 | 3060 | goto restart; |
e22bee78 TH |
3061 | } |
3062 | ||
73f53c4a | 3063 | /** |
e22bee78 TH |
3064 | * manage_workers - manage worker pool |
3065 | * @worker: self | |
73f53c4a | 3066 | * |
706026c2 | 3067 | * Assume the manager role and manage the worker pool @worker belongs |
e22bee78 | 3068 | * to. At any given time, there can be only zero or one manager per |
706026c2 | 3069 | * pool. The exclusion is handled automatically by this function. |
e22bee78 TH |
3070 | * |
3071 | * The caller can safely start processing works on false return. On | |
3072 | * true return, it's guaranteed that need_to_create_worker() is false | |
3073 | * and may_start_working() is true. | |
73f53c4a TH |
3074 | * |
3075 | * CONTEXT: | |
a9b8a985 | 3076 | * raw_spin_lock_irq(pool->lock) which may be released and regrabbed |
e22bee78 TH |
3077 | * multiple times. Does GFP_KERNEL allocations. |
3078 | * | |
d185af30 | 3079 | * Return: |
29187a9e TH |
3080 | * %false if the pool doesn't need management and the caller can safely |
3081 | * start processing works, %true if management function was performed and | |
3082 | * the conditions that the caller verified before calling the function may | |
3083 | * no longer be true. | |
73f53c4a | 3084 | */ |
e22bee78 | 3085 | static bool manage_workers(struct worker *worker) |
73f53c4a | 3086 | { |
63d95a91 | 3087 | struct worker_pool *pool = worker->pool; |
73f53c4a | 3088 | |
692b4825 | 3089 | if (pool->flags & POOL_MANAGER_ACTIVE) |
29187a9e | 3090 | return false; |
692b4825 TH |
3091 | |
3092 | pool->flags |= POOL_MANAGER_ACTIVE; | |
2607d7a6 | 3093 | pool->manager = worker; |
1e19ffc6 | 3094 | |
29187a9e | 3095 | maybe_create_worker(pool); |
e22bee78 | 3096 | |
2607d7a6 | 3097 | pool->manager = NULL; |
692b4825 | 3098 | pool->flags &= ~POOL_MANAGER_ACTIVE; |
d8bb65ab | 3099 | rcuwait_wake_up(&manager_wait); |
29187a9e | 3100 | return true; |
73f53c4a TH |
3101 | } |
3102 | ||
a62428c0 TH |
3103 | /** |
3104 | * process_one_work - process single work | |
c34056a3 | 3105 | * @worker: self |
a62428c0 TH |
3106 | * @work: work to process |
3107 | * | |
3108 | * Process @work. This function contains all the logics necessary to | |
3109 | * process a single work including synchronization against and | |
3110 | * interaction with other workers on the same cpu, queueing and | |
3111 | * flushing. As long as context requirement is met, any worker can | |
3112 | * call this function to process a work. | |
3113 | * | |
3114 | * CONTEXT: | |
a9b8a985 | 3115 | * raw_spin_lock_irq(pool->lock) which is released and regrabbed. |
a62428c0 | 3116 | */ |
c34056a3 | 3117 | static void process_one_work(struct worker *worker, struct work_struct *work) |
d565ed63 TH |
3118 | __releases(&pool->lock) |
3119 | __acquires(&pool->lock) | |
a62428c0 | 3120 | { |
112202d9 | 3121 | struct pool_workqueue *pwq = get_work_pwq(work); |
bd7bdd43 | 3122 | struct worker_pool *pool = worker->pool; |
c4560c2c | 3123 | unsigned long work_data; |
c35aea39 | 3124 | int lockdep_start_depth, rcu_start_depth; |
1acd92d9 | 3125 | bool bh_draining = pool->flags & POOL_BH_DRAINING; |
a62428c0 TH |
3126 | #ifdef CONFIG_LOCKDEP |
3127 | /* | |
3128 | * It is permissible to free the struct work_struct from | |
3129 | * inside the function that is called from it, this we need to | |
3130 | * take into account for lockdep too. To avoid bogus "held | |
3131 | * lock freed" warnings as well as problems when looking into | |
3132 | * work->lockdep_map, make a copy and use that here. | |
3133 | */ | |
4d82a1de PZ |
3134 | struct lockdep_map lockdep_map; |
3135 | ||
3136 | lockdep_copy_map(&lockdep_map, &work->lockdep_map); | |
a62428c0 | 3137 | #endif |
807407c0 | 3138 | /* ensure we're on the correct CPU */ |
85327af6 | 3139 | WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && |
ec22ca5e | 3140 | raw_smp_processor_id() != pool->cpu); |
25511a47 | 3141 | |
8930caba | 3142 | /* claim and dequeue */ |
a62428c0 | 3143 | debug_work_deactivate(work); |
c9e7cf27 | 3144 | hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); |
c34056a3 | 3145 | worker->current_work = work; |
a2c1c57b | 3146 | worker->current_func = work->func; |
112202d9 | 3147 | worker->current_pwq = pwq; |
4cb1ef64 TH |
3148 | if (worker->task) |
3149 | worker->current_at = worker->task->se.sum_exec_runtime; | |
c4560c2c | 3150 | work_data = *work_data_bits(work); |
d812796e | 3151 | worker->current_color = get_work_color(work_data); |
7a22ad75 | 3152 | |
8bf89593 TH |
3153 | /* |
3154 | * Record wq name for cmdline and debug reporting, may get | |
3155 | * overridden through set_worker_desc(). | |
3156 | */ | |
3157 | strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN); | |
3158 | ||
a62428c0 TH |
3159 | list_del_init(&work->entry); |
3160 | ||
fb0e7beb | 3161 | /* |
228f1d00 LJ |
3162 | * CPU intensive works don't participate in concurrency management. |
3163 | * They're the scheduler's responsibility. This takes @worker out | |
3164 | * of concurrency management and the next code block will chain | |
3165 | * execution of the pending work items. | |
fb0e7beb | 3166 | */ |
616db877 | 3167 | if (unlikely(pwq->wq->flags & WQ_CPU_INTENSIVE)) |
228f1d00 | 3168 | worker_set_flags(worker, WORKER_CPU_INTENSIVE); |
fb0e7beb | 3169 | |
974271c4 | 3170 | /* |
0219a352 TH |
3171 | * Kick @pool if necessary. It's always noop for per-cpu worker pools |
3172 | * since nr_running would always be >= 1 at this point. This is used to | |
3173 | * chain execution of the pending work items for WORKER_NOT_RUNNING | |
3174 | * workers such as the UNBOUND and CPU_INTENSIVE ones. | |
974271c4 | 3175 | */ |
0219a352 | 3176 | kick_pool(pool); |
974271c4 | 3177 | |
8930caba | 3178 | /* |
7c3eed5c | 3179 | * Record the last pool and clear PENDING which should be the last |
d565ed63 | 3180 | * update to @work. Also, do this inside @pool->lock so that |
23657bb1 TH |
3181 | * PENDING and queued state changes happen together while IRQ is |
3182 | * disabled. | |
8930caba | 3183 | */ |
456a78ee | 3184 | set_work_pool_and_clear_pending(work, pool->id, pool_offq_flags(pool)); |
a62428c0 | 3185 | |
fe48ba7d | 3186 | pwq->stats[PWQ_STAT_STARTED]++; |
a9b8a985 | 3187 | raw_spin_unlock_irq(&pool->lock); |
a62428c0 | 3188 | |
c35aea39 TH |
3189 | rcu_start_depth = rcu_preempt_depth(); |
3190 | lockdep_start_depth = lockdep_depth(current); | |
1acd92d9 TH |
3191 | /* see drain_dead_softirq_workfn() */ |
3192 | if (!bh_draining) | |
3193 | lock_map_acquire(&pwq->wq->lockdep_map); | |
a62428c0 | 3194 | lock_map_acquire(&lockdep_map); |
e6f3faa7 | 3195 | /* |
f52be570 PZ |
3196 | * Strictly speaking we should mark the invariant state without holding |
3197 | * any locks, that is, before these two lock_map_acquire()'s. | |
e6f3faa7 PZ |
3198 | * |
3199 | * However, that would result in: | |
3200 | * | |
3201 | * A(W1) | |
3202 | * WFC(C) | |
3203 | * A(W1) | |
3204 | * C(C) | |
3205 | * | |
3206 | * Which would create W1->C->W1 dependencies, even though there is no | |
3207 | * actual deadlock possible. There are two solutions, using a | |
3208 | * read-recursive acquire on the work(queue) 'locks', but this will then | |
f52be570 | 3209 | * hit the lockdep limitation on recursive locks, or simply discard |
e6f3faa7 PZ |
3210 | * these locks. |
3211 | * | |
3212 | * AFAICT there is no possible deadlock scenario between the | |
3213 | * flush_work() and complete() primitives (except for single-threaded | |
3214 | * workqueues), so hiding them isn't a problem. | |
3215 | */ | |
f52be570 | 3216 | lockdep_invariant_state(true); |
e36c886a | 3217 | trace_workqueue_execute_start(work); |
a2c1c57b | 3218 | worker->current_func(work); |
e36c886a AV |
3219 | /* |
3220 | * While we must be careful to not use "work" after this, the trace | |
3221 | * point will only record its address. | |
3222 | */ | |
1c5da0ec | 3223 | trace_workqueue_execute_end(work, worker->current_func); |
725e8ec5 | 3224 | pwq->stats[PWQ_STAT_COMPLETED]++; |
a62428c0 | 3225 | lock_map_release(&lockdep_map); |
1acd92d9 TH |
3226 | if (!bh_draining) |
3227 | lock_map_release(&pwq->wq->lockdep_map); | |
a62428c0 | 3228 | |
c35aea39 TH |
3229 | if (unlikely((worker->task && in_atomic()) || |
3230 | lockdep_depth(current) != lockdep_start_depth || | |
3231 | rcu_preempt_depth() != rcu_start_depth)) { | |
3232 | pr_err("BUG: workqueue leaked atomic, lock or RCU: %s[%d]\n" | |
3233 | " preempt=0x%08x lock=%d->%d RCU=%d->%d workfn=%ps\n", | |
3234 | current->comm, task_pid_nr(current), preempt_count(), | |
3235 | lockdep_start_depth, lockdep_depth(current), | |
3236 | rcu_start_depth, rcu_preempt_depth(), | |
3237 | worker->current_func); | |
a62428c0 TH |
3238 | debug_show_held_locks(current); |
3239 | dump_stack(); | |
3240 | } | |
3241 | ||
b22ce278 | 3242 | /* |
025f50f3 | 3243 | * The following prevents a kworker from hogging CPU on !PREEMPTION |
b22ce278 TH |
3244 | * kernels, where a requeueing work item waiting for something to |
3245 | * happen could deadlock with stop_machine as such work item could | |
3246 | * indefinitely requeue itself while all other CPUs are trapped in | |
789cbbec JL |
3247 | * stop_machine. At the same time, report a quiescent RCU state so |
3248 | * the same condition doesn't freeze RCU. | |
b22ce278 | 3249 | */ |
4cb1ef64 TH |
3250 | if (worker->task) |
3251 | cond_resched(); | |
b22ce278 | 3252 | |
a9b8a985 | 3253 | raw_spin_lock_irq(&pool->lock); |
a62428c0 | 3254 | |
616db877 TH |
3255 | /* |
3256 | * In addition to %WQ_CPU_INTENSIVE, @worker may also have been marked | |
3257 | * CPU intensive by wq_worker_tick() if @work hogged CPU longer than | |
3258 | * wq_cpu_intensive_thresh_us. Clear it. | |
3259 | */ | |
3260 | worker_clr_flags(worker, WORKER_CPU_INTENSIVE); | |
fb0e7beb | 3261 | |
1b69ac6b JW |
3262 | /* tag the worker for identification in schedule() */ |
3263 | worker->last_func = worker->current_func; | |
3264 | ||
a62428c0 | 3265 | /* we're done with it, release */ |
42f8570f | 3266 | hash_del(&worker->hentry); |
c34056a3 | 3267 | worker->current_work = NULL; |
a2c1c57b | 3268 | worker->current_func = NULL; |
112202d9 | 3269 | worker->current_pwq = NULL; |
d812796e | 3270 | worker->current_color = INT_MAX; |
dd6c3c54 TH |
3271 | |
3272 | /* must be the last step, see the function comment */ | |
c4560c2c | 3273 | pwq_dec_nr_in_flight(pwq, work_data); |
a62428c0 TH |
3274 | } |
3275 | ||
affee4b2 TH |
3276 | /** |
3277 | * process_scheduled_works - process scheduled works | |
3278 | * @worker: self | |
3279 | * | |
3280 | * Process all scheduled works. Please note that the scheduled list | |
3281 | * may change while processing a work, so this function repeatedly | |
3282 | * fetches a work from the top and executes it. | |
3283 | * | |
3284 | * CONTEXT: | |
a9b8a985 | 3285 | * raw_spin_lock_irq(pool->lock) which may be released and regrabbed |
affee4b2 TH |
3286 | * multiple times. |
3287 | */ | |
3288 | static void process_scheduled_works(struct worker *worker) | |
1da177e4 | 3289 | { |
c0ab017d TH |
3290 | struct work_struct *work; |
3291 | bool first = true; | |
3292 | ||
3293 | while ((work = list_first_entry_or_null(&worker->scheduled, | |
3294 | struct work_struct, entry))) { | |
3295 | if (first) { | |
3296 | worker->pool->watchdog_ts = jiffies; | |
3297 | first = false; | |
3298 | } | |
c34056a3 | 3299 | process_one_work(worker, work); |
1da177e4 | 3300 | } |
1da177e4 LT |
3301 | } |
3302 | ||
197f6acc TH |
3303 | static void set_pf_worker(bool val) |
3304 | { | |
3305 | mutex_lock(&wq_pool_attach_mutex); | |
3306 | if (val) | |
3307 | current->flags |= PF_WQ_WORKER; | |
3308 | else | |
3309 | current->flags &= ~PF_WQ_WORKER; | |
3310 | mutex_unlock(&wq_pool_attach_mutex); | |
3311 | } | |
3312 | ||
4690c4ab TH |
3313 | /** |
3314 | * worker_thread - the worker thread function | |
c34056a3 | 3315 | * @__worker: self |
4690c4ab | 3316 | * |
c5aa87bb TH |
3317 | * The worker thread function. All workers belong to a worker_pool - |
3318 | * either a per-cpu one or dynamic unbound one. These workers process all | |
3319 | * work items regardless of their specific target workqueue. The only | |
3320 | * exception is work items which belong to workqueues with a rescuer which | |
3321 | * will be explained in rescuer_thread(). | |
d185af30 YB |
3322 | * |
3323 | * Return: 0 | |
4690c4ab | 3324 | */ |
c34056a3 | 3325 | static int worker_thread(void *__worker) |
1da177e4 | 3326 | { |
c34056a3 | 3327 | struct worker *worker = __worker; |
bd7bdd43 | 3328 | struct worker_pool *pool = worker->pool; |
1da177e4 | 3329 | |
e22bee78 | 3330 | /* tell the scheduler that this is a workqueue worker */ |
197f6acc | 3331 | set_pf_worker(true); |
c8e55f36 | 3332 | woke_up: |
a9b8a985 | 3333 | raw_spin_lock_irq(&pool->lock); |
1da177e4 | 3334 | |
a9ab775b TH |
3335 | /* am I supposed to die? */ |
3336 | if (unlikely(worker->flags & WORKER_DIE)) { | |
a9b8a985 | 3337 | raw_spin_unlock_irq(&pool->lock); |
197f6acc | 3338 | set_pf_worker(false); |
60f5a4bc LJ |
3339 | |
3340 | set_task_comm(worker->task, "kworker/dying"); | |
e441b56f | 3341 | ida_free(&pool->worker_ida, worker->id); |
a2d812a2 | 3342 | worker_detach_from_pool(worker); |
e02b9312 | 3343 | WARN_ON_ONCE(!list_empty(&worker->entry)); |
60f5a4bc | 3344 | kfree(worker); |
a9ab775b | 3345 | return 0; |
c8e55f36 | 3346 | } |
affee4b2 | 3347 | |
c8e55f36 | 3348 | worker_leave_idle(worker); |
db7bccf4 | 3349 | recheck: |
e22bee78 | 3350 | /* no more worker necessary? */ |
63d95a91 | 3351 | if (!need_more_worker(pool)) |
e22bee78 TH |
3352 | goto sleep; |
3353 | ||
3354 | /* do we need to manage? */ | |
63d95a91 | 3355 | if (unlikely(!may_start_working(pool)) && manage_workers(worker)) |
e22bee78 TH |
3356 | goto recheck; |
3357 | ||
c8e55f36 TH |
3358 | /* |
3359 | * ->scheduled list can only be filled while a worker is | |
3360 | * preparing to process a work or actually processing it. | |
3361 | * Make sure nobody diddled with it while I was sleeping. | |
3362 | */ | |
6183c009 | 3363 | WARN_ON_ONCE(!list_empty(&worker->scheduled)); |
c8e55f36 | 3364 | |
e22bee78 | 3365 | /* |
a9ab775b TH |
3366 | * Finish PREP stage. We're guaranteed to have at least one idle |
3367 | * worker or that someone else has already assumed the manager | |
3368 | * role. This is where @worker starts participating in concurrency | |
3369 | * management if applicable and concurrency management is restored | |
3370 | * after being rebound. See rebind_workers() for details. | |
e22bee78 | 3371 | */ |
a9ab775b | 3372 | worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); |
e22bee78 TH |
3373 | |
3374 | do { | |
c8e55f36 | 3375 | struct work_struct *work = |
bd7bdd43 | 3376 | list_first_entry(&pool->worklist, |
c8e55f36 TH |
3377 | struct work_struct, entry); |
3378 | ||
873eaca6 TH |
3379 | if (assign_work(work, worker, NULL)) |
3380 | process_scheduled_works(worker); | |
63d95a91 | 3381 | } while (keep_working(pool)); |
e22bee78 | 3382 | |
228f1d00 | 3383 | worker_set_flags(worker, WORKER_PREP); |
d313dd85 | 3384 | sleep: |
c8e55f36 | 3385 | /* |
d565ed63 TH |
3386 | * pool->lock is held and there's no work to process and no need to |
3387 | * manage, sleep. Workers are woken up only while holding | |
3388 | * pool->lock or from local cpu, so setting the current state | |
3389 | * before releasing pool->lock is enough to prevent losing any | |
3390 | * event. | |
c8e55f36 TH |
3391 | */ |
3392 | worker_enter_idle(worker); | |
c5a94a61 | 3393 | __set_current_state(TASK_IDLE); |
a9b8a985 | 3394 | raw_spin_unlock_irq(&pool->lock); |
c8e55f36 TH |
3395 | schedule(); |
3396 | goto woke_up; | |
1da177e4 LT |
3397 | } |
3398 | ||
e22bee78 TH |
3399 | /** |
3400 | * rescuer_thread - the rescuer thread function | |
111c225a | 3401 | * @__rescuer: self |
e22bee78 TH |
3402 | * |
3403 | * Workqueue rescuer thread function. There's one rescuer for each | |
493008a8 | 3404 | * workqueue which has WQ_MEM_RECLAIM set. |
e22bee78 | 3405 | * |
706026c2 | 3406 | * Regular work processing on a pool may block trying to create a new |
e22bee78 TH |
3407 | * worker which uses GFP_KERNEL allocation which has slight chance of |
3408 | * developing into deadlock if some works currently on the same queue | |
3409 | * need to be processed to satisfy the GFP_KERNEL allocation. This is | |
3410 | * the problem rescuer solves. | |
3411 | * | |
706026c2 TH |
3412 | * When such condition is possible, the pool summons rescuers of all |
3413 | * workqueues which have works queued on the pool and let them process | |
e22bee78 TH |
3414 | * those works so that forward progress can be guaranteed. |
3415 | * | |
3416 | * This should happen rarely. | |
d185af30 YB |
3417 | * |
3418 | * Return: 0 | |
e22bee78 | 3419 | */ |
111c225a | 3420 | static int rescuer_thread(void *__rescuer) |
e22bee78 | 3421 | { |
111c225a TH |
3422 | struct worker *rescuer = __rescuer; |
3423 | struct workqueue_struct *wq = rescuer->rescue_wq; | |
4d595b86 | 3424 | bool should_stop; |
e22bee78 TH |
3425 | |
3426 | set_user_nice(current, RESCUER_NICE_LEVEL); | |
111c225a TH |
3427 | |
3428 | /* | |
3429 | * Mark rescuer as worker too. As WORKER_PREP is never cleared, it | |
3430 | * doesn't participate in concurrency management. | |
3431 | */ | |
197f6acc | 3432 | set_pf_worker(true); |
e22bee78 | 3433 | repeat: |
c5a94a61 | 3434 | set_current_state(TASK_IDLE); |
e22bee78 | 3435 | |
4d595b86 LJ |
3436 | /* |
3437 | * By the time the rescuer is requested to stop, the workqueue | |
3438 | * shouldn't have any work pending, but @wq->maydays may still have | |
3439 | * pwq(s) queued. This can happen by non-rescuer workers consuming | |
3440 | * all the work items before the rescuer got to them. Go through | |
3441 | * @wq->maydays processing before acting on should_stop so that the | |
3442 | * list is always empty on exit. | |
3443 | */ | |
3444 | should_stop = kthread_should_stop(); | |
e22bee78 | 3445 | |
493a1724 | 3446 | /* see whether any pwq is asking for help */ |
a9b8a985 | 3447 | raw_spin_lock_irq(&wq_mayday_lock); |
493a1724 TH |
3448 | |
3449 | while (!list_empty(&wq->maydays)) { | |
3450 | struct pool_workqueue *pwq = list_first_entry(&wq->maydays, | |
3451 | struct pool_workqueue, mayday_node); | |
112202d9 | 3452 | struct worker_pool *pool = pwq->pool; |
e22bee78 TH |
3453 | struct work_struct *work, *n; |
3454 | ||
3455 | __set_current_state(TASK_RUNNING); | |
493a1724 TH |
3456 | list_del_init(&pwq->mayday_node); |
3457 | ||
a9b8a985 | 3458 | raw_spin_unlock_irq(&wq_mayday_lock); |
e22bee78 | 3459 | |
51697d39 LJ |
3460 | worker_attach_to_pool(rescuer, pool); |
3461 | ||
a9b8a985 | 3462 | raw_spin_lock_irq(&pool->lock); |
e22bee78 TH |
3463 | |
3464 | /* | |
3465 | * Slurp in all works issued via this workqueue and | |
3466 | * process'em. | |
3467 | */ | |
873eaca6 | 3468 | WARN_ON_ONCE(!list_empty(&rescuer->scheduled)); |
82607adc | 3469 | list_for_each_entry_safe(work, n, &pool->worklist, entry) { |
873eaca6 TH |
3470 | if (get_work_pwq(work) == pwq && |
3471 | assign_work(work, rescuer, &n)) | |
725e8ec5 | 3472 | pwq->stats[PWQ_STAT_RESCUED]++; |
82607adc | 3473 | } |
e22bee78 | 3474 | |
873eaca6 | 3475 | if (!list_empty(&rescuer->scheduled)) { |
008847f6 N |
3476 | process_scheduled_works(rescuer); |
3477 | ||
3478 | /* | |
3479 | * The above execution of rescued work items could | |
3480 | * have created more to rescue through | |
f97a4a1a | 3481 | * pwq_activate_first_inactive() or chained |
008847f6 N |
3482 | * queueing. Let's put @pwq back on mayday list so |
3483 | * that such back-to-back work items, which may be | |
3484 | * being used to relieve memory pressure, don't | |
3485 | * incur MAYDAY_INTERVAL delay inbetween. | |
3486 | */ | |
4f3f4cf3 | 3487 | if (pwq->nr_active && need_to_create_worker(pool)) { |
a9b8a985 | 3488 | raw_spin_lock(&wq_mayday_lock); |
e66b39af TH |
3489 | /* |
3490 | * Queue iff we aren't racing destruction | |
3491 | * and somebody else hasn't queued it already. | |
3492 | */ | |
3493 | if (wq->rescuer && list_empty(&pwq->mayday_node)) { | |
3494 | get_pwq(pwq); | |
3495 | list_add_tail(&pwq->mayday_node, &wq->maydays); | |
3496 | } | |
a9b8a985 | 3497 | raw_spin_unlock(&wq_mayday_lock); |
008847f6 N |
3498 | } |
3499 | } | |
7576958a | 3500 | |
77668c8b LJ |
3501 | /* |
3502 | * Put the reference grabbed by send_mayday(). @pool won't | |
13b1d625 | 3503 | * go away while we're still attached to it. |
77668c8b LJ |
3504 | */ |
3505 | put_pwq(pwq); | |
3506 | ||
7576958a | 3507 | /* |
0219a352 TH |
3508 | * Leave this pool. Notify regular workers; otherwise, we end up |
3509 | * with 0 concurrency and stalling the execution. | |
7576958a | 3510 | */ |
0219a352 | 3511 | kick_pool(pool); |
7576958a | 3512 | |
a9b8a985 | 3513 | raw_spin_unlock_irq(&pool->lock); |
13b1d625 | 3514 | |
a2d812a2 | 3515 | worker_detach_from_pool(rescuer); |
13b1d625 | 3516 | |
a9b8a985 | 3517 | raw_spin_lock_irq(&wq_mayday_lock); |
e22bee78 TH |
3518 | } |
3519 | ||
a9b8a985 | 3520 | raw_spin_unlock_irq(&wq_mayday_lock); |
493a1724 | 3521 | |
4d595b86 LJ |
3522 | if (should_stop) { |
3523 | __set_current_state(TASK_RUNNING); | |
197f6acc | 3524 | set_pf_worker(false); |
4d595b86 LJ |
3525 | return 0; |
3526 | } | |
3527 | ||
111c225a TH |
3528 | /* rescuers should never participate in concurrency management */ |
3529 | WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); | |
e22bee78 TH |
3530 | schedule(); |
3531 | goto repeat; | |
1da177e4 LT |
3532 | } |
3533 | ||
4cb1ef64 TH |
3534 | static void bh_worker(struct worker *worker) |
3535 | { | |
3536 | struct worker_pool *pool = worker->pool; | |
3537 | int nr_restarts = BH_WORKER_RESTARTS; | |
3538 | unsigned long end = jiffies + BH_WORKER_JIFFIES; | |
3539 | ||
3540 | raw_spin_lock_irq(&pool->lock); | |
3541 | worker_leave_idle(worker); | |
3542 | ||
3543 | /* | |
3544 | * This function follows the structure of worker_thread(). See there for | |
3545 | * explanations on each step. | |
3546 | */ | |
3547 | if (!need_more_worker(pool)) | |
3548 | goto done; | |
3549 | ||
3550 | WARN_ON_ONCE(!list_empty(&worker->scheduled)); | |
3551 | worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); | |
3552 | ||
3553 | do { | |
3554 | struct work_struct *work = | |
3555 | list_first_entry(&pool->worklist, | |
3556 | struct work_struct, entry); | |
3557 | ||
3558 | if (assign_work(work, worker, NULL)) | |
3559 | process_scheduled_works(worker); | |
3560 | } while (keep_working(pool) && | |
3561 | --nr_restarts && time_before(jiffies, end)); | |
3562 | ||
3563 | worker_set_flags(worker, WORKER_PREP); | |
3564 | done: | |
3565 | worker_enter_idle(worker); | |
3566 | kick_pool(pool); | |
3567 | raw_spin_unlock_irq(&pool->lock); | |
3568 | } | |
3569 | ||
3570 | /* | |
3571 | * TODO: Convert all tasklet users to workqueue and use softirq directly. | |
3572 | * | |
3573 | * This is currently called from tasklet[_hi]action() and thus is also called | |
3574 | * whenever there are tasklets to run. Let's do an early exit if there's nothing | |
3575 | * queued. Once conversion from tasklet is complete, the need_more_worker() test | |
3576 | * can be dropped. | |
3577 | * | |
3578 | * After full conversion, we'll add worker->softirq_action, directly use the | |
3579 | * softirq action and obtain the worker pointer from the softirq_action pointer. | |
3580 | */ | |
3581 | void workqueue_softirq_action(bool highpri) | |
3582 | { | |
3583 | struct worker_pool *pool = | |
3584 | &per_cpu(bh_worker_pools, smp_processor_id())[highpri]; | |
3585 | if (need_more_worker(pool)) | |
3586 | bh_worker(list_first_entry(&pool->workers, struct worker, node)); | |
3587 | } | |
3588 | ||
1acd92d9 TH |
3589 | struct wq_drain_dead_softirq_work { |
3590 | struct work_struct work; | |
3591 | struct worker_pool *pool; | |
3592 | struct completion done; | |
3593 | }; | |
3594 | ||
3595 | static void drain_dead_softirq_workfn(struct work_struct *work) | |
3596 | { | |
3597 | struct wq_drain_dead_softirq_work *dead_work = | |
3598 | container_of(work, struct wq_drain_dead_softirq_work, work); | |
3599 | struct worker_pool *pool = dead_work->pool; | |
3600 | bool repeat; | |
3601 | ||
3602 | /* | |
3603 | * @pool's CPU is dead and we want to execute its still pending work | |
3604 | * items from this BH work item which is running on a different CPU. As | |
3605 | * its CPU is dead, @pool can't be kicked and, as work execution path | |
3606 | * will be nested, a lockdep annotation needs to be suppressed. Mark | |
3607 | * @pool with %POOL_BH_DRAINING for the special treatments. | |
3608 | */ | |
3609 | raw_spin_lock_irq(&pool->lock); | |
3610 | pool->flags |= POOL_BH_DRAINING; | |
3611 | raw_spin_unlock_irq(&pool->lock); | |
3612 | ||
3613 | bh_worker(list_first_entry(&pool->workers, struct worker, node)); | |
3614 | ||
3615 | raw_spin_lock_irq(&pool->lock); | |
3616 | pool->flags &= ~POOL_BH_DRAINING; | |
3617 | repeat = need_more_worker(pool); | |
3618 | raw_spin_unlock_irq(&pool->lock); | |
3619 | ||
3620 | /* | |
3621 | * bh_worker() might hit consecutive execution limit and bail. If there | |
3622 | * still are pending work items, reschedule self and return so that we | |
3623 | * don't hog this CPU's BH. | |
3624 | */ | |
3625 | if (repeat) { | |
3626 | if (pool->attrs->nice == HIGHPRI_NICE_LEVEL) | |
3627 | queue_work(system_bh_highpri_wq, work); | |
3628 | else | |
3629 | queue_work(system_bh_wq, work); | |
3630 | } else { | |
3631 | complete(&dead_work->done); | |
3632 | } | |
3633 | } | |
3634 | ||
3635 | /* | |
3636 | * @cpu is dead. Drain the remaining BH work items on the current CPU. It's | |
3637 | * possible to allocate dead_work per CPU and avoid flushing. However, then we | |
3638 | * have to worry about draining overlapping with CPU coming back online or | |
3639 | * nesting (one CPU's dead_work queued on another CPU which is also dead and so | |
3640 | * on). Let's keep it simple and drain them synchronously. These are BH work | |
3641 | * items which shouldn't be requeued on the same pool. Shouldn't take long. | |
3642 | */ | |
3643 | void workqueue_softirq_dead(unsigned int cpu) | |
3644 | { | |
3645 | int i; | |
3646 | ||
3647 | for (i = 0; i < NR_STD_WORKER_POOLS; i++) { | |
3648 | struct worker_pool *pool = &per_cpu(bh_worker_pools, cpu)[i]; | |
3649 | struct wq_drain_dead_softirq_work dead_work; | |
3650 | ||
3651 | if (!need_more_worker(pool)) | |
3652 | continue; | |
3653 | ||
e7cc3be6 | 3654 | INIT_WORK_ONSTACK(&dead_work.work, drain_dead_softirq_workfn); |
1acd92d9 TH |
3655 | dead_work.pool = pool; |
3656 | init_completion(&dead_work.done); | |
3657 | ||
3658 | if (pool->attrs->nice == HIGHPRI_NICE_LEVEL) | |
3659 | queue_work(system_bh_highpri_wq, &dead_work.work); | |
3660 | else | |
3661 | queue_work(system_bh_wq, &dead_work.work); | |
3662 | ||
3663 | wait_for_completion(&dead_work.done); | |
3664 | } | |
3665 | } | |
3666 | ||
fca839c0 TH |
3667 | /** |
3668 | * check_flush_dependency - check for flush dependency sanity | |
3669 | * @target_wq: workqueue being flushed | |
3670 | * @target_work: work item being flushed (NULL for workqueue flushes) | |
3671 | * | |
3672 | * %current is trying to flush the whole @target_wq or @target_work on it. | |
3673 | * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not | |
3674 | * reclaiming memory or running on a workqueue which doesn't have | |
3675 | * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to | |
3676 | * a deadlock. | |
3677 | */ | |
3678 | static void check_flush_dependency(struct workqueue_struct *target_wq, | |
3679 | struct work_struct *target_work) | |
3680 | { | |
3681 | work_func_t target_func = target_work ? target_work->func : NULL; | |
3682 | struct worker *worker; | |
3683 | ||
3684 | if (target_wq->flags & WQ_MEM_RECLAIM) | |
3685 | return; | |
3686 | ||
3687 | worker = current_wq_worker(); | |
3688 | ||
3689 | WARN_ONCE(current->flags & PF_MEMALLOC, | |
d75f773c | 3690 | "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps", |
fca839c0 | 3691 | current->pid, current->comm, target_wq->name, target_func); |
23d11a58 TH |
3692 | WARN_ONCE(worker && ((worker->current_pwq->wq->flags & |
3693 | (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM), | |
d75f773c | 3694 | "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps", |
fca839c0 TH |
3695 | worker->current_pwq->wq->name, worker->current_func, |
3696 | target_wq->name, target_func); | |
3697 | } | |
3698 | ||
fc2e4d70 ON |
3699 | struct wq_barrier { |
3700 | struct work_struct work; | |
3701 | struct completion done; | |
2607d7a6 | 3702 | struct task_struct *task; /* purely informational */ |
fc2e4d70 ON |
3703 | }; |
3704 | ||
3705 | static void wq_barrier_func(struct work_struct *work) | |
3706 | { | |
3707 | struct wq_barrier *barr = container_of(work, struct wq_barrier, work); | |
3708 | complete(&barr->done); | |
3709 | } | |
3710 | ||
4690c4ab TH |
3711 | /** |
3712 | * insert_wq_barrier - insert a barrier work | |
112202d9 | 3713 | * @pwq: pwq to insert barrier into |
4690c4ab | 3714 | * @barr: wq_barrier to insert |
affee4b2 TH |
3715 | * @target: target work to attach @barr to |
3716 | * @worker: worker currently executing @target, NULL if @target is not executing | |
4690c4ab | 3717 | * |
affee4b2 TH |
3718 | * @barr is linked to @target such that @barr is completed only after |
3719 | * @target finishes execution. Please note that the ordering | |
3720 | * guarantee is observed only with respect to @target and on the local | |
3721 | * cpu. | |
3722 | * | |
3723 | * Currently, a queued barrier can't be canceled. This is because | |
3724 | * try_to_grab_pending() can't determine whether the work to be | |
3725 | * grabbed is at the head of the queue and thus can't clear LINKED | |
3726 | * flag of the previous work while there must be a valid next work | |
3727 | * after a work with LINKED flag set. | |
3728 | * | |
3729 | * Note that when @worker is non-NULL, @target may be modified | |
112202d9 | 3730 | * underneath us, so we can't reliably determine pwq from @target. |
4690c4ab TH |
3731 | * |
3732 | * CONTEXT: | |
a9b8a985 | 3733 | * raw_spin_lock_irq(pool->lock). |
4690c4ab | 3734 | */ |
112202d9 | 3735 | static void insert_wq_barrier(struct pool_workqueue *pwq, |
affee4b2 TH |
3736 | struct wq_barrier *barr, |
3737 | struct work_struct *target, struct worker *worker) | |
fc2e4d70 | 3738 | { |
4cb1ef64 | 3739 | static __maybe_unused struct lock_class_key bh_key, thr_key; |
d812796e LJ |
3740 | unsigned int work_flags = 0; |
3741 | unsigned int work_color; | |
affee4b2 | 3742 | struct list_head *head; |
affee4b2 | 3743 | |
dc186ad7 | 3744 | /* |
d565ed63 | 3745 | * debugobject calls are safe here even with pool->lock locked |
dc186ad7 TG |
3746 | * as we know for sure that this will not trigger any of the |
3747 | * checks and call back into the fixup functions where we | |
3748 | * might deadlock. | |
4cb1ef64 TH |
3749 | * |
3750 | * BH and threaded workqueues need separate lockdep keys to avoid | |
3751 | * spuriously triggering "inconsistent {SOFTIRQ-ON-W} -> {IN-SOFTIRQ-W} | |
3752 | * usage". | |
dc186ad7 | 3753 | */ |
4cb1ef64 TH |
3754 | INIT_WORK_ONSTACK_KEY(&barr->work, wq_barrier_func, |
3755 | (pwq->wq->flags & WQ_BH) ? &bh_key : &thr_key); | |
22df02bb | 3756 | __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); |
52fa5bc5 | 3757 | |
fd1a5b04 BP |
3758 | init_completion_map(&barr->done, &target->lockdep_map); |
3759 | ||
2607d7a6 | 3760 | barr->task = current; |
83c22520 | 3761 | |
5797b1c1 | 3762 | /* The barrier work item does not participate in nr_active. */ |
018f3a13 LJ |
3763 | work_flags |= WORK_STRUCT_INACTIVE; |
3764 | ||
affee4b2 TH |
3765 | /* |
3766 | * If @target is currently being executed, schedule the | |
3767 | * barrier to the worker; otherwise, put it after @target. | |
3768 | */ | |
d812796e | 3769 | if (worker) { |
affee4b2 | 3770 | head = worker->scheduled.next; |
d812796e LJ |
3771 | work_color = worker->current_color; |
3772 | } else { | |
affee4b2 TH |
3773 | unsigned long *bits = work_data_bits(target); |
3774 | ||
3775 | head = target->entry.next; | |
3776 | /* there can already be other linked works, inherit and set */ | |
d21cece0 | 3777 | work_flags |= *bits & WORK_STRUCT_LINKED; |
d812796e | 3778 | work_color = get_work_color(*bits); |
affee4b2 TH |
3779 | __set_bit(WORK_STRUCT_LINKED_BIT, bits); |
3780 | } | |
3781 | ||
d812796e LJ |
3782 | pwq->nr_in_flight[work_color]++; |
3783 | work_flags |= work_color_to_flags(work_color); | |
3784 | ||
d21cece0 | 3785 | insert_work(pwq, &barr->work, head, work_flags); |
fc2e4d70 ON |
3786 | } |
3787 | ||
73f53c4a | 3788 | /** |
112202d9 | 3789 | * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing |
73f53c4a TH |
3790 | * @wq: workqueue being flushed |
3791 | * @flush_color: new flush color, < 0 for no-op | |
3792 | * @work_color: new work color, < 0 for no-op | |
3793 | * | |
112202d9 | 3794 | * Prepare pwqs for workqueue flushing. |
73f53c4a | 3795 | * |
112202d9 TH |
3796 | * If @flush_color is non-negative, flush_color on all pwqs should be |
3797 | * -1. If no pwq has in-flight commands at the specified color, all | |
3798 | * pwq->flush_color's stay at -1 and %false is returned. If any pwq | |
3799 | * has in flight commands, its pwq->flush_color is set to | |
3800 | * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq | |
73f53c4a TH |
3801 | * wakeup logic is armed and %true is returned. |
3802 | * | |
3803 | * The caller should have initialized @wq->first_flusher prior to | |
3804 | * calling this function with non-negative @flush_color. If | |
3805 | * @flush_color is negative, no flush color update is done and %false | |
3806 | * is returned. | |
3807 | * | |
112202d9 | 3808 | * If @work_color is non-negative, all pwqs should have the same |
73f53c4a TH |
3809 | * work_color which is previous to @work_color and all will be |
3810 | * advanced to @work_color. | |
3811 | * | |
3812 | * CONTEXT: | |
3c25a55d | 3813 | * mutex_lock(wq->mutex). |
73f53c4a | 3814 | * |
d185af30 | 3815 | * Return: |
73f53c4a TH |
3816 | * %true if @flush_color >= 0 and there's something to flush. %false |
3817 | * otherwise. | |
3818 | */ | |
112202d9 | 3819 | static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, |
73f53c4a | 3820 | int flush_color, int work_color) |
1da177e4 | 3821 | { |
73f53c4a | 3822 | bool wait = false; |
49e3cf44 | 3823 | struct pool_workqueue *pwq; |
1da177e4 | 3824 | |
73f53c4a | 3825 | if (flush_color >= 0) { |
6183c009 | 3826 | WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); |
112202d9 | 3827 | atomic_set(&wq->nr_pwqs_to_flush, 1); |
1da177e4 | 3828 | } |
2355b70f | 3829 | |
49e3cf44 | 3830 | for_each_pwq(pwq, wq) { |
112202d9 | 3831 | struct worker_pool *pool = pwq->pool; |
fc2e4d70 | 3832 | |
a9b8a985 | 3833 | raw_spin_lock_irq(&pool->lock); |
83c22520 | 3834 | |
73f53c4a | 3835 | if (flush_color >= 0) { |
6183c009 | 3836 | WARN_ON_ONCE(pwq->flush_color != -1); |
fc2e4d70 | 3837 | |
112202d9 TH |
3838 | if (pwq->nr_in_flight[flush_color]) { |
3839 | pwq->flush_color = flush_color; | |
3840 | atomic_inc(&wq->nr_pwqs_to_flush); | |
73f53c4a TH |
3841 | wait = true; |
3842 | } | |
3843 | } | |
1da177e4 | 3844 | |
73f53c4a | 3845 | if (work_color >= 0) { |
6183c009 | 3846 | WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); |
112202d9 | 3847 | pwq->work_color = work_color; |
73f53c4a | 3848 | } |
1da177e4 | 3849 | |
a9b8a985 | 3850 | raw_spin_unlock_irq(&pool->lock); |
1da177e4 | 3851 | } |
2355b70f | 3852 | |
112202d9 | 3853 | if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) |
73f53c4a | 3854 | complete(&wq->first_flusher->done); |
14441960 | 3855 | |
73f53c4a | 3856 | return wait; |
1da177e4 LT |
3857 | } |
3858 | ||
c35aea39 TH |
3859 | static void touch_wq_lockdep_map(struct workqueue_struct *wq) |
3860 | { | |
4cb1ef64 TH |
3861 | #ifdef CONFIG_LOCKDEP |
3862 | if (wq->flags & WQ_BH) | |
3863 | local_bh_disable(); | |
3864 | ||
c35aea39 TH |
3865 | lock_map_acquire(&wq->lockdep_map); |
3866 | lock_map_release(&wq->lockdep_map); | |
4cb1ef64 TH |
3867 | |
3868 | if (wq->flags & WQ_BH) | |
3869 | local_bh_enable(); | |
3870 | #endif | |
c35aea39 TH |
3871 | } |
3872 | ||
3873 | static void touch_work_lockdep_map(struct work_struct *work, | |
3874 | struct workqueue_struct *wq) | |
3875 | { | |
4cb1ef64 TH |
3876 | #ifdef CONFIG_LOCKDEP |
3877 | if (wq->flags & WQ_BH) | |
3878 | local_bh_disable(); | |
3879 | ||
c35aea39 TH |
3880 | lock_map_acquire(&work->lockdep_map); |
3881 | lock_map_release(&work->lockdep_map); | |
4cb1ef64 TH |
3882 | |
3883 | if (wq->flags & WQ_BH) | |
3884 | local_bh_enable(); | |
3885 | #endif | |
c35aea39 TH |
3886 | } |
3887 | ||
0fcb78c2 | 3888 | /** |
c4f135d6 | 3889 | * __flush_workqueue - ensure that any scheduled work has run to completion. |
0fcb78c2 | 3890 | * @wq: workqueue to flush |
1da177e4 | 3891 | * |
c5aa87bb TH |
3892 | * This function sleeps until all work items which were queued on entry |
3893 | * have finished execution, but it is not livelocked by new incoming ones. | |
1da177e4 | 3894 | */ |
c4f135d6 | 3895 | void __flush_workqueue(struct workqueue_struct *wq) |
1da177e4 | 3896 | { |
73f53c4a TH |
3897 | struct wq_flusher this_flusher = { |
3898 | .list = LIST_HEAD_INIT(this_flusher.list), | |
3899 | .flush_color = -1, | |
fd1a5b04 | 3900 | .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map), |
73f53c4a TH |
3901 | }; |
3902 | int next_color; | |
1da177e4 | 3903 | |
3347fa09 TH |
3904 | if (WARN_ON(!wq_online)) |
3905 | return; | |
3906 | ||
c35aea39 | 3907 | touch_wq_lockdep_map(wq); |
87915adc | 3908 | |
3c25a55d | 3909 | mutex_lock(&wq->mutex); |
73f53c4a TH |
3910 | |
3911 | /* | |
3912 | * Start-to-wait phase | |
3913 | */ | |
3914 | next_color = work_next_color(wq->work_color); | |
3915 | ||
3916 | if (next_color != wq->flush_color) { | |
3917 | /* | |
3918 | * Color space is not full. The current work_color | |
3919 | * becomes our flush_color and work_color is advanced | |
3920 | * by one. | |
3921 | */ | |
6183c009 | 3922 | WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); |
73f53c4a TH |
3923 | this_flusher.flush_color = wq->work_color; |
3924 | wq->work_color = next_color; | |
3925 | ||
3926 | if (!wq->first_flusher) { | |
3927 | /* no flush in progress, become the first flusher */ | |
6183c009 | 3928 | WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); |
73f53c4a TH |
3929 | |
3930 | wq->first_flusher = &this_flusher; | |
3931 | ||
112202d9 | 3932 | if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, |
73f53c4a TH |
3933 | wq->work_color)) { |
3934 | /* nothing to flush, done */ | |
3935 | wq->flush_color = next_color; | |
3936 | wq->first_flusher = NULL; | |
3937 | goto out_unlock; | |
3938 | } | |
3939 | } else { | |
3940 | /* wait in queue */ | |
6183c009 | 3941 | WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); |
73f53c4a | 3942 | list_add_tail(&this_flusher.list, &wq->flusher_queue); |
112202d9 | 3943 | flush_workqueue_prep_pwqs(wq, -1, wq->work_color); |
73f53c4a TH |
3944 | } |
3945 | } else { | |
3946 | /* | |
3947 | * Oops, color space is full, wait on overflow queue. | |
3948 | * The next flush completion will assign us | |
3949 | * flush_color and transfer to flusher_queue. | |
3950 | */ | |
3951 | list_add_tail(&this_flusher.list, &wq->flusher_overflow); | |
3952 | } | |
3953 | ||
fca839c0 TH |
3954 | check_flush_dependency(wq, NULL); |
3955 | ||
3c25a55d | 3956 | mutex_unlock(&wq->mutex); |
73f53c4a TH |
3957 | |
3958 | wait_for_completion(&this_flusher.done); | |
3959 | ||
3960 | /* | |
3961 | * Wake-up-and-cascade phase | |
3962 | * | |
3963 | * First flushers are responsible for cascading flushes and | |
3964 | * handling overflow. Non-first flushers can simply return. | |
3965 | */ | |
00d5d15b | 3966 | if (READ_ONCE(wq->first_flusher) != &this_flusher) |
73f53c4a TH |
3967 | return; |
3968 | ||
3c25a55d | 3969 | mutex_lock(&wq->mutex); |
73f53c4a | 3970 | |
4ce48b37 TH |
3971 | /* we might have raced, check again with mutex held */ |
3972 | if (wq->first_flusher != &this_flusher) | |
3973 | goto out_unlock; | |
3974 | ||
00d5d15b | 3975 | WRITE_ONCE(wq->first_flusher, NULL); |
73f53c4a | 3976 | |
6183c009 TH |
3977 | WARN_ON_ONCE(!list_empty(&this_flusher.list)); |
3978 | WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); | |
73f53c4a TH |
3979 | |
3980 | while (true) { | |
3981 | struct wq_flusher *next, *tmp; | |
3982 | ||
3983 | /* complete all the flushers sharing the current flush color */ | |
3984 | list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { | |
3985 | if (next->flush_color != wq->flush_color) | |
3986 | break; | |
3987 | list_del_init(&next->list); | |
3988 | complete(&next->done); | |
3989 | } | |
3990 | ||
6183c009 TH |
3991 | WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && |
3992 | wq->flush_color != work_next_color(wq->work_color)); | |
73f53c4a TH |
3993 | |
3994 | /* this flush_color is finished, advance by one */ | |
3995 | wq->flush_color = work_next_color(wq->flush_color); | |
3996 | ||
3997 | /* one color has been freed, handle overflow queue */ | |
3998 | if (!list_empty(&wq->flusher_overflow)) { | |
3999 | /* | |
4000 | * Assign the same color to all overflowed | |
4001 | * flushers, advance work_color and append to | |
4002 | * flusher_queue. This is the start-to-wait | |
4003 | * phase for these overflowed flushers. | |
4004 | */ | |
4005 | list_for_each_entry(tmp, &wq->flusher_overflow, list) | |
4006 | tmp->flush_color = wq->work_color; | |
4007 | ||
4008 | wq->work_color = work_next_color(wq->work_color); | |
4009 | ||
4010 | list_splice_tail_init(&wq->flusher_overflow, | |
4011 | &wq->flusher_queue); | |
112202d9 | 4012 | flush_workqueue_prep_pwqs(wq, -1, wq->work_color); |
73f53c4a TH |
4013 | } |
4014 | ||
4015 | if (list_empty(&wq->flusher_queue)) { | |
6183c009 | 4016 | WARN_ON_ONCE(wq->flush_color != wq->work_color); |
73f53c4a TH |
4017 | break; |
4018 | } | |
4019 | ||
4020 | /* | |
4021 | * Need to flush more colors. Make the next flusher | |
112202d9 | 4022 | * the new first flusher and arm pwqs. |
73f53c4a | 4023 | */ |
6183c009 TH |
4024 | WARN_ON_ONCE(wq->flush_color == wq->work_color); |
4025 | WARN_ON_ONCE(wq->flush_color != next->flush_color); | |
73f53c4a TH |
4026 | |
4027 | list_del_init(&next->list); | |
4028 | wq->first_flusher = next; | |
4029 | ||
112202d9 | 4030 | if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) |
73f53c4a TH |
4031 | break; |
4032 | ||
4033 | /* | |
4034 | * Meh... this color is already done, clear first | |
4035 | * flusher and repeat cascading. | |
4036 | */ | |
4037 | wq->first_flusher = NULL; | |
4038 | } | |
4039 | ||
4040 | out_unlock: | |
3c25a55d | 4041 | mutex_unlock(&wq->mutex); |
1da177e4 | 4042 | } |
c4f135d6 | 4043 | EXPORT_SYMBOL(__flush_workqueue); |
1da177e4 | 4044 | |
9c5a2ba7 TH |
4045 | /** |
4046 | * drain_workqueue - drain a workqueue | |
4047 | * @wq: workqueue to drain | |
4048 | * | |
4049 | * Wait until the workqueue becomes empty. While draining is in progress, | |
4050 | * only chain queueing is allowed. IOW, only currently pending or running | |
4051 | * work items on @wq can queue further work items on it. @wq is flushed | |
b749b1b6 | 4052 | * repeatedly until it becomes empty. The number of flushing is determined |
9c5a2ba7 TH |
4053 | * by the depth of chaining and should be relatively short. Whine if it |
4054 | * takes too long. | |
4055 | */ | |
4056 | void drain_workqueue(struct workqueue_struct *wq) | |
4057 | { | |
4058 | unsigned int flush_cnt = 0; | |
49e3cf44 | 4059 | struct pool_workqueue *pwq; |
9c5a2ba7 TH |
4060 | |
4061 | /* | |
4062 | * __queue_work() needs to test whether there are drainers, is much | |
4063 | * hotter than drain_workqueue() and already looks at @wq->flags. | |
618b01eb | 4064 | * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. |
9c5a2ba7 | 4065 | */ |
87fc741e | 4066 | mutex_lock(&wq->mutex); |
9c5a2ba7 | 4067 | if (!wq->nr_drainers++) |
618b01eb | 4068 | wq->flags |= __WQ_DRAINING; |
87fc741e | 4069 | mutex_unlock(&wq->mutex); |
9c5a2ba7 | 4070 | reflush: |
c4f135d6 | 4071 | __flush_workqueue(wq); |
9c5a2ba7 | 4072 | |
b09f4fd3 | 4073 | mutex_lock(&wq->mutex); |
76af4d93 | 4074 | |
49e3cf44 | 4075 | for_each_pwq(pwq, wq) { |
fa2563e4 | 4076 | bool drained; |
9c5a2ba7 | 4077 | |
a9b8a985 | 4078 | raw_spin_lock_irq(&pwq->pool->lock); |
afa87ce8 | 4079 | drained = pwq_is_empty(pwq); |
a9b8a985 | 4080 | raw_spin_unlock_irq(&pwq->pool->lock); |
fa2563e4 TT |
4081 | |
4082 | if (drained) | |
9c5a2ba7 TH |
4083 | continue; |
4084 | ||
4085 | if (++flush_cnt == 10 || | |
4086 | (flush_cnt % 100 == 0 && flush_cnt <= 1000)) | |
e9ad2eb3 SZ |
4087 | pr_warn("workqueue %s: %s() isn't complete after %u tries\n", |
4088 | wq->name, __func__, flush_cnt); | |
76af4d93 | 4089 | |
b09f4fd3 | 4090 | mutex_unlock(&wq->mutex); |
9c5a2ba7 TH |
4091 | goto reflush; |
4092 | } | |
4093 | ||
9c5a2ba7 | 4094 | if (!--wq->nr_drainers) |
618b01eb | 4095 | wq->flags &= ~__WQ_DRAINING; |
87fc741e | 4096 | mutex_unlock(&wq->mutex); |
9c5a2ba7 TH |
4097 | } |
4098 | EXPORT_SYMBOL_GPL(drain_workqueue); | |
4099 | ||
d6e89786 JB |
4100 | static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr, |
4101 | bool from_cancel) | |
db700897 | 4102 | { |
affee4b2 | 4103 | struct worker *worker = NULL; |
c9e7cf27 | 4104 | struct worker_pool *pool; |
112202d9 | 4105 | struct pool_workqueue *pwq; |
c35aea39 | 4106 | struct workqueue_struct *wq; |
db700897 | 4107 | |
24acfb71 | 4108 | rcu_read_lock(); |
c9e7cf27 | 4109 | pool = get_work_pool(work); |
fa1b54e6 | 4110 | if (!pool) { |
24acfb71 | 4111 | rcu_read_unlock(); |
baf59022 | 4112 | return false; |
fa1b54e6 | 4113 | } |
db700897 | 4114 | |
a9b8a985 | 4115 | raw_spin_lock_irq(&pool->lock); |
0b3dae68 | 4116 | /* see the comment in try_to_grab_pending() with the same code */ |
112202d9 TH |
4117 | pwq = get_work_pwq(work); |
4118 | if (pwq) { | |
4119 | if (unlikely(pwq->pool != pool)) | |
4690c4ab | 4120 | goto already_gone; |
606a5020 | 4121 | } else { |
c9e7cf27 | 4122 | worker = find_worker_executing_work(pool, work); |
affee4b2 | 4123 | if (!worker) |
4690c4ab | 4124 | goto already_gone; |
112202d9 | 4125 | pwq = worker->current_pwq; |
606a5020 | 4126 | } |
db700897 | 4127 | |
c35aea39 TH |
4128 | wq = pwq->wq; |
4129 | check_flush_dependency(wq, work); | |
fca839c0 | 4130 | |
112202d9 | 4131 | insert_wq_barrier(pwq, barr, work, worker); |
a9b8a985 | 4132 | raw_spin_unlock_irq(&pool->lock); |
7a22ad75 | 4133 | |
c35aea39 TH |
4134 | touch_work_lockdep_map(work, wq); |
4135 | ||
e159489b | 4136 | /* |
a1d14934 PZ |
4137 | * Force a lock recursion deadlock when using flush_work() inside a |
4138 | * single-threaded or rescuer equipped workqueue. | |
4139 | * | |
4140 | * For single threaded workqueues the deadlock happens when the work | |
4141 | * is after the work issuing the flush_work(). For rescuer equipped | |
4142 | * workqueues the deadlock happens when the rescuer stalls, blocking | |
4143 | * forward progress. | |
e159489b | 4144 | */ |
c35aea39 TH |
4145 | if (!from_cancel && (wq->saved_max_active == 1 || wq->rescuer)) |
4146 | touch_wq_lockdep_map(wq); | |
4147 | ||
24acfb71 | 4148 | rcu_read_unlock(); |
401a8d04 | 4149 | return true; |
4690c4ab | 4150 | already_gone: |
a9b8a985 | 4151 | raw_spin_unlock_irq(&pool->lock); |
24acfb71 | 4152 | rcu_read_unlock(); |
401a8d04 | 4153 | return false; |
db700897 | 4154 | } |
baf59022 | 4155 | |
d6e89786 JB |
4156 | static bool __flush_work(struct work_struct *work, bool from_cancel) |
4157 | { | |
4158 | struct wq_barrier barr; | |
134874e2 | 4159 | unsigned long data; |
d6e89786 JB |
4160 | |
4161 | if (WARN_ON(!wq_online)) | |
4162 | return false; | |
4163 | ||
4d43d395 TH |
4164 | if (WARN_ON(!work->func)) |
4165 | return false; | |
4166 | ||
134874e2 | 4167 | if (!start_flush_work(work, &barr, from_cancel)) |
d6e89786 | 4168 | return false; |
134874e2 TH |
4169 | |
4170 | /* | |
4171 | * start_flush_work() returned %true. If @from_cancel is set, we know | |
4172 | * that @work must have been executing during start_flush_work() and | |
4173 | * can't currently be queued. Its data must contain OFFQ bits. If @work | |
4174 | * was queued on a BH workqueue, we also know that it was running in the | |
4175 | * BH context and thus can be busy-waited. | |
4176 | */ | |
4177 | data = *work_data_bits(work); | |
4178 | if (from_cancel && | |
4179 | !WARN_ON_ONCE(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_BH)) { | |
4180 | /* | |
4181 | * On RT, prevent a live lock when %current preempted soft | |
4182 | * interrupt processing or prevents ksoftirqd from running by | |
4183 | * keeping flipping BH. If the BH work item runs on a different | |
4184 | * CPU then this has no effect other than doing the BH | |
4185 | * disable/enable dance for nothing. This is copied from | |
4186 | * kernel/softirq.c::tasklet_unlock_spin_wait(). | |
4187 | */ | |
4188 | while (!try_wait_for_completion(&barr.done)) { | |
4189 | if (IS_ENABLED(CONFIG_PREEMPT_RT)) { | |
4190 | local_bh_disable(); | |
4191 | local_bh_enable(); | |
4192 | } else { | |
4193 | cpu_relax(); | |
4194 | } | |
4195 | } | |
4196 | } else { | |
4197 | wait_for_completion(&barr.done); | |
d6e89786 | 4198 | } |
134874e2 TH |
4199 | |
4200 | destroy_work_on_stack(&barr.work); | |
4201 | return true; | |
d6e89786 JB |
4202 | } |
4203 | ||
baf59022 TH |
4204 | /** |
4205 | * flush_work - wait for a work to finish executing the last queueing instance | |
4206 | * @work: the work to flush | |
4207 | * | |
606a5020 TH |
4208 | * Wait until @work has finished execution. @work is guaranteed to be idle |
4209 | * on return if it hasn't been requeued since flush started. | |
baf59022 | 4210 | * |
d185af30 | 4211 | * Return: |
baf59022 TH |
4212 | * %true if flush_work() waited for the work to finish execution, |
4213 | * %false if it was already idle. | |
4214 | */ | |
4215 | bool flush_work(struct work_struct *work) | |
4216 | { | |
134874e2 | 4217 | might_sleep(); |
d6e89786 | 4218 | return __flush_work(work, false); |
6e84d644 | 4219 | } |
606a5020 | 4220 | EXPORT_SYMBOL_GPL(flush_work); |
6e84d644 | 4221 | |
cdc6e4b3 TH |
4222 | /** |
4223 | * flush_delayed_work - wait for a dwork to finish executing the last queueing | |
4224 | * @dwork: the delayed work to flush | |
4225 | * | |
4226 | * Delayed timer is cancelled and the pending work is queued for | |
4227 | * immediate execution. Like flush_work(), this function only | |
4228 | * considers the last queueing instance of @dwork. | |
4229 | * | |
4230 | * Return: | |
4231 | * %true if flush_work() waited for the work to finish execution, | |
4232 | * %false if it was already idle. | |
4233 | */ | |
4234 | bool flush_delayed_work(struct delayed_work *dwork) | |
4235 | { | |
4236 | local_irq_disable(); | |
4237 | if (del_timer_sync(&dwork->timer)) | |
4238 | __queue_work(dwork->cpu, dwork->wq, &dwork->work); | |
4239 | local_irq_enable(); | |
4240 | return flush_work(&dwork->work); | |
4241 | } | |
4242 | EXPORT_SYMBOL(flush_delayed_work); | |
4243 | ||
4244 | /** | |
4245 | * flush_rcu_work - wait for a rwork to finish executing the last queueing | |
4246 | * @rwork: the rcu work to flush | |
4247 | * | |
4248 | * Return: | |
4249 | * %true if flush_rcu_work() waited for the work to finish execution, | |
4250 | * %false if it was already idle. | |
4251 | */ | |
4252 | bool flush_rcu_work(struct rcu_work *rwork) | |
4253 | { | |
4254 | if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) { | |
4255 | rcu_barrier(); | |
4256 | flush_work(&rwork->work); | |
4257 | return true; | |
4258 | } else { | |
4259 | return flush_work(&rwork->work); | |
4260 | } | |
4261 | } | |
4262 | EXPORT_SYMBOL(flush_rcu_work); | |
4263 | ||
86898fa6 TH |
4264 | static void work_offqd_disable(struct work_offq_data *offqd) |
4265 | { | |
4266 | const unsigned long max = (1lu << WORK_OFFQ_DISABLE_BITS) - 1; | |
4267 | ||
4268 | if (likely(offqd->disable < max)) | |
4269 | offqd->disable++; | |
4270 | else | |
4271 | WARN_ONCE(true, "workqueue: work disable count overflowed\n"); | |
4272 | } | |
4273 | ||
4274 | static void work_offqd_enable(struct work_offq_data *offqd) | |
4275 | { | |
4276 | if (likely(offqd->disable > 0)) | |
4277 | offqd->disable--; | |
4278 | else | |
4279 | WARN_ONCE(true, "workqueue: work disable count underflowed\n"); | |
4280 | } | |
4281 | ||
c5f5b942 | 4282 | static bool __cancel_work(struct work_struct *work, u32 cflags) |
cdc6e4b3 | 4283 | { |
1211f3b2 | 4284 | struct work_offq_data offqd; |
c26e2f2e | 4285 | unsigned long irq_flags; |
cdc6e4b3 TH |
4286 | int ret; |
4287 | ||
f09b10b6 | 4288 | ret = work_grab_pending(work, cflags, &irq_flags); |
cdc6e4b3 | 4289 | |
1211f3b2 | 4290 | work_offqd_unpack(&offqd, *work_data_bits(work)); |
86898fa6 TH |
4291 | |
4292 | if (cflags & WORK_CANCEL_DISABLE) | |
4293 | work_offqd_disable(&offqd); | |
4294 | ||
1211f3b2 TH |
4295 | set_work_pool_and_clear_pending(work, offqd.pool_id, |
4296 | work_offqd_pack_flags(&offqd)); | |
c26e2f2e | 4297 | local_irq_restore(irq_flags); |
cdc6e4b3 TH |
4298 | return ret; |
4299 | } | |
4300 | ||
c5f5b942 | 4301 | static bool __cancel_work_sync(struct work_struct *work, u32 cflags) |
1f1f642e | 4302 | { |
978b8409 | 4303 | bool ret; |
1f1f642e | 4304 | |
f09b10b6 | 4305 | ret = __cancel_work(work, cflags | WORK_CANCEL_DISABLE); |
bbb68dfa | 4306 | |
134874e2 TH |
4307 | if (*work_data_bits(work) & WORK_OFFQ_BH) |
4308 | WARN_ON_ONCE(in_hardirq()); | |
4309 | else | |
4310 | might_sleep(); | |
4311 | ||
3347fa09 | 4312 | /* |
c7a40c49 TH |
4313 | * Skip __flush_work() during early boot when we know that @work isn't |
4314 | * executing. This allows canceling during early boot. | |
3347fa09 TH |
4315 | */ |
4316 | if (wq_online) | |
d6e89786 | 4317 | __flush_work(work, true); |
3347fa09 | 4318 | |
f09b10b6 TH |
4319 | if (!(cflags & WORK_CANCEL_DISABLE)) |
4320 | enable_work(work); | |
8603e1b3 | 4321 | |
1f1f642e ON |
4322 | return ret; |
4323 | } | |
4324 | ||
cdc6e4b3 TH |
4325 | /* |
4326 | * See cancel_delayed_work() | |
4327 | */ | |
4328 | bool cancel_work(struct work_struct *work) | |
4329 | { | |
c5f5b942 | 4330 | return __cancel_work(work, 0); |
cdc6e4b3 TH |
4331 | } |
4332 | EXPORT_SYMBOL(cancel_work); | |
4333 | ||
6e84d644 | 4334 | /** |
401a8d04 TH |
4335 | * cancel_work_sync - cancel a work and wait for it to finish |
4336 | * @work: the work to cancel | |
6e84d644 | 4337 | * |
134874e2 TH |
4338 | * Cancel @work and wait for its execution to finish. This function can be used |
4339 | * even if the work re-queues itself or migrates to another workqueue. On return | |
4340 | * from this function, @work is guaranteed to be not pending or executing on any | |
4341 | * CPU as long as there aren't racing enqueues. | |
1f1f642e | 4342 | * |
134874e2 TH |
4343 | * cancel_work_sync(&delayed_work->work) must not be used for delayed_work's. |
4344 | * Use cancel_delayed_work_sync() instead. | |
6e84d644 | 4345 | * |
134874e2 TH |
4346 | * Must be called from a sleepable context if @work was last queued on a non-BH |
4347 | * workqueue. Can also be called from non-hardirq atomic contexts including BH | |
4348 | * if @work was last queued on a BH workqueue. | |
401a8d04 | 4349 | * |
134874e2 | 4350 | * Returns %true if @work was pending, %false otherwise. |
6e84d644 | 4351 | */ |
401a8d04 | 4352 | bool cancel_work_sync(struct work_struct *work) |
6e84d644 | 4353 | { |
c5f5b942 | 4354 | return __cancel_work_sync(work, 0); |
b89deed3 | 4355 | } |
28e53bdd | 4356 | EXPORT_SYMBOL_GPL(cancel_work_sync); |
b89deed3 | 4357 | |
09383498 | 4358 | /** |
57b30ae7 TH |
4359 | * cancel_delayed_work - cancel a delayed work |
4360 | * @dwork: delayed_work to cancel | |
09383498 | 4361 | * |
d185af30 YB |
4362 | * Kill off a pending delayed_work. |
4363 | * | |
4364 | * Return: %true if @dwork was pending and canceled; %false if it wasn't | |
4365 | * pending. | |
4366 | * | |
4367 | * Note: | |
4368 | * The work callback function may still be running on return, unless | |
4369 | * it returns %true and the work doesn't re-arm itself. Explicitly flush or | |
4370 | * use cancel_delayed_work_sync() to wait on it. | |
09383498 | 4371 | * |
57b30ae7 | 4372 | * This function is safe to call from any context including IRQ handler. |
09383498 | 4373 | */ |
57b30ae7 | 4374 | bool cancel_delayed_work(struct delayed_work *dwork) |
09383498 | 4375 | { |
c5f5b942 | 4376 | return __cancel_work(&dwork->work, WORK_CANCEL_DELAYED); |
09383498 | 4377 | } |
57b30ae7 | 4378 | EXPORT_SYMBOL(cancel_delayed_work); |
09383498 | 4379 | |
401a8d04 TH |
4380 | /** |
4381 | * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish | |
4382 | * @dwork: the delayed work cancel | |
4383 | * | |
4384 | * This is cancel_work_sync() for delayed works. | |
4385 | * | |
d185af30 | 4386 | * Return: |
401a8d04 TH |
4387 | * %true if @dwork was pending, %false otherwise. |
4388 | */ | |
4389 | bool cancel_delayed_work_sync(struct delayed_work *dwork) | |
6e84d644 | 4390 | { |
c5f5b942 | 4391 | return __cancel_work_sync(&dwork->work, WORK_CANCEL_DELAYED); |
6e84d644 | 4392 | } |
f5a421a4 | 4393 | EXPORT_SYMBOL(cancel_delayed_work_sync); |
1da177e4 | 4394 | |
86898fa6 TH |
4395 | /** |
4396 | * disable_work - Disable and cancel a work item | |
4397 | * @work: work item to disable | |
4398 | * | |
4399 | * Disable @work by incrementing its disable count and cancel it if currently | |
4400 | * pending. As long as the disable count is non-zero, any attempt to queue @work | |
4401 | * will fail and return %false. The maximum supported disable depth is 2 to the | |
4402 | * power of %WORK_OFFQ_DISABLE_BITS, currently 65536. | |
4403 | * | |
f09b10b6 TH |
4404 | * Can be called from any context. Returns %true if @work was pending, %false |
4405 | * otherwise. | |
86898fa6 TH |
4406 | */ |
4407 | bool disable_work(struct work_struct *work) | |
4408 | { | |
4409 | return __cancel_work(work, WORK_CANCEL_DISABLE); | |
4410 | } | |
4411 | EXPORT_SYMBOL_GPL(disable_work); | |
4412 | ||
4413 | /** | |
4414 | * disable_work_sync - Disable, cancel and drain a work item | |
4415 | * @work: work item to disable | |
4416 | * | |
4417 | * Similar to disable_work() but also wait for @work to finish if currently | |
4418 | * executing. | |
4419 | * | |
134874e2 TH |
4420 | * Must be called from a sleepable context if @work was last queued on a non-BH |
4421 | * workqueue. Can also be called from non-hardirq atomic contexts including BH | |
4422 | * if @work was last queued on a BH workqueue. | |
4423 | * | |
4424 | * Returns %true if @work was pending, %false otherwise. | |
86898fa6 TH |
4425 | */ |
4426 | bool disable_work_sync(struct work_struct *work) | |
4427 | { | |
4428 | return __cancel_work_sync(work, WORK_CANCEL_DISABLE); | |
4429 | } | |
4430 | EXPORT_SYMBOL_GPL(disable_work_sync); | |
4431 | ||
4432 | /** | |
4433 | * enable_work - Enable a work item | |
4434 | * @work: work item to enable | |
4435 | * | |
4436 | * Undo disable_work[_sync]() by decrementing @work's disable count. @work can | |
4437 | * only be queued if its disable count is 0. | |
4438 | * | |
f09b10b6 TH |
4439 | * Can be called from any context. Returns %true if the disable count reached 0. |
4440 | * Otherwise, %false. | |
86898fa6 TH |
4441 | */ |
4442 | bool enable_work(struct work_struct *work) | |
4443 | { | |
4444 | struct work_offq_data offqd; | |
4445 | unsigned long irq_flags; | |
4446 | ||
4447 | work_grab_pending(work, 0, &irq_flags); | |
4448 | ||
4449 | work_offqd_unpack(&offqd, *work_data_bits(work)); | |
4450 | work_offqd_enable(&offqd); | |
4451 | set_work_pool_and_clear_pending(work, offqd.pool_id, | |
4452 | work_offqd_pack_flags(&offqd)); | |
4453 | local_irq_restore(irq_flags); | |
4454 | ||
4455 | return !offqd.disable; | |
4456 | } | |
4457 | EXPORT_SYMBOL_GPL(enable_work); | |
4458 | ||
4459 | /** | |
4460 | * disable_delayed_work - Disable and cancel a delayed work item | |
4461 | * @dwork: delayed work item to disable | |
4462 | * | |
4463 | * disable_work() for delayed work items. | |
4464 | */ | |
4465 | bool disable_delayed_work(struct delayed_work *dwork) | |
4466 | { | |
4467 | return __cancel_work(&dwork->work, | |
4468 | WORK_CANCEL_DELAYED | WORK_CANCEL_DISABLE); | |
4469 | } | |
4470 | EXPORT_SYMBOL_GPL(disable_delayed_work); | |
4471 | ||
4472 | /** | |
4473 | * disable_delayed_work_sync - Disable, cancel and drain a delayed work item | |
4474 | * @dwork: delayed work item to disable | |
4475 | * | |
4476 | * disable_work_sync() for delayed work items. | |
4477 | */ | |
4478 | bool disable_delayed_work_sync(struct delayed_work *dwork) | |
4479 | { | |
4480 | return __cancel_work_sync(&dwork->work, | |
4481 | WORK_CANCEL_DELAYED | WORK_CANCEL_DISABLE); | |
4482 | } | |
4483 | EXPORT_SYMBOL_GPL(disable_delayed_work_sync); | |
4484 | ||
4485 | /** | |
4486 | * enable_delayed_work - Enable a delayed work item | |
4487 | * @dwork: delayed work item to enable | |
4488 | * | |
4489 | * enable_work() for delayed work items. | |
4490 | */ | |
4491 | bool enable_delayed_work(struct delayed_work *dwork) | |
4492 | { | |
4493 | return enable_work(&dwork->work); | |
4494 | } | |
4495 | EXPORT_SYMBOL_GPL(enable_delayed_work); | |
4496 | ||
b6136773 | 4497 | /** |
31ddd871 | 4498 | * schedule_on_each_cpu - execute a function synchronously on each online CPU |
b6136773 | 4499 | * @func: the function to call |
b6136773 | 4500 | * |
31ddd871 TH |
4501 | * schedule_on_each_cpu() executes @func on each online CPU using the |
4502 | * system workqueue and blocks until all CPUs have completed. | |
b6136773 | 4503 | * schedule_on_each_cpu() is very slow. |
31ddd871 | 4504 | * |
d185af30 | 4505 | * Return: |
31ddd871 | 4506 | * 0 on success, -errno on failure. |
b6136773 | 4507 | */ |
65f27f38 | 4508 | int schedule_on_each_cpu(work_func_t func) |
15316ba8 CL |
4509 | { |
4510 | int cpu; | |
38f51568 | 4511 | struct work_struct __percpu *works; |
15316ba8 | 4512 | |
b6136773 AM |
4513 | works = alloc_percpu(struct work_struct); |
4514 | if (!works) | |
15316ba8 | 4515 | return -ENOMEM; |
b6136773 | 4516 | |
ffd8bea8 | 4517 | cpus_read_lock(); |
93981800 | 4518 | |
15316ba8 | 4519 | for_each_online_cpu(cpu) { |
9bfb1839 IM |
4520 | struct work_struct *work = per_cpu_ptr(works, cpu); |
4521 | ||
4522 | INIT_WORK(work, func); | |
b71ab8c2 | 4523 | schedule_work_on(cpu, work); |
65a64464 | 4524 | } |
93981800 TH |
4525 | |
4526 | for_each_online_cpu(cpu) | |
4527 | flush_work(per_cpu_ptr(works, cpu)); | |
4528 | ||
ffd8bea8 | 4529 | cpus_read_unlock(); |
b6136773 | 4530 | free_percpu(works); |
15316ba8 CL |
4531 | return 0; |
4532 | } | |
4533 | ||
1fa44eca JB |
4534 | /** |
4535 | * execute_in_process_context - reliably execute the routine with user context | |
4536 | * @fn: the function to execute | |
1fa44eca JB |
4537 | * @ew: guaranteed storage for the execute work structure (must |
4538 | * be available when the work executes) | |
4539 | * | |
4540 | * Executes the function immediately if process context is available, | |
4541 | * otherwise schedules the function for delayed execution. | |
4542 | * | |
d185af30 | 4543 | * Return: 0 - function was executed |
1fa44eca JB |
4544 | * 1 - function was scheduled for execution |
4545 | */ | |
65f27f38 | 4546 | int execute_in_process_context(work_func_t fn, struct execute_work *ew) |
1fa44eca JB |
4547 | { |
4548 | if (!in_interrupt()) { | |
65f27f38 | 4549 | fn(&ew->work); |
1fa44eca JB |
4550 | return 0; |
4551 | } | |
4552 | ||
65f27f38 | 4553 | INIT_WORK(&ew->work, fn); |
1fa44eca JB |
4554 | schedule_work(&ew->work); |
4555 | ||
4556 | return 1; | |
4557 | } | |
4558 | EXPORT_SYMBOL_GPL(execute_in_process_context); | |
4559 | ||
6ba94429 FW |
4560 | /** |
4561 | * free_workqueue_attrs - free a workqueue_attrs | |
4562 | * @attrs: workqueue_attrs to free | |
226223ab | 4563 | * |
6ba94429 | 4564 | * Undo alloc_workqueue_attrs(). |
226223ab | 4565 | */ |
513c98d0 | 4566 | void free_workqueue_attrs(struct workqueue_attrs *attrs) |
226223ab | 4567 | { |
6ba94429 FW |
4568 | if (attrs) { |
4569 | free_cpumask_var(attrs->cpumask); | |
9546b29e | 4570 | free_cpumask_var(attrs->__pod_cpumask); |
6ba94429 FW |
4571 | kfree(attrs); |
4572 | } | |
226223ab TH |
4573 | } |
4574 | ||
6ba94429 FW |
4575 | /** |
4576 | * alloc_workqueue_attrs - allocate a workqueue_attrs | |
6ba94429 FW |
4577 | * |
4578 | * Allocate a new workqueue_attrs, initialize with default settings and | |
4579 | * return it. | |
4580 | * | |
4581 | * Return: The allocated new workqueue_attr on success. %NULL on failure. | |
4582 | */ | |
513c98d0 | 4583 | struct workqueue_attrs *alloc_workqueue_attrs(void) |
226223ab | 4584 | { |
6ba94429 | 4585 | struct workqueue_attrs *attrs; |
226223ab | 4586 | |
be69d00d | 4587 | attrs = kzalloc(sizeof(*attrs), GFP_KERNEL); |
6ba94429 FW |
4588 | if (!attrs) |
4589 | goto fail; | |
be69d00d | 4590 | if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL)) |
6ba94429 | 4591 | goto fail; |
9546b29e TH |
4592 | if (!alloc_cpumask_var(&attrs->__pod_cpumask, GFP_KERNEL)) |
4593 | goto fail; | |
6ba94429 FW |
4594 | |
4595 | cpumask_copy(attrs->cpumask, cpu_possible_mask); | |
523a301e | 4596 | attrs->affn_scope = WQ_AFFN_DFL; |
6ba94429 FW |
4597 | return attrs; |
4598 | fail: | |
4599 | free_workqueue_attrs(attrs); | |
4600 | return NULL; | |
226223ab TH |
4601 | } |
4602 | ||
6ba94429 FW |
4603 | static void copy_workqueue_attrs(struct workqueue_attrs *to, |
4604 | const struct workqueue_attrs *from) | |
226223ab | 4605 | { |
6ba94429 FW |
4606 | to->nice = from->nice; |
4607 | cpumask_copy(to->cpumask, from->cpumask); | |
9546b29e | 4608 | cpumask_copy(to->__pod_cpumask, from->__pod_cpumask); |
8639eceb | 4609 | to->affn_strict = from->affn_strict; |
84193c07 | 4610 | |
6ba94429 | 4611 | /* |
84193c07 TH |
4612 | * Unlike hash and equality test, copying shouldn't ignore wq-only |
4613 | * fields as copying is used for both pool and wq attrs. Instead, | |
4614 | * get_unbound_pool() explicitly clears the fields. | |
6ba94429 | 4615 | */ |
84193c07 | 4616 | to->affn_scope = from->affn_scope; |
af73f5c9 | 4617 | to->ordered = from->ordered; |
226223ab TH |
4618 | } |
4619 | ||
5de7a03c TH |
4620 | /* |
4621 | * Some attrs fields are workqueue-only. Clear them for worker_pool's. See the | |
4622 | * comments in 'struct workqueue_attrs' definition. | |
4623 | */ | |
4624 | static void wqattrs_clear_for_pool(struct workqueue_attrs *attrs) | |
4625 | { | |
84193c07 | 4626 | attrs->affn_scope = WQ_AFFN_NR_TYPES; |
5de7a03c | 4627 | attrs->ordered = false; |
ae1296a7 LJ |
4628 | if (attrs->affn_strict) |
4629 | cpumask_copy(attrs->cpumask, cpu_possible_mask); | |
5de7a03c TH |
4630 | } |
4631 | ||
6ba94429 FW |
4632 | /* hash value of the content of @attr */ |
4633 | static u32 wqattrs_hash(const struct workqueue_attrs *attrs) | |
226223ab | 4634 | { |
6ba94429 | 4635 | u32 hash = 0; |
226223ab | 4636 | |
6ba94429 | 4637 | hash = jhash_1word(attrs->nice, hash); |
ae1296a7 | 4638 | hash = jhash_1word(attrs->affn_strict, hash); |
9546b29e TH |
4639 | hash = jhash(cpumask_bits(attrs->__pod_cpumask), |
4640 | BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); | |
ae1296a7 LJ |
4641 | if (!attrs->affn_strict) |
4642 | hash = jhash(cpumask_bits(attrs->cpumask), | |
4643 | BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); | |
6ba94429 | 4644 | return hash; |
226223ab | 4645 | } |
226223ab | 4646 | |
6ba94429 FW |
4647 | /* content equality test */ |
4648 | static bool wqattrs_equal(const struct workqueue_attrs *a, | |
4649 | const struct workqueue_attrs *b) | |
226223ab | 4650 | { |
6ba94429 FW |
4651 | if (a->nice != b->nice) |
4652 | return false; | |
ae1296a7 | 4653 | if (a->affn_strict != b->affn_strict) |
6ba94429 | 4654 | return false; |
9546b29e TH |
4655 | if (!cpumask_equal(a->__pod_cpumask, b->__pod_cpumask)) |
4656 | return false; | |
ae1296a7 | 4657 | if (!a->affn_strict && !cpumask_equal(a->cpumask, b->cpumask)) |
8639eceb | 4658 | return false; |
6ba94429 | 4659 | return true; |
226223ab TH |
4660 | } |
4661 | ||
0f36ee24 TH |
4662 | /* Update @attrs with actually available CPUs */ |
4663 | static void wqattrs_actualize_cpumask(struct workqueue_attrs *attrs, | |
4664 | const cpumask_t *unbound_cpumask) | |
4665 | { | |
4666 | /* | |
4667 | * Calculate the effective CPU mask of @attrs given @unbound_cpumask. If | |
4668 | * @attrs->cpumask doesn't overlap with @unbound_cpumask, we fallback to | |
4669 | * @unbound_cpumask. | |
4670 | */ | |
4671 | cpumask_and(attrs->cpumask, attrs->cpumask, unbound_cpumask); | |
4672 | if (unlikely(cpumask_empty(attrs->cpumask))) | |
4673 | cpumask_copy(attrs->cpumask, unbound_cpumask); | |
4674 | } | |
4675 | ||
84193c07 TH |
4676 | /* find wq_pod_type to use for @attrs */ |
4677 | static const struct wq_pod_type * | |
4678 | wqattrs_pod_type(const struct workqueue_attrs *attrs) | |
4679 | { | |
523a301e TH |
4680 | enum wq_affn_scope scope; |
4681 | struct wq_pod_type *pt; | |
4682 | ||
4683 | /* to synchronize access to wq_affn_dfl */ | |
4684 | lockdep_assert_held(&wq_pool_mutex); | |
4685 | ||
4686 | if (attrs->affn_scope == WQ_AFFN_DFL) | |
4687 | scope = wq_affn_dfl; | |
4688 | else | |
4689 | scope = attrs->affn_scope; | |
4690 | ||
4691 | pt = &wq_pod_types[scope]; | |
84193c07 TH |
4692 | |
4693 | if (!WARN_ON_ONCE(attrs->affn_scope == WQ_AFFN_NR_TYPES) && | |
4694 | likely(pt->nr_pods)) | |
4695 | return pt; | |
4696 | ||
4697 | /* | |
4698 | * Before workqueue_init_topology(), only SYSTEM is available which is | |
4699 | * initialized in workqueue_init_early(). | |
4700 | */ | |
4701 | pt = &wq_pod_types[WQ_AFFN_SYSTEM]; | |
4702 | BUG_ON(!pt->nr_pods); | |
4703 | return pt; | |
4704 | } | |
4705 | ||
6ba94429 FW |
4706 | /** |
4707 | * init_worker_pool - initialize a newly zalloc'd worker_pool | |
4708 | * @pool: worker_pool to initialize | |
4709 | * | |
402dd89d | 4710 | * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs. |
6ba94429 FW |
4711 | * |
4712 | * Return: 0 on success, -errno on failure. Even on failure, all fields | |
4713 | * inside @pool proper are initialized and put_unbound_pool() can be called | |
4714 | * on @pool safely to release it. | |
4715 | */ | |
4716 | static int init_worker_pool(struct worker_pool *pool) | |
226223ab | 4717 | { |
a9b8a985 | 4718 | raw_spin_lock_init(&pool->lock); |
6ba94429 FW |
4719 | pool->id = -1; |
4720 | pool->cpu = -1; | |
4721 | pool->node = NUMA_NO_NODE; | |
4722 | pool->flags |= POOL_DISASSOCIATED; | |
82607adc | 4723 | pool->watchdog_ts = jiffies; |
6ba94429 FW |
4724 | INIT_LIST_HEAD(&pool->worklist); |
4725 | INIT_LIST_HEAD(&pool->idle_list); | |
4726 | hash_init(pool->busy_hash); | |
226223ab | 4727 | |
32a6c723 | 4728 | timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE); |
3f959aa3 | 4729 | INIT_WORK(&pool->idle_cull_work, idle_cull_fn); |
226223ab | 4730 | |
32a6c723 | 4731 | timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0); |
226223ab | 4732 | |
6ba94429 | 4733 | INIT_LIST_HEAD(&pool->workers); |
e02b9312 | 4734 | INIT_LIST_HEAD(&pool->dying_workers); |
226223ab | 4735 | |
6ba94429 FW |
4736 | ida_init(&pool->worker_ida); |
4737 | INIT_HLIST_NODE(&pool->hash_node); | |
4738 | pool->refcnt = 1; | |
226223ab | 4739 | |
6ba94429 | 4740 | /* shouldn't fail above this point */ |
be69d00d | 4741 | pool->attrs = alloc_workqueue_attrs(); |
6ba94429 FW |
4742 | if (!pool->attrs) |
4743 | return -ENOMEM; | |
5de7a03c TH |
4744 | |
4745 | wqattrs_clear_for_pool(pool->attrs); | |
4746 | ||
6ba94429 | 4747 | return 0; |
226223ab TH |
4748 | } |
4749 | ||
669de8bd BVA |
4750 | #ifdef CONFIG_LOCKDEP |
4751 | static void wq_init_lockdep(struct workqueue_struct *wq) | |
4752 | { | |
4753 | char *lock_name; | |
4754 | ||
4755 | lockdep_register_key(&wq->key); | |
4756 | lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name); | |
4757 | if (!lock_name) | |
4758 | lock_name = wq->name; | |
69a106c0 QC |
4759 | |
4760 | wq->lock_name = lock_name; | |
669de8bd BVA |
4761 | lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0); |
4762 | } | |
4763 | ||
4764 | static void wq_unregister_lockdep(struct workqueue_struct *wq) | |
4765 | { | |
4766 | lockdep_unregister_key(&wq->key); | |
4767 | } | |
4768 | ||
4769 | static void wq_free_lockdep(struct workqueue_struct *wq) | |
4770 | { | |
4771 | if (wq->lock_name != wq->name) | |
4772 | kfree(wq->lock_name); | |
4773 | } | |
4774 | #else | |
4775 | static void wq_init_lockdep(struct workqueue_struct *wq) | |
4776 | { | |
4777 | } | |
4778 | ||
4779 | static void wq_unregister_lockdep(struct workqueue_struct *wq) | |
4780 | { | |
4781 | } | |
4782 | ||
4783 | static void wq_free_lockdep(struct workqueue_struct *wq) | |
4784 | { | |
4785 | } | |
4786 | #endif | |
4787 | ||
91ccc6e7 TH |
4788 | static void free_node_nr_active(struct wq_node_nr_active **nna_ar) |
4789 | { | |
4790 | int node; | |
4791 | ||
4792 | for_each_node(node) { | |
4793 | kfree(nna_ar[node]); | |
4794 | nna_ar[node] = NULL; | |
4795 | } | |
4796 | ||
4797 | kfree(nna_ar[nr_node_ids]); | |
4798 | nna_ar[nr_node_ids] = NULL; | |
4799 | } | |
4800 | ||
4801 | static void init_node_nr_active(struct wq_node_nr_active *nna) | |
4802 | { | |
c5f8cd6c | 4803 | nna->max = WQ_DFL_MIN_ACTIVE; |
91ccc6e7 | 4804 | atomic_set(&nna->nr, 0); |
5797b1c1 TH |
4805 | raw_spin_lock_init(&nna->lock); |
4806 | INIT_LIST_HEAD(&nna->pending_pwqs); | |
91ccc6e7 TH |
4807 | } |
4808 | ||
4809 | /* | |
4810 | * Each node's nr_active counter will be accessed mostly from its own node and | |
4811 | * should be allocated in the node. | |
4812 | */ | |
4813 | static int alloc_node_nr_active(struct wq_node_nr_active **nna_ar) | |
4814 | { | |
4815 | struct wq_node_nr_active *nna; | |
4816 | int node; | |
4817 | ||
4818 | for_each_node(node) { | |
4819 | nna = kzalloc_node(sizeof(*nna), GFP_KERNEL, node); | |
4820 | if (!nna) | |
4821 | goto err_free; | |
4822 | init_node_nr_active(nna); | |
4823 | nna_ar[node] = nna; | |
4824 | } | |
4825 | ||
4826 | /* [nr_node_ids] is used as the fallback */ | |
4827 | nna = kzalloc_node(sizeof(*nna), GFP_KERNEL, NUMA_NO_NODE); | |
4828 | if (!nna) | |
4829 | goto err_free; | |
4830 | init_node_nr_active(nna); | |
4831 | nna_ar[nr_node_ids] = nna; | |
4832 | ||
4833 | return 0; | |
4834 | ||
4835 | err_free: | |
4836 | free_node_nr_active(nna_ar); | |
4837 | return -ENOMEM; | |
4838 | } | |
4839 | ||
6ba94429 | 4840 | static void rcu_free_wq(struct rcu_head *rcu) |
226223ab | 4841 | { |
6ba94429 FW |
4842 | struct workqueue_struct *wq = |
4843 | container_of(rcu, struct workqueue_struct, rcu); | |
226223ab | 4844 | |
91ccc6e7 TH |
4845 | if (wq->flags & WQ_UNBOUND) |
4846 | free_node_nr_active(wq->node_nr_active); | |
4847 | ||
669de8bd | 4848 | wq_free_lockdep(wq); |
636b927e TH |
4849 | free_percpu(wq->cpu_pwq); |
4850 | free_workqueue_attrs(wq->unbound_attrs); | |
6ba94429 | 4851 | kfree(wq); |
226223ab TH |
4852 | } |
4853 | ||
6ba94429 | 4854 | static void rcu_free_pool(struct rcu_head *rcu) |
226223ab | 4855 | { |
6ba94429 | 4856 | struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); |
226223ab | 4857 | |
6ba94429 FW |
4858 | ida_destroy(&pool->worker_ida); |
4859 | free_workqueue_attrs(pool->attrs); | |
4860 | kfree(pool); | |
226223ab TH |
4861 | } |
4862 | ||
6ba94429 FW |
4863 | /** |
4864 | * put_unbound_pool - put a worker_pool | |
4865 | * @pool: worker_pool to put | |
4866 | * | |
24acfb71 | 4867 | * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU |
6ba94429 FW |
4868 | * safe manner. get_unbound_pool() calls this function on its failure path |
4869 | * and this function should be able to release pools which went through, | |
4870 | * successfully or not, init_worker_pool(). | |
4871 | * | |
4872 | * Should be called with wq_pool_mutex held. | |
4873 | */ | |
4874 | static void put_unbound_pool(struct worker_pool *pool) | |
226223ab | 4875 | { |
6ba94429 FW |
4876 | DECLARE_COMPLETION_ONSTACK(detach_completion); |
4877 | struct worker *worker; | |
9680540c | 4878 | LIST_HEAD(cull_list); |
e02b9312 | 4879 | |
6ba94429 | 4880 | lockdep_assert_held(&wq_pool_mutex); |
226223ab | 4881 | |
6ba94429 FW |
4882 | if (--pool->refcnt) |
4883 | return; | |
226223ab | 4884 | |
6ba94429 FW |
4885 | /* sanity checks */ |
4886 | if (WARN_ON(!(pool->cpu < 0)) || | |
4887 | WARN_ON(!list_empty(&pool->worklist))) | |
4888 | return; | |
226223ab | 4889 | |
6ba94429 FW |
4890 | /* release id and unhash */ |
4891 | if (pool->id >= 0) | |
4892 | idr_remove(&worker_pool_idr, pool->id); | |
4893 | hash_del(&pool->hash_node); | |
d55262c4 | 4894 | |
6ba94429 | 4895 | /* |
692b4825 TH |
4896 | * Become the manager and destroy all workers. This prevents |
4897 | * @pool's workers from blocking on attach_mutex. We're the last | |
4898 | * manager and @pool gets freed with the flag set. | |
9ab03be4 VS |
4899 | * |
4900 | * Having a concurrent manager is quite unlikely to happen as we can | |
4901 | * only get here with | |
4902 | * pwq->refcnt == pool->refcnt == 0 | |
4903 | * which implies no work queued to the pool, which implies no worker can | |
4904 | * become the manager. However a worker could have taken the role of | |
4905 | * manager before the refcnts dropped to 0, since maybe_create_worker() | |
4906 | * drops pool->lock | |
6ba94429 | 4907 | */ |
9ab03be4 VS |
4908 | while (true) { |
4909 | rcuwait_wait_event(&manager_wait, | |
4910 | !(pool->flags & POOL_MANAGER_ACTIVE), | |
4911 | TASK_UNINTERRUPTIBLE); | |
e02b9312 VS |
4912 | |
4913 | mutex_lock(&wq_pool_attach_mutex); | |
9ab03be4 VS |
4914 | raw_spin_lock_irq(&pool->lock); |
4915 | if (!(pool->flags & POOL_MANAGER_ACTIVE)) { | |
4916 | pool->flags |= POOL_MANAGER_ACTIVE; | |
4917 | break; | |
4918 | } | |
4919 | raw_spin_unlock_irq(&pool->lock); | |
e02b9312 | 4920 | mutex_unlock(&wq_pool_attach_mutex); |
9ab03be4 | 4921 | } |
692b4825 | 4922 | |
6ba94429 | 4923 | while ((worker = first_idle_worker(pool))) |
e02b9312 | 4924 | set_worker_dying(worker, &cull_list); |
6ba94429 | 4925 | WARN_ON(pool->nr_workers || pool->nr_idle); |
a9b8a985 | 4926 | raw_spin_unlock_irq(&pool->lock); |
d55262c4 | 4927 | |
e02b9312 VS |
4928 | wake_dying_workers(&cull_list); |
4929 | ||
4930 | if (!list_empty(&pool->workers) || !list_empty(&pool->dying_workers)) | |
6ba94429 | 4931 | pool->detach_completion = &detach_completion; |
1258fae7 | 4932 | mutex_unlock(&wq_pool_attach_mutex); |
226223ab | 4933 | |
6ba94429 FW |
4934 | if (pool->detach_completion) |
4935 | wait_for_completion(pool->detach_completion); | |
226223ab | 4936 | |
6ba94429 FW |
4937 | /* shut down the timers */ |
4938 | del_timer_sync(&pool->idle_timer); | |
3f959aa3 | 4939 | cancel_work_sync(&pool->idle_cull_work); |
6ba94429 | 4940 | del_timer_sync(&pool->mayday_timer); |
226223ab | 4941 | |
24acfb71 | 4942 | /* RCU protected to allow dereferences from get_work_pool() */ |
25b00775 | 4943 | call_rcu(&pool->rcu, rcu_free_pool); |
226223ab TH |
4944 | } |
4945 | ||
4946 | /** | |
6ba94429 FW |
4947 | * get_unbound_pool - get a worker_pool with the specified attributes |
4948 | * @attrs: the attributes of the worker_pool to get | |
226223ab | 4949 | * |
6ba94429 FW |
4950 | * Obtain a worker_pool which has the same attributes as @attrs, bump the |
4951 | * reference count and return it. If there already is a matching | |
4952 | * worker_pool, it will be used; otherwise, this function attempts to | |
4953 | * create a new one. | |
226223ab | 4954 | * |
6ba94429 | 4955 | * Should be called with wq_pool_mutex held. |
226223ab | 4956 | * |
6ba94429 FW |
4957 | * Return: On success, a worker_pool with the same attributes as @attrs. |
4958 | * On failure, %NULL. | |
226223ab | 4959 | */ |
6ba94429 | 4960 | static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) |
226223ab | 4961 | { |
84193c07 | 4962 | struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_NUMA]; |
6ba94429 FW |
4963 | u32 hash = wqattrs_hash(attrs); |
4964 | struct worker_pool *pool; | |
84193c07 | 4965 | int pod, node = NUMA_NO_NODE; |
226223ab | 4966 | |
6ba94429 | 4967 | lockdep_assert_held(&wq_pool_mutex); |
226223ab | 4968 | |
6ba94429 FW |
4969 | /* do we already have a matching pool? */ |
4970 | hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { | |
4971 | if (wqattrs_equal(pool->attrs, attrs)) { | |
4972 | pool->refcnt++; | |
4973 | return pool; | |
4974 | } | |
4975 | } | |
226223ab | 4976 | |
9546b29e | 4977 | /* If __pod_cpumask is contained inside a NUMA pod, that's our node */ |
84193c07 | 4978 | for (pod = 0; pod < pt->nr_pods; pod++) { |
9546b29e | 4979 | if (cpumask_subset(attrs->__pod_cpumask, pt->pod_cpus[pod])) { |
84193c07 TH |
4980 | node = pt->pod_node[pod]; |
4981 | break; | |
e2273584 XP |
4982 | } |
4983 | } | |
4984 | ||
6ba94429 | 4985 | /* nope, create a new one */ |
84193c07 | 4986 | pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, node); |
6ba94429 FW |
4987 | if (!pool || init_worker_pool(pool) < 0) |
4988 | goto fail; | |
4989 | ||
84193c07 | 4990 | pool->node = node; |
5de7a03c TH |
4991 | copy_workqueue_attrs(pool->attrs, attrs); |
4992 | wqattrs_clear_for_pool(pool->attrs); | |
226223ab | 4993 | |
6ba94429 FW |
4994 | if (worker_pool_assign_id(pool) < 0) |
4995 | goto fail; | |
226223ab | 4996 | |
6ba94429 | 4997 | /* create and start the initial worker */ |
3347fa09 | 4998 | if (wq_online && !create_worker(pool)) |
6ba94429 | 4999 | goto fail; |
226223ab | 5000 | |
6ba94429 FW |
5001 | /* install */ |
5002 | hash_add(unbound_pool_hash, &pool->hash_node, hash); | |
226223ab | 5003 | |
6ba94429 FW |
5004 | return pool; |
5005 | fail: | |
5006 | if (pool) | |
5007 | put_unbound_pool(pool); | |
5008 | return NULL; | |
226223ab | 5009 | } |
226223ab | 5010 | |
6ba94429 | 5011 | static void rcu_free_pwq(struct rcu_head *rcu) |
7a4e344c | 5012 | { |
6ba94429 FW |
5013 | kmem_cache_free(pwq_cache, |
5014 | container_of(rcu, struct pool_workqueue, rcu)); | |
7a4e344c TH |
5015 | } |
5016 | ||
6ba94429 | 5017 | /* |
967b494e TH |
5018 | * Scheduled on pwq_release_worker by put_pwq() when an unbound pwq hits zero |
5019 | * refcnt and needs to be destroyed. | |
7a4e344c | 5020 | */ |
687a9aa5 | 5021 | static void pwq_release_workfn(struct kthread_work *work) |
7a4e344c | 5022 | { |
6ba94429 | 5023 | struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, |
687a9aa5 | 5024 | release_work); |
6ba94429 FW |
5025 | struct workqueue_struct *wq = pwq->wq; |
5026 | struct worker_pool *pool = pwq->pool; | |
b42b0bdd | 5027 | bool is_last = false; |
7a4e344c | 5028 | |
b42b0bdd | 5029 | /* |
687a9aa5 | 5030 | * When @pwq is not linked, it doesn't hold any reference to the |
b42b0bdd YY |
5031 | * @wq, and @wq is invalid to access. |
5032 | */ | |
5033 | if (!list_empty(&pwq->pwqs_node)) { | |
b42b0bdd YY |
5034 | mutex_lock(&wq->mutex); |
5035 | list_del_rcu(&pwq->pwqs_node); | |
5036 | is_last = list_empty(&wq->pwqs); | |
4c065dbc WL |
5037 | |
5038 | /* | |
5039 | * For ordered workqueue with a plugged dfl_pwq, restart it now. | |
5040 | */ | |
5041 | if (!is_last && (wq->flags & __WQ_ORDERED)) | |
5042 | unplug_oldest_pwq(wq); | |
5043 | ||
b42b0bdd YY |
5044 | mutex_unlock(&wq->mutex); |
5045 | } | |
6ba94429 | 5046 | |
687a9aa5 TH |
5047 | if (wq->flags & WQ_UNBOUND) { |
5048 | mutex_lock(&wq_pool_mutex); | |
5049 | put_unbound_pool(pool); | |
5050 | mutex_unlock(&wq_pool_mutex); | |
5051 | } | |
6ba94429 | 5052 | |
5797b1c1 TH |
5053 | if (!list_empty(&pwq->pending_node)) { |
5054 | struct wq_node_nr_active *nna = | |
5055 | wq_node_nr_active(pwq->wq, pwq->pool->node); | |
5056 | ||
5057 | raw_spin_lock_irq(&nna->lock); | |
5058 | list_del_init(&pwq->pending_node); | |
5059 | raw_spin_unlock_irq(&nna->lock); | |
5060 | } | |
5061 | ||
25b00775 | 5062 | call_rcu(&pwq->rcu, rcu_free_pwq); |
7a4e344c | 5063 | |
2865a8fb | 5064 | /* |
6ba94429 FW |
5065 | * If we're the last pwq going away, @wq is already dead and no one |
5066 | * is gonna access it anymore. Schedule RCU free. | |
2865a8fb | 5067 | */ |
669de8bd BVA |
5068 | if (is_last) { |
5069 | wq_unregister_lockdep(wq); | |
25b00775 | 5070 | call_rcu(&wq->rcu, rcu_free_wq); |
669de8bd | 5071 | } |
29c91e99 TH |
5072 | } |
5073 | ||
67dc8325 | 5074 | /* initialize newly allocated @pwq which is associated with @wq and @pool */ |
6ba94429 FW |
5075 | static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, |
5076 | struct worker_pool *pool) | |
29c91e99 | 5077 | { |
e9a8e01f | 5078 | BUG_ON((unsigned long)pwq & ~WORK_STRUCT_PWQ_MASK); |
29c91e99 | 5079 | |
6ba94429 FW |
5080 | memset(pwq, 0, sizeof(*pwq)); |
5081 | ||
5082 | pwq->pool = pool; | |
5083 | pwq->wq = wq; | |
5084 | pwq->flush_color = -1; | |
5085 | pwq->refcnt = 1; | |
f97a4a1a | 5086 | INIT_LIST_HEAD(&pwq->inactive_works); |
5797b1c1 | 5087 | INIT_LIST_HEAD(&pwq->pending_node); |
6ba94429 FW |
5088 | INIT_LIST_HEAD(&pwq->pwqs_node); |
5089 | INIT_LIST_HEAD(&pwq->mayday_node); | |
687a9aa5 | 5090 | kthread_init_work(&pwq->release_work, pwq_release_workfn); |
29c91e99 TH |
5091 | } |
5092 | ||
6ba94429 FW |
5093 | /* sync @pwq with the current state of its associated wq and link it */ |
5094 | static void link_pwq(struct pool_workqueue *pwq) | |
29c91e99 | 5095 | { |
6ba94429 | 5096 | struct workqueue_struct *wq = pwq->wq; |
29c91e99 | 5097 | |
6ba94429 | 5098 | lockdep_assert_held(&wq->mutex); |
a892cacc | 5099 | |
6ba94429 FW |
5100 | /* may be called multiple times, ignore if already linked */ |
5101 | if (!list_empty(&pwq->pwqs_node)) | |
29c91e99 | 5102 | return; |
29c91e99 | 5103 | |
6ba94429 FW |
5104 | /* set the matching work_color */ |
5105 | pwq->work_color = wq->work_color; | |
29c91e99 | 5106 | |
6ba94429 | 5107 | /* link in @pwq */ |
26fb7e3d | 5108 | list_add_tail_rcu(&pwq->pwqs_node, &wq->pwqs); |
6ba94429 | 5109 | } |
29c91e99 | 5110 | |
6ba94429 FW |
5111 | /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ |
5112 | static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, | |
5113 | const struct workqueue_attrs *attrs) | |
5114 | { | |
5115 | struct worker_pool *pool; | |
5116 | struct pool_workqueue *pwq; | |
60f5a4bc | 5117 | |
6ba94429 | 5118 | lockdep_assert_held(&wq_pool_mutex); |
60f5a4bc | 5119 | |
6ba94429 FW |
5120 | pool = get_unbound_pool(attrs); |
5121 | if (!pool) | |
5122 | return NULL; | |
60f5a4bc | 5123 | |
6ba94429 FW |
5124 | pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); |
5125 | if (!pwq) { | |
5126 | put_unbound_pool(pool); | |
5127 | return NULL; | |
5128 | } | |
29c91e99 | 5129 | |
6ba94429 FW |
5130 | init_pwq(pwq, wq, pool); |
5131 | return pwq; | |
5132 | } | |
29c91e99 | 5133 | |
29c91e99 | 5134 | /** |
fef59c9c | 5135 | * wq_calc_pod_cpumask - calculate a wq_attrs' cpumask for a pod |
042f7df1 | 5136 | * @attrs: the wq_attrs of the default pwq of the target workqueue |
84193c07 | 5137 | * @cpu: the target CPU |
6ba94429 | 5138 | * @cpu_going_down: if >= 0, the CPU to consider as offline |
29c91e99 | 5139 | * |
fef59c9c TH |
5140 | * Calculate the cpumask a workqueue with @attrs should use on @pod. If |
5141 | * @cpu_going_down is >= 0, that cpu is considered offline during calculation. | |
9546b29e | 5142 | * The result is stored in @attrs->__pod_cpumask. |
a892cacc | 5143 | * |
fef59c9c TH |
5144 | * If pod affinity is not enabled, @attrs->cpumask is always used. If enabled |
5145 | * and @pod has online CPUs requested by @attrs, the returned cpumask is the | |
5146 | * intersection of the possible CPUs of @pod and @attrs->cpumask. | |
d185af30 | 5147 | * |
fef59c9c | 5148 | * The caller is responsible for ensuring that the cpumask of @pod stays stable. |
29c91e99 | 5149 | */ |
9546b29e TH |
5150 | static void wq_calc_pod_cpumask(struct workqueue_attrs *attrs, int cpu, |
5151 | int cpu_going_down) | |
29c91e99 | 5152 | { |
84193c07 TH |
5153 | const struct wq_pod_type *pt = wqattrs_pod_type(attrs); |
5154 | int pod = pt->cpu_pod[cpu]; | |
29c91e99 | 5155 | |
fef59c9c | 5156 | /* does @pod have any online CPUs @attrs wants? */ |
9546b29e TH |
5157 | cpumask_and(attrs->__pod_cpumask, pt->pod_cpus[pod], attrs->cpumask); |
5158 | cpumask_and(attrs->__pod_cpumask, attrs->__pod_cpumask, cpu_online_mask); | |
6ba94429 | 5159 | if (cpu_going_down >= 0) |
9546b29e | 5160 | cpumask_clear_cpu(cpu_going_down, attrs->__pod_cpumask); |
29c91e99 | 5161 | |
9546b29e TH |
5162 | if (cpumask_empty(attrs->__pod_cpumask)) { |
5163 | cpumask_copy(attrs->__pod_cpumask, attrs->cpumask); | |
84193c07 TH |
5164 | return; |
5165 | } | |
4c16bd32 | 5166 | |
fef59c9c | 5167 | /* yeap, return possible CPUs in @pod that @attrs wants */ |
9546b29e | 5168 | cpumask_and(attrs->__pod_cpumask, attrs->cpumask, pt->pod_cpus[pod]); |
1ad0f0a7 | 5169 | |
9546b29e | 5170 | if (cpumask_empty(attrs->__pod_cpumask)) |
1ad0f0a7 MB |
5171 | pr_warn_once("WARNING: workqueue cpumask: online intersect > " |
5172 | "possible intersect\n"); | |
4c16bd32 TH |
5173 | } |
5174 | ||
9f66cff2 | 5175 | /* install @pwq into @wq and return the old pwq, @cpu < 0 for dfl_pwq */ |
636b927e TH |
5176 | static struct pool_workqueue *install_unbound_pwq(struct workqueue_struct *wq, |
5177 | int cpu, struct pool_workqueue *pwq) | |
1befcf30 | 5178 | { |
9f66cff2 | 5179 | struct pool_workqueue __rcu **slot = unbound_pwq_slot(wq, cpu); |
1befcf30 TH |
5180 | struct pool_workqueue *old_pwq; |
5181 | ||
5b95e1af | 5182 | lockdep_assert_held(&wq_pool_mutex); |
1befcf30 TH |
5183 | lockdep_assert_held(&wq->mutex); |
5184 | ||
5185 | /* link_pwq() can handle duplicate calls */ | |
5186 | link_pwq(pwq); | |
5187 | ||
9f66cff2 TH |
5188 | old_pwq = rcu_access_pointer(*slot); |
5189 | rcu_assign_pointer(*slot, pwq); | |
1befcf30 TH |
5190 | return old_pwq; |
5191 | } | |
5192 | ||
2d5f0764 LJ |
5193 | /* context to store the prepared attrs & pwqs before applying */ |
5194 | struct apply_wqattrs_ctx { | |
5195 | struct workqueue_struct *wq; /* target workqueue */ | |
5196 | struct workqueue_attrs *attrs; /* attrs to apply */ | |
042f7df1 | 5197 | struct list_head list; /* queued for batching commit */ |
2d5f0764 LJ |
5198 | struct pool_workqueue *dfl_pwq; |
5199 | struct pool_workqueue *pwq_tbl[]; | |
5200 | }; | |
5201 | ||
5202 | /* free the resources after success or abort */ | |
5203 | static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx) | |
5204 | { | |
5205 | if (ctx) { | |
636b927e | 5206 | int cpu; |
2d5f0764 | 5207 | |
636b927e TH |
5208 | for_each_possible_cpu(cpu) |
5209 | put_pwq_unlocked(ctx->pwq_tbl[cpu]); | |
2d5f0764 LJ |
5210 | put_pwq_unlocked(ctx->dfl_pwq); |
5211 | ||
5212 | free_workqueue_attrs(ctx->attrs); | |
5213 | ||
5214 | kfree(ctx); | |
5215 | } | |
5216 | } | |
5217 | ||
5218 | /* allocate the attrs and pwqs for later installation */ | |
5219 | static struct apply_wqattrs_ctx * | |
5220 | apply_wqattrs_prepare(struct workqueue_struct *wq, | |
99c621ef LJ |
5221 | const struct workqueue_attrs *attrs, |
5222 | const cpumask_var_t unbound_cpumask) | |
9e8cd2f5 | 5223 | { |
2d5f0764 | 5224 | struct apply_wqattrs_ctx *ctx; |
9546b29e | 5225 | struct workqueue_attrs *new_attrs; |
636b927e | 5226 | int cpu; |
9e8cd2f5 | 5227 | |
2d5f0764 | 5228 | lockdep_assert_held(&wq_pool_mutex); |
9e8cd2f5 | 5229 | |
84193c07 TH |
5230 | if (WARN_ON(attrs->affn_scope < 0 || |
5231 | attrs->affn_scope >= WQ_AFFN_NR_TYPES)) | |
5232 | return ERR_PTR(-EINVAL); | |
5233 | ||
636b927e | 5234 | ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_cpu_ids), GFP_KERNEL); |
8719dcea | 5235 | |
be69d00d | 5236 | new_attrs = alloc_workqueue_attrs(); |
9546b29e | 5237 | if (!ctx || !new_attrs) |
2d5f0764 | 5238 | goto out_free; |
13e2e556 | 5239 | |
4c16bd32 TH |
5240 | /* |
5241 | * If something goes wrong during CPU up/down, we'll fall back to | |
5242 | * the default pwq covering whole @attrs->cpumask. Always create | |
5243 | * it even if we don't use it immediately. | |
5244 | */ | |
0f36ee24 TH |
5245 | copy_workqueue_attrs(new_attrs, attrs); |
5246 | wqattrs_actualize_cpumask(new_attrs, unbound_cpumask); | |
9546b29e | 5247 | cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask); |
2d5f0764 LJ |
5248 | ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs); |
5249 | if (!ctx->dfl_pwq) | |
5250 | goto out_free; | |
4c16bd32 | 5251 | |
636b927e | 5252 | for_each_possible_cpu(cpu) { |
af73f5c9 | 5253 | if (new_attrs->ordered) { |
2d5f0764 | 5254 | ctx->dfl_pwq->refcnt++; |
636b927e TH |
5255 | ctx->pwq_tbl[cpu] = ctx->dfl_pwq; |
5256 | } else { | |
9546b29e TH |
5257 | wq_calc_pod_cpumask(new_attrs, cpu, -1); |
5258 | ctx->pwq_tbl[cpu] = alloc_unbound_pwq(wq, new_attrs); | |
636b927e TH |
5259 | if (!ctx->pwq_tbl[cpu]) |
5260 | goto out_free; | |
4c16bd32 TH |
5261 | } |
5262 | } | |
5263 | ||
042f7df1 LJ |
5264 | /* save the user configured attrs and sanitize it. */ |
5265 | copy_workqueue_attrs(new_attrs, attrs); | |
5266 | cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); | |
9546b29e | 5267 | cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask); |
2d5f0764 | 5268 | ctx->attrs = new_attrs; |
042f7df1 | 5269 | |
4c065dbc WL |
5270 | /* |
5271 | * For initialized ordered workqueues, there should only be one pwq | |
5272 | * (dfl_pwq). Set the plugged flag of ctx->dfl_pwq to suspend execution | |
5273 | * of newly queued work items until execution of older work items in | |
5274 | * the old pwq's have completed. | |
5275 | */ | |
5276 | if ((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs)) | |
5277 | ctx->dfl_pwq->plugged = true; | |
5278 | ||
2d5f0764 | 5279 | ctx->wq = wq; |
2d5f0764 LJ |
5280 | return ctx; |
5281 | ||
5282 | out_free: | |
2d5f0764 LJ |
5283 | free_workqueue_attrs(new_attrs); |
5284 | apply_wqattrs_cleanup(ctx); | |
84193c07 | 5285 | return ERR_PTR(-ENOMEM); |
2d5f0764 LJ |
5286 | } |
5287 | ||
5288 | /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */ | |
5289 | static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx) | |
5290 | { | |
636b927e | 5291 | int cpu; |
9e8cd2f5 | 5292 | |
4c16bd32 | 5293 | /* all pwqs have been created successfully, let's install'em */ |
2d5f0764 | 5294 | mutex_lock(&ctx->wq->mutex); |
a892cacc | 5295 | |
2d5f0764 | 5296 | copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs); |
4c16bd32 | 5297 | |
9f66cff2 | 5298 | /* save the previous pwqs and install the new ones */ |
636b927e TH |
5299 | for_each_possible_cpu(cpu) |
5300 | ctx->pwq_tbl[cpu] = install_unbound_pwq(ctx->wq, cpu, | |
5301 | ctx->pwq_tbl[cpu]); | |
9f66cff2 | 5302 | ctx->dfl_pwq = install_unbound_pwq(ctx->wq, -1, ctx->dfl_pwq); |
f147f29e | 5303 | |
5797b1c1 TH |
5304 | /* update node_nr_active->max */ |
5305 | wq_update_node_max_active(ctx->wq, -1); | |
5306 | ||
d64f2fa0 JL |
5307 | /* rescuer needs to respect wq cpumask changes */ |
5308 | if (ctx->wq->rescuer) | |
5309 | set_cpus_allowed_ptr(ctx->wq->rescuer->task, | |
5310 | unbound_effective_cpumask(ctx->wq)); | |
5311 | ||
2d5f0764 LJ |
5312 | mutex_unlock(&ctx->wq->mutex); |
5313 | } | |
9e8cd2f5 | 5314 | |
a0111cf6 LJ |
5315 | static int apply_workqueue_attrs_locked(struct workqueue_struct *wq, |
5316 | const struct workqueue_attrs *attrs) | |
2d5f0764 LJ |
5317 | { |
5318 | struct apply_wqattrs_ctx *ctx; | |
4c16bd32 | 5319 | |
2d5f0764 LJ |
5320 | /* only unbound workqueues can change attributes */ |
5321 | if (WARN_ON(!(wq->flags & WQ_UNBOUND))) | |
5322 | return -EINVAL; | |
13e2e556 | 5323 | |
99c621ef | 5324 | ctx = apply_wqattrs_prepare(wq, attrs, wq_unbound_cpumask); |
84193c07 TH |
5325 | if (IS_ERR(ctx)) |
5326 | return PTR_ERR(ctx); | |
2d5f0764 LJ |
5327 | |
5328 | /* the ctx has been prepared successfully, let's commit it */ | |
6201171e | 5329 | apply_wqattrs_commit(ctx); |
2d5f0764 LJ |
5330 | apply_wqattrs_cleanup(ctx); |
5331 | ||
6201171e | 5332 | return 0; |
9e8cd2f5 TH |
5333 | } |
5334 | ||
a0111cf6 LJ |
5335 | /** |
5336 | * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue | |
5337 | * @wq: the target workqueue | |
5338 | * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() | |
5339 | * | |
fef59c9c TH |
5340 | * Apply @attrs to an unbound workqueue @wq. Unless disabled, this function maps |
5341 | * a separate pwq to each CPU pod with possibles CPUs in @attrs->cpumask so that | |
5342 | * work items are affine to the pod it was issued on. Older pwqs are released as | |
5343 | * in-flight work items finish. Note that a work item which repeatedly requeues | |
5344 | * itself back-to-back will stay on its current pwq. | |
a0111cf6 LJ |
5345 | * |
5346 | * Performs GFP_KERNEL allocations. | |
5347 | * | |
ffd8bea8 | 5348 | * Assumes caller has CPU hotplug read exclusion, i.e. cpus_read_lock(). |
509b3204 | 5349 | * |
a0111cf6 LJ |
5350 | * Return: 0 on success and -errno on failure. |
5351 | */ | |
513c98d0 | 5352 | int apply_workqueue_attrs(struct workqueue_struct *wq, |
a0111cf6 LJ |
5353 | const struct workqueue_attrs *attrs) |
5354 | { | |
5355 | int ret; | |
5356 | ||
509b3204 DJ |
5357 | lockdep_assert_cpus_held(); |
5358 | ||
5359 | mutex_lock(&wq_pool_mutex); | |
a0111cf6 | 5360 | ret = apply_workqueue_attrs_locked(wq, attrs); |
509b3204 | 5361 | mutex_unlock(&wq_pool_mutex); |
a0111cf6 LJ |
5362 | |
5363 | return ret; | |
5364 | } | |
5365 | ||
4c16bd32 | 5366 | /** |
fef59c9c | 5367 | * wq_update_pod - update pod affinity of a wq for CPU hot[un]plug |
4c16bd32 | 5368 | * @wq: the target workqueue |
4cbfd3de TH |
5369 | * @cpu: the CPU to update pool association for |
5370 | * @hotplug_cpu: the CPU coming up or going down | |
4c16bd32 TH |
5371 | * @online: whether @cpu is coming up or going down |
5372 | * | |
5373 | * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and | |
fef59c9c | 5374 | * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update pod affinity of |
4c16bd32 TH |
5375 | * @wq accordingly. |
5376 | * | |
fef59c9c TH |
5377 | * |
5378 | * If pod affinity can't be adjusted due to memory allocation failure, it falls | |
5379 | * back to @wq->dfl_pwq which may not be optimal but is always correct. | |
5380 | * | |
5381 | * Note that when the last allowed CPU of a pod goes offline for a workqueue | |
5382 | * with a cpumask spanning multiple pods, the workers which were already | |
5383 | * executing the work items for the workqueue will lose their CPU affinity and | |
5384 | * may execute on any CPU. This is similar to how per-cpu workqueues behave on | |
5385 | * CPU_DOWN. If a workqueue user wants strict affinity, it's the user's | |
5386 | * responsibility to flush the work item from CPU_DOWN_PREPARE. | |
4c16bd32 | 5387 | */ |
fef59c9c TH |
5388 | static void wq_update_pod(struct workqueue_struct *wq, int cpu, |
5389 | int hotplug_cpu, bool online) | |
4c16bd32 | 5390 | { |
4cbfd3de | 5391 | int off_cpu = online ? -1 : hotplug_cpu; |
4c16bd32 TH |
5392 | struct pool_workqueue *old_pwq = NULL, *pwq; |
5393 | struct workqueue_attrs *target_attrs; | |
4c16bd32 TH |
5394 | |
5395 | lockdep_assert_held(&wq_pool_mutex); | |
5396 | ||
84193c07 | 5397 | if (!(wq->flags & WQ_UNBOUND) || wq->unbound_attrs->ordered) |
4c16bd32 TH |
5398 | return; |
5399 | ||
5400 | /* | |
5401 | * We don't wanna alloc/free wq_attrs for each wq for each CPU. | |
5402 | * Let's use a preallocated one. The following buf is protected by | |
5403 | * CPU hotplug exclusion. | |
5404 | */ | |
fef59c9c | 5405 | target_attrs = wq_update_pod_attrs_buf; |
4c16bd32 | 5406 | |
4c16bd32 | 5407 | copy_workqueue_attrs(target_attrs, wq->unbound_attrs); |
0f36ee24 | 5408 | wqattrs_actualize_cpumask(target_attrs, wq_unbound_cpumask); |
4c16bd32 | 5409 | |
636b927e | 5410 | /* nothing to do if the target cpumask matches the current pwq */ |
9546b29e | 5411 | wq_calc_pod_cpumask(target_attrs, cpu, off_cpu); |
9f66cff2 | 5412 | if (wqattrs_equal(target_attrs, unbound_pwq(wq, cpu)->pool->attrs)) |
636b927e | 5413 | return; |
4c16bd32 | 5414 | |
4c16bd32 TH |
5415 | /* create a new pwq */ |
5416 | pwq = alloc_unbound_pwq(wq, target_attrs); | |
5417 | if (!pwq) { | |
fef59c9c | 5418 | pr_warn("workqueue: allocation failed while updating CPU pod affinity of \"%s\"\n", |
2d916033 | 5419 | wq->name); |
77f300b1 | 5420 | goto use_dfl_pwq; |
4c16bd32 TH |
5421 | } |
5422 | ||
f7142ed4 | 5423 | /* Install the new pwq. */ |
4c16bd32 | 5424 | mutex_lock(&wq->mutex); |
636b927e | 5425 | old_pwq = install_unbound_pwq(wq, cpu, pwq); |
4c16bd32 TH |
5426 | goto out_unlock; |
5427 | ||
5428 | use_dfl_pwq: | |
f7142ed4 | 5429 | mutex_lock(&wq->mutex); |
9f66cff2 TH |
5430 | pwq = unbound_pwq(wq, -1); |
5431 | raw_spin_lock_irq(&pwq->pool->lock); | |
5432 | get_pwq(pwq); | |
5433 | raw_spin_unlock_irq(&pwq->pool->lock); | |
5434 | old_pwq = install_unbound_pwq(wq, cpu, pwq); | |
4c16bd32 TH |
5435 | out_unlock: |
5436 | mutex_unlock(&wq->mutex); | |
5437 | put_pwq_unlocked(old_pwq); | |
5438 | } | |
5439 | ||
30cdf249 | 5440 | static int alloc_and_link_pwqs(struct workqueue_struct *wq) |
0f900049 | 5441 | { |
49e3cf44 | 5442 | bool highpri = wq->flags & WQ_HIGHPRI; |
8a2b7538 | 5443 | int cpu, ret; |
30cdf249 | 5444 | |
636b927e TH |
5445 | wq->cpu_pwq = alloc_percpu(struct pool_workqueue *); |
5446 | if (!wq->cpu_pwq) | |
5447 | goto enomem; | |
30cdf249 | 5448 | |
636b927e | 5449 | if (!(wq->flags & WQ_UNBOUND)) { |
30cdf249 | 5450 | for_each_possible_cpu(cpu) { |
4cb1ef64 TH |
5451 | struct pool_workqueue **pwq_p; |
5452 | struct worker_pool __percpu *pools; | |
5453 | struct worker_pool *pool; | |
5454 | ||
5455 | if (wq->flags & WQ_BH) | |
5456 | pools = bh_worker_pools; | |
5457 | else | |
5458 | pools = cpu_worker_pools; | |
5459 | ||
5460 | pool = &(per_cpu_ptr(pools, cpu)[highpri]); | |
5461 | pwq_p = per_cpu_ptr(wq->cpu_pwq, cpu); | |
687a9aa5 TH |
5462 | |
5463 | *pwq_p = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, | |
5464 | pool->node); | |
5465 | if (!*pwq_p) | |
5466 | goto enomem; | |
f3421797 | 5467 | |
687a9aa5 | 5468 | init_pwq(*pwq_p, wq, pool); |
f147f29e TH |
5469 | |
5470 | mutex_lock(&wq->mutex); | |
687a9aa5 | 5471 | link_pwq(*pwq_p); |
f147f29e | 5472 | mutex_unlock(&wq->mutex); |
30cdf249 | 5473 | } |
9e8cd2f5 | 5474 | return 0; |
509b3204 DJ |
5475 | } |
5476 | ||
ffd8bea8 | 5477 | cpus_read_lock(); |
509b3204 | 5478 | if (wq->flags & __WQ_ORDERED) { |
9f66cff2 TH |
5479 | struct pool_workqueue *dfl_pwq; |
5480 | ||
8a2b7538 TH |
5481 | ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]); |
5482 | /* there should only be single pwq for ordering guarantee */ | |
9f66cff2 TH |
5483 | dfl_pwq = rcu_access_pointer(wq->dfl_pwq); |
5484 | WARN(!ret && (wq->pwqs.next != &dfl_pwq->pwqs_node || | |
5485 | wq->pwqs.prev != &dfl_pwq->pwqs_node), | |
8a2b7538 | 5486 | "ordering guarantee broken for workqueue %s\n", wq->name); |
30cdf249 | 5487 | } else { |
509b3204 | 5488 | ret = apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); |
30cdf249 | 5489 | } |
ffd8bea8 | 5490 | cpus_read_unlock(); |
509b3204 | 5491 | |
64344553 Z |
5492 | /* for unbound pwq, flush the pwq_release_worker ensures that the |
5493 | * pwq_release_workfn() completes before calling kfree(wq). | |
5494 | */ | |
5495 | if (ret) | |
5496 | kthread_flush_worker(pwq_release_worker); | |
5497 | ||
509b3204 | 5498 | return ret; |
687a9aa5 TH |
5499 | |
5500 | enomem: | |
5501 | if (wq->cpu_pwq) { | |
7b42f401 Z |
5502 | for_each_possible_cpu(cpu) { |
5503 | struct pool_workqueue *pwq = *per_cpu_ptr(wq->cpu_pwq, cpu); | |
5504 | ||
5505 | if (pwq) | |
5506 | kmem_cache_free(pwq_cache, pwq); | |
5507 | } | |
687a9aa5 TH |
5508 | free_percpu(wq->cpu_pwq); |
5509 | wq->cpu_pwq = NULL; | |
5510 | } | |
5511 | return -ENOMEM; | |
0f900049 TH |
5512 | } |
5513 | ||
f3421797 TH |
5514 | static int wq_clamp_max_active(int max_active, unsigned int flags, |
5515 | const char *name) | |
b71ab8c2 | 5516 | { |
636b927e | 5517 | if (max_active < 1 || max_active > WQ_MAX_ACTIVE) |
044c782c | 5518 | pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", |
636b927e | 5519 | max_active, name, 1, WQ_MAX_ACTIVE); |
b71ab8c2 | 5520 | |
636b927e | 5521 | return clamp_val(max_active, 1, WQ_MAX_ACTIVE); |
b71ab8c2 TH |
5522 | } |
5523 | ||
983c7515 TH |
5524 | /* |
5525 | * Workqueues which may be used during memory reclaim should have a rescuer | |
5526 | * to guarantee forward progress. | |
5527 | */ | |
5528 | static int init_rescuer(struct workqueue_struct *wq) | |
5529 | { | |
5530 | struct worker *rescuer; | |
b92b36ea | 5531 | int ret; |
983c7515 TH |
5532 | |
5533 | if (!(wq->flags & WQ_MEM_RECLAIM)) | |
5534 | return 0; | |
5535 | ||
5536 | rescuer = alloc_worker(NUMA_NO_NODE); | |
4c0736a7 PM |
5537 | if (!rescuer) { |
5538 | pr_err("workqueue: Failed to allocate a rescuer for wq \"%s\"\n", | |
5539 | wq->name); | |
983c7515 | 5540 | return -ENOMEM; |
4c0736a7 | 5541 | } |
983c7515 TH |
5542 | |
5543 | rescuer->rescue_wq = wq; | |
b6a46f72 | 5544 | rescuer->task = kthread_create(rescuer_thread, rescuer, "kworker/R-%s", wq->name); |
f187b697 | 5545 | if (IS_ERR(rescuer->task)) { |
b92b36ea | 5546 | ret = PTR_ERR(rescuer->task); |
4c0736a7 PM |
5547 | pr_err("workqueue: Failed to create a rescuer kthread for wq \"%s\": %pe", |
5548 | wq->name, ERR_PTR(ret)); | |
983c7515 | 5549 | kfree(rescuer); |
b92b36ea | 5550 | return ret; |
983c7515 TH |
5551 | } |
5552 | ||
5553 | wq->rescuer = rescuer; | |
85f0ab43 | 5554 | if (wq->flags & WQ_UNBOUND) |
49584bb8 | 5555 | kthread_bind_mask(rescuer->task, wq_unbound_cpumask); |
85f0ab43 JL |
5556 | else |
5557 | kthread_bind_mask(rescuer->task, cpu_possible_mask); | |
983c7515 TH |
5558 | wake_up_process(rescuer->task); |
5559 | ||
5560 | return 0; | |
5561 | } | |
5562 | ||
a045a272 TH |
5563 | /** |
5564 | * wq_adjust_max_active - update a wq's max_active to the current setting | |
5565 | * @wq: target workqueue | |
5566 | * | |
5567 | * If @wq isn't freezing, set @wq->max_active to the saved_max_active and | |
5568 | * activate inactive work items accordingly. If @wq is freezing, clear | |
5569 | * @wq->max_active to zero. | |
5570 | */ | |
5571 | static void wq_adjust_max_active(struct workqueue_struct *wq) | |
5572 | { | |
c5404d4e | 5573 | bool activated; |
5797b1c1 | 5574 | int new_max, new_min; |
a045a272 TH |
5575 | |
5576 | lockdep_assert_held(&wq->mutex); | |
5577 | ||
5578 | if ((wq->flags & WQ_FREEZABLE) && workqueue_freezing) { | |
5797b1c1 TH |
5579 | new_max = 0; |
5580 | new_min = 0; | |
5581 | } else { | |
5582 | new_max = wq->saved_max_active; | |
5583 | new_min = wq->saved_min_active; | |
a045a272 TH |
5584 | } |
5585 | ||
5797b1c1 | 5586 | if (wq->max_active == new_max && wq->min_active == new_min) |
a045a272 TH |
5587 | return; |
5588 | ||
5589 | /* | |
5797b1c1 | 5590 | * Update @wq->max/min_active and then kick inactive work items if more |
a045a272 TH |
5591 | * active work items are allowed. This doesn't break work item ordering |
5592 | * because new work items are always queued behind existing inactive | |
5593 | * work items if there are any. | |
5594 | */ | |
5797b1c1 TH |
5595 | WRITE_ONCE(wq->max_active, new_max); |
5596 | WRITE_ONCE(wq->min_active, new_min); | |
5597 | ||
5598 | if (wq->flags & WQ_UNBOUND) | |
5599 | wq_update_node_max_active(wq, -1); | |
5600 | ||
5601 | if (new_max == 0) | |
5602 | return; | |
a045a272 | 5603 | |
c5404d4e TH |
5604 | /* |
5605 | * Round-robin through pwq's activating the first inactive work item | |
5606 | * until max_active is filled. | |
5607 | */ | |
5608 | do { | |
5609 | struct pool_workqueue *pwq; | |
a045a272 | 5610 | |
c5404d4e TH |
5611 | activated = false; |
5612 | for_each_pwq(pwq, wq) { | |
c26e2f2e | 5613 | unsigned long irq_flags; |
a045a272 | 5614 | |
c5404d4e | 5615 | /* can be called during early boot w/ irq disabled */ |
c26e2f2e | 5616 | raw_spin_lock_irqsave(&pwq->pool->lock, irq_flags); |
5797b1c1 | 5617 | if (pwq_activate_first_inactive(pwq, true)) { |
c5404d4e TH |
5618 | activated = true; |
5619 | kick_pool(pwq->pool); | |
5620 | } | |
c26e2f2e | 5621 | raw_spin_unlock_irqrestore(&pwq->pool->lock, irq_flags); |
c5404d4e TH |
5622 | } |
5623 | } while (activated); | |
a045a272 TH |
5624 | } |
5625 | ||
a2775bbc | 5626 | __printf(1, 4) |
669de8bd BVA |
5627 | struct workqueue_struct *alloc_workqueue(const char *fmt, |
5628 | unsigned int flags, | |
5629 | int max_active, ...) | |
1da177e4 | 5630 | { |
ecf6881f | 5631 | va_list args; |
1da177e4 | 5632 | struct workqueue_struct *wq; |
91ccc6e7 TH |
5633 | size_t wq_size; |
5634 | int name_len; | |
b196be89 | 5635 | |
4cb1ef64 TH |
5636 | if (flags & WQ_BH) { |
5637 | if (WARN_ON_ONCE(flags & ~__WQ_BH_ALLOWS)) | |
5638 | return NULL; | |
5639 | if (WARN_ON_ONCE(max_active)) | |
5640 | return NULL; | |
5641 | } | |
5642 | ||
cee22a15 VK |
5643 | /* see the comment above the definition of WQ_POWER_EFFICIENT */ |
5644 | if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) | |
5645 | flags |= WQ_UNBOUND; | |
5646 | ||
ecf6881f | 5647 | /* allocate wq and format name */ |
91ccc6e7 TH |
5648 | if (flags & WQ_UNBOUND) |
5649 | wq_size = struct_size(wq, node_nr_active, nr_node_ids + 1); | |
5650 | else | |
5651 | wq_size = sizeof(*wq); | |
5652 | ||
5653 | wq = kzalloc(wq_size, GFP_KERNEL); | |
b196be89 | 5654 | if (!wq) |
d2c1d404 | 5655 | return NULL; |
b196be89 | 5656 | |
6029a918 | 5657 | if (flags & WQ_UNBOUND) { |
be69d00d | 5658 | wq->unbound_attrs = alloc_workqueue_attrs(); |
6029a918 TH |
5659 | if (!wq->unbound_attrs) |
5660 | goto err_free_wq; | |
5661 | } | |
5662 | ||
669de8bd | 5663 | va_start(args, max_active); |
91ccc6e7 | 5664 | name_len = vsnprintf(wq->name, sizeof(wq->name), fmt, args); |
b196be89 | 5665 | va_end(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 | |
669de8bd | 5695 | wq_init_lockdep(wq); |
cce1a165 | 5696 | INIT_LIST_HEAD(&wq->list); |
3af24433 | 5697 | |
91ccc6e7 TH |
5698 | if (flags & WQ_UNBOUND) { |
5699 | if (alloc_node_nr_active(wq->node_nr_active) < 0) | |
5700 | goto err_unreg_lockdep; | |
5701 | } | |
5702 | ||
30cdf249 | 5703 | if (alloc_and_link_pwqs(wq) < 0) |
91ccc6e7 | 5704 | goto err_free_node_nr_active; |
1537663f | 5705 | |
40c17f75 | 5706 | if (wq_online && init_rescuer(wq) < 0) |
983c7515 | 5707 | goto err_destroy; |
3af24433 | 5708 | |
226223ab TH |
5709 | if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) |
5710 | goto err_destroy; | |
5711 | ||
a0a1a5fd | 5712 | /* |
68e13a67 LJ |
5713 | * wq_pool_mutex protects global freeze state and workqueues list. |
5714 | * Grab it, adjust max_active and add the new @wq to workqueues | |
5715 | * list. | |
a0a1a5fd | 5716 | */ |
68e13a67 | 5717 | mutex_lock(&wq_pool_mutex); |
a0a1a5fd | 5718 | |
a357fc03 | 5719 | mutex_lock(&wq->mutex); |
a045a272 | 5720 | wq_adjust_max_active(wq); |
a357fc03 | 5721 | mutex_unlock(&wq->mutex); |
a0a1a5fd | 5722 | |
e2dca7ad | 5723 | list_add_tail_rcu(&wq->list, &workqueues); |
a0a1a5fd | 5724 | |
68e13a67 | 5725 | mutex_unlock(&wq_pool_mutex); |
1537663f | 5726 | |
3af24433 | 5727 | return wq; |
d2c1d404 | 5728 | |
91ccc6e7 TH |
5729 | err_free_node_nr_active: |
5730 | if (wq->flags & WQ_UNBOUND) | |
5731 | free_node_nr_active(wq->node_nr_active); | |
82efcab3 | 5732 | err_unreg_lockdep: |
009bb421 BVA |
5733 | wq_unregister_lockdep(wq); |
5734 | wq_free_lockdep(wq); | |
82efcab3 | 5735 | err_free_wq: |
6029a918 | 5736 | free_workqueue_attrs(wq->unbound_attrs); |
d2c1d404 TH |
5737 | kfree(wq); |
5738 | return NULL; | |
5739 | err_destroy: | |
5740 | destroy_workqueue(wq); | |
4690c4ab | 5741 | return NULL; |
3af24433 | 5742 | } |
669de8bd | 5743 | EXPORT_SYMBOL_GPL(alloc_workqueue); |
1da177e4 | 5744 | |
c29eb853 TH |
5745 | static bool pwq_busy(struct pool_workqueue *pwq) |
5746 | { | |
5747 | int i; | |
5748 | ||
5749 | for (i = 0; i < WORK_NR_COLORS; i++) | |
5750 | if (pwq->nr_in_flight[i]) | |
5751 | return true; | |
5752 | ||
9f66cff2 | 5753 | if ((pwq != rcu_access_pointer(pwq->wq->dfl_pwq)) && (pwq->refcnt > 1)) |
c29eb853 | 5754 | return true; |
afa87ce8 | 5755 | if (!pwq_is_empty(pwq)) |
c29eb853 TH |
5756 | return true; |
5757 | ||
5758 | return false; | |
5759 | } | |
5760 | ||
3af24433 ON |
5761 | /** |
5762 | * destroy_workqueue - safely terminate a workqueue | |
5763 | * @wq: target workqueue | |
5764 | * | |
5765 | * Safely destroy a workqueue. All work currently pending will be done first. | |
5766 | */ | |
5767 | void destroy_workqueue(struct workqueue_struct *wq) | |
5768 | { | |
49e3cf44 | 5769 | struct pool_workqueue *pwq; |
636b927e | 5770 | int cpu; |
3af24433 | 5771 | |
def98c84 TH |
5772 | /* |
5773 | * Remove it from sysfs first so that sanity check failure doesn't | |
5774 | * lead to sysfs name conflicts. | |
5775 | */ | |
5776 | workqueue_sysfs_unregister(wq); | |
5777 | ||
33e3f0a3 RC |
5778 | /* mark the workqueue destruction is in progress */ |
5779 | mutex_lock(&wq->mutex); | |
5780 | wq->flags |= __WQ_DESTROYING; | |
5781 | mutex_unlock(&wq->mutex); | |
5782 | ||
9c5a2ba7 TH |
5783 | /* drain it before proceeding with destruction */ |
5784 | drain_workqueue(wq); | |
c8efcc25 | 5785 | |
def98c84 TH |
5786 | /* kill rescuer, if sanity checks fail, leave it w/o rescuer */ |
5787 | if (wq->rescuer) { | |
5788 | struct worker *rescuer = wq->rescuer; | |
5789 | ||
5790 | /* this prevents new queueing */ | |
a9b8a985 | 5791 | raw_spin_lock_irq(&wq_mayday_lock); |
def98c84 | 5792 | wq->rescuer = NULL; |
a9b8a985 | 5793 | raw_spin_unlock_irq(&wq_mayday_lock); |
def98c84 TH |
5794 | |
5795 | /* rescuer will empty maydays list before exiting */ | |
5796 | kthread_stop(rescuer->task); | |
8efe1223 | 5797 | kfree(rescuer); |
def98c84 TH |
5798 | } |
5799 | ||
c29eb853 TH |
5800 | /* |
5801 | * Sanity checks - grab all the locks so that we wait for all | |
5802 | * in-flight operations which may do put_pwq(). | |
5803 | */ | |
5804 | mutex_lock(&wq_pool_mutex); | |
b09f4fd3 | 5805 | mutex_lock(&wq->mutex); |
49e3cf44 | 5806 | for_each_pwq(pwq, wq) { |
a9b8a985 | 5807 | raw_spin_lock_irq(&pwq->pool->lock); |
c29eb853 | 5808 | if (WARN_ON(pwq_busy(pwq))) { |
1d9a6159 KW |
5809 | pr_warn("%s: %s has the following busy pwq\n", |
5810 | __func__, wq->name); | |
c29eb853 | 5811 | show_pwq(pwq); |
a9b8a985 | 5812 | raw_spin_unlock_irq(&pwq->pool->lock); |
b09f4fd3 | 5813 | mutex_unlock(&wq->mutex); |
c29eb853 | 5814 | mutex_unlock(&wq_pool_mutex); |
55df0933 | 5815 | show_one_workqueue(wq); |
6183c009 | 5816 | return; |
76af4d93 | 5817 | } |
a9b8a985 | 5818 | raw_spin_unlock_irq(&pwq->pool->lock); |
6183c009 | 5819 | } |
b09f4fd3 | 5820 | mutex_unlock(&wq->mutex); |
6183c009 | 5821 | |
a0a1a5fd TH |
5822 | /* |
5823 | * wq list is used to freeze wq, remove from list after | |
5824 | * flushing is complete in case freeze races us. | |
5825 | */ | |
e2dca7ad | 5826 | list_del_rcu(&wq->list); |
68e13a67 | 5827 | mutex_unlock(&wq_pool_mutex); |
3af24433 | 5828 | |
636b927e TH |
5829 | /* |
5830 | * We're the sole accessor of @wq. Directly access cpu_pwq and dfl_pwq | |
5831 | * to put the base refs. @wq will be auto-destroyed from the last | |
5832 | * pwq_put. RCU read lock prevents @wq from going away from under us. | |
5833 | */ | |
5834 | rcu_read_lock(); | |
4c16bd32 | 5835 | |
636b927e | 5836 | for_each_possible_cpu(cpu) { |
9f66cff2 TH |
5837 | put_pwq_unlocked(unbound_pwq(wq, cpu)); |
5838 | RCU_INIT_POINTER(*unbound_pwq_slot(wq, cpu), NULL); | |
29c91e99 | 5839 | } |
636b927e | 5840 | |
9f66cff2 TH |
5841 | put_pwq_unlocked(unbound_pwq(wq, -1)); |
5842 | RCU_INIT_POINTER(*unbound_pwq_slot(wq, -1), NULL); | |
636b927e TH |
5843 | |
5844 | rcu_read_unlock(); | |
3af24433 ON |
5845 | } |
5846 | EXPORT_SYMBOL_GPL(destroy_workqueue); | |
5847 | ||
dcd989cb TH |
5848 | /** |
5849 | * workqueue_set_max_active - adjust max_active of a workqueue | |
5850 | * @wq: target workqueue | |
5851 | * @max_active: new max_active value. | |
5852 | * | |
5797b1c1 TH |
5853 | * Set max_active of @wq to @max_active. See the alloc_workqueue() function |
5854 | * comment. | |
dcd989cb TH |
5855 | * |
5856 | * CONTEXT: | |
5857 | * Don't call from IRQ context. | |
5858 | */ | |
5859 | void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) | |
5860 | { | |
4cb1ef64 TH |
5861 | /* max_active doesn't mean anything for BH workqueues */ |
5862 | if (WARN_ON(wq->flags & WQ_BH)) | |
5863 | return; | |
8719dcea | 5864 | /* disallow meddling with max_active for ordered workqueues */ |
3bc1e711 | 5865 | if (WARN_ON(wq->flags & __WQ_ORDERED)) |
8719dcea TH |
5866 | return; |
5867 | ||
f3421797 | 5868 | max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); |
dcd989cb | 5869 | |
a357fc03 | 5870 | mutex_lock(&wq->mutex); |
dcd989cb TH |
5871 | |
5872 | wq->saved_max_active = max_active; | |
5797b1c1 TH |
5873 | if (wq->flags & WQ_UNBOUND) |
5874 | wq->saved_min_active = min(wq->saved_min_active, max_active); | |
5875 | ||
a045a272 | 5876 | wq_adjust_max_active(wq); |
93981800 | 5877 | |
a357fc03 | 5878 | mutex_unlock(&wq->mutex); |
15316ba8 | 5879 | } |
dcd989cb | 5880 | EXPORT_SYMBOL_GPL(workqueue_set_max_active); |
15316ba8 | 5881 | |
8f172181 TH |
5882 | /** |
5883 | * workqueue_set_min_active - adjust min_active of an unbound workqueue | |
5884 | * @wq: target unbound workqueue | |
5885 | * @min_active: new min_active value | |
5886 | * | |
5887 | * Set min_active of an unbound workqueue. Unlike other types of workqueues, an | |
5888 | * unbound workqueue is not guaranteed to be able to process max_active | |
5889 | * interdependent work items. Instead, an unbound workqueue is guaranteed to be | |
5890 | * able to process min_active number of interdependent work items which is | |
5891 | * %WQ_DFL_MIN_ACTIVE by default. | |
5892 | * | |
5893 | * Use this function to adjust the min_active value between 0 and the current | |
5894 | * max_active. | |
5895 | */ | |
5896 | void workqueue_set_min_active(struct workqueue_struct *wq, int min_active) | |
5897 | { | |
5898 | /* min_active is only meaningful for non-ordered unbound workqueues */ | |
5899 | if (WARN_ON((wq->flags & (WQ_BH | WQ_UNBOUND | __WQ_ORDERED)) != | |
5900 | WQ_UNBOUND)) | |
5901 | return; | |
5902 | ||
5903 | mutex_lock(&wq->mutex); | |
5904 | wq->saved_min_active = clamp(min_active, 0, wq->saved_max_active); | |
5905 | wq_adjust_max_active(wq); | |
5906 | mutex_unlock(&wq->mutex); | |
5907 | } | |
5908 | ||
27d4ee03 LW |
5909 | /** |
5910 | * current_work - retrieve %current task's work struct | |
5911 | * | |
5912 | * Determine if %current task is a workqueue worker and what it's working on. | |
5913 | * Useful to find out the context that the %current task is running in. | |
5914 | * | |
5915 | * Return: work struct if %current task is a workqueue worker, %NULL otherwise. | |
5916 | */ | |
5917 | struct work_struct *current_work(void) | |
5918 | { | |
5919 | struct worker *worker = current_wq_worker(); | |
5920 | ||
5921 | return worker ? worker->current_work : NULL; | |
5922 | } | |
5923 | EXPORT_SYMBOL(current_work); | |
5924 | ||
e6267616 TH |
5925 | /** |
5926 | * current_is_workqueue_rescuer - is %current workqueue rescuer? | |
5927 | * | |
5928 | * Determine whether %current is a workqueue rescuer. Can be used from | |
5929 | * work functions to determine whether it's being run off the rescuer task. | |
d185af30 YB |
5930 | * |
5931 | * Return: %true if %current is a workqueue rescuer. %false otherwise. | |
e6267616 TH |
5932 | */ |
5933 | bool current_is_workqueue_rescuer(void) | |
5934 | { | |
5935 | struct worker *worker = current_wq_worker(); | |
5936 | ||
6a092dfd | 5937 | return worker && worker->rescue_wq; |
e6267616 TH |
5938 | } |
5939 | ||
eef6a7d5 | 5940 | /** |
dcd989cb TH |
5941 | * workqueue_congested - test whether a workqueue is congested |
5942 | * @cpu: CPU in question | |
5943 | * @wq: target workqueue | |
eef6a7d5 | 5944 | * |
dcd989cb TH |
5945 | * Test whether @wq's cpu workqueue for @cpu is congested. There is |
5946 | * no synchronization around this function and the test result is | |
5947 | * unreliable and only useful as advisory hints or for debugging. | |
eef6a7d5 | 5948 | * |
d3251859 | 5949 | * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. |
636b927e TH |
5950 | * |
5951 | * With the exception of ordered workqueues, all workqueues have per-cpu | |
5952 | * pool_workqueues, each with its own congested state. A workqueue being | |
5953 | * congested on one CPU doesn't mean that the workqueue is contested on any | |
5954 | * other CPUs. | |
d3251859 | 5955 | * |
d185af30 | 5956 | * Return: |
dcd989cb | 5957 | * %true if congested, %false otherwise. |
eef6a7d5 | 5958 | */ |
d84ff051 | 5959 | bool workqueue_congested(int cpu, struct workqueue_struct *wq) |
1da177e4 | 5960 | { |
7fb98ea7 | 5961 | struct pool_workqueue *pwq; |
76af4d93 TH |
5962 | bool ret; |
5963 | ||
24acfb71 TG |
5964 | rcu_read_lock(); |
5965 | preempt_disable(); | |
7fb98ea7 | 5966 | |
d3251859 TH |
5967 | if (cpu == WORK_CPU_UNBOUND) |
5968 | cpu = smp_processor_id(); | |
5969 | ||
636b927e | 5970 | pwq = *per_cpu_ptr(wq->cpu_pwq, cpu); |
f97a4a1a | 5971 | ret = !list_empty(&pwq->inactive_works); |
636b927e | 5972 | |
24acfb71 TG |
5973 | preempt_enable(); |
5974 | rcu_read_unlock(); | |
76af4d93 TH |
5975 | |
5976 | return ret; | |
1da177e4 | 5977 | } |
dcd989cb | 5978 | EXPORT_SYMBOL_GPL(workqueue_congested); |
1da177e4 | 5979 | |
dcd989cb TH |
5980 | /** |
5981 | * work_busy - test whether a work is currently pending or running | |
5982 | * @work: the work to be tested | |
5983 | * | |
5984 | * Test whether @work is currently pending or running. There is no | |
5985 | * synchronization around this function and the test result is | |
5986 | * unreliable and only useful as advisory hints or for debugging. | |
dcd989cb | 5987 | * |
d185af30 | 5988 | * Return: |
dcd989cb TH |
5989 | * OR'd bitmask of WORK_BUSY_* bits. |
5990 | */ | |
5991 | unsigned int work_busy(struct work_struct *work) | |
1da177e4 | 5992 | { |
fa1b54e6 | 5993 | struct worker_pool *pool; |
c26e2f2e | 5994 | unsigned long irq_flags; |
dcd989cb | 5995 | unsigned int ret = 0; |
1da177e4 | 5996 | |
dcd989cb TH |
5997 | if (work_pending(work)) |
5998 | ret |= WORK_BUSY_PENDING; | |
1da177e4 | 5999 | |
24acfb71 | 6000 | rcu_read_lock(); |
fa1b54e6 | 6001 | pool = get_work_pool(work); |
038366c5 | 6002 | if (pool) { |
c26e2f2e | 6003 | raw_spin_lock_irqsave(&pool->lock, irq_flags); |
038366c5 LJ |
6004 | if (find_worker_executing_work(pool, work)) |
6005 | ret |= WORK_BUSY_RUNNING; | |
c26e2f2e | 6006 | raw_spin_unlock_irqrestore(&pool->lock, irq_flags); |
038366c5 | 6007 | } |
24acfb71 | 6008 | rcu_read_unlock(); |
1da177e4 | 6009 | |
dcd989cb | 6010 | return ret; |
1da177e4 | 6011 | } |
dcd989cb | 6012 | EXPORT_SYMBOL_GPL(work_busy); |
1da177e4 | 6013 | |
3d1cb205 TH |
6014 | /** |
6015 | * set_worker_desc - set description for the current work item | |
6016 | * @fmt: printf-style format string | |
6017 | * @...: arguments for the format string | |
6018 | * | |
6019 | * This function can be called by a running work function to describe what | |
6020 | * the work item is about. If the worker task gets dumped, this | |
6021 | * information will be printed out together to help debugging. The | |
6022 | * description can be at most WORKER_DESC_LEN including the trailing '\0'. | |
6023 | */ | |
6024 | void set_worker_desc(const char *fmt, ...) | |
6025 | { | |
6026 | struct worker *worker = current_wq_worker(); | |
6027 | va_list args; | |
6028 | ||
6029 | if (worker) { | |
6030 | va_start(args, fmt); | |
6031 | vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); | |
6032 | va_end(args); | |
3d1cb205 TH |
6033 | } |
6034 | } | |
5c750d58 | 6035 | EXPORT_SYMBOL_GPL(set_worker_desc); |
3d1cb205 TH |
6036 | |
6037 | /** | |
6038 | * print_worker_info - print out worker information and description | |
6039 | * @log_lvl: the log level to use when printing | |
6040 | * @task: target task | |
6041 | * | |
6042 | * If @task is a worker and currently executing a work item, print out the | |
6043 | * name of the workqueue being serviced and worker description set with | |
6044 | * set_worker_desc() by the currently executing work item. | |
6045 | * | |
6046 | * This function can be safely called on any task as long as the | |
6047 | * task_struct itself is accessible. While safe, this function isn't | |
6048 | * synchronized and may print out mixups or garbages of limited length. | |
6049 | */ | |
6050 | void print_worker_info(const char *log_lvl, struct task_struct *task) | |
6051 | { | |
6052 | work_func_t *fn = NULL; | |
6053 | char name[WQ_NAME_LEN] = { }; | |
6054 | char desc[WORKER_DESC_LEN] = { }; | |
6055 | struct pool_workqueue *pwq = NULL; | |
6056 | struct workqueue_struct *wq = NULL; | |
3d1cb205 TH |
6057 | struct worker *worker; |
6058 | ||
6059 | if (!(task->flags & PF_WQ_WORKER)) | |
6060 | return; | |
6061 | ||
6062 | /* | |
6063 | * This function is called without any synchronization and @task | |
6064 | * could be in any state. Be careful with dereferences. | |
6065 | */ | |
e700591a | 6066 | worker = kthread_probe_data(task); |
3d1cb205 TH |
6067 | |
6068 | /* | |
8bf89593 TH |
6069 | * Carefully copy the associated workqueue's workfn, name and desc. |
6070 | * Keep the original last '\0' in case the original is garbage. | |
3d1cb205 | 6071 | */ |
fe557319 CH |
6072 | copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn)); |
6073 | copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq)); | |
6074 | copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq)); | |
6075 | copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1); | |
6076 | copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1); | |
3d1cb205 TH |
6077 | |
6078 | if (fn || name[0] || desc[0]) { | |
d75f773c | 6079 | printk("%sWorkqueue: %s %ps", log_lvl, name, fn); |
8bf89593 | 6080 | if (strcmp(name, desc)) |
3d1cb205 TH |
6081 | pr_cont(" (%s)", desc); |
6082 | pr_cont("\n"); | |
6083 | } | |
6084 | } | |
6085 | ||
3494fc30 TH |
6086 | static void pr_cont_pool_info(struct worker_pool *pool) |
6087 | { | |
6088 | pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask); | |
6089 | if (pool->node != NUMA_NO_NODE) | |
6090 | pr_cont(" node=%d", pool->node); | |
4cb1ef64 TH |
6091 | pr_cont(" flags=0x%x", pool->flags); |
6092 | if (pool->flags & POOL_BH) | |
6093 | pr_cont(" bh%s", | |
6094 | pool->attrs->nice == HIGHPRI_NICE_LEVEL ? "-hi" : ""); | |
6095 | else | |
6096 | pr_cont(" nice=%d", pool->attrs->nice); | |
6097 | } | |
6098 | ||
6099 | static void pr_cont_worker_id(struct worker *worker) | |
6100 | { | |
6101 | struct worker_pool *pool = worker->pool; | |
6102 | ||
6103 | if (pool->flags & WQ_BH) | |
6104 | pr_cont("bh%s", | |
6105 | pool->attrs->nice == HIGHPRI_NICE_LEVEL ? "-hi" : ""); | |
6106 | else | |
6107 | pr_cont("%d%s", task_pid_nr(worker->task), | |
6108 | worker->rescue_wq ? "(RESCUER)" : ""); | |
3494fc30 TH |
6109 | } |
6110 | ||
c76feb0d PM |
6111 | struct pr_cont_work_struct { |
6112 | bool comma; | |
6113 | work_func_t func; | |
6114 | long ctr; | |
6115 | }; | |
6116 | ||
6117 | static void pr_cont_work_flush(bool comma, work_func_t func, struct pr_cont_work_struct *pcwsp) | |
6118 | { | |
6119 | if (!pcwsp->ctr) | |
6120 | goto out_record; | |
6121 | if (func == pcwsp->func) { | |
6122 | pcwsp->ctr++; | |
6123 | return; | |
6124 | } | |
6125 | if (pcwsp->ctr == 1) | |
6126 | pr_cont("%s %ps", pcwsp->comma ? "," : "", pcwsp->func); | |
6127 | else | |
6128 | pr_cont("%s %ld*%ps", pcwsp->comma ? "," : "", pcwsp->ctr, pcwsp->func); | |
6129 | pcwsp->ctr = 0; | |
6130 | out_record: | |
6131 | if ((long)func == -1L) | |
6132 | return; | |
6133 | pcwsp->comma = comma; | |
6134 | pcwsp->func = func; | |
6135 | pcwsp->ctr = 1; | |
6136 | } | |
6137 | ||
6138 | static void pr_cont_work(bool comma, struct work_struct *work, struct pr_cont_work_struct *pcwsp) | |
3494fc30 TH |
6139 | { |
6140 | if (work->func == wq_barrier_func) { | |
6141 | struct wq_barrier *barr; | |
6142 | ||
6143 | barr = container_of(work, struct wq_barrier, work); | |
6144 | ||
c76feb0d | 6145 | pr_cont_work_flush(comma, (work_func_t)-1, pcwsp); |
3494fc30 TH |
6146 | pr_cont("%s BAR(%d)", comma ? "," : "", |
6147 | task_pid_nr(barr->task)); | |
6148 | } else { | |
c76feb0d PM |
6149 | if (!comma) |
6150 | pr_cont_work_flush(comma, (work_func_t)-1, pcwsp); | |
6151 | pr_cont_work_flush(comma, work->func, pcwsp); | |
3494fc30 TH |
6152 | } |
6153 | } | |
6154 | ||
6155 | static void show_pwq(struct pool_workqueue *pwq) | |
6156 | { | |
c76feb0d | 6157 | struct pr_cont_work_struct pcws = { .ctr = 0, }; |
3494fc30 TH |
6158 | struct worker_pool *pool = pwq->pool; |
6159 | struct work_struct *work; | |
6160 | struct worker *worker; | |
6161 | bool has_in_flight = false, has_pending = false; | |
6162 | int bkt; | |
6163 | ||
6164 | pr_info(" pwq %d:", pool->id); | |
6165 | pr_cont_pool_info(pool); | |
6166 | ||
a045a272 TH |
6167 | pr_cont(" active=%d refcnt=%d%s\n", |
6168 | pwq->nr_active, pwq->refcnt, | |
3494fc30 TH |
6169 | !list_empty(&pwq->mayday_node) ? " MAYDAY" : ""); |
6170 | ||
6171 | hash_for_each(pool->busy_hash, bkt, worker, hentry) { | |
6172 | if (worker->current_pwq == pwq) { | |
6173 | has_in_flight = true; | |
6174 | break; | |
6175 | } | |
6176 | } | |
6177 | if (has_in_flight) { | |
6178 | bool comma = false; | |
6179 | ||
6180 | pr_info(" in-flight:"); | |
6181 | hash_for_each(pool->busy_hash, bkt, worker, hentry) { | |
6182 | if (worker->current_pwq != pwq) | |
6183 | continue; | |
6184 | ||
4cb1ef64 TH |
6185 | pr_cont(" %s", comma ? "," : ""); |
6186 | pr_cont_worker_id(worker); | |
6187 | pr_cont(":%ps", worker->current_func); | |
3494fc30 | 6188 | list_for_each_entry(work, &worker->scheduled, entry) |
c76feb0d PM |
6189 | pr_cont_work(false, work, &pcws); |
6190 | pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); | |
3494fc30 TH |
6191 | comma = true; |
6192 | } | |
6193 | pr_cont("\n"); | |
6194 | } | |
6195 | ||
6196 | list_for_each_entry(work, &pool->worklist, entry) { | |
6197 | if (get_work_pwq(work) == pwq) { | |
6198 | has_pending = true; | |
6199 | break; | |
6200 | } | |
6201 | } | |
6202 | if (has_pending) { | |
6203 | bool comma = false; | |
6204 | ||
6205 | pr_info(" pending:"); | |
6206 | list_for_each_entry(work, &pool->worklist, entry) { | |
6207 | if (get_work_pwq(work) != pwq) | |
6208 | continue; | |
6209 | ||
c76feb0d | 6210 | pr_cont_work(comma, work, &pcws); |
3494fc30 TH |
6211 | comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); |
6212 | } | |
c76feb0d | 6213 | pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); |
3494fc30 TH |
6214 | pr_cont("\n"); |
6215 | } | |
6216 | ||
f97a4a1a | 6217 | if (!list_empty(&pwq->inactive_works)) { |
3494fc30 TH |
6218 | bool comma = false; |
6219 | ||
f97a4a1a LJ |
6220 | pr_info(" inactive:"); |
6221 | list_for_each_entry(work, &pwq->inactive_works, entry) { | |
c76feb0d | 6222 | pr_cont_work(comma, work, &pcws); |
3494fc30 TH |
6223 | comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); |
6224 | } | |
c76feb0d | 6225 | pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); |
3494fc30 TH |
6226 | pr_cont("\n"); |
6227 | } | |
6228 | } | |
6229 | ||
6230 | /** | |
55df0933 IK |
6231 | * show_one_workqueue - dump state of specified workqueue |
6232 | * @wq: workqueue whose state will be printed | |
3494fc30 | 6233 | */ |
55df0933 | 6234 | void show_one_workqueue(struct workqueue_struct *wq) |
3494fc30 | 6235 | { |
55df0933 IK |
6236 | struct pool_workqueue *pwq; |
6237 | bool idle = true; | |
c26e2f2e | 6238 | unsigned long irq_flags; |
3494fc30 | 6239 | |
55df0933 | 6240 | for_each_pwq(pwq, wq) { |
afa87ce8 | 6241 | if (!pwq_is_empty(pwq)) { |
55df0933 IK |
6242 | idle = false; |
6243 | break; | |
3494fc30 | 6244 | } |
55df0933 IK |
6245 | } |
6246 | if (idle) /* Nothing to print for idle workqueue */ | |
6247 | return; | |
3494fc30 | 6248 | |
55df0933 | 6249 | pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags); |
3494fc30 | 6250 | |
55df0933 | 6251 | for_each_pwq(pwq, wq) { |
c26e2f2e | 6252 | raw_spin_lock_irqsave(&pwq->pool->lock, irq_flags); |
afa87ce8 | 6253 | if (!pwq_is_empty(pwq)) { |
62635ea8 | 6254 | /* |
55df0933 IK |
6255 | * Defer printing to avoid deadlocks in console |
6256 | * drivers that queue work while holding locks | |
6257 | * also taken in their write paths. | |
62635ea8 | 6258 | */ |
55df0933 IK |
6259 | printk_deferred_enter(); |
6260 | show_pwq(pwq); | |
6261 | printk_deferred_exit(); | |
3494fc30 | 6262 | } |
c26e2f2e | 6263 | raw_spin_unlock_irqrestore(&pwq->pool->lock, irq_flags); |
62635ea8 SS |
6264 | /* |
6265 | * We could be printing a lot from atomic context, e.g. | |
55df0933 | 6266 | * sysrq-t -> show_all_workqueues(). Avoid triggering |
62635ea8 SS |
6267 | * hard lockup. |
6268 | */ | |
6269 | touch_nmi_watchdog(); | |
3494fc30 TH |
6270 | } |
6271 | ||
55df0933 IK |
6272 | } |
6273 | ||
6274 | /** | |
6275 | * show_one_worker_pool - dump state of specified worker pool | |
6276 | * @pool: worker pool whose state will be printed | |
6277 | */ | |
6278 | static void show_one_worker_pool(struct worker_pool *pool) | |
6279 | { | |
6280 | struct worker *worker; | |
6281 | bool first = true; | |
c26e2f2e | 6282 | unsigned long irq_flags; |
335a42eb | 6283 | unsigned long hung = 0; |
55df0933 | 6284 | |
c26e2f2e | 6285 | raw_spin_lock_irqsave(&pool->lock, irq_flags); |
55df0933 IK |
6286 | if (pool->nr_workers == pool->nr_idle) |
6287 | goto next_pool; | |
335a42eb PM |
6288 | |
6289 | /* How long the first pending work is waiting for a worker. */ | |
6290 | if (!list_empty(&pool->worklist)) | |
6291 | hung = jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000; | |
6292 | ||
55df0933 IK |
6293 | /* |
6294 | * Defer printing to avoid deadlocks in console drivers that | |
6295 | * queue work while holding locks also taken in their write | |
6296 | * paths. | |
6297 | */ | |
6298 | printk_deferred_enter(); | |
6299 | pr_info("pool %d:", pool->id); | |
6300 | pr_cont_pool_info(pool); | |
335a42eb | 6301 | pr_cont(" hung=%lus workers=%d", hung, pool->nr_workers); |
55df0933 IK |
6302 | if (pool->manager) |
6303 | pr_cont(" manager: %d", | |
6304 | task_pid_nr(pool->manager->task)); | |
6305 | list_for_each_entry(worker, &pool->idle_list, entry) { | |
4cb1ef64 TH |
6306 | pr_cont(" %s", first ? "idle: " : ""); |
6307 | pr_cont_worker_id(worker); | |
55df0933 IK |
6308 | first = false; |
6309 | } | |
6310 | pr_cont("\n"); | |
6311 | printk_deferred_exit(); | |
6312 | next_pool: | |
c26e2f2e | 6313 | raw_spin_unlock_irqrestore(&pool->lock, irq_flags); |
55df0933 IK |
6314 | /* |
6315 | * We could be printing a lot from atomic context, e.g. | |
6316 | * sysrq-t -> show_all_workqueues(). Avoid triggering | |
6317 | * hard lockup. | |
6318 | */ | |
6319 | touch_nmi_watchdog(); | |
6320 | ||
6321 | } | |
6322 | ||
6323 | /** | |
6324 | * show_all_workqueues - dump workqueue state | |
6325 | * | |
704bc669 | 6326 | * Called from a sysrq handler and prints out all busy workqueues and pools. |
55df0933 IK |
6327 | */ |
6328 | void show_all_workqueues(void) | |
6329 | { | |
6330 | struct workqueue_struct *wq; | |
6331 | struct worker_pool *pool; | |
6332 | int pi; | |
6333 | ||
6334 | rcu_read_lock(); | |
6335 | ||
6336 | pr_info("Showing busy workqueues and worker pools:\n"); | |
6337 | ||
6338 | list_for_each_entry_rcu(wq, &workqueues, list) | |
6339 | show_one_workqueue(wq); | |
6340 | ||
6341 | for_each_pool(pool, pi) | |
6342 | show_one_worker_pool(pool); | |
6343 | ||
24acfb71 | 6344 | rcu_read_unlock(); |
3494fc30 TH |
6345 | } |
6346 | ||
704bc669 JL |
6347 | /** |
6348 | * show_freezable_workqueues - dump freezable workqueue state | |
6349 | * | |
6350 | * Called from try_to_freeze_tasks() and prints out all freezable workqueues | |
6351 | * still busy. | |
6352 | */ | |
6353 | void show_freezable_workqueues(void) | |
6354 | { | |
6355 | struct workqueue_struct *wq; | |
6356 | ||
6357 | rcu_read_lock(); | |
6358 | ||
6359 | pr_info("Showing freezable workqueues that are still busy:\n"); | |
6360 | ||
6361 | list_for_each_entry_rcu(wq, &workqueues, list) { | |
6362 | if (!(wq->flags & WQ_FREEZABLE)) | |
6363 | continue; | |
6364 | show_one_workqueue(wq); | |
6365 | } | |
6366 | ||
6367 | rcu_read_unlock(); | |
6368 | } | |
6369 | ||
6b59808b TH |
6370 | /* used to show worker information through /proc/PID/{comm,stat,status} */ |
6371 | void wq_worker_comm(char *buf, size_t size, struct task_struct *task) | |
6372 | { | |
6b59808b TH |
6373 | int off; |
6374 | ||
6375 | /* always show the actual comm */ | |
6376 | off = strscpy(buf, task->comm, size); | |
6377 | if (off < 0) | |
6378 | return; | |
6379 | ||
197f6acc | 6380 | /* stabilize PF_WQ_WORKER and worker pool association */ |
6b59808b TH |
6381 | mutex_lock(&wq_pool_attach_mutex); |
6382 | ||
197f6acc TH |
6383 | if (task->flags & PF_WQ_WORKER) { |
6384 | struct worker *worker = kthread_data(task); | |
6385 | struct worker_pool *pool = worker->pool; | |
6b59808b | 6386 | |
197f6acc | 6387 | if (pool) { |
a9b8a985 | 6388 | raw_spin_lock_irq(&pool->lock); |
197f6acc TH |
6389 | /* |
6390 | * ->desc tracks information (wq name or | |
6391 | * set_worker_desc()) for the latest execution. If | |
6392 | * current, prepend '+', otherwise '-'. | |
6393 | */ | |
6394 | if (worker->desc[0] != '\0') { | |
6395 | if (worker->current_work) | |
6396 | scnprintf(buf + off, size - off, "+%s", | |
6397 | worker->desc); | |
6398 | else | |
6399 | scnprintf(buf + off, size - off, "-%s", | |
6400 | worker->desc); | |
6401 | } | |
a9b8a985 | 6402 | raw_spin_unlock_irq(&pool->lock); |
6b59808b | 6403 | } |
6b59808b TH |
6404 | } |
6405 | ||
6406 | mutex_unlock(&wq_pool_attach_mutex); | |
6407 | } | |
6408 | ||
66448bc2 MM |
6409 | #ifdef CONFIG_SMP |
6410 | ||
db7bccf4 TH |
6411 | /* |
6412 | * CPU hotplug. | |
6413 | * | |
e22bee78 | 6414 | * There are two challenges in supporting CPU hotplug. Firstly, there |
112202d9 | 6415 | * are a lot of assumptions on strong associations among work, pwq and |
706026c2 | 6416 | * pool which make migrating pending and scheduled works very |
e22bee78 | 6417 | * difficult to implement without impacting hot paths. Secondly, |
94cf58bb | 6418 | * worker pools serve mix of short, long and very long running works making |
e22bee78 TH |
6419 | * blocked draining impractical. |
6420 | * | |
24647570 | 6421 | * This is solved by allowing the pools to be disassociated from the CPU |
628c78e7 TH |
6422 | * running as an unbound one and allowing it to be reattached later if the |
6423 | * cpu comes back online. | |
db7bccf4 | 6424 | */ |
1da177e4 | 6425 | |
e8b3f8db | 6426 | static void unbind_workers(int cpu) |
3af24433 | 6427 | { |
4ce62e9e | 6428 | struct worker_pool *pool; |
db7bccf4 | 6429 | struct worker *worker; |
3af24433 | 6430 | |
f02ae73a | 6431 | for_each_cpu_worker_pool(pool, cpu) { |
1258fae7 | 6432 | mutex_lock(&wq_pool_attach_mutex); |
a9b8a985 | 6433 | raw_spin_lock_irq(&pool->lock); |
3af24433 | 6434 | |
94cf58bb | 6435 | /* |
92f9c5c4 | 6436 | * We've blocked all attach/detach operations. Make all workers |
94cf58bb | 6437 | * unbound and set DISASSOCIATED. Before this, all workers |
11b45b0b | 6438 | * must be on the cpu. After this, they may become diasporas. |
b4ac9384 LJ |
6439 | * And the preemption disabled section in their sched callbacks |
6440 | * are guaranteed to see WORKER_UNBOUND since the code here | |
6441 | * is on the same cpu. | |
94cf58bb | 6442 | */ |
da028469 | 6443 | for_each_pool_worker(worker, pool) |
c9e7cf27 | 6444 | worker->flags |= WORKER_UNBOUND; |
06ba38a9 | 6445 | |
24647570 | 6446 | pool->flags |= POOL_DISASSOCIATED; |
f2d5a0ee | 6447 | |
eb283428 | 6448 | /* |
989442d7 LJ |
6449 | * The handling of nr_running in sched callbacks are disabled |
6450 | * now. Zap nr_running. After this, nr_running stays zero and | |
6451 | * need_more_worker() and keep_working() are always true as | |
6452 | * long as the worklist is not empty. This pool now behaves as | |
6453 | * an unbound (in terms of concurrency management) pool which | |
eb283428 LJ |
6454 | * are served by workers tied to the pool. |
6455 | */ | |
bc35f7ef | 6456 | pool->nr_running = 0; |
eb283428 LJ |
6457 | |
6458 | /* | |
6459 | * With concurrency management just turned off, a busy | |
6460 | * worker blocking could lead to lengthy stalls. Kick off | |
6461 | * unbound chain execution of currently pending work items. | |
6462 | */ | |
0219a352 | 6463 | kick_pool(pool); |
989442d7 | 6464 | |
a9b8a985 | 6465 | raw_spin_unlock_irq(&pool->lock); |
989442d7 | 6466 | |
793777bc VS |
6467 | for_each_pool_worker(worker, pool) |
6468 | unbind_worker(worker); | |
989442d7 LJ |
6469 | |
6470 | mutex_unlock(&wq_pool_attach_mutex); | |
eb283428 | 6471 | } |
3af24433 | 6472 | } |
3af24433 | 6473 | |
bd7c089e TH |
6474 | /** |
6475 | * rebind_workers - rebind all workers of a pool to the associated CPU | |
6476 | * @pool: pool of interest | |
6477 | * | |
a9ab775b | 6478 | * @pool->cpu is coming online. Rebind all workers to the CPU. |
bd7c089e TH |
6479 | */ |
6480 | static void rebind_workers(struct worker_pool *pool) | |
6481 | { | |
a9ab775b | 6482 | struct worker *worker; |
bd7c089e | 6483 | |
1258fae7 | 6484 | lockdep_assert_held(&wq_pool_attach_mutex); |
bd7c089e | 6485 | |
a9ab775b TH |
6486 | /* |
6487 | * Restore CPU affinity of all workers. As all idle workers should | |
6488 | * be on the run-queue of the associated CPU before any local | |
402dd89d | 6489 | * wake-ups for concurrency management happen, restore CPU affinity |
a9ab775b TH |
6490 | * of all workers first and then clear UNBOUND. As we're called |
6491 | * from CPU_ONLINE, the following shouldn't fail. | |
6492 | */ | |
c63a2e52 VS |
6493 | for_each_pool_worker(worker, pool) { |
6494 | kthread_set_per_cpu(worker->task, pool->cpu); | |
6495 | WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, | |
9546b29e | 6496 | pool_allowed_cpus(pool)) < 0); |
c63a2e52 | 6497 | } |
bd7c089e | 6498 | |
a9b8a985 | 6499 | raw_spin_lock_irq(&pool->lock); |
f7c17d26 | 6500 | |
3de5e884 | 6501 | pool->flags &= ~POOL_DISASSOCIATED; |
bd7c089e | 6502 | |
da028469 | 6503 | for_each_pool_worker(worker, pool) { |
a9ab775b | 6504 | unsigned int worker_flags = worker->flags; |
bd7c089e | 6505 | |
a9ab775b TH |
6506 | /* |
6507 | * We want to clear UNBOUND but can't directly call | |
6508 | * worker_clr_flags() or adjust nr_running. Atomically | |
6509 | * replace UNBOUND with another NOT_RUNNING flag REBOUND. | |
6510 | * @worker will clear REBOUND using worker_clr_flags() when | |
6511 | * it initiates the next execution cycle thus restoring | |
6512 | * concurrency management. Note that when or whether | |
6513 | * @worker clears REBOUND doesn't affect correctness. | |
6514 | * | |
c95491ed | 6515 | * WRITE_ONCE() is necessary because @worker->flags may be |
a9ab775b | 6516 | * tested without holding any lock in |
6d25be57 | 6517 | * wq_worker_running(). Without it, NOT_RUNNING test may |
a9ab775b TH |
6518 | * fail incorrectly leading to premature concurrency |
6519 | * management operations. | |
6520 | */ | |
6521 | WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); | |
6522 | worker_flags |= WORKER_REBOUND; | |
6523 | worker_flags &= ~WORKER_UNBOUND; | |
c95491ed | 6524 | WRITE_ONCE(worker->flags, worker_flags); |
bd7c089e | 6525 | } |
a9ab775b | 6526 | |
a9b8a985 | 6527 | raw_spin_unlock_irq(&pool->lock); |
bd7c089e TH |
6528 | } |
6529 | ||
7dbc725e TH |
6530 | /** |
6531 | * restore_unbound_workers_cpumask - restore cpumask of unbound workers | |
6532 | * @pool: unbound pool of interest | |
6533 | * @cpu: the CPU which is coming up | |
6534 | * | |
6535 | * An unbound pool may end up with a cpumask which doesn't have any online | |
6536 | * CPUs. When a worker of such pool get scheduled, the scheduler resets | |
6537 | * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any | |
6538 | * online CPU before, cpus_allowed of all its workers should be restored. | |
6539 | */ | |
6540 | static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) | |
6541 | { | |
6542 | static cpumask_t cpumask; | |
6543 | struct worker *worker; | |
7dbc725e | 6544 | |
1258fae7 | 6545 | lockdep_assert_held(&wq_pool_attach_mutex); |
7dbc725e TH |
6546 | |
6547 | /* is @cpu allowed for @pool? */ | |
6548 | if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) | |
6549 | return; | |
6550 | ||
7dbc725e | 6551 | cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); |
7dbc725e TH |
6552 | |
6553 | /* as we're called from CPU_ONLINE, the following shouldn't fail */ | |
da028469 | 6554 | for_each_pool_worker(worker, pool) |
d945b5e9 | 6555 | WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0); |
7dbc725e TH |
6556 | } |
6557 | ||
7ee681b2 TG |
6558 | int workqueue_prepare_cpu(unsigned int cpu) |
6559 | { | |
6560 | struct worker_pool *pool; | |
6561 | ||
6562 | for_each_cpu_worker_pool(pool, cpu) { | |
6563 | if (pool->nr_workers) | |
6564 | continue; | |
6565 | if (!create_worker(pool)) | |
6566 | return -ENOMEM; | |
6567 | } | |
6568 | return 0; | |
6569 | } | |
6570 | ||
6571 | int workqueue_online_cpu(unsigned int cpu) | |
3af24433 | 6572 | { |
4ce62e9e | 6573 | struct worker_pool *pool; |
4c16bd32 | 6574 | struct workqueue_struct *wq; |
7dbc725e | 6575 | int pi; |
3ce63377 | 6576 | |
7ee681b2 | 6577 | mutex_lock(&wq_pool_mutex); |
7dbc725e | 6578 | |
7ee681b2 | 6579 | for_each_pool(pool, pi) { |
4cb1ef64 TH |
6580 | /* BH pools aren't affected by hotplug */ |
6581 | if (pool->flags & POOL_BH) | |
6582 | continue; | |
94cf58bb | 6583 | |
4cb1ef64 | 6584 | mutex_lock(&wq_pool_attach_mutex); |
7ee681b2 TG |
6585 | if (pool->cpu == cpu) |
6586 | rebind_workers(pool); | |
6587 | else if (pool->cpu < 0) | |
6588 | restore_unbound_workers_cpumask(pool, cpu); | |
1258fae7 | 6589 | mutex_unlock(&wq_pool_attach_mutex); |
7ee681b2 | 6590 | } |
6ba94429 | 6591 | |
fef59c9c | 6592 | /* update pod affinity of unbound workqueues */ |
4cbfd3de | 6593 | list_for_each_entry(wq, &workqueues, list) { |
84193c07 TH |
6594 | struct workqueue_attrs *attrs = wq->unbound_attrs; |
6595 | ||
6596 | if (attrs) { | |
6597 | const struct wq_pod_type *pt = wqattrs_pod_type(attrs); | |
6598 | int tcpu; | |
4cbfd3de | 6599 | |
84193c07 | 6600 | for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]]) |
fef59c9c | 6601 | wq_update_pod(wq, tcpu, cpu, true); |
5797b1c1 TH |
6602 | |
6603 | mutex_lock(&wq->mutex); | |
6604 | wq_update_node_max_active(wq, -1); | |
6605 | mutex_unlock(&wq->mutex); | |
4cbfd3de TH |
6606 | } |
6607 | } | |
6ba94429 | 6608 | |
7ee681b2 TG |
6609 | mutex_unlock(&wq_pool_mutex); |
6610 | return 0; | |
6ba94429 FW |
6611 | } |
6612 | ||
7ee681b2 | 6613 | int workqueue_offline_cpu(unsigned int cpu) |
6ba94429 | 6614 | { |
6ba94429 FW |
6615 | struct workqueue_struct *wq; |
6616 | ||
7ee681b2 | 6617 | /* unbinding per-cpu workers should happen on the local CPU */ |
e8b3f8db LJ |
6618 | if (WARN_ON(cpu != smp_processor_id())) |
6619 | return -1; | |
6620 | ||
6621 | unbind_workers(cpu); | |
7ee681b2 | 6622 | |
fef59c9c | 6623 | /* update pod affinity of unbound workqueues */ |
7ee681b2 | 6624 | mutex_lock(&wq_pool_mutex); |
4cbfd3de | 6625 | list_for_each_entry(wq, &workqueues, list) { |
84193c07 TH |
6626 | struct workqueue_attrs *attrs = wq->unbound_attrs; |
6627 | ||
6628 | if (attrs) { | |
6629 | const struct wq_pod_type *pt = wqattrs_pod_type(attrs); | |
6630 | int tcpu; | |
4cbfd3de | 6631 | |
84193c07 | 6632 | for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]]) |
fef59c9c | 6633 | wq_update_pod(wq, tcpu, cpu, false); |
5797b1c1 TH |
6634 | |
6635 | mutex_lock(&wq->mutex); | |
6636 | wq_update_node_max_active(wq, cpu); | |
6637 | mutex_unlock(&wq->mutex); | |
4cbfd3de TH |
6638 | } |
6639 | } | |
7ee681b2 TG |
6640 | mutex_unlock(&wq_pool_mutex); |
6641 | ||
7ee681b2 | 6642 | return 0; |
6ba94429 FW |
6643 | } |
6644 | ||
6ba94429 FW |
6645 | struct work_for_cpu { |
6646 | struct work_struct work; | |
6647 | long (*fn)(void *); | |
6648 | void *arg; | |
6649 | long ret; | |
6650 | }; | |
6651 | ||
6652 | static void work_for_cpu_fn(struct work_struct *work) | |
6653 | { | |
6654 | struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); | |
6655 | ||
6656 | wfc->ret = wfc->fn(wfc->arg); | |
6657 | } | |
6658 | ||
6659 | /** | |
265f3ed0 | 6660 | * work_on_cpu_key - run a function in thread context on a particular cpu |
6ba94429 FW |
6661 | * @cpu: the cpu to run on |
6662 | * @fn: the function to run | |
6663 | * @arg: the function arg | |
265f3ed0 | 6664 | * @key: The lock class key for lock debugging purposes |
6ba94429 FW |
6665 | * |
6666 | * It is up to the caller to ensure that the cpu doesn't go offline. | |
6667 | * The caller must not hold any locks which would prevent @fn from completing. | |
6668 | * | |
6669 | * Return: The value @fn returns. | |
6670 | */ | |
265f3ed0 FW |
6671 | long work_on_cpu_key(int cpu, long (*fn)(void *), |
6672 | void *arg, struct lock_class_key *key) | |
6ba94429 FW |
6673 | { |
6674 | struct work_for_cpu wfc = { .fn = fn, .arg = arg }; | |
6675 | ||
265f3ed0 | 6676 | INIT_WORK_ONSTACK_KEY(&wfc.work, work_for_cpu_fn, key); |
6ba94429 FW |
6677 | schedule_work_on(cpu, &wfc.work); |
6678 | flush_work(&wfc.work); | |
6679 | destroy_work_on_stack(&wfc.work); | |
6680 | return wfc.ret; | |
6681 | } | |
265f3ed0 | 6682 | EXPORT_SYMBOL_GPL(work_on_cpu_key); |
0e8d6a93 TG |
6683 | |
6684 | /** | |
265f3ed0 | 6685 | * work_on_cpu_safe_key - run a function in thread context on a particular cpu |
0e8d6a93 TG |
6686 | * @cpu: the cpu to run on |
6687 | * @fn: the function to run | |
6688 | * @arg: the function argument | |
265f3ed0 | 6689 | * @key: The lock class key for lock debugging purposes |
0e8d6a93 TG |
6690 | * |
6691 | * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold | |
6692 | * any locks which would prevent @fn from completing. | |
6693 | * | |
6694 | * Return: The value @fn returns. | |
6695 | */ | |
265f3ed0 FW |
6696 | long work_on_cpu_safe_key(int cpu, long (*fn)(void *), |
6697 | void *arg, struct lock_class_key *key) | |
0e8d6a93 TG |
6698 | { |
6699 | long ret = -ENODEV; | |
6700 | ||
ffd8bea8 | 6701 | cpus_read_lock(); |
0e8d6a93 | 6702 | if (cpu_online(cpu)) |
265f3ed0 | 6703 | ret = work_on_cpu_key(cpu, fn, arg, key); |
ffd8bea8 | 6704 | cpus_read_unlock(); |
0e8d6a93 TG |
6705 | return ret; |
6706 | } | |
265f3ed0 | 6707 | EXPORT_SYMBOL_GPL(work_on_cpu_safe_key); |
6ba94429 FW |
6708 | #endif /* CONFIG_SMP */ |
6709 | ||
6710 | #ifdef CONFIG_FREEZER | |
6711 | ||
6712 | /** | |
6713 | * freeze_workqueues_begin - begin freezing workqueues | |
6714 | * | |
6715 | * Start freezing workqueues. After this function returns, all freezable | |
f97a4a1a | 6716 | * workqueues will queue new works to their inactive_works list instead of |
6ba94429 FW |
6717 | * pool->worklist. |
6718 | * | |
6719 | * CONTEXT: | |
6720 | * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. | |
6721 | */ | |
6722 | void freeze_workqueues_begin(void) | |
6723 | { | |
6724 | struct workqueue_struct *wq; | |
6ba94429 FW |
6725 | |
6726 | mutex_lock(&wq_pool_mutex); | |
6727 | ||
6728 | WARN_ON_ONCE(workqueue_freezing); | |
6729 | workqueue_freezing = true; | |
6730 | ||
6731 | list_for_each_entry(wq, &workqueues, list) { | |
6732 | mutex_lock(&wq->mutex); | |
a045a272 | 6733 | wq_adjust_max_active(wq); |
6ba94429 FW |
6734 | mutex_unlock(&wq->mutex); |
6735 | } | |
6736 | ||
6737 | mutex_unlock(&wq_pool_mutex); | |
6738 | } | |
6739 | ||
6740 | /** | |
6741 | * freeze_workqueues_busy - are freezable workqueues still busy? | |
6742 | * | |
6743 | * Check whether freezing is complete. This function must be called | |
6744 | * between freeze_workqueues_begin() and thaw_workqueues(). | |
6745 | * | |
6746 | * CONTEXT: | |
6747 | * Grabs and releases wq_pool_mutex. | |
6748 | * | |
6749 | * Return: | |
6750 | * %true if some freezable workqueues are still busy. %false if freezing | |
6751 | * is complete. | |
6752 | */ | |
6753 | bool freeze_workqueues_busy(void) | |
6754 | { | |
6755 | bool busy = false; | |
6756 | struct workqueue_struct *wq; | |
6757 | struct pool_workqueue *pwq; | |
6758 | ||
6759 | mutex_lock(&wq_pool_mutex); | |
6760 | ||
6761 | WARN_ON_ONCE(!workqueue_freezing); | |
6762 | ||
6763 | list_for_each_entry(wq, &workqueues, list) { | |
6764 | if (!(wq->flags & WQ_FREEZABLE)) | |
6765 | continue; | |
6766 | /* | |
6767 | * nr_active is monotonically decreasing. It's safe | |
6768 | * to peek without lock. | |
6769 | */ | |
24acfb71 | 6770 | rcu_read_lock(); |
6ba94429 FW |
6771 | for_each_pwq(pwq, wq) { |
6772 | WARN_ON_ONCE(pwq->nr_active < 0); | |
6773 | if (pwq->nr_active) { | |
6774 | busy = true; | |
24acfb71 | 6775 | rcu_read_unlock(); |
6ba94429 FW |
6776 | goto out_unlock; |
6777 | } | |
6778 | } | |
24acfb71 | 6779 | rcu_read_unlock(); |
6ba94429 FW |
6780 | } |
6781 | out_unlock: | |
6782 | mutex_unlock(&wq_pool_mutex); | |
6783 | return busy; | |
6784 | } | |
6785 | ||
6786 | /** | |
6787 | * thaw_workqueues - thaw workqueues | |
6788 | * | |
6789 | * Thaw workqueues. Normal queueing is restored and all collected | |
6790 | * frozen works are transferred to their respective pool worklists. | |
6791 | * | |
6792 | * CONTEXT: | |
6793 | * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. | |
6794 | */ | |
6795 | void thaw_workqueues(void) | |
6796 | { | |
6797 | struct workqueue_struct *wq; | |
6ba94429 FW |
6798 | |
6799 | mutex_lock(&wq_pool_mutex); | |
6800 | ||
6801 | if (!workqueue_freezing) | |
6802 | goto out_unlock; | |
6803 | ||
6804 | workqueue_freezing = false; | |
6805 | ||
6806 | /* restore max_active and repopulate worklist */ | |
6807 | list_for_each_entry(wq, &workqueues, list) { | |
6808 | mutex_lock(&wq->mutex); | |
a045a272 | 6809 | wq_adjust_max_active(wq); |
6ba94429 FW |
6810 | mutex_unlock(&wq->mutex); |
6811 | } | |
6812 | ||
6813 | out_unlock: | |
6814 | mutex_unlock(&wq_pool_mutex); | |
6815 | } | |
6816 | #endif /* CONFIG_FREEZER */ | |
6817 | ||
99c621ef | 6818 | static int workqueue_apply_unbound_cpumask(const cpumask_var_t unbound_cpumask) |
042f7df1 LJ |
6819 | { |
6820 | LIST_HEAD(ctxs); | |
6821 | int ret = 0; | |
6822 | struct workqueue_struct *wq; | |
6823 | struct apply_wqattrs_ctx *ctx, *n; | |
6824 | ||
6825 | lockdep_assert_held(&wq_pool_mutex); | |
6826 | ||
6827 | list_for_each_entry(wq, &workqueues, list) { | |
8eb17dc1 | 6828 | if (!(wq->flags & WQ_UNBOUND) || (wq->flags & __WQ_DESTROYING)) |
042f7df1 | 6829 | continue; |
042f7df1 | 6830 | |
99c621ef | 6831 | ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs, unbound_cpumask); |
84193c07 TH |
6832 | if (IS_ERR(ctx)) { |
6833 | ret = PTR_ERR(ctx); | |
042f7df1 LJ |
6834 | break; |
6835 | } | |
6836 | ||
6837 | list_add_tail(&ctx->list, &ctxs); | |
6838 | } | |
6839 | ||
6840 | list_for_each_entry_safe(ctx, n, &ctxs, list) { | |
6841 | if (!ret) | |
6842 | apply_wqattrs_commit(ctx); | |
6843 | apply_wqattrs_cleanup(ctx); | |
6844 | } | |
6845 | ||
99c621ef LJ |
6846 | if (!ret) { |
6847 | mutex_lock(&wq_pool_attach_mutex); | |
6848 | cpumask_copy(wq_unbound_cpumask, unbound_cpumask); | |
6849 | mutex_unlock(&wq_pool_attach_mutex); | |
6850 | } | |
042f7df1 LJ |
6851 | return ret; |
6852 | } | |
6853 | ||
fe28f631 WL |
6854 | /** |
6855 | * workqueue_unbound_exclude_cpumask - Exclude given CPUs from unbound cpumask | |
6856 | * @exclude_cpumask: the cpumask to be excluded from wq_unbound_cpumask | |
6857 | * | |
6858 | * This function can be called from cpuset code to provide a set of isolated | |
6859 | * CPUs that should be excluded from wq_unbound_cpumask. The caller must hold | |
6860 | * either cpus_read_lock or cpus_write_lock. | |
6861 | */ | |
6862 | int workqueue_unbound_exclude_cpumask(cpumask_var_t exclude_cpumask) | |
6863 | { | |
6864 | cpumask_var_t cpumask; | |
6865 | int ret = 0; | |
6866 | ||
6867 | if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) | |
6868 | return -ENOMEM; | |
6869 | ||
6870 | lockdep_assert_cpus_held(); | |
6871 | mutex_lock(&wq_pool_mutex); | |
6872 | ||
6873 | /* Save the current isolated cpumask & export it via sysfs */ | |
6874 | cpumask_copy(wq_isolated_cpumask, exclude_cpumask); | |
6875 | ||
6876 | /* | |
6877 | * If the operation fails, it will fall back to | |
6878 | * wq_requested_unbound_cpumask which is initially set to | |
6879 | * (HK_TYPE_WQ ∩ HK_TYPE_DOMAIN) house keeping mask and rewritten | |
6880 | * by any subsequent write to workqueue/cpumask sysfs file. | |
6881 | */ | |
6882 | if (!cpumask_andnot(cpumask, wq_requested_unbound_cpumask, exclude_cpumask)) | |
6883 | cpumask_copy(cpumask, wq_requested_unbound_cpumask); | |
6884 | if (!cpumask_equal(cpumask, wq_unbound_cpumask)) | |
6885 | ret = workqueue_apply_unbound_cpumask(cpumask); | |
6886 | ||
6887 | mutex_unlock(&wq_pool_mutex); | |
6888 | free_cpumask_var(cpumask); | |
6889 | return ret; | |
6890 | } | |
6891 | ||
63c5484e TH |
6892 | static int parse_affn_scope(const char *val) |
6893 | { | |
6894 | int i; | |
6895 | ||
6896 | for (i = 0; i < ARRAY_SIZE(wq_affn_names); i++) { | |
6897 | if (!strncasecmp(val, wq_affn_names[i], strlen(wq_affn_names[i]))) | |
6898 | return i; | |
6899 | } | |
6900 | return -EINVAL; | |
6901 | } | |
6902 | ||
6903 | static int wq_affn_dfl_set(const char *val, const struct kernel_param *kp) | |
6904 | { | |
523a301e TH |
6905 | struct workqueue_struct *wq; |
6906 | int affn, cpu; | |
63c5484e TH |
6907 | |
6908 | affn = parse_affn_scope(val); | |
6909 | if (affn < 0) | |
6910 | return affn; | |
523a301e TH |
6911 | if (affn == WQ_AFFN_DFL) |
6912 | return -EINVAL; | |
6913 | ||
6914 | cpus_read_lock(); | |
6915 | mutex_lock(&wq_pool_mutex); | |
63c5484e TH |
6916 | |
6917 | wq_affn_dfl = affn; | |
523a301e TH |
6918 | |
6919 | list_for_each_entry(wq, &workqueues, list) { | |
6920 | for_each_online_cpu(cpu) { | |
6921 | wq_update_pod(wq, cpu, cpu, true); | |
6922 | } | |
6923 | } | |
6924 | ||
6925 | mutex_unlock(&wq_pool_mutex); | |
6926 | cpus_read_unlock(); | |
6927 | ||
63c5484e TH |
6928 | return 0; |
6929 | } | |
6930 | ||
6931 | static int wq_affn_dfl_get(char *buffer, const struct kernel_param *kp) | |
6932 | { | |
6933 | return scnprintf(buffer, PAGE_SIZE, "%s\n", wq_affn_names[wq_affn_dfl]); | |
6934 | } | |
6935 | ||
6936 | static const struct kernel_param_ops wq_affn_dfl_ops = { | |
6937 | .set = wq_affn_dfl_set, | |
6938 | .get = wq_affn_dfl_get, | |
6939 | }; | |
6940 | ||
6941 | module_param_cb(default_affinity_scope, &wq_affn_dfl_ops, NULL, 0644); | |
6942 | ||
6ba94429 FW |
6943 | #ifdef CONFIG_SYSFS |
6944 | /* | |
6945 | * Workqueues with WQ_SYSFS flag set is visible to userland via | |
6946 | * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the | |
6947 | * following attributes. | |
6948 | * | |
63c5484e TH |
6949 | * per_cpu RO bool : whether the workqueue is per-cpu or unbound |
6950 | * max_active RW int : maximum number of in-flight work items | |
6ba94429 FW |
6951 | * |
6952 | * Unbound workqueues have the following extra attributes. | |
6953 | * | |
63c5484e TH |
6954 | * nice RW int : nice value of the workers |
6955 | * cpumask RW mask : bitmask of allowed CPUs for the workers | |
6956 | * affinity_scope RW str : worker CPU affinity scope (cache, numa, none) | |
8639eceb | 6957 | * affinity_strict RW bool : worker CPU affinity is strict |
6ba94429 FW |
6958 | */ |
6959 | struct wq_device { | |
6960 | struct workqueue_struct *wq; | |
6961 | struct device dev; | |
6962 | }; | |
6963 | ||
6964 | static struct workqueue_struct *dev_to_wq(struct device *dev) | |
6965 | { | |
6966 | struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); | |
6967 | ||
6968 | return wq_dev->wq; | |
6969 | } | |
6970 | ||
6971 | static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, | |
6972 | char *buf) | |
6973 | { | |
6974 | struct workqueue_struct *wq = dev_to_wq(dev); | |
6975 | ||
6976 | return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); | |
6977 | } | |
6978 | static DEVICE_ATTR_RO(per_cpu); | |
6979 | ||
6980 | static ssize_t max_active_show(struct device *dev, | |
6981 | struct device_attribute *attr, char *buf) | |
6982 | { | |
6983 | struct workqueue_struct *wq = dev_to_wq(dev); | |
6984 | ||
6985 | return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); | |
6986 | } | |
6987 | ||
6988 | static ssize_t max_active_store(struct device *dev, | |
6989 | struct device_attribute *attr, const char *buf, | |
6990 | size_t count) | |
6991 | { | |
6992 | struct workqueue_struct *wq = dev_to_wq(dev); | |
6993 | int val; | |
6994 | ||
6995 | if (sscanf(buf, "%d", &val) != 1 || val <= 0) | |
6996 | return -EINVAL; | |
6997 | ||
6998 | workqueue_set_max_active(wq, val); | |
6999 | return count; | |
7000 | } | |
7001 | static DEVICE_ATTR_RW(max_active); | |
7002 | ||
7003 | static struct attribute *wq_sysfs_attrs[] = { | |
7004 | &dev_attr_per_cpu.attr, | |
7005 | &dev_attr_max_active.attr, | |
7006 | NULL, | |
7007 | }; | |
7008 | ATTRIBUTE_GROUPS(wq_sysfs); | |
7009 | ||
49277a5b WL |
7010 | static void apply_wqattrs_lock(void) |
7011 | { | |
7012 | /* CPUs should stay stable across pwq creations and installations */ | |
7013 | cpus_read_lock(); | |
7014 | mutex_lock(&wq_pool_mutex); | |
7015 | } | |
7016 | ||
7017 | static void apply_wqattrs_unlock(void) | |
7018 | { | |
7019 | mutex_unlock(&wq_pool_mutex); | |
7020 | cpus_read_unlock(); | |
7021 | } | |
7022 | ||
6ba94429 FW |
7023 | static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, |
7024 | char *buf) | |
7025 | { | |
7026 | struct workqueue_struct *wq = dev_to_wq(dev); | |
7027 | int written; | |
7028 | ||
7029 | mutex_lock(&wq->mutex); | |
7030 | written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); | |
7031 | mutex_unlock(&wq->mutex); | |
7032 | ||
7033 | return written; | |
7034 | } | |
7035 | ||
7036 | /* prepare workqueue_attrs for sysfs store operations */ | |
7037 | static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) | |
7038 | { | |
7039 | struct workqueue_attrs *attrs; | |
7040 | ||
899a94fe LJ |
7041 | lockdep_assert_held(&wq_pool_mutex); |
7042 | ||
be69d00d | 7043 | attrs = alloc_workqueue_attrs(); |
6ba94429 FW |
7044 | if (!attrs) |
7045 | return NULL; | |
7046 | ||
6ba94429 | 7047 | copy_workqueue_attrs(attrs, wq->unbound_attrs); |
6ba94429 FW |
7048 | return attrs; |
7049 | } | |
7050 | ||
7051 | static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, | |
7052 | const char *buf, size_t count) | |
7053 | { | |
7054 | struct workqueue_struct *wq = dev_to_wq(dev); | |
7055 | struct workqueue_attrs *attrs; | |
d4d3e257 LJ |
7056 | int ret = -ENOMEM; |
7057 | ||
7058 | apply_wqattrs_lock(); | |
6ba94429 FW |
7059 | |
7060 | attrs = wq_sysfs_prep_attrs(wq); | |
7061 | if (!attrs) | |
d4d3e257 | 7062 | goto out_unlock; |
6ba94429 FW |
7063 | |
7064 | if (sscanf(buf, "%d", &attrs->nice) == 1 && | |
7065 | attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) | |
d4d3e257 | 7066 | ret = apply_workqueue_attrs_locked(wq, attrs); |
6ba94429 FW |
7067 | else |
7068 | ret = -EINVAL; | |
7069 | ||
d4d3e257 LJ |
7070 | out_unlock: |
7071 | apply_wqattrs_unlock(); | |
6ba94429 FW |
7072 | free_workqueue_attrs(attrs); |
7073 | return ret ?: count; | |
7074 | } | |
7075 | ||
7076 | static ssize_t wq_cpumask_show(struct device *dev, | |
7077 | struct device_attribute *attr, char *buf) | |
7078 | { | |
7079 | struct workqueue_struct *wq = dev_to_wq(dev); | |
7080 | int written; | |
7081 | ||
7082 | mutex_lock(&wq->mutex); | |
7083 | written = scnprintf(buf, PAGE_SIZE, "%*pb\n", | |
7084 | cpumask_pr_args(wq->unbound_attrs->cpumask)); | |
7085 | mutex_unlock(&wq->mutex); | |
7086 | return written; | |
7087 | } | |
7088 | ||
7089 | static ssize_t wq_cpumask_store(struct device *dev, | |
7090 | struct device_attribute *attr, | |
7091 | const char *buf, size_t count) | |
7092 | { | |
7093 | struct workqueue_struct *wq = dev_to_wq(dev); | |
7094 | struct workqueue_attrs *attrs; | |
d4d3e257 LJ |
7095 | int ret = -ENOMEM; |
7096 | ||
7097 | apply_wqattrs_lock(); | |
6ba94429 FW |
7098 | |
7099 | attrs = wq_sysfs_prep_attrs(wq); | |
7100 | if (!attrs) | |
d4d3e257 | 7101 | goto out_unlock; |
6ba94429 FW |
7102 | |
7103 | ret = cpumask_parse(buf, attrs->cpumask); | |
7104 | if (!ret) | |
d4d3e257 | 7105 | ret = apply_workqueue_attrs_locked(wq, attrs); |
6ba94429 | 7106 | |
d4d3e257 LJ |
7107 | out_unlock: |
7108 | apply_wqattrs_unlock(); | |
6ba94429 FW |
7109 | free_workqueue_attrs(attrs); |
7110 | return ret ?: count; | |
7111 | } | |
7112 | ||
63c5484e TH |
7113 | static ssize_t wq_affn_scope_show(struct device *dev, |
7114 | struct device_attribute *attr, char *buf) | |
7115 | { | |
7116 | struct workqueue_struct *wq = dev_to_wq(dev); | |
7117 | int written; | |
7118 | ||
7119 | mutex_lock(&wq->mutex); | |
523a301e TH |
7120 | if (wq->unbound_attrs->affn_scope == WQ_AFFN_DFL) |
7121 | written = scnprintf(buf, PAGE_SIZE, "%s (%s)\n", | |
7122 | wq_affn_names[WQ_AFFN_DFL], | |
7123 | wq_affn_names[wq_affn_dfl]); | |
7124 | else | |
7125 | written = scnprintf(buf, PAGE_SIZE, "%s\n", | |
7126 | wq_affn_names[wq->unbound_attrs->affn_scope]); | |
63c5484e TH |
7127 | mutex_unlock(&wq->mutex); |
7128 | ||
7129 | return written; | |
7130 | } | |
7131 | ||
7132 | static ssize_t wq_affn_scope_store(struct device *dev, | |
7133 | struct device_attribute *attr, | |
7134 | const char *buf, size_t count) | |
7135 | { | |
7136 | struct workqueue_struct *wq = dev_to_wq(dev); | |
7137 | struct workqueue_attrs *attrs; | |
7138 | int affn, ret = -ENOMEM; | |
7139 | ||
7140 | affn = parse_affn_scope(buf); | |
7141 | if (affn < 0) | |
7142 | return affn; | |
7143 | ||
7144 | apply_wqattrs_lock(); | |
7145 | attrs = wq_sysfs_prep_attrs(wq); | |
7146 | if (attrs) { | |
7147 | attrs->affn_scope = affn; | |
7148 | ret = apply_workqueue_attrs_locked(wq, attrs); | |
7149 | } | |
7150 | apply_wqattrs_unlock(); | |
7151 | free_workqueue_attrs(attrs); | |
7152 | return ret ?: count; | |
7153 | } | |
7154 | ||
8639eceb TH |
7155 | static ssize_t wq_affinity_strict_show(struct device *dev, |
7156 | struct device_attribute *attr, char *buf) | |
7157 | { | |
7158 | struct workqueue_struct *wq = dev_to_wq(dev); | |
7159 | ||
7160 | return scnprintf(buf, PAGE_SIZE, "%d\n", | |
7161 | wq->unbound_attrs->affn_strict); | |
7162 | } | |
7163 | ||
7164 | static ssize_t wq_affinity_strict_store(struct device *dev, | |
7165 | struct device_attribute *attr, | |
7166 | const char *buf, size_t count) | |
7167 | { | |
7168 | struct workqueue_struct *wq = dev_to_wq(dev); | |
7169 | struct workqueue_attrs *attrs; | |
7170 | int v, ret = -ENOMEM; | |
7171 | ||
7172 | if (sscanf(buf, "%d", &v) != 1) | |
7173 | return -EINVAL; | |
7174 | ||
7175 | apply_wqattrs_lock(); | |
7176 | attrs = wq_sysfs_prep_attrs(wq); | |
7177 | if (attrs) { | |
7178 | attrs->affn_strict = (bool)v; | |
7179 | ret = apply_workqueue_attrs_locked(wq, attrs); | |
7180 | } | |
7181 | apply_wqattrs_unlock(); | |
7182 | free_workqueue_attrs(attrs); | |
7183 | return ret ?: count; | |
7184 | } | |
7185 | ||
6ba94429 | 7186 | static struct device_attribute wq_sysfs_unbound_attrs[] = { |
6ba94429 FW |
7187 | __ATTR(nice, 0644, wq_nice_show, wq_nice_store), |
7188 | __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), | |
63c5484e | 7189 | __ATTR(affinity_scope, 0644, wq_affn_scope_show, wq_affn_scope_store), |
8639eceb | 7190 | __ATTR(affinity_strict, 0644, wq_affinity_strict_show, wq_affinity_strict_store), |
6ba94429 FW |
7191 | __ATTR_NULL, |
7192 | }; | |
8ccad40d | 7193 | |
5df9197e | 7194 | static const struct bus_type wq_subsys = { |
6ba94429 FW |
7195 | .name = "workqueue", |
7196 | .dev_groups = wq_sysfs_groups, | |
2d3854a3 RR |
7197 | }; |
7198 | ||
49277a5b WL |
7199 | /** |
7200 | * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask | |
7201 | * @cpumask: the cpumask to set | |
7202 | * | |
7203 | * The low-level workqueues cpumask is a global cpumask that limits | |
7204 | * the affinity of all unbound workqueues. This function check the @cpumask | |
7205 | * and apply it to all unbound workqueues and updates all pwqs of them. | |
7206 | * | |
7207 | * Return: 0 - Success | |
7208 | * -EINVAL - Invalid @cpumask | |
7209 | * -ENOMEM - Failed to allocate memory for attrs or pwqs. | |
7210 | */ | |
7211 | static int workqueue_set_unbound_cpumask(cpumask_var_t cpumask) | |
7212 | { | |
7213 | int ret = -EINVAL; | |
7214 | ||
7215 | /* | |
7216 | * Not excluding isolated cpus on purpose. | |
7217 | * If the user wishes to include them, we allow that. | |
7218 | */ | |
7219 | cpumask_and(cpumask, cpumask, cpu_possible_mask); | |
7220 | if (!cpumask_empty(cpumask)) { | |
7221 | apply_wqattrs_lock(); | |
7222 | cpumask_copy(wq_requested_unbound_cpumask, cpumask); | |
7223 | if (cpumask_equal(cpumask, wq_unbound_cpumask)) { | |
7224 | ret = 0; | |
7225 | goto out_unlock; | |
7226 | } | |
7227 | ||
7228 | ret = workqueue_apply_unbound_cpumask(cpumask); | |
7229 | ||
7230 | out_unlock: | |
7231 | apply_wqattrs_unlock(); | |
7232 | } | |
7233 | ||
7234 | return ret; | |
7235 | } | |
7236 | ||
fe28f631 WL |
7237 | static ssize_t __wq_cpumask_show(struct device *dev, |
7238 | struct device_attribute *attr, char *buf, cpumask_var_t mask) | |
b05a7928 FW |
7239 | { |
7240 | int written; | |
7241 | ||
042f7df1 | 7242 | mutex_lock(&wq_pool_mutex); |
fe28f631 | 7243 | written = scnprintf(buf, PAGE_SIZE, "%*pb\n", cpumask_pr_args(mask)); |
042f7df1 | 7244 | mutex_unlock(&wq_pool_mutex); |
b05a7928 FW |
7245 | |
7246 | return written; | |
7247 | } | |
7248 | ||
fe28f631 WL |
7249 | static ssize_t wq_unbound_cpumask_show(struct device *dev, |
7250 | struct device_attribute *attr, char *buf) | |
7251 | { | |
7252 | return __wq_cpumask_show(dev, attr, buf, wq_unbound_cpumask); | |
7253 | } | |
7254 | ||
7255 | static ssize_t wq_requested_cpumask_show(struct device *dev, | |
7256 | struct device_attribute *attr, char *buf) | |
7257 | { | |
7258 | return __wq_cpumask_show(dev, attr, buf, wq_requested_unbound_cpumask); | |
7259 | } | |
7260 | ||
7261 | static ssize_t wq_isolated_cpumask_show(struct device *dev, | |
7262 | struct device_attribute *attr, char *buf) | |
7263 | { | |
7264 | return __wq_cpumask_show(dev, attr, buf, wq_isolated_cpumask); | |
7265 | } | |
7266 | ||
042f7df1 LJ |
7267 | static ssize_t wq_unbound_cpumask_store(struct device *dev, |
7268 | struct device_attribute *attr, const char *buf, size_t count) | |
7269 | { | |
7270 | cpumask_var_t cpumask; | |
7271 | int ret; | |
7272 | ||
7273 | if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) | |
7274 | return -ENOMEM; | |
7275 | ||
7276 | ret = cpumask_parse(buf, cpumask); | |
7277 | if (!ret) | |
7278 | ret = workqueue_set_unbound_cpumask(cpumask); | |
7279 | ||
7280 | free_cpumask_var(cpumask); | |
7281 | return ret ? ret : count; | |
7282 | } | |
7283 | ||
fe28f631 | 7284 | static struct device_attribute wq_sysfs_cpumask_attrs[] = { |
042f7df1 | 7285 | __ATTR(cpumask, 0644, wq_unbound_cpumask_show, |
fe28f631 WL |
7286 | wq_unbound_cpumask_store), |
7287 | __ATTR(cpumask_requested, 0444, wq_requested_cpumask_show, NULL), | |
7288 | __ATTR(cpumask_isolated, 0444, wq_isolated_cpumask_show, NULL), | |
7289 | __ATTR_NULL, | |
7290 | }; | |
b05a7928 | 7291 | |
6ba94429 | 7292 | static int __init wq_sysfs_init(void) |
2d3854a3 | 7293 | { |
686f6697 | 7294 | struct device *dev_root; |
b05a7928 FW |
7295 | int err; |
7296 | ||
7297 | err = subsys_virtual_register(&wq_subsys, NULL); | |
7298 | if (err) | |
7299 | return err; | |
7300 | ||
686f6697 GKH |
7301 | dev_root = bus_get_dev_root(&wq_subsys); |
7302 | if (dev_root) { | |
fe28f631 WL |
7303 | struct device_attribute *attr; |
7304 | ||
7305 | for (attr = wq_sysfs_cpumask_attrs; attr->attr.name; attr++) { | |
7306 | err = device_create_file(dev_root, attr); | |
7307 | if (err) | |
7308 | break; | |
7309 | } | |
686f6697 GKH |
7310 | put_device(dev_root); |
7311 | } | |
7312 | return err; | |
2d3854a3 | 7313 | } |
6ba94429 | 7314 | core_initcall(wq_sysfs_init); |
2d3854a3 | 7315 | |
6ba94429 | 7316 | static void wq_device_release(struct device *dev) |
2d3854a3 | 7317 | { |
6ba94429 | 7318 | struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); |
6b44003e | 7319 | |
6ba94429 | 7320 | kfree(wq_dev); |
2d3854a3 | 7321 | } |
a0a1a5fd TH |
7322 | |
7323 | /** | |
6ba94429 FW |
7324 | * workqueue_sysfs_register - make a workqueue visible in sysfs |
7325 | * @wq: the workqueue to register | |
a0a1a5fd | 7326 | * |
6ba94429 FW |
7327 | * Expose @wq in sysfs under /sys/bus/workqueue/devices. |
7328 | * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set | |
7329 | * which is the preferred method. | |
a0a1a5fd | 7330 | * |
6ba94429 FW |
7331 | * Workqueue user should use this function directly iff it wants to apply |
7332 | * workqueue_attrs before making the workqueue visible in sysfs; otherwise, | |
7333 | * apply_workqueue_attrs() may race against userland updating the | |
7334 | * attributes. | |
7335 | * | |
7336 | * Return: 0 on success, -errno on failure. | |
a0a1a5fd | 7337 | */ |
6ba94429 | 7338 | int workqueue_sysfs_register(struct workqueue_struct *wq) |
a0a1a5fd | 7339 | { |
6ba94429 FW |
7340 | struct wq_device *wq_dev; |
7341 | int ret; | |
a0a1a5fd | 7342 | |
6ba94429 | 7343 | /* |
4c065dbc WL |
7344 | * Adjusting max_active breaks ordering guarantee. Disallow exposing |
7345 | * ordered workqueues. | |
6ba94429 | 7346 | */ |
3bc1e711 | 7347 | if (WARN_ON(wq->flags & __WQ_ORDERED)) |
6ba94429 | 7348 | return -EINVAL; |
a0a1a5fd | 7349 | |
6ba94429 FW |
7350 | wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); |
7351 | if (!wq_dev) | |
7352 | return -ENOMEM; | |
5bcab335 | 7353 | |
6ba94429 FW |
7354 | wq_dev->wq = wq; |
7355 | wq_dev->dev.bus = &wq_subsys; | |
6ba94429 | 7356 | wq_dev->dev.release = wq_device_release; |
23217b44 | 7357 | dev_set_name(&wq_dev->dev, "%s", wq->name); |
a0a1a5fd | 7358 | |
6ba94429 FW |
7359 | /* |
7360 | * unbound_attrs are created separately. Suppress uevent until | |
7361 | * everything is ready. | |
7362 | */ | |
7363 | dev_set_uevent_suppress(&wq_dev->dev, true); | |
a0a1a5fd | 7364 | |
6ba94429 FW |
7365 | ret = device_register(&wq_dev->dev); |
7366 | if (ret) { | |
537f4146 | 7367 | put_device(&wq_dev->dev); |
6ba94429 FW |
7368 | wq->wq_dev = NULL; |
7369 | return ret; | |
7370 | } | |
a0a1a5fd | 7371 | |
6ba94429 FW |
7372 | if (wq->flags & WQ_UNBOUND) { |
7373 | struct device_attribute *attr; | |
a0a1a5fd | 7374 | |
6ba94429 FW |
7375 | for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { |
7376 | ret = device_create_file(&wq_dev->dev, attr); | |
7377 | if (ret) { | |
7378 | device_unregister(&wq_dev->dev); | |
7379 | wq->wq_dev = NULL; | |
7380 | return ret; | |
a0a1a5fd TH |
7381 | } |
7382 | } | |
7383 | } | |
6ba94429 FW |
7384 | |
7385 | dev_set_uevent_suppress(&wq_dev->dev, false); | |
7386 | kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); | |
7387 | return 0; | |
a0a1a5fd TH |
7388 | } |
7389 | ||
7390 | /** | |
6ba94429 FW |
7391 | * workqueue_sysfs_unregister - undo workqueue_sysfs_register() |
7392 | * @wq: the workqueue to unregister | |
a0a1a5fd | 7393 | * |
6ba94429 | 7394 | * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. |
a0a1a5fd | 7395 | */ |
6ba94429 | 7396 | static void workqueue_sysfs_unregister(struct workqueue_struct *wq) |
a0a1a5fd | 7397 | { |
6ba94429 | 7398 | struct wq_device *wq_dev = wq->wq_dev; |
8b03ae3c | 7399 | |
6ba94429 FW |
7400 | if (!wq->wq_dev) |
7401 | return; | |
a0a1a5fd | 7402 | |
6ba94429 FW |
7403 | wq->wq_dev = NULL; |
7404 | device_unregister(&wq_dev->dev); | |
a0a1a5fd | 7405 | } |
6ba94429 FW |
7406 | #else /* CONFIG_SYSFS */ |
7407 | static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } | |
7408 | #endif /* CONFIG_SYSFS */ | |
a0a1a5fd | 7409 | |
82607adc TH |
7410 | /* |
7411 | * Workqueue watchdog. | |
7412 | * | |
7413 | * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal | |
7414 | * flush dependency, a concurrency managed work item which stays RUNNING | |
7415 | * indefinitely. Workqueue stalls can be very difficult to debug as the | |
7416 | * usual warning mechanisms don't trigger and internal workqueue state is | |
7417 | * largely opaque. | |
7418 | * | |
7419 | * Workqueue watchdog monitors all worker pools periodically and dumps | |
7420 | * state if some pools failed to make forward progress for a while where | |
7421 | * forward progress is defined as the first item on ->worklist changing. | |
7422 | * | |
7423 | * This mechanism is controlled through the kernel parameter | |
7424 | * "workqueue.watchdog_thresh" which can be updated at runtime through the | |
7425 | * corresponding sysfs parameter file. | |
7426 | */ | |
7427 | #ifdef CONFIG_WQ_WATCHDOG | |
7428 | ||
82607adc | 7429 | static unsigned long wq_watchdog_thresh = 30; |
5cd79d6a | 7430 | static struct timer_list wq_watchdog_timer; |
82607adc TH |
7431 | |
7432 | static unsigned long wq_watchdog_touched = INITIAL_JIFFIES; | |
7433 | static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES; | |
7434 | ||
cd2440d6 PM |
7435 | /* |
7436 | * Show workers that might prevent the processing of pending work items. | |
7437 | * The only candidates are CPU-bound workers in the running state. | |
7438 | * Pending work items should be handled by another idle worker | |
7439 | * in all other situations. | |
7440 | */ | |
7441 | static void show_cpu_pool_hog(struct worker_pool *pool) | |
7442 | { | |
7443 | struct worker *worker; | |
c26e2f2e | 7444 | unsigned long irq_flags; |
cd2440d6 PM |
7445 | int bkt; |
7446 | ||
c26e2f2e | 7447 | raw_spin_lock_irqsave(&pool->lock, irq_flags); |
cd2440d6 PM |
7448 | |
7449 | hash_for_each(pool->busy_hash, bkt, worker, hentry) { | |
7450 | if (task_is_running(worker->task)) { | |
7451 | /* | |
7452 | * Defer printing to avoid deadlocks in console | |
7453 | * drivers that queue work while holding locks | |
7454 | * also taken in their write paths. | |
7455 | */ | |
7456 | printk_deferred_enter(); | |
7457 | ||
7458 | pr_info("pool %d:\n", pool->id); | |
7459 | sched_show_task(worker->task); | |
7460 | ||
7461 | printk_deferred_exit(); | |
7462 | } | |
7463 | } | |
7464 | ||
c26e2f2e | 7465 | raw_spin_unlock_irqrestore(&pool->lock, irq_flags); |
cd2440d6 PM |
7466 | } |
7467 | ||
7468 | static void show_cpu_pools_hogs(void) | |
7469 | { | |
7470 | struct worker_pool *pool; | |
7471 | int pi; | |
7472 | ||
7473 | pr_info("Showing backtraces of running workers in stalled CPU-bound worker pools:\n"); | |
7474 | ||
7475 | rcu_read_lock(); | |
7476 | ||
7477 | for_each_pool(pool, pi) { | |
7478 | if (pool->cpu_stall) | |
7479 | show_cpu_pool_hog(pool); | |
7480 | ||
7481 | } | |
7482 | ||
7483 | rcu_read_unlock(); | |
7484 | } | |
7485 | ||
82607adc TH |
7486 | static void wq_watchdog_reset_touched(void) |
7487 | { | |
7488 | int cpu; | |
7489 | ||
7490 | wq_watchdog_touched = jiffies; | |
7491 | for_each_possible_cpu(cpu) | |
7492 | per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; | |
7493 | } | |
7494 | ||
5cd79d6a | 7495 | static void wq_watchdog_timer_fn(struct timer_list *unused) |
82607adc TH |
7496 | { |
7497 | unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ; | |
7498 | bool lockup_detected = false; | |
cd2440d6 | 7499 | bool cpu_pool_stall = false; |
940d71c6 | 7500 | unsigned long now = jiffies; |
82607adc TH |
7501 | struct worker_pool *pool; |
7502 | int pi; | |
7503 | ||
7504 | if (!thresh) | |
7505 | return; | |
7506 | ||
7507 | rcu_read_lock(); | |
7508 | ||
7509 | for_each_pool(pool, pi) { | |
7510 | unsigned long pool_ts, touched, ts; | |
7511 | ||
cd2440d6 | 7512 | pool->cpu_stall = false; |
82607adc TH |
7513 | if (list_empty(&pool->worklist)) |
7514 | continue; | |
7515 | ||
940d71c6 SS |
7516 | /* |
7517 | * If a virtual machine is stopped by the host it can look to | |
7518 | * the watchdog like a stall. | |
7519 | */ | |
7520 | kvm_check_and_clear_guest_paused(); | |
7521 | ||
82607adc | 7522 | /* get the latest of pool and touched timestamps */ |
89e28ce6 WQ |
7523 | if (pool->cpu >= 0) |
7524 | touched = READ_ONCE(per_cpu(wq_watchdog_touched_cpu, pool->cpu)); | |
7525 | else | |
7526 | touched = READ_ONCE(wq_watchdog_touched); | |
82607adc | 7527 | pool_ts = READ_ONCE(pool->watchdog_ts); |
82607adc TH |
7528 | |
7529 | if (time_after(pool_ts, touched)) | |
7530 | ts = pool_ts; | |
7531 | else | |
7532 | ts = touched; | |
7533 | ||
82607adc | 7534 | /* did we stall? */ |
940d71c6 | 7535 | if (time_after(now, ts + thresh)) { |
82607adc | 7536 | lockup_detected = true; |
4cb1ef64 | 7537 | if (pool->cpu >= 0 && !(pool->flags & POOL_BH)) { |
cd2440d6 PM |
7538 | pool->cpu_stall = true; |
7539 | cpu_pool_stall = true; | |
7540 | } | |
82607adc TH |
7541 | pr_emerg("BUG: workqueue lockup - pool"); |
7542 | pr_cont_pool_info(pool); | |
7543 | pr_cont(" stuck for %us!\n", | |
940d71c6 | 7544 | jiffies_to_msecs(now - pool_ts) / 1000); |
82607adc | 7545 | } |
cd2440d6 PM |
7546 | |
7547 | ||
82607adc TH |
7548 | } |
7549 | ||
7550 | rcu_read_unlock(); | |
7551 | ||
7552 | if (lockup_detected) | |
55df0933 | 7553 | show_all_workqueues(); |
82607adc | 7554 | |
cd2440d6 PM |
7555 | if (cpu_pool_stall) |
7556 | show_cpu_pools_hogs(); | |
7557 | ||
82607adc TH |
7558 | wq_watchdog_reset_touched(); |
7559 | mod_timer(&wq_watchdog_timer, jiffies + thresh); | |
7560 | } | |
7561 | ||
cb9d7fd5 | 7562 | notrace void wq_watchdog_touch(int cpu) |
82607adc TH |
7563 | { |
7564 | if (cpu >= 0) | |
7565 | per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; | |
89e28ce6 WQ |
7566 | |
7567 | wq_watchdog_touched = jiffies; | |
82607adc TH |
7568 | } |
7569 | ||
7570 | static void wq_watchdog_set_thresh(unsigned long thresh) | |
7571 | { | |
7572 | wq_watchdog_thresh = 0; | |
7573 | del_timer_sync(&wq_watchdog_timer); | |
7574 | ||
7575 | if (thresh) { | |
7576 | wq_watchdog_thresh = thresh; | |
7577 | wq_watchdog_reset_touched(); | |
7578 | mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ); | |
7579 | } | |
7580 | } | |
7581 | ||
7582 | static int wq_watchdog_param_set_thresh(const char *val, | |
7583 | const struct kernel_param *kp) | |
7584 | { | |
7585 | unsigned long thresh; | |
7586 | int ret; | |
7587 | ||
7588 | ret = kstrtoul(val, 0, &thresh); | |
7589 | if (ret) | |
7590 | return ret; | |
7591 | ||
7592 | if (system_wq) | |
7593 | wq_watchdog_set_thresh(thresh); | |
7594 | else | |
7595 | wq_watchdog_thresh = thresh; | |
7596 | ||
7597 | return 0; | |
7598 | } | |
7599 | ||
7600 | static const struct kernel_param_ops wq_watchdog_thresh_ops = { | |
7601 | .set = wq_watchdog_param_set_thresh, | |
7602 | .get = param_get_ulong, | |
7603 | }; | |
7604 | ||
7605 | module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh, | |
7606 | 0644); | |
7607 | ||
7608 | static void wq_watchdog_init(void) | |
7609 | { | |
5cd79d6a | 7610 | timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE); |
82607adc TH |
7611 | wq_watchdog_set_thresh(wq_watchdog_thresh); |
7612 | } | |
7613 | ||
7614 | #else /* CONFIG_WQ_WATCHDOG */ | |
7615 | ||
7616 | static inline void wq_watchdog_init(void) { } | |
7617 | ||
7618 | #endif /* CONFIG_WQ_WATCHDOG */ | |
7619 | ||
2f34d733 TH |
7620 | static void bh_pool_kick_normal(struct irq_work *irq_work) |
7621 | { | |
7622 | raise_softirq_irqoff(TASKLET_SOFTIRQ); | |
7623 | } | |
7624 | ||
7625 | static void bh_pool_kick_highpri(struct irq_work *irq_work) | |
7626 | { | |
7627 | raise_softirq_irqoff(HI_SOFTIRQ); | |
7628 | } | |
7629 | ||
4a6c5607 TH |
7630 | static void __init restrict_unbound_cpumask(const char *name, const struct cpumask *mask) |
7631 | { | |
7632 | if (!cpumask_intersects(wq_unbound_cpumask, mask)) { | |
7633 | pr_warn("workqueue: Restricting unbound_cpumask (%*pb) with %s (%*pb) leaves no CPU, ignoring\n", | |
7634 | cpumask_pr_args(wq_unbound_cpumask), name, cpumask_pr_args(mask)); | |
7635 | return; | |
7636 | } | |
7637 | ||
7638 | cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, mask); | |
7639 | } | |
7640 | ||
2fcdb1b4 TH |
7641 | static void __init init_cpu_worker_pool(struct worker_pool *pool, int cpu, int nice) |
7642 | { | |
7643 | BUG_ON(init_worker_pool(pool)); | |
7644 | pool->cpu = cpu; | |
7645 | cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); | |
7646 | cpumask_copy(pool->attrs->__pod_cpumask, cpumask_of(cpu)); | |
7647 | pool->attrs->nice = nice; | |
7648 | pool->attrs->affn_strict = true; | |
7649 | pool->node = cpu_to_node(cpu); | |
7650 | ||
7651 | /* alloc pool ID */ | |
7652 | mutex_lock(&wq_pool_mutex); | |
7653 | BUG_ON(worker_pool_assign_id(pool)); | |
7654 | mutex_unlock(&wq_pool_mutex); | |
7655 | } | |
7656 | ||
3347fa09 TH |
7657 | /** |
7658 | * workqueue_init_early - early init for workqueue subsystem | |
7659 | * | |
2930155b TH |
7660 | * This is the first step of three-staged workqueue subsystem initialization and |
7661 | * invoked as soon as the bare basics - memory allocation, cpumasks and idr are | |
7662 | * up. It sets up all the data structures and system workqueues and allows early | |
7663 | * boot code to create workqueues and queue/cancel work items. Actual work item | |
7664 | * execution starts only after kthreads can be created and scheduled right | |
7665 | * before early initcalls. | |
3347fa09 | 7666 | */ |
2333e829 | 7667 | void __init workqueue_init_early(void) |
1da177e4 | 7668 | { |
84193c07 | 7669 | struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_SYSTEM]; |
7a4e344c | 7670 | int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; |
2f34d733 TH |
7671 | void (*irq_work_fns[2])(struct irq_work *) = { bh_pool_kick_normal, |
7672 | bh_pool_kick_highpri }; | |
7a4e344c | 7673 | int i, cpu; |
c34056a3 | 7674 | |
10cdb157 | 7675 | BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); |
e904e6c2 | 7676 | |
b05a7928 | 7677 | BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL)); |
fe28f631 WL |
7678 | BUG_ON(!alloc_cpumask_var(&wq_requested_unbound_cpumask, GFP_KERNEL)); |
7679 | BUG_ON(!zalloc_cpumask_var(&wq_isolated_cpumask, GFP_KERNEL)); | |
b05a7928 | 7680 | |
4a6c5607 TH |
7681 | cpumask_copy(wq_unbound_cpumask, cpu_possible_mask); |
7682 | restrict_unbound_cpumask("HK_TYPE_WQ", housekeeping_cpumask(HK_TYPE_WQ)); | |
7683 | restrict_unbound_cpumask("HK_TYPE_DOMAIN", housekeeping_cpumask(HK_TYPE_DOMAIN)); | |
ace3c549 | 7684 | if (!cpumask_empty(&wq_cmdline_cpumask)) |
4a6c5607 | 7685 | restrict_unbound_cpumask("workqueue.unbound_cpus", &wq_cmdline_cpumask); |
ace3c549 | 7686 | |
fe28f631 | 7687 | cpumask_copy(wq_requested_unbound_cpumask, wq_unbound_cpumask); |
ace3c549 | 7688 | |
e904e6c2 TH |
7689 | pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); |
7690 | ||
2930155b TH |
7691 | wq_update_pod_attrs_buf = alloc_workqueue_attrs(); |
7692 | BUG_ON(!wq_update_pod_attrs_buf); | |
7693 | ||
7bd20b6b MT |
7694 | /* |
7695 | * If nohz_full is enabled, set power efficient workqueue as unbound. | |
7696 | * This allows workqueue items to be moved to HK CPUs. | |
7697 | */ | |
7698 | if (housekeeping_enabled(HK_TYPE_TICK)) | |
7699 | wq_power_efficient = true; | |
7700 | ||
84193c07 TH |
7701 | /* initialize WQ_AFFN_SYSTEM pods */ |
7702 | pt->pod_cpus = kcalloc(1, sizeof(pt->pod_cpus[0]), GFP_KERNEL); | |
7703 | pt->pod_node = kcalloc(1, sizeof(pt->pod_node[0]), GFP_KERNEL); | |
7704 | pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL); | |
7705 | BUG_ON(!pt->pod_cpus || !pt->pod_node || !pt->cpu_pod); | |
7706 | ||
7707 | BUG_ON(!zalloc_cpumask_var_node(&pt->pod_cpus[0], GFP_KERNEL, NUMA_NO_NODE)); | |
7708 | ||
84193c07 TH |
7709 | pt->nr_pods = 1; |
7710 | cpumask_copy(pt->pod_cpus[0], cpu_possible_mask); | |
7711 | pt->pod_node[0] = NUMA_NO_NODE; | |
7712 | pt->cpu_pod[0] = 0; | |
7713 | ||
4cb1ef64 | 7714 | /* initialize BH and CPU pools */ |
29c91e99 | 7715 | for_each_possible_cpu(cpu) { |
4ce62e9e | 7716 | struct worker_pool *pool; |
8b03ae3c | 7717 | |
4cb1ef64 TH |
7718 | i = 0; |
7719 | for_each_bh_worker_pool(pool, cpu) { | |
2f34d733 | 7720 | init_cpu_worker_pool(pool, cpu, std_nice[i]); |
4cb1ef64 | 7721 | pool->flags |= POOL_BH; |
2f34d733 TH |
7722 | init_irq_work(bh_pool_irq_work(pool), irq_work_fns[i]); |
7723 | i++; | |
4cb1ef64 TH |
7724 | } |
7725 | ||
7a4e344c | 7726 | i = 0; |
2fcdb1b4 TH |
7727 | for_each_cpu_worker_pool(pool, cpu) |
7728 | init_cpu_worker_pool(pool, cpu, std_nice[i++]); | |
8b03ae3c TH |
7729 | } |
7730 | ||
8a2b7538 | 7731 | /* create default unbound and ordered wq attrs */ |
29c91e99 TH |
7732 | for (i = 0; i < NR_STD_WORKER_POOLS; i++) { |
7733 | struct workqueue_attrs *attrs; | |
7734 | ||
be69d00d | 7735 | BUG_ON(!(attrs = alloc_workqueue_attrs())); |
29c91e99 | 7736 | attrs->nice = std_nice[i]; |
29c91e99 | 7737 | unbound_std_wq_attrs[i] = attrs; |
8a2b7538 TH |
7738 | |
7739 | /* | |
7740 | * An ordered wq should have only one pwq as ordering is | |
7741 | * guaranteed by max_active which is enforced by pwqs. | |
8a2b7538 | 7742 | */ |
be69d00d | 7743 | BUG_ON(!(attrs = alloc_workqueue_attrs())); |
8a2b7538 | 7744 | attrs->nice = std_nice[i]; |
af73f5c9 | 7745 | attrs->ordered = true; |
8a2b7538 | 7746 | ordered_wq_attrs[i] = attrs; |
29c91e99 TH |
7747 | } |
7748 | ||
d320c038 | 7749 | system_wq = alloc_workqueue("events", 0, 0); |
1aabe902 | 7750 | system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); |
d320c038 | 7751 | system_long_wq = alloc_workqueue("events_long", 0, 0); |
f3421797 | 7752 | system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, |
636b927e | 7753 | WQ_MAX_ACTIVE); |
24d51add TH |
7754 | system_freezable_wq = alloc_workqueue("events_freezable", |
7755 | WQ_FREEZABLE, 0); | |
0668106c VK |
7756 | system_power_efficient_wq = alloc_workqueue("events_power_efficient", |
7757 | WQ_POWER_EFFICIENT, 0); | |
8318d6a6 | 7758 | system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_pwr_efficient", |
0668106c VK |
7759 | WQ_FREEZABLE | WQ_POWER_EFFICIENT, |
7760 | 0); | |
4cb1ef64 TH |
7761 | system_bh_wq = alloc_workqueue("events_bh", WQ_BH, 0); |
7762 | system_bh_highpri_wq = alloc_workqueue("events_bh_highpri", | |
7763 | WQ_BH | WQ_HIGHPRI, 0); | |
1aabe902 | 7764 | BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || |
0668106c VK |
7765 | !system_unbound_wq || !system_freezable_wq || |
7766 | !system_power_efficient_wq || | |
4cb1ef64 TH |
7767 | !system_freezable_power_efficient_wq || |
7768 | !system_bh_wq || !system_bh_highpri_wq); | |
3347fa09 TH |
7769 | } |
7770 | ||
aa6fde93 TH |
7771 | static void __init wq_cpu_intensive_thresh_init(void) |
7772 | { | |
7773 | unsigned long thresh; | |
7774 | unsigned long bogo; | |
7775 | ||
dd64c873 Z |
7776 | pwq_release_worker = kthread_create_worker(0, "pool_workqueue_release"); |
7777 | BUG_ON(IS_ERR(pwq_release_worker)); | |
7778 | ||
aa6fde93 TH |
7779 | /* if the user set it to a specific value, keep it */ |
7780 | if (wq_cpu_intensive_thresh_us != ULONG_MAX) | |
7781 | return; | |
7782 | ||
7783 | /* | |
7784 | * The default of 10ms is derived from the fact that most modern (as of | |
7785 | * 2023) processors can do a lot in 10ms and that it's just below what | |
7786 | * most consider human-perceivable. However, the kernel also runs on a | |
7787 | * lot slower CPUs including microcontrollers where the threshold is way | |
7788 | * too low. | |
7789 | * | |
7790 | * Let's scale up the threshold upto 1 second if BogoMips is below 4000. | |
7791 | * This is by no means accurate but it doesn't have to be. The mechanism | |
7792 | * is still useful even when the threshold is fully scaled up. Also, as | |
7793 | * the reports would usually be applicable to everyone, some machines | |
7794 | * operating on longer thresholds won't significantly diminish their | |
7795 | * usefulness. | |
7796 | */ | |
7797 | thresh = 10 * USEC_PER_MSEC; | |
7798 | ||
7799 | /* see init/calibrate.c for lpj -> BogoMIPS calculation */ | |
7800 | bogo = max_t(unsigned long, loops_per_jiffy / 500000 * HZ, 1); | |
7801 | if (bogo < 4000) | |
7802 | thresh = min_t(unsigned long, thresh * 4000 / bogo, USEC_PER_SEC); | |
7803 | ||
7804 | pr_debug("wq_cpu_intensive_thresh: lpj=%lu BogoMIPS=%lu thresh_us=%lu\n", | |
7805 | loops_per_jiffy, bogo, thresh); | |
7806 | ||
7807 | wq_cpu_intensive_thresh_us = thresh; | |
7808 | } | |
7809 | ||
3347fa09 TH |
7810 | /** |
7811 | * workqueue_init - bring workqueue subsystem fully online | |
7812 | * | |
2930155b TH |
7813 | * This is the second step of three-staged workqueue subsystem initialization |
7814 | * and invoked as soon as kthreads can be created and scheduled. Workqueues have | |
7815 | * been created and work items queued on them, but there are no kworkers | |
7816 | * executing the work items yet. Populate the worker pools with the initial | |
7817 | * workers and enable future kworker creations. | |
3347fa09 | 7818 | */ |
2333e829 | 7819 | void __init workqueue_init(void) |
3347fa09 | 7820 | { |
2186d9f9 | 7821 | struct workqueue_struct *wq; |
3347fa09 TH |
7822 | struct worker_pool *pool; |
7823 | int cpu, bkt; | |
7824 | ||
aa6fde93 TH |
7825 | wq_cpu_intensive_thresh_init(); |
7826 | ||
2186d9f9 TH |
7827 | mutex_lock(&wq_pool_mutex); |
7828 | ||
2930155b TH |
7829 | /* |
7830 | * Per-cpu pools created earlier could be missing node hint. Fix them | |
7831 | * up. Also, create a rescuer for workqueues that requested it. | |
7832 | */ | |
2186d9f9 | 7833 | for_each_possible_cpu(cpu) { |
4cb1ef64 TH |
7834 | for_each_bh_worker_pool(pool, cpu) |
7835 | pool->node = cpu_to_node(cpu); | |
7836 | for_each_cpu_worker_pool(pool, cpu) | |
2186d9f9 | 7837 | pool->node = cpu_to_node(cpu); |
2186d9f9 TH |
7838 | } |
7839 | ||
40c17f75 | 7840 | list_for_each_entry(wq, &workqueues, list) { |
40c17f75 TH |
7841 | WARN(init_rescuer(wq), |
7842 | "workqueue: failed to create early rescuer for %s", | |
7843 | wq->name); | |
7844 | } | |
2186d9f9 TH |
7845 | |
7846 | mutex_unlock(&wq_pool_mutex); | |
7847 | ||
4cb1ef64 TH |
7848 | /* |
7849 | * Create the initial workers. A BH pool has one pseudo worker that | |
7850 | * represents the shared BH execution context and thus doesn't get | |
7851 | * affected by hotplug events. Create the BH pseudo workers for all | |
7852 | * possible CPUs here. | |
7853 | */ | |
7854 | for_each_possible_cpu(cpu) | |
7855 | for_each_bh_worker_pool(pool, cpu) | |
7856 | BUG_ON(!create_worker(pool)); | |
7857 | ||
3347fa09 TH |
7858 | for_each_online_cpu(cpu) { |
7859 | for_each_cpu_worker_pool(pool, cpu) { | |
7860 | pool->flags &= ~POOL_DISASSOCIATED; | |
7861 | BUG_ON(!create_worker(pool)); | |
7862 | } | |
7863 | } | |
7864 | ||
7865 | hash_for_each(unbound_pool_hash, bkt, pool, hash_node) | |
7866 | BUG_ON(!create_worker(pool)); | |
7867 | ||
7868 | wq_online = true; | |
82607adc | 7869 | wq_watchdog_init(); |
1da177e4 | 7870 | } |
c4f135d6 | 7871 | |
025e1684 TH |
7872 | /* |
7873 | * Initialize @pt by first initializing @pt->cpu_pod[] with pod IDs according to | |
7874 | * @cpu_shares_pod(). Each subset of CPUs that share a pod is assigned a unique | |
7875 | * and consecutive pod ID. The rest of @pt is initialized accordingly. | |
7876 | */ | |
7877 | static void __init init_pod_type(struct wq_pod_type *pt, | |
7878 | bool (*cpus_share_pod)(int, int)) | |
7879 | { | |
7880 | int cur, pre, cpu, pod; | |
7881 | ||
7882 | pt->nr_pods = 0; | |
7883 | ||
7884 | /* init @pt->cpu_pod[] according to @cpus_share_pod() */ | |
7885 | pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL); | |
7886 | BUG_ON(!pt->cpu_pod); | |
7887 | ||
7888 | for_each_possible_cpu(cur) { | |
7889 | for_each_possible_cpu(pre) { | |
7890 | if (pre >= cur) { | |
7891 | pt->cpu_pod[cur] = pt->nr_pods++; | |
7892 | break; | |
7893 | } | |
7894 | if (cpus_share_pod(cur, pre)) { | |
7895 | pt->cpu_pod[cur] = pt->cpu_pod[pre]; | |
7896 | break; | |
7897 | } | |
7898 | } | |
7899 | } | |
7900 | ||
7901 | /* init the rest to match @pt->cpu_pod[] */ | |
7902 | pt->pod_cpus = kcalloc(pt->nr_pods, sizeof(pt->pod_cpus[0]), GFP_KERNEL); | |
7903 | pt->pod_node = kcalloc(pt->nr_pods, sizeof(pt->pod_node[0]), GFP_KERNEL); | |
7904 | BUG_ON(!pt->pod_cpus || !pt->pod_node); | |
7905 | ||
7906 | for (pod = 0; pod < pt->nr_pods; pod++) | |
7907 | BUG_ON(!zalloc_cpumask_var(&pt->pod_cpus[pod], GFP_KERNEL)); | |
7908 | ||
7909 | for_each_possible_cpu(cpu) { | |
7910 | cpumask_set_cpu(cpu, pt->pod_cpus[pt->cpu_pod[cpu]]); | |
7911 | pt->pod_node[pt->cpu_pod[cpu]] = cpu_to_node(cpu); | |
7912 | } | |
7913 | } | |
7914 | ||
63c5484e TH |
7915 | static bool __init cpus_dont_share(int cpu0, int cpu1) |
7916 | { | |
7917 | return false; | |
7918 | } | |
7919 | ||
7920 | static bool __init cpus_share_smt(int cpu0, int cpu1) | |
7921 | { | |
7922 | #ifdef CONFIG_SCHED_SMT | |
7923 | return cpumask_test_cpu(cpu0, cpu_smt_mask(cpu1)); | |
7924 | #else | |
7925 | return false; | |
7926 | #endif | |
7927 | } | |
7928 | ||
025e1684 TH |
7929 | static bool __init cpus_share_numa(int cpu0, int cpu1) |
7930 | { | |
7931 | return cpu_to_node(cpu0) == cpu_to_node(cpu1); | |
7932 | } | |
7933 | ||
2930155b TH |
7934 | /** |
7935 | * workqueue_init_topology - initialize CPU pods for unbound workqueues | |
7936 | * | |
96068b60 | 7937 | * This is the third step of three-staged workqueue subsystem initialization and |
2930155b TH |
7938 | * invoked after SMP and topology information are fully initialized. It |
7939 | * initializes the unbound CPU pods accordingly. | |
7940 | */ | |
7941 | void __init workqueue_init_topology(void) | |
a86feae6 | 7942 | { |
2930155b | 7943 | struct workqueue_struct *wq; |
025e1684 | 7944 | int cpu; |
a86feae6 | 7945 | |
63c5484e TH |
7946 | init_pod_type(&wq_pod_types[WQ_AFFN_CPU], cpus_dont_share); |
7947 | init_pod_type(&wq_pod_types[WQ_AFFN_SMT], cpus_share_smt); | |
7948 | init_pod_type(&wq_pod_types[WQ_AFFN_CACHE], cpus_share_cache); | |
025e1684 | 7949 | init_pod_type(&wq_pod_types[WQ_AFFN_NUMA], cpus_share_numa); |
a86feae6 | 7950 | |
c5f8cd6c TH |
7951 | wq_topo_initialized = true; |
7952 | ||
2930155b | 7953 | mutex_lock(&wq_pool_mutex); |
a86feae6 | 7954 | |
2930155b TH |
7955 | /* |
7956 | * Workqueues allocated earlier would have all CPUs sharing the default | |
7957 | * worker pool. Explicitly call wq_update_pod() on all workqueue and CPU | |
7958 | * combinations to apply per-pod sharing. | |
7959 | */ | |
7960 | list_for_each_entry(wq, &workqueues, list) { | |
5797b1c1 | 7961 | for_each_online_cpu(cpu) |
2930155b | 7962 | wq_update_pod(wq, cpu, cpu, true); |
5797b1c1 TH |
7963 | if (wq->flags & WQ_UNBOUND) { |
7964 | mutex_lock(&wq->mutex); | |
7965 | wq_update_node_max_active(wq, -1); | |
7966 | mutex_unlock(&wq->mutex); | |
2930155b TH |
7967 | } |
7968 | } | |
7969 | ||
7970 | mutex_unlock(&wq_pool_mutex); | |
a86feae6 TH |
7971 | } |
7972 | ||
20bdedaf TH |
7973 | void __warn_flushing_systemwide_wq(void) |
7974 | { | |
7975 | pr_warn("WARNING: Flushing system-wide workqueues will be prohibited in near future.\n"); | |
7976 | dump_stack(); | |
7977 | } | |
c4f135d6 | 7978 | EXPORT_SYMBOL(__warn_flushing_systemwide_wq); |
ace3c549 | 7979 | |
7980 | static int __init workqueue_unbound_cpus_setup(char *str) | |
7981 | { | |
7982 | if (cpulist_parse(str, &wq_cmdline_cpumask) < 0) { | |
7983 | cpumask_clear(&wq_cmdline_cpumask); | |
7984 | pr_warn("workqueue.unbound_cpus: incorrect CPU range, using default\n"); | |
7985 | } | |
7986 | ||
7987 | return 1; | |
7988 | } | |
7989 | __setup("workqueue.unbound_cpus=", workqueue_unbound_cpus_setup); |