Commit | Line | Data |
---|---|---|
457c8996 | 1 | // SPDX-License-Identifier: GPL-2.0-only |
1da177e4 | 2 | /* |
c54fce6e | 3 | * kernel/workqueue.c - generic async execution with shared worker pool |
1da177e4 | 4 | * |
c54fce6e | 5 | * Copyright (C) 2002 Ingo Molnar |
1da177e4 | 6 | * |
c54fce6e TH |
7 | * Derived from the taskqueue/keventd code by: |
8 | * David Woodhouse <dwmw2@infradead.org> | |
9 | * Andrew Morton | |
10 | * Kai Petzke <wpp@marie.physik.tu-berlin.de> | |
11 | * Theodore Ts'o <tytso@mit.edu> | |
1da177e4 | 12 | * |
c54fce6e | 13 | * Made to use alloc_percpu by Christoph Lameter. |
1da177e4 | 14 | * |
c54fce6e TH |
15 | * Copyright (C) 2010 SUSE Linux Products GmbH |
16 | * Copyright (C) 2010 Tejun Heo <tj@kernel.org> | |
89ada679 | 17 | * |
c54fce6e TH |
18 | * This is the generic async execution mechanism. Work items as are |
19 | * executed in process context. The worker pool is shared and | |
b11895c4 L |
20 | * automatically managed. There are two worker pools for each CPU (one for |
21 | * normal work items and the other for high priority ones) and some extra | |
22 | * pools for workqueues which are not bound to any specific CPU - the | |
23 | * number of these backing pools is dynamic. | |
c54fce6e | 24 | * |
9a261491 | 25 | * Please read Documentation/core-api/workqueue.rst for details. |
1da177e4 LT |
26 | */ |
27 | ||
9984de1a | 28 | #include <linux/export.h> |
1da177e4 LT |
29 | #include <linux/kernel.h> |
30 | #include <linux/sched.h> | |
31 | #include <linux/init.h> | |
32 | #include <linux/signal.h> | |
33 | #include <linux/completion.h> | |
34 | #include <linux/workqueue.h> | |
35 | #include <linux/slab.h> | |
36 | #include <linux/cpu.h> | |
37 | #include <linux/notifier.h> | |
38 | #include <linux/kthread.h> | |
1fa44eca | 39 | #include <linux/hardirq.h> |
46934023 | 40 | #include <linux/mempolicy.h> |
341a5958 | 41 | #include <linux/freezer.h> |
d5abe669 | 42 | #include <linux/debug_locks.h> |
4e6045f1 | 43 | #include <linux/lockdep.h> |
c34056a3 | 44 | #include <linux/idr.h> |
29c91e99 | 45 | #include <linux/jhash.h> |
42f8570f | 46 | #include <linux/hashtable.h> |
76af4d93 | 47 | #include <linux/rculist.h> |
bce90380 | 48 | #include <linux/nodemask.h> |
4c16bd32 | 49 | #include <linux/moduleparam.h> |
3d1cb205 | 50 | #include <linux/uaccess.h> |
c98a9805 | 51 | #include <linux/sched/isolation.h> |
cd2440d6 | 52 | #include <linux/sched/debug.h> |
62635ea8 | 53 | #include <linux/nmi.h> |
940d71c6 | 54 | #include <linux/kvm_para.h> |
e22bee78 | 55 | |
ea138446 | 56 | #include "workqueue_internal.h" |
1da177e4 | 57 | |
c8e55f36 | 58 | enum { |
24647570 TH |
59 | /* |
60 | * worker_pool flags | |
bc2ae0f5 | 61 | * |
24647570 | 62 | * A bound pool is either associated or disassociated with its CPU. |
bc2ae0f5 TH |
63 | * While associated (!DISASSOCIATED), all workers are bound to the |
64 | * CPU and none has %WORKER_UNBOUND set and concurrency management | |
65 | * is in effect. | |
66 | * | |
67 | * While DISASSOCIATED, the cpu may be offline and all workers have | |
68 | * %WORKER_UNBOUND set and concurrency management disabled, and may | |
24647570 | 69 | * be executing on any CPU. The pool behaves as an unbound one. |
bc2ae0f5 | 70 | * |
bc3a1afc | 71 | * Note that DISASSOCIATED should be flipped only while holding |
1258fae7 | 72 | * wq_pool_attach_mutex to avoid changing binding state while |
4736cbf7 | 73 | * worker_attach_to_pool() is in progress. |
bc2ae0f5 | 74 | */ |
692b4825 | 75 | POOL_MANAGER_ACTIVE = 1 << 0, /* being managed */ |
24647570 | 76 | POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */ |
db7bccf4 | 77 | |
c8e55f36 | 78 | /* worker flags */ |
c8e55f36 TH |
79 | WORKER_DIE = 1 << 1, /* die die die */ |
80 | WORKER_IDLE = 1 << 2, /* is idle */ | |
e22bee78 | 81 | WORKER_PREP = 1 << 3, /* preparing to run works */ |
fb0e7beb | 82 | WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */ |
f3421797 | 83 | WORKER_UNBOUND = 1 << 7, /* worker is unbound */ |
a9ab775b | 84 | WORKER_REBOUND = 1 << 8, /* worker was rebound */ |
e22bee78 | 85 | |
a9ab775b TH |
86 | WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE | |
87 | WORKER_UNBOUND | WORKER_REBOUND, | |
db7bccf4 | 88 | |
e34cdddb | 89 | NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */ |
4ce62e9e | 90 | |
29c91e99 | 91 | UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */ |
c8e55f36 | 92 | BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */ |
db7bccf4 | 93 | |
e22bee78 TH |
94 | MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */ |
95 | IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */ | |
96 | ||
3233cdbd TH |
97 | MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2, |
98 | /* call for help after 10ms | |
99 | (min two ticks) */ | |
e22bee78 TH |
100 | MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */ |
101 | CREATE_COOLDOWN = HZ, /* time to breath after fail */ | |
e22bee78 TH |
102 | |
103 | /* | |
104 | * Rescue workers are used only on emergencies and shared by | |
8698a745 | 105 | * all cpus. Give MIN_NICE. |
e22bee78 | 106 | */ |
8698a745 DY |
107 | RESCUER_NICE_LEVEL = MIN_NICE, |
108 | HIGHPRI_NICE_LEVEL = MIN_NICE, | |
ecf6881f TH |
109 | |
110 | WQ_NAME_LEN = 24, | |
c8e55f36 | 111 | }; |
1da177e4 LT |
112 | |
113 | /* | |
4690c4ab TH |
114 | * Structure fields follow one of the following exclusion rules. |
115 | * | |
e41e704b TH |
116 | * I: Modifiable by initialization/destruction paths and read-only for |
117 | * everyone else. | |
4690c4ab | 118 | * |
e22bee78 TH |
119 | * P: Preemption protected. Disabling preemption is enough and should |
120 | * only be modified and accessed from the local cpu. | |
121 | * | |
d565ed63 | 122 | * L: pool->lock protected. Access with pool->lock held. |
4690c4ab | 123 | * |
d565ed63 TH |
124 | * X: During normal operation, modification requires pool->lock and should |
125 | * be done only from local cpu. Either disabling preemption on local | |
126 | * cpu or grabbing pool->lock is enough for read access. If | |
127 | * POOL_DISASSOCIATED is set, it's identical to L. | |
e22bee78 | 128 | * |
bdf8b9bf TH |
129 | * K: Only modified by worker while holding pool->lock. Can be safely read by |
130 | * self, while holding pool->lock or from IRQ context if %current is the | |
131 | * kworker. | |
132 | * | |
133 | * S: Only modified by worker self. | |
134 | * | |
1258fae7 | 135 | * A: wq_pool_attach_mutex protected. |
822d8405 | 136 | * |
68e13a67 | 137 | * PL: wq_pool_mutex protected. |
5bcab335 | 138 | * |
24acfb71 | 139 | * PR: wq_pool_mutex protected for writes. RCU protected for reads. |
76af4d93 | 140 | * |
5b95e1af LJ |
141 | * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads. |
142 | * | |
143 | * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or | |
24acfb71 | 144 | * RCU for reads. |
5b95e1af | 145 | * |
3c25a55d LJ |
146 | * WQ: wq->mutex protected. |
147 | * | |
24acfb71 | 148 | * WR: wq->mutex protected for writes. RCU protected for reads. |
2e109a28 TH |
149 | * |
150 | * MD: wq_mayday_lock protected. | |
cd2440d6 PM |
151 | * |
152 | * WD: Used internally by the watchdog. | |
1da177e4 | 153 | */ |
1da177e4 | 154 | |
2eaebdb3 | 155 | /* struct worker is defined in workqueue_internal.h */ |
c34056a3 | 156 | |
bd7bdd43 | 157 | struct worker_pool { |
a9b8a985 | 158 | raw_spinlock_t lock; /* the pool lock */ |
d84ff051 | 159 | int cpu; /* I: the associated cpu */ |
f3f90ad4 | 160 | int node; /* I: the associated node ID */ |
9daf9e67 | 161 | int id; /* I: pool ID */ |
11ebea50 | 162 | unsigned int flags; /* X: flags */ |
bd7bdd43 | 163 | |
82607adc | 164 | unsigned long watchdog_ts; /* L: watchdog timestamp */ |
cd2440d6 | 165 | bool cpu_stall; /* WD: stalled cpu bound pool */ |
82607adc | 166 | |
bc35f7ef LJ |
167 | /* |
168 | * The counter is incremented in a process context on the associated CPU | |
169 | * w/ preemption disabled, and decremented or reset in the same context | |
170 | * but w/ pool->lock held. The readers grab pool->lock and are | |
171 | * guaranteed to see if the counter reached zero. | |
172 | */ | |
173 | int nr_running; | |
84f91c62 | 174 | |
bd7bdd43 | 175 | struct list_head worklist; /* L: list of pending works */ |
ea1abd61 | 176 | |
5826cc8f LJ |
177 | int nr_workers; /* L: total number of workers */ |
178 | int nr_idle; /* L: currently idle workers */ | |
bd7bdd43 | 179 | |
2c1f1a91 | 180 | struct list_head idle_list; /* L: list of idle workers */ |
bd7bdd43 | 181 | struct timer_list idle_timer; /* L: worker idle timeout */ |
3f959aa3 VS |
182 | struct work_struct idle_cull_work; /* L: worker idle cleanup */ |
183 | ||
184 | struct timer_list mayday_timer; /* L: SOS timer for workers */ | |
bd7bdd43 | 185 | |
c5aa87bb | 186 | /* a workers is either on busy_hash or idle_list, or the manager */ |
c9e7cf27 TH |
187 | DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER); |
188 | /* L: hash of busy workers */ | |
189 | ||
2607d7a6 | 190 | struct worker *manager; /* L: purely informational */ |
92f9c5c4 | 191 | struct list_head workers; /* A: attached workers */ |
e02b9312 | 192 | struct list_head dying_workers; /* A: workers about to die */ |
60f5a4bc | 193 | struct completion *detach_completion; /* all workers detached */ |
e19e397a | 194 | |
7cda9aae | 195 | struct ida worker_ida; /* worker IDs for task name */ |
e19e397a | 196 | |
7a4e344c | 197 | struct workqueue_attrs *attrs; /* I: worker attributes */ |
68e13a67 LJ |
198 | struct hlist_node hash_node; /* PL: unbound_pool_hash node */ |
199 | int refcnt; /* PL: refcnt for unbound pools */ | |
7a4e344c | 200 | |
29c91e99 | 201 | /* |
24acfb71 | 202 | * Destruction of pool is RCU protected to allow dereferences |
29c91e99 TH |
203 | * from get_work_pool(). |
204 | */ | |
205 | struct rcu_head rcu; | |
84f91c62 | 206 | }; |
8b03ae3c | 207 | |
725e8ec5 TH |
208 | /* |
209 | * Per-pool_workqueue statistics. These can be monitored using | |
210 | * tools/workqueue/wq_monitor.py. | |
211 | */ | |
212 | enum pool_workqueue_stats { | |
213 | PWQ_STAT_STARTED, /* work items started execution */ | |
214 | PWQ_STAT_COMPLETED, /* work items completed execution */ | |
8a1dd1e5 | 215 | PWQ_STAT_CPU_TIME, /* total CPU time consumed */ |
616db877 | 216 | PWQ_STAT_CPU_INTENSIVE, /* wq_cpu_intensive_thresh_us violations */ |
725e8ec5 TH |
217 | PWQ_STAT_CM_WAKEUP, /* concurrency-management worker wakeups */ |
218 | PWQ_STAT_MAYDAY, /* maydays to rescuer */ | |
219 | PWQ_STAT_RESCUED, /* linked work items executed by rescuer */ | |
220 | ||
221 | PWQ_NR_STATS, | |
222 | }; | |
223 | ||
1da177e4 | 224 | /* |
112202d9 TH |
225 | * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS |
226 | * of work_struct->data are used for flags and the remaining high bits | |
227 | * point to the pwq; thus, pwqs need to be aligned at two's power of the | |
228 | * number of flag bits. | |
1da177e4 | 229 | */ |
112202d9 | 230 | struct pool_workqueue { |
bd7bdd43 | 231 | struct worker_pool *pool; /* I: the associated pool */ |
4690c4ab | 232 | struct workqueue_struct *wq; /* I: the owning workqueue */ |
73f53c4a TH |
233 | int work_color; /* L: current color */ |
234 | int flush_color; /* L: flushing color */ | |
8864b4e5 | 235 | int refcnt; /* L: reference count */ |
73f53c4a TH |
236 | int nr_in_flight[WORK_NR_COLORS]; |
237 | /* L: nr of in_flight works */ | |
018f3a13 LJ |
238 | |
239 | /* | |
240 | * nr_active management and WORK_STRUCT_INACTIVE: | |
241 | * | |
242 | * When pwq->nr_active >= max_active, new work item is queued to | |
243 | * pwq->inactive_works instead of pool->worklist and marked with | |
244 | * WORK_STRUCT_INACTIVE. | |
245 | * | |
246 | * All work items marked with WORK_STRUCT_INACTIVE do not participate | |
247 | * in pwq->nr_active and all work items in pwq->inactive_works are | |
248 | * marked with WORK_STRUCT_INACTIVE. But not all WORK_STRUCT_INACTIVE | |
249 | * work items are in pwq->inactive_works. Some of them are ready to | |
250 | * run in pool->worklist or worker->scheduled. Those work itmes are | |
251 | * only struct wq_barrier which is used for flush_work() and should | |
252 | * not participate in pwq->nr_active. For non-barrier work item, it | |
253 | * is marked with WORK_STRUCT_INACTIVE iff it is in pwq->inactive_works. | |
254 | */ | |
1e19ffc6 | 255 | int nr_active; /* L: nr of active works */ |
a0a1a5fd | 256 | int max_active; /* L: max active works */ |
f97a4a1a | 257 | struct list_head inactive_works; /* L: inactive works */ |
3c25a55d | 258 | struct list_head pwqs_node; /* WR: node on wq->pwqs */ |
2e109a28 | 259 | struct list_head mayday_node; /* MD: node on wq->maydays */ |
8864b4e5 | 260 | |
725e8ec5 TH |
261 | u64 stats[PWQ_NR_STATS]; |
262 | ||
8864b4e5 TH |
263 | /* |
264 | * Release of unbound pwq is punted to system_wq. See put_pwq() | |
265 | * and pwq_unbound_release_workfn() for details. pool_workqueue | |
24acfb71 | 266 | * itself is also RCU protected so that the first pwq can be |
b09f4fd3 | 267 | * determined without grabbing wq->mutex. |
8864b4e5 TH |
268 | */ |
269 | struct work_struct unbound_release_work; | |
270 | struct rcu_head rcu; | |
e904e6c2 | 271 | } __aligned(1 << WORK_STRUCT_FLAG_BITS); |
1da177e4 | 272 | |
73f53c4a TH |
273 | /* |
274 | * Structure used to wait for workqueue flush. | |
275 | */ | |
276 | struct wq_flusher { | |
3c25a55d LJ |
277 | struct list_head list; /* WQ: list of flushers */ |
278 | int flush_color; /* WQ: flush color waiting for */ | |
73f53c4a TH |
279 | struct completion done; /* flush completion */ |
280 | }; | |
281 | ||
226223ab TH |
282 | struct wq_device; |
283 | ||
1da177e4 | 284 | /* |
c5aa87bb TH |
285 | * The externally visible workqueue. It relays the issued work items to |
286 | * the appropriate worker_pool through its pool_workqueues. | |
1da177e4 LT |
287 | */ |
288 | struct workqueue_struct { | |
3c25a55d | 289 | struct list_head pwqs; /* WR: all pwqs of this wq */ |
e2dca7ad | 290 | struct list_head list; /* PR: list of all workqueues */ |
73f53c4a | 291 | |
3c25a55d LJ |
292 | struct mutex mutex; /* protects this wq */ |
293 | int work_color; /* WQ: current work color */ | |
294 | int flush_color; /* WQ: current flush color */ | |
112202d9 | 295 | atomic_t nr_pwqs_to_flush; /* flush in progress */ |
3c25a55d LJ |
296 | struct wq_flusher *first_flusher; /* WQ: first flusher */ |
297 | struct list_head flusher_queue; /* WQ: flush waiters */ | |
298 | struct list_head flusher_overflow; /* WQ: flush overflow list */ | |
73f53c4a | 299 | |
2e109a28 | 300 | struct list_head maydays; /* MD: pwqs requesting rescue */ |
30ae2fc0 | 301 | struct worker *rescuer; /* MD: rescue worker */ |
e22bee78 | 302 | |
87fc741e | 303 | int nr_drainers; /* WQ: drain in progress */ |
a357fc03 | 304 | int saved_max_active; /* WQ: saved pwq max_active */ |
226223ab | 305 | |
5b95e1af LJ |
306 | struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */ |
307 | struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */ | |
6029a918 | 308 | |
226223ab TH |
309 | #ifdef CONFIG_SYSFS |
310 | struct wq_device *wq_dev; /* I: for sysfs interface */ | |
311 | #endif | |
4e6045f1 | 312 | #ifdef CONFIG_LOCKDEP |
669de8bd BVA |
313 | char *lock_name; |
314 | struct lock_class_key key; | |
4690c4ab | 315 | struct lockdep_map lockdep_map; |
4e6045f1 | 316 | #endif |
ecf6881f | 317 | char name[WQ_NAME_LEN]; /* I: workqueue name */ |
2728fd2f | 318 | |
e2dca7ad | 319 | /* |
24acfb71 TG |
320 | * Destruction of workqueue_struct is RCU protected to allow walking |
321 | * the workqueues list without grabbing wq_pool_mutex. | |
e2dca7ad TH |
322 | * This is used to dump all workqueues from sysrq. |
323 | */ | |
324 | struct rcu_head rcu; | |
325 | ||
2728fd2f TH |
326 | /* hot fields used during command issue, aligned to cacheline */ |
327 | unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */ | |
328 | struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */ | |
5b95e1af | 329 | struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */ |
1da177e4 LT |
330 | }; |
331 | ||
e904e6c2 TH |
332 | static struct kmem_cache *pwq_cache; |
333 | ||
bce90380 TH |
334 | static cpumask_var_t *wq_numa_possible_cpumask; |
335 | /* possible CPUs of each node */ | |
336 | ||
616db877 TH |
337 | /* |
338 | * Per-cpu work items which run for longer than the following threshold are | |
339 | * automatically considered CPU intensive and excluded from concurrency | |
340 | * management to prevent them from noticeably delaying other per-cpu work items. | |
341 | */ | |
342 | static unsigned long wq_cpu_intensive_thresh_us = 10000; | |
343 | module_param_named(cpu_intensive_thresh_us, wq_cpu_intensive_thresh_us, ulong, 0644); | |
344 | ||
d55262c4 TH |
345 | static bool wq_disable_numa; |
346 | module_param_named(disable_numa, wq_disable_numa, bool, 0444); | |
347 | ||
cee22a15 | 348 | /* see the comment above the definition of WQ_POWER_EFFICIENT */ |
552f530c | 349 | static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT); |
cee22a15 VK |
350 | module_param_named(power_efficient, wq_power_efficient, bool, 0444); |
351 | ||
863b710b | 352 | static bool wq_online; /* can kworkers be created yet? */ |
3347fa09 | 353 | |
bce90380 TH |
354 | static bool wq_numa_enabled; /* unbound NUMA affinity enabled */ |
355 | ||
4c16bd32 TH |
356 | /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */ |
357 | static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf; | |
358 | ||
68e13a67 | 359 | static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ |
1258fae7 | 360 | static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */ |
a9b8a985 | 361 | static DEFINE_RAW_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ |
d8bb65ab SAS |
362 | /* wait for manager to go away */ |
363 | static struct rcuwait manager_wait = __RCUWAIT_INITIALIZER(manager_wait); | |
5bcab335 | 364 | |
e2dca7ad | 365 | static LIST_HEAD(workqueues); /* PR: list of all workqueues */ |
68e13a67 | 366 | static bool workqueue_freezing; /* PL: have wqs started freezing? */ |
7d19c5ce | 367 | |
99c621ef | 368 | /* PL&A: allowable cpus for unbound wqs and work items */ |
ef557180 MG |
369 | static cpumask_var_t wq_unbound_cpumask; |
370 | ||
371 | /* CPU where unbound work was last round robin scheduled from this CPU */ | |
372 | static DEFINE_PER_CPU(int, wq_rr_cpu_last); | |
b05a7928 | 373 | |
f303fccb TH |
374 | /* |
375 | * Local execution of unbound work items is no longer guaranteed. The | |
376 | * following always forces round-robin CPU selection on unbound work items | |
377 | * to uncover usages which depend on it. | |
378 | */ | |
379 | #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU | |
380 | static bool wq_debug_force_rr_cpu = true; | |
381 | #else | |
382 | static bool wq_debug_force_rr_cpu = false; | |
383 | #endif | |
384 | module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644); | |
385 | ||
7d19c5ce | 386 | /* the per-cpu worker pools */ |
25528213 | 387 | static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools); |
7d19c5ce | 388 | |
68e13a67 | 389 | static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ |
7d19c5ce | 390 | |
68e13a67 | 391 | /* PL: hash of all unbound pools keyed by pool->attrs */ |
29c91e99 TH |
392 | static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); |
393 | ||
c5aa87bb | 394 | /* I: attributes used when instantiating standard unbound pools on demand */ |
29c91e99 TH |
395 | static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; |
396 | ||
8a2b7538 TH |
397 | /* I: attributes used when instantiating ordered pools on demand */ |
398 | static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS]; | |
399 | ||
d320c038 | 400 | struct workqueue_struct *system_wq __read_mostly; |
ad7b1f84 | 401 | EXPORT_SYMBOL(system_wq); |
044c782c | 402 | struct workqueue_struct *system_highpri_wq __read_mostly; |
1aabe902 | 403 | EXPORT_SYMBOL_GPL(system_highpri_wq); |
044c782c | 404 | struct workqueue_struct *system_long_wq __read_mostly; |
d320c038 | 405 | EXPORT_SYMBOL_GPL(system_long_wq); |
044c782c | 406 | struct workqueue_struct *system_unbound_wq __read_mostly; |
f3421797 | 407 | EXPORT_SYMBOL_GPL(system_unbound_wq); |
044c782c | 408 | struct workqueue_struct *system_freezable_wq __read_mostly; |
24d51add | 409 | EXPORT_SYMBOL_GPL(system_freezable_wq); |
0668106c VK |
410 | struct workqueue_struct *system_power_efficient_wq __read_mostly; |
411 | EXPORT_SYMBOL_GPL(system_power_efficient_wq); | |
412 | struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly; | |
413 | EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq); | |
d320c038 | 414 | |
7d19c5ce | 415 | static int worker_thread(void *__worker); |
6ba94429 | 416 | static void workqueue_sysfs_unregister(struct workqueue_struct *wq); |
c29eb853 | 417 | static void show_pwq(struct pool_workqueue *pwq); |
55df0933 | 418 | static void show_one_worker_pool(struct worker_pool *pool); |
7d19c5ce | 419 | |
97bd2347 TH |
420 | #define CREATE_TRACE_POINTS |
421 | #include <trace/events/workqueue.h> | |
422 | ||
68e13a67 | 423 | #define assert_rcu_or_pool_mutex() \ |
24acfb71 | 424 | RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \ |
f78f5b90 | 425 | !lockdep_is_held(&wq_pool_mutex), \ |
24acfb71 | 426 | "RCU or wq_pool_mutex should be held") |
5bcab335 | 427 | |
5b95e1af | 428 | #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \ |
24acfb71 | 429 | RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \ |
f78f5b90 PM |
430 | !lockdep_is_held(&wq->mutex) && \ |
431 | !lockdep_is_held(&wq_pool_mutex), \ | |
24acfb71 | 432 | "RCU, wq->mutex or wq_pool_mutex should be held") |
5b95e1af | 433 | |
f02ae73a TH |
434 | #define for_each_cpu_worker_pool(pool, cpu) \ |
435 | for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ | |
436 | (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ | |
7a62c2c8 | 437 | (pool)++) |
4ce62e9e | 438 | |
17116969 TH |
439 | /** |
440 | * for_each_pool - iterate through all worker_pools in the system | |
441 | * @pool: iteration cursor | |
611c92a0 | 442 | * @pi: integer used for iteration |
fa1b54e6 | 443 | * |
24acfb71 | 444 | * This must be called either with wq_pool_mutex held or RCU read |
68e13a67 LJ |
445 | * locked. If the pool needs to be used beyond the locking in effect, the |
446 | * caller is responsible for guaranteeing that the pool stays online. | |
fa1b54e6 TH |
447 | * |
448 | * The if/else clause exists only for the lockdep assertion and can be | |
449 | * ignored. | |
17116969 | 450 | */ |
611c92a0 TH |
451 | #define for_each_pool(pool, pi) \ |
452 | idr_for_each_entry(&worker_pool_idr, pool, pi) \ | |
68e13a67 | 453 | if (({ assert_rcu_or_pool_mutex(); false; })) { } \ |
fa1b54e6 | 454 | else |
17116969 | 455 | |
822d8405 TH |
456 | /** |
457 | * for_each_pool_worker - iterate through all workers of a worker_pool | |
458 | * @worker: iteration cursor | |
822d8405 TH |
459 | * @pool: worker_pool to iterate workers of |
460 | * | |
1258fae7 | 461 | * This must be called with wq_pool_attach_mutex. |
822d8405 TH |
462 | * |
463 | * The if/else clause exists only for the lockdep assertion and can be | |
464 | * ignored. | |
465 | */ | |
da028469 LJ |
466 | #define for_each_pool_worker(worker, pool) \ |
467 | list_for_each_entry((worker), &(pool)->workers, node) \ | |
1258fae7 | 468 | if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \ |
822d8405 TH |
469 | else |
470 | ||
49e3cf44 TH |
471 | /** |
472 | * for_each_pwq - iterate through all pool_workqueues of the specified workqueue | |
473 | * @pwq: iteration cursor | |
474 | * @wq: the target workqueue | |
76af4d93 | 475 | * |
24acfb71 | 476 | * This must be called either with wq->mutex held or RCU read locked. |
794b18bc TH |
477 | * If the pwq needs to be used beyond the locking in effect, the caller is |
478 | * responsible for guaranteeing that the pwq stays online. | |
76af4d93 TH |
479 | * |
480 | * The if/else clause exists only for the lockdep assertion and can be | |
481 | * ignored. | |
49e3cf44 TH |
482 | */ |
483 | #define for_each_pwq(pwq, wq) \ | |
49e9d1a9 | 484 | list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \ |
5a644662 | 485 | lockdep_is_held(&(wq->mutex))) |
f3421797 | 486 | |
dc186ad7 TG |
487 | #ifdef CONFIG_DEBUG_OBJECTS_WORK |
488 | ||
f9e62f31 | 489 | static const struct debug_obj_descr work_debug_descr; |
dc186ad7 | 490 | |
99777288 SG |
491 | static void *work_debug_hint(void *addr) |
492 | { | |
493 | return ((struct work_struct *) addr)->func; | |
494 | } | |
495 | ||
b9fdac7f DC |
496 | static bool work_is_static_object(void *addr) |
497 | { | |
498 | struct work_struct *work = addr; | |
499 | ||
500 | return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work)); | |
501 | } | |
502 | ||
dc186ad7 TG |
503 | /* |
504 | * fixup_init is called when: | |
505 | * - an active object is initialized | |
506 | */ | |
02a982a6 | 507 | static bool work_fixup_init(void *addr, enum debug_obj_state state) |
dc186ad7 TG |
508 | { |
509 | struct work_struct *work = addr; | |
510 | ||
511 | switch (state) { | |
512 | case ODEBUG_STATE_ACTIVE: | |
513 | cancel_work_sync(work); | |
514 | debug_object_init(work, &work_debug_descr); | |
02a982a6 | 515 | return true; |
dc186ad7 | 516 | default: |
02a982a6 | 517 | return false; |
dc186ad7 TG |
518 | } |
519 | } | |
520 | ||
dc186ad7 TG |
521 | /* |
522 | * fixup_free is called when: | |
523 | * - an active object is freed | |
524 | */ | |
02a982a6 | 525 | static bool work_fixup_free(void *addr, enum debug_obj_state state) |
dc186ad7 TG |
526 | { |
527 | struct work_struct *work = addr; | |
528 | ||
529 | switch (state) { | |
530 | case ODEBUG_STATE_ACTIVE: | |
531 | cancel_work_sync(work); | |
532 | debug_object_free(work, &work_debug_descr); | |
02a982a6 | 533 | return true; |
dc186ad7 | 534 | default: |
02a982a6 | 535 | return false; |
dc186ad7 TG |
536 | } |
537 | } | |
538 | ||
f9e62f31 | 539 | static const struct debug_obj_descr work_debug_descr = { |
dc186ad7 | 540 | .name = "work_struct", |
99777288 | 541 | .debug_hint = work_debug_hint, |
b9fdac7f | 542 | .is_static_object = work_is_static_object, |
dc186ad7 | 543 | .fixup_init = work_fixup_init, |
dc186ad7 TG |
544 | .fixup_free = work_fixup_free, |
545 | }; | |
546 | ||
547 | static inline void debug_work_activate(struct work_struct *work) | |
548 | { | |
549 | debug_object_activate(work, &work_debug_descr); | |
550 | } | |
551 | ||
552 | static inline void debug_work_deactivate(struct work_struct *work) | |
553 | { | |
554 | debug_object_deactivate(work, &work_debug_descr); | |
555 | } | |
556 | ||
557 | void __init_work(struct work_struct *work, int onstack) | |
558 | { | |
559 | if (onstack) | |
560 | debug_object_init_on_stack(work, &work_debug_descr); | |
561 | else | |
562 | debug_object_init(work, &work_debug_descr); | |
563 | } | |
564 | EXPORT_SYMBOL_GPL(__init_work); | |
565 | ||
566 | void destroy_work_on_stack(struct work_struct *work) | |
567 | { | |
568 | debug_object_free(work, &work_debug_descr); | |
569 | } | |
570 | EXPORT_SYMBOL_GPL(destroy_work_on_stack); | |
571 | ||
ea2e64f2 TG |
572 | void destroy_delayed_work_on_stack(struct delayed_work *work) |
573 | { | |
574 | destroy_timer_on_stack(&work->timer); | |
575 | debug_object_free(&work->work, &work_debug_descr); | |
576 | } | |
577 | EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack); | |
578 | ||
dc186ad7 TG |
579 | #else |
580 | static inline void debug_work_activate(struct work_struct *work) { } | |
581 | static inline void debug_work_deactivate(struct work_struct *work) { } | |
582 | #endif | |
583 | ||
4e8b22bd | 584 | /** |
67dc8325 | 585 | * worker_pool_assign_id - allocate ID and assign it to @pool |
4e8b22bd LB |
586 | * @pool: the pool pointer of interest |
587 | * | |
588 | * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned | |
589 | * successfully, -errno on failure. | |
590 | */ | |
9daf9e67 TH |
591 | static int worker_pool_assign_id(struct worker_pool *pool) |
592 | { | |
593 | int ret; | |
594 | ||
68e13a67 | 595 | lockdep_assert_held(&wq_pool_mutex); |
5bcab335 | 596 | |
4e8b22bd LB |
597 | ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE, |
598 | GFP_KERNEL); | |
229641a6 | 599 | if (ret >= 0) { |
e68035fb | 600 | pool->id = ret; |
229641a6 TH |
601 | return 0; |
602 | } | |
fa1b54e6 | 603 | return ret; |
7c3eed5c TH |
604 | } |
605 | ||
df2d5ae4 TH |
606 | /** |
607 | * unbound_pwq_by_node - return the unbound pool_workqueue for the given node | |
608 | * @wq: the target workqueue | |
609 | * @node: the node ID | |
610 | * | |
24acfb71 | 611 | * This must be called with any of wq_pool_mutex, wq->mutex or RCU |
5b95e1af | 612 | * read locked. |
df2d5ae4 TH |
613 | * If the pwq needs to be used beyond the locking in effect, the caller is |
614 | * responsible for guaranteeing that the pwq stays online. | |
d185af30 YB |
615 | * |
616 | * Return: The unbound pool_workqueue for @node. | |
df2d5ae4 TH |
617 | */ |
618 | static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq, | |
619 | int node) | |
620 | { | |
5b95e1af | 621 | assert_rcu_or_wq_mutex_or_pool_mutex(wq); |
d6e022f1 TH |
622 | |
623 | /* | |
624 | * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a | |
625 | * delayed item is pending. The plan is to keep CPU -> NODE | |
626 | * mapping valid and stable across CPU on/offlines. Once that | |
627 | * happens, this workaround can be removed. | |
628 | */ | |
629 | if (unlikely(node == NUMA_NO_NODE)) | |
630 | return wq->dfl_pwq; | |
631 | ||
df2d5ae4 TH |
632 | return rcu_dereference_raw(wq->numa_pwq_tbl[node]); |
633 | } | |
634 | ||
73f53c4a TH |
635 | static unsigned int work_color_to_flags(int color) |
636 | { | |
637 | return color << WORK_STRUCT_COLOR_SHIFT; | |
638 | } | |
639 | ||
c4560c2c | 640 | static int get_work_color(unsigned long work_data) |
73f53c4a | 641 | { |
c4560c2c | 642 | return (work_data >> WORK_STRUCT_COLOR_SHIFT) & |
73f53c4a TH |
643 | ((1 << WORK_STRUCT_COLOR_BITS) - 1); |
644 | } | |
645 | ||
646 | static int work_next_color(int color) | |
647 | { | |
648 | return (color + 1) % WORK_NR_COLORS; | |
649 | } | |
1da177e4 | 650 | |
14441960 | 651 | /* |
112202d9 TH |
652 | * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data |
653 | * contain the pointer to the queued pwq. Once execution starts, the flag | |
7c3eed5c | 654 | * is cleared and the high bits contain OFFQ flags and pool ID. |
7a22ad75 | 655 | * |
112202d9 TH |
656 | * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() |
657 | * and clear_work_data() can be used to set the pwq, pool or clear | |
bbb68dfa TH |
658 | * work->data. These functions should only be called while the work is |
659 | * owned - ie. while the PENDING bit is set. | |
7a22ad75 | 660 | * |
112202d9 | 661 | * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq |
7c3eed5c | 662 | * corresponding to a work. Pool is available once the work has been |
112202d9 | 663 | * queued anywhere after initialization until it is sync canceled. pwq is |
7c3eed5c | 664 | * available only while the work item is queued. |
7a22ad75 | 665 | * |
bbb68dfa TH |
666 | * %WORK_OFFQ_CANCELING is used to mark a work item which is being |
667 | * canceled. While being canceled, a work item may have its PENDING set | |
668 | * but stay off timer and worklist for arbitrarily long and nobody should | |
669 | * try to steal the PENDING bit. | |
14441960 | 670 | */ |
7a22ad75 TH |
671 | static inline void set_work_data(struct work_struct *work, unsigned long data, |
672 | unsigned long flags) | |
365970a1 | 673 | { |
6183c009 | 674 | WARN_ON_ONCE(!work_pending(work)); |
7a22ad75 TH |
675 | atomic_long_set(&work->data, data | flags | work_static(work)); |
676 | } | |
365970a1 | 677 | |
112202d9 | 678 | static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, |
7a22ad75 TH |
679 | unsigned long extra_flags) |
680 | { | |
112202d9 TH |
681 | set_work_data(work, (unsigned long)pwq, |
682 | WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); | |
365970a1 DH |
683 | } |
684 | ||
4468a00f LJ |
685 | static void set_work_pool_and_keep_pending(struct work_struct *work, |
686 | int pool_id) | |
687 | { | |
688 | set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, | |
689 | WORK_STRUCT_PENDING); | |
690 | } | |
691 | ||
7c3eed5c TH |
692 | static void set_work_pool_and_clear_pending(struct work_struct *work, |
693 | int pool_id) | |
7a22ad75 | 694 | { |
23657bb1 TH |
695 | /* |
696 | * The following wmb is paired with the implied mb in | |
697 | * test_and_set_bit(PENDING) and ensures all updates to @work made | |
698 | * here are visible to and precede any updates by the next PENDING | |
699 | * owner. | |
700 | */ | |
701 | smp_wmb(); | |
7c3eed5c | 702 | set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); |
346c09f8 RP |
703 | /* |
704 | * The following mb guarantees that previous clear of a PENDING bit | |
705 | * will not be reordered with any speculative LOADS or STORES from | |
706 | * work->current_func, which is executed afterwards. This possible | |
8bdc6201 | 707 | * reordering can lead to a missed execution on attempt to queue |
346c09f8 RP |
708 | * the same @work. E.g. consider this case: |
709 | * | |
710 | * CPU#0 CPU#1 | |
711 | * ---------------------------- -------------------------------- | |
712 | * | |
713 | * 1 STORE event_indicated | |
714 | * 2 queue_work_on() { | |
715 | * 3 test_and_set_bit(PENDING) | |
716 | * 4 } set_..._and_clear_pending() { | |
717 | * 5 set_work_data() # clear bit | |
718 | * 6 smp_mb() | |
719 | * 7 work->current_func() { | |
720 | * 8 LOAD event_indicated | |
721 | * } | |
722 | * | |
723 | * Without an explicit full barrier speculative LOAD on line 8 can | |
724 | * be executed before CPU#0 does STORE on line 1. If that happens, | |
725 | * CPU#0 observes the PENDING bit is still set and new execution of | |
726 | * a @work is not queued in a hope, that CPU#1 will eventually | |
727 | * finish the queued @work. Meanwhile CPU#1 does not see | |
728 | * event_indicated is set, because speculative LOAD was executed | |
729 | * before actual STORE. | |
730 | */ | |
731 | smp_mb(); | |
7a22ad75 | 732 | } |
f756d5e2 | 733 | |
7a22ad75 | 734 | static void clear_work_data(struct work_struct *work) |
1da177e4 | 735 | { |
7c3eed5c TH |
736 | smp_wmb(); /* see set_work_pool_and_clear_pending() */ |
737 | set_work_data(work, WORK_STRUCT_NO_POOL, 0); | |
1da177e4 LT |
738 | } |
739 | ||
afa4bb77 LT |
740 | static inline struct pool_workqueue *work_struct_pwq(unsigned long data) |
741 | { | |
742 | return (struct pool_workqueue *)(data & WORK_STRUCT_WQ_DATA_MASK); | |
743 | } | |
744 | ||
112202d9 | 745 | static struct pool_workqueue *get_work_pwq(struct work_struct *work) |
b1f4ec17 | 746 | { |
e120153d | 747 | unsigned long data = atomic_long_read(&work->data); |
7a22ad75 | 748 | |
112202d9 | 749 | if (data & WORK_STRUCT_PWQ) |
afa4bb77 | 750 | return work_struct_pwq(data); |
e120153d TH |
751 | else |
752 | return NULL; | |
4d707b9f ON |
753 | } |
754 | ||
7c3eed5c TH |
755 | /** |
756 | * get_work_pool - return the worker_pool a given work was associated with | |
757 | * @work: the work item of interest | |
758 | * | |
68e13a67 | 759 | * Pools are created and destroyed under wq_pool_mutex, and allows read |
24acfb71 TG |
760 | * access under RCU read lock. As such, this function should be |
761 | * called under wq_pool_mutex or inside of a rcu_read_lock() region. | |
fa1b54e6 TH |
762 | * |
763 | * All fields of the returned pool are accessible as long as the above | |
764 | * mentioned locking is in effect. If the returned pool needs to be used | |
765 | * beyond the critical section, the caller is responsible for ensuring the | |
766 | * returned pool is and stays online. | |
d185af30 YB |
767 | * |
768 | * Return: The worker_pool @work was last associated with. %NULL if none. | |
7c3eed5c TH |
769 | */ |
770 | static struct worker_pool *get_work_pool(struct work_struct *work) | |
365970a1 | 771 | { |
e120153d | 772 | unsigned long data = atomic_long_read(&work->data); |
7c3eed5c | 773 | int pool_id; |
7a22ad75 | 774 | |
68e13a67 | 775 | assert_rcu_or_pool_mutex(); |
fa1b54e6 | 776 | |
112202d9 | 777 | if (data & WORK_STRUCT_PWQ) |
afa4bb77 | 778 | return work_struct_pwq(data)->pool; |
7a22ad75 | 779 | |
7c3eed5c TH |
780 | pool_id = data >> WORK_OFFQ_POOL_SHIFT; |
781 | if (pool_id == WORK_OFFQ_POOL_NONE) | |
7a22ad75 TH |
782 | return NULL; |
783 | ||
fa1b54e6 | 784 | return idr_find(&worker_pool_idr, pool_id); |
7c3eed5c TH |
785 | } |
786 | ||
787 | /** | |
788 | * get_work_pool_id - return the worker pool ID a given work is associated with | |
789 | * @work: the work item of interest | |
790 | * | |
d185af30 | 791 | * Return: The worker_pool ID @work was last associated with. |
7c3eed5c TH |
792 | * %WORK_OFFQ_POOL_NONE if none. |
793 | */ | |
794 | static int get_work_pool_id(struct work_struct *work) | |
795 | { | |
54d5b7d0 LJ |
796 | unsigned long data = atomic_long_read(&work->data); |
797 | ||
112202d9 | 798 | if (data & WORK_STRUCT_PWQ) |
afa4bb77 | 799 | return work_struct_pwq(data)->pool->id; |
7c3eed5c | 800 | |
54d5b7d0 | 801 | return data >> WORK_OFFQ_POOL_SHIFT; |
7c3eed5c TH |
802 | } |
803 | ||
bbb68dfa TH |
804 | static void mark_work_canceling(struct work_struct *work) |
805 | { | |
7c3eed5c | 806 | unsigned long pool_id = get_work_pool_id(work); |
bbb68dfa | 807 | |
7c3eed5c TH |
808 | pool_id <<= WORK_OFFQ_POOL_SHIFT; |
809 | set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); | |
bbb68dfa TH |
810 | } |
811 | ||
812 | static bool work_is_canceling(struct work_struct *work) | |
813 | { | |
814 | unsigned long data = atomic_long_read(&work->data); | |
815 | ||
112202d9 | 816 | return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); |
bbb68dfa TH |
817 | } |
818 | ||
e22bee78 | 819 | /* |
3270476a TH |
820 | * Policy functions. These define the policies on how the global worker |
821 | * pools are managed. Unless noted otherwise, these functions assume that | |
d565ed63 | 822 | * they're being called with pool->lock held. |
e22bee78 TH |
823 | */ |
824 | ||
63d95a91 | 825 | static bool __need_more_worker(struct worker_pool *pool) |
a848e3b6 | 826 | { |
bc35f7ef | 827 | return !pool->nr_running; |
a848e3b6 ON |
828 | } |
829 | ||
4594bf15 | 830 | /* |
e22bee78 TH |
831 | * Need to wake up a worker? Called from anything but currently |
832 | * running workers. | |
974271c4 TH |
833 | * |
834 | * Note that, because unbound workers never contribute to nr_running, this | |
706026c2 | 835 | * function will always return %true for unbound pools as long as the |
974271c4 | 836 | * worklist isn't empty. |
4594bf15 | 837 | */ |
63d95a91 | 838 | static bool need_more_worker(struct worker_pool *pool) |
365970a1 | 839 | { |
63d95a91 | 840 | return !list_empty(&pool->worklist) && __need_more_worker(pool); |
e22bee78 | 841 | } |
4594bf15 | 842 | |
e22bee78 | 843 | /* Can I start working? Called from busy but !running workers. */ |
63d95a91 | 844 | static bool may_start_working(struct worker_pool *pool) |
e22bee78 | 845 | { |
63d95a91 | 846 | return pool->nr_idle; |
e22bee78 TH |
847 | } |
848 | ||
849 | /* Do I need to keep working? Called from currently running workers. */ | |
63d95a91 | 850 | static bool keep_working(struct worker_pool *pool) |
e22bee78 | 851 | { |
bc35f7ef | 852 | return !list_empty(&pool->worklist) && (pool->nr_running <= 1); |
e22bee78 TH |
853 | } |
854 | ||
855 | /* Do we need a new worker? Called from manager. */ | |
63d95a91 | 856 | static bool need_to_create_worker(struct worker_pool *pool) |
e22bee78 | 857 | { |
63d95a91 | 858 | return need_more_worker(pool) && !may_start_working(pool); |
e22bee78 | 859 | } |
365970a1 | 860 | |
e22bee78 | 861 | /* Do we have too many workers and should some go away? */ |
63d95a91 | 862 | static bool too_many_workers(struct worker_pool *pool) |
e22bee78 | 863 | { |
692b4825 | 864 | bool managing = pool->flags & POOL_MANAGER_ACTIVE; |
63d95a91 TH |
865 | int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ |
866 | int nr_busy = pool->nr_workers - nr_idle; | |
e22bee78 TH |
867 | |
868 | return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; | |
365970a1 DH |
869 | } |
870 | ||
4d707b9f | 871 | /* |
e22bee78 TH |
872 | * Wake up functions. |
873 | */ | |
874 | ||
2c1f1a91 | 875 | /* Return the first idle worker. Called with pool->lock held. */ |
1037de36 | 876 | static struct worker *first_idle_worker(struct worker_pool *pool) |
7e11629d | 877 | { |
63d95a91 | 878 | if (unlikely(list_empty(&pool->idle_list))) |
7e11629d TH |
879 | return NULL; |
880 | ||
63d95a91 | 881 | return list_first_entry(&pool->idle_list, struct worker, entry); |
7e11629d TH |
882 | } |
883 | ||
884 | /** | |
885 | * wake_up_worker - wake up an idle worker | |
63d95a91 | 886 | * @pool: worker pool to wake worker from |
7e11629d | 887 | * |
63d95a91 | 888 | * Wake up the first idle worker of @pool. |
7e11629d TH |
889 | * |
890 | * CONTEXT: | |
a9b8a985 | 891 | * raw_spin_lock_irq(pool->lock). |
7e11629d | 892 | */ |
63d95a91 | 893 | static void wake_up_worker(struct worker_pool *pool) |
7e11629d | 894 | { |
1037de36 | 895 | struct worker *worker = first_idle_worker(pool); |
7e11629d TH |
896 | |
897 | if (likely(worker)) | |
898 | wake_up_process(worker->task); | |
899 | } | |
900 | ||
c54d5046 TH |
901 | /** |
902 | * worker_set_flags - set worker flags and adjust nr_running accordingly | |
903 | * @worker: self | |
904 | * @flags: flags to set | |
905 | * | |
906 | * Set @flags in @worker->flags and adjust nr_running accordingly. | |
907 | * | |
908 | * CONTEXT: | |
909 | * raw_spin_lock_irq(pool->lock) | |
910 | */ | |
911 | static inline void worker_set_flags(struct worker *worker, unsigned int flags) | |
912 | { | |
913 | struct worker_pool *pool = worker->pool; | |
914 | ||
915 | WARN_ON_ONCE(worker->task != current); | |
916 | ||
917 | /* If transitioning into NOT_RUNNING, adjust nr_running. */ | |
918 | if ((flags & WORKER_NOT_RUNNING) && | |
919 | !(worker->flags & WORKER_NOT_RUNNING)) { | |
920 | pool->nr_running--; | |
921 | } | |
922 | ||
923 | worker->flags |= flags; | |
924 | } | |
925 | ||
926 | /** | |
927 | * worker_clr_flags - clear worker flags and adjust nr_running accordingly | |
928 | * @worker: self | |
929 | * @flags: flags to clear | |
930 | * | |
931 | * Clear @flags in @worker->flags and adjust nr_running accordingly. | |
932 | * | |
933 | * CONTEXT: | |
934 | * raw_spin_lock_irq(pool->lock) | |
935 | */ | |
936 | static inline void worker_clr_flags(struct worker *worker, unsigned int flags) | |
937 | { | |
938 | struct worker_pool *pool = worker->pool; | |
939 | unsigned int oflags = worker->flags; | |
940 | ||
941 | WARN_ON_ONCE(worker->task != current); | |
942 | ||
943 | worker->flags &= ~flags; | |
944 | ||
945 | /* | |
946 | * If transitioning out of NOT_RUNNING, increment nr_running. Note | |
947 | * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask | |
948 | * of multiple flags, not a single flag. | |
949 | */ | |
950 | if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) | |
951 | if (!(worker->flags & WORKER_NOT_RUNNING)) | |
952 | pool->nr_running++; | |
953 | } | |
954 | ||
63638450 TH |
955 | #ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT |
956 | ||
957 | /* | |
958 | * Concurrency-managed per-cpu work items that hog CPU for longer than | |
959 | * wq_cpu_intensive_thresh_us trigger the automatic CPU_INTENSIVE mechanism, | |
960 | * which prevents them from stalling other concurrency-managed work items. If a | |
961 | * work function keeps triggering this mechanism, it's likely that the work item | |
962 | * should be using an unbound workqueue instead. | |
963 | * | |
964 | * wq_cpu_intensive_report() tracks work functions which trigger such conditions | |
965 | * and report them so that they can be examined and converted to use unbound | |
966 | * workqueues as appropriate. To avoid flooding the console, each violating work | |
967 | * function is tracked and reported with exponential backoff. | |
968 | */ | |
969 | #define WCI_MAX_ENTS 128 | |
970 | ||
971 | struct wci_ent { | |
972 | work_func_t func; | |
973 | atomic64_t cnt; | |
974 | struct hlist_node hash_node; | |
975 | }; | |
976 | ||
977 | static struct wci_ent wci_ents[WCI_MAX_ENTS]; | |
978 | static int wci_nr_ents; | |
979 | static DEFINE_RAW_SPINLOCK(wci_lock); | |
980 | static DEFINE_HASHTABLE(wci_hash, ilog2(WCI_MAX_ENTS)); | |
981 | ||
982 | static struct wci_ent *wci_find_ent(work_func_t func) | |
983 | { | |
984 | struct wci_ent *ent; | |
985 | ||
986 | hash_for_each_possible_rcu(wci_hash, ent, hash_node, | |
987 | (unsigned long)func) { | |
988 | if (ent->func == func) | |
989 | return ent; | |
990 | } | |
991 | return NULL; | |
992 | } | |
993 | ||
994 | static void wq_cpu_intensive_report(work_func_t func) | |
995 | { | |
996 | struct wci_ent *ent; | |
997 | ||
998 | restart: | |
999 | ent = wci_find_ent(func); | |
1000 | if (ent) { | |
1001 | u64 cnt; | |
1002 | ||
1003 | /* | |
1004 | * Start reporting from the fourth time and back off | |
1005 | * exponentially. | |
1006 | */ | |
1007 | cnt = atomic64_inc_return_relaxed(&ent->cnt); | |
1008 | if (cnt >= 4 && is_power_of_2(cnt)) | |
1009 | printk_deferred(KERN_WARNING "workqueue: %ps hogged CPU for >%luus %llu times, consider switching to WQ_UNBOUND\n", | |
1010 | ent->func, wq_cpu_intensive_thresh_us, | |
1011 | atomic64_read(&ent->cnt)); | |
1012 | return; | |
1013 | } | |
1014 | ||
1015 | /* | |
1016 | * @func is a new violation. Allocate a new entry for it. If wcn_ents[] | |
1017 | * is exhausted, something went really wrong and we probably made enough | |
1018 | * noise already. | |
1019 | */ | |
1020 | if (wci_nr_ents >= WCI_MAX_ENTS) | |
1021 | return; | |
1022 | ||
1023 | raw_spin_lock(&wci_lock); | |
1024 | ||
1025 | if (wci_nr_ents >= WCI_MAX_ENTS) { | |
1026 | raw_spin_unlock(&wci_lock); | |
1027 | return; | |
1028 | } | |
1029 | ||
1030 | if (wci_find_ent(func)) { | |
1031 | raw_spin_unlock(&wci_lock); | |
1032 | goto restart; | |
1033 | } | |
1034 | ||
1035 | ent = &wci_ents[wci_nr_ents++]; | |
1036 | ent->func = func; | |
1037 | atomic64_set(&ent->cnt, 1); | |
1038 | hash_add_rcu(wci_hash, &ent->hash_node, (unsigned long)func); | |
1039 | ||
1040 | raw_spin_unlock(&wci_lock); | |
1041 | } | |
1042 | ||
1043 | #else /* CONFIG_WQ_CPU_INTENSIVE_REPORT */ | |
1044 | static void wq_cpu_intensive_report(work_func_t func) {} | |
1045 | #endif /* CONFIG_WQ_CPU_INTENSIVE_REPORT */ | |
1046 | ||
d302f017 | 1047 | /** |
6d25be57 | 1048 | * wq_worker_running - a worker is running again |
e22bee78 | 1049 | * @task: task waking up |
e22bee78 | 1050 | * |
6d25be57 | 1051 | * This function is called when a worker returns from schedule() |
e22bee78 | 1052 | */ |
6d25be57 | 1053 | void wq_worker_running(struct task_struct *task) |
e22bee78 TH |
1054 | { |
1055 | struct worker *worker = kthread_data(task); | |
1056 | ||
c8f6219b | 1057 | if (!READ_ONCE(worker->sleeping)) |
6d25be57 | 1058 | return; |
07edfece FW |
1059 | |
1060 | /* | |
1061 | * If preempted by unbind_workers() between the WORKER_NOT_RUNNING check | |
1062 | * and the nr_running increment below, we may ruin the nr_running reset | |
1063 | * and leave with an unexpected pool->nr_running == 1 on the newly unbound | |
1064 | * pool. Protect against such race. | |
1065 | */ | |
1066 | preempt_disable(); | |
6d25be57 | 1067 | if (!(worker->flags & WORKER_NOT_RUNNING)) |
bc35f7ef | 1068 | worker->pool->nr_running++; |
07edfece | 1069 | preempt_enable(); |
616db877 TH |
1070 | |
1071 | /* | |
1072 | * CPU intensive auto-detection cares about how long a work item hogged | |
1073 | * CPU without sleeping. Reset the starting timestamp on wakeup. | |
1074 | */ | |
1075 | worker->current_at = worker->task->se.sum_exec_runtime; | |
1076 | ||
c8f6219b | 1077 | WRITE_ONCE(worker->sleeping, 0); |
e22bee78 TH |
1078 | } |
1079 | ||
1080 | /** | |
1081 | * wq_worker_sleeping - a worker is going to sleep | |
1082 | * @task: task going to sleep | |
e22bee78 | 1083 | * |
6d25be57 | 1084 | * This function is called from schedule() when a busy worker is |
ccf45156 | 1085 | * going to sleep. |
e22bee78 | 1086 | */ |
6d25be57 | 1087 | void wq_worker_sleeping(struct task_struct *task) |
e22bee78 | 1088 | { |
cc5bff38 | 1089 | struct worker *worker = kthread_data(task); |
111c225a | 1090 | struct worker_pool *pool; |
e22bee78 | 1091 | |
111c225a TH |
1092 | /* |
1093 | * Rescuers, which may not have all the fields set up like normal | |
1094 | * workers, also reach here, let's not access anything before | |
1095 | * checking NOT_RUNNING. | |
1096 | */ | |
2d64672e | 1097 | if (worker->flags & WORKER_NOT_RUNNING) |
6d25be57 | 1098 | return; |
e22bee78 | 1099 | |
111c225a | 1100 | pool = worker->pool; |
111c225a | 1101 | |
62849a96 | 1102 | /* Return if preempted before wq_worker_running() was reached */ |
c8f6219b | 1103 | if (READ_ONCE(worker->sleeping)) |
6d25be57 TG |
1104 | return; |
1105 | ||
c8f6219b | 1106 | WRITE_ONCE(worker->sleeping, 1); |
a9b8a985 | 1107 | raw_spin_lock_irq(&pool->lock); |
e22bee78 | 1108 | |
45c753f5 FW |
1109 | /* |
1110 | * Recheck in case unbind_workers() preempted us. We don't | |
1111 | * want to decrement nr_running after the worker is unbound | |
1112 | * and nr_running has been reset. | |
1113 | */ | |
1114 | if (worker->flags & WORKER_NOT_RUNNING) { | |
1115 | raw_spin_unlock_irq(&pool->lock); | |
1116 | return; | |
1117 | } | |
1118 | ||
bc35f7ef | 1119 | pool->nr_running--; |
725e8ec5 TH |
1120 | if (need_more_worker(pool)) { |
1121 | worker->current_pwq->stats[PWQ_STAT_CM_WAKEUP]++; | |
cc5bff38 | 1122 | wake_up_worker(pool); |
725e8ec5 | 1123 | } |
a9b8a985 | 1124 | raw_spin_unlock_irq(&pool->lock); |
e22bee78 TH |
1125 | } |
1126 | ||
616db877 TH |
1127 | /** |
1128 | * wq_worker_tick - a scheduler tick occurred while a kworker is running | |
1129 | * @task: task currently running | |
1130 | * | |
1131 | * Called from scheduler_tick(). We're in the IRQ context and the current | |
1132 | * worker's fields which follow the 'K' locking rule can be accessed safely. | |
1133 | */ | |
1134 | void wq_worker_tick(struct task_struct *task) | |
1135 | { | |
1136 | struct worker *worker = kthread_data(task); | |
1137 | struct pool_workqueue *pwq = worker->current_pwq; | |
1138 | struct worker_pool *pool = worker->pool; | |
1139 | ||
1140 | if (!pwq) | |
1141 | return; | |
1142 | ||
8a1dd1e5 TH |
1143 | pwq->stats[PWQ_STAT_CPU_TIME] += TICK_USEC; |
1144 | ||
18c8ae81 Z |
1145 | if (!wq_cpu_intensive_thresh_us) |
1146 | return; | |
1147 | ||
616db877 TH |
1148 | /* |
1149 | * If the current worker is concurrency managed and hogged the CPU for | |
1150 | * longer than wq_cpu_intensive_thresh_us, it's automatically marked | |
1151 | * CPU_INTENSIVE to avoid stalling other concurrency-managed work items. | |
c8f6219b Z |
1152 | * |
1153 | * Set @worker->sleeping means that @worker is in the process of | |
1154 | * switching out voluntarily and won't be contributing to | |
1155 | * @pool->nr_running until it wakes up. As wq_worker_sleeping() also | |
1156 | * decrements ->nr_running, setting CPU_INTENSIVE here can lead to | |
1157 | * double decrements. The task is releasing the CPU anyway. Let's skip. | |
1158 | * We probably want to make this prettier in the future. | |
616db877 | 1159 | */ |
c8f6219b | 1160 | if ((worker->flags & WORKER_NOT_RUNNING) || READ_ONCE(worker->sleeping) || |
616db877 TH |
1161 | worker->task->se.sum_exec_runtime - worker->current_at < |
1162 | wq_cpu_intensive_thresh_us * NSEC_PER_USEC) | |
1163 | return; | |
1164 | ||
1165 | raw_spin_lock(&pool->lock); | |
1166 | ||
1167 | worker_set_flags(worker, WORKER_CPU_INTENSIVE); | |
63638450 | 1168 | wq_cpu_intensive_report(worker->current_func); |
616db877 TH |
1169 | pwq->stats[PWQ_STAT_CPU_INTENSIVE]++; |
1170 | ||
1171 | if (need_more_worker(pool)) { | |
1172 | pwq->stats[PWQ_STAT_CM_WAKEUP]++; | |
1173 | wake_up_worker(pool); | |
1174 | } | |
1175 | ||
1176 | raw_spin_unlock(&pool->lock); | |
1177 | } | |
1178 | ||
1b69ac6b JW |
1179 | /** |
1180 | * wq_worker_last_func - retrieve worker's last work function | |
8194fe94 | 1181 | * @task: Task to retrieve last work function of. |
1b69ac6b JW |
1182 | * |
1183 | * Determine the last function a worker executed. This is called from | |
1184 | * the scheduler to get a worker's last known identity. | |
1185 | * | |
1186 | * CONTEXT: | |
a9b8a985 | 1187 | * raw_spin_lock_irq(rq->lock) |
1b69ac6b | 1188 | * |
4b047002 JW |
1189 | * This function is called during schedule() when a kworker is going |
1190 | * to sleep. It's used by psi to identify aggregation workers during | |
1191 | * dequeuing, to allow periodic aggregation to shut-off when that | |
1192 | * worker is the last task in the system or cgroup to go to sleep. | |
1193 | * | |
1194 | * As this function doesn't involve any workqueue-related locking, it | |
1195 | * only returns stable values when called from inside the scheduler's | |
1196 | * queuing and dequeuing paths, when @task, which must be a kworker, | |
1197 | * is guaranteed to not be processing any works. | |
1198 | * | |
1b69ac6b JW |
1199 | * Return: |
1200 | * The last work function %current executed as a worker, NULL if it | |
1201 | * hasn't executed any work yet. | |
1202 | */ | |
1203 | work_func_t wq_worker_last_func(struct task_struct *task) | |
1204 | { | |
1205 | struct worker *worker = kthread_data(task); | |
1206 | ||
1207 | return worker->last_func; | |
1208 | } | |
1209 | ||
8cca0eea TH |
1210 | /** |
1211 | * find_worker_executing_work - find worker which is executing a work | |
c9e7cf27 | 1212 | * @pool: pool of interest |
8cca0eea TH |
1213 | * @work: work to find worker for |
1214 | * | |
c9e7cf27 TH |
1215 | * Find a worker which is executing @work on @pool by searching |
1216 | * @pool->busy_hash which is keyed by the address of @work. For a worker | |
a2c1c57b TH |
1217 | * to match, its current execution should match the address of @work and |
1218 | * its work function. This is to avoid unwanted dependency between | |
1219 | * unrelated work executions through a work item being recycled while still | |
1220 | * being executed. | |
1221 | * | |
1222 | * This is a bit tricky. A work item may be freed once its execution | |
1223 | * starts and nothing prevents the freed area from being recycled for | |
1224 | * another work item. If the same work item address ends up being reused | |
1225 | * before the original execution finishes, workqueue will identify the | |
1226 | * recycled work item as currently executing and make it wait until the | |
1227 | * current execution finishes, introducing an unwanted dependency. | |
1228 | * | |
c5aa87bb TH |
1229 | * This function checks the work item address and work function to avoid |
1230 | * false positives. Note that this isn't complete as one may construct a | |
1231 | * work function which can introduce dependency onto itself through a | |
1232 | * recycled work item. Well, if somebody wants to shoot oneself in the | |
1233 | * foot that badly, there's only so much we can do, and if such deadlock | |
1234 | * actually occurs, it should be easy to locate the culprit work function. | |
8cca0eea TH |
1235 | * |
1236 | * CONTEXT: | |
a9b8a985 | 1237 | * raw_spin_lock_irq(pool->lock). |
8cca0eea | 1238 | * |
d185af30 YB |
1239 | * Return: |
1240 | * Pointer to worker which is executing @work if found, %NULL | |
8cca0eea | 1241 | * otherwise. |
4d707b9f | 1242 | */ |
c9e7cf27 | 1243 | static struct worker *find_worker_executing_work(struct worker_pool *pool, |
8cca0eea | 1244 | struct work_struct *work) |
4d707b9f | 1245 | { |
42f8570f | 1246 | struct worker *worker; |
42f8570f | 1247 | |
b67bfe0d | 1248 | hash_for_each_possible(pool->busy_hash, worker, hentry, |
a2c1c57b TH |
1249 | (unsigned long)work) |
1250 | if (worker->current_work == work && | |
1251 | worker->current_func == work->func) | |
42f8570f SL |
1252 | return worker; |
1253 | ||
1254 | return NULL; | |
4d707b9f ON |
1255 | } |
1256 | ||
bf4ede01 TH |
1257 | /** |
1258 | * move_linked_works - move linked works to a list | |
1259 | * @work: start of series of works to be scheduled | |
1260 | * @head: target list to append @work to | |
402dd89d | 1261 | * @nextp: out parameter for nested worklist walking |
bf4ede01 TH |
1262 | * |
1263 | * Schedule linked works starting from @work to @head. Work series to | |
1264 | * be scheduled starts at @work and includes any consecutive work with | |
1265 | * WORK_STRUCT_LINKED set in its predecessor. | |
1266 | * | |
1267 | * If @nextp is not NULL, it's updated to point to the next work of | |
1268 | * the last scheduled work. This allows move_linked_works() to be | |
1269 | * nested inside outer list_for_each_entry_safe(). | |
1270 | * | |
1271 | * CONTEXT: | |
a9b8a985 | 1272 | * raw_spin_lock_irq(pool->lock). |
bf4ede01 TH |
1273 | */ |
1274 | static void move_linked_works(struct work_struct *work, struct list_head *head, | |
1275 | struct work_struct **nextp) | |
1276 | { | |
1277 | struct work_struct *n; | |
1278 | ||
1279 | /* | |
1280 | * Linked worklist will always end before the end of the list, | |
1281 | * use NULL for list head. | |
1282 | */ | |
1283 | list_for_each_entry_safe_from(work, n, NULL, entry) { | |
1284 | list_move_tail(&work->entry, head); | |
1285 | if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) | |
1286 | break; | |
1287 | } | |
1288 | ||
1289 | /* | |
1290 | * If we're already inside safe list traversal and have moved | |
1291 | * multiple works to the scheduled queue, the next position | |
1292 | * needs to be updated. | |
1293 | */ | |
1294 | if (nextp) | |
1295 | *nextp = n; | |
1296 | } | |
1297 | ||
8864b4e5 TH |
1298 | /** |
1299 | * get_pwq - get an extra reference on the specified pool_workqueue | |
1300 | * @pwq: pool_workqueue to get | |
1301 | * | |
1302 | * Obtain an extra reference on @pwq. The caller should guarantee that | |
1303 | * @pwq has positive refcnt and be holding the matching pool->lock. | |
1304 | */ | |
1305 | static void get_pwq(struct pool_workqueue *pwq) | |
1306 | { | |
1307 | lockdep_assert_held(&pwq->pool->lock); | |
1308 | WARN_ON_ONCE(pwq->refcnt <= 0); | |
1309 | pwq->refcnt++; | |
1310 | } | |
1311 | ||
1312 | /** | |
1313 | * put_pwq - put a pool_workqueue reference | |
1314 | * @pwq: pool_workqueue to put | |
1315 | * | |
1316 | * Drop a reference of @pwq. If its refcnt reaches zero, schedule its | |
1317 | * destruction. The caller should be holding the matching pool->lock. | |
1318 | */ | |
1319 | static void put_pwq(struct pool_workqueue *pwq) | |
1320 | { | |
1321 | lockdep_assert_held(&pwq->pool->lock); | |
1322 | if (likely(--pwq->refcnt)) | |
1323 | return; | |
1324 | if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND))) | |
1325 | return; | |
1326 | /* | |
1327 | * @pwq can't be released under pool->lock, bounce to | |
1328 | * pwq_unbound_release_workfn(). This never recurses on the same | |
1329 | * pool->lock as this path is taken only for unbound workqueues and | |
1330 | * the release work item is scheduled on a per-cpu workqueue. To | |
1331 | * avoid lockdep warning, unbound pool->locks are given lockdep | |
1332 | * subclass of 1 in get_unbound_pool(). | |
1333 | */ | |
1334 | schedule_work(&pwq->unbound_release_work); | |
1335 | } | |
1336 | ||
dce90d47 TH |
1337 | /** |
1338 | * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock | |
1339 | * @pwq: pool_workqueue to put (can be %NULL) | |
1340 | * | |
1341 | * put_pwq() with locking. This function also allows %NULL @pwq. | |
1342 | */ | |
1343 | static void put_pwq_unlocked(struct pool_workqueue *pwq) | |
1344 | { | |
1345 | if (pwq) { | |
1346 | /* | |
24acfb71 | 1347 | * As both pwqs and pools are RCU protected, the |
dce90d47 TH |
1348 | * following lock operations are safe. |
1349 | */ | |
a9b8a985 | 1350 | raw_spin_lock_irq(&pwq->pool->lock); |
dce90d47 | 1351 | put_pwq(pwq); |
a9b8a985 | 1352 | raw_spin_unlock_irq(&pwq->pool->lock); |
dce90d47 TH |
1353 | } |
1354 | } | |
1355 | ||
f97a4a1a | 1356 | static void pwq_activate_inactive_work(struct work_struct *work) |
bf4ede01 | 1357 | { |
112202d9 | 1358 | struct pool_workqueue *pwq = get_work_pwq(work); |
bf4ede01 TH |
1359 | |
1360 | trace_workqueue_activate_work(work); | |
82607adc TH |
1361 | if (list_empty(&pwq->pool->worklist)) |
1362 | pwq->pool->watchdog_ts = jiffies; | |
112202d9 | 1363 | move_linked_works(work, &pwq->pool->worklist, NULL); |
f97a4a1a | 1364 | __clear_bit(WORK_STRUCT_INACTIVE_BIT, work_data_bits(work)); |
112202d9 | 1365 | pwq->nr_active++; |
bf4ede01 TH |
1366 | } |
1367 | ||
f97a4a1a | 1368 | static void pwq_activate_first_inactive(struct pool_workqueue *pwq) |
3aa62497 | 1369 | { |
f97a4a1a | 1370 | struct work_struct *work = list_first_entry(&pwq->inactive_works, |
3aa62497 LJ |
1371 | struct work_struct, entry); |
1372 | ||
f97a4a1a | 1373 | pwq_activate_inactive_work(work); |
3aa62497 LJ |
1374 | } |
1375 | ||
bf4ede01 | 1376 | /** |
112202d9 TH |
1377 | * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight |
1378 | * @pwq: pwq of interest | |
c4560c2c | 1379 | * @work_data: work_data of work which left the queue |
bf4ede01 TH |
1380 | * |
1381 | * A work either has completed or is removed from pending queue, | |
112202d9 | 1382 | * decrement nr_in_flight of its pwq and handle workqueue flushing. |
bf4ede01 TH |
1383 | * |
1384 | * CONTEXT: | |
a9b8a985 | 1385 | * raw_spin_lock_irq(pool->lock). |
bf4ede01 | 1386 | */ |
c4560c2c | 1387 | static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, unsigned long work_data) |
bf4ede01 | 1388 | { |
c4560c2c LJ |
1389 | int color = get_work_color(work_data); |
1390 | ||
018f3a13 LJ |
1391 | if (!(work_data & WORK_STRUCT_INACTIVE)) { |
1392 | pwq->nr_active--; | |
1393 | if (!list_empty(&pwq->inactive_works)) { | |
1394 | /* one down, submit an inactive one */ | |
1395 | if (pwq->nr_active < pwq->max_active) | |
1396 | pwq_activate_first_inactive(pwq); | |
1397 | } | |
1398 | } | |
1399 | ||
112202d9 | 1400 | pwq->nr_in_flight[color]--; |
bf4ede01 | 1401 | |
bf4ede01 | 1402 | /* is flush in progress and are we at the flushing tip? */ |
112202d9 | 1403 | if (likely(pwq->flush_color != color)) |
8864b4e5 | 1404 | goto out_put; |
bf4ede01 TH |
1405 | |
1406 | /* are there still in-flight works? */ | |
112202d9 | 1407 | if (pwq->nr_in_flight[color]) |
8864b4e5 | 1408 | goto out_put; |
bf4ede01 | 1409 | |
112202d9 TH |
1410 | /* this pwq is done, clear flush_color */ |
1411 | pwq->flush_color = -1; | |
bf4ede01 TH |
1412 | |
1413 | /* | |
112202d9 | 1414 | * If this was the last pwq, wake up the first flusher. It |
bf4ede01 TH |
1415 | * will handle the rest. |
1416 | */ | |
112202d9 TH |
1417 | if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) |
1418 | complete(&pwq->wq->first_flusher->done); | |
8864b4e5 TH |
1419 | out_put: |
1420 | put_pwq(pwq); | |
bf4ede01 TH |
1421 | } |
1422 | ||
36e227d2 | 1423 | /** |
bbb68dfa | 1424 | * try_to_grab_pending - steal work item from worklist and disable irq |
36e227d2 TH |
1425 | * @work: work item to steal |
1426 | * @is_dwork: @work is a delayed_work | |
bbb68dfa | 1427 | * @flags: place to store irq state |
36e227d2 TH |
1428 | * |
1429 | * Try to grab PENDING bit of @work. This function can handle @work in any | |
d185af30 | 1430 | * stable state - idle, on timer or on worklist. |
36e227d2 | 1431 | * |
d185af30 | 1432 | * Return: |
3eb6b31b MCC |
1433 | * |
1434 | * ======== ================================================================ | |
36e227d2 TH |
1435 | * 1 if @work was pending and we successfully stole PENDING |
1436 | * 0 if @work was idle and we claimed PENDING | |
1437 | * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry | |
bbb68dfa TH |
1438 | * -ENOENT if someone else is canceling @work, this state may persist |
1439 | * for arbitrarily long | |
3eb6b31b | 1440 | * ======== ================================================================ |
36e227d2 | 1441 | * |
d185af30 | 1442 | * Note: |
bbb68dfa | 1443 | * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting |
e0aecdd8 TH |
1444 | * interrupted while holding PENDING and @work off queue, irq must be |
1445 | * disabled on entry. This, combined with delayed_work->timer being | |
1446 | * irqsafe, ensures that we return -EAGAIN for finite short period of time. | |
bbb68dfa TH |
1447 | * |
1448 | * On successful return, >= 0, irq is disabled and the caller is | |
1449 | * responsible for releasing it using local_irq_restore(*@flags). | |
1450 | * | |
e0aecdd8 | 1451 | * This function is safe to call from any context including IRQ handler. |
bf4ede01 | 1452 | */ |
bbb68dfa TH |
1453 | static int try_to_grab_pending(struct work_struct *work, bool is_dwork, |
1454 | unsigned long *flags) | |
bf4ede01 | 1455 | { |
d565ed63 | 1456 | struct worker_pool *pool; |
112202d9 | 1457 | struct pool_workqueue *pwq; |
bf4ede01 | 1458 | |
bbb68dfa TH |
1459 | local_irq_save(*flags); |
1460 | ||
36e227d2 TH |
1461 | /* try to steal the timer if it exists */ |
1462 | if (is_dwork) { | |
1463 | struct delayed_work *dwork = to_delayed_work(work); | |
1464 | ||
e0aecdd8 TH |
1465 | /* |
1466 | * dwork->timer is irqsafe. If del_timer() fails, it's | |
1467 | * guaranteed that the timer is not queued anywhere and not | |
1468 | * running on the local CPU. | |
1469 | */ | |
36e227d2 TH |
1470 | if (likely(del_timer(&dwork->timer))) |
1471 | return 1; | |
1472 | } | |
1473 | ||
1474 | /* try to claim PENDING the normal way */ | |
bf4ede01 TH |
1475 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) |
1476 | return 0; | |
1477 | ||
24acfb71 | 1478 | rcu_read_lock(); |
bf4ede01 TH |
1479 | /* |
1480 | * The queueing is in progress, or it is already queued. Try to | |
1481 | * steal it from ->worklist without clearing WORK_STRUCT_PENDING. | |
1482 | */ | |
d565ed63 TH |
1483 | pool = get_work_pool(work); |
1484 | if (!pool) | |
bbb68dfa | 1485 | goto fail; |
bf4ede01 | 1486 | |
a9b8a985 | 1487 | raw_spin_lock(&pool->lock); |
0b3dae68 | 1488 | /* |
112202d9 TH |
1489 | * work->data is guaranteed to point to pwq only while the work |
1490 | * item is queued on pwq->wq, and both updating work->data to point | |
1491 | * to pwq on queueing and to pool on dequeueing are done under | |
1492 | * pwq->pool->lock. This in turn guarantees that, if work->data | |
1493 | * points to pwq which is associated with a locked pool, the work | |
0b3dae68 LJ |
1494 | * item is currently queued on that pool. |
1495 | */ | |
112202d9 TH |
1496 | pwq = get_work_pwq(work); |
1497 | if (pwq && pwq->pool == pool) { | |
16062836 TH |
1498 | debug_work_deactivate(work); |
1499 | ||
1500 | /* | |
018f3a13 LJ |
1501 | * A cancelable inactive work item must be in the |
1502 | * pwq->inactive_works since a queued barrier can't be | |
1503 | * canceled (see the comments in insert_wq_barrier()). | |
1504 | * | |
f97a4a1a | 1505 | * An inactive work item cannot be grabbed directly because |
d812796e | 1506 | * it might have linked barrier work items which, if left |
f97a4a1a | 1507 | * on the inactive_works list, will confuse pwq->nr_active |
16062836 TH |
1508 | * management later on and cause stall. Make sure the work |
1509 | * item is activated before grabbing. | |
1510 | */ | |
f97a4a1a LJ |
1511 | if (*work_data_bits(work) & WORK_STRUCT_INACTIVE) |
1512 | pwq_activate_inactive_work(work); | |
16062836 TH |
1513 | |
1514 | list_del_init(&work->entry); | |
c4560c2c | 1515 | pwq_dec_nr_in_flight(pwq, *work_data_bits(work)); |
16062836 | 1516 | |
112202d9 | 1517 | /* work->data points to pwq iff queued, point to pool */ |
16062836 TH |
1518 | set_work_pool_and_keep_pending(work, pool->id); |
1519 | ||
a9b8a985 | 1520 | raw_spin_unlock(&pool->lock); |
24acfb71 | 1521 | rcu_read_unlock(); |
16062836 | 1522 | return 1; |
bf4ede01 | 1523 | } |
a9b8a985 | 1524 | raw_spin_unlock(&pool->lock); |
bbb68dfa | 1525 | fail: |
24acfb71 | 1526 | rcu_read_unlock(); |
bbb68dfa TH |
1527 | local_irq_restore(*flags); |
1528 | if (work_is_canceling(work)) | |
1529 | return -ENOENT; | |
1530 | cpu_relax(); | |
36e227d2 | 1531 | return -EAGAIN; |
bf4ede01 TH |
1532 | } |
1533 | ||
4690c4ab | 1534 | /** |
706026c2 | 1535 | * insert_work - insert a work into a pool |
112202d9 | 1536 | * @pwq: pwq @work belongs to |
4690c4ab TH |
1537 | * @work: work to insert |
1538 | * @head: insertion point | |
1539 | * @extra_flags: extra WORK_STRUCT_* flags to set | |
1540 | * | |
112202d9 | 1541 | * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to |
706026c2 | 1542 | * work_struct flags. |
4690c4ab TH |
1543 | * |
1544 | * CONTEXT: | |
a9b8a985 | 1545 | * raw_spin_lock_irq(pool->lock). |
4690c4ab | 1546 | */ |
112202d9 TH |
1547 | static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, |
1548 | struct list_head *head, unsigned int extra_flags) | |
b89deed3 | 1549 | { |
112202d9 | 1550 | struct worker_pool *pool = pwq->pool; |
e22bee78 | 1551 | |
e89a85d6 | 1552 | /* record the work call stack in order to print it in KASAN reports */ |
f70da745 | 1553 | kasan_record_aux_stack_noalloc(work); |
e89a85d6 | 1554 | |
4690c4ab | 1555 | /* we own @work, set data and link */ |
112202d9 | 1556 | set_work_pwq(work, pwq, extra_flags); |
1a4d9b0a | 1557 | list_add_tail(&work->entry, head); |
8864b4e5 | 1558 | get_pwq(pwq); |
e22bee78 | 1559 | |
63d95a91 TH |
1560 | if (__need_more_worker(pool)) |
1561 | wake_up_worker(pool); | |
b89deed3 ON |
1562 | } |
1563 | ||
c8efcc25 TH |
1564 | /* |
1565 | * Test whether @work is being queued from another work executing on the | |
8d03ecfe | 1566 | * same workqueue. |
c8efcc25 TH |
1567 | */ |
1568 | static bool is_chained_work(struct workqueue_struct *wq) | |
1569 | { | |
8d03ecfe TH |
1570 | struct worker *worker; |
1571 | ||
1572 | worker = current_wq_worker(); | |
1573 | /* | |
bf393fd4 | 1574 | * Return %true iff I'm a worker executing a work item on @wq. If |
8d03ecfe TH |
1575 | * I'm @worker, it's safe to dereference it without locking. |
1576 | */ | |
112202d9 | 1577 | return worker && worker->current_pwq->wq == wq; |
c8efcc25 TH |
1578 | } |
1579 | ||
ef557180 MG |
1580 | /* |
1581 | * When queueing an unbound work item to a wq, prefer local CPU if allowed | |
1582 | * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to | |
1583 | * avoid perturbing sensitive tasks. | |
1584 | */ | |
1585 | static int wq_select_unbound_cpu(int cpu) | |
1586 | { | |
1587 | int new_cpu; | |
1588 | ||
f303fccb TH |
1589 | if (likely(!wq_debug_force_rr_cpu)) { |
1590 | if (cpumask_test_cpu(cpu, wq_unbound_cpumask)) | |
1591 | return cpu; | |
a8ec5880 AF |
1592 | } else { |
1593 | pr_warn_once("workqueue: round-robin CPU selection forced, expect performance impact\n"); | |
f303fccb TH |
1594 | } |
1595 | ||
ef557180 MG |
1596 | if (cpumask_empty(wq_unbound_cpumask)) |
1597 | return cpu; | |
1598 | ||
1599 | new_cpu = __this_cpu_read(wq_rr_cpu_last); | |
1600 | new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask); | |
1601 | if (unlikely(new_cpu >= nr_cpu_ids)) { | |
1602 | new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask); | |
1603 | if (unlikely(new_cpu >= nr_cpu_ids)) | |
1604 | return cpu; | |
1605 | } | |
1606 | __this_cpu_write(wq_rr_cpu_last, new_cpu); | |
1607 | ||
1608 | return new_cpu; | |
1609 | } | |
1610 | ||
d84ff051 | 1611 | static void __queue_work(int cpu, struct workqueue_struct *wq, |
1da177e4 LT |
1612 | struct work_struct *work) |
1613 | { | |
112202d9 | 1614 | struct pool_workqueue *pwq; |
c9178087 | 1615 | struct worker_pool *last_pool; |
1e19ffc6 | 1616 | struct list_head *worklist; |
8a2e8e5d | 1617 | unsigned int work_flags; |
b75cac93 | 1618 | unsigned int req_cpu = cpu; |
8930caba TH |
1619 | |
1620 | /* | |
1621 | * While a work item is PENDING && off queue, a task trying to | |
1622 | * steal the PENDING will busy-loop waiting for it to either get | |
1623 | * queued or lose PENDING. Grabbing PENDING and queueing should | |
1624 | * happen with IRQ disabled. | |
1625 | */ | |
8e8eb730 | 1626 | lockdep_assert_irqs_disabled(); |
1da177e4 | 1627 | |
1e19ffc6 | 1628 | |
33e3f0a3 RC |
1629 | /* |
1630 | * For a draining wq, only works from the same workqueue are | |
1631 | * allowed. The __WQ_DESTROYING helps to spot the issue that | |
1632 | * queues a new work item to a wq after destroy_workqueue(wq). | |
1633 | */ | |
1634 | if (unlikely(wq->flags & (__WQ_DESTROYING | __WQ_DRAINING) && | |
1635 | WARN_ON_ONCE(!is_chained_work(wq)))) | |
e41e704b | 1636 | return; |
24acfb71 | 1637 | rcu_read_lock(); |
9e8cd2f5 | 1638 | retry: |
c9178087 | 1639 | /* pwq which will be used unless @work is executing elsewhere */ |
aa202f1f HD |
1640 | if (wq->flags & WQ_UNBOUND) { |
1641 | if (req_cpu == WORK_CPU_UNBOUND) | |
1642 | cpu = wq_select_unbound_cpu(raw_smp_processor_id()); | |
df2d5ae4 | 1643 | pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); |
aa202f1f HD |
1644 | } else { |
1645 | if (req_cpu == WORK_CPU_UNBOUND) | |
1646 | cpu = raw_smp_processor_id(); | |
1647 | pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); | |
1648 | } | |
dbf2576e | 1649 | |
c9178087 TH |
1650 | /* |
1651 | * If @work was previously on a different pool, it might still be | |
1652 | * running there, in which case the work needs to be queued on that | |
1653 | * pool to guarantee non-reentrancy. | |
1654 | */ | |
1655 | last_pool = get_work_pool(work); | |
1656 | if (last_pool && last_pool != pwq->pool) { | |
1657 | struct worker *worker; | |
18aa9eff | 1658 | |
a9b8a985 | 1659 | raw_spin_lock(&last_pool->lock); |
18aa9eff | 1660 | |
c9178087 | 1661 | worker = find_worker_executing_work(last_pool, work); |
18aa9eff | 1662 | |
c9178087 TH |
1663 | if (worker && worker->current_pwq->wq == wq) { |
1664 | pwq = worker->current_pwq; | |
8930caba | 1665 | } else { |
c9178087 | 1666 | /* meh... not running there, queue here */ |
a9b8a985 SAS |
1667 | raw_spin_unlock(&last_pool->lock); |
1668 | raw_spin_lock(&pwq->pool->lock); | |
8930caba | 1669 | } |
f3421797 | 1670 | } else { |
a9b8a985 | 1671 | raw_spin_lock(&pwq->pool->lock); |
502ca9d8 TH |
1672 | } |
1673 | ||
9e8cd2f5 TH |
1674 | /* |
1675 | * pwq is determined and locked. For unbound pools, we could have | |
1676 | * raced with pwq release and it could already be dead. If its | |
1677 | * refcnt is zero, repeat pwq selection. Note that pwqs never die | |
df2d5ae4 TH |
1678 | * without another pwq replacing it in the numa_pwq_tbl or while |
1679 | * work items are executing on it, so the retrying is guaranteed to | |
9e8cd2f5 TH |
1680 | * make forward-progress. |
1681 | */ | |
1682 | if (unlikely(!pwq->refcnt)) { | |
1683 | if (wq->flags & WQ_UNBOUND) { | |
a9b8a985 | 1684 | raw_spin_unlock(&pwq->pool->lock); |
9e8cd2f5 TH |
1685 | cpu_relax(); |
1686 | goto retry; | |
1687 | } | |
1688 | /* oops */ | |
1689 | WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", | |
1690 | wq->name, cpu); | |
1691 | } | |
1692 | ||
112202d9 TH |
1693 | /* pwq determined, queue */ |
1694 | trace_workqueue_queue_work(req_cpu, pwq, work); | |
502ca9d8 | 1695 | |
24acfb71 TG |
1696 | if (WARN_ON(!list_empty(&work->entry))) |
1697 | goto out; | |
1e19ffc6 | 1698 | |
112202d9 TH |
1699 | pwq->nr_in_flight[pwq->work_color]++; |
1700 | work_flags = work_color_to_flags(pwq->work_color); | |
1e19ffc6 | 1701 | |
112202d9 | 1702 | if (likely(pwq->nr_active < pwq->max_active)) { |
cdadf009 | 1703 | trace_workqueue_activate_work(work); |
112202d9 TH |
1704 | pwq->nr_active++; |
1705 | worklist = &pwq->pool->worklist; | |
82607adc TH |
1706 | if (list_empty(worklist)) |
1707 | pwq->pool->watchdog_ts = jiffies; | |
8a2e8e5d | 1708 | } else { |
f97a4a1a LJ |
1709 | work_flags |= WORK_STRUCT_INACTIVE; |
1710 | worklist = &pwq->inactive_works; | |
8a2e8e5d | 1711 | } |
1e19ffc6 | 1712 | |
0687c66b | 1713 | debug_work_activate(work); |
112202d9 | 1714 | insert_work(pwq, work, worklist, work_flags); |
1e19ffc6 | 1715 | |
24acfb71 | 1716 | out: |
a9b8a985 | 1717 | raw_spin_unlock(&pwq->pool->lock); |
24acfb71 | 1718 | rcu_read_unlock(); |
1da177e4 LT |
1719 | } |
1720 | ||
0fcb78c2 | 1721 | /** |
c1a220e7 ZR |
1722 | * queue_work_on - queue work on specific cpu |
1723 | * @cpu: CPU number to execute work on | |
0fcb78c2 REB |
1724 | * @wq: workqueue to use |
1725 | * @work: work to queue | |
1726 | * | |
c1a220e7 | 1727 | * We queue the work to a specific CPU, the caller must ensure it |
443378f0 PM |
1728 | * can't go away. Callers that fail to ensure that the specified |
1729 | * CPU cannot go away will execute on a randomly chosen CPU. | |
854f5cc5 PM |
1730 | * But note well that callers specifying a CPU that never has been |
1731 | * online will get a splat. | |
d185af30 YB |
1732 | * |
1733 | * Return: %false if @work was already on a queue, %true otherwise. | |
1da177e4 | 1734 | */ |
d4283e93 TH |
1735 | bool queue_work_on(int cpu, struct workqueue_struct *wq, |
1736 | struct work_struct *work) | |
1da177e4 | 1737 | { |
d4283e93 | 1738 | bool ret = false; |
8930caba | 1739 | unsigned long flags; |
ef1ca236 | 1740 | |
8930caba | 1741 | local_irq_save(flags); |
c1a220e7 | 1742 | |
22df02bb | 1743 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { |
4690c4ab | 1744 | __queue_work(cpu, wq, work); |
d4283e93 | 1745 | ret = true; |
c1a220e7 | 1746 | } |
ef1ca236 | 1747 | |
8930caba | 1748 | local_irq_restore(flags); |
1da177e4 LT |
1749 | return ret; |
1750 | } | |
ad7b1f84 | 1751 | EXPORT_SYMBOL(queue_work_on); |
1da177e4 | 1752 | |
8204e0c1 AD |
1753 | /** |
1754 | * workqueue_select_cpu_near - Select a CPU based on NUMA node | |
1755 | * @node: NUMA node ID that we want to select a CPU from | |
1756 | * | |
1757 | * This function will attempt to find a "random" cpu available on a given | |
1758 | * node. If there are no CPUs available on the given node it will return | |
1759 | * WORK_CPU_UNBOUND indicating that we should just schedule to any | |
1760 | * available CPU if we need to schedule this work. | |
1761 | */ | |
1762 | static int workqueue_select_cpu_near(int node) | |
1763 | { | |
1764 | int cpu; | |
1765 | ||
1766 | /* No point in doing this if NUMA isn't enabled for workqueues */ | |
1767 | if (!wq_numa_enabled) | |
1768 | return WORK_CPU_UNBOUND; | |
1769 | ||
1770 | /* Delay binding to CPU if node is not valid or online */ | |
1771 | if (node < 0 || node >= MAX_NUMNODES || !node_online(node)) | |
1772 | return WORK_CPU_UNBOUND; | |
1773 | ||
1774 | /* Use local node/cpu if we are already there */ | |
1775 | cpu = raw_smp_processor_id(); | |
1776 | if (node == cpu_to_node(cpu)) | |
1777 | return cpu; | |
1778 | ||
1779 | /* Use "random" otherwise know as "first" online CPU of node */ | |
1780 | cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask); | |
1781 | ||
1782 | /* If CPU is valid return that, otherwise just defer */ | |
1783 | return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND; | |
1784 | } | |
1785 | ||
1786 | /** | |
1787 | * queue_work_node - queue work on a "random" cpu for a given NUMA node | |
1788 | * @node: NUMA node that we are targeting the work for | |
1789 | * @wq: workqueue to use | |
1790 | * @work: work to queue | |
1791 | * | |
1792 | * We queue the work to a "random" CPU within a given NUMA node. The basic | |
1793 | * idea here is to provide a way to somehow associate work with a given | |
1794 | * NUMA node. | |
1795 | * | |
1796 | * This function will only make a best effort attempt at getting this onto | |
1797 | * the right NUMA node. If no node is requested or the requested node is | |
1798 | * offline then we just fall back to standard queue_work behavior. | |
1799 | * | |
1800 | * Currently the "random" CPU ends up being the first available CPU in the | |
1801 | * intersection of cpu_online_mask and the cpumask of the node, unless we | |
1802 | * are running on the node. In that case we just use the current CPU. | |
1803 | * | |
1804 | * Return: %false if @work was already on a queue, %true otherwise. | |
1805 | */ | |
1806 | bool queue_work_node(int node, struct workqueue_struct *wq, | |
1807 | struct work_struct *work) | |
1808 | { | |
1809 | unsigned long flags; | |
1810 | bool ret = false; | |
1811 | ||
1812 | /* | |
1813 | * This current implementation is specific to unbound workqueues. | |
1814 | * Specifically we only return the first available CPU for a given | |
1815 | * node instead of cycling through individual CPUs within the node. | |
1816 | * | |
1817 | * If this is used with a per-cpu workqueue then the logic in | |
1818 | * workqueue_select_cpu_near would need to be updated to allow for | |
1819 | * some round robin type logic. | |
1820 | */ | |
1821 | WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)); | |
1822 | ||
1823 | local_irq_save(flags); | |
1824 | ||
1825 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { | |
1826 | int cpu = workqueue_select_cpu_near(node); | |
1827 | ||
1828 | __queue_work(cpu, wq, work); | |
1829 | ret = true; | |
1830 | } | |
1831 | ||
1832 | local_irq_restore(flags); | |
1833 | return ret; | |
1834 | } | |
1835 | EXPORT_SYMBOL_GPL(queue_work_node); | |
1836 | ||
8c20feb6 | 1837 | void delayed_work_timer_fn(struct timer_list *t) |
1da177e4 | 1838 | { |
8c20feb6 | 1839 | struct delayed_work *dwork = from_timer(dwork, t, timer); |
1da177e4 | 1840 | |
e0aecdd8 | 1841 | /* should have been called from irqsafe timer with irq already off */ |
60c057bc | 1842 | __queue_work(dwork->cpu, dwork->wq, &dwork->work); |
1da177e4 | 1843 | } |
1438ade5 | 1844 | EXPORT_SYMBOL(delayed_work_timer_fn); |
1da177e4 | 1845 | |
7beb2edf TH |
1846 | static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, |
1847 | struct delayed_work *dwork, unsigned long delay) | |
1da177e4 | 1848 | { |
7beb2edf TH |
1849 | struct timer_list *timer = &dwork->timer; |
1850 | struct work_struct *work = &dwork->work; | |
7beb2edf | 1851 | |
637fdbae | 1852 | WARN_ON_ONCE(!wq); |
4b243563 | 1853 | WARN_ON_ONCE(timer->function != delayed_work_timer_fn); |
fc4b514f TH |
1854 | WARN_ON_ONCE(timer_pending(timer)); |
1855 | WARN_ON_ONCE(!list_empty(&work->entry)); | |
7beb2edf | 1856 | |
8852aac2 TH |
1857 | /* |
1858 | * If @delay is 0, queue @dwork->work immediately. This is for | |
1859 | * both optimization and correctness. The earliest @timer can | |
1860 | * expire is on the closest next tick and delayed_work users depend | |
1861 | * on that there's no such delay when @delay is 0. | |
1862 | */ | |
1863 | if (!delay) { | |
1864 | __queue_work(cpu, wq, &dwork->work); | |
1865 | return; | |
1866 | } | |
1867 | ||
60c057bc | 1868 | dwork->wq = wq; |
1265057f | 1869 | dwork->cpu = cpu; |
7beb2edf TH |
1870 | timer->expires = jiffies + delay; |
1871 | ||
041bd12e TH |
1872 | if (unlikely(cpu != WORK_CPU_UNBOUND)) |
1873 | add_timer_on(timer, cpu); | |
1874 | else | |
1875 | add_timer(timer); | |
1da177e4 LT |
1876 | } |
1877 | ||
0fcb78c2 REB |
1878 | /** |
1879 | * queue_delayed_work_on - queue work on specific CPU after delay | |
1880 | * @cpu: CPU number to execute work on | |
1881 | * @wq: workqueue to use | |
af9997e4 | 1882 | * @dwork: work to queue |
0fcb78c2 REB |
1883 | * @delay: number of jiffies to wait before queueing |
1884 | * | |
d185af30 | 1885 | * Return: %false if @work was already on a queue, %true otherwise. If |
715f1300 TH |
1886 | * @delay is zero and @dwork is idle, it will be scheduled for immediate |
1887 | * execution. | |
0fcb78c2 | 1888 | */ |
d4283e93 TH |
1889 | bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, |
1890 | struct delayed_work *dwork, unsigned long delay) | |
7a6bc1cd | 1891 | { |
52bad64d | 1892 | struct work_struct *work = &dwork->work; |
d4283e93 | 1893 | bool ret = false; |
8930caba | 1894 | unsigned long flags; |
7a6bc1cd | 1895 | |
8930caba TH |
1896 | /* read the comment in __queue_work() */ |
1897 | local_irq_save(flags); | |
7a6bc1cd | 1898 | |
22df02bb | 1899 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { |
7beb2edf | 1900 | __queue_delayed_work(cpu, wq, dwork, delay); |
d4283e93 | 1901 | ret = true; |
7a6bc1cd | 1902 | } |
8a3e77cc | 1903 | |
8930caba | 1904 | local_irq_restore(flags); |
7a6bc1cd VP |
1905 | return ret; |
1906 | } | |
ad7b1f84 | 1907 | EXPORT_SYMBOL(queue_delayed_work_on); |
c7fc77f7 | 1908 | |
8376fe22 TH |
1909 | /** |
1910 | * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU | |
1911 | * @cpu: CPU number to execute work on | |
1912 | * @wq: workqueue to use | |
1913 | * @dwork: work to queue | |
1914 | * @delay: number of jiffies to wait before queueing | |
1915 | * | |
1916 | * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, | |
1917 | * modify @dwork's timer so that it expires after @delay. If @delay is | |
1918 | * zero, @work is guaranteed to be scheduled immediately regardless of its | |
1919 | * current state. | |
1920 | * | |
d185af30 | 1921 | * Return: %false if @dwork was idle and queued, %true if @dwork was |
8376fe22 TH |
1922 | * pending and its timer was modified. |
1923 | * | |
e0aecdd8 | 1924 | * This function is safe to call from any context including IRQ handler. |
8376fe22 TH |
1925 | * See try_to_grab_pending() for details. |
1926 | */ | |
1927 | bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, | |
1928 | struct delayed_work *dwork, unsigned long delay) | |
1929 | { | |
1930 | unsigned long flags; | |
1931 | int ret; | |
c7fc77f7 | 1932 | |
8376fe22 TH |
1933 | do { |
1934 | ret = try_to_grab_pending(&dwork->work, true, &flags); | |
1935 | } while (unlikely(ret == -EAGAIN)); | |
63bc0362 | 1936 | |
8376fe22 TH |
1937 | if (likely(ret >= 0)) { |
1938 | __queue_delayed_work(cpu, wq, dwork, delay); | |
1939 | local_irq_restore(flags); | |
7a6bc1cd | 1940 | } |
8376fe22 TH |
1941 | |
1942 | /* -ENOENT from try_to_grab_pending() becomes %true */ | |
7a6bc1cd VP |
1943 | return ret; |
1944 | } | |
8376fe22 TH |
1945 | EXPORT_SYMBOL_GPL(mod_delayed_work_on); |
1946 | ||
05f0fe6b TH |
1947 | static void rcu_work_rcufn(struct rcu_head *rcu) |
1948 | { | |
1949 | struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu); | |
1950 | ||
1951 | /* read the comment in __queue_work() */ | |
1952 | local_irq_disable(); | |
1953 | __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work); | |
1954 | local_irq_enable(); | |
1955 | } | |
1956 | ||
1957 | /** | |
1958 | * queue_rcu_work - queue work after a RCU grace period | |
1959 | * @wq: workqueue to use | |
1960 | * @rwork: work to queue | |
1961 | * | |
1962 | * Return: %false if @rwork was already pending, %true otherwise. Note | |
1963 | * that a full RCU grace period is guaranteed only after a %true return. | |
bf393fd4 | 1964 | * While @rwork is guaranteed to be executed after a %false return, the |
05f0fe6b TH |
1965 | * execution may happen before a full RCU grace period has passed. |
1966 | */ | |
1967 | bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork) | |
1968 | { | |
1969 | struct work_struct *work = &rwork->work; | |
1970 | ||
1971 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { | |
1972 | rwork->wq = wq; | |
a7e30c0e | 1973 | call_rcu_hurry(&rwork->rcu, rcu_work_rcufn); |
05f0fe6b TH |
1974 | return true; |
1975 | } | |
1976 | ||
1977 | return false; | |
1978 | } | |
1979 | EXPORT_SYMBOL(queue_rcu_work); | |
1980 | ||
c8e55f36 TH |
1981 | /** |
1982 | * worker_enter_idle - enter idle state | |
1983 | * @worker: worker which is entering idle state | |
1984 | * | |
1985 | * @worker is entering idle state. Update stats and idle timer if | |
1986 | * necessary. | |
1987 | * | |
1988 | * LOCKING: | |
a9b8a985 | 1989 | * raw_spin_lock_irq(pool->lock). |
c8e55f36 TH |
1990 | */ |
1991 | static void worker_enter_idle(struct worker *worker) | |
1da177e4 | 1992 | { |
bd7bdd43 | 1993 | struct worker_pool *pool = worker->pool; |
c8e55f36 | 1994 | |
6183c009 TH |
1995 | if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || |
1996 | WARN_ON_ONCE(!list_empty(&worker->entry) && | |
1997 | (worker->hentry.next || worker->hentry.pprev))) | |
1998 | return; | |
c8e55f36 | 1999 | |
051e1850 | 2000 | /* can't use worker_set_flags(), also called from create_worker() */ |
cb444766 | 2001 | worker->flags |= WORKER_IDLE; |
bd7bdd43 | 2002 | pool->nr_idle++; |
e22bee78 | 2003 | worker->last_active = jiffies; |
c8e55f36 TH |
2004 | |
2005 | /* idle_list is LIFO */ | |
bd7bdd43 | 2006 | list_add(&worker->entry, &pool->idle_list); |
db7bccf4 | 2007 | |
628c78e7 TH |
2008 | if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) |
2009 | mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); | |
cb444766 | 2010 | |
989442d7 | 2011 | /* Sanity check nr_running. */ |
bc35f7ef | 2012 | WARN_ON_ONCE(pool->nr_workers == pool->nr_idle && pool->nr_running); |
c8e55f36 TH |
2013 | } |
2014 | ||
2015 | /** | |
2016 | * worker_leave_idle - leave idle state | |
2017 | * @worker: worker which is leaving idle state | |
2018 | * | |
2019 | * @worker is leaving idle state. Update stats. | |
2020 | * | |
2021 | * LOCKING: | |
a9b8a985 | 2022 | * raw_spin_lock_irq(pool->lock). |
c8e55f36 TH |
2023 | */ |
2024 | static void worker_leave_idle(struct worker *worker) | |
2025 | { | |
bd7bdd43 | 2026 | struct worker_pool *pool = worker->pool; |
c8e55f36 | 2027 | |
6183c009 TH |
2028 | if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) |
2029 | return; | |
d302f017 | 2030 | worker_clr_flags(worker, WORKER_IDLE); |
bd7bdd43 | 2031 | pool->nr_idle--; |
c8e55f36 TH |
2032 | list_del_init(&worker->entry); |
2033 | } | |
2034 | ||
f7537df5 | 2035 | static struct worker *alloc_worker(int node) |
c34056a3 TH |
2036 | { |
2037 | struct worker *worker; | |
2038 | ||
f7537df5 | 2039 | worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node); |
c8e55f36 TH |
2040 | if (worker) { |
2041 | INIT_LIST_HEAD(&worker->entry); | |
affee4b2 | 2042 | INIT_LIST_HEAD(&worker->scheduled); |
da028469 | 2043 | INIT_LIST_HEAD(&worker->node); |
e22bee78 TH |
2044 | /* on creation a worker is in !idle && prep state */ |
2045 | worker->flags = WORKER_PREP; | |
c8e55f36 | 2046 | } |
c34056a3 TH |
2047 | return worker; |
2048 | } | |
2049 | ||
4736cbf7 LJ |
2050 | /** |
2051 | * worker_attach_to_pool() - attach a worker to a pool | |
2052 | * @worker: worker to be attached | |
2053 | * @pool: the target pool | |
2054 | * | |
2055 | * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and | |
2056 | * cpu-binding of @worker are kept coordinated with the pool across | |
2057 | * cpu-[un]hotplugs. | |
2058 | */ | |
2059 | static void worker_attach_to_pool(struct worker *worker, | |
2060 | struct worker_pool *pool) | |
2061 | { | |
1258fae7 | 2062 | mutex_lock(&wq_pool_attach_mutex); |
4736cbf7 | 2063 | |
4736cbf7 | 2064 | /* |
1258fae7 TH |
2065 | * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains |
2066 | * stable across this function. See the comments above the flag | |
2067 | * definition for details. | |
4736cbf7 LJ |
2068 | */ |
2069 | if (pool->flags & POOL_DISASSOCIATED) | |
2070 | worker->flags |= WORKER_UNBOUND; | |
5c25b5ff PZ |
2071 | else |
2072 | kthread_set_per_cpu(worker->task, pool->cpu); | |
4736cbf7 | 2073 | |
640f17c8 PZ |
2074 | if (worker->rescue_wq) |
2075 | set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask); | |
2076 | ||
4736cbf7 | 2077 | list_add_tail(&worker->node, &pool->workers); |
a2d812a2 | 2078 | worker->pool = pool; |
4736cbf7 | 2079 | |
1258fae7 | 2080 | mutex_unlock(&wq_pool_attach_mutex); |
4736cbf7 LJ |
2081 | } |
2082 | ||
60f5a4bc LJ |
2083 | /** |
2084 | * worker_detach_from_pool() - detach a worker from its pool | |
2085 | * @worker: worker which is attached to its pool | |
60f5a4bc | 2086 | * |
4736cbf7 LJ |
2087 | * Undo the attaching which had been done in worker_attach_to_pool(). The |
2088 | * caller worker shouldn't access to the pool after detached except it has | |
2089 | * other reference to the pool. | |
60f5a4bc | 2090 | */ |
a2d812a2 | 2091 | static void worker_detach_from_pool(struct worker *worker) |
60f5a4bc | 2092 | { |
a2d812a2 | 2093 | struct worker_pool *pool = worker->pool; |
60f5a4bc LJ |
2094 | struct completion *detach_completion = NULL; |
2095 | ||
1258fae7 | 2096 | mutex_lock(&wq_pool_attach_mutex); |
a2d812a2 | 2097 | |
5c25b5ff | 2098 | kthread_set_per_cpu(worker->task, -1); |
da028469 | 2099 | list_del(&worker->node); |
a2d812a2 TH |
2100 | worker->pool = NULL; |
2101 | ||
e02b9312 | 2102 | if (list_empty(&pool->workers) && list_empty(&pool->dying_workers)) |
60f5a4bc | 2103 | detach_completion = pool->detach_completion; |
1258fae7 | 2104 | mutex_unlock(&wq_pool_attach_mutex); |
60f5a4bc | 2105 | |
b62c0751 LJ |
2106 | /* clear leftover flags without pool->lock after it is detached */ |
2107 | worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND); | |
2108 | ||
60f5a4bc LJ |
2109 | if (detach_completion) |
2110 | complete(detach_completion); | |
2111 | } | |
2112 | ||
c34056a3 TH |
2113 | /** |
2114 | * create_worker - create a new workqueue worker | |
63d95a91 | 2115 | * @pool: pool the new worker will belong to |
c34056a3 | 2116 | * |
051e1850 | 2117 | * Create and start a new worker which is attached to @pool. |
c34056a3 TH |
2118 | * |
2119 | * CONTEXT: | |
2120 | * Might sleep. Does GFP_KERNEL allocations. | |
2121 | * | |
d185af30 | 2122 | * Return: |
c34056a3 TH |
2123 | * Pointer to the newly created worker. |
2124 | */ | |
bc2ae0f5 | 2125 | static struct worker *create_worker(struct worker_pool *pool) |
c34056a3 | 2126 | { |
e441b56f ZL |
2127 | struct worker *worker; |
2128 | int id; | |
e3c916a4 | 2129 | char id_buf[16]; |
c34056a3 | 2130 | |
7cda9aae | 2131 | /* ID is needed to determine kthread name */ |
e441b56f | 2132 | id = ida_alloc(&pool->worker_ida, GFP_KERNEL); |
3f0ea0b8 PM |
2133 | if (id < 0) { |
2134 | pr_err_once("workqueue: Failed to allocate a worker ID: %pe\n", | |
2135 | ERR_PTR(id)); | |
e441b56f | 2136 | return NULL; |
3f0ea0b8 | 2137 | } |
c34056a3 | 2138 | |
f7537df5 | 2139 | worker = alloc_worker(pool->node); |
3f0ea0b8 PM |
2140 | if (!worker) { |
2141 | pr_err_once("workqueue: Failed to allocate a worker\n"); | |
c34056a3 | 2142 | goto fail; |
3f0ea0b8 | 2143 | } |
c34056a3 | 2144 | |
c34056a3 TH |
2145 | worker->id = id; |
2146 | ||
29c91e99 | 2147 | if (pool->cpu >= 0) |
e3c916a4 TH |
2148 | snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, |
2149 | pool->attrs->nice < 0 ? "H" : ""); | |
f3421797 | 2150 | else |
e3c916a4 TH |
2151 | snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); |
2152 | ||
f3f90ad4 | 2153 | worker->task = kthread_create_on_node(worker_thread, worker, pool->node, |
e3c916a4 | 2154 | "kworker/%s", id_buf); |
3f0ea0b8 | 2155 | if (IS_ERR(worker->task)) { |
60f54038 PM |
2156 | if (PTR_ERR(worker->task) == -EINTR) { |
2157 | pr_err("workqueue: Interrupted when creating a worker thread \"kworker/%s\"\n", | |
2158 | id_buf); | |
2159 | } else { | |
2160 | pr_err_once("workqueue: Failed to create a worker thread: %pe", | |
2161 | worker->task); | |
2162 | } | |
c34056a3 | 2163 | goto fail; |
3f0ea0b8 | 2164 | } |
c34056a3 | 2165 | |
91151228 | 2166 | set_user_nice(worker->task, pool->attrs->nice); |
25834c73 | 2167 | kthread_bind_mask(worker->task, pool->attrs->cpumask); |
91151228 | 2168 | |
da028469 | 2169 | /* successful, attach the worker to the pool */ |
4736cbf7 | 2170 | worker_attach_to_pool(worker, pool); |
822d8405 | 2171 | |
051e1850 | 2172 | /* start the newly created worker */ |
a9b8a985 | 2173 | raw_spin_lock_irq(&pool->lock); |
051e1850 LJ |
2174 | worker->pool->nr_workers++; |
2175 | worker_enter_idle(worker); | |
2176 | wake_up_process(worker->task); | |
a9b8a985 | 2177 | raw_spin_unlock_irq(&pool->lock); |
051e1850 | 2178 | |
c34056a3 | 2179 | return worker; |
822d8405 | 2180 | |
c34056a3 | 2181 | fail: |
e441b56f | 2182 | ida_free(&pool->worker_ida, id); |
c34056a3 TH |
2183 | kfree(worker); |
2184 | return NULL; | |
2185 | } | |
2186 | ||
793777bc VS |
2187 | static void unbind_worker(struct worker *worker) |
2188 | { | |
2189 | lockdep_assert_held(&wq_pool_attach_mutex); | |
2190 | ||
2191 | kthread_set_per_cpu(worker->task, -1); | |
2192 | if (cpumask_intersects(wq_unbound_cpumask, cpu_active_mask)) | |
2193 | WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, wq_unbound_cpumask) < 0); | |
2194 | else | |
2195 | WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0); | |
2196 | } | |
2197 | ||
e02b9312 VS |
2198 | static void wake_dying_workers(struct list_head *cull_list) |
2199 | { | |
2200 | struct worker *worker, *tmp; | |
2201 | ||
2202 | list_for_each_entry_safe(worker, tmp, cull_list, entry) { | |
2203 | list_del_init(&worker->entry); | |
2204 | unbind_worker(worker); | |
2205 | /* | |
2206 | * If the worker was somehow already running, then it had to be | |
2207 | * in pool->idle_list when set_worker_dying() happened or we | |
2208 | * wouldn't have gotten here. | |
2209 | * | |
2210 | * Thus, the worker must either have observed the WORKER_DIE | |
2211 | * flag, or have set its state to TASK_IDLE. Either way, the | |
2212 | * below will be observed by the worker and is safe to do | |
2213 | * outside of pool->lock. | |
2214 | */ | |
2215 | wake_up_process(worker->task); | |
2216 | } | |
2217 | } | |
2218 | ||
c34056a3 | 2219 | /** |
e02b9312 | 2220 | * set_worker_dying - Tag a worker for destruction |
c34056a3 | 2221 | * @worker: worker to be destroyed |
e02b9312 | 2222 | * @list: transfer worker away from its pool->idle_list and into list |
c34056a3 | 2223 | * |
e02b9312 VS |
2224 | * Tag @worker for destruction and adjust @pool stats accordingly. The worker |
2225 | * should be idle. | |
c8e55f36 TH |
2226 | * |
2227 | * CONTEXT: | |
a9b8a985 | 2228 | * raw_spin_lock_irq(pool->lock). |
c34056a3 | 2229 | */ |
e02b9312 | 2230 | static void set_worker_dying(struct worker *worker, struct list_head *list) |
c34056a3 | 2231 | { |
bd7bdd43 | 2232 | struct worker_pool *pool = worker->pool; |
c34056a3 | 2233 | |
cd549687 | 2234 | lockdep_assert_held(&pool->lock); |
e02b9312 | 2235 | lockdep_assert_held(&wq_pool_attach_mutex); |
cd549687 | 2236 | |
c34056a3 | 2237 | /* sanity check frenzy */ |
6183c009 | 2238 | if (WARN_ON(worker->current_work) || |
73eb7fe7 LJ |
2239 | WARN_ON(!list_empty(&worker->scheduled)) || |
2240 | WARN_ON(!(worker->flags & WORKER_IDLE))) | |
6183c009 | 2241 | return; |
c34056a3 | 2242 | |
73eb7fe7 LJ |
2243 | pool->nr_workers--; |
2244 | pool->nr_idle--; | |
5bdfff96 | 2245 | |
cb444766 | 2246 | worker->flags |= WORKER_DIE; |
e02b9312 VS |
2247 | |
2248 | list_move(&worker->entry, list); | |
2249 | list_move(&worker->node, &pool->dying_workers); | |
c34056a3 TH |
2250 | } |
2251 | ||
3f959aa3 VS |
2252 | /** |
2253 | * idle_worker_timeout - check if some idle workers can now be deleted. | |
2254 | * @t: The pool's idle_timer that just expired | |
2255 | * | |
2256 | * The timer is armed in worker_enter_idle(). Note that it isn't disarmed in | |
2257 | * worker_leave_idle(), as a worker flicking between idle and active while its | |
2258 | * pool is at the too_many_workers() tipping point would cause too much timer | |
2259 | * housekeeping overhead. Since IDLE_WORKER_TIMEOUT is long enough, we just let | |
2260 | * it expire and re-evaluate things from there. | |
2261 | */ | |
32a6c723 | 2262 | static void idle_worker_timeout(struct timer_list *t) |
e22bee78 | 2263 | { |
32a6c723 | 2264 | struct worker_pool *pool = from_timer(pool, t, idle_timer); |
3f959aa3 VS |
2265 | bool do_cull = false; |
2266 | ||
2267 | if (work_pending(&pool->idle_cull_work)) | |
2268 | return; | |
e22bee78 | 2269 | |
a9b8a985 | 2270 | raw_spin_lock_irq(&pool->lock); |
e22bee78 | 2271 | |
3f959aa3 | 2272 | if (too_many_workers(pool)) { |
e22bee78 TH |
2273 | struct worker *worker; |
2274 | unsigned long expires; | |
2275 | ||
2276 | /* idle_list is kept in LIFO order, check the last one */ | |
3f959aa3 VS |
2277 | worker = list_entry(pool->idle_list.prev, struct worker, entry); |
2278 | expires = worker->last_active + IDLE_WORKER_TIMEOUT; | |
2279 | do_cull = !time_before(jiffies, expires); | |
2280 | ||
2281 | if (!do_cull) | |
2282 | mod_timer(&pool->idle_timer, expires); | |
2283 | } | |
2284 | raw_spin_unlock_irq(&pool->lock); | |
2285 | ||
2286 | if (do_cull) | |
2287 | queue_work(system_unbound_wq, &pool->idle_cull_work); | |
2288 | } | |
2289 | ||
2290 | /** | |
2291 | * idle_cull_fn - cull workers that have been idle for too long. | |
2292 | * @work: the pool's work for handling these idle workers | |
2293 | * | |
2294 | * This goes through a pool's idle workers and gets rid of those that have been | |
2295 | * idle for at least IDLE_WORKER_TIMEOUT seconds. | |
e02b9312 VS |
2296 | * |
2297 | * We don't want to disturb isolated CPUs because of a pcpu kworker being | |
2298 | * culled, so this also resets worker affinity. This requires a sleepable | |
2299 | * context, hence the split between timer callback and work item. | |
3f959aa3 VS |
2300 | */ |
2301 | static void idle_cull_fn(struct work_struct *work) | |
2302 | { | |
2303 | struct worker_pool *pool = container_of(work, struct worker_pool, idle_cull_work); | |
e02b9312 | 2304 | struct list_head cull_list; |
3f959aa3 | 2305 | |
e02b9312 VS |
2306 | INIT_LIST_HEAD(&cull_list); |
2307 | /* | |
2308 | * Grabbing wq_pool_attach_mutex here ensures an already-running worker | |
2309 | * cannot proceed beyong worker_detach_from_pool() in its self-destruct | |
2310 | * path. This is required as a previously-preempted worker could run after | |
2311 | * set_worker_dying() has happened but before wake_dying_workers() did. | |
2312 | */ | |
2313 | mutex_lock(&wq_pool_attach_mutex); | |
3f959aa3 VS |
2314 | raw_spin_lock_irq(&pool->lock); |
2315 | ||
2316 | while (too_many_workers(pool)) { | |
2317 | struct worker *worker; | |
2318 | unsigned long expires; | |
2319 | ||
63d95a91 | 2320 | worker = list_entry(pool->idle_list.prev, struct worker, entry); |
e22bee78 TH |
2321 | expires = worker->last_active + IDLE_WORKER_TIMEOUT; |
2322 | ||
3347fc9f | 2323 | if (time_before(jiffies, expires)) { |
63d95a91 | 2324 | mod_timer(&pool->idle_timer, expires); |
3347fc9f | 2325 | break; |
d5abe669 | 2326 | } |
3347fc9f | 2327 | |
e02b9312 | 2328 | set_worker_dying(worker, &cull_list); |
e22bee78 TH |
2329 | } |
2330 | ||
a9b8a985 | 2331 | raw_spin_unlock_irq(&pool->lock); |
e02b9312 VS |
2332 | wake_dying_workers(&cull_list); |
2333 | mutex_unlock(&wq_pool_attach_mutex); | |
e22bee78 | 2334 | } |
d5abe669 | 2335 | |
493a1724 | 2336 | static void send_mayday(struct work_struct *work) |
e22bee78 | 2337 | { |
112202d9 TH |
2338 | struct pool_workqueue *pwq = get_work_pwq(work); |
2339 | struct workqueue_struct *wq = pwq->wq; | |
493a1724 | 2340 | |
2e109a28 | 2341 | lockdep_assert_held(&wq_mayday_lock); |
e22bee78 | 2342 | |
493008a8 | 2343 | if (!wq->rescuer) |
493a1724 | 2344 | return; |
e22bee78 TH |
2345 | |
2346 | /* mayday mayday mayday */ | |
493a1724 | 2347 | if (list_empty(&pwq->mayday_node)) { |
77668c8b LJ |
2348 | /* |
2349 | * If @pwq is for an unbound wq, its base ref may be put at | |
2350 | * any time due to an attribute change. Pin @pwq until the | |
2351 | * rescuer is done with it. | |
2352 | */ | |
2353 | get_pwq(pwq); | |
493a1724 | 2354 | list_add_tail(&pwq->mayday_node, &wq->maydays); |
e22bee78 | 2355 | wake_up_process(wq->rescuer->task); |
725e8ec5 | 2356 | pwq->stats[PWQ_STAT_MAYDAY]++; |
493a1724 | 2357 | } |
e22bee78 TH |
2358 | } |
2359 | ||
32a6c723 | 2360 | static void pool_mayday_timeout(struct timer_list *t) |
e22bee78 | 2361 | { |
32a6c723 | 2362 | struct worker_pool *pool = from_timer(pool, t, mayday_timer); |
e22bee78 TH |
2363 | struct work_struct *work; |
2364 | ||
a9b8a985 SAS |
2365 | raw_spin_lock_irq(&pool->lock); |
2366 | raw_spin_lock(&wq_mayday_lock); /* for wq->maydays */ | |
e22bee78 | 2367 | |
63d95a91 | 2368 | if (need_to_create_worker(pool)) { |
e22bee78 TH |
2369 | /* |
2370 | * We've been trying to create a new worker but | |
2371 | * haven't been successful. We might be hitting an | |
2372 | * allocation deadlock. Send distress signals to | |
2373 | * rescuers. | |
2374 | */ | |
63d95a91 | 2375 | list_for_each_entry(work, &pool->worklist, entry) |
e22bee78 | 2376 | send_mayday(work); |
1da177e4 | 2377 | } |
e22bee78 | 2378 | |
a9b8a985 SAS |
2379 | raw_spin_unlock(&wq_mayday_lock); |
2380 | raw_spin_unlock_irq(&pool->lock); | |
e22bee78 | 2381 | |
63d95a91 | 2382 | mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); |
1da177e4 LT |
2383 | } |
2384 | ||
e22bee78 TH |
2385 | /** |
2386 | * maybe_create_worker - create a new worker if necessary | |
63d95a91 | 2387 | * @pool: pool to create a new worker for |
e22bee78 | 2388 | * |
63d95a91 | 2389 | * Create a new worker for @pool if necessary. @pool is guaranteed to |
e22bee78 TH |
2390 | * have at least one idle worker on return from this function. If |
2391 | * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is | |
63d95a91 | 2392 | * sent to all rescuers with works scheduled on @pool to resolve |
e22bee78 TH |
2393 | * possible allocation deadlock. |
2394 | * | |
c5aa87bb TH |
2395 | * On return, need_to_create_worker() is guaranteed to be %false and |
2396 | * may_start_working() %true. | |
e22bee78 TH |
2397 | * |
2398 | * LOCKING: | |
a9b8a985 | 2399 | * raw_spin_lock_irq(pool->lock) which may be released and regrabbed |
e22bee78 TH |
2400 | * multiple times. Does GFP_KERNEL allocations. Called only from |
2401 | * manager. | |
e22bee78 | 2402 | */ |
29187a9e | 2403 | static void maybe_create_worker(struct worker_pool *pool) |
d565ed63 TH |
2404 | __releases(&pool->lock) |
2405 | __acquires(&pool->lock) | |
1da177e4 | 2406 | { |
e22bee78 | 2407 | restart: |
a9b8a985 | 2408 | raw_spin_unlock_irq(&pool->lock); |
9f9c2364 | 2409 | |
e22bee78 | 2410 | /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ |
63d95a91 | 2411 | mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); |
e22bee78 TH |
2412 | |
2413 | while (true) { | |
051e1850 | 2414 | if (create_worker(pool) || !need_to_create_worker(pool)) |
e22bee78 | 2415 | break; |
1da177e4 | 2416 | |
e212f361 | 2417 | schedule_timeout_interruptible(CREATE_COOLDOWN); |
9f9c2364 | 2418 | |
63d95a91 | 2419 | if (!need_to_create_worker(pool)) |
e22bee78 TH |
2420 | break; |
2421 | } | |
2422 | ||
63d95a91 | 2423 | del_timer_sync(&pool->mayday_timer); |
a9b8a985 | 2424 | raw_spin_lock_irq(&pool->lock); |
051e1850 LJ |
2425 | /* |
2426 | * This is necessary even after a new worker was just successfully | |
2427 | * created as @pool->lock was dropped and the new worker might have | |
2428 | * already become busy. | |
2429 | */ | |
63d95a91 | 2430 | if (need_to_create_worker(pool)) |
e22bee78 | 2431 | goto restart; |
e22bee78 TH |
2432 | } |
2433 | ||
73f53c4a | 2434 | /** |
e22bee78 TH |
2435 | * manage_workers - manage worker pool |
2436 | * @worker: self | |
73f53c4a | 2437 | * |
706026c2 | 2438 | * Assume the manager role and manage the worker pool @worker belongs |
e22bee78 | 2439 | * to. At any given time, there can be only zero or one manager per |
706026c2 | 2440 | * pool. The exclusion is handled automatically by this function. |
e22bee78 TH |
2441 | * |
2442 | * The caller can safely start processing works on false return. On | |
2443 | * true return, it's guaranteed that need_to_create_worker() is false | |
2444 | * and may_start_working() is true. | |
73f53c4a TH |
2445 | * |
2446 | * CONTEXT: | |
a9b8a985 | 2447 | * raw_spin_lock_irq(pool->lock) which may be released and regrabbed |
e22bee78 TH |
2448 | * multiple times. Does GFP_KERNEL allocations. |
2449 | * | |
d185af30 | 2450 | * Return: |
29187a9e TH |
2451 | * %false if the pool doesn't need management and the caller can safely |
2452 | * start processing works, %true if management function was performed and | |
2453 | * the conditions that the caller verified before calling the function may | |
2454 | * no longer be true. | |
73f53c4a | 2455 | */ |
e22bee78 | 2456 | static bool manage_workers(struct worker *worker) |
73f53c4a | 2457 | { |
63d95a91 | 2458 | struct worker_pool *pool = worker->pool; |
73f53c4a | 2459 | |
692b4825 | 2460 | if (pool->flags & POOL_MANAGER_ACTIVE) |
29187a9e | 2461 | return false; |
692b4825 TH |
2462 | |
2463 | pool->flags |= POOL_MANAGER_ACTIVE; | |
2607d7a6 | 2464 | pool->manager = worker; |
1e19ffc6 | 2465 | |
29187a9e | 2466 | maybe_create_worker(pool); |
e22bee78 | 2467 | |
2607d7a6 | 2468 | pool->manager = NULL; |
692b4825 | 2469 | pool->flags &= ~POOL_MANAGER_ACTIVE; |
d8bb65ab | 2470 | rcuwait_wake_up(&manager_wait); |
29187a9e | 2471 | return true; |
73f53c4a TH |
2472 | } |
2473 | ||
a62428c0 TH |
2474 | /** |
2475 | * process_one_work - process single work | |
c34056a3 | 2476 | * @worker: self |
a62428c0 TH |
2477 | * @work: work to process |
2478 | * | |
2479 | * Process @work. This function contains all the logics necessary to | |
2480 | * process a single work including synchronization against and | |
2481 | * interaction with other workers on the same cpu, queueing and | |
2482 | * flushing. As long as context requirement is met, any worker can | |
2483 | * call this function to process a work. | |
2484 | * | |
2485 | * CONTEXT: | |
a9b8a985 | 2486 | * raw_spin_lock_irq(pool->lock) which is released and regrabbed. |
a62428c0 | 2487 | */ |
c34056a3 | 2488 | static void process_one_work(struct worker *worker, struct work_struct *work) |
d565ed63 TH |
2489 | __releases(&pool->lock) |
2490 | __acquires(&pool->lock) | |
a62428c0 | 2491 | { |
112202d9 | 2492 | struct pool_workqueue *pwq = get_work_pwq(work); |
bd7bdd43 | 2493 | struct worker_pool *pool = worker->pool; |
c4560c2c | 2494 | unsigned long work_data; |
7e11629d | 2495 | struct worker *collision; |
a62428c0 TH |
2496 | #ifdef CONFIG_LOCKDEP |
2497 | /* | |
2498 | * It is permissible to free the struct work_struct from | |
2499 | * inside the function that is called from it, this we need to | |
2500 | * take into account for lockdep too. To avoid bogus "held | |
2501 | * lock freed" warnings as well as problems when looking into | |
2502 | * work->lockdep_map, make a copy and use that here. | |
2503 | */ | |
4d82a1de PZ |
2504 | struct lockdep_map lockdep_map; |
2505 | ||
2506 | lockdep_copy_map(&lockdep_map, &work->lockdep_map); | |
a62428c0 | 2507 | #endif |
807407c0 | 2508 | /* ensure we're on the correct CPU */ |
85327af6 | 2509 | WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && |
ec22ca5e | 2510 | raw_smp_processor_id() != pool->cpu); |
25511a47 | 2511 | |
7e11629d TH |
2512 | /* |
2513 | * A single work shouldn't be executed concurrently by | |
2514 | * multiple workers on a single cpu. Check whether anyone is | |
2515 | * already processing the work. If so, defer the work to the | |
2516 | * currently executing one. | |
2517 | */ | |
c9e7cf27 | 2518 | collision = find_worker_executing_work(pool, work); |
7e11629d TH |
2519 | if (unlikely(collision)) { |
2520 | move_linked_works(work, &collision->scheduled, NULL); | |
2521 | return; | |
2522 | } | |
2523 | ||
8930caba | 2524 | /* claim and dequeue */ |
a62428c0 | 2525 | debug_work_deactivate(work); |
c9e7cf27 | 2526 | hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); |
c34056a3 | 2527 | worker->current_work = work; |
a2c1c57b | 2528 | worker->current_func = work->func; |
112202d9 | 2529 | worker->current_pwq = pwq; |
616db877 | 2530 | worker->current_at = worker->task->se.sum_exec_runtime; |
c4560c2c | 2531 | work_data = *work_data_bits(work); |
d812796e | 2532 | worker->current_color = get_work_color(work_data); |
7a22ad75 | 2533 | |
8bf89593 TH |
2534 | /* |
2535 | * Record wq name for cmdline and debug reporting, may get | |
2536 | * overridden through set_worker_desc(). | |
2537 | */ | |
2538 | strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN); | |
2539 | ||
a62428c0 TH |
2540 | list_del_init(&work->entry); |
2541 | ||
fb0e7beb | 2542 | /* |
228f1d00 LJ |
2543 | * CPU intensive works don't participate in concurrency management. |
2544 | * They're the scheduler's responsibility. This takes @worker out | |
2545 | * of concurrency management and the next code block will chain | |
2546 | * execution of the pending work items. | |
fb0e7beb | 2547 | */ |
616db877 | 2548 | if (unlikely(pwq->wq->flags & WQ_CPU_INTENSIVE)) |
228f1d00 | 2549 | worker_set_flags(worker, WORKER_CPU_INTENSIVE); |
fb0e7beb | 2550 | |
974271c4 | 2551 | /* |
a489a03e LJ |
2552 | * Wake up another worker if necessary. The condition is always |
2553 | * false for normal per-cpu workers since nr_running would always | |
2554 | * be >= 1 at this point. This is used to chain execution of the | |
2555 | * pending work items for WORKER_NOT_RUNNING workers such as the | |
228f1d00 | 2556 | * UNBOUND and CPU_INTENSIVE ones. |
974271c4 | 2557 | */ |
a489a03e | 2558 | if (need_more_worker(pool)) |
63d95a91 | 2559 | wake_up_worker(pool); |
974271c4 | 2560 | |
8930caba | 2561 | /* |
7c3eed5c | 2562 | * Record the last pool and clear PENDING which should be the last |
d565ed63 | 2563 | * update to @work. Also, do this inside @pool->lock so that |
23657bb1 TH |
2564 | * PENDING and queued state changes happen together while IRQ is |
2565 | * disabled. | |
8930caba | 2566 | */ |
7c3eed5c | 2567 | set_work_pool_and_clear_pending(work, pool->id); |
a62428c0 | 2568 | |
a9b8a985 | 2569 | raw_spin_unlock_irq(&pool->lock); |
a62428c0 | 2570 | |
a1d14934 | 2571 | lock_map_acquire(&pwq->wq->lockdep_map); |
a62428c0 | 2572 | lock_map_acquire(&lockdep_map); |
e6f3faa7 | 2573 | /* |
f52be570 PZ |
2574 | * Strictly speaking we should mark the invariant state without holding |
2575 | * any locks, that is, before these two lock_map_acquire()'s. | |
e6f3faa7 PZ |
2576 | * |
2577 | * However, that would result in: | |
2578 | * | |
2579 | * A(W1) | |
2580 | * WFC(C) | |
2581 | * A(W1) | |
2582 | * C(C) | |
2583 | * | |
2584 | * Which would create W1->C->W1 dependencies, even though there is no | |
2585 | * actual deadlock possible. There are two solutions, using a | |
2586 | * read-recursive acquire on the work(queue) 'locks', but this will then | |
f52be570 | 2587 | * hit the lockdep limitation on recursive locks, or simply discard |
e6f3faa7 PZ |
2588 | * these locks. |
2589 | * | |
2590 | * AFAICT there is no possible deadlock scenario between the | |
2591 | * flush_work() and complete() primitives (except for single-threaded | |
2592 | * workqueues), so hiding them isn't a problem. | |
2593 | */ | |
f52be570 | 2594 | lockdep_invariant_state(true); |
725e8ec5 | 2595 | pwq->stats[PWQ_STAT_STARTED]++; |
e36c886a | 2596 | trace_workqueue_execute_start(work); |
a2c1c57b | 2597 | worker->current_func(work); |
e36c886a AV |
2598 | /* |
2599 | * While we must be careful to not use "work" after this, the trace | |
2600 | * point will only record its address. | |
2601 | */ | |
1c5da0ec | 2602 | trace_workqueue_execute_end(work, worker->current_func); |
725e8ec5 | 2603 | pwq->stats[PWQ_STAT_COMPLETED]++; |
a62428c0 | 2604 | lock_map_release(&lockdep_map); |
112202d9 | 2605 | lock_map_release(&pwq->wq->lockdep_map); |
a62428c0 TH |
2606 | |
2607 | if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { | |
044c782c | 2608 | pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" |
d75f773c | 2609 | " last function: %ps\n", |
a2c1c57b TH |
2610 | current->comm, preempt_count(), task_pid_nr(current), |
2611 | worker->current_func); | |
a62428c0 TH |
2612 | debug_show_held_locks(current); |
2613 | dump_stack(); | |
2614 | } | |
2615 | ||
b22ce278 | 2616 | /* |
025f50f3 | 2617 | * The following prevents a kworker from hogging CPU on !PREEMPTION |
b22ce278 TH |
2618 | * kernels, where a requeueing work item waiting for something to |
2619 | * happen could deadlock with stop_machine as such work item could | |
2620 | * indefinitely requeue itself while all other CPUs are trapped in | |
789cbbec JL |
2621 | * stop_machine. At the same time, report a quiescent RCU state so |
2622 | * the same condition doesn't freeze RCU. | |
b22ce278 | 2623 | */ |
a7e6425e | 2624 | cond_resched(); |
b22ce278 | 2625 | |
a9b8a985 | 2626 | raw_spin_lock_irq(&pool->lock); |
a62428c0 | 2627 | |
616db877 TH |
2628 | /* |
2629 | * In addition to %WQ_CPU_INTENSIVE, @worker may also have been marked | |
2630 | * CPU intensive by wq_worker_tick() if @work hogged CPU longer than | |
2631 | * wq_cpu_intensive_thresh_us. Clear it. | |
2632 | */ | |
2633 | worker_clr_flags(worker, WORKER_CPU_INTENSIVE); | |
fb0e7beb | 2634 | |
1b69ac6b JW |
2635 | /* tag the worker for identification in schedule() */ |
2636 | worker->last_func = worker->current_func; | |
2637 | ||
a62428c0 | 2638 | /* we're done with it, release */ |
42f8570f | 2639 | hash_del(&worker->hentry); |
c34056a3 | 2640 | worker->current_work = NULL; |
a2c1c57b | 2641 | worker->current_func = NULL; |
112202d9 | 2642 | worker->current_pwq = NULL; |
d812796e | 2643 | worker->current_color = INT_MAX; |
c4560c2c | 2644 | pwq_dec_nr_in_flight(pwq, work_data); |
a62428c0 TH |
2645 | } |
2646 | ||
affee4b2 TH |
2647 | /** |
2648 | * process_scheduled_works - process scheduled works | |
2649 | * @worker: self | |
2650 | * | |
2651 | * Process all scheduled works. Please note that the scheduled list | |
2652 | * may change while processing a work, so this function repeatedly | |
2653 | * fetches a work from the top and executes it. | |
2654 | * | |
2655 | * CONTEXT: | |
a9b8a985 | 2656 | * raw_spin_lock_irq(pool->lock) which may be released and regrabbed |
affee4b2 TH |
2657 | * multiple times. |
2658 | */ | |
2659 | static void process_scheduled_works(struct worker *worker) | |
1da177e4 | 2660 | { |
affee4b2 TH |
2661 | while (!list_empty(&worker->scheduled)) { |
2662 | struct work_struct *work = list_first_entry(&worker->scheduled, | |
1da177e4 | 2663 | struct work_struct, entry); |
c34056a3 | 2664 | process_one_work(worker, work); |
1da177e4 | 2665 | } |
1da177e4 LT |
2666 | } |
2667 | ||
197f6acc TH |
2668 | static void set_pf_worker(bool val) |
2669 | { | |
2670 | mutex_lock(&wq_pool_attach_mutex); | |
2671 | if (val) | |
2672 | current->flags |= PF_WQ_WORKER; | |
2673 | else | |
2674 | current->flags &= ~PF_WQ_WORKER; | |
2675 | mutex_unlock(&wq_pool_attach_mutex); | |
2676 | } | |
2677 | ||
4690c4ab TH |
2678 | /** |
2679 | * worker_thread - the worker thread function | |
c34056a3 | 2680 | * @__worker: self |
4690c4ab | 2681 | * |
c5aa87bb TH |
2682 | * The worker thread function. All workers belong to a worker_pool - |
2683 | * either a per-cpu one or dynamic unbound one. These workers process all | |
2684 | * work items regardless of their specific target workqueue. The only | |
2685 | * exception is work items which belong to workqueues with a rescuer which | |
2686 | * will be explained in rescuer_thread(). | |
d185af30 YB |
2687 | * |
2688 | * Return: 0 | |
4690c4ab | 2689 | */ |
c34056a3 | 2690 | static int worker_thread(void *__worker) |
1da177e4 | 2691 | { |
c34056a3 | 2692 | struct worker *worker = __worker; |
bd7bdd43 | 2693 | struct worker_pool *pool = worker->pool; |
1da177e4 | 2694 | |
e22bee78 | 2695 | /* tell the scheduler that this is a workqueue worker */ |
197f6acc | 2696 | set_pf_worker(true); |
c8e55f36 | 2697 | woke_up: |
a9b8a985 | 2698 | raw_spin_lock_irq(&pool->lock); |
1da177e4 | 2699 | |
a9ab775b TH |
2700 | /* am I supposed to die? */ |
2701 | if (unlikely(worker->flags & WORKER_DIE)) { | |
a9b8a985 | 2702 | raw_spin_unlock_irq(&pool->lock); |
197f6acc | 2703 | set_pf_worker(false); |
60f5a4bc LJ |
2704 | |
2705 | set_task_comm(worker->task, "kworker/dying"); | |
e441b56f | 2706 | ida_free(&pool->worker_ida, worker->id); |
a2d812a2 | 2707 | worker_detach_from_pool(worker); |
e02b9312 | 2708 | WARN_ON_ONCE(!list_empty(&worker->entry)); |
60f5a4bc | 2709 | kfree(worker); |
a9ab775b | 2710 | return 0; |
c8e55f36 | 2711 | } |
affee4b2 | 2712 | |
c8e55f36 | 2713 | worker_leave_idle(worker); |
db7bccf4 | 2714 | recheck: |
e22bee78 | 2715 | /* no more worker necessary? */ |
63d95a91 | 2716 | if (!need_more_worker(pool)) |
e22bee78 TH |
2717 | goto sleep; |
2718 | ||
2719 | /* do we need to manage? */ | |
63d95a91 | 2720 | if (unlikely(!may_start_working(pool)) && manage_workers(worker)) |
e22bee78 TH |
2721 | goto recheck; |
2722 | ||
c8e55f36 TH |
2723 | /* |
2724 | * ->scheduled list can only be filled while a worker is | |
2725 | * preparing to process a work or actually processing it. | |
2726 | * Make sure nobody diddled with it while I was sleeping. | |
2727 | */ | |
6183c009 | 2728 | WARN_ON_ONCE(!list_empty(&worker->scheduled)); |
c8e55f36 | 2729 | |
e22bee78 | 2730 | /* |
a9ab775b TH |
2731 | * Finish PREP stage. We're guaranteed to have at least one idle |
2732 | * worker or that someone else has already assumed the manager | |
2733 | * role. This is where @worker starts participating in concurrency | |
2734 | * management if applicable and concurrency management is restored | |
2735 | * after being rebound. See rebind_workers() for details. | |
e22bee78 | 2736 | */ |
a9ab775b | 2737 | worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); |
e22bee78 TH |
2738 | |
2739 | do { | |
c8e55f36 | 2740 | struct work_struct *work = |
bd7bdd43 | 2741 | list_first_entry(&pool->worklist, |
c8e55f36 TH |
2742 | struct work_struct, entry); |
2743 | ||
82607adc TH |
2744 | pool->watchdog_ts = jiffies; |
2745 | ||
c8e55f36 TH |
2746 | if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { |
2747 | /* optimization path, not strictly necessary */ | |
2748 | process_one_work(worker, work); | |
2749 | if (unlikely(!list_empty(&worker->scheduled))) | |
affee4b2 | 2750 | process_scheduled_works(worker); |
c8e55f36 TH |
2751 | } else { |
2752 | move_linked_works(work, &worker->scheduled, NULL); | |
2753 | process_scheduled_works(worker); | |
affee4b2 | 2754 | } |
63d95a91 | 2755 | } while (keep_working(pool)); |
e22bee78 | 2756 | |
228f1d00 | 2757 | worker_set_flags(worker, WORKER_PREP); |
d313dd85 | 2758 | sleep: |
c8e55f36 | 2759 | /* |
d565ed63 TH |
2760 | * pool->lock is held and there's no work to process and no need to |
2761 | * manage, sleep. Workers are woken up only while holding | |
2762 | * pool->lock or from local cpu, so setting the current state | |
2763 | * before releasing pool->lock is enough to prevent losing any | |
2764 | * event. | |
c8e55f36 TH |
2765 | */ |
2766 | worker_enter_idle(worker); | |
c5a94a61 | 2767 | __set_current_state(TASK_IDLE); |
a9b8a985 | 2768 | raw_spin_unlock_irq(&pool->lock); |
c8e55f36 TH |
2769 | schedule(); |
2770 | goto woke_up; | |
1da177e4 LT |
2771 | } |
2772 | ||
e22bee78 TH |
2773 | /** |
2774 | * rescuer_thread - the rescuer thread function | |
111c225a | 2775 | * @__rescuer: self |
e22bee78 TH |
2776 | * |
2777 | * Workqueue rescuer thread function. There's one rescuer for each | |
493008a8 | 2778 | * workqueue which has WQ_MEM_RECLAIM set. |
e22bee78 | 2779 | * |
706026c2 | 2780 | * Regular work processing on a pool may block trying to create a new |
e22bee78 TH |
2781 | * worker which uses GFP_KERNEL allocation which has slight chance of |
2782 | * developing into deadlock if some works currently on the same queue | |
2783 | * need to be processed to satisfy the GFP_KERNEL allocation. This is | |
2784 | * the problem rescuer solves. | |
2785 | * | |
706026c2 TH |
2786 | * When such condition is possible, the pool summons rescuers of all |
2787 | * workqueues which have works queued on the pool and let them process | |
e22bee78 TH |
2788 | * those works so that forward progress can be guaranteed. |
2789 | * | |
2790 | * This should happen rarely. | |
d185af30 YB |
2791 | * |
2792 | * Return: 0 | |
e22bee78 | 2793 | */ |
111c225a | 2794 | static int rescuer_thread(void *__rescuer) |
e22bee78 | 2795 | { |
111c225a TH |
2796 | struct worker *rescuer = __rescuer; |
2797 | struct workqueue_struct *wq = rescuer->rescue_wq; | |
e22bee78 | 2798 | struct list_head *scheduled = &rescuer->scheduled; |
4d595b86 | 2799 | bool should_stop; |
e22bee78 TH |
2800 | |
2801 | set_user_nice(current, RESCUER_NICE_LEVEL); | |
111c225a TH |
2802 | |
2803 | /* | |
2804 | * Mark rescuer as worker too. As WORKER_PREP is never cleared, it | |
2805 | * doesn't participate in concurrency management. | |
2806 | */ | |
197f6acc | 2807 | set_pf_worker(true); |
e22bee78 | 2808 | repeat: |
c5a94a61 | 2809 | set_current_state(TASK_IDLE); |
e22bee78 | 2810 | |
4d595b86 LJ |
2811 | /* |
2812 | * By the time the rescuer is requested to stop, the workqueue | |
2813 | * shouldn't have any work pending, but @wq->maydays may still have | |
2814 | * pwq(s) queued. This can happen by non-rescuer workers consuming | |
2815 | * all the work items before the rescuer got to them. Go through | |
2816 | * @wq->maydays processing before acting on should_stop so that the | |
2817 | * list is always empty on exit. | |
2818 | */ | |
2819 | should_stop = kthread_should_stop(); | |
e22bee78 | 2820 | |
493a1724 | 2821 | /* see whether any pwq is asking for help */ |
a9b8a985 | 2822 | raw_spin_lock_irq(&wq_mayday_lock); |
493a1724 TH |
2823 | |
2824 | while (!list_empty(&wq->maydays)) { | |
2825 | struct pool_workqueue *pwq = list_first_entry(&wq->maydays, | |
2826 | struct pool_workqueue, mayday_node); | |
112202d9 | 2827 | struct worker_pool *pool = pwq->pool; |
e22bee78 | 2828 | struct work_struct *work, *n; |
82607adc | 2829 | bool first = true; |
e22bee78 TH |
2830 | |
2831 | __set_current_state(TASK_RUNNING); | |
493a1724 TH |
2832 | list_del_init(&pwq->mayday_node); |
2833 | ||
a9b8a985 | 2834 | raw_spin_unlock_irq(&wq_mayday_lock); |
e22bee78 | 2835 | |
51697d39 LJ |
2836 | worker_attach_to_pool(rescuer, pool); |
2837 | ||
a9b8a985 | 2838 | raw_spin_lock_irq(&pool->lock); |
e22bee78 TH |
2839 | |
2840 | /* | |
2841 | * Slurp in all works issued via this workqueue and | |
2842 | * process'em. | |
2843 | */ | |
0479c8c5 | 2844 | WARN_ON_ONCE(!list_empty(scheduled)); |
82607adc TH |
2845 | list_for_each_entry_safe(work, n, &pool->worklist, entry) { |
2846 | if (get_work_pwq(work) == pwq) { | |
2847 | if (first) | |
2848 | pool->watchdog_ts = jiffies; | |
e22bee78 | 2849 | move_linked_works(work, scheduled, &n); |
725e8ec5 | 2850 | pwq->stats[PWQ_STAT_RESCUED]++; |
82607adc TH |
2851 | } |
2852 | first = false; | |
2853 | } | |
e22bee78 | 2854 | |
008847f6 N |
2855 | if (!list_empty(scheduled)) { |
2856 | process_scheduled_works(rescuer); | |
2857 | ||
2858 | /* | |
2859 | * The above execution of rescued work items could | |
2860 | * have created more to rescue through | |
f97a4a1a | 2861 | * pwq_activate_first_inactive() or chained |
008847f6 N |
2862 | * queueing. Let's put @pwq back on mayday list so |
2863 | * that such back-to-back work items, which may be | |
2864 | * being used to relieve memory pressure, don't | |
2865 | * incur MAYDAY_INTERVAL delay inbetween. | |
2866 | */ | |
4f3f4cf3 | 2867 | if (pwq->nr_active && need_to_create_worker(pool)) { |
a9b8a985 | 2868 | raw_spin_lock(&wq_mayday_lock); |
e66b39af TH |
2869 | /* |
2870 | * Queue iff we aren't racing destruction | |
2871 | * and somebody else hasn't queued it already. | |
2872 | */ | |
2873 | if (wq->rescuer && list_empty(&pwq->mayday_node)) { | |
2874 | get_pwq(pwq); | |
2875 | list_add_tail(&pwq->mayday_node, &wq->maydays); | |
2876 | } | |
a9b8a985 | 2877 | raw_spin_unlock(&wq_mayday_lock); |
008847f6 N |
2878 | } |
2879 | } | |
7576958a | 2880 | |
77668c8b LJ |
2881 | /* |
2882 | * Put the reference grabbed by send_mayday(). @pool won't | |
13b1d625 | 2883 | * go away while we're still attached to it. |
77668c8b LJ |
2884 | */ |
2885 | put_pwq(pwq); | |
2886 | ||
7576958a | 2887 | /* |
d8ca83e6 | 2888 | * Leave this pool. If need_more_worker() is %true, notify a |
7576958a TH |
2889 | * regular worker; otherwise, we end up with 0 concurrency |
2890 | * and stalling the execution. | |
2891 | */ | |
d8ca83e6 | 2892 | if (need_more_worker(pool)) |
63d95a91 | 2893 | wake_up_worker(pool); |
7576958a | 2894 | |
a9b8a985 | 2895 | raw_spin_unlock_irq(&pool->lock); |
13b1d625 | 2896 | |
a2d812a2 | 2897 | worker_detach_from_pool(rescuer); |
13b1d625 | 2898 | |
a9b8a985 | 2899 | raw_spin_lock_irq(&wq_mayday_lock); |
e22bee78 TH |
2900 | } |
2901 | ||
a9b8a985 | 2902 | raw_spin_unlock_irq(&wq_mayday_lock); |
493a1724 | 2903 | |
4d595b86 LJ |
2904 | if (should_stop) { |
2905 | __set_current_state(TASK_RUNNING); | |
197f6acc | 2906 | set_pf_worker(false); |
4d595b86 LJ |
2907 | return 0; |
2908 | } | |
2909 | ||
111c225a TH |
2910 | /* rescuers should never participate in concurrency management */ |
2911 | WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); | |
e22bee78 TH |
2912 | schedule(); |
2913 | goto repeat; | |
1da177e4 LT |
2914 | } |
2915 | ||
fca839c0 TH |
2916 | /** |
2917 | * check_flush_dependency - check for flush dependency sanity | |
2918 | * @target_wq: workqueue being flushed | |
2919 | * @target_work: work item being flushed (NULL for workqueue flushes) | |
2920 | * | |
2921 | * %current is trying to flush the whole @target_wq or @target_work on it. | |
2922 | * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not | |
2923 | * reclaiming memory or running on a workqueue which doesn't have | |
2924 | * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to | |
2925 | * a deadlock. | |
2926 | */ | |
2927 | static void check_flush_dependency(struct workqueue_struct *target_wq, | |
2928 | struct work_struct *target_work) | |
2929 | { | |
2930 | work_func_t target_func = target_work ? target_work->func : NULL; | |
2931 | struct worker *worker; | |
2932 | ||
2933 | if (target_wq->flags & WQ_MEM_RECLAIM) | |
2934 | return; | |
2935 | ||
2936 | worker = current_wq_worker(); | |
2937 | ||
2938 | WARN_ONCE(current->flags & PF_MEMALLOC, | |
d75f773c | 2939 | "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps", |
fca839c0 | 2940 | current->pid, current->comm, target_wq->name, target_func); |
23d11a58 TH |
2941 | WARN_ONCE(worker && ((worker->current_pwq->wq->flags & |
2942 | (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM), | |
d75f773c | 2943 | "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps", |
fca839c0 TH |
2944 | worker->current_pwq->wq->name, worker->current_func, |
2945 | target_wq->name, target_func); | |
2946 | } | |
2947 | ||
fc2e4d70 ON |
2948 | struct wq_barrier { |
2949 | struct work_struct work; | |
2950 | struct completion done; | |
2607d7a6 | 2951 | struct task_struct *task; /* purely informational */ |
fc2e4d70 ON |
2952 | }; |
2953 | ||
2954 | static void wq_barrier_func(struct work_struct *work) | |
2955 | { | |
2956 | struct wq_barrier *barr = container_of(work, struct wq_barrier, work); | |
2957 | complete(&barr->done); | |
2958 | } | |
2959 | ||
4690c4ab TH |
2960 | /** |
2961 | * insert_wq_barrier - insert a barrier work | |
112202d9 | 2962 | * @pwq: pwq to insert barrier into |
4690c4ab | 2963 | * @barr: wq_barrier to insert |
affee4b2 TH |
2964 | * @target: target work to attach @barr to |
2965 | * @worker: worker currently executing @target, NULL if @target is not executing | |
4690c4ab | 2966 | * |
affee4b2 TH |
2967 | * @barr is linked to @target such that @barr is completed only after |
2968 | * @target finishes execution. Please note that the ordering | |
2969 | * guarantee is observed only with respect to @target and on the local | |
2970 | * cpu. | |
2971 | * | |
2972 | * Currently, a queued barrier can't be canceled. This is because | |
2973 | * try_to_grab_pending() can't determine whether the work to be | |
2974 | * grabbed is at the head of the queue and thus can't clear LINKED | |
2975 | * flag of the previous work while there must be a valid next work | |
2976 | * after a work with LINKED flag set. | |
2977 | * | |
2978 | * Note that when @worker is non-NULL, @target may be modified | |
112202d9 | 2979 | * underneath us, so we can't reliably determine pwq from @target. |
4690c4ab TH |
2980 | * |
2981 | * CONTEXT: | |
a9b8a985 | 2982 | * raw_spin_lock_irq(pool->lock). |
4690c4ab | 2983 | */ |
112202d9 | 2984 | static void insert_wq_barrier(struct pool_workqueue *pwq, |
affee4b2 TH |
2985 | struct wq_barrier *barr, |
2986 | struct work_struct *target, struct worker *worker) | |
fc2e4d70 | 2987 | { |
d812796e LJ |
2988 | unsigned int work_flags = 0; |
2989 | unsigned int work_color; | |
affee4b2 | 2990 | struct list_head *head; |
affee4b2 | 2991 | |
dc186ad7 | 2992 | /* |
d565ed63 | 2993 | * debugobject calls are safe here even with pool->lock locked |
dc186ad7 TG |
2994 | * as we know for sure that this will not trigger any of the |
2995 | * checks and call back into the fixup functions where we | |
2996 | * might deadlock. | |
2997 | */ | |
ca1cab37 | 2998 | INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); |
22df02bb | 2999 | __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); |
52fa5bc5 | 3000 | |
fd1a5b04 BP |
3001 | init_completion_map(&barr->done, &target->lockdep_map); |
3002 | ||
2607d7a6 | 3003 | barr->task = current; |
83c22520 | 3004 | |
018f3a13 LJ |
3005 | /* The barrier work item does not participate in pwq->nr_active. */ |
3006 | work_flags |= WORK_STRUCT_INACTIVE; | |
3007 | ||
affee4b2 TH |
3008 | /* |
3009 | * If @target is currently being executed, schedule the | |
3010 | * barrier to the worker; otherwise, put it after @target. | |
3011 | */ | |
d812796e | 3012 | if (worker) { |
affee4b2 | 3013 | head = worker->scheduled.next; |
d812796e LJ |
3014 | work_color = worker->current_color; |
3015 | } else { | |
affee4b2 TH |
3016 | unsigned long *bits = work_data_bits(target); |
3017 | ||
3018 | head = target->entry.next; | |
3019 | /* there can already be other linked works, inherit and set */ | |
d21cece0 | 3020 | work_flags |= *bits & WORK_STRUCT_LINKED; |
d812796e | 3021 | work_color = get_work_color(*bits); |
affee4b2 TH |
3022 | __set_bit(WORK_STRUCT_LINKED_BIT, bits); |
3023 | } | |
3024 | ||
d812796e LJ |
3025 | pwq->nr_in_flight[work_color]++; |
3026 | work_flags |= work_color_to_flags(work_color); | |
3027 | ||
dc186ad7 | 3028 | debug_work_activate(&barr->work); |
d21cece0 | 3029 | insert_work(pwq, &barr->work, head, work_flags); |
fc2e4d70 ON |
3030 | } |
3031 | ||
73f53c4a | 3032 | /** |
112202d9 | 3033 | * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing |
73f53c4a TH |
3034 | * @wq: workqueue being flushed |
3035 | * @flush_color: new flush color, < 0 for no-op | |
3036 | * @work_color: new work color, < 0 for no-op | |
3037 | * | |
112202d9 | 3038 | * Prepare pwqs for workqueue flushing. |
73f53c4a | 3039 | * |
112202d9 TH |
3040 | * If @flush_color is non-negative, flush_color on all pwqs should be |
3041 | * -1. If no pwq has in-flight commands at the specified color, all | |
3042 | * pwq->flush_color's stay at -1 and %false is returned. If any pwq | |
3043 | * has in flight commands, its pwq->flush_color is set to | |
3044 | * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq | |
73f53c4a TH |
3045 | * wakeup logic is armed and %true is returned. |
3046 | * | |
3047 | * The caller should have initialized @wq->first_flusher prior to | |
3048 | * calling this function with non-negative @flush_color. If | |
3049 | * @flush_color is negative, no flush color update is done and %false | |
3050 | * is returned. | |
3051 | * | |
112202d9 | 3052 | * If @work_color is non-negative, all pwqs should have the same |
73f53c4a TH |
3053 | * work_color which is previous to @work_color and all will be |
3054 | * advanced to @work_color. | |
3055 | * | |
3056 | * CONTEXT: | |
3c25a55d | 3057 | * mutex_lock(wq->mutex). |
73f53c4a | 3058 | * |
d185af30 | 3059 | * Return: |
73f53c4a TH |
3060 | * %true if @flush_color >= 0 and there's something to flush. %false |
3061 | * otherwise. | |
3062 | */ | |
112202d9 | 3063 | static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, |
73f53c4a | 3064 | int flush_color, int work_color) |
1da177e4 | 3065 | { |
73f53c4a | 3066 | bool wait = false; |
49e3cf44 | 3067 | struct pool_workqueue *pwq; |
1da177e4 | 3068 | |
73f53c4a | 3069 | if (flush_color >= 0) { |
6183c009 | 3070 | WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); |
112202d9 | 3071 | atomic_set(&wq->nr_pwqs_to_flush, 1); |
1da177e4 | 3072 | } |
2355b70f | 3073 | |
49e3cf44 | 3074 | for_each_pwq(pwq, wq) { |
112202d9 | 3075 | struct worker_pool *pool = pwq->pool; |
fc2e4d70 | 3076 | |
a9b8a985 | 3077 | raw_spin_lock_irq(&pool->lock); |
83c22520 | 3078 | |
73f53c4a | 3079 | if (flush_color >= 0) { |
6183c009 | 3080 | WARN_ON_ONCE(pwq->flush_color != -1); |
fc2e4d70 | 3081 | |
112202d9 TH |
3082 | if (pwq->nr_in_flight[flush_color]) { |
3083 | pwq->flush_color = flush_color; | |
3084 | atomic_inc(&wq->nr_pwqs_to_flush); | |
73f53c4a TH |
3085 | wait = true; |
3086 | } | |
3087 | } | |
1da177e4 | 3088 | |
73f53c4a | 3089 | if (work_color >= 0) { |
6183c009 | 3090 | WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); |
112202d9 | 3091 | pwq->work_color = work_color; |
73f53c4a | 3092 | } |
1da177e4 | 3093 | |
a9b8a985 | 3094 | raw_spin_unlock_irq(&pool->lock); |
1da177e4 | 3095 | } |
2355b70f | 3096 | |
112202d9 | 3097 | if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) |
73f53c4a | 3098 | complete(&wq->first_flusher->done); |
14441960 | 3099 | |
73f53c4a | 3100 | return wait; |
1da177e4 LT |
3101 | } |
3102 | ||
0fcb78c2 | 3103 | /** |
c4f135d6 | 3104 | * __flush_workqueue - ensure that any scheduled work has run to completion. |
0fcb78c2 | 3105 | * @wq: workqueue to flush |
1da177e4 | 3106 | * |
c5aa87bb TH |
3107 | * This function sleeps until all work items which were queued on entry |
3108 | * have finished execution, but it is not livelocked by new incoming ones. | |
1da177e4 | 3109 | */ |
c4f135d6 | 3110 | void __flush_workqueue(struct workqueue_struct *wq) |
1da177e4 | 3111 | { |
73f53c4a TH |
3112 | struct wq_flusher this_flusher = { |
3113 | .list = LIST_HEAD_INIT(this_flusher.list), | |
3114 | .flush_color = -1, | |
fd1a5b04 | 3115 | .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map), |
73f53c4a TH |
3116 | }; |
3117 | int next_color; | |
1da177e4 | 3118 | |
3347fa09 TH |
3119 | if (WARN_ON(!wq_online)) |
3120 | return; | |
3121 | ||
87915adc JB |
3122 | lock_map_acquire(&wq->lockdep_map); |
3123 | lock_map_release(&wq->lockdep_map); | |
3124 | ||
3c25a55d | 3125 | mutex_lock(&wq->mutex); |
73f53c4a TH |
3126 | |
3127 | /* | |
3128 | * Start-to-wait phase | |
3129 | */ | |
3130 | next_color = work_next_color(wq->work_color); | |
3131 | ||
3132 | if (next_color != wq->flush_color) { | |
3133 | /* | |
3134 | * Color space is not full. The current work_color | |
3135 | * becomes our flush_color and work_color is advanced | |
3136 | * by one. | |
3137 | */ | |
6183c009 | 3138 | WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); |
73f53c4a TH |
3139 | this_flusher.flush_color = wq->work_color; |
3140 | wq->work_color = next_color; | |
3141 | ||
3142 | if (!wq->first_flusher) { | |
3143 | /* no flush in progress, become the first flusher */ | |
6183c009 | 3144 | WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); |
73f53c4a TH |
3145 | |
3146 | wq->first_flusher = &this_flusher; | |
3147 | ||
112202d9 | 3148 | if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, |
73f53c4a TH |
3149 | wq->work_color)) { |
3150 | /* nothing to flush, done */ | |
3151 | wq->flush_color = next_color; | |
3152 | wq->first_flusher = NULL; | |
3153 | goto out_unlock; | |
3154 | } | |
3155 | } else { | |
3156 | /* wait in queue */ | |
6183c009 | 3157 | WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); |
73f53c4a | 3158 | list_add_tail(&this_flusher.list, &wq->flusher_queue); |
112202d9 | 3159 | flush_workqueue_prep_pwqs(wq, -1, wq->work_color); |
73f53c4a TH |
3160 | } |
3161 | } else { | |
3162 | /* | |
3163 | * Oops, color space is full, wait on overflow queue. | |
3164 | * The next flush completion will assign us | |
3165 | * flush_color and transfer to flusher_queue. | |
3166 | */ | |
3167 | list_add_tail(&this_flusher.list, &wq->flusher_overflow); | |
3168 | } | |
3169 | ||
fca839c0 TH |
3170 | check_flush_dependency(wq, NULL); |
3171 | ||
3c25a55d | 3172 | mutex_unlock(&wq->mutex); |
73f53c4a TH |
3173 | |
3174 | wait_for_completion(&this_flusher.done); | |
3175 | ||
3176 | /* | |
3177 | * Wake-up-and-cascade phase | |
3178 | * | |
3179 | * First flushers are responsible for cascading flushes and | |
3180 | * handling overflow. Non-first flushers can simply return. | |
3181 | */ | |
00d5d15b | 3182 | if (READ_ONCE(wq->first_flusher) != &this_flusher) |
73f53c4a TH |
3183 | return; |
3184 | ||
3c25a55d | 3185 | mutex_lock(&wq->mutex); |
73f53c4a | 3186 | |
4ce48b37 TH |
3187 | /* we might have raced, check again with mutex held */ |
3188 | if (wq->first_flusher != &this_flusher) | |
3189 | goto out_unlock; | |
3190 | ||
00d5d15b | 3191 | WRITE_ONCE(wq->first_flusher, NULL); |
73f53c4a | 3192 | |
6183c009 TH |
3193 | WARN_ON_ONCE(!list_empty(&this_flusher.list)); |
3194 | WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); | |
73f53c4a TH |
3195 | |
3196 | while (true) { | |
3197 | struct wq_flusher *next, *tmp; | |
3198 | ||
3199 | /* complete all the flushers sharing the current flush color */ | |
3200 | list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { | |
3201 | if (next->flush_color != wq->flush_color) | |
3202 | break; | |
3203 | list_del_init(&next->list); | |
3204 | complete(&next->done); | |
3205 | } | |
3206 | ||
6183c009 TH |
3207 | WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && |
3208 | wq->flush_color != work_next_color(wq->work_color)); | |
73f53c4a TH |
3209 | |
3210 | /* this flush_color is finished, advance by one */ | |
3211 | wq->flush_color = work_next_color(wq->flush_color); | |
3212 | ||
3213 | /* one color has been freed, handle overflow queue */ | |
3214 | if (!list_empty(&wq->flusher_overflow)) { | |
3215 | /* | |
3216 | * Assign the same color to all overflowed | |
3217 | * flushers, advance work_color and append to | |
3218 | * flusher_queue. This is the start-to-wait | |
3219 | * phase for these overflowed flushers. | |
3220 | */ | |
3221 | list_for_each_entry(tmp, &wq->flusher_overflow, list) | |
3222 | tmp->flush_color = wq->work_color; | |
3223 | ||
3224 | wq->work_color = work_next_color(wq->work_color); | |
3225 | ||
3226 | list_splice_tail_init(&wq->flusher_overflow, | |
3227 | &wq->flusher_queue); | |
112202d9 | 3228 | flush_workqueue_prep_pwqs(wq, -1, wq->work_color); |
73f53c4a TH |
3229 | } |
3230 | ||
3231 | if (list_empty(&wq->flusher_queue)) { | |
6183c009 | 3232 | WARN_ON_ONCE(wq->flush_color != wq->work_color); |
73f53c4a TH |
3233 | break; |
3234 | } | |
3235 | ||
3236 | /* | |
3237 | * Need to flush more colors. Make the next flusher | |
112202d9 | 3238 | * the new first flusher and arm pwqs. |
73f53c4a | 3239 | */ |
6183c009 TH |
3240 | WARN_ON_ONCE(wq->flush_color == wq->work_color); |
3241 | WARN_ON_ONCE(wq->flush_color != next->flush_color); | |
73f53c4a TH |
3242 | |
3243 | list_del_init(&next->list); | |
3244 | wq->first_flusher = next; | |
3245 | ||
112202d9 | 3246 | if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) |
73f53c4a TH |
3247 | break; |
3248 | ||
3249 | /* | |
3250 | * Meh... this color is already done, clear first | |
3251 | * flusher and repeat cascading. | |
3252 | */ | |
3253 | wq->first_flusher = NULL; | |
3254 | } | |
3255 | ||
3256 | out_unlock: | |
3c25a55d | 3257 | mutex_unlock(&wq->mutex); |
1da177e4 | 3258 | } |
c4f135d6 | 3259 | EXPORT_SYMBOL(__flush_workqueue); |
1da177e4 | 3260 | |
9c5a2ba7 TH |
3261 | /** |
3262 | * drain_workqueue - drain a workqueue | |
3263 | * @wq: workqueue to drain | |
3264 | * | |
3265 | * Wait until the workqueue becomes empty. While draining is in progress, | |
3266 | * only chain queueing is allowed. IOW, only currently pending or running | |
3267 | * work items on @wq can queue further work items on it. @wq is flushed | |
b749b1b6 | 3268 | * repeatedly until it becomes empty. The number of flushing is determined |
9c5a2ba7 TH |
3269 | * by the depth of chaining and should be relatively short. Whine if it |
3270 | * takes too long. | |
3271 | */ | |
3272 | void drain_workqueue(struct workqueue_struct *wq) | |
3273 | { | |
3274 | unsigned int flush_cnt = 0; | |
49e3cf44 | 3275 | struct pool_workqueue *pwq; |
9c5a2ba7 TH |
3276 | |
3277 | /* | |
3278 | * __queue_work() needs to test whether there are drainers, is much | |
3279 | * hotter than drain_workqueue() and already looks at @wq->flags. | |
618b01eb | 3280 | * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. |
9c5a2ba7 | 3281 | */ |
87fc741e | 3282 | mutex_lock(&wq->mutex); |
9c5a2ba7 | 3283 | if (!wq->nr_drainers++) |
618b01eb | 3284 | wq->flags |= __WQ_DRAINING; |
87fc741e | 3285 | mutex_unlock(&wq->mutex); |
9c5a2ba7 | 3286 | reflush: |
c4f135d6 | 3287 | __flush_workqueue(wq); |
9c5a2ba7 | 3288 | |
b09f4fd3 | 3289 | mutex_lock(&wq->mutex); |
76af4d93 | 3290 | |
49e3cf44 | 3291 | for_each_pwq(pwq, wq) { |
fa2563e4 | 3292 | bool drained; |
9c5a2ba7 | 3293 | |
a9b8a985 | 3294 | raw_spin_lock_irq(&pwq->pool->lock); |
f97a4a1a | 3295 | drained = !pwq->nr_active && list_empty(&pwq->inactive_works); |
a9b8a985 | 3296 | raw_spin_unlock_irq(&pwq->pool->lock); |
fa2563e4 TT |
3297 | |
3298 | if (drained) | |
9c5a2ba7 TH |
3299 | continue; |
3300 | ||
3301 | if (++flush_cnt == 10 || | |
3302 | (flush_cnt % 100 == 0 && flush_cnt <= 1000)) | |
e9ad2eb3 SZ |
3303 | pr_warn("workqueue %s: %s() isn't complete after %u tries\n", |
3304 | wq->name, __func__, flush_cnt); | |
76af4d93 | 3305 | |
b09f4fd3 | 3306 | mutex_unlock(&wq->mutex); |
9c5a2ba7 TH |
3307 | goto reflush; |
3308 | } | |
3309 | ||
9c5a2ba7 | 3310 | if (!--wq->nr_drainers) |
618b01eb | 3311 | wq->flags &= ~__WQ_DRAINING; |
87fc741e | 3312 | mutex_unlock(&wq->mutex); |
9c5a2ba7 TH |
3313 | } |
3314 | EXPORT_SYMBOL_GPL(drain_workqueue); | |
3315 | ||
d6e89786 JB |
3316 | static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr, |
3317 | bool from_cancel) | |
db700897 | 3318 | { |
affee4b2 | 3319 | struct worker *worker = NULL; |
c9e7cf27 | 3320 | struct worker_pool *pool; |
112202d9 | 3321 | struct pool_workqueue *pwq; |
db700897 ON |
3322 | |
3323 | might_sleep(); | |
fa1b54e6 | 3324 | |
24acfb71 | 3325 | rcu_read_lock(); |
c9e7cf27 | 3326 | pool = get_work_pool(work); |
fa1b54e6 | 3327 | if (!pool) { |
24acfb71 | 3328 | rcu_read_unlock(); |
baf59022 | 3329 | return false; |
fa1b54e6 | 3330 | } |
db700897 | 3331 | |
a9b8a985 | 3332 | raw_spin_lock_irq(&pool->lock); |
0b3dae68 | 3333 | /* see the comment in try_to_grab_pending() with the same code */ |
112202d9 TH |
3334 | pwq = get_work_pwq(work); |
3335 | if (pwq) { | |
3336 | if (unlikely(pwq->pool != pool)) | |
4690c4ab | 3337 | goto already_gone; |
606a5020 | 3338 | } else { |
c9e7cf27 | 3339 | worker = find_worker_executing_work(pool, work); |
affee4b2 | 3340 | if (!worker) |
4690c4ab | 3341 | goto already_gone; |
112202d9 | 3342 | pwq = worker->current_pwq; |
606a5020 | 3343 | } |
db700897 | 3344 | |
fca839c0 TH |
3345 | check_flush_dependency(pwq->wq, work); |
3346 | ||
112202d9 | 3347 | insert_wq_barrier(pwq, barr, work, worker); |
a9b8a985 | 3348 | raw_spin_unlock_irq(&pool->lock); |
7a22ad75 | 3349 | |
e159489b | 3350 | /* |
a1d14934 PZ |
3351 | * Force a lock recursion deadlock when using flush_work() inside a |
3352 | * single-threaded or rescuer equipped workqueue. | |
3353 | * | |
3354 | * For single threaded workqueues the deadlock happens when the work | |
3355 | * is after the work issuing the flush_work(). For rescuer equipped | |
3356 | * workqueues the deadlock happens when the rescuer stalls, blocking | |
3357 | * forward progress. | |
e159489b | 3358 | */ |
d6e89786 JB |
3359 | if (!from_cancel && |
3360 | (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) { | |
112202d9 | 3361 | lock_map_acquire(&pwq->wq->lockdep_map); |
a1d14934 PZ |
3362 | lock_map_release(&pwq->wq->lockdep_map); |
3363 | } | |
24acfb71 | 3364 | rcu_read_unlock(); |
401a8d04 | 3365 | return true; |
4690c4ab | 3366 | already_gone: |
a9b8a985 | 3367 | raw_spin_unlock_irq(&pool->lock); |
24acfb71 | 3368 | rcu_read_unlock(); |
401a8d04 | 3369 | return false; |
db700897 | 3370 | } |
baf59022 | 3371 | |
d6e89786 JB |
3372 | static bool __flush_work(struct work_struct *work, bool from_cancel) |
3373 | { | |
3374 | struct wq_barrier barr; | |
3375 | ||
3376 | if (WARN_ON(!wq_online)) | |
3377 | return false; | |
3378 | ||
4d43d395 TH |
3379 | if (WARN_ON(!work->func)) |
3380 | return false; | |
3381 | ||
c0feea59 TH |
3382 | lock_map_acquire(&work->lockdep_map); |
3383 | lock_map_release(&work->lockdep_map); | |
87915adc | 3384 | |
d6e89786 JB |
3385 | if (start_flush_work(work, &barr, from_cancel)) { |
3386 | wait_for_completion(&barr.done); | |
3387 | destroy_work_on_stack(&barr.work); | |
3388 | return true; | |
3389 | } else { | |
3390 | return false; | |
3391 | } | |
3392 | } | |
3393 | ||
baf59022 TH |
3394 | /** |
3395 | * flush_work - wait for a work to finish executing the last queueing instance | |
3396 | * @work: the work to flush | |
3397 | * | |
606a5020 TH |
3398 | * Wait until @work has finished execution. @work is guaranteed to be idle |
3399 | * on return if it hasn't been requeued since flush started. | |
baf59022 | 3400 | * |
d185af30 | 3401 | * Return: |
baf59022 TH |
3402 | * %true if flush_work() waited for the work to finish execution, |
3403 | * %false if it was already idle. | |
3404 | */ | |
3405 | bool flush_work(struct work_struct *work) | |
3406 | { | |
d6e89786 | 3407 | return __flush_work(work, false); |
6e84d644 | 3408 | } |
606a5020 | 3409 | EXPORT_SYMBOL_GPL(flush_work); |
6e84d644 | 3410 | |
8603e1b3 | 3411 | struct cwt_wait { |
ac6424b9 | 3412 | wait_queue_entry_t wait; |
8603e1b3 TH |
3413 | struct work_struct *work; |
3414 | }; | |
3415 | ||
ac6424b9 | 3416 | static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key) |
8603e1b3 TH |
3417 | { |
3418 | struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait); | |
3419 | ||
3420 | if (cwait->work != key) | |
3421 | return 0; | |
3422 | return autoremove_wake_function(wait, mode, sync, key); | |
3423 | } | |
3424 | ||
36e227d2 | 3425 | static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) |
1f1f642e | 3426 | { |
8603e1b3 | 3427 | static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq); |
bbb68dfa | 3428 | unsigned long flags; |
1f1f642e ON |
3429 | int ret; |
3430 | ||
3431 | do { | |
bbb68dfa TH |
3432 | ret = try_to_grab_pending(work, is_dwork, &flags); |
3433 | /* | |
8603e1b3 TH |
3434 | * If someone else is already canceling, wait for it to |
3435 | * finish. flush_work() doesn't work for PREEMPT_NONE | |
3436 | * because we may get scheduled between @work's completion | |
3437 | * and the other canceling task resuming and clearing | |
3438 | * CANCELING - flush_work() will return false immediately | |
3439 | * as @work is no longer busy, try_to_grab_pending() will | |
3440 | * return -ENOENT as @work is still being canceled and the | |
3441 | * other canceling task won't be able to clear CANCELING as | |
3442 | * we're hogging the CPU. | |
3443 | * | |
3444 | * Let's wait for completion using a waitqueue. As this | |
3445 | * may lead to the thundering herd problem, use a custom | |
3446 | * wake function which matches @work along with exclusive | |
3447 | * wait and wakeup. | |
bbb68dfa | 3448 | */ |
8603e1b3 TH |
3449 | if (unlikely(ret == -ENOENT)) { |
3450 | struct cwt_wait cwait; | |
3451 | ||
3452 | init_wait(&cwait.wait); | |
3453 | cwait.wait.func = cwt_wakefn; | |
3454 | cwait.work = work; | |
3455 | ||
3456 | prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait, | |
3457 | TASK_UNINTERRUPTIBLE); | |
3458 | if (work_is_canceling(work)) | |
3459 | schedule(); | |
3460 | finish_wait(&cancel_waitq, &cwait.wait); | |
3461 | } | |
1f1f642e ON |
3462 | } while (unlikely(ret < 0)); |
3463 | ||
bbb68dfa TH |
3464 | /* tell other tasks trying to grab @work to back off */ |
3465 | mark_work_canceling(work); | |
3466 | local_irq_restore(flags); | |
3467 | ||
3347fa09 TH |
3468 | /* |
3469 | * This allows canceling during early boot. We know that @work | |
3470 | * isn't executing. | |
3471 | */ | |
3472 | if (wq_online) | |
d6e89786 | 3473 | __flush_work(work, true); |
3347fa09 | 3474 | |
7a22ad75 | 3475 | clear_work_data(work); |
8603e1b3 TH |
3476 | |
3477 | /* | |
3478 | * Paired with prepare_to_wait() above so that either | |
3479 | * waitqueue_active() is visible here or !work_is_canceling() is | |
3480 | * visible there. | |
3481 | */ | |
3482 | smp_mb(); | |
3483 | if (waitqueue_active(&cancel_waitq)) | |
3484 | __wake_up(&cancel_waitq, TASK_NORMAL, 1, work); | |
3485 | ||
1f1f642e ON |
3486 | return ret; |
3487 | } | |
3488 | ||
6e84d644 | 3489 | /** |
401a8d04 TH |
3490 | * cancel_work_sync - cancel a work and wait for it to finish |
3491 | * @work: the work to cancel | |
6e84d644 | 3492 | * |
401a8d04 TH |
3493 | * Cancel @work and wait for its execution to finish. This function |
3494 | * can be used even if the work re-queues itself or migrates to | |
3495 | * another workqueue. On return from this function, @work is | |
3496 | * guaranteed to be not pending or executing on any CPU. | |
1f1f642e | 3497 | * |
401a8d04 TH |
3498 | * cancel_work_sync(&delayed_work->work) must not be used for |
3499 | * delayed_work's. Use cancel_delayed_work_sync() instead. | |
6e84d644 | 3500 | * |
401a8d04 | 3501 | * The caller must ensure that the workqueue on which @work was last |
6e84d644 | 3502 | * queued can't be destroyed before this function returns. |
401a8d04 | 3503 | * |
d185af30 | 3504 | * Return: |
401a8d04 | 3505 | * %true if @work was pending, %false otherwise. |
6e84d644 | 3506 | */ |
401a8d04 | 3507 | bool cancel_work_sync(struct work_struct *work) |
6e84d644 | 3508 | { |
36e227d2 | 3509 | return __cancel_work_timer(work, false); |
b89deed3 | 3510 | } |
28e53bdd | 3511 | EXPORT_SYMBOL_GPL(cancel_work_sync); |
b89deed3 | 3512 | |
6e84d644 | 3513 | /** |
401a8d04 TH |
3514 | * flush_delayed_work - wait for a dwork to finish executing the last queueing |
3515 | * @dwork: the delayed work to flush | |
6e84d644 | 3516 | * |
401a8d04 TH |
3517 | * Delayed timer is cancelled and the pending work is queued for |
3518 | * immediate execution. Like flush_work(), this function only | |
3519 | * considers the last queueing instance of @dwork. | |
1f1f642e | 3520 | * |
d185af30 | 3521 | * Return: |
401a8d04 TH |
3522 | * %true if flush_work() waited for the work to finish execution, |
3523 | * %false if it was already idle. | |
6e84d644 | 3524 | */ |
401a8d04 TH |
3525 | bool flush_delayed_work(struct delayed_work *dwork) |
3526 | { | |
8930caba | 3527 | local_irq_disable(); |
401a8d04 | 3528 | if (del_timer_sync(&dwork->timer)) |
60c057bc | 3529 | __queue_work(dwork->cpu, dwork->wq, &dwork->work); |
8930caba | 3530 | local_irq_enable(); |
401a8d04 TH |
3531 | return flush_work(&dwork->work); |
3532 | } | |
3533 | EXPORT_SYMBOL(flush_delayed_work); | |
3534 | ||
05f0fe6b TH |
3535 | /** |
3536 | * flush_rcu_work - wait for a rwork to finish executing the last queueing | |
3537 | * @rwork: the rcu work to flush | |
3538 | * | |
3539 | * Return: | |
3540 | * %true if flush_rcu_work() waited for the work to finish execution, | |
3541 | * %false if it was already idle. | |
3542 | */ | |
3543 | bool flush_rcu_work(struct rcu_work *rwork) | |
3544 | { | |
3545 | if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) { | |
3546 | rcu_barrier(); | |
3547 | flush_work(&rwork->work); | |
3548 | return true; | |
3549 | } else { | |
3550 | return flush_work(&rwork->work); | |
3551 | } | |
3552 | } | |
3553 | EXPORT_SYMBOL(flush_rcu_work); | |
3554 | ||
f72b8792 JA |
3555 | static bool __cancel_work(struct work_struct *work, bool is_dwork) |
3556 | { | |
3557 | unsigned long flags; | |
3558 | int ret; | |
3559 | ||
3560 | do { | |
3561 | ret = try_to_grab_pending(work, is_dwork, &flags); | |
3562 | } while (unlikely(ret == -EAGAIN)); | |
3563 | ||
3564 | if (unlikely(ret < 0)) | |
3565 | return false; | |
3566 | ||
3567 | set_work_pool_and_clear_pending(work, get_work_pool_id(work)); | |
3568 | local_irq_restore(flags); | |
3569 | return ret; | |
3570 | } | |
3571 | ||
73b4b532 AG |
3572 | /* |
3573 | * See cancel_delayed_work() | |
3574 | */ | |
3575 | bool cancel_work(struct work_struct *work) | |
3576 | { | |
3577 | return __cancel_work(work, false); | |
3578 | } | |
3579 | EXPORT_SYMBOL(cancel_work); | |
3580 | ||
09383498 | 3581 | /** |
57b30ae7 TH |
3582 | * cancel_delayed_work - cancel a delayed work |
3583 | * @dwork: delayed_work to cancel | |
09383498 | 3584 | * |
d185af30 YB |
3585 | * Kill off a pending delayed_work. |
3586 | * | |
3587 | * Return: %true if @dwork was pending and canceled; %false if it wasn't | |
3588 | * pending. | |
3589 | * | |
3590 | * Note: | |
3591 | * The work callback function may still be running on return, unless | |
3592 | * it returns %true and the work doesn't re-arm itself. Explicitly flush or | |
3593 | * use cancel_delayed_work_sync() to wait on it. | |
09383498 | 3594 | * |
57b30ae7 | 3595 | * This function is safe to call from any context including IRQ handler. |
09383498 | 3596 | */ |
57b30ae7 | 3597 | bool cancel_delayed_work(struct delayed_work *dwork) |
09383498 | 3598 | { |
f72b8792 | 3599 | return __cancel_work(&dwork->work, true); |
09383498 | 3600 | } |
57b30ae7 | 3601 | EXPORT_SYMBOL(cancel_delayed_work); |
09383498 | 3602 | |
401a8d04 TH |
3603 | /** |
3604 | * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish | |
3605 | * @dwork: the delayed work cancel | |
3606 | * | |
3607 | * This is cancel_work_sync() for delayed works. | |
3608 | * | |
d185af30 | 3609 | * Return: |
401a8d04 TH |
3610 | * %true if @dwork was pending, %false otherwise. |
3611 | */ | |
3612 | bool cancel_delayed_work_sync(struct delayed_work *dwork) | |
6e84d644 | 3613 | { |
36e227d2 | 3614 | return __cancel_work_timer(&dwork->work, true); |
6e84d644 | 3615 | } |
f5a421a4 | 3616 | EXPORT_SYMBOL(cancel_delayed_work_sync); |
1da177e4 | 3617 | |
b6136773 | 3618 | /** |
31ddd871 | 3619 | * schedule_on_each_cpu - execute a function synchronously on each online CPU |
b6136773 | 3620 | * @func: the function to call |
b6136773 | 3621 | * |
31ddd871 TH |
3622 | * schedule_on_each_cpu() executes @func on each online CPU using the |
3623 | * system workqueue and blocks until all CPUs have completed. | |
b6136773 | 3624 | * schedule_on_each_cpu() is very slow. |
31ddd871 | 3625 | * |
d185af30 | 3626 | * Return: |
31ddd871 | 3627 | * 0 on success, -errno on failure. |
b6136773 | 3628 | */ |
65f27f38 | 3629 | int schedule_on_each_cpu(work_func_t func) |
15316ba8 CL |
3630 | { |
3631 | int cpu; | |
38f51568 | 3632 | struct work_struct __percpu *works; |
15316ba8 | 3633 | |
b6136773 AM |
3634 | works = alloc_percpu(struct work_struct); |
3635 | if (!works) | |
15316ba8 | 3636 | return -ENOMEM; |
b6136773 | 3637 | |
ffd8bea8 | 3638 | cpus_read_lock(); |
93981800 | 3639 | |
15316ba8 | 3640 | for_each_online_cpu(cpu) { |
9bfb1839 IM |
3641 | struct work_struct *work = per_cpu_ptr(works, cpu); |
3642 | ||
3643 | INIT_WORK(work, func); | |
b71ab8c2 | 3644 | schedule_work_on(cpu, work); |
65a64464 | 3645 | } |
93981800 TH |
3646 | |
3647 | for_each_online_cpu(cpu) | |
3648 | flush_work(per_cpu_ptr(works, cpu)); | |
3649 | ||
ffd8bea8 | 3650 | cpus_read_unlock(); |
b6136773 | 3651 | free_percpu(works); |
15316ba8 CL |
3652 | return 0; |
3653 | } | |
3654 | ||
1fa44eca JB |
3655 | /** |
3656 | * execute_in_process_context - reliably execute the routine with user context | |
3657 | * @fn: the function to execute | |
1fa44eca JB |
3658 | * @ew: guaranteed storage for the execute work structure (must |
3659 | * be available when the work executes) | |
3660 | * | |
3661 | * Executes the function immediately if process context is available, | |
3662 | * otherwise schedules the function for delayed execution. | |
3663 | * | |
d185af30 | 3664 | * Return: 0 - function was executed |
1fa44eca JB |
3665 | * 1 - function was scheduled for execution |
3666 | */ | |
65f27f38 | 3667 | int execute_in_process_context(work_func_t fn, struct execute_work *ew) |
1fa44eca JB |
3668 | { |
3669 | if (!in_interrupt()) { | |
65f27f38 | 3670 | fn(&ew->work); |
1fa44eca JB |
3671 | return 0; |
3672 | } | |
3673 | ||
65f27f38 | 3674 | INIT_WORK(&ew->work, fn); |
1fa44eca JB |
3675 | schedule_work(&ew->work); |
3676 | ||
3677 | return 1; | |
3678 | } | |
3679 | EXPORT_SYMBOL_GPL(execute_in_process_context); | |
3680 | ||
6ba94429 FW |
3681 | /** |
3682 | * free_workqueue_attrs - free a workqueue_attrs | |
3683 | * @attrs: workqueue_attrs to free | |
226223ab | 3684 | * |
6ba94429 | 3685 | * Undo alloc_workqueue_attrs(). |
226223ab | 3686 | */ |
513c98d0 | 3687 | void free_workqueue_attrs(struct workqueue_attrs *attrs) |
226223ab | 3688 | { |
6ba94429 FW |
3689 | if (attrs) { |
3690 | free_cpumask_var(attrs->cpumask); | |
3691 | kfree(attrs); | |
3692 | } | |
226223ab TH |
3693 | } |
3694 | ||
6ba94429 FW |
3695 | /** |
3696 | * alloc_workqueue_attrs - allocate a workqueue_attrs | |
6ba94429 FW |
3697 | * |
3698 | * Allocate a new workqueue_attrs, initialize with default settings and | |
3699 | * return it. | |
3700 | * | |
3701 | * Return: The allocated new workqueue_attr on success. %NULL on failure. | |
3702 | */ | |
513c98d0 | 3703 | struct workqueue_attrs *alloc_workqueue_attrs(void) |
226223ab | 3704 | { |
6ba94429 | 3705 | struct workqueue_attrs *attrs; |
226223ab | 3706 | |
be69d00d | 3707 | attrs = kzalloc(sizeof(*attrs), GFP_KERNEL); |
6ba94429 FW |
3708 | if (!attrs) |
3709 | goto fail; | |
be69d00d | 3710 | if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL)) |
6ba94429 FW |
3711 | goto fail; |
3712 | ||
3713 | cpumask_copy(attrs->cpumask, cpu_possible_mask); | |
3714 | return attrs; | |
3715 | fail: | |
3716 | free_workqueue_attrs(attrs); | |
3717 | return NULL; | |
226223ab TH |
3718 | } |
3719 | ||
6ba94429 FW |
3720 | static void copy_workqueue_attrs(struct workqueue_attrs *to, |
3721 | const struct workqueue_attrs *from) | |
226223ab | 3722 | { |
6ba94429 FW |
3723 | to->nice = from->nice; |
3724 | cpumask_copy(to->cpumask, from->cpumask); | |
3725 | /* | |
3726 | * Unlike hash and equality test, this function doesn't ignore | |
3727 | * ->no_numa as it is used for both pool and wq attrs. Instead, | |
3728 | * get_unbound_pool() explicitly clears ->no_numa after copying. | |
3729 | */ | |
3730 | to->no_numa = from->no_numa; | |
226223ab TH |
3731 | } |
3732 | ||
6ba94429 FW |
3733 | /* hash value of the content of @attr */ |
3734 | static u32 wqattrs_hash(const struct workqueue_attrs *attrs) | |
226223ab | 3735 | { |
6ba94429 | 3736 | u32 hash = 0; |
226223ab | 3737 | |
6ba94429 FW |
3738 | hash = jhash_1word(attrs->nice, hash); |
3739 | hash = jhash(cpumask_bits(attrs->cpumask), | |
3740 | BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); | |
3741 | return hash; | |
226223ab | 3742 | } |
226223ab | 3743 | |
6ba94429 FW |
3744 | /* content equality test */ |
3745 | static bool wqattrs_equal(const struct workqueue_attrs *a, | |
3746 | const struct workqueue_attrs *b) | |
226223ab | 3747 | { |
6ba94429 FW |
3748 | if (a->nice != b->nice) |
3749 | return false; | |
3750 | if (!cpumask_equal(a->cpumask, b->cpumask)) | |
3751 | return false; | |
3752 | return true; | |
226223ab TH |
3753 | } |
3754 | ||
6ba94429 FW |
3755 | /** |
3756 | * init_worker_pool - initialize a newly zalloc'd worker_pool | |
3757 | * @pool: worker_pool to initialize | |
3758 | * | |
402dd89d | 3759 | * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs. |
6ba94429 FW |
3760 | * |
3761 | * Return: 0 on success, -errno on failure. Even on failure, all fields | |
3762 | * inside @pool proper are initialized and put_unbound_pool() can be called | |
3763 | * on @pool safely to release it. | |
3764 | */ | |
3765 | static int init_worker_pool(struct worker_pool *pool) | |
226223ab | 3766 | { |
a9b8a985 | 3767 | raw_spin_lock_init(&pool->lock); |
6ba94429 FW |
3768 | pool->id = -1; |
3769 | pool->cpu = -1; | |
3770 | pool->node = NUMA_NO_NODE; | |
3771 | pool->flags |= POOL_DISASSOCIATED; | |
82607adc | 3772 | pool->watchdog_ts = jiffies; |
6ba94429 FW |
3773 | INIT_LIST_HEAD(&pool->worklist); |
3774 | INIT_LIST_HEAD(&pool->idle_list); | |
3775 | hash_init(pool->busy_hash); | |
226223ab | 3776 | |
32a6c723 | 3777 | timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE); |
3f959aa3 | 3778 | INIT_WORK(&pool->idle_cull_work, idle_cull_fn); |
226223ab | 3779 | |
32a6c723 | 3780 | timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0); |
226223ab | 3781 | |
6ba94429 | 3782 | INIT_LIST_HEAD(&pool->workers); |
e02b9312 | 3783 | INIT_LIST_HEAD(&pool->dying_workers); |
226223ab | 3784 | |
6ba94429 FW |
3785 | ida_init(&pool->worker_ida); |
3786 | INIT_HLIST_NODE(&pool->hash_node); | |
3787 | pool->refcnt = 1; | |
226223ab | 3788 | |
6ba94429 | 3789 | /* shouldn't fail above this point */ |
be69d00d | 3790 | pool->attrs = alloc_workqueue_attrs(); |
6ba94429 FW |
3791 | if (!pool->attrs) |
3792 | return -ENOMEM; | |
3793 | return 0; | |
226223ab TH |
3794 | } |
3795 | ||
669de8bd BVA |
3796 | #ifdef CONFIG_LOCKDEP |
3797 | static void wq_init_lockdep(struct workqueue_struct *wq) | |
3798 | { | |
3799 | char *lock_name; | |
3800 | ||
3801 | lockdep_register_key(&wq->key); | |
3802 | lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name); | |
3803 | if (!lock_name) | |
3804 | lock_name = wq->name; | |
69a106c0 QC |
3805 | |
3806 | wq->lock_name = lock_name; | |
669de8bd BVA |
3807 | lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0); |
3808 | } | |
3809 | ||
3810 | static void wq_unregister_lockdep(struct workqueue_struct *wq) | |
3811 | { | |
3812 | lockdep_unregister_key(&wq->key); | |
3813 | } | |
3814 | ||
3815 | static void wq_free_lockdep(struct workqueue_struct *wq) | |
3816 | { | |
3817 | if (wq->lock_name != wq->name) | |
3818 | kfree(wq->lock_name); | |
3819 | } | |
3820 | #else | |
3821 | static void wq_init_lockdep(struct workqueue_struct *wq) | |
3822 | { | |
3823 | } | |
3824 | ||
3825 | static void wq_unregister_lockdep(struct workqueue_struct *wq) | |
3826 | { | |
3827 | } | |
3828 | ||
3829 | static void wq_free_lockdep(struct workqueue_struct *wq) | |
3830 | { | |
3831 | } | |
3832 | #endif | |
3833 | ||
6ba94429 | 3834 | static void rcu_free_wq(struct rcu_head *rcu) |
226223ab | 3835 | { |
6ba94429 FW |
3836 | struct workqueue_struct *wq = |
3837 | container_of(rcu, struct workqueue_struct, rcu); | |
226223ab | 3838 | |
669de8bd BVA |
3839 | wq_free_lockdep(wq); |
3840 | ||
6ba94429 FW |
3841 | if (!(wq->flags & WQ_UNBOUND)) |
3842 | free_percpu(wq->cpu_pwqs); | |
226223ab | 3843 | else |
6ba94429 | 3844 | free_workqueue_attrs(wq->unbound_attrs); |
226223ab | 3845 | |
6ba94429 | 3846 | kfree(wq); |
226223ab TH |
3847 | } |
3848 | ||
6ba94429 | 3849 | static void rcu_free_pool(struct rcu_head *rcu) |
226223ab | 3850 | { |
6ba94429 | 3851 | struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); |
226223ab | 3852 | |
6ba94429 FW |
3853 | ida_destroy(&pool->worker_ida); |
3854 | free_workqueue_attrs(pool->attrs); | |
3855 | kfree(pool); | |
226223ab TH |
3856 | } |
3857 | ||
6ba94429 FW |
3858 | /** |
3859 | * put_unbound_pool - put a worker_pool | |
3860 | * @pool: worker_pool to put | |
3861 | * | |
24acfb71 | 3862 | * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU |
6ba94429 FW |
3863 | * safe manner. get_unbound_pool() calls this function on its failure path |
3864 | * and this function should be able to release pools which went through, | |
3865 | * successfully or not, init_worker_pool(). | |
3866 | * | |
3867 | * Should be called with wq_pool_mutex held. | |
3868 | */ | |
3869 | static void put_unbound_pool(struct worker_pool *pool) | |
226223ab | 3870 | { |
6ba94429 | 3871 | DECLARE_COMPLETION_ONSTACK(detach_completion); |
e02b9312 | 3872 | struct list_head cull_list; |
6ba94429 | 3873 | struct worker *worker; |
226223ab | 3874 | |
e02b9312 VS |
3875 | INIT_LIST_HEAD(&cull_list); |
3876 | ||
6ba94429 | 3877 | lockdep_assert_held(&wq_pool_mutex); |
226223ab | 3878 | |
6ba94429 FW |
3879 | if (--pool->refcnt) |
3880 | return; | |
226223ab | 3881 | |
6ba94429 FW |
3882 | /* sanity checks */ |
3883 | if (WARN_ON(!(pool->cpu < 0)) || | |
3884 | WARN_ON(!list_empty(&pool->worklist))) | |
3885 | return; | |
226223ab | 3886 | |
6ba94429 FW |
3887 | /* release id and unhash */ |
3888 | if (pool->id >= 0) | |
3889 | idr_remove(&worker_pool_idr, pool->id); | |
3890 | hash_del(&pool->hash_node); | |
d55262c4 | 3891 | |
6ba94429 | 3892 | /* |
692b4825 TH |
3893 | * Become the manager and destroy all workers. This prevents |
3894 | * @pool's workers from blocking on attach_mutex. We're the last | |
3895 | * manager and @pool gets freed with the flag set. | |
9ab03be4 VS |
3896 | * |
3897 | * Having a concurrent manager is quite unlikely to happen as we can | |
3898 | * only get here with | |
3899 | * pwq->refcnt == pool->refcnt == 0 | |
3900 | * which implies no work queued to the pool, which implies no worker can | |
3901 | * become the manager. However a worker could have taken the role of | |
3902 | * manager before the refcnts dropped to 0, since maybe_create_worker() | |
3903 | * drops pool->lock | |
6ba94429 | 3904 | */ |
9ab03be4 VS |
3905 | while (true) { |
3906 | rcuwait_wait_event(&manager_wait, | |
3907 | !(pool->flags & POOL_MANAGER_ACTIVE), | |
3908 | TASK_UNINTERRUPTIBLE); | |
e02b9312 VS |
3909 | |
3910 | mutex_lock(&wq_pool_attach_mutex); | |
9ab03be4 VS |
3911 | raw_spin_lock_irq(&pool->lock); |
3912 | if (!(pool->flags & POOL_MANAGER_ACTIVE)) { | |
3913 | pool->flags |= POOL_MANAGER_ACTIVE; | |
3914 | break; | |
3915 | } | |
3916 | raw_spin_unlock_irq(&pool->lock); | |
e02b9312 | 3917 | mutex_unlock(&wq_pool_attach_mutex); |
9ab03be4 | 3918 | } |
692b4825 | 3919 | |
6ba94429 | 3920 | while ((worker = first_idle_worker(pool))) |
e02b9312 | 3921 | set_worker_dying(worker, &cull_list); |
6ba94429 | 3922 | WARN_ON(pool->nr_workers || pool->nr_idle); |
a9b8a985 | 3923 | raw_spin_unlock_irq(&pool->lock); |
d55262c4 | 3924 | |
e02b9312 VS |
3925 | wake_dying_workers(&cull_list); |
3926 | ||
3927 | if (!list_empty(&pool->workers) || !list_empty(&pool->dying_workers)) | |
6ba94429 | 3928 | pool->detach_completion = &detach_completion; |
1258fae7 | 3929 | mutex_unlock(&wq_pool_attach_mutex); |
226223ab | 3930 | |
6ba94429 FW |
3931 | if (pool->detach_completion) |
3932 | wait_for_completion(pool->detach_completion); | |
226223ab | 3933 | |
6ba94429 FW |
3934 | /* shut down the timers */ |
3935 | del_timer_sync(&pool->idle_timer); | |
3f959aa3 | 3936 | cancel_work_sync(&pool->idle_cull_work); |
6ba94429 | 3937 | del_timer_sync(&pool->mayday_timer); |
226223ab | 3938 | |
24acfb71 | 3939 | /* RCU protected to allow dereferences from get_work_pool() */ |
25b00775 | 3940 | call_rcu(&pool->rcu, rcu_free_pool); |
226223ab TH |
3941 | } |
3942 | ||
3943 | /** | |
6ba94429 FW |
3944 | * get_unbound_pool - get a worker_pool with the specified attributes |
3945 | * @attrs: the attributes of the worker_pool to get | |
226223ab | 3946 | * |
6ba94429 FW |
3947 | * Obtain a worker_pool which has the same attributes as @attrs, bump the |
3948 | * reference count and return it. If there already is a matching | |
3949 | * worker_pool, it will be used; otherwise, this function attempts to | |
3950 | * create a new one. | |
226223ab | 3951 | * |
6ba94429 | 3952 | * Should be called with wq_pool_mutex held. |
226223ab | 3953 | * |
6ba94429 FW |
3954 | * Return: On success, a worker_pool with the same attributes as @attrs. |
3955 | * On failure, %NULL. | |
226223ab | 3956 | */ |
6ba94429 | 3957 | static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) |
226223ab | 3958 | { |
6ba94429 FW |
3959 | u32 hash = wqattrs_hash(attrs); |
3960 | struct worker_pool *pool; | |
3961 | int node; | |
e2273584 | 3962 | int target_node = NUMA_NO_NODE; |
226223ab | 3963 | |
6ba94429 | 3964 | lockdep_assert_held(&wq_pool_mutex); |
226223ab | 3965 | |
6ba94429 FW |
3966 | /* do we already have a matching pool? */ |
3967 | hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { | |
3968 | if (wqattrs_equal(pool->attrs, attrs)) { | |
3969 | pool->refcnt++; | |
3970 | return pool; | |
3971 | } | |
3972 | } | |
226223ab | 3973 | |
e2273584 XP |
3974 | /* if cpumask is contained inside a NUMA node, we belong to that node */ |
3975 | if (wq_numa_enabled) { | |
3976 | for_each_node(node) { | |
3977 | if (cpumask_subset(attrs->cpumask, | |
3978 | wq_numa_possible_cpumask[node])) { | |
3979 | target_node = node; | |
3980 | break; | |
3981 | } | |
3982 | } | |
3983 | } | |
3984 | ||
6ba94429 | 3985 | /* nope, create a new one */ |
e2273584 | 3986 | pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node); |
6ba94429 FW |
3987 | if (!pool || init_worker_pool(pool) < 0) |
3988 | goto fail; | |
3989 | ||
3990 | lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */ | |
3991 | copy_workqueue_attrs(pool->attrs, attrs); | |
e2273584 | 3992 | pool->node = target_node; |
226223ab TH |
3993 | |
3994 | /* | |
6ba94429 FW |
3995 | * no_numa isn't a worker_pool attribute, always clear it. See |
3996 | * 'struct workqueue_attrs' comments for detail. | |
226223ab | 3997 | */ |
6ba94429 | 3998 | pool->attrs->no_numa = false; |
226223ab | 3999 | |
6ba94429 FW |
4000 | if (worker_pool_assign_id(pool) < 0) |
4001 | goto fail; | |
226223ab | 4002 | |
6ba94429 | 4003 | /* create and start the initial worker */ |
3347fa09 | 4004 | if (wq_online && !create_worker(pool)) |
6ba94429 | 4005 | goto fail; |
226223ab | 4006 | |
6ba94429 FW |
4007 | /* install */ |
4008 | hash_add(unbound_pool_hash, &pool->hash_node, hash); | |
226223ab | 4009 | |
6ba94429 FW |
4010 | return pool; |
4011 | fail: | |
4012 | if (pool) | |
4013 | put_unbound_pool(pool); | |
4014 | return NULL; | |
226223ab | 4015 | } |
226223ab | 4016 | |
6ba94429 | 4017 | static void rcu_free_pwq(struct rcu_head *rcu) |
7a4e344c | 4018 | { |
6ba94429 FW |
4019 | kmem_cache_free(pwq_cache, |
4020 | container_of(rcu, struct pool_workqueue, rcu)); | |
7a4e344c TH |
4021 | } |
4022 | ||
6ba94429 FW |
4023 | /* |
4024 | * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt | |
4025 | * and needs to be destroyed. | |
7a4e344c | 4026 | */ |
6ba94429 | 4027 | static void pwq_unbound_release_workfn(struct work_struct *work) |
7a4e344c | 4028 | { |
6ba94429 FW |
4029 | struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, |
4030 | unbound_release_work); | |
4031 | struct workqueue_struct *wq = pwq->wq; | |
4032 | struct worker_pool *pool = pwq->pool; | |
b42b0bdd | 4033 | bool is_last = false; |
7a4e344c | 4034 | |
b42b0bdd YY |
4035 | /* |
4036 | * when @pwq is not linked, it doesn't hold any reference to the | |
4037 | * @wq, and @wq is invalid to access. | |
4038 | */ | |
4039 | if (!list_empty(&pwq->pwqs_node)) { | |
4040 | if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND))) | |
4041 | return; | |
7a4e344c | 4042 | |
b42b0bdd YY |
4043 | mutex_lock(&wq->mutex); |
4044 | list_del_rcu(&pwq->pwqs_node); | |
4045 | is_last = list_empty(&wq->pwqs); | |
4046 | mutex_unlock(&wq->mutex); | |
4047 | } | |
6ba94429 FW |
4048 | |
4049 | mutex_lock(&wq_pool_mutex); | |
4050 | put_unbound_pool(pool); | |
4051 | mutex_unlock(&wq_pool_mutex); | |
4052 | ||
25b00775 | 4053 | call_rcu(&pwq->rcu, rcu_free_pwq); |
7a4e344c | 4054 | |
2865a8fb | 4055 | /* |
6ba94429 FW |
4056 | * If we're the last pwq going away, @wq is already dead and no one |
4057 | * is gonna access it anymore. Schedule RCU free. | |
2865a8fb | 4058 | */ |
669de8bd BVA |
4059 | if (is_last) { |
4060 | wq_unregister_lockdep(wq); | |
25b00775 | 4061 | call_rcu(&wq->rcu, rcu_free_wq); |
669de8bd | 4062 | } |
29c91e99 TH |
4063 | } |
4064 | ||
7a4e344c | 4065 | /** |
6ba94429 FW |
4066 | * pwq_adjust_max_active - update a pwq's max_active to the current setting |
4067 | * @pwq: target pool_workqueue | |
d185af30 | 4068 | * |
6ba94429 | 4069 | * If @pwq isn't freezing, set @pwq->max_active to the associated |
f97a4a1a | 4070 | * workqueue's saved_max_active and activate inactive work items |
6ba94429 | 4071 | * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. |
7a4e344c | 4072 | */ |
6ba94429 | 4073 | static void pwq_adjust_max_active(struct pool_workqueue *pwq) |
4e1a1f9a | 4074 | { |
6ba94429 FW |
4075 | struct workqueue_struct *wq = pwq->wq; |
4076 | bool freezable = wq->flags & WQ_FREEZABLE; | |
3347fa09 | 4077 | unsigned long flags; |
4e1a1f9a | 4078 | |
6ba94429 FW |
4079 | /* for @wq->saved_max_active */ |
4080 | lockdep_assert_held(&wq->mutex); | |
4e1a1f9a | 4081 | |
6ba94429 FW |
4082 | /* fast exit for non-freezable wqs */ |
4083 | if (!freezable && pwq->max_active == wq->saved_max_active) | |
4084 | return; | |
7a4e344c | 4085 | |
3347fa09 | 4086 | /* this function can be called during early boot w/ irq disabled */ |
a9b8a985 | 4087 | raw_spin_lock_irqsave(&pwq->pool->lock, flags); |
29c91e99 | 4088 | |
6ba94429 FW |
4089 | /* |
4090 | * During [un]freezing, the caller is responsible for ensuring that | |
4091 | * this function is called at least once after @workqueue_freezing | |
4092 | * is updated and visible. | |
4093 | */ | |
4094 | if (!freezable || !workqueue_freezing) { | |
01341fbd YY |
4095 | bool kick = false; |
4096 | ||
6ba94429 | 4097 | pwq->max_active = wq->saved_max_active; |
4e1a1f9a | 4098 | |
f97a4a1a | 4099 | while (!list_empty(&pwq->inactive_works) && |
01341fbd | 4100 | pwq->nr_active < pwq->max_active) { |
f97a4a1a | 4101 | pwq_activate_first_inactive(pwq); |
01341fbd YY |
4102 | kick = true; |
4103 | } | |
e2dca7ad | 4104 | |
6ba94429 FW |
4105 | /* |
4106 | * Need to kick a worker after thawed or an unbound wq's | |
01341fbd YY |
4107 | * max_active is bumped. In realtime scenarios, always kicking a |
4108 | * worker will cause interference on the isolated cpu cores, so | |
4109 | * let's kick iff work items were activated. | |
6ba94429 | 4110 | */ |
01341fbd YY |
4111 | if (kick) |
4112 | wake_up_worker(pwq->pool); | |
6ba94429 FW |
4113 | } else { |
4114 | pwq->max_active = 0; | |
4115 | } | |
e2dca7ad | 4116 | |
a9b8a985 | 4117 | raw_spin_unlock_irqrestore(&pwq->pool->lock, flags); |
e2dca7ad TH |
4118 | } |
4119 | ||
67dc8325 | 4120 | /* initialize newly allocated @pwq which is associated with @wq and @pool */ |
6ba94429 FW |
4121 | static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, |
4122 | struct worker_pool *pool) | |
29c91e99 | 4123 | { |
6ba94429 | 4124 | BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); |
29c91e99 | 4125 | |
6ba94429 FW |
4126 | memset(pwq, 0, sizeof(*pwq)); |
4127 | ||
4128 | pwq->pool = pool; | |
4129 | pwq->wq = wq; | |
4130 | pwq->flush_color = -1; | |
4131 | pwq->refcnt = 1; | |
f97a4a1a | 4132 | INIT_LIST_HEAD(&pwq->inactive_works); |
6ba94429 FW |
4133 | INIT_LIST_HEAD(&pwq->pwqs_node); |
4134 | INIT_LIST_HEAD(&pwq->mayday_node); | |
4135 | INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn); | |
29c91e99 TH |
4136 | } |
4137 | ||
6ba94429 FW |
4138 | /* sync @pwq with the current state of its associated wq and link it */ |
4139 | static void link_pwq(struct pool_workqueue *pwq) | |
29c91e99 | 4140 | { |
6ba94429 | 4141 | struct workqueue_struct *wq = pwq->wq; |
29c91e99 | 4142 | |
6ba94429 | 4143 | lockdep_assert_held(&wq->mutex); |
a892cacc | 4144 | |
6ba94429 FW |
4145 | /* may be called multiple times, ignore if already linked */ |
4146 | if (!list_empty(&pwq->pwqs_node)) | |
29c91e99 | 4147 | return; |
29c91e99 | 4148 | |
6ba94429 FW |
4149 | /* set the matching work_color */ |
4150 | pwq->work_color = wq->work_color; | |
29c91e99 | 4151 | |
6ba94429 FW |
4152 | /* sync max_active to the current setting */ |
4153 | pwq_adjust_max_active(pwq); | |
29c91e99 | 4154 | |
6ba94429 FW |
4155 | /* link in @pwq */ |
4156 | list_add_rcu(&pwq->pwqs_node, &wq->pwqs); | |
4157 | } | |
29c91e99 | 4158 | |
6ba94429 FW |
4159 | /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ |
4160 | static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, | |
4161 | const struct workqueue_attrs *attrs) | |
4162 | { | |
4163 | struct worker_pool *pool; | |
4164 | struct pool_workqueue *pwq; | |
60f5a4bc | 4165 | |
6ba94429 | 4166 | lockdep_assert_held(&wq_pool_mutex); |
60f5a4bc | 4167 | |
6ba94429 FW |
4168 | pool = get_unbound_pool(attrs); |
4169 | if (!pool) | |
4170 | return NULL; | |
60f5a4bc | 4171 | |
6ba94429 FW |
4172 | pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); |
4173 | if (!pwq) { | |
4174 | put_unbound_pool(pool); | |
4175 | return NULL; | |
4176 | } | |
29c91e99 | 4177 | |
6ba94429 FW |
4178 | init_pwq(pwq, wq, pool); |
4179 | return pwq; | |
4180 | } | |
29c91e99 | 4181 | |
29c91e99 | 4182 | /** |
30186c6f | 4183 | * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node |
042f7df1 | 4184 | * @attrs: the wq_attrs of the default pwq of the target workqueue |
6ba94429 FW |
4185 | * @node: the target NUMA node |
4186 | * @cpu_going_down: if >= 0, the CPU to consider as offline | |
4187 | * @cpumask: outarg, the resulting cpumask | |
29c91e99 | 4188 | * |
6ba94429 FW |
4189 | * Calculate the cpumask a workqueue with @attrs should use on @node. If |
4190 | * @cpu_going_down is >= 0, that cpu is considered offline during | |
4191 | * calculation. The result is stored in @cpumask. | |
a892cacc | 4192 | * |
6ba94429 FW |
4193 | * If NUMA affinity is not enabled, @attrs->cpumask is always used. If |
4194 | * enabled and @node has online CPUs requested by @attrs, the returned | |
4195 | * cpumask is the intersection of the possible CPUs of @node and | |
4196 | * @attrs->cpumask. | |
d185af30 | 4197 | * |
6ba94429 FW |
4198 | * The caller is responsible for ensuring that the cpumask of @node stays |
4199 | * stable. | |
4200 | * | |
4201 | * Return: %true if the resulting @cpumask is different from @attrs->cpumask, | |
4202 | * %false if equal. | |
29c91e99 | 4203 | */ |
6ba94429 FW |
4204 | static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node, |
4205 | int cpu_going_down, cpumask_t *cpumask) | |
29c91e99 | 4206 | { |
6ba94429 FW |
4207 | if (!wq_numa_enabled || attrs->no_numa) |
4208 | goto use_dfl; | |
29c91e99 | 4209 | |
6ba94429 FW |
4210 | /* does @node have any online CPUs @attrs wants? */ |
4211 | cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask); | |
4212 | if (cpu_going_down >= 0) | |
4213 | cpumask_clear_cpu(cpu_going_down, cpumask); | |
29c91e99 | 4214 | |
6ba94429 FW |
4215 | if (cpumask_empty(cpumask)) |
4216 | goto use_dfl; | |
4c16bd32 TH |
4217 | |
4218 | /* yeap, return possible CPUs in @node that @attrs wants */ | |
4219 | cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]); | |
1ad0f0a7 MB |
4220 | |
4221 | if (cpumask_empty(cpumask)) { | |
4222 | pr_warn_once("WARNING: workqueue cpumask: online intersect > " | |
4223 | "possible intersect\n"); | |
4224 | return false; | |
4225 | } | |
4226 | ||
4c16bd32 TH |
4227 | return !cpumask_equal(cpumask, attrs->cpumask); |
4228 | ||
4229 | use_dfl: | |
4230 | cpumask_copy(cpumask, attrs->cpumask); | |
4231 | return false; | |
4232 | } | |
4233 | ||
1befcf30 TH |
4234 | /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */ |
4235 | static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq, | |
4236 | int node, | |
4237 | struct pool_workqueue *pwq) | |
4238 | { | |
4239 | struct pool_workqueue *old_pwq; | |
4240 | ||
5b95e1af | 4241 | lockdep_assert_held(&wq_pool_mutex); |
1befcf30 TH |
4242 | lockdep_assert_held(&wq->mutex); |
4243 | ||
4244 | /* link_pwq() can handle duplicate calls */ | |
4245 | link_pwq(pwq); | |
4246 | ||
4247 | old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); | |
4248 | rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq); | |
4249 | return old_pwq; | |
4250 | } | |
4251 | ||
2d5f0764 LJ |
4252 | /* context to store the prepared attrs & pwqs before applying */ |
4253 | struct apply_wqattrs_ctx { | |
4254 | struct workqueue_struct *wq; /* target workqueue */ | |
4255 | struct workqueue_attrs *attrs; /* attrs to apply */ | |
042f7df1 | 4256 | struct list_head list; /* queued for batching commit */ |
2d5f0764 LJ |
4257 | struct pool_workqueue *dfl_pwq; |
4258 | struct pool_workqueue *pwq_tbl[]; | |
4259 | }; | |
4260 | ||
4261 | /* free the resources after success or abort */ | |
4262 | static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx) | |
4263 | { | |
4264 | if (ctx) { | |
4265 | int node; | |
4266 | ||
4267 | for_each_node(node) | |
4268 | put_pwq_unlocked(ctx->pwq_tbl[node]); | |
4269 | put_pwq_unlocked(ctx->dfl_pwq); | |
4270 | ||
4271 | free_workqueue_attrs(ctx->attrs); | |
4272 | ||
4273 | kfree(ctx); | |
4274 | } | |
4275 | } | |
4276 | ||
4277 | /* allocate the attrs and pwqs for later installation */ | |
4278 | static struct apply_wqattrs_ctx * | |
4279 | apply_wqattrs_prepare(struct workqueue_struct *wq, | |
99c621ef LJ |
4280 | const struct workqueue_attrs *attrs, |
4281 | const cpumask_var_t unbound_cpumask) | |
9e8cd2f5 | 4282 | { |
2d5f0764 | 4283 | struct apply_wqattrs_ctx *ctx; |
4c16bd32 | 4284 | struct workqueue_attrs *new_attrs, *tmp_attrs; |
2d5f0764 | 4285 | int node; |
9e8cd2f5 | 4286 | |
2d5f0764 | 4287 | lockdep_assert_held(&wq_pool_mutex); |
9e8cd2f5 | 4288 | |
acafe7e3 | 4289 | ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_node_ids), GFP_KERNEL); |
8719dcea | 4290 | |
be69d00d TG |
4291 | new_attrs = alloc_workqueue_attrs(); |
4292 | tmp_attrs = alloc_workqueue_attrs(); | |
2d5f0764 LJ |
4293 | if (!ctx || !new_attrs || !tmp_attrs) |
4294 | goto out_free; | |
13e2e556 | 4295 | |
042f7df1 | 4296 | /* |
99c621ef LJ |
4297 | * Calculate the attrs of the default pwq with unbound_cpumask |
4298 | * which is wq_unbound_cpumask or to set to wq_unbound_cpumask. | |
042f7df1 LJ |
4299 | * If the user configured cpumask doesn't overlap with the |
4300 | * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask. | |
4301 | */ | |
13e2e556 | 4302 | copy_workqueue_attrs(new_attrs, attrs); |
99c621ef | 4303 | cpumask_and(new_attrs->cpumask, new_attrs->cpumask, unbound_cpumask); |
042f7df1 | 4304 | if (unlikely(cpumask_empty(new_attrs->cpumask))) |
99c621ef | 4305 | cpumask_copy(new_attrs->cpumask, unbound_cpumask); |
13e2e556 | 4306 | |
4c16bd32 TH |
4307 | /* |
4308 | * We may create multiple pwqs with differing cpumasks. Make a | |
4309 | * copy of @new_attrs which will be modified and used to obtain | |
4310 | * pools. | |
4311 | */ | |
4312 | copy_workqueue_attrs(tmp_attrs, new_attrs); | |
4313 | ||
4c16bd32 TH |
4314 | /* |
4315 | * If something goes wrong during CPU up/down, we'll fall back to | |
4316 | * the default pwq covering whole @attrs->cpumask. Always create | |
4317 | * it even if we don't use it immediately. | |
4318 | */ | |
2d5f0764 LJ |
4319 | ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs); |
4320 | if (!ctx->dfl_pwq) | |
4321 | goto out_free; | |
4c16bd32 TH |
4322 | |
4323 | for_each_node(node) { | |
042f7df1 | 4324 | if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) { |
2d5f0764 LJ |
4325 | ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs); |
4326 | if (!ctx->pwq_tbl[node]) | |
4327 | goto out_free; | |
4c16bd32 | 4328 | } else { |
2d5f0764 LJ |
4329 | ctx->dfl_pwq->refcnt++; |
4330 | ctx->pwq_tbl[node] = ctx->dfl_pwq; | |
4c16bd32 TH |
4331 | } |
4332 | } | |
4333 | ||
042f7df1 LJ |
4334 | /* save the user configured attrs and sanitize it. */ |
4335 | copy_workqueue_attrs(new_attrs, attrs); | |
4336 | cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); | |
2d5f0764 | 4337 | ctx->attrs = new_attrs; |
042f7df1 | 4338 | |
2d5f0764 LJ |
4339 | ctx->wq = wq; |
4340 | free_workqueue_attrs(tmp_attrs); | |
4341 | return ctx; | |
4342 | ||
4343 | out_free: | |
4344 | free_workqueue_attrs(tmp_attrs); | |
4345 | free_workqueue_attrs(new_attrs); | |
4346 | apply_wqattrs_cleanup(ctx); | |
4347 | return NULL; | |
4348 | } | |
4349 | ||
4350 | /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */ | |
4351 | static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx) | |
4352 | { | |
4353 | int node; | |
9e8cd2f5 | 4354 | |
4c16bd32 | 4355 | /* all pwqs have been created successfully, let's install'em */ |
2d5f0764 | 4356 | mutex_lock(&ctx->wq->mutex); |
a892cacc | 4357 | |
2d5f0764 | 4358 | copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs); |
4c16bd32 TH |
4359 | |
4360 | /* save the previous pwq and install the new one */ | |
f147f29e | 4361 | for_each_node(node) |
2d5f0764 LJ |
4362 | ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node, |
4363 | ctx->pwq_tbl[node]); | |
4c16bd32 TH |
4364 | |
4365 | /* @dfl_pwq might not have been used, ensure it's linked */ | |
2d5f0764 LJ |
4366 | link_pwq(ctx->dfl_pwq); |
4367 | swap(ctx->wq->dfl_pwq, ctx->dfl_pwq); | |
f147f29e | 4368 | |
2d5f0764 LJ |
4369 | mutex_unlock(&ctx->wq->mutex); |
4370 | } | |
9e8cd2f5 | 4371 | |
a0111cf6 LJ |
4372 | static void apply_wqattrs_lock(void) |
4373 | { | |
4374 | /* CPUs should stay stable across pwq creations and installations */ | |
ffd8bea8 | 4375 | cpus_read_lock(); |
a0111cf6 LJ |
4376 | mutex_lock(&wq_pool_mutex); |
4377 | } | |
4378 | ||
4379 | static void apply_wqattrs_unlock(void) | |
4380 | { | |
4381 | mutex_unlock(&wq_pool_mutex); | |
ffd8bea8 | 4382 | cpus_read_unlock(); |
a0111cf6 LJ |
4383 | } |
4384 | ||
4385 | static int apply_workqueue_attrs_locked(struct workqueue_struct *wq, | |
4386 | const struct workqueue_attrs *attrs) | |
2d5f0764 LJ |
4387 | { |
4388 | struct apply_wqattrs_ctx *ctx; | |
4c16bd32 | 4389 | |
2d5f0764 LJ |
4390 | /* only unbound workqueues can change attributes */ |
4391 | if (WARN_ON(!(wq->flags & WQ_UNBOUND))) | |
4392 | return -EINVAL; | |
13e2e556 | 4393 | |
2d5f0764 | 4394 | /* creating multiple pwqs breaks ordering guarantee */ |
0a94efb5 TH |
4395 | if (!list_empty(&wq->pwqs)) { |
4396 | if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) | |
4397 | return -EINVAL; | |
4398 | ||
4399 | wq->flags &= ~__WQ_ORDERED; | |
4400 | } | |
2d5f0764 | 4401 | |
99c621ef | 4402 | ctx = apply_wqattrs_prepare(wq, attrs, wq_unbound_cpumask); |
6201171e | 4403 | if (!ctx) |
4404 | return -ENOMEM; | |
2d5f0764 LJ |
4405 | |
4406 | /* the ctx has been prepared successfully, let's commit it */ | |
6201171e | 4407 | apply_wqattrs_commit(ctx); |
2d5f0764 LJ |
4408 | apply_wqattrs_cleanup(ctx); |
4409 | ||
6201171e | 4410 | return 0; |
9e8cd2f5 TH |
4411 | } |
4412 | ||
a0111cf6 LJ |
4413 | /** |
4414 | * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue | |
4415 | * @wq: the target workqueue | |
4416 | * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() | |
4417 | * | |
4418 | * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA | |
4419 | * machines, this function maps a separate pwq to each NUMA node with | |
4420 | * possibles CPUs in @attrs->cpumask so that work items are affine to the | |
4421 | * NUMA node it was issued on. Older pwqs are released as in-flight work | |
4422 | * items finish. Note that a work item which repeatedly requeues itself | |
4423 | * back-to-back will stay on its current pwq. | |
4424 | * | |
4425 | * Performs GFP_KERNEL allocations. | |
4426 | * | |
ffd8bea8 | 4427 | * Assumes caller has CPU hotplug read exclusion, i.e. cpus_read_lock(). |
509b3204 | 4428 | * |
a0111cf6 LJ |
4429 | * Return: 0 on success and -errno on failure. |
4430 | */ | |
513c98d0 | 4431 | int apply_workqueue_attrs(struct workqueue_struct *wq, |
a0111cf6 LJ |
4432 | const struct workqueue_attrs *attrs) |
4433 | { | |
4434 | int ret; | |
4435 | ||
509b3204 DJ |
4436 | lockdep_assert_cpus_held(); |
4437 | ||
4438 | mutex_lock(&wq_pool_mutex); | |
a0111cf6 | 4439 | ret = apply_workqueue_attrs_locked(wq, attrs); |
509b3204 | 4440 | mutex_unlock(&wq_pool_mutex); |
a0111cf6 LJ |
4441 | |
4442 | return ret; | |
4443 | } | |
4444 | ||
4c16bd32 TH |
4445 | /** |
4446 | * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug | |
4447 | * @wq: the target workqueue | |
4448 | * @cpu: the CPU coming up or going down | |
4449 | * @online: whether @cpu is coming up or going down | |
4450 | * | |
4451 | * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and | |
4452 | * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of | |
4453 | * @wq accordingly. | |
4454 | * | |
4455 | * If NUMA affinity can't be adjusted due to memory allocation failure, it | |
4456 | * falls back to @wq->dfl_pwq which may not be optimal but is always | |
4457 | * correct. | |
4458 | * | |
4459 | * Note that when the last allowed CPU of a NUMA node goes offline for a | |
4460 | * workqueue with a cpumask spanning multiple nodes, the workers which were | |
4461 | * already executing the work items for the workqueue will lose their CPU | |
4462 | * affinity and may execute on any CPU. This is similar to how per-cpu | |
4463 | * workqueues behave on CPU_DOWN. If a workqueue user wants strict | |
4464 | * affinity, it's the user's responsibility to flush the work item from | |
4465 | * CPU_DOWN_PREPARE. | |
4466 | */ | |
4467 | static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu, | |
4468 | bool online) | |
4469 | { | |
4470 | int node = cpu_to_node(cpu); | |
4471 | int cpu_off = online ? -1 : cpu; | |
4472 | struct pool_workqueue *old_pwq = NULL, *pwq; | |
4473 | struct workqueue_attrs *target_attrs; | |
4474 | cpumask_t *cpumask; | |
4475 | ||
4476 | lockdep_assert_held(&wq_pool_mutex); | |
4477 | ||
f7142ed4 LJ |
4478 | if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) || |
4479 | wq->unbound_attrs->no_numa) | |
4c16bd32 TH |
4480 | return; |
4481 | ||
4482 | /* | |
4483 | * We don't wanna alloc/free wq_attrs for each wq for each CPU. | |
4484 | * Let's use a preallocated one. The following buf is protected by | |
4485 | * CPU hotplug exclusion. | |
4486 | */ | |
4487 | target_attrs = wq_update_unbound_numa_attrs_buf; | |
4488 | cpumask = target_attrs->cpumask; | |
4489 | ||
4c16bd32 TH |
4490 | copy_workqueue_attrs(target_attrs, wq->unbound_attrs); |
4491 | pwq = unbound_pwq_by_node(wq, node); | |
4492 | ||
4493 | /* | |
4494 | * Let's determine what needs to be done. If the target cpumask is | |
042f7df1 LJ |
4495 | * different from the default pwq's, we need to compare it to @pwq's |
4496 | * and create a new one if they don't match. If the target cpumask | |
4497 | * equals the default pwq's, the default pwq should be used. | |
4c16bd32 | 4498 | */ |
042f7df1 | 4499 | if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) { |
4c16bd32 | 4500 | if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask)) |
f7142ed4 | 4501 | return; |
4c16bd32 | 4502 | } else { |
534a3fbb | 4503 | goto use_dfl_pwq; |
4c16bd32 TH |
4504 | } |
4505 | ||
4c16bd32 TH |
4506 | /* create a new pwq */ |
4507 | pwq = alloc_unbound_pwq(wq, target_attrs); | |
4508 | if (!pwq) { | |
2d916033 FF |
4509 | pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n", |
4510 | wq->name); | |
77f300b1 | 4511 | goto use_dfl_pwq; |
4c16bd32 TH |
4512 | } |
4513 | ||
f7142ed4 | 4514 | /* Install the new pwq. */ |
4c16bd32 TH |
4515 | mutex_lock(&wq->mutex); |
4516 | old_pwq = numa_pwq_tbl_install(wq, node, pwq); | |
4517 | goto out_unlock; | |
4518 | ||
4519 | use_dfl_pwq: | |
f7142ed4 | 4520 | mutex_lock(&wq->mutex); |
a9b8a985 | 4521 | raw_spin_lock_irq(&wq->dfl_pwq->pool->lock); |
4c16bd32 | 4522 | get_pwq(wq->dfl_pwq); |
a9b8a985 | 4523 | raw_spin_unlock_irq(&wq->dfl_pwq->pool->lock); |
4c16bd32 TH |
4524 | old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq); |
4525 | out_unlock: | |
4526 | mutex_unlock(&wq->mutex); | |
4527 | put_pwq_unlocked(old_pwq); | |
4528 | } | |
4529 | ||
30cdf249 | 4530 | static int alloc_and_link_pwqs(struct workqueue_struct *wq) |
0f900049 | 4531 | { |
49e3cf44 | 4532 | bool highpri = wq->flags & WQ_HIGHPRI; |
8a2b7538 | 4533 | int cpu, ret; |
30cdf249 TH |
4534 | |
4535 | if (!(wq->flags & WQ_UNBOUND)) { | |
420c0ddb TH |
4536 | wq->cpu_pwqs = alloc_percpu(struct pool_workqueue); |
4537 | if (!wq->cpu_pwqs) | |
30cdf249 TH |
4538 | return -ENOMEM; |
4539 | ||
4540 | for_each_possible_cpu(cpu) { | |
7fb98ea7 TH |
4541 | struct pool_workqueue *pwq = |
4542 | per_cpu_ptr(wq->cpu_pwqs, cpu); | |
7a62c2c8 | 4543 | struct worker_pool *cpu_pools = |
f02ae73a | 4544 | per_cpu(cpu_worker_pools, cpu); |
f3421797 | 4545 | |
f147f29e TH |
4546 | init_pwq(pwq, wq, &cpu_pools[highpri]); |
4547 | ||
4548 | mutex_lock(&wq->mutex); | |
1befcf30 | 4549 | link_pwq(pwq); |
f147f29e | 4550 | mutex_unlock(&wq->mutex); |
30cdf249 | 4551 | } |
9e8cd2f5 | 4552 | return 0; |
509b3204 DJ |
4553 | } |
4554 | ||
ffd8bea8 | 4555 | cpus_read_lock(); |
509b3204 | 4556 | if (wq->flags & __WQ_ORDERED) { |
8a2b7538 TH |
4557 | ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]); |
4558 | /* there should only be single pwq for ordering guarantee */ | |
4559 | WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node || | |
4560 | wq->pwqs.prev != &wq->dfl_pwq->pwqs_node), | |
4561 | "ordering guarantee broken for workqueue %s\n", wq->name); | |
30cdf249 | 4562 | } else { |
509b3204 | 4563 | ret = apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); |
30cdf249 | 4564 | } |
ffd8bea8 | 4565 | cpus_read_unlock(); |
509b3204 DJ |
4566 | |
4567 | return ret; | |
0f900049 TH |
4568 | } |
4569 | ||
f3421797 TH |
4570 | static int wq_clamp_max_active(int max_active, unsigned int flags, |
4571 | const char *name) | |
b71ab8c2 | 4572 | { |
f3421797 TH |
4573 | int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; |
4574 | ||
4575 | if (max_active < 1 || max_active > lim) | |
044c782c VI |
4576 | pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", |
4577 | max_active, name, 1, lim); | |
b71ab8c2 | 4578 | |
f3421797 | 4579 | return clamp_val(max_active, 1, lim); |
b71ab8c2 TH |
4580 | } |
4581 | ||
983c7515 TH |
4582 | /* |
4583 | * Workqueues which may be used during memory reclaim should have a rescuer | |
4584 | * to guarantee forward progress. | |
4585 | */ | |
4586 | static int init_rescuer(struct workqueue_struct *wq) | |
4587 | { | |
4588 | struct worker *rescuer; | |
b92b36ea | 4589 | int ret; |
983c7515 TH |
4590 | |
4591 | if (!(wq->flags & WQ_MEM_RECLAIM)) | |
4592 | return 0; | |
4593 | ||
4594 | rescuer = alloc_worker(NUMA_NO_NODE); | |
4c0736a7 PM |
4595 | if (!rescuer) { |
4596 | pr_err("workqueue: Failed to allocate a rescuer for wq \"%s\"\n", | |
4597 | wq->name); | |
983c7515 | 4598 | return -ENOMEM; |
4c0736a7 | 4599 | } |
983c7515 TH |
4600 | |
4601 | rescuer->rescue_wq = wq; | |
4602 | rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", wq->name); | |
f187b697 | 4603 | if (IS_ERR(rescuer->task)) { |
b92b36ea | 4604 | ret = PTR_ERR(rescuer->task); |
4c0736a7 PM |
4605 | pr_err("workqueue: Failed to create a rescuer kthread for wq \"%s\": %pe", |
4606 | wq->name, ERR_PTR(ret)); | |
983c7515 | 4607 | kfree(rescuer); |
b92b36ea | 4608 | return ret; |
983c7515 TH |
4609 | } |
4610 | ||
4611 | wq->rescuer = rescuer; | |
4612 | kthread_bind_mask(rescuer->task, cpu_possible_mask); | |
4613 | wake_up_process(rescuer->task); | |
4614 | ||
4615 | return 0; | |
4616 | } | |
4617 | ||
a2775bbc | 4618 | __printf(1, 4) |
669de8bd BVA |
4619 | struct workqueue_struct *alloc_workqueue(const char *fmt, |
4620 | unsigned int flags, | |
4621 | int max_active, ...) | |
1da177e4 | 4622 | { |
df2d5ae4 | 4623 | size_t tbl_size = 0; |
ecf6881f | 4624 | va_list args; |
1da177e4 | 4625 | struct workqueue_struct *wq; |
49e3cf44 | 4626 | struct pool_workqueue *pwq; |
b196be89 | 4627 | |
5c0338c6 TH |
4628 | /* |
4629 | * Unbound && max_active == 1 used to imply ordered, which is no | |
4630 | * longer the case on NUMA machines due to per-node pools. While | |
4631 | * alloc_ordered_workqueue() is the right way to create an ordered | |
4632 | * workqueue, keep the previous behavior to avoid subtle breakages | |
4633 | * on NUMA. | |
4634 | */ | |
4635 | if ((flags & WQ_UNBOUND) && max_active == 1) | |
4636 | flags |= __WQ_ORDERED; | |
4637 | ||
cee22a15 VK |
4638 | /* see the comment above the definition of WQ_POWER_EFFICIENT */ |
4639 | if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) | |
4640 | flags |= WQ_UNBOUND; | |
4641 | ||
ecf6881f | 4642 | /* allocate wq and format name */ |
df2d5ae4 | 4643 | if (flags & WQ_UNBOUND) |
ddcb57e2 | 4644 | tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]); |
df2d5ae4 TH |
4645 | |
4646 | wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL); | |
b196be89 | 4647 | if (!wq) |
d2c1d404 | 4648 | return NULL; |
b196be89 | 4649 | |
6029a918 | 4650 | if (flags & WQ_UNBOUND) { |
be69d00d | 4651 | wq->unbound_attrs = alloc_workqueue_attrs(); |
6029a918 TH |
4652 | if (!wq->unbound_attrs) |
4653 | goto err_free_wq; | |
4654 | } | |
4655 | ||
669de8bd | 4656 | va_start(args, max_active); |
ecf6881f | 4657 | vsnprintf(wq->name, sizeof(wq->name), fmt, args); |
b196be89 | 4658 | va_end(args); |
1da177e4 | 4659 | |
d320c038 | 4660 | max_active = max_active ?: WQ_DFL_ACTIVE; |
b196be89 | 4661 | max_active = wq_clamp_max_active(max_active, flags, wq->name); |
3af24433 | 4662 | |
b196be89 | 4663 | /* init wq */ |
97e37d7b | 4664 | wq->flags = flags; |
a0a1a5fd | 4665 | wq->saved_max_active = max_active; |
3c25a55d | 4666 | mutex_init(&wq->mutex); |
112202d9 | 4667 | atomic_set(&wq->nr_pwqs_to_flush, 0); |
30cdf249 | 4668 | INIT_LIST_HEAD(&wq->pwqs); |
73f53c4a TH |
4669 | INIT_LIST_HEAD(&wq->flusher_queue); |
4670 | INIT_LIST_HEAD(&wq->flusher_overflow); | |
493a1724 | 4671 | INIT_LIST_HEAD(&wq->maydays); |
502ca9d8 | 4672 | |
669de8bd | 4673 | wq_init_lockdep(wq); |
cce1a165 | 4674 | INIT_LIST_HEAD(&wq->list); |
3af24433 | 4675 | |
30cdf249 | 4676 | if (alloc_and_link_pwqs(wq) < 0) |
82efcab3 | 4677 | goto err_unreg_lockdep; |
1537663f | 4678 | |
40c17f75 | 4679 | if (wq_online && init_rescuer(wq) < 0) |
983c7515 | 4680 | goto err_destroy; |
3af24433 | 4681 | |
226223ab TH |
4682 | if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) |
4683 | goto err_destroy; | |
4684 | ||
a0a1a5fd | 4685 | /* |
68e13a67 LJ |
4686 | * wq_pool_mutex protects global freeze state and workqueues list. |
4687 | * Grab it, adjust max_active and add the new @wq to workqueues | |
4688 | * list. | |
a0a1a5fd | 4689 | */ |
68e13a67 | 4690 | mutex_lock(&wq_pool_mutex); |
a0a1a5fd | 4691 | |
a357fc03 | 4692 | mutex_lock(&wq->mutex); |
699ce097 TH |
4693 | for_each_pwq(pwq, wq) |
4694 | pwq_adjust_max_active(pwq); | |
a357fc03 | 4695 | mutex_unlock(&wq->mutex); |
a0a1a5fd | 4696 | |
e2dca7ad | 4697 | list_add_tail_rcu(&wq->list, &workqueues); |
a0a1a5fd | 4698 | |
68e13a67 | 4699 | mutex_unlock(&wq_pool_mutex); |
1537663f | 4700 | |
3af24433 | 4701 | return wq; |
d2c1d404 | 4702 | |
82efcab3 | 4703 | err_unreg_lockdep: |
009bb421 BVA |
4704 | wq_unregister_lockdep(wq); |
4705 | wq_free_lockdep(wq); | |
82efcab3 | 4706 | err_free_wq: |
6029a918 | 4707 | free_workqueue_attrs(wq->unbound_attrs); |
d2c1d404 TH |
4708 | kfree(wq); |
4709 | return NULL; | |
4710 | err_destroy: | |
4711 | destroy_workqueue(wq); | |
4690c4ab | 4712 | return NULL; |
3af24433 | 4713 | } |
669de8bd | 4714 | EXPORT_SYMBOL_GPL(alloc_workqueue); |
1da177e4 | 4715 | |
c29eb853 TH |
4716 | static bool pwq_busy(struct pool_workqueue *pwq) |
4717 | { | |
4718 | int i; | |
4719 | ||
4720 | for (i = 0; i < WORK_NR_COLORS; i++) | |
4721 | if (pwq->nr_in_flight[i]) | |
4722 | return true; | |
4723 | ||
4724 | if ((pwq != pwq->wq->dfl_pwq) && (pwq->refcnt > 1)) | |
4725 | return true; | |
f97a4a1a | 4726 | if (pwq->nr_active || !list_empty(&pwq->inactive_works)) |
c29eb853 TH |
4727 | return true; |
4728 | ||
4729 | return false; | |
4730 | } | |
4731 | ||
3af24433 ON |
4732 | /** |
4733 | * destroy_workqueue - safely terminate a workqueue | |
4734 | * @wq: target workqueue | |
4735 | * | |
4736 | * Safely destroy a workqueue. All work currently pending will be done first. | |
4737 | */ | |
4738 | void destroy_workqueue(struct workqueue_struct *wq) | |
4739 | { | |
49e3cf44 | 4740 | struct pool_workqueue *pwq; |
4c16bd32 | 4741 | int node; |
3af24433 | 4742 | |
def98c84 TH |
4743 | /* |
4744 | * Remove it from sysfs first so that sanity check failure doesn't | |
4745 | * lead to sysfs name conflicts. | |
4746 | */ | |
4747 | workqueue_sysfs_unregister(wq); | |
4748 | ||
33e3f0a3 RC |
4749 | /* mark the workqueue destruction is in progress */ |
4750 | mutex_lock(&wq->mutex); | |
4751 | wq->flags |= __WQ_DESTROYING; | |
4752 | mutex_unlock(&wq->mutex); | |
4753 | ||
9c5a2ba7 TH |
4754 | /* drain it before proceeding with destruction */ |
4755 | drain_workqueue(wq); | |
c8efcc25 | 4756 | |
def98c84 TH |
4757 | /* kill rescuer, if sanity checks fail, leave it w/o rescuer */ |
4758 | if (wq->rescuer) { | |
4759 | struct worker *rescuer = wq->rescuer; | |
4760 | ||
4761 | /* this prevents new queueing */ | |
a9b8a985 | 4762 | raw_spin_lock_irq(&wq_mayday_lock); |
def98c84 | 4763 | wq->rescuer = NULL; |
a9b8a985 | 4764 | raw_spin_unlock_irq(&wq_mayday_lock); |
def98c84 TH |
4765 | |
4766 | /* rescuer will empty maydays list before exiting */ | |
4767 | kthread_stop(rescuer->task); | |
8efe1223 | 4768 | kfree(rescuer); |
def98c84 TH |
4769 | } |
4770 | ||
c29eb853 TH |
4771 | /* |
4772 | * Sanity checks - grab all the locks so that we wait for all | |
4773 | * in-flight operations which may do put_pwq(). | |
4774 | */ | |
4775 | mutex_lock(&wq_pool_mutex); | |
b09f4fd3 | 4776 | mutex_lock(&wq->mutex); |
49e3cf44 | 4777 | for_each_pwq(pwq, wq) { |
a9b8a985 | 4778 | raw_spin_lock_irq(&pwq->pool->lock); |
c29eb853 | 4779 | if (WARN_ON(pwq_busy(pwq))) { |
1d9a6159 KW |
4780 | pr_warn("%s: %s has the following busy pwq\n", |
4781 | __func__, wq->name); | |
c29eb853 | 4782 | show_pwq(pwq); |
a9b8a985 | 4783 | raw_spin_unlock_irq(&pwq->pool->lock); |
b09f4fd3 | 4784 | mutex_unlock(&wq->mutex); |
c29eb853 | 4785 | mutex_unlock(&wq_pool_mutex); |
55df0933 | 4786 | show_one_workqueue(wq); |
6183c009 | 4787 | return; |
76af4d93 | 4788 | } |
a9b8a985 | 4789 | raw_spin_unlock_irq(&pwq->pool->lock); |
6183c009 | 4790 | } |
b09f4fd3 | 4791 | mutex_unlock(&wq->mutex); |
6183c009 | 4792 | |
a0a1a5fd TH |
4793 | /* |
4794 | * wq list is used to freeze wq, remove from list after | |
4795 | * flushing is complete in case freeze races us. | |
4796 | */ | |
e2dca7ad | 4797 | list_del_rcu(&wq->list); |
68e13a67 | 4798 | mutex_unlock(&wq_pool_mutex); |
3af24433 | 4799 | |
8864b4e5 | 4800 | if (!(wq->flags & WQ_UNBOUND)) { |
669de8bd | 4801 | wq_unregister_lockdep(wq); |
8864b4e5 TH |
4802 | /* |
4803 | * The base ref is never dropped on per-cpu pwqs. Directly | |
e2dca7ad | 4804 | * schedule RCU free. |
8864b4e5 | 4805 | */ |
25b00775 | 4806 | call_rcu(&wq->rcu, rcu_free_wq); |
8864b4e5 TH |
4807 | } else { |
4808 | /* | |
4809 | * We're the sole accessor of @wq at this point. Directly | |
4c16bd32 TH |
4810 | * access numa_pwq_tbl[] and dfl_pwq to put the base refs. |
4811 | * @wq will be freed when the last pwq is released. | |
8864b4e5 | 4812 | */ |
4c16bd32 TH |
4813 | for_each_node(node) { |
4814 | pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); | |
4815 | RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL); | |
4816 | put_pwq_unlocked(pwq); | |
4817 | } | |
4818 | ||
4819 | /* | |
4820 | * Put dfl_pwq. @wq may be freed any time after dfl_pwq is | |
4821 | * put. Don't access it afterwards. | |
4822 | */ | |
4823 | pwq = wq->dfl_pwq; | |
4824 | wq->dfl_pwq = NULL; | |
dce90d47 | 4825 | put_pwq_unlocked(pwq); |
29c91e99 | 4826 | } |
3af24433 ON |
4827 | } |
4828 | EXPORT_SYMBOL_GPL(destroy_workqueue); | |
4829 | ||
dcd989cb TH |
4830 | /** |
4831 | * workqueue_set_max_active - adjust max_active of a workqueue | |
4832 | * @wq: target workqueue | |
4833 | * @max_active: new max_active value. | |
4834 | * | |
4835 | * Set max_active of @wq to @max_active. | |
4836 | * | |
4837 | * CONTEXT: | |
4838 | * Don't call from IRQ context. | |
4839 | */ | |
4840 | void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) | |
4841 | { | |
49e3cf44 | 4842 | struct pool_workqueue *pwq; |
dcd989cb | 4843 | |
8719dcea | 4844 | /* disallow meddling with max_active for ordered workqueues */ |
0a94efb5 | 4845 | if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) |
8719dcea TH |
4846 | return; |
4847 | ||
f3421797 | 4848 | max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); |
dcd989cb | 4849 | |
a357fc03 | 4850 | mutex_lock(&wq->mutex); |
dcd989cb | 4851 | |
0a94efb5 | 4852 | wq->flags &= ~__WQ_ORDERED; |
dcd989cb TH |
4853 | wq->saved_max_active = max_active; |
4854 | ||
699ce097 TH |
4855 | for_each_pwq(pwq, wq) |
4856 | pwq_adjust_max_active(pwq); | |
93981800 | 4857 | |
a357fc03 | 4858 | mutex_unlock(&wq->mutex); |
15316ba8 | 4859 | } |
dcd989cb | 4860 | EXPORT_SYMBOL_GPL(workqueue_set_max_active); |
15316ba8 | 4861 | |
27d4ee03 LW |
4862 | /** |
4863 | * current_work - retrieve %current task's work struct | |
4864 | * | |
4865 | * Determine if %current task is a workqueue worker and what it's working on. | |
4866 | * Useful to find out the context that the %current task is running in. | |
4867 | * | |
4868 | * Return: work struct if %current task is a workqueue worker, %NULL otherwise. | |
4869 | */ | |
4870 | struct work_struct *current_work(void) | |
4871 | { | |
4872 | struct worker *worker = current_wq_worker(); | |
4873 | ||
4874 | return worker ? worker->current_work : NULL; | |
4875 | } | |
4876 | EXPORT_SYMBOL(current_work); | |
4877 | ||
e6267616 TH |
4878 | /** |
4879 | * current_is_workqueue_rescuer - is %current workqueue rescuer? | |
4880 | * | |
4881 | * Determine whether %current is a workqueue rescuer. Can be used from | |
4882 | * work functions to determine whether it's being run off the rescuer task. | |
d185af30 YB |
4883 | * |
4884 | * Return: %true if %current is a workqueue rescuer. %false otherwise. | |
e6267616 TH |
4885 | */ |
4886 | bool current_is_workqueue_rescuer(void) | |
4887 | { | |
4888 | struct worker *worker = current_wq_worker(); | |
4889 | ||
6a092dfd | 4890 | return worker && worker->rescue_wq; |
e6267616 TH |
4891 | } |
4892 | ||
eef6a7d5 | 4893 | /** |
dcd989cb TH |
4894 | * workqueue_congested - test whether a workqueue is congested |
4895 | * @cpu: CPU in question | |
4896 | * @wq: target workqueue | |
eef6a7d5 | 4897 | * |
dcd989cb TH |
4898 | * Test whether @wq's cpu workqueue for @cpu is congested. There is |
4899 | * no synchronization around this function and the test result is | |
4900 | * unreliable and only useful as advisory hints or for debugging. | |
eef6a7d5 | 4901 | * |
d3251859 TH |
4902 | * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. |
4903 | * Note that both per-cpu and unbound workqueues may be associated with | |
4904 | * multiple pool_workqueues which have separate congested states. A | |
4905 | * workqueue being congested on one CPU doesn't mean the workqueue is also | |
4906 | * contested on other CPUs / NUMA nodes. | |
4907 | * | |
d185af30 | 4908 | * Return: |
dcd989cb | 4909 | * %true if congested, %false otherwise. |
eef6a7d5 | 4910 | */ |
d84ff051 | 4911 | bool workqueue_congested(int cpu, struct workqueue_struct *wq) |
1da177e4 | 4912 | { |
7fb98ea7 | 4913 | struct pool_workqueue *pwq; |
76af4d93 TH |
4914 | bool ret; |
4915 | ||
24acfb71 TG |
4916 | rcu_read_lock(); |
4917 | preempt_disable(); | |
7fb98ea7 | 4918 | |
d3251859 TH |
4919 | if (cpu == WORK_CPU_UNBOUND) |
4920 | cpu = smp_processor_id(); | |
4921 | ||
7fb98ea7 TH |
4922 | if (!(wq->flags & WQ_UNBOUND)) |
4923 | pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); | |
4924 | else | |
df2d5ae4 | 4925 | pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); |
dcd989cb | 4926 | |
f97a4a1a | 4927 | ret = !list_empty(&pwq->inactive_works); |
24acfb71 TG |
4928 | preempt_enable(); |
4929 | rcu_read_unlock(); | |
76af4d93 TH |
4930 | |
4931 | return ret; | |
1da177e4 | 4932 | } |
dcd989cb | 4933 | EXPORT_SYMBOL_GPL(workqueue_congested); |
1da177e4 | 4934 | |
dcd989cb TH |
4935 | /** |
4936 | * work_busy - test whether a work is currently pending or running | |
4937 | * @work: the work to be tested | |
4938 | * | |
4939 | * Test whether @work is currently pending or running. There is no | |
4940 | * synchronization around this function and the test result is | |
4941 | * unreliable and only useful as advisory hints or for debugging. | |
dcd989cb | 4942 | * |
d185af30 | 4943 | * Return: |
dcd989cb TH |
4944 | * OR'd bitmask of WORK_BUSY_* bits. |
4945 | */ | |
4946 | unsigned int work_busy(struct work_struct *work) | |
1da177e4 | 4947 | { |
fa1b54e6 | 4948 | struct worker_pool *pool; |
dcd989cb TH |
4949 | unsigned long flags; |
4950 | unsigned int ret = 0; | |
1da177e4 | 4951 | |
dcd989cb TH |
4952 | if (work_pending(work)) |
4953 | ret |= WORK_BUSY_PENDING; | |
1da177e4 | 4954 | |
24acfb71 | 4955 | rcu_read_lock(); |
fa1b54e6 | 4956 | pool = get_work_pool(work); |
038366c5 | 4957 | if (pool) { |
a9b8a985 | 4958 | raw_spin_lock_irqsave(&pool->lock, flags); |
038366c5 LJ |
4959 | if (find_worker_executing_work(pool, work)) |
4960 | ret |= WORK_BUSY_RUNNING; | |
a9b8a985 | 4961 | raw_spin_unlock_irqrestore(&pool->lock, flags); |
038366c5 | 4962 | } |
24acfb71 | 4963 | rcu_read_unlock(); |
1da177e4 | 4964 | |
dcd989cb | 4965 | return ret; |
1da177e4 | 4966 | } |
dcd989cb | 4967 | EXPORT_SYMBOL_GPL(work_busy); |
1da177e4 | 4968 | |
3d1cb205 TH |
4969 | /** |
4970 | * set_worker_desc - set description for the current work item | |
4971 | * @fmt: printf-style format string | |
4972 | * @...: arguments for the format string | |
4973 | * | |
4974 | * This function can be called by a running work function to describe what | |
4975 | * the work item is about. If the worker task gets dumped, this | |
4976 | * information will be printed out together to help debugging. The | |
4977 | * description can be at most WORKER_DESC_LEN including the trailing '\0'. | |
4978 | */ | |
4979 | void set_worker_desc(const char *fmt, ...) | |
4980 | { | |
4981 | struct worker *worker = current_wq_worker(); | |
4982 | va_list args; | |
4983 | ||
4984 | if (worker) { | |
4985 | va_start(args, fmt); | |
4986 | vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); | |
4987 | va_end(args); | |
3d1cb205 TH |
4988 | } |
4989 | } | |
5c750d58 | 4990 | EXPORT_SYMBOL_GPL(set_worker_desc); |
3d1cb205 TH |
4991 | |
4992 | /** | |
4993 | * print_worker_info - print out worker information and description | |
4994 | * @log_lvl: the log level to use when printing | |
4995 | * @task: target task | |
4996 | * | |
4997 | * If @task is a worker and currently executing a work item, print out the | |
4998 | * name of the workqueue being serviced and worker description set with | |
4999 | * set_worker_desc() by the currently executing work item. | |
5000 | * | |
5001 | * This function can be safely called on any task as long as the | |
5002 | * task_struct itself is accessible. While safe, this function isn't | |
5003 | * synchronized and may print out mixups or garbages of limited length. | |
5004 | */ | |
5005 | void print_worker_info(const char *log_lvl, struct task_struct *task) | |
5006 | { | |
5007 | work_func_t *fn = NULL; | |
5008 | char name[WQ_NAME_LEN] = { }; | |
5009 | char desc[WORKER_DESC_LEN] = { }; | |
5010 | struct pool_workqueue *pwq = NULL; | |
5011 | struct workqueue_struct *wq = NULL; | |
3d1cb205 TH |
5012 | struct worker *worker; |
5013 | ||
5014 | if (!(task->flags & PF_WQ_WORKER)) | |
5015 | return; | |
5016 | ||
5017 | /* | |
5018 | * This function is called without any synchronization and @task | |
5019 | * could be in any state. Be careful with dereferences. | |
5020 | */ | |
e700591a | 5021 | worker = kthread_probe_data(task); |
3d1cb205 TH |
5022 | |
5023 | /* | |
8bf89593 TH |
5024 | * Carefully copy the associated workqueue's workfn, name and desc. |
5025 | * Keep the original last '\0' in case the original is garbage. | |
3d1cb205 | 5026 | */ |
fe557319 CH |
5027 | copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn)); |
5028 | copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq)); | |
5029 | copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq)); | |
5030 | copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1); | |
5031 | copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1); | |
3d1cb205 TH |
5032 | |
5033 | if (fn || name[0] || desc[0]) { | |
d75f773c | 5034 | printk("%sWorkqueue: %s %ps", log_lvl, name, fn); |
8bf89593 | 5035 | if (strcmp(name, desc)) |
3d1cb205 TH |
5036 | pr_cont(" (%s)", desc); |
5037 | pr_cont("\n"); | |
5038 | } | |
5039 | } | |
5040 | ||
3494fc30 TH |
5041 | static void pr_cont_pool_info(struct worker_pool *pool) |
5042 | { | |
5043 | pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask); | |
5044 | if (pool->node != NUMA_NO_NODE) | |
5045 | pr_cont(" node=%d", pool->node); | |
5046 | pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice); | |
5047 | } | |
5048 | ||
c76feb0d PM |
5049 | struct pr_cont_work_struct { |
5050 | bool comma; | |
5051 | work_func_t func; | |
5052 | long ctr; | |
5053 | }; | |
5054 | ||
5055 | static void pr_cont_work_flush(bool comma, work_func_t func, struct pr_cont_work_struct *pcwsp) | |
5056 | { | |
5057 | if (!pcwsp->ctr) | |
5058 | goto out_record; | |
5059 | if (func == pcwsp->func) { | |
5060 | pcwsp->ctr++; | |
5061 | return; | |
5062 | } | |
5063 | if (pcwsp->ctr == 1) | |
5064 | pr_cont("%s %ps", pcwsp->comma ? "," : "", pcwsp->func); | |
5065 | else | |
5066 | pr_cont("%s %ld*%ps", pcwsp->comma ? "," : "", pcwsp->ctr, pcwsp->func); | |
5067 | pcwsp->ctr = 0; | |
5068 | out_record: | |
5069 | if ((long)func == -1L) | |
5070 | return; | |
5071 | pcwsp->comma = comma; | |
5072 | pcwsp->func = func; | |
5073 | pcwsp->ctr = 1; | |
5074 | } | |
5075 | ||
5076 | static void pr_cont_work(bool comma, struct work_struct *work, struct pr_cont_work_struct *pcwsp) | |
3494fc30 TH |
5077 | { |
5078 | if (work->func == wq_barrier_func) { | |
5079 | struct wq_barrier *barr; | |
5080 | ||
5081 | barr = container_of(work, struct wq_barrier, work); | |
5082 | ||
c76feb0d | 5083 | pr_cont_work_flush(comma, (work_func_t)-1, pcwsp); |
3494fc30 TH |
5084 | pr_cont("%s BAR(%d)", comma ? "," : "", |
5085 | task_pid_nr(barr->task)); | |
5086 | } else { | |
c76feb0d PM |
5087 | if (!comma) |
5088 | pr_cont_work_flush(comma, (work_func_t)-1, pcwsp); | |
5089 | pr_cont_work_flush(comma, work->func, pcwsp); | |
3494fc30 TH |
5090 | } |
5091 | } | |
5092 | ||
5093 | static void show_pwq(struct pool_workqueue *pwq) | |
5094 | { | |
c76feb0d | 5095 | struct pr_cont_work_struct pcws = { .ctr = 0, }; |
3494fc30 TH |
5096 | struct worker_pool *pool = pwq->pool; |
5097 | struct work_struct *work; | |
5098 | struct worker *worker; | |
5099 | bool has_in_flight = false, has_pending = false; | |
5100 | int bkt; | |
5101 | ||
5102 | pr_info(" pwq %d:", pool->id); | |
5103 | pr_cont_pool_info(pool); | |
5104 | ||
e66b39af TH |
5105 | pr_cont(" active=%d/%d refcnt=%d%s\n", |
5106 | pwq->nr_active, pwq->max_active, pwq->refcnt, | |
3494fc30 TH |
5107 | !list_empty(&pwq->mayday_node) ? " MAYDAY" : ""); |
5108 | ||
5109 | hash_for_each(pool->busy_hash, bkt, worker, hentry) { | |
5110 | if (worker->current_pwq == pwq) { | |
5111 | has_in_flight = true; | |
5112 | break; | |
5113 | } | |
5114 | } | |
5115 | if (has_in_flight) { | |
5116 | bool comma = false; | |
5117 | ||
5118 | pr_info(" in-flight:"); | |
5119 | hash_for_each(pool->busy_hash, bkt, worker, hentry) { | |
5120 | if (worker->current_pwq != pwq) | |
5121 | continue; | |
5122 | ||
d75f773c | 5123 | pr_cont("%s %d%s:%ps", comma ? "," : "", |
3494fc30 | 5124 | task_pid_nr(worker->task), |
30ae2fc0 | 5125 | worker->rescue_wq ? "(RESCUER)" : "", |
3494fc30 TH |
5126 | worker->current_func); |
5127 | list_for_each_entry(work, &worker->scheduled, entry) | |
c76feb0d PM |
5128 | pr_cont_work(false, work, &pcws); |
5129 | pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); | |
3494fc30 TH |
5130 | comma = true; |
5131 | } | |
5132 | pr_cont("\n"); | |
5133 | } | |
5134 | ||
5135 | list_for_each_entry(work, &pool->worklist, entry) { | |
5136 | if (get_work_pwq(work) == pwq) { | |
5137 | has_pending = true; | |
5138 | break; | |
5139 | } | |
5140 | } | |
5141 | if (has_pending) { | |
5142 | bool comma = false; | |
5143 | ||
5144 | pr_info(" pending:"); | |
5145 | list_for_each_entry(work, &pool->worklist, entry) { | |
5146 | if (get_work_pwq(work) != pwq) | |
5147 | continue; | |
5148 | ||
c76feb0d | 5149 | pr_cont_work(comma, work, &pcws); |
3494fc30 TH |
5150 | comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); |
5151 | } | |
c76feb0d | 5152 | pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); |
3494fc30 TH |
5153 | pr_cont("\n"); |
5154 | } | |
5155 | ||
f97a4a1a | 5156 | if (!list_empty(&pwq->inactive_works)) { |
3494fc30 TH |
5157 | bool comma = false; |
5158 | ||
f97a4a1a LJ |
5159 | pr_info(" inactive:"); |
5160 | list_for_each_entry(work, &pwq->inactive_works, entry) { | |
c76feb0d | 5161 | pr_cont_work(comma, work, &pcws); |
3494fc30 TH |
5162 | comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); |
5163 | } | |
c76feb0d | 5164 | pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); |
3494fc30 TH |
5165 | pr_cont("\n"); |
5166 | } | |
5167 | } | |
5168 | ||
5169 | /** | |
55df0933 IK |
5170 | * show_one_workqueue - dump state of specified workqueue |
5171 | * @wq: workqueue whose state will be printed | |
3494fc30 | 5172 | */ |
55df0933 | 5173 | void show_one_workqueue(struct workqueue_struct *wq) |
3494fc30 | 5174 | { |
55df0933 IK |
5175 | struct pool_workqueue *pwq; |
5176 | bool idle = true; | |
3494fc30 | 5177 | unsigned long flags; |
3494fc30 | 5178 | |
55df0933 IK |
5179 | for_each_pwq(pwq, wq) { |
5180 | if (pwq->nr_active || !list_empty(&pwq->inactive_works)) { | |
5181 | idle = false; | |
5182 | break; | |
3494fc30 | 5183 | } |
55df0933 IK |
5184 | } |
5185 | if (idle) /* Nothing to print for idle workqueue */ | |
5186 | return; | |
3494fc30 | 5187 | |
55df0933 | 5188 | pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags); |
3494fc30 | 5189 | |
55df0933 IK |
5190 | for_each_pwq(pwq, wq) { |
5191 | raw_spin_lock_irqsave(&pwq->pool->lock, flags); | |
5192 | if (pwq->nr_active || !list_empty(&pwq->inactive_works)) { | |
62635ea8 | 5193 | /* |
55df0933 IK |
5194 | * Defer printing to avoid deadlocks in console |
5195 | * drivers that queue work while holding locks | |
5196 | * also taken in their write paths. | |
62635ea8 | 5197 | */ |
55df0933 IK |
5198 | printk_deferred_enter(); |
5199 | show_pwq(pwq); | |
5200 | printk_deferred_exit(); | |
3494fc30 | 5201 | } |
55df0933 | 5202 | raw_spin_unlock_irqrestore(&pwq->pool->lock, flags); |
62635ea8 SS |
5203 | /* |
5204 | * We could be printing a lot from atomic context, e.g. | |
55df0933 | 5205 | * sysrq-t -> show_all_workqueues(). Avoid triggering |
62635ea8 SS |
5206 | * hard lockup. |
5207 | */ | |
5208 | touch_nmi_watchdog(); | |
3494fc30 TH |
5209 | } |
5210 | ||
55df0933 IK |
5211 | } |
5212 | ||
5213 | /** | |
5214 | * show_one_worker_pool - dump state of specified worker pool | |
5215 | * @pool: worker pool whose state will be printed | |
5216 | */ | |
5217 | static void show_one_worker_pool(struct worker_pool *pool) | |
5218 | { | |
5219 | struct worker *worker; | |
5220 | bool first = true; | |
5221 | unsigned long flags; | |
335a42eb | 5222 | unsigned long hung = 0; |
55df0933 IK |
5223 | |
5224 | raw_spin_lock_irqsave(&pool->lock, flags); | |
5225 | if (pool->nr_workers == pool->nr_idle) | |
5226 | goto next_pool; | |
335a42eb PM |
5227 | |
5228 | /* How long the first pending work is waiting for a worker. */ | |
5229 | if (!list_empty(&pool->worklist)) | |
5230 | hung = jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000; | |
5231 | ||
55df0933 IK |
5232 | /* |
5233 | * Defer printing to avoid deadlocks in console drivers that | |
5234 | * queue work while holding locks also taken in their write | |
5235 | * paths. | |
5236 | */ | |
5237 | printk_deferred_enter(); | |
5238 | pr_info("pool %d:", pool->id); | |
5239 | pr_cont_pool_info(pool); | |
335a42eb | 5240 | pr_cont(" hung=%lus workers=%d", hung, pool->nr_workers); |
55df0933 IK |
5241 | if (pool->manager) |
5242 | pr_cont(" manager: %d", | |
5243 | task_pid_nr(pool->manager->task)); | |
5244 | list_for_each_entry(worker, &pool->idle_list, entry) { | |
5245 | pr_cont(" %s%d", first ? "idle: " : "", | |
5246 | task_pid_nr(worker->task)); | |
5247 | first = false; | |
5248 | } | |
5249 | pr_cont("\n"); | |
5250 | printk_deferred_exit(); | |
5251 | next_pool: | |
5252 | raw_spin_unlock_irqrestore(&pool->lock, flags); | |
5253 | /* | |
5254 | * We could be printing a lot from atomic context, e.g. | |
5255 | * sysrq-t -> show_all_workqueues(). Avoid triggering | |
5256 | * hard lockup. | |
5257 | */ | |
5258 | touch_nmi_watchdog(); | |
5259 | ||
5260 | } | |
5261 | ||
5262 | /** | |
5263 | * show_all_workqueues - dump workqueue state | |
5264 | * | |
704bc669 | 5265 | * Called from a sysrq handler and prints out all busy workqueues and pools. |
55df0933 IK |
5266 | */ |
5267 | void show_all_workqueues(void) | |
5268 | { | |
5269 | struct workqueue_struct *wq; | |
5270 | struct worker_pool *pool; | |
5271 | int pi; | |
5272 | ||
5273 | rcu_read_lock(); | |
5274 | ||
5275 | pr_info("Showing busy workqueues and worker pools:\n"); | |
5276 | ||
5277 | list_for_each_entry_rcu(wq, &workqueues, list) | |
5278 | show_one_workqueue(wq); | |
5279 | ||
5280 | for_each_pool(pool, pi) | |
5281 | show_one_worker_pool(pool); | |
5282 | ||
24acfb71 | 5283 | rcu_read_unlock(); |
3494fc30 TH |
5284 | } |
5285 | ||
704bc669 JL |
5286 | /** |
5287 | * show_freezable_workqueues - dump freezable workqueue state | |
5288 | * | |
5289 | * Called from try_to_freeze_tasks() and prints out all freezable workqueues | |
5290 | * still busy. | |
5291 | */ | |
5292 | void show_freezable_workqueues(void) | |
5293 | { | |
5294 | struct workqueue_struct *wq; | |
5295 | ||
5296 | rcu_read_lock(); | |
5297 | ||
5298 | pr_info("Showing freezable workqueues that are still busy:\n"); | |
5299 | ||
5300 | list_for_each_entry_rcu(wq, &workqueues, list) { | |
5301 | if (!(wq->flags & WQ_FREEZABLE)) | |
5302 | continue; | |
5303 | show_one_workqueue(wq); | |
5304 | } | |
5305 | ||
5306 | rcu_read_unlock(); | |
5307 | } | |
5308 | ||
6b59808b TH |
5309 | /* used to show worker information through /proc/PID/{comm,stat,status} */ |
5310 | void wq_worker_comm(char *buf, size_t size, struct task_struct *task) | |
5311 | { | |
6b59808b TH |
5312 | int off; |
5313 | ||
5314 | /* always show the actual comm */ | |
5315 | off = strscpy(buf, task->comm, size); | |
5316 | if (off < 0) | |
5317 | return; | |
5318 | ||
197f6acc | 5319 | /* stabilize PF_WQ_WORKER and worker pool association */ |
6b59808b TH |
5320 | mutex_lock(&wq_pool_attach_mutex); |
5321 | ||
197f6acc TH |
5322 | if (task->flags & PF_WQ_WORKER) { |
5323 | struct worker *worker = kthread_data(task); | |
5324 | struct worker_pool *pool = worker->pool; | |
6b59808b | 5325 | |
197f6acc | 5326 | if (pool) { |
a9b8a985 | 5327 | raw_spin_lock_irq(&pool->lock); |
197f6acc TH |
5328 | /* |
5329 | * ->desc tracks information (wq name or | |
5330 | * set_worker_desc()) for the latest execution. If | |
5331 | * current, prepend '+', otherwise '-'. | |
5332 | */ | |
5333 | if (worker->desc[0] != '\0') { | |
5334 | if (worker->current_work) | |
5335 | scnprintf(buf + off, size - off, "+%s", | |
5336 | worker->desc); | |
5337 | else | |
5338 | scnprintf(buf + off, size - off, "-%s", | |
5339 | worker->desc); | |
5340 | } | |
a9b8a985 | 5341 | raw_spin_unlock_irq(&pool->lock); |
6b59808b | 5342 | } |
6b59808b TH |
5343 | } |
5344 | ||
5345 | mutex_unlock(&wq_pool_attach_mutex); | |
5346 | } | |
5347 | ||
66448bc2 MM |
5348 | #ifdef CONFIG_SMP |
5349 | ||
db7bccf4 TH |
5350 | /* |
5351 | * CPU hotplug. | |
5352 | * | |
e22bee78 | 5353 | * There are two challenges in supporting CPU hotplug. Firstly, there |
112202d9 | 5354 | * are a lot of assumptions on strong associations among work, pwq and |
706026c2 | 5355 | * pool which make migrating pending and scheduled works very |
e22bee78 | 5356 | * difficult to implement without impacting hot paths. Secondly, |
94cf58bb | 5357 | * worker pools serve mix of short, long and very long running works making |
e22bee78 TH |
5358 | * blocked draining impractical. |
5359 | * | |
24647570 | 5360 | * This is solved by allowing the pools to be disassociated from the CPU |
628c78e7 TH |
5361 | * running as an unbound one and allowing it to be reattached later if the |
5362 | * cpu comes back online. | |
db7bccf4 | 5363 | */ |
1da177e4 | 5364 | |
e8b3f8db | 5365 | static void unbind_workers(int cpu) |
3af24433 | 5366 | { |
4ce62e9e | 5367 | struct worker_pool *pool; |
db7bccf4 | 5368 | struct worker *worker; |
3af24433 | 5369 | |
f02ae73a | 5370 | for_each_cpu_worker_pool(pool, cpu) { |
1258fae7 | 5371 | mutex_lock(&wq_pool_attach_mutex); |
a9b8a985 | 5372 | raw_spin_lock_irq(&pool->lock); |
3af24433 | 5373 | |
94cf58bb | 5374 | /* |
92f9c5c4 | 5375 | * We've blocked all attach/detach operations. Make all workers |
94cf58bb | 5376 | * unbound and set DISASSOCIATED. Before this, all workers |
11b45b0b | 5377 | * must be on the cpu. After this, they may become diasporas. |
b4ac9384 LJ |
5378 | * And the preemption disabled section in their sched callbacks |
5379 | * are guaranteed to see WORKER_UNBOUND since the code here | |
5380 | * is on the same cpu. | |
94cf58bb | 5381 | */ |
da028469 | 5382 | for_each_pool_worker(worker, pool) |
c9e7cf27 | 5383 | worker->flags |= WORKER_UNBOUND; |
06ba38a9 | 5384 | |
24647570 | 5385 | pool->flags |= POOL_DISASSOCIATED; |
f2d5a0ee | 5386 | |
eb283428 | 5387 | /* |
989442d7 LJ |
5388 | * The handling of nr_running in sched callbacks are disabled |
5389 | * now. Zap nr_running. After this, nr_running stays zero and | |
5390 | * need_more_worker() and keep_working() are always true as | |
5391 | * long as the worklist is not empty. This pool now behaves as | |
5392 | * an unbound (in terms of concurrency management) pool which | |
eb283428 LJ |
5393 | * are served by workers tied to the pool. |
5394 | */ | |
bc35f7ef | 5395 | pool->nr_running = 0; |
eb283428 LJ |
5396 | |
5397 | /* | |
5398 | * With concurrency management just turned off, a busy | |
5399 | * worker blocking could lead to lengthy stalls. Kick off | |
5400 | * unbound chain execution of currently pending work items. | |
5401 | */ | |
eb283428 | 5402 | wake_up_worker(pool); |
989442d7 | 5403 | |
a9b8a985 | 5404 | raw_spin_unlock_irq(&pool->lock); |
989442d7 | 5405 | |
793777bc VS |
5406 | for_each_pool_worker(worker, pool) |
5407 | unbind_worker(worker); | |
989442d7 LJ |
5408 | |
5409 | mutex_unlock(&wq_pool_attach_mutex); | |
eb283428 | 5410 | } |
3af24433 | 5411 | } |
3af24433 | 5412 | |
bd7c089e TH |
5413 | /** |
5414 | * rebind_workers - rebind all workers of a pool to the associated CPU | |
5415 | * @pool: pool of interest | |
5416 | * | |
a9ab775b | 5417 | * @pool->cpu is coming online. Rebind all workers to the CPU. |
bd7c089e TH |
5418 | */ |
5419 | static void rebind_workers(struct worker_pool *pool) | |
5420 | { | |
a9ab775b | 5421 | struct worker *worker; |
bd7c089e | 5422 | |
1258fae7 | 5423 | lockdep_assert_held(&wq_pool_attach_mutex); |
bd7c089e | 5424 | |
a9ab775b TH |
5425 | /* |
5426 | * Restore CPU affinity of all workers. As all idle workers should | |
5427 | * be on the run-queue of the associated CPU before any local | |
402dd89d | 5428 | * wake-ups for concurrency management happen, restore CPU affinity |
a9ab775b TH |
5429 | * of all workers first and then clear UNBOUND. As we're called |
5430 | * from CPU_ONLINE, the following shouldn't fail. | |
5431 | */ | |
c63a2e52 VS |
5432 | for_each_pool_worker(worker, pool) { |
5433 | kthread_set_per_cpu(worker->task, pool->cpu); | |
5434 | WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, | |
5435 | pool->attrs->cpumask) < 0); | |
5436 | } | |
bd7c089e | 5437 | |
a9b8a985 | 5438 | raw_spin_lock_irq(&pool->lock); |
f7c17d26 | 5439 | |
3de5e884 | 5440 | pool->flags &= ~POOL_DISASSOCIATED; |
bd7c089e | 5441 | |
da028469 | 5442 | for_each_pool_worker(worker, pool) { |
a9ab775b | 5443 | unsigned int worker_flags = worker->flags; |
bd7c089e | 5444 | |
a9ab775b TH |
5445 | /* |
5446 | * We want to clear UNBOUND but can't directly call | |
5447 | * worker_clr_flags() or adjust nr_running. Atomically | |
5448 | * replace UNBOUND with another NOT_RUNNING flag REBOUND. | |
5449 | * @worker will clear REBOUND using worker_clr_flags() when | |
5450 | * it initiates the next execution cycle thus restoring | |
5451 | * concurrency management. Note that when or whether | |
5452 | * @worker clears REBOUND doesn't affect correctness. | |
5453 | * | |
c95491ed | 5454 | * WRITE_ONCE() is necessary because @worker->flags may be |
a9ab775b | 5455 | * tested without holding any lock in |
6d25be57 | 5456 | * wq_worker_running(). Without it, NOT_RUNNING test may |
a9ab775b TH |
5457 | * fail incorrectly leading to premature concurrency |
5458 | * management operations. | |
5459 | */ | |
5460 | WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); | |
5461 | worker_flags |= WORKER_REBOUND; | |
5462 | worker_flags &= ~WORKER_UNBOUND; | |
c95491ed | 5463 | WRITE_ONCE(worker->flags, worker_flags); |
bd7c089e | 5464 | } |
a9ab775b | 5465 | |
a9b8a985 | 5466 | raw_spin_unlock_irq(&pool->lock); |
bd7c089e TH |
5467 | } |
5468 | ||
7dbc725e TH |
5469 | /** |
5470 | * restore_unbound_workers_cpumask - restore cpumask of unbound workers | |
5471 | * @pool: unbound pool of interest | |
5472 | * @cpu: the CPU which is coming up | |
5473 | * | |
5474 | * An unbound pool may end up with a cpumask which doesn't have any online | |
5475 | * CPUs. When a worker of such pool get scheduled, the scheduler resets | |
5476 | * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any | |
5477 | * online CPU before, cpus_allowed of all its workers should be restored. | |
5478 | */ | |
5479 | static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) | |
5480 | { | |
5481 | static cpumask_t cpumask; | |
5482 | struct worker *worker; | |
7dbc725e | 5483 | |
1258fae7 | 5484 | lockdep_assert_held(&wq_pool_attach_mutex); |
7dbc725e TH |
5485 | |
5486 | /* is @cpu allowed for @pool? */ | |
5487 | if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) | |
5488 | return; | |
5489 | ||
7dbc725e | 5490 | cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); |
7dbc725e TH |
5491 | |
5492 | /* as we're called from CPU_ONLINE, the following shouldn't fail */ | |
da028469 | 5493 | for_each_pool_worker(worker, pool) |
d945b5e9 | 5494 | WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0); |
7dbc725e TH |
5495 | } |
5496 | ||
7ee681b2 TG |
5497 | int workqueue_prepare_cpu(unsigned int cpu) |
5498 | { | |
5499 | struct worker_pool *pool; | |
5500 | ||
5501 | for_each_cpu_worker_pool(pool, cpu) { | |
5502 | if (pool->nr_workers) | |
5503 | continue; | |
5504 | if (!create_worker(pool)) | |
5505 | return -ENOMEM; | |
5506 | } | |
5507 | return 0; | |
5508 | } | |
5509 | ||
5510 | int workqueue_online_cpu(unsigned int cpu) | |
3af24433 | 5511 | { |
4ce62e9e | 5512 | struct worker_pool *pool; |
4c16bd32 | 5513 | struct workqueue_struct *wq; |
7dbc725e | 5514 | int pi; |
3ce63377 | 5515 | |
7ee681b2 | 5516 | mutex_lock(&wq_pool_mutex); |
7dbc725e | 5517 | |
7ee681b2 | 5518 | for_each_pool(pool, pi) { |
1258fae7 | 5519 | mutex_lock(&wq_pool_attach_mutex); |
94cf58bb | 5520 | |
7ee681b2 TG |
5521 | if (pool->cpu == cpu) |
5522 | rebind_workers(pool); | |
5523 | else if (pool->cpu < 0) | |
5524 | restore_unbound_workers_cpumask(pool, cpu); | |
94cf58bb | 5525 | |
1258fae7 | 5526 | mutex_unlock(&wq_pool_attach_mutex); |
7ee681b2 | 5527 | } |
6ba94429 | 5528 | |
7ee681b2 TG |
5529 | /* update NUMA affinity of unbound workqueues */ |
5530 | list_for_each_entry(wq, &workqueues, list) | |
5531 | wq_update_unbound_numa(wq, cpu, true); | |
6ba94429 | 5532 | |
7ee681b2 TG |
5533 | mutex_unlock(&wq_pool_mutex); |
5534 | return 0; | |
6ba94429 FW |
5535 | } |
5536 | ||
7ee681b2 | 5537 | int workqueue_offline_cpu(unsigned int cpu) |
6ba94429 | 5538 | { |
6ba94429 FW |
5539 | struct workqueue_struct *wq; |
5540 | ||
7ee681b2 | 5541 | /* unbinding per-cpu workers should happen on the local CPU */ |
e8b3f8db LJ |
5542 | if (WARN_ON(cpu != smp_processor_id())) |
5543 | return -1; | |
5544 | ||
5545 | unbind_workers(cpu); | |
7ee681b2 TG |
5546 | |
5547 | /* update NUMA affinity of unbound workqueues */ | |
5548 | mutex_lock(&wq_pool_mutex); | |
5549 | list_for_each_entry(wq, &workqueues, list) | |
5550 | wq_update_unbound_numa(wq, cpu, false); | |
5551 | mutex_unlock(&wq_pool_mutex); | |
5552 | ||
7ee681b2 | 5553 | return 0; |
6ba94429 FW |
5554 | } |
5555 | ||
6ba94429 FW |
5556 | struct work_for_cpu { |
5557 | struct work_struct work; | |
5558 | long (*fn)(void *); | |
5559 | void *arg; | |
5560 | long ret; | |
5561 | }; | |
5562 | ||
5563 | static void work_for_cpu_fn(struct work_struct *work) | |
5564 | { | |
5565 | struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); | |
5566 | ||
5567 | wfc->ret = wfc->fn(wfc->arg); | |
5568 | } | |
5569 | ||
5570 | /** | |
22aceb31 | 5571 | * work_on_cpu - run a function in thread context on a particular cpu |
6ba94429 FW |
5572 | * @cpu: the cpu to run on |
5573 | * @fn: the function to run | |
5574 | * @arg: the function arg | |
5575 | * | |
5576 | * It is up to the caller to ensure that the cpu doesn't go offline. | |
5577 | * The caller must not hold any locks which would prevent @fn from completing. | |
5578 | * | |
5579 | * Return: The value @fn returns. | |
5580 | */ | |
5581 | long work_on_cpu(int cpu, long (*fn)(void *), void *arg) | |
5582 | { | |
5583 | struct work_for_cpu wfc = { .fn = fn, .arg = arg }; | |
5584 | ||
5585 | INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); | |
5586 | schedule_work_on(cpu, &wfc.work); | |
5587 | flush_work(&wfc.work); | |
5588 | destroy_work_on_stack(&wfc.work); | |
5589 | return wfc.ret; | |
5590 | } | |
5591 | EXPORT_SYMBOL_GPL(work_on_cpu); | |
0e8d6a93 TG |
5592 | |
5593 | /** | |
5594 | * work_on_cpu_safe - run a function in thread context on a particular cpu | |
5595 | * @cpu: the cpu to run on | |
5596 | * @fn: the function to run | |
5597 | * @arg: the function argument | |
5598 | * | |
5599 | * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold | |
5600 | * any locks which would prevent @fn from completing. | |
5601 | * | |
5602 | * Return: The value @fn returns. | |
5603 | */ | |
5604 | long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg) | |
5605 | { | |
5606 | long ret = -ENODEV; | |
5607 | ||
ffd8bea8 | 5608 | cpus_read_lock(); |
0e8d6a93 TG |
5609 | if (cpu_online(cpu)) |
5610 | ret = work_on_cpu(cpu, fn, arg); | |
ffd8bea8 | 5611 | cpus_read_unlock(); |
0e8d6a93 TG |
5612 | return ret; |
5613 | } | |
5614 | EXPORT_SYMBOL_GPL(work_on_cpu_safe); | |
6ba94429 FW |
5615 | #endif /* CONFIG_SMP */ |
5616 | ||
5617 | #ifdef CONFIG_FREEZER | |
5618 | ||
5619 | /** | |
5620 | * freeze_workqueues_begin - begin freezing workqueues | |
5621 | * | |
5622 | * Start freezing workqueues. After this function returns, all freezable | |
f97a4a1a | 5623 | * workqueues will queue new works to their inactive_works list instead of |
6ba94429 FW |
5624 | * pool->worklist. |
5625 | * | |
5626 | * CONTEXT: | |
5627 | * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. | |
5628 | */ | |
5629 | void freeze_workqueues_begin(void) | |
5630 | { | |
5631 | struct workqueue_struct *wq; | |
5632 | struct pool_workqueue *pwq; | |
5633 | ||
5634 | mutex_lock(&wq_pool_mutex); | |
5635 | ||
5636 | WARN_ON_ONCE(workqueue_freezing); | |
5637 | workqueue_freezing = true; | |
5638 | ||
5639 | list_for_each_entry(wq, &workqueues, list) { | |
5640 | mutex_lock(&wq->mutex); | |
5641 | for_each_pwq(pwq, wq) | |
5642 | pwq_adjust_max_active(pwq); | |
5643 | mutex_unlock(&wq->mutex); | |
5644 | } | |
5645 | ||
5646 | mutex_unlock(&wq_pool_mutex); | |
5647 | } | |
5648 | ||
5649 | /** | |
5650 | * freeze_workqueues_busy - are freezable workqueues still busy? | |
5651 | * | |
5652 | * Check whether freezing is complete. This function must be called | |
5653 | * between freeze_workqueues_begin() and thaw_workqueues(). | |
5654 | * | |
5655 | * CONTEXT: | |
5656 | * Grabs and releases wq_pool_mutex. | |
5657 | * | |
5658 | * Return: | |
5659 | * %true if some freezable workqueues are still busy. %false if freezing | |
5660 | * is complete. | |
5661 | */ | |
5662 | bool freeze_workqueues_busy(void) | |
5663 | { | |
5664 | bool busy = false; | |
5665 | struct workqueue_struct *wq; | |
5666 | struct pool_workqueue *pwq; | |
5667 | ||
5668 | mutex_lock(&wq_pool_mutex); | |
5669 | ||
5670 | WARN_ON_ONCE(!workqueue_freezing); | |
5671 | ||
5672 | list_for_each_entry(wq, &workqueues, list) { | |
5673 | if (!(wq->flags & WQ_FREEZABLE)) | |
5674 | continue; | |
5675 | /* | |
5676 | * nr_active is monotonically decreasing. It's safe | |
5677 | * to peek without lock. | |
5678 | */ | |
24acfb71 | 5679 | rcu_read_lock(); |
6ba94429 FW |
5680 | for_each_pwq(pwq, wq) { |
5681 | WARN_ON_ONCE(pwq->nr_active < 0); | |
5682 | if (pwq->nr_active) { | |
5683 | busy = true; | |
24acfb71 | 5684 | rcu_read_unlock(); |
6ba94429 FW |
5685 | goto out_unlock; |
5686 | } | |
5687 | } | |
24acfb71 | 5688 | rcu_read_unlock(); |
6ba94429 FW |
5689 | } |
5690 | out_unlock: | |
5691 | mutex_unlock(&wq_pool_mutex); | |
5692 | return busy; | |
5693 | } | |
5694 | ||
5695 | /** | |
5696 | * thaw_workqueues - thaw workqueues | |
5697 | * | |
5698 | * Thaw workqueues. Normal queueing is restored and all collected | |
5699 | * frozen works are transferred to their respective pool worklists. | |
5700 | * | |
5701 | * CONTEXT: | |
5702 | * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. | |
5703 | */ | |
5704 | void thaw_workqueues(void) | |
5705 | { | |
5706 | struct workqueue_struct *wq; | |
5707 | struct pool_workqueue *pwq; | |
5708 | ||
5709 | mutex_lock(&wq_pool_mutex); | |
5710 | ||
5711 | if (!workqueue_freezing) | |
5712 | goto out_unlock; | |
5713 | ||
5714 | workqueue_freezing = false; | |
5715 | ||
5716 | /* restore max_active and repopulate worklist */ | |
5717 | list_for_each_entry(wq, &workqueues, list) { | |
5718 | mutex_lock(&wq->mutex); | |
5719 | for_each_pwq(pwq, wq) | |
5720 | pwq_adjust_max_active(pwq); | |
5721 | mutex_unlock(&wq->mutex); | |
5722 | } | |
5723 | ||
5724 | out_unlock: | |
5725 | mutex_unlock(&wq_pool_mutex); | |
5726 | } | |
5727 | #endif /* CONFIG_FREEZER */ | |
5728 | ||
99c621ef | 5729 | static int workqueue_apply_unbound_cpumask(const cpumask_var_t unbound_cpumask) |
042f7df1 LJ |
5730 | { |
5731 | LIST_HEAD(ctxs); | |
5732 | int ret = 0; | |
5733 | struct workqueue_struct *wq; | |
5734 | struct apply_wqattrs_ctx *ctx, *n; | |
5735 | ||
5736 | lockdep_assert_held(&wq_pool_mutex); | |
5737 | ||
5738 | list_for_each_entry(wq, &workqueues, list) { | |
5739 | if (!(wq->flags & WQ_UNBOUND)) | |
5740 | continue; | |
5741 | /* creating multiple pwqs breaks ordering guarantee */ | |
5742 | if (wq->flags & __WQ_ORDERED) | |
5743 | continue; | |
5744 | ||
99c621ef | 5745 | ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs, unbound_cpumask); |
042f7df1 LJ |
5746 | if (!ctx) { |
5747 | ret = -ENOMEM; | |
5748 | break; | |
5749 | } | |
5750 | ||
5751 | list_add_tail(&ctx->list, &ctxs); | |
5752 | } | |
5753 | ||
5754 | list_for_each_entry_safe(ctx, n, &ctxs, list) { | |
5755 | if (!ret) | |
5756 | apply_wqattrs_commit(ctx); | |
5757 | apply_wqattrs_cleanup(ctx); | |
5758 | } | |
5759 | ||
99c621ef LJ |
5760 | if (!ret) { |
5761 | mutex_lock(&wq_pool_attach_mutex); | |
5762 | cpumask_copy(wq_unbound_cpumask, unbound_cpumask); | |
5763 | mutex_unlock(&wq_pool_attach_mutex); | |
5764 | } | |
042f7df1 LJ |
5765 | return ret; |
5766 | } | |
5767 | ||
5768 | /** | |
5769 | * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask | |
5770 | * @cpumask: the cpumask to set | |
5771 | * | |
5772 | * The low-level workqueues cpumask is a global cpumask that limits | |
5773 | * the affinity of all unbound workqueues. This function check the @cpumask | |
5774 | * and apply it to all unbound workqueues and updates all pwqs of them. | |
5775 | * | |
67dc8325 | 5776 | * Return: 0 - Success |
042f7df1 LJ |
5777 | * -EINVAL - Invalid @cpumask |
5778 | * -ENOMEM - Failed to allocate memory for attrs or pwqs. | |
5779 | */ | |
5780 | int workqueue_set_unbound_cpumask(cpumask_var_t cpumask) | |
5781 | { | |
5782 | int ret = -EINVAL; | |
042f7df1 | 5783 | |
c98a9805 TS |
5784 | /* |
5785 | * Not excluding isolated cpus on purpose. | |
5786 | * If the user wishes to include them, we allow that. | |
5787 | */ | |
042f7df1 LJ |
5788 | cpumask_and(cpumask, cpumask, cpu_possible_mask); |
5789 | if (!cpumask_empty(cpumask)) { | |
a0111cf6 | 5790 | apply_wqattrs_lock(); |
d25302e4 MD |
5791 | if (cpumask_equal(cpumask, wq_unbound_cpumask)) { |
5792 | ret = 0; | |
5793 | goto out_unlock; | |
5794 | } | |
5795 | ||
99c621ef | 5796 | ret = workqueue_apply_unbound_cpumask(cpumask); |
042f7df1 | 5797 | |
d25302e4 | 5798 | out_unlock: |
a0111cf6 | 5799 | apply_wqattrs_unlock(); |
042f7df1 | 5800 | } |
042f7df1 | 5801 | |
042f7df1 LJ |
5802 | return ret; |
5803 | } | |
5804 | ||
6ba94429 FW |
5805 | #ifdef CONFIG_SYSFS |
5806 | /* | |
5807 | * Workqueues with WQ_SYSFS flag set is visible to userland via | |
5808 | * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the | |
5809 | * following attributes. | |
5810 | * | |
5811 | * per_cpu RO bool : whether the workqueue is per-cpu or unbound | |
5812 | * max_active RW int : maximum number of in-flight work items | |
5813 | * | |
5814 | * Unbound workqueues have the following extra attributes. | |
5815 | * | |
9a19b463 | 5816 | * pool_ids RO int : the associated pool IDs for each node |
6ba94429 FW |
5817 | * nice RW int : nice value of the workers |
5818 | * cpumask RW mask : bitmask of allowed CPUs for the workers | |
9a19b463 | 5819 | * numa RW bool : whether enable NUMA affinity |
6ba94429 FW |
5820 | */ |
5821 | struct wq_device { | |
5822 | struct workqueue_struct *wq; | |
5823 | struct device dev; | |
5824 | }; | |
5825 | ||
5826 | static struct workqueue_struct *dev_to_wq(struct device *dev) | |
5827 | { | |
5828 | struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); | |
5829 | ||
5830 | return wq_dev->wq; | |
5831 | } | |
5832 | ||
5833 | static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, | |
5834 | char *buf) | |
5835 | { | |
5836 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5837 | ||
5838 | return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); | |
5839 | } | |
5840 | static DEVICE_ATTR_RO(per_cpu); | |
5841 | ||
5842 | static ssize_t max_active_show(struct device *dev, | |
5843 | struct device_attribute *attr, char *buf) | |
5844 | { | |
5845 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5846 | ||
5847 | return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); | |
5848 | } | |
5849 | ||
5850 | static ssize_t max_active_store(struct device *dev, | |
5851 | struct device_attribute *attr, const char *buf, | |
5852 | size_t count) | |
5853 | { | |
5854 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5855 | int val; | |
5856 | ||
5857 | if (sscanf(buf, "%d", &val) != 1 || val <= 0) | |
5858 | return -EINVAL; | |
5859 | ||
5860 | workqueue_set_max_active(wq, val); | |
5861 | return count; | |
5862 | } | |
5863 | static DEVICE_ATTR_RW(max_active); | |
5864 | ||
5865 | static struct attribute *wq_sysfs_attrs[] = { | |
5866 | &dev_attr_per_cpu.attr, | |
5867 | &dev_attr_max_active.attr, | |
5868 | NULL, | |
5869 | }; | |
5870 | ATTRIBUTE_GROUPS(wq_sysfs); | |
5871 | ||
5872 | static ssize_t wq_pool_ids_show(struct device *dev, | |
5873 | struct device_attribute *attr, char *buf) | |
5874 | { | |
5875 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5876 | const char *delim = ""; | |
5877 | int node, written = 0; | |
5878 | ||
ffd8bea8 | 5879 | cpus_read_lock(); |
24acfb71 | 5880 | rcu_read_lock(); |
6ba94429 FW |
5881 | for_each_node(node) { |
5882 | written += scnprintf(buf + written, PAGE_SIZE - written, | |
5883 | "%s%d:%d", delim, node, | |
5884 | unbound_pwq_by_node(wq, node)->pool->id); | |
5885 | delim = " "; | |
5886 | } | |
5887 | written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); | |
24acfb71 | 5888 | rcu_read_unlock(); |
ffd8bea8 | 5889 | cpus_read_unlock(); |
6ba94429 FW |
5890 | |
5891 | return written; | |
5892 | } | |
5893 | ||
5894 | static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, | |
5895 | char *buf) | |
5896 | { | |
5897 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5898 | int written; | |
5899 | ||
5900 | mutex_lock(&wq->mutex); | |
5901 | written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); | |
5902 | mutex_unlock(&wq->mutex); | |
5903 | ||
5904 | return written; | |
5905 | } | |
5906 | ||
5907 | /* prepare workqueue_attrs for sysfs store operations */ | |
5908 | static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) | |
5909 | { | |
5910 | struct workqueue_attrs *attrs; | |
5911 | ||
899a94fe LJ |
5912 | lockdep_assert_held(&wq_pool_mutex); |
5913 | ||
be69d00d | 5914 | attrs = alloc_workqueue_attrs(); |
6ba94429 FW |
5915 | if (!attrs) |
5916 | return NULL; | |
5917 | ||
6ba94429 | 5918 | copy_workqueue_attrs(attrs, wq->unbound_attrs); |
6ba94429 FW |
5919 | return attrs; |
5920 | } | |
5921 | ||
5922 | static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, | |
5923 | const char *buf, size_t count) | |
5924 | { | |
5925 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5926 | struct workqueue_attrs *attrs; | |
d4d3e257 LJ |
5927 | int ret = -ENOMEM; |
5928 | ||
5929 | apply_wqattrs_lock(); | |
6ba94429 FW |
5930 | |
5931 | attrs = wq_sysfs_prep_attrs(wq); | |
5932 | if (!attrs) | |
d4d3e257 | 5933 | goto out_unlock; |
6ba94429 FW |
5934 | |
5935 | if (sscanf(buf, "%d", &attrs->nice) == 1 && | |
5936 | attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) | |
d4d3e257 | 5937 | ret = apply_workqueue_attrs_locked(wq, attrs); |
6ba94429 FW |
5938 | else |
5939 | ret = -EINVAL; | |
5940 | ||
d4d3e257 LJ |
5941 | out_unlock: |
5942 | apply_wqattrs_unlock(); | |
6ba94429 FW |
5943 | free_workqueue_attrs(attrs); |
5944 | return ret ?: count; | |
5945 | } | |
5946 | ||
5947 | static ssize_t wq_cpumask_show(struct device *dev, | |
5948 | struct device_attribute *attr, char *buf) | |
5949 | { | |
5950 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5951 | int written; | |
5952 | ||
5953 | mutex_lock(&wq->mutex); | |
5954 | written = scnprintf(buf, PAGE_SIZE, "%*pb\n", | |
5955 | cpumask_pr_args(wq->unbound_attrs->cpumask)); | |
5956 | mutex_unlock(&wq->mutex); | |
5957 | return written; | |
5958 | } | |
5959 | ||
5960 | static ssize_t wq_cpumask_store(struct device *dev, | |
5961 | struct device_attribute *attr, | |
5962 | const char *buf, size_t count) | |
5963 | { | |
5964 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5965 | struct workqueue_attrs *attrs; | |
d4d3e257 LJ |
5966 | int ret = -ENOMEM; |
5967 | ||
5968 | apply_wqattrs_lock(); | |
6ba94429 FW |
5969 | |
5970 | attrs = wq_sysfs_prep_attrs(wq); | |
5971 | if (!attrs) | |
d4d3e257 | 5972 | goto out_unlock; |
6ba94429 FW |
5973 | |
5974 | ret = cpumask_parse(buf, attrs->cpumask); | |
5975 | if (!ret) | |
d4d3e257 | 5976 | ret = apply_workqueue_attrs_locked(wq, attrs); |
6ba94429 | 5977 | |
d4d3e257 LJ |
5978 | out_unlock: |
5979 | apply_wqattrs_unlock(); | |
6ba94429 FW |
5980 | free_workqueue_attrs(attrs); |
5981 | return ret ?: count; | |
5982 | } | |
5983 | ||
5984 | static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr, | |
5985 | char *buf) | |
5986 | { | |
5987 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5988 | int written; | |
7dbc725e | 5989 | |
6ba94429 FW |
5990 | mutex_lock(&wq->mutex); |
5991 | written = scnprintf(buf, PAGE_SIZE, "%d\n", | |
5992 | !wq->unbound_attrs->no_numa); | |
5993 | mutex_unlock(&wq->mutex); | |
4c16bd32 | 5994 | |
6ba94429 | 5995 | return written; |
65758202 TH |
5996 | } |
5997 | ||
6ba94429 FW |
5998 | static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr, |
5999 | const char *buf, size_t count) | |
65758202 | 6000 | { |
6ba94429 FW |
6001 | struct workqueue_struct *wq = dev_to_wq(dev); |
6002 | struct workqueue_attrs *attrs; | |
d4d3e257 LJ |
6003 | int v, ret = -ENOMEM; |
6004 | ||
6005 | apply_wqattrs_lock(); | |
4c16bd32 | 6006 | |
6ba94429 FW |
6007 | attrs = wq_sysfs_prep_attrs(wq); |
6008 | if (!attrs) | |
d4d3e257 | 6009 | goto out_unlock; |
4c16bd32 | 6010 | |
6ba94429 FW |
6011 | ret = -EINVAL; |
6012 | if (sscanf(buf, "%d", &v) == 1) { | |
6013 | attrs->no_numa = !v; | |
d4d3e257 | 6014 | ret = apply_workqueue_attrs_locked(wq, attrs); |
65758202 | 6015 | } |
6ba94429 | 6016 | |
d4d3e257 LJ |
6017 | out_unlock: |
6018 | apply_wqattrs_unlock(); | |
6ba94429 FW |
6019 | free_workqueue_attrs(attrs); |
6020 | return ret ?: count; | |
65758202 TH |
6021 | } |
6022 | ||
6ba94429 FW |
6023 | static struct device_attribute wq_sysfs_unbound_attrs[] = { |
6024 | __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL), | |
6025 | __ATTR(nice, 0644, wq_nice_show, wq_nice_store), | |
6026 | __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), | |
6027 | __ATTR(numa, 0644, wq_numa_show, wq_numa_store), | |
6028 | __ATTR_NULL, | |
6029 | }; | |
8ccad40d | 6030 | |
6ba94429 FW |
6031 | static struct bus_type wq_subsys = { |
6032 | .name = "workqueue", | |
6033 | .dev_groups = wq_sysfs_groups, | |
2d3854a3 RR |
6034 | }; |
6035 | ||
b05a7928 FW |
6036 | static ssize_t wq_unbound_cpumask_show(struct device *dev, |
6037 | struct device_attribute *attr, char *buf) | |
6038 | { | |
6039 | int written; | |
6040 | ||
042f7df1 | 6041 | mutex_lock(&wq_pool_mutex); |
b05a7928 FW |
6042 | written = scnprintf(buf, PAGE_SIZE, "%*pb\n", |
6043 | cpumask_pr_args(wq_unbound_cpumask)); | |
042f7df1 | 6044 | mutex_unlock(&wq_pool_mutex); |
b05a7928 FW |
6045 | |
6046 | return written; | |
6047 | } | |
6048 | ||
042f7df1 LJ |
6049 | static ssize_t wq_unbound_cpumask_store(struct device *dev, |
6050 | struct device_attribute *attr, const char *buf, size_t count) | |
6051 | { | |
6052 | cpumask_var_t cpumask; | |
6053 | int ret; | |
6054 | ||
6055 | if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) | |
6056 | return -ENOMEM; | |
6057 | ||
6058 | ret = cpumask_parse(buf, cpumask); | |
6059 | if (!ret) | |
6060 | ret = workqueue_set_unbound_cpumask(cpumask); | |
6061 | ||
6062 | free_cpumask_var(cpumask); | |
6063 | return ret ? ret : count; | |
6064 | } | |
6065 | ||
b05a7928 | 6066 | static struct device_attribute wq_sysfs_cpumask_attr = |
042f7df1 LJ |
6067 | __ATTR(cpumask, 0644, wq_unbound_cpumask_show, |
6068 | wq_unbound_cpumask_store); | |
b05a7928 | 6069 | |
6ba94429 | 6070 | static int __init wq_sysfs_init(void) |
2d3854a3 | 6071 | { |
686f6697 | 6072 | struct device *dev_root; |
b05a7928 FW |
6073 | int err; |
6074 | ||
6075 | err = subsys_virtual_register(&wq_subsys, NULL); | |
6076 | if (err) | |
6077 | return err; | |
6078 | ||
686f6697 GKH |
6079 | dev_root = bus_get_dev_root(&wq_subsys); |
6080 | if (dev_root) { | |
6081 | err = device_create_file(dev_root, &wq_sysfs_cpumask_attr); | |
6082 | put_device(dev_root); | |
6083 | } | |
6084 | return err; | |
2d3854a3 | 6085 | } |
6ba94429 | 6086 | core_initcall(wq_sysfs_init); |
2d3854a3 | 6087 | |
6ba94429 | 6088 | static void wq_device_release(struct device *dev) |
2d3854a3 | 6089 | { |
6ba94429 | 6090 | struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); |
6b44003e | 6091 | |
6ba94429 | 6092 | kfree(wq_dev); |
2d3854a3 | 6093 | } |
a0a1a5fd TH |
6094 | |
6095 | /** | |
6ba94429 FW |
6096 | * workqueue_sysfs_register - make a workqueue visible in sysfs |
6097 | * @wq: the workqueue to register | |
a0a1a5fd | 6098 | * |
6ba94429 FW |
6099 | * Expose @wq in sysfs under /sys/bus/workqueue/devices. |
6100 | * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set | |
6101 | * which is the preferred method. | |
a0a1a5fd | 6102 | * |
6ba94429 FW |
6103 | * Workqueue user should use this function directly iff it wants to apply |
6104 | * workqueue_attrs before making the workqueue visible in sysfs; otherwise, | |
6105 | * apply_workqueue_attrs() may race against userland updating the | |
6106 | * attributes. | |
6107 | * | |
6108 | * Return: 0 on success, -errno on failure. | |
a0a1a5fd | 6109 | */ |
6ba94429 | 6110 | int workqueue_sysfs_register(struct workqueue_struct *wq) |
a0a1a5fd | 6111 | { |
6ba94429 FW |
6112 | struct wq_device *wq_dev; |
6113 | int ret; | |
a0a1a5fd | 6114 | |
6ba94429 | 6115 | /* |
402dd89d | 6116 | * Adjusting max_active or creating new pwqs by applying |
6ba94429 FW |
6117 | * attributes breaks ordering guarantee. Disallow exposing ordered |
6118 | * workqueues. | |
6119 | */ | |
0a94efb5 | 6120 | if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) |
6ba94429 | 6121 | return -EINVAL; |
a0a1a5fd | 6122 | |
6ba94429 FW |
6123 | wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); |
6124 | if (!wq_dev) | |
6125 | return -ENOMEM; | |
5bcab335 | 6126 | |
6ba94429 FW |
6127 | wq_dev->wq = wq; |
6128 | wq_dev->dev.bus = &wq_subsys; | |
6ba94429 | 6129 | wq_dev->dev.release = wq_device_release; |
23217b44 | 6130 | dev_set_name(&wq_dev->dev, "%s", wq->name); |
a0a1a5fd | 6131 | |
6ba94429 FW |
6132 | /* |
6133 | * unbound_attrs are created separately. Suppress uevent until | |
6134 | * everything is ready. | |
6135 | */ | |
6136 | dev_set_uevent_suppress(&wq_dev->dev, true); | |
a0a1a5fd | 6137 | |
6ba94429 FW |
6138 | ret = device_register(&wq_dev->dev); |
6139 | if (ret) { | |
537f4146 | 6140 | put_device(&wq_dev->dev); |
6ba94429 FW |
6141 | wq->wq_dev = NULL; |
6142 | return ret; | |
6143 | } | |
a0a1a5fd | 6144 | |
6ba94429 FW |
6145 | if (wq->flags & WQ_UNBOUND) { |
6146 | struct device_attribute *attr; | |
a0a1a5fd | 6147 | |
6ba94429 FW |
6148 | for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { |
6149 | ret = device_create_file(&wq_dev->dev, attr); | |
6150 | if (ret) { | |
6151 | device_unregister(&wq_dev->dev); | |
6152 | wq->wq_dev = NULL; | |
6153 | return ret; | |
a0a1a5fd TH |
6154 | } |
6155 | } | |
6156 | } | |
6ba94429 FW |
6157 | |
6158 | dev_set_uevent_suppress(&wq_dev->dev, false); | |
6159 | kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); | |
6160 | return 0; | |
a0a1a5fd TH |
6161 | } |
6162 | ||
6163 | /** | |
6ba94429 FW |
6164 | * workqueue_sysfs_unregister - undo workqueue_sysfs_register() |
6165 | * @wq: the workqueue to unregister | |
a0a1a5fd | 6166 | * |
6ba94429 | 6167 | * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. |
a0a1a5fd | 6168 | */ |
6ba94429 | 6169 | static void workqueue_sysfs_unregister(struct workqueue_struct *wq) |
a0a1a5fd | 6170 | { |
6ba94429 | 6171 | struct wq_device *wq_dev = wq->wq_dev; |
8b03ae3c | 6172 | |
6ba94429 FW |
6173 | if (!wq->wq_dev) |
6174 | return; | |
a0a1a5fd | 6175 | |
6ba94429 FW |
6176 | wq->wq_dev = NULL; |
6177 | device_unregister(&wq_dev->dev); | |
a0a1a5fd | 6178 | } |
6ba94429 FW |
6179 | #else /* CONFIG_SYSFS */ |
6180 | static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } | |
6181 | #endif /* CONFIG_SYSFS */ | |
a0a1a5fd | 6182 | |
82607adc TH |
6183 | /* |
6184 | * Workqueue watchdog. | |
6185 | * | |
6186 | * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal | |
6187 | * flush dependency, a concurrency managed work item which stays RUNNING | |
6188 | * indefinitely. Workqueue stalls can be very difficult to debug as the | |
6189 | * usual warning mechanisms don't trigger and internal workqueue state is | |
6190 | * largely opaque. | |
6191 | * | |
6192 | * Workqueue watchdog monitors all worker pools periodically and dumps | |
6193 | * state if some pools failed to make forward progress for a while where | |
6194 | * forward progress is defined as the first item on ->worklist changing. | |
6195 | * | |
6196 | * This mechanism is controlled through the kernel parameter | |
6197 | * "workqueue.watchdog_thresh" which can be updated at runtime through the | |
6198 | * corresponding sysfs parameter file. | |
6199 | */ | |
6200 | #ifdef CONFIG_WQ_WATCHDOG | |
6201 | ||
82607adc | 6202 | static unsigned long wq_watchdog_thresh = 30; |
5cd79d6a | 6203 | static struct timer_list wq_watchdog_timer; |
82607adc TH |
6204 | |
6205 | static unsigned long wq_watchdog_touched = INITIAL_JIFFIES; | |
6206 | static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES; | |
6207 | ||
cd2440d6 PM |
6208 | /* |
6209 | * Show workers that might prevent the processing of pending work items. | |
6210 | * The only candidates are CPU-bound workers in the running state. | |
6211 | * Pending work items should be handled by another idle worker | |
6212 | * in all other situations. | |
6213 | */ | |
6214 | static void show_cpu_pool_hog(struct worker_pool *pool) | |
6215 | { | |
6216 | struct worker *worker; | |
6217 | unsigned long flags; | |
6218 | int bkt; | |
6219 | ||
6220 | raw_spin_lock_irqsave(&pool->lock, flags); | |
6221 | ||
6222 | hash_for_each(pool->busy_hash, bkt, worker, hentry) { | |
6223 | if (task_is_running(worker->task)) { | |
6224 | /* | |
6225 | * Defer printing to avoid deadlocks in console | |
6226 | * drivers that queue work while holding locks | |
6227 | * also taken in their write paths. | |
6228 | */ | |
6229 | printk_deferred_enter(); | |
6230 | ||
6231 | pr_info("pool %d:\n", pool->id); | |
6232 | sched_show_task(worker->task); | |
6233 | ||
6234 | printk_deferred_exit(); | |
6235 | } | |
6236 | } | |
6237 | ||
6238 | raw_spin_unlock_irqrestore(&pool->lock, flags); | |
6239 | } | |
6240 | ||
6241 | static void show_cpu_pools_hogs(void) | |
6242 | { | |
6243 | struct worker_pool *pool; | |
6244 | int pi; | |
6245 | ||
6246 | pr_info("Showing backtraces of running workers in stalled CPU-bound worker pools:\n"); | |
6247 | ||
6248 | rcu_read_lock(); | |
6249 | ||
6250 | for_each_pool(pool, pi) { | |
6251 | if (pool->cpu_stall) | |
6252 | show_cpu_pool_hog(pool); | |
6253 | ||
6254 | } | |
6255 | ||
6256 | rcu_read_unlock(); | |
6257 | } | |
6258 | ||
82607adc TH |
6259 | static void wq_watchdog_reset_touched(void) |
6260 | { | |
6261 | int cpu; | |
6262 | ||
6263 | wq_watchdog_touched = jiffies; | |
6264 | for_each_possible_cpu(cpu) | |
6265 | per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; | |
6266 | } | |
6267 | ||
5cd79d6a | 6268 | static void wq_watchdog_timer_fn(struct timer_list *unused) |
82607adc TH |
6269 | { |
6270 | unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ; | |
6271 | bool lockup_detected = false; | |
cd2440d6 | 6272 | bool cpu_pool_stall = false; |
940d71c6 | 6273 | unsigned long now = jiffies; |
82607adc TH |
6274 | struct worker_pool *pool; |
6275 | int pi; | |
6276 | ||
6277 | if (!thresh) | |
6278 | return; | |
6279 | ||
6280 | rcu_read_lock(); | |
6281 | ||
6282 | for_each_pool(pool, pi) { | |
6283 | unsigned long pool_ts, touched, ts; | |
6284 | ||
cd2440d6 | 6285 | pool->cpu_stall = false; |
82607adc TH |
6286 | if (list_empty(&pool->worklist)) |
6287 | continue; | |
6288 | ||
940d71c6 SS |
6289 | /* |
6290 | * If a virtual machine is stopped by the host it can look to | |
6291 | * the watchdog like a stall. | |
6292 | */ | |
6293 | kvm_check_and_clear_guest_paused(); | |
6294 | ||
82607adc | 6295 | /* get the latest of pool and touched timestamps */ |
89e28ce6 WQ |
6296 | if (pool->cpu >= 0) |
6297 | touched = READ_ONCE(per_cpu(wq_watchdog_touched_cpu, pool->cpu)); | |
6298 | else | |
6299 | touched = READ_ONCE(wq_watchdog_touched); | |
82607adc | 6300 | pool_ts = READ_ONCE(pool->watchdog_ts); |
82607adc TH |
6301 | |
6302 | if (time_after(pool_ts, touched)) | |
6303 | ts = pool_ts; | |
6304 | else | |
6305 | ts = touched; | |
6306 | ||
82607adc | 6307 | /* did we stall? */ |
940d71c6 | 6308 | if (time_after(now, ts + thresh)) { |
82607adc | 6309 | lockup_detected = true; |
cd2440d6 PM |
6310 | if (pool->cpu >= 0) { |
6311 | pool->cpu_stall = true; | |
6312 | cpu_pool_stall = true; | |
6313 | } | |
82607adc TH |
6314 | pr_emerg("BUG: workqueue lockup - pool"); |
6315 | pr_cont_pool_info(pool); | |
6316 | pr_cont(" stuck for %us!\n", | |
940d71c6 | 6317 | jiffies_to_msecs(now - pool_ts) / 1000); |
82607adc | 6318 | } |
cd2440d6 PM |
6319 | |
6320 | ||
82607adc TH |
6321 | } |
6322 | ||
6323 | rcu_read_unlock(); | |
6324 | ||
6325 | if (lockup_detected) | |
55df0933 | 6326 | show_all_workqueues(); |
82607adc | 6327 | |
cd2440d6 PM |
6328 | if (cpu_pool_stall) |
6329 | show_cpu_pools_hogs(); | |
6330 | ||
82607adc TH |
6331 | wq_watchdog_reset_touched(); |
6332 | mod_timer(&wq_watchdog_timer, jiffies + thresh); | |
6333 | } | |
6334 | ||
cb9d7fd5 | 6335 | notrace void wq_watchdog_touch(int cpu) |
82607adc TH |
6336 | { |
6337 | if (cpu >= 0) | |
6338 | per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; | |
89e28ce6 WQ |
6339 | |
6340 | wq_watchdog_touched = jiffies; | |
82607adc TH |
6341 | } |
6342 | ||
6343 | static void wq_watchdog_set_thresh(unsigned long thresh) | |
6344 | { | |
6345 | wq_watchdog_thresh = 0; | |
6346 | del_timer_sync(&wq_watchdog_timer); | |
6347 | ||
6348 | if (thresh) { | |
6349 | wq_watchdog_thresh = thresh; | |
6350 | wq_watchdog_reset_touched(); | |
6351 | mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ); | |
6352 | } | |
6353 | } | |
6354 | ||
6355 | static int wq_watchdog_param_set_thresh(const char *val, | |
6356 | const struct kernel_param *kp) | |
6357 | { | |
6358 | unsigned long thresh; | |
6359 | int ret; | |
6360 | ||
6361 | ret = kstrtoul(val, 0, &thresh); | |
6362 | if (ret) | |
6363 | return ret; | |
6364 | ||
6365 | if (system_wq) | |
6366 | wq_watchdog_set_thresh(thresh); | |
6367 | else | |
6368 | wq_watchdog_thresh = thresh; | |
6369 | ||
6370 | return 0; | |
6371 | } | |
6372 | ||
6373 | static const struct kernel_param_ops wq_watchdog_thresh_ops = { | |
6374 | .set = wq_watchdog_param_set_thresh, | |
6375 | .get = param_get_ulong, | |
6376 | }; | |
6377 | ||
6378 | module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh, | |
6379 | 0644); | |
6380 | ||
6381 | static void wq_watchdog_init(void) | |
6382 | { | |
5cd79d6a | 6383 | timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE); |
82607adc TH |
6384 | wq_watchdog_set_thresh(wq_watchdog_thresh); |
6385 | } | |
6386 | ||
6387 | #else /* CONFIG_WQ_WATCHDOG */ | |
6388 | ||
6389 | static inline void wq_watchdog_init(void) { } | |
6390 | ||
6391 | #endif /* CONFIG_WQ_WATCHDOG */ | |
6392 | ||
bce90380 TH |
6393 | static void __init wq_numa_init(void) |
6394 | { | |
6395 | cpumask_var_t *tbl; | |
6396 | int node, cpu; | |
6397 | ||
bce90380 TH |
6398 | if (num_possible_nodes() <= 1) |
6399 | return; | |
6400 | ||
d55262c4 TH |
6401 | if (wq_disable_numa) { |
6402 | pr_info("workqueue: NUMA affinity support disabled\n"); | |
6403 | return; | |
6404 | } | |
6405 | ||
f728c4a9 ZL |
6406 | for_each_possible_cpu(cpu) { |
6407 | if (WARN_ON(cpu_to_node(cpu) == NUMA_NO_NODE)) { | |
6408 | pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu); | |
6409 | return; | |
6410 | } | |
6411 | } | |
6412 | ||
be69d00d | 6413 | wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(); |
4c16bd32 TH |
6414 | BUG_ON(!wq_update_unbound_numa_attrs_buf); |
6415 | ||
bce90380 TH |
6416 | /* |
6417 | * We want masks of possible CPUs of each node which isn't readily | |
6418 | * available. Build one from cpu_to_node() which should have been | |
6419 | * fully initialized by now. | |
6420 | */ | |
6396bb22 | 6421 | tbl = kcalloc(nr_node_ids, sizeof(tbl[0]), GFP_KERNEL); |
bce90380 TH |
6422 | BUG_ON(!tbl); |
6423 | ||
6424 | for_each_node(node) | |
5a6024f1 | 6425 | BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL, |
1be0c25d | 6426 | node_online(node) ? node : NUMA_NO_NODE)); |
bce90380 TH |
6427 | |
6428 | for_each_possible_cpu(cpu) { | |
6429 | node = cpu_to_node(cpu); | |
bce90380 TH |
6430 | cpumask_set_cpu(cpu, tbl[node]); |
6431 | } | |
6432 | ||
6433 | wq_numa_possible_cpumask = tbl; | |
6434 | wq_numa_enabled = true; | |
6435 | } | |
6436 | ||
3347fa09 TH |
6437 | /** |
6438 | * workqueue_init_early - early init for workqueue subsystem | |
6439 | * | |
6440 | * This is the first half of two-staged workqueue subsystem initialization | |
6441 | * and invoked as soon as the bare basics - memory allocation, cpumasks and | |
6442 | * idr are up. It sets up all the data structures and system workqueues | |
6443 | * and allows early boot code to create workqueues and queue/cancel work | |
6444 | * items. Actual work item execution starts only after kthreads can be | |
6445 | * created and scheduled right before early initcalls. | |
6446 | */ | |
2333e829 | 6447 | void __init workqueue_init_early(void) |
1da177e4 | 6448 | { |
7a4e344c TH |
6449 | int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; |
6450 | int i, cpu; | |
c34056a3 | 6451 | |
10cdb157 | 6452 | BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); |
e904e6c2 | 6453 | |
b05a7928 | 6454 | BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL)); |
04d4e665 FW |
6455 | cpumask_copy(wq_unbound_cpumask, housekeeping_cpumask(HK_TYPE_WQ)); |
6456 | cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, housekeeping_cpumask(HK_TYPE_DOMAIN)); | |
b05a7928 | 6457 | |
e904e6c2 TH |
6458 | pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); |
6459 | ||
706026c2 | 6460 | /* initialize CPU pools */ |
29c91e99 | 6461 | for_each_possible_cpu(cpu) { |
4ce62e9e | 6462 | struct worker_pool *pool; |
8b03ae3c | 6463 | |
7a4e344c | 6464 | i = 0; |
f02ae73a | 6465 | for_each_cpu_worker_pool(pool, cpu) { |
7a4e344c | 6466 | BUG_ON(init_worker_pool(pool)); |
ec22ca5e | 6467 | pool->cpu = cpu; |
29c91e99 | 6468 | cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); |
7a4e344c | 6469 | pool->attrs->nice = std_nice[i++]; |
f3f90ad4 | 6470 | pool->node = cpu_to_node(cpu); |
7a4e344c | 6471 | |
9daf9e67 | 6472 | /* alloc pool ID */ |
68e13a67 | 6473 | mutex_lock(&wq_pool_mutex); |
9daf9e67 | 6474 | BUG_ON(worker_pool_assign_id(pool)); |
68e13a67 | 6475 | mutex_unlock(&wq_pool_mutex); |
4ce62e9e | 6476 | } |
8b03ae3c TH |
6477 | } |
6478 | ||
8a2b7538 | 6479 | /* create default unbound and ordered wq attrs */ |
29c91e99 TH |
6480 | for (i = 0; i < NR_STD_WORKER_POOLS; i++) { |
6481 | struct workqueue_attrs *attrs; | |
6482 | ||
be69d00d | 6483 | BUG_ON(!(attrs = alloc_workqueue_attrs())); |
29c91e99 | 6484 | attrs->nice = std_nice[i]; |
29c91e99 | 6485 | unbound_std_wq_attrs[i] = attrs; |
8a2b7538 TH |
6486 | |
6487 | /* | |
6488 | * An ordered wq should have only one pwq as ordering is | |
6489 | * guaranteed by max_active which is enforced by pwqs. | |
6490 | * Turn off NUMA so that dfl_pwq is used for all nodes. | |
6491 | */ | |
be69d00d | 6492 | BUG_ON(!(attrs = alloc_workqueue_attrs())); |
8a2b7538 TH |
6493 | attrs->nice = std_nice[i]; |
6494 | attrs->no_numa = true; | |
6495 | ordered_wq_attrs[i] = attrs; | |
29c91e99 TH |
6496 | } |
6497 | ||
d320c038 | 6498 | system_wq = alloc_workqueue("events", 0, 0); |
1aabe902 | 6499 | system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); |
d320c038 | 6500 | system_long_wq = alloc_workqueue("events_long", 0, 0); |
f3421797 TH |
6501 | system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, |
6502 | WQ_UNBOUND_MAX_ACTIVE); | |
24d51add TH |
6503 | system_freezable_wq = alloc_workqueue("events_freezable", |
6504 | WQ_FREEZABLE, 0); | |
0668106c VK |
6505 | system_power_efficient_wq = alloc_workqueue("events_power_efficient", |
6506 | WQ_POWER_EFFICIENT, 0); | |
6507 | system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient", | |
6508 | WQ_FREEZABLE | WQ_POWER_EFFICIENT, | |
6509 | 0); | |
1aabe902 | 6510 | BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || |
0668106c VK |
6511 | !system_unbound_wq || !system_freezable_wq || |
6512 | !system_power_efficient_wq || | |
6513 | !system_freezable_power_efficient_wq); | |
3347fa09 TH |
6514 | } |
6515 | ||
6516 | /** | |
6517 | * workqueue_init - bring workqueue subsystem fully online | |
6518 | * | |
6519 | * This is the latter half of two-staged workqueue subsystem initialization | |
6520 | * and invoked as soon as kthreads can be created and scheduled. | |
6521 | * Workqueues have been created and work items queued on them, but there | |
6522 | * are no kworkers executing the work items yet. Populate the worker pools | |
6523 | * with the initial workers and enable future kworker creations. | |
6524 | */ | |
2333e829 | 6525 | void __init workqueue_init(void) |
3347fa09 | 6526 | { |
2186d9f9 | 6527 | struct workqueue_struct *wq; |
3347fa09 TH |
6528 | struct worker_pool *pool; |
6529 | int cpu, bkt; | |
6530 | ||
2186d9f9 TH |
6531 | /* |
6532 | * It'd be simpler to initialize NUMA in workqueue_init_early() but | |
6533 | * CPU to node mapping may not be available that early on some | |
6534 | * archs such as power and arm64. As per-cpu pools created | |
6535 | * previously could be missing node hint and unbound pools NUMA | |
6536 | * affinity, fix them up. | |
40c17f75 TH |
6537 | * |
6538 | * Also, while iterating workqueues, create rescuers if requested. | |
2186d9f9 TH |
6539 | */ |
6540 | wq_numa_init(); | |
6541 | ||
6542 | mutex_lock(&wq_pool_mutex); | |
6543 | ||
6544 | for_each_possible_cpu(cpu) { | |
6545 | for_each_cpu_worker_pool(pool, cpu) { | |
6546 | pool->node = cpu_to_node(cpu); | |
6547 | } | |
6548 | } | |
6549 | ||
40c17f75 | 6550 | list_for_each_entry(wq, &workqueues, list) { |
2186d9f9 | 6551 | wq_update_unbound_numa(wq, smp_processor_id(), true); |
40c17f75 TH |
6552 | WARN(init_rescuer(wq), |
6553 | "workqueue: failed to create early rescuer for %s", | |
6554 | wq->name); | |
6555 | } | |
2186d9f9 TH |
6556 | |
6557 | mutex_unlock(&wq_pool_mutex); | |
6558 | ||
3347fa09 TH |
6559 | /* create the initial workers */ |
6560 | for_each_online_cpu(cpu) { | |
6561 | for_each_cpu_worker_pool(pool, cpu) { | |
6562 | pool->flags &= ~POOL_DISASSOCIATED; | |
6563 | BUG_ON(!create_worker(pool)); | |
6564 | } | |
6565 | } | |
6566 | ||
6567 | hash_for_each(unbound_pool_hash, bkt, pool, hash_node) | |
6568 | BUG_ON(!create_worker(pool)); | |
6569 | ||
6570 | wq_online = true; | |
82607adc | 6571 | wq_watchdog_init(); |
1da177e4 | 6572 | } |
c4f135d6 TH |
6573 | |
6574 | /* | |
6575 | * Despite the naming, this is a no-op function which is here only for avoiding | |
6576 | * link error. Since compile-time warning may fail to catch, we will need to | |
6577 | * emit run-time warning from __flush_workqueue(). | |
6578 | */ | |
6579 | void __warn_flushing_systemwide_wq(void) { } | |
6580 | EXPORT_SYMBOL(__warn_flushing_systemwide_wq); |