Commit | Line | Data |
---|---|---|
457c8996 | 1 | // SPDX-License-Identifier: GPL-2.0-only |
1da177e4 | 2 | /* |
c54fce6e | 3 | * kernel/workqueue.c - generic async execution with shared worker pool |
1da177e4 | 4 | * |
c54fce6e | 5 | * Copyright (C) 2002 Ingo Molnar |
1da177e4 | 6 | * |
c54fce6e TH |
7 | * Derived from the taskqueue/keventd code by: |
8 | * David Woodhouse <dwmw2@infradead.org> | |
9 | * Andrew Morton | |
10 | * Kai Petzke <wpp@marie.physik.tu-berlin.de> | |
11 | * Theodore Ts'o <tytso@mit.edu> | |
1da177e4 | 12 | * |
c54fce6e | 13 | * Made to use alloc_percpu by Christoph Lameter. |
1da177e4 | 14 | * |
c54fce6e TH |
15 | * Copyright (C) 2010 SUSE Linux Products GmbH |
16 | * Copyright (C) 2010 Tejun Heo <tj@kernel.org> | |
89ada679 | 17 | * |
c54fce6e TH |
18 | * This is the generic async execution mechanism. Work items as are |
19 | * executed in process context. The worker pool is shared and | |
b11895c4 L |
20 | * automatically managed. There are two worker pools for each CPU (one for |
21 | * normal work items and the other for high priority ones) and some extra | |
22 | * pools for workqueues which are not bound to any specific CPU - the | |
23 | * number of these backing pools is dynamic. | |
c54fce6e | 24 | * |
9a261491 | 25 | * Please read Documentation/core-api/workqueue.rst for details. |
1da177e4 LT |
26 | */ |
27 | ||
9984de1a | 28 | #include <linux/export.h> |
1da177e4 LT |
29 | #include <linux/kernel.h> |
30 | #include <linux/sched.h> | |
31 | #include <linux/init.h> | |
32 | #include <linux/signal.h> | |
33 | #include <linux/completion.h> | |
34 | #include <linux/workqueue.h> | |
35 | #include <linux/slab.h> | |
36 | #include <linux/cpu.h> | |
37 | #include <linux/notifier.h> | |
38 | #include <linux/kthread.h> | |
1fa44eca | 39 | #include <linux/hardirq.h> |
46934023 | 40 | #include <linux/mempolicy.h> |
341a5958 | 41 | #include <linux/freezer.h> |
d5abe669 | 42 | #include <linux/debug_locks.h> |
4e6045f1 | 43 | #include <linux/lockdep.h> |
c34056a3 | 44 | #include <linux/idr.h> |
29c91e99 | 45 | #include <linux/jhash.h> |
42f8570f | 46 | #include <linux/hashtable.h> |
76af4d93 | 47 | #include <linux/rculist.h> |
bce90380 | 48 | #include <linux/nodemask.h> |
4c16bd32 | 49 | #include <linux/moduleparam.h> |
3d1cb205 | 50 | #include <linux/uaccess.h> |
c98a9805 | 51 | #include <linux/sched/isolation.h> |
cd2440d6 | 52 | #include <linux/sched/debug.h> |
62635ea8 | 53 | #include <linux/nmi.h> |
940d71c6 | 54 | #include <linux/kvm_para.h> |
aa6fde93 | 55 | #include <linux/delay.h> |
e22bee78 | 56 | |
ea138446 | 57 | #include "workqueue_internal.h" |
1da177e4 | 58 | |
c8e55f36 | 59 | enum { |
24647570 TH |
60 | /* |
61 | * worker_pool flags | |
bc2ae0f5 | 62 | * |
24647570 | 63 | * A bound pool is either associated or disassociated with its CPU. |
bc2ae0f5 TH |
64 | * While associated (!DISASSOCIATED), all workers are bound to the |
65 | * CPU and none has %WORKER_UNBOUND set and concurrency management | |
66 | * is in effect. | |
67 | * | |
68 | * While DISASSOCIATED, the cpu may be offline and all workers have | |
69 | * %WORKER_UNBOUND set and concurrency management disabled, and may | |
24647570 | 70 | * be executing on any CPU. The pool behaves as an unbound one. |
bc2ae0f5 | 71 | * |
bc3a1afc | 72 | * Note that DISASSOCIATED should be flipped only while holding |
1258fae7 | 73 | * wq_pool_attach_mutex to avoid changing binding state while |
4736cbf7 | 74 | * worker_attach_to_pool() is in progress. |
bc2ae0f5 | 75 | */ |
692b4825 | 76 | POOL_MANAGER_ACTIVE = 1 << 0, /* being managed */ |
24647570 | 77 | POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */ |
db7bccf4 | 78 | |
c8e55f36 | 79 | /* worker flags */ |
c8e55f36 TH |
80 | WORKER_DIE = 1 << 1, /* die die die */ |
81 | WORKER_IDLE = 1 << 2, /* is idle */ | |
e22bee78 | 82 | WORKER_PREP = 1 << 3, /* preparing to run works */ |
fb0e7beb | 83 | WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */ |
f3421797 | 84 | WORKER_UNBOUND = 1 << 7, /* worker is unbound */ |
a9ab775b | 85 | WORKER_REBOUND = 1 << 8, /* worker was rebound */ |
e22bee78 | 86 | |
a9ab775b TH |
87 | WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE | |
88 | WORKER_UNBOUND | WORKER_REBOUND, | |
db7bccf4 | 89 | |
e34cdddb | 90 | NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */ |
4ce62e9e | 91 | |
29c91e99 | 92 | UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */ |
c8e55f36 | 93 | BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */ |
db7bccf4 | 94 | |
e22bee78 TH |
95 | MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */ |
96 | IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */ | |
97 | ||
3233cdbd TH |
98 | MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2, |
99 | /* call for help after 10ms | |
100 | (min two ticks) */ | |
e22bee78 TH |
101 | MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */ |
102 | CREATE_COOLDOWN = HZ, /* time to breath after fail */ | |
e22bee78 TH |
103 | |
104 | /* | |
105 | * Rescue workers are used only on emergencies and shared by | |
8698a745 | 106 | * all cpus. Give MIN_NICE. |
e22bee78 | 107 | */ |
8698a745 DY |
108 | RESCUER_NICE_LEVEL = MIN_NICE, |
109 | HIGHPRI_NICE_LEVEL = MIN_NICE, | |
ecf6881f TH |
110 | |
111 | WQ_NAME_LEN = 24, | |
c8e55f36 | 112 | }; |
1da177e4 LT |
113 | |
114 | /* | |
4690c4ab TH |
115 | * Structure fields follow one of the following exclusion rules. |
116 | * | |
e41e704b TH |
117 | * I: Modifiable by initialization/destruction paths and read-only for |
118 | * everyone else. | |
4690c4ab | 119 | * |
e22bee78 TH |
120 | * P: Preemption protected. Disabling preemption is enough and should |
121 | * only be modified and accessed from the local cpu. | |
122 | * | |
d565ed63 | 123 | * L: pool->lock protected. Access with pool->lock held. |
4690c4ab | 124 | * |
bdf8b9bf TH |
125 | * K: Only modified by worker while holding pool->lock. Can be safely read by |
126 | * self, while holding pool->lock or from IRQ context if %current is the | |
127 | * kworker. | |
128 | * | |
129 | * S: Only modified by worker self. | |
130 | * | |
1258fae7 | 131 | * A: wq_pool_attach_mutex protected. |
822d8405 | 132 | * |
68e13a67 | 133 | * PL: wq_pool_mutex protected. |
5bcab335 | 134 | * |
24acfb71 | 135 | * PR: wq_pool_mutex protected for writes. RCU protected for reads. |
76af4d93 | 136 | * |
5b95e1af LJ |
137 | * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads. |
138 | * | |
139 | * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or | |
24acfb71 | 140 | * RCU for reads. |
5b95e1af | 141 | * |
3c25a55d LJ |
142 | * WQ: wq->mutex protected. |
143 | * | |
24acfb71 | 144 | * WR: wq->mutex protected for writes. RCU protected for reads. |
2e109a28 TH |
145 | * |
146 | * MD: wq_mayday_lock protected. | |
cd2440d6 PM |
147 | * |
148 | * WD: Used internally by the watchdog. | |
1da177e4 | 149 | */ |
1da177e4 | 150 | |
2eaebdb3 | 151 | /* struct worker is defined in workqueue_internal.h */ |
c34056a3 | 152 | |
bd7bdd43 | 153 | struct worker_pool { |
a9b8a985 | 154 | raw_spinlock_t lock; /* the pool lock */ |
d84ff051 | 155 | int cpu; /* I: the associated cpu */ |
f3f90ad4 | 156 | int node; /* I: the associated node ID */ |
9daf9e67 | 157 | int id; /* I: pool ID */ |
bc8b50c2 | 158 | unsigned int flags; /* L: flags */ |
bd7bdd43 | 159 | |
82607adc | 160 | unsigned long watchdog_ts; /* L: watchdog timestamp */ |
cd2440d6 | 161 | bool cpu_stall; /* WD: stalled cpu bound pool */ |
82607adc | 162 | |
bc35f7ef LJ |
163 | /* |
164 | * The counter is incremented in a process context on the associated CPU | |
165 | * w/ preemption disabled, and decremented or reset in the same context | |
166 | * but w/ pool->lock held. The readers grab pool->lock and are | |
167 | * guaranteed to see if the counter reached zero. | |
168 | */ | |
169 | int nr_running; | |
84f91c62 | 170 | |
bd7bdd43 | 171 | struct list_head worklist; /* L: list of pending works */ |
ea1abd61 | 172 | |
5826cc8f LJ |
173 | int nr_workers; /* L: total number of workers */ |
174 | int nr_idle; /* L: currently idle workers */ | |
bd7bdd43 | 175 | |
2c1f1a91 | 176 | struct list_head idle_list; /* L: list of idle workers */ |
bd7bdd43 | 177 | struct timer_list idle_timer; /* L: worker idle timeout */ |
3f959aa3 VS |
178 | struct work_struct idle_cull_work; /* L: worker idle cleanup */ |
179 | ||
180 | struct timer_list mayday_timer; /* L: SOS timer for workers */ | |
bd7bdd43 | 181 | |
c5aa87bb | 182 | /* a workers is either on busy_hash or idle_list, or the manager */ |
c9e7cf27 TH |
183 | DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER); |
184 | /* L: hash of busy workers */ | |
185 | ||
2607d7a6 | 186 | struct worker *manager; /* L: purely informational */ |
92f9c5c4 | 187 | struct list_head workers; /* A: attached workers */ |
e02b9312 | 188 | struct list_head dying_workers; /* A: workers about to die */ |
60f5a4bc | 189 | struct completion *detach_completion; /* all workers detached */ |
e19e397a | 190 | |
7cda9aae | 191 | struct ida worker_ida; /* worker IDs for task name */ |
e19e397a | 192 | |
7a4e344c | 193 | struct workqueue_attrs *attrs; /* I: worker attributes */ |
68e13a67 LJ |
194 | struct hlist_node hash_node; /* PL: unbound_pool_hash node */ |
195 | int refcnt; /* PL: refcnt for unbound pools */ | |
7a4e344c | 196 | |
29c91e99 | 197 | /* |
24acfb71 | 198 | * Destruction of pool is RCU protected to allow dereferences |
29c91e99 TH |
199 | * from get_work_pool(). |
200 | */ | |
201 | struct rcu_head rcu; | |
84f91c62 | 202 | }; |
8b03ae3c | 203 | |
725e8ec5 TH |
204 | /* |
205 | * Per-pool_workqueue statistics. These can be monitored using | |
206 | * tools/workqueue/wq_monitor.py. | |
207 | */ | |
208 | enum pool_workqueue_stats { | |
209 | PWQ_STAT_STARTED, /* work items started execution */ | |
210 | PWQ_STAT_COMPLETED, /* work items completed execution */ | |
8a1dd1e5 | 211 | PWQ_STAT_CPU_TIME, /* total CPU time consumed */ |
616db877 | 212 | PWQ_STAT_CPU_INTENSIVE, /* wq_cpu_intensive_thresh_us violations */ |
725e8ec5 | 213 | PWQ_STAT_CM_WAKEUP, /* concurrency-management worker wakeups */ |
8639eceb | 214 | PWQ_STAT_REPATRIATED, /* unbound workers brought back into scope */ |
725e8ec5 TH |
215 | PWQ_STAT_MAYDAY, /* maydays to rescuer */ |
216 | PWQ_STAT_RESCUED, /* linked work items executed by rescuer */ | |
217 | ||
218 | PWQ_NR_STATS, | |
219 | }; | |
220 | ||
1da177e4 | 221 | /* |
112202d9 TH |
222 | * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS |
223 | * of work_struct->data are used for flags and the remaining high bits | |
224 | * point to the pwq; thus, pwqs need to be aligned at two's power of the | |
225 | * number of flag bits. | |
1da177e4 | 226 | */ |
112202d9 | 227 | struct pool_workqueue { |
bd7bdd43 | 228 | struct worker_pool *pool; /* I: the associated pool */ |
4690c4ab | 229 | struct workqueue_struct *wq; /* I: the owning workqueue */ |
73f53c4a TH |
230 | int work_color; /* L: current color */ |
231 | int flush_color; /* L: flushing color */ | |
8864b4e5 | 232 | int refcnt; /* L: reference count */ |
73f53c4a TH |
233 | int nr_in_flight[WORK_NR_COLORS]; |
234 | /* L: nr of in_flight works */ | |
018f3a13 LJ |
235 | |
236 | /* | |
237 | * nr_active management and WORK_STRUCT_INACTIVE: | |
238 | * | |
239 | * When pwq->nr_active >= max_active, new work item is queued to | |
240 | * pwq->inactive_works instead of pool->worklist and marked with | |
241 | * WORK_STRUCT_INACTIVE. | |
242 | * | |
243 | * All work items marked with WORK_STRUCT_INACTIVE do not participate | |
244 | * in pwq->nr_active and all work items in pwq->inactive_works are | |
245 | * marked with WORK_STRUCT_INACTIVE. But not all WORK_STRUCT_INACTIVE | |
246 | * work items are in pwq->inactive_works. Some of them are ready to | |
247 | * run in pool->worklist or worker->scheduled. Those work itmes are | |
248 | * only struct wq_barrier which is used for flush_work() and should | |
249 | * not participate in pwq->nr_active. For non-barrier work item, it | |
250 | * is marked with WORK_STRUCT_INACTIVE iff it is in pwq->inactive_works. | |
251 | */ | |
1e19ffc6 | 252 | int nr_active; /* L: nr of active works */ |
a0a1a5fd | 253 | int max_active; /* L: max active works */ |
f97a4a1a | 254 | struct list_head inactive_works; /* L: inactive works */ |
3c25a55d | 255 | struct list_head pwqs_node; /* WR: node on wq->pwqs */ |
2e109a28 | 256 | struct list_head mayday_node; /* MD: node on wq->maydays */ |
8864b4e5 | 257 | |
725e8ec5 TH |
258 | u64 stats[PWQ_NR_STATS]; |
259 | ||
8864b4e5 | 260 | /* |
967b494e | 261 | * Release of unbound pwq is punted to a kthread_worker. See put_pwq() |
687a9aa5 TH |
262 | * and pwq_release_workfn() for details. pool_workqueue itself is also |
263 | * RCU protected so that the first pwq can be determined without | |
967b494e | 264 | * grabbing wq->mutex. |
8864b4e5 | 265 | */ |
687a9aa5 | 266 | struct kthread_work release_work; |
8864b4e5 | 267 | struct rcu_head rcu; |
e904e6c2 | 268 | } __aligned(1 << WORK_STRUCT_FLAG_BITS); |
1da177e4 | 269 | |
73f53c4a TH |
270 | /* |
271 | * Structure used to wait for workqueue flush. | |
272 | */ | |
273 | struct wq_flusher { | |
3c25a55d LJ |
274 | struct list_head list; /* WQ: list of flushers */ |
275 | int flush_color; /* WQ: flush color waiting for */ | |
73f53c4a TH |
276 | struct completion done; /* flush completion */ |
277 | }; | |
278 | ||
226223ab TH |
279 | struct wq_device; |
280 | ||
1da177e4 | 281 | /* |
c5aa87bb TH |
282 | * The externally visible workqueue. It relays the issued work items to |
283 | * the appropriate worker_pool through its pool_workqueues. | |
1da177e4 LT |
284 | */ |
285 | struct workqueue_struct { | |
3c25a55d | 286 | struct list_head pwqs; /* WR: all pwqs of this wq */ |
e2dca7ad | 287 | struct list_head list; /* PR: list of all workqueues */ |
73f53c4a | 288 | |
3c25a55d LJ |
289 | struct mutex mutex; /* protects this wq */ |
290 | int work_color; /* WQ: current work color */ | |
291 | int flush_color; /* WQ: current flush color */ | |
112202d9 | 292 | atomic_t nr_pwqs_to_flush; /* flush in progress */ |
3c25a55d LJ |
293 | struct wq_flusher *first_flusher; /* WQ: first flusher */ |
294 | struct list_head flusher_queue; /* WQ: flush waiters */ | |
295 | struct list_head flusher_overflow; /* WQ: flush overflow list */ | |
73f53c4a | 296 | |
2e109a28 | 297 | struct list_head maydays; /* MD: pwqs requesting rescue */ |
30ae2fc0 | 298 | struct worker *rescuer; /* MD: rescue worker */ |
e22bee78 | 299 | |
87fc741e | 300 | int nr_drainers; /* WQ: drain in progress */ |
a357fc03 | 301 | int saved_max_active; /* WQ: saved pwq max_active */ |
226223ab | 302 | |
5b95e1af LJ |
303 | struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */ |
304 | struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */ | |
6029a918 | 305 | |
226223ab TH |
306 | #ifdef CONFIG_SYSFS |
307 | struct wq_device *wq_dev; /* I: for sysfs interface */ | |
308 | #endif | |
4e6045f1 | 309 | #ifdef CONFIG_LOCKDEP |
669de8bd BVA |
310 | char *lock_name; |
311 | struct lock_class_key key; | |
4690c4ab | 312 | struct lockdep_map lockdep_map; |
4e6045f1 | 313 | #endif |
ecf6881f | 314 | char name[WQ_NAME_LEN]; /* I: workqueue name */ |
2728fd2f | 315 | |
e2dca7ad | 316 | /* |
24acfb71 TG |
317 | * Destruction of workqueue_struct is RCU protected to allow walking |
318 | * the workqueues list without grabbing wq_pool_mutex. | |
e2dca7ad TH |
319 | * This is used to dump all workqueues from sysrq. |
320 | */ | |
321 | struct rcu_head rcu; | |
322 | ||
2728fd2f TH |
323 | /* hot fields used during command issue, aligned to cacheline */ |
324 | unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */ | |
636b927e | 325 | struct pool_workqueue __percpu __rcu **cpu_pwq; /* I: per-cpu pwqs */ |
1da177e4 LT |
326 | }; |
327 | ||
e904e6c2 TH |
328 | static struct kmem_cache *pwq_cache; |
329 | ||
84193c07 TH |
330 | /* |
331 | * Each pod type describes how CPUs should be grouped for unbound workqueues. | |
332 | * See the comment above workqueue_attrs->affn_scope. | |
333 | */ | |
334 | struct wq_pod_type { | |
335 | int nr_pods; /* number of pods */ | |
336 | cpumask_var_t *pod_cpus; /* pod -> cpus */ | |
337 | int *pod_node; /* pod -> node */ | |
338 | int *cpu_pod; /* cpu -> pod */ | |
339 | }; | |
340 | ||
341 | static struct wq_pod_type wq_pod_types[WQ_AFFN_NR_TYPES]; | |
523a301e | 342 | static enum wq_affn_scope wq_affn_dfl = WQ_AFFN_CACHE; |
63c5484e TH |
343 | |
344 | static const char *wq_affn_names[WQ_AFFN_NR_TYPES] = { | |
523a301e | 345 | [WQ_AFFN_DFL] = "default", |
63c5484e TH |
346 | [WQ_AFFN_CPU] = "cpu", |
347 | [WQ_AFFN_SMT] = "smt", | |
348 | [WQ_AFFN_CACHE] = "cache", | |
349 | [WQ_AFFN_NUMA] = "numa", | |
350 | [WQ_AFFN_SYSTEM] = "system", | |
351 | }; | |
bce90380 | 352 | |
616db877 TH |
353 | /* |
354 | * Per-cpu work items which run for longer than the following threshold are | |
355 | * automatically considered CPU intensive and excluded from concurrency | |
356 | * management to prevent them from noticeably delaying other per-cpu work items. | |
aa6fde93 TH |
357 | * ULONG_MAX indicates that the user hasn't overridden it with a boot parameter. |
358 | * The actual value is initialized in wq_cpu_intensive_thresh_init(). | |
616db877 | 359 | */ |
aa6fde93 | 360 | static unsigned long wq_cpu_intensive_thresh_us = ULONG_MAX; |
616db877 TH |
361 | module_param_named(cpu_intensive_thresh_us, wq_cpu_intensive_thresh_us, ulong, 0644); |
362 | ||
cee22a15 | 363 | /* see the comment above the definition of WQ_POWER_EFFICIENT */ |
552f530c | 364 | static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT); |
cee22a15 VK |
365 | module_param_named(power_efficient, wq_power_efficient, bool, 0444); |
366 | ||
863b710b | 367 | static bool wq_online; /* can kworkers be created yet? */ |
3347fa09 | 368 | |
fef59c9c TH |
369 | /* buf for wq_update_unbound_pod_attrs(), protected by CPU hotplug exclusion */ |
370 | static struct workqueue_attrs *wq_update_pod_attrs_buf; | |
4c16bd32 | 371 | |
68e13a67 | 372 | static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ |
1258fae7 | 373 | static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */ |
a9b8a985 | 374 | static DEFINE_RAW_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ |
d8bb65ab SAS |
375 | /* wait for manager to go away */ |
376 | static struct rcuwait manager_wait = __RCUWAIT_INITIALIZER(manager_wait); | |
5bcab335 | 377 | |
e2dca7ad | 378 | static LIST_HEAD(workqueues); /* PR: list of all workqueues */ |
68e13a67 | 379 | static bool workqueue_freezing; /* PL: have wqs started freezing? */ |
7d19c5ce | 380 | |
99c621ef | 381 | /* PL&A: allowable cpus for unbound wqs and work items */ |
ef557180 MG |
382 | static cpumask_var_t wq_unbound_cpumask; |
383 | ||
ace3c549 | 384 | /* for further constrain wq_unbound_cpumask by cmdline parameter*/ |
385 | static struct cpumask wq_cmdline_cpumask __initdata; | |
386 | ||
ef557180 MG |
387 | /* CPU where unbound work was last round robin scheduled from this CPU */ |
388 | static DEFINE_PER_CPU(int, wq_rr_cpu_last); | |
b05a7928 | 389 | |
f303fccb TH |
390 | /* |
391 | * Local execution of unbound work items is no longer guaranteed. The | |
392 | * following always forces round-robin CPU selection on unbound work items | |
393 | * to uncover usages which depend on it. | |
394 | */ | |
395 | #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU | |
396 | static bool wq_debug_force_rr_cpu = true; | |
397 | #else | |
398 | static bool wq_debug_force_rr_cpu = false; | |
399 | #endif | |
400 | module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644); | |
401 | ||
7d19c5ce | 402 | /* the per-cpu worker pools */ |
25528213 | 403 | static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools); |
7d19c5ce | 404 | |
68e13a67 | 405 | static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ |
7d19c5ce | 406 | |
68e13a67 | 407 | /* PL: hash of all unbound pools keyed by pool->attrs */ |
29c91e99 TH |
408 | static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); |
409 | ||
c5aa87bb | 410 | /* I: attributes used when instantiating standard unbound pools on demand */ |
29c91e99 TH |
411 | static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; |
412 | ||
8a2b7538 TH |
413 | /* I: attributes used when instantiating ordered pools on demand */ |
414 | static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS]; | |
415 | ||
967b494e TH |
416 | /* |
417 | * I: kthread_worker to release pwq's. pwq release needs to be bounced to a | |
418 | * process context while holding a pool lock. Bounce to a dedicated kthread | |
419 | * worker to avoid A-A deadlocks. | |
420 | */ | |
421 | static struct kthread_worker *pwq_release_worker; | |
422 | ||
d320c038 | 423 | struct workqueue_struct *system_wq __read_mostly; |
ad7b1f84 | 424 | EXPORT_SYMBOL(system_wq); |
044c782c | 425 | struct workqueue_struct *system_highpri_wq __read_mostly; |
1aabe902 | 426 | EXPORT_SYMBOL_GPL(system_highpri_wq); |
044c782c | 427 | struct workqueue_struct *system_long_wq __read_mostly; |
d320c038 | 428 | EXPORT_SYMBOL_GPL(system_long_wq); |
044c782c | 429 | struct workqueue_struct *system_unbound_wq __read_mostly; |
f3421797 | 430 | EXPORT_SYMBOL_GPL(system_unbound_wq); |
044c782c | 431 | struct workqueue_struct *system_freezable_wq __read_mostly; |
24d51add | 432 | EXPORT_SYMBOL_GPL(system_freezable_wq); |
0668106c VK |
433 | struct workqueue_struct *system_power_efficient_wq __read_mostly; |
434 | EXPORT_SYMBOL_GPL(system_power_efficient_wq); | |
435 | struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly; | |
436 | EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq); | |
d320c038 | 437 | |
7d19c5ce | 438 | static int worker_thread(void *__worker); |
6ba94429 | 439 | static void workqueue_sysfs_unregister(struct workqueue_struct *wq); |
c29eb853 | 440 | static void show_pwq(struct pool_workqueue *pwq); |
55df0933 | 441 | static void show_one_worker_pool(struct worker_pool *pool); |
7d19c5ce | 442 | |
97bd2347 TH |
443 | #define CREATE_TRACE_POINTS |
444 | #include <trace/events/workqueue.h> | |
445 | ||
68e13a67 | 446 | #define assert_rcu_or_pool_mutex() \ |
24acfb71 | 447 | RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \ |
f78f5b90 | 448 | !lockdep_is_held(&wq_pool_mutex), \ |
24acfb71 | 449 | "RCU or wq_pool_mutex should be held") |
5bcab335 | 450 | |
5b95e1af | 451 | #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \ |
24acfb71 | 452 | RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \ |
f78f5b90 PM |
453 | !lockdep_is_held(&wq->mutex) && \ |
454 | !lockdep_is_held(&wq_pool_mutex), \ | |
24acfb71 | 455 | "RCU, wq->mutex or wq_pool_mutex should be held") |
5b95e1af | 456 | |
f02ae73a TH |
457 | #define for_each_cpu_worker_pool(pool, cpu) \ |
458 | for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ | |
459 | (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ | |
7a62c2c8 | 460 | (pool)++) |
4ce62e9e | 461 | |
17116969 TH |
462 | /** |
463 | * for_each_pool - iterate through all worker_pools in the system | |
464 | * @pool: iteration cursor | |
611c92a0 | 465 | * @pi: integer used for iteration |
fa1b54e6 | 466 | * |
24acfb71 | 467 | * This must be called either with wq_pool_mutex held or RCU read |
68e13a67 LJ |
468 | * locked. If the pool needs to be used beyond the locking in effect, the |
469 | * caller is responsible for guaranteeing that the pool stays online. | |
fa1b54e6 TH |
470 | * |
471 | * The if/else clause exists only for the lockdep assertion and can be | |
472 | * ignored. | |
17116969 | 473 | */ |
611c92a0 TH |
474 | #define for_each_pool(pool, pi) \ |
475 | idr_for_each_entry(&worker_pool_idr, pool, pi) \ | |
68e13a67 | 476 | if (({ assert_rcu_or_pool_mutex(); false; })) { } \ |
fa1b54e6 | 477 | else |
17116969 | 478 | |
822d8405 TH |
479 | /** |
480 | * for_each_pool_worker - iterate through all workers of a worker_pool | |
481 | * @worker: iteration cursor | |
822d8405 TH |
482 | * @pool: worker_pool to iterate workers of |
483 | * | |
1258fae7 | 484 | * This must be called with wq_pool_attach_mutex. |
822d8405 TH |
485 | * |
486 | * The if/else clause exists only for the lockdep assertion and can be | |
487 | * ignored. | |
488 | */ | |
da028469 LJ |
489 | #define for_each_pool_worker(worker, pool) \ |
490 | list_for_each_entry((worker), &(pool)->workers, node) \ | |
1258fae7 | 491 | if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \ |
822d8405 TH |
492 | else |
493 | ||
49e3cf44 TH |
494 | /** |
495 | * for_each_pwq - iterate through all pool_workqueues of the specified workqueue | |
496 | * @pwq: iteration cursor | |
497 | * @wq: the target workqueue | |
76af4d93 | 498 | * |
24acfb71 | 499 | * This must be called either with wq->mutex held or RCU read locked. |
794b18bc TH |
500 | * If the pwq needs to be used beyond the locking in effect, the caller is |
501 | * responsible for guaranteeing that the pwq stays online. | |
76af4d93 TH |
502 | * |
503 | * The if/else clause exists only for the lockdep assertion and can be | |
504 | * ignored. | |
49e3cf44 TH |
505 | */ |
506 | #define for_each_pwq(pwq, wq) \ | |
49e9d1a9 | 507 | list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \ |
5a644662 | 508 | lockdep_is_held(&(wq->mutex))) |
f3421797 | 509 | |
dc186ad7 TG |
510 | #ifdef CONFIG_DEBUG_OBJECTS_WORK |
511 | ||
f9e62f31 | 512 | static const struct debug_obj_descr work_debug_descr; |
dc186ad7 | 513 | |
99777288 SG |
514 | static void *work_debug_hint(void *addr) |
515 | { | |
516 | return ((struct work_struct *) addr)->func; | |
517 | } | |
518 | ||
b9fdac7f DC |
519 | static bool work_is_static_object(void *addr) |
520 | { | |
521 | struct work_struct *work = addr; | |
522 | ||
523 | return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work)); | |
524 | } | |
525 | ||
dc186ad7 TG |
526 | /* |
527 | * fixup_init is called when: | |
528 | * - an active object is initialized | |
529 | */ | |
02a982a6 | 530 | static bool work_fixup_init(void *addr, enum debug_obj_state state) |
dc186ad7 TG |
531 | { |
532 | struct work_struct *work = addr; | |
533 | ||
534 | switch (state) { | |
535 | case ODEBUG_STATE_ACTIVE: | |
536 | cancel_work_sync(work); | |
537 | debug_object_init(work, &work_debug_descr); | |
02a982a6 | 538 | return true; |
dc186ad7 | 539 | default: |
02a982a6 | 540 | return false; |
dc186ad7 TG |
541 | } |
542 | } | |
543 | ||
dc186ad7 TG |
544 | /* |
545 | * fixup_free is called when: | |
546 | * - an active object is freed | |
547 | */ | |
02a982a6 | 548 | static bool work_fixup_free(void *addr, enum debug_obj_state state) |
dc186ad7 TG |
549 | { |
550 | struct work_struct *work = addr; | |
551 | ||
552 | switch (state) { | |
553 | case ODEBUG_STATE_ACTIVE: | |
554 | cancel_work_sync(work); | |
555 | debug_object_free(work, &work_debug_descr); | |
02a982a6 | 556 | return true; |
dc186ad7 | 557 | default: |
02a982a6 | 558 | return false; |
dc186ad7 TG |
559 | } |
560 | } | |
561 | ||
f9e62f31 | 562 | static const struct debug_obj_descr work_debug_descr = { |
dc186ad7 | 563 | .name = "work_struct", |
99777288 | 564 | .debug_hint = work_debug_hint, |
b9fdac7f | 565 | .is_static_object = work_is_static_object, |
dc186ad7 | 566 | .fixup_init = work_fixup_init, |
dc186ad7 TG |
567 | .fixup_free = work_fixup_free, |
568 | }; | |
569 | ||
570 | static inline void debug_work_activate(struct work_struct *work) | |
571 | { | |
572 | debug_object_activate(work, &work_debug_descr); | |
573 | } | |
574 | ||
575 | static inline void debug_work_deactivate(struct work_struct *work) | |
576 | { | |
577 | debug_object_deactivate(work, &work_debug_descr); | |
578 | } | |
579 | ||
580 | void __init_work(struct work_struct *work, int onstack) | |
581 | { | |
582 | if (onstack) | |
583 | debug_object_init_on_stack(work, &work_debug_descr); | |
584 | else | |
585 | debug_object_init(work, &work_debug_descr); | |
586 | } | |
587 | EXPORT_SYMBOL_GPL(__init_work); | |
588 | ||
589 | void destroy_work_on_stack(struct work_struct *work) | |
590 | { | |
591 | debug_object_free(work, &work_debug_descr); | |
592 | } | |
593 | EXPORT_SYMBOL_GPL(destroy_work_on_stack); | |
594 | ||
ea2e64f2 TG |
595 | void destroy_delayed_work_on_stack(struct delayed_work *work) |
596 | { | |
597 | destroy_timer_on_stack(&work->timer); | |
598 | debug_object_free(&work->work, &work_debug_descr); | |
599 | } | |
600 | EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack); | |
601 | ||
dc186ad7 TG |
602 | #else |
603 | static inline void debug_work_activate(struct work_struct *work) { } | |
604 | static inline void debug_work_deactivate(struct work_struct *work) { } | |
605 | #endif | |
606 | ||
4e8b22bd | 607 | /** |
67dc8325 | 608 | * worker_pool_assign_id - allocate ID and assign it to @pool |
4e8b22bd LB |
609 | * @pool: the pool pointer of interest |
610 | * | |
611 | * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned | |
612 | * successfully, -errno on failure. | |
613 | */ | |
9daf9e67 TH |
614 | static int worker_pool_assign_id(struct worker_pool *pool) |
615 | { | |
616 | int ret; | |
617 | ||
68e13a67 | 618 | lockdep_assert_held(&wq_pool_mutex); |
5bcab335 | 619 | |
4e8b22bd LB |
620 | ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE, |
621 | GFP_KERNEL); | |
229641a6 | 622 | if (ret >= 0) { |
e68035fb | 623 | pool->id = ret; |
229641a6 TH |
624 | return 0; |
625 | } | |
fa1b54e6 | 626 | return ret; |
7c3eed5c TH |
627 | } |
628 | ||
73f53c4a TH |
629 | static unsigned int work_color_to_flags(int color) |
630 | { | |
631 | return color << WORK_STRUCT_COLOR_SHIFT; | |
632 | } | |
633 | ||
c4560c2c | 634 | static int get_work_color(unsigned long work_data) |
73f53c4a | 635 | { |
c4560c2c | 636 | return (work_data >> WORK_STRUCT_COLOR_SHIFT) & |
73f53c4a TH |
637 | ((1 << WORK_STRUCT_COLOR_BITS) - 1); |
638 | } | |
639 | ||
640 | static int work_next_color(int color) | |
641 | { | |
642 | return (color + 1) % WORK_NR_COLORS; | |
643 | } | |
1da177e4 | 644 | |
14441960 | 645 | /* |
112202d9 TH |
646 | * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data |
647 | * contain the pointer to the queued pwq. Once execution starts, the flag | |
7c3eed5c | 648 | * is cleared and the high bits contain OFFQ flags and pool ID. |
7a22ad75 | 649 | * |
112202d9 TH |
650 | * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() |
651 | * and clear_work_data() can be used to set the pwq, pool or clear | |
bbb68dfa TH |
652 | * work->data. These functions should only be called while the work is |
653 | * owned - ie. while the PENDING bit is set. | |
7a22ad75 | 654 | * |
112202d9 | 655 | * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq |
7c3eed5c | 656 | * corresponding to a work. Pool is available once the work has been |
112202d9 | 657 | * queued anywhere after initialization until it is sync canceled. pwq is |
7c3eed5c | 658 | * available only while the work item is queued. |
7a22ad75 | 659 | * |
bbb68dfa TH |
660 | * %WORK_OFFQ_CANCELING is used to mark a work item which is being |
661 | * canceled. While being canceled, a work item may have its PENDING set | |
662 | * but stay off timer and worklist for arbitrarily long and nobody should | |
663 | * try to steal the PENDING bit. | |
14441960 | 664 | */ |
7a22ad75 TH |
665 | static inline void set_work_data(struct work_struct *work, unsigned long data, |
666 | unsigned long flags) | |
365970a1 | 667 | { |
6183c009 | 668 | WARN_ON_ONCE(!work_pending(work)); |
7a22ad75 TH |
669 | atomic_long_set(&work->data, data | flags | work_static(work)); |
670 | } | |
365970a1 | 671 | |
112202d9 | 672 | static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, |
7a22ad75 TH |
673 | unsigned long extra_flags) |
674 | { | |
112202d9 TH |
675 | set_work_data(work, (unsigned long)pwq, |
676 | WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); | |
365970a1 DH |
677 | } |
678 | ||
4468a00f LJ |
679 | static void set_work_pool_and_keep_pending(struct work_struct *work, |
680 | int pool_id) | |
681 | { | |
682 | set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, | |
683 | WORK_STRUCT_PENDING); | |
684 | } | |
685 | ||
7c3eed5c TH |
686 | static void set_work_pool_and_clear_pending(struct work_struct *work, |
687 | int pool_id) | |
7a22ad75 | 688 | { |
23657bb1 TH |
689 | /* |
690 | * The following wmb is paired with the implied mb in | |
691 | * test_and_set_bit(PENDING) and ensures all updates to @work made | |
692 | * here are visible to and precede any updates by the next PENDING | |
693 | * owner. | |
694 | */ | |
695 | smp_wmb(); | |
7c3eed5c | 696 | set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); |
346c09f8 RP |
697 | /* |
698 | * The following mb guarantees that previous clear of a PENDING bit | |
699 | * will not be reordered with any speculative LOADS or STORES from | |
700 | * work->current_func, which is executed afterwards. This possible | |
8bdc6201 | 701 | * reordering can lead to a missed execution on attempt to queue |
346c09f8 RP |
702 | * the same @work. E.g. consider this case: |
703 | * | |
704 | * CPU#0 CPU#1 | |
705 | * ---------------------------- -------------------------------- | |
706 | * | |
707 | * 1 STORE event_indicated | |
708 | * 2 queue_work_on() { | |
709 | * 3 test_and_set_bit(PENDING) | |
710 | * 4 } set_..._and_clear_pending() { | |
711 | * 5 set_work_data() # clear bit | |
712 | * 6 smp_mb() | |
713 | * 7 work->current_func() { | |
714 | * 8 LOAD event_indicated | |
715 | * } | |
716 | * | |
717 | * Without an explicit full barrier speculative LOAD on line 8 can | |
718 | * be executed before CPU#0 does STORE on line 1. If that happens, | |
719 | * CPU#0 observes the PENDING bit is still set and new execution of | |
720 | * a @work is not queued in a hope, that CPU#1 will eventually | |
721 | * finish the queued @work. Meanwhile CPU#1 does not see | |
722 | * event_indicated is set, because speculative LOAD was executed | |
723 | * before actual STORE. | |
724 | */ | |
725 | smp_mb(); | |
7a22ad75 | 726 | } |
f756d5e2 | 727 | |
7a22ad75 | 728 | static void clear_work_data(struct work_struct *work) |
1da177e4 | 729 | { |
7c3eed5c TH |
730 | smp_wmb(); /* see set_work_pool_and_clear_pending() */ |
731 | set_work_data(work, WORK_STRUCT_NO_POOL, 0); | |
1da177e4 LT |
732 | } |
733 | ||
afa4bb77 LT |
734 | static inline struct pool_workqueue *work_struct_pwq(unsigned long data) |
735 | { | |
736 | return (struct pool_workqueue *)(data & WORK_STRUCT_WQ_DATA_MASK); | |
737 | } | |
738 | ||
112202d9 | 739 | static struct pool_workqueue *get_work_pwq(struct work_struct *work) |
b1f4ec17 | 740 | { |
e120153d | 741 | unsigned long data = atomic_long_read(&work->data); |
7a22ad75 | 742 | |
112202d9 | 743 | if (data & WORK_STRUCT_PWQ) |
afa4bb77 | 744 | return work_struct_pwq(data); |
e120153d TH |
745 | else |
746 | return NULL; | |
4d707b9f ON |
747 | } |
748 | ||
7c3eed5c TH |
749 | /** |
750 | * get_work_pool - return the worker_pool a given work was associated with | |
751 | * @work: the work item of interest | |
752 | * | |
68e13a67 | 753 | * Pools are created and destroyed under wq_pool_mutex, and allows read |
24acfb71 TG |
754 | * access under RCU read lock. As such, this function should be |
755 | * called under wq_pool_mutex or inside of a rcu_read_lock() region. | |
fa1b54e6 TH |
756 | * |
757 | * All fields of the returned pool are accessible as long as the above | |
758 | * mentioned locking is in effect. If the returned pool needs to be used | |
759 | * beyond the critical section, the caller is responsible for ensuring the | |
760 | * returned pool is and stays online. | |
d185af30 YB |
761 | * |
762 | * Return: The worker_pool @work was last associated with. %NULL if none. | |
7c3eed5c TH |
763 | */ |
764 | static struct worker_pool *get_work_pool(struct work_struct *work) | |
365970a1 | 765 | { |
e120153d | 766 | unsigned long data = atomic_long_read(&work->data); |
7c3eed5c | 767 | int pool_id; |
7a22ad75 | 768 | |
68e13a67 | 769 | assert_rcu_or_pool_mutex(); |
fa1b54e6 | 770 | |
112202d9 | 771 | if (data & WORK_STRUCT_PWQ) |
afa4bb77 | 772 | return work_struct_pwq(data)->pool; |
7a22ad75 | 773 | |
7c3eed5c TH |
774 | pool_id = data >> WORK_OFFQ_POOL_SHIFT; |
775 | if (pool_id == WORK_OFFQ_POOL_NONE) | |
7a22ad75 TH |
776 | return NULL; |
777 | ||
fa1b54e6 | 778 | return idr_find(&worker_pool_idr, pool_id); |
7c3eed5c TH |
779 | } |
780 | ||
781 | /** | |
782 | * get_work_pool_id - return the worker pool ID a given work is associated with | |
783 | * @work: the work item of interest | |
784 | * | |
d185af30 | 785 | * Return: The worker_pool ID @work was last associated with. |
7c3eed5c TH |
786 | * %WORK_OFFQ_POOL_NONE if none. |
787 | */ | |
788 | static int get_work_pool_id(struct work_struct *work) | |
789 | { | |
54d5b7d0 LJ |
790 | unsigned long data = atomic_long_read(&work->data); |
791 | ||
112202d9 | 792 | if (data & WORK_STRUCT_PWQ) |
afa4bb77 | 793 | return work_struct_pwq(data)->pool->id; |
7c3eed5c | 794 | |
54d5b7d0 | 795 | return data >> WORK_OFFQ_POOL_SHIFT; |
7c3eed5c TH |
796 | } |
797 | ||
bbb68dfa TH |
798 | static void mark_work_canceling(struct work_struct *work) |
799 | { | |
7c3eed5c | 800 | unsigned long pool_id = get_work_pool_id(work); |
bbb68dfa | 801 | |
7c3eed5c TH |
802 | pool_id <<= WORK_OFFQ_POOL_SHIFT; |
803 | set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); | |
bbb68dfa TH |
804 | } |
805 | ||
806 | static bool work_is_canceling(struct work_struct *work) | |
807 | { | |
808 | unsigned long data = atomic_long_read(&work->data); | |
809 | ||
112202d9 | 810 | return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); |
bbb68dfa TH |
811 | } |
812 | ||
e22bee78 | 813 | /* |
3270476a TH |
814 | * Policy functions. These define the policies on how the global worker |
815 | * pools are managed. Unless noted otherwise, these functions assume that | |
d565ed63 | 816 | * they're being called with pool->lock held. |
e22bee78 TH |
817 | */ |
818 | ||
4594bf15 | 819 | /* |
e22bee78 TH |
820 | * Need to wake up a worker? Called from anything but currently |
821 | * running workers. | |
974271c4 TH |
822 | * |
823 | * Note that, because unbound workers never contribute to nr_running, this | |
706026c2 | 824 | * function will always return %true for unbound pools as long as the |
974271c4 | 825 | * worklist isn't empty. |
4594bf15 | 826 | */ |
63d95a91 | 827 | static bool need_more_worker(struct worker_pool *pool) |
365970a1 | 828 | { |
0219a352 | 829 | return !list_empty(&pool->worklist) && !pool->nr_running; |
e22bee78 | 830 | } |
4594bf15 | 831 | |
e22bee78 | 832 | /* Can I start working? Called from busy but !running workers. */ |
63d95a91 | 833 | static bool may_start_working(struct worker_pool *pool) |
e22bee78 | 834 | { |
63d95a91 | 835 | return pool->nr_idle; |
e22bee78 TH |
836 | } |
837 | ||
838 | /* Do I need to keep working? Called from currently running workers. */ | |
63d95a91 | 839 | static bool keep_working(struct worker_pool *pool) |
e22bee78 | 840 | { |
bc35f7ef | 841 | return !list_empty(&pool->worklist) && (pool->nr_running <= 1); |
e22bee78 TH |
842 | } |
843 | ||
844 | /* Do we need a new worker? Called from manager. */ | |
63d95a91 | 845 | static bool need_to_create_worker(struct worker_pool *pool) |
e22bee78 | 846 | { |
63d95a91 | 847 | return need_more_worker(pool) && !may_start_working(pool); |
e22bee78 | 848 | } |
365970a1 | 849 | |
e22bee78 | 850 | /* Do we have too many workers and should some go away? */ |
63d95a91 | 851 | static bool too_many_workers(struct worker_pool *pool) |
e22bee78 | 852 | { |
692b4825 | 853 | bool managing = pool->flags & POOL_MANAGER_ACTIVE; |
63d95a91 TH |
854 | int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ |
855 | int nr_busy = pool->nr_workers - nr_idle; | |
e22bee78 TH |
856 | |
857 | return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; | |
365970a1 DH |
858 | } |
859 | ||
c54d5046 TH |
860 | /** |
861 | * worker_set_flags - set worker flags and adjust nr_running accordingly | |
862 | * @worker: self | |
863 | * @flags: flags to set | |
864 | * | |
865 | * Set @flags in @worker->flags and adjust nr_running accordingly. | |
c54d5046 TH |
866 | */ |
867 | static inline void worker_set_flags(struct worker *worker, unsigned int flags) | |
868 | { | |
869 | struct worker_pool *pool = worker->pool; | |
870 | ||
bc8b50c2 | 871 | lockdep_assert_held(&pool->lock); |
c54d5046 TH |
872 | |
873 | /* If transitioning into NOT_RUNNING, adjust nr_running. */ | |
874 | if ((flags & WORKER_NOT_RUNNING) && | |
875 | !(worker->flags & WORKER_NOT_RUNNING)) { | |
876 | pool->nr_running--; | |
877 | } | |
878 | ||
879 | worker->flags |= flags; | |
880 | } | |
881 | ||
882 | /** | |
883 | * worker_clr_flags - clear worker flags and adjust nr_running accordingly | |
884 | * @worker: self | |
885 | * @flags: flags to clear | |
886 | * | |
887 | * Clear @flags in @worker->flags and adjust nr_running accordingly. | |
c54d5046 TH |
888 | */ |
889 | static inline void worker_clr_flags(struct worker *worker, unsigned int flags) | |
890 | { | |
891 | struct worker_pool *pool = worker->pool; | |
892 | unsigned int oflags = worker->flags; | |
893 | ||
bc8b50c2 | 894 | lockdep_assert_held(&pool->lock); |
c54d5046 TH |
895 | |
896 | worker->flags &= ~flags; | |
897 | ||
898 | /* | |
899 | * If transitioning out of NOT_RUNNING, increment nr_running. Note | |
900 | * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask | |
901 | * of multiple flags, not a single flag. | |
902 | */ | |
903 | if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) | |
904 | if (!(worker->flags & WORKER_NOT_RUNNING)) | |
905 | pool->nr_running++; | |
906 | } | |
907 | ||
797e8345 TH |
908 | /* Return the first idle worker. Called with pool->lock held. */ |
909 | static struct worker *first_idle_worker(struct worker_pool *pool) | |
910 | { | |
911 | if (unlikely(list_empty(&pool->idle_list))) | |
912 | return NULL; | |
913 | ||
914 | return list_first_entry(&pool->idle_list, struct worker, entry); | |
915 | } | |
916 | ||
917 | /** | |
918 | * worker_enter_idle - enter idle state | |
919 | * @worker: worker which is entering idle state | |
920 | * | |
921 | * @worker is entering idle state. Update stats and idle timer if | |
922 | * necessary. | |
923 | * | |
924 | * LOCKING: | |
925 | * raw_spin_lock_irq(pool->lock). | |
926 | */ | |
927 | static void worker_enter_idle(struct worker *worker) | |
928 | { | |
929 | struct worker_pool *pool = worker->pool; | |
930 | ||
931 | if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || | |
932 | WARN_ON_ONCE(!list_empty(&worker->entry) && | |
933 | (worker->hentry.next || worker->hentry.pprev))) | |
934 | return; | |
935 | ||
936 | /* can't use worker_set_flags(), also called from create_worker() */ | |
937 | worker->flags |= WORKER_IDLE; | |
938 | pool->nr_idle++; | |
939 | worker->last_active = jiffies; | |
940 | ||
941 | /* idle_list is LIFO */ | |
942 | list_add(&worker->entry, &pool->idle_list); | |
943 | ||
944 | if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) | |
945 | mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); | |
946 | ||
947 | /* Sanity check nr_running. */ | |
948 | WARN_ON_ONCE(pool->nr_workers == pool->nr_idle && pool->nr_running); | |
949 | } | |
950 | ||
951 | /** | |
952 | * worker_leave_idle - leave idle state | |
953 | * @worker: worker which is leaving idle state | |
954 | * | |
955 | * @worker is leaving idle state. Update stats. | |
956 | * | |
957 | * LOCKING: | |
958 | * raw_spin_lock_irq(pool->lock). | |
959 | */ | |
960 | static void worker_leave_idle(struct worker *worker) | |
961 | { | |
962 | struct worker_pool *pool = worker->pool; | |
963 | ||
964 | if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) | |
965 | return; | |
966 | worker_clr_flags(worker, WORKER_IDLE); | |
967 | pool->nr_idle--; | |
968 | list_del_init(&worker->entry); | |
969 | } | |
970 | ||
971 | /** | |
972 | * find_worker_executing_work - find worker which is executing a work | |
973 | * @pool: pool of interest | |
974 | * @work: work to find worker for | |
975 | * | |
976 | * Find a worker which is executing @work on @pool by searching | |
977 | * @pool->busy_hash which is keyed by the address of @work. For a worker | |
978 | * to match, its current execution should match the address of @work and | |
979 | * its work function. This is to avoid unwanted dependency between | |
980 | * unrelated work executions through a work item being recycled while still | |
981 | * being executed. | |
982 | * | |
983 | * This is a bit tricky. A work item may be freed once its execution | |
984 | * starts and nothing prevents the freed area from being recycled for | |
985 | * another work item. If the same work item address ends up being reused | |
986 | * before the original execution finishes, workqueue will identify the | |
987 | * recycled work item as currently executing and make it wait until the | |
988 | * current execution finishes, introducing an unwanted dependency. | |
989 | * | |
990 | * This function checks the work item address and work function to avoid | |
991 | * false positives. Note that this isn't complete as one may construct a | |
992 | * work function which can introduce dependency onto itself through a | |
993 | * recycled work item. Well, if somebody wants to shoot oneself in the | |
994 | * foot that badly, there's only so much we can do, and if such deadlock | |
995 | * actually occurs, it should be easy to locate the culprit work function. | |
996 | * | |
997 | * CONTEXT: | |
998 | * raw_spin_lock_irq(pool->lock). | |
999 | * | |
1000 | * Return: | |
1001 | * Pointer to worker which is executing @work if found, %NULL | |
1002 | * otherwise. | |
1003 | */ | |
1004 | static struct worker *find_worker_executing_work(struct worker_pool *pool, | |
1005 | struct work_struct *work) | |
1006 | { | |
1007 | struct worker *worker; | |
1008 | ||
1009 | hash_for_each_possible(pool->busy_hash, worker, hentry, | |
1010 | (unsigned long)work) | |
1011 | if (worker->current_work == work && | |
1012 | worker->current_func == work->func) | |
1013 | return worker; | |
1014 | ||
1015 | return NULL; | |
1016 | } | |
1017 | ||
1018 | /** | |
1019 | * move_linked_works - move linked works to a list | |
1020 | * @work: start of series of works to be scheduled | |
1021 | * @head: target list to append @work to | |
1022 | * @nextp: out parameter for nested worklist walking | |
1023 | * | |
873eaca6 TH |
1024 | * Schedule linked works starting from @work to @head. Work series to be |
1025 | * scheduled starts at @work and includes any consecutive work with | |
1026 | * WORK_STRUCT_LINKED set in its predecessor. See assign_work() for details on | |
1027 | * @nextp. | |
797e8345 TH |
1028 | * |
1029 | * CONTEXT: | |
1030 | * raw_spin_lock_irq(pool->lock). | |
1031 | */ | |
1032 | static void move_linked_works(struct work_struct *work, struct list_head *head, | |
1033 | struct work_struct **nextp) | |
1034 | { | |
1035 | struct work_struct *n; | |
1036 | ||
1037 | /* | |
1038 | * Linked worklist will always end before the end of the list, | |
1039 | * use NULL for list head. | |
1040 | */ | |
1041 | list_for_each_entry_safe_from(work, n, NULL, entry) { | |
1042 | list_move_tail(&work->entry, head); | |
1043 | if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) | |
1044 | break; | |
1045 | } | |
1046 | ||
1047 | /* | |
1048 | * If we're already inside safe list traversal and have moved | |
1049 | * multiple works to the scheduled queue, the next position | |
1050 | * needs to be updated. | |
1051 | */ | |
1052 | if (nextp) | |
1053 | *nextp = n; | |
1054 | } | |
1055 | ||
873eaca6 TH |
1056 | /** |
1057 | * assign_work - assign a work item and its linked work items to a worker | |
1058 | * @work: work to assign | |
1059 | * @worker: worker to assign to | |
1060 | * @nextp: out parameter for nested worklist walking | |
1061 | * | |
1062 | * Assign @work and its linked work items to @worker. If @work is already being | |
1063 | * executed by another worker in the same pool, it'll be punted there. | |
1064 | * | |
1065 | * If @nextp is not NULL, it's updated to point to the next work of the last | |
1066 | * scheduled work. This allows assign_work() to be nested inside | |
1067 | * list_for_each_entry_safe(). | |
1068 | * | |
1069 | * Returns %true if @work was successfully assigned to @worker. %false if @work | |
1070 | * was punted to another worker already executing it. | |
1071 | */ | |
1072 | static bool assign_work(struct work_struct *work, struct worker *worker, | |
1073 | struct work_struct **nextp) | |
1074 | { | |
1075 | struct worker_pool *pool = worker->pool; | |
1076 | struct worker *collision; | |
1077 | ||
1078 | lockdep_assert_held(&pool->lock); | |
1079 | ||
1080 | /* | |
1081 | * A single work shouldn't be executed concurrently by multiple workers. | |
1082 | * __queue_work() ensures that @work doesn't jump to a different pool | |
1083 | * while still running in the previous pool. Here, we should ensure that | |
1084 | * @work is not executed concurrently by multiple workers from the same | |
1085 | * pool. Check whether anyone is already processing the work. If so, | |
1086 | * defer the work to the currently executing one. | |
1087 | */ | |
1088 | collision = find_worker_executing_work(pool, work); | |
1089 | if (unlikely(collision)) { | |
1090 | move_linked_works(work, &collision->scheduled, nextp); | |
1091 | return false; | |
1092 | } | |
1093 | ||
1094 | move_linked_works(work, &worker->scheduled, nextp); | |
1095 | return true; | |
1096 | } | |
1097 | ||
797e8345 | 1098 | /** |
0219a352 TH |
1099 | * kick_pool - wake up an idle worker if necessary |
1100 | * @pool: pool to kick | |
797e8345 | 1101 | * |
0219a352 TH |
1102 | * @pool may have pending work items. Wake up worker if necessary. Returns |
1103 | * whether a worker was woken up. | |
797e8345 | 1104 | */ |
0219a352 | 1105 | static bool kick_pool(struct worker_pool *pool) |
797e8345 TH |
1106 | { |
1107 | struct worker *worker = first_idle_worker(pool); | |
8639eceb | 1108 | struct task_struct *p; |
797e8345 | 1109 | |
0219a352 TH |
1110 | lockdep_assert_held(&pool->lock); |
1111 | ||
1112 | if (!need_more_worker(pool) || !worker) | |
1113 | return false; | |
1114 | ||
8639eceb TH |
1115 | p = worker->task; |
1116 | ||
1117 | #ifdef CONFIG_SMP | |
1118 | /* | |
1119 | * Idle @worker is about to execute @work and waking up provides an | |
1120 | * opportunity to migrate @worker at a lower cost by setting the task's | |
1121 | * wake_cpu field. Let's see if we want to move @worker to improve | |
1122 | * execution locality. | |
1123 | * | |
1124 | * We're waking the worker that went idle the latest and there's some | |
1125 | * chance that @worker is marked idle but hasn't gone off CPU yet. If | |
1126 | * so, setting the wake_cpu won't do anything. As this is a best-effort | |
1127 | * optimization and the race window is narrow, let's leave as-is for | |
1128 | * now. If this becomes pronounced, we can skip over workers which are | |
1129 | * still on cpu when picking an idle worker. | |
1130 | * | |
1131 | * If @pool has non-strict affinity, @worker might have ended up outside | |
1132 | * its affinity scope. Repatriate. | |
1133 | */ | |
1134 | if (!pool->attrs->affn_strict && | |
1135 | !cpumask_test_cpu(p->wake_cpu, pool->attrs->__pod_cpumask)) { | |
1136 | struct work_struct *work = list_first_entry(&pool->worklist, | |
1137 | struct work_struct, entry); | |
1138 | p->wake_cpu = cpumask_any_distribute(pool->attrs->__pod_cpumask); | |
1139 | get_work_pwq(work)->stats[PWQ_STAT_REPATRIATED]++; | |
1140 | } | |
1141 | #endif | |
1142 | wake_up_process(p); | |
0219a352 | 1143 | return true; |
797e8345 TH |
1144 | } |
1145 | ||
63638450 TH |
1146 | #ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT |
1147 | ||
1148 | /* | |
1149 | * Concurrency-managed per-cpu work items that hog CPU for longer than | |
1150 | * wq_cpu_intensive_thresh_us trigger the automatic CPU_INTENSIVE mechanism, | |
1151 | * which prevents them from stalling other concurrency-managed work items. If a | |
1152 | * work function keeps triggering this mechanism, it's likely that the work item | |
1153 | * should be using an unbound workqueue instead. | |
1154 | * | |
1155 | * wq_cpu_intensive_report() tracks work functions which trigger such conditions | |
1156 | * and report them so that they can be examined and converted to use unbound | |
1157 | * workqueues as appropriate. To avoid flooding the console, each violating work | |
1158 | * function is tracked and reported with exponential backoff. | |
1159 | */ | |
1160 | #define WCI_MAX_ENTS 128 | |
1161 | ||
1162 | struct wci_ent { | |
1163 | work_func_t func; | |
1164 | atomic64_t cnt; | |
1165 | struct hlist_node hash_node; | |
1166 | }; | |
1167 | ||
1168 | static struct wci_ent wci_ents[WCI_MAX_ENTS]; | |
1169 | static int wci_nr_ents; | |
1170 | static DEFINE_RAW_SPINLOCK(wci_lock); | |
1171 | static DEFINE_HASHTABLE(wci_hash, ilog2(WCI_MAX_ENTS)); | |
1172 | ||
1173 | static struct wci_ent *wci_find_ent(work_func_t func) | |
1174 | { | |
1175 | struct wci_ent *ent; | |
1176 | ||
1177 | hash_for_each_possible_rcu(wci_hash, ent, hash_node, | |
1178 | (unsigned long)func) { | |
1179 | if (ent->func == func) | |
1180 | return ent; | |
1181 | } | |
1182 | return NULL; | |
1183 | } | |
1184 | ||
1185 | static void wq_cpu_intensive_report(work_func_t func) | |
1186 | { | |
1187 | struct wci_ent *ent; | |
1188 | ||
1189 | restart: | |
1190 | ent = wci_find_ent(func); | |
1191 | if (ent) { | |
1192 | u64 cnt; | |
1193 | ||
1194 | /* | |
1195 | * Start reporting from the fourth time and back off | |
1196 | * exponentially. | |
1197 | */ | |
1198 | cnt = atomic64_inc_return_relaxed(&ent->cnt); | |
1199 | if (cnt >= 4 && is_power_of_2(cnt)) | |
1200 | printk_deferred(KERN_WARNING "workqueue: %ps hogged CPU for >%luus %llu times, consider switching to WQ_UNBOUND\n", | |
1201 | ent->func, wq_cpu_intensive_thresh_us, | |
1202 | atomic64_read(&ent->cnt)); | |
1203 | return; | |
1204 | } | |
1205 | ||
1206 | /* | |
1207 | * @func is a new violation. Allocate a new entry for it. If wcn_ents[] | |
1208 | * is exhausted, something went really wrong and we probably made enough | |
1209 | * noise already. | |
1210 | */ | |
1211 | if (wci_nr_ents >= WCI_MAX_ENTS) | |
1212 | return; | |
1213 | ||
1214 | raw_spin_lock(&wci_lock); | |
1215 | ||
1216 | if (wci_nr_ents >= WCI_MAX_ENTS) { | |
1217 | raw_spin_unlock(&wci_lock); | |
1218 | return; | |
1219 | } | |
1220 | ||
1221 | if (wci_find_ent(func)) { | |
1222 | raw_spin_unlock(&wci_lock); | |
1223 | goto restart; | |
1224 | } | |
1225 | ||
1226 | ent = &wci_ents[wci_nr_ents++]; | |
1227 | ent->func = func; | |
1228 | atomic64_set(&ent->cnt, 1); | |
1229 | hash_add_rcu(wci_hash, &ent->hash_node, (unsigned long)func); | |
1230 | ||
1231 | raw_spin_unlock(&wci_lock); | |
1232 | } | |
1233 | ||
1234 | #else /* CONFIG_WQ_CPU_INTENSIVE_REPORT */ | |
1235 | static void wq_cpu_intensive_report(work_func_t func) {} | |
1236 | #endif /* CONFIG_WQ_CPU_INTENSIVE_REPORT */ | |
1237 | ||
d302f017 | 1238 | /** |
6d25be57 | 1239 | * wq_worker_running - a worker is running again |
e22bee78 | 1240 | * @task: task waking up |
e22bee78 | 1241 | * |
6d25be57 | 1242 | * This function is called when a worker returns from schedule() |
e22bee78 | 1243 | */ |
6d25be57 | 1244 | void wq_worker_running(struct task_struct *task) |
e22bee78 TH |
1245 | { |
1246 | struct worker *worker = kthread_data(task); | |
1247 | ||
c8f6219b | 1248 | if (!READ_ONCE(worker->sleeping)) |
6d25be57 | 1249 | return; |
07edfece FW |
1250 | |
1251 | /* | |
1252 | * If preempted by unbind_workers() between the WORKER_NOT_RUNNING check | |
1253 | * and the nr_running increment below, we may ruin the nr_running reset | |
1254 | * and leave with an unexpected pool->nr_running == 1 on the newly unbound | |
1255 | * pool. Protect against such race. | |
1256 | */ | |
1257 | preempt_disable(); | |
6d25be57 | 1258 | if (!(worker->flags & WORKER_NOT_RUNNING)) |
bc35f7ef | 1259 | worker->pool->nr_running++; |
07edfece | 1260 | preempt_enable(); |
616db877 TH |
1261 | |
1262 | /* | |
1263 | * CPU intensive auto-detection cares about how long a work item hogged | |
1264 | * CPU without sleeping. Reset the starting timestamp on wakeup. | |
1265 | */ | |
1266 | worker->current_at = worker->task->se.sum_exec_runtime; | |
1267 | ||
c8f6219b | 1268 | WRITE_ONCE(worker->sleeping, 0); |
e22bee78 TH |
1269 | } |
1270 | ||
1271 | /** | |
1272 | * wq_worker_sleeping - a worker is going to sleep | |
1273 | * @task: task going to sleep | |
e22bee78 | 1274 | * |
6d25be57 | 1275 | * This function is called from schedule() when a busy worker is |
ccf45156 | 1276 | * going to sleep. |
e22bee78 | 1277 | */ |
6d25be57 | 1278 | void wq_worker_sleeping(struct task_struct *task) |
e22bee78 | 1279 | { |
cc5bff38 | 1280 | struct worker *worker = kthread_data(task); |
111c225a | 1281 | struct worker_pool *pool; |
e22bee78 | 1282 | |
111c225a TH |
1283 | /* |
1284 | * Rescuers, which may not have all the fields set up like normal | |
1285 | * workers, also reach here, let's not access anything before | |
1286 | * checking NOT_RUNNING. | |
1287 | */ | |
2d64672e | 1288 | if (worker->flags & WORKER_NOT_RUNNING) |
6d25be57 | 1289 | return; |
e22bee78 | 1290 | |
111c225a | 1291 | pool = worker->pool; |
111c225a | 1292 | |
62849a96 | 1293 | /* Return if preempted before wq_worker_running() was reached */ |
c8f6219b | 1294 | if (READ_ONCE(worker->sleeping)) |
6d25be57 TG |
1295 | return; |
1296 | ||
c8f6219b | 1297 | WRITE_ONCE(worker->sleeping, 1); |
a9b8a985 | 1298 | raw_spin_lock_irq(&pool->lock); |
e22bee78 | 1299 | |
45c753f5 FW |
1300 | /* |
1301 | * Recheck in case unbind_workers() preempted us. We don't | |
1302 | * want to decrement nr_running after the worker is unbound | |
1303 | * and nr_running has been reset. | |
1304 | */ | |
1305 | if (worker->flags & WORKER_NOT_RUNNING) { | |
1306 | raw_spin_unlock_irq(&pool->lock); | |
1307 | return; | |
1308 | } | |
1309 | ||
bc35f7ef | 1310 | pool->nr_running--; |
0219a352 | 1311 | if (kick_pool(pool)) |
725e8ec5 | 1312 | worker->current_pwq->stats[PWQ_STAT_CM_WAKEUP]++; |
0219a352 | 1313 | |
a9b8a985 | 1314 | raw_spin_unlock_irq(&pool->lock); |
e22bee78 TH |
1315 | } |
1316 | ||
616db877 TH |
1317 | /** |
1318 | * wq_worker_tick - a scheduler tick occurred while a kworker is running | |
1319 | * @task: task currently running | |
1320 | * | |
1321 | * Called from scheduler_tick(). We're in the IRQ context and the current | |
1322 | * worker's fields which follow the 'K' locking rule can be accessed safely. | |
1323 | */ | |
1324 | void wq_worker_tick(struct task_struct *task) | |
1325 | { | |
1326 | struct worker *worker = kthread_data(task); | |
1327 | struct pool_workqueue *pwq = worker->current_pwq; | |
1328 | struct worker_pool *pool = worker->pool; | |
1329 | ||
1330 | if (!pwq) | |
1331 | return; | |
1332 | ||
8a1dd1e5 TH |
1333 | pwq->stats[PWQ_STAT_CPU_TIME] += TICK_USEC; |
1334 | ||
18c8ae81 Z |
1335 | if (!wq_cpu_intensive_thresh_us) |
1336 | return; | |
1337 | ||
616db877 TH |
1338 | /* |
1339 | * If the current worker is concurrency managed and hogged the CPU for | |
1340 | * longer than wq_cpu_intensive_thresh_us, it's automatically marked | |
1341 | * CPU_INTENSIVE to avoid stalling other concurrency-managed work items. | |
c8f6219b Z |
1342 | * |
1343 | * Set @worker->sleeping means that @worker is in the process of | |
1344 | * switching out voluntarily and won't be contributing to | |
1345 | * @pool->nr_running until it wakes up. As wq_worker_sleeping() also | |
1346 | * decrements ->nr_running, setting CPU_INTENSIVE here can lead to | |
1347 | * double decrements. The task is releasing the CPU anyway. Let's skip. | |
1348 | * We probably want to make this prettier in the future. | |
616db877 | 1349 | */ |
c8f6219b | 1350 | if ((worker->flags & WORKER_NOT_RUNNING) || READ_ONCE(worker->sleeping) || |
616db877 TH |
1351 | worker->task->se.sum_exec_runtime - worker->current_at < |
1352 | wq_cpu_intensive_thresh_us * NSEC_PER_USEC) | |
1353 | return; | |
1354 | ||
1355 | raw_spin_lock(&pool->lock); | |
1356 | ||
1357 | worker_set_flags(worker, WORKER_CPU_INTENSIVE); | |
63638450 | 1358 | wq_cpu_intensive_report(worker->current_func); |
616db877 TH |
1359 | pwq->stats[PWQ_STAT_CPU_INTENSIVE]++; |
1360 | ||
0219a352 | 1361 | if (kick_pool(pool)) |
616db877 | 1362 | pwq->stats[PWQ_STAT_CM_WAKEUP]++; |
616db877 TH |
1363 | |
1364 | raw_spin_unlock(&pool->lock); | |
1365 | } | |
1366 | ||
1b69ac6b JW |
1367 | /** |
1368 | * wq_worker_last_func - retrieve worker's last work function | |
8194fe94 | 1369 | * @task: Task to retrieve last work function of. |
1b69ac6b JW |
1370 | * |
1371 | * Determine the last function a worker executed. This is called from | |
1372 | * the scheduler to get a worker's last known identity. | |
1373 | * | |
1374 | * CONTEXT: | |
a9b8a985 | 1375 | * raw_spin_lock_irq(rq->lock) |
1b69ac6b | 1376 | * |
4b047002 JW |
1377 | * This function is called during schedule() when a kworker is going |
1378 | * to sleep. It's used by psi to identify aggregation workers during | |
1379 | * dequeuing, to allow periodic aggregation to shut-off when that | |
1380 | * worker is the last task in the system or cgroup to go to sleep. | |
1381 | * | |
1382 | * As this function doesn't involve any workqueue-related locking, it | |
1383 | * only returns stable values when called from inside the scheduler's | |
1384 | * queuing and dequeuing paths, when @task, which must be a kworker, | |
1385 | * is guaranteed to not be processing any works. | |
1386 | * | |
1b69ac6b JW |
1387 | * Return: |
1388 | * The last work function %current executed as a worker, NULL if it | |
1389 | * hasn't executed any work yet. | |
1390 | */ | |
1391 | work_func_t wq_worker_last_func(struct task_struct *task) | |
1392 | { | |
1393 | struct worker *worker = kthread_data(task); | |
1394 | ||
1395 | return worker->last_func; | |
1396 | } | |
1397 | ||
8864b4e5 TH |
1398 | /** |
1399 | * get_pwq - get an extra reference on the specified pool_workqueue | |
1400 | * @pwq: pool_workqueue to get | |
1401 | * | |
1402 | * Obtain an extra reference on @pwq. The caller should guarantee that | |
1403 | * @pwq has positive refcnt and be holding the matching pool->lock. | |
1404 | */ | |
1405 | static void get_pwq(struct pool_workqueue *pwq) | |
1406 | { | |
1407 | lockdep_assert_held(&pwq->pool->lock); | |
1408 | WARN_ON_ONCE(pwq->refcnt <= 0); | |
1409 | pwq->refcnt++; | |
1410 | } | |
1411 | ||
1412 | /** | |
1413 | * put_pwq - put a pool_workqueue reference | |
1414 | * @pwq: pool_workqueue to put | |
1415 | * | |
1416 | * Drop a reference of @pwq. If its refcnt reaches zero, schedule its | |
1417 | * destruction. The caller should be holding the matching pool->lock. | |
1418 | */ | |
1419 | static void put_pwq(struct pool_workqueue *pwq) | |
1420 | { | |
1421 | lockdep_assert_held(&pwq->pool->lock); | |
1422 | if (likely(--pwq->refcnt)) | |
1423 | return; | |
8864b4e5 | 1424 | /* |
967b494e TH |
1425 | * @pwq can't be released under pool->lock, bounce to a dedicated |
1426 | * kthread_worker to avoid A-A deadlocks. | |
8864b4e5 | 1427 | */ |
687a9aa5 | 1428 | kthread_queue_work(pwq_release_worker, &pwq->release_work); |
8864b4e5 TH |
1429 | } |
1430 | ||
dce90d47 TH |
1431 | /** |
1432 | * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock | |
1433 | * @pwq: pool_workqueue to put (can be %NULL) | |
1434 | * | |
1435 | * put_pwq() with locking. This function also allows %NULL @pwq. | |
1436 | */ | |
1437 | static void put_pwq_unlocked(struct pool_workqueue *pwq) | |
1438 | { | |
1439 | if (pwq) { | |
1440 | /* | |
24acfb71 | 1441 | * As both pwqs and pools are RCU protected, the |
dce90d47 TH |
1442 | * following lock operations are safe. |
1443 | */ | |
a9b8a985 | 1444 | raw_spin_lock_irq(&pwq->pool->lock); |
dce90d47 | 1445 | put_pwq(pwq); |
a9b8a985 | 1446 | raw_spin_unlock_irq(&pwq->pool->lock); |
dce90d47 TH |
1447 | } |
1448 | } | |
1449 | ||
f97a4a1a | 1450 | static void pwq_activate_inactive_work(struct work_struct *work) |
bf4ede01 | 1451 | { |
112202d9 | 1452 | struct pool_workqueue *pwq = get_work_pwq(work); |
bf4ede01 TH |
1453 | |
1454 | trace_workqueue_activate_work(work); | |
82607adc TH |
1455 | if (list_empty(&pwq->pool->worklist)) |
1456 | pwq->pool->watchdog_ts = jiffies; | |
112202d9 | 1457 | move_linked_works(work, &pwq->pool->worklist, NULL); |
f97a4a1a | 1458 | __clear_bit(WORK_STRUCT_INACTIVE_BIT, work_data_bits(work)); |
112202d9 | 1459 | pwq->nr_active++; |
bf4ede01 TH |
1460 | } |
1461 | ||
f97a4a1a | 1462 | static void pwq_activate_first_inactive(struct pool_workqueue *pwq) |
3aa62497 | 1463 | { |
f97a4a1a | 1464 | struct work_struct *work = list_first_entry(&pwq->inactive_works, |
3aa62497 LJ |
1465 | struct work_struct, entry); |
1466 | ||
f97a4a1a | 1467 | pwq_activate_inactive_work(work); |
3aa62497 LJ |
1468 | } |
1469 | ||
bf4ede01 | 1470 | /** |
112202d9 TH |
1471 | * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight |
1472 | * @pwq: pwq of interest | |
c4560c2c | 1473 | * @work_data: work_data of work which left the queue |
bf4ede01 TH |
1474 | * |
1475 | * A work either has completed or is removed from pending queue, | |
112202d9 | 1476 | * decrement nr_in_flight of its pwq and handle workqueue flushing. |
bf4ede01 TH |
1477 | * |
1478 | * CONTEXT: | |
a9b8a985 | 1479 | * raw_spin_lock_irq(pool->lock). |
bf4ede01 | 1480 | */ |
c4560c2c | 1481 | static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, unsigned long work_data) |
bf4ede01 | 1482 | { |
c4560c2c LJ |
1483 | int color = get_work_color(work_data); |
1484 | ||
018f3a13 LJ |
1485 | if (!(work_data & WORK_STRUCT_INACTIVE)) { |
1486 | pwq->nr_active--; | |
1487 | if (!list_empty(&pwq->inactive_works)) { | |
1488 | /* one down, submit an inactive one */ | |
1489 | if (pwq->nr_active < pwq->max_active) | |
1490 | pwq_activate_first_inactive(pwq); | |
1491 | } | |
1492 | } | |
1493 | ||
112202d9 | 1494 | pwq->nr_in_flight[color]--; |
bf4ede01 | 1495 | |
bf4ede01 | 1496 | /* is flush in progress and are we at the flushing tip? */ |
112202d9 | 1497 | if (likely(pwq->flush_color != color)) |
8864b4e5 | 1498 | goto out_put; |
bf4ede01 TH |
1499 | |
1500 | /* are there still in-flight works? */ | |
112202d9 | 1501 | if (pwq->nr_in_flight[color]) |
8864b4e5 | 1502 | goto out_put; |
bf4ede01 | 1503 | |
112202d9 TH |
1504 | /* this pwq is done, clear flush_color */ |
1505 | pwq->flush_color = -1; | |
bf4ede01 TH |
1506 | |
1507 | /* | |
112202d9 | 1508 | * If this was the last pwq, wake up the first flusher. It |
bf4ede01 TH |
1509 | * will handle the rest. |
1510 | */ | |
112202d9 TH |
1511 | if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) |
1512 | complete(&pwq->wq->first_flusher->done); | |
8864b4e5 TH |
1513 | out_put: |
1514 | put_pwq(pwq); | |
bf4ede01 TH |
1515 | } |
1516 | ||
36e227d2 | 1517 | /** |
bbb68dfa | 1518 | * try_to_grab_pending - steal work item from worklist and disable irq |
36e227d2 TH |
1519 | * @work: work item to steal |
1520 | * @is_dwork: @work is a delayed_work | |
bbb68dfa | 1521 | * @flags: place to store irq state |
36e227d2 TH |
1522 | * |
1523 | * Try to grab PENDING bit of @work. This function can handle @work in any | |
d185af30 | 1524 | * stable state - idle, on timer or on worklist. |
36e227d2 | 1525 | * |
d185af30 | 1526 | * Return: |
3eb6b31b MCC |
1527 | * |
1528 | * ======== ================================================================ | |
36e227d2 TH |
1529 | * 1 if @work was pending and we successfully stole PENDING |
1530 | * 0 if @work was idle and we claimed PENDING | |
1531 | * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry | |
bbb68dfa TH |
1532 | * -ENOENT if someone else is canceling @work, this state may persist |
1533 | * for arbitrarily long | |
3eb6b31b | 1534 | * ======== ================================================================ |
36e227d2 | 1535 | * |
d185af30 | 1536 | * Note: |
bbb68dfa | 1537 | * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting |
e0aecdd8 TH |
1538 | * interrupted while holding PENDING and @work off queue, irq must be |
1539 | * disabled on entry. This, combined with delayed_work->timer being | |
1540 | * irqsafe, ensures that we return -EAGAIN for finite short period of time. | |
bbb68dfa TH |
1541 | * |
1542 | * On successful return, >= 0, irq is disabled and the caller is | |
1543 | * responsible for releasing it using local_irq_restore(*@flags). | |
1544 | * | |
e0aecdd8 | 1545 | * This function is safe to call from any context including IRQ handler. |
bf4ede01 | 1546 | */ |
bbb68dfa TH |
1547 | static int try_to_grab_pending(struct work_struct *work, bool is_dwork, |
1548 | unsigned long *flags) | |
bf4ede01 | 1549 | { |
d565ed63 | 1550 | struct worker_pool *pool; |
112202d9 | 1551 | struct pool_workqueue *pwq; |
bf4ede01 | 1552 | |
bbb68dfa TH |
1553 | local_irq_save(*flags); |
1554 | ||
36e227d2 TH |
1555 | /* try to steal the timer if it exists */ |
1556 | if (is_dwork) { | |
1557 | struct delayed_work *dwork = to_delayed_work(work); | |
1558 | ||
e0aecdd8 TH |
1559 | /* |
1560 | * dwork->timer is irqsafe. If del_timer() fails, it's | |
1561 | * guaranteed that the timer is not queued anywhere and not | |
1562 | * running on the local CPU. | |
1563 | */ | |
36e227d2 TH |
1564 | if (likely(del_timer(&dwork->timer))) |
1565 | return 1; | |
1566 | } | |
1567 | ||
1568 | /* try to claim PENDING the normal way */ | |
bf4ede01 TH |
1569 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) |
1570 | return 0; | |
1571 | ||
24acfb71 | 1572 | rcu_read_lock(); |
bf4ede01 TH |
1573 | /* |
1574 | * The queueing is in progress, or it is already queued. Try to | |
1575 | * steal it from ->worklist without clearing WORK_STRUCT_PENDING. | |
1576 | */ | |
d565ed63 TH |
1577 | pool = get_work_pool(work); |
1578 | if (!pool) | |
bbb68dfa | 1579 | goto fail; |
bf4ede01 | 1580 | |
a9b8a985 | 1581 | raw_spin_lock(&pool->lock); |
0b3dae68 | 1582 | /* |
112202d9 TH |
1583 | * work->data is guaranteed to point to pwq only while the work |
1584 | * item is queued on pwq->wq, and both updating work->data to point | |
1585 | * to pwq on queueing and to pool on dequeueing are done under | |
1586 | * pwq->pool->lock. This in turn guarantees that, if work->data | |
1587 | * points to pwq which is associated with a locked pool, the work | |
0b3dae68 LJ |
1588 | * item is currently queued on that pool. |
1589 | */ | |
112202d9 TH |
1590 | pwq = get_work_pwq(work); |
1591 | if (pwq && pwq->pool == pool) { | |
16062836 TH |
1592 | debug_work_deactivate(work); |
1593 | ||
1594 | /* | |
018f3a13 LJ |
1595 | * A cancelable inactive work item must be in the |
1596 | * pwq->inactive_works since a queued barrier can't be | |
1597 | * canceled (see the comments in insert_wq_barrier()). | |
1598 | * | |
f97a4a1a | 1599 | * An inactive work item cannot be grabbed directly because |
d812796e | 1600 | * it might have linked barrier work items which, if left |
f97a4a1a | 1601 | * on the inactive_works list, will confuse pwq->nr_active |
16062836 TH |
1602 | * management later on and cause stall. Make sure the work |
1603 | * item is activated before grabbing. | |
1604 | */ | |
f97a4a1a LJ |
1605 | if (*work_data_bits(work) & WORK_STRUCT_INACTIVE) |
1606 | pwq_activate_inactive_work(work); | |
16062836 TH |
1607 | |
1608 | list_del_init(&work->entry); | |
c4560c2c | 1609 | pwq_dec_nr_in_flight(pwq, *work_data_bits(work)); |
16062836 | 1610 | |
112202d9 | 1611 | /* work->data points to pwq iff queued, point to pool */ |
16062836 TH |
1612 | set_work_pool_and_keep_pending(work, pool->id); |
1613 | ||
a9b8a985 | 1614 | raw_spin_unlock(&pool->lock); |
24acfb71 | 1615 | rcu_read_unlock(); |
16062836 | 1616 | return 1; |
bf4ede01 | 1617 | } |
a9b8a985 | 1618 | raw_spin_unlock(&pool->lock); |
bbb68dfa | 1619 | fail: |
24acfb71 | 1620 | rcu_read_unlock(); |
bbb68dfa TH |
1621 | local_irq_restore(*flags); |
1622 | if (work_is_canceling(work)) | |
1623 | return -ENOENT; | |
1624 | cpu_relax(); | |
36e227d2 | 1625 | return -EAGAIN; |
bf4ede01 TH |
1626 | } |
1627 | ||
4690c4ab | 1628 | /** |
706026c2 | 1629 | * insert_work - insert a work into a pool |
112202d9 | 1630 | * @pwq: pwq @work belongs to |
4690c4ab TH |
1631 | * @work: work to insert |
1632 | * @head: insertion point | |
1633 | * @extra_flags: extra WORK_STRUCT_* flags to set | |
1634 | * | |
112202d9 | 1635 | * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to |
706026c2 | 1636 | * work_struct flags. |
4690c4ab TH |
1637 | * |
1638 | * CONTEXT: | |
a9b8a985 | 1639 | * raw_spin_lock_irq(pool->lock). |
4690c4ab | 1640 | */ |
112202d9 TH |
1641 | static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, |
1642 | struct list_head *head, unsigned int extra_flags) | |
b89deed3 | 1643 | { |
fe089f87 | 1644 | debug_work_activate(work); |
e22bee78 | 1645 | |
e89a85d6 | 1646 | /* record the work call stack in order to print it in KASAN reports */ |
f70da745 | 1647 | kasan_record_aux_stack_noalloc(work); |
e89a85d6 | 1648 | |
4690c4ab | 1649 | /* we own @work, set data and link */ |
112202d9 | 1650 | set_work_pwq(work, pwq, extra_flags); |
1a4d9b0a | 1651 | list_add_tail(&work->entry, head); |
8864b4e5 | 1652 | get_pwq(pwq); |
b89deed3 ON |
1653 | } |
1654 | ||
c8efcc25 TH |
1655 | /* |
1656 | * Test whether @work is being queued from another work executing on the | |
8d03ecfe | 1657 | * same workqueue. |
c8efcc25 TH |
1658 | */ |
1659 | static bool is_chained_work(struct workqueue_struct *wq) | |
1660 | { | |
8d03ecfe TH |
1661 | struct worker *worker; |
1662 | ||
1663 | worker = current_wq_worker(); | |
1664 | /* | |
bf393fd4 | 1665 | * Return %true iff I'm a worker executing a work item on @wq. If |
8d03ecfe TH |
1666 | * I'm @worker, it's safe to dereference it without locking. |
1667 | */ | |
112202d9 | 1668 | return worker && worker->current_pwq->wq == wq; |
c8efcc25 TH |
1669 | } |
1670 | ||
ef557180 MG |
1671 | /* |
1672 | * When queueing an unbound work item to a wq, prefer local CPU if allowed | |
1673 | * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to | |
1674 | * avoid perturbing sensitive tasks. | |
1675 | */ | |
1676 | static int wq_select_unbound_cpu(int cpu) | |
1677 | { | |
1678 | int new_cpu; | |
1679 | ||
f303fccb TH |
1680 | if (likely(!wq_debug_force_rr_cpu)) { |
1681 | if (cpumask_test_cpu(cpu, wq_unbound_cpumask)) | |
1682 | return cpu; | |
a8ec5880 AF |
1683 | } else { |
1684 | pr_warn_once("workqueue: round-robin CPU selection forced, expect performance impact\n"); | |
f303fccb TH |
1685 | } |
1686 | ||
ef557180 MG |
1687 | if (cpumask_empty(wq_unbound_cpumask)) |
1688 | return cpu; | |
1689 | ||
1690 | new_cpu = __this_cpu_read(wq_rr_cpu_last); | |
1691 | new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask); | |
1692 | if (unlikely(new_cpu >= nr_cpu_ids)) { | |
1693 | new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask); | |
1694 | if (unlikely(new_cpu >= nr_cpu_ids)) | |
1695 | return cpu; | |
1696 | } | |
1697 | __this_cpu_write(wq_rr_cpu_last, new_cpu); | |
1698 | ||
1699 | return new_cpu; | |
1700 | } | |
1701 | ||
d84ff051 | 1702 | static void __queue_work(int cpu, struct workqueue_struct *wq, |
1da177e4 LT |
1703 | struct work_struct *work) |
1704 | { | |
112202d9 | 1705 | struct pool_workqueue *pwq; |
fe089f87 | 1706 | struct worker_pool *last_pool, *pool; |
8a2e8e5d | 1707 | unsigned int work_flags; |
b75cac93 | 1708 | unsigned int req_cpu = cpu; |
8930caba TH |
1709 | |
1710 | /* | |
1711 | * While a work item is PENDING && off queue, a task trying to | |
1712 | * steal the PENDING will busy-loop waiting for it to either get | |
1713 | * queued or lose PENDING. Grabbing PENDING and queueing should | |
1714 | * happen with IRQ disabled. | |
1715 | */ | |
8e8eb730 | 1716 | lockdep_assert_irqs_disabled(); |
1da177e4 | 1717 | |
1e19ffc6 | 1718 | |
33e3f0a3 RC |
1719 | /* |
1720 | * For a draining wq, only works from the same workqueue are | |
1721 | * allowed. The __WQ_DESTROYING helps to spot the issue that | |
1722 | * queues a new work item to a wq after destroy_workqueue(wq). | |
1723 | */ | |
1724 | if (unlikely(wq->flags & (__WQ_DESTROYING | __WQ_DRAINING) && | |
1725 | WARN_ON_ONCE(!is_chained_work(wq)))) | |
e41e704b | 1726 | return; |
24acfb71 | 1727 | rcu_read_lock(); |
9e8cd2f5 | 1728 | retry: |
c9178087 | 1729 | /* pwq which will be used unless @work is executing elsewhere */ |
636b927e TH |
1730 | if (req_cpu == WORK_CPU_UNBOUND) { |
1731 | if (wq->flags & WQ_UNBOUND) | |
aa202f1f | 1732 | cpu = wq_select_unbound_cpu(raw_smp_processor_id()); |
636b927e | 1733 | else |
aa202f1f | 1734 | cpu = raw_smp_processor_id(); |
aa202f1f | 1735 | } |
dbf2576e | 1736 | |
636b927e | 1737 | pwq = rcu_dereference(*per_cpu_ptr(wq->cpu_pwq, cpu)); |
fe089f87 TH |
1738 | pool = pwq->pool; |
1739 | ||
c9178087 TH |
1740 | /* |
1741 | * If @work was previously on a different pool, it might still be | |
1742 | * running there, in which case the work needs to be queued on that | |
1743 | * pool to guarantee non-reentrancy. | |
1744 | */ | |
1745 | last_pool = get_work_pool(work); | |
fe089f87 | 1746 | if (last_pool && last_pool != pool) { |
c9178087 | 1747 | struct worker *worker; |
18aa9eff | 1748 | |
a9b8a985 | 1749 | raw_spin_lock(&last_pool->lock); |
18aa9eff | 1750 | |
c9178087 | 1751 | worker = find_worker_executing_work(last_pool, work); |
18aa9eff | 1752 | |
c9178087 TH |
1753 | if (worker && worker->current_pwq->wq == wq) { |
1754 | pwq = worker->current_pwq; | |
fe089f87 TH |
1755 | pool = pwq->pool; |
1756 | WARN_ON_ONCE(pool != last_pool); | |
8930caba | 1757 | } else { |
c9178087 | 1758 | /* meh... not running there, queue here */ |
a9b8a985 | 1759 | raw_spin_unlock(&last_pool->lock); |
fe089f87 | 1760 | raw_spin_lock(&pool->lock); |
8930caba | 1761 | } |
f3421797 | 1762 | } else { |
fe089f87 | 1763 | raw_spin_lock(&pool->lock); |
502ca9d8 TH |
1764 | } |
1765 | ||
9e8cd2f5 | 1766 | /* |
636b927e TH |
1767 | * pwq is determined and locked. For unbound pools, we could have raced |
1768 | * with pwq release and it could already be dead. If its refcnt is zero, | |
1769 | * repeat pwq selection. Note that unbound pwqs never die without | |
1770 | * another pwq replacing it in cpu_pwq or while work items are executing | |
1771 | * on it, so the retrying is guaranteed to make forward-progress. | |
9e8cd2f5 TH |
1772 | */ |
1773 | if (unlikely(!pwq->refcnt)) { | |
1774 | if (wq->flags & WQ_UNBOUND) { | |
fe089f87 | 1775 | raw_spin_unlock(&pool->lock); |
9e8cd2f5 TH |
1776 | cpu_relax(); |
1777 | goto retry; | |
1778 | } | |
1779 | /* oops */ | |
1780 | WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", | |
1781 | wq->name, cpu); | |
1782 | } | |
1783 | ||
112202d9 TH |
1784 | /* pwq determined, queue */ |
1785 | trace_workqueue_queue_work(req_cpu, pwq, work); | |
502ca9d8 | 1786 | |
24acfb71 TG |
1787 | if (WARN_ON(!list_empty(&work->entry))) |
1788 | goto out; | |
1e19ffc6 | 1789 | |
112202d9 TH |
1790 | pwq->nr_in_flight[pwq->work_color]++; |
1791 | work_flags = work_color_to_flags(pwq->work_color); | |
1e19ffc6 | 1792 | |
112202d9 | 1793 | if (likely(pwq->nr_active < pwq->max_active)) { |
fe089f87 TH |
1794 | if (list_empty(&pool->worklist)) |
1795 | pool->watchdog_ts = jiffies; | |
1796 | ||
cdadf009 | 1797 | trace_workqueue_activate_work(work); |
112202d9 | 1798 | pwq->nr_active++; |
fe089f87 | 1799 | insert_work(pwq, work, &pool->worklist, work_flags); |
0219a352 | 1800 | kick_pool(pool); |
8a2e8e5d | 1801 | } else { |
f97a4a1a | 1802 | work_flags |= WORK_STRUCT_INACTIVE; |
fe089f87 | 1803 | insert_work(pwq, work, &pwq->inactive_works, work_flags); |
8a2e8e5d | 1804 | } |
1e19ffc6 | 1805 | |
24acfb71 | 1806 | out: |
fe089f87 | 1807 | raw_spin_unlock(&pool->lock); |
24acfb71 | 1808 | rcu_read_unlock(); |
1da177e4 LT |
1809 | } |
1810 | ||
0fcb78c2 | 1811 | /** |
c1a220e7 ZR |
1812 | * queue_work_on - queue work on specific cpu |
1813 | * @cpu: CPU number to execute work on | |
0fcb78c2 REB |
1814 | * @wq: workqueue to use |
1815 | * @work: work to queue | |
1816 | * | |
c1a220e7 | 1817 | * We queue the work to a specific CPU, the caller must ensure it |
443378f0 PM |
1818 | * can't go away. Callers that fail to ensure that the specified |
1819 | * CPU cannot go away will execute on a randomly chosen CPU. | |
854f5cc5 PM |
1820 | * But note well that callers specifying a CPU that never has been |
1821 | * online will get a splat. | |
d185af30 YB |
1822 | * |
1823 | * Return: %false if @work was already on a queue, %true otherwise. | |
1da177e4 | 1824 | */ |
d4283e93 TH |
1825 | bool queue_work_on(int cpu, struct workqueue_struct *wq, |
1826 | struct work_struct *work) | |
1da177e4 | 1827 | { |
d4283e93 | 1828 | bool ret = false; |
8930caba | 1829 | unsigned long flags; |
ef1ca236 | 1830 | |
8930caba | 1831 | local_irq_save(flags); |
c1a220e7 | 1832 | |
22df02bb | 1833 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { |
4690c4ab | 1834 | __queue_work(cpu, wq, work); |
d4283e93 | 1835 | ret = true; |
c1a220e7 | 1836 | } |
ef1ca236 | 1837 | |
8930caba | 1838 | local_irq_restore(flags); |
1da177e4 LT |
1839 | return ret; |
1840 | } | |
ad7b1f84 | 1841 | EXPORT_SYMBOL(queue_work_on); |
1da177e4 | 1842 | |
8204e0c1 | 1843 | /** |
fef59c9c | 1844 | * select_numa_node_cpu - Select a CPU based on NUMA node |
8204e0c1 AD |
1845 | * @node: NUMA node ID that we want to select a CPU from |
1846 | * | |
1847 | * This function will attempt to find a "random" cpu available on a given | |
1848 | * node. If there are no CPUs available on the given node it will return | |
1849 | * WORK_CPU_UNBOUND indicating that we should just schedule to any | |
1850 | * available CPU if we need to schedule this work. | |
1851 | */ | |
fef59c9c | 1852 | static int select_numa_node_cpu(int node) |
8204e0c1 AD |
1853 | { |
1854 | int cpu; | |
1855 | ||
8204e0c1 AD |
1856 | /* Delay binding to CPU if node is not valid or online */ |
1857 | if (node < 0 || node >= MAX_NUMNODES || !node_online(node)) | |
1858 | return WORK_CPU_UNBOUND; | |
1859 | ||
1860 | /* Use local node/cpu if we are already there */ | |
1861 | cpu = raw_smp_processor_id(); | |
1862 | if (node == cpu_to_node(cpu)) | |
1863 | return cpu; | |
1864 | ||
1865 | /* Use "random" otherwise know as "first" online CPU of node */ | |
1866 | cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask); | |
1867 | ||
1868 | /* If CPU is valid return that, otherwise just defer */ | |
1869 | return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND; | |
1870 | } | |
1871 | ||
1872 | /** | |
1873 | * queue_work_node - queue work on a "random" cpu for a given NUMA node | |
1874 | * @node: NUMA node that we are targeting the work for | |
1875 | * @wq: workqueue to use | |
1876 | * @work: work to queue | |
1877 | * | |
1878 | * We queue the work to a "random" CPU within a given NUMA node. The basic | |
1879 | * idea here is to provide a way to somehow associate work with a given | |
1880 | * NUMA node. | |
1881 | * | |
1882 | * This function will only make a best effort attempt at getting this onto | |
1883 | * the right NUMA node. If no node is requested or the requested node is | |
1884 | * offline then we just fall back to standard queue_work behavior. | |
1885 | * | |
1886 | * Currently the "random" CPU ends up being the first available CPU in the | |
1887 | * intersection of cpu_online_mask and the cpumask of the node, unless we | |
1888 | * are running on the node. In that case we just use the current CPU. | |
1889 | * | |
1890 | * Return: %false if @work was already on a queue, %true otherwise. | |
1891 | */ | |
1892 | bool queue_work_node(int node, struct workqueue_struct *wq, | |
1893 | struct work_struct *work) | |
1894 | { | |
1895 | unsigned long flags; | |
1896 | bool ret = false; | |
1897 | ||
1898 | /* | |
1899 | * This current implementation is specific to unbound workqueues. | |
1900 | * Specifically we only return the first available CPU for a given | |
1901 | * node instead of cycling through individual CPUs within the node. | |
1902 | * | |
1903 | * If this is used with a per-cpu workqueue then the logic in | |
1904 | * workqueue_select_cpu_near would need to be updated to allow for | |
1905 | * some round robin type logic. | |
1906 | */ | |
1907 | WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)); | |
1908 | ||
1909 | local_irq_save(flags); | |
1910 | ||
1911 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { | |
fef59c9c | 1912 | int cpu = select_numa_node_cpu(node); |
8204e0c1 AD |
1913 | |
1914 | __queue_work(cpu, wq, work); | |
1915 | ret = true; | |
1916 | } | |
1917 | ||
1918 | local_irq_restore(flags); | |
1919 | return ret; | |
1920 | } | |
1921 | EXPORT_SYMBOL_GPL(queue_work_node); | |
1922 | ||
8c20feb6 | 1923 | void delayed_work_timer_fn(struct timer_list *t) |
1da177e4 | 1924 | { |
8c20feb6 | 1925 | struct delayed_work *dwork = from_timer(dwork, t, timer); |
1da177e4 | 1926 | |
e0aecdd8 | 1927 | /* should have been called from irqsafe timer with irq already off */ |
60c057bc | 1928 | __queue_work(dwork->cpu, dwork->wq, &dwork->work); |
1da177e4 | 1929 | } |
1438ade5 | 1930 | EXPORT_SYMBOL(delayed_work_timer_fn); |
1da177e4 | 1931 | |
7beb2edf TH |
1932 | static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, |
1933 | struct delayed_work *dwork, unsigned long delay) | |
1da177e4 | 1934 | { |
7beb2edf TH |
1935 | struct timer_list *timer = &dwork->timer; |
1936 | struct work_struct *work = &dwork->work; | |
7beb2edf | 1937 | |
637fdbae | 1938 | WARN_ON_ONCE(!wq); |
4b243563 | 1939 | WARN_ON_ONCE(timer->function != delayed_work_timer_fn); |
fc4b514f TH |
1940 | WARN_ON_ONCE(timer_pending(timer)); |
1941 | WARN_ON_ONCE(!list_empty(&work->entry)); | |
7beb2edf | 1942 | |
8852aac2 TH |
1943 | /* |
1944 | * If @delay is 0, queue @dwork->work immediately. This is for | |
1945 | * both optimization and correctness. The earliest @timer can | |
1946 | * expire is on the closest next tick and delayed_work users depend | |
1947 | * on that there's no such delay when @delay is 0. | |
1948 | */ | |
1949 | if (!delay) { | |
1950 | __queue_work(cpu, wq, &dwork->work); | |
1951 | return; | |
1952 | } | |
1953 | ||
60c057bc | 1954 | dwork->wq = wq; |
1265057f | 1955 | dwork->cpu = cpu; |
7beb2edf TH |
1956 | timer->expires = jiffies + delay; |
1957 | ||
041bd12e TH |
1958 | if (unlikely(cpu != WORK_CPU_UNBOUND)) |
1959 | add_timer_on(timer, cpu); | |
1960 | else | |
1961 | add_timer(timer); | |
1da177e4 LT |
1962 | } |
1963 | ||
0fcb78c2 REB |
1964 | /** |
1965 | * queue_delayed_work_on - queue work on specific CPU after delay | |
1966 | * @cpu: CPU number to execute work on | |
1967 | * @wq: workqueue to use | |
af9997e4 | 1968 | * @dwork: work to queue |
0fcb78c2 REB |
1969 | * @delay: number of jiffies to wait before queueing |
1970 | * | |
d185af30 | 1971 | * Return: %false if @work was already on a queue, %true otherwise. If |
715f1300 TH |
1972 | * @delay is zero and @dwork is idle, it will be scheduled for immediate |
1973 | * execution. | |
0fcb78c2 | 1974 | */ |
d4283e93 TH |
1975 | bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, |
1976 | struct delayed_work *dwork, unsigned long delay) | |
7a6bc1cd | 1977 | { |
52bad64d | 1978 | struct work_struct *work = &dwork->work; |
d4283e93 | 1979 | bool ret = false; |
8930caba | 1980 | unsigned long flags; |
7a6bc1cd | 1981 | |
8930caba TH |
1982 | /* read the comment in __queue_work() */ |
1983 | local_irq_save(flags); | |
7a6bc1cd | 1984 | |
22df02bb | 1985 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { |
7beb2edf | 1986 | __queue_delayed_work(cpu, wq, dwork, delay); |
d4283e93 | 1987 | ret = true; |
7a6bc1cd | 1988 | } |
8a3e77cc | 1989 | |
8930caba | 1990 | local_irq_restore(flags); |
7a6bc1cd VP |
1991 | return ret; |
1992 | } | |
ad7b1f84 | 1993 | EXPORT_SYMBOL(queue_delayed_work_on); |
c7fc77f7 | 1994 | |
8376fe22 TH |
1995 | /** |
1996 | * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU | |
1997 | * @cpu: CPU number to execute work on | |
1998 | * @wq: workqueue to use | |
1999 | * @dwork: work to queue | |
2000 | * @delay: number of jiffies to wait before queueing | |
2001 | * | |
2002 | * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, | |
2003 | * modify @dwork's timer so that it expires after @delay. If @delay is | |
2004 | * zero, @work is guaranteed to be scheduled immediately regardless of its | |
2005 | * current state. | |
2006 | * | |
d185af30 | 2007 | * Return: %false if @dwork was idle and queued, %true if @dwork was |
8376fe22 TH |
2008 | * pending and its timer was modified. |
2009 | * | |
e0aecdd8 | 2010 | * This function is safe to call from any context including IRQ handler. |
8376fe22 TH |
2011 | * See try_to_grab_pending() for details. |
2012 | */ | |
2013 | bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, | |
2014 | struct delayed_work *dwork, unsigned long delay) | |
2015 | { | |
2016 | unsigned long flags; | |
2017 | int ret; | |
c7fc77f7 | 2018 | |
8376fe22 TH |
2019 | do { |
2020 | ret = try_to_grab_pending(&dwork->work, true, &flags); | |
2021 | } while (unlikely(ret == -EAGAIN)); | |
63bc0362 | 2022 | |
8376fe22 TH |
2023 | if (likely(ret >= 0)) { |
2024 | __queue_delayed_work(cpu, wq, dwork, delay); | |
2025 | local_irq_restore(flags); | |
7a6bc1cd | 2026 | } |
8376fe22 TH |
2027 | |
2028 | /* -ENOENT from try_to_grab_pending() becomes %true */ | |
7a6bc1cd VP |
2029 | return ret; |
2030 | } | |
8376fe22 TH |
2031 | EXPORT_SYMBOL_GPL(mod_delayed_work_on); |
2032 | ||
05f0fe6b TH |
2033 | static void rcu_work_rcufn(struct rcu_head *rcu) |
2034 | { | |
2035 | struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu); | |
2036 | ||
2037 | /* read the comment in __queue_work() */ | |
2038 | local_irq_disable(); | |
2039 | __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work); | |
2040 | local_irq_enable(); | |
2041 | } | |
2042 | ||
2043 | /** | |
2044 | * queue_rcu_work - queue work after a RCU grace period | |
2045 | * @wq: workqueue to use | |
2046 | * @rwork: work to queue | |
2047 | * | |
2048 | * Return: %false if @rwork was already pending, %true otherwise. Note | |
2049 | * that a full RCU grace period is guaranteed only after a %true return. | |
bf393fd4 | 2050 | * While @rwork is guaranteed to be executed after a %false return, the |
05f0fe6b TH |
2051 | * execution may happen before a full RCU grace period has passed. |
2052 | */ | |
2053 | bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork) | |
2054 | { | |
2055 | struct work_struct *work = &rwork->work; | |
2056 | ||
2057 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { | |
2058 | rwork->wq = wq; | |
a7e30c0e | 2059 | call_rcu_hurry(&rwork->rcu, rcu_work_rcufn); |
05f0fe6b TH |
2060 | return true; |
2061 | } | |
2062 | ||
2063 | return false; | |
2064 | } | |
2065 | EXPORT_SYMBOL(queue_rcu_work); | |
2066 | ||
f7537df5 | 2067 | static struct worker *alloc_worker(int node) |
c34056a3 TH |
2068 | { |
2069 | struct worker *worker; | |
2070 | ||
f7537df5 | 2071 | worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node); |
c8e55f36 TH |
2072 | if (worker) { |
2073 | INIT_LIST_HEAD(&worker->entry); | |
affee4b2 | 2074 | INIT_LIST_HEAD(&worker->scheduled); |
da028469 | 2075 | INIT_LIST_HEAD(&worker->node); |
e22bee78 TH |
2076 | /* on creation a worker is in !idle && prep state */ |
2077 | worker->flags = WORKER_PREP; | |
c8e55f36 | 2078 | } |
c34056a3 TH |
2079 | return worker; |
2080 | } | |
2081 | ||
9546b29e TH |
2082 | static cpumask_t *pool_allowed_cpus(struct worker_pool *pool) |
2083 | { | |
8639eceb TH |
2084 | if (pool->cpu < 0 && pool->attrs->affn_strict) |
2085 | return pool->attrs->__pod_cpumask; | |
2086 | else | |
2087 | return pool->attrs->cpumask; | |
9546b29e TH |
2088 | } |
2089 | ||
4736cbf7 LJ |
2090 | /** |
2091 | * worker_attach_to_pool() - attach a worker to a pool | |
2092 | * @worker: worker to be attached | |
2093 | * @pool: the target pool | |
2094 | * | |
2095 | * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and | |
2096 | * cpu-binding of @worker are kept coordinated with the pool across | |
2097 | * cpu-[un]hotplugs. | |
2098 | */ | |
2099 | static void worker_attach_to_pool(struct worker *worker, | |
2100 | struct worker_pool *pool) | |
2101 | { | |
1258fae7 | 2102 | mutex_lock(&wq_pool_attach_mutex); |
4736cbf7 | 2103 | |
4736cbf7 | 2104 | /* |
1258fae7 TH |
2105 | * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains |
2106 | * stable across this function. See the comments above the flag | |
2107 | * definition for details. | |
4736cbf7 LJ |
2108 | */ |
2109 | if (pool->flags & POOL_DISASSOCIATED) | |
2110 | worker->flags |= WORKER_UNBOUND; | |
5c25b5ff PZ |
2111 | else |
2112 | kthread_set_per_cpu(worker->task, pool->cpu); | |
4736cbf7 | 2113 | |
640f17c8 | 2114 | if (worker->rescue_wq) |
9546b29e | 2115 | set_cpus_allowed_ptr(worker->task, pool_allowed_cpus(pool)); |
640f17c8 | 2116 | |
4736cbf7 | 2117 | list_add_tail(&worker->node, &pool->workers); |
a2d812a2 | 2118 | worker->pool = pool; |
4736cbf7 | 2119 | |
1258fae7 | 2120 | mutex_unlock(&wq_pool_attach_mutex); |
4736cbf7 LJ |
2121 | } |
2122 | ||
60f5a4bc LJ |
2123 | /** |
2124 | * worker_detach_from_pool() - detach a worker from its pool | |
2125 | * @worker: worker which is attached to its pool | |
60f5a4bc | 2126 | * |
4736cbf7 LJ |
2127 | * Undo the attaching which had been done in worker_attach_to_pool(). The |
2128 | * caller worker shouldn't access to the pool after detached except it has | |
2129 | * other reference to the pool. | |
60f5a4bc | 2130 | */ |
a2d812a2 | 2131 | static void worker_detach_from_pool(struct worker *worker) |
60f5a4bc | 2132 | { |
a2d812a2 | 2133 | struct worker_pool *pool = worker->pool; |
60f5a4bc LJ |
2134 | struct completion *detach_completion = NULL; |
2135 | ||
1258fae7 | 2136 | mutex_lock(&wq_pool_attach_mutex); |
a2d812a2 | 2137 | |
5c25b5ff | 2138 | kthread_set_per_cpu(worker->task, -1); |
da028469 | 2139 | list_del(&worker->node); |
a2d812a2 TH |
2140 | worker->pool = NULL; |
2141 | ||
e02b9312 | 2142 | if (list_empty(&pool->workers) && list_empty(&pool->dying_workers)) |
60f5a4bc | 2143 | detach_completion = pool->detach_completion; |
1258fae7 | 2144 | mutex_unlock(&wq_pool_attach_mutex); |
60f5a4bc | 2145 | |
b62c0751 LJ |
2146 | /* clear leftover flags without pool->lock after it is detached */ |
2147 | worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND); | |
2148 | ||
60f5a4bc LJ |
2149 | if (detach_completion) |
2150 | complete(detach_completion); | |
2151 | } | |
2152 | ||
c34056a3 TH |
2153 | /** |
2154 | * create_worker - create a new workqueue worker | |
63d95a91 | 2155 | * @pool: pool the new worker will belong to |
c34056a3 | 2156 | * |
051e1850 | 2157 | * Create and start a new worker which is attached to @pool. |
c34056a3 TH |
2158 | * |
2159 | * CONTEXT: | |
2160 | * Might sleep. Does GFP_KERNEL allocations. | |
2161 | * | |
d185af30 | 2162 | * Return: |
c34056a3 TH |
2163 | * Pointer to the newly created worker. |
2164 | */ | |
bc2ae0f5 | 2165 | static struct worker *create_worker(struct worker_pool *pool) |
c34056a3 | 2166 | { |
e441b56f ZL |
2167 | struct worker *worker; |
2168 | int id; | |
e3c916a4 | 2169 | char id_buf[16]; |
c34056a3 | 2170 | |
7cda9aae | 2171 | /* ID is needed to determine kthread name */ |
e441b56f | 2172 | id = ida_alloc(&pool->worker_ida, GFP_KERNEL); |
3f0ea0b8 PM |
2173 | if (id < 0) { |
2174 | pr_err_once("workqueue: Failed to allocate a worker ID: %pe\n", | |
2175 | ERR_PTR(id)); | |
e441b56f | 2176 | return NULL; |
3f0ea0b8 | 2177 | } |
c34056a3 | 2178 | |
f7537df5 | 2179 | worker = alloc_worker(pool->node); |
3f0ea0b8 PM |
2180 | if (!worker) { |
2181 | pr_err_once("workqueue: Failed to allocate a worker\n"); | |
c34056a3 | 2182 | goto fail; |
3f0ea0b8 | 2183 | } |
c34056a3 | 2184 | |
c34056a3 TH |
2185 | worker->id = id; |
2186 | ||
29c91e99 | 2187 | if (pool->cpu >= 0) |
e3c916a4 TH |
2188 | snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, |
2189 | pool->attrs->nice < 0 ? "H" : ""); | |
f3421797 | 2190 | else |
e3c916a4 TH |
2191 | snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); |
2192 | ||
f3f90ad4 | 2193 | worker->task = kthread_create_on_node(worker_thread, worker, pool->node, |
e3c916a4 | 2194 | "kworker/%s", id_buf); |
3f0ea0b8 | 2195 | if (IS_ERR(worker->task)) { |
60f54038 PM |
2196 | if (PTR_ERR(worker->task) == -EINTR) { |
2197 | pr_err("workqueue: Interrupted when creating a worker thread \"kworker/%s\"\n", | |
2198 | id_buf); | |
2199 | } else { | |
2200 | pr_err_once("workqueue: Failed to create a worker thread: %pe", | |
2201 | worker->task); | |
2202 | } | |
c34056a3 | 2203 | goto fail; |
3f0ea0b8 | 2204 | } |
c34056a3 | 2205 | |
91151228 | 2206 | set_user_nice(worker->task, pool->attrs->nice); |
9546b29e | 2207 | kthread_bind_mask(worker->task, pool_allowed_cpus(pool)); |
91151228 | 2208 | |
da028469 | 2209 | /* successful, attach the worker to the pool */ |
4736cbf7 | 2210 | worker_attach_to_pool(worker, pool); |
822d8405 | 2211 | |
051e1850 | 2212 | /* start the newly created worker */ |
a9b8a985 | 2213 | raw_spin_lock_irq(&pool->lock); |
0219a352 | 2214 | |
051e1850 LJ |
2215 | worker->pool->nr_workers++; |
2216 | worker_enter_idle(worker); | |
0219a352 TH |
2217 | kick_pool(pool); |
2218 | ||
2219 | /* | |
2220 | * @worker is waiting on a completion in kthread() and will trigger hung | |
2221 | * check if not woken up soon. As kick_pool() might not have waken it | |
2222 | * up, wake it up explicitly once more. | |
2223 | */ | |
051e1850 | 2224 | wake_up_process(worker->task); |
0219a352 | 2225 | |
a9b8a985 | 2226 | raw_spin_unlock_irq(&pool->lock); |
051e1850 | 2227 | |
c34056a3 | 2228 | return worker; |
822d8405 | 2229 | |
c34056a3 | 2230 | fail: |
e441b56f | 2231 | ida_free(&pool->worker_ida, id); |
c34056a3 TH |
2232 | kfree(worker); |
2233 | return NULL; | |
2234 | } | |
2235 | ||
793777bc VS |
2236 | static void unbind_worker(struct worker *worker) |
2237 | { | |
2238 | lockdep_assert_held(&wq_pool_attach_mutex); | |
2239 | ||
2240 | kthread_set_per_cpu(worker->task, -1); | |
2241 | if (cpumask_intersects(wq_unbound_cpumask, cpu_active_mask)) | |
2242 | WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, wq_unbound_cpumask) < 0); | |
2243 | else | |
2244 | WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0); | |
2245 | } | |
2246 | ||
e02b9312 VS |
2247 | static void wake_dying_workers(struct list_head *cull_list) |
2248 | { | |
2249 | struct worker *worker, *tmp; | |
2250 | ||
2251 | list_for_each_entry_safe(worker, tmp, cull_list, entry) { | |
2252 | list_del_init(&worker->entry); | |
2253 | unbind_worker(worker); | |
2254 | /* | |
2255 | * If the worker was somehow already running, then it had to be | |
2256 | * in pool->idle_list when set_worker_dying() happened or we | |
2257 | * wouldn't have gotten here. | |
2258 | * | |
2259 | * Thus, the worker must either have observed the WORKER_DIE | |
2260 | * flag, or have set its state to TASK_IDLE. Either way, the | |
2261 | * below will be observed by the worker and is safe to do | |
2262 | * outside of pool->lock. | |
2263 | */ | |
2264 | wake_up_process(worker->task); | |
2265 | } | |
2266 | } | |
2267 | ||
c34056a3 | 2268 | /** |
e02b9312 | 2269 | * set_worker_dying - Tag a worker for destruction |
c34056a3 | 2270 | * @worker: worker to be destroyed |
e02b9312 | 2271 | * @list: transfer worker away from its pool->idle_list and into list |
c34056a3 | 2272 | * |
e02b9312 VS |
2273 | * Tag @worker for destruction and adjust @pool stats accordingly. The worker |
2274 | * should be idle. | |
c8e55f36 TH |
2275 | * |
2276 | * CONTEXT: | |
a9b8a985 | 2277 | * raw_spin_lock_irq(pool->lock). |
c34056a3 | 2278 | */ |
e02b9312 | 2279 | static void set_worker_dying(struct worker *worker, struct list_head *list) |
c34056a3 | 2280 | { |
bd7bdd43 | 2281 | struct worker_pool *pool = worker->pool; |
c34056a3 | 2282 | |
cd549687 | 2283 | lockdep_assert_held(&pool->lock); |
e02b9312 | 2284 | lockdep_assert_held(&wq_pool_attach_mutex); |
cd549687 | 2285 | |
c34056a3 | 2286 | /* sanity check frenzy */ |
6183c009 | 2287 | if (WARN_ON(worker->current_work) || |
73eb7fe7 LJ |
2288 | WARN_ON(!list_empty(&worker->scheduled)) || |
2289 | WARN_ON(!(worker->flags & WORKER_IDLE))) | |
6183c009 | 2290 | return; |
c34056a3 | 2291 | |
73eb7fe7 LJ |
2292 | pool->nr_workers--; |
2293 | pool->nr_idle--; | |
5bdfff96 | 2294 | |
cb444766 | 2295 | worker->flags |= WORKER_DIE; |
e02b9312 VS |
2296 | |
2297 | list_move(&worker->entry, list); | |
2298 | list_move(&worker->node, &pool->dying_workers); | |
c34056a3 TH |
2299 | } |
2300 | ||
3f959aa3 VS |
2301 | /** |
2302 | * idle_worker_timeout - check if some idle workers can now be deleted. | |
2303 | * @t: The pool's idle_timer that just expired | |
2304 | * | |
2305 | * The timer is armed in worker_enter_idle(). Note that it isn't disarmed in | |
2306 | * worker_leave_idle(), as a worker flicking between idle and active while its | |
2307 | * pool is at the too_many_workers() tipping point would cause too much timer | |
2308 | * housekeeping overhead. Since IDLE_WORKER_TIMEOUT is long enough, we just let | |
2309 | * it expire and re-evaluate things from there. | |
2310 | */ | |
32a6c723 | 2311 | static void idle_worker_timeout(struct timer_list *t) |
e22bee78 | 2312 | { |
32a6c723 | 2313 | struct worker_pool *pool = from_timer(pool, t, idle_timer); |
3f959aa3 VS |
2314 | bool do_cull = false; |
2315 | ||
2316 | if (work_pending(&pool->idle_cull_work)) | |
2317 | return; | |
e22bee78 | 2318 | |
a9b8a985 | 2319 | raw_spin_lock_irq(&pool->lock); |
e22bee78 | 2320 | |
3f959aa3 | 2321 | if (too_many_workers(pool)) { |
e22bee78 TH |
2322 | struct worker *worker; |
2323 | unsigned long expires; | |
2324 | ||
2325 | /* idle_list is kept in LIFO order, check the last one */ | |
3f959aa3 VS |
2326 | worker = list_entry(pool->idle_list.prev, struct worker, entry); |
2327 | expires = worker->last_active + IDLE_WORKER_TIMEOUT; | |
2328 | do_cull = !time_before(jiffies, expires); | |
2329 | ||
2330 | if (!do_cull) | |
2331 | mod_timer(&pool->idle_timer, expires); | |
2332 | } | |
2333 | raw_spin_unlock_irq(&pool->lock); | |
2334 | ||
2335 | if (do_cull) | |
2336 | queue_work(system_unbound_wq, &pool->idle_cull_work); | |
2337 | } | |
2338 | ||
2339 | /** | |
2340 | * idle_cull_fn - cull workers that have been idle for too long. | |
2341 | * @work: the pool's work for handling these idle workers | |
2342 | * | |
2343 | * This goes through a pool's idle workers and gets rid of those that have been | |
2344 | * idle for at least IDLE_WORKER_TIMEOUT seconds. | |
e02b9312 VS |
2345 | * |
2346 | * We don't want to disturb isolated CPUs because of a pcpu kworker being | |
2347 | * culled, so this also resets worker affinity. This requires a sleepable | |
2348 | * context, hence the split between timer callback and work item. | |
3f959aa3 VS |
2349 | */ |
2350 | static void idle_cull_fn(struct work_struct *work) | |
2351 | { | |
2352 | struct worker_pool *pool = container_of(work, struct worker_pool, idle_cull_work); | |
9680540c | 2353 | LIST_HEAD(cull_list); |
3f959aa3 | 2354 | |
e02b9312 VS |
2355 | /* |
2356 | * Grabbing wq_pool_attach_mutex here ensures an already-running worker | |
2357 | * cannot proceed beyong worker_detach_from_pool() in its self-destruct | |
2358 | * path. This is required as a previously-preempted worker could run after | |
2359 | * set_worker_dying() has happened but before wake_dying_workers() did. | |
2360 | */ | |
2361 | mutex_lock(&wq_pool_attach_mutex); | |
3f959aa3 VS |
2362 | raw_spin_lock_irq(&pool->lock); |
2363 | ||
2364 | while (too_many_workers(pool)) { | |
2365 | struct worker *worker; | |
2366 | unsigned long expires; | |
2367 | ||
63d95a91 | 2368 | worker = list_entry(pool->idle_list.prev, struct worker, entry); |
e22bee78 TH |
2369 | expires = worker->last_active + IDLE_WORKER_TIMEOUT; |
2370 | ||
3347fc9f | 2371 | if (time_before(jiffies, expires)) { |
63d95a91 | 2372 | mod_timer(&pool->idle_timer, expires); |
3347fc9f | 2373 | break; |
d5abe669 | 2374 | } |
3347fc9f | 2375 | |
e02b9312 | 2376 | set_worker_dying(worker, &cull_list); |
e22bee78 TH |
2377 | } |
2378 | ||
a9b8a985 | 2379 | raw_spin_unlock_irq(&pool->lock); |
e02b9312 VS |
2380 | wake_dying_workers(&cull_list); |
2381 | mutex_unlock(&wq_pool_attach_mutex); | |
e22bee78 | 2382 | } |
d5abe669 | 2383 | |
493a1724 | 2384 | static void send_mayday(struct work_struct *work) |
e22bee78 | 2385 | { |
112202d9 TH |
2386 | struct pool_workqueue *pwq = get_work_pwq(work); |
2387 | struct workqueue_struct *wq = pwq->wq; | |
493a1724 | 2388 | |
2e109a28 | 2389 | lockdep_assert_held(&wq_mayday_lock); |
e22bee78 | 2390 | |
493008a8 | 2391 | if (!wq->rescuer) |
493a1724 | 2392 | return; |
e22bee78 TH |
2393 | |
2394 | /* mayday mayday mayday */ | |
493a1724 | 2395 | if (list_empty(&pwq->mayday_node)) { |
77668c8b LJ |
2396 | /* |
2397 | * If @pwq is for an unbound wq, its base ref may be put at | |
2398 | * any time due to an attribute change. Pin @pwq until the | |
2399 | * rescuer is done with it. | |
2400 | */ | |
2401 | get_pwq(pwq); | |
493a1724 | 2402 | list_add_tail(&pwq->mayday_node, &wq->maydays); |
e22bee78 | 2403 | wake_up_process(wq->rescuer->task); |
725e8ec5 | 2404 | pwq->stats[PWQ_STAT_MAYDAY]++; |
493a1724 | 2405 | } |
e22bee78 TH |
2406 | } |
2407 | ||
32a6c723 | 2408 | static void pool_mayday_timeout(struct timer_list *t) |
e22bee78 | 2409 | { |
32a6c723 | 2410 | struct worker_pool *pool = from_timer(pool, t, mayday_timer); |
e22bee78 TH |
2411 | struct work_struct *work; |
2412 | ||
a9b8a985 SAS |
2413 | raw_spin_lock_irq(&pool->lock); |
2414 | raw_spin_lock(&wq_mayday_lock); /* for wq->maydays */ | |
e22bee78 | 2415 | |
63d95a91 | 2416 | if (need_to_create_worker(pool)) { |
e22bee78 TH |
2417 | /* |
2418 | * We've been trying to create a new worker but | |
2419 | * haven't been successful. We might be hitting an | |
2420 | * allocation deadlock. Send distress signals to | |
2421 | * rescuers. | |
2422 | */ | |
63d95a91 | 2423 | list_for_each_entry(work, &pool->worklist, entry) |
e22bee78 | 2424 | send_mayday(work); |
1da177e4 | 2425 | } |
e22bee78 | 2426 | |
a9b8a985 SAS |
2427 | raw_spin_unlock(&wq_mayday_lock); |
2428 | raw_spin_unlock_irq(&pool->lock); | |
e22bee78 | 2429 | |
63d95a91 | 2430 | mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); |
1da177e4 LT |
2431 | } |
2432 | ||
e22bee78 TH |
2433 | /** |
2434 | * maybe_create_worker - create a new worker if necessary | |
63d95a91 | 2435 | * @pool: pool to create a new worker for |
e22bee78 | 2436 | * |
63d95a91 | 2437 | * Create a new worker for @pool if necessary. @pool is guaranteed to |
e22bee78 TH |
2438 | * have at least one idle worker on return from this function. If |
2439 | * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is | |
63d95a91 | 2440 | * sent to all rescuers with works scheduled on @pool to resolve |
e22bee78 TH |
2441 | * possible allocation deadlock. |
2442 | * | |
c5aa87bb TH |
2443 | * On return, need_to_create_worker() is guaranteed to be %false and |
2444 | * may_start_working() %true. | |
e22bee78 TH |
2445 | * |
2446 | * LOCKING: | |
a9b8a985 | 2447 | * raw_spin_lock_irq(pool->lock) which may be released and regrabbed |
e22bee78 TH |
2448 | * multiple times. Does GFP_KERNEL allocations. Called only from |
2449 | * manager. | |
e22bee78 | 2450 | */ |
29187a9e | 2451 | static void maybe_create_worker(struct worker_pool *pool) |
d565ed63 TH |
2452 | __releases(&pool->lock) |
2453 | __acquires(&pool->lock) | |
1da177e4 | 2454 | { |
e22bee78 | 2455 | restart: |
a9b8a985 | 2456 | raw_spin_unlock_irq(&pool->lock); |
9f9c2364 | 2457 | |
e22bee78 | 2458 | /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ |
63d95a91 | 2459 | mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); |
e22bee78 TH |
2460 | |
2461 | while (true) { | |
051e1850 | 2462 | if (create_worker(pool) || !need_to_create_worker(pool)) |
e22bee78 | 2463 | break; |
1da177e4 | 2464 | |
e212f361 | 2465 | schedule_timeout_interruptible(CREATE_COOLDOWN); |
9f9c2364 | 2466 | |
63d95a91 | 2467 | if (!need_to_create_worker(pool)) |
e22bee78 TH |
2468 | break; |
2469 | } | |
2470 | ||
63d95a91 | 2471 | del_timer_sync(&pool->mayday_timer); |
a9b8a985 | 2472 | raw_spin_lock_irq(&pool->lock); |
051e1850 LJ |
2473 | /* |
2474 | * This is necessary even after a new worker was just successfully | |
2475 | * created as @pool->lock was dropped and the new worker might have | |
2476 | * already become busy. | |
2477 | */ | |
63d95a91 | 2478 | if (need_to_create_worker(pool)) |
e22bee78 | 2479 | goto restart; |
e22bee78 TH |
2480 | } |
2481 | ||
73f53c4a | 2482 | /** |
e22bee78 TH |
2483 | * manage_workers - manage worker pool |
2484 | * @worker: self | |
73f53c4a | 2485 | * |
706026c2 | 2486 | * Assume the manager role and manage the worker pool @worker belongs |
e22bee78 | 2487 | * to. At any given time, there can be only zero or one manager per |
706026c2 | 2488 | * pool. The exclusion is handled automatically by this function. |
e22bee78 TH |
2489 | * |
2490 | * The caller can safely start processing works on false return. On | |
2491 | * true return, it's guaranteed that need_to_create_worker() is false | |
2492 | * and may_start_working() is true. | |
73f53c4a TH |
2493 | * |
2494 | * CONTEXT: | |
a9b8a985 | 2495 | * raw_spin_lock_irq(pool->lock) which may be released and regrabbed |
e22bee78 TH |
2496 | * multiple times. Does GFP_KERNEL allocations. |
2497 | * | |
d185af30 | 2498 | * Return: |
29187a9e TH |
2499 | * %false if the pool doesn't need management and the caller can safely |
2500 | * start processing works, %true if management function was performed and | |
2501 | * the conditions that the caller verified before calling the function may | |
2502 | * no longer be true. | |
73f53c4a | 2503 | */ |
e22bee78 | 2504 | static bool manage_workers(struct worker *worker) |
73f53c4a | 2505 | { |
63d95a91 | 2506 | struct worker_pool *pool = worker->pool; |
73f53c4a | 2507 | |
692b4825 | 2508 | if (pool->flags & POOL_MANAGER_ACTIVE) |
29187a9e | 2509 | return false; |
692b4825 TH |
2510 | |
2511 | pool->flags |= POOL_MANAGER_ACTIVE; | |
2607d7a6 | 2512 | pool->manager = worker; |
1e19ffc6 | 2513 | |
29187a9e | 2514 | maybe_create_worker(pool); |
e22bee78 | 2515 | |
2607d7a6 | 2516 | pool->manager = NULL; |
692b4825 | 2517 | pool->flags &= ~POOL_MANAGER_ACTIVE; |
d8bb65ab | 2518 | rcuwait_wake_up(&manager_wait); |
29187a9e | 2519 | return true; |
73f53c4a TH |
2520 | } |
2521 | ||
a62428c0 TH |
2522 | /** |
2523 | * process_one_work - process single work | |
c34056a3 | 2524 | * @worker: self |
a62428c0 TH |
2525 | * @work: work to process |
2526 | * | |
2527 | * Process @work. This function contains all the logics necessary to | |
2528 | * process a single work including synchronization against and | |
2529 | * interaction with other workers on the same cpu, queueing and | |
2530 | * flushing. As long as context requirement is met, any worker can | |
2531 | * call this function to process a work. | |
2532 | * | |
2533 | * CONTEXT: | |
a9b8a985 | 2534 | * raw_spin_lock_irq(pool->lock) which is released and regrabbed. |
a62428c0 | 2535 | */ |
c34056a3 | 2536 | static void process_one_work(struct worker *worker, struct work_struct *work) |
d565ed63 TH |
2537 | __releases(&pool->lock) |
2538 | __acquires(&pool->lock) | |
a62428c0 | 2539 | { |
112202d9 | 2540 | struct pool_workqueue *pwq = get_work_pwq(work); |
bd7bdd43 | 2541 | struct worker_pool *pool = worker->pool; |
c4560c2c | 2542 | unsigned long work_data; |
a62428c0 TH |
2543 | #ifdef CONFIG_LOCKDEP |
2544 | /* | |
2545 | * It is permissible to free the struct work_struct from | |
2546 | * inside the function that is called from it, this we need to | |
2547 | * take into account for lockdep too. To avoid bogus "held | |
2548 | * lock freed" warnings as well as problems when looking into | |
2549 | * work->lockdep_map, make a copy and use that here. | |
2550 | */ | |
4d82a1de PZ |
2551 | struct lockdep_map lockdep_map; |
2552 | ||
2553 | lockdep_copy_map(&lockdep_map, &work->lockdep_map); | |
a62428c0 | 2554 | #endif |
807407c0 | 2555 | /* ensure we're on the correct CPU */ |
85327af6 | 2556 | WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && |
ec22ca5e | 2557 | raw_smp_processor_id() != pool->cpu); |
25511a47 | 2558 | |
8930caba | 2559 | /* claim and dequeue */ |
a62428c0 | 2560 | debug_work_deactivate(work); |
c9e7cf27 | 2561 | hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); |
c34056a3 | 2562 | worker->current_work = work; |
a2c1c57b | 2563 | worker->current_func = work->func; |
112202d9 | 2564 | worker->current_pwq = pwq; |
616db877 | 2565 | worker->current_at = worker->task->se.sum_exec_runtime; |
c4560c2c | 2566 | work_data = *work_data_bits(work); |
d812796e | 2567 | worker->current_color = get_work_color(work_data); |
7a22ad75 | 2568 | |
8bf89593 TH |
2569 | /* |
2570 | * Record wq name for cmdline and debug reporting, may get | |
2571 | * overridden through set_worker_desc(). | |
2572 | */ | |
2573 | strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN); | |
2574 | ||
a62428c0 TH |
2575 | list_del_init(&work->entry); |
2576 | ||
fb0e7beb | 2577 | /* |
228f1d00 LJ |
2578 | * CPU intensive works don't participate in concurrency management. |
2579 | * They're the scheduler's responsibility. This takes @worker out | |
2580 | * of concurrency management and the next code block will chain | |
2581 | * execution of the pending work items. | |
fb0e7beb | 2582 | */ |
616db877 | 2583 | if (unlikely(pwq->wq->flags & WQ_CPU_INTENSIVE)) |
228f1d00 | 2584 | worker_set_flags(worker, WORKER_CPU_INTENSIVE); |
fb0e7beb | 2585 | |
974271c4 | 2586 | /* |
0219a352 TH |
2587 | * Kick @pool if necessary. It's always noop for per-cpu worker pools |
2588 | * since nr_running would always be >= 1 at this point. This is used to | |
2589 | * chain execution of the pending work items for WORKER_NOT_RUNNING | |
2590 | * workers such as the UNBOUND and CPU_INTENSIVE ones. | |
974271c4 | 2591 | */ |
0219a352 | 2592 | kick_pool(pool); |
974271c4 | 2593 | |
8930caba | 2594 | /* |
7c3eed5c | 2595 | * Record the last pool and clear PENDING which should be the last |
d565ed63 | 2596 | * update to @work. Also, do this inside @pool->lock so that |
23657bb1 TH |
2597 | * PENDING and queued state changes happen together while IRQ is |
2598 | * disabled. | |
8930caba | 2599 | */ |
7c3eed5c | 2600 | set_work_pool_and_clear_pending(work, pool->id); |
a62428c0 | 2601 | |
fe48ba7d | 2602 | pwq->stats[PWQ_STAT_STARTED]++; |
a9b8a985 | 2603 | raw_spin_unlock_irq(&pool->lock); |
a62428c0 | 2604 | |
a1d14934 | 2605 | lock_map_acquire(&pwq->wq->lockdep_map); |
a62428c0 | 2606 | lock_map_acquire(&lockdep_map); |
e6f3faa7 | 2607 | /* |
f52be570 PZ |
2608 | * Strictly speaking we should mark the invariant state without holding |
2609 | * any locks, that is, before these two lock_map_acquire()'s. | |
e6f3faa7 PZ |
2610 | * |
2611 | * However, that would result in: | |
2612 | * | |
2613 | * A(W1) | |
2614 | * WFC(C) | |
2615 | * A(W1) | |
2616 | * C(C) | |
2617 | * | |
2618 | * Which would create W1->C->W1 dependencies, even though there is no | |
2619 | * actual deadlock possible. There are two solutions, using a | |
2620 | * read-recursive acquire on the work(queue) 'locks', but this will then | |
f52be570 | 2621 | * hit the lockdep limitation on recursive locks, or simply discard |
e6f3faa7 PZ |
2622 | * these locks. |
2623 | * | |
2624 | * AFAICT there is no possible deadlock scenario between the | |
2625 | * flush_work() and complete() primitives (except for single-threaded | |
2626 | * workqueues), so hiding them isn't a problem. | |
2627 | */ | |
f52be570 | 2628 | lockdep_invariant_state(true); |
e36c886a | 2629 | trace_workqueue_execute_start(work); |
a2c1c57b | 2630 | worker->current_func(work); |
e36c886a AV |
2631 | /* |
2632 | * While we must be careful to not use "work" after this, the trace | |
2633 | * point will only record its address. | |
2634 | */ | |
1c5da0ec | 2635 | trace_workqueue_execute_end(work, worker->current_func); |
725e8ec5 | 2636 | pwq->stats[PWQ_STAT_COMPLETED]++; |
a62428c0 | 2637 | lock_map_release(&lockdep_map); |
112202d9 | 2638 | lock_map_release(&pwq->wq->lockdep_map); |
a62428c0 TH |
2639 | |
2640 | if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { | |
044c782c | 2641 | pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" |
d75f773c | 2642 | " last function: %ps\n", |
a2c1c57b TH |
2643 | current->comm, preempt_count(), task_pid_nr(current), |
2644 | worker->current_func); | |
a62428c0 TH |
2645 | debug_show_held_locks(current); |
2646 | dump_stack(); | |
2647 | } | |
2648 | ||
b22ce278 | 2649 | /* |
025f50f3 | 2650 | * The following prevents a kworker from hogging CPU on !PREEMPTION |
b22ce278 TH |
2651 | * kernels, where a requeueing work item waiting for something to |
2652 | * happen could deadlock with stop_machine as such work item could | |
2653 | * indefinitely requeue itself while all other CPUs are trapped in | |
789cbbec JL |
2654 | * stop_machine. At the same time, report a quiescent RCU state so |
2655 | * the same condition doesn't freeze RCU. | |
b22ce278 | 2656 | */ |
a7e6425e | 2657 | cond_resched(); |
b22ce278 | 2658 | |
a9b8a985 | 2659 | raw_spin_lock_irq(&pool->lock); |
a62428c0 | 2660 | |
616db877 TH |
2661 | /* |
2662 | * In addition to %WQ_CPU_INTENSIVE, @worker may also have been marked | |
2663 | * CPU intensive by wq_worker_tick() if @work hogged CPU longer than | |
2664 | * wq_cpu_intensive_thresh_us. Clear it. | |
2665 | */ | |
2666 | worker_clr_flags(worker, WORKER_CPU_INTENSIVE); | |
fb0e7beb | 2667 | |
1b69ac6b JW |
2668 | /* tag the worker for identification in schedule() */ |
2669 | worker->last_func = worker->current_func; | |
2670 | ||
a62428c0 | 2671 | /* we're done with it, release */ |
42f8570f | 2672 | hash_del(&worker->hentry); |
c34056a3 | 2673 | worker->current_work = NULL; |
a2c1c57b | 2674 | worker->current_func = NULL; |
112202d9 | 2675 | worker->current_pwq = NULL; |
d812796e | 2676 | worker->current_color = INT_MAX; |
c4560c2c | 2677 | pwq_dec_nr_in_flight(pwq, work_data); |
a62428c0 TH |
2678 | } |
2679 | ||
affee4b2 TH |
2680 | /** |
2681 | * process_scheduled_works - process scheduled works | |
2682 | * @worker: self | |
2683 | * | |
2684 | * Process all scheduled works. Please note that the scheduled list | |
2685 | * may change while processing a work, so this function repeatedly | |
2686 | * fetches a work from the top and executes it. | |
2687 | * | |
2688 | * CONTEXT: | |
a9b8a985 | 2689 | * raw_spin_lock_irq(pool->lock) which may be released and regrabbed |
affee4b2 TH |
2690 | * multiple times. |
2691 | */ | |
2692 | static void process_scheduled_works(struct worker *worker) | |
1da177e4 | 2693 | { |
c0ab017d TH |
2694 | struct work_struct *work; |
2695 | bool first = true; | |
2696 | ||
2697 | while ((work = list_first_entry_or_null(&worker->scheduled, | |
2698 | struct work_struct, entry))) { | |
2699 | if (first) { | |
2700 | worker->pool->watchdog_ts = jiffies; | |
2701 | first = false; | |
2702 | } | |
c34056a3 | 2703 | process_one_work(worker, work); |
1da177e4 | 2704 | } |
1da177e4 LT |
2705 | } |
2706 | ||
197f6acc TH |
2707 | static void set_pf_worker(bool val) |
2708 | { | |
2709 | mutex_lock(&wq_pool_attach_mutex); | |
2710 | if (val) | |
2711 | current->flags |= PF_WQ_WORKER; | |
2712 | else | |
2713 | current->flags &= ~PF_WQ_WORKER; | |
2714 | mutex_unlock(&wq_pool_attach_mutex); | |
2715 | } | |
2716 | ||
4690c4ab TH |
2717 | /** |
2718 | * worker_thread - the worker thread function | |
c34056a3 | 2719 | * @__worker: self |
4690c4ab | 2720 | * |
c5aa87bb TH |
2721 | * The worker thread function. All workers belong to a worker_pool - |
2722 | * either a per-cpu one or dynamic unbound one. These workers process all | |
2723 | * work items regardless of their specific target workqueue. The only | |
2724 | * exception is work items which belong to workqueues with a rescuer which | |
2725 | * will be explained in rescuer_thread(). | |
d185af30 YB |
2726 | * |
2727 | * Return: 0 | |
4690c4ab | 2728 | */ |
c34056a3 | 2729 | static int worker_thread(void *__worker) |
1da177e4 | 2730 | { |
c34056a3 | 2731 | struct worker *worker = __worker; |
bd7bdd43 | 2732 | struct worker_pool *pool = worker->pool; |
1da177e4 | 2733 | |
e22bee78 | 2734 | /* tell the scheduler that this is a workqueue worker */ |
197f6acc | 2735 | set_pf_worker(true); |
c8e55f36 | 2736 | woke_up: |
a9b8a985 | 2737 | raw_spin_lock_irq(&pool->lock); |
1da177e4 | 2738 | |
a9ab775b TH |
2739 | /* am I supposed to die? */ |
2740 | if (unlikely(worker->flags & WORKER_DIE)) { | |
a9b8a985 | 2741 | raw_spin_unlock_irq(&pool->lock); |
197f6acc | 2742 | set_pf_worker(false); |
60f5a4bc LJ |
2743 | |
2744 | set_task_comm(worker->task, "kworker/dying"); | |
e441b56f | 2745 | ida_free(&pool->worker_ida, worker->id); |
a2d812a2 | 2746 | worker_detach_from_pool(worker); |
e02b9312 | 2747 | WARN_ON_ONCE(!list_empty(&worker->entry)); |
60f5a4bc | 2748 | kfree(worker); |
a9ab775b | 2749 | return 0; |
c8e55f36 | 2750 | } |
affee4b2 | 2751 | |
c8e55f36 | 2752 | worker_leave_idle(worker); |
db7bccf4 | 2753 | recheck: |
e22bee78 | 2754 | /* no more worker necessary? */ |
63d95a91 | 2755 | if (!need_more_worker(pool)) |
e22bee78 TH |
2756 | goto sleep; |
2757 | ||
2758 | /* do we need to manage? */ | |
63d95a91 | 2759 | if (unlikely(!may_start_working(pool)) && manage_workers(worker)) |
e22bee78 TH |
2760 | goto recheck; |
2761 | ||
c8e55f36 TH |
2762 | /* |
2763 | * ->scheduled list can only be filled while a worker is | |
2764 | * preparing to process a work or actually processing it. | |
2765 | * Make sure nobody diddled with it while I was sleeping. | |
2766 | */ | |
6183c009 | 2767 | WARN_ON_ONCE(!list_empty(&worker->scheduled)); |
c8e55f36 | 2768 | |
e22bee78 | 2769 | /* |
a9ab775b TH |
2770 | * Finish PREP stage. We're guaranteed to have at least one idle |
2771 | * worker or that someone else has already assumed the manager | |
2772 | * role. This is where @worker starts participating in concurrency | |
2773 | * management if applicable and concurrency management is restored | |
2774 | * after being rebound. See rebind_workers() for details. | |
e22bee78 | 2775 | */ |
a9ab775b | 2776 | worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); |
e22bee78 TH |
2777 | |
2778 | do { | |
c8e55f36 | 2779 | struct work_struct *work = |
bd7bdd43 | 2780 | list_first_entry(&pool->worklist, |
c8e55f36 TH |
2781 | struct work_struct, entry); |
2782 | ||
873eaca6 TH |
2783 | if (assign_work(work, worker, NULL)) |
2784 | process_scheduled_works(worker); | |
63d95a91 | 2785 | } while (keep_working(pool)); |
e22bee78 | 2786 | |
228f1d00 | 2787 | worker_set_flags(worker, WORKER_PREP); |
d313dd85 | 2788 | sleep: |
c8e55f36 | 2789 | /* |
d565ed63 TH |
2790 | * pool->lock is held and there's no work to process and no need to |
2791 | * manage, sleep. Workers are woken up only while holding | |
2792 | * pool->lock or from local cpu, so setting the current state | |
2793 | * before releasing pool->lock is enough to prevent losing any | |
2794 | * event. | |
c8e55f36 TH |
2795 | */ |
2796 | worker_enter_idle(worker); | |
c5a94a61 | 2797 | __set_current_state(TASK_IDLE); |
a9b8a985 | 2798 | raw_spin_unlock_irq(&pool->lock); |
c8e55f36 TH |
2799 | schedule(); |
2800 | goto woke_up; | |
1da177e4 LT |
2801 | } |
2802 | ||
e22bee78 TH |
2803 | /** |
2804 | * rescuer_thread - the rescuer thread function | |
111c225a | 2805 | * @__rescuer: self |
e22bee78 TH |
2806 | * |
2807 | * Workqueue rescuer thread function. There's one rescuer for each | |
493008a8 | 2808 | * workqueue which has WQ_MEM_RECLAIM set. |
e22bee78 | 2809 | * |
706026c2 | 2810 | * Regular work processing on a pool may block trying to create a new |
e22bee78 TH |
2811 | * worker which uses GFP_KERNEL allocation which has slight chance of |
2812 | * developing into deadlock if some works currently on the same queue | |
2813 | * need to be processed to satisfy the GFP_KERNEL allocation. This is | |
2814 | * the problem rescuer solves. | |
2815 | * | |
706026c2 TH |
2816 | * When such condition is possible, the pool summons rescuers of all |
2817 | * workqueues which have works queued on the pool and let them process | |
e22bee78 TH |
2818 | * those works so that forward progress can be guaranteed. |
2819 | * | |
2820 | * This should happen rarely. | |
d185af30 YB |
2821 | * |
2822 | * Return: 0 | |
e22bee78 | 2823 | */ |
111c225a | 2824 | static int rescuer_thread(void *__rescuer) |
e22bee78 | 2825 | { |
111c225a TH |
2826 | struct worker *rescuer = __rescuer; |
2827 | struct workqueue_struct *wq = rescuer->rescue_wq; | |
4d595b86 | 2828 | bool should_stop; |
e22bee78 TH |
2829 | |
2830 | set_user_nice(current, RESCUER_NICE_LEVEL); | |
111c225a TH |
2831 | |
2832 | /* | |
2833 | * Mark rescuer as worker too. As WORKER_PREP is never cleared, it | |
2834 | * doesn't participate in concurrency management. | |
2835 | */ | |
197f6acc | 2836 | set_pf_worker(true); |
e22bee78 | 2837 | repeat: |
c5a94a61 | 2838 | set_current_state(TASK_IDLE); |
e22bee78 | 2839 | |
4d595b86 LJ |
2840 | /* |
2841 | * By the time the rescuer is requested to stop, the workqueue | |
2842 | * shouldn't have any work pending, but @wq->maydays may still have | |
2843 | * pwq(s) queued. This can happen by non-rescuer workers consuming | |
2844 | * all the work items before the rescuer got to them. Go through | |
2845 | * @wq->maydays processing before acting on should_stop so that the | |
2846 | * list is always empty on exit. | |
2847 | */ | |
2848 | should_stop = kthread_should_stop(); | |
e22bee78 | 2849 | |
493a1724 | 2850 | /* see whether any pwq is asking for help */ |
a9b8a985 | 2851 | raw_spin_lock_irq(&wq_mayday_lock); |
493a1724 TH |
2852 | |
2853 | while (!list_empty(&wq->maydays)) { | |
2854 | struct pool_workqueue *pwq = list_first_entry(&wq->maydays, | |
2855 | struct pool_workqueue, mayday_node); | |
112202d9 | 2856 | struct worker_pool *pool = pwq->pool; |
e22bee78 TH |
2857 | struct work_struct *work, *n; |
2858 | ||
2859 | __set_current_state(TASK_RUNNING); | |
493a1724 TH |
2860 | list_del_init(&pwq->mayday_node); |
2861 | ||
a9b8a985 | 2862 | raw_spin_unlock_irq(&wq_mayday_lock); |
e22bee78 | 2863 | |
51697d39 LJ |
2864 | worker_attach_to_pool(rescuer, pool); |
2865 | ||
a9b8a985 | 2866 | raw_spin_lock_irq(&pool->lock); |
e22bee78 TH |
2867 | |
2868 | /* | |
2869 | * Slurp in all works issued via this workqueue and | |
2870 | * process'em. | |
2871 | */ | |
873eaca6 | 2872 | WARN_ON_ONCE(!list_empty(&rescuer->scheduled)); |
82607adc | 2873 | list_for_each_entry_safe(work, n, &pool->worklist, entry) { |
873eaca6 TH |
2874 | if (get_work_pwq(work) == pwq && |
2875 | assign_work(work, rescuer, &n)) | |
725e8ec5 | 2876 | pwq->stats[PWQ_STAT_RESCUED]++; |
82607adc | 2877 | } |
e22bee78 | 2878 | |
873eaca6 | 2879 | if (!list_empty(&rescuer->scheduled)) { |
008847f6 N |
2880 | process_scheduled_works(rescuer); |
2881 | ||
2882 | /* | |
2883 | * The above execution of rescued work items could | |
2884 | * have created more to rescue through | |
f97a4a1a | 2885 | * pwq_activate_first_inactive() or chained |
008847f6 N |
2886 | * queueing. Let's put @pwq back on mayday list so |
2887 | * that such back-to-back work items, which may be | |
2888 | * being used to relieve memory pressure, don't | |
2889 | * incur MAYDAY_INTERVAL delay inbetween. | |
2890 | */ | |
4f3f4cf3 | 2891 | if (pwq->nr_active && need_to_create_worker(pool)) { |
a9b8a985 | 2892 | raw_spin_lock(&wq_mayday_lock); |
e66b39af TH |
2893 | /* |
2894 | * Queue iff we aren't racing destruction | |
2895 | * and somebody else hasn't queued it already. | |
2896 | */ | |
2897 | if (wq->rescuer && list_empty(&pwq->mayday_node)) { | |
2898 | get_pwq(pwq); | |
2899 | list_add_tail(&pwq->mayday_node, &wq->maydays); | |
2900 | } | |
a9b8a985 | 2901 | raw_spin_unlock(&wq_mayday_lock); |
008847f6 N |
2902 | } |
2903 | } | |
7576958a | 2904 | |
77668c8b LJ |
2905 | /* |
2906 | * Put the reference grabbed by send_mayday(). @pool won't | |
13b1d625 | 2907 | * go away while we're still attached to it. |
77668c8b LJ |
2908 | */ |
2909 | put_pwq(pwq); | |
2910 | ||
7576958a | 2911 | /* |
0219a352 TH |
2912 | * Leave this pool. Notify regular workers; otherwise, we end up |
2913 | * with 0 concurrency and stalling the execution. | |
7576958a | 2914 | */ |
0219a352 | 2915 | kick_pool(pool); |
7576958a | 2916 | |
a9b8a985 | 2917 | raw_spin_unlock_irq(&pool->lock); |
13b1d625 | 2918 | |
a2d812a2 | 2919 | worker_detach_from_pool(rescuer); |
13b1d625 | 2920 | |
a9b8a985 | 2921 | raw_spin_lock_irq(&wq_mayday_lock); |
e22bee78 TH |
2922 | } |
2923 | ||
a9b8a985 | 2924 | raw_spin_unlock_irq(&wq_mayday_lock); |
493a1724 | 2925 | |
4d595b86 LJ |
2926 | if (should_stop) { |
2927 | __set_current_state(TASK_RUNNING); | |
197f6acc | 2928 | set_pf_worker(false); |
4d595b86 LJ |
2929 | return 0; |
2930 | } | |
2931 | ||
111c225a TH |
2932 | /* rescuers should never participate in concurrency management */ |
2933 | WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); | |
e22bee78 TH |
2934 | schedule(); |
2935 | goto repeat; | |
1da177e4 LT |
2936 | } |
2937 | ||
fca839c0 TH |
2938 | /** |
2939 | * check_flush_dependency - check for flush dependency sanity | |
2940 | * @target_wq: workqueue being flushed | |
2941 | * @target_work: work item being flushed (NULL for workqueue flushes) | |
2942 | * | |
2943 | * %current is trying to flush the whole @target_wq or @target_work on it. | |
2944 | * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not | |
2945 | * reclaiming memory or running on a workqueue which doesn't have | |
2946 | * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to | |
2947 | * a deadlock. | |
2948 | */ | |
2949 | static void check_flush_dependency(struct workqueue_struct *target_wq, | |
2950 | struct work_struct *target_work) | |
2951 | { | |
2952 | work_func_t target_func = target_work ? target_work->func : NULL; | |
2953 | struct worker *worker; | |
2954 | ||
2955 | if (target_wq->flags & WQ_MEM_RECLAIM) | |
2956 | return; | |
2957 | ||
2958 | worker = current_wq_worker(); | |
2959 | ||
2960 | WARN_ONCE(current->flags & PF_MEMALLOC, | |
d75f773c | 2961 | "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps", |
fca839c0 | 2962 | current->pid, current->comm, target_wq->name, target_func); |
23d11a58 TH |
2963 | WARN_ONCE(worker && ((worker->current_pwq->wq->flags & |
2964 | (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM), | |
d75f773c | 2965 | "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps", |
fca839c0 TH |
2966 | worker->current_pwq->wq->name, worker->current_func, |
2967 | target_wq->name, target_func); | |
2968 | } | |
2969 | ||
fc2e4d70 ON |
2970 | struct wq_barrier { |
2971 | struct work_struct work; | |
2972 | struct completion done; | |
2607d7a6 | 2973 | struct task_struct *task; /* purely informational */ |
fc2e4d70 ON |
2974 | }; |
2975 | ||
2976 | static void wq_barrier_func(struct work_struct *work) | |
2977 | { | |
2978 | struct wq_barrier *barr = container_of(work, struct wq_barrier, work); | |
2979 | complete(&barr->done); | |
2980 | } | |
2981 | ||
4690c4ab TH |
2982 | /** |
2983 | * insert_wq_barrier - insert a barrier work | |
112202d9 | 2984 | * @pwq: pwq to insert barrier into |
4690c4ab | 2985 | * @barr: wq_barrier to insert |
affee4b2 TH |
2986 | * @target: target work to attach @barr to |
2987 | * @worker: worker currently executing @target, NULL if @target is not executing | |
4690c4ab | 2988 | * |
affee4b2 TH |
2989 | * @barr is linked to @target such that @barr is completed only after |
2990 | * @target finishes execution. Please note that the ordering | |
2991 | * guarantee is observed only with respect to @target and on the local | |
2992 | * cpu. | |
2993 | * | |
2994 | * Currently, a queued barrier can't be canceled. This is because | |
2995 | * try_to_grab_pending() can't determine whether the work to be | |
2996 | * grabbed is at the head of the queue and thus can't clear LINKED | |
2997 | * flag of the previous work while there must be a valid next work | |
2998 | * after a work with LINKED flag set. | |
2999 | * | |
3000 | * Note that when @worker is non-NULL, @target may be modified | |
112202d9 | 3001 | * underneath us, so we can't reliably determine pwq from @target. |
4690c4ab TH |
3002 | * |
3003 | * CONTEXT: | |
a9b8a985 | 3004 | * raw_spin_lock_irq(pool->lock). |
4690c4ab | 3005 | */ |
112202d9 | 3006 | static void insert_wq_barrier(struct pool_workqueue *pwq, |
affee4b2 TH |
3007 | struct wq_barrier *barr, |
3008 | struct work_struct *target, struct worker *worker) | |
fc2e4d70 | 3009 | { |
d812796e LJ |
3010 | unsigned int work_flags = 0; |
3011 | unsigned int work_color; | |
affee4b2 | 3012 | struct list_head *head; |
affee4b2 | 3013 | |
dc186ad7 | 3014 | /* |
d565ed63 | 3015 | * debugobject calls are safe here even with pool->lock locked |
dc186ad7 TG |
3016 | * as we know for sure that this will not trigger any of the |
3017 | * checks and call back into the fixup functions where we | |
3018 | * might deadlock. | |
3019 | */ | |
ca1cab37 | 3020 | INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); |
22df02bb | 3021 | __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); |
52fa5bc5 | 3022 | |
fd1a5b04 BP |
3023 | init_completion_map(&barr->done, &target->lockdep_map); |
3024 | ||
2607d7a6 | 3025 | barr->task = current; |
83c22520 | 3026 | |
018f3a13 LJ |
3027 | /* The barrier work item does not participate in pwq->nr_active. */ |
3028 | work_flags |= WORK_STRUCT_INACTIVE; | |
3029 | ||
affee4b2 TH |
3030 | /* |
3031 | * If @target is currently being executed, schedule the | |
3032 | * barrier to the worker; otherwise, put it after @target. | |
3033 | */ | |
d812796e | 3034 | if (worker) { |
affee4b2 | 3035 | head = worker->scheduled.next; |
d812796e LJ |
3036 | work_color = worker->current_color; |
3037 | } else { | |
affee4b2 TH |
3038 | unsigned long *bits = work_data_bits(target); |
3039 | ||
3040 | head = target->entry.next; | |
3041 | /* there can already be other linked works, inherit and set */ | |
d21cece0 | 3042 | work_flags |= *bits & WORK_STRUCT_LINKED; |
d812796e | 3043 | work_color = get_work_color(*bits); |
affee4b2 TH |
3044 | __set_bit(WORK_STRUCT_LINKED_BIT, bits); |
3045 | } | |
3046 | ||
d812796e LJ |
3047 | pwq->nr_in_flight[work_color]++; |
3048 | work_flags |= work_color_to_flags(work_color); | |
3049 | ||
d21cece0 | 3050 | insert_work(pwq, &barr->work, head, work_flags); |
fc2e4d70 ON |
3051 | } |
3052 | ||
73f53c4a | 3053 | /** |
112202d9 | 3054 | * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing |
73f53c4a TH |
3055 | * @wq: workqueue being flushed |
3056 | * @flush_color: new flush color, < 0 for no-op | |
3057 | * @work_color: new work color, < 0 for no-op | |
3058 | * | |
112202d9 | 3059 | * Prepare pwqs for workqueue flushing. |
73f53c4a | 3060 | * |
112202d9 TH |
3061 | * If @flush_color is non-negative, flush_color on all pwqs should be |
3062 | * -1. If no pwq has in-flight commands at the specified color, all | |
3063 | * pwq->flush_color's stay at -1 and %false is returned. If any pwq | |
3064 | * has in flight commands, its pwq->flush_color is set to | |
3065 | * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq | |
73f53c4a TH |
3066 | * wakeup logic is armed and %true is returned. |
3067 | * | |
3068 | * The caller should have initialized @wq->first_flusher prior to | |
3069 | * calling this function with non-negative @flush_color. If | |
3070 | * @flush_color is negative, no flush color update is done and %false | |
3071 | * is returned. | |
3072 | * | |
112202d9 | 3073 | * If @work_color is non-negative, all pwqs should have the same |
73f53c4a TH |
3074 | * work_color which is previous to @work_color and all will be |
3075 | * advanced to @work_color. | |
3076 | * | |
3077 | * CONTEXT: | |
3c25a55d | 3078 | * mutex_lock(wq->mutex). |
73f53c4a | 3079 | * |
d185af30 | 3080 | * Return: |
73f53c4a TH |
3081 | * %true if @flush_color >= 0 and there's something to flush. %false |
3082 | * otherwise. | |
3083 | */ | |
112202d9 | 3084 | static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, |
73f53c4a | 3085 | int flush_color, int work_color) |
1da177e4 | 3086 | { |
73f53c4a | 3087 | bool wait = false; |
49e3cf44 | 3088 | struct pool_workqueue *pwq; |
1da177e4 | 3089 | |
73f53c4a | 3090 | if (flush_color >= 0) { |
6183c009 | 3091 | WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); |
112202d9 | 3092 | atomic_set(&wq->nr_pwqs_to_flush, 1); |
1da177e4 | 3093 | } |
2355b70f | 3094 | |
49e3cf44 | 3095 | for_each_pwq(pwq, wq) { |
112202d9 | 3096 | struct worker_pool *pool = pwq->pool; |
fc2e4d70 | 3097 | |
a9b8a985 | 3098 | raw_spin_lock_irq(&pool->lock); |
83c22520 | 3099 | |
73f53c4a | 3100 | if (flush_color >= 0) { |
6183c009 | 3101 | WARN_ON_ONCE(pwq->flush_color != -1); |
fc2e4d70 | 3102 | |
112202d9 TH |
3103 | if (pwq->nr_in_flight[flush_color]) { |
3104 | pwq->flush_color = flush_color; | |
3105 | atomic_inc(&wq->nr_pwqs_to_flush); | |
73f53c4a TH |
3106 | wait = true; |
3107 | } | |
3108 | } | |
1da177e4 | 3109 | |
73f53c4a | 3110 | if (work_color >= 0) { |
6183c009 | 3111 | WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); |
112202d9 | 3112 | pwq->work_color = work_color; |
73f53c4a | 3113 | } |
1da177e4 | 3114 | |
a9b8a985 | 3115 | raw_spin_unlock_irq(&pool->lock); |
1da177e4 | 3116 | } |
2355b70f | 3117 | |
112202d9 | 3118 | if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) |
73f53c4a | 3119 | complete(&wq->first_flusher->done); |
14441960 | 3120 | |
73f53c4a | 3121 | return wait; |
1da177e4 LT |
3122 | } |
3123 | ||
0fcb78c2 | 3124 | /** |
c4f135d6 | 3125 | * __flush_workqueue - ensure that any scheduled work has run to completion. |
0fcb78c2 | 3126 | * @wq: workqueue to flush |
1da177e4 | 3127 | * |
c5aa87bb TH |
3128 | * This function sleeps until all work items which were queued on entry |
3129 | * have finished execution, but it is not livelocked by new incoming ones. | |
1da177e4 | 3130 | */ |
c4f135d6 | 3131 | void __flush_workqueue(struct workqueue_struct *wq) |
1da177e4 | 3132 | { |
73f53c4a TH |
3133 | struct wq_flusher this_flusher = { |
3134 | .list = LIST_HEAD_INIT(this_flusher.list), | |
3135 | .flush_color = -1, | |
fd1a5b04 | 3136 | .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map), |
73f53c4a TH |
3137 | }; |
3138 | int next_color; | |
1da177e4 | 3139 | |
3347fa09 TH |
3140 | if (WARN_ON(!wq_online)) |
3141 | return; | |
3142 | ||
87915adc JB |
3143 | lock_map_acquire(&wq->lockdep_map); |
3144 | lock_map_release(&wq->lockdep_map); | |
3145 | ||
3c25a55d | 3146 | mutex_lock(&wq->mutex); |
73f53c4a TH |
3147 | |
3148 | /* | |
3149 | * Start-to-wait phase | |
3150 | */ | |
3151 | next_color = work_next_color(wq->work_color); | |
3152 | ||
3153 | if (next_color != wq->flush_color) { | |
3154 | /* | |
3155 | * Color space is not full. The current work_color | |
3156 | * becomes our flush_color and work_color is advanced | |
3157 | * by one. | |
3158 | */ | |
6183c009 | 3159 | WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); |
73f53c4a TH |
3160 | this_flusher.flush_color = wq->work_color; |
3161 | wq->work_color = next_color; | |
3162 | ||
3163 | if (!wq->first_flusher) { | |
3164 | /* no flush in progress, become the first flusher */ | |
6183c009 | 3165 | WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); |
73f53c4a TH |
3166 | |
3167 | wq->first_flusher = &this_flusher; | |
3168 | ||
112202d9 | 3169 | if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, |
73f53c4a TH |
3170 | wq->work_color)) { |
3171 | /* nothing to flush, done */ | |
3172 | wq->flush_color = next_color; | |
3173 | wq->first_flusher = NULL; | |
3174 | goto out_unlock; | |
3175 | } | |
3176 | } else { | |
3177 | /* wait in queue */ | |
6183c009 | 3178 | WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); |
73f53c4a | 3179 | list_add_tail(&this_flusher.list, &wq->flusher_queue); |
112202d9 | 3180 | flush_workqueue_prep_pwqs(wq, -1, wq->work_color); |
73f53c4a TH |
3181 | } |
3182 | } else { | |
3183 | /* | |
3184 | * Oops, color space is full, wait on overflow queue. | |
3185 | * The next flush completion will assign us | |
3186 | * flush_color and transfer to flusher_queue. | |
3187 | */ | |
3188 | list_add_tail(&this_flusher.list, &wq->flusher_overflow); | |
3189 | } | |
3190 | ||
fca839c0 TH |
3191 | check_flush_dependency(wq, NULL); |
3192 | ||
3c25a55d | 3193 | mutex_unlock(&wq->mutex); |
73f53c4a TH |
3194 | |
3195 | wait_for_completion(&this_flusher.done); | |
3196 | ||
3197 | /* | |
3198 | * Wake-up-and-cascade phase | |
3199 | * | |
3200 | * First flushers are responsible for cascading flushes and | |
3201 | * handling overflow. Non-first flushers can simply return. | |
3202 | */ | |
00d5d15b | 3203 | if (READ_ONCE(wq->first_flusher) != &this_flusher) |
73f53c4a TH |
3204 | return; |
3205 | ||
3c25a55d | 3206 | mutex_lock(&wq->mutex); |
73f53c4a | 3207 | |
4ce48b37 TH |
3208 | /* we might have raced, check again with mutex held */ |
3209 | if (wq->first_flusher != &this_flusher) | |
3210 | goto out_unlock; | |
3211 | ||
00d5d15b | 3212 | WRITE_ONCE(wq->first_flusher, NULL); |
73f53c4a | 3213 | |
6183c009 TH |
3214 | WARN_ON_ONCE(!list_empty(&this_flusher.list)); |
3215 | WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); | |
73f53c4a TH |
3216 | |
3217 | while (true) { | |
3218 | struct wq_flusher *next, *tmp; | |
3219 | ||
3220 | /* complete all the flushers sharing the current flush color */ | |
3221 | list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { | |
3222 | if (next->flush_color != wq->flush_color) | |
3223 | break; | |
3224 | list_del_init(&next->list); | |
3225 | complete(&next->done); | |
3226 | } | |
3227 | ||
6183c009 TH |
3228 | WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && |
3229 | wq->flush_color != work_next_color(wq->work_color)); | |
73f53c4a TH |
3230 | |
3231 | /* this flush_color is finished, advance by one */ | |
3232 | wq->flush_color = work_next_color(wq->flush_color); | |
3233 | ||
3234 | /* one color has been freed, handle overflow queue */ | |
3235 | if (!list_empty(&wq->flusher_overflow)) { | |
3236 | /* | |
3237 | * Assign the same color to all overflowed | |
3238 | * flushers, advance work_color and append to | |
3239 | * flusher_queue. This is the start-to-wait | |
3240 | * phase for these overflowed flushers. | |
3241 | */ | |
3242 | list_for_each_entry(tmp, &wq->flusher_overflow, list) | |
3243 | tmp->flush_color = wq->work_color; | |
3244 | ||
3245 | wq->work_color = work_next_color(wq->work_color); | |
3246 | ||
3247 | list_splice_tail_init(&wq->flusher_overflow, | |
3248 | &wq->flusher_queue); | |
112202d9 | 3249 | flush_workqueue_prep_pwqs(wq, -1, wq->work_color); |
73f53c4a TH |
3250 | } |
3251 | ||
3252 | if (list_empty(&wq->flusher_queue)) { | |
6183c009 | 3253 | WARN_ON_ONCE(wq->flush_color != wq->work_color); |
73f53c4a TH |
3254 | break; |
3255 | } | |
3256 | ||
3257 | /* | |
3258 | * Need to flush more colors. Make the next flusher | |
112202d9 | 3259 | * the new first flusher and arm pwqs. |
73f53c4a | 3260 | */ |
6183c009 TH |
3261 | WARN_ON_ONCE(wq->flush_color == wq->work_color); |
3262 | WARN_ON_ONCE(wq->flush_color != next->flush_color); | |
73f53c4a TH |
3263 | |
3264 | list_del_init(&next->list); | |
3265 | wq->first_flusher = next; | |
3266 | ||
112202d9 | 3267 | if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) |
73f53c4a TH |
3268 | break; |
3269 | ||
3270 | /* | |
3271 | * Meh... this color is already done, clear first | |
3272 | * flusher and repeat cascading. | |
3273 | */ | |
3274 | wq->first_flusher = NULL; | |
3275 | } | |
3276 | ||
3277 | out_unlock: | |
3c25a55d | 3278 | mutex_unlock(&wq->mutex); |
1da177e4 | 3279 | } |
c4f135d6 | 3280 | EXPORT_SYMBOL(__flush_workqueue); |
1da177e4 | 3281 | |
9c5a2ba7 TH |
3282 | /** |
3283 | * drain_workqueue - drain a workqueue | |
3284 | * @wq: workqueue to drain | |
3285 | * | |
3286 | * Wait until the workqueue becomes empty. While draining is in progress, | |
3287 | * only chain queueing is allowed. IOW, only currently pending or running | |
3288 | * work items on @wq can queue further work items on it. @wq is flushed | |
b749b1b6 | 3289 | * repeatedly until it becomes empty. The number of flushing is determined |
9c5a2ba7 TH |
3290 | * by the depth of chaining and should be relatively short. Whine if it |
3291 | * takes too long. | |
3292 | */ | |
3293 | void drain_workqueue(struct workqueue_struct *wq) | |
3294 | { | |
3295 | unsigned int flush_cnt = 0; | |
49e3cf44 | 3296 | struct pool_workqueue *pwq; |
9c5a2ba7 TH |
3297 | |
3298 | /* | |
3299 | * __queue_work() needs to test whether there are drainers, is much | |
3300 | * hotter than drain_workqueue() and already looks at @wq->flags. | |
618b01eb | 3301 | * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. |
9c5a2ba7 | 3302 | */ |
87fc741e | 3303 | mutex_lock(&wq->mutex); |
9c5a2ba7 | 3304 | if (!wq->nr_drainers++) |
618b01eb | 3305 | wq->flags |= __WQ_DRAINING; |
87fc741e | 3306 | mutex_unlock(&wq->mutex); |
9c5a2ba7 | 3307 | reflush: |
c4f135d6 | 3308 | __flush_workqueue(wq); |
9c5a2ba7 | 3309 | |
b09f4fd3 | 3310 | mutex_lock(&wq->mutex); |
76af4d93 | 3311 | |
49e3cf44 | 3312 | for_each_pwq(pwq, wq) { |
fa2563e4 | 3313 | bool drained; |
9c5a2ba7 | 3314 | |
a9b8a985 | 3315 | raw_spin_lock_irq(&pwq->pool->lock); |
f97a4a1a | 3316 | drained = !pwq->nr_active && list_empty(&pwq->inactive_works); |
a9b8a985 | 3317 | raw_spin_unlock_irq(&pwq->pool->lock); |
fa2563e4 TT |
3318 | |
3319 | if (drained) | |
9c5a2ba7 TH |
3320 | continue; |
3321 | ||
3322 | if (++flush_cnt == 10 || | |
3323 | (flush_cnt % 100 == 0 && flush_cnt <= 1000)) | |
e9ad2eb3 SZ |
3324 | pr_warn("workqueue %s: %s() isn't complete after %u tries\n", |
3325 | wq->name, __func__, flush_cnt); | |
76af4d93 | 3326 | |
b09f4fd3 | 3327 | mutex_unlock(&wq->mutex); |
9c5a2ba7 TH |
3328 | goto reflush; |
3329 | } | |
3330 | ||
9c5a2ba7 | 3331 | if (!--wq->nr_drainers) |
618b01eb | 3332 | wq->flags &= ~__WQ_DRAINING; |
87fc741e | 3333 | mutex_unlock(&wq->mutex); |
9c5a2ba7 TH |
3334 | } |
3335 | EXPORT_SYMBOL_GPL(drain_workqueue); | |
3336 | ||
d6e89786 JB |
3337 | static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr, |
3338 | bool from_cancel) | |
db700897 | 3339 | { |
affee4b2 | 3340 | struct worker *worker = NULL; |
c9e7cf27 | 3341 | struct worker_pool *pool; |
112202d9 | 3342 | struct pool_workqueue *pwq; |
db700897 ON |
3343 | |
3344 | might_sleep(); | |
fa1b54e6 | 3345 | |
24acfb71 | 3346 | rcu_read_lock(); |
c9e7cf27 | 3347 | pool = get_work_pool(work); |
fa1b54e6 | 3348 | if (!pool) { |
24acfb71 | 3349 | rcu_read_unlock(); |
baf59022 | 3350 | return false; |
fa1b54e6 | 3351 | } |
db700897 | 3352 | |
a9b8a985 | 3353 | raw_spin_lock_irq(&pool->lock); |
0b3dae68 | 3354 | /* see the comment in try_to_grab_pending() with the same code */ |
112202d9 TH |
3355 | pwq = get_work_pwq(work); |
3356 | if (pwq) { | |
3357 | if (unlikely(pwq->pool != pool)) | |
4690c4ab | 3358 | goto already_gone; |
606a5020 | 3359 | } else { |
c9e7cf27 | 3360 | worker = find_worker_executing_work(pool, work); |
affee4b2 | 3361 | if (!worker) |
4690c4ab | 3362 | goto already_gone; |
112202d9 | 3363 | pwq = worker->current_pwq; |
606a5020 | 3364 | } |
db700897 | 3365 | |
fca839c0 TH |
3366 | check_flush_dependency(pwq->wq, work); |
3367 | ||
112202d9 | 3368 | insert_wq_barrier(pwq, barr, work, worker); |
a9b8a985 | 3369 | raw_spin_unlock_irq(&pool->lock); |
7a22ad75 | 3370 | |
e159489b | 3371 | /* |
a1d14934 PZ |
3372 | * Force a lock recursion deadlock when using flush_work() inside a |
3373 | * single-threaded or rescuer equipped workqueue. | |
3374 | * | |
3375 | * For single threaded workqueues the deadlock happens when the work | |
3376 | * is after the work issuing the flush_work(). For rescuer equipped | |
3377 | * workqueues the deadlock happens when the rescuer stalls, blocking | |
3378 | * forward progress. | |
e159489b | 3379 | */ |
d6e89786 JB |
3380 | if (!from_cancel && |
3381 | (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) { | |
112202d9 | 3382 | lock_map_acquire(&pwq->wq->lockdep_map); |
a1d14934 PZ |
3383 | lock_map_release(&pwq->wq->lockdep_map); |
3384 | } | |
24acfb71 | 3385 | rcu_read_unlock(); |
401a8d04 | 3386 | return true; |
4690c4ab | 3387 | already_gone: |
a9b8a985 | 3388 | raw_spin_unlock_irq(&pool->lock); |
24acfb71 | 3389 | rcu_read_unlock(); |
401a8d04 | 3390 | return false; |
db700897 | 3391 | } |
baf59022 | 3392 | |
d6e89786 JB |
3393 | static bool __flush_work(struct work_struct *work, bool from_cancel) |
3394 | { | |
3395 | struct wq_barrier barr; | |
3396 | ||
3397 | if (WARN_ON(!wq_online)) | |
3398 | return false; | |
3399 | ||
4d43d395 TH |
3400 | if (WARN_ON(!work->func)) |
3401 | return false; | |
3402 | ||
c0feea59 TH |
3403 | lock_map_acquire(&work->lockdep_map); |
3404 | lock_map_release(&work->lockdep_map); | |
87915adc | 3405 | |
d6e89786 JB |
3406 | if (start_flush_work(work, &barr, from_cancel)) { |
3407 | wait_for_completion(&barr.done); | |
3408 | destroy_work_on_stack(&barr.work); | |
3409 | return true; | |
3410 | } else { | |
3411 | return false; | |
3412 | } | |
3413 | } | |
3414 | ||
baf59022 TH |
3415 | /** |
3416 | * flush_work - wait for a work to finish executing the last queueing instance | |
3417 | * @work: the work to flush | |
3418 | * | |
606a5020 TH |
3419 | * Wait until @work has finished execution. @work is guaranteed to be idle |
3420 | * on return if it hasn't been requeued since flush started. | |
baf59022 | 3421 | * |
d185af30 | 3422 | * Return: |
baf59022 TH |
3423 | * %true if flush_work() waited for the work to finish execution, |
3424 | * %false if it was already idle. | |
3425 | */ | |
3426 | bool flush_work(struct work_struct *work) | |
3427 | { | |
d6e89786 | 3428 | return __flush_work(work, false); |
6e84d644 | 3429 | } |
606a5020 | 3430 | EXPORT_SYMBOL_GPL(flush_work); |
6e84d644 | 3431 | |
8603e1b3 | 3432 | struct cwt_wait { |
ac6424b9 | 3433 | wait_queue_entry_t wait; |
8603e1b3 TH |
3434 | struct work_struct *work; |
3435 | }; | |
3436 | ||
ac6424b9 | 3437 | static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key) |
8603e1b3 TH |
3438 | { |
3439 | struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait); | |
3440 | ||
3441 | if (cwait->work != key) | |
3442 | return 0; | |
3443 | return autoremove_wake_function(wait, mode, sync, key); | |
3444 | } | |
3445 | ||
36e227d2 | 3446 | static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) |
1f1f642e | 3447 | { |
8603e1b3 | 3448 | static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq); |
bbb68dfa | 3449 | unsigned long flags; |
1f1f642e ON |
3450 | int ret; |
3451 | ||
3452 | do { | |
bbb68dfa TH |
3453 | ret = try_to_grab_pending(work, is_dwork, &flags); |
3454 | /* | |
8603e1b3 TH |
3455 | * If someone else is already canceling, wait for it to |
3456 | * finish. flush_work() doesn't work for PREEMPT_NONE | |
3457 | * because we may get scheduled between @work's completion | |
3458 | * and the other canceling task resuming and clearing | |
3459 | * CANCELING - flush_work() will return false immediately | |
3460 | * as @work is no longer busy, try_to_grab_pending() will | |
3461 | * return -ENOENT as @work is still being canceled and the | |
3462 | * other canceling task won't be able to clear CANCELING as | |
3463 | * we're hogging the CPU. | |
3464 | * | |
3465 | * Let's wait for completion using a waitqueue. As this | |
3466 | * may lead to the thundering herd problem, use a custom | |
3467 | * wake function which matches @work along with exclusive | |
3468 | * wait and wakeup. | |
bbb68dfa | 3469 | */ |
8603e1b3 TH |
3470 | if (unlikely(ret == -ENOENT)) { |
3471 | struct cwt_wait cwait; | |
3472 | ||
3473 | init_wait(&cwait.wait); | |
3474 | cwait.wait.func = cwt_wakefn; | |
3475 | cwait.work = work; | |
3476 | ||
3477 | prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait, | |
3478 | TASK_UNINTERRUPTIBLE); | |
3479 | if (work_is_canceling(work)) | |
3480 | schedule(); | |
3481 | finish_wait(&cancel_waitq, &cwait.wait); | |
3482 | } | |
1f1f642e ON |
3483 | } while (unlikely(ret < 0)); |
3484 | ||
bbb68dfa TH |
3485 | /* tell other tasks trying to grab @work to back off */ |
3486 | mark_work_canceling(work); | |
3487 | local_irq_restore(flags); | |
3488 | ||
3347fa09 TH |
3489 | /* |
3490 | * This allows canceling during early boot. We know that @work | |
3491 | * isn't executing. | |
3492 | */ | |
3493 | if (wq_online) | |
d6e89786 | 3494 | __flush_work(work, true); |
3347fa09 | 3495 | |
7a22ad75 | 3496 | clear_work_data(work); |
8603e1b3 TH |
3497 | |
3498 | /* | |
3499 | * Paired with prepare_to_wait() above so that either | |
3500 | * waitqueue_active() is visible here or !work_is_canceling() is | |
3501 | * visible there. | |
3502 | */ | |
3503 | smp_mb(); | |
3504 | if (waitqueue_active(&cancel_waitq)) | |
3505 | __wake_up(&cancel_waitq, TASK_NORMAL, 1, work); | |
3506 | ||
1f1f642e ON |
3507 | return ret; |
3508 | } | |
3509 | ||
6e84d644 | 3510 | /** |
401a8d04 TH |
3511 | * cancel_work_sync - cancel a work and wait for it to finish |
3512 | * @work: the work to cancel | |
6e84d644 | 3513 | * |
401a8d04 TH |
3514 | * Cancel @work and wait for its execution to finish. This function |
3515 | * can be used even if the work re-queues itself or migrates to | |
3516 | * another workqueue. On return from this function, @work is | |
3517 | * guaranteed to be not pending or executing on any CPU. | |
1f1f642e | 3518 | * |
401a8d04 TH |
3519 | * cancel_work_sync(&delayed_work->work) must not be used for |
3520 | * delayed_work's. Use cancel_delayed_work_sync() instead. | |
6e84d644 | 3521 | * |
401a8d04 | 3522 | * The caller must ensure that the workqueue on which @work was last |
6e84d644 | 3523 | * queued can't be destroyed before this function returns. |
401a8d04 | 3524 | * |
d185af30 | 3525 | * Return: |
401a8d04 | 3526 | * %true if @work was pending, %false otherwise. |
6e84d644 | 3527 | */ |
401a8d04 | 3528 | bool cancel_work_sync(struct work_struct *work) |
6e84d644 | 3529 | { |
36e227d2 | 3530 | return __cancel_work_timer(work, false); |
b89deed3 | 3531 | } |
28e53bdd | 3532 | EXPORT_SYMBOL_GPL(cancel_work_sync); |
b89deed3 | 3533 | |
6e84d644 | 3534 | /** |
401a8d04 TH |
3535 | * flush_delayed_work - wait for a dwork to finish executing the last queueing |
3536 | * @dwork: the delayed work to flush | |
6e84d644 | 3537 | * |
401a8d04 TH |
3538 | * Delayed timer is cancelled and the pending work is queued for |
3539 | * immediate execution. Like flush_work(), this function only | |
3540 | * considers the last queueing instance of @dwork. | |
1f1f642e | 3541 | * |
d185af30 | 3542 | * Return: |
401a8d04 TH |
3543 | * %true if flush_work() waited for the work to finish execution, |
3544 | * %false if it was already idle. | |
6e84d644 | 3545 | */ |
401a8d04 TH |
3546 | bool flush_delayed_work(struct delayed_work *dwork) |
3547 | { | |
8930caba | 3548 | local_irq_disable(); |
401a8d04 | 3549 | if (del_timer_sync(&dwork->timer)) |
60c057bc | 3550 | __queue_work(dwork->cpu, dwork->wq, &dwork->work); |
8930caba | 3551 | local_irq_enable(); |
401a8d04 TH |
3552 | return flush_work(&dwork->work); |
3553 | } | |
3554 | EXPORT_SYMBOL(flush_delayed_work); | |
3555 | ||
05f0fe6b TH |
3556 | /** |
3557 | * flush_rcu_work - wait for a rwork to finish executing the last queueing | |
3558 | * @rwork: the rcu work to flush | |
3559 | * | |
3560 | * Return: | |
3561 | * %true if flush_rcu_work() waited for the work to finish execution, | |
3562 | * %false if it was already idle. | |
3563 | */ | |
3564 | bool flush_rcu_work(struct rcu_work *rwork) | |
3565 | { | |
3566 | if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) { | |
3567 | rcu_barrier(); | |
3568 | flush_work(&rwork->work); | |
3569 | return true; | |
3570 | } else { | |
3571 | return flush_work(&rwork->work); | |
3572 | } | |
3573 | } | |
3574 | EXPORT_SYMBOL(flush_rcu_work); | |
3575 | ||
f72b8792 JA |
3576 | static bool __cancel_work(struct work_struct *work, bool is_dwork) |
3577 | { | |
3578 | unsigned long flags; | |
3579 | int ret; | |
3580 | ||
3581 | do { | |
3582 | ret = try_to_grab_pending(work, is_dwork, &flags); | |
3583 | } while (unlikely(ret == -EAGAIN)); | |
3584 | ||
3585 | if (unlikely(ret < 0)) | |
3586 | return false; | |
3587 | ||
3588 | set_work_pool_and_clear_pending(work, get_work_pool_id(work)); | |
3589 | local_irq_restore(flags); | |
3590 | return ret; | |
3591 | } | |
3592 | ||
73b4b532 AG |
3593 | /* |
3594 | * See cancel_delayed_work() | |
3595 | */ | |
3596 | bool cancel_work(struct work_struct *work) | |
3597 | { | |
3598 | return __cancel_work(work, false); | |
3599 | } | |
3600 | EXPORT_SYMBOL(cancel_work); | |
3601 | ||
09383498 | 3602 | /** |
57b30ae7 TH |
3603 | * cancel_delayed_work - cancel a delayed work |
3604 | * @dwork: delayed_work to cancel | |
09383498 | 3605 | * |
d185af30 YB |
3606 | * Kill off a pending delayed_work. |
3607 | * | |
3608 | * Return: %true if @dwork was pending and canceled; %false if it wasn't | |
3609 | * pending. | |
3610 | * | |
3611 | * Note: | |
3612 | * The work callback function may still be running on return, unless | |
3613 | * it returns %true and the work doesn't re-arm itself. Explicitly flush or | |
3614 | * use cancel_delayed_work_sync() to wait on it. | |
09383498 | 3615 | * |
57b30ae7 | 3616 | * This function is safe to call from any context including IRQ handler. |
09383498 | 3617 | */ |
57b30ae7 | 3618 | bool cancel_delayed_work(struct delayed_work *dwork) |
09383498 | 3619 | { |
f72b8792 | 3620 | return __cancel_work(&dwork->work, true); |
09383498 | 3621 | } |
57b30ae7 | 3622 | EXPORT_SYMBOL(cancel_delayed_work); |
09383498 | 3623 | |
401a8d04 TH |
3624 | /** |
3625 | * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish | |
3626 | * @dwork: the delayed work cancel | |
3627 | * | |
3628 | * This is cancel_work_sync() for delayed works. | |
3629 | * | |
d185af30 | 3630 | * Return: |
401a8d04 TH |
3631 | * %true if @dwork was pending, %false otherwise. |
3632 | */ | |
3633 | bool cancel_delayed_work_sync(struct delayed_work *dwork) | |
6e84d644 | 3634 | { |
36e227d2 | 3635 | return __cancel_work_timer(&dwork->work, true); |
6e84d644 | 3636 | } |
f5a421a4 | 3637 | EXPORT_SYMBOL(cancel_delayed_work_sync); |
1da177e4 | 3638 | |
b6136773 | 3639 | /** |
31ddd871 | 3640 | * schedule_on_each_cpu - execute a function synchronously on each online CPU |
b6136773 | 3641 | * @func: the function to call |
b6136773 | 3642 | * |
31ddd871 TH |
3643 | * schedule_on_each_cpu() executes @func on each online CPU using the |
3644 | * system workqueue and blocks until all CPUs have completed. | |
b6136773 | 3645 | * schedule_on_each_cpu() is very slow. |
31ddd871 | 3646 | * |
d185af30 | 3647 | * Return: |
31ddd871 | 3648 | * 0 on success, -errno on failure. |
b6136773 | 3649 | */ |
65f27f38 | 3650 | int schedule_on_each_cpu(work_func_t func) |
15316ba8 CL |
3651 | { |
3652 | int cpu; | |
38f51568 | 3653 | struct work_struct __percpu *works; |
15316ba8 | 3654 | |
b6136773 AM |
3655 | works = alloc_percpu(struct work_struct); |
3656 | if (!works) | |
15316ba8 | 3657 | return -ENOMEM; |
b6136773 | 3658 | |
ffd8bea8 | 3659 | cpus_read_lock(); |
93981800 | 3660 | |
15316ba8 | 3661 | for_each_online_cpu(cpu) { |
9bfb1839 IM |
3662 | struct work_struct *work = per_cpu_ptr(works, cpu); |
3663 | ||
3664 | INIT_WORK(work, func); | |
b71ab8c2 | 3665 | schedule_work_on(cpu, work); |
65a64464 | 3666 | } |
93981800 TH |
3667 | |
3668 | for_each_online_cpu(cpu) | |
3669 | flush_work(per_cpu_ptr(works, cpu)); | |
3670 | ||
ffd8bea8 | 3671 | cpus_read_unlock(); |
b6136773 | 3672 | free_percpu(works); |
15316ba8 CL |
3673 | return 0; |
3674 | } | |
3675 | ||
1fa44eca JB |
3676 | /** |
3677 | * execute_in_process_context - reliably execute the routine with user context | |
3678 | * @fn: the function to execute | |
1fa44eca JB |
3679 | * @ew: guaranteed storage for the execute work structure (must |
3680 | * be available when the work executes) | |
3681 | * | |
3682 | * Executes the function immediately if process context is available, | |
3683 | * otherwise schedules the function for delayed execution. | |
3684 | * | |
d185af30 | 3685 | * Return: 0 - function was executed |
1fa44eca JB |
3686 | * 1 - function was scheduled for execution |
3687 | */ | |
65f27f38 | 3688 | int execute_in_process_context(work_func_t fn, struct execute_work *ew) |
1fa44eca JB |
3689 | { |
3690 | if (!in_interrupt()) { | |
65f27f38 | 3691 | fn(&ew->work); |
1fa44eca JB |
3692 | return 0; |
3693 | } | |
3694 | ||
65f27f38 | 3695 | INIT_WORK(&ew->work, fn); |
1fa44eca JB |
3696 | schedule_work(&ew->work); |
3697 | ||
3698 | return 1; | |
3699 | } | |
3700 | EXPORT_SYMBOL_GPL(execute_in_process_context); | |
3701 | ||
6ba94429 FW |
3702 | /** |
3703 | * free_workqueue_attrs - free a workqueue_attrs | |
3704 | * @attrs: workqueue_attrs to free | |
226223ab | 3705 | * |
6ba94429 | 3706 | * Undo alloc_workqueue_attrs(). |
226223ab | 3707 | */ |
513c98d0 | 3708 | void free_workqueue_attrs(struct workqueue_attrs *attrs) |
226223ab | 3709 | { |
6ba94429 FW |
3710 | if (attrs) { |
3711 | free_cpumask_var(attrs->cpumask); | |
9546b29e | 3712 | free_cpumask_var(attrs->__pod_cpumask); |
6ba94429 FW |
3713 | kfree(attrs); |
3714 | } | |
226223ab TH |
3715 | } |
3716 | ||
6ba94429 FW |
3717 | /** |
3718 | * alloc_workqueue_attrs - allocate a workqueue_attrs | |
6ba94429 FW |
3719 | * |
3720 | * Allocate a new workqueue_attrs, initialize with default settings and | |
3721 | * return it. | |
3722 | * | |
3723 | * Return: The allocated new workqueue_attr on success. %NULL on failure. | |
3724 | */ | |
513c98d0 | 3725 | struct workqueue_attrs *alloc_workqueue_attrs(void) |
226223ab | 3726 | { |
6ba94429 | 3727 | struct workqueue_attrs *attrs; |
226223ab | 3728 | |
be69d00d | 3729 | attrs = kzalloc(sizeof(*attrs), GFP_KERNEL); |
6ba94429 FW |
3730 | if (!attrs) |
3731 | goto fail; | |
be69d00d | 3732 | if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL)) |
6ba94429 | 3733 | goto fail; |
9546b29e TH |
3734 | if (!alloc_cpumask_var(&attrs->__pod_cpumask, GFP_KERNEL)) |
3735 | goto fail; | |
6ba94429 FW |
3736 | |
3737 | cpumask_copy(attrs->cpumask, cpu_possible_mask); | |
523a301e | 3738 | attrs->affn_scope = WQ_AFFN_DFL; |
6ba94429 FW |
3739 | return attrs; |
3740 | fail: | |
3741 | free_workqueue_attrs(attrs); | |
3742 | return NULL; | |
226223ab TH |
3743 | } |
3744 | ||
6ba94429 FW |
3745 | static void copy_workqueue_attrs(struct workqueue_attrs *to, |
3746 | const struct workqueue_attrs *from) | |
226223ab | 3747 | { |
6ba94429 FW |
3748 | to->nice = from->nice; |
3749 | cpumask_copy(to->cpumask, from->cpumask); | |
9546b29e | 3750 | cpumask_copy(to->__pod_cpumask, from->__pod_cpumask); |
8639eceb | 3751 | to->affn_strict = from->affn_strict; |
84193c07 | 3752 | |
6ba94429 | 3753 | /* |
84193c07 TH |
3754 | * Unlike hash and equality test, copying shouldn't ignore wq-only |
3755 | * fields as copying is used for both pool and wq attrs. Instead, | |
3756 | * get_unbound_pool() explicitly clears the fields. | |
6ba94429 | 3757 | */ |
84193c07 | 3758 | to->affn_scope = from->affn_scope; |
af73f5c9 | 3759 | to->ordered = from->ordered; |
226223ab TH |
3760 | } |
3761 | ||
5de7a03c TH |
3762 | /* |
3763 | * Some attrs fields are workqueue-only. Clear them for worker_pool's. See the | |
3764 | * comments in 'struct workqueue_attrs' definition. | |
3765 | */ | |
3766 | static void wqattrs_clear_for_pool(struct workqueue_attrs *attrs) | |
3767 | { | |
84193c07 | 3768 | attrs->affn_scope = WQ_AFFN_NR_TYPES; |
5de7a03c TH |
3769 | attrs->ordered = false; |
3770 | } | |
3771 | ||
6ba94429 FW |
3772 | /* hash value of the content of @attr */ |
3773 | static u32 wqattrs_hash(const struct workqueue_attrs *attrs) | |
226223ab | 3774 | { |
6ba94429 | 3775 | u32 hash = 0; |
226223ab | 3776 | |
6ba94429 FW |
3777 | hash = jhash_1word(attrs->nice, hash); |
3778 | hash = jhash(cpumask_bits(attrs->cpumask), | |
3779 | BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); | |
9546b29e TH |
3780 | hash = jhash(cpumask_bits(attrs->__pod_cpumask), |
3781 | BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); | |
8639eceb | 3782 | hash = jhash_1word(attrs->affn_strict, hash); |
6ba94429 | 3783 | return hash; |
226223ab | 3784 | } |
226223ab | 3785 | |
6ba94429 FW |
3786 | /* content equality test */ |
3787 | static bool wqattrs_equal(const struct workqueue_attrs *a, | |
3788 | const struct workqueue_attrs *b) | |
226223ab | 3789 | { |
6ba94429 FW |
3790 | if (a->nice != b->nice) |
3791 | return false; | |
3792 | if (!cpumask_equal(a->cpumask, b->cpumask)) | |
3793 | return false; | |
9546b29e TH |
3794 | if (!cpumask_equal(a->__pod_cpumask, b->__pod_cpumask)) |
3795 | return false; | |
8639eceb TH |
3796 | if (a->affn_strict != b->affn_strict) |
3797 | return false; | |
6ba94429 | 3798 | return true; |
226223ab TH |
3799 | } |
3800 | ||
0f36ee24 TH |
3801 | /* Update @attrs with actually available CPUs */ |
3802 | static void wqattrs_actualize_cpumask(struct workqueue_attrs *attrs, | |
3803 | const cpumask_t *unbound_cpumask) | |
3804 | { | |
3805 | /* | |
3806 | * Calculate the effective CPU mask of @attrs given @unbound_cpumask. If | |
3807 | * @attrs->cpumask doesn't overlap with @unbound_cpumask, we fallback to | |
3808 | * @unbound_cpumask. | |
3809 | */ | |
3810 | cpumask_and(attrs->cpumask, attrs->cpumask, unbound_cpumask); | |
3811 | if (unlikely(cpumask_empty(attrs->cpumask))) | |
3812 | cpumask_copy(attrs->cpumask, unbound_cpumask); | |
3813 | } | |
3814 | ||
84193c07 TH |
3815 | /* find wq_pod_type to use for @attrs */ |
3816 | static const struct wq_pod_type * | |
3817 | wqattrs_pod_type(const struct workqueue_attrs *attrs) | |
3818 | { | |
523a301e TH |
3819 | enum wq_affn_scope scope; |
3820 | struct wq_pod_type *pt; | |
3821 | ||
3822 | /* to synchronize access to wq_affn_dfl */ | |
3823 | lockdep_assert_held(&wq_pool_mutex); | |
3824 | ||
3825 | if (attrs->affn_scope == WQ_AFFN_DFL) | |
3826 | scope = wq_affn_dfl; | |
3827 | else | |
3828 | scope = attrs->affn_scope; | |
3829 | ||
3830 | pt = &wq_pod_types[scope]; | |
84193c07 TH |
3831 | |
3832 | if (!WARN_ON_ONCE(attrs->affn_scope == WQ_AFFN_NR_TYPES) && | |
3833 | likely(pt->nr_pods)) | |
3834 | return pt; | |
3835 | ||
3836 | /* | |
3837 | * Before workqueue_init_topology(), only SYSTEM is available which is | |
3838 | * initialized in workqueue_init_early(). | |
3839 | */ | |
3840 | pt = &wq_pod_types[WQ_AFFN_SYSTEM]; | |
3841 | BUG_ON(!pt->nr_pods); | |
3842 | return pt; | |
3843 | } | |
3844 | ||
6ba94429 FW |
3845 | /** |
3846 | * init_worker_pool - initialize a newly zalloc'd worker_pool | |
3847 | * @pool: worker_pool to initialize | |
3848 | * | |
402dd89d | 3849 | * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs. |
6ba94429 FW |
3850 | * |
3851 | * Return: 0 on success, -errno on failure. Even on failure, all fields | |
3852 | * inside @pool proper are initialized and put_unbound_pool() can be called | |
3853 | * on @pool safely to release it. | |
3854 | */ | |
3855 | static int init_worker_pool(struct worker_pool *pool) | |
226223ab | 3856 | { |
a9b8a985 | 3857 | raw_spin_lock_init(&pool->lock); |
6ba94429 FW |
3858 | pool->id = -1; |
3859 | pool->cpu = -1; | |
3860 | pool->node = NUMA_NO_NODE; | |
3861 | pool->flags |= POOL_DISASSOCIATED; | |
82607adc | 3862 | pool->watchdog_ts = jiffies; |
6ba94429 FW |
3863 | INIT_LIST_HEAD(&pool->worklist); |
3864 | INIT_LIST_HEAD(&pool->idle_list); | |
3865 | hash_init(pool->busy_hash); | |
226223ab | 3866 | |
32a6c723 | 3867 | timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE); |
3f959aa3 | 3868 | INIT_WORK(&pool->idle_cull_work, idle_cull_fn); |
226223ab | 3869 | |
32a6c723 | 3870 | timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0); |
226223ab | 3871 | |
6ba94429 | 3872 | INIT_LIST_HEAD(&pool->workers); |
e02b9312 | 3873 | INIT_LIST_HEAD(&pool->dying_workers); |
226223ab | 3874 | |
6ba94429 FW |
3875 | ida_init(&pool->worker_ida); |
3876 | INIT_HLIST_NODE(&pool->hash_node); | |
3877 | pool->refcnt = 1; | |
226223ab | 3878 | |
6ba94429 | 3879 | /* shouldn't fail above this point */ |
be69d00d | 3880 | pool->attrs = alloc_workqueue_attrs(); |
6ba94429 FW |
3881 | if (!pool->attrs) |
3882 | return -ENOMEM; | |
5de7a03c TH |
3883 | |
3884 | wqattrs_clear_for_pool(pool->attrs); | |
3885 | ||
6ba94429 | 3886 | return 0; |
226223ab TH |
3887 | } |
3888 | ||
669de8bd BVA |
3889 | #ifdef CONFIG_LOCKDEP |
3890 | static void wq_init_lockdep(struct workqueue_struct *wq) | |
3891 | { | |
3892 | char *lock_name; | |
3893 | ||
3894 | lockdep_register_key(&wq->key); | |
3895 | lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name); | |
3896 | if (!lock_name) | |
3897 | lock_name = wq->name; | |
69a106c0 QC |
3898 | |
3899 | wq->lock_name = lock_name; | |
669de8bd BVA |
3900 | lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0); |
3901 | } | |
3902 | ||
3903 | static void wq_unregister_lockdep(struct workqueue_struct *wq) | |
3904 | { | |
3905 | lockdep_unregister_key(&wq->key); | |
3906 | } | |
3907 | ||
3908 | static void wq_free_lockdep(struct workqueue_struct *wq) | |
3909 | { | |
3910 | if (wq->lock_name != wq->name) | |
3911 | kfree(wq->lock_name); | |
3912 | } | |
3913 | #else | |
3914 | static void wq_init_lockdep(struct workqueue_struct *wq) | |
3915 | { | |
3916 | } | |
3917 | ||
3918 | static void wq_unregister_lockdep(struct workqueue_struct *wq) | |
3919 | { | |
3920 | } | |
3921 | ||
3922 | static void wq_free_lockdep(struct workqueue_struct *wq) | |
3923 | { | |
3924 | } | |
3925 | #endif | |
3926 | ||
6ba94429 | 3927 | static void rcu_free_wq(struct rcu_head *rcu) |
226223ab | 3928 | { |
6ba94429 FW |
3929 | struct workqueue_struct *wq = |
3930 | container_of(rcu, struct workqueue_struct, rcu); | |
226223ab | 3931 | |
669de8bd | 3932 | wq_free_lockdep(wq); |
636b927e TH |
3933 | free_percpu(wq->cpu_pwq); |
3934 | free_workqueue_attrs(wq->unbound_attrs); | |
6ba94429 | 3935 | kfree(wq); |
226223ab TH |
3936 | } |
3937 | ||
6ba94429 | 3938 | static void rcu_free_pool(struct rcu_head *rcu) |
226223ab | 3939 | { |
6ba94429 | 3940 | struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); |
226223ab | 3941 | |
6ba94429 FW |
3942 | ida_destroy(&pool->worker_ida); |
3943 | free_workqueue_attrs(pool->attrs); | |
3944 | kfree(pool); | |
226223ab TH |
3945 | } |
3946 | ||
6ba94429 FW |
3947 | /** |
3948 | * put_unbound_pool - put a worker_pool | |
3949 | * @pool: worker_pool to put | |
3950 | * | |
24acfb71 | 3951 | * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU |
6ba94429 FW |
3952 | * safe manner. get_unbound_pool() calls this function on its failure path |
3953 | * and this function should be able to release pools which went through, | |
3954 | * successfully or not, init_worker_pool(). | |
3955 | * | |
3956 | * Should be called with wq_pool_mutex held. | |
3957 | */ | |
3958 | static void put_unbound_pool(struct worker_pool *pool) | |
226223ab | 3959 | { |
6ba94429 FW |
3960 | DECLARE_COMPLETION_ONSTACK(detach_completion); |
3961 | struct worker *worker; | |
9680540c | 3962 | LIST_HEAD(cull_list); |
e02b9312 | 3963 | |
6ba94429 | 3964 | lockdep_assert_held(&wq_pool_mutex); |
226223ab | 3965 | |
6ba94429 FW |
3966 | if (--pool->refcnt) |
3967 | return; | |
226223ab | 3968 | |
6ba94429 FW |
3969 | /* sanity checks */ |
3970 | if (WARN_ON(!(pool->cpu < 0)) || | |
3971 | WARN_ON(!list_empty(&pool->worklist))) | |
3972 | return; | |
226223ab | 3973 | |
6ba94429 FW |
3974 | /* release id and unhash */ |
3975 | if (pool->id >= 0) | |
3976 | idr_remove(&worker_pool_idr, pool->id); | |
3977 | hash_del(&pool->hash_node); | |
d55262c4 | 3978 | |
6ba94429 | 3979 | /* |
692b4825 TH |
3980 | * Become the manager and destroy all workers. This prevents |
3981 | * @pool's workers from blocking on attach_mutex. We're the last | |
3982 | * manager and @pool gets freed with the flag set. | |
9ab03be4 VS |
3983 | * |
3984 | * Having a concurrent manager is quite unlikely to happen as we can | |
3985 | * only get here with | |
3986 | * pwq->refcnt == pool->refcnt == 0 | |
3987 | * which implies no work queued to the pool, which implies no worker can | |
3988 | * become the manager. However a worker could have taken the role of | |
3989 | * manager before the refcnts dropped to 0, since maybe_create_worker() | |
3990 | * drops pool->lock | |
6ba94429 | 3991 | */ |
9ab03be4 VS |
3992 | while (true) { |
3993 | rcuwait_wait_event(&manager_wait, | |
3994 | !(pool->flags & POOL_MANAGER_ACTIVE), | |
3995 | TASK_UNINTERRUPTIBLE); | |
e02b9312 VS |
3996 | |
3997 | mutex_lock(&wq_pool_attach_mutex); | |
9ab03be4 VS |
3998 | raw_spin_lock_irq(&pool->lock); |
3999 | if (!(pool->flags & POOL_MANAGER_ACTIVE)) { | |
4000 | pool->flags |= POOL_MANAGER_ACTIVE; | |
4001 | break; | |
4002 | } | |
4003 | raw_spin_unlock_irq(&pool->lock); | |
e02b9312 | 4004 | mutex_unlock(&wq_pool_attach_mutex); |
9ab03be4 | 4005 | } |
692b4825 | 4006 | |
6ba94429 | 4007 | while ((worker = first_idle_worker(pool))) |
e02b9312 | 4008 | set_worker_dying(worker, &cull_list); |
6ba94429 | 4009 | WARN_ON(pool->nr_workers || pool->nr_idle); |
a9b8a985 | 4010 | raw_spin_unlock_irq(&pool->lock); |
d55262c4 | 4011 | |
e02b9312 VS |
4012 | wake_dying_workers(&cull_list); |
4013 | ||
4014 | if (!list_empty(&pool->workers) || !list_empty(&pool->dying_workers)) | |
6ba94429 | 4015 | pool->detach_completion = &detach_completion; |
1258fae7 | 4016 | mutex_unlock(&wq_pool_attach_mutex); |
226223ab | 4017 | |
6ba94429 FW |
4018 | if (pool->detach_completion) |
4019 | wait_for_completion(pool->detach_completion); | |
226223ab | 4020 | |
6ba94429 FW |
4021 | /* shut down the timers */ |
4022 | del_timer_sync(&pool->idle_timer); | |
3f959aa3 | 4023 | cancel_work_sync(&pool->idle_cull_work); |
6ba94429 | 4024 | del_timer_sync(&pool->mayday_timer); |
226223ab | 4025 | |
24acfb71 | 4026 | /* RCU protected to allow dereferences from get_work_pool() */ |
25b00775 | 4027 | call_rcu(&pool->rcu, rcu_free_pool); |
226223ab TH |
4028 | } |
4029 | ||
4030 | /** | |
6ba94429 FW |
4031 | * get_unbound_pool - get a worker_pool with the specified attributes |
4032 | * @attrs: the attributes of the worker_pool to get | |
226223ab | 4033 | * |
6ba94429 FW |
4034 | * Obtain a worker_pool which has the same attributes as @attrs, bump the |
4035 | * reference count and return it. If there already is a matching | |
4036 | * worker_pool, it will be used; otherwise, this function attempts to | |
4037 | * create a new one. | |
226223ab | 4038 | * |
6ba94429 | 4039 | * Should be called with wq_pool_mutex held. |
226223ab | 4040 | * |
6ba94429 FW |
4041 | * Return: On success, a worker_pool with the same attributes as @attrs. |
4042 | * On failure, %NULL. | |
226223ab | 4043 | */ |
6ba94429 | 4044 | static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) |
226223ab | 4045 | { |
84193c07 | 4046 | struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_NUMA]; |
6ba94429 FW |
4047 | u32 hash = wqattrs_hash(attrs); |
4048 | struct worker_pool *pool; | |
84193c07 | 4049 | int pod, node = NUMA_NO_NODE; |
226223ab | 4050 | |
6ba94429 | 4051 | lockdep_assert_held(&wq_pool_mutex); |
226223ab | 4052 | |
6ba94429 FW |
4053 | /* do we already have a matching pool? */ |
4054 | hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { | |
4055 | if (wqattrs_equal(pool->attrs, attrs)) { | |
4056 | pool->refcnt++; | |
4057 | return pool; | |
4058 | } | |
4059 | } | |
226223ab | 4060 | |
9546b29e | 4061 | /* If __pod_cpumask is contained inside a NUMA pod, that's our node */ |
84193c07 | 4062 | for (pod = 0; pod < pt->nr_pods; pod++) { |
9546b29e | 4063 | if (cpumask_subset(attrs->__pod_cpumask, pt->pod_cpus[pod])) { |
84193c07 TH |
4064 | node = pt->pod_node[pod]; |
4065 | break; | |
e2273584 XP |
4066 | } |
4067 | } | |
4068 | ||
6ba94429 | 4069 | /* nope, create a new one */ |
84193c07 | 4070 | pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, node); |
6ba94429 FW |
4071 | if (!pool || init_worker_pool(pool) < 0) |
4072 | goto fail; | |
4073 | ||
84193c07 | 4074 | pool->node = node; |
5de7a03c TH |
4075 | copy_workqueue_attrs(pool->attrs, attrs); |
4076 | wqattrs_clear_for_pool(pool->attrs); | |
226223ab | 4077 | |
6ba94429 FW |
4078 | if (worker_pool_assign_id(pool) < 0) |
4079 | goto fail; | |
226223ab | 4080 | |
6ba94429 | 4081 | /* create and start the initial worker */ |
3347fa09 | 4082 | if (wq_online && !create_worker(pool)) |
6ba94429 | 4083 | goto fail; |
226223ab | 4084 | |
6ba94429 FW |
4085 | /* install */ |
4086 | hash_add(unbound_pool_hash, &pool->hash_node, hash); | |
226223ab | 4087 | |
6ba94429 FW |
4088 | return pool; |
4089 | fail: | |
4090 | if (pool) | |
4091 | put_unbound_pool(pool); | |
4092 | return NULL; | |
226223ab | 4093 | } |
226223ab | 4094 | |
6ba94429 | 4095 | static void rcu_free_pwq(struct rcu_head *rcu) |
7a4e344c | 4096 | { |
6ba94429 FW |
4097 | kmem_cache_free(pwq_cache, |
4098 | container_of(rcu, struct pool_workqueue, rcu)); | |
7a4e344c TH |
4099 | } |
4100 | ||
6ba94429 | 4101 | /* |
967b494e TH |
4102 | * Scheduled on pwq_release_worker by put_pwq() when an unbound pwq hits zero |
4103 | * refcnt and needs to be destroyed. | |
7a4e344c | 4104 | */ |
687a9aa5 | 4105 | static void pwq_release_workfn(struct kthread_work *work) |
7a4e344c | 4106 | { |
6ba94429 | 4107 | struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, |
687a9aa5 | 4108 | release_work); |
6ba94429 FW |
4109 | struct workqueue_struct *wq = pwq->wq; |
4110 | struct worker_pool *pool = pwq->pool; | |
b42b0bdd | 4111 | bool is_last = false; |
7a4e344c | 4112 | |
b42b0bdd | 4113 | /* |
687a9aa5 | 4114 | * When @pwq is not linked, it doesn't hold any reference to the |
b42b0bdd YY |
4115 | * @wq, and @wq is invalid to access. |
4116 | */ | |
4117 | if (!list_empty(&pwq->pwqs_node)) { | |
b42b0bdd YY |
4118 | mutex_lock(&wq->mutex); |
4119 | list_del_rcu(&pwq->pwqs_node); | |
4120 | is_last = list_empty(&wq->pwqs); | |
4121 | mutex_unlock(&wq->mutex); | |
4122 | } | |
6ba94429 | 4123 | |
687a9aa5 TH |
4124 | if (wq->flags & WQ_UNBOUND) { |
4125 | mutex_lock(&wq_pool_mutex); | |
4126 | put_unbound_pool(pool); | |
4127 | mutex_unlock(&wq_pool_mutex); | |
4128 | } | |
6ba94429 | 4129 | |
25b00775 | 4130 | call_rcu(&pwq->rcu, rcu_free_pwq); |
7a4e344c | 4131 | |
2865a8fb | 4132 | /* |
6ba94429 FW |
4133 | * If we're the last pwq going away, @wq is already dead and no one |
4134 | * is gonna access it anymore. Schedule RCU free. | |
2865a8fb | 4135 | */ |
669de8bd BVA |
4136 | if (is_last) { |
4137 | wq_unregister_lockdep(wq); | |
25b00775 | 4138 | call_rcu(&wq->rcu, rcu_free_wq); |
669de8bd | 4139 | } |
29c91e99 TH |
4140 | } |
4141 | ||
7a4e344c | 4142 | /** |
6ba94429 FW |
4143 | * pwq_adjust_max_active - update a pwq's max_active to the current setting |
4144 | * @pwq: target pool_workqueue | |
d185af30 | 4145 | * |
6ba94429 | 4146 | * If @pwq isn't freezing, set @pwq->max_active to the associated |
f97a4a1a | 4147 | * workqueue's saved_max_active and activate inactive work items |
6ba94429 | 4148 | * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. |
7a4e344c | 4149 | */ |
6ba94429 | 4150 | static void pwq_adjust_max_active(struct pool_workqueue *pwq) |
4e1a1f9a | 4151 | { |
6ba94429 FW |
4152 | struct workqueue_struct *wq = pwq->wq; |
4153 | bool freezable = wq->flags & WQ_FREEZABLE; | |
3347fa09 | 4154 | unsigned long flags; |
4e1a1f9a | 4155 | |
6ba94429 FW |
4156 | /* for @wq->saved_max_active */ |
4157 | lockdep_assert_held(&wq->mutex); | |
4e1a1f9a | 4158 | |
6ba94429 FW |
4159 | /* fast exit for non-freezable wqs */ |
4160 | if (!freezable && pwq->max_active == wq->saved_max_active) | |
4161 | return; | |
7a4e344c | 4162 | |
3347fa09 | 4163 | /* this function can be called during early boot w/ irq disabled */ |
a9b8a985 | 4164 | raw_spin_lock_irqsave(&pwq->pool->lock, flags); |
29c91e99 | 4165 | |
6ba94429 FW |
4166 | /* |
4167 | * During [un]freezing, the caller is responsible for ensuring that | |
4168 | * this function is called at least once after @workqueue_freezing | |
4169 | * is updated and visible. | |
4170 | */ | |
4171 | if (!freezable || !workqueue_freezing) { | |
4172 | pwq->max_active = wq->saved_max_active; | |
4e1a1f9a | 4173 | |
f97a4a1a | 4174 | while (!list_empty(&pwq->inactive_works) && |
0219a352 | 4175 | pwq->nr_active < pwq->max_active) |
f97a4a1a | 4176 | pwq_activate_first_inactive(pwq); |
e2dca7ad | 4177 | |
0219a352 | 4178 | kick_pool(pwq->pool); |
6ba94429 FW |
4179 | } else { |
4180 | pwq->max_active = 0; | |
4181 | } | |
e2dca7ad | 4182 | |
a9b8a985 | 4183 | raw_spin_unlock_irqrestore(&pwq->pool->lock, flags); |
e2dca7ad TH |
4184 | } |
4185 | ||
67dc8325 | 4186 | /* initialize newly allocated @pwq which is associated with @wq and @pool */ |
6ba94429 FW |
4187 | static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, |
4188 | struct worker_pool *pool) | |
29c91e99 | 4189 | { |
6ba94429 | 4190 | BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); |
29c91e99 | 4191 | |
6ba94429 FW |
4192 | memset(pwq, 0, sizeof(*pwq)); |
4193 | ||
4194 | pwq->pool = pool; | |
4195 | pwq->wq = wq; | |
4196 | pwq->flush_color = -1; | |
4197 | pwq->refcnt = 1; | |
f97a4a1a | 4198 | INIT_LIST_HEAD(&pwq->inactive_works); |
6ba94429 FW |
4199 | INIT_LIST_HEAD(&pwq->pwqs_node); |
4200 | INIT_LIST_HEAD(&pwq->mayday_node); | |
687a9aa5 | 4201 | kthread_init_work(&pwq->release_work, pwq_release_workfn); |
29c91e99 TH |
4202 | } |
4203 | ||
6ba94429 FW |
4204 | /* sync @pwq with the current state of its associated wq and link it */ |
4205 | static void link_pwq(struct pool_workqueue *pwq) | |
29c91e99 | 4206 | { |
6ba94429 | 4207 | struct workqueue_struct *wq = pwq->wq; |
29c91e99 | 4208 | |
6ba94429 | 4209 | lockdep_assert_held(&wq->mutex); |
a892cacc | 4210 | |
6ba94429 FW |
4211 | /* may be called multiple times, ignore if already linked */ |
4212 | if (!list_empty(&pwq->pwqs_node)) | |
29c91e99 | 4213 | return; |
29c91e99 | 4214 | |
6ba94429 FW |
4215 | /* set the matching work_color */ |
4216 | pwq->work_color = wq->work_color; | |
29c91e99 | 4217 | |
6ba94429 FW |
4218 | /* sync max_active to the current setting */ |
4219 | pwq_adjust_max_active(pwq); | |
29c91e99 | 4220 | |
6ba94429 FW |
4221 | /* link in @pwq */ |
4222 | list_add_rcu(&pwq->pwqs_node, &wq->pwqs); | |
4223 | } | |
29c91e99 | 4224 | |
6ba94429 FW |
4225 | /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ |
4226 | static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, | |
4227 | const struct workqueue_attrs *attrs) | |
4228 | { | |
4229 | struct worker_pool *pool; | |
4230 | struct pool_workqueue *pwq; | |
60f5a4bc | 4231 | |
6ba94429 | 4232 | lockdep_assert_held(&wq_pool_mutex); |
60f5a4bc | 4233 | |
6ba94429 FW |
4234 | pool = get_unbound_pool(attrs); |
4235 | if (!pool) | |
4236 | return NULL; | |
60f5a4bc | 4237 | |
6ba94429 FW |
4238 | pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); |
4239 | if (!pwq) { | |
4240 | put_unbound_pool(pool); | |
4241 | return NULL; | |
4242 | } | |
29c91e99 | 4243 | |
6ba94429 FW |
4244 | init_pwq(pwq, wq, pool); |
4245 | return pwq; | |
4246 | } | |
29c91e99 | 4247 | |
29c91e99 | 4248 | /** |
fef59c9c | 4249 | * wq_calc_pod_cpumask - calculate a wq_attrs' cpumask for a pod |
042f7df1 | 4250 | * @attrs: the wq_attrs of the default pwq of the target workqueue |
84193c07 | 4251 | * @cpu: the target CPU |
6ba94429 | 4252 | * @cpu_going_down: if >= 0, the CPU to consider as offline |
29c91e99 | 4253 | * |
fef59c9c TH |
4254 | * Calculate the cpumask a workqueue with @attrs should use on @pod. If |
4255 | * @cpu_going_down is >= 0, that cpu is considered offline during calculation. | |
9546b29e | 4256 | * The result is stored in @attrs->__pod_cpumask. |
a892cacc | 4257 | * |
fef59c9c TH |
4258 | * If pod affinity is not enabled, @attrs->cpumask is always used. If enabled |
4259 | * and @pod has online CPUs requested by @attrs, the returned cpumask is the | |
4260 | * intersection of the possible CPUs of @pod and @attrs->cpumask. | |
d185af30 | 4261 | * |
fef59c9c | 4262 | * The caller is responsible for ensuring that the cpumask of @pod stays stable. |
29c91e99 | 4263 | */ |
9546b29e TH |
4264 | static void wq_calc_pod_cpumask(struct workqueue_attrs *attrs, int cpu, |
4265 | int cpu_going_down) | |
29c91e99 | 4266 | { |
84193c07 TH |
4267 | const struct wq_pod_type *pt = wqattrs_pod_type(attrs); |
4268 | int pod = pt->cpu_pod[cpu]; | |
29c91e99 | 4269 | |
fef59c9c | 4270 | /* does @pod have any online CPUs @attrs wants? */ |
9546b29e TH |
4271 | cpumask_and(attrs->__pod_cpumask, pt->pod_cpus[pod], attrs->cpumask); |
4272 | cpumask_and(attrs->__pod_cpumask, attrs->__pod_cpumask, cpu_online_mask); | |
6ba94429 | 4273 | if (cpu_going_down >= 0) |
9546b29e | 4274 | cpumask_clear_cpu(cpu_going_down, attrs->__pod_cpumask); |
29c91e99 | 4275 | |
9546b29e TH |
4276 | if (cpumask_empty(attrs->__pod_cpumask)) { |
4277 | cpumask_copy(attrs->__pod_cpumask, attrs->cpumask); | |
84193c07 TH |
4278 | return; |
4279 | } | |
4c16bd32 | 4280 | |
fef59c9c | 4281 | /* yeap, return possible CPUs in @pod that @attrs wants */ |
9546b29e | 4282 | cpumask_and(attrs->__pod_cpumask, attrs->cpumask, pt->pod_cpus[pod]); |
1ad0f0a7 | 4283 | |
9546b29e | 4284 | if (cpumask_empty(attrs->__pod_cpumask)) |
1ad0f0a7 MB |
4285 | pr_warn_once("WARNING: workqueue cpumask: online intersect > " |
4286 | "possible intersect\n"); | |
4c16bd32 TH |
4287 | } |
4288 | ||
636b927e TH |
4289 | /* install @pwq into @wq's cpu_pwq and return the old pwq */ |
4290 | static struct pool_workqueue *install_unbound_pwq(struct workqueue_struct *wq, | |
4291 | int cpu, struct pool_workqueue *pwq) | |
1befcf30 TH |
4292 | { |
4293 | struct pool_workqueue *old_pwq; | |
4294 | ||
5b95e1af | 4295 | lockdep_assert_held(&wq_pool_mutex); |
1befcf30 TH |
4296 | lockdep_assert_held(&wq->mutex); |
4297 | ||
4298 | /* link_pwq() can handle duplicate calls */ | |
4299 | link_pwq(pwq); | |
4300 | ||
636b927e TH |
4301 | old_pwq = rcu_access_pointer(*per_cpu_ptr(wq->cpu_pwq, cpu)); |
4302 | rcu_assign_pointer(*per_cpu_ptr(wq->cpu_pwq, cpu), pwq); | |
1befcf30 TH |
4303 | return old_pwq; |
4304 | } | |
4305 | ||
2d5f0764 LJ |
4306 | /* context to store the prepared attrs & pwqs before applying */ |
4307 | struct apply_wqattrs_ctx { | |
4308 | struct workqueue_struct *wq; /* target workqueue */ | |
4309 | struct workqueue_attrs *attrs; /* attrs to apply */ | |
042f7df1 | 4310 | struct list_head list; /* queued for batching commit */ |
2d5f0764 LJ |
4311 | struct pool_workqueue *dfl_pwq; |
4312 | struct pool_workqueue *pwq_tbl[]; | |
4313 | }; | |
4314 | ||
4315 | /* free the resources after success or abort */ | |
4316 | static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx) | |
4317 | { | |
4318 | if (ctx) { | |
636b927e | 4319 | int cpu; |
2d5f0764 | 4320 | |
636b927e TH |
4321 | for_each_possible_cpu(cpu) |
4322 | put_pwq_unlocked(ctx->pwq_tbl[cpu]); | |
2d5f0764 LJ |
4323 | put_pwq_unlocked(ctx->dfl_pwq); |
4324 | ||
4325 | free_workqueue_attrs(ctx->attrs); | |
4326 | ||
4327 | kfree(ctx); | |
4328 | } | |
4329 | } | |
4330 | ||
4331 | /* allocate the attrs and pwqs for later installation */ | |
4332 | static struct apply_wqattrs_ctx * | |
4333 | apply_wqattrs_prepare(struct workqueue_struct *wq, | |
99c621ef LJ |
4334 | const struct workqueue_attrs *attrs, |
4335 | const cpumask_var_t unbound_cpumask) | |
9e8cd2f5 | 4336 | { |
2d5f0764 | 4337 | struct apply_wqattrs_ctx *ctx; |
9546b29e | 4338 | struct workqueue_attrs *new_attrs; |
636b927e | 4339 | int cpu; |
9e8cd2f5 | 4340 | |
2d5f0764 | 4341 | lockdep_assert_held(&wq_pool_mutex); |
9e8cd2f5 | 4342 | |
84193c07 TH |
4343 | if (WARN_ON(attrs->affn_scope < 0 || |
4344 | attrs->affn_scope >= WQ_AFFN_NR_TYPES)) | |
4345 | return ERR_PTR(-EINVAL); | |
4346 | ||
636b927e | 4347 | ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_cpu_ids), GFP_KERNEL); |
8719dcea | 4348 | |
be69d00d | 4349 | new_attrs = alloc_workqueue_attrs(); |
9546b29e | 4350 | if (!ctx || !new_attrs) |
2d5f0764 | 4351 | goto out_free; |
13e2e556 | 4352 | |
4c16bd32 TH |
4353 | /* |
4354 | * If something goes wrong during CPU up/down, we'll fall back to | |
4355 | * the default pwq covering whole @attrs->cpumask. Always create | |
4356 | * it even if we don't use it immediately. | |
4357 | */ | |
0f36ee24 TH |
4358 | copy_workqueue_attrs(new_attrs, attrs); |
4359 | wqattrs_actualize_cpumask(new_attrs, unbound_cpumask); | |
9546b29e | 4360 | cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask); |
2d5f0764 LJ |
4361 | ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs); |
4362 | if (!ctx->dfl_pwq) | |
4363 | goto out_free; | |
4c16bd32 | 4364 | |
636b927e | 4365 | for_each_possible_cpu(cpu) { |
af73f5c9 | 4366 | if (new_attrs->ordered) { |
2d5f0764 | 4367 | ctx->dfl_pwq->refcnt++; |
636b927e TH |
4368 | ctx->pwq_tbl[cpu] = ctx->dfl_pwq; |
4369 | } else { | |
9546b29e TH |
4370 | wq_calc_pod_cpumask(new_attrs, cpu, -1); |
4371 | ctx->pwq_tbl[cpu] = alloc_unbound_pwq(wq, new_attrs); | |
636b927e TH |
4372 | if (!ctx->pwq_tbl[cpu]) |
4373 | goto out_free; | |
4c16bd32 TH |
4374 | } |
4375 | } | |
4376 | ||
042f7df1 LJ |
4377 | /* save the user configured attrs and sanitize it. */ |
4378 | copy_workqueue_attrs(new_attrs, attrs); | |
4379 | cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); | |
9546b29e | 4380 | cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask); |
2d5f0764 | 4381 | ctx->attrs = new_attrs; |
042f7df1 | 4382 | |
2d5f0764 | 4383 | ctx->wq = wq; |
2d5f0764 LJ |
4384 | return ctx; |
4385 | ||
4386 | out_free: | |
2d5f0764 LJ |
4387 | free_workqueue_attrs(new_attrs); |
4388 | apply_wqattrs_cleanup(ctx); | |
84193c07 | 4389 | return ERR_PTR(-ENOMEM); |
2d5f0764 LJ |
4390 | } |
4391 | ||
4392 | /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */ | |
4393 | static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx) | |
4394 | { | |
636b927e | 4395 | int cpu; |
9e8cd2f5 | 4396 | |
4c16bd32 | 4397 | /* all pwqs have been created successfully, let's install'em */ |
2d5f0764 | 4398 | mutex_lock(&ctx->wq->mutex); |
a892cacc | 4399 | |
2d5f0764 | 4400 | copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs); |
4c16bd32 TH |
4401 | |
4402 | /* save the previous pwq and install the new one */ | |
636b927e TH |
4403 | for_each_possible_cpu(cpu) |
4404 | ctx->pwq_tbl[cpu] = install_unbound_pwq(ctx->wq, cpu, | |
4405 | ctx->pwq_tbl[cpu]); | |
4c16bd32 TH |
4406 | |
4407 | /* @dfl_pwq might not have been used, ensure it's linked */ | |
2d5f0764 LJ |
4408 | link_pwq(ctx->dfl_pwq); |
4409 | swap(ctx->wq->dfl_pwq, ctx->dfl_pwq); | |
f147f29e | 4410 | |
2d5f0764 LJ |
4411 | mutex_unlock(&ctx->wq->mutex); |
4412 | } | |
9e8cd2f5 | 4413 | |
a0111cf6 LJ |
4414 | static void apply_wqattrs_lock(void) |
4415 | { | |
4416 | /* CPUs should stay stable across pwq creations and installations */ | |
ffd8bea8 | 4417 | cpus_read_lock(); |
a0111cf6 LJ |
4418 | mutex_lock(&wq_pool_mutex); |
4419 | } | |
4420 | ||
4421 | static void apply_wqattrs_unlock(void) | |
4422 | { | |
4423 | mutex_unlock(&wq_pool_mutex); | |
ffd8bea8 | 4424 | cpus_read_unlock(); |
a0111cf6 LJ |
4425 | } |
4426 | ||
4427 | static int apply_workqueue_attrs_locked(struct workqueue_struct *wq, | |
4428 | const struct workqueue_attrs *attrs) | |
2d5f0764 LJ |
4429 | { |
4430 | struct apply_wqattrs_ctx *ctx; | |
4c16bd32 | 4431 | |
2d5f0764 LJ |
4432 | /* only unbound workqueues can change attributes */ |
4433 | if (WARN_ON(!(wq->flags & WQ_UNBOUND))) | |
4434 | return -EINVAL; | |
13e2e556 | 4435 | |
2d5f0764 | 4436 | /* creating multiple pwqs breaks ordering guarantee */ |
0a94efb5 TH |
4437 | if (!list_empty(&wq->pwqs)) { |
4438 | if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) | |
4439 | return -EINVAL; | |
4440 | ||
4441 | wq->flags &= ~__WQ_ORDERED; | |
4442 | } | |
2d5f0764 | 4443 | |
99c621ef | 4444 | ctx = apply_wqattrs_prepare(wq, attrs, wq_unbound_cpumask); |
84193c07 TH |
4445 | if (IS_ERR(ctx)) |
4446 | return PTR_ERR(ctx); | |
2d5f0764 LJ |
4447 | |
4448 | /* the ctx has been prepared successfully, let's commit it */ | |
6201171e | 4449 | apply_wqattrs_commit(ctx); |
2d5f0764 LJ |
4450 | apply_wqattrs_cleanup(ctx); |
4451 | ||
6201171e | 4452 | return 0; |
9e8cd2f5 TH |
4453 | } |
4454 | ||
a0111cf6 LJ |
4455 | /** |
4456 | * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue | |
4457 | * @wq: the target workqueue | |
4458 | * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() | |
4459 | * | |
fef59c9c TH |
4460 | * Apply @attrs to an unbound workqueue @wq. Unless disabled, this function maps |
4461 | * a separate pwq to each CPU pod with possibles CPUs in @attrs->cpumask so that | |
4462 | * work items are affine to the pod it was issued on. Older pwqs are released as | |
4463 | * in-flight work items finish. Note that a work item which repeatedly requeues | |
4464 | * itself back-to-back will stay on its current pwq. | |
a0111cf6 LJ |
4465 | * |
4466 | * Performs GFP_KERNEL allocations. | |
4467 | * | |
ffd8bea8 | 4468 | * Assumes caller has CPU hotplug read exclusion, i.e. cpus_read_lock(). |
509b3204 | 4469 | * |
a0111cf6 LJ |
4470 | * Return: 0 on success and -errno on failure. |
4471 | */ | |
513c98d0 | 4472 | int apply_workqueue_attrs(struct workqueue_struct *wq, |
a0111cf6 LJ |
4473 | const struct workqueue_attrs *attrs) |
4474 | { | |
4475 | int ret; | |
4476 | ||
509b3204 DJ |
4477 | lockdep_assert_cpus_held(); |
4478 | ||
4479 | mutex_lock(&wq_pool_mutex); | |
a0111cf6 | 4480 | ret = apply_workqueue_attrs_locked(wq, attrs); |
509b3204 | 4481 | mutex_unlock(&wq_pool_mutex); |
a0111cf6 LJ |
4482 | |
4483 | return ret; | |
4484 | } | |
4485 | ||
4c16bd32 | 4486 | /** |
fef59c9c | 4487 | * wq_update_pod - update pod affinity of a wq for CPU hot[un]plug |
4c16bd32 | 4488 | * @wq: the target workqueue |
4cbfd3de TH |
4489 | * @cpu: the CPU to update pool association for |
4490 | * @hotplug_cpu: the CPU coming up or going down | |
4c16bd32 TH |
4491 | * @online: whether @cpu is coming up or going down |
4492 | * | |
4493 | * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and | |
fef59c9c | 4494 | * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update pod affinity of |
4c16bd32 TH |
4495 | * @wq accordingly. |
4496 | * | |
fef59c9c TH |
4497 | * |
4498 | * If pod affinity can't be adjusted due to memory allocation failure, it falls | |
4499 | * back to @wq->dfl_pwq which may not be optimal but is always correct. | |
4500 | * | |
4501 | * Note that when the last allowed CPU of a pod goes offline for a workqueue | |
4502 | * with a cpumask spanning multiple pods, the workers which were already | |
4503 | * executing the work items for the workqueue will lose their CPU affinity and | |
4504 | * may execute on any CPU. This is similar to how per-cpu workqueues behave on | |
4505 | * CPU_DOWN. If a workqueue user wants strict affinity, it's the user's | |
4506 | * responsibility to flush the work item from CPU_DOWN_PREPARE. | |
4c16bd32 | 4507 | */ |
fef59c9c TH |
4508 | static void wq_update_pod(struct workqueue_struct *wq, int cpu, |
4509 | int hotplug_cpu, bool online) | |
4c16bd32 | 4510 | { |
4cbfd3de | 4511 | int off_cpu = online ? -1 : hotplug_cpu; |
4c16bd32 TH |
4512 | struct pool_workqueue *old_pwq = NULL, *pwq; |
4513 | struct workqueue_attrs *target_attrs; | |
4c16bd32 TH |
4514 | |
4515 | lockdep_assert_held(&wq_pool_mutex); | |
4516 | ||
84193c07 | 4517 | if (!(wq->flags & WQ_UNBOUND) || wq->unbound_attrs->ordered) |
4c16bd32 TH |
4518 | return; |
4519 | ||
4520 | /* | |
4521 | * We don't wanna alloc/free wq_attrs for each wq for each CPU. | |
4522 | * Let's use a preallocated one. The following buf is protected by | |
4523 | * CPU hotplug exclusion. | |
4524 | */ | |
fef59c9c | 4525 | target_attrs = wq_update_pod_attrs_buf; |
4c16bd32 | 4526 | |
4c16bd32 | 4527 | copy_workqueue_attrs(target_attrs, wq->unbound_attrs); |
0f36ee24 | 4528 | wqattrs_actualize_cpumask(target_attrs, wq_unbound_cpumask); |
4c16bd32 | 4529 | |
636b927e | 4530 | /* nothing to do if the target cpumask matches the current pwq */ |
9546b29e | 4531 | wq_calc_pod_cpumask(target_attrs, cpu, off_cpu); |
636b927e TH |
4532 | pwq = rcu_dereference_protected(*per_cpu_ptr(wq->cpu_pwq, cpu), |
4533 | lockdep_is_held(&wq_pool_mutex)); | |
9546b29e | 4534 | if (wqattrs_equal(target_attrs, pwq->pool->attrs)) |
636b927e | 4535 | return; |
4c16bd32 | 4536 | |
4c16bd32 TH |
4537 | /* create a new pwq */ |
4538 | pwq = alloc_unbound_pwq(wq, target_attrs); | |
4539 | if (!pwq) { | |
fef59c9c | 4540 | pr_warn("workqueue: allocation failed while updating CPU pod affinity of \"%s\"\n", |
2d916033 | 4541 | wq->name); |
77f300b1 | 4542 | goto use_dfl_pwq; |
4c16bd32 TH |
4543 | } |
4544 | ||
f7142ed4 | 4545 | /* Install the new pwq. */ |
4c16bd32 | 4546 | mutex_lock(&wq->mutex); |
636b927e | 4547 | old_pwq = install_unbound_pwq(wq, cpu, pwq); |
4c16bd32 TH |
4548 | goto out_unlock; |
4549 | ||
4550 | use_dfl_pwq: | |
f7142ed4 | 4551 | mutex_lock(&wq->mutex); |
a9b8a985 | 4552 | raw_spin_lock_irq(&wq->dfl_pwq->pool->lock); |
4c16bd32 | 4553 | get_pwq(wq->dfl_pwq); |
a9b8a985 | 4554 | raw_spin_unlock_irq(&wq->dfl_pwq->pool->lock); |
636b927e | 4555 | old_pwq = install_unbound_pwq(wq, cpu, wq->dfl_pwq); |
4c16bd32 TH |
4556 | out_unlock: |
4557 | mutex_unlock(&wq->mutex); | |
4558 | put_pwq_unlocked(old_pwq); | |
4559 | } | |
4560 | ||
30cdf249 | 4561 | static int alloc_and_link_pwqs(struct workqueue_struct *wq) |
0f900049 | 4562 | { |
49e3cf44 | 4563 | bool highpri = wq->flags & WQ_HIGHPRI; |
8a2b7538 | 4564 | int cpu, ret; |
30cdf249 | 4565 | |
636b927e TH |
4566 | wq->cpu_pwq = alloc_percpu(struct pool_workqueue *); |
4567 | if (!wq->cpu_pwq) | |
4568 | goto enomem; | |
30cdf249 | 4569 | |
636b927e | 4570 | if (!(wq->flags & WQ_UNBOUND)) { |
30cdf249 | 4571 | for_each_possible_cpu(cpu) { |
687a9aa5 | 4572 | struct pool_workqueue **pwq_p = |
ee1ceef7 | 4573 | per_cpu_ptr(wq->cpu_pwq, cpu); |
687a9aa5 TH |
4574 | struct worker_pool *pool = |
4575 | &(per_cpu_ptr(cpu_worker_pools, cpu)[highpri]); | |
4576 | ||
4577 | *pwq_p = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, | |
4578 | pool->node); | |
4579 | if (!*pwq_p) | |
4580 | goto enomem; | |
f3421797 | 4581 | |
687a9aa5 | 4582 | init_pwq(*pwq_p, wq, pool); |
f147f29e TH |
4583 | |
4584 | mutex_lock(&wq->mutex); | |
687a9aa5 | 4585 | link_pwq(*pwq_p); |
f147f29e | 4586 | mutex_unlock(&wq->mutex); |
30cdf249 | 4587 | } |
9e8cd2f5 | 4588 | return 0; |
509b3204 DJ |
4589 | } |
4590 | ||
ffd8bea8 | 4591 | cpus_read_lock(); |
509b3204 | 4592 | if (wq->flags & __WQ_ORDERED) { |
8a2b7538 TH |
4593 | ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]); |
4594 | /* there should only be single pwq for ordering guarantee */ | |
4595 | WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node || | |
4596 | wq->pwqs.prev != &wq->dfl_pwq->pwqs_node), | |
4597 | "ordering guarantee broken for workqueue %s\n", wq->name); | |
30cdf249 | 4598 | } else { |
509b3204 | 4599 | ret = apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); |
30cdf249 | 4600 | } |
ffd8bea8 | 4601 | cpus_read_unlock(); |
509b3204 DJ |
4602 | |
4603 | return ret; | |
687a9aa5 TH |
4604 | |
4605 | enomem: | |
4606 | if (wq->cpu_pwq) { | |
4607 | for_each_possible_cpu(cpu) | |
4608 | kfree(*per_cpu_ptr(wq->cpu_pwq, cpu)); | |
4609 | free_percpu(wq->cpu_pwq); | |
4610 | wq->cpu_pwq = NULL; | |
4611 | } | |
4612 | return -ENOMEM; | |
0f900049 TH |
4613 | } |
4614 | ||
f3421797 TH |
4615 | static int wq_clamp_max_active(int max_active, unsigned int flags, |
4616 | const char *name) | |
b71ab8c2 | 4617 | { |
636b927e | 4618 | if (max_active < 1 || max_active > WQ_MAX_ACTIVE) |
044c782c | 4619 | pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", |
636b927e | 4620 | max_active, name, 1, WQ_MAX_ACTIVE); |
b71ab8c2 | 4621 | |
636b927e | 4622 | return clamp_val(max_active, 1, WQ_MAX_ACTIVE); |
b71ab8c2 TH |
4623 | } |
4624 | ||
983c7515 TH |
4625 | /* |
4626 | * Workqueues which may be used during memory reclaim should have a rescuer | |
4627 | * to guarantee forward progress. | |
4628 | */ | |
4629 | static int init_rescuer(struct workqueue_struct *wq) | |
4630 | { | |
4631 | struct worker *rescuer; | |
b92b36ea | 4632 | int ret; |
983c7515 TH |
4633 | |
4634 | if (!(wq->flags & WQ_MEM_RECLAIM)) | |
4635 | return 0; | |
4636 | ||
4637 | rescuer = alloc_worker(NUMA_NO_NODE); | |
4c0736a7 PM |
4638 | if (!rescuer) { |
4639 | pr_err("workqueue: Failed to allocate a rescuer for wq \"%s\"\n", | |
4640 | wq->name); | |
983c7515 | 4641 | return -ENOMEM; |
4c0736a7 | 4642 | } |
983c7515 TH |
4643 | |
4644 | rescuer->rescue_wq = wq; | |
b6a46f72 | 4645 | rescuer->task = kthread_create(rescuer_thread, rescuer, "kworker/R-%s", wq->name); |
f187b697 | 4646 | if (IS_ERR(rescuer->task)) { |
b92b36ea | 4647 | ret = PTR_ERR(rescuer->task); |
4c0736a7 PM |
4648 | pr_err("workqueue: Failed to create a rescuer kthread for wq \"%s\": %pe", |
4649 | wq->name, ERR_PTR(ret)); | |
983c7515 | 4650 | kfree(rescuer); |
b92b36ea | 4651 | return ret; |
983c7515 TH |
4652 | } |
4653 | ||
4654 | wq->rescuer = rescuer; | |
4655 | kthread_bind_mask(rescuer->task, cpu_possible_mask); | |
4656 | wake_up_process(rescuer->task); | |
4657 | ||
4658 | return 0; | |
4659 | } | |
4660 | ||
a2775bbc | 4661 | __printf(1, 4) |
669de8bd BVA |
4662 | struct workqueue_struct *alloc_workqueue(const char *fmt, |
4663 | unsigned int flags, | |
4664 | int max_active, ...) | |
1da177e4 | 4665 | { |
ecf6881f | 4666 | va_list args; |
1da177e4 | 4667 | struct workqueue_struct *wq; |
49e3cf44 | 4668 | struct pool_workqueue *pwq; |
b196be89 | 4669 | |
5c0338c6 | 4670 | /* |
fef59c9c TH |
4671 | * Unbound && max_active == 1 used to imply ordered, which is no longer |
4672 | * the case on many machines due to per-pod pools. While | |
5c0338c6 | 4673 | * alloc_ordered_workqueue() is the right way to create an ordered |
fef59c9c | 4674 | * workqueue, keep the previous behavior to avoid subtle breakages. |
5c0338c6 TH |
4675 | */ |
4676 | if ((flags & WQ_UNBOUND) && max_active == 1) | |
4677 | flags |= __WQ_ORDERED; | |
4678 | ||
cee22a15 VK |
4679 | /* see the comment above the definition of WQ_POWER_EFFICIENT */ |
4680 | if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) | |
4681 | flags |= WQ_UNBOUND; | |
4682 | ||
ecf6881f | 4683 | /* allocate wq and format name */ |
636b927e | 4684 | wq = kzalloc(sizeof(*wq), GFP_KERNEL); |
b196be89 | 4685 | if (!wq) |
d2c1d404 | 4686 | return NULL; |
b196be89 | 4687 | |
6029a918 | 4688 | if (flags & WQ_UNBOUND) { |
be69d00d | 4689 | wq->unbound_attrs = alloc_workqueue_attrs(); |
6029a918 TH |
4690 | if (!wq->unbound_attrs) |
4691 | goto err_free_wq; | |
4692 | } | |
4693 | ||
669de8bd | 4694 | va_start(args, max_active); |
ecf6881f | 4695 | vsnprintf(wq->name, sizeof(wq->name), fmt, args); |
b196be89 | 4696 | va_end(args); |
1da177e4 | 4697 | |
d320c038 | 4698 | max_active = max_active ?: WQ_DFL_ACTIVE; |
b196be89 | 4699 | max_active = wq_clamp_max_active(max_active, flags, wq->name); |
3af24433 | 4700 | |
b196be89 | 4701 | /* init wq */ |
97e37d7b | 4702 | wq->flags = flags; |
a0a1a5fd | 4703 | wq->saved_max_active = max_active; |
3c25a55d | 4704 | mutex_init(&wq->mutex); |
112202d9 | 4705 | atomic_set(&wq->nr_pwqs_to_flush, 0); |
30cdf249 | 4706 | INIT_LIST_HEAD(&wq->pwqs); |
73f53c4a TH |
4707 | INIT_LIST_HEAD(&wq->flusher_queue); |
4708 | INIT_LIST_HEAD(&wq->flusher_overflow); | |
493a1724 | 4709 | INIT_LIST_HEAD(&wq->maydays); |
502ca9d8 | 4710 | |
669de8bd | 4711 | wq_init_lockdep(wq); |
cce1a165 | 4712 | INIT_LIST_HEAD(&wq->list); |
3af24433 | 4713 | |
30cdf249 | 4714 | if (alloc_and_link_pwqs(wq) < 0) |
82efcab3 | 4715 | goto err_unreg_lockdep; |
1537663f | 4716 | |
40c17f75 | 4717 | if (wq_online && init_rescuer(wq) < 0) |
983c7515 | 4718 | goto err_destroy; |
3af24433 | 4719 | |
226223ab TH |
4720 | if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) |
4721 | goto err_destroy; | |
4722 | ||
a0a1a5fd | 4723 | /* |
68e13a67 LJ |
4724 | * wq_pool_mutex protects global freeze state and workqueues list. |
4725 | * Grab it, adjust max_active and add the new @wq to workqueues | |
4726 | * list. | |
a0a1a5fd | 4727 | */ |
68e13a67 | 4728 | mutex_lock(&wq_pool_mutex); |
a0a1a5fd | 4729 | |
a357fc03 | 4730 | mutex_lock(&wq->mutex); |
699ce097 TH |
4731 | for_each_pwq(pwq, wq) |
4732 | pwq_adjust_max_active(pwq); | |
a357fc03 | 4733 | mutex_unlock(&wq->mutex); |
a0a1a5fd | 4734 | |
e2dca7ad | 4735 | list_add_tail_rcu(&wq->list, &workqueues); |
a0a1a5fd | 4736 | |
68e13a67 | 4737 | mutex_unlock(&wq_pool_mutex); |
1537663f | 4738 | |
3af24433 | 4739 | return wq; |
d2c1d404 | 4740 | |
82efcab3 | 4741 | err_unreg_lockdep: |
009bb421 BVA |
4742 | wq_unregister_lockdep(wq); |
4743 | wq_free_lockdep(wq); | |
82efcab3 | 4744 | err_free_wq: |
6029a918 | 4745 | free_workqueue_attrs(wq->unbound_attrs); |
d2c1d404 TH |
4746 | kfree(wq); |
4747 | return NULL; | |
4748 | err_destroy: | |
4749 | destroy_workqueue(wq); | |
4690c4ab | 4750 | return NULL; |
3af24433 | 4751 | } |
669de8bd | 4752 | EXPORT_SYMBOL_GPL(alloc_workqueue); |
1da177e4 | 4753 | |
c29eb853 TH |
4754 | static bool pwq_busy(struct pool_workqueue *pwq) |
4755 | { | |
4756 | int i; | |
4757 | ||
4758 | for (i = 0; i < WORK_NR_COLORS; i++) | |
4759 | if (pwq->nr_in_flight[i]) | |
4760 | return true; | |
4761 | ||
4762 | if ((pwq != pwq->wq->dfl_pwq) && (pwq->refcnt > 1)) | |
4763 | return true; | |
f97a4a1a | 4764 | if (pwq->nr_active || !list_empty(&pwq->inactive_works)) |
c29eb853 TH |
4765 | return true; |
4766 | ||
4767 | return false; | |
4768 | } | |
4769 | ||
3af24433 ON |
4770 | /** |
4771 | * destroy_workqueue - safely terminate a workqueue | |
4772 | * @wq: target workqueue | |
4773 | * | |
4774 | * Safely destroy a workqueue. All work currently pending will be done first. | |
4775 | */ | |
4776 | void destroy_workqueue(struct workqueue_struct *wq) | |
4777 | { | |
49e3cf44 | 4778 | struct pool_workqueue *pwq; |
636b927e | 4779 | int cpu; |
3af24433 | 4780 | |
def98c84 TH |
4781 | /* |
4782 | * Remove it from sysfs first so that sanity check failure doesn't | |
4783 | * lead to sysfs name conflicts. | |
4784 | */ | |
4785 | workqueue_sysfs_unregister(wq); | |
4786 | ||
33e3f0a3 RC |
4787 | /* mark the workqueue destruction is in progress */ |
4788 | mutex_lock(&wq->mutex); | |
4789 | wq->flags |= __WQ_DESTROYING; | |
4790 | mutex_unlock(&wq->mutex); | |
4791 | ||
9c5a2ba7 TH |
4792 | /* drain it before proceeding with destruction */ |
4793 | drain_workqueue(wq); | |
c8efcc25 | 4794 | |
def98c84 TH |
4795 | /* kill rescuer, if sanity checks fail, leave it w/o rescuer */ |
4796 | if (wq->rescuer) { | |
4797 | struct worker *rescuer = wq->rescuer; | |
4798 | ||
4799 | /* this prevents new queueing */ | |
a9b8a985 | 4800 | raw_spin_lock_irq(&wq_mayday_lock); |
def98c84 | 4801 | wq->rescuer = NULL; |
a9b8a985 | 4802 | raw_spin_unlock_irq(&wq_mayday_lock); |
def98c84 TH |
4803 | |
4804 | /* rescuer will empty maydays list before exiting */ | |
4805 | kthread_stop(rescuer->task); | |
8efe1223 | 4806 | kfree(rescuer); |
def98c84 TH |
4807 | } |
4808 | ||
c29eb853 TH |
4809 | /* |
4810 | * Sanity checks - grab all the locks so that we wait for all | |
4811 | * in-flight operations which may do put_pwq(). | |
4812 | */ | |
4813 | mutex_lock(&wq_pool_mutex); | |
b09f4fd3 | 4814 | mutex_lock(&wq->mutex); |
49e3cf44 | 4815 | for_each_pwq(pwq, wq) { |
a9b8a985 | 4816 | raw_spin_lock_irq(&pwq->pool->lock); |
c29eb853 | 4817 | if (WARN_ON(pwq_busy(pwq))) { |
1d9a6159 KW |
4818 | pr_warn("%s: %s has the following busy pwq\n", |
4819 | __func__, wq->name); | |
c29eb853 | 4820 | show_pwq(pwq); |
a9b8a985 | 4821 | raw_spin_unlock_irq(&pwq->pool->lock); |
b09f4fd3 | 4822 | mutex_unlock(&wq->mutex); |
c29eb853 | 4823 | mutex_unlock(&wq_pool_mutex); |
55df0933 | 4824 | show_one_workqueue(wq); |
6183c009 | 4825 | return; |
76af4d93 | 4826 | } |
a9b8a985 | 4827 | raw_spin_unlock_irq(&pwq->pool->lock); |
6183c009 | 4828 | } |
b09f4fd3 | 4829 | mutex_unlock(&wq->mutex); |
6183c009 | 4830 | |
a0a1a5fd TH |
4831 | /* |
4832 | * wq list is used to freeze wq, remove from list after | |
4833 | * flushing is complete in case freeze races us. | |
4834 | */ | |
e2dca7ad | 4835 | list_del_rcu(&wq->list); |
68e13a67 | 4836 | mutex_unlock(&wq_pool_mutex); |
3af24433 | 4837 | |
636b927e TH |
4838 | /* |
4839 | * We're the sole accessor of @wq. Directly access cpu_pwq and dfl_pwq | |
4840 | * to put the base refs. @wq will be auto-destroyed from the last | |
4841 | * pwq_put. RCU read lock prevents @wq from going away from under us. | |
4842 | */ | |
4843 | rcu_read_lock(); | |
4c16bd32 | 4844 | |
636b927e TH |
4845 | for_each_possible_cpu(cpu) { |
4846 | pwq = rcu_access_pointer(*per_cpu_ptr(wq->cpu_pwq, cpu)); | |
4847 | RCU_INIT_POINTER(*per_cpu_ptr(wq->cpu_pwq, cpu), NULL); | |
dce90d47 | 4848 | put_pwq_unlocked(pwq); |
29c91e99 | 4849 | } |
636b927e TH |
4850 | |
4851 | put_pwq_unlocked(wq->dfl_pwq); | |
4852 | wq->dfl_pwq = NULL; | |
4853 | ||
4854 | rcu_read_unlock(); | |
3af24433 ON |
4855 | } |
4856 | EXPORT_SYMBOL_GPL(destroy_workqueue); | |
4857 | ||
dcd989cb TH |
4858 | /** |
4859 | * workqueue_set_max_active - adjust max_active of a workqueue | |
4860 | * @wq: target workqueue | |
4861 | * @max_active: new max_active value. | |
4862 | * | |
4863 | * Set max_active of @wq to @max_active. | |
4864 | * | |
4865 | * CONTEXT: | |
4866 | * Don't call from IRQ context. | |
4867 | */ | |
4868 | void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) | |
4869 | { | |
49e3cf44 | 4870 | struct pool_workqueue *pwq; |
dcd989cb | 4871 | |
8719dcea | 4872 | /* disallow meddling with max_active for ordered workqueues */ |
0a94efb5 | 4873 | if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) |
8719dcea TH |
4874 | return; |
4875 | ||
f3421797 | 4876 | max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); |
dcd989cb | 4877 | |
a357fc03 | 4878 | mutex_lock(&wq->mutex); |
dcd989cb | 4879 | |
0a94efb5 | 4880 | wq->flags &= ~__WQ_ORDERED; |
dcd989cb TH |
4881 | wq->saved_max_active = max_active; |
4882 | ||
699ce097 TH |
4883 | for_each_pwq(pwq, wq) |
4884 | pwq_adjust_max_active(pwq); | |
93981800 | 4885 | |
a357fc03 | 4886 | mutex_unlock(&wq->mutex); |
15316ba8 | 4887 | } |
dcd989cb | 4888 | EXPORT_SYMBOL_GPL(workqueue_set_max_active); |
15316ba8 | 4889 | |
27d4ee03 LW |
4890 | /** |
4891 | * current_work - retrieve %current task's work struct | |
4892 | * | |
4893 | * Determine if %current task is a workqueue worker and what it's working on. | |
4894 | * Useful to find out the context that the %current task is running in. | |
4895 | * | |
4896 | * Return: work struct if %current task is a workqueue worker, %NULL otherwise. | |
4897 | */ | |
4898 | struct work_struct *current_work(void) | |
4899 | { | |
4900 | struct worker *worker = current_wq_worker(); | |
4901 | ||
4902 | return worker ? worker->current_work : NULL; | |
4903 | } | |
4904 | EXPORT_SYMBOL(current_work); | |
4905 | ||
e6267616 TH |
4906 | /** |
4907 | * current_is_workqueue_rescuer - is %current workqueue rescuer? | |
4908 | * | |
4909 | * Determine whether %current is a workqueue rescuer. Can be used from | |
4910 | * work functions to determine whether it's being run off the rescuer task. | |
d185af30 YB |
4911 | * |
4912 | * Return: %true if %current is a workqueue rescuer. %false otherwise. | |
e6267616 TH |
4913 | */ |
4914 | bool current_is_workqueue_rescuer(void) | |
4915 | { | |
4916 | struct worker *worker = current_wq_worker(); | |
4917 | ||
6a092dfd | 4918 | return worker && worker->rescue_wq; |
e6267616 TH |
4919 | } |
4920 | ||
eef6a7d5 | 4921 | /** |
dcd989cb TH |
4922 | * workqueue_congested - test whether a workqueue is congested |
4923 | * @cpu: CPU in question | |
4924 | * @wq: target workqueue | |
eef6a7d5 | 4925 | * |
dcd989cb TH |
4926 | * Test whether @wq's cpu workqueue for @cpu is congested. There is |
4927 | * no synchronization around this function and the test result is | |
4928 | * unreliable and only useful as advisory hints or for debugging. | |
eef6a7d5 | 4929 | * |
d3251859 | 4930 | * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. |
636b927e TH |
4931 | * |
4932 | * With the exception of ordered workqueues, all workqueues have per-cpu | |
4933 | * pool_workqueues, each with its own congested state. A workqueue being | |
4934 | * congested on one CPU doesn't mean that the workqueue is contested on any | |
4935 | * other CPUs. | |
d3251859 | 4936 | * |
d185af30 | 4937 | * Return: |
dcd989cb | 4938 | * %true if congested, %false otherwise. |
eef6a7d5 | 4939 | */ |
d84ff051 | 4940 | bool workqueue_congested(int cpu, struct workqueue_struct *wq) |
1da177e4 | 4941 | { |
7fb98ea7 | 4942 | struct pool_workqueue *pwq; |
76af4d93 TH |
4943 | bool ret; |
4944 | ||
24acfb71 TG |
4945 | rcu_read_lock(); |
4946 | preempt_disable(); | |
7fb98ea7 | 4947 | |
d3251859 TH |
4948 | if (cpu == WORK_CPU_UNBOUND) |
4949 | cpu = smp_processor_id(); | |
4950 | ||
636b927e | 4951 | pwq = *per_cpu_ptr(wq->cpu_pwq, cpu); |
f97a4a1a | 4952 | ret = !list_empty(&pwq->inactive_works); |
636b927e | 4953 | |
24acfb71 TG |
4954 | preempt_enable(); |
4955 | rcu_read_unlock(); | |
76af4d93 TH |
4956 | |
4957 | return ret; | |
1da177e4 | 4958 | } |
dcd989cb | 4959 | EXPORT_SYMBOL_GPL(workqueue_congested); |
1da177e4 | 4960 | |
dcd989cb TH |
4961 | /** |
4962 | * work_busy - test whether a work is currently pending or running | |
4963 | * @work: the work to be tested | |
4964 | * | |
4965 | * Test whether @work is currently pending or running. There is no | |
4966 | * synchronization around this function and the test result is | |
4967 | * unreliable and only useful as advisory hints or for debugging. | |
dcd989cb | 4968 | * |
d185af30 | 4969 | * Return: |
dcd989cb TH |
4970 | * OR'd bitmask of WORK_BUSY_* bits. |
4971 | */ | |
4972 | unsigned int work_busy(struct work_struct *work) | |
1da177e4 | 4973 | { |
fa1b54e6 | 4974 | struct worker_pool *pool; |
dcd989cb TH |
4975 | unsigned long flags; |
4976 | unsigned int ret = 0; | |
1da177e4 | 4977 | |
dcd989cb TH |
4978 | if (work_pending(work)) |
4979 | ret |= WORK_BUSY_PENDING; | |
1da177e4 | 4980 | |
24acfb71 | 4981 | rcu_read_lock(); |
fa1b54e6 | 4982 | pool = get_work_pool(work); |
038366c5 | 4983 | if (pool) { |
a9b8a985 | 4984 | raw_spin_lock_irqsave(&pool->lock, flags); |
038366c5 LJ |
4985 | if (find_worker_executing_work(pool, work)) |
4986 | ret |= WORK_BUSY_RUNNING; | |
a9b8a985 | 4987 | raw_spin_unlock_irqrestore(&pool->lock, flags); |
038366c5 | 4988 | } |
24acfb71 | 4989 | rcu_read_unlock(); |
1da177e4 | 4990 | |
dcd989cb | 4991 | return ret; |
1da177e4 | 4992 | } |
dcd989cb | 4993 | EXPORT_SYMBOL_GPL(work_busy); |
1da177e4 | 4994 | |
3d1cb205 TH |
4995 | /** |
4996 | * set_worker_desc - set description for the current work item | |
4997 | * @fmt: printf-style format string | |
4998 | * @...: arguments for the format string | |
4999 | * | |
5000 | * This function can be called by a running work function to describe what | |
5001 | * the work item is about. If the worker task gets dumped, this | |
5002 | * information will be printed out together to help debugging. The | |
5003 | * description can be at most WORKER_DESC_LEN including the trailing '\0'. | |
5004 | */ | |
5005 | void set_worker_desc(const char *fmt, ...) | |
5006 | { | |
5007 | struct worker *worker = current_wq_worker(); | |
5008 | va_list args; | |
5009 | ||
5010 | if (worker) { | |
5011 | va_start(args, fmt); | |
5012 | vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); | |
5013 | va_end(args); | |
3d1cb205 TH |
5014 | } |
5015 | } | |
5c750d58 | 5016 | EXPORT_SYMBOL_GPL(set_worker_desc); |
3d1cb205 TH |
5017 | |
5018 | /** | |
5019 | * print_worker_info - print out worker information and description | |
5020 | * @log_lvl: the log level to use when printing | |
5021 | * @task: target task | |
5022 | * | |
5023 | * If @task is a worker and currently executing a work item, print out the | |
5024 | * name of the workqueue being serviced and worker description set with | |
5025 | * set_worker_desc() by the currently executing work item. | |
5026 | * | |
5027 | * This function can be safely called on any task as long as the | |
5028 | * task_struct itself is accessible. While safe, this function isn't | |
5029 | * synchronized and may print out mixups or garbages of limited length. | |
5030 | */ | |
5031 | void print_worker_info(const char *log_lvl, struct task_struct *task) | |
5032 | { | |
5033 | work_func_t *fn = NULL; | |
5034 | char name[WQ_NAME_LEN] = { }; | |
5035 | char desc[WORKER_DESC_LEN] = { }; | |
5036 | struct pool_workqueue *pwq = NULL; | |
5037 | struct workqueue_struct *wq = NULL; | |
3d1cb205 TH |
5038 | struct worker *worker; |
5039 | ||
5040 | if (!(task->flags & PF_WQ_WORKER)) | |
5041 | return; | |
5042 | ||
5043 | /* | |
5044 | * This function is called without any synchronization and @task | |
5045 | * could be in any state. Be careful with dereferences. | |
5046 | */ | |
e700591a | 5047 | worker = kthread_probe_data(task); |
3d1cb205 TH |
5048 | |
5049 | /* | |
8bf89593 TH |
5050 | * Carefully copy the associated workqueue's workfn, name and desc. |
5051 | * Keep the original last '\0' in case the original is garbage. | |
3d1cb205 | 5052 | */ |
fe557319 CH |
5053 | copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn)); |
5054 | copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq)); | |
5055 | copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq)); | |
5056 | copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1); | |
5057 | copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1); | |
3d1cb205 TH |
5058 | |
5059 | if (fn || name[0] || desc[0]) { | |
d75f773c | 5060 | printk("%sWorkqueue: %s %ps", log_lvl, name, fn); |
8bf89593 | 5061 | if (strcmp(name, desc)) |
3d1cb205 TH |
5062 | pr_cont(" (%s)", desc); |
5063 | pr_cont("\n"); | |
5064 | } | |
5065 | } | |
5066 | ||
3494fc30 TH |
5067 | static void pr_cont_pool_info(struct worker_pool *pool) |
5068 | { | |
5069 | pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask); | |
5070 | if (pool->node != NUMA_NO_NODE) | |
5071 | pr_cont(" node=%d", pool->node); | |
5072 | pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice); | |
5073 | } | |
5074 | ||
c76feb0d PM |
5075 | struct pr_cont_work_struct { |
5076 | bool comma; | |
5077 | work_func_t func; | |
5078 | long ctr; | |
5079 | }; | |
5080 | ||
5081 | static void pr_cont_work_flush(bool comma, work_func_t func, struct pr_cont_work_struct *pcwsp) | |
5082 | { | |
5083 | if (!pcwsp->ctr) | |
5084 | goto out_record; | |
5085 | if (func == pcwsp->func) { | |
5086 | pcwsp->ctr++; | |
5087 | return; | |
5088 | } | |
5089 | if (pcwsp->ctr == 1) | |
5090 | pr_cont("%s %ps", pcwsp->comma ? "," : "", pcwsp->func); | |
5091 | else | |
5092 | pr_cont("%s %ld*%ps", pcwsp->comma ? "," : "", pcwsp->ctr, pcwsp->func); | |
5093 | pcwsp->ctr = 0; | |
5094 | out_record: | |
5095 | if ((long)func == -1L) | |
5096 | return; | |
5097 | pcwsp->comma = comma; | |
5098 | pcwsp->func = func; | |
5099 | pcwsp->ctr = 1; | |
5100 | } | |
5101 | ||
5102 | static void pr_cont_work(bool comma, struct work_struct *work, struct pr_cont_work_struct *pcwsp) | |
3494fc30 TH |
5103 | { |
5104 | if (work->func == wq_barrier_func) { | |
5105 | struct wq_barrier *barr; | |
5106 | ||
5107 | barr = container_of(work, struct wq_barrier, work); | |
5108 | ||
c76feb0d | 5109 | pr_cont_work_flush(comma, (work_func_t)-1, pcwsp); |
3494fc30 TH |
5110 | pr_cont("%s BAR(%d)", comma ? "," : "", |
5111 | task_pid_nr(barr->task)); | |
5112 | } else { | |
c76feb0d PM |
5113 | if (!comma) |
5114 | pr_cont_work_flush(comma, (work_func_t)-1, pcwsp); | |
5115 | pr_cont_work_flush(comma, work->func, pcwsp); | |
3494fc30 TH |
5116 | } |
5117 | } | |
5118 | ||
5119 | static void show_pwq(struct pool_workqueue *pwq) | |
5120 | { | |
c76feb0d | 5121 | struct pr_cont_work_struct pcws = { .ctr = 0, }; |
3494fc30 TH |
5122 | struct worker_pool *pool = pwq->pool; |
5123 | struct work_struct *work; | |
5124 | struct worker *worker; | |
5125 | bool has_in_flight = false, has_pending = false; | |
5126 | int bkt; | |
5127 | ||
5128 | pr_info(" pwq %d:", pool->id); | |
5129 | pr_cont_pool_info(pool); | |
5130 | ||
e66b39af TH |
5131 | pr_cont(" active=%d/%d refcnt=%d%s\n", |
5132 | pwq->nr_active, pwq->max_active, pwq->refcnt, | |
3494fc30 TH |
5133 | !list_empty(&pwq->mayday_node) ? " MAYDAY" : ""); |
5134 | ||
5135 | hash_for_each(pool->busy_hash, bkt, worker, hentry) { | |
5136 | if (worker->current_pwq == pwq) { | |
5137 | has_in_flight = true; | |
5138 | break; | |
5139 | } | |
5140 | } | |
5141 | if (has_in_flight) { | |
5142 | bool comma = false; | |
5143 | ||
5144 | pr_info(" in-flight:"); | |
5145 | hash_for_each(pool->busy_hash, bkt, worker, hentry) { | |
5146 | if (worker->current_pwq != pwq) | |
5147 | continue; | |
5148 | ||
d75f773c | 5149 | pr_cont("%s %d%s:%ps", comma ? "," : "", |
3494fc30 | 5150 | task_pid_nr(worker->task), |
30ae2fc0 | 5151 | worker->rescue_wq ? "(RESCUER)" : "", |
3494fc30 TH |
5152 | worker->current_func); |
5153 | list_for_each_entry(work, &worker->scheduled, entry) | |
c76feb0d PM |
5154 | pr_cont_work(false, work, &pcws); |
5155 | pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); | |
3494fc30 TH |
5156 | comma = true; |
5157 | } | |
5158 | pr_cont("\n"); | |
5159 | } | |
5160 | ||
5161 | list_for_each_entry(work, &pool->worklist, entry) { | |
5162 | if (get_work_pwq(work) == pwq) { | |
5163 | has_pending = true; | |
5164 | break; | |
5165 | } | |
5166 | } | |
5167 | if (has_pending) { | |
5168 | bool comma = false; | |
5169 | ||
5170 | pr_info(" pending:"); | |
5171 | list_for_each_entry(work, &pool->worklist, entry) { | |
5172 | if (get_work_pwq(work) != pwq) | |
5173 | continue; | |
5174 | ||
c76feb0d | 5175 | pr_cont_work(comma, work, &pcws); |
3494fc30 TH |
5176 | comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); |
5177 | } | |
c76feb0d | 5178 | pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); |
3494fc30 TH |
5179 | pr_cont("\n"); |
5180 | } | |
5181 | ||
f97a4a1a | 5182 | if (!list_empty(&pwq->inactive_works)) { |
3494fc30 TH |
5183 | bool comma = false; |
5184 | ||
f97a4a1a LJ |
5185 | pr_info(" inactive:"); |
5186 | list_for_each_entry(work, &pwq->inactive_works, entry) { | |
c76feb0d | 5187 | pr_cont_work(comma, work, &pcws); |
3494fc30 TH |
5188 | comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); |
5189 | } | |
c76feb0d | 5190 | pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); |
3494fc30 TH |
5191 | pr_cont("\n"); |
5192 | } | |
5193 | } | |
5194 | ||
5195 | /** | |
55df0933 IK |
5196 | * show_one_workqueue - dump state of specified workqueue |
5197 | * @wq: workqueue whose state will be printed | |
3494fc30 | 5198 | */ |
55df0933 | 5199 | void show_one_workqueue(struct workqueue_struct *wq) |
3494fc30 | 5200 | { |
55df0933 IK |
5201 | struct pool_workqueue *pwq; |
5202 | bool idle = true; | |
3494fc30 | 5203 | unsigned long flags; |
3494fc30 | 5204 | |
55df0933 IK |
5205 | for_each_pwq(pwq, wq) { |
5206 | if (pwq->nr_active || !list_empty(&pwq->inactive_works)) { | |
5207 | idle = false; | |
5208 | break; | |
3494fc30 | 5209 | } |
55df0933 IK |
5210 | } |
5211 | if (idle) /* Nothing to print for idle workqueue */ | |
5212 | return; | |
3494fc30 | 5213 | |
55df0933 | 5214 | pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags); |
3494fc30 | 5215 | |
55df0933 IK |
5216 | for_each_pwq(pwq, wq) { |
5217 | raw_spin_lock_irqsave(&pwq->pool->lock, flags); | |
5218 | if (pwq->nr_active || !list_empty(&pwq->inactive_works)) { | |
62635ea8 | 5219 | /* |
55df0933 IK |
5220 | * Defer printing to avoid deadlocks in console |
5221 | * drivers that queue work while holding locks | |
5222 | * also taken in their write paths. | |
62635ea8 | 5223 | */ |
55df0933 IK |
5224 | printk_deferred_enter(); |
5225 | show_pwq(pwq); | |
5226 | printk_deferred_exit(); | |
3494fc30 | 5227 | } |
55df0933 | 5228 | raw_spin_unlock_irqrestore(&pwq->pool->lock, flags); |
62635ea8 SS |
5229 | /* |
5230 | * We could be printing a lot from atomic context, e.g. | |
55df0933 | 5231 | * sysrq-t -> show_all_workqueues(). Avoid triggering |
62635ea8 SS |
5232 | * hard lockup. |
5233 | */ | |
5234 | touch_nmi_watchdog(); | |
3494fc30 TH |
5235 | } |
5236 | ||
55df0933 IK |
5237 | } |
5238 | ||
5239 | /** | |
5240 | * show_one_worker_pool - dump state of specified worker pool | |
5241 | * @pool: worker pool whose state will be printed | |
5242 | */ | |
5243 | static void show_one_worker_pool(struct worker_pool *pool) | |
5244 | { | |
5245 | struct worker *worker; | |
5246 | bool first = true; | |
5247 | unsigned long flags; | |
335a42eb | 5248 | unsigned long hung = 0; |
55df0933 IK |
5249 | |
5250 | raw_spin_lock_irqsave(&pool->lock, flags); | |
5251 | if (pool->nr_workers == pool->nr_idle) | |
5252 | goto next_pool; | |
335a42eb PM |
5253 | |
5254 | /* How long the first pending work is waiting for a worker. */ | |
5255 | if (!list_empty(&pool->worklist)) | |
5256 | hung = jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000; | |
5257 | ||
55df0933 IK |
5258 | /* |
5259 | * Defer printing to avoid deadlocks in console drivers that | |
5260 | * queue work while holding locks also taken in their write | |
5261 | * paths. | |
5262 | */ | |
5263 | printk_deferred_enter(); | |
5264 | pr_info("pool %d:", pool->id); | |
5265 | pr_cont_pool_info(pool); | |
335a42eb | 5266 | pr_cont(" hung=%lus workers=%d", hung, pool->nr_workers); |
55df0933 IK |
5267 | if (pool->manager) |
5268 | pr_cont(" manager: %d", | |
5269 | task_pid_nr(pool->manager->task)); | |
5270 | list_for_each_entry(worker, &pool->idle_list, entry) { | |
5271 | pr_cont(" %s%d", first ? "idle: " : "", | |
5272 | task_pid_nr(worker->task)); | |
5273 | first = false; | |
5274 | } | |
5275 | pr_cont("\n"); | |
5276 | printk_deferred_exit(); | |
5277 | next_pool: | |
5278 | raw_spin_unlock_irqrestore(&pool->lock, flags); | |
5279 | /* | |
5280 | * We could be printing a lot from atomic context, e.g. | |
5281 | * sysrq-t -> show_all_workqueues(). Avoid triggering | |
5282 | * hard lockup. | |
5283 | */ | |
5284 | touch_nmi_watchdog(); | |
5285 | ||
5286 | } | |
5287 | ||
5288 | /** | |
5289 | * show_all_workqueues - dump workqueue state | |
5290 | * | |
704bc669 | 5291 | * Called from a sysrq handler and prints out all busy workqueues and pools. |
55df0933 IK |
5292 | */ |
5293 | void show_all_workqueues(void) | |
5294 | { | |
5295 | struct workqueue_struct *wq; | |
5296 | struct worker_pool *pool; | |
5297 | int pi; | |
5298 | ||
5299 | rcu_read_lock(); | |
5300 | ||
5301 | pr_info("Showing busy workqueues and worker pools:\n"); | |
5302 | ||
5303 | list_for_each_entry_rcu(wq, &workqueues, list) | |
5304 | show_one_workqueue(wq); | |
5305 | ||
5306 | for_each_pool(pool, pi) | |
5307 | show_one_worker_pool(pool); | |
5308 | ||
24acfb71 | 5309 | rcu_read_unlock(); |
3494fc30 TH |
5310 | } |
5311 | ||
704bc669 JL |
5312 | /** |
5313 | * show_freezable_workqueues - dump freezable workqueue state | |
5314 | * | |
5315 | * Called from try_to_freeze_tasks() and prints out all freezable workqueues | |
5316 | * still busy. | |
5317 | */ | |
5318 | void show_freezable_workqueues(void) | |
5319 | { | |
5320 | struct workqueue_struct *wq; | |
5321 | ||
5322 | rcu_read_lock(); | |
5323 | ||
5324 | pr_info("Showing freezable workqueues that are still busy:\n"); | |
5325 | ||
5326 | list_for_each_entry_rcu(wq, &workqueues, list) { | |
5327 | if (!(wq->flags & WQ_FREEZABLE)) | |
5328 | continue; | |
5329 | show_one_workqueue(wq); | |
5330 | } | |
5331 | ||
5332 | rcu_read_unlock(); | |
5333 | } | |
5334 | ||
6b59808b TH |
5335 | /* used to show worker information through /proc/PID/{comm,stat,status} */ |
5336 | void wq_worker_comm(char *buf, size_t size, struct task_struct *task) | |
5337 | { | |
6b59808b TH |
5338 | int off; |
5339 | ||
5340 | /* always show the actual comm */ | |
5341 | off = strscpy(buf, task->comm, size); | |
5342 | if (off < 0) | |
5343 | return; | |
5344 | ||
197f6acc | 5345 | /* stabilize PF_WQ_WORKER and worker pool association */ |
6b59808b TH |
5346 | mutex_lock(&wq_pool_attach_mutex); |
5347 | ||
197f6acc TH |
5348 | if (task->flags & PF_WQ_WORKER) { |
5349 | struct worker *worker = kthread_data(task); | |
5350 | struct worker_pool *pool = worker->pool; | |
6b59808b | 5351 | |
197f6acc | 5352 | if (pool) { |
a9b8a985 | 5353 | raw_spin_lock_irq(&pool->lock); |
197f6acc TH |
5354 | /* |
5355 | * ->desc tracks information (wq name or | |
5356 | * set_worker_desc()) for the latest execution. If | |
5357 | * current, prepend '+', otherwise '-'. | |
5358 | */ | |
5359 | if (worker->desc[0] != '\0') { | |
5360 | if (worker->current_work) | |
5361 | scnprintf(buf + off, size - off, "+%s", | |
5362 | worker->desc); | |
5363 | else | |
5364 | scnprintf(buf + off, size - off, "-%s", | |
5365 | worker->desc); | |
5366 | } | |
a9b8a985 | 5367 | raw_spin_unlock_irq(&pool->lock); |
6b59808b | 5368 | } |
6b59808b TH |
5369 | } |
5370 | ||
5371 | mutex_unlock(&wq_pool_attach_mutex); | |
5372 | } | |
5373 | ||
66448bc2 MM |
5374 | #ifdef CONFIG_SMP |
5375 | ||
db7bccf4 TH |
5376 | /* |
5377 | * CPU hotplug. | |
5378 | * | |
e22bee78 | 5379 | * There are two challenges in supporting CPU hotplug. Firstly, there |
112202d9 | 5380 | * are a lot of assumptions on strong associations among work, pwq and |
706026c2 | 5381 | * pool which make migrating pending and scheduled works very |
e22bee78 | 5382 | * difficult to implement without impacting hot paths. Secondly, |
94cf58bb | 5383 | * worker pools serve mix of short, long and very long running works making |
e22bee78 TH |
5384 | * blocked draining impractical. |
5385 | * | |
24647570 | 5386 | * This is solved by allowing the pools to be disassociated from the CPU |
628c78e7 TH |
5387 | * running as an unbound one and allowing it to be reattached later if the |
5388 | * cpu comes back online. | |
db7bccf4 | 5389 | */ |
1da177e4 | 5390 | |
e8b3f8db | 5391 | static void unbind_workers(int cpu) |
3af24433 | 5392 | { |
4ce62e9e | 5393 | struct worker_pool *pool; |
db7bccf4 | 5394 | struct worker *worker; |
3af24433 | 5395 | |
f02ae73a | 5396 | for_each_cpu_worker_pool(pool, cpu) { |
1258fae7 | 5397 | mutex_lock(&wq_pool_attach_mutex); |
a9b8a985 | 5398 | raw_spin_lock_irq(&pool->lock); |
3af24433 | 5399 | |
94cf58bb | 5400 | /* |
92f9c5c4 | 5401 | * We've blocked all attach/detach operations. Make all workers |
94cf58bb | 5402 | * unbound and set DISASSOCIATED. Before this, all workers |
11b45b0b | 5403 | * must be on the cpu. After this, they may become diasporas. |
b4ac9384 LJ |
5404 | * And the preemption disabled section in their sched callbacks |
5405 | * are guaranteed to see WORKER_UNBOUND since the code here | |
5406 | * is on the same cpu. | |
94cf58bb | 5407 | */ |
da028469 | 5408 | for_each_pool_worker(worker, pool) |
c9e7cf27 | 5409 | worker->flags |= WORKER_UNBOUND; |
06ba38a9 | 5410 | |
24647570 | 5411 | pool->flags |= POOL_DISASSOCIATED; |
f2d5a0ee | 5412 | |
eb283428 | 5413 | /* |
989442d7 LJ |
5414 | * The handling of nr_running in sched callbacks are disabled |
5415 | * now. Zap nr_running. After this, nr_running stays zero and | |
5416 | * need_more_worker() and keep_working() are always true as | |
5417 | * long as the worklist is not empty. This pool now behaves as | |
5418 | * an unbound (in terms of concurrency management) pool which | |
eb283428 LJ |
5419 | * are served by workers tied to the pool. |
5420 | */ | |
bc35f7ef | 5421 | pool->nr_running = 0; |
eb283428 LJ |
5422 | |
5423 | /* | |
5424 | * With concurrency management just turned off, a busy | |
5425 | * worker blocking could lead to lengthy stalls. Kick off | |
5426 | * unbound chain execution of currently pending work items. | |
5427 | */ | |
0219a352 | 5428 | kick_pool(pool); |
989442d7 | 5429 | |
a9b8a985 | 5430 | raw_spin_unlock_irq(&pool->lock); |
989442d7 | 5431 | |
793777bc VS |
5432 | for_each_pool_worker(worker, pool) |
5433 | unbind_worker(worker); | |
989442d7 LJ |
5434 | |
5435 | mutex_unlock(&wq_pool_attach_mutex); | |
eb283428 | 5436 | } |
3af24433 | 5437 | } |
3af24433 | 5438 | |
bd7c089e TH |
5439 | /** |
5440 | * rebind_workers - rebind all workers of a pool to the associated CPU | |
5441 | * @pool: pool of interest | |
5442 | * | |
a9ab775b | 5443 | * @pool->cpu is coming online. Rebind all workers to the CPU. |
bd7c089e TH |
5444 | */ |
5445 | static void rebind_workers(struct worker_pool *pool) | |
5446 | { | |
a9ab775b | 5447 | struct worker *worker; |
bd7c089e | 5448 | |
1258fae7 | 5449 | lockdep_assert_held(&wq_pool_attach_mutex); |
bd7c089e | 5450 | |
a9ab775b TH |
5451 | /* |
5452 | * Restore CPU affinity of all workers. As all idle workers should | |
5453 | * be on the run-queue of the associated CPU before any local | |
402dd89d | 5454 | * wake-ups for concurrency management happen, restore CPU affinity |
a9ab775b TH |
5455 | * of all workers first and then clear UNBOUND. As we're called |
5456 | * from CPU_ONLINE, the following shouldn't fail. | |
5457 | */ | |
c63a2e52 VS |
5458 | for_each_pool_worker(worker, pool) { |
5459 | kthread_set_per_cpu(worker->task, pool->cpu); | |
5460 | WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, | |
9546b29e | 5461 | pool_allowed_cpus(pool)) < 0); |
c63a2e52 | 5462 | } |
bd7c089e | 5463 | |
a9b8a985 | 5464 | raw_spin_lock_irq(&pool->lock); |
f7c17d26 | 5465 | |
3de5e884 | 5466 | pool->flags &= ~POOL_DISASSOCIATED; |
bd7c089e | 5467 | |
da028469 | 5468 | for_each_pool_worker(worker, pool) { |
a9ab775b | 5469 | unsigned int worker_flags = worker->flags; |
bd7c089e | 5470 | |
a9ab775b TH |
5471 | /* |
5472 | * We want to clear UNBOUND but can't directly call | |
5473 | * worker_clr_flags() or adjust nr_running. Atomically | |
5474 | * replace UNBOUND with another NOT_RUNNING flag REBOUND. | |
5475 | * @worker will clear REBOUND using worker_clr_flags() when | |
5476 | * it initiates the next execution cycle thus restoring | |
5477 | * concurrency management. Note that when or whether | |
5478 | * @worker clears REBOUND doesn't affect correctness. | |
5479 | * | |
c95491ed | 5480 | * WRITE_ONCE() is necessary because @worker->flags may be |
a9ab775b | 5481 | * tested without holding any lock in |
6d25be57 | 5482 | * wq_worker_running(). Without it, NOT_RUNNING test may |
a9ab775b TH |
5483 | * fail incorrectly leading to premature concurrency |
5484 | * management operations. | |
5485 | */ | |
5486 | WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); | |
5487 | worker_flags |= WORKER_REBOUND; | |
5488 | worker_flags &= ~WORKER_UNBOUND; | |
c95491ed | 5489 | WRITE_ONCE(worker->flags, worker_flags); |
bd7c089e | 5490 | } |
a9ab775b | 5491 | |
a9b8a985 | 5492 | raw_spin_unlock_irq(&pool->lock); |
bd7c089e TH |
5493 | } |
5494 | ||
7dbc725e TH |
5495 | /** |
5496 | * restore_unbound_workers_cpumask - restore cpumask of unbound workers | |
5497 | * @pool: unbound pool of interest | |
5498 | * @cpu: the CPU which is coming up | |
5499 | * | |
5500 | * An unbound pool may end up with a cpumask which doesn't have any online | |
5501 | * CPUs. When a worker of such pool get scheduled, the scheduler resets | |
5502 | * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any | |
5503 | * online CPU before, cpus_allowed of all its workers should be restored. | |
5504 | */ | |
5505 | static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) | |
5506 | { | |
5507 | static cpumask_t cpumask; | |
5508 | struct worker *worker; | |
7dbc725e | 5509 | |
1258fae7 | 5510 | lockdep_assert_held(&wq_pool_attach_mutex); |
7dbc725e TH |
5511 | |
5512 | /* is @cpu allowed for @pool? */ | |
5513 | if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) | |
5514 | return; | |
5515 | ||
7dbc725e | 5516 | cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); |
7dbc725e TH |
5517 | |
5518 | /* as we're called from CPU_ONLINE, the following shouldn't fail */ | |
da028469 | 5519 | for_each_pool_worker(worker, pool) |
d945b5e9 | 5520 | WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0); |
7dbc725e TH |
5521 | } |
5522 | ||
7ee681b2 TG |
5523 | int workqueue_prepare_cpu(unsigned int cpu) |
5524 | { | |
5525 | struct worker_pool *pool; | |
5526 | ||
5527 | for_each_cpu_worker_pool(pool, cpu) { | |
5528 | if (pool->nr_workers) | |
5529 | continue; | |
5530 | if (!create_worker(pool)) | |
5531 | return -ENOMEM; | |
5532 | } | |
5533 | return 0; | |
5534 | } | |
5535 | ||
5536 | int workqueue_online_cpu(unsigned int cpu) | |
3af24433 | 5537 | { |
4ce62e9e | 5538 | struct worker_pool *pool; |
4c16bd32 | 5539 | struct workqueue_struct *wq; |
7dbc725e | 5540 | int pi; |
3ce63377 | 5541 | |
7ee681b2 | 5542 | mutex_lock(&wq_pool_mutex); |
7dbc725e | 5543 | |
7ee681b2 | 5544 | for_each_pool(pool, pi) { |
1258fae7 | 5545 | mutex_lock(&wq_pool_attach_mutex); |
94cf58bb | 5546 | |
7ee681b2 TG |
5547 | if (pool->cpu == cpu) |
5548 | rebind_workers(pool); | |
5549 | else if (pool->cpu < 0) | |
5550 | restore_unbound_workers_cpumask(pool, cpu); | |
94cf58bb | 5551 | |
1258fae7 | 5552 | mutex_unlock(&wq_pool_attach_mutex); |
7ee681b2 | 5553 | } |
6ba94429 | 5554 | |
fef59c9c | 5555 | /* update pod affinity of unbound workqueues */ |
4cbfd3de | 5556 | list_for_each_entry(wq, &workqueues, list) { |
84193c07 TH |
5557 | struct workqueue_attrs *attrs = wq->unbound_attrs; |
5558 | ||
5559 | if (attrs) { | |
5560 | const struct wq_pod_type *pt = wqattrs_pod_type(attrs); | |
5561 | int tcpu; | |
4cbfd3de | 5562 | |
84193c07 | 5563 | for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]]) |
fef59c9c | 5564 | wq_update_pod(wq, tcpu, cpu, true); |
4cbfd3de TH |
5565 | } |
5566 | } | |
6ba94429 | 5567 | |
7ee681b2 TG |
5568 | mutex_unlock(&wq_pool_mutex); |
5569 | return 0; | |
6ba94429 FW |
5570 | } |
5571 | ||
7ee681b2 | 5572 | int workqueue_offline_cpu(unsigned int cpu) |
6ba94429 | 5573 | { |
6ba94429 FW |
5574 | struct workqueue_struct *wq; |
5575 | ||
7ee681b2 | 5576 | /* unbinding per-cpu workers should happen on the local CPU */ |
e8b3f8db LJ |
5577 | if (WARN_ON(cpu != smp_processor_id())) |
5578 | return -1; | |
5579 | ||
5580 | unbind_workers(cpu); | |
7ee681b2 | 5581 | |
fef59c9c | 5582 | /* update pod affinity of unbound workqueues */ |
7ee681b2 | 5583 | mutex_lock(&wq_pool_mutex); |
4cbfd3de | 5584 | list_for_each_entry(wq, &workqueues, list) { |
84193c07 TH |
5585 | struct workqueue_attrs *attrs = wq->unbound_attrs; |
5586 | ||
5587 | if (attrs) { | |
5588 | const struct wq_pod_type *pt = wqattrs_pod_type(attrs); | |
5589 | int tcpu; | |
4cbfd3de | 5590 | |
84193c07 | 5591 | for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]]) |
fef59c9c | 5592 | wq_update_pod(wq, tcpu, cpu, false); |
4cbfd3de TH |
5593 | } |
5594 | } | |
7ee681b2 TG |
5595 | mutex_unlock(&wq_pool_mutex); |
5596 | ||
7ee681b2 | 5597 | return 0; |
6ba94429 FW |
5598 | } |
5599 | ||
6ba94429 FW |
5600 | struct work_for_cpu { |
5601 | struct work_struct work; | |
5602 | long (*fn)(void *); | |
5603 | void *arg; | |
5604 | long ret; | |
5605 | }; | |
5606 | ||
5607 | static void work_for_cpu_fn(struct work_struct *work) | |
5608 | { | |
5609 | struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); | |
5610 | ||
5611 | wfc->ret = wfc->fn(wfc->arg); | |
5612 | } | |
5613 | ||
5614 | /** | |
22aceb31 | 5615 | * work_on_cpu - run a function in thread context on a particular cpu |
6ba94429 FW |
5616 | * @cpu: the cpu to run on |
5617 | * @fn: the function to run | |
5618 | * @arg: the function arg | |
5619 | * | |
5620 | * It is up to the caller to ensure that the cpu doesn't go offline. | |
5621 | * The caller must not hold any locks which would prevent @fn from completing. | |
5622 | * | |
5623 | * Return: The value @fn returns. | |
5624 | */ | |
5625 | long work_on_cpu(int cpu, long (*fn)(void *), void *arg) | |
5626 | { | |
5627 | struct work_for_cpu wfc = { .fn = fn, .arg = arg }; | |
5628 | ||
5629 | INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); | |
5630 | schedule_work_on(cpu, &wfc.work); | |
5631 | flush_work(&wfc.work); | |
5632 | destroy_work_on_stack(&wfc.work); | |
5633 | return wfc.ret; | |
5634 | } | |
5635 | EXPORT_SYMBOL_GPL(work_on_cpu); | |
0e8d6a93 TG |
5636 | |
5637 | /** | |
5638 | * work_on_cpu_safe - run a function in thread context on a particular cpu | |
5639 | * @cpu: the cpu to run on | |
5640 | * @fn: the function to run | |
5641 | * @arg: the function argument | |
5642 | * | |
5643 | * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold | |
5644 | * any locks which would prevent @fn from completing. | |
5645 | * | |
5646 | * Return: The value @fn returns. | |
5647 | */ | |
5648 | long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg) | |
5649 | { | |
5650 | long ret = -ENODEV; | |
5651 | ||
ffd8bea8 | 5652 | cpus_read_lock(); |
0e8d6a93 TG |
5653 | if (cpu_online(cpu)) |
5654 | ret = work_on_cpu(cpu, fn, arg); | |
ffd8bea8 | 5655 | cpus_read_unlock(); |
0e8d6a93 TG |
5656 | return ret; |
5657 | } | |
5658 | EXPORT_SYMBOL_GPL(work_on_cpu_safe); | |
6ba94429 FW |
5659 | #endif /* CONFIG_SMP */ |
5660 | ||
5661 | #ifdef CONFIG_FREEZER | |
5662 | ||
5663 | /** | |
5664 | * freeze_workqueues_begin - begin freezing workqueues | |
5665 | * | |
5666 | * Start freezing workqueues. After this function returns, all freezable | |
f97a4a1a | 5667 | * workqueues will queue new works to their inactive_works list instead of |
6ba94429 FW |
5668 | * pool->worklist. |
5669 | * | |
5670 | * CONTEXT: | |
5671 | * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. | |
5672 | */ | |
5673 | void freeze_workqueues_begin(void) | |
5674 | { | |
5675 | struct workqueue_struct *wq; | |
5676 | struct pool_workqueue *pwq; | |
5677 | ||
5678 | mutex_lock(&wq_pool_mutex); | |
5679 | ||
5680 | WARN_ON_ONCE(workqueue_freezing); | |
5681 | workqueue_freezing = true; | |
5682 | ||
5683 | list_for_each_entry(wq, &workqueues, list) { | |
5684 | mutex_lock(&wq->mutex); | |
5685 | for_each_pwq(pwq, wq) | |
5686 | pwq_adjust_max_active(pwq); | |
5687 | mutex_unlock(&wq->mutex); | |
5688 | } | |
5689 | ||
5690 | mutex_unlock(&wq_pool_mutex); | |
5691 | } | |
5692 | ||
5693 | /** | |
5694 | * freeze_workqueues_busy - are freezable workqueues still busy? | |
5695 | * | |
5696 | * Check whether freezing is complete. This function must be called | |
5697 | * between freeze_workqueues_begin() and thaw_workqueues(). | |
5698 | * | |
5699 | * CONTEXT: | |
5700 | * Grabs and releases wq_pool_mutex. | |
5701 | * | |
5702 | * Return: | |
5703 | * %true if some freezable workqueues are still busy. %false if freezing | |
5704 | * is complete. | |
5705 | */ | |
5706 | bool freeze_workqueues_busy(void) | |
5707 | { | |
5708 | bool busy = false; | |
5709 | struct workqueue_struct *wq; | |
5710 | struct pool_workqueue *pwq; | |
5711 | ||
5712 | mutex_lock(&wq_pool_mutex); | |
5713 | ||
5714 | WARN_ON_ONCE(!workqueue_freezing); | |
5715 | ||
5716 | list_for_each_entry(wq, &workqueues, list) { | |
5717 | if (!(wq->flags & WQ_FREEZABLE)) | |
5718 | continue; | |
5719 | /* | |
5720 | * nr_active is monotonically decreasing. It's safe | |
5721 | * to peek without lock. | |
5722 | */ | |
24acfb71 | 5723 | rcu_read_lock(); |
6ba94429 FW |
5724 | for_each_pwq(pwq, wq) { |
5725 | WARN_ON_ONCE(pwq->nr_active < 0); | |
5726 | if (pwq->nr_active) { | |
5727 | busy = true; | |
24acfb71 | 5728 | rcu_read_unlock(); |
6ba94429 FW |
5729 | goto out_unlock; |
5730 | } | |
5731 | } | |
24acfb71 | 5732 | rcu_read_unlock(); |
6ba94429 FW |
5733 | } |
5734 | out_unlock: | |
5735 | mutex_unlock(&wq_pool_mutex); | |
5736 | return busy; | |
5737 | } | |
5738 | ||
5739 | /** | |
5740 | * thaw_workqueues - thaw workqueues | |
5741 | * | |
5742 | * Thaw workqueues. Normal queueing is restored and all collected | |
5743 | * frozen works are transferred to their respective pool worklists. | |
5744 | * | |
5745 | * CONTEXT: | |
5746 | * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. | |
5747 | */ | |
5748 | void thaw_workqueues(void) | |
5749 | { | |
5750 | struct workqueue_struct *wq; | |
5751 | struct pool_workqueue *pwq; | |
5752 | ||
5753 | mutex_lock(&wq_pool_mutex); | |
5754 | ||
5755 | if (!workqueue_freezing) | |
5756 | goto out_unlock; | |
5757 | ||
5758 | workqueue_freezing = false; | |
5759 | ||
5760 | /* restore max_active and repopulate worklist */ | |
5761 | list_for_each_entry(wq, &workqueues, list) { | |
5762 | mutex_lock(&wq->mutex); | |
5763 | for_each_pwq(pwq, wq) | |
5764 | pwq_adjust_max_active(pwq); | |
5765 | mutex_unlock(&wq->mutex); | |
5766 | } | |
5767 | ||
5768 | out_unlock: | |
5769 | mutex_unlock(&wq_pool_mutex); | |
5770 | } | |
5771 | #endif /* CONFIG_FREEZER */ | |
5772 | ||
99c621ef | 5773 | static int workqueue_apply_unbound_cpumask(const cpumask_var_t unbound_cpumask) |
042f7df1 LJ |
5774 | { |
5775 | LIST_HEAD(ctxs); | |
5776 | int ret = 0; | |
5777 | struct workqueue_struct *wq; | |
5778 | struct apply_wqattrs_ctx *ctx, *n; | |
5779 | ||
5780 | lockdep_assert_held(&wq_pool_mutex); | |
5781 | ||
5782 | list_for_each_entry(wq, &workqueues, list) { | |
5783 | if (!(wq->flags & WQ_UNBOUND)) | |
5784 | continue; | |
5785 | /* creating multiple pwqs breaks ordering guarantee */ | |
5786 | if (wq->flags & __WQ_ORDERED) | |
5787 | continue; | |
5788 | ||
99c621ef | 5789 | ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs, unbound_cpumask); |
84193c07 TH |
5790 | if (IS_ERR(ctx)) { |
5791 | ret = PTR_ERR(ctx); | |
042f7df1 LJ |
5792 | break; |
5793 | } | |
5794 | ||
5795 | list_add_tail(&ctx->list, &ctxs); | |
5796 | } | |
5797 | ||
5798 | list_for_each_entry_safe(ctx, n, &ctxs, list) { | |
5799 | if (!ret) | |
5800 | apply_wqattrs_commit(ctx); | |
5801 | apply_wqattrs_cleanup(ctx); | |
5802 | } | |
5803 | ||
99c621ef LJ |
5804 | if (!ret) { |
5805 | mutex_lock(&wq_pool_attach_mutex); | |
5806 | cpumask_copy(wq_unbound_cpumask, unbound_cpumask); | |
5807 | mutex_unlock(&wq_pool_attach_mutex); | |
5808 | } | |
042f7df1 LJ |
5809 | return ret; |
5810 | } | |
5811 | ||
5812 | /** | |
5813 | * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask | |
5814 | * @cpumask: the cpumask to set | |
5815 | * | |
5816 | * The low-level workqueues cpumask is a global cpumask that limits | |
5817 | * the affinity of all unbound workqueues. This function check the @cpumask | |
5818 | * and apply it to all unbound workqueues and updates all pwqs of them. | |
5819 | * | |
67dc8325 | 5820 | * Return: 0 - Success |
042f7df1 LJ |
5821 | * -EINVAL - Invalid @cpumask |
5822 | * -ENOMEM - Failed to allocate memory for attrs or pwqs. | |
5823 | */ | |
5824 | int workqueue_set_unbound_cpumask(cpumask_var_t cpumask) | |
5825 | { | |
5826 | int ret = -EINVAL; | |
042f7df1 | 5827 | |
c98a9805 TS |
5828 | /* |
5829 | * Not excluding isolated cpus on purpose. | |
5830 | * If the user wishes to include them, we allow that. | |
5831 | */ | |
042f7df1 LJ |
5832 | cpumask_and(cpumask, cpumask, cpu_possible_mask); |
5833 | if (!cpumask_empty(cpumask)) { | |
a0111cf6 | 5834 | apply_wqattrs_lock(); |
d25302e4 MD |
5835 | if (cpumask_equal(cpumask, wq_unbound_cpumask)) { |
5836 | ret = 0; | |
5837 | goto out_unlock; | |
5838 | } | |
5839 | ||
99c621ef | 5840 | ret = workqueue_apply_unbound_cpumask(cpumask); |
042f7df1 | 5841 | |
d25302e4 | 5842 | out_unlock: |
a0111cf6 | 5843 | apply_wqattrs_unlock(); |
042f7df1 | 5844 | } |
042f7df1 | 5845 | |
042f7df1 LJ |
5846 | return ret; |
5847 | } | |
5848 | ||
63c5484e TH |
5849 | static int parse_affn_scope(const char *val) |
5850 | { | |
5851 | int i; | |
5852 | ||
5853 | for (i = 0; i < ARRAY_SIZE(wq_affn_names); i++) { | |
5854 | if (!strncasecmp(val, wq_affn_names[i], strlen(wq_affn_names[i]))) | |
5855 | return i; | |
5856 | } | |
5857 | return -EINVAL; | |
5858 | } | |
5859 | ||
5860 | static int wq_affn_dfl_set(const char *val, const struct kernel_param *kp) | |
5861 | { | |
523a301e TH |
5862 | struct workqueue_struct *wq; |
5863 | int affn, cpu; | |
63c5484e TH |
5864 | |
5865 | affn = parse_affn_scope(val); | |
5866 | if (affn < 0) | |
5867 | return affn; | |
523a301e TH |
5868 | if (affn == WQ_AFFN_DFL) |
5869 | return -EINVAL; | |
5870 | ||
5871 | cpus_read_lock(); | |
5872 | mutex_lock(&wq_pool_mutex); | |
63c5484e TH |
5873 | |
5874 | wq_affn_dfl = affn; | |
523a301e TH |
5875 | |
5876 | list_for_each_entry(wq, &workqueues, list) { | |
5877 | for_each_online_cpu(cpu) { | |
5878 | wq_update_pod(wq, cpu, cpu, true); | |
5879 | } | |
5880 | } | |
5881 | ||
5882 | mutex_unlock(&wq_pool_mutex); | |
5883 | cpus_read_unlock(); | |
5884 | ||
63c5484e TH |
5885 | return 0; |
5886 | } | |
5887 | ||
5888 | static int wq_affn_dfl_get(char *buffer, const struct kernel_param *kp) | |
5889 | { | |
5890 | return scnprintf(buffer, PAGE_SIZE, "%s\n", wq_affn_names[wq_affn_dfl]); | |
5891 | } | |
5892 | ||
5893 | static const struct kernel_param_ops wq_affn_dfl_ops = { | |
5894 | .set = wq_affn_dfl_set, | |
5895 | .get = wq_affn_dfl_get, | |
5896 | }; | |
5897 | ||
5898 | module_param_cb(default_affinity_scope, &wq_affn_dfl_ops, NULL, 0644); | |
5899 | ||
6ba94429 FW |
5900 | #ifdef CONFIG_SYSFS |
5901 | /* | |
5902 | * Workqueues with WQ_SYSFS flag set is visible to userland via | |
5903 | * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the | |
5904 | * following attributes. | |
5905 | * | |
63c5484e TH |
5906 | * per_cpu RO bool : whether the workqueue is per-cpu or unbound |
5907 | * max_active RW int : maximum number of in-flight work items | |
6ba94429 FW |
5908 | * |
5909 | * Unbound workqueues have the following extra attributes. | |
5910 | * | |
63c5484e TH |
5911 | * nice RW int : nice value of the workers |
5912 | * cpumask RW mask : bitmask of allowed CPUs for the workers | |
5913 | * affinity_scope RW str : worker CPU affinity scope (cache, numa, none) | |
8639eceb | 5914 | * affinity_strict RW bool : worker CPU affinity is strict |
6ba94429 FW |
5915 | */ |
5916 | struct wq_device { | |
5917 | struct workqueue_struct *wq; | |
5918 | struct device dev; | |
5919 | }; | |
5920 | ||
5921 | static struct workqueue_struct *dev_to_wq(struct device *dev) | |
5922 | { | |
5923 | struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); | |
5924 | ||
5925 | return wq_dev->wq; | |
5926 | } | |
5927 | ||
5928 | static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, | |
5929 | char *buf) | |
5930 | { | |
5931 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5932 | ||
5933 | return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); | |
5934 | } | |
5935 | static DEVICE_ATTR_RO(per_cpu); | |
5936 | ||
5937 | static ssize_t max_active_show(struct device *dev, | |
5938 | struct device_attribute *attr, char *buf) | |
5939 | { | |
5940 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5941 | ||
5942 | return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); | |
5943 | } | |
5944 | ||
5945 | static ssize_t max_active_store(struct device *dev, | |
5946 | struct device_attribute *attr, const char *buf, | |
5947 | size_t count) | |
5948 | { | |
5949 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5950 | int val; | |
5951 | ||
5952 | if (sscanf(buf, "%d", &val) != 1 || val <= 0) | |
5953 | return -EINVAL; | |
5954 | ||
5955 | workqueue_set_max_active(wq, val); | |
5956 | return count; | |
5957 | } | |
5958 | static DEVICE_ATTR_RW(max_active); | |
5959 | ||
5960 | static struct attribute *wq_sysfs_attrs[] = { | |
5961 | &dev_attr_per_cpu.attr, | |
5962 | &dev_attr_max_active.attr, | |
5963 | NULL, | |
5964 | }; | |
5965 | ATTRIBUTE_GROUPS(wq_sysfs); | |
5966 | ||
6ba94429 FW |
5967 | static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, |
5968 | char *buf) | |
5969 | { | |
5970 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5971 | int written; | |
5972 | ||
5973 | mutex_lock(&wq->mutex); | |
5974 | written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); | |
5975 | mutex_unlock(&wq->mutex); | |
5976 | ||
5977 | return written; | |
5978 | } | |
5979 | ||
5980 | /* prepare workqueue_attrs for sysfs store operations */ | |
5981 | static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) | |
5982 | { | |
5983 | struct workqueue_attrs *attrs; | |
5984 | ||
899a94fe LJ |
5985 | lockdep_assert_held(&wq_pool_mutex); |
5986 | ||
be69d00d | 5987 | attrs = alloc_workqueue_attrs(); |
6ba94429 FW |
5988 | if (!attrs) |
5989 | return NULL; | |
5990 | ||
6ba94429 | 5991 | copy_workqueue_attrs(attrs, wq->unbound_attrs); |
6ba94429 FW |
5992 | return attrs; |
5993 | } | |
5994 | ||
5995 | static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, | |
5996 | const char *buf, size_t count) | |
5997 | { | |
5998 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5999 | struct workqueue_attrs *attrs; | |
d4d3e257 LJ |
6000 | int ret = -ENOMEM; |
6001 | ||
6002 | apply_wqattrs_lock(); | |
6ba94429 FW |
6003 | |
6004 | attrs = wq_sysfs_prep_attrs(wq); | |
6005 | if (!attrs) | |
d4d3e257 | 6006 | goto out_unlock; |
6ba94429 FW |
6007 | |
6008 | if (sscanf(buf, "%d", &attrs->nice) == 1 && | |
6009 | attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) | |
d4d3e257 | 6010 | ret = apply_workqueue_attrs_locked(wq, attrs); |
6ba94429 FW |
6011 | else |
6012 | ret = -EINVAL; | |
6013 | ||
d4d3e257 LJ |
6014 | out_unlock: |
6015 | apply_wqattrs_unlock(); | |
6ba94429 FW |
6016 | free_workqueue_attrs(attrs); |
6017 | return ret ?: count; | |
6018 | } | |
6019 | ||
6020 | static ssize_t wq_cpumask_show(struct device *dev, | |
6021 | struct device_attribute *attr, char *buf) | |
6022 | { | |
6023 | struct workqueue_struct *wq = dev_to_wq(dev); | |
6024 | int written; | |
6025 | ||
6026 | mutex_lock(&wq->mutex); | |
6027 | written = scnprintf(buf, PAGE_SIZE, "%*pb\n", | |
6028 | cpumask_pr_args(wq->unbound_attrs->cpumask)); | |
6029 | mutex_unlock(&wq->mutex); | |
6030 | return written; | |
6031 | } | |
6032 | ||
6033 | static ssize_t wq_cpumask_store(struct device *dev, | |
6034 | struct device_attribute *attr, | |
6035 | const char *buf, size_t count) | |
6036 | { | |
6037 | struct workqueue_struct *wq = dev_to_wq(dev); | |
6038 | struct workqueue_attrs *attrs; | |
d4d3e257 LJ |
6039 | int ret = -ENOMEM; |
6040 | ||
6041 | apply_wqattrs_lock(); | |
6ba94429 FW |
6042 | |
6043 | attrs = wq_sysfs_prep_attrs(wq); | |
6044 | if (!attrs) | |
d4d3e257 | 6045 | goto out_unlock; |
6ba94429 FW |
6046 | |
6047 | ret = cpumask_parse(buf, attrs->cpumask); | |
6048 | if (!ret) | |
d4d3e257 | 6049 | ret = apply_workqueue_attrs_locked(wq, attrs); |
6ba94429 | 6050 | |
d4d3e257 LJ |
6051 | out_unlock: |
6052 | apply_wqattrs_unlock(); | |
6ba94429 FW |
6053 | free_workqueue_attrs(attrs); |
6054 | return ret ?: count; | |
6055 | } | |
6056 | ||
63c5484e TH |
6057 | static ssize_t wq_affn_scope_show(struct device *dev, |
6058 | struct device_attribute *attr, char *buf) | |
6059 | { | |
6060 | struct workqueue_struct *wq = dev_to_wq(dev); | |
6061 | int written; | |
6062 | ||
6063 | mutex_lock(&wq->mutex); | |
523a301e TH |
6064 | if (wq->unbound_attrs->affn_scope == WQ_AFFN_DFL) |
6065 | written = scnprintf(buf, PAGE_SIZE, "%s (%s)\n", | |
6066 | wq_affn_names[WQ_AFFN_DFL], | |
6067 | wq_affn_names[wq_affn_dfl]); | |
6068 | else | |
6069 | written = scnprintf(buf, PAGE_SIZE, "%s\n", | |
6070 | wq_affn_names[wq->unbound_attrs->affn_scope]); | |
63c5484e TH |
6071 | mutex_unlock(&wq->mutex); |
6072 | ||
6073 | return written; | |
6074 | } | |
6075 | ||
6076 | static ssize_t wq_affn_scope_store(struct device *dev, | |
6077 | struct device_attribute *attr, | |
6078 | const char *buf, size_t count) | |
6079 | { | |
6080 | struct workqueue_struct *wq = dev_to_wq(dev); | |
6081 | struct workqueue_attrs *attrs; | |
6082 | int affn, ret = -ENOMEM; | |
6083 | ||
6084 | affn = parse_affn_scope(buf); | |
6085 | if (affn < 0) | |
6086 | return affn; | |
6087 | ||
6088 | apply_wqattrs_lock(); | |
6089 | attrs = wq_sysfs_prep_attrs(wq); | |
6090 | if (attrs) { | |
6091 | attrs->affn_scope = affn; | |
6092 | ret = apply_workqueue_attrs_locked(wq, attrs); | |
6093 | } | |
6094 | apply_wqattrs_unlock(); | |
6095 | free_workqueue_attrs(attrs); | |
6096 | return ret ?: count; | |
6097 | } | |
6098 | ||
8639eceb TH |
6099 | static ssize_t wq_affinity_strict_show(struct device *dev, |
6100 | struct device_attribute *attr, char *buf) | |
6101 | { | |
6102 | struct workqueue_struct *wq = dev_to_wq(dev); | |
6103 | ||
6104 | return scnprintf(buf, PAGE_SIZE, "%d\n", | |
6105 | wq->unbound_attrs->affn_strict); | |
6106 | } | |
6107 | ||
6108 | static ssize_t wq_affinity_strict_store(struct device *dev, | |
6109 | struct device_attribute *attr, | |
6110 | const char *buf, size_t count) | |
6111 | { | |
6112 | struct workqueue_struct *wq = dev_to_wq(dev); | |
6113 | struct workqueue_attrs *attrs; | |
6114 | int v, ret = -ENOMEM; | |
6115 | ||
6116 | if (sscanf(buf, "%d", &v) != 1) | |
6117 | return -EINVAL; | |
6118 | ||
6119 | apply_wqattrs_lock(); | |
6120 | attrs = wq_sysfs_prep_attrs(wq); | |
6121 | if (attrs) { | |
6122 | attrs->affn_strict = (bool)v; | |
6123 | ret = apply_workqueue_attrs_locked(wq, attrs); | |
6124 | } | |
6125 | apply_wqattrs_unlock(); | |
6126 | free_workqueue_attrs(attrs); | |
6127 | return ret ?: count; | |
6128 | } | |
6129 | ||
6ba94429 | 6130 | static struct device_attribute wq_sysfs_unbound_attrs[] = { |
6ba94429 FW |
6131 | __ATTR(nice, 0644, wq_nice_show, wq_nice_store), |
6132 | __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), | |
63c5484e | 6133 | __ATTR(affinity_scope, 0644, wq_affn_scope_show, wq_affn_scope_store), |
8639eceb | 6134 | __ATTR(affinity_strict, 0644, wq_affinity_strict_show, wq_affinity_strict_store), |
6ba94429 FW |
6135 | __ATTR_NULL, |
6136 | }; | |
8ccad40d | 6137 | |
6ba94429 FW |
6138 | static struct bus_type wq_subsys = { |
6139 | .name = "workqueue", | |
6140 | .dev_groups = wq_sysfs_groups, | |
2d3854a3 RR |
6141 | }; |
6142 | ||
b05a7928 FW |
6143 | static ssize_t wq_unbound_cpumask_show(struct device *dev, |
6144 | struct device_attribute *attr, char *buf) | |
6145 | { | |
6146 | int written; | |
6147 | ||
042f7df1 | 6148 | mutex_lock(&wq_pool_mutex); |
b05a7928 FW |
6149 | written = scnprintf(buf, PAGE_SIZE, "%*pb\n", |
6150 | cpumask_pr_args(wq_unbound_cpumask)); | |
042f7df1 | 6151 | mutex_unlock(&wq_pool_mutex); |
b05a7928 FW |
6152 | |
6153 | return written; | |
6154 | } | |
6155 | ||
042f7df1 LJ |
6156 | static ssize_t wq_unbound_cpumask_store(struct device *dev, |
6157 | struct device_attribute *attr, const char *buf, size_t count) | |
6158 | { | |
6159 | cpumask_var_t cpumask; | |
6160 | int ret; | |
6161 | ||
6162 | if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) | |
6163 | return -ENOMEM; | |
6164 | ||
6165 | ret = cpumask_parse(buf, cpumask); | |
6166 | if (!ret) | |
6167 | ret = workqueue_set_unbound_cpumask(cpumask); | |
6168 | ||
6169 | free_cpumask_var(cpumask); | |
6170 | return ret ? ret : count; | |
6171 | } | |
6172 | ||
b05a7928 | 6173 | static struct device_attribute wq_sysfs_cpumask_attr = |
042f7df1 LJ |
6174 | __ATTR(cpumask, 0644, wq_unbound_cpumask_show, |
6175 | wq_unbound_cpumask_store); | |
b05a7928 | 6176 | |
6ba94429 | 6177 | static int __init wq_sysfs_init(void) |
2d3854a3 | 6178 | { |
686f6697 | 6179 | struct device *dev_root; |
b05a7928 FW |
6180 | int err; |
6181 | ||
6182 | err = subsys_virtual_register(&wq_subsys, NULL); | |
6183 | if (err) | |
6184 | return err; | |
6185 | ||
686f6697 GKH |
6186 | dev_root = bus_get_dev_root(&wq_subsys); |
6187 | if (dev_root) { | |
6188 | err = device_create_file(dev_root, &wq_sysfs_cpumask_attr); | |
6189 | put_device(dev_root); | |
6190 | } | |
6191 | return err; | |
2d3854a3 | 6192 | } |
6ba94429 | 6193 | core_initcall(wq_sysfs_init); |
2d3854a3 | 6194 | |
6ba94429 | 6195 | static void wq_device_release(struct device *dev) |
2d3854a3 | 6196 | { |
6ba94429 | 6197 | struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); |
6b44003e | 6198 | |
6ba94429 | 6199 | kfree(wq_dev); |
2d3854a3 | 6200 | } |
a0a1a5fd TH |
6201 | |
6202 | /** | |
6ba94429 FW |
6203 | * workqueue_sysfs_register - make a workqueue visible in sysfs |
6204 | * @wq: the workqueue to register | |
a0a1a5fd | 6205 | * |
6ba94429 FW |
6206 | * Expose @wq in sysfs under /sys/bus/workqueue/devices. |
6207 | * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set | |
6208 | * which is the preferred method. | |
a0a1a5fd | 6209 | * |
6ba94429 FW |
6210 | * Workqueue user should use this function directly iff it wants to apply |
6211 | * workqueue_attrs before making the workqueue visible in sysfs; otherwise, | |
6212 | * apply_workqueue_attrs() may race against userland updating the | |
6213 | * attributes. | |
6214 | * | |
6215 | * Return: 0 on success, -errno on failure. | |
a0a1a5fd | 6216 | */ |
6ba94429 | 6217 | int workqueue_sysfs_register(struct workqueue_struct *wq) |
a0a1a5fd | 6218 | { |
6ba94429 FW |
6219 | struct wq_device *wq_dev; |
6220 | int ret; | |
a0a1a5fd | 6221 | |
6ba94429 | 6222 | /* |
402dd89d | 6223 | * Adjusting max_active or creating new pwqs by applying |
6ba94429 FW |
6224 | * attributes breaks ordering guarantee. Disallow exposing ordered |
6225 | * workqueues. | |
6226 | */ | |
0a94efb5 | 6227 | if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) |
6ba94429 | 6228 | return -EINVAL; |
a0a1a5fd | 6229 | |
6ba94429 FW |
6230 | wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); |
6231 | if (!wq_dev) | |
6232 | return -ENOMEM; | |
5bcab335 | 6233 | |
6ba94429 FW |
6234 | wq_dev->wq = wq; |
6235 | wq_dev->dev.bus = &wq_subsys; | |
6ba94429 | 6236 | wq_dev->dev.release = wq_device_release; |
23217b44 | 6237 | dev_set_name(&wq_dev->dev, "%s", wq->name); |
a0a1a5fd | 6238 | |
6ba94429 FW |
6239 | /* |
6240 | * unbound_attrs are created separately. Suppress uevent until | |
6241 | * everything is ready. | |
6242 | */ | |
6243 | dev_set_uevent_suppress(&wq_dev->dev, true); | |
a0a1a5fd | 6244 | |
6ba94429 FW |
6245 | ret = device_register(&wq_dev->dev); |
6246 | if (ret) { | |
537f4146 | 6247 | put_device(&wq_dev->dev); |
6ba94429 FW |
6248 | wq->wq_dev = NULL; |
6249 | return ret; | |
6250 | } | |
a0a1a5fd | 6251 | |
6ba94429 FW |
6252 | if (wq->flags & WQ_UNBOUND) { |
6253 | struct device_attribute *attr; | |
a0a1a5fd | 6254 | |
6ba94429 FW |
6255 | for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { |
6256 | ret = device_create_file(&wq_dev->dev, attr); | |
6257 | if (ret) { | |
6258 | device_unregister(&wq_dev->dev); | |
6259 | wq->wq_dev = NULL; | |
6260 | return ret; | |
a0a1a5fd TH |
6261 | } |
6262 | } | |
6263 | } | |
6ba94429 FW |
6264 | |
6265 | dev_set_uevent_suppress(&wq_dev->dev, false); | |
6266 | kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); | |
6267 | return 0; | |
a0a1a5fd TH |
6268 | } |
6269 | ||
6270 | /** | |
6ba94429 FW |
6271 | * workqueue_sysfs_unregister - undo workqueue_sysfs_register() |
6272 | * @wq: the workqueue to unregister | |
a0a1a5fd | 6273 | * |
6ba94429 | 6274 | * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. |
a0a1a5fd | 6275 | */ |
6ba94429 | 6276 | static void workqueue_sysfs_unregister(struct workqueue_struct *wq) |
a0a1a5fd | 6277 | { |
6ba94429 | 6278 | struct wq_device *wq_dev = wq->wq_dev; |
8b03ae3c | 6279 | |
6ba94429 FW |
6280 | if (!wq->wq_dev) |
6281 | return; | |
a0a1a5fd | 6282 | |
6ba94429 FW |
6283 | wq->wq_dev = NULL; |
6284 | device_unregister(&wq_dev->dev); | |
a0a1a5fd | 6285 | } |
6ba94429 FW |
6286 | #else /* CONFIG_SYSFS */ |
6287 | static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } | |
6288 | #endif /* CONFIG_SYSFS */ | |
a0a1a5fd | 6289 | |
82607adc TH |
6290 | /* |
6291 | * Workqueue watchdog. | |
6292 | * | |
6293 | * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal | |
6294 | * flush dependency, a concurrency managed work item which stays RUNNING | |
6295 | * indefinitely. Workqueue stalls can be very difficult to debug as the | |
6296 | * usual warning mechanisms don't trigger and internal workqueue state is | |
6297 | * largely opaque. | |
6298 | * | |
6299 | * Workqueue watchdog monitors all worker pools periodically and dumps | |
6300 | * state if some pools failed to make forward progress for a while where | |
6301 | * forward progress is defined as the first item on ->worklist changing. | |
6302 | * | |
6303 | * This mechanism is controlled through the kernel parameter | |
6304 | * "workqueue.watchdog_thresh" which can be updated at runtime through the | |
6305 | * corresponding sysfs parameter file. | |
6306 | */ | |
6307 | #ifdef CONFIG_WQ_WATCHDOG | |
6308 | ||
82607adc | 6309 | static unsigned long wq_watchdog_thresh = 30; |
5cd79d6a | 6310 | static struct timer_list wq_watchdog_timer; |
82607adc TH |
6311 | |
6312 | static unsigned long wq_watchdog_touched = INITIAL_JIFFIES; | |
6313 | static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES; | |
6314 | ||
cd2440d6 PM |
6315 | /* |
6316 | * Show workers that might prevent the processing of pending work items. | |
6317 | * The only candidates are CPU-bound workers in the running state. | |
6318 | * Pending work items should be handled by another idle worker | |
6319 | * in all other situations. | |
6320 | */ | |
6321 | static void show_cpu_pool_hog(struct worker_pool *pool) | |
6322 | { | |
6323 | struct worker *worker; | |
6324 | unsigned long flags; | |
6325 | int bkt; | |
6326 | ||
6327 | raw_spin_lock_irqsave(&pool->lock, flags); | |
6328 | ||
6329 | hash_for_each(pool->busy_hash, bkt, worker, hentry) { | |
6330 | if (task_is_running(worker->task)) { | |
6331 | /* | |
6332 | * Defer printing to avoid deadlocks in console | |
6333 | * drivers that queue work while holding locks | |
6334 | * also taken in their write paths. | |
6335 | */ | |
6336 | printk_deferred_enter(); | |
6337 | ||
6338 | pr_info("pool %d:\n", pool->id); | |
6339 | sched_show_task(worker->task); | |
6340 | ||
6341 | printk_deferred_exit(); | |
6342 | } | |
6343 | } | |
6344 | ||
6345 | raw_spin_unlock_irqrestore(&pool->lock, flags); | |
6346 | } | |
6347 | ||
6348 | static void show_cpu_pools_hogs(void) | |
6349 | { | |
6350 | struct worker_pool *pool; | |
6351 | int pi; | |
6352 | ||
6353 | pr_info("Showing backtraces of running workers in stalled CPU-bound worker pools:\n"); | |
6354 | ||
6355 | rcu_read_lock(); | |
6356 | ||
6357 | for_each_pool(pool, pi) { | |
6358 | if (pool->cpu_stall) | |
6359 | show_cpu_pool_hog(pool); | |
6360 | ||
6361 | } | |
6362 | ||
6363 | rcu_read_unlock(); | |
6364 | } | |
6365 | ||
82607adc TH |
6366 | static void wq_watchdog_reset_touched(void) |
6367 | { | |
6368 | int cpu; | |
6369 | ||
6370 | wq_watchdog_touched = jiffies; | |
6371 | for_each_possible_cpu(cpu) | |
6372 | per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; | |
6373 | } | |
6374 | ||
5cd79d6a | 6375 | static void wq_watchdog_timer_fn(struct timer_list *unused) |
82607adc TH |
6376 | { |
6377 | unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ; | |
6378 | bool lockup_detected = false; | |
cd2440d6 | 6379 | bool cpu_pool_stall = false; |
940d71c6 | 6380 | unsigned long now = jiffies; |
82607adc TH |
6381 | struct worker_pool *pool; |
6382 | int pi; | |
6383 | ||
6384 | if (!thresh) | |
6385 | return; | |
6386 | ||
6387 | rcu_read_lock(); | |
6388 | ||
6389 | for_each_pool(pool, pi) { | |
6390 | unsigned long pool_ts, touched, ts; | |
6391 | ||
cd2440d6 | 6392 | pool->cpu_stall = false; |
82607adc TH |
6393 | if (list_empty(&pool->worklist)) |
6394 | continue; | |
6395 | ||
940d71c6 SS |
6396 | /* |
6397 | * If a virtual machine is stopped by the host it can look to | |
6398 | * the watchdog like a stall. | |
6399 | */ | |
6400 | kvm_check_and_clear_guest_paused(); | |
6401 | ||
82607adc | 6402 | /* get the latest of pool and touched timestamps */ |
89e28ce6 WQ |
6403 | if (pool->cpu >= 0) |
6404 | touched = READ_ONCE(per_cpu(wq_watchdog_touched_cpu, pool->cpu)); | |
6405 | else | |
6406 | touched = READ_ONCE(wq_watchdog_touched); | |
82607adc | 6407 | pool_ts = READ_ONCE(pool->watchdog_ts); |
82607adc TH |
6408 | |
6409 | if (time_after(pool_ts, touched)) | |
6410 | ts = pool_ts; | |
6411 | else | |
6412 | ts = touched; | |
6413 | ||
82607adc | 6414 | /* did we stall? */ |
940d71c6 | 6415 | if (time_after(now, ts + thresh)) { |
82607adc | 6416 | lockup_detected = true; |
cd2440d6 PM |
6417 | if (pool->cpu >= 0) { |
6418 | pool->cpu_stall = true; | |
6419 | cpu_pool_stall = true; | |
6420 | } | |
82607adc TH |
6421 | pr_emerg("BUG: workqueue lockup - pool"); |
6422 | pr_cont_pool_info(pool); | |
6423 | pr_cont(" stuck for %us!\n", | |
940d71c6 | 6424 | jiffies_to_msecs(now - pool_ts) / 1000); |
82607adc | 6425 | } |
cd2440d6 PM |
6426 | |
6427 | ||
82607adc TH |
6428 | } |
6429 | ||
6430 | rcu_read_unlock(); | |
6431 | ||
6432 | if (lockup_detected) | |
55df0933 | 6433 | show_all_workqueues(); |
82607adc | 6434 | |
cd2440d6 PM |
6435 | if (cpu_pool_stall) |
6436 | show_cpu_pools_hogs(); | |
6437 | ||
82607adc TH |
6438 | wq_watchdog_reset_touched(); |
6439 | mod_timer(&wq_watchdog_timer, jiffies + thresh); | |
6440 | } | |
6441 | ||
cb9d7fd5 | 6442 | notrace void wq_watchdog_touch(int cpu) |
82607adc TH |
6443 | { |
6444 | if (cpu >= 0) | |
6445 | per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; | |
89e28ce6 WQ |
6446 | |
6447 | wq_watchdog_touched = jiffies; | |
82607adc TH |
6448 | } |
6449 | ||
6450 | static void wq_watchdog_set_thresh(unsigned long thresh) | |
6451 | { | |
6452 | wq_watchdog_thresh = 0; | |
6453 | del_timer_sync(&wq_watchdog_timer); | |
6454 | ||
6455 | if (thresh) { | |
6456 | wq_watchdog_thresh = thresh; | |
6457 | wq_watchdog_reset_touched(); | |
6458 | mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ); | |
6459 | } | |
6460 | } | |
6461 | ||
6462 | static int wq_watchdog_param_set_thresh(const char *val, | |
6463 | const struct kernel_param *kp) | |
6464 | { | |
6465 | unsigned long thresh; | |
6466 | int ret; | |
6467 | ||
6468 | ret = kstrtoul(val, 0, &thresh); | |
6469 | if (ret) | |
6470 | return ret; | |
6471 | ||
6472 | if (system_wq) | |
6473 | wq_watchdog_set_thresh(thresh); | |
6474 | else | |
6475 | wq_watchdog_thresh = thresh; | |
6476 | ||
6477 | return 0; | |
6478 | } | |
6479 | ||
6480 | static const struct kernel_param_ops wq_watchdog_thresh_ops = { | |
6481 | .set = wq_watchdog_param_set_thresh, | |
6482 | .get = param_get_ulong, | |
6483 | }; | |
6484 | ||
6485 | module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh, | |
6486 | 0644); | |
6487 | ||
6488 | static void wq_watchdog_init(void) | |
6489 | { | |
5cd79d6a | 6490 | timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE); |
82607adc TH |
6491 | wq_watchdog_set_thresh(wq_watchdog_thresh); |
6492 | } | |
6493 | ||
6494 | #else /* CONFIG_WQ_WATCHDOG */ | |
6495 | ||
6496 | static inline void wq_watchdog_init(void) { } | |
6497 | ||
6498 | #endif /* CONFIG_WQ_WATCHDOG */ | |
6499 | ||
3347fa09 TH |
6500 | /** |
6501 | * workqueue_init_early - early init for workqueue subsystem | |
6502 | * | |
2930155b TH |
6503 | * This is the first step of three-staged workqueue subsystem initialization and |
6504 | * invoked as soon as the bare basics - memory allocation, cpumasks and idr are | |
6505 | * up. It sets up all the data structures and system workqueues and allows early | |
6506 | * boot code to create workqueues and queue/cancel work items. Actual work item | |
6507 | * execution starts only after kthreads can be created and scheduled right | |
6508 | * before early initcalls. | |
3347fa09 | 6509 | */ |
2333e829 | 6510 | void __init workqueue_init_early(void) |
1da177e4 | 6511 | { |
84193c07 | 6512 | struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_SYSTEM]; |
7a4e344c TH |
6513 | int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; |
6514 | int i, cpu; | |
c34056a3 | 6515 | |
10cdb157 | 6516 | BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); |
e904e6c2 | 6517 | |
b05a7928 | 6518 | BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL)); |
04d4e665 FW |
6519 | cpumask_copy(wq_unbound_cpumask, housekeeping_cpumask(HK_TYPE_WQ)); |
6520 | cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, housekeeping_cpumask(HK_TYPE_DOMAIN)); | |
b05a7928 | 6521 | |
ace3c549 | 6522 | if (!cpumask_empty(&wq_cmdline_cpumask)) |
6523 | cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, &wq_cmdline_cpumask); | |
6524 | ||
e904e6c2 TH |
6525 | pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); |
6526 | ||
2930155b TH |
6527 | wq_update_pod_attrs_buf = alloc_workqueue_attrs(); |
6528 | BUG_ON(!wq_update_pod_attrs_buf); | |
6529 | ||
84193c07 TH |
6530 | /* initialize WQ_AFFN_SYSTEM pods */ |
6531 | pt->pod_cpus = kcalloc(1, sizeof(pt->pod_cpus[0]), GFP_KERNEL); | |
6532 | pt->pod_node = kcalloc(1, sizeof(pt->pod_node[0]), GFP_KERNEL); | |
6533 | pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL); | |
6534 | BUG_ON(!pt->pod_cpus || !pt->pod_node || !pt->cpu_pod); | |
6535 | ||
6536 | BUG_ON(!zalloc_cpumask_var_node(&pt->pod_cpus[0], GFP_KERNEL, NUMA_NO_NODE)); | |
6537 | ||
84193c07 TH |
6538 | pt->nr_pods = 1; |
6539 | cpumask_copy(pt->pod_cpus[0], cpu_possible_mask); | |
6540 | pt->pod_node[0] = NUMA_NO_NODE; | |
6541 | pt->cpu_pod[0] = 0; | |
6542 | ||
706026c2 | 6543 | /* initialize CPU pools */ |
29c91e99 | 6544 | for_each_possible_cpu(cpu) { |
4ce62e9e | 6545 | struct worker_pool *pool; |
8b03ae3c | 6546 | |
7a4e344c | 6547 | i = 0; |
f02ae73a | 6548 | for_each_cpu_worker_pool(pool, cpu) { |
7a4e344c | 6549 | BUG_ON(init_worker_pool(pool)); |
ec22ca5e | 6550 | pool->cpu = cpu; |
29c91e99 | 6551 | cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); |
9546b29e | 6552 | cpumask_copy(pool->attrs->__pod_cpumask, cpumask_of(cpu)); |
7a4e344c | 6553 | pool->attrs->nice = std_nice[i++]; |
8639eceb | 6554 | pool->attrs->affn_strict = true; |
f3f90ad4 | 6555 | pool->node = cpu_to_node(cpu); |
7a4e344c | 6556 | |
9daf9e67 | 6557 | /* alloc pool ID */ |
68e13a67 | 6558 | mutex_lock(&wq_pool_mutex); |
9daf9e67 | 6559 | BUG_ON(worker_pool_assign_id(pool)); |
68e13a67 | 6560 | mutex_unlock(&wq_pool_mutex); |
4ce62e9e | 6561 | } |
8b03ae3c TH |
6562 | } |
6563 | ||
8a2b7538 | 6564 | /* create default unbound and ordered wq attrs */ |
29c91e99 TH |
6565 | for (i = 0; i < NR_STD_WORKER_POOLS; i++) { |
6566 | struct workqueue_attrs *attrs; | |
6567 | ||
be69d00d | 6568 | BUG_ON(!(attrs = alloc_workqueue_attrs())); |
29c91e99 | 6569 | attrs->nice = std_nice[i]; |
29c91e99 | 6570 | unbound_std_wq_attrs[i] = attrs; |
8a2b7538 TH |
6571 | |
6572 | /* | |
6573 | * An ordered wq should have only one pwq as ordering is | |
6574 | * guaranteed by max_active which is enforced by pwqs. | |
8a2b7538 | 6575 | */ |
be69d00d | 6576 | BUG_ON(!(attrs = alloc_workqueue_attrs())); |
8a2b7538 | 6577 | attrs->nice = std_nice[i]; |
af73f5c9 | 6578 | attrs->ordered = true; |
8a2b7538 | 6579 | ordered_wq_attrs[i] = attrs; |
29c91e99 TH |
6580 | } |
6581 | ||
d320c038 | 6582 | system_wq = alloc_workqueue("events", 0, 0); |
1aabe902 | 6583 | system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); |
d320c038 | 6584 | system_long_wq = alloc_workqueue("events_long", 0, 0); |
f3421797 | 6585 | system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, |
636b927e | 6586 | WQ_MAX_ACTIVE); |
24d51add TH |
6587 | system_freezable_wq = alloc_workqueue("events_freezable", |
6588 | WQ_FREEZABLE, 0); | |
0668106c VK |
6589 | system_power_efficient_wq = alloc_workqueue("events_power_efficient", |
6590 | WQ_POWER_EFFICIENT, 0); | |
6591 | system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient", | |
6592 | WQ_FREEZABLE | WQ_POWER_EFFICIENT, | |
6593 | 0); | |
1aabe902 | 6594 | BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || |
0668106c VK |
6595 | !system_unbound_wq || !system_freezable_wq || |
6596 | !system_power_efficient_wq || | |
6597 | !system_freezable_power_efficient_wq); | |
3347fa09 TH |
6598 | } |
6599 | ||
aa6fde93 TH |
6600 | static void __init wq_cpu_intensive_thresh_init(void) |
6601 | { | |
6602 | unsigned long thresh; | |
6603 | unsigned long bogo; | |
6604 | ||
dd64c873 Z |
6605 | pwq_release_worker = kthread_create_worker(0, "pool_workqueue_release"); |
6606 | BUG_ON(IS_ERR(pwq_release_worker)); | |
6607 | ||
aa6fde93 TH |
6608 | /* if the user set it to a specific value, keep it */ |
6609 | if (wq_cpu_intensive_thresh_us != ULONG_MAX) | |
6610 | return; | |
6611 | ||
6612 | /* | |
6613 | * The default of 10ms is derived from the fact that most modern (as of | |
6614 | * 2023) processors can do a lot in 10ms and that it's just below what | |
6615 | * most consider human-perceivable. However, the kernel also runs on a | |
6616 | * lot slower CPUs including microcontrollers where the threshold is way | |
6617 | * too low. | |
6618 | * | |
6619 | * Let's scale up the threshold upto 1 second if BogoMips is below 4000. | |
6620 | * This is by no means accurate but it doesn't have to be. The mechanism | |
6621 | * is still useful even when the threshold is fully scaled up. Also, as | |
6622 | * the reports would usually be applicable to everyone, some machines | |
6623 | * operating on longer thresholds won't significantly diminish their | |
6624 | * usefulness. | |
6625 | */ | |
6626 | thresh = 10 * USEC_PER_MSEC; | |
6627 | ||
6628 | /* see init/calibrate.c for lpj -> BogoMIPS calculation */ | |
6629 | bogo = max_t(unsigned long, loops_per_jiffy / 500000 * HZ, 1); | |
6630 | if (bogo < 4000) | |
6631 | thresh = min_t(unsigned long, thresh * 4000 / bogo, USEC_PER_SEC); | |
6632 | ||
6633 | pr_debug("wq_cpu_intensive_thresh: lpj=%lu BogoMIPS=%lu thresh_us=%lu\n", | |
6634 | loops_per_jiffy, bogo, thresh); | |
6635 | ||
6636 | wq_cpu_intensive_thresh_us = thresh; | |
6637 | } | |
6638 | ||
3347fa09 TH |
6639 | /** |
6640 | * workqueue_init - bring workqueue subsystem fully online | |
6641 | * | |
2930155b TH |
6642 | * This is the second step of three-staged workqueue subsystem initialization |
6643 | * and invoked as soon as kthreads can be created and scheduled. Workqueues have | |
6644 | * been created and work items queued on them, but there are no kworkers | |
6645 | * executing the work items yet. Populate the worker pools with the initial | |
6646 | * workers and enable future kworker creations. | |
3347fa09 | 6647 | */ |
2333e829 | 6648 | void __init workqueue_init(void) |
3347fa09 | 6649 | { |
2186d9f9 | 6650 | struct workqueue_struct *wq; |
3347fa09 TH |
6651 | struct worker_pool *pool; |
6652 | int cpu, bkt; | |
6653 | ||
aa6fde93 TH |
6654 | wq_cpu_intensive_thresh_init(); |
6655 | ||
2186d9f9 TH |
6656 | mutex_lock(&wq_pool_mutex); |
6657 | ||
2930155b TH |
6658 | /* |
6659 | * Per-cpu pools created earlier could be missing node hint. Fix them | |
6660 | * up. Also, create a rescuer for workqueues that requested it. | |
6661 | */ | |
2186d9f9 TH |
6662 | for_each_possible_cpu(cpu) { |
6663 | for_each_cpu_worker_pool(pool, cpu) { | |
6664 | pool->node = cpu_to_node(cpu); | |
6665 | } | |
6666 | } | |
6667 | ||
40c17f75 | 6668 | list_for_each_entry(wq, &workqueues, list) { |
40c17f75 TH |
6669 | WARN(init_rescuer(wq), |
6670 | "workqueue: failed to create early rescuer for %s", | |
6671 | wq->name); | |
6672 | } | |
2186d9f9 TH |
6673 | |
6674 | mutex_unlock(&wq_pool_mutex); | |
6675 | ||
3347fa09 TH |
6676 | /* create the initial workers */ |
6677 | for_each_online_cpu(cpu) { | |
6678 | for_each_cpu_worker_pool(pool, cpu) { | |
6679 | pool->flags &= ~POOL_DISASSOCIATED; | |
6680 | BUG_ON(!create_worker(pool)); | |
6681 | } | |
6682 | } | |
6683 | ||
6684 | hash_for_each(unbound_pool_hash, bkt, pool, hash_node) | |
6685 | BUG_ON(!create_worker(pool)); | |
6686 | ||
6687 | wq_online = true; | |
82607adc | 6688 | wq_watchdog_init(); |
1da177e4 | 6689 | } |
c4f135d6 | 6690 | |
025e1684 TH |
6691 | /* |
6692 | * Initialize @pt by first initializing @pt->cpu_pod[] with pod IDs according to | |
6693 | * @cpu_shares_pod(). Each subset of CPUs that share a pod is assigned a unique | |
6694 | * and consecutive pod ID. The rest of @pt is initialized accordingly. | |
6695 | */ | |
6696 | static void __init init_pod_type(struct wq_pod_type *pt, | |
6697 | bool (*cpus_share_pod)(int, int)) | |
6698 | { | |
6699 | int cur, pre, cpu, pod; | |
6700 | ||
6701 | pt->nr_pods = 0; | |
6702 | ||
6703 | /* init @pt->cpu_pod[] according to @cpus_share_pod() */ | |
6704 | pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL); | |
6705 | BUG_ON(!pt->cpu_pod); | |
6706 | ||
6707 | for_each_possible_cpu(cur) { | |
6708 | for_each_possible_cpu(pre) { | |
6709 | if (pre >= cur) { | |
6710 | pt->cpu_pod[cur] = pt->nr_pods++; | |
6711 | break; | |
6712 | } | |
6713 | if (cpus_share_pod(cur, pre)) { | |
6714 | pt->cpu_pod[cur] = pt->cpu_pod[pre]; | |
6715 | break; | |
6716 | } | |
6717 | } | |
6718 | } | |
6719 | ||
6720 | /* init the rest to match @pt->cpu_pod[] */ | |
6721 | pt->pod_cpus = kcalloc(pt->nr_pods, sizeof(pt->pod_cpus[0]), GFP_KERNEL); | |
6722 | pt->pod_node = kcalloc(pt->nr_pods, sizeof(pt->pod_node[0]), GFP_KERNEL); | |
6723 | BUG_ON(!pt->pod_cpus || !pt->pod_node); | |
6724 | ||
6725 | for (pod = 0; pod < pt->nr_pods; pod++) | |
6726 | BUG_ON(!zalloc_cpumask_var(&pt->pod_cpus[pod], GFP_KERNEL)); | |
6727 | ||
6728 | for_each_possible_cpu(cpu) { | |
6729 | cpumask_set_cpu(cpu, pt->pod_cpus[pt->cpu_pod[cpu]]); | |
6730 | pt->pod_node[pt->cpu_pod[cpu]] = cpu_to_node(cpu); | |
6731 | } | |
6732 | } | |
6733 | ||
63c5484e TH |
6734 | static bool __init cpus_dont_share(int cpu0, int cpu1) |
6735 | { | |
6736 | return false; | |
6737 | } | |
6738 | ||
6739 | static bool __init cpus_share_smt(int cpu0, int cpu1) | |
6740 | { | |
6741 | #ifdef CONFIG_SCHED_SMT | |
6742 | return cpumask_test_cpu(cpu0, cpu_smt_mask(cpu1)); | |
6743 | #else | |
6744 | return false; | |
6745 | #endif | |
6746 | } | |
6747 | ||
025e1684 TH |
6748 | static bool __init cpus_share_numa(int cpu0, int cpu1) |
6749 | { | |
6750 | return cpu_to_node(cpu0) == cpu_to_node(cpu1); | |
6751 | } | |
6752 | ||
2930155b TH |
6753 | /** |
6754 | * workqueue_init_topology - initialize CPU pods for unbound workqueues | |
6755 | * | |
6756 | * This is the third step of there-staged workqueue subsystem initialization and | |
6757 | * invoked after SMP and topology information are fully initialized. It | |
6758 | * initializes the unbound CPU pods accordingly. | |
6759 | */ | |
6760 | void __init workqueue_init_topology(void) | |
a86feae6 | 6761 | { |
2930155b | 6762 | struct workqueue_struct *wq; |
025e1684 | 6763 | int cpu; |
a86feae6 | 6764 | |
63c5484e TH |
6765 | init_pod_type(&wq_pod_types[WQ_AFFN_CPU], cpus_dont_share); |
6766 | init_pod_type(&wq_pod_types[WQ_AFFN_SMT], cpus_share_smt); | |
6767 | init_pod_type(&wq_pod_types[WQ_AFFN_CACHE], cpus_share_cache); | |
025e1684 | 6768 | init_pod_type(&wq_pod_types[WQ_AFFN_NUMA], cpus_share_numa); |
a86feae6 | 6769 | |
2930155b | 6770 | mutex_lock(&wq_pool_mutex); |
a86feae6 | 6771 | |
2930155b TH |
6772 | /* |
6773 | * Workqueues allocated earlier would have all CPUs sharing the default | |
6774 | * worker pool. Explicitly call wq_update_pod() on all workqueue and CPU | |
6775 | * combinations to apply per-pod sharing. | |
6776 | */ | |
6777 | list_for_each_entry(wq, &workqueues, list) { | |
6778 | for_each_online_cpu(cpu) { | |
6779 | wq_update_pod(wq, cpu, cpu, true); | |
6780 | } | |
6781 | } | |
6782 | ||
6783 | mutex_unlock(&wq_pool_mutex); | |
a86feae6 TH |
6784 | } |
6785 | ||
20bdedaf TH |
6786 | void __warn_flushing_systemwide_wq(void) |
6787 | { | |
6788 | pr_warn("WARNING: Flushing system-wide workqueues will be prohibited in near future.\n"); | |
6789 | dump_stack(); | |
6790 | } | |
c4f135d6 | 6791 | EXPORT_SYMBOL(__warn_flushing_systemwide_wq); |
ace3c549 | 6792 | |
6793 | static int __init workqueue_unbound_cpus_setup(char *str) | |
6794 | { | |
6795 | if (cpulist_parse(str, &wq_cmdline_cpumask) < 0) { | |
6796 | cpumask_clear(&wq_cmdline_cpumask); | |
6797 | pr_warn("workqueue.unbound_cpus: incorrect CPU range, using default\n"); | |
6798 | } | |
6799 | ||
6800 | return 1; | |
6801 | } | |
6802 | __setup("workqueue.unbound_cpus=", workqueue_unbound_cpus_setup); |