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