]>
Commit | Line | Data |
---|---|---|
1da177e4 | 1 | /* |
c54fce6e | 2 | * kernel/workqueue.c - generic async execution with shared worker pool |
1da177e4 | 3 | * |
c54fce6e | 4 | * Copyright (C) 2002 Ingo Molnar |
1da177e4 | 5 | * |
c54fce6e TH |
6 | * Derived from the taskqueue/keventd code by: |
7 | * David Woodhouse <dwmw2@infradead.org> | |
8 | * Andrew Morton | |
9 | * Kai Petzke <wpp@marie.physik.tu-berlin.de> | |
10 | * Theodore Ts'o <tytso@mit.edu> | |
1da177e4 | 11 | * |
c54fce6e | 12 | * Made to use alloc_percpu by Christoph Lameter. |
1da177e4 | 13 | * |
c54fce6e TH |
14 | * Copyright (C) 2010 SUSE Linux Products GmbH |
15 | * Copyright (C) 2010 Tejun Heo <tj@kernel.org> | |
89ada679 | 16 | * |
c54fce6e TH |
17 | * This is the generic async execution mechanism. Work items as are |
18 | * executed in process context. The worker pool is shared and | |
b11895c4 L |
19 | * automatically managed. There are two worker pools for each CPU (one for |
20 | * normal work items and the other for high priority ones) and some extra | |
21 | * pools for workqueues which are not bound to any specific CPU - the | |
22 | * number of these backing pools is dynamic. | |
c54fce6e | 23 | * |
9a261491 | 24 | * Please read Documentation/core-api/workqueue.rst for details. |
1da177e4 LT |
25 | */ |
26 | ||
9984de1a | 27 | #include <linux/export.h> |
1da177e4 LT |
28 | #include <linux/kernel.h> |
29 | #include <linux/sched.h> | |
30 | #include <linux/init.h> | |
31 | #include <linux/signal.h> | |
32 | #include <linux/completion.h> | |
33 | #include <linux/workqueue.h> | |
34 | #include <linux/slab.h> | |
35 | #include <linux/cpu.h> | |
36 | #include <linux/notifier.h> | |
37 | #include <linux/kthread.h> | |
1fa44eca | 38 | #include <linux/hardirq.h> |
46934023 | 39 | #include <linux/mempolicy.h> |
341a5958 | 40 | #include <linux/freezer.h> |
d5abe669 | 41 | #include <linux/debug_locks.h> |
4e6045f1 | 42 | #include <linux/lockdep.h> |
c34056a3 | 43 | #include <linux/idr.h> |
29c91e99 | 44 | #include <linux/jhash.h> |
42f8570f | 45 | #include <linux/hashtable.h> |
76af4d93 | 46 | #include <linux/rculist.h> |
bce90380 | 47 | #include <linux/nodemask.h> |
4c16bd32 | 48 | #include <linux/moduleparam.h> |
3d1cb205 | 49 | #include <linux/uaccess.h> |
c98a9805 | 50 | #include <linux/sched/isolation.h> |
62635ea8 | 51 | #include <linux/nmi.h> |
e22bee78 | 52 | |
ea138446 | 53 | #include "workqueue_internal.h" |
1da177e4 | 54 | |
c8e55f36 | 55 | enum { |
24647570 TH |
56 | /* |
57 | * worker_pool flags | |
bc2ae0f5 | 58 | * |
24647570 | 59 | * A bound pool is either associated or disassociated with its CPU. |
bc2ae0f5 TH |
60 | * While associated (!DISASSOCIATED), all workers are bound to the |
61 | * CPU and none has %WORKER_UNBOUND set and concurrency management | |
62 | * is in effect. | |
63 | * | |
64 | * While DISASSOCIATED, the cpu may be offline and all workers have | |
65 | * %WORKER_UNBOUND set and concurrency management disabled, and may | |
24647570 | 66 | * be executing on any CPU. The pool behaves as an unbound one. |
bc2ae0f5 | 67 | * |
bc3a1afc | 68 | * Note that DISASSOCIATED should be flipped only while holding |
1258fae7 | 69 | * wq_pool_attach_mutex to avoid changing binding state while |
4736cbf7 | 70 | * worker_attach_to_pool() is in progress. |
bc2ae0f5 | 71 | */ |
692b4825 | 72 | POOL_MANAGER_ACTIVE = 1 << 0, /* being managed */ |
24647570 | 73 | POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */ |
db7bccf4 | 74 | |
c8e55f36 | 75 | /* worker flags */ |
c8e55f36 TH |
76 | WORKER_DIE = 1 << 1, /* die die die */ |
77 | WORKER_IDLE = 1 << 2, /* is idle */ | |
e22bee78 | 78 | WORKER_PREP = 1 << 3, /* preparing to run works */ |
fb0e7beb | 79 | WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */ |
f3421797 | 80 | WORKER_UNBOUND = 1 << 7, /* worker is unbound */ |
a9ab775b | 81 | WORKER_REBOUND = 1 << 8, /* worker was rebound */ |
e22bee78 | 82 | |
a9ab775b TH |
83 | WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE | |
84 | WORKER_UNBOUND | WORKER_REBOUND, | |
db7bccf4 | 85 | |
e34cdddb | 86 | NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */ |
4ce62e9e | 87 | |
29c91e99 | 88 | UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */ |
c8e55f36 | 89 | BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */ |
db7bccf4 | 90 | |
e22bee78 TH |
91 | MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */ |
92 | IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */ | |
93 | ||
3233cdbd TH |
94 | MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2, |
95 | /* call for help after 10ms | |
96 | (min two ticks) */ | |
e22bee78 TH |
97 | MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */ |
98 | CREATE_COOLDOWN = HZ, /* time to breath after fail */ | |
e22bee78 TH |
99 | |
100 | /* | |
101 | * Rescue workers are used only on emergencies and shared by | |
8698a745 | 102 | * all cpus. Give MIN_NICE. |
e22bee78 | 103 | */ |
8698a745 DY |
104 | RESCUER_NICE_LEVEL = MIN_NICE, |
105 | HIGHPRI_NICE_LEVEL = MIN_NICE, | |
ecf6881f TH |
106 | |
107 | WQ_NAME_LEN = 24, | |
c8e55f36 | 108 | }; |
1da177e4 LT |
109 | |
110 | /* | |
4690c4ab TH |
111 | * Structure fields follow one of the following exclusion rules. |
112 | * | |
e41e704b TH |
113 | * I: Modifiable by initialization/destruction paths and read-only for |
114 | * everyone else. | |
4690c4ab | 115 | * |
e22bee78 TH |
116 | * P: Preemption protected. Disabling preemption is enough and should |
117 | * only be modified and accessed from the local cpu. | |
118 | * | |
d565ed63 | 119 | * L: pool->lock protected. Access with pool->lock held. |
4690c4ab | 120 | * |
d565ed63 TH |
121 | * X: During normal operation, modification requires pool->lock and should |
122 | * be done only from local cpu. Either disabling preemption on local | |
123 | * cpu or grabbing pool->lock is enough for read access. If | |
124 | * POOL_DISASSOCIATED is set, it's identical to L. | |
e22bee78 | 125 | * |
1258fae7 | 126 | * A: wq_pool_attach_mutex protected. |
822d8405 | 127 | * |
68e13a67 | 128 | * PL: wq_pool_mutex protected. |
5bcab335 | 129 | * |
68e13a67 | 130 | * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads. |
76af4d93 | 131 | * |
5b95e1af LJ |
132 | * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads. |
133 | * | |
134 | * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or | |
135 | * sched-RCU for reads. | |
136 | * | |
3c25a55d LJ |
137 | * WQ: wq->mutex protected. |
138 | * | |
b5927605 | 139 | * WR: wq->mutex protected for writes. Sched-RCU protected for reads. |
2e109a28 TH |
140 | * |
141 | * MD: wq_mayday_lock protected. | |
1da177e4 | 142 | */ |
1da177e4 | 143 | |
2eaebdb3 | 144 | /* struct worker is defined in workqueue_internal.h */ |
c34056a3 | 145 | |
bd7bdd43 | 146 | struct worker_pool { |
d565ed63 | 147 | spinlock_t lock; /* the pool lock */ |
d84ff051 | 148 | int cpu; /* I: the associated cpu */ |
f3f90ad4 | 149 | int node; /* I: the associated node ID */ |
9daf9e67 | 150 | int id; /* I: pool ID */ |
11ebea50 | 151 | unsigned int flags; /* X: flags */ |
bd7bdd43 | 152 | |
82607adc TH |
153 | unsigned long watchdog_ts; /* L: watchdog timestamp */ |
154 | ||
bd7bdd43 | 155 | struct list_head worklist; /* L: list of pending works */ |
ea1abd61 | 156 | |
5826cc8f LJ |
157 | int nr_workers; /* L: total number of workers */ |
158 | int nr_idle; /* L: currently idle workers */ | |
bd7bdd43 TH |
159 | |
160 | struct list_head idle_list; /* X: list of idle workers */ | |
161 | struct timer_list idle_timer; /* L: worker idle timeout */ | |
162 | struct timer_list mayday_timer; /* L: SOS timer for workers */ | |
163 | ||
c5aa87bb | 164 | /* a workers is either on busy_hash or idle_list, or the manager */ |
c9e7cf27 TH |
165 | DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER); |
166 | /* L: hash of busy workers */ | |
167 | ||
2607d7a6 | 168 | struct worker *manager; /* L: purely informational */ |
92f9c5c4 | 169 | struct list_head workers; /* A: attached workers */ |
60f5a4bc | 170 | struct completion *detach_completion; /* all workers detached */ |
e19e397a | 171 | |
7cda9aae | 172 | struct ida worker_ida; /* worker IDs for task name */ |
e19e397a | 173 | |
7a4e344c | 174 | struct workqueue_attrs *attrs; /* I: worker attributes */ |
68e13a67 LJ |
175 | struct hlist_node hash_node; /* PL: unbound_pool_hash node */ |
176 | int refcnt; /* PL: refcnt for unbound pools */ | |
7a4e344c | 177 | |
e19e397a TH |
178 | /* |
179 | * The current concurrency level. As it's likely to be accessed | |
180 | * from other CPUs during try_to_wake_up(), put it in a separate | |
181 | * cacheline. | |
182 | */ | |
183 | atomic_t nr_running ____cacheline_aligned_in_smp; | |
29c91e99 TH |
184 | |
185 | /* | |
186 | * Destruction of pool is sched-RCU protected to allow dereferences | |
187 | * from get_work_pool(). | |
188 | */ | |
189 | struct rcu_head rcu; | |
8b03ae3c TH |
190 | } ____cacheline_aligned_in_smp; |
191 | ||
1da177e4 | 192 | /* |
112202d9 TH |
193 | * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS |
194 | * of work_struct->data are used for flags and the remaining high bits | |
195 | * point to the pwq; thus, pwqs need to be aligned at two's power of the | |
196 | * number of flag bits. | |
1da177e4 | 197 | */ |
112202d9 | 198 | struct pool_workqueue { |
bd7bdd43 | 199 | struct worker_pool *pool; /* I: the associated pool */ |
4690c4ab | 200 | struct workqueue_struct *wq; /* I: the owning workqueue */ |
73f53c4a TH |
201 | int work_color; /* L: current color */ |
202 | int flush_color; /* L: flushing color */ | |
8864b4e5 | 203 | int refcnt; /* L: reference count */ |
73f53c4a TH |
204 | int nr_in_flight[WORK_NR_COLORS]; |
205 | /* L: nr of in_flight works */ | |
1e19ffc6 | 206 | int nr_active; /* L: nr of active works */ |
a0a1a5fd | 207 | int max_active; /* L: max active works */ |
1e19ffc6 | 208 | struct list_head delayed_works; /* L: delayed works */ |
3c25a55d | 209 | struct list_head pwqs_node; /* WR: node on wq->pwqs */ |
2e109a28 | 210 | struct list_head mayday_node; /* MD: node on wq->maydays */ |
8864b4e5 TH |
211 | |
212 | /* | |
213 | * Release of unbound pwq is punted to system_wq. See put_pwq() | |
214 | * and pwq_unbound_release_workfn() for details. pool_workqueue | |
215 | * itself is also sched-RCU protected so that the first pwq can be | |
b09f4fd3 | 216 | * determined without grabbing wq->mutex. |
8864b4e5 TH |
217 | */ |
218 | struct work_struct unbound_release_work; | |
219 | struct rcu_head rcu; | |
e904e6c2 | 220 | } __aligned(1 << WORK_STRUCT_FLAG_BITS); |
1da177e4 | 221 | |
73f53c4a TH |
222 | /* |
223 | * Structure used to wait for workqueue flush. | |
224 | */ | |
225 | struct wq_flusher { | |
3c25a55d LJ |
226 | struct list_head list; /* WQ: list of flushers */ |
227 | int flush_color; /* WQ: flush color waiting for */ | |
73f53c4a TH |
228 | struct completion done; /* flush completion */ |
229 | }; | |
230 | ||
226223ab TH |
231 | struct wq_device; |
232 | ||
1da177e4 | 233 | /* |
c5aa87bb TH |
234 | * The externally visible workqueue. It relays the issued work items to |
235 | * the appropriate worker_pool through its pool_workqueues. | |
1da177e4 LT |
236 | */ |
237 | struct workqueue_struct { | |
3c25a55d | 238 | struct list_head pwqs; /* WR: all pwqs of this wq */ |
e2dca7ad | 239 | struct list_head list; /* PR: list of all workqueues */ |
73f53c4a | 240 | |
3c25a55d LJ |
241 | struct mutex mutex; /* protects this wq */ |
242 | int work_color; /* WQ: current work color */ | |
243 | int flush_color; /* WQ: current flush color */ | |
112202d9 | 244 | atomic_t nr_pwqs_to_flush; /* flush in progress */ |
3c25a55d LJ |
245 | struct wq_flusher *first_flusher; /* WQ: first flusher */ |
246 | struct list_head flusher_queue; /* WQ: flush waiters */ | |
247 | struct list_head flusher_overflow; /* WQ: flush overflow list */ | |
73f53c4a | 248 | |
2e109a28 | 249 | struct list_head maydays; /* MD: pwqs requesting rescue */ |
e22bee78 TH |
250 | struct worker *rescuer; /* I: rescue worker */ |
251 | ||
87fc741e | 252 | int nr_drainers; /* WQ: drain in progress */ |
a357fc03 | 253 | int saved_max_active; /* WQ: saved pwq max_active */ |
226223ab | 254 | |
5b95e1af LJ |
255 | struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */ |
256 | struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */ | |
6029a918 | 257 | |
226223ab TH |
258 | #ifdef CONFIG_SYSFS |
259 | struct wq_device *wq_dev; /* I: for sysfs interface */ | |
260 | #endif | |
4e6045f1 | 261 | #ifdef CONFIG_LOCKDEP |
669de8bd BVA |
262 | char *lock_name; |
263 | struct lock_class_key key; | |
4690c4ab | 264 | struct lockdep_map lockdep_map; |
4e6045f1 | 265 | #endif |
ecf6881f | 266 | char name[WQ_NAME_LEN]; /* I: workqueue name */ |
2728fd2f | 267 | |
e2dca7ad TH |
268 | /* |
269 | * Destruction of workqueue_struct is sched-RCU protected to allow | |
270 | * walking the workqueues list without grabbing wq_pool_mutex. | |
271 | * This is used to dump all workqueues from sysrq. | |
272 | */ | |
273 | struct rcu_head rcu; | |
274 | ||
2728fd2f TH |
275 | /* hot fields used during command issue, aligned to cacheline */ |
276 | unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */ | |
277 | struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */ | |
5b95e1af | 278 | struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */ |
1da177e4 LT |
279 | }; |
280 | ||
e904e6c2 TH |
281 | static struct kmem_cache *pwq_cache; |
282 | ||
bce90380 TH |
283 | static cpumask_var_t *wq_numa_possible_cpumask; |
284 | /* possible CPUs of each node */ | |
285 | ||
d55262c4 TH |
286 | static bool wq_disable_numa; |
287 | module_param_named(disable_numa, wq_disable_numa, bool, 0444); | |
288 | ||
cee22a15 | 289 | /* see the comment above the definition of WQ_POWER_EFFICIENT */ |
552f530c | 290 | static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT); |
cee22a15 VK |
291 | module_param_named(power_efficient, wq_power_efficient, bool, 0444); |
292 | ||
863b710b | 293 | static bool wq_online; /* can kworkers be created yet? */ |
3347fa09 | 294 | |
bce90380 TH |
295 | static bool wq_numa_enabled; /* unbound NUMA affinity enabled */ |
296 | ||
4c16bd32 TH |
297 | /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */ |
298 | static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf; | |
299 | ||
68e13a67 | 300 | static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ |
1258fae7 | 301 | static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */ |
2e109a28 | 302 | static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ |
692b4825 | 303 | static DECLARE_WAIT_QUEUE_HEAD(wq_manager_wait); /* wait for manager to go away */ |
5bcab335 | 304 | |
e2dca7ad | 305 | static LIST_HEAD(workqueues); /* PR: list of all workqueues */ |
68e13a67 | 306 | static bool workqueue_freezing; /* PL: have wqs started freezing? */ |
7d19c5ce | 307 | |
ef557180 MG |
308 | /* PL: allowable cpus for unbound wqs and work items */ |
309 | static cpumask_var_t wq_unbound_cpumask; | |
310 | ||
311 | /* CPU where unbound work was last round robin scheduled from this CPU */ | |
312 | static DEFINE_PER_CPU(int, wq_rr_cpu_last); | |
b05a7928 | 313 | |
f303fccb TH |
314 | /* |
315 | * Local execution of unbound work items is no longer guaranteed. The | |
316 | * following always forces round-robin CPU selection on unbound work items | |
317 | * to uncover usages which depend on it. | |
318 | */ | |
319 | #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU | |
320 | static bool wq_debug_force_rr_cpu = true; | |
321 | #else | |
322 | static bool wq_debug_force_rr_cpu = false; | |
323 | #endif | |
324 | module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644); | |
325 | ||
7d19c5ce | 326 | /* the per-cpu worker pools */ |
25528213 | 327 | static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools); |
7d19c5ce | 328 | |
68e13a67 | 329 | static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ |
7d19c5ce | 330 | |
68e13a67 | 331 | /* PL: hash of all unbound pools keyed by pool->attrs */ |
29c91e99 TH |
332 | static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); |
333 | ||
c5aa87bb | 334 | /* I: attributes used when instantiating standard unbound pools on demand */ |
29c91e99 TH |
335 | static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; |
336 | ||
8a2b7538 TH |
337 | /* I: attributes used when instantiating ordered pools on demand */ |
338 | static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS]; | |
339 | ||
d320c038 | 340 | struct workqueue_struct *system_wq __read_mostly; |
ad7b1f84 | 341 | EXPORT_SYMBOL(system_wq); |
044c782c | 342 | struct workqueue_struct *system_highpri_wq __read_mostly; |
1aabe902 | 343 | EXPORT_SYMBOL_GPL(system_highpri_wq); |
044c782c | 344 | struct workqueue_struct *system_long_wq __read_mostly; |
d320c038 | 345 | EXPORT_SYMBOL_GPL(system_long_wq); |
044c782c | 346 | struct workqueue_struct *system_unbound_wq __read_mostly; |
f3421797 | 347 | EXPORT_SYMBOL_GPL(system_unbound_wq); |
044c782c | 348 | struct workqueue_struct *system_freezable_wq __read_mostly; |
24d51add | 349 | EXPORT_SYMBOL_GPL(system_freezable_wq); |
0668106c VK |
350 | struct workqueue_struct *system_power_efficient_wq __read_mostly; |
351 | EXPORT_SYMBOL_GPL(system_power_efficient_wq); | |
352 | struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly; | |
353 | EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq); | |
d320c038 | 354 | |
7d19c5ce | 355 | static int worker_thread(void *__worker); |
6ba94429 | 356 | static void workqueue_sysfs_unregister(struct workqueue_struct *wq); |
7d19c5ce | 357 | |
97bd2347 TH |
358 | #define CREATE_TRACE_POINTS |
359 | #include <trace/events/workqueue.h> | |
360 | ||
68e13a67 | 361 | #define assert_rcu_or_pool_mutex() \ |
f78f5b90 PM |
362 | RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \ |
363 | !lockdep_is_held(&wq_pool_mutex), \ | |
364 | "sched RCU or wq_pool_mutex should be held") | |
5bcab335 | 365 | |
b09f4fd3 | 366 | #define assert_rcu_or_wq_mutex(wq) \ |
f78f5b90 PM |
367 | RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \ |
368 | !lockdep_is_held(&wq->mutex), \ | |
369 | "sched RCU or wq->mutex should be held") | |
76af4d93 | 370 | |
5b95e1af | 371 | #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \ |
f78f5b90 PM |
372 | RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \ |
373 | !lockdep_is_held(&wq->mutex) && \ | |
374 | !lockdep_is_held(&wq_pool_mutex), \ | |
375 | "sched RCU, wq->mutex or wq_pool_mutex should be held") | |
5b95e1af | 376 | |
f02ae73a TH |
377 | #define for_each_cpu_worker_pool(pool, cpu) \ |
378 | for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ | |
379 | (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ | |
7a62c2c8 | 380 | (pool)++) |
4ce62e9e | 381 | |
17116969 TH |
382 | /** |
383 | * for_each_pool - iterate through all worker_pools in the system | |
384 | * @pool: iteration cursor | |
611c92a0 | 385 | * @pi: integer used for iteration |
fa1b54e6 | 386 | * |
68e13a67 LJ |
387 | * This must be called either with wq_pool_mutex held or sched RCU read |
388 | * locked. If the pool needs to be used beyond the locking in effect, the | |
389 | * caller is responsible for guaranteeing that the pool stays online. | |
fa1b54e6 TH |
390 | * |
391 | * The if/else clause exists only for the lockdep assertion and can be | |
392 | * ignored. | |
17116969 | 393 | */ |
611c92a0 TH |
394 | #define for_each_pool(pool, pi) \ |
395 | idr_for_each_entry(&worker_pool_idr, pool, pi) \ | |
68e13a67 | 396 | if (({ assert_rcu_or_pool_mutex(); false; })) { } \ |
fa1b54e6 | 397 | else |
17116969 | 398 | |
822d8405 TH |
399 | /** |
400 | * for_each_pool_worker - iterate through all workers of a worker_pool | |
401 | * @worker: iteration cursor | |
822d8405 TH |
402 | * @pool: worker_pool to iterate workers of |
403 | * | |
1258fae7 | 404 | * This must be called with wq_pool_attach_mutex. |
822d8405 TH |
405 | * |
406 | * The if/else clause exists only for the lockdep assertion and can be | |
407 | * ignored. | |
408 | */ | |
da028469 LJ |
409 | #define for_each_pool_worker(worker, pool) \ |
410 | list_for_each_entry((worker), &(pool)->workers, node) \ | |
1258fae7 | 411 | if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \ |
822d8405 TH |
412 | else |
413 | ||
49e3cf44 TH |
414 | /** |
415 | * for_each_pwq - iterate through all pool_workqueues of the specified workqueue | |
416 | * @pwq: iteration cursor | |
417 | * @wq: the target workqueue | |
76af4d93 | 418 | * |
b09f4fd3 | 419 | * This must be called either with wq->mutex held or sched RCU read locked. |
794b18bc TH |
420 | * If the pwq needs to be used beyond the locking in effect, the caller is |
421 | * responsible for guaranteeing that the pwq stays online. | |
76af4d93 TH |
422 | * |
423 | * The if/else clause exists only for the lockdep assertion and can be | |
424 | * ignored. | |
49e3cf44 TH |
425 | */ |
426 | #define for_each_pwq(pwq, wq) \ | |
76af4d93 | 427 | list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \ |
b09f4fd3 | 428 | if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \ |
76af4d93 | 429 | else |
f3421797 | 430 | |
dc186ad7 TG |
431 | #ifdef CONFIG_DEBUG_OBJECTS_WORK |
432 | ||
433 | static struct debug_obj_descr work_debug_descr; | |
434 | ||
99777288 SG |
435 | static void *work_debug_hint(void *addr) |
436 | { | |
437 | return ((struct work_struct *) addr)->func; | |
438 | } | |
439 | ||
b9fdac7f CD |
440 | static bool work_is_static_object(void *addr) |
441 | { | |
442 | struct work_struct *work = addr; | |
443 | ||
444 | return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work)); | |
445 | } | |
446 | ||
dc186ad7 TG |
447 | /* |
448 | * fixup_init is called when: | |
449 | * - an active object is initialized | |
450 | */ | |
02a982a6 | 451 | static bool work_fixup_init(void *addr, enum debug_obj_state state) |
dc186ad7 TG |
452 | { |
453 | struct work_struct *work = addr; | |
454 | ||
455 | switch (state) { | |
456 | case ODEBUG_STATE_ACTIVE: | |
457 | cancel_work_sync(work); | |
458 | debug_object_init(work, &work_debug_descr); | |
02a982a6 | 459 | return true; |
dc186ad7 | 460 | default: |
02a982a6 | 461 | return false; |
dc186ad7 TG |
462 | } |
463 | } | |
464 | ||
dc186ad7 TG |
465 | /* |
466 | * fixup_free is called when: | |
467 | * - an active object is freed | |
468 | */ | |
02a982a6 | 469 | static bool work_fixup_free(void *addr, enum debug_obj_state state) |
dc186ad7 TG |
470 | { |
471 | struct work_struct *work = addr; | |
472 | ||
473 | switch (state) { | |
474 | case ODEBUG_STATE_ACTIVE: | |
475 | cancel_work_sync(work); | |
476 | debug_object_free(work, &work_debug_descr); | |
02a982a6 | 477 | return true; |
dc186ad7 | 478 | default: |
02a982a6 | 479 | return false; |
dc186ad7 TG |
480 | } |
481 | } | |
482 | ||
483 | static struct debug_obj_descr work_debug_descr = { | |
484 | .name = "work_struct", | |
99777288 | 485 | .debug_hint = work_debug_hint, |
b9fdac7f | 486 | .is_static_object = work_is_static_object, |
dc186ad7 | 487 | .fixup_init = work_fixup_init, |
dc186ad7 TG |
488 | .fixup_free = work_fixup_free, |
489 | }; | |
490 | ||
491 | static inline void debug_work_activate(struct work_struct *work) | |
492 | { | |
493 | debug_object_activate(work, &work_debug_descr); | |
494 | } | |
495 | ||
496 | static inline void debug_work_deactivate(struct work_struct *work) | |
497 | { | |
498 | debug_object_deactivate(work, &work_debug_descr); | |
499 | } | |
500 | ||
501 | void __init_work(struct work_struct *work, int onstack) | |
502 | { | |
503 | if (onstack) | |
504 | debug_object_init_on_stack(work, &work_debug_descr); | |
505 | else | |
506 | debug_object_init(work, &work_debug_descr); | |
507 | } | |
508 | EXPORT_SYMBOL_GPL(__init_work); | |
509 | ||
510 | void destroy_work_on_stack(struct work_struct *work) | |
511 | { | |
512 | debug_object_free(work, &work_debug_descr); | |
513 | } | |
514 | EXPORT_SYMBOL_GPL(destroy_work_on_stack); | |
515 | ||
ea2e64f2 TG |
516 | void destroy_delayed_work_on_stack(struct delayed_work *work) |
517 | { | |
518 | destroy_timer_on_stack(&work->timer); | |
519 | debug_object_free(&work->work, &work_debug_descr); | |
520 | } | |
521 | EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack); | |
522 | ||
dc186ad7 TG |
523 | #else |
524 | static inline void debug_work_activate(struct work_struct *work) { } | |
525 | static inline void debug_work_deactivate(struct work_struct *work) { } | |
526 | #endif | |
527 | ||
4e8b22bd LB |
528 | /** |
529 | * worker_pool_assign_id - allocate ID and assing it to @pool | |
530 | * @pool: the pool pointer of interest | |
531 | * | |
532 | * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned | |
533 | * successfully, -errno on failure. | |
534 | */ | |
9daf9e67 TH |
535 | static int worker_pool_assign_id(struct worker_pool *pool) |
536 | { | |
537 | int ret; | |
538 | ||
68e13a67 | 539 | lockdep_assert_held(&wq_pool_mutex); |
5bcab335 | 540 | |
4e8b22bd LB |
541 | ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE, |
542 | GFP_KERNEL); | |
229641a6 | 543 | if (ret >= 0) { |
e68035fb | 544 | pool->id = ret; |
229641a6 TH |
545 | return 0; |
546 | } | |
fa1b54e6 | 547 | return ret; |
7c3eed5c TH |
548 | } |
549 | ||
df2d5ae4 TH |
550 | /** |
551 | * unbound_pwq_by_node - return the unbound pool_workqueue for the given node | |
552 | * @wq: the target workqueue | |
553 | * @node: the node ID | |
554 | * | |
5b95e1af LJ |
555 | * This must be called with any of wq_pool_mutex, wq->mutex or sched RCU |
556 | * read locked. | |
df2d5ae4 TH |
557 | * If the pwq needs to be used beyond the locking in effect, the caller is |
558 | * responsible for guaranteeing that the pwq stays online. | |
d185af30 YB |
559 | * |
560 | * Return: The unbound pool_workqueue for @node. | |
df2d5ae4 TH |
561 | */ |
562 | static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq, | |
563 | int node) | |
564 | { | |
5b95e1af | 565 | assert_rcu_or_wq_mutex_or_pool_mutex(wq); |
d6e022f1 TH |
566 | |
567 | /* | |
568 | * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a | |
569 | * delayed item is pending. The plan is to keep CPU -> NODE | |
570 | * mapping valid and stable across CPU on/offlines. Once that | |
571 | * happens, this workaround can be removed. | |
572 | */ | |
573 | if (unlikely(node == NUMA_NO_NODE)) | |
574 | return wq->dfl_pwq; | |
575 | ||
df2d5ae4 TH |
576 | return rcu_dereference_raw(wq->numa_pwq_tbl[node]); |
577 | } | |
578 | ||
73f53c4a TH |
579 | static unsigned int work_color_to_flags(int color) |
580 | { | |
581 | return color << WORK_STRUCT_COLOR_SHIFT; | |
582 | } | |
583 | ||
584 | static int get_work_color(struct work_struct *work) | |
585 | { | |
586 | return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) & | |
587 | ((1 << WORK_STRUCT_COLOR_BITS) - 1); | |
588 | } | |
589 | ||
590 | static int work_next_color(int color) | |
591 | { | |
592 | return (color + 1) % WORK_NR_COLORS; | |
593 | } | |
1da177e4 | 594 | |
14441960 | 595 | /* |
112202d9 TH |
596 | * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data |
597 | * contain the pointer to the queued pwq. Once execution starts, the flag | |
7c3eed5c | 598 | * is cleared and the high bits contain OFFQ flags and pool ID. |
7a22ad75 | 599 | * |
112202d9 TH |
600 | * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() |
601 | * and clear_work_data() can be used to set the pwq, pool or clear | |
bbb68dfa TH |
602 | * work->data. These functions should only be called while the work is |
603 | * owned - ie. while the PENDING bit is set. | |
7a22ad75 | 604 | * |
112202d9 | 605 | * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq |
7c3eed5c | 606 | * corresponding to a work. Pool is available once the work has been |
112202d9 | 607 | * queued anywhere after initialization until it is sync canceled. pwq is |
7c3eed5c | 608 | * available only while the work item is queued. |
7a22ad75 | 609 | * |
bbb68dfa TH |
610 | * %WORK_OFFQ_CANCELING is used to mark a work item which is being |
611 | * canceled. While being canceled, a work item may have its PENDING set | |
612 | * but stay off timer and worklist for arbitrarily long and nobody should | |
613 | * try to steal the PENDING bit. | |
14441960 | 614 | */ |
7a22ad75 TH |
615 | static inline void set_work_data(struct work_struct *work, unsigned long data, |
616 | unsigned long flags) | |
365970a1 | 617 | { |
6183c009 | 618 | WARN_ON_ONCE(!work_pending(work)); |
7a22ad75 TH |
619 | atomic_long_set(&work->data, data | flags | work_static(work)); |
620 | } | |
365970a1 | 621 | |
112202d9 | 622 | static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, |
7a22ad75 TH |
623 | unsigned long extra_flags) |
624 | { | |
112202d9 TH |
625 | set_work_data(work, (unsigned long)pwq, |
626 | WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); | |
365970a1 DH |
627 | } |
628 | ||
4468a00f LJ |
629 | static void set_work_pool_and_keep_pending(struct work_struct *work, |
630 | int pool_id) | |
631 | { | |
632 | set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, | |
633 | WORK_STRUCT_PENDING); | |
634 | } | |
635 | ||
7c3eed5c TH |
636 | static void set_work_pool_and_clear_pending(struct work_struct *work, |
637 | int pool_id) | |
7a22ad75 | 638 | { |
23657bb1 TH |
639 | /* |
640 | * The following wmb is paired with the implied mb in | |
641 | * test_and_set_bit(PENDING) and ensures all updates to @work made | |
642 | * here are visible to and precede any updates by the next PENDING | |
643 | * owner. | |
644 | */ | |
645 | smp_wmb(); | |
7c3eed5c | 646 | set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); |
346c09f8 RP |
647 | /* |
648 | * The following mb guarantees that previous clear of a PENDING bit | |
649 | * will not be reordered with any speculative LOADS or STORES from | |
650 | * work->current_func, which is executed afterwards. This possible | |
651 | * reordering can lead to a missed execution on attempt to qeueue | |
652 | * the same @work. E.g. consider this case: | |
653 | * | |
654 | * CPU#0 CPU#1 | |
655 | * ---------------------------- -------------------------------- | |
656 | * | |
657 | * 1 STORE event_indicated | |
658 | * 2 queue_work_on() { | |
659 | * 3 test_and_set_bit(PENDING) | |
660 | * 4 } set_..._and_clear_pending() { | |
661 | * 5 set_work_data() # clear bit | |
662 | * 6 smp_mb() | |
663 | * 7 work->current_func() { | |
664 | * 8 LOAD event_indicated | |
665 | * } | |
666 | * | |
667 | * Without an explicit full barrier speculative LOAD on line 8 can | |
668 | * be executed before CPU#0 does STORE on line 1. If that happens, | |
669 | * CPU#0 observes the PENDING bit is still set and new execution of | |
670 | * a @work is not queued in a hope, that CPU#1 will eventually | |
671 | * finish the queued @work. Meanwhile CPU#1 does not see | |
672 | * event_indicated is set, because speculative LOAD was executed | |
673 | * before actual STORE. | |
674 | */ | |
675 | smp_mb(); | |
7a22ad75 | 676 | } |
f756d5e2 | 677 | |
7a22ad75 | 678 | static void clear_work_data(struct work_struct *work) |
1da177e4 | 679 | { |
7c3eed5c TH |
680 | smp_wmb(); /* see set_work_pool_and_clear_pending() */ |
681 | set_work_data(work, WORK_STRUCT_NO_POOL, 0); | |
1da177e4 LT |
682 | } |
683 | ||
112202d9 | 684 | static struct pool_workqueue *get_work_pwq(struct work_struct *work) |
b1f4ec17 | 685 | { |
e120153d | 686 | unsigned long data = atomic_long_read(&work->data); |
7a22ad75 | 687 | |
112202d9 | 688 | if (data & WORK_STRUCT_PWQ) |
e120153d TH |
689 | return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); |
690 | else | |
691 | return NULL; | |
4d707b9f ON |
692 | } |
693 | ||
7c3eed5c TH |
694 | /** |
695 | * get_work_pool - return the worker_pool a given work was associated with | |
696 | * @work: the work item of interest | |
697 | * | |
68e13a67 LJ |
698 | * Pools are created and destroyed under wq_pool_mutex, and allows read |
699 | * access under sched-RCU read lock. As such, this function should be | |
700 | * called under wq_pool_mutex or with preemption disabled. | |
fa1b54e6 TH |
701 | * |
702 | * All fields of the returned pool are accessible as long as the above | |
703 | * mentioned locking is in effect. If the returned pool needs to be used | |
704 | * beyond the critical section, the caller is responsible for ensuring the | |
705 | * returned pool is and stays online. | |
d185af30 YB |
706 | * |
707 | * Return: The worker_pool @work was last associated with. %NULL if none. | |
7c3eed5c TH |
708 | */ |
709 | static struct worker_pool *get_work_pool(struct work_struct *work) | |
365970a1 | 710 | { |
e120153d | 711 | unsigned long data = atomic_long_read(&work->data); |
7c3eed5c | 712 | int pool_id; |
7a22ad75 | 713 | |
68e13a67 | 714 | assert_rcu_or_pool_mutex(); |
fa1b54e6 | 715 | |
112202d9 TH |
716 | if (data & WORK_STRUCT_PWQ) |
717 | return ((struct pool_workqueue *) | |
7c3eed5c | 718 | (data & WORK_STRUCT_WQ_DATA_MASK))->pool; |
7a22ad75 | 719 | |
7c3eed5c TH |
720 | pool_id = data >> WORK_OFFQ_POOL_SHIFT; |
721 | if (pool_id == WORK_OFFQ_POOL_NONE) | |
7a22ad75 TH |
722 | return NULL; |
723 | ||
fa1b54e6 | 724 | return idr_find(&worker_pool_idr, pool_id); |
7c3eed5c TH |
725 | } |
726 | ||
727 | /** | |
728 | * get_work_pool_id - return the worker pool ID a given work is associated with | |
729 | * @work: the work item of interest | |
730 | * | |
d185af30 | 731 | * Return: The worker_pool ID @work was last associated with. |
7c3eed5c TH |
732 | * %WORK_OFFQ_POOL_NONE if none. |
733 | */ | |
734 | static int get_work_pool_id(struct work_struct *work) | |
735 | { | |
54d5b7d0 LJ |
736 | unsigned long data = atomic_long_read(&work->data); |
737 | ||
112202d9 TH |
738 | if (data & WORK_STRUCT_PWQ) |
739 | return ((struct pool_workqueue *) | |
54d5b7d0 | 740 | (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id; |
7c3eed5c | 741 | |
54d5b7d0 | 742 | return data >> WORK_OFFQ_POOL_SHIFT; |
7c3eed5c TH |
743 | } |
744 | ||
bbb68dfa TH |
745 | static void mark_work_canceling(struct work_struct *work) |
746 | { | |
7c3eed5c | 747 | unsigned long pool_id = get_work_pool_id(work); |
bbb68dfa | 748 | |
7c3eed5c TH |
749 | pool_id <<= WORK_OFFQ_POOL_SHIFT; |
750 | set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); | |
bbb68dfa TH |
751 | } |
752 | ||
753 | static bool work_is_canceling(struct work_struct *work) | |
754 | { | |
755 | unsigned long data = atomic_long_read(&work->data); | |
756 | ||
112202d9 | 757 | return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); |
bbb68dfa TH |
758 | } |
759 | ||
e22bee78 | 760 | /* |
3270476a TH |
761 | * Policy functions. These define the policies on how the global worker |
762 | * pools are managed. Unless noted otherwise, these functions assume that | |
d565ed63 | 763 | * they're being called with pool->lock held. |
e22bee78 TH |
764 | */ |
765 | ||
63d95a91 | 766 | static bool __need_more_worker(struct worker_pool *pool) |
a848e3b6 | 767 | { |
e19e397a | 768 | return !atomic_read(&pool->nr_running); |
a848e3b6 ON |
769 | } |
770 | ||
4594bf15 | 771 | /* |
e22bee78 TH |
772 | * Need to wake up a worker? Called from anything but currently |
773 | * running workers. | |
974271c4 TH |
774 | * |
775 | * Note that, because unbound workers never contribute to nr_running, this | |
706026c2 | 776 | * function will always return %true for unbound pools as long as the |
974271c4 | 777 | * worklist isn't empty. |
4594bf15 | 778 | */ |
63d95a91 | 779 | static bool need_more_worker(struct worker_pool *pool) |
365970a1 | 780 | { |
63d95a91 | 781 | return !list_empty(&pool->worklist) && __need_more_worker(pool); |
e22bee78 | 782 | } |
4594bf15 | 783 | |
e22bee78 | 784 | /* Can I start working? Called from busy but !running workers. */ |
63d95a91 | 785 | static bool may_start_working(struct worker_pool *pool) |
e22bee78 | 786 | { |
63d95a91 | 787 | return pool->nr_idle; |
e22bee78 TH |
788 | } |
789 | ||
790 | /* Do I need to keep working? Called from currently running workers. */ | |
63d95a91 | 791 | static bool keep_working(struct worker_pool *pool) |
e22bee78 | 792 | { |
e19e397a TH |
793 | return !list_empty(&pool->worklist) && |
794 | atomic_read(&pool->nr_running) <= 1; | |
e22bee78 TH |
795 | } |
796 | ||
797 | /* Do we need a new worker? Called from manager. */ | |
63d95a91 | 798 | static bool need_to_create_worker(struct worker_pool *pool) |
e22bee78 | 799 | { |
63d95a91 | 800 | return need_more_worker(pool) && !may_start_working(pool); |
e22bee78 | 801 | } |
365970a1 | 802 | |
e22bee78 | 803 | /* Do we have too many workers and should some go away? */ |
63d95a91 | 804 | static bool too_many_workers(struct worker_pool *pool) |
e22bee78 | 805 | { |
692b4825 | 806 | bool managing = pool->flags & POOL_MANAGER_ACTIVE; |
63d95a91 TH |
807 | int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ |
808 | int nr_busy = pool->nr_workers - nr_idle; | |
e22bee78 TH |
809 | |
810 | return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; | |
365970a1 DH |
811 | } |
812 | ||
4d707b9f | 813 | /* |
e22bee78 TH |
814 | * Wake up functions. |
815 | */ | |
816 | ||
1037de36 LJ |
817 | /* Return the first idle worker. Safe with preemption disabled */ |
818 | static struct worker *first_idle_worker(struct worker_pool *pool) | |
7e11629d | 819 | { |
63d95a91 | 820 | if (unlikely(list_empty(&pool->idle_list))) |
7e11629d TH |
821 | return NULL; |
822 | ||
63d95a91 | 823 | return list_first_entry(&pool->idle_list, struct worker, entry); |
7e11629d TH |
824 | } |
825 | ||
826 | /** | |
827 | * wake_up_worker - wake up an idle worker | |
63d95a91 | 828 | * @pool: worker pool to wake worker from |
7e11629d | 829 | * |
63d95a91 | 830 | * Wake up the first idle worker of @pool. |
7e11629d TH |
831 | * |
832 | * CONTEXT: | |
d565ed63 | 833 | * spin_lock_irq(pool->lock). |
7e11629d | 834 | */ |
63d95a91 | 835 | static void wake_up_worker(struct worker_pool *pool) |
7e11629d | 836 | { |
1037de36 | 837 | struct worker *worker = first_idle_worker(pool); |
7e11629d TH |
838 | |
839 | if (likely(worker)) | |
840 | wake_up_process(worker->task); | |
841 | } | |
842 | ||
d302f017 | 843 | /** |
e22bee78 TH |
844 | * wq_worker_waking_up - a worker is waking up |
845 | * @task: task waking up | |
846 | * @cpu: CPU @task is waking up to | |
847 | * | |
848 | * This function is called during try_to_wake_up() when a worker is | |
849 | * being awoken. | |
850 | * | |
851 | * CONTEXT: | |
852 | * spin_lock_irq(rq->lock) | |
853 | */ | |
d84ff051 | 854 | void wq_worker_waking_up(struct task_struct *task, int cpu) |
e22bee78 TH |
855 | { |
856 | struct worker *worker = kthread_data(task); | |
857 | ||
36576000 | 858 | if (!(worker->flags & WORKER_NOT_RUNNING)) { |
ec22ca5e | 859 | WARN_ON_ONCE(worker->pool->cpu != cpu); |
e19e397a | 860 | atomic_inc(&worker->pool->nr_running); |
36576000 | 861 | } |
e22bee78 TH |
862 | } |
863 | ||
864 | /** | |
865 | * wq_worker_sleeping - a worker is going to sleep | |
866 | * @task: task going to sleep | |
e22bee78 TH |
867 | * |
868 | * This function is called during schedule() when a busy worker is | |
869 | * going to sleep. Worker on the same cpu can be woken up by | |
870 | * returning pointer to its task. | |
871 | * | |
872 | * CONTEXT: | |
873 | * spin_lock_irq(rq->lock) | |
874 | * | |
d185af30 | 875 | * Return: |
e22bee78 TH |
876 | * Worker task on @cpu to wake up, %NULL if none. |
877 | */ | |
9b7f6597 | 878 | struct task_struct *wq_worker_sleeping(struct task_struct *task) |
e22bee78 TH |
879 | { |
880 | struct worker *worker = kthread_data(task), *to_wakeup = NULL; | |
111c225a | 881 | struct worker_pool *pool; |
e22bee78 | 882 | |
111c225a TH |
883 | /* |
884 | * Rescuers, which may not have all the fields set up like normal | |
885 | * workers, also reach here, let's not access anything before | |
886 | * checking NOT_RUNNING. | |
887 | */ | |
2d64672e | 888 | if (worker->flags & WORKER_NOT_RUNNING) |
e22bee78 TH |
889 | return NULL; |
890 | ||
111c225a | 891 | pool = worker->pool; |
111c225a | 892 | |
e22bee78 | 893 | /* this can only happen on the local cpu */ |
9b7f6597 | 894 | if (WARN_ON_ONCE(pool->cpu != raw_smp_processor_id())) |
6183c009 | 895 | return NULL; |
e22bee78 TH |
896 | |
897 | /* | |
898 | * The counterpart of the following dec_and_test, implied mb, | |
899 | * worklist not empty test sequence is in insert_work(). | |
900 | * Please read comment there. | |
901 | * | |
628c78e7 TH |
902 | * NOT_RUNNING is clear. This means that we're bound to and |
903 | * running on the local cpu w/ rq lock held and preemption | |
904 | * disabled, which in turn means that none else could be | |
d565ed63 | 905 | * manipulating idle_list, so dereferencing idle_list without pool |
628c78e7 | 906 | * lock is safe. |
e22bee78 | 907 | */ |
e19e397a TH |
908 | if (atomic_dec_and_test(&pool->nr_running) && |
909 | !list_empty(&pool->worklist)) | |
1037de36 | 910 | to_wakeup = first_idle_worker(pool); |
e22bee78 TH |
911 | return to_wakeup ? to_wakeup->task : NULL; |
912 | } | |
913 | ||
1b69ac6b JW |
914 | /** |
915 | * wq_worker_last_func - retrieve worker's last work function | |
916 | * | |
917 | * Determine the last function a worker executed. This is called from | |
918 | * the scheduler to get a worker's last known identity. | |
919 | * | |
920 | * CONTEXT: | |
921 | * spin_lock_irq(rq->lock) | |
922 | * | |
923 | * Return: | |
924 | * The last work function %current executed as a worker, NULL if it | |
925 | * hasn't executed any work yet. | |
926 | */ | |
927 | work_func_t wq_worker_last_func(struct task_struct *task) | |
928 | { | |
929 | struct worker *worker = kthread_data(task); | |
930 | ||
931 | return worker->last_func; | |
932 | } | |
933 | ||
e22bee78 TH |
934 | /** |
935 | * worker_set_flags - set worker flags and adjust nr_running accordingly | |
cb444766 | 936 | * @worker: self |
d302f017 | 937 | * @flags: flags to set |
d302f017 | 938 | * |
228f1d00 | 939 | * Set @flags in @worker->flags and adjust nr_running accordingly. |
d302f017 | 940 | * |
cb444766 | 941 | * CONTEXT: |
d565ed63 | 942 | * spin_lock_irq(pool->lock) |
d302f017 | 943 | */ |
228f1d00 | 944 | static inline void worker_set_flags(struct worker *worker, unsigned int flags) |
d302f017 | 945 | { |
bd7bdd43 | 946 | struct worker_pool *pool = worker->pool; |
e22bee78 | 947 | |
cb444766 TH |
948 | WARN_ON_ONCE(worker->task != current); |
949 | ||
228f1d00 | 950 | /* If transitioning into NOT_RUNNING, adjust nr_running. */ |
e22bee78 TH |
951 | if ((flags & WORKER_NOT_RUNNING) && |
952 | !(worker->flags & WORKER_NOT_RUNNING)) { | |
228f1d00 | 953 | atomic_dec(&pool->nr_running); |
e22bee78 TH |
954 | } |
955 | ||
d302f017 TH |
956 | worker->flags |= flags; |
957 | } | |
958 | ||
959 | /** | |
e22bee78 | 960 | * worker_clr_flags - clear worker flags and adjust nr_running accordingly |
cb444766 | 961 | * @worker: self |
d302f017 TH |
962 | * @flags: flags to clear |
963 | * | |
e22bee78 | 964 | * Clear @flags in @worker->flags and adjust nr_running accordingly. |
d302f017 | 965 | * |
cb444766 | 966 | * CONTEXT: |
d565ed63 | 967 | * spin_lock_irq(pool->lock) |
d302f017 TH |
968 | */ |
969 | static inline void worker_clr_flags(struct worker *worker, unsigned int flags) | |
970 | { | |
63d95a91 | 971 | struct worker_pool *pool = worker->pool; |
e22bee78 TH |
972 | unsigned int oflags = worker->flags; |
973 | ||
cb444766 TH |
974 | WARN_ON_ONCE(worker->task != current); |
975 | ||
d302f017 | 976 | worker->flags &= ~flags; |
e22bee78 | 977 | |
42c025f3 TH |
978 | /* |
979 | * If transitioning out of NOT_RUNNING, increment nr_running. Note | |
980 | * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask | |
981 | * of multiple flags, not a single flag. | |
982 | */ | |
e22bee78 TH |
983 | if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) |
984 | if (!(worker->flags & WORKER_NOT_RUNNING)) | |
e19e397a | 985 | atomic_inc(&pool->nr_running); |
d302f017 TH |
986 | } |
987 | ||
8cca0eea TH |
988 | /** |
989 | * find_worker_executing_work - find worker which is executing a work | |
c9e7cf27 | 990 | * @pool: pool of interest |
8cca0eea TH |
991 | * @work: work to find worker for |
992 | * | |
c9e7cf27 TH |
993 | * Find a worker which is executing @work on @pool by searching |
994 | * @pool->busy_hash which is keyed by the address of @work. For a worker | |
a2c1c57b TH |
995 | * to match, its current execution should match the address of @work and |
996 | * its work function. This is to avoid unwanted dependency between | |
997 | * unrelated work executions through a work item being recycled while still | |
998 | * being executed. | |
999 | * | |
1000 | * This is a bit tricky. A work item may be freed once its execution | |
1001 | * starts and nothing prevents the freed area from being recycled for | |
1002 | * another work item. If the same work item address ends up being reused | |
1003 | * before the original execution finishes, workqueue will identify the | |
1004 | * recycled work item as currently executing and make it wait until the | |
1005 | * current execution finishes, introducing an unwanted dependency. | |
1006 | * | |
c5aa87bb TH |
1007 | * This function checks the work item address and work function to avoid |
1008 | * false positives. Note that this isn't complete as one may construct a | |
1009 | * work function which can introduce dependency onto itself through a | |
1010 | * recycled work item. Well, if somebody wants to shoot oneself in the | |
1011 | * foot that badly, there's only so much we can do, and if such deadlock | |
1012 | * actually occurs, it should be easy to locate the culprit work function. | |
8cca0eea TH |
1013 | * |
1014 | * CONTEXT: | |
d565ed63 | 1015 | * spin_lock_irq(pool->lock). |
8cca0eea | 1016 | * |
d185af30 YB |
1017 | * Return: |
1018 | * Pointer to worker which is executing @work if found, %NULL | |
8cca0eea | 1019 | * otherwise. |
4d707b9f | 1020 | */ |
c9e7cf27 | 1021 | static struct worker *find_worker_executing_work(struct worker_pool *pool, |
8cca0eea | 1022 | struct work_struct *work) |
4d707b9f | 1023 | { |
42f8570f | 1024 | struct worker *worker; |
42f8570f | 1025 | |
b67bfe0d | 1026 | hash_for_each_possible(pool->busy_hash, worker, hentry, |
a2c1c57b TH |
1027 | (unsigned long)work) |
1028 | if (worker->current_work == work && | |
1029 | worker->current_func == work->func) | |
42f8570f SL |
1030 | return worker; |
1031 | ||
1032 | return NULL; | |
4d707b9f ON |
1033 | } |
1034 | ||
bf4ede01 TH |
1035 | /** |
1036 | * move_linked_works - move linked works to a list | |
1037 | * @work: start of series of works to be scheduled | |
1038 | * @head: target list to append @work to | |
402dd89d | 1039 | * @nextp: out parameter for nested worklist walking |
bf4ede01 TH |
1040 | * |
1041 | * Schedule linked works starting from @work to @head. Work series to | |
1042 | * be scheduled starts at @work and includes any consecutive work with | |
1043 | * WORK_STRUCT_LINKED set in its predecessor. | |
1044 | * | |
1045 | * If @nextp is not NULL, it's updated to point to the next work of | |
1046 | * the last scheduled work. This allows move_linked_works() to be | |
1047 | * nested inside outer list_for_each_entry_safe(). | |
1048 | * | |
1049 | * CONTEXT: | |
d565ed63 | 1050 | * spin_lock_irq(pool->lock). |
bf4ede01 TH |
1051 | */ |
1052 | static void move_linked_works(struct work_struct *work, struct list_head *head, | |
1053 | struct work_struct **nextp) | |
1054 | { | |
1055 | struct work_struct *n; | |
1056 | ||
1057 | /* | |
1058 | * Linked worklist will always end before the end of the list, | |
1059 | * use NULL for list head. | |
1060 | */ | |
1061 | list_for_each_entry_safe_from(work, n, NULL, entry) { | |
1062 | list_move_tail(&work->entry, head); | |
1063 | if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) | |
1064 | break; | |
1065 | } | |
1066 | ||
1067 | /* | |
1068 | * If we're already inside safe list traversal and have moved | |
1069 | * multiple works to the scheduled queue, the next position | |
1070 | * needs to be updated. | |
1071 | */ | |
1072 | if (nextp) | |
1073 | *nextp = n; | |
1074 | } | |
1075 | ||
8864b4e5 TH |
1076 | /** |
1077 | * get_pwq - get an extra reference on the specified pool_workqueue | |
1078 | * @pwq: pool_workqueue to get | |
1079 | * | |
1080 | * Obtain an extra reference on @pwq. The caller should guarantee that | |
1081 | * @pwq has positive refcnt and be holding the matching pool->lock. | |
1082 | */ | |
1083 | static void get_pwq(struct pool_workqueue *pwq) | |
1084 | { | |
1085 | lockdep_assert_held(&pwq->pool->lock); | |
1086 | WARN_ON_ONCE(pwq->refcnt <= 0); | |
1087 | pwq->refcnt++; | |
1088 | } | |
1089 | ||
1090 | /** | |
1091 | * put_pwq - put a pool_workqueue reference | |
1092 | * @pwq: pool_workqueue to put | |
1093 | * | |
1094 | * Drop a reference of @pwq. If its refcnt reaches zero, schedule its | |
1095 | * destruction. The caller should be holding the matching pool->lock. | |
1096 | */ | |
1097 | static void put_pwq(struct pool_workqueue *pwq) | |
1098 | { | |
1099 | lockdep_assert_held(&pwq->pool->lock); | |
1100 | if (likely(--pwq->refcnt)) | |
1101 | return; | |
1102 | if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND))) | |
1103 | return; | |
1104 | /* | |
1105 | * @pwq can't be released under pool->lock, bounce to | |
1106 | * pwq_unbound_release_workfn(). This never recurses on the same | |
1107 | * pool->lock as this path is taken only for unbound workqueues and | |
1108 | * the release work item is scheduled on a per-cpu workqueue. To | |
1109 | * avoid lockdep warning, unbound pool->locks are given lockdep | |
1110 | * subclass of 1 in get_unbound_pool(). | |
1111 | */ | |
1112 | schedule_work(&pwq->unbound_release_work); | |
1113 | } | |
1114 | ||
dce90d47 TH |
1115 | /** |
1116 | * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock | |
1117 | * @pwq: pool_workqueue to put (can be %NULL) | |
1118 | * | |
1119 | * put_pwq() with locking. This function also allows %NULL @pwq. | |
1120 | */ | |
1121 | static void put_pwq_unlocked(struct pool_workqueue *pwq) | |
1122 | { | |
1123 | if (pwq) { | |
1124 | /* | |
1125 | * As both pwqs and pools are sched-RCU protected, the | |
1126 | * following lock operations are safe. | |
1127 | */ | |
1128 | spin_lock_irq(&pwq->pool->lock); | |
1129 | put_pwq(pwq); | |
1130 | spin_unlock_irq(&pwq->pool->lock); | |
1131 | } | |
1132 | } | |
1133 | ||
112202d9 | 1134 | static void pwq_activate_delayed_work(struct work_struct *work) |
bf4ede01 | 1135 | { |
112202d9 | 1136 | struct pool_workqueue *pwq = get_work_pwq(work); |
bf4ede01 TH |
1137 | |
1138 | trace_workqueue_activate_work(work); | |
82607adc TH |
1139 | if (list_empty(&pwq->pool->worklist)) |
1140 | pwq->pool->watchdog_ts = jiffies; | |
112202d9 | 1141 | move_linked_works(work, &pwq->pool->worklist, NULL); |
bf4ede01 | 1142 | __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work)); |
112202d9 | 1143 | pwq->nr_active++; |
bf4ede01 TH |
1144 | } |
1145 | ||
112202d9 | 1146 | static void pwq_activate_first_delayed(struct pool_workqueue *pwq) |
3aa62497 | 1147 | { |
112202d9 | 1148 | struct work_struct *work = list_first_entry(&pwq->delayed_works, |
3aa62497 LJ |
1149 | struct work_struct, entry); |
1150 | ||
112202d9 | 1151 | pwq_activate_delayed_work(work); |
3aa62497 LJ |
1152 | } |
1153 | ||
bf4ede01 | 1154 | /** |
112202d9 TH |
1155 | * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight |
1156 | * @pwq: pwq of interest | |
bf4ede01 | 1157 | * @color: color of work which left the queue |
bf4ede01 TH |
1158 | * |
1159 | * A work either has completed or is removed from pending queue, | |
112202d9 | 1160 | * decrement nr_in_flight of its pwq and handle workqueue flushing. |
bf4ede01 TH |
1161 | * |
1162 | * CONTEXT: | |
d565ed63 | 1163 | * spin_lock_irq(pool->lock). |
bf4ede01 | 1164 | */ |
112202d9 | 1165 | static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color) |
bf4ede01 | 1166 | { |
8864b4e5 | 1167 | /* uncolored work items don't participate in flushing or nr_active */ |
bf4ede01 | 1168 | if (color == WORK_NO_COLOR) |
8864b4e5 | 1169 | goto out_put; |
bf4ede01 | 1170 | |
112202d9 | 1171 | pwq->nr_in_flight[color]--; |
bf4ede01 | 1172 | |
112202d9 TH |
1173 | pwq->nr_active--; |
1174 | if (!list_empty(&pwq->delayed_works)) { | |
b3f9f405 | 1175 | /* one down, submit a delayed one */ |
112202d9 TH |
1176 | if (pwq->nr_active < pwq->max_active) |
1177 | pwq_activate_first_delayed(pwq); | |
bf4ede01 TH |
1178 | } |
1179 | ||
1180 | /* is flush in progress and are we at the flushing tip? */ | |
112202d9 | 1181 | if (likely(pwq->flush_color != color)) |
8864b4e5 | 1182 | goto out_put; |
bf4ede01 TH |
1183 | |
1184 | /* are there still in-flight works? */ | |
112202d9 | 1185 | if (pwq->nr_in_flight[color]) |
8864b4e5 | 1186 | goto out_put; |
bf4ede01 | 1187 | |
112202d9 TH |
1188 | /* this pwq is done, clear flush_color */ |
1189 | pwq->flush_color = -1; | |
bf4ede01 TH |
1190 | |
1191 | /* | |
112202d9 | 1192 | * If this was the last pwq, wake up the first flusher. It |
bf4ede01 TH |
1193 | * will handle the rest. |
1194 | */ | |
112202d9 TH |
1195 | if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) |
1196 | complete(&pwq->wq->first_flusher->done); | |
8864b4e5 TH |
1197 | out_put: |
1198 | put_pwq(pwq); | |
bf4ede01 TH |
1199 | } |
1200 | ||
36e227d2 | 1201 | /** |
bbb68dfa | 1202 | * try_to_grab_pending - steal work item from worklist and disable irq |
36e227d2 TH |
1203 | * @work: work item to steal |
1204 | * @is_dwork: @work is a delayed_work | |
bbb68dfa | 1205 | * @flags: place to store irq state |
36e227d2 TH |
1206 | * |
1207 | * Try to grab PENDING bit of @work. This function can handle @work in any | |
d185af30 | 1208 | * stable state - idle, on timer or on worklist. |
36e227d2 | 1209 | * |
d185af30 | 1210 | * Return: |
36e227d2 TH |
1211 | * 1 if @work was pending and we successfully stole PENDING |
1212 | * 0 if @work was idle and we claimed PENDING | |
1213 | * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry | |
bbb68dfa TH |
1214 | * -ENOENT if someone else is canceling @work, this state may persist |
1215 | * for arbitrarily long | |
36e227d2 | 1216 | * |
d185af30 | 1217 | * Note: |
bbb68dfa | 1218 | * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting |
e0aecdd8 TH |
1219 | * interrupted while holding PENDING and @work off queue, irq must be |
1220 | * disabled on entry. This, combined with delayed_work->timer being | |
1221 | * irqsafe, ensures that we return -EAGAIN for finite short period of time. | |
bbb68dfa TH |
1222 | * |
1223 | * On successful return, >= 0, irq is disabled and the caller is | |
1224 | * responsible for releasing it using local_irq_restore(*@flags). | |
1225 | * | |
e0aecdd8 | 1226 | * This function is safe to call from any context including IRQ handler. |
bf4ede01 | 1227 | */ |
bbb68dfa TH |
1228 | static int try_to_grab_pending(struct work_struct *work, bool is_dwork, |
1229 | unsigned long *flags) | |
bf4ede01 | 1230 | { |
d565ed63 | 1231 | struct worker_pool *pool; |
112202d9 | 1232 | struct pool_workqueue *pwq; |
bf4ede01 | 1233 | |
bbb68dfa TH |
1234 | local_irq_save(*flags); |
1235 | ||
36e227d2 TH |
1236 | /* try to steal the timer if it exists */ |
1237 | if (is_dwork) { | |
1238 | struct delayed_work *dwork = to_delayed_work(work); | |
1239 | ||
e0aecdd8 TH |
1240 | /* |
1241 | * dwork->timer is irqsafe. If del_timer() fails, it's | |
1242 | * guaranteed that the timer is not queued anywhere and not | |
1243 | * running on the local CPU. | |
1244 | */ | |
36e227d2 TH |
1245 | if (likely(del_timer(&dwork->timer))) |
1246 | return 1; | |
1247 | } | |
1248 | ||
1249 | /* try to claim PENDING the normal way */ | |
bf4ede01 TH |
1250 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) |
1251 | return 0; | |
1252 | ||
1253 | /* | |
1254 | * The queueing is in progress, or it is already queued. Try to | |
1255 | * steal it from ->worklist without clearing WORK_STRUCT_PENDING. | |
1256 | */ | |
d565ed63 TH |
1257 | pool = get_work_pool(work); |
1258 | if (!pool) | |
bbb68dfa | 1259 | goto fail; |
bf4ede01 | 1260 | |
d565ed63 | 1261 | spin_lock(&pool->lock); |
0b3dae68 | 1262 | /* |
112202d9 TH |
1263 | * work->data is guaranteed to point to pwq only while the work |
1264 | * item is queued on pwq->wq, and both updating work->data to point | |
1265 | * to pwq on queueing and to pool on dequeueing are done under | |
1266 | * pwq->pool->lock. This in turn guarantees that, if work->data | |
1267 | * points to pwq which is associated with a locked pool, the work | |
0b3dae68 LJ |
1268 | * item is currently queued on that pool. |
1269 | */ | |
112202d9 TH |
1270 | pwq = get_work_pwq(work); |
1271 | if (pwq && pwq->pool == pool) { | |
16062836 TH |
1272 | debug_work_deactivate(work); |
1273 | ||
1274 | /* | |
1275 | * A delayed work item cannot be grabbed directly because | |
1276 | * it might have linked NO_COLOR work items which, if left | |
112202d9 | 1277 | * on the delayed_list, will confuse pwq->nr_active |
16062836 TH |
1278 | * management later on and cause stall. Make sure the work |
1279 | * item is activated before grabbing. | |
1280 | */ | |
1281 | if (*work_data_bits(work) & WORK_STRUCT_DELAYED) | |
112202d9 | 1282 | pwq_activate_delayed_work(work); |
16062836 TH |
1283 | |
1284 | list_del_init(&work->entry); | |
9c34a704 | 1285 | pwq_dec_nr_in_flight(pwq, get_work_color(work)); |
16062836 | 1286 | |
112202d9 | 1287 | /* work->data points to pwq iff queued, point to pool */ |
16062836 TH |
1288 | set_work_pool_and_keep_pending(work, pool->id); |
1289 | ||
1290 | spin_unlock(&pool->lock); | |
1291 | return 1; | |
bf4ede01 | 1292 | } |
d565ed63 | 1293 | spin_unlock(&pool->lock); |
bbb68dfa TH |
1294 | fail: |
1295 | local_irq_restore(*flags); | |
1296 | if (work_is_canceling(work)) | |
1297 | return -ENOENT; | |
1298 | cpu_relax(); | |
36e227d2 | 1299 | return -EAGAIN; |
bf4ede01 TH |
1300 | } |
1301 | ||
4690c4ab | 1302 | /** |
706026c2 | 1303 | * insert_work - insert a work into a pool |
112202d9 | 1304 | * @pwq: pwq @work belongs to |
4690c4ab TH |
1305 | * @work: work to insert |
1306 | * @head: insertion point | |
1307 | * @extra_flags: extra WORK_STRUCT_* flags to set | |
1308 | * | |
112202d9 | 1309 | * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to |
706026c2 | 1310 | * work_struct flags. |
4690c4ab TH |
1311 | * |
1312 | * CONTEXT: | |
d565ed63 | 1313 | * spin_lock_irq(pool->lock). |
4690c4ab | 1314 | */ |
112202d9 TH |
1315 | static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, |
1316 | struct list_head *head, unsigned int extra_flags) | |
b89deed3 | 1317 | { |
112202d9 | 1318 | struct worker_pool *pool = pwq->pool; |
e22bee78 | 1319 | |
4690c4ab | 1320 | /* we own @work, set data and link */ |
112202d9 | 1321 | set_work_pwq(work, pwq, extra_flags); |
1a4d9b0a | 1322 | list_add_tail(&work->entry, head); |
8864b4e5 | 1323 | get_pwq(pwq); |
e22bee78 TH |
1324 | |
1325 | /* | |
c5aa87bb TH |
1326 | * Ensure either wq_worker_sleeping() sees the above |
1327 | * list_add_tail() or we see zero nr_running to avoid workers lying | |
1328 | * around lazily while there are works to be processed. | |
e22bee78 TH |
1329 | */ |
1330 | smp_mb(); | |
1331 | ||
63d95a91 TH |
1332 | if (__need_more_worker(pool)) |
1333 | wake_up_worker(pool); | |
b89deed3 ON |
1334 | } |
1335 | ||
c8efcc25 TH |
1336 | /* |
1337 | * Test whether @work is being queued from another work executing on the | |
8d03ecfe | 1338 | * same workqueue. |
c8efcc25 TH |
1339 | */ |
1340 | static bool is_chained_work(struct workqueue_struct *wq) | |
1341 | { | |
8d03ecfe TH |
1342 | struct worker *worker; |
1343 | ||
1344 | worker = current_wq_worker(); | |
1345 | /* | |
1346 | * Return %true iff I'm a worker execuing a work item on @wq. If | |
1347 | * I'm @worker, it's safe to dereference it without locking. | |
1348 | */ | |
112202d9 | 1349 | return worker && worker->current_pwq->wq == wq; |
c8efcc25 TH |
1350 | } |
1351 | ||
ef557180 MG |
1352 | /* |
1353 | * When queueing an unbound work item to a wq, prefer local CPU if allowed | |
1354 | * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to | |
1355 | * avoid perturbing sensitive tasks. | |
1356 | */ | |
1357 | static int wq_select_unbound_cpu(int cpu) | |
1358 | { | |
f303fccb | 1359 | static bool printed_dbg_warning; |
ef557180 MG |
1360 | int new_cpu; |
1361 | ||
f303fccb TH |
1362 | if (likely(!wq_debug_force_rr_cpu)) { |
1363 | if (cpumask_test_cpu(cpu, wq_unbound_cpumask)) | |
1364 | return cpu; | |
1365 | } else if (!printed_dbg_warning) { | |
1366 | pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n"); | |
1367 | printed_dbg_warning = true; | |
1368 | } | |
1369 | ||
ef557180 MG |
1370 | if (cpumask_empty(wq_unbound_cpumask)) |
1371 | return cpu; | |
1372 | ||
1373 | new_cpu = __this_cpu_read(wq_rr_cpu_last); | |
1374 | new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask); | |
1375 | if (unlikely(new_cpu >= nr_cpu_ids)) { | |
1376 | new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask); | |
1377 | if (unlikely(new_cpu >= nr_cpu_ids)) | |
1378 | return cpu; | |
1379 | } | |
1380 | __this_cpu_write(wq_rr_cpu_last, new_cpu); | |
1381 | ||
1382 | return new_cpu; | |
1383 | } | |
1384 | ||
d84ff051 | 1385 | static void __queue_work(int cpu, struct workqueue_struct *wq, |
1da177e4 LT |
1386 | struct work_struct *work) |
1387 | { | |
112202d9 | 1388 | struct pool_workqueue *pwq; |
c9178087 | 1389 | struct worker_pool *last_pool; |
1e19ffc6 | 1390 | struct list_head *worklist; |
8a2e8e5d | 1391 | unsigned int work_flags; |
b75cac93 | 1392 | unsigned int req_cpu = cpu; |
8930caba TH |
1393 | |
1394 | /* | |
1395 | * While a work item is PENDING && off queue, a task trying to | |
1396 | * steal the PENDING will busy-loop waiting for it to either get | |
1397 | * queued or lose PENDING. Grabbing PENDING and queueing should | |
1398 | * happen with IRQ disabled. | |
1399 | */ | |
8e8eb730 | 1400 | lockdep_assert_irqs_disabled(); |
1da177e4 | 1401 | |
dc186ad7 | 1402 | debug_work_activate(work); |
1e19ffc6 | 1403 | |
9ef28a73 | 1404 | /* if draining, only works from the same workqueue are allowed */ |
618b01eb | 1405 | if (unlikely(wq->flags & __WQ_DRAINING) && |
c8efcc25 | 1406 | WARN_ON_ONCE(!is_chained_work(wq))) |
e41e704b | 1407 | return; |
9e8cd2f5 | 1408 | retry: |
df2d5ae4 | 1409 | if (req_cpu == WORK_CPU_UNBOUND) |
ef557180 | 1410 | cpu = wq_select_unbound_cpu(raw_smp_processor_id()); |
df2d5ae4 | 1411 | |
c9178087 | 1412 | /* pwq which will be used unless @work is executing elsewhere */ |
df2d5ae4 | 1413 | if (!(wq->flags & WQ_UNBOUND)) |
7fb98ea7 | 1414 | pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); |
df2d5ae4 TH |
1415 | else |
1416 | pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); | |
dbf2576e | 1417 | |
c9178087 TH |
1418 | /* |
1419 | * If @work was previously on a different pool, it might still be | |
1420 | * running there, in which case the work needs to be queued on that | |
1421 | * pool to guarantee non-reentrancy. | |
1422 | */ | |
1423 | last_pool = get_work_pool(work); | |
1424 | if (last_pool && last_pool != pwq->pool) { | |
1425 | struct worker *worker; | |
18aa9eff | 1426 | |
c9178087 | 1427 | spin_lock(&last_pool->lock); |
18aa9eff | 1428 | |
c9178087 | 1429 | worker = find_worker_executing_work(last_pool, work); |
18aa9eff | 1430 | |
c9178087 TH |
1431 | if (worker && worker->current_pwq->wq == wq) { |
1432 | pwq = worker->current_pwq; | |
8930caba | 1433 | } else { |
c9178087 TH |
1434 | /* meh... not running there, queue here */ |
1435 | spin_unlock(&last_pool->lock); | |
112202d9 | 1436 | spin_lock(&pwq->pool->lock); |
8930caba | 1437 | } |
f3421797 | 1438 | } else { |
112202d9 | 1439 | spin_lock(&pwq->pool->lock); |
502ca9d8 TH |
1440 | } |
1441 | ||
9e8cd2f5 TH |
1442 | /* |
1443 | * pwq is determined and locked. For unbound pools, we could have | |
1444 | * raced with pwq release and it could already be dead. If its | |
1445 | * refcnt is zero, repeat pwq selection. Note that pwqs never die | |
df2d5ae4 TH |
1446 | * without another pwq replacing it in the numa_pwq_tbl or while |
1447 | * work items are executing on it, so the retrying is guaranteed to | |
9e8cd2f5 TH |
1448 | * make forward-progress. |
1449 | */ | |
1450 | if (unlikely(!pwq->refcnt)) { | |
1451 | if (wq->flags & WQ_UNBOUND) { | |
1452 | spin_unlock(&pwq->pool->lock); | |
1453 | cpu_relax(); | |
1454 | goto retry; | |
1455 | } | |
1456 | /* oops */ | |
1457 | WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", | |
1458 | wq->name, cpu); | |
1459 | } | |
1460 | ||
112202d9 TH |
1461 | /* pwq determined, queue */ |
1462 | trace_workqueue_queue_work(req_cpu, pwq, work); | |
502ca9d8 | 1463 | |
f5b2552b | 1464 | if (WARN_ON(!list_empty(&work->entry))) { |
112202d9 | 1465 | spin_unlock(&pwq->pool->lock); |
f5b2552b DC |
1466 | return; |
1467 | } | |
1e19ffc6 | 1468 | |
112202d9 TH |
1469 | pwq->nr_in_flight[pwq->work_color]++; |
1470 | work_flags = work_color_to_flags(pwq->work_color); | |
1e19ffc6 | 1471 | |
112202d9 | 1472 | if (likely(pwq->nr_active < pwq->max_active)) { |
cdadf009 | 1473 | trace_workqueue_activate_work(work); |
112202d9 TH |
1474 | pwq->nr_active++; |
1475 | worklist = &pwq->pool->worklist; | |
82607adc TH |
1476 | if (list_empty(worklist)) |
1477 | pwq->pool->watchdog_ts = jiffies; | |
8a2e8e5d TH |
1478 | } else { |
1479 | work_flags |= WORK_STRUCT_DELAYED; | |
112202d9 | 1480 | worklist = &pwq->delayed_works; |
8a2e8e5d | 1481 | } |
1e19ffc6 | 1482 | |
112202d9 | 1483 | insert_work(pwq, work, worklist, work_flags); |
1e19ffc6 | 1484 | |
112202d9 | 1485 | spin_unlock(&pwq->pool->lock); |
1da177e4 LT |
1486 | } |
1487 | ||
0fcb78c2 | 1488 | /** |
c1a220e7 ZR |
1489 | * queue_work_on - queue work on specific cpu |
1490 | * @cpu: CPU number to execute work on | |
0fcb78c2 REB |
1491 | * @wq: workqueue to use |
1492 | * @work: work to queue | |
1493 | * | |
c1a220e7 ZR |
1494 | * We queue the work to a specific CPU, the caller must ensure it |
1495 | * can't go away. | |
d185af30 YB |
1496 | * |
1497 | * Return: %false if @work was already on a queue, %true otherwise. | |
1da177e4 | 1498 | */ |
d4283e93 TH |
1499 | bool queue_work_on(int cpu, struct workqueue_struct *wq, |
1500 | struct work_struct *work) | |
1da177e4 | 1501 | { |
d4283e93 | 1502 | bool ret = false; |
8930caba | 1503 | unsigned long flags; |
ef1ca236 | 1504 | |
8930caba | 1505 | local_irq_save(flags); |
c1a220e7 | 1506 | |
22df02bb | 1507 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { |
4690c4ab | 1508 | __queue_work(cpu, wq, work); |
d4283e93 | 1509 | ret = true; |
c1a220e7 | 1510 | } |
ef1ca236 | 1511 | |
8930caba | 1512 | local_irq_restore(flags); |
1da177e4 LT |
1513 | return ret; |
1514 | } | |
ad7b1f84 | 1515 | EXPORT_SYMBOL(queue_work_on); |
1da177e4 | 1516 | |
8c20feb6 | 1517 | void delayed_work_timer_fn(struct timer_list *t) |
1da177e4 | 1518 | { |
8c20feb6 | 1519 | struct delayed_work *dwork = from_timer(dwork, t, timer); |
1da177e4 | 1520 | |
e0aecdd8 | 1521 | /* should have been called from irqsafe timer with irq already off */ |
60c057bc | 1522 | __queue_work(dwork->cpu, dwork->wq, &dwork->work); |
1da177e4 | 1523 | } |
1438ade5 | 1524 | EXPORT_SYMBOL(delayed_work_timer_fn); |
1da177e4 | 1525 | |
7beb2edf TH |
1526 | static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, |
1527 | struct delayed_work *dwork, unsigned long delay) | |
1da177e4 | 1528 | { |
7beb2edf TH |
1529 | struct timer_list *timer = &dwork->timer; |
1530 | struct work_struct *work = &dwork->work; | |
7beb2edf | 1531 | |
637fdbae | 1532 | WARN_ON_ONCE(!wq); |
841b86f3 | 1533 | WARN_ON_ONCE(timer->function != delayed_work_timer_fn); |
fc4b514f TH |
1534 | WARN_ON_ONCE(timer_pending(timer)); |
1535 | WARN_ON_ONCE(!list_empty(&work->entry)); | |
7beb2edf | 1536 | |
8852aac2 TH |
1537 | /* |
1538 | * If @delay is 0, queue @dwork->work immediately. This is for | |
1539 | * both optimization and correctness. The earliest @timer can | |
1540 | * expire is on the closest next tick and delayed_work users depend | |
1541 | * on that there's no such delay when @delay is 0. | |
1542 | */ | |
1543 | if (!delay) { | |
1544 | __queue_work(cpu, wq, &dwork->work); | |
1545 | return; | |
1546 | } | |
1547 | ||
60c057bc | 1548 | dwork->wq = wq; |
1265057f | 1549 | dwork->cpu = cpu; |
7beb2edf TH |
1550 | timer->expires = jiffies + delay; |
1551 | ||
041bd12e TH |
1552 | if (unlikely(cpu != WORK_CPU_UNBOUND)) |
1553 | add_timer_on(timer, cpu); | |
1554 | else | |
1555 | add_timer(timer); | |
1da177e4 LT |
1556 | } |
1557 | ||
0fcb78c2 REB |
1558 | /** |
1559 | * queue_delayed_work_on - queue work on specific CPU after delay | |
1560 | * @cpu: CPU number to execute work on | |
1561 | * @wq: workqueue to use | |
af9997e4 | 1562 | * @dwork: work to queue |
0fcb78c2 REB |
1563 | * @delay: number of jiffies to wait before queueing |
1564 | * | |
d185af30 | 1565 | * Return: %false if @work was already on a queue, %true otherwise. If |
715f1300 TH |
1566 | * @delay is zero and @dwork is idle, it will be scheduled for immediate |
1567 | * execution. | |
0fcb78c2 | 1568 | */ |
d4283e93 TH |
1569 | bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, |
1570 | struct delayed_work *dwork, unsigned long delay) | |
7a6bc1cd | 1571 | { |
52bad64d | 1572 | struct work_struct *work = &dwork->work; |
d4283e93 | 1573 | bool ret = false; |
8930caba | 1574 | unsigned long flags; |
7a6bc1cd | 1575 | |
8930caba TH |
1576 | /* read the comment in __queue_work() */ |
1577 | local_irq_save(flags); | |
7a6bc1cd | 1578 | |
22df02bb | 1579 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { |
7beb2edf | 1580 | __queue_delayed_work(cpu, wq, dwork, delay); |
d4283e93 | 1581 | ret = true; |
7a6bc1cd | 1582 | } |
8a3e77cc | 1583 | |
8930caba | 1584 | local_irq_restore(flags); |
7a6bc1cd VP |
1585 | return ret; |
1586 | } | |
ad7b1f84 | 1587 | EXPORT_SYMBOL(queue_delayed_work_on); |
c7fc77f7 | 1588 | |
8376fe22 TH |
1589 | /** |
1590 | * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU | |
1591 | * @cpu: CPU number to execute work on | |
1592 | * @wq: workqueue to use | |
1593 | * @dwork: work to queue | |
1594 | * @delay: number of jiffies to wait before queueing | |
1595 | * | |
1596 | * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, | |
1597 | * modify @dwork's timer so that it expires after @delay. If @delay is | |
1598 | * zero, @work is guaranteed to be scheduled immediately regardless of its | |
1599 | * current state. | |
1600 | * | |
d185af30 | 1601 | * Return: %false if @dwork was idle and queued, %true if @dwork was |
8376fe22 TH |
1602 | * pending and its timer was modified. |
1603 | * | |
e0aecdd8 | 1604 | * This function is safe to call from any context including IRQ handler. |
8376fe22 TH |
1605 | * See try_to_grab_pending() for details. |
1606 | */ | |
1607 | bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, | |
1608 | struct delayed_work *dwork, unsigned long delay) | |
1609 | { | |
1610 | unsigned long flags; | |
1611 | int ret; | |
c7fc77f7 | 1612 | |
8376fe22 TH |
1613 | do { |
1614 | ret = try_to_grab_pending(&dwork->work, true, &flags); | |
1615 | } while (unlikely(ret == -EAGAIN)); | |
63bc0362 | 1616 | |
8376fe22 TH |
1617 | if (likely(ret >= 0)) { |
1618 | __queue_delayed_work(cpu, wq, dwork, delay); | |
1619 | local_irq_restore(flags); | |
7a6bc1cd | 1620 | } |
8376fe22 TH |
1621 | |
1622 | /* -ENOENT from try_to_grab_pending() becomes %true */ | |
7a6bc1cd VP |
1623 | return ret; |
1624 | } | |
8376fe22 TH |
1625 | EXPORT_SYMBOL_GPL(mod_delayed_work_on); |
1626 | ||
05f0fe6b TH |
1627 | static void rcu_work_rcufn(struct rcu_head *rcu) |
1628 | { | |
1629 | struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu); | |
1630 | ||
1631 | /* read the comment in __queue_work() */ | |
1632 | local_irq_disable(); | |
1633 | __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work); | |
1634 | local_irq_enable(); | |
1635 | } | |
1636 | ||
1637 | /** | |
1638 | * queue_rcu_work - queue work after a RCU grace period | |
1639 | * @wq: workqueue to use | |
1640 | * @rwork: work to queue | |
1641 | * | |
1642 | * Return: %false if @rwork was already pending, %true otherwise. Note | |
1643 | * that a full RCU grace period is guaranteed only after a %true return. | |
1644 | * While @rwork is guarnateed to be executed after a %false return, the | |
1645 | * execution may happen before a full RCU grace period has passed. | |
1646 | */ | |
1647 | bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork) | |
1648 | { | |
1649 | struct work_struct *work = &rwork->work; | |
1650 | ||
1651 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { | |
1652 | rwork->wq = wq; | |
1653 | call_rcu(&rwork->rcu, rcu_work_rcufn); | |
1654 | return true; | |
1655 | } | |
1656 | ||
1657 | return false; | |
1658 | } | |
1659 | EXPORT_SYMBOL(queue_rcu_work); | |
1660 | ||
c8e55f36 TH |
1661 | /** |
1662 | * worker_enter_idle - enter idle state | |
1663 | * @worker: worker which is entering idle state | |
1664 | * | |
1665 | * @worker is entering idle state. Update stats and idle timer if | |
1666 | * necessary. | |
1667 | * | |
1668 | * LOCKING: | |
d565ed63 | 1669 | * spin_lock_irq(pool->lock). |
c8e55f36 TH |
1670 | */ |
1671 | static void worker_enter_idle(struct worker *worker) | |
1da177e4 | 1672 | { |
bd7bdd43 | 1673 | struct worker_pool *pool = worker->pool; |
c8e55f36 | 1674 | |
6183c009 TH |
1675 | if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || |
1676 | WARN_ON_ONCE(!list_empty(&worker->entry) && | |
1677 | (worker->hentry.next || worker->hentry.pprev))) | |
1678 | return; | |
c8e55f36 | 1679 | |
051e1850 | 1680 | /* can't use worker_set_flags(), also called from create_worker() */ |
cb444766 | 1681 | worker->flags |= WORKER_IDLE; |
bd7bdd43 | 1682 | pool->nr_idle++; |
e22bee78 | 1683 | worker->last_active = jiffies; |
c8e55f36 TH |
1684 | |
1685 | /* idle_list is LIFO */ | |
bd7bdd43 | 1686 | list_add(&worker->entry, &pool->idle_list); |
db7bccf4 | 1687 | |
628c78e7 TH |
1688 | if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) |
1689 | mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); | |
cb444766 | 1690 | |
544ecf31 | 1691 | /* |
e8b3f8db | 1692 | * Sanity check nr_running. Because unbind_workers() releases |
d565ed63 | 1693 | * pool->lock between setting %WORKER_UNBOUND and zapping |
628c78e7 TH |
1694 | * nr_running, the warning may trigger spuriously. Check iff |
1695 | * unbind is not in progress. | |
544ecf31 | 1696 | */ |
24647570 | 1697 | WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && |
bd7bdd43 | 1698 | pool->nr_workers == pool->nr_idle && |
e19e397a | 1699 | atomic_read(&pool->nr_running)); |
c8e55f36 TH |
1700 | } |
1701 | ||
1702 | /** | |
1703 | * worker_leave_idle - leave idle state | |
1704 | * @worker: worker which is leaving idle state | |
1705 | * | |
1706 | * @worker is leaving idle state. Update stats. | |
1707 | * | |
1708 | * LOCKING: | |
d565ed63 | 1709 | * spin_lock_irq(pool->lock). |
c8e55f36 TH |
1710 | */ |
1711 | static void worker_leave_idle(struct worker *worker) | |
1712 | { | |
bd7bdd43 | 1713 | struct worker_pool *pool = worker->pool; |
c8e55f36 | 1714 | |
6183c009 TH |
1715 | if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) |
1716 | return; | |
d302f017 | 1717 | worker_clr_flags(worker, WORKER_IDLE); |
bd7bdd43 | 1718 | pool->nr_idle--; |
c8e55f36 TH |
1719 | list_del_init(&worker->entry); |
1720 | } | |
1721 | ||
f7537df5 | 1722 | static struct worker *alloc_worker(int node) |
c34056a3 TH |
1723 | { |
1724 | struct worker *worker; | |
1725 | ||
f7537df5 | 1726 | worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node); |
c8e55f36 TH |
1727 | if (worker) { |
1728 | INIT_LIST_HEAD(&worker->entry); | |
affee4b2 | 1729 | INIT_LIST_HEAD(&worker->scheduled); |
da028469 | 1730 | INIT_LIST_HEAD(&worker->node); |
e22bee78 TH |
1731 | /* on creation a worker is in !idle && prep state */ |
1732 | worker->flags = WORKER_PREP; | |
c8e55f36 | 1733 | } |
c34056a3 TH |
1734 | return worker; |
1735 | } | |
1736 | ||
4736cbf7 LJ |
1737 | /** |
1738 | * worker_attach_to_pool() - attach a worker to a pool | |
1739 | * @worker: worker to be attached | |
1740 | * @pool: the target pool | |
1741 | * | |
1742 | * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and | |
1743 | * cpu-binding of @worker are kept coordinated with the pool across | |
1744 | * cpu-[un]hotplugs. | |
1745 | */ | |
1746 | static void worker_attach_to_pool(struct worker *worker, | |
1747 | struct worker_pool *pool) | |
1748 | { | |
1258fae7 | 1749 | mutex_lock(&wq_pool_attach_mutex); |
4736cbf7 LJ |
1750 | |
1751 | /* | |
1752 | * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any | |
1753 | * online CPUs. It'll be re-applied when any of the CPUs come up. | |
1754 | */ | |
1755 | set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask); | |
1756 | ||
1757 | /* | |
1258fae7 TH |
1758 | * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains |
1759 | * stable across this function. See the comments above the flag | |
1760 | * definition for details. | |
4736cbf7 LJ |
1761 | */ |
1762 | if (pool->flags & POOL_DISASSOCIATED) | |
1763 | worker->flags |= WORKER_UNBOUND; | |
1764 | ||
1765 | list_add_tail(&worker->node, &pool->workers); | |
a2d812a2 | 1766 | worker->pool = pool; |
4736cbf7 | 1767 | |
1258fae7 | 1768 | mutex_unlock(&wq_pool_attach_mutex); |
4736cbf7 LJ |
1769 | } |
1770 | ||
60f5a4bc LJ |
1771 | /** |
1772 | * worker_detach_from_pool() - detach a worker from its pool | |
1773 | * @worker: worker which is attached to its pool | |
60f5a4bc | 1774 | * |
4736cbf7 LJ |
1775 | * Undo the attaching which had been done in worker_attach_to_pool(). The |
1776 | * caller worker shouldn't access to the pool after detached except it has | |
1777 | * other reference to the pool. | |
60f5a4bc | 1778 | */ |
a2d812a2 | 1779 | static void worker_detach_from_pool(struct worker *worker) |
60f5a4bc | 1780 | { |
a2d812a2 | 1781 | struct worker_pool *pool = worker->pool; |
60f5a4bc LJ |
1782 | struct completion *detach_completion = NULL; |
1783 | ||
1258fae7 | 1784 | mutex_lock(&wq_pool_attach_mutex); |
a2d812a2 | 1785 | |
da028469 | 1786 | list_del(&worker->node); |
a2d812a2 TH |
1787 | worker->pool = NULL; |
1788 | ||
da028469 | 1789 | if (list_empty(&pool->workers)) |
60f5a4bc | 1790 | detach_completion = pool->detach_completion; |
1258fae7 | 1791 | mutex_unlock(&wq_pool_attach_mutex); |
60f5a4bc | 1792 | |
b62c0751 LJ |
1793 | /* clear leftover flags without pool->lock after it is detached */ |
1794 | worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND); | |
1795 | ||
60f5a4bc LJ |
1796 | if (detach_completion) |
1797 | complete(detach_completion); | |
1798 | } | |
1799 | ||
c34056a3 TH |
1800 | /** |
1801 | * create_worker - create a new workqueue worker | |
63d95a91 | 1802 | * @pool: pool the new worker will belong to |
c34056a3 | 1803 | * |
051e1850 | 1804 | * Create and start a new worker which is attached to @pool. |
c34056a3 TH |
1805 | * |
1806 | * CONTEXT: | |
1807 | * Might sleep. Does GFP_KERNEL allocations. | |
1808 | * | |
d185af30 | 1809 | * Return: |
c34056a3 TH |
1810 | * Pointer to the newly created worker. |
1811 | */ | |
bc2ae0f5 | 1812 | static struct worker *create_worker(struct worker_pool *pool) |
c34056a3 | 1813 | { |
c34056a3 | 1814 | struct worker *worker = NULL; |
f3421797 | 1815 | int id = -1; |
e3c916a4 | 1816 | char id_buf[16]; |
c34056a3 | 1817 | |
7cda9aae LJ |
1818 | /* ID is needed to determine kthread name */ |
1819 | id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL); | |
822d8405 TH |
1820 | if (id < 0) |
1821 | goto fail; | |
c34056a3 | 1822 | |
f7537df5 | 1823 | worker = alloc_worker(pool->node); |
c34056a3 TH |
1824 | if (!worker) |
1825 | goto fail; | |
1826 | ||
c34056a3 TH |
1827 | worker->id = id; |
1828 | ||
29c91e99 | 1829 | if (pool->cpu >= 0) |
e3c916a4 TH |
1830 | snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, |
1831 | pool->attrs->nice < 0 ? "H" : ""); | |
f3421797 | 1832 | else |
e3c916a4 TH |
1833 | snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); |
1834 | ||
f3f90ad4 | 1835 | worker->task = kthread_create_on_node(worker_thread, worker, pool->node, |
e3c916a4 | 1836 | "kworker/%s", id_buf); |
c34056a3 TH |
1837 | if (IS_ERR(worker->task)) |
1838 | goto fail; | |
1839 | ||
91151228 | 1840 | set_user_nice(worker->task, pool->attrs->nice); |
25834c73 | 1841 | kthread_bind_mask(worker->task, pool->attrs->cpumask); |
91151228 | 1842 | |
da028469 | 1843 | /* successful, attach the worker to the pool */ |
4736cbf7 | 1844 | worker_attach_to_pool(worker, pool); |
822d8405 | 1845 | |
051e1850 LJ |
1846 | /* start the newly created worker */ |
1847 | spin_lock_irq(&pool->lock); | |
1848 | worker->pool->nr_workers++; | |
1849 | worker_enter_idle(worker); | |
1850 | wake_up_process(worker->task); | |
1851 | spin_unlock_irq(&pool->lock); | |
1852 | ||
c34056a3 | 1853 | return worker; |
822d8405 | 1854 | |
c34056a3 | 1855 | fail: |
9625ab17 | 1856 | if (id >= 0) |
7cda9aae | 1857 | ida_simple_remove(&pool->worker_ida, id); |
c34056a3 TH |
1858 | kfree(worker); |
1859 | return NULL; | |
1860 | } | |
1861 | ||
c34056a3 TH |
1862 | /** |
1863 | * destroy_worker - destroy a workqueue worker | |
1864 | * @worker: worker to be destroyed | |
1865 | * | |
73eb7fe7 LJ |
1866 | * Destroy @worker and adjust @pool stats accordingly. The worker should |
1867 | * be idle. | |
c8e55f36 TH |
1868 | * |
1869 | * CONTEXT: | |
60f5a4bc | 1870 | * spin_lock_irq(pool->lock). |
c34056a3 TH |
1871 | */ |
1872 | static void destroy_worker(struct worker *worker) | |
1873 | { | |
bd7bdd43 | 1874 | struct worker_pool *pool = worker->pool; |
c34056a3 | 1875 | |
cd549687 TH |
1876 | lockdep_assert_held(&pool->lock); |
1877 | ||
c34056a3 | 1878 | /* sanity check frenzy */ |
6183c009 | 1879 | if (WARN_ON(worker->current_work) || |
73eb7fe7 LJ |
1880 | WARN_ON(!list_empty(&worker->scheduled)) || |
1881 | WARN_ON(!(worker->flags & WORKER_IDLE))) | |
6183c009 | 1882 | return; |
c34056a3 | 1883 | |
73eb7fe7 LJ |
1884 | pool->nr_workers--; |
1885 | pool->nr_idle--; | |
5bdfff96 | 1886 | |
c8e55f36 | 1887 | list_del_init(&worker->entry); |
cb444766 | 1888 | worker->flags |= WORKER_DIE; |
60f5a4bc | 1889 | wake_up_process(worker->task); |
c34056a3 TH |
1890 | } |
1891 | ||
32a6c723 | 1892 | static void idle_worker_timeout(struct timer_list *t) |
e22bee78 | 1893 | { |
32a6c723 | 1894 | struct worker_pool *pool = from_timer(pool, t, idle_timer); |
e22bee78 | 1895 | |
d565ed63 | 1896 | spin_lock_irq(&pool->lock); |
e22bee78 | 1897 | |
3347fc9f | 1898 | while (too_many_workers(pool)) { |
e22bee78 TH |
1899 | struct worker *worker; |
1900 | unsigned long expires; | |
1901 | ||
1902 | /* idle_list is kept in LIFO order, check the last one */ | |
63d95a91 | 1903 | worker = list_entry(pool->idle_list.prev, struct worker, entry); |
e22bee78 TH |
1904 | expires = worker->last_active + IDLE_WORKER_TIMEOUT; |
1905 | ||
3347fc9f | 1906 | if (time_before(jiffies, expires)) { |
63d95a91 | 1907 | mod_timer(&pool->idle_timer, expires); |
3347fc9f | 1908 | break; |
d5abe669 | 1909 | } |
3347fc9f LJ |
1910 | |
1911 | destroy_worker(worker); | |
e22bee78 TH |
1912 | } |
1913 | ||
d565ed63 | 1914 | spin_unlock_irq(&pool->lock); |
e22bee78 | 1915 | } |
d5abe669 | 1916 | |
493a1724 | 1917 | static void send_mayday(struct work_struct *work) |
e22bee78 | 1918 | { |
112202d9 TH |
1919 | struct pool_workqueue *pwq = get_work_pwq(work); |
1920 | struct workqueue_struct *wq = pwq->wq; | |
493a1724 | 1921 | |
2e109a28 | 1922 | lockdep_assert_held(&wq_mayday_lock); |
e22bee78 | 1923 | |
493008a8 | 1924 | if (!wq->rescuer) |
493a1724 | 1925 | return; |
e22bee78 TH |
1926 | |
1927 | /* mayday mayday mayday */ | |
493a1724 | 1928 | if (list_empty(&pwq->mayday_node)) { |
77668c8b LJ |
1929 | /* |
1930 | * If @pwq is for an unbound wq, its base ref may be put at | |
1931 | * any time due to an attribute change. Pin @pwq until the | |
1932 | * rescuer is done with it. | |
1933 | */ | |
1934 | get_pwq(pwq); | |
493a1724 | 1935 | list_add_tail(&pwq->mayday_node, &wq->maydays); |
e22bee78 | 1936 | wake_up_process(wq->rescuer->task); |
493a1724 | 1937 | } |
e22bee78 TH |
1938 | } |
1939 | ||
32a6c723 | 1940 | static void pool_mayday_timeout(struct timer_list *t) |
e22bee78 | 1941 | { |
32a6c723 | 1942 | struct worker_pool *pool = from_timer(pool, t, mayday_timer); |
e22bee78 TH |
1943 | struct work_struct *work; |
1944 | ||
b2d82909 TH |
1945 | spin_lock_irq(&pool->lock); |
1946 | spin_lock(&wq_mayday_lock); /* for wq->maydays */ | |
e22bee78 | 1947 | |
63d95a91 | 1948 | if (need_to_create_worker(pool)) { |
e22bee78 TH |
1949 | /* |
1950 | * We've been trying to create a new worker but | |
1951 | * haven't been successful. We might be hitting an | |
1952 | * allocation deadlock. Send distress signals to | |
1953 | * rescuers. | |
1954 | */ | |
63d95a91 | 1955 | list_for_each_entry(work, &pool->worklist, entry) |
e22bee78 | 1956 | send_mayday(work); |
1da177e4 | 1957 | } |
e22bee78 | 1958 | |
b2d82909 TH |
1959 | spin_unlock(&wq_mayday_lock); |
1960 | spin_unlock_irq(&pool->lock); | |
e22bee78 | 1961 | |
63d95a91 | 1962 | mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); |
1da177e4 LT |
1963 | } |
1964 | ||
e22bee78 TH |
1965 | /** |
1966 | * maybe_create_worker - create a new worker if necessary | |
63d95a91 | 1967 | * @pool: pool to create a new worker for |
e22bee78 | 1968 | * |
63d95a91 | 1969 | * Create a new worker for @pool if necessary. @pool is guaranteed to |
e22bee78 TH |
1970 | * have at least one idle worker on return from this function. If |
1971 | * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is | |
63d95a91 | 1972 | * sent to all rescuers with works scheduled on @pool to resolve |
e22bee78 TH |
1973 | * possible allocation deadlock. |
1974 | * | |
c5aa87bb TH |
1975 | * On return, need_to_create_worker() is guaranteed to be %false and |
1976 | * may_start_working() %true. | |
e22bee78 TH |
1977 | * |
1978 | * LOCKING: | |
d565ed63 | 1979 | * spin_lock_irq(pool->lock) which may be released and regrabbed |
e22bee78 TH |
1980 | * multiple times. Does GFP_KERNEL allocations. Called only from |
1981 | * manager. | |
e22bee78 | 1982 | */ |
29187a9e | 1983 | static void maybe_create_worker(struct worker_pool *pool) |
d565ed63 TH |
1984 | __releases(&pool->lock) |
1985 | __acquires(&pool->lock) | |
1da177e4 | 1986 | { |
e22bee78 | 1987 | restart: |
d565ed63 | 1988 | spin_unlock_irq(&pool->lock); |
9f9c2364 | 1989 | |
e22bee78 | 1990 | /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ |
63d95a91 | 1991 | mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); |
e22bee78 TH |
1992 | |
1993 | while (true) { | |
051e1850 | 1994 | if (create_worker(pool) || !need_to_create_worker(pool)) |
e22bee78 | 1995 | break; |
1da177e4 | 1996 | |
e212f361 | 1997 | schedule_timeout_interruptible(CREATE_COOLDOWN); |
9f9c2364 | 1998 | |
63d95a91 | 1999 | if (!need_to_create_worker(pool)) |
e22bee78 TH |
2000 | break; |
2001 | } | |
2002 | ||
63d95a91 | 2003 | del_timer_sync(&pool->mayday_timer); |
d565ed63 | 2004 | spin_lock_irq(&pool->lock); |
051e1850 LJ |
2005 | /* |
2006 | * This is necessary even after a new worker was just successfully | |
2007 | * created as @pool->lock was dropped and the new worker might have | |
2008 | * already become busy. | |
2009 | */ | |
63d95a91 | 2010 | if (need_to_create_worker(pool)) |
e22bee78 | 2011 | goto restart; |
e22bee78 TH |
2012 | } |
2013 | ||
73f53c4a | 2014 | /** |
e22bee78 TH |
2015 | * manage_workers - manage worker pool |
2016 | * @worker: self | |
73f53c4a | 2017 | * |
706026c2 | 2018 | * Assume the manager role and manage the worker pool @worker belongs |
e22bee78 | 2019 | * to. At any given time, there can be only zero or one manager per |
706026c2 | 2020 | * pool. The exclusion is handled automatically by this function. |
e22bee78 TH |
2021 | * |
2022 | * The caller can safely start processing works on false return. On | |
2023 | * true return, it's guaranteed that need_to_create_worker() is false | |
2024 | * and may_start_working() is true. | |
73f53c4a TH |
2025 | * |
2026 | * CONTEXT: | |
d565ed63 | 2027 | * spin_lock_irq(pool->lock) which may be released and regrabbed |
e22bee78 TH |
2028 | * multiple times. Does GFP_KERNEL allocations. |
2029 | * | |
d185af30 | 2030 | * Return: |
29187a9e TH |
2031 | * %false if the pool doesn't need management and the caller can safely |
2032 | * start processing works, %true if management function was performed and | |
2033 | * the conditions that the caller verified before calling the function may | |
2034 | * no longer be true. | |
73f53c4a | 2035 | */ |
e22bee78 | 2036 | static bool manage_workers(struct worker *worker) |
73f53c4a | 2037 | { |
63d95a91 | 2038 | struct worker_pool *pool = worker->pool; |
73f53c4a | 2039 | |
692b4825 | 2040 | if (pool->flags & POOL_MANAGER_ACTIVE) |
29187a9e | 2041 | return false; |
692b4825 TH |
2042 | |
2043 | pool->flags |= POOL_MANAGER_ACTIVE; | |
2607d7a6 | 2044 | pool->manager = worker; |
1e19ffc6 | 2045 | |
29187a9e | 2046 | maybe_create_worker(pool); |
e22bee78 | 2047 | |
2607d7a6 | 2048 | pool->manager = NULL; |
692b4825 TH |
2049 | pool->flags &= ~POOL_MANAGER_ACTIVE; |
2050 | wake_up(&wq_manager_wait); | |
29187a9e | 2051 | return true; |
73f53c4a TH |
2052 | } |
2053 | ||
a62428c0 TH |
2054 | /** |
2055 | * process_one_work - process single work | |
c34056a3 | 2056 | * @worker: self |
a62428c0 TH |
2057 | * @work: work to process |
2058 | * | |
2059 | * Process @work. This function contains all the logics necessary to | |
2060 | * process a single work including synchronization against and | |
2061 | * interaction with other workers on the same cpu, queueing and | |
2062 | * flushing. As long as context requirement is met, any worker can | |
2063 | * call this function to process a work. | |
2064 | * | |
2065 | * CONTEXT: | |
d565ed63 | 2066 | * spin_lock_irq(pool->lock) which is released and regrabbed. |
a62428c0 | 2067 | */ |
c34056a3 | 2068 | static void process_one_work(struct worker *worker, struct work_struct *work) |
d565ed63 TH |
2069 | __releases(&pool->lock) |
2070 | __acquires(&pool->lock) | |
a62428c0 | 2071 | { |
112202d9 | 2072 | struct pool_workqueue *pwq = get_work_pwq(work); |
bd7bdd43 | 2073 | struct worker_pool *pool = worker->pool; |
112202d9 | 2074 | bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE; |
73f53c4a | 2075 | int work_color; |
7e11629d | 2076 | struct worker *collision; |
a62428c0 TH |
2077 | #ifdef CONFIG_LOCKDEP |
2078 | /* | |
2079 | * It is permissible to free the struct work_struct from | |
2080 | * inside the function that is called from it, this we need to | |
2081 | * take into account for lockdep too. To avoid bogus "held | |
2082 | * lock freed" warnings as well as problems when looking into | |
2083 | * work->lockdep_map, make a copy and use that here. | |
2084 | */ | |
4d82a1de PZ |
2085 | struct lockdep_map lockdep_map; |
2086 | ||
2087 | lockdep_copy_map(&lockdep_map, &work->lockdep_map); | |
a62428c0 | 2088 | #endif |
807407c0 | 2089 | /* ensure we're on the correct CPU */ |
85327af6 | 2090 | WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && |
ec22ca5e | 2091 | raw_smp_processor_id() != pool->cpu); |
25511a47 | 2092 | |
7e11629d TH |
2093 | /* |
2094 | * A single work shouldn't be executed concurrently by | |
2095 | * multiple workers on a single cpu. Check whether anyone is | |
2096 | * already processing the work. If so, defer the work to the | |
2097 | * currently executing one. | |
2098 | */ | |
c9e7cf27 | 2099 | collision = find_worker_executing_work(pool, work); |
7e11629d TH |
2100 | if (unlikely(collision)) { |
2101 | move_linked_works(work, &collision->scheduled, NULL); | |
2102 | return; | |
2103 | } | |
2104 | ||
8930caba | 2105 | /* claim and dequeue */ |
a62428c0 | 2106 | debug_work_deactivate(work); |
c9e7cf27 | 2107 | hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); |
c34056a3 | 2108 | worker->current_work = work; |
a2c1c57b | 2109 | worker->current_func = work->func; |
112202d9 | 2110 | worker->current_pwq = pwq; |
73f53c4a | 2111 | work_color = get_work_color(work); |
7a22ad75 | 2112 | |
8bf89593 TH |
2113 | /* |
2114 | * Record wq name for cmdline and debug reporting, may get | |
2115 | * overridden through set_worker_desc(). | |
2116 | */ | |
2117 | strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN); | |
2118 | ||
a62428c0 TH |
2119 | list_del_init(&work->entry); |
2120 | ||
fb0e7beb | 2121 | /* |
228f1d00 LJ |
2122 | * CPU intensive works don't participate in concurrency management. |
2123 | * They're the scheduler's responsibility. This takes @worker out | |
2124 | * of concurrency management and the next code block will chain | |
2125 | * execution of the pending work items. | |
fb0e7beb TH |
2126 | */ |
2127 | if (unlikely(cpu_intensive)) | |
228f1d00 | 2128 | worker_set_flags(worker, WORKER_CPU_INTENSIVE); |
fb0e7beb | 2129 | |
974271c4 | 2130 | /* |
a489a03e LJ |
2131 | * Wake up another worker if necessary. The condition is always |
2132 | * false for normal per-cpu workers since nr_running would always | |
2133 | * be >= 1 at this point. This is used to chain execution of the | |
2134 | * pending work items for WORKER_NOT_RUNNING workers such as the | |
228f1d00 | 2135 | * UNBOUND and CPU_INTENSIVE ones. |
974271c4 | 2136 | */ |
a489a03e | 2137 | if (need_more_worker(pool)) |
63d95a91 | 2138 | wake_up_worker(pool); |
974271c4 | 2139 | |
8930caba | 2140 | /* |
7c3eed5c | 2141 | * Record the last pool and clear PENDING which should be the last |
d565ed63 | 2142 | * update to @work. Also, do this inside @pool->lock so that |
23657bb1 TH |
2143 | * PENDING and queued state changes happen together while IRQ is |
2144 | * disabled. | |
8930caba | 2145 | */ |
7c3eed5c | 2146 | set_work_pool_and_clear_pending(work, pool->id); |
a62428c0 | 2147 | |
d565ed63 | 2148 | spin_unlock_irq(&pool->lock); |
a62428c0 | 2149 | |
a1d14934 | 2150 | lock_map_acquire(&pwq->wq->lockdep_map); |
a62428c0 | 2151 | lock_map_acquire(&lockdep_map); |
e6f3faa7 | 2152 | /* |
f52be570 PZ |
2153 | * Strictly speaking we should mark the invariant state without holding |
2154 | * any locks, that is, before these two lock_map_acquire()'s. | |
e6f3faa7 PZ |
2155 | * |
2156 | * However, that would result in: | |
2157 | * | |
2158 | * A(W1) | |
2159 | * WFC(C) | |
2160 | * A(W1) | |
2161 | * C(C) | |
2162 | * | |
2163 | * Which would create W1->C->W1 dependencies, even though there is no | |
2164 | * actual deadlock possible. There are two solutions, using a | |
2165 | * read-recursive acquire on the work(queue) 'locks', but this will then | |
f52be570 | 2166 | * hit the lockdep limitation on recursive locks, or simply discard |
e6f3faa7 PZ |
2167 | * these locks. |
2168 | * | |
2169 | * AFAICT there is no possible deadlock scenario between the | |
2170 | * flush_work() and complete() primitives (except for single-threaded | |
2171 | * workqueues), so hiding them isn't a problem. | |
2172 | */ | |
f52be570 | 2173 | lockdep_invariant_state(true); |
e36c886a | 2174 | trace_workqueue_execute_start(work); |
a2c1c57b | 2175 | worker->current_func(work); |
e36c886a AV |
2176 | /* |
2177 | * While we must be careful to not use "work" after this, the trace | |
2178 | * point will only record its address. | |
2179 | */ | |
2180 | trace_workqueue_execute_end(work); | |
a62428c0 | 2181 | lock_map_release(&lockdep_map); |
112202d9 | 2182 | lock_map_release(&pwq->wq->lockdep_map); |
a62428c0 TH |
2183 | |
2184 | if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { | |
044c782c VI |
2185 | pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" |
2186 | " last function: %pf\n", | |
a2c1c57b TH |
2187 | current->comm, preempt_count(), task_pid_nr(current), |
2188 | worker->current_func); | |
a62428c0 TH |
2189 | debug_show_held_locks(current); |
2190 | dump_stack(); | |
2191 | } | |
2192 | ||
b22ce278 TH |
2193 | /* |
2194 | * The following prevents a kworker from hogging CPU on !PREEMPT | |
2195 | * kernels, where a requeueing work item waiting for something to | |
2196 | * happen could deadlock with stop_machine as such work item could | |
2197 | * indefinitely requeue itself while all other CPUs are trapped in | |
789cbbec JL |
2198 | * stop_machine. At the same time, report a quiescent RCU state so |
2199 | * the same condition doesn't freeze RCU. | |
b22ce278 | 2200 | */ |
a7e6425e | 2201 | cond_resched(); |
b22ce278 | 2202 | |
d565ed63 | 2203 | spin_lock_irq(&pool->lock); |
a62428c0 | 2204 | |
fb0e7beb TH |
2205 | /* clear cpu intensive status */ |
2206 | if (unlikely(cpu_intensive)) | |
2207 | worker_clr_flags(worker, WORKER_CPU_INTENSIVE); | |
2208 | ||
1b69ac6b JW |
2209 | /* tag the worker for identification in schedule() */ |
2210 | worker->last_func = worker->current_func; | |
2211 | ||
a62428c0 | 2212 | /* we're done with it, release */ |
42f8570f | 2213 | hash_del(&worker->hentry); |
c34056a3 | 2214 | worker->current_work = NULL; |
a2c1c57b | 2215 | worker->current_func = NULL; |
112202d9 TH |
2216 | worker->current_pwq = NULL; |
2217 | pwq_dec_nr_in_flight(pwq, work_color); | |
a62428c0 TH |
2218 | } |
2219 | ||
affee4b2 TH |
2220 | /** |
2221 | * process_scheduled_works - process scheduled works | |
2222 | * @worker: self | |
2223 | * | |
2224 | * Process all scheduled works. Please note that the scheduled list | |
2225 | * may change while processing a work, so this function repeatedly | |
2226 | * fetches a work from the top and executes it. | |
2227 | * | |
2228 | * CONTEXT: | |
d565ed63 | 2229 | * spin_lock_irq(pool->lock) which may be released and regrabbed |
affee4b2 TH |
2230 | * multiple times. |
2231 | */ | |
2232 | static void process_scheduled_works(struct worker *worker) | |
1da177e4 | 2233 | { |
affee4b2 TH |
2234 | while (!list_empty(&worker->scheduled)) { |
2235 | struct work_struct *work = list_first_entry(&worker->scheduled, | |
1da177e4 | 2236 | struct work_struct, entry); |
c34056a3 | 2237 | process_one_work(worker, work); |
1da177e4 | 2238 | } |
1da177e4 LT |
2239 | } |
2240 | ||
197f6acc TH |
2241 | static void set_pf_worker(bool val) |
2242 | { | |
2243 | mutex_lock(&wq_pool_attach_mutex); | |
2244 | if (val) | |
2245 | current->flags |= PF_WQ_WORKER; | |
2246 | else | |
2247 | current->flags &= ~PF_WQ_WORKER; | |
2248 | mutex_unlock(&wq_pool_attach_mutex); | |
2249 | } | |
2250 | ||
4690c4ab TH |
2251 | /** |
2252 | * worker_thread - the worker thread function | |
c34056a3 | 2253 | * @__worker: self |
4690c4ab | 2254 | * |
c5aa87bb TH |
2255 | * The worker thread function. All workers belong to a worker_pool - |
2256 | * either a per-cpu one or dynamic unbound one. These workers process all | |
2257 | * work items regardless of their specific target workqueue. The only | |
2258 | * exception is work items which belong to workqueues with a rescuer which | |
2259 | * will be explained in rescuer_thread(). | |
d185af30 YB |
2260 | * |
2261 | * Return: 0 | |
4690c4ab | 2262 | */ |
c34056a3 | 2263 | static int worker_thread(void *__worker) |
1da177e4 | 2264 | { |
c34056a3 | 2265 | struct worker *worker = __worker; |
bd7bdd43 | 2266 | struct worker_pool *pool = worker->pool; |
1da177e4 | 2267 | |
e22bee78 | 2268 | /* tell the scheduler that this is a workqueue worker */ |
197f6acc | 2269 | set_pf_worker(true); |
c8e55f36 | 2270 | woke_up: |
d565ed63 | 2271 | spin_lock_irq(&pool->lock); |
1da177e4 | 2272 | |
a9ab775b TH |
2273 | /* am I supposed to die? */ |
2274 | if (unlikely(worker->flags & WORKER_DIE)) { | |
d565ed63 | 2275 | spin_unlock_irq(&pool->lock); |
a9ab775b | 2276 | WARN_ON_ONCE(!list_empty(&worker->entry)); |
197f6acc | 2277 | set_pf_worker(false); |
60f5a4bc LJ |
2278 | |
2279 | set_task_comm(worker->task, "kworker/dying"); | |
7cda9aae | 2280 | ida_simple_remove(&pool->worker_ida, worker->id); |
a2d812a2 | 2281 | worker_detach_from_pool(worker); |
60f5a4bc | 2282 | kfree(worker); |
a9ab775b | 2283 | return 0; |
c8e55f36 | 2284 | } |
affee4b2 | 2285 | |
c8e55f36 | 2286 | worker_leave_idle(worker); |
db7bccf4 | 2287 | recheck: |
e22bee78 | 2288 | /* no more worker necessary? */ |
63d95a91 | 2289 | if (!need_more_worker(pool)) |
e22bee78 TH |
2290 | goto sleep; |
2291 | ||
2292 | /* do we need to manage? */ | |
63d95a91 | 2293 | if (unlikely(!may_start_working(pool)) && manage_workers(worker)) |
e22bee78 TH |
2294 | goto recheck; |
2295 | ||
c8e55f36 TH |
2296 | /* |
2297 | * ->scheduled list can only be filled while a worker is | |
2298 | * preparing to process a work or actually processing it. | |
2299 | * Make sure nobody diddled with it while I was sleeping. | |
2300 | */ | |
6183c009 | 2301 | WARN_ON_ONCE(!list_empty(&worker->scheduled)); |
c8e55f36 | 2302 | |
e22bee78 | 2303 | /* |
a9ab775b TH |
2304 | * Finish PREP stage. We're guaranteed to have at least one idle |
2305 | * worker or that someone else has already assumed the manager | |
2306 | * role. This is where @worker starts participating in concurrency | |
2307 | * management if applicable and concurrency management is restored | |
2308 | * after being rebound. See rebind_workers() for details. | |
e22bee78 | 2309 | */ |
a9ab775b | 2310 | worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); |
e22bee78 TH |
2311 | |
2312 | do { | |
c8e55f36 | 2313 | struct work_struct *work = |
bd7bdd43 | 2314 | list_first_entry(&pool->worklist, |
c8e55f36 TH |
2315 | struct work_struct, entry); |
2316 | ||
82607adc TH |
2317 | pool->watchdog_ts = jiffies; |
2318 | ||
c8e55f36 TH |
2319 | if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { |
2320 | /* optimization path, not strictly necessary */ | |
2321 | process_one_work(worker, work); | |
2322 | if (unlikely(!list_empty(&worker->scheduled))) | |
affee4b2 | 2323 | process_scheduled_works(worker); |
c8e55f36 TH |
2324 | } else { |
2325 | move_linked_works(work, &worker->scheduled, NULL); | |
2326 | process_scheduled_works(worker); | |
affee4b2 | 2327 | } |
63d95a91 | 2328 | } while (keep_working(pool)); |
e22bee78 | 2329 | |
228f1d00 | 2330 | worker_set_flags(worker, WORKER_PREP); |
d313dd85 | 2331 | sleep: |
c8e55f36 | 2332 | /* |
d565ed63 TH |
2333 | * pool->lock is held and there's no work to process and no need to |
2334 | * manage, sleep. Workers are woken up only while holding | |
2335 | * pool->lock or from local cpu, so setting the current state | |
2336 | * before releasing pool->lock is enough to prevent losing any | |
2337 | * event. | |
c8e55f36 TH |
2338 | */ |
2339 | worker_enter_idle(worker); | |
c5a94a61 | 2340 | __set_current_state(TASK_IDLE); |
d565ed63 | 2341 | spin_unlock_irq(&pool->lock); |
c8e55f36 TH |
2342 | schedule(); |
2343 | goto woke_up; | |
1da177e4 LT |
2344 | } |
2345 | ||
e22bee78 TH |
2346 | /** |
2347 | * rescuer_thread - the rescuer thread function | |
111c225a | 2348 | * @__rescuer: self |
e22bee78 TH |
2349 | * |
2350 | * Workqueue rescuer thread function. There's one rescuer for each | |
493008a8 | 2351 | * workqueue which has WQ_MEM_RECLAIM set. |
e22bee78 | 2352 | * |
706026c2 | 2353 | * Regular work processing on a pool may block trying to create a new |
e22bee78 TH |
2354 | * worker which uses GFP_KERNEL allocation which has slight chance of |
2355 | * developing into deadlock if some works currently on the same queue | |
2356 | * need to be processed to satisfy the GFP_KERNEL allocation. This is | |
2357 | * the problem rescuer solves. | |
2358 | * | |
706026c2 TH |
2359 | * When such condition is possible, the pool summons rescuers of all |
2360 | * workqueues which have works queued on the pool and let them process | |
e22bee78 TH |
2361 | * those works so that forward progress can be guaranteed. |
2362 | * | |
2363 | * This should happen rarely. | |
d185af30 YB |
2364 | * |
2365 | * Return: 0 | |
e22bee78 | 2366 | */ |
111c225a | 2367 | static int rescuer_thread(void *__rescuer) |
e22bee78 | 2368 | { |
111c225a TH |
2369 | struct worker *rescuer = __rescuer; |
2370 | struct workqueue_struct *wq = rescuer->rescue_wq; | |
e22bee78 | 2371 | struct list_head *scheduled = &rescuer->scheduled; |
4d595b86 | 2372 | bool should_stop; |
e22bee78 TH |
2373 | |
2374 | set_user_nice(current, RESCUER_NICE_LEVEL); | |
111c225a TH |
2375 | |
2376 | /* | |
2377 | * Mark rescuer as worker too. As WORKER_PREP is never cleared, it | |
2378 | * doesn't participate in concurrency management. | |
2379 | */ | |
197f6acc | 2380 | set_pf_worker(true); |
e22bee78 | 2381 | repeat: |
c5a94a61 | 2382 | set_current_state(TASK_IDLE); |
e22bee78 | 2383 | |
4d595b86 LJ |
2384 | /* |
2385 | * By the time the rescuer is requested to stop, the workqueue | |
2386 | * shouldn't have any work pending, but @wq->maydays may still have | |
2387 | * pwq(s) queued. This can happen by non-rescuer workers consuming | |
2388 | * all the work items before the rescuer got to them. Go through | |
2389 | * @wq->maydays processing before acting on should_stop so that the | |
2390 | * list is always empty on exit. | |
2391 | */ | |
2392 | should_stop = kthread_should_stop(); | |
e22bee78 | 2393 | |
493a1724 | 2394 | /* see whether any pwq is asking for help */ |
2e109a28 | 2395 | spin_lock_irq(&wq_mayday_lock); |
493a1724 TH |
2396 | |
2397 | while (!list_empty(&wq->maydays)) { | |
2398 | struct pool_workqueue *pwq = list_first_entry(&wq->maydays, | |
2399 | struct pool_workqueue, mayday_node); | |
112202d9 | 2400 | struct worker_pool *pool = pwq->pool; |
e22bee78 | 2401 | struct work_struct *work, *n; |
82607adc | 2402 | bool first = true; |
e22bee78 TH |
2403 | |
2404 | __set_current_state(TASK_RUNNING); | |
493a1724 TH |
2405 | list_del_init(&pwq->mayday_node); |
2406 | ||
2e109a28 | 2407 | spin_unlock_irq(&wq_mayday_lock); |
e22bee78 | 2408 | |
51697d39 LJ |
2409 | worker_attach_to_pool(rescuer, pool); |
2410 | ||
2411 | spin_lock_irq(&pool->lock); | |
e22bee78 TH |
2412 | |
2413 | /* | |
2414 | * Slurp in all works issued via this workqueue and | |
2415 | * process'em. | |
2416 | */ | |
0479c8c5 | 2417 | WARN_ON_ONCE(!list_empty(scheduled)); |
82607adc TH |
2418 | list_for_each_entry_safe(work, n, &pool->worklist, entry) { |
2419 | if (get_work_pwq(work) == pwq) { | |
2420 | if (first) | |
2421 | pool->watchdog_ts = jiffies; | |
e22bee78 | 2422 | move_linked_works(work, scheduled, &n); |
82607adc TH |
2423 | } |
2424 | first = false; | |
2425 | } | |
e22bee78 | 2426 | |
008847f6 N |
2427 | if (!list_empty(scheduled)) { |
2428 | process_scheduled_works(rescuer); | |
2429 | ||
2430 | /* | |
2431 | * The above execution of rescued work items could | |
2432 | * have created more to rescue through | |
2433 | * pwq_activate_first_delayed() or chained | |
2434 | * queueing. Let's put @pwq back on mayday list so | |
2435 | * that such back-to-back work items, which may be | |
2436 | * being used to relieve memory pressure, don't | |
2437 | * incur MAYDAY_INTERVAL delay inbetween. | |
2438 | */ | |
2439 | if (need_to_create_worker(pool)) { | |
2440 | spin_lock(&wq_mayday_lock); | |
2441 | get_pwq(pwq); | |
2442 | list_move_tail(&pwq->mayday_node, &wq->maydays); | |
2443 | spin_unlock(&wq_mayday_lock); | |
2444 | } | |
2445 | } | |
7576958a | 2446 | |
77668c8b LJ |
2447 | /* |
2448 | * Put the reference grabbed by send_mayday(). @pool won't | |
13b1d625 | 2449 | * go away while we're still attached to it. |
77668c8b LJ |
2450 | */ |
2451 | put_pwq(pwq); | |
2452 | ||
7576958a | 2453 | /* |
d8ca83e6 | 2454 | * Leave this pool. If need_more_worker() is %true, notify a |
7576958a TH |
2455 | * regular worker; otherwise, we end up with 0 concurrency |
2456 | * and stalling the execution. | |
2457 | */ | |
d8ca83e6 | 2458 | if (need_more_worker(pool)) |
63d95a91 | 2459 | wake_up_worker(pool); |
7576958a | 2460 | |
13b1d625 LJ |
2461 | spin_unlock_irq(&pool->lock); |
2462 | ||
a2d812a2 | 2463 | worker_detach_from_pool(rescuer); |
13b1d625 LJ |
2464 | |
2465 | spin_lock_irq(&wq_mayday_lock); | |
e22bee78 TH |
2466 | } |
2467 | ||
2e109a28 | 2468 | spin_unlock_irq(&wq_mayday_lock); |
493a1724 | 2469 | |
4d595b86 LJ |
2470 | if (should_stop) { |
2471 | __set_current_state(TASK_RUNNING); | |
197f6acc | 2472 | set_pf_worker(false); |
4d595b86 LJ |
2473 | return 0; |
2474 | } | |
2475 | ||
111c225a TH |
2476 | /* rescuers should never participate in concurrency management */ |
2477 | WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); | |
e22bee78 TH |
2478 | schedule(); |
2479 | goto repeat; | |
1da177e4 LT |
2480 | } |
2481 | ||
fca839c0 TH |
2482 | /** |
2483 | * check_flush_dependency - check for flush dependency sanity | |
2484 | * @target_wq: workqueue being flushed | |
2485 | * @target_work: work item being flushed (NULL for workqueue flushes) | |
2486 | * | |
2487 | * %current is trying to flush the whole @target_wq or @target_work on it. | |
2488 | * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not | |
2489 | * reclaiming memory or running on a workqueue which doesn't have | |
2490 | * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to | |
2491 | * a deadlock. | |
2492 | */ | |
2493 | static void check_flush_dependency(struct workqueue_struct *target_wq, | |
2494 | struct work_struct *target_work) | |
2495 | { | |
2496 | work_func_t target_func = target_work ? target_work->func : NULL; | |
2497 | struct worker *worker; | |
2498 | ||
2499 | if (target_wq->flags & WQ_MEM_RECLAIM) | |
2500 | return; | |
2501 | ||
2502 | worker = current_wq_worker(); | |
2503 | ||
2504 | WARN_ONCE(current->flags & PF_MEMALLOC, | |
2505 | "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%pf", | |
2506 | current->pid, current->comm, target_wq->name, target_func); | |
23d11a58 TH |
2507 | WARN_ONCE(worker && ((worker->current_pwq->wq->flags & |
2508 | (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM), | |
fca839c0 TH |
2509 | "workqueue: WQ_MEM_RECLAIM %s:%pf is flushing !WQ_MEM_RECLAIM %s:%pf", |
2510 | worker->current_pwq->wq->name, worker->current_func, | |
2511 | target_wq->name, target_func); | |
2512 | } | |
2513 | ||
fc2e4d70 ON |
2514 | struct wq_barrier { |
2515 | struct work_struct work; | |
2516 | struct completion done; | |
2607d7a6 | 2517 | struct task_struct *task; /* purely informational */ |
fc2e4d70 ON |
2518 | }; |
2519 | ||
2520 | static void wq_barrier_func(struct work_struct *work) | |
2521 | { | |
2522 | struct wq_barrier *barr = container_of(work, struct wq_barrier, work); | |
2523 | complete(&barr->done); | |
2524 | } | |
2525 | ||
4690c4ab TH |
2526 | /** |
2527 | * insert_wq_barrier - insert a barrier work | |
112202d9 | 2528 | * @pwq: pwq to insert barrier into |
4690c4ab | 2529 | * @barr: wq_barrier to insert |
affee4b2 TH |
2530 | * @target: target work to attach @barr to |
2531 | * @worker: worker currently executing @target, NULL if @target is not executing | |
4690c4ab | 2532 | * |
affee4b2 TH |
2533 | * @barr is linked to @target such that @barr is completed only after |
2534 | * @target finishes execution. Please note that the ordering | |
2535 | * guarantee is observed only with respect to @target and on the local | |
2536 | * cpu. | |
2537 | * | |
2538 | * Currently, a queued barrier can't be canceled. This is because | |
2539 | * try_to_grab_pending() can't determine whether the work to be | |
2540 | * grabbed is at the head of the queue and thus can't clear LINKED | |
2541 | * flag of the previous work while there must be a valid next work | |
2542 | * after a work with LINKED flag set. | |
2543 | * | |
2544 | * Note that when @worker is non-NULL, @target may be modified | |
112202d9 | 2545 | * underneath us, so we can't reliably determine pwq from @target. |
4690c4ab TH |
2546 | * |
2547 | * CONTEXT: | |
d565ed63 | 2548 | * spin_lock_irq(pool->lock). |
4690c4ab | 2549 | */ |
112202d9 | 2550 | static void insert_wq_barrier(struct pool_workqueue *pwq, |
affee4b2 TH |
2551 | struct wq_barrier *barr, |
2552 | struct work_struct *target, struct worker *worker) | |
fc2e4d70 | 2553 | { |
affee4b2 TH |
2554 | struct list_head *head; |
2555 | unsigned int linked = 0; | |
2556 | ||
dc186ad7 | 2557 | /* |
d565ed63 | 2558 | * debugobject calls are safe here even with pool->lock locked |
dc186ad7 TG |
2559 | * as we know for sure that this will not trigger any of the |
2560 | * checks and call back into the fixup functions where we | |
2561 | * might deadlock. | |
2562 | */ | |
ca1cab37 | 2563 | INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); |
22df02bb | 2564 | __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); |
52fa5bc5 | 2565 | |
fd1a5b04 BP |
2566 | init_completion_map(&barr->done, &target->lockdep_map); |
2567 | ||
2607d7a6 | 2568 | barr->task = current; |
83c22520 | 2569 | |
affee4b2 TH |
2570 | /* |
2571 | * If @target is currently being executed, schedule the | |
2572 | * barrier to the worker; otherwise, put it after @target. | |
2573 | */ | |
2574 | if (worker) | |
2575 | head = worker->scheduled.next; | |
2576 | else { | |
2577 | unsigned long *bits = work_data_bits(target); | |
2578 | ||
2579 | head = target->entry.next; | |
2580 | /* there can already be other linked works, inherit and set */ | |
2581 | linked = *bits & WORK_STRUCT_LINKED; | |
2582 | __set_bit(WORK_STRUCT_LINKED_BIT, bits); | |
2583 | } | |
2584 | ||
dc186ad7 | 2585 | debug_work_activate(&barr->work); |
112202d9 | 2586 | insert_work(pwq, &barr->work, head, |
affee4b2 | 2587 | work_color_to_flags(WORK_NO_COLOR) | linked); |
fc2e4d70 ON |
2588 | } |
2589 | ||
73f53c4a | 2590 | /** |
112202d9 | 2591 | * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing |
73f53c4a TH |
2592 | * @wq: workqueue being flushed |
2593 | * @flush_color: new flush color, < 0 for no-op | |
2594 | * @work_color: new work color, < 0 for no-op | |
2595 | * | |
112202d9 | 2596 | * Prepare pwqs for workqueue flushing. |
73f53c4a | 2597 | * |
112202d9 TH |
2598 | * If @flush_color is non-negative, flush_color on all pwqs should be |
2599 | * -1. If no pwq has in-flight commands at the specified color, all | |
2600 | * pwq->flush_color's stay at -1 and %false is returned. If any pwq | |
2601 | * has in flight commands, its pwq->flush_color is set to | |
2602 | * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq | |
73f53c4a TH |
2603 | * wakeup logic is armed and %true is returned. |
2604 | * | |
2605 | * The caller should have initialized @wq->first_flusher prior to | |
2606 | * calling this function with non-negative @flush_color. If | |
2607 | * @flush_color is negative, no flush color update is done and %false | |
2608 | * is returned. | |
2609 | * | |
112202d9 | 2610 | * If @work_color is non-negative, all pwqs should have the same |
73f53c4a TH |
2611 | * work_color which is previous to @work_color and all will be |
2612 | * advanced to @work_color. | |
2613 | * | |
2614 | * CONTEXT: | |
3c25a55d | 2615 | * mutex_lock(wq->mutex). |
73f53c4a | 2616 | * |
d185af30 | 2617 | * Return: |
73f53c4a TH |
2618 | * %true if @flush_color >= 0 and there's something to flush. %false |
2619 | * otherwise. | |
2620 | */ | |
112202d9 | 2621 | static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, |
73f53c4a | 2622 | int flush_color, int work_color) |
1da177e4 | 2623 | { |
73f53c4a | 2624 | bool wait = false; |
49e3cf44 | 2625 | struct pool_workqueue *pwq; |
1da177e4 | 2626 | |
73f53c4a | 2627 | if (flush_color >= 0) { |
6183c009 | 2628 | WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); |
112202d9 | 2629 | atomic_set(&wq->nr_pwqs_to_flush, 1); |
1da177e4 | 2630 | } |
2355b70f | 2631 | |
49e3cf44 | 2632 | for_each_pwq(pwq, wq) { |
112202d9 | 2633 | struct worker_pool *pool = pwq->pool; |
fc2e4d70 | 2634 | |
b09f4fd3 | 2635 | spin_lock_irq(&pool->lock); |
83c22520 | 2636 | |
73f53c4a | 2637 | if (flush_color >= 0) { |
6183c009 | 2638 | WARN_ON_ONCE(pwq->flush_color != -1); |
fc2e4d70 | 2639 | |
112202d9 TH |
2640 | if (pwq->nr_in_flight[flush_color]) { |
2641 | pwq->flush_color = flush_color; | |
2642 | atomic_inc(&wq->nr_pwqs_to_flush); | |
73f53c4a TH |
2643 | wait = true; |
2644 | } | |
2645 | } | |
1da177e4 | 2646 | |
73f53c4a | 2647 | if (work_color >= 0) { |
6183c009 | 2648 | WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); |
112202d9 | 2649 | pwq->work_color = work_color; |
73f53c4a | 2650 | } |
1da177e4 | 2651 | |
b09f4fd3 | 2652 | spin_unlock_irq(&pool->lock); |
1da177e4 | 2653 | } |
2355b70f | 2654 | |
112202d9 | 2655 | if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) |
73f53c4a | 2656 | complete(&wq->first_flusher->done); |
14441960 | 2657 | |
73f53c4a | 2658 | return wait; |
1da177e4 LT |
2659 | } |
2660 | ||
0fcb78c2 | 2661 | /** |
1da177e4 | 2662 | * flush_workqueue - ensure that any scheduled work has run to completion. |
0fcb78c2 | 2663 | * @wq: workqueue to flush |
1da177e4 | 2664 | * |
c5aa87bb TH |
2665 | * This function sleeps until all work items which were queued on entry |
2666 | * have finished execution, but it is not livelocked by new incoming ones. | |
1da177e4 | 2667 | */ |
7ad5b3a5 | 2668 | void flush_workqueue(struct workqueue_struct *wq) |
1da177e4 | 2669 | { |
73f53c4a TH |
2670 | struct wq_flusher this_flusher = { |
2671 | .list = LIST_HEAD_INIT(this_flusher.list), | |
2672 | .flush_color = -1, | |
fd1a5b04 | 2673 | .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map), |
73f53c4a TH |
2674 | }; |
2675 | int next_color; | |
1da177e4 | 2676 | |
3347fa09 TH |
2677 | if (WARN_ON(!wq_online)) |
2678 | return; | |
2679 | ||
87915adc JB |
2680 | lock_map_acquire(&wq->lockdep_map); |
2681 | lock_map_release(&wq->lockdep_map); | |
2682 | ||
3c25a55d | 2683 | mutex_lock(&wq->mutex); |
73f53c4a TH |
2684 | |
2685 | /* | |
2686 | * Start-to-wait phase | |
2687 | */ | |
2688 | next_color = work_next_color(wq->work_color); | |
2689 | ||
2690 | if (next_color != wq->flush_color) { | |
2691 | /* | |
2692 | * Color space is not full. The current work_color | |
2693 | * becomes our flush_color and work_color is advanced | |
2694 | * by one. | |
2695 | */ | |
6183c009 | 2696 | WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); |
73f53c4a TH |
2697 | this_flusher.flush_color = wq->work_color; |
2698 | wq->work_color = next_color; | |
2699 | ||
2700 | if (!wq->first_flusher) { | |
2701 | /* no flush in progress, become the first flusher */ | |
6183c009 | 2702 | WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); |
73f53c4a TH |
2703 | |
2704 | wq->first_flusher = &this_flusher; | |
2705 | ||
112202d9 | 2706 | if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, |
73f53c4a TH |
2707 | wq->work_color)) { |
2708 | /* nothing to flush, done */ | |
2709 | wq->flush_color = next_color; | |
2710 | wq->first_flusher = NULL; | |
2711 | goto out_unlock; | |
2712 | } | |
2713 | } else { | |
2714 | /* wait in queue */ | |
6183c009 | 2715 | WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); |
73f53c4a | 2716 | list_add_tail(&this_flusher.list, &wq->flusher_queue); |
112202d9 | 2717 | flush_workqueue_prep_pwqs(wq, -1, wq->work_color); |
73f53c4a TH |
2718 | } |
2719 | } else { | |
2720 | /* | |
2721 | * Oops, color space is full, wait on overflow queue. | |
2722 | * The next flush completion will assign us | |
2723 | * flush_color and transfer to flusher_queue. | |
2724 | */ | |
2725 | list_add_tail(&this_flusher.list, &wq->flusher_overflow); | |
2726 | } | |
2727 | ||
fca839c0 TH |
2728 | check_flush_dependency(wq, NULL); |
2729 | ||
3c25a55d | 2730 | mutex_unlock(&wq->mutex); |
73f53c4a TH |
2731 | |
2732 | wait_for_completion(&this_flusher.done); | |
2733 | ||
2734 | /* | |
2735 | * Wake-up-and-cascade phase | |
2736 | * | |
2737 | * First flushers are responsible for cascading flushes and | |
2738 | * handling overflow. Non-first flushers can simply return. | |
2739 | */ | |
2740 | if (wq->first_flusher != &this_flusher) | |
2741 | return; | |
2742 | ||
3c25a55d | 2743 | mutex_lock(&wq->mutex); |
73f53c4a | 2744 | |
4ce48b37 TH |
2745 | /* we might have raced, check again with mutex held */ |
2746 | if (wq->first_flusher != &this_flusher) | |
2747 | goto out_unlock; | |
2748 | ||
73f53c4a TH |
2749 | wq->first_flusher = NULL; |
2750 | ||
6183c009 TH |
2751 | WARN_ON_ONCE(!list_empty(&this_flusher.list)); |
2752 | WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); | |
73f53c4a TH |
2753 | |
2754 | while (true) { | |
2755 | struct wq_flusher *next, *tmp; | |
2756 | ||
2757 | /* complete all the flushers sharing the current flush color */ | |
2758 | list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { | |
2759 | if (next->flush_color != wq->flush_color) | |
2760 | break; | |
2761 | list_del_init(&next->list); | |
2762 | complete(&next->done); | |
2763 | } | |
2764 | ||
6183c009 TH |
2765 | WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && |
2766 | wq->flush_color != work_next_color(wq->work_color)); | |
73f53c4a TH |
2767 | |
2768 | /* this flush_color is finished, advance by one */ | |
2769 | wq->flush_color = work_next_color(wq->flush_color); | |
2770 | ||
2771 | /* one color has been freed, handle overflow queue */ | |
2772 | if (!list_empty(&wq->flusher_overflow)) { | |
2773 | /* | |
2774 | * Assign the same color to all overflowed | |
2775 | * flushers, advance work_color and append to | |
2776 | * flusher_queue. This is the start-to-wait | |
2777 | * phase for these overflowed flushers. | |
2778 | */ | |
2779 | list_for_each_entry(tmp, &wq->flusher_overflow, list) | |
2780 | tmp->flush_color = wq->work_color; | |
2781 | ||
2782 | wq->work_color = work_next_color(wq->work_color); | |
2783 | ||
2784 | list_splice_tail_init(&wq->flusher_overflow, | |
2785 | &wq->flusher_queue); | |
112202d9 | 2786 | flush_workqueue_prep_pwqs(wq, -1, wq->work_color); |
73f53c4a TH |
2787 | } |
2788 | ||
2789 | if (list_empty(&wq->flusher_queue)) { | |
6183c009 | 2790 | WARN_ON_ONCE(wq->flush_color != wq->work_color); |
73f53c4a TH |
2791 | break; |
2792 | } | |
2793 | ||
2794 | /* | |
2795 | * Need to flush more colors. Make the next flusher | |
112202d9 | 2796 | * the new first flusher and arm pwqs. |
73f53c4a | 2797 | */ |
6183c009 TH |
2798 | WARN_ON_ONCE(wq->flush_color == wq->work_color); |
2799 | WARN_ON_ONCE(wq->flush_color != next->flush_color); | |
73f53c4a TH |
2800 | |
2801 | list_del_init(&next->list); | |
2802 | wq->first_flusher = next; | |
2803 | ||
112202d9 | 2804 | if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) |
73f53c4a TH |
2805 | break; |
2806 | ||
2807 | /* | |
2808 | * Meh... this color is already done, clear first | |
2809 | * flusher and repeat cascading. | |
2810 | */ | |
2811 | wq->first_flusher = NULL; | |
2812 | } | |
2813 | ||
2814 | out_unlock: | |
3c25a55d | 2815 | mutex_unlock(&wq->mutex); |
1da177e4 | 2816 | } |
1dadafa8 | 2817 | EXPORT_SYMBOL(flush_workqueue); |
1da177e4 | 2818 | |
9c5a2ba7 TH |
2819 | /** |
2820 | * drain_workqueue - drain a workqueue | |
2821 | * @wq: workqueue to drain | |
2822 | * | |
2823 | * Wait until the workqueue becomes empty. While draining is in progress, | |
2824 | * only chain queueing is allowed. IOW, only currently pending or running | |
2825 | * work items on @wq can queue further work items on it. @wq is flushed | |
b749b1b6 | 2826 | * repeatedly until it becomes empty. The number of flushing is determined |
9c5a2ba7 TH |
2827 | * by the depth of chaining and should be relatively short. Whine if it |
2828 | * takes too long. | |
2829 | */ | |
2830 | void drain_workqueue(struct workqueue_struct *wq) | |
2831 | { | |
2832 | unsigned int flush_cnt = 0; | |
49e3cf44 | 2833 | struct pool_workqueue *pwq; |
9c5a2ba7 TH |
2834 | |
2835 | /* | |
2836 | * __queue_work() needs to test whether there are drainers, is much | |
2837 | * hotter than drain_workqueue() and already looks at @wq->flags. | |
618b01eb | 2838 | * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. |
9c5a2ba7 | 2839 | */ |
87fc741e | 2840 | mutex_lock(&wq->mutex); |
9c5a2ba7 | 2841 | if (!wq->nr_drainers++) |
618b01eb | 2842 | wq->flags |= __WQ_DRAINING; |
87fc741e | 2843 | mutex_unlock(&wq->mutex); |
9c5a2ba7 TH |
2844 | reflush: |
2845 | flush_workqueue(wq); | |
2846 | ||
b09f4fd3 | 2847 | mutex_lock(&wq->mutex); |
76af4d93 | 2848 | |
49e3cf44 | 2849 | for_each_pwq(pwq, wq) { |
fa2563e4 | 2850 | bool drained; |
9c5a2ba7 | 2851 | |
b09f4fd3 | 2852 | spin_lock_irq(&pwq->pool->lock); |
112202d9 | 2853 | drained = !pwq->nr_active && list_empty(&pwq->delayed_works); |
b09f4fd3 | 2854 | spin_unlock_irq(&pwq->pool->lock); |
fa2563e4 TT |
2855 | |
2856 | if (drained) | |
9c5a2ba7 TH |
2857 | continue; |
2858 | ||
2859 | if (++flush_cnt == 10 || | |
2860 | (flush_cnt % 100 == 0 && flush_cnt <= 1000)) | |
c5aa87bb | 2861 | pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n", |
044c782c | 2862 | wq->name, flush_cnt); |
76af4d93 | 2863 | |
b09f4fd3 | 2864 | mutex_unlock(&wq->mutex); |
9c5a2ba7 TH |
2865 | goto reflush; |
2866 | } | |
2867 | ||
9c5a2ba7 | 2868 | if (!--wq->nr_drainers) |
618b01eb | 2869 | wq->flags &= ~__WQ_DRAINING; |
87fc741e | 2870 | mutex_unlock(&wq->mutex); |
9c5a2ba7 TH |
2871 | } |
2872 | EXPORT_SYMBOL_GPL(drain_workqueue); | |
2873 | ||
d6e89786 JB |
2874 | static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr, |
2875 | bool from_cancel) | |
db700897 | 2876 | { |
affee4b2 | 2877 | struct worker *worker = NULL; |
c9e7cf27 | 2878 | struct worker_pool *pool; |
112202d9 | 2879 | struct pool_workqueue *pwq; |
db700897 ON |
2880 | |
2881 | might_sleep(); | |
fa1b54e6 TH |
2882 | |
2883 | local_irq_disable(); | |
c9e7cf27 | 2884 | pool = get_work_pool(work); |
fa1b54e6 TH |
2885 | if (!pool) { |
2886 | local_irq_enable(); | |
baf59022 | 2887 | return false; |
fa1b54e6 | 2888 | } |
db700897 | 2889 | |
fa1b54e6 | 2890 | spin_lock(&pool->lock); |
0b3dae68 | 2891 | /* see the comment in try_to_grab_pending() with the same code */ |
112202d9 TH |
2892 | pwq = get_work_pwq(work); |
2893 | if (pwq) { | |
2894 | if (unlikely(pwq->pool != pool)) | |
4690c4ab | 2895 | goto already_gone; |
606a5020 | 2896 | } else { |
c9e7cf27 | 2897 | worker = find_worker_executing_work(pool, work); |
affee4b2 | 2898 | if (!worker) |
4690c4ab | 2899 | goto already_gone; |
112202d9 | 2900 | pwq = worker->current_pwq; |
606a5020 | 2901 | } |
db700897 | 2902 | |
fca839c0 TH |
2903 | check_flush_dependency(pwq->wq, work); |
2904 | ||
112202d9 | 2905 | insert_wq_barrier(pwq, barr, work, worker); |
d565ed63 | 2906 | spin_unlock_irq(&pool->lock); |
7a22ad75 | 2907 | |
e159489b | 2908 | /* |
a1d14934 PZ |
2909 | * Force a lock recursion deadlock when using flush_work() inside a |
2910 | * single-threaded or rescuer equipped workqueue. | |
2911 | * | |
2912 | * For single threaded workqueues the deadlock happens when the work | |
2913 | * is after the work issuing the flush_work(). For rescuer equipped | |
2914 | * workqueues the deadlock happens when the rescuer stalls, blocking | |
2915 | * forward progress. | |
e159489b | 2916 | */ |
d6e89786 JB |
2917 | if (!from_cancel && |
2918 | (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) { | |
112202d9 | 2919 | lock_map_acquire(&pwq->wq->lockdep_map); |
a1d14934 PZ |
2920 | lock_map_release(&pwq->wq->lockdep_map); |
2921 | } | |
e159489b | 2922 | |
401a8d04 | 2923 | return true; |
4690c4ab | 2924 | already_gone: |
d565ed63 | 2925 | spin_unlock_irq(&pool->lock); |
401a8d04 | 2926 | return false; |
db700897 | 2927 | } |
baf59022 | 2928 | |
d6e89786 JB |
2929 | static bool __flush_work(struct work_struct *work, bool from_cancel) |
2930 | { | |
2931 | struct wq_barrier barr; | |
2932 | ||
2933 | if (WARN_ON(!wq_online)) | |
2934 | return false; | |
2935 | ||
87915adc JB |
2936 | if (!from_cancel) { |
2937 | lock_map_acquire(&work->lockdep_map); | |
2938 | lock_map_release(&work->lockdep_map); | |
2939 | } | |
2940 | ||
d6e89786 JB |
2941 | if (start_flush_work(work, &barr, from_cancel)) { |
2942 | wait_for_completion(&barr.done); | |
2943 | destroy_work_on_stack(&barr.work); | |
2944 | return true; | |
2945 | } else { | |
2946 | return false; | |
2947 | } | |
2948 | } | |
2949 | ||
baf59022 TH |
2950 | /** |
2951 | * flush_work - wait for a work to finish executing the last queueing instance | |
2952 | * @work: the work to flush | |
2953 | * | |
606a5020 TH |
2954 | * Wait until @work has finished execution. @work is guaranteed to be idle |
2955 | * on return if it hasn't been requeued since flush started. | |
baf59022 | 2956 | * |
d185af30 | 2957 | * Return: |
baf59022 TH |
2958 | * %true if flush_work() waited for the work to finish execution, |
2959 | * %false if it was already idle. | |
2960 | */ | |
2961 | bool flush_work(struct work_struct *work) | |
2962 | { | |
d6e89786 | 2963 | return __flush_work(work, false); |
6e84d644 | 2964 | } |
606a5020 | 2965 | EXPORT_SYMBOL_GPL(flush_work); |
6e84d644 | 2966 | |
8603e1b3 | 2967 | struct cwt_wait { |
ac6424b9 | 2968 | wait_queue_entry_t wait; |
8603e1b3 TH |
2969 | struct work_struct *work; |
2970 | }; | |
2971 | ||
ac6424b9 | 2972 | static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key) |
8603e1b3 TH |
2973 | { |
2974 | struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait); | |
2975 | ||
2976 | if (cwait->work != key) | |
2977 | return 0; | |
2978 | return autoremove_wake_function(wait, mode, sync, key); | |
2979 | } | |
2980 | ||
36e227d2 | 2981 | static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) |
1f1f642e | 2982 | { |
8603e1b3 | 2983 | static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq); |
bbb68dfa | 2984 | unsigned long flags; |
1f1f642e ON |
2985 | int ret; |
2986 | ||
2987 | do { | |
bbb68dfa TH |
2988 | ret = try_to_grab_pending(work, is_dwork, &flags); |
2989 | /* | |
8603e1b3 TH |
2990 | * If someone else is already canceling, wait for it to |
2991 | * finish. flush_work() doesn't work for PREEMPT_NONE | |
2992 | * because we may get scheduled between @work's completion | |
2993 | * and the other canceling task resuming and clearing | |
2994 | * CANCELING - flush_work() will return false immediately | |
2995 | * as @work is no longer busy, try_to_grab_pending() will | |
2996 | * return -ENOENT as @work is still being canceled and the | |
2997 | * other canceling task won't be able to clear CANCELING as | |
2998 | * we're hogging the CPU. | |
2999 | * | |
3000 | * Let's wait for completion using a waitqueue. As this | |
3001 | * may lead to the thundering herd problem, use a custom | |
3002 | * wake function which matches @work along with exclusive | |
3003 | * wait and wakeup. | |
bbb68dfa | 3004 | */ |
8603e1b3 TH |
3005 | if (unlikely(ret == -ENOENT)) { |
3006 | struct cwt_wait cwait; | |
3007 | ||
3008 | init_wait(&cwait.wait); | |
3009 | cwait.wait.func = cwt_wakefn; | |
3010 | cwait.work = work; | |
3011 | ||
3012 | prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait, | |
3013 | TASK_UNINTERRUPTIBLE); | |
3014 | if (work_is_canceling(work)) | |
3015 | schedule(); | |
3016 | finish_wait(&cancel_waitq, &cwait.wait); | |
3017 | } | |
1f1f642e ON |
3018 | } while (unlikely(ret < 0)); |
3019 | ||
bbb68dfa TH |
3020 | /* tell other tasks trying to grab @work to back off */ |
3021 | mark_work_canceling(work); | |
3022 | local_irq_restore(flags); | |
3023 | ||
3347fa09 TH |
3024 | /* |
3025 | * This allows canceling during early boot. We know that @work | |
3026 | * isn't executing. | |
3027 | */ | |
3028 | if (wq_online) | |
d6e89786 | 3029 | __flush_work(work, true); |
3347fa09 | 3030 | |
7a22ad75 | 3031 | clear_work_data(work); |
8603e1b3 TH |
3032 | |
3033 | /* | |
3034 | * Paired with prepare_to_wait() above so that either | |
3035 | * waitqueue_active() is visible here or !work_is_canceling() is | |
3036 | * visible there. | |
3037 | */ | |
3038 | smp_mb(); | |
3039 | if (waitqueue_active(&cancel_waitq)) | |
3040 | __wake_up(&cancel_waitq, TASK_NORMAL, 1, work); | |
3041 | ||
1f1f642e ON |
3042 | return ret; |
3043 | } | |
3044 | ||
6e84d644 | 3045 | /** |
401a8d04 TH |
3046 | * cancel_work_sync - cancel a work and wait for it to finish |
3047 | * @work: the work to cancel | |
6e84d644 | 3048 | * |
401a8d04 TH |
3049 | * Cancel @work and wait for its execution to finish. This function |
3050 | * can be used even if the work re-queues itself or migrates to | |
3051 | * another workqueue. On return from this function, @work is | |
3052 | * guaranteed to be not pending or executing on any CPU. | |
1f1f642e | 3053 | * |
401a8d04 TH |
3054 | * cancel_work_sync(&delayed_work->work) must not be used for |
3055 | * delayed_work's. Use cancel_delayed_work_sync() instead. | |
6e84d644 | 3056 | * |
401a8d04 | 3057 | * The caller must ensure that the workqueue on which @work was last |
6e84d644 | 3058 | * queued can't be destroyed before this function returns. |
401a8d04 | 3059 | * |
d185af30 | 3060 | * Return: |
401a8d04 | 3061 | * %true if @work was pending, %false otherwise. |
6e84d644 | 3062 | */ |
401a8d04 | 3063 | bool cancel_work_sync(struct work_struct *work) |
6e84d644 | 3064 | { |
36e227d2 | 3065 | return __cancel_work_timer(work, false); |
b89deed3 | 3066 | } |
28e53bdd | 3067 | EXPORT_SYMBOL_GPL(cancel_work_sync); |
b89deed3 | 3068 | |
6e84d644 | 3069 | /** |
401a8d04 TH |
3070 | * flush_delayed_work - wait for a dwork to finish executing the last queueing |
3071 | * @dwork: the delayed work to flush | |
6e84d644 | 3072 | * |
401a8d04 TH |
3073 | * Delayed timer is cancelled and the pending work is queued for |
3074 | * immediate execution. Like flush_work(), this function only | |
3075 | * considers the last queueing instance of @dwork. | |
1f1f642e | 3076 | * |
d185af30 | 3077 | * Return: |
401a8d04 TH |
3078 | * %true if flush_work() waited for the work to finish execution, |
3079 | * %false if it was already idle. | |
6e84d644 | 3080 | */ |
401a8d04 TH |
3081 | bool flush_delayed_work(struct delayed_work *dwork) |
3082 | { | |
8930caba | 3083 | local_irq_disable(); |
401a8d04 | 3084 | if (del_timer_sync(&dwork->timer)) |
60c057bc | 3085 | __queue_work(dwork->cpu, dwork->wq, &dwork->work); |
8930caba | 3086 | local_irq_enable(); |
401a8d04 TH |
3087 | return flush_work(&dwork->work); |
3088 | } | |
3089 | EXPORT_SYMBOL(flush_delayed_work); | |
3090 | ||
05f0fe6b TH |
3091 | /** |
3092 | * flush_rcu_work - wait for a rwork to finish executing the last queueing | |
3093 | * @rwork: the rcu work to flush | |
3094 | * | |
3095 | * Return: | |
3096 | * %true if flush_rcu_work() waited for the work to finish execution, | |
3097 | * %false if it was already idle. | |
3098 | */ | |
3099 | bool flush_rcu_work(struct rcu_work *rwork) | |
3100 | { | |
3101 | if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) { | |
3102 | rcu_barrier(); | |
3103 | flush_work(&rwork->work); | |
3104 | return true; | |
3105 | } else { | |
3106 | return flush_work(&rwork->work); | |
3107 | } | |
3108 | } | |
3109 | EXPORT_SYMBOL(flush_rcu_work); | |
3110 | ||
f72b8792 JA |
3111 | static bool __cancel_work(struct work_struct *work, bool is_dwork) |
3112 | { | |
3113 | unsigned long flags; | |
3114 | int ret; | |
3115 | ||
3116 | do { | |
3117 | ret = try_to_grab_pending(work, is_dwork, &flags); | |
3118 | } while (unlikely(ret == -EAGAIN)); | |
3119 | ||
3120 | if (unlikely(ret < 0)) | |
3121 | return false; | |
3122 | ||
3123 | set_work_pool_and_clear_pending(work, get_work_pool_id(work)); | |
3124 | local_irq_restore(flags); | |
3125 | return ret; | |
3126 | } | |
3127 | ||
09383498 | 3128 | /** |
57b30ae7 TH |
3129 | * cancel_delayed_work - cancel a delayed work |
3130 | * @dwork: delayed_work to cancel | |
09383498 | 3131 | * |
d185af30 YB |
3132 | * Kill off a pending delayed_work. |
3133 | * | |
3134 | * Return: %true if @dwork was pending and canceled; %false if it wasn't | |
3135 | * pending. | |
3136 | * | |
3137 | * Note: | |
3138 | * The work callback function may still be running on return, unless | |
3139 | * it returns %true and the work doesn't re-arm itself. Explicitly flush or | |
3140 | * use cancel_delayed_work_sync() to wait on it. | |
09383498 | 3141 | * |
57b30ae7 | 3142 | * This function is safe to call from any context including IRQ handler. |
09383498 | 3143 | */ |
57b30ae7 | 3144 | bool cancel_delayed_work(struct delayed_work *dwork) |
09383498 | 3145 | { |
f72b8792 | 3146 | return __cancel_work(&dwork->work, true); |
09383498 | 3147 | } |
57b30ae7 | 3148 | EXPORT_SYMBOL(cancel_delayed_work); |
09383498 | 3149 | |
401a8d04 TH |
3150 | /** |
3151 | * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish | |
3152 | * @dwork: the delayed work cancel | |
3153 | * | |
3154 | * This is cancel_work_sync() for delayed works. | |
3155 | * | |
d185af30 | 3156 | * Return: |
401a8d04 TH |
3157 | * %true if @dwork was pending, %false otherwise. |
3158 | */ | |
3159 | bool cancel_delayed_work_sync(struct delayed_work *dwork) | |
6e84d644 | 3160 | { |
36e227d2 | 3161 | return __cancel_work_timer(&dwork->work, true); |
6e84d644 | 3162 | } |
f5a421a4 | 3163 | EXPORT_SYMBOL(cancel_delayed_work_sync); |
1da177e4 | 3164 | |
b6136773 | 3165 | /** |
31ddd871 | 3166 | * schedule_on_each_cpu - execute a function synchronously on each online CPU |
b6136773 | 3167 | * @func: the function to call |
b6136773 | 3168 | * |
31ddd871 TH |
3169 | * schedule_on_each_cpu() executes @func on each online CPU using the |
3170 | * system workqueue and blocks until all CPUs have completed. | |
b6136773 | 3171 | * schedule_on_each_cpu() is very slow. |
31ddd871 | 3172 | * |
d185af30 | 3173 | * Return: |
31ddd871 | 3174 | * 0 on success, -errno on failure. |
b6136773 | 3175 | */ |
65f27f38 | 3176 | int schedule_on_each_cpu(work_func_t func) |
15316ba8 CL |
3177 | { |
3178 | int cpu; | |
38f51568 | 3179 | struct work_struct __percpu *works; |
15316ba8 | 3180 | |
b6136773 AM |
3181 | works = alloc_percpu(struct work_struct); |
3182 | if (!works) | |
15316ba8 | 3183 | return -ENOMEM; |
b6136773 | 3184 | |
93981800 TH |
3185 | get_online_cpus(); |
3186 | ||
15316ba8 | 3187 | for_each_online_cpu(cpu) { |
9bfb1839 IM |
3188 | struct work_struct *work = per_cpu_ptr(works, cpu); |
3189 | ||
3190 | INIT_WORK(work, func); | |
b71ab8c2 | 3191 | schedule_work_on(cpu, work); |
65a64464 | 3192 | } |
93981800 TH |
3193 | |
3194 | for_each_online_cpu(cpu) | |
3195 | flush_work(per_cpu_ptr(works, cpu)); | |
3196 | ||
95402b38 | 3197 | put_online_cpus(); |
b6136773 | 3198 | free_percpu(works); |
15316ba8 CL |
3199 | return 0; |
3200 | } | |
3201 | ||
1fa44eca JB |
3202 | /** |
3203 | * execute_in_process_context - reliably execute the routine with user context | |
3204 | * @fn: the function to execute | |
1fa44eca JB |
3205 | * @ew: guaranteed storage for the execute work structure (must |
3206 | * be available when the work executes) | |
3207 | * | |
3208 | * Executes the function immediately if process context is available, | |
3209 | * otherwise schedules the function for delayed execution. | |
3210 | * | |
d185af30 | 3211 | * Return: 0 - function was executed |
1fa44eca JB |
3212 | * 1 - function was scheduled for execution |
3213 | */ | |
65f27f38 | 3214 | int execute_in_process_context(work_func_t fn, struct execute_work *ew) |
1fa44eca JB |
3215 | { |
3216 | if (!in_interrupt()) { | |
65f27f38 | 3217 | fn(&ew->work); |
1fa44eca JB |
3218 | return 0; |
3219 | } | |
3220 | ||
65f27f38 | 3221 | INIT_WORK(&ew->work, fn); |
1fa44eca JB |
3222 | schedule_work(&ew->work); |
3223 | ||
3224 | return 1; | |
3225 | } | |
3226 | EXPORT_SYMBOL_GPL(execute_in_process_context); | |
3227 | ||
6ba94429 FW |
3228 | /** |
3229 | * free_workqueue_attrs - free a workqueue_attrs | |
3230 | * @attrs: workqueue_attrs to free | |
226223ab | 3231 | * |
6ba94429 | 3232 | * Undo alloc_workqueue_attrs(). |
226223ab | 3233 | */ |
6ba94429 | 3234 | void free_workqueue_attrs(struct workqueue_attrs *attrs) |
226223ab | 3235 | { |
6ba94429 FW |
3236 | if (attrs) { |
3237 | free_cpumask_var(attrs->cpumask); | |
3238 | kfree(attrs); | |
3239 | } | |
226223ab TH |
3240 | } |
3241 | ||
6ba94429 FW |
3242 | /** |
3243 | * alloc_workqueue_attrs - allocate a workqueue_attrs | |
3244 | * @gfp_mask: allocation mask to use | |
3245 | * | |
3246 | * Allocate a new workqueue_attrs, initialize with default settings and | |
3247 | * return it. | |
3248 | * | |
3249 | * Return: The allocated new workqueue_attr on success. %NULL on failure. | |
3250 | */ | |
3251 | struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask) | |
226223ab | 3252 | { |
6ba94429 | 3253 | struct workqueue_attrs *attrs; |
226223ab | 3254 | |
6ba94429 FW |
3255 | attrs = kzalloc(sizeof(*attrs), gfp_mask); |
3256 | if (!attrs) | |
3257 | goto fail; | |
3258 | if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask)) | |
3259 | goto fail; | |
3260 | ||
3261 | cpumask_copy(attrs->cpumask, cpu_possible_mask); | |
3262 | return attrs; | |
3263 | fail: | |
3264 | free_workqueue_attrs(attrs); | |
3265 | return NULL; | |
226223ab TH |
3266 | } |
3267 | ||
6ba94429 FW |
3268 | static void copy_workqueue_attrs(struct workqueue_attrs *to, |
3269 | const struct workqueue_attrs *from) | |
226223ab | 3270 | { |
6ba94429 FW |
3271 | to->nice = from->nice; |
3272 | cpumask_copy(to->cpumask, from->cpumask); | |
3273 | /* | |
3274 | * Unlike hash and equality test, this function doesn't ignore | |
3275 | * ->no_numa as it is used for both pool and wq attrs. Instead, | |
3276 | * get_unbound_pool() explicitly clears ->no_numa after copying. | |
3277 | */ | |
3278 | to->no_numa = from->no_numa; | |
226223ab TH |
3279 | } |
3280 | ||
6ba94429 FW |
3281 | /* hash value of the content of @attr */ |
3282 | static u32 wqattrs_hash(const struct workqueue_attrs *attrs) | |
226223ab | 3283 | { |
6ba94429 | 3284 | u32 hash = 0; |
226223ab | 3285 | |
6ba94429 FW |
3286 | hash = jhash_1word(attrs->nice, hash); |
3287 | hash = jhash(cpumask_bits(attrs->cpumask), | |
3288 | BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); | |
3289 | return hash; | |
226223ab | 3290 | } |
226223ab | 3291 | |
6ba94429 FW |
3292 | /* content equality test */ |
3293 | static bool wqattrs_equal(const struct workqueue_attrs *a, | |
3294 | const struct workqueue_attrs *b) | |
226223ab | 3295 | { |
6ba94429 FW |
3296 | if (a->nice != b->nice) |
3297 | return false; | |
3298 | if (!cpumask_equal(a->cpumask, b->cpumask)) | |
3299 | return false; | |
3300 | return true; | |
226223ab TH |
3301 | } |
3302 | ||
6ba94429 FW |
3303 | /** |
3304 | * init_worker_pool - initialize a newly zalloc'd worker_pool | |
3305 | * @pool: worker_pool to initialize | |
3306 | * | |
402dd89d | 3307 | * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs. |
6ba94429 FW |
3308 | * |
3309 | * Return: 0 on success, -errno on failure. Even on failure, all fields | |
3310 | * inside @pool proper are initialized and put_unbound_pool() can be called | |
3311 | * on @pool safely to release it. | |
3312 | */ | |
3313 | static int init_worker_pool(struct worker_pool *pool) | |
226223ab | 3314 | { |
6ba94429 FW |
3315 | spin_lock_init(&pool->lock); |
3316 | pool->id = -1; | |
3317 | pool->cpu = -1; | |
3318 | pool->node = NUMA_NO_NODE; | |
3319 | pool->flags |= POOL_DISASSOCIATED; | |
82607adc | 3320 | pool->watchdog_ts = jiffies; |
6ba94429 FW |
3321 | INIT_LIST_HEAD(&pool->worklist); |
3322 | INIT_LIST_HEAD(&pool->idle_list); | |
3323 | hash_init(pool->busy_hash); | |
226223ab | 3324 | |
32a6c723 | 3325 | timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE); |
226223ab | 3326 | |
32a6c723 | 3327 | timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0); |
226223ab | 3328 | |
6ba94429 | 3329 | INIT_LIST_HEAD(&pool->workers); |
226223ab | 3330 | |
6ba94429 FW |
3331 | ida_init(&pool->worker_ida); |
3332 | INIT_HLIST_NODE(&pool->hash_node); | |
3333 | pool->refcnt = 1; | |
226223ab | 3334 | |
6ba94429 FW |
3335 | /* shouldn't fail above this point */ |
3336 | pool->attrs = alloc_workqueue_attrs(GFP_KERNEL); | |
3337 | if (!pool->attrs) | |
3338 | return -ENOMEM; | |
3339 | return 0; | |
226223ab TH |
3340 | } |
3341 | ||
669de8bd BVA |
3342 | #ifdef CONFIG_LOCKDEP |
3343 | static void wq_init_lockdep(struct workqueue_struct *wq) | |
3344 | { | |
3345 | char *lock_name; | |
3346 | ||
3347 | lockdep_register_key(&wq->key); | |
3348 | lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name); | |
3349 | if (!lock_name) | |
3350 | lock_name = wq->name; | |
3351 | lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0); | |
3352 | } | |
3353 | ||
3354 | static void wq_unregister_lockdep(struct workqueue_struct *wq) | |
3355 | { | |
3356 | lockdep_unregister_key(&wq->key); | |
3357 | } | |
3358 | ||
3359 | static void wq_free_lockdep(struct workqueue_struct *wq) | |
3360 | { | |
3361 | if (wq->lock_name != wq->name) | |
3362 | kfree(wq->lock_name); | |
3363 | } | |
3364 | #else | |
3365 | static void wq_init_lockdep(struct workqueue_struct *wq) | |
3366 | { | |
3367 | } | |
3368 | ||
3369 | static void wq_unregister_lockdep(struct workqueue_struct *wq) | |
3370 | { | |
3371 | } | |
3372 | ||
3373 | static void wq_free_lockdep(struct workqueue_struct *wq) | |
3374 | { | |
3375 | } | |
3376 | #endif | |
3377 | ||
6ba94429 | 3378 | static void rcu_free_wq(struct rcu_head *rcu) |
226223ab | 3379 | { |
6ba94429 FW |
3380 | struct workqueue_struct *wq = |
3381 | container_of(rcu, struct workqueue_struct, rcu); | |
226223ab | 3382 | |
669de8bd BVA |
3383 | wq_free_lockdep(wq); |
3384 | ||
6ba94429 FW |
3385 | if (!(wq->flags & WQ_UNBOUND)) |
3386 | free_percpu(wq->cpu_pwqs); | |
226223ab | 3387 | else |
6ba94429 | 3388 | free_workqueue_attrs(wq->unbound_attrs); |
226223ab | 3389 | |
6ba94429 FW |
3390 | kfree(wq->rescuer); |
3391 | kfree(wq); | |
226223ab TH |
3392 | } |
3393 | ||
6ba94429 | 3394 | static void rcu_free_pool(struct rcu_head *rcu) |
226223ab | 3395 | { |
6ba94429 | 3396 | struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); |
226223ab | 3397 | |
6ba94429 FW |
3398 | ida_destroy(&pool->worker_ida); |
3399 | free_workqueue_attrs(pool->attrs); | |
3400 | kfree(pool); | |
226223ab TH |
3401 | } |
3402 | ||
6ba94429 FW |
3403 | /** |
3404 | * put_unbound_pool - put a worker_pool | |
3405 | * @pool: worker_pool to put | |
3406 | * | |
3407 | * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU | |
3408 | * safe manner. get_unbound_pool() calls this function on its failure path | |
3409 | * and this function should be able to release pools which went through, | |
3410 | * successfully or not, init_worker_pool(). | |
3411 | * | |
3412 | * Should be called with wq_pool_mutex held. | |
3413 | */ | |
3414 | static void put_unbound_pool(struct worker_pool *pool) | |
226223ab | 3415 | { |
6ba94429 FW |
3416 | DECLARE_COMPLETION_ONSTACK(detach_completion); |
3417 | struct worker *worker; | |
226223ab | 3418 | |
6ba94429 | 3419 | lockdep_assert_held(&wq_pool_mutex); |
226223ab | 3420 | |
6ba94429 FW |
3421 | if (--pool->refcnt) |
3422 | return; | |
226223ab | 3423 | |
6ba94429 FW |
3424 | /* sanity checks */ |
3425 | if (WARN_ON(!(pool->cpu < 0)) || | |
3426 | WARN_ON(!list_empty(&pool->worklist))) | |
3427 | return; | |
226223ab | 3428 | |
6ba94429 FW |
3429 | /* release id and unhash */ |
3430 | if (pool->id >= 0) | |
3431 | idr_remove(&worker_pool_idr, pool->id); | |
3432 | hash_del(&pool->hash_node); | |
d55262c4 | 3433 | |
6ba94429 | 3434 | /* |
692b4825 TH |
3435 | * Become the manager and destroy all workers. This prevents |
3436 | * @pool's workers from blocking on attach_mutex. We're the last | |
3437 | * manager and @pool gets freed with the flag set. | |
6ba94429 | 3438 | */ |
6ba94429 | 3439 | spin_lock_irq(&pool->lock); |
692b4825 TH |
3440 | wait_event_lock_irq(wq_manager_wait, |
3441 | !(pool->flags & POOL_MANAGER_ACTIVE), pool->lock); | |
3442 | pool->flags |= POOL_MANAGER_ACTIVE; | |
3443 | ||
6ba94429 FW |
3444 | while ((worker = first_idle_worker(pool))) |
3445 | destroy_worker(worker); | |
3446 | WARN_ON(pool->nr_workers || pool->nr_idle); | |
3447 | spin_unlock_irq(&pool->lock); | |
d55262c4 | 3448 | |
1258fae7 | 3449 | mutex_lock(&wq_pool_attach_mutex); |
6ba94429 FW |
3450 | if (!list_empty(&pool->workers)) |
3451 | pool->detach_completion = &detach_completion; | |
1258fae7 | 3452 | mutex_unlock(&wq_pool_attach_mutex); |
226223ab | 3453 | |
6ba94429 FW |
3454 | if (pool->detach_completion) |
3455 | wait_for_completion(pool->detach_completion); | |
226223ab | 3456 | |
6ba94429 FW |
3457 | /* shut down the timers */ |
3458 | del_timer_sync(&pool->idle_timer); | |
3459 | del_timer_sync(&pool->mayday_timer); | |
226223ab | 3460 | |
6ba94429 | 3461 | /* sched-RCU protected to allow dereferences from get_work_pool() */ |
25b00775 | 3462 | call_rcu(&pool->rcu, rcu_free_pool); |
226223ab TH |
3463 | } |
3464 | ||
3465 | /** | |
6ba94429 FW |
3466 | * get_unbound_pool - get a worker_pool with the specified attributes |
3467 | * @attrs: the attributes of the worker_pool to get | |
226223ab | 3468 | * |
6ba94429 FW |
3469 | * Obtain a worker_pool which has the same attributes as @attrs, bump the |
3470 | * reference count and return it. If there already is a matching | |
3471 | * worker_pool, it will be used; otherwise, this function attempts to | |
3472 | * create a new one. | |
226223ab | 3473 | * |
6ba94429 | 3474 | * Should be called with wq_pool_mutex held. |
226223ab | 3475 | * |
6ba94429 FW |
3476 | * Return: On success, a worker_pool with the same attributes as @attrs. |
3477 | * On failure, %NULL. | |
226223ab | 3478 | */ |
6ba94429 | 3479 | static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) |
226223ab | 3480 | { |
6ba94429 FW |
3481 | u32 hash = wqattrs_hash(attrs); |
3482 | struct worker_pool *pool; | |
3483 | int node; | |
e2273584 | 3484 | int target_node = NUMA_NO_NODE; |
226223ab | 3485 | |
6ba94429 | 3486 | lockdep_assert_held(&wq_pool_mutex); |
226223ab | 3487 | |
6ba94429 FW |
3488 | /* do we already have a matching pool? */ |
3489 | hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { | |
3490 | if (wqattrs_equal(pool->attrs, attrs)) { | |
3491 | pool->refcnt++; | |
3492 | return pool; | |
3493 | } | |
3494 | } | |
226223ab | 3495 | |
e2273584 XP |
3496 | /* if cpumask is contained inside a NUMA node, we belong to that node */ |
3497 | if (wq_numa_enabled) { | |
3498 | for_each_node(node) { | |
3499 | if (cpumask_subset(attrs->cpumask, | |
3500 | wq_numa_possible_cpumask[node])) { | |
3501 | target_node = node; | |
3502 | break; | |
3503 | } | |
3504 | } | |
3505 | } | |
3506 | ||
6ba94429 | 3507 | /* nope, create a new one */ |
e2273584 | 3508 | pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node); |
6ba94429 FW |
3509 | if (!pool || init_worker_pool(pool) < 0) |
3510 | goto fail; | |
3511 | ||
3512 | lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */ | |
3513 | copy_workqueue_attrs(pool->attrs, attrs); | |
e2273584 | 3514 | pool->node = target_node; |
226223ab TH |
3515 | |
3516 | /* | |
6ba94429 FW |
3517 | * no_numa isn't a worker_pool attribute, always clear it. See |
3518 | * 'struct workqueue_attrs' comments for detail. | |
226223ab | 3519 | */ |
6ba94429 | 3520 | pool->attrs->no_numa = false; |
226223ab | 3521 | |
6ba94429 FW |
3522 | if (worker_pool_assign_id(pool) < 0) |
3523 | goto fail; | |
226223ab | 3524 | |
6ba94429 | 3525 | /* create and start the initial worker */ |
3347fa09 | 3526 | if (wq_online && !create_worker(pool)) |
6ba94429 | 3527 | goto fail; |
226223ab | 3528 | |
6ba94429 FW |
3529 | /* install */ |
3530 | hash_add(unbound_pool_hash, &pool->hash_node, hash); | |
226223ab | 3531 | |
6ba94429 FW |
3532 | return pool; |
3533 | fail: | |
3534 | if (pool) | |
3535 | put_unbound_pool(pool); | |
3536 | return NULL; | |
226223ab | 3537 | } |
226223ab | 3538 | |
6ba94429 | 3539 | static void rcu_free_pwq(struct rcu_head *rcu) |
7a4e344c | 3540 | { |
6ba94429 FW |
3541 | kmem_cache_free(pwq_cache, |
3542 | container_of(rcu, struct pool_workqueue, rcu)); | |
7a4e344c TH |
3543 | } |
3544 | ||
6ba94429 FW |
3545 | /* |
3546 | * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt | |
3547 | * and needs to be destroyed. | |
7a4e344c | 3548 | */ |
6ba94429 | 3549 | static void pwq_unbound_release_workfn(struct work_struct *work) |
7a4e344c | 3550 | { |
6ba94429 FW |
3551 | struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, |
3552 | unbound_release_work); | |
3553 | struct workqueue_struct *wq = pwq->wq; | |
3554 | struct worker_pool *pool = pwq->pool; | |
3555 | bool is_last; | |
7a4e344c | 3556 | |
6ba94429 FW |
3557 | if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND))) |
3558 | return; | |
7a4e344c | 3559 | |
6ba94429 FW |
3560 | mutex_lock(&wq->mutex); |
3561 | list_del_rcu(&pwq->pwqs_node); | |
3562 | is_last = list_empty(&wq->pwqs); | |
3563 | mutex_unlock(&wq->mutex); | |
3564 | ||
3565 | mutex_lock(&wq_pool_mutex); | |
3566 | put_unbound_pool(pool); | |
3567 | mutex_unlock(&wq_pool_mutex); | |
3568 | ||
25b00775 | 3569 | call_rcu(&pwq->rcu, rcu_free_pwq); |
7a4e344c | 3570 | |
2865a8fb | 3571 | /* |
6ba94429 FW |
3572 | * If we're the last pwq going away, @wq is already dead and no one |
3573 | * is gonna access it anymore. Schedule RCU free. | |
2865a8fb | 3574 | */ |
669de8bd BVA |
3575 | if (is_last) { |
3576 | wq_unregister_lockdep(wq); | |
25b00775 | 3577 | call_rcu(&wq->rcu, rcu_free_wq); |
669de8bd | 3578 | } |
29c91e99 TH |
3579 | } |
3580 | ||
7a4e344c | 3581 | /** |
6ba94429 FW |
3582 | * pwq_adjust_max_active - update a pwq's max_active to the current setting |
3583 | * @pwq: target pool_workqueue | |
d185af30 | 3584 | * |
6ba94429 FW |
3585 | * If @pwq isn't freezing, set @pwq->max_active to the associated |
3586 | * workqueue's saved_max_active and activate delayed work items | |
3587 | * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. | |
7a4e344c | 3588 | */ |
6ba94429 | 3589 | static void pwq_adjust_max_active(struct pool_workqueue *pwq) |
4e1a1f9a | 3590 | { |
6ba94429 FW |
3591 | struct workqueue_struct *wq = pwq->wq; |
3592 | bool freezable = wq->flags & WQ_FREEZABLE; | |
3347fa09 | 3593 | unsigned long flags; |
4e1a1f9a | 3594 | |
6ba94429 FW |
3595 | /* for @wq->saved_max_active */ |
3596 | lockdep_assert_held(&wq->mutex); | |
4e1a1f9a | 3597 | |
6ba94429 FW |
3598 | /* fast exit for non-freezable wqs */ |
3599 | if (!freezable && pwq->max_active == wq->saved_max_active) | |
3600 | return; | |
7a4e344c | 3601 | |
3347fa09 TH |
3602 | /* this function can be called during early boot w/ irq disabled */ |
3603 | spin_lock_irqsave(&pwq->pool->lock, flags); | |
29c91e99 | 3604 | |
6ba94429 FW |
3605 | /* |
3606 | * During [un]freezing, the caller is responsible for ensuring that | |
3607 | * this function is called at least once after @workqueue_freezing | |
3608 | * is updated and visible. | |
3609 | */ | |
3610 | if (!freezable || !workqueue_freezing) { | |
3611 | pwq->max_active = wq->saved_max_active; | |
4e1a1f9a | 3612 | |
6ba94429 FW |
3613 | while (!list_empty(&pwq->delayed_works) && |
3614 | pwq->nr_active < pwq->max_active) | |
3615 | pwq_activate_first_delayed(pwq); | |
e2dca7ad | 3616 | |
6ba94429 FW |
3617 | /* |
3618 | * Need to kick a worker after thawed or an unbound wq's | |
3619 | * max_active is bumped. It's a slow path. Do it always. | |
3620 | */ | |
3621 | wake_up_worker(pwq->pool); | |
3622 | } else { | |
3623 | pwq->max_active = 0; | |
3624 | } | |
e2dca7ad | 3625 | |
3347fa09 | 3626 | spin_unlock_irqrestore(&pwq->pool->lock, flags); |
e2dca7ad TH |
3627 | } |
3628 | ||
6ba94429 FW |
3629 | /* initialize newly alloced @pwq which is associated with @wq and @pool */ |
3630 | static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, | |
3631 | struct worker_pool *pool) | |
29c91e99 | 3632 | { |
6ba94429 | 3633 | BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); |
29c91e99 | 3634 | |
6ba94429 FW |
3635 | memset(pwq, 0, sizeof(*pwq)); |
3636 | ||
3637 | pwq->pool = pool; | |
3638 | pwq->wq = wq; | |
3639 | pwq->flush_color = -1; | |
3640 | pwq->refcnt = 1; | |
3641 | INIT_LIST_HEAD(&pwq->delayed_works); | |
3642 | INIT_LIST_HEAD(&pwq->pwqs_node); | |
3643 | INIT_LIST_HEAD(&pwq->mayday_node); | |
3644 | INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn); | |
29c91e99 TH |
3645 | } |
3646 | ||
6ba94429 FW |
3647 | /* sync @pwq with the current state of its associated wq and link it */ |
3648 | static void link_pwq(struct pool_workqueue *pwq) | |
29c91e99 | 3649 | { |
6ba94429 | 3650 | struct workqueue_struct *wq = pwq->wq; |
29c91e99 | 3651 | |
6ba94429 | 3652 | lockdep_assert_held(&wq->mutex); |
a892cacc | 3653 | |
6ba94429 FW |
3654 | /* may be called multiple times, ignore if already linked */ |
3655 | if (!list_empty(&pwq->pwqs_node)) | |
29c91e99 | 3656 | return; |
29c91e99 | 3657 | |
6ba94429 FW |
3658 | /* set the matching work_color */ |
3659 | pwq->work_color = wq->work_color; | |
29c91e99 | 3660 | |
6ba94429 FW |
3661 | /* sync max_active to the current setting */ |
3662 | pwq_adjust_max_active(pwq); | |
29c91e99 | 3663 | |
6ba94429 FW |
3664 | /* link in @pwq */ |
3665 | list_add_rcu(&pwq->pwqs_node, &wq->pwqs); | |
3666 | } | |
29c91e99 | 3667 | |
6ba94429 FW |
3668 | /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ |
3669 | static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, | |
3670 | const struct workqueue_attrs *attrs) | |
3671 | { | |
3672 | struct worker_pool *pool; | |
3673 | struct pool_workqueue *pwq; | |
60f5a4bc | 3674 | |
6ba94429 | 3675 | lockdep_assert_held(&wq_pool_mutex); |
60f5a4bc | 3676 | |
6ba94429 FW |
3677 | pool = get_unbound_pool(attrs); |
3678 | if (!pool) | |
3679 | return NULL; | |
60f5a4bc | 3680 | |
6ba94429 FW |
3681 | pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); |
3682 | if (!pwq) { | |
3683 | put_unbound_pool(pool); | |
3684 | return NULL; | |
3685 | } | |
29c91e99 | 3686 | |
6ba94429 FW |
3687 | init_pwq(pwq, wq, pool); |
3688 | return pwq; | |
3689 | } | |
29c91e99 | 3690 | |
29c91e99 | 3691 | /** |
30186c6f | 3692 | * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node |
042f7df1 | 3693 | * @attrs: the wq_attrs of the default pwq of the target workqueue |
6ba94429 FW |
3694 | * @node: the target NUMA node |
3695 | * @cpu_going_down: if >= 0, the CPU to consider as offline | |
3696 | * @cpumask: outarg, the resulting cpumask | |
29c91e99 | 3697 | * |
6ba94429 FW |
3698 | * Calculate the cpumask a workqueue with @attrs should use on @node. If |
3699 | * @cpu_going_down is >= 0, that cpu is considered offline during | |
3700 | * calculation. The result is stored in @cpumask. | |
a892cacc | 3701 | * |
6ba94429 FW |
3702 | * If NUMA affinity is not enabled, @attrs->cpumask is always used. If |
3703 | * enabled and @node has online CPUs requested by @attrs, the returned | |
3704 | * cpumask is the intersection of the possible CPUs of @node and | |
3705 | * @attrs->cpumask. | |
d185af30 | 3706 | * |
6ba94429 FW |
3707 | * The caller is responsible for ensuring that the cpumask of @node stays |
3708 | * stable. | |
3709 | * | |
3710 | * Return: %true if the resulting @cpumask is different from @attrs->cpumask, | |
3711 | * %false if equal. | |
29c91e99 | 3712 | */ |
6ba94429 FW |
3713 | static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node, |
3714 | int cpu_going_down, cpumask_t *cpumask) | |
29c91e99 | 3715 | { |
6ba94429 FW |
3716 | if (!wq_numa_enabled || attrs->no_numa) |
3717 | goto use_dfl; | |
29c91e99 | 3718 | |
6ba94429 FW |
3719 | /* does @node have any online CPUs @attrs wants? */ |
3720 | cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask); | |
3721 | if (cpu_going_down >= 0) | |
3722 | cpumask_clear_cpu(cpu_going_down, cpumask); | |
29c91e99 | 3723 | |
6ba94429 FW |
3724 | if (cpumask_empty(cpumask)) |
3725 | goto use_dfl; | |
4c16bd32 TH |
3726 | |
3727 | /* yeap, return possible CPUs in @node that @attrs wants */ | |
3728 | cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]); | |
1ad0f0a7 MB |
3729 | |
3730 | if (cpumask_empty(cpumask)) { | |
3731 | pr_warn_once("WARNING: workqueue cpumask: online intersect > " | |
3732 | "possible intersect\n"); | |
3733 | return false; | |
3734 | } | |
3735 | ||
4c16bd32 TH |
3736 | return !cpumask_equal(cpumask, attrs->cpumask); |
3737 | ||
3738 | use_dfl: | |
3739 | cpumask_copy(cpumask, attrs->cpumask); | |
3740 | return false; | |
3741 | } | |
3742 | ||
1befcf30 TH |
3743 | /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */ |
3744 | static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq, | |
3745 | int node, | |
3746 | struct pool_workqueue *pwq) | |
3747 | { | |
3748 | struct pool_workqueue *old_pwq; | |
3749 | ||
5b95e1af | 3750 | lockdep_assert_held(&wq_pool_mutex); |
1befcf30 TH |
3751 | lockdep_assert_held(&wq->mutex); |
3752 | ||
3753 | /* link_pwq() can handle duplicate calls */ | |
3754 | link_pwq(pwq); | |
3755 | ||
3756 | old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); | |
3757 | rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq); | |
3758 | return old_pwq; | |
3759 | } | |
3760 | ||
2d5f0764 LJ |
3761 | /* context to store the prepared attrs & pwqs before applying */ |
3762 | struct apply_wqattrs_ctx { | |
3763 | struct workqueue_struct *wq; /* target workqueue */ | |
3764 | struct workqueue_attrs *attrs; /* attrs to apply */ | |
042f7df1 | 3765 | struct list_head list; /* queued for batching commit */ |
2d5f0764 LJ |
3766 | struct pool_workqueue *dfl_pwq; |
3767 | struct pool_workqueue *pwq_tbl[]; | |
3768 | }; | |
3769 | ||
3770 | /* free the resources after success or abort */ | |
3771 | static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx) | |
3772 | { | |
3773 | if (ctx) { | |
3774 | int node; | |
3775 | ||
3776 | for_each_node(node) | |
3777 | put_pwq_unlocked(ctx->pwq_tbl[node]); | |
3778 | put_pwq_unlocked(ctx->dfl_pwq); | |
3779 | ||
3780 | free_workqueue_attrs(ctx->attrs); | |
3781 | ||
3782 | kfree(ctx); | |
3783 | } | |
3784 | } | |
3785 | ||
3786 | /* allocate the attrs and pwqs for later installation */ | |
3787 | static struct apply_wqattrs_ctx * | |
3788 | apply_wqattrs_prepare(struct workqueue_struct *wq, | |
3789 | const struct workqueue_attrs *attrs) | |
9e8cd2f5 | 3790 | { |
2d5f0764 | 3791 | struct apply_wqattrs_ctx *ctx; |
4c16bd32 | 3792 | struct workqueue_attrs *new_attrs, *tmp_attrs; |
2d5f0764 | 3793 | int node; |
9e8cd2f5 | 3794 | |
2d5f0764 | 3795 | lockdep_assert_held(&wq_pool_mutex); |
9e8cd2f5 | 3796 | |
acafe7e3 | 3797 | ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_node_ids), GFP_KERNEL); |
8719dcea | 3798 | |
13e2e556 | 3799 | new_attrs = alloc_workqueue_attrs(GFP_KERNEL); |
4c16bd32 | 3800 | tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL); |
2d5f0764 LJ |
3801 | if (!ctx || !new_attrs || !tmp_attrs) |
3802 | goto out_free; | |
13e2e556 | 3803 | |
042f7df1 LJ |
3804 | /* |
3805 | * Calculate the attrs of the default pwq. | |
3806 | * If the user configured cpumask doesn't overlap with the | |
3807 | * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask. | |
3808 | */ | |
13e2e556 | 3809 | copy_workqueue_attrs(new_attrs, attrs); |
b05a7928 | 3810 | cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask); |
042f7df1 LJ |
3811 | if (unlikely(cpumask_empty(new_attrs->cpumask))) |
3812 | cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask); | |
13e2e556 | 3813 | |
4c16bd32 TH |
3814 | /* |
3815 | * We may create multiple pwqs with differing cpumasks. Make a | |
3816 | * copy of @new_attrs which will be modified and used to obtain | |
3817 | * pools. | |
3818 | */ | |
3819 | copy_workqueue_attrs(tmp_attrs, new_attrs); | |
3820 | ||
4c16bd32 TH |
3821 | /* |
3822 | * If something goes wrong during CPU up/down, we'll fall back to | |
3823 | * the default pwq covering whole @attrs->cpumask. Always create | |
3824 | * it even if we don't use it immediately. | |
3825 | */ | |
2d5f0764 LJ |
3826 | ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs); |
3827 | if (!ctx->dfl_pwq) | |
3828 | goto out_free; | |
4c16bd32 TH |
3829 | |
3830 | for_each_node(node) { | |
042f7df1 | 3831 | if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) { |
2d5f0764 LJ |
3832 | ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs); |
3833 | if (!ctx->pwq_tbl[node]) | |
3834 | goto out_free; | |
4c16bd32 | 3835 | } else { |
2d5f0764 LJ |
3836 | ctx->dfl_pwq->refcnt++; |
3837 | ctx->pwq_tbl[node] = ctx->dfl_pwq; | |
4c16bd32 TH |
3838 | } |
3839 | } | |
3840 | ||
042f7df1 LJ |
3841 | /* save the user configured attrs and sanitize it. */ |
3842 | copy_workqueue_attrs(new_attrs, attrs); | |
3843 | cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); | |
2d5f0764 | 3844 | ctx->attrs = new_attrs; |
042f7df1 | 3845 | |
2d5f0764 LJ |
3846 | ctx->wq = wq; |
3847 | free_workqueue_attrs(tmp_attrs); | |
3848 | return ctx; | |
3849 | ||
3850 | out_free: | |
3851 | free_workqueue_attrs(tmp_attrs); | |
3852 | free_workqueue_attrs(new_attrs); | |
3853 | apply_wqattrs_cleanup(ctx); | |
3854 | return NULL; | |
3855 | } | |
3856 | ||
3857 | /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */ | |
3858 | static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx) | |
3859 | { | |
3860 | int node; | |
9e8cd2f5 | 3861 | |
4c16bd32 | 3862 | /* all pwqs have been created successfully, let's install'em */ |
2d5f0764 | 3863 | mutex_lock(&ctx->wq->mutex); |
a892cacc | 3864 | |
2d5f0764 | 3865 | copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs); |
4c16bd32 TH |
3866 | |
3867 | /* save the previous pwq and install the new one */ | |
f147f29e | 3868 | for_each_node(node) |
2d5f0764 LJ |
3869 | ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node, |
3870 | ctx->pwq_tbl[node]); | |
4c16bd32 TH |
3871 | |
3872 | /* @dfl_pwq might not have been used, ensure it's linked */ | |
2d5f0764 LJ |
3873 | link_pwq(ctx->dfl_pwq); |
3874 | swap(ctx->wq->dfl_pwq, ctx->dfl_pwq); | |
f147f29e | 3875 | |
2d5f0764 LJ |
3876 | mutex_unlock(&ctx->wq->mutex); |
3877 | } | |
9e8cd2f5 | 3878 | |
a0111cf6 LJ |
3879 | static void apply_wqattrs_lock(void) |
3880 | { | |
3881 | /* CPUs should stay stable across pwq creations and installations */ | |
3882 | get_online_cpus(); | |
3883 | mutex_lock(&wq_pool_mutex); | |
3884 | } | |
3885 | ||
3886 | static void apply_wqattrs_unlock(void) | |
3887 | { | |
3888 | mutex_unlock(&wq_pool_mutex); | |
3889 | put_online_cpus(); | |
3890 | } | |
3891 | ||
3892 | static int apply_workqueue_attrs_locked(struct workqueue_struct *wq, | |
3893 | const struct workqueue_attrs *attrs) | |
2d5f0764 LJ |
3894 | { |
3895 | struct apply_wqattrs_ctx *ctx; | |
4c16bd32 | 3896 | |
2d5f0764 LJ |
3897 | /* only unbound workqueues can change attributes */ |
3898 | if (WARN_ON(!(wq->flags & WQ_UNBOUND))) | |
3899 | return -EINVAL; | |
13e2e556 | 3900 | |
2d5f0764 | 3901 | /* creating multiple pwqs breaks ordering guarantee */ |
0a94efb5 TH |
3902 | if (!list_empty(&wq->pwqs)) { |
3903 | if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) | |
3904 | return -EINVAL; | |
3905 | ||
3906 | wq->flags &= ~__WQ_ORDERED; | |
3907 | } | |
2d5f0764 | 3908 | |
2d5f0764 | 3909 | ctx = apply_wqattrs_prepare(wq, attrs); |
6201171e | 3910 | if (!ctx) |
3911 | return -ENOMEM; | |
2d5f0764 LJ |
3912 | |
3913 | /* the ctx has been prepared successfully, let's commit it */ | |
6201171e | 3914 | apply_wqattrs_commit(ctx); |
2d5f0764 LJ |
3915 | apply_wqattrs_cleanup(ctx); |
3916 | ||
6201171e | 3917 | return 0; |
9e8cd2f5 TH |
3918 | } |
3919 | ||
a0111cf6 LJ |
3920 | /** |
3921 | * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue | |
3922 | * @wq: the target workqueue | |
3923 | * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() | |
3924 | * | |
3925 | * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA | |
3926 | * machines, this function maps a separate pwq to each NUMA node with | |
3927 | * possibles CPUs in @attrs->cpumask so that work items are affine to the | |
3928 | * NUMA node it was issued on. Older pwqs are released as in-flight work | |
3929 | * items finish. Note that a work item which repeatedly requeues itself | |
3930 | * back-to-back will stay on its current pwq. | |
3931 | * | |
3932 | * Performs GFP_KERNEL allocations. | |
3933 | * | |
3934 | * Return: 0 on success and -errno on failure. | |
3935 | */ | |
3936 | int apply_workqueue_attrs(struct workqueue_struct *wq, | |
3937 | const struct workqueue_attrs *attrs) | |
3938 | { | |
3939 | int ret; | |
3940 | ||
3941 | apply_wqattrs_lock(); | |
3942 | ret = apply_workqueue_attrs_locked(wq, attrs); | |
3943 | apply_wqattrs_unlock(); | |
3944 | ||
3945 | return ret; | |
3946 | } | |
6106c0f8 | 3947 | EXPORT_SYMBOL_GPL(apply_workqueue_attrs); |
a0111cf6 | 3948 | |
4c16bd32 TH |
3949 | /** |
3950 | * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug | |
3951 | * @wq: the target workqueue | |
3952 | * @cpu: the CPU coming up or going down | |
3953 | * @online: whether @cpu is coming up or going down | |
3954 | * | |
3955 | * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and | |
3956 | * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of | |
3957 | * @wq accordingly. | |
3958 | * | |
3959 | * If NUMA affinity can't be adjusted due to memory allocation failure, it | |
3960 | * falls back to @wq->dfl_pwq which may not be optimal but is always | |
3961 | * correct. | |
3962 | * | |
3963 | * Note that when the last allowed CPU of a NUMA node goes offline for a | |
3964 | * workqueue with a cpumask spanning multiple nodes, the workers which were | |
3965 | * already executing the work items for the workqueue will lose their CPU | |
3966 | * affinity and may execute on any CPU. This is similar to how per-cpu | |
3967 | * workqueues behave on CPU_DOWN. If a workqueue user wants strict | |
3968 | * affinity, it's the user's responsibility to flush the work item from | |
3969 | * CPU_DOWN_PREPARE. | |
3970 | */ | |
3971 | static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu, | |
3972 | bool online) | |
3973 | { | |
3974 | int node = cpu_to_node(cpu); | |
3975 | int cpu_off = online ? -1 : cpu; | |
3976 | struct pool_workqueue *old_pwq = NULL, *pwq; | |
3977 | struct workqueue_attrs *target_attrs; | |
3978 | cpumask_t *cpumask; | |
3979 | ||
3980 | lockdep_assert_held(&wq_pool_mutex); | |
3981 | ||
f7142ed4 LJ |
3982 | if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) || |
3983 | wq->unbound_attrs->no_numa) | |
4c16bd32 TH |
3984 | return; |
3985 | ||
3986 | /* | |
3987 | * We don't wanna alloc/free wq_attrs for each wq for each CPU. | |
3988 | * Let's use a preallocated one. The following buf is protected by | |
3989 | * CPU hotplug exclusion. | |
3990 | */ | |
3991 | target_attrs = wq_update_unbound_numa_attrs_buf; | |
3992 | cpumask = target_attrs->cpumask; | |
3993 | ||
4c16bd32 TH |
3994 | copy_workqueue_attrs(target_attrs, wq->unbound_attrs); |
3995 | pwq = unbound_pwq_by_node(wq, node); | |
3996 | ||
3997 | /* | |
3998 | * Let's determine what needs to be done. If the target cpumask is | |
042f7df1 LJ |
3999 | * different from the default pwq's, we need to compare it to @pwq's |
4000 | * and create a new one if they don't match. If the target cpumask | |
4001 | * equals the default pwq's, the default pwq should be used. | |
4c16bd32 | 4002 | */ |
042f7df1 | 4003 | if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) { |
4c16bd32 | 4004 | if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask)) |
f7142ed4 | 4005 | return; |
4c16bd32 | 4006 | } else { |
534a3fbb | 4007 | goto use_dfl_pwq; |
4c16bd32 TH |
4008 | } |
4009 | ||
4c16bd32 TH |
4010 | /* create a new pwq */ |
4011 | pwq = alloc_unbound_pwq(wq, target_attrs); | |
4012 | if (!pwq) { | |
2d916033 FF |
4013 | pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n", |
4014 | wq->name); | |
77f300b1 | 4015 | goto use_dfl_pwq; |
4c16bd32 TH |
4016 | } |
4017 | ||
f7142ed4 | 4018 | /* Install the new pwq. */ |
4c16bd32 TH |
4019 | mutex_lock(&wq->mutex); |
4020 | old_pwq = numa_pwq_tbl_install(wq, node, pwq); | |
4021 | goto out_unlock; | |
4022 | ||
4023 | use_dfl_pwq: | |
f7142ed4 | 4024 | mutex_lock(&wq->mutex); |
4c16bd32 TH |
4025 | spin_lock_irq(&wq->dfl_pwq->pool->lock); |
4026 | get_pwq(wq->dfl_pwq); | |
4027 | spin_unlock_irq(&wq->dfl_pwq->pool->lock); | |
4028 | old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq); | |
4029 | out_unlock: | |
4030 | mutex_unlock(&wq->mutex); | |
4031 | put_pwq_unlocked(old_pwq); | |
4032 | } | |
4033 | ||
30cdf249 | 4034 | static int alloc_and_link_pwqs(struct workqueue_struct *wq) |
0f900049 | 4035 | { |
49e3cf44 | 4036 | bool highpri = wq->flags & WQ_HIGHPRI; |
8a2b7538 | 4037 | int cpu, ret; |
30cdf249 TH |
4038 | |
4039 | if (!(wq->flags & WQ_UNBOUND)) { | |
420c0ddb TH |
4040 | wq->cpu_pwqs = alloc_percpu(struct pool_workqueue); |
4041 | if (!wq->cpu_pwqs) | |
30cdf249 TH |
4042 | return -ENOMEM; |
4043 | ||
4044 | for_each_possible_cpu(cpu) { | |
7fb98ea7 TH |
4045 | struct pool_workqueue *pwq = |
4046 | per_cpu_ptr(wq->cpu_pwqs, cpu); | |
7a62c2c8 | 4047 | struct worker_pool *cpu_pools = |
f02ae73a | 4048 | per_cpu(cpu_worker_pools, cpu); |
f3421797 | 4049 | |
f147f29e TH |
4050 | init_pwq(pwq, wq, &cpu_pools[highpri]); |
4051 | ||
4052 | mutex_lock(&wq->mutex); | |
1befcf30 | 4053 | link_pwq(pwq); |
f147f29e | 4054 | mutex_unlock(&wq->mutex); |
30cdf249 | 4055 | } |
9e8cd2f5 | 4056 | return 0; |
8a2b7538 TH |
4057 | } else if (wq->flags & __WQ_ORDERED) { |
4058 | ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]); | |
4059 | /* there should only be single pwq for ordering guarantee */ | |
4060 | WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node || | |
4061 | wq->pwqs.prev != &wq->dfl_pwq->pwqs_node), | |
4062 | "ordering guarantee broken for workqueue %s\n", wq->name); | |
4063 | return ret; | |
30cdf249 | 4064 | } else { |
9e8cd2f5 | 4065 | return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); |
30cdf249 | 4066 | } |
0f900049 TH |
4067 | } |
4068 | ||
f3421797 TH |
4069 | static int wq_clamp_max_active(int max_active, unsigned int flags, |
4070 | const char *name) | |
b71ab8c2 | 4071 | { |
f3421797 TH |
4072 | int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; |
4073 | ||
4074 | if (max_active < 1 || max_active > lim) | |
044c782c VI |
4075 | pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", |
4076 | max_active, name, 1, lim); | |
b71ab8c2 | 4077 | |
f3421797 | 4078 | return clamp_val(max_active, 1, lim); |
b71ab8c2 TH |
4079 | } |
4080 | ||
983c7515 TH |
4081 | /* |
4082 | * Workqueues which may be used during memory reclaim should have a rescuer | |
4083 | * to guarantee forward progress. | |
4084 | */ | |
4085 | static int init_rescuer(struct workqueue_struct *wq) | |
4086 | { | |
4087 | struct worker *rescuer; | |
4088 | int ret; | |
4089 | ||
4090 | if (!(wq->flags & WQ_MEM_RECLAIM)) | |
4091 | return 0; | |
4092 | ||
4093 | rescuer = alloc_worker(NUMA_NO_NODE); | |
4094 | if (!rescuer) | |
4095 | return -ENOMEM; | |
4096 | ||
4097 | rescuer->rescue_wq = wq; | |
4098 | rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", wq->name); | |
4099 | ret = PTR_ERR_OR_ZERO(rescuer->task); | |
4100 | if (ret) { | |
4101 | kfree(rescuer); | |
4102 | return ret; | |
4103 | } | |
4104 | ||
4105 | wq->rescuer = rescuer; | |
4106 | kthread_bind_mask(rescuer->task, cpu_possible_mask); | |
4107 | wake_up_process(rescuer->task); | |
4108 | ||
4109 | return 0; | |
4110 | } | |
4111 | ||
669de8bd BVA |
4112 | struct workqueue_struct *alloc_workqueue(const char *fmt, |
4113 | unsigned int flags, | |
4114 | int max_active, ...) | |
1da177e4 | 4115 | { |
df2d5ae4 | 4116 | size_t tbl_size = 0; |
ecf6881f | 4117 | va_list args; |
1da177e4 | 4118 | struct workqueue_struct *wq; |
49e3cf44 | 4119 | struct pool_workqueue *pwq; |
b196be89 | 4120 | |
5c0338c6 TH |
4121 | /* |
4122 | * Unbound && max_active == 1 used to imply ordered, which is no | |
4123 | * longer the case on NUMA machines due to per-node pools. While | |
4124 | * alloc_ordered_workqueue() is the right way to create an ordered | |
4125 | * workqueue, keep the previous behavior to avoid subtle breakages | |
4126 | * on NUMA. | |
4127 | */ | |
4128 | if ((flags & WQ_UNBOUND) && max_active == 1) | |
4129 | flags |= __WQ_ORDERED; | |
4130 | ||
cee22a15 VK |
4131 | /* see the comment above the definition of WQ_POWER_EFFICIENT */ |
4132 | if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) | |
4133 | flags |= WQ_UNBOUND; | |
4134 | ||
ecf6881f | 4135 | /* allocate wq and format name */ |
df2d5ae4 | 4136 | if (flags & WQ_UNBOUND) |
ddcb57e2 | 4137 | tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]); |
df2d5ae4 TH |
4138 | |
4139 | wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL); | |
b196be89 | 4140 | if (!wq) |
d2c1d404 | 4141 | return NULL; |
b196be89 | 4142 | |
6029a918 TH |
4143 | if (flags & WQ_UNBOUND) { |
4144 | wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL); | |
4145 | if (!wq->unbound_attrs) | |
4146 | goto err_free_wq; | |
4147 | } | |
4148 | ||
669de8bd | 4149 | va_start(args, max_active); |
ecf6881f | 4150 | vsnprintf(wq->name, sizeof(wq->name), fmt, args); |
b196be89 | 4151 | va_end(args); |
1da177e4 | 4152 | |
d320c038 | 4153 | max_active = max_active ?: WQ_DFL_ACTIVE; |
b196be89 | 4154 | max_active = wq_clamp_max_active(max_active, flags, wq->name); |
3af24433 | 4155 | |
b196be89 | 4156 | /* init wq */ |
97e37d7b | 4157 | wq->flags = flags; |
a0a1a5fd | 4158 | wq->saved_max_active = max_active; |
3c25a55d | 4159 | mutex_init(&wq->mutex); |
112202d9 | 4160 | atomic_set(&wq->nr_pwqs_to_flush, 0); |
30cdf249 | 4161 | INIT_LIST_HEAD(&wq->pwqs); |
73f53c4a TH |
4162 | INIT_LIST_HEAD(&wq->flusher_queue); |
4163 | INIT_LIST_HEAD(&wq->flusher_overflow); | |
493a1724 | 4164 | INIT_LIST_HEAD(&wq->maydays); |
502ca9d8 | 4165 | |
669de8bd | 4166 | wq_init_lockdep(wq); |
cce1a165 | 4167 | INIT_LIST_HEAD(&wq->list); |
3af24433 | 4168 | |
30cdf249 | 4169 | if (alloc_and_link_pwqs(wq) < 0) |
d2c1d404 | 4170 | goto err_free_wq; |
1537663f | 4171 | |
40c17f75 | 4172 | if (wq_online && init_rescuer(wq) < 0) |
983c7515 | 4173 | goto err_destroy; |
3af24433 | 4174 | |
226223ab TH |
4175 | if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) |
4176 | goto err_destroy; | |
4177 | ||
a0a1a5fd | 4178 | /* |
68e13a67 LJ |
4179 | * wq_pool_mutex protects global freeze state and workqueues list. |
4180 | * Grab it, adjust max_active and add the new @wq to workqueues | |
4181 | * list. | |
a0a1a5fd | 4182 | */ |
68e13a67 | 4183 | mutex_lock(&wq_pool_mutex); |
a0a1a5fd | 4184 | |
a357fc03 | 4185 | mutex_lock(&wq->mutex); |
699ce097 TH |
4186 | for_each_pwq(pwq, wq) |
4187 | pwq_adjust_max_active(pwq); | |
a357fc03 | 4188 | mutex_unlock(&wq->mutex); |
a0a1a5fd | 4189 | |
e2dca7ad | 4190 | list_add_tail_rcu(&wq->list, &workqueues); |
a0a1a5fd | 4191 | |
68e13a67 | 4192 | mutex_unlock(&wq_pool_mutex); |
1537663f | 4193 | |
3af24433 | 4194 | return wq; |
d2c1d404 TH |
4195 | |
4196 | err_free_wq: | |
6029a918 | 4197 | free_workqueue_attrs(wq->unbound_attrs); |
d2c1d404 TH |
4198 | kfree(wq); |
4199 | return NULL; | |
4200 | err_destroy: | |
4201 | destroy_workqueue(wq); | |
4690c4ab | 4202 | return NULL; |
3af24433 | 4203 | } |
669de8bd | 4204 | EXPORT_SYMBOL_GPL(alloc_workqueue); |
1da177e4 | 4205 | |
3af24433 ON |
4206 | /** |
4207 | * destroy_workqueue - safely terminate a workqueue | |
4208 | * @wq: target workqueue | |
4209 | * | |
4210 | * Safely destroy a workqueue. All work currently pending will be done first. | |
4211 | */ | |
4212 | void destroy_workqueue(struct workqueue_struct *wq) | |
4213 | { | |
49e3cf44 | 4214 | struct pool_workqueue *pwq; |
4c16bd32 | 4215 | int node; |
3af24433 | 4216 | |
9c5a2ba7 TH |
4217 | /* drain it before proceeding with destruction */ |
4218 | drain_workqueue(wq); | |
c8efcc25 | 4219 | |
6183c009 | 4220 | /* sanity checks */ |
b09f4fd3 | 4221 | mutex_lock(&wq->mutex); |
49e3cf44 | 4222 | for_each_pwq(pwq, wq) { |
6183c009 TH |
4223 | int i; |
4224 | ||
76af4d93 TH |
4225 | for (i = 0; i < WORK_NR_COLORS; i++) { |
4226 | if (WARN_ON(pwq->nr_in_flight[i])) { | |
b09f4fd3 | 4227 | mutex_unlock(&wq->mutex); |
fa07fb6a | 4228 | show_workqueue_state(); |
6183c009 | 4229 | return; |
76af4d93 TH |
4230 | } |
4231 | } | |
4232 | ||
5c529597 | 4233 | if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) || |
8864b4e5 | 4234 | WARN_ON(pwq->nr_active) || |
76af4d93 | 4235 | WARN_ON(!list_empty(&pwq->delayed_works))) { |
b09f4fd3 | 4236 | mutex_unlock(&wq->mutex); |
fa07fb6a | 4237 | show_workqueue_state(); |
6183c009 | 4238 | return; |
76af4d93 | 4239 | } |
6183c009 | 4240 | } |
b09f4fd3 | 4241 | mutex_unlock(&wq->mutex); |
6183c009 | 4242 | |
a0a1a5fd TH |
4243 | /* |
4244 | * wq list is used to freeze wq, remove from list after | |
4245 | * flushing is complete in case freeze races us. | |
4246 | */ | |
68e13a67 | 4247 | mutex_lock(&wq_pool_mutex); |
e2dca7ad | 4248 | list_del_rcu(&wq->list); |
68e13a67 | 4249 | mutex_unlock(&wq_pool_mutex); |
3af24433 | 4250 | |
226223ab TH |
4251 | workqueue_sysfs_unregister(wq); |
4252 | ||
e2dca7ad | 4253 | if (wq->rescuer) |
e22bee78 | 4254 | kthread_stop(wq->rescuer->task); |
e22bee78 | 4255 | |
8864b4e5 | 4256 | if (!(wq->flags & WQ_UNBOUND)) { |
669de8bd | 4257 | wq_unregister_lockdep(wq); |
8864b4e5 TH |
4258 | /* |
4259 | * The base ref is never dropped on per-cpu pwqs. Directly | |
e2dca7ad | 4260 | * schedule RCU free. |
8864b4e5 | 4261 | */ |
25b00775 | 4262 | call_rcu(&wq->rcu, rcu_free_wq); |
8864b4e5 TH |
4263 | } else { |
4264 | /* | |
4265 | * We're the sole accessor of @wq at this point. Directly | |
4c16bd32 TH |
4266 | * access numa_pwq_tbl[] and dfl_pwq to put the base refs. |
4267 | * @wq will be freed when the last pwq is released. | |
8864b4e5 | 4268 | */ |
4c16bd32 TH |
4269 | for_each_node(node) { |
4270 | pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); | |
4271 | RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL); | |
4272 | put_pwq_unlocked(pwq); | |
4273 | } | |
4274 | ||
4275 | /* | |
4276 | * Put dfl_pwq. @wq may be freed any time after dfl_pwq is | |
4277 | * put. Don't access it afterwards. | |
4278 | */ | |
4279 | pwq = wq->dfl_pwq; | |
4280 | wq->dfl_pwq = NULL; | |
dce90d47 | 4281 | put_pwq_unlocked(pwq); |
29c91e99 | 4282 | } |
3af24433 ON |
4283 | } |
4284 | EXPORT_SYMBOL_GPL(destroy_workqueue); | |
4285 | ||
dcd989cb TH |
4286 | /** |
4287 | * workqueue_set_max_active - adjust max_active of a workqueue | |
4288 | * @wq: target workqueue | |
4289 | * @max_active: new max_active value. | |
4290 | * | |
4291 | * Set max_active of @wq to @max_active. | |
4292 | * | |
4293 | * CONTEXT: | |
4294 | * Don't call from IRQ context. | |
4295 | */ | |
4296 | void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) | |
4297 | { | |
49e3cf44 | 4298 | struct pool_workqueue *pwq; |
dcd989cb | 4299 | |
8719dcea | 4300 | /* disallow meddling with max_active for ordered workqueues */ |
0a94efb5 | 4301 | if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) |
8719dcea TH |
4302 | return; |
4303 | ||
f3421797 | 4304 | max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); |
dcd989cb | 4305 | |
a357fc03 | 4306 | mutex_lock(&wq->mutex); |
dcd989cb | 4307 | |
0a94efb5 | 4308 | wq->flags &= ~__WQ_ORDERED; |
dcd989cb TH |
4309 | wq->saved_max_active = max_active; |
4310 | ||
699ce097 TH |
4311 | for_each_pwq(pwq, wq) |
4312 | pwq_adjust_max_active(pwq); | |
93981800 | 4313 | |
a357fc03 | 4314 | mutex_unlock(&wq->mutex); |
15316ba8 | 4315 | } |
dcd989cb | 4316 | EXPORT_SYMBOL_GPL(workqueue_set_max_active); |
15316ba8 | 4317 | |
27d4ee03 LW |
4318 | /** |
4319 | * current_work - retrieve %current task's work struct | |
4320 | * | |
4321 | * Determine if %current task is a workqueue worker and what it's working on. | |
4322 | * Useful to find out the context that the %current task is running in. | |
4323 | * | |
4324 | * Return: work struct if %current task is a workqueue worker, %NULL otherwise. | |
4325 | */ | |
4326 | struct work_struct *current_work(void) | |
4327 | { | |
4328 | struct worker *worker = current_wq_worker(); | |
4329 | ||
4330 | return worker ? worker->current_work : NULL; | |
4331 | } | |
4332 | EXPORT_SYMBOL(current_work); | |
4333 | ||
e6267616 TH |
4334 | /** |
4335 | * current_is_workqueue_rescuer - is %current workqueue rescuer? | |
4336 | * | |
4337 | * Determine whether %current is a workqueue rescuer. Can be used from | |
4338 | * work functions to determine whether it's being run off the rescuer task. | |
d185af30 YB |
4339 | * |
4340 | * Return: %true if %current is a workqueue rescuer. %false otherwise. | |
e6267616 TH |
4341 | */ |
4342 | bool current_is_workqueue_rescuer(void) | |
4343 | { | |
4344 | struct worker *worker = current_wq_worker(); | |
4345 | ||
6a092dfd | 4346 | return worker && worker->rescue_wq; |
e6267616 TH |
4347 | } |
4348 | ||
eef6a7d5 | 4349 | /** |
dcd989cb TH |
4350 | * workqueue_congested - test whether a workqueue is congested |
4351 | * @cpu: CPU in question | |
4352 | * @wq: target workqueue | |
eef6a7d5 | 4353 | * |
dcd989cb TH |
4354 | * Test whether @wq's cpu workqueue for @cpu is congested. There is |
4355 | * no synchronization around this function and the test result is | |
4356 | * unreliable and only useful as advisory hints or for debugging. | |
eef6a7d5 | 4357 | * |
d3251859 TH |
4358 | * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. |
4359 | * Note that both per-cpu and unbound workqueues may be associated with | |
4360 | * multiple pool_workqueues which have separate congested states. A | |
4361 | * workqueue being congested on one CPU doesn't mean the workqueue is also | |
4362 | * contested on other CPUs / NUMA nodes. | |
4363 | * | |
d185af30 | 4364 | * Return: |
dcd989cb | 4365 | * %true if congested, %false otherwise. |
eef6a7d5 | 4366 | */ |
d84ff051 | 4367 | bool workqueue_congested(int cpu, struct workqueue_struct *wq) |
1da177e4 | 4368 | { |
7fb98ea7 | 4369 | struct pool_workqueue *pwq; |
76af4d93 TH |
4370 | bool ret; |
4371 | ||
88109453 | 4372 | rcu_read_lock_sched(); |
7fb98ea7 | 4373 | |
d3251859 TH |
4374 | if (cpu == WORK_CPU_UNBOUND) |
4375 | cpu = smp_processor_id(); | |
4376 | ||
7fb98ea7 TH |
4377 | if (!(wq->flags & WQ_UNBOUND)) |
4378 | pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); | |
4379 | else | |
df2d5ae4 | 4380 | pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); |
dcd989cb | 4381 | |
76af4d93 | 4382 | ret = !list_empty(&pwq->delayed_works); |
88109453 | 4383 | rcu_read_unlock_sched(); |
76af4d93 TH |
4384 | |
4385 | return ret; | |
1da177e4 | 4386 | } |
dcd989cb | 4387 | EXPORT_SYMBOL_GPL(workqueue_congested); |
1da177e4 | 4388 | |
dcd989cb TH |
4389 | /** |
4390 | * work_busy - test whether a work is currently pending or running | |
4391 | * @work: the work to be tested | |
4392 | * | |
4393 | * Test whether @work is currently pending or running. There is no | |
4394 | * synchronization around this function and the test result is | |
4395 | * unreliable and only useful as advisory hints or for debugging. | |
dcd989cb | 4396 | * |
d185af30 | 4397 | * Return: |
dcd989cb TH |
4398 | * OR'd bitmask of WORK_BUSY_* bits. |
4399 | */ | |
4400 | unsigned int work_busy(struct work_struct *work) | |
1da177e4 | 4401 | { |
fa1b54e6 | 4402 | struct worker_pool *pool; |
dcd989cb TH |
4403 | unsigned long flags; |
4404 | unsigned int ret = 0; | |
1da177e4 | 4405 | |
dcd989cb TH |
4406 | if (work_pending(work)) |
4407 | ret |= WORK_BUSY_PENDING; | |
1da177e4 | 4408 | |
fa1b54e6 TH |
4409 | local_irq_save(flags); |
4410 | pool = get_work_pool(work); | |
038366c5 | 4411 | if (pool) { |
fa1b54e6 | 4412 | spin_lock(&pool->lock); |
038366c5 LJ |
4413 | if (find_worker_executing_work(pool, work)) |
4414 | ret |= WORK_BUSY_RUNNING; | |
fa1b54e6 | 4415 | spin_unlock(&pool->lock); |
038366c5 | 4416 | } |
fa1b54e6 | 4417 | local_irq_restore(flags); |
1da177e4 | 4418 | |
dcd989cb | 4419 | return ret; |
1da177e4 | 4420 | } |
dcd989cb | 4421 | EXPORT_SYMBOL_GPL(work_busy); |
1da177e4 | 4422 | |
3d1cb205 TH |
4423 | /** |
4424 | * set_worker_desc - set description for the current work item | |
4425 | * @fmt: printf-style format string | |
4426 | * @...: arguments for the format string | |
4427 | * | |
4428 | * This function can be called by a running work function to describe what | |
4429 | * the work item is about. If the worker task gets dumped, this | |
4430 | * information will be printed out together to help debugging. The | |
4431 | * description can be at most WORKER_DESC_LEN including the trailing '\0'. | |
4432 | */ | |
4433 | void set_worker_desc(const char *fmt, ...) | |
4434 | { | |
4435 | struct worker *worker = current_wq_worker(); | |
4436 | va_list args; | |
4437 | ||
4438 | if (worker) { | |
4439 | va_start(args, fmt); | |
4440 | vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); | |
4441 | va_end(args); | |
3d1cb205 TH |
4442 | } |
4443 | } | |
5c750d58 | 4444 | EXPORT_SYMBOL_GPL(set_worker_desc); |
3d1cb205 TH |
4445 | |
4446 | /** | |
4447 | * print_worker_info - print out worker information and description | |
4448 | * @log_lvl: the log level to use when printing | |
4449 | * @task: target task | |
4450 | * | |
4451 | * If @task is a worker and currently executing a work item, print out the | |
4452 | * name of the workqueue being serviced and worker description set with | |
4453 | * set_worker_desc() by the currently executing work item. | |
4454 | * | |
4455 | * This function can be safely called on any task as long as the | |
4456 | * task_struct itself is accessible. While safe, this function isn't | |
4457 | * synchronized and may print out mixups or garbages of limited length. | |
4458 | */ | |
4459 | void print_worker_info(const char *log_lvl, struct task_struct *task) | |
4460 | { | |
4461 | work_func_t *fn = NULL; | |
4462 | char name[WQ_NAME_LEN] = { }; | |
4463 | char desc[WORKER_DESC_LEN] = { }; | |
4464 | struct pool_workqueue *pwq = NULL; | |
4465 | struct workqueue_struct *wq = NULL; | |
3d1cb205 TH |
4466 | struct worker *worker; |
4467 | ||
4468 | if (!(task->flags & PF_WQ_WORKER)) | |
4469 | return; | |
4470 | ||
4471 | /* | |
4472 | * This function is called without any synchronization and @task | |
4473 | * could be in any state. Be careful with dereferences. | |
4474 | */ | |
e700591a | 4475 | worker = kthread_probe_data(task); |
3d1cb205 TH |
4476 | |
4477 | /* | |
8bf89593 TH |
4478 | * Carefully copy the associated workqueue's workfn, name and desc. |
4479 | * Keep the original last '\0' in case the original is garbage. | |
3d1cb205 TH |
4480 | */ |
4481 | probe_kernel_read(&fn, &worker->current_func, sizeof(fn)); | |
4482 | probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq)); | |
4483 | probe_kernel_read(&wq, &pwq->wq, sizeof(wq)); | |
4484 | probe_kernel_read(name, wq->name, sizeof(name) - 1); | |
8bf89593 | 4485 | probe_kernel_read(desc, worker->desc, sizeof(desc) - 1); |
3d1cb205 TH |
4486 | |
4487 | if (fn || name[0] || desc[0]) { | |
4488 | printk("%sWorkqueue: %s %pf", log_lvl, name, fn); | |
8bf89593 | 4489 | if (strcmp(name, desc)) |
3d1cb205 TH |
4490 | pr_cont(" (%s)", desc); |
4491 | pr_cont("\n"); | |
4492 | } | |
4493 | } | |
4494 | ||
3494fc30 TH |
4495 | static void pr_cont_pool_info(struct worker_pool *pool) |
4496 | { | |
4497 | pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask); | |
4498 | if (pool->node != NUMA_NO_NODE) | |
4499 | pr_cont(" node=%d", pool->node); | |
4500 | pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice); | |
4501 | } | |
4502 | ||
4503 | static void pr_cont_work(bool comma, struct work_struct *work) | |
4504 | { | |
4505 | if (work->func == wq_barrier_func) { | |
4506 | struct wq_barrier *barr; | |
4507 | ||
4508 | barr = container_of(work, struct wq_barrier, work); | |
4509 | ||
4510 | pr_cont("%s BAR(%d)", comma ? "," : "", | |
4511 | task_pid_nr(barr->task)); | |
4512 | } else { | |
4513 | pr_cont("%s %pf", comma ? "," : "", work->func); | |
4514 | } | |
4515 | } | |
4516 | ||
4517 | static void show_pwq(struct pool_workqueue *pwq) | |
4518 | { | |
4519 | struct worker_pool *pool = pwq->pool; | |
4520 | struct work_struct *work; | |
4521 | struct worker *worker; | |
4522 | bool has_in_flight = false, has_pending = false; | |
4523 | int bkt; | |
4524 | ||
4525 | pr_info(" pwq %d:", pool->id); | |
4526 | pr_cont_pool_info(pool); | |
4527 | ||
4528 | pr_cont(" active=%d/%d%s\n", pwq->nr_active, pwq->max_active, | |
4529 | !list_empty(&pwq->mayday_node) ? " MAYDAY" : ""); | |
4530 | ||
4531 | hash_for_each(pool->busy_hash, bkt, worker, hentry) { | |
4532 | if (worker->current_pwq == pwq) { | |
4533 | has_in_flight = true; | |
4534 | break; | |
4535 | } | |
4536 | } | |
4537 | if (has_in_flight) { | |
4538 | bool comma = false; | |
4539 | ||
4540 | pr_info(" in-flight:"); | |
4541 | hash_for_each(pool->busy_hash, bkt, worker, hentry) { | |
4542 | if (worker->current_pwq != pwq) | |
4543 | continue; | |
4544 | ||
4545 | pr_cont("%s %d%s:%pf", comma ? "," : "", | |
4546 | task_pid_nr(worker->task), | |
4547 | worker == pwq->wq->rescuer ? "(RESCUER)" : "", | |
4548 | worker->current_func); | |
4549 | list_for_each_entry(work, &worker->scheduled, entry) | |
4550 | pr_cont_work(false, work); | |
4551 | comma = true; | |
4552 | } | |
4553 | pr_cont("\n"); | |
4554 | } | |
4555 | ||
4556 | list_for_each_entry(work, &pool->worklist, entry) { | |
4557 | if (get_work_pwq(work) == pwq) { | |
4558 | has_pending = true; | |
4559 | break; | |
4560 | } | |
4561 | } | |
4562 | if (has_pending) { | |
4563 | bool comma = false; | |
4564 | ||
4565 | pr_info(" pending:"); | |
4566 | list_for_each_entry(work, &pool->worklist, entry) { | |
4567 | if (get_work_pwq(work) != pwq) | |
4568 | continue; | |
4569 | ||
4570 | pr_cont_work(comma, work); | |
4571 | comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); | |
4572 | } | |
4573 | pr_cont("\n"); | |
4574 | } | |
4575 | ||
4576 | if (!list_empty(&pwq->delayed_works)) { | |
4577 | bool comma = false; | |
4578 | ||
4579 | pr_info(" delayed:"); | |
4580 | list_for_each_entry(work, &pwq->delayed_works, entry) { | |
4581 | pr_cont_work(comma, work); | |
4582 | comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); | |
4583 | } | |
4584 | pr_cont("\n"); | |
4585 | } | |
4586 | } | |
4587 | ||
4588 | /** | |
4589 | * show_workqueue_state - dump workqueue state | |
4590 | * | |
7b776af6 RL |
4591 | * Called from a sysrq handler or try_to_freeze_tasks() and prints out |
4592 | * all busy workqueues and pools. | |
3494fc30 TH |
4593 | */ |
4594 | void show_workqueue_state(void) | |
4595 | { | |
4596 | struct workqueue_struct *wq; | |
4597 | struct worker_pool *pool; | |
4598 | unsigned long flags; | |
4599 | int pi; | |
4600 | ||
4601 | rcu_read_lock_sched(); | |
4602 | ||
4603 | pr_info("Showing busy workqueues and worker pools:\n"); | |
4604 | ||
4605 | list_for_each_entry_rcu(wq, &workqueues, list) { | |
4606 | struct pool_workqueue *pwq; | |
4607 | bool idle = true; | |
4608 | ||
4609 | for_each_pwq(pwq, wq) { | |
4610 | if (pwq->nr_active || !list_empty(&pwq->delayed_works)) { | |
4611 | idle = false; | |
4612 | break; | |
4613 | } | |
4614 | } | |
4615 | if (idle) | |
4616 | continue; | |
4617 | ||
4618 | pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags); | |
4619 | ||
4620 | for_each_pwq(pwq, wq) { | |
4621 | spin_lock_irqsave(&pwq->pool->lock, flags); | |
4622 | if (pwq->nr_active || !list_empty(&pwq->delayed_works)) | |
4623 | show_pwq(pwq); | |
4624 | spin_unlock_irqrestore(&pwq->pool->lock, flags); | |
62635ea8 SS |
4625 | /* |
4626 | * We could be printing a lot from atomic context, e.g. | |
4627 | * sysrq-t -> show_workqueue_state(). Avoid triggering | |
4628 | * hard lockup. | |
4629 | */ | |
4630 | touch_nmi_watchdog(); | |
3494fc30 TH |
4631 | } |
4632 | } | |
4633 | ||
4634 | for_each_pool(pool, pi) { | |
4635 | struct worker *worker; | |
4636 | bool first = true; | |
4637 | ||
4638 | spin_lock_irqsave(&pool->lock, flags); | |
4639 | if (pool->nr_workers == pool->nr_idle) | |
4640 | goto next_pool; | |
4641 | ||
4642 | pr_info("pool %d:", pool->id); | |
4643 | pr_cont_pool_info(pool); | |
82607adc TH |
4644 | pr_cont(" hung=%us workers=%d", |
4645 | jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000, | |
4646 | pool->nr_workers); | |
3494fc30 TH |
4647 | if (pool->manager) |
4648 | pr_cont(" manager: %d", | |
4649 | task_pid_nr(pool->manager->task)); | |
4650 | list_for_each_entry(worker, &pool->idle_list, entry) { | |
4651 | pr_cont(" %s%d", first ? "idle: " : "", | |
4652 | task_pid_nr(worker->task)); | |
4653 | first = false; | |
4654 | } | |
4655 | pr_cont("\n"); | |
4656 | next_pool: | |
4657 | spin_unlock_irqrestore(&pool->lock, flags); | |
62635ea8 SS |
4658 | /* |
4659 | * We could be printing a lot from atomic context, e.g. | |
4660 | * sysrq-t -> show_workqueue_state(). Avoid triggering | |
4661 | * hard lockup. | |
4662 | */ | |
4663 | touch_nmi_watchdog(); | |
3494fc30 TH |
4664 | } |
4665 | ||
4666 | rcu_read_unlock_sched(); | |
4667 | } | |
4668 | ||
6b59808b TH |
4669 | /* used to show worker information through /proc/PID/{comm,stat,status} */ |
4670 | void wq_worker_comm(char *buf, size_t size, struct task_struct *task) | |
4671 | { | |
6b59808b TH |
4672 | int off; |
4673 | ||
4674 | /* always show the actual comm */ | |
4675 | off = strscpy(buf, task->comm, size); | |
4676 | if (off < 0) | |
4677 | return; | |
4678 | ||
197f6acc | 4679 | /* stabilize PF_WQ_WORKER and worker pool association */ |
6b59808b TH |
4680 | mutex_lock(&wq_pool_attach_mutex); |
4681 | ||
197f6acc TH |
4682 | if (task->flags & PF_WQ_WORKER) { |
4683 | struct worker *worker = kthread_data(task); | |
4684 | struct worker_pool *pool = worker->pool; | |
6b59808b | 4685 | |
197f6acc TH |
4686 | if (pool) { |
4687 | spin_lock_irq(&pool->lock); | |
4688 | /* | |
4689 | * ->desc tracks information (wq name or | |
4690 | * set_worker_desc()) for the latest execution. If | |
4691 | * current, prepend '+', otherwise '-'. | |
4692 | */ | |
4693 | if (worker->desc[0] != '\0') { | |
4694 | if (worker->current_work) | |
4695 | scnprintf(buf + off, size - off, "+%s", | |
4696 | worker->desc); | |
4697 | else | |
4698 | scnprintf(buf + off, size - off, "-%s", | |
4699 | worker->desc); | |
4700 | } | |
4701 | spin_unlock_irq(&pool->lock); | |
6b59808b | 4702 | } |
6b59808b TH |
4703 | } |
4704 | ||
4705 | mutex_unlock(&wq_pool_attach_mutex); | |
4706 | } | |
4707 | ||
66448bc2 MM |
4708 | #ifdef CONFIG_SMP |
4709 | ||
db7bccf4 TH |
4710 | /* |
4711 | * CPU hotplug. | |
4712 | * | |
e22bee78 | 4713 | * There are two challenges in supporting CPU hotplug. Firstly, there |
112202d9 | 4714 | * are a lot of assumptions on strong associations among work, pwq and |
706026c2 | 4715 | * pool which make migrating pending and scheduled works very |
e22bee78 | 4716 | * difficult to implement without impacting hot paths. Secondly, |
94cf58bb | 4717 | * worker pools serve mix of short, long and very long running works making |
e22bee78 TH |
4718 | * blocked draining impractical. |
4719 | * | |
24647570 | 4720 | * This is solved by allowing the pools to be disassociated from the CPU |
628c78e7 TH |
4721 | * running as an unbound one and allowing it to be reattached later if the |
4722 | * cpu comes back online. | |
db7bccf4 | 4723 | */ |
1da177e4 | 4724 | |
e8b3f8db | 4725 | static void unbind_workers(int cpu) |
3af24433 | 4726 | { |
4ce62e9e | 4727 | struct worker_pool *pool; |
db7bccf4 | 4728 | struct worker *worker; |
3af24433 | 4729 | |
f02ae73a | 4730 | for_each_cpu_worker_pool(pool, cpu) { |
1258fae7 | 4731 | mutex_lock(&wq_pool_attach_mutex); |
94cf58bb | 4732 | spin_lock_irq(&pool->lock); |
3af24433 | 4733 | |
94cf58bb | 4734 | /* |
92f9c5c4 | 4735 | * We've blocked all attach/detach operations. Make all workers |
94cf58bb TH |
4736 | * unbound and set DISASSOCIATED. Before this, all workers |
4737 | * except for the ones which are still executing works from | |
4738 | * before the last CPU down must be on the cpu. After | |
4739 | * this, they may become diasporas. | |
4740 | */ | |
da028469 | 4741 | for_each_pool_worker(worker, pool) |
c9e7cf27 | 4742 | worker->flags |= WORKER_UNBOUND; |
06ba38a9 | 4743 | |
24647570 | 4744 | pool->flags |= POOL_DISASSOCIATED; |
f2d5a0ee | 4745 | |
94cf58bb | 4746 | spin_unlock_irq(&pool->lock); |
1258fae7 | 4747 | mutex_unlock(&wq_pool_attach_mutex); |
628c78e7 | 4748 | |
eb283428 LJ |
4749 | /* |
4750 | * Call schedule() so that we cross rq->lock and thus can | |
4751 | * guarantee sched callbacks see the %WORKER_UNBOUND flag. | |
4752 | * This is necessary as scheduler callbacks may be invoked | |
4753 | * from other cpus. | |
4754 | */ | |
4755 | schedule(); | |
06ba38a9 | 4756 | |
eb283428 LJ |
4757 | /* |
4758 | * Sched callbacks are disabled now. Zap nr_running. | |
4759 | * After this, nr_running stays zero and need_more_worker() | |
4760 | * and keep_working() are always true as long as the | |
4761 | * worklist is not empty. This pool now behaves as an | |
4762 | * unbound (in terms of concurrency management) pool which | |
4763 | * are served by workers tied to the pool. | |
4764 | */ | |
e19e397a | 4765 | atomic_set(&pool->nr_running, 0); |
eb283428 LJ |
4766 | |
4767 | /* | |
4768 | * With concurrency management just turned off, a busy | |
4769 | * worker blocking could lead to lengthy stalls. Kick off | |
4770 | * unbound chain execution of currently pending work items. | |
4771 | */ | |
4772 | spin_lock_irq(&pool->lock); | |
4773 | wake_up_worker(pool); | |
4774 | spin_unlock_irq(&pool->lock); | |
4775 | } | |
3af24433 | 4776 | } |
3af24433 | 4777 | |
bd7c089e TH |
4778 | /** |
4779 | * rebind_workers - rebind all workers of a pool to the associated CPU | |
4780 | * @pool: pool of interest | |
4781 | * | |
a9ab775b | 4782 | * @pool->cpu is coming online. Rebind all workers to the CPU. |
bd7c089e TH |
4783 | */ |
4784 | static void rebind_workers(struct worker_pool *pool) | |
4785 | { | |
a9ab775b | 4786 | struct worker *worker; |
bd7c089e | 4787 | |
1258fae7 | 4788 | lockdep_assert_held(&wq_pool_attach_mutex); |
bd7c089e | 4789 | |
a9ab775b TH |
4790 | /* |
4791 | * Restore CPU affinity of all workers. As all idle workers should | |
4792 | * be on the run-queue of the associated CPU before any local | |
402dd89d | 4793 | * wake-ups for concurrency management happen, restore CPU affinity |
a9ab775b TH |
4794 | * of all workers first and then clear UNBOUND. As we're called |
4795 | * from CPU_ONLINE, the following shouldn't fail. | |
4796 | */ | |
da028469 | 4797 | for_each_pool_worker(worker, pool) |
a9ab775b TH |
4798 | WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, |
4799 | pool->attrs->cpumask) < 0); | |
bd7c089e | 4800 | |
a9ab775b | 4801 | spin_lock_irq(&pool->lock); |
f7c17d26 | 4802 | |
3de5e884 | 4803 | pool->flags &= ~POOL_DISASSOCIATED; |
bd7c089e | 4804 | |
da028469 | 4805 | for_each_pool_worker(worker, pool) { |
a9ab775b | 4806 | unsigned int worker_flags = worker->flags; |
bd7c089e TH |
4807 | |
4808 | /* | |
a9ab775b TH |
4809 | * A bound idle worker should actually be on the runqueue |
4810 | * of the associated CPU for local wake-ups targeting it to | |
4811 | * work. Kick all idle workers so that they migrate to the | |
4812 | * associated CPU. Doing this in the same loop as | |
4813 | * replacing UNBOUND with REBOUND is safe as no worker will | |
4814 | * be bound before @pool->lock is released. | |
bd7c089e | 4815 | */ |
a9ab775b TH |
4816 | if (worker_flags & WORKER_IDLE) |
4817 | wake_up_process(worker->task); | |
bd7c089e | 4818 | |
a9ab775b TH |
4819 | /* |
4820 | * We want to clear UNBOUND but can't directly call | |
4821 | * worker_clr_flags() or adjust nr_running. Atomically | |
4822 | * replace UNBOUND with another NOT_RUNNING flag REBOUND. | |
4823 | * @worker will clear REBOUND using worker_clr_flags() when | |
4824 | * it initiates the next execution cycle thus restoring | |
4825 | * concurrency management. Note that when or whether | |
4826 | * @worker clears REBOUND doesn't affect correctness. | |
4827 | * | |
c95491ed | 4828 | * WRITE_ONCE() is necessary because @worker->flags may be |
a9ab775b TH |
4829 | * tested without holding any lock in |
4830 | * wq_worker_waking_up(). Without it, NOT_RUNNING test may | |
4831 | * fail incorrectly leading to premature concurrency | |
4832 | * management operations. | |
4833 | */ | |
4834 | WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); | |
4835 | worker_flags |= WORKER_REBOUND; | |
4836 | worker_flags &= ~WORKER_UNBOUND; | |
c95491ed | 4837 | WRITE_ONCE(worker->flags, worker_flags); |
bd7c089e | 4838 | } |
a9ab775b TH |
4839 | |
4840 | spin_unlock_irq(&pool->lock); | |
bd7c089e TH |
4841 | } |
4842 | ||
7dbc725e TH |
4843 | /** |
4844 | * restore_unbound_workers_cpumask - restore cpumask of unbound workers | |
4845 | * @pool: unbound pool of interest | |
4846 | * @cpu: the CPU which is coming up | |
4847 | * | |
4848 | * An unbound pool may end up with a cpumask which doesn't have any online | |
4849 | * CPUs. When a worker of such pool get scheduled, the scheduler resets | |
4850 | * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any | |
4851 | * online CPU before, cpus_allowed of all its workers should be restored. | |
4852 | */ | |
4853 | static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) | |
4854 | { | |
4855 | static cpumask_t cpumask; | |
4856 | struct worker *worker; | |
7dbc725e | 4857 | |
1258fae7 | 4858 | lockdep_assert_held(&wq_pool_attach_mutex); |
7dbc725e TH |
4859 | |
4860 | /* is @cpu allowed for @pool? */ | |
4861 | if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) | |
4862 | return; | |
4863 | ||
7dbc725e | 4864 | cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); |
7dbc725e TH |
4865 | |
4866 | /* as we're called from CPU_ONLINE, the following shouldn't fail */ | |
da028469 | 4867 | for_each_pool_worker(worker, pool) |
d945b5e9 | 4868 | WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0); |
7dbc725e TH |
4869 | } |
4870 | ||
7ee681b2 TG |
4871 | int workqueue_prepare_cpu(unsigned int cpu) |
4872 | { | |
4873 | struct worker_pool *pool; | |
4874 | ||
4875 | for_each_cpu_worker_pool(pool, cpu) { | |
4876 | if (pool->nr_workers) | |
4877 | continue; | |
4878 | if (!create_worker(pool)) | |
4879 | return -ENOMEM; | |
4880 | } | |
4881 | return 0; | |
4882 | } | |
4883 | ||
4884 | int workqueue_online_cpu(unsigned int cpu) | |
3af24433 | 4885 | { |
4ce62e9e | 4886 | struct worker_pool *pool; |
4c16bd32 | 4887 | struct workqueue_struct *wq; |
7dbc725e | 4888 | int pi; |
3ce63377 | 4889 | |
7ee681b2 | 4890 | mutex_lock(&wq_pool_mutex); |
7dbc725e | 4891 | |
7ee681b2 | 4892 | for_each_pool(pool, pi) { |
1258fae7 | 4893 | mutex_lock(&wq_pool_attach_mutex); |
94cf58bb | 4894 | |
7ee681b2 TG |
4895 | if (pool->cpu == cpu) |
4896 | rebind_workers(pool); | |
4897 | else if (pool->cpu < 0) | |
4898 | restore_unbound_workers_cpumask(pool, cpu); | |
94cf58bb | 4899 | |
1258fae7 | 4900 | mutex_unlock(&wq_pool_attach_mutex); |
7ee681b2 | 4901 | } |
6ba94429 | 4902 | |
7ee681b2 TG |
4903 | /* update NUMA affinity of unbound workqueues */ |
4904 | list_for_each_entry(wq, &workqueues, list) | |
4905 | wq_update_unbound_numa(wq, cpu, true); | |
6ba94429 | 4906 | |
7ee681b2 TG |
4907 | mutex_unlock(&wq_pool_mutex); |
4908 | return 0; | |
6ba94429 FW |
4909 | } |
4910 | ||
7ee681b2 | 4911 | int workqueue_offline_cpu(unsigned int cpu) |
6ba94429 | 4912 | { |
6ba94429 FW |
4913 | struct workqueue_struct *wq; |
4914 | ||
7ee681b2 | 4915 | /* unbinding per-cpu workers should happen on the local CPU */ |
e8b3f8db LJ |
4916 | if (WARN_ON(cpu != smp_processor_id())) |
4917 | return -1; | |
4918 | ||
4919 | unbind_workers(cpu); | |
7ee681b2 TG |
4920 | |
4921 | /* update NUMA affinity of unbound workqueues */ | |
4922 | mutex_lock(&wq_pool_mutex); | |
4923 | list_for_each_entry(wq, &workqueues, list) | |
4924 | wq_update_unbound_numa(wq, cpu, false); | |
4925 | mutex_unlock(&wq_pool_mutex); | |
4926 | ||
7ee681b2 | 4927 | return 0; |
6ba94429 FW |
4928 | } |
4929 | ||
6ba94429 FW |
4930 | struct work_for_cpu { |
4931 | struct work_struct work; | |
4932 | long (*fn)(void *); | |
4933 | void *arg; | |
4934 | long ret; | |
4935 | }; | |
4936 | ||
4937 | static void work_for_cpu_fn(struct work_struct *work) | |
4938 | { | |
4939 | struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); | |
4940 | ||
4941 | wfc->ret = wfc->fn(wfc->arg); | |
4942 | } | |
4943 | ||
4944 | /** | |
22aceb31 | 4945 | * work_on_cpu - run a function in thread context on a particular cpu |
6ba94429 FW |
4946 | * @cpu: the cpu to run on |
4947 | * @fn: the function to run | |
4948 | * @arg: the function arg | |
4949 | * | |
4950 | * It is up to the caller to ensure that the cpu doesn't go offline. | |
4951 | * The caller must not hold any locks which would prevent @fn from completing. | |
4952 | * | |
4953 | * Return: The value @fn returns. | |
4954 | */ | |
4955 | long work_on_cpu(int cpu, long (*fn)(void *), void *arg) | |
4956 | { | |
4957 | struct work_for_cpu wfc = { .fn = fn, .arg = arg }; | |
4958 | ||
4959 | INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); | |
4960 | schedule_work_on(cpu, &wfc.work); | |
4961 | flush_work(&wfc.work); | |
4962 | destroy_work_on_stack(&wfc.work); | |
4963 | return wfc.ret; | |
4964 | } | |
4965 | EXPORT_SYMBOL_GPL(work_on_cpu); | |
0e8d6a93 TG |
4966 | |
4967 | /** | |
4968 | * work_on_cpu_safe - run a function in thread context on a particular cpu | |
4969 | * @cpu: the cpu to run on | |
4970 | * @fn: the function to run | |
4971 | * @arg: the function argument | |
4972 | * | |
4973 | * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold | |
4974 | * any locks which would prevent @fn from completing. | |
4975 | * | |
4976 | * Return: The value @fn returns. | |
4977 | */ | |
4978 | long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg) | |
4979 | { | |
4980 | long ret = -ENODEV; | |
4981 | ||
4982 | get_online_cpus(); | |
4983 | if (cpu_online(cpu)) | |
4984 | ret = work_on_cpu(cpu, fn, arg); | |
4985 | put_online_cpus(); | |
4986 | return ret; | |
4987 | } | |
4988 | EXPORT_SYMBOL_GPL(work_on_cpu_safe); | |
6ba94429 FW |
4989 | #endif /* CONFIG_SMP */ |
4990 | ||
4991 | #ifdef CONFIG_FREEZER | |
4992 | ||
4993 | /** | |
4994 | * freeze_workqueues_begin - begin freezing workqueues | |
4995 | * | |
4996 | * Start freezing workqueues. After this function returns, all freezable | |
4997 | * workqueues will queue new works to their delayed_works list instead of | |
4998 | * pool->worklist. | |
4999 | * | |
5000 | * CONTEXT: | |
5001 | * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. | |
5002 | */ | |
5003 | void freeze_workqueues_begin(void) | |
5004 | { | |
5005 | struct workqueue_struct *wq; | |
5006 | struct pool_workqueue *pwq; | |
5007 | ||
5008 | mutex_lock(&wq_pool_mutex); | |
5009 | ||
5010 | WARN_ON_ONCE(workqueue_freezing); | |
5011 | workqueue_freezing = true; | |
5012 | ||
5013 | list_for_each_entry(wq, &workqueues, list) { | |
5014 | mutex_lock(&wq->mutex); | |
5015 | for_each_pwq(pwq, wq) | |
5016 | pwq_adjust_max_active(pwq); | |
5017 | mutex_unlock(&wq->mutex); | |
5018 | } | |
5019 | ||
5020 | mutex_unlock(&wq_pool_mutex); | |
5021 | } | |
5022 | ||
5023 | /** | |
5024 | * freeze_workqueues_busy - are freezable workqueues still busy? | |
5025 | * | |
5026 | * Check whether freezing is complete. This function must be called | |
5027 | * between freeze_workqueues_begin() and thaw_workqueues(). | |
5028 | * | |
5029 | * CONTEXT: | |
5030 | * Grabs and releases wq_pool_mutex. | |
5031 | * | |
5032 | * Return: | |
5033 | * %true if some freezable workqueues are still busy. %false if freezing | |
5034 | * is complete. | |
5035 | */ | |
5036 | bool freeze_workqueues_busy(void) | |
5037 | { | |
5038 | bool busy = false; | |
5039 | struct workqueue_struct *wq; | |
5040 | struct pool_workqueue *pwq; | |
5041 | ||
5042 | mutex_lock(&wq_pool_mutex); | |
5043 | ||
5044 | WARN_ON_ONCE(!workqueue_freezing); | |
5045 | ||
5046 | list_for_each_entry(wq, &workqueues, list) { | |
5047 | if (!(wq->flags & WQ_FREEZABLE)) | |
5048 | continue; | |
5049 | /* | |
5050 | * nr_active is monotonically decreasing. It's safe | |
5051 | * to peek without lock. | |
5052 | */ | |
5053 | rcu_read_lock_sched(); | |
5054 | for_each_pwq(pwq, wq) { | |
5055 | WARN_ON_ONCE(pwq->nr_active < 0); | |
5056 | if (pwq->nr_active) { | |
5057 | busy = true; | |
5058 | rcu_read_unlock_sched(); | |
5059 | goto out_unlock; | |
5060 | } | |
5061 | } | |
5062 | rcu_read_unlock_sched(); | |
5063 | } | |
5064 | out_unlock: | |
5065 | mutex_unlock(&wq_pool_mutex); | |
5066 | return busy; | |
5067 | } | |
5068 | ||
5069 | /** | |
5070 | * thaw_workqueues - thaw workqueues | |
5071 | * | |
5072 | * Thaw workqueues. Normal queueing is restored and all collected | |
5073 | * frozen works are transferred to their respective pool worklists. | |
5074 | * | |
5075 | * CONTEXT: | |
5076 | * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. | |
5077 | */ | |
5078 | void thaw_workqueues(void) | |
5079 | { | |
5080 | struct workqueue_struct *wq; | |
5081 | struct pool_workqueue *pwq; | |
5082 | ||
5083 | mutex_lock(&wq_pool_mutex); | |
5084 | ||
5085 | if (!workqueue_freezing) | |
5086 | goto out_unlock; | |
5087 | ||
5088 | workqueue_freezing = false; | |
5089 | ||
5090 | /* restore max_active and repopulate worklist */ | |
5091 | list_for_each_entry(wq, &workqueues, list) { | |
5092 | mutex_lock(&wq->mutex); | |
5093 | for_each_pwq(pwq, wq) | |
5094 | pwq_adjust_max_active(pwq); | |
5095 | mutex_unlock(&wq->mutex); | |
5096 | } | |
5097 | ||
5098 | out_unlock: | |
5099 | mutex_unlock(&wq_pool_mutex); | |
5100 | } | |
5101 | #endif /* CONFIG_FREEZER */ | |
5102 | ||
042f7df1 LJ |
5103 | static int workqueue_apply_unbound_cpumask(void) |
5104 | { | |
5105 | LIST_HEAD(ctxs); | |
5106 | int ret = 0; | |
5107 | struct workqueue_struct *wq; | |
5108 | struct apply_wqattrs_ctx *ctx, *n; | |
5109 | ||
5110 | lockdep_assert_held(&wq_pool_mutex); | |
5111 | ||
5112 | list_for_each_entry(wq, &workqueues, list) { | |
5113 | if (!(wq->flags & WQ_UNBOUND)) | |
5114 | continue; | |
5115 | /* creating multiple pwqs breaks ordering guarantee */ | |
5116 | if (wq->flags & __WQ_ORDERED) | |
5117 | continue; | |
5118 | ||
5119 | ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs); | |
5120 | if (!ctx) { | |
5121 | ret = -ENOMEM; | |
5122 | break; | |
5123 | } | |
5124 | ||
5125 | list_add_tail(&ctx->list, &ctxs); | |
5126 | } | |
5127 | ||
5128 | list_for_each_entry_safe(ctx, n, &ctxs, list) { | |
5129 | if (!ret) | |
5130 | apply_wqattrs_commit(ctx); | |
5131 | apply_wqattrs_cleanup(ctx); | |
5132 | } | |
5133 | ||
5134 | return ret; | |
5135 | } | |
5136 | ||
5137 | /** | |
5138 | * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask | |
5139 | * @cpumask: the cpumask to set | |
5140 | * | |
5141 | * The low-level workqueues cpumask is a global cpumask that limits | |
5142 | * the affinity of all unbound workqueues. This function check the @cpumask | |
5143 | * and apply it to all unbound workqueues and updates all pwqs of them. | |
5144 | * | |
5145 | * Retun: 0 - Success | |
5146 | * -EINVAL - Invalid @cpumask | |
5147 | * -ENOMEM - Failed to allocate memory for attrs or pwqs. | |
5148 | */ | |
5149 | int workqueue_set_unbound_cpumask(cpumask_var_t cpumask) | |
5150 | { | |
5151 | int ret = -EINVAL; | |
5152 | cpumask_var_t saved_cpumask; | |
5153 | ||
5154 | if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL)) | |
5155 | return -ENOMEM; | |
5156 | ||
c98a9805 TS |
5157 | /* |
5158 | * Not excluding isolated cpus on purpose. | |
5159 | * If the user wishes to include them, we allow that. | |
5160 | */ | |
042f7df1 LJ |
5161 | cpumask_and(cpumask, cpumask, cpu_possible_mask); |
5162 | if (!cpumask_empty(cpumask)) { | |
a0111cf6 | 5163 | apply_wqattrs_lock(); |
042f7df1 LJ |
5164 | |
5165 | /* save the old wq_unbound_cpumask. */ | |
5166 | cpumask_copy(saved_cpumask, wq_unbound_cpumask); | |
5167 | ||
5168 | /* update wq_unbound_cpumask at first and apply it to wqs. */ | |
5169 | cpumask_copy(wq_unbound_cpumask, cpumask); | |
5170 | ret = workqueue_apply_unbound_cpumask(); | |
5171 | ||
5172 | /* restore the wq_unbound_cpumask when failed. */ | |
5173 | if (ret < 0) | |
5174 | cpumask_copy(wq_unbound_cpumask, saved_cpumask); | |
5175 | ||
a0111cf6 | 5176 | apply_wqattrs_unlock(); |
042f7df1 | 5177 | } |
042f7df1 LJ |
5178 | |
5179 | free_cpumask_var(saved_cpumask); | |
5180 | return ret; | |
5181 | } | |
5182 | ||
6ba94429 FW |
5183 | #ifdef CONFIG_SYSFS |
5184 | /* | |
5185 | * Workqueues with WQ_SYSFS flag set is visible to userland via | |
5186 | * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the | |
5187 | * following attributes. | |
5188 | * | |
5189 | * per_cpu RO bool : whether the workqueue is per-cpu or unbound | |
5190 | * max_active RW int : maximum number of in-flight work items | |
5191 | * | |
5192 | * Unbound workqueues have the following extra attributes. | |
5193 | * | |
9a19b463 | 5194 | * pool_ids RO int : the associated pool IDs for each node |
6ba94429 FW |
5195 | * nice RW int : nice value of the workers |
5196 | * cpumask RW mask : bitmask of allowed CPUs for the workers | |
9a19b463 | 5197 | * numa RW bool : whether enable NUMA affinity |
6ba94429 FW |
5198 | */ |
5199 | struct wq_device { | |
5200 | struct workqueue_struct *wq; | |
5201 | struct device dev; | |
5202 | }; | |
5203 | ||
5204 | static struct workqueue_struct *dev_to_wq(struct device *dev) | |
5205 | { | |
5206 | struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); | |
5207 | ||
5208 | return wq_dev->wq; | |
5209 | } | |
5210 | ||
5211 | static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, | |
5212 | char *buf) | |
5213 | { | |
5214 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5215 | ||
5216 | return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); | |
5217 | } | |
5218 | static DEVICE_ATTR_RO(per_cpu); | |
5219 | ||
5220 | static ssize_t max_active_show(struct device *dev, | |
5221 | struct device_attribute *attr, char *buf) | |
5222 | { | |
5223 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5224 | ||
5225 | return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); | |
5226 | } | |
5227 | ||
5228 | static ssize_t max_active_store(struct device *dev, | |
5229 | struct device_attribute *attr, const char *buf, | |
5230 | size_t count) | |
5231 | { | |
5232 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5233 | int val; | |
5234 | ||
5235 | if (sscanf(buf, "%d", &val) != 1 || val <= 0) | |
5236 | return -EINVAL; | |
5237 | ||
5238 | workqueue_set_max_active(wq, val); | |
5239 | return count; | |
5240 | } | |
5241 | static DEVICE_ATTR_RW(max_active); | |
5242 | ||
5243 | static struct attribute *wq_sysfs_attrs[] = { | |
5244 | &dev_attr_per_cpu.attr, | |
5245 | &dev_attr_max_active.attr, | |
5246 | NULL, | |
5247 | }; | |
5248 | ATTRIBUTE_GROUPS(wq_sysfs); | |
5249 | ||
5250 | static ssize_t wq_pool_ids_show(struct device *dev, | |
5251 | struct device_attribute *attr, char *buf) | |
5252 | { | |
5253 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5254 | const char *delim = ""; | |
5255 | int node, written = 0; | |
5256 | ||
5257 | rcu_read_lock_sched(); | |
5258 | for_each_node(node) { | |
5259 | written += scnprintf(buf + written, PAGE_SIZE - written, | |
5260 | "%s%d:%d", delim, node, | |
5261 | unbound_pwq_by_node(wq, node)->pool->id); | |
5262 | delim = " "; | |
5263 | } | |
5264 | written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); | |
5265 | rcu_read_unlock_sched(); | |
5266 | ||
5267 | return written; | |
5268 | } | |
5269 | ||
5270 | static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, | |
5271 | char *buf) | |
5272 | { | |
5273 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5274 | int written; | |
5275 | ||
5276 | mutex_lock(&wq->mutex); | |
5277 | written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); | |
5278 | mutex_unlock(&wq->mutex); | |
5279 | ||
5280 | return written; | |
5281 | } | |
5282 | ||
5283 | /* prepare workqueue_attrs for sysfs store operations */ | |
5284 | static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) | |
5285 | { | |
5286 | struct workqueue_attrs *attrs; | |
5287 | ||
899a94fe LJ |
5288 | lockdep_assert_held(&wq_pool_mutex); |
5289 | ||
6ba94429 FW |
5290 | attrs = alloc_workqueue_attrs(GFP_KERNEL); |
5291 | if (!attrs) | |
5292 | return NULL; | |
5293 | ||
6ba94429 | 5294 | copy_workqueue_attrs(attrs, wq->unbound_attrs); |
6ba94429 FW |
5295 | return attrs; |
5296 | } | |
5297 | ||
5298 | static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, | |
5299 | const char *buf, size_t count) | |
5300 | { | |
5301 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5302 | struct workqueue_attrs *attrs; | |
d4d3e257 LJ |
5303 | int ret = -ENOMEM; |
5304 | ||
5305 | apply_wqattrs_lock(); | |
6ba94429 FW |
5306 | |
5307 | attrs = wq_sysfs_prep_attrs(wq); | |
5308 | if (!attrs) | |
d4d3e257 | 5309 | goto out_unlock; |
6ba94429 FW |
5310 | |
5311 | if (sscanf(buf, "%d", &attrs->nice) == 1 && | |
5312 | attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) | |
d4d3e257 | 5313 | ret = apply_workqueue_attrs_locked(wq, attrs); |
6ba94429 FW |
5314 | else |
5315 | ret = -EINVAL; | |
5316 | ||
d4d3e257 LJ |
5317 | out_unlock: |
5318 | apply_wqattrs_unlock(); | |
6ba94429 FW |
5319 | free_workqueue_attrs(attrs); |
5320 | return ret ?: count; | |
5321 | } | |
5322 | ||
5323 | static ssize_t wq_cpumask_show(struct device *dev, | |
5324 | struct device_attribute *attr, char *buf) | |
5325 | { | |
5326 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5327 | int written; | |
5328 | ||
5329 | mutex_lock(&wq->mutex); | |
5330 | written = scnprintf(buf, PAGE_SIZE, "%*pb\n", | |
5331 | cpumask_pr_args(wq->unbound_attrs->cpumask)); | |
5332 | mutex_unlock(&wq->mutex); | |
5333 | return written; | |
5334 | } | |
5335 | ||
5336 | static ssize_t wq_cpumask_store(struct device *dev, | |
5337 | struct device_attribute *attr, | |
5338 | const char *buf, size_t count) | |
5339 | { | |
5340 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5341 | struct workqueue_attrs *attrs; | |
d4d3e257 LJ |
5342 | int ret = -ENOMEM; |
5343 | ||
5344 | apply_wqattrs_lock(); | |
6ba94429 FW |
5345 | |
5346 | attrs = wq_sysfs_prep_attrs(wq); | |
5347 | if (!attrs) | |
d4d3e257 | 5348 | goto out_unlock; |
6ba94429 FW |
5349 | |
5350 | ret = cpumask_parse(buf, attrs->cpumask); | |
5351 | if (!ret) | |
d4d3e257 | 5352 | ret = apply_workqueue_attrs_locked(wq, attrs); |
6ba94429 | 5353 | |
d4d3e257 LJ |
5354 | out_unlock: |
5355 | apply_wqattrs_unlock(); | |
6ba94429 FW |
5356 | free_workqueue_attrs(attrs); |
5357 | return ret ?: count; | |
5358 | } | |
5359 | ||
5360 | static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr, | |
5361 | char *buf) | |
5362 | { | |
5363 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5364 | int written; | |
7dbc725e | 5365 | |
6ba94429 FW |
5366 | mutex_lock(&wq->mutex); |
5367 | written = scnprintf(buf, PAGE_SIZE, "%d\n", | |
5368 | !wq->unbound_attrs->no_numa); | |
5369 | mutex_unlock(&wq->mutex); | |
4c16bd32 | 5370 | |
6ba94429 | 5371 | return written; |
65758202 TH |
5372 | } |
5373 | ||
6ba94429 FW |
5374 | static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr, |
5375 | const char *buf, size_t count) | |
65758202 | 5376 | { |
6ba94429 FW |
5377 | struct workqueue_struct *wq = dev_to_wq(dev); |
5378 | struct workqueue_attrs *attrs; | |
d4d3e257 LJ |
5379 | int v, ret = -ENOMEM; |
5380 | ||
5381 | apply_wqattrs_lock(); | |
4c16bd32 | 5382 | |
6ba94429 FW |
5383 | attrs = wq_sysfs_prep_attrs(wq); |
5384 | if (!attrs) | |
d4d3e257 | 5385 | goto out_unlock; |
4c16bd32 | 5386 | |
6ba94429 FW |
5387 | ret = -EINVAL; |
5388 | if (sscanf(buf, "%d", &v) == 1) { | |
5389 | attrs->no_numa = !v; | |
d4d3e257 | 5390 | ret = apply_workqueue_attrs_locked(wq, attrs); |
65758202 | 5391 | } |
6ba94429 | 5392 | |
d4d3e257 LJ |
5393 | out_unlock: |
5394 | apply_wqattrs_unlock(); | |
6ba94429 FW |
5395 | free_workqueue_attrs(attrs); |
5396 | return ret ?: count; | |
65758202 TH |
5397 | } |
5398 | ||
6ba94429 FW |
5399 | static struct device_attribute wq_sysfs_unbound_attrs[] = { |
5400 | __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL), | |
5401 | __ATTR(nice, 0644, wq_nice_show, wq_nice_store), | |
5402 | __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), | |
5403 | __ATTR(numa, 0644, wq_numa_show, wq_numa_store), | |
5404 | __ATTR_NULL, | |
5405 | }; | |
8ccad40d | 5406 | |
6ba94429 FW |
5407 | static struct bus_type wq_subsys = { |
5408 | .name = "workqueue", | |
5409 | .dev_groups = wq_sysfs_groups, | |
2d3854a3 RR |
5410 | }; |
5411 | ||
b05a7928 FW |
5412 | static ssize_t wq_unbound_cpumask_show(struct device *dev, |
5413 | struct device_attribute *attr, char *buf) | |
5414 | { | |
5415 | int written; | |
5416 | ||
042f7df1 | 5417 | mutex_lock(&wq_pool_mutex); |
b05a7928 FW |
5418 | written = scnprintf(buf, PAGE_SIZE, "%*pb\n", |
5419 | cpumask_pr_args(wq_unbound_cpumask)); | |
042f7df1 | 5420 | mutex_unlock(&wq_pool_mutex); |
b05a7928 FW |
5421 | |
5422 | return written; | |
5423 | } | |
5424 | ||
042f7df1 LJ |
5425 | static ssize_t wq_unbound_cpumask_store(struct device *dev, |
5426 | struct device_attribute *attr, const char *buf, size_t count) | |
5427 | { | |
5428 | cpumask_var_t cpumask; | |
5429 | int ret; | |
5430 | ||
5431 | if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) | |
5432 | return -ENOMEM; | |
5433 | ||
5434 | ret = cpumask_parse(buf, cpumask); | |
5435 | if (!ret) | |
5436 | ret = workqueue_set_unbound_cpumask(cpumask); | |
5437 | ||
5438 | free_cpumask_var(cpumask); | |
5439 | return ret ? ret : count; | |
5440 | } | |
5441 | ||
b05a7928 | 5442 | static struct device_attribute wq_sysfs_cpumask_attr = |
042f7df1 LJ |
5443 | __ATTR(cpumask, 0644, wq_unbound_cpumask_show, |
5444 | wq_unbound_cpumask_store); | |
b05a7928 | 5445 | |
6ba94429 | 5446 | static int __init wq_sysfs_init(void) |
2d3854a3 | 5447 | { |
b05a7928 FW |
5448 | int err; |
5449 | ||
5450 | err = subsys_virtual_register(&wq_subsys, NULL); | |
5451 | if (err) | |
5452 | return err; | |
5453 | ||
5454 | return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr); | |
2d3854a3 | 5455 | } |
6ba94429 | 5456 | core_initcall(wq_sysfs_init); |
2d3854a3 | 5457 | |
6ba94429 | 5458 | static void wq_device_release(struct device *dev) |
2d3854a3 | 5459 | { |
6ba94429 | 5460 | struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); |
6b44003e | 5461 | |
6ba94429 | 5462 | kfree(wq_dev); |
2d3854a3 | 5463 | } |
a0a1a5fd TH |
5464 | |
5465 | /** | |
6ba94429 FW |
5466 | * workqueue_sysfs_register - make a workqueue visible in sysfs |
5467 | * @wq: the workqueue to register | |
a0a1a5fd | 5468 | * |
6ba94429 FW |
5469 | * Expose @wq in sysfs under /sys/bus/workqueue/devices. |
5470 | * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set | |
5471 | * which is the preferred method. | |
a0a1a5fd | 5472 | * |
6ba94429 FW |
5473 | * Workqueue user should use this function directly iff it wants to apply |
5474 | * workqueue_attrs before making the workqueue visible in sysfs; otherwise, | |
5475 | * apply_workqueue_attrs() may race against userland updating the | |
5476 | * attributes. | |
5477 | * | |
5478 | * Return: 0 on success, -errno on failure. | |
a0a1a5fd | 5479 | */ |
6ba94429 | 5480 | int workqueue_sysfs_register(struct workqueue_struct *wq) |
a0a1a5fd | 5481 | { |
6ba94429 FW |
5482 | struct wq_device *wq_dev; |
5483 | int ret; | |
a0a1a5fd | 5484 | |
6ba94429 | 5485 | /* |
402dd89d | 5486 | * Adjusting max_active or creating new pwqs by applying |
6ba94429 FW |
5487 | * attributes breaks ordering guarantee. Disallow exposing ordered |
5488 | * workqueues. | |
5489 | */ | |
0a94efb5 | 5490 | if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) |
6ba94429 | 5491 | return -EINVAL; |
a0a1a5fd | 5492 | |
6ba94429 FW |
5493 | wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); |
5494 | if (!wq_dev) | |
5495 | return -ENOMEM; | |
5bcab335 | 5496 | |
6ba94429 FW |
5497 | wq_dev->wq = wq; |
5498 | wq_dev->dev.bus = &wq_subsys; | |
6ba94429 | 5499 | wq_dev->dev.release = wq_device_release; |
23217b44 | 5500 | dev_set_name(&wq_dev->dev, "%s", wq->name); |
a0a1a5fd | 5501 | |
6ba94429 FW |
5502 | /* |
5503 | * unbound_attrs are created separately. Suppress uevent until | |
5504 | * everything is ready. | |
5505 | */ | |
5506 | dev_set_uevent_suppress(&wq_dev->dev, true); | |
a0a1a5fd | 5507 | |
6ba94429 FW |
5508 | ret = device_register(&wq_dev->dev); |
5509 | if (ret) { | |
537f4146 | 5510 | put_device(&wq_dev->dev); |
6ba94429 FW |
5511 | wq->wq_dev = NULL; |
5512 | return ret; | |
5513 | } | |
a0a1a5fd | 5514 | |
6ba94429 FW |
5515 | if (wq->flags & WQ_UNBOUND) { |
5516 | struct device_attribute *attr; | |
a0a1a5fd | 5517 | |
6ba94429 FW |
5518 | for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { |
5519 | ret = device_create_file(&wq_dev->dev, attr); | |
5520 | if (ret) { | |
5521 | device_unregister(&wq_dev->dev); | |
5522 | wq->wq_dev = NULL; | |
5523 | return ret; | |
a0a1a5fd TH |
5524 | } |
5525 | } | |
5526 | } | |
6ba94429 FW |
5527 | |
5528 | dev_set_uevent_suppress(&wq_dev->dev, false); | |
5529 | kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); | |
5530 | return 0; | |
a0a1a5fd TH |
5531 | } |
5532 | ||
5533 | /** | |
6ba94429 FW |
5534 | * workqueue_sysfs_unregister - undo workqueue_sysfs_register() |
5535 | * @wq: the workqueue to unregister | |
a0a1a5fd | 5536 | * |
6ba94429 | 5537 | * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. |
a0a1a5fd | 5538 | */ |
6ba94429 | 5539 | static void workqueue_sysfs_unregister(struct workqueue_struct *wq) |
a0a1a5fd | 5540 | { |
6ba94429 | 5541 | struct wq_device *wq_dev = wq->wq_dev; |
8b03ae3c | 5542 | |
6ba94429 FW |
5543 | if (!wq->wq_dev) |
5544 | return; | |
a0a1a5fd | 5545 | |
6ba94429 FW |
5546 | wq->wq_dev = NULL; |
5547 | device_unregister(&wq_dev->dev); | |
a0a1a5fd | 5548 | } |
6ba94429 FW |
5549 | #else /* CONFIG_SYSFS */ |
5550 | static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } | |
5551 | #endif /* CONFIG_SYSFS */ | |
a0a1a5fd | 5552 | |
82607adc TH |
5553 | /* |
5554 | * Workqueue watchdog. | |
5555 | * | |
5556 | * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal | |
5557 | * flush dependency, a concurrency managed work item which stays RUNNING | |
5558 | * indefinitely. Workqueue stalls can be very difficult to debug as the | |
5559 | * usual warning mechanisms don't trigger and internal workqueue state is | |
5560 | * largely opaque. | |
5561 | * | |
5562 | * Workqueue watchdog monitors all worker pools periodically and dumps | |
5563 | * state if some pools failed to make forward progress for a while where | |
5564 | * forward progress is defined as the first item on ->worklist changing. | |
5565 | * | |
5566 | * This mechanism is controlled through the kernel parameter | |
5567 | * "workqueue.watchdog_thresh" which can be updated at runtime through the | |
5568 | * corresponding sysfs parameter file. | |
5569 | */ | |
5570 | #ifdef CONFIG_WQ_WATCHDOG | |
5571 | ||
82607adc | 5572 | static unsigned long wq_watchdog_thresh = 30; |
5cd79d6a | 5573 | static struct timer_list wq_watchdog_timer; |
82607adc TH |
5574 | |
5575 | static unsigned long wq_watchdog_touched = INITIAL_JIFFIES; | |
5576 | static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES; | |
5577 | ||
5578 | static void wq_watchdog_reset_touched(void) | |
5579 | { | |
5580 | int cpu; | |
5581 | ||
5582 | wq_watchdog_touched = jiffies; | |
5583 | for_each_possible_cpu(cpu) | |
5584 | per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; | |
5585 | } | |
5586 | ||
5cd79d6a | 5587 | static void wq_watchdog_timer_fn(struct timer_list *unused) |
82607adc TH |
5588 | { |
5589 | unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ; | |
5590 | bool lockup_detected = false; | |
5591 | struct worker_pool *pool; | |
5592 | int pi; | |
5593 | ||
5594 | if (!thresh) | |
5595 | return; | |
5596 | ||
5597 | rcu_read_lock(); | |
5598 | ||
5599 | for_each_pool(pool, pi) { | |
5600 | unsigned long pool_ts, touched, ts; | |
5601 | ||
5602 | if (list_empty(&pool->worklist)) | |
5603 | continue; | |
5604 | ||
5605 | /* get the latest of pool and touched timestamps */ | |
5606 | pool_ts = READ_ONCE(pool->watchdog_ts); | |
5607 | touched = READ_ONCE(wq_watchdog_touched); | |
5608 | ||
5609 | if (time_after(pool_ts, touched)) | |
5610 | ts = pool_ts; | |
5611 | else | |
5612 | ts = touched; | |
5613 | ||
5614 | if (pool->cpu >= 0) { | |
5615 | unsigned long cpu_touched = | |
5616 | READ_ONCE(per_cpu(wq_watchdog_touched_cpu, | |
5617 | pool->cpu)); | |
5618 | if (time_after(cpu_touched, ts)) | |
5619 | ts = cpu_touched; | |
5620 | } | |
5621 | ||
5622 | /* did we stall? */ | |
5623 | if (time_after(jiffies, ts + thresh)) { | |
5624 | lockup_detected = true; | |
5625 | pr_emerg("BUG: workqueue lockup - pool"); | |
5626 | pr_cont_pool_info(pool); | |
5627 | pr_cont(" stuck for %us!\n", | |
5628 | jiffies_to_msecs(jiffies - pool_ts) / 1000); | |
5629 | } | |
5630 | } | |
5631 | ||
5632 | rcu_read_unlock(); | |
5633 | ||
5634 | if (lockup_detected) | |
5635 | show_workqueue_state(); | |
5636 | ||
5637 | wq_watchdog_reset_touched(); | |
5638 | mod_timer(&wq_watchdog_timer, jiffies + thresh); | |
5639 | } | |
5640 | ||
cb9d7fd5 | 5641 | notrace void wq_watchdog_touch(int cpu) |
82607adc TH |
5642 | { |
5643 | if (cpu >= 0) | |
5644 | per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; | |
5645 | else | |
5646 | wq_watchdog_touched = jiffies; | |
5647 | } | |
5648 | ||
5649 | static void wq_watchdog_set_thresh(unsigned long thresh) | |
5650 | { | |
5651 | wq_watchdog_thresh = 0; | |
5652 | del_timer_sync(&wq_watchdog_timer); | |
5653 | ||
5654 | if (thresh) { | |
5655 | wq_watchdog_thresh = thresh; | |
5656 | wq_watchdog_reset_touched(); | |
5657 | mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ); | |
5658 | } | |
5659 | } | |
5660 | ||
5661 | static int wq_watchdog_param_set_thresh(const char *val, | |
5662 | const struct kernel_param *kp) | |
5663 | { | |
5664 | unsigned long thresh; | |
5665 | int ret; | |
5666 | ||
5667 | ret = kstrtoul(val, 0, &thresh); | |
5668 | if (ret) | |
5669 | return ret; | |
5670 | ||
5671 | if (system_wq) | |
5672 | wq_watchdog_set_thresh(thresh); | |
5673 | else | |
5674 | wq_watchdog_thresh = thresh; | |
5675 | ||
5676 | return 0; | |
5677 | } | |
5678 | ||
5679 | static const struct kernel_param_ops wq_watchdog_thresh_ops = { | |
5680 | .set = wq_watchdog_param_set_thresh, | |
5681 | .get = param_get_ulong, | |
5682 | }; | |
5683 | ||
5684 | module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh, | |
5685 | 0644); | |
5686 | ||
5687 | static void wq_watchdog_init(void) | |
5688 | { | |
5cd79d6a | 5689 | timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE); |
82607adc TH |
5690 | wq_watchdog_set_thresh(wq_watchdog_thresh); |
5691 | } | |
5692 | ||
5693 | #else /* CONFIG_WQ_WATCHDOG */ | |
5694 | ||
5695 | static inline void wq_watchdog_init(void) { } | |
5696 | ||
5697 | #endif /* CONFIG_WQ_WATCHDOG */ | |
5698 | ||
bce90380 TH |
5699 | static void __init wq_numa_init(void) |
5700 | { | |
5701 | cpumask_var_t *tbl; | |
5702 | int node, cpu; | |
5703 | ||
bce90380 TH |
5704 | if (num_possible_nodes() <= 1) |
5705 | return; | |
5706 | ||
d55262c4 TH |
5707 | if (wq_disable_numa) { |
5708 | pr_info("workqueue: NUMA affinity support disabled\n"); | |
5709 | return; | |
5710 | } | |
5711 | ||
4c16bd32 TH |
5712 | wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL); |
5713 | BUG_ON(!wq_update_unbound_numa_attrs_buf); | |
5714 | ||
bce90380 TH |
5715 | /* |
5716 | * We want masks of possible CPUs of each node which isn't readily | |
5717 | * available. Build one from cpu_to_node() which should have been | |
5718 | * fully initialized by now. | |
5719 | */ | |
6396bb22 | 5720 | tbl = kcalloc(nr_node_ids, sizeof(tbl[0]), GFP_KERNEL); |
bce90380 TH |
5721 | BUG_ON(!tbl); |
5722 | ||
5723 | for_each_node(node) | |
5a6024f1 | 5724 | BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL, |
1be0c25d | 5725 | node_online(node) ? node : NUMA_NO_NODE)); |
bce90380 TH |
5726 | |
5727 | for_each_possible_cpu(cpu) { | |
5728 | node = cpu_to_node(cpu); | |
5729 | if (WARN_ON(node == NUMA_NO_NODE)) { | |
5730 | pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu); | |
5731 | /* happens iff arch is bonkers, let's just proceed */ | |
5732 | return; | |
5733 | } | |
5734 | cpumask_set_cpu(cpu, tbl[node]); | |
5735 | } | |
5736 | ||
5737 | wq_numa_possible_cpumask = tbl; | |
5738 | wq_numa_enabled = true; | |
5739 | } | |
5740 | ||
3347fa09 TH |
5741 | /** |
5742 | * workqueue_init_early - early init for workqueue subsystem | |
5743 | * | |
5744 | * This is the first half of two-staged workqueue subsystem initialization | |
5745 | * and invoked as soon as the bare basics - memory allocation, cpumasks and | |
5746 | * idr are up. It sets up all the data structures and system workqueues | |
5747 | * and allows early boot code to create workqueues and queue/cancel work | |
5748 | * items. Actual work item execution starts only after kthreads can be | |
5749 | * created and scheduled right before early initcalls. | |
5750 | */ | |
5751 | int __init workqueue_init_early(void) | |
1da177e4 | 5752 | { |
7a4e344c | 5753 | int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; |
1bda3f80 | 5754 | int hk_flags = HK_FLAG_DOMAIN | HK_FLAG_WQ; |
7a4e344c | 5755 | int i, cpu; |
c34056a3 | 5756 | |
e904e6c2 TH |
5757 | WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); |
5758 | ||
b05a7928 | 5759 | BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL)); |
1bda3f80 | 5760 | cpumask_copy(wq_unbound_cpumask, housekeeping_cpumask(hk_flags)); |
b05a7928 | 5761 | |
e904e6c2 TH |
5762 | pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); |
5763 | ||
706026c2 | 5764 | /* initialize CPU pools */ |
29c91e99 | 5765 | for_each_possible_cpu(cpu) { |
4ce62e9e | 5766 | struct worker_pool *pool; |
8b03ae3c | 5767 | |
7a4e344c | 5768 | i = 0; |
f02ae73a | 5769 | for_each_cpu_worker_pool(pool, cpu) { |
7a4e344c | 5770 | BUG_ON(init_worker_pool(pool)); |
ec22ca5e | 5771 | pool->cpu = cpu; |
29c91e99 | 5772 | cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); |
7a4e344c | 5773 | pool->attrs->nice = std_nice[i++]; |
f3f90ad4 | 5774 | pool->node = cpu_to_node(cpu); |
7a4e344c | 5775 | |
9daf9e67 | 5776 | /* alloc pool ID */ |
68e13a67 | 5777 | mutex_lock(&wq_pool_mutex); |
9daf9e67 | 5778 | BUG_ON(worker_pool_assign_id(pool)); |
68e13a67 | 5779 | mutex_unlock(&wq_pool_mutex); |
4ce62e9e | 5780 | } |
8b03ae3c TH |
5781 | } |
5782 | ||
8a2b7538 | 5783 | /* create default unbound and ordered wq attrs */ |
29c91e99 TH |
5784 | for (i = 0; i < NR_STD_WORKER_POOLS; i++) { |
5785 | struct workqueue_attrs *attrs; | |
5786 | ||
5787 | BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); | |
29c91e99 | 5788 | attrs->nice = std_nice[i]; |
29c91e99 | 5789 | unbound_std_wq_attrs[i] = attrs; |
8a2b7538 TH |
5790 | |
5791 | /* | |
5792 | * An ordered wq should have only one pwq as ordering is | |
5793 | * guaranteed by max_active which is enforced by pwqs. | |
5794 | * Turn off NUMA so that dfl_pwq is used for all nodes. | |
5795 | */ | |
5796 | BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); | |
5797 | attrs->nice = std_nice[i]; | |
5798 | attrs->no_numa = true; | |
5799 | ordered_wq_attrs[i] = attrs; | |
29c91e99 TH |
5800 | } |
5801 | ||
d320c038 | 5802 | system_wq = alloc_workqueue("events", 0, 0); |
1aabe902 | 5803 | system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); |
d320c038 | 5804 | system_long_wq = alloc_workqueue("events_long", 0, 0); |
f3421797 TH |
5805 | system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, |
5806 | WQ_UNBOUND_MAX_ACTIVE); | |
24d51add TH |
5807 | system_freezable_wq = alloc_workqueue("events_freezable", |
5808 | WQ_FREEZABLE, 0); | |
0668106c VK |
5809 | system_power_efficient_wq = alloc_workqueue("events_power_efficient", |
5810 | WQ_POWER_EFFICIENT, 0); | |
5811 | system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient", | |
5812 | WQ_FREEZABLE | WQ_POWER_EFFICIENT, | |
5813 | 0); | |
1aabe902 | 5814 | BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || |
0668106c VK |
5815 | !system_unbound_wq || !system_freezable_wq || |
5816 | !system_power_efficient_wq || | |
5817 | !system_freezable_power_efficient_wq); | |
82607adc | 5818 | |
3347fa09 TH |
5819 | return 0; |
5820 | } | |
5821 | ||
5822 | /** | |
5823 | * workqueue_init - bring workqueue subsystem fully online | |
5824 | * | |
5825 | * This is the latter half of two-staged workqueue subsystem initialization | |
5826 | * and invoked as soon as kthreads can be created and scheduled. | |
5827 | * Workqueues have been created and work items queued on them, but there | |
5828 | * are no kworkers executing the work items yet. Populate the worker pools | |
5829 | * with the initial workers and enable future kworker creations. | |
5830 | */ | |
5831 | int __init workqueue_init(void) | |
5832 | { | |
2186d9f9 | 5833 | struct workqueue_struct *wq; |
3347fa09 TH |
5834 | struct worker_pool *pool; |
5835 | int cpu, bkt; | |
5836 | ||
2186d9f9 TH |
5837 | /* |
5838 | * It'd be simpler to initialize NUMA in workqueue_init_early() but | |
5839 | * CPU to node mapping may not be available that early on some | |
5840 | * archs such as power and arm64. As per-cpu pools created | |
5841 | * previously could be missing node hint and unbound pools NUMA | |
5842 | * affinity, fix them up. | |
40c17f75 TH |
5843 | * |
5844 | * Also, while iterating workqueues, create rescuers if requested. | |
2186d9f9 TH |
5845 | */ |
5846 | wq_numa_init(); | |
5847 | ||
5848 | mutex_lock(&wq_pool_mutex); | |
5849 | ||
5850 | for_each_possible_cpu(cpu) { | |
5851 | for_each_cpu_worker_pool(pool, cpu) { | |
5852 | pool->node = cpu_to_node(cpu); | |
5853 | } | |
5854 | } | |
5855 | ||
40c17f75 | 5856 | list_for_each_entry(wq, &workqueues, list) { |
2186d9f9 | 5857 | wq_update_unbound_numa(wq, smp_processor_id(), true); |
40c17f75 TH |
5858 | WARN(init_rescuer(wq), |
5859 | "workqueue: failed to create early rescuer for %s", | |
5860 | wq->name); | |
5861 | } | |
2186d9f9 TH |
5862 | |
5863 | mutex_unlock(&wq_pool_mutex); | |
5864 | ||
3347fa09 TH |
5865 | /* create the initial workers */ |
5866 | for_each_online_cpu(cpu) { | |
5867 | for_each_cpu_worker_pool(pool, cpu) { | |
5868 | pool->flags &= ~POOL_DISASSOCIATED; | |
5869 | BUG_ON(!create_worker(pool)); | |
5870 | } | |
5871 | } | |
5872 | ||
5873 | hash_for_each(unbound_pool_hash, bkt, pool, hash_node) | |
5874 | BUG_ON(!create_worker(pool)); | |
5875 | ||
5876 | wq_online = true; | |
82607adc TH |
5877 | wq_watchdog_init(); |
5878 | ||
6ee0578b | 5879 | return 0; |
1da177e4 | 5880 | } |