]>
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 TH |
23 | * |
24 | * Please read Documentation/workqueue.txt 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 PZ |
41 | #include <linux/kallsyms.h> |
42 | #include <linux/debug_locks.h> | |
4e6045f1 | 43 | #include <linux/lockdep.h> |
c34056a3 | 44 | #include <linux/idr.h> |
29c91e99 | 45 | #include <linux/jhash.h> |
42f8570f | 46 | #include <linux/hashtable.h> |
76af4d93 | 47 | #include <linux/rculist.h> |
bce90380 | 48 | #include <linux/nodemask.h> |
4c16bd32 | 49 | #include <linux/moduleparam.h> |
3d1cb205 | 50 | #include <linux/uaccess.h> |
e22bee78 | 51 | |
ea138446 | 52 | #include "workqueue_internal.h" |
1da177e4 | 53 | |
c8e55f36 | 54 | enum { |
24647570 TH |
55 | /* |
56 | * worker_pool flags | |
bc2ae0f5 | 57 | * |
24647570 | 58 | * A bound pool is either associated or disassociated with its CPU. |
bc2ae0f5 TH |
59 | * While associated (!DISASSOCIATED), all workers are bound to the |
60 | * CPU and none has %WORKER_UNBOUND set and concurrency management | |
61 | * is in effect. | |
62 | * | |
63 | * While DISASSOCIATED, the cpu may be offline and all workers have | |
64 | * %WORKER_UNBOUND set and concurrency management disabled, and may | |
24647570 | 65 | * be executing on any CPU. The pool behaves as an unbound one. |
bc2ae0f5 | 66 | * |
bc3a1afc | 67 | * Note that DISASSOCIATED should be flipped only while holding |
92f9c5c4 | 68 | * attach_mutex to avoid changing binding state while |
4736cbf7 | 69 | * worker_attach_to_pool() is in progress. |
bc2ae0f5 | 70 | */ |
24647570 | 71 | POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */ |
db7bccf4 | 72 | |
c8e55f36 | 73 | /* worker flags */ |
c8e55f36 TH |
74 | WORKER_DIE = 1 << 1, /* die die die */ |
75 | WORKER_IDLE = 1 << 2, /* is idle */ | |
e22bee78 | 76 | WORKER_PREP = 1 << 3, /* preparing to run works */ |
fb0e7beb | 77 | WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */ |
f3421797 | 78 | WORKER_UNBOUND = 1 << 7, /* worker is unbound */ |
a9ab775b | 79 | WORKER_REBOUND = 1 << 8, /* worker was rebound */ |
e22bee78 | 80 | |
a9ab775b TH |
81 | WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE | |
82 | WORKER_UNBOUND | WORKER_REBOUND, | |
db7bccf4 | 83 | |
e34cdddb | 84 | NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */ |
4ce62e9e | 85 | |
29c91e99 | 86 | UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */ |
c8e55f36 | 87 | BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */ |
db7bccf4 | 88 | |
e22bee78 TH |
89 | MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */ |
90 | IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */ | |
91 | ||
3233cdbd TH |
92 | MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2, |
93 | /* call for help after 10ms | |
94 | (min two ticks) */ | |
e22bee78 TH |
95 | MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */ |
96 | CREATE_COOLDOWN = HZ, /* time to breath after fail */ | |
e22bee78 TH |
97 | |
98 | /* | |
99 | * Rescue workers are used only on emergencies and shared by | |
8698a745 | 100 | * all cpus. Give MIN_NICE. |
e22bee78 | 101 | */ |
8698a745 DY |
102 | RESCUER_NICE_LEVEL = MIN_NICE, |
103 | HIGHPRI_NICE_LEVEL = MIN_NICE, | |
ecf6881f TH |
104 | |
105 | WQ_NAME_LEN = 24, | |
c8e55f36 | 106 | }; |
1da177e4 LT |
107 | |
108 | /* | |
4690c4ab TH |
109 | * Structure fields follow one of the following exclusion rules. |
110 | * | |
e41e704b TH |
111 | * I: Modifiable by initialization/destruction paths and read-only for |
112 | * everyone else. | |
4690c4ab | 113 | * |
e22bee78 TH |
114 | * P: Preemption protected. Disabling preemption is enough and should |
115 | * only be modified and accessed from the local cpu. | |
116 | * | |
d565ed63 | 117 | * L: pool->lock protected. Access with pool->lock held. |
4690c4ab | 118 | * |
d565ed63 TH |
119 | * X: During normal operation, modification requires pool->lock and should |
120 | * be done only from local cpu. Either disabling preemption on local | |
121 | * cpu or grabbing pool->lock is enough for read access. If | |
122 | * POOL_DISASSOCIATED is set, it's identical to L. | |
e22bee78 | 123 | * |
92f9c5c4 | 124 | * A: pool->attach_mutex protected. |
822d8405 | 125 | * |
68e13a67 | 126 | * PL: wq_pool_mutex protected. |
5bcab335 | 127 | * |
68e13a67 | 128 | * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads. |
76af4d93 | 129 | * |
5b95e1af LJ |
130 | * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads. |
131 | * | |
132 | * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or | |
133 | * sched-RCU for reads. | |
134 | * | |
3c25a55d LJ |
135 | * WQ: wq->mutex protected. |
136 | * | |
b5927605 | 137 | * WR: wq->mutex protected for writes. Sched-RCU protected for reads. |
2e109a28 TH |
138 | * |
139 | * MD: wq_mayday_lock protected. | |
1da177e4 | 140 | */ |
1da177e4 | 141 | |
2eaebdb3 | 142 | /* struct worker is defined in workqueue_internal.h */ |
c34056a3 | 143 | |
bd7bdd43 | 144 | struct worker_pool { |
d565ed63 | 145 | spinlock_t lock; /* the pool lock */ |
d84ff051 | 146 | int cpu; /* I: the associated cpu */ |
f3f90ad4 | 147 | int node; /* I: the associated node ID */ |
9daf9e67 | 148 | int id; /* I: pool ID */ |
11ebea50 | 149 | unsigned int flags; /* X: flags */ |
bd7bdd43 | 150 | |
82607adc TH |
151 | unsigned long watchdog_ts; /* L: watchdog timestamp */ |
152 | ||
bd7bdd43 TH |
153 | struct list_head worklist; /* L: list of pending works */ |
154 | int nr_workers; /* L: total number of workers */ | |
ea1abd61 LJ |
155 | |
156 | /* nr_idle includes the ones off idle_list for rebinding */ | |
bd7bdd43 TH |
157 | int nr_idle; /* L: currently idle ones */ |
158 | ||
159 | struct list_head idle_list; /* X: list of idle workers */ | |
160 | struct timer_list idle_timer; /* L: worker idle timeout */ | |
161 | struct timer_list mayday_timer; /* L: SOS timer for workers */ | |
162 | ||
c5aa87bb | 163 | /* a workers is either on busy_hash or idle_list, or the manager */ |
c9e7cf27 TH |
164 | DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER); |
165 | /* L: hash of busy workers */ | |
166 | ||
bc3a1afc | 167 | /* see manage_workers() for details on the two manager mutexes */ |
34a06bd6 | 168 | struct mutex manager_arb; /* manager arbitration */ |
2607d7a6 | 169 | struct worker *manager; /* L: purely informational */ |
92f9c5c4 LJ |
170 | struct mutex attach_mutex; /* attach/detach exclusion */ |
171 | struct list_head workers; /* A: attached workers */ | |
60f5a4bc | 172 | struct completion *detach_completion; /* all workers detached */ |
e19e397a | 173 | |
7cda9aae | 174 | struct ida worker_ida; /* worker IDs for task name */ |
e19e397a | 175 | |
7a4e344c | 176 | struct workqueue_attrs *attrs; /* I: worker attributes */ |
68e13a67 LJ |
177 | struct hlist_node hash_node; /* PL: unbound_pool_hash node */ |
178 | int refcnt; /* PL: refcnt for unbound pools */ | |
7a4e344c | 179 | |
e19e397a TH |
180 | /* |
181 | * The current concurrency level. As it's likely to be accessed | |
182 | * from other CPUs during try_to_wake_up(), put it in a separate | |
183 | * cacheline. | |
184 | */ | |
185 | atomic_t nr_running ____cacheline_aligned_in_smp; | |
29c91e99 TH |
186 | |
187 | /* | |
188 | * Destruction of pool is sched-RCU protected to allow dereferences | |
189 | * from get_work_pool(). | |
190 | */ | |
191 | struct rcu_head rcu; | |
8b03ae3c TH |
192 | } ____cacheline_aligned_in_smp; |
193 | ||
1da177e4 | 194 | /* |
112202d9 TH |
195 | * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS |
196 | * of work_struct->data are used for flags and the remaining high bits | |
197 | * point to the pwq; thus, pwqs need to be aligned at two's power of the | |
198 | * number of flag bits. | |
1da177e4 | 199 | */ |
112202d9 | 200 | struct pool_workqueue { |
bd7bdd43 | 201 | struct worker_pool *pool; /* I: the associated pool */ |
4690c4ab | 202 | struct workqueue_struct *wq; /* I: the owning workqueue */ |
73f53c4a TH |
203 | int work_color; /* L: current color */ |
204 | int flush_color; /* L: flushing color */ | |
8864b4e5 | 205 | int refcnt; /* L: reference count */ |
73f53c4a TH |
206 | int nr_in_flight[WORK_NR_COLORS]; |
207 | /* L: nr of in_flight works */ | |
1e19ffc6 | 208 | int nr_active; /* L: nr of active works */ |
a0a1a5fd | 209 | int max_active; /* L: max active works */ |
1e19ffc6 | 210 | struct list_head delayed_works; /* L: delayed works */ |
3c25a55d | 211 | struct list_head pwqs_node; /* WR: node on wq->pwqs */ |
2e109a28 | 212 | struct list_head mayday_node; /* MD: node on wq->maydays */ |
8864b4e5 TH |
213 | |
214 | /* | |
215 | * Release of unbound pwq is punted to system_wq. See put_pwq() | |
216 | * and pwq_unbound_release_workfn() for details. pool_workqueue | |
217 | * itself is also sched-RCU protected so that the first pwq can be | |
b09f4fd3 | 218 | * determined without grabbing wq->mutex. |
8864b4e5 TH |
219 | */ |
220 | struct work_struct unbound_release_work; | |
221 | struct rcu_head rcu; | |
e904e6c2 | 222 | } __aligned(1 << WORK_STRUCT_FLAG_BITS); |
1da177e4 | 223 | |
73f53c4a TH |
224 | /* |
225 | * Structure used to wait for workqueue flush. | |
226 | */ | |
227 | struct wq_flusher { | |
3c25a55d LJ |
228 | struct list_head list; /* WQ: list of flushers */ |
229 | int flush_color; /* WQ: flush color waiting for */ | |
73f53c4a TH |
230 | struct completion done; /* flush completion */ |
231 | }; | |
232 | ||
226223ab TH |
233 | struct wq_device; |
234 | ||
1da177e4 | 235 | /* |
c5aa87bb TH |
236 | * The externally visible workqueue. It relays the issued work items to |
237 | * the appropriate worker_pool through its pool_workqueues. | |
1da177e4 LT |
238 | */ |
239 | struct workqueue_struct { | |
3c25a55d | 240 | struct list_head pwqs; /* WR: all pwqs of this wq */ |
e2dca7ad | 241 | struct list_head list; /* PR: list of all workqueues */ |
73f53c4a | 242 | |
3c25a55d LJ |
243 | struct mutex mutex; /* protects this wq */ |
244 | int work_color; /* WQ: current work color */ | |
245 | int flush_color; /* WQ: current flush color */ | |
112202d9 | 246 | atomic_t nr_pwqs_to_flush; /* flush in progress */ |
3c25a55d LJ |
247 | struct wq_flusher *first_flusher; /* WQ: first flusher */ |
248 | struct list_head flusher_queue; /* WQ: flush waiters */ | |
249 | struct list_head flusher_overflow; /* WQ: flush overflow list */ | |
73f53c4a | 250 | |
2e109a28 | 251 | struct list_head maydays; /* MD: pwqs requesting rescue */ |
e22bee78 TH |
252 | struct worker *rescuer; /* I: rescue worker */ |
253 | ||
87fc741e | 254 | int nr_drainers; /* WQ: drain in progress */ |
a357fc03 | 255 | int saved_max_active; /* WQ: saved pwq max_active */ |
226223ab | 256 | |
5b95e1af LJ |
257 | struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */ |
258 | struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */ | |
6029a918 | 259 | |
226223ab TH |
260 | #ifdef CONFIG_SYSFS |
261 | struct wq_device *wq_dev; /* I: for sysfs interface */ | |
262 | #endif | |
4e6045f1 | 263 | #ifdef CONFIG_LOCKDEP |
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 | ||
bce90380 TH |
293 | static bool wq_numa_enabled; /* unbound NUMA affinity enabled */ |
294 | ||
4c16bd32 TH |
295 | /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */ |
296 | static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf; | |
297 | ||
68e13a67 | 298 | static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ |
2e109a28 | 299 | static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ |
5bcab335 | 300 | |
e2dca7ad | 301 | static LIST_HEAD(workqueues); /* PR: list of all workqueues */ |
68e13a67 | 302 | static bool workqueue_freezing; /* PL: have wqs started freezing? */ |
7d19c5ce | 303 | |
ef557180 MG |
304 | /* PL: allowable cpus for unbound wqs and work items */ |
305 | static cpumask_var_t wq_unbound_cpumask; | |
306 | ||
307 | /* CPU where unbound work was last round robin scheduled from this CPU */ | |
308 | static DEFINE_PER_CPU(int, wq_rr_cpu_last); | |
b05a7928 | 309 | |
f303fccb TH |
310 | /* |
311 | * Local execution of unbound work items is no longer guaranteed. The | |
312 | * following always forces round-robin CPU selection on unbound work items | |
313 | * to uncover usages which depend on it. | |
314 | */ | |
315 | #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU | |
316 | static bool wq_debug_force_rr_cpu = true; | |
317 | #else | |
318 | static bool wq_debug_force_rr_cpu = false; | |
319 | #endif | |
320 | module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644); | |
321 | ||
7d19c5ce | 322 | /* the per-cpu worker pools */ |
25528213 | 323 | static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools); |
7d19c5ce | 324 | |
68e13a67 | 325 | static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ |
7d19c5ce | 326 | |
68e13a67 | 327 | /* PL: hash of all unbound pools keyed by pool->attrs */ |
29c91e99 TH |
328 | static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); |
329 | ||
c5aa87bb | 330 | /* I: attributes used when instantiating standard unbound pools on demand */ |
29c91e99 TH |
331 | static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; |
332 | ||
8a2b7538 TH |
333 | /* I: attributes used when instantiating ordered pools on demand */ |
334 | static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS]; | |
335 | ||
d320c038 | 336 | struct workqueue_struct *system_wq __read_mostly; |
ad7b1f84 | 337 | EXPORT_SYMBOL(system_wq); |
044c782c | 338 | struct workqueue_struct *system_highpri_wq __read_mostly; |
1aabe902 | 339 | EXPORT_SYMBOL_GPL(system_highpri_wq); |
044c782c | 340 | struct workqueue_struct *system_long_wq __read_mostly; |
d320c038 | 341 | EXPORT_SYMBOL_GPL(system_long_wq); |
044c782c | 342 | struct workqueue_struct *system_unbound_wq __read_mostly; |
f3421797 | 343 | EXPORT_SYMBOL_GPL(system_unbound_wq); |
044c782c | 344 | struct workqueue_struct *system_freezable_wq __read_mostly; |
24d51add | 345 | EXPORT_SYMBOL_GPL(system_freezable_wq); |
0668106c VK |
346 | struct workqueue_struct *system_power_efficient_wq __read_mostly; |
347 | EXPORT_SYMBOL_GPL(system_power_efficient_wq); | |
348 | struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly; | |
349 | EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq); | |
d320c038 | 350 | |
7d19c5ce | 351 | static int worker_thread(void *__worker); |
6ba94429 | 352 | static void workqueue_sysfs_unregister(struct workqueue_struct *wq); |
7d19c5ce | 353 | |
97bd2347 TH |
354 | #define CREATE_TRACE_POINTS |
355 | #include <trace/events/workqueue.h> | |
356 | ||
68e13a67 | 357 | #define assert_rcu_or_pool_mutex() \ |
f78f5b90 PM |
358 | RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \ |
359 | !lockdep_is_held(&wq_pool_mutex), \ | |
360 | "sched RCU or wq_pool_mutex should be held") | |
5bcab335 | 361 | |
b09f4fd3 | 362 | #define assert_rcu_or_wq_mutex(wq) \ |
f78f5b90 PM |
363 | RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \ |
364 | !lockdep_is_held(&wq->mutex), \ | |
365 | "sched RCU or wq->mutex should be held") | |
76af4d93 | 366 | |
5b95e1af | 367 | #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \ |
f78f5b90 PM |
368 | RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \ |
369 | !lockdep_is_held(&wq->mutex) && \ | |
370 | !lockdep_is_held(&wq_pool_mutex), \ | |
371 | "sched RCU, wq->mutex or wq_pool_mutex should be held") | |
5b95e1af | 372 | |
f02ae73a TH |
373 | #define for_each_cpu_worker_pool(pool, cpu) \ |
374 | for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ | |
375 | (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ | |
7a62c2c8 | 376 | (pool)++) |
4ce62e9e | 377 | |
17116969 TH |
378 | /** |
379 | * for_each_pool - iterate through all worker_pools in the system | |
380 | * @pool: iteration cursor | |
611c92a0 | 381 | * @pi: integer used for iteration |
fa1b54e6 | 382 | * |
68e13a67 LJ |
383 | * This must be called either with wq_pool_mutex held or sched RCU read |
384 | * locked. If the pool needs to be used beyond the locking in effect, the | |
385 | * caller is responsible for guaranteeing that the pool stays online. | |
fa1b54e6 TH |
386 | * |
387 | * The if/else clause exists only for the lockdep assertion and can be | |
388 | * ignored. | |
17116969 | 389 | */ |
611c92a0 TH |
390 | #define for_each_pool(pool, pi) \ |
391 | idr_for_each_entry(&worker_pool_idr, pool, pi) \ | |
68e13a67 | 392 | if (({ assert_rcu_or_pool_mutex(); false; })) { } \ |
fa1b54e6 | 393 | else |
17116969 | 394 | |
822d8405 TH |
395 | /** |
396 | * for_each_pool_worker - iterate through all workers of a worker_pool | |
397 | * @worker: iteration cursor | |
822d8405 TH |
398 | * @pool: worker_pool to iterate workers of |
399 | * | |
92f9c5c4 | 400 | * This must be called with @pool->attach_mutex. |
822d8405 TH |
401 | * |
402 | * The if/else clause exists only for the lockdep assertion and can be | |
403 | * ignored. | |
404 | */ | |
da028469 LJ |
405 | #define for_each_pool_worker(worker, pool) \ |
406 | list_for_each_entry((worker), &(pool)->workers, node) \ | |
92f9c5c4 | 407 | if (({ lockdep_assert_held(&pool->attach_mutex); false; })) { } \ |
822d8405 TH |
408 | else |
409 | ||
49e3cf44 TH |
410 | /** |
411 | * for_each_pwq - iterate through all pool_workqueues of the specified workqueue | |
412 | * @pwq: iteration cursor | |
413 | * @wq: the target workqueue | |
76af4d93 | 414 | * |
b09f4fd3 | 415 | * This must be called either with wq->mutex held or sched RCU read locked. |
794b18bc TH |
416 | * If the pwq needs to be used beyond the locking in effect, the caller is |
417 | * responsible for guaranteeing that the pwq stays online. | |
76af4d93 TH |
418 | * |
419 | * The if/else clause exists only for the lockdep assertion and can be | |
420 | * ignored. | |
49e3cf44 TH |
421 | */ |
422 | #define for_each_pwq(pwq, wq) \ | |
76af4d93 | 423 | list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \ |
b09f4fd3 | 424 | if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \ |
76af4d93 | 425 | else |
f3421797 | 426 | |
dc186ad7 TG |
427 | #ifdef CONFIG_DEBUG_OBJECTS_WORK |
428 | ||
429 | static struct debug_obj_descr work_debug_descr; | |
430 | ||
99777288 SG |
431 | static void *work_debug_hint(void *addr) |
432 | { | |
433 | return ((struct work_struct *) addr)->func; | |
434 | } | |
435 | ||
dc186ad7 TG |
436 | /* |
437 | * fixup_init is called when: | |
438 | * - an active object is initialized | |
439 | */ | |
440 | static int work_fixup_init(void *addr, enum debug_obj_state state) | |
441 | { | |
442 | struct work_struct *work = addr; | |
443 | ||
444 | switch (state) { | |
445 | case ODEBUG_STATE_ACTIVE: | |
446 | cancel_work_sync(work); | |
447 | debug_object_init(work, &work_debug_descr); | |
448 | return 1; | |
449 | default: | |
450 | return 0; | |
451 | } | |
452 | } | |
453 | ||
454 | /* | |
455 | * fixup_activate is called when: | |
456 | * - an active object is activated | |
457 | * - an unknown object is activated (might be a statically initialized object) | |
458 | */ | |
459 | static int work_fixup_activate(void *addr, enum debug_obj_state state) | |
460 | { | |
461 | struct work_struct *work = addr; | |
462 | ||
463 | switch (state) { | |
464 | ||
465 | case ODEBUG_STATE_NOTAVAILABLE: | |
466 | /* | |
467 | * This is not really a fixup. The work struct was | |
468 | * statically initialized. We just make sure that it | |
469 | * is tracked in the object tracker. | |
470 | */ | |
22df02bb | 471 | if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) { |
dc186ad7 TG |
472 | debug_object_init(work, &work_debug_descr); |
473 | debug_object_activate(work, &work_debug_descr); | |
474 | return 0; | |
475 | } | |
476 | WARN_ON_ONCE(1); | |
477 | return 0; | |
478 | ||
479 | case ODEBUG_STATE_ACTIVE: | |
480 | WARN_ON(1); | |
481 | ||
482 | default: | |
483 | return 0; | |
484 | } | |
485 | } | |
486 | ||
487 | /* | |
488 | * fixup_free is called when: | |
489 | * - an active object is freed | |
490 | */ | |
491 | static int work_fixup_free(void *addr, enum debug_obj_state state) | |
492 | { | |
493 | struct work_struct *work = addr; | |
494 | ||
495 | switch (state) { | |
496 | case ODEBUG_STATE_ACTIVE: | |
497 | cancel_work_sync(work); | |
498 | debug_object_free(work, &work_debug_descr); | |
499 | return 1; | |
500 | default: | |
501 | return 0; | |
502 | } | |
503 | } | |
504 | ||
505 | static struct debug_obj_descr work_debug_descr = { | |
506 | .name = "work_struct", | |
99777288 | 507 | .debug_hint = work_debug_hint, |
dc186ad7 TG |
508 | .fixup_init = work_fixup_init, |
509 | .fixup_activate = work_fixup_activate, | |
510 | .fixup_free = work_fixup_free, | |
511 | }; | |
512 | ||
513 | static inline void debug_work_activate(struct work_struct *work) | |
514 | { | |
515 | debug_object_activate(work, &work_debug_descr); | |
516 | } | |
517 | ||
518 | static inline void debug_work_deactivate(struct work_struct *work) | |
519 | { | |
520 | debug_object_deactivate(work, &work_debug_descr); | |
521 | } | |
522 | ||
523 | void __init_work(struct work_struct *work, int onstack) | |
524 | { | |
525 | if (onstack) | |
526 | debug_object_init_on_stack(work, &work_debug_descr); | |
527 | else | |
528 | debug_object_init(work, &work_debug_descr); | |
529 | } | |
530 | EXPORT_SYMBOL_GPL(__init_work); | |
531 | ||
532 | void destroy_work_on_stack(struct work_struct *work) | |
533 | { | |
534 | debug_object_free(work, &work_debug_descr); | |
535 | } | |
536 | EXPORT_SYMBOL_GPL(destroy_work_on_stack); | |
537 | ||
ea2e64f2 TG |
538 | void destroy_delayed_work_on_stack(struct delayed_work *work) |
539 | { | |
540 | destroy_timer_on_stack(&work->timer); | |
541 | debug_object_free(&work->work, &work_debug_descr); | |
542 | } | |
543 | EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack); | |
544 | ||
dc186ad7 TG |
545 | #else |
546 | static inline void debug_work_activate(struct work_struct *work) { } | |
547 | static inline void debug_work_deactivate(struct work_struct *work) { } | |
548 | #endif | |
549 | ||
4e8b22bd LB |
550 | /** |
551 | * worker_pool_assign_id - allocate ID and assing it to @pool | |
552 | * @pool: the pool pointer of interest | |
553 | * | |
554 | * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned | |
555 | * successfully, -errno on failure. | |
556 | */ | |
9daf9e67 TH |
557 | static int worker_pool_assign_id(struct worker_pool *pool) |
558 | { | |
559 | int ret; | |
560 | ||
68e13a67 | 561 | lockdep_assert_held(&wq_pool_mutex); |
5bcab335 | 562 | |
4e8b22bd LB |
563 | ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE, |
564 | GFP_KERNEL); | |
229641a6 | 565 | if (ret >= 0) { |
e68035fb | 566 | pool->id = ret; |
229641a6 TH |
567 | return 0; |
568 | } | |
fa1b54e6 | 569 | return ret; |
7c3eed5c TH |
570 | } |
571 | ||
df2d5ae4 TH |
572 | /** |
573 | * unbound_pwq_by_node - return the unbound pool_workqueue for the given node | |
574 | * @wq: the target workqueue | |
575 | * @node: the node ID | |
576 | * | |
5b95e1af LJ |
577 | * This must be called with any of wq_pool_mutex, wq->mutex or sched RCU |
578 | * read locked. | |
df2d5ae4 TH |
579 | * If the pwq needs to be used beyond the locking in effect, the caller is |
580 | * responsible for guaranteeing that the pwq stays online. | |
d185af30 YB |
581 | * |
582 | * Return: The unbound pool_workqueue for @node. | |
df2d5ae4 TH |
583 | */ |
584 | static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq, | |
585 | int node) | |
586 | { | |
5b95e1af | 587 | assert_rcu_or_wq_mutex_or_pool_mutex(wq); |
d6e022f1 TH |
588 | |
589 | /* | |
590 | * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a | |
591 | * delayed item is pending. The plan is to keep CPU -> NODE | |
592 | * mapping valid and stable across CPU on/offlines. Once that | |
593 | * happens, this workaround can be removed. | |
594 | */ | |
595 | if (unlikely(node == NUMA_NO_NODE)) | |
596 | return wq->dfl_pwq; | |
597 | ||
df2d5ae4 TH |
598 | return rcu_dereference_raw(wq->numa_pwq_tbl[node]); |
599 | } | |
600 | ||
73f53c4a TH |
601 | static unsigned int work_color_to_flags(int color) |
602 | { | |
603 | return color << WORK_STRUCT_COLOR_SHIFT; | |
604 | } | |
605 | ||
606 | static int get_work_color(struct work_struct *work) | |
607 | { | |
608 | return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) & | |
609 | ((1 << WORK_STRUCT_COLOR_BITS) - 1); | |
610 | } | |
611 | ||
612 | static int work_next_color(int color) | |
613 | { | |
614 | return (color + 1) % WORK_NR_COLORS; | |
615 | } | |
1da177e4 | 616 | |
14441960 | 617 | /* |
112202d9 TH |
618 | * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data |
619 | * contain the pointer to the queued pwq. Once execution starts, the flag | |
7c3eed5c | 620 | * is cleared and the high bits contain OFFQ flags and pool ID. |
7a22ad75 | 621 | * |
112202d9 TH |
622 | * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() |
623 | * and clear_work_data() can be used to set the pwq, pool or clear | |
bbb68dfa TH |
624 | * work->data. These functions should only be called while the work is |
625 | * owned - ie. while the PENDING bit is set. | |
7a22ad75 | 626 | * |
112202d9 | 627 | * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq |
7c3eed5c | 628 | * corresponding to a work. Pool is available once the work has been |
112202d9 | 629 | * queued anywhere after initialization until it is sync canceled. pwq is |
7c3eed5c | 630 | * available only while the work item is queued. |
7a22ad75 | 631 | * |
bbb68dfa TH |
632 | * %WORK_OFFQ_CANCELING is used to mark a work item which is being |
633 | * canceled. While being canceled, a work item may have its PENDING set | |
634 | * but stay off timer and worklist for arbitrarily long and nobody should | |
635 | * try to steal the PENDING bit. | |
14441960 | 636 | */ |
7a22ad75 TH |
637 | static inline void set_work_data(struct work_struct *work, unsigned long data, |
638 | unsigned long flags) | |
365970a1 | 639 | { |
6183c009 | 640 | WARN_ON_ONCE(!work_pending(work)); |
7a22ad75 TH |
641 | atomic_long_set(&work->data, data | flags | work_static(work)); |
642 | } | |
365970a1 | 643 | |
112202d9 | 644 | static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, |
7a22ad75 TH |
645 | unsigned long extra_flags) |
646 | { | |
112202d9 TH |
647 | set_work_data(work, (unsigned long)pwq, |
648 | WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); | |
365970a1 DH |
649 | } |
650 | ||
4468a00f LJ |
651 | static void set_work_pool_and_keep_pending(struct work_struct *work, |
652 | int pool_id) | |
653 | { | |
654 | set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, | |
655 | WORK_STRUCT_PENDING); | |
656 | } | |
657 | ||
7c3eed5c TH |
658 | static void set_work_pool_and_clear_pending(struct work_struct *work, |
659 | int pool_id) | |
7a22ad75 | 660 | { |
23657bb1 TH |
661 | /* |
662 | * The following wmb is paired with the implied mb in | |
663 | * test_and_set_bit(PENDING) and ensures all updates to @work made | |
664 | * here are visible to and precede any updates by the next PENDING | |
665 | * owner. | |
666 | */ | |
667 | smp_wmb(); | |
7c3eed5c | 668 | set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); |
346c09f8 RP |
669 | /* |
670 | * The following mb guarantees that previous clear of a PENDING bit | |
671 | * will not be reordered with any speculative LOADS or STORES from | |
672 | * work->current_func, which is executed afterwards. This possible | |
673 | * reordering can lead to a missed execution on attempt to qeueue | |
674 | * the same @work. E.g. consider this case: | |
675 | * | |
676 | * CPU#0 CPU#1 | |
677 | * ---------------------------- -------------------------------- | |
678 | * | |
679 | * 1 STORE event_indicated | |
680 | * 2 queue_work_on() { | |
681 | * 3 test_and_set_bit(PENDING) | |
682 | * 4 } set_..._and_clear_pending() { | |
683 | * 5 set_work_data() # clear bit | |
684 | * 6 smp_mb() | |
685 | * 7 work->current_func() { | |
686 | * 8 LOAD event_indicated | |
687 | * } | |
688 | * | |
689 | * Without an explicit full barrier speculative LOAD on line 8 can | |
690 | * be executed before CPU#0 does STORE on line 1. If that happens, | |
691 | * CPU#0 observes the PENDING bit is still set and new execution of | |
692 | * a @work is not queued in a hope, that CPU#1 will eventually | |
693 | * finish the queued @work. Meanwhile CPU#1 does not see | |
694 | * event_indicated is set, because speculative LOAD was executed | |
695 | * before actual STORE. | |
696 | */ | |
697 | smp_mb(); | |
7a22ad75 | 698 | } |
f756d5e2 | 699 | |
7a22ad75 | 700 | static void clear_work_data(struct work_struct *work) |
1da177e4 | 701 | { |
7c3eed5c TH |
702 | smp_wmb(); /* see set_work_pool_and_clear_pending() */ |
703 | set_work_data(work, WORK_STRUCT_NO_POOL, 0); | |
1da177e4 LT |
704 | } |
705 | ||
112202d9 | 706 | static struct pool_workqueue *get_work_pwq(struct work_struct *work) |
b1f4ec17 | 707 | { |
e120153d | 708 | unsigned long data = atomic_long_read(&work->data); |
7a22ad75 | 709 | |
112202d9 | 710 | if (data & WORK_STRUCT_PWQ) |
e120153d TH |
711 | return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); |
712 | else | |
713 | return NULL; | |
4d707b9f ON |
714 | } |
715 | ||
7c3eed5c TH |
716 | /** |
717 | * get_work_pool - return the worker_pool a given work was associated with | |
718 | * @work: the work item of interest | |
719 | * | |
68e13a67 LJ |
720 | * Pools are created and destroyed under wq_pool_mutex, and allows read |
721 | * access under sched-RCU read lock. As such, this function should be | |
722 | * called under wq_pool_mutex or with preemption disabled. | |
fa1b54e6 TH |
723 | * |
724 | * All fields of the returned pool are accessible as long as the above | |
725 | * mentioned locking is in effect. If the returned pool needs to be used | |
726 | * beyond the critical section, the caller is responsible for ensuring the | |
727 | * returned pool is and stays online. | |
d185af30 YB |
728 | * |
729 | * Return: The worker_pool @work was last associated with. %NULL if none. | |
7c3eed5c TH |
730 | */ |
731 | static struct worker_pool *get_work_pool(struct work_struct *work) | |
365970a1 | 732 | { |
e120153d | 733 | unsigned long data = atomic_long_read(&work->data); |
7c3eed5c | 734 | int pool_id; |
7a22ad75 | 735 | |
68e13a67 | 736 | assert_rcu_or_pool_mutex(); |
fa1b54e6 | 737 | |
112202d9 TH |
738 | if (data & WORK_STRUCT_PWQ) |
739 | return ((struct pool_workqueue *) | |
7c3eed5c | 740 | (data & WORK_STRUCT_WQ_DATA_MASK))->pool; |
7a22ad75 | 741 | |
7c3eed5c TH |
742 | pool_id = data >> WORK_OFFQ_POOL_SHIFT; |
743 | if (pool_id == WORK_OFFQ_POOL_NONE) | |
7a22ad75 TH |
744 | return NULL; |
745 | ||
fa1b54e6 | 746 | return idr_find(&worker_pool_idr, pool_id); |
7c3eed5c TH |
747 | } |
748 | ||
749 | /** | |
750 | * get_work_pool_id - return the worker pool ID a given work is associated with | |
751 | * @work: the work item of interest | |
752 | * | |
d185af30 | 753 | * Return: The worker_pool ID @work was last associated with. |
7c3eed5c TH |
754 | * %WORK_OFFQ_POOL_NONE if none. |
755 | */ | |
756 | static int get_work_pool_id(struct work_struct *work) | |
757 | { | |
54d5b7d0 LJ |
758 | unsigned long data = atomic_long_read(&work->data); |
759 | ||
112202d9 TH |
760 | if (data & WORK_STRUCT_PWQ) |
761 | return ((struct pool_workqueue *) | |
54d5b7d0 | 762 | (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id; |
7c3eed5c | 763 | |
54d5b7d0 | 764 | return data >> WORK_OFFQ_POOL_SHIFT; |
7c3eed5c TH |
765 | } |
766 | ||
bbb68dfa TH |
767 | static void mark_work_canceling(struct work_struct *work) |
768 | { | |
7c3eed5c | 769 | unsigned long pool_id = get_work_pool_id(work); |
bbb68dfa | 770 | |
7c3eed5c TH |
771 | pool_id <<= WORK_OFFQ_POOL_SHIFT; |
772 | set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); | |
bbb68dfa TH |
773 | } |
774 | ||
775 | static bool work_is_canceling(struct work_struct *work) | |
776 | { | |
777 | unsigned long data = atomic_long_read(&work->data); | |
778 | ||
112202d9 | 779 | return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); |
bbb68dfa TH |
780 | } |
781 | ||
e22bee78 | 782 | /* |
3270476a TH |
783 | * Policy functions. These define the policies on how the global worker |
784 | * pools are managed. Unless noted otherwise, these functions assume that | |
d565ed63 | 785 | * they're being called with pool->lock held. |
e22bee78 TH |
786 | */ |
787 | ||
63d95a91 | 788 | static bool __need_more_worker(struct worker_pool *pool) |
a848e3b6 | 789 | { |
e19e397a | 790 | return !atomic_read(&pool->nr_running); |
a848e3b6 ON |
791 | } |
792 | ||
4594bf15 | 793 | /* |
e22bee78 TH |
794 | * Need to wake up a worker? Called from anything but currently |
795 | * running workers. | |
974271c4 TH |
796 | * |
797 | * Note that, because unbound workers never contribute to nr_running, this | |
706026c2 | 798 | * function will always return %true for unbound pools as long as the |
974271c4 | 799 | * worklist isn't empty. |
4594bf15 | 800 | */ |
63d95a91 | 801 | static bool need_more_worker(struct worker_pool *pool) |
365970a1 | 802 | { |
63d95a91 | 803 | return !list_empty(&pool->worklist) && __need_more_worker(pool); |
e22bee78 | 804 | } |
4594bf15 | 805 | |
e22bee78 | 806 | /* Can I start working? Called from busy but !running workers. */ |
63d95a91 | 807 | static bool may_start_working(struct worker_pool *pool) |
e22bee78 | 808 | { |
63d95a91 | 809 | return pool->nr_idle; |
e22bee78 TH |
810 | } |
811 | ||
812 | /* Do I need to keep working? Called from currently running workers. */ | |
63d95a91 | 813 | static bool keep_working(struct worker_pool *pool) |
e22bee78 | 814 | { |
e19e397a TH |
815 | return !list_empty(&pool->worklist) && |
816 | atomic_read(&pool->nr_running) <= 1; | |
e22bee78 TH |
817 | } |
818 | ||
819 | /* Do we need a new worker? Called from manager. */ | |
63d95a91 | 820 | static bool need_to_create_worker(struct worker_pool *pool) |
e22bee78 | 821 | { |
63d95a91 | 822 | return need_more_worker(pool) && !may_start_working(pool); |
e22bee78 | 823 | } |
365970a1 | 824 | |
e22bee78 | 825 | /* Do we have too many workers and should some go away? */ |
63d95a91 | 826 | static bool too_many_workers(struct worker_pool *pool) |
e22bee78 | 827 | { |
34a06bd6 | 828 | bool managing = mutex_is_locked(&pool->manager_arb); |
63d95a91 TH |
829 | int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ |
830 | int nr_busy = pool->nr_workers - nr_idle; | |
e22bee78 TH |
831 | |
832 | return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; | |
365970a1 DH |
833 | } |
834 | ||
4d707b9f | 835 | /* |
e22bee78 TH |
836 | * Wake up functions. |
837 | */ | |
838 | ||
1037de36 LJ |
839 | /* Return the first idle worker. Safe with preemption disabled */ |
840 | static struct worker *first_idle_worker(struct worker_pool *pool) | |
7e11629d | 841 | { |
63d95a91 | 842 | if (unlikely(list_empty(&pool->idle_list))) |
7e11629d TH |
843 | return NULL; |
844 | ||
63d95a91 | 845 | return list_first_entry(&pool->idle_list, struct worker, entry); |
7e11629d TH |
846 | } |
847 | ||
848 | /** | |
849 | * wake_up_worker - wake up an idle worker | |
63d95a91 | 850 | * @pool: worker pool to wake worker from |
7e11629d | 851 | * |
63d95a91 | 852 | * Wake up the first idle worker of @pool. |
7e11629d TH |
853 | * |
854 | * CONTEXT: | |
d565ed63 | 855 | * spin_lock_irq(pool->lock). |
7e11629d | 856 | */ |
63d95a91 | 857 | static void wake_up_worker(struct worker_pool *pool) |
7e11629d | 858 | { |
1037de36 | 859 | struct worker *worker = first_idle_worker(pool); |
7e11629d TH |
860 | |
861 | if (likely(worker)) | |
862 | wake_up_process(worker->task); | |
863 | } | |
864 | ||
d302f017 | 865 | /** |
e22bee78 TH |
866 | * wq_worker_waking_up - a worker is waking up |
867 | * @task: task waking up | |
868 | * @cpu: CPU @task is waking up to | |
869 | * | |
870 | * This function is called during try_to_wake_up() when a worker is | |
871 | * being awoken. | |
872 | * | |
873 | * CONTEXT: | |
874 | * spin_lock_irq(rq->lock) | |
875 | */ | |
d84ff051 | 876 | void wq_worker_waking_up(struct task_struct *task, int cpu) |
e22bee78 TH |
877 | { |
878 | struct worker *worker = kthread_data(task); | |
879 | ||
36576000 | 880 | if (!(worker->flags & WORKER_NOT_RUNNING)) { |
ec22ca5e | 881 | WARN_ON_ONCE(worker->pool->cpu != cpu); |
e19e397a | 882 | atomic_inc(&worker->pool->nr_running); |
36576000 | 883 | } |
e22bee78 TH |
884 | } |
885 | ||
886 | /** | |
887 | * wq_worker_sleeping - a worker is going to sleep | |
888 | * @task: task going to sleep | |
889 | * @cpu: CPU in question, must be the current CPU number | |
890 | * | |
891 | * This function is called during schedule() when a busy worker is | |
892 | * going to sleep. Worker on the same cpu can be woken up by | |
893 | * returning pointer to its task. | |
894 | * | |
895 | * CONTEXT: | |
896 | * spin_lock_irq(rq->lock) | |
897 | * | |
d185af30 | 898 | * Return: |
e22bee78 TH |
899 | * Worker task on @cpu to wake up, %NULL if none. |
900 | */ | |
d84ff051 | 901 | struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu) |
e22bee78 TH |
902 | { |
903 | struct worker *worker = kthread_data(task), *to_wakeup = NULL; | |
111c225a | 904 | struct worker_pool *pool; |
e22bee78 | 905 | |
111c225a TH |
906 | /* |
907 | * Rescuers, which may not have all the fields set up like normal | |
908 | * workers, also reach here, let's not access anything before | |
909 | * checking NOT_RUNNING. | |
910 | */ | |
2d64672e | 911 | if (worker->flags & WORKER_NOT_RUNNING) |
e22bee78 TH |
912 | return NULL; |
913 | ||
111c225a | 914 | pool = worker->pool; |
111c225a | 915 | |
e22bee78 | 916 | /* this can only happen on the local cpu */ |
92b69f50 | 917 | if (WARN_ON_ONCE(cpu != raw_smp_processor_id() || pool->cpu != cpu)) |
6183c009 | 918 | return NULL; |
e22bee78 TH |
919 | |
920 | /* | |
921 | * The counterpart of the following dec_and_test, implied mb, | |
922 | * worklist not empty test sequence is in insert_work(). | |
923 | * Please read comment there. | |
924 | * | |
628c78e7 TH |
925 | * NOT_RUNNING is clear. This means that we're bound to and |
926 | * running on the local cpu w/ rq lock held and preemption | |
927 | * disabled, which in turn means that none else could be | |
d565ed63 | 928 | * manipulating idle_list, so dereferencing idle_list without pool |
628c78e7 | 929 | * lock is safe. |
e22bee78 | 930 | */ |
e19e397a TH |
931 | if (atomic_dec_and_test(&pool->nr_running) && |
932 | !list_empty(&pool->worklist)) | |
1037de36 | 933 | to_wakeup = first_idle_worker(pool); |
e22bee78 TH |
934 | return to_wakeup ? to_wakeup->task : NULL; |
935 | } | |
936 | ||
937 | /** | |
938 | * worker_set_flags - set worker flags and adjust nr_running accordingly | |
cb444766 | 939 | * @worker: self |
d302f017 | 940 | * @flags: flags to set |
d302f017 | 941 | * |
228f1d00 | 942 | * Set @flags in @worker->flags and adjust nr_running accordingly. |
d302f017 | 943 | * |
cb444766 | 944 | * CONTEXT: |
d565ed63 | 945 | * spin_lock_irq(pool->lock) |
d302f017 | 946 | */ |
228f1d00 | 947 | static inline void worker_set_flags(struct worker *worker, unsigned int flags) |
d302f017 | 948 | { |
bd7bdd43 | 949 | struct worker_pool *pool = worker->pool; |
e22bee78 | 950 | |
cb444766 TH |
951 | WARN_ON_ONCE(worker->task != current); |
952 | ||
228f1d00 | 953 | /* If transitioning into NOT_RUNNING, adjust nr_running. */ |
e22bee78 TH |
954 | if ((flags & WORKER_NOT_RUNNING) && |
955 | !(worker->flags & WORKER_NOT_RUNNING)) { | |
228f1d00 | 956 | atomic_dec(&pool->nr_running); |
e22bee78 TH |
957 | } |
958 | ||
d302f017 TH |
959 | worker->flags |= flags; |
960 | } | |
961 | ||
962 | /** | |
e22bee78 | 963 | * worker_clr_flags - clear worker flags and adjust nr_running accordingly |
cb444766 | 964 | * @worker: self |
d302f017 TH |
965 | * @flags: flags to clear |
966 | * | |
e22bee78 | 967 | * Clear @flags in @worker->flags and adjust nr_running accordingly. |
d302f017 | 968 | * |
cb444766 | 969 | * CONTEXT: |
d565ed63 | 970 | * spin_lock_irq(pool->lock) |
d302f017 TH |
971 | */ |
972 | static inline void worker_clr_flags(struct worker *worker, unsigned int flags) | |
973 | { | |
63d95a91 | 974 | struct worker_pool *pool = worker->pool; |
e22bee78 TH |
975 | unsigned int oflags = worker->flags; |
976 | ||
cb444766 TH |
977 | WARN_ON_ONCE(worker->task != current); |
978 | ||
d302f017 | 979 | worker->flags &= ~flags; |
e22bee78 | 980 | |
42c025f3 TH |
981 | /* |
982 | * If transitioning out of NOT_RUNNING, increment nr_running. Note | |
983 | * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask | |
984 | * of multiple flags, not a single flag. | |
985 | */ | |
e22bee78 TH |
986 | if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) |
987 | if (!(worker->flags & WORKER_NOT_RUNNING)) | |
e19e397a | 988 | atomic_inc(&pool->nr_running); |
d302f017 TH |
989 | } |
990 | ||
8cca0eea TH |
991 | /** |
992 | * find_worker_executing_work - find worker which is executing a work | |
c9e7cf27 | 993 | * @pool: pool of interest |
8cca0eea TH |
994 | * @work: work to find worker for |
995 | * | |
c9e7cf27 TH |
996 | * Find a worker which is executing @work on @pool by searching |
997 | * @pool->busy_hash which is keyed by the address of @work. For a worker | |
a2c1c57b TH |
998 | * to match, its current execution should match the address of @work and |
999 | * its work function. This is to avoid unwanted dependency between | |
1000 | * unrelated work executions through a work item being recycled while still | |
1001 | * being executed. | |
1002 | * | |
1003 | * This is a bit tricky. A work item may be freed once its execution | |
1004 | * starts and nothing prevents the freed area from being recycled for | |
1005 | * another work item. If the same work item address ends up being reused | |
1006 | * before the original execution finishes, workqueue will identify the | |
1007 | * recycled work item as currently executing and make it wait until the | |
1008 | * current execution finishes, introducing an unwanted dependency. | |
1009 | * | |
c5aa87bb TH |
1010 | * This function checks the work item address and work function to avoid |
1011 | * false positives. Note that this isn't complete as one may construct a | |
1012 | * work function which can introduce dependency onto itself through a | |
1013 | * recycled work item. Well, if somebody wants to shoot oneself in the | |
1014 | * foot that badly, there's only so much we can do, and if such deadlock | |
1015 | * actually occurs, it should be easy to locate the culprit work function. | |
8cca0eea TH |
1016 | * |
1017 | * CONTEXT: | |
d565ed63 | 1018 | * spin_lock_irq(pool->lock). |
8cca0eea | 1019 | * |
d185af30 YB |
1020 | * Return: |
1021 | * Pointer to worker which is executing @work if found, %NULL | |
8cca0eea | 1022 | * otherwise. |
4d707b9f | 1023 | */ |
c9e7cf27 | 1024 | static struct worker *find_worker_executing_work(struct worker_pool *pool, |
8cca0eea | 1025 | struct work_struct *work) |
4d707b9f | 1026 | { |
42f8570f | 1027 | struct worker *worker; |
42f8570f | 1028 | |
b67bfe0d | 1029 | hash_for_each_possible(pool->busy_hash, worker, hentry, |
a2c1c57b TH |
1030 | (unsigned long)work) |
1031 | if (worker->current_work == work && | |
1032 | worker->current_func == work->func) | |
42f8570f SL |
1033 | return worker; |
1034 | ||
1035 | return NULL; | |
4d707b9f ON |
1036 | } |
1037 | ||
bf4ede01 TH |
1038 | /** |
1039 | * move_linked_works - move linked works to a list | |
1040 | * @work: start of series of works to be scheduled | |
1041 | * @head: target list to append @work to | |
402dd89d | 1042 | * @nextp: out parameter for nested worklist walking |
bf4ede01 TH |
1043 | * |
1044 | * Schedule linked works starting from @work to @head. Work series to | |
1045 | * be scheduled starts at @work and includes any consecutive work with | |
1046 | * WORK_STRUCT_LINKED set in its predecessor. | |
1047 | * | |
1048 | * If @nextp is not NULL, it's updated to point to the next work of | |
1049 | * the last scheduled work. This allows move_linked_works() to be | |
1050 | * nested inside outer list_for_each_entry_safe(). | |
1051 | * | |
1052 | * CONTEXT: | |
d565ed63 | 1053 | * spin_lock_irq(pool->lock). |
bf4ede01 TH |
1054 | */ |
1055 | static void move_linked_works(struct work_struct *work, struct list_head *head, | |
1056 | struct work_struct **nextp) | |
1057 | { | |
1058 | struct work_struct *n; | |
1059 | ||
1060 | /* | |
1061 | * Linked worklist will always end before the end of the list, | |
1062 | * use NULL for list head. | |
1063 | */ | |
1064 | list_for_each_entry_safe_from(work, n, NULL, entry) { | |
1065 | list_move_tail(&work->entry, head); | |
1066 | if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) | |
1067 | break; | |
1068 | } | |
1069 | ||
1070 | /* | |
1071 | * If we're already inside safe list traversal and have moved | |
1072 | * multiple works to the scheduled queue, the next position | |
1073 | * needs to be updated. | |
1074 | */ | |
1075 | if (nextp) | |
1076 | *nextp = n; | |
1077 | } | |
1078 | ||
8864b4e5 TH |
1079 | /** |
1080 | * get_pwq - get an extra reference on the specified pool_workqueue | |
1081 | * @pwq: pool_workqueue to get | |
1082 | * | |
1083 | * Obtain an extra reference on @pwq. The caller should guarantee that | |
1084 | * @pwq has positive refcnt and be holding the matching pool->lock. | |
1085 | */ | |
1086 | static void get_pwq(struct pool_workqueue *pwq) | |
1087 | { | |
1088 | lockdep_assert_held(&pwq->pool->lock); | |
1089 | WARN_ON_ONCE(pwq->refcnt <= 0); | |
1090 | pwq->refcnt++; | |
1091 | } | |
1092 | ||
1093 | /** | |
1094 | * put_pwq - put a pool_workqueue reference | |
1095 | * @pwq: pool_workqueue to put | |
1096 | * | |
1097 | * Drop a reference of @pwq. If its refcnt reaches zero, schedule its | |
1098 | * destruction. The caller should be holding the matching pool->lock. | |
1099 | */ | |
1100 | static void put_pwq(struct pool_workqueue *pwq) | |
1101 | { | |
1102 | lockdep_assert_held(&pwq->pool->lock); | |
1103 | if (likely(--pwq->refcnt)) | |
1104 | return; | |
1105 | if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND))) | |
1106 | return; | |
1107 | /* | |
1108 | * @pwq can't be released under pool->lock, bounce to | |
1109 | * pwq_unbound_release_workfn(). This never recurses on the same | |
1110 | * pool->lock as this path is taken only for unbound workqueues and | |
1111 | * the release work item is scheduled on a per-cpu workqueue. To | |
1112 | * avoid lockdep warning, unbound pool->locks are given lockdep | |
1113 | * subclass of 1 in get_unbound_pool(). | |
1114 | */ | |
1115 | schedule_work(&pwq->unbound_release_work); | |
1116 | } | |
1117 | ||
dce90d47 TH |
1118 | /** |
1119 | * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock | |
1120 | * @pwq: pool_workqueue to put (can be %NULL) | |
1121 | * | |
1122 | * put_pwq() with locking. This function also allows %NULL @pwq. | |
1123 | */ | |
1124 | static void put_pwq_unlocked(struct pool_workqueue *pwq) | |
1125 | { | |
1126 | if (pwq) { | |
1127 | /* | |
1128 | * As both pwqs and pools are sched-RCU protected, the | |
1129 | * following lock operations are safe. | |
1130 | */ | |
1131 | spin_lock_irq(&pwq->pool->lock); | |
1132 | put_pwq(pwq); | |
1133 | spin_unlock_irq(&pwq->pool->lock); | |
1134 | } | |
1135 | } | |
1136 | ||
112202d9 | 1137 | static void pwq_activate_delayed_work(struct work_struct *work) |
bf4ede01 | 1138 | { |
112202d9 | 1139 | struct pool_workqueue *pwq = get_work_pwq(work); |
bf4ede01 TH |
1140 | |
1141 | trace_workqueue_activate_work(work); | |
82607adc TH |
1142 | if (list_empty(&pwq->pool->worklist)) |
1143 | pwq->pool->watchdog_ts = jiffies; | |
112202d9 | 1144 | move_linked_works(work, &pwq->pool->worklist, NULL); |
bf4ede01 | 1145 | __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work)); |
112202d9 | 1146 | pwq->nr_active++; |
bf4ede01 TH |
1147 | } |
1148 | ||
112202d9 | 1149 | static void pwq_activate_first_delayed(struct pool_workqueue *pwq) |
3aa62497 | 1150 | { |
112202d9 | 1151 | struct work_struct *work = list_first_entry(&pwq->delayed_works, |
3aa62497 LJ |
1152 | struct work_struct, entry); |
1153 | ||
112202d9 | 1154 | pwq_activate_delayed_work(work); |
3aa62497 LJ |
1155 | } |
1156 | ||
bf4ede01 | 1157 | /** |
112202d9 TH |
1158 | * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight |
1159 | * @pwq: pwq of interest | |
bf4ede01 | 1160 | * @color: color of work which left the queue |
bf4ede01 TH |
1161 | * |
1162 | * A work either has completed or is removed from pending queue, | |
112202d9 | 1163 | * decrement nr_in_flight of its pwq and handle workqueue flushing. |
bf4ede01 TH |
1164 | * |
1165 | * CONTEXT: | |
d565ed63 | 1166 | * spin_lock_irq(pool->lock). |
bf4ede01 | 1167 | */ |
112202d9 | 1168 | static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color) |
bf4ede01 | 1169 | { |
8864b4e5 | 1170 | /* uncolored work items don't participate in flushing or nr_active */ |
bf4ede01 | 1171 | if (color == WORK_NO_COLOR) |
8864b4e5 | 1172 | goto out_put; |
bf4ede01 | 1173 | |
112202d9 | 1174 | pwq->nr_in_flight[color]--; |
bf4ede01 | 1175 | |
112202d9 TH |
1176 | pwq->nr_active--; |
1177 | if (!list_empty(&pwq->delayed_works)) { | |
b3f9f405 | 1178 | /* one down, submit a delayed one */ |
112202d9 TH |
1179 | if (pwq->nr_active < pwq->max_active) |
1180 | pwq_activate_first_delayed(pwq); | |
bf4ede01 TH |
1181 | } |
1182 | ||
1183 | /* is flush in progress and are we at the flushing tip? */ | |
112202d9 | 1184 | if (likely(pwq->flush_color != color)) |
8864b4e5 | 1185 | goto out_put; |
bf4ede01 TH |
1186 | |
1187 | /* are there still in-flight works? */ | |
112202d9 | 1188 | if (pwq->nr_in_flight[color]) |
8864b4e5 | 1189 | goto out_put; |
bf4ede01 | 1190 | |
112202d9 TH |
1191 | /* this pwq is done, clear flush_color */ |
1192 | pwq->flush_color = -1; | |
bf4ede01 TH |
1193 | |
1194 | /* | |
112202d9 | 1195 | * If this was the last pwq, wake up the first flusher. It |
bf4ede01 TH |
1196 | * will handle the rest. |
1197 | */ | |
112202d9 TH |
1198 | if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) |
1199 | complete(&pwq->wq->first_flusher->done); | |
8864b4e5 TH |
1200 | out_put: |
1201 | put_pwq(pwq); | |
bf4ede01 TH |
1202 | } |
1203 | ||
36e227d2 | 1204 | /** |
bbb68dfa | 1205 | * try_to_grab_pending - steal work item from worklist and disable irq |
36e227d2 TH |
1206 | * @work: work item to steal |
1207 | * @is_dwork: @work is a delayed_work | |
bbb68dfa | 1208 | * @flags: place to store irq state |
36e227d2 TH |
1209 | * |
1210 | * Try to grab PENDING bit of @work. This function can handle @work in any | |
d185af30 | 1211 | * stable state - idle, on timer or on worklist. |
36e227d2 | 1212 | * |
d185af30 | 1213 | * Return: |
36e227d2 TH |
1214 | * 1 if @work was pending and we successfully stole PENDING |
1215 | * 0 if @work was idle and we claimed PENDING | |
1216 | * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry | |
bbb68dfa TH |
1217 | * -ENOENT if someone else is canceling @work, this state may persist |
1218 | * for arbitrarily long | |
36e227d2 | 1219 | * |
d185af30 | 1220 | * Note: |
bbb68dfa | 1221 | * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting |
e0aecdd8 TH |
1222 | * interrupted while holding PENDING and @work off queue, irq must be |
1223 | * disabled on entry. This, combined with delayed_work->timer being | |
1224 | * irqsafe, ensures that we return -EAGAIN for finite short period of time. | |
bbb68dfa TH |
1225 | * |
1226 | * On successful return, >= 0, irq is disabled and the caller is | |
1227 | * responsible for releasing it using local_irq_restore(*@flags). | |
1228 | * | |
e0aecdd8 | 1229 | * This function is safe to call from any context including IRQ handler. |
bf4ede01 | 1230 | */ |
bbb68dfa TH |
1231 | static int try_to_grab_pending(struct work_struct *work, bool is_dwork, |
1232 | unsigned long *flags) | |
bf4ede01 | 1233 | { |
d565ed63 | 1234 | struct worker_pool *pool; |
112202d9 | 1235 | struct pool_workqueue *pwq; |
bf4ede01 | 1236 | |
bbb68dfa TH |
1237 | local_irq_save(*flags); |
1238 | ||
36e227d2 TH |
1239 | /* try to steal the timer if it exists */ |
1240 | if (is_dwork) { | |
1241 | struct delayed_work *dwork = to_delayed_work(work); | |
1242 | ||
e0aecdd8 TH |
1243 | /* |
1244 | * dwork->timer is irqsafe. If del_timer() fails, it's | |
1245 | * guaranteed that the timer is not queued anywhere and not | |
1246 | * running on the local CPU. | |
1247 | */ | |
36e227d2 TH |
1248 | if (likely(del_timer(&dwork->timer))) |
1249 | return 1; | |
1250 | } | |
1251 | ||
1252 | /* try to claim PENDING the normal way */ | |
bf4ede01 TH |
1253 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) |
1254 | return 0; | |
1255 | ||
1256 | /* | |
1257 | * The queueing is in progress, or it is already queued. Try to | |
1258 | * steal it from ->worklist without clearing WORK_STRUCT_PENDING. | |
1259 | */ | |
d565ed63 TH |
1260 | pool = get_work_pool(work); |
1261 | if (!pool) | |
bbb68dfa | 1262 | goto fail; |
bf4ede01 | 1263 | |
d565ed63 | 1264 | spin_lock(&pool->lock); |
0b3dae68 | 1265 | /* |
112202d9 TH |
1266 | * work->data is guaranteed to point to pwq only while the work |
1267 | * item is queued on pwq->wq, and both updating work->data to point | |
1268 | * to pwq on queueing and to pool on dequeueing are done under | |
1269 | * pwq->pool->lock. This in turn guarantees that, if work->data | |
1270 | * points to pwq which is associated with a locked pool, the work | |
0b3dae68 LJ |
1271 | * item is currently queued on that pool. |
1272 | */ | |
112202d9 TH |
1273 | pwq = get_work_pwq(work); |
1274 | if (pwq && pwq->pool == pool) { | |
16062836 TH |
1275 | debug_work_deactivate(work); |
1276 | ||
1277 | /* | |
1278 | * A delayed work item cannot be grabbed directly because | |
1279 | * it might have linked NO_COLOR work items which, if left | |
112202d9 | 1280 | * on the delayed_list, will confuse pwq->nr_active |
16062836 TH |
1281 | * management later on and cause stall. Make sure the work |
1282 | * item is activated before grabbing. | |
1283 | */ | |
1284 | if (*work_data_bits(work) & WORK_STRUCT_DELAYED) | |
112202d9 | 1285 | pwq_activate_delayed_work(work); |
16062836 TH |
1286 | |
1287 | list_del_init(&work->entry); | |
9c34a704 | 1288 | pwq_dec_nr_in_flight(pwq, get_work_color(work)); |
16062836 | 1289 | |
112202d9 | 1290 | /* work->data points to pwq iff queued, point to pool */ |
16062836 TH |
1291 | set_work_pool_and_keep_pending(work, pool->id); |
1292 | ||
1293 | spin_unlock(&pool->lock); | |
1294 | return 1; | |
bf4ede01 | 1295 | } |
d565ed63 | 1296 | spin_unlock(&pool->lock); |
bbb68dfa TH |
1297 | fail: |
1298 | local_irq_restore(*flags); | |
1299 | if (work_is_canceling(work)) | |
1300 | return -ENOENT; | |
1301 | cpu_relax(); | |
36e227d2 | 1302 | return -EAGAIN; |
bf4ede01 TH |
1303 | } |
1304 | ||
4690c4ab | 1305 | /** |
706026c2 | 1306 | * insert_work - insert a work into a pool |
112202d9 | 1307 | * @pwq: pwq @work belongs to |
4690c4ab TH |
1308 | * @work: work to insert |
1309 | * @head: insertion point | |
1310 | * @extra_flags: extra WORK_STRUCT_* flags to set | |
1311 | * | |
112202d9 | 1312 | * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to |
706026c2 | 1313 | * work_struct flags. |
4690c4ab TH |
1314 | * |
1315 | * CONTEXT: | |
d565ed63 | 1316 | * spin_lock_irq(pool->lock). |
4690c4ab | 1317 | */ |
112202d9 TH |
1318 | static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, |
1319 | struct list_head *head, unsigned int extra_flags) | |
b89deed3 | 1320 | { |
112202d9 | 1321 | struct worker_pool *pool = pwq->pool; |
e22bee78 | 1322 | |
4690c4ab | 1323 | /* we own @work, set data and link */ |
112202d9 | 1324 | set_work_pwq(work, pwq, extra_flags); |
1a4d9b0a | 1325 | list_add_tail(&work->entry, head); |
8864b4e5 | 1326 | get_pwq(pwq); |
e22bee78 TH |
1327 | |
1328 | /* | |
c5aa87bb TH |
1329 | * Ensure either wq_worker_sleeping() sees the above |
1330 | * list_add_tail() or we see zero nr_running to avoid workers lying | |
1331 | * around lazily while there are works to be processed. | |
e22bee78 TH |
1332 | */ |
1333 | smp_mb(); | |
1334 | ||
63d95a91 TH |
1335 | if (__need_more_worker(pool)) |
1336 | wake_up_worker(pool); | |
b89deed3 ON |
1337 | } |
1338 | ||
c8efcc25 TH |
1339 | /* |
1340 | * Test whether @work is being queued from another work executing on the | |
8d03ecfe | 1341 | * same workqueue. |
c8efcc25 TH |
1342 | */ |
1343 | static bool is_chained_work(struct workqueue_struct *wq) | |
1344 | { | |
8d03ecfe TH |
1345 | struct worker *worker; |
1346 | ||
1347 | worker = current_wq_worker(); | |
1348 | /* | |
1349 | * Return %true iff I'm a worker execuing a work item on @wq. If | |
1350 | * I'm @worker, it's safe to dereference it without locking. | |
1351 | */ | |
112202d9 | 1352 | return worker && worker->current_pwq->wq == wq; |
c8efcc25 TH |
1353 | } |
1354 | ||
ef557180 MG |
1355 | /* |
1356 | * When queueing an unbound work item to a wq, prefer local CPU if allowed | |
1357 | * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to | |
1358 | * avoid perturbing sensitive tasks. | |
1359 | */ | |
1360 | static int wq_select_unbound_cpu(int cpu) | |
1361 | { | |
f303fccb | 1362 | static bool printed_dbg_warning; |
ef557180 MG |
1363 | int new_cpu; |
1364 | ||
f303fccb TH |
1365 | if (likely(!wq_debug_force_rr_cpu)) { |
1366 | if (cpumask_test_cpu(cpu, wq_unbound_cpumask)) | |
1367 | return cpu; | |
1368 | } else if (!printed_dbg_warning) { | |
1369 | pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n"); | |
1370 | printed_dbg_warning = true; | |
1371 | } | |
1372 | ||
ef557180 MG |
1373 | if (cpumask_empty(wq_unbound_cpumask)) |
1374 | return cpu; | |
1375 | ||
1376 | new_cpu = __this_cpu_read(wq_rr_cpu_last); | |
1377 | new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask); | |
1378 | if (unlikely(new_cpu >= nr_cpu_ids)) { | |
1379 | new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask); | |
1380 | if (unlikely(new_cpu >= nr_cpu_ids)) | |
1381 | return cpu; | |
1382 | } | |
1383 | __this_cpu_write(wq_rr_cpu_last, new_cpu); | |
1384 | ||
1385 | return new_cpu; | |
1386 | } | |
1387 | ||
d84ff051 | 1388 | static void __queue_work(int cpu, struct workqueue_struct *wq, |
1da177e4 LT |
1389 | struct work_struct *work) |
1390 | { | |
112202d9 | 1391 | struct pool_workqueue *pwq; |
c9178087 | 1392 | struct worker_pool *last_pool; |
1e19ffc6 | 1393 | struct list_head *worklist; |
8a2e8e5d | 1394 | unsigned int work_flags; |
b75cac93 | 1395 | unsigned int req_cpu = cpu; |
8930caba TH |
1396 | |
1397 | /* | |
1398 | * While a work item is PENDING && off queue, a task trying to | |
1399 | * steal the PENDING will busy-loop waiting for it to either get | |
1400 | * queued or lose PENDING. Grabbing PENDING and queueing should | |
1401 | * happen with IRQ disabled. | |
1402 | */ | |
1403 | WARN_ON_ONCE(!irqs_disabled()); | |
1da177e4 | 1404 | |
dc186ad7 | 1405 | debug_work_activate(work); |
1e19ffc6 | 1406 | |
9ef28a73 | 1407 | /* if draining, only works from the same workqueue are allowed */ |
618b01eb | 1408 | if (unlikely(wq->flags & __WQ_DRAINING) && |
c8efcc25 | 1409 | WARN_ON_ONCE(!is_chained_work(wq))) |
e41e704b | 1410 | return; |
9e8cd2f5 | 1411 | retry: |
df2d5ae4 | 1412 | if (req_cpu == WORK_CPU_UNBOUND) |
ef557180 | 1413 | cpu = wq_select_unbound_cpu(raw_smp_processor_id()); |
df2d5ae4 | 1414 | |
c9178087 | 1415 | /* pwq which will be used unless @work is executing elsewhere */ |
df2d5ae4 | 1416 | if (!(wq->flags & WQ_UNBOUND)) |
7fb98ea7 | 1417 | pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); |
df2d5ae4 TH |
1418 | else |
1419 | pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); | |
dbf2576e | 1420 | |
c9178087 TH |
1421 | /* |
1422 | * If @work was previously on a different pool, it might still be | |
1423 | * running there, in which case the work needs to be queued on that | |
1424 | * pool to guarantee non-reentrancy. | |
1425 | */ | |
1426 | last_pool = get_work_pool(work); | |
1427 | if (last_pool && last_pool != pwq->pool) { | |
1428 | struct worker *worker; | |
18aa9eff | 1429 | |
c9178087 | 1430 | spin_lock(&last_pool->lock); |
18aa9eff | 1431 | |
c9178087 | 1432 | worker = find_worker_executing_work(last_pool, work); |
18aa9eff | 1433 | |
c9178087 TH |
1434 | if (worker && worker->current_pwq->wq == wq) { |
1435 | pwq = worker->current_pwq; | |
8930caba | 1436 | } else { |
c9178087 TH |
1437 | /* meh... not running there, queue here */ |
1438 | spin_unlock(&last_pool->lock); | |
112202d9 | 1439 | spin_lock(&pwq->pool->lock); |
8930caba | 1440 | } |
f3421797 | 1441 | } else { |
112202d9 | 1442 | spin_lock(&pwq->pool->lock); |
502ca9d8 TH |
1443 | } |
1444 | ||
9e8cd2f5 TH |
1445 | /* |
1446 | * pwq is determined and locked. For unbound pools, we could have | |
1447 | * raced with pwq release and it could already be dead. If its | |
1448 | * refcnt is zero, repeat pwq selection. Note that pwqs never die | |
df2d5ae4 TH |
1449 | * without another pwq replacing it in the numa_pwq_tbl or while |
1450 | * work items are executing on it, so the retrying is guaranteed to | |
9e8cd2f5 TH |
1451 | * make forward-progress. |
1452 | */ | |
1453 | if (unlikely(!pwq->refcnt)) { | |
1454 | if (wq->flags & WQ_UNBOUND) { | |
1455 | spin_unlock(&pwq->pool->lock); | |
1456 | cpu_relax(); | |
1457 | goto retry; | |
1458 | } | |
1459 | /* oops */ | |
1460 | WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", | |
1461 | wq->name, cpu); | |
1462 | } | |
1463 | ||
112202d9 TH |
1464 | /* pwq determined, queue */ |
1465 | trace_workqueue_queue_work(req_cpu, pwq, work); | |
502ca9d8 | 1466 | |
f5b2552b | 1467 | if (WARN_ON(!list_empty(&work->entry))) { |
112202d9 | 1468 | spin_unlock(&pwq->pool->lock); |
f5b2552b DC |
1469 | return; |
1470 | } | |
1e19ffc6 | 1471 | |
112202d9 TH |
1472 | pwq->nr_in_flight[pwq->work_color]++; |
1473 | work_flags = work_color_to_flags(pwq->work_color); | |
1e19ffc6 | 1474 | |
112202d9 | 1475 | if (likely(pwq->nr_active < pwq->max_active)) { |
cdadf009 | 1476 | trace_workqueue_activate_work(work); |
112202d9 TH |
1477 | pwq->nr_active++; |
1478 | worklist = &pwq->pool->worklist; | |
82607adc TH |
1479 | if (list_empty(worklist)) |
1480 | pwq->pool->watchdog_ts = jiffies; | |
8a2e8e5d TH |
1481 | } else { |
1482 | work_flags |= WORK_STRUCT_DELAYED; | |
112202d9 | 1483 | worklist = &pwq->delayed_works; |
8a2e8e5d | 1484 | } |
1e19ffc6 | 1485 | |
112202d9 | 1486 | insert_work(pwq, work, worklist, work_flags); |
1e19ffc6 | 1487 | |
112202d9 | 1488 | spin_unlock(&pwq->pool->lock); |
1da177e4 LT |
1489 | } |
1490 | ||
0fcb78c2 | 1491 | /** |
c1a220e7 ZR |
1492 | * queue_work_on - queue work on specific cpu |
1493 | * @cpu: CPU number to execute work on | |
0fcb78c2 REB |
1494 | * @wq: workqueue to use |
1495 | * @work: work to queue | |
1496 | * | |
c1a220e7 ZR |
1497 | * We queue the work to a specific CPU, the caller must ensure it |
1498 | * can't go away. | |
d185af30 YB |
1499 | * |
1500 | * Return: %false if @work was already on a queue, %true otherwise. | |
1da177e4 | 1501 | */ |
d4283e93 TH |
1502 | bool queue_work_on(int cpu, struct workqueue_struct *wq, |
1503 | struct work_struct *work) | |
1da177e4 | 1504 | { |
d4283e93 | 1505 | bool ret = false; |
8930caba | 1506 | unsigned long flags; |
ef1ca236 | 1507 | |
8930caba | 1508 | local_irq_save(flags); |
c1a220e7 | 1509 | |
22df02bb | 1510 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { |
4690c4ab | 1511 | __queue_work(cpu, wq, work); |
d4283e93 | 1512 | ret = true; |
c1a220e7 | 1513 | } |
ef1ca236 | 1514 | |
8930caba | 1515 | local_irq_restore(flags); |
1da177e4 LT |
1516 | return ret; |
1517 | } | |
ad7b1f84 | 1518 | EXPORT_SYMBOL(queue_work_on); |
1da177e4 | 1519 | |
d8e794df | 1520 | void delayed_work_timer_fn(unsigned long __data) |
1da177e4 | 1521 | { |
52bad64d | 1522 | struct delayed_work *dwork = (struct delayed_work *)__data; |
1da177e4 | 1523 | |
e0aecdd8 | 1524 | /* should have been called from irqsafe timer with irq already off */ |
60c057bc | 1525 | __queue_work(dwork->cpu, dwork->wq, &dwork->work); |
1da177e4 | 1526 | } |
1438ade5 | 1527 | EXPORT_SYMBOL(delayed_work_timer_fn); |
1da177e4 | 1528 | |
7beb2edf TH |
1529 | static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, |
1530 | struct delayed_work *dwork, unsigned long delay) | |
1da177e4 | 1531 | { |
7beb2edf TH |
1532 | struct timer_list *timer = &dwork->timer; |
1533 | struct work_struct *work = &dwork->work; | |
7beb2edf TH |
1534 | |
1535 | WARN_ON_ONCE(timer->function != delayed_work_timer_fn || | |
1536 | timer->data != (unsigned long)dwork); | |
fc4b514f TH |
1537 | WARN_ON_ONCE(timer_pending(timer)); |
1538 | WARN_ON_ONCE(!list_empty(&work->entry)); | |
7beb2edf | 1539 | |
8852aac2 TH |
1540 | /* |
1541 | * If @delay is 0, queue @dwork->work immediately. This is for | |
1542 | * both optimization and correctness. The earliest @timer can | |
1543 | * expire is on the closest next tick and delayed_work users depend | |
1544 | * on that there's no such delay when @delay is 0. | |
1545 | */ | |
1546 | if (!delay) { | |
1547 | __queue_work(cpu, wq, &dwork->work); | |
1548 | return; | |
1549 | } | |
1550 | ||
7beb2edf | 1551 | timer_stats_timer_set_start_info(&dwork->timer); |
1da177e4 | 1552 | |
60c057bc | 1553 | dwork->wq = wq; |
1265057f | 1554 | dwork->cpu = cpu; |
7beb2edf TH |
1555 | timer->expires = jiffies + delay; |
1556 | ||
041bd12e TH |
1557 | if (unlikely(cpu != WORK_CPU_UNBOUND)) |
1558 | add_timer_on(timer, cpu); | |
1559 | else | |
1560 | add_timer(timer); | |
1da177e4 LT |
1561 | } |
1562 | ||
0fcb78c2 REB |
1563 | /** |
1564 | * queue_delayed_work_on - queue work on specific CPU after delay | |
1565 | * @cpu: CPU number to execute work on | |
1566 | * @wq: workqueue to use | |
af9997e4 | 1567 | * @dwork: work to queue |
0fcb78c2 REB |
1568 | * @delay: number of jiffies to wait before queueing |
1569 | * | |
d185af30 | 1570 | * Return: %false if @work was already on a queue, %true otherwise. If |
715f1300 TH |
1571 | * @delay is zero and @dwork is idle, it will be scheduled for immediate |
1572 | * execution. | |
0fcb78c2 | 1573 | */ |
d4283e93 TH |
1574 | bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, |
1575 | struct delayed_work *dwork, unsigned long delay) | |
7a6bc1cd | 1576 | { |
52bad64d | 1577 | struct work_struct *work = &dwork->work; |
d4283e93 | 1578 | bool ret = false; |
8930caba | 1579 | unsigned long flags; |
7a6bc1cd | 1580 | |
8930caba TH |
1581 | /* read the comment in __queue_work() */ |
1582 | local_irq_save(flags); | |
7a6bc1cd | 1583 | |
22df02bb | 1584 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { |
7beb2edf | 1585 | __queue_delayed_work(cpu, wq, dwork, delay); |
d4283e93 | 1586 | ret = true; |
7a6bc1cd | 1587 | } |
8a3e77cc | 1588 | |
8930caba | 1589 | local_irq_restore(flags); |
7a6bc1cd VP |
1590 | return ret; |
1591 | } | |
ad7b1f84 | 1592 | EXPORT_SYMBOL(queue_delayed_work_on); |
c7fc77f7 | 1593 | |
8376fe22 TH |
1594 | /** |
1595 | * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU | |
1596 | * @cpu: CPU number to execute work on | |
1597 | * @wq: workqueue to use | |
1598 | * @dwork: work to queue | |
1599 | * @delay: number of jiffies to wait before queueing | |
1600 | * | |
1601 | * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, | |
1602 | * modify @dwork's timer so that it expires after @delay. If @delay is | |
1603 | * zero, @work is guaranteed to be scheduled immediately regardless of its | |
1604 | * current state. | |
1605 | * | |
d185af30 | 1606 | * Return: %false if @dwork was idle and queued, %true if @dwork was |
8376fe22 TH |
1607 | * pending and its timer was modified. |
1608 | * | |
e0aecdd8 | 1609 | * This function is safe to call from any context including IRQ handler. |
8376fe22 TH |
1610 | * See try_to_grab_pending() for details. |
1611 | */ | |
1612 | bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, | |
1613 | struct delayed_work *dwork, unsigned long delay) | |
1614 | { | |
1615 | unsigned long flags; | |
1616 | int ret; | |
c7fc77f7 | 1617 | |
8376fe22 TH |
1618 | do { |
1619 | ret = try_to_grab_pending(&dwork->work, true, &flags); | |
1620 | } while (unlikely(ret == -EAGAIN)); | |
63bc0362 | 1621 | |
8376fe22 TH |
1622 | if (likely(ret >= 0)) { |
1623 | __queue_delayed_work(cpu, wq, dwork, delay); | |
1624 | local_irq_restore(flags); | |
7a6bc1cd | 1625 | } |
8376fe22 TH |
1626 | |
1627 | /* -ENOENT from try_to_grab_pending() becomes %true */ | |
7a6bc1cd VP |
1628 | return ret; |
1629 | } | |
8376fe22 TH |
1630 | EXPORT_SYMBOL_GPL(mod_delayed_work_on); |
1631 | ||
c8e55f36 TH |
1632 | /** |
1633 | * worker_enter_idle - enter idle state | |
1634 | * @worker: worker which is entering idle state | |
1635 | * | |
1636 | * @worker is entering idle state. Update stats and idle timer if | |
1637 | * necessary. | |
1638 | * | |
1639 | * LOCKING: | |
d565ed63 | 1640 | * spin_lock_irq(pool->lock). |
c8e55f36 TH |
1641 | */ |
1642 | static void worker_enter_idle(struct worker *worker) | |
1da177e4 | 1643 | { |
bd7bdd43 | 1644 | struct worker_pool *pool = worker->pool; |
c8e55f36 | 1645 | |
6183c009 TH |
1646 | if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || |
1647 | WARN_ON_ONCE(!list_empty(&worker->entry) && | |
1648 | (worker->hentry.next || worker->hentry.pprev))) | |
1649 | return; | |
c8e55f36 | 1650 | |
051e1850 | 1651 | /* can't use worker_set_flags(), also called from create_worker() */ |
cb444766 | 1652 | worker->flags |= WORKER_IDLE; |
bd7bdd43 | 1653 | pool->nr_idle++; |
e22bee78 | 1654 | worker->last_active = jiffies; |
c8e55f36 TH |
1655 | |
1656 | /* idle_list is LIFO */ | |
bd7bdd43 | 1657 | list_add(&worker->entry, &pool->idle_list); |
db7bccf4 | 1658 | |
628c78e7 TH |
1659 | if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) |
1660 | mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); | |
cb444766 | 1661 | |
544ecf31 | 1662 | /* |
706026c2 | 1663 | * Sanity check nr_running. Because wq_unbind_fn() releases |
d565ed63 | 1664 | * pool->lock between setting %WORKER_UNBOUND and zapping |
628c78e7 TH |
1665 | * nr_running, the warning may trigger spuriously. Check iff |
1666 | * unbind is not in progress. | |
544ecf31 | 1667 | */ |
24647570 | 1668 | WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && |
bd7bdd43 | 1669 | pool->nr_workers == pool->nr_idle && |
e19e397a | 1670 | atomic_read(&pool->nr_running)); |
c8e55f36 TH |
1671 | } |
1672 | ||
1673 | /** | |
1674 | * worker_leave_idle - leave idle state | |
1675 | * @worker: worker which is leaving idle state | |
1676 | * | |
1677 | * @worker is leaving idle state. Update stats. | |
1678 | * | |
1679 | * LOCKING: | |
d565ed63 | 1680 | * spin_lock_irq(pool->lock). |
c8e55f36 TH |
1681 | */ |
1682 | static void worker_leave_idle(struct worker *worker) | |
1683 | { | |
bd7bdd43 | 1684 | struct worker_pool *pool = worker->pool; |
c8e55f36 | 1685 | |
6183c009 TH |
1686 | if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) |
1687 | return; | |
d302f017 | 1688 | worker_clr_flags(worker, WORKER_IDLE); |
bd7bdd43 | 1689 | pool->nr_idle--; |
c8e55f36 TH |
1690 | list_del_init(&worker->entry); |
1691 | } | |
1692 | ||
f7537df5 | 1693 | static struct worker *alloc_worker(int node) |
c34056a3 TH |
1694 | { |
1695 | struct worker *worker; | |
1696 | ||
f7537df5 | 1697 | worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node); |
c8e55f36 TH |
1698 | if (worker) { |
1699 | INIT_LIST_HEAD(&worker->entry); | |
affee4b2 | 1700 | INIT_LIST_HEAD(&worker->scheduled); |
da028469 | 1701 | INIT_LIST_HEAD(&worker->node); |
e22bee78 TH |
1702 | /* on creation a worker is in !idle && prep state */ |
1703 | worker->flags = WORKER_PREP; | |
c8e55f36 | 1704 | } |
c34056a3 TH |
1705 | return worker; |
1706 | } | |
1707 | ||
4736cbf7 LJ |
1708 | /** |
1709 | * worker_attach_to_pool() - attach a worker to a pool | |
1710 | * @worker: worker to be attached | |
1711 | * @pool: the target pool | |
1712 | * | |
1713 | * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and | |
1714 | * cpu-binding of @worker are kept coordinated with the pool across | |
1715 | * cpu-[un]hotplugs. | |
1716 | */ | |
1717 | static void worker_attach_to_pool(struct worker *worker, | |
1718 | struct worker_pool *pool) | |
1719 | { | |
1720 | mutex_lock(&pool->attach_mutex); | |
1721 | ||
1722 | /* | |
1723 | * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any | |
1724 | * online CPUs. It'll be re-applied when any of the CPUs come up. | |
1725 | */ | |
1726 | set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask); | |
1727 | ||
1728 | /* | |
1729 | * The pool->attach_mutex ensures %POOL_DISASSOCIATED remains | |
1730 | * stable across this function. See the comments above the | |
1731 | * flag definition for details. | |
1732 | */ | |
1733 | if (pool->flags & POOL_DISASSOCIATED) | |
1734 | worker->flags |= WORKER_UNBOUND; | |
1735 | ||
1736 | list_add_tail(&worker->node, &pool->workers); | |
1737 | ||
1738 | mutex_unlock(&pool->attach_mutex); | |
1739 | } | |
1740 | ||
60f5a4bc LJ |
1741 | /** |
1742 | * worker_detach_from_pool() - detach a worker from its pool | |
1743 | * @worker: worker which is attached to its pool | |
1744 | * @pool: the pool @worker is attached to | |
1745 | * | |
4736cbf7 LJ |
1746 | * Undo the attaching which had been done in worker_attach_to_pool(). The |
1747 | * caller worker shouldn't access to the pool after detached except it has | |
1748 | * other reference to the pool. | |
60f5a4bc LJ |
1749 | */ |
1750 | static void worker_detach_from_pool(struct worker *worker, | |
1751 | struct worker_pool *pool) | |
1752 | { | |
1753 | struct completion *detach_completion = NULL; | |
1754 | ||
92f9c5c4 | 1755 | mutex_lock(&pool->attach_mutex); |
da028469 LJ |
1756 | list_del(&worker->node); |
1757 | if (list_empty(&pool->workers)) | |
60f5a4bc | 1758 | detach_completion = pool->detach_completion; |
92f9c5c4 | 1759 | mutex_unlock(&pool->attach_mutex); |
60f5a4bc | 1760 | |
b62c0751 LJ |
1761 | /* clear leftover flags without pool->lock after it is detached */ |
1762 | worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND); | |
1763 | ||
60f5a4bc LJ |
1764 | if (detach_completion) |
1765 | complete(detach_completion); | |
1766 | } | |
1767 | ||
c34056a3 TH |
1768 | /** |
1769 | * create_worker - create a new workqueue worker | |
63d95a91 | 1770 | * @pool: pool the new worker will belong to |
c34056a3 | 1771 | * |
051e1850 | 1772 | * Create and start a new worker which is attached to @pool. |
c34056a3 TH |
1773 | * |
1774 | * CONTEXT: | |
1775 | * Might sleep. Does GFP_KERNEL allocations. | |
1776 | * | |
d185af30 | 1777 | * Return: |
c34056a3 TH |
1778 | * Pointer to the newly created worker. |
1779 | */ | |
bc2ae0f5 | 1780 | static struct worker *create_worker(struct worker_pool *pool) |
c34056a3 | 1781 | { |
c34056a3 | 1782 | struct worker *worker = NULL; |
f3421797 | 1783 | int id = -1; |
e3c916a4 | 1784 | char id_buf[16]; |
c34056a3 | 1785 | |
7cda9aae LJ |
1786 | /* ID is needed to determine kthread name */ |
1787 | id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL); | |
822d8405 TH |
1788 | if (id < 0) |
1789 | goto fail; | |
c34056a3 | 1790 | |
f7537df5 | 1791 | worker = alloc_worker(pool->node); |
c34056a3 TH |
1792 | if (!worker) |
1793 | goto fail; | |
1794 | ||
bd7bdd43 | 1795 | worker->pool = pool; |
c34056a3 TH |
1796 | worker->id = id; |
1797 | ||
29c91e99 | 1798 | if (pool->cpu >= 0) |
e3c916a4 TH |
1799 | snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, |
1800 | pool->attrs->nice < 0 ? "H" : ""); | |
f3421797 | 1801 | else |
e3c916a4 TH |
1802 | snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); |
1803 | ||
f3f90ad4 | 1804 | worker->task = kthread_create_on_node(worker_thread, worker, pool->node, |
e3c916a4 | 1805 | "kworker/%s", id_buf); |
c34056a3 TH |
1806 | if (IS_ERR(worker->task)) |
1807 | goto fail; | |
1808 | ||
91151228 | 1809 | set_user_nice(worker->task, pool->attrs->nice); |
25834c73 | 1810 | kthread_bind_mask(worker->task, pool->attrs->cpumask); |
91151228 | 1811 | |
da028469 | 1812 | /* successful, attach the worker to the pool */ |
4736cbf7 | 1813 | worker_attach_to_pool(worker, pool); |
822d8405 | 1814 | |
051e1850 LJ |
1815 | /* start the newly created worker */ |
1816 | spin_lock_irq(&pool->lock); | |
1817 | worker->pool->nr_workers++; | |
1818 | worker_enter_idle(worker); | |
1819 | wake_up_process(worker->task); | |
1820 | spin_unlock_irq(&pool->lock); | |
1821 | ||
c34056a3 | 1822 | return worker; |
822d8405 | 1823 | |
c34056a3 | 1824 | fail: |
9625ab17 | 1825 | if (id >= 0) |
7cda9aae | 1826 | ida_simple_remove(&pool->worker_ida, id); |
c34056a3 TH |
1827 | kfree(worker); |
1828 | return NULL; | |
1829 | } | |
1830 | ||
c34056a3 TH |
1831 | /** |
1832 | * destroy_worker - destroy a workqueue worker | |
1833 | * @worker: worker to be destroyed | |
1834 | * | |
73eb7fe7 LJ |
1835 | * Destroy @worker and adjust @pool stats accordingly. The worker should |
1836 | * be idle. | |
c8e55f36 TH |
1837 | * |
1838 | * CONTEXT: | |
60f5a4bc | 1839 | * spin_lock_irq(pool->lock). |
c34056a3 TH |
1840 | */ |
1841 | static void destroy_worker(struct worker *worker) | |
1842 | { | |
bd7bdd43 | 1843 | struct worker_pool *pool = worker->pool; |
c34056a3 | 1844 | |
cd549687 TH |
1845 | lockdep_assert_held(&pool->lock); |
1846 | ||
c34056a3 | 1847 | /* sanity check frenzy */ |
6183c009 | 1848 | if (WARN_ON(worker->current_work) || |
73eb7fe7 LJ |
1849 | WARN_ON(!list_empty(&worker->scheduled)) || |
1850 | WARN_ON(!(worker->flags & WORKER_IDLE))) | |
6183c009 | 1851 | return; |
c34056a3 | 1852 | |
73eb7fe7 LJ |
1853 | pool->nr_workers--; |
1854 | pool->nr_idle--; | |
5bdfff96 | 1855 | |
c8e55f36 | 1856 | list_del_init(&worker->entry); |
cb444766 | 1857 | worker->flags |= WORKER_DIE; |
60f5a4bc | 1858 | wake_up_process(worker->task); |
c34056a3 TH |
1859 | } |
1860 | ||
63d95a91 | 1861 | static void idle_worker_timeout(unsigned long __pool) |
e22bee78 | 1862 | { |
63d95a91 | 1863 | struct worker_pool *pool = (void *)__pool; |
e22bee78 | 1864 | |
d565ed63 | 1865 | spin_lock_irq(&pool->lock); |
e22bee78 | 1866 | |
3347fc9f | 1867 | while (too_many_workers(pool)) { |
e22bee78 TH |
1868 | struct worker *worker; |
1869 | unsigned long expires; | |
1870 | ||
1871 | /* idle_list is kept in LIFO order, check the last one */ | |
63d95a91 | 1872 | worker = list_entry(pool->idle_list.prev, struct worker, entry); |
e22bee78 TH |
1873 | expires = worker->last_active + IDLE_WORKER_TIMEOUT; |
1874 | ||
3347fc9f | 1875 | if (time_before(jiffies, expires)) { |
63d95a91 | 1876 | mod_timer(&pool->idle_timer, expires); |
3347fc9f | 1877 | break; |
d5abe669 | 1878 | } |
3347fc9f LJ |
1879 | |
1880 | destroy_worker(worker); | |
e22bee78 TH |
1881 | } |
1882 | ||
d565ed63 | 1883 | spin_unlock_irq(&pool->lock); |
e22bee78 | 1884 | } |
d5abe669 | 1885 | |
493a1724 | 1886 | static void send_mayday(struct work_struct *work) |
e22bee78 | 1887 | { |
112202d9 TH |
1888 | struct pool_workqueue *pwq = get_work_pwq(work); |
1889 | struct workqueue_struct *wq = pwq->wq; | |
493a1724 | 1890 | |
2e109a28 | 1891 | lockdep_assert_held(&wq_mayday_lock); |
e22bee78 | 1892 | |
493008a8 | 1893 | if (!wq->rescuer) |
493a1724 | 1894 | return; |
e22bee78 TH |
1895 | |
1896 | /* mayday mayday mayday */ | |
493a1724 | 1897 | if (list_empty(&pwq->mayday_node)) { |
77668c8b LJ |
1898 | /* |
1899 | * If @pwq is for an unbound wq, its base ref may be put at | |
1900 | * any time due to an attribute change. Pin @pwq until the | |
1901 | * rescuer is done with it. | |
1902 | */ | |
1903 | get_pwq(pwq); | |
493a1724 | 1904 | list_add_tail(&pwq->mayday_node, &wq->maydays); |
e22bee78 | 1905 | wake_up_process(wq->rescuer->task); |
493a1724 | 1906 | } |
e22bee78 TH |
1907 | } |
1908 | ||
706026c2 | 1909 | static void pool_mayday_timeout(unsigned long __pool) |
e22bee78 | 1910 | { |
63d95a91 | 1911 | struct worker_pool *pool = (void *)__pool; |
e22bee78 TH |
1912 | struct work_struct *work; |
1913 | ||
b2d82909 TH |
1914 | spin_lock_irq(&pool->lock); |
1915 | spin_lock(&wq_mayday_lock); /* for wq->maydays */ | |
e22bee78 | 1916 | |
63d95a91 | 1917 | if (need_to_create_worker(pool)) { |
e22bee78 TH |
1918 | /* |
1919 | * We've been trying to create a new worker but | |
1920 | * haven't been successful. We might be hitting an | |
1921 | * allocation deadlock. Send distress signals to | |
1922 | * rescuers. | |
1923 | */ | |
63d95a91 | 1924 | list_for_each_entry(work, &pool->worklist, entry) |
e22bee78 | 1925 | send_mayday(work); |
1da177e4 | 1926 | } |
e22bee78 | 1927 | |
b2d82909 TH |
1928 | spin_unlock(&wq_mayday_lock); |
1929 | spin_unlock_irq(&pool->lock); | |
e22bee78 | 1930 | |
63d95a91 | 1931 | mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); |
1da177e4 LT |
1932 | } |
1933 | ||
e22bee78 TH |
1934 | /** |
1935 | * maybe_create_worker - create a new worker if necessary | |
63d95a91 | 1936 | * @pool: pool to create a new worker for |
e22bee78 | 1937 | * |
63d95a91 | 1938 | * Create a new worker for @pool if necessary. @pool is guaranteed to |
e22bee78 TH |
1939 | * have at least one idle worker on return from this function. If |
1940 | * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is | |
63d95a91 | 1941 | * sent to all rescuers with works scheduled on @pool to resolve |
e22bee78 TH |
1942 | * possible allocation deadlock. |
1943 | * | |
c5aa87bb TH |
1944 | * On return, need_to_create_worker() is guaranteed to be %false and |
1945 | * may_start_working() %true. | |
e22bee78 TH |
1946 | * |
1947 | * LOCKING: | |
d565ed63 | 1948 | * spin_lock_irq(pool->lock) which may be released and regrabbed |
e22bee78 TH |
1949 | * multiple times. Does GFP_KERNEL allocations. Called only from |
1950 | * manager. | |
e22bee78 | 1951 | */ |
29187a9e | 1952 | static void maybe_create_worker(struct worker_pool *pool) |
d565ed63 TH |
1953 | __releases(&pool->lock) |
1954 | __acquires(&pool->lock) | |
1da177e4 | 1955 | { |
e22bee78 | 1956 | restart: |
d565ed63 | 1957 | spin_unlock_irq(&pool->lock); |
9f9c2364 | 1958 | |
e22bee78 | 1959 | /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ |
63d95a91 | 1960 | mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); |
e22bee78 TH |
1961 | |
1962 | while (true) { | |
051e1850 | 1963 | if (create_worker(pool) || !need_to_create_worker(pool)) |
e22bee78 | 1964 | break; |
1da177e4 | 1965 | |
e212f361 | 1966 | schedule_timeout_interruptible(CREATE_COOLDOWN); |
9f9c2364 | 1967 | |
63d95a91 | 1968 | if (!need_to_create_worker(pool)) |
e22bee78 TH |
1969 | break; |
1970 | } | |
1971 | ||
63d95a91 | 1972 | del_timer_sync(&pool->mayday_timer); |
d565ed63 | 1973 | spin_lock_irq(&pool->lock); |
051e1850 LJ |
1974 | /* |
1975 | * This is necessary even after a new worker was just successfully | |
1976 | * created as @pool->lock was dropped and the new worker might have | |
1977 | * already become busy. | |
1978 | */ | |
63d95a91 | 1979 | if (need_to_create_worker(pool)) |
e22bee78 | 1980 | goto restart; |
e22bee78 TH |
1981 | } |
1982 | ||
73f53c4a | 1983 | /** |
e22bee78 TH |
1984 | * manage_workers - manage worker pool |
1985 | * @worker: self | |
73f53c4a | 1986 | * |
706026c2 | 1987 | * Assume the manager role and manage the worker pool @worker belongs |
e22bee78 | 1988 | * to. At any given time, there can be only zero or one manager per |
706026c2 | 1989 | * pool. The exclusion is handled automatically by this function. |
e22bee78 TH |
1990 | * |
1991 | * The caller can safely start processing works on false return. On | |
1992 | * true return, it's guaranteed that need_to_create_worker() is false | |
1993 | * and may_start_working() is true. | |
73f53c4a TH |
1994 | * |
1995 | * CONTEXT: | |
d565ed63 | 1996 | * spin_lock_irq(pool->lock) which may be released and regrabbed |
e22bee78 TH |
1997 | * multiple times. Does GFP_KERNEL allocations. |
1998 | * | |
d185af30 | 1999 | * Return: |
29187a9e TH |
2000 | * %false if the pool doesn't need management and the caller can safely |
2001 | * start processing works, %true if management function was performed and | |
2002 | * the conditions that the caller verified before calling the function may | |
2003 | * no longer be true. | |
73f53c4a | 2004 | */ |
e22bee78 | 2005 | static bool manage_workers(struct worker *worker) |
73f53c4a | 2006 | { |
63d95a91 | 2007 | struct worker_pool *pool = worker->pool; |
73f53c4a | 2008 | |
bc3a1afc | 2009 | /* |
bc3a1afc TH |
2010 | * Anyone who successfully grabs manager_arb wins the arbitration |
2011 | * and becomes the manager. mutex_trylock() on pool->manager_arb | |
2012 | * failure while holding pool->lock reliably indicates that someone | |
2013 | * else is managing the pool and the worker which failed trylock | |
2014 | * can proceed to executing work items. This means that anyone | |
2015 | * grabbing manager_arb is responsible for actually performing | |
2016 | * manager duties. If manager_arb is grabbed and released without | |
2017 | * actual management, the pool may stall indefinitely. | |
bc3a1afc | 2018 | */ |
34a06bd6 | 2019 | if (!mutex_trylock(&pool->manager_arb)) |
29187a9e | 2020 | return false; |
2607d7a6 | 2021 | pool->manager = worker; |
1e19ffc6 | 2022 | |
29187a9e | 2023 | maybe_create_worker(pool); |
e22bee78 | 2024 | |
2607d7a6 | 2025 | pool->manager = NULL; |
34a06bd6 | 2026 | mutex_unlock(&pool->manager_arb); |
29187a9e | 2027 | return true; |
73f53c4a TH |
2028 | } |
2029 | ||
a62428c0 TH |
2030 | /** |
2031 | * process_one_work - process single work | |
c34056a3 | 2032 | * @worker: self |
a62428c0 TH |
2033 | * @work: work to process |
2034 | * | |
2035 | * Process @work. This function contains all the logics necessary to | |
2036 | * process a single work including synchronization against and | |
2037 | * interaction with other workers on the same cpu, queueing and | |
2038 | * flushing. As long as context requirement is met, any worker can | |
2039 | * call this function to process a work. | |
2040 | * | |
2041 | * CONTEXT: | |
d565ed63 | 2042 | * spin_lock_irq(pool->lock) which is released and regrabbed. |
a62428c0 | 2043 | */ |
c34056a3 | 2044 | static void process_one_work(struct worker *worker, struct work_struct *work) |
d565ed63 TH |
2045 | __releases(&pool->lock) |
2046 | __acquires(&pool->lock) | |
a62428c0 | 2047 | { |
112202d9 | 2048 | struct pool_workqueue *pwq = get_work_pwq(work); |
bd7bdd43 | 2049 | struct worker_pool *pool = worker->pool; |
112202d9 | 2050 | bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE; |
73f53c4a | 2051 | int work_color; |
7e11629d | 2052 | struct worker *collision; |
a62428c0 TH |
2053 | #ifdef CONFIG_LOCKDEP |
2054 | /* | |
2055 | * It is permissible to free the struct work_struct from | |
2056 | * inside the function that is called from it, this we need to | |
2057 | * take into account for lockdep too. To avoid bogus "held | |
2058 | * lock freed" warnings as well as problems when looking into | |
2059 | * work->lockdep_map, make a copy and use that here. | |
2060 | */ | |
4d82a1de PZ |
2061 | struct lockdep_map lockdep_map; |
2062 | ||
2063 | lockdep_copy_map(&lockdep_map, &work->lockdep_map); | |
a62428c0 | 2064 | #endif |
807407c0 | 2065 | /* ensure we're on the correct CPU */ |
85327af6 | 2066 | WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && |
ec22ca5e | 2067 | raw_smp_processor_id() != pool->cpu); |
25511a47 | 2068 | |
7e11629d TH |
2069 | /* |
2070 | * A single work shouldn't be executed concurrently by | |
2071 | * multiple workers on a single cpu. Check whether anyone is | |
2072 | * already processing the work. If so, defer the work to the | |
2073 | * currently executing one. | |
2074 | */ | |
c9e7cf27 | 2075 | collision = find_worker_executing_work(pool, work); |
7e11629d TH |
2076 | if (unlikely(collision)) { |
2077 | move_linked_works(work, &collision->scheduled, NULL); | |
2078 | return; | |
2079 | } | |
2080 | ||
8930caba | 2081 | /* claim and dequeue */ |
a62428c0 | 2082 | debug_work_deactivate(work); |
c9e7cf27 | 2083 | hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); |
c34056a3 | 2084 | worker->current_work = work; |
a2c1c57b | 2085 | worker->current_func = work->func; |
112202d9 | 2086 | worker->current_pwq = pwq; |
73f53c4a | 2087 | work_color = get_work_color(work); |
7a22ad75 | 2088 | |
a62428c0 TH |
2089 | list_del_init(&work->entry); |
2090 | ||
fb0e7beb | 2091 | /* |
228f1d00 LJ |
2092 | * CPU intensive works don't participate in concurrency management. |
2093 | * They're the scheduler's responsibility. This takes @worker out | |
2094 | * of concurrency management and the next code block will chain | |
2095 | * execution of the pending work items. | |
fb0e7beb TH |
2096 | */ |
2097 | if (unlikely(cpu_intensive)) | |
228f1d00 | 2098 | worker_set_flags(worker, WORKER_CPU_INTENSIVE); |
fb0e7beb | 2099 | |
974271c4 | 2100 | /* |
a489a03e LJ |
2101 | * Wake up another worker if necessary. The condition is always |
2102 | * false for normal per-cpu workers since nr_running would always | |
2103 | * be >= 1 at this point. This is used to chain execution of the | |
2104 | * pending work items for WORKER_NOT_RUNNING workers such as the | |
228f1d00 | 2105 | * UNBOUND and CPU_INTENSIVE ones. |
974271c4 | 2106 | */ |
a489a03e | 2107 | if (need_more_worker(pool)) |
63d95a91 | 2108 | wake_up_worker(pool); |
974271c4 | 2109 | |
8930caba | 2110 | /* |
7c3eed5c | 2111 | * Record the last pool and clear PENDING which should be the last |
d565ed63 | 2112 | * update to @work. Also, do this inside @pool->lock so that |
23657bb1 TH |
2113 | * PENDING and queued state changes happen together while IRQ is |
2114 | * disabled. | |
8930caba | 2115 | */ |
7c3eed5c | 2116 | set_work_pool_and_clear_pending(work, pool->id); |
a62428c0 | 2117 | |
d565ed63 | 2118 | spin_unlock_irq(&pool->lock); |
a62428c0 | 2119 | |
112202d9 | 2120 | lock_map_acquire_read(&pwq->wq->lockdep_map); |
a62428c0 | 2121 | lock_map_acquire(&lockdep_map); |
e36c886a | 2122 | trace_workqueue_execute_start(work); |
a2c1c57b | 2123 | worker->current_func(work); |
e36c886a AV |
2124 | /* |
2125 | * While we must be careful to not use "work" after this, the trace | |
2126 | * point will only record its address. | |
2127 | */ | |
2128 | trace_workqueue_execute_end(work); | |
a62428c0 | 2129 | lock_map_release(&lockdep_map); |
112202d9 | 2130 | lock_map_release(&pwq->wq->lockdep_map); |
a62428c0 TH |
2131 | |
2132 | if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { | |
044c782c VI |
2133 | pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" |
2134 | " last function: %pf\n", | |
a2c1c57b TH |
2135 | current->comm, preempt_count(), task_pid_nr(current), |
2136 | worker->current_func); | |
a62428c0 TH |
2137 | debug_show_held_locks(current); |
2138 | dump_stack(); | |
2139 | } | |
2140 | ||
b22ce278 TH |
2141 | /* |
2142 | * The following prevents a kworker from hogging CPU on !PREEMPT | |
2143 | * kernels, where a requeueing work item waiting for something to | |
2144 | * happen could deadlock with stop_machine as such work item could | |
2145 | * indefinitely requeue itself while all other CPUs are trapped in | |
789cbbec JL |
2146 | * stop_machine. At the same time, report a quiescent RCU state so |
2147 | * the same condition doesn't freeze RCU. | |
b22ce278 | 2148 | */ |
3e28e377 | 2149 | cond_resched_rcu_qs(); |
b22ce278 | 2150 | |
d565ed63 | 2151 | spin_lock_irq(&pool->lock); |
a62428c0 | 2152 | |
fb0e7beb TH |
2153 | /* clear cpu intensive status */ |
2154 | if (unlikely(cpu_intensive)) | |
2155 | worker_clr_flags(worker, WORKER_CPU_INTENSIVE); | |
2156 | ||
a62428c0 | 2157 | /* we're done with it, release */ |
42f8570f | 2158 | hash_del(&worker->hentry); |
c34056a3 | 2159 | worker->current_work = NULL; |
a2c1c57b | 2160 | worker->current_func = NULL; |
112202d9 | 2161 | worker->current_pwq = NULL; |
3d1cb205 | 2162 | worker->desc_valid = false; |
112202d9 | 2163 | pwq_dec_nr_in_flight(pwq, work_color); |
a62428c0 TH |
2164 | } |
2165 | ||
affee4b2 TH |
2166 | /** |
2167 | * process_scheduled_works - process scheduled works | |
2168 | * @worker: self | |
2169 | * | |
2170 | * Process all scheduled works. Please note that the scheduled list | |
2171 | * may change while processing a work, so this function repeatedly | |
2172 | * fetches a work from the top and executes it. | |
2173 | * | |
2174 | * CONTEXT: | |
d565ed63 | 2175 | * spin_lock_irq(pool->lock) which may be released and regrabbed |
affee4b2 TH |
2176 | * multiple times. |
2177 | */ | |
2178 | static void process_scheduled_works(struct worker *worker) | |
1da177e4 | 2179 | { |
affee4b2 TH |
2180 | while (!list_empty(&worker->scheduled)) { |
2181 | struct work_struct *work = list_first_entry(&worker->scheduled, | |
1da177e4 | 2182 | struct work_struct, entry); |
c34056a3 | 2183 | process_one_work(worker, work); |
1da177e4 | 2184 | } |
1da177e4 LT |
2185 | } |
2186 | ||
4690c4ab TH |
2187 | /** |
2188 | * worker_thread - the worker thread function | |
c34056a3 | 2189 | * @__worker: self |
4690c4ab | 2190 | * |
c5aa87bb TH |
2191 | * The worker thread function. All workers belong to a worker_pool - |
2192 | * either a per-cpu one or dynamic unbound one. These workers process all | |
2193 | * work items regardless of their specific target workqueue. The only | |
2194 | * exception is work items which belong to workqueues with a rescuer which | |
2195 | * will be explained in rescuer_thread(). | |
d185af30 YB |
2196 | * |
2197 | * Return: 0 | |
4690c4ab | 2198 | */ |
c34056a3 | 2199 | static int worker_thread(void *__worker) |
1da177e4 | 2200 | { |
c34056a3 | 2201 | struct worker *worker = __worker; |
bd7bdd43 | 2202 | struct worker_pool *pool = worker->pool; |
1da177e4 | 2203 | |
e22bee78 TH |
2204 | /* tell the scheduler that this is a workqueue worker */ |
2205 | worker->task->flags |= PF_WQ_WORKER; | |
c8e55f36 | 2206 | woke_up: |
d565ed63 | 2207 | spin_lock_irq(&pool->lock); |
1da177e4 | 2208 | |
a9ab775b TH |
2209 | /* am I supposed to die? */ |
2210 | if (unlikely(worker->flags & WORKER_DIE)) { | |
d565ed63 | 2211 | spin_unlock_irq(&pool->lock); |
a9ab775b TH |
2212 | WARN_ON_ONCE(!list_empty(&worker->entry)); |
2213 | worker->task->flags &= ~PF_WQ_WORKER; | |
60f5a4bc LJ |
2214 | |
2215 | set_task_comm(worker->task, "kworker/dying"); | |
7cda9aae | 2216 | ida_simple_remove(&pool->worker_ida, worker->id); |
60f5a4bc LJ |
2217 | worker_detach_from_pool(worker, pool); |
2218 | kfree(worker); | |
a9ab775b | 2219 | return 0; |
c8e55f36 | 2220 | } |
affee4b2 | 2221 | |
c8e55f36 | 2222 | worker_leave_idle(worker); |
db7bccf4 | 2223 | recheck: |
e22bee78 | 2224 | /* no more worker necessary? */ |
63d95a91 | 2225 | if (!need_more_worker(pool)) |
e22bee78 TH |
2226 | goto sleep; |
2227 | ||
2228 | /* do we need to manage? */ | |
63d95a91 | 2229 | if (unlikely(!may_start_working(pool)) && manage_workers(worker)) |
e22bee78 TH |
2230 | goto recheck; |
2231 | ||
c8e55f36 TH |
2232 | /* |
2233 | * ->scheduled list can only be filled while a worker is | |
2234 | * preparing to process a work or actually processing it. | |
2235 | * Make sure nobody diddled with it while I was sleeping. | |
2236 | */ | |
6183c009 | 2237 | WARN_ON_ONCE(!list_empty(&worker->scheduled)); |
c8e55f36 | 2238 | |
e22bee78 | 2239 | /* |
a9ab775b TH |
2240 | * Finish PREP stage. We're guaranteed to have at least one idle |
2241 | * worker or that someone else has already assumed the manager | |
2242 | * role. This is where @worker starts participating in concurrency | |
2243 | * management if applicable and concurrency management is restored | |
2244 | * after being rebound. See rebind_workers() for details. | |
e22bee78 | 2245 | */ |
a9ab775b | 2246 | worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); |
e22bee78 TH |
2247 | |
2248 | do { | |
c8e55f36 | 2249 | struct work_struct *work = |
bd7bdd43 | 2250 | list_first_entry(&pool->worklist, |
c8e55f36 TH |
2251 | struct work_struct, entry); |
2252 | ||
82607adc TH |
2253 | pool->watchdog_ts = jiffies; |
2254 | ||
c8e55f36 TH |
2255 | if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { |
2256 | /* optimization path, not strictly necessary */ | |
2257 | process_one_work(worker, work); | |
2258 | if (unlikely(!list_empty(&worker->scheduled))) | |
affee4b2 | 2259 | process_scheduled_works(worker); |
c8e55f36 TH |
2260 | } else { |
2261 | move_linked_works(work, &worker->scheduled, NULL); | |
2262 | process_scheduled_works(worker); | |
affee4b2 | 2263 | } |
63d95a91 | 2264 | } while (keep_working(pool)); |
e22bee78 | 2265 | |
228f1d00 | 2266 | worker_set_flags(worker, WORKER_PREP); |
d313dd85 | 2267 | sleep: |
c8e55f36 | 2268 | /* |
d565ed63 TH |
2269 | * pool->lock is held and there's no work to process and no need to |
2270 | * manage, sleep. Workers are woken up only while holding | |
2271 | * pool->lock or from local cpu, so setting the current state | |
2272 | * before releasing pool->lock is enough to prevent losing any | |
2273 | * event. | |
c8e55f36 TH |
2274 | */ |
2275 | worker_enter_idle(worker); | |
2276 | __set_current_state(TASK_INTERRUPTIBLE); | |
d565ed63 | 2277 | spin_unlock_irq(&pool->lock); |
c8e55f36 TH |
2278 | schedule(); |
2279 | goto woke_up; | |
1da177e4 LT |
2280 | } |
2281 | ||
e22bee78 TH |
2282 | /** |
2283 | * rescuer_thread - the rescuer thread function | |
111c225a | 2284 | * @__rescuer: self |
e22bee78 TH |
2285 | * |
2286 | * Workqueue rescuer thread function. There's one rescuer for each | |
493008a8 | 2287 | * workqueue which has WQ_MEM_RECLAIM set. |
e22bee78 | 2288 | * |
706026c2 | 2289 | * Regular work processing on a pool may block trying to create a new |
e22bee78 TH |
2290 | * worker which uses GFP_KERNEL allocation which has slight chance of |
2291 | * developing into deadlock if some works currently on the same queue | |
2292 | * need to be processed to satisfy the GFP_KERNEL allocation. This is | |
2293 | * the problem rescuer solves. | |
2294 | * | |
706026c2 TH |
2295 | * When such condition is possible, the pool summons rescuers of all |
2296 | * workqueues which have works queued on the pool and let them process | |
e22bee78 TH |
2297 | * those works so that forward progress can be guaranteed. |
2298 | * | |
2299 | * This should happen rarely. | |
d185af30 YB |
2300 | * |
2301 | * Return: 0 | |
e22bee78 | 2302 | */ |
111c225a | 2303 | static int rescuer_thread(void *__rescuer) |
e22bee78 | 2304 | { |
111c225a TH |
2305 | struct worker *rescuer = __rescuer; |
2306 | struct workqueue_struct *wq = rescuer->rescue_wq; | |
e22bee78 | 2307 | struct list_head *scheduled = &rescuer->scheduled; |
4d595b86 | 2308 | bool should_stop; |
e22bee78 TH |
2309 | |
2310 | set_user_nice(current, RESCUER_NICE_LEVEL); | |
111c225a TH |
2311 | |
2312 | /* | |
2313 | * Mark rescuer as worker too. As WORKER_PREP is never cleared, it | |
2314 | * doesn't participate in concurrency management. | |
2315 | */ | |
2316 | rescuer->task->flags |= PF_WQ_WORKER; | |
e22bee78 TH |
2317 | repeat: |
2318 | set_current_state(TASK_INTERRUPTIBLE); | |
2319 | ||
4d595b86 LJ |
2320 | /* |
2321 | * By the time the rescuer is requested to stop, the workqueue | |
2322 | * shouldn't have any work pending, but @wq->maydays may still have | |
2323 | * pwq(s) queued. This can happen by non-rescuer workers consuming | |
2324 | * all the work items before the rescuer got to them. Go through | |
2325 | * @wq->maydays processing before acting on should_stop so that the | |
2326 | * list is always empty on exit. | |
2327 | */ | |
2328 | should_stop = kthread_should_stop(); | |
e22bee78 | 2329 | |
493a1724 | 2330 | /* see whether any pwq is asking for help */ |
2e109a28 | 2331 | spin_lock_irq(&wq_mayday_lock); |
493a1724 TH |
2332 | |
2333 | while (!list_empty(&wq->maydays)) { | |
2334 | struct pool_workqueue *pwq = list_first_entry(&wq->maydays, | |
2335 | struct pool_workqueue, mayday_node); | |
112202d9 | 2336 | struct worker_pool *pool = pwq->pool; |
e22bee78 | 2337 | struct work_struct *work, *n; |
82607adc | 2338 | bool first = true; |
e22bee78 TH |
2339 | |
2340 | __set_current_state(TASK_RUNNING); | |
493a1724 TH |
2341 | list_del_init(&pwq->mayday_node); |
2342 | ||
2e109a28 | 2343 | spin_unlock_irq(&wq_mayday_lock); |
e22bee78 | 2344 | |
51697d39 LJ |
2345 | worker_attach_to_pool(rescuer, pool); |
2346 | ||
2347 | spin_lock_irq(&pool->lock); | |
b3104104 | 2348 | rescuer->pool = pool; |
e22bee78 TH |
2349 | |
2350 | /* | |
2351 | * Slurp in all works issued via this workqueue and | |
2352 | * process'em. | |
2353 | */ | |
0479c8c5 | 2354 | WARN_ON_ONCE(!list_empty(scheduled)); |
82607adc TH |
2355 | list_for_each_entry_safe(work, n, &pool->worklist, entry) { |
2356 | if (get_work_pwq(work) == pwq) { | |
2357 | if (first) | |
2358 | pool->watchdog_ts = jiffies; | |
e22bee78 | 2359 | move_linked_works(work, scheduled, &n); |
82607adc TH |
2360 | } |
2361 | first = false; | |
2362 | } | |
e22bee78 | 2363 | |
008847f6 N |
2364 | if (!list_empty(scheduled)) { |
2365 | process_scheduled_works(rescuer); | |
2366 | ||
2367 | /* | |
2368 | * The above execution of rescued work items could | |
2369 | * have created more to rescue through | |
2370 | * pwq_activate_first_delayed() or chained | |
2371 | * queueing. Let's put @pwq back on mayday list so | |
2372 | * that such back-to-back work items, which may be | |
2373 | * being used to relieve memory pressure, don't | |
2374 | * incur MAYDAY_INTERVAL delay inbetween. | |
2375 | */ | |
2376 | if (need_to_create_worker(pool)) { | |
2377 | spin_lock(&wq_mayday_lock); | |
2378 | get_pwq(pwq); | |
2379 | list_move_tail(&pwq->mayday_node, &wq->maydays); | |
2380 | spin_unlock(&wq_mayday_lock); | |
2381 | } | |
2382 | } | |
7576958a | 2383 | |
77668c8b LJ |
2384 | /* |
2385 | * Put the reference grabbed by send_mayday(). @pool won't | |
13b1d625 | 2386 | * go away while we're still attached to it. |
77668c8b LJ |
2387 | */ |
2388 | put_pwq(pwq); | |
2389 | ||
7576958a | 2390 | /* |
d8ca83e6 | 2391 | * Leave this pool. If need_more_worker() is %true, notify a |
7576958a TH |
2392 | * regular worker; otherwise, we end up with 0 concurrency |
2393 | * and stalling the execution. | |
2394 | */ | |
d8ca83e6 | 2395 | if (need_more_worker(pool)) |
63d95a91 | 2396 | wake_up_worker(pool); |
7576958a | 2397 | |
b3104104 | 2398 | rescuer->pool = NULL; |
13b1d625 LJ |
2399 | spin_unlock_irq(&pool->lock); |
2400 | ||
2401 | worker_detach_from_pool(rescuer, pool); | |
2402 | ||
2403 | spin_lock_irq(&wq_mayday_lock); | |
e22bee78 TH |
2404 | } |
2405 | ||
2e109a28 | 2406 | spin_unlock_irq(&wq_mayday_lock); |
493a1724 | 2407 | |
4d595b86 LJ |
2408 | if (should_stop) { |
2409 | __set_current_state(TASK_RUNNING); | |
2410 | rescuer->task->flags &= ~PF_WQ_WORKER; | |
2411 | return 0; | |
2412 | } | |
2413 | ||
111c225a TH |
2414 | /* rescuers should never participate in concurrency management */ |
2415 | WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); | |
e22bee78 TH |
2416 | schedule(); |
2417 | goto repeat; | |
1da177e4 LT |
2418 | } |
2419 | ||
fca839c0 TH |
2420 | /** |
2421 | * check_flush_dependency - check for flush dependency sanity | |
2422 | * @target_wq: workqueue being flushed | |
2423 | * @target_work: work item being flushed (NULL for workqueue flushes) | |
2424 | * | |
2425 | * %current is trying to flush the whole @target_wq or @target_work on it. | |
2426 | * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not | |
2427 | * reclaiming memory or running on a workqueue which doesn't have | |
2428 | * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to | |
2429 | * a deadlock. | |
2430 | */ | |
2431 | static void check_flush_dependency(struct workqueue_struct *target_wq, | |
2432 | struct work_struct *target_work) | |
2433 | { | |
2434 | work_func_t target_func = target_work ? target_work->func : NULL; | |
2435 | struct worker *worker; | |
2436 | ||
2437 | if (target_wq->flags & WQ_MEM_RECLAIM) | |
2438 | return; | |
2439 | ||
2440 | worker = current_wq_worker(); | |
2441 | ||
2442 | WARN_ONCE(current->flags & PF_MEMALLOC, | |
2443 | "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%pf", | |
2444 | current->pid, current->comm, target_wq->name, target_func); | |
23d11a58 TH |
2445 | WARN_ONCE(worker && ((worker->current_pwq->wq->flags & |
2446 | (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM), | |
fca839c0 TH |
2447 | "workqueue: WQ_MEM_RECLAIM %s:%pf is flushing !WQ_MEM_RECLAIM %s:%pf", |
2448 | worker->current_pwq->wq->name, worker->current_func, | |
2449 | target_wq->name, target_func); | |
2450 | } | |
2451 | ||
fc2e4d70 ON |
2452 | struct wq_barrier { |
2453 | struct work_struct work; | |
2454 | struct completion done; | |
2607d7a6 | 2455 | struct task_struct *task; /* purely informational */ |
fc2e4d70 ON |
2456 | }; |
2457 | ||
2458 | static void wq_barrier_func(struct work_struct *work) | |
2459 | { | |
2460 | struct wq_barrier *barr = container_of(work, struct wq_barrier, work); | |
2461 | complete(&barr->done); | |
2462 | } | |
2463 | ||
4690c4ab TH |
2464 | /** |
2465 | * insert_wq_barrier - insert a barrier work | |
112202d9 | 2466 | * @pwq: pwq to insert barrier into |
4690c4ab | 2467 | * @barr: wq_barrier to insert |
affee4b2 TH |
2468 | * @target: target work to attach @barr to |
2469 | * @worker: worker currently executing @target, NULL if @target is not executing | |
4690c4ab | 2470 | * |
affee4b2 TH |
2471 | * @barr is linked to @target such that @barr is completed only after |
2472 | * @target finishes execution. Please note that the ordering | |
2473 | * guarantee is observed only with respect to @target and on the local | |
2474 | * cpu. | |
2475 | * | |
2476 | * Currently, a queued barrier can't be canceled. This is because | |
2477 | * try_to_grab_pending() can't determine whether the work to be | |
2478 | * grabbed is at the head of the queue and thus can't clear LINKED | |
2479 | * flag of the previous work while there must be a valid next work | |
2480 | * after a work with LINKED flag set. | |
2481 | * | |
2482 | * Note that when @worker is non-NULL, @target may be modified | |
112202d9 | 2483 | * underneath us, so we can't reliably determine pwq from @target. |
4690c4ab TH |
2484 | * |
2485 | * CONTEXT: | |
d565ed63 | 2486 | * spin_lock_irq(pool->lock). |
4690c4ab | 2487 | */ |
112202d9 | 2488 | static void insert_wq_barrier(struct pool_workqueue *pwq, |
affee4b2 TH |
2489 | struct wq_barrier *barr, |
2490 | struct work_struct *target, struct worker *worker) | |
fc2e4d70 | 2491 | { |
affee4b2 TH |
2492 | struct list_head *head; |
2493 | unsigned int linked = 0; | |
2494 | ||
dc186ad7 | 2495 | /* |
d565ed63 | 2496 | * debugobject calls are safe here even with pool->lock locked |
dc186ad7 TG |
2497 | * as we know for sure that this will not trigger any of the |
2498 | * checks and call back into the fixup functions where we | |
2499 | * might deadlock. | |
2500 | */ | |
ca1cab37 | 2501 | INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); |
22df02bb | 2502 | __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); |
fc2e4d70 | 2503 | init_completion(&barr->done); |
2607d7a6 | 2504 | barr->task = current; |
83c22520 | 2505 | |
affee4b2 TH |
2506 | /* |
2507 | * If @target is currently being executed, schedule the | |
2508 | * barrier to the worker; otherwise, put it after @target. | |
2509 | */ | |
2510 | if (worker) | |
2511 | head = worker->scheduled.next; | |
2512 | else { | |
2513 | unsigned long *bits = work_data_bits(target); | |
2514 | ||
2515 | head = target->entry.next; | |
2516 | /* there can already be other linked works, inherit and set */ | |
2517 | linked = *bits & WORK_STRUCT_LINKED; | |
2518 | __set_bit(WORK_STRUCT_LINKED_BIT, bits); | |
2519 | } | |
2520 | ||
dc186ad7 | 2521 | debug_work_activate(&barr->work); |
112202d9 | 2522 | insert_work(pwq, &barr->work, head, |
affee4b2 | 2523 | work_color_to_flags(WORK_NO_COLOR) | linked); |
fc2e4d70 ON |
2524 | } |
2525 | ||
73f53c4a | 2526 | /** |
112202d9 | 2527 | * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing |
73f53c4a TH |
2528 | * @wq: workqueue being flushed |
2529 | * @flush_color: new flush color, < 0 for no-op | |
2530 | * @work_color: new work color, < 0 for no-op | |
2531 | * | |
112202d9 | 2532 | * Prepare pwqs for workqueue flushing. |
73f53c4a | 2533 | * |
112202d9 TH |
2534 | * If @flush_color is non-negative, flush_color on all pwqs should be |
2535 | * -1. If no pwq has in-flight commands at the specified color, all | |
2536 | * pwq->flush_color's stay at -1 and %false is returned. If any pwq | |
2537 | * has in flight commands, its pwq->flush_color is set to | |
2538 | * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq | |
73f53c4a TH |
2539 | * wakeup logic is armed and %true is returned. |
2540 | * | |
2541 | * The caller should have initialized @wq->first_flusher prior to | |
2542 | * calling this function with non-negative @flush_color. If | |
2543 | * @flush_color is negative, no flush color update is done and %false | |
2544 | * is returned. | |
2545 | * | |
112202d9 | 2546 | * If @work_color is non-negative, all pwqs should have the same |
73f53c4a TH |
2547 | * work_color which is previous to @work_color and all will be |
2548 | * advanced to @work_color. | |
2549 | * | |
2550 | * CONTEXT: | |
3c25a55d | 2551 | * mutex_lock(wq->mutex). |
73f53c4a | 2552 | * |
d185af30 | 2553 | * Return: |
73f53c4a TH |
2554 | * %true if @flush_color >= 0 and there's something to flush. %false |
2555 | * otherwise. | |
2556 | */ | |
112202d9 | 2557 | static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, |
73f53c4a | 2558 | int flush_color, int work_color) |
1da177e4 | 2559 | { |
73f53c4a | 2560 | bool wait = false; |
49e3cf44 | 2561 | struct pool_workqueue *pwq; |
1da177e4 | 2562 | |
73f53c4a | 2563 | if (flush_color >= 0) { |
6183c009 | 2564 | WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); |
112202d9 | 2565 | atomic_set(&wq->nr_pwqs_to_flush, 1); |
1da177e4 | 2566 | } |
2355b70f | 2567 | |
49e3cf44 | 2568 | for_each_pwq(pwq, wq) { |
112202d9 | 2569 | struct worker_pool *pool = pwq->pool; |
fc2e4d70 | 2570 | |
b09f4fd3 | 2571 | spin_lock_irq(&pool->lock); |
83c22520 | 2572 | |
73f53c4a | 2573 | if (flush_color >= 0) { |
6183c009 | 2574 | WARN_ON_ONCE(pwq->flush_color != -1); |
fc2e4d70 | 2575 | |
112202d9 TH |
2576 | if (pwq->nr_in_flight[flush_color]) { |
2577 | pwq->flush_color = flush_color; | |
2578 | atomic_inc(&wq->nr_pwqs_to_flush); | |
73f53c4a TH |
2579 | wait = true; |
2580 | } | |
2581 | } | |
1da177e4 | 2582 | |
73f53c4a | 2583 | if (work_color >= 0) { |
6183c009 | 2584 | WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); |
112202d9 | 2585 | pwq->work_color = work_color; |
73f53c4a | 2586 | } |
1da177e4 | 2587 | |
b09f4fd3 | 2588 | spin_unlock_irq(&pool->lock); |
1da177e4 | 2589 | } |
2355b70f | 2590 | |
112202d9 | 2591 | if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) |
73f53c4a | 2592 | complete(&wq->first_flusher->done); |
14441960 | 2593 | |
73f53c4a | 2594 | return wait; |
1da177e4 LT |
2595 | } |
2596 | ||
0fcb78c2 | 2597 | /** |
1da177e4 | 2598 | * flush_workqueue - ensure that any scheduled work has run to completion. |
0fcb78c2 | 2599 | * @wq: workqueue to flush |
1da177e4 | 2600 | * |
c5aa87bb TH |
2601 | * This function sleeps until all work items which were queued on entry |
2602 | * have finished execution, but it is not livelocked by new incoming ones. | |
1da177e4 | 2603 | */ |
7ad5b3a5 | 2604 | void flush_workqueue(struct workqueue_struct *wq) |
1da177e4 | 2605 | { |
73f53c4a TH |
2606 | struct wq_flusher this_flusher = { |
2607 | .list = LIST_HEAD_INIT(this_flusher.list), | |
2608 | .flush_color = -1, | |
2609 | .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done), | |
2610 | }; | |
2611 | int next_color; | |
1da177e4 | 2612 | |
3295f0ef IM |
2613 | lock_map_acquire(&wq->lockdep_map); |
2614 | lock_map_release(&wq->lockdep_map); | |
73f53c4a | 2615 | |
3c25a55d | 2616 | mutex_lock(&wq->mutex); |
73f53c4a TH |
2617 | |
2618 | /* | |
2619 | * Start-to-wait phase | |
2620 | */ | |
2621 | next_color = work_next_color(wq->work_color); | |
2622 | ||
2623 | if (next_color != wq->flush_color) { | |
2624 | /* | |
2625 | * Color space is not full. The current work_color | |
2626 | * becomes our flush_color and work_color is advanced | |
2627 | * by one. | |
2628 | */ | |
6183c009 | 2629 | WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); |
73f53c4a TH |
2630 | this_flusher.flush_color = wq->work_color; |
2631 | wq->work_color = next_color; | |
2632 | ||
2633 | if (!wq->first_flusher) { | |
2634 | /* no flush in progress, become the first flusher */ | |
6183c009 | 2635 | WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); |
73f53c4a TH |
2636 | |
2637 | wq->first_flusher = &this_flusher; | |
2638 | ||
112202d9 | 2639 | if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, |
73f53c4a TH |
2640 | wq->work_color)) { |
2641 | /* nothing to flush, done */ | |
2642 | wq->flush_color = next_color; | |
2643 | wq->first_flusher = NULL; | |
2644 | goto out_unlock; | |
2645 | } | |
2646 | } else { | |
2647 | /* wait in queue */ | |
6183c009 | 2648 | WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); |
73f53c4a | 2649 | list_add_tail(&this_flusher.list, &wq->flusher_queue); |
112202d9 | 2650 | flush_workqueue_prep_pwqs(wq, -1, wq->work_color); |
73f53c4a TH |
2651 | } |
2652 | } else { | |
2653 | /* | |
2654 | * Oops, color space is full, wait on overflow queue. | |
2655 | * The next flush completion will assign us | |
2656 | * flush_color and transfer to flusher_queue. | |
2657 | */ | |
2658 | list_add_tail(&this_flusher.list, &wq->flusher_overflow); | |
2659 | } | |
2660 | ||
fca839c0 TH |
2661 | check_flush_dependency(wq, NULL); |
2662 | ||
3c25a55d | 2663 | mutex_unlock(&wq->mutex); |
73f53c4a TH |
2664 | |
2665 | wait_for_completion(&this_flusher.done); | |
2666 | ||
2667 | /* | |
2668 | * Wake-up-and-cascade phase | |
2669 | * | |
2670 | * First flushers are responsible for cascading flushes and | |
2671 | * handling overflow. Non-first flushers can simply return. | |
2672 | */ | |
2673 | if (wq->first_flusher != &this_flusher) | |
2674 | return; | |
2675 | ||
3c25a55d | 2676 | mutex_lock(&wq->mutex); |
73f53c4a | 2677 | |
4ce48b37 TH |
2678 | /* we might have raced, check again with mutex held */ |
2679 | if (wq->first_flusher != &this_flusher) | |
2680 | goto out_unlock; | |
2681 | ||
73f53c4a TH |
2682 | wq->first_flusher = NULL; |
2683 | ||
6183c009 TH |
2684 | WARN_ON_ONCE(!list_empty(&this_flusher.list)); |
2685 | WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); | |
73f53c4a TH |
2686 | |
2687 | while (true) { | |
2688 | struct wq_flusher *next, *tmp; | |
2689 | ||
2690 | /* complete all the flushers sharing the current flush color */ | |
2691 | list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { | |
2692 | if (next->flush_color != wq->flush_color) | |
2693 | break; | |
2694 | list_del_init(&next->list); | |
2695 | complete(&next->done); | |
2696 | } | |
2697 | ||
6183c009 TH |
2698 | WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && |
2699 | wq->flush_color != work_next_color(wq->work_color)); | |
73f53c4a TH |
2700 | |
2701 | /* this flush_color is finished, advance by one */ | |
2702 | wq->flush_color = work_next_color(wq->flush_color); | |
2703 | ||
2704 | /* one color has been freed, handle overflow queue */ | |
2705 | if (!list_empty(&wq->flusher_overflow)) { | |
2706 | /* | |
2707 | * Assign the same color to all overflowed | |
2708 | * flushers, advance work_color and append to | |
2709 | * flusher_queue. This is the start-to-wait | |
2710 | * phase for these overflowed flushers. | |
2711 | */ | |
2712 | list_for_each_entry(tmp, &wq->flusher_overflow, list) | |
2713 | tmp->flush_color = wq->work_color; | |
2714 | ||
2715 | wq->work_color = work_next_color(wq->work_color); | |
2716 | ||
2717 | list_splice_tail_init(&wq->flusher_overflow, | |
2718 | &wq->flusher_queue); | |
112202d9 | 2719 | flush_workqueue_prep_pwqs(wq, -1, wq->work_color); |
73f53c4a TH |
2720 | } |
2721 | ||
2722 | if (list_empty(&wq->flusher_queue)) { | |
6183c009 | 2723 | WARN_ON_ONCE(wq->flush_color != wq->work_color); |
73f53c4a TH |
2724 | break; |
2725 | } | |
2726 | ||
2727 | /* | |
2728 | * Need to flush more colors. Make the next flusher | |
112202d9 | 2729 | * the new first flusher and arm pwqs. |
73f53c4a | 2730 | */ |
6183c009 TH |
2731 | WARN_ON_ONCE(wq->flush_color == wq->work_color); |
2732 | WARN_ON_ONCE(wq->flush_color != next->flush_color); | |
73f53c4a TH |
2733 | |
2734 | list_del_init(&next->list); | |
2735 | wq->first_flusher = next; | |
2736 | ||
112202d9 | 2737 | if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) |
73f53c4a TH |
2738 | break; |
2739 | ||
2740 | /* | |
2741 | * Meh... this color is already done, clear first | |
2742 | * flusher and repeat cascading. | |
2743 | */ | |
2744 | wq->first_flusher = NULL; | |
2745 | } | |
2746 | ||
2747 | out_unlock: | |
3c25a55d | 2748 | mutex_unlock(&wq->mutex); |
1da177e4 | 2749 | } |
1dadafa8 | 2750 | EXPORT_SYMBOL(flush_workqueue); |
1da177e4 | 2751 | |
9c5a2ba7 TH |
2752 | /** |
2753 | * drain_workqueue - drain a workqueue | |
2754 | * @wq: workqueue to drain | |
2755 | * | |
2756 | * Wait until the workqueue becomes empty. While draining is in progress, | |
2757 | * only chain queueing is allowed. IOW, only currently pending or running | |
2758 | * work items on @wq can queue further work items on it. @wq is flushed | |
b749b1b6 | 2759 | * repeatedly until it becomes empty. The number of flushing is determined |
9c5a2ba7 TH |
2760 | * by the depth of chaining and should be relatively short. Whine if it |
2761 | * takes too long. | |
2762 | */ | |
2763 | void drain_workqueue(struct workqueue_struct *wq) | |
2764 | { | |
2765 | unsigned int flush_cnt = 0; | |
49e3cf44 | 2766 | struct pool_workqueue *pwq; |
9c5a2ba7 TH |
2767 | |
2768 | /* | |
2769 | * __queue_work() needs to test whether there are drainers, is much | |
2770 | * hotter than drain_workqueue() and already looks at @wq->flags. | |
618b01eb | 2771 | * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. |
9c5a2ba7 | 2772 | */ |
87fc741e | 2773 | mutex_lock(&wq->mutex); |
9c5a2ba7 | 2774 | if (!wq->nr_drainers++) |
618b01eb | 2775 | wq->flags |= __WQ_DRAINING; |
87fc741e | 2776 | mutex_unlock(&wq->mutex); |
9c5a2ba7 TH |
2777 | reflush: |
2778 | flush_workqueue(wq); | |
2779 | ||
b09f4fd3 | 2780 | mutex_lock(&wq->mutex); |
76af4d93 | 2781 | |
49e3cf44 | 2782 | for_each_pwq(pwq, wq) { |
fa2563e4 | 2783 | bool drained; |
9c5a2ba7 | 2784 | |
b09f4fd3 | 2785 | spin_lock_irq(&pwq->pool->lock); |
112202d9 | 2786 | drained = !pwq->nr_active && list_empty(&pwq->delayed_works); |
b09f4fd3 | 2787 | spin_unlock_irq(&pwq->pool->lock); |
fa2563e4 TT |
2788 | |
2789 | if (drained) | |
9c5a2ba7 TH |
2790 | continue; |
2791 | ||
2792 | if (++flush_cnt == 10 || | |
2793 | (flush_cnt % 100 == 0 && flush_cnt <= 1000)) | |
c5aa87bb | 2794 | pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n", |
044c782c | 2795 | wq->name, flush_cnt); |
76af4d93 | 2796 | |
b09f4fd3 | 2797 | mutex_unlock(&wq->mutex); |
9c5a2ba7 TH |
2798 | goto reflush; |
2799 | } | |
2800 | ||
9c5a2ba7 | 2801 | if (!--wq->nr_drainers) |
618b01eb | 2802 | wq->flags &= ~__WQ_DRAINING; |
87fc741e | 2803 | mutex_unlock(&wq->mutex); |
9c5a2ba7 TH |
2804 | } |
2805 | EXPORT_SYMBOL_GPL(drain_workqueue); | |
2806 | ||
606a5020 | 2807 | static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr) |
db700897 | 2808 | { |
affee4b2 | 2809 | struct worker *worker = NULL; |
c9e7cf27 | 2810 | struct worker_pool *pool; |
112202d9 | 2811 | struct pool_workqueue *pwq; |
db700897 ON |
2812 | |
2813 | might_sleep(); | |
fa1b54e6 TH |
2814 | |
2815 | local_irq_disable(); | |
c9e7cf27 | 2816 | pool = get_work_pool(work); |
fa1b54e6 TH |
2817 | if (!pool) { |
2818 | local_irq_enable(); | |
baf59022 | 2819 | return false; |
fa1b54e6 | 2820 | } |
db700897 | 2821 | |
fa1b54e6 | 2822 | spin_lock(&pool->lock); |
0b3dae68 | 2823 | /* see the comment in try_to_grab_pending() with the same code */ |
112202d9 TH |
2824 | pwq = get_work_pwq(work); |
2825 | if (pwq) { | |
2826 | if (unlikely(pwq->pool != pool)) | |
4690c4ab | 2827 | goto already_gone; |
606a5020 | 2828 | } else { |
c9e7cf27 | 2829 | worker = find_worker_executing_work(pool, work); |
affee4b2 | 2830 | if (!worker) |
4690c4ab | 2831 | goto already_gone; |
112202d9 | 2832 | pwq = worker->current_pwq; |
606a5020 | 2833 | } |
db700897 | 2834 | |
fca839c0 TH |
2835 | check_flush_dependency(pwq->wq, work); |
2836 | ||
112202d9 | 2837 | insert_wq_barrier(pwq, barr, work, worker); |
d565ed63 | 2838 | spin_unlock_irq(&pool->lock); |
7a22ad75 | 2839 | |
e159489b TH |
2840 | /* |
2841 | * If @max_active is 1 or rescuer is in use, flushing another work | |
2842 | * item on the same workqueue may lead to deadlock. Make sure the | |
2843 | * flusher is not running on the same workqueue by verifying write | |
2844 | * access. | |
2845 | */ | |
493008a8 | 2846 | if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer) |
112202d9 | 2847 | lock_map_acquire(&pwq->wq->lockdep_map); |
e159489b | 2848 | else |
112202d9 TH |
2849 | lock_map_acquire_read(&pwq->wq->lockdep_map); |
2850 | lock_map_release(&pwq->wq->lockdep_map); | |
e159489b | 2851 | |
401a8d04 | 2852 | return true; |
4690c4ab | 2853 | already_gone: |
d565ed63 | 2854 | spin_unlock_irq(&pool->lock); |
401a8d04 | 2855 | return false; |
db700897 | 2856 | } |
baf59022 TH |
2857 | |
2858 | /** | |
2859 | * flush_work - wait for a work to finish executing the last queueing instance | |
2860 | * @work: the work to flush | |
2861 | * | |
606a5020 TH |
2862 | * Wait until @work has finished execution. @work is guaranteed to be idle |
2863 | * on return if it hasn't been requeued since flush started. | |
baf59022 | 2864 | * |
d185af30 | 2865 | * Return: |
baf59022 TH |
2866 | * %true if flush_work() waited for the work to finish execution, |
2867 | * %false if it was already idle. | |
2868 | */ | |
2869 | bool flush_work(struct work_struct *work) | |
2870 | { | |
12997d1a BH |
2871 | struct wq_barrier barr; |
2872 | ||
0976dfc1 SB |
2873 | lock_map_acquire(&work->lockdep_map); |
2874 | lock_map_release(&work->lockdep_map); | |
2875 | ||
12997d1a BH |
2876 | if (start_flush_work(work, &barr)) { |
2877 | wait_for_completion(&barr.done); | |
2878 | destroy_work_on_stack(&barr.work); | |
2879 | return true; | |
2880 | } else { | |
2881 | return false; | |
2882 | } | |
6e84d644 | 2883 | } |
606a5020 | 2884 | EXPORT_SYMBOL_GPL(flush_work); |
6e84d644 | 2885 | |
8603e1b3 TH |
2886 | struct cwt_wait { |
2887 | wait_queue_t wait; | |
2888 | struct work_struct *work; | |
2889 | }; | |
2890 | ||
2891 | static int cwt_wakefn(wait_queue_t *wait, unsigned mode, int sync, void *key) | |
2892 | { | |
2893 | struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait); | |
2894 | ||
2895 | if (cwait->work != key) | |
2896 | return 0; | |
2897 | return autoremove_wake_function(wait, mode, sync, key); | |
2898 | } | |
2899 | ||
36e227d2 | 2900 | static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) |
1f1f642e | 2901 | { |
8603e1b3 | 2902 | static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq); |
bbb68dfa | 2903 | unsigned long flags; |
1f1f642e ON |
2904 | int ret; |
2905 | ||
2906 | do { | |
bbb68dfa TH |
2907 | ret = try_to_grab_pending(work, is_dwork, &flags); |
2908 | /* | |
8603e1b3 TH |
2909 | * If someone else is already canceling, wait for it to |
2910 | * finish. flush_work() doesn't work for PREEMPT_NONE | |
2911 | * because we may get scheduled between @work's completion | |
2912 | * and the other canceling task resuming and clearing | |
2913 | * CANCELING - flush_work() will return false immediately | |
2914 | * as @work is no longer busy, try_to_grab_pending() will | |
2915 | * return -ENOENT as @work is still being canceled and the | |
2916 | * other canceling task won't be able to clear CANCELING as | |
2917 | * we're hogging the CPU. | |
2918 | * | |
2919 | * Let's wait for completion using a waitqueue. As this | |
2920 | * may lead to the thundering herd problem, use a custom | |
2921 | * wake function which matches @work along with exclusive | |
2922 | * wait and wakeup. | |
bbb68dfa | 2923 | */ |
8603e1b3 TH |
2924 | if (unlikely(ret == -ENOENT)) { |
2925 | struct cwt_wait cwait; | |
2926 | ||
2927 | init_wait(&cwait.wait); | |
2928 | cwait.wait.func = cwt_wakefn; | |
2929 | cwait.work = work; | |
2930 | ||
2931 | prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait, | |
2932 | TASK_UNINTERRUPTIBLE); | |
2933 | if (work_is_canceling(work)) | |
2934 | schedule(); | |
2935 | finish_wait(&cancel_waitq, &cwait.wait); | |
2936 | } | |
1f1f642e ON |
2937 | } while (unlikely(ret < 0)); |
2938 | ||
bbb68dfa TH |
2939 | /* tell other tasks trying to grab @work to back off */ |
2940 | mark_work_canceling(work); | |
2941 | local_irq_restore(flags); | |
2942 | ||
606a5020 | 2943 | flush_work(work); |
7a22ad75 | 2944 | clear_work_data(work); |
8603e1b3 TH |
2945 | |
2946 | /* | |
2947 | * Paired with prepare_to_wait() above so that either | |
2948 | * waitqueue_active() is visible here or !work_is_canceling() is | |
2949 | * visible there. | |
2950 | */ | |
2951 | smp_mb(); | |
2952 | if (waitqueue_active(&cancel_waitq)) | |
2953 | __wake_up(&cancel_waitq, TASK_NORMAL, 1, work); | |
2954 | ||
1f1f642e ON |
2955 | return ret; |
2956 | } | |
2957 | ||
6e84d644 | 2958 | /** |
401a8d04 TH |
2959 | * cancel_work_sync - cancel a work and wait for it to finish |
2960 | * @work: the work to cancel | |
6e84d644 | 2961 | * |
401a8d04 TH |
2962 | * Cancel @work and wait for its execution to finish. This function |
2963 | * can be used even if the work re-queues itself or migrates to | |
2964 | * another workqueue. On return from this function, @work is | |
2965 | * guaranteed to be not pending or executing on any CPU. | |
1f1f642e | 2966 | * |
401a8d04 TH |
2967 | * cancel_work_sync(&delayed_work->work) must not be used for |
2968 | * delayed_work's. Use cancel_delayed_work_sync() instead. | |
6e84d644 | 2969 | * |
401a8d04 | 2970 | * The caller must ensure that the workqueue on which @work was last |
6e84d644 | 2971 | * queued can't be destroyed before this function returns. |
401a8d04 | 2972 | * |
d185af30 | 2973 | * Return: |
401a8d04 | 2974 | * %true if @work was pending, %false otherwise. |
6e84d644 | 2975 | */ |
401a8d04 | 2976 | bool cancel_work_sync(struct work_struct *work) |
6e84d644 | 2977 | { |
36e227d2 | 2978 | return __cancel_work_timer(work, false); |
b89deed3 | 2979 | } |
28e53bdd | 2980 | EXPORT_SYMBOL_GPL(cancel_work_sync); |
b89deed3 | 2981 | |
6e84d644 | 2982 | /** |
401a8d04 TH |
2983 | * flush_delayed_work - wait for a dwork to finish executing the last queueing |
2984 | * @dwork: the delayed work to flush | |
6e84d644 | 2985 | * |
401a8d04 TH |
2986 | * Delayed timer is cancelled and the pending work is queued for |
2987 | * immediate execution. Like flush_work(), this function only | |
2988 | * considers the last queueing instance of @dwork. | |
1f1f642e | 2989 | * |
d185af30 | 2990 | * Return: |
401a8d04 TH |
2991 | * %true if flush_work() waited for the work to finish execution, |
2992 | * %false if it was already idle. | |
6e84d644 | 2993 | */ |
401a8d04 TH |
2994 | bool flush_delayed_work(struct delayed_work *dwork) |
2995 | { | |
8930caba | 2996 | local_irq_disable(); |
401a8d04 | 2997 | if (del_timer_sync(&dwork->timer)) |
60c057bc | 2998 | __queue_work(dwork->cpu, dwork->wq, &dwork->work); |
8930caba | 2999 | local_irq_enable(); |
401a8d04 TH |
3000 | return flush_work(&dwork->work); |
3001 | } | |
3002 | EXPORT_SYMBOL(flush_delayed_work); | |
3003 | ||
09383498 | 3004 | /** |
57b30ae7 TH |
3005 | * cancel_delayed_work - cancel a delayed work |
3006 | * @dwork: delayed_work to cancel | |
09383498 | 3007 | * |
d185af30 YB |
3008 | * Kill off a pending delayed_work. |
3009 | * | |
3010 | * Return: %true if @dwork was pending and canceled; %false if it wasn't | |
3011 | * pending. | |
3012 | * | |
3013 | * Note: | |
3014 | * The work callback function may still be running on return, unless | |
3015 | * it returns %true and the work doesn't re-arm itself. Explicitly flush or | |
3016 | * use cancel_delayed_work_sync() to wait on it. | |
09383498 | 3017 | * |
57b30ae7 | 3018 | * This function is safe to call from any context including IRQ handler. |
09383498 | 3019 | */ |
57b30ae7 | 3020 | bool cancel_delayed_work(struct delayed_work *dwork) |
09383498 | 3021 | { |
57b30ae7 TH |
3022 | unsigned long flags; |
3023 | int ret; | |
3024 | ||
3025 | do { | |
3026 | ret = try_to_grab_pending(&dwork->work, true, &flags); | |
3027 | } while (unlikely(ret == -EAGAIN)); | |
3028 | ||
3029 | if (unlikely(ret < 0)) | |
3030 | return false; | |
3031 | ||
7c3eed5c TH |
3032 | set_work_pool_and_clear_pending(&dwork->work, |
3033 | get_work_pool_id(&dwork->work)); | |
57b30ae7 | 3034 | local_irq_restore(flags); |
c0158ca6 | 3035 | return ret; |
09383498 | 3036 | } |
57b30ae7 | 3037 | EXPORT_SYMBOL(cancel_delayed_work); |
09383498 | 3038 | |
401a8d04 TH |
3039 | /** |
3040 | * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish | |
3041 | * @dwork: the delayed work cancel | |
3042 | * | |
3043 | * This is cancel_work_sync() for delayed works. | |
3044 | * | |
d185af30 | 3045 | * Return: |
401a8d04 TH |
3046 | * %true if @dwork was pending, %false otherwise. |
3047 | */ | |
3048 | bool cancel_delayed_work_sync(struct delayed_work *dwork) | |
6e84d644 | 3049 | { |
36e227d2 | 3050 | return __cancel_work_timer(&dwork->work, true); |
6e84d644 | 3051 | } |
f5a421a4 | 3052 | EXPORT_SYMBOL(cancel_delayed_work_sync); |
1da177e4 | 3053 | |
b6136773 | 3054 | /** |
31ddd871 | 3055 | * schedule_on_each_cpu - execute a function synchronously on each online CPU |
b6136773 | 3056 | * @func: the function to call |
b6136773 | 3057 | * |
31ddd871 TH |
3058 | * schedule_on_each_cpu() executes @func on each online CPU using the |
3059 | * system workqueue and blocks until all CPUs have completed. | |
b6136773 | 3060 | * schedule_on_each_cpu() is very slow. |
31ddd871 | 3061 | * |
d185af30 | 3062 | * Return: |
31ddd871 | 3063 | * 0 on success, -errno on failure. |
b6136773 | 3064 | */ |
65f27f38 | 3065 | int schedule_on_each_cpu(work_func_t func) |
15316ba8 CL |
3066 | { |
3067 | int cpu; | |
38f51568 | 3068 | struct work_struct __percpu *works; |
15316ba8 | 3069 | |
b6136773 AM |
3070 | works = alloc_percpu(struct work_struct); |
3071 | if (!works) | |
15316ba8 | 3072 | return -ENOMEM; |
b6136773 | 3073 | |
93981800 TH |
3074 | get_online_cpus(); |
3075 | ||
15316ba8 | 3076 | for_each_online_cpu(cpu) { |
9bfb1839 IM |
3077 | struct work_struct *work = per_cpu_ptr(works, cpu); |
3078 | ||
3079 | INIT_WORK(work, func); | |
b71ab8c2 | 3080 | schedule_work_on(cpu, work); |
65a64464 | 3081 | } |
93981800 TH |
3082 | |
3083 | for_each_online_cpu(cpu) | |
3084 | flush_work(per_cpu_ptr(works, cpu)); | |
3085 | ||
95402b38 | 3086 | put_online_cpus(); |
b6136773 | 3087 | free_percpu(works); |
15316ba8 CL |
3088 | return 0; |
3089 | } | |
3090 | ||
1fa44eca JB |
3091 | /** |
3092 | * execute_in_process_context - reliably execute the routine with user context | |
3093 | * @fn: the function to execute | |
1fa44eca JB |
3094 | * @ew: guaranteed storage for the execute work structure (must |
3095 | * be available when the work executes) | |
3096 | * | |
3097 | * Executes the function immediately if process context is available, | |
3098 | * otherwise schedules the function for delayed execution. | |
3099 | * | |
d185af30 | 3100 | * Return: 0 - function was executed |
1fa44eca JB |
3101 | * 1 - function was scheduled for execution |
3102 | */ | |
65f27f38 | 3103 | int execute_in_process_context(work_func_t fn, struct execute_work *ew) |
1fa44eca JB |
3104 | { |
3105 | if (!in_interrupt()) { | |
65f27f38 | 3106 | fn(&ew->work); |
1fa44eca JB |
3107 | return 0; |
3108 | } | |
3109 | ||
65f27f38 | 3110 | INIT_WORK(&ew->work, fn); |
1fa44eca JB |
3111 | schedule_work(&ew->work); |
3112 | ||
3113 | return 1; | |
3114 | } | |
3115 | EXPORT_SYMBOL_GPL(execute_in_process_context); | |
3116 | ||
6ba94429 FW |
3117 | /** |
3118 | * free_workqueue_attrs - free a workqueue_attrs | |
3119 | * @attrs: workqueue_attrs to free | |
226223ab | 3120 | * |
6ba94429 | 3121 | * Undo alloc_workqueue_attrs(). |
226223ab | 3122 | */ |
6ba94429 | 3123 | void free_workqueue_attrs(struct workqueue_attrs *attrs) |
226223ab | 3124 | { |
6ba94429 FW |
3125 | if (attrs) { |
3126 | free_cpumask_var(attrs->cpumask); | |
3127 | kfree(attrs); | |
3128 | } | |
226223ab TH |
3129 | } |
3130 | ||
6ba94429 FW |
3131 | /** |
3132 | * alloc_workqueue_attrs - allocate a workqueue_attrs | |
3133 | * @gfp_mask: allocation mask to use | |
3134 | * | |
3135 | * Allocate a new workqueue_attrs, initialize with default settings and | |
3136 | * return it. | |
3137 | * | |
3138 | * Return: The allocated new workqueue_attr on success. %NULL on failure. | |
3139 | */ | |
3140 | struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask) | |
226223ab | 3141 | { |
6ba94429 | 3142 | struct workqueue_attrs *attrs; |
226223ab | 3143 | |
6ba94429 FW |
3144 | attrs = kzalloc(sizeof(*attrs), gfp_mask); |
3145 | if (!attrs) | |
3146 | goto fail; | |
3147 | if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask)) | |
3148 | goto fail; | |
3149 | ||
3150 | cpumask_copy(attrs->cpumask, cpu_possible_mask); | |
3151 | return attrs; | |
3152 | fail: | |
3153 | free_workqueue_attrs(attrs); | |
3154 | return NULL; | |
226223ab TH |
3155 | } |
3156 | ||
6ba94429 FW |
3157 | static void copy_workqueue_attrs(struct workqueue_attrs *to, |
3158 | const struct workqueue_attrs *from) | |
226223ab | 3159 | { |
6ba94429 FW |
3160 | to->nice = from->nice; |
3161 | cpumask_copy(to->cpumask, from->cpumask); | |
3162 | /* | |
3163 | * Unlike hash and equality test, this function doesn't ignore | |
3164 | * ->no_numa as it is used for both pool and wq attrs. Instead, | |
3165 | * get_unbound_pool() explicitly clears ->no_numa after copying. | |
3166 | */ | |
3167 | to->no_numa = from->no_numa; | |
226223ab TH |
3168 | } |
3169 | ||
6ba94429 FW |
3170 | /* hash value of the content of @attr */ |
3171 | static u32 wqattrs_hash(const struct workqueue_attrs *attrs) | |
226223ab | 3172 | { |
6ba94429 | 3173 | u32 hash = 0; |
226223ab | 3174 | |
6ba94429 FW |
3175 | hash = jhash_1word(attrs->nice, hash); |
3176 | hash = jhash(cpumask_bits(attrs->cpumask), | |
3177 | BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); | |
3178 | return hash; | |
226223ab | 3179 | } |
226223ab | 3180 | |
6ba94429 FW |
3181 | /* content equality test */ |
3182 | static bool wqattrs_equal(const struct workqueue_attrs *a, | |
3183 | const struct workqueue_attrs *b) | |
226223ab | 3184 | { |
6ba94429 FW |
3185 | if (a->nice != b->nice) |
3186 | return false; | |
3187 | if (!cpumask_equal(a->cpumask, b->cpumask)) | |
3188 | return false; | |
3189 | return true; | |
226223ab TH |
3190 | } |
3191 | ||
6ba94429 FW |
3192 | /** |
3193 | * init_worker_pool - initialize a newly zalloc'd worker_pool | |
3194 | * @pool: worker_pool to initialize | |
3195 | * | |
402dd89d | 3196 | * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs. |
6ba94429 FW |
3197 | * |
3198 | * Return: 0 on success, -errno on failure. Even on failure, all fields | |
3199 | * inside @pool proper are initialized and put_unbound_pool() can be called | |
3200 | * on @pool safely to release it. | |
3201 | */ | |
3202 | static int init_worker_pool(struct worker_pool *pool) | |
226223ab | 3203 | { |
6ba94429 FW |
3204 | spin_lock_init(&pool->lock); |
3205 | pool->id = -1; | |
3206 | pool->cpu = -1; | |
3207 | pool->node = NUMA_NO_NODE; | |
3208 | pool->flags |= POOL_DISASSOCIATED; | |
82607adc | 3209 | pool->watchdog_ts = jiffies; |
6ba94429 FW |
3210 | INIT_LIST_HEAD(&pool->worklist); |
3211 | INIT_LIST_HEAD(&pool->idle_list); | |
3212 | hash_init(pool->busy_hash); | |
226223ab | 3213 | |
6ba94429 FW |
3214 | init_timer_deferrable(&pool->idle_timer); |
3215 | pool->idle_timer.function = idle_worker_timeout; | |
3216 | pool->idle_timer.data = (unsigned long)pool; | |
226223ab | 3217 | |
6ba94429 FW |
3218 | setup_timer(&pool->mayday_timer, pool_mayday_timeout, |
3219 | (unsigned long)pool); | |
226223ab | 3220 | |
6ba94429 FW |
3221 | mutex_init(&pool->manager_arb); |
3222 | mutex_init(&pool->attach_mutex); | |
3223 | INIT_LIST_HEAD(&pool->workers); | |
226223ab | 3224 | |
6ba94429 FW |
3225 | ida_init(&pool->worker_ida); |
3226 | INIT_HLIST_NODE(&pool->hash_node); | |
3227 | pool->refcnt = 1; | |
226223ab | 3228 | |
6ba94429 FW |
3229 | /* shouldn't fail above this point */ |
3230 | pool->attrs = alloc_workqueue_attrs(GFP_KERNEL); | |
3231 | if (!pool->attrs) | |
3232 | return -ENOMEM; | |
3233 | return 0; | |
226223ab TH |
3234 | } |
3235 | ||
6ba94429 | 3236 | static void rcu_free_wq(struct rcu_head *rcu) |
226223ab | 3237 | { |
6ba94429 FW |
3238 | struct workqueue_struct *wq = |
3239 | container_of(rcu, struct workqueue_struct, rcu); | |
226223ab | 3240 | |
6ba94429 FW |
3241 | if (!(wq->flags & WQ_UNBOUND)) |
3242 | free_percpu(wq->cpu_pwqs); | |
226223ab | 3243 | else |
6ba94429 | 3244 | free_workqueue_attrs(wq->unbound_attrs); |
226223ab | 3245 | |
6ba94429 FW |
3246 | kfree(wq->rescuer); |
3247 | kfree(wq); | |
226223ab TH |
3248 | } |
3249 | ||
6ba94429 | 3250 | static void rcu_free_pool(struct rcu_head *rcu) |
226223ab | 3251 | { |
6ba94429 | 3252 | struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); |
226223ab | 3253 | |
6ba94429 FW |
3254 | ida_destroy(&pool->worker_ida); |
3255 | free_workqueue_attrs(pool->attrs); | |
3256 | kfree(pool); | |
226223ab TH |
3257 | } |
3258 | ||
6ba94429 FW |
3259 | /** |
3260 | * put_unbound_pool - put a worker_pool | |
3261 | * @pool: worker_pool to put | |
3262 | * | |
3263 | * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU | |
3264 | * safe manner. get_unbound_pool() calls this function on its failure path | |
3265 | * and this function should be able to release pools which went through, | |
3266 | * successfully or not, init_worker_pool(). | |
3267 | * | |
3268 | * Should be called with wq_pool_mutex held. | |
3269 | */ | |
3270 | static void put_unbound_pool(struct worker_pool *pool) | |
226223ab | 3271 | { |
6ba94429 FW |
3272 | DECLARE_COMPLETION_ONSTACK(detach_completion); |
3273 | struct worker *worker; | |
226223ab | 3274 | |
6ba94429 | 3275 | lockdep_assert_held(&wq_pool_mutex); |
226223ab | 3276 | |
6ba94429 FW |
3277 | if (--pool->refcnt) |
3278 | return; | |
226223ab | 3279 | |
6ba94429 FW |
3280 | /* sanity checks */ |
3281 | if (WARN_ON(!(pool->cpu < 0)) || | |
3282 | WARN_ON(!list_empty(&pool->worklist))) | |
3283 | return; | |
226223ab | 3284 | |
6ba94429 FW |
3285 | /* release id and unhash */ |
3286 | if (pool->id >= 0) | |
3287 | idr_remove(&worker_pool_idr, pool->id); | |
3288 | hash_del(&pool->hash_node); | |
d55262c4 | 3289 | |
6ba94429 FW |
3290 | /* |
3291 | * Become the manager and destroy all workers. Grabbing | |
3292 | * manager_arb prevents @pool's workers from blocking on | |
3293 | * attach_mutex. | |
3294 | */ | |
3295 | mutex_lock(&pool->manager_arb); | |
d55262c4 | 3296 | |
6ba94429 FW |
3297 | spin_lock_irq(&pool->lock); |
3298 | while ((worker = first_idle_worker(pool))) | |
3299 | destroy_worker(worker); | |
3300 | WARN_ON(pool->nr_workers || pool->nr_idle); | |
3301 | spin_unlock_irq(&pool->lock); | |
d55262c4 | 3302 | |
6ba94429 FW |
3303 | mutex_lock(&pool->attach_mutex); |
3304 | if (!list_empty(&pool->workers)) | |
3305 | pool->detach_completion = &detach_completion; | |
3306 | mutex_unlock(&pool->attach_mutex); | |
226223ab | 3307 | |
6ba94429 FW |
3308 | if (pool->detach_completion) |
3309 | wait_for_completion(pool->detach_completion); | |
226223ab | 3310 | |
6ba94429 | 3311 | mutex_unlock(&pool->manager_arb); |
226223ab | 3312 | |
6ba94429 FW |
3313 | /* shut down the timers */ |
3314 | del_timer_sync(&pool->idle_timer); | |
3315 | del_timer_sync(&pool->mayday_timer); | |
226223ab | 3316 | |
6ba94429 FW |
3317 | /* sched-RCU protected to allow dereferences from get_work_pool() */ |
3318 | call_rcu_sched(&pool->rcu, rcu_free_pool); | |
226223ab TH |
3319 | } |
3320 | ||
3321 | /** | |
6ba94429 FW |
3322 | * get_unbound_pool - get a worker_pool with the specified attributes |
3323 | * @attrs: the attributes of the worker_pool to get | |
226223ab | 3324 | * |
6ba94429 FW |
3325 | * Obtain a worker_pool which has the same attributes as @attrs, bump the |
3326 | * reference count and return it. If there already is a matching | |
3327 | * worker_pool, it will be used; otherwise, this function attempts to | |
3328 | * create a new one. | |
226223ab | 3329 | * |
6ba94429 | 3330 | * Should be called with wq_pool_mutex held. |
226223ab | 3331 | * |
6ba94429 FW |
3332 | * Return: On success, a worker_pool with the same attributes as @attrs. |
3333 | * On failure, %NULL. | |
226223ab | 3334 | */ |
6ba94429 | 3335 | static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) |
226223ab | 3336 | { |
6ba94429 FW |
3337 | u32 hash = wqattrs_hash(attrs); |
3338 | struct worker_pool *pool; | |
3339 | int node; | |
e2273584 | 3340 | int target_node = NUMA_NO_NODE; |
226223ab | 3341 | |
6ba94429 | 3342 | lockdep_assert_held(&wq_pool_mutex); |
226223ab | 3343 | |
6ba94429 FW |
3344 | /* do we already have a matching pool? */ |
3345 | hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { | |
3346 | if (wqattrs_equal(pool->attrs, attrs)) { | |
3347 | pool->refcnt++; | |
3348 | return pool; | |
3349 | } | |
3350 | } | |
226223ab | 3351 | |
e2273584 XP |
3352 | /* if cpumask is contained inside a NUMA node, we belong to that node */ |
3353 | if (wq_numa_enabled) { | |
3354 | for_each_node(node) { | |
3355 | if (cpumask_subset(attrs->cpumask, | |
3356 | wq_numa_possible_cpumask[node])) { | |
3357 | target_node = node; | |
3358 | break; | |
3359 | } | |
3360 | } | |
3361 | } | |
3362 | ||
6ba94429 | 3363 | /* nope, create a new one */ |
e2273584 | 3364 | pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node); |
6ba94429 FW |
3365 | if (!pool || init_worker_pool(pool) < 0) |
3366 | goto fail; | |
3367 | ||
3368 | lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */ | |
3369 | copy_workqueue_attrs(pool->attrs, attrs); | |
e2273584 | 3370 | pool->node = target_node; |
226223ab TH |
3371 | |
3372 | /* | |
6ba94429 FW |
3373 | * no_numa isn't a worker_pool attribute, always clear it. See |
3374 | * 'struct workqueue_attrs' comments for detail. | |
226223ab | 3375 | */ |
6ba94429 | 3376 | pool->attrs->no_numa = false; |
226223ab | 3377 | |
6ba94429 FW |
3378 | if (worker_pool_assign_id(pool) < 0) |
3379 | goto fail; | |
226223ab | 3380 | |
6ba94429 FW |
3381 | /* create and start the initial worker */ |
3382 | if (!create_worker(pool)) | |
3383 | goto fail; | |
226223ab | 3384 | |
6ba94429 FW |
3385 | /* install */ |
3386 | hash_add(unbound_pool_hash, &pool->hash_node, hash); | |
226223ab | 3387 | |
6ba94429 FW |
3388 | return pool; |
3389 | fail: | |
3390 | if (pool) | |
3391 | put_unbound_pool(pool); | |
3392 | return NULL; | |
226223ab | 3393 | } |
226223ab | 3394 | |
6ba94429 | 3395 | static void rcu_free_pwq(struct rcu_head *rcu) |
7a4e344c | 3396 | { |
6ba94429 FW |
3397 | kmem_cache_free(pwq_cache, |
3398 | container_of(rcu, struct pool_workqueue, rcu)); | |
7a4e344c TH |
3399 | } |
3400 | ||
6ba94429 FW |
3401 | /* |
3402 | * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt | |
3403 | * and needs to be destroyed. | |
7a4e344c | 3404 | */ |
6ba94429 | 3405 | static void pwq_unbound_release_workfn(struct work_struct *work) |
7a4e344c | 3406 | { |
6ba94429 FW |
3407 | struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, |
3408 | unbound_release_work); | |
3409 | struct workqueue_struct *wq = pwq->wq; | |
3410 | struct worker_pool *pool = pwq->pool; | |
3411 | bool is_last; | |
7a4e344c | 3412 | |
6ba94429 FW |
3413 | if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND))) |
3414 | return; | |
7a4e344c | 3415 | |
6ba94429 FW |
3416 | mutex_lock(&wq->mutex); |
3417 | list_del_rcu(&pwq->pwqs_node); | |
3418 | is_last = list_empty(&wq->pwqs); | |
3419 | mutex_unlock(&wq->mutex); | |
3420 | ||
3421 | mutex_lock(&wq_pool_mutex); | |
3422 | put_unbound_pool(pool); | |
3423 | mutex_unlock(&wq_pool_mutex); | |
3424 | ||
3425 | call_rcu_sched(&pwq->rcu, rcu_free_pwq); | |
7a4e344c | 3426 | |
2865a8fb | 3427 | /* |
6ba94429 FW |
3428 | * If we're the last pwq going away, @wq is already dead and no one |
3429 | * is gonna access it anymore. Schedule RCU free. | |
2865a8fb | 3430 | */ |
6ba94429 FW |
3431 | if (is_last) |
3432 | call_rcu_sched(&wq->rcu, rcu_free_wq); | |
29c91e99 TH |
3433 | } |
3434 | ||
7a4e344c | 3435 | /** |
6ba94429 FW |
3436 | * pwq_adjust_max_active - update a pwq's max_active to the current setting |
3437 | * @pwq: target pool_workqueue | |
d185af30 | 3438 | * |
6ba94429 FW |
3439 | * If @pwq isn't freezing, set @pwq->max_active to the associated |
3440 | * workqueue's saved_max_active and activate delayed work items | |
3441 | * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. | |
7a4e344c | 3442 | */ |
6ba94429 | 3443 | static void pwq_adjust_max_active(struct pool_workqueue *pwq) |
4e1a1f9a | 3444 | { |
6ba94429 FW |
3445 | struct workqueue_struct *wq = pwq->wq; |
3446 | bool freezable = wq->flags & WQ_FREEZABLE; | |
4e1a1f9a | 3447 | |
6ba94429 FW |
3448 | /* for @wq->saved_max_active */ |
3449 | lockdep_assert_held(&wq->mutex); | |
4e1a1f9a | 3450 | |
6ba94429 FW |
3451 | /* fast exit for non-freezable wqs */ |
3452 | if (!freezable && pwq->max_active == wq->saved_max_active) | |
3453 | return; | |
7a4e344c | 3454 | |
6ba94429 | 3455 | spin_lock_irq(&pwq->pool->lock); |
29c91e99 | 3456 | |
6ba94429 FW |
3457 | /* |
3458 | * During [un]freezing, the caller is responsible for ensuring that | |
3459 | * this function is called at least once after @workqueue_freezing | |
3460 | * is updated and visible. | |
3461 | */ | |
3462 | if (!freezable || !workqueue_freezing) { | |
3463 | pwq->max_active = wq->saved_max_active; | |
4e1a1f9a | 3464 | |
6ba94429 FW |
3465 | while (!list_empty(&pwq->delayed_works) && |
3466 | pwq->nr_active < pwq->max_active) | |
3467 | pwq_activate_first_delayed(pwq); | |
e2dca7ad | 3468 | |
6ba94429 FW |
3469 | /* |
3470 | * Need to kick a worker after thawed or an unbound wq's | |
3471 | * max_active is bumped. It's a slow path. Do it always. | |
3472 | */ | |
3473 | wake_up_worker(pwq->pool); | |
3474 | } else { | |
3475 | pwq->max_active = 0; | |
3476 | } | |
e2dca7ad | 3477 | |
6ba94429 | 3478 | spin_unlock_irq(&pwq->pool->lock); |
e2dca7ad TH |
3479 | } |
3480 | ||
6ba94429 FW |
3481 | /* initialize newly alloced @pwq which is associated with @wq and @pool */ |
3482 | static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, | |
3483 | struct worker_pool *pool) | |
29c91e99 | 3484 | { |
6ba94429 | 3485 | BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); |
29c91e99 | 3486 | |
6ba94429 FW |
3487 | memset(pwq, 0, sizeof(*pwq)); |
3488 | ||
3489 | pwq->pool = pool; | |
3490 | pwq->wq = wq; | |
3491 | pwq->flush_color = -1; | |
3492 | pwq->refcnt = 1; | |
3493 | INIT_LIST_HEAD(&pwq->delayed_works); | |
3494 | INIT_LIST_HEAD(&pwq->pwqs_node); | |
3495 | INIT_LIST_HEAD(&pwq->mayday_node); | |
3496 | INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn); | |
29c91e99 TH |
3497 | } |
3498 | ||
6ba94429 FW |
3499 | /* sync @pwq with the current state of its associated wq and link it */ |
3500 | static void link_pwq(struct pool_workqueue *pwq) | |
29c91e99 | 3501 | { |
6ba94429 | 3502 | struct workqueue_struct *wq = pwq->wq; |
29c91e99 | 3503 | |
6ba94429 | 3504 | lockdep_assert_held(&wq->mutex); |
a892cacc | 3505 | |
6ba94429 FW |
3506 | /* may be called multiple times, ignore if already linked */ |
3507 | if (!list_empty(&pwq->pwqs_node)) | |
29c91e99 | 3508 | return; |
29c91e99 | 3509 | |
6ba94429 FW |
3510 | /* set the matching work_color */ |
3511 | pwq->work_color = wq->work_color; | |
29c91e99 | 3512 | |
6ba94429 FW |
3513 | /* sync max_active to the current setting */ |
3514 | pwq_adjust_max_active(pwq); | |
29c91e99 | 3515 | |
6ba94429 FW |
3516 | /* link in @pwq */ |
3517 | list_add_rcu(&pwq->pwqs_node, &wq->pwqs); | |
3518 | } | |
29c91e99 | 3519 | |
6ba94429 FW |
3520 | /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ |
3521 | static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, | |
3522 | const struct workqueue_attrs *attrs) | |
3523 | { | |
3524 | struct worker_pool *pool; | |
3525 | struct pool_workqueue *pwq; | |
60f5a4bc | 3526 | |
6ba94429 | 3527 | lockdep_assert_held(&wq_pool_mutex); |
60f5a4bc | 3528 | |
6ba94429 FW |
3529 | pool = get_unbound_pool(attrs); |
3530 | if (!pool) | |
3531 | return NULL; | |
60f5a4bc | 3532 | |
6ba94429 FW |
3533 | pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); |
3534 | if (!pwq) { | |
3535 | put_unbound_pool(pool); | |
3536 | return NULL; | |
3537 | } | |
29c91e99 | 3538 | |
6ba94429 FW |
3539 | init_pwq(pwq, wq, pool); |
3540 | return pwq; | |
3541 | } | |
29c91e99 | 3542 | |
29c91e99 | 3543 | /** |
30186c6f | 3544 | * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node |
042f7df1 | 3545 | * @attrs: the wq_attrs of the default pwq of the target workqueue |
6ba94429 FW |
3546 | * @node: the target NUMA node |
3547 | * @cpu_going_down: if >= 0, the CPU to consider as offline | |
3548 | * @cpumask: outarg, the resulting cpumask | |
29c91e99 | 3549 | * |
6ba94429 FW |
3550 | * Calculate the cpumask a workqueue with @attrs should use on @node. If |
3551 | * @cpu_going_down is >= 0, that cpu is considered offline during | |
3552 | * calculation. The result is stored in @cpumask. | |
a892cacc | 3553 | * |
6ba94429 FW |
3554 | * If NUMA affinity is not enabled, @attrs->cpumask is always used. If |
3555 | * enabled and @node has online CPUs requested by @attrs, the returned | |
3556 | * cpumask is the intersection of the possible CPUs of @node and | |
3557 | * @attrs->cpumask. | |
d185af30 | 3558 | * |
6ba94429 FW |
3559 | * The caller is responsible for ensuring that the cpumask of @node stays |
3560 | * stable. | |
3561 | * | |
3562 | * Return: %true if the resulting @cpumask is different from @attrs->cpumask, | |
3563 | * %false if equal. | |
29c91e99 | 3564 | */ |
6ba94429 FW |
3565 | static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node, |
3566 | int cpu_going_down, cpumask_t *cpumask) | |
29c91e99 | 3567 | { |
6ba94429 FW |
3568 | if (!wq_numa_enabled || attrs->no_numa) |
3569 | goto use_dfl; | |
29c91e99 | 3570 | |
6ba94429 FW |
3571 | /* does @node have any online CPUs @attrs wants? */ |
3572 | cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask); | |
3573 | if (cpu_going_down >= 0) | |
3574 | cpumask_clear_cpu(cpu_going_down, cpumask); | |
29c91e99 | 3575 | |
6ba94429 FW |
3576 | if (cpumask_empty(cpumask)) |
3577 | goto use_dfl; | |
4c16bd32 TH |
3578 | |
3579 | /* yeap, return possible CPUs in @node that @attrs wants */ | |
3580 | cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]); | |
3581 | return !cpumask_equal(cpumask, attrs->cpumask); | |
3582 | ||
3583 | use_dfl: | |
3584 | cpumask_copy(cpumask, attrs->cpumask); | |
3585 | return false; | |
3586 | } | |
3587 | ||
1befcf30 TH |
3588 | /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */ |
3589 | static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq, | |
3590 | int node, | |
3591 | struct pool_workqueue *pwq) | |
3592 | { | |
3593 | struct pool_workqueue *old_pwq; | |
3594 | ||
5b95e1af | 3595 | lockdep_assert_held(&wq_pool_mutex); |
1befcf30 TH |
3596 | lockdep_assert_held(&wq->mutex); |
3597 | ||
3598 | /* link_pwq() can handle duplicate calls */ | |
3599 | link_pwq(pwq); | |
3600 | ||
3601 | old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); | |
3602 | rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq); | |
3603 | return old_pwq; | |
3604 | } | |
3605 | ||
2d5f0764 LJ |
3606 | /* context to store the prepared attrs & pwqs before applying */ |
3607 | struct apply_wqattrs_ctx { | |
3608 | struct workqueue_struct *wq; /* target workqueue */ | |
3609 | struct workqueue_attrs *attrs; /* attrs to apply */ | |
042f7df1 | 3610 | struct list_head list; /* queued for batching commit */ |
2d5f0764 LJ |
3611 | struct pool_workqueue *dfl_pwq; |
3612 | struct pool_workqueue *pwq_tbl[]; | |
3613 | }; | |
3614 | ||
3615 | /* free the resources after success or abort */ | |
3616 | static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx) | |
3617 | { | |
3618 | if (ctx) { | |
3619 | int node; | |
3620 | ||
3621 | for_each_node(node) | |
3622 | put_pwq_unlocked(ctx->pwq_tbl[node]); | |
3623 | put_pwq_unlocked(ctx->dfl_pwq); | |
3624 | ||
3625 | free_workqueue_attrs(ctx->attrs); | |
3626 | ||
3627 | kfree(ctx); | |
3628 | } | |
3629 | } | |
3630 | ||
3631 | /* allocate the attrs and pwqs for later installation */ | |
3632 | static struct apply_wqattrs_ctx * | |
3633 | apply_wqattrs_prepare(struct workqueue_struct *wq, | |
3634 | const struct workqueue_attrs *attrs) | |
9e8cd2f5 | 3635 | { |
2d5f0764 | 3636 | struct apply_wqattrs_ctx *ctx; |
4c16bd32 | 3637 | struct workqueue_attrs *new_attrs, *tmp_attrs; |
2d5f0764 | 3638 | int node; |
9e8cd2f5 | 3639 | |
2d5f0764 | 3640 | lockdep_assert_held(&wq_pool_mutex); |
9e8cd2f5 | 3641 | |
2d5f0764 LJ |
3642 | ctx = kzalloc(sizeof(*ctx) + nr_node_ids * sizeof(ctx->pwq_tbl[0]), |
3643 | GFP_KERNEL); | |
8719dcea | 3644 | |
13e2e556 | 3645 | new_attrs = alloc_workqueue_attrs(GFP_KERNEL); |
4c16bd32 | 3646 | tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL); |
2d5f0764 LJ |
3647 | if (!ctx || !new_attrs || !tmp_attrs) |
3648 | goto out_free; | |
13e2e556 | 3649 | |
042f7df1 LJ |
3650 | /* |
3651 | * Calculate the attrs of the default pwq. | |
3652 | * If the user configured cpumask doesn't overlap with the | |
3653 | * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask. | |
3654 | */ | |
13e2e556 | 3655 | copy_workqueue_attrs(new_attrs, attrs); |
b05a7928 | 3656 | cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask); |
042f7df1 LJ |
3657 | if (unlikely(cpumask_empty(new_attrs->cpumask))) |
3658 | cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask); | |
13e2e556 | 3659 | |
4c16bd32 TH |
3660 | /* |
3661 | * We may create multiple pwqs with differing cpumasks. Make a | |
3662 | * copy of @new_attrs which will be modified and used to obtain | |
3663 | * pools. | |
3664 | */ | |
3665 | copy_workqueue_attrs(tmp_attrs, new_attrs); | |
3666 | ||
4c16bd32 TH |
3667 | /* |
3668 | * If something goes wrong during CPU up/down, we'll fall back to | |
3669 | * the default pwq covering whole @attrs->cpumask. Always create | |
3670 | * it even if we don't use it immediately. | |
3671 | */ | |
2d5f0764 LJ |
3672 | ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs); |
3673 | if (!ctx->dfl_pwq) | |
3674 | goto out_free; | |
4c16bd32 TH |
3675 | |
3676 | for_each_node(node) { | |
042f7df1 | 3677 | if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) { |
2d5f0764 LJ |
3678 | ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs); |
3679 | if (!ctx->pwq_tbl[node]) | |
3680 | goto out_free; | |
4c16bd32 | 3681 | } else { |
2d5f0764 LJ |
3682 | ctx->dfl_pwq->refcnt++; |
3683 | ctx->pwq_tbl[node] = ctx->dfl_pwq; | |
4c16bd32 TH |
3684 | } |
3685 | } | |
3686 | ||
042f7df1 LJ |
3687 | /* save the user configured attrs and sanitize it. */ |
3688 | copy_workqueue_attrs(new_attrs, attrs); | |
3689 | cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); | |
2d5f0764 | 3690 | ctx->attrs = new_attrs; |
042f7df1 | 3691 | |
2d5f0764 LJ |
3692 | ctx->wq = wq; |
3693 | free_workqueue_attrs(tmp_attrs); | |
3694 | return ctx; | |
3695 | ||
3696 | out_free: | |
3697 | free_workqueue_attrs(tmp_attrs); | |
3698 | free_workqueue_attrs(new_attrs); | |
3699 | apply_wqattrs_cleanup(ctx); | |
3700 | return NULL; | |
3701 | } | |
3702 | ||
3703 | /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */ | |
3704 | static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx) | |
3705 | { | |
3706 | int node; | |
9e8cd2f5 | 3707 | |
4c16bd32 | 3708 | /* all pwqs have been created successfully, let's install'em */ |
2d5f0764 | 3709 | mutex_lock(&ctx->wq->mutex); |
a892cacc | 3710 | |
2d5f0764 | 3711 | copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs); |
4c16bd32 TH |
3712 | |
3713 | /* save the previous pwq and install the new one */ | |
f147f29e | 3714 | for_each_node(node) |
2d5f0764 LJ |
3715 | ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node, |
3716 | ctx->pwq_tbl[node]); | |
4c16bd32 TH |
3717 | |
3718 | /* @dfl_pwq might not have been used, ensure it's linked */ | |
2d5f0764 LJ |
3719 | link_pwq(ctx->dfl_pwq); |
3720 | swap(ctx->wq->dfl_pwq, ctx->dfl_pwq); | |
f147f29e | 3721 | |
2d5f0764 LJ |
3722 | mutex_unlock(&ctx->wq->mutex); |
3723 | } | |
9e8cd2f5 | 3724 | |
a0111cf6 LJ |
3725 | static void apply_wqattrs_lock(void) |
3726 | { | |
3727 | /* CPUs should stay stable across pwq creations and installations */ | |
3728 | get_online_cpus(); | |
3729 | mutex_lock(&wq_pool_mutex); | |
3730 | } | |
3731 | ||
3732 | static void apply_wqattrs_unlock(void) | |
3733 | { | |
3734 | mutex_unlock(&wq_pool_mutex); | |
3735 | put_online_cpus(); | |
3736 | } | |
3737 | ||
3738 | static int apply_workqueue_attrs_locked(struct workqueue_struct *wq, | |
3739 | const struct workqueue_attrs *attrs) | |
2d5f0764 LJ |
3740 | { |
3741 | struct apply_wqattrs_ctx *ctx; | |
4c16bd32 | 3742 | |
2d5f0764 LJ |
3743 | /* only unbound workqueues can change attributes */ |
3744 | if (WARN_ON(!(wq->flags & WQ_UNBOUND))) | |
3745 | return -EINVAL; | |
13e2e556 | 3746 | |
2d5f0764 LJ |
3747 | /* creating multiple pwqs breaks ordering guarantee */ |
3748 | if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs))) | |
3749 | return -EINVAL; | |
3750 | ||
2d5f0764 | 3751 | ctx = apply_wqattrs_prepare(wq, attrs); |
6201171e | 3752 | if (!ctx) |
3753 | return -ENOMEM; | |
2d5f0764 LJ |
3754 | |
3755 | /* the ctx has been prepared successfully, let's commit it */ | |
6201171e | 3756 | apply_wqattrs_commit(ctx); |
2d5f0764 LJ |
3757 | apply_wqattrs_cleanup(ctx); |
3758 | ||
6201171e | 3759 | return 0; |
9e8cd2f5 TH |
3760 | } |
3761 | ||
a0111cf6 LJ |
3762 | /** |
3763 | * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue | |
3764 | * @wq: the target workqueue | |
3765 | * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() | |
3766 | * | |
3767 | * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA | |
3768 | * machines, this function maps a separate pwq to each NUMA node with | |
3769 | * possibles CPUs in @attrs->cpumask so that work items are affine to the | |
3770 | * NUMA node it was issued on. Older pwqs are released as in-flight work | |
3771 | * items finish. Note that a work item which repeatedly requeues itself | |
3772 | * back-to-back will stay on its current pwq. | |
3773 | * | |
3774 | * Performs GFP_KERNEL allocations. | |
3775 | * | |
3776 | * Return: 0 on success and -errno on failure. | |
3777 | */ | |
3778 | int apply_workqueue_attrs(struct workqueue_struct *wq, | |
3779 | const struct workqueue_attrs *attrs) | |
3780 | { | |
3781 | int ret; | |
3782 | ||
3783 | apply_wqattrs_lock(); | |
3784 | ret = apply_workqueue_attrs_locked(wq, attrs); | |
3785 | apply_wqattrs_unlock(); | |
3786 | ||
3787 | return ret; | |
3788 | } | |
3789 | ||
4c16bd32 TH |
3790 | /** |
3791 | * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug | |
3792 | * @wq: the target workqueue | |
3793 | * @cpu: the CPU coming up or going down | |
3794 | * @online: whether @cpu is coming up or going down | |
3795 | * | |
3796 | * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and | |
3797 | * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of | |
3798 | * @wq accordingly. | |
3799 | * | |
3800 | * If NUMA affinity can't be adjusted due to memory allocation failure, it | |
3801 | * falls back to @wq->dfl_pwq which may not be optimal but is always | |
3802 | * correct. | |
3803 | * | |
3804 | * Note that when the last allowed CPU of a NUMA node goes offline for a | |
3805 | * workqueue with a cpumask spanning multiple nodes, the workers which were | |
3806 | * already executing the work items for the workqueue will lose their CPU | |
3807 | * affinity and may execute on any CPU. This is similar to how per-cpu | |
3808 | * workqueues behave on CPU_DOWN. If a workqueue user wants strict | |
3809 | * affinity, it's the user's responsibility to flush the work item from | |
3810 | * CPU_DOWN_PREPARE. | |
3811 | */ | |
3812 | static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu, | |
3813 | bool online) | |
3814 | { | |
3815 | int node = cpu_to_node(cpu); | |
3816 | int cpu_off = online ? -1 : cpu; | |
3817 | struct pool_workqueue *old_pwq = NULL, *pwq; | |
3818 | struct workqueue_attrs *target_attrs; | |
3819 | cpumask_t *cpumask; | |
3820 | ||
3821 | lockdep_assert_held(&wq_pool_mutex); | |
3822 | ||
f7142ed4 LJ |
3823 | if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) || |
3824 | wq->unbound_attrs->no_numa) | |
4c16bd32 TH |
3825 | return; |
3826 | ||
3827 | /* | |
3828 | * We don't wanna alloc/free wq_attrs for each wq for each CPU. | |
3829 | * Let's use a preallocated one. The following buf is protected by | |
3830 | * CPU hotplug exclusion. | |
3831 | */ | |
3832 | target_attrs = wq_update_unbound_numa_attrs_buf; | |
3833 | cpumask = target_attrs->cpumask; | |
3834 | ||
4c16bd32 TH |
3835 | copy_workqueue_attrs(target_attrs, wq->unbound_attrs); |
3836 | pwq = unbound_pwq_by_node(wq, node); | |
3837 | ||
3838 | /* | |
3839 | * Let's determine what needs to be done. If the target cpumask is | |
042f7df1 LJ |
3840 | * different from the default pwq's, we need to compare it to @pwq's |
3841 | * and create a new one if they don't match. If the target cpumask | |
3842 | * equals the default pwq's, the default pwq should be used. | |
4c16bd32 | 3843 | */ |
042f7df1 | 3844 | if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) { |
4c16bd32 | 3845 | if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask)) |
f7142ed4 | 3846 | return; |
4c16bd32 | 3847 | } else { |
534a3fbb | 3848 | goto use_dfl_pwq; |
4c16bd32 TH |
3849 | } |
3850 | ||
4c16bd32 TH |
3851 | /* create a new pwq */ |
3852 | pwq = alloc_unbound_pwq(wq, target_attrs); | |
3853 | if (!pwq) { | |
2d916033 FF |
3854 | pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n", |
3855 | wq->name); | |
77f300b1 | 3856 | goto use_dfl_pwq; |
4c16bd32 TH |
3857 | } |
3858 | ||
f7142ed4 | 3859 | /* Install the new pwq. */ |
4c16bd32 TH |
3860 | mutex_lock(&wq->mutex); |
3861 | old_pwq = numa_pwq_tbl_install(wq, node, pwq); | |
3862 | goto out_unlock; | |
3863 | ||
3864 | use_dfl_pwq: | |
f7142ed4 | 3865 | mutex_lock(&wq->mutex); |
4c16bd32 TH |
3866 | spin_lock_irq(&wq->dfl_pwq->pool->lock); |
3867 | get_pwq(wq->dfl_pwq); | |
3868 | spin_unlock_irq(&wq->dfl_pwq->pool->lock); | |
3869 | old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq); | |
3870 | out_unlock: | |
3871 | mutex_unlock(&wq->mutex); | |
3872 | put_pwq_unlocked(old_pwq); | |
3873 | } | |
3874 | ||
30cdf249 | 3875 | static int alloc_and_link_pwqs(struct workqueue_struct *wq) |
0f900049 | 3876 | { |
49e3cf44 | 3877 | bool highpri = wq->flags & WQ_HIGHPRI; |
8a2b7538 | 3878 | int cpu, ret; |
30cdf249 TH |
3879 | |
3880 | if (!(wq->flags & WQ_UNBOUND)) { | |
420c0ddb TH |
3881 | wq->cpu_pwqs = alloc_percpu(struct pool_workqueue); |
3882 | if (!wq->cpu_pwqs) | |
30cdf249 TH |
3883 | return -ENOMEM; |
3884 | ||
3885 | for_each_possible_cpu(cpu) { | |
7fb98ea7 TH |
3886 | struct pool_workqueue *pwq = |
3887 | per_cpu_ptr(wq->cpu_pwqs, cpu); | |
7a62c2c8 | 3888 | struct worker_pool *cpu_pools = |
f02ae73a | 3889 | per_cpu(cpu_worker_pools, cpu); |
f3421797 | 3890 | |
f147f29e TH |
3891 | init_pwq(pwq, wq, &cpu_pools[highpri]); |
3892 | ||
3893 | mutex_lock(&wq->mutex); | |
1befcf30 | 3894 | link_pwq(pwq); |
f147f29e | 3895 | mutex_unlock(&wq->mutex); |
30cdf249 | 3896 | } |
9e8cd2f5 | 3897 | return 0; |
8a2b7538 TH |
3898 | } else if (wq->flags & __WQ_ORDERED) { |
3899 | ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]); | |
3900 | /* there should only be single pwq for ordering guarantee */ | |
3901 | WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node || | |
3902 | wq->pwqs.prev != &wq->dfl_pwq->pwqs_node), | |
3903 | "ordering guarantee broken for workqueue %s\n", wq->name); | |
3904 | return ret; | |
30cdf249 | 3905 | } else { |
9e8cd2f5 | 3906 | return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); |
30cdf249 | 3907 | } |
0f900049 TH |
3908 | } |
3909 | ||
f3421797 TH |
3910 | static int wq_clamp_max_active(int max_active, unsigned int flags, |
3911 | const char *name) | |
b71ab8c2 | 3912 | { |
f3421797 TH |
3913 | int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; |
3914 | ||
3915 | if (max_active < 1 || max_active > lim) | |
044c782c VI |
3916 | pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", |
3917 | max_active, name, 1, lim); | |
b71ab8c2 | 3918 | |
f3421797 | 3919 | return clamp_val(max_active, 1, lim); |
b71ab8c2 TH |
3920 | } |
3921 | ||
b196be89 | 3922 | struct workqueue_struct *__alloc_workqueue_key(const char *fmt, |
d320c038 TH |
3923 | unsigned int flags, |
3924 | int max_active, | |
3925 | struct lock_class_key *key, | |
b196be89 | 3926 | const char *lock_name, ...) |
1da177e4 | 3927 | { |
df2d5ae4 | 3928 | size_t tbl_size = 0; |
ecf6881f | 3929 | va_list args; |
1da177e4 | 3930 | struct workqueue_struct *wq; |
49e3cf44 | 3931 | struct pool_workqueue *pwq; |
b196be89 | 3932 | |
cee22a15 VK |
3933 | /* see the comment above the definition of WQ_POWER_EFFICIENT */ |
3934 | if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) | |
3935 | flags |= WQ_UNBOUND; | |
3936 | ||
ecf6881f | 3937 | /* allocate wq and format name */ |
df2d5ae4 | 3938 | if (flags & WQ_UNBOUND) |
ddcb57e2 | 3939 | tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]); |
df2d5ae4 TH |
3940 | |
3941 | wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL); | |
b196be89 | 3942 | if (!wq) |
d2c1d404 | 3943 | return NULL; |
b196be89 | 3944 | |
6029a918 TH |
3945 | if (flags & WQ_UNBOUND) { |
3946 | wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL); | |
3947 | if (!wq->unbound_attrs) | |
3948 | goto err_free_wq; | |
3949 | } | |
3950 | ||
ecf6881f TH |
3951 | va_start(args, lock_name); |
3952 | vsnprintf(wq->name, sizeof(wq->name), fmt, args); | |
b196be89 | 3953 | va_end(args); |
1da177e4 | 3954 | |
d320c038 | 3955 | max_active = max_active ?: WQ_DFL_ACTIVE; |
b196be89 | 3956 | max_active = wq_clamp_max_active(max_active, flags, wq->name); |
3af24433 | 3957 | |
b196be89 | 3958 | /* init wq */ |
97e37d7b | 3959 | wq->flags = flags; |
a0a1a5fd | 3960 | wq->saved_max_active = max_active; |
3c25a55d | 3961 | mutex_init(&wq->mutex); |
112202d9 | 3962 | atomic_set(&wq->nr_pwqs_to_flush, 0); |
30cdf249 | 3963 | INIT_LIST_HEAD(&wq->pwqs); |
73f53c4a TH |
3964 | INIT_LIST_HEAD(&wq->flusher_queue); |
3965 | INIT_LIST_HEAD(&wq->flusher_overflow); | |
493a1724 | 3966 | INIT_LIST_HEAD(&wq->maydays); |
502ca9d8 | 3967 | |
eb13ba87 | 3968 | lockdep_init_map(&wq->lockdep_map, lock_name, key, 0); |
cce1a165 | 3969 | INIT_LIST_HEAD(&wq->list); |
3af24433 | 3970 | |
30cdf249 | 3971 | if (alloc_and_link_pwqs(wq) < 0) |
d2c1d404 | 3972 | goto err_free_wq; |
1537663f | 3973 | |
493008a8 TH |
3974 | /* |
3975 | * Workqueues which may be used during memory reclaim should | |
3976 | * have a rescuer to guarantee forward progress. | |
3977 | */ | |
3978 | if (flags & WQ_MEM_RECLAIM) { | |
e22bee78 TH |
3979 | struct worker *rescuer; |
3980 | ||
f7537df5 | 3981 | rescuer = alloc_worker(NUMA_NO_NODE); |
e22bee78 | 3982 | if (!rescuer) |
d2c1d404 | 3983 | goto err_destroy; |
e22bee78 | 3984 | |
111c225a TH |
3985 | rescuer->rescue_wq = wq; |
3986 | rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", | |
b196be89 | 3987 | wq->name); |
d2c1d404 TH |
3988 | if (IS_ERR(rescuer->task)) { |
3989 | kfree(rescuer); | |
3990 | goto err_destroy; | |
3991 | } | |
e22bee78 | 3992 | |
d2c1d404 | 3993 | wq->rescuer = rescuer; |
25834c73 | 3994 | kthread_bind_mask(rescuer->task, cpu_possible_mask); |
e22bee78 | 3995 | wake_up_process(rescuer->task); |
3af24433 ON |
3996 | } |
3997 | ||
226223ab TH |
3998 | if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) |
3999 | goto err_destroy; | |
4000 | ||
a0a1a5fd | 4001 | /* |
68e13a67 LJ |
4002 | * wq_pool_mutex protects global freeze state and workqueues list. |
4003 | * Grab it, adjust max_active and add the new @wq to workqueues | |
4004 | * list. | |
a0a1a5fd | 4005 | */ |
68e13a67 | 4006 | mutex_lock(&wq_pool_mutex); |
a0a1a5fd | 4007 | |
a357fc03 | 4008 | mutex_lock(&wq->mutex); |
699ce097 TH |
4009 | for_each_pwq(pwq, wq) |
4010 | pwq_adjust_max_active(pwq); | |
a357fc03 | 4011 | mutex_unlock(&wq->mutex); |
a0a1a5fd | 4012 | |
e2dca7ad | 4013 | list_add_tail_rcu(&wq->list, &workqueues); |
a0a1a5fd | 4014 | |
68e13a67 | 4015 | mutex_unlock(&wq_pool_mutex); |
1537663f | 4016 | |
3af24433 | 4017 | return wq; |
d2c1d404 TH |
4018 | |
4019 | err_free_wq: | |
6029a918 | 4020 | free_workqueue_attrs(wq->unbound_attrs); |
d2c1d404 TH |
4021 | kfree(wq); |
4022 | return NULL; | |
4023 | err_destroy: | |
4024 | destroy_workqueue(wq); | |
4690c4ab | 4025 | return NULL; |
3af24433 | 4026 | } |
d320c038 | 4027 | EXPORT_SYMBOL_GPL(__alloc_workqueue_key); |
1da177e4 | 4028 | |
3af24433 ON |
4029 | /** |
4030 | * destroy_workqueue - safely terminate a workqueue | |
4031 | * @wq: target workqueue | |
4032 | * | |
4033 | * Safely destroy a workqueue. All work currently pending will be done first. | |
4034 | */ | |
4035 | void destroy_workqueue(struct workqueue_struct *wq) | |
4036 | { | |
49e3cf44 | 4037 | struct pool_workqueue *pwq; |
4c16bd32 | 4038 | int node; |
3af24433 | 4039 | |
9c5a2ba7 TH |
4040 | /* drain it before proceeding with destruction */ |
4041 | drain_workqueue(wq); | |
c8efcc25 | 4042 | |
6183c009 | 4043 | /* sanity checks */ |
b09f4fd3 | 4044 | mutex_lock(&wq->mutex); |
49e3cf44 | 4045 | for_each_pwq(pwq, wq) { |
6183c009 TH |
4046 | int i; |
4047 | ||
76af4d93 TH |
4048 | for (i = 0; i < WORK_NR_COLORS; i++) { |
4049 | if (WARN_ON(pwq->nr_in_flight[i])) { | |
b09f4fd3 | 4050 | mutex_unlock(&wq->mutex); |
6183c009 | 4051 | return; |
76af4d93 TH |
4052 | } |
4053 | } | |
4054 | ||
5c529597 | 4055 | if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) || |
8864b4e5 | 4056 | WARN_ON(pwq->nr_active) || |
76af4d93 | 4057 | WARN_ON(!list_empty(&pwq->delayed_works))) { |
b09f4fd3 | 4058 | mutex_unlock(&wq->mutex); |
6183c009 | 4059 | return; |
76af4d93 | 4060 | } |
6183c009 | 4061 | } |
b09f4fd3 | 4062 | mutex_unlock(&wq->mutex); |
6183c009 | 4063 | |
a0a1a5fd TH |
4064 | /* |
4065 | * wq list is used to freeze wq, remove from list after | |
4066 | * flushing is complete in case freeze races us. | |
4067 | */ | |
68e13a67 | 4068 | mutex_lock(&wq_pool_mutex); |
e2dca7ad | 4069 | list_del_rcu(&wq->list); |
68e13a67 | 4070 | mutex_unlock(&wq_pool_mutex); |
3af24433 | 4071 | |
226223ab TH |
4072 | workqueue_sysfs_unregister(wq); |
4073 | ||
e2dca7ad | 4074 | if (wq->rescuer) |
e22bee78 | 4075 | kthread_stop(wq->rescuer->task); |
e22bee78 | 4076 | |
8864b4e5 TH |
4077 | if (!(wq->flags & WQ_UNBOUND)) { |
4078 | /* | |
4079 | * The base ref is never dropped on per-cpu pwqs. Directly | |
e2dca7ad | 4080 | * schedule RCU free. |
8864b4e5 | 4081 | */ |
e2dca7ad | 4082 | call_rcu_sched(&wq->rcu, rcu_free_wq); |
8864b4e5 TH |
4083 | } else { |
4084 | /* | |
4085 | * We're the sole accessor of @wq at this point. Directly | |
4c16bd32 TH |
4086 | * access numa_pwq_tbl[] and dfl_pwq to put the base refs. |
4087 | * @wq will be freed when the last pwq is released. | |
8864b4e5 | 4088 | */ |
4c16bd32 TH |
4089 | for_each_node(node) { |
4090 | pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); | |
4091 | RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL); | |
4092 | put_pwq_unlocked(pwq); | |
4093 | } | |
4094 | ||
4095 | /* | |
4096 | * Put dfl_pwq. @wq may be freed any time after dfl_pwq is | |
4097 | * put. Don't access it afterwards. | |
4098 | */ | |
4099 | pwq = wq->dfl_pwq; | |
4100 | wq->dfl_pwq = NULL; | |
dce90d47 | 4101 | put_pwq_unlocked(pwq); |
29c91e99 | 4102 | } |
3af24433 ON |
4103 | } |
4104 | EXPORT_SYMBOL_GPL(destroy_workqueue); | |
4105 | ||
dcd989cb TH |
4106 | /** |
4107 | * workqueue_set_max_active - adjust max_active of a workqueue | |
4108 | * @wq: target workqueue | |
4109 | * @max_active: new max_active value. | |
4110 | * | |
4111 | * Set max_active of @wq to @max_active. | |
4112 | * | |
4113 | * CONTEXT: | |
4114 | * Don't call from IRQ context. | |
4115 | */ | |
4116 | void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) | |
4117 | { | |
49e3cf44 | 4118 | struct pool_workqueue *pwq; |
dcd989cb | 4119 | |
8719dcea TH |
4120 | /* disallow meddling with max_active for ordered workqueues */ |
4121 | if (WARN_ON(wq->flags & __WQ_ORDERED)) | |
4122 | return; | |
4123 | ||
f3421797 | 4124 | max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); |
dcd989cb | 4125 | |
a357fc03 | 4126 | mutex_lock(&wq->mutex); |
dcd989cb TH |
4127 | |
4128 | wq->saved_max_active = max_active; | |
4129 | ||
699ce097 TH |
4130 | for_each_pwq(pwq, wq) |
4131 | pwq_adjust_max_active(pwq); | |
93981800 | 4132 | |
a357fc03 | 4133 | mutex_unlock(&wq->mutex); |
15316ba8 | 4134 | } |
dcd989cb | 4135 | EXPORT_SYMBOL_GPL(workqueue_set_max_active); |
15316ba8 | 4136 | |
e6267616 TH |
4137 | /** |
4138 | * current_is_workqueue_rescuer - is %current workqueue rescuer? | |
4139 | * | |
4140 | * Determine whether %current is a workqueue rescuer. Can be used from | |
4141 | * work functions to determine whether it's being run off the rescuer task. | |
d185af30 YB |
4142 | * |
4143 | * Return: %true if %current is a workqueue rescuer. %false otherwise. | |
e6267616 TH |
4144 | */ |
4145 | bool current_is_workqueue_rescuer(void) | |
4146 | { | |
4147 | struct worker *worker = current_wq_worker(); | |
4148 | ||
6a092dfd | 4149 | return worker && worker->rescue_wq; |
e6267616 TH |
4150 | } |
4151 | ||
eef6a7d5 | 4152 | /** |
dcd989cb TH |
4153 | * workqueue_congested - test whether a workqueue is congested |
4154 | * @cpu: CPU in question | |
4155 | * @wq: target workqueue | |
eef6a7d5 | 4156 | * |
dcd989cb TH |
4157 | * Test whether @wq's cpu workqueue for @cpu is congested. There is |
4158 | * no synchronization around this function and the test result is | |
4159 | * unreliable and only useful as advisory hints or for debugging. | |
eef6a7d5 | 4160 | * |
d3251859 TH |
4161 | * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. |
4162 | * Note that both per-cpu and unbound workqueues may be associated with | |
4163 | * multiple pool_workqueues which have separate congested states. A | |
4164 | * workqueue being congested on one CPU doesn't mean the workqueue is also | |
4165 | * contested on other CPUs / NUMA nodes. | |
4166 | * | |
d185af30 | 4167 | * Return: |
dcd989cb | 4168 | * %true if congested, %false otherwise. |
eef6a7d5 | 4169 | */ |
d84ff051 | 4170 | bool workqueue_congested(int cpu, struct workqueue_struct *wq) |
1da177e4 | 4171 | { |
7fb98ea7 | 4172 | struct pool_workqueue *pwq; |
76af4d93 TH |
4173 | bool ret; |
4174 | ||
88109453 | 4175 | rcu_read_lock_sched(); |
7fb98ea7 | 4176 | |
d3251859 TH |
4177 | if (cpu == WORK_CPU_UNBOUND) |
4178 | cpu = smp_processor_id(); | |
4179 | ||
7fb98ea7 TH |
4180 | if (!(wq->flags & WQ_UNBOUND)) |
4181 | pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); | |
4182 | else | |
df2d5ae4 | 4183 | pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); |
dcd989cb | 4184 | |
76af4d93 | 4185 | ret = !list_empty(&pwq->delayed_works); |
88109453 | 4186 | rcu_read_unlock_sched(); |
76af4d93 TH |
4187 | |
4188 | return ret; | |
1da177e4 | 4189 | } |
dcd989cb | 4190 | EXPORT_SYMBOL_GPL(workqueue_congested); |
1da177e4 | 4191 | |
dcd989cb TH |
4192 | /** |
4193 | * work_busy - test whether a work is currently pending or running | |
4194 | * @work: the work to be tested | |
4195 | * | |
4196 | * Test whether @work is currently pending or running. There is no | |
4197 | * synchronization around this function and the test result is | |
4198 | * unreliable and only useful as advisory hints or for debugging. | |
dcd989cb | 4199 | * |
d185af30 | 4200 | * Return: |
dcd989cb TH |
4201 | * OR'd bitmask of WORK_BUSY_* bits. |
4202 | */ | |
4203 | unsigned int work_busy(struct work_struct *work) | |
1da177e4 | 4204 | { |
fa1b54e6 | 4205 | struct worker_pool *pool; |
dcd989cb TH |
4206 | unsigned long flags; |
4207 | unsigned int ret = 0; | |
1da177e4 | 4208 | |
dcd989cb TH |
4209 | if (work_pending(work)) |
4210 | ret |= WORK_BUSY_PENDING; | |
1da177e4 | 4211 | |
fa1b54e6 TH |
4212 | local_irq_save(flags); |
4213 | pool = get_work_pool(work); | |
038366c5 | 4214 | if (pool) { |
fa1b54e6 | 4215 | spin_lock(&pool->lock); |
038366c5 LJ |
4216 | if (find_worker_executing_work(pool, work)) |
4217 | ret |= WORK_BUSY_RUNNING; | |
fa1b54e6 | 4218 | spin_unlock(&pool->lock); |
038366c5 | 4219 | } |
fa1b54e6 | 4220 | local_irq_restore(flags); |
1da177e4 | 4221 | |
dcd989cb | 4222 | return ret; |
1da177e4 | 4223 | } |
dcd989cb | 4224 | EXPORT_SYMBOL_GPL(work_busy); |
1da177e4 | 4225 | |
3d1cb205 TH |
4226 | /** |
4227 | * set_worker_desc - set description for the current work item | |
4228 | * @fmt: printf-style format string | |
4229 | * @...: arguments for the format string | |
4230 | * | |
4231 | * This function can be called by a running work function to describe what | |
4232 | * the work item is about. If the worker task gets dumped, this | |
4233 | * information will be printed out together to help debugging. The | |
4234 | * description can be at most WORKER_DESC_LEN including the trailing '\0'. | |
4235 | */ | |
4236 | void set_worker_desc(const char *fmt, ...) | |
4237 | { | |
4238 | struct worker *worker = current_wq_worker(); | |
4239 | va_list args; | |
4240 | ||
4241 | if (worker) { | |
4242 | va_start(args, fmt); | |
4243 | vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); | |
4244 | va_end(args); | |
4245 | worker->desc_valid = true; | |
4246 | } | |
4247 | } | |
4248 | ||
4249 | /** | |
4250 | * print_worker_info - print out worker information and description | |
4251 | * @log_lvl: the log level to use when printing | |
4252 | * @task: target task | |
4253 | * | |
4254 | * If @task is a worker and currently executing a work item, print out the | |
4255 | * name of the workqueue being serviced and worker description set with | |
4256 | * set_worker_desc() by the currently executing work item. | |
4257 | * | |
4258 | * This function can be safely called on any task as long as the | |
4259 | * task_struct itself is accessible. While safe, this function isn't | |
4260 | * synchronized and may print out mixups or garbages of limited length. | |
4261 | */ | |
4262 | void print_worker_info(const char *log_lvl, struct task_struct *task) | |
4263 | { | |
4264 | work_func_t *fn = NULL; | |
4265 | char name[WQ_NAME_LEN] = { }; | |
4266 | char desc[WORKER_DESC_LEN] = { }; | |
4267 | struct pool_workqueue *pwq = NULL; | |
4268 | struct workqueue_struct *wq = NULL; | |
4269 | bool desc_valid = false; | |
4270 | struct worker *worker; | |
4271 | ||
4272 | if (!(task->flags & PF_WQ_WORKER)) | |
4273 | return; | |
4274 | ||
4275 | /* | |
4276 | * This function is called without any synchronization and @task | |
4277 | * could be in any state. Be careful with dereferences. | |
4278 | */ | |
4279 | worker = probe_kthread_data(task); | |
4280 | ||
4281 | /* | |
4282 | * Carefully copy the associated workqueue's workfn and name. Keep | |
4283 | * the original last '\0' in case the original contains garbage. | |
4284 | */ | |
4285 | probe_kernel_read(&fn, &worker->current_func, sizeof(fn)); | |
4286 | probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq)); | |
4287 | probe_kernel_read(&wq, &pwq->wq, sizeof(wq)); | |
4288 | probe_kernel_read(name, wq->name, sizeof(name) - 1); | |
4289 | ||
4290 | /* copy worker description */ | |
4291 | probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid)); | |
4292 | if (desc_valid) | |
4293 | probe_kernel_read(desc, worker->desc, sizeof(desc) - 1); | |
4294 | ||
4295 | if (fn || name[0] || desc[0]) { | |
4296 | printk("%sWorkqueue: %s %pf", log_lvl, name, fn); | |
4297 | if (desc[0]) | |
4298 | pr_cont(" (%s)", desc); | |
4299 | pr_cont("\n"); | |
4300 | } | |
4301 | } | |
4302 | ||
3494fc30 TH |
4303 | static void pr_cont_pool_info(struct worker_pool *pool) |
4304 | { | |
4305 | pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask); | |
4306 | if (pool->node != NUMA_NO_NODE) | |
4307 | pr_cont(" node=%d", pool->node); | |
4308 | pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice); | |
4309 | } | |
4310 | ||
4311 | static void pr_cont_work(bool comma, struct work_struct *work) | |
4312 | { | |
4313 | if (work->func == wq_barrier_func) { | |
4314 | struct wq_barrier *barr; | |
4315 | ||
4316 | barr = container_of(work, struct wq_barrier, work); | |
4317 | ||
4318 | pr_cont("%s BAR(%d)", comma ? "," : "", | |
4319 | task_pid_nr(barr->task)); | |
4320 | } else { | |
4321 | pr_cont("%s %pf", comma ? "," : "", work->func); | |
4322 | } | |
4323 | } | |
4324 | ||
4325 | static void show_pwq(struct pool_workqueue *pwq) | |
4326 | { | |
4327 | struct worker_pool *pool = pwq->pool; | |
4328 | struct work_struct *work; | |
4329 | struct worker *worker; | |
4330 | bool has_in_flight = false, has_pending = false; | |
4331 | int bkt; | |
4332 | ||
4333 | pr_info(" pwq %d:", pool->id); | |
4334 | pr_cont_pool_info(pool); | |
4335 | ||
4336 | pr_cont(" active=%d/%d%s\n", pwq->nr_active, pwq->max_active, | |
4337 | !list_empty(&pwq->mayday_node) ? " MAYDAY" : ""); | |
4338 | ||
4339 | hash_for_each(pool->busy_hash, bkt, worker, hentry) { | |
4340 | if (worker->current_pwq == pwq) { | |
4341 | has_in_flight = true; | |
4342 | break; | |
4343 | } | |
4344 | } | |
4345 | if (has_in_flight) { | |
4346 | bool comma = false; | |
4347 | ||
4348 | pr_info(" in-flight:"); | |
4349 | hash_for_each(pool->busy_hash, bkt, worker, hentry) { | |
4350 | if (worker->current_pwq != pwq) | |
4351 | continue; | |
4352 | ||
4353 | pr_cont("%s %d%s:%pf", comma ? "," : "", | |
4354 | task_pid_nr(worker->task), | |
4355 | worker == pwq->wq->rescuer ? "(RESCUER)" : "", | |
4356 | worker->current_func); | |
4357 | list_for_each_entry(work, &worker->scheduled, entry) | |
4358 | pr_cont_work(false, work); | |
4359 | comma = true; | |
4360 | } | |
4361 | pr_cont("\n"); | |
4362 | } | |
4363 | ||
4364 | list_for_each_entry(work, &pool->worklist, entry) { | |
4365 | if (get_work_pwq(work) == pwq) { | |
4366 | has_pending = true; | |
4367 | break; | |
4368 | } | |
4369 | } | |
4370 | if (has_pending) { | |
4371 | bool comma = false; | |
4372 | ||
4373 | pr_info(" pending:"); | |
4374 | list_for_each_entry(work, &pool->worklist, entry) { | |
4375 | if (get_work_pwq(work) != pwq) | |
4376 | continue; | |
4377 | ||
4378 | pr_cont_work(comma, work); | |
4379 | comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); | |
4380 | } | |
4381 | pr_cont("\n"); | |
4382 | } | |
4383 | ||
4384 | if (!list_empty(&pwq->delayed_works)) { | |
4385 | bool comma = false; | |
4386 | ||
4387 | pr_info(" delayed:"); | |
4388 | list_for_each_entry(work, &pwq->delayed_works, entry) { | |
4389 | pr_cont_work(comma, work); | |
4390 | comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); | |
4391 | } | |
4392 | pr_cont("\n"); | |
4393 | } | |
4394 | } | |
4395 | ||
4396 | /** | |
4397 | * show_workqueue_state - dump workqueue state | |
4398 | * | |
4399 | * Called from a sysrq handler and prints out all busy workqueues and | |
4400 | * pools. | |
4401 | */ | |
4402 | void show_workqueue_state(void) | |
4403 | { | |
4404 | struct workqueue_struct *wq; | |
4405 | struct worker_pool *pool; | |
4406 | unsigned long flags; | |
4407 | int pi; | |
4408 | ||
4409 | rcu_read_lock_sched(); | |
4410 | ||
4411 | pr_info("Showing busy workqueues and worker pools:\n"); | |
4412 | ||
4413 | list_for_each_entry_rcu(wq, &workqueues, list) { | |
4414 | struct pool_workqueue *pwq; | |
4415 | bool idle = true; | |
4416 | ||
4417 | for_each_pwq(pwq, wq) { | |
4418 | if (pwq->nr_active || !list_empty(&pwq->delayed_works)) { | |
4419 | idle = false; | |
4420 | break; | |
4421 | } | |
4422 | } | |
4423 | if (idle) | |
4424 | continue; | |
4425 | ||
4426 | pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags); | |
4427 | ||
4428 | for_each_pwq(pwq, wq) { | |
4429 | spin_lock_irqsave(&pwq->pool->lock, flags); | |
4430 | if (pwq->nr_active || !list_empty(&pwq->delayed_works)) | |
4431 | show_pwq(pwq); | |
4432 | spin_unlock_irqrestore(&pwq->pool->lock, flags); | |
4433 | } | |
4434 | } | |
4435 | ||
4436 | for_each_pool(pool, pi) { | |
4437 | struct worker *worker; | |
4438 | bool first = true; | |
4439 | ||
4440 | spin_lock_irqsave(&pool->lock, flags); | |
4441 | if (pool->nr_workers == pool->nr_idle) | |
4442 | goto next_pool; | |
4443 | ||
4444 | pr_info("pool %d:", pool->id); | |
4445 | pr_cont_pool_info(pool); | |
82607adc TH |
4446 | pr_cont(" hung=%us workers=%d", |
4447 | jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000, | |
4448 | pool->nr_workers); | |
3494fc30 TH |
4449 | if (pool->manager) |
4450 | pr_cont(" manager: %d", | |
4451 | task_pid_nr(pool->manager->task)); | |
4452 | list_for_each_entry(worker, &pool->idle_list, entry) { | |
4453 | pr_cont(" %s%d", first ? "idle: " : "", | |
4454 | task_pid_nr(worker->task)); | |
4455 | first = false; | |
4456 | } | |
4457 | pr_cont("\n"); | |
4458 | next_pool: | |
4459 | spin_unlock_irqrestore(&pool->lock, flags); | |
4460 | } | |
4461 | ||
4462 | rcu_read_unlock_sched(); | |
4463 | } | |
4464 | ||
db7bccf4 TH |
4465 | /* |
4466 | * CPU hotplug. | |
4467 | * | |
e22bee78 | 4468 | * There are two challenges in supporting CPU hotplug. Firstly, there |
112202d9 | 4469 | * are a lot of assumptions on strong associations among work, pwq and |
706026c2 | 4470 | * pool which make migrating pending and scheduled works very |
e22bee78 | 4471 | * difficult to implement without impacting hot paths. Secondly, |
94cf58bb | 4472 | * worker pools serve mix of short, long and very long running works making |
e22bee78 TH |
4473 | * blocked draining impractical. |
4474 | * | |
24647570 | 4475 | * This is solved by allowing the pools to be disassociated from the CPU |
628c78e7 TH |
4476 | * running as an unbound one and allowing it to be reattached later if the |
4477 | * cpu comes back online. | |
db7bccf4 | 4478 | */ |
1da177e4 | 4479 | |
706026c2 | 4480 | static void wq_unbind_fn(struct work_struct *work) |
3af24433 | 4481 | { |
38db41d9 | 4482 | int cpu = smp_processor_id(); |
4ce62e9e | 4483 | struct worker_pool *pool; |
db7bccf4 | 4484 | struct worker *worker; |
3af24433 | 4485 | |
f02ae73a | 4486 | for_each_cpu_worker_pool(pool, cpu) { |
92f9c5c4 | 4487 | mutex_lock(&pool->attach_mutex); |
94cf58bb | 4488 | spin_lock_irq(&pool->lock); |
3af24433 | 4489 | |
94cf58bb | 4490 | /* |
92f9c5c4 | 4491 | * We've blocked all attach/detach operations. Make all workers |
94cf58bb TH |
4492 | * unbound and set DISASSOCIATED. Before this, all workers |
4493 | * except for the ones which are still executing works from | |
4494 | * before the last CPU down must be on the cpu. After | |
4495 | * this, they may become diasporas. | |
4496 | */ | |
da028469 | 4497 | for_each_pool_worker(worker, pool) |
c9e7cf27 | 4498 | worker->flags |= WORKER_UNBOUND; |
06ba38a9 | 4499 | |
24647570 | 4500 | pool->flags |= POOL_DISASSOCIATED; |
f2d5a0ee | 4501 | |
94cf58bb | 4502 | spin_unlock_irq(&pool->lock); |
92f9c5c4 | 4503 | mutex_unlock(&pool->attach_mutex); |
628c78e7 | 4504 | |
eb283428 LJ |
4505 | /* |
4506 | * Call schedule() so that we cross rq->lock and thus can | |
4507 | * guarantee sched callbacks see the %WORKER_UNBOUND flag. | |
4508 | * This is necessary as scheduler callbacks may be invoked | |
4509 | * from other cpus. | |
4510 | */ | |
4511 | schedule(); | |
06ba38a9 | 4512 | |
eb283428 LJ |
4513 | /* |
4514 | * Sched callbacks are disabled now. Zap nr_running. | |
4515 | * After this, nr_running stays zero and need_more_worker() | |
4516 | * and keep_working() are always true as long as the | |
4517 | * worklist is not empty. This pool now behaves as an | |
4518 | * unbound (in terms of concurrency management) pool which | |
4519 | * are served by workers tied to the pool. | |
4520 | */ | |
e19e397a | 4521 | atomic_set(&pool->nr_running, 0); |
eb283428 LJ |
4522 | |
4523 | /* | |
4524 | * With concurrency management just turned off, a busy | |
4525 | * worker blocking could lead to lengthy stalls. Kick off | |
4526 | * unbound chain execution of currently pending work items. | |
4527 | */ | |
4528 | spin_lock_irq(&pool->lock); | |
4529 | wake_up_worker(pool); | |
4530 | spin_unlock_irq(&pool->lock); | |
4531 | } | |
3af24433 | 4532 | } |
3af24433 | 4533 | |
bd7c089e TH |
4534 | /** |
4535 | * rebind_workers - rebind all workers of a pool to the associated CPU | |
4536 | * @pool: pool of interest | |
4537 | * | |
a9ab775b | 4538 | * @pool->cpu is coming online. Rebind all workers to the CPU. |
bd7c089e TH |
4539 | */ |
4540 | static void rebind_workers(struct worker_pool *pool) | |
4541 | { | |
a9ab775b | 4542 | struct worker *worker; |
bd7c089e | 4543 | |
92f9c5c4 | 4544 | lockdep_assert_held(&pool->attach_mutex); |
bd7c089e | 4545 | |
a9ab775b TH |
4546 | /* |
4547 | * Restore CPU affinity of all workers. As all idle workers should | |
4548 | * be on the run-queue of the associated CPU before any local | |
402dd89d | 4549 | * wake-ups for concurrency management happen, restore CPU affinity |
a9ab775b TH |
4550 | * of all workers first and then clear UNBOUND. As we're called |
4551 | * from CPU_ONLINE, the following shouldn't fail. | |
4552 | */ | |
da028469 | 4553 | for_each_pool_worker(worker, pool) |
a9ab775b TH |
4554 | WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, |
4555 | pool->attrs->cpumask) < 0); | |
bd7c089e | 4556 | |
a9ab775b | 4557 | spin_lock_irq(&pool->lock); |
3de5e884 | 4558 | pool->flags &= ~POOL_DISASSOCIATED; |
bd7c089e | 4559 | |
da028469 | 4560 | for_each_pool_worker(worker, pool) { |
a9ab775b | 4561 | unsigned int worker_flags = worker->flags; |
bd7c089e TH |
4562 | |
4563 | /* | |
a9ab775b TH |
4564 | * A bound idle worker should actually be on the runqueue |
4565 | * of the associated CPU for local wake-ups targeting it to | |
4566 | * work. Kick all idle workers so that they migrate to the | |
4567 | * associated CPU. Doing this in the same loop as | |
4568 | * replacing UNBOUND with REBOUND is safe as no worker will | |
4569 | * be bound before @pool->lock is released. | |
bd7c089e | 4570 | */ |
a9ab775b TH |
4571 | if (worker_flags & WORKER_IDLE) |
4572 | wake_up_process(worker->task); | |
bd7c089e | 4573 | |
a9ab775b TH |
4574 | /* |
4575 | * We want to clear UNBOUND but can't directly call | |
4576 | * worker_clr_flags() or adjust nr_running. Atomically | |
4577 | * replace UNBOUND with another NOT_RUNNING flag REBOUND. | |
4578 | * @worker will clear REBOUND using worker_clr_flags() when | |
4579 | * it initiates the next execution cycle thus restoring | |
4580 | * concurrency management. Note that when or whether | |
4581 | * @worker clears REBOUND doesn't affect correctness. | |
4582 | * | |
4583 | * ACCESS_ONCE() is necessary because @worker->flags may be | |
4584 | * tested without holding any lock in | |
4585 | * wq_worker_waking_up(). Without it, NOT_RUNNING test may | |
4586 | * fail incorrectly leading to premature concurrency | |
4587 | * management operations. | |
4588 | */ | |
4589 | WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); | |
4590 | worker_flags |= WORKER_REBOUND; | |
4591 | worker_flags &= ~WORKER_UNBOUND; | |
4592 | ACCESS_ONCE(worker->flags) = worker_flags; | |
bd7c089e | 4593 | } |
a9ab775b TH |
4594 | |
4595 | spin_unlock_irq(&pool->lock); | |
bd7c089e TH |
4596 | } |
4597 | ||
7dbc725e TH |
4598 | /** |
4599 | * restore_unbound_workers_cpumask - restore cpumask of unbound workers | |
4600 | * @pool: unbound pool of interest | |
4601 | * @cpu: the CPU which is coming up | |
4602 | * | |
4603 | * An unbound pool may end up with a cpumask which doesn't have any online | |
4604 | * CPUs. When a worker of such pool get scheduled, the scheduler resets | |
4605 | * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any | |
4606 | * online CPU before, cpus_allowed of all its workers should be restored. | |
4607 | */ | |
4608 | static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) | |
4609 | { | |
4610 | static cpumask_t cpumask; | |
4611 | struct worker *worker; | |
7dbc725e | 4612 | |
92f9c5c4 | 4613 | lockdep_assert_held(&pool->attach_mutex); |
7dbc725e TH |
4614 | |
4615 | /* is @cpu allowed for @pool? */ | |
4616 | if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) | |
4617 | return; | |
4618 | ||
4619 | /* is @cpu the only online CPU? */ | |
4620 | cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); | |
4621 | if (cpumask_weight(&cpumask) != 1) | |
4622 | return; | |
4623 | ||
4624 | /* as we're called from CPU_ONLINE, the following shouldn't fail */ | |
da028469 | 4625 | for_each_pool_worker(worker, pool) |
7dbc725e TH |
4626 | WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, |
4627 | pool->attrs->cpumask) < 0); | |
4628 | } | |
4629 | ||
8db25e78 TH |
4630 | /* |
4631 | * Workqueues should be brought up before normal priority CPU notifiers. | |
4632 | * This will be registered high priority CPU notifier. | |
4633 | */ | |
0db0628d | 4634 | static int workqueue_cpu_up_callback(struct notifier_block *nfb, |
8db25e78 TH |
4635 | unsigned long action, |
4636 | void *hcpu) | |
3af24433 | 4637 | { |
d84ff051 | 4638 | int cpu = (unsigned long)hcpu; |
4ce62e9e | 4639 | struct worker_pool *pool; |
4c16bd32 | 4640 | struct workqueue_struct *wq; |
7dbc725e | 4641 | int pi; |
3ce63377 | 4642 | |
8db25e78 | 4643 | switch (action & ~CPU_TASKS_FROZEN) { |
3af24433 | 4644 | case CPU_UP_PREPARE: |
f02ae73a | 4645 | for_each_cpu_worker_pool(pool, cpu) { |
3ce63377 TH |
4646 | if (pool->nr_workers) |
4647 | continue; | |
051e1850 | 4648 | if (!create_worker(pool)) |
3ce63377 | 4649 | return NOTIFY_BAD; |
3af24433 | 4650 | } |
8db25e78 | 4651 | break; |
3af24433 | 4652 | |
db7bccf4 TH |
4653 | case CPU_DOWN_FAILED: |
4654 | case CPU_ONLINE: | |
68e13a67 | 4655 | mutex_lock(&wq_pool_mutex); |
7dbc725e TH |
4656 | |
4657 | for_each_pool(pool, pi) { | |
92f9c5c4 | 4658 | mutex_lock(&pool->attach_mutex); |
94cf58bb | 4659 | |
f05b558d | 4660 | if (pool->cpu == cpu) |
7dbc725e | 4661 | rebind_workers(pool); |
f05b558d | 4662 | else if (pool->cpu < 0) |
7dbc725e | 4663 | restore_unbound_workers_cpumask(pool, cpu); |
94cf58bb | 4664 | |
6ba94429 FW |
4665 | mutex_unlock(&pool->attach_mutex); |
4666 | } | |
4667 | ||
4668 | /* update NUMA affinity of unbound workqueues */ | |
4669 | list_for_each_entry(wq, &workqueues, list) | |
4670 | wq_update_unbound_numa(wq, cpu, true); | |
4671 | ||
4672 | mutex_unlock(&wq_pool_mutex); | |
4673 | break; | |
4674 | } | |
4675 | return NOTIFY_OK; | |
4676 | } | |
4677 | ||
4678 | /* | |
4679 | * Workqueues should be brought down after normal priority CPU notifiers. | |
4680 | * This will be registered as low priority CPU notifier. | |
4681 | */ | |
4682 | static int workqueue_cpu_down_callback(struct notifier_block *nfb, | |
4683 | unsigned long action, | |
4684 | void *hcpu) | |
4685 | { | |
4686 | int cpu = (unsigned long)hcpu; | |
4687 | struct work_struct unbind_work; | |
4688 | struct workqueue_struct *wq; | |
4689 | ||
4690 | switch (action & ~CPU_TASKS_FROZEN) { | |
4691 | case CPU_DOWN_PREPARE: | |
4692 | /* unbinding per-cpu workers should happen on the local CPU */ | |
4693 | INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn); | |
4694 | queue_work_on(cpu, system_highpri_wq, &unbind_work); | |
4695 | ||
4696 | /* update NUMA affinity of unbound workqueues */ | |
4697 | mutex_lock(&wq_pool_mutex); | |
4698 | list_for_each_entry(wq, &workqueues, list) | |
4699 | wq_update_unbound_numa(wq, cpu, false); | |
4700 | mutex_unlock(&wq_pool_mutex); | |
4701 | ||
4702 | /* wait for per-cpu unbinding to finish */ | |
4703 | flush_work(&unbind_work); | |
4704 | destroy_work_on_stack(&unbind_work); | |
4705 | break; | |
4706 | } | |
4707 | return NOTIFY_OK; | |
4708 | } | |
4709 | ||
4710 | #ifdef CONFIG_SMP | |
4711 | ||
4712 | struct work_for_cpu { | |
4713 | struct work_struct work; | |
4714 | long (*fn)(void *); | |
4715 | void *arg; | |
4716 | long ret; | |
4717 | }; | |
4718 | ||
4719 | static void work_for_cpu_fn(struct work_struct *work) | |
4720 | { | |
4721 | struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); | |
4722 | ||
4723 | wfc->ret = wfc->fn(wfc->arg); | |
4724 | } | |
4725 | ||
4726 | /** | |
4727 | * work_on_cpu - run a function in user context on a particular cpu | |
4728 | * @cpu: the cpu to run on | |
4729 | * @fn: the function to run | |
4730 | * @arg: the function arg | |
4731 | * | |
4732 | * It is up to the caller to ensure that the cpu doesn't go offline. | |
4733 | * The caller must not hold any locks which would prevent @fn from completing. | |
4734 | * | |
4735 | * Return: The value @fn returns. | |
4736 | */ | |
4737 | long work_on_cpu(int cpu, long (*fn)(void *), void *arg) | |
4738 | { | |
4739 | struct work_for_cpu wfc = { .fn = fn, .arg = arg }; | |
4740 | ||
4741 | INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); | |
4742 | schedule_work_on(cpu, &wfc.work); | |
4743 | flush_work(&wfc.work); | |
4744 | destroy_work_on_stack(&wfc.work); | |
4745 | return wfc.ret; | |
4746 | } | |
4747 | EXPORT_SYMBOL_GPL(work_on_cpu); | |
4748 | #endif /* CONFIG_SMP */ | |
4749 | ||
4750 | #ifdef CONFIG_FREEZER | |
4751 | ||
4752 | /** | |
4753 | * freeze_workqueues_begin - begin freezing workqueues | |
4754 | * | |
4755 | * Start freezing workqueues. After this function returns, all freezable | |
4756 | * workqueues will queue new works to their delayed_works list instead of | |
4757 | * pool->worklist. | |
4758 | * | |
4759 | * CONTEXT: | |
4760 | * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. | |
4761 | */ | |
4762 | void freeze_workqueues_begin(void) | |
4763 | { | |
4764 | struct workqueue_struct *wq; | |
4765 | struct pool_workqueue *pwq; | |
4766 | ||
4767 | mutex_lock(&wq_pool_mutex); | |
4768 | ||
4769 | WARN_ON_ONCE(workqueue_freezing); | |
4770 | workqueue_freezing = true; | |
4771 | ||
4772 | list_for_each_entry(wq, &workqueues, list) { | |
4773 | mutex_lock(&wq->mutex); | |
4774 | for_each_pwq(pwq, wq) | |
4775 | pwq_adjust_max_active(pwq); | |
4776 | mutex_unlock(&wq->mutex); | |
4777 | } | |
4778 | ||
4779 | mutex_unlock(&wq_pool_mutex); | |
4780 | } | |
4781 | ||
4782 | /** | |
4783 | * freeze_workqueues_busy - are freezable workqueues still busy? | |
4784 | * | |
4785 | * Check whether freezing is complete. This function must be called | |
4786 | * between freeze_workqueues_begin() and thaw_workqueues(). | |
4787 | * | |
4788 | * CONTEXT: | |
4789 | * Grabs and releases wq_pool_mutex. | |
4790 | * | |
4791 | * Return: | |
4792 | * %true if some freezable workqueues are still busy. %false if freezing | |
4793 | * is complete. | |
4794 | */ | |
4795 | bool freeze_workqueues_busy(void) | |
4796 | { | |
4797 | bool busy = false; | |
4798 | struct workqueue_struct *wq; | |
4799 | struct pool_workqueue *pwq; | |
4800 | ||
4801 | mutex_lock(&wq_pool_mutex); | |
4802 | ||
4803 | WARN_ON_ONCE(!workqueue_freezing); | |
4804 | ||
4805 | list_for_each_entry(wq, &workqueues, list) { | |
4806 | if (!(wq->flags & WQ_FREEZABLE)) | |
4807 | continue; | |
4808 | /* | |
4809 | * nr_active is monotonically decreasing. It's safe | |
4810 | * to peek without lock. | |
4811 | */ | |
4812 | rcu_read_lock_sched(); | |
4813 | for_each_pwq(pwq, wq) { | |
4814 | WARN_ON_ONCE(pwq->nr_active < 0); | |
4815 | if (pwq->nr_active) { | |
4816 | busy = true; | |
4817 | rcu_read_unlock_sched(); | |
4818 | goto out_unlock; | |
4819 | } | |
4820 | } | |
4821 | rcu_read_unlock_sched(); | |
4822 | } | |
4823 | out_unlock: | |
4824 | mutex_unlock(&wq_pool_mutex); | |
4825 | return busy; | |
4826 | } | |
4827 | ||
4828 | /** | |
4829 | * thaw_workqueues - thaw workqueues | |
4830 | * | |
4831 | * Thaw workqueues. Normal queueing is restored and all collected | |
4832 | * frozen works are transferred to their respective pool worklists. | |
4833 | * | |
4834 | * CONTEXT: | |
4835 | * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. | |
4836 | */ | |
4837 | void thaw_workqueues(void) | |
4838 | { | |
4839 | struct workqueue_struct *wq; | |
4840 | struct pool_workqueue *pwq; | |
4841 | ||
4842 | mutex_lock(&wq_pool_mutex); | |
4843 | ||
4844 | if (!workqueue_freezing) | |
4845 | goto out_unlock; | |
4846 | ||
4847 | workqueue_freezing = false; | |
4848 | ||
4849 | /* restore max_active and repopulate worklist */ | |
4850 | list_for_each_entry(wq, &workqueues, list) { | |
4851 | mutex_lock(&wq->mutex); | |
4852 | for_each_pwq(pwq, wq) | |
4853 | pwq_adjust_max_active(pwq); | |
4854 | mutex_unlock(&wq->mutex); | |
4855 | } | |
4856 | ||
4857 | out_unlock: | |
4858 | mutex_unlock(&wq_pool_mutex); | |
4859 | } | |
4860 | #endif /* CONFIG_FREEZER */ | |
4861 | ||
042f7df1 LJ |
4862 | static int workqueue_apply_unbound_cpumask(void) |
4863 | { | |
4864 | LIST_HEAD(ctxs); | |
4865 | int ret = 0; | |
4866 | struct workqueue_struct *wq; | |
4867 | struct apply_wqattrs_ctx *ctx, *n; | |
4868 | ||
4869 | lockdep_assert_held(&wq_pool_mutex); | |
4870 | ||
4871 | list_for_each_entry(wq, &workqueues, list) { | |
4872 | if (!(wq->flags & WQ_UNBOUND)) | |
4873 | continue; | |
4874 | /* creating multiple pwqs breaks ordering guarantee */ | |
4875 | if (wq->flags & __WQ_ORDERED) | |
4876 | continue; | |
4877 | ||
4878 | ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs); | |
4879 | if (!ctx) { | |
4880 | ret = -ENOMEM; | |
4881 | break; | |
4882 | } | |
4883 | ||
4884 | list_add_tail(&ctx->list, &ctxs); | |
4885 | } | |
4886 | ||
4887 | list_for_each_entry_safe(ctx, n, &ctxs, list) { | |
4888 | if (!ret) | |
4889 | apply_wqattrs_commit(ctx); | |
4890 | apply_wqattrs_cleanup(ctx); | |
4891 | } | |
4892 | ||
4893 | return ret; | |
4894 | } | |
4895 | ||
4896 | /** | |
4897 | * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask | |
4898 | * @cpumask: the cpumask to set | |
4899 | * | |
4900 | * The low-level workqueues cpumask is a global cpumask that limits | |
4901 | * the affinity of all unbound workqueues. This function check the @cpumask | |
4902 | * and apply it to all unbound workqueues and updates all pwqs of them. | |
4903 | * | |
4904 | * Retun: 0 - Success | |
4905 | * -EINVAL - Invalid @cpumask | |
4906 | * -ENOMEM - Failed to allocate memory for attrs or pwqs. | |
4907 | */ | |
4908 | int workqueue_set_unbound_cpumask(cpumask_var_t cpumask) | |
4909 | { | |
4910 | int ret = -EINVAL; | |
4911 | cpumask_var_t saved_cpumask; | |
4912 | ||
4913 | if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL)) | |
4914 | return -ENOMEM; | |
4915 | ||
042f7df1 LJ |
4916 | cpumask_and(cpumask, cpumask, cpu_possible_mask); |
4917 | if (!cpumask_empty(cpumask)) { | |
a0111cf6 | 4918 | apply_wqattrs_lock(); |
042f7df1 LJ |
4919 | |
4920 | /* save the old wq_unbound_cpumask. */ | |
4921 | cpumask_copy(saved_cpumask, wq_unbound_cpumask); | |
4922 | ||
4923 | /* update wq_unbound_cpumask at first and apply it to wqs. */ | |
4924 | cpumask_copy(wq_unbound_cpumask, cpumask); | |
4925 | ret = workqueue_apply_unbound_cpumask(); | |
4926 | ||
4927 | /* restore the wq_unbound_cpumask when failed. */ | |
4928 | if (ret < 0) | |
4929 | cpumask_copy(wq_unbound_cpumask, saved_cpumask); | |
4930 | ||
a0111cf6 | 4931 | apply_wqattrs_unlock(); |
042f7df1 | 4932 | } |
042f7df1 LJ |
4933 | |
4934 | free_cpumask_var(saved_cpumask); | |
4935 | return ret; | |
4936 | } | |
4937 | ||
6ba94429 FW |
4938 | #ifdef CONFIG_SYSFS |
4939 | /* | |
4940 | * Workqueues with WQ_SYSFS flag set is visible to userland via | |
4941 | * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the | |
4942 | * following attributes. | |
4943 | * | |
4944 | * per_cpu RO bool : whether the workqueue is per-cpu or unbound | |
4945 | * max_active RW int : maximum number of in-flight work items | |
4946 | * | |
4947 | * Unbound workqueues have the following extra attributes. | |
4948 | * | |
4949 | * id RO int : the associated pool ID | |
4950 | * nice RW int : nice value of the workers | |
4951 | * cpumask RW mask : bitmask of allowed CPUs for the workers | |
4952 | */ | |
4953 | struct wq_device { | |
4954 | struct workqueue_struct *wq; | |
4955 | struct device dev; | |
4956 | }; | |
4957 | ||
4958 | static struct workqueue_struct *dev_to_wq(struct device *dev) | |
4959 | { | |
4960 | struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); | |
4961 | ||
4962 | return wq_dev->wq; | |
4963 | } | |
4964 | ||
4965 | static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, | |
4966 | char *buf) | |
4967 | { | |
4968 | struct workqueue_struct *wq = dev_to_wq(dev); | |
4969 | ||
4970 | return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); | |
4971 | } | |
4972 | static DEVICE_ATTR_RO(per_cpu); | |
4973 | ||
4974 | static ssize_t max_active_show(struct device *dev, | |
4975 | struct device_attribute *attr, char *buf) | |
4976 | { | |
4977 | struct workqueue_struct *wq = dev_to_wq(dev); | |
4978 | ||
4979 | return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); | |
4980 | } | |
4981 | ||
4982 | static ssize_t max_active_store(struct device *dev, | |
4983 | struct device_attribute *attr, const char *buf, | |
4984 | size_t count) | |
4985 | { | |
4986 | struct workqueue_struct *wq = dev_to_wq(dev); | |
4987 | int val; | |
4988 | ||
4989 | if (sscanf(buf, "%d", &val) != 1 || val <= 0) | |
4990 | return -EINVAL; | |
4991 | ||
4992 | workqueue_set_max_active(wq, val); | |
4993 | return count; | |
4994 | } | |
4995 | static DEVICE_ATTR_RW(max_active); | |
4996 | ||
4997 | static struct attribute *wq_sysfs_attrs[] = { | |
4998 | &dev_attr_per_cpu.attr, | |
4999 | &dev_attr_max_active.attr, | |
5000 | NULL, | |
5001 | }; | |
5002 | ATTRIBUTE_GROUPS(wq_sysfs); | |
5003 | ||
5004 | static ssize_t wq_pool_ids_show(struct device *dev, | |
5005 | struct device_attribute *attr, char *buf) | |
5006 | { | |
5007 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5008 | const char *delim = ""; | |
5009 | int node, written = 0; | |
5010 | ||
5011 | rcu_read_lock_sched(); | |
5012 | for_each_node(node) { | |
5013 | written += scnprintf(buf + written, PAGE_SIZE - written, | |
5014 | "%s%d:%d", delim, node, | |
5015 | unbound_pwq_by_node(wq, node)->pool->id); | |
5016 | delim = " "; | |
5017 | } | |
5018 | written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); | |
5019 | rcu_read_unlock_sched(); | |
5020 | ||
5021 | return written; | |
5022 | } | |
5023 | ||
5024 | static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, | |
5025 | char *buf) | |
5026 | { | |
5027 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5028 | int written; | |
5029 | ||
5030 | mutex_lock(&wq->mutex); | |
5031 | written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); | |
5032 | mutex_unlock(&wq->mutex); | |
5033 | ||
5034 | return written; | |
5035 | } | |
5036 | ||
5037 | /* prepare workqueue_attrs for sysfs store operations */ | |
5038 | static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) | |
5039 | { | |
5040 | struct workqueue_attrs *attrs; | |
5041 | ||
899a94fe LJ |
5042 | lockdep_assert_held(&wq_pool_mutex); |
5043 | ||
6ba94429 FW |
5044 | attrs = alloc_workqueue_attrs(GFP_KERNEL); |
5045 | if (!attrs) | |
5046 | return NULL; | |
5047 | ||
6ba94429 | 5048 | copy_workqueue_attrs(attrs, wq->unbound_attrs); |
6ba94429 FW |
5049 | return attrs; |
5050 | } | |
5051 | ||
5052 | static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, | |
5053 | const char *buf, size_t count) | |
5054 | { | |
5055 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5056 | struct workqueue_attrs *attrs; | |
d4d3e257 LJ |
5057 | int ret = -ENOMEM; |
5058 | ||
5059 | apply_wqattrs_lock(); | |
6ba94429 FW |
5060 | |
5061 | attrs = wq_sysfs_prep_attrs(wq); | |
5062 | if (!attrs) | |
d4d3e257 | 5063 | goto out_unlock; |
6ba94429 FW |
5064 | |
5065 | if (sscanf(buf, "%d", &attrs->nice) == 1 && | |
5066 | attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) | |
d4d3e257 | 5067 | ret = apply_workqueue_attrs_locked(wq, attrs); |
6ba94429 FW |
5068 | else |
5069 | ret = -EINVAL; | |
5070 | ||
d4d3e257 LJ |
5071 | out_unlock: |
5072 | apply_wqattrs_unlock(); | |
6ba94429 FW |
5073 | free_workqueue_attrs(attrs); |
5074 | return ret ?: count; | |
5075 | } | |
5076 | ||
5077 | static ssize_t wq_cpumask_show(struct device *dev, | |
5078 | struct device_attribute *attr, char *buf) | |
5079 | { | |
5080 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5081 | int written; | |
5082 | ||
5083 | mutex_lock(&wq->mutex); | |
5084 | written = scnprintf(buf, PAGE_SIZE, "%*pb\n", | |
5085 | cpumask_pr_args(wq->unbound_attrs->cpumask)); | |
5086 | mutex_unlock(&wq->mutex); | |
5087 | return written; | |
5088 | } | |
5089 | ||
5090 | static ssize_t wq_cpumask_store(struct device *dev, | |
5091 | struct device_attribute *attr, | |
5092 | const char *buf, size_t count) | |
5093 | { | |
5094 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5095 | struct workqueue_attrs *attrs; | |
d4d3e257 LJ |
5096 | int ret = -ENOMEM; |
5097 | ||
5098 | apply_wqattrs_lock(); | |
6ba94429 FW |
5099 | |
5100 | attrs = wq_sysfs_prep_attrs(wq); | |
5101 | if (!attrs) | |
d4d3e257 | 5102 | goto out_unlock; |
6ba94429 FW |
5103 | |
5104 | ret = cpumask_parse(buf, attrs->cpumask); | |
5105 | if (!ret) | |
d4d3e257 | 5106 | ret = apply_workqueue_attrs_locked(wq, attrs); |
6ba94429 | 5107 | |
d4d3e257 LJ |
5108 | out_unlock: |
5109 | apply_wqattrs_unlock(); | |
6ba94429 FW |
5110 | free_workqueue_attrs(attrs); |
5111 | return ret ?: count; | |
5112 | } | |
5113 | ||
5114 | static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr, | |
5115 | char *buf) | |
5116 | { | |
5117 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5118 | int written; | |
7dbc725e | 5119 | |
6ba94429 FW |
5120 | mutex_lock(&wq->mutex); |
5121 | written = scnprintf(buf, PAGE_SIZE, "%d\n", | |
5122 | !wq->unbound_attrs->no_numa); | |
5123 | mutex_unlock(&wq->mutex); | |
4c16bd32 | 5124 | |
6ba94429 | 5125 | return written; |
65758202 TH |
5126 | } |
5127 | ||
6ba94429 FW |
5128 | static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr, |
5129 | const char *buf, size_t count) | |
65758202 | 5130 | { |
6ba94429 FW |
5131 | struct workqueue_struct *wq = dev_to_wq(dev); |
5132 | struct workqueue_attrs *attrs; | |
d4d3e257 LJ |
5133 | int v, ret = -ENOMEM; |
5134 | ||
5135 | apply_wqattrs_lock(); | |
4c16bd32 | 5136 | |
6ba94429 FW |
5137 | attrs = wq_sysfs_prep_attrs(wq); |
5138 | if (!attrs) | |
d4d3e257 | 5139 | goto out_unlock; |
4c16bd32 | 5140 | |
6ba94429 FW |
5141 | ret = -EINVAL; |
5142 | if (sscanf(buf, "%d", &v) == 1) { | |
5143 | attrs->no_numa = !v; | |
d4d3e257 | 5144 | ret = apply_workqueue_attrs_locked(wq, attrs); |
65758202 | 5145 | } |
6ba94429 | 5146 | |
d4d3e257 LJ |
5147 | out_unlock: |
5148 | apply_wqattrs_unlock(); | |
6ba94429 FW |
5149 | free_workqueue_attrs(attrs); |
5150 | return ret ?: count; | |
65758202 TH |
5151 | } |
5152 | ||
6ba94429 FW |
5153 | static struct device_attribute wq_sysfs_unbound_attrs[] = { |
5154 | __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL), | |
5155 | __ATTR(nice, 0644, wq_nice_show, wq_nice_store), | |
5156 | __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), | |
5157 | __ATTR(numa, 0644, wq_numa_show, wq_numa_store), | |
5158 | __ATTR_NULL, | |
5159 | }; | |
8ccad40d | 5160 | |
6ba94429 FW |
5161 | static struct bus_type wq_subsys = { |
5162 | .name = "workqueue", | |
5163 | .dev_groups = wq_sysfs_groups, | |
2d3854a3 RR |
5164 | }; |
5165 | ||
b05a7928 FW |
5166 | static ssize_t wq_unbound_cpumask_show(struct device *dev, |
5167 | struct device_attribute *attr, char *buf) | |
5168 | { | |
5169 | int written; | |
5170 | ||
042f7df1 | 5171 | mutex_lock(&wq_pool_mutex); |
b05a7928 FW |
5172 | written = scnprintf(buf, PAGE_SIZE, "%*pb\n", |
5173 | cpumask_pr_args(wq_unbound_cpumask)); | |
042f7df1 | 5174 | mutex_unlock(&wq_pool_mutex); |
b05a7928 FW |
5175 | |
5176 | return written; | |
5177 | } | |
5178 | ||
042f7df1 LJ |
5179 | static ssize_t wq_unbound_cpumask_store(struct device *dev, |
5180 | struct device_attribute *attr, const char *buf, size_t count) | |
5181 | { | |
5182 | cpumask_var_t cpumask; | |
5183 | int ret; | |
5184 | ||
5185 | if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) | |
5186 | return -ENOMEM; | |
5187 | ||
5188 | ret = cpumask_parse(buf, cpumask); | |
5189 | if (!ret) | |
5190 | ret = workqueue_set_unbound_cpumask(cpumask); | |
5191 | ||
5192 | free_cpumask_var(cpumask); | |
5193 | return ret ? ret : count; | |
5194 | } | |
5195 | ||
b05a7928 | 5196 | static struct device_attribute wq_sysfs_cpumask_attr = |
042f7df1 LJ |
5197 | __ATTR(cpumask, 0644, wq_unbound_cpumask_show, |
5198 | wq_unbound_cpumask_store); | |
b05a7928 | 5199 | |
6ba94429 | 5200 | static int __init wq_sysfs_init(void) |
2d3854a3 | 5201 | { |
b05a7928 FW |
5202 | int err; |
5203 | ||
5204 | err = subsys_virtual_register(&wq_subsys, NULL); | |
5205 | if (err) | |
5206 | return err; | |
5207 | ||
5208 | return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr); | |
2d3854a3 | 5209 | } |
6ba94429 | 5210 | core_initcall(wq_sysfs_init); |
2d3854a3 | 5211 | |
6ba94429 | 5212 | static void wq_device_release(struct device *dev) |
2d3854a3 | 5213 | { |
6ba94429 | 5214 | struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); |
6b44003e | 5215 | |
6ba94429 | 5216 | kfree(wq_dev); |
2d3854a3 | 5217 | } |
a0a1a5fd TH |
5218 | |
5219 | /** | |
6ba94429 FW |
5220 | * workqueue_sysfs_register - make a workqueue visible in sysfs |
5221 | * @wq: the workqueue to register | |
a0a1a5fd | 5222 | * |
6ba94429 FW |
5223 | * Expose @wq in sysfs under /sys/bus/workqueue/devices. |
5224 | * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set | |
5225 | * which is the preferred method. | |
a0a1a5fd | 5226 | * |
6ba94429 FW |
5227 | * Workqueue user should use this function directly iff it wants to apply |
5228 | * workqueue_attrs before making the workqueue visible in sysfs; otherwise, | |
5229 | * apply_workqueue_attrs() may race against userland updating the | |
5230 | * attributes. | |
5231 | * | |
5232 | * Return: 0 on success, -errno on failure. | |
a0a1a5fd | 5233 | */ |
6ba94429 | 5234 | int workqueue_sysfs_register(struct workqueue_struct *wq) |
a0a1a5fd | 5235 | { |
6ba94429 FW |
5236 | struct wq_device *wq_dev; |
5237 | int ret; | |
a0a1a5fd | 5238 | |
6ba94429 | 5239 | /* |
402dd89d | 5240 | * Adjusting max_active or creating new pwqs by applying |
6ba94429 FW |
5241 | * attributes breaks ordering guarantee. Disallow exposing ordered |
5242 | * workqueues. | |
5243 | */ | |
5244 | if (WARN_ON(wq->flags & __WQ_ORDERED)) | |
5245 | return -EINVAL; | |
a0a1a5fd | 5246 | |
6ba94429 FW |
5247 | wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); |
5248 | if (!wq_dev) | |
5249 | return -ENOMEM; | |
5bcab335 | 5250 | |
6ba94429 FW |
5251 | wq_dev->wq = wq; |
5252 | wq_dev->dev.bus = &wq_subsys; | |
5253 | wq_dev->dev.init_name = wq->name; | |
5254 | wq_dev->dev.release = wq_device_release; | |
a0a1a5fd | 5255 | |
6ba94429 FW |
5256 | /* |
5257 | * unbound_attrs are created separately. Suppress uevent until | |
5258 | * everything is ready. | |
5259 | */ | |
5260 | dev_set_uevent_suppress(&wq_dev->dev, true); | |
a0a1a5fd | 5261 | |
6ba94429 FW |
5262 | ret = device_register(&wq_dev->dev); |
5263 | if (ret) { | |
5264 | kfree(wq_dev); | |
5265 | wq->wq_dev = NULL; | |
5266 | return ret; | |
5267 | } | |
a0a1a5fd | 5268 | |
6ba94429 FW |
5269 | if (wq->flags & WQ_UNBOUND) { |
5270 | struct device_attribute *attr; | |
a0a1a5fd | 5271 | |
6ba94429 FW |
5272 | for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { |
5273 | ret = device_create_file(&wq_dev->dev, attr); | |
5274 | if (ret) { | |
5275 | device_unregister(&wq_dev->dev); | |
5276 | wq->wq_dev = NULL; | |
5277 | return ret; | |
a0a1a5fd TH |
5278 | } |
5279 | } | |
5280 | } | |
6ba94429 FW |
5281 | |
5282 | dev_set_uevent_suppress(&wq_dev->dev, false); | |
5283 | kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); | |
5284 | return 0; | |
a0a1a5fd TH |
5285 | } |
5286 | ||
5287 | /** | |
6ba94429 FW |
5288 | * workqueue_sysfs_unregister - undo workqueue_sysfs_register() |
5289 | * @wq: the workqueue to unregister | |
a0a1a5fd | 5290 | * |
6ba94429 | 5291 | * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. |
a0a1a5fd | 5292 | */ |
6ba94429 | 5293 | static void workqueue_sysfs_unregister(struct workqueue_struct *wq) |
a0a1a5fd | 5294 | { |
6ba94429 | 5295 | struct wq_device *wq_dev = wq->wq_dev; |
8b03ae3c | 5296 | |
6ba94429 FW |
5297 | if (!wq->wq_dev) |
5298 | return; | |
a0a1a5fd | 5299 | |
6ba94429 FW |
5300 | wq->wq_dev = NULL; |
5301 | device_unregister(&wq_dev->dev); | |
a0a1a5fd | 5302 | } |
6ba94429 FW |
5303 | #else /* CONFIG_SYSFS */ |
5304 | static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } | |
5305 | #endif /* CONFIG_SYSFS */ | |
a0a1a5fd | 5306 | |
82607adc TH |
5307 | /* |
5308 | * Workqueue watchdog. | |
5309 | * | |
5310 | * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal | |
5311 | * flush dependency, a concurrency managed work item which stays RUNNING | |
5312 | * indefinitely. Workqueue stalls can be very difficult to debug as the | |
5313 | * usual warning mechanisms don't trigger and internal workqueue state is | |
5314 | * largely opaque. | |
5315 | * | |
5316 | * Workqueue watchdog monitors all worker pools periodically and dumps | |
5317 | * state if some pools failed to make forward progress for a while where | |
5318 | * forward progress is defined as the first item on ->worklist changing. | |
5319 | * | |
5320 | * This mechanism is controlled through the kernel parameter | |
5321 | * "workqueue.watchdog_thresh" which can be updated at runtime through the | |
5322 | * corresponding sysfs parameter file. | |
5323 | */ | |
5324 | #ifdef CONFIG_WQ_WATCHDOG | |
5325 | ||
5326 | static void wq_watchdog_timer_fn(unsigned long data); | |
5327 | ||
5328 | static unsigned long wq_watchdog_thresh = 30; | |
5329 | static struct timer_list wq_watchdog_timer = | |
5330 | TIMER_DEFERRED_INITIALIZER(wq_watchdog_timer_fn, 0, 0); | |
5331 | ||
5332 | static unsigned long wq_watchdog_touched = INITIAL_JIFFIES; | |
5333 | static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES; | |
5334 | ||
5335 | static void wq_watchdog_reset_touched(void) | |
5336 | { | |
5337 | int cpu; | |
5338 | ||
5339 | wq_watchdog_touched = jiffies; | |
5340 | for_each_possible_cpu(cpu) | |
5341 | per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; | |
5342 | } | |
5343 | ||
5344 | static void wq_watchdog_timer_fn(unsigned long data) | |
5345 | { | |
5346 | unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ; | |
5347 | bool lockup_detected = false; | |
5348 | struct worker_pool *pool; | |
5349 | int pi; | |
5350 | ||
5351 | if (!thresh) | |
5352 | return; | |
5353 | ||
5354 | rcu_read_lock(); | |
5355 | ||
5356 | for_each_pool(pool, pi) { | |
5357 | unsigned long pool_ts, touched, ts; | |
5358 | ||
5359 | if (list_empty(&pool->worklist)) | |
5360 | continue; | |
5361 | ||
5362 | /* get the latest of pool and touched timestamps */ | |
5363 | pool_ts = READ_ONCE(pool->watchdog_ts); | |
5364 | touched = READ_ONCE(wq_watchdog_touched); | |
5365 | ||
5366 | if (time_after(pool_ts, touched)) | |
5367 | ts = pool_ts; | |
5368 | else | |
5369 | ts = touched; | |
5370 | ||
5371 | if (pool->cpu >= 0) { | |
5372 | unsigned long cpu_touched = | |
5373 | READ_ONCE(per_cpu(wq_watchdog_touched_cpu, | |
5374 | pool->cpu)); | |
5375 | if (time_after(cpu_touched, ts)) | |
5376 | ts = cpu_touched; | |
5377 | } | |
5378 | ||
5379 | /* did we stall? */ | |
5380 | if (time_after(jiffies, ts + thresh)) { | |
5381 | lockup_detected = true; | |
5382 | pr_emerg("BUG: workqueue lockup - pool"); | |
5383 | pr_cont_pool_info(pool); | |
5384 | pr_cont(" stuck for %us!\n", | |
5385 | jiffies_to_msecs(jiffies - pool_ts) / 1000); | |
5386 | } | |
5387 | } | |
5388 | ||
5389 | rcu_read_unlock(); | |
5390 | ||
5391 | if (lockup_detected) | |
5392 | show_workqueue_state(); | |
5393 | ||
5394 | wq_watchdog_reset_touched(); | |
5395 | mod_timer(&wq_watchdog_timer, jiffies + thresh); | |
5396 | } | |
5397 | ||
5398 | void wq_watchdog_touch(int cpu) | |
5399 | { | |
5400 | if (cpu >= 0) | |
5401 | per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; | |
5402 | else | |
5403 | wq_watchdog_touched = jiffies; | |
5404 | } | |
5405 | ||
5406 | static void wq_watchdog_set_thresh(unsigned long thresh) | |
5407 | { | |
5408 | wq_watchdog_thresh = 0; | |
5409 | del_timer_sync(&wq_watchdog_timer); | |
5410 | ||
5411 | if (thresh) { | |
5412 | wq_watchdog_thresh = thresh; | |
5413 | wq_watchdog_reset_touched(); | |
5414 | mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ); | |
5415 | } | |
5416 | } | |
5417 | ||
5418 | static int wq_watchdog_param_set_thresh(const char *val, | |
5419 | const struct kernel_param *kp) | |
5420 | { | |
5421 | unsigned long thresh; | |
5422 | int ret; | |
5423 | ||
5424 | ret = kstrtoul(val, 0, &thresh); | |
5425 | if (ret) | |
5426 | return ret; | |
5427 | ||
5428 | if (system_wq) | |
5429 | wq_watchdog_set_thresh(thresh); | |
5430 | else | |
5431 | wq_watchdog_thresh = thresh; | |
5432 | ||
5433 | return 0; | |
5434 | } | |
5435 | ||
5436 | static const struct kernel_param_ops wq_watchdog_thresh_ops = { | |
5437 | .set = wq_watchdog_param_set_thresh, | |
5438 | .get = param_get_ulong, | |
5439 | }; | |
5440 | ||
5441 | module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh, | |
5442 | 0644); | |
5443 | ||
5444 | static void wq_watchdog_init(void) | |
5445 | { | |
5446 | wq_watchdog_set_thresh(wq_watchdog_thresh); | |
5447 | } | |
5448 | ||
5449 | #else /* CONFIG_WQ_WATCHDOG */ | |
5450 | ||
5451 | static inline void wq_watchdog_init(void) { } | |
5452 | ||
5453 | #endif /* CONFIG_WQ_WATCHDOG */ | |
5454 | ||
bce90380 TH |
5455 | static void __init wq_numa_init(void) |
5456 | { | |
5457 | cpumask_var_t *tbl; | |
5458 | int node, cpu; | |
5459 | ||
bce90380 TH |
5460 | if (num_possible_nodes() <= 1) |
5461 | return; | |
5462 | ||
d55262c4 TH |
5463 | if (wq_disable_numa) { |
5464 | pr_info("workqueue: NUMA affinity support disabled\n"); | |
5465 | return; | |
5466 | } | |
5467 | ||
4c16bd32 TH |
5468 | wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL); |
5469 | BUG_ON(!wq_update_unbound_numa_attrs_buf); | |
5470 | ||
bce90380 TH |
5471 | /* |
5472 | * We want masks of possible CPUs of each node which isn't readily | |
5473 | * available. Build one from cpu_to_node() which should have been | |
5474 | * fully initialized by now. | |
5475 | */ | |
ddcb57e2 | 5476 | tbl = kzalloc(nr_node_ids * sizeof(tbl[0]), GFP_KERNEL); |
bce90380 TH |
5477 | BUG_ON(!tbl); |
5478 | ||
5479 | for_each_node(node) | |
5a6024f1 | 5480 | BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL, |
1be0c25d | 5481 | node_online(node) ? node : NUMA_NO_NODE)); |
bce90380 TH |
5482 | |
5483 | for_each_possible_cpu(cpu) { | |
5484 | node = cpu_to_node(cpu); | |
5485 | if (WARN_ON(node == NUMA_NO_NODE)) { | |
5486 | pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu); | |
5487 | /* happens iff arch is bonkers, let's just proceed */ | |
5488 | return; | |
5489 | } | |
5490 | cpumask_set_cpu(cpu, tbl[node]); | |
5491 | } | |
5492 | ||
5493 | wq_numa_possible_cpumask = tbl; | |
5494 | wq_numa_enabled = true; | |
5495 | } | |
5496 | ||
6ee0578b | 5497 | static int __init init_workqueues(void) |
1da177e4 | 5498 | { |
7a4e344c TH |
5499 | int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; |
5500 | int i, cpu; | |
c34056a3 | 5501 | |
e904e6c2 TH |
5502 | WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); |
5503 | ||
b05a7928 FW |
5504 | BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL)); |
5505 | cpumask_copy(wq_unbound_cpumask, cpu_possible_mask); | |
5506 | ||
e904e6c2 TH |
5507 | pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); |
5508 | ||
65758202 | 5509 | cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP); |
a5b4e57d | 5510 | hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN); |
8b03ae3c | 5511 | |
bce90380 TH |
5512 | wq_numa_init(); |
5513 | ||
706026c2 | 5514 | /* initialize CPU pools */ |
29c91e99 | 5515 | for_each_possible_cpu(cpu) { |
4ce62e9e | 5516 | struct worker_pool *pool; |
8b03ae3c | 5517 | |
7a4e344c | 5518 | i = 0; |
f02ae73a | 5519 | for_each_cpu_worker_pool(pool, cpu) { |
7a4e344c | 5520 | BUG_ON(init_worker_pool(pool)); |
ec22ca5e | 5521 | pool->cpu = cpu; |
29c91e99 | 5522 | cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); |
7a4e344c | 5523 | pool->attrs->nice = std_nice[i++]; |
f3f90ad4 | 5524 | pool->node = cpu_to_node(cpu); |
7a4e344c | 5525 | |
9daf9e67 | 5526 | /* alloc pool ID */ |
68e13a67 | 5527 | mutex_lock(&wq_pool_mutex); |
9daf9e67 | 5528 | BUG_ON(worker_pool_assign_id(pool)); |
68e13a67 | 5529 | mutex_unlock(&wq_pool_mutex); |
4ce62e9e | 5530 | } |
8b03ae3c TH |
5531 | } |
5532 | ||
e22bee78 | 5533 | /* create the initial worker */ |
29c91e99 | 5534 | for_each_online_cpu(cpu) { |
4ce62e9e | 5535 | struct worker_pool *pool; |
e22bee78 | 5536 | |
f02ae73a | 5537 | for_each_cpu_worker_pool(pool, cpu) { |
29c91e99 | 5538 | pool->flags &= ~POOL_DISASSOCIATED; |
051e1850 | 5539 | BUG_ON(!create_worker(pool)); |
4ce62e9e | 5540 | } |
e22bee78 TH |
5541 | } |
5542 | ||
8a2b7538 | 5543 | /* create default unbound and ordered wq attrs */ |
29c91e99 TH |
5544 | for (i = 0; i < NR_STD_WORKER_POOLS; i++) { |
5545 | struct workqueue_attrs *attrs; | |
5546 | ||
5547 | BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); | |
29c91e99 | 5548 | attrs->nice = std_nice[i]; |
29c91e99 | 5549 | unbound_std_wq_attrs[i] = attrs; |
8a2b7538 TH |
5550 | |
5551 | /* | |
5552 | * An ordered wq should have only one pwq as ordering is | |
5553 | * guaranteed by max_active which is enforced by pwqs. | |
5554 | * Turn off NUMA so that dfl_pwq is used for all nodes. | |
5555 | */ | |
5556 | BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); | |
5557 | attrs->nice = std_nice[i]; | |
5558 | attrs->no_numa = true; | |
5559 | ordered_wq_attrs[i] = attrs; | |
29c91e99 TH |
5560 | } |
5561 | ||
d320c038 | 5562 | system_wq = alloc_workqueue("events", 0, 0); |
1aabe902 | 5563 | system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); |
d320c038 | 5564 | system_long_wq = alloc_workqueue("events_long", 0, 0); |
f3421797 TH |
5565 | system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, |
5566 | WQ_UNBOUND_MAX_ACTIVE); | |
24d51add TH |
5567 | system_freezable_wq = alloc_workqueue("events_freezable", |
5568 | WQ_FREEZABLE, 0); | |
0668106c VK |
5569 | system_power_efficient_wq = alloc_workqueue("events_power_efficient", |
5570 | WQ_POWER_EFFICIENT, 0); | |
5571 | system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient", | |
5572 | WQ_FREEZABLE | WQ_POWER_EFFICIENT, | |
5573 | 0); | |
1aabe902 | 5574 | BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || |
0668106c VK |
5575 | !system_unbound_wq || !system_freezable_wq || |
5576 | !system_power_efficient_wq || | |
5577 | !system_freezable_power_efficient_wq); | |
82607adc TH |
5578 | |
5579 | wq_watchdog_init(); | |
5580 | ||
6ee0578b | 5581 | return 0; |
1da177e4 | 5582 | } |
6ee0578b | 5583 | early_initcall(init_workqueues); |