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 | ||
863b710b | 293 | static bool wq_online; /* can kworkers be created yet? */ |
3347fa09 | 294 | |
bce90380 TH |
295 | static bool wq_numa_enabled; /* unbound NUMA affinity enabled */ |
296 | ||
4c16bd32 TH |
297 | /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */ |
298 | static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf; | |
299 | ||
68e13a67 | 300 | static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ |
2e109a28 | 301 | static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ |
5bcab335 | 302 | |
e2dca7ad | 303 | static LIST_HEAD(workqueues); /* PR: list of all workqueues */ |
68e13a67 | 304 | static bool workqueue_freezing; /* PL: have wqs started freezing? */ |
7d19c5ce | 305 | |
ef557180 MG |
306 | /* PL: allowable cpus for unbound wqs and work items */ |
307 | static cpumask_var_t wq_unbound_cpumask; | |
308 | ||
309 | /* CPU where unbound work was last round robin scheduled from this CPU */ | |
310 | static DEFINE_PER_CPU(int, wq_rr_cpu_last); | |
b05a7928 | 311 | |
f303fccb TH |
312 | /* |
313 | * Local execution of unbound work items is no longer guaranteed. The | |
314 | * following always forces round-robin CPU selection on unbound work items | |
315 | * to uncover usages which depend on it. | |
316 | */ | |
317 | #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU | |
318 | static bool wq_debug_force_rr_cpu = true; | |
319 | #else | |
320 | static bool wq_debug_force_rr_cpu = false; | |
321 | #endif | |
322 | module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644); | |
323 | ||
7d19c5ce | 324 | /* the per-cpu worker pools */ |
25528213 | 325 | static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools); |
7d19c5ce | 326 | |
68e13a67 | 327 | static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ |
7d19c5ce | 328 | |
68e13a67 | 329 | /* PL: hash of all unbound pools keyed by pool->attrs */ |
29c91e99 TH |
330 | static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); |
331 | ||
c5aa87bb | 332 | /* I: attributes used when instantiating standard unbound pools on demand */ |
29c91e99 TH |
333 | static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; |
334 | ||
8a2b7538 TH |
335 | /* I: attributes used when instantiating ordered pools on demand */ |
336 | static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS]; | |
337 | ||
d320c038 | 338 | struct workqueue_struct *system_wq __read_mostly; |
ad7b1f84 | 339 | EXPORT_SYMBOL(system_wq); |
044c782c | 340 | struct workqueue_struct *system_highpri_wq __read_mostly; |
1aabe902 | 341 | EXPORT_SYMBOL_GPL(system_highpri_wq); |
044c782c | 342 | struct workqueue_struct *system_long_wq __read_mostly; |
d320c038 | 343 | EXPORT_SYMBOL_GPL(system_long_wq); |
044c782c | 344 | struct workqueue_struct *system_unbound_wq __read_mostly; |
f3421797 | 345 | EXPORT_SYMBOL_GPL(system_unbound_wq); |
044c782c | 346 | struct workqueue_struct *system_freezable_wq __read_mostly; |
24d51add | 347 | EXPORT_SYMBOL_GPL(system_freezable_wq); |
0668106c VK |
348 | struct workqueue_struct *system_power_efficient_wq __read_mostly; |
349 | EXPORT_SYMBOL_GPL(system_power_efficient_wq); | |
350 | struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly; | |
351 | EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq); | |
d320c038 | 352 | |
7d19c5ce | 353 | static int worker_thread(void *__worker); |
6ba94429 | 354 | static void workqueue_sysfs_unregister(struct workqueue_struct *wq); |
7d19c5ce | 355 | |
97bd2347 TH |
356 | #define CREATE_TRACE_POINTS |
357 | #include <trace/events/workqueue.h> | |
358 | ||
68e13a67 | 359 | #define assert_rcu_or_pool_mutex() \ |
f78f5b90 PM |
360 | RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \ |
361 | !lockdep_is_held(&wq_pool_mutex), \ | |
362 | "sched RCU or wq_pool_mutex should be held") | |
5bcab335 | 363 | |
b09f4fd3 | 364 | #define assert_rcu_or_wq_mutex(wq) \ |
f78f5b90 PM |
365 | RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \ |
366 | !lockdep_is_held(&wq->mutex), \ | |
367 | "sched RCU or wq->mutex should be held") | |
76af4d93 | 368 | |
5b95e1af | 369 | #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \ |
f78f5b90 PM |
370 | RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \ |
371 | !lockdep_is_held(&wq->mutex) && \ | |
372 | !lockdep_is_held(&wq_pool_mutex), \ | |
373 | "sched RCU, wq->mutex or wq_pool_mutex should be held") | |
5b95e1af | 374 | |
f02ae73a TH |
375 | #define for_each_cpu_worker_pool(pool, cpu) \ |
376 | for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ | |
377 | (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ | |
7a62c2c8 | 378 | (pool)++) |
4ce62e9e | 379 | |
17116969 TH |
380 | /** |
381 | * for_each_pool - iterate through all worker_pools in the system | |
382 | * @pool: iteration cursor | |
611c92a0 | 383 | * @pi: integer used for iteration |
fa1b54e6 | 384 | * |
68e13a67 LJ |
385 | * This must be called either with wq_pool_mutex held or sched RCU read |
386 | * locked. If the pool needs to be used beyond the locking in effect, the | |
387 | * caller is responsible for guaranteeing that the pool stays online. | |
fa1b54e6 TH |
388 | * |
389 | * The if/else clause exists only for the lockdep assertion and can be | |
390 | * ignored. | |
17116969 | 391 | */ |
611c92a0 TH |
392 | #define for_each_pool(pool, pi) \ |
393 | idr_for_each_entry(&worker_pool_idr, pool, pi) \ | |
68e13a67 | 394 | if (({ assert_rcu_or_pool_mutex(); false; })) { } \ |
fa1b54e6 | 395 | else |
17116969 | 396 | |
822d8405 TH |
397 | /** |
398 | * for_each_pool_worker - iterate through all workers of a worker_pool | |
399 | * @worker: iteration cursor | |
822d8405 TH |
400 | * @pool: worker_pool to iterate workers of |
401 | * | |
92f9c5c4 | 402 | * This must be called with @pool->attach_mutex. |
822d8405 TH |
403 | * |
404 | * The if/else clause exists only for the lockdep assertion and can be | |
405 | * ignored. | |
406 | */ | |
da028469 LJ |
407 | #define for_each_pool_worker(worker, pool) \ |
408 | list_for_each_entry((worker), &(pool)->workers, node) \ | |
92f9c5c4 | 409 | if (({ lockdep_assert_held(&pool->attach_mutex); false; })) { } \ |
822d8405 TH |
410 | else |
411 | ||
49e3cf44 TH |
412 | /** |
413 | * for_each_pwq - iterate through all pool_workqueues of the specified workqueue | |
414 | * @pwq: iteration cursor | |
415 | * @wq: the target workqueue | |
76af4d93 | 416 | * |
b09f4fd3 | 417 | * This must be called either with wq->mutex held or sched RCU read locked. |
794b18bc TH |
418 | * If the pwq needs to be used beyond the locking in effect, the caller is |
419 | * responsible for guaranteeing that the pwq stays online. | |
76af4d93 TH |
420 | * |
421 | * The if/else clause exists only for the lockdep assertion and can be | |
422 | * ignored. | |
49e3cf44 TH |
423 | */ |
424 | #define for_each_pwq(pwq, wq) \ | |
76af4d93 | 425 | list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \ |
b09f4fd3 | 426 | if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \ |
76af4d93 | 427 | else |
f3421797 | 428 | |
dc186ad7 TG |
429 | #ifdef CONFIG_DEBUG_OBJECTS_WORK |
430 | ||
431 | static struct debug_obj_descr work_debug_descr; | |
432 | ||
99777288 SG |
433 | static void *work_debug_hint(void *addr) |
434 | { | |
435 | return ((struct work_struct *) addr)->func; | |
436 | } | |
437 | ||
b9fdac7f DC |
438 | static bool work_is_static_object(void *addr) |
439 | { | |
440 | struct work_struct *work = addr; | |
441 | ||
442 | return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work)); | |
443 | } | |
444 | ||
dc186ad7 TG |
445 | /* |
446 | * fixup_init is called when: | |
447 | * - an active object is initialized | |
448 | */ | |
02a982a6 | 449 | static bool work_fixup_init(void *addr, enum debug_obj_state state) |
dc186ad7 TG |
450 | { |
451 | struct work_struct *work = addr; | |
452 | ||
453 | switch (state) { | |
454 | case ODEBUG_STATE_ACTIVE: | |
455 | cancel_work_sync(work); | |
456 | debug_object_init(work, &work_debug_descr); | |
02a982a6 | 457 | return true; |
dc186ad7 | 458 | default: |
02a982a6 | 459 | return false; |
dc186ad7 TG |
460 | } |
461 | } | |
462 | ||
dc186ad7 TG |
463 | /* |
464 | * fixup_free is called when: | |
465 | * - an active object is freed | |
466 | */ | |
02a982a6 | 467 | static bool work_fixup_free(void *addr, enum debug_obj_state state) |
dc186ad7 TG |
468 | { |
469 | struct work_struct *work = addr; | |
470 | ||
471 | switch (state) { | |
472 | case ODEBUG_STATE_ACTIVE: | |
473 | cancel_work_sync(work); | |
474 | debug_object_free(work, &work_debug_descr); | |
02a982a6 | 475 | return true; |
dc186ad7 | 476 | default: |
02a982a6 | 477 | return false; |
dc186ad7 TG |
478 | } |
479 | } | |
480 | ||
481 | static struct debug_obj_descr work_debug_descr = { | |
482 | .name = "work_struct", | |
99777288 | 483 | .debug_hint = work_debug_hint, |
b9fdac7f | 484 | .is_static_object = work_is_static_object, |
dc186ad7 | 485 | .fixup_init = work_fixup_init, |
dc186ad7 TG |
486 | .fixup_free = work_fixup_free, |
487 | }; | |
488 | ||
489 | static inline void debug_work_activate(struct work_struct *work) | |
490 | { | |
491 | debug_object_activate(work, &work_debug_descr); | |
492 | } | |
493 | ||
494 | static inline void debug_work_deactivate(struct work_struct *work) | |
495 | { | |
496 | debug_object_deactivate(work, &work_debug_descr); | |
497 | } | |
498 | ||
499 | void __init_work(struct work_struct *work, int onstack) | |
500 | { | |
501 | if (onstack) | |
502 | debug_object_init_on_stack(work, &work_debug_descr); | |
503 | else | |
504 | debug_object_init(work, &work_debug_descr); | |
505 | } | |
506 | EXPORT_SYMBOL_GPL(__init_work); | |
507 | ||
508 | void destroy_work_on_stack(struct work_struct *work) | |
509 | { | |
510 | debug_object_free(work, &work_debug_descr); | |
511 | } | |
512 | EXPORT_SYMBOL_GPL(destroy_work_on_stack); | |
513 | ||
ea2e64f2 TG |
514 | void destroy_delayed_work_on_stack(struct delayed_work *work) |
515 | { | |
516 | destroy_timer_on_stack(&work->timer); | |
517 | debug_object_free(&work->work, &work_debug_descr); | |
518 | } | |
519 | EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack); | |
520 | ||
dc186ad7 TG |
521 | #else |
522 | static inline void debug_work_activate(struct work_struct *work) { } | |
523 | static inline void debug_work_deactivate(struct work_struct *work) { } | |
524 | #endif | |
525 | ||
4e8b22bd LB |
526 | /** |
527 | * worker_pool_assign_id - allocate ID and assing it to @pool | |
528 | * @pool: the pool pointer of interest | |
529 | * | |
530 | * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned | |
531 | * successfully, -errno on failure. | |
532 | */ | |
9daf9e67 TH |
533 | static int worker_pool_assign_id(struct worker_pool *pool) |
534 | { | |
535 | int ret; | |
536 | ||
68e13a67 | 537 | lockdep_assert_held(&wq_pool_mutex); |
5bcab335 | 538 | |
4e8b22bd LB |
539 | ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE, |
540 | GFP_KERNEL); | |
229641a6 | 541 | if (ret >= 0) { |
e68035fb | 542 | pool->id = ret; |
229641a6 TH |
543 | return 0; |
544 | } | |
fa1b54e6 | 545 | return ret; |
7c3eed5c TH |
546 | } |
547 | ||
df2d5ae4 TH |
548 | /** |
549 | * unbound_pwq_by_node - return the unbound pool_workqueue for the given node | |
550 | * @wq: the target workqueue | |
551 | * @node: the node ID | |
552 | * | |
5b95e1af LJ |
553 | * This must be called with any of wq_pool_mutex, wq->mutex or sched RCU |
554 | * read locked. | |
df2d5ae4 TH |
555 | * If the pwq needs to be used beyond the locking in effect, the caller is |
556 | * responsible for guaranteeing that the pwq stays online. | |
d185af30 YB |
557 | * |
558 | * Return: The unbound pool_workqueue for @node. | |
df2d5ae4 TH |
559 | */ |
560 | static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq, | |
561 | int node) | |
562 | { | |
5b95e1af | 563 | assert_rcu_or_wq_mutex_or_pool_mutex(wq); |
d6e022f1 TH |
564 | |
565 | /* | |
566 | * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a | |
567 | * delayed item is pending. The plan is to keep CPU -> NODE | |
568 | * mapping valid and stable across CPU on/offlines. Once that | |
569 | * happens, this workaround can be removed. | |
570 | */ | |
571 | if (unlikely(node == NUMA_NO_NODE)) | |
572 | return wq->dfl_pwq; | |
573 | ||
df2d5ae4 TH |
574 | return rcu_dereference_raw(wq->numa_pwq_tbl[node]); |
575 | } | |
576 | ||
73f53c4a TH |
577 | static unsigned int work_color_to_flags(int color) |
578 | { | |
579 | return color << WORK_STRUCT_COLOR_SHIFT; | |
580 | } | |
581 | ||
582 | static int get_work_color(struct work_struct *work) | |
583 | { | |
584 | return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) & | |
585 | ((1 << WORK_STRUCT_COLOR_BITS) - 1); | |
586 | } | |
587 | ||
588 | static int work_next_color(int color) | |
589 | { | |
590 | return (color + 1) % WORK_NR_COLORS; | |
591 | } | |
1da177e4 | 592 | |
14441960 | 593 | /* |
112202d9 TH |
594 | * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data |
595 | * contain the pointer to the queued pwq. Once execution starts, the flag | |
7c3eed5c | 596 | * is cleared and the high bits contain OFFQ flags and pool ID. |
7a22ad75 | 597 | * |
112202d9 TH |
598 | * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() |
599 | * and clear_work_data() can be used to set the pwq, pool or clear | |
bbb68dfa TH |
600 | * work->data. These functions should only be called while the work is |
601 | * owned - ie. while the PENDING bit is set. | |
7a22ad75 | 602 | * |
112202d9 | 603 | * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq |
7c3eed5c | 604 | * corresponding to a work. Pool is available once the work has been |
112202d9 | 605 | * queued anywhere after initialization until it is sync canceled. pwq is |
7c3eed5c | 606 | * available only while the work item is queued. |
7a22ad75 | 607 | * |
bbb68dfa TH |
608 | * %WORK_OFFQ_CANCELING is used to mark a work item which is being |
609 | * canceled. While being canceled, a work item may have its PENDING set | |
610 | * but stay off timer and worklist for arbitrarily long and nobody should | |
611 | * try to steal the PENDING bit. | |
14441960 | 612 | */ |
7a22ad75 TH |
613 | static inline void set_work_data(struct work_struct *work, unsigned long data, |
614 | unsigned long flags) | |
365970a1 | 615 | { |
6183c009 | 616 | WARN_ON_ONCE(!work_pending(work)); |
7a22ad75 TH |
617 | atomic_long_set(&work->data, data | flags | work_static(work)); |
618 | } | |
365970a1 | 619 | |
112202d9 | 620 | static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, |
7a22ad75 TH |
621 | unsigned long extra_flags) |
622 | { | |
112202d9 TH |
623 | set_work_data(work, (unsigned long)pwq, |
624 | WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); | |
365970a1 DH |
625 | } |
626 | ||
4468a00f LJ |
627 | static void set_work_pool_and_keep_pending(struct work_struct *work, |
628 | int pool_id) | |
629 | { | |
630 | set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, | |
631 | WORK_STRUCT_PENDING); | |
632 | } | |
633 | ||
7c3eed5c TH |
634 | static void set_work_pool_and_clear_pending(struct work_struct *work, |
635 | int pool_id) | |
7a22ad75 | 636 | { |
23657bb1 TH |
637 | /* |
638 | * The following wmb is paired with the implied mb in | |
639 | * test_and_set_bit(PENDING) and ensures all updates to @work made | |
640 | * here are visible to and precede any updates by the next PENDING | |
641 | * owner. | |
642 | */ | |
643 | smp_wmb(); | |
7c3eed5c | 644 | set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); |
346c09f8 RP |
645 | /* |
646 | * The following mb guarantees that previous clear of a PENDING bit | |
647 | * will not be reordered with any speculative LOADS or STORES from | |
648 | * work->current_func, which is executed afterwards. This possible | |
649 | * reordering can lead to a missed execution on attempt to qeueue | |
650 | * the same @work. E.g. consider this case: | |
651 | * | |
652 | * CPU#0 CPU#1 | |
653 | * ---------------------------- -------------------------------- | |
654 | * | |
655 | * 1 STORE event_indicated | |
656 | * 2 queue_work_on() { | |
657 | * 3 test_and_set_bit(PENDING) | |
658 | * 4 } set_..._and_clear_pending() { | |
659 | * 5 set_work_data() # clear bit | |
660 | * 6 smp_mb() | |
661 | * 7 work->current_func() { | |
662 | * 8 LOAD event_indicated | |
663 | * } | |
664 | * | |
665 | * Without an explicit full barrier speculative LOAD on line 8 can | |
666 | * be executed before CPU#0 does STORE on line 1. If that happens, | |
667 | * CPU#0 observes the PENDING bit is still set and new execution of | |
668 | * a @work is not queued in a hope, that CPU#1 will eventually | |
669 | * finish the queued @work. Meanwhile CPU#1 does not see | |
670 | * event_indicated is set, because speculative LOAD was executed | |
671 | * before actual STORE. | |
672 | */ | |
673 | smp_mb(); | |
7a22ad75 | 674 | } |
f756d5e2 | 675 | |
7a22ad75 | 676 | static void clear_work_data(struct work_struct *work) |
1da177e4 | 677 | { |
7c3eed5c TH |
678 | smp_wmb(); /* see set_work_pool_and_clear_pending() */ |
679 | set_work_data(work, WORK_STRUCT_NO_POOL, 0); | |
1da177e4 LT |
680 | } |
681 | ||
112202d9 | 682 | static struct pool_workqueue *get_work_pwq(struct work_struct *work) |
b1f4ec17 | 683 | { |
e120153d | 684 | unsigned long data = atomic_long_read(&work->data); |
7a22ad75 | 685 | |
112202d9 | 686 | if (data & WORK_STRUCT_PWQ) |
e120153d TH |
687 | return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); |
688 | else | |
689 | return NULL; | |
4d707b9f ON |
690 | } |
691 | ||
7c3eed5c TH |
692 | /** |
693 | * get_work_pool - return the worker_pool a given work was associated with | |
694 | * @work: the work item of interest | |
695 | * | |
68e13a67 LJ |
696 | * Pools are created and destroyed under wq_pool_mutex, and allows read |
697 | * access under sched-RCU read lock. As such, this function should be | |
698 | * called under wq_pool_mutex or with preemption disabled. | |
fa1b54e6 TH |
699 | * |
700 | * All fields of the returned pool are accessible as long as the above | |
701 | * mentioned locking is in effect. If the returned pool needs to be used | |
702 | * beyond the critical section, the caller is responsible for ensuring the | |
703 | * returned pool is and stays online. | |
d185af30 YB |
704 | * |
705 | * Return: The worker_pool @work was last associated with. %NULL if none. | |
7c3eed5c TH |
706 | */ |
707 | static struct worker_pool *get_work_pool(struct work_struct *work) | |
365970a1 | 708 | { |
e120153d | 709 | unsigned long data = atomic_long_read(&work->data); |
7c3eed5c | 710 | int pool_id; |
7a22ad75 | 711 | |
68e13a67 | 712 | assert_rcu_or_pool_mutex(); |
fa1b54e6 | 713 | |
112202d9 TH |
714 | if (data & WORK_STRUCT_PWQ) |
715 | return ((struct pool_workqueue *) | |
7c3eed5c | 716 | (data & WORK_STRUCT_WQ_DATA_MASK))->pool; |
7a22ad75 | 717 | |
7c3eed5c TH |
718 | pool_id = data >> WORK_OFFQ_POOL_SHIFT; |
719 | if (pool_id == WORK_OFFQ_POOL_NONE) | |
7a22ad75 TH |
720 | return NULL; |
721 | ||
fa1b54e6 | 722 | return idr_find(&worker_pool_idr, pool_id); |
7c3eed5c TH |
723 | } |
724 | ||
725 | /** | |
726 | * get_work_pool_id - return the worker pool ID a given work is associated with | |
727 | * @work: the work item of interest | |
728 | * | |
d185af30 | 729 | * Return: The worker_pool ID @work was last associated with. |
7c3eed5c TH |
730 | * %WORK_OFFQ_POOL_NONE if none. |
731 | */ | |
732 | static int get_work_pool_id(struct work_struct *work) | |
733 | { | |
54d5b7d0 LJ |
734 | unsigned long data = atomic_long_read(&work->data); |
735 | ||
112202d9 TH |
736 | if (data & WORK_STRUCT_PWQ) |
737 | return ((struct pool_workqueue *) | |
54d5b7d0 | 738 | (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id; |
7c3eed5c | 739 | |
54d5b7d0 | 740 | return data >> WORK_OFFQ_POOL_SHIFT; |
7c3eed5c TH |
741 | } |
742 | ||
bbb68dfa TH |
743 | static void mark_work_canceling(struct work_struct *work) |
744 | { | |
7c3eed5c | 745 | unsigned long pool_id = get_work_pool_id(work); |
bbb68dfa | 746 | |
7c3eed5c TH |
747 | pool_id <<= WORK_OFFQ_POOL_SHIFT; |
748 | set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); | |
bbb68dfa TH |
749 | } |
750 | ||
751 | static bool work_is_canceling(struct work_struct *work) | |
752 | { | |
753 | unsigned long data = atomic_long_read(&work->data); | |
754 | ||
112202d9 | 755 | return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); |
bbb68dfa TH |
756 | } |
757 | ||
e22bee78 | 758 | /* |
3270476a TH |
759 | * Policy functions. These define the policies on how the global worker |
760 | * pools are managed. Unless noted otherwise, these functions assume that | |
d565ed63 | 761 | * they're being called with pool->lock held. |
e22bee78 TH |
762 | */ |
763 | ||
63d95a91 | 764 | static bool __need_more_worker(struct worker_pool *pool) |
a848e3b6 | 765 | { |
e19e397a | 766 | return !atomic_read(&pool->nr_running); |
a848e3b6 ON |
767 | } |
768 | ||
4594bf15 | 769 | /* |
e22bee78 TH |
770 | * Need to wake up a worker? Called from anything but currently |
771 | * running workers. | |
974271c4 TH |
772 | * |
773 | * Note that, because unbound workers never contribute to nr_running, this | |
706026c2 | 774 | * function will always return %true for unbound pools as long as the |
974271c4 | 775 | * worklist isn't empty. |
4594bf15 | 776 | */ |
63d95a91 | 777 | static bool need_more_worker(struct worker_pool *pool) |
365970a1 | 778 | { |
63d95a91 | 779 | return !list_empty(&pool->worklist) && __need_more_worker(pool); |
e22bee78 | 780 | } |
4594bf15 | 781 | |
e22bee78 | 782 | /* Can I start working? Called from busy but !running workers. */ |
63d95a91 | 783 | static bool may_start_working(struct worker_pool *pool) |
e22bee78 | 784 | { |
63d95a91 | 785 | return pool->nr_idle; |
e22bee78 TH |
786 | } |
787 | ||
788 | /* Do I need to keep working? Called from currently running workers. */ | |
63d95a91 | 789 | static bool keep_working(struct worker_pool *pool) |
e22bee78 | 790 | { |
e19e397a TH |
791 | return !list_empty(&pool->worklist) && |
792 | atomic_read(&pool->nr_running) <= 1; | |
e22bee78 TH |
793 | } |
794 | ||
795 | /* Do we need a new worker? Called from manager. */ | |
63d95a91 | 796 | static bool need_to_create_worker(struct worker_pool *pool) |
e22bee78 | 797 | { |
63d95a91 | 798 | return need_more_worker(pool) && !may_start_working(pool); |
e22bee78 | 799 | } |
365970a1 | 800 | |
e22bee78 | 801 | /* Do we have too many workers and should some go away? */ |
63d95a91 | 802 | static bool too_many_workers(struct worker_pool *pool) |
e22bee78 | 803 | { |
34a06bd6 | 804 | bool managing = mutex_is_locked(&pool->manager_arb); |
63d95a91 TH |
805 | int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ |
806 | int nr_busy = pool->nr_workers - nr_idle; | |
e22bee78 TH |
807 | |
808 | return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; | |
365970a1 DH |
809 | } |
810 | ||
4d707b9f | 811 | /* |
e22bee78 TH |
812 | * Wake up functions. |
813 | */ | |
814 | ||
1037de36 LJ |
815 | /* Return the first idle worker. Safe with preemption disabled */ |
816 | static struct worker *first_idle_worker(struct worker_pool *pool) | |
7e11629d | 817 | { |
63d95a91 | 818 | if (unlikely(list_empty(&pool->idle_list))) |
7e11629d TH |
819 | return NULL; |
820 | ||
63d95a91 | 821 | return list_first_entry(&pool->idle_list, struct worker, entry); |
7e11629d TH |
822 | } |
823 | ||
824 | /** | |
825 | * wake_up_worker - wake up an idle worker | |
63d95a91 | 826 | * @pool: worker pool to wake worker from |
7e11629d | 827 | * |
63d95a91 | 828 | * Wake up the first idle worker of @pool. |
7e11629d TH |
829 | * |
830 | * CONTEXT: | |
d565ed63 | 831 | * spin_lock_irq(pool->lock). |
7e11629d | 832 | */ |
63d95a91 | 833 | static void wake_up_worker(struct worker_pool *pool) |
7e11629d | 834 | { |
1037de36 | 835 | struct worker *worker = first_idle_worker(pool); |
7e11629d TH |
836 | |
837 | if (likely(worker)) | |
838 | wake_up_process(worker->task); | |
839 | } | |
840 | ||
d302f017 | 841 | /** |
e22bee78 TH |
842 | * wq_worker_waking_up - a worker is waking up |
843 | * @task: task waking up | |
844 | * @cpu: CPU @task is waking up to | |
845 | * | |
846 | * This function is called during try_to_wake_up() when a worker is | |
847 | * being awoken. | |
848 | * | |
849 | * CONTEXT: | |
850 | * spin_lock_irq(rq->lock) | |
851 | */ | |
d84ff051 | 852 | void wq_worker_waking_up(struct task_struct *task, int cpu) |
e22bee78 TH |
853 | { |
854 | struct worker *worker = kthread_data(task); | |
855 | ||
36576000 | 856 | if (!(worker->flags & WORKER_NOT_RUNNING)) { |
ec22ca5e | 857 | WARN_ON_ONCE(worker->pool->cpu != cpu); |
e19e397a | 858 | atomic_inc(&worker->pool->nr_running); |
36576000 | 859 | } |
e22bee78 TH |
860 | } |
861 | ||
862 | /** | |
863 | * wq_worker_sleeping - a worker is going to sleep | |
864 | * @task: task going to sleep | |
e22bee78 TH |
865 | * |
866 | * This function is called during schedule() when a busy worker is | |
867 | * going to sleep. Worker on the same cpu can be woken up by | |
868 | * returning pointer to its task. | |
869 | * | |
870 | * CONTEXT: | |
871 | * spin_lock_irq(rq->lock) | |
872 | * | |
d185af30 | 873 | * Return: |
e22bee78 TH |
874 | * Worker task on @cpu to wake up, %NULL if none. |
875 | */ | |
9b7f6597 | 876 | struct task_struct *wq_worker_sleeping(struct task_struct *task) |
e22bee78 TH |
877 | { |
878 | struct worker *worker = kthread_data(task), *to_wakeup = NULL; | |
111c225a | 879 | struct worker_pool *pool; |
e22bee78 | 880 | |
111c225a TH |
881 | /* |
882 | * Rescuers, which may not have all the fields set up like normal | |
883 | * workers, also reach here, let's not access anything before | |
884 | * checking NOT_RUNNING. | |
885 | */ | |
2d64672e | 886 | if (worker->flags & WORKER_NOT_RUNNING) |
e22bee78 TH |
887 | return NULL; |
888 | ||
111c225a | 889 | pool = worker->pool; |
111c225a | 890 | |
e22bee78 | 891 | /* this can only happen on the local cpu */ |
9b7f6597 | 892 | if (WARN_ON_ONCE(pool->cpu != raw_smp_processor_id())) |
6183c009 | 893 | return NULL; |
e22bee78 TH |
894 | |
895 | /* | |
896 | * The counterpart of the following dec_and_test, implied mb, | |
897 | * worklist not empty test sequence is in insert_work(). | |
898 | * Please read comment there. | |
899 | * | |
628c78e7 TH |
900 | * NOT_RUNNING is clear. This means that we're bound to and |
901 | * running on the local cpu w/ rq lock held and preemption | |
902 | * disabled, which in turn means that none else could be | |
d565ed63 | 903 | * manipulating idle_list, so dereferencing idle_list without pool |
628c78e7 | 904 | * lock is safe. |
e22bee78 | 905 | */ |
e19e397a TH |
906 | if (atomic_dec_and_test(&pool->nr_running) && |
907 | !list_empty(&pool->worklist)) | |
1037de36 | 908 | to_wakeup = first_idle_worker(pool); |
e22bee78 TH |
909 | return to_wakeup ? to_wakeup->task : NULL; |
910 | } | |
911 | ||
912 | /** | |
913 | * worker_set_flags - set worker flags and adjust nr_running accordingly | |
cb444766 | 914 | * @worker: self |
d302f017 | 915 | * @flags: flags to set |
d302f017 | 916 | * |
228f1d00 | 917 | * Set @flags in @worker->flags and adjust nr_running accordingly. |
d302f017 | 918 | * |
cb444766 | 919 | * CONTEXT: |
d565ed63 | 920 | * spin_lock_irq(pool->lock) |
d302f017 | 921 | */ |
228f1d00 | 922 | static inline void worker_set_flags(struct worker *worker, unsigned int flags) |
d302f017 | 923 | { |
bd7bdd43 | 924 | struct worker_pool *pool = worker->pool; |
e22bee78 | 925 | |
cb444766 TH |
926 | WARN_ON_ONCE(worker->task != current); |
927 | ||
228f1d00 | 928 | /* If transitioning into NOT_RUNNING, adjust nr_running. */ |
e22bee78 TH |
929 | if ((flags & WORKER_NOT_RUNNING) && |
930 | !(worker->flags & WORKER_NOT_RUNNING)) { | |
228f1d00 | 931 | atomic_dec(&pool->nr_running); |
e22bee78 TH |
932 | } |
933 | ||
d302f017 TH |
934 | worker->flags |= flags; |
935 | } | |
936 | ||
937 | /** | |
e22bee78 | 938 | * worker_clr_flags - clear worker flags and adjust nr_running accordingly |
cb444766 | 939 | * @worker: self |
d302f017 TH |
940 | * @flags: flags to clear |
941 | * | |
e22bee78 | 942 | * Clear @flags in @worker->flags and adjust nr_running accordingly. |
d302f017 | 943 | * |
cb444766 | 944 | * CONTEXT: |
d565ed63 | 945 | * spin_lock_irq(pool->lock) |
d302f017 TH |
946 | */ |
947 | static inline void worker_clr_flags(struct worker *worker, unsigned int flags) | |
948 | { | |
63d95a91 | 949 | struct worker_pool *pool = worker->pool; |
e22bee78 TH |
950 | unsigned int oflags = worker->flags; |
951 | ||
cb444766 TH |
952 | WARN_ON_ONCE(worker->task != current); |
953 | ||
d302f017 | 954 | worker->flags &= ~flags; |
e22bee78 | 955 | |
42c025f3 TH |
956 | /* |
957 | * If transitioning out of NOT_RUNNING, increment nr_running. Note | |
958 | * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask | |
959 | * of multiple flags, not a single flag. | |
960 | */ | |
e22bee78 TH |
961 | if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) |
962 | if (!(worker->flags & WORKER_NOT_RUNNING)) | |
e19e397a | 963 | atomic_inc(&pool->nr_running); |
d302f017 TH |
964 | } |
965 | ||
8cca0eea TH |
966 | /** |
967 | * find_worker_executing_work - find worker which is executing a work | |
c9e7cf27 | 968 | * @pool: pool of interest |
8cca0eea TH |
969 | * @work: work to find worker for |
970 | * | |
c9e7cf27 TH |
971 | * Find a worker which is executing @work on @pool by searching |
972 | * @pool->busy_hash which is keyed by the address of @work. For a worker | |
a2c1c57b TH |
973 | * to match, its current execution should match the address of @work and |
974 | * its work function. This is to avoid unwanted dependency between | |
975 | * unrelated work executions through a work item being recycled while still | |
976 | * being executed. | |
977 | * | |
978 | * This is a bit tricky. A work item may be freed once its execution | |
979 | * starts and nothing prevents the freed area from being recycled for | |
980 | * another work item. If the same work item address ends up being reused | |
981 | * before the original execution finishes, workqueue will identify the | |
982 | * recycled work item as currently executing and make it wait until the | |
983 | * current execution finishes, introducing an unwanted dependency. | |
984 | * | |
c5aa87bb TH |
985 | * This function checks the work item address and work function to avoid |
986 | * false positives. Note that this isn't complete as one may construct a | |
987 | * work function which can introduce dependency onto itself through a | |
988 | * recycled work item. Well, if somebody wants to shoot oneself in the | |
989 | * foot that badly, there's only so much we can do, and if such deadlock | |
990 | * actually occurs, it should be easy to locate the culprit work function. | |
8cca0eea TH |
991 | * |
992 | * CONTEXT: | |
d565ed63 | 993 | * spin_lock_irq(pool->lock). |
8cca0eea | 994 | * |
d185af30 YB |
995 | * Return: |
996 | * Pointer to worker which is executing @work if found, %NULL | |
8cca0eea | 997 | * otherwise. |
4d707b9f | 998 | */ |
c9e7cf27 | 999 | static struct worker *find_worker_executing_work(struct worker_pool *pool, |
8cca0eea | 1000 | struct work_struct *work) |
4d707b9f | 1001 | { |
42f8570f | 1002 | struct worker *worker; |
42f8570f | 1003 | |
b67bfe0d | 1004 | hash_for_each_possible(pool->busy_hash, worker, hentry, |
a2c1c57b TH |
1005 | (unsigned long)work) |
1006 | if (worker->current_work == work && | |
1007 | worker->current_func == work->func) | |
42f8570f SL |
1008 | return worker; |
1009 | ||
1010 | return NULL; | |
4d707b9f ON |
1011 | } |
1012 | ||
bf4ede01 TH |
1013 | /** |
1014 | * move_linked_works - move linked works to a list | |
1015 | * @work: start of series of works to be scheduled | |
1016 | * @head: target list to append @work to | |
402dd89d | 1017 | * @nextp: out parameter for nested worklist walking |
bf4ede01 TH |
1018 | * |
1019 | * Schedule linked works starting from @work to @head. Work series to | |
1020 | * be scheduled starts at @work and includes any consecutive work with | |
1021 | * WORK_STRUCT_LINKED set in its predecessor. | |
1022 | * | |
1023 | * If @nextp is not NULL, it's updated to point to the next work of | |
1024 | * the last scheduled work. This allows move_linked_works() to be | |
1025 | * nested inside outer list_for_each_entry_safe(). | |
1026 | * | |
1027 | * CONTEXT: | |
d565ed63 | 1028 | * spin_lock_irq(pool->lock). |
bf4ede01 TH |
1029 | */ |
1030 | static void move_linked_works(struct work_struct *work, struct list_head *head, | |
1031 | struct work_struct **nextp) | |
1032 | { | |
1033 | struct work_struct *n; | |
1034 | ||
1035 | /* | |
1036 | * Linked worklist will always end before the end of the list, | |
1037 | * use NULL for list head. | |
1038 | */ | |
1039 | list_for_each_entry_safe_from(work, n, NULL, entry) { | |
1040 | list_move_tail(&work->entry, head); | |
1041 | if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) | |
1042 | break; | |
1043 | } | |
1044 | ||
1045 | /* | |
1046 | * If we're already inside safe list traversal and have moved | |
1047 | * multiple works to the scheduled queue, the next position | |
1048 | * needs to be updated. | |
1049 | */ | |
1050 | if (nextp) | |
1051 | *nextp = n; | |
1052 | } | |
1053 | ||
8864b4e5 TH |
1054 | /** |
1055 | * get_pwq - get an extra reference on the specified pool_workqueue | |
1056 | * @pwq: pool_workqueue to get | |
1057 | * | |
1058 | * Obtain an extra reference on @pwq. The caller should guarantee that | |
1059 | * @pwq has positive refcnt and be holding the matching pool->lock. | |
1060 | */ | |
1061 | static void get_pwq(struct pool_workqueue *pwq) | |
1062 | { | |
1063 | lockdep_assert_held(&pwq->pool->lock); | |
1064 | WARN_ON_ONCE(pwq->refcnt <= 0); | |
1065 | pwq->refcnt++; | |
1066 | } | |
1067 | ||
1068 | /** | |
1069 | * put_pwq - put a pool_workqueue reference | |
1070 | * @pwq: pool_workqueue to put | |
1071 | * | |
1072 | * Drop a reference of @pwq. If its refcnt reaches zero, schedule its | |
1073 | * destruction. The caller should be holding the matching pool->lock. | |
1074 | */ | |
1075 | static void put_pwq(struct pool_workqueue *pwq) | |
1076 | { | |
1077 | lockdep_assert_held(&pwq->pool->lock); | |
1078 | if (likely(--pwq->refcnt)) | |
1079 | return; | |
1080 | if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND))) | |
1081 | return; | |
1082 | /* | |
1083 | * @pwq can't be released under pool->lock, bounce to | |
1084 | * pwq_unbound_release_workfn(). This never recurses on the same | |
1085 | * pool->lock as this path is taken only for unbound workqueues and | |
1086 | * the release work item is scheduled on a per-cpu workqueue. To | |
1087 | * avoid lockdep warning, unbound pool->locks are given lockdep | |
1088 | * subclass of 1 in get_unbound_pool(). | |
1089 | */ | |
1090 | schedule_work(&pwq->unbound_release_work); | |
1091 | } | |
1092 | ||
dce90d47 TH |
1093 | /** |
1094 | * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock | |
1095 | * @pwq: pool_workqueue to put (can be %NULL) | |
1096 | * | |
1097 | * put_pwq() with locking. This function also allows %NULL @pwq. | |
1098 | */ | |
1099 | static void put_pwq_unlocked(struct pool_workqueue *pwq) | |
1100 | { | |
1101 | if (pwq) { | |
1102 | /* | |
1103 | * As both pwqs and pools are sched-RCU protected, the | |
1104 | * following lock operations are safe. | |
1105 | */ | |
1106 | spin_lock_irq(&pwq->pool->lock); | |
1107 | put_pwq(pwq); | |
1108 | spin_unlock_irq(&pwq->pool->lock); | |
1109 | } | |
1110 | } | |
1111 | ||
112202d9 | 1112 | static void pwq_activate_delayed_work(struct work_struct *work) |
bf4ede01 | 1113 | { |
112202d9 | 1114 | struct pool_workqueue *pwq = get_work_pwq(work); |
bf4ede01 TH |
1115 | |
1116 | trace_workqueue_activate_work(work); | |
82607adc TH |
1117 | if (list_empty(&pwq->pool->worklist)) |
1118 | pwq->pool->watchdog_ts = jiffies; | |
112202d9 | 1119 | move_linked_works(work, &pwq->pool->worklist, NULL); |
bf4ede01 | 1120 | __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work)); |
112202d9 | 1121 | pwq->nr_active++; |
bf4ede01 TH |
1122 | } |
1123 | ||
112202d9 | 1124 | static void pwq_activate_first_delayed(struct pool_workqueue *pwq) |
3aa62497 | 1125 | { |
112202d9 | 1126 | struct work_struct *work = list_first_entry(&pwq->delayed_works, |
3aa62497 LJ |
1127 | struct work_struct, entry); |
1128 | ||
112202d9 | 1129 | pwq_activate_delayed_work(work); |
3aa62497 LJ |
1130 | } |
1131 | ||
bf4ede01 | 1132 | /** |
112202d9 TH |
1133 | * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight |
1134 | * @pwq: pwq of interest | |
bf4ede01 | 1135 | * @color: color of work which left the queue |
bf4ede01 TH |
1136 | * |
1137 | * A work either has completed or is removed from pending queue, | |
112202d9 | 1138 | * decrement nr_in_flight of its pwq and handle workqueue flushing. |
bf4ede01 TH |
1139 | * |
1140 | * CONTEXT: | |
d565ed63 | 1141 | * spin_lock_irq(pool->lock). |
bf4ede01 | 1142 | */ |
112202d9 | 1143 | static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color) |
bf4ede01 | 1144 | { |
8864b4e5 | 1145 | /* uncolored work items don't participate in flushing or nr_active */ |
bf4ede01 | 1146 | if (color == WORK_NO_COLOR) |
8864b4e5 | 1147 | goto out_put; |
bf4ede01 | 1148 | |
112202d9 | 1149 | pwq->nr_in_flight[color]--; |
bf4ede01 | 1150 | |
112202d9 TH |
1151 | pwq->nr_active--; |
1152 | if (!list_empty(&pwq->delayed_works)) { | |
b3f9f405 | 1153 | /* one down, submit a delayed one */ |
112202d9 TH |
1154 | if (pwq->nr_active < pwq->max_active) |
1155 | pwq_activate_first_delayed(pwq); | |
bf4ede01 TH |
1156 | } |
1157 | ||
1158 | /* is flush in progress and are we at the flushing tip? */ | |
112202d9 | 1159 | if (likely(pwq->flush_color != color)) |
8864b4e5 | 1160 | goto out_put; |
bf4ede01 TH |
1161 | |
1162 | /* are there still in-flight works? */ | |
112202d9 | 1163 | if (pwq->nr_in_flight[color]) |
8864b4e5 | 1164 | goto out_put; |
bf4ede01 | 1165 | |
112202d9 TH |
1166 | /* this pwq is done, clear flush_color */ |
1167 | pwq->flush_color = -1; | |
bf4ede01 TH |
1168 | |
1169 | /* | |
112202d9 | 1170 | * If this was the last pwq, wake up the first flusher. It |
bf4ede01 TH |
1171 | * will handle the rest. |
1172 | */ | |
112202d9 TH |
1173 | if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) |
1174 | complete(&pwq->wq->first_flusher->done); | |
8864b4e5 TH |
1175 | out_put: |
1176 | put_pwq(pwq); | |
bf4ede01 TH |
1177 | } |
1178 | ||
36e227d2 | 1179 | /** |
bbb68dfa | 1180 | * try_to_grab_pending - steal work item from worklist and disable irq |
36e227d2 TH |
1181 | * @work: work item to steal |
1182 | * @is_dwork: @work is a delayed_work | |
bbb68dfa | 1183 | * @flags: place to store irq state |
36e227d2 TH |
1184 | * |
1185 | * Try to grab PENDING bit of @work. This function can handle @work in any | |
d185af30 | 1186 | * stable state - idle, on timer or on worklist. |
36e227d2 | 1187 | * |
d185af30 | 1188 | * Return: |
36e227d2 TH |
1189 | * 1 if @work was pending and we successfully stole PENDING |
1190 | * 0 if @work was idle and we claimed PENDING | |
1191 | * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry | |
bbb68dfa TH |
1192 | * -ENOENT if someone else is canceling @work, this state may persist |
1193 | * for arbitrarily long | |
36e227d2 | 1194 | * |
d185af30 | 1195 | * Note: |
bbb68dfa | 1196 | * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting |
e0aecdd8 TH |
1197 | * interrupted while holding PENDING and @work off queue, irq must be |
1198 | * disabled on entry. This, combined with delayed_work->timer being | |
1199 | * irqsafe, ensures that we return -EAGAIN for finite short period of time. | |
bbb68dfa TH |
1200 | * |
1201 | * On successful return, >= 0, irq is disabled and the caller is | |
1202 | * responsible for releasing it using local_irq_restore(*@flags). | |
1203 | * | |
e0aecdd8 | 1204 | * This function is safe to call from any context including IRQ handler. |
bf4ede01 | 1205 | */ |
bbb68dfa TH |
1206 | static int try_to_grab_pending(struct work_struct *work, bool is_dwork, |
1207 | unsigned long *flags) | |
bf4ede01 | 1208 | { |
d565ed63 | 1209 | struct worker_pool *pool; |
112202d9 | 1210 | struct pool_workqueue *pwq; |
bf4ede01 | 1211 | |
bbb68dfa TH |
1212 | local_irq_save(*flags); |
1213 | ||
36e227d2 TH |
1214 | /* try to steal the timer if it exists */ |
1215 | if (is_dwork) { | |
1216 | struct delayed_work *dwork = to_delayed_work(work); | |
1217 | ||
e0aecdd8 TH |
1218 | /* |
1219 | * dwork->timer is irqsafe. If del_timer() fails, it's | |
1220 | * guaranteed that the timer is not queued anywhere and not | |
1221 | * running on the local CPU. | |
1222 | */ | |
36e227d2 TH |
1223 | if (likely(del_timer(&dwork->timer))) |
1224 | return 1; | |
1225 | } | |
1226 | ||
1227 | /* try to claim PENDING the normal way */ | |
bf4ede01 TH |
1228 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) |
1229 | return 0; | |
1230 | ||
1231 | /* | |
1232 | * The queueing is in progress, or it is already queued. Try to | |
1233 | * steal it from ->worklist without clearing WORK_STRUCT_PENDING. | |
1234 | */ | |
d565ed63 TH |
1235 | pool = get_work_pool(work); |
1236 | if (!pool) | |
bbb68dfa | 1237 | goto fail; |
bf4ede01 | 1238 | |
d565ed63 | 1239 | spin_lock(&pool->lock); |
0b3dae68 | 1240 | /* |
112202d9 TH |
1241 | * work->data is guaranteed to point to pwq only while the work |
1242 | * item is queued on pwq->wq, and both updating work->data to point | |
1243 | * to pwq on queueing and to pool on dequeueing are done under | |
1244 | * pwq->pool->lock. This in turn guarantees that, if work->data | |
1245 | * points to pwq which is associated with a locked pool, the work | |
0b3dae68 LJ |
1246 | * item is currently queued on that pool. |
1247 | */ | |
112202d9 TH |
1248 | pwq = get_work_pwq(work); |
1249 | if (pwq && pwq->pool == pool) { | |
16062836 TH |
1250 | debug_work_deactivate(work); |
1251 | ||
1252 | /* | |
1253 | * A delayed work item cannot be grabbed directly because | |
1254 | * it might have linked NO_COLOR work items which, if left | |
112202d9 | 1255 | * on the delayed_list, will confuse pwq->nr_active |
16062836 TH |
1256 | * management later on and cause stall. Make sure the work |
1257 | * item is activated before grabbing. | |
1258 | */ | |
1259 | if (*work_data_bits(work) & WORK_STRUCT_DELAYED) | |
112202d9 | 1260 | pwq_activate_delayed_work(work); |
16062836 TH |
1261 | |
1262 | list_del_init(&work->entry); | |
9c34a704 | 1263 | pwq_dec_nr_in_flight(pwq, get_work_color(work)); |
16062836 | 1264 | |
112202d9 | 1265 | /* work->data points to pwq iff queued, point to pool */ |
16062836 TH |
1266 | set_work_pool_and_keep_pending(work, pool->id); |
1267 | ||
1268 | spin_unlock(&pool->lock); | |
1269 | return 1; | |
bf4ede01 | 1270 | } |
d565ed63 | 1271 | spin_unlock(&pool->lock); |
bbb68dfa TH |
1272 | fail: |
1273 | local_irq_restore(*flags); | |
1274 | if (work_is_canceling(work)) | |
1275 | return -ENOENT; | |
1276 | cpu_relax(); | |
36e227d2 | 1277 | return -EAGAIN; |
bf4ede01 TH |
1278 | } |
1279 | ||
4690c4ab | 1280 | /** |
706026c2 | 1281 | * insert_work - insert a work into a pool |
112202d9 | 1282 | * @pwq: pwq @work belongs to |
4690c4ab TH |
1283 | * @work: work to insert |
1284 | * @head: insertion point | |
1285 | * @extra_flags: extra WORK_STRUCT_* flags to set | |
1286 | * | |
112202d9 | 1287 | * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to |
706026c2 | 1288 | * work_struct flags. |
4690c4ab TH |
1289 | * |
1290 | * CONTEXT: | |
d565ed63 | 1291 | * spin_lock_irq(pool->lock). |
4690c4ab | 1292 | */ |
112202d9 TH |
1293 | static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, |
1294 | struct list_head *head, unsigned int extra_flags) | |
b89deed3 | 1295 | { |
112202d9 | 1296 | struct worker_pool *pool = pwq->pool; |
e22bee78 | 1297 | |
4690c4ab | 1298 | /* we own @work, set data and link */ |
112202d9 | 1299 | set_work_pwq(work, pwq, extra_flags); |
1a4d9b0a | 1300 | list_add_tail(&work->entry, head); |
8864b4e5 | 1301 | get_pwq(pwq); |
e22bee78 TH |
1302 | |
1303 | /* | |
c5aa87bb TH |
1304 | * Ensure either wq_worker_sleeping() sees the above |
1305 | * list_add_tail() or we see zero nr_running to avoid workers lying | |
1306 | * around lazily while there are works to be processed. | |
e22bee78 TH |
1307 | */ |
1308 | smp_mb(); | |
1309 | ||
63d95a91 TH |
1310 | if (__need_more_worker(pool)) |
1311 | wake_up_worker(pool); | |
b89deed3 ON |
1312 | } |
1313 | ||
c8efcc25 TH |
1314 | /* |
1315 | * Test whether @work is being queued from another work executing on the | |
8d03ecfe | 1316 | * same workqueue. |
c8efcc25 TH |
1317 | */ |
1318 | static bool is_chained_work(struct workqueue_struct *wq) | |
1319 | { | |
8d03ecfe TH |
1320 | struct worker *worker; |
1321 | ||
1322 | worker = current_wq_worker(); | |
1323 | /* | |
1324 | * Return %true iff I'm a worker execuing a work item on @wq. If | |
1325 | * I'm @worker, it's safe to dereference it without locking. | |
1326 | */ | |
112202d9 | 1327 | return worker && worker->current_pwq->wq == wq; |
c8efcc25 TH |
1328 | } |
1329 | ||
ef557180 MG |
1330 | /* |
1331 | * When queueing an unbound work item to a wq, prefer local CPU if allowed | |
1332 | * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to | |
1333 | * avoid perturbing sensitive tasks. | |
1334 | */ | |
1335 | static int wq_select_unbound_cpu(int cpu) | |
1336 | { | |
f303fccb | 1337 | static bool printed_dbg_warning; |
ef557180 MG |
1338 | int new_cpu; |
1339 | ||
f303fccb TH |
1340 | if (likely(!wq_debug_force_rr_cpu)) { |
1341 | if (cpumask_test_cpu(cpu, wq_unbound_cpumask)) | |
1342 | return cpu; | |
1343 | } else if (!printed_dbg_warning) { | |
1344 | pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n"); | |
1345 | printed_dbg_warning = true; | |
1346 | } | |
1347 | ||
ef557180 MG |
1348 | if (cpumask_empty(wq_unbound_cpumask)) |
1349 | return cpu; | |
1350 | ||
1351 | new_cpu = __this_cpu_read(wq_rr_cpu_last); | |
1352 | new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask); | |
1353 | if (unlikely(new_cpu >= nr_cpu_ids)) { | |
1354 | new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask); | |
1355 | if (unlikely(new_cpu >= nr_cpu_ids)) | |
1356 | return cpu; | |
1357 | } | |
1358 | __this_cpu_write(wq_rr_cpu_last, new_cpu); | |
1359 | ||
1360 | return new_cpu; | |
1361 | } | |
1362 | ||
d84ff051 | 1363 | static void __queue_work(int cpu, struct workqueue_struct *wq, |
1da177e4 LT |
1364 | struct work_struct *work) |
1365 | { | |
112202d9 | 1366 | struct pool_workqueue *pwq; |
c9178087 | 1367 | struct worker_pool *last_pool; |
1e19ffc6 | 1368 | struct list_head *worklist; |
8a2e8e5d | 1369 | unsigned int work_flags; |
b75cac93 | 1370 | unsigned int req_cpu = cpu; |
8930caba TH |
1371 | |
1372 | /* | |
1373 | * While a work item is PENDING && off queue, a task trying to | |
1374 | * steal the PENDING will busy-loop waiting for it to either get | |
1375 | * queued or lose PENDING. Grabbing PENDING and queueing should | |
1376 | * happen with IRQ disabled. | |
1377 | */ | |
1378 | WARN_ON_ONCE(!irqs_disabled()); | |
1da177e4 | 1379 | |
dc186ad7 | 1380 | debug_work_activate(work); |
1e19ffc6 | 1381 | |
9ef28a73 | 1382 | /* if draining, only works from the same workqueue are allowed */ |
618b01eb | 1383 | if (unlikely(wq->flags & __WQ_DRAINING) && |
c8efcc25 | 1384 | WARN_ON_ONCE(!is_chained_work(wq))) |
e41e704b | 1385 | return; |
9e8cd2f5 | 1386 | retry: |
df2d5ae4 | 1387 | if (req_cpu == WORK_CPU_UNBOUND) |
ef557180 | 1388 | cpu = wq_select_unbound_cpu(raw_smp_processor_id()); |
df2d5ae4 | 1389 | |
c9178087 | 1390 | /* pwq which will be used unless @work is executing elsewhere */ |
df2d5ae4 | 1391 | if (!(wq->flags & WQ_UNBOUND)) |
7fb98ea7 | 1392 | pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); |
df2d5ae4 TH |
1393 | else |
1394 | pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); | |
dbf2576e | 1395 | |
c9178087 TH |
1396 | /* |
1397 | * If @work was previously on a different pool, it might still be | |
1398 | * running there, in which case the work needs to be queued on that | |
1399 | * pool to guarantee non-reentrancy. | |
1400 | */ | |
1401 | last_pool = get_work_pool(work); | |
1402 | if (last_pool && last_pool != pwq->pool) { | |
1403 | struct worker *worker; | |
18aa9eff | 1404 | |
c9178087 | 1405 | spin_lock(&last_pool->lock); |
18aa9eff | 1406 | |
c9178087 | 1407 | worker = find_worker_executing_work(last_pool, work); |
18aa9eff | 1408 | |
c9178087 TH |
1409 | if (worker && worker->current_pwq->wq == wq) { |
1410 | pwq = worker->current_pwq; | |
8930caba | 1411 | } else { |
c9178087 TH |
1412 | /* meh... not running there, queue here */ |
1413 | spin_unlock(&last_pool->lock); | |
112202d9 | 1414 | spin_lock(&pwq->pool->lock); |
8930caba | 1415 | } |
f3421797 | 1416 | } else { |
112202d9 | 1417 | spin_lock(&pwq->pool->lock); |
502ca9d8 TH |
1418 | } |
1419 | ||
9e8cd2f5 TH |
1420 | /* |
1421 | * pwq is determined and locked. For unbound pools, we could have | |
1422 | * raced with pwq release and it could already be dead. If its | |
1423 | * refcnt is zero, repeat pwq selection. Note that pwqs never die | |
df2d5ae4 TH |
1424 | * without another pwq replacing it in the numa_pwq_tbl or while |
1425 | * work items are executing on it, so the retrying is guaranteed to | |
9e8cd2f5 TH |
1426 | * make forward-progress. |
1427 | */ | |
1428 | if (unlikely(!pwq->refcnt)) { | |
1429 | if (wq->flags & WQ_UNBOUND) { | |
1430 | spin_unlock(&pwq->pool->lock); | |
1431 | cpu_relax(); | |
1432 | goto retry; | |
1433 | } | |
1434 | /* oops */ | |
1435 | WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", | |
1436 | wq->name, cpu); | |
1437 | } | |
1438 | ||
112202d9 TH |
1439 | /* pwq determined, queue */ |
1440 | trace_workqueue_queue_work(req_cpu, pwq, work); | |
502ca9d8 | 1441 | |
f5b2552b | 1442 | if (WARN_ON(!list_empty(&work->entry))) { |
112202d9 | 1443 | spin_unlock(&pwq->pool->lock); |
f5b2552b DC |
1444 | return; |
1445 | } | |
1e19ffc6 | 1446 | |
112202d9 TH |
1447 | pwq->nr_in_flight[pwq->work_color]++; |
1448 | work_flags = work_color_to_flags(pwq->work_color); | |
1e19ffc6 | 1449 | |
112202d9 | 1450 | if (likely(pwq->nr_active < pwq->max_active)) { |
cdadf009 | 1451 | trace_workqueue_activate_work(work); |
112202d9 TH |
1452 | pwq->nr_active++; |
1453 | worklist = &pwq->pool->worklist; | |
82607adc TH |
1454 | if (list_empty(worklist)) |
1455 | pwq->pool->watchdog_ts = jiffies; | |
8a2e8e5d TH |
1456 | } else { |
1457 | work_flags |= WORK_STRUCT_DELAYED; | |
112202d9 | 1458 | worklist = &pwq->delayed_works; |
8a2e8e5d | 1459 | } |
1e19ffc6 | 1460 | |
112202d9 | 1461 | insert_work(pwq, work, worklist, work_flags); |
1e19ffc6 | 1462 | |
112202d9 | 1463 | spin_unlock(&pwq->pool->lock); |
1da177e4 LT |
1464 | } |
1465 | ||
0fcb78c2 | 1466 | /** |
c1a220e7 ZR |
1467 | * queue_work_on - queue work on specific cpu |
1468 | * @cpu: CPU number to execute work on | |
0fcb78c2 REB |
1469 | * @wq: workqueue to use |
1470 | * @work: work to queue | |
1471 | * | |
c1a220e7 ZR |
1472 | * We queue the work to a specific CPU, the caller must ensure it |
1473 | * can't go away. | |
d185af30 YB |
1474 | * |
1475 | * Return: %false if @work was already on a queue, %true otherwise. | |
1da177e4 | 1476 | */ |
d4283e93 TH |
1477 | bool queue_work_on(int cpu, struct workqueue_struct *wq, |
1478 | struct work_struct *work) | |
1da177e4 | 1479 | { |
d4283e93 | 1480 | bool ret = false; |
8930caba | 1481 | unsigned long flags; |
ef1ca236 | 1482 | |
8930caba | 1483 | local_irq_save(flags); |
c1a220e7 | 1484 | |
22df02bb | 1485 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { |
4690c4ab | 1486 | __queue_work(cpu, wq, work); |
d4283e93 | 1487 | ret = true; |
c1a220e7 | 1488 | } |
ef1ca236 | 1489 | |
8930caba | 1490 | local_irq_restore(flags); |
1da177e4 LT |
1491 | return ret; |
1492 | } | |
ad7b1f84 | 1493 | EXPORT_SYMBOL(queue_work_on); |
1da177e4 | 1494 | |
d8e794df | 1495 | void delayed_work_timer_fn(unsigned long __data) |
1da177e4 | 1496 | { |
52bad64d | 1497 | struct delayed_work *dwork = (struct delayed_work *)__data; |
1da177e4 | 1498 | |
e0aecdd8 | 1499 | /* should have been called from irqsafe timer with irq already off */ |
60c057bc | 1500 | __queue_work(dwork->cpu, dwork->wq, &dwork->work); |
1da177e4 | 1501 | } |
1438ade5 | 1502 | EXPORT_SYMBOL(delayed_work_timer_fn); |
1da177e4 | 1503 | |
7beb2edf TH |
1504 | static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, |
1505 | struct delayed_work *dwork, unsigned long delay) | |
1da177e4 | 1506 | { |
7beb2edf TH |
1507 | struct timer_list *timer = &dwork->timer; |
1508 | struct work_struct *work = &dwork->work; | |
7beb2edf TH |
1509 | |
1510 | WARN_ON_ONCE(timer->function != delayed_work_timer_fn || | |
1511 | timer->data != (unsigned long)dwork); | |
fc4b514f TH |
1512 | WARN_ON_ONCE(timer_pending(timer)); |
1513 | WARN_ON_ONCE(!list_empty(&work->entry)); | |
7beb2edf | 1514 | |
8852aac2 TH |
1515 | /* |
1516 | * If @delay is 0, queue @dwork->work immediately. This is for | |
1517 | * both optimization and correctness. The earliest @timer can | |
1518 | * expire is on the closest next tick and delayed_work users depend | |
1519 | * on that there's no such delay when @delay is 0. | |
1520 | */ | |
1521 | if (!delay) { | |
1522 | __queue_work(cpu, wq, &dwork->work); | |
1523 | return; | |
1524 | } | |
1525 | ||
7beb2edf | 1526 | timer_stats_timer_set_start_info(&dwork->timer); |
1da177e4 | 1527 | |
60c057bc | 1528 | dwork->wq = wq; |
1265057f | 1529 | dwork->cpu = cpu; |
7beb2edf TH |
1530 | timer->expires = jiffies + delay; |
1531 | ||
041bd12e TH |
1532 | if (unlikely(cpu != WORK_CPU_UNBOUND)) |
1533 | add_timer_on(timer, cpu); | |
1534 | else | |
1535 | add_timer(timer); | |
1da177e4 LT |
1536 | } |
1537 | ||
0fcb78c2 REB |
1538 | /** |
1539 | * queue_delayed_work_on - queue work on specific CPU after delay | |
1540 | * @cpu: CPU number to execute work on | |
1541 | * @wq: workqueue to use | |
af9997e4 | 1542 | * @dwork: work to queue |
0fcb78c2 REB |
1543 | * @delay: number of jiffies to wait before queueing |
1544 | * | |
d185af30 | 1545 | * Return: %false if @work was already on a queue, %true otherwise. If |
715f1300 TH |
1546 | * @delay is zero and @dwork is idle, it will be scheduled for immediate |
1547 | * execution. | |
0fcb78c2 | 1548 | */ |
d4283e93 TH |
1549 | bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, |
1550 | struct delayed_work *dwork, unsigned long delay) | |
7a6bc1cd | 1551 | { |
52bad64d | 1552 | struct work_struct *work = &dwork->work; |
d4283e93 | 1553 | bool ret = false; |
8930caba | 1554 | unsigned long flags; |
7a6bc1cd | 1555 | |
8930caba TH |
1556 | /* read the comment in __queue_work() */ |
1557 | local_irq_save(flags); | |
7a6bc1cd | 1558 | |
22df02bb | 1559 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { |
7beb2edf | 1560 | __queue_delayed_work(cpu, wq, dwork, delay); |
d4283e93 | 1561 | ret = true; |
7a6bc1cd | 1562 | } |
8a3e77cc | 1563 | |
8930caba | 1564 | local_irq_restore(flags); |
7a6bc1cd VP |
1565 | return ret; |
1566 | } | |
ad7b1f84 | 1567 | EXPORT_SYMBOL(queue_delayed_work_on); |
c7fc77f7 | 1568 | |
8376fe22 TH |
1569 | /** |
1570 | * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU | |
1571 | * @cpu: CPU number to execute work on | |
1572 | * @wq: workqueue to use | |
1573 | * @dwork: work to queue | |
1574 | * @delay: number of jiffies to wait before queueing | |
1575 | * | |
1576 | * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, | |
1577 | * modify @dwork's timer so that it expires after @delay. If @delay is | |
1578 | * zero, @work is guaranteed to be scheduled immediately regardless of its | |
1579 | * current state. | |
1580 | * | |
d185af30 | 1581 | * Return: %false if @dwork was idle and queued, %true if @dwork was |
8376fe22 TH |
1582 | * pending and its timer was modified. |
1583 | * | |
e0aecdd8 | 1584 | * This function is safe to call from any context including IRQ handler. |
8376fe22 TH |
1585 | * See try_to_grab_pending() for details. |
1586 | */ | |
1587 | bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, | |
1588 | struct delayed_work *dwork, unsigned long delay) | |
1589 | { | |
1590 | unsigned long flags; | |
1591 | int ret; | |
c7fc77f7 | 1592 | |
8376fe22 TH |
1593 | do { |
1594 | ret = try_to_grab_pending(&dwork->work, true, &flags); | |
1595 | } while (unlikely(ret == -EAGAIN)); | |
63bc0362 | 1596 | |
8376fe22 TH |
1597 | if (likely(ret >= 0)) { |
1598 | __queue_delayed_work(cpu, wq, dwork, delay); | |
1599 | local_irq_restore(flags); | |
7a6bc1cd | 1600 | } |
8376fe22 TH |
1601 | |
1602 | /* -ENOENT from try_to_grab_pending() becomes %true */ | |
7a6bc1cd VP |
1603 | return ret; |
1604 | } | |
8376fe22 TH |
1605 | EXPORT_SYMBOL_GPL(mod_delayed_work_on); |
1606 | ||
c8e55f36 TH |
1607 | /** |
1608 | * worker_enter_idle - enter idle state | |
1609 | * @worker: worker which is entering idle state | |
1610 | * | |
1611 | * @worker is entering idle state. Update stats and idle timer if | |
1612 | * necessary. | |
1613 | * | |
1614 | * LOCKING: | |
d565ed63 | 1615 | * spin_lock_irq(pool->lock). |
c8e55f36 TH |
1616 | */ |
1617 | static void worker_enter_idle(struct worker *worker) | |
1da177e4 | 1618 | { |
bd7bdd43 | 1619 | struct worker_pool *pool = worker->pool; |
c8e55f36 | 1620 | |
6183c009 TH |
1621 | if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || |
1622 | WARN_ON_ONCE(!list_empty(&worker->entry) && | |
1623 | (worker->hentry.next || worker->hentry.pprev))) | |
1624 | return; | |
c8e55f36 | 1625 | |
051e1850 | 1626 | /* can't use worker_set_flags(), also called from create_worker() */ |
cb444766 | 1627 | worker->flags |= WORKER_IDLE; |
bd7bdd43 | 1628 | pool->nr_idle++; |
e22bee78 | 1629 | worker->last_active = jiffies; |
c8e55f36 TH |
1630 | |
1631 | /* idle_list is LIFO */ | |
bd7bdd43 | 1632 | list_add(&worker->entry, &pool->idle_list); |
db7bccf4 | 1633 | |
628c78e7 TH |
1634 | if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) |
1635 | mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); | |
cb444766 | 1636 | |
544ecf31 | 1637 | /* |
706026c2 | 1638 | * Sanity check nr_running. Because wq_unbind_fn() releases |
d565ed63 | 1639 | * pool->lock between setting %WORKER_UNBOUND and zapping |
628c78e7 TH |
1640 | * nr_running, the warning may trigger spuriously. Check iff |
1641 | * unbind is not in progress. | |
544ecf31 | 1642 | */ |
24647570 | 1643 | WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && |
bd7bdd43 | 1644 | pool->nr_workers == pool->nr_idle && |
e19e397a | 1645 | atomic_read(&pool->nr_running)); |
c8e55f36 TH |
1646 | } |
1647 | ||
1648 | /** | |
1649 | * worker_leave_idle - leave idle state | |
1650 | * @worker: worker which is leaving idle state | |
1651 | * | |
1652 | * @worker is leaving idle state. Update stats. | |
1653 | * | |
1654 | * LOCKING: | |
d565ed63 | 1655 | * spin_lock_irq(pool->lock). |
c8e55f36 TH |
1656 | */ |
1657 | static void worker_leave_idle(struct worker *worker) | |
1658 | { | |
bd7bdd43 | 1659 | struct worker_pool *pool = worker->pool; |
c8e55f36 | 1660 | |
6183c009 TH |
1661 | if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) |
1662 | return; | |
d302f017 | 1663 | worker_clr_flags(worker, WORKER_IDLE); |
bd7bdd43 | 1664 | pool->nr_idle--; |
c8e55f36 TH |
1665 | list_del_init(&worker->entry); |
1666 | } | |
1667 | ||
f7537df5 | 1668 | static struct worker *alloc_worker(int node) |
c34056a3 TH |
1669 | { |
1670 | struct worker *worker; | |
1671 | ||
f7537df5 | 1672 | worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node); |
c8e55f36 TH |
1673 | if (worker) { |
1674 | INIT_LIST_HEAD(&worker->entry); | |
affee4b2 | 1675 | INIT_LIST_HEAD(&worker->scheduled); |
da028469 | 1676 | INIT_LIST_HEAD(&worker->node); |
e22bee78 TH |
1677 | /* on creation a worker is in !idle && prep state */ |
1678 | worker->flags = WORKER_PREP; | |
c8e55f36 | 1679 | } |
c34056a3 TH |
1680 | return worker; |
1681 | } | |
1682 | ||
4736cbf7 LJ |
1683 | /** |
1684 | * worker_attach_to_pool() - attach a worker to a pool | |
1685 | * @worker: worker to be attached | |
1686 | * @pool: the target pool | |
1687 | * | |
1688 | * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and | |
1689 | * cpu-binding of @worker are kept coordinated with the pool across | |
1690 | * cpu-[un]hotplugs. | |
1691 | */ | |
1692 | static void worker_attach_to_pool(struct worker *worker, | |
1693 | struct worker_pool *pool) | |
1694 | { | |
1695 | mutex_lock(&pool->attach_mutex); | |
1696 | ||
1697 | /* | |
1698 | * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any | |
1699 | * online CPUs. It'll be re-applied when any of the CPUs come up. | |
1700 | */ | |
1701 | set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask); | |
1702 | ||
1703 | /* | |
1704 | * The pool->attach_mutex ensures %POOL_DISASSOCIATED remains | |
1705 | * stable across this function. See the comments above the | |
1706 | * flag definition for details. | |
1707 | */ | |
1708 | if (pool->flags & POOL_DISASSOCIATED) | |
1709 | worker->flags |= WORKER_UNBOUND; | |
1710 | ||
1711 | list_add_tail(&worker->node, &pool->workers); | |
1712 | ||
1713 | mutex_unlock(&pool->attach_mutex); | |
1714 | } | |
1715 | ||
60f5a4bc LJ |
1716 | /** |
1717 | * worker_detach_from_pool() - detach a worker from its pool | |
1718 | * @worker: worker which is attached to its pool | |
1719 | * @pool: the pool @worker is attached to | |
1720 | * | |
4736cbf7 LJ |
1721 | * Undo the attaching which had been done in worker_attach_to_pool(). The |
1722 | * caller worker shouldn't access to the pool after detached except it has | |
1723 | * other reference to the pool. | |
60f5a4bc LJ |
1724 | */ |
1725 | static void worker_detach_from_pool(struct worker *worker, | |
1726 | struct worker_pool *pool) | |
1727 | { | |
1728 | struct completion *detach_completion = NULL; | |
1729 | ||
92f9c5c4 | 1730 | mutex_lock(&pool->attach_mutex); |
da028469 LJ |
1731 | list_del(&worker->node); |
1732 | if (list_empty(&pool->workers)) | |
60f5a4bc | 1733 | detach_completion = pool->detach_completion; |
92f9c5c4 | 1734 | mutex_unlock(&pool->attach_mutex); |
60f5a4bc | 1735 | |
b62c0751 LJ |
1736 | /* clear leftover flags without pool->lock after it is detached */ |
1737 | worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND); | |
1738 | ||
60f5a4bc LJ |
1739 | if (detach_completion) |
1740 | complete(detach_completion); | |
1741 | } | |
1742 | ||
c34056a3 TH |
1743 | /** |
1744 | * create_worker - create a new workqueue worker | |
63d95a91 | 1745 | * @pool: pool the new worker will belong to |
c34056a3 | 1746 | * |
051e1850 | 1747 | * Create and start a new worker which is attached to @pool. |
c34056a3 TH |
1748 | * |
1749 | * CONTEXT: | |
1750 | * Might sleep. Does GFP_KERNEL allocations. | |
1751 | * | |
d185af30 | 1752 | * Return: |
c34056a3 TH |
1753 | * Pointer to the newly created worker. |
1754 | */ | |
bc2ae0f5 | 1755 | static struct worker *create_worker(struct worker_pool *pool) |
c34056a3 | 1756 | { |
c34056a3 | 1757 | struct worker *worker = NULL; |
f3421797 | 1758 | int id = -1; |
e3c916a4 | 1759 | char id_buf[16]; |
c34056a3 | 1760 | |
7cda9aae LJ |
1761 | /* ID is needed to determine kthread name */ |
1762 | id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL); | |
822d8405 TH |
1763 | if (id < 0) |
1764 | goto fail; | |
c34056a3 | 1765 | |
f7537df5 | 1766 | worker = alloc_worker(pool->node); |
c34056a3 TH |
1767 | if (!worker) |
1768 | goto fail; | |
1769 | ||
bd7bdd43 | 1770 | worker->pool = pool; |
c34056a3 TH |
1771 | worker->id = id; |
1772 | ||
29c91e99 | 1773 | if (pool->cpu >= 0) |
e3c916a4 TH |
1774 | snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, |
1775 | pool->attrs->nice < 0 ? "H" : ""); | |
f3421797 | 1776 | else |
e3c916a4 TH |
1777 | snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); |
1778 | ||
f3f90ad4 | 1779 | worker->task = kthread_create_on_node(worker_thread, worker, pool->node, |
e3c916a4 | 1780 | "kworker/%s", id_buf); |
c34056a3 TH |
1781 | if (IS_ERR(worker->task)) |
1782 | goto fail; | |
1783 | ||
91151228 | 1784 | set_user_nice(worker->task, pool->attrs->nice); |
25834c73 | 1785 | kthread_bind_mask(worker->task, pool->attrs->cpumask); |
91151228 | 1786 | |
da028469 | 1787 | /* successful, attach the worker to the pool */ |
4736cbf7 | 1788 | worker_attach_to_pool(worker, pool); |
822d8405 | 1789 | |
051e1850 LJ |
1790 | /* start the newly created worker */ |
1791 | spin_lock_irq(&pool->lock); | |
1792 | worker->pool->nr_workers++; | |
1793 | worker_enter_idle(worker); | |
1794 | wake_up_process(worker->task); | |
1795 | spin_unlock_irq(&pool->lock); | |
1796 | ||
c34056a3 | 1797 | return worker; |
822d8405 | 1798 | |
c34056a3 | 1799 | fail: |
9625ab17 | 1800 | if (id >= 0) |
7cda9aae | 1801 | ida_simple_remove(&pool->worker_ida, id); |
c34056a3 TH |
1802 | kfree(worker); |
1803 | return NULL; | |
1804 | } | |
1805 | ||
c34056a3 TH |
1806 | /** |
1807 | * destroy_worker - destroy a workqueue worker | |
1808 | * @worker: worker to be destroyed | |
1809 | * | |
73eb7fe7 LJ |
1810 | * Destroy @worker and adjust @pool stats accordingly. The worker should |
1811 | * be idle. | |
c8e55f36 TH |
1812 | * |
1813 | * CONTEXT: | |
60f5a4bc | 1814 | * spin_lock_irq(pool->lock). |
c34056a3 TH |
1815 | */ |
1816 | static void destroy_worker(struct worker *worker) | |
1817 | { | |
bd7bdd43 | 1818 | struct worker_pool *pool = worker->pool; |
c34056a3 | 1819 | |
cd549687 TH |
1820 | lockdep_assert_held(&pool->lock); |
1821 | ||
c34056a3 | 1822 | /* sanity check frenzy */ |
6183c009 | 1823 | if (WARN_ON(worker->current_work) || |
73eb7fe7 LJ |
1824 | WARN_ON(!list_empty(&worker->scheduled)) || |
1825 | WARN_ON(!(worker->flags & WORKER_IDLE))) | |
6183c009 | 1826 | return; |
c34056a3 | 1827 | |
73eb7fe7 LJ |
1828 | pool->nr_workers--; |
1829 | pool->nr_idle--; | |
5bdfff96 | 1830 | |
c8e55f36 | 1831 | list_del_init(&worker->entry); |
cb444766 | 1832 | worker->flags |= WORKER_DIE; |
60f5a4bc | 1833 | wake_up_process(worker->task); |
c34056a3 TH |
1834 | } |
1835 | ||
63d95a91 | 1836 | static void idle_worker_timeout(unsigned long __pool) |
e22bee78 | 1837 | { |
63d95a91 | 1838 | struct worker_pool *pool = (void *)__pool; |
e22bee78 | 1839 | |
d565ed63 | 1840 | spin_lock_irq(&pool->lock); |
e22bee78 | 1841 | |
3347fc9f | 1842 | while (too_many_workers(pool)) { |
e22bee78 TH |
1843 | struct worker *worker; |
1844 | unsigned long expires; | |
1845 | ||
1846 | /* idle_list is kept in LIFO order, check the last one */ | |
63d95a91 | 1847 | worker = list_entry(pool->idle_list.prev, struct worker, entry); |
e22bee78 TH |
1848 | expires = worker->last_active + IDLE_WORKER_TIMEOUT; |
1849 | ||
3347fc9f | 1850 | if (time_before(jiffies, expires)) { |
63d95a91 | 1851 | mod_timer(&pool->idle_timer, expires); |
3347fc9f | 1852 | break; |
d5abe669 | 1853 | } |
3347fc9f LJ |
1854 | |
1855 | destroy_worker(worker); | |
e22bee78 TH |
1856 | } |
1857 | ||
d565ed63 | 1858 | spin_unlock_irq(&pool->lock); |
e22bee78 | 1859 | } |
d5abe669 | 1860 | |
493a1724 | 1861 | static void send_mayday(struct work_struct *work) |
e22bee78 | 1862 | { |
112202d9 TH |
1863 | struct pool_workqueue *pwq = get_work_pwq(work); |
1864 | struct workqueue_struct *wq = pwq->wq; | |
493a1724 | 1865 | |
2e109a28 | 1866 | lockdep_assert_held(&wq_mayday_lock); |
e22bee78 | 1867 | |
493008a8 | 1868 | if (!wq->rescuer) |
493a1724 | 1869 | return; |
e22bee78 TH |
1870 | |
1871 | /* mayday mayday mayday */ | |
493a1724 | 1872 | if (list_empty(&pwq->mayday_node)) { |
77668c8b LJ |
1873 | /* |
1874 | * If @pwq is for an unbound wq, its base ref may be put at | |
1875 | * any time due to an attribute change. Pin @pwq until the | |
1876 | * rescuer is done with it. | |
1877 | */ | |
1878 | get_pwq(pwq); | |
493a1724 | 1879 | list_add_tail(&pwq->mayday_node, &wq->maydays); |
e22bee78 | 1880 | wake_up_process(wq->rescuer->task); |
493a1724 | 1881 | } |
e22bee78 TH |
1882 | } |
1883 | ||
706026c2 | 1884 | static void pool_mayday_timeout(unsigned long __pool) |
e22bee78 | 1885 | { |
63d95a91 | 1886 | struct worker_pool *pool = (void *)__pool; |
e22bee78 TH |
1887 | struct work_struct *work; |
1888 | ||
b2d82909 TH |
1889 | spin_lock_irq(&pool->lock); |
1890 | spin_lock(&wq_mayday_lock); /* for wq->maydays */ | |
e22bee78 | 1891 | |
63d95a91 | 1892 | if (need_to_create_worker(pool)) { |
e22bee78 TH |
1893 | /* |
1894 | * We've been trying to create a new worker but | |
1895 | * haven't been successful. We might be hitting an | |
1896 | * allocation deadlock. Send distress signals to | |
1897 | * rescuers. | |
1898 | */ | |
63d95a91 | 1899 | list_for_each_entry(work, &pool->worklist, entry) |
e22bee78 | 1900 | send_mayday(work); |
1da177e4 | 1901 | } |
e22bee78 | 1902 | |
b2d82909 TH |
1903 | spin_unlock(&wq_mayday_lock); |
1904 | spin_unlock_irq(&pool->lock); | |
e22bee78 | 1905 | |
63d95a91 | 1906 | mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); |
1da177e4 LT |
1907 | } |
1908 | ||
e22bee78 TH |
1909 | /** |
1910 | * maybe_create_worker - create a new worker if necessary | |
63d95a91 | 1911 | * @pool: pool to create a new worker for |
e22bee78 | 1912 | * |
63d95a91 | 1913 | * Create a new worker for @pool if necessary. @pool is guaranteed to |
e22bee78 TH |
1914 | * have at least one idle worker on return from this function. If |
1915 | * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is | |
63d95a91 | 1916 | * sent to all rescuers with works scheduled on @pool to resolve |
e22bee78 TH |
1917 | * possible allocation deadlock. |
1918 | * | |
c5aa87bb TH |
1919 | * On return, need_to_create_worker() is guaranteed to be %false and |
1920 | * may_start_working() %true. | |
e22bee78 TH |
1921 | * |
1922 | * LOCKING: | |
d565ed63 | 1923 | * spin_lock_irq(pool->lock) which may be released and regrabbed |
e22bee78 TH |
1924 | * multiple times. Does GFP_KERNEL allocations. Called only from |
1925 | * manager. | |
e22bee78 | 1926 | */ |
29187a9e | 1927 | static void maybe_create_worker(struct worker_pool *pool) |
d565ed63 TH |
1928 | __releases(&pool->lock) |
1929 | __acquires(&pool->lock) | |
1da177e4 | 1930 | { |
e22bee78 | 1931 | restart: |
d565ed63 | 1932 | spin_unlock_irq(&pool->lock); |
9f9c2364 | 1933 | |
e22bee78 | 1934 | /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ |
63d95a91 | 1935 | mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); |
e22bee78 TH |
1936 | |
1937 | while (true) { | |
051e1850 | 1938 | if (create_worker(pool) || !need_to_create_worker(pool)) |
e22bee78 | 1939 | break; |
1da177e4 | 1940 | |
e212f361 | 1941 | schedule_timeout_interruptible(CREATE_COOLDOWN); |
9f9c2364 | 1942 | |
63d95a91 | 1943 | if (!need_to_create_worker(pool)) |
e22bee78 TH |
1944 | break; |
1945 | } | |
1946 | ||
63d95a91 | 1947 | del_timer_sync(&pool->mayday_timer); |
d565ed63 | 1948 | spin_lock_irq(&pool->lock); |
051e1850 LJ |
1949 | /* |
1950 | * This is necessary even after a new worker was just successfully | |
1951 | * created as @pool->lock was dropped and the new worker might have | |
1952 | * already become busy. | |
1953 | */ | |
63d95a91 | 1954 | if (need_to_create_worker(pool)) |
e22bee78 | 1955 | goto restart; |
e22bee78 TH |
1956 | } |
1957 | ||
73f53c4a | 1958 | /** |
e22bee78 TH |
1959 | * manage_workers - manage worker pool |
1960 | * @worker: self | |
73f53c4a | 1961 | * |
706026c2 | 1962 | * Assume the manager role and manage the worker pool @worker belongs |
e22bee78 | 1963 | * to. At any given time, there can be only zero or one manager per |
706026c2 | 1964 | * pool. The exclusion is handled automatically by this function. |
e22bee78 TH |
1965 | * |
1966 | * The caller can safely start processing works on false return. On | |
1967 | * true return, it's guaranteed that need_to_create_worker() is false | |
1968 | * and may_start_working() is true. | |
73f53c4a TH |
1969 | * |
1970 | * CONTEXT: | |
d565ed63 | 1971 | * spin_lock_irq(pool->lock) which may be released and regrabbed |
e22bee78 TH |
1972 | * multiple times. Does GFP_KERNEL allocations. |
1973 | * | |
d185af30 | 1974 | * Return: |
29187a9e TH |
1975 | * %false if the pool doesn't need management and the caller can safely |
1976 | * start processing works, %true if management function was performed and | |
1977 | * the conditions that the caller verified before calling the function may | |
1978 | * no longer be true. | |
73f53c4a | 1979 | */ |
e22bee78 | 1980 | static bool manage_workers(struct worker *worker) |
73f53c4a | 1981 | { |
63d95a91 | 1982 | struct worker_pool *pool = worker->pool; |
73f53c4a | 1983 | |
bc3a1afc | 1984 | /* |
bc3a1afc TH |
1985 | * Anyone who successfully grabs manager_arb wins the arbitration |
1986 | * and becomes the manager. mutex_trylock() on pool->manager_arb | |
1987 | * failure while holding pool->lock reliably indicates that someone | |
1988 | * else is managing the pool and the worker which failed trylock | |
1989 | * can proceed to executing work items. This means that anyone | |
1990 | * grabbing manager_arb is responsible for actually performing | |
1991 | * manager duties. If manager_arb is grabbed and released without | |
1992 | * actual management, the pool may stall indefinitely. | |
bc3a1afc | 1993 | */ |
34a06bd6 | 1994 | if (!mutex_trylock(&pool->manager_arb)) |
29187a9e | 1995 | return false; |
2607d7a6 | 1996 | pool->manager = worker; |
1e19ffc6 | 1997 | |
29187a9e | 1998 | maybe_create_worker(pool); |
e22bee78 | 1999 | |
2607d7a6 | 2000 | pool->manager = NULL; |
34a06bd6 | 2001 | mutex_unlock(&pool->manager_arb); |
29187a9e | 2002 | return true; |
73f53c4a TH |
2003 | } |
2004 | ||
a62428c0 TH |
2005 | /** |
2006 | * process_one_work - process single work | |
c34056a3 | 2007 | * @worker: self |
a62428c0 TH |
2008 | * @work: work to process |
2009 | * | |
2010 | * Process @work. This function contains all the logics necessary to | |
2011 | * process a single work including synchronization against and | |
2012 | * interaction with other workers on the same cpu, queueing and | |
2013 | * flushing. As long as context requirement is met, any worker can | |
2014 | * call this function to process a work. | |
2015 | * | |
2016 | * CONTEXT: | |
d565ed63 | 2017 | * spin_lock_irq(pool->lock) which is released and regrabbed. |
a62428c0 | 2018 | */ |
c34056a3 | 2019 | static void process_one_work(struct worker *worker, struct work_struct *work) |
d565ed63 TH |
2020 | __releases(&pool->lock) |
2021 | __acquires(&pool->lock) | |
a62428c0 | 2022 | { |
112202d9 | 2023 | struct pool_workqueue *pwq = get_work_pwq(work); |
bd7bdd43 | 2024 | struct worker_pool *pool = worker->pool; |
112202d9 | 2025 | bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE; |
73f53c4a | 2026 | int work_color; |
7e11629d | 2027 | struct worker *collision; |
a62428c0 TH |
2028 | #ifdef CONFIG_LOCKDEP |
2029 | /* | |
2030 | * It is permissible to free the struct work_struct from | |
2031 | * inside the function that is called from it, this we need to | |
2032 | * take into account for lockdep too. To avoid bogus "held | |
2033 | * lock freed" warnings as well as problems when looking into | |
2034 | * work->lockdep_map, make a copy and use that here. | |
2035 | */ | |
4d82a1de PZ |
2036 | struct lockdep_map lockdep_map; |
2037 | ||
2038 | lockdep_copy_map(&lockdep_map, &work->lockdep_map); | |
a62428c0 | 2039 | #endif |
807407c0 | 2040 | /* ensure we're on the correct CPU */ |
85327af6 | 2041 | WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && |
ec22ca5e | 2042 | raw_smp_processor_id() != pool->cpu); |
25511a47 | 2043 | |
7e11629d TH |
2044 | /* |
2045 | * A single work shouldn't be executed concurrently by | |
2046 | * multiple workers on a single cpu. Check whether anyone is | |
2047 | * already processing the work. If so, defer the work to the | |
2048 | * currently executing one. | |
2049 | */ | |
c9e7cf27 | 2050 | collision = find_worker_executing_work(pool, work); |
7e11629d TH |
2051 | if (unlikely(collision)) { |
2052 | move_linked_works(work, &collision->scheduled, NULL); | |
2053 | return; | |
2054 | } | |
2055 | ||
8930caba | 2056 | /* claim and dequeue */ |
a62428c0 | 2057 | debug_work_deactivate(work); |
c9e7cf27 | 2058 | hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); |
c34056a3 | 2059 | worker->current_work = work; |
a2c1c57b | 2060 | worker->current_func = work->func; |
112202d9 | 2061 | worker->current_pwq = pwq; |
73f53c4a | 2062 | work_color = get_work_color(work); |
7a22ad75 | 2063 | |
a62428c0 TH |
2064 | list_del_init(&work->entry); |
2065 | ||
fb0e7beb | 2066 | /* |
228f1d00 LJ |
2067 | * CPU intensive works don't participate in concurrency management. |
2068 | * They're the scheduler's responsibility. This takes @worker out | |
2069 | * of concurrency management and the next code block will chain | |
2070 | * execution of the pending work items. | |
fb0e7beb TH |
2071 | */ |
2072 | if (unlikely(cpu_intensive)) | |
228f1d00 | 2073 | worker_set_flags(worker, WORKER_CPU_INTENSIVE); |
fb0e7beb | 2074 | |
974271c4 | 2075 | /* |
a489a03e LJ |
2076 | * Wake up another worker if necessary. The condition is always |
2077 | * false for normal per-cpu workers since nr_running would always | |
2078 | * be >= 1 at this point. This is used to chain execution of the | |
2079 | * pending work items for WORKER_NOT_RUNNING workers such as the | |
228f1d00 | 2080 | * UNBOUND and CPU_INTENSIVE ones. |
974271c4 | 2081 | */ |
a489a03e | 2082 | if (need_more_worker(pool)) |
63d95a91 | 2083 | wake_up_worker(pool); |
974271c4 | 2084 | |
8930caba | 2085 | /* |
7c3eed5c | 2086 | * Record the last pool and clear PENDING which should be the last |
d565ed63 | 2087 | * update to @work. Also, do this inside @pool->lock so that |
23657bb1 TH |
2088 | * PENDING and queued state changes happen together while IRQ is |
2089 | * disabled. | |
8930caba | 2090 | */ |
7c3eed5c | 2091 | set_work_pool_and_clear_pending(work, pool->id); |
a62428c0 | 2092 | |
d565ed63 | 2093 | spin_unlock_irq(&pool->lock); |
a62428c0 | 2094 | |
112202d9 | 2095 | lock_map_acquire_read(&pwq->wq->lockdep_map); |
a62428c0 | 2096 | lock_map_acquire(&lockdep_map); |
e36c886a | 2097 | trace_workqueue_execute_start(work); |
a2c1c57b | 2098 | worker->current_func(work); |
e36c886a AV |
2099 | /* |
2100 | * While we must be careful to not use "work" after this, the trace | |
2101 | * point will only record its address. | |
2102 | */ | |
2103 | trace_workqueue_execute_end(work); | |
a62428c0 | 2104 | lock_map_release(&lockdep_map); |
112202d9 | 2105 | lock_map_release(&pwq->wq->lockdep_map); |
a62428c0 TH |
2106 | |
2107 | if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { | |
044c782c VI |
2108 | pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" |
2109 | " last function: %pf\n", | |
a2c1c57b TH |
2110 | current->comm, preempt_count(), task_pid_nr(current), |
2111 | worker->current_func); | |
a62428c0 TH |
2112 | debug_show_held_locks(current); |
2113 | dump_stack(); | |
2114 | } | |
2115 | ||
b22ce278 TH |
2116 | /* |
2117 | * The following prevents a kworker from hogging CPU on !PREEMPT | |
2118 | * kernels, where a requeueing work item waiting for something to | |
2119 | * happen could deadlock with stop_machine as such work item could | |
2120 | * indefinitely requeue itself while all other CPUs are trapped in | |
789cbbec JL |
2121 | * stop_machine. At the same time, report a quiescent RCU state so |
2122 | * the same condition doesn't freeze RCU. | |
b22ce278 | 2123 | */ |
3e28e377 | 2124 | cond_resched_rcu_qs(); |
b22ce278 | 2125 | |
d565ed63 | 2126 | spin_lock_irq(&pool->lock); |
a62428c0 | 2127 | |
fb0e7beb TH |
2128 | /* clear cpu intensive status */ |
2129 | if (unlikely(cpu_intensive)) | |
2130 | worker_clr_flags(worker, WORKER_CPU_INTENSIVE); | |
2131 | ||
a62428c0 | 2132 | /* we're done with it, release */ |
42f8570f | 2133 | hash_del(&worker->hentry); |
c34056a3 | 2134 | worker->current_work = NULL; |
a2c1c57b | 2135 | worker->current_func = NULL; |
112202d9 | 2136 | worker->current_pwq = NULL; |
3d1cb205 | 2137 | worker->desc_valid = false; |
112202d9 | 2138 | pwq_dec_nr_in_flight(pwq, work_color); |
a62428c0 TH |
2139 | } |
2140 | ||
affee4b2 TH |
2141 | /** |
2142 | * process_scheduled_works - process scheduled works | |
2143 | * @worker: self | |
2144 | * | |
2145 | * Process all scheduled works. Please note that the scheduled list | |
2146 | * may change while processing a work, so this function repeatedly | |
2147 | * fetches a work from the top and executes it. | |
2148 | * | |
2149 | * CONTEXT: | |
d565ed63 | 2150 | * spin_lock_irq(pool->lock) which may be released and regrabbed |
affee4b2 TH |
2151 | * multiple times. |
2152 | */ | |
2153 | static void process_scheduled_works(struct worker *worker) | |
1da177e4 | 2154 | { |
affee4b2 TH |
2155 | while (!list_empty(&worker->scheduled)) { |
2156 | struct work_struct *work = list_first_entry(&worker->scheduled, | |
1da177e4 | 2157 | struct work_struct, entry); |
c34056a3 | 2158 | process_one_work(worker, work); |
1da177e4 | 2159 | } |
1da177e4 LT |
2160 | } |
2161 | ||
4690c4ab TH |
2162 | /** |
2163 | * worker_thread - the worker thread function | |
c34056a3 | 2164 | * @__worker: self |
4690c4ab | 2165 | * |
c5aa87bb TH |
2166 | * The worker thread function. All workers belong to a worker_pool - |
2167 | * either a per-cpu one or dynamic unbound one. These workers process all | |
2168 | * work items regardless of their specific target workqueue. The only | |
2169 | * exception is work items which belong to workqueues with a rescuer which | |
2170 | * will be explained in rescuer_thread(). | |
d185af30 YB |
2171 | * |
2172 | * Return: 0 | |
4690c4ab | 2173 | */ |
c34056a3 | 2174 | static int worker_thread(void *__worker) |
1da177e4 | 2175 | { |
c34056a3 | 2176 | struct worker *worker = __worker; |
bd7bdd43 | 2177 | struct worker_pool *pool = worker->pool; |
1da177e4 | 2178 | |
e22bee78 TH |
2179 | /* tell the scheduler that this is a workqueue worker */ |
2180 | worker->task->flags |= PF_WQ_WORKER; | |
c8e55f36 | 2181 | woke_up: |
d565ed63 | 2182 | spin_lock_irq(&pool->lock); |
1da177e4 | 2183 | |
a9ab775b TH |
2184 | /* am I supposed to die? */ |
2185 | if (unlikely(worker->flags & WORKER_DIE)) { | |
d565ed63 | 2186 | spin_unlock_irq(&pool->lock); |
a9ab775b TH |
2187 | WARN_ON_ONCE(!list_empty(&worker->entry)); |
2188 | worker->task->flags &= ~PF_WQ_WORKER; | |
60f5a4bc LJ |
2189 | |
2190 | set_task_comm(worker->task, "kworker/dying"); | |
7cda9aae | 2191 | ida_simple_remove(&pool->worker_ida, worker->id); |
60f5a4bc LJ |
2192 | worker_detach_from_pool(worker, pool); |
2193 | kfree(worker); | |
a9ab775b | 2194 | return 0; |
c8e55f36 | 2195 | } |
affee4b2 | 2196 | |
c8e55f36 | 2197 | worker_leave_idle(worker); |
db7bccf4 | 2198 | recheck: |
e22bee78 | 2199 | /* no more worker necessary? */ |
63d95a91 | 2200 | if (!need_more_worker(pool)) |
e22bee78 TH |
2201 | goto sleep; |
2202 | ||
2203 | /* do we need to manage? */ | |
63d95a91 | 2204 | if (unlikely(!may_start_working(pool)) && manage_workers(worker)) |
e22bee78 TH |
2205 | goto recheck; |
2206 | ||
c8e55f36 TH |
2207 | /* |
2208 | * ->scheduled list can only be filled while a worker is | |
2209 | * preparing to process a work or actually processing it. | |
2210 | * Make sure nobody diddled with it while I was sleeping. | |
2211 | */ | |
6183c009 | 2212 | WARN_ON_ONCE(!list_empty(&worker->scheduled)); |
c8e55f36 | 2213 | |
e22bee78 | 2214 | /* |
a9ab775b TH |
2215 | * Finish PREP stage. We're guaranteed to have at least one idle |
2216 | * worker or that someone else has already assumed the manager | |
2217 | * role. This is where @worker starts participating in concurrency | |
2218 | * management if applicable and concurrency management is restored | |
2219 | * after being rebound. See rebind_workers() for details. | |
e22bee78 | 2220 | */ |
a9ab775b | 2221 | worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); |
e22bee78 TH |
2222 | |
2223 | do { | |
c8e55f36 | 2224 | struct work_struct *work = |
bd7bdd43 | 2225 | list_first_entry(&pool->worklist, |
c8e55f36 TH |
2226 | struct work_struct, entry); |
2227 | ||
82607adc TH |
2228 | pool->watchdog_ts = jiffies; |
2229 | ||
c8e55f36 TH |
2230 | if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { |
2231 | /* optimization path, not strictly necessary */ | |
2232 | process_one_work(worker, work); | |
2233 | if (unlikely(!list_empty(&worker->scheduled))) | |
affee4b2 | 2234 | process_scheduled_works(worker); |
c8e55f36 TH |
2235 | } else { |
2236 | move_linked_works(work, &worker->scheduled, NULL); | |
2237 | process_scheduled_works(worker); | |
affee4b2 | 2238 | } |
63d95a91 | 2239 | } while (keep_working(pool)); |
e22bee78 | 2240 | |
228f1d00 | 2241 | worker_set_flags(worker, WORKER_PREP); |
d313dd85 | 2242 | sleep: |
c8e55f36 | 2243 | /* |
d565ed63 TH |
2244 | * pool->lock is held and there's no work to process and no need to |
2245 | * manage, sleep. Workers are woken up only while holding | |
2246 | * pool->lock or from local cpu, so setting the current state | |
2247 | * before releasing pool->lock is enough to prevent losing any | |
2248 | * event. | |
c8e55f36 TH |
2249 | */ |
2250 | worker_enter_idle(worker); | |
2251 | __set_current_state(TASK_INTERRUPTIBLE); | |
d565ed63 | 2252 | spin_unlock_irq(&pool->lock); |
c8e55f36 TH |
2253 | schedule(); |
2254 | goto woke_up; | |
1da177e4 LT |
2255 | } |
2256 | ||
e22bee78 TH |
2257 | /** |
2258 | * rescuer_thread - the rescuer thread function | |
111c225a | 2259 | * @__rescuer: self |
e22bee78 TH |
2260 | * |
2261 | * Workqueue rescuer thread function. There's one rescuer for each | |
493008a8 | 2262 | * workqueue which has WQ_MEM_RECLAIM set. |
e22bee78 | 2263 | * |
706026c2 | 2264 | * Regular work processing on a pool may block trying to create a new |
e22bee78 TH |
2265 | * worker which uses GFP_KERNEL allocation which has slight chance of |
2266 | * developing into deadlock if some works currently on the same queue | |
2267 | * need to be processed to satisfy the GFP_KERNEL allocation. This is | |
2268 | * the problem rescuer solves. | |
2269 | * | |
706026c2 TH |
2270 | * When such condition is possible, the pool summons rescuers of all |
2271 | * workqueues which have works queued on the pool and let them process | |
e22bee78 TH |
2272 | * those works so that forward progress can be guaranteed. |
2273 | * | |
2274 | * This should happen rarely. | |
d185af30 YB |
2275 | * |
2276 | * Return: 0 | |
e22bee78 | 2277 | */ |
111c225a | 2278 | static int rescuer_thread(void *__rescuer) |
e22bee78 | 2279 | { |
111c225a TH |
2280 | struct worker *rescuer = __rescuer; |
2281 | struct workqueue_struct *wq = rescuer->rescue_wq; | |
e22bee78 | 2282 | struct list_head *scheduled = &rescuer->scheduled; |
4d595b86 | 2283 | bool should_stop; |
e22bee78 TH |
2284 | |
2285 | set_user_nice(current, RESCUER_NICE_LEVEL); | |
111c225a TH |
2286 | |
2287 | /* | |
2288 | * Mark rescuer as worker too. As WORKER_PREP is never cleared, it | |
2289 | * doesn't participate in concurrency management. | |
2290 | */ | |
2291 | rescuer->task->flags |= PF_WQ_WORKER; | |
e22bee78 TH |
2292 | repeat: |
2293 | set_current_state(TASK_INTERRUPTIBLE); | |
2294 | ||
4d595b86 LJ |
2295 | /* |
2296 | * By the time the rescuer is requested to stop, the workqueue | |
2297 | * shouldn't have any work pending, but @wq->maydays may still have | |
2298 | * pwq(s) queued. This can happen by non-rescuer workers consuming | |
2299 | * all the work items before the rescuer got to them. Go through | |
2300 | * @wq->maydays processing before acting on should_stop so that the | |
2301 | * list is always empty on exit. | |
2302 | */ | |
2303 | should_stop = kthread_should_stop(); | |
e22bee78 | 2304 | |
493a1724 | 2305 | /* see whether any pwq is asking for help */ |
2e109a28 | 2306 | spin_lock_irq(&wq_mayday_lock); |
493a1724 TH |
2307 | |
2308 | while (!list_empty(&wq->maydays)) { | |
2309 | struct pool_workqueue *pwq = list_first_entry(&wq->maydays, | |
2310 | struct pool_workqueue, mayday_node); | |
112202d9 | 2311 | struct worker_pool *pool = pwq->pool; |
e22bee78 | 2312 | struct work_struct *work, *n; |
82607adc | 2313 | bool first = true; |
e22bee78 TH |
2314 | |
2315 | __set_current_state(TASK_RUNNING); | |
493a1724 TH |
2316 | list_del_init(&pwq->mayday_node); |
2317 | ||
2e109a28 | 2318 | spin_unlock_irq(&wq_mayday_lock); |
e22bee78 | 2319 | |
51697d39 LJ |
2320 | worker_attach_to_pool(rescuer, pool); |
2321 | ||
2322 | spin_lock_irq(&pool->lock); | |
b3104104 | 2323 | rescuer->pool = pool; |
e22bee78 TH |
2324 | |
2325 | /* | |
2326 | * Slurp in all works issued via this workqueue and | |
2327 | * process'em. | |
2328 | */ | |
0479c8c5 | 2329 | WARN_ON_ONCE(!list_empty(scheduled)); |
82607adc TH |
2330 | list_for_each_entry_safe(work, n, &pool->worklist, entry) { |
2331 | if (get_work_pwq(work) == pwq) { | |
2332 | if (first) | |
2333 | pool->watchdog_ts = jiffies; | |
e22bee78 | 2334 | move_linked_works(work, scheduled, &n); |
82607adc TH |
2335 | } |
2336 | first = false; | |
2337 | } | |
e22bee78 | 2338 | |
008847f6 N |
2339 | if (!list_empty(scheduled)) { |
2340 | process_scheduled_works(rescuer); | |
2341 | ||
2342 | /* | |
2343 | * The above execution of rescued work items could | |
2344 | * have created more to rescue through | |
2345 | * pwq_activate_first_delayed() or chained | |
2346 | * queueing. Let's put @pwq back on mayday list so | |
2347 | * that such back-to-back work items, which may be | |
2348 | * being used to relieve memory pressure, don't | |
2349 | * incur MAYDAY_INTERVAL delay inbetween. | |
2350 | */ | |
2351 | if (need_to_create_worker(pool)) { | |
2352 | spin_lock(&wq_mayday_lock); | |
2353 | get_pwq(pwq); | |
2354 | list_move_tail(&pwq->mayday_node, &wq->maydays); | |
2355 | spin_unlock(&wq_mayday_lock); | |
2356 | } | |
2357 | } | |
7576958a | 2358 | |
77668c8b LJ |
2359 | /* |
2360 | * Put the reference grabbed by send_mayday(). @pool won't | |
13b1d625 | 2361 | * go away while we're still attached to it. |
77668c8b LJ |
2362 | */ |
2363 | put_pwq(pwq); | |
2364 | ||
7576958a | 2365 | /* |
d8ca83e6 | 2366 | * Leave this pool. If need_more_worker() is %true, notify a |
7576958a TH |
2367 | * regular worker; otherwise, we end up with 0 concurrency |
2368 | * and stalling the execution. | |
2369 | */ | |
d8ca83e6 | 2370 | if (need_more_worker(pool)) |
63d95a91 | 2371 | wake_up_worker(pool); |
7576958a | 2372 | |
b3104104 | 2373 | rescuer->pool = NULL; |
13b1d625 LJ |
2374 | spin_unlock_irq(&pool->lock); |
2375 | ||
2376 | worker_detach_from_pool(rescuer, pool); | |
2377 | ||
2378 | spin_lock_irq(&wq_mayday_lock); | |
e22bee78 TH |
2379 | } |
2380 | ||
2e109a28 | 2381 | spin_unlock_irq(&wq_mayday_lock); |
493a1724 | 2382 | |
4d595b86 LJ |
2383 | if (should_stop) { |
2384 | __set_current_state(TASK_RUNNING); | |
2385 | rescuer->task->flags &= ~PF_WQ_WORKER; | |
2386 | return 0; | |
2387 | } | |
2388 | ||
111c225a TH |
2389 | /* rescuers should never participate in concurrency management */ |
2390 | WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); | |
e22bee78 TH |
2391 | schedule(); |
2392 | goto repeat; | |
1da177e4 LT |
2393 | } |
2394 | ||
fca839c0 TH |
2395 | /** |
2396 | * check_flush_dependency - check for flush dependency sanity | |
2397 | * @target_wq: workqueue being flushed | |
2398 | * @target_work: work item being flushed (NULL for workqueue flushes) | |
2399 | * | |
2400 | * %current is trying to flush the whole @target_wq or @target_work on it. | |
2401 | * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not | |
2402 | * reclaiming memory or running on a workqueue which doesn't have | |
2403 | * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to | |
2404 | * a deadlock. | |
2405 | */ | |
2406 | static void check_flush_dependency(struct workqueue_struct *target_wq, | |
2407 | struct work_struct *target_work) | |
2408 | { | |
2409 | work_func_t target_func = target_work ? target_work->func : NULL; | |
2410 | struct worker *worker; | |
2411 | ||
2412 | if (target_wq->flags & WQ_MEM_RECLAIM) | |
2413 | return; | |
2414 | ||
2415 | worker = current_wq_worker(); | |
2416 | ||
2417 | WARN_ONCE(current->flags & PF_MEMALLOC, | |
2418 | "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%pf", | |
2419 | current->pid, current->comm, target_wq->name, target_func); | |
23d11a58 TH |
2420 | WARN_ONCE(worker && ((worker->current_pwq->wq->flags & |
2421 | (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM), | |
fca839c0 TH |
2422 | "workqueue: WQ_MEM_RECLAIM %s:%pf is flushing !WQ_MEM_RECLAIM %s:%pf", |
2423 | worker->current_pwq->wq->name, worker->current_func, | |
2424 | target_wq->name, target_func); | |
2425 | } | |
2426 | ||
fc2e4d70 ON |
2427 | struct wq_barrier { |
2428 | struct work_struct work; | |
2429 | struct completion done; | |
2607d7a6 | 2430 | struct task_struct *task; /* purely informational */ |
fc2e4d70 ON |
2431 | }; |
2432 | ||
2433 | static void wq_barrier_func(struct work_struct *work) | |
2434 | { | |
2435 | struct wq_barrier *barr = container_of(work, struct wq_barrier, work); | |
2436 | complete(&barr->done); | |
2437 | } | |
2438 | ||
4690c4ab TH |
2439 | /** |
2440 | * insert_wq_barrier - insert a barrier work | |
112202d9 | 2441 | * @pwq: pwq to insert barrier into |
4690c4ab | 2442 | * @barr: wq_barrier to insert |
affee4b2 TH |
2443 | * @target: target work to attach @barr to |
2444 | * @worker: worker currently executing @target, NULL if @target is not executing | |
4690c4ab | 2445 | * |
affee4b2 TH |
2446 | * @barr is linked to @target such that @barr is completed only after |
2447 | * @target finishes execution. Please note that the ordering | |
2448 | * guarantee is observed only with respect to @target and on the local | |
2449 | * cpu. | |
2450 | * | |
2451 | * Currently, a queued barrier can't be canceled. This is because | |
2452 | * try_to_grab_pending() can't determine whether the work to be | |
2453 | * grabbed is at the head of the queue and thus can't clear LINKED | |
2454 | * flag of the previous work while there must be a valid next work | |
2455 | * after a work with LINKED flag set. | |
2456 | * | |
2457 | * Note that when @worker is non-NULL, @target may be modified | |
112202d9 | 2458 | * underneath us, so we can't reliably determine pwq from @target. |
4690c4ab TH |
2459 | * |
2460 | * CONTEXT: | |
d565ed63 | 2461 | * spin_lock_irq(pool->lock). |
4690c4ab | 2462 | */ |
112202d9 | 2463 | static void insert_wq_barrier(struct pool_workqueue *pwq, |
affee4b2 TH |
2464 | struct wq_barrier *barr, |
2465 | struct work_struct *target, struct worker *worker) | |
fc2e4d70 | 2466 | { |
affee4b2 TH |
2467 | struct list_head *head; |
2468 | unsigned int linked = 0; | |
2469 | ||
dc186ad7 | 2470 | /* |
d565ed63 | 2471 | * debugobject calls are safe here even with pool->lock locked |
dc186ad7 TG |
2472 | * as we know for sure that this will not trigger any of the |
2473 | * checks and call back into the fixup functions where we | |
2474 | * might deadlock. | |
2475 | */ | |
ca1cab37 | 2476 | INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); |
22df02bb | 2477 | __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); |
fc2e4d70 | 2478 | init_completion(&barr->done); |
2607d7a6 | 2479 | barr->task = current; |
83c22520 | 2480 | |
affee4b2 TH |
2481 | /* |
2482 | * If @target is currently being executed, schedule the | |
2483 | * barrier to the worker; otherwise, put it after @target. | |
2484 | */ | |
2485 | if (worker) | |
2486 | head = worker->scheduled.next; | |
2487 | else { | |
2488 | unsigned long *bits = work_data_bits(target); | |
2489 | ||
2490 | head = target->entry.next; | |
2491 | /* there can already be other linked works, inherit and set */ | |
2492 | linked = *bits & WORK_STRUCT_LINKED; | |
2493 | __set_bit(WORK_STRUCT_LINKED_BIT, bits); | |
2494 | } | |
2495 | ||
dc186ad7 | 2496 | debug_work_activate(&barr->work); |
112202d9 | 2497 | insert_work(pwq, &barr->work, head, |
affee4b2 | 2498 | work_color_to_flags(WORK_NO_COLOR) | linked); |
fc2e4d70 ON |
2499 | } |
2500 | ||
73f53c4a | 2501 | /** |
112202d9 | 2502 | * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing |
73f53c4a TH |
2503 | * @wq: workqueue being flushed |
2504 | * @flush_color: new flush color, < 0 for no-op | |
2505 | * @work_color: new work color, < 0 for no-op | |
2506 | * | |
112202d9 | 2507 | * Prepare pwqs for workqueue flushing. |
73f53c4a | 2508 | * |
112202d9 TH |
2509 | * If @flush_color is non-negative, flush_color on all pwqs should be |
2510 | * -1. If no pwq has in-flight commands at the specified color, all | |
2511 | * pwq->flush_color's stay at -1 and %false is returned. If any pwq | |
2512 | * has in flight commands, its pwq->flush_color is set to | |
2513 | * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq | |
73f53c4a TH |
2514 | * wakeup logic is armed and %true is returned. |
2515 | * | |
2516 | * The caller should have initialized @wq->first_flusher prior to | |
2517 | * calling this function with non-negative @flush_color. If | |
2518 | * @flush_color is negative, no flush color update is done and %false | |
2519 | * is returned. | |
2520 | * | |
112202d9 | 2521 | * If @work_color is non-negative, all pwqs should have the same |
73f53c4a TH |
2522 | * work_color which is previous to @work_color and all will be |
2523 | * advanced to @work_color. | |
2524 | * | |
2525 | * CONTEXT: | |
3c25a55d | 2526 | * mutex_lock(wq->mutex). |
73f53c4a | 2527 | * |
d185af30 | 2528 | * Return: |
73f53c4a TH |
2529 | * %true if @flush_color >= 0 and there's something to flush. %false |
2530 | * otherwise. | |
2531 | */ | |
112202d9 | 2532 | static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, |
73f53c4a | 2533 | int flush_color, int work_color) |
1da177e4 | 2534 | { |
73f53c4a | 2535 | bool wait = false; |
49e3cf44 | 2536 | struct pool_workqueue *pwq; |
1da177e4 | 2537 | |
73f53c4a | 2538 | if (flush_color >= 0) { |
6183c009 | 2539 | WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); |
112202d9 | 2540 | atomic_set(&wq->nr_pwqs_to_flush, 1); |
1da177e4 | 2541 | } |
2355b70f | 2542 | |
49e3cf44 | 2543 | for_each_pwq(pwq, wq) { |
112202d9 | 2544 | struct worker_pool *pool = pwq->pool; |
fc2e4d70 | 2545 | |
b09f4fd3 | 2546 | spin_lock_irq(&pool->lock); |
83c22520 | 2547 | |
73f53c4a | 2548 | if (flush_color >= 0) { |
6183c009 | 2549 | WARN_ON_ONCE(pwq->flush_color != -1); |
fc2e4d70 | 2550 | |
112202d9 TH |
2551 | if (pwq->nr_in_flight[flush_color]) { |
2552 | pwq->flush_color = flush_color; | |
2553 | atomic_inc(&wq->nr_pwqs_to_flush); | |
73f53c4a TH |
2554 | wait = true; |
2555 | } | |
2556 | } | |
1da177e4 | 2557 | |
73f53c4a | 2558 | if (work_color >= 0) { |
6183c009 | 2559 | WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); |
112202d9 | 2560 | pwq->work_color = work_color; |
73f53c4a | 2561 | } |
1da177e4 | 2562 | |
b09f4fd3 | 2563 | spin_unlock_irq(&pool->lock); |
1da177e4 | 2564 | } |
2355b70f | 2565 | |
112202d9 | 2566 | if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) |
73f53c4a | 2567 | complete(&wq->first_flusher->done); |
14441960 | 2568 | |
73f53c4a | 2569 | return wait; |
1da177e4 LT |
2570 | } |
2571 | ||
0fcb78c2 | 2572 | /** |
1da177e4 | 2573 | * flush_workqueue - ensure that any scheduled work has run to completion. |
0fcb78c2 | 2574 | * @wq: workqueue to flush |
1da177e4 | 2575 | * |
c5aa87bb TH |
2576 | * This function sleeps until all work items which were queued on entry |
2577 | * have finished execution, but it is not livelocked by new incoming ones. | |
1da177e4 | 2578 | */ |
7ad5b3a5 | 2579 | void flush_workqueue(struct workqueue_struct *wq) |
1da177e4 | 2580 | { |
73f53c4a TH |
2581 | struct wq_flusher this_flusher = { |
2582 | .list = LIST_HEAD_INIT(this_flusher.list), | |
2583 | .flush_color = -1, | |
2584 | .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done), | |
2585 | }; | |
2586 | int next_color; | |
1da177e4 | 2587 | |
3347fa09 TH |
2588 | if (WARN_ON(!wq_online)) |
2589 | return; | |
2590 | ||
3295f0ef IM |
2591 | lock_map_acquire(&wq->lockdep_map); |
2592 | lock_map_release(&wq->lockdep_map); | |
73f53c4a | 2593 | |
3c25a55d | 2594 | mutex_lock(&wq->mutex); |
73f53c4a TH |
2595 | |
2596 | /* | |
2597 | * Start-to-wait phase | |
2598 | */ | |
2599 | next_color = work_next_color(wq->work_color); | |
2600 | ||
2601 | if (next_color != wq->flush_color) { | |
2602 | /* | |
2603 | * Color space is not full. The current work_color | |
2604 | * becomes our flush_color and work_color is advanced | |
2605 | * by one. | |
2606 | */ | |
6183c009 | 2607 | WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); |
73f53c4a TH |
2608 | this_flusher.flush_color = wq->work_color; |
2609 | wq->work_color = next_color; | |
2610 | ||
2611 | if (!wq->first_flusher) { | |
2612 | /* no flush in progress, become the first flusher */ | |
6183c009 | 2613 | WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); |
73f53c4a TH |
2614 | |
2615 | wq->first_flusher = &this_flusher; | |
2616 | ||
112202d9 | 2617 | if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, |
73f53c4a TH |
2618 | wq->work_color)) { |
2619 | /* nothing to flush, done */ | |
2620 | wq->flush_color = next_color; | |
2621 | wq->first_flusher = NULL; | |
2622 | goto out_unlock; | |
2623 | } | |
2624 | } else { | |
2625 | /* wait in queue */ | |
6183c009 | 2626 | WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); |
73f53c4a | 2627 | list_add_tail(&this_flusher.list, &wq->flusher_queue); |
112202d9 | 2628 | flush_workqueue_prep_pwqs(wq, -1, wq->work_color); |
73f53c4a TH |
2629 | } |
2630 | } else { | |
2631 | /* | |
2632 | * Oops, color space is full, wait on overflow queue. | |
2633 | * The next flush completion will assign us | |
2634 | * flush_color and transfer to flusher_queue. | |
2635 | */ | |
2636 | list_add_tail(&this_flusher.list, &wq->flusher_overflow); | |
2637 | } | |
2638 | ||
fca839c0 TH |
2639 | check_flush_dependency(wq, NULL); |
2640 | ||
3c25a55d | 2641 | mutex_unlock(&wq->mutex); |
73f53c4a TH |
2642 | |
2643 | wait_for_completion(&this_flusher.done); | |
2644 | ||
2645 | /* | |
2646 | * Wake-up-and-cascade phase | |
2647 | * | |
2648 | * First flushers are responsible for cascading flushes and | |
2649 | * handling overflow. Non-first flushers can simply return. | |
2650 | */ | |
2651 | if (wq->first_flusher != &this_flusher) | |
2652 | return; | |
2653 | ||
3c25a55d | 2654 | mutex_lock(&wq->mutex); |
73f53c4a | 2655 | |
4ce48b37 TH |
2656 | /* we might have raced, check again with mutex held */ |
2657 | if (wq->first_flusher != &this_flusher) | |
2658 | goto out_unlock; | |
2659 | ||
73f53c4a TH |
2660 | wq->first_flusher = NULL; |
2661 | ||
6183c009 TH |
2662 | WARN_ON_ONCE(!list_empty(&this_flusher.list)); |
2663 | WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); | |
73f53c4a TH |
2664 | |
2665 | while (true) { | |
2666 | struct wq_flusher *next, *tmp; | |
2667 | ||
2668 | /* complete all the flushers sharing the current flush color */ | |
2669 | list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { | |
2670 | if (next->flush_color != wq->flush_color) | |
2671 | break; | |
2672 | list_del_init(&next->list); | |
2673 | complete(&next->done); | |
2674 | } | |
2675 | ||
6183c009 TH |
2676 | WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && |
2677 | wq->flush_color != work_next_color(wq->work_color)); | |
73f53c4a TH |
2678 | |
2679 | /* this flush_color is finished, advance by one */ | |
2680 | wq->flush_color = work_next_color(wq->flush_color); | |
2681 | ||
2682 | /* one color has been freed, handle overflow queue */ | |
2683 | if (!list_empty(&wq->flusher_overflow)) { | |
2684 | /* | |
2685 | * Assign the same color to all overflowed | |
2686 | * flushers, advance work_color and append to | |
2687 | * flusher_queue. This is the start-to-wait | |
2688 | * phase for these overflowed flushers. | |
2689 | */ | |
2690 | list_for_each_entry(tmp, &wq->flusher_overflow, list) | |
2691 | tmp->flush_color = wq->work_color; | |
2692 | ||
2693 | wq->work_color = work_next_color(wq->work_color); | |
2694 | ||
2695 | list_splice_tail_init(&wq->flusher_overflow, | |
2696 | &wq->flusher_queue); | |
112202d9 | 2697 | flush_workqueue_prep_pwqs(wq, -1, wq->work_color); |
73f53c4a TH |
2698 | } |
2699 | ||
2700 | if (list_empty(&wq->flusher_queue)) { | |
6183c009 | 2701 | WARN_ON_ONCE(wq->flush_color != wq->work_color); |
73f53c4a TH |
2702 | break; |
2703 | } | |
2704 | ||
2705 | /* | |
2706 | * Need to flush more colors. Make the next flusher | |
112202d9 | 2707 | * the new first flusher and arm pwqs. |
73f53c4a | 2708 | */ |
6183c009 TH |
2709 | WARN_ON_ONCE(wq->flush_color == wq->work_color); |
2710 | WARN_ON_ONCE(wq->flush_color != next->flush_color); | |
73f53c4a TH |
2711 | |
2712 | list_del_init(&next->list); | |
2713 | wq->first_flusher = next; | |
2714 | ||
112202d9 | 2715 | if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) |
73f53c4a TH |
2716 | break; |
2717 | ||
2718 | /* | |
2719 | * Meh... this color is already done, clear first | |
2720 | * flusher and repeat cascading. | |
2721 | */ | |
2722 | wq->first_flusher = NULL; | |
2723 | } | |
2724 | ||
2725 | out_unlock: | |
3c25a55d | 2726 | mutex_unlock(&wq->mutex); |
1da177e4 | 2727 | } |
1dadafa8 | 2728 | EXPORT_SYMBOL(flush_workqueue); |
1da177e4 | 2729 | |
9c5a2ba7 TH |
2730 | /** |
2731 | * drain_workqueue - drain a workqueue | |
2732 | * @wq: workqueue to drain | |
2733 | * | |
2734 | * Wait until the workqueue becomes empty. While draining is in progress, | |
2735 | * only chain queueing is allowed. IOW, only currently pending or running | |
2736 | * work items on @wq can queue further work items on it. @wq is flushed | |
b749b1b6 | 2737 | * repeatedly until it becomes empty. The number of flushing is determined |
9c5a2ba7 TH |
2738 | * by the depth of chaining and should be relatively short. Whine if it |
2739 | * takes too long. | |
2740 | */ | |
2741 | void drain_workqueue(struct workqueue_struct *wq) | |
2742 | { | |
2743 | unsigned int flush_cnt = 0; | |
49e3cf44 | 2744 | struct pool_workqueue *pwq; |
9c5a2ba7 TH |
2745 | |
2746 | /* | |
2747 | * __queue_work() needs to test whether there are drainers, is much | |
2748 | * hotter than drain_workqueue() and already looks at @wq->flags. | |
618b01eb | 2749 | * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. |
9c5a2ba7 | 2750 | */ |
87fc741e | 2751 | mutex_lock(&wq->mutex); |
9c5a2ba7 | 2752 | if (!wq->nr_drainers++) |
618b01eb | 2753 | wq->flags |= __WQ_DRAINING; |
87fc741e | 2754 | mutex_unlock(&wq->mutex); |
9c5a2ba7 TH |
2755 | reflush: |
2756 | flush_workqueue(wq); | |
2757 | ||
b09f4fd3 | 2758 | mutex_lock(&wq->mutex); |
76af4d93 | 2759 | |
49e3cf44 | 2760 | for_each_pwq(pwq, wq) { |
fa2563e4 | 2761 | bool drained; |
9c5a2ba7 | 2762 | |
b09f4fd3 | 2763 | spin_lock_irq(&pwq->pool->lock); |
112202d9 | 2764 | drained = !pwq->nr_active && list_empty(&pwq->delayed_works); |
b09f4fd3 | 2765 | spin_unlock_irq(&pwq->pool->lock); |
fa2563e4 TT |
2766 | |
2767 | if (drained) | |
9c5a2ba7 TH |
2768 | continue; |
2769 | ||
2770 | if (++flush_cnt == 10 || | |
2771 | (flush_cnt % 100 == 0 && flush_cnt <= 1000)) | |
c5aa87bb | 2772 | pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n", |
044c782c | 2773 | wq->name, flush_cnt); |
76af4d93 | 2774 | |
b09f4fd3 | 2775 | mutex_unlock(&wq->mutex); |
9c5a2ba7 TH |
2776 | goto reflush; |
2777 | } | |
2778 | ||
9c5a2ba7 | 2779 | if (!--wq->nr_drainers) |
618b01eb | 2780 | wq->flags &= ~__WQ_DRAINING; |
87fc741e | 2781 | mutex_unlock(&wq->mutex); |
9c5a2ba7 TH |
2782 | } |
2783 | EXPORT_SYMBOL_GPL(drain_workqueue); | |
2784 | ||
606a5020 | 2785 | static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr) |
db700897 | 2786 | { |
affee4b2 | 2787 | struct worker *worker = NULL; |
c9e7cf27 | 2788 | struct worker_pool *pool; |
112202d9 | 2789 | struct pool_workqueue *pwq; |
db700897 ON |
2790 | |
2791 | might_sleep(); | |
fa1b54e6 TH |
2792 | |
2793 | local_irq_disable(); | |
c9e7cf27 | 2794 | pool = get_work_pool(work); |
fa1b54e6 TH |
2795 | if (!pool) { |
2796 | local_irq_enable(); | |
baf59022 | 2797 | return false; |
fa1b54e6 | 2798 | } |
db700897 | 2799 | |
fa1b54e6 | 2800 | spin_lock(&pool->lock); |
0b3dae68 | 2801 | /* see the comment in try_to_grab_pending() with the same code */ |
112202d9 TH |
2802 | pwq = get_work_pwq(work); |
2803 | if (pwq) { | |
2804 | if (unlikely(pwq->pool != pool)) | |
4690c4ab | 2805 | goto already_gone; |
606a5020 | 2806 | } else { |
c9e7cf27 | 2807 | worker = find_worker_executing_work(pool, work); |
affee4b2 | 2808 | if (!worker) |
4690c4ab | 2809 | goto already_gone; |
112202d9 | 2810 | pwq = worker->current_pwq; |
606a5020 | 2811 | } |
db700897 | 2812 | |
fca839c0 TH |
2813 | check_flush_dependency(pwq->wq, work); |
2814 | ||
112202d9 | 2815 | insert_wq_barrier(pwq, barr, work, worker); |
d565ed63 | 2816 | spin_unlock_irq(&pool->lock); |
7a22ad75 | 2817 | |
e159489b TH |
2818 | /* |
2819 | * If @max_active is 1 or rescuer is in use, flushing another work | |
2820 | * item on the same workqueue may lead to deadlock. Make sure the | |
2821 | * flusher is not running on the same workqueue by verifying write | |
2822 | * access. | |
2823 | */ | |
493008a8 | 2824 | if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer) |
112202d9 | 2825 | lock_map_acquire(&pwq->wq->lockdep_map); |
e159489b | 2826 | else |
112202d9 TH |
2827 | lock_map_acquire_read(&pwq->wq->lockdep_map); |
2828 | lock_map_release(&pwq->wq->lockdep_map); | |
e159489b | 2829 | |
401a8d04 | 2830 | return true; |
4690c4ab | 2831 | already_gone: |
d565ed63 | 2832 | spin_unlock_irq(&pool->lock); |
401a8d04 | 2833 | return false; |
db700897 | 2834 | } |
baf59022 TH |
2835 | |
2836 | /** | |
2837 | * flush_work - wait for a work to finish executing the last queueing instance | |
2838 | * @work: the work to flush | |
2839 | * | |
606a5020 TH |
2840 | * Wait until @work has finished execution. @work is guaranteed to be idle |
2841 | * on return if it hasn't been requeued since flush started. | |
baf59022 | 2842 | * |
d185af30 | 2843 | * Return: |
baf59022 TH |
2844 | * %true if flush_work() waited for the work to finish execution, |
2845 | * %false if it was already idle. | |
2846 | */ | |
2847 | bool flush_work(struct work_struct *work) | |
2848 | { | |
12997d1a BH |
2849 | struct wq_barrier barr; |
2850 | ||
3347fa09 TH |
2851 | if (WARN_ON(!wq_online)) |
2852 | return false; | |
2853 | ||
0976dfc1 SB |
2854 | lock_map_acquire(&work->lockdep_map); |
2855 | lock_map_release(&work->lockdep_map); | |
2856 | ||
12997d1a BH |
2857 | if (start_flush_work(work, &barr)) { |
2858 | wait_for_completion(&barr.done); | |
2859 | destroy_work_on_stack(&barr.work); | |
2860 | return true; | |
2861 | } else { | |
2862 | return false; | |
2863 | } | |
6e84d644 | 2864 | } |
606a5020 | 2865 | EXPORT_SYMBOL_GPL(flush_work); |
6e84d644 | 2866 | |
8603e1b3 TH |
2867 | struct cwt_wait { |
2868 | wait_queue_t wait; | |
2869 | struct work_struct *work; | |
2870 | }; | |
2871 | ||
2872 | static int cwt_wakefn(wait_queue_t *wait, unsigned mode, int sync, void *key) | |
2873 | { | |
2874 | struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait); | |
2875 | ||
2876 | if (cwait->work != key) | |
2877 | return 0; | |
2878 | return autoremove_wake_function(wait, mode, sync, key); | |
2879 | } | |
2880 | ||
36e227d2 | 2881 | static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) |
1f1f642e | 2882 | { |
8603e1b3 | 2883 | static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq); |
bbb68dfa | 2884 | unsigned long flags; |
1f1f642e ON |
2885 | int ret; |
2886 | ||
2887 | do { | |
bbb68dfa TH |
2888 | ret = try_to_grab_pending(work, is_dwork, &flags); |
2889 | /* | |
8603e1b3 TH |
2890 | * If someone else is already canceling, wait for it to |
2891 | * finish. flush_work() doesn't work for PREEMPT_NONE | |
2892 | * because we may get scheduled between @work's completion | |
2893 | * and the other canceling task resuming and clearing | |
2894 | * CANCELING - flush_work() will return false immediately | |
2895 | * as @work is no longer busy, try_to_grab_pending() will | |
2896 | * return -ENOENT as @work is still being canceled and the | |
2897 | * other canceling task won't be able to clear CANCELING as | |
2898 | * we're hogging the CPU. | |
2899 | * | |
2900 | * Let's wait for completion using a waitqueue. As this | |
2901 | * may lead to the thundering herd problem, use a custom | |
2902 | * wake function which matches @work along with exclusive | |
2903 | * wait and wakeup. | |
bbb68dfa | 2904 | */ |
8603e1b3 TH |
2905 | if (unlikely(ret == -ENOENT)) { |
2906 | struct cwt_wait cwait; | |
2907 | ||
2908 | init_wait(&cwait.wait); | |
2909 | cwait.wait.func = cwt_wakefn; | |
2910 | cwait.work = work; | |
2911 | ||
2912 | prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait, | |
2913 | TASK_UNINTERRUPTIBLE); | |
2914 | if (work_is_canceling(work)) | |
2915 | schedule(); | |
2916 | finish_wait(&cancel_waitq, &cwait.wait); | |
2917 | } | |
1f1f642e ON |
2918 | } while (unlikely(ret < 0)); |
2919 | ||
bbb68dfa TH |
2920 | /* tell other tasks trying to grab @work to back off */ |
2921 | mark_work_canceling(work); | |
2922 | local_irq_restore(flags); | |
2923 | ||
3347fa09 TH |
2924 | /* |
2925 | * This allows canceling during early boot. We know that @work | |
2926 | * isn't executing. | |
2927 | */ | |
2928 | if (wq_online) | |
2929 | flush_work(work); | |
2930 | ||
7a22ad75 | 2931 | clear_work_data(work); |
8603e1b3 TH |
2932 | |
2933 | /* | |
2934 | * Paired with prepare_to_wait() above so that either | |
2935 | * waitqueue_active() is visible here or !work_is_canceling() is | |
2936 | * visible there. | |
2937 | */ | |
2938 | smp_mb(); | |
2939 | if (waitqueue_active(&cancel_waitq)) | |
2940 | __wake_up(&cancel_waitq, TASK_NORMAL, 1, work); | |
2941 | ||
1f1f642e ON |
2942 | return ret; |
2943 | } | |
2944 | ||
6e84d644 | 2945 | /** |
401a8d04 TH |
2946 | * cancel_work_sync - cancel a work and wait for it to finish |
2947 | * @work: the work to cancel | |
6e84d644 | 2948 | * |
401a8d04 TH |
2949 | * Cancel @work and wait for its execution to finish. This function |
2950 | * can be used even if the work re-queues itself or migrates to | |
2951 | * another workqueue. On return from this function, @work is | |
2952 | * guaranteed to be not pending or executing on any CPU. | |
1f1f642e | 2953 | * |
401a8d04 TH |
2954 | * cancel_work_sync(&delayed_work->work) must not be used for |
2955 | * delayed_work's. Use cancel_delayed_work_sync() instead. | |
6e84d644 | 2956 | * |
401a8d04 | 2957 | * The caller must ensure that the workqueue on which @work was last |
6e84d644 | 2958 | * queued can't be destroyed before this function returns. |
401a8d04 | 2959 | * |
d185af30 | 2960 | * Return: |
401a8d04 | 2961 | * %true if @work was pending, %false otherwise. |
6e84d644 | 2962 | */ |
401a8d04 | 2963 | bool cancel_work_sync(struct work_struct *work) |
6e84d644 | 2964 | { |
36e227d2 | 2965 | return __cancel_work_timer(work, false); |
b89deed3 | 2966 | } |
28e53bdd | 2967 | EXPORT_SYMBOL_GPL(cancel_work_sync); |
b89deed3 | 2968 | |
6e84d644 | 2969 | /** |
401a8d04 TH |
2970 | * flush_delayed_work - wait for a dwork to finish executing the last queueing |
2971 | * @dwork: the delayed work to flush | |
6e84d644 | 2972 | * |
401a8d04 TH |
2973 | * Delayed timer is cancelled and the pending work is queued for |
2974 | * immediate execution. Like flush_work(), this function only | |
2975 | * considers the last queueing instance of @dwork. | |
1f1f642e | 2976 | * |
d185af30 | 2977 | * Return: |
401a8d04 TH |
2978 | * %true if flush_work() waited for the work to finish execution, |
2979 | * %false if it was already idle. | |
6e84d644 | 2980 | */ |
401a8d04 TH |
2981 | bool flush_delayed_work(struct delayed_work *dwork) |
2982 | { | |
8930caba | 2983 | local_irq_disable(); |
401a8d04 | 2984 | if (del_timer_sync(&dwork->timer)) |
60c057bc | 2985 | __queue_work(dwork->cpu, dwork->wq, &dwork->work); |
8930caba | 2986 | local_irq_enable(); |
401a8d04 TH |
2987 | return flush_work(&dwork->work); |
2988 | } | |
2989 | EXPORT_SYMBOL(flush_delayed_work); | |
2990 | ||
f72b8792 JA |
2991 | static bool __cancel_work(struct work_struct *work, bool is_dwork) |
2992 | { | |
2993 | unsigned long flags; | |
2994 | int ret; | |
2995 | ||
2996 | do { | |
2997 | ret = try_to_grab_pending(work, is_dwork, &flags); | |
2998 | } while (unlikely(ret == -EAGAIN)); | |
2999 | ||
3000 | if (unlikely(ret < 0)) | |
3001 | return false; | |
3002 | ||
3003 | set_work_pool_and_clear_pending(work, get_work_pool_id(work)); | |
3004 | local_irq_restore(flags); | |
3005 | return ret; | |
3006 | } | |
3007 | ||
3008 | /* | |
3009 | * See cancel_delayed_work() | |
3010 | */ | |
3011 | bool cancel_work(struct work_struct *work) | |
3012 | { | |
3013 | return __cancel_work(work, false); | |
3014 | } | |
3015 | ||
09383498 | 3016 | /** |
57b30ae7 TH |
3017 | * cancel_delayed_work - cancel a delayed work |
3018 | * @dwork: delayed_work to cancel | |
09383498 | 3019 | * |
d185af30 YB |
3020 | * Kill off a pending delayed_work. |
3021 | * | |
3022 | * Return: %true if @dwork was pending and canceled; %false if it wasn't | |
3023 | * pending. | |
3024 | * | |
3025 | * Note: | |
3026 | * The work callback function may still be running on return, unless | |
3027 | * it returns %true and the work doesn't re-arm itself. Explicitly flush or | |
3028 | * use cancel_delayed_work_sync() to wait on it. | |
09383498 | 3029 | * |
57b30ae7 | 3030 | * This function is safe to call from any context including IRQ handler. |
09383498 | 3031 | */ |
57b30ae7 | 3032 | bool cancel_delayed_work(struct delayed_work *dwork) |
09383498 | 3033 | { |
f72b8792 | 3034 | return __cancel_work(&dwork->work, true); |
09383498 | 3035 | } |
57b30ae7 | 3036 | EXPORT_SYMBOL(cancel_delayed_work); |
09383498 | 3037 | |
401a8d04 TH |
3038 | /** |
3039 | * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish | |
3040 | * @dwork: the delayed work cancel | |
3041 | * | |
3042 | * This is cancel_work_sync() for delayed works. | |
3043 | * | |
d185af30 | 3044 | * Return: |
401a8d04 TH |
3045 | * %true if @dwork was pending, %false otherwise. |
3046 | */ | |
3047 | bool cancel_delayed_work_sync(struct delayed_work *dwork) | |
6e84d644 | 3048 | { |
36e227d2 | 3049 | return __cancel_work_timer(&dwork->work, true); |
6e84d644 | 3050 | } |
f5a421a4 | 3051 | EXPORT_SYMBOL(cancel_delayed_work_sync); |
1da177e4 | 3052 | |
b6136773 | 3053 | /** |
31ddd871 | 3054 | * schedule_on_each_cpu - execute a function synchronously on each online CPU |
b6136773 | 3055 | * @func: the function to call |
b6136773 | 3056 | * |
31ddd871 TH |
3057 | * schedule_on_each_cpu() executes @func on each online CPU using the |
3058 | * system workqueue and blocks until all CPUs have completed. | |
b6136773 | 3059 | * schedule_on_each_cpu() is very slow. |
31ddd871 | 3060 | * |
d185af30 | 3061 | * Return: |
31ddd871 | 3062 | * 0 on success, -errno on failure. |
b6136773 | 3063 | */ |
65f27f38 | 3064 | int schedule_on_each_cpu(work_func_t func) |
15316ba8 CL |
3065 | { |
3066 | int cpu; | |
38f51568 | 3067 | struct work_struct __percpu *works; |
15316ba8 | 3068 | |
b6136773 AM |
3069 | works = alloc_percpu(struct work_struct); |
3070 | if (!works) | |
15316ba8 | 3071 | return -ENOMEM; |
b6136773 | 3072 | |
93981800 TH |
3073 | get_online_cpus(); |
3074 | ||
15316ba8 | 3075 | for_each_online_cpu(cpu) { |
9bfb1839 IM |
3076 | struct work_struct *work = per_cpu_ptr(works, cpu); |
3077 | ||
3078 | INIT_WORK(work, func); | |
b71ab8c2 | 3079 | schedule_work_on(cpu, work); |
65a64464 | 3080 | } |
93981800 TH |
3081 | |
3082 | for_each_online_cpu(cpu) | |
3083 | flush_work(per_cpu_ptr(works, cpu)); | |
3084 | ||
95402b38 | 3085 | put_online_cpus(); |
b6136773 | 3086 | free_percpu(works); |
15316ba8 CL |
3087 | return 0; |
3088 | } | |
3089 | ||
1fa44eca JB |
3090 | /** |
3091 | * execute_in_process_context - reliably execute the routine with user context | |
3092 | * @fn: the function to execute | |
1fa44eca JB |
3093 | * @ew: guaranteed storage for the execute work structure (must |
3094 | * be available when the work executes) | |
3095 | * | |
3096 | * Executes the function immediately if process context is available, | |
3097 | * otherwise schedules the function for delayed execution. | |
3098 | * | |
d185af30 | 3099 | * Return: 0 - function was executed |
1fa44eca JB |
3100 | * 1 - function was scheduled for execution |
3101 | */ | |
65f27f38 | 3102 | int execute_in_process_context(work_func_t fn, struct execute_work *ew) |
1fa44eca JB |
3103 | { |
3104 | if (!in_interrupt()) { | |
65f27f38 | 3105 | fn(&ew->work); |
1fa44eca JB |
3106 | return 0; |
3107 | } | |
3108 | ||
65f27f38 | 3109 | INIT_WORK(&ew->work, fn); |
1fa44eca JB |
3110 | schedule_work(&ew->work); |
3111 | ||
3112 | return 1; | |
3113 | } | |
3114 | EXPORT_SYMBOL_GPL(execute_in_process_context); | |
3115 | ||
6ba94429 FW |
3116 | /** |
3117 | * free_workqueue_attrs - free a workqueue_attrs | |
3118 | * @attrs: workqueue_attrs to free | |
226223ab | 3119 | * |
6ba94429 | 3120 | * Undo alloc_workqueue_attrs(). |
226223ab | 3121 | */ |
6ba94429 | 3122 | void free_workqueue_attrs(struct workqueue_attrs *attrs) |
226223ab | 3123 | { |
6ba94429 FW |
3124 | if (attrs) { |
3125 | free_cpumask_var(attrs->cpumask); | |
3126 | kfree(attrs); | |
3127 | } | |
226223ab TH |
3128 | } |
3129 | ||
6ba94429 FW |
3130 | /** |
3131 | * alloc_workqueue_attrs - allocate a workqueue_attrs | |
3132 | * @gfp_mask: allocation mask to use | |
3133 | * | |
3134 | * Allocate a new workqueue_attrs, initialize with default settings and | |
3135 | * return it. | |
3136 | * | |
3137 | * Return: The allocated new workqueue_attr on success. %NULL on failure. | |
3138 | */ | |
3139 | struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask) | |
226223ab | 3140 | { |
6ba94429 | 3141 | struct workqueue_attrs *attrs; |
226223ab | 3142 | |
6ba94429 FW |
3143 | attrs = kzalloc(sizeof(*attrs), gfp_mask); |
3144 | if (!attrs) | |
3145 | goto fail; | |
3146 | if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask)) | |
3147 | goto fail; | |
3148 | ||
3149 | cpumask_copy(attrs->cpumask, cpu_possible_mask); | |
3150 | return attrs; | |
3151 | fail: | |
3152 | free_workqueue_attrs(attrs); | |
3153 | return NULL; | |
226223ab TH |
3154 | } |
3155 | ||
6ba94429 FW |
3156 | static void copy_workqueue_attrs(struct workqueue_attrs *to, |
3157 | const struct workqueue_attrs *from) | |
226223ab | 3158 | { |
6ba94429 FW |
3159 | to->nice = from->nice; |
3160 | cpumask_copy(to->cpumask, from->cpumask); | |
3161 | /* | |
3162 | * Unlike hash and equality test, this function doesn't ignore | |
3163 | * ->no_numa as it is used for both pool and wq attrs. Instead, | |
3164 | * get_unbound_pool() explicitly clears ->no_numa after copying. | |
3165 | */ | |
3166 | to->no_numa = from->no_numa; | |
226223ab TH |
3167 | } |
3168 | ||
6ba94429 FW |
3169 | /* hash value of the content of @attr */ |
3170 | static u32 wqattrs_hash(const struct workqueue_attrs *attrs) | |
226223ab | 3171 | { |
6ba94429 | 3172 | u32 hash = 0; |
226223ab | 3173 | |
6ba94429 FW |
3174 | hash = jhash_1word(attrs->nice, hash); |
3175 | hash = jhash(cpumask_bits(attrs->cpumask), | |
3176 | BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); | |
3177 | return hash; | |
226223ab | 3178 | } |
226223ab | 3179 | |
6ba94429 FW |
3180 | /* content equality test */ |
3181 | static bool wqattrs_equal(const struct workqueue_attrs *a, | |
3182 | const struct workqueue_attrs *b) | |
226223ab | 3183 | { |
6ba94429 FW |
3184 | if (a->nice != b->nice) |
3185 | return false; | |
3186 | if (!cpumask_equal(a->cpumask, b->cpumask)) | |
3187 | return false; | |
3188 | return true; | |
226223ab TH |
3189 | } |
3190 | ||
6ba94429 FW |
3191 | /** |
3192 | * init_worker_pool - initialize a newly zalloc'd worker_pool | |
3193 | * @pool: worker_pool to initialize | |
3194 | * | |
402dd89d | 3195 | * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs. |
6ba94429 FW |
3196 | * |
3197 | * Return: 0 on success, -errno on failure. Even on failure, all fields | |
3198 | * inside @pool proper are initialized and put_unbound_pool() can be called | |
3199 | * on @pool safely to release it. | |
3200 | */ | |
3201 | static int init_worker_pool(struct worker_pool *pool) | |
226223ab | 3202 | { |
6ba94429 FW |
3203 | spin_lock_init(&pool->lock); |
3204 | pool->id = -1; | |
3205 | pool->cpu = -1; | |
3206 | pool->node = NUMA_NO_NODE; | |
3207 | pool->flags |= POOL_DISASSOCIATED; | |
82607adc | 3208 | pool->watchdog_ts = jiffies; |
6ba94429 FW |
3209 | INIT_LIST_HEAD(&pool->worklist); |
3210 | INIT_LIST_HEAD(&pool->idle_list); | |
3211 | hash_init(pool->busy_hash); | |
226223ab | 3212 | |
6ba94429 FW |
3213 | init_timer_deferrable(&pool->idle_timer); |
3214 | pool->idle_timer.function = idle_worker_timeout; | |
3215 | pool->idle_timer.data = (unsigned long)pool; | |
226223ab | 3216 | |
6ba94429 FW |
3217 | setup_timer(&pool->mayday_timer, pool_mayday_timeout, |
3218 | (unsigned long)pool); | |
226223ab | 3219 | |
6ba94429 FW |
3220 | mutex_init(&pool->manager_arb); |
3221 | mutex_init(&pool->attach_mutex); | |
3222 | INIT_LIST_HEAD(&pool->workers); | |
226223ab | 3223 | |
6ba94429 FW |
3224 | ida_init(&pool->worker_ida); |
3225 | INIT_HLIST_NODE(&pool->hash_node); | |
3226 | pool->refcnt = 1; | |
226223ab | 3227 | |
6ba94429 FW |
3228 | /* shouldn't fail above this point */ |
3229 | pool->attrs = alloc_workqueue_attrs(GFP_KERNEL); | |
3230 | if (!pool->attrs) | |
3231 | return -ENOMEM; | |
3232 | return 0; | |
226223ab TH |
3233 | } |
3234 | ||
6ba94429 | 3235 | static void rcu_free_wq(struct rcu_head *rcu) |
226223ab | 3236 | { |
6ba94429 FW |
3237 | struct workqueue_struct *wq = |
3238 | container_of(rcu, struct workqueue_struct, rcu); | |
226223ab | 3239 | |
6ba94429 FW |
3240 | if (!(wq->flags & WQ_UNBOUND)) |
3241 | free_percpu(wq->cpu_pwqs); | |
226223ab | 3242 | else |
6ba94429 | 3243 | free_workqueue_attrs(wq->unbound_attrs); |
226223ab | 3244 | |
6ba94429 FW |
3245 | kfree(wq->rescuer); |
3246 | kfree(wq); | |
226223ab TH |
3247 | } |
3248 | ||
6ba94429 | 3249 | static void rcu_free_pool(struct rcu_head *rcu) |
226223ab | 3250 | { |
6ba94429 | 3251 | struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); |
226223ab | 3252 | |
6ba94429 FW |
3253 | ida_destroy(&pool->worker_ida); |
3254 | free_workqueue_attrs(pool->attrs); | |
3255 | kfree(pool); | |
226223ab TH |
3256 | } |
3257 | ||
6ba94429 FW |
3258 | /** |
3259 | * put_unbound_pool - put a worker_pool | |
3260 | * @pool: worker_pool to put | |
3261 | * | |
3262 | * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU | |
3263 | * safe manner. get_unbound_pool() calls this function on its failure path | |
3264 | * and this function should be able to release pools which went through, | |
3265 | * successfully or not, init_worker_pool(). | |
3266 | * | |
3267 | * Should be called with wq_pool_mutex held. | |
3268 | */ | |
3269 | static void put_unbound_pool(struct worker_pool *pool) | |
226223ab | 3270 | { |
6ba94429 FW |
3271 | DECLARE_COMPLETION_ONSTACK(detach_completion); |
3272 | struct worker *worker; | |
226223ab | 3273 | |
6ba94429 | 3274 | lockdep_assert_held(&wq_pool_mutex); |
226223ab | 3275 | |
6ba94429 FW |
3276 | if (--pool->refcnt) |
3277 | return; | |
226223ab | 3278 | |
6ba94429 FW |
3279 | /* sanity checks */ |
3280 | if (WARN_ON(!(pool->cpu < 0)) || | |
3281 | WARN_ON(!list_empty(&pool->worklist))) | |
3282 | return; | |
226223ab | 3283 | |
6ba94429 FW |
3284 | /* release id and unhash */ |
3285 | if (pool->id >= 0) | |
3286 | idr_remove(&worker_pool_idr, pool->id); | |
3287 | hash_del(&pool->hash_node); | |
d55262c4 | 3288 | |
6ba94429 FW |
3289 | /* |
3290 | * Become the manager and destroy all workers. Grabbing | |
3291 | * manager_arb prevents @pool's workers from blocking on | |
3292 | * attach_mutex. | |
3293 | */ | |
3294 | mutex_lock(&pool->manager_arb); | |
d55262c4 | 3295 | |
6ba94429 FW |
3296 | spin_lock_irq(&pool->lock); |
3297 | while ((worker = first_idle_worker(pool))) | |
3298 | destroy_worker(worker); | |
3299 | WARN_ON(pool->nr_workers || pool->nr_idle); | |
3300 | spin_unlock_irq(&pool->lock); | |
d55262c4 | 3301 | |
6ba94429 FW |
3302 | mutex_lock(&pool->attach_mutex); |
3303 | if (!list_empty(&pool->workers)) | |
3304 | pool->detach_completion = &detach_completion; | |
3305 | mutex_unlock(&pool->attach_mutex); | |
226223ab | 3306 | |
6ba94429 FW |
3307 | if (pool->detach_completion) |
3308 | wait_for_completion(pool->detach_completion); | |
226223ab | 3309 | |
6ba94429 | 3310 | mutex_unlock(&pool->manager_arb); |
226223ab | 3311 | |
6ba94429 FW |
3312 | /* shut down the timers */ |
3313 | del_timer_sync(&pool->idle_timer); | |
3314 | del_timer_sync(&pool->mayday_timer); | |
226223ab | 3315 | |
6ba94429 FW |
3316 | /* sched-RCU protected to allow dereferences from get_work_pool() */ |
3317 | call_rcu_sched(&pool->rcu, rcu_free_pool); | |
226223ab TH |
3318 | } |
3319 | ||
3320 | /** | |
6ba94429 FW |
3321 | * get_unbound_pool - get a worker_pool with the specified attributes |
3322 | * @attrs: the attributes of the worker_pool to get | |
226223ab | 3323 | * |
6ba94429 FW |
3324 | * Obtain a worker_pool which has the same attributes as @attrs, bump the |
3325 | * reference count and return it. If there already is a matching | |
3326 | * worker_pool, it will be used; otherwise, this function attempts to | |
3327 | * create a new one. | |
226223ab | 3328 | * |
6ba94429 | 3329 | * Should be called with wq_pool_mutex held. |
226223ab | 3330 | * |
6ba94429 FW |
3331 | * Return: On success, a worker_pool with the same attributes as @attrs. |
3332 | * On failure, %NULL. | |
226223ab | 3333 | */ |
6ba94429 | 3334 | static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) |
226223ab | 3335 | { |
6ba94429 FW |
3336 | u32 hash = wqattrs_hash(attrs); |
3337 | struct worker_pool *pool; | |
3338 | int node; | |
e2273584 | 3339 | int target_node = NUMA_NO_NODE; |
226223ab | 3340 | |
6ba94429 | 3341 | lockdep_assert_held(&wq_pool_mutex); |
226223ab | 3342 | |
6ba94429 FW |
3343 | /* do we already have a matching pool? */ |
3344 | hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { | |
3345 | if (wqattrs_equal(pool->attrs, attrs)) { | |
3346 | pool->refcnt++; | |
3347 | return pool; | |
3348 | } | |
3349 | } | |
226223ab | 3350 | |
e2273584 XP |
3351 | /* if cpumask is contained inside a NUMA node, we belong to that node */ |
3352 | if (wq_numa_enabled) { | |
3353 | for_each_node(node) { | |
3354 | if (cpumask_subset(attrs->cpumask, | |
3355 | wq_numa_possible_cpumask[node])) { | |
3356 | target_node = node; | |
3357 | break; | |
3358 | } | |
3359 | } | |
3360 | } | |
3361 | ||
6ba94429 | 3362 | /* nope, create a new one */ |
e2273584 | 3363 | pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node); |
6ba94429 FW |
3364 | if (!pool || init_worker_pool(pool) < 0) |
3365 | goto fail; | |
3366 | ||
3367 | lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */ | |
3368 | copy_workqueue_attrs(pool->attrs, attrs); | |
e2273584 | 3369 | pool->node = target_node; |
226223ab TH |
3370 | |
3371 | /* | |
6ba94429 FW |
3372 | * no_numa isn't a worker_pool attribute, always clear it. See |
3373 | * 'struct workqueue_attrs' comments for detail. | |
226223ab | 3374 | */ |
6ba94429 | 3375 | pool->attrs->no_numa = false; |
226223ab | 3376 | |
6ba94429 FW |
3377 | if (worker_pool_assign_id(pool) < 0) |
3378 | goto fail; | |
226223ab | 3379 | |
6ba94429 | 3380 | /* create and start the initial worker */ |
3347fa09 | 3381 | if (wq_online && !create_worker(pool)) |
6ba94429 | 3382 | goto fail; |
226223ab | 3383 | |
6ba94429 FW |
3384 | /* install */ |
3385 | hash_add(unbound_pool_hash, &pool->hash_node, hash); | |
226223ab | 3386 | |
6ba94429 FW |
3387 | return pool; |
3388 | fail: | |
3389 | if (pool) | |
3390 | put_unbound_pool(pool); | |
3391 | return NULL; | |
226223ab | 3392 | } |
226223ab | 3393 | |
6ba94429 | 3394 | static void rcu_free_pwq(struct rcu_head *rcu) |
7a4e344c | 3395 | { |
6ba94429 FW |
3396 | kmem_cache_free(pwq_cache, |
3397 | container_of(rcu, struct pool_workqueue, rcu)); | |
7a4e344c TH |
3398 | } |
3399 | ||
6ba94429 FW |
3400 | /* |
3401 | * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt | |
3402 | * and needs to be destroyed. | |
7a4e344c | 3403 | */ |
6ba94429 | 3404 | static void pwq_unbound_release_workfn(struct work_struct *work) |
7a4e344c | 3405 | { |
6ba94429 FW |
3406 | struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, |
3407 | unbound_release_work); | |
3408 | struct workqueue_struct *wq = pwq->wq; | |
3409 | struct worker_pool *pool = pwq->pool; | |
3410 | bool is_last; | |
7a4e344c | 3411 | |
6ba94429 FW |
3412 | if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND))) |
3413 | return; | |
7a4e344c | 3414 | |
6ba94429 FW |
3415 | mutex_lock(&wq->mutex); |
3416 | list_del_rcu(&pwq->pwqs_node); | |
3417 | is_last = list_empty(&wq->pwqs); | |
3418 | mutex_unlock(&wq->mutex); | |
3419 | ||
3420 | mutex_lock(&wq_pool_mutex); | |
3421 | put_unbound_pool(pool); | |
3422 | mutex_unlock(&wq_pool_mutex); | |
3423 | ||
3424 | call_rcu_sched(&pwq->rcu, rcu_free_pwq); | |
7a4e344c | 3425 | |
2865a8fb | 3426 | /* |
6ba94429 FW |
3427 | * If we're the last pwq going away, @wq is already dead and no one |
3428 | * is gonna access it anymore. Schedule RCU free. | |
2865a8fb | 3429 | */ |
6ba94429 FW |
3430 | if (is_last) |
3431 | call_rcu_sched(&wq->rcu, rcu_free_wq); | |
29c91e99 TH |
3432 | } |
3433 | ||
7a4e344c | 3434 | /** |
6ba94429 FW |
3435 | * pwq_adjust_max_active - update a pwq's max_active to the current setting |
3436 | * @pwq: target pool_workqueue | |
d185af30 | 3437 | * |
6ba94429 FW |
3438 | * If @pwq isn't freezing, set @pwq->max_active to the associated |
3439 | * workqueue's saved_max_active and activate delayed work items | |
3440 | * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. | |
7a4e344c | 3441 | */ |
6ba94429 | 3442 | static void pwq_adjust_max_active(struct pool_workqueue *pwq) |
4e1a1f9a | 3443 | { |
6ba94429 FW |
3444 | struct workqueue_struct *wq = pwq->wq; |
3445 | bool freezable = wq->flags & WQ_FREEZABLE; | |
3347fa09 | 3446 | unsigned long flags; |
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 | |
3347fa09 TH |
3455 | /* this function can be called during early boot w/ irq disabled */ |
3456 | spin_lock_irqsave(&pwq->pool->lock, flags); | |
29c91e99 | 3457 | |
6ba94429 FW |
3458 | /* |
3459 | * During [un]freezing, the caller is responsible for ensuring that | |
3460 | * this function is called at least once after @workqueue_freezing | |
3461 | * is updated and visible. | |
3462 | */ | |
3463 | if (!freezable || !workqueue_freezing) { | |
3464 | pwq->max_active = wq->saved_max_active; | |
4e1a1f9a | 3465 | |
6ba94429 FW |
3466 | while (!list_empty(&pwq->delayed_works) && |
3467 | pwq->nr_active < pwq->max_active) | |
3468 | pwq_activate_first_delayed(pwq); | |
e2dca7ad | 3469 | |
6ba94429 FW |
3470 | /* |
3471 | * Need to kick a worker after thawed or an unbound wq's | |
3472 | * max_active is bumped. It's a slow path. Do it always. | |
3473 | */ | |
3474 | wake_up_worker(pwq->pool); | |
3475 | } else { | |
3476 | pwq->max_active = 0; | |
3477 | } | |
e2dca7ad | 3478 | |
3347fa09 | 3479 | spin_unlock_irqrestore(&pwq->pool->lock, flags); |
e2dca7ad TH |
3480 | } |
3481 | ||
6ba94429 FW |
3482 | /* initialize newly alloced @pwq which is associated with @wq and @pool */ |
3483 | static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, | |
3484 | struct worker_pool *pool) | |
29c91e99 | 3485 | { |
6ba94429 | 3486 | BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); |
29c91e99 | 3487 | |
6ba94429 FW |
3488 | memset(pwq, 0, sizeof(*pwq)); |
3489 | ||
3490 | pwq->pool = pool; | |
3491 | pwq->wq = wq; | |
3492 | pwq->flush_color = -1; | |
3493 | pwq->refcnt = 1; | |
3494 | INIT_LIST_HEAD(&pwq->delayed_works); | |
3495 | INIT_LIST_HEAD(&pwq->pwqs_node); | |
3496 | INIT_LIST_HEAD(&pwq->mayday_node); | |
3497 | INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn); | |
29c91e99 TH |
3498 | } |
3499 | ||
6ba94429 FW |
3500 | /* sync @pwq with the current state of its associated wq and link it */ |
3501 | static void link_pwq(struct pool_workqueue *pwq) | |
29c91e99 | 3502 | { |
6ba94429 | 3503 | struct workqueue_struct *wq = pwq->wq; |
29c91e99 | 3504 | |
6ba94429 | 3505 | lockdep_assert_held(&wq->mutex); |
a892cacc | 3506 | |
6ba94429 FW |
3507 | /* may be called multiple times, ignore if already linked */ |
3508 | if (!list_empty(&pwq->pwqs_node)) | |
29c91e99 | 3509 | return; |
29c91e99 | 3510 | |
6ba94429 FW |
3511 | /* set the matching work_color */ |
3512 | pwq->work_color = wq->work_color; | |
29c91e99 | 3513 | |
6ba94429 FW |
3514 | /* sync max_active to the current setting */ |
3515 | pwq_adjust_max_active(pwq); | |
29c91e99 | 3516 | |
6ba94429 FW |
3517 | /* link in @pwq */ |
3518 | list_add_rcu(&pwq->pwqs_node, &wq->pwqs); | |
3519 | } | |
29c91e99 | 3520 | |
6ba94429 FW |
3521 | /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ |
3522 | static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, | |
3523 | const struct workqueue_attrs *attrs) | |
3524 | { | |
3525 | struct worker_pool *pool; | |
3526 | struct pool_workqueue *pwq; | |
60f5a4bc | 3527 | |
6ba94429 | 3528 | lockdep_assert_held(&wq_pool_mutex); |
60f5a4bc | 3529 | |
6ba94429 FW |
3530 | pool = get_unbound_pool(attrs); |
3531 | if (!pool) | |
3532 | return NULL; | |
60f5a4bc | 3533 | |
6ba94429 FW |
3534 | pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); |
3535 | if (!pwq) { | |
3536 | put_unbound_pool(pool); | |
3537 | return NULL; | |
3538 | } | |
29c91e99 | 3539 | |
6ba94429 FW |
3540 | init_pwq(pwq, wq, pool); |
3541 | return pwq; | |
3542 | } | |
29c91e99 | 3543 | |
29c91e99 | 3544 | /** |
30186c6f | 3545 | * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node |
042f7df1 | 3546 | * @attrs: the wq_attrs of the default pwq of the target workqueue |
6ba94429 FW |
3547 | * @node: the target NUMA node |
3548 | * @cpu_going_down: if >= 0, the CPU to consider as offline | |
3549 | * @cpumask: outarg, the resulting cpumask | |
29c91e99 | 3550 | * |
6ba94429 FW |
3551 | * Calculate the cpumask a workqueue with @attrs should use on @node. If |
3552 | * @cpu_going_down is >= 0, that cpu is considered offline during | |
3553 | * calculation. The result is stored in @cpumask. | |
a892cacc | 3554 | * |
6ba94429 FW |
3555 | * If NUMA affinity is not enabled, @attrs->cpumask is always used. If |
3556 | * enabled and @node has online CPUs requested by @attrs, the returned | |
3557 | * cpumask is the intersection of the possible CPUs of @node and | |
3558 | * @attrs->cpumask. | |
d185af30 | 3559 | * |
6ba94429 FW |
3560 | * The caller is responsible for ensuring that the cpumask of @node stays |
3561 | * stable. | |
3562 | * | |
3563 | * Return: %true if the resulting @cpumask is different from @attrs->cpumask, | |
3564 | * %false if equal. | |
29c91e99 | 3565 | */ |
6ba94429 FW |
3566 | static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node, |
3567 | int cpu_going_down, cpumask_t *cpumask) | |
29c91e99 | 3568 | { |
6ba94429 FW |
3569 | if (!wq_numa_enabled || attrs->no_numa) |
3570 | goto use_dfl; | |
29c91e99 | 3571 | |
6ba94429 FW |
3572 | /* does @node have any online CPUs @attrs wants? */ |
3573 | cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask); | |
3574 | if (cpu_going_down >= 0) | |
3575 | cpumask_clear_cpu(cpu_going_down, cpumask); | |
29c91e99 | 3576 | |
6ba94429 FW |
3577 | if (cpumask_empty(cpumask)) |
3578 | goto use_dfl; | |
4c16bd32 TH |
3579 | |
3580 | /* yeap, return possible CPUs in @node that @attrs wants */ | |
3581 | cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]); | |
3582 | return !cpumask_equal(cpumask, attrs->cpumask); | |
3583 | ||
3584 | use_dfl: | |
3585 | cpumask_copy(cpumask, attrs->cpumask); | |
3586 | return false; | |
3587 | } | |
3588 | ||
1befcf30 TH |
3589 | /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */ |
3590 | static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq, | |
3591 | int node, | |
3592 | struct pool_workqueue *pwq) | |
3593 | { | |
3594 | struct pool_workqueue *old_pwq; | |
3595 | ||
5b95e1af | 3596 | lockdep_assert_held(&wq_pool_mutex); |
1befcf30 TH |
3597 | lockdep_assert_held(&wq->mutex); |
3598 | ||
3599 | /* link_pwq() can handle duplicate calls */ | |
3600 | link_pwq(pwq); | |
3601 | ||
3602 | old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); | |
3603 | rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq); | |
3604 | return old_pwq; | |
3605 | } | |
3606 | ||
2d5f0764 LJ |
3607 | /* context to store the prepared attrs & pwqs before applying */ |
3608 | struct apply_wqattrs_ctx { | |
3609 | struct workqueue_struct *wq; /* target workqueue */ | |
3610 | struct workqueue_attrs *attrs; /* attrs to apply */ | |
042f7df1 | 3611 | struct list_head list; /* queued for batching commit */ |
2d5f0764 LJ |
3612 | struct pool_workqueue *dfl_pwq; |
3613 | struct pool_workqueue *pwq_tbl[]; | |
3614 | }; | |
3615 | ||
3616 | /* free the resources after success or abort */ | |
3617 | static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx) | |
3618 | { | |
3619 | if (ctx) { | |
3620 | int node; | |
3621 | ||
3622 | for_each_node(node) | |
3623 | put_pwq_unlocked(ctx->pwq_tbl[node]); | |
3624 | put_pwq_unlocked(ctx->dfl_pwq); | |
3625 | ||
3626 | free_workqueue_attrs(ctx->attrs); | |
3627 | ||
3628 | kfree(ctx); | |
3629 | } | |
3630 | } | |
3631 | ||
3632 | /* allocate the attrs and pwqs for later installation */ | |
3633 | static struct apply_wqattrs_ctx * | |
3634 | apply_wqattrs_prepare(struct workqueue_struct *wq, | |
3635 | const struct workqueue_attrs *attrs) | |
9e8cd2f5 | 3636 | { |
2d5f0764 | 3637 | struct apply_wqattrs_ctx *ctx; |
4c16bd32 | 3638 | struct workqueue_attrs *new_attrs, *tmp_attrs; |
2d5f0764 | 3639 | int node; |
9e8cd2f5 | 3640 | |
2d5f0764 | 3641 | lockdep_assert_held(&wq_pool_mutex); |
9e8cd2f5 | 3642 | |
2d5f0764 LJ |
3643 | ctx = kzalloc(sizeof(*ctx) + nr_node_ids * sizeof(ctx->pwq_tbl[0]), |
3644 | GFP_KERNEL); | |
8719dcea | 3645 | |
13e2e556 | 3646 | new_attrs = alloc_workqueue_attrs(GFP_KERNEL); |
4c16bd32 | 3647 | tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL); |
2d5f0764 LJ |
3648 | if (!ctx || !new_attrs || !tmp_attrs) |
3649 | goto out_free; | |
13e2e556 | 3650 | |
042f7df1 LJ |
3651 | /* |
3652 | * Calculate the attrs of the default pwq. | |
3653 | * If the user configured cpumask doesn't overlap with the | |
3654 | * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask. | |
3655 | */ | |
13e2e556 | 3656 | copy_workqueue_attrs(new_attrs, attrs); |
b05a7928 | 3657 | cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask); |
042f7df1 LJ |
3658 | if (unlikely(cpumask_empty(new_attrs->cpumask))) |
3659 | cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask); | |
13e2e556 | 3660 | |
4c16bd32 TH |
3661 | /* |
3662 | * We may create multiple pwqs with differing cpumasks. Make a | |
3663 | * copy of @new_attrs which will be modified and used to obtain | |
3664 | * pools. | |
3665 | */ | |
3666 | copy_workqueue_attrs(tmp_attrs, new_attrs); | |
3667 | ||
4c16bd32 TH |
3668 | /* |
3669 | * If something goes wrong during CPU up/down, we'll fall back to | |
3670 | * the default pwq covering whole @attrs->cpumask. Always create | |
3671 | * it even if we don't use it immediately. | |
3672 | */ | |
2d5f0764 LJ |
3673 | ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs); |
3674 | if (!ctx->dfl_pwq) | |
3675 | goto out_free; | |
4c16bd32 TH |
3676 | |
3677 | for_each_node(node) { | |
042f7df1 | 3678 | if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) { |
2d5f0764 LJ |
3679 | ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs); |
3680 | if (!ctx->pwq_tbl[node]) | |
3681 | goto out_free; | |
4c16bd32 | 3682 | } else { |
2d5f0764 LJ |
3683 | ctx->dfl_pwq->refcnt++; |
3684 | ctx->pwq_tbl[node] = ctx->dfl_pwq; | |
4c16bd32 TH |
3685 | } |
3686 | } | |
3687 | ||
042f7df1 LJ |
3688 | /* save the user configured attrs and sanitize it. */ |
3689 | copy_workqueue_attrs(new_attrs, attrs); | |
3690 | cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); | |
2d5f0764 | 3691 | ctx->attrs = new_attrs; |
042f7df1 | 3692 | |
2d5f0764 LJ |
3693 | ctx->wq = wq; |
3694 | free_workqueue_attrs(tmp_attrs); | |
3695 | return ctx; | |
3696 | ||
3697 | out_free: | |
3698 | free_workqueue_attrs(tmp_attrs); | |
3699 | free_workqueue_attrs(new_attrs); | |
3700 | apply_wqattrs_cleanup(ctx); | |
3701 | return NULL; | |
3702 | } | |
3703 | ||
3704 | /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */ | |
3705 | static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx) | |
3706 | { | |
3707 | int node; | |
9e8cd2f5 | 3708 | |
4c16bd32 | 3709 | /* all pwqs have been created successfully, let's install'em */ |
2d5f0764 | 3710 | mutex_lock(&ctx->wq->mutex); |
a892cacc | 3711 | |
2d5f0764 | 3712 | copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs); |
4c16bd32 TH |
3713 | |
3714 | /* save the previous pwq and install the new one */ | |
f147f29e | 3715 | for_each_node(node) |
2d5f0764 LJ |
3716 | ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node, |
3717 | ctx->pwq_tbl[node]); | |
4c16bd32 TH |
3718 | |
3719 | /* @dfl_pwq might not have been used, ensure it's linked */ | |
2d5f0764 LJ |
3720 | link_pwq(ctx->dfl_pwq); |
3721 | swap(ctx->wq->dfl_pwq, ctx->dfl_pwq); | |
f147f29e | 3722 | |
2d5f0764 LJ |
3723 | mutex_unlock(&ctx->wq->mutex); |
3724 | } | |
9e8cd2f5 | 3725 | |
a0111cf6 LJ |
3726 | static void apply_wqattrs_lock(void) |
3727 | { | |
3728 | /* CPUs should stay stable across pwq creations and installations */ | |
3729 | get_online_cpus(); | |
3730 | mutex_lock(&wq_pool_mutex); | |
3731 | } | |
3732 | ||
3733 | static void apply_wqattrs_unlock(void) | |
3734 | { | |
3735 | mutex_unlock(&wq_pool_mutex); | |
3736 | put_online_cpus(); | |
3737 | } | |
3738 | ||
3739 | static int apply_workqueue_attrs_locked(struct workqueue_struct *wq, | |
3740 | const struct workqueue_attrs *attrs) | |
2d5f0764 LJ |
3741 | { |
3742 | struct apply_wqattrs_ctx *ctx; | |
4c16bd32 | 3743 | |
2d5f0764 LJ |
3744 | /* only unbound workqueues can change attributes */ |
3745 | if (WARN_ON(!(wq->flags & WQ_UNBOUND))) | |
3746 | return -EINVAL; | |
13e2e556 | 3747 | |
2d5f0764 LJ |
3748 | /* creating multiple pwqs breaks ordering guarantee */ |
3749 | if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs))) | |
3750 | return -EINVAL; | |
3751 | ||
2d5f0764 | 3752 | ctx = apply_wqattrs_prepare(wq, attrs); |
6201171e | 3753 | if (!ctx) |
3754 | return -ENOMEM; | |
2d5f0764 LJ |
3755 | |
3756 | /* the ctx has been prepared successfully, let's commit it */ | |
6201171e | 3757 | apply_wqattrs_commit(ctx); |
2d5f0764 LJ |
3758 | apply_wqattrs_cleanup(ctx); |
3759 | ||
6201171e | 3760 | return 0; |
9e8cd2f5 TH |
3761 | } |
3762 | ||
a0111cf6 LJ |
3763 | /** |
3764 | * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue | |
3765 | * @wq: the target workqueue | |
3766 | * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() | |
3767 | * | |
3768 | * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA | |
3769 | * machines, this function maps a separate pwq to each NUMA node with | |
3770 | * possibles CPUs in @attrs->cpumask so that work items are affine to the | |
3771 | * NUMA node it was issued on. Older pwqs are released as in-flight work | |
3772 | * items finish. Note that a work item which repeatedly requeues itself | |
3773 | * back-to-back will stay on its current pwq. | |
3774 | * | |
3775 | * Performs GFP_KERNEL allocations. | |
3776 | * | |
3777 | * Return: 0 on success and -errno on failure. | |
3778 | */ | |
3779 | int apply_workqueue_attrs(struct workqueue_struct *wq, | |
3780 | const struct workqueue_attrs *attrs) | |
3781 | { | |
3782 | int ret; | |
3783 | ||
3784 | apply_wqattrs_lock(); | |
3785 | ret = apply_workqueue_attrs_locked(wq, attrs); | |
3786 | apply_wqattrs_unlock(); | |
3787 | ||
3788 | return ret; | |
3789 | } | |
3790 | ||
4c16bd32 TH |
3791 | /** |
3792 | * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug | |
3793 | * @wq: the target workqueue | |
3794 | * @cpu: the CPU coming up or going down | |
3795 | * @online: whether @cpu is coming up or going down | |
3796 | * | |
3797 | * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and | |
3798 | * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of | |
3799 | * @wq accordingly. | |
3800 | * | |
3801 | * If NUMA affinity can't be adjusted due to memory allocation failure, it | |
3802 | * falls back to @wq->dfl_pwq which may not be optimal but is always | |
3803 | * correct. | |
3804 | * | |
3805 | * Note that when the last allowed CPU of a NUMA node goes offline for a | |
3806 | * workqueue with a cpumask spanning multiple nodes, the workers which were | |
3807 | * already executing the work items for the workqueue will lose their CPU | |
3808 | * affinity and may execute on any CPU. This is similar to how per-cpu | |
3809 | * workqueues behave on CPU_DOWN. If a workqueue user wants strict | |
3810 | * affinity, it's the user's responsibility to flush the work item from | |
3811 | * CPU_DOWN_PREPARE. | |
3812 | */ | |
3813 | static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu, | |
3814 | bool online) | |
3815 | { | |
3816 | int node = cpu_to_node(cpu); | |
3817 | int cpu_off = online ? -1 : cpu; | |
3818 | struct pool_workqueue *old_pwq = NULL, *pwq; | |
3819 | struct workqueue_attrs *target_attrs; | |
3820 | cpumask_t *cpumask; | |
3821 | ||
3822 | lockdep_assert_held(&wq_pool_mutex); | |
3823 | ||
f7142ed4 LJ |
3824 | if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) || |
3825 | wq->unbound_attrs->no_numa) | |
4c16bd32 TH |
3826 | return; |
3827 | ||
3828 | /* | |
3829 | * We don't wanna alloc/free wq_attrs for each wq for each CPU. | |
3830 | * Let's use a preallocated one. The following buf is protected by | |
3831 | * CPU hotplug exclusion. | |
3832 | */ | |
3833 | target_attrs = wq_update_unbound_numa_attrs_buf; | |
3834 | cpumask = target_attrs->cpumask; | |
3835 | ||
4c16bd32 TH |
3836 | copy_workqueue_attrs(target_attrs, wq->unbound_attrs); |
3837 | pwq = unbound_pwq_by_node(wq, node); | |
3838 | ||
3839 | /* | |
3840 | * Let's determine what needs to be done. If the target cpumask is | |
042f7df1 LJ |
3841 | * different from the default pwq's, we need to compare it to @pwq's |
3842 | * and create a new one if they don't match. If the target cpumask | |
3843 | * equals the default pwq's, the default pwq should be used. | |
4c16bd32 | 3844 | */ |
042f7df1 | 3845 | if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) { |
4c16bd32 | 3846 | if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask)) |
f7142ed4 | 3847 | return; |
4c16bd32 | 3848 | } else { |
534a3fbb | 3849 | goto use_dfl_pwq; |
4c16bd32 TH |
3850 | } |
3851 | ||
4c16bd32 TH |
3852 | /* create a new pwq */ |
3853 | pwq = alloc_unbound_pwq(wq, target_attrs); | |
3854 | if (!pwq) { | |
2d916033 FF |
3855 | pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n", |
3856 | wq->name); | |
77f300b1 | 3857 | goto use_dfl_pwq; |
4c16bd32 TH |
3858 | } |
3859 | ||
f7142ed4 | 3860 | /* Install the new pwq. */ |
4c16bd32 TH |
3861 | mutex_lock(&wq->mutex); |
3862 | old_pwq = numa_pwq_tbl_install(wq, node, pwq); | |
3863 | goto out_unlock; | |
3864 | ||
3865 | use_dfl_pwq: | |
f7142ed4 | 3866 | mutex_lock(&wq->mutex); |
4c16bd32 TH |
3867 | spin_lock_irq(&wq->dfl_pwq->pool->lock); |
3868 | get_pwq(wq->dfl_pwq); | |
3869 | spin_unlock_irq(&wq->dfl_pwq->pool->lock); | |
3870 | old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq); | |
3871 | out_unlock: | |
3872 | mutex_unlock(&wq->mutex); | |
3873 | put_pwq_unlocked(old_pwq); | |
3874 | } | |
3875 | ||
30cdf249 | 3876 | static int alloc_and_link_pwqs(struct workqueue_struct *wq) |
0f900049 | 3877 | { |
49e3cf44 | 3878 | bool highpri = wq->flags & WQ_HIGHPRI; |
8a2b7538 | 3879 | int cpu, ret; |
30cdf249 TH |
3880 | |
3881 | if (!(wq->flags & WQ_UNBOUND)) { | |
420c0ddb TH |
3882 | wq->cpu_pwqs = alloc_percpu(struct pool_workqueue); |
3883 | if (!wq->cpu_pwqs) | |
30cdf249 TH |
3884 | return -ENOMEM; |
3885 | ||
3886 | for_each_possible_cpu(cpu) { | |
7fb98ea7 TH |
3887 | struct pool_workqueue *pwq = |
3888 | per_cpu_ptr(wq->cpu_pwqs, cpu); | |
7a62c2c8 | 3889 | struct worker_pool *cpu_pools = |
f02ae73a | 3890 | per_cpu(cpu_worker_pools, cpu); |
f3421797 | 3891 | |
f147f29e TH |
3892 | init_pwq(pwq, wq, &cpu_pools[highpri]); |
3893 | ||
3894 | mutex_lock(&wq->mutex); | |
1befcf30 | 3895 | link_pwq(pwq); |
f147f29e | 3896 | mutex_unlock(&wq->mutex); |
30cdf249 | 3897 | } |
9e8cd2f5 | 3898 | return 0; |
8a2b7538 TH |
3899 | } else if (wq->flags & __WQ_ORDERED) { |
3900 | ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]); | |
3901 | /* there should only be single pwq for ordering guarantee */ | |
3902 | WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node || | |
3903 | wq->pwqs.prev != &wq->dfl_pwq->pwqs_node), | |
3904 | "ordering guarantee broken for workqueue %s\n", wq->name); | |
3905 | return ret; | |
30cdf249 | 3906 | } else { |
9e8cd2f5 | 3907 | return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); |
30cdf249 | 3908 | } |
0f900049 TH |
3909 | } |
3910 | ||
f3421797 TH |
3911 | static int wq_clamp_max_active(int max_active, unsigned int flags, |
3912 | const char *name) | |
b71ab8c2 | 3913 | { |
f3421797 TH |
3914 | int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; |
3915 | ||
3916 | if (max_active < 1 || max_active > lim) | |
044c782c VI |
3917 | pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", |
3918 | max_active, name, 1, lim); | |
b71ab8c2 | 3919 | |
f3421797 | 3920 | return clamp_val(max_active, 1, lim); |
b71ab8c2 TH |
3921 | } |
3922 | ||
b196be89 | 3923 | struct workqueue_struct *__alloc_workqueue_key(const char *fmt, |
d320c038 TH |
3924 | unsigned int flags, |
3925 | int max_active, | |
3926 | struct lock_class_key *key, | |
b196be89 | 3927 | const char *lock_name, ...) |
1da177e4 | 3928 | { |
df2d5ae4 | 3929 | size_t tbl_size = 0; |
ecf6881f | 3930 | va_list args; |
1da177e4 | 3931 | struct workqueue_struct *wq; |
49e3cf44 | 3932 | struct pool_workqueue *pwq; |
b196be89 | 3933 | |
cee22a15 VK |
3934 | /* see the comment above the definition of WQ_POWER_EFFICIENT */ |
3935 | if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) | |
3936 | flags |= WQ_UNBOUND; | |
3937 | ||
ecf6881f | 3938 | /* allocate wq and format name */ |
df2d5ae4 | 3939 | if (flags & WQ_UNBOUND) |
ddcb57e2 | 3940 | tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]); |
df2d5ae4 TH |
3941 | |
3942 | wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL); | |
b196be89 | 3943 | if (!wq) |
d2c1d404 | 3944 | return NULL; |
b196be89 | 3945 | |
6029a918 TH |
3946 | if (flags & WQ_UNBOUND) { |
3947 | wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL); | |
3948 | if (!wq->unbound_attrs) | |
3949 | goto err_free_wq; | |
3950 | } | |
3951 | ||
ecf6881f TH |
3952 | va_start(args, lock_name); |
3953 | vsnprintf(wq->name, sizeof(wq->name), fmt, args); | |
b196be89 | 3954 | va_end(args); |
1da177e4 | 3955 | |
d320c038 | 3956 | max_active = max_active ?: WQ_DFL_ACTIVE; |
b196be89 | 3957 | max_active = wq_clamp_max_active(max_active, flags, wq->name); |
3af24433 | 3958 | |
b196be89 | 3959 | /* init wq */ |
97e37d7b | 3960 | wq->flags = flags; |
a0a1a5fd | 3961 | wq->saved_max_active = max_active; |
3c25a55d | 3962 | mutex_init(&wq->mutex); |
112202d9 | 3963 | atomic_set(&wq->nr_pwqs_to_flush, 0); |
30cdf249 | 3964 | INIT_LIST_HEAD(&wq->pwqs); |
73f53c4a TH |
3965 | INIT_LIST_HEAD(&wq->flusher_queue); |
3966 | INIT_LIST_HEAD(&wq->flusher_overflow); | |
493a1724 | 3967 | INIT_LIST_HEAD(&wq->maydays); |
502ca9d8 | 3968 | |
eb13ba87 | 3969 | lockdep_init_map(&wq->lockdep_map, lock_name, key, 0); |
cce1a165 | 3970 | INIT_LIST_HEAD(&wq->list); |
3af24433 | 3971 | |
30cdf249 | 3972 | if (alloc_and_link_pwqs(wq) < 0) |
d2c1d404 | 3973 | goto err_free_wq; |
1537663f | 3974 | |
493008a8 TH |
3975 | /* |
3976 | * Workqueues which may be used during memory reclaim should | |
3977 | * have a rescuer to guarantee forward progress. | |
3978 | */ | |
3979 | if (flags & WQ_MEM_RECLAIM) { | |
e22bee78 TH |
3980 | struct worker *rescuer; |
3981 | ||
f7537df5 | 3982 | rescuer = alloc_worker(NUMA_NO_NODE); |
e22bee78 | 3983 | if (!rescuer) |
d2c1d404 | 3984 | goto err_destroy; |
e22bee78 | 3985 | |
111c225a TH |
3986 | rescuer->rescue_wq = wq; |
3987 | rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", | |
b196be89 | 3988 | wq->name); |
d2c1d404 TH |
3989 | if (IS_ERR(rescuer->task)) { |
3990 | kfree(rescuer); | |
3991 | goto err_destroy; | |
3992 | } | |
e22bee78 | 3993 | |
d2c1d404 | 3994 | wq->rescuer = rescuer; |
25834c73 | 3995 | kthread_bind_mask(rescuer->task, cpu_possible_mask); |
e22bee78 | 3996 | wake_up_process(rescuer->task); |
3af24433 ON |
3997 | } |
3998 | ||
226223ab TH |
3999 | if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) |
4000 | goto err_destroy; | |
4001 | ||
a0a1a5fd | 4002 | /* |
68e13a67 LJ |
4003 | * wq_pool_mutex protects global freeze state and workqueues list. |
4004 | * Grab it, adjust max_active and add the new @wq to workqueues | |
4005 | * list. | |
a0a1a5fd | 4006 | */ |
68e13a67 | 4007 | mutex_lock(&wq_pool_mutex); |
a0a1a5fd | 4008 | |
a357fc03 | 4009 | mutex_lock(&wq->mutex); |
699ce097 TH |
4010 | for_each_pwq(pwq, wq) |
4011 | pwq_adjust_max_active(pwq); | |
a357fc03 | 4012 | mutex_unlock(&wq->mutex); |
a0a1a5fd | 4013 | |
e2dca7ad | 4014 | list_add_tail_rcu(&wq->list, &workqueues); |
a0a1a5fd | 4015 | |
68e13a67 | 4016 | mutex_unlock(&wq_pool_mutex); |
1537663f | 4017 | |
3af24433 | 4018 | return wq; |
d2c1d404 TH |
4019 | |
4020 | err_free_wq: | |
6029a918 | 4021 | free_workqueue_attrs(wq->unbound_attrs); |
d2c1d404 TH |
4022 | kfree(wq); |
4023 | return NULL; | |
4024 | err_destroy: | |
4025 | destroy_workqueue(wq); | |
4690c4ab | 4026 | return NULL; |
3af24433 | 4027 | } |
d320c038 | 4028 | EXPORT_SYMBOL_GPL(__alloc_workqueue_key); |
1da177e4 | 4029 | |
3af24433 ON |
4030 | /** |
4031 | * destroy_workqueue - safely terminate a workqueue | |
4032 | * @wq: target workqueue | |
4033 | * | |
4034 | * Safely destroy a workqueue. All work currently pending will be done first. | |
4035 | */ | |
4036 | void destroy_workqueue(struct workqueue_struct *wq) | |
4037 | { | |
49e3cf44 | 4038 | struct pool_workqueue *pwq; |
4c16bd32 | 4039 | int node; |
3af24433 | 4040 | |
9c5a2ba7 TH |
4041 | /* drain it before proceeding with destruction */ |
4042 | drain_workqueue(wq); | |
c8efcc25 | 4043 | |
6183c009 | 4044 | /* sanity checks */ |
b09f4fd3 | 4045 | mutex_lock(&wq->mutex); |
49e3cf44 | 4046 | for_each_pwq(pwq, wq) { |
6183c009 TH |
4047 | int i; |
4048 | ||
76af4d93 TH |
4049 | for (i = 0; i < WORK_NR_COLORS; i++) { |
4050 | if (WARN_ON(pwq->nr_in_flight[i])) { | |
b09f4fd3 | 4051 | mutex_unlock(&wq->mutex); |
fa07fb6a | 4052 | show_workqueue_state(); |
6183c009 | 4053 | return; |
76af4d93 TH |
4054 | } |
4055 | } | |
4056 | ||
5c529597 | 4057 | if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) || |
8864b4e5 | 4058 | WARN_ON(pwq->nr_active) || |
76af4d93 | 4059 | WARN_ON(!list_empty(&pwq->delayed_works))) { |
b09f4fd3 | 4060 | mutex_unlock(&wq->mutex); |
fa07fb6a | 4061 | show_workqueue_state(); |
6183c009 | 4062 | return; |
76af4d93 | 4063 | } |
6183c009 | 4064 | } |
b09f4fd3 | 4065 | mutex_unlock(&wq->mutex); |
6183c009 | 4066 | |
a0a1a5fd TH |
4067 | /* |
4068 | * wq list is used to freeze wq, remove from list after | |
4069 | * flushing is complete in case freeze races us. | |
4070 | */ | |
68e13a67 | 4071 | mutex_lock(&wq_pool_mutex); |
e2dca7ad | 4072 | list_del_rcu(&wq->list); |
68e13a67 | 4073 | mutex_unlock(&wq_pool_mutex); |
3af24433 | 4074 | |
226223ab TH |
4075 | workqueue_sysfs_unregister(wq); |
4076 | ||
e2dca7ad | 4077 | if (wq->rescuer) |
e22bee78 | 4078 | kthread_stop(wq->rescuer->task); |
e22bee78 | 4079 | |
8864b4e5 TH |
4080 | if (!(wq->flags & WQ_UNBOUND)) { |
4081 | /* | |
4082 | * The base ref is never dropped on per-cpu pwqs. Directly | |
e2dca7ad | 4083 | * schedule RCU free. |
8864b4e5 | 4084 | */ |
e2dca7ad | 4085 | call_rcu_sched(&wq->rcu, rcu_free_wq); |
8864b4e5 TH |
4086 | } else { |
4087 | /* | |
4088 | * We're the sole accessor of @wq at this point. Directly | |
4c16bd32 TH |
4089 | * access numa_pwq_tbl[] and dfl_pwq to put the base refs. |
4090 | * @wq will be freed when the last pwq is released. | |
8864b4e5 | 4091 | */ |
4c16bd32 TH |
4092 | for_each_node(node) { |
4093 | pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); | |
4094 | RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL); | |
4095 | put_pwq_unlocked(pwq); | |
4096 | } | |
4097 | ||
4098 | /* | |
4099 | * Put dfl_pwq. @wq may be freed any time after dfl_pwq is | |
4100 | * put. Don't access it afterwards. | |
4101 | */ | |
4102 | pwq = wq->dfl_pwq; | |
4103 | wq->dfl_pwq = NULL; | |
dce90d47 | 4104 | put_pwq_unlocked(pwq); |
29c91e99 | 4105 | } |
3af24433 ON |
4106 | } |
4107 | EXPORT_SYMBOL_GPL(destroy_workqueue); | |
4108 | ||
dcd989cb TH |
4109 | /** |
4110 | * workqueue_set_max_active - adjust max_active of a workqueue | |
4111 | * @wq: target workqueue | |
4112 | * @max_active: new max_active value. | |
4113 | * | |
4114 | * Set max_active of @wq to @max_active. | |
4115 | * | |
4116 | * CONTEXT: | |
4117 | * Don't call from IRQ context. | |
4118 | */ | |
4119 | void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) | |
4120 | { | |
49e3cf44 | 4121 | struct pool_workqueue *pwq; |
dcd989cb | 4122 | |
8719dcea TH |
4123 | /* disallow meddling with max_active for ordered workqueues */ |
4124 | if (WARN_ON(wq->flags & __WQ_ORDERED)) | |
4125 | return; | |
4126 | ||
f3421797 | 4127 | max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); |
dcd989cb | 4128 | |
a357fc03 | 4129 | mutex_lock(&wq->mutex); |
dcd989cb TH |
4130 | |
4131 | wq->saved_max_active = max_active; | |
4132 | ||
699ce097 TH |
4133 | for_each_pwq(pwq, wq) |
4134 | pwq_adjust_max_active(pwq); | |
93981800 | 4135 | |
a357fc03 | 4136 | mutex_unlock(&wq->mutex); |
15316ba8 | 4137 | } |
dcd989cb | 4138 | EXPORT_SYMBOL_GPL(workqueue_set_max_active); |
15316ba8 | 4139 | |
e6267616 TH |
4140 | /** |
4141 | * current_is_workqueue_rescuer - is %current workqueue rescuer? | |
4142 | * | |
4143 | * Determine whether %current is a workqueue rescuer. Can be used from | |
4144 | * work functions to determine whether it's being run off the rescuer task. | |
d185af30 YB |
4145 | * |
4146 | * Return: %true if %current is a workqueue rescuer. %false otherwise. | |
e6267616 TH |
4147 | */ |
4148 | bool current_is_workqueue_rescuer(void) | |
4149 | { | |
4150 | struct worker *worker = current_wq_worker(); | |
4151 | ||
6a092dfd | 4152 | return worker && worker->rescue_wq; |
e6267616 TH |
4153 | } |
4154 | ||
eef6a7d5 | 4155 | /** |
dcd989cb TH |
4156 | * workqueue_congested - test whether a workqueue is congested |
4157 | * @cpu: CPU in question | |
4158 | * @wq: target workqueue | |
eef6a7d5 | 4159 | * |
dcd989cb TH |
4160 | * Test whether @wq's cpu workqueue for @cpu is congested. There is |
4161 | * no synchronization around this function and the test result is | |
4162 | * unreliable and only useful as advisory hints or for debugging. | |
eef6a7d5 | 4163 | * |
d3251859 TH |
4164 | * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. |
4165 | * Note that both per-cpu and unbound workqueues may be associated with | |
4166 | * multiple pool_workqueues which have separate congested states. A | |
4167 | * workqueue being congested on one CPU doesn't mean the workqueue is also | |
4168 | * contested on other CPUs / NUMA nodes. | |
4169 | * | |
d185af30 | 4170 | * Return: |
dcd989cb | 4171 | * %true if congested, %false otherwise. |
eef6a7d5 | 4172 | */ |
d84ff051 | 4173 | bool workqueue_congested(int cpu, struct workqueue_struct *wq) |
1da177e4 | 4174 | { |
7fb98ea7 | 4175 | struct pool_workqueue *pwq; |
76af4d93 TH |
4176 | bool ret; |
4177 | ||
88109453 | 4178 | rcu_read_lock_sched(); |
7fb98ea7 | 4179 | |
d3251859 TH |
4180 | if (cpu == WORK_CPU_UNBOUND) |
4181 | cpu = smp_processor_id(); | |
4182 | ||
7fb98ea7 TH |
4183 | if (!(wq->flags & WQ_UNBOUND)) |
4184 | pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); | |
4185 | else | |
df2d5ae4 | 4186 | pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); |
dcd989cb | 4187 | |
76af4d93 | 4188 | ret = !list_empty(&pwq->delayed_works); |
88109453 | 4189 | rcu_read_unlock_sched(); |
76af4d93 TH |
4190 | |
4191 | return ret; | |
1da177e4 | 4192 | } |
dcd989cb | 4193 | EXPORT_SYMBOL_GPL(workqueue_congested); |
1da177e4 | 4194 | |
dcd989cb TH |
4195 | /** |
4196 | * work_busy - test whether a work is currently pending or running | |
4197 | * @work: the work to be tested | |
4198 | * | |
4199 | * Test whether @work is currently pending or running. There is no | |
4200 | * synchronization around this function and the test result is | |
4201 | * unreliable and only useful as advisory hints or for debugging. | |
dcd989cb | 4202 | * |
d185af30 | 4203 | * Return: |
dcd989cb TH |
4204 | * OR'd bitmask of WORK_BUSY_* bits. |
4205 | */ | |
4206 | unsigned int work_busy(struct work_struct *work) | |
1da177e4 | 4207 | { |
fa1b54e6 | 4208 | struct worker_pool *pool; |
dcd989cb TH |
4209 | unsigned long flags; |
4210 | unsigned int ret = 0; | |
1da177e4 | 4211 | |
dcd989cb TH |
4212 | if (work_pending(work)) |
4213 | ret |= WORK_BUSY_PENDING; | |
1da177e4 | 4214 | |
fa1b54e6 TH |
4215 | local_irq_save(flags); |
4216 | pool = get_work_pool(work); | |
038366c5 | 4217 | if (pool) { |
fa1b54e6 | 4218 | spin_lock(&pool->lock); |
038366c5 LJ |
4219 | if (find_worker_executing_work(pool, work)) |
4220 | ret |= WORK_BUSY_RUNNING; | |
fa1b54e6 | 4221 | spin_unlock(&pool->lock); |
038366c5 | 4222 | } |
fa1b54e6 | 4223 | local_irq_restore(flags); |
1da177e4 | 4224 | |
dcd989cb | 4225 | return ret; |
1da177e4 | 4226 | } |
dcd989cb | 4227 | EXPORT_SYMBOL_GPL(work_busy); |
1da177e4 | 4228 | |
3d1cb205 TH |
4229 | /** |
4230 | * set_worker_desc - set description for the current work item | |
4231 | * @fmt: printf-style format string | |
4232 | * @...: arguments for the format string | |
4233 | * | |
4234 | * This function can be called by a running work function to describe what | |
4235 | * the work item is about. If the worker task gets dumped, this | |
4236 | * information will be printed out together to help debugging. The | |
4237 | * description can be at most WORKER_DESC_LEN including the trailing '\0'. | |
4238 | */ | |
4239 | void set_worker_desc(const char *fmt, ...) | |
4240 | { | |
4241 | struct worker *worker = current_wq_worker(); | |
4242 | va_list args; | |
4243 | ||
4244 | if (worker) { | |
4245 | va_start(args, fmt); | |
4246 | vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); | |
4247 | va_end(args); | |
4248 | worker->desc_valid = true; | |
4249 | } | |
4250 | } | |
4251 | ||
4252 | /** | |
4253 | * print_worker_info - print out worker information and description | |
4254 | * @log_lvl: the log level to use when printing | |
4255 | * @task: target task | |
4256 | * | |
4257 | * If @task is a worker and currently executing a work item, print out the | |
4258 | * name of the workqueue being serviced and worker description set with | |
4259 | * set_worker_desc() by the currently executing work item. | |
4260 | * | |
4261 | * This function can be safely called on any task as long as the | |
4262 | * task_struct itself is accessible. While safe, this function isn't | |
4263 | * synchronized and may print out mixups or garbages of limited length. | |
4264 | */ | |
4265 | void print_worker_info(const char *log_lvl, struct task_struct *task) | |
4266 | { | |
4267 | work_func_t *fn = NULL; | |
4268 | char name[WQ_NAME_LEN] = { }; | |
4269 | char desc[WORKER_DESC_LEN] = { }; | |
4270 | struct pool_workqueue *pwq = NULL; | |
4271 | struct workqueue_struct *wq = NULL; | |
4272 | bool desc_valid = false; | |
4273 | struct worker *worker; | |
4274 | ||
4275 | if (!(task->flags & PF_WQ_WORKER)) | |
4276 | return; | |
4277 | ||
4278 | /* | |
4279 | * This function is called without any synchronization and @task | |
4280 | * could be in any state. Be careful with dereferences. | |
4281 | */ | |
e700591a | 4282 | worker = kthread_probe_data(task); |
3d1cb205 TH |
4283 | |
4284 | /* | |
4285 | * Carefully copy the associated workqueue's workfn and name. Keep | |
4286 | * the original last '\0' in case the original contains garbage. | |
4287 | */ | |
4288 | probe_kernel_read(&fn, &worker->current_func, sizeof(fn)); | |
4289 | probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq)); | |
4290 | probe_kernel_read(&wq, &pwq->wq, sizeof(wq)); | |
4291 | probe_kernel_read(name, wq->name, sizeof(name) - 1); | |
4292 | ||
4293 | /* copy worker description */ | |
4294 | probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid)); | |
4295 | if (desc_valid) | |
4296 | probe_kernel_read(desc, worker->desc, sizeof(desc) - 1); | |
4297 | ||
4298 | if (fn || name[0] || desc[0]) { | |
4299 | printk("%sWorkqueue: %s %pf", log_lvl, name, fn); | |
4300 | if (desc[0]) | |
4301 | pr_cont(" (%s)", desc); | |
4302 | pr_cont("\n"); | |
4303 | } | |
4304 | } | |
4305 | ||
3494fc30 TH |
4306 | static void pr_cont_pool_info(struct worker_pool *pool) |
4307 | { | |
4308 | pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask); | |
4309 | if (pool->node != NUMA_NO_NODE) | |
4310 | pr_cont(" node=%d", pool->node); | |
4311 | pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice); | |
4312 | } | |
4313 | ||
4314 | static void pr_cont_work(bool comma, struct work_struct *work) | |
4315 | { | |
4316 | if (work->func == wq_barrier_func) { | |
4317 | struct wq_barrier *barr; | |
4318 | ||
4319 | barr = container_of(work, struct wq_barrier, work); | |
4320 | ||
4321 | pr_cont("%s BAR(%d)", comma ? "," : "", | |
4322 | task_pid_nr(barr->task)); | |
4323 | } else { | |
4324 | pr_cont("%s %pf", comma ? "," : "", work->func); | |
4325 | } | |
4326 | } | |
4327 | ||
4328 | static void show_pwq(struct pool_workqueue *pwq) | |
4329 | { | |
4330 | struct worker_pool *pool = pwq->pool; | |
4331 | struct work_struct *work; | |
4332 | struct worker *worker; | |
4333 | bool has_in_flight = false, has_pending = false; | |
4334 | int bkt; | |
4335 | ||
4336 | pr_info(" pwq %d:", pool->id); | |
4337 | pr_cont_pool_info(pool); | |
4338 | ||
4339 | pr_cont(" active=%d/%d%s\n", pwq->nr_active, pwq->max_active, | |
4340 | !list_empty(&pwq->mayday_node) ? " MAYDAY" : ""); | |
4341 | ||
4342 | hash_for_each(pool->busy_hash, bkt, worker, hentry) { | |
4343 | if (worker->current_pwq == pwq) { | |
4344 | has_in_flight = true; | |
4345 | break; | |
4346 | } | |
4347 | } | |
4348 | if (has_in_flight) { | |
4349 | bool comma = false; | |
4350 | ||
4351 | pr_info(" in-flight:"); | |
4352 | hash_for_each(pool->busy_hash, bkt, worker, hentry) { | |
4353 | if (worker->current_pwq != pwq) | |
4354 | continue; | |
4355 | ||
4356 | pr_cont("%s %d%s:%pf", comma ? "," : "", | |
4357 | task_pid_nr(worker->task), | |
4358 | worker == pwq->wq->rescuer ? "(RESCUER)" : "", | |
4359 | worker->current_func); | |
4360 | list_for_each_entry(work, &worker->scheduled, entry) | |
4361 | pr_cont_work(false, work); | |
4362 | comma = true; | |
4363 | } | |
4364 | pr_cont("\n"); | |
4365 | } | |
4366 | ||
4367 | list_for_each_entry(work, &pool->worklist, entry) { | |
4368 | if (get_work_pwq(work) == pwq) { | |
4369 | has_pending = true; | |
4370 | break; | |
4371 | } | |
4372 | } | |
4373 | if (has_pending) { | |
4374 | bool comma = false; | |
4375 | ||
4376 | pr_info(" pending:"); | |
4377 | list_for_each_entry(work, &pool->worklist, entry) { | |
4378 | if (get_work_pwq(work) != pwq) | |
4379 | continue; | |
4380 | ||
4381 | pr_cont_work(comma, work); | |
4382 | comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); | |
4383 | } | |
4384 | pr_cont("\n"); | |
4385 | } | |
4386 | ||
4387 | if (!list_empty(&pwq->delayed_works)) { | |
4388 | bool comma = false; | |
4389 | ||
4390 | pr_info(" delayed:"); | |
4391 | list_for_each_entry(work, &pwq->delayed_works, entry) { | |
4392 | pr_cont_work(comma, work); | |
4393 | comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); | |
4394 | } | |
4395 | pr_cont("\n"); | |
4396 | } | |
4397 | } | |
4398 | ||
4399 | /** | |
4400 | * show_workqueue_state - dump workqueue state | |
4401 | * | |
7b776af6 RL |
4402 | * Called from a sysrq handler or try_to_freeze_tasks() and prints out |
4403 | * all busy workqueues and pools. | |
3494fc30 TH |
4404 | */ |
4405 | void show_workqueue_state(void) | |
4406 | { | |
4407 | struct workqueue_struct *wq; | |
4408 | struct worker_pool *pool; | |
4409 | unsigned long flags; | |
4410 | int pi; | |
4411 | ||
4412 | rcu_read_lock_sched(); | |
4413 | ||
4414 | pr_info("Showing busy workqueues and worker pools:\n"); | |
4415 | ||
4416 | list_for_each_entry_rcu(wq, &workqueues, list) { | |
4417 | struct pool_workqueue *pwq; | |
4418 | bool idle = true; | |
4419 | ||
4420 | for_each_pwq(pwq, wq) { | |
4421 | if (pwq->nr_active || !list_empty(&pwq->delayed_works)) { | |
4422 | idle = false; | |
4423 | break; | |
4424 | } | |
4425 | } | |
4426 | if (idle) | |
4427 | continue; | |
4428 | ||
4429 | pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags); | |
4430 | ||
4431 | for_each_pwq(pwq, wq) { | |
4432 | spin_lock_irqsave(&pwq->pool->lock, flags); | |
4433 | if (pwq->nr_active || !list_empty(&pwq->delayed_works)) | |
4434 | show_pwq(pwq); | |
4435 | spin_unlock_irqrestore(&pwq->pool->lock, flags); | |
4436 | } | |
4437 | } | |
4438 | ||
4439 | for_each_pool(pool, pi) { | |
4440 | struct worker *worker; | |
4441 | bool first = true; | |
4442 | ||
4443 | spin_lock_irqsave(&pool->lock, flags); | |
4444 | if (pool->nr_workers == pool->nr_idle) | |
4445 | goto next_pool; | |
4446 | ||
4447 | pr_info("pool %d:", pool->id); | |
4448 | pr_cont_pool_info(pool); | |
82607adc TH |
4449 | pr_cont(" hung=%us workers=%d", |
4450 | jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000, | |
4451 | pool->nr_workers); | |
3494fc30 TH |
4452 | if (pool->manager) |
4453 | pr_cont(" manager: %d", | |
4454 | task_pid_nr(pool->manager->task)); | |
4455 | list_for_each_entry(worker, &pool->idle_list, entry) { | |
4456 | pr_cont(" %s%d", first ? "idle: " : "", | |
4457 | task_pid_nr(worker->task)); | |
4458 | first = false; | |
4459 | } | |
4460 | pr_cont("\n"); | |
4461 | next_pool: | |
4462 | spin_unlock_irqrestore(&pool->lock, flags); | |
4463 | } | |
4464 | ||
4465 | rcu_read_unlock_sched(); | |
4466 | } | |
4467 | ||
db7bccf4 TH |
4468 | /* |
4469 | * CPU hotplug. | |
4470 | * | |
e22bee78 | 4471 | * There are two challenges in supporting CPU hotplug. Firstly, there |
112202d9 | 4472 | * are a lot of assumptions on strong associations among work, pwq and |
706026c2 | 4473 | * pool which make migrating pending and scheduled works very |
e22bee78 | 4474 | * difficult to implement without impacting hot paths. Secondly, |
94cf58bb | 4475 | * worker pools serve mix of short, long and very long running works making |
e22bee78 TH |
4476 | * blocked draining impractical. |
4477 | * | |
24647570 | 4478 | * This is solved by allowing the pools to be disassociated from the CPU |
628c78e7 TH |
4479 | * running as an unbound one and allowing it to be reattached later if the |
4480 | * cpu comes back online. | |
db7bccf4 | 4481 | */ |
1da177e4 | 4482 | |
706026c2 | 4483 | static void wq_unbind_fn(struct work_struct *work) |
3af24433 | 4484 | { |
38db41d9 | 4485 | int cpu = smp_processor_id(); |
4ce62e9e | 4486 | struct worker_pool *pool; |
db7bccf4 | 4487 | struct worker *worker; |
3af24433 | 4488 | |
f02ae73a | 4489 | for_each_cpu_worker_pool(pool, cpu) { |
92f9c5c4 | 4490 | mutex_lock(&pool->attach_mutex); |
94cf58bb | 4491 | spin_lock_irq(&pool->lock); |
3af24433 | 4492 | |
94cf58bb | 4493 | /* |
92f9c5c4 | 4494 | * We've blocked all attach/detach operations. Make all workers |
94cf58bb TH |
4495 | * unbound and set DISASSOCIATED. Before this, all workers |
4496 | * except for the ones which are still executing works from | |
4497 | * before the last CPU down must be on the cpu. After | |
4498 | * this, they may become diasporas. | |
4499 | */ | |
da028469 | 4500 | for_each_pool_worker(worker, pool) |
c9e7cf27 | 4501 | worker->flags |= WORKER_UNBOUND; |
06ba38a9 | 4502 | |
24647570 | 4503 | pool->flags |= POOL_DISASSOCIATED; |
f2d5a0ee | 4504 | |
94cf58bb | 4505 | spin_unlock_irq(&pool->lock); |
92f9c5c4 | 4506 | mutex_unlock(&pool->attach_mutex); |
628c78e7 | 4507 | |
eb283428 LJ |
4508 | /* |
4509 | * Call schedule() so that we cross rq->lock and thus can | |
4510 | * guarantee sched callbacks see the %WORKER_UNBOUND flag. | |
4511 | * This is necessary as scheduler callbacks may be invoked | |
4512 | * from other cpus. | |
4513 | */ | |
4514 | schedule(); | |
06ba38a9 | 4515 | |
eb283428 LJ |
4516 | /* |
4517 | * Sched callbacks are disabled now. Zap nr_running. | |
4518 | * After this, nr_running stays zero and need_more_worker() | |
4519 | * and keep_working() are always true as long as the | |
4520 | * worklist is not empty. This pool now behaves as an | |
4521 | * unbound (in terms of concurrency management) pool which | |
4522 | * are served by workers tied to the pool. | |
4523 | */ | |
e19e397a | 4524 | atomic_set(&pool->nr_running, 0); |
eb283428 LJ |
4525 | |
4526 | /* | |
4527 | * With concurrency management just turned off, a busy | |
4528 | * worker blocking could lead to lengthy stalls. Kick off | |
4529 | * unbound chain execution of currently pending work items. | |
4530 | */ | |
4531 | spin_lock_irq(&pool->lock); | |
4532 | wake_up_worker(pool); | |
4533 | spin_unlock_irq(&pool->lock); | |
4534 | } | |
3af24433 | 4535 | } |
3af24433 | 4536 | |
bd7c089e TH |
4537 | /** |
4538 | * rebind_workers - rebind all workers of a pool to the associated CPU | |
4539 | * @pool: pool of interest | |
4540 | * | |
a9ab775b | 4541 | * @pool->cpu is coming online. Rebind all workers to the CPU. |
bd7c089e TH |
4542 | */ |
4543 | static void rebind_workers(struct worker_pool *pool) | |
4544 | { | |
a9ab775b | 4545 | struct worker *worker; |
bd7c089e | 4546 | |
92f9c5c4 | 4547 | lockdep_assert_held(&pool->attach_mutex); |
bd7c089e | 4548 | |
a9ab775b TH |
4549 | /* |
4550 | * Restore CPU affinity of all workers. As all idle workers should | |
4551 | * be on the run-queue of the associated CPU before any local | |
402dd89d | 4552 | * wake-ups for concurrency management happen, restore CPU affinity |
a9ab775b TH |
4553 | * of all workers first and then clear UNBOUND. As we're called |
4554 | * from CPU_ONLINE, the following shouldn't fail. | |
4555 | */ | |
da028469 | 4556 | for_each_pool_worker(worker, pool) |
a9ab775b TH |
4557 | WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, |
4558 | pool->attrs->cpumask) < 0); | |
bd7c089e | 4559 | |
a9ab775b | 4560 | spin_lock_irq(&pool->lock); |
f7c17d26 WL |
4561 | |
4562 | /* | |
4563 | * XXX: CPU hotplug notifiers are weird and can call DOWN_FAILED | |
4564 | * w/o preceding DOWN_PREPARE. Work around it. CPU hotplug is | |
4565 | * being reworked and this can go away in time. | |
4566 | */ | |
4567 | if (!(pool->flags & POOL_DISASSOCIATED)) { | |
4568 | spin_unlock_irq(&pool->lock); | |
4569 | return; | |
4570 | } | |
4571 | ||
3de5e884 | 4572 | pool->flags &= ~POOL_DISASSOCIATED; |
bd7c089e | 4573 | |
da028469 | 4574 | for_each_pool_worker(worker, pool) { |
a9ab775b | 4575 | unsigned int worker_flags = worker->flags; |
bd7c089e TH |
4576 | |
4577 | /* | |
a9ab775b TH |
4578 | * A bound idle worker should actually be on the runqueue |
4579 | * of the associated CPU for local wake-ups targeting it to | |
4580 | * work. Kick all idle workers so that they migrate to the | |
4581 | * associated CPU. Doing this in the same loop as | |
4582 | * replacing UNBOUND with REBOUND is safe as no worker will | |
4583 | * be bound before @pool->lock is released. | |
bd7c089e | 4584 | */ |
a9ab775b TH |
4585 | if (worker_flags & WORKER_IDLE) |
4586 | wake_up_process(worker->task); | |
bd7c089e | 4587 | |
a9ab775b TH |
4588 | /* |
4589 | * We want to clear UNBOUND but can't directly call | |
4590 | * worker_clr_flags() or adjust nr_running. Atomically | |
4591 | * replace UNBOUND with another NOT_RUNNING flag REBOUND. | |
4592 | * @worker will clear REBOUND using worker_clr_flags() when | |
4593 | * it initiates the next execution cycle thus restoring | |
4594 | * concurrency management. Note that when or whether | |
4595 | * @worker clears REBOUND doesn't affect correctness. | |
4596 | * | |
4597 | * ACCESS_ONCE() is necessary because @worker->flags may be | |
4598 | * tested without holding any lock in | |
4599 | * wq_worker_waking_up(). Without it, NOT_RUNNING test may | |
4600 | * fail incorrectly leading to premature concurrency | |
4601 | * management operations. | |
4602 | */ | |
4603 | WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); | |
4604 | worker_flags |= WORKER_REBOUND; | |
4605 | worker_flags &= ~WORKER_UNBOUND; | |
4606 | ACCESS_ONCE(worker->flags) = worker_flags; | |
bd7c089e | 4607 | } |
a9ab775b TH |
4608 | |
4609 | spin_unlock_irq(&pool->lock); | |
bd7c089e TH |
4610 | } |
4611 | ||
7dbc725e TH |
4612 | /** |
4613 | * restore_unbound_workers_cpumask - restore cpumask of unbound workers | |
4614 | * @pool: unbound pool of interest | |
4615 | * @cpu: the CPU which is coming up | |
4616 | * | |
4617 | * An unbound pool may end up with a cpumask which doesn't have any online | |
4618 | * CPUs. When a worker of such pool get scheduled, the scheduler resets | |
4619 | * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any | |
4620 | * online CPU before, cpus_allowed of all its workers should be restored. | |
4621 | */ | |
4622 | static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) | |
4623 | { | |
4624 | static cpumask_t cpumask; | |
4625 | struct worker *worker; | |
7dbc725e | 4626 | |
92f9c5c4 | 4627 | lockdep_assert_held(&pool->attach_mutex); |
7dbc725e TH |
4628 | |
4629 | /* is @cpu allowed for @pool? */ | |
4630 | if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) | |
4631 | return; | |
4632 | ||
7dbc725e | 4633 | cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); |
7dbc725e TH |
4634 | |
4635 | /* as we're called from CPU_ONLINE, the following shouldn't fail */ | |
da028469 | 4636 | for_each_pool_worker(worker, pool) |
d945b5e9 | 4637 | WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0); |
7dbc725e TH |
4638 | } |
4639 | ||
7ee681b2 TG |
4640 | int workqueue_prepare_cpu(unsigned int cpu) |
4641 | { | |
4642 | struct worker_pool *pool; | |
4643 | ||
4644 | for_each_cpu_worker_pool(pool, cpu) { | |
4645 | if (pool->nr_workers) | |
4646 | continue; | |
4647 | if (!create_worker(pool)) | |
4648 | return -ENOMEM; | |
4649 | } | |
4650 | return 0; | |
4651 | } | |
4652 | ||
4653 | int workqueue_online_cpu(unsigned int cpu) | |
3af24433 | 4654 | { |
4ce62e9e | 4655 | struct worker_pool *pool; |
4c16bd32 | 4656 | struct workqueue_struct *wq; |
7dbc725e | 4657 | int pi; |
3ce63377 | 4658 | |
7ee681b2 | 4659 | mutex_lock(&wq_pool_mutex); |
7dbc725e | 4660 | |
7ee681b2 TG |
4661 | for_each_pool(pool, pi) { |
4662 | mutex_lock(&pool->attach_mutex); | |
94cf58bb | 4663 | |
7ee681b2 TG |
4664 | if (pool->cpu == cpu) |
4665 | rebind_workers(pool); | |
4666 | else if (pool->cpu < 0) | |
4667 | restore_unbound_workers_cpumask(pool, cpu); | |
94cf58bb | 4668 | |
7ee681b2 TG |
4669 | mutex_unlock(&pool->attach_mutex); |
4670 | } | |
6ba94429 | 4671 | |
7ee681b2 TG |
4672 | /* update NUMA affinity of unbound workqueues */ |
4673 | list_for_each_entry(wq, &workqueues, list) | |
4674 | wq_update_unbound_numa(wq, cpu, true); | |
6ba94429 | 4675 | |
7ee681b2 TG |
4676 | mutex_unlock(&wq_pool_mutex); |
4677 | return 0; | |
6ba94429 FW |
4678 | } |
4679 | ||
7ee681b2 | 4680 | int workqueue_offline_cpu(unsigned int cpu) |
6ba94429 | 4681 | { |
6ba94429 FW |
4682 | struct work_struct unbind_work; |
4683 | struct workqueue_struct *wq; | |
4684 | ||
7ee681b2 TG |
4685 | /* unbinding per-cpu workers should happen on the local CPU */ |
4686 | INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn); | |
4687 | queue_work_on(cpu, system_highpri_wq, &unbind_work); | |
4688 | ||
4689 | /* update NUMA affinity of unbound workqueues */ | |
4690 | mutex_lock(&wq_pool_mutex); | |
4691 | list_for_each_entry(wq, &workqueues, list) | |
4692 | wq_update_unbound_numa(wq, cpu, false); | |
4693 | mutex_unlock(&wq_pool_mutex); | |
4694 | ||
4695 | /* wait for per-cpu unbinding to finish */ | |
4696 | flush_work(&unbind_work); | |
4697 | destroy_work_on_stack(&unbind_work); | |
4698 | return 0; | |
6ba94429 FW |
4699 | } |
4700 | ||
4701 | #ifdef CONFIG_SMP | |
4702 | ||
4703 | struct work_for_cpu { | |
4704 | struct work_struct work; | |
4705 | long (*fn)(void *); | |
4706 | void *arg; | |
4707 | long ret; | |
4708 | }; | |
4709 | ||
4710 | static void work_for_cpu_fn(struct work_struct *work) | |
4711 | { | |
4712 | struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); | |
4713 | ||
4714 | wfc->ret = wfc->fn(wfc->arg); | |
4715 | } | |
4716 | ||
4717 | /** | |
22aceb31 | 4718 | * work_on_cpu - run a function in thread context on a particular cpu |
6ba94429 FW |
4719 | * @cpu: the cpu to run on |
4720 | * @fn: the function to run | |
4721 | * @arg: the function arg | |
4722 | * | |
4723 | * It is up to the caller to ensure that the cpu doesn't go offline. | |
4724 | * The caller must not hold any locks which would prevent @fn from completing. | |
4725 | * | |
4726 | * Return: The value @fn returns. | |
4727 | */ | |
4728 | long work_on_cpu(int cpu, long (*fn)(void *), void *arg) | |
4729 | { | |
4730 | struct work_for_cpu wfc = { .fn = fn, .arg = arg }; | |
4731 | ||
4732 | INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); | |
4733 | schedule_work_on(cpu, &wfc.work); | |
4734 | flush_work(&wfc.work); | |
4735 | destroy_work_on_stack(&wfc.work); | |
4736 | return wfc.ret; | |
4737 | } | |
4738 | EXPORT_SYMBOL_GPL(work_on_cpu); | |
4739 | #endif /* CONFIG_SMP */ | |
4740 | ||
4741 | #ifdef CONFIG_FREEZER | |
4742 | ||
4743 | /** | |
4744 | * freeze_workqueues_begin - begin freezing workqueues | |
4745 | * | |
4746 | * Start freezing workqueues. After this function returns, all freezable | |
4747 | * workqueues will queue new works to their delayed_works list instead of | |
4748 | * pool->worklist. | |
4749 | * | |
4750 | * CONTEXT: | |
4751 | * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. | |
4752 | */ | |
4753 | void freeze_workqueues_begin(void) | |
4754 | { | |
4755 | struct workqueue_struct *wq; | |
4756 | struct pool_workqueue *pwq; | |
4757 | ||
4758 | mutex_lock(&wq_pool_mutex); | |
4759 | ||
4760 | WARN_ON_ONCE(workqueue_freezing); | |
4761 | workqueue_freezing = true; | |
4762 | ||
4763 | list_for_each_entry(wq, &workqueues, list) { | |
4764 | mutex_lock(&wq->mutex); | |
4765 | for_each_pwq(pwq, wq) | |
4766 | pwq_adjust_max_active(pwq); | |
4767 | mutex_unlock(&wq->mutex); | |
4768 | } | |
4769 | ||
4770 | mutex_unlock(&wq_pool_mutex); | |
4771 | } | |
4772 | ||
4773 | /** | |
4774 | * freeze_workqueues_busy - are freezable workqueues still busy? | |
4775 | * | |
4776 | * Check whether freezing is complete. This function must be called | |
4777 | * between freeze_workqueues_begin() and thaw_workqueues(). | |
4778 | * | |
4779 | * CONTEXT: | |
4780 | * Grabs and releases wq_pool_mutex. | |
4781 | * | |
4782 | * Return: | |
4783 | * %true if some freezable workqueues are still busy. %false if freezing | |
4784 | * is complete. | |
4785 | */ | |
4786 | bool freeze_workqueues_busy(void) | |
4787 | { | |
4788 | bool busy = false; | |
4789 | struct workqueue_struct *wq; | |
4790 | struct pool_workqueue *pwq; | |
4791 | ||
4792 | mutex_lock(&wq_pool_mutex); | |
4793 | ||
4794 | WARN_ON_ONCE(!workqueue_freezing); | |
4795 | ||
4796 | list_for_each_entry(wq, &workqueues, list) { | |
4797 | if (!(wq->flags & WQ_FREEZABLE)) | |
4798 | continue; | |
4799 | /* | |
4800 | * nr_active is monotonically decreasing. It's safe | |
4801 | * to peek without lock. | |
4802 | */ | |
4803 | rcu_read_lock_sched(); | |
4804 | for_each_pwq(pwq, wq) { | |
4805 | WARN_ON_ONCE(pwq->nr_active < 0); | |
4806 | if (pwq->nr_active) { | |
4807 | busy = true; | |
4808 | rcu_read_unlock_sched(); | |
4809 | goto out_unlock; | |
4810 | } | |
4811 | } | |
4812 | rcu_read_unlock_sched(); | |
4813 | } | |
4814 | out_unlock: | |
4815 | mutex_unlock(&wq_pool_mutex); | |
4816 | return busy; | |
4817 | } | |
4818 | ||
4819 | /** | |
4820 | * thaw_workqueues - thaw workqueues | |
4821 | * | |
4822 | * Thaw workqueues. Normal queueing is restored and all collected | |
4823 | * frozen works are transferred to their respective pool worklists. | |
4824 | * | |
4825 | * CONTEXT: | |
4826 | * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. | |
4827 | */ | |
4828 | void thaw_workqueues(void) | |
4829 | { | |
4830 | struct workqueue_struct *wq; | |
4831 | struct pool_workqueue *pwq; | |
4832 | ||
4833 | mutex_lock(&wq_pool_mutex); | |
4834 | ||
4835 | if (!workqueue_freezing) | |
4836 | goto out_unlock; | |
4837 | ||
4838 | workqueue_freezing = false; | |
4839 | ||
4840 | /* restore max_active and repopulate worklist */ | |
4841 | list_for_each_entry(wq, &workqueues, list) { | |
4842 | mutex_lock(&wq->mutex); | |
4843 | for_each_pwq(pwq, wq) | |
4844 | pwq_adjust_max_active(pwq); | |
4845 | mutex_unlock(&wq->mutex); | |
4846 | } | |
4847 | ||
4848 | out_unlock: | |
4849 | mutex_unlock(&wq_pool_mutex); | |
4850 | } | |
4851 | #endif /* CONFIG_FREEZER */ | |
4852 | ||
042f7df1 LJ |
4853 | static int workqueue_apply_unbound_cpumask(void) |
4854 | { | |
4855 | LIST_HEAD(ctxs); | |
4856 | int ret = 0; | |
4857 | struct workqueue_struct *wq; | |
4858 | struct apply_wqattrs_ctx *ctx, *n; | |
4859 | ||
4860 | lockdep_assert_held(&wq_pool_mutex); | |
4861 | ||
4862 | list_for_each_entry(wq, &workqueues, list) { | |
4863 | if (!(wq->flags & WQ_UNBOUND)) | |
4864 | continue; | |
4865 | /* creating multiple pwqs breaks ordering guarantee */ | |
4866 | if (wq->flags & __WQ_ORDERED) | |
4867 | continue; | |
4868 | ||
4869 | ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs); | |
4870 | if (!ctx) { | |
4871 | ret = -ENOMEM; | |
4872 | break; | |
4873 | } | |
4874 | ||
4875 | list_add_tail(&ctx->list, &ctxs); | |
4876 | } | |
4877 | ||
4878 | list_for_each_entry_safe(ctx, n, &ctxs, list) { | |
4879 | if (!ret) | |
4880 | apply_wqattrs_commit(ctx); | |
4881 | apply_wqattrs_cleanup(ctx); | |
4882 | } | |
4883 | ||
4884 | return ret; | |
4885 | } | |
4886 | ||
4887 | /** | |
4888 | * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask | |
4889 | * @cpumask: the cpumask to set | |
4890 | * | |
4891 | * The low-level workqueues cpumask is a global cpumask that limits | |
4892 | * the affinity of all unbound workqueues. This function check the @cpumask | |
4893 | * and apply it to all unbound workqueues and updates all pwqs of them. | |
4894 | * | |
4895 | * Retun: 0 - Success | |
4896 | * -EINVAL - Invalid @cpumask | |
4897 | * -ENOMEM - Failed to allocate memory for attrs or pwqs. | |
4898 | */ | |
4899 | int workqueue_set_unbound_cpumask(cpumask_var_t cpumask) | |
4900 | { | |
4901 | int ret = -EINVAL; | |
4902 | cpumask_var_t saved_cpumask; | |
4903 | ||
4904 | if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL)) | |
4905 | return -ENOMEM; | |
4906 | ||
042f7df1 LJ |
4907 | cpumask_and(cpumask, cpumask, cpu_possible_mask); |
4908 | if (!cpumask_empty(cpumask)) { | |
a0111cf6 | 4909 | apply_wqattrs_lock(); |
042f7df1 LJ |
4910 | |
4911 | /* save the old wq_unbound_cpumask. */ | |
4912 | cpumask_copy(saved_cpumask, wq_unbound_cpumask); | |
4913 | ||
4914 | /* update wq_unbound_cpumask at first and apply it to wqs. */ | |
4915 | cpumask_copy(wq_unbound_cpumask, cpumask); | |
4916 | ret = workqueue_apply_unbound_cpumask(); | |
4917 | ||
4918 | /* restore the wq_unbound_cpumask when failed. */ | |
4919 | if (ret < 0) | |
4920 | cpumask_copy(wq_unbound_cpumask, saved_cpumask); | |
4921 | ||
a0111cf6 | 4922 | apply_wqattrs_unlock(); |
042f7df1 | 4923 | } |
042f7df1 LJ |
4924 | |
4925 | free_cpumask_var(saved_cpumask); | |
4926 | return ret; | |
4927 | } | |
4928 | ||
6ba94429 FW |
4929 | #ifdef CONFIG_SYSFS |
4930 | /* | |
4931 | * Workqueues with WQ_SYSFS flag set is visible to userland via | |
4932 | * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the | |
4933 | * following attributes. | |
4934 | * | |
4935 | * per_cpu RO bool : whether the workqueue is per-cpu or unbound | |
4936 | * max_active RW int : maximum number of in-flight work items | |
4937 | * | |
4938 | * Unbound workqueues have the following extra attributes. | |
4939 | * | |
4940 | * id RO int : the associated pool ID | |
4941 | * nice RW int : nice value of the workers | |
4942 | * cpumask RW mask : bitmask of allowed CPUs for the workers | |
4943 | */ | |
4944 | struct wq_device { | |
4945 | struct workqueue_struct *wq; | |
4946 | struct device dev; | |
4947 | }; | |
4948 | ||
4949 | static struct workqueue_struct *dev_to_wq(struct device *dev) | |
4950 | { | |
4951 | struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); | |
4952 | ||
4953 | return wq_dev->wq; | |
4954 | } | |
4955 | ||
4956 | static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, | |
4957 | char *buf) | |
4958 | { | |
4959 | struct workqueue_struct *wq = dev_to_wq(dev); | |
4960 | ||
4961 | return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); | |
4962 | } | |
4963 | static DEVICE_ATTR_RO(per_cpu); | |
4964 | ||
4965 | static ssize_t max_active_show(struct device *dev, | |
4966 | struct device_attribute *attr, char *buf) | |
4967 | { | |
4968 | struct workqueue_struct *wq = dev_to_wq(dev); | |
4969 | ||
4970 | return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); | |
4971 | } | |
4972 | ||
4973 | static ssize_t max_active_store(struct device *dev, | |
4974 | struct device_attribute *attr, const char *buf, | |
4975 | size_t count) | |
4976 | { | |
4977 | struct workqueue_struct *wq = dev_to_wq(dev); | |
4978 | int val; | |
4979 | ||
4980 | if (sscanf(buf, "%d", &val) != 1 || val <= 0) | |
4981 | return -EINVAL; | |
4982 | ||
4983 | workqueue_set_max_active(wq, val); | |
4984 | return count; | |
4985 | } | |
4986 | static DEVICE_ATTR_RW(max_active); | |
4987 | ||
4988 | static struct attribute *wq_sysfs_attrs[] = { | |
4989 | &dev_attr_per_cpu.attr, | |
4990 | &dev_attr_max_active.attr, | |
4991 | NULL, | |
4992 | }; | |
4993 | ATTRIBUTE_GROUPS(wq_sysfs); | |
4994 | ||
4995 | static ssize_t wq_pool_ids_show(struct device *dev, | |
4996 | struct device_attribute *attr, char *buf) | |
4997 | { | |
4998 | struct workqueue_struct *wq = dev_to_wq(dev); | |
4999 | const char *delim = ""; | |
5000 | int node, written = 0; | |
5001 | ||
5002 | rcu_read_lock_sched(); | |
5003 | for_each_node(node) { | |
5004 | written += scnprintf(buf + written, PAGE_SIZE - written, | |
5005 | "%s%d:%d", delim, node, | |
5006 | unbound_pwq_by_node(wq, node)->pool->id); | |
5007 | delim = " "; | |
5008 | } | |
5009 | written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); | |
5010 | rcu_read_unlock_sched(); | |
5011 | ||
5012 | return written; | |
5013 | } | |
5014 | ||
5015 | static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, | |
5016 | char *buf) | |
5017 | { | |
5018 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5019 | int written; | |
5020 | ||
5021 | mutex_lock(&wq->mutex); | |
5022 | written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); | |
5023 | mutex_unlock(&wq->mutex); | |
5024 | ||
5025 | return written; | |
5026 | } | |
5027 | ||
5028 | /* prepare workqueue_attrs for sysfs store operations */ | |
5029 | static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) | |
5030 | { | |
5031 | struct workqueue_attrs *attrs; | |
5032 | ||
899a94fe LJ |
5033 | lockdep_assert_held(&wq_pool_mutex); |
5034 | ||
6ba94429 FW |
5035 | attrs = alloc_workqueue_attrs(GFP_KERNEL); |
5036 | if (!attrs) | |
5037 | return NULL; | |
5038 | ||
6ba94429 | 5039 | copy_workqueue_attrs(attrs, wq->unbound_attrs); |
6ba94429 FW |
5040 | return attrs; |
5041 | } | |
5042 | ||
5043 | static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, | |
5044 | const char *buf, size_t count) | |
5045 | { | |
5046 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5047 | struct workqueue_attrs *attrs; | |
d4d3e257 LJ |
5048 | int ret = -ENOMEM; |
5049 | ||
5050 | apply_wqattrs_lock(); | |
6ba94429 FW |
5051 | |
5052 | attrs = wq_sysfs_prep_attrs(wq); | |
5053 | if (!attrs) | |
d4d3e257 | 5054 | goto out_unlock; |
6ba94429 FW |
5055 | |
5056 | if (sscanf(buf, "%d", &attrs->nice) == 1 && | |
5057 | attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) | |
d4d3e257 | 5058 | ret = apply_workqueue_attrs_locked(wq, attrs); |
6ba94429 FW |
5059 | else |
5060 | ret = -EINVAL; | |
5061 | ||
d4d3e257 LJ |
5062 | out_unlock: |
5063 | apply_wqattrs_unlock(); | |
6ba94429 FW |
5064 | free_workqueue_attrs(attrs); |
5065 | return ret ?: count; | |
5066 | } | |
5067 | ||
5068 | static ssize_t wq_cpumask_show(struct device *dev, | |
5069 | struct device_attribute *attr, char *buf) | |
5070 | { | |
5071 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5072 | int written; | |
5073 | ||
5074 | mutex_lock(&wq->mutex); | |
5075 | written = scnprintf(buf, PAGE_SIZE, "%*pb\n", | |
5076 | cpumask_pr_args(wq->unbound_attrs->cpumask)); | |
5077 | mutex_unlock(&wq->mutex); | |
5078 | return written; | |
5079 | } | |
5080 | ||
5081 | static ssize_t wq_cpumask_store(struct device *dev, | |
5082 | struct device_attribute *attr, | |
5083 | const char *buf, size_t count) | |
5084 | { | |
5085 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5086 | struct workqueue_attrs *attrs; | |
d4d3e257 LJ |
5087 | int ret = -ENOMEM; |
5088 | ||
5089 | apply_wqattrs_lock(); | |
6ba94429 FW |
5090 | |
5091 | attrs = wq_sysfs_prep_attrs(wq); | |
5092 | if (!attrs) | |
d4d3e257 | 5093 | goto out_unlock; |
6ba94429 FW |
5094 | |
5095 | ret = cpumask_parse(buf, attrs->cpumask); | |
5096 | if (!ret) | |
d4d3e257 | 5097 | ret = apply_workqueue_attrs_locked(wq, attrs); |
6ba94429 | 5098 | |
d4d3e257 LJ |
5099 | out_unlock: |
5100 | apply_wqattrs_unlock(); | |
6ba94429 FW |
5101 | free_workqueue_attrs(attrs); |
5102 | return ret ?: count; | |
5103 | } | |
5104 | ||
5105 | static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr, | |
5106 | char *buf) | |
5107 | { | |
5108 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5109 | int written; | |
7dbc725e | 5110 | |
6ba94429 FW |
5111 | mutex_lock(&wq->mutex); |
5112 | written = scnprintf(buf, PAGE_SIZE, "%d\n", | |
5113 | !wq->unbound_attrs->no_numa); | |
5114 | mutex_unlock(&wq->mutex); | |
4c16bd32 | 5115 | |
6ba94429 | 5116 | return written; |
65758202 TH |
5117 | } |
5118 | ||
6ba94429 FW |
5119 | static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr, |
5120 | const char *buf, size_t count) | |
65758202 | 5121 | { |
6ba94429 FW |
5122 | struct workqueue_struct *wq = dev_to_wq(dev); |
5123 | struct workqueue_attrs *attrs; | |
d4d3e257 LJ |
5124 | int v, ret = -ENOMEM; |
5125 | ||
5126 | apply_wqattrs_lock(); | |
4c16bd32 | 5127 | |
6ba94429 FW |
5128 | attrs = wq_sysfs_prep_attrs(wq); |
5129 | if (!attrs) | |
d4d3e257 | 5130 | goto out_unlock; |
4c16bd32 | 5131 | |
6ba94429 FW |
5132 | ret = -EINVAL; |
5133 | if (sscanf(buf, "%d", &v) == 1) { | |
5134 | attrs->no_numa = !v; | |
d4d3e257 | 5135 | ret = apply_workqueue_attrs_locked(wq, attrs); |
65758202 | 5136 | } |
6ba94429 | 5137 | |
d4d3e257 LJ |
5138 | out_unlock: |
5139 | apply_wqattrs_unlock(); | |
6ba94429 FW |
5140 | free_workqueue_attrs(attrs); |
5141 | return ret ?: count; | |
65758202 TH |
5142 | } |
5143 | ||
6ba94429 FW |
5144 | static struct device_attribute wq_sysfs_unbound_attrs[] = { |
5145 | __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL), | |
5146 | __ATTR(nice, 0644, wq_nice_show, wq_nice_store), | |
5147 | __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), | |
5148 | __ATTR(numa, 0644, wq_numa_show, wq_numa_store), | |
5149 | __ATTR_NULL, | |
5150 | }; | |
8ccad40d | 5151 | |
6ba94429 FW |
5152 | static struct bus_type wq_subsys = { |
5153 | .name = "workqueue", | |
5154 | .dev_groups = wq_sysfs_groups, | |
2d3854a3 RR |
5155 | }; |
5156 | ||
b05a7928 FW |
5157 | static ssize_t wq_unbound_cpumask_show(struct device *dev, |
5158 | struct device_attribute *attr, char *buf) | |
5159 | { | |
5160 | int written; | |
5161 | ||
042f7df1 | 5162 | mutex_lock(&wq_pool_mutex); |
b05a7928 FW |
5163 | written = scnprintf(buf, PAGE_SIZE, "%*pb\n", |
5164 | cpumask_pr_args(wq_unbound_cpumask)); | |
042f7df1 | 5165 | mutex_unlock(&wq_pool_mutex); |
b05a7928 FW |
5166 | |
5167 | return written; | |
5168 | } | |
5169 | ||
042f7df1 LJ |
5170 | static ssize_t wq_unbound_cpumask_store(struct device *dev, |
5171 | struct device_attribute *attr, const char *buf, size_t count) | |
5172 | { | |
5173 | cpumask_var_t cpumask; | |
5174 | int ret; | |
5175 | ||
5176 | if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) | |
5177 | return -ENOMEM; | |
5178 | ||
5179 | ret = cpumask_parse(buf, cpumask); | |
5180 | if (!ret) | |
5181 | ret = workqueue_set_unbound_cpumask(cpumask); | |
5182 | ||
5183 | free_cpumask_var(cpumask); | |
5184 | return ret ? ret : count; | |
5185 | } | |
5186 | ||
b05a7928 | 5187 | static struct device_attribute wq_sysfs_cpumask_attr = |
042f7df1 LJ |
5188 | __ATTR(cpumask, 0644, wq_unbound_cpumask_show, |
5189 | wq_unbound_cpumask_store); | |
b05a7928 | 5190 | |
6ba94429 | 5191 | static int __init wq_sysfs_init(void) |
2d3854a3 | 5192 | { |
b05a7928 FW |
5193 | int err; |
5194 | ||
5195 | err = subsys_virtual_register(&wq_subsys, NULL); | |
5196 | if (err) | |
5197 | return err; | |
5198 | ||
5199 | return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr); | |
2d3854a3 | 5200 | } |
6ba94429 | 5201 | core_initcall(wq_sysfs_init); |
2d3854a3 | 5202 | |
6ba94429 | 5203 | static void wq_device_release(struct device *dev) |
2d3854a3 | 5204 | { |
6ba94429 | 5205 | struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); |
6b44003e | 5206 | |
6ba94429 | 5207 | kfree(wq_dev); |
2d3854a3 | 5208 | } |
a0a1a5fd TH |
5209 | |
5210 | /** | |
6ba94429 FW |
5211 | * workqueue_sysfs_register - make a workqueue visible in sysfs |
5212 | * @wq: the workqueue to register | |
a0a1a5fd | 5213 | * |
6ba94429 FW |
5214 | * Expose @wq in sysfs under /sys/bus/workqueue/devices. |
5215 | * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set | |
5216 | * which is the preferred method. | |
a0a1a5fd | 5217 | * |
6ba94429 FW |
5218 | * Workqueue user should use this function directly iff it wants to apply |
5219 | * workqueue_attrs before making the workqueue visible in sysfs; otherwise, | |
5220 | * apply_workqueue_attrs() may race against userland updating the | |
5221 | * attributes. | |
5222 | * | |
5223 | * Return: 0 on success, -errno on failure. | |
a0a1a5fd | 5224 | */ |
6ba94429 | 5225 | int workqueue_sysfs_register(struct workqueue_struct *wq) |
a0a1a5fd | 5226 | { |
6ba94429 FW |
5227 | struct wq_device *wq_dev; |
5228 | int ret; | |
a0a1a5fd | 5229 | |
6ba94429 | 5230 | /* |
402dd89d | 5231 | * Adjusting max_active or creating new pwqs by applying |
6ba94429 FW |
5232 | * attributes breaks ordering guarantee. Disallow exposing ordered |
5233 | * workqueues. | |
5234 | */ | |
5235 | if (WARN_ON(wq->flags & __WQ_ORDERED)) | |
5236 | return -EINVAL; | |
a0a1a5fd | 5237 | |
6ba94429 FW |
5238 | wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); |
5239 | if (!wq_dev) | |
5240 | return -ENOMEM; | |
5bcab335 | 5241 | |
6ba94429 FW |
5242 | wq_dev->wq = wq; |
5243 | wq_dev->dev.bus = &wq_subsys; | |
6ba94429 | 5244 | wq_dev->dev.release = wq_device_release; |
23217b44 | 5245 | dev_set_name(&wq_dev->dev, "%s", wq->name); |
a0a1a5fd | 5246 | |
6ba94429 FW |
5247 | /* |
5248 | * unbound_attrs are created separately. Suppress uevent until | |
5249 | * everything is ready. | |
5250 | */ | |
5251 | dev_set_uevent_suppress(&wq_dev->dev, true); | |
a0a1a5fd | 5252 | |
6ba94429 FW |
5253 | ret = device_register(&wq_dev->dev); |
5254 | if (ret) { | |
5255 | kfree(wq_dev); | |
5256 | wq->wq_dev = NULL; | |
5257 | return ret; | |
5258 | } | |
a0a1a5fd | 5259 | |
6ba94429 FW |
5260 | if (wq->flags & WQ_UNBOUND) { |
5261 | struct device_attribute *attr; | |
a0a1a5fd | 5262 | |
6ba94429 FW |
5263 | for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { |
5264 | ret = device_create_file(&wq_dev->dev, attr); | |
5265 | if (ret) { | |
5266 | device_unregister(&wq_dev->dev); | |
5267 | wq->wq_dev = NULL; | |
5268 | return ret; | |
a0a1a5fd TH |
5269 | } |
5270 | } | |
5271 | } | |
6ba94429 FW |
5272 | |
5273 | dev_set_uevent_suppress(&wq_dev->dev, false); | |
5274 | kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); | |
5275 | return 0; | |
a0a1a5fd TH |
5276 | } |
5277 | ||
5278 | /** | |
6ba94429 FW |
5279 | * workqueue_sysfs_unregister - undo workqueue_sysfs_register() |
5280 | * @wq: the workqueue to unregister | |
a0a1a5fd | 5281 | * |
6ba94429 | 5282 | * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. |
a0a1a5fd | 5283 | */ |
6ba94429 | 5284 | static void workqueue_sysfs_unregister(struct workqueue_struct *wq) |
a0a1a5fd | 5285 | { |
6ba94429 | 5286 | struct wq_device *wq_dev = wq->wq_dev; |
8b03ae3c | 5287 | |
6ba94429 FW |
5288 | if (!wq->wq_dev) |
5289 | return; | |
a0a1a5fd | 5290 | |
6ba94429 FW |
5291 | wq->wq_dev = NULL; |
5292 | device_unregister(&wq_dev->dev); | |
a0a1a5fd | 5293 | } |
6ba94429 FW |
5294 | #else /* CONFIG_SYSFS */ |
5295 | static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } | |
5296 | #endif /* CONFIG_SYSFS */ | |
a0a1a5fd | 5297 | |
82607adc TH |
5298 | /* |
5299 | * Workqueue watchdog. | |
5300 | * | |
5301 | * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal | |
5302 | * flush dependency, a concurrency managed work item which stays RUNNING | |
5303 | * indefinitely. Workqueue stalls can be very difficult to debug as the | |
5304 | * usual warning mechanisms don't trigger and internal workqueue state is | |
5305 | * largely opaque. | |
5306 | * | |
5307 | * Workqueue watchdog monitors all worker pools periodically and dumps | |
5308 | * state if some pools failed to make forward progress for a while where | |
5309 | * forward progress is defined as the first item on ->worklist changing. | |
5310 | * | |
5311 | * This mechanism is controlled through the kernel parameter | |
5312 | * "workqueue.watchdog_thresh" which can be updated at runtime through the | |
5313 | * corresponding sysfs parameter file. | |
5314 | */ | |
5315 | #ifdef CONFIG_WQ_WATCHDOG | |
5316 | ||
5317 | static void wq_watchdog_timer_fn(unsigned long data); | |
5318 | ||
5319 | static unsigned long wq_watchdog_thresh = 30; | |
5320 | static struct timer_list wq_watchdog_timer = | |
5321 | TIMER_DEFERRED_INITIALIZER(wq_watchdog_timer_fn, 0, 0); | |
5322 | ||
5323 | static unsigned long wq_watchdog_touched = INITIAL_JIFFIES; | |
5324 | static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES; | |
5325 | ||
5326 | static void wq_watchdog_reset_touched(void) | |
5327 | { | |
5328 | int cpu; | |
5329 | ||
5330 | wq_watchdog_touched = jiffies; | |
5331 | for_each_possible_cpu(cpu) | |
5332 | per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; | |
5333 | } | |
5334 | ||
5335 | static void wq_watchdog_timer_fn(unsigned long data) | |
5336 | { | |
5337 | unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ; | |
5338 | bool lockup_detected = false; | |
5339 | struct worker_pool *pool; | |
5340 | int pi; | |
5341 | ||
5342 | if (!thresh) | |
5343 | return; | |
5344 | ||
5345 | rcu_read_lock(); | |
5346 | ||
5347 | for_each_pool(pool, pi) { | |
5348 | unsigned long pool_ts, touched, ts; | |
5349 | ||
5350 | if (list_empty(&pool->worklist)) | |
5351 | continue; | |
5352 | ||
5353 | /* get the latest of pool and touched timestamps */ | |
5354 | pool_ts = READ_ONCE(pool->watchdog_ts); | |
5355 | touched = READ_ONCE(wq_watchdog_touched); | |
5356 | ||
5357 | if (time_after(pool_ts, touched)) | |
5358 | ts = pool_ts; | |
5359 | else | |
5360 | ts = touched; | |
5361 | ||
5362 | if (pool->cpu >= 0) { | |
5363 | unsigned long cpu_touched = | |
5364 | READ_ONCE(per_cpu(wq_watchdog_touched_cpu, | |
5365 | pool->cpu)); | |
5366 | if (time_after(cpu_touched, ts)) | |
5367 | ts = cpu_touched; | |
5368 | } | |
5369 | ||
5370 | /* did we stall? */ | |
5371 | if (time_after(jiffies, ts + thresh)) { | |
5372 | lockup_detected = true; | |
5373 | pr_emerg("BUG: workqueue lockup - pool"); | |
5374 | pr_cont_pool_info(pool); | |
5375 | pr_cont(" stuck for %us!\n", | |
5376 | jiffies_to_msecs(jiffies - pool_ts) / 1000); | |
5377 | } | |
5378 | } | |
5379 | ||
5380 | rcu_read_unlock(); | |
5381 | ||
5382 | if (lockup_detected) | |
5383 | show_workqueue_state(); | |
5384 | ||
5385 | wq_watchdog_reset_touched(); | |
5386 | mod_timer(&wq_watchdog_timer, jiffies + thresh); | |
5387 | } | |
5388 | ||
5389 | void wq_watchdog_touch(int cpu) | |
5390 | { | |
5391 | if (cpu >= 0) | |
5392 | per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; | |
5393 | else | |
5394 | wq_watchdog_touched = jiffies; | |
5395 | } | |
5396 | ||
5397 | static void wq_watchdog_set_thresh(unsigned long thresh) | |
5398 | { | |
5399 | wq_watchdog_thresh = 0; | |
5400 | del_timer_sync(&wq_watchdog_timer); | |
5401 | ||
5402 | if (thresh) { | |
5403 | wq_watchdog_thresh = thresh; | |
5404 | wq_watchdog_reset_touched(); | |
5405 | mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ); | |
5406 | } | |
5407 | } | |
5408 | ||
5409 | static int wq_watchdog_param_set_thresh(const char *val, | |
5410 | const struct kernel_param *kp) | |
5411 | { | |
5412 | unsigned long thresh; | |
5413 | int ret; | |
5414 | ||
5415 | ret = kstrtoul(val, 0, &thresh); | |
5416 | if (ret) | |
5417 | return ret; | |
5418 | ||
5419 | if (system_wq) | |
5420 | wq_watchdog_set_thresh(thresh); | |
5421 | else | |
5422 | wq_watchdog_thresh = thresh; | |
5423 | ||
5424 | return 0; | |
5425 | } | |
5426 | ||
5427 | static const struct kernel_param_ops wq_watchdog_thresh_ops = { | |
5428 | .set = wq_watchdog_param_set_thresh, | |
5429 | .get = param_get_ulong, | |
5430 | }; | |
5431 | ||
5432 | module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh, | |
5433 | 0644); | |
5434 | ||
5435 | static void wq_watchdog_init(void) | |
5436 | { | |
5437 | wq_watchdog_set_thresh(wq_watchdog_thresh); | |
5438 | } | |
5439 | ||
5440 | #else /* CONFIG_WQ_WATCHDOG */ | |
5441 | ||
5442 | static inline void wq_watchdog_init(void) { } | |
5443 | ||
5444 | #endif /* CONFIG_WQ_WATCHDOG */ | |
5445 | ||
bce90380 TH |
5446 | static void __init wq_numa_init(void) |
5447 | { | |
5448 | cpumask_var_t *tbl; | |
5449 | int node, cpu; | |
5450 | ||
bce90380 TH |
5451 | if (num_possible_nodes() <= 1) |
5452 | return; | |
5453 | ||
d55262c4 TH |
5454 | if (wq_disable_numa) { |
5455 | pr_info("workqueue: NUMA affinity support disabled\n"); | |
5456 | return; | |
5457 | } | |
5458 | ||
4c16bd32 TH |
5459 | wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL); |
5460 | BUG_ON(!wq_update_unbound_numa_attrs_buf); | |
5461 | ||
bce90380 TH |
5462 | /* |
5463 | * We want masks of possible CPUs of each node which isn't readily | |
5464 | * available. Build one from cpu_to_node() which should have been | |
5465 | * fully initialized by now. | |
5466 | */ | |
ddcb57e2 | 5467 | tbl = kzalloc(nr_node_ids * sizeof(tbl[0]), GFP_KERNEL); |
bce90380 TH |
5468 | BUG_ON(!tbl); |
5469 | ||
5470 | for_each_node(node) | |
5a6024f1 | 5471 | BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL, |
1be0c25d | 5472 | node_online(node) ? node : NUMA_NO_NODE)); |
bce90380 TH |
5473 | |
5474 | for_each_possible_cpu(cpu) { | |
5475 | node = cpu_to_node(cpu); | |
5476 | if (WARN_ON(node == NUMA_NO_NODE)) { | |
5477 | pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu); | |
5478 | /* happens iff arch is bonkers, let's just proceed */ | |
5479 | return; | |
5480 | } | |
5481 | cpumask_set_cpu(cpu, tbl[node]); | |
5482 | } | |
5483 | ||
5484 | wq_numa_possible_cpumask = tbl; | |
5485 | wq_numa_enabled = true; | |
5486 | } | |
5487 | ||
3347fa09 TH |
5488 | /** |
5489 | * workqueue_init_early - early init for workqueue subsystem | |
5490 | * | |
5491 | * This is the first half of two-staged workqueue subsystem initialization | |
5492 | * and invoked as soon as the bare basics - memory allocation, cpumasks and | |
5493 | * idr are up. It sets up all the data structures and system workqueues | |
5494 | * and allows early boot code to create workqueues and queue/cancel work | |
5495 | * items. Actual work item execution starts only after kthreads can be | |
5496 | * created and scheduled right before early initcalls. | |
5497 | */ | |
5498 | int __init workqueue_init_early(void) | |
1da177e4 | 5499 | { |
7a4e344c TH |
5500 | int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; |
5501 | int i, cpu; | |
c34056a3 | 5502 | |
e904e6c2 TH |
5503 | WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); |
5504 | ||
b05a7928 FW |
5505 | BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL)); |
5506 | cpumask_copy(wq_unbound_cpumask, cpu_possible_mask); | |
5507 | ||
e904e6c2 TH |
5508 | pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); |
5509 | ||
706026c2 | 5510 | /* initialize CPU pools */ |
29c91e99 | 5511 | for_each_possible_cpu(cpu) { |
4ce62e9e | 5512 | struct worker_pool *pool; |
8b03ae3c | 5513 | |
7a4e344c | 5514 | i = 0; |
f02ae73a | 5515 | for_each_cpu_worker_pool(pool, cpu) { |
7a4e344c | 5516 | BUG_ON(init_worker_pool(pool)); |
ec22ca5e | 5517 | pool->cpu = cpu; |
29c91e99 | 5518 | cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); |
7a4e344c | 5519 | pool->attrs->nice = std_nice[i++]; |
f3f90ad4 | 5520 | pool->node = cpu_to_node(cpu); |
7a4e344c | 5521 | |
9daf9e67 | 5522 | /* alloc pool ID */ |
68e13a67 | 5523 | mutex_lock(&wq_pool_mutex); |
9daf9e67 | 5524 | BUG_ON(worker_pool_assign_id(pool)); |
68e13a67 | 5525 | mutex_unlock(&wq_pool_mutex); |
4ce62e9e | 5526 | } |
8b03ae3c TH |
5527 | } |
5528 | ||
8a2b7538 | 5529 | /* create default unbound and ordered wq attrs */ |
29c91e99 TH |
5530 | for (i = 0; i < NR_STD_WORKER_POOLS; i++) { |
5531 | struct workqueue_attrs *attrs; | |
5532 | ||
5533 | BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); | |
29c91e99 | 5534 | attrs->nice = std_nice[i]; |
29c91e99 | 5535 | unbound_std_wq_attrs[i] = attrs; |
8a2b7538 TH |
5536 | |
5537 | /* | |
5538 | * An ordered wq should have only one pwq as ordering is | |
5539 | * guaranteed by max_active which is enforced by pwqs. | |
5540 | * Turn off NUMA so that dfl_pwq is used for all nodes. | |
5541 | */ | |
5542 | BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); | |
5543 | attrs->nice = std_nice[i]; | |
5544 | attrs->no_numa = true; | |
5545 | ordered_wq_attrs[i] = attrs; | |
29c91e99 TH |
5546 | } |
5547 | ||
d320c038 | 5548 | system_wq = alloc_workqueue("events", 0, 0); |
1aabe902 | 5549 | system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); |
d320c038 | 5550 | system_long_wq = alloc_workqueue("events_long", 0, 0); |
f3421797 TH |
5551 | system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, |
5552 | WQ_UNBOUND_MAX_ACTIVE); | |
24d51add TH |
5553 | system_freezable_wq = alloc_workqueue("events_freezable", |
5554 | WQ_FREEZABLE, 0); | |
0668106c VK |
5555 | system_power_efficient_wq = alloc_workqueue("events_power_efficient", |
5556 | WQ_POWER_EFFICIENT, 0); | |
5557 | system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient", | |
5558 | WQ_FREEZABLE | WQ_POWER_EFFICIENT, | |
5559 | 0); | |
1aabe902 | 5560 | BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || |
0668106c VK |
5561 | !system_unbound_wq || !system_freezable_wq || |
5562 | !system_power_efficient_wq || | |
5563 | !system_freezable_power_efficient_wq); | |
82607adc | 5564 | |
3347fa09 TH |
5565 | return 0; |
5566 | } | |
5567 | ||
5568 | /** | |
5569 | * workqueue_init - bring workqueue subsystem fully online | |
5570 | * | |
5571 | * This is the latter half of two-staged workqueue subsystem initialization | |
5572 | * and invoked as soon as kthreads can be created and scheduled. | |
5573 | * Workqueues have been created and work items queued on them, but there | |
5574 | * are no kworkers executing the work items yet. Populate the worker pools | |
5575 | * with the initial workers and enable future kworker creations. | |
5576 | */ | |
5577 | int __init workqueue_init(void) | |
5578 | { | |
2186d9f9 | 5579 | struct workqueue_struct *wq; |
3347fa09 TH |
5580 | struct worker_pool *pool; |
5581 | int cpu, bkt; | |
5582 | ||
2186d9f9 TH |
5583 | /* |
5584 | * It'd be simpler to initialize NUMA in workqueue_init_early() but | |
5585 | * CPU to node mapping may not be available that early on some | |
5586 | * archs such as power and arm64. As per-cpu pools created | |
5587 | * previously could be missing node hint and unbound pools NUMA | |
5588 | * affinity, fix them up. | |
5589 | */ | |
5590 | wq_numa_init(); | |
5591 | ||
5592 | mutex_lock(&wq_pool_mutex); | |
5593 | ||
5594 | for_each_possible_cpu(cpu) { | |
5595 | for_each_cpu_worker_pool(pool, cpu) { | |
5596 | pool->node = cpu_to_node(cpu); | |
5597 | } | |
5598 | } | |
5599 | ||
5600 | list_for_each_entry(wq, &workqueues, list) | |
5601 | wq_update_unbound_numa(wq, smp_processor_id(), true); | |
5602 | ||
5603 | mutex_unlock(&wq_pool_mutex); | |
5604 | ||
3347fa09 TH |
5605 | /* create the initial workers */ |
5606 | for_each_online_cpu(cpu) { | |
5607 | for_each_cpu_worker_pool(pool, cpu) { | |
5608 | pool->flags &= ~POOL_DISASSOCIATED; | |
5609 | BUG_ON(!create_worker(pool)); | |
5610 | } | |
5611 | } | |
5612 | ||
5613 | hash_for_each(unbound_pool_hash, bkt, pool, hash_node) | |
5614 | BUG_ON(!create_worker(pool)); | |
5615 | ||
5616 | wq_online = true; | |
82607adc TH |
5617 | wq_watchdog_init(); |
5618 | ||
6ee0578b | 5619 | return 0; |
1da177e4 | 5620 | } |