Commit | Line | Data |
---|---|---|
1da177e4 | 1 | /* |
c54fce6e | 2 | * kernel/workqueue.c - generic async execution with shared worker pool |
1da177e4 | 3 | * |
c54fce6e | 4 | * Copyright (C) 2002 Ingo Molnar |
1da177e4 | 5 | * |
c54fce6e TH |
6 | * Derived from the taskqueue/keventd code by: |
7 | * David Woodhouse <dwmw2@infradead.org> | |
8 | * Andrew Morton | |
9 | * Kai Petzke <wpp@marie.physik.tu-berlin.de> | |
10 | * Theodore Ts'o <tytso@mit.edu> | |
1da177e4 | 11 | * |
c54fce6e | 12 | * Made to use alloc_percpu by Christoph Lameter. |
1da177e4 | 13 | * |
c54fce6e TH |
14 | * Copyright (C) 2010 SUSE Linux Products GmbH |
15 | * Copyright (C) 2010 Tejun Heo <tj@kernel.org> | |
89ada679 | 16 | * |
c54fce6e TH |
17 | * This is the generic async execution mechanism. Work items as are |
18 | * executed in process context. The worker pool is shared and | |
b11895c4 L |
19 | * automatically managed. There are two worker pools for each CPU (one for |
20 | * normal work items and the other for high priority ones) and some extra | |
21 | * pools for workqueues which are not bound to any specific CPU - the | |
22 | * number of these backing pools is dynamic. | |
c54fce6e | 23 | * |
9a261491 | 24 | * Please read Documentation/core-api/workqueue.rst for details. |
1da177e4 LT |
25 | */ |
26 | ||
9984de1a | 27 | #include <linux/export.h> |
1da177e4 LT |
28 | #include <linux/kernel.h> |
29 | #include <linux/sched.h> | |
30 | #include <linux/init.h> | |
31 | #include <linux/signal.h> | |
32 | #include <linux/completion.h> | |
33 | #include <linux/workqueue.h> | |
34 | #include <linux/slab.h> | |
35 | #include <linux/cpu.h> | |
36 | #include <linux/notifier.h> | |
37 | #include <linux/kthread.h> | |
1fa44eca | 38 | #include <linux/hardirq.h> |
46934023 | 39 | #include <linux/mempolicy.h> |
341a5958 | 40 | #include <linux/freezer.h> |
d5abe669 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> |
a6d5930c | 51 | #include <linux/nmi.h> |
dd101ca5 | 52 | #include <linux/debug-snapshot.h> |
e22bee78 | 53 | |
ea138446 | 54 | #include "workqueue_internal.h" |
1da177e4 | 55 | |
c8e55f36 | 56 | enum { |
24647570 TH |
57 | /* |
58 | * worker_pool flags | |
bc2ae0f5 | 59 | * |
24647570 | 60 | * A bound pool is either associated or disassociated with its CPU. |
bc2ae0f5 TH |
61 | * While associated (!DISASSOCIATED), all workers are bound to the |
62 | * CPU and none has %WORKER_UNBOUND set and concurrency management | |
63 | * is in effect. | |
64 | * | |
65 | * While DISASSOCIATED, the cpu may be offline and all workers have | |
66 | * %WORKER_UNBOUND set and concurrency management disabled, and may | |
24647570 | 67 | * be executing on any CPU. The pool behaves as an unbound one. |
bc2ae0f5 | 68 | * |
bc3a1afc | 69 | * Note that DISASSOCIATED should be flipped only while holding |
92f9c5c4 | 70 | * attach_mutex to avoid changing binding state while |
4736cbf7 | 71 | * worker_attach_to_pool() is in progress. |
bc2ae0f5 | 72 | */ |
692b4825 | 73 | POOL_MANAGER_ACTIVE = 1 << 0, /* being managed */ |
24647570 | 74 | POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */ |
db7bccf4 | 75 | |
c8e55f36 | 76 | /* worker flags */ |
c8e55f36 TH |
77 | WORKER_DIE = 1 << 1, /* die die die */ |
78 | WORKER_IDLE = 1 << 2, /* is idle */ | |
e22bee78 | 79 | WORKER_PREP = 1 << 3, /* preparing to run works */ |
fb0e7beb | 80 | WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */ |
f3421797 | 81 | WORKER_UNBOUND = 1 << 7, /* worker is unbound */ |
a9ab775b | 82 | WORKER_REBOUND = 1 << 8, /* worker was rebound */ |
e22bee78 | 83 | |
a9ab775b TH |
84 | WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE | |
85 | WORKER_UNBOUND | WORKER_REBOUND, | |
db7bccf4 | 86 | |
e34cdddb | 87 | NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */ |
4ce62e9e | 88 | |
29c91e99 | 89 | UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */ |
c8e55f36 | 90 | BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */ |
db7bccf4 | 91 | |
e22bee78 TH |
92 | MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */ |
93 | IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */ | |
94 | ||
3233cdbd TH |
95 | MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2, |
96 | /* call for help after 10ms | |
97 | (min two ticks) */ | |
e22bee78 TH |
98 | MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */ |
99 | CREATE_COOLDOWN = HZ, /* time to breath after fail */ | |
e22bee78 TH |
100 | |
101 | /* | |
102 | * Rescue workers are used only on emergencies and shared by | |
8698a745 | 103 | * all cpus. Give MIN_NICE. |
e22bee78 | 104 | */ |
8698a745 DY |
105 | RESCUER_NICE_LEVEL = MIN_NICE, |
106 | HIGHPRI_NICE_LEVEL = MIN_NICE, | |
ecf6881f TH |
107 | |
108 | WQ_NAME_LEN = 24, | |
c8e55f36 | 109 | }; |
1da177e4 LT |
110 | |
111 | /* | |
4690c4ab TH |
112 | * Structure fields follow one of the following exclusion rules. |
113 | * | |
e41e704b TH |
114 | * I: Modifiable by initialization/destruction paths and read-only for |
115 | * everyone else. | |
4690c4ab | 116 | * |
e22bee78 TH |
117 | * P: Preemption protected. Disabling preemption is enough and should |
118 | * only be modified and accessed from the local cpu. | |
119 | * | |
d565ed63 | 120 | * L: pool->lock protected. Access with pool->lock held. |
4690c4ab | 121 | * |
d565ed63 TH |
122 | * X: During normal operation, modification requires pool->lock and should |
123 | * be done only from local cpu. Either disabling preemption on local | |
124 | * cpu or grabbing pool->lock is enough for read access. If | |
125 | * POOL_DISASSOCIATED is set, it's identical to L. | |
e22bee78 | 126 | * |
92f9c5c4 | 127 | * A: pool->attach_mutex protected. |
822d8405 | 128 | * |
68e13a67 | 129 | * PL: wq_pool_mutex protected. |
5bcab335 | 130 | * |
68e13a67 | 131 | * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads. |
76af4d93 | 132 | * |
5b95e1af LJ |
133 | * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads. |
134 | * | |
135 | * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or | |
136 | * sched-RCU for reads. | |
137 | * | |
3c25a55d LJ |
138 | * WQ: wq->mutex protected. |
139 | * | |
b5927605 | 140 | * WR: wq->mutex protected for writes. Sched-RCU protected for reads. |
2e109a28 TH |
141 | * |
142 | * MD: wq_mayday_lock protected. | |
1da177e4 | 143 | */ |
1da177e4 | 144 | |
2eaebdb3 | 145 | /* struct worker is defined in workqueue_internal.h */ |
c34056a3 | 146 | |
bd7bdd43 | 147 | struct worker_pool { |
d565ed63 | 148 | spinlock_t lock; /* the pool lock */ |
d84ff051 | 149 | int cpu; /* I: the associated cpu */ |
f3f90ad4 | 150 | int node; /* I: the associated node ID */ |
9daf9e67 | 151 | int id; /* I: pool ID */ |
11ebea50 | 152 | unsigned int flags; /* X: flags */ |
bd7bdd43 | 153 | |
82607adc TH |
154 | unsigned long watchdog_ts; /* L: watchdog timestamp */ |
155 | ||
bd7bdd43 TH |
156 | struct list_head worklist; /* L: list of pending works */ |
157 | int nr_workers; /* L: total number of workers */ | |
ea1abd61 LJ |
158 | |
159 | /* nr_idle includes the ones off idle_list for rebinding */ | |
bd7bdd43 TH |
160 | int nr_idle; /* L: currently idle ones */ |
161 | ||
162 | struct list_head idle_list; /* X: list of idle workers */ | |
163 | struct timer_list idle_timer; /* L: worker idle timeout */ | |
164 | struct timer_list mayday_timer; /* L: SOS timer for workers */ | |
165 | ||
c5aa87bb | 166 | /* a workers is either on busy_hash or idle_list, or the manager */ |
c9e7cf27 TH |
167 | DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER); |
168 | /* L: hash of busy workers */ | |
169 | ||
bc3a1afc | 170 | /* see manage_workers() for details on the two manager mutexes */ |
2607d7a6 | 171 | struct worker *manager; /* L: purely informational */ |
92f9c5c4 LJ |
172 | struct mutex attach_mutex; /* attach/detach exclusion */ |
173 | struct list_head workers; /* A: attached workers */ | |
60f5a4bc | 174 | struct completion *detach_completion; /* all workers detached */ |
e19e397a | 175 | |
7cda9aae | 176 | struct ida worker_ida; /* worker IDs for task name */ |
e19e397a | 177 | |
7a4e344c | 178 | struct workqueue_attrs *attrs; /* I: worker attributes */ |
68e13a67 LJ |
179 | struct hlist_node hash_node; /* PL: unbound_pool_hash node */ |
180 | int refcnt; /* PL: refcnt for unbound pools */ | |
7a4e344c | 181 | |
e19e397a TH |
182 | /* |
183 | * The current concurrency level. As it's likely to be accessed | |
184 | * from other CPUs during try_to_wake_up(), put it in a separate | |
185 | * cacheline. | |
186 | */ | |
187 | atomic_t nr_running ____cacheline_aligned_in_smp; | |
29c91e99 TH |
188 | |
189 | /* | |
190 | * Destruction of pool is sched-RCU protected to allow dereferences | |
191 | * from get_work_pool(). | |
192 | */ | |
193 | struct rcu_head rcu; | |
8b03ae3c TH |
194 | } ____cacheline_aligned_in_smp; |
195 | ||
1da177e4 | 196 | /* |
112202d9 TH |
197 | * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS |
198 | * of work_struct->data are used for flags and the remaining high bits | |
199 | * point to the pwq; thus, pwqs need to be aligned at two's power of the | |
200 | * number of flag bits. | |
1da177e4 | 201 | */ |
112202d9 | 202 | struct pool_workqueue { |
bd7bdd43 | 203 | struct worker_pool *pool; /* I: the associated pool */ |
4690c4ab | 204 | struct workqueue_struct *wq; /* I: the owning workqueue */ |
73f53c4a TH |
205 | int work_color; /* L: current color */ |
206 | int flush_color; /* L: flushing color */ | |
8864b4e5 | 207 | int refcnt; /* L: reference count */ |
73f53c4a TH |
208 | int nr_in_flight[WORK_NR_COLORS]; |
209 | /* L: nr of in_flight works */ | |
1e19ffc6 | 210 | int nr_active; /* L: nr of active works */ |
a0a1a5fd | 211 | int max_active; /* L: max active works */ |
1e19ffc6 | 212 | struct list_head delayed_works; /* L: delayed works */ |
3c25a55d | 213 | struct list_head pwqs_node; /* WR: node on wq->pwqs */ |
2e109a28 | 214 | struct list_head mayday_node; /* MD: node on wq->maydays */ |
8864b4e5 TH |
215 | |
216 | /* | |
217 | * Release of unbound pwq is punted to system_wq. See put_pwq() | |
218 | * and pwq_unbound_release_workfn() for details. pool_workqueue | |
219 | * itself is also sched-RCU protected so that the first pwq can be | |
b09f4fd3 | 220 | * determined without grabbing wq->mutex. |
8864b4e5 TH |
221 | */ |
222 | struct work_struct unbound_release_work; | |
223 | struct rcu_head rcu; | |
e904e6c2 | 224 | } __aligned(1 << WORK_STRUCT_FLAG_BITS); |
1da177e4 | 225 | |
73f53c4a TH |
226 | /* |
227 | * Structure used to wait for workqueue flush. | |
228 | */ | |
229 | struct wq_flusher { | |
3c25a55d LJ |
230 | struct list_head list; /* WQ: list of flushers */ |
231 | int flush_color; /* WQ: flush color waiting for */ | |
73f53c4a TH |
232 | struct completion done; /* flush completion */ |
233 | }; | |
234 | ||
226223ab TH |
235 | struct wq_device; |
236 | ||
1da177e4 | 237 | /* |
c5aa87bb TH |
238 | * The externally visible workqueue. It relays the issued work items to |
239 | * the appropriate worker_pool through its pool_workqueues. | |
1da177e4 LT |
240 | */ |
241 | struct workqueue_struct { | |
3c25a55d | 242 | struct list_head pwqs; /* WR: all pwqs of this wq */ |
e2dca7ad | 243 | struct list_head list; /* PR: list of all workqueues */ |
73f53c4a | 244 | |
3c25a55d LJ |
245 | struct mutex mutex; /* protects this wq */ |
246 | int work_color; /* WQ: current work color */ | |
247 | int flush_color; /* WQ: current flush color */ | |
112202d9 | 248 | atomic_t nr_pwqs_to_flush; /* flush in progress */ |
3c25a55d LJ |
249 | struct wq_flusher *first_flusher; /* WQ: first flusher */ |
250 | struct list_head flusher_queue; /* WQ: flush waiters */ | |
251 | struct list_head flusher_overflow; /* WQ: flush overflow list */ | |
73f53c4a | 252 | |
2e109a28 | 253 | struct list_head maydays; /* MD: pwqs requesting rescue */ |
e22bee78 TH |
254 | struct worker *rescuer; /* I: rescue worker */ |
255 | ||
87fc741e | 256 | int nr_drainers; /* WQ: drain in progress */ |
a357fc03 | 257 | int saved_max_active; /* WQ: saved pwq max_active */ |
226223ab | 258 | |
5b95e1af LJ |
259 | struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */ |
260 | struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */ | |
6029a918 | 261 | |
226223ab TH |
262 | #ifdef CONFIG_SYSFS |
263 | struct wq_device *wq_dev; /* I: for sysfs interface */ | |
264 | #endif | |
4e6045f1 | 265 | #ifdef CONFIG_LOCKDEP |
4690c4ab | 266 | struct lockdep_map lockdep_map; |
4e6045f1 | 267 | #endif |
ecf6881f | 268 | char name[WQ_NAME_LEN]; /* I: workqueue name */ |
2728fd2f | 269 | |
e2dca7ad TH |
270 | /* |
271 | * Destruction of workqueue_struct is sched-RCU protected to allow | |
272 | * walking the workqueues list without grabbing wq_pool_mutex. | |
273 | * This is used to dump all workqueues from sysrq. | |
274 | */ | |
275 | struct rcu_head rcu; | |
276 | ||
2728fd2f TH |
277 | /* hot fields used during command issue, aligned to cacheline */ |
278 | unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */ | |
279 | struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */ | |
5b95e1af | 280 | struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */ |
1da177e4 LT |
281 | }; |
282 | ||
e904e6c2 TH |
283 | static struct kmem_cache *pwq_cache; |
284 | ||
bce90380 TH |
285 | static cpumask_var_t *wq_numa_possible_cpumask; |
286 | /* possible CPUs of each node */ | |
287 | ||
d55262c4 TH |
288 | static bool wq_disable_numa; |
289 | module_param_named(disable_numa, wq_disable_numa, bool, 0444); | |
290 | ||
cee22a15 | 291 | /* see the comment above the definition of WQ_POWER_EFFICIENT */ |
552f530c | 292 | static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT); |
cee22a15 VK |
293 | module_param_named(power_efficient, wq_power_efficient, bool, 0444); |
294 | ||
863b710b | 295 | static bool wq_online; /* can kworkers be created yet? */ |
3347fa09 | 296 | |
bce90380 TH |
297 | static bool wq_numa_enabled; /* unbound NUMA affinity enabled */ |
298 | ||
4c16bd32 TH |
299 | /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */ |
300 | static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf; | |
301 | ||
68e13a67 | 302 | static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ |
2e109a28 | 303 | static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ |
692b4825 | 304 | static DECLARE_WAIT_QUEUE_HEAD(wq_manager_wait); /* wait for manager to go away */ |
5bcab335 | 305 | |
e2dca7ad | 306 | static LIST_HEAD(workqueues); /* PR: list of all workqueues */ |
68e13a67 | 307 | static bool workqueue_freezing; /* PL: have wqs started freezing? */ |
7d19c5ce | 308 | |
ef557180 MG |
309 | /* PL: allowable cpus for unbound wqs and work items */ |
310 | static cpumask_var_t wq_unbound_cpumask; | |
311 | ||
312 | /* CPU where unbound work was last round robin scheduled from this CPU */ | |
313 | static DEFINE_PER_CPU(int, wq_rr_cpu_last); | |
b05a7928 | 314 | |
f303fccb TH |
315 | /* |
316 | * Local execution of unbound work items is no longer guaranteed. The | |
317 | * following always forces round-robin CPU selection on unbound work items | |
318 | * to uncover usages which depend on it. | |
319 | */ | |
320 | #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU | |
321 | static bool wq_debug_force_rr_cpu = true; | |
322 | #else | |
323 | static bool wq_debug_force_rr_cpu = false; | |
324 | #endif | |
325 | module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644); | |
326 | ||
7d19c5ce | 327 | /* the per-cpu worker pools */ |
25528213 | 328 | static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools); |
7d19c5ce | 329 | |
68e13a67 | 330 | static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ |
7d19c5ce | 331 | |
68e13a67 | 332 | /* PL: hash of all unbound pools keyed by pool->attrs */ |
29c91e99 TH |
333 | static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); |
334 | ||
c5aa87bb | 335 | /* I: attributes used when instantiating standard unbound pools on demand */ |
29c91e99 TH |
336 | static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; |
337 | ||
8a2b7538 TH |
338 | /* I: attributes used when instantiating ordered pools on demand */ |
339 | static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS]; | |
340 | ||
d320c038 | 341 | struct workqueue_struct *system_wq __read_mostly; |
ad7b1f84 | 342 | EXPORT_SYMBOL(system_wq); |
044c782c | 343 | struct workqueue_struct *system_highpri_wq __read_mostly; |
1aabe902 | 344 | EXPORT_SYMBOL_GPL(system_highpri_wq); |
044c782c | 345 | struct workqueue_struct *system_long_wq __read_mostly; |
d320c038 | 346 | EXPORT_SYMBOL_GPL(system_long_wq); |
044c782c | 347 | struct workqueue_struct *system_unbound_wq __read_mostly; |
f3421797 | 348 | EXPORT_SYMBOL_GPL(system_unbound_wq); |
044c782c | 349 | struct workqueue_struct *system_freezable_wq __read_mostly; |
24d51add | 350 | EXPORT_SYMBOL_GPL(system_freezable_wq); |
0668106c VK |
351 | struct workqueue_struct *system_power_efficient_wq __read_mostly; |
352 | EXPORT_SYMBOL_GPL(system_power_efficient_wq); | |
353 | struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly; | |
354 | EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq); | |
d320c038 | 355 | |
7d19c5ce | 356 | static int worker_thread(void *__worker); |
6ba94429 | 357 | static void workqueue_sysfs_unregister(struct workqueue_struct *wq); |
7d19c5ce | 358 | |
97bd2347 TH |
359 | #define CREATE_TRACE_POINTS |
360 | #include <trace/events/workqueue.h> | |
361 | ||
68e13a67 | 362 | #define assert_rcu_or_pool_mutex() \ |
f78f5b90 PM |
363 | RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \ |
364 | !lockdep_is_held(&wq_pool_mutex), \ | |
365 | "sched RCU or wq_pool_mutex should be held") | |
5bcab335 | 366 | |
b09f4fd3 | 367 | #define assert_rcu_or_wq_mutex(wq) \ |
f78f5b90 PM |
368 | RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \ |
369 | !lockdep_is_held(&wq->mutex), \ | |
370 | "sched RCU or wq->mutex should be held") | |
76af4d93 | 371 | |
5b95e1af | 372 | #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \ |
f78f5b90 PM |
373 | RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \ |
374 | !lockdep_is_held(&wq->mutex) && \ | |
375 | !lockdep_is_held(&wq_pool_mutex), \ | |
376 | "sched RCU, wq->mutex or wq_pool_mutex should be held") | |
5b95e1af | 377 | |
f02ae73a TH |
378 | #define for_each_cpu_worker_pool(pool, cpu) \ |
379 | for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ | |
380 | (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ | |
7a62c2c8 | 381 | (pool)++) |
4ce62e9e | 382 | |
17116969 TH |
383 | /** |
384 | * for_each_pool - iterate through all worker_pools in the system | |
385 | * @pool: iteration cursor | |
611c92a0 | 386 | * @pi: integer used for iteration |
fa1b54e6 | 387 | * |
68e13a67 LJ |
388 | * This must be called either with wq_pool_mutex held or sched RCU read |
389 | * locked. If the pool needs to be used beyond the locking in effect, the | |
390 | * caller is responsible for guaranteeing that the pool stays online. | |
fa1b54e6 TH |
391 | * |
392 | * The if/else clause exists only for the lockdep assertion and can be | |
393 | * ignored. | |
17116969 | 394 | */ |
611c92a0 TH |
395 | #define for_each_pool(pool, pi) \ |
396 | idr_for_each_entry(&worker_pool_idr, pool, pi) \ | |
68e13a67 | 397 | if (({ assert_rcu_or_pool_mutex(); false; })) { } \ |
fa1b54e6 | 398 | else |
17116969 | 399 | |
822d8405 TH |
400 | /** |
401 | * for_each_pool_worker - iterate through all workers of a worker_pool | |
402 | * @worker: iteration cursor | |
822d8405 TH |
403 | * @pool: worker_pool to iterate workers of |
404 | * | |
92f9c5c4 | 405 | * This must be called with @pool->attach_mutex. |
822d8405 TH |
406 | * |
407 | * The if/else clause exists only for the lockdep assertion and can be | |
408 | * ignored. | |
409 | */ | |
da028469 LJ |
410 | #define for_each_pool_worker(worker, pool) \ |
411 | list_for_each_entry((worker), &(pool)->workers, node) \ | |
92f9c5c4 | 412 | if (({ lockdep_assert_held(&pool->attach_mutex); false; })) { } \ |
822d8405 TH |
413 | else |
414 | ||
49e3cf44 TH |
415 | /** |
416 | * for_each_pwq - iterate through all pool_workqueues of the specified workqueue | |
417 | * @pwq: iteration cursor | |
418 | * @wq: the target workqueue | |
76af4d93 | 419 | * |
b09f4fd3 | 420 | * This must be called either with wq->mutex held or sched RCU read locked. |
794b18bc TH |
421 | * If the pwq needs to be used beyond the locking in effect, the caller is |
422 | * responsible for guaranteeing that the pwq stays online. | |
76af4d93 TH |
423 | * |
424 | * The if/else clause exists only for the lockdep assertion and can be | |
425 | * ignored. | |
49e3cf44 TH |
426 | */ |
427 | #define for_each_pwq(pwq, wq) \ | |
76af4d93 | 428 | list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \ |
b09f4fd3 | 429 | if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \ |
76af4d93 | 430 | else |
f3421797 | 431 | |
dc186ad7 TG |
432 | #ifdef CONFIG_DEBUG_OBJECTS_WORK |
433 | ||
434 | static struct debug_obj_descr work_debug_descr; | |
435 | ||
99777288 SG |
436 | static void *work_debug_hint(void *addr) |
437 | { | |
438 | return ((struct work_struct *) addr)->func; | |
439 | } | |
440 | ||
b9fdac7f DC |
441 | static bool work_is_static_object(void *addr) |
442 | { | |
443 | struct work_struct *work = addr; | |
444 | ||
445 | return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work)); | |
446 | } | |
447 | ||
dc186ad7 TG |
448 | /* |
449 | * fixup_init is called when: | |
450 | * - an active object is initialized | |
451 | */ | |
02a982a6 | 452 | static bool work_fixup_init(void *addr, enum debug_obj_state state) |
dc186ad7 TG |
453 | { |
454 | struct work_struct *work = addr; | |
455 | ||
456 | switch (state) { | |
457 | case ODEBUG_STATE_ACTIVE: | |
458 | cancel_work_sync(work); | |
459 | debug_object_init(work, &work_debug_descr); | |
02a982a6 | 460 | return true; |
dc186ad7 | 461 | default: |
02a982a6 | 462 | return false; |
dc186ad7 TG |
463 | } |
464 | } | |
465 | ||
dc186ad7 TG |
466 | /* |
467 | * fixup_free is called when: | |
468 | * - an active object is freed | |
469 | */ | |
02a982a6 | 470 | static bool work_fixup_free(void *addr, enum debug_obj_state state) |
dc186ad7 TG |
471 | { |
472 | struct work_struct *work = addr; | |
473 | ||
474 | switch (state) { | |
475 | case ODEBUG_STATE_ACTIVE: | |
476 | cancel_work_sync(work); | |
477 | debug_object_free(work, &work_debug_descr); | |
02a982a6 | 478 | return true; |
dc186ad7 | 479 | default: |
02a982a6 | 480 | return false; |
dc186ad7 TG |
481 | } |
482 | } | |
483 | ||
484 | static struct debug_obj_descr work_debug_descr = { | |
485 | .name = "work_struct", | |
99777288 | 486 | .debug_hint = work_debug_hint, |
b9fdac7f | 487 | .is_static_object = work_is_static_object, |
dc186ad7 | 488 | .fixup_init = work_fixup_init, |
dc186ad7 TG |
489 | .fixup_free = work_fixup_free, |
490 | }; | |
491 | ||
492 | static inline void debug_work_activate(struct work_struct *work) | |
493 | { | |
494 | debug_object_activate(work, &work_debug_descr); | |
495 | } | |
496 | ||
497 | static inline void debug_work_deactivate(struct work_struct *work) | |
498 | { | |
499 | debug_object_deactivate(work, &work_debug_descr); | |
500 | } | |
501 | ||
502 | void __init_work(struct work_struct *work, int onstack) | |
503 | { | |
504 | if (onstack) | |
505 | debug_object_init_on_stack(work, &work_debug_descr); | |
506 | else | |
507 | debug_object_init(work, &work_debug_descr); | |
508 | } | |
509 | EXPORT_SYMBOL_GPL(__init_work); | |
510 | ||
511 | void destroy_work_on_stack(struct work_struct *work) | |
512 | { | |
513 | debug_object_free(work, &work_debug_descr); | |
514 | } | |
515 | EXPORT_SYMBOL_GPL(destroy_work_on_stack); | |
516 | ||
ea2e64f2 TG |
517 | void destroy_delayed_work_on_stack(struct delayed_work *work) |
518 | { | |
519 | destroy_timer_on_stack(&work->timer); | |
520 | debug_object_free(&work->work, &work_debug_descr); | |
521 | } | |
522 | EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack); | |
523 | ||
dc186ad7 TG |
524 | #else |
525 | static inline void debug_work_activate(struct work_struct *work) { } | |
526 | static inline void debug_work_deactivate(struct work_struct *work) { } | |
527 | #endif | |
528 | ||
4e8b22bd LB |
529 | /** |
530 | * worker_pool_assign_id - allocate ID and assing it to @pool | |
531 | * @pool: the pool pointer of interest | |
532 | * | |
533 | * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned | |
534 | * successfully, -errno on failure. | |
535 | */ | |
9daf9e67 TH |
536 | static int worker_pool_assign_id(struct worker_pool *pool) |
537 | { | |
538 | int ret; | |
539 | ||
68e13a67 | 540 | lockdep_assert_held(&wq_pool_mutex); |
5bcab335 | 541 | |
4e8b22bd LB |
542 | ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE, |
543 | GFP_KERNEL); | |
229641a6 | 544 | if (ret >= 0) { |
e68035fb | 545 | pool->id = ret; |
229641a6 TH |
546 | return 0; |
547 | } | |
fa1b54e6 | 548 | return ret; |
7c3eed5c TH |
549 | } |
550 | ||
df2d5ae4 TH |
551 | /** |
552 | * unbound_pwq_by_node - return the unbound pool_workqueue for the given node | |
553 | * @wq: the target workqueue | |
554 | * @node: the node ID | |
555 | * | |
5b95e1af LJ |
556 | * This must be called with any of wq_pool_mutex, wq->mutex or sched RCU |
557 | * read locked. | |
df2d5ae4 TH |
558 | * If the pwq needs to be used beyond the locking in effect, the caller is |
559 | * responsible for guaranteeing that the pwq stays online. | |
d185af30 YB |
560 | * |
561 | * Return: The unbound pool_workqueue for @node. | |
df2d5ae4 TH |
562 | */ |
563 | static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq, | |
564 | int node) | |
565 | { | |
5b95e1af | 566 | assert_rcu_or_wq_mutex_or_pool_mutex(wq); |
d6e022f1 TH |
567 | |
568 | /* | |
569 | * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a | |
570 | * delayed item is pending. The plan is to keep CPU -> NODE | |
571 | * mapping valid and stable across CPU on/offlines. Once that | |
572 | * happens, this workaround can be removed. | |
573 | */ | |
574 | if (unlikely(node == NUMA_NO_NODE)) | |
575 | return wq->dfl_pwq; | |
576 | ||
df2d5ae4 TH |
577 | return rcu_dereference_raw(wq->numa_pwq_tbl[node]); |
578 | } | |
579 | ||
73f53c4a TH |
580 | static unsigned int work_color_to_flags(int color) |
581 | { | |
582 | return color << WORK_STRUCT_COLOR_SHIFT; | |
583 | } | |
584 | ||
585 | static int get_work_color(struct work_struct *work) | |
586 | { | |
587 | return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) & | |
588 | ((1 << WORK_STRUCT_COLOR_BITS) - 1); | |
589 | } | |
590 | ||
591 | static int work_next_color(int color) | |
592 | { | |
593 | return (color + 1) % WORK_NR_COLORS; | |
594 | } | |
1da177e4 | 595 | |
14441960 | 596 | /* |
112202d9 TH |
597 | * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data |
598 | * contain the pointer to the queued pwq. Once execution starts, the flag | |
7c3eed5c | 599 | * is cleared and the high bits contain OFFQ flags and pool ID. |
7a22ad75 | 600 | * |
112202d9 TH |
601 | * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() |
602 | * and clear_work_data() can be used to set the pwq, pool or clear | |
bbb68dfa TH |
603 | * work->data. These functions should only be called while the work is |
604 | * owned - ie. while the PENDING bit is set. | |
7a22ad75 | 605 | * |
112202d9 | 606 | * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq |
7c3eed5c | 607 | * corresponding to a work. Pool is available once the work has been |
112202d9 | 608 | * queued anywhere after initialization until it is sync canceled. pwq is |
7c3eed5c | 609 | * available only while the work item is queued. |
7a22ad75 | 610 | * |
bbb68dfa TH |
611 | * %WORK_OFFQ_CANCELING is used to mark a work item which is being |
612 | * canceled. While being canceled, a work item may have its PENDING set | |
613 | * but stay off timer and worklist for arbitrarily long and nobody should | |
614 | * try to steal the PENDING bit. | |
14441960 | 615 | */ |
7a22ad75 TH |
616 | static inline void set_work_data(struct work_struct *work, unsigned long data, |
617 | unsigned long flags) | |
365970a1 | 618 | { |
6183c009 | 619 | WARN_ON_ONCE(!work_pending(work)); |
7a22ad75 TH |
620 | atomic_long_set(&work->data, data | flags | work_static(work)); |
621 | } | |
365970a1 | 622 | |
112202d9 | 623 | static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, |
7a22ad75 TH |
624 | unsigned long extra_flags) |
625 | { | |
112202d9 TH |
626 | set_work_data(work, (unsigned long)pwq, |
627 | WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); | |
365970a1 DH |
628 | } |
629 | ||
4468a00f LJ |
630 | static void set_work_pool_and_keep_pending(struct work_struct *work, |
631 | int pool_id) | |
632 | { | |
633 | set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, | |
634 | WORK_STRUCT_PENDING); | |
635 | } | |
636 | ||
7c3eed5c TH |
637 | static void set_work_pool_and_clear_pending(struct work_struct *work, |
638 | int pool_id) | |
7a22ad75 | 639 | { |
23657bb1 TH |
640 | /* |
641 | * The following wmb is paired with the implied mb in | |
642 | * test_and_set_bit(PENDING) and ensures all updates to @work made | |
643 | * here are visible to and precede any updates by the next PENDING | |
644 | * owner. | |
645 | */ | |
646 | smp_wmb(); | |
7c3eed5c | 647 | set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); |
346c09f8 RP |
648 | /* |
649 | * The following mb guarantees that previous clear of a PENDING bit | |
650 | * will not be reordered with any speculative LOADS or STORES from | |
651 | * work->current_func, which is executed afterwards. This possible | |
652 | * reordering can lead to a missed execution on attempt to qeueue | |
653 | * the same @work. E.g. consider this case: | |
654 | * | |
655 | * CPU#0 CPU#1 | |
656 | * ---------------------------- -------------------------------- | |
657 | * | |
658 | * 1 STORE event_indicated | |
659 | * 2 queue_work_on() { | |
660 | * 3 test_and_set_bit(PENDING) | |
661 | * 4 } set_..._and_clear_pending() { | |
662 | * 5 set_work_data() # clear bit | |
663 | * 6 smp_mb() | |
664 | * 7 work->current_func() { | |
665 | * 8 LOAD event_indicated | |
666 | * } | |
667 | * | |
668 | * Without an explicit full barrier speculative LOAD on line 8 can | |
669 | * be executed before CPU#0 does STORE on line 1. If that happens, | |
670 | * CPU#0 observes the PENDING bit is still set and new execution of | |
671 | * a @work is not queued in a hope, that CPU#1 will eventually | |
672 | * finish the queued @work. Meanwhile CPU#1 does not see | |
673 | * event_indicated is set, because speculative LOAD was executed | |
674 | * before actual STORE. | |
675 | */ | |
676 | smp_mb(); | |
7a22ad75 | 677 | } |
f756d5e2 | 678 | |
7a22ad75 | 679 | static void clear_work_data(struct work_struct *work) |
1da177e4 | 680 | { |
7c3eed5c TH |
681 | smp_wmb(); /* see set_work_pool_and_clear_pending() */ |
682 | set_work_data(work, WORK_STRUCT_NO_POOL, 0); | |
1da177e4 LT |
683 | } |
684 | ||
112202d9 | 685 | static struct pool_workqueue *get_work_pwq(struct work_struct *work) |
b1f4ec17 | 686 | { |
e120153d | 687 | unsigned long data = atomic_long_read(&work->data); |
7a22ad75 | 688 | |
112202d9 | 689 | if (data & WORK_STRUCT_PWQ) |
e120153d TH |
690 | return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); |
691 | else | |
692 | return NULL; | |
4d707b9f ON |
693 | } |
694 | ||
7c3eed5c TH |
695 | /** |
696 | * get_work_pool - return the worker_pool a given work was associated with | |
697 | * @work: the work item of interest | |
698 | * | |
68e13a67 LJ |
699 | * Pools are created and destroyed under wq_pool_mutex, and allows read |
700 | * access under sched-RCU read lock. As such, this function should be | |
701 | * called under wq_pool_mutex or with preemption disabled. | |
fa1b54e6 TH |
702 | * |
703 | * All fields of the returned pool are accessible as long as the above | |
704 | * mentioned locking is in effect. If the returned pool needs to be used | |
705 | * beyond the critical section, the caller is responsible for ensuring the | |
706 | * returned pool is and stays online. | |
d185af30 YB |
707 | * |
708 | * Return: The worker_pool @work was last associated with. %NULL if none. | |
7c3eed5c TH |
709 | */ |
710 | static struct worker_pool *get_work_pool(struct work_struct *work) | |
365970a1 | 711 | { |
e120153d | 712 | unsigned long data = atomic_long_read(&work->data); |
7c3eed5c | 713 | int pool_id; |
7a22ad75 | 714 | |
68e13a67 | 715 | assert_rcu_or_pool_mutex(); |
fa1b54e6 | 716 | |
112202d9 TH |
717 | if (data & WORK_STRUCT_PWQ) |
718 | return ((struct pool_workqueue *) | |
7c3eed5c | 719 | (data & WORK_STRUCT_WQ_DATA_MASK))->pool; |
7a22ad75 | 720 | |
7c3eed5c TH |
721 | pool_id = data >> WORK_OFFQ_POOL_SHIFT; |
722 | if (pool_id == WORK_OFFQ_POOL_NONE) | |
7a22ad75 TH |
723 | return NULL; |
724 | ||
fa1b54e6 | 725 | return idr_find(&worker_pool_idr, pool_id); |
7c3eed5c TH |
726 | } |
727 | ||
728 | /** | |
729 | * get_work_pool_id - return the worker pool ID a given work is associated with | |
730 | * @work: the work item of interest | |
731 | * | |
d185af30 | 732 | * Return: The worker_pool ID @work was last associated with. |
7c3eed5c TH |
733 | * %WORK_OFFQ_POOL_NONE if none. |
734 | */ | |
735 | static int get_work_pool_id(struct work_struct *work) | |
736 | { | |
54d5b7d0 LJ |
737 | unsigned long data = atomic_long_read(&work->data); |
738 | ||
112202d9 TH |
739 | if (data & WORK_STRUCT_PWQ) |
740 | return ((struct pool_workqueue *) | |
54d5b7d0 | 741 | (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id; |
7c3eed5c | 742 | |
54d5b7d0 | 743 | return data >> WORK_OFFQ_POOL_SHIFT; |
7c3eed5c TH |
744 | } |
745 | ||
bbb68dfa TH |
746 | static void mark_work_canceling(struct work_struct *work) |
747 | { | |
7c3eed5c | 748 | unsigned long pool_id = get_work_pool_id(work); |
bbb68dfa | 749 | |
7c3eed5c TH |
750 | pool_id <<= WORK_OFFQ_POOL_SHIFT; |
751 | set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); | |
bbb68dfa TH |
752 | } |
753 | ||
754 | static bool work_is_canceling(struct work_struct *work) | |
755 | { | |
756 | unsigned long data = atomic_long_read(&work->data); | |
757 | ||
112202d9 | 758 | return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); |
bbb68dfa TH |
759 | } |
760 | ||
e22bee78 | 761 | /* |
3270476a TH |
762 | * Policy functions. These define the policies on how the global worker |
763 | * pools are managed. Unless noted otherwise, these functions assume that | |
d565ed63 | 764 | * they're being called with pool->lock held. |
e22bee78 TH |
765 | */ |
766 | ||
63d95a91 | 767 | static bool __need_more_worker(struct worker_pool *pool) |
a848e3b6 | 768 | { |
e19e397a | 769 | return !atomic_read(&pool->nr_running); |
a848e3b6 ON |
770 | } |
771 | ||
4594bf15 | 772 | /* |
e22bee78 TH |
773 | * Need to wake up a worker? Called from anything but currently |
774 | * running workers. | |
974271c4 TH |
775 | * |
776 | * Note that, because unbound workers never contribute to nr_running, this | |
706026c2 | 777 | * function will always return %true for unbound pools as long as the |
974271c4 | 778 | * worklist isn't empty. |
4594bf15 | 779 | */ |
63d95a91 | 780 | static bool need_more_worker(struct worker_pool *pool) |
365970a1 | 781 | { |
63d95a91 | 782 | return !list_empty(&pool->worklist) && __need_more_worker(pool); |
e22bee78 | 783 | } |
4594bf15 | 784 | |
e22bee78 | 785 | /* Can I start working? Called from busy but !running workers. */ |
63d95a91 | 786 | static bool may_start_working(struct worker_pool *pool) |
e22bee78 | 787 | { |
63d95a91 | 788 | return pool->nr_idle; |
e22bee78 TH |
789 | } |
790 | ||
791 | /* Do I need to keep working? Called from currently running workers. */ | |
63d95a91 | 792 | static bool keep_working(struct worker_pool *pool) |
e22bee78 | 793 | { |
e19e397a TH |
794 | return !list_empty(&pool->worklist) && |
795 | atomic_read(&pool->nr_running) <= 1; | |
e22bee78 TH |
796 | } |
797 | ||
798 | /* Do we need a new worker? Called from manager. */ | |
63d95a91 | 799 | static bool need_to_create_worker(struct worker_pool *pool) |
e22bee78 | 800 | { |
63d95a91 | 801 | return need_more_worker(pool) && !may_start_working(pool); |
e22bee78 | 802 | } |
365970a1 | 803 | |
e22bee78 | 804 | /* Do we have too many workers and should some go away? */ |
63d95a91 | 805 | static bool too_many_workers(struct worker_pool *pool) |
e22bee78 | 806 | { |
692b4825 | 807 | bool managing = pool->flags & POOL_MANAGER_ACTIVE; |
63d95a91 TH |
808 | int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ |
809 | int nr_busy = pool->nr_workers - nr_idle; | |
e22bee78 TH |
810 | |
811 | return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; | |
365970a1 DH |
812 | } |
813 | ||
4d707b9f | 814 | /* |
e22bee78 TH |
815 | * Wake up functions. |
816 | */ | |
817 | ||
1037de36 LJ |
818 | /* Return the first idle worker. Safe with preemption disabled */ |
819 | static struct worker *first_idle_worker(struct worker_pool *pool) | |
7e11629d | 820 | { |
63d95a91 | 821 | if (unlikely(list_empty(&pool->idle_list))) |
7e11629d TH |
822 | return NULL; |
823 | ||
63d95a91 | 824 | return list_first_entry(&pool->idle_list, struct worker, entry); |
7e11629d TH |
825 | } |
826 | ||
827 | /** | |
828 | * wake_up_worker - wake up an idle worker | |
63d95a91 | 829 | * @pool: worker pool to wake worker from |
7e11629d | 830 | * |
63d95a91 | 831 | * Wake up the first idle worker of @pool. |
7e11629d TH |
832 | * |
833 | * CONTEXT: | |
d565ed63 | 834 | * spin_lock_irq(pool->lock). |
7e11629d | 835 | */ |
63d95a91 | 836 | static void wake_up_worker(struct worker_pool *pool) |
7e11629d | 837 | { |
1037de36 | 838 | struct worker *worker = first_idle_worker(pool); |
7e11629d TH |
839 | |
840 | if (likely(worker)) | |
841 | wake_up_process(worker->task); | |
842 | } | |
843 | ||
d302f017 | 844 | /** |
e22bee78 TH |
845 | * wq_worker_waking_up - a worker is waking up |
846 | * @task: task waking up | |
847 | * @cpu: CPU @task is waking up to | |
848 | * | |
849 | * This function is called during try_to_wake_up() when a worker is | |
850 | * being awoken. | |
851 | * | |
852 | * CONTEXT: | |
853 | * spin_lock_irq(rq->lock) | |
854 | */ | |
d84ff051 | 855 | void wq_worker_waking_up(struct task_struct *task, int cpu) |
e22bee78 TH |
856 | { |
857 | struct worker *worker = kthread_data(task); | |
858 | ||
36576000 | 859 | if (!(worker->flags & WORKER_NOT_RUNNING)) { |
ec22ca5e | 860 | WARN_ON_ONCE(worker->pool->cpu != cpu); |
e19e397a | 861 | atomic_inc(&worker->pool->nr_running); |
36576000 | 862 | } |
e22bee78 TH |
863 | } |
864 | ||
865 | /** | |
866 | * wq_worker_sleeping - a worker is going to sleep | |
867 | * @task: task going to sleep | |
e22bee78 TH |
868 | * |
869 | * This function is called during schedule() when a busy worker is | |
870 | * going to sleep. Worker on the same cpu can be woken up by | |
871 | * returning pointer to its task. | |
872 | * | |
873 | * CONTEXT: | |
874 | * spin_lock_irq(rq->lock) | |
875 | * | |
d185af30 | 876 | * Return: |
e22bee78 TH |
877 | * Worker task on @cpu to wake up, %NULL if none. |
878 | */ | |
9b7f6597 | 879 | struct task_struct *wq_worker_sleeping(struct task_struct *task) |
e22bee78 TH |
880 | { |
881 | struct worker *worker = kthread_data(task), *to_wakeup = NULL; | |
111c225a | 882 | struct worker_pool *pool; |
e22bee78 | 883 | |
111c225a TH |
884 | /* |
885 | * Rescuers, which may not have all the fields set up like normal | |
886 | * workers, also reach here, let's not access anything before | |
887 | * checking NOT_RUNNING. | |
888 | */ | |
2d64672e | 889 | if (worker->flags & WORKER_NOT_RUNNING) |
e22bee78 TH |
890 | return NULL; |
891 | ||
111c225a | 892 | pool = worker->pool; |
111c225a | 893 | |
e22bee78 | 894 | /* this can only happen on the local cpu */ |
9b7f6597 | 895 | if (WARN_ON_ONCE(pool->cpu != raw_smp_processor_id())) |
6183c009 | 896 | return NULL; |
e22bee78 TH |
897 | |
898 | /* | |
899 | * The counterpart of the following dec_and_test, implied mb, | |
900 | * worklist not empty test sequence is in insert_work(). | |
901 | * Please read comment there. | |
902 | * | |
628c78e7 TH |
903 | * NOT_RUNNING is clear. This means that we're bound to and |
904 | * running on the local cpu w/ rq lock held and preemption | |
905 | * disabled, which in turn means that none else could be | |
d565ed63 | 906 | * manipulating idle_list, so dereferencing idle_list without pool |
628c78e7 | 907 | * lock is safe. |
e22bee78 | 908 | */ |
e19e397a TH |
909 | if (atomic_dec_and_test(&pool->nr_running) && |
910 | !list_empty(&pool->worklist)) | |
1037de36 | 911 | to_wakeup = first_idle_worker(pool); |
e22bee78 TH |
912 | return to_wakeup ? to_wakeup->task : NULL; |
913 | } | |
914 | ||
7ab429b6 JW |
915 | /** |
916 | * wq_worker_last_func - retrieve worker's last work function | |
917 | * | |
918 | * Determine the last function a worker executed. This is called from | |
919 | * the scheduler to get a worker's last known identity. | |
920 | * | |
921 | * CONTEXT: | |
922 | * spin_lock_irq(rq->lock) | |
923 | * | |
924 | * Return: | |
925 | * The last work function %current executed as a worker, NULL if it | |
926 | * hasn't executed any work yet. | |
927 | */ | |
928 | work_func_t wq_worker_last_func(struct task_struct *task) | |
929 | { | |
930 | struct worker *worker = kthread_data(task); | |
931 | ||
932 | return worker->last_func; | |
933 | } | |
934 | ||
e22bee78 TH |
935 | /** |
936 | * worker_set_flags - set worker flags and adjust nr_running accordingly | |
cb444766 | 937 | * @worker: self |
d302f017 | 938 | * @flags: flags to set |
d302f017 | 939 | * |
228f1d00 | 940 | * Set @flags in @worker->flags and adjust nr_running accordingly. |
d302f017 | 941 | * |
cb444766 | 942 | * CONTEXT: |
d565ed63 | 943 | * spin_lock_irq(pool->lock) |
d302f017 | 944 | */ |
228f1d00 | 945 | static inline void worker_set_flags(struct worker *worker, unsigned int flags) |
d302f017 | 946 | { |
bd7bdd43 | 947 | struct worker_pool *pool = worker->pool; |
e22bee78 | 948 | |
cb444766 TH |
949 | WARN_ON_ONCE(worker->task != current); |
950 | ||
228f1d00 | 951 | /* If transitioning into NOT_RUNNING, adjust nr_running. */ |
e22bee78 TH |
952 | if ((flags & WORKER_NOT_RUNNING) && |
953 | !(worker->flags & WORKER_NOT_RUNNING)) { | |
228f1d00 | 954 | atomic_dec(&pool->nr_running); |
e22bee78 TH |
955 | } |
956 | ||
d302f017 TH |
957 | worker->flags |= flags; |
958 | } | |
959 | ||
960 | /** | |
e22bee78 | 961 | * worker_clr_flags - clear worker flags and adjust nr_running accordingly |
cb444766 | 962 | * @worker: self |
d302f017 TH |
963 | * @flags: flags to clear |
964 | * | |
e22bee78 | 965 | * Clear @flags in @worker->flags and adjust nr_running accordingly. |
d302f017 | 966 | * |
cb444766 | 967 | * CONTEXT: |
d565ed63 | 968 | * spin_lock_irq(pool->lock) |
d302f017 TH |
969 | */ |
970 | static inline void worker_clr_flags(struct worker *worker, unsigned int flags) | |
971 | { | |
63d95a91 | 972 | struct worker_pool *pool = worker->pool; |
e22bee78 TH |
973 | unsigned int oflags = worker->flags; |
974 | ||
cb444766 TH |
975 | WARN_ON_ONCE(worker->task != current); |
976 | ||
d302f017 | 977 | worker->flags &= ~flags; |
e22bee78 | 978 | |
42c025f3 TH |
979 | /* |
980 | * If transitioning out of NOT_RUNNING, increment nr_running. Note | |
981 | * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask | |
982 | * of multiple flags, not a single flag. | |
983 | */ | |
e22bee78 TH |
984 | if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) |
985 | if (!(worker->flags & WORKER_NOT_RUNNING)) | |
e19e397a | 986 | atomic_inc(&pool->nr_running); |
d302f017 TH |
987 | } |
988 | ||
8cca0eea TH |
989 | /** |
990 | * find_worker_executing_work - find worker which is executing a work | |
c9e7cf27 | 991 | * @pool: pool of interest |
8cca0eea TH |
992 | * @work: work to find worker for |
993 | * | |
c9e7cf27 TH |
994 | * Find a worker which is executing @work on @pool by searching |
995 | * @pool->busy_hash which is keyed by the address of @work. For a worker | |
a2c1c57b TH |
996 | * to match, its current execution should match the address of @work and |
997 | * its work function. This is to avoid unwanted dependency between | |
998 | * unrelated work executions through a work item being recycled while still | |
999 | * being executed. | |
1000 | * | |
1001 | * This is a bit tricky. A work item may be freed once its execution | |
1002 | * starts and nothing prevents the freed area from being recycled for | |
1003 | * another work item. If the same work item address ends up being reused | |
1004 | * before the original execution finishes, workqueue will identify the | |
1005 | * recycled work item as currently executing and make it wait until the | |
1006 | * current execution finishes, introducing an unwanted dependency. | |
1007 | * | |
c5aa87bb TH |
1008 | * This function checks the work item address and work function to avoid |
1009 | * false positives. Note that this isn't complete as one may construct a | |
1010 | * work function which can introduce dependency onto itself through a | |
1011 | * recycled work item. Well, if somebody wants to shoot oneself in the | |
1012 | * foot that badly, there's only so much we can do, and if such deadlock | |
1013 | * actually occurs, it should be easy to locate the culprit work function. | |
8cca0eea TH |
1014 | * |
1015 | * CONTEXT: | |
d565ed63 | 1016 | * spin_lock_irq(pool->lock). |
8cca0eea | 1017 | * |
d185af30 YB |
1018 | * Return: |
1019 | * Pointer to worker which is executing @work if found, %NULL | |
8cca0eea | 1020 | * otherwise. |
4d707b9f | 1021 | */ |
c9e7cf27 | 1022 | static struct worker *find_worker_executing_work(struct worker_pool *pool, |
8cca0eea | 1023 | struct work_struct *work) |
4d707b9f | 1024 | { |
42f8570f | 1025 | struct worker *worker; |
42f8570f | 1026 | |
b67bfe0d | 1027 | hash_for_each_possible(pool->busy_hash, worker, hentry, |
a2c1c57b TH |
1028 | (unsigned long)work) |
1029 | if (worker->current_work == work && | |
1030 | worker->current_func == work->func) | |
42f8570f SL |
1031 | return worker; |
1032 | ||
1033 | return NULL; | |
4d707b9f ON |
1034 | } |
1035 | ||
bf4ede01 TH |
1036 | /** |
1037 | * move_linked_works - move linked works to a list | |
1038 | * @work: start of series of works to be scheduled | |
1039 | * @head: target list to append @work to | |
402dd89d | 1040 | * @nextp: out parameter for nested worklist walking |
bf4ede01 TH |
1041 | * |
1042 | * Schedule linked works starting from @work to @head. Work series to | |
1043 | * be scheduled starts at @work and includes any consecutive work with | |
1044 | * WORK_STRUCT_LINKED set in its predecessor. | |
1045 | * | |
1046 | * If @nextp is not NULL, it's updated to point to the next work of | |
1047 | * the last scheduled work. This allows move_linked_works() to be | |
1048 | * nested inside outer list_for_each_entry_safe(). | |
1049 | * | |
1050 | * CONTEXT: | |
d565ed63 | 1051 | * spin_lock_irq(pool->lock). |
bf4ede01 TH |
1052 | */ |
1053 | static void move_linked_works(struct work_struct *work, struct list_head *head, | |
1054 | struct work_struct **nextp) | |
1055 | { | |
1056 | struct work_struct *n; | |
1057 | ||
1058 | /* | |
1059 | * Linked worklist will always end before the end of the list, | |
1060 | * use NULL for list head. | |
1061 | */ | |
1062 | list_for_each_entry_safe_from(work, n, NULL, entry) { | |
1063 | list_move_tail(&work->entry, head); | |
1064 | if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) | |
1065 | break; | |
1066 | } | |
1067 | ||
1068 | /* | |
1069 | * If we're already inside safe list traversal and have moved | |
1070 | * multiple works to the scheduled queue, the next position | |
1071 | * needs to be updated. | |
1072 | */ | |
1073 | if (nextp) | |
1074 | *nextp = n; | |
1075 | } | |
1076 | ||
8864b4e5 TH |
1077 | /** |
1078 | * get_pwq - get an extra reference on the specified pool_workqueue | |
1079 | * @pwq: pool_workqueue to get | |
1080 | * | |
1081 | * Obtain an extra reference on @pwq. The caller should guarantee that | |
1082 | * @pwq has positive refcnt and be holding the matching pool->lock. | |
1083 | */ | |
1084 | static void get_pwq(struct pool_workqueue *pwq) | |
1085 | { | |
1086 | lockdep_assert_held(&pwq->pool->lock); | |
1087 | WARN_ON_ONCE(pwq->refcnt <= 0); | |
1088 | pwq->refcnt++; | |
1089 | } | |
1090 | ||
1091 | /** | |
1092 | * put_pwq - put a pool_workqueue reference | |
1093 | * @pwq: pool_workqueue to put | |
1094 | * | |
1095 | * Drop a reference of @pwq. If its refcnt reaches zero, schedule its | |
1096 | * destruction. The caller should be holding the matching pool->lock. | |
1097 | */ | |
1098 | static void put_pwq(struct pool_workqueue *pwq) | |
1099 | { | |
1100 | lockdep_assert_held(&pwq->pool->lock); | |
1101 | if (likely(--pwq->refcnt)) | |
1102 | return; | |
1103 | if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND))) | |
1104 | return; | |
1105 | /* | |
1106 | * @pwq can't be released under pool->lock, bounce to | |
1107 | * pwq_unbound_release_workfn(). This never recurses on the same | |
1108 | * pool->lock as this path is taken only for unbound workqueues and | |
1109 | * the release work item is scheduled on a per-cpu workqueue. To | |
1110 | * avoid lockdep warning, unbound pool->locks are given lockdep | |
1111 | * subclass of 1 in get_unbound_pool(). | |
1112 | */ | |
1113 | schedule_work(&pwq->unbound_release_work); | |
1114 | } | |
1115 | ||
dce90d47 TH |
1116 | /** |
1117 | * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock | |
1118 | * @pwq: pool_workqueue to put (can be %NULL) | |
1119 | * | |
1120 | * put_pwq() with locking. This function also allows %NULL @pwq. | |
1121 | */ | |
1122 | static void put_pwq_unlocked(struct pool_workqueue *pwq) | |
1123 | { | |
1124 | if (pwq) { | |
1125 | /* | |
1126 | * As both pwqs and pools are sched-RCU protected, the | |
1127 | * following lock operations are safe. | |
1128 | */ | |
1129 | spin_lock_irq(&pwq->pool->lock); | |
1130 | put_pwq(pwq); | |
1131 | spin_unlock_irq(&pwq->pool->lock); | |
1132 | } | |
1133 | } | |
1134 | ||
112202d9 | 1135 | static void pwq_activate_delayed_work(struct work_struct *work) |
bf4ede01 | 1136 | { |
112202d9 | 1137 | struct pool_workqueue *pwq = get_work_pwq(work); |
bf4ede01 TH |
1138 | |
1139 | trace_workqueue_activate_work(work); | |
82607adc TH |
1140 | if (list_empty(&pwq->pool->worklist)) |
1141 | pwq->pool->watchdog_ts = jiffies; | |
112202d9 | 1142 | move_linked_works(work, &pwq->pool->worklist, NULL); |
bf4ede01 | 1143 | __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work)); |
112202d9 | 1144 | pwq->nr_active++; |
bf4ede01 TH |
1145 | } |
1146 | ||
112202d9 | 1147 | static void pwq_activate_first_delayed(struct pool_workqueue *pwq) |
3aa62497 | 1148 | { |
112202d9 | 1149 | struct work_struct *work = list_first_entry(&pwq->delayed_works, |
3aa62497 LJ |
1150 | struct work_struct, entry); |
1151 | ||
112202d9 | 1152 | pwq_activate_delayed_work(work); |
3aa62497 LJ |
1153 | } |
1154 | ||
bf4ede01 | 1155 | /** |
112202d9 TH |
1156 | * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight |
1157 | * @pwq: pwq of interest | |
bf4ede01 | 1158 | * @color: color of work which left the queue |
bf4ede01 TH |
1159 | * |
1160 | * A work either has completed or is removed from pending queue, | |
112202d9 | 1161 | * decrement nr_in_flight of its pwq and handle workqueue flushing. |
bf4ede01 TH |
1162 | * |
1163 | * CONTEXT: | |
d565ed63 | 1164 | * spin_lock_irq(pool->lock). |
bf4ede01 | 1165 | */ |
112202d9 | 1166 | static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color) |
bf4ede01 | 1167 | { |
8864b4e5 | 1168 | /* uncolored work items don't participate in flushing or nr_active */ |
bf4ede01 | 1169 | if (color == WORK_NO_COLOR) |
8864b4e5 | 1170 | goto out_put; |
bf4ede01 | 1171 | |
112202d9 | 1172 | pwq->nr_in_flight[color]--; |
bf4ede01 | 1173 | |
112202d9 TH |
1174 | pwq->nr_active--; |
1175 | if (!list_empty(&pwq->delayed_works)) { | |
b3f9f405 | 1176 | /* one down, submit a delayed one */ |
112202d9 TH |
1177 | if (pwq->nr_active < pwq->max_active) |
1178 | pwq_activate_first_delayed(pwq); | |
bf4ede01 TH |
1179 | } |
1180 | ||
1181 | /* is flush in progress and are we at the flushing tip? */ | |
112202d9 | 1182 | if (likely(pwq->flush_color != color)) |
8864b4e5 | 1183 | goto out_put; |
bf4ede01 TH |
1184 | |
1185 | /* are there still in-flight works? */ | |
112202d9 | 1186 | if (pwq->nr_in_flight[color]) |
8864b4e5 | 1187 | goto out_put; |
bf4ede01 | 1188 | |
112202d9 TH |
1189 | /* this pwq is done, clear flush_color */ |
1190 | pwq->flush_color = -1; | |
bf4ede01 TH |
1191 | |
1192 | /* | |
112202d9 | 1193 | * If this was the last pwq, wake up the first flusher. It |
bf4ede01 TH |
1194 | * will handle the rest. |
1195 | */ | |
112202d9 TH |
1196 | if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) |
1197 | complete(&pwq->wq->first_flusher->done); | |
8864b4e5 TH |
1198 | out_put: |
1199 | put_pwq(pwq); | |
bf4ede01 TH |
1200 | } |
1201 | ||
36e227d2 | 1202 | /** |
bbb68dfa | 1203 | * try_to_grab_pending - steal work item from worklist and disable irq |
36e227d2 TH |
1204 | * @work: work item to steal |
1205 | * @is_dwork: @work is a delayed_work | |
bbb68dfa | 1206 | * @flags: place to store irq state |
36e227d2 TH |
1207 | * |
1208 | * Try to grab PENDING bit of @work. This function can handle @work in any | |
d185af30 | 1209 | * stable state - idle, on timer or on worklist. |
36e227d2 | 1210 | * |
d185af30 | 1211 | * Return: |
36e227d2 TH |
1212 | * 1 if @work was pending and we successfully stole PENDING |
1213 | * 0 if @work was idle and we claimed PENDING | |
1214 | * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry | |
bbb68dfa TH |
1215 | * -ENOENT if someone else is canceling @work, this state may persist |
1216 | * for arbitrarily long | |
36e227d2 | 1217 | * |
d185af30 | 1218 | * Note: |
bbb68dfa | 1219 | * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting |
e0aecdd8 TH |
1220 | * interrupted while holding PENDING and @work off queue, irq must be |
1221 | * disabled on entry. This, combined with delayed_work->timer being | |
1222 | * irqsafe, ensures that we return -EAGAIN for finite short period of time. | |
bbb68dfa TH |
1223 | * |
1224 | * On successful return, >= 0, irq is disabled and the caller is | |
1225 | * responsible for releasing it using local_irq_restore(*@flags). | |
1226 | * | |
e0aecdd8 | 1227 | * This function is safe to call from any context including IRQ handler. |
bf4ede01 | 1228 | */ |
bbb68dfa TH |
1229 | static int try_to_grab_pending(struct work_struct *work, bool is_dwork, |
1230 | unsigned long *flags) | |
bf4ede01 | 1231 | { |
d565ed63 | 1232 | struct worker_pool *pool; |
112202d9 | 1233 | struct pool_workqueue *pwq; |
bf4ede01 | 1234 | |
bbb68dfa TH |
1235 | local_irq_save(*flags); |
1236 | ||
36e227d2 TH |
1237 | /* try to steal the timer if it exists */ |
1238 | if (is_dwork) { | |
1239 | struct delayed_work *dwork = to_delayed_work(work); | |
1240 | ||
e0aecdd8 TH |
1241 | /* |
1242 | * dwork->timer is irqsafe. If del_timer() fails, it's | |
1243 | * guaranteed that the timer is not queued anywhere and not | |
1244 | * running on the local CPU. | |
1245 | */ | |
36e227d2 TH |
1246 | if (likely(del_timer(&dwork->timer))) |
1247 | return 1; | |
1248 | } | |
1249 | ||
1250 | /* try to claim PENDING the normal way */ | |
bf4ede01 TH |
1251 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) |
1252 | return 0; | |
1253 | ||
1254 | /* | |
1255 | * The queueing is in progress, or it is already queued. Try to | |
1256 | * steal it from ->worklist without clearing WORK_STRUCT_PENDING. | |
1257 | */ | |
d565ed63 TH |
1258 | pool = get_work_pool(work); |
1259 | if (!pool) | |
bbb68dfa | 1260 | goto fail; |
bf4ede01 | 1261 | |
d565ed63 | 1262 | spin_lock(&pool->lock); |
0b3dae68 | 1263 | /* |
112202d9 TH |
1264 | * work->data is guaranteed to point to pwq only while the work |
1265 | * item is queued on pwq->wq, and both updating work->data to point | |
1266 | * to pwq on queueing and to pool on dequeueing are done under | |
1267 | * pwq->pool->lock. This in turn guarantees that, if work->data | |
1268 | * points to pwq which is associated with a locked pool, the work | |
0b3dae68 LJ |
1269 | * item is currently queued on that pool. |
1270 | */ | |
112202d9 TH |
1271 | pwq = get_work_pwq(work); |
1272 | if (pwq && pwq->pool == pool) { | |
16062836 TH |
1273 | debug_work_deactivate(work); |
1274 | ||
1275 | /* | |
1276 | * A delayed work item cannot be grabbed directly because | |
1277 | * it might have linked NO_COLOR work items which, if left | |
112202d9 | 1278 | * on the delayed_list, will confuse pwq->nr_active |
16062836 TH |
1279 | * management later on and cause stall. Make sure the work |
1280 | * item is activated before grabbing. | |
1281 | */ | |
1282 | if (*work_data_bits(work) & WORK_STRUCT_DELAYED) | |
112202d9 | 1283 | pwq_activate_delayed_work(work); |
16062836 TH |
1284 | |
1285 | list_del_init(&work->entry); | |
9c34a704 | 1286 | pwq_dec_nr_in_flight(pwq, get_work_color(work)); |
16062836 | 1287 | |
112202d9 | 1288 | /* work->data points to pwq iff queued, point to pool */ |
16062836 TH |
1289 | set_work_pool_and_keep_pending(work, pool->id); |
1290 | ||
1291 | spin_unlock(&pool->lock); | |
1292 | return 1; | |
bf4ede01 | 1293 | } |
d565ed63 | 1294 | spin_unlock(&pool->lock); |
bbb68dfa TH |
1295 | fail: |
1296 | local_irq_restore(*flags); | |
1297 | if (work_is_canceling(work)) | |
1298 | return -ENOENT; | |
1299 | cpu_relax(); | |
36e227d2 | 1300 | return -EAGAIN; |
bf4ede01 TH |
1301 | } |
1302 | ||
4690c4ab | 1303 | /** |
706026c2 | 1304 | * insert_work - insert a work into a pool |
112202d9 | 1305 | * @pwq: pwq @work belongs to |
4690c4ab TH |
1306 | * @work: work to insert |
1307 | * @head: insertion point | |
1308 | * @extra_flags: extra WORK_STRUCT_* flags to set | |
1309 | * | |
112202d9 | 1310 | * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to |
706026c2 | 1311 | * work_struct flags. |
4690c4ab TH |
1312 | * |
1313 | * CONTEXT: | |
d565ed63 | 1314 | * spin_lock_irq(pool->lock). |
4690c4ab | 1315 | */ |
112202d9 TH |
1316 | static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, |
1317 | struct list_head *head, unsigned int extra_flags) | |
b89deed3 | 1318 | { |
112202d9 | 1319 | struct worker_pool *pool = pwq->pool; |
e22bee78 | 1320 | |
4690c4ab | 1321 | /* we own @work, set data and link */ |
112202d9 | 1322 | set_work_pwq(work, pwq, extra_flags); |
1a4d9b0a | 1323 | list_add_tail(&work->entry, head); |
8864b4e5 | 1324 | get_pwq(pwq); |
e22bee78 TH |
1325 | |
1326 | /* | |
c5aa87bb TH |
1327 | * Ensure either wq_worker_sleeping() sees the above |
1328 | * list_add_tail() or we see zero nr_running to avoid workers lying | |
1329 | * around lazily while there are works to be processed. | |
e22bee78 TH |
1330 | */ |
1331 | smp_mb(); | |
1332 | ||
63d95a91 TH |
1333 | if (__need_more_worker(pool)) |
1334 | wake_up_worker(pool); | |
b89deed3 ON |
1335 | } |
1336 | ||
c8efcc25 TH |
1337 | /* |
1338 | * Test whether @work is being queued from another work executing on the | |
8d03ecfe | 1339 | * same workqueue. |
c8efcc25 TH |
1340 | */ |
1341 | static bool is_chained_work(struct workqueue_struct *wq) | |
1342 | { | |
8d03ecfe TH |
1343 | struct worker *worker; |
1344 | ||
1345 | worker = current_wq_worker(); | |
1346 | /* | |
1347 | * Return %true iff I'm a worker execuing a work item on @wq. If | |
1348 | * I'm @worker, it's safe to dereference it without locking. | |
1349 | */ | |
112202d9 | 1350 | return worker && worker->current_pwq->wq == wq; |
c8efcc25 TH |
1351 | } |
1352 | ||
ef557180 MG |
1353 | /* |
1354 | * When queueing an unbound work item to a wq, prefer local CPU if allowed | |
1355 | * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to | |
1356 | * avoid perturbing sensitive tasks. | |
1357 | */ | |
1358 | static int wq_select_unbound_cpu(int cpu) | |
1359 | { | |
f303fccb | 1360 | static bool printed_dbg_warning; |
ef557180 MG |
1361 | int new_cpu; |
1362 | ||
f303fccb TH |
1363 | if (likely(!wq_debug_force_rr_cpu)) { |
1364 | if (cpumask_test_cpu(cpu, wq_unbound_cpumask)) | |
1365 | return cpu; | |
1366 | } else if (!printed_dbg_warning) { | |
1367 | pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n"); | |
1368 | printed_dbg_warning = true; | |
1369 | } | |
1370 | ||
ef557180 MG |
1371 | if (cpumask_empty(wq_unbound_cpumask)) |
1372 | return cpu; | |
1373 | ||
1374 | new_cpu = __this_cpu_read(wq_rr_cpu_last); | |
1375 | new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask); | |
1376 | if (unlikely(new_cpu >= nr_cpu_ids)) { | |
1377 | new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask); | |
1378 | if (unlikely(new_cpu >= nr_cpu_ids)) | |
1379 | return cpu; | |
1380 | } | |
1381 | __this_cpu_write(wq_rr_cpu_last, new_cpu); | |
1382 | ||
1383 | return new_cpu; | |
1384 | } | |
1385 | ||
d84ff051 | 1386 | static void __queue_work(int cpu, struct workqueue_struct *wq, |
1da177e4 LT |
1387 | struct work_struct *work) |
1388 | { | |
112202d9 | 1389 | struct pool_workqueue *pwq; |
c9178087 | 1390 | struct worker_pool *last_pool; |
1e19ffc6 | 1391 | struct list_head *worklist; |
8a2e8e5d | 1392 | unsigned int work_flags; |
b75cac93 | 1393 | unsigned int req_cpu = cpu; |
8930caba TH |
1394 | |
1395 | /* | |
1396 | * While a work item is PENDING && off queue, a task trying to | |
1397 | * steal the PENDING will busy-loop waiting for it to either get | |
1398 | * queued or lose PENDING. Grabbing PENDING and queueing should | |
1399 | * happen with IRQ disabled. | |
1400 | */ | |
1401 | WARN_ON_ONCE(!irqs_disabled()); | |
1da177e4 | 1402 | |
dc186ad7 | 1403 | debug_work_activate(work); |
1e19ffc6 | 1404 | |
9ef28a73 | 1405 | /* if draining, only works from the same workqueue are allowed */ |
618b01eb | 1406 | if (unlikely(wq->flags & __WQ_DRAINING) && |
c8efcc25 | 1407 | WARN_ON_ONCE(!is_chained_work(wq))) |
e41e704b | 1408 | return; |
9e8cd2f5 | 1409 | retry: |
df2d5ae4 | 1410 | if (req_cpu == WORK_CPU_UNBOUND) |
ef557180 | 1411 | cpu = wq_select_unbound_cpu(raw_smp_processor_id()); |
df2d5ae4 | 1412 | |
c9178087 | 1413 | /* pwq which will be used unless @work is executing elsewhere */ |
df2d5ae4 | 1414 | if (!(wq->flags & WQ_UNBOUND)) |
7fb98ea7 | 1415 | pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); |
df2d5ae4 TH |
1416 | else |
1417 | pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); | |
dbf2576e | 1418 | |
c9178087 TH |
1419 | /* |
1420 | * If @work was previously on a different pool, it might still be | |
1421 | * running there, in which case the work needs to be queued on that | |
1422 | * pool to guarantee non-reentrancy. | |
1423 | */ | |
1424 | last_pool = get_work_pool(work); | |
1425 | if (last_pool && last_pool != pwq->pool) { | |
1426 | struct worker *worker; | |
18aa9eff | 1427 | |
c9178087 | 1428 | spin_lock(&last_pool->lock); |
18aa9eff | 1429 | |
c9178087 | 1430 | worker = find_worker_executing_work(last_pool, work); |
18aa9eff | 1431 | |
c9178087 TH |
1432 | if (worker && worker->current_pwq->wq == wq) { |
1433 | pwq = worker->current_pwq; | |
8930caba | 1434 | } else { |
c9178087 TH |
1435 | /* meh... not running there, queue here */ |
1436 | spin_unlock(&last_pool->lock); | |
112202d9 | 1437 | spin_lock(&pwq->pool->lock); |
8930caba | 1438 | } |
f3421797 | 1439 | } else { |
112202d9 | 1440 | spin_lock(&pwq->pool->lock); |
502ca9d8 TH |
1441 | } |
1442 | ||
9e8cd2f5 TH |
1443 | /* |
1444 | * pwq is determined and locked. For unbound pools, we could have | |
1445 | * raced with pwq release and it could already be dead. If its | |
1446 | * refcnt is zero, repeat pwq selection. Note that pwqs never die | |
df2d5ae4 TH |
1447 | * without another pwq replacing it in the numa_pwq_tbl or while |
1448 | * work items are executing on it, so the retrying is guaranteed to | |
9e8cd2f5 TH |
1449 | * make forward-progress. |
1450 | */ | |
1451 | if (unlikely(!pwq->refcnt)) { | |
1452 | if (wq->flags & WQ_UNBOUND) { | |
1453 | spin_unlock(&pwq->pool->lock); | |
1454 | cpu_relax(); | |
1455 | goto retry; | |
1456 | } | |
1457 | /* oops */ | |
1458 | WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", | |
1459 | wq->name, cpu); | |
1460 | } | |
1461 | ||
112202d9 TH |
1462 | /* pwq determined, queue */ |
1463 | trace_workqueue_queue_work(req_cpu, pwq, work); | |
502ca9d8 | 1464 | |
f5b2552b | 1465 | if (WARN_ON(!list_empty(&work->entry))) { |
112202d9 | 1466 | spin_unlock(&pwq->pool->lock); |
f5b2552b DC |
1467 | return; |
1468 | } | |
1e19ffc6 | 1469 | |
112202d9 TH |
1470 | pwq->nr_in_flight[pwq->work_color]++; |
1471 | work_flags = work_color_to_flags(pwq->work_color); | |
1e19ffc6 | 1472 | |
112202d9 | 1473 | if (likely(pwq->nr_active < pwq->max_active)) { |
cdadf009 | 1474 | trace_workqueue_activate_work(work); |
112202d9 TH |
1475 | pwq->nr_active++; |
1476 | worklist = &pwq->pool->worklist; | |
82607adc TH |
1477 | if (list_empty(worklist)) |
1478 | pwq->pool->watchdog_ts = jiffies; | |
8a2e8e5d TH |
1479 | } else { |
1480 | work_flags |= WORK_STRUCT_DELAYED; | |
112202d9 | 1481 | worklist = &pwq->delayed_works; |
8a2e8e5d | 1482 | } |
1e19ffc6 | 1483 | |
112202d9 | 1484 | insert_work(pwq, work, worklist, work_flags); |
1e19ffc6 | 1485 | |
112202d9 | 1486 | spin_unlock(&pwq->pool->lock); |
1da177e4 LT |
1487 | } |
1488 | ||
0fcb78c2 | 1489 | /** |
c1a220e7 ZR |
1490 | * queue_work_on - queue work on specific cpu |
1491 | * @cpu: CPU number to execute work on | |
0fcb78c2 REB |
1492 | * @wq: workqueue to use |
1493 | * @work: work to queue | |
1494 | * | |
c1a220e7 ZR |
1495 | * We queue the work to a specific CPU, the caller must ensure it |
1496 | * can't go away. | |
d185af30 YB |
1497 | * |
1498 | * Return: %false if @work was already on a queue, %true otherwise. | |
1da177e4 | 1499 | */ |
d4283e93 TH |
1500 | bool queue_work_on(int cpu, struct workqueue_struct *wq, |
1501 | struct work_struct *work) | |
1da177e4 | 1502 | { |
d4283e93 | 1503 | bool ret = false; |
8930caba | 1504 | unsigned long flags; |
ef1ca236 | 1505 | |
8930caba | 1506 | local_irq_save(flags); |
c1a220e7 | 1507 | |
22df02bb | 1508 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { |
4690c4ab | 1509 | __queue_work(cpu, wq, work); |
d4283e93 | 1510 | ret = true; |
c1a220e7 | 1511 | } |
ef1ca236 | 1512 | |
8930caba | 1513 | local_irq_restore(flags); |
1da177e4 LT |
1514 | return ret; |
1515 | } | |
ad7b1f84 | 1516 | EXPORT_SYMBOL(queue_work_on); |
1da177e4 | 1517 | |
d8e794df | 1518 | void delayed_work_timer_fn(unsigned long __data) |
1da177e4 | 1519 | { |
52bad64d | 1520 | struct delayed_work *dwork = (struct delayed_work *)__data; |
1da177e4 | 1521 | |
e0aecdd8 | 1522 | /* should have been called from irqsafe timer with irq already off */ |
60c057bc | 1523 | __queue_work(dwork->cpu, dwork->wq, &dwork->work); |
1da177e4 | 1524 | } |
1438ade5 | 1525 | EXPORT_SYMBOL(delayed_work_timer_fn); |
1da177e4 | 1526 | |
7beb2edf TH |
1527 | static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, |
1528 | struct delayed_work *dwork, unsigned long delay) | |
1da177e4 | 1529 | { |
7beb2edf TH |
1530 | struct timer_list *timer = &dwork->timer; |
1531 | struct work_struct *work = &dwork->work; | |
7beb2edf | 1532 | |
637fdbae | 1533 | WARN_ON_ONCE(!wq); |
7beb2edf TH |
1534 | WARN_ON_ONCE(timer->function != delayed_work_timer_fn || |
1535 | timer->data != (unsigned long)dwork); | |
fc4b514f TH |
1536 | WARN_ON_ONCE(timer_pending(timer)); |
1537 | WARN_ON_ONCE(!list_empty(&work->entry)); | |
7beb2edf | 1538 | |
8852aac2 TH |
1539 | /* |
1540 | * If @delay is 0, queue @dwork->work immediately. This is for | |
1541 | * both optimization and correctness. The earliest @timer can | |
1542 | * expire is on the closest next tick and delayed_work users depend | |
1543 | * on that there's no such delay when @delay is 0. | |
1544 | */ | |
1545 | if (!delay) { | |
1546 | __queue_work(cpu, wq, &dwork->work); | |
1547 | return; | |
1548 | } | |
1549 | ||
60c057bc | 1550 | dwork->wq = wq; |
1265057f | 1551 | dwork->cpu = cpu; |
7beb2edf TH |
1552 | timer->expires = jiffies + delay; |
1553 | ||
041bd12e TH |
1554 | if (unlikely(cpu != WORK_CPU_UNBOUND)) |
1555 | add_timer_on(timer, cpu); | |
1556 | else | |
1557 | add_timer(timer); | |
1da177e4 LT |
1558 | } |
1559 | ||
0fcb78c2 REB |
1560 | /** |
1561 | * queue_delayed_work_on - queue work on specific CPU after delay | |
1562 | * @cpu: CPU number to execute work on | |
1563 | * @wq: workqueue to use | |
af9997e4 | 1564 | * @dwork: work to queue |
0fcb78c2 REB |
1565 | * @delay: number of jiffies to wait before queueing |
1566 | * | |
d185af30 | 1567 | * Return: %false if @work was already on a queue, %true otherwise. If |
715f1300 TH |
1568 | * @delay is zero and @dwork is idle, it will be scheduled for immediate |
1569 | * execution. | |
0fcb78c2 | 1570 | */ |
d4283e93 TH |
1571 | bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, |
1572 | struct delayed_work *dwork, unsigned long delay) | |
7a6bc1cd | 1573 | { |
52bad64d | 1574 | struct work_struct *work = &dwork->work; |
d4283e93 | 1575 | bool ret = false; |
8930caba | 1576 | unsigned long flags; |
7a6bc1cd | 1577 | |
8930caba TH |
1578 | /* read the comment in __queue_work() */ |
1579 | local_irq_save(flags); | |
7a6bc1cd | 1580 | |
22df02bb | 1581 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { |
7beb2edf | 1582 | __queue_delayed_work(cpu, wq, dwork, delay); |
d4283e93 | 1583 | ret = true; |
7a6bc1cd | 1584 | } |
8a3e77cc | 1585 | |
8930caba | 1586 | local_irq_restore(flags); |
7a6bc1cd VP |
1587 | return ret; |
1588 | } | |
ad7b1f84 | 1589 | EXPORT_SYMBOL(queue_delayed_work_on); |
c7fc77f7 | 1590 | |
8376fe22 TH |
1591 | /** |
1592 | * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU | |
1593 | * @cpu: CPU number to execute work on | |
1594 | * @wq: workqueue to use | |
1595 | * @dwork: work to queue | |
1596 | * @delay: number of jiffies to wait before queueing | |
1597 | * | |
1598 | * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, | |
1599 | * modify @dwork's timer so that it expires after @delay. If @delay is | |
1600 | * zero, @work is guaranteed to be scheduled immediately regardless of its | |
1601 | * current state. | |
1602 | * | |
d185af30 | 1603 | * Return: %false if @dwork was idle and queued, %true if @dwork was |
8376fe22 TH |
1604 | * pending and its timer was modified. |
1605 | * | |
e0aecdd8 | 1606 | * This function is safe to call from any context including IRQ handler. |
8376fe22 TH |
1607 | * See try_to_grab_pending() for details. |
1608 | */ | |
1609 | bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, | |
1610 | struct delayed_work *dwork, unsigned long delay) | |
1611 | { | |
1612 | unsigned long flags; | |
1613 | int ret; | |
c7fc77f7 | 1614 | |
8376fe22 TH |
1615 | do { |
1616 | ret = try_to_grab_pending(&dwork->work, true, &flags); | |
1617 | } while (unlikely(ret == -EAGAIN)); | |
63bc0362 | 1618 | |
8376fe22 TH |
1619 | if (likely(ret >= 0)) { |
1620 | __queue_delayed_work(cpu, wq, dwork, delay); | |
1621 | local_irq_restore(flags); | |
7a6bc1cd | 1622 | } |
8376fe22 TH |
1623 | |
1624 | /* -ENOENT from try_to_grab_pending() becomes %true */ | |
7a6bc1cd VP |
1625 | return ret; |
1626 | } | |
8376fe22 TH |
1627 | EXPORT_SYMBOL_GPL(mod_delayed_work_on); |
1628 | ||
c8e55f36 TH |
1629 | /** |
1630 | * worker_enter_idle - enter idle state | |
1631 | * @worker: worker which is entering idle state | |
1632 | * | |
1633 | * @worker is entering idle state. Update stats and idle timer if | |
1634 | * necessary. | |
1635 | * | |
1636 | * LOCKING: | |
d565ed63 | 1637 | * spin_lock_irq(pool->lock). |
c8e55f36 TH |
1638 | */ |
1639 | static void worker_enter_idle(struct worker *worker) | |
1da177e4 | 1640 | { |
bd7bdd43 | 1641 | struct worker_pool *pool = worker->pool; |
c8e55f36 | 1642 | |
6183c009 TH |
1643 | if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || |
1644 | WARN_ON_ONCE(!list_empty(&worker->entry) && | |
1645 | (worker->hentry.next || worker->hentry.pprev))) | |
1646 | return; | |
c8e55f36 | 1647 | |
051e1850 | 1648 | /* can't use worker_set_flags(), also called from create_worker() */ |
cb444766 | 1649 | worker->flags |= WORKER_IDLE; |
bd7bdd43 | 1650 | pool->nr_idle++; |
e22bee78 | 1651 | worker->last_active = jiffies; |
c8e55f36 TH |
1652 | |
1653 | /* idle_list is LIFO */ | |
bd7bdd43 | 1654 | list_add(&worker->entry, &pool->idle_list); |
db7bccf4 | 1655 | |
628c78e7 TH |
1656 | if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) |
1657 | mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); | |
cb444766 | 1658 | |
544ecf31 | 1659 | /* |
706026c2 | 1660 | * Sanity check nr_running. Because wq_unbind_fn() releases |
d565ed63 | 1661 | * pool->lock between setting %WORKER_UNBOUND and zapping |
628c78e7 TH |
1662 | * nr_running, the warning may trigger spuriously. Check iff |
1663 | * unbind is not in progress. | |
544ecf31 | 1664 | */ |
24647570 | 1665 | WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && |
bd7bdd43 | 1666 | pool->nr_workers == pool->nr_idle && |
e19e397a | 1667 | atomic_read(&pool->nr_running)); |
c8e55f36 TH |
1668 | } |
1669 | ||
1670 | /** | |
1671 | * worker_leave_idle - leave idle state | |
1672 | * @worker: worker which is leaving idle state | |
1673 | * | |
1674 | * @worker is leaving idle state. Update stats. | |
1675 | * | |
1676 | * LOCKING: | |
d565ed63 | 1677 | * spin_lock_irq(pool->lock). |
c8e55f36 TH |
1678 | */ |
1679 | static void worker_leave_idle(struct worker *worker) | |
1680 | { | |
bd7bdd43 | 1681 | struct worker_pool *pool = worker->pool; |
c8e55f36 | 1682 | |
6183c009 TH |
1683 | if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) |
1684 | return; | |
d302f017 | 1685 | worker_clr_flags(worker, WORKER_IDLE); |
bd7bdd43 | 1686 | pool->nr_idle--; |
c8e55f36 TH |
1687 | list_del_init(&worker->entry); |
1688 | } | |
1689 | ||
f7537df5 | 1690 | static struct worker *alloc_worker(int node) |
c34056a3 TH |
1691 | { |
1692 | struct worker *worker; | |
1693 | ||
f7537df5 | 1694 | worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node); |
c8e55f36 TH |
1695 | if (worker) { |
1696 | INIT_LIST_HEAD(&worker->entry); | |
affee4b2 | 1697 | INIT_LIST_HEAD(&worker->scheduled); |
da028469 | 1698 | INIT_LIST_HEAD(&worker->node); |
e22bee78 TH |
1699 | /* on creation a worker is in !idle && prep state */ |
1700 | worker->flags = WORKER_PREP; | |
c8e55f36 | 1701 | } |
c34056a3 TH |
1702 | return worker; |
1703 | } | |
1704 | ||
4736cbf7 LJ |
1705 | /** |
1706 | * worker_attach_to_pool() - attach a worker to a pool | |
1707 | * @worker: worker to be attached | |
1708 | * @pool: the target pool | |
1709 | * | |
1710 | * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and | |
1711 | * cpu-binding of @worker are kept coordinated with the pool across | |
1712 | * cpu-[un]hotplugs. | |
1713 | */ | |
1714 | static void worker_attach_to_pool(struct worker *worker, | |
1715 | struct worker_pool *pool) | |
1716 | { | |
1717 | mutex_lock(&pool->attach_mutex); | |
1718 | ||
1719 | /* | |
1720 | * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any | |
1721 | * online CPUs. It'll be re-applied when any of the CPUs come up. | |
1722 | */ | |
1723 | set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask); | |
1724 | ||
1725 | /* | |
1726 | * The pool->attach_mutex ensures %POOL_DISASSOCIATED remains | |
1727 | * stable across this function. See the comments above the | |
1728 | * flag definition for details. | |
1729 | */ | |
1730 | if (pool->flags & POOL_DISASSOCIATED) | |
1731 | worker->flags |= WORKER_UNBOUND; | |
1732 | ||
1733 | list_add_tail(&worker->node, &pool->workers); | |
1734 | ||
1735 | mutex_unlock(&pool->attach_mutex); | |
1736 | } | |
1737 | ||
60f5a4bc LJ |
1738 | /** |
1739 | * worker_detach_from_pool() - detach a worker from its pool | |
1740 | * @worker: worker which is attached to its pool | |
1741 | * @pool: the pool @worker is attached to | |
1742 | * | |
4736cbf7 LJ |
1743 | * Undo the attaching which had been done in worker_attach_to_pool(). The |
1744 | * caller worker shouldn't access to the pool after detached except it has | |
1745 | * other reference to the pool. | |
60f5a4bc LJ |
1746 | */ |
1747 | static void worker_detach_from_pool(struct worker *worker, | |
1748 | struct worker_pool *pool) | |
1749 | { | |
1750 | struct completion *detach_completion = NULL; | |
1751 | ||
92f9c5c4 | 1752 | mutex_lock(&pool->attach_mutex); |
da028469 LJ |
1753 | list_del(&worker->node); |
1754 | if (list_empty(&pool->workers)) | |
60f5a4bc | 1755 | detach_completion = pool->detach_completion; |
92f9c5c4 | 1756 | mutex_unlock(&pool->attach_mutex); |
60f5a4bc | 1757 | |
b62c0751 LJ |
1758 | /* clear leftover flags without pool->lock after it is detached */ |
1759 | worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND); | |
1760 | ||
60f5a4bc LJ |
1761 | if (detach_completion) |
1762 | complete(detach_completion); | |
1763 | } | |
1764 | ||
c34056a3 TH |
1765 | /** |
1766 | * create_worker - create a new workqueue worker | |
63d95a91 | 1767 | * @pool: pool the new worker will belong to |
c34056a3 | 1768 | * |
051e1850 | 1769 | * Create and start a new worker which is attached to @pool. |
c34056a3 TH |
1770 | * |
1771 | * CONTEXT: | |
1772 | * Might sleep. Does GFP_KERNEL allocations. | |
1773 | * | |
d185af30 | 1774 | * Return: |
c34056a3 TH |
1775 | * Pointer to the newly created worker. |
1776 | */ | |
bc2ae0f5 | 1777 | static struct worker *create_worker(struct worker_pool *pool) |
c34056a3 | 1778 | { |
c34056a3 | 1779 | struct worker *worker = NULL; |
f3421797 | 1780 | int id = -1; |
e3c916a4 | 1781 | char id_buf[16]; |
c34056a3 | 1782 | |
7cda9aae LJ |
1783 | /* ID is needed to determine kthread name */ |
1784 | id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL); | |
822d8405 TH |
1785 | if (id < 0) |
1786 | goto fail; | |
c34056a3 | 1787 | |
f7537df5 | 1788 | worker = alloc_worker(pool->node); |
c34056a3 TH |
1789 | if (!worker) |
1790 | goto fail; | |
1791 | ||
bd7bdd43 | 1792 | worker->pool = pool; |
c34056a3 TH |
1793 | worker->id = id; |
1794 | ||
29c91e99 | 1795 | if (pool->cpu >= 0) |
e3c916a4 TH |
1796 | snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, |
1797 | pool->attrs->nice < 0 ? "H" : ""); | |
f3421797 | 1798 | else |
e3c916a4 TH |
1799 | snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); |
1800 | ||
f3f90ad4 | 1801 | worker->task = kthread_create_on_node(worker_thread, worker, pool->node, |
e3c916a4 | 1802 | "kworker/%s", id_buf); |
c34056a3 TH |
1803 | if (IS_ERR(worker->task)) |
1804 | goto fail; | |
1805 | ||
91151228 | 1806 | set_user_nice(worker->task, pool->attrs->nice); |
25834c73 | 1807 | kthread_bind_mask(worker->task, pool->attrs->cpumask); |
91151228 | 1808 | |
da028469 | 1809 | /* successful, attach the worker to the pool */ |
4736cbf7 | 1810 | worker_attach_to_pool(worker, pool); |
822d8405 | 1811 | |
051e1850 LJ |
1812 | /* start the newly created worker */ |
1813 | spin_lock_irq(&pool->lock); | |
1814 | worker->pool->nr_workers++; | |
1815 | worker_enter_idle(worker); | |
1816 | wake_up_process(worker->task); | |
1817 | spin_unlock_irq(&pool->lock); | |
1818 | ||
c34056a3 | 1819 | return worker; |
822d8405 | 1820 | |
c34056a3 | 1821 | fail: |
9625ab17 | 1822 | if (id >= 0) |
7cda9aae | 1823 | ida_simple_remove(&pool->worker_ida, id); |
c34056a3 TH |
1824 | kfree(worker); |
1825 | return NULL; | |
1826 | } | |
1827 | ||
c34056a3 TH |
1828 | /** |
1829 | * destroy_worker - destroy a workqueue worker | |
1830 | * @worker: worker to be destroyed | |
1831 | * | |
73eb7fe7 LJ |
1832 | * Destroy @worker and adjust @pool stats accordingly. The worker should |
1833 | * be idle. | |
c8e55f36 TH |
1834 | * |
1835 | * CONTEXT: | |
60f5a4bc | 1836 | * spin_lock_irq(pool->lock). |
c34056a3 TH |
1837 | */ |
1838 | static void destroy_worker(struct worker *worker) | |
1839 | { | |
bd7bdd43 | 1840 | struct worker_pool *pool = worker->pool; |
c34056a3 | 1841 | |
cd549687 TH |
1842 | lockdep_assert_held(&pool->lock); |
1843 | ||
c34056a3 | 1844 | /* sanity check frenzy */ |
6183c009 | 1845 | if (WARN_ON(worker->current_work) || |
73eb7fe7 LJ |
1846 | WARN_ON(!list_empty(&worker->scheduled)) || |
1847 | WARN_ON(!(worker->flags & WORKER_IDLE))) | |
6183c009 | 1848 | return; |
c34056a3 | 1849 | |
73eb7fe7 LJ |
1850 | pool->nr_workers--; |
1851 | pool->nr_idle--; | |
5bdfff96 | 1852 | |
c8e55f36 | 1853 | list_del_init(&worker->entry); |
cb444766 | 1854 | worker->flags |= WORKER_DIE; |
60f5a4bc | 1855 | wake_up_process(worker->task); |
c34056a3 TH |
1856 | } |
1857 | ||
63d95a91 | 1858 | static void idle_worker_timeout(unsigned long __pool) |
e22bee78 | 1859 | { |
63d95a91 | 1860 | struct worker_pool *pool = (void *)__pool; |
e22bee78 | 1861 | |
d565ed63 | 1862 | spin_lock_irq(&pool->lock); |
e22bee78 | 1863 | |
3347fc9f | 1864 | while (too_many_workers(pool)) { |
e22bee78 TH |
1865 | struct worker *worker; |
1866 | unsigned long expires; | |
1867 | ||
1868 | /* idle_list is kept in LIFO order, check the last one */ | |
63d95a91 | 1869 | worker = list_entry(pool->idle_list.prev, struct worker, entry); |
e22bee78 TH |
1870 | expires = worker->last_active + IDLE_WORKER_TIMEOUT; |
1871 | ||
3347fc9f | 1872 | if (time_before(jiffies, expires)) { |
63d95a91 | 1873 | mod_timer(&pool->idle_timer, expires); |
3347fc9f | 1874 | break; |
d5abe669 | 1875 | } |
3347fc9f LJ |
1876 | |
1877 | destroy_worker(worker); | |
e22bee78 TH |
1878 | } |
1879 | ||
d565ed63 | 1880 | spin_unlock_irq(&pool->lock); |
e22bee78 | 1881 | } |
d5abe669 | 1882 | |
493a1724 | 1883 | static void send_mayday(struct work_struct *work) |
e22bee78 | 1884 | { |
112202d9 TH |
1885 | struct pool_workqueue *pwq = get_work_pwq(work); |
1886 | struct workqueue_struct *wq = pwq->wq; | |
493a1724 | 1887 | |
2e109a28 | 1888 | lockdep_assert_held(&wq_mayday_lock); |
e22bee78 | 1889 | |
493008a8 | 1890 | if (!wq->rescuer) |
493a1724 | 1891 | return; |
e22bee78 TH |
1892 | |
1893 | /* mayday mayday mayday */ | |
493a1724 | 1894 | if (list_empty(&pwq->mayday_node)) { |
77668c8b LJ |
1895 | /* |
1896 | * If @pwq is for an unbound wq, its base ref may be put at | |
1897 | * any time due to an attribute change. Pin @pwq until the | |
1898 | * rescuer is done with it. | |
1899 | */ | |
1900 | get_pwq(pwq); | |
493a1724 | 1901 | list_add_tail(&pwq->mayday_node, &wq->maydays); |
e22bee78 | 1902 | wake_up_process(wq->rescuer->task); |
493a1724 | 1903 | } |
e22bee78 TH |
1904 | } |
1905 | ||
706026c2 | 1906 | static void pool_mayday_timeout(unsigned long __pool) |
e22bee78 | 1907 | { |
63d95a91 | 1908 | struct worker_pool *pool = (void *)__pool; |
e22bee78 TH |
1909 | struct work_struct *work; |
1910 | ||
b2d82909 TH |
1911 | spin_lock_irq(&pool->lock); |
1912 | spin_lock(&wq_mayday_lock); /* for wq->maydays */ | |
e22bee78 | 1913 | |
63d95a91 | 1914 | if (need_to_create_worker(pool)) { |
e22bee78 TH |
1915 | /* |
1916 | * We've been trying to create a new worker but | |
1917 | * haven't been successful. We might be hitting an | |
1918 | * allocation deadlock. Send distress signals to | |
1919 | * rescuers. | |
1920 | */ | |
63d95a91 | 1921 | list_for_each_entry(work, &pool->worklist, entry) |
e22bee78 | 1922 | send_mayday(work); |
1da177e4 | 1923 | } |
e22bee78 | 1924 | |
b2d82909 TH |
1925 | spin_unlock(&wq_mayday_lock); |
1926 | spin_unlock_irq(&pool->lock); | |
e22bee78 | 1927 | |
63d95a91 | 1928 | mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); |
1da177e4 LT |
1929 | } |
1930 | ||
e22bee78 TH |
1931 | /** |
1932 | * maybe_create_worker - create a new worker if necessary | |
63d95a91 | 1933 | * @pool: pool to create a new worker for |
e22bee78 | 1934 | * |
63d95a91 | 1935 | * Create a new worker for @pool if necessary. @pool is guaranteed to |
e22bee78 TH |
1936 | * have at least one idle worker on return from this function. If |
1937 | * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is | |
63d95a91 | 1938 | * sent to all rescuers with works scheduled on @pool to resolve |
e22bee78 TH |
1939 | * possible allocation deadlock. |
1940 | * | |
c5aa87bb TH |
1941 | * On return, need_to_create_worker() is guaranteed to be %false and |
1942 | * may_start_working() %true. | |
e22bee78 TH |
1943 | * |
1944 | * LOCKING: | |
d565ed63 | 1945 | * spin_lock_irq(pool->lock) which may be released and regrabbed |
e22bee78 TH |
1946 | * multiple times. Does GFP_KERNEL allocations. Called only from |
1947 | * manager. | |
e22bee78 | 1948 | */ |
29187a9e | 1949 | static void maybe_create_worker(struct worker_pool *pool) |
d565ed63 TH |
1950 | __releases(&pool->lock) |
1951 | __acquires(&pool->lock) | |
1da177e4 | 1952 | { |
e22bee78 | 1953 | restart: |
d565ed63 | 1954 | spin_unlock_irq(&pool->lock); |
9f9c2364 | 1955 | |
e22bee78 | 1956 | /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ |
63d95a91 | 1957 | mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); |
e22bee78 TH |
1958 | |
1959 | while (true) { | |
051e1850 | 1960 | if (create_worker(pool) || !need_to_create_worker(pool)) |
e22bee78 | 1961 | break; |
1da177e4 | 1962 | |
e212f361 | 1963 | schedule_timeout_interruptible(CREATE_COOLDOWN); |
9f9c2364 | 1964 | |
63d95a91 | 1965 | if (!need_to_create_worker(pool)) |
e22bee78 TH |
1966 | break; |
1967 | } | |
1968 | ||
63d95a91 | 1969 | del_timer_sync(&pool->mayday_timer); |
d565ed63 | 1970 | spin_lock_irq(&pool->lock); |
051e1850 LJ |
1971 | /* |
1972 | * This is necessary even after a new worker was just successfully | |
1973 | * created as @pool->lock was dropped and the new worker might have | |
1974 | * already become busy. | |
1975 | */ | |
63d95a91 | 1976 | if (need_to_create_worker(pool)) |
e22bee78 | 1977 | goto restart; |
e22bee78 TH |
1978 | } |
1979 | ||
73f53c4a | 1980 | /** |
e22bee78 TH |
1981 | * manage_workers - manage worker pool |
1982 | * @worker: self | |
73f53c4a | 1983 | * |
706026c2 | 1984 | * Assume the manager role and manage the worker pool @worker belongs |
e22bee78 | 1985 | * to. At any given time, there can be only zero or one manager per |
706026c2 | 1986 | * pool. The exclusion is handled automatically by this function. |
e22bee78 TH |
1987 | * |
1988 | * The caller can safely start processing works on false return. On | |
1989 | * true return, it's guaranteed that need_to_create_worker() is false | |
1990 | * and may_start_working() is true. | |
73f53c4a TH |
1991 | * |
1992 | * CONTEXT: | |
d565ed63 | 1993 | * spin_lock_irq(pool->lock) which may be released and regrabbed |
e22bee78 TH |
1994 | * multiple times. Does GFP_KERNEL allocations. |
1995 | * | |
d185af30 | 1996 | * Return: |
29187a9e TH |
1997 | * %false if the pool doesn't need management and the caller can safely |
1998 | * start processing works, %true if management function was performed and | |
1999 | * the conditions that the caller verified before calling the function may | |
2000 | * no longer be true. | |
73f53c4a | 2001 | */ |
e22bee78 | 2002 | static bool manage_workers(struct worker *worker) |
73f53c4a | 2003 | { |
63d95a91 | 2004 | struct worker_pool *pool = worker->pool; |
73f53c4a | 2005 | |
692b4825 | 2006 | if (pool->flags & POOL_MANAGER_ACTIVE) |
29187a9e | 2007 | return false; |
692b4825 TH |
2008 | |
2009 | pool->flags |= POOL_MANAGER_ACTIVE; | |
2607d7a6 | 2010 | pool->manager = worker; |
1e19ffc6 | 2011 | |
29187a9e | 2012 | maybe_create_worker(pool); |
e22bee78 | 2013 | |
2607d7a6 | 2014 | pool->manager = NULL; |
692b4825 TH |
2015 | pool->flags &= ~POOL_MANAGER_ACTIVE; |
2016 | wake_up(&wq_manager_wait); | |
29187a9e | 2017 | return true; |
73f53c4a TH |
2018 | } |
2019 | ||
a62428c0 TH |
2020 | /** |
2021 | * process_one_work - process single work | |
c34056a3 | 2022 | * @worker: self |
a62428c0 TH |
2023 | * @work: work to process |
2024 | * | |
2025 | * Process @work. This function contains all the logics necessary to | |
2026 | * process a single work including synchronization against and | |
2027 | * interaction with other workers on the same cpu, queueing and | |
2028 | * flushing. As long as context requirement is met, any worker can | |
2029 | * call this function to process a work. | |
2030 | * | |
2031 | * CONTEXT: | |
d565ed63 | 2032 | * spin_lock_irq(pool->lock) which is released and regrabbed. |
a62428c0 | 2033 | */ |
c34056a3 | 2034 | static void process_one_work(struct worker *worker, struct work_struct *work) |
d565ed63 TH |
2035 | __releases(&pool->lock) |
2036 | __acquires(&pool->lock) | |
a62428c0 | 2037 | { |
112202d9 | 2038 | struct pool_workqueue *pwq = get_work_pwq(work); |
bd7bdd43 | 2039 | struct worker_pool *pool = worker->pool; |
112202d9 | 2040 | bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE; |
73f53c4a | 2041 | int work_color; |
7e11629d | 2042 | struct worker *collision; |
a62428c0 TH |
2043 | #ifdef CONFIG_LOCKDEP |
2044 | /* | |
2045 | * It is permissible to free the struct work_struct from | |
2046 | * inside the function that is called from it, this we need to | |
2047 | * take into account for lockdep too. To avoid bogus "held | |
2048 | * lock freed" warnings as well as problems when looking into | |
2049 | * work->lockdep_map, make a copy and use that here. | |
2050 | */ | |
4d82a1de PZ |
2051 | struct lockdep_map lockdep_map; |
2052 | ||
2053 | lockdep_copy_map(&lockdep_map, &work->lockdep_map); | |
a62428c0 | 2054 | #endif |
807407c0 | 2055 | /* ensure we're on the correct CPU */ |
85327af6 | 2056 | WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && |
ec22ca5e | 2057 | raw_smp_processor_id() != pool->cpu); |
25511a47 | 2058 | |
7e11629d TH |
2059 | /* |
2060 | * A single work shouldn't be executed concurrently by | |
2061 | * multiple workers on a single cpu. Check whether anyone is | |
2062 | * already processing the work. If so, defer the work to the | |
2063 | * currently executing one. | |
2064 | */ | |
c9e7cf27 | 2065 | collision = find_worker_executing_work(pool, work); |
7e11629d TH |
2066 | if (unlikely(collision)) { |
2067 | move_linked_works(work, &collision->scheduled, NULL); | |
2068 | return; | |
2069 | } | |
2070 | ||
8930caba | 2071 | /* claim and dequeue */ |
a62428c0 | 2072 | debug_work_deactivate(work); |
c9e7cf27 | 2073 | hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); |
c34056a3 | 2074 | worker->current_work = work; |
a2c1c57b | 2075 | worker->current_func = work->func; |
112202d9 | 2076 | worker->current_pwq = pwq; |
73f53c4a | 2077 | work_color = get_work_color(work); |
7a22ad75 | 2078 | |
a62428c0 TH |
2079 | list_del_init(&work->entry); |
2080 | ||
fb0e7beb | 2081 | /* |
228f1d00 LJ |
2082 | * CPU intensive works don't participate in concurrency management. |
2083 | * They're the scheduler's responsibility. This takes @worker out | |
2084 | * of concurrency management and the next code block will chain | |
2085 | * execution of the pending work items. | |
fb0e7beb TH |
2086 | */ |
2087 | if (unlikely(cpu_intensive)) | |
228f1d00 | 2088 | worker_set_flags(worker, WORKER_CPU_INTENSIVE); |
fb0e7beb | 2089 | |
974271c4 | 2090 | /* |
a489a03e LJ |
2091 | * Wake up another worker if necessary. The condition is always |
2092 | * false for normal per-cpu workers since nr_running would always | |
2093 | * be >= 1 at this point. This is used to chain execution of the | |
2094 | * pending work items for WORKER_NOT_RUNNING workers such as the | |
228f1d00 | 2095 | * UNBOUND and CPU_INTENSIVE ones. |
974271c4 | 2096 | */ |
a489a03e | 2097 | if (need_more_worker(pool)) |
63d95a91 | 2098 | wake_up_worker(pool); |
974271c4 | 2099 | |
8930caba | 2100 | /* |
7c3eed5c | 2101 | * Record the last pool and clear PENDING which should be the last |
d565ed63 | 2102 | * update to @work. Also, do this inside @pool->lock so that |
23657bb1 TH |
2103 | * PENDING and queued state changes happen together while IRQ is |
2104 | * disabled. | |
8930caba | 2105 | */ |
7c3eed5c | 2106 | set_work_pool_and_clear_pending(work, pool->id); |
a62428c0 | 2107 | |
d565ed63 | 2108 | spin_unlock_irq(&pool->lock); |
a62428c0 | 2109 | |
a1d14934 | 2110 | lock_map_acquire(&pwq->wq->lockdep_map); |
a62428c0 | 2111 | lock_map_acquire(&lockdep_map); |
e6f3faa7 | 2112 | /* |
f52be570 PZ |
2113 | * Strictly speaking we should mark the invariant state without holding |
2114 | * any locks, that is, before these two lock_map_acquire()'s. | |
e6f3faa7 PZ |
2115 | * |
2116 | * However, that would result in: | |
2117 | * | |
2118 | * A(W1) | |
2119 | * WFC(C) | |
2120 | * A(W1) | |
2121 | * C(C) | |
2122 | * | |
2123 | * Which would create W1->C->W1 dependencies, even though there is no | |
2124 | * actual deadlock possible. There are two solutions, using a | |
2125 | * read-recursive acquire on the work(queue) 'locks', but this will then | |
f52be570 | 2126 | * hit the lockdep limitation on recursive locks, or simply discard |
e6f3faa7 PZ |
2127 | * these locks. |
2128 | * | |
2129 | * AFAICT there is no possible deadlock scenario between the | |
2130 | * flush_work() and complete() primitives (except for single-threaded | |
2131 | * workqueues), so hiding them isn't a problem. | |
2132 | */ | |
f52be570 | 2133 | lockdep_invariant_state(true); |
e36c886a | 2134 | trace_workqueue_execute_start(work); |
dd101ca5 | 2135 | dbg_snapshot_work(worker, worker->task, worker->current_func, DSS_FLAG_IN); |
a2c1c57b | 2136 | worker->current_func(work); |
dd101ca5 | 2137 | dbg_snapshot_work(worker, worker->task, worker->current_func, DSS_FLAG_OUT); |
e36c886a AV |
2138 | /* |
2139 | * While we must be careful to not use "work" after this, the trace | |
2140 | * point will only record its address. | |
2141 | */ | |
2142 | trace_workqueue_execute_end(work); | |
a62428c0 | 2143 | lock_map_release(&lockdep_map); |
112202d9 | 2144 | lock_map_release(&pwq->wq->lockdep_map); |
a62428c0 TH |
2145 | |
2146 | if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { | |
044c782c VI |
2147 | pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" |
2148 | " last function: %pf\n", | |
a2c1c57b TH |
2149 | current->comm, preempt_count(), task_pid_nr(current), |
2150 | worker->current_func); | |
a62428c0 TH |
2151 | debug_show_held_locks(current); |
2152 | dump_stack(); | |
2153 | } | |
2154 | ||
b22ce278 TH |
2155 | /* |
2156 | * The following prevents a kworker from hogging CPU on !PREEMPT | |
2157 | * kernels, where a requeueing work item waiting for something to | |
2158 | * happen could deadlock with stop_machine as such work item could | |
2159 | * indefinitely requeue itself while all other CPUs are trapped in | |
789cbbec JL |
2160 | * stop_machine. At the same time, report a quiescent RCU state so |
2161 | * the same condition doesn't freeze RCU. | |
b22ce278 | 2162 | */ |
3e28e377 | 2163 | cond_resched_rcu_qs(); |
b22ce278 | 2164 | |
d565ed63 | 2165 | spin_lock_irq(&pool->lock); |
a62428c0 | 2166 | |
fb0e7beb TH |
2167 | /* clear cpu intensive status */ |
2168 | if (unlikely(cpu_intensive)) | |
2169 | worker_clr_flags(worker, WORKER_CPU_INTENSIVE); | |
2170 | ||
7ab429b6 JW |
2171 | /* tag the worker for identification in schedule() */ |
2172 | worker->last_func = worker->current_func; | |
2173 | ||
a62428c0 | 2174 | /* we're done with it, release */ |
42f8570f | 2175 | hash_del(&worker->hentry); |
c34056a3 | 2176 | worker->current_work = NULL; |
a2c1c57b | 2177 | worker->current_func = NULL; |
112202d9 | 2178 | worker->current_pwq = NULL; |
3d1cb205 | 2179 | worker->desc_valid = false; |
112202d9 | 2180 | pwq_dec_nr_in_flight(pwq, work_color); |
a62428c0 TH |
2181 | } |
2182 | ||
affee4b2 TH |
2183 | /** |
2184 | * process_scheduled_works - process scheduled works | |
2185 | * @worker: self | |
2186 | * | |
2187 | * Process all scheduled works. Please note that the scheduled list | |
2188 | * may change while processing a work, so this function repeatedly | |
2189 | * fetches a work from the top and executes it. | |
2190 | * | |
2191 | * CONTEXT: | |
d565ed63 | 2192 | * spin_lock_irq(pool->lock) which may be released and regrabbed |
affee4b2 TH |
2193 | * multiple times. |
2194 | */ | |
2195 | static void process_scheduled_works(struct worker *worker) | |
1da177e4 | 2196 | { |
affee4b2 TH |
2197 | while (!list_empty(&worker->scheduled)) { |
2198 | struct work_struct *work = list_first_entry(&worker->scheduled, | |
1da177e4 | 2199 | struct work_struct, entry); |
c34056a3 | 2200 | process_one_work(worker, work); |
1da177e4 | 2201 | } |
1da177e4 LT |
2202 | } |
2203 | ||
4690c4ab TH |
2204 | /** |
2205 | * worker_thread - the worker thread function | |
c34056a3 | 2206 | * @__worker: self |
4690c4ab | 2207 | * |
c5aa87bb TH |
2208 | * The worker thread function. All workers belong to a worker_pool - |
2209 | * either a per-cpu one or dynamic unbound one. These workers process all | |
2210 | * work items regardless of their specific target workqueue. The only | |
2211 | * exception is work items which belong to workqueues with a rescuer which | |
2212 | * will be explained in rescuer_thread(). | |
d185af30 YB |
2213 | * |
2214 | * Return: 0 | |
4690c4ab | 2215 | */ |
c34056a3 | 2216 | static int worker_thread(void *__worker) |
1da177e4 | 2217 | { |
c34056a3 | 2218 | struct worker *worker = __worker; |
bd7bdd43 | 2219 | struct worker_pool *pool = worker->pool; |
1da177e4 | 2220 | |
e22bee78 TH |
2221 | /* tell the scheduler that this is a workqueue worker */ |
2222 | worker->task->flags |= PF_WQ_WORKER; | |
c8e55f36 | 2223 | woke_up: |
d565ed63 | 2224 | spin_lock_irq(&pool->lock); |
1da177e4 | 2225 | |
a9ab775b TH |
2226 | /* am I supposed to die? */ |
2227 | if (unlikely(worker->flags & WORKER_DIE)) { | |
d565ed63 | 2228 | spin_unlock_irq(&pool->lock); |
a9ab775b TH |
2229 | WARN_ON_ONCE(!list_empty(&worker->entry)); |
2230 | worker->task->flags &= ~PF_WQ_WORKER; | |
60f5a4bc LJ |
2231 | |
2232 | set_task_comm(worker->task, "kworker/dying"); | |
7cda9aae | 2233 | ida_simple_remove(&pool->worker_ida, worker->id); |
60f5a4bc LJ |
2234 | worker_detach_from_pool(worker, pool); |
2235 | kfree(worker); | |
a9ab775b | 2236 | return 0; |
c8e55f36 | 2237 | } |
affee4b2 | 2238 | |
c8e55f36 | 2239 | worker_leave_idle(worker); |
db7bccf4 | 2240 | recheck: |
e22bee78 | 2241 | /* no more worker necessary? */ |
63d95a91 | 2242 | if (!need_more_worker(pool)) |
e22bee78 TH |
2243 | goto sleep; |
2244 | ||
2245 | /* do we need to manage? */ | |
63d95a91 | 2246 | if (unlikely(!may_start_working(pool)) && manage_workers(worker)) |
e22bee78 TH |
2247 | goto recheck; |
2248 | ||
c8e55f36 TH |
2249 | /* |
2250 | * ->scheduled list can only be filled while a worker is | |
2251 | * preparing to process a work or actually processing it. | |
2252 | * Make sure nobody diddled with it while I was sleeping. | |
2253 | */ | |
6183c009 | 2254 | WARN_ON_ONCE(!list_empty(&worker->scheduled)); |
c8e55f36 | 2255 | |
e22bee78 | 2256 | /* |
a9ab775b TH |
2257 | * Finish PREP stage. We're guaranteed to have at least one idle |
2258 | * worker or that someone else has already assumed the manager | |
2259 | * role. This is where @worker starts participating in concurrency | |
2260 | * management if applicable and concurrency management is restored | |
2261 | * after being rebound. See rebind_workers() for details. | |
e22bee78 | 2262 | */ |
a9ab775b | 2263 | worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); |
e22bee78 TH |
2264 | |
2265 | do { | |
c8e55f36 | 2266 | struct work_struct *work = |
bd7bdd43 | 2267 | list_first_entry(&pool->worklist, |
c8e55f36 TH |
2268 | struct work_struct, entry); |
2269 | ||
82607adc TH |
2270 | pool->watchdog_ts = jiffies; |
2271 | ||
c8e55f36 TH |
2272 | if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { |
2273 | /* optimization path, not strictly necessary */ | |
2274 | process_one_work(worker, work); | |
2275 | if (unlikely(!list_empty(&worker->scheduled))) | |
affee4b2 | 2276 | process_scheduled_works(worker); |
c8e55f36 TH |
2277 | } else { |
2278 | move_linked_works(work, &worker->scheduled, NULL); | |
2279 | process_scheduled_works(worker); | |
affee4b2 | 2280 | } |
63d95a91 | 2281 | } while (keep_working(pool)); |
e22bee78 | 2282 | |
228f1d00 | 2283 | worker_set_flags(worker, WORKER_PREP); |
d313dd85 | 2284 | sleep: |
c8e55f36 | 2285 | /* |
d565ed63 TH |
2286 | * pool->lock is held and there's no work to process and no need to |
2287 | * manage, sleep. Workers are woken up only while holding | |
2288 | * pool->lock or from local cpu, so setting the current state | |
2289 | * before releasing pool->lock is enough to prevent losing any | |
2290 | * event. | |
c8e55f36 TH |
2291 | */ |
2292 | worker_enter_idle(worker); | |
c5a94a61 | 2293 | __set_current_state(TASK_IDLE); |
d565ed63 | 2294 | spin_unlock_irq(&pool->lock); |
c8e55f36 TH |
2295 | schedule(); |
2296 | goto woke_up; | |
1da177e4 LT |
2297 | } |
2298 | ||
e22bee78 TH |
2299 | /** |
2300 | * rescuer_thread - the rescuer thread function | |
111c225a | 2301 | * @__rescuer: self |
e22bee78 TH |
2302 | * |
2303 | * Workqueue rescuer thread function. There's one rescuer for each | |
493008a8 | 2304 | * workqueue which has WQ_MEM_RECLAIM set. |
e22bee78 | 2305 | * |
706026c2 | 2306 | * Regular work processing on a pool may block trying to create a new |
e22bee78 TH |
2307 | * worker which uses GFP_KERNEL allocation which has slight chance of |
2308 | * developing into deadlock if some works currently on the same queue | |
2309 | * need to be processed to satisfy the GFP_KERNEL allocation. This is | |
2310 | * the problem rescuer solves. | |
2311 | * | |
706026c2 TH |
2312 | * When such condition is possible, the pool summons rescuers of all |
2313 | * workqueues which have works queued on the pool and let them process | |
e22bee78 TH |
2314 | * those works so that forward progress can be guaranteed. |
2315 | * | |
2316 | * This should happen rarely. | |
d185af30 YB |
2317 | * |
2318 | * Return: 0 | |
e22bee78 | 2319 | */ |
111c225a | 2320 | static int rescuer_thread(void *__rescuer) |
e22bee78 | 2321 | { |
111c225a TH |
2322 | struct worker *rescuer = __rescuer; |
2323 | struct workqueue_struct *wq = rescuer->rescue_wq; | |
e22bee78 | 2324 | struct list_head *scheduled = &rescuer->scheduled; |
4d595b86 | 2325 | bool should_stop; |
e22bee78 TH |
2326 | |
2327 | set_user_nice(current, RESCUER_NICE_LEVEL); | |
111c225a TH |
2328 | |
2329 | /* | |
2330 | * Mark rescuer as worker too. As WORKER_PREP is never cleared, it | |
2331 | * doesn't participate in concurrency management. | |
2332 | */ | |
2333 | rescuer->task->flags |= PF_WQ_WORKER; | |
e22bee78 | 2334 | repeat: |
c5a94a61 | 2335 | set_current_state(TASK_IDLE); |
e22bee78 | 2336 | |
4d595b86 LJ |
2337 | /* |
2338 | * By the time the rescuer is requested to stop, the workqueue | |
2339 | * shouldn't have any work pending, but @wq->maydays may still have | |
2340 | * pwq(s) queued. This can happen by non-rescuer workers consuming | |
2341 | * all the work items before the rescuer got to them. Go through | |
2342 | * @wq->maydays processing before acting on should_stop so that the | |
2343 | * list is always empty on exit. | |
2344 | */ | |
2345 | should_stop = kthread_should_stop(); | |
e22bee78 | 2346 | |
493a1724 | 2347 | /* see whether any pwq is asking for help */ |
2e109a28 | 2348 | spin_lock_irq(&wq_mayday_lock); |
493a1724 TH |
2349 | |
2350 | while (!list_empty(&wq->maydays)) { | |
2351 | struct pool_workqueue *pwq = list_first_entry(&wq->maydays, | |
2352 | struct pool_workqueue, mayday_node); | |
112202d9 | 2353 | struct worker_pool *pool = pwq->pool; |
e22bee78 | 2354 | struct work_struct *work, *n; |
82607adc | 2355 | bool first = true; |
e22bee78 TH |
2356 | |
2357 | __set_current_state(TASK_RUNNING); | |
493a1724 TH |
2358 | list_del_init(&pwq->mayday_node); |
2359 | ||
2e109a28 | 2360 | spin_unlock_irq(&wq_mayday_lock); |
e22bee78 | 2361 | |
51697d39 LJ |
2362 | worker_attach_to_pool(rescuer, pool); |
2363 | ||
2364 | spin_lock_irq(&pool->lock); | |
b3104104 | 2365 | rescuer->pool = pool; |
e22bee78 TH |
2366 | |
2367 | /* | |
2368 | * Slurp in all works issued via this workqueue and | |
2369 | * process'em. | |
2370 | */ | |
0479c8c5 | 2371 | WARN_ON_ONCE(!list_empty(scheduled)); |
82607adc TH |
2372 | list_for_each_entry_safe(work, n, &pool->worklist, entry) { |
2373 | if (get_work_pwq(work) == pwq) { | |
2374 | if (first) | |
2375 | pool->watchdog_ts = jiffies; | |
e22bee78 | 2376 | move_linked_works(work, scheduled, &n); |
82607adc TH |
2377 | } |
2378 | first = false; | |
2379 | } | |
e22bee78 | 2380 | |
008847f6 N |
2381 | if (!list_empty(scheduled)) { |
2382 | process_scheduled_works(rescuer); | |
2383 | ||
2384 | /* | |
2385 | * The above execution of rescued work items could | |
2386 | * have created more to rescue through | |
2387 | * pwq_activate_first_delayed() or chained | |
2388 | * queueing. Let's put @pwq back on mayday list so | |
2389 | * that such back-to-back work items, which may be | |
2390 | * being used to relieve memory pressure, don't | |
2391 | * incur MAYDAY_INTERVAL delay inbetween. | |
2392 | */ | |
2393 | if (need_to_create_worker(pool)) { | |
2394 | spin_lock(&wq_mayday_lock); | |
2395 | get_pwq(pwq); | |
2396 | list_move_tail(&pwq->mayday_node, &wq->maydays); | |
2397 | spin_unlock(&wq_mayday_lock); | |
2398 | } | |
2399 | } | |
7576958a | 2400 | |
77668c8b LJ |
2401 | /* |
2402 | * Put the reference grabbed by send_mayday(). @pool won't | |
13b1d625 | 2403 | * go away while we're still attached to it. |
77668c8b LJ |
2404 | */ |
2405 | put_pwq(pwq); | |
2406 | ||
7576958a | 2407 | /* |
d8ca83e6 | 2408 | * Leave this pool. If need_more_worker() is %true, notify a |
7576958a TH |
2409 | * regular worker; otherwise, we end up with 0 concurrency |
2410 | * and stalling the execution. | |
2411 | */ | |
d8ca83e6 | 2412 | if (need_more_worker(pool)) |
63d95a91 | 2413 | wake_up_worker(pool); |
7576958a | 2414 | |
b3104104 | 2415 | rescuer->pool = NULL; |
13b1d625 LJ |
2416 | spin_unlock_irq(&pool->lock); |
2417 | ||
2418 | worker_detach_from_pool(rescuer, pool); | |
2419 | ||
2420 | spin_lock_irq(&wq_mayday_lock); | |
e22bee78 TH |
2421 | } |
2422 | ||
2e109a28 | 2423 | spin_unlock_irq(&wq_mayday_lock); |
493a1724 | 2424 | |
4d595b86 LJ |
2425 | if (should_stop) { |
2426 | __set_current_state(TASK_RUNNING); | |
2427 | rescuer->task->flags &= ~PF_WQ_WORKER; | |
2428 | return 0; | |
2429 | } | |
2430 | ||
111c225a TH |
2431 | /* rescuers should never participate in concurrency management */ |
2432 | WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); | |
e22bee78 TH |
2433 | schedule(); |
2434 | goto repeat; | |
1da177e4 LT |
2435 | } |
2436 | ||
fca839c0 TH |
2437 | /** |
2438 | * check_flush_dependency - check for flush dependency sanity | |
2439 | * @target_wq: workqueue being flushed | |
2440 | * @target_work: work item being flushed (NULL for workqueue flushes) | |
2441 | * | |
2442 | * %current is trying to flush the whole @target_wq or @target_work on it. | |
2443 | * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not | |
2444 | * reclaiming memory or running on a workqueue which doesn't have | |
2445 | * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to | |
2446 | * a deadlock. | |
2447 | */ | |
2448 | static void check_flush_dependency(struct workqueue_struct *target_wq, | |
2449 | struct work_struct *target_work) | |
2450 | { | |
2451 | work_func_t target_func = target_work ? target_work->func : NULL; | |
2452 | struct worker *worker; | |
2453 | ||
2454 | if (target_wq->flags & WQ_MEM_RECLAIM) | |
2455 | return; | |
2456 | ||
2457 | worker = current_wq_worker(); | |
2458 | ||
2459 | WARN_ONCE(current->flags & PF_MEMALLOC, | |
2460 | "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%pf", | |
2461 | current->pid, current->comm, target_wq->name, target_func); | |
23d11a58 TH |
2462 | WARN_ONCE(worker && ((worker->current_pwq->wq->flags & |
2463 | (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM), | |
fca839c0 TH |
2464 | "workqueue: WQ_MEM_RECLAIM %s:%pf is flushing !WQ_MEM_RECLAIM %s:%pf", |
2465 | worker->current_pwq->wq->name, worker->current_func, | |
2466 | target_wq->name, target_func); | |
2467 | } | |
2468 | ||
fc2e4d70 ON |
2469 | struct wq_barrier { |
2470 | struct work_struct work; | |
2471 | struct completion done; | |
2607d7a6 | 2472 | struct task_struct *task; /* purely informational */ |
fc2e4d70 ON |
2473 | }; |
2474 | ||
2475 | static void wq_barrier_func(struct work_struct *work) | |
2476 | { | |
2477 | struct wq_barrier *barr = container_of(work, struct wq_barrier, work); | |
2478 | complete(&barr->done); | |
2479 | } | |
2480 | ||
4690c4ab TH |
2481 | /** |
2482 | * insert_wq_barrier - insert a barrier work | |
112202d9 | 2483 | * @pwq: pwq to insert barrier into |
4690c4ab | 2484 | * @barr: wq_barrier to insert |
affee4b2 TH |
2485 | * @target: target work to attach @barr to |
2486 | * @worker: worker currently executing @target, NULL if @target is not executing | |
4690c4ab | 2487 | * |
affee4b2 TH |
2488 | * @barr is linked to @target such that @barr is completed only after |
2489 | * @target finishes execution. Please note that the ordering | |
2490 | * guarantee is observed only with respect to @target and on the local | |
2491 | * cpu. | |
2492 | * | |
2493 | * Currently, a queued barrier can't be canceled. This is because | |
2494 | * try_to_grab_pending() can't determine whether the work to be | |
2495 | * grabbed is at the head of the queue and thus can't clear LINKED | |
2496 | * flag of the previous work while there must be a valid next work | |
2497 | * after a work with LINKED flag set. | |
2498 | * | |
2499 | * Note that when @worker is non-NULL, @target may be modified | |
112202d9 | 2500 | * underneath us, so we can't reliably determine pwq from @target. |
4690c4ab TH |
2501 | * |
2502 | * CONTEXT: | |
d565ed63 | 2503 | * spin_lock_irq(pool->lock). |
4690c4ab | 2504 | */ |
112202d9 | 2505 | static void insert_wq_barrier(struct pool_workqueue *pwq, |
affee4b2 TH |
2506 | struct wq_barrier *barr, |
2507 | struct work_struct *target, struct worker *worker) | |
fc2e4d70 | 2508 | { |
affee4b2 TH |
2509 | struct list_head *head; |
2510 | unsigned int linked = 0; | |
2511 | ||
dc186ad7 | 2512 | /* |
d565ed63 | 2513 | * debugobject calls are safe here even with pool->lock locked |
dc186ad7 TG |
2514 | * as we know for sure that this will not trigger any of the |
2515 | * checks and call back into the fixup functions where we | |
2516 | * might deadlock. | |
2517 | */ | |
ca1cab37 | 2518 | INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); |
22df02bb | 2519 | __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); |
52fa5bc5 BF |
2520 | |
2521 | /* | |
2522 | * Explicitly init the crosslock for wq_barrier::done, make its lock | |
2523 | * key a subkey of the corresponding work. As a result we won't | |
2524 | * build a dependency between wq_barrier::done and unrelated work. | |
2525 | */ | |
2526 | lockdep_init_map_crosslock((struct lockdep_map *)&barr->done.map, | |
2527 | "(complete)wq_barr::done", | |
2528 | target->lockdep_map.key, 1); | |
2529 | __init_completion(&barr->done); | |
2607d7a6 | 2530 | barr->task = current; |
83c22520 | 2531 | |
affee4b2 TH |
2532 | /* |
2533 | * If @target is currently being executed, schedule the | |
2534 | * barrier to the worker; otherwise, put it after @target. | |
2535 | */ | |
2536 | if (worker) | |
2537 | head = worker->scheduled.next; | |
2538 | else { | |
2539 | unsigned long *bits = work_data_bits(target); | |
2540 | ||
2541 | head = target->entry.next; | |
2542 | /* there can already be other linked works, inherit and set */ | |
2543 | linked = *bits & WORK_STRUCT_LINKED; | |
2544 | __set_bit(WORK_STRUCT_LINKED_BIT, bits); | |
2545 | } | |
2546 | ||
dc186ad7 | 2547 | debug_work_activate(&barr->work); |
112202d9 | 2548 | insert_work(pwq, &barr->work, head, |
affee4b2 | 2549 | work_color_to_flags(WORK_NO_COLOR) | linked); |
fc2e4d70 ON |
2550 | } |
2551 | ||
73f53c4a | 2552 | /** |
112202d9 | 2553 | * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing |
73f53c4a TH |
2554 | * @wq: workqueue being flushed |
2555 | * @flush_color: new flush color, < 0 for no-op | |
2556 | * @work_color: new work color, < 0 for no-op | |
2557 | * | |
112202d9 | 2558 | * Prepare pwqs for workqueue flushing. |
73f53c4a | 2559 | * |
112202d9 TH |
2560 | * If @flush_color is non-negative, flush_color on all pwqs should be |
2561 | * -1. If no pwq has in-flight commands at the specified color, all | |
2562 | * pwq->flush_color's stay at -1 and %false is returned. If any pwq | |
2563 | * has in flight commands, its pwq->flush_color is set to | |
2564 | * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq | |
73f53c4a TH |
2565 | * wakeup logic is armed and %true is returned. |
2566 | * | |
2567 | * The caller should have initialized @wq->first_flusher prior to | |
2568 | * calling this function with non-negative @flush_color. If | |
2569 | * @flush_color is negative, no flush color update is done and %false | |
2570 | * is returned. | |
2571 | * | |
112202d9 | 2572 | * If @work_color is non-negative, all pwqs should have the same |
73f53c4a TH |
2573 | * work_color which is previous to @work_color and all will be |
2574 | * advanced to @work_color. | |
2575 | * | |
2576 | * CONTEXT: | |
3c25a55d | 2577 | * mutex_lock(wq->mutex). |
73f53c4a | 2578 | * |
d185af30 | 2579 | * Return: |
73f53c4a TH |
2580 | * %true if @flush_color >= 0 and there's something to flush. %false |
2581 | * otherwise. | |
2582 | */ | |
112202d9 | 2583 | static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, |
73f53c4a | 2584 | int flush_color, int work_color) |
1da177e4 | 2585 | { |
73f53c4a | 2586 | bool wait = false; |
49e3cf44 | 2587 | struct pool_workqueue *pwq; |
1da177e4 | 2588 | |
73f53c4a | 2589 | if (flush_color >= 0) { |
6183c009 | 2590 | WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); |
112202d9 | 2591 | atomic_set(&wq->nr_pwqs_to_flush, 1); |
1da177e4 | 2592 | } |
2355b70f | 2593 | |
49e3cf44 | 2594 | for_each_pwq(pwq, wq) { |
112202d9 | 2595 | struct worker_pool *pool = pwq->pool; |
fc2e4d70 | 2596 | |
b09f4fd3 | 2597 | spin_lock_irq(&pool->lock); |
83c22520 | 2598 | |
73f53c4a | 2599 | if (flush_color >= 0) { |
6183c009 | 2600 | WARN_ON_ONCE(pwq->flush_color != -1); |
fc2e4d70 | 2601 | |
112202d9 TH |
2602 | if (pwq->nr_in_flight[flush_color]) { |
2603 | pwq->flush_color = flush_color; | |
2604 | atomic_inc(&wq->nr_pwqs_to_flush); | |
73f53c4a TH |
2605 | wait = true; |
2606 | } | |
2607 | } | |
1da177e4 | 2608 | |
73f53c4a | 2609 | if (work_color >= 0) { |
6183c009 | 2610 | WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); |
112202d9 | 2611 | pwq->work_color = work_color; |
73f53c4a | 2612 | } |
1da177e4 | 2613 | |
b09f4fd3 | 2614 | spin_unlock_irq(&pool->lock); |
1da177e4 | 2615 | } |
2355b70f | 2616 | |
112202d9 | 2617 | if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) |
73f53c4a | 2618 | complete(&wq->first_flusher->done); |
14441960 | 2619 | |
73f53c4a | 2620 | return wait; |
1da177e4 LT |
2621 | } |
2622 | ||
0fcb78c2 | 2623 | /** |
1da177e4 | 2624 | * flush_workqueue - ensure that any scheduled work has run to completion. |
0fcb78c2 | 2625 | * @wq: workqueue to flush |
1da177e4 | 2626 | * |
c5aa87bb TH |
2627 | * This function sleeps until all work items which were queued on entry |
2628 | * have finished execution, but it is not livelocked by new incoming ones. | |
1da177e4 | 2629 | */ |
7ad5b3a5 | 2630 | void flush_workqueue(struct workqueue_struct *wq) |
1da177e4 | 2631 | { |
73f53c4a TH |
2632 | struct wq_flusher this_flusher = { |
2633 | .list = LIST_HEAD_INIT(this_flusher.list), | |
2634 | .flush_color = -1, | |
2635 | .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done), | |
2636 | }; | |
2637 | int next_color; | |
1da177e4 | 2638 | |
3347fa09 TH |
2639 | if (WARN_ON(!wq_online)) |
2640 | return; | |
2641 | ||
3295f0ef IM |
2642 | lock_map_acquire(&wq->lockdep_map); |
2643 | lock_map_release(&wq->lockdep_map); | |
73f53c4a | 2644 | |
3c25a55d | 2645 | mutex_lock(&wq->mutex); |
73f53c4a TH |
2646 | |
2647 | /* | |
2648 | * Start-to-wait phase | |
2649 | */ | |
2650 | next_color = work_next_color(wq->work_color); | |
2651 | ||
2652 | if (next_color != wq->flush_color) { | |
2653 | /* | |
2654 | * Color space is not full. The current work_color | |
2655 | * becomes our flush_color and work_color is advanced | |
2656 | * by one. | |
2657 | */ | |
6183c009 | 2658 | WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); |
73f53c4a TH |
2659 | this_flusher.flush_color = wq->work_color; |
2660 | wq->work_color = next_color; | |
2661 | ||
2662 | if (!wq->first_flusher) { | |
2663 | /* no flush in progress, become the first flusher */ | |
6183c009 | 2664 | WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); |
73f53c4a TH |
2665 | |
2666 | wq->first_flusher = &this_flusher; | |
2667 | ||
112202d9 | 2668 | if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, |
73f53c4a TH |
2669 | wq->work_color)) { |
2670 | /* nothing to flush, done */ | |
2671 | wq->flush_color = next_color; | |
2672 | wq->first_flusher = NULL; | |
2673 | goto out_unlock; | |
2674 | } | |
2675 | } else { | |
2676 | /* wait in queue */ | |
6183c009 | 2677 | WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); |
73f53c4a | 2678 | list_add_tail(&this_flusher.list, &wq->flusher_queue); |
112202d9 | 2679 | flush_workqueue_prep_pwqs(wq, -1, wq->work_color); |
73f53c4a TH |
2680 | } |
2681 | } else { | |
2682 | /* | |
2683 | * Oops, color space is full, wait on overflow queue. | |
2684 | * The next flush completion will assign us | |
2685 | * flush_color and transfer to flusher_queue. | |
2686 | */ | |
2687 | list_add_tail(&this_flusher.list, &wq->flusher_overflow); | |
2688 | } | |
2689 | ||
fca839c0 TH |
2690 | check_flush_dependency(wq, NULL); |
2691 | ||
3c25a55d | 2692 | mutex_unlock(&wq->mutex); |
73f53c4a TH |
2693 | |
2694 | wait_for_completion(&this_flusher.done); | |
2695 | ||
2696 | /* | |
2697 | * Wake-up-and-cascade phase | |
2698 | * | |
2699 | * First flushers are responsible for cascading flushes and | |
2700 | * handling overflow. Non-first flushers can simply return. | |
2701 | */ | |
2702 | if (wq->first_flusher != &this_flusher) | |
2703 | return; | |
2704 | ||
3c25a55d | 2705 | mutex_lock(&wq->mutex); |
73f53c4a | 2706 | |
4ce48b37 TH |
2707 | /* we might have raced, check again with mutex held */ |
2708 | if (wq->first_flusher != &this_flusher) | |
2709 | goto out_unlock; | |
2710 | ||
73f53c4a TH |
2711 | wq->first_flusher = NULL; |
2712 | ||
6183c009 TH |
2713 | WARN_ON_ONCE(!list_empty(&this_flusher.list)); |
2714 | WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); | |
73f53c4a TH |
2715 | |
2716 | while (true) { | |
2717 | struct wq_flusher *next, *tmp; | |
2718 | ||
2719 | /* complete all the flushers sharing the current flush color */ | |
2720 | list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { | |
2721 | if (next->flush_color != wq->flush_color) | |
2722 | break; | |
2723 | list_del_init(&next->list); | |
2724 | complete(&next->done); | |
2725 | } | |
2726 | ||
6183c009 TH |
2727 | WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && |
2728 | wq->flush_color != work_next_color(wq->work_color)); | |
73f53c4a TH |
2729 | |
2730 | /* this flush_color is finished, advance by one */ | |
2731 | wq->flush_color = work_next_color(wq->flush_color); | |
2732 | ||
2733 | /* one color has been freed, handle overflow queue */ | |
2734 | if (!list_empty(&wq->flusher_overflow)) { | |
2735 | /* | |
2736 | * Assign the same color to all overflowed | |
2737 | * flushers, advance work_color and append to | |
2738 | * flusher_queue. This is the start-to-wait | |
2739 | * phase for these overflowed flushers. | |
2740 | */ | |
2741 | list_for_each_entry(tmp, &wq->flusher_overflow, list) | |
2742 | tmp->flush_color = wq->work_color; | |
2743 | ||
2744 | wq->work_color = work_next_color(wq->work_color); | |
2745 | ||
2746 | list_splice_tail_init(&wq->flusher_overflow, | |
2747 | &wq->flusher_queue); | |
112202d9 | 2748 | flush_workqueue_prep_pwqs(wq, -1, wq->work_color); |
73f53c4a TH |
2749 | } |
2750 | ||
2751 | if (list_empty(&wq->flusher_queue)) { | |
6183c009 | 2752 | WARN_ON_ONCE(wq->flush_color != wq->work_color); |
73f53c4a TH |
2753 | break; |
2754 | } | |
2755 | ||
2756 | /* | |
2757 | * Need to flush more colors. Make the next flusher | |
112202d9 | 2758 | * the new first flusher and arm pwqs. |
73f53c4a | 2759 | */ |
6183c009 TH |
2760 | WARN_ON_ONCE(wq->flush_color == wq->work_color); |
2761 | WARN_ON_ONCE(wq->flush_color != next->flush_color); | |
73f53c4a TH |
2762 | |
2763 | list_del_init(&next->list); | |
2764 | wq->first_flusher = next; | |
2765 | ||
112202d9 | 2766 | if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) |
73f53c4a TH |
2767 | break; |
2768 | ||
2769 | /* | |
2770 | * Meh... this color is already done, clear first | |
2771 | * flusher and repeat cascading. | |
2772 | */ | |
2773 | wq->first_flusher = NULL; | |
2774 | } | |
2775 | ||
2776 | out_unlock: | |
3c25a55d | 2777 | mutex_unlock(&wq->mutex); |
1da177e4 | 2778 | } |
1dadafa8 | 2779 | EXPORT_SYMBOL(flush_workqueue); |
1da177e4 | 2780 | |
9c5a2ba7 TH |
2781 | /** |
2782 | * drain_workqueue - drain a workqueue | |
2783 | * @wq: workqueue to drain | |
2784 | * | |
2785 | * Wait until the workqueue becomes empty. While draining is in progress, | |
2786 | * only chain queueing is allowed. IOW, only currently pending or running | |
2787 | * work items on @wq can queue further work items on it. @wq is flushed | |
b749b1b6 | 2788 | * repeatedly until it becomes empty. The number of flushing is determined |
9c5a2ba7 TH |
2789 | * by the depth of chaining and should be relatively short. Whine if it |
2790 | * takes too long. | |
2791 | */ | |
2792 | void drain_workqueue(struct workqueue_struct *wq) | |
2793 | { | |
2794 | unsigned int flush_cnt = 0; | |
49e3cf44 | 2795 | struct pool_workqueue *pwq; |
9c5a2ba7 TH |
2796 | |
2797 | /* | |
2798 | * __queue_work() needs to test whether there are drainers, is much | |
2799 | * hotter than drain_workqueue() and already looks at @wq->flags. | |
618b01eb | 2800 | * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. |
9c5a2ba7 | 2801 | */ |
87fc741e | 2802 | mutex_lock(&wq->mutex); |
9c5a2ba7 | 2803 | if (!wq->nr_drainers++) |
618b01eb | 2804 | wq->flags |= __WQ_DRAINING; |
87fc741e | 2805 | mutex_unlock(&wq->mutex); |
9c5a2ba7 TH |
2806 | reflush: |
2807 | flush_workqueue(wq); | |
2808 | ||
b09f4fd3 | 2809 | mutex_lock(&wq->mutex); |
76af4d93 | 2810 | |
49e3cf44 | 2811 | for_each_pwq(pwq, wq) { |
fa2563e4 | 2812 | bool drained; |
9c5a2ba7 | 2813 | |
b09f4fd3 | 2814 | spin_lock_irq(&pwq->pool->lock); |
112202d9 | 2815 | drained = !pwq->nr_active && list_empty(&pwq->delayed_works); |
b09f4fd3 | 2816 | spin_unlock_irq(&pwq->pool->lock); |
fa2563e4 TT |
2817 | |
2818 | if (drained) | |
9c5a2ba7 TH |
2819 | continue; |
2820 | ||
2821 | if (++flush_cnt == 10 || | |
2822 | (flush_cnt % 100 == 0 && flush_cnt <= 1000)) | |
c5aa87bb | 2823 | pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n", |
044c782c | 2824 | wq->name, flush_cnt); |
76af4d93 | 2825 | |
b09f4fd3 | 2826 | mutex_unlock(&wq->mutex); |
9c5a2ba7 TH |
2827 | goto reflush; |
2828 | } | |
2829 | ||
9c5a2ba7 | 2830 | if (!--wq->nr_drainers) |
618b01eb | 2831 | wq->flags &= ~__WQ_DRAINING; |
87fc741e | 2832 | mutex_unlock(&wq->mutex); |
9c5a2ba7 TH |
2833 | } |
2834 | EXPORT_SYMBOL_GPL(drain_workqueue); | |
2835 | ||
606a5020 | 2836 | static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr) |
db700897 | 2837 | { |
affee4b2 | 2838 | struct worker *worker = NULL; |
c9e7cf27 | 2839 | struct worker_pool *pool; |
112202d9 | 2840 | struct pool_workqueue *pwq; |
db700897 ON |
2841 | |
2842 | might_sleep(); | |
fa1b54e6 TH |
2843 | |
2844 | local_irq_disable(); | |
c9e7cf27 | 2845 | pool = get_work_pool(work); |
fa1b54e6 TH |
2846 | if (!pool) { |
2847 | local_irq_enable(); | |
baf59022 | 2848 | return false; |
fa1b54e6 | 2849 | } |
db700897 | 2850 | |
fa1b54e6 | 2851 | spin_lock(&pool->lock); |
0b3dae68 | 2852 | /* see the comment in try_to_grab_pending() with the same code */ |
112202d9 TH |
2853 | pwq = get_work_pwq(work); |
2854 | if (pwq) { | |
2855 | if (unlikely(pwq->pool != pool)) | |
4690c4ab | 2856 | goto already_gone; |
606a5020 | 2857 | } else { |
c9e7cf27 | 2858 | worker = find_worker_executing_work(pool, work); |
affee4b2 | 2859 | if (!worker) |
4690c4ab | 2860 | goto already_gone; |
112202d9 | 2861 | pwq = worker->current_pwq; |
606a5020 | 2862 | } |
db700897 | 2863 | |
fca839c0 TH |
2864 | check_flush_dependency(pwq->wq, work); |
2865 | ||
112202d9 | 2866 | insert_wq_barrier(pwq, barr, work, worker); |
d565ed63 | 2867 | spin_unlock_irq(&pool->lock); |
7a22ad75 | 2868 | |
e159489b | 2869 | /* |
a1d14934 PZ |
2870 | * Force a lock recursion deadlock when using flush_work() inside a |
2871 | * single-threaded or rescuer equipped workqueue. | |
2872 | * | |
2873 | * For single threaded workqueues the deadlock happens when the work | |
2874 | * is after the work issuing the flush_work(). For rescuer equipped | |
2875 | * workqueues the deadlock happens when the rescuer stalls, blocking | |
2876 | * forward progress. | |
e159489b | 2877 | */ |
a1d14934 | 2878 | if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer) { |
112202d9 | 2879 | lock_map_acquire(&pwq->wq->lockdep_map); |
a1d14934 PZ |
2880 | lock_map_release(&pwq->wq->lockdep_map); |
2881 | } | |
e159489b | 2882 | |
401a8d04 | 2883 | return true; |
4690c4ab | 2884 | already_gone: |
d565ed63 | 2885 | spin_unlock_irq(&pool->lock); |
401a8d04 | 2886 | return false; |
db700897 | 2887 | } |
baf59022 TH |
2888 | |
2889 | /** | |
2890 | * flush_work - wait for a work to finish executing the last queueing instance | |
2891 | * @work: the work to flush | |
2892 | * | |
606a5020 TH |
2893 | * Wait until @work has finished execution. @work is guaranteed to be idle |
2894 | * on return if it hasn't been requeued since flush started. | |
baf59022 | 2895 | * |
d185af30 | 2896 | * Return: |
baf59022 TH |
2897 | * %true if flush_work() waited for the work to finish execution, |
2898 | * %false if it was already idle. | |
2899 | */ | |
2900 | bool flush_work(struct work_struct *work) | |
2901 | { | |
12997d1a BH |
2902 | struct wq_barrier barr; |
2903 | ||
3347fa09 TH |
2904 | if (WARN_ON(!wq_online)) |
2905 | return false; | |
2906 | ||
0976dfc1 SB |
2907 | lock_map_acquire(&work->lockdep_map); |
2908 | lock_map_release(&work->lockdep_map); | |
2909 | ||
12997d1a BH |
2910 | if (start_flush_work(work, &barr)) { |
2911 | wait_for_completion(&barr.done); | |
2912 | destroy_work_on_stack(&barr.work); | |
2913 | return true; | |
2914 | } else { | |
2915 | return false; | |
2916 | } | |
6e84d644 | 2917 | } |
606a5020 | 2918 | EXPORT_SYMBOL_GPL(flush_work); |
6e84d644 | 2919 | |
8603e1b3 | 2920 | struct cwt_wait { |
ac6424b9 | 2921 | wait_queue_entry_t wait; |
8603e1b3 TH |
2922 | struct work_struct *work; |
2923 | }; | |
2924 | ||
ac6424b9 | 2925 | static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key) |
8603e1b3 TH |
2926 | { |
2927 | struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait); | |
2928 | ||
2929 | if (cwait->work != key) | |
2930 | return 0; | |
2931 | return autoremove_wake_function(wait, mode, sync, key); | |
2932 | } | |
2933 | ||
36e227d2 | 2934 | static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) |
1f1f642e | 2935 | { |
8603e1b3 | 2936 | static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq); |
bbb68dfa | 2937 | unsigned long flags; |
1f1f642e ON |
2938 | int ret; |
2939 | ||
2940 | do { | |
bbb68dfa TH |
2941 | ret = try_to_grab_pending(work, is_dwork, &flags); |
2942 | /* | |
8603e1b3 TH |
2943 | * If someone else is already canceling, wait for it to |
2944 | * finish. flush_work() doesn't work for PREEMPT_NONE | |
2945 | * because we may get scheduled between @work's completion | |
2946 | * and the other canceling task resuming and clearing | |
2947 | * CANCELING - flush_work() will return false immediately | |
2948 | * as @work is no longer busy, try_to_grab_pending() will | |
2949 | * return -ENOENT as @work is still being canceled and the | |
2950 | * other canceling task won't be able to clear CANCELING as | |
2951 | * we're hogging the CPU. | |
2952 | * | |
2953 | * Let's wait for completion using a waitqueue. As this | |
2954 | * may lead to the thundering herd problem, use a custom | |
2955 | * wake function which matches @work along with exclusive | |
2956 | * wait and wakeup. | |
bbb68dfa | 2957 | */ |
8603e1b3 TH |
2958 | if (unlikely(ret == -ENOENT)) { |
2959 | struct cwt_wait cwait; | |
2960 | ||
2961 | init_wait(&cwait.wait); | |
2962 | cwait.wait.func = cwt_wakefn; | |
2963 | cwait.work = work; | |
2964 | ||
2965 | prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait, | |
2966 | TASK_UNINTERRUPTIBLE); | |
2967 | if (work_is_canceling(work)) | |
2968 | schedule(); | |
2969 | finish_wait(&cancel_waitq, &cwait.wait); | |
2970 | } | |
1f1f642e ON |
2971 | } while (unlikely(ret < 0)); |
2972 | ||
bbb68dfa TH |
2973 | /* tell other tasks trying to grab @work to back off */ |
2974 | mark_work_canceling(work); | |
2975 | local_irq_restore(flags); | |
2976 | ||
3347fa09 TH |
2977 | /* |
2978 | * This allows canceling during early boot. We know that @work | |
2979 | * isn't executing. | |
2980 | */ | |
2981 | if (wq_online) | |
2982 | flush_work(work); | |
2983 | ||
7a22ad75 | 2984 | clear_work_data(work); |
8603e1b3 TH |
2985 | |
2986 | /* | |
2987 | * Paired with prepare_to_wait() above so that either | |
2988 | * waitqueue_active() is visible here or !work_is_canceling() is | |
2989 | * visible there. | |
2990 | */ | |
2991 | smp_mb(); | |
2992 | if (waitqueue_active(&cancel_waitq)) | |
2993 | __wake_up(&cancel_waitq, TASK_NORMAL, 1, work); | |
2994 | ||
1f1f642e ON |
2995 | return ret; |
2996 | } | |
2997 | ||
6e84d644 | 2998 | /** |
401a8d04 TH |
2999 | * cancel_work_sync - cancel a work and wait for it to finish |
3000 | * @work: the work to cancel | |
6e84d644 | 3001 | * |
401a8d04 TH |
3002 | * Cancel @work and wait for its execution to finish. This function |
3003 | * can be used even if the work re-queues itself or migrates to | |
3004 | * another workqueue. On return from this function, @work is | |
3005 | * guaranteed to be not pending or executing on any CPU. | |
1f1f642e | 3006 | * |
401a8d04 TH |
3007 | * cancel_work_sync(&delayed_work->work) must not be used for |
3008 | * delayed_work's. Use cancel_delayed_work_sync() instead. | |
6e84d644 | 3009 | * |
401a8d04 | 3010 | * The caller must ensure that the workqueue on which @work was last |
6e84d644 | 3011 | * queued can't be destroyed before this function returns. |
401a8d04 | 3012 | * |
d185af30 | 3013 | * Return: |
401a8d04 | 3014 | * %true if @work was pending, %false otherwise. |
6e84d644 | 3015 | */ |
401a8d04 | 3016 | bool cancel_work_sync(struct work_struct *work) |
6e84d644 | 3017 | { |
36e227d2 | 3018 | return __cancel_work_timer(work, false); |
b89deed3 | 3019 | } |
28e53bdd | 3020 | EXPORT_SYMBOL_GPL(cancel_work_sync); |
b89deed3 | 3021 | |
6e84d644 | 3022 | /** |
401a8d04 TH |
3023 | * flush_delayed_work - wait for a dwork to finish executing the last queueing |
3024 | * @dwork: the delayed work to flush | |
6e84d644 | 3025 | * |
401a8d04 TH |
3026 | * Delayed timer is cancelled and the pending work is queued for |
3027 | * immediate execution. Like flush_work(), this function only | |
3028 | * considers the last queueing instance of @dwork. | |
1f1f642e | 3029 | * |
d185af30 | 3030 | * Return: |
401a8d04 TH |
3031 | * %true if flush_work() waited for the work to finish execution, |
3032 | * %false if it was already idle. | |
6e84d644 | 3033 | */ |
401a8d04 TH |
3034 | bool flush_delayed_work(struct delayed_work *dwork) |
3035 | { | |
8930caba | 3036 | local_irq_disable(); |
401a8d04 | 3037 | if (del_timer_sync(&dwork->timer)) |
60c057bc | 3038 | __queue_work(dwork->cpu, dwork->wq, &dwork->work); |
8930caba | 3039 | local_irq_enable(); |
401a8d04 TH |
3040 | return flush_work(&dwork->work); |
3041 | } | |
3042 | EXPORT_SYMBOL(flush_delayed_work); | |
3043 | ||
f72b8792 JA |
3044 | static bool __cancel_work(struct work_struct *work, bool is_dwork) |
3045 | { | |
3046 | unsigned long flags; | |
3047 | int ret; | |
3048 | ||
3049 | do { | |
3050 | ret = try_to_grab_pending(work, is_dwork, &flags); | |
3051 | } while (unlikely(ret == -EAGAIN)); | |
3052 | ||
3053 | if (unlikely(ret < 0)) | |
3054 | return false; | |
3055 | ||
3056 | set_work_pool_and_clear_pending(work, get_work_pool_id(work)); | |
3057 | local_irq_restore(flags); | |
3058 | return ret; | |
3059 | } | |
3060 | ||
3061 | /* | |
3062 | * See cancel_delayed_work() | |
3063 | */ | |
3064 | bool cancel_work(struct work_struct *work) | |
3065 | { | |
3066 | return __cancel_work(work, false); | |
3067 | } | |
3068 | ||
09383498 | 3069 | /** |
57b30ae7 TH |
3070 | * cancel_delayed_work - cancel a delayed work |
3071 | * @dwork: delayed_work to cancel | |
09383498 | 3072 | * |
d185af30 YB |
3073 | * Kill off a pending delayed_work. |
3074 | * | |
3075 | * Return: %true if @dwork was pending and canceled; %false if it wasn't | |
3076 | * pending. | |
3077 | * | |
3078 | * Note: | |
3079 | * The work callback function may still be running on return, unless | |
3080 | * it returns %true and the work doesn't re-arm itself. Explicitly flush or | |
3081 | * use cancel_delayed_work_sync() to wait on it. | |
09383498 | 3082 | * |
57b30ae7 | 3083 | * This function is safe to call from any context including IRQ handler. |
09383498 | 3084 | */ |
57b30ae7 | 3085 | bool cancel_delayed_work(struct delayed_work *dwork) |
09383498 | 3086 | { |
f72b8792 | 3087 | return __cancel_work(&dwork->work, true); |
09383498 | 3088 | } |
57b30ae7 | 3089 | EXPORT_SYMBOL(cancel_delayed_work); |
09383498 | 3090 | |
401a8d04 TH |
3091 | /** |
3092 | * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish | |
3093 | * @dwork: the delayed work cancel | |
3094 | * | |
3095 | * This is cancel_work_sync() for delayed works. | |
3096 | * | |
d185af30 | 3097 | * Return: |
401a8d04 TH |
3098 | * %true if @dwork was pending, %false otherwise. |
3099 | */ | |
3100 | bool cancel_delayed_work_sync(struct delayed_work *dwork) | |
6e84d644 | 3101 | { |
36e227d2 | 3102 | return __cancel_work_timer(&dwork->work, true); |
6e84d644 | 3103 | } |
f5a421a4 | 3104 | EXPORT_SYMBOL(cancel_delayed_work_sync); |
1da177e4 | 3105 | |
b6136773 | 3106 | /** |
31ddd871 | 3107 | * schedule_on_each_cpu - execute a function synchronously on each online CPU |
b6136773 | 3108 | * @func: the function to call |
b6136773 | 3109 | * |
31ddd871 TH |
3110 | * schedule_on_each_cpu() executes @func on each online CPU using the |
3111 | * system workqueue and blocks until all CPUs have completed. | |
b6136773 | 3112 | * schedule_on_each_cpu() is very slow. |
31ddd871 | 3113 | * |
d185af30 | 3114 | * Return: |
31ddd871 | 3115 | * 0 on success, -errno on failure. |
b6136773 | 3116 | */ |
65f27f38 | 3117 | int schedule_on_each_cpu(work_func_t func) |
15316ba8 CL |
3118 | { |
3119 | int cpu; | |
38f51568 | 3120 | struct work_struct __percpu *works; |
15316ba8 | 3121 | |
b6136773 AM |
3122 | works = alloc_percpu(struct work_struct); |
3123 | if (!works) | |
15316ba8 | 3124 | return -ENOMEM; |
b6136773 | 3125 | |
93981800 TH |
3126 | get_online_cpus(); |
3127 | ||
15316ba8 | 3128 | for_each_online_cpu(cpu) { |
9bfb1839 IM |
3129 | struct work_struct *work = per_cpu_ptr(works, cpu); |
3130 | ||
3131 | INIT_WORK(work, func); | |
b71ab8c2 | 3132 | schedule_work_on(cpu, work); |
65a64464 | 3133 | } |
93981800 TH |
3134 | |
3135 | for_each_online_cpu(cpu) | |
3136 | flush_work(per_cpu_ptr(works, cpu)); | |
3137 | ||
95402b38 | 3138 | put_online_cpus(); |
b6136773 | 3139 | free_percpu(works); |
15316ba8 CL |
3140 | return 0; |
3141 | } | |
3142 | ||
1fa44eca JB |
3143 | /** |
3144 | * execute_in_process_context - reliably execute the routine with user context | |
3145 | * @fn: the function to execute | |
1fa44eca JB |
3146 | * @ew: guaranteed storage for the execute work structure (must |
3147 | * be available when the work executes) | |
3148 | * | |
3149 | * Executes the function immediately if process context is available, | |
3150 | * otherwise schedules the function for delayed execution. | |
3151 | * | |
d185af30 | 3152 | * Return: 0 - function was executed |
1fa44eca JB |
3153 | * 1 - function was scheduled for execution |
3154 | */ | |
65f27f38 | 3155 | int execute_in_process_context(work_func_t fn, struct execute_work *ew) |
1fa44eca JB |
3156 | { |
3157 | if (!in_interrupt()) { | |
65f27f38 | 3158 | fn(&ew->work); |
1fa44eca JB |
3159 | return 0; |
3160 | } | |
3161 | ||
65f27f38 | 3162 | INIT_WORK(&ew->work, fn); |
1fa44eca JB |
3163 | schedule_work(&ew->work); |
3164 | ||
3165 | return 1; | |
3166 | } | |
3167 | EXPORT_SYMBOL_GPL(execute_in_process_context); | |
3168 | ||
6ba94429 FW |
3169 | /** |
3170 | * free_workqueue_attrs - free a workqueue_attrs | |
3171 | * @attrs: workqueue_attrs to free | |
226223ab | 3172 | * |
6ba94429 | 3173 | * Undo alloc_workqueue_attrs(). |
226223ab | 3174 | */ |
6ba94429 | 3175 | void free_workqueue_attrs(struct workqueue_attrs *attrs) |
226223ab | 3176 | { |
6ba94429 FW |
3177 | if (attrs) { |
3178 | free_cpumask_var(attrs->cpumask); | |
3179 | kfree(attrs); | |
3180 | } | |
226223ab TH |
3181 | } |
3182 | ||
6ba94429 FW |
3183 | /** |
3184 | * alloc_workqueue_attrs - allocate a workqueue_attrs | |
3185 | * @gfp_mask: allocation mask to use | |
3186 | * | |
3187 | * Allocate a new workqueue_attrs, initialize with default settings and | |
3188 | * return it. | |
3189 | * | |
3190 | * Return: The allocated new workqueue_attr on success. %NULL on failure. | |
3191 | */ | |
3192 | struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask) | |
226223ab | 3193 | { |
6ba94429 | 3194 | struct workqueue_attrs *attrs; |
226223ab | 3195 | |
6ba94429 FW |
3196 | attrs = kzalloc(sizeof(*attrs), gfp_mask); |
3197 | if (!attrs) | |
3198 | goto fail; | |
3199 | if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask)) | |
3200 | goto fail; | |
3201 | ||
3202 | cpumask_copy(attrs->cpumask, cpu_possible_mask); | |
3203 | return attrs; | |
3204 | fail: | |
3205 | free_workqueue_attrs(attrs); | |
3206 | return NULL; | |
226223ab TH |
3207 | } |
3208 | ||
6ba94429 FW |
3209 | static void copy_workqueue_attrs(struct workqueue_attrs *to, |
3210 | const struct workqueue_attrs *from) | |
226223ab | 3211 | { |
6ba94429 FW |
3212 | to->nice = from->nice; |
3213 | cpumask_copy(to->cpumask, from->cpumask); | |
3214 | /* | |
3215 | * Unlike hash and equality test, this function doesn't ignore | |
3216 | * ->no_numa as it is used for both pool and wq attrs. Instead, | |
3217 | * get_unbound_pool() explicitly clears ->no_numa after copying. | |
3218 | */ | |
3219 | to->no_numa = from->no_numa; | |
226223ab TH |
3220 | } |
3221 | ||
6ba94429 FW |
3222 | /* hash value of the content of @attr */ |
3223 | static u32 wqattrs_hash(const struct workqueue_attrs *attrs) | |
226223ab | 3224 | { |
6ba94429 | 3225 | u32 hash = 0; |
226223ab | 3226 | |
6ba94429 FW |
3227 | hash = jhash_1word(attrs->nice, hash); |
3228 | hash = jhash(cpumask_bits(attrs->cpumask), | |
3229 | BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); | |
3230 | return hash; | |
226223ab | 3231 | } |
226223ab | 3232 | |
6ba94429 FW |
3233 | /* content equality test */ |
3234 | static bool wqattrs_equal(const struct workqueue_attrs *a, | |
3235 | const struct workqueue_attrs *b) | |
226223ab | 3236 | { |
6ba94429 FW |
3237 | if (a->nice != b->nice) |
3238 | return false; | |
3239 | if (!cpumask_equal(a->cpumask, b->cpumask)) | |
3240 | return false; | |
3241 | return true; | |
226223ab TH |
3242 | } |
3243 | ||
6ba94429 FW |
3244 | /** |
3245 | * init_worker_pool - initialize a newly zalloc'd worker_pool | |
3246 | * @pool: worker_pool to initialize | |
3247 | * | |
402dd89d | 3248 | * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs. |
6ba94429 FW |
3249 | * |
3250 | * Return: 0 on success, -errno on failure. Even on failure, all fields | |
3251 | * inside @pool proper are initialized and put_unbound_pool() can be called | |
3252 | * on @pool safely to release it. | |
3253 | */ | |
3254 | static int init_worker_pool(struct worker_pool *pool) | |
226223ab | 3255 | { |
6ba94429 FW |
3256 | spin_lock_init(&pool->lock); |
3257 | pool->id = -1; | |
3258 | pool->cpu = -1; | |
3259 | pool->node = NUMA_NO_NODE; | |
3260 | pool->flags |= POOL_DISASSOCIATED; | |
82607adc | 3261 | pool->watchdog_ts = jiffies; |
6ba94429 FW |
3262 | INIT_LIST_HEAD(&pool->worklist); |
3263 | INIT_LIST_HEAD(&pool->idle_list); | |
3264 | hash_init(pool->busy_hash); | |
226223ab | 3265 | |
c30fb26b GT |
3266 | setup_deferrable_timer(&pool->idle_timer, idle_worker_timeout, |
3267 | (unsigned long)pool); | |
226223ab | 3268 | |
6ba94429 FW |
3269 | setup_timer(&pool->mayday_timer, pool_mayday_timeout, |
3270 | (unsigned long)pool); | |
226223ab | 3271 | |
6ba94429 FW |
3272 | mutex_init(&pool->attach_mutex); |
3273 | INIT_LIST_HEAD(&pool->workers); | |
226223ab | 3274 | |
6ba94429 FW |
3275 | ida_init(&pool->worker_ida); |
3276 | INIT_HLIST_NODE(&pool->hash_node); | |
3277 | pool->refcnt = 1; | |
226223ab | 3278 | |
6ba94429 FW |
3279 | /* shouldn't fail above this point */ |
3280 | pool->attrs = alloc_workqueue_attrs(GFP_KERNEL); | |
3281 | if (!pool->attrs) | |
3282 | return -ENOMEM; | |
3283 | return 0; | |
226223ab TH |
3284 | } |
3285 | ||
6ba94429 | 3286 | static void rcu_free_wq(struct rcu_head *rcu) |
226223ab | 3287 | { |
6ba94429 FW |
3288 | struct workqueue_struct *wq = |
3289 | container_of(rcu, struct workqueue_struct, rcu); | |
226223ab | 3290 | |
6ba94429 FW |
3291 | if (!(wq->flags & WQ_UNBOUND)) |
3292 | free_percpu(wq->cpu_pwqs); | |
226223ab | 3293 | else |
6ba94429 | 3294 | free_workqueue_attrs(wq->unbound_attrs); |
226223ab | 3295 | |
6ba94429 FW |
3296 | kfree(wq->rescuer); |
3297 | kfree(wq); | |
226223ab TH |
3298 | } |
3299 | ||
6ba94429 | 3300 | static void rcu_free_pool(struct rcu_head *rcu) |
226223ab | 3301 | { |
6ba94429 | 3302 | struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); |
226223ab | 3303 | |
6ba94429 FW |
3304 | ida_destroy(&pool->worker_ida); |
3305 | free_workqueue_attrs(pool->attrs); | |
3306 | kfree(pool); | |
226223ab TH |
3307 | } |
3308 | ||
6ba94429 FW |
3309 | /** |
3310 | * put_unbound_pool - put a worker_pool | |
3311 | * @pool: worker_pool to put | |
3312 | * | |
3313 | * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU | |
3314 | * safe manner. get_unbound_pool() calls this function on its failure path | |
3315 | * and this function should be able to release pools which went through, | |
3316 | * successfully or not, init_worker_pool(). | |
3317 | * | |
3318 | * Should be called with wq_pool_mutex held. | |
3319 | */ | |
3320 | static void put_unbound_pool(struct worker_pool *pool) | |
226223ab | 3321 | { |
6ba94429 FW |
3322 | DECLARE_COMPLETION_ONSTACK(detach_completion); |
3323 | struct worker *worker; | |
226223ab | 3324 | |
6ba94429 | 3325 | lockdep_assert_held(&wq_pool_mutex); |
226223ab | 3326 | |
6ba94429 FW |
3327 | if (--pool->refcnt) |
3328 | return; | |
226223ab | 3329 | |
6ba94429 FW |
3330 | /* sanity checks */ |
3331 | if (WARN_ON(!(pool->cpu < 0)) || | |
3332 | WARN_ON(!list_empty(&pool->worklist))) | |
3333 | return; | |
226223ab | 3334 | |
6ba94429 FW |
3335 | /* release id and unhash */ |
3336 | if (pool->id >= 0) | |
3337 | idr_remove(&worker_pool_idr, pool->id); | |
3338 | hash_del(&pool->hash_node); | |
d55262c4 | 3339 | |
6ba94429 | 3340 | /* |
692b4825 TH |
3341 | * Become the manager and destroy all workers. This prevents |
3342 | * @pool's workers from blocking on attach_mutex. We're the last | |
3343 | * manager and @pool gets freed with the flag set. | |
6ba94429 | 3344 | */ |
6ba94429 | 3345 | spin_lock_irq(&pool->lock); |
692b4825 TH |
3346 | wait_event_lock_irq(wq_manager_wait, |
3347 | !(pool->flags & POOL_MANAGER_ACTIVE), pool->lock); | |
3348 | pool->flags |= POOL_MANAGER_ACTIVE; | |
3349 | ||
6ba94429 FW |
3350 | while ((worker = first_idle_worker(pool))) |
3351 | destroy_worker(worker); | |
3352 | WARN_ON(pool->nr_workers || pool->nr_idle); | |
3353 | spin_unlock_irq(&pool->lock); | |
d55262c4 | 3354 | |
6ba94429 FW |
3355 | mutex_lock(&pool->attach_mutex); |
3356 | if (!list_empty(&pool->workers)) | |
3357 | pool->detach_completion = &detach_completion; | |
3358 | mutex_unlock(&pool->attach_mutex); | |
226223ab | 3359 | |
6ba94429 FW |
3360 | if (pool->detach_completion) |
3361 | wait_for_completion(pool->detach_completion); | |
226223ab | 3362 | |
6ba94429 FW |
3363 | /* shut down the timers */ |
3364 | del_timer_sync(&pool->idle_timer); | |
3365 | del_timer_sync(&pool->mayday_timer); | |
226223ab | 3366 | |
6ba94429 FW |
3367 | /* sched-RCU protected to allow dereferences from get_work_pool() */ |
3368 | call_rcu_sched(&pool->rcu, rcu_free_pool); | |
226223ab TH |
3369 | } |
3370 | ||
3371 | /** | |
6ba94429 FW |
3372 | * get_unbound_pool - get a worker_pool with the specified attributes |
3373 | * @attrs: the attributes of the worker_pool to get | |
226223ab | 3374 | * |
6ba94429 FW |
3375 | * Obtain a worker_pool which has the same attributes as @attrs, bump the |
3376 | * reference count and return it. If there already is a matching | |
3377 | * worker_pool, it will be used; otherwise, this function attempts to | |
3378 | * create a new one. | |
226223ab | 3379 | * |
6ba94429 | 3380 | * Should be called with wq_pool_mutex held. |
226223ab | 3381 | * |
6ba94429 FW |
3382 | * Return: On success, a worker_pool with the same attributes as @attrs. |
3383 | * On failure, %NULL. | |
226223ab | 3384 | */ |
6ba94429 | 3385 | static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) |
226223ab | 3386 | { |
6ba94429 FW |
3387 | u32 hash = wqattrs_hash(attrs); |
3388 | struct worker_pool *pool; | |
3389 | int node; | |
e2273584 | 3390 | int target_node = NUMA_NO_NODE; |
226223ab | 3391 | |
6ba94429 | 3392 | lockdep_assert_held(&wq_pool_mutex); |
226223ab | 3393 | |
6ba94429 FW |
3394 | /* do we already have a matching pool? */ |
3395 | hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { | |
3396 | if (wqattrs_equal(pool->attrs, attrs)) { | |
3397 | pool->refcnt++; | |
3398 | return pool; | |
3399 | } | |
3400 | } | |
226223ab | 3401 | |
e2273584 XP |
3402 | /* if cpumask is contained inside a NUMA node, we belong to that node */ |
3403 | if (wq_numa_enabled) { | |
3404 | for_each_node(node) { | |
3405 | if (cpumask_subset(attrs->cpumask, | |
3406 | wq_numa_possible_cpumask[node])) { | |
3407 | target_node = node; | |
3408 | break; | |
3409 | } | |
3410 | } | |
3411 | } | |
3412 | ||
6ba94429 | 3413 | /* nope, create a new one */ |
e2273584 | 3414 | pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node); |
6ba94429 FW |
3415 | if (!pool || init_worker_pool(pool) < 0) |
3416 | goto fail; | |
3417 | ||
3418 | lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */ | |
3419 | copy_workqueue_attrs(pool->attrs, attrs); | |
e2273584 | 3420 | pool->node = target_node; |
226223ab TH |
3421 | |
3422 | /* | |
6ba94429 FW |
3423 | * no_numa isn't a worker_pool attribute, always clear it. See |
3424 | * 'struct workqueue_attrs' comments for detail. | |
226223ab | 3425 | */ |
6ba94429 | 3426 | pool->attrs->no_numa = false; |
226223ab | 3427 | |
6ba94429 FW |
3428 | if (worker_pool_assign_id(pool) < 0) |
3429 | goto fail; | |
226223ab | 3430 | |
6ba94429 | 3431 | /* create and start the initial worker */ |
3347fa09 | 3432 | if (wq_online && !create_worker(pool)) |
6ba94429 | 3433 | goto fail; |
226223ab | 3434 | |
6ba94429 FW |
3435 | /* install */ |
3436 | hash_add(unbound_pool_hash, &pool->hash_node, hash); | |
226223ab | 3437 | |
6ba94429 FW |
3438 | return pool; |
3439 | fail: | |
3440 | if (pool) | |
3441 | put_unbound_pool(pool); | |
3442 | return NULL; | |
226223ab | 3443 | } |
226223ab | 3444 | |
6ba94429 | 3445 | static void rcu_free_pwq(struct rcu_head *rcu) |
7a4e344c | 3446 | { |
6ba94429 FW |
3447 | kmem_cache_free(pwq_cache, |
3448 | container_of(rcu, struct pool_workqueue, rcu)); | |
7a4e344c TH |
3449 | } |
3450 | ||
6ba94429 FW |
3451 | /* |
3452 | * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt | |
3453 | * and needs to be destroyed. | |
7a4e344c | 3454 | */ |
6ba94429 | 3455 | static void pwq_unbound_release_workfn(struct work_struct *work) |
7a4e344c | 3456 | { |
6ba94429 FW |
3457 | struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, |
3458 | unbound_release_work); | |
3459 | struct workqueue_struct *wq = pwq->wq; | |
3460 | struct worker_pool *pool = pwq->pool; | |
3461 | bool is_last; | |
7a4e344c | 3462 | |
6ba94429 FW |
3463 | if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND))) |
3464 | return; | |
7a4e344c | 3465 | |
6ba94429 FW |
3466 | mutex_lock(&wq->mutex); |
3467 | list_del_rcu(&pwq->pwqs_node); | |
3468 | is_last = list_empty(&wq->pwqs); | |
3469 | mutex_unlock(&wq->mutex); | |
3470 | ||
3471 | mutex_lock(&wq_pool_mutex); | |
3472 | put_unbound_pool(pool); | |
3473 | mutex_unlock(&wq_pool_mutex); | |
3474 | ||
3475 | call_rcu_sched(&pwq->rcu, rcu_free_pwq); | |
7a4e344c | 3476 | |
2865a8fb | 3477 | /* |
6ba94429 FW |
3478 | * If we're the last pwq going away, @wq is already dead and no one |
3479 | * is gonna access it anymore. Schedule RCU free. | |
2865a8fb | 3480 | */ |
6ba94429 FW |
3481 | if (is_last) |
3482 | call_rcu_sched(&wq->rcu, rcu_free_wq); | |
29c91e99 TH |
3483 | } |
3484 | ||
7a4e344c | 3485 | /** |
6ba94429 FW |
3486 | * pwq_adjust_max_active - update a pwq's max_active to the current setting |
3487 | * @pwq: target pool_workqueue | |
d185af30 | 3488 | * |
6ba94429 FW |
3489 | * If @pwq isn't freezing, set @pwq->max_active to the associated |
3490 | * workqueue's saved_max_active and activate delayed work items | |
3491 | * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. | |
7a4e344c | 3492 | */ |
6ba94429 | 3493 | static void pwq_adjust_max_active(struct pool_workqueue *pwq) |
4e1a1f9a | 3494 | { |
6ba94429 FW |
3495 | struct workqueue_struct *wq = pwq->wq; |
3496 | bool freezable = wq->flags & WQ_FREEZABLE; | |
3347fa09 | 3497 | unsigned long flags; |
4e1a1f9a | 3498 | |
6ba94429 FW |
3499 | /* for @wq->saved_max_active */ |
3500 | lockdep_assert_held(&wq->mutex); | |
4e1a1f9a | 3501 | |
6ba94429 FW |
3502 | /* fast exit for non-freezable wqs */ |
3503 | if (!freezable && pwq->max_active == wq->saved_max_active) | |
3504 | return; | |
7a4e344c | 3505 | |
3347fa09 TH |
3506 | /* this function can be called during early boot w/ irq disabled */ |
3507 | spin_lock_irqsave(&pwq->pool->lock, flags); | |
29c91e99 | 3508 | |
6ba94429 FW |
3509 | /* |
3510 | * During [un]freezing, the caller is responsible for ensuring that | |
3511 | * this function is called at least once after @workqueue_freezing | |
3512 | * is updated and visible. | |
3513 | */ | |
3514 | if (!freezable || !workqueue_freezing) { | |
3515 | pwq->max_active = wq->saved_max_active; | |
4e1a1f9a | 3516 | |
6ba94429 FW |
3517 | while (!list_empty(&pwq->delayed_works) && |
3518 | pwq->nr_active < pwq->max_active) | |
3519 | pwq_activate_first_delayed(pwq); | |
e2dca7ad | 3520 | |
6ba94429 FW |
3521 | /* |
3522 | * Need to kick a worker after thawed or an unbound wq's | |
3523 | * max_active is bumped. It's a slow path. Do it always. | |
3524 | */ | |
3525 | wake_up_worker(pwq->pool); | |
3526 | } else { | |
3527 | pwq->max_active = 0; | |
3528 | } | |
e2dca7ad | 3529 | |
3347fa09 | 3530 | spin_unlock_irqrestore(&pwq->pool->lock, flags); |
e2dca7ad TH |
3531 | } |
3532 | ||
6ba94429 FW |
3533 | /* initialize newly alloced @pwq which is associated with @wq and @pool */ |
3534 | static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, | |
3535 | struct worker_pool *pool) | |
29c91e99 | 3536 | { |
6ba94429 | 3537 | BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); |
29c91e99 | 3538 | |
6ba94429 FW |
3539 | memset(pwq, 0, sizeof(*pwq)); |
3540 | ||
3541 | pwq->pool = pool; | |
3542 | pwq->wq = wq; | |
3543 | pwq->flush_color = -1; | |
3544 | pwq->refcnt = 1; | |
3545 | INIT_LIST_HEAD(&pwq->delayed_works); | |
3546 | INIT_LIST_HEAD(&pwq->pwqs_node); | |
3547 | INIT_LIST_HEAD(&pwq->mayday_node); | |
3548 | INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn); | |
29c91e99 TH |
3549 | } |
3550 | ||
6ba94429 FW |
3551 | /* sync @pwq with the current state of its associated wq and link it */ |
3552 | static void link_pwq(struct pool_workqueue *pwq) | |
29c91e99 | 3553 | { |
6ba94429 | 3554 | struct workqueue_struct *wq = pwq->wq; |
29c91e99 | 3555 | |
6ba94429 | 3556 | lockdep_assert_held(&wq->mutex); |
a892cacc | 3557 | |
6ba94429 FW |
3558 | /* may be called multiple times, ignore if already linked */ |
3559 | if (!list_empty(&pwq->pwqs_node)) | |
29c91e99 | 3560 | return; |
29c91e99 | 3561 | |
6ba94429 FW |
3562 | /* set the matching work_color */ |
3563 | pwq->work_color = wq->work_color; | |
29c91e99 | 3564 | |
6ba94429 FW |
3565 | /* sync max_active to the current setting */ |
3566 | pwq_adjust_max_active(pwq); | |
29c91e99 | 3567 | |
6ba94429 FW |
3568 | /* link in @pwq */ |
3569 | list_add_rcu(&pwq->pwqs_node, &wq->pwqs); | |
3570 | } | |
29c91e99 | 3571 | |
6ba94429 FW |
3572 | /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ |
3573 | static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, | |
3574 | const struct workqueue_attrs *attrs) | |
3575 | { | |
3576 | struct worker_pool *pool; | |
3577 | struct pool_workqueue *pwq; | |
60f5a4bc | 3578 | |
6ba94429 | 3579 | lockdep_assert_held(&wq_pool_mutex); |
60f5a4bc | 3580 | |
6ba94429 FW |
3581 | pool = get_unbound_pool(attrs); |
3582 | if (!pool) | |
3583 | return NULL; | |
60f5a4bc | 3584 | |
6ba94429 FW |
3585 | pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); |
3586 | if (!pwq) { | |
3587 | put_unbound_pool(pool); | |
3588 | return NULL; | |
3589 | } | |
29c91e99 | 3590 | |
6ba94429 FW |
3591 | init_pwq(pwq, wq, pool); |
3592 | return pwq; | |
3593 | } | |
29c91e99 | 3594 | |
29c91e99 | 3595 | /** |
30186c6f | 3596 | * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node |
042f7df1 | 3597 | * @attrs: the wq_attrs of the default pwq of the target workqueue |
6ba94429 FW |
3598 | * @node: the target NUMA node |
3599 | * @cpu_going_down: if >= 0, the CPU to consider as offline | |
3600 | * @cpumask: outarg, the resulting cpumask | |
29c91e99 | 3601 | * |
6ba94429 FW |
3602 | * Calculate the cpumask a workqueue with @attrs should use on @node. If |
3603 | * @cpu_going_down is >= 0, that cpu is considered offline during | |
3604 | * calculation. The result is stored in @cpumask. | |
a892cacc | 3605 | * |
6ba94429 FW |
3606 | * If NUMA affinity is not enabled, @attrs->cpumask is always used. If |
3607 | * enabled and @node has online CPUs requested by @attrs, the returned | |
3608 | * cpumask is the intersection of the possible CPUs of @node and | |
3609 | * @attrs->cpumask. | |
d185af30 | 3610 | * |
6ba94429 FW |
3611 | * The caller is responsible for ensuring that the cpumask of @node stays |
3612 | * stable. | |
3613 | * | |
3614 | * Return: %true if the resulting @cpumask is different from @attrs->cpumask, | |
3615 | * %false if equal. | |
29c91e99 | 3616 | */ |
6ba94429 FW |
3617 | static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node, |
3618 | int cpu_going_down, cpumask_t *cpumask) | |
29c91e99 | 3619 | { |
6ba94429 FW |
3620 | if (!wq_numa_enabled || attrs->no_numa) |
3621 | goto use_dfl; | |
29c91e99 | 3622 | |
6ba94429 FW |
3623 | /* does @node have any online CPUs @attrs wants? */ |
3624 | cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask); | |
3625 | if (cpu_going_down >= 0) | |
3626 | cpumask_clear_cpu(cpu_going_down, cpumask); | |
29c91e99 | 3627 | |
6ba94429 FW |
3628 | if (cpumask_empty(cpumask)) |
3629 | goto use_dfl; | |
4c16bd32 TH |
3630 | |
3631 | /* yeap, return possible CPUs in @node that @attrs wants */ | |
3632 | cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]); | |
1ad0f0a7 MB |
3633 | |
3634 | if (cpumask_empty(cpumask)) { | |
3635 | pr_warn_once("WARNING: workqueue cpumask: online intersect > " | |
3636 | "possible intersect\n"); | |
3637 | return false; | |
3638 | } | |
3639 | ||
4c16bd32 TH |
3640 | return !cpumask_equal(cpumask, attrs->cpumask); |
3641 | ||
3642 | use_dfl: | |
3643 | cpumask_copy(cpumask, attrs->cpumask); | |
3644 | return false; | |
3645 | } | |
3646 | ||
1befcf30 TH |
3647 | /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */ |
3648 | static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq, | |
3649 | int node, | |
3650 | struct pool_workqueue *pwq) | |
3651 | { | |
3652 | struct pool_workqueue *old_pwq; | |
3653 | ||
5b95e1af | 3654 | lockdep_assert_held(&wq_pool_mutex); |
1befcf30 TH |
3655 | lockdep_assert_held(&wq->mutex); |
3656 | ||
3657 | /* link_pwq() can handle duplicate calls */ | |
3658 | link_pwq(pwq); | |
3659 | ||
3660 | old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); | |
3661 | rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq); | |
3662 | return old_pwq; | |
3663 | } | |
3664 | ||
2d5f0764 LJ |
3665 | /* context to store the prepared attrs & pwqs before applying */ |
3666 | struct apply_wqattrs_ctx { | |
3667 | struct workqueue_struct *wq; /* target workqueue */ | |
3668 | struct workqueue_attrs *attrs; /* attrs to apply */ | |
042f7df1 | 3669 | struct list_head list; /* queued for batching commit */ |
2d5f0764 LJ |
3670 | struct pool_workqueue *dfl_pwq; |
3671 | struct pool_workqueue *pwq_tbl[]; | |
3672 | }; | |
3673 | ||
3674 | /* free the resources after success or abort */ | |
3675 | static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx) | |
3676 | { | |
3677 | if (ctx) { | |
3678 | int node; | |
3679 | ||
3680 | for_each_node(node) | |
3681 | put_pwq_unlocked(ctx->pwq_tbl[node]); | |
3682 | put_pwq_unlocked(ctx->dfl_pwq); | |
3683 | ||
3684 | free_workqueue_attrs(ctx->attrs); | |
3685 | ||
3686 | kfree(ctx); | |
3687 | } | |
3688 | } | |
3689 | ||
3690 | /* allocate the attrs and pwqs for later installation */ | |
3691 | static struct apply_wqattrs_ctx * | |
3692 | apply_wqattrs_prepare(struct workqueue_struct *wq, | |
3693 | const struct workqueue_attrs *attrs) | |
9e8cd2f5 | 3694 | { |
2d5f0764 | 3695 | struct apply_wqattrs_ctx *ctx; |
4c16bd32 | 3696 | struct workqueue_attrs *new_attrs, *tmp_attrs; |
2d5f0764 | 3697 | int node; |
9e8cd2f5 | 3698 | |
2d5f0764 | 3699 | lockdep_assert_held(&wq_pool_mutex); |
9e8cd2f5 | 3700 | |
2d5f0764 LJ |
3701 | ctx = kzalloc(sizeof(*ctx) + nr_node_ids * sizeof(ctx->pwq_tbl[0]), |
3702 | GFP_KERNEL); | |
8719dcea | 3703 | |
13e2e556 | 3704 | new_attrs = alloc_workqueue_attrs(GFP_KERNEL); |
4c16bd32 | 3705 | tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL); |
2d5f0764 LJ |
3706 | if (!ctx || !new_attrs || !tmp_attrs) |
3707 | goto out_free; | |
13e2e556 | 3708 | |
042f7df1 LJ |
3709 | /* |
3710 | * Calculate the attrs of the default pwq. | |
3711 | * If the user configured cpumask doesn't overlap with the | |
3712 | * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask. | |
3713 | */ | |
13e2e556 | 3714 | copy_workqueue_attrs(new_attrs, attrs); |
b05a7928 | 3715 | cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask); |
042f7df1 LJ |
3716 | if (unlikely(cpumask_empty(new_attrs->cpumask))) |
3717 | cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask); | |
13e2e556 | 3718 | |
4c16bd32 TH |
3719 | /* |
3720 | * We may create multiple pwqs with differing cpumasks. Make a | |
3721 | * copy of @new_attrs which will be modified and used to obtain | |
3722 | * pools. | |
3723 | */ | |
3724 | copy_workqueue_attrs(tmp_attrs, new_attrs); | |
3725 | ||
4c16bd32 TH |
3726 | /* |
3727 | * If something goes wrong during CPU up/down, we'll fall back to | |
3728 | * the default pwq covering whole @attrs->cpumask. Always create | |
3729 | * it even if we don't use it immediately. | |
3730 | */ | |
2d5f0764 LJ |
3731 | ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs); |
3732 | if (!ctx->dfl_pwq) | |
3733 | goto out_free; | |
4c16bd32 TH |
3734 | |
3735 | for_each_node(node) { | |
042f7df1 | 3736 | if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) { |
2d5f0764 LJ |
3737 | ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs); |
3738 | if (!ctx->pwq_tbl[node]) | |
3739 | goto out_free; | |
4c16bd32 | 3740 | } else { |
2d5f0764 LJ |
3741 | ctx->dfl_pwq->refcnt++; |
3742 | ctx->pwq_tbl[node] = ctx->dfl_pwq; | |
4c16bd32 TH |
3743 | } |
3744 | } | |
3745 | ||
042f7df1 LJ |
3746 | /* save the user configured attrs and sanitize it. */ |
3747 | copy_workqueue_attrs(new_attrs, attrs); | |
3748 | cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); | |
2d5f0764 | 3749 | ctx->attrs = new_attrs; |
042f7df1 | 3750 | |
2d5f0764 LJ |
3751 | ctx->wq = wq; |
3752 | free_workqueue_attrs(tmp_attrs); | |
3753 | return ctx; | |
3754 | ||
3755 | out_free: | |
3756 | free_workqueue_attrs(tmp_attrs); | |
3757 | free_workqueue_attrs(new_attrs); | |
3758 | apply_wqattrs_cleanup(ctx); | |
3759 | return NULL; | |
3760 | } | |
3761 | ||
3762 | /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */ | |
3763 | static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx) | |
3764 | { | |
3765 | int node; | |
9e8cd2f5 | 3766 | |
4c16bd32 | 3767 | /* all pwqs have been created successfully, let's install'em */ |
2d5f0764 | 3768 | mutex_lock(&ctx->wq->mutex); |
a892cacc | 3769 | |
2d5f0764 | 3770 | copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs); |
4c16bd32 TH |
3771 | |
3772 | /* save the previous pwq and install the new one */ | |
f147f29e | 3773 | for_each_node(node) |
2d5f0764 LJ |
3774 | ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node, |
3775 | ctx->pwq_tbl[node]); | |
4c16bd32 TH |
3776 | |
3777 | /* @dfl_pwq might not have been used, ensure it's linked */ | |
2d5f0764 LJ |
3778 | link_pwq(ctx->dfl_pwq); |
3779 | swap(ctx->wq->dfl_pwq, ctx->dfl_pwq); | |
f147f29e | 3780 | |
2d5f0764 LJ |
3781 | mutex_unlock(&ctx->wq->mutex); |
3782 | } | |
9e8cd2f5 | 3783 | |
a0111cf6 LJ |
3784 | static void apply_wqattrs_lock(void) |
3785 | { | |
3786 | /* CPUs should stay stable across pwq creations and installations */ | |
3787 | get_online_cpus(); | |
3788 | mutex_lock(&wq_pool_mutex); | |
3789 | } | |
3790 | ||
3791 | static void apply_wqattrs_unlock(void) | |
3792 | { | |
3793 | mutex_unlock(&wq_pool_mutex); | |
3794 | put_online_cpus(); | |
3795 | } | |
3796 | ||
3797 | static int apply_workqueue_attrs_locked(struct workqueue_struct *wq, | |
3798 | const struct workqueue_attrs *attrs) | |
2d5f0764 LJ |
3799 | { |
3800 | struct apply_wqattrs_ctx *ctx; | |
4c16bd32 | 3801 | |
2d5f0764 LJ |
3802 | /* only unbound workqueues can change attributes */ |
3803 | if (WARN_ON(!(wq->flags & WQ_UNBOUND))) | |
3804 | return -EINVAL; | |
13e2e556 | 3805 | |
2d5f0764 | 3806 | /* creating multiple pwqs breaks ordering guarantee */ |
0a94efb5 TH |
3807 | if (!list_empty(&wq->pwqs)) { |
3808 | if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) | |
3809 | return -EINVAL; | |
3810 | ||
3811 | wq->flags &= ~__WQ_ORDERED; | |
3812 | } | |
2d5f0764 | 3813 | |
2d5f0764 | 3814 | ctx = apply_wqattrs_prepare(wq, attrs); |
6201171e | 3815 | if (!ctx) |
3816 | return -ENOMEM; | |
2d5f0764 LJ |
3817 | |
3818 | /* the ctx has been prepared successfully, let's commit it */ | |
6201171e | 3819 | apply_wqattrs_commit(ctx); |
2d5f0764 LJ |
3820 | apply_wqattrs_cleanup(ctx); |
3821 | ||
6201171e | 3822 | return 0; |
9e8cd2f5 TH |
3823 | } |
3824 | ||
a0111cf6 LJ |
3825 | /** |
3826 | * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue | |
3827 | * @wq: the target workqueue | |
3828 | * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() | |
3829 | * | |
3830 | * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA | |
3831 | * machines, this function maps a separate pwq to each NUMA node with | |
3832 | * possibles CPUs in @attrs->cpumask so that work items are affine to the | |
3833 | * NUMA node it was issued on. Older pwqs are released as in-flight work | |
3834 | * items finish. Note that a work item which repeatedly requeues itself | |
3835 | * back-to-back will stay on its current pwq. | |
3836 | * | |
3837 | * Performs GFP_KERNEL allocations. | |
3838 | * | |
3839 | * Return: 0 on success and -errno on failure. | |
3840 | */ | |
3841 | int apply_workqueue_attrs(struct workqueue_struct *wq, | |
3842 | const struct workqueue_attrs *attrs) | |
3843 | { | |
3844 | int ret; | |
3845 | ||
3846 | apply_wqattrs_lock(); | |
3847 | ret = apply_workqueue_attrs_locked(wq, attrs); | |
3848 | apply_wqattrs_unlock(); | |
3849 | ||
3850 | return ret; | |
3851 | } | |
3852 | ||
4c16bd32 TH |
3853 | /** |
3854 | * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug | |
3855 | * @wq: the target workqueue | |
3856 | * @cpu: the CPU coming up or going down | |
3857 | * @online: whether @cpu is coming up or going down | |
3858 | * | |
3859 | * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and | |
3860 | * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of | |
3861 | * @wq accordingly. | |
3862 | * | |
3863 | * If NUMA affinity can't be adjusted due to memory allocation failure, it | |
3864 | * falls back to @wq->dfl_pwq which may not be optimal but is always | |
3865 | * correct. | |
3866 | * | |
3867 | * Note that when the last allowed CPU of a NUMA node goes offline for a | |
3868 | * workqueue with a cpumask spanning multiple nodes, the workers which were | |
3869 | * already executing the work items for the workqueue will lose their CPU | |
3870 | * affinity and may execute on any CPU. This is similar to how per-cpu | |
3871 | * workqueues behave on CPU_DOWN. If a workqueue user wants strict | |
3872 | * affinity, it's the user's responsibility to flush the work item from | |
3873 | * CPU_DOWN_PREPARE. | |
3874 | */ | |
3875 | static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu, | |
3876 | bool online) | |
3877 | { | |
3878 | int node = cpu_to_node(cpu); | |
3879 | int cpu_off = online ? -1 : cpu; | |
3880 | struct pool_workqueue *old_pwq = NULL, *pwq; | |
3881 | struct workqueue_attrs *target_attrs; | |
3882 | cpumask_t *cpumask; | |
3883 | ||
3884 | lockdep_assert_held(&wq_pool_mutex); | |
3885 | ||
f7142ed4 LJ |
3886 | if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) || |
3887 | wq->unbound_attrs->no_numa) | |
4c16bd32 TH |
3888 | return; |
3889 | ||
3890 | /* | |
3891 | * We don't wanna alloc/free wq_attrs for each wq for each CPU. | |
3892 | * Let's use a preallocated one. The following buf is protected by | |
3893 | * CPU hotplug exclusion. | |
3894 | */ | |
3895 | target_attrs = wq_update_unbound_numa_attrs_buf; | |
3896 | cpumask = target_attrs->cpumask; | |
3897 | ||
4c16bd32 TH |
3898 | copy_workqueue_attrs(target_attrs, wq->unbound_attrs); |
3899 | pwq = unbound_pwq_by_node(wq, node); | |
3900 | ||
3901 | /* | |
3902 | * Let's determine what needs to be done. If the target cpumask is | |
042f7df1 LJ |
3903 | * different from the default pwq's, we need to compare it to @pwq's |
3904 | * and create a new one if they don't match. If the target cpumask | |
3905 | * equals the default pwq's, the default pwq should be used. | |
4c16bd32 | 3906 | */ |
042f7df1 | 3907 | if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) { |
4c16bd32 | 3908 | if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask)) |
f7142ed4 | 3909 | return; |
4c16bd32 | 3910 | } else { |
534a3fbb | 3911 | goto use_dfl_pwq; |
4c16bd32 TH |
3912 | } |
3913 | ||
4c16bd32 TH |
3914 | /* create a new pwq */ |
3915 | pwq = alloc_unbound_pwq(wq, target_attrs); | |
3916 | if (!pwq) { | |
2d916033 FF |
3917 | pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n", |
3918 | wq->name); | |
77f300b1 | 3919 | goto use_dfl_pwq; |
4c16bd32 TH |
3920 | } |
3921 | ||
f7142ed4 | 3922 | /* Install the new pwq. */ |
4c16bd32 TH |
3923 | mutex_lock(&wq->mutex); |
3924 | old_pwq = numa_pwq_tbl_install(wq, node, pwq); | |
3925 | goto out_unlock; | |
3926 | ||
3927 | use_dfl_pwq: | |
f7142ed4 | 3928 | mutex_lock(&wq->mutex); |
4c16bd32 TH |
3929 | spin_lock_irq(&wq->dfl_pwq->pool->lock); |
3930 | get_pwq(wq->dfl_pwq); | |
3931 | spin_unlock_irq(&wq->dfl_pwq->pool->lock); | |
3932 | old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq); | |
3933 | out_unlock: | |
3934 | mutex_unlock(&wq->mutex); | |
3935 | put_pwq_unlocked(old_pwq); | |
3936 | } | |
3937 | ||
30cdf249 | 3938 | static int alloc_and_link_pwqs(struct workqueue_struct *wq) |
0f900049 | 3939 | { |
49e3cf44 | 3940 | bool highpri = wq->flags & WQ_HIGHPRI; |
8a2b7538 | 3941 | int cpu, ret; |
30cdf249 TH |
3942 | |
3943 | if (!(wq->flags & WQ_UNBOUND)) { | |
420c0ddb TH |
3944 | wq->cpu_pwqs = alloc_percpu(struct pool_workqueue); |
3945 | if (!wq->cpu_pwqs) | |
30cdf249 TH |
3946 | return -ENOMEM; |
3947 | ||
3948 | for_each_possible_cpu(cpu) { | |
7fb98ea7 TH |
3949 | struct pool_workqueue *pwq = |
3950 | per_cpu_ptr(wq->cpu_pwqs, cpu); | |
7a62c2c8 | 3951 | struct worker_pool *cpu_pools = |
f02ae73a | 3952 | per_cpu(cpu_worker_pools, cpu); |
f3421797 | 3953 | |
f147f29e TH |
3954 | init_pwq(pwq, wq, &cpu_pools[highpri]); |
3955 | ||
3956 | mutex_lock(&wq->mutex); | |
1befcf30 | 3957 | link_pwq(pwq); |
f147f29e | 3958 | mutex_unlock(&wq->mutex); |
30cdf249 | 3959 | } |
9e8cd2f5 | 3960 | return 0; |
8a2b7538 TH |
3961 | } else if (wq->flags & __WQ_ORDERED) { |
3962 | ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]); | |
3963 | /* there should only be single pwq for ordering guarantee */ | |
3964 | WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node || | |
3965 | wq->pwqs.prev != &wq->dfl_pwq->pwqs_node), | |
3966 | "ordering guarantee broken for workqueue %s\n", wq->name); | |
3967 | return ret; | |
30cdf249 | 3968 | } else { |
9e8cd2f5 | 3969 | return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); |
30cdf249 | 3970 | } |
0f900049 TH |
3971 | } |
3972 | ||
f3421797 TH |
3973 | static int wq_clamp_max_active(int max_active, unsigned int flags, |
3974 | const char *name) | |
b71ab8c2 | 3975 | { |
f3421797 TH |
3976 | int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; |
3977 | ||
3978 | if (max_active < 1 || max_active > lim) | |
044c782c VI |
3979 | pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", |
3980 | max_active, name, 1, lim); | |
b71ab8c2 | 3981 | |
f3421797 | 3982 | return clamp_val(max_active, 1, lim); |
b71ab8c2 TH |
3983 | } |
3984 | ||
b196be89 | 3985 | struct workqueue_struct *__alloc_workqueue_key(const char *fmt, |
d320c038 TH |
3986 | unsigned int flags, |
3987 | int max_active, | |
3988 | struct lock_class_key *key, | |
b196be89 | 3989 | const char *lock_name, ...) |
1da177e4 | 3990 | { |
df2d5ae4 | 3991 | size_t tbl_size = 0; |
ecf6881f | 3992 | va_list args; |
1da177e4 | 3993 | struct workqueue_struct *wq; |
49e3cf44 | 3994 | struct pool_workqueue *pwq; |
b196be89 | 3995 | |
5c0338c6 TH |
3996 | /* |
3997 | * Unbound && max_active == 1 used to imply ordered, which is no | |
3998 | * longer the case on NUMA machines due to per-node pools. While | |
3999 | * alloc_ordered_workqueue() is the right way to create an ordered | |
4000 | * workqueue, keep the previous behavior to avoid subtle breakages | |
4001 | * on NUMA. | |
4002 | */ | |
4003 | if ((flags & WQ_UNBOUND) && max_active == 1) | |
4004 | flags |= __WQ_ORDERED; | |
4005 | ||
cee22a15 VK |
4006 | /* see the comment above the definition of WQ_POWER_EFFICIENT */ |
4007 | if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) | |
4008 | flags |= WQ_UNBOUND; | |
4009 | ||
ecf6881f | 4010 | /* allocate wq and format name */ |
df2d5ae4 | 4011 | if (flags & WQ_UNBOUND) |
ddcb57e2 | 4012 | tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]); |
df2d5ae4 TH |
4013 | |
4014 | wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL); | |
b196be89 | 4015 | if (!wq) |
d2c1d404 | 4016 | return NULL; |
b196be89 | 4017 | |
6029a918 TH |
4018 | if (flags & WQ_UNBOUND) { |
4019 | wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL); | |
4020 | if (!wq->unbound_attrs) | |
4021 | goto err_free_wq; | |
4022 | } | |
4023 | ||
ecf6881f TH |
4024 | va_start(args, lock_name); |
4025 | vsnprintf(wq->name, sizeof(wq->name), fmt, args); | |
b196be89 | 4026 | va_end(args); |
1da177e4 | 4027 | |
d320c038 | 4028 | max_active = max_active ?: WQ_DFL_ACTIVE; |
b196be89 | 4029 | max_active = wq_clamp_max_active(max_active, flags, wq->name); |
3af24433 | 4030 | |
b196be89 | 4031 | /* init wq */ |
97e37d7b | 4032 | wq->flags = flags; |
a0a1a5fd | 4033 | wq->saved_max_active = max_active; |
3c25a55d | 4034 | mutex_init(&wq->mutex); |
112202d9 | 4035 | atomic_set(&wq->nr_pwqs_to_flush, 0); |
30cdf249 | 4036 | INIT_LIST_HEAD(&wq->pwqs); |
73f53c4a TH |
4037 | INIT_LIST_HEAD(&wq->flusher_queue); |
4038 | INIT_LIST_HEAD(&wq->flusher_overflow); | |
493a1724 | 4039 | INIT_LIST_HEAD(&wq->maydays); |
502ca9d8 | 4040 | |
eb13ba87 | 4041 | lockdep_init_map(&wq->lockdep_map, lock_name, key, 0); |
cce1a165 | 4042 | INIT_LIST_HEAD(&wq->list); |
3af24433 | 4043 | |
30cdf249 | 4044 | if (alloc_and_link_pwqs(wq) < 0) |
d2c1d404 | 4045 | goto err_free_wq; |
1537663f | 4046 | |
493008a8 TH |
4047 | /* |
4048 | * Workqueues which may be used during memory reclaim should | |
4049 | * have a rescuer to guarantee forward progress. | |
4050 | */ | |
4051 | if (flags & WQ_MEM_RECLAIM) { | |
e22bee78 TH |
4052 | struct worker *rescuer; |
4053 | ||
f7537df5 | 4054 | rescuer = alloc_worker(NUMA_NO_NODE); |
e22bee78 | 4055 | if (!rescuer) |
d2c1d404 | 4056 | goto err_destroy; |
e22bee78 | 4057 | |
111c225a TH |
4058 | rescuer->rescue_wq = wq; |
4059 | rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", | |
b196be89 | 4060 | wq->name); |
d2c1d404 TH |
4061 | if (IS_ERR(rescuer->task)) { |
4062 | kfree(rescuer); | |
4063 | goto err_destroy; | |
4064 | } | |
e22bee78 | 4065 | |
d2c1d404 | 4066 | wq->rescuer = rescuer; |
25834c73 | 4067 | kthread_bind_mask(rescuer->task, cpu_possible_mask); |
e22bee78 | 4068 | wake_up_process(rescuer->task); |
3af24433 ON |
4069 | } |
4070 | ||
226223ab TH |
4071 | if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) |
4072 | goto err_destroy; | |
4073 | ||
a0a1a5fd | 4074 | /* |
68e13a67 LJ |
4075 | * wq_pool_mutex protects global freeze state and workqueues list. |
4076 | * Grab it, adjust max_active and add the new @wq to workqueues | |
4077 | * list. | |
a0a1a5fd | 4078 | */ |
68e13a67 | 4079 | mutex_lock(&wq_pool_mutex); |
a0a1a5fd | 4080 | |
a357fc03 | 4081 | mutex_lock(&wq->mutex); |
699ce097 TH |
4082 | for_each_pwq(pwq, wq) |
4083 | pwq_adjust_max_active(pwq); | |
a357fc03 | 4084 | mutex_unlock(&wq->mutex); |
a0a1a5fd | 4085 | |
e2dca7ad | 4086 | list_add_tail_rcu(&wq->list, &workqueues); |
a0a1a5fd | 4087 | |
68e13a67 | 4088 | mutex_unlock(&wq_pool_mutex); |
1537663f | 4089 | |
3af24433 | 4090 | return wq; |
d2c1d404 TH |
4091 | |
4092 | err_free_wq: | |
6029a918 | 4093 | free_workqueue_attrs(wq->unbound_attrs); |
d2c1d404 TH |
4094 | kfree(wq); |
4095 | return NULL; | |
4096 | err_destroy: | |
4097 | destroy_workqueue(wq); | |
4690c4ab | 4098 | return NULL; |
3af24433 | 4099 | } |
d320c038 | 4100 | EXPORT_SYMBOL_GPL(__alloc_workqueue_key); |
1da177e4 | 4101 | |
3af24433 ON |
4102 | /** |
4103 | * destroy_workqueue - safely terminate a workqueue | |
4104 | * @wq: target workqueue | |
4105 | * | |
4106 | * Safely destroy a workqueue. All work currently pending will be done first. | |
4107 | */ | |
4108 | void destroy_workqueue(struct workqueue_struct *wq) | |
4109 | { | |
49e3cf44 | 4110 | struct pool_workqueue *pwq; |
4c16bd32 | 4111 | int node; |
3af24433 | 4112 | |
9c5a2ba7 TH |
4113 | /* drain it before proceeding with destruction */ |
4114 | drain_workqueue(wq); | |
c8efcc25 | 4115 | |
6183c009 | 4116 | /* sanity checks */ |
b09f4fd3 | 4117 | mutex_lock(&wq->mutex); |
49e3cf44 | 4118 | for_each_pwq(pwq, wq) { |
6183c009 TH |
4119 | int i; |
4120 | ||
76af4d93 TH |
4121 | for (i = 0; i < WORK_NR_COLORS; i++) { |
4122 | if (WARN_ON(pwq->nr_in_flight[i])) { | |
b09f4fd3 | 4123 | mutex_unlock(&wq->mutex); |
fa07fb6a | 4124 | show_workqueue_state(); |
6183c009 | 4125 | return; |
76af4d93 TH |
4126 | } |
4127 | } | |
4128 | ||
5c529597 | 4129 | if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) || |
8864b4e5 | 4130 | WARN_ON(pwq->nr_active) || |
76af4d93 | 4131 | WARN_ON(!list_empty(&pwq->delayed_works))) { |
b09f4fd3 | 4132 | mutex_unlock(&wq->mutex); |
fa07fb6a | 4133 | show_workqueue_state(); |
6183c009 | 4134 | return; |
76af4d93 | 4135 | } |
6183c009 | 4136 | } |
b09f4fd3 | 4137 | mutex_unlock(&wq->mutex); |
6183c009 | 4138 | |
a0a1a5fd TH |
4139 | /* |
4140 | * wq list is used to freeze wq, remove from list after | |
4141 | * flushing is complete in case freeze races us. | |
4142 | */ | |
68e13a67 | 4143 | mutex_lock(&wq_pool_mutex); |
e2dca7ad | 4144 | list_del_rcu(&wq->list); |
68e13a67 | 4145 | mutex_unlock(&wq_pool_mutex); |
3af24433 | 4146 | |
226223ab TH |
4147 | workqueue_sysfs_unregister(wq); |
4148 | ||
e2dca7ad | 4149 | if (wq->rescuer) |
e22bee78 | 4150 | kthread_stop(wq->rescuer->task); |
e22bee78 | 4151 | |
8864b4e5 TH |
4152 | if (!(wq->flags & WQ_UNBOUND)) { |
4153 | /* | |
4154 | * The base ref is never dropped on per-cpu pwqs. Directly | |
e2dca7ad | 4155 | * schedule RCU free. |
8864b4e5 | 4156 | */ |
e2dca7ad | 4157 | call_rcu_sched(&wq->rcu, rcu_free_wq); |
8864b4e5 TH |
4158 | } else { |
4159 | /* | |
4160 | * We're the sole accessor of @wq at this point. Directly | |
4c16bd32 TH |
4161 | * access numa_pwq_tbl[] and dfl_pwq to put the base refs. |
4162 | * @wq will be freed when the last pwq is released. | |
8864b4e5 | 4163 | */ |
4c16bd32 TH |
4164 | for_each_node(node) { |
4165 | pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); | |
4166 | RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL); | |
4167 | put_pwq_unlocked(pwq); | |
4168 | } | |
4169 | ||
4170 | /* | |
4171 | * Put dfl_pwq. @wq may be freed any time after dfl_pwq is | |
4172 | * put. Don't access it afterwards. | |
4173 | */ | |
4174 | pwq = wq->dfl_pwq; | |
4175 | wq->dfl_pwq = NULL; | |
dce90d47 | 4176 | put_pwq_unlocked(pwq); |
29c91e99 | 4177 | } |
3af24433 ON |
4178 | } |
4179 | EXPORT_SYMBOL_GPL(destroy_workqueue); | |
4180 | ||
dcd989cb TH |
4181 | /** |
4182 | * workqueue_set_max_active - adjust max_active of a workqueue | |
4183 | * @wq: target workqueue | |
4184 | * @max_active: new max_active value. | |
4185 | * | |
4186 | * Set max_active of @wq to @max_active. | |
4187 | * | |
4188 | * CONTEXT: | |
4189 | * Don't call from IRQ context. | |
4190 | */ | |
4191 | void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) | |
4192 | { | |
49e3cf44 | 4193 | struct pool_workqueue *pwq; |
dcd989cb | 4194 | |
8719dcea | 4195 | /* disallow meddling with max_active for ordered workqueues */ |
0a94efb5 | 4196 | if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) |
8719dcea TH |
4197 | return; |
4198 | ||
f3421797 | 4199 | max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); |
dcd989cb | 4200 | |
a357fc03 | 4201 | mutex_lock(&wq->mutex); |
dcd989cb | 4202 | |
0a94efb5 | 4203 | wq->flags &= ~__WQ_ORDERED; |
dcd989cb TH |
4204 | wq->saved_max_active = max_active; |
4205 | ||
699ce097 TH |
4206 | for_each_pwq(pwq, wq) |
4207 | pwq_adjust_max_active(pwq); | |
93981800 | 4208 | |
a357fc03 | 4209 | mutex_unlock(&wq->mutex); |
15316ba8 | 4210 | } |
dcd989cb | 4211 | EXPORT_SYMBOL_GPL(workqueue_set_max_active); |
15316ba8 | 4212 | |
363e3fd5 LW |
4213 | /** |
4214 | * current_work - retrieve %current task's work struct | |
4215 | * | |
4216 | * Determine if %current task is a workqueue worker and what it's working on. | |
4217 | * Useful to find out the context that the %current task is running in. | |
4218 | * | |
4219 | * Return: work struct if %current task is a workqueue worker, %NULL otherwise. | |
4220 | */ | |
4221 | struct work_struct *current_work(void) | |
4222 | { | |
4223 | struct worker *worker = current_wq_worker(); | |
4224 | ||
4225 | return worker ? worker->current_work : NULL; | |
4226 | } | |
4227 | EXPORT_SYMBOL(current_work); | |
4228 | ||
e6267616 TH |
4229 | /** |
4230 | * current_is_workqueue_rescuer - is %current workqueue rescuer? | |
4231 | * | |
4232 | * Determine whether %current is a workqueue rescuer. Can be used from | |
4233 | * work functions to determine whether it's being run off the rescuer task. | |
d185af30 YB |
4234 | * |
4235 | * Return: %true if %current is a workqueue rescuer. %false otherwise. | |
e6267616 TH |
4236 | */ |
4237 | bool current_is_workqueue_rescuer(void) | |
4238 | { | |
4239 | struct worker *worker = current_wq_worker(); | |
4240 | ||
6a092dfd | 4241 | return worker && worker->rescue_wq; |
e6267616 TH |
4242 | } |
4243 | ||
eef6a7d5 | 4244 | /** |
dcd989cb TH |
4245 | * workqueue_congested - test whether a workqueue is congested |
4246 | * @cpu: CPU in question | |
4247 | * @wq: target workqueue | |
eef6a7d5 | 4248 | * |
dcd989cb TH |
4249 | * Test whether @wq's cpu workqueue for @cpu is congested. There is |
4250 | * no synchronization around this function and the test result is | |
4251 | * unreliable and only useful as advisory hints or for debugging. | |
eef6a7d5 | 4252 | * |
d3251859 TH |
4253 | * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. |
4254 | * Note that both per-cpu and unbound workqueues may be associated with | |
4255 | * multiple pool_workqueues which have separate congested states. A | |
4256 | * workqueue being congested on one CPU doesn't mean the workqueue is also | |
4257 | * contested on other CPUs / NUMA nodes. | |
4258 | * | |
d185af30 | 4259 | * Return: |
dcd989cb | 4260 | * %true if congested, %false otherwise. |
eef6a7d5 | 4261 | */ |
d84ff051 | 4262 | bool workqueue_congested(int cpu, struct workqueue_struct *wq) |
1da177e4 | 4263 | { |
7fb98ea7 | 4264 | struct pool_workqueue *pwq; |
76af4d93 TH |
4265 | bool ret; |
4266 | ||
88109453 | 4267 | rcu_read_lock_sched(); |
7fb98ea7 | 4268 | |
d3251859 TH |
4269 | if (cpu == WORK_CPU_UNBOUND) |
4270 | cpu = smp_processor_id(); | |
4271 | ||
7fb98ea7 TH |
4272 | if (!(wq->flags & WQ_UNBOUND)) |
4273 | pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); | |
4274 | else | |
df2d5ae4 | 4275 | pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); |
dcd989cb | 4276 | |
76af4d93 | 4277 | ret = !list_empty(&pwq->delayed_works); |
88109453 | 4278 | rcu_read_unlock_sched(); |
76af4d93 TH |
4279 | |
4280 | return ret; | |
1da177e4 | 4281 | } |
dcd989cb | 4282 | EXPORT_SYMBOL_GPL(workqueue_congested); |
1da177e4 | 4283 | |
dcd989cb TH |
4284 | /** |
4285 | * work_busy - test whether a work is currently pending or running | |
4286 | * @work: the work to be tested | |
4287 | * | |
4288 | * Test whether @work is currently pending or running. There is no | |
4289 | * synchronization around this function and the test result is | |
4290 | * unreliable and only useful as advisory hints or for debugging. | |
dcd989cb | 4291 | * |
d185af30 | 4292 | * Return: |
dcd989cb TH |
4293 | * OR'd bitmask of WORK_BUSY_* bits. |
4294 | */ | |
4295 | unsigned int work_busy(struct work_struct *work) | |
1da177e4 | 4296 | { |
fa1b54e6 | 4297 | struct worker_pool *pool; |
dcd989cb TH |
4298 | unsigned long flags; |
4299 | unsigned int ret = 0; | |
1da177e4 | 4300 | |
dcd989cb TH |
4301 | if (work_pending(work)) |
4302 | ret |= WORK_BUSY_PENDING; | |
1da177e4 | 4303 | |
fa1b54e6 TH |
4304 | local_irq_save(flags); |
4305 | pool = get_work_pool(work); | |
038366c5 | 4306 | if (pool) { |
fa1b54e6 | 4307 | spin_lock(&pool->lock); |
038366c5 LJ |
4308 | if (find_worker_executing_work(pool, work)) |
4309 | ret |= WORK_BUSY_RUNNING; | |
fa1b54e6 | 4310 | spin_unlock(&pool->lock); |
038366c5 | 4311 | } |
fa1b54e6 | 4312 | local_irq_restore(flags); |
1da177e4 | 4313 | |
dcd989cb | 4314 | return ret; |
1da177e4 | 4315 | } |
dcd989cb | 4316 | EXPORT_SYMBOL_GPL(work_busy); |
1da177e4 | 4317 | |
3d1cb205 TH |
4318 | /** |
4319 | * set_worker_desc - set description for the current work item | |
4320 | * @fmt: printf-style format string | |
4321 | * @...: arguments for the format string | |
4322 | * | |
4323 | * This function can be called by a running work function to describe what | |
4324 | * the work item is about. If the worker task gets dumped, this | |
4325 | * information will be printed out together to help debugging. The | |
4326 | * description can be at most WORKER_DESC_LEN including the trailing '\0'. | |
4327 | */ | |
4328 | void set_worker_desc(const char *fmt, ...) | |
4329 | { | |
4330 | struct worker *worker = current_wq_worker(); | |
4331 | va_list args; | |
4332 | ||
4333 | if (worker) { | |
4334 | va_start(args, fmt); | |
4335 | vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); | |
4336 | va_end(args); | |
4337 | worker->desc_valid = true; | |
4338 | } | |
4339 | } | |
4340 | ||
4341 | /** | |
4342 | * print_worker_info - print out worker information and description | |
4343 | * @log_lvl: the log level to use when printing | |
4344 | * @task: target task | |
4345 | * | |
4346 | * If @task is a worker and currently executing a work item, print out the | |
4347 | * name of the workqueue being serviced and worker description set with | |
4348 | * set_worker_desc() by the currently executing work item. | |
4349 | * | |
4350 | * This function can be safely called on any task as long as the | |
4351 | * task_struct itself is accessible. While safe, this function isn't | |
4352 | * synchronized and may print out mixups or garbages of limited length. | |
4353 | */ | |
4354 | void print_worker_info(const char *log_lvl, struct task_struct *task) | |
4355 | { | |
4356 | work_func_t *fn = NULL; | |
4357 | char name[WQ_NAME_LEN] = { }; | |
4358 | char desc[WORKER_DESC_LEN] = { }; | |
4359 | struct pool_workqueue *pwq = NULL; | |
4360 | struct workqueue_struct *wq = NULL; | |
4361 | bool desc_valid = false; | |
4362 | struct worker *worker; | |
4363 | ||
4364 | if (!(task->flags & PF_WQ_WORKER)) | |
4365 | return; | |
4366 | ||
4367 | /* | |
4368 | * This function is called without any synchronization and @task | |
4369 | * could be in any state. Be careful with dereferences. | |
4370 | */ | |
e700591a | 4371 | worker = kthread_probe_data(task); |
3d1cb205 TH |
4372 | |
4373 | /* | |
4374 | * Carefully copy the associated workqueue's workfn and name. Keep | |
4375 | * the original last '\0' in case the original contains garbage. | |
4376 | */ | |
4377 | probe_kernel_read(&fn, &worker->current_func, sizeof(fn)); | |
4378 | probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq)); | |
4379 | probe_kernel_read(&wq, &pwq->wq, sizeof(wq)); | |
4380 | probe_kernel_read(name, wq->name, sizeof(name) - 1); | |
4381 | ||
4382 | /* copy worker description */ | |
4383 | probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid)); | |
4384 | if (desc_valid) | |
4385 | probe_kernel_read(desc, worker->desc, sizeof(desc) - 1); | |
4386 | ||
4387 | if (fn || name[0] || desc[0]) { | |
4388 | printk("%sWorkqueue: %s %pf", log_lvl, name, fn); | |
4389 | if (desc[0]) | |
4390 | pr_cont(" (%s)", desc); | |
4391 | pr_cont("\n"); | |
4392 | } | |
4393 | } | |
4394 | ||
3494fc30 TH |
4395 | static void pr_cont_pool_info(struct worker_pool *pool) |
4396 | { | |
4397 | pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask); | |
4398 | if (pool->node != NUMA_NO_NODE) | |
4399 | pr_cont(" node=%d", pool->node); | |
4400 | pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice); | |
4401 | } | |
4402 | ||
4403 | static void pr_cont_work(bool comma, struct work_struct *work) | |
4404 | { | |
4405 | if (work->func == wq_barrier_func) { | |
4406 | struct wq_barrier *barr; | |
4407 | ||
4408 | barr = container_of(work, struct wq_barrier, work); | |
4409 | ||
4410 | pr_cont("%s BAR(%d)", comma ? "," : "", | |
4411 | task_pid_nr(barr->task)); | |
4412 | } else { | |
4413 | pr_cont("%s %pf", comma ? "," : "", work->func); | |
4414 | } | |
4415 | } | |
4416 | ||
4417 | static void show_pwq(struct pool_workqueue *pwq) | |
4418 | { | |
4419 | struct worker_pool *pool = pwq->pool; | |
4420 | struct work_struct *work; | |
4421 | struct worker *worker; | |
4422 | bool has_in_flight = false, has_pending = false; | |
4423 | int bkt; | |
4424 | ||
4425 | pr_info(" pwq %d:", pool->id); | |
4426 | pr_cont_pool_info(pool); | |
4427 | ||
4428 | pr_cont(" active=%d/%d%s\n", pwq->nr_active, pwq->max_active, | |
4429 | !list_empty(&pwq->mayday_node) ? " MAYDAY" : ""); | |
4430 | ||
4431 | hash_for_each(pool->busy_hash, bkt, worker, hentry) { | |
4432 | if (worker->current_pwq == pwq) { | |
4433 | has_in_flight = true; | |
4434 | break; | |
4435 | } | |
4436 | } | |
4437 | if (has_in_flight) { | |
4438 | bool comma = false; | |
4439 | ||
4440 | pr_info(" in-flight:"); | |
4441 | hash_for_each(pool->busy_hash, bkt, worker, hentry) { | |
4442 | if (worker->current_pwq != pwq) | |
4443 | continue; | |
4444 | ||
4445 | pr_cont("%s %d%s:%pf", comma ? "," : "", | |
4446 | task_pid_nr(worker->task), | |
4447 | worker == pwq->wq->rescuer ? "(RESCUER)" : "", | |
4448 | worker->current_func); | |
4449 | list_for_each_entry(work, &worker->scheduled, entry) | |
4450 | pr_cont_work(false, work); | |
4451 | comma = true; | |
4452 | } | |
4453 | pr_cont("\n"); | |
4454 | } | |
4455 | ||
4456 | list_for_each_entry(work, &pool->worklist, entry) { | |
4457 | if (get_work_pwq(work) == pwq) { | |
4458 | has_pending = true; | |
4459 | break; | |
4460 | } | |
4461 | } | |
4462 | if (has_pending) { | |
4463 | bool comma = false; | |
4464 | ||
4465 | pr_info(" pending:"); | |
4466 | list_for_each_entry(work, &pool->worklist, entry) { | |
4467 | if (get_work_pwq(work) != pwq) | |
4468 | continue; | |
4469 | ||
4470 | pr_cont_work(comma, work); | |
4471 | comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); | |
4472 | } | |
4473 | pr_cont("\n"); | |
4474 | } | |
4475 | ||
4476 | if (!list_empty(&pwq->delayed_works)) { | |
4477 | bool comma = false; | |
4478 | ||
4479 | pr_info(" delayed:"); | |
4480 | list_for_each_entry(work, &pwq->delayed_works, entry) { | |
4481 | pr_cont_work(comma, work); | |
4482 | comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); | |
4483 | } | |
4484 | pr_cont("\n"); | |
4485 | } | |
4486 | } | |
4487 | ||
4488 | /** | |
4489 | * show_workqueue_state - dump workqueue state | |
4490 | * | |
7b776af6 RL |
4491 | * Called from a sysrq handler or try_to_freeze_tasks() and prints out |
4492 | * all busy workqueues and pools. | |
3494fc30 TH |
4493 | */ |
4494 | void show_workqueue_state(void) | |
4495 | { | |
4496 | struct workqueue_struct *wq; | |
4497 | struct worker_pool *pool; | |
4498 | unsigned long flags; | |
4499 | int pi; | |
4500 | ||
4501 | rcu_read_lock_sched(); | |
4502 | ||
4503 | pr_info("Showing busy workqueues and worker pools:\n"); | |
4504 | ||
4505 | list_for_each_entry_rcu(wq, &workqueues, list) { | |
4506 | struct pool_workqueue *pwq; | |
4507 | bool idle = true; | |
4508 | ||
4509 | for_each_pwq(pwq, wq) { | |
4510 | if (pwq->nr_active || !list_empty(&pwq->delayed_works)) { | |
4511 | idle = false; | |
4512 | break; | |
4513 | } | |
4514 | } | |
4515 | if (idle) | |
4516 | continue; | |
4517 | ||
4518 | pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags); | |
4519 | ||
4520 | for_each_pwq(pwq, wq) { | |
4521 | spin_lock_irqsave(&pwq->pool->lock, flags); | |
4522 | if (pwq->nr_active || !list_empty(&pwq->delayed_works)) | |
4523 | show_pwq(pwq); | |
4524 | spin_unlock_irqrestore(&pwq->pool->lock, flags); | |
a6d5930c SS |
4525 | /* |
4526 | * We could be printing a lot from atomic context, e.g. | |
4527 | * sysrq-t -> show_workqueue_state(). Avoid triggering | |
4528 | * hard lockup. | |
4529 | */ | |
4530 | touch_nmi_watchdog(); | |
3494fc30 TH |
4531 | } |
4532 | } | |
4533 | ||
4534 | for_each_pool(pool, pi) { | |
4535 | struct worker *worker; | |
4536 | bool first = true; | |
4537 | ||
4538 | spin_lock_irqsave(&pool->lock, flags); | |
4539 | if (pool->nr_workers == pool->nr_idle) | |
4540 | goto next_pool; | |
4541 | ||
4542 | pr_info("pool %d:", pool->id); | |
4543 | pr_cont_pool_info(pool); | |
82607adc TH |
4544 | pr_cont(" hung=%us workers=%d", |
4545 | jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000, | |
4546 | pool->nr_workers); | |
3494fc30 TH |
4547 | if (pool->manager) |
4548 | pr_cont(" manager: %d", | |
4549 | task_pid_nr(pool->manager->task)); | |
4550 | list_for_each_entry(worker, &pool->idle_list, entry) { | |
4551 | pr_cont(" %s%d", first ? "idle: " : "", | |
4552 | task_pid_nr(worker->task)); | |
4553 | first = false; | |
4554 | } | |
4555 | pr_cont("\n"); | |
4556 | next_pool: | |
4557 | spin_unlock_irqrestore(&pool->lock, flags); | |
a6d5930c SS |
4558 | /* |
4559 | * We could be printing a lot from atomic context, e.g. | |
4560 | * sysrq-t -> show_workqueue_state(). Avoid triggering | |
4561 | * hard lockup. | |
4562 | */ | |
4563 | touch_nmi_watchdog(); | |
3494fc30 TH |
4564 | } |
4565 | ||
4566 | rcu_read_unlock_sched(); | |
4567 | } | |
4568 | ||
db7bccf4 TH |
4569 | /* |
4570 | * CPU hotplug. | |
4571 | * | |
e22bee78 | 4572 | * There are two challenges in supporting CPU hotplug. Firstly, there |
112202d9 | 4573 | * are a lot of assumptions on strong associations among work, pwq and |
706026c2 | 4574 | * pool which make migrating pending and scheduled works very |
e22bee78 | 4575 | * difficult to implement without impacting hot paths. Secondly, |
94cf58bb | 4576 | * worker pools serve mix of short, long and very long running works making |
e22bee78 TH |
4577 | * blocked draining impractical. |
4578 | * | |
24647570 | 4579 | * This is solved by allowing the pools to be disassociated from the CPU |
628c78e7 TH |
4580 | * running as an unbound one and allowing it to be reattached later if the |
4581 | * cpu comes back online. | |
db7bccf4 | 4582 | */ |
1da177e4 | 4583 | |
706026c2 | 4584 | static void wq_unbind_fn(struct work_struct *work) |
3af24433 | 4585 | { |
38db41d9 | 4586 | int cpu = smp_processor_id(); |
4ce62e9e | 4587 | struct worker_pool *pool; |
db7bccf4 | 4588 | struct worker *worker; |
3af24433 | 4589 | |
f02ae73a | 4590 | for_each_cpu_worker_pool(pool, cpu) { |
92f9c5c4 | 4591 | mutex_lock(&pool->attach_mutex); |
94cf58bb | 4592 | spin_lock_irq(&pool->lock); |
3af24433 | 4593 | |
94cf58bb | 4594 | /* |
92f9c5c4 | 4595 | * We've blocked all attach/detach operations. Make all workers |
94cf58bb TH |
4596 | * unbound and set DISASSOCIATED. Before this, all workers |
4597 | * except for the ones which are still executing works from | |
4598 | * before the last CPU down must be on the cpu. After | |
4599 | * this, they may become diasporas. | |
4600 | */ | |
da028469 | 4601 | for_each_pool_worker(worker, pool) |
c9e7cf27 | 4602 | worker->flags |= WORKER_UNBOUND; |
06ba38a9 | 4603 | |
24647570 | 4604 | pool->flags |= POOL_DISASSOCIATED; |
f2d5a0ee | 4605 | |
94cf58bb | 4606 | spin_unlock_irq(&pool->lock); |
92f9c5c4 | 4607 | mutex_unlock(&pool->attach_mutex); |
628c78e7 | 4608 | |
eb283428 LJ |
4609 | /* |
4610 | * Call schedule() so that we cross rq->lock and thus can | |
4611 | * guarantee sched callbacks see the %WORKER_UNBOUND flag. | |
4612 | * This is necessary as scheduler callbacks may be invoked | |
4613 | * from other cpus. | |
4614 | */ | |
4615 | schedule(); | |
06ba38a9 | 4616 | |
eb283428 LJ |
4617 | /* |
4618 | * Sched callbacks are disabled now. Zap nr_running. | |
4619 | * After this, nr_running stays zero and need_more_worker() | |
4620 | * and keep_working() are always true as long as the | |
4621 | * worklist is not empty. This pool now behaves as an | |
4622 | * unbound (in terms of concurrency management) pool which | |
4623 | * are served by workers tied to the pool. | |
4624 | */ | |
e19e397a | 4625 | atomic_set(&pool->nr_running, 0); |
eb283428 LJ |
4626 | |
4627 | /* | |
4628 | * With concurrency management just turned off, a busy | |
4629 | * worker blocking could lead to lengthy stalls. Kick off | |
4630 | * unbound chain execution of currently pending work items. | |
4631 | */ | |
4632 | spin_lock_irq(&pool->lock); | |
4633 | wake_up_worker(pool); | |
4634 | spin_unlock_irq(&pool->lock); | |
4635 | } | |
3af24433 | 4636 | } |
3af24433 | 4637 | |
bd7c089e TH |
4638 | /** |
4639 | * rebind_workers - rebind all workers of a pool to the associated CPU | |
4640 | * @pool: pool of interest | |
4641 | * | |
a9ab775b | 4642 | * @pool->cpu is coming online. Rebind all workers to the CPU. |
bd7c089e TH |
4643 | */ |
4644 | static void rebind_workers(struct worker_pool *pool) | |
4645 | { | |
a9ab775b | 4646 | struct worker *worker; |
bd7c089e | 4647 | |
92f9c5c4 | 4648 | lockdep_assert_held(&pool->attach_mutex); |
bd7c089e | 4649 | |
a9ab775b TH |
4650 | /* |
4651 | * Restore CPU affinity of all workers. As all idle workers should | |
4652 | * be on the run-queue of the associated CPU before any local | |
402dd89d | 4653 | * wake-ups for concurrency management happen, restore CPU affinity |
a9ab775b TH |
4654 | * of all workers first and then clear UNBOUND. As we're called |
4655 | * from CPU_ONLINE, the following shouldn't fail. | |
4656 | */ | |
da028469 | 4657 | for_each_pool_worker(worker, pool) |
a9ab775b TH |
4658 | WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, |
4659 | pool->attrs->cpumask) < 0); | |
bd7c089e | 4660 | |
a9ab775b | 4661 | spin_lock_irq(&pool->lock); |
f7c17d26 WL |
4662 | |
4663 | /* | |
4664 | * XXX: CPU hotplug notifiers are weird and can call DOWN_FAILED | |
4665 | * w/o preceding DOWN_PREPARE. Work around it. CPU hotplug is | |
4666 | * being reworked and this can go away in time. | |
4667 | */ | |
4668 | if (!(pool->flags & POOL_DISASSOCIATED)) { | |
4669 | spin_unlock_irq(&pool->lock); | |
4670 | return; | |
4671 | } | |
4672 | ||
3de5e884 | 4673 | pool->flags &= ~POOL_DISASSOCIATED; |
bd7c089e | 4674 | |
da028469 | 4675 | for_each_pool_worker(worker, pool) { |
a9ab775b | 4676 | unsigned int worker_flags = worker->flags; |
bd7c089e TH |
4677 | |
4678 | /* | |
a9ab775b TH |
4679 | * A bound idle worker should actually be on the runqueue |
4680 | * of the associated CPU for local wake-ups targeting it to | |
4681 | * work. Kick all idle workers so that they migrate to the | |
4682 | * associated CPU. Doing this in the same loop as | |
4683 | * replacing UNBOUND with REBOUND is safe as no worker will | |
4684 | * be bound before @pool->lock is released. | |
bd7c089e | 4685 | */ |
a9ab775b TH |
4686 | if (worker_flags & WORKER_IDLE) |
4687 | wake_up_process(worker->task); | |
bd7c089e | 4688 | |
a9ab775b TH |
4689 | /* |
4690 | * We want to clear UNBOUND but can't directly call | |
4691 | * worker_clr_flags() or adjust nr_running. Atomically | |
4692 | * replace UNBOUND with another NOT_RUNNING flag REBOUND. | |
4693 | * @worker will clear REBOUND using worker_clr_flags() when | |
4694 | * it initiates the next execution cycle thus restoring | |
4695 | * concurrency management. Note that when or whether | |
4696 | * @worker clears REBOUND doesn't affect correctness. | |
4697 | * | |
4698 | * ACCESS_ONCE() is necessary because @worker->flags may be | |
4699 | * tested without holding any lock in | |
4700 | * wq_worker_waking_up(). Without it, NOT_RUNNING test may | |
4701 | * fail incorrectly leading to premature concurrency | |
4702 | * management operations. | |
4703 | */ | |
4704 | WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); | |
4705 | worker_flags |= WORKER_REBOUND; | |
4706 | worker_flags &= ~WORKER_UNBOUND; | |
4707 | ACCESS_ONCE(worker->flags) = worker_flags; | |
bd7c089e | 4708 | } |
a9ab775b TH |
4709 | |
4710 | spin_unlock_irq(&pool->lock); | |
bd7c089e TH |
4711 | } |
4712 | ||
7dbc725e TH |
4713 | /** |
4714 | * restore_unbound_workers_cpumask - restore cpumask of unbound workers | |
4715 | * @pool: unbound pool of interest | |
4716 | * @cpu: the CPU which is coming up | |
4717 | * | |
4718 | * An unbound pool may end up with a cpumask which doesn't have any online | |
4719 | * CPUs. When a worker of such pool get scheduled, the scheduler resets | |
4720 | * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any | |
4721 | * online CPU before, cpus_allowed of all its workers should be restored. | |
4722 | */ | |
4723 | static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) | |
4724 | { | |
4725 | static cpumask_t cpumask; | |
4726 | struct worker *worker; | |
7dbc725e | 4727 | |
92f9c5c4 | 4728 | lockdep_assert_held(&pool->attach_mutex); |
7dbc725e TH |
4729 | |
4730 | /* is @cpu allowed for @pool? */ | |
4731 | if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) | |
4732 | return; | |
4733 | ||
7dbc725e | 4734 | cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); |
7dbc725e TH |
4735 | |
4736 | /* as we're called from CPU_ONLINE, the following shouldn't fail */ | |
da028469 | 4737 | for_each_pool_worker(worker, pool) |
d945b5e9 | 4738 | WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0); |
7dbc725e TH |
4739 | } |
4740 | ||
7ee681b2 TG |
4741 | int workqueue_prepare_cpu(unsigned int cpu) |
4742 | { | |
4743 | struct worker_pool *pool; | |
4744 | ||
4745 | for_each_cpu_worker_pool(pool, cpu) { | |
4746 | if (pool->nr_workers) | |
4747 | continue; | |
4748 | if (!create_worker(pool)) | |
4749 | return -ENOMEM; | |
4750 | } | |
4751 | return 0; | |
4752 | } | |
4753 | ||
4754 | int workqueue_online_cpu(unsigned int cpu) | |
3af24433 | 4755 | { |
4ce62e9e | 4756 | struct worker_pool *pool; |
4c16bd32 | 4757 | struct workqueue_struct *wq; |
7dbc725e | 4758 | int pi; |
3ce63377 | 4759 | |
7ee681b2 | 4760 | mutex_lock(&wq_pool_mutex); |
7dbc725e | 4761 | |
7ee681b2 TG |
4762 | for_each_pool(pool, pi) { |
4763 | mutex_lock(&pool->attach_mutex); | |
94cf58bb | 4764 | |
7ee681b2 TG |
4765 | if (pool->cpu == cpu) |
4766 | rebind_workers(pool); | |
4767 | else if (pool->cpu < 0) | |
4768 | restore_unbound_workers_cpumask(pool, cpu); | |
94cf58bb | 4769 | |
7ee681b2 TG |
4770 | mutex_unlock(&pool->attach_mutex); |
4771 | } | |
6ba94429 | 4772 | |
7ee681b2 TG |
4773 | /* update NUMA affinity of unbound workqueues */ |
4774 | list_for_each_entry(wq, &workqueues, list) | |
4775 | wq_update_unbound_numa(wq, cpu, true); | |
6ba94429 | 4776 | |
7ee681b2 TG |
4777 | mutex_unlock(&wq_pool_mutex); |
4778 | return 0; | |
6ba94429 FW |
4779 | } |
4780 | ||
7ee681b2 | 4781 | int workqueue_offline_cpu(unsigned int cpu) |
6ba94429 | 4782 | { |
6ba94429 FW |
4783 | struct work_struct unbind_work; |
4784 | struct workqueue_struct *wq; | |
4785 | ||
7ee681b2 TG |
4786 | /* unbinding per-cpu workers should happen on the local CPU */ |
4787 | INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn); | |
4788 | queue_work_on(cpu, system_highpri_wq, &unbind_work); | |
4789 | ||
4790 | /* update NUMA affinity of unbound workqueues */ | |
4791 | mutex_lock(&wq_pool_mutex); | |
4792 | list_for_each_entry(wq, &workqueues, list) | |
4793 | wq_update_unbound_numa(wq, cpu, false); | |
4794 | mutex_unlock(&wq_pool_mutex); | |
4795 | ||
4796 | /* wait for per-cpu unbinding to finish */ | |
4797 | flush_work(&unbind_work); | |
4798 | destroy_work_on_stack(&unbind_work); | |
4799 | return 0; | |
6ba94429 FW |
4800 | } |
4801 | ||
4802 | #ifdef CONFIG_SMP | |
4803 | ||
4804 | struct work_for_cpu { | |
4805 | struct work_struct work; | |
4806 | long (*fn)(void *); | |
4807 | void *arg; | |
4808 | long ret; | |
4809 | }; | |
4810 | ||
4811 | static void work_for_cpu_fn(struct work_struct *work) | |
4812 | { | |
4813 | struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); | |
4814 | ||
4815 | wfc->ret = wfc->fn(wfc->arg); | |
4816 | } | |
4817 | ||
4818 | /** | |
22aceb31 | 4819 | * work_on_cpu - run a function in thread context on a particular cpu |
6ba94429 FW |
4820 | * @cpu: the cpu to run on |
4821 | * @fn: the function to run | |
4822 | * @arg: the function arg | |
4823 | * | |
4824 | * It is up to the caller to ensure that the cpu doesn't go offline. | |
4825 | * The caller must not hold any locks which would prevent @fn from completing. | |
4826 | * | |
4827 | * Return: The value @fn returns. | |
4828 | */ | |
4829 | long work_on_cpu(int cpu, long (*fn)(void *), void *arg) | |
4830 | { | |
4831 | struct work_for_cpu wfc = { .fn = fn, .arg = arg }; | |
4832 | ||
4833 | INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); | |
4834 | schedule_work_on(cpu, &wfc.work); | |
4835 | flush_work(&wfc.work); | |
4836 | destroy_work_on_stack(&wfc.work); | |
4837 | return wfc.ret; | |
4838 | } | |
4839 | EXPORT_SYMBOL_GPL(work_on_cpu); | |
0e8d6a93 TG |
4840 | |
4841 | /** | |
4842 | * work_on_cpu_safe - run a function in thread context on a particular cpu | |
4843 | * @cpu: the cpu to run on | |
4844 | * @fn: the function to run | |
4845 | * @arg: the function argument | |
4846 | * | |
4847 | * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold | |
4848 | * any locks which would prevent @fn from completing. | |
4849 | * | |
4850 | * Return: The value @fn returns. | |
4851 | */ | |
4852 | long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg) | |
4853 | { | |
4854 | long ret = -ENODEV; | |
4855 | ||
4856 | get_online_cpus(); | |
4857 | if (cpu_online(cpu)) | |
4858 | ret = work_on_cpu(cpu, fn, arg); | |
4859 | put_online_cpus(); | |
4860 | return ret; | |
4861 | } | |
4862 | EXPORT_SYMBOL_GPL(work_on_cpu_safe); | |
6ba94429 FW |
4863 | #endif /* CONFIG_SMP */ |
4864 | ||
4865 | #ifdef CONFIG_FREEZER | |
4866 | ||
4867 | /** | |
4868 | * freeze_workqueues_begin - begin freezing workqueues | |
4869 | * | |
4870 | * Start freezing workqueues. After this function returns, all freezable | |
4871 | * workqueues will queue new works to their delayed_works list instead of | |
4872 | * pool->worklist. | |
4873 | * | |
4874 | * CONTEXT: | |
4875 | * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. | |
4876 | */ | |
4877 | void freeze_workqueues_begin(void) | |
4878 | { | |
4879 | struct workqueue_struct *wq; | |
4880 | struct pool_workqueue *pwq; | |
4881 | ||
4882 | mutex_lock(&wq_pool_mutex); | |
4883 | ||
4884 | WARN_ON_ONCE(workqueue_freezing); | |
4885 | workqueue_freezing = true; | |
4886 | ||
4887 | list_for_each_entry(wq, &workqueues, list) { | |
4888 | mutex_lock(&wq->mutex); | |
4889 | for_each_pwq(pwq, wq) | |
4890 | pwq_adjust_max_active(pwq); | |
4891 | mutex_unlock(&wq->mutex); | |
4892 | } | |
4893 | ||
4894 | mutex_unlock(&wq_pool_mutex); | |
4895 | } | |
4896 | ||
4897 | /** | |
4898 | * freeze_workqueues_busy - are freezable workqueues still busy? | |
4899 | * | |
4900 | * Check whether freezing is complete. This function must be called | |
4901 | * between freeze_workqueues_begin() and thaw_workqueues(). | |
4902 | * | |
4903 | * CONTEXT: | |
4904 | * Grabs and releases wq_pool_mutex. | |
4905 | * | |
4906 | * Return: | |
4907 | * %true if some freezable workqueues are still busy. %false if freezing | |
4908 | * is complete. | |
4909 | */ | |
4910 | bool freeze_workqueues_busy(void) | |
4911 | { | |
4912 | bool busy = false; | |
4913 | struct workqueue_struct *wq; | |
4914 | struct pool_workqueue *pwq; | |
4915 | ||
4916 | mutex_lock(&wq_pool_mutex); | |
4917 | ||
4918 | WARN_ON_ONCE(!workqueue_freezing); | |
4919 | ||
4920 | list_for_each_entry(wq, &workqueues, list) { | |
4921 | if (!(wq->flags & WQ_FREEZABLE)) | |
4922 | continue; | |
4923 | /* | |
4924 | * nr_active is monotonically decreasing. It's safe | |
4925 | * to peek without lock. | |
4926 | */ | |
4927 | rcu_read_lock_sched(); | |
4928 | for_each_pwq(pwq, wq) { | |
4929 | WARN_ON_ONCE(pwq->nr_active < 0); | |
4930 | if (pwq->nr_active) { | |
4931 | busy = true; | |
4932 | rcu_read_unlock_sched(); | |
4933 | goto out_unlock; | |
4934 | } | |
4935 | } | |
4936 | rcu_read_unlock_sched(); | |
4937 | } | |
4938 | out_unlock: | |
4939 | mutex_unlock(&wq_pool_mutex); | |
4940 | return busy; | |
4941 | } | |
4942 | ||
4943 | /** | |
4944 | * thaw_workqueues - thaw workqueues | |
4945 | * | |
4946 | * Thaw workqueues. Normal queueing is restored and all collected | |
4947 | * frozen works are transferred to their respective pool worklists. | |
4948 | * | |
4949 | * CONTEXT: | |
4950 | * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. | |
4951 | */ | |
4952 | void thaw_workqueues(void) | |
4953 | { | |
4954 | struct workqueue_struct *wq; | |
4955 | struct pool_workqueue *pwq; | |
4956 | ||
4957 | mutex_lock(&wq_pool_mutex); | |
4958 | ||
4959 | if (!workqueue_freezing) | |
4960 | goto out_unlock; | |
4961 | ||
4962 | workqueue_freezing = false; | |
4963 | ||
4964 | /* restore max_active and repopulate worklist */ | |
4965 | list_for_each_entry(wq, &workqueues, list) { | |
4966 | mutex_lock(&wq->mutex); | |
4967 | for_each_pwq(pwq, wq) | |
4968 | pwq_adjust_max_active(pwq); | |
4969 | mutex_unlock(&wq->mutex); | |
4970 | } | |
4971 | ||
4972 | out_unlock: | |
4973 | mutex_unlock(&wq_pool_mutex); | |
4974 | } | |
4975 | #endif /* CONFIG_FREEZER */ | |
4976 | ||
042f7df1 LJ |
4977 | static int workqueue_apply_unbound_cpumask(void) |
4978 | { | |
4979 | LIST_HEAD(ctxs); | |
4980 | int ret = 0; | |
4981 | struct workqueue_struct *wq; | |
4982 | struct apply_wqattrs_ctx *ctx, *n; | |
4983 | ||
4984 | lockdep_assert_held(&wq_pool_mutex); | |
4985 | ||
4986 | list_for_each_entry(wq, &workqueues, list) { | |
4987 | if (!(wq->flags & WQ_UNBOUND)) | |
4988 | continue; | |
4989 | /* creating multiple pwqs breaks ordering guarantee */ | |
4990 | if (wq->flags & __WQ_ORDERED) | |
4991 | continue; | |
4992 | ||
4993 | ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs); | |
4994 | if (!ctx) { | |
4995 | ret = -ENOMEM; | |
4996 | break; | |
4997 | } | |
4998 | ||
4999 | list_add_tail(&ctx->list, &ctxs); | |
5000 | } | |
5001 | ||
5002 | list_for_each_entry_safe(ctx, n, &ctxs, list) { | |
5003 | if (!ret) | |
5004 | apply_wqattrs_commit(ctx); | |
5005 | apply_wqattrs_cleanup(ctx); | |
5006 | } | |
5007 | ||
5008 | return ret; | |
5009 | } | |
5010 | ||
5011 | /** | |
5012 | * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask | |
5013 | * @cpumask: the cpumask to set | |
5014 | * | |
5015 | * The low-level workqueues cpumask is a global cpumask that limits | |
5016 | * the affinity of all unbound workqueues. This function check the @cpumask | |
5017 | * and apply it to all unbound workqueues and updates all pwqs of them. | |
5018 | * | |
5019 | * Retun: 0 - Success | |
5020 | * -EINVAL - Invalid @cpumask | |
5021 | * -ENOMEM - Failed to allocate memory for attrs or pwqs. | |
5022 | */ | |
5023 | int workqueue_set_unbound_cpumask(cpumask_var_t cpumask) | |
5024 | { | |
5025 | int ret = -EINVAL; | |
5026 | cpumask_var_t saved_cpumask; | |
5027 | ||
5028 | if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL)) | |
5029 | return -ENOMEM; | |
5030 | ||
042f7df1 LJ |
5031 | cpumask_and(cpumask, cpumask, cpu_possible_mask); |
5032 | if (!cpumask_empty(cpumask)) { | |
a0111cf6 | 5033 | apply_wqattrs_lock(); |
042f7df1 LJ |
5034 | |
5035 | /* save the old wq_unbound_cpumask. */ | |
5036 | cpumask_copy(saved_cpumask, wq_unbound_cpumask); | |
5037 | ||
5038 | /* update wq_unbound_cpumask at first and apply it to wqs. */ | |
5039 | cpumask_copy(wq_unbound_cpumask, cpumask); | |
5040 | ret = workqueue_apply_unbound_cpumask(); | |
5041 | ||
5042 | /* restore the wq_unbound_cpumask when failed. */ | |
5043 | if (ret < 0) | |
5044 | cpumask_copy(wq_unbound_cpumask, saved_cpumask); | |
5045 | ||
a0111cf6 | 5046 | apply_wqattrs_unlock(); |
042f7df1 | 5047 | } |
042f7df1 LJ |
5048 | |
5049 | free_cpumask_var(saved_cpumask); | |
5050 | return ret; | |
5051 | } | |
5052 | ||
6ba94429 FW |
5053 | #ifdef CONFIG_SYSFS |
5054 | /* | |
5055 | * Workqueues with WQ_SYSFS flag set is visible to userland via | |
5056 | * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the | |
5057 | * following attributes. | |
5058 | * | |
5059 | * per_cpu RO bool : whether the workqueue is per-cpu or unbound | |
5060 | * max_active RW int : maximum number of in-flight work items | |
5061 | * | |
5062 | * Unbound workqueues have the following extra attributes. | |
5063 | * | |
5064 | * id RO int : the associated pool ID | |
5065 | * nice RW int : nice value of the workers | |
5066 | * cpumask RW mask : bitmask of allowed CPUs for the workers | |
5067 | */ | |
5068 | struct wq_device { | |
5069 | struct workqueue_struct *wq; | |
5070 | struct device dev; | |
5071 | }; | |
5072 | ||
5073 | static struct workqueue_struct *dev_to_wq(struct device *dev) | |
5074 | { | |
5075 | struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); | |
5076 | ||
5077 | return wq_dev->wq; | |
5078 | } | |
5079 | ||
5080 | static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, | |
5081 | char *buf) | |
5082 | { | |
5083 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5084 | ||
5085 | return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); | |
5086 | } | |
5087 | static DEVICE_ATTR_RO(per_cpu); | |
5088 | ||
5089 | static ssize_t max_active_show(struct device *dev, | |
5090 | struct device_attribute *attr, char *buf) | |
5091 | { | |
5092 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5093 | ||
5094 | return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); | |
5095 | } | |
5096 | ||
5097 | static ssize_t max_active_store(struct device *dev, | |
5098 | struct device_attribute *attr, const char *buf, | |
5099 | size_t count) | |
5100 | { | |
5101 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5102 | int val; | |
5103 | ||
5104 | if (sscanf(buf, "%d", &val) != 1 || val <= 0) | |
5105 | return -EINVAL; | |
5106 | ||
5107 | workqueue_set_max_active(wq, val); | |
5108 | return count; | |
5109 | } | |
5110 | static DEVICE_ATTR_RW(max_active); | |
5111 | ||
5112 | static struct attribute *wq_sysfs_attrs[] = { | |
5113 | &dev_attr_per_cpu.attr, | |
5114 | &dev_attr_max_active.attr, | |
5115 | NULL, | |
5116 | }; | |
5117 | ATTRIBUTE_GROUPS(wq_sysfs); | |
5118 | ||
5119 | static ssize_t wq_pool_ids_show(struct device *dev, | |
5120 | struct device_attribute *attr, char *buf) | |
5121 | { | |
5122 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5123 | const char *delim = ""; | |
5124 | int node, written = 0; | |
5125 | ||
5126 | rcu_read_lock_sched(); | |
5127 | for_each_node(node) { | |
5128 | written += scnprintf(buf + written, PAGE_SIZE - written, | |
5129 | "%s%d:%d", delim, node, | |
5130 | unbound_pwq_by_node(wq, node)->pool->id); | |
5131 | delim = " "; | |
5132 | } | |
5133 | written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); | |
5134 | rcu_read_unlock_sched(); | |
5135 | ||
5136 | return written; | |
5137 | } | |
5138 | ||
5139 | static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, | |
5140 | char *buf) | |
5141 | { | |
5142 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5143 | int written; | |
5144 | ||
5145 | mutex_lock(&wq->mutex); | |
5146 | written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); | |
5147 | mutex_unlock(&wq->mutex); | |
5148 | ||
5149 | return written; | |
5150 | } | |
5151 | ||
5152 | /* prepare workqueue_attrs for sysfs store operations */ | |
5153 | static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) | |
5154 | { | |
5155 | struct workqueue_attrs *attrs; | |
5156 | ||
899a94fe LJ |
5157 | lockdep_assert_held(&wq_pool_mutex); |
5158 | ||
6ba94429 FW |
5159 | attrs = alloc_workqueue_attrs(GFP_KERNEL); |
5160 | if (!attrs) | |
5161 | return NULL; | |
5162 | ||
6ba94429 | 5163 | copy_workqueue_attrs(attrs, wq->unbound_attrs); |
6ba94429 FW |
5164 | return attrs; |
5165 | } | |
5166 | ||
5167 | static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, | |
5168 | const char *buf, size_t count) | |
5169 | { | |
5170 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5171 | struct workqueue_attrs *attrs; | |
d4d3e257 LJ |
5172 | int ret = -ENOMEM; |
5173 | ||
5174 | apply_wqattrs_lock(); | |
6ba94429 FW |
5175 | |
5176 | attrs = wq_sysfs_prep_attrs(wq); | |
5177 | if (!attrs) | |
d4d3e257 | 5178 | goto out_unlock; |
6ba94429 FW |
5179 | |
5180 | if (sscanf(buf, "%d", &attrs->nice) == 1 && | |
5181 | attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) | |
d4d3e257 | 5182 | ret = apply_workqueue_attrs_locked(wq, attrs); |
6ba94429 FW |
5183 | else |
5184 | ret = -EINVAL; | |
5185 | ||
d4d3e257 LJ |
5186 | out_unlock: |
5187 | apply_wqattrs_unlock(); | |
6ba94429 FW |
5188 | free_workqueue_attrs(attrs); |
5189 | return ret ?: count; | |
5190 | } | |
5191 | ||
5192 | static ssize_t wq_cpumask_show(struct device *dev, | |
5193 | struct device_attribute *attr, char *buf) | |
5194 | { | |
5195 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5196 | int written; | |
5197 | ||
5198 | mutex_lock(&wq->mutex); | |
5199 | written = scnprintf(buf, PAGE_SIZE, "%*pb\n", | |
5200 | cpumask_pr_args(wq->unbound_attrs->cpumask)); | |
5201 | mutex_unlock(&wq->mutex); | |
5202 | return written; | |
5203 | } | |
5204 | ||
5205 | static ssize_t wq_cpumask_store(struct device *dev, | |
5206 | struct device_attribute *attr, | |
5207 | const char *buf, size_t count) | |
5208 | { | |
5209 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5210 | struct workqueue_attrs *attrs; | |
d4d3e257 LJ |
5211 | int ret = -ENOMEM; |
5212 | ||
5213 | apply_wqattrs_lock(); | |
6ba94429 FW |
5214 | |
5215 | attrs = wq_sysfs_prep_attrs(wq); | |
5216 | if (!attrs) | |
d4d3e257 | 5217 | goto out_unlock; |
6ba94429 FW |
5218 | |
5219 | ret = cpumask_parse(buf, attrs->cpumask); | |
5220 | if (!ret) | |
d4d3e257 | 5221 | ret = apply_workqueue_attrs_locked(wq, attrs); |
6ba94429 | 5222 | |
d4d3e257 LJ |
5223 | out_unlock: |
5224 | apply_wqattrs_unlock(); | |
6ba94429 FW |
5225 | free_workqueue_attrs(attrs); |
5226 | return ret ?: count; | |
5227 | } | |
5228 | ||
5229 | static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr, | |
5230 | char *buf) | |
5231 | { | |
5232 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5233 | int written; | |
7dbc725e | 5234 | |
6ba94429 FW |
5235 | mutex_lock(&wq->mutex); |
5236 | written = scnprintf(buf, PAGE_SIZE, "%d\n", | |
5237 | !wq->unbound_attrs->no_numa); | |
5238 | mutex_unlock(&wq->mutex); | |
4c16bd32 | 5239 | |
6ba94429 | 5240 | return written; |
65758202 TH |
5241 | } |
5242 | ||
6ba94429 FW |
5243 | static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr, |
5244 | const char *buf, size_t count) | |
65758202 | 5245 | { |
6ba94429 FW |
5246 | struct workqueue_struct *wq = dev_to_wq(dev); |
5247 | struct workqueue_attrs *attrs; | |
d4d3e257 LJ |
5248 | int v, ret = -ENOMEM; |
5249 | ||
5250 | apply_wqattrs_lock(); | |
4c16bd32 | 5251 | |
6ba94429 FW |
5252 | attrs = wq_sysfs_prep_attrs(wq); |
5253 | if (!attrs) | |
d4d3e257 | 5254 | goto out_unlock; |
4c16bd32 | 5255 | |
6ba94429 FW |
5256 | ret = -EINVAL; |
5257 | if (sscanf(buf, "%d", &v) == 1) { | |
5258 | attrs->no_numa = !v; | |
d4d3e257 | 5259 | ret = apply_workqueue_attrs_locked(wq, attrs); |
65758202 | 5260 | } |
6ba94429 | 5261 | |
d4d3e257 LJ |
5262 | out_unlock: |
5263 | apply_wqattrs_unlock(); | |
6ba94429 FW |
5264 | free_workqueue_attrs(attrs); |
5265 | return ret ?: count; | |
65758202 TH |
5266 | } |
5267 | ||
6ba94429 FW |
5268 | static struct device_attribute wq_sysfs_unbound_attrs[] = { |
5269 | __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL), | |
5270 | __ATTR(nice, 0644, wq_nice_show, wq_nice_store), | |
5271 | __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), | |
5272 | __ATTR(numa, 0644, wq_numa_show, wq_numa_store), | |
5273 | __ATTR_NULL, | |
5274 | }; | |
8ccad40d | 5275 | |
6ba94429 FW |
5276 | static struct bus_type wq_subsys = { |
5277 | .name = "workqueue", | |
5278 | .dev_groups = wq_sysfs_groups, | |
2d3854a3 RR |
5279 | }; |
5280 | ||
b05a7928 FW |
5281 | static ssize_t wq_unbound_cpumask_show(struct device *dev, |
5282 | struct device_attribute *attr, char *buf) | |
5283 | { | |
5284 | int written; | |
5285 | ||
042f7df1 | 5286 | mutex_lock(&wq_pool_mutex); |
b05a7928 FW |
5287 | written = scnprintf(buf, PAGE_SIZE, "%*pb\n", |
5288 | cpumask_pr_args(wq_unbound_cpumask)); | |
042f7df1 | 5289 | mutex_unlock(&wq_pool_mutex); |
b05a7928 FW |
5290 | |
5291 | return written; | |
5292 | } | |
5293 | ||
042f7df1 LJ |
5294 | static ssize_t wq_unbound_cpumask_store(struct device *dev, |
5295 | struct device_attribute *attr, const char *buf, size_t count) | |
5296 | { | |
5297 | cpumask_var_t cpumask; | |
5298 | int ret; | |
5299 | ||
5300 | if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) | |
5301 | return -ENOMEM; | |
5302 | ||
5303 | ret = cpumask_parse(buf, cpumask); | |
5304 | if (!ret) | |
5305 | ret = workqueue_set_unbound_cpumask(cpumask); | |
5306 | ||
5307 | free_cpumask_var(cpumask); | |
5308 | return ret ? ret : count; | |
5309 | } | |
5310 | ||
b05a7928 | 5311 | static struct device_attribute wq_sysfs_cpumask_attr = |
042f7df1 LJ |
5312 | __ATTR(cpumask, 0644, wq_unbound_cpumask_show, |
5313 | wq_unbound_cpumask_store); | |
b05a7928 | 5314 | |
6ba94429 | 5315 | static int __init wq_sysfs_init(void) |
2d3854a3 | 5316 | { |
b05a7928 FW |
5317 | int err; |
5318 | ||
5319 | err = subsys_virtual_register(&wq_subsys, NULL); | |
5320 | if (err) | |
5321 | return err; | |
5322 | ||
5323 | return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr); | |
2d3854a3 | 5324 | } |
6ba94429 | 5325 | core_initcall(wq_sysfs_init); |
2d3854a3 | 5326 | |
6ba94429 | 5327 | static void wq_device_release(struct device *dev) |
2d3854a3 | 5328 | { |
6ba94429 | 5329 | struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); |
6b44003e | 5330 | |
6ba94429 | 5331 | kfree(wq_dev); |
2d3854a3 | 5332 | } |
a0a1a5fd TH |
5333 | |
5334 | /** | |
6ba94429 FW |
5335 | * workqueue_sysfs_register - make a workqueue visible in sysfs |
5336 | * @wq: the workqueue to register | |
a0a1a5fd | 5337 | * |
6ba94429 FW |
5338 | * Expose @wq in sysfs under /sys/bus/workqueue/devices. |
5339 | * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set | |
5340 | * which is the preferred method. | |
a0a1a5fd | 5341 | * |
6ba94429 FW |
5342 | * Workqueue user should use this function directly iff it wants to apply |
5343 | * workqueue_attrs before making the workqueue visible in sysfs; otherwise, | |
5344 | * apply_workqueue_attrs() may race against userland updating the | |
5345 | * attributes. | |
5346 | * | |
5347 | * Return: 0 on success, -errno on failure. | |
a0a1a5fd | 5348 | */ |
6ba94429 | 5349 | int workqueue_sysfs_register(struct workqueue_struct *wq) |
a0a1a5fd | 5350 | { |
6ba94429 FW |
5351 | struct wq_device *wq_dev; |
5352 | int ret; | |
a0a1a5fd | 5353 | |
6ba94429 | 5354 | /* |
402dd89d | 5355 | * Adjusting max_active or creating new pwqs by applying |
6ba94429 FW |
5356 | * attributes breaks ordering guarantee. Disallow exposing ordered |
5357 | * workqueues. | |
5358 | */ | |
0a94efb5 | 5359 | if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) |
6ba94429 | 5360 | return -EINVAL; |
a0a1a5fd | 5361 | |
6ba94429 FW |
5362 | wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); |
5363 | if (!wq_dev) | |
5364 | return -ENOMEM; | |
5bcab335 | 5365 | |
6ba94429 FW |
5366 | wq_dev->wq = wq; |
5367 | wq_dev->dev.bus = &wq_subsys; | |
6ba94429 | 5368 | wq_dev->dev.release = wq_device_release; |
23217b44 | 5369 | dev_set_name(&wq_dev->dev, "%s", wq->name); |
a0a1a5fd | 5370 | |
6ba94429 FW |
5371 | /* |
5372 | * unbound_attrs are created separately. Suppress uevent until | |
5373 | * everything is ready. | |
5374 | */ | |
5375 | dev_set_uevent_suppress(&wq_dev->dev, true); | |
a0a1a5fd | 5376 | |
6ba94429 FW |
5377 | ret = device_register(&wq_dev->dev); |
5378 | if (ret) { | |
7c84e5e9 | 5379 | put_device(&wq_dev->dev); |
6ba94429 FW |
5380 | wq->wq_dev = NULL; |
5381 | return ret; | |
5382 | } | |
a0a1a5fd | 5383 | |
6ba94429 FW |
5384 | if (wq->flags & WQ_UNBOUND) { |
5385 | struct device_attribute *attr; | |
a0a1a5fd | 5386 | |
6ba94429 FW |
5387 | for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { |
5388 | ret = device_create_file(&wq_dev->dev, attr); | |
5389 | if (ret) { | |
5390 | device_unregister(&wq_dev->dev); | |
5391 | wq->wq_dev = NULL; | |
5392 | return ret; | |
a0a1a5fd TH |
5393 | } |
5394 | } | |
5395 | } | |
6ba94429 FW |
5396 | |
5397 | dev_set_uevent_suppress(&wq_dev->dev, false); | |
5398 | kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); | |
5399 | return 0; | |
a0a1a5fd TH |
5400 | } |
5401 | ||
5402 | /** | |
6ba94429 FW |
5403 | * workqueue_sysfs_unregister - undo workqueue_sysfs_register() |
5404 | * @wq: the workqueue to unregister | |
a0a1a5fd | 5405 | * |
6ba94429 | 5406 | * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. |
a0a1a5fd | 5407 | */ |
6ba94429 | 5408 | static void workqueue_sysfs_unregister(struct workqueue_struct *wq) |
a0a1a5fd | 5409 | { |
6ba94429 | 5410 | struct wq_device *wq_dev = wq->wq_dev; |
8b03ae3c | 5411 | |
6ba94429 FW |
5412 | if (!wq->wq_dev) |
5413 | return; | |
a0a1a5fd | 5414 | |
6ba94429 FW |
5415 | wq->wq_dev = NULL; |
5416 | device_unregister(&wq_dev->dev); | |
a0a1a5fd | 5417 | } |
6ba94429 FW |
5418 | #else /* CONFIG_SYSFS */ |
5419 | static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } | |
5420 | #endif /* CONFIG_SYSFS */ | |
a0a1a5fd | 5421 | |
82607adc TH |
5422 | /* |
5423 | * Workqueue watchdog. | |
5424 | * | |
5425 | * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal | |
5426 | * flush dependency, a concurrency managed work item which stays RUNNING | |
5427 | * indefinitely. Workqueue stalls can be very difficult to debug as the | |
5428 | * usual warning mechanisms don't trigger and internal workqueue state is | |
5429 | * largely opaque. | |
5430 | * | |
5431 | * Workqueue watchdog monitors all worker pools periodically and dumps | |
5432 | * state if some pools failed to make forward progress for a while where | |
5433 | * forward progress is defined as the first item on ->worklist changing. | |
5434 | * | |
5435 | * This mechanism is controlled through the kernel parameter | |
5436 | * "workqueue.watchdog_thresh" which can be updated at runtime through the | |
5437 | * corresponding sysfs parameter file. | |
5438 | */ | |
5439 | #ifdef CONFIG_WQ_WATCHDOG | |
5440 | ||
5441 | static void wq_watchdog_timer_fn(unsigned long data); | |
5442 | ||
5443 | static unsigned long wq_watchdog_thresh = 30; | |
5444 | static struct timer_list wq_watchdog_timer = | |
5445 | TIMER_DEFERRED_INITIALIZER(wq_watchdog_timer_fn, 0, 0); | |
5446 | ||
5447 | static unsigned long wq_watchdog_touched = INITIAL_JIFFIES; | |
5448 | static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES; | |
5449 | ||
5450 | static void wq_watchdog_reset_touched(void) | |
5451 | { | |
5452 | int cpu; | |
5453 | ||
5454 | wq_watchdog_touched = jiffies; | |
5455 | for_each_possible_cpu(cpu) | |
5456 | per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; | |
5457 | } | |
5458 | ||
5459 | static void wq_watchdog_timer_fn(unsigned long data) | |
5460 | { | |
5461 | unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ; | |
5462 | bool lockup_detected = false; | |
5463 | struct worker_pool *pool; | |
5464 | int pi; | |
5465 | ||
5466 | if (!thresh) | |
5467 | return; | |
5468 | ||
5469 | rcu_read_lock(); | |
5470 | ||
5471 | for_each_pool(pool, pi) { | |
5472 | unsigned long pool_ts, touched, ts; | |
5473 | ||
5474 | if (list_empty(&pool->worklist)) | |
5475 | continue; | |
5476 | ||
5477 | /* get the latest of pool and touched timestamps */ | |
5478 | pool_ts = READ_ONCE(pool->watchdog_ts); | |
5479 | touched = READ_ONCE(wq_watchdog_touched); | |
5480 | ||
5481 | if (time_after(pool_ts, touched)) | |
5482 | ts = pool_ts; | |
5483 | else | |
5484 | ts = touched; | |
5485 | ||
5486 | if (pool->cpu >= 0) { | |
5487 | unsigned long cpu_touched = | |
5488 | READ_ONCE(per_cpu(wq_watchdog_touched_cpu, | |
5489 | pool->cpu)); | |
5490 | if (time_after(cpu_touched, ts)) | |
5491 | ts = cpu_touched; | |
5492 | } | |
5493 | ||
5494 | /* did we stall? */ | |
5495 | if (time_after(jiffies, ts + thresh)) { | |
5496 | lockup_detected = true; | |
5497 | pr_emerg("BUG: workqueue lockup - pool"); | |
5498 | pr_cont_pool_info(pool); | |
5499 | pr_cont(" stuck for %us!\n", | |
5500 | jiffies_to_msecs(jiffies - pool_ts) / 1000); | |
5501 | } | |
5502 | } | |
5503 | ||
5504 | rcu_read_unlock(); | |
5505 | ||
5506 | if (lockup_detected) | |
5507 | show_workqueue_state(); | |
5508 | ||
5509 | wq_watchdog_reset_touched(); | |
5510 | mod_timer(&wq_watchdog_timer, jiffies + thresh); | |
5511 | } | |
5512 | ||
63a0f9de | 5513 | notrace void wq_watchdog_touch(int cpu) |
82607adc TH |
5514 | { |
5515 | if (cpu >= 0) | |
5516 | per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; | |
5517 | else | |
5518 | wq_watchdog_touched = jiffies; | |
5519 | } | |
5520 | ||
5521 | static void wq_watchdog_set_thresh(unsigned long thresh) | |
5522 | { | |
5523 | wq_watchdog_thresh = 0; | |
5524 | del_timer_sync(&wq_watchdog_timer); | |
5525 | ||
5526 | if (thresh) { | |
5527 | wq_watchdog_thresh = thresh; | |
5528 | wq_watchdog_reset_touched(); | |
5529 | mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ); | |
5530 | } | |
5531 | } | |
5532 | ||
5533 | static int wq_watchdog_param_set_thresh(const char *val, | |
5534 | const struct kernel_param *kp) | |
5535 | { | |
5536 | unsigned long thresh; | |
5537 | int ret; | |
5538 | ||
5539 | ret = kstrtoul(val, 0, &thresh); | |
5540 | if (ret) | |
5541 | return ret; | |
5542 | ||
5543 | if (system_wq) | |
5544 | wq_watchdog_set_thresh(thresh); | |
5545 | else | |
5546 | wq_watchdog_thresh = thresh; | |
5547 | ||
5548 | return 0; | |
5549 | } | |
5550 | ||
5551 | static const struct kernel_param_ops wq_watchdog_thresh_ops = { | |
5552 | .set = wq_watchdog_param_set_thresh, | |
5553 | .get = param_get_ulong, | |
5554 | }; | |
5555 | ||
5556 | module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh, | |
5557 | 0644); | |
5558 | ||
5559 | static void wq_watchdog_init(void) | |
5560 | { | |
5561 | wq_watchdog_set_thresh(wq_watchdog_thresh); | |
5562 | } | |
5563 | ||
5564 | #else /* CONFIG_WQ_WATCHDOG */ | |
5565 | ||
5566 | static inline void wq_watchdog_init(void) { } | |
5567 | ||
5568 | #endif /* CONFIG_WQ_WATCHDOG */ | |
5569 | ||
bce90380 TH |
5570 | static void __init wq_numa_init(void) |
5571 | { | |
5572 | cpumask_var_t *tbl; | |
5573 | int node, cpu; | |
5574 | ||
bce90380 TH |
5575 | if (num_possible_nodes() <= 1) |
5576 | return; | |
5577 | ||
d55262c4 TH |
5578 | if (wq_disable_numa) { |
5579 | pr_info("workqueue: NUMA affinity support disabled\n"); | |
5580 | return; | |
5581 | } | |
5582 | ||
4c16bd32 TH |
5583 | wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL); |
5584 | BUG_ON(!wq_update_unbound_numa_attrs_buf); | |
5585 | ||
bce90380 TH |
5586 | /* |
5587 | * We want masks of possible CPUs of each node which isn't readily | |
5588 | * available. Build one from cpu_to_node() which should have been | |
5589 | * fully initialized by now. | |
5590 | */ | |
ddcb57e2 | 5591 | tbl = kzalloc(nr_node_ids * sizeof(tbl[0]), GFP_KERNEL); |
bce90380 TH |
5592 | BUG_ON(!tbl); |
5593 | ||
5594 | for_each_node(node) | |
5a6024f1 | 5595 | BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL, |
1be0c25d | 5596 | node_online(node) ? node : NUMA_NO_NODE)); |
bce90380 TH |
5597 | |
5598 | for_each_possible_cpu(cpu) { | |
5599 | node = cpu_to_node(cpu); | |
5600 | if (WARN_ON(node == NUMA_NO_NODE)) { | |
5601 | pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu); | |
5602 | /* happens iff arch is bonkers, let's just proceed */ | |
5603 | return; | |
5604 | } | |
5605 | cpumask_set_cpu(cpu, tbl[node]); | |
5606 | } | |
5607 | ||
5608 | wq_numa_possible_cpumask = tbl; | |
5609 | wq_numa_enabled = true; | |
5610 | } | |
5611 | ||
3347fa09 TH |
5612 | /** |
5613 | * workqueue_init_early - early init for workqueue subsystem | |
5614 | * | |
5615 | * This is the first half of two-staged workqueue subsystem initialization | |
5616 | * and invoked as soon as the bare basics - memory allocation, cpumasks and | |
5617 | * idr are up. It sets up all the data structures and system workqueues | |
5618 | * and allows early boot code to create workqueues and queue/cancel work | |
5619 | * items. Actual work item execution starts only after kthreads can be | |
5620 | * created and scheduled right before early initcalls. | |
5621 | */ | |
5622 | int __init workqueue_init_early(void) | |
1da177e4 | 5623 | { |
7a4e344c TH |
5624 | int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; |
5625 | int i, cpu; | |
c34056a3 | 5626 | |
e904e6c2 TH |
5627 | WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); |
5628 | ||
b05a7928 FW |
5629 | BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL)); |
5630 | cpumask_copy(wq_unbound_cpumask, cpu_possible_mask); | |
5631 | ||
e904e6c2 TH |
5632 | pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); |
5633 | ||
706026c2 | 5634 | /* initialize CPU pools */ |
29c91e99 | 5635 | for_each_possible_cpu(cpu) { |
4ce62e9e | 5636 | struct worker_pool *pool; |
8b03ae3c | 5637 | |
7a4e344c | 5638 | i = 0; |
f02ae73a | 5639 | for_each_cpu_worker_pool(pool, cpu) { |
7a4e344c | 5640 | BUG_ON(init_worker_pool(pool)); |
ec22ca5e | 5641 | pool->cpu = cpu; |
29c91e99 | 5642 | cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); |
7a4e344c | 5643 | pool->attrs->nice = std_nice[i++]; |
f3f90ad4 | 5644 | pool->node = cpu_to_node(cpu); |
7a4e344c | 5645 | |
9daf9e67 | 5646 | /* alloc pool ID */ |
68e13a67 | 5647 | mutex_lock(&wq_pool_mutex); |
9daf9e67 | 5648 | BUG_ON(worker_pool_assign_id(pool)); |
68e13a67 | 5649 | mutex_unlock(&wq_pool_mutex); |
4ce62e9e | 5650 | } |
8b03ae3c TH |
5651 | } |
5652 | ||
8a2b7538 | 5653 | /* create default unbound and ordered wq attrs */ |
29c91e99 TH |
5654 | for (i = 0; i < NR_STD_WORKER_POOLS; i++) { |
5655 | struct workqueue_attrs *attrs; | |
5656 | ||
5657 | BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); | |
29c91e99 | 5658 | attrs->nice = std_nice[i]; |
29c91e99 | 5659 | unbound_std_wq_attrs[i] = attrs; |
8a2b7538 TH |
5660 | |
5661 | /* | |
5662 | * An ordered wq should have only one pwq as ordering is | |
5663 | * guaranteed by max_active which is enforced by pwqs. | |
5664 | * Turn off NUMA so that dfl_pwq is used for all nodes. | |
5665 | */ | |
5666 | BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); | |
5667 | attrs->nice = std_nice[i]; | |
5668 | attrs->no_numa = true; | |
5669 | ordered_wq_attrs[i] = attrs; | |
29c91e99 TH |
5670 | } |
5671 | ||
d320c038 | 5672 | system_wq = alloc_workqueue("events", 0, 0); |
1aabe902 | 5673 | system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); |
d320c038 | 5674 | system_long_wq = alloc_workqueue("events_long", 0, 0); |
f3421797 TH |
5675 | system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, |
5676 | WQ_UNBOUND_MAX_ACTIVE); | |
24d51add TH |
5677 | system_freezable_wq = alloc_workqueue("events_freezable", |
5678 | WQ_FREEZABLE, 0); | |
0668106c VK |
5679 | system_power_efficient_wq = alloc_workqueue("events_power_efficient", |
5680 | WQ_POWER_EFFICIENT, 0); | |
5681 | system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient", | |
5682 | WQ_FREEZABLE | WQ_POWER_EFFICIENT, | |
5683 | 0); | |
1aabe902 | 5684 | BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || |
0668106c VK |
5685 | !system_unbound_wq || !system_freezable_wq || |
5686 | !system_power_efficient_wq || | |
5687 | !system_freezable_power_efficient_wq); | |
82607adc | 5688 | |
3347fa09 TH |
5689 | return 0; |
5690 | } | |
5691 | ||
5692 | /** | |
5693 | * workqueue_init - bring workqueue subsystem fully online | |
5694 | * | |
5695 | * This is the latter half of two-staged workqueue subsystem initialization | |
5696 | * and invoked as soon as kthreads can be created and scheduled. | |
5697 | * Workqueues have been created and work items queued on them, but there | |
5698 | * are no kworkers executing the work items yet. Populate the worker pools | |
5699 | * with the initial workers and enable future kworker creations. | |
5700 | */ | |
5701 | int __init workqueue_init(void) | |
5702 | { | |
2186d9f9 | 5703 | struct workqueue_struct *wq; |
3347fa09 TH |
5704 | struct worker_pool *pool; |
5705 | int cpu, bkt; | |
5706 | ||
2186d9f9 TH |
5707 | /* |
5708 | * It'd be simpler to initialize NUMA in workqueue_init_early() but | |
5709 | * CPU to node mapping may not be available that early on some | |
5710 | * archs such as power and arm64. As per-cpu pools created | |
5711 | * previously could be missing node hint and unbound pools NUMA | |
5712 | * affinity, fix them up. | |
5713 | */ | |
5714 | wq_numa_init(); | |
5715 | ||
5716 | mutex_lock(&wq_pool_mutex); | |
5717 | ||
5718 | for_each_possible_cpu(cpu) { | |
5719 | for_each_cpu_worker_pool(pool, cpu) { | |
5720 | pool->node = cpu_to_node(cpu); | |
5721 | } | |
5722 | } | |
5723 | ||
5724 | list_for_each_entry(wq, &workqueues, list) | |
5725 | wq_update_unbound_numa(wq, smp_processor_id(), true); | |
5726 | ||
5727 | mutex_unlock(&wq_pool_mutex); | |
5728 | ||
3347fa09 TH |
5729 | /* create the initial workers */ |
5730 | for_each_online_cpu(cpu) { | |
5731 | for_each_cpu_worker_pool(pool, cpu) { | |
5732 | pool->flags &= ~POOL_DISASSOCIATED; | |
5733 | BUG_ON(!create_worker(pool)); | |
5734 | } | |
5735 | } | |
5736 | ||
5737 | hash_for_each(unbound_pool_hash, bkt, pool, hash_node) | |
5738 | BUG_ON(!create_worker(pool)); | |
5739 | ||
5740 | wq_online = true; | |
82607adc TH |
5741 | wq_watchdog_init(); |
5742 | ||
6ee0578b | 5743 | return 0; |
1da177e4 | 5744 | } |