Commit | Line | Data |
---|---|---|
457c8996 | 1 | // SPDX-License-Identifier: GPL-2.0-only |
1da177e4 | 2 | /* |
c54fce6e | 3 | * kernel/workqueue.c - generic async execution with shared worker pool |
1da177e4 | 4 | * |
c54fce6e | 5 | * Copyright (C) 2002 Ingo Molnar |
1da177e4 | 6 | * |
c54fce6e TH |
7 | * Derived from the taskqueue/keventd code by: |
8 | * David Woodhouse <dwmw2@infradead.org> | |
9 | * Andrew Morton | |
10 | * Kai Petzke <wpp@marie.physik.tu-berlin.de> | |
11 | * Theodore Ts'o <tytso@mit.edu> | |
1da177e4 | 12 | * |
c54fce6e | 13 | * Made to use alloc_percpu by Christoph Lameter. |
1da177e4 | 14 | * |
c54fce6e TH |
15 | * Copyright (C) 2010 SUSE Linux Products GmbH |
16 | * Copyright (C) 2010 Tejun Heo <tj@kernel.org> | |
89ada679 | 17 | * |
c54fce6e TH |
18 | * This is the generic async execution mechanism. Work items as are |
19 | * executed in process context. The worker pool is shared and | |
b11895c4 L |
20 | * automatically managed. There are two worker pools for each CPU (one for |
21 | * normal work items and the other for high priority ones) and some extra | |
22 | * pools for workqueues which are not bound to any specific CPU - the | |
23 | * number of these backing pools is dynamic. | |
c54fce6e | 24 | * |
9a261491 | 25 | * Please read Documentation/core-api/workqueue.rst for details. |
1da177e4 LT |
26 | */ |
27 | ||
9984de1a | 28 | #include <linux/export.h> |
1da177e4 LT |
29 | #include <linux/kernel.h> |
30 | #include <linux/sched.h> | |
31 | #include <linux/init.h> | |
32 | #include <linux/signal.h> | |
33 | #include <linux/completion.h> | |
34 | #include <linux/workqueue.h> | |
35 | #include <linux/slab.h> | |
36 | #include <linux/cpu.h> | |
37 | #include <linux/notifier.h> | |
38 | #include <linux/kthread.h> | |
1fa44eca | 39 | #include <linux/hardirq.h> |
46934023 | 40 | #include <linux/mempolicy.h> |
341a5958 | 41 | #include <linux/freezer.h> |
d5abe669 | 42 | #include <linux/debug_locks.h> |
4e6045f1 | 43 | #include <linux/lockdep.h> |
c34056a3 | 44 | #include <linux/idr.h> |
29c91e99 | 45 | #include <linux/jhash.h> |
42f8570f | 46 | #include <linux/hashtable.h> |
76af4d93 | 47 | #include <linux/rculist.h> |
bce90380 | 48 | #include <linux/nodemask.h> |
4c16bd32 | 49 | #include <linux/moduleparam.h> |
3d1cb205 | 50 | #include <linux/uaccess.h> |
c98a9805 | 51 | #include <linux/sched/isolation.h> |
62635ea8 | 52 | #include <linux/nmi.h> |
940d71c6 | 53 | #include <linux/kvm_para.h> |
e22bee78 | 54 | |
ea138446 | 55 | #include "workqueue_internal.h" |
1da177e4 | 56 | |
c8e55f36 | 57 | enum { |
24647570 TH |
58 | /* |
59 | * worker_pool flags | |
bc2ae0f5 | 60 | * |
24647570 | 61 | * A bound pool is either associated or disassociated with its CPU. |
bc2ae0f5 TH |
62 | * While associated (!DISASSOCIATED), all workers are bound to the |
63 | * CPU and none has %WORKER_UNBOUND set and concurrency management | |
64 | * is in effect. | |
65 | * | |
66 | * While DISASSOCIATED, the cpu may be offline and all workers have | |
67 | * %WORKER_UNBOUND set and concurrency management disabled, and may | |
24647570 | 68 | * be executing on any CPU. The pool behaves as an unbound one. |
bc2ae0f5 | 69 | * |
bc3a1afc | 70 | * Note that DISASSOCIATED should be flipped only while holding |
1258fae7 | 71 | * wq_pool_attach_mutex to avoid changing binding state while |
4736cbf7 | 72 | * worker_attach_to_pool() is in progress. |
bc2ae0f5 | 73 | */ |
692b4825 | 74 | POOL_MANAGER_ACTIVE = 1 << 0, /* being managed */ |
24647570 | 75 | POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */ |
db7bccf4 | 76 | |
c8e55f36 | 77 | /* worker flags */ |
c8e55f36 TH |
78 | WORKER_DIE = 1 << 1, /* die die die */ |
79 | WORKER_IDLE = 1 << 2, /* is idle */ | |
e22bee78 | 80 | WORKER_PREP = 1 << 3, /* preparing to run works */ |
fb0e7beb | 81 | WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */ |
f3421797 | 82 | WORKER_UNBOUND = 1 << 7, /* worker is unbound */ |
a9ab775b | 83 | WORKER_REBOUND = 1 << 8, /* worker was rebound */ |
e22bee78 | 84 | |
a9ab775b TH |
85 | WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE | |
86 | WORKER_UNBOUND | WORKER_REBOUND, | |
db7bccf4 | 87 | |
e34cdddb | 88 | NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */ |
4ce62e9e | 89 | |
29c91e99 | 90 | UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */ |
c8e55f36 | 91 | BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */ |
db7bccf4 | 92 | |
e22bee78 TH |
93 | MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */ |
94 | IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */ | |
95 | ||
3233cdbd TH |
96 | MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2, |
97 | /* call for help after 10ms | |
98 | (min two ticks) */ | |
e22bee78 TH |
99 | MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */ |
100 | CREATE_COOLDOWN = HZ, /* time to breath after fail */ | |
e22bee78 TH |
101 | |
102 | /* | |
103 | * Rescue workers are used only on emergencies and shared by | |
8698a745 | 104 | * all cpus. Give MIN_NICE. |
e22bee78 | 105 | */ |
8698a745 DY |
106 | RESCUER_NICE_LEVEL = MIN_NICE, |
107 | HIGHPRI_NICE_LEVEL = MIN_NICE, | |
ecf6881f TH |
108 | |
109 | WQ_NAME_LEN = 24, | |
c8e55f36 | 110 | }; |
1da177e4 LT |
111 | |
112 | /* | |
4690c4ab TH |
113 | * Structure fields follow one of the following exclusion rules. |
114 | * | |
e41e704b TH |
115 | * I: Modifiable by initialization/destruction paths and read-only for |
116 | * everyone else. | |
4690c4ab | 117 | * |
e22bee78 TH |
118 | * P: Preemption protected. Disabling preemption is enough and should |
119 | * only be modified and accessed from the local cpu. | |
120 | * | |
d565ed63 | 121 | * L: pool->lock protected. Access with pool->lock held. |
4690c4ab | 122 | * |
d565ed63 TH |
123 | * X: During normal operation, modification requires pool->lock and should |
124 | * be done only from local cpu. Either disabling preemption on local | |
125 | * cpu or grabbing pool->lock is enough for read access. If | |
126 | * POOL_DISASSOCIATED is set, it's identical to L. | |
e22bee78 | 127 | * |
1258fae7 | 128 | * A: wq_pool_attach_mutex protected. |
822d8405 | 129 | * |
68e13a67 | 130 | * PL: wq_pool_mutex protected. |
5bcab335 | 131 | * |
24acfb71 | 132 | * PR: wq_pool_mutex protected for writes. RCU protected for reads. |
76af4d93 | 133 | * |
5b95e1af LJ |
134 | * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads. |
135 | * | |
136 | * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or | |
24acfb71 | 137 | * RCU for reads. |
5b95e1af | 138 | * |
3c25a55d LJ |
139 | * WQ: wq->mutex protected. |
140 | * | |
24acfb71 | 141 | * WR: wq->mutex protected for writes. RCU protected for reads. |
2e109a28 TH |
142 | * |
143 | * MD: wq_mayday_lock protected. | |
1da177e4 | 144 | */ |
1da177e4 | 145 | |
2eaebdb3 | 146 | /* struct worker is defined in workqueue_internal.h */ |
c34056a3 | 147 | |
bd7bdd43 | 148 | struct worker_pool { |
a9b8a985 | 149 | raw_spinlock_t lock; /* the pool lock */ |
d84ff051 | 150 | int cpu; /* I: the associated cpu */ |
f3f90ad4 | 151 | int node; /* I: the associated node ID */ |
9daf9e67 | 152 | int id; /* I: pool ID */ |
11ebea50 | 153 | unsigned int flags; /* X: flags */ |
bd7bdd43 | 154 | |
82607adc TH |
155 | unsigned long watchdog_ts; /* L: watchdog timestamp */ |
156 | ||
bc35f7ef LJ |
157 | /* |
158 | * The counter is incremented in a process context on the associated CPU | |
159 | * w/ preemption disabled, and decremented or reset in the same context | |
160 | * but w/ pool->lock held. The readers grab pool->lock and are | |
161 | * guaranteed to see if the counter reached zero. | |
162 | */ | |
163 | int nr_running; | |
84f91c62 | 164 | |
bd7bdd43 | 165 | struct list_head worklist; /* L: list of pending works */ |
ea1abd61 | 166 | |
5826cc8f LJ |
167 | int nr_workers; /* L: total number of workers */ |
168 | int nr_idle; /* L: currently idle workers */ | |
bd7bdd43 | 169 | |
2c1f1a91 | 170 | struct list_head idle_list; /* L: list of idle workers */ |
bd7bdd43 | 171 | struct timer_list idle_timer; /* L: worker idle timeout */ |
3f959aa3 VS |
172 | struct work_struct idle_cull_work; /* L: worker idle cleanup */ |
173 | ||
174 | struct timer_list mayday_timer; /* L: SOS timer for workers */ | |
bd7bdd43 | 175 | |
c5aa87bb | 176 | /* a workers is either on busy_hash or idle_list, or the manager */ |
c9e7cf27 TH |
177 | DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER); |
178 | /* L: hash of busy workers */ | |
179 | ||
2607d7a6 | 180 | struct worker *manager; /* L: purely informational */ |
92f9c5c4 | 181 | struct list_head workers; /* A: attached workers */ |
e02b9312 | 182 | struct list_head dying_workers; /* A: workers about to die */ |
60f5a4bc | 183 | struct completion *detach_completion; /* all workers detached */ |
e19e397a | 184 | |
7cda9aae | 185 | struct ida worker_ida; /* worker IDs for task name */ |
e19e397a | 186 | |
7a4e344c | 187 | struct workqueue_attrs *attrs; /* I: worker attributes */ |
68e13a67 LJ |
188 | struct hlist_node hash_node; /* PL: unbound_pool_hash node */ |
189 | int refcnt; /* PL: refcnt for unbound pools */ | |
7a4e344c | 190 | |
29c91e99 | 191 | /* |
24acfb71 | 192 | * Destruction of pool is RCU protected to allow dereferences |
29c91e99 TH |
193 | * from get_work_pool(). |
194 | */ | |
195 | struct rcu_head rcu; | |
84f91c62 | 196 | }; |
8b03ae3c | 197 | |
1da177e4 | 198 | /* |
112202d9 TH |
199 | * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS |
200 | * of work_struct->data are used for flags and the remaining high bits | |
201 | * point to the pwq; thus, pwqs need to be aligned at two's power of the | |
202 | * number of flag bits. | |
1da177e4 | 203 | */ |
112202d9 | 204 | struct pool_workqueue { |
bd7bdd43 | 205 | struct worker_pool *pool; /* I: the associated pool */ |
4690c4ab | 206 | struct workqueue_struct *wq; /* I: the owning workqueue */ |
73f53c4a TH |
207 | int work_color; /* L: current color */ |
208 | int flush_color; /* L: flushing color */ | |
8864b4e5 | 209 | int refcnt; /* L: reference count */ |
73f53c4a TH |
210 | int nr_in_flight[WORK_NR_COLORS]; |
211 | /* L: nr of in_flight works */ | |
018f3a13 LJ |
212 | |
213 | /* | |
214 | * nr_active management and WORK_STRUCT_INACTIVE: | |
215 | * | |
216 | * When pwq->nr_active >= max_active, new work item is queued to | |
217 | * pwq->inactive_works instead of pool->worklist and marked with | |
218 | * WORK_STRUCT_INACTIVE. | |
219 | * | |
220 | * All work items marked with WORK_STRUCT_INACTIVE do not participate | |
221 | * in pwq->nr_active and all work items in pwq->inactive_works are | |
222 | * marked with WORK_STRUCT_INACTIVE. But not all WORK_STRUCT_INACTIVE | |
223 | * work items are in pwq->inactive_works. Some of them are ready to | |
224 | * run in pool->worklist or worker->scheduled. Those work itmes are | |
225 | * only struct wq_barrier which is used for flush_work() and should | |
226 | * not participate in pwq->nr_active. For non-barrier work item, it | |
227 | * is marked with WORK_STRUCT_INACTIVE iff it is in pwq->inactive_works. | |
228 | */ | |
1e19ffc6 | 229 | int nr_active; /* L: nr of active works */ |
a0a1a5fd | 230 | int max_active; /* L: max active works */ |
f97a4a1a | 231 | struct list_head inactive_works; /* L: inactive works */ |
3c25a55d | 232 | struct list_head pwqs_node; /* WR: node on wq->pwqs */ |
2e109a28 | 233 | struct list_head mayday_node; /* MD: node on wq->maydays */ |
8864b4e5 TH |
234 | |
235 | /* | |
236 | * Release of unbound pwq is punted to system_wq. See put_pwq() | |
237 | * and pwq_unbound_release_workfn() for details. pool_workqueue | |
24acfb71 | 238 | * itself is also RCU protected so that the first pwq can be |
b09f4fd3 | 239 | * determined without grabbing wq->mutex. |
8864b4e5 TH |
240 | */ |
241 | struct work_struct unbound_release_work; | |
242 | struct rcu_head rcu; | |
e904e6c2 | 243 | } __aligned(1 << WORK_STRUCT_FLAG_BITS); |
1da177e4 | 244 | |
73f53c4a TH |
245 | /* |
246 | * Structure used to wait for workqueue flush. | |
247 | */ | |
248 | struct wq_flusher { | |
3c25a55d LJ |
249 | struct list_head list; /* WQ: list of flushers */ |
250 | int flush_color; /* WQ: flush color waiting for */ | |
73f53c4a TH |
251 | struct completion done; /* flush completion */ |
252 | }; | |
253 | ||
226223ab TH |
254 | struct wq_device; |
255 | ||
1da177e4 | 256 | /* |
c5aa87bb TH |
257 | * The externally visible workqueue. It relays the issued work items to |
258 | * the appropriate worker_pool through its pool_workqueues. | |
1da177e4 LT |
259 | */ |
260 | struct workqueue_struct { | |
3c25a55d | 261 | struct list_head pwqs; /* WR: all pwqs of this wq */ |
e2dca7ad | 262 | struct list_head list; /* PR: list of all workqueues */ |
73f53c4a | 263 | |
3c25a55d LJ |
264 | struct mutex mutex; /* protects this wq */ |
265 | int work_color; /* WQ: current work color */ | |
266 | int flush_color; /* WQ: current flush color */ | |
112202d9 | 267 | atomic_t nr_pwqs_to_flush; /* flush in progress */ |
3c25a55d LJ |
268 | struct wq_flusher *first_flusher; /* WQ: first flusher */ |
269 | struct list_head flusher_queue; /* WQ: flush waiters */ | |
270 | struct list_head flusher_overflow; /* WQ: flush overflow list */ | |
73f53c4a | 271 | |
2e109a28 | 272 | struct list_head maydays; /* MD: pwqs requesting rescue */ |
30ae2fc0 | 273 | struct worker *rescuer; /* MD: rescue worker */ |
e22bee78 | 274 | |
87fc741e | 275 | int nr_drainers; /* WQ: drain in progress */ |
a357fc03 | 276 | int saved_max_active; /* WQ: saved pwq max_active */ |
226223ab | 277 | |
5b95e1af LJ |
278 | struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */ |
279 | struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */ | |
6029a918 | 280 | |
226223ab TH |
281 | #ifdef CONFIG_SYSFS |
282 | struct wq_device *wq_dev; /* I: for sysfs interface */ | |
283 | #endif | |
4e6045f1 | 284 | #ifdef CONFIG_LOCKDEP |
669de8bd BVA |
285 | char *lock_name; |
286 | struct lock_class_key key; | |
4690c4ab | 287 | struct lockdep_map lockdep_map; |
4e6045f1 | 288 | #endif |
ecf6881f | 289 | char name[WQ_NAME_LEN]; /* I: workqueue name */ |
2728fd2f | 290 | |
e2dca7ad | 291 | /* |
24acfb71 TG |
292 | * Destruction of workqueue_struct is RCU protected to allow walking |
293 | * the workqueues list without grabbing wq_pool_mutex. | |
e2dca7ad TH |
294 | * This is used to dump all workqueues from sysrq. |
295 | */ | |
296 | struct rcu_head rcu; | |
297 | ||
2728fd2f TH |
298 | /* hot fields used during command issue, aligned to cacheline */ |
299 | unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */ | |
300 | struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */ | |
5b95e1af | 301 | struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */ |
1da177e4 LT |
302 | }; |
303 | ||
e904e6c2 TH |
304 | static struct kmem_cache *pwq_cache; |
305 | ||
bce90380 TH |
306 | static cpumask_var_t *wq_numa_possible_cpumask; |
307 | /* possible CPUs of each node */ | |
308 | ||
d55262c4 TH |
309 | static bool wq_disable_numa; |
310 | module_param_named(disable_numa, wq_disable_numa, bool, 0444); | |
311 | ||
cee22a15 | 312 | /* see the comment above the definition of WQ_POWER_EFFICIENT */ |
552f530c | 313 | static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT); |
cee22a15 VK |
314 | module_param_named(power_efficient, wq_power_efficient, bool, 0444); |
315 | ||
863b710b | 316 | static bool wq_online; /* can kworkers be created yet? */ |
3347fa09 | 317 | |
bce90380 TH |
318 | static bool wq_numa_enabled; /* unbound NUMA affinity enabled */ |
319 | ||
4c16bd32 TH |
320 | /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */ |
321 | static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf; | |
322 | ||
68e13a67 | 323 | static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ |
1258fae7 | 324 | static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */ |
a9b8a985 | 325 | static DEFINE_RAW_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ |
d8bb65ab SAS |
326 | /* wait for manager to go away */ |
327 | static struct rcuwait manager_wait = __RCUWAIT_INITIALIZER(manager_wait); | |
5bcab335 | 328 | |
e2dca7ad | 329 | static LIST_HEAD(workqueues); /* PR: list of all workqueues */ |
68e13a67 | 330 | static bool workqueue_freezing; /* PL: have wqs started freezing? */ |
7d19c5ce | 331 | |
99c621ef | 332 | /* PL&A: allowable cpus for unbound wqs and work items */ |
ef557180 MG |
333 | static cpumask_var_t wq_unbound_cpumask; |
334 | ||
335 | /* CPU where unbound work was last round robin scheduled from this CPU */ | |
336 | static DEFINE_PER_CPU(int, wq_rr_cpu_last); | |
b05a7928 | 337 | |
f303fccb TH |
338 | /* |
339 | * Local execution of unbound work items is no longer guaranteed. The | |
340 | * following always forces round-robin CPU selection on unbound work items | |
341 | * to uncover usages which depend on it. | |
342 | */ | |
343 | #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU | |
344 | static bool wq_debug_force_rr_cpu = true; | |
345 | #else | |
346 | static bool wq_debug_force_rr_cpu = false; | |
347 | #endif | |
348 | module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644); | |
349 | ||
7d19c5ce | 350 | /* the per-cpu worker pools */ |
25528213 | 351 | static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools); |
7d19c5ce | 352 | |
68e13a67 | 353 | static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ |
7d19c5ce | 354 | |
68e13a67 | 355 | /* PL: hash of all unbound pools keyed by pool->attrs */ |
29c91e99 TH |
356 | static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); |
357 | ||
c5aa87bb | 358 | /* I: attributes used when instantiating standard unbound pools on demand */ |
29c91e99 TH |
359 | static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; |
360 | ||
8a2b7538 TH |
361 | /* I: attributes used when instantiating ordered pools on demand */ |
362 | static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS]; | |
363 | ||
d320c038 | 364 | struct workqueue_struct *system_wq __read_mostly; |
ad7b1f84 | 365 | EXPORT_SYMBOL(system_wq); |
044c782c | 366 | struct workqueue_struct *system_highpri_wq __read_mostly; |
1aabe902 | 367 | EXPORT_SYMBOL_GPL(system_highpri_wq); |
044c782c | 368 | struct workqueue_struct *system_long_wq __read_mostly; |
d320c038 | 369 | EXPORT_SYMBOL_GPL(system_long_wq); |
044c782c | 370 | struct workqueue_struct *system_unbound_wq __read_mostly; |
f3421797 | 371 | EXPORT_SYMBOL_GPL(system_unbound_wq); |
044c782c | 372 | struct workqueue_struct *system_freezable_wq __read_mostly; |
24d51add | 373 | EXPORT_SYMBOL_GPL(system_freezable_wq); |
0668106c VK |
374 | struct workqueue_struct *system_power_efficient_wq __read_mostly; |
375 | EXPORT_SYMBOL_GPL(system_power_efficient_wq); | |
376 | struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly; | |
377 | EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq); | |
d320c038 | 378 | |
7d19c5ce | 379 | static int worker_thread(void *__worker); |
6ba94429 | 380 | static void workqueue_sysfs_unregister(struct workqueue_struct *wq); |
c29eb853 | 381 | static void show_pwq(struct pool_workqueue *pwq); |
55df0933 | 382 | static void show_one_worker_pool(struct worker_pool *pool); |
7d19c5ce | 383 | |
97bd2347 TH |
384 | #define CREATE_TRACE_POINTS |
385 | #include <trace/events/workqueue.h> | |
386 | ||
68e13a67 | 387 | #define assert_rcu_or_pool_mutex() \ |
24acfb71 | 388 | RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \ |
f78f5b90 | 389 | !lockdep_is_held(&wq_pool_mutex), \ |
24acfb71 | 390 | "RCU or wq_pool_mutex should be held") |
5bcab335 | 391 | |
5b95e1af | 392 | #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \ |
24acfb71 | 393 | RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \ |
f78f5b90 PM |
394 | !lockdep_is_held(&wq->mutex) && \ |
395 | !lockdep_is_held(&wq_pool_mutex), \ | |
24acfb71 | 396 | "RCU, wq->mutex or wq_pool_mutex should be held") |
5b95e1af | 397 | |
f02ae73a TH |
398 | #define for_each_cpu_worker_pool(pool, cpu) \ |
399 | for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ | |
400 | (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ | |
7a62c2c8 | 401 | (pool)++) |
4ce62e9e | 402 | |
17116969 TH |
403 | /** |
404 | * for_each_pool - iterate through all worker_pools in the system | |
405 | * @pool: iteration cursor | |
611c92a0 | 406 | * @pi: integer used for iteration |
fa1b54e6 | 407 | * |
24acfb71 | 408 | * This must be called either with wq_pool_mutex held or RCU read |
68e13a67 LJ |
409 | * locked. If the pool needs to be used beyond the locking in effect, the |
410 | * caller is responsible for guaranteeing that the pool stays online. | |
fa1b54e6 TH |
411 | * |
412 | * The if/else clause exists only for the lockdep assertion and can be | |
413 | * ignored. | |
17116969 | 414 | */ |
611c92a0 TH |
415 | #define for_each_pool(pool, pi) \ |
416 | idr_for_each_entry(&worker_pool_idr, pool, pi) \ | |
68e13a67 | 417 | if (({ assert_rcu_or_pool_mutex(); false; })) { } \ |
fa1b54e6 | 418 | else |
17116969 | 419 | |
822d8405 TH |
420 | /** |
421 | * for_each_pool_worker - iterate through all workers of a worker_pool | |
422 | * @worker: iteration cursor | |
822d8405 TH |
423 | * @pool: worker_pool to iterate workers of |
424 | * | |
1258fae7 | 425 | * This must be called with wq_pool_attach_mutex. |
822d8405 TH |
426 | * |
427 | * The if/else clause exists only for the lockdep assertion and can be | |
428 | * ignored. | |
429 | */ | |
da028469 LJ |
430 | #define for_each_pool_worker(worker, pool) \ |
431 | list_for_each_entry((worker), &(pool)->workers, node) \ | |
1258fae7 | 432 | if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \ |
822d8405 TH |
433 | else |
434 | ||
49e3cf44 TH |
435 | /** |
436 | * for_each_pwq - iterate through all pool_workqueues of the specified workqueue | |
437 | * @pwq: iteration cursor | |
438 | * @wq: the target workqueue | |
76af4d93 | 439 | * |
24acfb71 | 440 | * This must be called either with wq->mutex held or RCU read locked. |
794b18bc TH |
441 | * If the pwq needs to be used beyond the locking in effect, the caller is |
442 | * responsible for guaranteeing that the pwq stays online. | |
76af4d93 TH |
443 | * |
444 | * The if/else clause exists only for the lockdep assertion and can be | |
445 | * ignored. | |
49e3cf44 TH |
446 | */ |
447 | #define for_each_pwq(pwq, wq) \ | |
49e9d1a9 | 448 | list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \ |
5a644662 | 449 | lockdep_is_held(&(wq->mutex))) |
f3421797 | 450 | |
dc186ad7 TG |
451 | #ifdef CONFIG_DEBUG_OBJECTS_WORK |
452 | ||
f9e62f31 | 453 | static const struct debug_obj_descr work_debug_descr; |
dc186ad7 | 454 | |
99777288 SG |
455 | static void *work_debug_hint(void *addr) |
456 | { | |
457 | return ((struct work_struct *) addr)->func; | |
458 | } | |
459 | ||
b9fdac7f DC |
460 | static bool work_is_static_object(void *addr) |
461 | { | |
462 | struct work_struct *work = addr; | |
463 | ||
464 | return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work)); | |
465 | } | |
466 | ||
dc186ad7 TG |
467 | /* |
468 | * fixup_init is called when: | |
469 | * - an active object is initialized | |
470 | */ | |
02a982a6 | 471 | static bool work_fixup_init(void *addr, enum debug_obj_state state) |
dc186ad7 TG |
472 | { |
473 | struct work_struct *work = addr; | |
474 | ||
475 | switch (state) { | |
476 | case ODEBUG_STATE_ACTIVE: | |
477 | cancel_work_sync(work); | |
478 | debug_object_init(work, &work_debug_descr); | |
02a982a6 | 479 | return true; |
dc186ad7 | 480 | default: |
02a982a6 | 481 | return false; |
dc186ad7 TG |
482 | } |
483 | } | |
484 | ||
dc186ad7 TG |
485 | /* |
486 | * fixup_free is called when: | |
487 | * - an active object is freed | |
488 | */ | |
02a982a6 | 489 | static bool work_fixup_free(void *addr, enum debug_obj_state state) |
dc186ad7 TG |
490 | { |
491 | struct work_struct *work = addr; | |
492 | ||
493 | switch (state) { | |
494 | case ODEBUG_STATE_ACTIVE: | |
495 | cancel_work_sync(work); | |
496 | debug_object_free(work, &work_debug_descr); | |
02a982a6 | 497 | return true; |
dc186ad7 | 498 | default: |
02a982a6 | 499 | return false; |
dc186ad7 TG |
500 | } |
501 | } | |
502 | ||
f9e62f31 | 503 | static const struct debug_obj_descr work_debug_descr = { |
dc186ad7 | 504 | .name = "work_struct", |
99777288 | 505 | .debug_hint = work_debug_hint, |
b9fdac7f | 506 | .is_static_object = work_is_static_object, |
dc186ad7 | 507 | .fixup_init = work_fixup_init, |
dc186ad7 TG |
508 | .fixup_free = work_fixup_free, |
509 | }; | |
510 | ||
511 | static inline void debug_work_activate(struct work_struct *work) | |
512 | { | |
513 | debug_object_activate(work, &work_debug_descr); | |
514 | } | |
515 | ||
516 | static inline void debug_work_deactivate(struct work_struct *work) | |
517 | { | |
518 | debug_object_deactivate(work, &work_debug_descr); | |
519 | } | |
520 | ||
521 | void __init_work(struct work_struct *work, int onstack) | |
522 | { | |
523 | if (onstack) | |
524 | debug_object_init_on_stack(work, &work_debug_descr); | |
525 | else | |
526 | debug_object_init(work, &work_debug_descr); | |
527 | } | |
528 | EXPORT_SYMBOL_GPL(__init_work); | |
529 | ||
530 | void destroy_work_on_stack(struct work_struct *work) | |
531 | { | |
532 | debug_object_free(work, &work_debug_descr); | |
533 | } | |
534 | EXPORT_SYMBOL_GPL(destroy_work_on_stack); | |
535 | ||
ea2e64f2 TG |
536 | void destroy_delayed_work_on_stack(struct delayed_work *work) |
537 | { | |
538 | destroy_timer_on_stack(&work->timer); | |
539 | debug_object_free(&work->work, &work_debug_descr); | |
540 | } | |
541 | EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack); | |
542 | ||
dc186ad7 TG |
543 | #else |
544 | static inline void debug_work_activate(struct work_struct *work) { } | |
545 | static inline void debug_work_deactivate(struct work_struct *work) { } | |
546 | #endif | |
547 | ||
4e8b22bd | 548 | /** |
67dc8325 | 549 | * worker_pool_assign_id - allocate ID and assign it to @pool |
4e8b22bd LB |
550 | * @pool: the pool pointer of interest |
551 | * | |
552 | * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned | |
553 | * successfully, -errno on failure. | |
554 | */ | |
9daf9e67 TH |
555 | static int worker_pool_assign_id(struct worker_pool *pool) |
556 | { | |
557 | int ret; | |
558 | ||
68e13a67 | 559 | lockdep_assert_held(&wq_pool_mutex); |
5bcab335 | 560 | |
4e8b22bd LB |
561 | ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE, |
562 | GFP_KERNEL); | |
229641a6 | 563 | if (ret >= 0) { |
e68035fb | 564 | pool->id = ret; |
229641a6 TH |
565 | return 0; |
566 | } | |
fa1b54e6 | 567 | return ret; |
7c3eed5c TH |
568 | } |
569 | ||
df2d5ae4 TH |
570 | /** |
571 | * unbound_pwq_by_node - return the unbound pool_workqueue for the given node | |
572 | * @wq: the target workqueue | |
573 | * @node: the node ID | |
574 | * | |
24acfb71 | 575 | * This must be called with any of wq_pool_mutex, wq->mutex or RCU |
5b95e1af | 576 | * read locked. |
df2d5ae4 TH |
577 | * If the pwq needs to be used beyond the locking in effect, the caller is |
578 | * responsible for guaranteeing that the pwq stays online. | |
d185af30 YB |
579 | * |
580 | * Return: The unbound pool_workqueue for @node. | |
df2d5ae4 TH |
581 | */ |
582 | static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq, | |
583 | int node) | |
584 | { | |
5b95e1af | 585 | assert_rcu_or_wq_mutex_or_pool_mutex(wq); |
d6e022f1 TH |
586 | |
587 | /* | |
588 | * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a | |
589 | * delayed item is pending. The plan is to keep CPU -> NODE | |
590 | * mapping valid and stable across CPU on/offlines. Once that | |
591 | * happens, this workaround can be removed. | |
592 | */ | |
593 | if (unlikely(node == NUMA_NO_NODE)) | |
594 | return wq->dfl_pwq; | |
595 | ||
df2d5ae4 TH |
596 | return rcu_dereference_raw(wq->numa_pwq_tbl[node]); |
597 | } | |
598 | ||
73f53c4a TH |
599 | static unsigned int work_color_to_flags(int color) |
600 | { | |
601 | return color << WORK_STRUCT_COLOR_SHIFT; | |
602 | } | |
603 | ||
c4560c2c | 604 | static int get_work_color(unsigned long work_data) |
73f53c4a | 605 | { |
c4560c2c | 606 | return (work_data >> WORK_STRUCT_COLOR_SHIFT) & |
73f53c4a TH |
607 | ((1 << WORK_STRUCT_COLOR_BITS) - 1); |
608 | } | |
609 | ||
610 | static int work_next_color(int color) | |
611 | { | |
612 | return (color + 1) % WORK_NR_COLORS; | |
613 | } | |
1da177e4 | 614 | |
14441960 | 615 | /* |
112202d9 TH |
616 | * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data |
617 | * contain the pointer to the queued pwq. Once execution starts, the flag | |
7c3eed5c | 618 | * is cleared and the high bits contain OFFQ flags and pool ID. |
7a22ad75 | 619 | * |
112202d9 TH |
620 | * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() |
621 | * and clear_work_data() can be used to set the pwq, pool or clear | |
bbb68dfa TH |
622 | * work->data. These functions should only be called while the work is |
623 | * owned - ie. while the PENDING bit is set. | |
7a22ad75 | 624 | * |
112202d9 | 625 | * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq |
7c3eed5c | 626 | * corresponding to a work. Pool is available once the work has been |
112202d9 | 627 | * queued anywhere after initialization until it is sync canceled. pwq is |
7c3eed5c | 628 | * available only while the work item is queued. |
7a22ad75 | 629 | * |
bbb68dfa TH |
630 | * %WORK_OFFQ_CANCELING is used to mark a work item which is being |
631 | * canceled. While being canceled, a work item may have its PENDING set | |
632 | * but stay off timer and worklist for arbitrarily long and nobody should | |
633 | * try to steal the PENDING bit. | |
14441960 | 634 | */ |
7a22ad75 TH |
635 | static inline void set_work_data(struct work_struct *work, unsigned long data, |
636 | unsigned long flags) | |
365970a1 | 637 | { |
6183c009 | 638 | WARN_ON_ONCE(!work_pending(work)); |
7a22ad75 TH |
639 | atomic_long_set(&work->data, data | flags | work_static(work)); |
640 | } | |
365970a1 | 641 | |
112202d9 | 642 | static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, |
7a22ad75 TH |
643 | unsigned long extra_flags) |
644 | { | |
112202d9 TH |
645 | set_work_data(work, (unsigned long)pwq, |
646 | WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); | |
365970a1 DH |
647 | } |
648 | ||
4468a00f LJ |
649 | static void set_work_pool_and_keep_pending(struct work_struct *work, |
650 | int pool_id) | |
651 | { | |
652 | set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, | |
653 | WORK_STRUCT_PENDING); | |
654 | } | |
655 | ||
7c3eed5c TH |
656 | static void set_work_pool_and_clear_pending(struct work_struct *work, |
657 | int pool_id) | |
7a22ad75 | 658 | { |
23657bb1 TH |
659 | /* |
660 | * The following wmb is paired with the implied mb in | |
661 | * test_and_set_bit(PENDING) and ensures all updates to @work made | |
662 | * here are visible to and precede any updates by the next PENDING | |
663 | * owner. | |
664 | */ | |
665 | smp_wmb(); | |
7c3eed5c | 666 | set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); |
346c09f8 RP |
667 | /* |
668 | * The following mb guarantees that previous clear of a PENDING bit | |
669 | * will not be reordered with any speculative LOADS or STORES from | |
670 | * work->current_func, which is executed afterwards. This possible | |
8bdc6201 | 671 | * reordering can lead to a missed execution on attempt to queue |
346c09f8 RP |
672 | * the same @work. E.g. consider this case: |
673 | * | |
674 | * CPU#0 CPU#1 | |
675 | * ---------------------------- -------------------------------- | |
676 | * | |
677 | * 1 STORE event_indicated | |
678 | * 2 queue_work_on() { | |
679 | * 3 test_and_set_bit(PENDING) | |
680 | * 4 } set_..._and_clear_pending() { | |
681 | * 5 set_work_data() # clear bit | |
682 | * 6 smp_mb() | |
683 | * 7 work->current_func() { | |
684 | * 8 LOAD event_indicated | |
685 | * } | |
686 | * | |
687 | * Without an explicit full barrier speculative LOAD on line 8 can | |
688 | * be executed before CPU#0 does STORE on line 1. If that happens, | |
689 | * CPU#0 observes the PENDING bit is still set and new execution of | |
690 | * a @work is not queued in a hope, that CPU#1 will eventually | |
691 | * finish the queued @work. Meanwhile CPU#1 does not see | |
692 | * event_indicated is set, because speculative LOAD was executed | |
693 | * before actual STORE. | |
694 | */ | |
695 | smp_mb(); | |
7a22ad75 | 696 | } |
f756d5e2 | 697 | |
7a22ad75 | 698 | static void clear_work_data(struct work_struct *work) |
1da177e4 | 699 | { |
7c3eed5c TH |
700 | smp_wmb(); /* see set_work_pool_and_clear_pending() */ |
701 | set_work_data(work, WORK_STRUCT_NO_POOL, 0); | |
1da177e4 LT |
702 | } |
703 | ||
112202d9 | 704 | static struct pool_workqueue *get_work_pwq(struct work_struct *work) |
b1f4ec17 | 705 | { |
e120153d | 706 | unsigned long data = atomic_long_read(&work->data); |
7a22ad75 | 707 | |
112202d9 | 708 | if (data & WORK_STRUCT_PWQ) |
e120153d TH |
709 | return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); |
710 | else | |
711 | return NULL; | |
4d707b9f ON |
712 | } |
713 | ||
7c3eed5c TH |
714 | /** |
715 | * get_work_pool - return the worker_pool a given work was associated with | |
716 | * @work: the work item of interest | |
717 | * | |
68e13a67 | 718 | * Pools are created and destroyed under wq_pool_mutex, and allows read |
24acfb71 TG |
719 | * access under RCU read lock. As such, this function should be |
720 | * called under wq_pool_mutex or inside of a rcu_read_lock() region. | |
fa1b54e6 TH |
721 | * |
722 | * All fields of the returned pool are accessible as long as the above | |
723 | * mentioned locking is in effect. If the returned pool needs to be used | |
724 | * beyond the critical section, the caller is responsible for ensuring the | |
725 | * returned pool is and stays online. | |
d185af30 YB |
726 | * |
727 | * Return: The worker_pool @work was last associated with. %NULL if none. | |
7c3eed5c TH |
728 | */ |
729 | static struct worker_pool *get_work_pool(struct work_struct *work) | |
365970a1 | 730 | { |
e120153d | 731 | unsigned long data = atomic_long_read(&work->data); |
7c3eed5c | 732 | int pool_id; |
7a22ad75 | 733 | |
68e13a67 | 734 | assert_rcu_or_pool_mutex(); |
fa1b54e6 | 735 | |
112202d9 TH |
736 | if (data & WORK_STRUCT_PWQ) |
737 | return ((struct pool_workqueue *) | |
7c3eed5c | 738 | (data & WORK_STRUCT_WQ_DATA_MASK))->pool; |
7a22ad75 | 739 | |
7c3eed5c TH |
740 | pool_id = data >> WORK_OFFQ_POOL_SHIFT; |
741 | if (pool_id == WORK_OFFQ_POOL_NONE) | |
7a22ad75 TH |
742 | return NULL; |
743 | ||
fa1b54e6 | 744 | return idr_find(&worker_pool_idr, pool_id); |
7c3eed5c TH |
745 | } |
746 | ||
747 | /** | |
748 | * get_work_pool_id - return the worker pool ID a given work is associated with | |
749 | * @work: the work item of interest | |
750 | * | |
d185af30 | 751 | * Return: The worker_pool ID @work was last associated with. |
7c3eed5c TH |
752 | * %WORK_OFFQ_POOL_NONE if none. |
753 | */ | |
754 | static int get_work_pool_id(struct work_struct *work) | |
755 | { | |
54d5b7d0 LJ |
756 | unsigned long data = atomic_long_read(&work->data); |
757 | ||
112202d9 TH |
758 | if (data & WORK_STRUCT_PWQ) |
759 | return ((struct pool_workqueue *) | |
54d5b7d0 | 760 | (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id; |
7c3eed5c | 761 | |
54d5b7d0 | 762 | return data >> WORK_OFFQ_POOL_SHIFT; |
7c3eed5c TH |
763 | } |
764 | ||
bbb68dfa TH |
765 | static void mark_work_canceling(struct work_struct *work) |
766 | { | |
7c3eed5c | 767 | unsigned long pool_id = get_work_pool_id(work); |
bbb68dfa | 768 | |
7c3eed5c TH |
769 | pool_id <<= WORK_OFFQ_POOL_SHIFT; |
770 | set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); | |
bbb68dfa TH |
771 | } |
772 | ||
773 | static bool work_is_canceling(struct work_struct *work) | |
774 | { | |
775 | unsigned long data = atomic_long_read(&work->data); | |
776 | ||
112202d9 | 777 | return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); |
bbb68dfa TH |
778 | } |
779 | ||
e22bee78 | 780 | /* |
3270476a TH |
781 | * Policy functions. These define the policies on how the global worker |
782 | * pools are managed. Unless noted otherwise, these functions assume that | |
d565ed63 | 783 | * they're being called with pool->lock held. |
e22bee78 TH |
784 | */ |
785 | ||
63d95a91 | 786 | static bool __need_more_worker(struct worker_pool *pool) |
a848e3b6 | 787 | { |
bc35f7ef | 788 | return !pool->nr_running; |
a848e3b6 ON |
789 | } |
790 | ||
4594bf15 | 791 | /* |
e22bee78 TH |
792 | * Need to wake up a worker? Called from anything but currently |
793 | * running workers. | |
974271c4 TH |
794 | * |
795 | * Note that, because unbound workers never contribute to nr_running, this | |
706026c2 | 796 | * function will always return %true for unbound pools as long as the |
974271c4 | 797 | * worklist isn't empty. |
4594bf15 | 798 | */ |
63d95a91 | 799 | static bool need_more_worker(struct worker_pool *pool) |
365970a1 | 800 | { |
63d95a91 | 801 | return !list_empty(&pool->worklist) && __need_more_worker(pool); |
e22bee78 | 802 | } |
4594bf15 | 803 | |
e22bee78 | 804 | /* Can I start working? Called from busy but !running workers. */ |
63d95a91 | 805 | static bool may_start_working(struct worker_pool *pool) |
e22bee78 | 806 | { |
63d95a91 | 807 | return pool->nr_idle; |
e22bee78 TH |
808 | } |
809 | ||
810 | /* Do I need to keep working? Called from currently running workers. */ | |
63d95a91 | 811 | static bool keep_working(struct worker_pool *pool) |
e22bee78 | 812 | { |
bc35f7ef | 813 | return !list_empty(&pool->worklist) && (pool->nr_running <= 1); |
e22bee78 TH |
814 | } |
815 | ||
816 | /* Do we need a new worker? Called from manager. */ | |
63d95a91 | 817 | static bool need_to_create_worker(struct worker_pool *pool) |
e22bee78 | 818 | { |
63d95a91 | 819 | return need_more_worker(pool) && !may_start_working(pool); |
e22bee78 | 820 | } |
365970a1 | 821 | |
e22bee78 | 822 | /* Do we have too many workers and should some go away? */ |
63d95a91 | 823 | static bool too_many_workers(struct worker_pool *pool) |
e22bee78 | 824 | { |
692b4825 | 825 | bool managing = pool->flags & POOL_MANAGER_ACTIVE; |
63d95a91 TH |
826 | int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ |
827 | int nr_busy = pool->nr_workers - nr_idle; | |
e22bee78 TH |
828 | |
829 | return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; | |
365970a1 DH |
830 | } |
831 | ||
4d707b9f | 832 | /* |
e22bee78 TH |
833 | * Wake up functions. |
834 | */ | |
835 | ||
2c1f1a91 | 836 | /* Return the first idle worker. Called with pool->lock held. */ |
1037de36 | 837 | static struct worker *first_idle_worker(struct worker_pool *pool) |
7e11629d | 838 | { |
63d95a91 | 839 | if (unlikely(list_empty(&pool->idle_list))) |
7e11629d TH |
840 | return NULL; |
841 | ||
63d95a91 | 842 | return list_first_entry(&pool->idle_list, struct worker, entry); |
7e11629d TH |
843 | } |
844 | ||
845 | /** | |
846 | * wake_up_worker - wake up an idle worker | |
63d95a91 | 847 | * @pool: worker pool to wake worker from |
7e11629d | 848 | * |
63d95a91 | 849 | * Wake up the first idle worker of @pool. |
7e11629d TH |
850 | * |
851 | * CONTEXT: | |
a9b8a985 | 852 | * raw_spin_lock_irq(pool->lock). |
7e11629d | 853 | */ |
63d95a91 | 854 | static void wake_up_worker(struct worker_pool *pool) |
7e11629d | 855 | { |
1037de36 | 856 | struct worker *worker = first_idle_worker(pool); |
7e11629d TH |
857 | |
858 | if (likely(worker)) | |
859 | wake_up_process(worker->task); | |
860 | } | |
861 | ||
d302f017 | 862 | /** |
6d25be57 | 863 | * wq_worker_running - a worker is running again |
e22bee78 | 864 | * @task: task waking up |
e22bee78 | 865 | * |
6d25be57 | 866 | * This function is called when a worker returns from schedule() |
e22bee78 | 867 | */ |
6d25be57 | 868 | void wq_worker_running(struct task_struct *task) |
e22bee78 TH |
869 | { |
870 | struct worker *worker = kthread_data(task); | |
871 | ||
6d25be57 TG |
872 | if (!worker->sleeping) |
873 | return; | |
07edfece FW |
874 | |
875 | /* | |
876 | * If preempted by unbind_workers() between the WORKER_NOT_RUNNING check | |
877 | * and the nr_running increment below, we may ruin the nr_running reset | |
878 | * and leave with an unexpected pool->nr_running == 1 on the newly unbound | |
879 | * pool. Protect against such race. | |
880 | */ | |
881 | preempt_disable(); | |
6d25be57 | 882 | if (!(worker->flags & WORKER_NOT_RUNNING)) |
bc35f7ef | 883 | worker->pool->nr_running++; |
07edfece | 884 | preempt_enable(); |
6d25be57 | 885 | worker->sleeping = 0; |
e22bee78 TH |
886 | } |
887 | ||
888 | /** | |
889 | * wq_worker_sleeping - a worker is going to sleep | |
890 | * @task: task going to sleep | |
e22bee78 | 891 | * |
6d25be57 | 892 | * This function is called from schedule() when a busy worker is |
ccf45156 | 893 | * going to sleep. |
e22bee78 | 894 | */ |
6d25be57 | 895 | void wq_worker_sleeping(struct task_struct *task) |
e22bee78 | 896 | { |
cc5bff38 | 897 | struct worker *worker = kthread_data(task); |
111c225a | 898 | struct worker_pool *pool; |
e22bee78 | 899 | |
111c225a TH |
900 | /* |
901 | * Rescuers, which may not have all the fields set up like normal | |
902 | * workers, also reach here, let's not access anything before | |
903 | * checking NOT_RUNNING. | |
904 | */ | |
2d64672e | 905 | if (worker->flags & WORKER_NOT_RUNNING) |
6d25be57 | 906 | return; |
e22bee78 | 907 | |
111c225a | 908 | pool = worker->pool; |
111c225a | 909 | |
62849a96 SAS |
910 | /* Return if preempted before wq_worker_running() was reached */ |
911 | if (worker->sleeping) | |
6d25be57 TG |
912 | return; |
913 | ||
914 | worker->sleeping = 1; | |
a9b8a985 | 915 | raw_spin_lock_irq(&pool->lock); |
e22bee78 | 916 | |
45c753f5 FW |
917 | /* |
918 | * Recheck in case unbind_workers() preempted us. We don't | |
919 | * want to decrement nr_running after the worker is unbound | |
920 | * and nr_running has been reset. | |
921 | */ | |
922 | if (worker->flags & WORKER_NOT_RUNNING) { | |
923 | raw_spin_unlock_irq(&pool->lock); | |
924 | return; | |
925 | } | |
926 | ||
bc35f7ef LJ |
927 | pool->nr_running--; |
928 | if (need_more_worker(pool)) | |
cc5bff38 | 929 | wake_up_worker(pool); |
a9b8a985 | 930 | raw_spin_unlock_irq(&pool->lock); |
e22bee78 TH |
931 | } |
932 | ||
1b69ac6b JW |
933 | /** |
934 | * wq_worker_last_func - retrieve worker's last work function | |
8194fe94 | 935 | * @task: Task to retrieve last work function of. |
1b69ac6b JW |
936 | * |
937 | * Determine the last function a worker executed. This is called from | |
938 | * the scheduler to get a worker's last known identity. | |
939 | * | |
940 | * CONTEXT: | |
a9b8a985 | 941 | * raw_spin_lock_irq(rq->lock) |
1b69ac6b | 942 | * |
4b047002 JW |
943 | * This function is called during schedule() when a kworker is going |
944 | * to sleep. It's used by psi to identify aggregation workers during | |
945 | * dequeuing, to allow periodic aggregation to shut-off when that | |
946 | * worker is the last task in the system or cgroup to go to sleep. | |
947 | * | |
948 | * As this function doesn't involve any workqueue-related locking, it | |
949 | * only returns stable values when called from inside the scheduler's | |
950 | * queuing and dequeuing paths, when @task, which must be a kworker, | |
951 | * is guaranteed to not be processing any works. | |
952 | * | |
1b69ac6b JW |
953 | * Return: |
954 | * The last work function %current executed as a worker, NULL if it | |
955 | * hasn't executed any work yet. | |
956 | */ | |
957 | work_func_t wq_worker_last_func(struct task_struct *task) | |
958 | { | |
959 | struct worker *worker = kthread_data(task); | |
960 | ||
961 | return worker->last_func; | |
962 | } | |
963 | ||
e22bee78 TH |
964 | /** |
965 | * worker_set_flags - set worker flags and adjust nr_running accordingly | |
cb444766 | 966 | * @worker: self |
d302f017 | 967 | * @flags: flags to set |
d302f017 | 968 | * |
228f1d00 | 969 | * Set @flags in @worker->flags and adjust nr_running accordingly. |
d302f017 | 970 | * |
cb444766 | 971 | * CONTEXT: |
a9b8a985 | 972 | * raw_spin_lock_irq(pool->lock) |
d302f017 | 973 | */ |
228f1d00 | 974 | static inline void worker_set_flags(struct worker *worker, unsigned int flags) |
d302f017 | 975 | { |
bd7bdd43 | 976 | struct worker_pool *pool = worker->pool; |
e22bee78 | 977 | |
cb444766 TH |
978 | WARN_ON_ONCE(worker->task != current); |
979 | ||
228f1d00 | 980 | /* If transitioning into NOT_RUNNING, adjust nr_running. */ |
e22bee78 TH |
981 | if ((flags & WORKER_NOT_RUNNING) && |
982 | !(worker->flags & WORKER_NOT_RUNNING)) { | |
bc35f7ef | 983 | pool->nr_running--; |
e22bee78 TH |
984 | } |
985 | ||
d302f017 TH |
986 | worker->flags |= flags; |
987 | } | |
988 | ||
989 | /** | |
e22bee78 | 990 | * worker_clr_flags - clear worker flags and adjust nr_running accordingly |
cb444766 | 991 | * @worker: self |
d302f017 TH |
992 | * @flags: flags to clear |
993 | * | |
e22bee78 | 994 | * Clear @flags in @worker->flags and adjust nr_running accordingly. |
d302f017 | 995 | * |
cb444766 | 996 | * CONTEXT: |
a9b8a985 | 997 | * raw_spin_lock_irq(pool->lock) |
d302f017 TH |
998 | */ |
999 | static inline void worker_clr_flags(struct worker *worker, unsigned int flags) | |
1000 | { | |
63d95a91 | 1001 | struct worker_pool *pool = worker->pool; |
e22bee78 TH |
1002 | unsigned int oflags = worker->flags; |
1003 | ||
cb444766 TH |
1004 | WARN_ON_ONCE(worker->task != current); |
1005 | ||
d302f017 | 1006 | worker->flags &= ~flags; |
e22bee78 | 1007 | |
42c025f3 TH |
1008 | /* |
1009 | * If transitioning out of NOT_RUNNING, increment nr_running. Note | |
1010 | * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask | |
1011 | * of multiple flags, not a single flag. | |
1012 | */ | |
e22bee78 TH |
1013 | if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) |
1014 | if (!(worker->flags & WORKER_NOT_RUNNING)) | |
bc35f7ef | 1015 | pool->nr_running++; |
d302f017 TH |
1016 | } |
1017 | ||
8cca0eea TH |
1018 | /** |
1019 | * find_worker_executing_work - find worker which is executing a work | |
c9e7cf27 | 1020 | * @pool: pool of interest |
8cca0eea TH |
1021 | * @work: work to find worker for |
1022 | * | |
c9e7cf27 TH |
1023 | * Find a worker which is executing @work on @pool by searching |
1024 | * @pool->busy_hash which is keyed by the address of @work. For a worker | |
a2c1c57b TH |
1025 | * to match, its current execution should match the address of @work and |
1026 | * its work function. This is to avoid unwanted dependency between | |
1027 | * unrelated work executions through a work item being recycled while still | |
1028 | * being executed. | |
1029 | * | |
1030 | * This is a bit tricky. A work item may be freed once its execution | |
1031 | * starts and nothing prevents the freed area from being recycled for | |
1032 | * another work item. If the same work item address ends up being reused | |
1033 | * before the original execution finishes, workqueue will identify the | |
1034 | * recycled work item as currently executing and make it wait until the | |
1035 | * current execution finishes, introducing an unwanted dependency. | |
1036 | * | |
c5aa87bb TH |
1037 | * This function checks the work item address and work function to avoid |
1038 | * false positives. Note that this isn't complete as one may construct a | |
1039 | * work function which can introduce dependency onto itself through a | |
1040 | * recycled work item. Well, if somebody wants to shoot oneself in the | |
1041 | * foot that badly, there's only so much we can do, and if such deadlock | |
1042 | * actually occurs, it should be easy to locate the culprit work function. | |
8cca0eea TH |
1043 | * |
1044 | * CONTEXT: | |
a9b8a985 | 1045 | * raw_spin_lock_irq(pool->lock). |
8cca0eea | 1046 | * |
d185af30 YB |
1047 | * Return: |
1048 | * Pointer to worker which is executing @work if found, %NULL | |
8cca0eea | 1049 | * otherwise. |
4d707b9f | 1050 | */ |
c9e7cf27 | 1051 | static struct worker *find_worker_executing_work(struct worker_pool *pool, |
8cca0eea | 1052 | struct work_struct *work) |
4d707b9f | 1053 | { |
42f8570f | 1054 | struct worker *worker; |
42f8570f | 1055 | |
b67bfe0d | 1056 | hash_for_each_possible(pool->busy_hash, worker, hentry, |
a2c1c57b TH |
1057 | (unsigned long)work) |
1058 | if (worker->current_work == work && | |
1059 | worker->current_func == work->func) | |
42f8570f SL |
1060 | return worker; |
1061 | ||
1062 | return NULL; | |
4d707b9f ON |
1063 | } |
1064 | ||
bf4ede01 TH |
1065 | /** |
1066 | * move_linked_works - move linked works to a list | |
1067 | * @work: start of series of works to be scheduled | |
1068 | * @head: target list to append @work to | |
402dd89d | 1069 | * @nextp: out parameter for nested worklist walking |
bf4ede01 TH |
1070 | * |
1071 | * Schedule linked works starting from @work to @head. Work series to | |
1072 | * be scheduled starts at @work and includes any consecutive work with | |
1073 | * WORK_STRUCT_LINKED set in its predecessor. | |
1074 | * | |
1075 | * If @nextp is not NULL, it's updated to point to the next work of | |
1076 | * the last scheduled work. This allows move_linked_works() to be | |
1077 | * nested inside outer list_for_each_entry_safe(). | |
1078 | * | |
1079 | * CONTEXT: | |
a9b8a985 | 1080 | * raw_spin_lock_irq(pool->lock). |
bf4ede01 TH |
1081 | */ |
1082 | static void move_linked_works(struct work_struct *work, struct list_head *head, | |
1083 | struct work_struct **nextp) | |
1084 | { | |
1085 | struct work_struct *n; | |
1086 | ||
1087 | /* | |
1088 | * Linked worklist will always end before the end of the list, | |
1089 | * use NULL for list head. | |
1090 | */ | |
1091 | list_for_each_entry_safe_from(work, n, NULL, entry) { | |
1092 | list_move_tail(&work->entry, head); | |
1093 | if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) | |
1094 | break; | |
1095 | } | |
1096 | ||
1097 | /* | |
1098 | * If we're already inside safe list traversal and have moved | |
1099 | * multiple works to the scheduled queue, the next position | |
1100 | * needs to be updated. | |
1101 | */ | |
1102 | if (nextp) | |
1103 | *nextp = n; | |
1104 | } | |
1105 | ||
8864b4e5 TH |
1106 | /** |
1107 | * get_pwq - get an extra reference on the specified pool_workqueue | |
1108 | * @pwq: pool_workqueue to get | |
1109 | * | |
1110 | * Obtain an extra reference on @pwq. The caller should guarantee that | |
1111 | * @pwq has positive refcnt and be holding the matching pool->lock. | |
1112 | */ | |
1113 | static void get_pwq(struct pool_workqueue *pwq) | |
1114 | { | |
1115 | lockdep_assert_held(&pwq->pool->lock); | |
1116 | WARN_ON_ONCE(pwq->refcnt <= 0); | |
1117 | pwq->refcnt++; | |
1118 | } | |
1119 | ||
1120 | /** | |
1121 | * put_pwq - put a pool_workqueue reference | |
1122 | * @pwq: pool_workqueue to put | |
1123 | * | |
1124 | * Drop a reference of @pwq. If its refcnt reaches zero, schedule its | |
1125 | * destruction. The caller should be holding the matching pool->lock. | |
1126 | */ | |
1127 | static void put_pwq(struct pool_workqueue *pwq) | |
1128 | { | |
1129 | lockdep_assert_held(&pwq->pool->lock); | |
1130 | if (likely(--pwq->refcnt)) | |
1131 | return; | |
1132 | if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND))) | |
1133 | return; | |
1134 | /* | |
1135 | * @pwq can't be released under pool->lock, bounce to | |
1136 | * pwq_unbound_release_workfn(). This never recurses on the same | |
1137 | * pool->lock as this path is taken only for unbound workqueues and | |
1138 | * the release work item is scheduled on a per-cpu workqueue. To | |
1139 | * avoid lockdep warning, unbound pool->locks are given lockdep | |
1140 | * subclass of 1 in get_unbound_pool(). | |
1141 | */ | |
1142 | schedule_work(&pwq->unbound_release_work); | |
1143 | } | |
1144 | ||
dce90d47 TH |
1145 | /** |
1146 | * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock | |
1147 | * @pwq: pool_workqueue to put (can be %NULL) | |
1148 | * | |
1149 | * put_pwq() with locking. This function also allows %NULL @pwq. | |
1150 | */ | |
1151 | static void put_pwq_unlocked(struct pool_workqueue *pwq) | |
1152 | { | |
1153 | if (pwq) { | |
1154 | /* | |
24acfb71 | 1155 | * As both pwqs and pools are RCU protected, the |
dce90d47 TH |
1156 | * following lock operations are safe. |
1157 | */ | |
a9b8a985 | 1158 | raw_spin_lock_irq(&pwq->pool->lock); |
dce90d47 | 1159 | put_pwq(pwq); |
a9b8a985 | 1160 | raw_spin_unlock_irq(&pwq->pool->lock); |
dce90d47 TH |
1161 | } |
1162 | } | |
1163 | ||
f97a4a1a | 1164 | static void pwq_activate_inactive_work(struct work_struct *work) |
bf4ede01 | 1165 | { |
112202d9 | 1166 | struct pool_workqueue *pwq = get_work_pwq(work); |
bf4ede01 TH |
1167 | |
1168 | trace_workqueue_activate_work(work); | |
82607adc TH |
1169 | if (list_empty(&pwq->pool->worklist)) |
1170 | pwq->pool->watchdog_ts = jiffies; | |
112202d9 | 1171 | move_linked_works(work, &pwq->pool->worklist, NULL); |
f97a4a1a | 1172 | __clear_bit(WORK_STRUCT_INACTIVE_BIT, work_data_bits(work)); |
112202d9 | 1173 | pwq->nr_active++; |
bf4ede01 TH |
1174 | } |
1175 | ||
f97a4a1a | 1176 | static void pwq_activate_first_inactive(struct pool_workqueue *pwq) |
3aa62497 | 1177 | { |
f97a4a1a | 1178 | struct work_struct *work = list_first_entry(&pwq->inactive_works, |
3aa62497 LJ |
1179 | struct work_struct, entry); |
1180 | ||
f97a4a1a | 1181 | pwq_activate_inactive_work(work); |
3aa62497 LJ |
1182 | } |
1183 | ||
bf4ede01 | 1184 | /** |
112202d9 TH |
1185 | * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight |
1186 | * @pwq: pwq of interest | |
c4560c2c | 1187 | * @work_data: work_data of work which left the queue |
bf4ede01 TH |
1188 | * |
1189 | * A work either has completed or is removed from pending queue, | |
112202d9 | 1190 | * decrement nr_in_flight of its pwq and handle workqueue flushing. |
bf4ede01 TH |
1191 | * |
1192 | * CONTEXT: | |
a9b8a985 | 1193 | * raw_spin_lock_irq(pool->lock). |
bf4ede01 | 1194 | */ |
c4560c2c | 1195 | static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, unsigned long work_data) |
bf4ede01 | 1196 | { |
c4560c2c LJ |
1197 | int color = get_work_color(work_data); |
1198 | ||
018f3a13 LJ |
1199 | if (!(work_data & WORK_STRUCT_INACTIVE)) { |
1200 | pwq->nr_active--; | |
1201 | if (!list_empty(&pwq->inactive_works)) { | |
1202 | /* one down, submit an inactive one */ | |
1203 | if (pwq->nr_active < pwq->max_active) | |
1204 | pwq_activate_first_inactive(pwq); | |
1205 | } | |
1206 | } | |
1207 | ||
112202d9 | 1208 | pwq->nr_in_flight[color]--; |
bf4ede01 | 1209 | |
bf4ede01 | 1210 | /* is flush in progress and are we at the flushing tip? */ |
112202d9 | 1211 | if (likely(pwq->flush_color != color)) |
8864b4e5 | 1212 | goto out_put; |
bf4ede01 TH |
1213 | |
1214 | /* are there still in-flight works? */ | |
112202d9 | 1215 | if (pwq->nr_in_flight[color]) |
8864b4e5 | 1216 | goto out_put; |
bf4ede01 | 1217 | |
112202d9 TH |
1218 | /* this pwq is done, clear flush_color */ |
1219 | pwq->flush_color = -1; | |
bf4ede01 TH |
1220 | |
1221 | /* | |
112202d9 | 1222 | * If this was the last pwq, wake up the first flusher. It |
bf4ede01 TH |
1223 | * will handle the rest. |
1224 | */ | |
112202d9 TH |
1225 | if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) |
1226 | complete(&pwq->wq->first_flusher->done); | |
8864b4e5 TH |
1227 | out_put: |
1228 | put_pwq(pwq); | |
bf4ede01 TH |
1229 | } |
1230 | ||
36e227d2 | 1231 | /** |
bbb68dfa | 1232 | * try_to_grab_pending - steal work item from worklist and disable irq |
36e227d2 TH |
1233 | * @work: work item to steal |
1234 | * @is_dwork: @work is a delayed_work | |
bbb68dfa | 1235 | * @flags: place to store irq state |
36e227d2 TH |
1236 | * |
1237 | * Try to grab PENDING bit of @work. This function can handle @work in any | |
d185af30 | 1238 | * stable state - idle, on timer or on worklist. |
36e227d2 | 1239 | * |
d185af30 | 1240 | * Return: |
3eb6b31b MCC |
1241 | * |
1242 | * ======== ================================================================ | |
36e227d2 TH |
1243 | * 1 if @work was pending and we successfully stole PENDING |
1244 | * 0 if @work was idle and we claimed PENDING | |
1245 | * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry | |
bbb68dfa TH |
1246 | * -ENOENT if someone else is canceling @work, this state may persist |
1247 | * for arbitrarily long | |
3eb6b31b | 1248 | * ======== ================================================================ |
36e227d2 | 1249 | * |
d185af30 | 1250 | * Note: |
bbb68dfa | 1251 | * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting |
e0aecdd8 TH |
1252 | * interrupted while holding PENDING and @work off queue, irq must be |
1253 | * disabled on entry. This, combined with delayed_work->timer being | |
1254 | * irqsafe, ensures that we return -EAGAIN for finite short period of time. | |
bbb68dfa TH |
1255 | * |
1256 | * On successful return, >= 0, irq is disabled and the caller is | |
1257 | * responsible for releasing it using local_irq_restore(*@flags). | |
1258 | * | |
e0aecdd8 | 1259 | * This function is safe to call from any context including IRQ handler. |
bf4ede01 | 1260 | */ |
bbb68dfa TH |
1261 | static int try_to_grab_pending(struct work_struct *work, bool is_dwork, |
1262 | unsigned long *flags) | |
bf4ede01 | 1263 | { |
d565ed63 | 1264 | struct worker_pool *pool; |
112202d9 | 1265 | struct pool_workqueue *pwq; |
bf4ede01 | 1266 | |
bbb68dfa TH |
1267 | local_irq_save(*flags); |
1268 | ||
36e227d2 TH |
1269 | /* try to steal the timer if it exists */ |
1270 | if (is_dwork) { | |
1271 | struct delayed_work *dwork = to_delayed_work(work); | |
1272 | ||
e0aecdd8 TH |
1273 | /* |
1274 | * dwork->timer is irqsafe. If del_timer() fails, it's | |
1275 | * guaranteed that the timer is not queued anywhere and not | |
1276 | * running on the local CPU. | |
1277 | */ | |
36e227d2 TH |
1278 | if (likely(del_timer(&dwork->timer))) |
1279 | return 1; | |
1280 | } | |
1281 | ||
1282 | /* try to claim PENDING the normal way */ | |
bf4ede01 TH |
1283 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) |
1284 | return 0; | |
1285 | ||
24acfb71 | 1286 | rcu_read_lock(); |
bf4ede01 TH |
1287 | /* |
1288 | * The queueing is in progress, or it is already queued. Try to | |
1289 | * steal it from ->worklist without clearing WORK_STRUCT_PENDING. | |
1290 | */ | |
d565ed63 TH |
1291 | pool = get_work_pool(work); |
1292 | if (!pool) | |
bbb68dfa | 1293 | goto fail; |
bf4ede01 | 1294 | |
a9b8a985 | 1295 | raw_spin_lock(&pool->lock); |
0b3dae68 | 1296 | /* |
112202d9 TH |
1297 | * work->data is guaranteed to point to pwq only while the work |
1298 | * item is queued on pwq->wq, and both updating work->data to point | |
1299 | * to pwq on queueing and to pool on dequeueing are done under | |
1300 | * pwq->pool->lock. This in turn guarantees that, if work->data | |
1301 | * points to pwq which is associated with a locked pool, the work | |
0b3dae68 LJ |
1302 | * item is currently queued on that pool. |
1303 | */ | |
112202d9 TH |
1304 | pwq = get_work_pwq(work); |
1305 | if (pwq && pwq->pool == pool) { | |
16062836 TH |
1306 | debug_work_deactivate(work); |
1307 | ||
1308 | /* | |
018f3a13 LJ |
1309 | * A cancelable inactive work item must be in the |
1310 | * pwq->inactive_works since a queued barrier can't be | |
1311 | * canceled (see the comments in insert_wq_barrier()). | |
1312 | * | |
f97a4a1a | 1313 | * An inactive work item cannot be grabbed directly because |
d812796e | 1314 | * it might have linked barrier work items which, if left |
f97a4a1a | 1315 | * on the inactive_works list, will confuse pwq->nr_active |
16062836 TH |
1316 | * management later on and cause stall. Make sure the work |
1317 | * item is activated before grabbing. | |
1318 | */ | |
f97a4a1a LJ |
1319 | if (*work_data_bits(work) & WORK_STRUCT_INACTIVE) |
1320 | pwq_activate_inactive_work(work); | |
16062836 TH |
1321 | |
1322 | list_del_init(&work->entry); | |
c4560c2c | 1323 | pwq_dec_nr_in_flight(pwq, *work_data_bits(work)); |
16062836 | 1324 | |
112202d9 | 1325 | /* work->data points to pwq iff queued, point to pool */ |
16062836 TH |
1326 | set_work_pool_and_keep_pending(work, pool->id); |
1327 | ||
a9b8a985 | 1328 | raw_spin_unlock(&pool->lock); |
24acfb71 | 1329 | rcu_read_unlock(); |
16062836 | 1330 | return 1; |
bf4ede01 | 1331 | } |
a9b8a985 | 1332 | raw_spin_unlock(&pool->lock); |
bbb68dfa | 1333 | fail: |
24acfb71 | 1334 | rcu_read_unlock(); |
bbb68dfa TH |
1335 | local_irq_restore(*flags); |
1336 | if (work_is_canceling(work)) | |
1337 | return -ENOENT; | |
1338 | cpu_relax(); | |
36e227d2 | 1339 | return -EAGAIN; |
bf4ede01 TH |
1340 | } |
1341 | ||
4690c4ab | 1342 | /** |
706026c2 | 1343 | * insert_work - insert a work into a pool |
112202d9 | 1344 | * @pwq: pwq @work belongs to |
4690c4ab TH |
1345 | * @work: work to insert |
1346 | * @head: insertion point | |
1347 | * @extra_flags: extra WORK_STRUCT_* flags to set | |
1348 | * | |
112202d9 | 1349 | * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to |
706026c2 | 1350 | * work_struct flags. |
4690c4ab TH |
1351 | * |
1352 | * CONTEXT: | |
a9b8a985 | 1353 | * raw_spin_lock_irq(pool->lock). |
4690c4ab | 1354 | */ |
112202d9 TH |
1355 | static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, |
1356 | struct list_head *head, unsigned int extra_flags) | |
b89deed3 | 1357 | { |
112202d9 | 1358 | struct worker_pool *pool = pwq->pool; |
e22bee78 | 1359 | |
e89a85d6 | 1360 | /* record the work call stack in order to print it in KASAN reports */ |
f70da745 | 1361 | kasan_record_aux_stack_noalloc(work); |
e89a85d6 | 1362 | |
4690c4ab | 1363 | /* we own @work, set data and link */ |
112202d9 | 1364 | set_work_pwq(work, pwq, extra_flags); |
1a4d9b0a | 1365 | list_add_tail(&work->entry, head); |
8864b4e5 | 1366 | get_pwq(pwq); |
e22bee78 | 1367 | |
63d95a91 TH |
1368 | if (__need_more_worker(pool)) |
1369 | wake_up_worker(pool); | |
b89deed3 ON |
1370 | } |
1371 | ||
c8efcc25 TH |
1372 | /* |
1373 | * Test whether @work is being queued from another work executing on the | |
8d03ecfe | 1374 | * same workqueue. |
c8efcc25 TH |
1375 | */ |
1376 | static bool is_chained_work(struct workqueue_struct *wq) | |
1377 | { | |
8d03ecfe TH |
1378 | struct worker *worker; |
1379 | ||
1380 | worker = current_wq_worker(); | |
1381 | /* | |
bf393fd4 | 1382 | * Return %true iff I'm a worker executing a work item on @wq. If |
8d03ecfe TH |
1383 | * I'm @worker, it's safe to dereference it without locking. |
1384 | */ | |
112202d9 | 1385 | return worker && worker->current_pwq->wq == wq; |
c8efcc25 TH |
1386 | } |
1387 | ||
ef557180 MG |
1388 | /* |
1389 | * When queueing an unbound work item to a wq, prefer local CPU if allowed | |
1390 | * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to | |
1391 | * avoid perturbing sensitive tasks. | |
1392 | */ | |
1393 | static int wq_select_unbound_cpu(int cpu) | |
1394 | { | |
f303fccb | 1395 | static bool printed_dbg_warning; |
ef557180 MG |
1396 | int new_cpu; |
1397 | ||
f303fccb TH |
1398 | if (likely(!wq_debug_force_rr_cpu)) { |
1399 | if (cpumask_test_cpu(cpu, wq_unbound_cpumask)) | |
1400 | return cpu; | |
1401 | } else if (!printed_dbg_warning) { | |
1402 | pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n"); | |
1403 | printed_dbg_warning = true; | |
1404 | } | |
1405 | ||
ef557180 MG |
1406 | if (cpumask_empty(wq_unbound_cpumask)) |
1407 | return cpu; | |
1408 | ||
1409 | new_cpu = __this_cpu_read(wq_rr_cpu_last); | |
1410 | new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask); | |
1411 | if (unlikely(new_cpu >= nr_cpu_ids)) { | |
1412 | new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask); | |
1413 | if (unlikely(new_cpu >= nr_cpu_ids)) | |
1414 | return cpu; | |
1415 | } | |
1416 | __this_cpu_write(wq_rr_cpu_last, new_cpu); | |
1417 | ||
1418 | return new_cpu; | |
1419 | } | |
1420 | ||
d84ff051 | 1421 | static void __queue_work(int cpu, struct workqueue_struct *wq, |
1da177e4 LT |
1422 | struct work_struct *work) |
1423 | { | |
112202d9 | 1424 | struct pool_workqueue *pwq; |
c9178087 | 1425 | struct worker_pool *last_pool; |
1e19ffc6 | 1426 | struct list_head *worklist; |
8a2e8e5d | 1427 | unsigned int work_flags; |
b75cac93 | 1428 | unsigned int req_cpu = cpu; |
8930caba TH |
1429 | |
1430 | /* | |
1431 | * While a work item is PENDING && off queue, a task trying to | |
1432 | * steal the PENDING will busy-loop waiting for it to either get | |
1433 | * queued or lose PENDING. Grabbing PENDING and queueing should | |
1434 | * happen with IRQ disabled. | |
1435 | */ | |
8e8eb730 | 1436 | lockdep_assert_irqs_disabled(); |
1da177e4 | 1437 | |
1e19ffc6 | 1438 | |
33e3f0a3 RC |
1439 | /* |
1440 | * For a draining wq, only works from the same workqueue are | |
1441 | * allowed. The __WQ_DESTROYING helps to spot the issue that | |
1442 | * queues a new work item to a wq after destroy_workqueue(wq). | |
1443 | */ | |
1444 | if (unlikely(wq->flags & (__WQ_DESTROYING | __WQ_DRAINING) && | |
1445 | WARN_ON_ONCE(!is_chained_work(wq)))) | |
e41e704b | 1446 | return; |
24acfb71 | 1447 | rcu_read_lock(); |
9e8cd2f5 | 1448 | retry: |
c9178087 | 1449 | /* pwq which will be used unless @work is executing elsewhere */ |
aa202f1f HD |
1450 | if (wq->flags & WQ_UNBOUND) { |
1451 | if (req_cpu == WORK_CPU_UNBOUND) | |
1452 | cpu = wq_select_unbound_cpu(raw_smp_processor_id()); | |
df2d5ae4 | 1453 | pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); |
aa202f1f HD |
1454 | } else { |
1455 | if (req_cpu == WORK_CPU_UNBOUND) | |
1456 | cpu = raw_smp_processor_id(); | |
1457 | pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); | |
1458 | } | |
dbf2576e | 1459 | |
c9178087 TH |
1460 | /* |
1461 | * If @work was previously on a different pool, it might still be | |
1462 | * running there, in which case the work needs to be queued on that | |
1463 | * pool to guarantee non-reentrancy. | |
1464 | */ | |
1465 | last_pool = get_work_pool(work); | |
1466 | if (last_pool && last_pool != pwq->pool) { | |
1467 | struct worker *worker; | |
18aa9eff | 1468 | |
a9b8a985 | 1469 | raw_spin_lock(&last_pool->lock); |
18aa9eff | 1470 | |
c9178087 | 1471 | worker = find_worker_executing_work(last_pool, work); |
18aa9eff | 1472 | |
c9178087 TH |
1473 | if (worker && worker->current_pwq->wq == wq) { |
1474 | pwq = worker->current_pwq; | |
8930caba | 1475 | } else { |
c9178087 | 1476 | /* meh... not running there, queue here */ |
a9b8a985 SAS |
1477 | raw_spin_unlock(&last_pool->lock); |
1478 | raw_spin_lock(&pwq->pool->lock); | |
8930caba | 1479 | } |
f3421797 | 1480 | } else { |
a9b8a985 | 1481 | raw_spin_lock(&pwq->pool->lock); |
502ca9d8 TH |
1482 | } |
1483 | ||
9e8cd2f5 TH |
1484 | /* |
1485 | * pwq is determined and locked. For unbound pools, we could have | |
1486 | * raced with pwq release and it could already be dead. If its | |
1487 | * refcnt is zero, repeat pwq selection. Note that pwqs never die | |
df2d5ae4 TH |
1488 | * without another pwq replacing it in the numa_pwq_tbl or while |
1489 | * work items are executing on it, so the retrying is guaranteed to | |
9e8cd2f5 TH |
1490 | * make forward-progress. |
1491 | */ | |
1492 | if (unlikely(!pwq->refcnt)) { | |
1493 | if (wq->flags & WQ_UNBOUND) { | |
a9b8a985 | 1494 | raw_spin_unlock(&pwq->pool->lock); |
9e8cd2f5 TH |
1495 | cpu_relax(); |
1496 | goto retry; | |
1497 | } | |
1498 | /* oops */ | |
1499 | WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", | |
1500 | wq->name, cpu); | |
1501 | } | |
1502 | ||
112202d9 TH |
1503 | /* pwq determined, queue */ |
1504 | trace_workqueue_queue_work(req_cpu, pwq, work); | |
502ca9d8 | 1505 | |
24acfb71 TG |
1506 | if (WARN_ON(!list_empty(&work->entry))) |
1507 | goto out; | |
1e19ffc6 | 1508 | |
112202d9 TH |
1509 | pwq->nr_in_flight[pwq->work_color]++; |
1510 | work_flags = work_color_to_flags(pwq->work_color); | |
1e19ffc6 | 1511 | |
112202d9 | 1512 | if (likely(pwq->nr_active < pwq->max_active)) { |
cdadf009 | 1513 | trace_workqueue_activate_work(work); |
112202d9 TH |
1514 | pwq->nr_active++; |
1515 | worklist = &pwq->pool->worklist; | |
82607adc TH |
1516 | if (list_empty(worklist)) |
1517 | pwq->pool->watchdog_ts = jiffies; | |
8a2e8e5d | 1518 | } else { |
f97a4a1a LJ |
1519 | work_flags |= WORK_STRUCT_INACTIVE; |
1520 | worklist = &pwq->inactive_works; | |
8a2e8e5d | 1521 | } |
1e19ffc6 | 1522 | |
0687c66b | 1523 | debug_work_activate(work); |
112202d9 | 1524 | insert_work(pwq, work, worklist, work_flags); |
1e19ffc6 | 1525 | |
24acfb71 | 1526 | out: |
a9b8a985 | 1527 | raw_spin_unlock(&pwq->pool->lock); |
24acfb71 | 1528 | rcu_read_unlock(); |
1da177e4 LT |
1529 | } |
1530 | ||
0fcb78c2 | 1531 | /** |
c1a220e7 ZR |
1532 | * queue_work_on - queue work on specific cpu |
1533 | * @cpu: CPU number to execute work on | |
0fcb78c2 REB |
1534 | * @wq: workqueue to use |
1535 | * @work: work to queue | |
1536 | * | |
c1a220e7 | 1537 | * We queue the work to a specific CPU, the caller must ensure it |
443378f0 PM |
1538 | * can't go away. Callers that fail to ensure that the specified |
1539 | * CPU cannot go away will execute on a randomly chosen CPU. | |
d185af30 YB |
1540 | * |
1541 | * Return: %false if @work was already on a queue, %true otherwise. | |
1da177e4 | 1542 | */ |
d4283e93 TH |
1543 | bool queue_work_on(int cpu, struct workqueue_struct *wq, |
1544 | struct work_struct *work) | |
1da177e4 | 1545 | { |
d4283e93 | 1546 | bool ret = false; |
8930caba | 1547 | unsigned long flags; |
ef1ca236 | 1548 | |
8930caba | 1549 | local_irq_save(flags); |
c1a220e7 | 1550 | |
22df02bb | 1551 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { |
4690c4ab | 1552 | __queue_work(cpu, wq, work); |
d4283e93 | 1553 | ret = true; |
c1a220e7 | 1554 | } |
ef1ca236 | 1555 | |
8930caba | 1556 | local_irq_restore(flags); |
1da177e4 LT |
1557 | return ret; |
1558 | } | |
ad7b1f84 | 1559 | EXPORT_SYMBOL(queue_work_on); |
1da177e4 | 1560 | |
8204e0c1 AD |
1561 | /** |
1562 | * workqueue_select_cpu_near - Select a CPU based on NUMA node | |
1563 | * @node: NUMA node ID that we want to select a CPU from | |
1564 | * | |
1565 | * This function will attempt to find a "random" cpu available on a given | |
1566 | * node. If there are no CPUs available on the given node it will return | |
1567 | * WORK_CPU_UNBOUND indicating that we should just schedule to any | |
1568 | * available CPU if we need to schedule this work. | |
1569 | */ | |
1570 | static int workqueue_select_cpu_near(int node) | |
1571 | { | |
1572 | int cpu; | |
1573 | ||
1574 | /* No point in doing this if NUMA isn't enabled for workqueues */ | |
1575 | if (!wq_numa_enabled) | |
1576 | return WORK_CPU_UNBOUND; | |
1577 | ||
1578 | /* Delay binding to CPU if node is not valid or online */ | |
1579 | if (node < 0 || node >= MAX_NUMNODES || !node_online(node)) | |
1580 | return WORK_CPU_UNBOUND; | |
1581 | ||
1582 | /* Use local node/cpu if we are already there */ | |
1583 | cpu = raw_smp_processor_id(); | |
1584 | if (node == cpu_to_node(cpu)) | |
1585 | return cpu; | |
1586 | ||
1587 | /* Use "random" otherwise know as "first" online CPU of node */ | |
1588 | cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask); | |
1589 | ||
1590 | /* If CPU is valid return that, otherwise just defer */ | |
1591 | return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND; | |
1592 | } | |
1593 | ||
1594 | /** | |
1595 | * queue_work_node - queue work on a "random" cpu for a given NUMA node | |
1596 | * @node: NUMA node that we are targeting the work for | |
1597 | * @wq: workqueue to use | |
1598 | * @work: work to queue | |
1599 | * | |
1600 | * We queue the work to a "random" CPU within a given NUMA node. The basic | |
1601 | * idea here is to provide a way to somehow associate work with a given | |
1602 | * NUMA node. | |
1603 | * | |
1604 | * This function will only make a best effort attempt at getting this onto | |
1605 | * the right NUMA node. If no node is requested or the requested node is | |
1606 | * offline then we just fall back to standard queue_work behavior. | |
1607 | * | |
1608 | * Currently the "random" CPU ends up being the first available CPU in the | |
1609 | * intersection of cpu_online_mask and the cpumask of the node, unless we | |
1610 | * are running on the node. In that case we just use the current CPU. | |
1611 | * | |
1612 | * Return: %false if @work was already on a queue, %true otherwise. | |
1613 | */ | |
1614 | bool queue_work_node(int node, struct workqueue_struct *wq, | |
1615 | struct work_struct *work) | |
1616 | { | |
1617 | unsigned long flags; | |
1618 | bool ret = false; | |
1619 | ||
1620 | /* | |
1621 | * This current implementation is specific to unbound workqueues. | |
1622 | * Specifically we only return the first available CPU for a given | |
1623 | * node instead of cycling through individual CPUs within the node. | |
1624 | * | |
1625 | * If this is used with a per-cpu workqueue then the logic in | |
1626 | * workqueue_select_cpu_near would need to be updated to allow for | |
1627 | * some round robin type logic. | |
1628 | */ | |
1629 | WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)); | |
1630 | ||
1631 | local_irq_save(flags); | |
1632 | ||
1633 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { | |
1634 | int cpu = workqueue_select_cpu_near(node); | |
1635 | ||
1636 | __queue_work(cpu, wq, work); | |
1637 | ret = true; | |
1638 | } | |
1639 | ||
1640 | local_irq_restore(flags); | |
1641 | return ret; | |
1642 | } | |
1643 | EXPORT_SYMBOL_GPL(queue_work_node); | |
1644 | ||
8c20feb6 | 1645 | void delayed_work_timer_fn(struct timer_list *t) |
1da177e4 | 1646 | { |
8c20feb6 | 1647 | struct delayed_work *dwork = from_timer(dwork, t, timer); |
1da177e4 | 1648 | |
e0aecdd8 | 1649 | /* should have been called from irqsafe timer with irq already off */ |
60c057bc | 1650 | __queue_work(dwork->cpu, dwork->wq, &dwork->work); |
1da177e4 | 1651 | } |
1438ade5 | 1652 | EXPORT_SYMBOL(delayed_work_timer_fn); |
1da177e4 | 1653 | |
7beb2edf TH |
1654 | static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, |
1655 | struct delayed_work *dwork, unsigned long delay) | |
1da177e4 | 1656 | { |
7beb2edf TH |
1657 | struct timer_list *timer = &dwork->timer; |
1658 | struct work_struct *work = &dwork->work; | |
7beb2edf | 1659 | |
637fdbae | 1660 | WARN_ON_ONCE(!wq); |
4b243563 | 1661 | WARN_ON_ONCE(timer->function != delayed_work_timer_fn); |
fc4b514f TH |
1662 | WARN_ON_ONCE(timer_pending(timer)); |
1663 | WARN_ON_ONCE(!list_empty(&work->entry)); | |
7beb2edf | 1664 | |
8852aac2 TH |
1665 | /* |
1666 | * If @delay is 0, queue @dwork->work immediately. This is for | |
1667 | * both optimization and correctness. The earliest @timer can | |
1668 | * expire is on the closest next tick and delayed_work users depend | |
1669 | * on that there's no such delay when @delay is 0. | |
1670 | */ | |
1671 | if (!delay) { | |
1672 | __queue_work(cpu, wq, &dwork->work); | |
1673 | return; | |
1674 | } | |
1675 | ||
60c057bc | 1676 | dwork->wq = wq; |
1265057f | 1677 | dwork->cpu = cpu; |
7beb2edf TH |
1678 | timer->expires = jiffies + delay; |
1679 | ||
041bd12e TH |
1680 | if (unlikely(cpu != WORK_CPU_UNBOUND)) |
1681 | add_timer_on(timer, cpu); | |
1682 | else | |
1683 | add_timer(timer); | |
1da177e4 LT |
1684 | } |
1685 | ||
0fcb78c2 REB |
1686 | /** |
1687 | * queue_delayed_work_on - queue work on specific CPU after delay | |
1688 | * @cpu: CPU number to execute work on | |
1689 | * @wq: workqueue to use | |
af9997e4 | 1690 | * @dwork: work to queue |
0fcb78c2 REB |
1691 | * @delay: number of jiffies to wait before queueing |
1692 | * | |
d185af30 | 1693 | * Return: %false if @work was already on a queue, %true otherwise. If |
715f1300 TH |
1694 | * @delay is zero and @dwork is idle, it will be scheduled for immediate |
1695 | * execution. | |
0fcb78c2 | 1696 | */ |
d4283e93 TH |
1697 | bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, |
1698 | struct delayed_work *dwork, unsigned long delay) | |
7a6bc1cd | 1699 | { |
52bad64d | 1700 | struct work_struct *work = &dwork->work; |
d4283e93 | 1701 | bool ret = false; |
8930caba | 1702 | unsigned long flags; |
7a6bc1cd | 1703 | |
8930caba TH |
1704 | /* read the comment in __queue_work() */ |
1705 | local_irq_save(flags); | |
7a6bc1cd | 1706 | |
22df02bb | 1707 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { |
7beb2edf | 1708 | __queue_delayed_work(cpu, wq, dwork, delay); |
d4283e93 | 1709 | ret = true; |
7a6bc1cd | 1710 | } |
8a3e77cc | 1711 | |
8930caba | 1712 | local_irq_restore(flags); |
7a6bc1cd VP |
1713 | return ret; |
1714 | } | |
ad7b1f84 | 1715 | EXPORT_SYMBOL(queue_delayed_work_on); |
c7fc77f7 | 1716 | |
8376fe22 TH |
1717 | /** |
1718 | * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU | |
1719 | * @cpu: CPU number to execute work on | |
1720 | * @wq: workqueue to use | |
1721 | * @dwork: work to queue | |
1722 | * @delay: number of jiffies to wait before queueing | |
1723 | * | |
1724 | * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, | |
1725 | * modify @dwork's timer so that it expires after @delay. If @delay is | |
1726 | * zero, @work is guaranteed to be scheduled immediately regardless of its | |
1727 | * current state. | |
1728 | * | |
d185af30 | 1729 | * Return: %false if @dwork was idle and queued, %true if @dwork was |
8376fe22 TH |
1730 | * pending and its timer was modified. |
1731 | * | |
e0aecdd8 | 1732 | * This function is safe to call from any context including IRQ handler. |
8376fe22 TH |
1733 | * See try_to_grab_pending() for details. |
1734 | */ | |
1735 | bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, | |
1736 | struct delayed_work *dwork, unsigned long delay) | |
1737 | { | |
1738 | unsigned long flags; | |
1739 | int ret; | |
c7fc77f7 | 1740 | |
8376fe22 TH |
1741 | do { |
1742 | ret = try_to_grab_pending(&dwork->work, true, &flags); | |
1743 | } while (unlikely(ret == -EAGAIN)); | |
63bc0362 | 1744 | |
8376fe22 TH |
1745 | if (likely(ret >= 0)) { |
1746 | __queue_delayed_work(cpu, wq, dwork, delay); | |
1747 | local_irq_restore(flags); | |
7a6bc1cd | 1748 | } |
8376fe22 TH |
1749 | |
1750 | /* -ENOENT from try_to_grab_pending() becomes %true */ | |
7a6bc1cd VP |
1751 | return ret; |
1752 | } | |
8376fe22 TH |
1753 | EXPORT_SYMBOL_GPL(mod_delayed_work_on); |
1754 | ||
05f0fe6b TH |
1755 | static void rcu_work_rcufn(struct rcu_head *rcu) |
1756 | { | |
1757 | struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu); | |
1758 | ||
1759 | /* read the comment in __queue_work() */ | |
1760 | local_irq_disable(); | |
1761 | __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work); | |
1762 | local_irq_enable(); | |
1763 | } | |
1764 | ||
1765 | /** | |
1766 | * queue_rcu_work - queue work after a RCU grace period | |
1767 | * @wq: workqueue to use | |
1768 | * @rwork: work to queue | |
1769 | * | |
1770 | * Return: %false if @rwork was already pending, %true otherwise. Note | |
1771 | * that a full RCU grace period is guaranteed only after a %true return. | |
bf393fd4 | 1772 | * While @rwork is guaranteed to be executed after a %false return, the |
05f0fe6b TH |
1773 | * execution may happen before a full RCU grace period has passed. |
1774 | */ | |
1775 | bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork) | |
1776 | { | |
1777 | struct work_struct *work = &rwork->work; | |
1778 | ||
1779 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { | |
1780 | rwork->wq = wq; | |
a7e30c0e | 1781 | call_rcu_hurry(&rwork->rcu, rcu_work_rcufn); |
05f0fe6b TH |
1782 | return true; |
1783 | } | |
1784 | ||
1785 | return false; | |
1786 | } | |
1787 | EXPORT_SYMBOL(queue_rcu_work); | |
1788 | ||
c8e55f36 TH |
1789 | /** |
1790 | * worker_enter_idle - enter idle state | |
1791 | * @worker: worker which is entering idle state | |
1792 | * | |
1793 | * @worker is entering idle state. Update stats and idle timer if | |
1794 | * necessary. | |
1795 | * | |
1796 | * LOCKING: | |
a9b8a985 | 1797 | * raw_spin_lock_irq(pool->lock). |
c8e55f36 TH |
1798 | */ |
1799 | static void worker_enter_idle(struct worker *worker) | |
1da177e4 | 1800 | { |
bd7bdd43 | 1801 | struct worker_pool *pool = worker->pool; |
c8e55f36 | 1802 | |
6183c009 TH |
1803 | if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || |
1804 | WARN_ON_ONCE(!list_empty(&worker->entry) && | |
1805 | (worker->hentry.next || worker->hentry.pprev))) | |
1806 | return; | |
c8e55f36 | 1807 | |
051e1850 | 1808 | /* can't use worker_set_flags(), also called from create_worker() */ |
cb444766 | 1809 | worker->flags |= WORKER_IDLE; |
bd7bdd43 | 1810 | pool->nr_idle++; |
e22bee78 | 1811 | worker->last_active = jiffies; |
c8e55f36 TH |
1812 | |
1813 | /* idle_list is LIFO */ | |
bd7bdd43 | 1814 | list_add(&worker->entry, &pool->idle_list); |
db7bccf4 | 1815 | |
628c78e7 TH |
1816 | if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) |
1817 | mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); | |
cb444766 | 1818 | |
989442d7 | 1819 | /* Sanity check nr_running. */ |
bc35f7ef | 1820 | WARN_ON_ONCE(pool->nr_workers == pool->nr_idle && pool->nr_running); |
c8e55f36 TH |
1821 | } |
1822 | ||
1823 | /** | |
1824 | * worker_leave_idle - leave idle state | |
1825 | * @worker: worker which is leaving idle state | |
1826 | * | |
1827 | * @worker is leaving idle state. Update stats. | |
1828 | * | |
1829 | * LOCKING: | |
a9b8a985 | 1830 | * raw_spin_lock_irq(pool->lock). |
c8e55f36 TH |
1831 | */ |
1832 | static void worker_leave_idle(struct worker *worker) | |
1833 | { | |
bd7bdd43 | 1834 | struct worker_pool *pool = worker->pool; |
c8e55f36 | 1835 | |
6183c009 TH |
1836 | if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) |
1837 | return; | |
d302f017 | 1838 | worker_clr_flags(worker, WORKER_IDLE); |
bd7bdd43 | 1839 | pool->nr_idle--; |
c8e55f36 TH |
1840 | list_del_init(&worker->entry); |
1841 | } | |
1842 | ||
f7537df5 | 1843 | static struct worker *alloc_worker(int node) |
c34056a3 TH |
1844 | { |
1845 | struct worker *worker; | |
1846 | ||
f7537df5 | 1847 | worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node); |
c8e55f36 TH |
1848 | if (worker) { |
1849 | INIT_LIST_HEAD(&worker->entry); | |
affee4b2 | 1850 | INIT_LIST_HEAD(&worker->scheduled); |
da028469 | 1851 | INIT_LIST_HEAD(&worker->node); |
e22bee78 TH |
1852 | /* on creation a worker is in !idle && prep state */ |
1853 | worker->flags = WORKER_PREP; | |
c8e55f36 | 1854 | } |
c34056a3 TH |
1855 | return worker; |
1856 | } | |
1857 | ||
4736cbf7 LJ |
1858 | /** |
1859 | * worker_attach_to_pool() - attach a worker to a pool | |
1860 | * @worker: worker to be attached | |
1861 | * @pool: the target pool | |
1862 | * | |
1863 | * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and | |
1864 | * cpu-binding of @worker are kept coordinated with the pool across | |
1865 | * cpu-[un]hotplugs. | |
1866 | */ | |
1867 | static void worker_attach_to_pool(struct worker *worker, | |
1868 | struct worker_pool *pool) | |
1869 | { | |
1258fae7 | 1870 | mutex_lock(&wq_pool_attach_mutex); |
4736cbf7 | 1871 | |
4736cbf7 | 1872 | /* |
1258fae7 TH |
1873 | * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains |
1874 | * stable across this function. See the comments above the flag | |
1875 | * definition for details. | |
4736cbf7 LJ |
1876 | */ |
1877 | if (pool->flags & POOL_DISASSOCIATED) | |
1878 | worker->flags |= WORKER_UNBOUND; | |
5c25b5ff PZ |
1879 | else |
1880 | kthread_set_per_cpu(worker->task, pool->cpu); | |
4736cbf7 | 1881 | |
640f17c8 PZ |
1882 | if (worker->rescue_wq) |
1883 | set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask); | |
1884 | ||
4736cbf7 | 1885 | list_add_tail(&worker->node, &pool->workers); |
a2d812a2 | 1886 | worker->pool = pool; |
4736cbf7 | 1887 | |
1258fae7 | 1888 | mutex_unlock(&wq_pool_attach_mutex); |
4736cbf7 LJ |
1889 | } |
1890 | ||
60f5a4bc LJ |
1891 | /** |
1892 | * worker_detach_from_pool() - detach a worker from its pool | |
1893 | * @worker: worker which is attached to its pool | |
60f5a4bc | 1894 | * |
4736cbf7 LJ |
1895 | * Undo the attaching which had been done in worker_attach_to_pool(). The |
1896 | * caller worker shouldn't access to the pool after detached except it has | |
1897 | * other reference to the pool. | |
60f5a4bc | 1898 | */ |
a2d812a2 | 1899 | static void worker_detach_from_pool(struct worker *worker) |
60f5a4bc | 1900 | { |
a2d812a2 | 1901 | struct worker_pool *pool = worker->pool; |
60f5a4bc LJ |
1902 | struct completion *detach_completion = NULL; |
1903 | ||
1258fae7 | 1904 | mutex_lock(&wq_pool_attach_mutex); |
a2d812a2 | 1905 | |
5c25b5ff | 1906 | kthread_set_per_cpu(worker->task, -1); |
da028469 | 1907 | list_del(&worker->node); |
a2d812a2 TH |
1908 | worker->pool = NULL; |
1909 | ||
e02b9312 | 1910 | if (list_empty(&pool->workers) && list_empty(&pool->dying_workers)) |
60f5a4bc | 1911 | detach_completion = pool->detach_completion; |
1258fae7 | 1912 | mutex_unlock(&wq_pool_attach_mutex); |
60f5a4bc | 1913 | |
b62c0751 LJ |
1914 | /* clear leftover flags without pool->lock after it is detached */ |
1915 | worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND); | |
1916 | ||
60f5a4bc LJ |
1917 | if (detach_completion) |
1918 | complete(detach_completion); | |
1919 | } | |
1920 | ||
c34056a3 TH |
1921 | /** |
1922 | * create_worker - create a new workqueue worker | |
63d95a91 | 1923 | * @pool: pool the new worker will belong to |
c34056a3 | 1924 | * |
051e1850 | 1925 | * Create and start a new worker which is attached to @pool. |
c34056a3 TH |
1926 | * |
1927 | * CONTEXT: | |
1928 | * Might sleep. Does GFP_KERNEL allocations. | |
1929 | * | |
d185af30 | 1930 | * Return: |
c34056a3 TH |
1931 | * Pointer to the newly created worker. |
1932 | */ | |
bc2ae0f5 | 1933 | static struct worker *create_worker(struct worker_pool *pool) |
c34056a3 | 1934 | { |
e441b56f ZL |
1935 | struct worker *worker; |
1936 | int id; | |
e3c916a4 | 1937 | char id_buf[16]; |
c34056a3 | 1938 | |
7cda9aae | 1939 | /* ID is needed to determine kthread name */ |
e441b56f | 1940 | id = ida_alloc(&pool->worker_ida, GFP_KERNEL); |
822d8405 | 1941 | if (id < 0) |
e441b56f | 1942 | return NULL; |
c34056a3 | 1943 | |
f7537df5 | 1944 | worker = alloc_worker(pool->node); |
c34056a3 TH |
1945 | if (!worker) |
1946 | goto fail; | |
1947 | ||
c34056a3 TH |
1948 | worker->id = id; |
1949 | ||
29c91e99 | 1950 | if (pool->cpu >= 0) |
e3c916a4 TH |
1951 | snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, |
1952 | pool->attrs->nice < 0 ? "H" : ""); | |
f3421797 | 1953 | else |
e3c916a4 TH |
1954 | snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); |
1955 | ||
f3f90ad4 | 1956 | worker->task = kthread_create_on_node(worker_thread, worker, pool->node, |
e3c916a4 | 1957 | "kworker/%s", id_buf); |
c34056a3 TH |
1958 | if (IS_ERR(worker->task)) |
1959 | goto fail; | |
1960 | ||
91151228 | 1961 | set_user_nice(worker->task, pool->attrs->nice); |
25834c73 | 1962 | kthread_bind_mask(worker->task, pool->attrs->cpumask); |
91151228 | 1963 | |
da028469 | 1964 | /* successful, attach the worker to the pool */ |
4736cbf7 | 1965 | worker_attach_to_pool(worker, pool); |
822d8405 | 1966 | |
051e1850 | 1967 | /* start the newly created worker */ |
a9b8a985 | 1968 | raw_spin_lock_irq(&pool->lock); |
051e1850 LJ |
1969 | worker->pool->nr_workers++; |
1970 | worker_enter_idle(worker); | |
1971 | wake_up_process(worker->task); | |
a9b8a985 | 1972 | raw_spin_unlock_irq(&pool->lock); |
051e1850 | 1973 | |
c34056a3 | 1974 | return worker; |
822d8405 | 1975 | |
c34056a3 | 1976 | fail: |
e441b56f | 1977 | ida_free(&pool->worker_ida, id); |
c34056a3 TH |
1978 | kfree(worker); |
1979 | return NULL; | |
1980 | } | |
1981 | ||
793777bc VS |
1982 | static void unbind_worker(struct worker *worker) |
1983 | { | |
1984 | lockdep_assert_held(&wq_pool_attach_mutex); | |
1985 | ||
1986 | kthread_set_per_cpu(worker->task, -1); | |
1987 | if (cpumask_intersects(wq_unbound_cpumask, cpu_active_mask)) | |
1988 | WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, wq_unbound_cpumask) < 0); | |
1989 | else | |
1990 | WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0); | |
1991 | } | |
1992 | ||
e02b9312 VS |
1993 | static void wake_dying_workers(struct list_head *cull_list) |
1994 | { | |
1995 | struct worker *worker, *tmp; | |
1996 | ||
1997 | list_for_each_entry_safe(worker, tmp, cull_list, entry) { | |
1998 | list_del_init(&worker->entry); | |
1999 | unbind_worker(worker); | |
2000 | /* | |
2001 | * If the worker was somehow already running, then it had to be | |
2002 | * in pool->idle_list when set_worker_dying() happened or we | |
2003 | * wouldn't have gotten here. | |
2004 | * | |
2005 | * Thus, the worker must either have observed the WORKER_DIE | |
2006 | * flag, or have set its state to TASK_IDLE. Either way, the | |
2007 | * below will be observed by the worker and is safe to do | |
2008 | * outside of pool->lock. | |
2009 | */ | |
2010 | wake_up_process(worker->task); | |
2011 | } | |
2012 | } | |
2013 | ||
c34056a3 | 2014 | /** |
e02b9312 | 2015 | * set_worker_dying - Tag a worker for destruction |
c34056a3 | 2016 | * @worker: worker to be destroyed |
e02b9312 | 2017 | * @list: transfer worker away from its pool->idle_list and into list |
c34056a3 | 2018 | * |
e02b9312 VS |
2019 | * Tag @worker for destruction and adjust @pool stats accordingly. The worker |
2020 | * should be idle. | |
c8e55f36 TH |
2021 | * |
2022 | * CONTEXT: | |
a9b8a985 | 2023 | * raw_spin_lock_irq(pool->lock). |
c34056a3 | 2024 | */ |
e02b9312 | 2025 | static void set_worker_dying(struct worker *worker, struct list_head *list) |
c34056a3 | 2026 | { |
bd7bdd43 | 2027 | struct worker_pool *pool = worker->pool; |
c34056a3 | 2028 | |
cd549687 | 2029 | lockdep_assert_held(&pool->lock); |
e02b9312 | 2030 | lockdep_assert_held(&wq_pool_attach_mutex); |
cd549687 | 2031 | |
c34056a3 | 2032 | /* sanity check frenzy */ |
6183c009 | 2033 | if (WARN_ON(worker->current_work) || |
73eb7fe7 LJ |
2034 | WARN_ON(!list_empty(&worker->scheduled)) || |
2035 | WARN_ON(!(worker->flags & WORKER_IDLE))) | |
6183c009 | 2036 | return; |
c34056a3 | 2037 | |
73eb7fe7 LJ |
2038 | pool->nr_workers--; |
2039 | pool->nr_idle--; | |
5bdfff96 | 2040 | |
cb444766 | 2041 | worker->flags |= WORKER_DIE; |
e02b9312 VS |
2042 | |
2043 | list_move(&worker->entry, list); | |
2044 | list_move(&worker->node, &pool->dying_workers); | |
c34056a3 TH |
2045 | } |
2046 | ||
3f959aa3 VS |
2047 | /** |
2048 | * idle_worker_timeout - check if some idle workers can now be deleted. | |
2049 | * @t: The pool's idle_timer that just expired | |
2050 | * | |
2051 | * The timer is armed in worker_enter_idle(). Note that it isn't disarmed in | |
2052 | * worker_leave_idle(), as a worker flicking between idle and active while its | |
2053 | * pool is at the too_many_workers() tipping point would cause too much timer | |
2054 | * housekeeping overhead. Since IDLE_WORKER_TIMEOUT is long enough, we just let | |
2055 | * it expire and re-evaluate things from there. | |
2056 | */ | |
32a6c723 | 2057 | static void idle_worker_timeout(struct timer_list *t) |
e22bee78 | 2058 | { |
32a6c723 | 2059 | struct worker_pool *pool = from_timer(pool, t, idle_timer); |
3f959aa3 VS |
2060 | bool do_cull = false; |
2061 | ||
2062 | if (work_pending(&pool->idle_cull_work)) | |
2063 | return; | |
e22bee78 | 2064 | |
a9b8a985 | 2065 | raw_spin_lock_irq(&pool->lock); |
e22bee78 | 2066 | |
3f959aa3 | 2067 | if (too_many_workers(pool)) { |
e22bee78 TH |
2068 | struct worker *worker; |
2069 | unsigned long expires; | |
2070 | ||
2071 | /* idle_list is kept in LIFO order, check the last one */ | |
3f959aa3 VS |
2072 | worker = list_entry(pool->idle_list.prev, struct worker, entry); |
2073 | expires = worker->last_active + IDLE_WORKER_TIMEOUT; | |
2074 | do_cull = !time_before(jiffies, expires); | |
2075 | ||
2076 | if (!do_cull) | |
2077 | mod_timer(&pool->idle_timer, expires); | |
2078 | } | |
2079 | raw_spin_unlock_irq(&pool->lock); | |
2080 | ||
2081 | if (do_cull) | |
2082 | queue_work(system_unbound_wq, &pool->idle_cull_work); | |
2083 | } | |
2084 | ||
2085 | /** | |
2086 | * idle_cull_fn - cull workers that have been idle for too long. | |
2087 | * @work: the pool's work for handling these idle workers | |
2088 | * | |
2089 | * This goes through a pool's idle workers and gets rid of those that have been | |
2090 | * idle for at least IDLE_WORKER_TIMEOUT seconds. | |
e02b9312 VS |
2091 | * |
2092 | * We don't want to disturb isolated CPUs because of a pcpu kworker being | |
2093 | * culled, so this also resets worker affinity. This requires a sleepable | |
2094 | * context, hence the split between timer callback and work item. | |
3f959aa3 VS |
2095 | */ |
2096 | static void idle_cull_fn(struct work_struct *work) | |
2097 | { | |
2098 | struct worker_pool *pool = container_of(work, struct worker_pool, idle_cull_work); | |
e02b9312 | 2099 | struct list_head cull_list; |
3f959aa3 | 2100 | |
e02b9312 VS |
2101 | INIT_LIST_HEAD(&cull_list); |
2102 | /* | |
2103 | * Grabbing wq_pool_attach_mutex here ensures an already-running worker | |
2104 | * cannot proceed beyong worker_detach_from_pool() in its self-destruct | |
2105 | * path. This is required as a previously-preempted worker could run after | |
2106 | * set_worker_dying() has happened but before wake_dying_workers() did. | |
2107 | */ | |
2108 | mutex_lock(&wq_pool_attach_mutex); | |
3f959aa3 VS |
2109 | raw_spin_lock_irq(&pool->lock); |
2110 | ||
2111 | while (too_many_workers(pool)) { | |
2112 | struct worker *worker; | |
2113 | unsigned long expires; | |
2114 | ||
63d95a91 | 2115 | worker = list_entry(pool->idle_list.prev, struct worker, entry); |
e22bee78 TH |
2116 | expires = worker->last_active + IDLE_WORKER_TIMEOUT; |
2117 | ||
3347fc9f | 2118 | if (time_before(jiffies, expires)) { |
63d95a91 | 2119 | mod_timer(&pool->idle_timer, expires); |
3347fc9f | 2120 | break; |
d5abe669 | 2121 | } |
3347fc9f | 2122 | |
e02b9312 | 2123 | set_worker_dying(worker, &cull_list); |
e22bee78 TH |
2124 | } |
2125 | ||
a9b8a985 | 2126 | raw_spin_unlock_irq(&pool->lock); |
e02b9312 VS |
2127 | wake_dying_workers(&cull_list); |
2128 | mutex_unlock(&wq_pool_attach_mutex); | |
e22bee78 | 2129 | } |
d5abe669 | 2130 | |
493a1724 | 2131 | static void send_mayday(struct work_struct *work) |
e22bee78 | 2132 | { |
112202d9 TH |
2133 | struct pool_workqueue *pwq = get_work_pwq(work); |
2134 | struct workqueue_struct *wq = pwq->wq; | |
493a1724 | 2135 | |
2e109a28 | 2136 | lockdep_assert_held(&wq_mayday_lock); |
e22bee78 | 2137 | |
493008a8 | 2138 | if (!wq->rescuer) |
493a1724 | 2139 | return; |
e22bee78 TH |
2140 | |
2141 | /* mayday mayday mayday */ | |
493a1724 | 2142 | if (list_empty(&pwq->mayday_node)) { |
77668c8b LJ |
2143 | /* |
2144 | * If @pwq is for an unbound wq, its base ref may be put at | |
2145 | * any time due to an attribute change. Pin @pwq until the | |
2146 | * rescuer is done with it. | |
2147 | */ | |
2148 | get_pwq(pwq); | |
493a1724 | 2149 | list_add_tail(&pwq->mayday_node, &wq->maydays); |
e22bee78 | 2150 | wake_up_process(wq->rescuer->task); |
493a1724 | 2151 | } |
e22bee78 TH |
2152 | } |
2153 | ||
32a6c723 | 2154 | static void pool_mayday_timeout(struct timer_list *t) |
e22bee78 | 2155 | { |
32a6c723 | 2156 | struct worker_pool *pool = from_timer(pool, t, mayday_timer); |
e22bee78 TH |
2157 | struct work_struct *work; |
2158 | ||
a9b8a985 SAS |
2159 | raw_spin_lock_irq(&pool->lock); |
2160 | raw_spin_lock(&wq_mayday_lock); /* for wq->maydays */ | |
e22bee78 | 2161 | |
63d95a91 | 2162 | if (need_to_create_worker(pool)) { |
e22bee78 TH |
2163 | /* |
2164 | * We've been trying to create a new worker but | |
2165 | * haven't been successful. We might be hitting an | |
2166 | * allocation deadlock. Send distress signals to | |
2167 | * rescuers. | |
2168 | */ | |
63d95a91 | 2169 | list_for_each_entry(work, &pool->worklist, entry) |
e22bee78 | 2170 | send_mayday(work); |
1da177e4 | 2171 | } |
e22bee78 | 2172 | |
a9b8a985 SAS |
2173 | raw_spin_unlock(&wq_mayday_lock); |
2174 | raw_spin_unlock_irq(&pool->lock); | |
e22bee78 | 2175 | |
63d95a91 | 2176 | mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); |
1da177e4 LT |
2177 | } |
2178 | ||
e22bee78 TH |
2179 | /** |
2180 | * maybe_create_worker - create a new worker if necessary | |
63d95a91 | 2181 | * @pool: pool to create a new worker for |
e22bee78 | 2182 | * |
63d95a91 | 2183 | * Create a new worker for @pool if necessary. @pool is guaranteed to |
e22bee78 TH |
2184 | * have at least one idle worker on return from this function. If |
2185 | * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is | |
63d95a91 | 2186 | * sent to all rescuers with works scheduled on @pool to resolve |
e22bee78 TH |
2187 | * possible allocation deadlock. |
2188 | * | |
c5aa87bb TH |
2189 | * On return, need_to_create_worker() is guaranteed to be %false and |
2190 | * may_start_working() %true. | |
e22bee78 TH |
2191 | * |
2192 | * LOCKING: | |
a9b8a985 | 2193 | * raw_spin_lock_irq(pool->lock) which may be released and regrabbed |
e22bee78 TH |
2194 | * multiple times. Does GFP_KERNEL allocations. Called only from |
2195 | * manager. | |
e22bee78 | 2196 | */ |
29187a9e | 2197 | static void maybe_create_worker(struct worker_pool *pool) |
d565ed63 TH |
2198 | __releases(&pool->lock) |
2199 | __acquires(&pool->lock) | |
1da177e4 | 2200 | { |
e22bee78 | 2201 | restart: |
a9b8a985 | 2202 | raw_spin_unlock_irq(&pool->lock); |
9f9c2364 | 2203 | |
e22bee78 | 2204 | /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ |
63d95a91 | 2205 | mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); |
e22bee78 TH |
2206 | |
2207 | while (true) { | |
051e1850 | 2208 | if (create_worker(pool) || !need_to_create_worker(pool)) |
e22bee78 | 2209 | break; |
1da177e4 | 2210 | |
e212f361 | 2211 | schedule_timeout_interruptible(CREATE_COOLDOWN); |
9f9c2364 | 2212 | |
63d95a91 | 2213 | if (!need_to_create_worker(pool)) |
e22bee78 TH |
2214 | break; |
2215 | } | |
2216 | ||
63d95a91 | 2217 | del_timer_sync(&pool->mayday_timer); |
a9b8a985 | 2218 | raw_spin_lock_irq(&pool->lock); |
051e1850 LJ |
2219 | /* |
2220 | * This is necessary even after a new worker was just successfully | |
2221 | * created as @pool->lock was dropped and the new worker might have | |
2222 | * already become busy. | |
2223 | */ | |
63d95a91 | 2224 | if (need_to_create_worker(pool)) |
e22bee78 | 2225 | goto restart; |
e22bee78 TH |
2226 | } |
2227 | ||
73f53c4a | 2228 | /** |
e22bee78 TH |
2229 | * manage_workers - manage worker pool |
2230 | * @worker: self | |
73f53c4a | 2231 | * |
706026c2 | 2232 | * Assume the manager role and manage the worker pool @worker belongs |
e22bee78 | 2233 | * to. At any given time, there can be only zero or one manager per |
706026c2 | 2234 | * pool. The exclusion is handled automatically by this function. |
e22bee78 TH |
2235 | * |
2236 | * The caller can safely start processing works on false return. On | |
2237 | * true return, it's guaranteed that need_to_create_worker() is false | |
2238 | * and may_start_working() is true. | |
73f53c4a TH |
2239 | * |
2240 | * CONTEXT: | |
a9b8a985 | 2241 | * raw_spin_lock_irq(pool->lock) which may be released and regrabbed |
e22bee78 TH |
2242 | * multiple times. Does GFP_KERNEL allocations. |
2243 | * | |
d185af30 | 2244 | * Return: |
29187a9e TH |
2245 | * %false if the pool doesn't need management and the caller can safely |
2246 | * start processing works, %true if management function was performed and | |
2247 | * the conditions that the caller verified before calling the function may | |
2248 | * no longer be true. | |
73f53c4a | 2249 | */ |
e22bee78 | 2250 | static bool manage_workers(struct worker *worker) |
73f53c4a | 2251 | { |
63d95a91 | 2252 | struct worker_pool *pool = worker->pool; |
73f53c4a | 2253 | |
692b4825 | 2254 | if (pool->flags & POOL_MANAGER_ACTIVE) |
29187a9e | 2255 | return false; |
692b4825 TH |
2256 | |
2257 | pool->flags |= POOL_MANAGER_ACTIVE; | |
2607d7a6 | 2258 | pool->manager = worker; |
1e19ffc6 | 2259 | |
29187a9e | 2260 | maybe_create_worker(pool); |
e22bee78 | 2261 | |
2607d7a6 | 2262 | pool->manager = NULL; |
692b4825 | 2263 | pool->flags &= ~POOL_MANAGER_ACTIVE; |
d8bb65ab | 2264 | rcuwait_wake_up(&manager_wait); |
29187a9e | 2265 | return true; |
73f53c4a TH |
2266 | } |
2267 | ||
a62428c0 TH |
2268 | /** |
2269 | * process_one_work - process single work | |
c34056a3 | 2270 | * @worker: self |
a62428c0 TH |
2271 | * @work: work to process |
2272 | * | |
2273 | * Process @work. This function contains all the logics necessary to | |
2274 | * process a single work including synchronization against and | |
2275 | * interaction with other workers on the same cpu, queueing and | |
2276 | * flushing. As long as context requirement is met, any worker can | |
2277 | * call this function to process a work. | |
2278 | * | |
2279 | * CONTEXT: | |
a9b8a985 | 2280 | * raw_spin_lock_irq(pool->lock) which is released and regrabbed. |
a62428c0 | 2281 | */ |
c34056a3 | 2282 | static void process_one_work(struct worker *worker, struct work_struct *work) |
d565ed63 TH |
2283 | __releases(&pool->lock) |
2284 | __acquires(&pool->lock) | |
a62428c0 | 2285 | { |
112202d9 | 2286 | struct pool_workqueue *pwq = get_work_pwq(work); |
bd7bdd43 | 2287 | struct worker_pool *pool = worker->pool; |
112202d9 | 2288 | bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE; |
c4560c2c | 2289 | unsigned long work_data; |
7e11629d | 2290 | struct worker *collision; |
a62428c0 TH |
2291 | #ifdef CONFIG_LOCKDEP |
2292 | /* | |
2293 | * It is permissible to free the struct work_struct from | |
2294 | * inside the function that is called from it, this we need to | |
2295 | * take into account for lockdep too. To avoid bogus "held | |
2296 | * lock freed" warnings as well as problems when looking into | |
2297 | * work->lockdep_map, make a copy and use that here. | |
2298 | */ | |
4d82a1de PZ |
2299 | struct lockdep_map lockdep_map; |
2300 | ||
2301 | lockdep_copy_map(&lockdep_map, &work->lockdep_map); | |
a62428c0 | 2302 | #endif |
807407c0 | 2303 | /* ensure we're on the correct CPU */ |
85327af6 | 2304 | WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && |
ec22ca5e | 2305 | raw_smp_processor_id() != pool->cpu); |
25511a47 | 2306 | |
7e11629d TH |
2307 | /* |
2308 | * A single work shouldn't be executed concurrently by | |
2309 | * multiple workers on a single cpu. Check whether anyone is | |
2310 | * already processing the work. If so, defer the work to the | |
2311 | * currently executing one. | |
2312 | */ | |
c9e7cf27 | 2313 | collision = find_worker_executing_work(pool, work); |
7e11629d TH |
2314 | if (unlikely(collision)) { |
2315 | move_linked_works(work, &collision->scheduled, NULL); | |
2316 | return; | |
2317 | } | |
2318 | ||
8930caba | 2319 | /* claim and dequeue */ |
a62428c0 | 2320 | debug_work_deactivate(work); |
c9e7cf27 | 2321 | hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); |
c34056a3 | 2322 | worker->current_work = work; |
a2c1c57b | 2323 | worker->current_func = work->func; |
112202d9 | 2324 | worker->current_pwq = pwq; |
c4560c2c | 2325 | work_data = *work_data_bits(work); |
d812796e | 2326 | worker->current_color = get_work_color(work_data); |
7a22ad75 | 2327 | |
8bf89593 TH |
2328 | /* |
2329 | * Record wq name for cmdline and debug reporting, may get | |
2330 | * overridden through set_worker_desc(). | |
2331 | */ | |
2332 | strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN); | |
2333 | ||
a62428c0 TH |
2334 | list_del_init(&work->entry); |
2335 | ||
fb0e7beb | 2336 | /* |
228f1d00 LJ |
2337 | * CPU intensive works don't participate in concurrency management. |
2338 | * They're the scheduler's responsibility. This takes @worker out | |
2339 | * of concurrency management and the next code block will chain | |
2340 | * execution of the pending work items. | |
fb0e7beb TH |
2341 | */ |
2342 | if (unlikely(cpu_intensive)) | |
228f1d00 | 2343 | worker_set_flags(worker, WORKER_CPU_INTENSIVE); |
fb0e7beb | 2344 | |
974271c4 | 2345 | /* |
a489a03e LJ |
2346 | * Wake up another worker if necessary. The condition is always |
2347 | * false for normal per-cpu workers since nr_running would always | |
2348 | * be >= 1 at this point. This is used to chain execution of the | |
2349 | * pending work items for WORKER_NOT_RUNNING workers such as the | |
228f1d00 | 2350 | * UNBOUND and CPU_INTENSIVE ones. |
974271c4 | 2351 | */ |
a489a03e | 2352 | if (need_more_worker(pool)) |
63d95a91 | 2353 | wake_up_worker(pool); |
974271c4 | 2354 | |
8930caba | 2355 | /* |
7c3eed5c | 2356 | * Record the last pool and clear PENDING which should be the last |
d565ed63 | 2357 | * update to @work. Also, do this inside @pool->lock so that |
23657bb1 TH |
2358 | * PENDING and queued state changes happen together while IRQ is |
2359 | * disabled. | |
8930caba | 2360 | */ |
7c3eed5c | 2361 | set_work_pool_and_clear_pending(work, pool->id); |
a62428c0 | 2362 | |
a9b8a985 | 2363 | raw_spin_unlock_irq(&pool->lock); |
a62428c0 | 2364 | |
a1d14934 | 2365 | lock_map_acquire(&pwq->wq->lockdep_map); |
a62428c0 | 2366 | lock_map_acquire(&lockdep_map); |
e6f3faa7 | 2367 | /* |
f52be570 PZ |
2368 | * Strictly speaking we should mark the invariant state without holding |
2369 | * any locks, that is, before these two lock_map_acquire()'s. | |
e6f3faa7 PZ |
2370 | * |
2371 | * However, that would result in: | |
2372 | * | |
2373 | * A(W1) | |
2374 | * WFC(C) | |
2375 | * A(W1) | |
2376 | * C(C) | |
2377 | * | |
2378 | * Which would create W1->C->W1 dependencies, even though there is no | |
2379 | * actual deadlock possible. There are two solutions, using a | |
2380 | * read-recursive acquire on the work(queue) 'locks', but this will then | |
f52be570 | 2381 | * hit the lockdep limitation on recursive locks, or simply discard |
e6f3faa7 PZ |
2382 | * these locks. |
2383 | * | |
2384 | * AFAICT there is no possible deadlock scenario between the | |
2385 | * flush_work() and complete() primitives (except for single-threaded | |
2386 | * workqueues), so hiding them isn't a problem. | |
2387 | */ | |
f52be570 | 2388 | lockdep_invariant_state(true); |
e36c886a | 2389 | trace_workqueue_execute_start(work); |
a2c1c57b | 2390 | worker->current_func(work); |
e36c886a AV |
2391 | /* |
2392 | * While we must be careful to not use "work" after this, the trace | |
2393 | * point will only record its address. | |
2394 | */ | |
1c5da0ec | 2395 | trace_workqueue_execute_end(work, worker->current_func); |
a62428c0 | 2396 | lock_map_release(&lockdep_map); |
112202d9 | 2397 | lock_map_release(&pwq->wq->lockdep_map); |
a62428c0 TH |
2398 | |
2399 | if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { | |
044c782c | 2400 | pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" |
d75f773c | 2401 | " last function: %ps\n", |
a2c1c57b TH |
2402 | current->comm, preempt_count(), task_pid_nr(current), |
2403 | worker->current_func); | |
a62428c0 TH |
2404 | debug_show_held_locks(current); |
2405 | dump_stack(); | |
2406 | } | |
2407 | ||
b22ce278 | 2408 | /* |
025f50f3 | 2409 | * The following prevents a kworker from hogging CPU on !PREEMPTION |
b22ce278 TH |
2410 | * kernels, where a requeueing work item waiting for something to |
2411 | * happen could deadlock with stop_machine as such work item could | |
2412 | * indefinitely requeue itself while all other CPUs are trapped in | |
789cbbec JL |
2413 | * stop_machine. At the same time, report a quiescent RCU state so |
2414 | * the same condition doesn't freeze RCU. | |
b22ce278 | 2415 | */ |
a7e6425e | 2416 | cond_resched(); |
b22ce278 | 2417 | |
a9b8a985 | 2418 | raw_spin_lock_irq(&pool->lock); |
a62428c0 | 2419 | |
fb0e7beb TH |
2420 | /* clear cpu intensive status */ |
2421 | if (unlikely(cpu_intensive)) | |
2422 | worker_clr_flags(worker, WORKER_CPU_INTENSIVE); | |
2423 | ||
1b69ac6b JW |
2424 | /* tag the worker for identification in schedule() */ |
2425 | worker->last_func = worker->current_func; | |
2426 | ||
a62428c0 | 2427 | /* we're done with it, release */ |
42f8570f | 2428 | hash_del(&worker->hentry); |
c34056a3 | 2429 | worker->current_work = NULL; |
a2c1c57b | 2430 | worker->current_func = NULL; |
112202d9 | 2431 | worker->current_pwq = NULL; |
d812796e | 2432 | worker->current_color = INT_MAX; |
c4560c2c | 2433 | pwq_dec_nr_in_flight(pwq, work_data); |
a62428c0 TH |
2434 | } |
2435 | ||
affee4b2 TH |
2436 | /** |
2437 | * process_scheduled_works - process scheduled works | |
2438 | * @worker: self | |
2439 | * | |
2440 | * Process all scheduled works. Please note that the scheduled list | |
2441 | * may change while processing a work, so this function repeatedly | |
2442 | * fetches a work from the top and executes it. | |
2443 | * | |
2444 | * CONTEXT: | |
a9b8a985 | 2445 | * raw_spin_lock_irq(pool->lock) which may be released and regrabbed |
affee4b2 TH |
2446 | * multiple times. |
2447 | */ | |
2448 | static void process_scheduled_works(struct worker *worker) | |
1da177e4 | 2449 | { |
affee4b2 TH |
2450 | while (!list_empty(&worker->scheduled)) { |
2451 | struct work_struct *work = list_first_entry(&worker->scheduled, | |
1da177e4 | 2452 | struct work_struct, entry); |
c34056a3 | 2453 | process_one_work(worker, work); |
1da177e4 | 2454 | } |
1da177e4 LT |
2455 | } |
2456 | ||
197f6acc TH |
2457 | static void set_pf_worker(bool val) |
2458 | { | |
2459 | mutex_lock(&wq_pool_attach_mutex); | |
2460 | if (val) | |
2461 | current->flags |= PF_WQ_WORKER; | |
2462 | else | |
2463 | current->flags &= ~PF_WQ_WORKER; | |
2464 | mutex_unlock(&wq_pool_attach_mutex); | |
2465 | } | |
2466 | ||
4690c4ab TH |
2467 | /** |
2468 | * worker_thread - the worker thread function | |
c34056a3 | 2469 | * @__worker: self |
4690c4ab | 2470 | * |
c5aa87bb TH |
2471 | * The worker thread function. All workers belong to a worker_pool - |
2472 | * either a per-cpu one or dynamic unbound one. These workers process all | |
2473 | * work items regardless of their specific target workqueue. The only | |
2474 | * exception is work items which belong to workqueues with a rescuer which | |
2475 | * will be explained in rescuer_thread(). | |
d185af30 YB |
2476 | * |
2477 | * Return: 0 | |
4690c4ab | 2478 | */ |
c34056a3 | 2479 | static int worker_thread(void *__worker) |
1da177e4 | 2480 | { |
c34056a3 | 2481 | struct worker *worker = __worker; |
bd7bdd43 | 2482 | struct worker_pool *pool = worker->pool; |
1da177e4 | 2483 | |
e22bee78 | 2484 | /* tell the scheduler that this is a workqueue worker */ |
197f6acc | 2485 | set_pf_worker(true); |
c8e55f36 | 2486 | woke_up: |
a9b8a985 | 2487 | raw_spin_lock_irq(&pool->lock); |
1da177e4 | 2488 | |
a9ab775b TH |
2489 | /* am I supposed to die? */ |
2490 | if (unlikely(worker->flags & WORKER_DIE)) { | |
a9b8a985 | 2491 | raw_spin_unlock_irq(&pool->lock); |
197f6acc | 2492 | set_pf_worker(false); |
60f5a4bc LJ |
2493 | |
2494 | set_task_comm(worker->task, "kworker/dying"); | |
e441b56f | 2495 | ida_free(&pool->worker_ida, worker->id); |
a2d812a2 | 2496 | worker_detach_from_pool(worker); |
e02b9312 | 2497 | WARN_ON_ONCE(!list_empty(&worker->entry)); |
60f5a4bc | 2498 | kfree(worker); |
a9ab775b | 2499 | return 0; |
c8e55f36 | 2500 | } |
affee4b2 | 2501 | |
c8e55f36 | 2502 | worker_leave_idle(worker); |
db7bccf4 | 2503 | recheck: |
e22bee78 | 2504 | /* no more worker necessary? */ |
63d95a91 | 2505 | if (!need_more_worker(pool)) |
e22bee78 TH |
2506 | goto sleep; |
2507 | ||
2508 | /* do we need to manage? */ | |
63d95a91 | 2509 | if (unlikely(!may_start_working(pool)) && manage_workers(worker)) |
e22bee78 TH |
2510 | goto recheck; |
2511 | ||
c8e55f36 TH |
2512 | /* |
2513 | * ->scheduled list can only be filled while a worker is | |
2514 | * preparing to process a work or actually processing it. | |
2515 | * Make sure nobody diddled with it while I was sleeping. | |
2516 | */ | |
6183c009 | 2517 | WARN_ON_ONCE(!list_empty(&worker->scheduled)); |
c8e55f36 | 2518 | |
e22bee78 | 2519 | /* |
a9ab775b TH |
2520 | * Finish PREP stage. We're guaranteed to have at least one idle |
2521 | * worker or that someone else has already assumed the manager | |
2522 | * role. This is where @worker starts participating in concurrency | |
2523 | * management if applicable and concurrency management is restored | |
2524 | * after being rebound. See rebind_workers() for details. | |
e22bee78 | 2525 | */ |
a9ab775b | 2526 | worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); |
e22bee78 TH |
2527 | |
2528 | do { | |
c8e55f36 | 2529 | struct work_struct *work = |
bd7bdd43 | 2530 | list_first_entry(&pool->worklist, |
c8e55f36 TH |
2531 | struct work_struct, entry); |
2532 | ||
82607adc TH |
2533 | pool->watchdog_ts = jiffies; |
2534 | ||
c8e55f36 TH |
2535 | if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { |
2536 | /* optimization path, not strictly necessary */ | |
2537 | process_one_work(worker, work); | |
2538 | if (unlikely(!list_empty(&worker->scheduled))) | |
affee4b2 | 2539 | process_scheduled_works(worker); |
c8e55f36 TH |
2540 | } else { |
2541 | move_linked_works(work, &worker->scheduled, NULL); | |
2542 | process_scheduled_works(worker); | |
affee4b2 | 2543 | } |
63d95a91 | 2544 | } while (keep_working(pool)); |
e22bee78 | 2545 | |
228f1d00 | 2546 | worker_set_flags(worker, WORKER_PREP); |
d313dd85 | 2547 | sleep: |
c8e55f36 | 2548 | /* |
d565ed63 TH |
2549 | * pool->lock is held and there's no work to process and no need to |
2550 | * manage, sleep. Workers are woken up only while holding | |
2551 | * pool->lock or from local cpu, so setting the current state | |
2552 | * before releasing pool->lock is enough to prevent losing any | |
2553 | * event. | |
c8e55f36 TH |
2554 | */ |
2555 | worker_enter_idle(worker); | |
c5a94a61 | 2556 | __set_current_state(TASK_IDLE); |
a9b8a985 | 2557 | raw_spin_unlock_irq(&pool->lock); |
c8e55f36 TH |
2558 | schedule(); |
2559 | goto woke_up; | |
1da177e4 LT |
2560 | } |
2561 | ||
e22bee78 TH |
2562 | /** |
2563 | * rescuer_thread - the rescuer thread function | |
111c225a | 2564 | * @__rescuer: self |
e22bee78 TH |
2565 | * |
2566 | * Workqueue rescuer thread function. There's one rescuer for each | |
493008a8 | 2567 | * workqueue which has WQ_MEM_RECLAIM set. |
e22bee78 | 2568 | * |
706026c2 | 2569 | * Regular work processing on a pool may block trying to create a new |
e22bee78 TH |
2570 | * worker which uses GFP_KERNEL allocation which has slight chance of |
2571 | * developing into deadlock if some works currently on the same queue | |
2572 | * need to be processed to satisfy the GFP_KERNEL allocation. This is | |
2573 | * the problem rescuer solves. | |
2574 | * | |
706026c2 TH |
2575 | * When such condition is possible, the pool summons rescuers of all |
2576 | * workqueues which have works queued on the pool and let them process | |
e22bee78 TH |
2577 | * those works so that forward progress can be guaranteed. |
2578 | * | |
2579 | * This should happen rarely. | |
d185af30 YB |
2580 | * |
2581 | * Return: 0 | |
e22bee78 | 2582 | */ |
111c225a | 2583 | static int rescuer_thread(void *__rescuer) |
e22bee78 | 2584 | { |
111c225a TH |
2585 | struct worker *rescuer = __rescuer; |
2586 | struct workqueue_struct *wq = rescuer->rescue_wq; | |
e22bee78 | 2587 | struct list_head *scheduled = &rescuer->scheduled; |
4d595b86 | 2588 | bool should_stop; |
e22bee78 TH |
2589 | |
2590 | set_user_nice(current, RESCUER_NICE_LEVEL); | |
111c225a TH |
2591 | |
2592 | /* | |
2593 | * Mark rescuer as worker too. As WORKER_PREP is never cleared, it | |
2594 | * doesn't participate in concurrency management. | |
2595 | */ | |
197f6acc | 2596 | set_pf_worker(true); |
e22bee78 | 2597 | repeat: |
c5a94a61 | 2598 | set_current_state(TASK_IDLE); |
e22bee78 | 2599 | |
4d595b86 LJ |
2600 | /* |
2601 | * By the time the rescuer is requested to stop, the workqueue | |
2602 | * shouldn't have any work pending, but @wq->maydays may still have | |
2603 | * pwq(s) queued. This can happen by non-rescuer workers consuming | |
2604 | * all the work items before the rescuer got to them. Go through | |
2605 | * @wq->maydays processing before acting on should_stop so that the | |
2606 | * list is always empty on exit. | |
2607 | */ | |
2608 | should_stop = kthread_should_stop(); | |
e22bee78 | 2609 | |
493a1724 | 2610 | /* see whether any pwq is asking for help */ |
a9b8a985 | 2611 | raw_spin_lock_irq(&wq_mayday_lock); |
493a1724 TH |
2612 | |
2613 | while (!list_empty(&wq->maydays)) { | |
2614 | struct pool_workqueue *pwq = list_first_entry(&wq->maydays, | |
2615 | struct pool_workqueue, mayday_node); | |
112202d9 | 2616 | struct worker_pool *pool = pwq->pool; |
e22bee78 | 2617 | struct work_struct *work, *n; |
82607adc | 2618 | bool first = true; |
e22bee78 TH |
2619 | |
2620 | __set_current_state(TASK_RUNNING); | |
493a1724 TH |
2621 | list_del_init(&pwq->mayday_node); |
2622 | ||
a9b8a985 | 2623 | raw_spin_unlock_irq(&wq_mayday_lock); |
e22bee78 | 2624 | |
51697d39 LJ |
2625 | worker_attach_to_pool(rescuer, pool); |
2626 | ||
a9b8a985 | 2627 | raw_spin_lock_irq(&pool->lock); |
e22bee78 TH |
2628 | |
2629 | /* | |
2630 | * Slurp in all works issued via this workqueue and | |
2631 | * process'em. | |
2632 | */ | |
0479c8c5 | 2633 | WARN_ON_ONCE(!list_empty(scheduled)); |
82607adc TH |
2634 | list_for_each_entry_safe(work, n, &pool->worklist, entry) { |
2635 | if (get_work_pwq(work) == pwq) { | |
2636 | if (first) | |
2637 | pool->watchdog_ts = jiffies; | |
e22bee78 | 2638 | move_linked_works(work, scheduled, &n); |
82607adc TH |
2639 | } |
2640 | first = false; | |
2641 | } | |
e22bee78 | 2642 | |
008847f6 N |
2643 | if (!list_empty(scheduled)) { |
2644 | process_scheduled_works(rescuer); | |
2645 | ||
2646 | /* | |
2647 | * The above execution of rescued work items could | |
2648 | * have created more to rescue through | |
f97a4a1a | 2649 | * pwq_activate_first_inactive() or chained |
008847f6 N |
2650 | * queueing. Let's put @pwq back on mayday list so |
2651 | * that such back-to-back work items, which may be | |
2652 | * being used to relieve memory pressure, don't | |
2653 | * incur MAYDAY_INTERVAL delay inbetween. | |
2654 | */ | |
4f3f4cf3 | 2655 | if (pwq->nr_active && need_to_create_worker(pool)) { |
a9b8a985 | 2656 | raw_spin_lock(&wq_mayday_lock); |
e66b39af TH |
2657 | /* |
2658 | * Queue iff we aren't racing destruction | |
2659 | * and somebody else hasn't queued it already. | |
2660 | */ | |
2661 | if (wq->rescuer && list_empty(&pwq->mayday_node)) { | |
2662 | get_pwq(pwq); | |
2663 | list_add_tail(&pwq->mayday_node, &wq->maydays); | |
2664 | } | |
a9b8a985 | 2665 | raw_spin_unlock(&wq_mayday_lock); |
008847f6 N |
2666 | } |
2667 | } | |
7576958a | 2668 | |
77668c8b LJ |
2669 | /* |
2670 | * Put the reference grabbed by send_mayday(). @pool won't | |
13b1d625 | 2671 | * go away while we're still attached to it. |
77668c8b LJ |
2672 | */ |
2673 | put_pwq(pwq); | |
2674 | ||
7576958a | 2675 | /* |
d8ca83e6 | 2676 | * Leave this pool. If need_more_worker() is %true, notify a |
7576958a TH |
2677 | * regular worker; otherwise, we end up with 0 concurrency |
2678 | * and stalling the execution. | |
2679 | */ | |
d8ca83e6 | 2680 | if (need_more_worker(pool)) |
63d95a91 | 2681 | wake_up_worker(pool); |
7576958a | 2682 | |
a9b8a985 | 2683 | raw_spin_unlock_irq(&pool->lock); |
13b1d625 | 2684 | |
a2d812a2 | 2685 | worker_detach_from_pool(rescuer); |
13b1d625 | 2686 | |
a9b8a985 | 2687 | raw_spin_lock_irq(&wq_mayday_lock); |
e22bee78 TH |
2688 | } |
2689 | ||
a9b8a985 | 2690 | raw_spin_unlock_irq(&wq_mayday_lock); |
493a1724 | 2691 | |
4d595b86 LJ |
2692 | if (should_stop) { |
2693 | __set_current_state(TASK_RUNNING); | |
197f6acc | 2694 | set_pf_worker(false); |
4d595b86 LJ |
2695 | return 0; |
2696 | } | |
2697 | ||
111c225a TH |
2698 | /* rescuers should never participate in concurrency management */ |
2699 | WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); | |
e22bee78 TH |
2700 | schedule(); |
2701 | goto repeat; | |
1da177e4 LT |
2702 | } |
2703 | ||
fca839c0 TH |
2704 | /** |
2705 | * check_flush_dependency - check for flush dependency sanity | |
2706 | * @target_wq: workqueue being flushed | |
2707 | * @target_work: work item being flushed (NULL for workqueue flushes) | |
2708 | * | |
2709 | * %current is trying to flush the whole @target_wq or @target_work on it. | |
2710 | * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not | |
2711 | * reclaiming memory or running on a workqueue which doesn't have | |
2712 | * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to | |
2713 | * a deadlock. | |
2714 | */ | |
2715 | static void check_flush_dependency(struct workqueue_struct *target_wq, | |
2716 | struct work_struct *target_work) | |
2717 | { | |
2718 | work_func_t target_func = target_work ? target_work->func : NULL; | |
2719 | struct worker *worker; | |
2720 | ||
2721 | if (target_wq->flags & WQ_MEM_RECLAIM) | |
2722 | return; | |
2723 | ||
2724 | worker = current_wq_worker(); | |
2725 | ||
2726 | WARN_ONCE(current->flags & PF_MEMALLOC, | |
d75f773c | 2727 | "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps", |
fca839c0 | 2728 | current->pid, current->comm, target_wq->name, target_func); |
23d11a58 TH |
2729 | WARN_ONCE(worker && ((worker->current_pwq->wq->flags & |
2730 | (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM), | |
d75f773c | 2731 | "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps", |
fca839c0 TH |
2732 | worker->current_pwq->wq->name, worker->current_func, |
2733 | target_wq->name, target_func); | |
2734 | } | |
2735 | ||
fc2e4d70 ON |
2736 | struct wq_barrier { |
2737 | struct work_struct work; | |
2738 | struct completion done; | |
2607d7a6 | 2739 | struct task_struct *task; /* purely informational */ |
fc2e4d70 ON |
2740 | }; |
2741 | ||
2742 | static void wq_barrier_func(struct work_struct *work) | |
2743 | { | |
2744 | struct wq_barrier *barr = container_of(work, struct wq_barrier, work); | |
2745 | complete(&barr->done); | |
2746 | } | |
2747 | ||
4690c4ab TH |
2748 | /** |
2749 | * insert_wq_barrier - insert a barrier work | |
112202d9 | 2750 | * @pwq: pwq to insert barrier into |
4690c4ab | 2751 | * @barr: wq_barrier to insert |
affee4b2 TH |
2752 | * @target: target work to attach @barr to |
2753 | * @worker: worker currently executing @target, NULL if @target is not executing | |
4690c4ab | 2754 | * |
affee4b2 TH |
2755 | * @barr is linked to @target such that @barr is completed only after |
2756 | * @target finishes execution. Please note that the ordering | |
2757 | * guarantee is observed only with respect to @target and on the local | |
2758 | * cpu. | |
2759 | * | |
2760 | * Currently, a queued barrier can't be canceled. This is because | |
2761 | * try_to_grab_pending() can't determine whether the work to be | |
2762 | * grabbed is at the head of the queue and thus can't clear LINKED | |
2763 | * flag of the previous work while there must be a valid next work | |
2764 | * after a work with LINKED flag set. | |
2765 | * | |
2766 | * Note that when @worker is non-NULL, @target may be modified | |
112202d9 | 2767 | * underneath us, so we can't reliably determine pwq from @target. |
4690c4ab TH |
2768 | * |
2769 | * CONTEXT: | |
a9b8a985 | 2770 | * raw_spin_lock_irq(pool->lock). |
4690c4ab | 2771 | */ |
112202d9 | 2772 | static void insert_wq_barrier(struct pool_workqueue *pwq, |
affee4b2 TH |
2773 | struct wq_barrier *barr, |
2774 | struct work_struct *target, struct worker *worker) | |
fc2e4d70 | 2775 | { |
d812796e LJ |
2776 | unsigned int work_flags = 0; |
2777 | unsigned int work_color; | |
affee4b2 | 2778 | struct list_head *head; |
affee4b2 | 2779 | |
dc186ad7 | 2780 | /* |
d565ed63 | 2781 | * debugobject calls are safe here even with pool->lock locked |
dc186ad7 TG |
2782 | * as we know for sure that this will not trigger any of the |
2783 | * checks and call back into the fixup functions where we | |
2784 | * might deadlock. | |
2785 | */ | |
ca1cab37 | 2786 | INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); |
22df02bb | 2787 | __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); |
52fa5bc5 | 2788 | |
fd1a5b04 BP |
2789 | init_completion_map(&barr->done, &target->lockdep_map); |
2790 | ||
2607d7a6 | 2791 | barr->task = current; |
83c22520 | 2792 | |
018f3a13 LJ |
2793 | /* The barrier work item does not participate in pwq->nr_active. */ |
2794 | work_flags |= WORK_STRUCT_INACTIVE; | |
2795 | ||
affee4b2 TH |
2796 | /* |
2797 | * If @target is currently being executed, schedule the | |
2798 | * barrier to the worker; otherwise, put it after @target. | |
2799 | */ | |
d812796e | 2800 | if (worker) { |
affee4b2 | 2801 | head = worker->scheduled.next; |
d812796e LJ |
2802 | work_color = worker->current_color; |
2803 | } else { | |
affee4b2 TH |
2804 | unsigned long *bits = work_data_bits(target); |
2805 | ||
2806 | head = target->entry.next; | |
2807 | /* there can already be other linked works, inherit and set */ | |
d21cece0 | 2808 | work_flags |= *bits & WORK_STRUCT_LINKED; |
d812796e | 2809 | work_color = get_work_color(*bits); |
affee4b2 TH |
2810 | __set_bit(WORK_STRUCT_LINKED_BIT, bits); |
2811 | } | |
2812 | ||
d812796e LJ |
2813 | pwq->nr_in_flight[work_color]++; |
2814 | work_flags |= work_color_to_flags(work_color); | |
2815 | ||
dc186ad7 | 2816 | debug_work_activate(&barr->work); |
d21cece0 | 2817 | insert_work(pwq, &barr->work, head, work_flags); |
fc2e4d70 ON |
2818 | } |
2819 | ||
73f53c4a | 2820 | /** |
112202d9 | 2821 | * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing |
73f53c4a TH |
2822 | * @wq: workqueue being flushed |
2823 | * @flush_color: new flush color, < 0 for no-op | |
2824 | * @work_color: new work color, < 0 for no-op | |
2825 | * | |
112202d9 | 2826 | * Prepare pwqs for workqueue flushing. |
73f53c4a | 2827 | * |
112202d9 TH |
2828 | * If @flush_color is non-negative, flush_color on all pwqs should be |
2829 | * -1. If no pwq has in-flight commands at the specified color, all | |
2830 | * pwq->flush_color's stay at -1 and %false is returned. If any pwq | |
2831 | * has in flight commands, its pwq->flush_color is set to | |
2832 | * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq | |
73f53c4a TH |
2833 | * wakeup logic is armed and %true is returned. |
2834 | * | |
2835 | * The caller should have initialized @wq->first_flusher prior to | |
2836 | * calling this function with non-negative @flush_color. If | |
2837 | * @flush_color is negative, no flush color update is done and %false | |
2838 | * is returned. | |
2839 | * | |
112202d9 | 2840 | * If @work_color is non-negative, all pwqs should have the same |
73f53c4a TH |
2841 | * work_color which is previous to @work_color and all will be |
2842 | * advanced to @work_color. | |
2843 | * | |
2844 | * CONTEXT: | |
3c25a55d | 2845 | * mutex_lock(wq->mutex). |
73f53c4a | 2846 | * |
d185af30 | 2847 | * Return: |
73f53c4a TH |
2848 | * %true if @flush_color >= 0 and there's something to flush. %false |
2849 | * otherwise. | |
2850 | */ | |
112202d9 | 2851 | static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, |
73f53c4a | 2852 | int flush_color, int work_color) |
1da177e4 | 2853 | { |
73f53c4a | 2854 | bool wait = false; |
49e3cf44 | 2855 | struct pool_workqueue *pwq; |
1da177e4 | 2856 | |
73f53c4a | 2857 | if (flush_color >= 0) { |
6183c009 | 2858 | WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); |
112202d9 | 2859 | atomic_set(&wq->nr_pwqs_to_flush, 1); |
1da177e4 | 2860 | } |
2355b70f | 2861 | |
49e3cf44 | 2862 | for_each_pwq(pwq, wq) { |
112202d9 | 2863 | struct worker_pool *pool = pwq->pool; |
fc2e4d70 | 2864 | |
a9b8a985 | 2865 | raw_spin_lock_irq(&pool->lock); |
83c22520 | 2866 | |
73f53c4a | 2867 | if (flush_color >= 0) { |
6183c009 | 2868 | WARN_ON_ONCE(pwq->flush_color != -1); |
fc2e4d70 | 2869 | |
112202d9 TH |
2870 | if (pwq->nr_in_flight[flush_color]) { |
2871 | pwq->flush_color = flush_color; | |
2872 | atomic_inc(&wq->nr_pwqs_to_flush); | |
73f53c4a TH |
2873 | wait = true; |
2874 | } | |
2875 | } | |
1da177e4 | 2876 | |
73f53c4a | 2877 | if (work_color >= 0) { |
6183c009 | 2878 | WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); |
112202d9 | 2879 | pwq->work_color = work_color; |
73f53c4a | 2880 | } |
1da177e4 | 2881 | |
a9b8a985 | 2882 | raw_spin_unlock_irq(&pool->lock); |
1da177e4 | 2883 | } |
2355b70f | 2884 | |
112202d9 | 2885 | if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) |
73f53c4a | 2886 | complete(&wq->first_flusher->done); |
14441960 | 2887 | |
73f53c4a | 2888 | return wait; |
1da177e4 LT |
2889 | } |
2890 | ||
0fcb78c2 | 2891 | /** |
c4f135d6 | 2892 | * __flush_workqueue - ensure that any scheduled work has run to completion. |
0fcb78c2 | 2893 | * @wq: workqueue to flush |
1da177e4 | 2894 | * |
c5aa87bb TH |
2895 | * This function sleeps until all work items which were queued on entry |
2896 | * have finished execution, but it is not livelocked by new incoming ones. | |
1da177e4 | 2897 | */ |
c4f135d6 | 2898 | void __flush_workqueue(struct workqueue_struct *wq) |
1da177e4 | 2899 | { |
73f53c4a TH |
2900 | struct wq_flusher this_flusher = { |
2901 | .list = LIST_HEAD_INIT(this_flusher.list), | |
2902 | .flush_color = -1, | |
fd1a5b04 | 2903 | .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map), |
73f53c4a TH |
2904 | }; |
2905 | int next_color; | |
1da177e4 | 2906 | |
3347fa09 TH |
2907 | if (WARN_ON(!wq_online)) |
2908 | return; | |
2909 | ||
87915adc JB |
2910 | lock_map_acquire(&wq->lockdep_map); |
2911 | lock_map_release(&wq->lockdep_map); | |
2912 | ||
3c25a55d | 2913 | mutex_lock(&wq->mutex); |
73f53c4a TH |
2914 | |
2915 | /* | |
2916 | * Start-to-wait phase | |
2917 | */ | |
2918 | next_color = work_next_color(wq->work_color); | |
2919 | ||
2920 | if (next_color != wq->flush_color) { | |
2921 | /* | |
2922 | * Color space is not full. The current work_color | |
2923 | * becomes our flush_color and work_color is advanced | |
2924 | * by one. | |
2925 | */ | |
6183c009 | 2926 | WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); |
73f53c4a TH |
2927 | this_flusher.flush_color = wq->work_color; |
2928 | wq->work_color = next_color; | |
2929 | ||
2930 | if (!wq->first_flusher) { | |
2931 | /* no flush in progress, become the first flusher */ | |
6183c009 | 2932 | WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); |
73f53c4a TH |
2933 | |
2934 | wq->first_flusher = &this_flusher; | |
2935 | ||
112202d9 | 2936 | if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, |
73f53c4a TH |
2937 | wq->work_color)) { |
2938 | /* nothing to flush, done */ | |
2939 | wq->flush_color = next_color; | |
2940 | wq->first_flusher = NULL; | |
2941 | goto out_unlock; | |
2942 | } | |
2943 | } else { | |
2944 | /* wait in queue */ | |
6183c009 | 2945 | WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); |
73f53c4a | 2946 | list_add_tail(&this_flusher.list, &wq->flusher_queue); |
112202d9 | 2947 | flush_workqueue_prep_pwqs(wq, -1, wq->work_color); |
73f53c4a TH |
2948 | } |
2949 | } else { | |
2950 | /* | |
2951 | * Oops, color space is full, wait on overflow queue. | |
2952 | * The next flush completion will assign us | |
2953 | * flush_color and transfer to flusher_queue. | |
2954 | */ | |
2955 | list_add_tail(&this_flusher.list, &wq->flusher_overflow); | |
2956 | } | |
2957 | ||
fca839c0 TH |
2958 | check_flush_dependency(wq, NULL); |
2959 | ||
3c25a55d | 2960 | mutex_unlock(&wq->mutex); |
73f53c4a TH |
2961 | |
2962 | wait_for_completion(&this_flusher.done); | |
2963 | ||
2964 | /* | |
2965 | * Wake-up-and-cascade phase | |
2966 | * | |
2967 | * First flushers are responsible for cascading flushes and | |
2968 | * handling overflow. Non-first flushers can simply return. | |
2969 | */ | |
00d5d15b | 2970 | if (READ_ONCE(wq->first_flusher) != &this_flusher) |
73f53c4a TH |
2971 | return; |
2972 | ||
3c25a55d | 2973 | mutex_lock(&wq->mutex); |
73f53c4a | 2974 | |
4ce48b37 TH |
2975 | /* we might have raced, check again with mutex held */ |
2976 | if (wq->first_flusher != &this_flusher) | |
2977 | goto out_unlock; | |
2978 | ||
00d5d15b | 2979 | WRITE_ONCE(wq->first_flusher, NULL); |
73f53c4a | 2980 | |
6183c009 TH |
2981 | WARN_ON_ONCE(!list_empty(&this_flusher.list)); |
2982 | WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); | |
73f53c4a TH |
2983 | |
2984 | while (true) { | |
2985 | struct wq_flusher *next, *tmp; | |
2986 | ||
2987 | /* complete all the flushers sharing the current flush color */ | |
2988 | list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { | |
2989 | if (next->flush_color != wq->flush_color) | |
2990 | break; | |
2991 | list_del_init(&next->list); | |
2992 | complete(&next->done); | |
2993 | } | |
2994 | ||
6183c009 TH |
2995 | WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && |
2996 | wq->flush_color != work_next_color(wq->work_color)); | |
73f53c4a TH |
2997 | |
2998 | /* this flush_color is finished, advance by one */ | |
2999 | wq->flush_color = work_next_color(wq->flush_color); | |
3000 | ||
3001 | /* one color has been freed, handle overflow queue */ | |
3002 | if (!list_empty(&wq->flusher_overflow)) { | |
3003 | /* | |
3004 | * Assign the same color to all overflowed | |
3005 | * flushers, advance work_color and append to | |
3006 | * flusher_queue. This is the start-to-wait | |
3007 | * phase for these overflowed flushers. | |
3008 | */ | |
3009 | list_for_each_entry(tmp, &wq->flusher_overflow, list) | |
3010 | tmp->flush_color = wq->work_color; | |
3011 | ||
3012 | wq->work_color = work_next_color(wq->work_color); | |
3013 | ||
3014 | list_splice_tail_init(&wq->flusher_overflow, | |
3015 | &wq->flusher_queue); | |
112202d9 | 3016 | flush_workqueue_prep_pwqs(wq, -1, wq->work_color); |
73f53c4a TH |
3017 | } |
3018 | ||
3019 | if (list_empty(&wq->flusher_queue)) { | |
6183c009 | 3020 | WARN_ON_ONCE(wq->flush_color != wq->work_color); |
73f53c4a TH |
3021 | break; |
3022 | } | |
3023 | ||
3024 | /* | |
3025 | * Need to flush more colors. Make the next flusher | |
112202d9 | 3026 | * the new first flusher and arm pwqs. |
73f53c4a | 3027 | */ |
6183c009 TH |
3028 | WARN_ON_ONCE(wq->flush_color == wq->work_color); |
3029 | WARN_ON_ONCE(wq->flush_color != next->flush_color); | |
73f53c4a TH |
3030 | |
3031 | list_del_init(&next->list); | |
3032 | wq->first_flusher = next; | |
3033 | ||
112202d9 | 3034 | if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) |
73f53c4a TH |
3035 | break; |
3036 | ||
3037 | /* | |
3038 | * Meh... this color is already done, clear first | |
3039 | * flusher and repeat cascading. | |
3040 | */ | |
3041 | wq->first_flusher = NULL; | |
3042 | } | |
3043 | ||
3044 | out_unlock: | |
3c25a55d | 3045 | mutex_unlock(&wq->mutex); |
1da177e4 | 3046 | } |
c4f135d6 | 3047 | EXPORT_SYMBOL(__flush_workqueue); |
1da177e4 | 3048 | |
9c5a2ba7 TH |
3049 | /** |
3050 | * drain_workqueue - drain a workqueue | |
3051 | * @wq: workqueue to drain | |
3052 | * | |
3053 | * Wait until the workqueue becomes empty. While draining is in progress, | |
3054 | * only chain queueing is allowed. IOW, only currently pending or running | |
3055 | * work items on @wq can queue further work items on it. @wq is flushed | |
b749b1b6 | 3056 | * repeatedly until it becomes empty. The number of flushing is determined |
9c5a2ba7 TH |
3057 | * by the depth of chaining and should be relatively short. Whine if it |
3058 | * takes too long. | |
3059 | */ | |
3060 | void drain_workqueue(struct workqueue_struct *wq) | |
3061 | { | |
3062 | unsigned int flush_cnt = 0; | |
49e3cf44 | 3063 | struct pool_workqueue *pwq; |
9c5a2ba7 TH |
3064 | |
3065 | /* | |
3066 | * __queue_work() needs to test whether there are drainers, is much | |
3067 | * hotter than drain_workqueue() and already looks at @wq->flags. | |
618b01eb | 3068 | * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. |
9c5a2ba7 | 3069 | */ |
87fc741e | 3070 | mutex_lock(&wq->mutex); |
9c5a2ba7 | 3071 | if (!wq->nr_drainers++) |
618b01eb | 3072 | wq->flags |= __WQ_DRAINING; |
87fc741e | 3073 | mutex_unlock(&wq->mutex); |
9c5a2ba7 | 3074 | reflush: |
c4f135d6 | 3075 | __flush_workqueue(wq); |
9c5a2ba7 | 3076 | |
b09f4fd3 | 3077 | mutex_lock(&wq->mutex); |
76af4d93 | 3078 | |
49e3cf44 | 3079 | for_each_pwq(pwq, wq) { |
fa2563e4 | 3080 | bool drained; |
9c5a2ba7 | 3081 | |
a9b8a985 | 3082 | raw_spin_lock_irq(&pwq->pool->lock); |
f97a4a1a | 3083 | drained = !pwq->nr_active && list_empty(&pwq->inactive_works); |
a9b8a985 | 3084 | raw_spin_unlock_irq(&pwq->pool->lock); |
fa2563e4 TT |
3085 | |
3086 | if (drained) | |
9c5a2ba7 TH |
3087 | continue; |
3088 | ||
3089 | if (++flush_cnt == 10 || | |
3090 | (flush_cnt % 100 == 0 && flush_cnt <= 1000)) | |
e9ad2eb3 SZ |
3091 | pr_warn("workqueue %s: %s() isn't complete after %u tries\n", |
3092 | wq->name, __func__, flush_cnt); | |
76af4d93 | 3093 | |
b09f4fd3 | 3094 | mutex_unlock(&wq->mutex); |
9c5a2ba7 TH |
3095 | goto reflush; |
3096 | } | |
3097 | ||
9c5a2ba7 | 3098 | if (!--wq->nr_drainers) |
618b01eb | 3099 | wq->flags &= ~__WQ_DRAINING; |
87fc741e | 3100 | mutex_unlock(&wq->mutex); |
9c5a2ba7 TH |
3101 | } |
3102 | EXPORT_SYMBOL_GPL(drain_workqueue); | |
3103 | ||
d6e89786 JB |
3104 | static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr, |
3105 | bool from_cancel) | |
db700897 | 3106 | { |
affee4b2 | 3107 | struct worker *worker = NULL; |
c9e7cf27 | 3108 | struct worker_pool *pool; |
112202d9 | 3109 | struct pool_workqueue *pwq; |
db700897 ON |
3110 | |
3111 | might_sleep(); | |
fa1b54e6 | 3112 | |
24acfb71 | 3113 | rcu_read_lock(); |
c9e7cf27 | 3114 | pool = get_work_pool(work); |
fa1b54e6 | 3115 | if (!pool) { |
24acfb71 | 3116 | rcu_read_unlock(); |
baf59022 | 3117 | return false; |
fa1b54e6 | 3118 | } |
db700897 | 3119 | |
a9b8a985 | 3120 | raw_spin_lock_irq(&pool->lock); |
0b3dae68 | 3121 | /* see the comment in try_to_grab_pending() with the same code */ |
112202d9 TH |
3122 | pwq = get_work_pwq(work); |
3123 | if (pwq) { | |
3124 | if (unlikely(pwq->pool != pool)) | |
4690c4ab | 3125 | goto already_gone; |
606a5020 | 3126 | } else { |
c9e7cf27 | 3127 | worker = find_worker_executing_work(pool, work); |
affee4b2 | 3128 | if (!worker) |
4690c4ab | 3129 | goto already_gone; |
112202d9 | 3130 | pwq = worker->current_pwq; |
606a5020 | 3131 | } |
db700897 | 3132 | |
fca839c0 TH |
3133 | check_flush_dependency(pwq->wq, work); |
3134 | ||
112202d9 | 3135 | insert_wq_barrier(pwq, barr, work, worker); |
a9b8a985 | 3136 | raw_spin_unlock_irq(&pool->lock); |
7a22ad75 | 3137 | |
e159489b | 3138 | /* |
a1d14934 PZ |
3139 | * Force a lock recursion deadlock when using flush_work() inside a |
3140 | * single-threaded or rescuer equipped workqueue. | |
3141 | * | |
3142 | * For single threaded workqueues the deadlock happens when the work | |
3143 | * is after the work issuing the flush_work(). For rescuer equipped | |
3144 | * workqueues the deadlock happens when the rescuer stalls, blocking | |
3145 | * forward progress. | |
e159489b | 3146 | */ |
d6e89786 JB |
3147 | if (!from_cancel && |
3148 | (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) { | |
112202d9 | 3149 | lock_map_acquire(&pwq->wq->lockdep_map); |
a1d14934 PZ |
3150 | lock_map_release(&pwq->wq->lockdep_map); |
3151 | } | |
24acfb71 | 3152 | rcu_read_unlock(); |
401a8d04 | 3153 | return true; |
4690c4ab | 3154 | already_gone: |
a9b8a985 | 3155 | raw_spin_unlock_irq(&pool->lock); |
24acfb71 | 3156 | rcu_read_unlock(); |
401a8d04 | 3157 | return false; |
db700897 | 3158 | } |
baf59022 | 3159 | |
d6e89786 JB |
3160 | static bool __flush_work(struct work_struct *work, bool from_cancel) |
3161 | { | |
3162 | struct wq_barrier barr; | |
3163 | ||
3164 | if (WARN_ON(!wq_online)) | |
3165 | return false; | |
3166 | ||
4d43d395 TH |
3167 | if (WARN_ON(!work->func)) |
3168 | return false; | |
3169 | ||
c0feea59 TH |
3170 | lock_map_acquire(&work->lockdep_map); |
3171 | lock_map_release(&work->lockdep_map); | |
87915adc | 3172 | |
d6e89786 JB |
3173 | if (start_flush_work(work, &barr, from_cancel)) { |
3174 | wait_for_completion(&barr.done); | |
3175 | destroy_work_on_stack(&barr.work); | |
3176 | return true; | |
3177 | } else { | |
3178 | return false; | |
3179 | } | |
3180 | } | |
3181 | ||
baf59022 TH |
3182 | /** |
3183 | * flush_work - wait for a work to finish executing the last queueing instance | |
3184 | * @work: the work to flush | |
3185 | * | |
606a5020 TH |
3186 | * Wait until @work has finished execution. @work is guaranteed to be idle |
3187 | * on return if it hasn't been requeued since flush started. | |
baf59022 | 3188 | * |
d185af30 | 3189 | * Return: |
baf59022 TH |
3190 | * %true if flush_work() waited for the work to finish execution, |
3191 | * %false if it was already idle. | |
3192 | */ | |
3193 | bool flush_work(struct work_struct *work) | |
3194 | { | |
d6e89786 | 3195 | return __flush_work(work, false); |
6e84d644 | 3196 | } |
606a5020 | 3197 | EXPORT_SYMBOL_GPL(flush_work); |
6e84d644 | 3198 | |
8603e1b3 | 3199 | struct cwt_wait { |
ac6424b9 | 3200 | wait_queue_entry_t wait; |
8603e1b3 TH |
3201 | struct work_struct *work; |
3202 | }; | |
3203 | ||
ac6424b9 | 3204 | static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key) |
8603e1b3 TH |
3205 | { |
3206 | struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait); | |
3207 | ||
3208 | if (cwait->work != key) | |
3209 | return 0; | |
3210 | return autoremove_wake_function(wait, mode, sync, key); | |
3211 | } | |
3212 | ||
36e227d2 | 3213 | static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) |
1f1f642e | 3214 | { |
8603e1b3 | 3215 | static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq); |
bbb68dfa | 3216 | unsigned long flags; |
1f1f642e ON |
3217 | int ret; |
3218 | ||
3219 | do { | |
bbb68dfa TH |
3220 | ret = try_to_grab_pending(work, is_dwork, &flags); |
3221 | /* | |
8603e1b3 TH |
3222 | * If someone else is already canceling, wait for it to |
3223 | * finish. flush_work() doesn't work for PREEMPT_NONE | |
3224 | * because we may get scheduled between @work's completion | |
3225 | * and the other canceling task resuming and clearing | |
3226 | * CANCELING - flush_work() will return false immediately | |
3227 | * as @work is no longer busy, try_to_grab_pending() will | |
3228 | * return -ENOENT as @work is still being canceled and the | |
3229 | * other canceling task won't be able to clear CANCELING as | |
3230 | * we're hogging the CPU. | |
3231 | * | |
3232 | * Let's wait for completion using a waitqueue. As this | |
3233 | * may lead to the thundering herd problem, use a custom | |
3234 | * wake function which matches @work along with exclusive | |
3235 | * wait and wakeup. | |
bbb68dfa | 3236 | */ |
8603e1b3 TH |
3237 | if (unlikely(ret == -ENOENT)) { |
3238 | struct cwt_wait cwait; | |
3239 | ||
3240 | init_wait(&cwait.wait); | |
3241 | cwait.wait.func = cwt_wakefn; | |
3242 | cwait.work = work; | |
3243 | ||
3244 | prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait, | |
3245 | TASK_UNINTERRUPTIBLE); | |
3246 | if (work_is_canceling(work)) | |
3247 | schedule(); | |
3248 | finish_wait(&cancel_waitq, &cwait.wait); | |
3249 | } | |
1f1f642e ON |
3250 | } while (unlikely(ret < 0)); |
3251 | ||
bbb68dfa TH |
3252 | /* tell other tasks trying to grab @work to back off */ |
3253 | mark_work_canceling(work); | |
3254 | local_irq_restore(flags); | |
3255 | ||
3347fa09 TH |
3256 | /* |
3257 | * This allows canceling during early boot. We know that @work | |
3258 | * isn't executing. | |
3259 | */ | |
3260 | if (wq_online) | |
d6e89786 | 3261 | __flush_work(work, true); |
3347fa09 | 3262 | |
7a22ad75 | 3263 | clear_work_data(work); |
8603e1b3 TH |
3264 | |
3265 | /* | |
3266 | * Paired with prepare_to_wait() above so that either | |
3267 | * waitqueue_active() is visible here or !work_is_canceling() is | |
3268 | * visible there. | |
3269 | */ | |
3270 | smp_mb(); | |
3271 | if (waitqueue_active(&cancel_waitq)) | |
3272 | __wake_up(&cancel_waitq, TASK_NORMAL, 1, work); | |
3273 | ||
1f1f642e ON |
3274 | return ret; |
3275 | } | |
3276 | ||
6e84d644 | 3277 | /** |
401a8d04 TH |
3278 | * cancel_work_sync - cancel a work and wait for it to finish |
3279 | * @work: the work to cancel | |
6e84d644 | 3280 | * |
401a8d04 TH |
3281 | * Cancel @work and wait for its execution to finish. This function |
3282 | * can be used even if the work re-queues itself or migrates to | |
3283 | * another workqueue. On return from this function, @work is | |
3284 | * guaranteed to be not pending or executing on any CPU. | |
1f1f642e | 3285 | * |
401a8d04 TH |
3286 | * cancel_work_sync(&delayed_work->work) must not be used for |
3287 | * delayed_work's. Use cancel_delayed_work_sync() instead. | |
6e84d644 | 3288 | * |
401a8d04 | 3289 | * The caller must ensure that the workqueue on which @work was last |
6e84d644 | 3290 | * queued can't be destroyed before this function returns. |
401a8d04 | 3291 | * |
d185af30 | 3292 | * Return: |
401a8d04 | 3293 | * %true if @work was pending, %false otherwise. |
6e84d644 | 3294 | */ |
401a8d04 | 3295 | bool cancel_work_sync(struct work_struct *work) |
6e84d644 | 3296 | { |
36e227d2 | 3297 | return __cancel_work_timer(work, false); |
b89deed3 | 3298 | } |
28e53bdd | 3299 | EXPORT_SYMBOL_GPL(cancel_work_sync); |
b89deed3 | 3300 | |
6e84d644 | 3301 | /** |
401a8d04 TH |
3302 | * flush_delayed_work - wait for a dwork to finish executing the last queueing |
3303 | * @dwork: the delayed work to flush | |
6e84d644 | 3304 | * |
401a8d04 TH |
3305 | * Delayed timer is cancelled and the pending work is queued for |
3306 | * immediate execution. Like flush_work(), this function only | |
3307 | * considers the last queueing instance of @dwork. | |
1f1f642e | 3308 | * |
d185af30 | 3309 | * Return: |
401a8d04 TH |
3310 | * %true if flush_work() waited for the work to finish execution, |
3311 | * %false if it was already idle. | |
6e84d644 | 3312 | */ |
401a8d04 TH |
3313 | bool flush_delayed_work(struct delayed_work *dwork) |
3314 | { | |
8930caba | 3315 | local_irq_disable(); |
401a8d04 | 3316 | if (del_timer_sync(&dwork->timer)) |
60c057bc | 3317 | __queue_work(dwork->cpu, dwork->wq, &dwork->work); |
8930caba | 3318 | local_irq_enable(); |
401a8d04 TH |
3319 | return flush_work(&dwork->work); |
3320 | } | |
3321 | EXPORT_SYMBOL(flush_delayed_work); | |
3322 | ||
05f0fe6b TH |
3323 | /** |
3324 | * flush_rcu_work - wait for a rwork to finish executing the last queueing | |
3325 | * @rwork: the rcu work to flush | |
3326 | * | |
3327 | * Return: | |
3328 | * %true if flush_rcu_work() waited for the work to finish execution, | |
3329 | * %false if it was already idle. | |
3330 | */ | |
3331 | bool flush_rcu_work(struct rcu_work *rwork) | |
3332 | { | |
3333 | if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) { | |
3334 | rcu_barrier(); | |
3335 | flush_work(&rwork->work); | |
3336 | return true; | |
3337 | } else { | |
3338 | return flush_work(&rwork->work); | |
3339 | } | |
3340 | } | |
3341 | EXPORT_SYMBOL(flush_rcu_work); | |
3342 | ||
f72b8792 JA |
3343 | static bool __cancel_work(struct work_struct *work, bool is_dwork) |
3344 | { | |
3345 | unsigned long flags; | |
3346 | int ret; | |
3347 | ||
3348 | do { | |
3349 | ret = try_to_grab_pending(work, is_dwork, &flags); | |
3350 | } while (unlikely(ret == -EAGAIN)); | |
3351 | ||
3352 | if (unlikely(ret < 0)) | |
3353 | return false; | |
3354 | ||
3355 | set_work_pool_and_clear_pending(work, get_work_pool_id(work)); | |
3356 | local_irq_restore(flags); | |
3357 | return ret; | |
3358 | } | |
3359 | ||
73b4b532 AG |
3360 | /* |
3361 | * See cancel_delayed_work() | |
3362 | */ | |
3363 | bool cancel_work(struct work_struct *work) | |
3364 | { | |
3365 | return __cancel_work(work, false); | |
3366 | } | |
3367 | EXPORT_SYMBOL(cancel_work); | |
3368 | ||
09383498 | 3369 | /** |
57b30ae7 TH |
3370 | * cancel_delayed_work - cancel a delayed work |
3371 | * @dwork: delayed_work to cancel | |
09383498 | 3372 | * |
d185af30 YB |
3373 | * Kill off a pending delayed_work. |
3374 | * | |
3375 | * Return: %true if @dwork was pending and canceled; %false if it wasn't | |
3376 | * pending. | |
3377 | * | |
3378 | * Note: | |
3379 | * The work callback function may still be running on return, unless | |
3380 | * it returns %true and the work doesn't re-arm itself. Explicitly flush or | |
3381 | * use cancel_delayed_work_sync() to wait on it. | |
09383498 | 3382 | * |
57b30ae7 | 3383 | * This function is safe to call from any context including IRQ handler. |
09383498 | 3384 | */ |
57b30ae7 | 3385 | bool cancel_delayed_work(struct delayed_work *dwork) |
09383498 | 3386 | { |
f72b8792 | 3387 | return __cancel_work(&dwork->work, true); |
09383498 | 3388 | } |
57b30ae7 | 3389 | EXPORT_SYMBOL(cancel_delayed_work); |
09383498 | 3390 | |
401a8d04 TH |
3391 | /** |
3392 | * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish | |
3393 | * @dwork: the delayed work cancel | |
3394 | * | |
3395 | * This is cancel_work_sync() for delayed works. | |
3396 | * | |
d185af30 | 3397 | * Return: |
401a8d04 TH |
3398 | * %true if @dwork was pending, %false otherwise. |
3399 | */ | |
3400 | bool cancel_delayed_work_sync(struct delayed_work *dwork) | |
6e84d644 | 3401 | { |
36e227d2 | 3402 | return __cancel_work_timer(&dwork->work, true); |
6e84d644 | 3403 | } |
f5a421a4 | 3404 | EXPORT_SYMBOL(cancel_delayed_work_sync); |
1da177e4 | 3405 | |
b6136773 | 3406 | /** |
31ddd871 | 3407 | * schedule_on_each_cpu - execute a function synchronously on each online CPU |
b6136773 | 3408 | * @func: the function to call |
b6136773 | 3409 | * |
31ddd871 TH |
3410 | * schedule_on_each_cpu() executes @func on each online CPU using the |
3411 | * system workqueue and blocks until all CPUs have completed. | |
b6136773 | 3412 | * schedule_on_each_cpu() is very slow. |
31ddd871 | 3413 | * |
d185af30 | 3414 | * Return: |
31ddd871 | 3415 | * 0 on success, -errno on failure. |
b6136773 | 3416 | */ |
65f27f38 | 3417 | int schedule_on_each_cpu(work_func_t func) |
15316ba8 CL |
3418 | { |
3419 | int cpu; | |
38f51568 | 3420 | struct work_struct __percpu *works; |
15316ba8 | 3421 | |
b6136773 AM |
3422 | works = alloc_percpu(struct work_struct); |
3423 | if (!works) | |
15316ba8 | 3424 | return -ENOMEM; |
b6136773 | 3425 | |
ffd8bea8 | 3426 | cpus_read_lock(); |
93981800 | 3427 | |
15316ba8 | 3428 | for_each_online_cpu(cpu) { |
9bfb1839 IM |
3429 | struct work_struct *work = per_cpu_ptr(works, cpu); |
3430 | ||
3431 | INIT_WORK(work, func); | |
b71ab8c2 | 3432 | schedule_work_on(cpu, work); |
65a64464 | 3433 | } |
93981800 TH |
3434 | |
3435 | for_each_online_cpu(cpu) | |
3436 | flush_work(per_cpu_ptr(works, cpu)); | |
3437 | ||
ffd8bea8 | 3438 | cpus_read_unlock(); |
b6136773 | 3439 | free_percpu(works); |
15316ba8 CL |
3440 | return 0; |
3441 | } | |
3442 | ||
1fa44eca JB |
3443 | /** |
3444 | * execute_in_process_context - reliably execute the routine with user context | |
3445 | * @fn: the function to execute | |
1fa44eca JB |
3446 | * @ew: guaranteed storage for the execute work structure (must |
3447 | * be available when the work executes) | |
3448 | * | |
3449 | * Executes the function immediately if process context is available, | |
3450 | * otherwise schedules the function for delayed execution. | |
3451 | * | |
d185af30 | 3452 | * Return: 0 - function was executed |
1fa44eca JB |
3453 | * 1 - function was scheduled for execution |
3454 | */ | |
65f27f38 | 3455 | int execute_in_process_context(work_func_t fn, struct execute_work *ew) |
1fa44eca JB |
3456 | { |
3457 | if (!in_interrupt()) { | |
65f27f38 | 3458 | fn(&ew->work); |
1fa44eca JB |
3459 | return 0; |
3460 | } | |
3461 | ||
65f27f38 | 3462 | INIT_WORK(&ew->work, fn); |
1fa44eca JB |
3463 | schedule_work(&ew->work); |
3464 | ||
3465 | return 1; | |
3466 | } | |
3467 | EXPORT_SYMBOL_GPL(execute_in_process_context); | |
3468 | ||
6ba94429 FW |
3469 | /** |
3470 | * free_workqueue_attrs - free a workqueue_attrs | |
3471 | * @attrs: workqueue_attrs to free | |
226223ab | 3472 | * |
6ba94429 | 3473 | * Undo alloc_workqueue_attrs(). |
226223ab | 3474 | */ |
513c98d0 | 3475 | void free_workqueue_attrs(struct workqueue_attrs *attrs) |
226223ab | 3476 | { |
6ba94429 FW |
3477 | if (attrs) { |
3478 | free_cpumask_var(attrs->cpumask); | |
3479 | kfree(attrs); | |
3480 | } | |
226223ab TH |
3481 | } |
3482 | ||
6ba94429 FW |
3483 | /** |
3484 | * alloc_workqueue_attrs - allocate a workqueue_attrs | |
6ba94429 FW |
3485 | * |
3486 | * Allocate a new workqueue_attrs, initialize with default settings and | |
3487 | * return it. | |
3488 | * | |
3489 | * Return: The allocated new workqueue_attr on success. %NULL on failure. | |
3490 | */ | |
513c98d0 | 3491 | struct workqueue_attrs *alloc_workqueue_attrs(void) |
226223ab | 3492 | { |
6ba94429 | 3493 | struct workqueue_attrs *attrs; |
226223ab | 3494 | |
be69d00d | 3495 | attrs = kzalloc(sizeof(*attrs), GFP_KERNEL); |
6ba94429 FW |
3496 | if (!attrs) |
3497 | goto fail; | |
be69d00d | 3498 | if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL)) |
6ba94429 FW |
3499 | goto fail; |
3500 | ||
3501 | cpumask_copy(attrs->cpumask, cpu_possible_mask); | |
3502 | return attrs; | |
3503 | fail: | |
3504 | free_workqueue_attrs(attrs); | |
3505 | return NULL; | |
226223ab TH |
3506 | } |
3507 | ||
6ba94429 FW |
3508 | static void copy_workqueue_attrs(struct workqueue_attrs *to, |
3509 | const struct workqueue_attrs *from) | |
226223ab | 3510 | { |
6ba94429 FW |
3511 | to->nice = from->nice; |
3512 | cpumask_copy(to->cpumask, from->cpumask); | |
3513 | /* | |
3514 | * Unlike hash and equality test, this function doesn't ignore | |
3515 | * ->no_numa as it is used for both pool and wq attrs. Instead, | |
3516 | * get_unbound_pool() explicitly clears ->no_numa after copying. | |
3517 | */ | |
3518 | to->no_numa = from->no_numa; | |
226223ab TH |
3519 | } |
3520 | ||
6ba94429 FW |
3521 | /* hash value of the content of @attr */ |
3522 | static u32 wqattrs_hash(const struct workqueue_attrs *attrs) | |
226223ab | 3523 | { |
6ba94429 | 3524 | u32 hash = 0; |
226223ab | 3525 | |
6ba94429 FW |
3526 | hash = jhash_1word(attrs->nice, hash); |
3527 | hash = jhash(cpumask_bits(attrs->cpumask), | |
3528 | BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); | |
3529 | return hash; | |
226223ab | 3530 | } |
226223ab | 3531 | |
6ba94429 FW |
3532 | /* content equality test */ |
3533 | static bool wqattrs_equal(const struct workqueue_attrs *a, | |
3534 | const struct workqueue_attrs *b) | |
226223ab | 3535 | { |
6ba94429 FW |
3536 | if (a->nice != b->nice) |
3537 | return false; | |
3538 | if (!cpumask_equal(a->cpumask, b->cpumask)) | |
3539 | return false; | |
3540 | return true; | |
226223ab TH |
3541 | } |
3542 | ||
6ba94429 FW |
3543 | /** |
3544 | * init_worker_pool - initialize a newly zalloc'd worker_pool | |
3545 | * @pool: worker_pool to initialize | |
3546 | * | |
402dd89d | 3547 | * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs. |
6ba94429 FW |
3548 | * |
3549 | * Return: 0 on success, -errno on failure. Even on failure, all fields | |
3550 | * inside @pool proper are initialized and put_unbound_pool() can be called | |
3551 | * on @pool safely to release it. | |
3552 | */ | |
3553 | static int init_worker_pool(struct worker_pool *pool) | |
226223ab | 3554 | { |
a9b8a985 | 3555 | raw_spin_lock_init(&pool->lock); |
6ba94429 FW |
3556 | pool->id = -1; |
3557 | pool->cpu = -1; | |
3558 | pool->node = NUMA_NO_NODE; | |
3559 | pool->flags |= POOL_DISASSOCIATED; | |
82607adc | 3560 | pool->watchdog_ts = jiffies; |
6ba94429 FW |
3561 | INIT_LIST_HEAD(&pool->worklist); |
3562 | INIT_LIST_HEAD(&pool->idle_list); | |
3563 | hash_init(pool->busy_hash); | |
226223ab | 3564 | |
32a6c723 | 3565 | timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE); |
3f959aa3 | 3566 | INIT_WORK(&pool->idle_cull_work, idle_cull_fn); |
226223ab | 3567 | |
32a6c723 | 3568 | timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0); |
226223ab | 3569 | |
6ba94429 | 3570 | INIT_LIST_HEAD(&pool->workers); |
e02b9312 | 3571 | INIT_LIST_HEAD(&pool->dying_workers); |
226223ab | 3572 | |
6ba94429 FW |
3573 | ida_init(&pool->worker_ida); |
3574 | INIT_HLIST_NODE(&pool->hash_node); | |
3575 | pool->refcnt = 1; | |
226223ab | 3576 | |
6ba94429 | 3577 | /* shouldn't fail above this point */ |
be69d00d | 3578 | pool->attrs = alloc_workqueue_attrs(); |
6ba94429 FW |
3579 | if (!pool->attrs) |
3580 | return -ENOMEM; | |
3581 | return 0; | |
226223ab TH |
3582 | } |
3583 | ||
669de8bd BVA |
3584 | #ifdef CONFIG_LOCKDEP |
3585 | static void wq_init_lockdep(struct workqueue_struct *wq) | |
3586 | { | |
3587 | char *lock_name; | |
3588 | ||
3589 | lockdep_register_key(&wq->key); | |
3590 | lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name); | |
3591 | if (!lock_name) | |
3592 | lock_name = wq->name; | |
69a106c0 QC |
3593 | |
3594 | wq->lock_name = lock_name; | |
669de8bd BVA |
3595 | lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0); |
3596 | } | |
3597 | ||
3598 | static void wq_unregister_lockdep(struct workqueue_struct *wq) | |
3599 | { | |
3600 | lockdep_unregister_key(&wq->key); | |
3601 | } | |
3602 | ||
3603 | static void wq_free_lockdep(struct workqueue_struct *wq) | |
3604 | { | |
3605 | if (wq->lock_name != wq->name) | |
3606 | kfree(wq->lock_name); | |
3607 | } | |
3608 | #else | |
3609 | static void wq_init_lockdep(struct workqueue_struct *wq) | |
3610 | { | |
3611 | } | |
3612 | ||
3613 | static void wq_unregister_lockdep(struct workqueue_struct *wq) | |
3614 | { | |
3615 | } | |
3616 | ||
3617 | static void wq_free_lockdep(struct workqueue_struct *wq) | |
3618 | { | |
3619 | } | |
3620 | #endif | |
3621 | ||
6ba94429 | 3622 | static void rcu_free_wq(struct rcu_head *rcu) |
226223ab | 3623 | { |
6ba94429 FW |
3624 | struct workqueue_struct *wq = |
3625 | container_of(rcu, struct workqueue_struct, rcu); | |
226223ab | 3626 | |
669de8bd BVA |
3627 | wq_free_lockdep(wq); |
3628 | ||
6ba94429 FW |
3629 | if (!(wq->flags & WQ_UNBOUND)) |
3630 | free_percpu(wq->cpu_pwqs); | |
226223ab | 3631 | else |
6ba94429 | 3632 | free_workqueue_attrs(wq->unbound_attrs); |
226223ab | 3633 | |
6ba94429 | 3634 | kfree(wq); |
226223ab TH |
3635 | } |
3636 | ||
6ba94429 | 3637 | static void rcu_free_pool(struct rcu_head *rcu) |
226223ab | 3638 | { |
6ba94429 | 3639 | struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); |
226223ab | 3640 | |
6ba94429 FW |
3641 | ida_destroy(&pool->worker_ida); |
3642 | free_workqueue_attrs(pool->attrs); | |
3643 | kfree(pool); | |
226223ab TH |
3644 | } |
3645 | ||
6ba94429 FW |
3646 | /** |
3647 | * put_unbound_pool - put a worker_pool | |
3648 | * @pool: worker_pool to put | |
3649 | * | |
24acfb71 | 3650 | * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU |
6ba94429 FW |
3651 | * safe manner. get_unbound_pool() calls this function on its failure path |
3652 | * and this function should be able to release pools which went through, | |
3653 | * successfully or not, init_worker_pool(). | |
3654 | * | |
3655 | * Should be called with wq_pool_mutex held. | |
3656 | */ | |
3657 | static void put_unbound_pool(struct worker_pool *pool) | |
226223ab | 3658 | { |
6ba94429 | 3659 | DECLARE_COMPLETION_ONSTACK(detach_completion); |
e02b9312 | 3660 | struct list_head cull_list; |
6ba94429 | 3661 | struct worker *worker; |
226223ab | 3662 | |
e02b9312 VS |
3663 | INIT_LIST_HEAD(&cull_list); |
3664 | ||
6ba94429 | 3665 | lockdep_assert_held(&wq_pool_mutex); |
226223ab | 3666 | |
6ba94429 FW |
3667 | if (--pool->refcnt) |
3668 | return; | |
226223ab | 3669 | |
6ba94429 FW |
3670 | /* sanity checks */ |
3671 | if (WARN_ON(!(pool->cpu < 0)) || | |
3672 | WARN_ON(!list_empty(&pool->worklist))) | |
3673 | return; | |
226223ab | 3674 | |
6ba94429 FW |
3675 | /* release id and unhash */ |
3676 | if (pool->id >= 0) | |
3677 | idr_remove(&worker_pool_idr, pool->id); | |
3678 | hash_del(&pool->hash_node); | |
d55262c4 | 3679 | |
6ba94429 | 3680 | /* |
692b4825 TH |
3681 | * Become the manager and destroy all workers. This prevents |
3682 | * @pool's workers from blocking on attach_mutex. We're the last | |
3683 | * manager and @pool gets freed with the flag set. | |
9ab03be4 VS |
3684 | * |
3685 | * Having a concurrent manager is quite unlikely to happen as we can | |
3686 | * only get here with | |
3687 | * pwq->refcnt == pool->refcnt == 0 | |
3688 | * which implies no work queued to the pool, which implies no worker can | |
3689 | * become the manager. However a worker could have taken the role of | |
3690 | * manager before the refcnts dropped to 0, since maybe_create_worker() | |
3691 | * drops pool->lock | |
6ba94429 | 3692 | */ |
9ab03be4 VS |
3693 | while (true) { |
3694 | rcuwait_wait_event(&manager_wait, | |
3695 | !(pool->flags & POOL_MANAGER_ACTIVE), | |
3696 | TASK_UNINTERRUPTIBLE); | |
e02b9312 VS |
3697 | |
3698 | mutex_lock(&wq_pool_attach_mutex); | |
9ab03be4 VS |
3699 | raw_spin_lock_irq(&pool->lock); |
3700 | if (!(pool->flags & POOL_MANAGER_ACTIVE)) { | |
3701 | pool->flags |= POOL_MANAGER_ACTIVE; | |
3702 | break; | |
3703 | } | |
3704 | raw_spin_unlock_irq(&pool->lock); | |
e02b9312 | 3705 | mutex_unlock(&wq_pool_attach_mutex); |
9ab03be4 | 3706 | } |
692b4825 | 3707 | |
6ba94429 | 3708 | while ((worker = first_idle_worker(pool))) |
e02b9312 | 3709 | set_worker_dying(worker, &cull_list); |
6ba94429 | 3710 | WARN_ON(pool->nr_workers || pool->nr_idle); |
a9b8a985 | 3711 | raw_spin_unlock_irq(&pool->lock); |
d55262c4 | 3712 | |
e02b9312 VS |
3713 | wake_dying_workers(&cull_list); |
3714 | ||
3715 | if (!list_empty(&pool->workers) || !list_empty(&pool->dying_workers)) | |
6ba94429 | 3716 | pool->detach_completion = &detach_completion; |
1258fae7 | 3717 | mutex_unlock(&wq_pool_attach_mutex); |
226223ab | 3718 | |
6ba94429 FW |
3719 | if (pool->detach_completion) |
3720 | wait_for_completion(pool->detach_completion); | |
226223ab | 3721 | |
6ba94429 FW |
3722 | /* shut down the timers */ |
3723 | del_timer_sync(&pool->idle_timer); | |
3f959aa3 | 3724 | cancel_work_sync(&pool->idle_cull_work); |
6ba94429 | 3725 | del_timer_sync(&pool->mayday_timer); |
226223ab | 3726 | |
24acfb71 | 3727 | /* RCU protected to allow dereferences from get_work_pool() */ |
25b00775 | 3728 | call_rcu(&pool->rcu, rcu_free_pool); |
226223ab TH |
3729 | } |
3730 | ||
3731 | /** | |
6ba94429 FW |
3732 | * get_unbound_pool - get a worker_pool with the specified attributes |
3733 | * @attrs: the attributes of the worker_pool to get | |
226223ab | 3734 | * |
6ba94429 FW |
3735 | * Obtain a worker_pool which has the same attributes as @attrs, bump the |
3736 | * reference count and return it. If there already is a matching | |
3737 | * worker_pool, it will be used; otherwise, this function attempts to | |
3738 | * create a new one. | |
226223ab | 3739 | * |
6ba94429 | 3740 | * Should be called with wq_pool_mutex held. |
226223ab | 3741 | * |
6ba94429 FW |
3742 | * Return: On success, a worker_pool with the same attributes as @attrs. |
3743 | * On failure, %NULL. | |
226223ab | 3744 | */ |
6ba94429 | 3745 | static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) |
226223ab | 3746 | { |
6ba94429 FW |
3747 | u32 hash = wqattrs_hash(attrs); |
3748 | struct worker_pool *pool; | |
3749 | int node; | |
e2273584 | 3750 | int target_node = NUMA_NO_NODE; |
226223ab | 3751 | |
6ba94429 | 3752 | lockdep_assert_held(&wq_pool_mutex); |
226223ab | 3753 | |
6ba94429 FW |
3754 | /* do we already have a matching pool? */ |
3755 | hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { | |
3756 | if (wqattrs_equal(pool->attrs, attrs)) { | |
3757 | pool->refcnt++; | |
3758 | return pool; | |
3759 | } | |
3760 | } | |
226223ab | 3761 | |
e2273584 XP |
3762 | /* if cpumask is contained inside a NUMA node, we belong to that node */ |
3763 | if (wq_numa_enabled) { | |
3764 | for_each_node(node) { | |
3765 | if (cpumask_subset(attrs->cpumask, | |
3766 | wq_numa_possible_cpumask[node])) { | |
3767 | target_node = node; | |
3768 | break; | |
3769 | } | |
3770 | } | |
3771 | } | |
3772 | ||
6ba94429 | 3773 | /* nope, create a new one */ |
e2273584 | 3774 | pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node); |
6ba94429 FW |
3775 | if (!pool || init_worker_pool(pool) < 0) |
3776 | goto fail; | |
3777 | ||
3778 | lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */ | |
3779 | copy_workqueue_attrs(pool->attrs, attrs); | |
e2273584 | 3780 | pool->node = target_node; |
226223ab TH |
3781 | |
3782 | /* | |
6ba94429 FW |
3783 | * no_numa isn't a worker_pool attribute, always clear it. See |
3784 | * 'struct workqueue_attrs' comments for detail. | |
226223ab | 3785 | */ |
6ba94429 | 3786 | pool->attrs->no_numa = false; |
226223ab | 3787 | |
6ba94429 FW |
3788 | if (worker_pool_assign_id(pool) < 0) |
3789 | goto fail; | |
226223ab | 3790 | |
6ba94429 | 3791 | /* create and start the initial worker */ |
3347fa09 | 3792 | if (wq_online && !create_worker(pool)) |
6ba94429 | 3793 | goto fail; |
226223ab | 3794 | |
6ba94429 FW |
3795 | /* install */ |
3796 | hash_add(unbound_pool_hash, &pool->hash_node, hash); | |
226223ab | 3797 | |
6ba94429 FW |
3798 | return pool; |
3799 | fail: | |
3800 | if (pool) | |
3801 | put_unbound_pool(pool); | |
3802 | return NULL; | |
226223ab | 3803 | } |
226223ab | 3804 | |
6ba94429 | 3805 | static void rcu_free_pwq(struct rcu_head *rcu) |
7a4e344c | 3806 | { |
6ba94429 FW |
3807 | kmem_cache_free(pwq_cache, |
3808 | container_of(rcu, struct pool_workqueue, rcu)); | |
7a4e344c TH |
3809 | } |
3810 | ||
6ba94429 FW |
3811 | /* |
3812 | * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt | |
3813 | * and needs to be destroyed. | |
7a4e344c | 3814 | */ |
6ba94429 | 3815 | static void pwq_unbound_release_workfn(struct work_struct *work) |
7a4e344c | 3816 | { |
6ba94429 FW |
3817 | struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, |
3818 | unbound_release_work); | |
3819 | struct workqueue_struct *wq = pwq->wq; | |
3820 | struct worker_pool *pool = pwq->pool; | |
b42b0bdd | 3821 | bool is_last = false; |
7a4e344c | 3822 | |
b42b0bdd YY |
3823 | /* |
3824 | * when @pwq is not linked, it doesn't hold any reference to the | |
3825 | * @wq, and @wq is invalid to access. | |
3826 | */ | |
3827 | if (!list_empty(&pwq->pwqs_node)) { | |
3828 | if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND))) | |
3829 | return; | |
7a4e344c | 3830 | |
b42b0bdd YY |
3831 | mutex_lock(&wq->mutex); |
3832 | list_del_rcu(&pwq->pwqs_node); | |
3833 | is_last = list_empty(&wq->pwqs); | |
3834 | mutex_unlock(&wq->mutex); | |
3835 | } | |
6ba94429 FW |
3836 | |
3837 | mutex_lock(&wq_pool_mutex); | |
3838 | put_unbound_pool(pool); | |
3839 | mutex_unlock(&wq_pool_mutex); | |
3840 | ||
25b00775 | 3841 | call_rcu(&pwq->rcu, rcu_free_pwq); |
7a4e344c | 3842 | |
2865a8fb | 3843 | /* |
6ba94429 FW |
3844 | * If we're the last pwq going away, @wq is already dead and no one |
3845 | * is gonna access it anymore. Schedule RCU free. | |
2865a8fb | 3846 | */ |
669de8bd BVA |
3847 | if (is_last) { |
3848 | wq_unregister_lockdep(wq); | |
25b00775 | 3849 | call_rcu(&wq->rcu, rcu_free_wq); |
669de8bd | 3850 | } |
29c91e99 TH |
3851 | } |
3852 | ||
7a4e344c | 3853 | /** |
6ba94429 FW |
3854 | * pwq_adjust_max_active - update a pwq's max_active to the current setting |
3855 | * @pwq: target pool_workqueue | |
d185af30 | 3856 | * |
6ba94429 | 3857 | * If @pwq isn't freezing, set @pwq->max_active to the associated |
f97a4a1a | 3858 | * workqueue's saved_max_active and activate inactive work items |
6ba94429 | 3859 | * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. |
7a4e344c | 3860 | */ |
6ba94429 | 3861 | static void pwq_adjust_max_active(struct pool_workqueue *pwq) |
4e1a1f9a | 3862 | { |
6ba94429 FW |
3863 | struct workqueue_struct *wq = pwq->wq; |
3864 | bool freezable = wq->flags & WQ_FREEZABLE; | |
3347fa09 | 3865 | unsigned long flags; |
4e1a1f9a | 3866 | |
6ba94429 FW |
3867 | /* for @wq->saved_max_active */ |
3868 | lockdep_assert_held(&wq->mutex); | |
4e1a1f9a | 3869 | |
6ba94429 FW |
3870 | /* fast exit for non-freezable wqs */ |
3871 | if (!freezable && pwq->max_active == wq->saved_max_active) | |
3872 | return; | |
7a4e344c | 3873 | |
3347fa09 | 3874 | /* this function can be called during early boot w/ irq disabled */ |
a9b8a985 | 3875 | raw_spin_lock_irqsave(&pwq->pool->lock, flags); |
29c91e99 | 3876 | |
6ba94429 FW |
3877 | /* |
3878 | * During [un]freezing, the caller is responsible for ensuring that | |
3879 | * this function is called at least once after @workqueue_freezing | |
3880 | * is updated and visible. | |
3881 | */ | |
3882 | if (!freezable || !workqueue_freezing) { | |
01341fbd YY |
3883 | bool kick = false; |
3884 | ||
6ba94429 | 3885 | pwq->max_active = wq->saved_max_active; |
4e1a1f9a | 3886 | |
f97a4a1a | 3887 | while (!list_empty(&pwq->inactive_works) && |
01341fbd | 3888 | pwq->nr_active < pwq->max_active) { |
f97a4a1a | 3889 | pwq_activate_first_inactive(pwq); |
01341fbd YY |
3890 | kick = true; |
3891 | } | |
e2dca7ad | 3892 | |
6ba94429 FW |
3893 | /* |
3894 | * Need to kick a worker after thawed or an unbound wq's | |
01341fbd YY |
3895 | * max_active is bumped. In realtime scenarios, always kicking a |
3896 | * worker will cause interference on the isolated cpu cores, so | |
3897 | * let's kick iff work items were activated. | |
6ba94429 | 3898 | */ |
01341fbd YY |
3899 | if (kick) |
3900 | wake_up_worker(pwq->pool); | |
6ba94429 FW |
3901 | } else { |
3902 | pwq->max_active = 0; | |
3903 | } | |
e2dca7ad | 3904 | |
a9b8a985 | 3905 | raw_spin_unlock_irqrestore(&pwq->pool->lock, flags); |
e2dca7ad TH |
3906 | } |
3907 | ||
67dc8325 | 3908 | /* initialize newly allocated @pwq which is associated with @wq and @pool */ |
6ba94429 FW |
3909 | static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, |
3910 | struct worker_pool *pool) | |
29c91e99 | 3911 | { |
6ba94429 | 3912 | BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); |
29c91e99 | 3913 | |
6ba94429 FW |
3914 | memset(pwq, 0, sizeof(*pwq)); |
3915 | ||
3916 | pwq->pool = pool; | |
3917 | pwq->wq = wq; | |
3918 | pwq->flush_color = -1; | |
3919 | pwq->refcnt = 1; | |
f97a4a1a | 3920 | INIT_LIST_HEAD(&pwq->inactive_works); |
6ba94429 FW |
3921 | INIT_LIST_HEAD(&pwq->pwqs_node); |
3922 | INIT_LIST_HEAD(&pwq->mayday_node); | |
3923 | INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn); | |
29c91e99 TH |
3924 | } |
3925 | ||
6ba94429 FW |
3926 | /* sync @pwq with the current state of its associated wq and link it */ |
3927 | static void link_pwq(struct pool_workqueue *pwq) | |
29c91e99 | 3928 | { |
6ba94429 | 3929 | struct workqueue_struct *wq = pwq->wq; |
29c91e99 | 3930 | |
6ba94429 | 3931 | lockdep_assert_held(&wq->mutex); |
a892cacc | 3932 | |
6ba94429 FW |
3933 | /* may be called multiple times, ignore if already linked */ |
3934 | if (!list_empty(&pwq->pwqs_node)) | |
29c91e99 | 3935 | return; |
29c91e99 | 3936 | |
6ba94429 FW |
3937 | /* set the matching work_color */ |
3938 | pwq->work_color = wq->work_color; | |
29c91e99 | 3939 | |
6ba94429 FW |
3940 | /* sync max_active to the current setting */ |
3941 | pwq_adjust_max_active(pwq); | |
29c91e99 | 3942 | |
6ba94429 FW |
3943 | /* link in @pwq */ |
3944 | list_add_rcu(&pwq->pwqs_node, &wq->pwqs); | |
3945 | } | |
29c91e99 | 3946 | |
6ba94429 FW |
3947 | /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ |
3948 | static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, | |
3949 | const struct workqueue_attrs *attrs) | |
3950 | { | |
3951 | struct worker_pool *pool; | |
3952 | struct pool_workqueue *pwq; | |
60f5a4bc | 3953 | |
6ba94429 | 3954 | lockdep_assert_held(&wq_pool_mutex); |
60f5a4bc | 3955 | |
6ba94429 FW |
3956 | pool = get_unbound_pool(attrs); |
3957 | if (!pool) | |
3958 | return NULL; | |
60f5a4bc | 3959 | |
6ba94429 FW |
3960 | pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); |
3961 | if (!pwq) { | |
3962 | put_unbound_pool(pool); | |
3963 | return NULL; | |
3964 | } | |
29c91e99 | 3965 | |
6ba94429 FW |
3966 | init_pwq(pwq, wq, pool); |
3967 | return pwq; | |
3968 | } | |
29c91e99 | 3969 | |
29c91e99 | 3970 | /** |
30186c6f | 3971 | * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node |
042f7df1 | 3972 | * @attrs: the wq_attrs of the default pwq of the target workqueue |
6ba94429 FW |
3973 | * @node: the target NUMA node |
3974 | * @cpu_going_down: if >= 0, the CPU to consider as offline | |
3975 | * @cpumask: outarg, the resulting cpumask | |
29c91e99 | 3976 | * |
6ba94429 FW |
3977 | * Calculate the cpumask a workqueue with @attrs should use on @node. If |
3978 | * @cpu_going_down is >= 0, that cpu is considered offline during | |
3979 | * calculation. The result is stored in @cpumask. | |
a892cacc | 3980 | * |
6ba94429 FW |
3981 | * If NUMA affinity is not enabled, @attrs->cpumask is always used. If |
3982 | * enabled and @node has online CPUs requested by @attrs, the returned | |
3983 | * cpumask is the intersection of the possible CPUs of @node and | |
3984 | * @attrs->cpumask. | |
d185af30 | 3985 | * |
6ba94429 FW |
3986 | * The caller is responsible for ensuring that the cpumask of @node stays |
3987 | * stable. | |
3988 | * | |
3989 | * Return: %true if the resulting @cpumask is different from @attrs->cpumask, | |
3990 | * %false if equal. | |
29c91e99 | 3991 | */ |
6ba94429 FW |
3992 | static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node, |
3993 | int cpu_going_down, cpumask_t *cpumask) | |
29c91e99 | 3994 | { |
6ba94429 FW |
3995 | if (!wq_numa_enabled || attrs->no_numa) |
3996 | goto use_dfl; | |
29c91e99 | 3997 | |
6ba94429 FW |
3998 | /* does @node have any online CPUs @attrs wants? */ |
3999 | cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask); | |
4000 | if (cpu_going_down >= 0) | |
4001 | cpumask_clear_cpu(cpu_going_down, cpumask); | |
29c91e99 | 4002 | |
6ba94429 FW |
4003 | if (cpumask_empty(cpumask)) |
4004 | goto use_dfl; | |
4c16bd32 TH |
4005 | |
4006 | /* yeap, return possible CPUs in @node that @attrs wants */ | |
4007 | cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]); | |
1ad0f0a7 MB |
4008 | |
4009 | if (cpumask_empty(cpumask)) { | |
4010 | pr_warn_once("WARNING: workqueue cpumask: online intersect > " | |
4011 | "possible intersect\n"); | |
4012 | return false; | |
4013 | } | |
4014 | ||
4c16bd32 TH |
4015 | return !cpumask_equal(cpumask, attrs->cpumask); |
4016 | ||
4017 | use_dfl: | |
4018 | cpumask_copy(cpumask, attrs->cpumask); | |
4019 | return false; | |
4020 | } | |
4021 | ||
1befcf30 TH |
4022 | /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */ |
4023 | static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq, | |
4024 | int node, | |
4025 | struct pool_workqueue *pwq) | |
4026 | { | |
4027 | struct pool_workqueue *old_pwq; | |
4028 | ||
5b95e1af | 4029 | lockdep_assert_held(&wq_pool_mutex); |
1befcf30 TH |
4030 | lockdep_assert_held(&wq->mutex); |
4031 | ||
4032 | /* link_pwq() can handle duplicate calls */ | |
4033 | link_pwq(pwq); | |
4034 | ||
4035 | old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); | |
4036 | rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq); | |
4037 | return old_pwq; | |
4038 | } | |
4039 | ||
2d5f0764 LJ |
4040 | /* context to store the prepared attrs & pwqs before applying */ |
4041 | struct apply_wqattrs_ctx { | |
4042 | struct workqueue_struct *wq; /* target workqueue */ | |
4043 | struct workqueue_attrs *attrs; /* attrs to apply */ | |
042f7df1 | 4044 | struct list_head list; /* queued for batching commit */ |
2d5f0764 LJ |
4045 | struct pool_workqueue *dfl_pwq; |
4046 | struct pool_workqueue *pwq_tbl[]; | |
4047 | }; | |
4048 | ||
4049 | /* free the resources after success or abort */ | |
4050 | static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx) | |
4051 | { | |
4052 | if (ctx) { | |
4053 | int node; | |
4054 | ||
4055 | for_each_node(node) | |
4056 | put_pwq_unlocked(ctx->pwq_tbl[node]); | |
4057 | put_pwq_unlocked(ctx->dfl_pwq); | |
4058 | ||
4059 | free_workqueue_attrs(ctx->attrs); | |
4060 | ||
4061 | kfree(ctx); | |
4062 | } | |
4063 | } | |
4064 | ||
4065 | /* allocate the attrs and pwqs for later installation */ | |
4066 | static struct apply_wqattrs_ctx * | |
4067 | apply_wqattrs_prepare(struct workqueue_struct *wq, | |
99c621ef LJ |
4068 | const struct workqueue_attrs *attrs, |
4069 | const cpumask_var_t unbound_cpumask) | |
9e8cd2f5 | 4070 | { |
2d5f0764 | 4071 | struct apply_wqattrs_ctx *ctx; |
4c16bd32 | 4072 | struct workqueue_attrs *new_attrs, *tmp_attrs; |
2d5f0764 | 4073 | int node; |
9e8cd2f5 | 4074 | |
2d5f0764 | 4075 | lockdep_assert_held(&wq_pool_mutex); |
9e8cd2f5 | 4076 | |
acafe7e3 | 4077 | ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_node_ids), GFP_KERNEL); |
8719dcea | 4078 | |
be69d00d TG |
4079 | new_attrs = alloc_workqueue_attrs(); |
4080 | tmp_attrs = alloc_workqueue_attrs(); | |
2d5f0764 LJ |
4081 | if (!ctx || !new_attrs || !tmp_attrs) |
4082 | goto out_free; | |
13e2e556 | 4083 | |
042f7df1 | 4084 | /* |
99c621ef LJ |
4085 | * Calculate the attrs of the default pwq with unbound_cpumask |
4086 | * which is wq_unbound_cpumask or to set to wq_unbound_cpumask. | |
042f7df1 LJ |
4087 | * If the user configured cpumask doesn't overlap with the |
4088 | * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask. | |
4089 | */ | |
13e2e556 | 4090 | copy_workqueue_attrs(new_attrs, attrs); |
99c621ef | 4091 | cpumask_and(new_attrs->cpumask, new_attrs->cpumask, unbound_cpumask); |
042f7df1 | 4092 | if (unlikely(cpumask_empty(new_attrs->cpumask))) |
99c621ef | 4093 | cpumask_copy(new_attrs->cpumask, unbound_cpumask); |
13e2e556 | 4094 | |
4c16bd32 TH |
4095 | /* |
4096 | * We may create multiple pwqs with differing cpumasks. Make a | |
4097 | * copy of @new_attrs which will be modified and used to obtain | |
4098 | * pools. | |
4099 | */ | |
4100 | copy_workqueue_attrs(tmp_attrs, new_attrs); | |
4101 | ||
4c16bd32 TH |
4102 | /* |
4103 | * If something goes wrong during CPU up/down, we'll fall back to | |
4104 | * the default pwq covering whole @attrs->cpumask. Always create | |
4105 | * it even if we don't use it immediately. | |
4106 | */ | |
2d5f0764 LJ |
4107 | ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs); |
4108 | if (!ctx->dfl_pwq) | |
4109 | goto out_free; | |
4c16bd32 TH |
4110 | |
4111 | for_each_node(node) { | |
042f7df1 | 4112 | if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) { |
2d5f0764 LJ |
4113 | ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs); |
4114 | if (!ctx->pwq_tbl[node]) | |
4115 | goto out_free; | |
4c16bd32 | 4116 | } else { |
2d5f0764 LJ |
4117 | ctx->dfl_pwq->refcnt++; |
4118 | ctx->pwq_tbl[node] = ctx->dfl_pwq; | |
4c16bd32 TH |
4119 | } |
4120 | } | |
4121 | ||
042f7df1 LJ |
4122 | /* save the user configured attrs and sanitize it. */ |
4123 | copy_workqueue_attrs(new_attrs, attrs); | |
4124 | cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); | |
2d5f0764 | 4125 | ctx->attrs = new_attrs; |
042f7df1 | 4126 | |
2d5f0764 LJ |
4127 | ctx->wq = wq; |
4128 | free_workqueue_attrs(tmp_attrs); | |
4129 | return ctx; | |
4130 | ||
4131 | out_free: | |
4132 | free_workqueue_attrs(tmp_attrs); | |
4133 | free_workqueue_attrs(new_attrs); | |
4134 | apply_wqattrs_cleanup(ctx); | |
4135 | return NULL; | |
4136 | } | |
4137 | ||
4138 | /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */ | |
4139 | static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx) | |
4140 | { | |
4141 | int node; | |
9e8cd2f5 | 4142 | |
4c16bd32 | 4143 | /* all pwqs have been created successfully, let's install'em */ |
2d5f0764 | 4144 | mutex_lock(&ctx->wq->mutex); |
a892cacc | 4145 | |
2d5f0764 | 4146 | copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs); |
4c16bd32 TH |
4147 | |
4148 | /* save the previous pwq and install the new one */ | |
f147f29e | 4149 | for_each_node(node) |
2d5f0764 LJ |
4150 | ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node, |
4151 | ctx->pwq_tbl[node]); | |
4c16bd32 TH |
4152 | |
4153 | /* @dfl_pwq might not have been used, ensure it's linked */ | |
2d5f0764 LJ |
4154 | link_pwq(ctx->dfl_pwq); |
4155 | swap(ctx->wq->dfl_pwq, ctx->dfl_pwq); | |
f147f29e | 4156 | |
2d5f0764 LJ |
4157 | mutex_unlock(&ctx->wq->mutex); |
4158 | } | |
9e8cd2f5 | 4159 | |
a0111cf6 LJ |
4160 | static void apply_wqattrs_lock(void) |
4161 | { | |
4162 | /* CPUs should stay stable across pwq creations and installations */ | |
ffd8bea8 | 4163 | cpus_read_lock(); |
a0111cf6 LJ |
4164 | mutex_lock(&wq_pool_mutex); |
4165 | } | |
4166 | ||
4167 | static void apply_wqattrs_unlock(void) | |
4168 | { | |
4169 | mutex_unlock(&wq_pool_mutex); | |
ffd8bea8 | 4170 | cpus_read_unlock(); |
a0111cf6 LJ |
4171 | } |
4172 | ||
4173 | static int apply_workqueue_attrs_locked(struct workqueue_struct *wq, | |
4174 | const struct workqueue_attrs *attrs) | |
2d5f0764 LJ |
4175 | { |
4176 | struct apply_wqattrs_ctx *ctx; | |
4c16bd32 | 4177 | |
2d5f0764 LJ |
4178 | /* only unbound workqueues can change attributes */ |
4179 | if (WARN_ON(!(wq->flags & WQ_UNBOUND))) | |
4180 | return -EINVAL; | |
13e2e556 | 4181 | |
2d5f0764 | 4182 | /* creating multiple pwqs breaks ordering guarantee */ |
0a94efb5 TH |
4183 | if (!list_empty(&wq->pwqs)) { |
4184 | if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) | |
4185 | return -EINVAL; | |
4186 | ||
4187 | wq->flags &= ~__WQ_ORDERED; | |
4188 | } | |
2d5f0764 | 4189 | |
99c621ef | 4190 | ctx = apply_wqattrs_prepare(wq, attrs, wq_unbound_cpumask); |
6201171e | 4191 | if (!ctx) |
4192 | return -ENOMEM; | |
2d5f0764 LJ |
4193 | |
4194 | /* the ctx has been prepared successfully, let's commit it */ | |
6201171e | 4195 | apply_wqattrs_commit(ctx); |
2d5f0764 LJ |
4196 | apply_wqattrs_cleanup(ctx); |
4197 | ||
6201171e | 4198 | return 0; |
9e8cd2f5 TH |
4199 | } |
4200 | ||
a0111cf6 LJ |
4201 | /** |
4202 | * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue | |
4203 | * @wq: the target workqueue | |
4204 | * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() | |
4205 | * | |
4206 | * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA | |
4207 | * machines, this function maps a separate pwq to each NUMA node with | |
4208 | * possibles CPUs in @attrs->cpumask so that work items are affine to the | |
4209 | * NUMA node it was issued on. Older pwqs are released as in-flight work | |
4210 | * items finish. Note that a work item which repeatedly requeues itself | |
4211 | * back-to-back will stay on its current pwq. | |
4212 | * | |
4213 | * Performs GFP_KERNEL allocations. | |
4214 | * | |
ffd8bea8 | 4215 | * Assumes caller has CPU hotplug read exclusion, i.e. cpus_read_lock(). |
509b3204 | 4216 | * |
a0111cf6 LJ |
4217 | * Return: 0 on success and -errno on failure. |
4218 | */ | |
513c98d0 | 4219 | int apply_workqueue_attrs(struct workqueue_struct *wq, |
a0111cf6 LJ |
4220 | const struct workqueue_attrs *attrs) |
4221 | { | |
4222 | int ret; | |
4223 | ||
509b3204 DJ |
4224 | lockdep_assert_cpus_held(); |
4225 | ||
4226 | mutex_lock(&wq_pool_mutex); | |
a0111cf6 | 4227 | ret = apply_workqueue_attrs_locked(wq, attrs); |
509b3204 | 4228 | mutex_unlock(&wq_pool_mutex); |
a0111cf6 LJ |
4229 | |
4230 | return ret; | |
4231 | } | |
4232 | ||
4c16bd32 TH |
4233 | /** |
4234 | * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug | |
4235 | * @wq: the target workqueue | |
4236 | * @cpu: the CPU coming up or going down | |
4237 | * @online: whether @cpu is coming up or going down | |
4238 | * | |
4239 | * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and | |
4240 | * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of | |
4241 | * @wq accordingly. | |
4242 | * | |
4243 | * If NUMA affinity can't be adjusted due to memory allocation failure, it | |
4244 | * falls back to @wq->dfl_pwq which may not be optimal but is always | |
4245 | * correct. | |
4246 | * | |
4247 | * Note that when the last allowed CPU of a NUMA node goes offline for a | |
4248 | * workqueue with a cpumask spanning multiple nodes, the workers which were | |
4249 | * already executing the work items for the workqueue will lose their CPU | |
4250 | * affinity and may execute on any CPU. This is similar to how per-cpu | |
4251 | * workqueues behave on CPU_DOWN. If a workqueue user wants strict | |
4252 | * affinity, it's the user's responsibility to flush the work item from | |
4253 | * CPU_DOWN_PREPARE. | |
4254 | */ | |
4255 | static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu, | |
4256 | bool online) | |
4257 | { | |
4258 | int node = cpu_to_node(cpu); | |
4259 | int cpu_off = online ? -1 : cpu; | |
4260 | struct pool_workqueue *old_pwq = NULL, *pwq; | |
4261 | struct workqueue_attrs *target_attrs; | |
4262 | cpumask_t *cpumask; | |
4263 | ||
4264 | lockdep_assert_held(&wq_pool_mutex); | |
4265 | ||
f7142ed4 LJ |
4266 | if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) || |
4267 | wq->unbound_attrs->no_numa) | |
4c16bd32 TH |
4268 | return; |
4269 | ||
4270 | /* | |
4271 | * We don't wanna alloc/free wq_attrs for each wq for each CPU. | |
4272 | * Let's use a preallocated one. The following buf is protected by | |
4273 | * CPU hotplug exclusion. | |
4274 | */ | |
4275 | target_attrs = wq_update_unbound_numa_attrs_buf; | |
4276 | cpumask = target_attrs->cpumask; | |
4277 | ||
4c16bd32 TH |
4278 | copy_workqueue_attrs(target_attrs, wq->unbound_attrs); |
4279 | pwq = unbound_pwq_by_node(wq, node); | |
4280 | ||
4281 | /* | |
4282 | * Let's determine what needs to be done. If the target cpumask is | |
042f7df1 LJ |
4283 | * different from the default pwq's, we need to compare it to @pwq's |
4284 | * and create a new one if they don't match. If the target cpumask | |
4285 | * equals the default pwq's, the default pwq should be used. | |
4c16bd32 | 4286 | */ |
042f7df1 | 4287 | if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) { |
4c16bd32 | 4288 | if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask)) |
f7142ed4 | 4289 | return; |
4c16bd32 | 4290 | } else { |
534a3fbb | 4291 | goto use_dfl_pwq; |
4c16bd32 TH |
4292 | } |
4293 | ||
4c16bd32 TH |
4294 | /* create a new pwq */ |
4295 | pwq = alloc_unbound_pwq(wq, target_attrs); | |
4296 | if (!pwq) { | |
2d916033 FF |
4297 | pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n", |
4298 | wq->name); | |
77f300b1 | 4299 | goto use_dfl_pwq; |
4c16bd32 TH |
4300 | } |
4301 | ||
f7142ed4 | 4302 | /* Install the new pwq. */ |
4c16bd32 TH |
4303 | mutex_lock(&wq->mutex); |
4304 | old_pwq = numa_pwq_tbl_install(wq, node, pwq); | |
4305 | goto out_unlock; | |
4306 | ||
4307 | use_dfl_pwq: | |
f7142ed4 | 4308 | mutex_lock(&wq->mutex); |
a9b8a985 | 4309 | raw_spin_lock_irq(&wq->dfl_pwq->pool->lock); |
4c16bd32 | 4310 | get_pwq(wq->dfl_pwq); |
a9b8a985 | 4311 | raw_spin_unlock_irq(&wq->dfl_pwq->pool->lock); |
4c16bd32 TH |
4312 | old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq); |
4313 | out_unlock: | |
4314 | mutex_unlock(&wq->mutex); | |
4315 | put_pwq_unlocked(old_pwq); | |
4316 | } | |
4317 | ||
30cdf249 | 4318 | static int alloc_and_link_pwqs(struct workqueue_struct *wq) |
0f900049 | 4319 | { |
49e3cf44 | 4320 | bool highpri = wq->flags & WQ_HIGHPRI; |
8a2b7538 | 4321 | int cpu, ret; |
30cdf249 TH |
4322 | |
4323 | if (!(wq->flags & WQ_UNBOUND)) { | |
420c0ddb TH |
4324 | wq->cpu_pwqs = alloc_percpu(struct pool_workqueue); |
4325 | if (!wq->cpu_pwqs) | |
30cdf249 TH |
4326 | return -ENOMEM; |
4327 | ||
4328 | for_each_possible_cpu(cpu) { | |
7fb98ea7 TH |
4329 | struct pool_workqueue *pwq = |
4330 | per_cpu_ptr(wq->cpu_pwqs, cpu); | |
7a62c2c8 | 4331 | struct worker_pool *cpu_pools = |
f02ae73a | 4332 | per_cpu(cpu_worker_pools, cpu); |
f3421797 | 4333 | |
f147f29e TH |
4334 | init_pwq(pwq, wq, &cpu_pools[highpri]); |
4335 | ||
4336 | mutex_lock(&wq->mutex); | |
1befcf30 | 4337 | link_pwq(pwq); |
f147f29e | 4338 | mutex_unlock(&wq->mutex); |
30cdf249 | 4339 | } |
9e8cd2f5 | 4340 | return 0; |
509b3204 DJ |
4341 | } |
4342 | ||
ffd8bea8 | 4343 | cpus_read_lock(); |
509b3204 | 4344 | if (wq->flags & __WQ_ORDERED) { |
8a2b7538 TH |
4345 | ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]); |
4346 | /* there should only be single pwq for ordering guarantee */ | |
4347 | WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node || | |
4348 | wq->pwqs.prev != &wq->dfl_pwq->pwqs_node), | |
4349 | "ordering guarantee broken for workqueue %s\n", wq->name); | |
30cdf249 | 4350 | } else { |
509b3204 | 4351 | ret = apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); |
30cdf249 | 4352 | } |
ffd8bea8 | 4353 | cpus_read_unlock(); |
509b3204 DJ |
4354 | |
4355 | return ret; | |
0f900049 TH |
4356 | } |
4357 | ||
f3421797 TH |
4358 | static int wq_clamp_max_active(int max_active, unsigned int flags, |
4359 | const char *name) | |
b71ab8c2 | 4360 | { |
f3421797 TH |
4361 | int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; |
4362 | ||
4363 | if (max_active < 1 || max_active > lim) | |
044c782c VI |
4364 | pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", |
4365 | max_active, name, 1, lim); | |
b71ab8c2 | 4366 | |
f3421797 | 4367 | return clamp_val(max_active, 1, lim); |
b71ab8c2 TH |
4368 | } |
4369 | ||
983c7515 TH |
4370 | /* |
4371 | * Workqueues which may be used during memory reclaim should have a rescuer | |
4372 | * to guarantee forward progress. | |
4373 | */ | |
4374 | static int init_rescuer(struct workqueue_struct *wq) | |
4375 | { | |
4376 | struct worker *rescuer; | |
b92b36ea | 4377 | int ret; |
983c7515 TH |
4378 | |
4379 | if (!(wq->flags & WQ_MEM_RECLAIM)) | |
4380 | return 0; | |
4381 | ||
4382 | rescuer = alloc_worker(NUMA_NO_NODE); | |
4383 | if (!rescuer) | |
4384 | return -ENOMEM; | |
4385 | ||
4386 | rescuer->rescue_wq = wq; | |
4387 | rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", wq->name); | |
f187b697 | 4388 | if (IS_ERR(rescuer->task)) { |
b92b36ea | 4389 | ret = PTR_ERR(rescuer->task); |
983c7515 | 4390 | kfree(rescuer); |
b92b36ea | 4391 | return ret; |
983c7515 TH |
4392 | } |
4393 | ||
4394 | wq->rescuer = rescuer; | |
4395 | kthread_bind_mask(rescuer->task, cpu_possible_mask); | |
4396 | wake_up_process(rescuer->task); | |
4397 | ||
4398 | return 0; | |
4399 | } | |
4400 | ||
a2775bbc | 4401 | __printf(1, 4) |
669de8bd BVA |
4402 | struct workqueue_struct *alloc_workqueue(const char *fmt, |
4403 | unsigned int flags, | |
4404 | int max_active, ...) | |
1da177e4 | 4405 | { |
df2d5ae4 | 4406 | size_t tbl_size = 0; |
ecf6881f | 4407 | va_list args; |
1da177e4 | 4408 | struct workqueue_struct *wq; |
49e3cf44 | 4409 | struct pool_workqueue *pwq; |
b196be89 | 4410 | |
5c0338c6 TH |
4411 | /* |
4412 | * Unbound && max_active == 1 used to imply ordered, which is no | |
4413 | * longer the case on NUMA machines due to per-node pools. While | |
4414 | * alloc_ordered_workqueue() is the right way to create an ordered | |
4415 | * workqueue, keep the previous behavior to avoid subtle breakages | |
4416 | * on NUMA. | |
4417 | */ | |
4418 | if ((flags & WQ_UNBOUND) && max_active == 1) | |
4419 | flags |= __WQ_ORDERED; | |
4420 | ||
cee22a15 VK |
4421 | /* see the comment above the definition of WQ_POWER_EFFICIENT */ |
4422 | if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) | |
4423 | flags |= WQ_UNBOUND; | |
4424 | ||
ecf6881f | 4425 | /* allocate wq and format name */ |
df2d5ae4 | 4426 | if (flags & WQ_UNBOUND) |
ddcb57e2 | 4427 | tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]); |
df2d5ae4 TH |
4428 | |
4429 | wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL); | |
b196be89 | 4430 | if (!wq) |
d2c1d404 | 4431 | return NULL; |
b196be89 | 4432 | |
6029a918 | 4433 | if (flags & WQ_UNBOUND) { |
be69d00d | 4434 | wq->unbound_attrs = alloc_workqueue_attrs(); |
6029a918 TH |
4435 | if (!wq->unbound_attrs) |
4436 | goto err_free_wq; | |
4437 | } | |
4438 | ||
669de8bd | 4439 | va_start(args, max_active); |
ecf6881f | 4440 | vsnprintf(wq->name, sizeof(wq->name), fmt, args); |
b196be89 | 4441 | va_end(args); |
1da177e4 | 4442 | |
d320c038 | 4443 | max_active = max_active ?: WQ_DFL_ACTIVE; |
b196be89 | 4444 | max_active = wq_clamp_max_active(max_active, flags, wq->name); |
3af24433 | 4445 | |
b196be89 | 4446 | /* init wq */ |
97e37d7b | 4447 | wq->flags = flags; |
a0a1a5fd | 4448 | wq->saved_max_active = max_active; |
3c25a55d | 4449 | mutex_init(&wq->mutex); |
112202d9 | 4450 | atomic_set(&wq->nr_pwqs_to_flush, 0); |
30cdf249 | 4451 | INIT_LIST_HEAD(&wq->pwqs); |
73f53c4a TH |
4452 | INIT_LIST_HEAD(&wq->flusher_queue); |
4453 | INIT_LIST_HEAD(&wq->flusher_overflow); | |
493a1724 | 4454 | INIT_LIST_HEAD(&wq->maydays); |
502ca9d8 | 4455 | |
669de8bd | 4456 | wq_init_lockdep(wq); |
cce1a165 | 4457 | INIT_LIST_HEAD(&wq->list); |
3af24433 | 4458 | |
30cdf249 | 4459 | if (alloc_and_link_pwqs(wq) < 0) |
82efcab3 | 4460 | goto err_unreg_lockdep; |
1537663f | 4461 | |
40c17f75 | 4462 | if (wq_online && init_rescuer(wq) < 0) |
983c7515 | 4463 | goto err_destroy; |
3af24433 | 4464 | |
226223ab TH |
4465 | if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) |
4466 | goto err_destroy; | |
4467 | ||
a0a1a5fd | 4468 | /* |
68e13a67 LJ |
4469 | * wq_pool_mutex protects global freeze state and workqueues list. |
4470 | * Grab it, adjust max_active and add the new @wq to workqueues | |
4471 | * list. | |
a0a1a5fd | 4472 | */ |
68e13a67 | 4473 | mutex_lock(&wq_pool_mutex); |
a0a1a5fd | 4474 | |
a357fc03 | 4475 | mutex_lock(&wq->mutex); |
699ce097 TH |
4476 | for_each_pwq(pwq, wq) |
4477 | pwq_adjust_max_active(pwq); | |
a357fc03 | 4478 | mutex_unlock(&wq->mutex); |
a0a1a5fd | 4479 | |
e2dca7ad | 4480 | list_add_tail_rcu(&wq->list, &workqueues); |
a0a1a5fd | 4481 | |
68e13a67 | 4482 | mutex_unlock(&wq_pool_mutex); |
1537663f | 4483 | |
3af24433 | 4484 | return wq; |
d2c1d404 | 4485 | |
82efcab3 | 4486 | err_unreg_lockdep: |
009bb421 BVA |
4487 | wq_unregister_lockdep(wq); |
4488 | wq_free_lockdep(wq); | |
82efcab3 | 4489 | err_free_wq: |
6029a918 | 4490 | free_workqueue_attrs(wq->unbound_attrs); |
d2c1d404 TH |
4491 | kfree(wq); |
4492 | return NULL; | |
4493 | err_destroy: | |
4494 | destroy_workqueue(wq); | |
4690c4ab | 4495 | return NULL; |
3af24433 | 4496 | } |
669de8bd | 4497 | EXPORT_SYMBOL_GPL(alloc_workqueue); |
1da177e4 | 4498 | |
c29eb853 TH |
4499 | static bool pwq_busy(struct pool_workqueue *pwq) |
4500 | { | |
4501 | int i; | |
4502 | ||
4503 | for (i = 0; i < WORK_NR_COLORS; i++) | |
4504 | if (pwq->nr_in_flight[i]) | |
4505 | return true; | |
4506 | ||
4507 | if ((pwq != pwq->wq->dfl_pwq) && (pwq->refcnt > 1)) | |
4508 | return true; | |
f97a4a1a | 4509 | if (pwq->nr_active || !list_empty(&pwq->inactive_works)) |
c29eb853 TH |
4510 | return true; |
4511 | ||
4512 | return false; | |
4513 | } | |
4514 | ||
3af24433 ON |
4515 | /** |
4516 | * destroy_workqueue - safely terminate a workqueue | |
4517 | * @wq: target workqueue | |
4518 | * | |
4519 | * Safely destroy a workqueue. All work currently pending will be done first. | |
4520 | */ | |
4521 | void destroy_workqueue(struct workqueue_struct *wq) | |
4522 | { | |
49e3cf44 | 4523 | struct pool_workqueue *pwq; |
4c16bd32 | 4524 | int node; |
3af24433 | 4525 | |
def98c84 TH |
4526 | /* |
4527 | * Remove it from sysfs first so that sanity check failure doesn't | |
4528 | * lead to sysfs name conflicts. | |
4529 | */ | |
4530 | workqueue_sysfs_unregister(wq); | |
4531 | ||
33e3f0a3 RC |
4532 | /* mark the workqueue destruction is in progress */ |
4533 | mutex_lock(&wq->mutex); | |
4534 | wq->flags |= __WQ_DESTROYING; | |
4535 | mutex_unlock(&wq->mutex); | |
4536 | ||
9c5a2ba7 TH |
4537 | /* drain it before proceeding with destruction */ |
4538 | drain_workqueue(wq); | |
c8efcc25 | 4539 | |
def98c84 TH |
4540 | /* kill rescuer, if sanity checks fail, leave it w/o rescuer */ |
4541 | if (wq->rescuer) { | |
4542 | struct worker *rescuer = wq->rescuer; | |
4543 | ||
4544 | /* this prevents new queueing */ | |
a9b8a985 | 4545 | raw_spin_lock_irq(&wq_mayday_lock); |
def98c84 | 4546 | wq->rescuer = NULL; |
a9b8a985 | 4547 | raw_spin_unlock_irq(&wq_mayday_lock); |
def98c84 TH |
4548 | |
4549 | /* rescuer will empty maydays list before exiting */ | |
4550 | kthread_stop(rescuer->task); | |
8efe1223 | 4551 | kfree(rescuer); |
def98c84 TH |
4552 | } |
4553 | ||
c29eb853 TH |
4554 | /* |
4555 | * Sanity checks - grab all the locks so that we wait for all | |
4556 | * in-flight operations which may do put_pwq(). | |
4557 | */ | |
4558 | mutex_lock(&wq_pool_mutex); | |
b09f4fd3 | 4559 | mutex_lock(&wq->mutex); |
49e3cf44 | 4560 | for_each_pwq(pwq, wq) { |
a9b8a985 | 4561 | raw_spin_lock_irq(&pwq->pool->lock); |
c29eb853 | 4562 | if (WARN_ON(pwq_busy(pwq))) { |
1d9a6159 KW |
4563 | pr_warn("%s: %s has the following busy pwq\n", |
4564 | __func__, wq->name); | |
c29eb853 | 4565 | show_pwq(pwq); |
a9b8a985 | 4566 | raw_spin_unlock_irq(&pwq->pool->lock); |
b09f4fd3 | 4567 | mutex_unlock(&wq->mutex); |
c29eb853 | 4568 | mutex_unlock(&wq_pool_mutex); |
55df0933 | 4569 | show_one_workqueue(wq); |
6183c009 | 4570 | return; |
76af4d93 | 4571 | } |
a9b8a985 | 4572 | raw_spin_unlock_irq(&pwq->pool->lock); |
6183c009 | 4573 | } |
b09f4fd3 | 4574 | mutex_unlock(&wq->mutex); |
6183c009 | 4575 | |
a0a1a5fd TH |
4576 | /* |
4577 | * wq list is used to freeze wq, remove from list after | |
4578 | * flushing is complete in case freeze races us. | |
4579 | */ | |
e2dca7ad | 4580 | list_del_rcu(&wq->list); |
68e13a67 | 4581 | mutex_unlock(&wq_pool_mutex); |
3af24433 | 4582 | |
8864b4e5 | 4583 | if (!(wq->flags & WQ_UNBOUND)) { |
669de8bd | 4584 | wq_unregister_lockdep(wq); |
8864b4e5 TH |
4585 | /* |
4586 | * The base ref is never dropped on per-cpu pwqs. Directly | |
e2dca7ad | 4587 | * schedule RCU free. |
8864b4e5 | 4588 | */ |
25b00775 | 4589 | call_rcu(&wq->rcu, rcu_free_wq); |
8864b4e5 TH |
4590 | } else { |
4591 | /* | |
4592 | * We're the sole accessor of @wq at this point. Directly | |
4c16bd32 TH |
4593 | * access numa_pwq_tbl[] and dfl_pwq to put the base refs. |
4594 | * @wq will be freed when the last pwq is released. | |
8864b4e5 | 4595 | */ |
4c16bd32 TH |
4596 | for_each_node(node) { |
4597 | pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); | |
4598 | RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL); | |
4599 | put_pwq_unlocked(pwq); | |
4600 | } | |
4601 | ||
4602 | /* | |
4603 | * Put dfl_pwq. @wq may be freed any time after dfl_pwq is | |
4604 | * put. Don't access it afterwards. | |
4605 | */ | |
4606 | pwq = wq->dfl_pwq; | |
4607 | wq->dfl_pwq = NULL; | |
dce90d47 | 4608 | put_pwq_unlocked(pwq); |
29c91e99 | 4609 | } |
3af24433 ON |
4610 | } |
4611 | EXPORT_SYMBOL_GPL(destroy_workqueue); | |
4612 | ||
dcd989cb TH |
4613 | /** |
4614 | * workqueue_set_max_active - adjust max_active of a workqueue | |
4615 | * @wq: target workqueue | |
4616 | * @max_active: new max_active value. | |
4617 | * | |
4618 | * Set max_active of @wq to @max_active. | |
4619 | * | |
4620 | * CONTEXT: | |
4621 | * Don't call from IRQ context. | |
4622 | */ | |
4623 | void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) | |
4624 | { | |
49e3cf44 | 4625 | struct pool_workqueue *pwq; |
dcd989cb | 4626 | |
8719dcea | 4627 | /* disallow meddling with max_active for ordered workqueues */ |
0a94efb5 | 4628 | if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) |
8719dcea TH |
4629 | return; |
4630 | ||
f3421797 | 4631 | max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); |
dcd989cb | 4632 | |
a357fc03 | 4633 | mutex_lock(&wq->mutex); |
dcd989cb | 4634 | |
0a94efb5 | 4635 | wq->flags &= ~__WQ_ORDERED; |
dcd989cb TH |
4636 | wq->saved_max_active = max_active; |
4637 | ||
699ce097 TH |
4638 | for_each_pwq(pwq, wq) |
4639 | pwq_adjust_max_active(pwq); | |
93981800 | 4640 | |
a357fc03 | 4641 | mutex_unlock(&wq->mutex); |
15316ba8 | 4642 | } |
dcd989cb | 4643 | EXPORT_SYMBOL_GPL(workqueue_set_max_active); |
15316ba8 | 4644 | |
27d4ee03 LW |
4645 | /** |
4646 | * current_work - retrieve %current task's work struct | |
4647 | * | |
4648 | * Determine if %current task is a workqueue worker and what it's working on. | |
4649 | * Useful to find out the context that the %current task is running in. | |
4650 | * | |
4651 | * Return: work struct if %current task is a workqueue worker, %NULL otherwise. | |
4652 | */ | |
4653 | struct work_struct *current_work(void) | |
4654 | { | |
4655 | struct worker *worker = current_wq_worker(); | |
4656 | ||
4657 | return worker ? worker->current_work : NULL; | |
4658 | } | |
4659 | EXPORT_SYMBOL(current_work); | |
4660 | ||
e6267616 TH |
4661 | /** |
4662 | * current_is_workqueue_rescuer - is %current workqueue rescuer? | |
4663 | * | |
4664 | * Determine whether %current is a workqueue rescuer. Can be used from | |
4665 | * work functions to determine whether it's being run off the rescuer task. | |
d185af30 YB |
4666 | * |
4667 | * Return: %true if %current is a workqueue rescuer. %false otherwise. | |
e6267616 TH |
4668 | */ |
4669 | bool current_is_workqueue_rescuer(void) | |
4670 | { | |
4671 | struct worker *worker = current_wq_worker(); | |
4672 | ||
6a092dfd | 4673 | return worker && worker->rescue_wq; |
e6267616 TH |
4674 | } |
4675 | ||
eef6a7d5 | 4676 | /** |
dcd989cb TH |
4677 | * workqueue_congested - test whether a workqueue is congested |
4678 | * @cpu: CPU in question | |
4679 | * @wq: target workqueue | |
eef6a7d5 | 4680 | * |
dcd989cb TH |
4681 | * Test whether @wq's cpu workqueue for @cpu is congested. There is |
4682 | * no synchronization around this function and the test result is | |
4683 | * unreliable and only useful as advisory hints or for debugging. | |
eef6a7d5 | 4684 | * |
d3251859 TH |
4685 | * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. |
4686 | * Note that both per-cpu and unbound workqueues may be associated with | |
4687 | * multiple pool_workqueues which have separate congested states. A | |
4688 | * workqueue being congested on one CPU doesn't mean the workqueue is also | |
4689 | * contested on other CPUs / NUMA nodes. | |
4690 | * | |
d185af30 | 4691 | * Return: |
dcd989cb | 4692 | * %true if congested, %false otherwise. |
eef6a7d5 | 4693 | */ |
d84ff051 | 4694 | bool workqueue_congested(int cpu, struct workqueue_struct *wq) |
1da177e4 | 4695 | { |
7fb98ea7 | 4696 | struct pool_workqueue *pwq; |
76af4d93 TH |
4697 | bool ret; |
4698 | ||
24acfb71 TG |
4699 | rcu_read_lock(); |
4700 | preempt_disable(); | |
7fb98ea7 | 4701 | |
d3251859 TH |
4702 | if (cpu == WORK_CPU_UNBOUND) |
4703 | cpu = smp_processor_id(); | |
4704 | ||
7fb98ea7 TH |
4705 | if (!(wq->flags & WQ_UNBOUND)) |
4706 | pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); | |
4707 | else | |
df2d5ae4 | 4708 | pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); |
dcd989cb | 4709 | |
f97a4a1a | 4710 | ret = !list_empty(&pwq->inactive_works); |
24acfb71 TG |
4711 | preempt_enable(); |
4712 | rcu_read_unlock(); | |
76af4d93 TH |
4713 | |
4714 | return ret; | |
1da177e4 | 4715 | } |
dcd989cb | 4716 | EXPORT_SYMBOL_GPL(workqueue_congested); |
1da177e4 | 4717 | |
dcd989cb TH |
4718 | /** |
4719 | * work_busy - test whether a work is currently pending or running | |
4720 | * @work: the work to be tested | |
4721 | * | |
4722 | * Test whether @work is currently pending or running. There is no | |
4723 | * synchronization around this function and the test result is | |
4724 | * unreliable and only useful as advisory hints or for debugging. | |
dcd989cb | 4725 | * |
d185af30 | 4726 | * Return: |
dcd989cb TH |
4727 | * OR'd bitmask of WORK_BUSY_* bits. |
4728 | */ | |
4729 | unsigned int work_busy(struct work_struct *work) | |
1da177e4 | 4730 | { |
fa1b54e6 | 4731 | struct worker_pool *pool; |
dcd989cb TH |
4732 | unsigned long flags; |
4733 | unsigned int ret = 0; | |
1da177e4 | 4734 | |
dcd989cb TH |
4735 | if (work_pending(work)) |
4736 | ret |= WORK_BUSY_PENDING; | |
1da177e4 | 4737 | |
24acfb71 | 4738 | rcu_read_lock(); |
fa1b54e6 | 4739 | pool = get_work_pool(work); |
038366c5 | 4740 | if (pool) { |
a9b8a985 | 4741 | raw_spin_lock_irqsave(&pool->lock, flags); |
038366c5 LJ |
4742 | if (find_worker_executing_work(pool, work)) |
4743 | ret |= WORK_BUSY_RUNNING; | |
a9b8a985 | 4744 | raw_spin_unlock_irqrestore(&pool->lock, flags); |
038366c5 | 4745 | } |
24acfb71 | 4746 | rcu_read_unlock(); |
1da177e4 | 4747 | |
dcd989cb | 4748 | return ret; |
1da177e4 | 4749 | } |
dcd989cb | 4750 | EXPORT_SYMBOL_GPL(work_busy); |
1da177e4 | 4751 | |
3d1cb205 TH |
4752 | /** |
4753 | * set_worker_desc - set description for the current work item | |
4754 | * @fmt: printf-style format string | |
4755 | * @...: arguments for the format string | |
4756 | * | |
4757 | * This function can be called by a running work function to describe what | |
4758 | * the work item is about. If the worker task gets dumped, this | |
4759 | * information will be printed out together to help debugging. The | |
4760 | * description can be at most WORKER_DESC_LEN including the trailing '\0'. | |
4761 | */ | |
4762 | void set_worker_desc(const char *fmt, ...) | |
4763 | { | |
4764 | struct worker *worker = current_wq_worker(); | |
4765 | va_list args; | |
4766 | ||
4767 | if (worker) { | |
4768 | va_start(args, fmt); | |
4769 | vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); | |
4770 | va_end(args); | |
3d1cb205 TH |
4771 | } |
4772 | } | |
5c750d58 | 4773 | EXPORT_SYMBOL_GPL(set_worker_desc); |
3d1cb205 TH |
4774 | |
4775 | /** | |
4776 | * print_worker_info - print out worker information and description | |
4777 | * @log_lvl: the log level to use when printing | |
4778 | * @task: target task | |
4779 | * | |
4780 | * If @task is a worker and currently executing a work item, print out the | |
4781 | * name of the workqueue being serviced and worker description set with | |
4782 | * set_worker_desc() by the currently executing work item. | |
4783 | * | |
4784 | * This function can be safely called on any task as long as the | |
4785 | * task_struct itself is accessible. While safe, this function isn't | |
4786 | * synchronized and may print out mixups or garbages of limited length. | |
4787 | */ | |
4788 | void print_worker_info(const char *log_lvl, struct task_struct *task) | |
4789 | { | |
4790 | work_func_t *fn = NULL; | |
4791 | char name[WQ_NAME_LEN] = { }; | |
4792 | char desc[WORKER_DESC_LEN] = { }; | |
4793 | struct pool_workqueue *pwq = NULL; | |
4794 | struct workqueue_struct *wq = NULL; | |
3d1cb205 TH |
4795 | struct worker *worker; |
4796 | ||
4797 | if (!(task->flags & PF_WQ_WORKER)) | |
4798 | return; | |
4799 | ||
4800 | /* | |
4801 | * This function is called without any synchronization and @task | |
4802 | * could be in any state. Be careful with dereferences. | |
4803 | */ | |
e700591a | 4804 | worker = kthread_probe_data(task); |
3d1cb205 TH |
4805 | |
4806 | /* | |
8bf89593 TH |
4807 | * Carefully copy the associated workqueue's workfn, name and desc. |
4808 | * Keep the original last '\0' in case the original is garbage. | |
3d1cb205 | 4809 | */ |
fe557319 CH |
4810 | copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn)); |
4811 | copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq)); | |
4812 | copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq)); | |
4813 | copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1); | |
4814 | copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1); | |
3d1cb205 TH |
4815 | |
4816 | if (fn || name[0] || desc[0]) { | |
d75f773c | 4817 | printk("%sWorkqueue: %s %ps", log_lvl, name, fn); |
8bf89593 | 4818 | if (strcmp(name, desc)) |
3d1cb205 TH |
4819 | pr_cont(" (%s)", desc); |
4820 | pr_cont("\n"); | |
4821 | } | |
4822 | } | |
4823 | ||
3494fc30 TH |
4824 | static void pr_cont_pool_info(struct worker_pool *pool) |
4825 | { | |
4826 | pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask); | |
4827 | if (pool->node != NUMA_NO_NODE) | |
4828 | pr_cont(" node=%d", pool->node); | |
4829 | pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice); | |
4830 | } | |
4831 | ||
c76feb0d PM |
4832 | struct pr_cont_work_struct { |
4833 | bool comma; | |
4834 | work_func_t func; | |
4835 | long ctr; | |
4836 | }; | |
4837 | ||
4838 | static void pr_cont_work_flush(bool comma, work_func_t func, struct pr_cont_work_struct *pcwsp) | |
4839 | { | |
4840 | if (!pcwsp->ctr) | |
4841 | goto out_record; | |
4842 | if (func == pcwsp->func) { | |
4843 | pcwsp->ctr++; | |
4844 | return; | |
4845 | } | |
4846 | if (pcwsp->ctr == 1) | |
4847 | pr_cont("%s %ps", pcwsp->comma ? "," : "", pcwsp->func); | |
4848 | else | |
4849 | pr_cont("%s %ld*%ps", pcwsp->comma ? "," : "", pcwsp->ctr, pcwsp->func); | |
4850 | pcwsp->ctr = 0; | |
4851 | out_record: | |
4852 | if ((long)func == -1L) | |
4853 | return; | |
4854 | pcwsp->comma = comma; | |
4855 | pcwsp->func = func; | |
4856 | pcwsp->ctr = 1; | |
4857 | } | |
4858 | ||
4859 | static void pr_cont_work(bool comma, struct work_struct *work, struct pr_cont_work_struct *pcwsp) | |
3494fc30 TH |
4860 | { |
4861 | if (work->func == wq_barrier_func) { | |
4862 | struct wq_barrier *barr; | |
4863 | ||
4864 | barr = container_of(work, struct wq_barrier, work); | |
4865 | ||
c76feb0d | 4866 | pr_cont_work_flush(comma, (work_func_t)-1, pcwsp); |
3494fc30 TH |
4867 | pr_cont("%s BAR(%d)", comma ? "," : "", |
4868 | task_pid_nr(barr->task)); | |
4869 | } else { | |
c76feb0d PM |
4870 | if (!comma) |
4871 | pr_cont_work_flush(comma, (work_func_t)-1, pcwsp); | |
4872 | pr_cont_work_flush(comma, work->func, pcwsp); | |
3494fc30 TH |
4873 | } |
4874 | } | |
4875 | ||
4876 | static void show_pwq(struct pool_workqueue *pwq) | |
4877 | { | |
c76feb0d | 4878 | struct pr_cont_work_struct pcws = { .ctr = 0, }; |
3494fc30 TH |
4879 | struct worker_pool *pool = pwq->pool; |
4880 | struct work_struct *work; | |
4881 | struct worker *worker; | |
4882 | bool has_in_flight = false, has_pending = false; | |
4883 | int bkt; | |
4884 | ||
4885 | pr_info(" pwq %d:", pool->id); | |
4886 | pr_cont_pool_info(pool); | |
4887 | ||
e66b39af TH |
4888 | pr_cont(" active=%d/%d refcnt=%d%s\n", |
4889 | pwq->nr_active, pwq->max_active, pwq->refcnt, | |
3494fc30 TH |
4890 | !list_empty(&pwq->mayday_node) ? " MAYDAY" : ""); |
4891 | ||
4892 | hash_for_each(pool->busy_hash, bkt, worker, hentry) { | |
4893 | if (worker->current_pwq == pwq) { | |
4894 | has_in_flight = true; | |
4895 | break; | |
4896 | } | |
4897 | } | |
4898 | if (has_in_flight) { | |
4899 | bool comma = false; | |
4900 | ||
4901 | pr_info(" in-flight:"); | |
4902 | hash_for_each(pool->busy_hash, bkt, worker, hentry) { | |
4903 | if (worker->current_pwq != pwq) | |
4904 | continue; | |
4905 | ||
d75f773c | 4906 | pr_cont("%s %d%s:%ps", comma ? "," : "", |
3494fc30 | 4907 | task_pid_nr(worker->task), |
30ae2fc0 | 4908 | worker->rescue_wq ? "(RESCUER)" : "", |
3494fc30 TH |
4909 | worker->current_func); |
4910 | list_for_each_entry(work, &worker->scheduled, entry) | |
c76feb0d PM |
4911 | pr_cont_work(false, work, &pcws); |
4912 | pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); | |
3494fc30 TH |
4913 | comma = true; |
4914 | } | |
4915 | pr_cont("\n"); | |
4916 | } | |
4917 | ||
4918 | list_for_each_entry(work, &pool->worklist, entry) { | |
4919 | if (get_work_pwq(work) == pwq) { | |
4920 | has_pending = true; | |
4921 | break; | |
4922 | } | |
4923 | } | |
4924 | if (has_pending) { | |
4925 | bool comma = false; | |
4926 | ||
4927 | pr_info(" pending:"); | |
4928 | list_for_each_entry(work, &pool->worklist, entry) { | |
4929 | if (get_work_pwq(work) != pwq) | |
4930 | continue; | |
4931 | ||
c76feb0d | 4932 | pr_cont_work(comma, work, &pcws); |
3494fc30 TH |
4933 | comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); |
4934 | } | |
c76feb0d | 4935 | pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); |
3494fc30 TH |
4936 | pr_cont("\n"); |
4937 | } | |
4938 | ||
f97a4a1a | 4939 | if (!list_empty(&pwq->inactive_works)) { |
3494fc30 TH |
4940 | bool comma = false; |
4941 | ||
f97a4a1a LJ |
4942 | pr_info(" inactive:"); |
4943 | list_for_each_entry(work, &pwq->inactive_works, entry) { | |
c76feb0d | 4944 | pr_cont_work(comma, work, &pcws); |
3494fc30 TH |
4945 | comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); |
4946 | } | |
c76feb0d | 4947 | pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); |
3494fc30 TH |
4948 | pr_cont("\n"); |
4949 | } | |
4950 | } | |
4951 | ||
4952 | /** | |
55df0933 IK |
4953 | * show_one_workqueue - dump state of specified workqueue |
4954 | * @wq: workqueue whose state will be printed | |
3494fc30 | 4955 | */ |
55df0933 | 4956 | void show_one_workqueue(struct workqueue_struct *wq) |
3494fc30 | 4957 | { |
55df0933 IK |
4958 | struct pool_workqueue *pwq; |
4959 | bool idle = true; | |
3494fc30 | 4960 | unsigned long flags; |
3494fc30 | 4961 | |
55df0933 IK |
4962 | for_each_pwq(pwq, wq) { |
4963 | if (pwq->nr_active || !list_empty(&pwq->inactive_works)) { | |
4964 | idle = false; | |
4965 | break; | |
3494fc30 | 4966 | } |
55df0933 IK |
4967 | } |
4968 | if (idle) /* Nothing to print for idle workqueue */ | |
4969 | return; | |
3494fc30 | 4970 | |
55df0933 | 4971 | pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags); |
3494fc30 | 4972 | |
55df0933 IK |
4973 | for_each_pwq(pwq, wq) { |
4974 | raw_spin_lock_irqsave(&pwq->pool->lock, flags); | |
4975 | if (pwq->nr_active || !list_empty(&pwq->inactive_works)) { | |
62635ea8 | 4976 | /* |
55df0933 IK |
4977 | * Defer printing to avoid deadlocks in console |
4978 | * drivers that queue work while holding locks | |
4979 | * also taken in their write paths. | |
62635ea8 | 4980 | */ |
55df0933 IK |
4981 | printk_deferred_enter(); |
4982 | show_pwq(pwq); | |
4983 | printk_deferred_exit(); | |
3494fc30 | 4984 | } |
55df0933 | 4985 | raw_spin_unlock_irqrestore(&pwq->pool->lock, flags); |
62635ea8 SS |
4986 | /* |
4987 | * We could be printing a lot from atomic context, e.g. | |
55df0933 | 4988 | * sysrq-t -> show_all_workqueues(). Avoid triggering |
62635ea8 SS |
4989 | * hard lockup. |
4990 | */ | |
4991 | touch_nmi_watchdog(); | |
3494fc30 TH |
4992 | } |
4993 | ||
55df0933 IK |
4994 | } |
4995 | ||
4996 | /** | |
4997 | * show_one_worker_pool - dump state of specified worker pool | |
4998 | * @pool: worker pool whose state will be printed | |
4999 | */ | |
5000 | static void show_one_worker_pool(struct worker_pool *pool) | |
5001 | { | |
5002 | struct worker *worker; | |
5003 | bool first = true; | |
5004 | unsigned long flags; | |
5005 | ||
5006 | raw_spin_lock_irqsave(&pool->lock, flags); | |
5007 | if (pool->nr_workers == pool->nr_idle) | |
5008 | goto next_pool; | |
5009 | /* | |
5010 | * Defer printing to avoid deadlocks in console drivers that | |
5011 | * queue work while holding locks also taken in their write | |
5012 | * paths. | |
5013 | */ | |
5014 | printk_deferred_enter(); | |
5015 | pr_info("pool %d:", pool->id); | |
5016 | pr_cont_pool_info(pool); | |
5017 | pr_cont(" hung=%us workers=%d", | |
5018 | jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000, | |
5019 | pool->nr_workers); | |
5020 | if (pool->manager) | |
5021 | pr_cont(" manager: %d", | |
5022 | task_pid_nr(pool->manager->task)); | |
5023 | list_for_each_entry(worker, &pool->idle_list, entry) { | |
5024 | pr_cont(" %s%d", first ? "idle: " : "", | |
5025 | task_pid_nr(worker->task)); | |
5026 | first = false; | |
5027 | } | |
5028 | pr_cont("\n"); | |
5029 | printk_deferred_exit(); | |
5030 | next_pool: | |
5031 | raw_spin_unlock_irqrestore(&pool->lock, flags); | |
5032 | /* | |
5033 | * We could be printing a lot from atomic context, e.g. | |
5034 | * sysrq-t -> show_all_workqueues(). Avoid triggering | |
5035 | * hard lockup. | |
5036 | */ | |
5037 | touch_nmi_watchdog(); | |
5038 | ||
5039 | } | |
5040 | ||
5041 | /** | |
5042 | * show_all_workqueues - dump workqueue state | |
5043 | * | |
5044 | * Called from a sysrq handler or try_to_freeze_tasks() and prints out | |
5045 | * all busy workqueues and pools. | |
5046 | */ | |
5047 | void show_all_workqueues(void) | |
5048 | { | |
5049 | struct workqueue_struct *wq; | |
5050 | struct worker_pool *pool; | |
5051 | int pi; | |
5052 | ||
5053 | rcu_read_lock(); | |
5054 | ||
5055 | pr_info("Showing busy workqueues and worker pools:\n"); | |
5056 | ||
5057 | list_for_each_entry_rcu(wq, &workqueues, list) | |
5058 | show_one_workqueue(wq); | |
5059 | ||
5060 | for_each_pool(pool, pi) | |
5061 | show_one_worker_pool(pool); | |
5062 | ||
24acfb71 | 5063 | rcu_read_unlock(); |
3494fc30 TH |
5064 | } |
5065 | ||
6b59808b TH |
5066 | /* used to show worker information through /proc/PID/{comm,stat,status} */ |
5067 | void wq_worker_comm(char *buf, size_t size, struct task_struct *task) | |
5068 | { | |
6b59808b TH |
5069 | int off; |
5070 | ||
5071 | /* always show the actual comm */ | |
5072 | off = strscpy(buf, task->comm, size); | |
5073 | if (off < 0) | |
5074 | return; | |
5075 | ||
197f6acc | 5076 | /* stabilize PF_WQ_WORKER and worker pool association */ |
6b59808b TH |
5077 | mutex_lock(&wq_pool_attach_mutex); |
5078 | ||
197f6acc TH |
5079 | if (task->flags & PF_WQ_WORKER) { |
5080 | struct worker *worker = kthread_data(task); | |
5081 | struct worker_pool *pool = worker->pool; | |
6b59808b | 5082 | |
197f6acc | 5083 | if (pool) { |
a9b8a985 | 5084 | raw_spin_lock_irq(&pool->lock); |
197f6acc TH |
5085 | /* |
5086 | * ->desc tracks information (wq name or | |
5087 | * set_worker_desc()) for the latest execution. If | |
5088 | * current, prepend '+', otherwise '-'. | |
5089 | */ | |
5090 | if (worker->desc[0] != '\0') { | |
5091 | if (worker->current_work) | |
5092 | scnprintf(buf + off, size - off, "+%s", | |
5093 | worker->desc); | |
5094 | else | |
5095 | scnprintf(buf + off, size - off, "-%s", | |
5096 | worker->desc); | |
5097 | } | |
a9b8a985 | 5098 | raw_spin_unlock_irq(&pool->lock); |
6b59808b | 5099 | } |
6b59808b TH |
5100 | } |
5101 | ||
5102 | mutex_unlock(&wq_pool_attach_mutex); | |
5103 | } | |
5104 | ||
66448bc2 MM |
5105 | #ifdef CONFIG_SMP |
5106 | ||
db7bccf4 TH |
5107 | /* |
5108 | * CPU hotplug. | |
5109 | * | |
e22bee78 | 5110 | * There are two challenges in supporting CPU hotplug. Firstly, there |
112202d9 | 5111 | * are a lot of assumptions on strong associations among work, pwq and |
706026c2 | 5112 | * pool which make migrating pending and scheduled works very |
e22bee78 | 5113 | * difficult to implement without impacting hot paths. Secondly, |
94cf58bb | 5114 | * worker pools serve mix of short, long and very long running works making |
e22bee78 TH |
5115 | * blocked draining impractical. |
5116 | * | |
24647570 | 5117 | * This is solved by allowing the pools to be disassociated from the CPU |
628c78e7 TH |
5118 | * running as an unbound one and allowing it to be reattached later if the |
5119 | * cpu comes back online. | |
db7bccf4 | 5120 | */ |
1da177e4 | 5121 | |
e8b3f8db | 5122 | static void unbind_workers(int cpu) |
3af24433 | 5123 | { |
4ce62e9e | 5124 | struct worker_pool *pool; |
db7bccf4 | 5125 | struct worker *worker; |
3af24433 | 5126 | |
f02ae73a | 5127 | for_each_cpu_worker_pool(pool, cpu) { |
1258fae7 | 5128 | mutex_lock(&wq_pool_attach_mutex); |
a9b8a985 | 5129 | raw_spin_lock_irq(&pool->lock); |
3af24433 | 5130 | |
94cf58bb | 5131 | /* |
92f9c5c4 | 5132 | * We've blocked all attach/detach operations. Make all workers |
94cf58bb | 5133 | * unbound and set DISASSOCIATED. Before this, all workers |
11b45b0b | 5134 | * must be on the cpu. After this, they may become diasporas. |
b4ac9384 LJ |
5135 | * And the preemption disabled section in their sched callbacks |
5136 | * are guaranteed to see WORKER_UNBOUND since the code here | |
5137 | * is on the same cpu. | |
94cf58bb | 5138 | */ |
da028469 | 5139 | for_each_pool_worker(worker, pool) |
c9e7cf27 | 5140 | worker->flags |= WORKER_UNBOUND; |
06ba38a9 | 5141 | |
24647570 | 5142 | pool->flags |= POOL_DISASSOCIATED; |
f2d5a0ee | 5143 | |
eb283428 | 5144 | /* |
989442d7 LJ |
5145 | * The handling of nr_running in sched callbacks are disabled |
5146 | * now. Zap nr_running. After this, nr_running stays zero and | |
5147 | * need_more_worker() and keep_working() are always true as | |
5148 | * long as the worklist is not empty. This pool now behaves as | |
5149 | * an unbound (in terms of concurrency management) pool which | |
eb283428 LJ |
5150 | * are served by workers tied to the pool. |
5151 | */ | |
bc35f7ef | 5152 | pool->nr_running = 0; |
eb283428 LJ |
5153 | |
5154 | /* | |
5155 | * With concurrency management just turned off, a busy | |
5156 | * worker blocking could lead to lengthy stalls. Kick off | |
5157 | * unbound chain execution of currently pending work items. | |
5158 | */ | |
eb283428 | 5159 | wake_up_worker(pool); |
989442d7 | 5160 | |
a9b8a985 | 5161 | raw_spin_unlock_irq(&pool->lock); |
989442d7 | 5162 | |
793777bc VS |
5163 | for_each_pool_worker(worker, pool) |
5164 | unbind_worker(worker); | |
989442d7 LJ |
5165 | |
5166 | mutex_unlock(&wq_pool_attach_mutex); | |
eb283428 | 5167 | } |
3af24433 | 5168 | } |
3af24433 | 5169 | |
bd7c089e TH |
5170 | /** |
5171 | * rebind_workers - rebind all workers of a pool to the associated CPU | |
5172 | * @pool: pool of interest | |
5173 | * | |
a9ab775b | 5174 | * @pool->cpu is coming online. Rebind all workers to the CPU. |
bd7c089e TH |
5175 | */ |
5176 | static void rebind_workers(struct worker_pool *pool) | |
5177 | { | |
a9ab775b | 5178 | struct worker *worker; |
bd7c089e | 5179 | |
1258fae7 | 5180 | lockdep_assert_held(&wq_pool_attach_mutex); |
bd7c089e | 5181 | |
a9ab775b TH |
5182 | /* |
5183 | * Restore CPU affinity of all workers. As all idle workers should | |
5184 | * be on the run-queue of the associated CPU before any local | |
402dd89d | 5185 | * wake-ups for concurrency management happen, restore CPU affinity |
a9ab775b TH |
5186 | * of all workers first and then clear UNBOUND. As we're called |
5187 | * from CPU_ONLINE, the following shouldn't fail. | |
5188 | */ | |
c63a2e52 VS |
5189 | for_each_pool_worker(worker, pool) { |
5190 | kthread_set_per_cpu(worker->task, pool->cpu); | |
5191 | WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, | |
5192 | pool->attrs->cpumask) < 0); | |
5193 | } | |
bd7c089e | 5194 | |
a9b8a985 | 5195 | raw_spin_lock_irq(&pool->lock); |
f7c17d26 | 5196 | |
3de5e884 | 5197 | pool->flags &= ~POOL_DISASSOCIATED; |
bd7c089e | 5198 | |
da028469 | 5199 | for_each_pool_worker(worker, pool) { |
a9ab775b | 5200 | unsigned int worker_flags = worker->flags; |
bd7c089e | 5201 | |
a9ab775b TH |
5202 | /* |
5203 | * We want to clear UNBOUND but can't directly call | |
5204 | * worker_clr_flags() or adjust nr_running. Atomically | |
5205 | * replace UNBOUND with another NOT_RUNNING flag REBOUND. | |
5206 | * @worker will clear REBOUND using worker_clr_flags() when | |
5207 | * it initiates the next execution cycle thus restoring | |
5208 | * concurrency management. Note that when or whether | |
5209 | * @worker clears REBOUND doesn't affect correctness. | |
5210 | * | |
c95491ed | 5211 | * WRITE_ONCE() is necessary because @worker->flags may be |
a9ab775b | 5212 | * tested without holding any lock in |
6d25be57 | 5213 | * wq_worker_running(). Without it, NOT_RUNNING test may |
a9ab775b TH |
5214 | * fail incorrectly leading to premature concurrency |
5215 | * management operations. | |
5216 | */ | |
5217 | WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); | |
5218 | worker_flags |= WORKER_REBOUND; | |
5219 | worker_flags &= ~WORKER_UNBOUND; | |
c95491ed | 5220 | WRITE_ONCE(worker->flags, worker_flags); |
bd7c089e | 5221 | } |
a9ab775b | 5222 | |
a9b8a985 | 5223 | raw_spin_unlock_irq(&pool->lock); |
bd7c089e TH |
5224 | } |
5225 | ||
7dbc725e TH |
5226 | /** |
5227 | * restore_unbound_workers_cpumask - restore cpumask of unbound workers | |
5228 | * @pool: unbound pool of interest | |
5229 | * @cpu: the CPU which is coming up | |
5230 | * | |
5231 | * An unbound pool may end up with a cpumask which doesn't have any online | |
5232 | * CPUs. When a worker of such pool get scheduled, the scheduler resets | |
5233 | * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any | |
5234 | * online CPU before, cpus_allowed of all its workers should be restored. | |
5235 | */ | |
5236 | static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) | |
5237 | { | |
5238 | static cpumask_t cpumask; | |
5239 | struct worker *worker; | |
7dbc725e | 5240 | |
1258fae7 | 5241 | lockdep_assert_held(&wq_pool_attach_mutex); |
7dbc725e TH |
5242 | |
5243 | /* is @cpu allowed for @pool? */ | |
5244 | if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) | |
5245 | return; | |
5246 | ||
7dbc725e | 5247 | cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); |
7dbc725e TH |
5248 | |
5249 | /* as we're called from CPU_ONLINE, the following shouldn't fail */ | |
da028469 | 5250 | for_each_pool_worker(worker, pool) |
d945b5e9 | 5251 | WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0); |
7dbc725e TH |
5252 | } |
5253 | ||
7ee681b2 TG |
5254 | int workqueue_prepare_cpu(unsigned int cpu) |
5255 | { | |
5256 | struct worker_pool *pool; | |
5257 | ||
5258 | for_each_cpu_worker_pool(pool, cpu) { | |
5259 | if (pool->nr_workers) | |
5260 | continue; | |
5261 | if (!create_worker(pool)) | |
5262 | return -ENOMEM; | |
5263 | } | |
5264 | return 0; | |
5265 | } | |
5266 | ||
5267 | int workqueue_online_cpu(unsigned int cpu) | |
3af24433 | 5268 | { |
4ce62e9e | 5269 | struct worker_pool *pool; |
4c16bd32 | 5270 | struct workqueue_struct *wq; |
7dbc725e | 5271 | int pi; |
3ce63377 | 5272 | |
7ee681b2 | 5273 | mutex_lock(&wq_pool_mutex); |
7dbc725e | 5274 | |
7ee681b2 | 5275 | for_each_pool(pool, pi) { |
1258fae7 | 5276 | mutex_lock(&wq_pool_attach_mutex); |
94cf58bb | 5277 | |
7ee681b2 TG |
5278 | if (pool->cpu == cpu) |
5279 | rebind_workers(pool); | |
5280 | else if (pool->cpu < 0) | |
5281 | restore_unbound_workers_cpumask(pool, cpu); | |
94cf58bb | 5282 | |
1258fae7 | 5283 | mutex_unlock(&wq_pool_attach_mutex); |
7ee681b2 | 5284 | } |
6ba94429 | 5285 | |
7ee681b2 TG |
5286 | /* update NUMA affinity of unbound workqueues */ |
5287 | list_for_each_entry(wq, &workqueues, list) | |
5288 | wq_update_unbound_numa(wq, cpu, true); | |
6ba94429 | 5289 | |
7ee681b2 TG |
5290 | mutex_unlock(&wq_pool_mutex); |
5291 | return 0; | |
6ba94429 FW |
5292 | } |
5293 | ||
7ee681b2 | 5294 | int workqueue_offline_cpu(unsigned int cpu) |
6ba94429 | 5295 | { |
6ba94429 FW |
5296 | struct workqueue_struct *wq; |
5297 | ||
7ee681b2 | 5298 | /* unbinding per-cpu workers should happen on the local CPU */ |
e8b3f8db LJ |
5299 | if (WARN_ON(cpu != smp_processor_id())) |
5300 | return -1; | |
5301 | ||
5302 | unbind_workers(cpu); | |
7ee681b2 TG |
5303 | |
5304 | /* update NUMA affinity of unbound workqueues */ | |
5305 | mutex_lock(&wq_pool_mutex); | |
5306 | list_for_each_entry(wq, &workqueues, list) | |
5307 | wq_update_unbound_numa(wq, cpu, false); | |
5308 | mutex_unlock(&wq_pool_mutex); | |
5309 | ||
7ee681b2 | 5310 | return 0; |
6ba94429 FW |
5311 | } |
5312 | ||
6ba94429 FW |
5313 | struct work_for_cpu { |
5314 | struct work_struct work; | |
5315 | long (*fn)(void *); | |
5316 | void *arg; | |
5317 | long ret; | |
5318 | }; | |
5319 | ||
5320 | static void work_for_cpu_fn(struct work_struct *work) | |
5321 | { | |
5322 | struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); | |
5323 | ||
5324 | wfc->ret = wfc->fn(wfc->arg); | |
5325 | } | |
5326 | ||
5327 | /** | |
22aceb31 | 5328 | * work_on_cpu - run a function in thread context on a particular cpu |
6ba94429 FW |
5329 | * @cpu: the cpu to run on |
5330 | * @fn: the function to run | |
5331 | * @arg: the function arg | |
5332 | * | |
5333 | * It is up to the caller to ensure that the cpu doesn't go offline. | |
5334 | * The caller must not hold any locks which would prevent @fn from completing. | |
5335 | * | |
5336 | * Return: The value @fn returns. | |
5337 | */ | |
5338 | long work_on_cpu(int cpu, long (*fn)(void *), void *arg) | |
5339 | { | |
5340 | struct work_for_cpu wfc = { .fn = fn, .arg = arg }; | |
5341 | ||
5342 | INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); | |
5343 | schedule_work_on(cpu, &wfc.work); | |
5344 | flush_work(&wfc.work); | |
5345 | destroy_work_on_stack(&wfc.work); | |
5346 | return wfc.ret; | |
5347 | } | |
5348 | EXPORT_SYMBOL_GPL(work_on_cpu); | |
0e8d6a93 TG |
5349 | |
5350 | /** | |
5351 | * work_on_cpu_safe - run a function in thread context on a particular cpu | |
5352 | * @cpu: the cpu to run on | |
5353 | * @fn: the function to run | |
5354 | * @arg: the function argument | |
5355 | * | |
5356 | * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold | |
5357 | * any locks which would prevent @fn from completing. | |
5358 | * | |
5359 | * Return: The value @fn returns. | |
5360 | */ | |
5361 | long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg) | |
5362 | { | |
5363 | long ret = -ENODEV; | |
5364 | ||
ffd8bea8 | 5365 | cpus_read_lock(); |
0e8d6a93 TG |
5366 | if (cpu_online(cpu)) |
5367 | ret = work_on_cpu(cpu, fn, arg); | |
ffd8bea8 | 5368 | cpus_read_unlock(); |
0e8d6a93 TG |
5369 | return ret; |
5370 | } | |
5371 | EXPORT_SYMBOL_GPL(work_on_cpu_safe); | |
6ba94429 FW |
5372 | #endif /* CONFIG_SMP */ |
5373 | ||
5374 | #ifdef CONFIG_FREEZER | |
5375 | ||
5376 | /** | |
5377 | * freeze_workqueues_begin - begin freezing workqueues | |
5378 | * | |
5379 | * Start freezing workqueues. After this function returns, all freezable | |
f97a4a1a | 5380 | * workqueues will queue new works to their inactive_works list instead of |
6ba94429 FW |
5381 | * pool->worklist. |
5382 | * | |
5383 | * CONTEXT: | |
5384 | * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. | |
5385 | */ | |
5386 | void freeze_workqueues_begin(void) | |
5387 | { | |
5388 | struct workqueue_struct *wq; | |
5389 | struct pool_workqueue *pwq; | |
5390 | ||
5391 | mutex_lock(&wq_pool_mutex); | |
5392 | ||
5393 | WARN_ON_ONCE(workqueue_freezing); | |
5394 | workqueue_freezing = true; | |
5395 | ||
5396 | list_for_each_entry(wq, &workqueues, list) { | |
5397 | mutex_lock(&wq->mutex); | |
5398 | for_each_pwq(pwq, wq) | |
5399 | pwq_adjust_max_active(pwq); | |
5400 | mutex_unlock(&wq->mutex); | |
5401 | } | |
5402 | ||
5403 | mutex_unlock(&wq_pool_mutex); | |
5404 | } | |
5405 | ||
5406 | /** | |
5407 | * freeze_workqueues_busy - are freezable workqueues still busy? | |
5408 | * | |
5409 | * Check whether freezing is complete. This function must be called | |
5410 | * between freeze_workqueues_begin() and thaw_workqueues(). | |
5411 | * | |
5412 | * CONTEXT: | |
5413 | * Grabs and releases wq_pool_mutex. | |
5414 | * | |
5415 | * Return: | |
5416 | * %true if some freezable workqueues are still busy. %false if freezing | |
5417 | * is complete. | |
5418 | */ | |
5419 | bool freeze_workqueues_busy(void) | |
5420 | { | |
5421 | bool busy = false; | |
5422 | struct workqueue_struct *wq; | |
5423 | struct pool_workqueue *pwq; | |
5424 | ||
5425 | mutex_lock(&wq_pool_mutex); | |
5426 | ||
5427 | WARN_ON_ONCE(!workqueue_freezing); | |
5428 | ||
5429 | list_for_each_entry(wq, &workqueues, list) { | |
5430 | if (!(wq->flags & WQ_FREEZABLE)) | |
5431 | continue; | |
5432 | /* | |
5433 | * nr_active is monotonically decreasing. It's safe | |
5434 | * to peek without lock. | |
5435 | */ | |
24acfb71 | 5436 | rcu_read_lock(); |
6ba94429 FW |
5437 | for_each_pwq(pwq, wq) { |
5438 | WARN_ON_ONCE(pwq->nr_active < 0); | |
5439 | if (pwq->nr_active) { | |
5440 | busy = true; | |
24acfb71 | 5441 | rcu_read_unlock(); |
6ba94429 FW |
5442 | goto out_unlock; |
5443 | } | |
5444 | } | |
24acfb71 | 5445 | rcu_read_unlock(); |
6ba94429 FW |
5446 | } |
5447 | out_unlock: | |
5448 | mutex_unlock(&wq_pool_mutex); | |
5449 | return busy; | |
5450 | } | |
5451 | ||
5452 | /** | |
5453 | * thaw_workqueues - thaw workqueues | |
5454 | * | |
5455 | * Thaw workqueues. Normal queueing is restored and all collected | |
5456 | * frozen works are transferred to their respective pool worklists. | |
5457 | * | |
5458 | * CONTEXT: | |
5459 | * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. | |
5460 | */ | |
5461 | void thaw_workqueues(void) | |
5462 | { | |
5463 | struct workqueue_struct *wq; | |
5464 | struct pool_workqueue *pwq; | |
5465 | ||
5466 | mutex_lock(&wq_pool_mutex); | |
5467 | ||
5468 | if (!workqueue_freezing) | |
5469 | goto out_unlock; | |
5470 | ||
5471 | workqueue_freezing = false; | |
5472 | ||
5473 | /* restore max_active and repopulate worklist */ | |
5474 | list_for_each_entry(wq, &workqueues, list) { | |
5475 | mutex_lock(&wq->mutex); | |
5476 | for_each_pwq(pwq, wq) | |
5477 | pwq_adjust_max_active(pwq); | |
5478 | mutex_unlock(&wq->mutex); | |
5479 | } | |
5480 | ||
5481 | out_unlock: | |
5482 | mutex_unlock(&wq_pool_mutex); | |
5483 | } | |
5484 | #endif /* CONFIG_FREEZER */ | |
5485 | ||
99c621ef | 5486 | static int workqueue_apply_unbound_cpumask(const cpumask_var_t unbound_cpumask) |
042f7df1 LJ |
5487 | { |
5488 | LIST_HEAD(ctxs); | |
5489 | int ret = 0; | |
5490 | struct workqueue_struct *wq; | |
5491 | struct apply_wqattrs_ctx *ctx, *n; | |
5492 | ||
5493 | lockdep_assert_held(&wq_pool_mutex); | |
5494 | ||
5495 | list_for_each_entry(wq, &workqueues, list) { | |
5496 | if (!(wq->flags & WQ_UNBOUND)) | |
5497 | continue; | |
5498 | /* creating multiple pwqs breaks ordering guarantee */ | |
5499 | if (wq->flags & __WQ_ORDERED) | |
5500 | continue; | |
5501 | ||
99c621ef | 5502 | ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs, unbound_cpumask); |
042f7df1 LJ |
5503 | if (!ctx) { |
5504 | ret = -ENOMEM; | |
5505 | break; | |
5506 | } | |
5507 | ||
5508 | list_add_tail(&ctx->list, &ctxs); | |
5509 | } | |
5510 | ||
5511 | list_for_each_entry_safe(ctx, n, &ctxs, list) { | |
5512 | if (!ret) | |
5513 | apply_wqattrs_commit(ctx); | |
5514 | apply_wqattrs_cleanup(ctx); | |
5515 | } | |
5516 | ||
99c621ef LJ |
5517 | if (!ret) { |
5518 | mutex_lock(&wq_pool_attach_mutex); | |
5519 | cpumask_copy(wq_unbound_cpumask, unbound_cpumask); | |
5520 | mutex_unlock(&wq_pool_attach_mutex); | |
5521 | } | |
042f7df1 LJ |
5522 | return ret; |
5523 | } | |
5524 | ||
5525 | /** | |
5526 | * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask | |
5527 | * @cpumask: the cpumask to set | |
5528 | * | |
5529 | * The low-level workqueues cpumask is a global cpumask that limits | |
5530 | * the affinity of all unbound workqueues. This function check the @cpumask | |
5531 | * and apply it to all unbound workqueues and updates all pwqs of them. | |
5532 | * | |
67dc8325 | 5533 | * Return: 0 - Success |
042f7df1 LJ |
5534 | * -EINVAL - Invalid @cpumask |
5535 | * -ENOMEM - Failed to allocate memory for attrs or pwqs. | |
5536 | */ | |
5537 | int workqueue_set_unbound_cpumask(cpumask_var_t cpumask) | |
5538 | { | |
5539 | int ret = -EINVAL; | |
042f7df1 | 5540 | |
c98a9805 TS |
5541 | /* |
5542 | * Not excluding isolated cpus on purpose. | |
5543 | * If the user wishes to include them, we allow that. | |
5544 | */ | |
042f7df1 LJ |
5545 | cpumask_and(cpumask, cpumask, cpu_possible_mask); |
5546 | if (!cpumask_empty(cpumask)) { | |
a0111cf6 | 5547 | apply_wqattrs_lock(); |
d25302e4 MD |
5548 | if (cpumask_equal(cpumask, wq_unbound_cpumask)) { |
5549 | ret = 0; | |
5550 | goto out_unlock; | |
5551 | } | |
5552 | ||
99c621ef | 5553 | ret = workqueue_apply_unbound_cpumask(cpumask); |
042f7df1 | 5554 | |
d25302e4 | 5555 | out_unlock: |
a0111cf6 | 5556 | apply_wqattrs_unlock(); |
042f7df1 | 5557 | } |
042f7df1 | 5558 | |
042f7df1 LJ |
5559 | return ret; |
5560 | } | |
5561 | ||
6ba94429 FW |
5562 | #ifdef CONFIG_SYSFS |
5563 | /* | |
5564 | * Workqueues with WQ_SYSFS flag set is visible to userland via | |
5565 | * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the | |
5566 | * following attributes. | |
5567 | * | |
5568 | * per_cpu RO bool : whether the workqueue is per-cpu or unbound | |
5569 | * max_active RW int : maximum number of in-flight work items | |
5570 | * | |
5571 | * Unbound workqueues have the following extra attributes. | |
5572 | * | |
9a19b463 | 5573 | * pool_ids RO int : the associated pool IDs for each node |
6ba94429 FW |
5574 | * nice RW int : nice value of the workers |
5575 | * cpumask RW mask : bitmask of allowed CPUs for the workers | |
9a19b463 | 5576 | * numa RW bool : whether enable NUMA affinity |
6ba94429 FW |
5577 | */ |
5578 | struct wq_device { | |
5579 | struct workqueue_struct *wq; | |
5580 | struct device dev; | |
5581 | }; | |
5582 | ||
5583 | static struct workqueue_struct *dev_to_wq(struct device *dev) | |
5584 | { | |
5585 | struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); | |
5586 | ||
5587 | return wq_dev->wq; | |
5588 | } | |
5589 | ||
5590 | static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, | |
5591 | char *buf) | |
5592 | { | |
5593 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5594 | ||
5595 | return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); | |
5596 | } | |
5597 | static DEVICE_ATTR_RO(per_cpu); | |
5598 | ||
5599 | static ssize_t max_active_show(struct device *dev, | |
5600 | struct device_attribute *attr, char *buf) | |
5601 | { | |
5602 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5603 | ||
5604 | return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); | |
5605 | } | |
5606 | ||
5607 | static ssize_t max_active_store(struct device *dev, | |
5608 | struct device_attribute *attr, const char *buf, | |
5609 | size_t count) | |
5610 | { | |
5611 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5612 | int val; | |
5613 | ||
5614 | if (sscanf(buf, "%d", &val) != 1 || val <= 0) | |
5615 | return -EINVAL; | |
5616 | ||
5617 | workqueue_set_max_active(wq, val); | |
5618 | return count; | |
5619 | } | |
5620 | static DEVICE_ATTR_RW(max_active); | |
5621 | ||
5622 | static struct attribute *wq_sysfs_attrs[] = { | |
5623 | &dev_attr_per_cpu.attr, | |
5624 | &dev_attr_max_active.attr, | |
5625 | NULL, | |
5626 | }; | |
5627 | ATTRIBUTE_GROUPS(wq_sysfs); | |
5628 | ||
5629 | static ssize_t wq_pool_ids_show(struct device *dev, | |
5630 | struct device_attribute *attr, char *buf) | |
5631 | { | |
5632 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5633 | const char *delim = ""; | |
5634 | int node, written = 0; | |
5635 | ||
ffd8bea8 | 5636 | cpus_read_lock(); |
24acfb71 | 5637 | rcu_read_lock(); |
6ba94429 FW |
5638 | for_each_node(node) { |
5639 | written += scnprintf(buf + written, PAGE_SIZE - written, | |
5640 | "%s%d:%d", delim, node, | |
5641 | unbound_pwq_by_node(wq, node)->pool->id); | |
5642 | delim = " "; | |
5643 | } | |
5644 | written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); | |
24acfb71 | 5645 | rcu_read_unlock(); |
ffd8bea8 | 5646 | cpus_read_unlock(); |
6ba94429 FW |
5647 | |
5648 | return written; | |
5649 | } | |
5650 | ||
5651 | static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, | |
5652 | char *buf) | |
5653 | { | |
5654 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5655 | int written; | |
5656 | ||
5657 | mutex_lock(&wq->mutex); | |
5658 | written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); | |
5659 | mutex_unlock(&wq->mutex); | |
5660 | ||
5661 | return written; | |
5662 | } | |
5663 | ||
5664 | /* prepare workqueue_attrs for sysfs store operations */ | |
5665 | static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) | |
5666 | { | |
5667 | struct workqueue_attrs *attrs; | |
5668 | ||
899a94fe LJ |
5669 | lockdep_assert_held(&wq_pool_mutex); |
5670 | ||
be69d00d | 5671 | attrs = alloc_workqueue_attrs(); |
6ba94429 FW |
5672 | if (!attrs) |
5673 | return NULL; | |
5674 | ||
6ba94429 | 5675 | copy_workqueue_attrs(attrs, wq->unbound_attrs); |
6ba94429 FW |
5676 | return attrs; |
5677 | } | |
5678 | ||
5679 | static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, | |
5680 | const char *buf, size_t count) | |
5681 | { | |
5682 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5683 | struct workqueue_attrs *attrs; | |
d4d3e257 LJ |
5684 | int ret = -ENOMEM; |
5685 | ||
5686 | apply_wqattrs_lock(); | |
6ba94429 FW |
5687 | |
5688 | attrs = wq_sysfs_prep_attrs(wq); | |
5689 | if (!attrs) | |
d4d3e257 | 5690 | goto out_unlock; |
6ba94429 FW |
5691 | |
5692 | if (sscanf(buf, "%d", &attrs->nice) == 1 && | |
5693 | attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) | |
d4d3e257 | 5694 | ret = apply_workqueue_attrs_locked(wq, attrs); |
6ba94429 FW |
5695 | else |
5696 | ret = -EINVAL; | |
5697 | ||
d4d3e257 LJ |
5698 | out_unlock: |
5699 | apply_wqattrs_unlock(); | |
6ba94429 FW |
5700 | free_workqueue_attrs(attrs); |
5701 | return ret ?: count; | |
5702 | } | |
5703 | ||
5704 | static ssize_t wq_cpumask_show(struct device *dev, | |
5705 | struct device_attribute *attr, char *buf) | |
5706 | { | |
5707 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5708 | int written; | |
5709 | ||
5710 | mutex_lock(&wq->mutex); | |
5711 | written = scnprintf(buf, PAGE_SIZE, "%*pb\n", | |
5712 | cpumask_pr_args(wq->unbound_attrs->cpumask)); | |
5713 | mutex_unlock(&wq->mutex); | |
5714 | return written; | |
5715 | } | |
5716 | ||
5717 | static ssize_t wq_cpumask_store(struct device *dev, | |
5718 | struct device_attribute *attr, | |
5719 | const char *buf, size_t count) | |
5720 | { | |
5721 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5722 | struct workqueue_attrs *attrs; | |
d4d3e257 LJ |
5723 | int ret = -ENOMEM; |
5724 | ||
5725 | apply_wqattrs_lock(); | |
6ba94429 FW |
5726 | |
5727 | attrs = wq_sysfs_prep_attrs(wq); | |
5728 | if (!attrs) | |
d4d3e257 | 5729 | goto out_unlock; |
6ba94429 FW |
5730 | |
5731 | ret = cpumask_parse(buf, attrs->cpumask); | |
5732 | if (!ret) | |
d4d3e257 | 5733 | ret = apply_workqueue_attrs_locked(wq, attrs); |
6ba94429 | 5734 | |
d4d3e257 LJ |
5735 | out_unlock: |
5736 | apply_wqattrs_unlock(); | |
6ba94429 FW |
5737 | free_workqueue_attrs(attrs); |
5738 | return ret ?: count; | |
5739 | } | |
5740 | ||
5741 | static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr, | |
5742 | char *buf) | |
5743 | { | |
5744 | struct workqueue_struct *wq = dev_to_wq(dev); | |
5745 | int written; | |
7dbc725e | 5746 | |
6ba94429 FW |
5747 | mutex_lock(&wq->mutex); |
5748 | written = scnprintf(buf, PAGE_SIZE, "%d\n", | |
5749 | !wq->unbound_attrs->no_numa); | |
5750 | mutex_unlock(&wq->mutex); | |
4c16bd32 | 5751 | |
6ba94429 | 5752 | return written; |
65758202 TH |
5753 | } |
5754 | ||
6ba94429 FW |
5755 | static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr, |
5756 | const char *buf, size_t count) | |
65758202 | 5757 | { |
6ba94429 FW |
5758 | struct workqueue_struct *wq = dev_to_wq(dev); |
5759 | struct workqueue_attrs *attrs; | |
d4d3e257 LJ |
5760 | int v, ret = -ENOMEM; |
5761 | ||
5762 | apply_wqattrs_lock(); | |
4c16bd32 | 5763 | |
6ba94429 FW |
5764 | attrs = wq_sysfs_prep_attrs(wq); |
5765 | if (!attrs) | |
d4d3e257 | 5766 | goto out_unlock; |
4c16bd32 | 5767 | |
6ba94429 FW |
5768 | ret = -EINVAL; |
5769 | if (sscanf(buf, "%d", &v) == 1) { | |
5770 | attrs->no_numa = !v; | |
d4d3e257 | 5771 | ret = apply_workqueue_attrs_locked(wq, attrs); |
65758202 | 5772 | } |
6ba94429 | 5773 | |
d4d3e257 LJ |
5774 | out_unlock: |
5775 | apply_wqattrs_unlock(); | |
6ba94429 FW |
5776 | free_workqueue_attrs(attrs); |
5777 | return ret ?: count; | |
65758202 TH |
5778 | } |
5779 | ||
6ba94429 FW |
5780 | static struct device_attribute wq_sysfs_unbound_attrs[] = { |
5781 | __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL), | |
5782 | __ATTR(nice, 0644, wq_nice_show, wq_nice_store), | |
5783 | __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), | |
5784 | __ATTR(numa, 0644, wq_numa_show, wq_numa_store), | |
5785 | __ATTR_NULL, | |
5786 | }; | |
8ccad40d | 5787 | |
6ba94429 FW |
5788 | static struct bus_type wq_subsys = { |
5789 | .name = "workqueue", | |
5790 | .dev_groups = wq_sysfs_groups, | |
2d3854a3 RR |
5791 | }; |
5792 | ||
b05a7928 FW |
5793 | static ssize_t wq_unbound_cpumask_show(struct device *dev, |
5794 | struct device_attribute *attr, char *buf) | |
5795 | { | |
5796 | int written; | |
5797 | ||
042f7df1 | 5798 | mutex_lock(&wq_pool_mutex); |
b05a7928 FW |
5799 | written = scnprintf(buf, PAGE_SIZE, "%*pb\n", |
5800 | cpumask_pr_args(wq_unbound_cpumask)); | |
042f7df1 | 5801 | mutex_unlock(&wq_pool_mutex); |
b05a7928 FW |
5802 | |
5803 | return written; | |
5804 | } | |
5805 | ||
042f7df1 LJ |
5806 | static ssize_t wq_unbound_cpumask_store(struct device *dev, |
5807 | struct device_attribute *attr, const char *buf, size_t count) | |
5808 | { | |
5809 | cpumask_var_t cpumask; | |
5810 | int ret; | |
5811 | ||
5812 | if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) | |
5813 | return -ENOMEM; | |
5814 | ||
5815 | ret = cpumask_parse(buf, cpumask); | |
5816 | if (!ret) | |
5817 | ret = workqueue_set_unbound_cpumask(cpumask); | |
5818 | ||
5819 | free_cpumask_var(cpumask); | |
5820 | return ret ? ret : count; | |
5821 | } | |
5822 | ||
b05a7928 | 5823 | static struct device_attribute wq_sysfs_cpumask_attr = |
042f7df1 LJ |
5824 | __ATTR(cpumask, 0644, wq_unbound_cpumask_show, |
5825 | wq_unbound_cpumask_store); | |
b05a7928 | 5826 | |
6ba94429 | 5827 | static int __init wq_sysfs_init(void) |
2d3854a3 | 5828 | { |
b05a7928 FW |
5829 | int err; |
5830 | ||
5831 | err = subsys_virtual_register(&wq_subsys, NULL); | |
5832 | if (err) | |
5833 | return err; | |
5834 | ||
5835 | return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr); | |
2d3854a3 | 5836 | } |
6ba94429 | 5837 | core_initcall(wq_sysfs_init); |
2d3854a3 | 5838 | |
6ba94429 | 5839 | static void wq_device_release(struct device *dev) |
2d3854a3 | 5840 | { |
6ba94429 | 5841 | struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); |
6b44003e | 5842 | |
6ba94429 | 5843 | kfree(wq_dev); |
2d3854a3 | 5844 | } |
a0a1a5fd TH |
5845 | |
5846 | /** | |
6ba94429 FW |
5847 | * workqueue_sysfs_register - make a workqueue visible in sysfs |
5848 | * @wq: the workqueue to register | |
a0a1a5fd | 5849 | * |
6ba94429 FW |
5850 | * Expose @wq in sysfs under /sys/bus/workqueue/devices. |
5851 | * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set | |
5852 | * which is the preferred method. | |
a0a1a5fd | 5853 | * |
6ba94429 FW |
5854 | * Workqueue user should use this function directly iff it wants to apply |
5855 | * workqueue_attrs before making the workqueue visible in sysfs; otherwise, | |
5856 | * apply_workqueue_attrs() may race against userland updating the | |
5857 | * attributes. | |
5858 | * | |
5859 | * Return: 0 on success, -errno on failure. | |
a0a1a5fd | 5860 | */ |
6ba94429 | 5861 | int workqueue_sysfs_register(struct workqueue_struct *wq) |
a0a1a5fd | 5862 | { |
6ba94429 FW |
5863 | struct wq_device *wq_dev; |
5864 | int ret; | |
a0a1a5fd | 5865 | |
6ba94429 | 5866 | /* |
402dd89d | 5867 | * Adjusting max_active or creating new pwqs by applying |
6ba94429 FW |
5868 | * attributes breaks ordering guarantee. Disallow exposing ordered |
5869 | * workqueues. | |
5870 | */ | |
0a94efb5 | 5871 | if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) |
6ba94429 | 5872 | return -EINVAL; |
a0a1a5fd | 5873 | |
6ba94429 FW |
5874 | wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); |
5875 | if (!wq_dev) | |
5876 | return -ENOMEM; | |
5bcab335 | 5877 | |
6ba94429 FW |
5878 | wq_dev->wq = wq; |
5879 | wq_dev->dev.bus = &wq_subsys; | |
6ba94429 | 5880 | wq_dev->dev.release = wq_device_release; |
23217b44 | 5881 | dev_set_name(&wq_dev->dev, "%s", wq->name); |
a0a1a5fd | 5882 | |
6ba94429 FW |
5883 | /* |
5884 | * unbound_attrs are created separately. Suppress uevent until | |
5885 | * everything is ready. | |
5886 | */ | |
5887 | dev_set_uevent_suppress(&wq_dev->dev, true); | |
a0a1a5fd | 5888 | |
6ba94429 FW |
5889 | ret = device_register(&wq_dev->dev); |
5890 | if (ret) { | |
537f4146 | 5891 | put_device(&wq_dev->dev); |
6ba94429 FW |
5892 | wq->wq_dev = NULL; |
5893 | return ret; | |
5894 | } | |
a0a1a5fd | 5895 | |
6ba94429 FW |
5896 | if (wq->flags & WQ_UNBOUND) { |
5897 | struct device_attribute *attr; | |
a0a1a5fd | 5898 | |
6ba94429 FW |
5899 | for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { |
5900 | ret = device_create_file(&wq_dev->dev, attr); | |
5901 | if (ret) { | |
5902 | device_unregister(&wq_dev->dev); | |
5903 | wq->wq_dev = NULL; | |
5904 | return ret; | |
a0a1a5fd TH |
5905 | } |
5906 | } | |
5907 | } | |
6ba94429 FW |
5908 | |
5909 | dev_set_uevent_suppress(&wq_dev->dev, false); | |
5910 | kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); | |
5911 | return 0; | |
a0a1a5fd TH |
5912 | } |
5913 | ||
5914 | /** | |
6ba94429 FW |
5915 | * workqueue_sysfs_unregister - undo workqueue_sysfs_register() |
5916 | * @wq: the workqueue to unregister | |
a0a1a5fd | 5917 | * |
6ba94429 | 5918 | * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. |
a0a1a5fd | 5919 | */ |
6ba94429 | 5920 | static void workqueue_sysfs_unregister(struct workqueue_struct *wq) |
a0a1a5fd | 5921 | { |
6ba94429 | 5922 | struct wq_device *wq_dev = wq->wq_dev; |
8b03ae3c | 5923 | |
6ba94429 FW |
5924 | if (!wq->wq_dev) |
5925 | return; | |
a0a1a5fd | 5926 | |
6ba94429 FW |
5927 | wq->wq_dev = NULL; |
5928 | device_unregister(&wq_dev->dev); | |
a0a1a5fd | 5929 | } |
6ba94429 FW |
5930 | #else /* CONFIG_SYSFS */ |
5931 | static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } | |
5932 | #endif /* CONFIG_SYSFS */ | |
a0a1a5fd | 5933 | |
82607adc TH |
5934 | /* |
5935 | * Workqueue watchdog. | |
5936 | * | |
5937 | * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal | |
5938 | * flush dependency, a concurrency managed work item which stays RUNNING | |
5939 | * indefinitely. Workqueue stalls can be very difficult to debug as the | |
5940 | * usual warning mechanisms don't trigger and internal workqueue state is | |
5941 | * largely opaque. | |
5942 | * | |
5943 | * Workqueue watchdog monitors all worker pools periodically and dumps | |
5944 | * state if some pools failed to make forward progress for a while where | |
5945 | * forward progress is defined as the first item on ->worklist changing. | |
5946 | * | |
5947 | * This mechanism is controlled through the kernel parameter | |
5948 | * "workqueue.watchdog_thresh" which can be updated at runtime through the | |
5949 | * corresponding sysfs parameter file. | |
5950 | */ | |
5951 | #ifdef CONFIG_WQ_WATCHDOG | |
5952 | ||
82607adc | 5953 | static unsigned long wq_watchdog_thresh = 30; |
5cd79d6a | 5954 | static struct timer_list wq_watchdog_timer; |
82607adc TH |
5955 | |
5956 | static unsigned long wq_watchdog_touched = INITIAL_JIFFIES; | |
5957 | static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES; | |
5958 | ||
5959 | static void wq_watchdog_reset_touched(void) | |
5960 | { | |
5961 | int cpu; | |
5962 | ||
5963 | wq_watchdog_touched = jiffies; | |
5964 | for_each_possible_cpu(cpu) | |
5965 | per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; | |
5966 | } | |
5967 | ||
5cd79d6a | 5968 | static void wq_watchdog_timer_fn(struct timer_list *unused) |
82607adc TH |
5969 | { |
5970 | unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ; | |
5971 | bool lockup_detected = false; | |
940d71c6 | 5972 | unsigned long now = jiffies; |
82607adc TH |
5973 | struct worker_pool *pool; |
5974 | int pi; | |
5975 | ||
5976 | if (!thresh) | |
5977 | return; | |
5978 | ||
5979 | rcu_read_lock(); | |
5980 | ||
5981 | for_each_pool(pool, pi) { | |
5982 | unsigned long pool_ts, touched, ts; | |
5983 | ||
5984 | if (list_empty(&pool->worklist)) | |
5985 | continue; | |
5986 | ||
940d71c6 SS |
5987 | /* |
5988 | * If a virtual machine is stopped by the host it can look to | |
5989 | * the watchdog like a stall. | |
5990 | */ | |
5991 | kvm_check_and_clear_guest_paused(); | |
5992 | ||
82607adc | 5993 | /* get the latest of pool and touched timestamps */ |
89e28ce6 WQ |
5994 | if (pool->cpu >= 0) |
5995 | touched = READ_ONCE(per_cpu(wq_watchdog_touched_cpu, pool->cpu)); | |
5996 | else | |
5997 | touched = READ_ONCE(wq_watchdog_touched); | |
82607adc | 5998 | pool_ts = READ_ONCE(pool->watchdog_ts); |
82607adc TH |
5999 | |
6000 | if (time_after(pool_ts, touched)) | |
6001 | ts = pool_ts; | |
6002 | else | |
6003 | ts = touched; | |
6004 | ||
82607adc | 6005 | /* did we stall? */ |
940d71c6 | 6006 | if (time_after(now, ts + thresh)) { |
82607adc TH |
6007 | lockup_detected = true; |
6008 | pr_emerg("BUG: workqueue lockup - pool"); | |
6009 | pr_cont_pool_info(pool); | |
6010 | pr_cont(" stuck for %us!\n", | |
940d71c6 | 6011 | jiffies_to_msecs(now - pool_ts) / 1000); |
82607adc TH |
6012 | } |
6013 | } | |
6014 | ||
6015 | rcu_read_unlock(); | |
6016 | ||
6017 | if (lockup_detected) | |
55df0933 | 6018 | show_all_workqueues(); |
82607adc TH |
6019 | |
6020 | wq_watchdog_reset_touched(); | |
6021 | mod_timer(&wq_watchdog_timer, jiffies + thresh); | |
6022 | } | |
6023 | ||
cb9d7fd5 | 6024 | notrace void wq_watchdog_touch(int cpu) |
82607adc TH |
6025 | { |
6026 | if (cpu >= 0) | |
6027 | per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; | |
89e28ce6 WQ |
6028 | |
6029 | wq_watchdog_touched = jiffies; | |
82607adc TH |
6030 | } |
6031 | ||
6032 | static void wq_watchdog_set_thresh(unsigned long thresh) | |
6033 | { | |
6034 | wq_watchdog_thresh = 0; | |
6035 | del_timer_sync(&wq_watchdog_timer); | |
6036 | ||
6037 | if (thresh) { | |
6038 | wq_watchdog_thresh = thresh; | |
6039 | wq_watchdog_reset_touched(); | |
6040 | mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ); | |
6041 | } | |
6042 | } | |
6043 | ||
6044 | static int wq_watchdog_param_set_thresh(const char *val, | |
6045 | const struct kernel_param *kp) | |
6046 | { | |
6047 | unsigned long thresh; | |
6048 | int ret; | |
6049 | ||
6050 | ret = kstrtoul(val, 0, &thresh); | |
6051 | if (ret) | |
6052 | return ret; | |
6053 | ||
6054 | if (system_wq) | |
6055 | wq_watchdog_set_thresh(thresh); | |
6056 | else | |
6057 | wq_watchdog_thresh = thresh; | |
6058 | ||
6059 | return 0; | |
6060 | } | |
6061 | ||
6062 | static const struct kernel_param_ops wq_watchdog_thresh_ops = { | |
6063 | .set = wq_watchdog_param_set_thresh, | |
6064 | .get = param_get_ulong, | |
6065 | }; | |
6066 | ||
6067 | module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh, | |
6068 | 0644); | |
6069 | ||
6070 | static void wq_watchdog_init(void) | |
6071 | { | |
5cd79d6a | 6072 | timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE); |
82607adc TH |
6073 | wq_watchdog_set_thresh(wq_watchdog_thresh); |
6074 | } | |
6075 | ||
6076 | #else /* CONFIG_WQ_WATCHDOG */ | |
6077 | ||
6078 | static inline void wq_watchdog_init(void) { } | |
6079 | ||
6080 | #endif /* CONFIG_WQ_WATCHDOG */ | |
6081 | ||
bce90380 TH |
6082 | static void __init wq_numa_init(void) |
6083 | { | |
6084 | cpumask_var_t *tbl; | |
6085 | int node, cpu; | |
6086 | ||
bce90380 TH |
6087 | if (num_possible_nodes() <= 1) |
6088 | return; | |
6089 | ||
d55262c4 TH |
6090 | if (wq_disable_numa) { |
6091 | pr_info("workqueue: NUMA affinity support disabled\n"); | |
6092 | return; | |
6093 | } | |
6094 | ||
f728c4a9 ZL |
6095 | for_each_possible_cpu(cpu) { |
6096 | if (WARN_ON(cpu_to_node(cpu) == NUMA_NO_NODE)) { | |
6097 | pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu); | |
6098 | return; | |
6099 | } | |
6100 | } | |
6101 | ||
be69d00d | 6102 | wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(); |
4c16bd32 TH |
6103 | BUG_ON(!wq_update_unbound_numa_attrs_buf); |
6104 | ||
bce90380 TH |
6105 | /* |
6106 | * We want masks of possible CPUs of each node which isn't readily | |
6107 | * available. Build one from cpu_to_node() which should have been | |
6108 | * fully initialized by now. | |
6109 | */ | |
6396bb22 | 6110 | tbl = kcalloc(nr_node_ids, sizeof(tbl[0]), GFP_KERNEL); |
bce90380 TH |
6111 | BUG_ON(!tbl); |
6112 | ||
6113 | for_each_node(node) | |
5a6024f1 | 6114 | BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL, |
1be0c25d | 6115 | node_online(node) ? node : NUMA_NO_NODE)); |
bce90380 TH |
6116 | |
6117 | for_each_possible_cpu(cpu) { | |
6118 | node = cpu_to_node(cpu); | |
bce90380 TH |
6119 | cpumask_set_cpu(cpu, tbl[node]); |
6120 | } | |
6121 | ||
6122 | wq_numa_possible_cpumask = tbl; | |
6123 | wq_numa_enabled = true; | |
6124 | } | |
6125 | ||
3347fa09 TH |
6126 | /** |
6127 | * workqueue_init_early - early init for workqueue subsystem | |
6128 | * | |
6129 | * This is the first half of two-staged workqueue subsystem initialization | |
6130 | * and invoked as soon as the bare basics - memory allocation, cpumasks and | |
6131 | * idr are up. It sets up all the data structures and system workqueues | |
6132 | * and allows early boot code to create workqueues and queue/cancel work | |
6133 | * items. Actual work item execution starts only after kthreads can be | |
6134 | * created and scheduled right before early initcalls. | |
6135 | */ | |
2333e829 | 6136 | void __init workqueue_init_early(void) |
1da177e4 | 6137 | { |
7a4e344c TH |
6138 | int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; |
6139 | int i, cpu; | |
c34056a3 | 6140 | |
10cdb157 | 6141 | BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); |
e904e6c2 | 6142 | |
b05a7928 | 6143 | BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL)); |
04d4e665 FW |
6144 | cpumask_copy(wq_unbound_cpumask, housekeeping_cpumask(HK_TYPE_WQ)); |
6145 | cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, housekeeping_cpumask(HK_TYPE_DOMAIN)); | |
b05a7928 | 6146 | |
e904e6c2 TH |
6147 | pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); |
6148 | ||
706026c2 | 6149 | /* initialize CPU pools */ |
29c91e99 | 6150 | for_each_possible_cpu(cpu) { |
4ce62e9e | 6151 | struct worker_pool *pool; |
8b03ae3c | 6152 | |
7a4e344c | 6153 | i = 0; |
f02ae73a | 6154 | for_each_cpu_worker_pool(pool, cpu) { |
7a4e344c | 6155 | BUG_ON(init_worker_pool(pool)); |
ec22ca5e | 6156 | pool->cpu = cpu; |
29c91e99 | 6157 | cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); |
7a4e344c | 6158 | pool->attrs->nice = std_nice[i++]; |
f3f90ad4 | 6159 | pool->node = cpu_to_node(cpu); |
7a4e344c | 6160 | |
9daf9e67 | 6161 | /* alloc pool ID */ |
68e13a67 | 6162 | mutex_lock(&wq_pool_mutex); |
9daf9e67 | 6163 | BUG_ON(worker_pool_assign_id(pool)); |
68e13a67 | 6164 | mutex_unlock(&wq_pool_mutex); |
4ce62e9e | 6165 | } |
8b03ae3c TH |
6166 | } |
6167 | ||
8a2b7538 | 6168 | /* create default unbound and ordered wq attrs */ |
29c91e99 TH |
6169 | for (i = 0; i < NR_STD_WORKER_POOLS; i++) { |
6170 | struct workqueue_attrs *attrs; | |
6171 | ||
be69d00d | 6172 | BUG_ON(!(attrs = alloc_workqueue_attrs())); |
29c91e99 | 6173 | attrs->nice = std_nice[i]; |
29c91e99 | 6174 | unbound_std_wq_attrs[i] = attrs; |
8a2b7538 TH |
6175 | |
6176 | /* | |
6177 | * An ordered wq should have only one pwq as ordering is | |
6178 | * guaranteed by max_active which is enforced by pwqs. | |
6179 | * Turn off NUMA so that dfl_pwq is used for all nodes. | |
6180 | */ | |
be69d00d | 6181 | BUG_ON(!(attrs = alloc_workqueue_attrs())); |
8a2b7538 TH |
6182 | attrs->nice = std_nice[i]; |
6183 | attrs->no_numa = true; | |
6184 | ordered_wq_attrs[i] = attrs; | |
29c91e99 TH |
6185 | } |
6186 | ||
d320c038 | 6187 | system_wq = alloc_workqueue("events", 0, 0); |
1aabe902 | 6188 | system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); |
d320c038 | 6189 | system_long_wq = alloc_workqueue("events_long", 0, 0); |
f3421797 TH |
6190 | system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, |
6191 | WQ_UNBOUND_MAX_ACTIVE); | |
24d51add TH |
6192 | system_freezable_wq = alloc_workqueue("events_freezable", |
6193 | WQ_FREEZABLE, 0); | |
0668106c VK |
6194 | system_power_efficient_wq = alloc_workqueue("events_power_efficient", |
6195 | WQ_POWER_EFFICIENT, 0); | |
6196 | system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient", | |
6197 | WQ_FREEZABLE | WQ_POWER_EFFICIENT, | |
6198 | 0); | |
1aabe902 | 6199 | BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || |
0668106c VK |
6200 | !system_unbound_wq || !system_freezable_wq || |
6201 | !system_power_efficient_wq || | |
6202 | !system_freezable_power_efficient_wq); | |
3347fa09 TH |
6203 | } |
6204 | ||
6205 | /** | |
6206 | * workqueue_init - bring workqueue subsystem fully online | |
6207 | * | |
6208 | * This is the latter half of two-staged workqueue subsystem initialization | |
6209 | * and invoked as soon as kthreads can be created and scheduled. | |
6210 | * Workqueues have been created and work items queued on them, but there | |
6211 | * are no kworkers executing the work items yet. Populate the worker pools | |
6212 | * with the initial workers and enable future kworker creations. | |
6213 | */ | |
2333e829 | 6214 | void __init workqueue_init(void) |
3347fa09 | 6215 | { |
2186d9f9 | 6216 | struct workqueue_struct *wq; |
3347fa09 TH |
6217 | struct worker_pool *pool; |
6218 | int cpu, bkt; | |
6219 | ||
2186d9f9 TH |
6220 | /* |
6221 | * It'd be simpler to initialize NUMA in workqueue_init_early() but | |
6222 | * CPU to node mapping may not be available that early on some | |
6223 | * archs such as power and arm64. As per-cpu pools created | |
6224 | * previously could be missing node hint and unbound pools NUMA | |
6225 | * affinity, fix them up. | |
40c17f75 TH |
6226 | * |
6227 | * Also, while iterating workqueues, create rescuers if requested. | |
2186d9f9 TH |
6228 | */ |
6229 | wq_numa_init(); | |
6230 | ||
6231 | mutex_lock(&wq_pool_mutex); | |
6232 | ||
6233 | for_each_possible_cpu(cpu) { | |
6234 | for_each_cpu_worker_pool(pool, cpu) { | |
6235 | pool->node = cpu_to_node(cpu); | |
6236 | } | |
6237 | } | |
6238 | ||
40c17f75 | 6239 | list_for_each_entry(wq, &workqueues, list) { |
2186d9f9 | 6240 | wq_update_unbound_numa(wq, smp_processor_id(), true); |
40c17f75 TH |
6241 | WARN(init_rescuer(wq), |
6242 | "workqueue: failed to create early rescuer for %s", | |
6243 | wq->name); | |
6244 | } | |
2186d9f9 TH |
6245 | |
6246 | mutex_unlock(&wq_pool_mutex); | |
6247 | ||
3347fa09 TH |
6248 | /* create the initial workers */ |
6249 | for_each_online_cpu(cpu) { | |
6250 | for_each_cpu_worker_pool(pool, cpu) { | |
6251 | pool->flags &= ~POOL_DISASSOCIATED; | |
6252 | BUG_ON(!create_worker(pool)); | |
6253 | } | |
6254 | } | |
6255 | ||
6256 | hash_for_each(unbound_pool_hash, bkt, pool, hash_node) | |
6257 | BUG_ON(!create_worker(pool)); | |
6258 | ||
6259 | wq_online = true; | |
82607adc | 6260 | wq_watchdog_init(); |
1da177e4 | 6261 | } |
c4f135d6 TH |
6262 | |
6263 | /* | |
6264 | * Despite the naming, this is a no-op function which is here only for avoiding | |
6265 | * link error. Since compile-time warning may fail to catch, we will need to | |
6266 | * emit run-time warning from __flush_workqueue(). | |
6267 | */ | |
6268 | void __warn_flushing_systemwide_wq(void) { } | |
6269 | EXPORT_SYMBOL(__warn_flushing_systemwide_wq); |