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