Line data Source code
1 : /*
2 : * kernel/workqueue.c - generic async execution with shared worker pool
3 : *
4 : * Copyright (C) 2002 Ingo Molnar
5 : *
6 : * Derived from the taskqueue/keventd code by:
7 : * David Woodhouse <dwmw2@infradead.org>
8 : * Andrew Morton
9 : * Kai Petzke <wpp@marie.physik.tu-berlin.de>
10 : * Theodore Ts'o <tytso@mit.edu>
11 : *
12 : * Made to use alloc_percpu by Christoph Lameter.
13 : *
14 : * Copyright (C) 2010 SUSE Linux Products GmbH
15 : * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
16 : *
17 : * This is the generic async execution mechanism. Work items as are
18 : * executed in process context. The worker pool is shared and
19 : * automatically managed. There are two worker pools for each CPU (one for
20 : * normal work items and the other for high priority ones) and some extra
21 : * pools for workqueues which are not bound to any specific CPU - the
22 : * number of these backing pools is dynamic.
23 : *
24 : * Please read Documentation/workqueue.txt for details.
25 : */
26 :
27 : #include <linux/export.h>
28 : #include <linux/kernel.h>
29 : #include <linux/sched.h>
30 : #include <linux/init.h>
31 : #include <linux/signal.h>
32 : #include <linux/completion.h>
33 : #include <linux/workqueue.h>
34 : #include <linux/slab.h>
35 : #include <linux/cpu.h>
36 : #include <linux/notifier.h>
37 : #include <linux/kthread.h>
38 : #include <linux/hardirq.h>
39 : #include <linux/mempolicy.h>
40 : #include <linux/freezer.h>
41 : #include <linux/kallsyms.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 :
52 : #include "workqueue_internal.h"
53 :
54 : enum {
55 : /*
56 : * worker_pool flags
57 : *
58 : * A bound pool is either associated or disassociated with its CPU.
59 : * While associated (!DISASSOCIATED), all workers are bound to the
60 : * CPU and none has %WORKER_UNBOUND set and concurrency management
61 : * is in effect.
62 : *
63 : * While DISASSOCIATED, the cpu may be offline and all workers have
64 : * %WORKER_UNBOUND set and concurrency management disabled, and may
65 : * be executing on any CPU. The pool behaves as an unbound one.
66 : *
67 : * Note that DISASSOCIATED should be flipped only while holding
68 : * attach_mutex to avoid changing binding state while
69 : * worker_attach_to_pool() is in progress.
70 : */
71 : POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
72 :
73 : /* worker flags */
74 : WORKER_DIE = 1 << 1, /* die die die */
75 : WORKER_IDLE = 1 << 2, /* is idle */
76 : WORKER_PREP = 1 << 3, /* preparing to run works */
77 : WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
78 : WORKER_UNBOUND = 1 << 7, /* worker is unbound */
79 : WORKER_REBOUND = 1 << 8, /* worker was rebound */
80 :
81 : WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
82 : WORKER_UNBOUND | WORKER_REBOUND,
83 :
84 : NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
85 :
86 : UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
87 : BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
88 :
89 : MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
90 : IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
91 :
92 : MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
93 : /* call for help after 10ms
94 : (min two ticks) */
95 : MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
96 : CREATE_COOLDOWN = HZ, /* time to breath after fail */
97 :
98 : /*
99 : * Rescue workers are used only on emergencies and shared by
100 : * all cpus. Give MIN_NICE.
101 : */
102 : RESCUER_NICE_LEVEL = MIN_NICE,
103 : HIGHPRI_NICE_LEVEL = MIN_NICE,
104 :
105 : WQ_NAME_LEN = 24,
106 : };
107 :
108 : /*
109 : * Structure fields follow one of the following exclusion rules.
110 : *
111 : * I: Modifiable by initialization/destruction paths and read-only for
112 : * everyone else.
113 : *
114 : * P: Preemption protected. Disabling preemption is enough and should
115 : * only be modified and accessed from the local cpu.
116 : *
117 : * L: pool->lock protected. Access with pool->lock held.
118 : *
119 : * X: During normal operation, modification requires pool->lock and should
120 : * be done only from local cpu. Either disabling preemption on local
121 : * cpu or grabbing pool->lock is enough for read access. If
122 : * POOL_DISASSOCIATED is set, it's identical to L.
123 : *
124 : * A: pool->attach_mutex protected.
125 : *
126 : * PL: wq_pool_mutex protected.
127 : *
128 : * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads.
129 : *
130 : * WQ: wq->mutex protected.
131 : *
132 : * WR: wq->mutex protected for writes. Sched-RCU protected for reads.
133 : *
134 : * MD: wq_mayday_lock protected.
135 : */
136 :
137 : /* struct worker is defined in workqueue_internal.h */
138 :
139 : struct worker_pool {
140 : spinlock_t lock; /* the pool lock */
141 : int cpu; /* I: the associated cpu */
142 : int node; /* I: the associated node ID */
143 : int id; /* I: pool ID */
144 : unsigned int flags; /* X: flags */
145 :
146 : struct list_head worklist; /* L: list of pending works */
147 : int nr_workers; /* L: total number of workers */
148 :
149 : /* nr_idle includes the ones off idle_list for rebinding */
150 : int nr_idle; /* L: currently idle ones */
151 :
152 : struct list_head idle_list; /* X: list of idle workers */
153 : struct timer_list idle_timer; /* L: worker idle timeout */
154 : struct timer_list mayday_timer; /* L: SOS timer for workers */
155 :
156 : /* a workers is either on busy_hash or idle_list, or the manager */
157 : DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
158 : /* L: hash of busy workers */
159 :
160 : /* see manage_workers() for details on the two manager mutexes */
161 : struct mutex manager_arb; /* manager arbitration */
162 : struct mutex attach_mutex; /* attach/detach exclusion */
163 : struct list_head workers; /* A: attached workers */
164 : struct completion *detach_completion; /* all workers detached */
165 :
166 : struct ida worker_ida; /* worker IDs for task name */
167 :
168 : struct workqueue_attrs *attrs; /* I: worker attributes */
169 : struct hlist_node hash_node; /* PL: unbound_pool_hash node */
170 : int refcnt; /* PL: refcnt for unbound pools */
171 :
172 : /*
173 : * The current concurrency level. As it's likely to be accessed
174 : * from other CPUs during try_to_wake_up(), put it in a separate
175 : * cacheline.
176 : */
177 : atomic_t nr_running ____cacheline_aligned_in_smp;
178 :
179 : /*
180 : * Destruction of pool is sched-RCU protected to allow dereferences
181 : * from get_work_pool().
182 : */
183 : struct rcu_head rcu;
184 : } ____cacheline_aligned_in_smp;
185 :
186 : /*
187 : * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
188 : * of work_struct->data are used for flags and the remaining high bits
189 : * point to the pwq; thus, pwqs need to be aligned at two's power of the
190 : * number of flag bits.
191 : */
192 : struct pool_workqueue {
193 : struct worker_pool *pool; /* I: the associated pool */
194 : struct workqueue_struct *wq; /* I: the owning workqueue */
195 : int work_color; /* L: current color */
196 : int flush_color; /* L: flushing color */
197 : int refcnt; /* L: reference count */
198 : int nr_in_flight[WORK_NR_COLORS];
199 : /* L: nr of in_flight works */
200 : int nr_active; /* L: nr of active works */
201 : int max_active; /* L: max active works */
202 : struct list_head delayed_works; /* L: delayed works */
203 : struct list_head pwqs_node; /* WR: node on wq->pwqs */
204 : struct list_head mayday_node; /* MD: node on wq->maydays */
205 :
206 : /*
207 : * Release of unbound pwq is punted to system_wq. See put_pwq()
208 : * and pwq_unbound_release_workfn() for details. pool_workqueue
209 : * itself is also sched-RCU protected so that the first pwq can be
210 : * determined without grabbing wq->mutex.
211 : */
212 : struct work_struct unbound_release_work;
213 : struct rcu_head rcu;
214 : } __aligned(1 << WORK_STRUCT_FLAG_BITS);
215 :
216 : /*
217 : * Structure used to wait for workqueue flush.
218 : */
219 : struct wq_flusher {
220 : struct list_head list; /* WQ: list of flushers */
221 : int flush_color; /* WQ: flush color waiting for */
222 : struct completion done; /* flush completion */
223 : };
224 :
225 : struct wq_device;
226 :
227 : /*
228 : * The externally visible workqueue. It relays the issued work items to
229 : * the appropriate worker_pool through its pool_workqueues.
230 : */
231 : struct workqueue_struct {
232 : struct list_head pwqs; /* WR: all pwqs of this wq */
233 : struct list_head list; /* PL: list of all workqueues */
234 :
235 : struct mutex mutex; /* protects this wq */
236 : int work_color; /* WQ: current work color */
237 : int flush_color; /* WQ: current flush color */
238 : atomic_t nr_pwqs_to_flush; /* flush in progress */
239 : struct wq_flusher *first_flusher; /* WQ: first flusher */
240 : struct list_head flusher_queue; /* WQ: flush waiters */
241 : struct list_head flusher_overflow; /* WQ: flush overflow list */
242 :
243 : struct list_head maydays; /* MD: pwqs requesting rescue */
244 : struct worker *rescuer; /* I: rescue worker */
245 :
246 : int nr_drainers; /* WQ: drain in progress */
247 : int saved_max_active; /* WQ: saved pwq max_active */
248 :
249 : struct workqueue_attrs *unbound_attrs; /* WQ: only for unbound wqs */
250 : struct pool_workqueue *dfl_pwq; /* WQ: only for unbound wqs */
251 :
252 : #ifdef CONFIG_SYSFS
253 : struct wq_device *wq_dev; /* I: for sysfs interface */
254 : #endif
255 : #ifdef CONFIG_LOCKDEP
256 : struct lockdep_map lockdep_map;
257 : #endif
258 : char name[WQ_NAME_LEN]; /* I: workqueue name */
259 :
260 : /* hot fields used during command issue, aligned to cacheline */
261 : unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
262 : struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
263 : struct pool_workqueue __rcu *numa_pwq_tbl[]; /* FR: unbound pwqs indexed by node */
264 : };
265 :
266 : static struct kmem_cache *pwq_cache;
267 :
268 : static cpumask_var_t *wq_numa_possible_cpumask;
269 : /* possible CPUs of each node */
270 :
271 : static bool wq_disable_numa;
272 : module_param_named(disable_numa, wq_disable_numa, bool, 0444);
273 :
274 : /* see the comment above the definition of WQ_POWER_EFFICIENT */
275 : #ifdef CONFIG_WQ_POWER_EFFICIENT_DEFAULT
276 : static bool wq_power_efficient = true;
277 : #else
278 : static bool wq_power_efficient;
279 : #endif
280 :
281 : module_param_named(power_efficient, wq_power_efficient, bool, 0444);
282 :
283 : static bool wq_numa_enabled; /* unbound NUMA affinity enabled */
284 :
285 : /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
286 : static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
287 :
288 : static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
289 : static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
290 :
291 : static LIST_HEAD(workqueues); /* PL: list of all workqueues */
292 : static bool workqueue_freezing; /* PL: have wqs started freezing? */
293 :
294 : /* the per-cpu worker pools */
295 : static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS],
296 : cpu_worker_pools);
297 :
298 : static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
299 :
300 : /* PL: hash of all unbound pools keyed by pool->attrs */
301 : static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
302 :
303 : /* I: attributes used when instantiating standard unbound pools on demand */
304 : static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
305 :
306 : /* I: attributes used when instantiating ordered pools on demand */
307 : static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
308 :
309 : struct workqueue_struct *system_wq __read_mostly;
310 : EXPORT_SYMBOL(system_wq);
311 : struct workqueue_struct *system_highpri_wq __read_mostly;
312 : EXPORT_SYMBOL_GPL(system_highpri_wq);
313 : struct workqueue_struct *system_long_wq __read_mostly;
314 : EXPORT_SYMBOL_GPL(system_long_wq);
315 : struct workqueue_struct *system_unbound_wq __read_mostly;
316 : EXPORT_SYMBOL_GPL(system_unbound_wq);
317 : struct workqueue_struct *system_freezable_wq __read_mostly;
318 : EXPORT_SYMBOL_GPL(system_freezable_wq);
319 : struct workqueue_struct *system_power_efficient_wq __read_mostly;
320 : EXPORT_SYMBOL_GPL(system_power_efficient_wq);
321 : struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
322 : EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
323 :
324 : static int worker_thread(void *__worker);
325 : static void copy_workqueue_attrs(struct workqueue_attrs *to,
326 : const struct workqueue_attrs *from);
327 :
328 : #define CREATE_TRACE_POINTS
329 : #include <trace/events/workqueue.h>
330 :
331 : #define assert_rcu_or_pool_mutex() \
332 : rcu_lockdep_assert(rcu_read_lock_sched_held() || \
333 : lockdep_is_held(&wq_pool_mutex), \
334 : "sched RCU or wq_pool_mutex should be held")
335 :
336 : #define assert_rcu_or_wq_mutex(wq) \
337 : rcu_lockdep_assert(rcu_read_lock_sched_held() || \
338 : lockdep_is_held(&wq->mutex), \
339 : "sched RCU or wq->mutex should be held")
340 :
341 : #define for_each_cpu_worker_pool(pool, cpu) \
342 : for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
343 : (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
344 : (pool)++)
345 :
346 : /**
347 : * for_each_pool - iterate through all worker_pools in the system
348 : * @pool: iteration cursor
349 : * @pi: integer used for iteration
350 : *
351 : * This must be called either with wq_pool_mutex held or sched RCU read
352 : * locked. If the pool needs to be used beyond the locking in effect, the
353 : * caller is responsible for guaranteeing that the pool stays online.
354 : *
355 : * The if/else clause exists only for the lockdep assertion and can be
356 : * ignored.
357 : */
358 : #define for_each_pool(pool, pi) \
359 : idr_for_each_entry(&worker_pool_idr, pool, pi) \
360 : if (({ assert_rcu_or_pool_mutex(); false; })) { } \
361 : else
362 :
363 : /**
364 : * for_each_pool_worker - iterate through all workers of a worker_pool
365 : * @worker: iteration cursor
366 : * @pool: worker_pool to iterate workers of
367 : *
368 : * This must be called with @pool->attach_mutex.
369 : *
370 : * The if/else clause exists only for the lockdep assertion and can be
371 : * ignored.
372 : */
373 : #define for_each_pool_worker(worker, pool) \
374 : list_for_each_entry((worker), &(pool)->workers, node) \
375 : if (({ lockdep_assert_held(&pool->attach_mutex); false; })) { } \
376 : else
377 :
378 : /**
379 : * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
380 : * @pwq: iteration cursor
381 : * @wq: the target workqueue
382 : *
383 : * This must be called either with wq->mutex held or sched RCU read locked.
384 : * If the pwq needs to be used beyond the locking in effect, the caller is
385 : * responsible for guaranteeing that the pwq stays online.
386 : *
387 : * The if/else clause exists only for the lockdep assertion and can be
388 : * ignored.
389 : */
390 : #define for_each_pwq(pwq, wq) \
391 : list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
392 : if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
393 : else
394 :
395 : #ifdef CONFIG_DEBUG_OBJECTS_WORK
396 :
397 : static struct debug_obj_descr work_debug_descr;
398 :
399 : static void *work_debug_hint(void *addr)
400 : {
401 : return ((struct work_struct *) addr)->func;
402 : }
403 :
404 : /*
405 : * fixup_init is called when:
406 : * - an active object is initialized
407 : */
408 : static int work_fixup_init(void *addr, enum debug_obj_state state)
409 : {
410 : struct work_struct *work = addr;
411 :
412 : switch (state) {
413 : case ODEBUG_STATE_ACTIVE:
414 : cancel_work_sync(work);
415 : debug_object_init(work, &work_debug_descr);
416 : return 1;
417 : default:
418 : return 0;
419 : }
420 : }
421 :
422 : /*
423 : * fixup_activate is called when:
424 : * - an active object is activated
425 : * - an unknown object is activated (might be a statically initialized object)
426 : */
427 : static int work_fixup_activate(void *addr, enum debug_obj_state state)
428 : {
429 : struct work_struct *work = addr;
430 :
431 : switch (state) {
432 :
433 : case ODEBUG_STATE_NOTAVAILABLE:
434 : /*
435 : * This is not really a fixup. The work struct was
436 : * statically initialized. We just make sure that it
437 : * is tracked in the object tracker.
438 : */
439 : if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
440 : debug_object_init(work, &work_debug_descr);
441 : debug_object_activate(work, &work_debug_descr);
442 : return 0;
443 : }
444 : WARN_ON_ONCE(1);
445 : return 0;
446 :
447 : case ODEBUG_STATE_ACTIVE:
448 : WARN_ON(1);
449 :
450 : default:
451 : return 0;
452 : }
453 : }
454 :
455 : /*
456 : * fixup_free is called when:
457 : * - an active object is freed
458 : */
459 : static int work_fixup_free(void *addr, enum debug_obj_state state)
460 : {
461 : struct work_struct *work = addr;
462 :
463 : switch (state) {
464 : case ODEBUG_STATE_ACTIVE:
465 : cancel_work_sync(work);
466 : debug_object_free(work, &work_debug_descr);
467 : return 1;
468 : default:
469 : return 0;
470 : }
471 : }
472 :
473 : static struct debug_obj_descr work_debug_descr = {
474 : .name = "work_struct",
475 : .debug_hint = work_debug_hint,
476 : .fixup_init = work_fixup_init,
477 : .fixup_activate = work_fixup_activate,
478 : .fixup_free = work_fixup_free,
479 : };
480 :
481 : static inline void debug_work_activate(struct work_struct *work)
482 : {
483 : debug_object_activate(work, &work_debug_descr);
484 : }
485 :
486 : static inline void debug_work_deactivate(struct work_struct *work)
487 : {
488 : debug_object_deactivate(work, &work_debug_descr);
489 : }
490 :
491 : void __init_work(struct work_struct *work, int onstack)
492 : {
493 : if (onstack)
494 : debug_object_init_on_stack(work, &work_debug_descr);
495 : else
496 : debug_object_init(work, &work_debug_descr);
497 : }
498 : EXPORT_SYMBOL_GPL(__init_work);
499 :
500 : void destroy_work_on_stack(struct work_struct *work)
501 : {
502 : debug_object_free(work, &work_debug_descr);
503 : }
504 : EXPORT_SYMBOL_GPL(destroy_work_on_stack);
505 :
506 : void destroy_delayed_work_on_stack(struct delayed_work *work)
507 : {
508 : destroy_timer_on_stack(&work->timer);
509 : debug_object_free(&work->work, &work_debug_descr);
510 : }
511 : EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
512 :
513 : #else
514 : static inline void debug_work_activate(struct work_struct *work) { }
515 : static inline void debug_work_deactivate(struct work_struct *work) { }
516 : #endif
517 :
518 : /**
519 : * worker_pool_assign_id - allocate ID and assing it to @pool
520 : * @pool: the pool pointer of interest
521 : *
522 : * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
523 : * successfully, -errno on failure.
524 : */
525 3 : static int worker_pool_assign_id(struct worker_pool *pool)
526 : {
527 : int ret;
528 :
529 : lockdep_assert_held(&wq_pool_mutex);
530 :
531 3 : ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
532 : GFP_KERNEL);
533 3 : if (ret >= 0) {
534 3 : pool->id = ret;
535 3 : return 0;
536 : }
537 : return ret;
538 : }
539 :
540 : /**
541 : * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
542 : * @wq: the target workqueue
543 : * @node: the node ID
544 : *
545 : * This must be called either with pwq_lock held or sched RCU read locked.
546 : * If the pwq needs to be used beyond the locking in effect, the caller is
547 : * responsible for guaranteeing that the pwq stays online.
548 : *
549 : * Return: The unbound pool_workqueue for @node.
550 : */
551 : static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
552 : int node)
553 : {
554 : assert_rcu_or_wq_mutex(wq);
555 2763 : return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
556 : }
557 :
558 : static unsigned int work_color_to_flags(int color)
559 : {
560 42745 : return color << WORK_STRUCT_COLOR_SHIFT;
561 : }
562 :
563 : static int get_work_color(struct work_struct *work)
564 : {
565 42754 : return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
566 : ((1 << WORK_STRUCT_COLOR_BITS) - 1);
567 : }
568 :
569 : static int work_next_color(int color)
570 : {
571 2 : return (color + 1) % WORK_NR_COLORS;
572 : }
573 :
574 : /*
575 : * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
576 : * contain the pointer to the queued pwq. Once execution starts, the flag
577 : * is cleared and the high bits contain OFFQ flags and pool ID.
578 : *
579 : * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
580 : * and clear_work_data() can be used to set the pwq, pool or clear
581 : * work->data. These functions should only be called while the work is
582 : * owned - ie. while the PENDING bit is set.
583 : *
584 : * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
585 : * corresponding to a work. Pool is available once the work has been
586 : * queued anywhere after initialization until it is sync canceled. pwq is
587 : * available only while the work item is queued.
588 : *
589 : * %WORK_OFFQ_CANCELING is used to mark a work item which is being
590 : * canceled. While being canceled, a work item may have its PENDING set
591 : * but stay off timer and worklist for arbitrarily long and nobody should
592 : * try to steal the PENDING bit.
593 : */
594 : static inline void set_work_data(struct work_struct *work, unsigned long data,
595 : unsigned long flags)
596 : {
597 : WARN_ON_ONCE(!work_pending(work));
598 86909 : atomic_long_set(&work->data, data | flags | work_static(work));
599 : }
600 :
601 : static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
602 : unsigned long extra_flags)
603 : {
604 42754 : set_work_data(work, (unsigned long)pwq,
605 : WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
606 : }
607 :
608 : static void set_work_pool_and_keep_pending(struct work_struct *work,
609 : int pool_id)
610 : {
611 0 : set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
612 : WORK_STRUCT_PENDING);
613 : }
614 :
615 : static void set_work_pool_and_clear_pending(struct work_struct *work,
616 : int pool_id)
617 : {
618 : /*
619 : * The following wmb is paired with the implied mb in
620 : * test_and_set_bit(PENDING) and ensures all updates to @work made
621 : * here are visible to and precede any updates by the next PENDING
622 : * owner.
623 : */
624 42757 : smp_wmb();
625 42757 : set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
626 : }
627 :
628 : static void clear_work_data(struct work_struct *work)
629 : {
630 1398 : smp_wmb(); /* see set_work_pool_and_clear_pending() */
631 : set_work_data(work, WORK_STRUCT_NO_POOL, 0);
632 : }
633 :
634 : static struct pool_workqueue *get_work_pwq(struct work_struct *work)
635 : {
636 42840 : unsigned long data = atomic_long_read(&work->data);
637 :
638 42840 : if (data & WORK_STRUCT_PWQ)
639 42767 : return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
640 : else
641 : return NULL;
642 : }
643 :
644 : /**
645 : * get_work_pool - return the worker_pool a given work was associated with
646 : * @work: the work item of interest
647 : *
648 : * Pools are created and destroyed under wq_pool_mutex, and allows read
649 : * access under sched-RCU read lock. As such, this function should be
650 : * called under wq_pool_mutex or with preemption disabled.
651 : *
652 : * All fields of the returned pool are accessible as long as the above
653 : * mentioned locking is in effect. If the returned pool needs to be used
654 : * beyond the critical section, the caller is responsible for ensuring the
655 : * returned pool is and stays online.
656 : *
657 : * Return: The worker_pool @work was last associated with. %NULL if none.
658 : */
659 46540 : static struct worker_pool *get_work_pool(struct work_struct *work)
660 : {
661 46540 : unsigned long data = atomic_long_read(&work->data);
662 : int pool_id;
663 :
664 : assert_rcu_or_pool_mutex();
665 :
666 46540 : if (data & WORK_STRUCT_PWQ)
667 9 : return ((struct pool_workqueue *)
668 9 : (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
669 :
670 46531 : pool_id = data >> WORK_OFFQ_POOL_SHIFT;
671 46531 : if (pool_id == WORK_OFFQ_POOL_NONE)
672 : return NULL;
673 :
674 41798 : return idr_find(&worker_pool_idr, pool_id);
675 : }
676 :
677 : /**
678 : * get_work_pool_id - return the worker pool ID a given work is associated with
679 : * @work: the work item of interest
680 : *
681 : * Return: The worker_pool ID @work was last associated with.
682 : * %WORK_OFFQ_POOL_NONE if none.
683 : */
684 : static int get_work_pool_id(struct work_struct *work)
685 : {
686 1401 : unsigned long data = atomic_long_read(&work->data);
687 :
688 1401 : if (data & WORK_STRUCT_PWQ)
689 0 : return ((struct pool_workqueue *)
690 0 : (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
691 :
692 1401 : return data >> WORK_OFFQ_POOL_SHIFT;
693 : }
694 :
695 : static void mark_work_canceling(struct work_struct *work)
696 : {
697 1398 : unsigned long pool_id = get_work_pool_id(work);
698 :
699 1398 : pool_id <<= WORK_OFFQ_POOL_SHIFT;
700 : set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
701 : }
702 :
703 : static bool work_is_canceling(struct work_struct *work)
704 : {
705 0 : unsigned long data = atomic_long_read(&work->data);
706 :
707 0 : return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
708 : }
709 :
710 : /*
711 : * Policy functions. These define the policies on how the global worker
712 : * pools are managed. Unless noted otherwise, these functions assume that
713 : * they're being called with pool->lock held.
714 : */
715 :
716 : static bool __need_more_worker(struct worker_pool *pool)
717 : {
718 85483 : return !atomic_read(&pool->nr_running);
719 : }
720 :
721 : /*
722 : * Need to wake up a worker? Called from anything but currently
723 : * running workers.
724 : *
725 : * Note that, because unbound workers never contribute to nr_running, this
726 : * function will always return %true for unbound pools as long as the
727 : * worklist isn't empty.
728 : */
729 : static bool need_more_worker(struct worker_pool *pool)
730 : {
731 212839 : return !list_empty(&pool->worklist) && __need_more_worker(pool);
732 : }
733 :
734 : /* Can I start working? Called from busy but !running workers. */
735 : static bool may_start_working(struct worker_pool *pool)
736 : {
737 : return pool->nr_idle;
738 : }
739 :
740 : /* Do I need to keep working? Called from currently running workers. */
741 : static bool keep_working(struct worker_pool *pool)
742 : {
743 43247 : return !list_empty(&pool->worklist) &&
744 502 : atomic_read(&pool->nr_running) <= 1;
745 : }
746 :
747 : /* Do we need a new worker? Called from manager. */
748 7 : static bool need_to_create_worker(struct worker_pool *pool)
749 : {
750 16 : return need_more_worker(pool) && !may_start_working(pool);
751 : }
752 :
753 : /* Do we have too many workers and should some go away? */
754 : static bool too_many_workers(struct worker_pool *pool)
755 : {
756 : bool managing = mutex_is_locked(&pool->manager_arb);
757 42307 : int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
758 42307 : int nr_busy = pool->nr_workers - nr_idle;
759 :
760 42307 : return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
761 : }
762 :
763 : /*
764 : * Wake up functions.
765 : */
766 :
767 : /* Return the first idle worker. Safe with preemption disabled */
768 : static struct worker *first_idle_worker(struct worker_pool *pool)
769 : {
770 85508 : if (unlikely(list_empty(&pool->idle_list)))
771 : return NULL;
772 :
773 : return list_first_entry(&pool->idle_list, struct worker, entry);
774 : }
775 :
776 : /**
777 : * wake_up_worker - wake up an idle worker
778 : * @pool: worker pool to wake worker from
779 : *
780 : * Wake up the first idle worker of @pool.
781 : *
782 : * CONTEXT:
783 : * spin_lock_irq(pool->lock).
784 : */
785 42747 : static void wake_up_worker(struct worker_pool *pool)
786 : {
787 : struct worker *worker = first_idle_worker(pool);
788 :
789 42747 : if (likely(worker))
790 42746 : wake_up_process(worker->task);
791 42747 : }
792 :
793 : /**
794 : * wq_worker_waking_up - a worker is waking up
795 : * @task: task waking up
796 : * @cpu: CPU @task is waking up to
797 : *
798 : * This function is called during try_to_wake_up() when a worker is
799 : * being awoken.
800 : *
801 : * CONTEXT:
802 : * spin_lock_irq(rq->lock)
803 : */
804 42469 : void wq_worker_waking_up(struct task_struct *task, int cpu)
805 : {
806 42469 : struct worker *worker = kthread_data(task);
807 :
808 42469 : if (!(worker->flags & WORKER_NOT_RUNNING)) {
809 : WARN_ON_ONCE(worker->pool->cpu != cpu);
810 111 : atomic_inc(&worker->pool->nr_running);
811 : }
812 42469 : }
813 :
814 : /**
815 : * wq_worker_sleeping - a worker is going to sleep
816 : * @task: task going to sleep
817 : * @cpu: CPU in question, must be the current CPU number
818 : *
819 : * This function is called during schedule() when a busy worker is
820 : * going to sleep. Worker on the same cpu can be woken up by
821 : * returning pointer to its task.
822 : *
823 : * CONTEXT:
824 : * spin_lock_irq(rq->lock)
825 : *
826 : * Return:
827 : * Worker task on @cpu to wake up, %NULL if none.
828 : */
829 42489 : struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu)
830 : {
831 42489 : struct worker *worker = kthread_data(task), *to_wakeup = NULL;
832 : struct worker_pool *pool;
833 :
834 : /*
835 : * Rescuers, which may not have all the fields set up like normal
836 : * workers, also reach here, let's not access anything before
837 : * checking NOT_RUNNING.
838 : */
839 42489 : if (worker->flags & WORKER_NOT_RUNNING)
840 : return NULL;
841 :
842 111 : pool = worker->pool;
843 :
844 : /* this can only happen on the local cpu */
845 111 : if (WARN_ON_ONCE(cpu != raw_smp_processor_id() || pool->cpu != cpu))
846 : return NULL;
847 :
848 : /*
849 : * The counterpart of the following dec_and_test, implied mb,
850 : * worklist not empty test sequence is in insert_work().
851 : * Please read comment there.
852 : *
853 : * NOT_RUNNING is clear. This means that we're bound to and
854 : * running on the local cpu w/ rq lock held and preemption
855 : * disabled, which in turn means that none else could be
856 : * manipulating idle_list, so dereferencing idle_list without pool
857 : * lock is safe.
858 : */
859 333 : if (atomic_dec_and_test(&pool->nr_running) &&
860 111 : !list_empty(&pool->worklist))
861 : to_wakeup = first_idle_worker(pool);
862 111 : return to_wakeup ? to_wakeup->task : NULL;
863 : }
864 :
865 : /**
866 : * worker_set_flags - set worker flags and adjust nr_running accordingly
867 : * @worker: self
868 : * @flags: flags to set
869 : *
870 : * Set @flags in @worker->flags and adjust nr_running accordingly.
871 : *
872 : * CONTEXT:
873 : * spin_lock_irq(pool->lock)
874 : */
875 : static inline void worker_set_flags(struct worker *worker, unsigned int flags)
876 : {
877 : struct worker_pool *pool = worker->pool;
878 :
879 : WARN_ON_ONCE(worker->task != current);
880 :
881 : /* If transitioning into NOT_RUNNING, adjust nr_running. */
882 42243 : if ((flags & WORKER_NOT_RUNNING) &&
883 42243 : !(worker->flags & WORKER_NOT_RUNNING)) {
884 39484 : atomic_dec(&pool->nr_running);
885 : }
886 :
887 42243 : worker->flags |= flags;
888 : }
889 :
890 : /**
891 : * worker_clr_flags - clear worker flags and adjust nr_running accordingly
892 : * @worker: self
893 : * @flags: flags to clear
894 : *
895 : * Clear @flags in @worker->flags and adjust nr_running accordingly.
896 : *
897 : * CONTEXT:
898 : * spin_lock_irq(pool->lock)
899 : */
900 : static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
901 : {
902 : struct worker_pool *pool = worker->pool;
903 42243 : unsigned int oflags = worker->flags;
904 :
905 : WARN_ON_ONCE(worker->task != current);
906 :
907 84527 : worker->flags &= ~flags;
908 :
909 : /*
910 : * If transitioning out of NOT_RUNNING, increment nr_running. Note
911 : * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
912 : * of multiple flags, not a single flag.
913 : */
914 42243 : if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
915 42243 : if (!(worker->flags & WORKER_NOT_RUNNING))
916 39484 : atomic_inc(&pool->nr_running);
917 : }
918 :
919 : /**
920 : * find_worker_executing_work - find worker which is executing a work
921 : * @pool: pool of interest
922 : * @work: work to find worker for
923 : *
924 : * Find a worker which is executing @work on @pool by searching
925 : * @pool->busy_hash which is keyed by the address of @work. For a worker
926 : * to match, its current execution should match the address of @work and
927 : * its work function. This is to avoid unwanted dependency between
928 : * unrelated work executions through a work item being recycled while still
929 : * being executed.
930 : *
931 : * This is a bit tricky. A work item may be freed once its execution
932 : * starts and nothing prevents the freed area from being recycled for
933 : * another work item. If the same work item address ends up being reused
934 : * before the original execution finishes, workqueue will identify the
935 : * recycled work item as currently executing and make it wait until the
936 : * current execution finishes, introducing an unwanted dependency.
937 : *
938 : * This function checks the work item address and work function to avoid
939 : * false positives. Note that this isn't complete as one may construct a
940 : * work function which can introduce dependency onto itself through a
941 : * recycled work item. Well, if somebody wants to shoot oneself in the
942 : * foot that badly, there's only so much we can do, and if such deadlock
943 : * actually occurs, it should be easy to locate the culprit work function.
944 : *
945 : * CONTEXT:
946 : * spin_lock_irq(pool->lock).
947 : *
948 : * Return:
949 : * Pointer to worker which is executing @work if found, %NULL
950 : * otherwise.
951 : */
952 42832 : static struct worker *find_worker_executing_work(struct worker_pool *pool,
953 : struct work_struct *work)
954 : {
955 : struct worker *worker;
956 :
957 85664 : hash_for_each_possible(pool->busy_hash, worker, hentry,
958 : (unsigned long)work)
959 6 : if (worker->current_work == work &&
960 3 : worker->current_func == work->func)
961 : return worker;
962 :
963 : return NULL;
964 : }
965 :
966 : /**
967 : * move_linked_works - move linked works to a list
968 : * @work: start of series of works to be scheduled
969 : * @head: target list to append @work to
970 : * @nextp: out paramter for nested worklist walking
971 : *
972 : * Schedule linked works starting from @work to @head. Work series to
973 : * be scheduled starts at @work and includes any consecutive work with
974 : * WORK_STRUCT_LINKED set in its predecessor.
975 : *
976 : * If @nextp is not NULL, it's updated to point to the next work of
977 : * the last scheduled work. This allows move_linked_works() to be
978 : * nested inside outer list_for_each_entry_safe().
979 : *
980 : * CONTEXT:
981 : * spin_lock_irq(pool->lock).
982 : */
983 12 : static void move_linked_works(struct work_struct *work, struct list_head *head,
984 : struct work_struct **nextp)
985 : {
986 : struct work_struct *n;
987 :
988 : /*
989 : * Linked worklist will always end before the end of the list,
990 : * use NULL for list head.
991 : */
992 21 : list_for_each_entry_safe_from(work, n, NULL, entry) {
993 : list_move_tail(&work->entry, head);
994 21 : if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
995 : break;
996 : }
997 :
998 : /*
999 : * If we're already inside safe list traversal and have moved
1000 : * multiple works to the scheduled queue, the next position
1001 : * needs to be updated.
1002 : */
1003 12 : if (nextp)
1004 0 : *nextp = n;
1005 12 : }
1006 :
1007 : /**
1008 : * get_pwq - get an extra reference on the specified pool_workqueue
1009 : * @pwq: pool_workqueue to get
1010 : *
1011 : * Obtain an extra reference on @pwq. The caller should guarantee that
1012 : * @pwq has positive refcnt and be holding the matching pool->lock.
1013 : */
1014 : static void get_pwq(struct pool_workqueue *pwq)
1015 : {
1016 : lockdep_assert_held(&pwq->pool->lock);
1017 : WARN_ON_ONCE(pwq->refcnt <= 0);
1018 42754 : pwq->refcnt++;
1019 : }
1020 :
1021 : /**
1022 : * put_pwq - put a pool_workqueue reference
1023 : * @pwq: pool_workqueue to put
1024 : *
1025 : * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1026 : * destruction. The caller should be holding the matching pool->lock.
1027 : */
1028 42754 : static void put_pwq(struct pool_workqueue *pwq)
1029 : {
1030 : lockdep_assert_held(&pwq->pool->lock);
1031 42754 : if (likely(--pwq->refcnt))
1032 : return;
1033 0 : if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1034 : return;
1035 : /*
1036 : * @pwq can't be released under pool->lock, bounce to
1037 : * pwq_unbound_release_workfn(). This never recurses on the same
1038 : * pool->lock as this path is taken only for unbound workqueues and
1039 : * the release work item is scheduled on a per-cpu workqueue. To
1040 : * avoid lockdep warning, unbound pool->locks are given lockdep
1041 : * subclass of 1 in get_unbound_pool().
1042 : */
1043 0 : schedule_work(&pwq->unbound_release_work);
1044 : }
1045 :
1046 : /**
1047 : * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1048 : * @pwq: pool_workqueue to put (can be %NULL)
1049 : *
1050 : * put_pwq() with locking. This function also allows %NULL @pwq.
1051 : */
1052 26 : static void put_pwq_unlocked(struct pool_workqueue *pwq)
1053 : {
1054 26 : if (pwq) {
1055 : /*
1056 : * As both pwqs and pools are sched-RCU protected, the
1057 : * following lock operations are safe.
1058 : */
1059 : spin_lock_irq(&pwq->pool->lock);
1060 0 : put_pwq(pwq);
1061 : spin_unlock_irq(&pwq->pool->lock);
1062 : }
1063 26 : }
1064 :
1065 0 : static void pwq_activate_delayed_work(struct work_struct *work)
1066 : {
1067 : struct pool_workqueue *pwq = get_work_pwq(work);
1068 :
1069 : trace_workqueue_activate_work(work);
1070 0 : move_linked_works(work, &pwq->pool->worklist, NULL);
1071 : __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1072 0 : pwq->nr_active++;
1073 0 : }
1074 :
1075 : static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1076 : {
1077 0 : struct work_struct *work = list_first_entry(&pwq->delayed_works,
1078 : struct work_struct, entry);
1079 :
1080 0 : pwq_activate_delayed_work(work);
1081 : }
1082 :
1083 : /**
1084 : * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1085 : * @pwq: pwq of interest
1086 : * @color: color of work which left the queue
1087 : *
1088 : * A work either has completed or is removed from pending queue,
1089 : * decrement nr_in_flight of its pwq and handle workqueue flushing.
1090 : *
1091 : * CONTEXT:
1092 : * spin_lock_irq(pool->lock).
1093 : */
1094 42754 : static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1095 : {
1096 : /* uncolored work items don't participate in flushing or nr_active */
1097 42754 : if (color == WORK_NO_COLOR)
1098 : goto out_put;
1099 :
1100 42745 : pwq->nr_in_flight[color]--;
1101 :
1102 42745 : pwq->nr_active--;
1103 85490 : if (!list_empty(&pwq->delayed_works)) {
1104 : /* one down, submit a delayed one */
1105 0 : if (pwq->nr_active < pwq->max_active)
1106 : pwq_activate_first_delayed(pwq);
1107 : }
1108 :
1109 : /* is flush in progress and are we at the flushing tip? */
1110 42745 : if (likely(pwq->flush_color != color))
1111 : goto out_put;
1112 :
1113 : /* are there still in-flight works? */
1114 1 : if (pwq->nr_in_flight[color])
1115 : goto out_put;
1116 :
1117 : /* this pwq is done, clear flush_color */
1118 1 : pwq->flush_color = -1;
1119 :
1120 : /*
1121 : * If this was the last pwq, wake up the first flusher. It
1122 : * will handle the rest.
1123 : */
1124 2 : if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1125 1 : complete(&pwq->wq->first_flusher->done);
1126 : out_put:
1127 42754 : put_pwq(pwq);
1128 42754 : }
1129 :
1130 : /**
1131 : * try_to_grab_pending - steal work item from worklist and disable irq
1132 : * @work: work item to steal
1133 : * @is_dwork: @work is a delayed_work
1134 : * @flags: place to store irq state
1135 : *
1136 : * Try to grab PENDING bit of @work. This function can handle @work in any
1137 : * stable state - idle, on timer or on worklist.
1138 : *
1139 : * Return:
1140 : * 1 if @work was pending and we successfully stole PENDING
1141 : * 0 if @work was idle and we claimed PENDING
1142 : * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1143 : * -ENOENT if someone else is canceling @work, this state may persist
1144 : * for arbitrarily long
1145 : *
1146 : * Note:
1147 : * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1148 : * interrupted while holding PENDING and @work off queue, irq must be
1149 : * disabled on entry. This, combined with delayed_work->timer being
1150 : * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1151 : *
1152 : * On successful return, >= 0, irq is disabled and the caller is
1153 : * responsible for releasing it using local_irq_restore(*@flags).
1154 : *
1155 : * This function is safe to call from any context including IRQ handler.
1156 : */
1157 2204 : static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1158 : unsigned long *flags)
1159 : {
1160 : struct worker_pool *pool;
1161 : struct pool_workqueue *pwq;
1162 :
1163 2204 : local_irq_save(*flags);
1164 :
1165 : /* try to steal the timer if it exists */
1166 2204 : if (is_dwork) {
1167 : struct delayed_work *dwork = to_delayed_work(work);
1168 :
1169 : /*
1170 : * dwork->timer is irqsafe. If del_timer() fails, it's
1171 : * guaranteed that the timer is not queued anywhere and not
1172 : * running on the local CPU.
1173 : */
1174 811 : if (likely(del_timer(&dwork->timer)))
1175 : return 1;
1176 : }
1177 :
1178 : /* try to claim PENDING the normal way */
1179 2201 : if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1180 : return 0;
1181 :
1182 : /*
1183 : * The queueing is in progress, or it is already queued. Try to
1184 : * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1185 : */
1186 0 : pool = get_work_pool(work);
1187 0 : if (!pool)
1188 : goto fail;
1189 :
1190 : spin_lock(&pool->lock);
1191 : /*
1192 : * work->data is guaranteed to point to pwq only while the work
1193 : * item is queued on pwq->wq, and both updating work->data to point
1194 : * to pwq on queueing and to pool on dequeueing are done under
1195 : * pwq->pool->lock. This in turn guarantees that, if work->data
1196 : * points to pwq which is associated with a locked pool, the work
1197 : * item is currently queued on that pool.
1198 : */
1199 : pwq = get_work_pwq(work);
1200 0 : if (pwq && pwq->pool == pool) {
1201 : debug_work_deactivate(work);
1202 :
1203 : /*
1204 : * A delayed work item cannot be grabbed directly because
1205 : * it might have linked NO_COLOR work items which, if left
1206 : * on the delayed_list, will confuse pwq->nr_active
1207 : * management later on and cause stall. Make sure the work
1208 : * item is activated before grabbing.
1209 : */
1210 0 : if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1211 0 : pwq_activate_delayed_work(work);
1212 :
1213 0 : list_del_init(&work->entry);
1214 0 : pwq_dec_nr_in_flight(pwq, get_work_color(work));
1215 :
1216 : /* work->data points to pwq iff queued, point to pool */
1217 0 : set_work_pool_and_keep_pending(work, pool->id);
1218 :
1219 : spin_unlock(&pool->lock);
1220 : return 1;
1221 : }
1222 : spin_unlock(&pool->lock);
1223 : fail:
1224 0 : local_irq_restore(*flags);
1225 0 : if (work_is_canceling(work))
1226 : return -ENOENT;
1227 0 : cpu_relax();
1228 0 : return -EAGAIN;
1229 : }
1230 :
1231 : /**
1232 : * insert_work - insert a work into a pool
1233 : * @pwq: pwq @work belongs to
1234 : * @work: work to insert
1235 : * @head: insertion point
1236 : * @extra_flags: extra WORK_STRUCT_* flags to set
1237 : *
1238 : * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1239 : * work_struct flags.
1240 : *
1241 : * CONTEXT:
1242 : * spin_lock_irq(pool->lock).
1243 : */
1244 42754 : static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1245 : struct list_head *head, unsigned int extra_flags)
1246 : {
1247 42754 : struct worker_pool *pool = pwq->pool;
1248 :
1249 : /* we own @work, set data and link */
1250 : set_work_pwq(work, pwq, extra_flags);
1251 42754 : list_add_tail(&work->entry, head);
1252 : get_pwq(pwq);
1253 :
1254 : /*
1255 : * Ensure either wq_worker_sleeping() sees the above
1256 : * list_add_tail() or we see zero nr_running to avoid workers lying
1257 : * around lazily while there are works to be processed.
1258 : */
1259 42754 : smp_mb();
1260 :
1261 42754 : if (__need_more_worker(pool))
1262 42711 : wake_up_worker(pool);
1263 42754 : }
1264 :
1265 : /*
1266 : * Test whether @work is being queued from another work executing on the
1267 : * same workqueue.
1268 : */
1269 0 : static bool is_chained_work(struct workqueue_struct *wq)
1270 : {
1271 : struct worker *worker;
1272 :
1273 : worker = current_wq_worker();
1274 : /*
1275 : * Return %true iff I'm a worker execuing a work item on @wq. If
1276 : * I'm @worker, it's safe to dereference it without locking.
1277 : */
1278 0 : return worker && worker->current_pwq->wq == wq;
1279 : }
1280 :
1281 42745 : static void __queue_work(int cpu, struct workqueue_struct *wq,
1282 : struct work_struct *work)
1283 : {
1284 : struct pool_workqueue *pwq;
1285 : struct worker_pool *last_pool;
1286 : struct list_head *worklist;
1287 : unsigned int work_flags;
1288 : unsigned int req_cpu = cpu;
1289 :
1290 : /*
1291 : * While a work item is PENDING && off queue, a task trying to
1292 : * steal the PENDING will busy-loop waiting for it to either get
1293 : * queued or lose PENDING. Grabbing PENDING and queueing should
1294 : * happen with IRQ disabled.
1295 : */
1296 : WARN_ON_ONCE(!irqs_disabled());
1297 :
1298 : debug_work_activate(work);
1299 :
1300 : /* if draining, only works from the same workqueue are allowed */
1301 42745 : if (unlikely(wq->flags & __WQ_DRAINING) &&
1302 0 : WARN_ON_ONCE(!is_chained_work(wq)))
1303 : return;
1304 : retry:
1305 : if (req_cpu == WORK_CPU_UNBOUND)
1306 : cpu = raw_smp_processor_id();
1307 :
1308 : /* pwq which will be used unless @work is executing elsewhere */
1309 42745 : if (!(wq->flags & WQ_UNBOUND))
1310 39982 : pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1311 : else
1312 : pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1313 :
1314 : /*
1315 : * If @work was previously on a different pool, it might still be
1316 : * running there, in which case the work needs to be queued on that
1317 : * pool to guarantee non-reentrancy.
1318 : */
1319 42745 : last_pool = get_work_pool(work);
1320 42745 : if (last_pool && last_pool != pwq->pool) {
1321 : struct worker *worker;
1322 :
1323 : spin_lock(&last_pool->lock);
1324 :
1325 2 : worker = find_worker_executing_work(last_pool, work);
1326 :
1327 2 : if (worker && worker->current_pwq->wq == wq) {
1328 : pwq = worker->current_pwq;
1329 : } else {
1330 : /* meh... not running there, queue here */
1331 : spin_unlock(&last_pool->lock);
1332 : spin_lock(&pwq->pool->lock);
1333 : }
1334 : } else {
1335 : spin_lock(&pwq->pool->lock);
1336 : }
1337 :
1338 : /*
1339 : * pwq is determined and locked. For unbound pools, we could have
1340 : * raced with pwq release and it could already be dead. If its
1341 : * refcnt is zero, repeat pwq selection. Note that pwqs never die
1342 : * without another pwq replacing it in the numa_pwq_tbl or while
1343 : * work items are executing on it, so the retrying is guaranteed to
1344 : * make forward-progress.
1345 : */
1346 42745 : if (unlikely(!pwq->refcnt)) {
1347 0 : if (wq->flags & WQ_UNBOUND) {
1348 : spin_unlock(&pwq->pool->lock);
1349 0 : cpu_relax();
1350 : goto retry;
1351 : }
1352 : /* oops */
1353 : WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1354 : wq->name, cpu);
1355 : }
1356 :
1357 : /* pwq determined, queue */
1358 : trace_workqueue_queue_work(req_cpu, pwq, work);
1359 :
1360 85490 : if (WARN_ON(!list_empty(&work->entry))) {
1361 : spin_unlock(&pwq->pool->lock);
1362 : return;
1363 : }
1364 :
1365 42745 : pwq->nr_in_flight[pwq->work_color]++;
1366 : work_flags = work_color_to_flags(pwq->work_color);
1367 :
1368 42745 : if (likely(pwq->nr_active < pwq->max_active)) {
1369 : trace_workqueue_activate_work(work);
1370 42745 : pwq->nr_active++;
1371 42745 : worklist = &pwq->pool->worklist;
1372 : } else {
1373 0 : work_flags |= WORK_STRUCT_DELAYED;
1374 0 : worklist = &pwq->delayed_works;
1375 : }
1376 :
1377 42745 : insert_work(pwq, work, worklist, work_flags);
1378 :
1379 : spin_unlock(&pwq->pool->lock);
1380 : }
1381 :
1382 : /**
1383 : * queue_work_on - queue work on specific cpu
1384 : * @cpu: CPU number to execute work on
1385 : * @wq: workqueue to use
1386 : * @work: work to queue
1387 : *
1388 : * We queue the work to a specific CPU, the caller must ensure it
1389 : * can't go away.
1390 : *
1391 : * Return: %false if @work was already on a queue, %true otherwise.
1392 : */
1393 49149 : bool queue_work_on(int cpu, struct workqueue_struct *wq,
1394 : struct work_struct *work)
1395 : {
1396 : bool ret = false;
1397 : unsigned long flags;
1398 :
1399 : local_irq_save(flags);
1400 :
1401 49149 : if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1402 38902 : __queue_work(cpu, wq, work);
1403 : ret = true;
1404 : }
1405 :
1406 49149 : local_irq_restore(flags);
1407 49149 : return ret;
1408 : }
1409 : EXPORT_SYMBOL(queue_work_on);
1410 :
1411 3806 : void delayed_work_timer_fn(unsigned long __data)
1412 : {
1413 3806 : struct delayed_work *dwork = (struct delayed_work *)__data;
1414 :
1415 : /* should have been called from irqsafe timer with irq already off */
1416 3806 : __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1417 3806 : }
1418 : EXPORT_SYMBOL(delayed_work_timer_fn);
1419 :
1420 3853 : static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1421 : struct delayed_work *dwork, unsigned long delay)
1422 : {
1423 3853 : struct timer_list *timer = &dwork->timer;
1424 : struct work_struct *work = &dwork->work;
1425 :
1426 : WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1427 : timer->data != (unsigned long)dwork);
1428 : WARN_ON_ONCE(timer_pending(timer));
1429 : WARN_ON_ONCE(!list_empty(&work->entry));
1430 :
1431 : /*
1432 : * If @delay is 0, queue @dwork->work immediately. This is for
1433 : * both optimization and correctness. The earliest @timer can
1434 : * expire is on the closest next tick and delayed_work users depend
1435 : * on that there's no such delay when @delay is 0.
1436 : */
1437 3853 : if (!delay) {
1438 37 : __queue_work(cpu, wq, &dwork->work);
1439 3890 : return;
1440 : }
1441 :
1442 : timer_stats_timer_set_start_info(&dwork->timer);
1443 :
1444 3816 : dwork->wq = wq;
1445 3816 : dwork->cpu = cpu;
1446 3816 : timer->expires = jiffies + delay;
1447 :
1448 3816 : if (unlikely(cpu != WORK_CPU_UNBOUND))
1449 802 : add_timer_on(timer, cpu);
1450 : else
1451 3014 : add_timer(timer);
1452 : }
1453 :
1454 : /**
1455 : * queue_delayed_work_on - queue work on specific CPU after delay
1456 : * @cpu: CPU number to execute work on
1457 : * @wq: workqueue to use
1458 : * @dwork: work to queue
1459 : * @delay: number of jiffies to wait before queueing
1460 : *
1461 : * Return: %false if @work was already on a queue, %true otherwise. If
1462 : * @delay is zero and @dwork is idle, it will be scheduled for immediate
1463 : * execution.
1464 : */
1465 3058 : bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1466 : struct delayed_work *dwork, unsigned long delay)
1467 : {
1468 : struct work_struct *work = &dwork->work;
1469 : bool ret = false;
1470 : unsigned long flags;
1471 :
1472 : /* read the comment in __queue_work() */
1473 : local_irq_save(flags);
1474 :
1475 3058 : if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1476 3050 : __queue_delayed_work(cpu, wq, dwork, delay);
1477 : ret = true;
1478 : }
1479 :
1480 3058 : local_irq_restore(flags);
1481 3058 : return ret;
1482 : }
1483 : EXPORT_SYMBOL(queue_delayed_work_on);
1484 :
1485 : /**
1486 : * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1487 : * @cpu: CPU number to execute work on
1488 : * @wq: workqueue to use
1489 : * @dwork: work to queue
1490 : * @delay: number of jiffies to wait before queueing
1491 : *
1492 : * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1493 : * modify @dwork's timer so that it expires after @delay. If @delay is
1494 : * zero, @work is guaranteed to be scheduled immediately regardless of its
1495 : * current state.
1496 : *
1497 : * Return: %false if @dwork was idle and queued, %true if @dwork was
1498 : * pending and its timer was modified.
1499 : *
1500 : * This function is safe to call from any context including IRQ handler.
1501 : * See try_to_grab_pending() for details.
1502 : */
1503 803 : bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1504 : struct delayed_work *dwork, unsigned long delay)
1505 : {
1506 : unsigned long flags;
1507 : int ret;
1508 :
1509 : do {
1510 803 : ret = try_to_grab_pending(&dwork->work, true, &flags);
1511 803 : } while (unlikely(ret == -EAGAIN));
1512 :
1513 803 : if (likely(ret >= 0)) {
1514 803 : __queue_delayed_work(cpu, wq, dwork, delay);
1515 1606 : local_irq_restore(flags);
1516 : }
1517 :
1518 : /* -ENOENT from try_to_grab_pending() becomes %true */
1519 803 : return ret;
1520 : }
1521 : EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1522 :
1523 : /**
1524 : * worker_enter_idle - enter idle state
1525 : * @worker: worker which is entering idle state
1526 : *
1527 : * @worker is entering idle state. Update stats and idle timer if
1528 : * necessary.
1529 : *
1530 : * LOCKING:
1531 : * spin_lock_irq(pool->lock).
1532 : */
1533 42295 : static void worker_enter_idle(struct worker *worker)
1534 : {
1535 42295 : struct worker_pool *pool = worker->pool;
1536 :
1537 84590 : if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1538 84590 : WARN_ON_ONCE(!list_empty(&worker->entry) &&
1539 : (worker->hentry.next || worker->hentry.pprev)))
1540 42295 : return;
1541 :
1542 : /* can't use worker_set_flags(), also called from create_worker() */
1543 42295 : worker->flags |= WORKER_IDLE;
1544 42295 : pool->nr_idle++;
1545 42295 : worker->last_active = jiffies;
1546 :
1547 : /* idle_list is LIFO */
1548 42295 : list_add(&worker->entry, &pool->idle_list);
1549 :
1550 42295 : if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1551 6 : mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1552 :
1553 : /*
1554 : * Sanity check nr_running. Because wq_unbind_fn() releases
1555 : * pool->lock between setting %WORKER_UNBOUND and zapping
1556 : * nr_running, the warning may trigger spuriously. Check iff
1557 : * unbind is not in progress.
1558 : */
1559 42295 : WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1560 : pool->nr_workers == pool->nr_idle &&
1561 : atomic_read(&pool->nr_running));
1562 : }
1563 :
1564 : /**
1565 : * worker_leave_idle - leave idle state
1566 : * @worker: worker which is leaving idle state
1567 : *
1568 : * @worker is leaving idle state. Update stats.
1569 : *
1570 : * LOCKING:
1571 : * spin_lock_irq(pool->lock).
1572 : */
1573 42284 : static void worker_leave_idle(struct worker *worker)
1574 : {
1575 42284 : struct worker_pool *pool = worker->pool;
1576 :
1577 42284 : if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1578 42284 : return;
1579 : worker_clr_flags(worker, WORKER_IDLE);
1580 42284 : pool->nr_idle--;
1581 42284 : list_del_init(&worker->entry);
1582 : }
1583 :
1584 26 : static struct worker *alloc_worker(int node)
1585 : {
1586 : struct worker *worker;
1587 :
1588 : worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
1589 26 : if (worker) {
1590 26 : INIT_LIST_HEAD(&worker->entry);
1591 26 : INIT_LIST_HEAD(&worker->scheduled);
1592 26 : INIT_LIST_HEAD(&worker->node);
1593 : /* on creation a worker is in !idle && prep state */
1594 26 : worker->flags = WORKER_PREP;
1595 : }
1596 26 : return worker;
1597 : }
1598 :
1599 : /**
1600 : * worker_attach_to_pool() - attach a worker to a pool
1601 : * @worker: worker to be attached
1602 : * @pool: the target pool
1603 : *
1604 : * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1605 : * cpu-binding of @worker are kept coordinated with the pool across
1606 : * cpu-[un]hotplugs.
1607 : */
1608 11 : static void worker_attach_to_pool(struct worker *worker,
1609 : struct worker_pool *pool)
1610 : {
1611 11 : mutex_lock(&pool->attach_mutex);
1612 :
1613 : /*
1614 : * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1615 : * online CPUs. It'll be re-applied when any of the CPUs come up.
1616 : */
1617 11 : set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1618 :
1619 : /*
1620 : * The pool->attach_mutex ensures %POOL_DISASSOCIATED remains
1621 : * stable across this function. See the comments above the
1622 : * flag definition for details.
1623 : */
1624 11 : if (pool->flags & POOL_DISASSOCIATED)
1625 6 : worker->flags |= WORKER_UNBOUND;
1626 :
1627 11 : list_add_tail(&worker->node, &pool->workers);
1628 :
1629 11 : mutex_unlock(&pool->attach_mutex);
1630 11 : }
1631 :
1632 : /**
1633 : * worker_detach_from_pool() - detach a worker from its pool
1634 : * @worker: worker which is attached to its pool
1635 : * @pool: the pool @worker is attached to
1636 : *
1637 : * Undo the attaching which had been done in worker_attach_to_pool(). The
1638 : * caller worker shouldn't access to the pool after detached except it has
1639 : * other reference to the pool.
1640 : */
1641 6 : static void worker_detach_from_pool(struct worker *worker,
1642 : struct worker_pool *pool)
1643 : {
1644 : struct completion *detach_completion = NULL;
1645 :
1646 6 : mutex_lock(&pool->attach_mutex);
1647 : list_del(&worker->node);
1648 12 : if (list_empty(&pool->workers))
1649 0 : detach_completion = pool->detach_completion;
1650 6 : mutex_unlock(&pool->attach_mutex);
1651 :
1652 : /* clear leftover flags without pool->lock after it is detached */
1653 6 : worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
1654 :
1655 6 : if (detach_completion)
1656 0 : complete(detach_completion);
1657 6 : }
1658 :
1659 : /**
1660 : * create_worker - create a new workqueue worker
1661 : * @pool: pool the new worker will belong to
1662 : *
1663 : * Create and start a new worker which is attached to @pool.
1664 : *
1665 : * CONTEXT:
1666 : * Might sleep. Does GFP_KERNEL allocations.
1667 : *
1668 : * Return:
1669 : * Pointer to the newly created worker.
1670 : */
1671 11 : static struct worker *create_worker(struct worker_pool *pool)
1672 : {
1673 : struct worker *worker = NULL;
1674 : int id = -1;
1675 : char id_buf[16];
1676 :
1677 : /* ID is needed to determine kthread name */
1678 11 : id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL);
1679 11 : if (id < 0)
1680 : goto fail;
1681 :
1682 11 : worker = alloc_worker(pool->node);
1683 11 : if (!worker)
1684 : goto fail;
1685 :
1686 11 : worker->pool = pool;
1687 11 : worker->id = id;
1688 :
1689 11 : if (pool->cpu >= 0)
1690 5 : snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1691 5 : pool->attrs->nice < 0 ? "H" : "");
1692 : else
1693 6 : snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1694 :
1695 11 : worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1696 : "kworker/%s", id_buf);
1697 11 : if (IS_ERR(worker->task))
1698 : goto fail;
1699 :
1700 11 : set_user_nice(worker->task, pool->attrs->nice);
1701 :
1702 : /* prevent userland from meddling with cpumask of workqueue workers */
1703 11 : worker->task->flags |= PF_NO_SETAFFINITY;
1704 :
1705 : /* successful, attach the worker to the pool */
1706 11 : worker_attach_to_pool(worker, pool);
1707 :
1708 : /* start the newly created worker */
1709 : spin_lock_irq(&pool->lock);
1710 11 : worker->pool->nr_workers++;
1711 11 : worker_enter_idle(worker);
1712 11 : wake_up_process(worker->task);
1713 : spin_unlock_irq(&pool->lock);
1714 :
1715 11 : return worker;
1716 :
1717 : fail:
1718 0 : if (id >= 0)
1719 0 : ida_simple_remove(&pool->worker_ida, id);
1720 0 : kfree(worker);
1721 0 : return NULL;
1722 : }
1723 :
1724 : /**
1725 : * destroy_worker - destroy a workqueue worker
1726 : * @worker: worker to be destroyed
1727 : *
1728 : * Destroy @worker and adjust @pool stats accordingly. The worker should
1729 : * be idle.
1730 : *
1731 : * CONTEXT:
1732 : * spin_lock_irq(pool->lock).
1733 : */
1734 6 : static void destroy_worker(struct worker *worker)
1735 : {
1736 6 : struct worker_pool *pool = worker->pool;
1737 :
1738 : lockdep_assert_held(&pool->lock);
1739 :
1740 : /* sanity check frenzy */
1741 12 : if (WARN_ON(worker->current_work) ||
1742 18 : WARN_ON(!list_empty(&worker->scheduled)) ||
1743 6 : WARN_ON(!(worker->flags & WORKER_IDLE)))
1744 6 : return;
1745 :
1746 6 : pool->nr_workers--;
1747 6 : pool->nr_idle--;
1748 :
1749 6 : list_del_init(&worker->entry);
1750 6 : worker->flags |= WORKER_DIE;
1751 6 : wake_up_process(worker->task);
1752 : }
1753 :
1754 6 : static void idle_worker_timeout(unsigned long __pool)
1755 : {
1756 6 : struct worker_pool *pool = (void *)__pool;
1757 :
1758 : spin_lock_irq(&pool->lock);
1759 :
1760 12 : while (too_many_workers(pool)) {
1761 : struct worker *worker;
1762 : unsigned long expires;
1763 :
1764 : /* idle_list is kept in LIFO order, check the last one */
1765 6 : worker = list_entry(pool->idle_list.prev, struct worker, entry);
1766 6 : expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1767 :
1768 6 : if (time_before(jiffies, expires)) {
1769 0 : mod_timer(&pool->idle_timer, expires);
1770 0 : break;
1771 : }
1772 :
1773 6 : destroy_worker(worker);
1774 : }
1775 :
1776 : spin_unlock_irq(&pool->lock);
1777 6 : }
1778 :
1779 1 : static void send_mayday(struct work_struct *work)
1780 : {
1781 : struct pool_workqueue *pwq = get_work_pwq(work);
1782 1 : struct workqueue_struct *wq = pwq->wq;
1783 :
1784 : lockdep_assert_held(&wq_mayday_lock);
1785 :
1786 1 : if (!wq->rescuer)
1787 1 : return;
1788 :
1789 : /* mayday mayday mayday */
1790 0 : if (list_empty(&pwq->mayday_node)) {
1791 : /*
1792 : * If @pwq is for an unbound wq, its base ref may be put at
1793 : * any time due to an attribute change. Pin @pwq until the
1794 : * rescuer is done with it.
1795 : */
1796 : get_pwq(pwq);
1797 0 : list_add_tail(&pwq->mayday_node, &wq->maydays);
1798 0 : wake_up_process(wq->rescuer->task);
1799 : }
1800 : }
1801 :
1802 1 : static void pool_mayday_timeout(unsigned long __pool)
1803 : {
1804 1 : struct worker_pool *pool = (void *)__pool;
1805 : struct work_struct *work;
1806 :
1807 : spin_lock_irq(&pool->lock);
1808 : spin_lock(&wq_mayday_lock); /* for wq->maydays */
1809 :
1810 1 : if (need_to_create_worker(pool)) {
1811 : /*
1812 : * We've been trying to create a new worker but
1813 : * haven't been successful. We might be hitting an
1814 : * allocation deadlock. Send distress signals to
1815 : * rescuers.
1816 : */
1817 2 : list_for_each_entry(work, &pool->worklist, entry)
1818 1 : send_mayday(work);
1819 : }
1820 :
1821 : spin_unlock(&wq_mayday_lock);
1822 : spin_unlock_irq(&pool->lock);
1823 :
1824 1 : mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1825 1 : }
1826 :
1827 : /**
1828 : * maybe_create_worker - create a new worker if necessary
1829 : * @pool: pool to create a new worker for
1830 : *
1831 : * Create a new worker for @pool if necessary. @pool is guaranteed to
1832 : * have at least one idle worker on return from this function. If
1833 : * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1834 : * sent to all rescuers with works scheduled on @pool to resolve
1835 : * possible allocation deadlock.
1836 : *
1837 : * On return, need_to_create_worker() is guaranteed to be %false and
1838 : * may_start_working() %true.
1839 : *
1840 : * LOCKING:
1841 : * spin_lock_irq(pool->lock) which may be released and regrabbed
1842 : * multiple times. Does GFP_KERNEL allocations. Called only from
1843 : * manager.
1844 : */
1845 8 : static void maybe_create_worker(struct worker_pool *pool)
1846 : __releases(&pool->lock)
1847 : __acquires(&pool->lock)
1848 : {
1849 : restart:
1850 : spin_unlock_irq(&pool->lock);
1851 :
1852 : /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1853 8 : mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1854 :
1855 : while (true) {
1856 8 : if (create_worker(pool) || !need_to_create_worker(pool))
1857 : break;
1858 :
1859 0 : schedule_timeout_interruptible(CREATE_COOLDOWN);
1860 :
1861 0 : if (!need_to_create_worker(pool))
1862 : break;
1863 : }
1864 :
1865 8 : del_timer_sync(&pool->mayday_timer);
1866 : spin_lock_irq(&pool->lock);
1867 : /*
1868 : * This is necessary even after a new worker was just successfully
1869 : * created as @pool->lock was dropped and the new worker might have
1870 : * already become busy.
1871 : */
1872 8 : if (need_to_create_worker(pool))
1873 : goto restart;
1874 8 : }
1875 :
1876 : /**
1877 : * manage_workers - manage worker pool
1878 : * @worker: self
1879 : *
1880 : * Assume the manager role and manage the worker pool @worker belongs
1881 : * to. At any given time, there can be only zero or one manager per
1882 : * pool. The exclusion is handled automatically by this function.
1883 : *
1884 : * The caller can safely start processing works on false return. On
1885 : * true return, it's guaranteed that need_to_create_worker() is false
1886 : * and may_start_working() is true.
1887 : *
1888 : * CONTEXT:
1889 : * spin_lock_irq(pool->lock) which may be released and regrabbed
1890 : * multiple times. Does GFP_KERNEL allocations.
1891 : *
1892 : * Return:
1893 : * %false if the pool doesn't need management and the caller can safely
1894 : * start processing works, %true if management function was performed and
1895 : * the conditions that the caller verified before calling the function may
1896 : * no longer be true.
1897 : */
1898 8 : static bool manage_workers(struct worker *worker)
1899 : {
1900 8 : struct worker_pool *pool = worker->pool;
1901 :
1902 : /*
1903 : * Anyone who successfully grabs manager_arb wins the arbitration
1904 : * and becomes the manager. mutex_trylock() on pool->manager_arb
1905 : * failure while holding pool->lock reliably indicates that someone
1906 : * else is managing the pool and the worker which failed trylock
1907 : * can proceed to executing work items. This means that anyone
1908 : * grabbing manager_arb is responsible for actually performing
1909 : * manager duties. If manager_arb is grabbed and released without
1910 : * actual management, the pool may stall indefinitely.
1911 : */
1912 8 : if (!mutex_trylock(&pool->manager_arb))
1913 : return false;
1914 :
1915 8 : maybe_create_worker(pool);
1916 :
1917 8 : mutex_unlock(&pool->manager_arb);
1918 : return true;
1919 : }
1920 :
1921 : /**
1922 : * process_one_work - process single work
1923 : * @worker: self
1924 : * @work: work to process
1925 : *
1926 : * Process @work. This function contains all the logics necessary to
1927 : * process a single work including synchronization against and
1928 : * interaction with other workers on the same cpu, queueing and
1929 : * flushing. As long as context requirement is met, any worker can
1930 : * call this function to process a work.
1931 : *
1932 : * CONTEXT:
1933 : * spin_lock_irq(pool->lock) which is released and regrabbed.
1934 : */
1935 42757 : static void process_one_work(struct worker *worker, struct work_struct *work)
1936 : __releases(&pool->lock)
1937 : __acquires(&pool->lock)
1938 : {
1939 : struct pool_workqueue *pwq = get_work_pwq(work);
1940 42757 : struct worker_pool *pool = worker->pool;
1941 42757 : bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
1942 : int work_color;
1943 : struct worker *collision;
1944 : #ifdef CONFIG_LOCKDEP
1945 : /*
1946 : * It is permissible to free the struct work_struct from
1947 : * inside the function that is called from it, this we need to
1948 : * take into account for lockdep too. To avoid bogus "held
1949 : * lock freed" warnings as well as problems when looking into
1950 : * work->lockdep_map, make a copy and use that here.
1951 : */
1952 : struct lockdep_map lockdep_map;
1953 :
1954 : lockdep_copy_map(&lockdep_map, &work->lockdep_map);
1955 : #endif
1956 : /* ensure we're on the correct CPU */
1957 : WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1958 : raw_smp_processor_id() != pool->cpu);
1959 :
1960 : /*
1961 : * A single work shouldn't be executed concurrently by
1962 : * multiple workers on a single cpu. Check whether anyone is
1963 : * already processing the work. If so, defer the work to the
1964 : * currently executing one.
1965 : */
1966 42757 : collision = find_worker_executing_work(pool, work);
1967 42757 : if (unlikely(collision)) {
1968 3 : move_linked_works(work, &collision->scheduled, NULL);
1969 42760 : return;
1970 : }
1971 :
1972 : /* claim and dequeue */
1973 : debug_work_deactivate(work);
1974 85508 : hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
1975 42754 : worker->current_work = work;
1976 42754 : worker->current_func = work->func;
1977 42754 : worker->current_pwq = pwq;
1978 : work_color = get_work_color(work);
1979 :
1980 42754 : list_del_init(&work->entry);
1981 :
1982 : /*
1983 : * CPU intensive works don't participate in concurrency management.
1984 : * They're the scheduler's responsibility. This takes @worker out
1985 : * of concurrency management and the next code block will chain
1986 : * execution of the pending work items.
1987 : */
1988 42754 : if (unlikely(cpu_intensive))
1989 : worker_set_flags(worker, WORKER_CPU_INTENSIVE);
1990 :
1991 : /*
1992 : * Wake up another worker if necessary. The condition is always
1993 : * false for normal per-cpu workers since nr_running would always
1994 : * be >= 1 at this point. This is used to chain execution of the
1995 : * pending work items for WORKER_NOT_RUNNING workers such as the
1996 : * UNBOUND and CPU_INTENSIVE ones.
1997 : */
1998 42754 : if (need_more_worker(pool))
1999 4 : wake_up_worker(pool);
2000 :
2001 : /*
2002 : * Record the last pool and clear PENDING which should be the last
2003 : * update to @work. Also, do this inside @pool->lock so that
2004 : * PENDING and queued state changes happen together while IRQ is
2005 : * disabled.
2006 : */
2007 42754 : set_work_pool_and_clear_pending(work, pool->id);
2008 :
2009 : spin_unlock_irq(&pool->lock);
2010 :
2011 : lock_map_acquire_read(&pwq->wq->lockdep_map);
2012 : lock_map_acquire(&lockdep_map);
2013 : trace_workqueue_execute_start(work);
2014 42754 : worker->current_func(work);
2015 : /*
2016 : * While we must be careful to not use "work" after this, the trace
2017 : * point will only record its address.
2018 : */
2019 : trace_workqueue_execute_end(work);
2020 : lock_map_release(&lockdep_map);
2021 : lock_map_release(&pwq->wq->lockdep_map);
2022 :
2023 42754 : if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2024 0 : pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2025 : " last function: %pf\n",
2026 : current->comm, preempt_count(), task_pid_nr(current),
2027 : worker->current_func);
2028 : debug_show_held_locks(current);
2029 0 : dump_stack();
2030 : }
2031 :
2032 : /*
2033 : * The following prevents a kworker from hogging CPU on !PREEMPT
2034 : * kernels, where a requeueing work item waiting for something to
2035 : * happen could deadlock with stop_machine as such work item could
2036 : * indefinitely requeue itself while all other CPUs are trapped in
2037 : * stop_machine. At the same time, report a quiescent RCU state so
2038 : * the same condition doesn't freeze RCU.
2039 : */
2040 42754 : cond_resched_rcu_qs();
2041 :
2042 : spin_lock_irq(&pool->lock);
2043 :
2044 : /* clear cpu intensive status */
2045 42754 : if (unlikely(cpu_intensive))
2046 : worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2047 :
2048 : /* we're done with it, release */
2049 : hash_del(&worker->hentry);
2050 42754 : worker->current_work = NULL;
2051 42754 : worker->current_func = NULL;
2052 42754 : worker->current_pwq = NULL;
2053 42754 : worker->desc_valid = false;
2054 42754 : pwq_dec_nr_in_flight(pwq, work_color);
2055 : }
2056 :
2057 : /**
2058 : * process_scheduled_works - process scheduled works
2059 : * @worker: self
2060 : *
2061 : * Process all scheduled works. Please note that the scheduled list
2062 : * may change while processing a work, so this function repeatedly
2063 : * fetches a work from the top and executes it.
2064 : *
2065 : * CONTEXT:
2066 : * spin_lock_irq(pool->lock) which may be released and regrabbed
2067 : * multiple times.
2068 : */
2069 : static void process_scheduled_works(struct worker *worker)
2070 : {
2071 33 : while (!list_empty(&worker->scheduled)) {
2072 21 : struct work_struct *work = list_first_entry(&worker->scheduled,
2073 : struct work_struct, entry);
2074 21 : process_one_work(worker, work);
2075 : }
2076 : }
2077 :
2078 : /**
2079 : * worker_thread - the worker thread function
2080 : * @__worker: self
2081 : *
2082 : * The worker thread function. All workers belong to a worker_pool -
2083 : * either a per-cpu one or dynamic unbound one. These workers process all
2084 : * work items regardless of their specific target workqueue. The only
2085 : * exception is work items which belong to workqueues with a rescuer which
2086 : * will be explained in rescuer_thread().
2087 : *
2088 : * Return: 0
2089 : */
2090 11 : static int worker_thread(void *__worker)
2091 : {
2092 84494 : struct worker *worker = __worker;
2093 42262 : struct worker_pool *pool = worker->pool;
2094 :
2095 : /* tell the scheduler that this is a workqueue worker */
2096 11 : worker->task->flags |= PF_WQ_WORKER;
2097 : woke_up:
2098 : spin_lock_irq(&pool->lock);
2099 :
2100 : /* am I supposed to die? */
2101 42290 : if (unlikely(worker->flags & WORKER_DIE)) {
2102 : spin_unlock_irq(&pool->lock);
2103 : WARN_ON_ONCE(!list_empty(&worker->entry));
2104 6 : worker->task->flags &= ~PF_WQ_WORKER;
2105 :
2106 6 : set_task_comm(worker->task, "kworker/dying");
2107 6 : ida_simple_remove(&pool->worker_ida, worker->id);
2108 6 : worker_detach_from_pool(worker, pool);
2109 6 : kfree(worker);
2110 6 : return 0;
2111 : }
2112 :
2113 42284 : worker_leave_idle(worker);
2114 : recheck:
2115 : /* no more worker necessary? */
2116 42292 : if (!need_more_worker(pool))
2117 : goto sleep;
2118 :
2119 : /* do we need to manage? */
2120 42259 : if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2121 : goto recheck;
2122 :
2123 : /*
2124 : * ->scheduled list can only be filled while a worker is
2125 : * preparing to process a work or actually processing it.
2126 : * Make sure nobody diddled with it while I was sleeping.
2127 : */
2128 : WARN_ON_ONCE(!list_empty(&worker->scheduled));
2129 :
2130 : /*
2131 : * Finish PREP stage. We're guaranteed to have at least one idle
2132 : * worker or that someone else has already assumed the manager
2133 : * role. This is where @worker starts participating in concurrency
2134 : * management if applicable and concurrency management is restored
2135 : * after being rebound. See rebind_workers() for details.
2136 : */
2137 : worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2138 :
2139 : do {
2140 42745 : struct work_struct *work =
2141 42745 : list_first_entry(&pool->worklist,
2142 : struct work_struct, entry);
2143 :
2144 42745 : if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2145 : /* optimization path, not strictly necessary */
2146 42736 : process_one_work(worker, work);
2147 85472 : if (unlikely(!list_empty(&worker->scheduled)))
2148 : process_scheduled_works(worker);
2149 : } else {
2150 9 : move_linked_works(work, &worker->scheduled, NULL);
2151 : process_scheduled_works(worker);
2152 : }
2153 42745 : } while (keep_working(pool));
2154 :
2155 : worker_set_flags(worker, WORKER_PREP);
2156 : sleep:
2157 : /*
2158 : * pool->lock is held and there's no work to process and no need to
2159 : * manage, sleep. Workers are woken up only while holding
2160 : * pool->lock or from local cpu, so setting the current state
2161 : * before releasing pool->lock is enough to prevent losing any
2162 : * event.
2163 : */
2164 42284 : worker_enter_idle(worker);
2165 42284 : __set_current_state(TASK_INTERRUPTIBLE);
2166 : spin_unlock_irq(&pool->lock);
2167 42284 : schedule();
2168 42279 : goto woke_up;
2169 : }
2170 :
2171 : /**
2172 : * rescuer_thread - the rescuer thread function
2173 : * @__rescuer: self
2174 : *
2175 : * Workqueue rescuer thread function. There's one rescuer for each
2176 : * workqueue which has WQ_MEM_RECLAIM set.
2177 : *
2178 : * Regular work processing on a pool may block trying to create a new
2179 : * worker which uses GFP_KERNEL allocation which has slight chance of
2180 : * developing into deadlock if some works currently on the same queue
2181 : * need to be processed to satisfy the GFP_KERNEL allocation. This is
2182 : * the problem rescuer solves.
2183 : *
2184 : * When such condition is possible, the pool summons rescuers of all
2185 : * workqueues which have works queued on the pool and let them process
2186 : * those works so that forward progress can be guaranteed.
2187 : *
2188 : * This should happen rarely.
2189 : *
2190 : * Return: 0
2191 : */
2192 15 : static int rescuer_thread(void *__rescuer)
2193 : {
2194 : struct worker *rescuer = __rescuer;
2195 15 : struct workqueue_struct *wq = rescuer->rescue_wq;
2196 15 : struct list_head *scheduled = &rescuer->scheduled;
2197 : bool should_stop;
2198 :
2199 15 : set_user_nice(current, RESCUER_NICE_LEVEL);
2200 :
2201 : /*
2202 : * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2203 : * doesn't participate in concurrency management.
2204 : */
2205 15 : rescuer->task->flags |= PF_WQ_WORKER;
2206 : repeat:
2207 15 : set_current_state(TASK_INTERRUPTIBLE);
2208 :
2209 : /*
2210 : * By the time the rescuer is requested to stop, the workqueue
2211 : * shouldn't have any work pending, but @wq->maydays may still have
2212 : * pwq(s) queued. This can happen by non-rescuer workers consuming
2213 : * all the work items before the rescuer got to them. Go through
2214 : * @wq->maydays processing before acting on should_stop so that the
2215 : * list is always empty on exit.
2216 : */
2217 15 : should_stop = kthread_should_stop();
2218 :
2219 : /* see whether any pwq is asking for help */
2220 : spin_lock_irq(&wq_mayday_lock);
2221 :
2222 30 : while (!list_empty(&wq->maydays)) {
2223 0 : struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2224 : struct pool_workqueue, mayday_node);
2225 0 : struct worker_pool *pool = pwq->pool;
2226 : struct work_struct *work, *n;
2227 :
2228 0 : __set_current_state(TASK_RUNNING);
2229 0 : list_del_init(&pwq->mayday_node);
2230 :
2231 : spin_unlock_irq(&wq_mayday_lock);
2232 :
2233 0 : worker_attach_to_pool(rescuer, pool);
2234 :
2235 : spin_lock_irq(&pool->lock);
2236 0 : rescuer->pool = pool;
2237 :
2238 : /*
2239 : * Slurp in all works issued via this workqueue and
2240 : * process'em.
2241 : */
2242 : WARN_ON_ONCE(!list_empty(scheduled));
2243 0 : list_for_each_entry_safe(work, n, &pool->worklist, entry)
2244 0 : if (get_work_pwq(work) == pwq)
2245 0 : move_linked_works(work, scheduled, &n);
2246 :
2247 0 : if (!list_empty(scheduled)) {
2248 : process_scheduled_works(rescuer);
2249 :
2250 : /*
2251 : * The above execution of rescued work items could
2252 : * have created more to rescue through
2253 : * pwq_activate_first_delayed() or chained
2254 : * queueing. Let's put @pwq back on mayday list so
2255 : * that such back-to-back work items, which may be
2256 : * being used to relieve memory pressure, don't
2257 : * incur MAYDAY_INTERVAL delay inbetween.
2258 : */
2259 0 : if (need_to_create_worker(pool)) {
2260 : spin_lock(&wq_mayday_lock);
2261 : get_pwq(pwq);
2262 : list_move_tail(&pwq->mayday_node, &wq->maydays);
2263 : spin_unlock(&wq_mayday_lock);
2264 : }
2265 : }
2266 :
2267 : /*
2268 : * Put the reference grabbed by send_mayday(). @pool won't
2269 : * go away while we're still attached to it.
2270 : */
2271 0 : put_pwq(pwq);
2272 :
2273 : /*
2274 : * Leave this pool. If need_more_worker() is %true, notify a
2275 : * regular worker; otherwise, we end up with 0 concurrency
2276 : * and stalling the execution.
2277 : */
2278 0 : if (need_more_worker(pool))
2279 0 : wake_up_worker(pool);
2280 :
2281 0 : rescuer->pool = NULL;
2282 : spin_unlock_irq(&pool->lock);
2283 :
2284 0 : worker_detach_from_pool(rescuer, pool);
2285 :
2286 : spin_lock_irq(&wq_mayday_lock);
2287 : }
2288 :
2289 : spin_unlock_irq(&wq_mayday_lock);
2290 :
2291 15 : if (should_stop) {
2292 0 : __set_current_state(TASK_RUNNING);
2293 0 : rescuer->task->flags &= ~PF_WQ_WORKER;
2294 0 : return 0;
2295 : }
2296 :
2297 : /* rescuers should never participate in concurrency management */
2298 : WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2299 15 : schedule();
2300 0 : goto repeat;
2301 : }
2302 :
2303 : struct wq_barrier {
2304 : struct work_struct work;
2305 : struct completion done;
2306 : };
2307 :
2308 9 : static void wq_barrier_func(struct work_struct *work)
2309 : {
2310 : struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2311 9 : complete(&barr->done);
2312 9 : }
2313 :
2314 : /**
2315 : * insert_wq_barrier - insert a barrier work
2316 : * @pwq: pwq to insert barrier into
2317 : * @barr: wq_barrier to insert
2318 : * @target: target work to attach @barr to
2319 : * @worker: worker currently executing @target, NULL if @target is not executing
2320 : *
2321 : * @barr is linked to @target such that @barr is completed only after
2322 : * @target finishes execution. Please note that the ordering
2323 : * guarantee is observed only with respect to @target and on the local
2324 : * cpu.
2325 : *
2326 : * Currently, a queued barrier can't be canceled. This is because
2327 : * try_to_grab_pending() can't determine whether the work to be
2328 : * grabbed is at the head of the queue and thus can't clear LINKED
2329 : * flag of the previous work while there must be a valid next work
2330 : * after a work with LINKED flag set.
2331 : *
2332 : * Note that when @worker is non-NULL, @target may be modified
2333 : * underneath us, so we can't reliably determine pwq from @target.
2334 : *
2335 : * CONTEXT:
2336 : * spin_lock_irq(pool->lock).
2337 : */
2338 9 : static void insert_wq_barrier(struct pool_workqueue *pwq,
2339 : struct wq_barrier *barr,
2340 : struct work_struct *target, struct worker *worker)
2341 : {
2342 : struct list_head *head;
2343 : unsigned int linked = 0;
2344 :
2345 : /*
2346 : * debugobject calls are safe here even with pool->lock locked
2347 : * as we know for sure that this will not trigger any of the
2348 : * checks and call back into the fixup functions where we
2349 : * might deadlock.
2350 : */
2351 18 : INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2352 : __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2353 : init_completion(&barr->done);
2354 :
2355 : /*
2356 : * If @target is currently being executed, schedule the
2357 : * barrier to the worker; otherwise, put it after @target.
2358 : */
2359 9 : if (worker)
2360 0 : head = worker->scheduled.next;
2361 : else {
2362 : unsigned long *bits = work_data_bits(target);
2363 :
2364 9 : head = target->entry.next;
2365 : /* there can already be other linked works, inherit and set */
2366 9 : linked = *bits & WORK_STRUCT_LINKED;
2367 : __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2368 : }
2369 :
2370 : debug_work_activate(&barr->work);
2371 9 : insert_work(pwq, &barr->work, head,
2372 : work_color_to_flags(WORK_NO_COLOR) | linked);
2373 9 : }
2374 :
2375 : /**
2376 : * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2377 : * @wq: workqueue being flushed
2378 : * @flush_color: new flush color, < 0 for no-op
2379 : * @work_color: new work color, < 0 for no-op
2380 : *
2381 : * Prepare pwqs for workqueue flushing.
2382 : *
2383 : * If @flush_color is non-negative, flush_color on all pwqs should be
2384 : * -1. If no pwq has in-flight commands at the specified color, all
2385 : * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2386 : * has in flight commands, its pwq->flush_color is set to
2387 : * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2388 : * wakeup logic is armed and %true is returned.
2389 : *
2390 : * The caller should have initialized @wq->first_flusher prior to
2391 : * calling this function with non-negative @flush_color. If
2392 : * @flush_color is negative, no flush color update is done and %false
2393 : * is returned.
2394 : *
2395 : * If @work_color is non-negative, all pwqs should have the same
2396 : * work_color which is previous to @work_color and all will be
2397 : * advanced to @work_color.
2398 : *
2399 : * CONTEXT:
2400 : * mutex_lock(wq->mutex).
2401 : *
2402 : * Return:
2403 : * %true if @flush_color >= 0 and there's something to flush. %false
2404 : * otherwise.
2405 : */
2406 1 : static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2407 : int flush_color, int work_color)
2408 : {
2409 : bool wait = false;
2410 : struct pool_workqueue *pwq;
2411 :
2412 1 : if (flush_color >= 0) {
2413 1 : WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2414 1 : atomic_set(&wq->nr_pwqs_to_flush, 1);
2415 : }
2416 :
2417 2 : for_each_pwq(pwq, wq) {
2418 : struct worker_pool *pool = pwq->pool;
2419 :
2420 : spin_lock_irq(&pool->lock);
2421 :
2422 1 : if (flush_color >= 0) {
2423 : WARN_ON_ONCE(pwq->flush_color != -1);
2424 :
2425 1 : if (pwq->nr_in_flight[flush_color]) {
2426 1 : pwq->flush_color = flush_color;
2427 1 : atomic_inc(&wq->nr_pwqs_to_flush);
2428 : wait = true;
2429 : }
2430 : }
2431 :
2432 1 : if (work_color >= 0) {
2433 : WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2434 1 : pwq->work_color = work_color;
2435 : }
2436 :
2437 : spin_unlock_irq(&pool->lock);
2438 : }
2439 :
2440 2 : if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2441 0 : complete(&wq->first_flusher->done);
2442 :
2443 1 : return wait;
2444 : }
2445 :
2446 : /**
2447 : * flush_workqueue - ensure that any scheduled work has run to completion.
2448 : * @wq: workqueue to flush
2449 : *
2450 : * This function sleeps until all work items which were queued on entry
2451 : * have finished execution, but it is not livelocked by new incoming ones.
2452 : */
2453 1 : void flush_workqueue(struct workqueue_struct *wq)
2454 : {
2455 2 : struct wq_flusher this_flusher = {
2456 : .list = LIST_HEAD_INIT(this_flusher.list),
2457 : .flush_color = -1,
2458 1 : .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2459 : };
2460 : int next_color;
2461 :
2462 : lock_map_acquire(&wq->lockdep_map);
2463 : lock_map_release(&wq->lockdep_map);
2464 :
2465 1 : mutex_lock(&wq->mutex);
2466 :
2467 : /*
2468 : * Start-to-wait phase
2469 : */
2470 1 : next_color = work_next_color(wq->work_color);
2471 :
2472 1 : if (next_color != wq->flush_color) {
2473 : /*
2474 : * Color space is not full. The current work_color
2475 : * becomes our flush_color and work_color is advanced
2476 : * by one.
2477 : */
2478 : WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2479 1 : this_flusher.flush_color = wq->work_color;
2480 1 : wq->work_color = next_color;
2481 :
2482 1 : if (!wq->first_flusher) {
2483 : /* no flush in progress, become the first flusher */
2484 : WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2485 :
2486 1 : wq->first_flusher = &this_flusher;
2487 :
2488 1 : if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2489 : wq->work_color)) {
2490 : /* nothing to flush, done */
2491 0 : wq->flush_color = next_color;
2492 0 : wq->first_flusher = NULL;
2493 0 : goto out_unlock;
2494 : }
2495 : } else {
2496 : /* wait in queue */
2497 : WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2498 0 : list_add_tail(&this_flusher.list, &wq->flusher_queue);
2499 0 : flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2500 : }
2501 : } else {
2502 : /*
2503 : * Oops, color space is full, wait on overflow queue.
2504 : * The next flush completion will assign us
2505 : * flush_color and transfer to flusher_queue.
2506 : */
2507 0 : list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2508 : }
2509 :
2510 1 : mutex_unlock(&wq->mutex);
2511 :
2512 1 : wait_for_completion(&this_flusher.done);
2513 :
2514 : /*
2515 : * Wake-up-and-cascade phase
2516 : *
2517 : * First flushers are responsible for cascading flushes and
2518 : * handling overflow. Non-first flushers can simply return.
2519 : */
2520 1 : if (wq->first_flusher != &this_flusher)
2521 0 : return;
2522 :
2523 1 : mutex_lock(&wq->mutex);
2524 :
2525 : /* we might have raced, check again with mutex held */
2526 1 : if (wq->first_flusher != &this_flusher)
2527 : goto out_unlock;
2528 :
2529 1 : wq->first_flusher = NULL;
2530 :
2531 : WARN_ON_ONCE(!list_empty(&this_flusher.list));
2532 : WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2533 :
2534 : while (true) {
2535 : struct wq_flusher *next, *tmp;
2536 :
2537 : /* complete all the flushers sharing the current flush color */
2538 1 : list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2539 0 : if (next->flush_color != wq->flush_color)
2540 : break;
2541 : list_del_init(&next->list);
2542 0 : complete(&next->done);
2543 : }
2544 :
2545 1 : WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2546 : wq->flush_color != work_next_color(wq->work_color));
2547 :
2548 : /* this flush_color is finished, advance by one */
2549 2 : wq->flush_color = work_next_color(wq->flush_color);
2550 :
2551 : /* one color has been freed, handle overflow queue */
2552 1 : if (!list_empty(&wq->flusher_overflow)) {
2553 : /*
2554 : * Assign the same color to all overflowed
2555 : * flushers, advance work_color and append to
2556 : * flusher_queue. This is the start-to-wait
2557 : * phase for these overflowed flushers.
2558 : */
2559 0 : list_for_each_entry(tmp, &wq->flusher_overflow, list)
2560 0 : tmp->flush_color = wq->work_color;
2561 :
2562 0 : wq->work_color = work_next_color(wq->work_color);
2563 :
2564 : list_splice_tail_init(&wq->flusher_overflow,
2565 : &wq->flusher_queue);
2566 0 : flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2567 : }
2568 :
2569 1 : if (list_empty(&wq->flusher_queue)) {
2570 : WARN_ON_ONCE(wq->flush_color != wq->work_color);
2571 : break;
2572 : }
2573 :
2574 : /*
2575 : * Need to flush more colors. Make the next flusher
2576 : * the new first flusher and arm pwqs.
2577 : */
2578 : WARN_ON_ONCE(wq->flush_color == wq->work_color);
2579 : WARN_ON_ONCE(wq->flush_color != next->flush_color);
2580 :
2581 : list_del_init(&next->list);
2582 0 : wq->first_flusher = next;
2583 :
2584 0 : if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2585 : break;
2586 :
2587 : /*
2588 : * Meh... this color is already done, clear first
2589 : * flusher and repeat cascading.
2590 : */
2591 0 : wq->first_flusher = NULL;
2592 0 : }
2593 :
2594 : out_unlock:
2595 1 : mutex_unlock(&wq->mutex);
2596 : }
2597 : EXPORT_SYMBOL_GPL(flush_workqueue);
2598 :
2599 : /**
2600 : * drain_workqueue - drain a workqueue
2601 : * @wq: workqueue to drain
2602 : *
2603 : * Wait until the workqueue becomes empty. While draining is in progress,
2604 : * only chain queueing is allowed. IOW, only currently pending or running
2605 : * work items on @wq can queue further work items on it. @wq is flushed
2606 : * repeatedly until it becomes empty. The number of flushing is detemined
2607 : * by the depth of chaining and should be relatively short. Whine if it
2608 : * takes too long.
2609 : */
2610 0 : void drain_workqueue(struct workqueue_struct *wq)
2611 : {
2612 : unsigned int flush_cnt = 0;
2613 : struct pool_workqueue *pwq;
2614 :
2615 : /*
2616 : * __queue_work() needs to test whether there are drainers, is much
2617 : * hotter than drain_workqueue() and already looks at @wq->flags.
2618 : * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2619 : */
2620 0 : mutex_lock(&wq->mutex);
2621 0 : if (!wq->nr_drainers++)
2622 0 : wq->flags |= __WQ_DRAINING;
2623 0 : mutex_unlock(&wq->mutex);
2624 : reflush:
2625 0 : flush_workqueue(wq);
2626 :
2627 0 : mutex_lock(&wq->mutex);
2628 :
2629 0 : for_each_pwq(pwq, wq) {
2630 : bool drained;
2631 :
2632 : spin_lock_irq(&pwq->pool->lock);
2633 0 : drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2634 : spin_unlock_irq(&pwq->pool->lock);
2635 :
2636 0 : if (drained)
2637 0 : continue;
2638 :
2639 0 : if (++flush_cnt == 10 ||
2640 0 : (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2641 0 : pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2642 : wq->name, flush_cnt);
2643 :
2644 0 : mutex_unlock(&wq->mutex);
2645 0 : goto reflush;
2646 : }
2647 :
2648 0 : if (!--wq->nr_drainers)
2649 0 : wq->flags &= ~__WQ_DRAINING;
2650 0 : mutex_unlock(&wq->mutex);
2651 0 : }
2652 : EXPORT_SYMBOL_GPL(drain_workqueue);
2653 :
2654 3795 : static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2655 : {
2656 : struct worker *worker = NULL;
2657 : struct worker_pool *pool;
2658 : struct pool_workqueue *pwq;
2659 :
2660 : might_sleep();
2661 :
2662 : local_irq_disable();
2663 3795 : pool = get_work_pool(work);
2664 3795 : if (!pool) {
2665 : local_irq_enable();
2666 3713 : return false;
2667 : }
2668 :
2669 : spin_lock(&pool->lock);
2670 : /* see the comment in try_to_grab_pending() with the same code */
2671 : pwq = get_work_pwq(work);
2672 82 : if (pwq) {
2673 9 : if (unlikely(pwq->pool != pool))
2674 : goto already_gone;
2675 : } else {
2676 73 : worker = find_worker_executing_work(pool, work);
2677 73 : if (!worker)
2678 : goto already_gone;
2679 0 : pwq = worker->current_pwq;
2680 : }
2681 :
2682 9 : insert_wq_barrier(pwq, barr, work, worker);
2683 : spin_unlock_irq(&pool->lock);
2684 :
2685 : /*
2686 : * If @max_active is 1 or rescuer is in use, flushing another work
2687 : * item on the same workqueue may lead to deadlock. Make sure the
2688 : * flusher is not running on the same workqueue by verifying write
2689 : * access.
2690 : */
2691 : if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)
2692 : lock_map_acquire(&pwq->wq->lockdep_map);
2693 : else
2694 : lock_map_acquire_read(&pwq->wq->lockdep_map);
2695 : lock_map_release(&pwq->wq->lockdep_map);
2696 :
2697 : return true;
2698 : already_gone:
2699 : spin_unlock_irq(&pool->lock);
2700 : return false;
2701 : }
2702 :
2703 : /**
2704 : * flush_work - wait for a work to finish executing the last queueing instance
2705 : * @work: the work to flush
2706 : *
2707 : * Wait until @work has finished execution. @work is guaranteed to be idle
2708 : * on return if it hasn't been requeued since flush started.
2709 : *
2710 : * Return:
2711 : * %true if flush_work() waited for the work to finish execution,
2712 : * %false if it was already idle.
2713 : */
2714 3795 : bool flush_work(struct work_struct *work)
2715 : {
2716 : struct wq_barrier barr;
2717 :
2718 : lock_map_acquire(&work->lockdep_map);
2719 : lock_map_release(&work->lockdep_map);
2720 :
2721 3795 : if (start_flush_work(work, &barr)) {
2722 9 : wait_for_completion(&barr.done);
2723 : destroy_work_on_stack(&barr.work);
2724 9 : return true;
2725 : } else {
2726 : return false;
2727 : }
2728 : }
2729 : EXPORT_SYMBOL_GPL(flush_work);
2730 :
2731 : struct cwt_wait {
2732 : wait_queue_t wait;
2733 : struct work_struct *work;
2734 : };
2735 :
2736 0 : static int cwt_wakefn(wait_queue_t *wait, unsigned mode, int sync, void *key)
2737 : {
2738 : struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
2739 :
2740 0 : if (cwait->work != key)
2741 : return 0;
2742 0 : return autoremove_wake_function(wait, mode, sync, key);
2743 : }
2744 :
2745 1398 : static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2746 : {
2747 : static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
2748 : unsigned long flags;
2749 : int ret;
2750 :
2751 : do {
2752 1398 : ret = try_to_grab_pending(work, is_dwork, &flags);
2753 : /*
2754 : * If someone else is already canceling, wait for it to
2755 : * finish. flush_work() doesn't work for PREEMPT_NONE
2756 : * because we may get scheduled between @work's completion
2757 : * and the other canceling task resuming and clearing
2758 : * CANCELING - flush_work() will return false immediately
2759 : * as @work is no longer busy, try_to_grab_pending() will
2760 : * return -ENOENT as @work is still being canceled and the
2761 : * other canceling task won't be able to clear CANCELING as
2762 : * we're hogging the CPU.
2763 : *
2764 : * Let's wait for completion using a waitqueue. As this
2765 : * may lead to the thundering herd problem, use a custom
2766 : * wake function which matches @work along with exclusive
2767 : * wait and wakeup.
2768 : */
2769 1398 : if (unlikely(ret == -ENOENT)) {
2770 : struct cwt_wait cwait;
2771 :
2772 0 : init_wait(&cwait.wait);
2773 0 : cwait.wait.func = cwt_wakefn;
2774 0 : cwait.work = work;
2775 :
2776 0 : prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
2777 : TASK_UNINTERRUPTIBLE);
2778 0 : if (work_is_canceling(work))
2779 0 : schedule();
2780 0 : finish_wait(&cancel_waitq, &cwait.wait);
2781 : }
2782 1398 : } while (unlikely(ret < 0));
2783 :
2784 : /* tell other tasks trying to grab @work to back off */
2785 : mark_work_canceling(work);
2786 2796 : local_irq_restore(flags);
2787 :
2788 1398 : flush_work(work);
2789 : clear_work_data(work);
2790 :
2791 : /*
2792 : * Paired with prepare_to_wait() above so that either
2793 : * waitqueue_active() is visible here or !work_is_canceling() is
2794 : * visible there.
2795 : */
2796 1398 : smp_mb();
2797 1398 : if (waitqueue_active(&cancel_waitq))
2798 0 : __wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
2799 :
2800 1398 : return ret;
2801 : }
2802 :
2803 : /**
2804 : * cancel_work_sync - cancel a work and wait for it to finish
2805 : * @work: the work to cancel
2806 : *
2807 : * Cancel @work and wait for its execution to finish. This function
2808 : * can be used even if the work re-queues itself or migrates to
2809 : * another workqueue. On return from this function, @work is
2810 : * guaranteed to be not pending or executing on any CPU.
2811 : *
2812 : * cancel_work_sync(&delayed_work->work) must not be used for
2813 : * delayed_work's. Use cancel_delayed_work_sync() instead.
2814 : *
2815 : * The caller must ensure that the workqueue on which @work was last
2816 : * queued can't be destroyed before this function returns.
2817 : *
2818 : * Return:
2819 : * %true if @work was pending, %false otherwise.
2820 : */
2821 1393 : bool cancel_work_sync(struct work_struct *work)
2822 : {
2823 1393 : return __cancel_work_timer(work, false);
2824 : }
2825 : EXPORT_SYMBOL_GPL(cancel_work_sync);
2826 :
2827 : /**
2828 : * flush_delayed_work - wait for a dwork to finish executing the last queueing
2829 : * @dwork: the delayed work to flush
2830 : *
2831 : * Delayed timer is cancelled and the pending work is queued for
2832 : * immediate execution. Like flush_work(), this function only
2833 : * considers the last queueing instance of @dwork.
2834 : *
2835 : * Return:
2836 : * %true if flush_work() waited for the work to finish execution,
2837 : * %false if it was already idle.
2838 : */
2839 0 : bool flush_delayed_work(struct delayed_work *dwork)
2840 : {
2841 : local_irq_disable();
2842 0 : if (del_timer_sync(&dwork->timer))
2843 0 : __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2844 : local_irq_enable();
2845 0 : return flush_work(&dwork->work);
2846 : }
2847 : EXPORT_SYMBOL(flush_delayed_work);
2848 :
2849 : /**
2850 : * cancel_delayed_work - cancel a delayed work
2851 : * @dwork: delayed_work to cancel
2852 : *
2853 : * Kill off a pending delayed_work.
2854 : *
2855 : * Return: %true if @dwork was pending and canceled; %false if it wasn't
2856 : * pending.
2857 : *
2858 : * Note:
2859 : * The work callback function may still be running on return, unless
2860 : * it returns %true and the work doesn't re-arm itself. Explicitly flush or
2861 : * use cancel_delayed_work_sync() to wait on it.
2862 : *
2863 : * This function is safe to call from any context including IRQ handler.
2864 : */
2865 3 : bool cancel_delayed_work(struct delayed_work *dwork)
2866 : {
2867 : unsigned long flags;
2868 : int ret;
2869 :
2870 : do {
2871 3 : ret = try_to_grab_pending(&dwork->work, true, &flags);
2872 3 : } while (unlikely(ret == -EAGAIN));
2873 :
2874 3 : if (unlikely(ret < 0))
2875 : return false;
2876 :
2877 : set_work_pool_and_clear_pending(&dwork->work,
2878 : get_work_pool_id(&dwork->work));
2879 6 : local_irq_restore(flags);
2880 3 : return ret;
2881 : }
2882 : EXPORT_SYMBOL(cancel_delayed_work);
2883 :
2884 : /**
2885 : * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2886 : * @dwork: the delayed work cancel
2887 : *
2888 : * This is cancel_work_sync() for delayed works.
2889 : *
2890 : * Return:
2891 : * %true if @dwork was pending, %false otherwise.
2892 : */
2893 5 : bool cancel_delayed_work_sync(struct delayed_work *dwork)
2894 : {
2895 5 : return __cancel_work_timer(&dwork->work, true);
2896 : }
2897 : EXPORT_SYMBOL(cancel_delayed_work_sync);
2898 :
2899 : /**
2900 : * schedule_on_each_cpu - execute a function synchronously on each online CPU
2901 : * @func: the function to call
2902 : *
2903 : * schedule_on_each_cpu() executes @func on each online CPU using the
2904 : * system workqueue and blocks until all CPUs have completed.
2905 : * schedule_on_each_cpu() is very slow.
2906 : *
2907 : * Return:
2908 : * 0 on success, -errno on failure.
2909 : */
2910 0 : int schedule_on_each_cpu(work_func_t func)
2911 : {
2912 : int cpu;
2913 : struct work_struct __percpu *works;
2914 :
2915 0 : works = alloc_percpu(struct work_struct);
2916 0 : if (!works)
2917 : return -ENOMEM;
2918 :
2919 : get_online_cpus();
2920 :
2921 0 : for_each_online_cpu(cpu) {
2922 : struct work_struct *work = per_cpu_ptr(works, cpu);
2923 :
2924 0 : INIT_WORK(work, func);
2925 : schedule_work_on(cpu, work);
2926 : }
2927 :
2928 0 : for_each_online_cpu(cpu)
2929 0 : flush_work(per_cpu_ptr(works, cpu));
2930 :
2931 : put_online_cpus();
2932 0 : free_percpu(works);
2933 0 : return 0;
2934 : }
2935 :
2936 : /**
2937 : * flush_scheduled_work - ensure that any scheduled work has run to completion.
2938 : *
2939 : * Forces execution of the kernel-global workqueue and blocks until its
2940 : * completion.
2941 : *
2942 : * Think twice before calling this function! It's very easy to get into
2943 : * trouble if you don't take great care. Either of the following situations
2944 : * will lead to deadlock:
2945 : *
2946 : * One of the work items currently on the workqueue needs to acquire
2947 : * a lock held by your code or its caller.
2948 : *
2949 : * Your code is running in the context of a work routine.
2950 : *
2951 : * They will be detected by lockdep when they occur, but the first might not
2952 : * occur very often. It depends on what work items are on the workqueue and
2953 : * what locks they need, which you have no control over.
2954 : *
2955 : * In most situations flushing the entire workqueue is overkill; you merely
2956 : * need to know that a particular work item isn't queued and isn't running.
2957 : * In such cases you should use cancel_delayed_work_sync() or
2958 : * cancel_work_sync() instead.
2959 : */
2960 0 : void flush_scheduled_work(void)
2961 : {
2962 0 : flush_workqueue(system_wq);
2963 0 : }
2964 : EXPORT_SYMBOL(flush_scheduled_work);
2965 :
2966 : /**
2967 : * execute_in_process_context - reliably execute the routine with user context
2968 : * @fn: the function to execute
2969 : * @ew: guaranteed storage for the execute work structure (must
2970 : * be available when the work executes)
2971 : *
2972 : * Executes the function immediately if process context is available,
2973 : * otherwise schedules the function for delayed execution.
2974 : *
2975 : * Return: 0 - function was executed
2976 : * 1 - function was scheduled for execution
2977 : */
2978 0 : int execute_in_process_context(work_func_t fn, struct execute_work *ew)
2979 : {
2980 0 : if (!in_interrupt()) {
2981 0 : fn(&ew->work);
2982 0 : return 0;
2983 : }
2984 :
2985 0 : INIT_WORK(&ew->work, fn);
2986 0 : schedule_work(&ew->work);
2987 :
2988 0 : return 1;
2989 : }
2990 : EXPORT_SYMBOL_GPL(execute_in_process_context);
2991 :
2992 : #ifdef CONFIG_SYSFS
2993 : /*
2994 : * Workqueues with WQ_SYSFS flag set is visible to userland via
2995 : * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
2996 : * following attributes.
2997 : *
2998 : * per_cpu RO bool : whether the workqueue is per-cpu or unbound
2999 : * max_active RW int : maximum number of in-flight work items
3000 : *
3001 : * Unbound workqueues have the following extra attributes.
3002 : *
3003 : * id RO int : the associated pool ID
3004 : * nice RW int : nice value of the workers
3005 : * cpumask RW mask : bitmask of allowed CPUs for the workers
3006 : */
3007 : struct wq_device {
3008 : struct workqueue_struct *wq;
3009 : struct device dev;
3010 : };
3011 :
3012 : static struct workqueue_struct *dev_to_wq(struct device *dev)
3013 : {
3014 : struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
3015 :
3016 0 : return wq_dev->wq;
3017 : }
3018 :
3019 0 : static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
3020 : char *buf)
3021 : {
3022 : struct workqueue_struct *wq = dev_to_wq(dev);
3023 :
3024 0 : return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
3025 : }
3026 : static DEVICE_ATTR_RO(per_cpu);
3027 :
3028 0 : static ssize_t max_active_show(struct device *dev,
3029 : struct device_attribute *attr, char *buf)
3030 : {
3031 : struct workqueue_struct *wq = dev_to_wq(dev);
3032 :
3033 0 : return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
3034 : }
3035 :
3036 0 : static ssize_t max_active_store(struct device *dev,
3037 : struct device_attribute *attr, const char *buf,
3038 : size_t count)
3039 : {
3040 : struct workqueue_struct *wq = dev_to_wq(dev);
3041 : int val;
3042 :
3043 0 : if (sscanf(buf, "%d", &val) != 1 || val <= 0)
3044 : return -EINVAL;
3045 :
3046 0 : workqueue_set_max_active(wq, val);
3047 0 : return count;
3048 : }
3049 : static DEVICE_ATTR_RW(max_active);
3050 :
3051 : static struct attribute *wq_sysfs_attrs[] = {
3052 : &dev_attr_per_cpu.attr,
3053 : &dev_attr_max_active.attr,
3054 : NULL,
3055 : };
3056 : ATTRIBUTE_GROUPS(wq_sysfs);
3057 :
3058 0 : static ssize_t wq_pool_ids_show(struct device *dev,
3059 : struct device_attribute *attr, char *buf)
3060 : {
3061 : struct workqueue_struct *wq = dev_to_wq(dev);
3062 : const char *delim = "";
3063 : int node, written = 0;
3064 :
3065 : rcu_read_lock_sched();
3066 0 : for_each_node(node) {
3067 0 : written += scnprintf(buf + written, PAGE_SIZE - written,
3068 : "%s%d:%d", delim, node,
3069 0 : unbound_pwq_by_node(wq, node)->pool->id);
3070 : delim = " ";
3071 : }
3072 0 : written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
3073 : rcu_read_unlock_sched();
3074 :
3075 0 : return written;
3076 : }
3077 :
3078 0 : static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
3079 : char *buf)
3080 : {
3081 : struct workqueue_struct *wq = dev_to_wq(dev);
3082 : int written;
3083 :
3084 0 : mutex_lock(&wq->mutex);
3085 0 : written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
3086 0 : mutex_unlock(&wq->mutex);
3087 :
3088 0 : return written;
3089 : }
3090 :
3091 : /* prepare workqueue_attrs for sysfs store operations */
3092 0 : static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
3093 : {
3094 : struct workqueue_attrs *attrs;
3095 :
3096 0 : attrs = alloc_workqueue_attrs(GFP_KERNEL);
3097 0 : if (!attrs)
3098 : return NULL;
3099 :
3100 0 : mutex_lock(&wq->mutex);
3101 0 : copy_workqueue_attrs(attrs, wq->unbound_attrs);
3102 0 : mutex_unlock(&wq->mutex);
3103 0 : return attrs;
3104 : }
3105 :
3106 0 : static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
3107 : const char *buf, size_t count)
3108 : {
3109 : struct workqueue_struct *wq = dev_to_wq(dev);
3110 : struct workqueue_attrs *attrs;
3111 : int ret;
3112 :
3113 0 : attrs = wq_sysfs_prep_attrs(wq);
3114 0 : if (!attrs)
3115 : return -ENOMEM;
3116 :
3117 0 : if (sscanf(buf, "%d", &attrs->nice) == 1 &&
3118 0 : attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
3119 0 : ret = apply_workqueue_attrs(wq, attrs);
3120 : else
3121 : ret = -EINVAL;
3122 :
3123 : free_workqueue_attrs(attrs);
3124 0 : return ret ?: count;
3125 : }
3126 :
3127 0 : static ssize_t wq_cpumask_show(struct device *dev,
3128 : struct device_attribute *attr, char *buf)
3129 : {
3130 : struct workqueue_struct *wq = dev_to_wq(dev);
3131 : int written;
3132 :
3133 0 : mutex_lock(&wq->mutex);
3134 0 : written = cpumask_scnprintf(buf, PAGE_SIZE, wq->unbound_attrs->cpumask);
3135 0 : mutex_unlock(&wq->mutex);
3136 :
3137 0 : written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
3138 0 : return written;
3139 : }
3140 :
3141 0 : static ssize_t wq_cpumask_store(struct device *dev,
3142 : struct device_attribute *attr,
3143 : const char *buf, size_t count)
3144 : {
3145 : struct workqueue_struct *wq = dev_to_wq(dev);
3146 : struct workqueue_attrs *attrs;
3147 : int ret;
3148 :
3149 0 : attrs = wq_sysfs_prep_attrs(wq);
3150 0 : if (!attrs)
3151 : return -ENOMEM;
3152 :
3153 : ret = cpumask_parse(buf, attrs->cpumask);
3154 0 : if (!ret)
3155 0 : ret = apply_workqueue_attrs(wq, attrs);
3156 :
3157 : free_workqueue_attrs(attrs);
3158 0 : return ret ?: count;
3159 : }
3160 :
3161 0 : static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
3162 : char *buf)
3163 : {
3164 : struct workqueue_struct *wq = dev_to_wq(dev);
3165 : int written;
3166 :
3167 0 : mutex_lock(&wq->mutex);
3168 0 : written = scnprintf(buf, PAGE_SIZE, "%d\n",
3169 0 : !wq->unbound_attrs->no_numa);
3170 0 : mutex_unlock(&wq->mutex);
3171 :
3172 0 : return written;
3173 : }
3174 :
3175 0 : static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
3176 : const char *buf, size_t count)
3177 : {
3178 : struct workqueue_struct *wq = dev_to_wq(dev);
3179 : struct workqueue_attrs *attrs;
3180 : int v, ret;
3181 :
3182 0 : attrs = wq_sysfs_prep_attrs(wq);
3183 0 : if (!attrs)
3184 : return -ENOMEM;
3185 :
3186 : ret = -EINVAL;
3187 0 : if (sscanf(buf, "%d", &v) == 1) {
3188 0 : attrs->no_numa = !v;
3189 0 : ret = apply_workqueue_attrs(wq, attrs);
3190 : }
3191 :
3192 : free_workqueue_attrs(attrs);
3193 0 : return ret ?: count;
3194 : }
3195 :
3196 : static struct device_attribute wq_sysfs_unbound_attrs[] = {
3197 : __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
3198 : __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
3199 : __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
3200 : __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
3201 : __ATTR_NULL,
3202 : };
3203 :
3204 : static struct bus_type wq_subsys = {
3205 : .name = "workqueue",
3206 : .dev_groups = wq_sysfs_groups,
3207 : };
3208 :
3209 1 : static int __init wq_sysfs_init(void)
3210 : {
3211 1 : return subsys_virtual_register(&wq_subsys, NULL);
3212 : }
3213 : core_initcall(wq_sysfs_init);
3214 :
3215 0 : static void wq_device_release(struct device *dev)
3216 : {
3217 0 : struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
3218 :
3219 0 : kfree(wq_dev);
3220 0 : }
3221 :
3222 : /**
3223 : * workqueue_sysfs_register - make a workqueue visible in sysfs
3224 : * @wq: the workqueue to register
3225 : *
3226 : * Expose @wq in sysfs under /sys/bus/workqueue/devices.
3227 : * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
3228 : * which is the preferred method.
3229 : *
3230 : * Workqueue user should use this function directly iff it wants to apply
3231 : * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
3232 : * apply_workqueue_attrs() may race against userland updating the
3233 : * attributes.
3234 : *
3235 : * Return: 0 on success, -errno on failure.
3236 : */
3237 1 : int workqueue_sysfs_register(struct workqueue_struct *wq)
3238 : {
3239 : struct wq_device *wq_dev;
3240 : int ret;
3241 :
3242 : /*
3243 : * Adjusting max_active or creating new pwqs by applyting
3244 : * attributes breaks ordering guarantee. Disallow exposing ordered
3245 : * workqueues.
3246 : */
3247 1 : if (WARN_ON(wq->flags & __WQ_ORDERED))
3248 : return -EINVAL;
3249 :
3250 1 : wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
3251 1 : if (!wq_dev)
3252 : return -ENOMEM;
3253 :
3254 1 : wq_dev->wq = wq;
3255 1 : wq_dev->dev.bus = &wq_subsys;
3256 1 : wq_dev->dev.init_name = wq->name;
3257 1 : wq_dev->dev.release = wq_device_release;
3258 :
3259 : /*
3260 : * unbound_attrs are created separately. Suppress uevent until
3261 : * everything is ready.
3262 : */
3263 : dev_set_uevent_suppress(&wq_dev->dev, true);
3264 :
3265 1 : ret = device_register(&wq_dev->dev);
3266 1 : if (ret) {
3267 0 : kfree(wq_dev);
3268 0 : wq->wq_dev = NULL;
3269 0 : return ret;
3270 : }
3271 :
3272 1 : if (wq->flags & WQ_UNBOUND) {
3273 : struct device_attribute *attr;
3274 :
3275 4 : for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
3276 4 : ret = device_create_file(&wq_dev->dev, attr);
3277 4 : if (ret) {
3278 0 : device_unregister(&wq_dev->dev);
3279 0 : wq->wq_dev = NULL;
3280 0 : return ret;
3281 : }
3282 : }
3283 : }
3284 :
3285 : dev_set_uevent_suppress(&wq_dev->dev, false);
3286 1 : kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
3287 1 : return 0;
3288 : }
3289 :
3290 : /**
3291 : * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
3292 : * @wq: the workqueue to unregister
3293 : *
3294 : * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
3295 : */
3296 : static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
3297 : {
3298 0 : struct wq_device *wq_dev = wq->wq_dev;
3299 :
3300 0 : if (!wq->wq_dev)
3301 : return;
3302 :
3303 0 : wq->wq_dev = NULL;
3304 0 : device_unregister(&wq_dev->dev);
3305 : }
3306 : #else /* CONFIG_SYSFS */
3307 : static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
3308 : #endif /* CONFIG_SYSFS */
3309 :
3310 : /**
3311 : * free_workqueue_attrs - free a workqueue_attrs
3312 : * @attrs: workqueue_attrs to free
3313 : *
3314 : * Undo alloc_workqueue_attrs().
3315 : */
3316 0 : void free_workqueue_attrs(struct workqueue_attrs *attrs)
3317 : {
3318 26 : if (attrs) {
3319 : free_cpumask_var(attrs->cpumask);
3320 26 : kfree(attrs);
3321 : }
3322 0 : }
3323 :
3324 : /**
3325 : * alloc_workqueue_attrs - allocate a workqueue_attrs
3326 : * @gfp_mask: allocation mask to use
3327 : *
3328 : * Allocate a new workqueue_attrs, initialize with default settings and
3329 : * return it.
3330 : *
3331 : * Return: The allocated new workqueue_attr on success. %NULL on failure.
3332 : */
3333 46 : struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3334 : {
3335 : struct workqueue_attrs *attrs;
3336 :
3337 : attrs = kzalloc(sizeof(*attrs), gfp_mask);
3338 46 : if (!attrs)
3339 : goto fail;
3340 : if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3341 : goto fail;
3342 :
3343 46 : cpumask_copy(attrs->cpumask, cpu_possible_mask);
3344 46 : return attrs;
3345 : fail:
3346 : free_workqueue_attrs(attrs);
3347 : return NULL;
3348 : }
3349 :
3350 : static void copy_workqueue_attrs(struct workqueue_attrs *to,
3351 : const struct workqueue_attrs *from)
3352 : {
3353 40 : to->nice = from->nice;
3354 : cpumask_copy(to->cpumask, from->cpumask);
3355 : /*
3356 : * Unlike hash and equality test, this function doesn't ignore
3357 : * ->no_numa as it is used for both pool and wq attrs. Instead,
3358 : * get_unbound_pool() explicitly clears ->no_numa after copying.
3359 : */
3360 40 : to->no_numa = from->no_numa;
3361 : }
3362 :
3363 : /* hash value of the content of @attr */
3364 13 : static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3365 : {
3366 : u32 hash = 0;
3367 :
3368 13 : hash = jhash_1word(attrs->nice, hash);
3369 13 : hash = jhash(cpumask_bits(attrs->cpumask),
3370 : BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3371 13 : return hash;
3372 : }
3373 :
3374 : /* content equality test */
3375 : static bool wqattrs_equal(const struct workqueue_attrs *a,
3376 : const struct workqueue_attrs *b)
3377 : {
3378 12 : if (a->nice != b->nice)
3379 : return false;
3380 12 : if (!cpumask_equal(a->cpumask, b->cpumask))
3381 : return false;
3382 : return true;
3383 : }
3384 :
3385 : /**
3386 : * init_worker_pool - initialize a newly zalloc'd worker_pool
3387 : * @pool: worker_pool to initialize
3388 : *
3389 : * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs.
3390 : *
3391 : * Return: 0 on success, -errno on failure. Even on failure, all fields
3392 : * inside @pool proper are initialized and put_unbound_pool() can be called
3393 : * on @pool safely to release it.
3394 : */
3395 3 : static int init_worker_pool(struct worker_pool *pool)
3396 : {
3397 : spin_lock_init(&pool->lock);
3398 3 : pool->id = -1;
3399 3 : pool->cpu = -1;
3400 3 : pool->node = NUMA_NO_NODE;
3401 3 : pool->flags |= POOL_DISASSOCIATED;
3402 3 : INIT_LIST_HEAD(&pool->worklist);
3403 3 : INIT_LIST_HEAD(&pool->idle_list);
3404 3 : hash_init(pool->busy_hash);
3405 :
3406 3 : init_timer_deferrable(&pool->idle_timer);
3407 3 : pool->idle_timer.function = idle_worker_timeout;
3408 3 : pool->idle_timer.data = (unsigned long)pool;
3409 :
3410 3 : setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3411 : (unsigned long)pool);
3412 :
3413 3 : mutex_init(&pool->manager_arb);
3414 3 : mutex_init(&pool->attach_mutex);
3415 3 : INIT_LIST_HEAD(&pool->workers);
3416 :
3417 3 : ida_init(&pool->worker_ida);
3418 : INIT_HLIST_NODE(&pool->hash_node);
3419 3 : pool->refcnt = 1;
3420 :
3421 : /* shouldn't fail above this point */
3422 3 : pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3423 3 : if (!pool->attrs)
3424 : return -ENOMEM;
3425 3 : return 0;
3426 : }
3427 :
3428 0 : static void rcu_free_pool(struct rcu_head *rcu)
3429 : {
3430 0 : struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3431 :
3432 0 : ida_destroy(&pool->worker_ida);
3433 0 : free_workqueue_attrs(pool->attrs);
3434 0 : kfree(pool);
3435 0 : }
3436 :
3437 : /**
3438 : * put_unbound_pool - put a worker_pool
3439 : * @pool: worker_pool to put
3440 : *
3441 : * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3442 : * safe manner. get_unbound_pool() calls this function on its failure path
3443 : * and this function should be able to release pools which went through,
3444 : * successfully or not, init_worker_pool().
3445 : *
3446 : * Should be called with wq_pool_mutex held.
3447 : */
3448 0 : static void put_unbound_pool(struct worker_pool *pool)
3449 : {
3450 0 : DECLARE_COMPLETION_ONSTACK(detach_completion);
3451 : struct worker *worker;
3452 :
3453 : lockdep_assert_held(&wq_pool_mutex);
3454 :
3455 0 : if (--pool->refcnt)
3456 0 : return;
3457 :
3458 : /* sanity checks */
3459 0 : if (WARN_ON(!(pool->cpu < 0)) ||
3460 0 : WARN_ON(!list_empty(&pool->worklist)))
3461 : return;
3462 :
3463 : /* release id and unhash */
3464 0 : if (pool->id >= 0)
3465 0 : idr_remove(&worker_pool_idr, pool->id);
3466 : hash_del(&pool->hash_node);
3467 :
3468 : /*
3469 : * Become the manager and destroy all workers. Grabbing
3470 : * manager_arb prevents @pool's workers from blocking on
3471 : * attach_mutex.
3472 : */
3473 0 : mutex_lock(&pool->manager_arb);
3474 :
3475 : spin_lock_irq(&pool->lock);
3476 0 : while ((worker = first_idle_worker(pool)))
3477 0 : destroy_worker(worker);
3478 : WARN_ON(pool->nr_workers || pool->nr_idle);
3479 : spin_unlock_irq(&pool->lock);
3480 :
3481 0 : mutex_lock(&pool->attach_mutex);
3482 0 : if (!list_empty(&pool->workers))
3483 0 : pool->detach_completion = &detach_completion;
3484 0 : mutex_unlock(&pool->attach_mutex);
3485 :
3486 0 : if (pool->detach_completion)
3487 0 : wait_for_completion(pool->detach_completion);
3488 :
3489 0 : mutex_unlock(&pool->manager_arb);
3490 :
3491 : /* shut down the timers */
3492 0 : del_timer_sync(&pool->idle_timer);
3493 0 : del_timer_sync(&pool->mayday_timer);
3494 :
3495 : /* sched-RCU protected to allow dereferences from get_work_pool() */
3496 0 : call_rcu_sched(&pool->rcu, rcu_free_pool);
3497 : }
3498 :
3499 : /**
3500 : * get_unbound_pool - get a worker_pool with the specified attributes
3501 : * @attrs: the attributes of the worker_pool to get
3502 : *
3503 : * Obtain a worker_pool which has the same attributes as @attrs, bump the
3504 : * reference count and return it. If there already is a matching
3505 : * worker_pool, it will be used; otherwise, this function attempts to
3506 : * create a new one.
3507 : *
3508 : * Should be called with wq_pool_mutex held.
3509 : *
3510 : * Return: On success, a worker_pool with the same attributes as @attrs.
3511 : * On failure, %NULL.
3512 : */
3513 13 : static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3514 : {
3515 13 : u32 hash = wqattrs_hash(attrs);
3516 : struct worker_pool *pool;
3517 : int node;
3518 :
3519 : lockdep_assert_held(&wq_pool_mutex);
3520 :
3521 : /* do we already have a matching pool? */
3522 13 : hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3523 24 : if (wqattrs_equal(pool->attrs, attrs)) {
3524 12 : pool->refcnt++;
3525 12 : return pool;
3526 : }
3527 : }
3528 :
3529 : /* nope, create a new one */
3530 : pool = kzalloc(sizeof(*pool), GFP_KERNEL);
3531 1 : if (!pool || init_worker_pool(pool) < 0)
3532 : goto fail;
3533 :
3534 : lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3535 1 : copy_workqueue_attrs(pool->attrs, attrs);
3536 :
3537 : /*
3538 : * no_numa isn't a worker_pool attribute, always clear it. See
3539 : * 'struct workqueue_attrs' comments for detail.
3540 : */
3541 1 : pool->attrs->no_numa = false;
3542 :
3543 : /* if cpumask is contained inside a NUMA node, we belong to that node */
3544 1 : if (wq_numa_enabled) {
3545 0 : for_each_node(node) {
3546 0 : if (cpumask_subset(pool->attrs->cpumask,
3547 0 : wq_numa_possible_cpumask[node])) {
3548 0 : pool->node = node;
3549 0 : break;
3550 : }
3551 : }
3552 : }
3553 :
3554 1 : if (worker_pool_assign_id(pool) < 0)
3555 : goto fail;
3556 :
3557 : /* create and start the initial worker */
3558 1 : if (!create_worker(pool))
3559 : goto fail;
3560 :
3561 : /* install */
3562 1 : hash_add(unbound_pool_hash, &pool->hash_node, hash);
3563 :
3564 1 : return pool;
3565 : fail:
3566 0 : if (pool)
3567 0 : put_unbound_pool(pool);
3568 : return NULL;
3569 : }
3570 :
3571 0 : static void rcu_free_pwq(struct rcu_head *rcu)
3572 : {
3573 0 : kmem_cache_free(pwq_cache,
3574 0 : container_of(rcu, struct pool_workqueue, rcu));
3575 0 : }
3576 :
3577 : /*
3578 : * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3579 : * and needs to be destroyed.
3580 : */
3581 0 : static void pwq_unbound_release_workfn(struct work_struct *work)
3582 : {
3583 : struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3584 : unbound_release_work);
3585 0 : struct workqueue_struct *wq = pwq->wq;
3586 0 : struct worker_pool *pool = pwq->pool;
3587 : bool is_last;
3588 :
3589 0 : if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3590 0 : return;
3591 :
3592 0 : mutex_lock(&wq->mutex);
3593 : list_del_rcu(&pwq->pwqs_node);
3594 0 : is_last = list_empty(&wq->pwqs);
3595 0 : mutex_unlock(&wq->mutex);
3596 :
3597 0 : mutex_lock(&wq_pool_mutex);
3598 0 : put_unbound_pool(pool);
3599 0 : mutex_unlock(&wq_pool_mutex);
3600 :
3601 0 : call_rcu_sched(&pwq->rcu, rcu_free_pwq);
3602 :
3603 : /*
3604 : * If we're the last pwq going away, @wq is already dead and no one
3605 : * is gonna access it anymore. Free it.
3606 : */
3607 0 : if (is_last) {
3608 0 : free_workqueue_attrs(wq->unbound_attrs);
3609 0 : kfree(wq);
3610 : }
3611 : }
3612 :
3613 : /**
3614 : * pwq_adjust_max_active - update a pwq's max_active to the current setting
3615 : * @pwq: target pool_workqueue
3616 : *
3617 : * If @pwq isn't freezing, set @pwq->max_active to the associated
3618 : * workqueue's saved_max_active and activate delayed work items
3619 : * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3620 : */
3621 56 : static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3622 : {
3623 56 : struct workqueue_struct *wq = pwq->wq;
3624 56 : bool freezable = wq->flags & WQ_FREEZABLE;
3625 :
3626 : /* for @wq->saved_max_active */
3627 : lockdep_assert_held(&wq->mutex);
3628 :
3629 : /* fast exit for non-freezable wqs */
3630 56 : if (!freezable && pwq->max_active == wq->saved_max_active)
3631 56 : return;
3632 :
3633 : spin_lock_irq(&pwq->pool->lock);
3634 :
3635 : /*
3636 : * During [un]freezing, the caller is responsible for ensuring that
3637 : * this function is called at least once after @workqueue_freezing
3638 : * is updated and visible.
3639 : */
3640 32 : if (!freezable || !workqueue_freezing) {
3641 32 : pwq->max_active = wq->saved_max_active;
3642 :
3643 96 : while (!list_empty(&pwq->delayed_works) &&
3644 0 : pwq->nr_active < pwq->max_active)
3645 : pwq_activate_first_delayed(pwq);
3646 :
3647 : /*
3648 : * Need to kick a worker after thawed or an unbound wq's
3649 : * max_active is bumped. It's a slow path. Do it always.
3650 : */
3651 32 : wake_up_worker(pwq->pool);
3652 : } else {
3653 0 : pwq->max_active = 0;
3654 : }
3655 :
3656 : spin_unlock_irq(&pwq->pool->lock);
3657 : }
3658 :
3659 : /* initialize newly alloced @pwq which is associated with @wq and @pool */
3660 28 : static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3661 : struct worker_pool *pool)
3662 : {
3663 : BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3664 :
3665 28 : memset(pwq, 0, sizeof(*pwq));
3666 :
3667 28 : pwq->pool = pool;
3668 28 : pwq->wq = wq;
3669 28 : pwq->flush_color = -1;
3670 28 : pwq->refcnt = 1;
3671 28 : INIT_LIST_HEAD(&pwq->delayed_works);
3672 28 : INIT_LIST_HEAD(&pwq->pwqs_node);
3673 28 : INIT_LIST_HEAD(&pwq->mayday_node);
3674 56 : INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3675 28 : }
3676 :
3677 : /* sync @pwq with the current state of its associated wq and link it */
3678 41 : static void link_pwq(struct pool_workqueue *pwq)
3679 : {
3680 41 : struct workqueue_struct *wq = pwq->wq;
3681 :
3682 : lockdep_assert_held(&wq->mutex);
3683 :
3684 : /* may be called multiple times, ignore if already linked */
3685 82 : if (!list_empty(&pwq->pwqs_node))
3686 41 : return;
3687 :
3688 : /* set the matching work_color */
3689 28 : pwq->work_color = wq->work_color;
3690 :
3691 : /* sync max_active to the current setting */
3692 28 : pwq_adjust_max_active(pwq);
3693 :
3694 : /* link in @pwq */
3695 28 : list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3696 : }
3697 :
3698 : /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3699 13 : static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3700 : const struct workqueue_attrs *attrs)
3701 : {
3702 : struct worker_pool *pool;
3703 : struct pool_workqueue *pwq;
3704 :
3705 : lockdep_assert_held(&wq_pool_mutex);
3706 :
3707 13 : pool = get_unbound_pool(attrs);
3708 13 : if (!pool)
3709 : return NULL;
3710 :
3711 13 : pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3712 13 : if (!pwq) {
3713 0 : put_unbound_pool(pool);
3714 0 : return NULL;
3715 : }
3716 :
3717 13 : init_pwq(pwq, wq, pool);
3718 13 : return pwq;
3719 : }
3720 :
3721 : /* undo alloc_unbound_pwq(), used only in the error path */
3722 0 : static void free_unbound_pwq(struct pool_workqueue *pwq)
3723 : {
3724 : lockdep_assert_held(&wq_pool_mutex);
3725 :
3726 0 : if (pwq) {
3727 0 : put_unbound_pool(pwq->pool);
3728 0 : kmem_cache_free(pwq_cache, pwq);
3729 : }
3730 0 : }
3731 :
3732 : /**
3733 : * wq_calc_node_mask - calculate a wq_attrs' cpumask for the specified node
3734 : * @attrs: the wq_attrs of interest
3735 : * @node: the target NUMA node
3736 : * @cpu_going_down: if >= 0, the CPU to consider as offline
3737 : * @cpumask: outarg, the resulting cpumask
3738 : *
3739 : * Calculate the cpumask a workqueue with @attrs should use on @node. If
3740 : * @cpu_going_down is >= 0, that cpu is considered offline during
3741 : * calculation. The result is stored in @cpumask.
3742 : *
3743 : * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3744 : * enabled and @node has online CPUs requested by @attrs, the returned
3745 : * cpumask is the intersection of the possible CPUs of @node and
3746 : * @attrs->cpumask.
3747 : *
3748 : * The caller is responsible for ensuring that the cpumask of @node stays
3749 : * stable.
3750 : *
3751 : * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3752 : * %false if equal.
3753 : */
3754 13 : static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3755 : int cpu_going_down, cpumask_t *cpumask)
3756 : {
3757 13 : if (!wq_numa_enabled || attrs->no_numa)
3758 : goto use_dfl;
3759 :
3760 : /* does @node have any online CPUs @attrs wants? */
3761 0 : cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3762 0 : if (cpu_going_down >= 0)
3763 : cpumask_clear_cpu(cpu_going_down, cpumask);
3764 :
3765 0 : if (cpumask_empty(cpumask))
3766 : goto use_dfl;
3767 :
3768 : /* yeap, return possible CPUs in @node that @attrs wants */
3769 0 : cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3770 0 : return !cpumask_equal(cpumask, attrs->cpumask);
3771 :
3772 : use_dfl:
3773 : cpumask_copy(cpumask, attrs->cpumask);
3774 13 : return false;
3775 : }
3776 :
3777 : /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3778 : static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3779 : int node,
3780 : struct pool_workqueue *pwq)
3781 : {
3782 : struct pool_workqueue *old_pwq;
3783 :
3784 : lockdep_assert_held(&wq->mutex);
3785 :
3786 : /* link_pwq() can handle duplicate calls */
3787 13 : link_pwq(pwq);
3788 :
3789 13 : old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3790 13 : rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3791 : return old_pwq;
3792 : }
3793 :
3794 : /**
3795 : * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3796 : * @wq: the target workqueue
3797 : * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3798 : *
3799 : * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3800 : * machines, this function maps a separate pwq to each NUMA node with
3801 : * possibles CPUs in @attrs->cpumask so that work items are affine to the
3802 : * NUMA node it was issued on. Older pwqs are released as in-flight work
3803 : * items finish. Note that a work item which repeatedly requeues itself
3804 : * back-to-back will stay on its current pwq.
3805 : *
3806 : * Performs GFP_KERNEL allocations.
3807 : *
3808 : * Return: 0 on success and -errno on failure.
3809 : */
3810 13 : int apply_workqueue_attrs(struct workqueue_struct *wq,
3811 : const struct workqueue_attrs *attrs)
3812 : {
3813 : struct workqueue_attrs *new_attrs, *tmp_attrs;
3814 : struct pool_workqueue **pwq_tbl, *dfl_pwq;
3815 : int node, ret;
3816 :
3817 : /* only unbound workqueues can change attributes */
3818 13 : if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3819 : return -EINVAL;
3820 :
3821 : /* creating multiple pwqs breaks ordering guarantee */
3822 21 : if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs)))
3823 : return -EINVAL;
3824 :
3825 : pwq_tbl = kzalloc(nr_node_ids * sizeof(pwq_tbl[0]), GFP_KERNEL);
3826 13 : new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3827 13 : tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3828 13 : if (!pwq_tbl || !new_attrs || !tmp_attrs)
3829 : goto enomem;
3830 :
3831 : /* make a copy of @attrs and sanitize it */
3832 : copy_workqueue_attrs(new_attrs, attrs);
3833 13 : cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3834 :
3835 : /*
3836 : * We may create multiple pwqs with differing cpumasks. Make a
3837 : * copy of @new_attrs which will be modified and used to obtain
3838 : * pools.
3839 : */
3840 : copy_workqueue_attrs(tmp_attrs, new_attrs);
3841 :
3842 : /*
3843 : * CPUs should stay stable across pwq creations and installations.
3844 : * Pin CPUs, determine the target cpumask for each node and create
3845 : * pwqs accordingly.
3846 : */
3847 : get_online_cpus();
3848 :
3849 13 : mutex_lock(&wq_pool_mutex);
3850 :
3851 : /*
3852 : * If something goes wrong during CPU up/down, we'll fall back to
3853 : * the default pwq covering whole @attrs->cpumask. Always create
3854 : * it even if we don't use it immediately.
3855 : */
3856 13 : dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3857 13 : if (!dfl_pwq)
3858 : goto enomem_pwq;
3859 :
3860 26 : for_each_node(node) {
3861 13 : if (wq_calc_node_cpumask(attrs, node, -1, tmp_attrs->cpumask)) {
3862 0 : pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3863 0 : if (!pwq_tbl[node])
3864 : goto enomem_pwq;
3865 : } else {
3866 13 : dfl_pwq->refcnt++;
3867 13 : pwq_tbl[node] = dfl_pwq;
3868 : }
3869 : }
3870 :
3871 13 : mutex_unlock(&wq_pool_mutex);
3872 :
3873 : /* all pwqs have been created successfully, let's install'em */
3874 13 : mutex_lock(&wq->mutex);
3875 :
3876 13 : copy_workqueue_attrs(wq->unbound_attrs, new_attrs);
3877 :
3878 : /* save the previous pwq and install the new one */
3879 39 : for_each_node(node)
3880 26 : pwq_tbl[node] = numa_pwq_tbl_install(wq, node, pwq_tbl[node]);
3881 :
3882 : /* @dfl_pwq might not have been used, ensure it's linked */
3883 13 : link_pwq(dfl_pwq);
3884 13 : swap(wq->dfl_pwq, dfl_pwq);
3885 :
3886 13 : mutex_unlock(&wq->mutex);
3887 :
3888 : /* put the old pwqs */
3889 39 : for_each_node(node)
3890 13 : put_pwq_unlocked(pwq_tbl[node]);
3891 13 : put_pwq_unlocked(dfl_pwq);
3892 :
3893 : put_online_cpus();
3894 : ret = 0;
3895 : /* fall through */
3896 : out_free:
3897 : free_workqueue_attrs(tmp_attrs);
3898 : free_workqueue_attrs(new_attrs);
3899 13 : kfree(pwq_tbl);
3900 13 : return ret;
3901 :
3902 : enomem_pwq:
3903 0 : free_unbound_pwq(dfl_pwq);
3904 0 : for_each_node(node)
3905 0 : if (pwq_tbl && pwq_tbl[node] != dfl_pwq)
3906 0 : free_unbound_pwq(pwq_tbl[node]);
3907 0 : mutex_unlock(&wq_pool_mutex);
3908 : put_online_cpus();
3909 : enomem:
3910 : ret = -ENOMEM;
3911 : goto out_free;
3912 : }
3913 :
3914 : /**
3915 : * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
3916 : * @wq: the target workqueue
3917 : * @cpu: the CPU coming up or going down
3918 : * @online: whether @cpu is coming up or going down
3919 : *
3920 : * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
3921 : * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
3922 : * @wq accordingly.
3923 : *
3924 : * If NUMA affinity can't be adjusted due to memory allocation failure, it
3925 : * falls back to @wq->dfl_pwq which may not be optimal but is always
3926 : * correct.
3927 : *
3928 : * Note that when the last allowed CPU of a NUMA node goes offline for a
3929 : * workqueue with a cpumask spanning multiple nodes, the workers which were
3930 : * already executing the work items for the workqueue will lose their CPU
3931 : * affinity and may execute on any CPU. This is similar to how per-cpu
3932 : * workqueues behave on CPU_DOWN. If a workqueue user wants strict
3933 : * affinity, it's the user's responsibility to flush the work item from
3934 : * CPU_DOWN_PREPARE.
3935 : */
3936 : static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
3937 : bool online)
3938 : {
3939 : int node = cpu_to_node(cpu);
3940 : int cpu_off = online ? -1 : cpu;
3941 : struct pool_workqueue *old_pwq = NULL, *pwq;
3942 : struct workqueue_attrs *target_attrs;
3943 : cpumask_t *cpumask;
3944 :
3945 : lockdep_assert_held(&wq_pool_mutex);
3946 :
3947 : if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND))
3948 : return;
3949 :
3950 : /*
3951 : * We don't wanna alloc/free wq_attrs for each wq for each CPU.
3952 : * Let's use a preallocated one. The following buf is protected by
3953 : * CPU hotplug exclusion.
3954 : */
3955 : target_attrs = wq_update_unbound_numa_attrs_buf;
3956 : cpumask = target_attrs->cpumask;
3957 :
3958 : mutex_lock(&wq->mutex);
3959 : if (wq->unbound_attrs->no_numa)
3960 : goto out_unlock;
3961 :
3962 : copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
3963 : pwq = unbound_pwq_by_node(wq, node);
3964 :
3965 : /*
3966 : * Let's determine what needs to be done. If the target cpumask is
3967 : * different from wq's, we need to compare it to @pwq's and create
3968 : * a new one if they don't match. If the target cpumask equals
3969 : * wq's, the default pwq should be used.
3970 : */
3971 : if (wq_calc_node_cpumask(wq->unbound_attrs, node, cpu_off, cpumask)) {
3972 : if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
3973 : goto out_unlock;
3974 : } else {
3975 : goto use_dfl_pwq;
3976 : }
3977 :
3978 : mutex_unlock(&wq->mutex);
3979 :
3980 : /* create a new pwq */
3981 : pwq = alloc_unbound_pwq(wq, target_attrs);
3982 : if (!pwq) {
3983 : pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
3984 : wq->name);
3985 : mutex_lock(&wq->mutex);
3986 : goto use_dfl_pwq;
3987 : }
3988 :
3989 : /*
3990 : * Install the new pwq. As this function is called only from CPU
3991 : * hotplug callbacks and applying a new attrs is wrapped with
3992 : * get/put_online_cpus(), @wq->unbound_attrs couldn't have changed
3993 : * inbetween.
3994 : */
3995 : mutex_lock(&wq->mutex);
3996 : old_pwq = numa_pwq_tbl_install(wq, node, pwq);
3997 : goto out_unlock;
3998 :
3999 : use_dfl_pwq:
4000 : spin_lock_irq(&wq->dfl_pwq->pool->lock);
4001 : get_pwq(wq->dfl_pwq);
4002 : spin_unlock_irq(&wq->dfl_pwq->pool->lock);
4003 : old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
4004 : out_unlock:
4005 : mutex_unlock(&wq->mutex);
4006 : put_pwq_unlocked(old_pwq);
4007 : }
4008 :
4009 28 : static int alloc_and_link_pwqs(struct workqueue_struct *wq)
4010 : {
4011 28 : bool highpri = wq->flags & WQ_HIGHPRI;
4012 : int cpu, ret;
4013 :
4014 28 : if (!(wq->flags & WQ_UNBOUND)) {
4015 15 : wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
4016 15 : if (!wq->cpu_pwqs)
4017 : return -ENOMEM;
4018 :
4019 15 : for_each_possible_cpu(cpu) {
4020 : struct pool_workqueue *pwq =
4021 15 : per_cpu_ptr(wq->cpu_pwqs, cpu);
4022 : struct worker_pool *cpu_pools =
4023 : per_cpu(cpu_worker_pools, cpu);
4024 :
4025 15 : init_pwq(pwq, wq, &cpu_pools[highpri]);
4026 :
4027 15 : mutex_lock(&wq->mutex);
4028 15 : link_pwq(pwq);
4029 15 : mutex_unlock(&wq->mutex);
4030 : }
4031 : return 0;
4032 13 : } else if (wq->flags & __WQ_ORDERED) {
4033 8 : ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
4034 : /* there should only be single pwq for ordering guarantee */
4035 : WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
4036 : wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
4037 : "ordering guarantee broken for workqueue %s\n", wq->name);
4038 8 : return ret;
4039 : } else {
4040 5 : return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
4041 : }
4042 : }
4043 :
4044 28 : static int wq_clamp_max_active(int max_active, unsigned int flags,
4045 : const char *name)
4046 : {
4047 : int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
4048 :
4049 28 : if (max_active < 1 || max_active > lim)
4050 0 : pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
4051 : max_active, name, 1, lim);
4052 :
4053 28 : return clamp_val(max_active, 1, lim);
4054 : }
4055 :
4056 28 : struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
4057 : unsigned int flags,
4058 : int max_active,
4059 : struct lock_class_key *key,
4060 : const char *lock_name, ...)
4061 : {
4062 : size_t tbl_size = 0;
4063 : va_list args;
4064 : struct workqueue_struct *wq;
4065 : struct pool_workqueue *pwq;
4066 :
4067 : /* see the comment above the definition of WQ_POWER_EFFICIENT */
4068 28 : if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
4069 0 : flags |= WQ_UNBOUND;
4070 :
4071 : /* allocate wq and format name */
4072 28 : if (flags & WQ_UNBOUND)
4073 : tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]);
4074 :
4075 28 : wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
4076 28 : if (!wq)
4077 : return NULL;
4078 :
4079 28 : if (flags & WQ_UNBOUND) {
4080 13 : wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
4081 13 : if (!wq->unbound_attrs)
4082 : goto err_free_wq;
4083 : }
4084 :
4085 28 : va_start(args, lock_name);
4086 28 : vsnprintf(wq->name, sizeof(wq->name), fmt, args);
4087 28 : va_end(args);
4088 :
4089 28 : max_active = max_active ?: WQ_DFL_ACTIVE;
4090 28 : max_active = wq_clamp_max_active(max_active, flags, wq->name);
4091 :
4092 : /* init wq */
4093 28 : wq->flags = flags;
4094 28 : wq->saved_max_active = max_active;
4095 28 : mutex_init(&wq->mutex);
4096 28 : atomic_set(&wq->nr_pwqs_to_flush, 0);
4097 28 : INIT_LIST_HEAD(&wq->pwqs);
4098 28 : INIT_LIST_HEAD(&wq->flusher_queue);
4099 28 : INIT_LIST_HEAD(&wq->flusher_overflow);
4100 28 : INIT_LIST_HEAD(&wq->maydays);
4101 :
4102 : lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
4103 28 : INIT_LIST_HEAD(&wq->list);
4104 :
4105 28 : if (alloc_and_link_pwqs(wq) < 0)
4106 : goto err_free_wq;
4107 :
4108 : /*
4109 : * Workqueues which may be used during memory reclaim should
4110 : * have a rescuer to guarantee forward progress.
4111 : */
4112 28 : if (flags & WQ_MEM_RECLAIM) {
4113 : struct worker *rescuer;
4114 :
4115 15 : rescuer = alloc_worker(NUMA_NO_NODE);
4116 15 : if (!rescuer)
4117 : goto err_destroy;
4118 :
4119 15 : rescuer->rescue_wq = wq;
4120 15 : rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
4121 : wq->name);
4122 15 : if (IS_ERR(rescuer->task)) {
4123 0 : kfree(rescuer);
4124 0 : goto err_destroy;
4125 : }
4126 :
4127 15 : wq->rescuer = rescuer;
4128 15 : rescuer->task->flags |= PF_NO_SETAFFINITY;
4129 15 : wake_up_process(rescuer->task);
4130 : }
4131 :
4132 28 : if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
4133 : goto err_destroy;
4134 :
4135 : /*
4136 : * wq_pool_mutex protects global freeze state and workqueues list.
4137 : * Grab it, adjust max_active and add the new @wq to workqueues
4138 : * list.
4139 : */
4140 28 : mutex_lock(&wq_pool_mutex);
4141 :
4142 28 : mutex_lock(&wq->mutex);
4143 56 : for_each_pwq(pwq, wq)
4144 28 : pwq_adjust_max_active(pwq);
4145 28 : mutex_unlock(&wq->mutex);
4146 :
4147 : list_add(&wq->list, &workqueues);
4148 :
4149 28 : mutex_unlock(&wq_pool_mutex);
4150 :
4151 28 : return wq;
4152 :
4153 : err_free_wq:
4154 0 : free_workqueue_attrs(wq->unbound_attrs);
4155 0 : kfree(wq);
4156 0 : return NULL;
4157 : err_destroy:
4158 0 : destroy_workqueue(wq);
4159 0 : return NULL;
4160 : }
4161 : EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
4162 :
4163 : /**
4164 : * destroy_workqueue - safely terminate a workqueue
4165 : * @wq: target workqueue
4166 : *
4167 : * Safely destroy a workqueue. All work currently pending will be done first.
4168 : */
4169 0 : void destroy_workqueue(struct workqueue_struct *wq)
4170 : {
4171 : struct pool_workqueue *pwq;
4172 : int node;
4173 :
4174 : /* drain it before proceeding with destruction */
4175 0 : drain_workqueue(wq);
4176 :
4177 : /* sanity checks */
4178 0 : mutex_lock(&wq->mutex);
4179 0 : for_each_pwq(pwq, wq) {
4180 : int i;
4181 :
4182 0 : for (i = 0; i < WORK_NR_COLORS; i++) {
4183 0 : if (WARN_ON(pwq->nr_in_flight[i])) {
4184 0 : mutex_unlock(&wq->mutex);
4185 0 : return;
4186 : }
4187 : }
4188 :
4189 0 : if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) ||
4190 0 : WARN_ON(pwq->nr_active) ||
4191 0 : WARN_ON(!list_empty(&pwq->delayed_works))) {
4192 0 : mutex_unlock(&wq->mutex);
4193 0 : return;
4194 : }
4195 : }
4196 0 : mutex_unlock(&wq->mutex);
4197 :
4198 : /*
4199 : * wq list is used to freeze wq, remove from list after
4200 : * flushing is complete in case freeze races us.
4201 : */
4202 0 : mutex_lock(&wq_pool_mutex);
4203 0 : list_del_init(&wq->list);
4204 0 : mutex_unlock(&wq_pool_mutex);
4205 :
4206 : workqueue_sysfs_unregister(wq);
4207 :
4208 0 : if (wq->rescuer) {
4209 0 : kthread_stop(wq->rescuer->task);
4210 0 : kfree(wq->rescuer);
4211 0 : wq->rescuer = NULL;
4212 : }
4213 :
4214 0 : if (!(wq->flags & WQ_UNBOUND)) {
4215 : /*
4216 : * The base ref is never dropped on per-cpu pwqs. Directly
4217 : * free the pwqs and wq.
4218 : */
4219 0 : free_percpu(wq->cpu_pwqs);
4220 0 : kfree(wq);
4221 : } else {
4222 : /*
4223 : * We're the sole accessor of @wq at this point. Directly
4224 : * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4225 : * @wq will be freed when the last pwq is released.
4226 : */
4227 0 : for_each_node(node) {
4228 0 : pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4229 0 : RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4230 0 : put_pwq_unlocked(pwq);
4231 : }
4232 :
4233 : /*
4234 : * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4235 : * put. Don't access it afterwards.
4236 : */
4237 0 : pwq = wq->dfl_pwq;
4238 0 : wq->dfl_pwq = NULL;
4239 0 : put_pwq_unlocked(pwq);
4240 : }
4241 : }
4242 : EXPORT_SYMBOL_GPL(destroy_workqueue);
4243 :
4244 : /**
4245 : * workqueue_set_max_active - adjust max_active of a workqueue
4246 : * @wq: target workqueue
4247 : * @max_active: new max_active value.
4248 : *
4249 : * Set max_active of @wq to @max_active.
4250 : *
4251 : * CONTEXT:
4252 : * Don't call from IRQ context.
4253 : */
4254 0 : void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4255 : {
4256 : struct pool_workqueue *pwq;
4257 :
4258 : /* disallow meddling with max_active for ordered workqueues */
4259 0 : if (WARN_ON(wq->flags & __WQ_ORDERED))
4260 0 : return;
4261 :
4262 0 : max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4263 :
4264 0 : mutex_lock(&wq->mutex);
4265 :
4266 0 : wq->saved_max_active = max_active;
4267 :
4268 0 : for_each_pwq(pwq, wq)
4269 0 : pwq_adjust_max_active(pwq);
4270 :
4271 0 : mutex_unlock(&wq->mutex);
4272 : }
4273 : EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4274 :
4275 : /**
4276 : * current_is_workqueue_rescuer - is %current workqueue rescuer?
4277 : *
4278 : * Determine whether %current is a workqueue rescuer. Can be used from
4279 : * work functions to determine whether it's being run off the rescuer task.
4280 : *
4281 : * Return: %true if %current is a workqueue rescuer. %false otherwise.
4282 : */
4283 238 : bool current_is_workqueue_rescuer(void)
4284 : {
4285 : struct worker *worker = current_wq_worker();
4286 :
4287 238 : return worker && worker->rescue_wq;
4288 : }
4289 :
4290 : /**
4291 : * workqueue_congested - test whether a workqueue is congested
4292 : * @cpu: CPU in question
4293 : * @wq: target workqueue
4294 : *
4295 : * Test whether @wq's cpu workqueue for @cpu is congested. There is
4296 : * no synchronization around this function and the test result is
4297 : * unreliable and only useful as advisory hints or for debugging.
4298 : *
4299 : * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4300 : * Note that both per-cpu and unbound workqueues may be associated with
4301 : * multiple pool_workqueues which have separate congested states. A
4302 : * workqueue being congested on one CPU doesn't mean the workqueue is also
4303 : * contested on other CPUs / NUMA nodes.
4304 : *
4305 : * Return:
4306 : * %true if congested, %false otherwise.
4307 : */
4308 0 : bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4309 : {
4310 : struct pool_workqueue *pwq;
4311 : bool ret;
4312 :
4313 : rcu_read_lock_sched();
4314 :
4315 : if (cpu == WORK_CPU_UNBOUND)
4316 : cpu = smp_processor_id();
4317 :
4318 0 : if (!(wq->flags & WQ_UNBOUND))
4319 0 : pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4320 : else
4321 : pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4322 :
4323 0 : ret = !list_empty(&pwq->delayed_works);
4324 : rcu_read_unlock_sched();
4325 :
4326 0 : return ret;
4327 : }
4328 : EXPORT_SYMBOL_GPL(workqueue_congested);
4329 :
4330 : /**
4331 : * work_busy - test whether a work is currently pending or running
4332 : * @work: the work to be tested
4333 : *
4334 : * Test whether @work is currently pending or running. There is no
4335 : * synchronization around this function and the test result is
4336 : * unreliable and only useful as advisory hints or for debugging.
4337 : *
4338 : * Return:
4339 : * OR'd bitmask of WORK_BUSY_* bits.
4340 : */
4341 0 : unsigned int work_busy(struct work_struct *work)
4342 : {
4343 : struct worker_pool *pool;
4344 : unsigned long flags;
4345 : unsigned int ret = 0;
4346 :
4347 0 : if (work_pending(work))
4348 : ret |= WORK_BUSY_PENDING;
4349 :
4350 : local_irq_save(flags);
4351 0 : pool = get_work_pool(work);
4352 0 : if (pool) {
4353 : spin_lock(&pool->lock);
4354 0 : if (find_worker_executing_work(pool, work))
4355 0 : ret |= WORK_BUSY_RUNNING;
4356 : spin_unlock(&pool->lock);
4357 : }
4358 0 : local_irq_restore(flags);
4359 :
4360 0 : return ret;
4361 : }
4362 : EXPORT_SYMBOL_GPL(work_busy);
4363 :
4364 : /**
4365 : * set_worker_desc - set description for the current work item
4366 : * @fmt: printf-style format string
4367 : * @...: arguments for the format string
4368 : *
4369 : * This function can be called by a running work function to describe what
4370 : * the work item is about. If the worker task gets dumped, this
4371 : * information will be printed out together to help debugging. The
4372 : * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4373 : */
4374 238 : void set_worker_desc(const char *fmt, ...)
4375 : {
4376 : struct worker *worker = current_wq_worker();
4377 : va_list args;
4378 :
4379 238 : if (worker) {
4380 238 : va_start(args, fmt);
4381 238 : vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4382 238 : va_end(args);
4383 238 : worker->desc_valid = true;
4384 : }
4385 238 : }
4386 :
4387 : /**
4388 : * print_worker_info - print out worker information and description
4389 : * @log_lvl: the log level to use when printing
4390 : * @task: target task
4391 : *
4392 : * If @task is a worker and currently executing a work item, print out the
4393 : * name of the workqueue being serviced and worker description set with
4394 : * set_worker_desc() by the currently executing work item.
4395 : *
4396 : * This function can be safely called on any task as long as the
4397 : * task_struct itself is accessible. While safe, this function isn't
4398 : * synchronized and may print out mixups or garbages of limited length.
4399 : */
4400 0 : void print_worker_info(const char *log_lvl, struct task_struct *task)
4401 : {
4402 0 : work_func_t *fn = NULL;
4403 0 : char name[WQ_NAME_LEN] = { };
4404 0 : char desc[WORKER_DESC_LEN] = { };
4405 0 : struct pool_workqueue *pwq = NULL;
4406 0 : struct workqueue_struct *wq = NULL;
4407 0 : bool desc_valid = false;
4408 : struct worker *worker;
4409 :
4410 0 : if (!(task->flags & PF_WQ_WORKER))
4411 0 : return;
4412 :
4413 : /*
4414 : * This function is called without any synchronization and @task
4415 : * could be in any state. Be careful with dereferences.
4416 : */
4417 0 : worker = probe_kthread_data(task);
4418 :
4419 : /*
4420 : * Carefully copy the associated workqueue's workfn and name. Keep
4421 : * the original last '\0' in case the original contains garbage.
4422 : */
4423 0 : probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4424 0 : probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4425 0 : probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4426 0 : probe_kernel_read(name, wq->name, sizeof(name) - 1);
4427 :
4428 : /* copy worker description */
4429 0 : probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid));
4430 0 : if (desc_valid)
4431 0 : probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
4432 :
4433 0 : if (fn || name[0] || desc[0]) {
4434 0 : printk("%sWorkqueue: %s %pf", log_lvl, name, fn);
4435 0 : if (desc[0])
4436 0 : pr_cont(" (%s)", desc);
4437 0 : pr_cont("\n");
4438 : }
4439 : }
4440 :
4441 : /*
4442 : * CPU hotplug.
4443 : *
4444 : * There are two challenges in supporting CPU hotplug. Firstly, there
4445 : * are a lot of assumptions on strong associations among work, pwq and
4446 : * pool which make migrating pending and scheduled works very
4447 : * difficult to implement without impacting hot paths. Secondly,
4448 : * worker pools serve mix of short, long and very long running works making
4449 : * blocked draining impractical.
4450 : *
4451 : * This is solved by allowing the pools to be disassociated from the CPU
4452 : * running as an unbound one and allowing it to be reattached later if the
4453 : * cpu comes back online.
4454 : */
4455 :
4456 : static void wq_unbind_fn(struct work_struct *work)
4457 : {
4458 : int cpu = smp_processor_id();
4459 : struct worker_pool *pool;
4460 : struct worker *worker;
4461 :
4462 : for_each_cpu_worker_pool(pool, cpu) {
4463 : mutex_lock(&pool->attach_mutex);
4464 : spin_lock_irq(&pool->lock);
4465 :
4466 : /*
4467 : * We've blocked all attach/detach operations. Make all workers
4468 : * unbound and set DISASSOCIATED. Before this, all workers
4469 : * except for the ones which are still executing works from
4470 : * before the last CPU down must be on the cpu. After
4471 : * this, they may become diasporas.
4472 : */
4473 : for_each_pool_worker(worker, pool)
4474 : worker->flags |= WORKER_UNBOUND;
4475 :
4476 : pool->flags |= POOL_DISASSOCIATED;
4477 :
4478 : spin_unlock_irq(&pool->lock);
4479 : mutex_unlock(&pool->attach_mutex);
4480 :
4481 : /*
4482 : * Call schedule() so that we cross rq->lock and thus can
4483 : * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4484 : * This is necessary as scheduler callbacks may be invoked
4485 : * from other cpus.
4486 : */
4487 : schedule();
4488 :
4489 : /*
4490 : * Sched callbacks are disabled now. Zap nr_running.
4491 : * After this, nr_running stays zero and need_more_worker()
4492 : * and keep_working() are always true as long as the
4493 : * worklist is not empty. This pool now behaves as an
4494 : * unbound (in terms of concurrency management) pool which
4495 : * are served by workers tied to the pool.
4496 : */
4497 : atomic_set(&pool->nr_running, 0);
4498 :
4499 : /*
4500 : * With concurrency management just turned off, a busy
4501 : * worker blocking could lead to lengthy stalls. Kick off
4502 : * unbound chain execution of currently pending work items.
4503 : */
4504 : spin_lock_irq(&pool->lock);
4505 : wake_up_worker(pool);
4506 : spin_unlock_irq(&pool->lock);
4507 : }
4508 : }
4509 :
4510 : /**
4511 : * rebind_workers - rebind all workers of a pool to the associated CPU
4512 : * @pool: pool of interest
4513 : *
4514 : * @pool->cpu is coming online. Rebind all workers to the CPU.
4515 : */
4516 : static void rebind_workers(struct worker_pool *pool)
4517 : {
4518 : struct worker *worker;
4519 :
4520 : lockdep_assert_held(&pool->attach_mutex);
4521 :
4522 : /*
4523 : * Restore CPU affinity of all workers. As all idle workers should
4524 : * be on the run-queue of the associated CPU before any local
4525 : * wake-ups for concurrency management happen, restore CPU affinty
4526 : * of all workers first and then clear UNBOUND. As we're called
4527 : * from CPU_ONLINE, the following shouldn't fail.
4528 : */
4529 : for_each_pool_worker(worker, pool)
4530 : WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4531 : pool->attrs->cpumask) < 0);
4532 :
4533 : spin_lock_irq(&pool->lock);
4534 : pool->flags &= ~POOL_DISASSOCIATED;
4535 :
4536 : for_each_pool_worker(worker, pool) {
4537 : unsigned int worker_flags = worker->flags;
4538 :
4539 : /*
4540 : * A bound idle worker should actually be on the runqueue
4541 : * of the associated CPU for local wake-ups targeting it to
4542 : * work. Kick all idle workers so that they migrate to the
4543 : * associated CPU. Doing this in the same loop as
4544 : * replacing UNBOUND with REBOUND is safe as no worker will
4545 : * be bound before @pool->lock is released.
4546 : */
4547 : if (worker_flags & WORKER_IDLE)
4548 : wake_up_process(worker->task);
4549 :
4550 : /*
4551 : * We want to clear UNBOUND but can't directly call
4552 : * worker_clr_flags() or adjust nr_running. Atomically
4553 : * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4554 : * @worker will clear REBOUND using worker_clr_flags() when
4555 : * it initiates the next execution cycle thus restoring
4556 : * concurrency management. Note that when or whether
4557 : * @worker clears REBOUND doesn't affect correctness.
4558 : *
4559 : * ACCESS_ONCE() is necessary because @worker->flags may be
4560 : * tested without holding any lock in
4561 : * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4562 : * fail incorrectly leading to premature concurrency
4563 : * management operations.
4564 : */
4565 : WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4566 : worker_flags |= WORKER_REBOUND;
4567 : worker_flags &= ~WORKER_UNBOUND;
4568 : ACCESS_ONCE(worker->flags) = worker_flags;
4569 : }
4570 :
4571 : spin_unlock_irq(&pool->lock);
4572 : }
4573 :
4574 : /**
4575 : * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4576 : * @pool: unbound pool of interest
4577 : * @cpu: the CPU which is coming up
4578 : *
4579 : * An unbound pool may end up with a cpumask which doesn't have any online
4580 : * CPUs. When a worker of such pool get scheduled, the scheduler resets
4581 : * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4582 : * online CPU before, cpus_allowed of all its workers should be restored.
4583 : */
4584 : static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4585 : {
4586 : static cpumask_t cpumask;
4587 : struct worker *worker;
4588 :
4589 : lockdep_assert_held(&pool->attach_mutex);
4590 :
4591 : /* is @cpu allowed for @pool? */
4592 : if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4593 : return;
4594 :
4595 : /* is @cpu the only online CPU? */
4596 : cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
4597 : if (cpumask_weight(&cpumask) != 1)
4598 : return;
4599 :
4600 : /* as we're called from CPU_ONLINE, the following shouldn't fail */
4601 : for_each_pool_worker(worker, pool)
4602 : WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4603 : pool->attrs->cpumask) < 0);
4604 : }
4605 :
4606 : /*
4607 : * Workqueues should be brought up before normal priority CPU notifiers.
4608 : * This will be registered high priority CPU notifier.
4609 : */
4610 : static int workqueue_cpu_up_callback(struct notifier_block *nfb,
4611 : unsigned long action,
4612 : void *hcpu)
4613 : {
4614 : int cpu = (unsigned long)hcpu;
4615 : struct worker_pool *pool;
4616 : struct workqueue_struct *wq;
4617 : int pi;
4618 :
4619 : switch (action & ~CPU_TASKS_FROZEN) {
4620 : case CPU_UP_PREPARE:
4621 : for_each_cpu_worker_pool(pool, cpu) {
4622 : if (pool->nr_workers)
4623 : continue;
4624 : if (!create_worker(pool))
4625 : return NOTIFY_BAD;
4626 : }
4627 : break;
4628 :
4629 : case CPU_DOWN_FAILED:
4630 : case CPU_ONLINE:
4631 : mutex_lock(&wq_pool_mutex);
4632 :
4633 : for_each_pool(pool, pi) {
4634 : mutex_lock(&pool->attach_mutex);
4635 :
4636 : if (pool->cpu == cpu)
4637 : rebind_workers(pool);
4638 : else if (pool->cpu < 0)
4639 : restore_unbound_workers_cpumask(pool, cpu);
4640 :
4641 : mutex_unlock(&pool->attach_mutex);
4642 : }
4643 :
4644 : /* update NUMA affinity of unbound workqueues */
4645 : list_for_each_entry(wq, &workqueues, list)
4646 : wq_update_unbound_numa(wq, cpu, true);
4647 :
4648 : mutex_unlock(&wq_pool_mutex);
4649 : break;
4650 : }
4651 : return NOTIFY_OK;
4652 : }
4653 :
4654 : /*
4655 : * Workqueues should be brought down after normal priority CPU notifiers.
4656 : * This will be registered as low priority CPU notifier.
4657 : */
4658 : static int workqueue_cpu_down_callback(struct notifier_block *nfb,
4659 : unsigned long action,
4660 : void *hcpu)
4661 : {
4662 : int cpu = (unsigned long)hcpu;
4663 : struct work_struct unbind_work;
4664 : struct workqueue_struct *wq;
4665 :
4666 : switch (action & ~CPU_TASKS_FROZEN) {
4667 : case CPU_DOWN_PREPARE:
4668 : /* unbinding per-cpu workers should happen on the local CPU */
4669 : INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
4670 : queue_work_on(cpu, system_highpri_wq, &unbind_work);
4671 :
4672 : /* update NUMA affinity of unbound workqueues */
4673 : mutex_lock(&wq_pool_mutex);
4674 : list_for_each_entry(wq, &workqueues, list)
4675 : wq_update_unbound_numa(wq, cpu, false);
4676 : mutex_unlock(&wq_pool_mutex);
4677 :
4678 : /* wait for per-cpu unbinding to finish */
4679 : flush_work(&unbind_work);
4680 : destroy_work_on_stack(&unbind_work);
4681 : break;
4682 : }
4683 : return NOTIFY_OK;
4684 : }
4685 :
4686 : #ifdef CONFIG_SMP
4687 :
4688 : struct work_for_cpu {
4689 : struct work_struct work;
4690 : long (*fn)(void *);
4691 : void *arg;
4692 : long ret;
4693 : };
4694 :
4695 : static void work_for_cpu_fn(struct work_struct *work)
4696 : {
4697 : struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
4698 :
4699 : wfc->ret = wfc->fn(wfc->arg);
4700 : }
4701 :
4702 : /**
4703 : * work_on_cpu - run a function in user context on a particular cpu
4704 : * @cpu: the cpu to run on
4705 : * @fn: the function to run
4706 : * @arg: the function arg
4707 : *
4708 : * It is up to the caller to ensure that the cpu doesn't go offline.
4709 : * The caller must not hold any locks which would prevent @fn from completing.
4710 : *
4711 : * Return: The value @fn returns.
4712 : */
4713 : long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
4714 : {
4715 : struct work_for_cpu wfc = { .fn = fn, .arg = arg };
4716 :
4717 : INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
4718 : schedule_work_on(cpu, &wfc.work);
4719 : flush_work(&wfc.work);
4720 : destroy_work_on_stack(&wfc.work);
4721 : return wfc.ret;
4722 : }
4723 : EXPORT_SYMBOL_GPL(work_on_cpu);
4724 : #endif /* CONFIG_SMP */
4725 :
4726 : #ifdef CONFIG_FREEZER
4727 :
4728 : /**
4729 : * freeze_workqueues_begin - begin freezing workqueues
4730 : *
4731 : * Start freezing workqueues. After this function returns, all freezable
4732 : * workqueues will queue new works to their delayed_works list instead of
4733 : * pool->worklist.
4734 : *
4735 : * CONTEXT:
4736 : * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4737 : */
4738 0 : void freeze_workqueues_begin(void)
4739 : {
4740 : struct workqueue_struct *wq;
4741 : struct pool_workqueue *pwq;
4742 :
4743 0 : mutex_lock(&wq_pool_mutex);
4744 :
4745 : WARN_ON_ONCE(workqueue_freezing);
4746 0 : workqueue_freezing = true;
4747 :
4748 0 : list_for_each_entry(wq, &workqueues, list) {
4749 0 : mutex_lock(&wq->mutex);
4750 0 : for_each_pwq(pwq, wq)
4751 0 : pwq_adjust_max_active(pwq);
4752 0 : mutex_unlock(&wq->mutex);
4753 : }
4754 :
4755 0 : mutex_unlock(&wq_pool_mutex);
4756 0 : }
4757 :
4758 : /**
4759 : * freeze_workqueues_busy - are freezable workqueues still busy?
4760 : *
4761 : * Check whether freezing is complete. This function must be called
4762 : * between freeze_workqueues_begin() and thaw_workqueues().
4763 : *
4764 : * CONTEXT:
4765 : * Grabs and releases wq_pool_mutex.
4766 : *
4767 : * Return:
4768 : * %true if some freezable workqueues are still busy. %false if freezing
4769 : * is complete.
4770 : */
4771 0 : bool freeze_workqueues_busy(void)
4772 : {
4773 : bool busy = false;
4774 : struct workqueue_struct *wq;
4775 : struct pool_workqueue *pwq;
4776 :
4777 0 : mutex_lock(&wq_pool_mutex);
4778 :
4779 : WARN_ON_ONCE(!workqueue_freezing);
4780 :
4781 0 : list_for_each_entry(wq, &workqueues, list) {
4782 0 : if (!(wq->flags & WQ_FREEZABLE))
4783 0 : continue;
4784 : /*
4785 : * nr_active is monotonically decreasing. It's safe
4786 : * to peek without lock.
4787 : */
4788 : rcu_read_lock_sched();
4789 0 : for_each_pwq(pwq, wq) {
4790 : WARN_ON_ONCE(pwq->nr_active < 0);
4791 0 : if (pwq->nr_active) {
4792 : busy = true;
4793 : rcu_read_unlock_sched();
4794 : goto out_unlock;
4795 : }
4796 : }
4797 : rcu_read_unlock_sched();
4798 : }
4799 : out_unlock:
4800 0 : mutex_unlock(&wq_pool_mutex);
4801 0 : return busy;
4802 : }
4803 :
4804 : /**
4805 : * thaw_workqueues - thaw workqueues
4806 : *
4807 : * Thaw workqueues. Normal queueing is restored and all collected
4808 : * frozen works are transferred to their respective pool worklists.
4809 : *
4810 : * CONTEXT:
4811 : * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4812 : */
4813 0 : void thaw_workqueues(void)
4814 : {
4815 : struct workqueue_struct *wq;
4816 : struct pool_workqueue *pwq;
4817 :
4818 0 : mutex_lock(&wq_pool_mutex);
4819 :
4820 0 : if (!workqueue_freezing)
4821 : goto out_unlock;
4822 :
4823 0 : workqueue_freezing = false;
4824 :
4825 : /* restore max_active and repopulate worklist */
4826 0 : list_for_each_entry(wq, &workqueues, list) {
4827 0 : mutex_lock(&wq->mutex);
4828 0 : for_each_pwq(pwq, wq)
4829 0 : pwq_adjust_max_active(pwq);
4830 0 : mutex_unlock(&wq->mutex);
4831 : }
4832 :
4833 : out_unlock:
4834 0 : mutex_unlock(&wq_pool_mutex);
4835 0 : }
4836 : #endif /* CONFIG_FREEZER */
4837 :
4838 : static void __init wq_numa_init(void)
4839 : {
4840 : cpumask_var_t *tbl;
4841 : int node, cpu;
4842 :
4843 : if (num_possible_nodes() <= 1)
4844 : return;
4845 :
4846 : if (wq_disable_numa) {
4847 : pr_info("workqueue: NUMA affinity support disabled\n");
4848 : return;
4849 : }
4850 :
4851 : wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL);
4852 : BUG_ON(!wq_update_unbound_numa_attrs_buf);
4853 :
4854 : /*
4855 : * We want masks of possible CPUs of each node which isn't readily
4856 : * available. Build one from cpu_to_node() which should have been
4857 : * fully initialized by now.
4858 : */
4859 : tbl = kzalloc(nr_node_ids * sizeof(tbl[0]), GFP_KERNEL);
4860 : BUG_ON(!tbl);
4861 :
4862 : for_each_node(node)
4863 : BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
4864 : node_online(node) ? node : NUMA_NO_NODE));
4865 :
4866 : for_each_possible_cpu(cpu) {
4867 : node = cpu_to_node(cpu);
4868 : if (WARN_ON(node == NUMA_NO_NODE)) {
4869 : pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
4870 : /* happens iff arch is bonkers, let's just proceed */
4871 : return;
4872 : }
4873 : cpumask_set_cpu(cpu, tbl[node]);
4874 : }
4875 :
4876 : wq_numa_possible_cpumask = tbl;
4877 : wq_numa_enabled = true;
4878 : }
4879 :
4880 1 : static int __init init_workqueues(void)
4881 : {
4882 1 : int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
4883 : int i, cpu;
4884 :
4885 : WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
4886 :
4887 1 : pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
4888 :
4889 : cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
4890 : hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
4891 :
4892 : wq_numa_init();
4893 :
4894 : /* initialize CPU pools */
4895 2 : for_each_possible_cpu(cpu) {
4896 : struct worker_pool *pool;
4897 :
4898 : i = 0;
4899 2 : for_each_cpu_worker_pool(pool, cpu) {
4900 2 : BUG_ON(init_worker_pool(pool));
4901 2 : pool->cpu = cpu;
4902 2 : cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
4903 2 : pool->attrs->nice = std_nice[i++];
4904 2 : pool->node = cpu_to_node(cpu);
4905 :
4906 : /* alloc pool ID */
4907 2 : mutex_lock(&wq_pool_mutex);
4908 2 : BUG_ON(worker_pool_assign_id(pool));
4909 2 : mutex_unlock(&wq_pool_mutex);
4910 : }
4911 : }
4912 :
4913 : /* create the initial worker */
4914 1 : for_each_online_cpu(cpu) {
4915 : struct worker_pool *pool;
4916 :
4917 2 : for_each_cpu_worker_pool(pool, cpu) {
4918 2 : pool->flags &= ~POOL_DISASSOCIATED;
4919 2 : BUG_ON(!create_worker(pool));
4920 : }
4921 : }
4922 :
4923 : /* create default unbound and ordered wq attrs */
4924 2 : for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
4925 : struct workqueue_attrs *attrs;
4926 :
4927 2 : BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
4928 2 : attrs->nice = std_nice[i];
4929 2 : unbound_std_wq_attrs[i] = attrs;
4930 :
4931 : /*
4932 : * An ordered wq should have only one pwq as ordering is
4933 : * guaranteed by max_active which is enforced by pwqs.
4934 : * Turn off NUMA so that dfl_pwq is used for all nodes.
4935 : */
4936 2 : BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
4937 2 : attrs->nice = std_nice[i];
4938 2 : attrs->no_numa = true;
4939 2 : ordered_wq_attrs[i] = attrs;
4940 : }
4941 :
4942 1 : system_wq = alloc_workqueue("events", 0, 0);
4943 1 : system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
4944 1 : system_long_wq = alloc_workqueue("events_long", 0, 0);
4945 1 : system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
4946 : WQ_UNBOUND_MAX_ACTIVE);
4947 1 : system_freezable_wq = alloc_workqueue("events_freezable",
4948 : WQ_FREEZABLE, 0);
4949 1 : system_power_efficient_wq = alloc_workqueue("events_power_efficient",
4950 : WQ_POWER_EFFICIENT, 0);
4951 1 : system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
4952 : WQ_FREEZABLE | WQ_POWER_EFFICIENT,
4953 : 0);
4954 : BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
4955 : !system_unbound_wq || !system_freezable_wq ||
4956 : !system_power_efficient_wq ||
4957 : !system_freezable_power_efficient_wq);
4958 1 : return 0;
4959 : }
4960 : early_initcall(init_workqueues);
|
