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