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/interrupt.h>
33 #include <linux/signal.h>
34 #include <linux/completion.h>
35 #include <linux/workqueue.h>
36 #include <linux/slab.h>
37 #include <linux/cpu.h>
38 #include <linux/notifier.h>
39 #include <linux/kthread.h>
40 #include <linux/hardirq.h>
41 #include <linux/mempolicy.h>
42 #include <linux/freezer.h>
43 #include <linux/debug_locks.h>
44 #include <linux/lockdep.h>
45 #include <linux/idr.h>
46 #include <linux/jhash.h>
47 #include <linux/hashtable.h>
48 #include <linux/rculist.h>
49 #include <linux/nodemask.h>
50 #include <linux/moduleparam.h>
51 #include <linux/uaccess.h>
52 #include <linux/sched/isolation.h>
53 #include <linux/sched/debug.h>
54 #include <linux/nmi.h>
55 #include <linux/kvm_para.h>
56 #include <linux/delay.h>
57 #include <linux/irq_work.h>
58
59 #include "workqueue_internal.h"
60
61 enum worker_pool_flags {
62 /*
63 * worker_pool flags
64 *
65 * A bound pool is either associated or disassociated with its CPU.
66 * While associated (!DISASSOCIATED), all workers are bound to the
67 * CPU and none has %WORKER_UNBOUND set and concurrency management
68 * is in effect.
69 *
70 * While DISASSOCIATED, the cpu may be offline and all workers have
71 * %WORKER_UNBOUND set and concurrency management disabled, and may
72 * be executing on any CPU. The pool behaves as an unbound one.
73 *
74 * Note that DISASSOCIATED should be flipped only while holding
75 * wq_pool_attach_mutex to avoid changing binding state while
76 * worker_attach_to_pool() is in progress.
77 *
78 * As there can only be one concurrent BH execution context per CPU, a
79 * BH pool is per-CPU and always DISASSOCIATED.
80 */
81 POOL_BH = 1 << 0, /* is a BH pool */
82 POOL_MANAGER_ACTIVE = 1 << 1, /* being managed */
83 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
84 POOL_BH_DRAINING = 1 << 3, /* draining after CPU offline */
85 };
86
87 enum worker_flags {
88 /* worker flags */
89 WORKER_DIE = 1 << 1, /* die die die */
90 WORKER_IDLE = 1 << 2, /* is idle */
91 WORKER_PREP = 1 << 3, /* preparing to run works */
92 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
93 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
94 WORKER_REBOUND = 1 << 8, /* worker was rebound */
95
96 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
97 WORKER_UNBOUND | WORKER_REBOUND,
98 };
99
100 enum work_cancel_flags {
101 WORK_CANCEL_DELAYED = 1 << 0, /* canceling a delayed_work */
102 WORK_CANCEL_DISABLE = 1 << 1, /* canceling to disable */
103 };
104
105 enum wq_internal_consts {
106 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
107
108 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
109 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
110
111 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
112 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
113
114 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
115 /* call for help after 10ms
116 (min two ticks) */
117 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
118 CREATE_COOLDOWN = HZ, /* time to breath after fail */
119
120 /*
121 * Rescue workers are used only on emergencies and shared by
122 * all cpus. Give MIN_NICE.
123 */
124 RESCUER_NICE_LEVEL = MIN_NICE,
125 HIGHPRI_NICE_LEVEL = MIN_NICE,
126
127 WQ_NAME_LEN = 32,
128 WORKER_ID_LEN = 10 + WQ_NAME_LEN, /* "kworker/R-" + WQ_NAME_LEN */
129 };
130
131 /*
132 * We don't want to trap softirq for too long. See MAX_SOFTIRQ_TIME and
133 * MAX_SOFTIRQ_RESTART in kernel/softirq.c. These are macros because
134 * msecs_to_jiffies() can't be an initializer.
135 */
136 #define BH_WORKER_JIFFIES msecs_to_jiffies(2)
137 #define BH_WORKER_RESTARTS 10
138
139 /*
140 * Structure fields follow one of the following exclusion rules.
141 *
142 * I: Modifiable by initialization/destruction paths and read-only for
143 * everyone else.
144 *
145 * P: Preemption protected. Disabling preemption is enough and should
146 * only be modified and accessed from the local cpu.
147 *
148 * L: pool->lock protected. Access with pool->lock held.
149 *
150 * LN: pool->lock and wq_node_nr_active->lock protected for writes. Either for
151 * reads.
152 *
153 * K: Only modified by worker while holding pool->lock. Can be safely read by
154 * self, while holding pool->lock or from IRQ context if %current is the
155 * kworker.
156 *
157 * S: Only modified by worker self.
158 *
159 * A: wq_pool_attach_mutex protected.
160 *
161 * PL: wq_pool_mutex protected.
162 *
163 * PR: wq_pool_mutex protected for writes. RCU protected for reads.
164 *
165 * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads.
166 *
167 * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or
168 * RCU for reads.
169 *
170 * WQ: wq->mutex protected.
171 *
172 * WR: wq->mutex protected for writes. RCU protected for reads.
173 *
174 * WO: wq->mutex protected for writes. Updated with WRITE_ONCE() and can be read
175 * with READ_ONCE() without locking.
176 *
177 * MD: wq_mayday_lock protected.
178 *
179 * WD: Used internally by the watchdog.
180 */
181
182 /* struct worker is defined in workqueue_internal.h */
183
184 struct worker_pool {
185 raw_spinlock_t lock; /* the pool lock */
186 int cpu; /* I: the associated cpu */
187 int node; /* I: the associated node ID */
188 int id; /* I: pool ID */
189 unsigned int flags; /* L: flags */
190
191 unsigned long watchdog_ts; /* L: watchdog timestamp */
192 bool cpu_stall; /* WD: stalled cpu bound pool */
193
194 /*
195 * The counter is incremented in a process context on the associated CPU
196 * w/ preemption disabled, and decremented or reset in the same context
197 * but w/ pool->lock held. The readers grab pool->lock and are
198 * guaranteed to see if the counter reached zero.
199 */
200 int nr_running;
201
202 struct list_head worklist; /* L: list of pending works */
203
204 int nr_workers; /* L: total number of workers */
205 int nr_idle; /* L: currently idle workers */
206
207 struct list_head idle_list; /* L: list of idle workers */
208 struct timer_list idle_timer; /* L: worker idle timeout */
209 struct work_struct idle_cull_work; /* L: worker idle cleanup */
210
211 struct timer_list mayday_timer; /* L: SOS timer for workers */
212
213 /* a workers is either on busy_hash or idle_list, or the manager */
214 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
215 /* L: hash of busy workers */
216
217 struct worker *manager; /* L: purely informational */
218 struct list_head workers; /* A: attached workers */
219
220 struct ida worker_ida; /* worker IDs for task name */
221
222 struct workqueue_attrs *attrs; /* I: worker attributes */
223 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
224 int refcnt; /* PL: refcnt for unbound pools */
225
226 /*
227 * Destruction of pool is RCU protected to allow dereferences
228 * from get_work_pool().
229 */
230 struct rcu_head rcu;
231 };
232
233 /*
234 * Per-pool_workqueue statistics. These can be monitored using
235 * tools/workqueue/wq_monitor.py.
236 */
237 enum pool_workqueue_stats {
238 PWQ_STAT_STARTED, /* work items started execution */
239 PWQ_STAT_COMPLETED, /* work items completed execution */
240 PWQ_STAT_CPU_TIME, /* total CPU time consumed */
241 PWQ_STAT_CPU_INTENSIVE, /* wq_cpu_intensive_thresh_us violations */
242 PWQ_STAT_CM_WAKEUP, /* concurrency-management worker wakeups */
243 PWQ_STAT_REPATRIATED, /* unbound workers brought back into scope */
244 PWQ_STAT_MAYDAY, /* maydays to rescuer */
245 PWQ_STAT_RESCUED, /* linked work items executed by rescuer */
246
247 PWQ_NR_STATS,
248 };
249
250 /*
251 * The per-pool workqueue. While queued, bits below WORK_PWQ_SHIFT
252 * of work_struct->data are used for flags and the remaining high bits
253 * point to the pwq; thus, pwqs need to be aligned at two's power of the
254 * number of flag bits.
255 */
256 struct pool_workqueue {
257 struct worker_pool *pool; /* I: the associated pool */
258 struct workqueue_struct *wq; /* I: the owning workqueue */
259 int work_color; /* L: current color */
260 int flush_color; /* L: flushing color */
261 int refcnt; /* L: reference count */
262 int nr_in_flight[WORK_NR_COLORS];
263 /* L: nr of in_flight works */
264 bool plugged; /* L: execution suspended */
265
266 /*
267 * nr_active management and WORK_STRUCT_INACTIVE:
268 *
269 * When pwq->nr_active >= max_active, new work item is queued to
270 * pwq->inactive_works instead of pool->worklist and marked with
271 * WORK_STRUCT_INACTIVE.
272 *
273 * All work items marked with WORK_STRUCT_INACTIVE do not participate in
274 * nr_active and all work items in pwq->inactive_works are marked with
275 * WORK_STRUCT_INACTIVE. But not all WORK_STRUCT_INACTIVE work items are
276 * in pwq->inactive_works. Some of them are ready to run in
277 * pool->worklist or worker->scheduled. Those work itmes are only struct
278 * wq_barrier which is used for flush_work() and should not participate
279 * in nr_active. For non-barrier work item, it is marked with
280 * WORK_STRUCT_INACTIVE iff it is in pwq->inactive_works.
281 */
282 int nr_active; /* L: nr of active works */
283 struct list_head inactive_works; /* L: inactive works */
284 struct list_head pending_node; /* LN: node on wq_node_nr_active->pending_pwqs */
285 struct list_head pwqs_node; /* WR: node on wq->pwqs */
286 struct list_head mayday_node; /* MD: node on wq->maydays */
287
288 u64 stats[PWQ_NR_STATS];
289
290 /*
291 * Release of unbound pwq is punted to a kthread_worker. See put_pwq()
292 * and pwq_release_workfn() for details. pool_workqueue itself is also
293 * RCU protected so that the first pwq can be determined without
294 * grabbing wq->mutex.
295 */
296 struct kthread_work release_work;
297 struct rcu_head rcu;
298 } __aligned(1 << WORK_STRUCT_PWQ_SHIFT);
299
300 /*
301 * Structure used to wait for workqueue flush.
302 */
303 struct wq_flusher {
304 struct list_head list; /* WQ: list of flushers */
305 int flush_color; /* WQ: flush color waiting for */
306 struct completion done; /* flush completion */
307 };
308
309 struct wq_device;
310
311 /*
312 * Unlike in a per-cpu workqueue where max_active limits its concurrency level
313 * on each CPU, in an unbound workqueue, max_active applies to the whole system.
314 * As sharing a single nr_active across multiple sockets can be very expensive,
315 * the counting and enforcement is per NUMA node.
316 *
317 * The following struct is used to enforce per-node max_active. When a pwq wants
318 * to start executing a work item, it should increment ->nr using
319 * tryinc_node_nr_active(). If acquisition fails due to ->nr already being over
320 * ->max, the pwq is queued on ->pending_pwqs. As in-flight work items finish
321 * and decrement ->nr, node_activate_pending_pwq() activates the pending pwqs in
322 * round-robin order.
323 */
324 struct wq_node_nr_active {
325 int max; /* per-node max_active */
326 atomic_t nr; /* per-node nr_active */
327 raw_spinlock_t lock; /* nests inside pool locks */
328 struct list_head pending_pwqs; /* LN: pwqs with inactive works */
329 };
330
331 /*
332 * The externally visible workqueue. It relays the issued work items to
333 * the appropriate worker_pool through its pool_workqueues.
334 */
335 struct workqueue_struct {
336 struct list_head pwqs; /* WR: all pwqs of this wq */
337 struct list_head list; /* PR: list of all workqueues */
338
339 struct mutex mutex; /* protects this wq */
340 int work_color; /* WQ: current work color */
341 int flush_color; /* WQ: current flush color */
342 atomic_t nr_pwqs_to_flush; /* flush in progress */
343 struct wq_flusher *first_flusher; /* WQ: first flusher */
344 struct list_head flusher_queue; /* WQ: flush waiters */
345 struct list_head flusher_overflow; /* WQ: flush overflow list */
346
347 struct list_head maydays; /* MD: pwqs requesting rescue */
348 struct worker *rescuer; /* MD: rescue worker */
349
350 int nr_drainers; /* WQ: drain in progress */
351
352 /* See alloc_workqueue() function comment for info on min/max_active */
353 int max_active; /* WO: max active works */
354 int min_active; /* WO: min active works */
355 int saved_max_active; /* WQ: saved max_active */
356 int saved_min_active; /* WQ: saved min_active */
357
358 struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */
359 struct pool_workqueue __rcu *dfl_pwq; /* PW: only for unbound wqs */
360
361 #ifdef CONFIG_SYSFS
362 struct wq_device *wq_dev; /* I: for sysfs interface */
363 #endif
364 #ifdef CONFIG_LOCKDEP
365 char *lock_name;
366 struct lock_class_key key;
367 struct lockdep_map lockdep_map;
368 #endif
369 char name[WQ_NAME_LEN]; /* I: workqueue name */
370
371 /*
372 * Destruction of workqueue_struct is RCU protected to allow walking
373 * the workqueues list without grabbing wq_pool_mutex.
374 * This is used to dump all workqueues from sysrq.
375 */
376 struct rcu_head rcu;
377
378 /* hot fields used during command issue, aligned to cacheline */
379 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
380 struct pool_workqueue __percpu __rcu **cpu_pwq; /* I: per-cpu pwqs */
381 struct wq_node_nr_active *node_nr_active[]; /* I: per-node nr_active */
382 };
383
384 /*
385 * Each pod type describes how CPUs should be grouped for unbound workqueues.
386 * See the comment above workqueue_attrs->affn_scope.
387 */
388 struct wq_pod_type {
389 int nr_pods; /* number of pods */
390 cpumask_var_t *pod_cpus; /* pod -> cpus */
391 int *pod_node; /* pod -> node */
392 int *cpu_pod; /* cpu -> pod */
393 };
394
395 struct work_offq_data {
396 u32 pool_id;
397 u32 disable;
398 u32 flags;
399 };
400
401 static const char *wq_affn_names[WQ_AFFN_NR_TYPES] = {
402 [WQ_AFFN_DFL] = "default",
403 [WQ_AFFN_CPU] = "cpu",
404 [WQ_AFFN_SMT] = "smt",
405 [WQ_AFFN_CACHE] = "cache",
406 [WQ_AFFN_NUMA] = "numa",
407 [WQ_AFFN_SYSTEM] = "system",
408 };
409
410 /*
411 * Per-cpu work items which run for longer than the following threshold are
412 * automatically considered CPU intensive and excluded from concurrency
413 * management to prevent them from noticeably delaying other per-cpu work items.
414 * ULONG_MAX indicates that the user hasn't overridden it with a boot parameter.
415 * The actual value is initialized in wq_cpu_intensive_thresh_init().
416 */
417 static unsigned long wq_cpu_intensive_thresh_us = ULONG_MAX;
418 module_param_named(cpu_intensive_thresh_us, wq_cpu_intensive_thresh_us, ulong, 0644);
419 #ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT
420 static unsigned int wq_cpu_intensive_warning_thresh = 4;
421 module_param_named(cpu_intensive_warning_thresh, wq_cpu_intensive_warning_thresh, uint, 0644);
422 #endif
423
424 /* see the comment above the definition of WQ_POWER_EFFICIENT */
425 static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
426 module_param_named(power_efficient, wq_power_efficient, bool, 0444);
427
428 static bool wq_online; /* can kworkers be created yet? */
429 static bool wq_topo_initialized __read_mostly = false;
430
431 static struct kmem_cache *pwq_cache;
432
433 static struct wq_pod_type wq_pod_types[WQ_AFFN_NR_TYPES];
434 static enum wq_affn_scope wq_affn_dfl = WQ_AFFN_CACHE;
435
436 /* buf for wq_update_unbound_pod_attrs(), protected by CPU hotplug exclusion */
437 static struct workqueue_attrs *unbound_wq_update_pwq_attrs_buf;
438
439 static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
440 static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */
441 static DEFINE_RAW_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
442 /* wait for manager to go away */
443 static struct rcuwait manager_wait = __RCUWAIT_INITIALIZER(manager_wait);
444
445 static LIST_HEAD(workqueues); /* PR: list of all workqueues */
446 static bool workqueue_freezing; /* PL: have wqs started freezing? */
447
448 /* PL: mirror the cpu_online_mask excluding the CPU in the midst of hotplugging */
449 static cpumask_var_t wq_online_cpumask;
450
451 /* PL&A: allowable cpus for unbound wqs and work items */
452 static cpumask_var_t wq_unbound_cpumask;
453
454 /* PL: user requested unbound cpumask via sysfs */
455 static cpumask_var_t wq_requested_unbound_cpumask;
456
457 /* PL: isolated cpumask to be excluded from unbound cpumask */
458 static cpumask_var_t wq_isolated_cpumask;
459
460 /* for further constrain wq_unbound_cpumask by cmdline parameter*/
461 static struct cpumask wq_cmdline_cpumask __initdata;
462
463 /* CPU where unbound work was last round robin scheduled from this CPU */
464 static DEFINE_PER_CPU(int, wq_rr_cpu_last);
465
466 /*
467 * Local execution of unbound work items is no longer guaranteed. The
468 * following always forces round-robin CPU selection on unbound work items
469 * to uncover usages which depend on it.
470 */
471 #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
472 static bool wq_debug_force_rr_cpu = true;
473 #else
474 static bool wq_debug_force_rr_cpu = false;
475 #endif
476 module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
477
478 /* to raise softirq for the BH worker pools on other CPUs */
479 static DEFINE_PER_CPU_SHARED_ALIGNED(struct irq_work [NR_STD_WORKER_POOLS],
480 bh_pool_irq_works);
481
482 /* the BH worker pools */
483 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS],
484 bh_worker_pools);
485
486 /* the per-cpu worker pools */
487 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS],
488 cpu_worker_pools);
489
490 static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
491
492 /* PL: hash of all unbound pools keyed by pool->attrs */
493 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
494
495 /* I: attributes used when instantiating standard unbound pools on demand */
496 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
497
498 /* I: attributes used when instantiating ordered pools on demand */
499 static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
500
501 /*
502 * I: kthread_worker to release pwq's. pwq release needs to be bounced to a
503 * process context while holding a pool lock. Bounce to a dedicated kthread
504 * worker to avoid A-A deadlocks.
505 */
506 static struct kthread_worker *pwq_release_worker __ro_after_init;
507
508 struct workqueue_struct *system_wq __ro_after_init;
509 EXPORT_SYMBOL(system_wq);
510 struct workqueue_struct *system_highpri_wq __ro_after_init;
511 EXPORT_SYMBOL_GPL(system_highpri_wq);
512 struct workqueue_struct *system_long_wq __ro_after_init;
513 EXPORT_SYMBOL_GPL(system_long_wq);
514 struct workqueue_struct *system_unbound_wq __ro_after_init;
515 EXPORT_SYMBOL_GPL(system_unbound_wq);
516 struct workqueue_struct *system_freezable_wq __ro_after_init;
517 EXPORT_SYMBOL_GPL(system_freezable_wq);
518 struct workqueue_struct *system_power_efficient_wq __ro_after_init;
519 EXPORT_SYMBOL_GPL(system_power_efficient_wq);
520 struct workqueue_struct *system_freezable_power_efficient_wq __ro_after_init;
521 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
522 struct workqueue_struct *system_bh_wq;
523 EXPORT_SYMBOL_GPL(system_bh_wq);
524 struct workqueue_struct *system_bh_highpri_wq;
525 EXPORT_SYMBOL_GPL(system_bh_highpri_wq);
526
527 static int worker_thread(void *__worker);
528 static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
529 static void show_pwq(struct pool_workqueue *pwq);
530 static void show_one_worker_pool(struct worker_pool *pool);
531
532 #define CREATE_TRACE_POINTS
533 #include <trace/events/workqueue.h>
534
535 #define assert_rcu_or_pool_mutex() \
536 RCU_LOCKDEP_WARN(!rcu_read_lock_any_held() && \
537 !lockdep_is_held(&wq_pool_mutex), \
538 "RCU or wq_pool_mutex should be held")
539
540 #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
541 RCU_LOCKDEP_WARN(!rcu_read_lock_any_held() && \
542 !lockdep_is_held(&wq->mutex) && \
543 !lockdep_is_held(&wq_pool_mutex), \
544 "RCU, wq->mutex or wq_pool_mutex should be held")
545
546 #define for_each_bh_worker_pool(pool, cpu) \
547 for ((pool) = &per_cpu(bh_worker_pools, cpu)[0]; \
548 (pool) < &per_cpu(bh_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
549 (pool)++)
550
551 #define for_each_cpu_worker_pool(pool, cpu) \
552 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
553 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
554 (pool)++)
555
556 /**
557 * for_each_pool - iterate through all worker_pools in the system
558 * @pool: iteration cursor
559 * @pi: integer used for iteration
560 *
561 * This must be called either with wq_pool_mutex held or RCU read
562 * locked. If the pool needs to be used beyond the locking in effect, the
563 * caller is responsible for guaranteeing that the pool stays online.
564 *
565 * The if/else clause exists only for the lockdep assertion and can be
566 * ignored.
567 */
568 #define for_each_pool(pool, pi) \
569 idr_for_each_entry(&worker_pool_idr, pool, pi) \
570 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
571 else
572
573 /**
574 * for_each_pool_worker - iterate through all workers of a worker_pool
575 * @worker: iteration cursor
576 * @pool: worker_pool to iterate workers of
577 *
578 * This must be called with wq_pool_attach_mutex.
579 *
580 * The if/else clause exists only for the lockdep assertion and can be
581 * ignored.
582 */
583 #define for_each_pool_worker(worker, pool) \
584 list_for_each_entry((worker), &(pool)->workers, node) \
585 if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \
586 else
587
588 /**
589 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
590 * @pwq: iteration cursor
591 * @wq: the target workqueue
592 *
593 * This must be called either with wq->mutex held or RCU read locked.
594 * If the pwq needs to be used beyond the locking in effect, the caller is
595 * responsible for guaranteeing that the pwq stays online.
596 *
597 * The if/else clause exists only for the lockdep assertion and can be
598 * ignored.
599 */
600 #define for_each_pwq(pwq, wq) \
601 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \
602 lockdep_is_held(&(wq->mutex)))
603
604 #ifdef CONFIG_DEBUG_OBJECTS_WORK
605
606 static const struct debug_obj_descr work_debug_descr;
607
work_debug_hint(void * addr)608 static void *work_debug_hint(void *addr)
609 {
610 return ((struct work_struct *) addr)->func;
611 }
612
work_is_static_object(void * addr)613 static bool work_is_static_object(void *addr)
614 {
615 struct work_struct *work = addr;
616
617 return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work));
618 }
619
620 /*
621 * fixup_init is called when:
622 * - an active object is initialized
623 */
work_fixup_init(void * addr,enum debug_obj_state state)624 static bool work_fixup_init(void *addr, enum debug_obj_state state)
625 {
626 struct work_struct *work = addr;
627
628 switch (state) {
629 case ODEBUG_STATE_ACTIVE:
630 cancel_work_sync(work);
631 debug_object_init(work, &work_debug_descr);
632 return true;
633 default:
634 return false;
635 }
636 }
637
638 /*
639 * fixup_free is called when:
640 * - an active object is freed
641 */
work_fixup_free(void * addr,enum debug_obj_state state)642 static bool work_fixup_free(void *addr, enum debug_obj_state state)
643 {
644 struct work_struct *work = addr;
645
646 switch (state) {
647 case ODEBUG_STATE_ACTIVE:
648 cancel_work_sync(work);
649 debug_object_free(work, &work_debug_descr);
650 return true;
651 default:
652 return false;
653 }
654 }
655
656 static const struct debug_obj_descr work_debug_descr = {
657 .name = "work_struct",
658 .debug_hint = work_debug_hint,
659 .is_static_object = work_is_static_object,
660 .fixup_init = work_fixup_init,
661 .fixup_free = work_fixup_free,
662 };
663
debug_work_activate(struct work_struct * work)664 static inline void debug_work_activate(struct work_struct *work)
665 {
666 debug_object_activate(work, &work_debug_descr);
667 }
668
debug_work_deactivate(struct work_struct * work)669 static inline void debug_work_deactivate(struct work_struct *work)
670 {
671 debug_object_deactivate(work, &work_debug_descr);
672 }
673
__init_work(struct work_struct * work,int onstack)674 void __init_work(struct work_struct *work, int onstack)
675 {
676 if (onstack)
677 debug_object_init_on_stack(work, &work_debug_descr);
678 else
679 debug_object_init(work, &work_debug_descr);
680 }
681 EXPORT_SYMBOL_GPL(__init_work);
682
destroy_work_on_stack(struct work_struct * work)683 void destroy_work_on_stack(struct work_struct *work)
684 {
685 debug_object_free(work, &work_debug_descr);
686 }
687 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
688
destroy_delayed_work_on_stack(struct delayed_work * work)689 void destroy_delayed_work_on_stack(struct delayed_work *work)
690 {
691 destroy_timer_on_stack(&work->timer);
692 debug_object_free(&work->work, &work_debug_descr);
693 }
694 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
695
696 #else
debug_work_activate(struct work_struct * work)697 static inline void debug_work_activate(struct work_struct *work) { }
debug_work_deactivate(struct work_struct * work)698 static inline void debug_work_deactivate(struct work_struct *work) { }
699 #endif
700
701 /**
702 * worker_pool_assign_id - allocate ID and assign it to @pool
703 * @pool: the pool pointer of interest
704 *
705 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
706 * successfully, -errno on failure.
707 */
worker_pool_assign_id(struct worker_pool * pool)708 static int worker_pool_assign_id(struct worker_pool *pool)
709 {
710 int ret;
711
712 lockdep_assert_held(&wq_pool_mutex);
713
714 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
715 GFP_KERNEL);
716 if (ret >= 0) {
717 pool->id = ret;
718 return 0;
719 }
720 return ret;
721 }
722
723 static struct pool_workqueue __rcu **
unbound_pwq_slot(struct workqueue_struct * wq,int cpu)724 unbound_pwq_slot(struct workqueue_struct *wq, int cpu)
725 {
726 if (cpu >= 0)
727 return per_cpu_ptr(wq->cpu_pwq, cpu);
728 else
729 return &wq->dfl_pwq;
730 }
731
732 /* @cpu < 0 for dfl_pwq */
unbound_pwq(struct workqueue_struct * wq,int cpu)733 static struct pool_workqueue *unbound_pwq(struct workqueue_struct *wq, int cpu)
734 {
735 return rcu_dereference_check(*unbound_pwq_slot(wq, cpu),
736 lockdep_is_held(&wq_pool_mutex) ||
737 lockdep_is_held(&wq->mutex));
738 }
739
740 /**
741 * unbound_effective_cpumask - effective cpumask of an unbound workqueue
742 * @wq: workqueue of interest
743 *
744 * @wq->unbound_attrs->cpumask contains the cpumask requested by the user which
745 * is masked with wq_unbound_cpumask to determine the effective cpumask. The
746 * default pwq is always mapped to the pool with the current effective cpumask.
747 */
unbound_effective_cpumask(struct workqueue_struct * wq)748 static struct cpumask *unbound_effective_cpumask(struct workqueue_struct *wq)
749 {
750 return unbound_pwq(wq, -1)->pool->attrs->__pod_cpumask;
751 }
752
work_color_to_flags(int color)753 static unsigned int work_color_to_flags(int color)
754 {
755 return color << WORK_STRUCT_COLOR_SHIFT;
756 }
757
get_work_color(unsigned long work_data)758 static int get_work_color(unsigned long work_data)
759 {
760 return (work_data >> WORK_STRUCT_COLOR_SHIFT) &
761 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
762 }
763
work_next_color(int color)764 static int work_next_color(int color)
765 {
766 return (color + 1) % WORK_NR_COLORS;
767 }
768
pool_offq_flags(struct worker_pool * pool)769 static unsigned long pool_offq_flags(struct worker_pool *pool)
770 {
771 return (pool->flags & POOL_BH) ? WORK_OFFQ_BH : 0;
772 }
773
774 /*
775 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
776 * contain the pointer to the queued pwq. Once execution starts, the flag
777 * is cleared and the high bits contain OFFQ flags and pool ID.
778 *
779 * set_work_pwq(), set_work_pool_and_clear_pending() and mark_work_canceling()
780 * can be used to set the pwq, pool or clear work->data. These functions should
781 * only be called while the work is owned - ie. while the PENDING bit is set.
782 *
783 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
784 * corresponding to a work. Pool is available once the work has been
785 * queued anywhere after initialization until it is sync canceled. pwq is
786 * available only while the work item is queued.
787 */
set_work_data(struct work_struct * work,unsigned long data)788 static inline void set_work_data(struct work_struct *work, unsigned long data)
789 {
790 WARN_ON_ONCE(!work_pending(work));
791 atomic_long_set(&work->data, data | work_static(work));
792 }
793
set_work_pwq(struct work_struct * work,struct pool_workqueue * pwq,unsigned long flags)794 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
795 unsigned long flags)
796 {
797 set_work_data(work, (unsigned long)pwq | WORK_STRUCT_PENDING |
798 WORK_STRUCT_PWQ | flags);
799 }
800
set_work_pool_and_keep_pending(struct work_struct * work,int pool_id,unsigned long flags)801 static void set_work_pool_and_keep_pending(struct work_struct *work,
802 int pool_id, unsigned long flags)
803 {
804 set_work_data(work, ((unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT) |
805 WORK_STRUCT_PENDING | flags);
806 }
807
set_work_pool_and_clear_pending(struct work_struct * work,int pool_id,unsigned long flags)808 static void set_work_pool_and_clear_pending(struct work_struct *work,
809 int pool_id, unsigned long flags)
810 {
811 /*
812 * The following wmb is paired with the implied mb in
813 * test_and_set_bit(PENDING) and ensures all updates to @work made
814 * here are visible to and precede any updates by the next PENDING
815 * owner.
816 */
817 smp_wmb();
818 set_work_data(work, ((unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT) |
819 flags);
820 /*
821 * The following mb guarantees that previous clear of a PENDING bit
822 * will not be reordered with any speculative LOADS or STORES from
823 * work->current_func, which is executed afterwards. This possible
824 * reordering can lead to a missed execution on attempt to queue
825 * the same @work. E.g. consider this case:
826 *
827 * CPU#0 CPU#1
828 * ---------------------------- --------------------------------
829 *
830 * 1 STORE event_indicated
831 * 2 queue_work_on() {
832 * 3 test_and_set_bit(PENDING)
833 * 4 } set_..._and_clear_pending() {
834 * 5 set_work_data() # clear bit
835 * 6 smp_mb()
836 * 7 work->current_func() {
837 * 8 LOAD event_indicated
838 * }
839 *
840 * Without an explicit full barrier speculative LOAD on line 8 can
841 * be executed before CPU#0 does STORE on line 1. If that happens,
842 * CPU#0 observes the PENDING bit is still set and new execution of
843 * a @work is not queued in a hope, that CPU#1 will eventually
844 * finish the queued @work. Meanwhile CPU#1 does not see
845 * event_indicated is set, because speculative LOAD was executed
846 * before actual STORE.
847 */
848 smp_mb();
849 }
850
work_struct_pwq(unsigned long data)851 static inline struct pool_workqueue *work_struct_pwq(unsigned long data)
852 {
853 return (struct pool_workqueue *)(data & WORK_STRUCT_PWQ_MASK);
854 }
855
get_work_pwq(struct work_struct * work)856 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
857 {
858 unsigned long data = atomic_long_read(&work->data);
859
860 if (data & WORK_STRUCT_PWQ)
861 return work_struct_pwq(data);
862 else
863 return NULL;
864 }
865
866 /**
867 * get_work_pool - return the worker_pool a given work was associated with
868 * @work: the work item of interest
869 *
870 * Pools are created and destroyed under wq_pool_mutex, and allows read
871 * access under RCU read lock. As such, this function should be
872 * called under wq_pool_mutex or inside of a rcu_read_lock() region.
873 *
874 * All fields of the returned pool are accessible as long as the above
875 * mentioned locking is in effect. If the returned pool needs to be used
876 * beyond the critical section, the caller is responsible for ensuring the
877 * returned pool is and stays online.
878 *
879 * Return: The worker_pool @work was last associated with. %NULL if none.
880 */
get_work_pool(struct work_struct * work)881 static struct worker_pool *get_work_pool(struct work_struct *work)
882 {
883 unsigned long data = atomic_long_read(&work->data);
884 int pool_id;
885
886 assert_rcu_or_pool_mutex();
887
888 if (data & WORK_STRUCT_PWQ)
889 return work_struct_pwq(data)->pool;
890
891 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
892 if (pool_id == WORK_OFFQ_POOL_NONE)
893 return NULL;
894
895 return idr_find(&worker_pool_idr, pool_id);
896 }
897
shift_and_mask(unsigned long v,u32 shift,u32 bits)898 static unsigned long shift_and_mask(unsigned long v, u32 shift, u32 bits)
899 {
900 return (v >> shift) & ((1 << bits) - 1);
901 }
902
work_offqd_unpack(struct work_offq_data * offqd,unsigned long data)903 static void work_offqd_unpack(struct work_offq_data *offqd, unsigned long data)
904 {
905 WARN_ON_ONCE(data & WORK_STRUCT_PWQ);
906
907 offqd->pool_id = shift_and_mask(data, WORK_OFFQ_POOL_SHIFT,
908 WORK_OFFQ_POOL_BITS);
909 offqd->disable = shift_and_mask(data, WORK_OFFQ_DISABLE_SHIFT,
910 WORK_OFFQ_DISABLE_BITS);
911 offqd->flags = data & WORK_OFFQ_FLAG_MASK;
912 }
913
work_offqd_pack_flags(struct work_offq_data * offqd)914 static unsigned long work_offqd_pack_flags(struct work_offq_data *offqd)
915 {
916 return ((unsigned long)offqd->disable << WORK_OFFQ_DISABLE_SHIFT) |
917 ((unsigned long)offqd->flags);
918 }
919
920 /*
921 * Policy functions. These define the policies on how the global worker
922 * pools are managed. Unless noted otherwise, these functions assume that
923 * they're being called with pool->lock held.
924 */
925
926 /*
927 * Need to wake up a worker? Called from anything but currently
928 * running workers.
929 *
930 * Note that, because unbound workers never contribute to nr_running, this
931 * function will always return %true for unbound pools as long as the
932 * worklist isn't empty.
933 */
need_more_worker(struct worker_pool * pool)934 static bool need_more_worker(struct worker_pool *pool)
935 {
936 return !list_empty(&pool->worklist) && !pool->nr_running;
937 }
938
939 /* Can I start working? Called from busy but !running workers. */
may_start_working(struct worker_pool * pool)940 static bool may_start_working(struct worker_pool *pool)
941 {
942 return pool->nr_idle;
943 }
944
945 /* Do I need to keep working? Called from currently running workers. */
keep_working(struct worker_pool * pool)946 static bool keep_working(struct worker_pool *pool)
947 {
948 return !list_empty(&pool->worklist) && (pool->nr_running <= 1);
949 }
950
951 /* Do we need a new worker? Called from manager. */
need_to_create_worker(struct worker_pool * pool)952 static bool need_to_create_worker(struct worker_pool *pool)
953 {
954 return need_more_worker(pool) && !may_start_working(pool);
955 }
956
957 /* Do we have too many workers and should some go away? */
too_many_workers(struct worker_pool * pool)958 static bool too_many_workers(struct worker_pool *pool)
959 {
960 bool managing = pool->flags & POOL_MANAGER_ACTIVE;
961 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
962 int nr_busy = pool->nr_workers - nr_idle;
963
964 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
965 }
966
967 /**
968 * worker_set_flags - set worker flags and adjust nr_running accordingly
969 * @worker: self
970 * @flags: flags to set
971 *
972 * Set @flags in @worker->flags and adjust nr_running accordingly.
973 */
worker_set_flags(struct worker * worker,unsigned int flags)974 static inline void worker_set_flags(struct worker *worker, unsigned int flags)
975 {
976 struct worker_pool *pool = worker->pool;
977
978 lockdep_assert_held(&pool->lock);
979
980 /* If transitioning into NOT_RUNNING, adjust nr_running. */
981 if ((flags & WORKER_NOT_RUNNING) &&
982 !(worker->flags & WORKER_NOT_RUNNING)) {
983 pool->nr_running--;
984 }
985
986 worker->flags |= flags;
987 }
988
989 /**
990 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
991 * @worker: self
992 * @flags: flags to clear
993 *
994 * Clear @flags in @worker->flags and adjust nr_running accordingly.
995 */
worker_clr_flags(struct worker * worker,unsigned int flags)996 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
997 {
998 struct worker_pool *pool = worker->pool;
999 unsigned int oflags = worker->flags;
1000
1001 lockdep_assert_held(&pool->lock);
1002
1003 worker->flags &= ~flags;
1004
1005 /*
1006 * If transitioning out of NOT_RUNNING, increment nr_running. Note
1007 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
1008 * of multiple flags, not a single flag.
1009 */
1010 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
1011 if (!(worker->flags & WORKER_NOT_RUNNING))
1012 pool->nr_running++;
1013 }
1014
1015 /* Return the first idle worker. Called with pool->lock held. */
first_idle_worker(struct worker_pool * pool)1016 static struct worker *first_idle_worker(struct worker_pool *pool)
1017 {
1018 if (unlikely(list_empty(&pool->idle_list)))
1019 return NULL;
1020
1021 return list_first_entry(&pool->idle_list, struct worker, entry);
1022 }
1023
1024 /**
1025 * worker_enter_idle - enter idle state
1026 * @worker: worker which is entering idle state
1027 *
1028 * @worker is entering idle state. Update stats and idle timer if
1029 * necessary.
1030 *
1031 * LOCKING:
1032 * raw_spin_lock_irq(pool->lock).
1033 */
worker_enter_idle(struct worker * worker)1034 static void worker_enter_idle(struct worker *worker)
1035 {
1036 struct worker_pool *pool = worker->pool;
1037
1038 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1039 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1040 (worker->hentry.next || worker->hentry.pprev)))
1041 return;
1042
1043 /* can't use worker_set_flags(), also called from create_worker() */
1044 worker->flags |= WORKER_IDLE;
1045 pool->nr_idle++;
1046 worker->last_active = jiffies;
1047
1048 /* idle_list is LIFO */
1049 list_add(&worker->entry, &pool->idle_list);
1050
1051 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1052 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1053
1054 /* Sanity check nr_running. */
1055 WARN_ON_ONCE(pool->nr_workers == pool->nr_idle && pool->nr_running);
1056 }
1057
1058 /**
1059 * worker_leave_idle - leave idle state
1060 * @worker: worker which is leaving idle state
1061 *
1062 * @worker is leaving idle state. Update stats.
1063 *
1064 * LOCKING:
1065 * raw_spin_lock_irq(pool->lock).
1066 */
worker_leave_idle(struct worker * worker)1067 static void worker_leave_idle(struct worker *worker)
1068 {
1069 struct worker_pool *pool = worker->pool;
1070
1071 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1072 return;
1073 worker_clr_flags(worker, WORKER_IDLE);
1074 pool->nr_idle--;
1075 list_del_init(&worker->entry);
1076 }
1077
1078 /**
1079 * find_worker_executing_work - find worker which is executing a work
1080 * @pool: pool of interest
1081 * @work: work to find worker for
1082 *
1083 * Find a worker which is executing @work on @pool by searching
1084 * @pool->busy_hash which is keyed by the address of @work. For a worker
1085 * to match, its current execution should match the address of @work and
1086 * its work function. This is to avoid unwanted dependency between
1087 * unrelated work executions through a work item being recycled while still
1088 * being executed.
1089 *
1090 * This is a bit tricky. A work item may be freed once its execution
1091 * starts and nothing prevents the freed area from being recycled for
1092 * another work item. If the same work item address ends up being reused
1093 * before the original execution finishes, workqueue will identify the
1094 * recycled work item as currently executing and make it wait until the
1095 * current execution finishes, introducing an unwanted dependency.
1096 *
1097 * This function checks the work item address and work function to avoid
1098 * false positives. Note that this isn't complete as one may construct a
1099 * work function which can introduce dependency onto itself through a
1100 * recycled work item. Well, if somebody wants to shoot oneself in the
1101 * foot that badly, there's only so much we can do, and if such deadlock
1102 * actually occurs, it should be easy to locate the culprit work function.
1103 *
1104 * CONTEXT:
1105 * raw_spin_lock_irq(pool->lock).
1106 *
1107 * Return:
1108 * Pointer to worker which is executing @work if found, %NULL
1109 * otherwise.
1110 */
find_worker_executing_work(struct worker_pool * pool,struct work_struct * work)1111 static struct worker *find_worker_executing_work(struct worker_pool *pool,
1112 struct work_struct *work)
1113 {
1114 struct worker *worker;
1115
1116 hash_for_each_possible(pool->busy_hash, worker, hentry,
1117 (unsigned long)work)
1118 if (worker->current_work == work &&
1119 worker->current_func == work->func)
1120 return worker;
1121
1122 return NULL;
1123 }
1124
1125 /**
1126 * move_linked_works - move linked works to a list
1127 * @work: start of series of works to be scheduled
1128 * @head: target list to append @work to
1129 * @nextp: out parameter for nested worklist walking
1130 *
1131 * Schedule linked works starting from @work to @head. Work series to be
1132 * scheduled starts at @work and includes any consecutive work with
1133 * WORK_STRUCT_LINKED set in its predecessor. See assign_work() for details on
1134 * @nextp.
1135 *
1136 * CONTEXT:
1137 * raw_spin_lock_irq(pool->lock).
1138 */
move_linked_works(struct work_struct * work,struct list_head * head,struct work_struct ** nextp)1139 static void move_linked_works(struct work_struct *work, struct list_head *head,
1140 struct work_struct **nextp)
1141 {
1142 struct work_struct *n;
1143
1144 /*
1145 * Linked worklist will always end before the end of the list,
1146 * use NULL for list head.
1147 */
1148 list_for_each_entry_safe_from(work, n, NULL, entry) {
1149 list_move_tail(&work->entry, head);
1150 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1151 break;
1152 }
1153
1154 /*
1155 * If we're already inside safe list traversal and have moved
1156 * multiple works to the scheduled queue, the next position
1157 * needs to be updated.
1158 */
1159 if (nextp)
1160 *nextp = n;
1161 }
1162
1163 /**
1164 * assign_work - assign a work item and its linked work items to a worker
1165 * @work: work to assign
1166 * @worker: worker to assign to
1167 * @nextp: out parameter for nested worklist walking
1168 *
1169 * Assign @work and its linked work items to @worker. If @work is already being
1170 * executed by another worker in the same pool, it'll be punted there.
1171 *
1172 * If @nextp is not NULL, it's updated to point to the next work of the last
1173 * scheduled work. This allows assign_work() to be nested inside
1174 * list_for_each_entry_safe().
1175 *
1176 * Returns %true if @work was successfully assigned to @worker. %false if @work
1177 * was punted to another worker already executing it.
1178 */
assign_work(struct work_struct * work,struct worker * worker,struct work_struct ** nextp)1179 static bool assign_work(struct work_struct *work, struct worker *worker,
1180 struct work_struct **nextp)
1181 {
1182 struct worker_pool *pool = worker->pool;
1183 struct worker *collision;
1184
1185 lockdep_assert_held(&pool->lock);
1186
1187 /*
1188 * A single work shouldn't be executed concurrently by multiple workers.
1189 * __queue_work() ensures that @work doesn't jump to a different pool
1190 * while still running in the previous pool. Here, we should ensure that
1191 * @work is not executed concurrently by multiple workers from the same
1192 * pool. Check whether anyone is already processing the work. If so,
1193 * defer the work to the currently executing one.
1194 */
1195 collision = find_worker_executing_work(pool, work);
1196 if (unlikely(collision)) {
1197 move_linked_works(work, &collision->scheduled, nextp);
1198 return false;
1199 }
1200
1201 move_linked_works(work, &worker->scheduled, nextp);
1202 return true;
1203 }
1204
bh_pool_irq_work(struct worker_pool * pool)1205 static struct irq_work *bh_pool_irq_work(struct worker_pool *pool)
1206 {
1207 int high = pool->attrs->nice == HIGHPRI_NICE_LEVEL ? 1 : 0;
1208
1209 return &per_cpu(bh_pool_irq_works, pool->cpu)[high];
1210 }
1211
kick_bh_pool(struct worker_pool * pool)1212 static void kick_bh_pool(struct worker_pool *pool)
1213 {
1214 #ifdef CONFIG_SMP
1215 /* see drain_dead_softirq_workfn() for BH_DRAINING */
1216 if (unlikely(pool->cpu != smp_processor_id() &&
1217 !(pool->flags & POOL_BH_DRAINING))) {
1218 irq_work_queue_on(bh_pool_irq_work(pool), pool->cpu);
1219 return;
1220 }
1221 #endif
1222 if (pool->attrs->nice == HIGHPRI_NICE_LEVEL)
1223 raise_softirq_irqoff(HI_SOFTIRQ);
1224 else
1225 raise_softirq_irqoff(TASKLET_SOFTIRQ);
1226 }
1227
1228 /**
1229 * kick_pool - wake up an idle worker if necessary
1230 * @pool: pool to kick
1231 *
1232 * @pool may have pending work items. Wake up worker if necessary. Returns
1233 * whether a worker was woken up.
1234 */
kick_pool(struct worker_pool * pool)1235 static bool kick_pool(struct worker_pool *pool)
1236 {
1237 struct worker *worker = first_idle_worker(pool);
1238 struct task_struct *p;
1239
1240 lockdep_assert_held(&pool->lock);
1241
1242 if (!need_more_worker(pool) || !worker)
1243 return false;
1244
1245 if (pool->flags & POOL_BH) {
1246 kick_bh_pool(pool);
1247 return true;
1248 }
1249
1250 p = worker->task;
1251
1252 #ifdef CONFIG_SMP
1253 /*
1254 * Idle @worker is about to execute @work and waking up provides an
1255 * opportunity to migrate @worker at a lower cost by setting the task's
1256 * wake_cpu field. Let's see if we want to move @worker to improve
1257 * execution locality.
1258 *
1259 * We're waking the worker that went idle the latest and there's some
1260 * chance that @worker is marked idle but hasn't gone off CPU yet. If
1261 * so, setting the wake_cpu won't do anything. As this is a best-effort
1262 * optimization and the race window is narrow, let's leave as-is for
1263 * now. If this becomes pronounced, we can skip over workers which are
1264 * still on cpu when picking an idle worker.
1265 *
1266 * If @pool has non-strict affinity, @worker might have ended up outside
1267 * its affinity scope. Repatriate.
1268 */
1269 if (!pool->attrs->affn_strict &&
1270 !cpumask_test_cpu(p->wake_cpu, pool->attrs->__pod_cpumask)) {
1271 struct work_struct *work = list_first_entry(&pool->worklist,
1272 struct work_struct, entry);
1273 int wake_cpu = cpumask_any_and_distribute(pool->attrs->__pod_cpumask,
1274 cpu_online_mask);
1275 if (wake_cpu < nr_cpu_ids) {
1276 p->wake_cpu = wake_cpu;
1277 get_work_pwq(work)->stats[PWQ_STAT_REPATRIATED]++;
1278 }
1279 }
1280 #endif
1281 wake_up_process(p);
1282 return true;
1283 }
1284
1285 #ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT
1286
1287 /*
1288 * Concurrency-managed per-cpu work items that hog CPU for longer than
1289 * wq_cpu_intensive_thresh_us trigger the automatic CPU_INTENSIVE mechanism,
1290 * which prevents them from stalling other concurrency-managed work items. If a
1291 * work function keeps triggering this mechanism, it's likely that the work item
1292 * should be using an unbound workqueue instead.
1293 *
1294 * wq_cpu_intensive_report() tracks work functions which trigger such conditions
1295 * and report them so that they can be examined and converted to use unbound
1296 * workqueues as appropriate. To avoid flooding the console, each violating work
1297 * function is tracked and reported with exponential backoff.
1298 */
1299 #define WCI_MAX_ENTS 128
1300
1301 struct wci_ent {
1302 work_func_t func;
1303 atomic64_t cnt;
1304 struct hlist_node hash_node;
1305 };
1306
1307 static struct wci_ent wci_ents[WCI_MAX_ENTS];
1308 static int wci_nr_ents;
1309 static DEFINE_RAW_SPINLOCK(wci_lock);
1310 static DEFINE_HASHTABLE(wci_hash, ilog2(WCI_MAX_ENTS));
1311
wci_find_ent(work_func_t func)1312 static struct wci_ent *wci_find_ent(work_func_t func)
1313 {
1314 struct wci_ent *ent;
1315
1316 hash_for_each_possible_rcu(wci_hash, ent, hash_node,
1317 (unsigned long)func) {
1318 if (ent->func == func)
1319 return ent;
1320 }
1321 return NULL;
1322 }
1323
wq_cpu_intensive_report(work_func_t func)1324 static void wq_cpu_intensive_report(work_func_t func)
1325 {
1326 struct wci_ent *ent;
1327
1328 restart:
1329 ent = wci_find_ent(func);
1330 if (ent) {
1331 u64 cnt;
1332
1333 /*
1334 * Start reporting from the warning_thresh and back off
1335 * exponentially.
1336 */
1337 cnt = atomic64_inc_return_relaxed(&ent->cnt);
1338 if (wq_cpu_intensive_warning_thresh &&
1339 cnt >= wq_cpu_intensive_warning_thresh &&
1340 is_power_of_2(cnt + 1 - wq_cpu_intensive_warning_thresh))
1341 printk_deferred(KERN_WARNING "workqueue: %ps hogged CPU for >%luus %llu times, consider switching to WQ_UNBOUND\n",
1342 ent->func, wq_cpu_intensive_thresh_us,
1343 atomic64_read(&ent->cnt));
1344 return;
1345 }
1346
1347 /*
1348 * @func is a new violation. Allocate a new entry for it. If wcn_ents[]
1349 * is exhausted, something went really wrong and we probably made enough
1350 * noise already.
1351 */
1352 if (wci_nr_ents >= WCI_MAX_ENTS)
1353 return;
1354
1355 raw_spin_lock(&wci_lock);
1356
1357 if (wci_nr_ents >= WCI_MAX_ENTS) {
1358 raw_spin_unlock(&wci_lock);
1359 return;
1360 }
1361
1362 if (wci_find_ent(func)) {
1363 raw_spin_unlock(&wci_lock);
1364 goto restart;
1365 }
1366
1367 ent = &wci_ents[wci_nr_ents++];
1368 ent->func = func;
1369 atomic64_set(&ent->cnt, 0);
1370 hash_add_rcu(wci_hash, &ent->hash_node, (unsigned long)func);
1371
1372 raw_spin_unlock(&wci_lock);
1373
1374 goto restart;
1375 }
1376
1377 #else /* CONFIG_WQ_CPU_INTENSIVE_REPORT */
wq_cpu_intensive_report(work_func_t func)1378 static void wq_cpu_intensive_report(work_func_t func) {}
1379 #endif /* CONFIG_WQ_CPU_INTENSIVE_REPORT */
1380
1381 /**
1382 * wq_worker_running - a worker is running again
1383 * @task: task waking up
1384 *
1385 * This function is called when a worker returns from schedule()
1386 */
wq_worker_running(struct task_struct * task)1387 void wq_worker_running(struct task_struct *task)
1388 {
1389 struct worker *worker = kthread_data(task);
1390
1391 if (!READ_ONCE(worker->sleeping))
1392 return;
1393
1394 /*
1395 * If preempted by unbind_workers() between the WORKER_NOT_RUNNING check
1396 * and the nr_running increment below, we may ruin the nr_running reset
1397 * and leave with an unexpected pool->nr_running == 1 on the newly unbound
1398 * pool. Protect against such race.
1399 */
1400 preempt_disable();
1401 if (!(worker->flags & WORKER_NOT_RUNNING))
1402 worker->pool->nr_running++;
1403 preempt_enable();
1404
1405 /*
1406 * CPU intensive auto-detection cares about how long a work item hogged
1407 * CPU without sleeping. Reset the starting timestamp on wakeup.
1408 */
1409 worker->current_at = worker->task->se.sum_exec_runtime;
1410
1411 WRITE_ONCE(worker->sleeping, 0);
1412 }
1413
1414 /**
1415 * wq_worker_sleeping - a worker is going to sleep
1416 * @task: task going to sleep
1417 *
1418 * This function is called from schedule() when a busy worker is
1419 * going to sleep.
1420 */
wq_worker_sleeping(struct task_struct * task)1421 void wq_worker_sleeping(struct task_struct *task)
1422 {
1423 struct worker *worker = kthread_data(task);
1424 struct worker_pool *pool;
1425
1426 /*
1427 * Rescuers, which may not have all the fields set up like normal
1428 * workers, also reach here, let's not access anything before
1429 * checking NOT_RUNNING.
1430 */
1431 if (worker->flags & WORKER_NOT_RUNNING)
1432 return;
1433
1434 pool = worker->pool;
1435
1436 /* Return if preempted before wq_worker_running() was reached */
1437 if (READ_ONCE(worker->sleeping))
1438 return;
1439
1440 WRITE_ONCE(worker->sleeping, 1);
1441 raw_spin_lock_irq(&pool->lock);
1442
1443 /*
1444 * Recheck in case unbind_workers() preempted us. We don't
1445 * want to decrement nr_running after the worker is unbound
1446 * and nr_running has been reset.
1447 */
1448 if (worker->flags & WORKER_NOT_RUNNING) {
1449 raw_spin_unlock_irq(&pool->lock);
1450 return;
1451 }
1452
1453 pool->nr_running--;
1454 if (kick_pool(pool))
1455 worker->current_pwq->stats[PWQ_STAT_CM_WAKEUP]++;
1456
1457 raw_spin_unlock_irq(&pool->lock);
1458 }
1459
1460 /**
1461 * wq_worker_tick - a scheduler tick occurred while a kworker is running
1462 * @task: task currently running
1463 *
1464 * Called from sched_tick(). We're in the IRQ context and the current
1465 * worker's fields which follow the 'K' locking rule can be accessed safely.
1466 */
wq_worker_tick(struct task_struct * task)1467 void wq_worker_tick(struct task_struct *task)
1468 {
1469 struct worker *worker = kthread_data(task);
1470 struct pool_workqueue *pwq = worker->current_pwq;
1471 struct worker_pool *pool = worker->pool;
1472
1473 if (!pwq)
1474 return;
1475
1476 pwq->stats[PWQ_STAT_CPU_TIME] += TICK_USEC;
1477
1478 if (!wq_cpu_intensive_thresh_us)
1479 return;
1480
1481 /*
1482 * If the current worker is concurrency managed and hogged the CPU for
1483 * longer than wq_cpu_intensive_thresh_us, it's automatically marked
1484 * CPU_INTENSIVE to avoid stalling other concurrency-managed work items.
1485 *
1486 * Set @worker->sleeping means that @worker is in the process of
1487 * switching out voluntarily and won't be contributing to
1488 * @pool->nr_running until it wakes up. As wq_worker_sleeping() also
1489 * decrements ->nr_running, setting CPU_INTENSIVE here can lead to
1490 * double decrements. The task is releasing the CPU anyway. Let's skip.
1491 * We probably want to make this prettier in the future.
1492 */
1493 if ((worker->flags & WORKER_NOT_RUNNING) || READ_ONCE(worker->sleeping) ||
1494 worker->task->se.sum_exec_runtime - worker->current_at <
1495 wq_cpu_intensive_thresh_us * NSEC_PER_USEC)
1496 return;
1497
1498 raw_spin_lock(&pool->lock);
1499
1500 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
1501 wq_cpu_intensive_report(worker->current_func);
1502 pwq->stats[PWQ_STAT_CPU_INTENSIVE]++;
1503
1504 if (kick_pool(pool))
1505 pwq->stats[PWQ_STAT_CM_WAKEUP]++;
1506
1507 raw_spin_unlock(&pool->lock);
1508 }
1509
1510 /**
1511 * wq_worker_last_func - retrieve worker's last work function
1512 * @task: Task to retrieve last work function of.
1513 *
1514 * Determine the last function a worker executed. This is called from
1515 * the scheduler to get a worker's last known identity.
1516 *
1517 * CONTEXT:
1518 * raw_spin_lock_irq(rq->lock)
1519 *
1520 * This function is called during schedule() when a kworker is going
1521 * to sleep. It's used by psi to identify aggregation workers during
1522 * dequeuing, to allow periodic aggregation to shut-off when that
1523 * worker is the last task in the system or cgroup to go to sleep.
1524 *
1525 * As this function doesn't involve any workqueue-related locking, it
1526 * only returns stable values when called from inside the scheduler's
1527 * queuing and dequeuing paths, when @task, which must be a kworker,
1528 * is guaranteed to not be processing any works.
1529 *
1530 * Return:
1531 * The last work function %current executed as a worker, NULL if it
1532 * hasn't executed any work yet.
1533 */
wq_worker_last_func(struct task_struct * task)1534 work_func_t wq_worker_last_func(struct task_struct *task)
1535 {
1536 struct worker *worker = kthread_data(task);
1537
1538 return worker->last_func;
1539 }
1540
1541 /**
1542 * wq_node_nr_active - Determine wq_node_nr_active to use
1543 * @wq: workqueue of interest
1544 * @node: NUMA node, can be %NUMA_NO_NODE
1545 *
1546 * Determine wq_node_nr_active to use for @wq on @node. Returns:
1547 *
1548 * - %NULL for per-cpu workqueues as they don't need to use shared nr_active.
1549 *
1550 * - node_nr_active[nr_node_ids] if @node is %NUMA_NO_NODE.
1551 *
1552 * - Otherwise, node_nr_active[@node].
1553 */
wq_node_nr_active(struct workqueue_struct * wq,int node)1554 static struct wq_node_nr_active *wq_node_nr_active(struct workqueue_struct *wq,
1555 int node)
1556 {
1557 if (!(wq->flags & WQ_UNBOUND))
1558 return NULL;
1559
1560 if (node == NUMA_NO_NODE)
1561 node = nr_node_ids;
1562
1563 return wq->node_nr_active[node];
1564 }
1565
1566 /**
1567 * wq_update_node_max_active - Update per-node max_actives to use
1568 * @wq: workqueue to update
1569 * @off_cpu: CPU that's going down, -1 if a CPU is not going down
1570 *
1571 * Update @wq->node_nr_active[]->max. @wq must be unbound. max_active is
1572 * distributed among nodes according to the proportions of numbers of online
1573 * cpus. The result is always between @wq->min_active and max_active.
1574 */
wq_update_node_max_active(struct workqueue_struct * wq,int off_cpu)1575 static void wq_update_node_max_active(struct workqueue_struct *wq, int off_cpu)
1576 {
1577 struct cpumask *effective = unbound_effective_cpumask(wq);
1578 int min_active = READ_ONCE(wq->min_active);
1579 int max_active = READ_ONCE(wq->max_active);
1580 int total_cpus, node;
1581
1582 lockdep_assert_held(&wq->mutex);
1583
1584 if (!wq_topo_initialized)
1585 return;
1586
1587 if (off_cpu >= 0 && !cpumask_test_cpu(off_cpu, effective))
1588 off_cpu = -1;
1589
1590 total_cpus = cpumask_weight_and(effective, cpu_online_mask);
1591 if (off_cpu >= 0)
1592 total_cpus--;
1593
1594 /* If all CPUs of the wq get offline, use the default values */
1595 if (unlikely(!total_cpus)) {
1596 for_each_node(node)
1597 wq_node_nr_active(wq, node)->max = min_active;
1598
1599 wq_node_nr_active(wq, NUMA_NO_NODE)->max = max_active;
1600 return;
1601 }
1602
1603 for_each_node(node) {
1604 int node_cpus;
1605
1606 node_cpus = cpumask_weight_and(effective, cpumask_of_node(node));
1607 if (off_cpu >= 0 && cpu_to_node(off_cpu) == node)
1608 node_cpus--;
1609
1610 wq_node_nr_active(wq, node)->max =
1611 clamp(DIV_ROUND_UP(max_active * node_cpus, total_cpus),
1612 min_active, max_active);
1613 }
1614
1615 wq_node_nr_active(wq, NUMA_NO_NODE)->max = max_active;
1616 }
1617
1618 /**
1619 * get_pwq - get an extra reference on the specified pool_workqueue
1620 * @pwq: pool_workqueue to get
1621 *
1622 * Obtain an extra reference on @pwq. The caller should guarantee that
1623 * @pwq has positive refcnt and be holding the matching pool->lock.
1624 */
get_pwq(struct pool_workqueue * pwq)1625 static void get_pwq(struct pool_workqueue *pwq)
1626 {
1627 lockdep_assert_held(&pwq->pool->lock);
1628 WARN_ON_ONCE(pwq->refcnt <= 0);
1629 pwq->refcnt++;
1630 }
1631
1632 /**
1633 * put_pwq - put a pool_workqueue reference
1634 * @pwq: pool_workqueue to put
1635 *
1636 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1637 * destruction. The caller should be holding the matching pool->lock.
1638 */
put_pwq(struct pool_workqueue * pwq)1639 static void put_pwq(struct pool_workqueue *pwq)
1640 {
1641 lockdep_assert_held(&pwq->pool->lock);
1642 if (likely(--pwq->refcnt))
1643 return;
1644 /*
1645 * @pwq can't be released under pool->lock, bounce to a dedicated
1646 * kthread_worker to avoid A-A deadlocks.
1647 */
1648 kthread_queue_work(pwq_release_worker, &pwq->release_work);
1649 }
1650
1651 /**
1652 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1653 * @pwq: pool_workqueue to put (can be %NULL)
1654 *
1655 * put_pwq() with locking. This function also allows %NULL @pwq.
1656 */
put_pwq_unlocked(struct pool_workqueue * pwq)1657 static void put_pwq_unlocked(struct pool_workqueue *pwq)
1658 {
1659 if (pwq) {
1660 /*
1661 * As both pwqs and pools are RCU protected, the
1662 * following lock operations are safe.
1663 */
1664 raw_spin_lock_irq(&pwq->pool->lock);
1665 put_pwq(pwq);
1666 raw_spin_unlock_irq(&pwq->pool->lock);
1667 }
1668 }
1669
pwq_is_empty(struct pool_workqueue * pwq)1670 static bool pwq_is_empty(struct pool_workqueue *pwq)
1671 {
1672 return !pwq->nr_active && list_empty(&pwq->inactive_works);
1673 }
1674
__pwq_activate_work(struct pool_workqueue * pwq,struct work_struct * work)1675 static void __pwq_activate_work(struct pool_workqueue *pwq,
1676 struct work_struct *work)
1677 {
1678 unsigned long *wdb = work_data_bits(work);
1679
1680 WARN_ON_ONCE(!(*wdb & WORK_STRUCT_INACTIVE));
1681 trace_workqueue_activate_work(work);
1682 if (list_empty(&pwq->pool->worklist))
1683 pwq->pool->watchdog_ts = jiffies;
1684 move_linked_works(work, &pwq->pool->worklist, NULL);
1685 __clear_bit(WORK_STRUCT_INACTIVE_BIT, wdb);
1686 }
1687
tryinc_node_nr_active(struct wq_node_nr_active * nna)1688 static bool tryinc_node_nr_active(struct wq_node_nr_active *nna)
1689 {
1690 int max = READ_ONCE(nna->max);
1691
1692 while (true) {
1693 int old, tmp;
1694
1695 old = atomic_read(&nna->nr);
1696 if (old >= max)
1697 return false;
1698 tmp = atomic_cmpxchg_relaxed(&nna->nr, old, old + 1);
1699 if (tmp == old)
1700 return true;
1701 }
1702 }
1703
1704 /**
1705 * pwq_tryinc_nr_active - Try to increment nr_active for a pwq
1706 * @pwq: pool_workqueue of interest
1707 * @fill: max_active may have increased, try to increase concurrency level
1708 *
1709 * Try to increment nr_active for @pwq. Returns %true if an nr_active count is
1710 * successfully obtained. %false otherwise.
1711 */
pwq_tryinc_nr_active(struct pool_workqueue * pwq,bool fill)1712 static bool pwq_tryinc_nr_active(struct pool_workqueue *pwq, bool fill)
1713 {
1714 struct workqueue_struct *wq = pwq->wq;
1715 struct worker_pool *pool = pwq->pool;
1716 struct wq_node_nr_active *nna = wq_node_nr_active(wq, pool->node);
1717 bool obtained = false;
1718
1719 lockdep_assert_held(&pool->lock);
1720
1721 if (!nna) {
1722 /* BH or per-cpu workqueue, pwq->nr_active is sufficient */
1723 obtained = pwq->nr_active < READ_ONCE(wq->max_active);
1724 goto out;
1725 }
1726
1727 if (unlikely(pwq->plugged))
1728 return false;
1729
1730 /*
1731 * Unbound workqueue uses per-node shared nr_active $nna. If @pwq is
1732 * already waiting on $nna, pwq_dec_nr_active() will maintain the
1733 * concurrency level. Don't jump the line.
1734 *
1735 * We need to ignore the pending test after max_active has increased as
1736 * pwq_dec_nr_active() can only maintain the concurrency level but not
1737 * increase it. This is indicated by @fill.
1738 */
1739 if (!list_empty(&pwq->pending_node) && likely(!fill))
1740 goto out;
1741
1742 obtained = tryinc_node_nr_active(nna);
1743 if (obtained)
1744 goto out;
1745
1746 /*
1747 * Lockless acquisition failed. Lock, add ourself to $nna->pending_pwqs
1748 * and try again. The smp_mb() is paired with the implied memory barrier
1749 * of atomic_dec_return() in pwq_dec_nr_active() to ensure that either
1750 * we see the decremented $nna->nr or they see non-empty
1751 * $nna->pending_pwqs.
1752 */
1753 raw_spin_lock(&nna->lock);
1754
1755 if (list_empty(&pwq->pending_node))
1756 list_add_tail(&pwq->pending_node, &nna->pending_pwqs);
1757 else if (likely(!fill))
1758 goto out_unlock;
1759
1760 smp_mb();
1761
1762 obtained = tryinc_node_nr_active(nna);
1763
1764 /*
1765 * If @fill, @pwq might have already been pending. Being spuriously
1766 * pending in cold paths doesn't affect anything. Let's leave it be.
1767 */
1768 if (obtained && likely(!fill))
1769 list_del_init(&pwq->pending_node);
1770
1771 out_unlock:
1772 raw_spin_unlock(&nna->lock);
1773 out:
1774 if (obtained)
1775 pwq->nr_active++;
1776 return obtained;
1777 }
1778
1779 /**
1780 * pwq_activate_first_inactive - Activate the first inactive work item on a pwq
1781 * @pwq: pool_workqueue of interest
1782 * @fill: max_active may have increased, try to increase concurrency level
1783 *
1784 * Activate the first inactive work item of @pwq if available and allowed by
1785 * max_active limit.
1786 *
1787 * Returns %true if an inactive work item has been activated. %false if no
1788 * inactive work item is found or max_active limit is reached.
1789 */
pwq_activate_first_inactive(struct pool_workqueue * pwq,bool fill)1790 static bool pwq_activate_first_inactive(struct pool_workqueue *pwq, bool fill)
1791 {
1792 struct work_struct *work =
1793 list_first_entry_or_null(&pwq->inactive_works,
1794 struct work_struct, entry);
1795
1796 if (work && pwq_tryinc_nr_active(pwq, fill)) {
1797 __pwq_activate_work(pwq, work);
1798 return true;
1799 } else {
1800 return false;
1801 }
1802 }
1803
1804 /**
1805 * unplug_oldest_pwq - unplug the oldest pool_workqueue
1806 * @wq: workqueue_struct where its oldest pwq is to be unplugged
1807 *
1808 * This function should only be called for ordered workqueues where only the
1809 * oldest pwq is unplugged, the others are plugged to suspend execution to
1810 * ensure proper work item ordering::
1811 *
1812 * dfl_pwq --------------+ [P] - plugged
1813 * |
1814 * v
1815 * pwqs -> A -> B [P] -> C [P] (newest)
1816 * | | |
1817 * 1 3 5
1818 * | | |
1819 * 2 4 6
1820 *
1821 * When the oldest pwq is drained and removed, this function should be called
1822 * to unplug the next oldest one to start its work item execution. Note that
1823 * pwq's are linked into wq->pwqs with the oldest first, so the first one in
1824 * the list is the oldest.
1825 */
unplug_oldest_pwq(struct workqueue_struct * wq)1826 static void unplug_oldest_pwq(struct workqueue_struct *wq)
1827 {
1828 struct pool_workqueue *pwq;
1829
1830 lockdep_assert_held(&wq->mutex);
1831
1832 /* Caller should make sure that pwqs isn't empty before calling */
1833 pwq = list_first_entry_or_null(&wq->pwqs, struct pool_workqueue,
1834 pwqs_node);
1835 raw_spin_lock_irq(&pwq->pool->lock);
1836 if (pwq->plugged) {
1837 pwq->plugged = false;
1838 if (pwq_activate_first_inactive(pwq, true))
1839 kick_pool(pwq->pool);
1840 }
1841 raw_spin_unlock_irq(&pwq->pool->lock);
1842 }
1843
1844 /**
1845 * node_activate_pending_pwq - Activate a pending pwq on a wq_node_nr_active
1846 * @nna: wq_node_nr_active to activate a pending pwq for
1847 * @caller_pool: worker_pool the caller is locking
1848 *
1849 * Activate a pwq in @nna->pending_pwqs. Called with @caller_pool locked.
1850 * @caller_pool may be unlocked and relocked to lock other worker_pools.
1851 */
node_activate_pending_pwq(struct wq_node_nr_active * nna,struct worker_pool * caller_pool)1852 static void node_activate_pending_pwq(struct wq_node_nr_active *nna,
1853 struct worker_pool *caller_pool)
1854 {
1855 struct worker_pool *locked_pool = caller_pool;
1856 struct pool_workqueue *pwq;
1857 struct work_struct *work;
1858
1859 lockdep_assert_held(&caller_pool->lock);
1860
1861 raw_spin_lock(&nna->lock);
1862 retry:
1863 pwq = list_first_entry_or_null(&nna->pending_pwqs,
1864 struct pool_workqueue, pending_node);
1865 if (!pwq)
1866 goto out_unlock;
1867
1868 /*
1869 * If @pwq is for a different pool than @locked_pool, we need to lock
1870 * @pwq->pool->lock. Let's trylock first. If unsuccessful, do the unlock
1871 * / lock dance. For that, we also need to release @nna->lock as it's
1872 * nested inside pool locks.
1873 */
1874 if (pwq->pool != locked_pool) {
1875 raw_spin_unlock(&locked_pool->lock);
1876 locked_pool = pwq->pool;
1877 if (!raw_spin_trylock(&locked_pool->lock)) {
1878 raw_spin_unlock(&nna->lock);
1879 raw_spin_lock(&locked_pool->lock);
1880 raw_spin_lock(&nna->lock);
1881 goto retry;
1882 }
1883 }
1884
1885 /*
1886 * $pwq may not have any inactive work items due to e.g. cancellations.
1887 * Drop it from pending_pwqs and see if there's another one.
1888 */
1889 work = list_first_entry_or_null(&pwq->inactive_works,
1890 struct work_struct, entry);
1891 if (!work) {
1892 list_del_init(&pwq->pending_node);
1893 goto retry;
1894 }
1895
1896 /*
1897 * Acquire an nr_active count and activate the inactive work item. If
1898 * $pwq still has inactive work items, rotate it to the end of the
1899 * pending_pwqs so that we round-robin through them. This means that
1900 * inactive work items are not activated in queueing order which is fine
1901 * given that there has never been any ordering across different pwqs.
1902 */
1903 if (likely(tryinc_node_nr_active(nna))) {
1904 pwq->nr_active++;
1905 __pwq_activate_work(pwq, work);
1906
1907 if (list_empty(&pwq->inactive_works))
1908 list_del_init(&pwq->pending_node);
1909 else
1910 list_move_tail(&pwq->pending_node, &nna->pending_pwqs);
1911
1912 /* if activating a foreign pool, make sure it's running */
1913 if (pwq->pool != caller_pool)
1914 kick_pool(pwq->pool);
1915 }
1916
1917 out_unlock:
1918 raw_spin_unlock(&nna->lock);
1919 if (locked_pool != caller_pool) {
1920 raw_spin_unlock(&locked_pool->lock);
1921 raw_spin_lock(&caller_pool->lock);
1922 }
1923 }
1924
1925 /**
1926 * pwq_dec_nr_active - Retire an active count
1927 * @pwq: pool_workqueue of interest
1928 *
1929 * Decrement @pwq's nr_active and try to activate the first inactive work item.
1930 * For unbound workqueues, this function may temporarily drop @pwq->pool->lock.
1931 */
pwq_dec_nr_active(struct pool_workqueue * pwq)1932 static void pwq_dec_nr_active(struct pool_workqueue *pwq)
1933 {
1934 struct worker_pool *pool = pwq->pool;
1935 struct wq_node_nr_active *nna = wq_node_nr_active(pwq->wq, pool->node);
1936
1937 lockdep_assert_held(&pool->lock);
1938
1939 /*
1940 * @pwq->nr_active should be decremented for both percpu and unbound
1941 * workqueues.
1942 */
1943 pwq->nr_active--;
1944
1945 /*
1946 * For a percpu workqueue, it's simple. Just need to kick the first
1947 * inactive work item on @pwq itself.
1948 */
1949 if (!nna) {
1950 pwq_activate_first_inactive(pwq, false);
1951 return;
1952 }
1953
1954 /*
1955 * If @pwq is for an unbound workqueue, it's more complicated because
1956 * multiple pwqs and pools may be sharing the nr_active count. When a
1957 * pwq needs to wait for an nr_active count, it puts itself on
1958 * $nna->pending_pwqs. The following atomic_dec_return()'s implied
1959 * memory barrier is paired with smp_mb() in pwq_tryinc_nr_active() to
1960 * guarantee that either we see non-empty pending_pwqs or they see
1961 * decremented $nna->nr.
1962 *
1963 * $nna->max may change as CPUs come online/offline and @pwq->wq's
1964 * max_active gets updated. However, it is guaranteed to be equal to or
1965 * larger than @pwq->wq->min_active which is above zero unless freezing.
1966 * This maintains the forward progress guarantee.
1967 */
1968 if (atomic_dec_return(&nna->nr) >= READ_ONCE(nna->max))
1969 return;
1970
1971 if (!list_empty(&nna->pending_pwqs))
1972 node_activate_pending_pwq(nna, pool);
1973 }
1974
1975 /**
1976 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1977 * @pwq: pwq of interest
1978 * @work_data: work_data of work which left the queue
1979 *
1980 * A work either has completed or is removed from pending queue,
1981 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1982 *
1983 * NOTE:
1984 * For unbound workqueues, this function may temporarily drop @pwq->pool->lock
1985 * and thus should be called after all other state updates for the in-flight
1986 * work item is complete.
1987 *
1988 * CONTEXT:
1989 * raw_spin_lock_irq(pool->lock).
1990 */
pwq_dec_nr_in_flight(struct pool_workqueue * pwq,unsigned long work_data)1991 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, unsigned long work_data)
1992 {
1993 int color = get_work_color(work_data);
1994
1995 if (!(work_data & WORK_STRUCT_INACTIVE))
1996 pwq_dec_nr_active(pwq);
1997
1998 pwq->nr_in_flight[color]--;
1999
2000 /* is flush in progress and are we at the flushing tip? */
2001 if (likely(pwq->flush_color != color))
2002 goto out_put;
2003
2004 /* are there still in-flight works? */
2005 if (pwq->nr_in_flight[color])
2006 goto out_put;
2007
2008 /* this pwq is done, clear flush_color */
2009 pwq->flush_color = -1;
2010
2011 /*
2012 * If this was the last pwq, wake up the first flusher. It
2013 * will handle the rest.
2014 */
2015 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
2016 complete(&pwq->wq->first_flusher->done);
2017 out_put:
2018 put_pwq(pwq);
2019 }
2020
2021 /**
2022 * try_to_grab_pending - steal work item from worklist and disable irq
2023 * @work: work item to steal
2024 * @cflags: %WORK_CANCEL_ flags
2025 * @irq_flags: place to store irq state
2026 *
2027 * Try to grab PENDING bit of @work. This function can handle @work in any
2028 * stable state - idle, on timer or on worklist.
2029 *
2030 * Return:
2031 *
2032 * ======== ================================================================
2033 * 1 if @work was pending and we successfully stole PENDING
2034 * 0 if @work was idle and we claimed PENDING
2035 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
2036 * ======== ================================================================
2037 *
2038 * Note:
2039 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
2040 * interrupted while holding PENDING and @work off queue, irq must be
2041 * disabled on entry. This, combined with delayed_work->timer being
2042 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
2043 *
2044 * On successful return, >= 0, irq is disabled and the caller is
2045 * responsible for releasing it using local_irq_restore(*@irq_flags).
2046 *
2047 * This function is safe to call from any context including IRQ handler.
2048 */
try_to_grab_pending(struct work_struct * work,u32 cflags,unsigned long * irq_flags)2049 static int try_to_grab_pending(struct work_struct *work, u32 cflags,
2050 unsigned long *irq_flags)
2051 {
2052 struct worker_pool *pool;
2053 struct pool_workqueue *pwq;
2054
2055 local_irq_save(*irq_flags);
2056
2057 /* try to steal the timer if it exists */
2058 if (cflags & WORK_CANCEL_DELAYED) {
2059 struct delayed_work *dwork = to_delayed_work(work);
2060
2061 /*
2062 * dwork->timer is irqsafe. If del_timer() fails, it's
2063 * guaranteed that the timer is not queued anywhere and not
2064 * running on the local CPU.
2065 */
2066 if (likely(del_timer(&dwork->timer)))
2067 return 1;
2068 }
2069
2070 /* try to claim PENDING the normal way */
2071 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
2072 return 0;
2073
2074 rcu_read_lock();
2075 /*
2076 * The queueing is in progress, or it is already queued. Try to
2077 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
2078 */
2079 pool = get_work_pool(work);
2080 if (!pool)
2081 goto fail;
2082
2083 raw_spin_lock(&pool->lock);
2084 /*
2085 * work->data is guaranteed to point to pwq only while the work
2086 * item is queued on pwq->wq, and both updating work->data to point
2087 * to pwq on queueing and to pool on dequeueing are done under
2088 * pwq->pool->lock. This in turn guarantees that, if work->data
2089 * points to pwq which is associated with a locked pool, the work
2090 * item is currently queued on that pool.
2091 */
2092 pwq = get_work_pwq(work);
2093 if (pwq && pwq->pool == pool) {
2094 unsigned long work_data = *work_data_bits(work);
2095
2096 debug_work_deactivate(work);
2097
2098 /*
2099 * A cancelable inactive work item must be in the
2100 * pwq->inactive_works since a queued barrier can't be
2101 * canceled (see the comments in insert_wq_barrier()).
2102 *
2103 * An inactive work item cannot be deleted directly because
2104 * it might have linked barrier work items which, if left
2105 * on the inactive_works list, will confuse pwq->nr_active
2106 * management later on and cause stall. Move the linked
2107 * barrier work items to the worklist when deleting the grabbed
2108 * item. Also keep WORK_STRUCT_INACTIVE in work_data, so that
2109 * it doesn't participate in nr_active management in later
2110 * pwq_dec_nr_in_flight().
2111 */
2112 if (work_data & WORK_STRUCT_INACTIVE)
2113 move_linked_works(work, &pwq->pool->worklist, NULL);
2114
2115 list_del_init(&work->entry);
2116
2117 /*
2118 * work->data points to pwq iff queued. Let's point to pool. As
2119 * this destroys work->data needed by the next step, stash it.
2120 */
2121 set_work_pool_and_keep_pending(work, pool->id,
2122 pool_offq_flags(pool));
2123
2124 /* must be the last step, see the function comment */
2125 pwq_dec_nr_in_flight(pwq, work_data);
2126
2127 raw_spin_unlock(&pool->lock);
2128 rcu_read_unlock();
2129 return 1;
2130 }
2131 raw_spin_unlock(&pool->lock);
2132 fail:
2133 rcu_read_unlock();
2134 local_irq_restore(*irq_flags);
2135 return -EAGAIN;
2136 }
2137
2138 /**
2139 * work_grab_pending - steal work item from worklist and disable irq
2140 * @work: work item to steal
2141 * @cflags: %WORK_CANCEL_ flags
2142 * @irq_flags: place to store IRQ state
2143 *
2144 * Grab PENDING bit of @work. @work can be in any stable state - idle, on timer
2145 * or on worklist.
2146 *
2147 * Can be called from any context. IRQ is disabled on return with IRQ state
2148 * stored in *@irq_flags. The caller is responsible for re-enabling it using
2149 * local_irq_restore().
2150 *
2151 * Returns %true if @work was pending. %false if idle.
2152 */
work_grab_pending(struct work_struct * work,u32 cflags,unsigned long * irq_flags)2153 static bool work_grab_pending(struct work_struct *work, u32 cflags,
2154 unsigned long *irq_flags)
2155 {
2156 int ret;
2157
2158 while (true) {
2159 ret = try_to_grab_pending(work, cflags, irq_flags);
2160 if (ret >= 0)
2161 return ret;
2162 cpu_relax();
2163 }
2164 }
2165
2166 /**
2167 * insert_work - insert a work into a pool
2168 * @pwq: pwq @work belongs to
2169 * @work: work to insert
2170 * @head: insertion point
2171 * @extra_flags: extra WORK_STRUCT_* flags to set
2172 *
2173 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
2174 * work_struct flags.
2175 *
2176 * CONTEXT:
2177 * raw_spin_lock_irq(pool->lock).
2178 */
insert_work(struct pool_workqueue * pwq,struct work_struct * work,struct list_head * head,unsigned int extra_flags)2179 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
2180 struct list_head *head, unsigned int extra_flags)
2181 {
2182 debug_work_activate(work);
2183
2184 /* record the work call stack in order to print it in KASAN reports */
2185 kasan_record_aux_stack_noalloc(work);
2186
2187 /* we own @work, set data and link */
2188 set_work_pwq(work, pwq, extra_flags);
2189 list_add_tail(&work->entry, head);
2190 get_pwq(pwq);
2191 }
2192
2193 /*
2194 * Test whether @work is being queued from another work executing on the
2195 * same workqueue.
2196 */
is_chained_work(struct workqueue_struct * wq)2197 static bool is_chained_work(struct workqueue_struct *wq)
2198 {
2199 struct worker *worker;
2200
2201 worker = current_wq_worker();
2202 /*
2203 * Return %true iff I'm a worker executing a work item on @wq. If
2204 * I'm @worker, it's safe to dereference it without locking.
2205 */
2206 return worker && worker->current_pwq->wq == wq;
2207 }
2208
2209 /*
2210 * When queueing an unbound work item to a wq, prefer local CPU if allowed
2211 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
2212 * avoid perturbing sensitive tasks.
2213 */
wq_select_unbound_cpu(int cpu)2214 static int wq_select_unbound_cpu(int cpu)
2215 {
2216 int new_cpu;
2217
2218 if (likely(!wq_debug_force_rr_cpu)) {
2219 if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
2220 return cpu;
2221 } else {
2222 pr_warn_once("workqueue: round-robin CPU selection forced, expect performance impact\n");
2223 }
2224
2225 new_cpu = __this_cpu_read(wq_rr_cpu_last);
2226 new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
2227 if (unlikely(new_cpu >= nr_cpu_ids)) {
2228 new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
2229 if (unlikely(new_cpu >= nr_cpu_ids))
2230 return cpu;
2231 }
2232 __this_cpu_write(wq_rr_cpu_last, new_cpu);
2233
2234 return new_cpu;
2235 }
2236
__queue_work(int cpu,struct workqueue_struct * wq,struct work_struct * work)2237 static void __queue_work(int cpu, struct workqueue_struct *wq,
2238 struct work_struct *work)
2239 {
2240 struct pool_workqueue *pwq;
2241 struct worker_pool *last_pool, *pool;
2242 unsigned int work_flags;
2243 unsigned int req_cpu = cpu;
2244
2245 /*
2246 * While a work item is PENDING && off queue, a task trying to
2247 * steal the PENDING will busy-loop waiting for it to either get
2248 * queued or lose PENDING. Grabbing PENDING and queueing should
2249 * happen with IRQ disabled.
2250 */
2251 lockdep_assert_irqs_disabled();
2252
2253 /*
2254 * For a draining wq, only works from the same workqueue are
2255 * allowed. The __WQ_DESTROYING helps to spot the issue that
2256 * queues a new work item to a wq after destroy_workqueue(wq).
2257 */
2258 if (unlikely(wq->flags & (__WQ_DESTROYING | __WQ_DRAINING) &&
2259 WARN_ON_ONCE(!is_chained_work(wq))))
2260 return;
2261 rcu_read_lock();
2262 retry:
2263 /* pwq which will be used unless @work is executing elsewhere */
2264 if (req_cpu == WORK_CPU_UNBOUND) {
2265 if (wq->flags & WQ_UNBOUND)
2266 cpu = wq_select_unbound_cpu(raw_smp_processor_id());
2267 else
2268 cpu = raw_smp_processor_id();
2269 }
2270
2271 pwq = rcu_dereference(*per_cpu_ptr(wq->cpu_pwq, cpu));
2272 pool = pwq->pool;
2273
2274 /*
2275 * If @work was previously on a different pool, it might still be
2276 * running there, in which case the work needs to be queued on that
2277 * pool to guarantee non-reentrancy.
2278 *
2279 * For ordered workqueue, work items must be queued on the newest pwq
2280 * for accurate order management. Guaranteed order also guarantees
2281 * non-reentrancy. See the comments above unplug_oldest_pwq().
2282 */
2283 last_pool = get_work_pool(work);
2284 if (last_pool && last_pool != pool && !(wq->flags & __WQ_ORDERED)) {
2285 struct worker *worker;
2286
2287 raw_spin_lock(&last_pool->lock);
2288
2289 worker = find_worker_executing_work(last_pool, work);
2290
2291 if (worker && worker->current_pwq->wq == wq) {
2292 pwq = worker->current_pwq;
2293 pool = pwq->pool;
2294 WARN_ON_ONCE(pool != last_pool);
2295 } else {
2296 /* meh... not running there, queue here */
2297 raw_spin_unlock(&last_pool->lock);
2298 raw_spin_lock(&pool->lock);
2299 }
2300 } else {
2301 raw_spin_lock(&pool->lock);
2302 }
2303
2304 /*
2305 * pwq is determined and locked. For unbound pools, we could have raced
2306 * with pwq release and it could already be dead. If its refcnt is zero,
2307 * repeat pwq selection. Note that unbound pwqs never die without
2308 * another pwq replacing it in cpu_pwq or while work items are executing
2309 * on it, so the retrying is guaranteed to make forward-progress.
2310 */
2311 if (unlikely(!pwq->refcnt)) {
2312 if (wq->flags & WQ_UNBOUND) {
2313 raw_spin_unlock(&pool->lock);
2314 cpu_relax();
2315 goto retry;
2316 }
2317 /* oops */
2318 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
2319 wq->name, cpu);
2320 }
2321
2322 /* pwq determined, queue */
2323 trace_workqueue_queue_work(req_cpu, pwq, work);
2324
2325 if (WARN_ON(!list_empty(&work->entry)))
2326 goto out;
2327
2328 pwq->nr_in_flight[pwq->work_color]++;
2329 work_flags = work_color_to_flags(pwq->work_color);
2330
2331 /*
2332 * Limit the number of concurrently active work items to max_active.
2333 * @work must also queue behind existing inactive work items to maintain
2334 * ordering when max_active changes. See wq_adjust_max_active().
2335 */
2336 if (list_empty(&pwq->inactive_works) && pwq_tryinc_nr_active(pwq, false)) {
2337 if (list_empty(&pool->worklist))
2338 pool->watchdog_ts = jiffies;
2339
2340 trace_workqueue_activate_work(work);
2341 insert_work(pwq, work, &pool->worklist, work_flags);
2342 kick_pool(pool);
2343 } else {
2344 work_flags |= WORK_STRUCT_INACTIVE;
2345 insert_work(pwq, work, &pwq->inactive_works, work_flags);
2346 }
2347
2348 out:
2349 raw_spin_unlock(&pool->lock);
2350 rcu_read_unlock();
2351 }
2352
clear_pending_if_disabled(struct work_struct * work)2353 static bool clear_pending_if_disabled(struct work_struct *work)
2354 {
2355 unsigned long data = *work_data_bits(work);
2356 struct work_offq_data offqd;
2357
2358 if (likely((data & WORK_STRUCT_PWQ) ||
2359 !(data & WORK_OFFQ_DISABLE_MASK)))
2360 return false;
2361
2362 work_offqd_unpack(&offqd, data);
2363 set_work_pool_and_clear_pending(work, offqd.pool_id,
2364 work_offqd_pack_flags(&offqd));
2365 return true;
2366 }
2367
2368 /**
2369 * queue_work_on - queue work on specific cpu
2370 * @cpu: CPU number to execute work on
2371 * @wq: workqueue to use
2372 * @work: work to queue
2373 *
2374 * We queue the work to a specific CPU, the caller must ensure it
2375 * can't go away. Callers that fail to ensure that the specified
2376 * CPU cannot go away will execute on a randomly chosen CPU.
2377 * But note well that callers specifying a CPU that never has been
2378 * online will get a splat.
2379 *
2380 * Return: %false if @work was already on a queue, %true otherwise.
2381 */
queue_work_on(int cpu,struct workqueue_struct * wq,struct work_struct * work)2382 bool queue_work_on(int cpu, struct workqueue_struct *wq,
2383 struct work_struct *work)
2384 {
2385 bool ret = false;
2386 unsigned long irq_flags;
2387
2388 local_irq_save(irq_flags);
2389
2390 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)) &&
2391 !clear_pending_if_disabled(work)) {
2392 __queue_work(cpu, wq, work);
2393 ret = true;
2394 }
2395
2396 local_irq_restore(irq_flags);
2397 return ret;
2398 }
2399 EXPORT_SYMBOL(queue_work_on);
2400
2401 /**
2402 * select_numa_node_cpu - Select a CPU based on NUMA node
2403 * @node: NUMA node ID that we want to select a CPU from
2404 *
2405 * This function will attempt to find a "random" cpu available on a given
2406 * node. If there are no CPUs available on the given node it will return
2407 * WORK_CPU_UNBOUND indicating that we should just schedule to any
2408 * available CPU if we need to schedule this work.
2409 */
select_numa_node_cpu(int node)2410 static int select_numa_node_cpu(int node)
2411 {
2412 int cpu;
2413
2414 /* Delay binding to CPU if node is not valid or online */
2415 if (node < 0 || node >= MAX_NUMNODES || !node_online(node))
2416 return WORK_CPU_UNBOUND;
2417
2418 /* Use local node/cpu if we are already there */
2419 cpu = raw_smp_processor_id();
2420 if (node == cpu_to_node(cpu))
2421 return cpu;
2422
2423 /* Use "random" otherwise know as "first" online CPU of node */
2424 cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask);
2425
2426 /* If CPU is valid return that, otherwise just defer */
2427 return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND;
2428 }
2429
2430 /**
2431 * queue_work_node - queue work on a "random" cpu for a given NUMA node
2432 * @node: NUMA node that we are targeting the work for
2433 * @wq: workqueue to use
2434 * @work: work to queue
2435 *
2436 * We queue the work to a "random" CPU within a given NUMA node. The basic
2437 * idea here is to provide a way to somehow associate work with a given
2438 * NUMA node.
2439 *
2440 * This function will only make a best effort attempt at getting this onto
2441 * the right NUMA node. If no node is requested or the requested node is
2442 * offline then we just fall back to standard queue_work behavior.
2443 *
2444 * Currently the "random" CPU ends up being the first available CPU in the
2445 * intersection of cpu_online_mask and the cpumask of the node, unless we
2446 * are running on the node. In that case we just use the current CPU.
2447 *
2448 * Return: %false if @work was already on a queue, %true otherwise.
2449 */
queue_work_node(int node,struct workqueue_struct * wq,struct work_struct * work)2450 bool queue_work_node(int node, struct workqueue_struct *wq,
2451 struct work_struct *work)
2452 {
2453 unsigned long irq_flags;
2454 bool ret = false;
2455
2456 /*
2457 * This current implementation is specific to unbound workqueues.
2458 * Specifically we only return the first available CPU for a given
2459 * node instead of cycling through individual CPUs within the node.
2460 *
2461 * If this is used with a per-cpu workqueue then the logic in
2462 * workqueue_select_cpu_near would need to be updated to allow for
2463 * some round robin type logic.
2464 */
2465 WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND));
2466
2467 local_irq_save(irq_flags);
2468
2469 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)) &&
2470 !clear_pending_if_disabled(work)) {
2471 int cpu = select_numa_node_cpu(node);
2472
2473 __queue_work(cpu, wq, work);
2474 ret = true;
2475 }
2476
2477 local_irq_restore(irq_flags);
2478 return ret;
2479 }
2480 EXPORT_SYMBOL_GPL(queue_work_node);
2481
delayed_work_timer_fn(struct timer_list * t)2482 void delayed_work_timer_fn(struct timer_list *t)
2483 {
2484 struct delayed_work *dwork = from_timer(dwork, t, timer);
2485
2486 /* should have been called from irqsafe timer with irq already off */
2487 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2488 }
2489 EXPORT_SYMBOL(delayed_work_timer_fn);
2490
__queue_delayed_work(int cpu,struct workqueue_struct * wq,struct delayed_work * dwork,unsigned long delay)2491 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
2492 struct delayed_work *dwork, unsigned long delay)
2493 {
2494 struct timer_list *timer = &dwork->timer;
2495 struct work_struct *work = &dwork->work;
2496
2497 WARN_ON_ONCE(!wq);
2498 WARN_ON_ONCE(timer->function != delayed_work_timer_fn);
2499 WARN_ON_ONCE(timer_pending(timer));
2500 WARN_ON_ONCE(!list_empty(&work->entry));
2501
2502 /*
2503 * If @delay is 0, queue @dwork->work immediately. This is for
2504 * both optimization and correctness. The earliest @timer can
2505 * expire is on the closest next tick and delayed_work users depend
2506 * on that there's no such delay when @delay is 0.
2507 */
2508 if (!delay) {
2509 __queue_work(cpu, wq, &dwork->work);
2510 return;
2511 }
2512
2513 dwork->wq = wq;
2514 dwork->cpu = cpu;
2515 timer->expires = jiffies + delay;
2516
2517 if (housekeeping_enabled(HK_TYPE_TIMER)) {
2518 /* If the current cpu is a housekeeping cpu, use it. */
2519 cpu = smp_processor_id();
2520 if (!housekeeping_test_cpu(cpu, HK_TYPE_TIMER))
2521 cpu = housekeeping_any_cpu(HK_TYPE_TIMER);
2522 add_timer_on(timer, cpu);
2523 } else {
2524 if (likely(cpu == WORK_CPU_UNBOUND))
2525 add_timer_global(timer);
2526 else
2527 add_timer_on(timer, cpu);
2528 }
2529 }
2530
2531 /**
2532 * queue_delayed_work_on - queue work on specific CPU after delay
2533 * @cpu: CPU number to execute work on
2534 * @wq: workqueue to use
2535 * @dwork: work to queue
2536 * @delay: number of jiffies to wait before queueing
2537 *
2538 * Return: %false if @work was already on a queue, %true otherwise. If
2539 * @delay is zero and @dwork is idle, it will be scheduled for immediate
2540 * execution.
2541 */
queue_delayed_work_on(int cpu,struct workqueue_struct * wq,struct delayed_work * dwork,unsigned long delay)2542 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
2543 struct delayed_work *dwork, unsigned long delay)
2544 {
2545 struct work_struct *work = &dwork->work;
2546 bool ret = false;
2547 unsigned long irq_flags;
2548
2549 /* read the comment in __queue_work() */
2550 local_irq_save(irq_flags);
2551
2552 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)) &&
2553 !clear_pending_if_disabled(work)) {
2554 __queue_delayed_work(cpu, wq, dwork, delay);
2555 ret = true;
2556 }
2557
2558 local_irq_restore(irq_flags);
2559 return ret;
2560 }
2561 EXPORT_SYMBOL(queue_delayed_work_on);
2562
2563 /**
2564 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
2565 * @cpu: CPU number to execute work on
2566 * @wq: workqueue to use
2567 * @dwork: work to queue
2568 * @delay: number of jiffies to wait before queueing
2569 *
2570 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
2571 * modify @dwork's timer so that it expires after @delay. If @delay is
2572 * zero, @work is guaranteed to be scheduled immediately regardless of its
2573 * current state.
2574 *
2575 * Return: %false if @dwork was idle and queued, %true if @dwork was
2576 * pending and its timer was modified.
2577 *
2578 * This function is safe to call from any context including IRQ handler.
2579 * See try_to_grab_pending() for details.
2580 */
mod_delayed_work_on(int cpu,struct workqueue_struct * wq,struct delayed_work * dwork,unsigned long delay)2581 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
2582 struct delayed_work *dwork, unsigned long delay)
2583 {
2584 unsigned long irq_flags;
2585 bool ret;
2586
2587 ret = work_grab_pending(&dwork->work, WORK_CANCEL_DELAYED, &irq_flags);
2588
2589 if (!clear_pending_if_disabled(&dwork->work))
2590 __queue_delayed_work(cpu, wq, dwork, delay);
2591
2592 local_irq_restore(irq_flags);
2593 return ret;
2594 }
2595 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
2596
rcu_work_rcufn(struct rcu_head * rcu)2597 static void rcu_work_rcufn(struct rcu_head *rcu)
2598 {
2599 struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu);
2600
2601 /* read the comment in __queue_work() */
2602 local_irq_disable();
2603 __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work);
2604 local_irq_enable();
2605 }
2606
2607 /**
2608 * queue_rcu_work - queue work after a RCU grace period
2609 * @wq: workqueue to use
2610 * @rwork: work to queue
2611 *
2612 * Return: %false if @rwork was already pending, %true otherwise. Note
2613 * that a full RCU grace period is guaranteed only after a %true return.
2614 * While @rwork is guaranteed to be executed after a %false return, the
2615 * execution may happen before a full RCU grace period has passed.
2616 */
queue_rcu_work(struct workqueue_struct * wq,struct rcu_work * rwork)2617 bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork)
2618 {
2619 struct work_struct *work = &rwork->work;
2620
2621 /*
2622 * rcu_work can't be canceled or disabled. Warn if the user reached
2623 * inside @rwork and disabled the inner work.
2624 */
2625 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)) &&
2626 !WARN_ON_ONCE(clear_pending_if_disabled(work))) {
2627 rwork->wq = wq;
2628 call_rcu_hurry(&rwork->rcu, rcu_work_rcufn);
2629 return true;
2630 }
2631
2632 return false;
2633 }
2634 EXPORT_SYMBOL(queue_rcu_work);
2635
alloc_worker(int node)2636 static struct worker *alloc_worker(int node)
2637 {
2638 struct worker *worker;
2639
2640 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
2641 if (worker) {
2642 INIT_LIST_HEAD(&worker->entry);
2643 INIT_LIST_HEAD(&worker->scheduled);
2644 INIT_LIST_HEAD(&worker->node);
2645 /* on creation a worker is in !idle && prep state */
2646 worker->flags = WORKER_PREP;
2647 }
2648 return worker;
2649 }
2650
pool_allowed_cpus(struct worker_pool * pool)2651 static cpumask_t *pool_allowed_cpus(struct worker_pool *pool)
2652 {
2653 if (pool->cpu < 0 && pool->attrs->affn_strict)
2654 return pool->attrs->__pod_cpumask;
2655 else
2656 return pool->attrs->cpumask;
2657 }
2658
2659 /**
2660 * worker_attach_to_pool() - attach a worker to a pool
2661 * @worker: worker to be attached
2662 * @pool: the target pool
2663 *
2664 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
2665 * cpu-binding of @worker are kept coordinated with the pool across
2666 * cpu-[un]hotplugs.
2667 */
worker_attach_to_pool(struct worker * worker,struct worker_pool * pool)2668 static void worker_attach_to_pool(struct worker *worker,
2669 struct worker_pool *pool)
2670 {
2671 mutex_lock(&wq_pool_attach_mutex);
2672
2673 /*
2674 * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains stable
2675 * across this function. See the comments above the flag definition for
2676 * details. BH workers are, while per-CPU, always DISASSOCIATED.
2677 */
2678 if (pool->flags & POOL_DISASSOCIATED) {
2679 worker->flags |= WORKER_UNBOUND;
2680 } else {
2681 WARN_ON_ONCE(pool->flags & POOL_BH);
2682 kthread_set_per_cpu(worker->task, pool->cpu);
2683 }
2684
2685 if (worker->rescue_wq)
2686 set_cpus_allowed_ptr(worker->task, pool_allowed_cpus(pool));
2687
2688 list_add_tail(&worker->node, &pool->workers);
2689 worker->pool = pool;
2690
2691 mutex_unlock(&wq_pool_attach_mutex);
2692 }
2693
unbind_worker(struct worker * worker)2694 static void unbind_worker(struct worker *worker)
2695 {
2696 lockdep_assert_held(&wq_pool_attach_mutex);
2697
2698 kthread_set_per_cpu(worker->task, -1);
2699 if (cpumask_intersects(wq_unbound_cpumask, cpu_active_mask))
2700 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, wq_unbound_cpumask) < 0);
2701 else
2702 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0);
2703 }
2704
2705
detach_worker(struct worker * worker)2706 static void detach_worker(struct worker *worker)
2707 {
2708 lockdep_assert_held(&wq_pool_attach_mutex);
2709
2710 unbind_worker(worker);
2711 list_del(&worker->node);
2712 worker->pool = NULL;
2713 }
2714
2715 /**
2716 * worker_detach_from_pool() - detach a worker from its pool
2717 * @worker: worker which is attached to its pool
2718 *
2719 * Undo the attaching which had been done in worker_attach_to_pool(). The
2720 * caller worker shouldn't access to the pool after detached except it has
2721 * other reference to the pool.
2722 */
worker_detach_from_pool(struct worker * worker)2723 static void worker_detach_from_pool(struct worker *worker)
2724 {
2725 struct worker_pool *pool = worker->pool;
2726
2727 /* there is one permanent BH worker per CPU which should never detach */
2728 WARN_ON_ONCE(pool->flags & POOL_BH);
2729
2730 mutex_lock(&wq_pool_attach_mutex);
2731 detach_worker(worker);
2732 mutex_unlock(&wq_pool_attach_mutex);
2733
2734 /* clear leftover flags without pool->lock after it is detached */
2735 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
2736 }
2737
format_worker_id(char * buf,size_t size,struct worker * worker,struct worker_pool * pool)2738 static int format_worker_id(char *buf, size_t size, struct worker *worker,
2739 struct worker_pool *pool)
2740 {
2741 if (worker->rescue_wq)
2742 return scnprintf(buf, size, "kworker/R-%s",
2743 worker->rescue_wq->name);
2744
2745 if (pool) {
2746 if (pool->cpu >= 0)
2747 return scnprintf(buf, size, "kworker/%d:%d%s",
2748 pool->cpu, worker->id,
2749 pool->attrs->nice < 0 ? "H" : "");
2750 else
2751 return scnprintf(buf, size, "kworker/u%d:%d",
2752 pool->id, worker->id);
2753 } else {
2754 return scnprintf(buf, size, "kworker/dying");
2755 }
2756 }
2757
2758 /**
2759 * create_worker - create a new workqueue worker
2760 * @pool: pool the new worker will belong to
2761 *
2762 * Create and start a new worker which is attached to @pool.
2763 *
2764 * CONTEXT:
2765 * Might sleep. Does GFP_KERNEL allocations.
2766 *
2767 * Return:
2768 * Pointer to the newly created worker.
2769 */
create_worker(struct worker_pool * pool)2770 static struct worker *create_worker(struct worker_pool *pool)
2771 {
2772 struct worker *worker;
2773 int id;
2774
2775 /* ID is needed to determine kthread name */
2776 id = ida_alloc(&pool->worker_ida, GFP_KERNEL);
2777 if (id < 0) {
2778 pr_err_once("workqueue: Failed to allocate a worker ID: %pe\n",
2779 ERR_PTR(id));
2780 return NULL;
2781 }
2782
2783 worker = alloc_worker(pool->node);
2784 if (!worker) {
2785 pr_err_once("workqueue: Failed to allocate a worker\n");
2786 goto fail;
2787 }
2788
2789 worker->id = id;
2790
2791 if (!(pool->flags & POOL_BH)) {
2792 char id_buf[WORKER_ID_LEN];
2793
2794 format_worker_id(id_buf, sizeof(id_buf), worker, pool);
2795 worker->task = kthread_create_on_node(worker_thread, worker,
2796 pool->node, "%s", id_buf);
2797 if (IS_ERR(worker->task)) {
2798 if (PTR_ERR(worker->task) == -EINTR) {
2799 pr_err("workqueue: Interrupted when creating a worker thread \"%s\"\n",
2800 id_buf);
2801 } else {
2802 pr_err_once("workqueue: Failed to create a worker thread: %pe",
2803 worker->task);
2804 }
2805 goto fail;
2806 }
2807
2808 set_user_nice(worker->task, pool->attrs->nice);
2809 kthread_bind_mask(worker->task, pool_allowed_cpus(pool));
2810 }
2811
2812 /* successful, attach the worker to the pool */
2813 worker_attach_to_pool(worker, pool);
2814
2815 /* start the newly created worker */
2816 raw_spin_lock_irq(&pool->lock);
2817
2818 worker->pool->nr_workers++;
2819 worker_enter_idle(worker);
2820
2821 /*
2822 * @worker is waiting on a completion in kthread() and will trigger hung
2823 * check if not woken up soon. As kick_pool() is noop if @pool is empty,
2824 * wake it up explicitly.
2825 */
2826 if (worker->task)
2827 wake_up_process(worker->task);
2828
2829 raw_spin_unlock_irq(&pool->lock);
2830
2831 return worker;
2832
2833 fail:
2834 ida_free(&pool->worker_ida, id);
2835 kfree(worker);
2836 return NULL;
2837 }
2838
detach_dying_workers(struct list_head * cull_list)2839 static void detach_dying_workers(struct list_head *cull_list)
2840 {
2841 struct worker *worker;
2842
2843 list_for_each_entry(worker, cull_list, entry)
2844 detach_worker(worker);
2845 }
2846
reap_dying_workers(struct list_head * cull_list)2847 static void reap_dying_workers(struct list_head *cull_list)
2848 {
2849 struct worker *worker, *tmp;
2850
2851 list_for_each_entry_safe(worker, tmp, cull_list, entry) {
2852 list_del_init(&worker->entry);
2853 kthread_stop_put(worker->task);
2854 kfree(worker);
2855 }
2856 }
2857
2858 /**
2859 * set_worker_dying - Tag a worker for destruction
2860 * @worker: worker to be destroyed
2861 * @list: transfer worker away from its pool->idle_list and into list
2862 *
2863 * Tag @worker for destruction and adjust @pool stats accordingly. The worker
2864 * should be idle.
2865 *
2866 * CONTEXT:
2867 * raw_spin_lock_irq(pool->lock).
2868 */
set_worker_dying(struct worker * worker,struct list_head * list)2869 static void set_worker_dying(struct worker *worker, struct list_head *list)
2870 {
2871 struct worker_pool *pool = worker->pool;
2872
2873 lockdep_assert_held(&pool->lock);
2874 lockdep_assert_held(&wq_pool_attach_mutex);
2875
2876 /* sanity check frenzy */
2877 if (WARN_ON(worker->current_work) ||
2878 WARN_ON(!list_empty(&worker->scheduled)) ||
2879 WARN_ON(!(worker->flags & WORKER_IDLE)))
2880 return;
2881
2882 pool->nr_workers--;
2883 pool->nr_idle--;
2884
2885 worker->flags |= WORKER_DIE;
2886
2887 list_move(&worker->entry, list);
2888
2889 /* get an extra task struct reference for later kthread_stop_put() */
2890 get_task_struct(worker->task);
2891 }
2892
2893 /**
2894 * idle_worker_timeout - check if some idle workers can now be deleted.
2895 * @t: The pool's idle_timer that just expired
2896 *
2897 * The timer is armed in worker_enter_idle(). Note that it isn't disarmed in
2898 * worker_leave_idle(), as a worker flicking between idle and active while its
2899 * pool is at the too_many_workers() tipping point would cause too much timer
2900 * housekeeping overhead. Since IDLE_WORKER_TIMEOUT is long enough, we just let
2901 * it expire and re-evaluate things from there.
2902 */
idle_worker_timeout(struct timer_list * t)2903 static void idle_worker_timeout(struct timer_list *t)
2904 {
2905 struct worker_pool *pool = from_timer(pool, t, idle_timer);
2906 bool do_cull = false;
2907
2908 if (work_pending(&pool->idle_cull_work))
2909 return;
2910
2911 raw_spin_lock_irq(&pool->lock);
2912
2913 if (too_many_workers(pool)) {
2914 struct worker *worker;
2915 unsigned long expires;
2916
2917 /* idle_list is kept in LIFO order, check the last one */
2918 worker = list_last_entry(&pool->idle_list, struct worker, entry);
2919 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2920 do_cull = !time_before(jiffies, expires);
2921
2922 if (!do_cull)
2923 mod_timer(&pool->idle_timer, expires);
2924 }
2925 raw_spin_unlock_irq(&pool->lock);
2926
2927 if (do_cull)
2928 queue_work(system_unbound_wq, &pool->idle_cull_work);
2929 }
2930
2931 /**
2932 * idle_cull_fn - cull workers that have been idle for too long.
2933 * @work: the pool's work for handling these idle workers
2934 *
2935 * This goes through a pool's idle workers and gets rid of those that have been
2936 * idle for at least IDLE_WORKER_TIMEOUT seconds.
2937 *
2938 * We don't want to disturb isolated CPUs because of a pcpu kworker being
2939 * culled, so this also resets worker affinity. This requires a sleepable
2940 * context, hence the split between timer callback and work item.
2941 */
idle_cull_fn(struct work_struct * work)2942 static void idle_cull_fn(struct work_struct *work)
2943 {
2944 struct worker_pool *pool = container_of(work, struct worker_pool, idle_cull_work);
2945 LIST_HEAD(cull_list);
2946
2947 /*
2948 * Grabbing wq_pool_attach_mutex here ensures an already-running worker
2949 * cannot proceed beyong set_pf_worker() in its self-destruct path.
2950 * This is required as a previously-preempted worker could run after
2951 * set_worker_dying() has happened but before detach_dying_workers() did.
2952 */
2953 mutex_lock(&wq_pool_attach_mutex);
2954 raw_spin_lock_irq(&pool->lock);
2955
2956 while (too_many_workers(pool)) {
2957 struct worker *worker;
2958 unsigned long expires;
2959
2960 worker = list_last_entry(&pool->idle_list, struct worker, entry);
2961 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2962
2963 if (time_before(jiffies, expires)) {
2964 mod_timer(&pool->idle_timer, expires);
2965 break;
2966 }
2967
2968 set_worker_dying(worker, &cull_list);
2969 }
2970
2971 raw_spin_unlock_irq(&pool->lock);
2972 detach_dying_workers(&cull_list);
2973 mutex_unlock(&wq_pool_attach_mutex);
2974
2975 reap_dying_workers(&cull_list);
2976 }
2977
send_mayday(struct work_struct * work)2978 static void send_mayday(struct work_struct *work)
2979 {
2980 struct pool_workqueue *pwq = get_work_pwq(work);
2981 struct workqueue_struct *wq = pwq->wq;
2982
2983 lockdep_assert_held(&wq_mayday_lock);
2984
2985 if (!wq->rescuer)
2986 return;
2987
2988 /* mayday mayday mayday */
2989 if (list_empty(&pwq->mayday_node)) {
2990 /*
2991 * If @pwq is for an unbound wq, its base ref may be put at
2992 * any time due to an attribute change. Pin @pwq until the
2993 * rescuer is done with it.
2994 */
2995 get_pwq(pwq);
2996 list_add_tail(&pwq->mayday_node, &wq->maydays);
2997 wake_up_process(wq->rescuer->task);
2998 pwq->stats[PWQ_STAT_MAYDAY]++;
2999 }
3000 }
3001
pool_mayday_timeout(struct timer_list * t)3002 static void pool_mayday_timeout(struct timer_list *t)
3003 {
3004 struct worker_pool *pool = from_timer(pool, t, mayday_timer);
3005 struct work_struct *work;
3006
3007 raw_spin_lock_irq(&pool->lock);
3008 raw_spin_lock(&wq_mayday_lock); /* for wq->maydays */
3009
3010 if (need_to_create_worker(pool)) {
3011 /*
3012 * We've been trying to create a new worker but
3013 * haven't been successful. We might be hitting an
3014 * allocation deadlock. Send distress signals to
3015 * rescuers.
3016 */
3017 list_for_each_entry(work, &pool->worklist, entry)
3018 send_mayday(work);
3019 }
3020
3021 raw_spin_unlock(&wq_mayday_lock);
3022 raw_spin_unlock_irq(&pool->lock);
3023
3024 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
3025 }
3026
3027 /**
3028 * maybe_create_worker - create a new worker if necessary
3029 * @pool: pool to create a new worker for
3030 *
3031 * Create a new worker for @pool if necessary. @pool is guaranteed to
3032 * have at least one idle worker on return from this function. If
3033 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
3034 * sent to all rescuers with works scheduled on @pool to resolve
3035 * possible allocation deadlock.
3036 *
3037 * On return, need_to_create_worker() is guaranteed to be %false and
3038 * may_start_working() %true.
3039 *
3040 * LOCKING:
3041 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
3042 * multiple times. Does GFP_KERNEL allocations. Called only from
3043 * manager.
3044 */
maybe_create_worker(struct worker_pool * pool)3045 static void maybe_create_worker(struct worker_pool *pool)
3046 __releases(&pool->lock)
3047 __acquires(&pool->lock)
3048 {
3049 restart:
3050 raw_spin_unlock_irq(&pool->lock);
3051
3052 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
3053 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
3054
3055 while (true) {
3056 if (create_worker(pool) || !need_to_create_worker(pool))
3057 break;
3058
3059 schedule_timeout_interruptible(CREATE_COOLDOWN);
3060
3061 if (!need_to_create_worker(pool))
3062 break;
3063 }
3064
3065 del_timer_sync(&pool->mayday_timer);
3066 raw_spin_lock_irq(&pool->lock);
3067 /*
3068 * This is necessary even after a new worker was just successfully
3069 * created as @pool->lock was dropped and the new worker might have
3070 * already become busy.
3071 */
3072 if (need_to_create_worker(pool))
3073 goto restart;
3074 }
3075
3076 /**
3077 * manage_workers - manage worker pool
3078 * @worker: self
3079 *
3080 * Assume the manager role and manage the worker pool @worker belongs
3081 * to. At any given time, there can be only zero or one manager per
3082 * pool. The exclusion is handled automatically by this function.
3083 *
3084 * The caller can safely start processing works on false return. On
3085 * true return, it's guaranteed that need_to_create_worker() is false
3086 * and may_start_working() is true.
3087 *
3088 * CONTEXT:
3089 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
3090 * multiple times. Does GFP_KERNEL allocations.
3091 *
3092 * Return:
3093 * %false if the pool doesn't need management and the caller can safely
3094 * start processing works, %true if management function was performed and
3095 * the conditions that the caller verified before calling the function may
3096 * no longer be true.
3097 */
manage_workers(struct worker * worker)3098 static bool manage_workers(struct worker *worker)
3099 {
3100 struct worker_pool *pool = worker->pool;
3101
3102 if (pool->flags & POOL_MANAGER_ACTIVE)
3103 return false;
3104
3105 pool->flags |= POOL_MANAGER_ACTIVE;
3106 pool->manager = worker;
3107
3108 maybe_create_worker(pool);
3109
3110 pool->manager = NULL;
3111 pool->flags &= ~POOL_MANAGER_ACTIVE;
3112 rcuwait_wake_up(&manager_wait);
3113 return true;
3114 }
3115
3116 /**
3117 * process_one_work - process single work
3118 * @worker: self
3119 * @work: work to process
3120 *
3121 * Process @work. This function contains all the logics necessary to
3122 * process a single work including synchronization against and
3123 * interaction with other workers on the same cpu, queueing and
3124 * flushing. As long as context requirement is met, any worker can
3125 * call this function to process a work.
3126 *
3127 * CONTEXT:
3128 * raw_spin_lock_irq(pool->lock) which is released and regrabbed.
3129 */
process_one_work(struct worker * worker,struct work_struct * work)3130 static void process_one_work(struct worker *worker, struct work_struct *work)
3131 __releases(&pool->lock)
3132 __acquires(&pool->lock)
3133 {
3134 struct pool_workqueue *pwq = get_work_pwq(work);
3135 struct worker_pool *pool = worker->pool;
3136 unsigned long work_data;
3137 int lockdep_start_depth, rcu_start_depth;
3138 bool bh_draining = pool->flags & POOL_BH_DRAINING;
3139 #ifdef CONFIG_LOCKDEP
3140 /*
3141 * It is permissible to free the struct work_struct from
3142 * inside the function that is called from it, this we need to
3143 * take into account for lockdep too. To avoid bogus "held
3144 * lock freed" warnings as well as problems when looking into
3145 * work->lockdep_map, make a copy and use that here.
3146 */
3147 struct lockdep_map lockdep_map;
3148
3149 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
3150 #endif
3151 /* ensure we're on the correct CPU */
3152 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
3153 raw_smp_processor_id() != pool->cpu);
3154
3155 /* claim and dequeue */
3156 debug_work_deactivate(work);
3157 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
3158 worker->current_work = work;
3159 worker->current_func = work->func;
3160 worker->current_pwq = pwq;
3161 if (worker->task)
3162 worker->current_at = worker->task->se.sum_exec_runtime;
3163 work_data = *work_data_bits(work);
3164 worker->current_color = get_work_color(work_data);
3165
3166 /*
3167 * Record wq name for cmdline and debug reporting, may get
3168 * overridden through set_worker_desc().
3169 */
3170 strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN);
3171
3172 list_del_init(&work->entry);
3173
3174 /*
3175 * CPU intensive works don't participate in concurrency management.
3176 * They're the scheduler's responsibility. This takes @worker out
3177 * of concurrency management and the next code block will chain
3178 * execution of the pending work items.
3179 */
3180 if (unlikely(pwq->wq->flags & WQ_CPU_INTENSIVE))
3181 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
3182
3183 /*
3184 * Kick @pool if necessary. It's always noop for per-cpu worker pools
3185 * since nr_running would always be >= 1 at this point. This is used to
3186 * chain execution of the pending work items for WORKER_NOT_RUNNING
3187 * workers such as the UNBOUND and CPU_INTENSIVE ones.
3188 */
3189 kick_pool(pool);
3190
3191 /*
3192 * Record the last pool and clear PENDING which should be the last
3193 * update to @work. Also, do this inside @pool->lock so that
3194 * PENDING and queued state changes happen together while IRQ is
3195 * disabled.
3196 */
3197 set_work_pool_and_clear_pending(work, pool->id, pool_offq_flags(pool));
3198
3199 pwq->stats[PWQ_STAT_STARTED]++;
3200 raw_spin_unlock_irq(&pool->lock);
3201
3202 rcu_start_depth = rcu_preempt_depth();
3203 lockdep_start_depth = lockdep_depth(current);
3204 /* see drain_dead_softirq_workfn() */
3205 if (!bh_draining)
3206 lock_map_acquire(&pwq->wq->lockdep_map);
3207 lock_map_acquire(&lockdep_map);
3208 /*
3209 * Strictly speaking we should mark the invariant state without holding
3210 * any locks, that is, before these two lock_map_acquire()'s.
3211 *
3212 * However, that would result in:
3213 *
3214 * A(W1)
3215 * WFC(C)
3216 * A(W1)
3217 * C(C)
3218 *
3219 * Which would create W1->C->W1 dependencies, even though there is no
3220 * actual deadlock possible. There are two solutions, using a
3221 * read-recursive acquire on the work(queue) 'locks', but this will then
3222 * hit the lockdep limitation on recursive locks, or simply discard
3223 * these locks.
3224 *
3225 * AFAICT there is no possible deadlock scenario between the
3226 * flush_work() and complete() primitives (except for single-threaded
3227 * workqueues), so hiding them isn't a problem.
3228 */
3229 lockdep_invariant_state(true);
3230 trace_workqueue_execute_start(work);
3231 worker->current_func(work);
3232 /*
3233 * While we must be careful to not use "work" after this, the trace
3234 * point will only record its address.
3235 */
3236 trace_workqueue_execute_end(work, worker->current_func);
3237 pwq->stats[PWQ_STAT_COMPLETED]++;
3238 lock_map_release(&lockdep_map);
3239 if (!bh_draining)
3240 lock_map_release(&pwq->wq->lockdep_map);
3241
3242 if (unlikely((worker->task && in_atomic()) ||
3243 lockdep_depth(current) != lockdep_start_depth ||
3244 rcu_preempt_depth() != rcu_start_depth)) {
3245 pr_err("BUG: workqueue leaked atomic, lock or RCU: %s[%d]\n"
3246 " preempt=0x%08x lock=%d->%d RCU=%d->%d workfn=%ps\n",
3247 current->comm, task_pid_nr(current), preempt_count(),
3248 lockdep_start_depth, lockdep_depth(current),
3249 rcu_start_depth, rcu_preempt_depth(),
3250 worker->current_func);
3251 debug_show_held_locks(current);
3252 dump_stack();
3253 }
3254
3255 /*
3256 * The following prevents a kworker from hogging CPU on !PREEMPTION
3257 * kernels, where a requeueing work item waiting for something to
3258 * happen could deadlock with stop_machine as such work item could
3259 * indefinitely requeue itself while all other CPUs are trapped in
3260 * stop_machine. At the same time, report a quiescent RCU state so
3261 * the same condition doesn't freeze RCU.
3262 */
3263 if (worker->task)
3264 cond_resched();
3265
3266 raw_spin_lock_irq(&pool->lock);
3267
3268 /*
3269 * In addition to %WQ_CPU_INTENSIVE, @worker may also have been marked
3270 * CPU intensive by wq_worker_tick() if @work hogged CPU longer than
3271 * wq_cpu_intensive_thresh_us. Clear it.
3272 */
3273 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
3274
3275 /* tag the worker for identification in schedule() */
3276 worker->last_func = worker->current_func;
3277
3278 /* we're done with it, release */
3279 hash_del(&worker->hentry);
3280 worker->current_work = NULL;
3281 worker->current_func = NULL;
3282 worker->current_pwq = NULL;
3283 worker->current_color = INT_MAX;
3284
3285 /* must be the last step, see the function comment */
3286 pwq_dec_nr_in_flight(pwq, work_data);
3287 }
3288
3289 /**
3290 * process_scheduled_works - process scheduled works
3291 * @worker: self
3292 *
3293 * Process all scheduled works. Please note that the scheduled list
3294 * may change while processing a work, so this function repeatedly
3295 * fetches a work from the top and executes it.
3296 *
3297 * CONTEXT:
3298 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
3299 * multiple times.
3300 */
process_scheduled_works(struct worker * worker)3301 static void process_scheduled_works(struct worker *worker)
3302 {
3303 struct work_struct *work;
3304 bool first = true;
3305
3306 while ((work = list_first_entry_or_null(&worker->scheduled,
3307 struct work_struct, entry))) {
3308 if (first) {
3309 worker->pool->watchdog_ts = jiffies;
3310 first = false;
3311 }
3312 process_one_work(worker, work);
3313 }
3314 }
3315
set_pf_worker(bool val)3316 static void set_pf_worker(bool val)
3317 {
3318 mutex_lock(&wq_pool_attach_mutex);
3319 if (val)
3320 current->flags |= PF_WQ_WORKER;
3321 else
3322 current->flags &= ~PF_WQ_WORKER;
3323 mutex_unlock(&wq_pool_attach_mutex);
3324 }
3325
3326 /**
3327 * worker_thread - the worker thread function
3328 * @__worker: self
3329 *
3330 * The worker thread function. All workers belong to a worker_pool -
3331 * either a per-cpu one or dynamic unbound one. These workers process all
3332 * work items regardless of their specific target workqueue. The only
3333 * exception is work items which belong to workqueues with a rescuer which
3334 * will be explained in rescuer_thread().
3335 *
3336 * Return: 0
3337 */
worker_thread(void * __worker)3338 static int worker_thread(void *__worker)
3339 {
3340 struct worker *worker = __worker;
3341 struct worker_pool *pool = worker->pool;
3342
3343 /* tell the scheduler that this is a workqueue worker */
3344 set_pf_worker(true);
3345 woke_up:
3346 raw_spin_lock_irq(&pool->lock);
3347
3348 /* am I supposed to die? */
3349 if (unlikely(worker->flags & WORKER_DIE)) {
3350 raw_spin_unlock_irq(&pool->lock);
3351 set_pf_worker(false);
3352
3353 ida_free(&pool->worker_ida, worker->id);
3354 WARN_ON_ONCE(!list_empty(&worker->entry));
3355 return 0;
3356 }
3357
3358 worker_leave_idle(worker);
3359 recheck:
3360 /* no more worker necessary? */
3361 if (!need_more_worker(pool))
3362 goto sleep;
3363
3364 /* do we need to manage? */
3365 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
3366 goto recheck;
3367
3368 /*
3369 * ->scheduled list can only be filled while a worker is
3370 * preparing to process a work or actually processing it.
3371 * Make sure nobody diddled with it while I was sleeping.
3372 */
3373 WARN_ON_ONCE(!list_empty(&worker->scheduled));
3374
3375 /*
3376 * Finish PREP stage. We're guaranteed to have at least one idle
3377 * worker or that someone else has already assumed the manager
3378 * role. This is where @worker starts participating in concurrency
3379 * management if applicable and concurrency management is restored
3380 * after being rebound. See rebind_workers() for details.
3381 */
3382 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
3383
3384 do {
3385 struct work_struct *work =
3386 list_first_entry(&pool->worklist,
3387 struct work_struct, entry);
3388
3389 if (assign_work(work, worker, NULL))
3390 process_scheduled_works(worker);
3391 } while (keep_working(pool));
3392
3393 worker_set_flags(worker, WORKER_PREP);
3394 sleep:
3395 /*
3396 * pool->lock is held and there's no work to process and no need to
3397 * manage, sleep. Workers are woken up only while holding
3398 * pool->lock or from local cpu, so setting the current state
3399 * before releasing pool->lock is enough to prevent losing any
3400 * event.
3401 */
3402 worker_enter_idle(worker);
3403 __set_current_state(TASK_IDLE);
3404 raw_spin_unlock_irq(&pool->lock);
3405 schedule();
3406 goto woke_up;
3407 }
3408
3409 /**
3410 * rescuer_thread - the rescuer thread function
3411 * @__rescuer: self
3412 *
3413 * Workqueue rescuer thread function. There's one rescuer for each
3414 * workqueue which has WQ_MEM_RECLAIM set.
3415 *
3416 * Regular work processing on a pool may block trying to create a new
3417 * worker which uses GFP_KERNEL allocation which has slight chance of
3418 * developing into deadlock if some works currently on the same queue
3419 * need to be processed to satisfy the GFP_KERNEL allocation. This is
3420 * the problem rescuer solves.
3421 *
3422 * When such condition is possible, the pool summons rescuers of all
3423 * workqueues which have works queued on the pool and let them process
3424 * those works so that forward progress can be guaranteed.
3425 *
3426 * This should happen rarely.
3427 *
3428 * Return: 0
3429 */
rescuer_thread(void * __rescuer)3430 static int rescuer_thread(void *__rescuer)
3431 {
3432 struct worker *rescuer = __rescuer;
3433 struct workqueue_struct *wq = rescuer->rescue_wq;
3434 bool should_stop;
3435
3436 set_user_nice(current, RESCUER_NICE_LEVEL);
3437
3438 /*
3439 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
3440 * doesn't participate in concurrency management.
3441 */
3442 set_pf_worker(true);
3443 repeat:
3444 set_current_state(TASK_IDLE);
3445
3446 /*
3447 * By the time the rescuer is requested to stop, the workqueue
3448 * shouldn't have any work pending, but @wq->maydays may still have
3449 * pwq(s) queued. This can happen by non-rescuer workers consuming
3450 * all the work items before the rescuer got to them. Go through
3451 * @wq->maydays processing before acting on should_stop so that the
3452 * list is always empty on exit.
3453 */
3454 should_stop = kthread_should_stop();
3455
3456 /* see whether any pwq is asking for help */
3457 raw_spin_lock_irq(&wq_mayday_lock);
3458
3459 while (!list_empty(&wq->maydays)) {
3460 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
3461 struct pool_workqueue, mayday_node);
3462 struct worker_pool *pool = pwq->pool;
3463 struct work_struct *work, *n;
3464
3465 __set_current_state(TASK_RUNNING);
3466 list_del_init(&pwq->mayday_node);
3467
3468 raw_spin_unlock_irq(&wq_mayday_lock);
3469
3470 worker_attach_to_pool(rescuer, pool);
3471
3472 raw_spin_lock_irq(&pool->lock);
3473
3474 /*
3475 * Slurp in all works issued via this workqueue and
3476 * process'em.
3477 */
3478 WARN_ON_ONCE(!list_empty(&rescuer->scheduled));
3479 list_for_each_entry_safe(work, n, &pool->worklist, entry) {
3480 if (get_work_pwq(work) == pwq &&
3481 assign_work(work, rescuer, &n))
3482 pwq->stats[PWQ_STAT_RESCUED]++;
3483 }
3484
3485 if (!list_empty(&rescuer->scheduled)) {
3486 process_scheduled_works(rescuer);
3487
3488 /*
3489 * The above execution of rescued work items could
3490 * have created more to rescue through
3491 * pwq_activate_first_inactive() or chained
3492 * queueing. Let's put @pwq back on mayday list so
3493 * that such back-to-back work items, which may be
3494 * being used to relieve memory pressure, don't
3495 * incur MAYDAY_INTERVAL delay inbetween.
3496 */
3497 if (pwq->nr_active && need_to_create_worker(pool)) {
3498 raw_spin_lock(&wq_mayday_lock);
3499 /*
3500 * Queue iff we aren't racing destruction
3501 * and somebody else hasn't queued it already.
3502 */
3503 if (wq->rescuer && list_empty(&pwq->mayday_node)) {
3504 get_pwq(pwq);
3505 list_add_tail(&pwq->mayday_node, &wq->maydays);
3506 }
3507 raw_spin_unlock(&wq_mayday_lock);
3508 }
3509 }
3510
3511 /*
3512 * Put the reference grabbed by send_mayday(). @pool won't
3513 * go away while we're still attached to it.
3514 */
3515 put_pwq(pwq);
3516
3517 /*
3518 * Leave this pool. Notify regular workers; otherwise, we end up
3519 * with 0 concurrency and stalling the execution.
3520 */
3521 kick_pool(pool);
3522
3523 raw_spin_unlock_irq(&pool->lock);
3524
3525 worker_detach_from_pool(rescuer);
3526
3527 raw_spin_lock_irq(&wq_mayday_lock);
3528 }
3529
3530 raw_spin_unlock_irq(&wq_mayday_lock);
3531
3532 if (should_stop) {
3533 __set_current_state(TASK_RUNNING);
3534 set_pf_worker(false);
3535 return 0;
3536 }
3537
3538 /* rescuers should never participate in concurrency management */
3539 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
3540 schedule();
3541 goto repeat;
3542 }
3543
bh_worker(struct worker * worker)3544 static void bh_worker(struct worker *worker)
3545 {
3546 struct worker_pool *pool = worker->pool;
3547 int nr_restarts = BH_WORKER_RESTARTS;
3548 unsigned long end = jiffies + BH_WORKER_JIFFIES;
3549
3550 raw_spin_lock_irq(&pool->lock);
3551 worker_leave_idle(worker);
3552
3553 /*
3554 * This function follows the structure of worker_thread(). See there for
3555 * explanations on each step.
3556 */
3557 if (!need_more_worker(pool))
3558 goto done;
3559
3560 WARN_ON_ONCE(!list_empty(&worker->scheduled));
3561 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
3562
3563 do {
3564 struct work_struct *work =
3565 list_first_entry(&pool->worklist,
3566 struct work_struct, entry);
3567
3568 if (assign_work(work, worker, NULL))
3569 process_scheduled_works(worker);
3570 } while (keep_working(pool) &&
3571 --nr_restarts && time_before(jiffies, end));
3572
3573 worker_set_flags(worker, WORKER_PREP);
3574 done:
3575 worker_enter_idle(worker);
3576 kick_pool(pool);
3577 raw_spin_unlock_irq(&pool->lock);
3578 }
3579
3580 /*
3581 * TODO: Convert all tasklet users to workqueue and use softirq directly.
3582 *
3583 * This is currently called from tasklet[_hi]action() and thus is also called
3584 * whenever there are tasklets to run. Let's do an early exit if there's nothing
3585 * queued. Once conversion from tasklet is complete, the need_more_worker() test
3586 * can be dropped.
3587 *
3588 * After full conversion, we'll add worker->softirq_action, directly use the
3589 * softirq action and obtain the worker pointer from the softirq_action pointer.
3590 */
workqueue_softirq_action(bool highpri)3591 void workqueue_softirq_action(bool highpri)
3592 {
3593 struct worker_pool *pool =
3594 &per_cpu(bh_worker_pools, smp_processor_id())[highpri];
3595 if (need_more_worker(pool))
3596 bh_worker(list_first_entry(&pool->workers, struct worker, node));
3597 }
3598
3599 struct wq_drain_dead_softirq_work {
3600 struct work_struct work;
3601 struct worker_pool *pool;
3602 struct completion done;
3603 };
3604
drain_dead_softirq_workfn(struct work_struct * work)3605 static void drain_dead_softirq_workfn(struct work_struct *work)
3606 {
3607 struct wq_drain_dead_softirq_work *dead_work =
3608 container_of(work, struct wq_drain_dead_softirq_work, work);
3609 struct worker_pool *pool = dead_work->pool;
3610 bool repeat;
3611
3612 /*
3613 * @pool's CPU is dead and we want to execute its still pending work
3614 * items from this BH work item which is running on a different CPU. As
3615 * its CPU is dead, @pool can't be kicked and, as work execution path
3616 * will be nested, a lockdep annotation needs to be suppressed. Mark
3617 * @pool with %POOL_BH_DRAINING for the special treatments.
3618 */
3619 raw_spin_lock_irq(&pool->lock);
3620 pool->flags |= POOL_BH_DRAINING;
3621 raw_spin_unlock_irq(&pool->lock);
3622
3623 bh_worker(list_first_entry(&pool->workers, struct worker, node));
3624
3625 raw_spin_lock_irq(&pool->lock);
3626 pool->flags &= ~POOL_BH_DRAINING;
3627 repeat = need_more_worker(pool);
3628 raw_spin_unlock_irq(&pool->lock);
3629
3630 /*
3631 * bh_worker() might hit consecutive execution limit and bail. If there
3632 * still are pending work items, reschedule self and return so that we
3633 * don't hog this CPU's BH.
3634 */
3635 if (repeat) {
3636 if (pool->attrs->nice == HIGHPRI_NICE_LEVEL)
3637 queue_work(system_bh_highpri_wq, work);
3638 else
3639 queue_work(system_bh_wq, work);
3640 } else {
3641 complete(&dead_work->done);
3642 }
3643 }
3644
3645 /*
3646 * @cpu is dead. Drain the remaining BH work items on the current CPU. It's
3647 * possible to allocate dead_work per CPU and avoid flushing. However, then we
3648 * have to worry about draining overlapping with CPU coming back online or
3649 * nesting (one CPU's dead_work queued on another CPU which is also dead and so
3650 * on). Let's keep it simple and drain them synchronously. These are BH work
3651 * items which shouldn't be requeued on the same pool. Shouldn't take long.
3652 */
workqueue_softirq_dead(unsigned int cpu)3653 void workqueue_softirq_dead(unsigned int cpu)
3654 {
3655 int i;
3656
3657 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
3658 struct worker_pool *pool = &per_cpu(bh_worker_pools, cpu)[i];
3659 struct wq_drain_dead_softirq_work dead_work;
3660
3661 if (!need_more_worker(pool))
3662 continue;
3663
3664 INIT_WORK_ONSTACK(&dead_work.work, drain_dead_softirq_workfn);
3665 dead_work.pool = pool;
3666 init_completion(&dead_work.done);
3667
3668 if (pool->attrs->nice == HIGHPRI_NICE_LEVEL)
3669 queue_work(system_bh_highpri_wq, &dead_work.work);
3670 else
3671 queue_work(system_bh_wq, &dead_work.work);
3672
3673 wait_for_completion(&dead_work.done);
3674 destroy_work_on_stack(&dead_work.work);
3675 }
3676 }
3677
3678 /**
3679 * check_flush_dependency - check for flush dependency sanity
3680 * @target_wq: workqueue being flushed
3681 * @target_work: work item being flushed (NULL for workqueue flushes)
3682 *
3683 * %current is trying to flush the whole @target_wq or @target_work on it.
3684 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
3685 * reclaiming memory or running on a workqueue which doesn't have
3686 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
3687 * a deadlock.
3688 */
check_flush_dependency(struct workqueue_struct * target_wq,struct work_struct * target_work)3689 static void check_flush_dependency(struct workqueue_struct *target_wq,
3690 struct work_struct *target_work)
3691 {
3692 work_func_t target_func = target_work ? target_work->func : NULL;
3693 struct worker *worker;
3694
3695 if (target_wq->flags & WQ_MEM_RECLAIM)
3696 return;
3697
3698 worker = current_wq_worker();
3699
3700 WARN_ONCE(current->flags & PF_MEMALLOC,
3701 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps",
3702 current->pid, current->comm, target_wq->name, target_func);
3703 WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
3704 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
3705 "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps",
3706 worker->current_pwq->wq->name, worker->current_func,
3707 target_wq->name, target_func);
3708 }
3709
3710 struct wq_barrier {
3711 struct work_struct work;
3712 struct completion done;
3713 struct task_struct *task; /* purely informational */
3714 };
3715
wq_barrier_func(struct work_struct * work)3716 static void wq_barrier_func(struct work_struct *work)
3717 {
3718 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
3719 complete(&barr->done);
3720 }
3721
3722 /**
3723 * insert_wq_barrier - insert a barrier work
3724 * @pwq: pwq to insert barrier into
3725 * @barr: wq_barrier to insert
3726 * @target: target work to attach @barr to
3727 * @worker: worker currently executing @target, NULL if @target is not executing
3728 *
3729 * @barr is linked to @target such that @barr is completed only after
3730 * @target finishes execution. Please note that the ordering
3731 * guarantee is observed only with respect to @target and on the local
3732 * cpu.
3733 *
3734 * Currently, a queued barrier can't be canceled. This is because
3735 * try_to_grab_pending() can't determine whether the work to be
3736 * grabbed is at the head of the queue and thus can't clear LINKED
3737 * flag of the previous work while there must be a valid next work
3738 * after a work with LINKED flag set.
3739 *
3740 * Note that when @worker is non-NULL, @target may be modified
3741 * underneath us, so we can't reliably determine pwq from @target.
3742 *
3743 * CONTEXT:
3744 * raw_spin_lock_irq(pool->lock).
3745 */
insert_wq_barrier(struct pool_workqueue * pwq,struct wq_barrier * barr,struct work_struct * target,struct worker * worker)3746 static void insert_wq_barrier(struct pool_workqueue *pwq,
3747 struct wq_barrier *barr,
3748 struct work_struct *target, struct worker *worker)
3749 {
3750 static __maybe_unused struct lock_class_key bh_key, thr_key;
3751 unsigned int work_flags = 0;
3752 unsigned int work_color;
3753 struct list_head *head;
3754
3755 /*
3756 * debugobject calls are safe here even with pool->lock locked
3757 * as we know for sure that this will not trigger any of the
3758 * checks and call back into the fixup functions where we
3759 * might deadlock.
3760 *
3761 * BH and threaded workqueues need separate lockdep keys to avoid
3762 * spuriously triggering "inconsistent {SOFTIRQ-ON-W} -> {IN-SOFTIRQ-W}
3763 * usage".
3764 */
3765 INIT_WORK_ONSTACK_KEY(&barr->work, wq_barrier_func,
3766 (pwq->wq->flags & WQ_BH) ? &bh_key : &thr_key);
3767 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
3768
3769 init_completion_map(&barr->done, &target->lockdep_map);
3770
3771 barr->task = current;
3772
3773 /* The barrier work item does not participate in nr_active. */
3774 work_flags |= WORK_STRUCT_INACTIVE;
3775
3776 /*
3777 * If @target is currently being executed, schedule the
3778 * barrier to the worker; otherwise, put it after @target.
3779 */
3780 if (worker) {
3781 head = worker->scheduled.next;
3782 work_color = worker->current_color;
3783 } else {
3784 unsigned long *bits = work_data_bits(target);
3785
3786 head = target->entry.next;
3787 /* there can already be other linked works, inherit and set */
3788 work_flags |= *bits & WORK_STRUCT_LINKED;
3789 work_color = get_work_color(*bits);
3790 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
3791 }
3792
3793 pwq->nr_in_flight[work_color]++;
3794 work_flags |= work_color_to_flags(work_color);
3795
3796 insert_work(pwq, &barr->work, head, work_flags);
3797 }
3798
3799 /**
3800 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
3801 * @wq: workqueue being flushed
3802 * @flush_color: new flush color, < 0 for no-op
3803 * @work_color: new work color, < 0 for no-op
3804 *
3805 * Prepare pwqs for workqueue flushing.
3806 *
3807 * If @flush_color is non-negative, flush_color on all pwqs should be
3808 * -1. If no pwq has in-flight commands at the specified color, all
3809 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
3810 * has in flight commands, its pwq->flush_color is set to
3811 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
3812 * wakeup logic is armed and %true is returned.
3813 *
3814 * The caller should have initialized @wq->first_flusher prior to
3815 * calling this function with non-negative @flush_color. If
3816 * @flush_color is negative, no flush color update is done and %false
3817 * is returned.
3818 *
3819 * If @work_color is non-negative, all pwqs should have the same
3820 * work_color which is previous to @work_color and all will be
3821 * advanced to @work_color.
3822 *
3823 * CONTEXT:
3824 * mutex_lock(wq->mutex).
3825 *
3826 * Return:
3827 * %true if @flush_color >= 0 and there's something to flush. %false
3828 * otherwise.
3829 */
flush_workqueue_prep_pwqs(struct workqueue_struct * wq,int flush_color,int work_color)3830 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
3831 int flush_color, int work_color)
3832 {
3833 bool wait = false;
3834 struct pool_workqueue *pwq;
3835
3836 if (flush_color >= 0) {
3837 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
3838 atomic_set(&wq->nr_pwqs_to_flush, 1);
3839 }
3840
3841 for_each_pwq(pwq, wq) {
3842 struct worker_pool *pool = pwq->pool;
3843
3844 raw_spin_lock_irq(&pool->lock);
3845
3846 if (flush_color >= 0) {
3847 WARN_ON_ONCE(pwq->flush_color != -1);
3848
3849 if (pwq->nr_in_flight[flush_color]) {
3850 pwq->flush_color = flush_color;
3851 atomic_inc(&wq->nr_pwqs_to_flush);
3852 wait = true;
3853 }
3854 }
3855
3856 if (work_color >= 0) {
3857 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
3858 pwq->work_color = work_color;
3859 }
3860
3861 raw_spin_unlock_irq(&pool->lock);
3862 }
3863
3864 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
3865 complete(&wq->first_flusher->done);
3866
3867 return wait;
3868 }
3869
touch_wq_lockdep_map(struct workqueue_struct * wq)3870 static void touch_wq_lockdep_map(struct workqueue_struct *wq)
3871 {
3872 #ifdef CONFIG_LOCKDEP
3873 if (wq->flags & WQ_BH)
3874 local_bh_disable();
3875
3876 lock_map_acquire(&wq->lockdep_map);
3877 lock_map_release(&wq->lockdep_map);
3878
3879 if (wq->flags & WQ_BH)
3880 local_bh_enable();
3881 #endif
3882 }
3883
touch_work_lockdep_map(struct work_struct * work,struct workqueue_struct * wq)3884 static void touch_work_lockdep_map(struct work_struct *work,
3885 struct workqueue_struct *wq)
3886 {
3887 #ifdef CONFIG_LOCKDEP
3888 if (wq->flags & WQ_BH)
3889 local_bh_disable();
3890
3891 lock_map_acquire(&work->lockdep_map);
3892 lock_map_release(&work->lockdep_map);
3893
3894 if (wq->flags & WQ_BH)
3895 local_bh_enable();
3896 #endif
3897 }
3898
3899 /**
3900 * __flush_workqueue - ensure that any scheduled work has run to completion.
3901 * @wq: workqueue to flush
3902 *
3903 * This function sleeps until all work items which were queued on entry
3904 * have finished execution, but it is not livelocked by new incoming ones.
3905 */
__flush_workqueue(struct workqueue_struct * wq)3906 void __flush_workqueue(struct workqueue_struct *wq)
3907 {
3908 struct wq_flusher this_flusher = {
3909 .list = LIST_HEAD_INIT(this_flusher.list),
3910 .flush_color = -1,
3911 .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map),
3912 };
3913 int next_color;
3914
3915 if (WARN_ON(!wq_online))
3916 return;
3917
3918 touch_wq_lockdep_map(wq);
3919
3920 mutex_lock(&wq->mutex);
3921
3922 /*
3923 * Start-to-wait phase
3924 */
3925 next_color = work_next_color(wq->work_color);
3926
3927 if (next_color != wq->flush_color) {
3928 /*
3929 * Color space is not full. The current work_color
3930 * becomes our flush_color and work_color is advanced
3931 * by one.
3932 */
3933 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
3934 this_flusher.flush_color = wq->work_color;
3935 wq->work_color = next_color;
3936
3937 if (!wq->first_flusher) {
3938 /* no flush in progress, become the first flusher */
3939 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
3940
3941 wq->first_flusher = &this_flusher;
3942
3943 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
3944 wq->work_color)) {
3945 /* nothing to flush, done */
3946 wq->flush_color = next_color;
3947 wq->first_flusher = NULL;
3948 goto out_unlock;
3949 }
3950 } else {
3951 /* wait in queue */
3952 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
3953 list_add_tail(&this_flusher.list, &wq->flusher_queue);
3954 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
3955 }
3956 } else {
3957 /*
3958 * Oops, color space is full, wait on overflow queue.
3959 * The next flush completion will assign us
3960 * flush_color and transfer to flusher_queue.
3961 */
3962 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
3963 }
3964
3965 check_flush_dependency(wq, NULL);
3966
3967 mutex_unlock(&wq->mutex);
3968
3969 wait_for_completion(&this_flusher.done);
3970
3971 /*
3972 * Wake-up-and-cascade phase
3973 *
3974 * First flushers are responsible for cascading flushes and
3975 * handling overflow. Non-first flushers can simply return.
3976 */
3977 if (READ_ONCE(wq->first_flusher) != &this_flusher)
3978 return;
3979
3980 mutex_lock(&wq->mutex);
3981
3982 /* we might have raced, check again with mutex held */
3983 if (wq->first_flusher != &this_flusher)
3984 goto out_unlock;
3985
3986 WRITE_ONCE(wq->first_flusher, NULL);
3987
3988 WARN_ON_ONCE(!list_empty(&this_flusher.list));
3989 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
3990
3991 while (true) {
3992 struct wq_flusher *next, *tmp;
3993
3994 /* complete all the flushers sharing the current flush color */
3995 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
3996 if (next->flush_color != wq->flush_color)
3997 break;
3998 list_del_init(&next->list);
3999 complete(&next->done);
4000 }
4001
4002 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
4003 wq->flush_color != work_next_color(wq->work_color));
4004
4005 /* this flush_color is finished, advance by one */
4006 wq->flush_color = work_next_color(wq->flush_color);
4007
4008 /* one color has been freed, handle overflow queue */
4009 if (!list_empty(&wq->flusher_overflow)) {
4010 /*
4011 * Assign the same color to all overflowed
4012 * flushers, advance work_color and append to
4013 * flusher_queue. This is the start-to-wait
4014 * phase for these overflowed flushers.
4015 */
4016 list_for_each_entry(tmp, &wq->flusher_overflow, list)
4017 tmp->flush_color = wq->work_color;
4018
4019 wq->work_color = work_next_color(wq->work_color);
4020
4021 list_splice_tail_init(&wq->flusher_overflow,
4022 &wq->flusher_queue);
4023 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
4024 }
4025
4026 if (list_empty(&wq->flusher_queue)) {
4027 WARN_ON_ONCE(wq->flush_color != wq->work_color);
4028 break;
4029 }
4030
4031 /*
4032 * Need to flush more colors. Make the next flusher
4033 * the new first flusher and arm pwqs.
4034 */
4035 WARN_ON_ONCE(wq->flush_color == wq->work_color);
4036 WARN_ON_ONCE(wq->flush_color != next->flush_color);
4037
4038 list_del_init(&next->list);
4039 wq->first_flusher = next;
4040
4041 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
4042 break;
4043
4044 /*
4045 * Meh... this color is already done, clear first
4046 * flusher and repeat cascading.
4047 */
4048 wq->first_flusher = NULL;
4049 }
4050
4051 out_unlock:
4052 mutex_unlock(&wq->mutex);
4053 }
4054 EXPORT_SYMBOL(__flush_workqueue);
4055
4056 /**
4057 * drain_workqueue - drain a workqueue
4058 * @wq: workqueue to drain
4059 *
4060 * Wait until the workqueue becomes empty. While draining is in progress,
4061 * only chain queueing is allowed. IOW, only currently pending or running
4062 * work items on @wq can queue further work items on it. @wq is flushed
4063 * repeatedly until it becomes empty. The number of flushing is determined
4064 * by the depth of chaining and should be relatively short. Whine if it
4065 * takes too long.
4066 */
drain_workqueue(struct workqueue_struct * wq)4067 void drain_workqueue(struct workqueue_struct *wq)
4068 {
4069 unsigned int flush_cnt = 0;
4070 struct pool_workqueue *pwq;
4071
4072 /*
4073 * __queue_work() needs to test whether there are drainers, is much
4074 * hotter than drain_workqueue() and already looks at @wq->flags.
4075 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
4076 */
4077 mutex_lock(&wq->mutex);
4078 if (!wq->nr_drainers++)
4079 wq->flags |= __WQ_DRAINING;
4080 mutex_unlock(&wq->mutex);
4081 reflush:
4082 __flush_workqueue(wq);
4083
4084 mutex_lock(&wq->mutex);
4085
4086 for_each_pwq(pwq, wq) {
4087 bool drained;
4088
4089 raw_spin_lock_irq(&pwq->pool->lock);
4090 drained = pwq_is_empty(pwq);
4091 raw_spin_unlock_irq(&pwq->pool->lock);
4092
4093 if (drained)
4094 continue;
4095
4096 if (++flush_cnt == 10 ||
4097 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
4098 pr_warn("workqueue %s: %s() isn't complete after %u tries\n",
4099 wq->name, __func__, flush_cnt);
4100
4101 mutex_unlock(&wq->mutex);
4102 goto reflush;
4103 }
4104
4105 if (!--wq->nr_drainers)
4106 wq->flags &= ~__WQ_DRAINING;
4107 mutex_unlock(&wq->mutex);
4108 }
4109 EXPORT_SYMBOL_GPL(drain_workqueue);
4110
start_flush_work(struct work_struct * work,struct wq_barrier * barr,bool from_cancel)4111 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
4112 bool from_cancel)
4113 {
4114 struct worker *worker = NULL;
4115 struct worker_pool *pool;
4116 struct pool_workqueue *pwq;
4117 struct workqueue_struct *wq;
4118
4119 rcu_read_lock();
4120 pool = get_work_pool(work);
4121 if (!pool) {
4122 rcu_read_unlock();
4123 return false;
4124 }
4125
4126 raw_spin_lock_irq(&pool->lock);
4127 /* see the comment in try_to_grab_pending() with the same code */
4128 pwq = get_work_pwq(work);
4129 if (pwq) {
4130 if (unlikely(pwq->pool != pool))
4131 goto already_gone;
4132 } else {
4133 worker = find_worker_executing_work(pool, work);
4134 if (!worker)
4135 goto already_gone;
4136 pwq = worker->current_pwq;
4137 }
4138
4139 wq = pwq->wq;
4140 check_flush_dependency(wq, work);
4141
4142 insert_wq_barrier(pwq, barr, work, worker);
4143 raw_spin_unlock_irq(&pool->lock);
4144
4145 touch_work_lockdep_map(work, wq);
4146
4147 /*
4148 * Force a lock recursion deadlock when using flush_work() inside a
4149 * single-threaded or rescuer equipped workqueue.
4150 *
4151 * For single threaded workqueues the deadlock happens when the work
4152 * is after the work issuing the flush_work(). For rescuer equipped
4153 * workqueues the deadlock happens when the rescuer stalls, blocking
4154 * forward progress.
4155 */
4156 if (!from_cancel && (wq->saved_max_active == 1 || wq->rescuer))
4157 touch_wq_lockdep_map(wq);
4158
4159 rcu_read_unlock();
4160 return true;
4161 already_gone:
4162 raw_spin_unlock_irq(&pool->lock);
4163 rcu_read_unlock();
4164 return false;
4165 }
4166
__flush_work(struct work_struct * work,bool from_cancel)4167 static bool __flush_work(struct work_struct *work, bool from_cancel)
4168 {
4169 struct wq_barrier barr;
4170 unsigned long data;
4171
4172 if (WARN_ON(!wq_online))
4173 return false;
4174
4175 if (WARN_ON(!work->func))
4176 return false;
4177
4178 if (!start_flush_work(work, &barr, from_cancel))
4179 return false;
4180
4181 /*
4182 * start_flush_work() returned %true. If @from_cancel is set, we know
4183 * that @work must have been executing during start_flush_work() and
4184 * can't currently be queued. Its data must contain OFFQ bits. If @work
4185 * was queued on a BH workqueue, we also know that it was running in the
4186 * BH context and thus can be busy-waited.
4187 */
4188 data = *work_data_bits(work);
4189 if (from_cancel &&
4190 !WARN_ON_ONCE(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_BH)) {
4191 /*
4192 * On RT, prevent a live lock when %current preempted soft
4193 * interrupt processing or prevents ksoftirqd from running by
4194 * keeping flipping BH. If the BH work item runs on a different
4195 * CPU then this has no effect other than doing the BH
4196 * disable/enable dance for nothing. This is copied from
4197 * kernel/softirq.c::tasklet_unlock_spin_wait().
4198 */
4199 while (!try_wait_for_completion(&barr.done)) {
4200 if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
4201 local_bh_disable();
4202 local_bh_enable();
4203 } else {
4204 cpu_relax();
4205 }
4206 }
4207 } else {
4208 wait_for_completion(&barr.done);
4209 }
4210
4211 destroy_work_on_stack(&barr.work);
4212 return true;
4213 }
4214
4215 /**
4216 * flush_work - wait for a work to finish executing the last queueing instance
4217 * @work: the work to flush
4218 *
4219 * Wait until @work has finished execution. @work is guaranteed to be idle
4220 * on return if it hasn't been requeued since flush started.
4221 *
4222 * Return:
4223 * %true if flush_work() waited for the work to finish execution,
4224 * %false if it was already idle.
4225 */
flush_work(struct work_struct * work)4226 bool flush_work(struct work_struct *work)
4227 {
4228 might_sleep();
4229 return __flush_work(work, false);
4230 }
4231 EXPORT_SYMBOL_GPL(flush_work);
4232
4233 /**
4234 * flush_delayed_work - wait for a dwork to finish executing the last queueing
4235 * @dwork: the delayed work to flush
4236 *
4237 * Delayed timer is cancelled and the pending work is queued for
4238 * immediate execution. Like flush_work(), this function only
4239 * considers the last queueing instance of @dwork.
4240 *
4241 * Return:
4242 * %true if flush_work() waited for the work to finish execution,
4243 * %false if it was already idle.
4244 */
flush_delayed_work(struct delayed_work * dwork)4245 bool flush_delayed_work(struct delayed_work *dwork)
4246 {
4247 local_irq_disable();
4248 if (del_timer_sync(&dwork->timer))
4249 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
4250 local_irq_enable();
4251 return flush_work(&dwork->work);
4252 }
4253 EXPORT_SYMBOL(flush_delayed_work);
4254
4255 /**
4256 * flush_rcu_work - wait for a rwork to finish executing the last queueing
4257 * @rwork: the rcu work to flush
4258 *
4259 * Return:
4260 * %true if flush_rcu_work() waited for the work to finish execution,
4261 * %false if it was already idle.
4262 */
flush_rcu_work(struct rcu_work * rwork)4263 bool flush_rcu_work(struct rcu_work *rwork)
4264 {
4265 if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) {
4266 rcu_barrier();
4267 flush_work(&rwork->work);
4268 return true;
4269 } else {
4270 return flush_work(&rwork->work);
4271 }
4272 }
4273 EXPORT_SYMBOL(flush_rcu_work);
4274
work_offqd_disable(struct work_offq_data * offqd)4275 static void work_offqd_disable(struct work_offq_data *offqd)
4276 {
4277 const unsigned long max = (1lu << WORK_OFFQ_DISABLE_BITS) - 1;
4278
4279 if (likely(offqd->disable < max))
4280 offqd->disable++;
4281 else
4282 WARN_ONCE(true, "workqueue: work disable count overflowed\n");
4283 }
4284
work_offqd_enable(struct work_offq_data * offqd)4285 static void work_offqd_enable(struct work_offq_data *offqd)
4286 {
4287 if (likely(offqd->disable > 0))
4288 offqd->disable--;
4289 else
4290 WARN_ONCE(true, "workqueue: work disable count underflowed\n");
4291 }
4292
__cancel_work(struct work_struct * work,u32 cflags)4293 static bool __cancel_work(struct work_struct *work, u32 cflags)
4294 {
4295 struct work_offq_data offqd;
4296 unsigned long irq_flags;
4297 int ret;
4298
4299 ret = work_grab_pending(work, cflags, &irq_flags);
4300
4301 work_offqd_unpack(&offqd, *work_data_bits(work));
4302
4303 if (cflags & WORK_CANCEL_DISABLE)
4304 work_offqd_disable(&offqd);
4305
4306 set_work_pool_and_clear_pending(work, offqd.pool_id,
4307 work_offqd_pack_flags(&offqd));
4308 local_irq_restore(irq_flags);
4309 return ret;
4310 }
4311
__cancel_work_sync(struct work_struct * work,u32 cflags)4312 static bool __cancel_work_sync(struct work_struct *work, u32 cflags)
4313 {
4314 bool ret;
4315
4316 ret = __cancel_work(work, cflags | WORK_CANCEL_DISABLE);
4317
4318 if (*work_data_bits(work) & WORK_OFFQ_BH)
4319 WARN_ON_ONCE(in_hardirq());
4320 else
4321 might_sleep();
4322
4323 /*
4324 * Skip __flush_work() during early boot when we know that @work isn't
4325 * executing. This allows canceling during early boot.
4326 */
4327 if (wq_online)
4328 __flush_work(work, true);
4329
4330 if (!(cflags & WORK_CANCEL_DISABLE))
4331 enable_work(work);
4332
4333 return ret;
4334 }
4335
4336 /*
4337 * See cancel_delayed_work()
4338 */
cancel_work(struct work_struct * work)4339 bool cancel_work(struct work_struct *work)
4340 {
4341 return __cancel_work(work, 0);
4342 }
4343 EXPORT_SYMBOL(cancel_work);
4344
4345 /**
4346 * cancel_work_sync - cancel a work and wait for it to finish
4347 * @work: the work to cancel
4348 *
4349 * Cancel @work and wait for its execution to finish. This function can be used
4350 * even if the work re-queues itself or migrates to another workqueue. On return
4351 * from this function, @work is guaranteed to be not pending or executing on any
4352 * CPU as long as there aren't racing enqueues.
4353 *
4354 * cancel_work_sync(&delayed_work->work) must not be used for delayed_work's.
4355 * Use cancel_delayed_work_sync() instead.
4356 *
4357 * Must be called from a sleepable context if @work was last queued on a non-BH
4358 * workqueue. Can also be called from non-hardirq atomic contexts including BH
4359 * if @work was last queued on a BH workqueue.
4360 *
4361 * Returns %true if @work was pending, %false otherwise.
4362 */
cancel_work_sync(struct work_struct * work)4363 bool cancel_work_sync(struct work_struct *work)
4364 {
4365 return __cancel_work_sync(work, 0);
4366 }
4367 EXPORT_SYMBOL_GPL(cancel_work_sync);
4368
4369 /**
4370 * cancel_delayed_work - cancel a delayed work
4371 * @dwork: delayed_work to cancel
4372 *
4373 * Kill off a pending delayed_work.
4374 *
4375 * Return: %true if @dwork was pending and canceled; %false if it wasn't
4376 * pending.
4377 *
4378 * Note:
4379 * The work callback function may still be running on return, unless
4380 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
4381 * use cancel_delayed_work_sync() to wait on it.
4382 *
4383 * This function is safe to call from any context including IRQ handler.
4384 */
cancel_delayed_work(struct delayed_work * dwork)4385 bool cancel_delayed_work(struct delayed_work *dwork)
4386 {
4387 return __cancel_work(&dwork->work, WORK_CANCEL_DELAYED);
4388 }
4389 EXPORT_SYMBOL(cancel_delayed_work);
4390
4391 /**
4392 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
4393 * @dwork: the delayed work cancel
4394 *
4395 * This is cancel_work_sync() for delayed works.
4396 *
4397 * Return:
4398 * %true if @dwork was pending, %false otherwise.
4399 */
cancel_delayed_work_sync(struct delayed_work * dwork)4400 bool cancel_delayed_work_sync(struct delayed_work *dwork)
4401 {
4402 return __cancel_work_sync(&dwork->work, WORK_CANCEL_DELAYED);
4403 }
4404 EXPORT_SYMBOL(cancel_delayed_work_sync);
4405
4406 /**
4407 * disable_work - Disable and cancel a work item
4408 * @work: work item to disable
4409 *
4410 * Disable @work by incrementing its disable count and cancel it if currently
4411 * pending. As long as the disable count is non-zero, any attempt to queue @work
4412 * will fail and return %false. The maximum supported disable depth is 2 to the
4413 * power of %WORK_OFFQ_DISABLE_BITS, currently 65536.
4414 *
4415 * Can be called from any context. Returns %true if @work was pending, %false
4416 * otherwise.
4417 */
disable_work(struct work_struct * work)4418 bool disable_work(struct work_struct *work)
4419 {
4420 return __cancel_work(work, WORK_CANCEL_DISABLE);
4421 }
4422 EXPORT_SYMBOL_GPL(disable_work);
4423
4424 /**
4425 * disable_work_sync - Disable, cancel and drain a work item
4426 * @work: work item to disable
4427 *
4428 * Similar to disable_work() but also wait for @work to finish if currently
4429 * executing.
4430 *
4431 * Must be called from a sleepable context if @work was last queued on a non-BH
4432 * workqueue. Can also be called from non-hardirq atomic contexts including BH
4433 * if @work was last queued on a BH workqueue.
4434 *
4435 * Returns %true if @work was pending, %false otherwise.
4436 */
disable_work_sync(struct work_struct * work)4437 bool disable_work_sync(struct work_struct *work)
4438 {
4439 return __cancel_work_sync(work, WORK_CANCEL_DISABLE);
4440 }
4441 EXPORT_SYMBOL_GPL(disable_work_sync);
4442
4443 /**
4444 * enable_work - Enable a work item
4445 * @work: work item to enable
4446 *
4447 * Undo disable_work[_sync]() by decrementing @work's disable count. @work can
4448 * only be queued if its disable count is 0.
4449 *
4450 * Can be called from any context. Returns %true if the disable count reached 0.
4451 * Otherwise, %false.
4452 */
enable_work(struct work_struct * work)4453 bool enable_work(struct work_struct *work)
4454 {
4455 struct work_offq_data offqd;
4456 unsigned long irq_flags;
4457
4458 work_grab_pending(work, 0, &irq_flags);
4459
4460 work_offqd_unpack(&offqd, *work_data_bits(work));
4461 work_offqd_enable(&offqd);
4462 set_work_pool_and_clear_pending(work, offqd.pool_id,
4463 work_offqd_pack_flags(&offqd));
4464 local_irq_restore(irq_flags);
4465
4466 return !offqd.disable;
4467 }
4468 EXPORT_SYMBOL_GPL(enable_work);
4469
4470 /**
4471 * disable_delayed_work - Disable and cancel a delayed work item
4472 * @dwork: delayed work item to disable
4473 *
4474 * disable_work() for delayed work items.
4475 */
disable_delayed_work(struct delayed_work * dwork)4476 bool disable_delayed_work(struct delayed_work *dwork)
4477 {
4478 return __cancel_work(&dwork->work,
4479 WORK_CANCEL_DELAYED | WORK_CANCEL_DISABLE);
4480 }
4481 EXPORT_SYMBOL_GPL(disable_delayed_work);
4482
4483 /**
4484 * disable_delayed_work_sync - Disable, cancel and drain a delayed work item
4485 * @dwork: delayed work item to disable
4486 *
4487 * disable_work_sync() for delayed work items.
4488 */
disable_delayed_work_sync(struct delayed_work * dwork)4489 bool disable_delayed_work_sync(struct delayed_work *dwork)
4490 {
4491 return __cancel_work_sync(&dwork->work,
4492 WORK_CANCEL_DELAYED | WORK_CANCEL_DISABLE);
4493 }
4494 EXPORT_SYMBOL_GPL(disable_delayed_work_sync);
4495
4496 /**
4497 * enable_delayed_work - Enable a delayed work item
4498 * @dwork: delayed work item to enable
4499 *
4500 * enable_work() for delayed work items.
4501 */
enable_delayed_work(struct delayed_work * dwork)4502 bool enable_delayed_work(struct delayed_work *dwork)
4503 {
4504 return enable_work(&dwork->work);
4505 }
4506 EXPORT_SYMBOL_GPL(enable_delayed_work);
4507
4508 /**
4509 * schedule_on_each_cpu - execute a function synchronously on each online CPU
4510 * @func: the function to call
4511 *
4512 * schedule_on_each_cpu() executes @func on each online CPU using the
4513 * system workqueue and blocks until all CPUs have completed.
4514 * schedule_on_each_cpu() is very slow.
4515 *
4516 * Return:
4517 * 0 on success, -errno on failure.
4518 */
schedule_on_each_cpu(work_func_t func)4519 int schedule_on_each_cpu(work_func_t func)
4520 {
4521 int cpu;
4522 struct work_struct __percpu *works;
4523
4524 works = alloc_percpu(struct work_struct);
4525 if (!works)
4526 return -ENOMEM;
4527
4528 cpus_read_lock();
4529
4530 for_each_online_cpu(cpu) {
4531 struct work_struct *work = per_cpu_ptr(works, cpu);
4532
4533 INIT_WORK(work, func);
4534 schedule_work_on(cpu, work);
4535 }
4536
4537 for_each_online_cpu(cpu)
4538 flush_work(per_cpu_ptr(works, cpu));
4539
4540 cpus_read_unlock();
4541 free_percpu(works);
4542 return 0;
4543 }
4544
4545 /**
4546 * execute_in_process_context - reliably execute the routine with user context
4547 * @fn: the function to execute
4548 * @ew: guaranteed storage for the execute work structure (must
4549 * be available when the work executes)
4550 *
4551 * Executes the function immediately if process context is available,
4552 * otherwise schedules the function for delayed execution.
4553 *
4554 * Return: 0 - function was executed
4555 * 1 - function was scheduled for execution
4556 */
execute_in_process_context(work_func_t fn,struct execute_work * ew)4557 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
4558 {
4559 if (!in_interrupt()) {
4560 fn(&ew->work);
4561 return 0;
4562 }
4563
4564 INIT_WORK(&ew->work, fn);
4565 schedule_work(&ew->work);
4566
4567 return 1;
4568 }
4569 EXPORT_SYMBOL_GPL(execute_in_process_context);
4570
4571 /**
4572 * free_workqueue_attrs - free a workqueue_attrs
4573 * @attrs: workqueue_attrs to free
4574 *
4575 * Undo alloc_workqueue_attrs().
4576 */
free_workqueue_attrs(struct workqueue_attrs * attrs)4577 void free_workqueue_attrs(struct workqueue_attrs *attrs)
4578 {
4579 if (attrs) {
4580 free_cpumask_var(attrs->cpumask);
4581 free_cpumask_var(attrs->__pod_cpumask);
4582 kfree(attrs);
4583 }
4584 }
4585
4586 /**
4587 * alloc_workqueue_attrs - allocate a workqueue_attrs
4588 *
4589 * Allocate a new workqueue_attrs, initialize with default settings and
4590 * return it.
4591 *
4592 * Return: The allocated new workqueue_attr on success. %NULL on failure.
4593 */
alloc_workqueue_attrs(void)4594 struct workqueue_attrs *alloc_workqueue_attrs(void)
4595 {
4596 struct workqueue_attrs *attrs;
4597
4598 attrs = kzalloc(sizeof(*attrs), GFP_KERNEL);
4599 if (!attrs)
4600 goto fail;
4601 if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL))
4602 goto fail;
4603 if (!alloc_cpumask_var(&attrs->__pod_cpumask, GFP_KERNEL))
4604 goto fail;
4605
4606 cpumask_copy(attrs->cpumask, cpu_possible_mask);
4607 attrs->affn_scope = WQ_AFFN_DFL;
4608 return attrs;
4609 fail:
4610 free_workqueue_attrs(attrs);
4611 return NULL;
4612 }
4613
copy_workqueue_attrs(struct workqueue_attrs * to,const struct workqueue_attrs * from)4614 static void copy_workqueue_attrs(struct workqueue_attrs *to,
4615 const struct workqueue_attrs *from)
4616 {
4617 to->nice = from->nice;
4618 cpumask_copy(to->cpumask, from->cpumask);
4619 cpumask_copy(to->__pod_cpumask, from->__pod_cpumask);
4620 to->affn_strict = from->affn_strict;
4621
4622 /*
4623 * Unlike hash and equality test, copying shouldn't ignore wq-only
4624 * fields as copying is used for both pool and wq attrs. Instead,
4625 * get_unbound_pool() explicitly clears the fields.
4626 */
4627 to->affn_scope = from->affn_scope;
4628 to->ordered = from->ordered;
4629 }
4630
4631 /*
4632 * Some attrs fields are workqueue-only. Clear them for worker_pool's. See the
4633 * comments in 'struct workqueue_attrs' definition.
4634 */
wqattrs_clear_for_pool(struct workqueue_attrs * attrs)4635 static void wqattrs_clear_for_pool(struct workqueue_attrs *attrs)
4636 {
4637 attrs->affn_scope = WQ_AFFN_NR_TYPES;
4638 attrs->ordered = false;
4639 if (attrs->affn_strict)
4640 cpumask_copy(attrs->cpumask, cpu_possible_mask);
4641 }
4642
4643 /* hash value of the content of @attr */
wqattrs_hash(const struct workqueue_attrs * attrs)4644 static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
4645 {
4646 u32 hash = 0;
4647
4648 hash = jhash_1word(attrs->nice, hash);
4649 hash = jhash_1word(attrs->affn_strict, hash);
4650 hash = jhash(cpumask_bits(attrs->__pod_cpumask),
4651 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
4652 if (!attrs->affn_strict)
4653 hash = jhash(cpumask_bits(attrs->cpumask),
4654 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
4655 return hash;
4656 }
4657
4658 /* content equality test */
wqattrs_equal(const struct workqueue_attrs * a,const struct workqueue_attrs * b)4659 static bool wqattrs_equal(const struct workqueue_attrs *a,
4660 const struct workqueue_attrs *b)
4661 {
4662 if (a->nice != b->nice)
4663 return false;
4664 if (a->affn_strict != b->affn_strict)
4665 return false;
4666 if (!cpumask_equal(a->__pod_cpumask, b->__pod_cpumask))
4667 return false;
4668 if (!a->affn_strict && !cpumask_equal(a->cpumask, b->cpumask))
4669 return false;
4670 return true;
4671 }
4672
4673 /* Update @attrs with actually available CPUs */
wqattrs_actualize_cpumask(struct workqueue_attrs * attrs,const cpumask_t * unbound_cpumask)4674 static void wqattrs_actualize_cpumask(struct workqueue_attrs *attrs,
4675 const cpumask_t *unbound_cpumask)
4676 {
4677 /*
4678 * Calculate the effective CPU mask of @attrs given @unbound_cpumask. If
4679 * @attrs->cpumask doesn't overlap with @unbound_cpumask, we fallback to
4680 * @unbound_cpumask.
4681 */
4682 cpumask_and(attrs->cpumask, attrs->cpumask, unbound_cpumask);
4683 if (unlikely(cpumask_empty(attrs->cpumask)))
4684 cpumask_copy(attrs->cpumask, unbound_cpumask);
4685 }
4686
4687 /* find wq_pod_type to use for @attrs */
4688 static const struct wq_pod_type *
wqattrs_pod_type(const struct workqueue_attrs * attrs)4689 wqattrs_pod_type(const struct workqueue_attrs *attrs)
4690 {
4691 enum wq_affn_scope scope;
4692 struct wq_pod_type *pt;
4693
4694 /* to synchronize access to wq_affn_dfl */
4695 lockdep_assert_held(&wq_pool_mutex);
4696
4697 if (attrs->affn_scope == WQ_AFFN_DFL)
4698 scope = wq_affn_dfl;
4699 else
4700 scope = attrs->affn_scope;
4701
4702 pt = &wq_pod_types[scope];
4703
4704 if (!WARN_ON_ONCE(attrs->affn_scope == WQ_AFFN_NR_TYPES) &&
4705 likely(pt->nr_pods))
4706 return pt;
4707
4708 /*
4709 * Before workqueue_init_topology(), only SYSTEM is available which is
4710 * initialized in workqueue_init_early().
4711 */
4712 pt = &wq_pod_types[WQ_AFFN_SYSTEM];
4713 BUG_ON(!pt->nr_pods);
4714 return pt;
4715 }
4716
4717 /**
4718 * init_worker_pool - initialize a newly zalloc'd worker_pool
4719 * @pool: worker_pool to initialize
4720 *
4721 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
4722 *
4723 * Return: 0 on success, -errno on failure. Even on failure, all fields
4724 * inside @pool proper are initialized and put_unbound_pool() can be called
4725 * on @pool safely to release it.
4726 */
init_worker_pool(struct worker_pool * pool)4727 static int init_worker_pool(struct worker_pool *pool)
4728 {
4729 raw_spin_lock_init(&pool->lock);
4730 pool->id = -1;
4731 pool->cpu = -1;
4732 pool->node = NUMA_NO_NODE;
4733 pool->flags |= POOL_DISASSOCIATED;
4734 pool->watchdog_ts = jiffies;
4735 INIT_LIST_HEAD(&pool->worklist);
4736 INIT_LIST_HEAD(&pool->idle_list);
4737 hash_init(pool->busy_hash);
4738
4739 timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE);
4740 INIT_WORK(&pool->idle_cull_work, idle_cull_fn);
4741
4742 timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0);
4743
4744 INIT_LIST_HEAD(&pool->workers);
4745
4746 ida_init(&pool->worker_ida);
4747 INIT_HLIST_NODE(&pool->hash_node);
4748 pool->refcnt = 1;
4749
4750 /* shouldn't fail above this point */
4751 pool->attrs = alloc_workqueue_attrs();
4752 if (!pool->attrs)
4753 return -ENOMEM;
4754
4755 wqattrs_clear_for_pool(pool->attrs);
4756
4757 return 0;
4758 }
4759
4760 #ifdef CONFIG_LOCKDEP
wq_init_lockdep(struct workqueue_struct * wq)4761 static void wq_init_lockdep(struct workqueue_struct *wq)
4762 {
4763 char *lock_name;
4764
4765 lockdep_register_key(&wq->key);
4766 lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name);
4767 if (!lock_name)
4768 lock_name = wq->name;
4769
4770 wq->lock_name = lock_name;
4771 lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0);
4772 }
4773
wq_unregister_lockdep(struct workqueue_struct * wq)4774 static void wq_unregister_lockdep(struct workqueue_struct *wq)
4775 {
4776 lockdep_unregister_key(&wq->key);
4777 }
4778
wq_free_lockdep(struct workqueue_struct * wq)4779 static void wq_free_lockdep(struct workqueue_struct *wq)
4780 {
4781 if (wq->lock_name != wq->name)
4782 kfree(wq->lock_name);
4783 }
4784 #else
wq_init_lockdep(struct workqueue_struct * wq)4785 static void wq_init_lockdep(struct workqueue_struct *wq)
4786 {
4787 }
4788
wq_unregister_lockdep(struct workqueue_struct * wq)4789 static void wq_unregister_lockdep(struct workqueue_struct *wq)
4790 {
4791 }
4792
wq_free_lockdep(struct workqueue_struct * wq)4793 static void wq_free_lockdep(struct workqueue_struct *wq)
4794 {
4795 }
4796 #endif
4797
free_node_nr_active(struct wq_node_nr_active ** nna_ar)4798 static void free_node_nr_active(struct wq_node_nr_active **nna_ar)
4799 {
4800 int node;
4801
4802 for_each_node(node) {
4803 kfree(nna_ar[node]);
4804 nna_ar[node] = NULL;
4805 }
4806
4807 kfree(nna_ar[nr_node_ids]);
4808 nna_ar[nr_node_ids] = NULL;
4809 }
4810
init_node_nr_active(struct wq_node_nr_active * nna)4811 static void init_node_nr_active(struct wq_node_nr_active *nna)
4812 {
4813 nna->max = WQ_DFL_MIN_ACTIVE;
4814 atomic_set(&nna->nr, 0);
4815 raw_spin_lock_init(&nna->lock);
4816 INIT_LIST_HEAD(&nna->pending_pwqs);
4817 }
4818
4819 /*
4820 * Each node's nr_active counter will be accessed mostly from its own node and
4821 * should be allocated in the node.
4822 */
alloc_node_nr_active(struct wq_node_nr_active ** nna_ar)4823 static int alloc_node_nr_active(struct wq_node_nr_active **nna_ar)
4824 {
4825 struct wq_node_nr_active *nna;
4826 int node;
4827
4828 for_each_node(node) {
4829 nna = kzalloc_node(sizeof(*nna), GFP_KERNEL, node);
4830 if (!nna)
4831 goto err_free;
4832 init_node_nr_active(nna);
4833 nna_ar[node] = nna;
4834 }
4835
4836 /* [nr_node_ids] is used as the fallback */
4837 nna = kzalloc_node(sizeof(*nna), GFP_KERNEL, NUMA_NO_NODE);
4838 if (!nna)
4839 goto err_free;
4840 init_node_nr_active(nna);
4841 nna_ar[nr_node_ids] = nna;
4842
4843 return 0;
4844
4845 err_free:
4846 free_node_nr_active(nna_ar);
4847 return -ENOMEM;
4848 }
4849
rcu_free_wq(struct rcu_head * rcu)4850 static void rcu_free_wq(struct rcu_head *rcu)
4851 {
4852 struct workqueue_struct *wq =
4853 container_of(rcu, struct workqueue_struct, rcu);
4854
4855 if (wq->flags & WQ_UNBOUND)
4856 free_node_nr_active(wq->node_nr_active);
4857
4858 wq_free_lockdep(wq);
4859 free_percpu(wq->cpu_pwq);
4860 free_workqueue_attrs(wq->unbound_attrs);
4861 kfree(wq);
4862 }
4863
rcu_free_pool(struct rcu_head * rcu)4864 static void rcu_free_pool(struct rcu_head *rcu)
4865 {
4866 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
4867
4868 ida_destroy(&pool->worker_ida);
4869 free_workqueue_attrs(pool->attrs);
4870 kfree(pool);
4871 }
4872
4873 /**
4874 * put_unbound_pool - put a worker_pool
4875 * @pool: worker_pool to put
4876 *
4877 * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU
4878 * safe manner. get_unbound_pool() calls this function on its failure path
4879 * and this function should be able to release pools which went through,
4880 * successfully or not, init_worker_pool().
4881 *
4882 * Should be called with wq_pool_mutex held.
4883 */
put_unbound_pool(struct worker_pool * pool)4884 static void put_unbound_pool(struct worker_pool *pool)
4885 {
4886 struct worker *worker;
4887 LIST_HEAD(cull_list);
4888
4889 lockdep_assert_held(&wq_pool_mutex);
4890
4891 if (--pool->refcnt)
4892 return;
4893
4894 /* sanity checks */
4895 if (WARN_ON(!(pool->cpu < 0)) ||
4896 WARN_ON(!list_empty(&pool->worklist)))
4897 return;
4898
4899 /* release id and unhash */
4900 if (pool->id >= 0)
4901 idr_remove(&worker_pool_idr, pool->id);
4902 hash_del(&pool->hash_node);
4903
4904 /*
4905 * Become the manager and destroy all workers. This prevents
4906 * @pool's workers from blocking on attach_mutex. We're the last
4907 * manager and @pool gets freed with the flag set.
4908 *
4909 * Having a concurrent manager is quite unlikely to happen as we can
4910 * only get here with
4911 * pwq->refcnt == pool->refcnt == 0
4912 * which implies no work queued to the pool, which implies no worker can
4913 * become the manager. However a worker could have taken the role of
4914 * manager before the refcnts dropped to 0, since maybe_create_worker()
4915 * drops pool->lock
4916 */
4917 while (true) {
4918 rcuwait_wait_event(&manager_wait,
4919 !(pool->flags & POOL_MANAGER_ACTIVE),
4920 TASK_UNINTERRUPTIBLE);
4921
4922 mutex_lock(&wq_pool_attach_mutex);
4923 raw_spin_lock_irq(&pool->lock);
4924 if (!(pool->flags & POOL_MANAGER_ACTIVE)) {
4925 pool->flags |= POOL_MANAGER_ACTIVE;
4926 break;
4927 }
4928 raw_spin_unlock_irq(&pool->lock);
4929 mutex_unlock(&wq_pool_attach_mutex);
4930 }
4931
4932 while ((worker = first_idle_worker(pool)))
4933 set_worker_dying(worker, &cull_list);
4934 WARN_ON(pool->nr_workers || pool->nr_idle);
4935 raw_spin_unlock_irq(&pool->lock);
4936
4937 detach_dying_workers(&cull_list);
4938
4939 mutex_unlock(&wq_pool_attach_mutex);
4940
4941 reap_dying_workers(&cull_list);
4942
4943 /* shut down the timers */
4944 del_timer_sync(&pool->idle_timer);
4945 cancel_work_sync(&pool->idle_cull_work);
4946 del_timer_sync(&pool->mayday_timer);
4947
4948 /* RCU protected to allow dereferences from get_work_pool() */
4949 call_rcu(&pool->rcu, rcu_free_pool);
4950 }
4951
4952 /**
4953 * get_unbound_pool - get a worker_pool with the specified attributes
4954 * @attrs: the attributes of the worker_pool to get
4955 *
4956 * Obtain a worker_pool which has the same attributes as @attrs, bump the
4957 * reference count and return it. If there already is a matching
4958 * worker_pool, it will be used; otherwise, this function attempts to
4959 * create a new one.
4960 *
4961 * Should be called with wq_pool_mutex held.
4962 *
4963 * Return: On success, a worker_pool with the same attributes as @attrs.
4964 * On failure, %NULL.
4965 */
get_unbound_pool(const struct workqueue_attrs * attrs)4966 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
4967 {
4968 struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_NUMA];
4969 u32 hash = wqattrs_hash(attrs);
4970 struct worker_pool *pool;
4971 int pod, node = NUMA_NO_NODE;
4972
4973 lockdep_assert_held(&wq_pool_mutex);
4974
4975 /* do we already have a matching pool? */
4976 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
4977 if (wqattrs_equal(pool->attrs, attrs)) {
4978 pool->refcnt++;
4979 return pool;
4980 }
4981 }
4982
4983 /* If __pod_cpumask is contained inside a NUMA pod, that's our node */
4984 for (pod = 0; pod < pt->nr_pods; pod++) {
4985 if (cpumask_subset(attrs->__pod_cpumask, pt->pod_cpus[pod])) {
4986 node = pt->pod_node[pod];
4987 break;
4988 }
4989 }
4990
4991 /* nope, create a new one */
4992 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, node);
4993 if (!pool || init_worker_pool(pool) < 0)
4994 goto fail;
4995
4996 pool->node = node;
4997 copy_workqueue_attrs(pool->attrs, attrs);
4998 wqattrs_clear_for_pool(pool->attrs);
4999
5000 if (worker_pool_assign_id(pool) < 0)
5001 goto fail;
5002
5003 /* create and start the initial worker */
5004 if (wq_online && !create_worker(pool))
5005 goto fail;
5006
5007 /* install */
5008 hash_add(unbound_pool_hash, &pool->hash_node, hash);
5009
5010 return pool;
5011 fail:
5012 if (pool)
5013 put_unbound_pool(pool);
5014 return NULL;
5015 }
5016
5017 /*
5018 * Scheduled on pwq_release_worker by put_pwq() when an unbound pwq hits zero
5019 * refcnt and needs to be destroyed.
5020 */
pwq_release_workfn(struct kthread_work * work)5021 static void pwq_release_workfn(struct kthread_work *work)
5022 {
5023 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
5024 release_work);
5025 struct workqueue_struct *wq = pwq->wq;
5026 struct worker_pool *pool = pwq->pool;
5027 bool is_last = false;
5028
5029 /*
5030 * When @pwq is not linked, it doesn't hold any reference to the
5031 * @wq, and @wq is invalid to access.
5032 */
5033 if (!list_empty(&pwq->pwqs_node)) {
5034 mutex_lock(&wq->mutex);
5035 list_del_rcu(&pwq->pwqs_node);
5036 is_last = list_empty(&wq->pwqs);
5037
5038 /*
5039 * For ordered workqueue with a plugged dfl_pwq, restart it now.
5040 */
5041 if (!is_last && (wq->flags & __WQ_ORDERED))
5042 unplug_oldest_pwq(wq);
5043
5044 mutex_unlock(&wq->mutex);
5045 }
5046
5047 if (wq->flags & WQ_UNBOUND) {
5048 mutex_lock(&wq_pool_mutex);
5049 put_unbound_pool(pool);
5050 mutex_unlock(&wq_pool_mutex);
5051 }
5052
5053 if (!list_empty(&pwq->pending_node)) {
5054 struct wq_node_nr_active *nna =
5055 wq_node_nr_active(pwq->wq, pwq->pool->node);
5056
5057 raw_spin_lock_irq(&nna->lock);
5058 list_del_init(&pwq->pending_node);
5059 raw_spin_unlock_irq(&nna->lock);
5060 }
5061
5062 kfree_rcu(pwq, rcu);
5063
5064 /*
5065 * If we're the last pwq going away, @wq is already dead and no one
5066 * is gonna access it anymore. Schedule RCU free.
5067 */
5068 if (is_last) {
5069 wq_unregister_lockdep(wq);
5070 call_rcu(&wq->rcu, rcu_free_wq);
5071 }
5072 }
5073
5074 /* initialize newly allocated @pwq which is associated with @wq and @pool */
init_pwq(struct pool_workqueue * pwq,struct workqueue_struct * wq,struct worker_pool * pool)5075 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
5076 struct worker_pool *pool)
5077 {
5078 BUG_ON((unsigned long)pwq & ~WORK_STRUCT_PWQ_MASK);
5079
5080 memset(pwq, 0, sizeof(*pwq));
5081
5082 pwq->pool = pool;
5083 pwq->wq = wq;
5084 pwq->flush_color = -1;
5085 pwq->refcnt = 1;
5086 INIT_LIST_HEAD(&pwq->inactive_works);
5087 INIT_LIST_HEAD(&pwq->pending_node);
5088 INIT_LIST_HEAD(&pwq->pwqs_node);
5089 INIT_LIST_HEAD(&pwq->mayday_node);
5090 kthread_init_work(&pwq->release_work, pwq_release_workfn);
5091 }
5092
5093 /* sync @pwq with the current state of its associated wq and link it */
link_pwq(struct pool_workqueue * pwq)5094 static void link_pwq(struct pool_workqueue *pwq)
5095 {
5096 struct workqueue_struct *wq = pwq->wq;
5097
5098 lockdep_assert_held(&wq->mutex);
5099
5100 /* may be called multiple times, ignore if already linked */
5101 if (!list_empty(&pwq->pwqs_node))
5102 return;
5103
5104 /* set the matching work_color */
5105 pwq->work_color = wq->work_color;
5106
5107 /* link in @pwq */
5108 list_add_tail_rcu(&pwq->pwqs_node, &wq->pwqs);
5109 }
5110
5111 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */
alloc_unbound_pwq(struct workqueue_struct * wq,const struct workqueue_attrs * attrs)5112 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
5113 const struct workqueue_attrs *attrs)
5114 {
5115 struct worker_pool *pool;
5116 struct pool_workqueue *pwq;
5117
5118 lockdep_assert_held(&wq_pool_mutex);
5119
5120 pool = get_unbound_pool(attrs);
5121 if (!pool)
5122 return NULL;
5123
5124 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
5125 if (!pwq) {
5126 put_unbound_pool(pool);
5127 return NULL;
5128 }
5129
5130 init_pwq(pwq, wq, pool);
5131 return pwq;
5132 }
5133
apply_wqattrs_lock(void)5134 static void apply_wqattrs_lock(void)
5135 {
5136 mutex_lock(&wq_pool_mutex);
5137 }
5138
apply_wqattrs_unlock(void)5139 static void apply_wqattrs_unlock(void)
5140 {
5141 mutex_unlock(&wq_pool_mutex);
5142 }
5143
5144 /**
5145 * wq_calc_pod_cpumask - calculate a wq_attrs' cpumask for a pod
5146 * @attrs: the wq_attrs of the default pwq of the target workqueue
5147 * @cpu: the target CPU
5148 *
5149 * Calculate the cpumask a workqueue with @attrs should use on @pod.
5150 * The result is stored in @attrs->__pod_cpumask.
5151 *
5152 * If pod affinity is not enabled, @attrs->cpumask is always used. If enabled
5153 * and @pod has online CPUs requested by @attrs, the returned cpumask is the
5154 * intersection of the possible CPUs of @pod and @attrs->cpumask.
5155 *
5156 * The caller is responsible for ensuring that the cpumask of @pod stays stable.
5157 */
wq_calc_pod_cpumask(struct workqueue_attrs * attrs,int cpu)5158 static void wq_calc_pod_cpumask(struct workqueue_attrs *attrs, int cpu)
5159 {
5160 const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
5161 int pod = pt->cpu_pod[cpu];
5162
5163 /* calculate possible CPUs in @pod that @attrs wants */
5164 cpumask_and(attrs->__pod_cpumask, pt->pod_cpus[pod], attrs->cpumask);
5165 /* does @pod have any online CPUs @attrs wants? */
5166 if (!cpumask_intersects(attrs->__pod_cpumask, wq_online_cpumask)) {
5167 cpumask_copy(attrs->__pod_cpumask, attrs->cpumask);
5168 return;
5169 }
5170 }
5171
5172 /* install @pwq into @wq and return the old pwq, @cpu < 0 for dfl_pwq */
install_unbound_pwq(struct workqueue_struct * wq,int cpu,struct pool_workqueue * pwq)5173 static struct pool_workqueue *install_unbound_pwq(struct workqueue_struct *wq,
5174 int cpu, struct pool_workqueue *pwq)
5175 {
5176 struct pool_workqueue __rcu **slot = unbound_pwq_slot(wq, cpu);
5177 struct pool_workqueue *old_pwq;
5178
5179 lockdep_assert_held(&wq_pool_mutex);
5180 lockdep_assert_held(&wq->mutex);
5181
5182 /* link_pwq() can handle duplicate calls */
5183 link_pwq(pwq);
5184
5185 old_pwq = rcu_access_pointer(*slot);
5186 rcu_assign_pointer(*slot, pwq);
5187 return old_pwq;
5188 }
5189
5190 /* context to store the prepared attrs & pwqs before applying */
5191 struct apply_wqattrs_ctx {
5192 struct workqueue_struct *wq; /* target workqueue */
5193 struct workqueue_attrs *attrs; /* attrs to apply */
5194 struct list_head list; /* queued for batching commit */
5195 struct pool_workqueue *dfl_pwq;
5196 struct pool_workqueue *pwq_tbl[];
5197 };
5198
5199 /* free the resources after success or abort */
apply_wqattrs_cleanup(struct apply_wqattrs_ctx * ctx)5200 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
5201 {
5202 if (ctx) {
5203 int cpu;
5204
5205 for_each_possible_cpu(cpu)
5206 put_pwq_unlocked(ctx->pwq_tbl[cpu]);
5207 put_pwq_unlocked(ctx->dfl_pwq);
5208
5209 free_workqueue_attrs(ctx->attrs);
5210
5211 kfree(ctx);
5212 }
5213 }
5214
5215 /* allocate the attrs and pwqs for later installation */
5216 static struct apply_wqattrs_ctx *
apply_wqattrs_prepare(struct workqueue_struct * wq,const struct workqueue_attrs * attrs,const cpumask_var_t unbound_cpumask)5217 apply_wqattrs_prepare(struct workqueue_struct *wq,
5218 const struct workqueue_attrs *attrs,
5219 const cpumask_var_t unbound_cpumask)
5220 {
5221 struct apply_wqattrs_ctx *ctx;
5222 struct workqueue_attrs *new_attrs;
5223 int cpu;
5224
5225 lockdep_assert_held(&wq_pool_mutex);
5226
5227 if (WARN_ON(attrs->affn_scope < 0 ||
5228 attrs->affn_scope >= WQ_AFFN_NR_TYPES))
5229 return ERR_PTR(-EINVAL);
5230
5231 ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_cpu_ids), GFP_KERNEL);
5232
5233 new_attrs = alloc_workqueue_attrs();
5234 if (!ctx || !new_attrs)
5235 goto out_free;
5236
5237 /*
5238 * If something goes wrong during CPU up/down, we'll fall back to
5239 * the default pwq covering whole @attrs->cpumask. Always create
5240 * it even if we don't use it immediately.
5241 */
5242 copy_workqueue_attrs(new_attrs, attrs);
5243 wqattrs_actualize_cpumask(new_attrs, unbound_cpumask);
5244 cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask);
5245 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
5246 if (!ctx->dfl_pwq)
5247 goto out_free;
5248
5249 for_each_possible_cpu(cpu) {
5250 if (new_attrs->ordered) {
5251 ctx->dfl_pwq->refcnt++;
5252 ctx->pwq_tbl[cpu] = ctx->dfl_pwq;
5253 } else {
5254 wq_calc_pod_cpumask(new_attrs, cpu);
5255 ctx->pwq_tbl[cpu] = alloc_unbound_pwq(wq, new_attrs);
5256 if (!ctx->pwq_tbl[cpu])
5257 goto out_free;
5258 }
5259 }
5260
5261 /* save the user configured attrs and sanitize it. */
5262 copy_workqueue_attrs(new_attrs, attrs);
5263 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
5264 cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask);
5265 ctx->attrs = new_attrs;
5266
5267 /*
5268 * For initialized ordered workqueues, there should only be one pwq
5269 * (dfl_pwq). Set the plugged flag of ctx->dfl_pwq to suspend execution
5270 * of newly queued work items until execution of older work items in
5271 * the old pwq's have completed.
5272 */
5273 if ((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs))
5274 ctx->dfl_pwq->plugged = true;
5275
5276 ctx->wq = wq;
5277 return ctx;
5278
5279 out_free:
5280 free_workqueue_attrs(new_attrs);
5281 apply_wqattrs_cleanup(ctx);
5282 return ERR_PTR(-ENOMEM);
5283 }
5284
5285 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
apply_wqattrs_commit(struct apply_wqattrs_ctx * ctx)5286 static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
5287 {
5288 int cpu;
5289
5290 /* all pwqs have been created successfully, let's install'em */
5291 mutex_lock(&ctx->wq->mutex);
5292
5293 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
5294
5295 /* save the previous pwqs and install the new ones */
5296 for_each_possible_cpu(cpu)
5297 ctx->pwq_tbl[cpu] = install_unbound_pwq(ctx->wq, cpu,
5298 ctx->pwq_tbl[cpu]);
5299 ctx->dfl_pwq = install_unbound_pwq(ctx->wq, -1, ctx->dfl_pwq);
5300
5301 /* update node_nr_active->max */
5302 wq_update_node_max_active(ctx->wq, -1);
5303
5304 /* rescuer needs to respect wq cpumask changes */
5305 if (ctx->wq->rescuer)
5306 set_cpus_allowed_ptr(ctx->wq->rescuer->task,
5307 unbound_effective_cpumask(ctx->wq));
5308
5309 mutex_unlock(&ctx->wq->mutex);
5310 }
5311
apply_workqueue_attrs_locked(struct workqueue_struct * wq,const struct workqueue_attrs * attrs)5312 static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
5313 const struct workqueue_attrs *attrs)
5314 {
5315 struct apply_wqattrs_ctx *ctx;
5316
5317 /* only unbound workqueues can change attributes */
5318 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
5319 return -EINVAL;
5320
5321 ctx = apply_wqattrs_prepare(wq, attrs, wq_unbound_cpumask);
5322 if (IS_ERR(ctx))
5323 return PTR_ERR(ctx);
5324
5325 /* the ctx has been prepared successfully, let's commit it */
5326 apply_wqattrs_commit(ctx);
5327 apply_wqattrs_cleanup(ctx);
5328
5329 return 0;
5330 }
5331
5332 /**
5333 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
5334 * @wq: the target workqueue
5335 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
5336 *
5337 * Apply @attrs to an unbound workqueue @wq. Unless disabled, this function maps
5338 * a separate pwq to each CPU pod with possibles CPUs in @attrs->cpumask so that
5339 * work items are affine to the pod it was issued on. Older pwqs are released as
5340 * in-flight work items finish. Note that a work item which repeatedly requeues
5341 * itself back-to-back will stay on its current pwq.
5342 *
5343 * Performs GFP_KERNEL allocations.
5344 *
5345 * Return: 0 on success and -errno on failure.
5346 */
apply_workqueue_attrs(struct workqueue_struct * wq,const struct workqueue_attrs * attrs)5347 int apply_workqueue_attrs(struct workqueue_struct *wq,
5348 const struct workqueue_attrs *attrs)
5349 {
5350 int ret;
5351
5352 mutex_lock(&wq_pool_mutex);
5353 ret = apply_workqueue_attrs_locked(wq, attrs);
5354 mutex_unlock(&wq_pool_mutex);
5355
5356 return ret;
5357 }
5358
5359 /**
5360 * unbound_wq_update_pwq - update a pwq slot for CPU hot[un]plug
5361 * @wq: the target workqueue
5362 * @cpu: the CPU to update the pwq slot for
5363 *
5364 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
5365 * %CPU_DOWN_FAILED. @cpu is in the same pod of the CPU being hot[un]plugged.
5366 *
5367 *
5368 * If pod affinity can't be adjusted due to memory allocation failure, it falls
5369 * back to @wq->dfl_pwq which may not be optimal but is always correct.
5370 *
5371 * Note that when the last allowed CPU of a pod goes offline for a workqueue
5372 * with a cpumask spanning multiple pods, the workers which were already
5373 * executing the work items for the workqueue will lose their CPU affinity and
5374 * may execute on any CPU. This is similar to how per-cpu workqueues behave on
5375 * CPU_DOWN. If a workqueue user wants strict affinity, it's the user's
5376 * responsibility to flush the work item from CPU_DOWN_PREPARE.
5377 */
unbound_wq_update_pwq(struct workqueue_struct * wq,int cpu)5378 static void unbound_wq_update_pwq(struct workqueue_struct *wq, int cpu)
5379 {
5380 struct pool_workqueue *old_pwq = NULL, *pwq;
5381 struct workqueue_attrs *target_attrs;
5382
5383 lockdep_assert_held(&wq_pool_mutex);
5384
5385 if (!(wq->flags & WQ_UNBOUND) || wq->unbound_attrs->ordered)
5386 return;
5387
5388 /*
5389 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
5390 * Let's use a preallocated one. The following buf is protected by
5391 * CPU hotplug exclusion.
5392 */
5393 target_attrs = unbound_wq_update_pwq_attrs_buf;
5394
5395 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
5396 wqattrs_actualize_cpumask(target_attrs, wq_unbound_cpumask);
5397
5398 /* nothing to do if the target cpumask matches the current pwq */
5399 wq_calc_pod_cpumask(target_attrs, cpu);
5400 if (wqattrs_equal(target_attrs, unbound_pwq(wq, cpu)->pool->attrs))
5401 return;
5402
5403 /* create a new pwq */
5404 pwq = alloc_unbound_pwq(wq, target_attrs);
5405 if (!pwq) {
5406 pr_warn("workqueue: allocation failed while updating CPU pod affinity of \"%s\"\n",
5407 wq->name);
5408 goto use_dfl_pwq;
5409 }
5410
5411 /* Install the new pwq. */
5412 mutex_lock(&wq->mutex);
5413 old_pwq = install_unbound_pwq(wq, cpu, pwq);
5414 goto out_unlock;
5415
5416 use_dfl_pwq:
5417 mutex_lock(&wq->mutex);
5418 pwq = unbound_pwq(wq, -1);
5419 raw_spin_lock_irq(&pwq->pool->lock);
5420 get_pwq(pwq);
5421 raw_spin_unlock_irq(&pwq->pool->lock);
5422 old_pwq = install_unbound_pwq(wq, cpu, pwq);
5423 out_unlock:
5424 mutex_unlock(&wq->mutex);
5425 put_pwq_unlocked(old_pwq);
5426 }
5427
alloc_and_link_pwqs(struct workqueue_struct * wq)5428 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
5429 {
5430 bool highpri = wq->flags & WQ_HIGHPRI;
5431 int cpu, ret;
5432
5433 lockdep_assert_held(&wq_pool_mutex);
5434
5435 wq->cpu_pwq = alloc_percpu(struct pool_workqueue *);
5436 if (!wq->cpu_pwq)
5437 goto enomem;
5438
5439 if (!(wq->flags & WQ_UNBOUND)) {
5440 struct worker_pool __percpu *pools;
5441
5442 if (wq->flags & WQ_BH)
5443 pools = bh_worker_pools;
5444 else
5445 pools = cpu_worker_pools;
5446
5447 for_each_possible_cpu(cpu) {
5448 struct pool_workqueue **pwq_p;
5449 struct worker_pool *pool;
5450
5451 pool = &(per_cpu_ptr(pools, cpu)[highpri]);
5452 pwq_p = per_cpu_ptr(wq->cpu_pwq, cpu);
5453
5454 *pwq_p = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL,
5455 pool->node);
5456 if (!*pwq_p)
5457 goto enomem;
5458
5459 init_pwq(*pwq_p, wq, pool);
5460
5461 mutex_lock(&wq->mutex);
5462 link_pwq(*pwq_p);
5463 mutex_unlock(&wq->mutex);
5464 }
5465 return 0;
5466 }
5467
5468 if (wq->flags & __WQ_ORDERED) {
5469 struct pool_workqueue *dfl_pwq;
5470
5471 ret = apply_workqueue_attrs_locked(wq, ordered_wq_attrs[highpri]);
5472 /* there should only be single pwq for ordering guarantee */
5473 dfl_pwq = rcu_access_pointer(wq->dfl_pwq);
5474 WARN(!ret && (wq->pwqs.next != &dfl_pwq->pwqs_node ||
5475 wq->pwqs.prev != &dfl_pwq->pwqs_node),
5476 "ordering guarantee broken for workqueue %s\n", wq->name);
5477 } else {
5478 ret = apply_workqueue_attrs_locked(wq, unbound_std_wq_attrs[highpri]);
5479 }
5480
5481 return ret;
5482
5483 enomem:
5484 if (wq->cpu_pwq) {
5485 for_each_possible_cpu(cpu) {
5486 struct pool_workqueue *pwq = *per_cpu_ptr(wq->cpu_pwq, cpu);
5487
5488 if (pwq)
5489 kmem_cache_free(pwq_cache, pwq);
5490 }
5491 free_percpu(wq->cpu_pwq);
5492 wq->cpu_pwq = NULL;
5493 }
5494 return -ENOMEM;
5495 }
5496
wq_clamp_max_active(int max_active,unsigned int flags,const char * name)5497 static int wq_clamp_max_active(int max_active, unsigned int flags,
5498 const char *name)
5499 {
5500 if (max_active < 1 || max_active > WQ_MAX_ACTIVE)
5501 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
5502 max_active, name, 1, WQ_MAX_ACTIVE);
5503
5504 return clamp_val(max_active, 1, WQ_MAX_ACTIVE);
5505 }
5506
5507 /*
5508 * Workqueues which may be used during memory reclaim should have a rescuer
5509 * to guarantee forward progress.
5510 */
init_rescuer(struct workqueue_struct * wq)5511 static int init_rescuer(struct workqueue_struct *wq)
5512 {
5513 struct worker *rescuer;
5514 char id_buf[WORKER_ID_LEN];
5515 int ret;
5516
5517 lockdep_assert_held(&wq_pool_mutex);
5518
5519 if (!(wq->flags & WQ_MEM_RECLAIM))
5520 return 0;
5521
5522 rescuer = alloc_worker(NUMA_NO_NODE);
5523 if (!rescuer) {
5524 pr_err("workqueue: Failed to allocate a rescuer for wq \"%s\"\n",
5525 wq->name);
5526 return -ENOMEM;
5527 }
5528
5529 rescuer->rescue_wq = wq;
5530 format_worker_id(id_buf, sizeof(id_buf), rescuer, NULL);
5531
5532 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", id_buf);
5533 if (IS_ERR(rescuer->task)) {
5534 ret = PTR_ERR(rescuer->task);
5535 pr_err("workqueue: Failed to create a rescuer kthread for wq \"%s\": %pe",
5536 wq->name, ERR_PTR(ret));
5537 kfree(rescuer);
5538 return ret;
5539 }
5540
5541 wq->rescuer = rescuer;
5542 if (wq->flags & WQ_UNBOUND)
5543 kthread_bind_mask(rescuer->task, unbound_effective_cpumask(wq));
5544 else
5545 kthread_bind_mask(rescuer->task, cpu_possible_mask);
5546 wake_up_process(rescuer->task);
5547
5548 return 0;
5549 }
5550
5551 /**
5552 * wq_adjust_max_active - update a wq's max_active to the current setting
5553 * @wq: target workqueue
5554 *
5555 * If @wq isn't freezing, set @wq->max_active to the saved_max_active and
5556 * activate inactive work items accordingly. If @wq is freezing, clear
5557 * @wq->max_active to zero.
5558 */
wq_adjust_max_active(struct workqueue_struct * wq)5559 static void wq_adjust_max_active(struct workqueue_struct *wq)
5560 {
5561 bool activated;
5562 int new_max, new_min;
5563
5564 lockdep_assert_held(&wq->mutex);
5565
5566 if ((wq->flags & WQ_FREEZABLE) && workqueue_freezing) {
5567 new_max = 0;
5568 new_min = 0;
5569 } else {
5570 new_max = wq->saved_max_active;
5571 new_min = wq->saved_min_active;
5572 }
5573
5574 if (wq->max_active == new_max && wq->min_active == new_min)
5575 return;
5576
5577 /*
5578 * Update @wq->max/min_active and then kick inactive work items if more
5579 * active work items are allowed. This doesn't break work item ordering
5580 * because new work items are always queued behind existing inactive
5581 * work items if there are any.
5582 */
5583 WRITE_ONCE(wq->max_active, new_max);
5584 WRITE_ONCE(wq->min_active, new_min);
5585
5586 if (wq->flags & WQ_UNBOUND)
5587 wq_update_node_max_active(wq, -1);
5588
5589 if (new_max == 0)
5590 return;
5591
5592 /*
5593 * Round-robin through pwq's activating the first inactive work item
5594 * until max_active is filled.
5595 */
5596 do {
5597 struct pool_workqueue *pwq;
5598
5599 activated = false;
5600 for_each_pwq(pwq, wq) {
5601 unsigned long irq_flags;
5602
5603 /* can be called during early boot w/ irq disabled */
5604 raw_spin_lock_irqsave(&pwq->pool->lock, irq_flags);
5605 if (pwq_activate_first_inactive(pwq, true)) {
5606 activated = true;
5607 kick_pool(pwq->pool);
5608 }
5609 raw_spin_unlock_irqrestore(&pwq->pool->lock, irq_flags);
5610 }
5611 } while (activated);
5612 }
5613
5614 __printf(1, 4)
alloc_workqueue(const char * fmt,unsigned int flags,int max_active,...)5615 struct workqueue_struct *alloc_workqueue(const char *fmt,
5616 unsigned int flags,
5617 int max_active, ...)
5618 {
5619 va_list args;
5620 struct workqueue_struct *wq;
5621 size_t wq_size;
5622 int name_len;
5623
5624 if (flags & WQ_BH) {
5625 if (WARN_ON_ONCE(flags & ~__WQ_BH_ALLOWS))
5626 return NULL;
5627 if (WARN_ON_ONCE(max_active))
5628 return NULL;
5629 }
5630
5631 /* see the comment above the definition of WQ_POWER_EFFICIENT */
5632 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
5633 flags |= WQ_UNBOUND;
5634
5635 /* allocate wq and format name */
5636 if (flags & WQ_UNBOUND)
5637 wq_size = struct_size(wq, node_nr_active, nr_node_ids + 1);
5638 else
5639 wq_size = sizeof(*wq);
5640
5641 wq = kzalloc(wq_size, GFP_KERNEL);
5642 if (!wq)
5643 return NULL;
5644
5645 if (flags & WQ_UNBOUND) {
5646 wq->unbound_attrs = alloc_workqueue_attrs();
5647 if (!wq->unbound_attrs)
5648 goto err_free_wq;
5649 }
5650
5651 va_start(args, max_active);
5652 name_len = vsnprintf(wq->name, sizeof(wq->name), fmt, args);
5653 va_end(args);
5654
5655 if (name_len >= WQ_NAME_LEN)
5656 pr_warn_once("workqueue: name exceeds WQ_NAME_LEN. Truncating to: %s\n",
5657 wq->name);
5658
5659 if (flags & WQ_BH) {
5660 /*
5661 * BH workqueues always share a single execution context per CPU
5662 * and don't impose any max_active limit.
5663 */
5664 max_active = INT_MAX;
5665 } else {
5666 max_active = max_active ?: WQ_DFL_ACTIVE;
5667 max_active = wq_clamp_max_active(max_active, flags, wq->name);
5668 }
5669
5670 /* init wq */
5671 wq->flags = flags;
5672 wq->max_active = max_active;
5673 wq->min_active = min(max_active, WQ_DFL_MIN_ACTIVE);
5674 wq->saved_max_active = wq->max_active;
5675 wq->saved_min_active = wq->min_active;
5676 mutex_init(&wq->mutex);
5677 atomic_set(&wq->nr_pwqs_to_flush, 0);
5678 INIT_LIST_HEAD(&wq->pwqs);
5679 INIT_LIST_HEAD(&wq->flusher_queue);
5680 INIT_LIST_HEAD(&wq->flusher_overflow);
5681 INIT_LIST_HEAD(&wq->maydays);
5682
5683 wq_init_lockdep(wq);
5684 INIT_LIST_HEAD(&wq->list);
5685
5686 if (flags & WQ_UNBOUND) {
5687 if (alloc_node_nr_active(wq->node_nr_active) < 0)
5688 goto err_unreg_lockdep;
5689 }
5690
5691 /*
5692 * wq_pool_mutex protects the workqueues list, allocations of PWQs,
5693 * and the global freeze state.
5694 */
5695 apply_wqattrs_lock();
5696
5697 if (alloc_and_link_pwqs(wq) < 0)
5698 goto err_unlock_free_node_nr_active;
5699
5700 mutex_lock(&wq->mutex);
5701 wq_adjust_max_active(wq);
5702 mutex_unlock(&wq->mutex);
5703
5704 list_add_tail_rcu(&wq->list, &workqueues);
5705
5706 if (wq_online && init_rescuer(wq) < 0)
5707 goto err_unlock_destroy;
5708
5709 apply_wqattrs_unlock();
5710
5711 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
5712 goto err_destroy;
5713
5714 return wq;
5715
5716 err_unlock_free_node_nr_active:
5717 apply_wqattrs_unlock();
5718 /*
5719 * Failed alloc_and_link_pwqs() may leave pending pwq->release_work,
5720 * flushing the pwq_release_worker ensures that the pwq_release_workfn()
5721 * completes before calling kfree(wq).
5722 */
5723 if (wq->flags & WQ_UNBOUND) {
5724 kthread_flush_worker(pwq_release_worker);
5725 free_node_nr_active(wq->node_nr_active);
5726 }
5727 err_unreg_lockdep:
5728 wq_unregister_lockdep(wq);
5729 wq_free_lockdep(wq);
5730 err_free_wq:
5731 free_workqueue_attrs(wq->unbound_attrs);
5732 kfree(wq);
5733 return NULL;
5734 err_unlock_destroy:
5735 apply_wqattrs_unlock();
5736 err_destroy:
5737 destroy_workqueue(wq);
5738 return NULL;
5739 }
5740 EXPORT_SYMBOL_GPL(alloc_workqueue);
5741
pwq_busy(struct pool_workqueue * pwq)5742 static bool pwq_busy(struct pool_workqueue *pwq)
5743 {
5744 int i;
5745
5746 for (i = 0; i < WORK_NR_COLORS; i++)
5747 if (pwq->nr_in_flight[i])
5748 return true;
5749
5750 if ((pwq != rcu_access_pointer(pwq->wq->dfl_pwq)) && (pwq->refcnt > 1))
5751 return true;
5752 if (!pwq_is_empty(pwq))
5753 return true;
5754
5755 return false;
5756 }
5757
5758 /**
5759 * destroy_workqueue - safely terminate a workqueue
5760 * @wq: target workqueue
5761 *
5762 * Safely destroy a workqueue. All work currently pending will be done first.
5763 */
destroy_workqueue(struct workqueue_struct * wq)5764 void destroy_workqueue(struct workqueue_struct *wq)
5765 {
5766 struct pool_workqueue *pwq;
5767 int cpu;
5768
5769 /*
5770 * Remove it from sysfs first so that sanity check failure doesn't
5771 * lead to sysfs name conflicts.
5772 */
5773 workqueue_sysfs_unregister(wq);
5774
5775 /* mark the workqueue destruction is in progress */
5776 mutex_lock(&wq->mutex);
5777 wq->flags |= __WQ_DESTROYING;
5778 mutex_unlock(&wq->mutex);
5779
5780 /* drain it before proceeding with destruction */
5781 drain_workqueue(wq);
5782
5783 /* kill rescuer, if sanity checks fail, leave it w/o rescuer */
5784 if (wq->rescuer) {
5785 struct worker *rescuer = wq->rescuer;
5786
5787 /* this prevents new queueing */
5788 raw_spin_lock_irq(&wq_mayday_lock);
5789 wq->rescuer = NULL;
5790 raw_spin_unlock_irq(&wq_mayday_lock);
5791
5792 /* rescuer will empty maydays list before exiting */
5793 kthread_stop(rescuer->task);
5794 kfree(rescuer);
5795 }
5796
5797 /*
5798 * Sanity checks - grab all the locks so that we wait for all
5799 * in-flight operations which may do put_pwq().
5800 */
5801 mutex_lock(&wq_pool_mutex);
5802 mutex_lock(&wq->mutex);
5803 for_each_pwq(pwq, wq) {
5804 raw_spin_lock_irq(&pwq->pool->lock);
5805 if (WARN_ON(pwq_busy(pwq))) {
5806 pr_warn("%s: %s has the following busy pwq\n",
5807 __func__, wq->name);
5808 show_pwq(pwq);
5809 raw_spin_unlock_irq(&pwq->pool->lock);
5810 mutex_unlock(&wq->mutex);
5811 mutex_unlock(&wq_pool_mutex);
5812 show_one_workqueue(wq);
5813 return;
5814 }
5815 raw_spin_unlock_irq(&pwq->pool->lock);
5816 }
5817 mutex_unlock(&wq->mutex);
5818
5819 /*
5820 * wq list is used to freeze wq, remove from list after
5821 * flushing is complete in case freeze races us.
5822 */
5823 list_del_rcu(&wq->list);
5824 mutex_unlock(&wq_pool_mutex);
5825
5826 /*
5827 * We're the sole accessor of @wq. Directly access cpu_pwq and dfl_pwq
5828 * to put the base refs. @wq will be auto-destroyed from the last
5829 * pwq_put. RCU read lock prevents @wq from going away from under us.
5830 */
5831 rcu_read_lock();
5832
5833 for_each_possible_cpu(cpu) {
5834 put_pwq_unlocked(unbound_pwq(wq, cpu));
5835 RCU_INIT_POINTER(*unbound_pwq_slot(wq, cpu), NULL);
5836 }
5837
5838 put_pwq_unlocked(unbound_pwq(wq, -1));
5839 RCU_INIT_POINTER(*unbound_pwq_slot(wq, -1), NULL);
5840
5841 rcu_read_unlock();
5842 }
5843 EXPORT_SYMBOL_GPL(destroy_workqueue);
5844
5845 /**
5846 * workqueue_set_max_active - adjust max_active of a workqueue
5847 * @wq: target workqueue
5848 * @max_active: new max_active value.
5849 *
5850 * Set max_active of @wq to @max_active. See the alloc_workqueue() function
5851 * comment.
5852 *
5853 * CONTEXT:
5854 * Don't call from IRQ context.
5855 */
workqueue_set_max_active(struct workqueue_struct * wq,int max_active)5856 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
5857 {
5858 /* max_active doesn't mean anything for BH workqueues */
5859 if (WARN_ON(wq->flags & WQ_BH))
5860 return;
5861 /* disallow meddling with max_active for ordered workqueues */
5862 if (WARN_ON(wq->flags & __WQ_ORDERED))
5863 return;
5864
5865 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
5866
5867 mutex_lock(&wq->mutex);
5868
5869 wq->saved_max_active = max_active;
5870 if (wq->flags & WQ_UNBOUND)
5871 wq->saved_min_active = min(wq->saved_min_active, max_active);
5872
5873 wq_adjust_max_active(wq);
5874
5875 mutex_unlock(&wq->mutex);
5876 }
5877 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
5878
5879 /**
5880 * workqueue_set_min_active - adjust min_active of an unbound workqueue
5881 * @wq: target unbound workqueue
5882 * @min_active: new min_active value
5883 *
5884 * Set min_active of an unbound workqueue. Unlike other types of workqueues, an
5885 * unbound workqueue is not guaranteed to be able to process max_active
5886 * interdependent work items. Instead, an unbound workqueue is guaranteed to be
5887 * able to process min_active number of interdependent work items which is
5888 * %WQ_DFL_MIN_ACTIVE by default.
5889 *
5890 * Use this function to adjust the min_active value between 0 and the current
5891 * max_active.
5892 */
workqueue_set_min_active(struct workqueue_struct * wq,int min_active)5893 void workqueue_set_min_active(struct workqueue_struct *wq, int min_active)
5894 {
5895 /* min_active is only meaningful for non-ordered unbound workqueues */
5896 if (WARN_ON((wq->flags & (WQ_BH | WQ_UNBOUND | __WQ_ORDERED)) !=
5897 WQ_UNBOUND))
5898 return;
5899
5900 mutex_lock(&wq->mutex);
5901 wq->saved_min_active = clamp(min_active, 0, wq->saved_max_active);
5902 wq_adjust_max_active(wq);
5903 mutex_unlock(&wq->mutex);
5904 }
5905
5906 /**
5907 * current_work - retrieve %current task's work struct
5908 *
5909 * Determine if %current task is a workqueue worker and what it's working on.
5910 * Useful to find out the context that the %current task is running in.
5911 *
5912 * Return: work struct if %current task is a workqueue worker, %NULL otherwise.
5913 */
current_work(void)5914 struct work_struct *current_work(void)
5915 {
5916 struct worker *worker = current_wq_worker();
5917
5918 return worker ? worker->current_work : NULL;
5919 }
5920 EXPORT_SYMBOL(current_work);
5921
5922 /**
5923 * current_is_workqueue_rescuer - is %current workqueue rescuer?
5924 *
5925 * Determine whether %current is a workqueue rescuer. Can be used from
5926 * work functions to determine whether it's being run off the rescuer task.
5927 *
5928 * Return: %true if %current is a workqueue rescuer. %false otherwise.
5929 */
current_is_workqueue_rescuer(void)5930 bool current_is_workqueue_rescuer(void)
5931 {
5932 struct worker *worker = current_wq_worker();
5933
5934 return worker && worker->rescue_wq;
5935 }
5936
5937 /**
5938 * workqueue_congested - test whether a workqueue is congested
5939 * @cpu: CPU in question
5940 * @wq: target workqueue
5941 *
5942 * Test whether @wq's cpu workqueue for @cpu is congested. There is
5943 * no synchronization around this function and the test result is
5944 * unreliable and only useful as advisory hints or for debugging.
5945 *
5946 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
5947 *
5948 * With the exception of ordered workqueues, all workqueues have per-cpu
5949 * pool_workqueues, each with its own congested state. A workqueue being
5950 * congested on one CPU doesn't mean that the workqueue is contested on any
5951 * other CPUs.
5952 *
5953 * Return:
5954 * %true if congested, %false otherwise.
5955 */
workqueue_congested(int cpu,struct workqueue_struct * wq)5956 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
5957 {
5958 struct pool_workqueue *pwq;
5959 bool ret;
5960
5961 rcu_read_lock();
5962 preempt_disable();
5963
5964 if (cpu == WORK_CPU_UNBOUND)
5965 cpu = smp_processor_id();
5966
5967 pwq = *per_cpu_ptr(wq->cpu_pwq, cpu);
5968 ret = !list_empty(&pwq->inactive_works);
5969
5970 preempt_enable();
5971 rcu_read_unlock();
5972
5973 return ret;
5974 }
5975 EXPORT_SYMBOL_GPL(workqueue_congested);
5976
5977 /**
5978 * work_busy - test whether a work is currently pending or running
5979 * @work: the work to be tested
5980 *
5981 * Test whether @work is currently pending or running. There is no
5982 * synchronization around this function and the test result is
5983 * unreliable and only useful as advisory hints or for debugging.
5984 *
5985 * Return:
5986 * OR'd bitmask of WORK_BUSY_* bits.
5987 */
work_busy(struct work_struct * work)5988 unsigned int work_busy(struct work_struct *work)
5989 {
5990 struct worker_pool *pool;
5991 unsigned long irq_flags;
5992 unsigned int ret = 0;
5993
5994 if (work_pending(work))
5995 ret |= WORK_BUSY_PENDING;
5996
5997 rcu_read_lock();
5998 pool = get_work_pool(work);
5999 if (pool) {
6000 raw_spin_lock_irqsave(&pool->lock, irq_flags);
6001 if (find_worker_executing_work(pool, work))
6002 ret |= WORK_BUSY_RUNNING;
6003 raw_spin_unlock_irqrestore(&pool->lock, irq_flags);
6004 }
6005 rcu_read_unlock();
6006
6007 return ret;
6008 }
6009 EXPORT_SYMBOL_GPL(work_busy);
6010
6011 /**
6012 * set_worker_desc - set description for the current work item
6013 * @fmt: printf-style format string
6014 * @...: arguments for the format string
6015 *
6016 * This function can be called by a running work function to describe what
6017 * the work item is about. If the worker task gets dumped, this
6018 * information will be printed out together to help debugging. The
6019 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
6020 */
set_worker_desc(const char * fmt,...)6021 void set_worker_desc(const char *fmt, ...)
6022 {
6023 struct worker *worker = current_wq_worker();
6024 va_list args;
6025
6026 if (worker) {
6027 va_start(args, fmt);
6028 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
6029 va_end(args);
6030 }
6031 }
6032 EXPORT_SYMBOL_GPL(set_worker_desc);
6033
6034 /**
6035 * print_worker_info - print out worker information and description
6036 * @log_lvl: the log level to use when printing
6037 * @task: target task
6038 *
6039 * If @task is a worker and currently executing a work item, print out the
6040 * name of the workqueue being serviced and worker description set with
6041 * set_worker_desc() by the currently executing work item.
6042 *
6043 * This function can be safely called on any task as long as the
6044 * task_struct itself is accessible. While safe, this function isn't
6045 * synchronized and may print out mixups or garbages of limited length.
6046 */
print_worker_info(const char * log_lvl,struct task_struct * task)6047 void print_worker_info(const char *log_lvl, struct task_struct *task)
6048 {
6049 work_func_t *fn = NULL;
6050 char name[WQ_NAME_LEN] = { };
6051 char desc[WORKER_DESC_LEN] = { };
6052 struct pool_workqueue *pwq = NULL;
6053 struct workqueue_struct *wq = NULL;
6054 struct worker *worker;
6055
6056 if (!(task->flags & PF_WQ_WORKER))
6057 return;
6058
6059 /*
6060 * This function is called without any synchronization and @task
6061 * could be in any state. Be careful with dereferences.
6062 */
6063 worker = kthread_probe_data(task);
6064
6065 /*
6066 * Carefully copy the associated workqueue's workfn, name and desc.
6067 * Keep the original last '\0' in case the original is garbage.
6068 */
6069 copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn));
6070 copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq));
6071 copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq));
6072 copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1);
6073 copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1);
6074
6075 if (fn || name[0] || desc[0]) {
6076 printk("%sWorkqueue: %s %ps", log_lvl, name, fn);
6077 if (strcmp(name, desc))
6078 pr_cont(" (%s)", desc);
6079 pr_cont("\n");
6080 }
6081 }
6082
pr_cont_pool_info(struct worker_pool * pool)6083 static void pr_cont_pool_info(struct worker_pool *pool)
6084 {
6085 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
6086 if (pool->node != NUMA_NO_NODE)
6087 pr_cont(" node=%d", pool->node);
6088 pr_cont(" flags=0x%x", pool->flags);
6089 if (pool->flags & POOL_BH)
6090 pr_cont(" bh%s",
6091 pool->attrs->nice == HIGHPRI_NICE_LEVEL ? "-hi" : "");
6092 else
6093 pr_cont(" nice=%d", pool->attrs->nice);
6094 }
6095
pr_cont_worker_id(struct worker * worker)6096 static void pr_cont_worker_id(struct worker *worker)
6097 {
6098 struct worker_pool *pool = worker->pool;
6099
6100 if (pool->flags & WQ_BH)
6101 pr_cont("bh%s",
6102 pool->attrs->nice == HIGHPRI_NICE_LEVEL ? "-hi" : "");
6103 else
6104 pr_cont("%d%s", task_pid_nr(worker->task),
6105 worker->rescue_wq ? "(RESCUER)" : "");
6106 }
6107
6108 struct pr_cont_work_struct {
6109 bool comma;
6110 work_func_t func;
6111 long ctr;
6112 };
6113
pr_cont_work_flush(bool comma,work_func_t func,struct pr_cont_work_struct * pcwsp)6114 static void pr_cont_work_flush(bool comma, work_func_t func, struct pr_cont_work_struct *pcwsp)
6115 {
6116 if (!pcwsp->ctr)
6117 goto out_record;
6118 if (func == pcwsp->func) {
6119 pcwsp->ctr++;
6120 return;
6121 }
6122 if (pcwsp->ctr == 1)
6123 pr_cont("%s %ps", pcwsp->comma ? "," : "", pcwsp->func);
6124 else
6125 pr_cont("%s %ld*%ps", pcwsp->comma ? "," : "", pcwsp->ctr, pcwsp->func);
6126 pcwsp->ctr = 0;
6127 out_record:
6128 if ((long)func == -1L)
6129 return;
6130 pcwsp->comma = comma;
6131 pcwsp->func = func;
6132 pcwsp->ctr = 1;
6133 }
6134
pr_cont_work(bool comma,struct work_struct * work,struct pr_cont_work_struct * pcwsp)6135 static void pr_cont_work(bool comma, struct work_struct *work, struct pr_cont_work_struct *pcwsp)
6136 {
6137 if (work->func == wq_barrier_func) {
6138 struct wq_barrier *barr;
6139
6140 barr = container_of(work, struct wq_barrier, work);
6141
6142 pr_cont_work_flush(comma, (work_func_t)-1, pcwsp);
6143 pr_cont("%s BAR(%d)", comma ? "," : "",
6144 task_pid_nr(barr->task));
6145 } else {
6146 if (!comma)
6147 pr_cont_work_flush(comma, (work_func_t)-1, pcwsp);
6148 pr_cont_work_flush(comma, work->func, pcwsp);
6149 }
6150 }
6151
show_pwq(struct pool_workqueue * pwq)6152 static void show_pwq(struct pool_workqueue *pwq)
6153 {
6154 struct pr_cont_work_struct pcws = { .ctr = 0, };
6155 struct worker_pool *pool = pwq->pool;
6156 struct work_struct *work;
6157 struct worker *worker;
6158 bool has_in_flight = false, has_pending = false;
6159 int bkt;
6160
6161 pr_info(" pwq %d:", pool->id);
6162 pr_cont_pool_info(pool);
6163
6164 pr_cont(" active=%d refcnt=%d%s\n",
6165 pwq->nr_active, pwq->refcnt,
6166 !list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
6167
6168 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
6169 if (worker->current_pwq == pwq) {
6170 has_in_flight = true;
6171 break;
6172 }
6173 }
6174 if (has_in_flight) {
6175 bool comma = false;
6176
6177 pr_info(" in-flight:");
6178 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
6179 if (worker->current_pwq != pwq)
6180 continue;
6181
6182 pr_cont(" %s", comma ? "," : "");
6183 pr_cont_worker_id(worker);
6184 pr_cont(":%ps", worker->current_func);
6185 list_for_each_entry(work, &worker->scheduled, entry)
6186 pr_cont_work(false, work, &pcws);
6187 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
6188 comma = true;
6189 }
6190 pr_cont("\n");
6191 }
6192
6193 list_for_each_entry(work, &pool->worklist, entry) {
6194 if (get_work_pwq(work) == pwq) {
6195 has_pending = true;
6196 break;
6197 }
6198 }
6199 if (has_pending) {
6200 bool comma = false;
6201
6202 pr_info(" pending:");
6203 list_for_each_entry(work, &pool->worklist, entry) {
6204 if (get_work_pwq(work) != pwq)
6205 continue;
6206
6207 pr_cont_work(comma, work, &pcws);
6208 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
6209 }
6210 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
6211 pr_cont("\n");
6212 }
6213
6214 if (!list_empty(&pwq->inactive_works)) {
6215 bool comma = false;
6216
6217 pr_info(" inactive:");
6218 list_for_each_entry(work, &pwq->inactive_works, entry) {
6219 pr_cont_work(comma, work, &pcws);
6220 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
6221 }
6222 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
6223 pr_cont("\n");
6224 }
6225 }
6226
6227 /**
6228 * show_one_workqueue - dump state of specified workqueue
6229 * @wq: workqueue whose state will be printed
6230 */
show_one_workqueue(struct workqueue_struct * wq)6231 void show_one_workqueue(struct workqueue_struct *wq)
6232 {
6233 struct pool_workqueue *pwq;
6234 bool idle = true;
6235 unsigned long irq_flags;
6236
6237 for_each_pwq(pwq, wq) {
6238 if (!pwq_is_empty(pwq)) {
6239 idle = false;
6240 break;
6241 }
6242 }
6243 if (idle) /* Nothing to print for idle workqueue */
6244 return;
6245
6246 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
6247
6248 for_each_pwq(pwq, wq) {
6249 raw_spin_lock_irqsave(&pwq->pool->lock, irq_flags);
6250 if (!pwq_is_empty(pwq)) {
6251 /*
6252 * Defer printing to avoid deadlocks in console
6253 * drivers that queue work while holding locks
6254 * also taken in their write paths.
6255 */
6256 printk_deferred_enter();
6257 show_pwq(pwq);
6258 printk_deferred_exit();
6259 }
6260 raw_spin_unlock_irqrestore(&pwq->pool->lock, irq_flags);
6261 /*
6262 * We could be printing a lot from atomic context, e.g.
6263 * sysrq-t -> show_all_workqueues(). Avoid triggering
6264 * hard lockup.
6265 */
6266 touch_nmi_watchdog();
6267 }
6268
6269 }
6270
6271 /**
6272 * show_one_worker_pool - dump state of specified worker pool
6273 * @pool: worker pool whose state will be printed
6274 */
show_one_worker_pool(struct worker_pool * pool)6275 static void show_one_worker_pool(struct worker_pool *pool)
6276 {
6277 struct worker *worker;
6278 bool first = true;
6279 unsigned long irq_flags;
6280 unsigned long hung = 0;
6281
6282 raw_spin_lock_irqsave(&pool->lock, irq_flags);
6283 if (pool->nr_workers == pool->nr_idle)
6284 goto next_pool;
6285
6286 /* How long the first pending work is waiting for a worker. */
6287 if (!list_empty(&pool->worklist))
6288 hung = jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000;
6289
6290 /*
6291 * Defer printing to avoid deadlocks in console drivers that
6292 * queue work while holding locks also taken in their write
6293 * paths.
6294 */
6295 printk_deferred_enter();
6296 pr_info("pool %d:", pool->id);
6297 pr_cont_pool_info(pool);
6298 pr_cont(" hung=%lus workers=%d", hung, pool->nr_workers);
6299 if (pool->manager)
6300 pr_cont(" manager: %d",
6301 task_pid_nr(pool->manager->task));
6302 list_for_each_entry(worker, &pool->idle_list, entry) {
6303 pr_cont(" %s", first ? "idle: " : "");
6304 pr_cont_worker_id(worker);
6305 first = false;
6306 }
6307 pr_cont("\n");
6308 printk_deferred_exit();
6309 next_pool:
6310 raw_spin_unlock_irqrestore(&pool->lock, irq_flags);
6311 /*
6312 * We could be printing a lot from atomic context, e.g.
6313 * sysrq-t -> show_all_workqueues(). Avoid triggering
6314 * hard lockup.
6315 */
6316 touch_nmi_watchdog();
6317
6318 }
6319
6320 /**
6321 * show_all_workqueues - dump workqueue state
6322 *
6323 * Called from a sysrq handler and prints out all busy workqueues and pools.
6324 */
show_all_workqueues(void)6325 void show_all_workqueues(void)
6326 {
6327 struct workqueue_struct *wq;
6328 struct worker_pool *pool;
6329 int pi;
6330
6331 rcu_read_lock();
6332
6333 pr_info("Showing busy workqueues and worker pools:\n");
6334
6335 list_for_each_entry_rcu(wq, &workqueues, list)
6336 show_one_workqueue(wq);
6337
6338 for_each_pool(pool, pi)
6339 show_one_worker_pool(pool);
6340
6341 rcu_read_unlock();
6342 }
6343
6344 /**
6345 * show_freezable_workqueues - dump freezable workqueue state
6346 *
6347 * Called from try_to_freeze_tasks() and prints out all freezable workqueues
6348 * still busy.
6349 */
show_freezable_workqueues(void)6350 void show_freezable_workqueues(void)
6351 {
6352 struct workqueue_struct *wq;
6353
6354 rcu_read_lock();
6355
6356 pr_info("Showing freezable workqueues that are still busy:\n");
6357
6358 list_for_each_entry_rcu(wq, &workqueues, list) {
6359 if (!(wq->flags & WQ_FREEZABLE))
6360 continue;
6361 show_one_workqueue(wq);
6362 }
6363
6364 rcu_read_unlock();
6365 }
6366
6367 /* used to show worker information through /proc/PID/{comm,stat,status} */
wq_worker_comm(char * buf,size_t size,struct task_struct * task)6368 void wq_worker_comm(char *buf, size_t size, struct task_struct *task)
6369 {
6370 /* stabilize PF_WQ_WORKER and worker pool association */
6371 mutex_lock(&wq_pool_attach_mutex);
6372
6373 if (task->flags & PF_WQ_WORKER) {
6374 struct worker *worker = kthread_data(task);
6375 struct worker_pool *pool = worker->pool;
6376 int off;
6377
6378 off = format_worker_id(buf, size, worker, pool);
6379
6380 if (pool) {
6381 raw_spin_lock_irq(&pool->lock);
6382 /*
6383 * ->desc tracks information (wq name or
6384 * set_worker_desc()) for the latest execution. If
6385 * current, prepend '+', otherwise '-'.
6386 */
6387 if (worker->desc[0] != '\0') {
6388 if (worker->current_work)
6389 scnprintf(buf + off, size - off, "+%s",
6390 worker->desc);
6391 else
6392 scnprintf(buf + off, size - off, "-%s",
6393 worker->desc);
6394 }
6395 raw_spin_unlock_irq(&pool->lock);
6396 }
6397 } else {
6398 strscpy(buf, task->comm, size);
6399 }
6400
6401 mutex_unlock(&wq_pool_attach_mutex);
6402 }
6403
6404 #ifdef CONFIG_SMP
6405
6406 /*
6407 * CPU hotplug.
6408 *
6409 * There are two challenges in supporting CPU hotplug. Firstly, there
6410 * are a lot of assumptions on strong associations among work, pwq and
6411 * pool which make migrating pending and scheduled works very
6412 * difficult to implement without impacting hot paths. Secondly,
6413 * worker pools serve mix of short, long and very long running works making
6414 * blocked draining impractical.
6415 *
6416 * This is solved by allowing the pools to be disassociated from the CPU
6417 * running as an unbound one and allowing it to be reattached later if the
6418 * cpu comes back online.
6419 */
6420
unbind_workers(int cpu)6421 static void unbind_workers(int cpu)
6422 {
6423 struct worker_pool *pool;
6424 struct worker *worker;
6425
6426 for_each_cpu_worker_pool(pool, cpu) {
6427 mutex_lock(&wq_pool_attach_mutex);
6428 raw_spin_lock_irq(&pool->lock);
6429
6430 /*
6431 * We've blocked all attach/detach operations. Make all workers
6432 * unbound and set DISASSOCIATED. Before this, all workers
6433 * must be on the cpu. After this, they may become diasporas.
6434 * And the preemption disabled section in their sched callbacks
6435 * are guaranteed to see WORKER_UNBOUND since the code here
6436 * is on the same cpu.
6437 */
6438 for_each_pool_worker(worker, pool)
6439 worker->flags |= WORKER_UNBOUND;
6440
6441 pool->flags |= POOL_DISASSOCIATED;
6442
6443 /*
6444 * The handling of nr_running in sched callbacks are disabled
6445 * now. Zap nr_running. After this, nr_running stays zero and
6446 * need_more_worker() and keep_working() are always true as
6447 * long as the worklist is not empty. This pool now behaves as
6448 * an unbound (in terms of concurrency management) pool which
6449 * are served by workers tied to the pool.
6450 */
6451 pool->nr_running = 0;
6452
6453 /*
6454 * With concurrency management just turned off, a busy
6455 * worker blocking could lead to lengthy stalls. Kick off
6456 * unbound chain execution of currently pending work items.
6457 */
6458 kick_pool(pool);
6459
6460 raw_spin_unlock_irq(&pool->lock);
6461
6462 for_each_pool_worker(worker, pool)
6463 unbind_worker(worker);
6464
6465 mutex_unlock(&wq_pool_attach_mutex);
6466 }
6467 }
6468
6469 /**
6470 * rebind_workers - rebind all workers of a pool to the associated CPU
6471 * @pool: pool of interest
6472 *
6473 * @pool->cpu is coming online. Rebind all workers to the CPU.
6474 */
rebind_workers(struct worker_pool * pool)6475 static void rebind_workers(struct worker_pool *pool)
6476 {
6477 struct worker *worker;
6478
6479 lockdep_assert_held(&wq_pool_attach_mutex);
6480
6481 /*
6482 * Restore CPU affinity of all workers. As all idle workers should
6483 * be on the run-queue of the associated CPU before any local
6484 * wake-ups for concurrency management happen, restore CPU affinity
6485 * of all workers first and then clear UNBOUND. As we're called
6486 * from CPU_ONLINE, the following shouldn't fail.
6487 */
6488 for_each_pool_worker(worker, pool) {
6489 kthread_set_per_cpu(worker->task, pool->cpu);
6490 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
6491 pool_allowed_cpus(pool)) < 0);
6492 }
6493
6494 raw_spin_lock_irq(&pool->lock);
6495
6496 pool->flags &= ~POOL_DISASSOCIATED;
6497
6498 for_each_pool_worker(worker, pool) {
6499 unsigned int worker_flags = worker->flags;
6500
6501 /*
6502 * We want to clear UNBOUND but can't directly call
6503 * worker_clr_flags() or adjust nr_running. Atomically
6504 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
6505 * @worker will clear REBOUND using worker_clr_flags() when
6506 * it initiates the next execution cycle thus restoring
6507 * concurrency management. Note that when or whether
6508 * @worker clears REBOUND doesn't affect correctness.
6509 *
6510 * WRITE_ONCE() is necessary because @worker->flags may be
6511 * tested without holding any lock in
6512 * wq_worker_running(). Without it, NOT_RUNNING test may
6513 * fail incorrectly leading to premature concurrency
6514 * management operations.
6515 */
6516 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
6517 worker_flags |= WORKER_REBOUND;
6518 worker_flags &= ~WORKER_UNBOUND;
6519 WRITE_ONCE(worker->flags, worker_flags);
6520 }
6521
6522 raw_spin_unlock_irq(&pool->lock);
6523 }
6524
6525 /**
6526 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
6527 * @pool: unbound pool of interest
6528 * @cpu: the CPU which is coming up
6529 *
6530 * An unbound pool may end up with a cpumask which doesn't have any online
6531 * CPUs. When a worker of such pool get scheduled, the scheduler resets
6532 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
6533 * online CPU before, cpus_allowed of all its workers should be restored.
6534 */
restore_unbound_workers_cpumask(struct worker_pool * pool,int cpu)6535 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
6536 {
6537 static cpumask_t cpumask;
6538 struct worker *worker;
6539
6540 lockdep_assert_held(&wq_pool_attach_mutex);
6541
6542 /* is @cpu allowed for @pool? */
6543 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
6544 return;
6545
6546 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
6547
6548 /* as we're called from CPU_ONLINE, the following shouldn't fail */
6549 for_each_pool_worker(worker, pool)
6550 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0);
6551 }
6552
workqueue_prepare_cpu(unsigned int cpu)6553 int workqueue_prepare_cpu(unsigned int cpu)
6554 {
6555 struct worker_pool *pool;
6556
6557 for_each_cpu_worker_pool(pool, cpu) {
6558 if (pool->nr_workers)
6559 continue;
6560 if (!create_worker(pool))
6561 return -ENOMEM;
6562 }
6563 return 0;
6564 }
6565
workqueue_online_cpu(unsigned int cpu)6566 int workqueue_online_cpu(unsigned int cpu)
6567 {
6568 struct worker_pool *pool;
6569 struct workqueue_struct *wq;
6570 int pi;
6571
6572 mutex_lock(&wq_pool_mutex);
6573
6574 cpumask_set_cpu(cpu, wq_online_cpumask);
6575
6576 for_each_pool(pool, pi) {
6577 /* BH pools aren't affected by hotplug */
6578 if (pool->flags & POOL_BH)
6579 continue;
6580
6581 mutex_lock(&wq_pool_attach_mutex);
6582 if (pool->cpu == cpu)
6583 rebind_workers(pool);
6584 else if (pool->cpu < 0)
6585 restore_unbound_workers_cpumask(pool, cpu);
6586 mutex_unlock(&wq_pool_attach_mutex);
6587 }
6588
6589 /* update pod affinity of unbound workqueues */
6590 list_for_each_entry(wq, &workqueues, list) {
6591 struct workqueue_attrs *attrs = wq->unbound_attrs;
6592
6593 if (attrs) {
6594 const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
6595 int tcpu;
6596
6597 for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]])
6598 unbound_wq_update_pwq(wq, tcpu);
6599
6600 mutex_lock(&wq->mutex);
6601 wq_update_node_max_active(wq, -1);
6602 mutex_unlock(&wq->mutex);
6603 }
6604 }
6605
6606 mutex_unlock(&wq_pool_mutex);
6607 return 0;
6608 }
6609
workqueue_offline_cpu(unsigned int cpu)6610 int workqueue_offline_cpu(unsigned int cpu)
6611 {
6612 struct workqueue_struct *wq;
6613
6614 /* unbinding per-cpu workers should happen on the local CPU */
6615 if (WARN_ON(cpu != smp_processor_id()))
6616 return -1;
6617
6618 unbind_workers(cpu);
6619
6620 /* update pod affinity of unbound workqueues */
6621 mutex_lock(&wq_pool_mutex);
6622
6623 cpumask_clear_cpu(cpu, wq_online_cpumask);
6624
6625 list_for_each_entry(wq, &workqueues, list) {
6626 struct workqueue_attrs *attrs = wq->unbound_attrs;
6627
6628 if (attrs) {
6629 const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
6630 int tcpu;
6631
6632 for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]])
6633 unbound_wq_update_pwq(wq, tcpu);
6634
6635 mutex_lock(&wq->mutex);
6636 wq_update_node_max_active(wq, cpu);
6637 mutex_unlock(&wq->mutex);
6638 }
6639 }
6640 mutex_unlock(&wq_pool_mutex);
6641
6642 return 0;
6643 }
6644
6645 struct work_for_cpu {
6646 struct work_struct work;
6647 long (*fn)(void *);
6648 void *arg;
6649 long ret;
6650 };
6651
work_for_cpu_fn(struct work_struct * work)6652 static void work_for_cpu_fn(struct work_struct *work)
6653 {
6654 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
6655
6656 wfc->ret = wfc->fn(wfc->arg);
6657 }
6658
6659 /**
6660 * work_on_cpu_key - run a function in thread context on a particular cpu
6661 * @cpu: the cpu to run on
6662 * @fn: the function to run
6663 * @arg: the function arg
6664 * @key: The lock class key for lock debugging purposes
6665 *
6666 * It is up to the caller to ensure that the cpu doesn't go offline.
6667 * The caller must not hold any locks which would prevent @fn from completing.
6668 *
6669 * Return: The value @fn returns.
6670 */
work_on_cpu_key(int cpu,long (* fn)(void *),void * arg,struct lock_class_key * key)6671 long work_on_cpu_key(int cpu, long (*fn)(void *),
6672 void *arg, struct lock_class_key *key)
6673 {
6674 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
6675
6676 INIT_WORK_ONSTACK_KEY(&wfc.work, work_for_cpu_fn, key);
6677 schedule_work_on(cpu, &wfc.work);
6678 flush_work(&wfc.work);
6679 destroy_work_on_stack(&wfc.work);
6680 return wfc.ret;
6681 }
6682 EXPORT_SYMBOL_GPL(work_on_cpu_key);
6683
6684 /**
6685 * work_on_cpu_safe_key - run a function in thread context on a particular cpu
6686 * @cpu: the cpu to run on
6687 * @fn: the function to run
6688 * @arg: the function argument
6689 * @key: The lock class key for lock debugging purposes
6690 *
6691 * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold
6692 * any locks which would prevent @fn from completing.
6693 *
6694 * Return: The value @fn returns.
6695 */
work_on_cpu_safe_key(int cpu,long (* fn)(void *),void * arg,struct lock_class_key * key)6696 long work_on_cpu_safe_key(int cpu, long (*fn)(void *),
6697 void *arg, struct lock_class_key *key)
6698 {
6699 long ret = -ENODEV;
6700
6701 cpus_read_lock();
6702 if (cpu_online(cpu))
6703 ret = work_on_cpu_key(cpu, fn, arg, key);
6704 cpus_read_unlock();
6705 return ret;
6706 }
6707 EXPORT_SYMBOL_GPL(work_on_cpu_safe_key);
6708 #endif /* CONFIG_SMP */
6709
6710 #ifdef CONFIG_FREEZER
6711
6712 /**
6713 * freeze_workqueues_begin - begin freezing workqueues
6714 *
6715 * Start freezing workqueues. After this function returns, all freezable
6716 * workqueues will queue new works to their inactive_works list instead of
6717 * pool->worklist.
6718 *
6719 * CONTEXT:
6720 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
6721 */
freeze_workqueues_begin(void)6722 void freeze_workqueues_begin(void)
6723 {
6724 struct workqueue_struct *wq;
6725
6726 mutex_lock(&wq_pool_mutex);
6727
6728 WARN_ON_ONCE(workqueue_freezing);
6729 workqueue_freezing = true;
6730
6731 list_for_each_entry(wq, &workqueues, list) {
6732 mutex_lock(&wq->mutex);
6733 wq_adjust_max_active(wq);
6734 mutex_unlock(&wq->mutex);
6735 }
6736
6737 mutex_unlock(&wq_pool_mutex);
6738 }
6739
6740 /**
6741 * freeze_workqueues_busy - are freezable workqueues still busy?
6742 *
6743 * Check whether freezing is complete. This function must be called
6744 * between freeze_workqueues_begin() and thaw_workqueues().
6745 *
6746 * CONTEXT:
6747 * Grabs and releases wq_pool_mutex.
6748 *
6749 * Return:
6750 * %true if some freezable workqueues are still busy. %false if freezing
6751 * is complete.
6752 */
freeze_workqueues_busy(void)6753 bool freeze_workqueues_busy(void)
6754 {
6755 bool busy = false;
6756 struct workqueue_struct *wq;
6757 struct pool_workqueue *pwq;
6758
6759 mutex_lock(&wq_pool_mutex);
6760
6761 WARN_ON_ONCE(!workqueue_freezing);
6762
6763 list_for_each_entry(wq, &workqueues, list) {
6764 if (!(wq->flags & WQ_FREEZABLE))
6765 continue;
6766 /*
6767 * nr_active is monotonically decreasing. It's safe
6768 * to peek without lock.
6769 */
6770 rcu_read_lock();
6771 for_each_pwq(pwq, wq) {
6772 WARN_ON_ONCE(pwq->nr_active < 0);
6773 if (pwq->nr_active) {
6774 busy = true;
6775 rcu_read_unlock();
6776 goto out_unlock;
6777 }
6778 }
6779 rcu_read_unlock();
6780 }
6781 out_unlock:
6782 mutex_unlock(&wq_pool_mutex);
6783 return busy;
6784 }
6785
6786 /**
6787 * thaw_workqueues - thaw workqueues
6788 *
6789 * Thaw workqueues. Normal queueing is restored and all collected
6790 * frozen works are transferred to their respective pool worklists.
6791 *
6792 * CONTEXT:
6793 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
6794 */
thaw_workqueues(void)6795 void thaw_workqueues(void)
6796 {
6797 struct workqueue_struct *wq;
6798
6799 mutex_lock(&wq_pool_mutex);
6800
6801 if (!workqueue_freezing)
6802 goto out_unlock;
6803
6804 workqueue_freezing = false;
6805
6806 /* restore max_active and repopulate worklist */
6807 list_for_each_entry(wq, &workqueues, list) {
6808 mutex_lock(&wq->mutex);
6809 wq_adjust_max_active(wq);
6810 mutex_unlock(&wq->mutex);
6811 }
6812
6813 out_unlock:
6814 mutex_unlock(&wq_pool_mutex);
6815 }
6816 #endif /* CONFIG_FREEZER */
6817
workqueue_apply_unbound_cpumask(const cpumask_var_t unbound_cpumask)6818 static int workqueue_apply_unbound_cpumask(const cpumask_var_t unbound_cpumask)
6819 {
6820 LIST_HEAD(ctxs);
6821 int ret = 0;
6822 struct workqueue_struct *wq;
6823 struct apply_wqattrs_ctx *ctx, *n;
6824
6825 lockdep_assert_held(&wq_pool_mutex);
6826
6827 list_for_each_entry(wq, &workqueues, list) {
6828 if (!(wq->flags & WQ_UNBOUND) || (wq->flags & __WQ_DESTROYING))
6829 continue;
6830
6831 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs, unbound_cpumask);
6832 if (IS_ERR(ctx)) {
6833 ret = PTR_ERR(ctx);
6834 break;
6835 }
6836
6837 list_add_tail(&ctx->list, &ctxs);
6838 }
6839
6840 list_for_each_entry_safe(ctx, n, &ctxs, list) {
6841 if (!ret)
6842 apply_wqattrs_commit(ctx);
6843 apply_wqattrs_cleanup(ctx);
6844 }
6845
6846 if (!ret) {
6847 mutex_lock(&wq_pool_attach_mutex);
6848 cpumask_copy(wq_unbound_cpumask, unbound_cpumask);
6849 mutex_unlock(&wq_pool_attach_mutex);
6850 }
6851 return ret;
6852 }
6853
6854 /**
6855 * workqueue_unbound_exclude_cpumask - Exclude given CPUs from unbound cpumask
6856 * @exclude_cpumask: the cpumask to be excluded from wq_unbound_cpumask
6857 *
6858 * This function can be called from cpuset code to provide a set of isolated
6859 * CPUs that should be excluded from wq_unbound_cpumask.
6860 */
workqueue_unbound_exclude_cpumask(cpumask_var_t exclude_cpumask)6861 int workqueue_unbound_exclude_cpumask(cpumask_var_t exclude_cpumask)
6862 {
6863 cpumask_var_t cpumask;
6864 int ret = 0;
6865
6866 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
6867 return -ENOMEM;
6868
6869 mutex_lock(&wq_pool_mutex);
6870
6871 /*
6872 * If the operation fails, it will fall back to
6873 * wq_requested_unbound_cpumask which is initially set to
6874 * (HK_TYPE_WQ ∩ HK_TYPE_DOMAIN) house keeping mask and rewritten
6875 * by any subsequent write to workqueue/cpumask sysfs file.
6876 */
6877 if (!cpumask_andnot(cpumask, wq_requested_unbound_cpumask, exclude_cpumask))
6878 cpumask_copy(cpumask, wq_requested_unbound_cpumask);
6879 if (!cpumask_equal(cpumask, wq_unbound_cpumask))
6880 ret = workqueue_apply_unbound_cpumask(cpumask);
6881
6882 /* Save the current isolated cpumask & export it via sysfs */
6883 if (!ret)
6884 cpumask_copy(wq_isolated_cpumask, exclude_cpumask);
6885
6886 mutex_unlock(&wq_pool_mutex);
6887 free_cpumask_var(cpumask);
6888 return ret;
6889 }
6890
parse_affn_scope(const char * val)6891 static int parse_affn_scope(const char *val)
6892 {
6893 int i;
6894
6895 for (i = 0; i < ARRAY_SIZE(wq_affn_names); i++) {
6896 if (!strncasecmp(val, wq_affn_names[i], strlen(wq_affn_names[i])))
6897 return i;
6898 }
6899 return -EINVAL;
6900 }
6901
wq_affn_dfl_set(const char * val,const struct kernel_param * kp)6902 static int wq_affn_dfl_set(const char *val, const struct kernel_param *kp)
6903 {
6904 struct workqueue_struct *wq;
6905 int affn, cpu;
6906
6907 affn = parse_affn_scope(val);
6908 if (affn < 0)
6909 return affn;
6910 if (affn == WQ_AFFN_DFL)
6911 return -EINVAL;
6912
6913 cpus_read_lock();
6914 mutex_lock(&wq_pool_mutex);
6915
6916 wq_affn_dfl = affn;
6917
6918 list_for_each_entry(wq, &workqueues, list) {
6919 for_each_online_cpu(cpu)
6920 unbound_wq_update_pwq(wq, cpu);
6921 }
6922
6923 mutex_unlock(&wq_pool_mutex);
6924 cpus_read_unlock();
6925
6926 return 0;
6927 }
6928
wq_affn_dfl_get(char * buffer,const struct kernel_param * kp)6929 static int wq_affn_dfl_get(char *buffer, const struct kernel_param *kp)
6930 {
6931 return scnprintf(buffer, PAGE_SIZE, "%s\n", wq_affn_names[wq_affn_dfl]);
6932 }
6933
6934 static const struct kernel_param_ops wq_affn_dfl_ops = {
6935 .set = wq_affn_dfl_set,
6936 .get = wq_affn_dfl_get,
6937 };
6938
6939 module_param_cb(default_affinity_scope, &wq_affn_dfl_ops, NULL, 0644);
6940
6941 #ifdef CONFIG_SYSFS
6942 /*
6943 * Workqueues with WQ_SYSFS flag set is visible to userland via
6944 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
6945 * following attributes.
6946 *
6947 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
6948 * max_active RW int : maximum number of in-flight work items
6949 *
6950 * Unbound workqueues have the following extra attributes.
6951 *
6952 * nice RW int : nice value of the workers
6953 * cpumask RW mask : bitmask of allowed CPUs for the workers
6954 * affinity_scope RW str : worker CPU affinity scope (cache, numa, none)
6955 * affinity_strict RW bool : worker CPU affinity is strict
6956 */
6957 struct wq_device {
6958 struct workqueue_struct *wq;
6959 struct device dev;
6960 };
6961
dev_to_wq(struct device * dev)6962 static struct workqueue_struct *dev_to_wq(struct device *dev)
6963 {
6964 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
6965
6966 return wq_dev->wq;
6967 }
6968
per_cpu_show(struct device * dev,struct device_attribute * attr,char * buf)6969 static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
6970 char *buf)
6971 {
6972 struct workqueue_struct *wq = dev_to_wq(dev);
6973
6974 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
6975 }
6976 static DEVICE_ATTR_RO(per_cpu);
6977
max_active_show(struct device * dev,struct device_attribute * attr,char * buf)6978 static ssize_t max_active_show(struct device *dev,
6979 struct device_attribute *attr, char *buf)
6980 {
6981 struct workqueue_struct *wq = dev_to_wq(dev);
6982
6983 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
6984 }
6985
max_active_store(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)6986 static ssize_t max_active_store(struct device *dev,
6987 struct device_attribute *attr, const char *buf,
6988 size_t count)
6989 {
6990 struct workqueue_struct *wq = dev_to_wq(dev);
6991 int val;
6992
6993 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
6994 return -EINVAL;
6995
6996 workqueue_set_max_active(wq, val);
6997 return count;
6998 }
6999 static DEVICE_ATTR_RW(max_active);
7000
7001 static struct attribute *wq_sysfs_attrs[] = {
7002 &dev_attr_per_cpu.attr,
7003 &dev_attr_max_active.attr,
7004 NULL,
7005 };
7006 ATTRIBUTE_GROUPS(wq_sysfs);
7007
wq_nice_show(struct device * dev,struct device_attribute * attr,char * buf)7008 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
7009 char *buf)
7010 {
7011 struct workqueue_struct *wq = dev_to_wq(dev);
7012 int written;
7013
7014 mutex_lock(&wq->mutex);
7015 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
7016 mutex_unlock(&wq->mutex);
7017
7018 return written;
7019 }
7020
7021 /* prepare workqueue_attrs for sysfs store operations */
wq_sysfs_prep_attrs(struct workqueue_struct * wq)7022 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
7023 {
7024 struct workqueue_attrs *attrs;
7025
7026 lockdep_assert_held(&wq_pool_mutex);
7027
7028 attrs = alloc_workqueue_attrs();
7029 if (!attrs)
7030 return NULL;
7031
7032 copy_workqueue_attrs(attrs, wq->unbound_attrs);
7033 return attrs;
7034 }
7035
wq_nice_store(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)7036 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
7037 const char *buf, size_t count)
7038 {
7039 struct workqueue_struct *wq = dev_to_wq(dev);
7040 struct workqueue_attrs *attrs;
7041 int ret = -ENOMEM;
7042
7043 apply_wqattrs_lock();
7044
7045 attrs = wq_sysfs_prep_attrs(wq);
7046 if (!attrs)
7047 goto out_unlock;
7048
7049 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
7050 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
7051 ret = apply_workqueue_attrs_locked(wq, attrs);
7052 else
7053 ret = -EINVAL;
7054
7055 out_unlock:
7056 apply_wqattrs_unlock();
7057 free_workqueue_attrs(attrs);
7058 return ret ?: count;
7059 }
7060
wq_cpumask_show(struct device * dev,struct device_attribute * attr,char * buf)7061 static ssize_t wq_cpumask_show(struct device *dev,
7062 struct device_attribute *attr, char *buf)
7063 {
7064 struct workqueue_struct *wq = dev_to_wq(dev);
7065 int written;
7066
7067 mutex_lock(&wq->mutex);
7068 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
7069 cpumask_pr_args(wq->unbound_attrs->cpumask));
7070 mutex_unlock(&wq->mutex);
7071 return written;
7072 }
7073
wq_cpumask_store(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)7074 static ssize_t wq_cpumask_store(struct device *dev,
7075 struct device_attribute *attr,
7076 const char *buf, size_t count)
7077 {
7078 struct workqueue_struct *wq = dev_to_wq(dev);
7079 struct workqueue_attrs *attrs;
7080 int ret = -ENOMEM;
7081
7082 apply_wqattrs_lock();
7083
7084 attrs = wq_sysfs_prep_attrs(wq);
7085 if (!attrs)
7086 goto out_unlock;
7087
7088 ret = cpumask_parse(buf, attrs->cpumask);
7089 if (!ret)
7090 ret = apply_workqueue_attrs_locked(wq, attrs);
7091
7092 out_unlock:
7093 apply_wqattrs_unlock();
7094 free_workqueue_attrs(attrs);
7095 return ret ?: count;
7096 }
7097
wq_affn_scope_show(struct device * dev,struct device_attribute * attr,char * buf)7098 static ssize_t wq_affn_scope_show(struct device *dev,
7099 struct device_attribute *attr, char *buf)
7100 {
7101 struct workqueue_struct *wq = dev_to_wq(dev);
7102 int written;
7103
7104 mutex_lock(&wq->mutex);
7105 if (wq->unbound_attrs->affn_scope == WQ_AFFN_DFL)
7106 written = scnprintf(buf, PAGE_SIZE, "%s (%s)\n",
7107 wq_affn_names[WQ_AFFN_DFL],
7108 wq_affn_names[wq_affn_dfl]);
7109 else
7110 written = scnprintf(buf, PAGE_SIZE, "%s\n",
7111 wq_affn_names[wq->unbound_attrs->affn_scope]);
7112 mutex_unlock(&wq->mutex);
7113
7114 return written;
7115 }
7116
wq_affn_scope_store(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)7117 static ssize_t wq_affn_scope_store(struct device *dev,
7118 struct device_attribute *attr,
7119 const char *buf, size_t count)
7120 {
7121 struct workqueue_struct *wq = dev_to_wq(dev);
7122 struct workqueue_attrs *attrs;
7123 int affn, ret = -ENOMEM;
7124
7125 affn = parse_affn_scope(buf);
7126 if (affn < 0)
7127 return affn;
7128
7129 apply_wqattrs_lock();
7130 attrs = wq_sysfs_prep_attrs(wq);
7131 if (attrs) {
7132 attrs->affn_scope = affn;
7133 ret = apply_workqueue_attrs_locked(wq, attrs);
7134 }
7135 apply_wqattrs_unlock();
7136 free_workqueue_attrs(attrs);
7137 return ret ?: count;
7138 }
7139
wq_affinity_strict_show(struct device * dev,struct device_attribute * attr,char * buf)7140 static ssize_t wq_affinity_strict_show(struct device *dev,
7141 struct device_attribute *attr, char *buf)
7142 {
7143 struct workqueue_struct *wq = dev_to_wq(dev);
7144
7145 return scnprintf(buf, PAGE_SIZE, "%d\n",
7146 wq->unbound_attrs->affn_strict);
7147 }
7148
wq_affinity_strict_store(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)7149 static ssize_t wq_affinity_strict_store(struct device *dev,
7150 struct device_attribute *attr,
7151 const char *buf, size_t count)
7152 {
7153 struct workqueue_struct *wq = dev_to_wq(dev);
7154 struct workqueue_attrs *attrs;
7155 int v, ret = -ENOMEM;
7156
7157 if (sscanf(buf, "%d", &v) != 1)
7158 return -EINVAL;
7159
7160 apply_wqattrs_lock();
7161 attrs = wq_sysfs_prep_attrs(wq);
7162 if (attrs) {
7163 attrs->affn_strict = (bool)v;
7164 ret = apply_workqueue_attrs_locked(wq, attrs);
7165 }
7166 apply_wqattrs_unlock();
7167 free_workqueue_attrs(attrs);
7168 return ret ?: count;
7169 }
7170
7171 static struct device_attribute wq_sysfs_unbound_attrs[] = {
7172 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
7173 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
7174 __ATTR(affinity_scope, 0644, wq_affn_scope_show, wq_affn_scope_store),
7175 __ATTR(affinity_strict, 0644, wq_affinity_strict_show, wq_affinity_strict_store),
7176 __ATTR_NULL,
7177 };
7178
7179 static const struct bus_type wq_subsys = {
7180 .name = "workqueue",
7181 .dev_groups = wq_sysfs_groups,
7182 };
7183
7184 /**
7185 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
7186 * @cpumask: the cpumask to set
7187 *
7188 * The low-level workqueues cpumask is a global cpumask that limits
7189 * the affinity of all unbound workqueues. This function check the @cpumask
7190 * and apply it to all unbound workqueues and updates all pwqs of them.
7191 *
7192 * Return: 0 - Success
7193 * -EINVAL - Invalid @cpumask
7194 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
7195 */
workqueue_set_unbound_cpumask(cpumask_var_t cpumask)7196 static int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
7197 {
7198 int ret = -EINVAL;
7199
7200 /*
7201 * Not excluding isolated cpus on purpose.
7202 * If the user wishes to include them, we allow that.
7203 */
7204 cpumask_and(cpumask, cpumask, cpu_possible_mask);
7205 if (!cpumask_empty(cpumask)) {
7206 ret = 0;
7207 apply_wqattrs_lock();
7208 if (!cpumask_equal(cpumask, wq_unbound_cpumask))
7209 ret = workqueue_apply_unbound_cpumask(cpumask);
7210 if (!ret)
7211 cpumask_copy(wq_requested_unbound_cpumask, cpumask);
7212 apply_wqattrs_unlock();
7213 }
7214
7215 return ret;
7216 }
7217
__wq_cpumask_show(struct device * dev,struct device_attribute * attr,char * buf,cpumask_var_t mask)7218 static ssize_t __wq_cpumask_show(struct device *dev,
7219 struct device_attribute *attr, char *buf, cpumask_var_t mask)
7220 {
7221 int written;
7222
7223 mutex_lock(&wq_pool_mutex);
7224 written = scnprintf(buf, PAGE_SIZE, "%*pb\n", cpumask_pr_args(mask));
7225 mutex_unlock(&wq_pool_mutex);
7226
7227 return written;
7228 }
7229
cpumask_requested_show(struct device * dev,struct device_attribute * attr,char * buf)7230 static ssize_t cpumask_requested_show(struct device *dev,
7231 struct device_attribute *attr, char *buf)
7232 {
7233 return __wq_cpumask_show(dev, attr, buf, wq_requested_unbound_cpumask);
7234 }
7235 static DEVICE_ATTR_RO(cpumask_requested);
7236
cpumask_isolated_show(struct device * dev,struct device_attribute * attr,char * buf)7237 static ssize_t cpumask_isolated_show(struct device *dev,
7238 struct device_attribute *attr, char *buf)
7239 {
7240 return __wq_cpumask_show(dev, attr, buf, wq_isolated_cpumask);
7241 }
7242 static DEVICE_ATTR_RO(cpumask_isolated);
7243
cpumask_show(struct device * dev,struct device_attribute * attr,char * buf)7244 static ssize_t cpumask_show(struct device *dev,
7245 struct device_attribute *attr, char *buf)
7246 {
7247 return __wq_cpumask_show(dev, attr, buf, wq_unbound_cpumask);
7248 }
7249
cpumask_store(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)7250 static ssize_t cpumask_store(struct device *dev,
7251 struct device_attribute *attr, const char *buf, size_t count)
7252 {
7253 cpumask_var_t cpumask;
7254 int ret;
7255
7256 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
7257 return -ENOMEM;
7258
7259 ret = cpumask_parse(buf, cpumask);
7260 if (!ret)
7261 ret = workqueue_set_unbound_cpumask(cpumask);
7262
7263 free_cpumask_var(cpumask);
7264 return ret ? ret : count;
7265 }
7266 static DEVICE_ATTR_RW(cpumask);
7267
7268 static struct attribute *wq_sysfs_cpumask_attrs[] = {
7269 &dev_attr_cpumask.attr,
7270 &dev_attr_cpumask_requested.attr,
7271 &dev_attr_cpumask_isolated.attr,
7272 NULL,
7273 };
7274 ATTRIBUTE_GROUPS(wq_sysfs_cpumask);
7275
wq_sysfs_init(void)7276 static int __init wq_sysfs_init(void)
7277 {
7278 return subsys_virtual_register(&wq_subsys, wq_sysfs_cpumask_groups);
7279 }
7280 core_initcall(wq_sysfs_init);
7281
wq_device_release(struct device * dev)7282 static void wq_device_release(struct device *dev)
7283 {
7284 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
7285
7286 kfree(wq_dev);
7287 }
7288
7289 /**
7290 * workqueue_sysfs_register - make a workqueue visible in sysfs
7291 * @wq: the workqueue to register
7292 *
7293 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
7294 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
7295 * which is the preferred method.
7296 *
7297 * Workqueue user should use this function directly iff it wants to apply
7298 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
7299 * apply_workqueue_attrs() may race against userland updating the
7300 * attributes.
7301 *
7302 * Return: 0 on success, -errno on failure.
7303 */
workqueue_sysfs_register(struct workqueue_struct * wq)7304 int workqueue_sysfs_register(struct workqueue_struct *wq)
7305 {
7306 struct wq_device *wq_dev;
7307 int ret;
7308
7309 /*
7310 * Adjusting max_active breaks ordering guarantee. Disallow exposing
7311 * ordered workqueues.
7312 */
7313 if (WARN_ON(wq->flags & __WQ_ORDERED))
7314 return -EINVAL;
7315
7316 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
7317 if (!wq_dev)
7318 return -ENOMEM;
7319
7320 wq_dev->wq = wq;
7321 wq_dev->dev.bus = &wq_subsys;
7322 wq_dev->dev.release = wq_device_release;
7323 dev_set_name(&wq_dev->dev, "%s", wq->name);
7324
7325 /*
7326 * unbound_attrs are created separately. Suppress uevent until
7327 * everything is ready.
7328 */
7329 dev_set_uevent_suppress(&wq_dev->dev, true);
7330
7331 ret = device_register(&wq_dev->dev);
7332 if (ret) {
7333 put_device(&wq_dev->dev);
7334 wq->wq_dev = NULL;
7335 return ret;
7336 }
7337
7338 if (wq->flags & WQ_UNBOUND) {
7339 struct device_attribute *attr;
7340
7341 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
7342 ret = device_create_file(&wq_dev->dev, attr);
7343 if (ret) {
7344 device_unregister(&wq_dev->dev);
7345 wq->wq_dev = NULL;
7346 return ret;
7347 }
7348 }
7349 }
7350
7351 dev_set_uevent_suppress(&wq_dev->dev, false);
7352 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
7353 return 0;
7354 }
7355
7356 /**
7357 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
7358 * @wq: the workqueue to unregister
7359 *
7360 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
7361 */
workqueue_sysfs_unregister(struct workqueue_struct * wq)7362 static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
7363 {
7364 struct wq_device *wq_dev = wq->wq_dev;
7365
7366 if (!wq->wq_dev)
7367 return;
7368
7369 wq->wq_dev = NULL;
7370 device_unregister(&wq_dev->dev);
7371 }
7372 #else /* CONFIG_SYSFS */
workqueue_sysfs_unregister(struct workqueue_struct * wq)7373 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
7374 #endif /* CONFIG_SYSFS */
7375
7376 /*
7377 * Workqueue watchdog.
7378 *
7379 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
7380 * flush dependency, a concurrency managed work item which stays RUNNING
7381 * indefinitely. Workqueue stalls can be very difficult to debug as the
7382 * usual warning mechanisms don't trigger and internal workqueue state is
7383 * largely opaque.
7384 *
7385 * Workqueue watchdog monitors all worker pools periodically and dumps
7386 * state if some pools failed to make forward progress for a while where
7387 * forward progress is defined as the first item on ->worklist changing.
7388 *
7389 * This mechanism is controlled through the kernel parameter
7390 * "workqueue.watchdog_thresh" which can be updated at runtime through the
7391 * corresponding sysfs parameter file.
7392 */
7393 #ifdef CONFIG_WQ_WATCHDOG
7394
7395 static unsigned long wq_watchdog_thresh = 30;
7396 static struct timer_list wq_watchdog_timer;
7397
7398 static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
7399 static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
7400
7401 /*
7402 * Show workers that might prevent the processing of pending work items.
7403 * The only candidates are CPU-bound workers in the running state.
7404 * Pending work items should be handled by another idle worker
7405 * in all other situations.
7406 */
show_cpu_pool_hog(struct worker_pool * pool)7407 static void show_cpu_pool_hog(struct worker_pool *pool)
7408 {
7409 struct worker *worker;
7410 unsigned long irq_flags;
7411 int bkt;
7412
7413 raw_spin_lock_irqsave(&pool->lock, irq_flags);
7414
7415 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
7416 if (task_is_running(worker->task)) {
7417 /*
7418 * Defer printing to avoid deadlocks in console
7419 * drivers that queue work while holding locks
7420 * also taken in their write paths.
7421 */
7422 printk_deferred_enter();
7423
7424 pr_info("pool %d:\n", pool->id);
7425 sched_show_task(worker->task);
7426
7427 printk_deferred_exit();
7428 }
7429 }
7430
7431 raw_spin_unlock_irqrestore(&pool->lock, irq_flags);
7432 }
7433
show_cpu_pools_hogs(void)7434 static void show_cpu_pools_hogs(void)
7435 {
7436 struct worker_pool *pool;
7437 int pi;
7438
7439 pr_info("Showing backtraces of running workers in stalled CPU-bound worker pools:\n");
7440
7441 rcu_read_lock();
7442
7443 for_each_pool(pool, pi) {
7444 if (pool->cpu_stall)
7445 show_cpu_pool_hog(pool);
7446
7447 }
7448
7449 rcu_read_unlock();
7450 }
7451
wq_watchdog_reset_touched(void)7452 static void wq_watchdog_reset_touched(void)
7453 {
7454 int cpu;
7455
7456 wq_watchdog_touched = jiffies;
7457 for_each_possible_cpu(cpu)
7458 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
7459 }
7460
wq_watchdog_timer_fn(struct timer_list * unused)7461 static void wq_watchdog_timer_fn(struct timer_list *unused)
7462 {
7463 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
7464 bool lockup_detected = false;
7465 bool cpu_pool_stall = false;
7466 unsigned long now = jiffies;
7467 struct worker_pool *pool;
7468 int pi;
7469
7470 if (!thresh)
7471 return;
7472
7473 rcu_read_lock();
7474
7475 for_each_pool(pool, pi) {
7476 unsigned long pool_ts, touched, ts;
7477
7478 pool->cpu_stall = false;
7479 if (list_empty(&pool->worklist))
7480 continue;
7481
7482 /*
7483 * If a virtual machine is stopped by the host it can look to
7484 * the watchdog like a stall.
7485 */
7486 kvm_check_and_clear_guest_paused();
7487
7488 /* get the latest of pool and touched timestamps */
7489 if (pool->cpu >= 0)
7490 touched = READ_ONCE(per_cpu(wq_watchdog_touched_cpu, pool->cpu));
7491 else
7492 touched = READ_ONCE(wq_watchdog_touched);
7493 pool_ts = READ_ONCE(pool->watchdog_ts);
7494
7495 if (time_after(pool_ts, touched))
7496 ts = pool_ts;
7497 else
7498 ts = touched;
7499
7500 /* did we stall? */
7501 if (time_after(now, ts + thresh)) {
7502 lockup_detected = true;
7503 if (pool->cpu >= 0 && !(pool->flags & POOL_BH)) {
7504 pool->cpu_stall = true;
7505 cpu_pool_stall = true;
7506 }
7507 pr_emerg("BUG: workqueue lockup - pool");
7508 pr_cont_pool_info(pool);
7509 pr_cont(" stuck for %us!\n",
7510 jiffies_to_msecs(now - pool_ts) / 1000);
7511 }
7512
7513
7514 }
7515
7516 rcu_read_unlock();
7517
7518 if (lockup_detected)
7519 show_all_workqueues();
7520
7521 if (cpu_pool_stall)
7522 show_cpu_pools_hogs();
7523
7524 wq_watchdog_reset_touched();
7525 mod_timer(&wq_watchdog_timer, jiffies + thresh);
7526 }
7527
wq_watchdog_touch(int cpu)7528 notrace void wq_watchdog_touch(int cpu)
7529 {
7530 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
7531 unsigned long touch_ts = READ_ONCE(wq_watchdog_touched);
7532 unsigned long now = jiffies;
7533
7534 if (cpu >= 0)
7535 per_cpu(wq_watchdog_touched_cpu, cpu) = now;
7536 else
7537 WARN_ONCE(1, "%s should be called with valid CPU", __func__);
7538
7539 /* Don't unnecessarily store to global cacheline */
7540 if (time_after(now, touch_ts + thresh / 4))
7541 WRITE_ONCE(wq_watchdog_touched, jiffies);
7542 }
7543
wq_watchdog_set_thresh(unsigned long thresh)7544 static void wq_watchdog_set_thresh(unsigned long thresh)
7545 {
7546 wq_watchdog_thresh = 0;
7547 del_timer_sync(&wq_watchdog_timer);
7548
7549 if (thresh) {
7550 wq_watchdog_thresh = thresh;
7551 wq_watchdog_reset_touched();
7552 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
7553 }
7554 }
7555
wq_watchdog_param_set_thresh(const char * val,const struct kernel_param * kp)7556 static int wq_watchdog_param_set_thresh(const char *val,
7557 const struct kernel_param *kp)
7558 {
7559 unsigned long thresh;
7560 int ret;
7561
7562 ret = kstrtoul(val, 0, &thresh);
7563 if (ret)
7564 return ret;
7565
7566 if (system_wq)
7567 wq_watchdog_set_thresh(thresh);
7568 else
7569 wq_watchdog_thresh = thresh;
7570
7571 return 0;
7572 }
7573
7574 static const struct kernel_param_ops wq_watchdog_thresh_ops = {
7575 .set = wq_watchdog_param_set_thresh,
7576 .get = param_get_ulong,
7577 };
7578
7579 module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
7580 0644);
7581
wq_watchdog_init(void)7582 static void wq_watchdog_init(void)
7583 {
7584 timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE);
7585 wq_watchdog_set_thresh(wq_watchdog_thresh);
7586 }
7587
7588 #else /* CONFIG_WQ_WATCHDOG */
7589
wq_watchdog_init(void)7590 static inline void wq_watchdog_init(void) { }
7591
7592 #endif /* CONFIG_WQ_WATCHDOG */
7593
bh_pool_kick_normal(struct irq_work * irq_work)7594 static void bh_pool_kick_normal(struct irq_work *irq_work)
7595 {
7596 raise_softirq_irqoff(TASKLET_SOFTIRQ);
7597 }
7598
bh_pool_kick_highpri(struct irq_work * irq_work)7599 static void bh_pool_kick_highpri(struct irq_work *irq_work)
7600 {
7601 raise_softirq_irqoff(HI_SOFTIRQ);
7602 }
7603
restrict_unbound_cpumask(const char * name,const struct cpumask * mask)7604 static void __init restrict_unbound_cpumask(const char *name, const struct cpumask *mask)
7605 {
7606 if (!cpumask_intersects(wq_unbound_cpumask, mask)) {
7607 pr_warn("workqueue: Restricting unbound_cpumask (%*pb) with %s (%*pb) leaves no CPU, ignoring\n",
7608 cpumask_pr_args(wq_unbound_cpumask), name, cpumask_pr_args(mask));
7609 return;
7610 }
7611
7612 cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, mask);
7613 }
7614
init_cpu_worker_pool(struct worker_pool * pool,int cpu,int nice)7615 static void __init init_cpu_worker_pool(struct worker_pool *pool, int cpu, int nice)
7616 {
7617 BUG_ON(init_worker_pool(pool));
7618 pool->cpu = cpu;
7619 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
7620 cpumask_copy(pool->attrs->__pod_cpumask, cpumask_of(cpu));
7621 pool->attrs->nice = nice;
7622 pool->attrs->affn_strict = true;
7623 pool->node = cpu_to_node(cpu);
7624
7625 /* alloc pool ID */
7626 mutex_lock(&wq_pool_mutex);
7627 BUG_ON(worker_pool_assign_id(pool));
7628 mutex_unlock(&wq_pool_mutex);
7629 }
7630
7631 /**
7632 * workqueue_init_early - early init for workqueue subsystem
7633 *
7634 * This is the first step of three-staged workqueue subsystem initialization and
7635 * invoked as soon as the bare basics - memory allocation, cpumasks and idr are
7636 * up. It sets up all the data structures and system workqueues and allows early
7637 * boot code to create workqueues and queue/cancel work items. Actual work item
7638 * execution starts only after kthreads can be created and scheduled right
7639 * before early initcalls.
7640 */
workqueue_init_early(void)7641 void __init workqueue_init_early(void)
7642 {
7643 struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_SYSTEM];
7644 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
7645 void (*irq_work_fns[2])(struct irq_work *) = { bh_pool_kick_normal,
7646 bh_pool_kick_highpri };
7647 int i, cpu;
7648
7649 BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
7650
7651 BUG_ON(!alloc_cpumask_var(&wq_online_cpumask, GFP_KERNEL));
7652 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
7653 BUG_ON(!alloc_cpumask_var(&wq_requested_unbound_cpumask, GFP_KERNEL));
7654 BUG_ON(!zalloc_cpumask_var(&wq_isolated_cpumask, GFP_KERNEL));
7655
7656 cpumask_copy(wq_online_cpumask, cpu_online_mask);
7657 cpumask_copy(wq_unbound_cpumask, cpu_possible_mask);
7658 restrict_unbound_cpumask("HK_TYPE_WQ", housekeeping_cpumask(HK_TYPE_WQ));
7659 restrict_unbound_cpumask("HK_TYPE_DOMAIN", housekeeping_cpumask(HK_TYPE_DOMAIN));
7660 if (!cpumask_empty(&wq_cmdline_cpumask))
7661 restrict_unbound_cpumask("workqueue.unbound_cpus", &wq_cmdline_cpumask);
7662
7663 cpumask_copy(wq_requested_unbound_cpumask, wq_unbound_cpumask);
7664
7665 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
7666
7667 unbound_wq_update_pwq_attrs_buf = alloc_workqueue_attrs();
7668 BUG_ON(!unbound_wq_update_pwq_attrs_buf);
7669
7670 /*
7671 * If nohz_full is enabled, set power efficient workqueue as unbound.
7672 * This allows workqueue items to be moved to HK CPUs.
7673 */
7674 if (housekeeping_enabled(HK_TYPE_TICK))
7675 wq_power_efficient = true;
7676
7677 /* initialize WQ_AFFN_SYSTEM pods */
7678 pt->pod_cpus = kcalloc(1, sizeof(pt->pod_cpus[0]), GFP_KERNEL);
7679 pt->pod_node = kcalloc(1, sizeof(pt->pod_node[0]), GFP_KERNEL);
7680 pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL);
7681 BUG_ON(!pt->pod_cpus || !pt->pod_node || !pt->cpu_pod);
7682
7683 BUG_ON(!zalloc_cpumask_var_node(&pt->pod_cpus[0], GFP_KERNEL, NUMA_NO_NODE));
7684
7685 pt->nr_pods = 1;
7686 cpumask_copy(pt->pod_cpus[0], cpu_possible_mask);
7687 pt->pod_node[0] = NUMA_NO_NODE;
7688 pt->cpu_pod[0] = 0;
7689
7690 /* initialize BH and CPU pools */
7691 for_each_possible_cpu(cpu) {
7692 struct worker_pool *pool;
7693
7694 i = 0;
7695 for_each_bh_worker_pool(pool, cpu) {
7696 init_cpu_worker_pool(pool, cpu, std_nice[i]);
7697 pool->flags |= POOL_BH;
7698 init_irq_work(bh_pool_irq_work(pool), irq_work_fns[i]);
7699 i++;
7700 }
7701
7702 i = 0;
7703 for_each_cpu_worker_pool(pool, cpu)
7704 init_cpu_worker_pool(pool, cpu, std_nice[i++]);
7705 }
7706
7707 /* create default unbound and ordered wq attrs */
7708 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
7709 struct workqueue_attrs *attrs;
7710
7711 BUG_ON(!(attrs = alloc_workqueue_attrs()));
7712 attrs->nice = std_nice[i];
7713 unbound_std_wq_attrs[i] = attrs;
7714
7715 /*
7716 * An ordered wq should have only one pwq as ordering is
7717 * guaranteed by max_active which is enforced by pwqs.
7718 */
7719 BUG_ON(!(attrs = alloc_workqueue_attrs()));
7720 attrs->nice = std_nice[i];
7721 attrs->ordered = true;
7722 ordered_wq_attrs[i] = attrs;
7723 }
7724
7725 system_wq = alloc_workqueue("events", 0, 0);
7726 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
7727 system_long_wq = alloc_workqueue("events_long", 0, 0);
7728 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
7729 WQ_MAX_ACTIVE);
7730 system_freezable_wq = alloc_workqueue("events_freezable",
7731 WQ_FREEZABLE, 0);
7732 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
7733 WQ_POWER_EFFICIENT, 0);
7734 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_pwr_efficient",
7735 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
7736 0);
7737 system_bh_wq = alloc_workqueue("events_bh", WQ_BH, 0);
7738 system_bh_highpri_wq = alloc_workqueue("events_bh_highpri",
7739 WQ_BH | WQ_HIGHPRI, 0);
7740 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
7741 !system_unbound_wq || !system_freezable_wq ||
7742 !system_power_efficient_wq ||
7743 !system_freezable_power_efficient_wq ||
7744 !system_bh_wq || !system_bh_highpri_wq);
7745 }
7746
wq_cpu_intensive_thresh_init(void)7747 static void __init wq_cpu_intensive_thresh_init(void)
7748 {
7749 unsigned long thresh;
7750 unsigned long bogo;
7751
7752 pwq_release_worker = kthread_create_worker(0, "pool_workqueue_release");
7753 BUG_ON(IS_ERR(pwq_release_worker));
7754
7755 /* if the user set it to a specific value, keep it */
7756 if (wq_cpu_intensive_thresh_us != ULONG_MAX)
7757 return;
7758
7759 /*
7760 * The default of 10ms is derived from the fact that most modern (as of
7761 * 2023) processors can do a lot in 10ms and that it's just below what
7762 * most consider human-perceivable. However, the kernel also runs on a
7763 * lot slower CPUs including microcontrollers where the threshold is way
7764 * too low.
7765 *
7766 * Let's scale up the threshold upto 1 second if BogoMips is below 4000.
7767 * This is by no means accurate but it doesn't have to be. The mechanism
7768 * is still useful even when the threshold is fully scaled up. Also, as
7769 * the reports would usually be applicable to everyone, some machines
7770 * operating on longer thresholds won't significantly diminish their
7771 * usefulness.
7772 */
7773 thresh = 10 * USEC_PER_MSEC;
7774
7775 /* see init/calibrate.c for lpj -> BogoMIPS calculation */
7776 bogo = max_t(unsigned long, loops_per_jiffy / 500000 * HZ, 1);
7777 if (bogo < 4000)
7778 thresh = min_t(unsigned long, thresh * 4000 / bogo, USEC_PER_SEC);
7779
7780 pr_debug("wq_cpu_intensive_thresh: lpj=%lu BogoMIPS=%lu thresh_us=%lu\n",
7781 loops_per_jiffy, bogo, thresh);
7782
7783 wq_cpu_intensive_thresh_us = thresh;
7784 }
7785
7786 /**
7787 * workqueue_init - bring workqueue subsystem fully online
7788 *
7789 * This is the second step of three-staged workqueue subsystem initialization
7790 * and invoked as soon as kthreads can be created and scheduled. Workqueues have
7791 * been created and work items queued on them, but there are no kworkers
7792 * executing the work items yet. Populate the worker pools with the initial
7793 * workers and enable future kworker creations.
7794 */
workqueue_init(void)7795 void __init workqueue_init(void)
7796 {
7797 struct workqueue_struct *wq;
7798 struct worker_pool *pool;
7799 int cpu, bkt;
7800
7801 wq_cpu_intensive_thresh_init();
7802
7803 mutex_lock(&wq_pool_mutex);
7804
7805 /*
7806 * Per-cpu pools created earlier could be missing node hint. Fix them
7807 * up. Also, create a rescuer for workqueues that requested it.
7808 */
7809 for_each_possible_cpu(cpu) {
7810 for_each_bh_worker_pool(pool, cpu)
7811 pool->node = cpu_to_node(cpu);
7812 for_each_cpu_worker_pool(pool, cpu)
7813 pool->node = cpu_to_node(cpu);
7814 }
7815
7816 list_for_each_entry(wq, &workqueues, list) {
7817 WARN(init_rescuer(wq),
7818 "workqueue: failed to create early rescuer for %s",
7819 wq->name);
7820 }
7821
7822 mutex_unlock(&wq_pool_mutex);
7823
7824 /*
7825 * Create the initial workers. A BH pool has one pseudo worker that
7826 * represents the shared BH execution context and thus doesn't get
7827 * affected by hotplug events. Create the BH pseudo workers for all
7828 * possible CPUs here.
7829 */
7830 for_each_possible_cpu(cpu)
7831 for_each_bh_worker_pool(pool, cpu)
7832 BUG_ON(!create_worker(pool));
7833
7834 for_each_online_cpu(cpu) {
7835 for_each_cpu_worker_pool(pool, cpu) {
7836 pool->flags &= ~POOL_DISASSOCIATED;
7837 BUG_ON(!create_worker(pool));
7838 }
7839 }
7840
7841 hash_for_each(unbound_pool_hash, bkt, pool, hash_node)
7842 BUG_ON(!create_worker(pool));
7843
7844 wq_online = true;
7845 wq_watchdog_init();
7846 }
7847
7848 /*
7849 * Initialize @pt by first initializing @pt->cpu_pod[] with pod IDs according to
7850 * @cpu_shares_pod(). Each subset of CPUs that share a pod is assigned a unique
7851 * and consecutive pod ID. The rest of @pt is initialized accordingly.
7852 */
init_pod_type(struct wq_pod_type * pt,bool (* cpus_share_pod)(int,int))7853 static void __init init_pod_type(struct wq_pod_type *pt,
7854 bool (*cpus_share_pod)(int, int))
7855 {
7856 int cur, pre, cpu, pod;
7857
7858 pt->nr_pods = 0;
7859
7860 /* init @pt->cpu_pod[] according to @cpus_share_pod() */
7861 pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL);
7862 BUG_ON(!pt->cpu_pod);
7863
7864 for_each_possible_cpu(cur) {
7865 for_each_possible_cpu(pre) {
7866 if (pre >= cur) {
7867 pt->cpu_pod[cur] = pt->nr_pods++;
7868 break;
7869 }
7870 if (cpus_share_pod(cur, pre)) {
7871 pt->cpu_pod[cur] = pt->cpu_pod[pre];
7872 break;
7873 }
7874 }
7875 }
7876
7877 /* init the rest to match @pt->cpu_pod[] */
7878 pt->pod_cpus = kcalloc(pt->nr_pods, sizeof(pt->pod_cpus[0]), GFP_KERNEL);
7879 pt->pod_node = kcalloc(pt->nr_pods, sizeof(pt->pod_node[0]), GFP_KERNEL);
7880 BUG_ON(!pt->pod_cpus || !pt->pod_node);
7881
7882 for (pod = 0; pod < pt->nr_pods; pod++)
7883 BUG_ON(!zalloc_cpumask_var(&pt->pod_cpus[pod], GFP_KERNEL));
7884
7885 for_each_possible_cpu(cpu) {
7886 cpumask_set_cpu(cpu, pt->pod_cpus[pt->cpu_pod[cpu]]);
7887 pt->pod_node[pt->cpu_pod[cpu]] = cpu_to_node(cpu);
7888 }
7889 }
7890
cpus_dont_share(int cpu0,int cpu1)7891 static bool __init cpus_dont_share(int cpu0, int cpu1)
7892 {
7893 return false;
7894 }
7895
cpus_share_smt(int cpu0,int cpu1)7896 static bool __init cpus_share_smt(int cpu0, int cpu1)
7897 {
7898 #ifdef CONFIG_SCHED_SMT
7899 return cpumask_test_cpu(cpu0, cpu_smt_mask(cpu1));
7900 #else
7901 return false;
7902 #endif
7903 }
7904
cpus_share_numa(int cpu0,int cpu1)7905 static bool __init cpus_share_numa(int cpu0, int cpu1)
7906 {
7907 return cpu_to_node(cpu0) == cpu_to_node(cpu1);
7908 }
7909
7910 /**
7911 * workqueue_init_topology - initialize CPU pods for unbound workqueues
7912 *
7913 * This is the third step of three-staged workqueue subsystem initialization and
7914 * invoked after SMP and topology information are fully initialized. It
7915 * initializes the unbound CPU pods accordingly.
7916 */
workqueue_init_topology(void)7917 void __init workqueue_init_topology(void)
7918 {
7919 struct workqueue_struct *wq;
7920 int cpu;
7921
7922 init_pod_type(&wq_pod_types[WQ_AFFN_CPU], cpus_dont_share);
7923 init_pod_type(&wq_pod_types[WQ_AFFN_SMT], cpus_share_smt);
7924 init_pod_type(&wq_pod_types[WQ_AFFN_CACHE], cpus_share_cache);
7925 init_pod_type(&wq_pod_types[WQ_AFFN_NUMA], cpus_share_numa);
7926
7927 wq_topo_initialized = true;
7928
7929 mutex_lock(&wq_pool_mutex);
7930
7931 /*
7932 * Workqueues allocated earlier would have all CPUs sharing the default
7933 * worker pool. Explicitly call unbound_wq_update_pwq() on all workqueue
7934 * and CPU combinations to apply per-pod sharing.
7935 */
7936 list_for_each_entry(wq, &workqueues, list) {
7937 for_each_online_cpu(cpu)
7938 unbound_wq_update_pwq(wq, cpu);
7939 if (wq->flags & WQ_UNBOUND) {
7940 mutex_lock(&wq->mutex);
7941 wq_update_node_max_active(wq, -1);
7942 mutex_unlock(&wq->mutex);
7943 }
7944 }
7945
7946 mutex_unlock(&wq_pool_mutex);
7947 }
7948
__warn_flushing_systemwide_wq(void)7949 void __warn_flushing_systemwide_wq(void)
7950 {
7951 pr_warn("WARNING: Flushing system-wide workqueues will be prohibited in near future.\n");
7952 dump_stack();
7953 }
7954 EXPORT_SYMBOL(__warn_flushing_systemwide_wq);
7955
workqueue_unbound_cpus_setup(char * str)7956 static int __init workqueue_unbound_cpus_setup(char *str)
7957 {
7958 if (cpulist_parse(str, &wq_cmdline_cpumask) < 0) {
7959 cpumask_clear(&wq_cmdline_cpumask);
7960 pr_warn("workqueue.unbound_cpus: incorrect CPU range, using default\n");
7961 }
7962
7963 return 1;
7964 }
7965 __setup("workqueue.unbound_cpus=", workqueue_unbound_cpus_setup);
7966