~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

TOMOYO Linux Cross Reference
Linux/kernel/workqueue.c

Version: ~ [ linux-5.8 ] ~ [ linux-5.7.14 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.57 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.138 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.193 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.232 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.232 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.19.8 ] ~ [ linux-3.18.140 ] ~ [ linux-3.17.8 ] ~ [ linux-3.16.85 ] ~ [ linux-3.15.10 ] ~ [ linux-3.14.79 ] ~ [ linux-3.13.11 ] ~ [ linux-3.12.74 ] ~ [ linux-3.11.10 ] ~ [ linux-3.10.108 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

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

~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

kernel.org | git.kernel.org | LWN.net | Project Home | Wiki (Japanese) | Wiki (English) | SVN repository | Mail admin

Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.

osdn.jp