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

TOMOYO Linux Cross Reference
Linux/kernel/workqueue.c

Version: ~ [ linux-5.1-rc1 ] ~ [ linux-5.0.3 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.30 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.107 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.164 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.176 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.19.8 ] ~ [ linux-3.18.136 ] ~ [ linux-3.17.8 ] ~ [ linux-3.16.63 ] ~ [ 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-3.9.11 ] ~ [ linux-3.8.13 ] ~ [ linux-3.7.10 ] ~ [ linux-3.6.11 ] ~ [ linux-3.5.7 ] ~ [ linux-3.4.113 ] ~ [ linux-3.3.8 ] ~ [ linux-3.2.102 ] ~ [ linux-3.1.10 ] ~ [ linux-3.0.101 ] ~ [ linux-2.6.39.4 ] ~ [ linux-2.6.38.8 ] ~ [ linux-2.6.37.6 ] ~ [ linux-2.6.36.4 ] ~ [ linux-2.6.35.14 ] ~ [ linux-2.6.34.15 ] ~ [ linux-2.6.33.20 ] ~ [ 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 is one worker pool for each CPU and
 20  * one extra for works which are better served by workers which are
 21  * not bound to any specific CPU.
 22  *
 23  * Please read Documentation/workqueue.txt for details.
 24  */
 25 
 26 #include <linux/module.h>
 27 #include <linux/kernel.h>
 28 #include <linux/sched.h>
 29 #include <linux/init.h>
 30 #include <linux/signal.h>
 31 #include <linux/completion.h>
 32 #include <linux/workqueue.h>
 33 #include <linux/slab.h>
 34 #include <linux/cpu.h>
 35 #include <linux/notifier.h>
 36 #include <linux/kthread.h>
 37 #include <linux/hardirq.h>
 38 #include <linux/mempolicy.h>
 39 #include <linux/freezer.h>
 40 #include <linux/kallsyms.h>
 41 #include <linux/debug_locks.h>
 42 #include <linux/lockdep.h>
 43 #include <linux/idr.h>
 44 
 45 #include "workqueue_sched.h"
 46 
 47 enum {
 48         /* global_cwq flags */
 49         GCWQ_MANAGE_WORKERS     = 1 << 0,       /* need to manage workers */
 50         GCWQ_MANAGING_WORKERS   = 1 << 1,       /* managing workers */
 51         GCWQ_DISASSOCIATED      = 1 << 2,       /* cpu can't serve workers */
 52         GCWQ_FREEZING           = 1 << 3,       /* freeze in progress */
 53         GCWQ_HIGHPRI_PENDING    = 1 << 4,       /* highpri works on queue */
 54 
 55         /* worker flags */
 56         WORKER_STARTED          = 1 << 0,       /* started */
 57         WORKER_DIE              = 1 << 1,       /* die die die */
 58         WORKER_IDLE             = 1 << 2,       /* is idle */
 59         WORKER_PREP             = 1 << 3,       /* preparing to run works */
 60         WORKER_ROGUE            = 1 << 4,       /* not bound to any cpu */
 61         WORKER_REBIND           = 1 << 5,       /* mom is home, come back */
 62         WORKER_CPU_INTENSIVE    = 1 << 6,       /* cpu intensive */
 63         WORKER_UNBOUND          = 1 << 7,       /* worker is unbound */
 64 
 65         WORKER_NOT_RUNNING      = WORKER_PREP | WORKER_ROGUE | WORKER_REBIND |
 66                                   WORKER_CPU_INTENSIVE | WORKER_UNBOUND,
 67 
 68         /* gcwq->trustee_state */
 69         TRUSTEE_START           = 0,            /* start */
 70         TRUSTEE_IN_CHARGE       = 1,            /* trustee in charge of gcwq */
 71         TRUSTEE_BUTCHER         = 2,            /* butcher workers */
 72         TRUSTEE_RELEASE         = 3,            /* release workers */
 73         TRUSTEE_DONE            = 4,            /* trustee is done */
 74 
 75         BUSY_WORKER_HASH_ORDER  = 6,            /* 64 pointers */
 76         BUSY_WORKER_HASH_SIZE   = 1 << BUSY_WORKER_HASH_ORDER,
 77         BUSY_WORKER_HASH_MASK   = BUSY_WORKER_HASH_SIZE - 1,
 78 
 79         MAX_IDLE_WORKERS_RATIO  = 4,            /* 1/4 of busy can be idle */
 80         IDLE_WORKER_TIMEOUT     = 300 * HZ,     /* keep idle ones for 5 mins */
 81 
 82         MAYDAY_INITIAL_TIMEOUT  = HZ / 100 >= 2 ? HZ / 100 : 2,
 83                                                 /* call for help after 10ms
 84                                                    (min two ticks) */
 85         MAYDAY_INTERVAL         = HZ / 10,      /* and then every 100ms */
 86         CREATE_COOLDOWN         = HZ,           /* time to breath after fail */
 87         TRUSTEE_COOLDOWN        = HZ / 10,      /* for trustee draining */
 88 
 89         /*
 90          * Rescue workers are used only on emergencies and shared by
 91          * all cpus.  Give -20.
 92          */
 93         RESCUER_NICE_LEVEL      = -20,
 94 };
 95 
 96 /*
 97  * Structure fields follow one of the following exclusion rules.
 98  *
 99  * I: Modifiable by initialization/destruction paths and read-only for
100  *    everyone else.
101  *
102  * P: Preemption protected.  Disabling preemption is enough and should
103  *    only be modified and accessed from the local cpu.
104  *
105  * L: gcwq->lock protected.  Access with gcwq->lock held.
106  *
107  * X: During normal operation, modification requires gcwq->lock and
108  *    should be done only from local cpu.  Either disabling preemption
109  *    on local cpu or grabbing gcwq->lock is enough for read access.
110  *    If GCWQ_DISASSOCIATED is set, it's identical to L.
111  *
112  * F: wq->flush_mutex protected.
113  *
114  * W: workqueue_lock protected.
115  */
116 
117 struct global_cwq;
118 
119 /*
120  * The poor guys doing the actual heavy lifting.  All on-duty workers
121  * are either serving the manager role, on idle list or on busy hash.
122  */
123 struct worker {
124         /* on idle list while idle, on busy hash table while busy */
125         union {
126                 struct list_head        entry;  /* L: while idle */
127                 struct hlist_node       hentry; /* L: while busy */
128         };
129 
130         struct work_struct      *current_work;  /* L: work being processed */
131         struct cpu_workqueue_struct *current_cwq; /* L: current_work's cwq */
132         struct list_head        scheduled;      /* L: scheduled works */
133         struct task_struct      *task;          /* I: worker task */
134         struct global_cwq       *gcwq;          /* I: the associated gcwq */
135         /* 64 bytes boundary on 64bit, 32 on 32bit */
136         unsigned long           last_active;    /* L: last active timestamp */
137         unsigned int            flags;          /* X: flags */
138         int                     id;             /* I: worker id */
139         struct work_struct      rebind_work;    /* L: rebind worker to cpu */
140 };
141 
142 /*
143  * Global per-cpu workqueue.  There's one and only one for each cpu
144  * and all works are queued and processed here regardless of their
145  * target workqueues.
146  */
147 struct global_cwq {
148         spinlock_t              lock;           /* the gcwq lock */
149         struct list_head        worklist;       /* L: list of pending works */
150         unsigned int            cpu;            /* I: the associated cpu */
151         unsigned int            flags;          /* L: GCWQ_* flags */
152 
153         int                     nr_workers;     /* L: total number of workers */
154         int                     nr_idle;        /* L: currently idle ones */
155 
156         /* workers are chained either in the idle_list or busy_hash */
157         struct list_head        idle_list;      /* X: list of idle workers */
158         struct hlist_head       busy_hash[BUSY_WORKER_HASH_SIZE];
159                                                 /* L: hash of busy workers */
160 
161         struct timer_list       idle_timer;     /* L: worker idle timeout */
162         struct timer_list       mayday_timer;   /* L: SOS timer for dworkers */
163 
164         struct ida              worker_ida;     /* L: for worker IDs */
165 
166         struct task_struct      *trustee;       /* L: for gcwq shutdown */
167         unsigned int            trustee_state;  /* L: trustee state */
168         wait_queue_head_t       trustee_wait;   /* trustee wait */
169         struct worker           *first_idle;    /* L: first idle worker */
170 } ____cacheline_aligned_in_smp;
171 
172 /*
173  * The per-CPU workqueue.  The lower WORK_STRUCT_FLAG_BITS of
174  * work_struct->data are used for flags and thus cwqs need to be
175  * aligned at two's power of the number of flag bits.
176  */
177 struct cpu_workqueue_struct {
178         struct global_cwq       *gcwq;          /* I: the associated gcwq */
179         struct workqueue_struct *wq;            /* I: the owning workqueue */
180         int                     work_color;     /* L: current color */
181         int                     flush_color;    /* L: flushing color */
182         int                     nr_in_flight[WORK_NR_COLORS];
183                                                 /* L: nr of in_flight works */
184         int                     nr_active;      /* L: nr of active works */
185         int                     max_active;     /* L: max active works */
186         struct list_head        delayed_works;  /* L: delayed works */
187 };
188 
189 /*
190  * Structure used to wait for workqueue flush.
191  */
192 struct wq_flusher {
193         struct list_head        list;           /* F: list of flushers */
194         int                     flush_color;    /* F: flush color waiting for */
195         struct completion       done;           /* flush completion */
196 };
197 
198 /*
199  * All cpumasks are assumed to be always set on UP and thus can't be
200  * used to determine whether there's something to be done.
201  */
202 #ifdef CONFIG_SMP
203 typedef cpumask_var_t mayday_mask_t;
204 #define mayday_test_and_set_cpu(cpu, mask)      \
205         cpumask_test_and_set_cpu((cpu), (mask))
206 #define mayday_clear_cpu(cpu, mask)             cpumask_clear_cpu((cpu), (mask))
207 #define for_each_mayday_cpu(cpu, mask)          for_each_cpu((cpu), (mask))
208 #define alloc_mayday_mask(maskp, gfp)           zalloc_cpumask_var((maskp), (gfp))
209 #define free_mayday_mask(mask)                  free_cpumask_var((mask))
210 #else
211 typedef unsigned long mayday_mask_t;
212 #define mayday_test_and_set_cpu(cpu, mask)      test_and_set_bit(0, &(mask))
213 #define mayday_clear_cpu(cpu, mask)             clear_bit(0, &(mask))
214 #define for_each_mayday_cpu(cpu, mask)          if ((cpu) = 0, (mask))
215 #define alloc_mayday_mask(maskp, gfp)           true
216 #define free_mayday_mask(mask)                  do { } while (0)
217 #endif
218 
219 /*
220  * The externally visible workqueue abstraction is an array of
221  * per-CPU workqueues:
222  */
223 struct workqueue_struct {
224         unsigned int            flags;          /* I: WQ_* flags */
225         union {
226                 struct cpu_workqueue_struct __percpu    *pcpu;
227                 struct cpu_workqueue_struct             *single;
228                 unsigned long                           v;
229         } cpu_wq;                               /* I: cwq's */
230         struct list_head        list;           /* W: list of all workqueues */
231 
232         struct mutex            flush_mutex;    /* protects wq flushing */
233         int                     work_color;     /* F: current work color */
234         int                     flush_color;    /* F: current flush color */
235         atomic_t                nr_cwqs_to_flush; /* flush in progress */
236         struct wq_flusher       *first_flusher; /* F: first flusher */
237         struct list_head        flusher_queue;  /* F: flush waiters */
238         struct list_head        flusher_overflow; /* F: flush overflow list */
239 
240         mayday_mask_t           mayday_mask;    /* cpus requesting rescue */
241         struct worker           *rescuer;       /* I: rescue worker */
242 
243         int                     saved_max_active; /* W: saved cwq max_active */
244         const char              *name;          /* I: workqueue name */
245 #ifdef CONFIG_LOCKDEP
246         struct lockdep_map      lockdep_map;
247 #endif
248 };
249 
250 struct workqueue_struct *system_wq __read_mostly;
251 struct workqueue_struct *system_long_wq __read_mostly;
252 struct workqueue_struct *system_nrt_wq __read_mostly;
253 struct workqueue_struct *system_unbound_wq __read_mostly;
254 struct workqueue_struct *system_freezable_wq __read_mostly;
255 struct workqueue_struct *system_nrt_freezable_wq __read_mostly;
256 EXPORT_SYMBOL_GPL(system_wq);
257 EXPORT_SYMBOL_GPL(system_long_wq);
258 EXPORT_SYMBOL_GPL(system_nrt_wq);
259 EXPORT_SYMBOL_GPL(system_unbound_wq);
260 EXPORT_SYMBOL_GPL(system_freezable_wq);
261 EXPORT_SYMBOL_GPL(system_nrt_freezable_wq);
262 
263 #define CREATE_TRACE_POINTS
264 #include <trace/events/workqueue.h>
265 
266 #define for_each_busy_worker(worker, i, pos, gcwq)                      \
267         for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++)                     \
268                 hlist_for_each_entry(worker, pos, &gcwq->busy_hash[i], hentry)
269 
270 static inline int __next_gcwq_cpu(int cpu, const struct cpumask *mask,
271                                   unsigned int sw)
272 {
273         if (cpu < nr_cpu_ids) {
274                 if (sw & 1) {
275                         cpu = cpumask_next(cpu, mask);
276                         if (cpu < nr_cpu_ids)
277                                 return cpu;
278                 }
279                 if (sw & 2)
280                         return WORK_CPU_UNBOUND;
281         }
282         return WORK_CPU_NONE;
283 }
284 
285 static inline int __next_wq_cpu(int cpu, const struct cpumask *mask,
286                                 struct workqueue_struct *wq)
287 {
288         return __next_gcwq_cpu(cpu, mask, !(wq->flags & WQ_UNBOUND) ? 1 : 2);
289 }
290 
291 /*
292  * CPU iterators
293  *
294  * An extra gcwq is defined for an invalid cpu number
295  * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any
296  * specific CPU.  The following iterators are similar to
297  * for_each_*_cpu() iterators but also considers the unbound gcwq.
298  *
299  * for_each_gcwq_cpu()          : possible CPUs + WORK_CPU_UNBOUND
300  * for_each_online_gcwq_cpu()   : online CPUs + WORK_CPU_UNBOUND
301  * for_each_cwq_cpu()           : possible CPUs for bound workqueues,
302  *                                WORK_CPU_UNBOUND for unbound workqueues
303  */
304 #define for_each_gcwq_cpu(cpu)                                          \
305         for ((cpu) = __next_gcwq_cpu(-1, cpu_possible_mask, 3);         \
306              (cpu) < WORK_CPU_NONE;                                     \
307              (cpu) = __next_gcwq_cpu((cpu), cpu_possible_mask, 3))
308 
309 #define for_each_online_gcwq_cpu(cpu)                                   \
310         for ((cpu) = __next_gcwq_cpu(-1, cpu_online_mask, 3);           \
311              (cpu) < WORK_CPU_NONE;                                     \
312              (cpu) = __next_gcwq_cpu((cpu), cpu_online_mask, 3))
313 
314 #define for_each_cwq_cpu(cpu, wq)                                       \
315         for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, (wq));        \
316              (cpu) < WORK_CPU_NONE;                                     \
317              (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, (wq)))
318 
319 #ifdef CONFIG_DEBUG_OBJECTS_WORK
320 
321 static struct debug_obj_descr work_debug_descr;
322 
323 static void *work_debug_hint(void *addr)
324 {
325         return ((struct work_struct *) addr)->func;
326 }
327 
328 /*
329  * fixup_init is called when:
330  * - an active object is initialized
331  */
332 static int work_fixup_init(void *addr, enum debug_obj_state state)
333 {
334         struct work_struct *work = addr;
335 
336         switch (state) {
337         case ODEBUG_STATE_ACTIVE:
338                 cancel_work_sync(work);
339                 debug_object_init(work, &work_debug_descr);
340                 return 1;
341         default:
342                 return 0;
343         }
344 }
345 
346 /*
347  * fixup_activate is called when:
348  * - an active object is activated
349  * - an unknown object is activated (might be a statically initialized object)
350  */
351 static int work_fixup_activate(void *addr, enum debug_obj_state state)
352 {
353         struct work_struct *work = addr;
354 
355         switch (state) {
356 
357         case ODEBUG_STATE_NOTAVAILABLE:
358                 /*
359                  * This is not really a fixup. The work struct was
360                  * statically initialized. We just make sure that it
361                  * is tracked in the object tracker.
362                  */
363                 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
364                         debug_object_init(work, &work_debug_descr);
365                         debug_object_activate(work, &work_debug_descr);
366                         return 0;
367                 }
368                 WARN_ON_ONCE(1);
369                 return 0;
370 
371         case ODEBUG_STATE_ACTIVE:
372                 WARN_ON(1);
373 
374         default:
375                 return 0;
376         }
377 }
378 
379 /*
380  * fixup_free is called when:
381  * - an active object is freed
382  */
383 static int work_fixup_free(void *addr, enum debug_obj_state state)
384 {
385         struct work_struct *work = addr;
386 
387         switch (state) {
388         case ODEBUG_STATE_ACTIVE:
389                 cancel_work_sync(work);
390                 debug_object_free(work, &work_debug_descr);
391                 return 1;
392         default:
393                 return 0;
394         }
395 }
396 
397 static struct debug_obj_descr work_debug_descr = {
398         .name           = "work_struct",
399         .debug_hint     = work_debug_hint,
400         .fixup_init     = work_fixup_init,
401         .fixup_activate = work_fixup_activate,
402         .fixup_free     = work_fixup_free,
403 };
404 
405 static inline void debug_work_activate(struct work_struct *work)
406 {
407         debug_object_activate(work, &work_debug_descr);
408 }
409 
410 static inline void debug_work_deactivate(struct work_struct *work)
411 {
412         debug_object_deactivate(work, &work_debug_descr);
413 }
414 
415 void __init_work(struct work_struct *work, int onstack)
416 {
417         if (onstack)
418                 debug_object_init_on_stack(work, &work_debug_descr);
419         else
420                 debug_object_init(work, &work_debug_descr);
421 }
422 EXPORT_SYMBOL_GPL(__init_work);
423 
424 void destroy_work_on_stack(struct work_struct *work)
425 {
426         debug_object_free(work, &work_debug_descr);
427 }
428 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
429 
430 #else
431 static inline void debug_work_activate(struct work_struct *work) { }
432 static inline void debug_work_deactivate(struct work_struct *work) { }
433 #endif
434 
435 /* Serializes the accesses to the list of workqueues. */
436 static DEFINE_SPINLOCK(workqueue_lock);
437 static LIST_HEAD(workqueues);
438 static bool workqueue_freezing;         /* W: have wqs started freezing? */
439 
440 /*
441  * The almighty global cpu workqueues.  nr_running is the only field
442  * which is expected to be used frequently by other cpus via
443  * try_to_wake_up().  Put it in a separate cacheline.
444  */
445 static DEFINE_PER_CPU(struct global_cwq, global_cwq);
446 static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t, gcwq_nr_running);
447 
448 /*
449  * Global cpu workqueue and nr_running counter for unbound gcwq.  The
450  * gcwq is always online, has GCWQ_DISASSOCIATED set, and all its
451  * workers have WORKER_UNBOUND set.
452  */
453 static struct global_cwq unbound_global_cwq;
454 static atomic_t unbound_gcwq_nr_running = ATOMIC_INIT(0);       /* always 0 */
455 
456 static int worker_thread(void *__worker);
457 
458 static struct global_cwq *get_gcwq(unsigned int cpu)
459 {
460         if (cpu != WORK_CPU_UNBOUND)
461                 return &per_cpu(global_cwq, cpu);
462         else
463                 return &unbound_global_cwq;
464 }
465 
466 static atomic_t *get_gcwq_nr_running(unsigned int cpu)
467 {
468         if (cpu != WORK_CPU_UNBOUND)
469                 return &per_cpu(gcwq_nr_running, cpu);
470         else
471                 return &unbound_gcwq_nr_running;
472 }
473 
474 static struct cpu_workqueue_struct *get_cwq(unsigned int cpu,
475                                             struct workqueue_struct *wq)
476 {
477         if (!(wq->flags & WQ_UNBOUND)) {
478                 if (likely(cpu < nr_cpu_ids)) {
479 #ifdef CONFIG_SMP
480                         return per_cpu_ptr(wq->cpu_wq.pcpu, cpu);
481 #else
482                         return wq->cpu_wq.single;
483 #endif
484                 }
485         } else if (likely(cpu == WORK_CPU_UNBOUND))
486                 return wq->cpu_wq.single;
487         return NULL;
488 }
489 
490 static unsigned int work_color_to_flags(int color)
491 {
492         return color << WORK_STRUCT_COLOR_SHIFT;
493 }
494 
495 static int get_work_color(struct work_struct *work)
496 {
497         return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
498                 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
499 }
500 
501 static int work_next_color(int color)
502 {
503         return (color + 1) % WORK_NR_COLORS;
504 }
505 
506 /*
507  * A work's data points to the cwq with WORK_STRUCT_CWQ set while the
508  * work is on queue.  Once execution starts, WORK_STRUCT_CWQ is
509  * cleared and the work data contains the cpu number it was last on.
510  *
511  * set_work_{cwq|cpu}() and clear_work_data() can be used to set the
512  * cwq, cpu or clear work->data.  These functions should only be
513  * called while the work is owned - ie. while the PENDING bit is set.
514  *
515  * get_work_[g]cwq() can be used to obtain the gcwq or cwq
516  * corresponding to a work.  gcwq is available once the work has been
517  * queued anywhere after initialization.  cwq is available only from
518  * queueing until execution starts.
519  */
520 static inline void set_work_data(struct work_struct *work, unsigned long data,
521                                  unsigned long flags)
522 {
523         BUG_ON(!work_pending(work));
524         atomic_long_set(&work->data, data | flags | work_static(work));
525 }
526 
527 static void set_work_cwq(struct work_struct *work,
528                          struct cpu_workqueue_struct *cwq,
529                          unsigned long extra_flags)
530 {
531         set_work_data(work, (unsigned long)cwq,
532                       WORK_STRUCT_PENDING | WORK_STRUCT_CWQ | extra_flags);
533 }
534 
535 static void set_work_cpu(struct work_struct *work, unsigned int cpu)
536 {
537         set_work_data(work, cpu << WORK_STRUCT_FLAG_BITS, WORK_STRUCT_PENDING);
538 }
539 
540 static void clear_work_data(struct work_struct *work)
541 {
542         set_work_data(work, WORK_STRUCT_NO_CPU, 0);
543 }
544 
545 static struct cpu_workqueue_struct *get_work_cwq(struct work_struct *work)
546 {
547         unsigned long data = atomic_long_read(&work->data);
548 
549         if (data & WORK_STRUCT_CWQ)
550                 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
551         else
552                 return NULL;
553 }
554 
555 static struct global_cwq *get_work_gcwq(struct work_struct *work)
556 {
557         unsigned long data = atomic_long_read(&work->data);
558         unsigned int cpu;
559 
560         if (data & WORK_STRUCT_CWQ)
561                 return ((struct cpu_workqueue_struct *)
562                         (data & WORK_STRUCT_WQ_DATA_MASK))->gcwq;
563 
564         cpu = data >> WORK_STRUCT_FLAG_BITS;
565         if (cpu == WORK_CPU_NONE)
566                 return NULL;
567 
568         BUG_ON(cpu >= nr_cpu_ids && cpu != WORK_CPU_UNBOUND);
569         return get_gcwq(cpu);
570 }
571 
572 /*
573  * Policy functions.  These define the policies on how the global
574  * worker pool is managed.  Unless noted otherwise, these functions
575  * assume that they're being called with gcwq->lock held.
576  */
577 
578 static bool __need_more_worker(struct global_cwq *gcwq)
579 {
580         return !atomic_read(get_gcwq_nr_running(gcwq->cpu)) ||
581                 gcwq->flags & GCWQ_HIGHPRI_PENDING;
582 }
583 
584 /*
585  * Need to wake up a worker?  Called from anything but currently
586  * running workers.
587  */
588 static bool need_more_worker(struct global_cwq *gcwq)
589 {
590         return !list_empty(&gcwq->worklist) && __need_more_worker(gcwq);
591 }
592 
593 /* Can I start working?  Called from busy but !running workers. */
594 static bool may_start_working(struct global_cwq *gcwq)
595 {
596         return gcwq->nr_idle;
597 }
598 
599 /* Do I need to keep working?  Called from currently running workers. */
600 static bool keep_working(struct global_cwq *gcwq)
601 {
602         atomic_t *nr_running = get_gcwq_nr_running(gcwq->cpu);
603 
604         return !list_empty(&gcwq->worklist) &&
605                 (atomic_read(nr_running) <= 1 ||
606                  gcwq->flags & GCWQ_HIGHPRI_PENDING);
607 }
608 
609 /* Do we need a new worker?  Called from manager. */
610 static bool need_to_create_worker(struct global_cwq *gcwq)
611 {
612         return need_more_worker(gcwq) && !may_start_working(gcwq);
613 }
614 
615 /* Do I need to be the manager? */
616 static bool need_to_manage_workers(struct global_cwq *gcwq)
617 {
618         return need_to_create_worker(gcwq) || gcwq->flags & GCWQ_MANAGE_WORKERS;
619 }
620 
621 /* Do we have too many workers and should some go away? */
622 static bool too_many_workers(struct global_cwq *gcwq)
623 {
624         bool managing = gcwq->flags & GCWQ_MANAGING_WORKERS;
625         int nr_idle = gcwq->nr_idle + managing; /* manager is considered idle */
626         int nr_busy = gcwq->nr_workers - nr_idle;
627 
628         return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
629 }
630 
631 /*
632  * Wake up functions.
633  */
634 
635 /* Return the first worker.  Safe with preemption disabled */
636 static struct worker *first_worker(struct global_cwq *gcwq)
637 {
638         if (unlikely(list_empty(&gcwq->idle_list)))
639                 return NULL;
640 
641         return list_first_entry(&gcwq->idle_list, struct worker, entry);
642 }
643 
644 /**
645  * wake_up_worker - wake up an idle worker
646  * @gcwq: gcwq to wake worker for
647  *
648  * Wake up the first idle worker of @gcwq.
649  *
650  * CONTEXT:
651  * spin_lock_irq(gcwq->lock).
652  */
653 static void wake_up_worker(struct global_cwq *gcwq)
654 {
655         struct worker *worker = first_worker(gcwq);
656 
657         if (likely(worker))
658                 wake_up_process(worker->task);
659 }
660 
661 /**
662  * wq_worker_waking_up - a worker is waking up
663  * @task: task waking up
664  * @cpu: CPU @task is waking up to
665  *
666  * This function is called during try_to_wake_up() when a worker is
667  * being awoken.
668  *
669  * CONTEXT:
670  * spin_lock_irq(rq->lock)
671  */
672 void wq_worker_waking_up(struct task_struct *task, unsigned int cpu)
673 {
674         struct worker *worker = kthread_data(task);
675 
676         if (!(worker->flags & WORKER_NOT_RUNNING))
677                 atomic_inc(get_gcwq_nr_running(cpu));
678 }
679 
680 /**
681  * wq_worker_sleeping - a worker is going to sleep
682  * @task: task going to sleep
683  * @cpu: CPU in question, must be the current CPU number
684  *
685  * This function is called during schedule() when a busy worker is
686  * going to sleep.  Worker on the same cpu can be woken up by
687  * returning pointer to its task.
688  *
689  * CONTEXT:
690  * spin_lock_irq(rq->lock)
691  *
692  * RETURNS:
693  * Worker task on @cpu to wake up, %NULL if none.
694  */
695 struct task_struct *wq_worker_sleeping(struct task_struct *task,
696                                        unsigned int cpu)
697 {
698         struct worker *worker = kthread_data(task), *to_wakeup = NULL;
699         struct global_cwq *gcwq = get_gcwq(cpu);
700         atomic_t *nr_running = get_gcwq_nr_running(cpu);
701 
702         if (worker->flags & WORKER_NOT_RUNNING)
703                 return NULL;
704 
705         /* this can only happen on the local cpu */
706         BUG_ON(cpu != raw_smp_processor_id());
707 
708         /*
709          * The counterpart of the following dec_and_test, implied mb,
710          * worklist not empty test sequence is in insert_work().
711          * Please read comment there.
712          *
713          * NOT_RUNNING is clear.  This means that trustee is not in
714          * charge and we're running on the local cpu w/ rq lock held
715          * and preemption disabled, which in turn means that none else
716          * could be manipulating idle_list, so dereferencing idle_list
717          * without gcwq lock is safe.
718          */
719         if (atomic_dec_and_test(nr_running) && !list_empty(&gcwq->worklist))
720                 to_wakeup = first_worker(gcwq);
721         return to_wakeup ? to_wakeup->task : NULL;
722 }
723 
724 /**
725  * worker_set_flags - set worker flags and adjust nr_running accordingly
726  * @worker: self
727  * @flags: flags to set
728  * @wakeup: wakeup an idle worker if necessary
729  *
730  * Set @flags in @worker->flags and adjust nr_running accordingly.  If
731  * nr_running becomes zero and @wakeup is %true, an idle worker is
732  * woken up.
733  *
734  * CONTEXT:
735  * spin_lock_irq(gcwq->lock)
736  */
737 static inline void worker_set_flags(struct worker *worker, unsigned int flags,
738                                     bool wakeup)
739 {
740         struct global_cwq *gcwq = worker->gcwq;
741 
742         WARN_ON_ONCE(worker->task != current);
743 
744         /*
745          * If transitioning into NOT_RUNNING, adjust nr_running and
746          * wake up an idle worker as necessary if requested by
747          * @wakeup.
748          */
749         if ((flags & WORKER_NOT_RUNNING) &&
750             !(worker->flags & WORKER_NOT_RUNNING)) {
751                 atomic_t *nr_running = get_gcwq_nr_running(gcwq->cpu);
752 
753                 if (wakeup) {
754                         if (atomic_dec_and_test(nr_running) &&
755                             !list_empty(&gcwq->worklist))
756                                 wake_up_worker(gcwq);
757                 } else
758                         atomic_dec(nr_running);
759         }
760 
761         worker->flags |= flags;
762 }
763 
764 /**
765  * worker_clr_flags - clear worker flags and adjust nr_running accordingly
766  * @worker: self
767  * @flags: flags to clear
768  *
769  * Clear @flags in @worker->flags and adjust nr_running accordingly.
770  *
771  * CONTEXT:
772  * spin_lock_irq(gcwq->lock)
773  */
774 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
775 {
776         struct global_cwq *gcwq = worker->gcwq;
777         unsigned int oflags = worker->flags;
778 
779         WARN_ON_ONCE(worker->task != current);
780 
781         worker->flags &= ~flags;
782 
783         /*
784          * If transitioning out of NOT_RUNNING, increment nr_running.  Note
785          * that the nested NOT_RUNNING is not a noop.  NOT_RUNNING is mask
786          * of multiple flags, not a single flag.
787          */
788         if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
789                 if (!(worker->flags & WORKER_NOT_RUNNING))
790                         atomic_inc(get_gcwq_nr_running(gcwq->cpu));
791 }
792 
793 /**
794  * busy_worker_head - return the busy hash head for a work
795  * @gcwq: gcwq of interest
796  * @work: work to be hashed
797  *
798  * Return hash head of @gcwq for @work.
799  *
800  * CONTEXT:
801  * spin_lock_irq(gcwq->lock).
802  *
803  * RETURNS:
804  * Pointer to the hash head.
805  */
806 static struct hlist_head *busy_worker_head(struct global_cwq *gcwq,
807                                            struct work_struct *work)
808 {
809         const int base_shift = ilog2(sizeof(struct work_struct));
810         unsigned long v = (unsigned long)work;
811 
812         /* simple shift and fold hash, do we need something better? */
813         v >>= base_shift;
814         v += v >> BUSY_WORKER_HASH_ORDER;
815         v &= BUSY_WORKER_HASH_MASK;
816 
817         return &gcwq->busy_hash[v];
818 }
819 
820 /**
821  * __find_worker_executing_work - find worker which is executing a work
822  * @gcwq: gcwq of interest
823  * @bwh: hash head as returned by busy_worker_head()
824  * @work: work to find worker for
825  *
826  * Find a worker which is executing @work on @gcwq.  @bwh should be
827  * the hash head obtained by calling busy_worker_head() with the same
828  * work.
829  *
830  * CONTEXT:
831  * spin_lock_irq(gcwq->lock).
832  *
833  * RETURNS:
834  * Pointer to worker which is executing @work if found, NULL
835  * otherwise.
836  */
837 static struct worker *__find_worker_executing_work(struct global_cwq *gcwq,
838                                                    struct hlist_head *bwh,
839                                                    struct work_struct *work)
840 {
841         struct worker *worker;
842         struct hlist_node *tmp;
843 
844         hlist_for_each_entry(worker, tmp, bwh, hentry)
845                 if (worker->current_work == work)
846                         return worker;
847         return NULL;
848 }
849 
850 /**
851  * find_worker_executing_work - find worker which is executing a work
852  * @gcwq: gcwq of interest
853  * @work: work to find worker for
854  *
855  * Find a worker which is executing @work on @gcwq.  This function is
856  * identical to __find_worker_executing_work() except that this
857  * function calculates @bwh itself.
858  *
859  * CONTEXT:
860  * spin_lock_irq(gcwq->lock).
861  *
862  * RETURNS:
863  * Pointer to worker which is executing @work if found, NULL
864  * otherwise.
865  */
866 static struct worker *find_worker_executing_work(struct global_cwq *gcwq,
867                                                  struct work_struct *work)
868 {
869         return __find_worker_executing_work(gcwq, busy_worker_head(gcwq, work),
870                                             work);
871 }
872 
873 /**
874  * gcwq_determine_ins_pos - find insertion position
875  * @gcwq: gcwq of interest
876  * @cwq: cwq a work is being queued for
877  *
878  * A work for @cwq is about to be queued on @gcwq, determine insertion
879  * position for the work.  If @cwq is for HIGHPRI wq, the work is
880  * queued at the head of the queue but in FIFO order with respect to
881  * other HIGHPRI works; otherwise, at the end of the queue.  This
882  * function also sets GCWQ_HIGHPRI_PENDING flag to hint @gcwq that
883  * there are HIGHPRI works pending.
884  *
885  * CONTEXT:
886  * spin_lock_irq(gcwq->lock).
887  *
888  * RETURNS:
889  * Pointer to inserstion position.
890  */
891 static inline struct list_head *gcwq_determine_ins_pos(struct global_cwq *gcwq,
892                                                struct cpu_workqueue_struct *cwq)
893 {
894         struct work_struct *twork;
895 
896         if (likely(!(cwq->wq->flags & WQ_HIGHPRI)))
897                 return &gcwq->worklist;
898 
899         list_for_each_entry(twork, &gcwq->worklist, entry) {
900                 struct cpu_workqueue_struct *tcwq = get_work_cwq(twork);
901 
902                 if (!(tcwq->wq->flags & WQ_HIGHPRI))
903                         break;
904         }
905 
906         gcwq->flags |= GCWQ_HIGHPRI_PENDING;
907         return &twork->entry;
908 }
909 
910 /**
911  * insert_work - insert a work into gcwq
912  * @cwq: cwq @work belongs to
913  * @work: work to insert
914  * @head: insertion point
915  * @extra_flags: extra WORK_STRUCT_* flags to set
916  *
917  * Insert @work which belongs to @cwq into @gcwq after @head.
918  * @extra_flags is or'd to work_struct flags.
919  *
920  * CONTEXT:
921  * spin_lock_irq(gcwq->lock).
922  */
923 static void insert_work(struct cpu_workqueue_struct *cwq,
924                         struct work_struct *work, struct list_head *head,
925                         unsigned int extra_flags)
926 {
927         struct global_cwq *gcwq = cwq->gcwq;
928 
929         /* we own @work, set data and link */
930         set_work_cwq(work, cwq, extra_flags);
931 
932         /*
933          * Ensure that we get the right work->data if we see the
934          * result of list_add() below, see try_to_grab_pending().
935          */
936         smp_wmb();
937 
938         list_add_tail(&work->entry, head);
939 
940         /*
941          * Ensure either worker_sched_deactivated() sees the above
942          * list_add_tail() or we see zero nr_running to avoid workers
943          * lying around lazily while there are works to be processed.
944          */
945         smp_mb();
946 
947         if (__need_more_worker(gcwq))
948                 wake_up_worker(gcwq);
949 }
950 
951 /*
952  * Test whether @work is being queued from another work executing on the
953  * same workqueue.  This is rather expensive and should only be used from
954  * cold paths.
955  */
956 static bool is_chained_work(struct workqueue_struct *wq)
957 {
958         unsigned long flags;
959         unsigned int cpu;
960 
961         for_each_gcwq_cpu(cpu) {
962                 struct global_cwq *gcwq = get_gcwq(cpu);
963                 struct worker *worker;
964                 struct hlist_node *pos;
965                 int i;
966 
967                 spin_lock_irqsave(&gcwq->lock, flags);
968                 for_each_busy_worker(worker, i, pos, gcwq) {
969                         if (worker->task != current)
970                                 continue;
971                         spin_unlock_irqrestore(&gcwq->lock, flags);
972                         /*
973                          * I'm @worker, no locking necessary.  See if @work
974                          * is headed to the same workqueue.
975                          */
976                         return worker->current_cwq->wq == wq;
977                 }
978                 spin_unlock_irqrestore(&gcwq->lock, flags);
979         }
980         return false;
981 }
982 
983 static void __queue_work(unsigned int cpu, struct workqueue_struct *wq,
984                          struct work_struct *work)
985 {
986         struct global_cwq *gcwq;
987         struct cpu_workqueue_struct *cwq;
988         struct list_head *worklist;
989         unsigned int work_flags;
990         unsigned long flags;
991 
992         debug_work_activate(work);
993 
994         /* if dying, only works from the same workqueue are allowed */
995         if (unlikely(wq->flags & WQ_DYING) &&
996             WARN_ON_ONCE(!is_chained_work(wq)))
997                 return;
998 
999         /* determine gcwq to use */
1000         if (!(wq->flags & WQ_UNBOUND)) {
1001                 struct global_cwq *last_gcwq;
1002 
1003                 if (unlikely(cpu == WORK_CPU_UNBOUND))
1004                         cpu = raw_smp_processor_id();
1005 
1006                 /*
1007                  * It's multi cpu.  If @wq is non-reentrant and @work
1008                  * was previously on a different cpu, it might still
1009                  * be running there, in which case the work needs to
1010                  * be queued on that cpu to guarantee non-reentrance.
1011                  */
1012                 gcwq = get_gcwq(cpu);
1013                 if (wq->flags & WQ_NON_REENTRANT &&
1014                     (last_gcwq = get_work_gcwq(work)) && last_gcwq != gcwq) {
1015                         struct worker *worker;
1016 
1017                         spin_lock_irqsave(&last_gcwq->lock, flags);
1018 
1019                         worker = find_worker_executing_work(last_gcwq, work);
1020 
1021                         if (worker && worker->current_cwq->wq == wq)
1022                                 gcwq = last_gcwq;
1023                         else {
1024                                 /* meh... not running there, queue here */
1025                                 spin_unlock_irqrestore(&last_gcwq->lock, flags);
1026                                 spin_lock_irqsave(&gcwq->lock, flags);
1027                         }
1028                 } else
1029                         spin_lock_irqsave(&gcwq->lock, flags);
1030         } else {
1031                 gcwq = get_gcwq(WORK_CPU_UNBOUND);
1032                 spin_lock_irqsave(&gcwq->lock, flags);
1033         }
1034 
1035         /* gcwq determined, get cwq and queue */
1036         cwq = get_cwq(gcwq->cpu, wq);
1037         trace_workqueue_queue_work(cpu, cwq, work);
1038 
1039         BUG_ON(!list_empty(&work->entry));
1040 
1041         cwq->nr_in_flight[cwq->work_color]++;
1042         work_flags = work_color_to_flags(cwq->work_color);
1043 
1044         if (likely(cwq->nr_active < cwq->max_active)) {
1045                 trace_workqueue_activate_work(work);
1046                 cwq->nr_active++;
1047                 worklist = gcwq_determine_ins_pos(gcwq, cwq);
1048         } else {
1049                 work_flags |= WORK_STRUCT_DELAYED;
1050                 worklist = &cwq->delayed_works;
1051         }
1052 
1053         insert_work(cwq, work, worklist, work_flags);
1054 
1055         spin_unlock_irqrestore(&gcwq->lock, flags);
1056 }
1057 
1058 /**
1059  * queue_work - queue work on a workqueue
1060  * @wq: workqueue to use
1061  * @work: work to queue
1062  *
1063  * Returns 0 if @work was already on a queue, non-zero otherwise.
1064  *
1065  * We queue the work to the CPU on which it was submitted, but if the CPU dies
1066  * it can be processed by another CPU.
1067  */
1068 int queue_work(struct workqueue_struct *wq, struct work_struct *work)
1069 {
1070         int ret;
1071 
1072         ret = queue_work_on(get_cpu(), wq, work);
1073         put_cpu();
1074 
1075         return ret;
1076 }
1077 EXPORT_SYMBOL_GPL(queue_work);
1078 
1079 /**
1080  * queue_work_on - queue work on specific cpu
1081  * @cpu: CPU number to execute work on
1082  * @wq: workqueue to use
1083  * @work: work to queue
1084  *
1085  * Returns 0 if @work was already on a queue, non-zero otherwise.
1086  *
1087  * We queue the work to a specific CPU, the caller must ensure it
1088  * can't go away.
1089  */
1090 int
1091 queue_work_on(int cpu, struct workqueue_struct *wq, struct work_struct *work)
1092 {
1093         int ret = 0;
1094 
1095         if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1096                 __queue_work(cpu, wq, work);
1097                 ret = 1;
1098         }
1099         return ret;
1100 }
1101 EXPORT_SYMBOL_GPL(queue_work_on);
1102 
1103 static void delayed_work_timer_fn(unsigned long __data)
1104 {
1105         struct delayed_work *dwork = (struct delayed_work *)__data;
1106         struct cpu_workqueue_struct *cwq = get_work_cwq(&dwork->work);
1107 
1108         __queue_work(smp_processor_id(), cwq->wq, &dwork->work);
1109 }
1110 
1111 /**
1112  * queue_delayed_work - queue work on a workqueue after delay
1113  * @wq: workqueue to use
1114  * @dwork: delayable work to queue
1115  * @delay: number of jiffies to wait before queueing
1116  *
1117  * Returns 0 if @work was already on a queue, non-zero otherwise.
1118  */
1119 int queue_delayed_work(struct workqueue_struct *wq,
1120                         struct delayed_work *dwork, unsigned long delay)
1121 {
1122         if (delay == 0)
1123                 return queue_work(wq, &dwork->work);
1124 
1125         return queue_delayed_work_on(-1, wq, dwork, delay);
1126 }
1127 EXPORT_SYMBOL_GPL(queue_delayed_work);
1128 
1129 /**
1130  * queue_delayed_work_on - queue work on specific CPU after delay
1131  * @cpu: CPU number to execute work on
1132  * @wq: workqueue to use
1133  * @dwork: work to queue
1134  * @delay: number of jiffies to wait before queueing
1135  *
1136  * Returns 0 if @work was already on a queue, non-zero otherwise.
1137  */
1138 int queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1139                         struct delayed_work *dwork, unsigned long delay)
1140 {
1141         int ret = 0;
1142         struct timer_list *timer = &dwork->timer;
1143         struct work_struct *work = &dwork->work;
1144 
1145         if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1146                 unsigned int lcpu;
1147 
1148                 WARN_ON_ONCE(timer_pending(timer));
1149                 WARN_ON_ONCE(!list_empty(&work->entry));
1150 
1151                 timer_stats_timer_set_start_info(&dwork->timer);
1152 
1153                 /*
1154                  * This stores cwq for the moment, for the timer_fn.
1155                  * Note that the work's gcwq is preserved to allow
1156                  * reentrance detection for delayed works.
1157                  */
1158                 if (!(wq->flags & WQ_UNBOUND)) {
1159                         struct global_cwq *gcwq = get_work_gcwq(work);
1160 
1161                         if (gcwq && gcwq->cpu != WORK_CPU_UNBOUND)
1162                                 lcpu = gcwq->cpu;
1163                         else
1164                                 lcpu = raw_smp_processor_id();
1165                 } else
1166                         lcpu = WORK_CPU_UNBOUND;
1167 
1168                 set_work_cwq(work, get_cwq(lcpu, wq), 0);
1169 
1170                 timer->expires = jiffies + delay;
1171                 timer->data = (unsigned long)dwork;
1172                 timer->function = delayed_work_timer_fn;
1173 
1174                 if (unlikely(cpu >= 0))
1175                         add_timer_on(timer, cpu);
1176                 else
1177                         add_timer(timer);
1178                 ret = 1;
1179         }
1180         return ret;
1181 }
1182 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
1183 
1184 /**
1185  * worker_enter_idle - enter idle state
1186  * @worker: worker which is entering idle state
1187  *
1188  * @worker is entering idle state.  Update stats and idle timer if
1189  * necessary.
1190  *
1191  * LOCKING:
1192  * spin_lock_irq(gcwq->lock).
1193  */
1194 static void worker_enter_idle(struct worker *worker)
1195 {
1196         struct global_cwq *gcwq = worker->gcwq;
1197 
1198         BUG_ON(worker->flags & WORKER_IDLE);
1199         BUG_ON(!list_empty(&worker->entry) &&
1200                (worker->hentry.next || worker->hentry.pprev));
1201 
1202         /* can't use worker_set_flags(), also called from start_worker() */
1203         worker->flags |= WORKER_IDLE;
1204         gcwq->nr_idle++;
1205         worker->last_active = jiffies;
1206 
1207         /* idle_list is LIFO */
1208         list_add(&worker->entry, &gcwq->idle_list);
1209 
1210         if (likely(!(worker->flags & WORKER_ROGUE))) {
1211                 if (too_many_workers(gcwq) && !timer_pending(&gcwq->idle_timer))
1212                         mod_timer(&gcwq->idle_timer,
1213                                   jiffies + IDLE_WORKER_TIMEOUT);
1214         } else
1215                 wake_up_all(&gcwq->trustee_wait);
1216 
1217         /*
1218          * Sanity check nr_running.  Because trustee releases gcwq->lock
1219          * between setting %WORKER_ROGUE and zapping nr_running, the
1220          * warning may trigger spuriously.  Check iff trustee is idle.
1221          */
1222         WARN_ON_ONCE(gcwq->trustee_state == TRUSTEE_DONE &&
1223                      gcwq->nr_workers == gcwq->nr_idle &&
1224                      atomic_read(get_gcwq_nr_running(gcwq->cpu)));
1225 }
1226 
1227 /**
1228  * worker_leave_idle - leave idle state
1229  * @worker: worker which is leaving idle state
1230  *
1231  * @worker is leaving idle state.  Update stats.
1232  *
1233  * LOCKING:
1234  * spin_lock_irq(gcwq->lock).
1235  */
1236 static void worker_leave_idle(struct worker *worker)
1237 {
1238         struct global_cwq *gcwq = worker->gcwq;
1239 
1240         BUG_ON(!(worker->flags & WORKER_IDLE));
1241         worker_clr_flags(worker, WORKER_IDLE);
1242         gcwq->nr_idle--;
1243         list_del_init(&worker->entry);
1244 }
1245 
1246 /**
1247  * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock gcwq
1248  * @worker: self
1249  *
1250  * Works which are scheduled while the cpu is online must at least be
1251  * scheduled to a worker which is bound to the cpu so that if they are
1252  * flushed from cpu callbacks while cpu is going down, they are
1253  * guaranteed to execute on the cpu.
1254  *
1255  * This function is to be used by rogue workers and rescuers to bind
1256  * themselves to the target cpu and may race with cpu going down or
1257  * coming online.  kthread_bind() can't be used because it may put the
1258  * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1259  * verbatim as it's best effort and blocking and gcwq may be
1260  * [dis]associated in the meantime.
1261  *
1262  * This function tries set_cpus_allowed() and locks gcwq and verifies
1263  * the binding against GCWQ_DISASSOCIATED which is set during
1264  * CPU_DYING and cleared during CPU_ONLINE, so if the worker enters
1265  * idle state or fetches works without dropping lock, it can guarantee
1266  * the scheduling requirement described in the first paragraph.
1267  *
1268  * CONTEXT:
1269  * Might sleep.  Called without any lock but returns with gcwq->lock
1270  * held.
1271  *
1272  * RETURNS:
1273  * %true if the associated gcwq is online (@worker is successfully
1274  * bound), %false if offline.
1275  */
1276 static bool worker_maybe_bind_and_lock(struct worker *worker)
1277 __acquires(&gcwq->lock)
1278 {
1279         struct global_cwq *gcwq = worker->gcwq;
1280         struct task_struct *task = worker->task;
1281 
1282         while (true) {
1283                 /*
1284                  * The following call may fail, succeed or succeed
1285                  * without actually migrating the task to the cpu if
1286                  * it races with cpu hotunplug operation.  Verify
1287                  * against GCWQ_DISASSOCIATED.
1288                  */
1289                 if (!(gcwq->flags & GCWQ_DISASSOCIATED))
1290                         set_cpus_allowed_ptr(task, get_cpu_mask(gcwq->cpu));
1291 
1292                 spin_lock_irq(&gcwq->lock);
1293                 if (gcwq->flags & GCWQ_DISASSOCIATED)
1294                         return false;
1295                 if (task_cpu(task) == gcwq->cpu &&
1296                     cpumask_equal(&current->cpus_allowed,
1297                                   get_cpu_mask(gcwq->cpu)))
1298                         return true;
1299                 spin_unlock_irq(&gcwq->lock);
1300 
1301                 /*
1302                  * We've raced with CPU hot[un]plug.  Give it a breather
1303                  * and retry migration.  cond_resched() is required here;
1304                  * otherwise, we might deadlock against cpu_stop trying to
1305                  * bring down the CPU on non-preemptive kernel.
1306                  */
1307                 cpu_relax();
1308                 cond_resched();
1309         }
1310 }
1311 
1312 /*
1313  * Function for worker->rebind_work used to rebind rogue busy workers
1314  * to the associated cpu which is coming back online.  This is
1315  * scheduled by cpu up but can race with other cpu hotplug operations
1316  * and may be executed twice without intervening cpu down.
1317  */
1318 static void worker_rebind_fn(struct work_struct *work)
1319 {
1320         struct worker *worker = container_of(work, struct worker, rebind_work);
1321         struct global_cwq *gcwq = worker->gcwq;
1322 
1323         if (worker_maybe_bind_and_lock(worker))
1324                 worker_clr_flags(worker, WORKER_REBIND);
1325 
1326         spin_unlock_irq(&gcwq->lock);
1327 }
1328 
1329 static struct worker *alloc_worker(void)
1330 {
1331         struct worker *worker;
1332 
1333         worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1334         if (worker) {
1335                 INIT_LIST_HEAD(&worker->entry);
1336                 INIT_LIST_HEAD(&worker->scheduled);
1337                 INIT_WORK(&worker->rebind_work, worker_rebind_fn);
1338                 /* on creation a worker is in !idle && prep state */
1339                 worker->flags = WORKER_PREP;
1340         }
1341         return worker;
1342 }
1343 
1344 /**
1345  * create_worker - create a new workqueue worker
1346  * @gcwq: gcwq the new worker will belong to
1347  * @bind: whether to set affinity to @cpu or not
1348  *
1349  * Create a new worker which is bound to @gcwq.  The returned worker
1350  * can be started by calling start_worker() or destroyed using
1351  * destroy_worker().
1352  *
1353  * CONTEXT:
1354  * Might sleep.  Does GFP_KERNEL allocations.
1355  *
1356  * RETURNS:
1357  * Pointer to the newly created worker.
1358  */
1359 static struct worker *create_worker(struct global_cwq *gcwq, bool bind)
1360 {
1361         bool on_unbound_cpu = gcwq->cpu == WORK_CPU_UNBOUND;
1362         struct worker *worker = NULL;
1363         int id = -1;
1364 
1365         spin_lock_irq(&gcwq->lock);
1366         while (ida_get_new(&gcwq->worker_ida, &id)) {
1367                 spin_unlock_irq(&gcwq->lock);
1368                 if (!ida_pre_get(&gcwq->worker_ida, GFP_KERNEL))
1369                         goto fail;
1370                 spin_lock_irq(&gcwq->lock);
1371         }
1372         spin_unlock_irq(&gcwq->lock);
1373 
1374         worker = alloc_worker();
1375         if (!worker)
1376                 goto fail;
1377 
1378         worker->gcwq = gcwq;
1379         worker->id = id;
1380 
1381         if (!on_unbound_cpu)
1382                 worker->task = kthread_create_on_node(worker_thread,
1383                                                       worker,
1384                                                       cpu_to_node(gcwq->cpu),
1385                                                       "kworker/%u:%d", gcwq->cpu, id);
1386         else
1387                 worker->task = kthread_create(worker_thread, worker,
1388                                               "kworker/u:%d", id);
1389         if (IS_ERR(worker->task))
1390                 goto fail;
1391 
1392         /*
1393          * A rogue worker will become a regular one if CPU comes
1394          * online later on.  Make sure every worker has
1395          * PF_THREAD_BOUND set.
1396          */
1397         if (bind && !on_unbound_cpu)
1398                 kthread_bind(worker->task, gcwq->cpu);
1399         else {
1400                 worker->task->flags |= PF_THREAD_BOUND;
1401                 if (on_unbound_cpu)
1402                         worker->flags |= WORKER_UNBOUND;
1403         }
1404 
1405         return worker;
1406 fail:
1407         if (id >= 0) {
1408                 spin_lock_irq(&gcwq->lock);
1409                 ida_remove(&gcwq->worker_ida, id);
1410                 spin_unlock_irq(&gcwq->lock);
1411         }
1412         kfree(worker);
1413         return NULL;
1414 }
1415 
1416 /**
1417  * start_worker - start a newly created worker
1418  * @worker: worker to start
1419  *
1420  * Make the gcwq aware of @worker and start it.
1421  *
1422  * CONTEXT:
1423  * spin_lock_irq(gcwq->lock).
1424  */
1425 static void start_worker(struct worker *worker)
1426 {
1427         worker->flags |= WORKER_STARTED;
1428         worker->gcwq->nr_workers++;
1429         worker_enter_idle(worker);
1430         wake_up_process(worker->task);
1431 }
1432 
1433 /**
1434  * destroy_worker - destroy a workqueue worker
1435  * @worker: worker to be destroyed
1436  *
1437  * Destroy @worker and adjust @gcwq stats accordingly.
1438  *
1439  * CONTEXT:
1440  * spin_lock_irq(gcwq->lock) which is released and regrabbed.
1441  */
1442 static void destroy_worker(struct worker *worker)
1443 {
1444         struct global_cwq *gcwq = worker->gcwq;
1445         int id = worker->id;
1446 
1447         /* sanity check frenzy */
1448         BUG_ON(worker->current_work);
1449         BUG_ON(!list_empty(&worker->scheduled));
1450 
1451         if (worker->flags & WORKER_STARTED)
1452                 gcwq->nr_workers--;
1453         if (worker->flags & WORKER_IDLE)
1454                 gcwq->nr_idle--;
1455 
1456         list_del_init(&worker->entry);
1457         worker->flags |= WORKER_DIE;
1458 
1459         spin_unlock_irq(&gcwq->lock);
1460 
1461         kthread_stop(worker->task);
1462         kfree(worker);
1463 
1464         spin_lock_irq(&gcwq->lock);
1465         ida_remove(&gcwq->worker_ida, id);
1466 }
1467 
1468 static void idle_worker_timeout(unsigned long __gcwq)
1469 {
1470         struct global_cwq *gcwq = (void *)__gcwq;
1471 
1472         spin_lock_irq(&gcwq->lock);
1473 
1474         if (too_many_workers(gcwq)) {
1475                 struct worker *worker;
1476                 unsigned long expires;
1477 
1478                 /* idle_list is kept in LIFO order, check the last one */
1479                 worker = list_entry(gcwq->idle_list.prev, struct worker, entry);
1480                 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1481 
1482                 if (time_before(jiffies, expires))
1483                         mod_timer(&gcwq->idle_timer, expires);
1484                 else {
1485                         /* it's been idle for too long, wake up manager */
1486                         gcwq->flags |= GCWQ_MANAGE_WORKERS;
1487                         wake_up_worker(gcwq);
1488                 }
1489         }
1490 
1491         spin_unlock_irq(&gcwq->lock);
1492 }
1493 
1494 static bool send_mayday(struct work_struct *work)
1495 {
1496         struct cpu_workqueue_struct *cwq = get_work_cwq(work);
1497         struct workqueue_struct *wq = cwq->wq;
1498         unsigned int cpu;
1499 
1500         if (!(wq->flags & WQ_RESCUER))
1501                 return false;
1502 
1503         /* mayday mayday mayday */
1504         cpu = cwq->gcwq->cpu;
1505         /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */
1506         if (cpu == WORK_CPU_UNBOUND)
1507                 cpu = 0;
1508         if (!mayday_test_and_set_cpu(cpu, wq->mayday_mask))
1509                 wake_up_process(wq->rescuer->task);
1510         return true;
1511 }
1512 
1513 static void gcwq_mayday_timeout(unsigned long __gcwq)
1514 {
1515         struct global_cwq *gcwq = (void *)__gcwq;
1516         struct work_struct *work;
1517 
1518         spin_lock_irq(&gcwq->lock);
1519 
1520         if (need_to_create_worker(gcwq)) {
1521                 /*
1522                  * We've been trying to create a new worker but
1523                  * haven't been successful.  We might be hitting an
1524                  * allocation deadlock.  Send distress signals to
1525                  * rescuers.
1526                  */
1527                 list_for_each_entry(work, &gcwq->worklist, entry)
1528                         send_mayday(work);
1529         }
1530 
1531         spin_unlock_irq(&gcwq->lock);
1532 
1533         mod_timer(&gcwq->mayday_timer, jiffies + MAYDAY_INTERVAL);
1534 }
1535 
1536 /**
1537  * maybe_create_worker - create a new worker if necessary
1538  * @gcwq: gcwq to create a new worker for
1539  *
1540  * Create a new worker for @gcwq if necessary.  @gcwq is guaranteed to
1541  * have at least one idle worker on return from this function.  If
1542  * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1543  * sent to all rescuers with works scheduled on @gcwq to resolve
1544  * possible allocation deadlock.
1545  *
1546  * On return, need_to_create_worker() is guaranteed to be false and
1547  * may_start_working() true.
1548  *
1549  * LOCKING:
1550  * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1551  * multiple times.  Does GFP_KERNEL allocations.  Called only from
1552  * manager.
1553  *
1554  * RETURNS:
1555  * false if no action was taken and gcwq->lock stayed locked, true
1556  * otherwise.
1557  */
1558 static bool maybe_create_worker(struct global_cwq *gcwq)
1559 __releases(&gcwq->lock)
1560 __acquires(&gcwq->lock)
1561 {
1562         if (!need_to_create_worker(gcwq))
1563                 return false;
1564 restart:
1565         spin_unlock_irq(&gcwq->lock);
1566 
1567         /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1568         mod_timer(&gcwq->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1569 
1570         while (true) {
1571                 struct worker *worker;
1572 
1573                 worker = create_worker(gcwq, true);
1574                 if (worker) {
1575                         del_timer_sync(&gcwq->mayday_timer);
1576                         spin_lock_irq(&gcwq->lock);
1577                         start_worker(worker);
1578                         BUG_ON(need_to_create_worker(gcwq));
1579                         return true;
1580                 }
1581 
1582                 if (!need_to_create_worker(gcwq))
1583                         break;
1584 
1585                 __set_current_state(TASK_INTERRUPTIBLE);
1586                 schedule_timeout(CREATE_COOLDOWN);
1587 
1588                 if (!need_to_create_worker(gcwq))
1589                         break;
1590         }
1591 
1592         del_timer_sync(&gcwq->mayday_timer);
1593         spin_lock_irq(&gcwq->lock);
1594         if (need_to_create_worker(gcwq))
1595                 goto restart;
1596         return true;
1597 }
1598 
1599 /**
1600  * maybe_destroy_worker - destroy workers which have been idle for a while
1601  * @gcwq: gcwq to destroy workers for
1602  *
1603  * Destroy @gcwq workers which have been idle for longer than
1604  * IDLE_WORKER_TIMEOUT.
1605  *
1606  * LOCKING:
1607  * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1608  * multiple times.  Called only from manager.
1609  *
1610  * RETURNS:
1611  * false if no action was taken and gcwq->lock stayed locked, true
1612  * otherwise.
1613  */
1614 static bool maybe_destroy_workers(struct global_cwq *gcwq)
1615 {
1616         bool ret = false;
1617 
1618         while (too_many_workers(gcwq)) {
1619                 struct worker *worker;
1620                 unsigned long expires;
1621 
1622                 worker = list_entry(gcwq->idle_list.prev, struct worker, entry);
1623                 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1624 
1625                 if (time_before(jiffies, expires)) {
1626                         mod_timer(&gcwq->idle_timer, expires);
1627                         break;
1628                 }
1629 
1630                 destroy_worker(worker);
1631                 ret = true;
1632         }
1633 
1634         return ret;
1635 }
1636 
1637 /**
1638  * manage_workers - manage worker pool
1639  * @worker: self
1640  *
1641  * Assume the manager role and manage gcwq worker pool @worker belongs
1642  * to.  At any given time, there can be only zero or one manager per
1643  * gcwq.  The exclusion is handled automatically by this function.
1644  *
1645  * The caller can safely start processing works on false return.  On
1646  * true return, it's guaranteed that need_to_create_worker() is false
1647  * and may_start_working() is true.
1648  *
1649  * CONTEXT:
1650  * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1651  * multiple times.  Does GFP_KERNEL allocations.
1652  *
1653  * RETURNS:
1654  * false if no action was taken and gcwq->lock stayed locked, true if
1655  * some action was taken.
1656  */
1657 static bool manage_workers(struct worker *worker)
1658 {
1659         struct global_cwq *gcwq = worker->gcwq;
1660         bool ret = false;
1661 
1662         if (gcwq->flags & GCWQ_MANAGING_WORKERS)
1663                 return ret;
1664 
1665         gcwq->flags &= ~GCWQ_MANAGE_WORKERS;
1666         gcwq->flags |= GCWQ_MANAGING_WORKERS;
1667 
1668         /*
1669          * Destroy and then create so that may_start_working() is true
1670          * on return.
1671          */
1672         ret |= maybe_destroy_workers(gcwq);
1673         ret |= maybe_create_worker(gcwq);
1674 
1675         gcwq->flags &= ~GCWQ_MANAGING_WORKERS;
1676 
1677         /*
1678          * The trustee might be waiting to take over the manager
1679          * position, tell it we're done.
1680          */
1681         if (unlikely(gcwq->trustee))
1682                 wake_up_all(&gcwq->trustee_wait);
1683 
1684         return ret;
1685 }
1686 
1687 /**
1688  * move_linked_works - move linked works to a list
1689  * @work: start of series of works to be scheduled
1690  * @head: target list to append @work to
1691  * @nextp: out paramter for nested worklist walking
1692  *
1693  * Schedule linked works starting from @work to @head.  Work series to
1694  * be scheduled starts at @work and includes any consecutive work with
1695  * WORK_STRUCT_LINKED set in its predecessor.
1696  *
1697  * If @nextp is not NULL, it's updated to point to the next work of
1698  * the last scheduled work.  This allows move_linked_works() to be
1699  * nested inside outer list_for_each_entry_safe().
1700  *
1701  * CONTEXT:
1702  * spin_lock_irq(gcwq->lock).
1703  */
1704 static void move_linked_works(struct work_struct *work, struct list_head *head,
1705                               struct work_struct **nextp)
1706 {
1707         struct work_struct *n;
1708 
1709         /*
1710          * Linked worklist will always end before the end of the list,
1711          * use NULL for list head.
1712          */
1713         list_for_each_entry_safe_from(work, n, NULL, entry) {
1714                 list_move_tail(&work->entry, head);
1715                 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1716                         break;
1717         }
1718 
1719         /*
1720          * If we're already inside safe list traversal and have moved
1721          * multiple works to the scheduled queue, the next position
1722          * needs to be updated.
1723          */
1724         if (nextp)
1725                 *nextp = n;
1726 }
1727 
1728 static void cwq_activate_first_delayed(struct cpu_workqueue_struct *cwq)
1729 {
1730         struct work_struct *work = list_first_entry(&cwq->delayed_works,
1731                                                     struct work_struct, entry);
1732         struct list_head *pos = gcwq_determine_ins_pos(cwq->gcwq, cwq);
1733 
1734         trace_workqueue_activate_work(work);
1735         move_linked_works(work, pos, NULL);
1736         __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1737         cwq->nr_active++;
1738 }
1739 
1740 /**
1741  * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight
1742  * @cwq: cwq of interest
1743  * @color: color of work which left the queue
1744  * @delayed: for a delayed work
1745  *
1746  * A work either has completed or is removed from pending queue,
1747  * decrement nr_in_flight of its cwq and handle workqueue flushing.
1748  *
1749  * CONTEXT:
1750  * spin_lock_irq(gcwq->lock).
1751  */
1752 static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct *cwq, int color,
1753                                  bool delayed)
1754 {
1755         /* ignore uncolored works */
1756         if (color == WORK_NO_COLOR)
1757                 return;
1758 
1759         cwq->nr_in_flight[color]--;
1760 
1761         if (!delayed) {
1762                 cwq->nr_active--;
1763                 if (!list_empty(&cwq->delayed_works)) {
1764                         /* one down, submit a delayed one */
1765                         if (cwq->nr_active < cwq->max_active)
1766                                 cwq_activate_first_delayed(cwq);
1767                 }
1768         }
1769 
1770         /* is flush in progress and are we at the flushing tip? */
1771         if (likely(cwq->flush_color != color))
1772                 return;
1773 
1774         /* are there still in-flight works? */
1775         if (cwq->nr_in_flight[color])
1776                 return;
1777 
1778         /* this cwq is done, clear flush_color */
1779         cwq->flush_color = -1;
1780 
1781         /*
1782          * If this was the last cwq, wake up the first flusher.  It
1783          * will handle the rest.
1784          */
1785         if (atomic_dec_and_test(&cwq->wq->nr_cwqs_to_flush))
1786                 complete(&cwq->wq->first_flusher->done);
1787 }
1788 
1789 /**
1790  * process_one_work - process single work
1791  * @worker: self
1792  * @work: work to process
1793  *
1794  * Process @work.  This function contains all the logics necessary to
1795  * process a single work including synchronization against and
1796  * interaction with other workers on the same cpu, queueing and
1797  * flushing.  As long as context requirement is met, any worker can
1798  * call this function to process a work.
1799  *
1800  * CONTEXT:
1801  * spin_lock_irq(gcwq->lock) which is released and regrabbed.
1802  */
1803 static void process_one_work(struct worker *worker, struct work_struct *work)
1804 __releases(&gcwq->lock)
1805 __acquires(&gcwq->lock)
1806 {
1807         struct cpu_workqueue_struct *cwq = get_work_cwq(work);
1808         struct global_cwq *gcwq = cwq->gcwq;
1809         struct hlist_head *bwh = busy_worker_head(gcwq, work);
1810         bool cpu_intensive = cwq->wq->flags & WQ_CPU_INTENSIVE;
1811         work_func_t f = work->func;
1812         int work_color;
1813         struct worker *collision;
1814 #ifdef CONFIG_LOCKDEP
1815         /*
1816          * It is permissible to free the struct work_struct from
1817          * inside the function that is called from it, this we need to
1818          * take into account for lockdep too.  To avoid bogus "held
1819          * lock freed" warnings as well as problems when looking into
1820          * work->lockdep_map, make a copy and use that here.
1821          */
1822         struct lockdep_map lockdep_map = work->lockdep_map;
1823 #endif
1824         /*
1825          * A single work shouldn't be executed concurrently by
1826          * multiple workers on a single cpu.  Check whether anyone is
1827          * already processing the work.  If so, defer the work to the
1828          * currently executing one.
1829          */
1830         collision = __find_worker_executing_work(gcwq, bwh, work);
1831         if (unlikely(collision)) {
1832                 move_linked_works(work, &collision->scheduled, NULL);
1833                 return;
1834         }
1835 
1836         /* claim and process */
1837         debug_work_deactivate(work);
1838         hlist_add_head(&worker->hentry, bwh);
1839         worker->current_work = work;
1840         worker->current_cwq = cwq;
1841         work_color = get_work_color(work);
1842 
1843         /* record the current cpu number in the work data and dequeue */
1844         set_work_cpu(work, gcwq->cpu);
1845         list_del_init(&work->entry);
1846 
1847         /*
1848          * If HIGHPRI_PENDING, check the next work, and, if HIGHPRI,
1849          * wake up another worker; otherwise, clear HIGHPRI_PENDING.
1850          */
1851         if (unlikely(gcwq->flags & GCWQ_HIGHPRI_PENDING)) {
1852                 struct work_struct *nwork = list_first_entry(&gcwq->worklist,
1853                                                 struct work_struct, entry);
1854 
1855                 if (!list_empty(&gcwq->worklist) &&
1856                     get_work_cwq(nwork)->wq->flags & WQ_HIGHPRI)
1857                         wake_up_worker(gcwq);
1858                 else
1859                         gcwq->flags &= ~GCWQ_HIGHPRI_PENDING;
1860         }
1861 
1862         /*
1863          * CPU intensive works don't participate in concurrency
1864          * management.  They're the scheduler's responsibility.
1865          */
1866         if (unlikely(cpu_intensive))
1867                 worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
1868 
1869         spin_unlock_irq(&gcwq->lock);
1870 
1871         smp_wmb();      /* paired with test_and_set_bit(PENDING) */
1872         work_clear_pending(work);
1873 
1874         lock_map_acquire_read(&cwq->wq->lockdep_map);
1875         lock_map_acquire(&lockdep_map);
1876         trace_workqueue_execute_start(work);
1877         f(work);
1878         /*
1879          * While we must be careful to not use "work" after this, the trace
1880          * point will only record its address.
1881          */
1882         trace_workqueue_execute_end(work);
1883         lock_map_release(&lockdep_map);
1884         lock_map_release(&cwq->wq->lockdep_map);
1885 
1886         if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
1887                 printk(KERN_ERR "BUG: workqueue leaked lock or atomic: "
1888                        "%s/0x%08x/%d\n",
1889                        current->comm, preempt_count(), task_pid_nr(current));
1890                 printk(KERN_ERR "    last function: ");
1891                 print_symbol("%s\n", (unsigned long)f);
1892                 debug_show_held_locks(current);
1893                 dump_stack();
1894         }
1895 
1896         spin_lock_irq(&gcwq->lock);
1897 
1898         /* clear cpu intensive status */
1899         if (unlikely(cpu_intensive))
1900                 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
1901 
1902         /* we're done with it, release */
1903         hlist_del_init(&worker->hentry);
1904         worker->current_work = NULL;
1905         worker->current_cwq = NULL;
1906         cwq_dec_nr_in_flight(cwq, work_color, false);
1907 }
1908 
1909 /**
1910  * process_scheduled_works - process scheduled works
1911  * @worker: self
1912  *
1913  * Process all scheduled works.  Please note that the scheduled list
1914  * may change while processing a work, so this function repeatedly
1915  * fetches a work from the top and executes it.
1916  *
1917  * CONTEXT:
1918  * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1919  * multiple times.
1920  */
1921 static void process_scheduled_works(struct worker *worker)
1922 {
1923         while (!list_empty(&worker->scheduled)) {
1924                 struct work_struct *work = list_first_entry(&worker->scheduled,
1925                                                 struct work_struct, entry);
1926                 process_one_work(worker, work);
1927         }
1928 }
1929 
1930 /**
1931  * worker_thread - the worker thread function
1932  * @__worker: self
1933  *
1934  * The gcwq worker thread function.  There's a single dynamic pool of
1935  * these per each cpu.  These workers process all works regardless of
1936  * their specific target workqueue.  The only exception is works which
1937  * belong to workqueues with a rescuer which will be explained in
1938  * rescuer_thread().
1939  */
1940 static int worker_thread(void *__worker)
1941 {
1942         struct worker *worker = __worker;
1943         struct global_cwq *gcwq = worker->gcwq;
1944 
1945         /* tell the scheduler that this is a workqueue worker */
1946         worker->task->flags |= PF_WQ_WORKER;
1947 woke_up:
1948         spin_lock_irq(&gcwq->lock);
1949 
1950         /* DIE can be set only while we're idle, checking here is enough */
1951         if (worker->flags & WORKER_DIE) {
1952                 spin_unlock_irq(&gcwq->lock);
1953                 worker->task->flags &= ~PF_WQ_WORKER;
1954                 return 0;
1955         }
1956 
1957         worker_leave_idle(worker);
1958 recheck:
1959         /* no more worker necessary? */
1960         if (!need_more_worker(gcwq))
1961                 goto sleep;
1962 
1963         /* do we need to manage? */
1964         if (unlikely(!may_start_working(gcwq)) && manage_workers(worker))
1965                 goto recheck;
1966 
1967         /*
1968          * ->scheduled list can only be filled while a worker is
1969          * preparing to process a work or actually processing it.
1970          * Make sure nobody diddled with it while I was sleeping.
1971          */
1972         BUG_ON(!list_empty(&worker->scheduled));
1973 
1974         /*
1975          * When control reaches this point, we're guaranteed to have
1976          * at least one idle worker or that someone else has already
1977          * assumed the manager role.
1978          */
1979         worker_clr_flags(worker, WORKER_PREP);
1980 
1981         do {
1982                 struct work_struct *work =
1983                         list_first_entry(&gcwq->worklist,
1984                                          struct work_struct, entry);
1985 
1986                 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
1987                         /* optimization path, not strictly necessary */
1988                         process_one_work(worker, work);
1989                         if (unlikely(!list_empty(&worker->scheduled)))
1990                                 process_scheduled_works(worker);
1991                 } else {
1992                         move_linked_works(work, &worker->scheduled, NULL);
1993                         process_scheduled_works(worker);
1994                 }
1995         } while (keep_working(gcwq));
1996 
1997         worker_set_flags(worker, WORKER_PREP, false);
1998 sleep:
1999         if (unlikely(need_to_manage_workers(gcwq)) && manage_workers(worker))
2000                 goto recheck;
2001 
2002         /*
2003          * gcwq->lock is held and there's no work to process and no
2004          * need to manage, sleep.  Workers are woken up only while
2005          * holding gcwq->lock or from local cpu, so setting the
2006          * current state before releasing gcwq->lock is enough to
2007          * prevent losing any event.
2008          */
2009         worker_enter_idle(worker);
2010         __set_current_state(TASK_INTERRUPTIBLE);
2011         spin_unlock_irq(&gcwq->lock);
2012         schedule();
2013         goto woke_up;
2014 }
2015 
2016 /**
2017  * rescuer_thread - the rescuer thread function
2018  * @__wq: the associated workqueue
2019  *
2020  * Workqueue rescuer thread function.  There's one rescuer for each
2021  * workqueue which has WQ_RESCUER set.
2022  *
2023  * Regular work processing on a gcwq may block trying to create a new
2024  * worker which uses GFP_KERNEL allocation which has slight chance of
2025  * developing into deadlock if some works currently on the same queue
2026  * need to be processed to satisfy the GFP_KERNEL allocation.  This is
2027  * the problem rescuer solves.
2028  *
2029  * When such condition is possible, the gcwq summons rescuers of all
2030  * workqueues which have works queued on the gcwq and let them process
2031  * those works so that forward progress can be guaranteed.
2032  *
2033  * This should happen rarely.
2034  */
2035 static int rescuer_thread(void *__wq)
2036 {
2037         struct workqueue_struct *wq = __wq;
2038         struct worker *rescuer = wq->rescuer;
2039         struct list_head *scheduled = &rescuer->scheduled;
2040         bool is_unbound = wq->flags & WQ_UNBOUND;
2041         unsigned int cpu;
2042 
2043         set_user_nice(current, RESCUER_NICE_LEVEL);
2044 repeat:
2045         set_current_state(TASK_INTERRUPTIBLE);
2046 
2047         if (kthread_should_stop()) {
2048                 __set_current_state(TASK_RUNNING);
2049                 return 0;
2050         }
2051 
2052         /*
2053          * See whether any cpu is asking for help.  Unbounded
2054          * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND.
2055          */
2056         for_each_mayday_cpu(cpu, wq->mayday_mask) {
2057                 unsigned int tcpu = is_unbound ? WORK_CPU_UNBOUND : cpu;
2058                 struct cpu_workqueue_struct *cwq = get_cwq(tcpu, wq);
2059                 struct global_cwq *gcwq = cwq->gcwq;
2060                 struct work_struct *work, *n;
2061 
2062                 __set_current_state(TASK_RUNNING);
2063                 mayday_clear_cpu(cpu, wq->mayday_mask);
2064 
2065                 /* migrate to the target cpu if possible */
2066                 rescuer->gcwq = gcwq;
2067                 worker_maybe_bind_and_lock(rescuer);
2068 
2069                 /*
2070                  * Slurp in all works issued via this workqueue and
2071                  * process'em.
2072                  */
2073                 BUG_ON(!list_empty(&rescuer->scheduled));
2074                 list_for_each_entry_safe(work, n, &gcwq->worklist, entry)
2075                         if (get_work_cwq(work) == cwq)
2076                                 move_linked_works(work, scheduled, &n);
2077 
2078                 process_scheduled_works(rescuer);
2079 
2080                 /*
2081                  * Leave this gcwq.  If keep_working() is %true, notify a
2082                  * regular worker; otherwise, we end up with 0 concurrency
2083                  * and stalling the execution.
2084                  */
2085                 if (keep_working(gcwq))
2086                         wake_up_worker(gcwq);
2087 
2088                 spin_unlock_irq(&gcwq->lock);
2089         }
2090 
2091         schedule();
2092         goto repeat;
2093 }
2094 
2095 struct wq_barrier {
2096         struct work_struct      work;
2097         struct completion       done;
2098 };
2099 
2100 static void wq_barrier_func(struct work_struct *work)
2101 {
2102         struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2103         complete(&barr->done);
2104 }
2105 
2106 /**
2107  * insert_wq_barrier - insert a barrier work
2108  * @cwq: cwq to insert barrier into
2109  * @barr: wq_barrier to insert
2110  * @target: target work to attach @barr to
2111  * @worker: worker currently executing @target, NULL if @target is not executing
2112  *
2113  * @barr is linked to @target such that @barr is completed only after
2114  * @target finishes execution.  Please note that the ordering
2115  * guarantee is observed only with respect to @target and on the local
2116  * cpu.
2117  *
2118  * Currently, a queued barrier can't be canceled.  This is because
2119  * try_to_grab_pending() can't determine whether the work to be
2120  * grabbed is at the head of the queue and thus can't clear LINKED
2121  * flag of the previous work while there must be a valid next work
2122  * after a work with LINKED flag set.
2123  *
2124  * Note that when @worker is non-NULL, @target may be modified
2125  * underneath us, so we can't reliably determine cwq from @target.
2126  *
2127  * CONTEXT:
2128  * spin_lock_irq(gcwq->lock).
2129  */
2130 static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
2131                               struct wq_barrier *barr,
2132                               struct work_struct *target, struct worker *worker)
2133 {
2134         struct list_head *head;
2135         unsigned int linked = 0;
2136 
2137         /*
2138          * debugobject calls are safe here even with gcwq->lock locked
2139          * as we know for sure that this will not trigger any of the
2140          * checks and call back into the fixup functions where we
2141          * might deadlock.
2142          */
2143         INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2144         __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2145         init_completion(&barr->done);
2146 
2147         /*
2148          * If @target is currently being executed, schedule the
2149          * barrier to the worker; otherwise, put it after @target.
2150          */
2151         if (worker)
2152                 head = worker->scheduled.next;
2153         else {
2154                 unsigned long *bits = work_data_bits(target);
2155 
2156                 head = target->entry.next;
2157                 /* there can already be other linked works, inherit and set */
2158                 linked = *bits & WORK_STRUCT_LINKED;
2159                 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2160         }
2161 
2162         debug_work_activate(&barr->work);
2163         insert_work(cwq, &barr->work, head,
2164                     work_color_to_flags(WORK_NO_COLOR) | linked);
2165 }
2166 
2167 /**
2168  * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing
2169  * @wq: workqueue being flushed
2170  * @flush_color: new flush color, < 0 for no-op
2171  * @work_color: new work color, < 0 for no-op
2172  *
2173  * Prepare cwqs for workqueue flushing.
2174  *
2175  * If @flush_color is non-negative, flush_color on all cwqs should be
2176  * -1.  If no cwq has in-flight commands at the specified color, all
2177  * cwq->flush_color's stay at -1 and %false is returned.  If any cwq
2178  * has in flight commands, its cwq->flush_color is set to
2179  * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq
2180  * wakeup logic is armed and %true is returned.
2181  *
2182  * The caller should have initialized @wq->first_flusher prior to
2183  * calling this function with non-negative @flush_color.  If
2184  * @flush_color is negative, no flush color update is done and %false
2185  * is returned.
2186  *
2187  * If @work_color is non-negative, all cwqs should have the same
2188  * work_color which is previous to @work_color and all will be
2189  * advanced to @work_color.
2190  *
2191  * CONTEXT:
2192  * mutex_lock(wq->flush_mutex).
2193  *
2194  * RETURNS:
2195  * %true if @flush_color >= 0 and there's something to flush.  %false
2196  * otherwise.
2197  */
2198 static bool flush_workqueue_prep_cwqs(struct workqueue_struct *wq,
2199                                       int flush_color, int work_color)
2200 {
2201         bool wait = false;
2202         unsigned int cpu;
2203 
2204         if (flush_color >= 0) {
2205                 BUG_ON(atomic_read(&wq->nr_cwqs_to_flush));
2206                 atomic_set(&wq->nr_cwqs_to_flush, 1);
2207         }
2208 
2209         for_each_cwq_cpu(cpu, wq) {
2210                 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2211                 struct global_cwq *gcwq = cwq->gcwq;
2212 
2213                 spin_lock_irq(&gcwq->lock);
2214 
2215                 if (flush_color >= 0) {
2216                         BUG_ON(cwq->flush_color != -1);
2217 
2218                         if (cwq->nr_in_flight[flush_color]) {
2219                                 cwq->flush_color = flush_color;
2220                                 atomic_inc(&wq->nr_cwqs_to_flush);
2221                                 wait = true;
2222                         }
2223                 }
2224 
2225                 if (work_color >= 0) {
2226                         BUG_ON(work_color != work_next_color(cwq->work_color));
2227                         cwq->work_color = work_color;
2228                 }
2229 
2230                 spin_unlock_irq(&gcwq->lock);
2231         }
2232 
2233         if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_cwqs_to_flush))
2234                 complete(&wq->first_flusher->done);
2235 
2236         return wait;
2237 }
2238 
2239 /**
2240  * flush_workqueue - ensure that any scheduled work has run to completion.
2241  * @wq: workqueue to flush
2242  *
2243  * Forces execution of the workqueue and blocks until its completion.
2244  * This is typically used in driver shutdown handlers.
2245  *
2246  * We sleep until all works which were queued on entry have been handled,
2247  * but we are not livelocked by new incoming ones.
2248  */
2249 void flush_workqueue(struct workqueue_struct *wq)
2250 {
2251         struct wq_flusher this_flusher = {
2252                 .list = LIST_HEAD_INIT(this_flusher.list),
2253                 .flush_color = -1,
2254                 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2255         };
2256         int next_color;
2257 
2258         lock_map_acquire(&wq->lockdep_map);
2259         lock_map_release(&wq->lockdep_map);
2260 
2261         mutex_lock(&wq->flush_mutex);
2262 
2263         /*
2264          * Start-to-wait phase
2265          */
2266         next_color = work_next_color(wq->work_color);
2267 
2268         if (next_color != wq->flush_color) {
2269                 /*
2270                  * Color space is not full.  The current work_color
2271                  * becomes our flush_color and work_color is advanced
2272                  * by one.
2273                  */
2274                 BUG_ON(!list_empty(&wq->flusher_overflow));
2275                 this_flusher.flush_color = wq->work_color;
2276                 wq->work_color = next_color;
2277 
2278                 if (!wq->first_flusher) {
2279                         /* no flush in progress, become the first flusher */
2280                         BUG_ON(wq->flush_color != this_flusher.flush_color);
2281 
2282                         wq->first_flusher = &this_flusher;
2283 
2284                         if (!flush_workqueue_prep_cwqs(wq, wq->flush_color,
2285                                                        wq->work_color)) {
2286                                 /* nothing to flush, done */
2287                                 wq->flush_color = next_color;
2288                                 wq->first_flusher = NULL;
2289                                 goto out_unlock;
2290                         }
2291                 } else {
2292                         /* wait in queue */
2293                         BUG_ON(wq->flush_color == this_flusher.flush_color);
2294                         list_add_tail(&this_flusher.list, &wq->flusher_queue);
2295                         flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2296                 }
2297         } else {
2298                 /*
2299                  * Oops, color space is full, wait on overflow queue.
2300                  * The next flush completion will assign us
2301                  * flush_color and transfer to flusher_queue.
2302                  */
2303                 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2304         }
2305 
2306         mutex_unlock(&wq->flush_mutex);
2307 
2308         wait_for_completion(&this_flusher.done);
2309 
2310         /*
2311          * Wake-up-and-cascade phase
2312          *
2313          * First flushers are responsible for cascading flushes and
2314          * handling overflow.  Non-first flushers can simply return.
2315          */
2316         if (wq->first_flusher != &this_flusher)
2317                 return;
2318 
2319         mutex_lock(&wq->flush_mutex);
2320 
2321         /* we might have raced, check again with mutex held */
2322         if (wq->first_flusher != &this_flusher)
2323                 goto out_unlock;
2324 
2325         wq->first_flusher = NULL;
2326 
2327         BUG_ON(!list_empty(&this_flusher.list));
2328         BUG_ON(wq->flush_color != this_flusher.flush_color);
2329 
2330         while (true) {
2331                 struct wq_flusher *next, *tmp;
2332 
2333                 /* complete all the flushers sharing the current flush color */
2334                 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2335                         if (next->flush_color != wq->flush_color)
2336                                 break;
2337                         list_del_init(&next->list);
2338                         complete(&next->done);
2339                 }
2340 
2341                 BUG_ON(!list_empty(&wq->flusher_overflow) &&
2342                        wq->flush_color != work_next_color(wq->work_color));
2343 
2344                 /* this flush_color is finished, advance by one */
2345                 wq->flush_color = work_next_color(wq->flush_color);
2346 
2347                 /* one color has been freed, handle overflow queue */
2348                 if (!list_empty(&wq->flusher_overflow)) {
2349                         /*
2350                          * Assign the same color to all overflowed
2351                          * flushers, advance work_color and append to
2352                          * flusher_queue.  This is the start-to-wait
2353                          * phase for these overflowed flushers.
2354                          */
2355                         list_for_each_entry(tmp, &wq->flusher_overflow, list)
2356                                 tmp->flush_color = wq->work_color;
2357 
2358                         wq->work_color = work_next_color(wq->work_color);
2359 
2360                         list_splice_tail_init(&wq->flusher_overflow,
2361                                               &wq->flusher_queue);
2362                         flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2363                 }
2364 
2365                 if (list_empty(&wq->flusher_queue)) {
2366                         BUG_ON(wq->flush_color != wq->work_color);
2367                         break;
2368                 }
2369 
2370                 /*
2371                  * Need to flush more colors.  Make the next flusher
2372                  * the new first flusher and arm cwqs.
2373                  */
2374                 BUG_ON(wq->flush_color == wq->work_color);
2375                 BUG_ON(wq->flush_color != next->flush_color);
2376 
2377                 list_del_init(&next->list);
2378                 wq->first_flusher = next;
2379 
2380                 if (flush_workqueue_prep_cwqs(wq, wq->flush_color, -1))
2381                         break;
2382 
2383                 /*
2384                  * Meh... this color is already done, clear first
2385                  * flusher and repeat cascading.
2386                  */
2387                 wq->first_flusher = NULL;
2388         }
2389 
2390 out_unlock:
2391         mutex_unlock(&wq->flush_mutex);
2392 }
2393 EXPORT_SYMBOL_GPL(flush_workqueue);
2394 
2395 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
2396                              bool wait_executing)
2397 {
2398         struct worker *worker = NULL;
2399         struct global_cwq *gcwq;
2400         struct cpu_workqueue_struct *cwq;
2401 
2402         might_sleep();
2403         gcwq = get_work_gcwq(work);
2404         if (!gcwq)
2405                 return false;
2406 
2407         spin_lock_irq(&gcwq->lock);
2408         if (!list_empty(&work->entry)) {
2409                 /*
2410                  * See the comment near try_to_grab_pending()->smp_rmb().
2411                  * If it was re-queued to a different gcwq under us, we
2412                  * are not going to wait.
2413                  */
2414                 smp_rmb();
2415                 cwq = get_work_cwq(work);
2416                 if (unlikely(!cwq || gcwq != cwq->gcwq))
2417                         goto already_gone;
2418         } else if (wait_executing) {
2419                 worker = find_worker_executing_work(gcwq, work);
2420                 if (!worker)
2421                         goto already_gone;
2422                 cwq = worker->current_cwq;
2423         } else
2424                 goto already_gone;
2425 
2426         insert_wq_barrier(cwq, barr, work, worker);
2427         spin_unlock_irq(&gcwq->lock);
2428 
2429         /*
2430          * If @max_active is 1 or rescuer is in use, flushing another work
2431          * item on the same workqueue may lead to deadlock.  Make sure the
2432          * flusher is not running on the same workqueue by verifying write
2433          * access.
2434          */
2435         if (cwq->wq->saved_max_active == 1 || cwq->wq->flags & WQ_RESCUER)
2436                 lock_map_acquire(&cwq->wq->lockdep_map);
2437         else
2438                 lock_map_acquire_read(&cwq->wq->lockdep_map);
2439         lock_map_release(&cwq->wq->lockdep_map);
2440 
2441         return true;
2442 already_gone:
2443         spin_unlock_irq(&gcwq->lock);
2444         return false;
2445 }
2446 
2447 /**
2448  * flush_work - wait for a work to finish executing the last queueing instance
2449  * @work: the work to flush
2450  *
2451  * Wait until @work has finished execution.  This function considers
2452  * only the last queueing instance of @work.  If @work has been
2453  * enqueued across different CPUs on a non-reentrant workqueue or on
2454  * multiple workqueues, @work might still be executing on return on
2455  * some of the CPUs from earlier queueing.
2456  *
2457  * If @work was queued only on a non-reentrant, ordered or unbound
2458  * workqueue, @work is guaranteed to be idle on return if it hasn't
2459  * been requeued since flush started.
2460  *
2461  * RETURNS:
2462  * %true if flush_work() waited for the work to finish execution,
2463  * %false if it was already idle.
2464  */
2465 bool flush_work(struct work_struct *work)
2466 {
2467         struct wq_barrier barr;
2468 
2469         if (start_flush_work(work, &barr, true)) {
2470                 wait_for_completion(&barr.done);
2471                 destroy_work_on_stack(&barr.work);
2472                 return true;
2473         } else
2474                 return false;
2475 }
2476 EXPORT_SYMBOL_GPL(flush_work);
2477 
2478 static bool wait_on_cpu_work(struct global_cwq *gcwq, struct work_struct *work)
2479 {
2480         struct wq_barrier barr;
2481         struct worker *worker;
2482 
2483         spin_lock_irq(&gcwq->lock);
2484 
2485         worker = find_worker_executing_work(gcwq, work);
2486         if (unlikely(worker))
2487                 insert_wq_barrier(worker->current_cwq, &barr, work, worker);
2488 
2489         spin_unlock_irq(&gcwq->lock);
2490 
2491         if (unlikely(worker)) {
2492                 wait_for_completion(&barr.done);
2493                 destroy_work_on_stack(&barr.work);
2494                 return true;
2495         } else
2496                 return false;
2497 }
2498 
2499 static bool wait_on_work(struct work_struct *work)
2500 {
2501         bool ret = false;
2502         int cpu;
2503 
2504         might_sleep();
2505 
2506         lock_map_acquire(&work->lockdep_map);
2507         lock_map_release(&work->lockdep_map);
2508 
2509         for_each_gcwq_cpu(cpu)
2510                 ret |= wait_on_cpu_work(get_gcwq(cpu), work);
2511         return ret;
2512 }
2513 
2514 /**
2515  * flush_work_sync - wait until a work has finished execution
2516  * @work: the work to flush
2517  *
2518  * Wait until @work has finished execution.  On return, it's
2519  * guaranteed that all queueing instances of @work which happened
2520  * before this function is called are finished.  In other words, if
2521  * @work hasn't been requeued since this function was called, @work is
2522  * guaranteed to be idle on return.
2523  *
2524  * RETURNS:
2525  * %true if flush_work_sync() waited for the work to finish execution,
2526  * %false if it was already idle.
2527  */
2528 bool flush_work_sync(struct work_struct *work)
2529 {
2530         struct wq_barrier barr;
2531         bool pending, waited;
2532 
2533         /* we'll wait for executions separately, queue barr only if pending */
2534         pending = start_flush_work(work, &barr, false);
2535 
2536         /* wait for executions to finish */
2537         waited = wait_on_work(work);
2538 
2539         /* wait for the pending one */
2540         if (pending) {
2541                 wait_for_completion(&barr.done);
2542                 destroy_work_on_stack(&barr.work);
2543         }
2544 
2545         return pending || waited;
2546 }
2547 EXPORT_SYMBOL_GPL(flush_work_sync);
2548 
2549 /*
2550  * Upon a successful return (>= 0), the caller "owns" WORK_STRUCT_PENDING bit,
2551  * so this work can't be re-armed in any way.
2552  */
2553 static int try_to_grab_pending(struct work_struct *work)
2554 {
2555         struct global_cwq *gcwq;
2556         int ret = -1;
2557 
2558         if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
2559                 return 0;
2560 
2561         /*
2562          * The queueing is in progress, or it is already queued. Try to
2563          * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
2564          */
2565         gcwq = get_work_gcwq(work);
2566         if (!gcwq)
2567                 return ret;
2568 
2569         spin_lock_irq(&gcwq->lock);
2570         if (!list_empty(&work->entry)) {
2571                 /*
2572                  * This work is queued, but perhaps we locked the wrong gcwq.
2573                  * In that case we must see the new value after rmb(), see
2574                  * insert_work()->wmb().
2575                  */
2576                 smp_rmb();
2577                 if (gcwq == get_work_gcwq(work)) {
2578                         debug_work_deactivate(work);
2579                         list_del_init(&work->entry);
2580                         cwq_dec_nr_in_flight(get_work_cwq(work),
2581                                 get_work_color(work),
2582                                 *work_data_bits(work) & WORK_STRUCT_DELAYED);
2583                         ret = 1;
2584                 }
2585         }
2586         spin_unlock_irq(&gcwq->lock);
2587 
2588         return ret;
2589 }
2590 
2591 static bool __cancel_work_timer(struct work_struct *work,
2592                                 struct timer_list* timer)
2593 {
2594         int ret;
2595 
2596         do {
2597                 ret = (timer && likely(del_timer(timer)));
2598                 if (!ret)
2599                         ret = try_to_grab_pending(work);
2600                 wait_on_work(work);
2601         } while (unlikely(ret < 0));
2602 
2603         clear_work_data(work);
2604         return ret;
2605 }
2606 
2607 /**
2608  * cancel_work_sync - cancel a work and wait for it to finish
2609  * @work: the work to cancel
2610  *
2611  * Cancel @work and wait for its execution to finish.  This function
2612  * can be used even if the work re-queues itself or migrates to
2613  * another workqueue.  On return from this function, @work is
2614  * guaranteed to be not pending or executing on any CPU.
2615  *
2616  * cancel_work_sync(&delayed_work->work) must not be used for
2617  * delayed_work's.  Use cancel_delayed_work_sync() instead.
2618  *
2619  * The caller must ensure that the workqueue on which @work was last
2620  * queued can't be destroyed before this function returns.
2621  *
2622  * RETURNS:
2623  * %true if @work was pending, %false otherwise.
2624  */
2625 bool cancel_work_sync(struct work_struct *work)
2626 {
2627         return __cancel_work_timer(work, NULL);
2628 }
2629 EXPORT_SYMBOL_GPL(cancel_work_sync);
2630 
2631 /**
2632  * flush_delayed_work - wait for a dwork to finish executing the last queueing
2633  * @dwork: the delayed work to flush
2634  *
2635  * Delayed timer is cancelled and the pending work is queued for
2636  * immediate execution.  Like flush_work(), this function only
2637  * considers the last queueing instance of @dwork.
2638  *
2639  * RETURNS:
2640  * %true if flush_work() waited for the work to finish execution,
2641  * %false if it was already idle.
2642  */
2643 bool flush_delayed_work(struct delayed_work *dwork)
2644 {
2645         if (del_timer_sync(&dwork->timer))
2646                 __queue_work(raw_smp_processor_id(),
2647                              get_work_cwq(&dwork->work)->wq, &dwork->work);
2648         return flush_work(&dwork->work);
2649 }
2650 EXPORT_SYMBOL(flush_delayed_work);
2651 
2652 /**
2653  * flush_delayed_work_sync - wait for a dwork to finish
2654  * @dwork: the delayed work to flush
2655  *
2656  * Delayed timer is cancelled and the pending work is queued for
2657  * execution immediately.  Other than timer handling, its behavior
2658  * is identical to flush_work_sync().
2659  *
2660  * RETURNS:
2661  * %true if flush_work_sync() waited for the work to finish execution,
2662  * %false if it was already idle.
2663  */
2664 bool flush_delayed_work_sync(struct delayed_work *dwork)
2665 {
2666         if (del_timer_sync(&dwork->timer))
2667                 __queue_work(raw_smp_processor_id(),
2668                              get_work_cwq(&dwork->work)->wq, &dwork->work);
2669         return flush_work_sync(&dwork->work);
2670 }
2671 EXPORT_SYMBOL(flush_delayed_work_sync);
2672 
2673 /**
2674  * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2675  * @dwork: the delayed work cancel
2676  *
2677  * This is cancel_work_sync() for delayed works.
2678  *
2679  * RETURNS:
2680  * %true if @dwork was pending, %false otherwise.
2681  */
2682 bool cancel_delayed_work_sync(struct delayed_work *dwork)
2683 {
2684         return __cancel_work_timer(&dwork->work, &dwork->timer);
2685 }
2686 EXPORT_SYMBOL(cancel_delayed_work_sync);
2687 
2688 /**
2689  * schedule_work - put work task in global workqueue
2690  * @work: job to be done
2691  *
2692  * Returns zero if @work was already on the kernel-global workqueue and
2693  * non-zero otherwise.
2694  *
2695  * This puts a job in the kernel-global workqueue if it was not already
2696  * queued and leaves it in the same position on the kernel-global
2697  * workqueue otherwise.
2698  */
2699 int schedule_work(struct work_struct *work)
2700 {
2701         return queue_work(system_wq, work);
2702 }
2703 EXPORT_SYMBOL(schedule_work);
2704 
2705 /*
2706  * schedule_work_on - put work task on a specific cpu
2707  * @cpu: cpu to put the work task on
2708  * @work: job to be done
2709  *
2710  * This puts a job on a specific cpu
2711  */
2712 int schedule_work_on(int cpu, struct work_struct *work)
2713 {
2714         return queue_work_on(cpu, system_wq, work);
2715 }
2716 EXPORT_SYMBOL(schedule_work_on);
2717 
2718 /**
2719  * schedule_delayed_work - put work task in global workqueue after delay
2720  * @dwork: job to be done
2721  * @delay: number of jiffies to wait or 0 for immediate execution
2722  *
2723  * After waiting for a given time this puts a job in the kernel-global
2724  * workqueue.
2725  */
2726 int schedule_delayed_work(struct delayed_work *dwork,
2727                                         unsigned long delay)
2728 {
2729         return queue_delayed_work(system_wq, dwork, delay);
2730 }
2731 EXPORT_SYMBOL(schedule_delayed_work);
2732 
2733 /**
2734  * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
2735  * @cpu: cpu to use
2736  * @dwork: job to be done
2737  * @delay: number of jiffies to wait
2738  *
2739  * After waiting for a given time this puts a job in the kernel-global
2740  * workqueue on the specified CPU.
2741  */
2742 int schedule_delayed_work_on(int cpu,
2743                         struct delayed_work *dwork, unsigned long delay)
2744 {
2745         return queue_delayed_work_on(cpu, system_wq, dwork, delay);
2746 }
2747 EXPORT_SYMBOL(schedule_delayed_work_on);
2748 
2749 /**
2750  * schedule_on_each_cpu - execute a function synchronously on each online CPU
2751  * @func: the function to call
2752  *
2753  * schedule_on_each_cpu() executes @func on each online CPU using the
2754  * system workqueue and blocks until all CPUs have completed.
2755  * schedule_on_each_cpu() is very slow.
2756  *
2757  * RETURNS:
2758  * 0 on success, -errno on failure.
2759  */
2760 int schedule_on_each_cpu(work_func_t func)
2761 {
2762         int cpu;
2763         struct work_struct __percpu *works;
2764 
2765         works = alloc_percpu(struct work_struct);
2766         if (!works)
2767                 return -ENOMEM;
2768 
2769         get_online_cpus();
2770 
2771         for_each_online_cpu(cpu) {
2772                 struct work_struct *work = per_cpu_ptr(works, cpu);
2773 
2774                 INIT_WORK(work, func);
2775                 schedule_work_on(cpu, work);
2776         }
2777 
2778         for_each_online_cpu(cpu)
2779                 flush_work(per_cpu_ptr(works, cpu));
2780 
2781         put_online_cpus();
2782         free_percpu(works);
2783         return 0;
2784 }
2785 
2786 /**
2787  * flush_scheduled_work - ensure that any scheduled work has run to completion.
2788  *
2789  * Forces execution of the kernel-global workqueue and blocks until its
2790  * completion.
2791  *
2792  * Think twice before calling this function!  It's very easy to get into
2793  * trouble if you don't take great care.  Either of the following situations
2794  * will lead to deadlock:
2795  *
2796  *      One of the work items currently on the workqueue needs to acquire
2797  *      a lock held by your code or its caller.
2798  *
2799  *      Your code is running in the context of a work routine.
2800  *
2801  * They will be detected by lockdep when they occur, but the first might not
2802  * occur very often.  It depends on what work items are on the workqueue and
2803  * what locks they need, which you have no control over.
2804  *
2805  * In most situations flushing the entire workqueue is overkill; you merely
2806  * need to know that a particular work item isn't queued and isn't running.
2807  * In such cases you should use cancel_delayed_work_sync() or
2808  * cancel_work_sync() instead.
2809  */
2810 void flush_scheduled_work(void)
2811 {
2812         flush_workqueue(system_wq);
2813 }
2814 EXPORT_SYMBOL(flush_scheduled_work);
2815 
2816 /**
2817  * execute_in_process_context - reliably execute the routine with user context
2818  * @fn:         the function to execute
2819  * @ew:         guaranteed storage for the execute work structure (must
2820  *              be available when the work executes)
2821  *
2822  * Executes the function immediately if process context is available,
2823  * otherwise schedules the function for delayed execution.
2824  *
2825  * Returns:     0 - function was executed
2826  *              1 - function was scheduled for execution
2827  */
2828 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
2829 {
2830         if (!in_interrupt()) {
2831                 fn(&ew->work);
2832                 return 0;
2833         }
2834 
2835         INIT_WORK(&ew->work, fn);
2836         schedule_work(&ew->work);
2837 
2838         return 1;
2839 }
2840 EXPORT_SYMBOL_GPL(execute_in_process_context);
2841 
2842 int keventd_up(void)
2843 {
2844         return system_wq != NULL;
2845 }
2846 
2847 static int alloc_cwqs(struct workqueue_struct *wq)
2848 {
2849         /*
2850          * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS.
2851          * Make sure that the alignment isn't lower than that of
2852          * unsigned long long.
2853          */
2854         const size_t size = sizeof(struct cpu_workqueue_struct);
2855         const size_t align = max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS,
2856                                    __alignof__(unsigned long long));
2857 #ifdef CONFIG_SMP
2858         bool percpu = !(wq->flags & WQ_UNBOUND);
2859 #else
2860         bool percpu = false;
2861 #endif
2862 
2863         if (percpu)
2864                 wq->cpu_wq.pcpu = __alloc_percpu(size, align);
2865         else {
2866                 void *ptr;
2867 
2868                 /*
2869                  * Allocate enough room to align cwq and put an extra
2870                  * pointer at the end pointing back to the originally
2871                  * allocated pointer which will be used for free.
2872                  */
2873                 ptr = kzalloc(size + align + sizeof(void *), GFP_KERNEL);
2874                 if (ptr) {
2875                         wq->cpu_wq.single = PTR_ALIGN(ptr, align);
2876                         *(void **)(wq->cpu_wq.single + 1) = ptr;
2877                 }
2878         }
2879 
2880         /* just in case, make sure it's actually aligned */
2881         BUG_ON(!IS_ALIGNED(wq->cpu_wq.v, align));
2882         return wq->cpu_wq.v ? 0 : -ENOMEM;
2883 }
2884 
2885 static void free_cwqs(struct workqueue_struct *wq)
2886 {
2887 #ifdef CONFIG_SMP
2888         bool percpu = !(wq->flags & WQ_UNBOUND);
2889 #else
2890         bool percpu = false;
2891 #endif
2892 
2893         if (percpu)
2894                 free_percpu(wq->cpu_wq.pcpu);
2895         else if (wq->cpu_wq.single) {
2896                 /* the pointer to free is stored right after the cwq */
2897                 kfree(*(void **)(wq->cpu_wq.single + 1));
2898         }
2899 }
2900 
2901 static int wq_clamp_max_active(int max_active, unsigned int flags,
2902                                const char *name)
2903 {
2904         int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
2905 
2906         if (max_active < 1 || max_active > lim)
2907                 printk(KERN_WARNING "workqueue: max_active %d requested for %s "
2908                        "is out of range, clamping between %d and %d\n",
2909                        max_active, name, 1, lim);
2910 
2911         return clamp_val(max_active, 1, lim);
2912 }
2913 
2914 struct workqueue_struct *__alloc_workqueue_key(const char *name,
2915                                                unsigned int flags,
2916                                                int max_active,
2917                                                struct lock_class_key *key,
2918                                                const char *lock_name)
2919 {
2920         struct workqueue_struct *wq;
2921         unsigned int cpu;
2922 
2923         /*
2924          * Workqueues which may be used during memory reclaim should
2925          * have a rescuer to guarantee forward progress.
2926          */
2927         if (flags & WQ_MEM_RECLAIM)
2928                 flags |= WQ_RESCUER;
2929 
2930         /*
2931          * Unbound workqueues aren't concurrency managed and should be
2932          * dispatched to workers immediately.
2933          */
2934         if (flags & WQ_UNBOUND)
2935                 flags |= WQ_HIGHPRI;
2936 
2937         max_active = max_active ?: WQ_DFL_ACTIVE;
2938         max_active = wq_clamp_max_active(max_active, flags, name);
2939 
2940         wq = kzalloc(sizeof(*wq), GFP_KERNEL);
2941         if (!wq)
2942                 goto err;
2943 
2944         wq->flags = flags;
2945         wq->saved_max_active = max_active;
2946         mutex_init(&wq->flush_mutex);
2947         atomic_set(&wq->nr_cwqs_to_flush, 0);
2948         INIT_LIST_HEAD(&wq->flusher_queue);
2949         INIT_LIST_HEAD(&wq->flusher_overflow);
2950 
2951         wq->name = name;
2952         lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
2953         INIT_LIST_HEAD(&wq->list);
2954 
2955         if (alloc_cwqs(wq) < 0)
2956                 goto err;
2957 
2958         for_each_cwq_cpu(cpu, wq) {
2959                 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2960                 struct global_cwq *gcwq = get_gcwq(cpu);
2961 
2962                 BUG_ON((unsigned long)cwq & WORK_STRUCT_FLAG_MASK);
2963                 cwq->gcwq = gcwq;
2964                 cwq->wq = wq;
2965                 cwq->flush_color = -1;
2966                 cwq->max_active = max_active;
2967                 INIT_LIST_HEAD(&cwq->delayed_works);
2968         }
2969 
2970         if (flags & WQ_RESCUER) {
2971                 struct worker *rescuer;
2972 
2973                 if (!alloc_mayday_mask(&wq->mayday_mask, GFP_KERNEL))
2974                         goto err;
2975 
2976                 wq->rescuer = rescuer = alloc_worker();
2977                 if (!rescuer)
2978                         goto err;
2979 
2980                 rescuer->task = kthread_create(rescuer_thread, wq, "%s", name);
2981                 if (IS_ERR(rescuer->task))
2982                         goto err;
2983 
2984                 rescuer->task->flags |= PF_THREAD_BOUND;
2985                 wake_up_process(rescuer->task);
2986         }
2987 
2988         /*
2989          * workqueue_lock protects global freeze state and workqueues
2990          * list.  Grab it, set max_active accordingly and add the new
2991          * workqueue to workqueues list.
2992          */
2993         spin_lock(&workqueue_lock);
2994 
2995         if (workqueue_freezing && wq->flags & WQ_FREEZABLE)
2996                 for_each_cwq_cpu(cpu, wq)
2997                         get_cwq(cpu, wq)->max_active = 0;
2998 
2999         list_add(&wq->list, &workqueues);
3000 
3001         spin_unlock(&workqueue_lock);
3002 
3003         return wq;
3004 err:
3005         if (wq) {
3006                 free_cwqs(wq);
3007                 free_mayday_mask(wq->mayday_mask);
3008                 kfree(wq->rescuer);
3009                 kfree(wq);
3010         }
3011         return NULL;
3012 }
3013 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
3014 
3015 /**
3016  * destroy_workqueue - safely terminate a workqueue
3017  * @wq: target workqueue
3018  *
3019  * Safely destroy a workqueue. All work currently pending will be done first.
3020  */
3021 void destroy_workqueue(struct workqueue_struct *wq)
3022 {
3023         unsigned int flush_cnt = 0;
3024         unsigned int cpu;
3025 
3026         /*
3027          * Mark @wq dying and drain all pending works.  Once WQ_DYING is
3028          * set, only chain queueing is allowed.  IOW, only currently
3029          * pending or running work items on @wq can queue further work
3030          * items on it.  @wq is flushed repeatedly until it becomes empty.
3031          * The number of flushing is detemined by the depth of chaining and
3032          * should be relatively short.  Whine if it takes too long.
3033          */
3034         wq->flags |= WQ_DYING;
3035 reflush:
3036         flush_workqueue(wq);
3037 
3038         for_each_cwq_cpu(cpu, wq) {
3039                 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3040                 bool drained;
3041 
3042                 spin_lock_irq(&cwq->gcwq->lock);
3043                 drained = !cwq->nr_active && list_empty(&cwq->delayed_works);
3044                 spin_unlock_irq(&cwq->gcwq->lock);
3045 
3046                 if (drained)
3047                         continue;
3048 
3049                 if (++flush_cnt == 10 ||
3050                     (flush_cnt % 100 == 0 && flush_cnt <= 1000))
3051                         printk(KERN_WARNING "workqueue %s: flush on "
3052                                "destruction isn't complete after %u tries\n",
3053                                wq->name, flush_cnt);
3054                 goto reflush;
3055         }
3056 
3057         /*
3058          * wq list is used to freeze wq, remove from list after
3059          * flushing is complete in case freeze races us.
3060          */
3061         spin_lock(&workqueue_lock);
3062         list_del(&wq->list);
3063         spin_unlock(&workqueue_lock);
3064 
3065         /* sanity check */
3066         for_each_cwq_cpu(cpu, wq) {
3067                 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3068                 int i;
3069 
3070                 for (i = 0; i < WORK_NR_COLORS; i++)
3071                         BUG_ON(cwq->nr_in_flight[i]);
3072                 BUG_ON(cwq->nr_active);
3073                 BUG_ON(!list_empty(&cwq->delayed_works));
3074         }
3075 
3076         if (wq->flags & WQ_RESCUER) {
3077                 kthread_stop(wq->rescuer->task);
3078                 free_mayday_mask(wq->mayday_mask);
3079                 kfree(wq->rescuer);
3080         }
3081 
3082         free_cwqs(wq);
3083         kfree(wq);
3084 }
3085 EXPORT_SYMBOL_GPL(destroy_workqueue);
3086 
3087 /**
3088  * workqueue_set_max_active - adjust max_active of a workqueue
3089  * @wq: target workqueue
3090  * @max_active: new max_active value.
3091  *
3092  * Set max_active of @wq to @max_active.
3093  *
3094  * CONTEXT:
3095  * Don't call from IRQ context.
3096  */
3097 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
3098 {
3099         unsigned int cpu;
3100 
3101         max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
3102 
3103         spin_lock(&workqueue_lock);
3104 
3105         wq->saved_max_active = max_active;
3106 
3107         for_each_cwq_cpu(cpu, wq) {
3108                 struct global_cwq *gcwq = get_gcwq(cpu);
3109 
3110                 spin_lock_irq(&gcwq->lock);
3111 
3112                 if (!(wq->flags & WQ_FREEZABLE) ||
3113                     !(gcwq->flags & GCWQ_FREEZING))
3114                         get_cwq(gcwq->cpu, wq)->max_active = max_active;
3115 
3116                 spin_unlock_irq(&gcwq->lock);
3117         }
3118 
3119         spin_unlock(&workqueue_lock);
3120 }
3121 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
3122 
3123 /**
3124  * workqueue_congested - test whether a workqueue is congested
3125  * @cpu: CPU in question
3126  * @wq: target workqueue
3127  *
3128  * Test whether @wq's cpu workqueue for @cpu is congested.  There is
3129  * no synchronization around this function and the test result is
3130  * unreliable and only useful as advisory hints or for debugging.
3131  *
3132  * RETURNS:
3133  * %true if congested, %false otherwise.
3134  */
3135 bool workqueue_congested(unsigned int cpu, struct workqueue_struct *wq)
3136 {
3137         struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3138 
3139         return !list_empty(&cwq->delayed_works);
3140 }
3141 EXPORT_SYMBOL_GPL(workqueue_congested);
3142 
3143 /**
3144  * work_cpu - return the last known associated cpu for @work
3145  * @work: the work of interest
3146  *
3147  * RETURNS:
3148  * CPU number if @work was ever queued.  WORK_CPU_NONE otherwise.
3149  */
3150 unsigned int work_cpu(struct work_struct *work)
3151 {
3152         struct global_cwq *gcwq = get_work_gcwq(work);
3153 
3154         return gcwq ? gcwq->cpu : WORK_CPU_NONE;
3155 }
3156 EXPORT_SYMBOL_GPL(work_cpu);
3157 
3158 /**
3159  * work_busy - test whether a work is currently pending or running
3160  * @work: the work to be tested
3161  *
3162  * Test whether @work is currently pending or running.  There is no
3163  * synchronization around this function and the test result is
3164  * unreliable and only useful as advisory hints or for debugging.
3165  * Especially for reentrant wqs, the pending state might hide the
3166  * running state.
3167  *
3168  * RETURNS:
3169  * OR'd bitmask of WORK_BUSY_* bits.
3170  */
3171 unsigned int work_busy(struct work_struct *work)
3172 {
3173         struct global_cwq *gcwq = get_work_gcwq(work);
3174         unsigned long flags;
3175         unsigned int ret = 0;
3176 
3177         if (!gcwq)
3178                 return false;
3179 
3180         spin_lock_irqsave(&gcwq->lock, flags);
3181 
3182         if (work_pending(work))
3183                 ret |= WORK_BUSY_PENDING;
3184         if (find_worker_executing_work(gcwq, work))
3185                 ret |= WORK_BUSY_RUNNING;
3186 
3187         spin_unlock_irqrestore(&gcwq->lock, flags);
3188 
3189         return ret;
3190 }
3191 EXPORT_SYMBOL_GPL(work_busy);
3192 
3193 /*
3194  * CPU hotplug.
3195  *
3196  * There are two challenges in supporting CPU hotplug.  Firstly, there
3197  * are a lot of assumptions on strong associations among work, cwq and
3198  * gcwq which make migrating pending and scheduled works very
3199  * difficult to implement without impacting hot paths.  Secondly,
3200  * gcwqs serve mix of short, long and very long running works making
3201  * blocked draining impractical.
3202  *
3203  * This is solved by allowing a gcwq to be detached from CPU, running
3204  * it with unbound (rogue) workers and allowing it to be reattached
3205  * later if the cpu comes back online.  A separate thread is created
3206  * to govern a gcwq in such state and is called the trustee of the
3207  * gcwq.
3208  *
3209  * Trustee states and their descriptions.
3210  *
3211  * START        Command state used on startup.  On CPU_DOWN_PREPARE, a
3212  *              new trustee is started with this state.
3213  *
3214  * IN_CHARGE    Once started, trustee will enter this state after
3215  *              assuming the manager role and making all existing
3216  *              workers rogue.  DOWN_PREPARE waits for trustee to
3217  *              enter this state.  After reaching IN_CHARGE, trustee
3218  *              tries to execute the pending worklist until it's empty
3219  *              and the state is set to BUTCHER, or the state is set
3220  *              to RELEASE.
3221  *
3222  * BUTCHER      Command state which is set by the cpu callback after
3223  *              the cpu has went down.  Once this state is set trustee
3224  *              knows that there will be no new works on the worklist
3225  *              and once the worklist is empty it can proceed to
3226  *              killing idle workers.
3227  *
3228  * RELEASE      Command state which is set by the cpu callback if the
3229  *              cpu down has been canceled or it has come online
3230  *              again.  After recognizing this state, trustee stops
3231  *              trying to drain or butcher and clears ROGUE, rebinds
3232  *              all remaining workers back to the cpu and releases
3233  *              manager role.
3234  *
3235  * DONE         Trustee will enter this state after BUTCHER or RELEASE
3236  *              is complete.
3237  *
3238  *          trustee                 CPU                draining
3239  *         took over                down               complete
3240  * START -----------> IN_CHARGE -----------> BUTCHER -----------> DONE
3241  *                        |                     |                  ^
3242  *                        | CPU is back online  v   return workers |
3243  *                         ----------------> RELEASE --------------
3244  */
3245 
3246 /**
3247  * trustee_wait_event_timeout - timed event wait for trustee
3248  * @cond: condition to wait for
3249  * @timeout: timeout in jiffies
3250  *
3251  * wait_event_timeout() for trustee to use.  Handles locking and
3252  * checks for RELEASE request.
3253  *
3254  * CONTEXT:
3255  * spin_lock_irq(gcwq->lock) which may be released and regrabbed
3256  * multiple times.  To be used by trustee.
3257  *
3258  * RETURNS:
3259  * Positive indicating left time if @cond is satisfied, 0 if timed
3260  * out, -1 if canceled.
3261  */
3262 #define trustee_wait_event_timeout(cond, timeout) ({                    \
3263         long __ret = (timeout);                                         \
3264         while (!((cond) || (gcwq->trustee_state == TRUSTEE_RELEASE)) && \
3265                __ret) {                                                 \
3266                 spin_unlock_irq(&gcwq->lock);                           \
3267                 __wait_event_timeout(gcwq->trustee_wait, (cond) ||      \
3268                         (gcwq->trustee_state == TRUSTEE_RELEASE),       \
3269                         __ret);                                         \
3270                 spin_lock_irq(&gcwq->lock);                             \
3271         }                                                               \
3272         gcwq->trustee_state == TRUSTEE_RELEASE ? -1 : (__ret);          \
3273 })
3274 
3275 /**
3276  * trustee_wait_event - event wait for trustee
3277  * @cond: condition to wait for
3278  *
3279  * wait_event() for trustee to use.  Automatically handles locking and
3280  * checks for CANCEL request.
3281  *
3282  * CONTEXT:
3283  * spin_lock_irq(gcwq->lock) which may be released and regrabbed
3284  * multiple times.  To be used by trustee.
3285  *
3286  * RETURNS:
3287  * 0 if @cond is satisfied, -1 if canceled.
3288  */
3289 #define trustee_wait_event(cond) ({                                     \
3290         long __ret1;                                                    \
3291         __ret1 = trustee_wait_event_timeout(cond, MAX_SCHEDULE_TIMEOUT);\
3292         __ret1 < 0 ? -1 : 0;                                            \
3293 })
3294 
3295 static int __cpuinit trustee_thread(void *__gcwq)
3296 {
3297         struct global_cwq *gcwq = __gcwq;
3298         struct worker *worker;
3299         struct work_struct *work;
3300         struct hlist_node *pos;
3301         long rc;
3302         int i;
3303 
3304         BUG_ON(gcwq->cpu != smp_processor_id());
3305 
3306         spin_lock_irq(&gcwq->lock);
3307         /*
3308          * Claim the manager position and make all workers rogue.
3309          * Trustee must be bound to the target cpu and can't be
3310          * cancelled.
3311          */
3312         BUG_ON(gcwq->cpu != smp_processor_id());
3313         rc = trustee_wait_event(!(gcwq->flags & GCWQ_MANAGING_WORKERS));
3314         BUG_ON(rc < 0);
3315 
3316         gcwq->flags |= GCWQ_MANAGING_WORKERS;
3317 
3318         list_for_each_entry(worker, &gcwq->idle_list, entry)
3319                 worker->flags |= WORKER_ROGUE;
3320 
3321         for_each_busy_worker(worker, i, pos, gcwq)
3322                 worker->flags |= WORKER_ROGUE;
3323 
3324         /*
3325          * Call schedule() so that we cross rq->lock and thus can
3326          * guarantee sched callbacks see the rogue flag.  This is
3327          * necessary as scheduler callbacks may be invoked from other
3328          * cpus.
3329          */
3330         spin_unlock_irq(&gcwq->lock);
3331         schedule();
3332         spin_lock_irq(&gcwq->lock);
3333 
3334         /*
3335          * Sched callbacks are disabled now.  Zap nr_running.  After
3336          * this, nr_running stays zero and need_more_worker() and
3337          * keep_working() are always true as long as the worklist is
3338          * not empty.
3339          */
3340         atomic_set(get_gcwq_nr_running(gcwq->cpu), 0);
3341 
3342         spin_unlock_irq(&gcwq->lock);
3343         del_timer_sync(&gcwq->idle_timer);
3344         spin_lock_irq(&gcwq->lock);
3345 
3346         /*
3347          * We're now in charge.  Notify and proceed to drain.  We need
3348          * to keep the gcwq running during the whole CPU down
3349          * procedure as other cpu hotunplug callbacks may need to
3350          * flush currently running tasks.
3351          */
3352         gcwq->trustee_state = TRUSTEE_IN_CHARGE;
3353         wake_up_all(&gcwq->trustee_wait);
3354 
3355         /*
3356          * The original cpu is in the process of dying and may go away
3357          * anytime now.  When that happens, we and all workers would
3358          * be migrated to other cpus.  Try draining any left work.  We
3359          * want to get it over with ASAP - spam rescuers, wake up as
3360          * many idlers as necessary and create new ones till the
3361          * worklist is empty.  Note that if the gcwq is frozen, there
3362          * may be frozen works in freezable cwqs.  Don't declare
3363          * completion while frozen.
3364          */
3365         while (gcwq->nr_workers != gcwq->nr_idle ||
3366                gcwq->flags & GCWQ_FREEZING ||
3367                gcwq->trustee_state == TRUSTEE_IN_CHARGE) {
3368                 int nr_works = 0;
3369 
3370                 list_for_each_entry(work, &gcwq->worklist, entry) {
3371                         send_mayday(work);
3372                         nr_works++;
3373                 }
3374 
3375                 list_for_each_entry(worker, &gcwq->idle_list, entry) {
3376                         if (!nr_works--)
3377                                 break;
3378                         wake_up_process(worker->task);
3379                 }
3380 
3381                 if (need_to_create_worker(gcwq)) {
3382                         spin_unlock_irq(&gcwq->lock);
3383                         worker = create_worker(gcwq, false);
3384                         spin_lock_irq(&gcwq->lock);
3385                         if (worker) {
3386                                 worker->flags |= WORKER_ROGUE;
3387                                 start_worker(worker);
3388                         }
3389                 }
3390 
3391                 /* give a breather */
3392                 if (trustee_wait_event_timeout(false, TRUSTEE_COOLDOWN) < 0)
3393                         break;
3394         }
3395 
3396         /*
3397          * Either all works have been scheduled and cpu is down, or
3398          * cpu down has already been canceled.  Wait for and butcher
3399          * all workers till we're canceled.
3400          */
3401         do {
3402                 rc = trustee_wait_event(!list_empty(&gcwq->idle_list));
3403                 while (!list_empty(&gcwq->idle_list))
3404                         destroy_worker(list_first_entry(&gcwq->idle_list,
3405                                                         struct worker, entry));
3406         } while (gcwq->nr_workers && rc >= 0);
3407 
3408         /*
3409          * At this point, either draining has completed and no worker
3410          * is left, or cpu down has been canceled or the cpu is being
3411          * brought back up.  There shouldn't be any idle one left.
3412          * Tell the remaining busy ones to rebind once it finishes the
3413          * currently scheduled works by scheduling the rebind_work.
3414          */
3415         WARN_ON(!list_empty(&gcwq->idle_list));
3416 
3417         for_each_busy_worker(worker, i, pos, gcwq) {
3418                 struct work_struct *rebind_work = &worker->rebind_work;
3419                 unsigned long worker_flags = worker->flags;
3420 
3421                 /*
3422                  * Rebind_work may race with future cpu hotplug
3423                  * operations.  Use a separate flag to mark that
3424                  * rebinding is scheduled.  The morphing should
3425                  * be atomic.
3426                  */
3427                 worker_flags |= WORKER_REBIND;
3428                 worker_flags &= ~WORKER_ROGUE;
3429                 ACCESS_ONCE(worker->flags) = worker_flags;
3430 
3431                 /* queue rebind_work, wq doesn't matter, use the default one */
3432                 if (test_and_set_bit(WORK_STRUCT_PENDING_BIT,
3433                                      work_data_bits(rebind_work)))
3434                         continue;
3435 
3436                 debug_work_activate(rebind_work);
3437                 insert_work(get_cwq(gcwq->cpu, system_wq), rebind_work,
3438                             worker->scheduled.next,
3439                             work_color_to_flags(WORK_NO_COLOR));
3440         }
3441 
3442         /* relinquish manager role */
3443         gcwq->flags &= ~GCWQ_MANAGING_WORKERS;
3444 
3445         /* notify completion */
3446         gcwq->trustee = NULL;
3447         gcwq->trustee_state = TRUSTEE_DONE;
3448         wake_up_all(&gcwq->trustee_wait);
3449         spin_unlock_irq(&gcwq->lock);
3450         return 0;
3451 }
3452 
3453 /**
3454  * wait_trustee_state - wait for trustee to enter the specified state
3455  * @gcwq: gcwq the trustee of interest belongs to
3456  * @state: target state to wait for
3457  *
3458  * Wait for the trustee to reach @state.  DONE is already matched.
3459  *
3460  * CONTEXT:
3461  * spin_lock_irq(gcwq->lock) which may be released and regrabbed
3462  * multiple times.  To be used by cpu_callback.
3463  */
3464 static void __cpuinit wait_trustee_state(struct global_cwq *gcwq, int state)
3465 __releases(&gcwq->lock)
3466 __acquires(&gcwq->lock)
3467 {
3468         if (!(gcwq->trustee_state == state ||
3469               gcwq->trustee_state == TRUSTEE_DONE)) {
3470                 spin_unlock_irq(&gcwq->lock);
3471                 __wait_event(gcwq->trustee_wait,
3472                              gcwq->trustee_state == state ||
3473                              gcwq->trustee_state == TRUSTEE_DONE);
3474                 spin_lock_irq(&gcwq->lock);
3475         }
3476 }
3477 
3478 static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
3479                                                 unsigned long action,
3480                                                 void *hcpu)
3481 {
3482         unsigned int cpu = (unsigned long)hcpu;
3483         struct global_cwq *gcwq = get_gcwq(cpu);
3484         struct task_struct *new_trustee = NULL;
3485         struct worker *uninitialized_var(new_worker);
3486         unsigned long flags;
3487 
3488         action &= ~CPU_TASKS_FROZEN;
3489 
3490         switch (action) {
3491         case CPU_DOWN_PREPARE:
3492                 new_trustee = kthread_create(trustee_thread, gcwq,
3493                                              "workqueue_trustee/%d\n", cpu);
3494                 if (IS_ERR(new_trustee))
3495                         return notifier_from_errno(PTR_ERR(new_trustee));
3496                 kthread_bind(new_trustee, cpu);
3497                 /* fall through */
3498         case CPU_UP_PREPARE:
3499                 BUG_ON(gcwq->first_idle);
3500                 new_worker = create_worker(gcwq, false);
3501                 if (!new_worker) {
3502                         if (new_trustee)
3503                                 kthread_stop(new_trustee);
3504                         return NOTIFY_BAD;
3505                 }
3506         }
3507 
3508         /* some are called w/ irq disabled, don't disturb irq status */
3509         spin_lock_irqsave(&gcwq->lock, flags);
3510 
3511         switch (action) {
3512         case CPU_DOWN_PREPARE:
3513                 /* initialize trustee and tell it to acquire the gcwq */
3514                 BUG_ON(gcwq->trustee || gcwq->trustee_state != TRUSTEE_DONE);
3515                 gcwq->trustee = new_trustee;
3516                 gcwq->trustee_state = TRUSTEE_START;
3517                 wake_up_process(gcwq->trustee);
3518                 wait_trustee_state(gcwq, TRUSTEE_IN_CHARGE);
3519                 /* fall through */
3520         case CPU_UP_PREPARE:
3521                 BUG_ON(gcwq->first_idle);
3522                 gcwq->first_idle = new_worker;
3523                 break;
3524 
3525         case CPU_DYING:
3526                 /*
3527                  * Before this, the trustee and all workers except for
3528                  * the ones which are still executing works from
3529                  * before the last CPU down must be on the cpu.  After
3530                  * this, they'll all be diasporas.
3531                  */
3532                 gcwq->flags |= GCWQ_DISASSOCIATED;
3533                 break;
3534 
3535         case CPU_POST_DEAD:
3536                 gcwq->trustee_state = TRUSTEE_BUTCHER;
3537                 /* fall through */
3538         case CPU_UP_CANCELED:
3539                 destroy_worker(gcwq->first_idle);
3540                 gcwq->first_idle = NULL;
3541                 break;
3542 
3543         case CPU_DOWN_FAILED:
3544         case CPU_ONLINE:
3545                 gcwq->flags &= ~GCWQ_DISASSOCIATED;
3546                 if (gcwq->trustee_state != TRUSTEE_DONE) {
3547                         gcwq->trustee_state = TRUSTEE_RELEASE;
3548                         wake_up_process(gcwq->trustee);
3549                         wait_trustee_state(gcwq, TRUSTEE_DONE);
3550                 }
3551 
3552                 /*
3553                  * Trustee is done and there might be no worker left.
3554                  * Put the first_idle in and request a real manager to
3555                  * take a look.
3556                  */
3557                 spin_unlock_irq(&gcwq->lock);
3558                 kthread_bind(gcwq->first_idle->task, cpu);
3559                 spin_lock_irq(&gcwq->lock);
3560                 gcwq->flags |= GCWQ_MANAGE_WORKERS;
3561                 start_worker(gcwq->first_idle);
3562                 gcwq->first_idle = NULL;
3563                 break;
3564         }
3565 
3566         spin_unlock_irqrestore(&gcwq->lock, flags);
3567 
3568         return notifier_from_errno(0);
3569 }
3570 
3571 /*
3572  * Workqueues should be brought up before normal priority CPU notifiers.
3573  * This will be registered high priority CPU notifier.
3574  */
3575 static int __devinit workqueue_cpu_up_callback(struct notifier_block *nfb,
3576                                                unsigned long action,
3577                                                void *hcpu)
3578 {
3579         switch (action & ~CPU_TASKS_FROZEN) {
3580         case CPU_UP_PREPARE:
3581         case CPU_UP_CANCELED:
3582         case CPU_DOWN_FAILED:
3583         case CPU_ONLINE:
3584                 return workqueue_cpu_callback(nfb, action, hcpu);
3585         }
3586         return NOTIFY_OK;
3587 }
3588 
3589 /*
3590  * Workqueues should be brought down after normal priority CPU notifiers.
3591  * This will be registered as low priority CPU notifier.
3592  */
3593 static int __devinit workqueue_cpu_down_callback(struct notifier_block *nfb,
3594                                                  unsigned long action,
3595                                                  void *hcpu)
3596 {
3597         switch (action & ~CPU_TASKS_FROZEN) {
3598         case CPU_DOWN_PREPARE:
3599         case CPU_DYING:
3600         case CPU_POST_DEAD:
3601                 return workqueue_cpu_callback(nfb, action, hcpu);
3602         }
3603         return NOTIFY_OK;
3604 }
3605 
3606 #ifdef CONFIG_SMP
3607 
3608 struct work_for_cpu {
3609         struct work_struct work;
3610         long (*fn)(void *);
3611         void *arg;
3612         long ret;
3613 };
3614 
3615 static void work_for_cpu_fn(struct work_struct *work)
3616 {
3617         struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
3618 
3619         wfc->ret = wfc->fn(wfc->arg);
3620 }
3621 
3622 /**
3623  * work_on_cpu - run a function in user context on a particular cpu
3624  * @cpu: the cpu to run on
3625  * @fn: the function to run
3626  * @arg: the function arg
3627  *
3628  * This will return the value @fn returns.
3629  * It is up to the caller to ensure that the cpu doesn't go offline.
3630  * The caller must not hold any locks which would prevent @fn from completing.
3631  */
3632 long work_on_cpu(unsigned int cpu, long (*fn)(void *), void *arg)
3633 {
3634         struct work_for_cpu wfc = { .fn = fn, .arg = arg };
3635 
3636         INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
3637         schedule_work_on(cpu, &wfc.work);
3638         flush_work(&wfc.work);
3639         return wfc.ret;
3640 }
3641 EXPORT_SYMBOL_GPL(work_on_cpu);
3642 #endif /* CONFIG_SMP */
3643 
3644 #ifdef CONFIG_FREEZER
3645 
3646 /**
3647  * freeze_workqueues_begin - begin freezing workqueues
3648  *
3649  * Start freezing workqueues.  After this function returns, all freezable
3650  * workqueues will queue new works to their frozen_works list instead of
3651  * gcwq->worklist.
3652  *
3653  * CONTEXT:
3654  * Grabs and releases workqueue_lock and gcwq->lock's.
3655  */
3656 void freeze_workqueues_begin(void)
3657 {
3658         unsigned int cpu;
3659 
3660         spin_lock(&workqueue_lock);
3661 
3662         BUG_ON(workqueue_freezing);
3663         workqueue_freezing = true;
3664 
3665         for_each_gcwq_cpu(cpu) {
3666                 struct global_cwq *gcwq = get_gcwq(cpu);
3667                 struct workqueue_struct *wq;
3668 
3669                 spin_lock_irq(&gcwq->lock);
3670 
3671                 BUG_ON(gcwq->flags & GCWQ_FREEZING);
3672                 gcwq->flags |= GCWQ_FREEZING;
3673 
3674                 list_for_each_entry(wq, &workqueues, list) {
3675                         struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3676 
3677                         if (cwq && wq->flags & WQ_FREEZABLE)
3678                                 cwq->max_active = 0;
3679                 }
3680 
3681                 spin_unlock_irq(&gcwq->lock);
3682         }
3683 
3684         spin_unlock(&workqueue_lock);
3685 }
3686 
3687 /**
3688  * freeze_workqueues_busy - are freezable workqueues still busy?
3689  *
3690  * Check whether freezing is complete.  This function must be called
3691  * between freeze_workqueues_begin() and thaw_workqueues().
3692  *
3693  * CONTEXT:
3694  * Grabs and releases workqueue_lock.
3695  *
3696  * RETURNS:
3697  * %true if some freezable workqueues are still busy.  %false if freezing
3698  * is complete.
3699  */
3700 bool freeze_workqueues_busy(void)
3701 {
3702         unsigned int cpu;
3703         bool busy = false;
3704 
3705         spin_lock(&workqueue_lock);
3706 
3707         BUG_ON(!workqueue_freezing);
3708 
3709         for_each_gcwq_cpu(cpu) {
3710                 struct workqueue_struct *wq;
3711                 /*
3712                  * nr_active is monotonically decreasing.  It's safe
3713                  * to peek without lock.
3714                  */
3715                 list_for_each_entry(wq, &workqueues, list) {
3716                         struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3717 
3718                         if (!cwq || !(wq->flags & WQ_FREEZABLE))
3719                                 continue;
3720 
3721                         BUG_ON(cwq->nr_active < 0);
3722                         if (cwq->nr_active) {
3723                                 busy = true;
3724                                 goto out_unlock;
3725                         }
3726                 }
3727         }
3728 out_unlock:
3729         spin_unlock(&workqueue_lock);
3730         return busy;
3731 }
3732 
3733 /**
3734  * thaw_workqueues - thaw workqueues
3735  *
3736  * Thaw workqueues.  Normal queueing is restored and all collected
3737  * frozen works are transferred to their respective gcwq worklists.
3738  *
3739  * CONTEXT:
3740  * Grabs and releases workqueue_lock and gcwq->lock's.
3741  */
3742 void thaw_workqueues(void)
3743 {
3744         unsigned int cpu;
3745 
3746         spin_lock(&workqueue_lock);
3747 
3748         if (!workqueue_freezing)
3749                 goto out_unlock;
3750 
3751         for_each_gcwq_cpu(cpu) {
3752                 struct global_cwq *gcwq = get_gcwq(cpu);
3753                 struct workqueue_struct *wq;
3754 
3755                 spin_lock_irq(&gcwq->lock);
3756 
3757                 BUG_ON(!(gcwq->flags & GCWQ_FREEZING));
3758                 gcwq->flags &= ~GCWQ_FREEZING;
3759 
3760                 list_for_each_entry(wq, &workqueues, list) {
3761                         struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3762 
3763                         if (!cwq || !(wq->flags & WQ_FREEZABLE))
3764                                 continue;
3765 
3766                         /* restore max_active and repopulate worklist */
3767                         cwq->max_active = wq->saved_max_active;
3768 
3769                         while (!list_empty(&cwq->delayed_works) &&
3770                                cwq->nr_active < cwq->max_active)
3771                                 cwq_activate_first_delayed(cwq);
3772                 }
3773 
3774                 wake_up_worker(gcwq);
3775 
3776                 spin_unlock_irq(&gcwq->lock);
3777         }
3778 
3779         workqueue_freezing = false;
3780 out_unlock:
3781         spin_unlock(&workqueue_lock);
3782 }
3783 #endif /* CONFIG_FREEZER */
3784 
3785 static int __init init_workqueues(void)
3786 {
3787         unsigned int cpu;
3788         int i;
3789 
3790         cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
3791         cpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
3792 
3793         /* initialize gcwqs */
3794         for_each_gcwq_cpu(cpu) {
3795                 struct global_cwq *gcwq = get_gcwq(cpu);
3796 
3797                 spin_lock_init(&gcwq->lock);
3798                 INIT_LIST_HEAD(&gcwq->worklist);
3799                 gcwq->cpu = cpu;
3800                 gcwq->flags |= GCWQ_DISASSOCIATED;
3801 
3802                 INIT_LIST_HEAD(&gcwq->idle_list);
3803                 for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++)
3804                         INIT_HLIST_HEAD(&gcwq->busy_hash[i]);
3805 
3806                 init_timer_deferrable(&gcwq->idle_timer);
3807                 gcwq->idle_timer.function = idle_worker_timeout;
3808                 gcwq->idle_timer.data = (unsigned long)gcwq;
3809 
3810                 setup_timer(&gcwq->mayday_timer, gcwq_mayday_timeout,
3811                             (unsigned long)gcwq);
3812 
3813                 ida_init(&gcwq->worker_ida);
3814 
3815                 gcwq->trustee_state = TRUSTEE_DONE;
3816                 init_waitqueue_head(&gcwq->trustee_wait);
3817         }
3818 
3819         /* create the initial worker */
3820         for_each_online_gcwq_cpu(cpu) {
3821                 struct global_cwq *gcwq = get_gcwq(cpu);
3822                 struct worker *worker;
3823 
3824                 if (cpu != WORK_CPU_UNBOUND)
3825                         gcwq->flags &= ~GCWQ_DISASSOCIATED;
3826                 worker = create_worker(gcwq, true);
3827                 BUG_ON(!worker);
3828                 spin_lock_irq(&gcwq->lock);
3829                 start_worker(worker);
3830                 spin_unlock_irq(&gcwq->lock);
3831         }
3832 
3833         system_wq = alloc_workqueue("events", 0, 0);
3834         system_long_wq = alloc_workqueue("events_long", 0, 0);
3835         system_nrt_wq = alloc_workqueue("events_nrt", WQ_NON_REENTRANT, 0);
3836         system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
3837                                             WQ_UNBOUND_MAX_ACTIVE);
3838         system_freezable_wq = alloc_workqueue("events_freezable",
3839                                               WQ_FREEZABLE, 0);
3840         system_nrt_freezable_wq = alloc_workqueue("events_nrt_freezable",
3841                         WQ_NON_REENTRANT | WQ_FREEZABLE, 0);
3842         BUG_ON(!system_wq || !system_long_wq || !system_nrt_wq ||
3843                !system_unbound_wq || !system_freezable_wq ||
3844                 !system_nrt_freezable_wq);
3845         return 0;
3846 }
3847 early_initcall(init_workqueues);
3848 

~ [ 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