kernel/workqueue.c

   1 /*
   2  * linux/kernel/workqueue.c
   3  *
   4  * Generic mechanism for defining kernel helper threads for running
   5  * arbitrary tasks in process context.
   6  *
   7  * Started by Ingo Molnar, Copyright (C) 2002
   8  *
   9  * Derived from the taskqueue/keventd code by:
  10  *
  11  *   David Woodhouse <dwmw2@infradead.org>
  12  *   Andrew Morton <andrewm@uow.edu.au>
  13  *   Kai Petzke <wpp@marie.physik.tu-berlin.de>
  14  *   Theodore Ts'o <tytso@mit.edu>
  15  *
  16  * Made to use alloc_percpu by Christoph Lameter <clameter@sgi.com>.
  17  */
  18
  19 #include <linux/module.h>
  20 #include <linux/kernel.h>
  21 #include <linux/sched.h>
  22 #include <linux/init.h>
  23 #include <linux/signal.h>
  24 #include <linux/completion.h>
  25 #include <linux/workqueue.h>
  26 #include <linux/slab.h>
  27 #include <linux/cpu.h>
  28 #include <linux/notifier.h>
  29 #include <linux/kthread.h>
  30 #include <linux/hardirq.h>
  31 #include <linux/mempolicy.h>
  32 #include <linux/freezer.h>
  33 #include <linux/kallsyms.h>
  34 #include <linux/debug_locks.h>
  35 #include <linux/lockdep.h>
  36
  37 /*
  38  * The per-CPU workqueue (if single thread, we always use the first
  39  * possible cpu).
  40  */
  41 struct cpu_workqueue_struct {
  42
  43         spinlock_t lock;
  44
  45         struct list_head worklist;
  46         wait_queue_head_t more_work;
  47         struct work_struct *current_work;
  48
  49         struct workqueue_struct *wq;
  50         struct task_struct *thread;
  51
  52         int run_depth;          /* Detect run_workqueue() recursion depth */
  53 } ____cacheline_aligned;
  54
  55 /*
  56  * The externally visible workqueue abstraction is an array of
  57  * per-CPU workqueues:
  58  */
  59 struct workqueue_struct {
  60         struct cpu_workqueue_struct *cpu_wq;
  61         struct list_head list;
  62         const char *name;
  63         int singlethread;
  64         int freezeable;         /* Freeze threads during suspend */
  65 #ifdef CONFIG_LOCKDEP
  66         struct lockdep_map lockdep_map;
  67 #endif
  68 };
  69
  70 /* Serializes the accesses to the list of workqueues. */
  71 static DEFINE_SPINLOCK(workqueue_lock);
  72 static LIST_HEAD(workqueues);
  73
  74 static int singlethread_cpu __read_mostly;
  75 static cpumask_t cpu_singlethread_map __read_mostly;
  76 /*
  77  * _cpu_down() first removes CPU from cpu_online_map, then CPU_DEAD
  78  * flushes cwq->worklist. This means that flush_workqueue/wait_on_work
  79  * which comes in between can't use for_each_online_cpu(). We could
  80  * use cpu_possible_map, the cpumask below is more a documentation
  81  * than optimization.
  82  */
  83 static cpumask_t cpu_populated_map __read_mostly;
  84
  85 /* If it's single threaded, it isn't in the list of workqueues. */
  86 static inline int is_single_threaded(struct workqueue_struct *wq)
  87 {
  88         return wq->singlethread;
  89 }
  90
  91 static const cpumask_t *wq_cpu_map(struct workqueue_struct *wq)
  92 {
  93         return is_single_threaded(wq)
  94                 ? &cpu_singlethread_map : &cpu_populated_map;
  95 }
  96
  97 static
  98 struct cpu_workqueue_struct *wq_per_cpu(struct workqueue_struct *wq, int cpu)
  99 {
 100         if (unlikely(is_single_threaded(wq)))
 101                 cpu = singlethread_cpu;
 102         return per_cpu_ptr(wq->cpu_wq, cpu);
 103 }
 104
 105 /*
 106  * Set the workqueue on which a work item is to be run
 107  * - Must *only* be called if the pending flag is set
 108  */
 109 static inline void set_wq_data(struct work_struct *work,
 110                                 struct cpu_workqueue_struct *cwq)
 111 {
 112         unsigned long new;
 113
 114         BUG_ON(!work_pending(work));
 115
 116         new = (unsigned long) cwq | (1UL << WORK_STRUCT_PENDING);
 117         new |= WORK_STRUCT_FLAG_MASK & *work_data_bits(work);
 118         atomic_long_set(&work->data, new);
 119 }
 120
 121 static inline
 122 struct cpu_workqueue_struct *get_wq_data(struct work_struct *work)
 123 {
 124         return (void *) (atomic_long_read(&work->data) & WORK_STRUCT_WQ_DATA_MASK);
 125 }
 126
 127 static void insert_work(struct cpu_workqueue_struct *cwq,
 128                                 struct work_struct *work, int tail)
 129 {
 130         set_wq_data(work, cwq);
 131         /*
 132          * Ensure that we get the right work->data if we see the
 133          * result of list_add() below, see try_to_grab_pending().
 134          */
 135         smp_wmb();
 136         if (tail)
 137                 list_add_tail(&work->entry, &cwq->worklist);
 138         else
 139                 list_add(&work->entry, &cwq->worklist);
 140         wake_up(&cwq->more_work);
 141 }
 142
 143 /* Preempt must be disabled. */
 144 static void __queue_work(struct cpu_workqueue_struct *cwq,
 145                          struct work_struct *work)
 146 {
 147         unsigned long flags;
 148
 149         spin_lock_irqsave(&cwq->lock, flags);
 150         insert_work(cwq, work, 1);
 151         spin_unlock_irqrestore(&cwq->lock, flags);
 152 }
 153
 154 /**
 155  * queue_work - queue work on a workqueue
 156  * @wq: workqueue to use
 157  * @work: work to queue
 158  *
 159  * Returns 0 if @work was already on a queue, non-zero otherwise.
 160  *
 161  * We queue the work to the CPU on which it was submitted, but if the CPU dies
 162  * it can be processed by another CPU.
 163  */
 164 int queue_work(struct workqueue_struct *wq, struct work_struct *work)
 165 {
 166         int ret = 0;
 167
 168         if (!test_and_set_bit(WORK_STRUCT_PENDING, work_data_bits(work))) {
 169                 BUG_ON(!list_empty(&work->entry));
 170                 __queue_work(wq_per_cpu(wq, get_cpu()), work);
 171                 put_cpu();
 172                 ret = 1;
 173         }
 174         return ret;
 175 }
 176 EXPORT_SYMBOL_GPL(queue_work);
 177
 178 static void delayed_work_timer_fn(unsigned long __data)
 179 {
 180         struct delayed_work *dwork = (struct delayed_work *)__data;
 181         struct cpu_workqueue_struct *cwq = get_wq_data(&dwork->work);
 182         struct workqueue_struct *wq = cwq->wq;
 183
 184         __queue_work(wq_per_cpu(wq, smp_processor_id()), &dwork->work);
 185 }
 186
 187 /**
 188  * queue_delayed_work - queue work on a workqueue after delay
 189  * @wq: workqueue to use
 190  * @dwork: delayable work to queue
 191  * @delay: number of jiffies to wait before queueing
 192  *
 193  * Returns 0 if @work was already on a queue, non-zero otherwise.
 194  */
 195 int queue_delayed_work(struct workqueue_struct *wq,
 196                         struct delayed_work *dwork, unsigned long delay)
 197 {
 198         if (delay == 0)
 199                 return queue_work(wq, &dwork->work);
 200
 201         return queue_delayed_work_on(-1, wq, dwork, delay);
 202 }
 203 EXPORT_SYMBOL_GPL(queue_delayed_work);
 204
 205 /**
 206  * queue_delayed_work_on - queue work on specific CPU after delay
 207  * @cpu: CPU number to execute work on
 208  * @wq: workqueue to use
 209  * @dwork: work to queue
 210  * @delay: number of jiffies to wait before queueing
 211  *
 212  * Returns 0 if @work was already on a queue, non-zero otherwise.
 213  */
 214 int queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
 215                         struct delayed_work *dwork, unsigned long delay)
 216 {
 217         int ret = 0;
 218         struct timer_list *timer = &dwork->timer;
 219         struct work_struct *work = &dwork->work;
 220
 221         if (!test_and_set_bit(WORK_STRUCT_PENDING, work_data_bits(work))) {
 222                 BUG_ON(timer_pending(timer));
 223                 BUG_ON(!list_empty(&work->entry));
 224
 225                 timer_stats_timer_set_start_info(&dwork->timer);
 226
 227                 /* This stores cwq for the moment, for the timer_fn */
 228                 set_wq_data(work, wq_per_cpu(wq, raw_smp_processor_id()));
 229                 timer->expires = jiffies + delay;
 230                 timer->data = (unsigned long)dwork;
 231                 timer->function = delayed_work_timer_fn;
 232
 233                 if (unlikely(cpu >= 0))
 234                         add_timer_on(timer, cpu);
 235                 else
 236                         add_timer(timer);
 237                 ret = 1;
 238         }
 239         return ret;
 240 }
 241 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
 242
 243 static void run_workqueue(struct cpu_workqueue_struct *cwq)
 244 {
 245         spin_lock_irq(&cwq->lock);
 246         cwq->run_depth++;
 247         if (cwq->run_depth > 3) {
 248                 /* morton gets to eat his hat */
 249                 printk("%s: recursion depth exceeded: %d\n",
 250                         __func__, cwq->run_depth);
 251                 dump_stack();
 252         }
 253         while (!list_empty(&cwq->worklist)) {
 254                 struct work_struct *work = list_entry(cwq->worklist.next,
 255                                                 struct work_struct, entry);
 256                 work_func_t f = work->func;
 257 #ifdef CONFIG_LOCKDEP
 258                 /*
 259                  * It is permissible to free the struct work_struct
 260                  * from inside the function that is called from it,
 261                  * this we need to take into account for lockdep too.
 262                  * To avoid bogus "held lock freed" warnings as well
 263                  * as problems when looking into work->lockdep_map,
 264                  * make a copy and use that here.
 265                  */
 266                 struct lockdep_map lockdep_map = work->lockdep_map;
 267 #endif
 268
 269                 cwq->current_work = work;
 270                 list_del_init(cwq->worklist.next);
 271                 spin_unlock_irq(&cwq->lock);
 272
 273                 BUG_ON(get_wq_data(work) != cwq);
 274                 work_clear_pending(work);
 275                 lock_acquire(&cwq->wq->lockdep_map, 0, 0, 0, 2, _THIS_IP_);
 276                 lock_acquire(&lockdep_map, 0, 0, 0, 2, _THIS_IP_);
 277                 f(work);
 278                 lock_release(&lockdep_map, 1, _THIS_IP_);
 279                 lock_release(&cwq->wq->lockdep_map, 1, _THIS_IP_);
 280
 281                 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
 282                         printk(KERN_ERR "BUG: workqueue leaked lock or atomic: "
 283                                         "%s/0x%08x/%d\n",
 284                                         current->comm, preempt_count(),
 285                                         task_pid_nr(current));
 286                         printk(KERN_ERR "    last function: ");
 287                         print_symbol("%s\n", (unsigned long)f);
 288                         debug_show_held_locks(current);
 289                         dump_stack();
 290                 }
 291
 292                 spin_lock_irq(&cwq->lock);
 293                 cwq->current_work = NULL;
 294         }
 295         cwq->run_depth--;
 296         spin_unlock_irq(&cwq->lock);
 297 }
 298
 299 static int worker_thread(void *__cwq)
 300 {
 301         struct cpu_workqueue_struct *cwq = __cwq;
 302         DEFINE_WAIT(wait);
 303
 304         if (cwq->wq->freezeable)
 305                 set_freezable();
 306
 307         set_user_nice(current, -5);
 308
 309         for (;;) {
 310                 prepare_to_wait(&cwq->more_work, &wait, TASK_INTERRUPTIBLE);
 311                 if (!freezing(current) &&
 312                     !kthread_should_stop() &&
 313                     list_empty(&cwq->worklist))
 314                         schedule();
 315                 finish_wait(&cwq->more_work, &wait);
 316
 317                 try_to_freeze();
 318
 319                 if (kthread_should_stop())
 320                         break;
 321
 322                 run_workqueue(cwq);
 323         }
 324
 325         return 0;
 326 }
 327
 328 struct wq_barrier {
 329         struct work_struct      work;
 330         struct completion       done;
 331 };
 332
 333 static void wq_barrier_func(struct work_struct *work)
 334 {
 335         struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
 336         complete(&barr->done);
 337 }
 338
 339 static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
 340                                         struct wq_barrier *barr, int tail)
 341 {
 342         INIT_WORK(&barr->work, wq_barrier_func);
 343         __set_bit(WORK_STRUCT_PENDING, work_data_bits(&barr->work));
 344
 345         init_completion(&barr->done);
 346
 347         insert_work(cwq, &barr->work, tail);
 348 }
 349
 350 static int flush_cpu_workqueue(struct cpu_workqueue_struct *cwq)
 351 {
 352         int active;
 353
 354         if (cwq->thread == current) {
 355                 /*
 356                  * Probably keventd trying to flush its own queue. So simply run
 357                  * it by hand rather than deadlocking.
 358                  */
 359                 run_workqueue(cwq);
 360                 active = 1;
 361         } else {
 362                 struct wq_barrier barr;
 363
 364                 active = 0;
 365                 spin_lock_irq(&cwq->lock);
 366                 if (!list_empty(&cwq->worklist) || cwq->current_work != NULL) {
 367                         insert_wq_barrier(cwq, &barr, 1);
 368                         active = 1;
 369                 }
 370                 spin_unlock_irq(&cwq->lock);
 371
 372                 if (active)
 373                         wait_for_completion(&barr.done);
 374         }
 375
 376         return active;
 377 }
 378
 379 /**
 380  * flush_workqueue - ensure that any scheduled work has run to completion.
 381  * @wq: workqueue to flush
 382  *
 383  * Forces execution of the workqueue and blocks until its completion.
 384  * This is typically used in driver shutdown handlers.
 385  *
 386  * We sleep until all works which were queued on entry have been handled,
 387  * but we are not livelocked by new incoming ones.
 388  *
 389  * This function used to run the workqueues itself.  Now we just wait for the
 390  * helper threads to do it.
 391  */
 392 void flush_workqueue(struct workqueue_struct *wq)
 393 {
 394         const cpumask_t *cpu_map = wq_cpu_map(wq);
 395         int cpu;
 396
 397         might_sleep();
 398         lock_acquire(&wq->lockdep_map, 0, 0, 0, 2, _THIS_IP_);
 399         lock_release(&wq->lockdep_map, 1, _THIS_IP_);
 400         for_each_cpu_mask(cpu, *cpu_map)
 401                 flush_cpu_workqueue(per_cpu_ptr(wq->cpu_wq, cpu));
 402 }
 403 EXPORT_SYMBOL_GPL(flush_workqueue);
 404
 405 /*
 406  * Upon a successful return (>= 0), the caller "owns" WORK_STRUCT_PENDING bit,
 407  * so this work can't be re-armed in any way.
 408  */
 409 static int try_to_grab_pending(struct work_struct *work)
 410 {
 411         struct cpu_workqueue_struct *cwq;
 412         int ret = -1;
 413
 414         if (!test_and_set_bit(WORK_STRUCT_PENDING, work_data_bits(work)))
 415                 return 0;
 416
 417         /*
 418          * The queueing is in progress, or it is already queued. Try to
 419          * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
 420          */
 421
 422         cwq = get_wq_data(work);
 423         if (!cwq)
 424                 return ret;
 425
 426         spin_lock_irq(&cwq->lock);
 427         if (!list_empty(&work->entry)) {
 428                 /*
 429                  * This work is queued, but perhaps we locked the wrong cwq.
 430                  * In that case we must see the new value after rmb(), see
 431                  * insert_work()->wmb().
 432                  */
 433                 smp_rmb();
 434                 if (cwq == get_wq_data(work)) {
 435                         list_del_init(&work->entry);
 436                         ret = 1;
 437                 }
 438         }
 439         spin_unlock_irq(&cwq->lock);
 440
 441         return ret;
 442 }
 443
 444 static void wait_on_cpu_work(struct cpu_workqueue_struct *cwq,
 445                                 struct work_struct *work)
 446 {
 447         struct wq_barrier barr;
 448         int running = 0;
 449
 450         spin_lock_irq(&cwq->lock);
 451         if (unlikely(cwq->current_work == work)) {
 452                 insert_wq_barrier(cwq, &barr, 0);
 453                 running = 1;
 454         }
 455         spin_unlock_irq(&cwq->lock);
 456
 457         if (unlikely(running))
 458                 wait_for_completion(&barr.done);
 459 }
 460
 461 static void wait_on_work(struct work_struct *work)
 462 {
 463         struct cpu_workqueue_struct *cwq;
 464         struct workqueue_struct *wq;
 465         const cpumask_t *cpu_map;
 466         int cpu;
 467
 468         might_sleep();
 469
 470         lock_acquire(&work->lockdep_map, 0, 0, 0, 2, _THIS_IP_);
 471         lock_release(&work->lockdep_map, 1, _THIS_IP_);
 472
 473         cwq = get_wq_data(work);
 474         if (!cwq)
 475                 return;
 476
 477         wq = cwq->wq;
 478         cpu_map = wq_cpu_map(wq);
 479
 480         for_each_cpu_mask(cpu, *cpu_map)
 481                 wait_on_cpu_work(per_cpu_ptr(wq->cpu_wq, cpu), work);
 482 }
 483
 484 static int __cancel_work_timer(struct work_struct *work,
 485                                 struct timer_list* timer)
 486 {
 487         int ret;
 488
 489         do {
 490                 ret = (timer && likely(del_timer(timer)));
 491                 if (!ret)
 492                         ret = try_to_grab_pending(work);
 493                 wait_on_work(work);
 494         } while (unlikely(ret < 0));
 495
 496         work_clear_pending(work);
 497         return ret;
 498 }
 499
 500 /**
 501  * cancel_work_sync - block until a work_struct's callback has terminated
 502  * @work: the work which is to be flushed
 503  *
 504  * Returns true if @work was pending.
 505  *
 506  * cancel_work_sync() will cancel the work if it is queued. If the work's
 507  * callback appears to be running, cancel_work_sync() will block until it
 508  * has completed.
 509  *
 510  * It is possible to use this function if the work re-queues itself. It can
 511  * cancel the work even if it migrates to another workqueue, however in that
 512  * case it only guarantees that work->func() has completed on the last queued
 513  * workqueue.
 514  *
 515  * cancel_work_sync(&delayed_work->work) should be used only if ->timer is not
 516  * pending, otherwise it goes into a busy-wait loop until the timer expires.
 517  *
 518  * The caller must ensure that workqueue_struct on which this work was last
 519  * queued can't be destroyed before this function returns.
 520  */
 521 int cancel_work_sync(struct work_struct *work)
 522 {
 523         return __cancel_work_timer(work, NULL);
 524 }
 525 EXPORT_SYMBOL_GPL(cancel_work_sync);
 526
 527 /**
 528  * cancel_delayed_work_sync - reliably kill off a delayed work.
 529  * @dwork: the delayed work struct
 530  *
 531  * Returns true if @dwork was pending.
 532  *
 533  * It is possible to use this function if @dwork rearms itself via queue_work()
 534  * or queue_delayed_work(). See also the comment for cancel_work_sync().
 535  */
 536 int cancel_delayed_work_sync(struct delayed_work *dwork)
 537 {
 538         return __cancel_work_timer(&dwork->work, &dwork->timer);
 539 }
 540 EXPORT_SYMBOL(cancel_delayed_work_sync);
 541
 542 static struct workqueue_struct *keventd_wq __read_mostly;
 543
 544 /**
 545  * schedule_work - put work task in global workqueue
 546  * @work: job to be done
 547  *
 548  * This puts a job in the kernel-global workqueue.
 549  */
 550 int schedule_work(struct work_struct *work)
 551 {
 552         return queue_work(keventd_wq, work);
 553 }
 554 EXPORT_SYMBOL(schedule_work);
 555
 556 /**
 557  * schedule_delayed_work - put work task in global workqueue after delay
 558  * @dwork: job to be done
 559  * @delay: number of jiffies to wait or 0 for immediate execution
 560  *
 561  * After waiting for a given time this puts a job in the kernel-global
 562  * workqueue.
 563  */
 564 int schedule_delayed_work(struct delayed_work *dwork,
 565                                         unsigned long delay)
 566 {
 567         return queue_delayed_work(keventd_wq, dwork, delay);
 568 }
 569 EXPORT_SYMBOL(schedule_delayed_work);
 570
 571 /**
 572  * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
 573  * @cpu: cpu to use
 574  * @dwork: job to be done
 575  * @delay: number of jiffies to wait
 576  *
 577  * After waiting for a given time this puts a job in the kernel-global
 578  * workqueue on the specified CPU.
 579  */
 580 int schedule_delayed_work_on(int cpu,
 581                         struct delayed_work *dwork, unsigned long delay)
 582 {
 583         return queue_delayed_work_on(cpu, keventd_wq, dwork, delay);
 584 }
 585 EXPORT_SYMBOL(schedule_delayed_work_on);
 586
 587 /**
 588  * schedule_on_each_cpu - call a function on each online CPU from keventd
 589  * @func: the function to call
 590  *
 591  * Returns zero on success.
 592  * Returns -ve errno on failure.
 593  *
 594  * schedule_on_each_cpu() is very slow.
 595  */
 596 int schedule_on_each_cpu(work_func_t func)
 597 {
 598         int cpu;
 599         struct work_struct *works;
 600
 601         works = alloc_percpu(struct work_struct);
 602         if (!works)
 603                 return -ENOMEM;
 604
 605         get_online_cpus();
 606         for_each_online_cpu(cpu) {
 607                 struct work_struct *work = per_cpu_ptr(works, cpu);
 608
 609                 INIT_WORK(work, func);
 610                 set_bit(WORK_STRUCT_PENDING, work_data_bits(work));
 611                 __queue_work(per_cpu_ptr(keventd_wq->cpu_wq, cpu), work);
 612         }
 613         flush_workqueue(keventd_wq);
 614         put_online_cpus();
 615         free_percpu(works);
 616         return 0;
 617 }
 618
 619 void flush_scheduled_work(void)
 620 {
 621         flush_workqueue(keventd_wq);
 622 }
 623 EXPORT_SYMBOL(flush_scheduled_work);
 624
 625 /**
 626  * execute_in_process_context - reliably execute the routine with user context
 627  * @fn:         the function to execute
 628  * @ew:         guaranteed storage for the execute work structure (must
 629  *              be available when the work executes)
 630  *
 631  * Executes the function immediately if process context is available,
 632  * otherwise schedules the function for delayed execution.
 633  *
 634  * Returns:     0 - function was executed
 635  *              1 - function was scheduled for execution
 636  */
 637 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
 638 {
 639         if (!in_interrupt()) {
 640                 fn(&ew->work);
 641                 return 0;
 642         }
 643
 644         INIT_WORK(&ew->work, fn);
 645         schedule_work(&ew->work);
 646
 647         return 1;
 648 }
 649 EXPORT_SYMBOL_GPL(execute_in_process_context);
 650
 651 int keventd_up(void)
 652 {
 653         return keventd_wq != NULL;
 654 }
 655
 656 int current_is_keventd(void)
 657 {
 658         struct cpu_workqueue_struct *cwq;
 659         int cpu = raw_smp_processor_id(); /* preempt-safe: keventd is per-cpu */
 660         int ret = 0;
 661
 662         BUG_ON(!keventd_wq);
 663
 664         cwq = per_cpu_ptr(keventd_wq->cpu_wq, cpu);
 665         if (current == cwq->thread)
 666                 ret = 1;
 667
 668         return ret;
 669
 670 }
 671
 672 static struct cpu_workqueue_struct *
 673 init_cpu_workqueue(struct workqueue_struct *wq, int cpu)
 674 {
 675         struct cpu_workqueue_struct *cwq = per_cpu_ptr(wq->cpu_wq, cpu);
 676
 677         cwq->wq = wq;
 678         spin_lock_init(&cwq->lock);
 679         INIT_LIST_HEAD(&cwq->worklist);
 680         init_waitqueue_head(&cwq->more_work);
 681
 682         return cwq;
 683 }
 684
 685 static int create_workqueue_thread(struct cpu_workqueue_struct *cwq, int cpu)
 686 {
 687         struct workqueue_struct *wq = cwq->wq;
 688         const char *fmt = is_single_threaded(wq) ? "%s" : "%s/%d";
 689         struct task_struct *p;
 690
 691         p = kthread_create(worker_thread, cwq, fmt, wq->name, cpu);
 692         /*
 693          * Nobody can add the work_struct to this cwq,
 694          *      if (caller is __create_workqueue)
 695          *              nobody should see this wq
 696          *      else // caller is CPU_UP_PREPARE
 697          *              cpu is not on cpu_online_map
 698          * so we can abort safely.
 699          */
 700         if (IS_ERR(p))
 701                 return PTR_ERR(p);
 702
 703         cwq->thread = p;
 704
 705         return 0;
 706 }
 707
 708 static void start_workqueue_thread(struct cpu_workqueue_struct *cwq, int cpu)
 709 {
 710         struct task_struct *p = cwq->thread;
 711
 712         if (p != NULL) {
 713                 if (cpu >= 0)
 714                         kthread_bind(p, cpu);
 715                 wake_up_process(p);
 716         }
 717 }
 718
 719 struct workqueue_struct *__create_workqueue_key(const char *name,
 720                                                 int singlethread,
 721                                                 int freezeable,
 722                                                 struct lock_class_key *key,
 723                                                 const char *lock_name)
 724 {
 725         struct workqueue_struct *wq;
 726         struct cpu_workqueue_struct *cwq;
 727         int err = 0, cpu;
 728
 729         wq = kzalloc(sizeof(*wq), GFP_KERNEL);
 730         if (!wq)
 731                 return NULL;
 732
 733         wq->cpu_wq = alloc_percpu(struct cpu_workqueue_struct);
 734         if (!wq->cpu_wq) {
 735                 kfree(wq);
 736                 return NULL;
 737         }
 738
 739         wq->name = name;
 740         lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
 741         wq->singlethread = singlethread;
 742         wq->freezeable = freezeable;
 743         INIT_LIST_HEAD(&wq->list);
 744
 745         if (singlethread) {
 746                 cwq = init_cpu_workqueue(wq, singlethread_cpu);
 747                 err = create_workqueue_thread(cwq, singlethread_cpu);
 748                 start_workqueue_thread(cwq, -1);
 749         } else {
 750                 get_online_cpus();
 751                 spin_lock(&workqueue_lock);
 752                 list_add(&wq->list, &workqueues);
 753                 spin_unlock(&workqueue_lock);
 754
 755                 for_each_possible_cpu(cpu) {
 756                         cwq = init_cpu_workqueue(wq, cpu);
 757                         if (err || !cpu_online(cpu))
 758                                 continue;
 759                         err = create_workqueue_thread(cwq, cpu);
 760                         start_workqueue_thread(cwq, cpu);
 761                 }
 762                 put_online_cpus();
 763         }
 764
 765         if (err) {
 766                 destroy_workqueue(wq);
 767                 wq = NULL;
 768         }
 769         return wq;
 770 }
 771 EXPORT_SYMBOL_GPL(__create_workqueue_key);
 772
 773 static void cleanup_workqueue_thread(struct cpu_workqueue_struct *cwq)
 774 {
 775         /*
 776          * Our caller is either destroy_workqueue() or CPU_DEAD,
 777          * get_online_cpus() protects cwq->thread.
 778          */
 779         if (cwq->thread == NULL)
 780                 return;
 781
 782         lock_acquire(&cwq->wq->lockdep_map, 0, 0, 0, 2, _THIS_IP_);
 783         lock_release(&cwq->wq->lockdep_map, 1, _THIS_IP_);
 784
 785         flush_cpu_workqueue(cwq);
 786         /*
 787          * If the caller is CPU_DEAD and cwq->worklist was not empty,
 788          * a concurrent flush_workqueue() can insert a barrier after us.
 789          * However, in that case run_workqueue() won't return and check
 790          * kthread_should_stop() until it flushes all work_struct's.
 791          * When ->worklist becomes empty it is safe to exit because no
 792          * more work_structs can be queued on this cwq: flush_workqueue
 793          * checks list_empty(), and a "normal" queue_work() can't use
 794          * a dead CPU.
 795          */
 796         kthread_stop(cwq->thread);
 797         cwq->thread = NULL;
 798 }
 799
 800 /**
 801  * destroy_workqueue - safely terminate a workqueue
 802  * @wq: target workqueue
 803  *
 804  * Safely destroy a workqueue. All work currently pending will be done first.
 805  */
 806 void destroy_workqueue(struct workqueue_struct *wq)
 807 {
 808         const cpumask_t *cpu_map = wq_cpu_map(wq);
 809         int cpu;
 810
 811         get_online_cpus();
 812         spin_lock(&workqueue_lock);
 813         list_del(&wq->list);
 814         spin_unlock(&workqueue_lock);
 815
 816         for_each_cpu_mask(cpu, *cpu_map)
 817                 cleanup_workqueue_thread(per_cpu_ptr(wq->cpu_wq, cpu));
 818         put_online_cpus();
 819
 820         free_percpu(wq->cpu_wq);
 821         kfree(wq);
 822 }
 823 EXPORT_SYMBOL_GPL(destroy_workqueue);
 824
 825 static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
 826                                                 unsigned long action,
 827                                                 void *hcpu)
 828 {
 829         unsigned int cpu = (unsigned long)hcpu;
 830         struct cpu_workqueue_struct *cwq;
 831         struct workqueue_struct *wq;
 832
 833         action &= ~CPU_TASKS_FROZEN;
 834
 835         switch (action) {
 836         case CPU_UP_PREPARE:
 837                 cpu_set(cpu, cpu_populated_map);
 838         }
 839
 840         list_for_each_entry(wq, &workqueues, list) {
 841                 cwq = per_cpu_ptr(wq->cpu_wq, cpu);
 842
 843                 switch (action) {
 844                 case CPU_UP_PREPARE:
 845                         if (!create_workqueue_thread(cwq, cpu))
 846                                 break;
 847                         printk(KERN_ERR "workqueue [%s] for %i failed\n",
 848                                 wq->name, cpu);
 849                         return NOTIFY_BAD;
 850
 851                 case CPU_ONLINE:
 852                         start_workqueue_thread(cwq, cpu);
 853                         break;
 854
 855                 case CPU_UP_CANCELED:
 856                         start_workqueue_thread(cwq, -1);
 857                 case CPU_DEAD:
 858                         cleanup_workqueue_thread(cwq);
 859                         break;
 860                 }
 861         }
 862
 863         switch (action) {
 864         case CPU_UP_CANCELED:
 865         case CPU_DEAD:
 866                 cpu_clear(cpu, cpu_populated_map);
 867         }
 868
 869         return NOTIFY_OK;
 870 }
 871
 872 void __init init_workqueues(void)
 873 {
 874         cpu_populated_map = cpu_online_map;
 875         singlethread_cpu = first_cpu(cpu_possible_map);
 876         cpu_singlethread_map = cpumask_of_cpu(singlethread_cpu);
 877         hotcpu_notifier(workqueue_cpu_callback, 0);
 878         keventd_wq = create_workqueue("events");
 879         BUG_ON(!keventd_wq);
 880 }