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
  36 /*
  37  * The per-CPU workqueue (if single thread, we always use the first
  38  * possible cpu).
  39  */
  40 struct cpu_workqueue_struct {
  41
  42         spinlock_t lock;
  43
  44         struct list_head worklist;
  45         wait_queue_head_t more_work;
  46         struct work_struct *current_work;
  47
  48         struct workqueue_struct *wq;
  49         struct task_struct *thread;
  50         int should_stop;
  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 };
  66
  67 /* All the per-cpu workqueues on the system, for hotplug cpu to add/remove
  68    threads to each one as cpus come/go. */
  69 static DEFINE_MUTEX(workqueue_mutex);
  70 static LIST_HEAD(workqueues);
  71
  72 static int singlethread_cpu __read_mostly;
  73 static cpumask_t cpu_singlethread_map __read_mostly;
  74 /* optimization, we could use cpu_possible_map */
  75 static cpumask_t cpu_populated_map __read_mostly;
  76
  77 /* If it's single threaded, it isn't in the list of workqueues. */
  78 static inline int is_single_threaded(struct workqueue_struct *wq)
  79 {
  80         return wq->singlethread;
  81 }
  82
  83 static const cpumask_t *wq_cpu_map(struct workqueue_struct *wq)
  84 {
  85         return is_single_threaded(wq)
  86                 ? &cpu_singlethread_map : &cpu_populated_map;
  87 }
  88
  89 static
  90 struct cpu_workqueue_struct *wq_per_cpu(struct workqueue_struct *wq, int cpu)
  91 {
  92         if (unlikely(is_single_threaded(wq)))
  93                 cpu = singlethread_cpu;
  94         return per_cpu_ptr(wq->cpu_wq, cpu);
  95 }
  96
  97 /*
  98  * Set the workqueue on which a work item is to be run
  99  * - Must *only* be called if the pending flag is set
 100  */
 101 static inline void set_wq_data(struct work_struct *work,
 102                                 struct cpu_workqueue_struct *cwq)
 103 {
 104         unsigned long new;
 105
 106         BUG_ON(!work_pending(work));
 107
 108         new = (unsigned long) cwq | (1UL << WORK_STRUCT_PENDING);
 109         new |= WORK_STRUCT_FLAG_MASK & *work_data_bits(work);
 110         atomic_long_set(&work->data, new);
 111 }
 112
 113 static inline
 114 struct cpu_workqueue_struct *get_wq_data(struct work_struct *work)
 115 {
 116         return (void *) (atomic_long_read(&work->data) & WORK_STRUCT_WQ_DATA_MASK);
 117 }
 118
 119 static void insert_work(struct cpu_workqueue_struct *cwq,
 120                                 struct work_struct *work, int tail)
 121 {
 122         set_wq_data(work, cwq);
 123         if (tail)
 124                 list_add_tail(&work->entry, &cwq->worklist);
 125         else
 126                 list_add(&work->entry, &cwq->worklist);
 127         wake_up(&cwq->more_work);
 128 }
 129
 130 /* Preempt must be disabled. */
 131 static void __queue_work(struct cpu_workqueue_struct *cwq,
 132                          struct work_struct *work)
 133 {
 134         unsigned long flags;
 135
 136         spin_lock_irqsave(&cwq->lock, flags);
 137         insert_work(cwq, work, 1);
 138         spin_unlock_irqrestore(&cwq->lock, flags);
 139 }
 140
 141 /**
 142  * queue_work - queue work on a workqueue
 143  * @wq: workqueue to use
 144  * @work: work to queue
 145  *
 146  * Returns 0 if @work was already on a queue, non-zero otherwise.
 147  *
 148  * We queue the work to the CPU it was submitted, but there is no
 149  * guarantee that it will be processed by that CPU.
 150  */
 151 int fastcall queue_work(struct workqueue_struct *wq, struct work_struct *work)
 152 {
 153         int ret = 0;
 154
 155         if (!test_and_set_bit(WORK_STRUCT_PENDING, work_data_bits(work))) {
 156                 BUG_ON(!list_empty(&work->entry));
 157                 __queue_work(wq_per_cpu(wq, get_cpu()), work);
 158                 put_cpu();
 159                 ret = 1;
 160         }
 161         return ret;
 162 }
 163 EXPORT_SYMBOL_GPL(queue_work);
 164
 165 void delayed_work_timer_fn(unsigned long __data)
 166 {
 167         struct delayed_work *dwork = (struct delayed_work *)__data;
 168         struct cpu_workqueue_struct *cwq = get_wq_data(&dwork->work);
 169         struct workqueue_struct *wq = cwq->wq;
 170
 171         __queue_work(wq_per_cpu(wq, smp_processor_id()), &dwork->work);
 172 }
 173
 174 /**
 175  * queue_delayed_work - queue work on a workqueue after delay
 176  * @wq: workqueue to use
 177  * @dwork: delayable work to queue
 178  * @delay: number of jiffies to wait before queueing
 179  *
 180  * Returns 0 if @work was already on a queue, non-zero otherwise.
 181  */
 182 int fastcall queue_delayed_work(struct workqueue_struct *wq,
 183                         struct delayed_work *dwork, unsigned long delay)
 184 {
 185         timer_stats_timer_set_start_info(&dwork->timer);
 186         if (delay == 0)
 187                 return queue_work(wq, &dwork->work);
 188
 189         return queue_delayed_work_on(-1, wq, dwork, delay);
 190 }
 191 EXPORT_SYMBOL_GPL(queue_delayed_work);
 192
 193 /**
 194  * queue_delayed_work_on - queue work on specific CPU after delay
 195  * @cpu: CPU number to execute work on
 196  * @wq: workqueue to use
 197  * @dwork: work to queue
 198  * @delay: number of jiffies to wait before queueing
 199  *
 200  * Returns 0 if @work was already on a queue, non-zero otherwise.
 201  */
 202 int queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
 203                         struct delayed_work *dwork, unsigned long delay)
 204 {
 205         int ret = 0;
 206         struct timer_list *timer = &dwork->timer;
 207         struct work_struct *work = &dwork->work;
 208
 209         if (!test_and_set_bit(WORK_STRUCT_PENDING, work_data_bits(work))) {
 210                 BUG_ON(timer_pending(timer));
 211                 BUG_ON(!list_empty(&work->entry));
 212
 213                 /* This stores cwq for the moment, for the timer_fn */
 214                 set_wq_data(work, wq_per_cpu(wq, raw_smp_processor_id()));
 215                 timer->expires = jiffies + delay;
 216                 timer->data = (unsigned long)dwork;
 217                 timer->function = delayed_work_timer_fn;
 218
 219                 if (unlikely(cpu >= 0))
 220                         add_timer_on(timer, cpu);
 221                 else
 222                         add_timer(timer);
 223                 ret = 1;
 224         }
 225         return ret;
 226 }
 227 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
 228
 229 static void run_workqueue(struct cpu_workqueue_struct *cwq)
 230 {
 231         spin_lock_irq(&cwq->lock);
 232         cwq->run_depth++;
 233         if (cwq->run_depth > 3) {
 234                 /* morton gets to eat his hat */
 235                 printk("%s: recursion depth exceeded: %d\n",
 236                         __FUNCTION__, cwq->run_depth);
 237                 dump_stack();
 238         }
 239         while (!list_empty(&cwq->worklist)) {
 240                 struct work_struct *work = list_entry(cwq->worklist.next,
 241                                                 struct work_struct, entry);
 242                 work_func_t f = work->func;
 243
 244                 cwq->current_work = work;
 245                 list_del_init(cwq->worklist.next);
 246                 spin_unlock_irq(&cwq->lock);
 247
 248                 BUG_ON(get_wq_data(work) != cwq);
 249                 if (!test_bit(WORK_STRUCT_NOAUTOREL, work_data_bits(work)))
 250                         work_release(work);
 251                 f(work);
 252
 253                 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
 254                         printk(KERN_ERR "BUG: workqueue leaked lock or atomic: "
 255                                         "%s/0x%08x/%d\n",
 256                                         current->comm, preempt_count(),
 257                                         current->pid);
 258                         printk(KERN_ERR "    last function: ");
 259                         print_symbol("%s\n", (unsigned long)f);
 260                         debug_show_held_locks(current);
 261                         dump_stack();
 262                 }
 263
 264                 spin_lock_irq(&cwq->lock);
 265                 cwq->current_work = NULL;
 266         }
 267         cwq->run_depth--;
 268         spin_unlock_irq(&cwq->lock);
 269 }
 270
 271 /*
 272  * NOTE: the caller must not touch *cwq if this func returns true
 273  */
 274 static int cwq_should_stop(struct cpu_workqueue_struct *cwq)
 275 {
 276         int should_stop = cwq->should_stop;
 277
 278         if (unlikely(should_stop)) {
 279                 spin_lock_irq(&cwq->lock);
 280                 should_stop = cwq->should_stop && list_empty(&cwq->worklist);
 281                 if (should_stop)
 282                         cwq->thread = NULL;
 283                 spin_unlock_irq(&cwq->lock);
 284         }
 285
 286         return should_stop;
 287 }
 288
 289 static int worker_thread(void *__cwq)
 290 {
 291         struct cpu_workqueue_struct *cwq = __cwq;
 292         DEFINE_WAIT(wait);
 293         struct k_sigaction sa;
 294         sigset_t blocked;
 295
 296         if (!cwq->wq->freezeable)
 297                 current->flags |= PF_NOFREEZE;
 298
 299         set_user_nice(current, -5);
 300
 301         /* Block and flush all signals */
 302         sigfillset(&blocked);
 303         sigprocmask(SIG_BLOCK, &blocked, NULL);
 304         flush_signals(current);
 305
 306         /*
 307          * We inherited MPOL_INTERLEAVE from the booting kernel.
 308          * Set MPOL_DEFAULT to insure node local allocations.
 309          */
 310         numa_default_policy();
 311
 312         /* SIG_IGN makes children autoreap: see do_notify_parent(). */
 313         sa.sa.sa_handler = SIG_IGN;
 314         sa.sa.sa_flags = 0;
 315         siginitset(&sa.sa.sa_mask, sigmask(SIGCHLD));
 316         do_sigaction(SIGCHLD, &sa, (struct k_sigaction *)0);
 317
 318         for (;;) {
 319                 if (cwq->wq->freezeable)
 320                         try_to_freeze();
 321
 322                 prepare_to_wait(&cwq->more_work, &wait, TASK_INTERRUPTIBLE);
 323                 if (!cwq->should_stop && list_empty(&cwq->worklist))
 324                         schedule();
 325                 finish_wait(&cwq->more_work, &wait);
 326
 327                 if (cwq_should_stop(cwq))
 328                         break;
 329
 330                 run_workqueue(cwq);
 331         }
 332
 333         return 0;
 334 }
 335
 336 struct wq_barrier {
 337         struct work_struct      work;
 338         struct completion       done;
 339 };
 340
 341 static void wq_barrier_func(struct work_struct *work)
 342 {
 343         struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
 344         complete(&barr->done);
 345 }
 346
 347 static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
 348                                         struct wq_barrier *barr, int tail)
 349 {
 350         INIT_WORK(&barr->work, wq_barrier_func);
 351         __set_bit(WORK_STRUCT_PENDING, work_data_bits(&barr->work));
 352
 353         init_completion(&barr->done);
 354
 355         insert_work(cwq, &barr->work, tail);
 356 }
 357
 358 static void flush_cpu_workqueue(struct cpu_workqueue_struct *cwq)
 359 {
 360         if (cwq->thread == current) {
 361                 /*
 362                  * Probably keventd trying to flush its own queue. So simply run
 363                  * it by hand rather than deadlocking.
 364                  */
 365                 run_workqueue(cwq);
 366         } else {
 367                 struct wq_barrier barr;
 368                 int active = 0;
 369
 370                 spin_lock_irq(&cwq->lock);
 371                 if (!list_empty(&cwq->worklist) || cwq->current_work != NULL) {
 372                         insert_wq_barrier(cwq, &barr, 1);
 373                         active = 1;
 374                 }
 375                 spin_unlock_irq(&cwq->lock);
 376
 377                 if (active)
 378                         wait_for_completion(&barr.done);
 379         }
 380 }
 381
 382 /**
 383  * flush_workqueue - ensure that any scheduled work has run to completion.
 384  * @wq: workqueue to flush
 385  *
 386  * Forces execution of the workqueue and blocks until its completion.
 387  * This is typically used in driver shutdown handlers.
 388  *
 389  * We sleep until all works which were queued on entry have been handled,
 390  * but we are not livelocked by new incoming ones.
 391  *
 392  * This function used to run the workqueues itself.  Now we just wait for the
 393  * helper threads to do it.
 394  */
 395 void fastcall flush_workqueue(struct workqueue_struct *wq)
 396 {
 397         const cpumask_t *cpu_map = wq_cpu_map(wq);
 398         int cpu;
 399
 400         might_sleep();
 401         for_each_cpu_mask(cpu, *cpu_map)
 402                 flush_cpu_workqueue(per_cpu_ptr(wq->cpu_wq, cpu));
 403 }
 404 EXPORT_SYMBOL_GPL(flush_workqueue);
 405
 406 static void wait_on_work(struct cpu_workqueue_struct *cwq,
 407                                 struct work_struct *work)
 408 {
 409         struct wq_barrier barr;
 410         int running = 0;
 411
 412         spin_lock_irq(&cwq->lock);
 413         if (unlikely(cwq->current_work == work)) {
 414                 insert_wq_barrier(cwq, &barr, 0);
 415                 running = 1;
 416         }
 417         spin_unlock_irq(&cwq->lock);
 418
 419         if (unlikely(running))
 420                 wait_for_completion(&barr.done);
 421 }
 422
 423 /**
 424  * flush_work - block until a work_struct's callback has terminated
 425  * @wq: the workqueue on which the work is queued
 426  * @work: the work which is to be flushed
 427  *
 428  * flush_work() will attempt to cancel the work if it is queued.  If the work's
 429  * callback appears to be running, flush_work() will block until it has
 430  * completed.
 431  *
 432  * flush_work() is designed to be used when the caller is tearing down data
 433  * structures which the callback function operates upon.  It is expected that,
 434  * prior to calling flush_work(), the caller has arranged for the work to not
 435  * be requeued.
 436  */
 437 void flush_work(struct workqueue_struct *wq, struct work_struct *work)
 438 {
 439         const cpumask_t *cpu_map = wq_cpu_map(wq);
 440         struct cpu_workqueue_struct *cwq;
 441         int cpu;
 442
 443         might_sleep();
 444
 445         cwq = get_wq_data(work);
 446         /* Was it ever queued ? */
 447         if (!cwq)
 448                 return;
 449
 450         /*
 451          * This work can't be re-queued, no need to re-check that
 452          * get_wq_data() is still the same when we take cwq->lock.
 453          */
 454         spin_lock_irq(&cwq->lock);
 455         list_del_init(&work->entry);
 456         work_release(work);
 457         spin_unlock_irq(&cwq->lock);
 458
 459         for_each_cpu_mask(cpu, *cpu_map)
 460                 wait_on_work(per_cpu_ptr(wq->cpu_wq, cpu), work);
 461 }
 462 EXPORT_SYMBOL_GPL(flush_work);
 463
 464
 465 static struct workqueue_struct *keventd_wq;
 466
 467 /**
 468  * schedule_work - put work task in global workqueue
 469  * @work: job to be done
 470  *
 471  * This puts a job in the kernel-global workqueue.
 472  */
 473 int fastcall schedule_work(struct work_struct *work)
 474 {
 475         return queue_work(keventd_wq, work);
 476 }
 477 EXPORT_SYMBOL(schedule_work);
 478
 479 /**
 480  * schedule_delayed_work - put work task in global workqueue after delay
 481  * @dwork: job to be done
 482  * @delay: number of jiffies to wait or 0 for immediate execution
 483  *
 484  * After waiting for a given time this puts a job in the kernel-global
 485  * workqueue.
 486  */
 487 int fastcall schedule_delayed_work(struct delayed_work *dwork,
 488                                         unsigned long delay)
 489 {
 490         timer_stats_timer_set_start_info(&dwork->timer);
 491         return queue_delayed_work(keventd_wq, dwork, delay);
 492 }
 493 EXPORT_SYMBOL(schedule_delayed_work);
 494
 495 /**
 496  * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
 497  * @cpu: cpu to use
 498  * @dwork: job to be done
 499  * @delay: number of jiffies to wait
 500  *
 501  * After waiting for a given time this puts a job in the kernel-global
 502  * workqueue on the specified CPU.
 503  */
 504 int schedule_delayed_work_on(int cpu,
 505                         struct delayed_work *dwork, unsigned long delay)
 506 {
 507         return queue_delayed_work_on(cpu, keventd_wq, dwork, delay);
 508 }
 509 EXPORT_SYMBOL(schedule_delayed_work_on);
 510
 511 /**
 512  * schedule_on_each_cpu - call a function on each online CPU from keventd
 513  * @func: the function to call
 514  *
 515  * Returns zero on success.
 516  * Returns -ve errno on failure.
 517  *
 518  * Appears to be racy against CPU hotplug.
 519  *
 520  * schedule_on_each_cpu() is very slow.
 521  */
 522 int schedule_on_each_cpu(work_func_t func)
 523 {
 524         int cpu;
 525         struct work_struct *works;
 526
 527         works = alloc_percpu(struct work_struct);
 528         if (!works)
 529                 return -ENOMEM;
 530
 531         preempt_disable();              /* CPU hotplug */
 532         for_each_online_cpu(cpu) {
 533                 struct work_struct *work = per_cpu_ptr(works, cpu);
 534
 535                 INIT_WORK(work, func);
 536                 set_bit(WORK_STRUCT_PENDING, work_data_bits(work));
 537                 __queue_work(per_cpu_ptr(keventd_wq->cpu_wq, cpu), work);
 538         }
 539         preempt_enable();
 540         flush_workqueue(keventd_wq);
 541         free_percpu(works);
 542         return 0;
 543 }
 544
 545 void flush_scheduled_work(void)
 546 {
 547         flush_workqueue(keventd_wq);
 548 }
 549 EXPORT_SYMBOL(flush_scheduled_work);
 550
 551 void flush_work_keventd(struct work_struct *work)
 552 {
 553         flush_work(keventd_wq, work);
 554 }
 555 EXPORT_SYMBOL(flush_work_keventd);
 556
 557 /**
 558  * cancel_rearming_delayed_work - kill off a delayed work whose handler rearms the delayed work.
 559  * @dwork: the delayed work struct
 560  *
 561  * Note that the work callback function may still be running on return from
 562  * cancel_delayed_work(). Run flush_workqueue() or flush_work() to wait on it.
 563  */
 564 void cancel_rearming_delayed_work(struct delayed_work *dwork)
 565 {
 566         struct cpu_workqueue_struct *cwq = get_wq_data(&dwork->work);
 567
 568         /* Was it ever queued ? */
 569         if (cwq != NULL) {
 570                 struct workqueue_struct *wq = cwq->wq;
 571
 572                 while (!cancel_delayed_work(dwork))
 573                         flush_workqueue(wq);
 574         }
 575 }
 576 EXPORT_SYMBOL(cancel_rearming_delayed_work);
 577
 578 /**
 579  * execute_in_process_context - reliably execute the routine with user context
 580  * @fn:         the function to execute
 581  * @ew:         guaranteed storage for the execute work structure (must
 582  *              be available when the work executes)
 583  *
 584  * Executes the function immediately if process context is available,
 585  * otherwise schedules the function for delayed execution.
 586  *
 587  * Returns:     0 - function was executed
 588  *              1 - function was scheduled for execution
 589  */
 590 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
 591 {
 592         if (!in_interrupt()) {
 593                 fn(&ew->work);
 594                 return 0;
 595         }
 596
 597         INIT_WORK(&ew->work, fn);
 598         schedule_work(&ew->work);
 599
 600         return 1;
 601 }
 602 EXPORT_SYMBOL_GPL(execute_in_process_context);
 603
 604 int keventd_up(void)
 605 {
 606         return keventd_wq != NULL;
 607 }
 608
 609 int current_is_keventd(void)
 610 {
 611         struct cpu_workqueue_struct *cwq;
 612         int cpu = smp_processor_id();   /* preempt-safe: keventd is per-cpu */
 613         int ret = 0;
 614
 615         BUG_ON(!keventd_wq);
 616
 617         cwq = per_cpu_ptr(keventd_wq->cpu_wq, cpu);
 618         if (current == cwq->thread)
 619                 ret = 1;
 620
 621         return ret;
 622
 623 }
 624
 625 static struct cpu_workqueue_struct *
 626 init_cpu_workqueue(struct workqueue_struct *wq, int cpu)
 627 {
 628         struct cpu_workqueue_struct *cwq = per_cpu_ptr(wq->cpu_wq, cpu);
 629
 630         cwq->wq = wq;
 631         spin_lock_init(&cwq->lock);
 632         INIT_LIST_HEAD(&cwq->worklist);
 633         init_waitqueue_head(&cwq->more_work);
 634
 635         return cwq;
 636 }
 637
 638 static int create_workqueue_thread(struct cpu_workqueue_struct *cwq, int cpu)
 639 {
 640         struct workqueue_struct *wq = cwq->wq;
 641         const char *fmt = is_single_threaded(wq) ? "%s" : "%s/%d";
 642         struct task_struct *p;
 643
 644         p = kthread_create(worker_thread, cwq, fmt, wq->name, cpu);
 645         /*
 646          * Nobody can add the work_struct to this cwq,
 647          *      if (caller is __create_workqueue)
 648          *              nobody should see this wq
 649          *      else // caller is CPU_UP_PREPARE
 650          *              cpu is not on cpu_online_map
 651          * so we can abort safely.
 652          */
 653         if (IS_ERR(p))
 654                 return PTR_ERR(p);
 655
 656         cwq->thread = p;
 657         cwq->should_stop = 0;
 658
 659         return 0;
 660 }
 661
 662 static void start_workqueue_thread(struct cpu_workqueue_struct *cwq, int cpu)
 663 {
 664         struct task_struct *p = cwq->thread;
 665
 666         if (p != NULL) {
 667                 if (cpu >= 0)
 668                         kthread_bind(p, cpu);
 669                 wake_up_process(p);
 670         }
 671 }
 672
 673 struct workqueue_struct *__create_workqueue(const char *name,
 674                                             int singlethread, int freezeable)
 675 {
 676         struct workqueue_struct *wq;
 677         struct cpu_workqueue_struct *cwq;
 678         int err = 0, cpu;
 679
 680         wq = kzalloc(sizeof(*wq), GFP_KERNEL);
 681         if (!wq)
 682                 return NULL;
 683
 684         wq->cpu_wq = alloc_percpu(struct cpu_workqueue_struct);
 685         if (!wq->cpu_wq) {
 686                 kfree(wq);
 687                 return NULL;
 688         }
 689
 690         wq->name = name;
 691         wq->singlethread = singlethread;
 692         wq->freezeable = freezeable;
 693         INIT_LIST_HEAD(&wq->list);
 694
 695         if (singlethread) {
 696                 cwq = init_cpu_workqueue(wq, singlethread_cpu);
 697                 err = create_workqueue_thread(cwq, singlethread_cpu);
 698                 start_workqueue_thread(cwq, -1);
 699         } else {
 700                 mutex_lock(&workqueue_mutex);
 701                 list_add(&wq->list, &workqueues);
 702
 703                 for_each_possible_cpu(cpu) {
 704                         cwq = init_cpu_workqueue(wq, cpu);
 705                         if (err || !cpu_online(cpu))
 706                                 continue;
 707                         err = create_workqueue_thread(cwq, cpu);
 708                         start_workqueue_thread(cwq, cpu);
 709                 }
 710                 mutex_unlock(&workqueue_mutex);
 711         }
 712
 713         if (err) {
 714                 destroy_workqueue(wq);
 715                 wq = NULL;
 716         }
 717         return wq;
 718 }
 719 EXPORT_SYMBOL_GPL(__create_workqueue);
 720
 721 static void cleanup_workqueue_thread(struct cpu_workqueue_struct *cwq, int cpu)
 722 {
 723         struct wq_barrier barr;
 724         int alive = 0;
 725
 726         spin_lock_irq(&cwq->lock);
 727         if (cwq->thread != NULL) {
 728                 insert_wq_barrier(cwq, &barr, 1);
 729                 cwq->should_stop = 1;
 730                 alive = 1;
 731         }
 732         spin_unlock_irq(&cwq->lock);
 733
 734         if (alive) {
 735                 wait_for_completion(&barr.done);
 736
 737                 while (unlikely(cwq->thread != NULL))
 738                         cpu_relax();
 739                 /*
 740                  * Wait until cwq->thread unlocks cwq->lock,
 741                  * it won't touch *cwq after that.
 742                  */
 743                 smp_rmb();
 744                 spin_unlock_wait(&cwq->lock);
 745         }
 746 }
 747
 748 /**
 749  * destroy_workqueue - safely terminate a workqueue
 750  * @wq: target workqueue
 751  *
 752  * Safely destroy a workqueue. All work currently pending will be done first.
 753  */
 754 void destroy_workqueue(struct workqueue_struct *wq)
 755 {
 756         const cpumask_t *cpu_map = wq_cpu_map(wq);
 757         struct cpu_workqueue_struct *cwq;
 758         int cpu;
 759
 760         mutex_lock(&workqueue_mutex);
 761         list_del(&wq->list);
 762         mutex_unlock(&workqueue_mutex);
 763
 764         for_each_cpu_mask(cpu, *cpu_map) {
 765                 cwq = per_cpu_ptr(wq->cpu_wq, cpu);
 766                 cleanup_workqueue_thread(cwq, cpu);
 767         }
 768
 769         free_percpu(wq->cpu_wq);
 770         kfree(wq);
 771 }
 772 EXPORT_SYMBOL_GPL(destroy_workqueue);
 773
 774 static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
 775                                                 unsigned long action,
 776                                                 void *hcpu)
 777 {
 778         unsigned int cpu = (unsigned long)hcpu;
 779         struct cpu_workqueue_struct *cwq;
 780         struct workqueue_struct *wq;
 781
 782         switch (action) {
 783         case CPU_LOCK_ACQUIRE:
 784                 mutex_lock(&workqueue_mutex);
 785                 return NOTIFY_OK;
 786
 787         case CPU_LOCK_RELEASE:
 788                 mutex_unlock(&workqueue_mutex);
 789                 return NOTIFY_OK;
 790
 791         case CPU_UP_PREPARE:
 792                 cpu_set(cpu, cpu_populated_map);
 793         }
 794
 795         list_for_each_entry(wq, &workqueues, list) {
 796                 cwq = per_cpu_ptr(wq->cpu_wq, cpu);
 797
 798                 switch (action) {
 799                 case CPU_UP_PREPARE:
 800                         if (!create_workqueue_thread(cwq, cpu))
 801                                 break;
 802                         printk(KERN_ERR "workqueue for %i failed\n", cpu);
 803                         return NOTIFY_BAD;
 804
 805                 case CPU_ONLINE:
 806                         start_workqueue_thread(cwq, cpu);
 807                         break;
 808
 809                 case CPU_UP_CANCELED:
 810                         start_workqueue_thread(cwq, -1);
 811                 case CPU_DEAD:
 812                         cleanup_workqueue_thread(cwq, cpu);
 813                         break;
 814                 }
 815         }
 816
 817         return NOTIFY_OK;
 818 }
 819
 820 void __init init_workqueues(void)
 821 {
 822         cpu_populated_map = cpu_online_map;
 823         singlethread_cpu = first_cpu(cpu_possible_map);
 824         cpu_singlethread_map = cpumask_of_cpu(singlethread_cpu);
 825         hotcpu_notifier(workqueue_cpu_callback, 0);
 826         keventd_wq = create_workqueue("events");
 827         BUG_ON(!keventd_wq);
 828 }