SAFE public projects git trees. - safe/jmp/linux-2.6/blob - 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                 work_clear_pending(work);
 250                 f(work);
 251
 252                 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
 253                         printk(KERN_ERR "BUG: workqueue leaked lock or atomic: "
 254                                         "%s/0x%08x/%d\n",
 255                                         current->comm, preempt_count(),
 256                                         current->pid);
 257                         printk(KERN_ERR "    last function: ");
 258                         print_symbol("%s\n", (unsigned long)f);
 259                         debug_show_held_locks(current);
 260                         dump_stack();
 261                 }
 262
 263                 spin_lock_irq(&cwq->lock);
 264                 cwq->current_work = NULL;
 265         }
 266         cwq->run_depth--;
 267         spin_unlock_irq(&cwq->lock);
 268 }
 269
 270 /*
 271  * NOTE: the caller must not touch *cwq if this func returns true
 272  */
 273 static int cwq_should_stop(struct cpu_workqueue_struct *cwq)
 274 {
 275         int should_stop = cwq->should_stop;
 276
 277         if (unlikely(should_stop)) {
 278                 spin_lock_irq(&cwq->lock);
 279                 should_stop = cwq->should_stop && list_empty(&cwq->worklist);
 280                 if (should_stop)
 281                         cwq->thread = NULL;
 282                 spin_unlock_irq(&cwq->lock);
 283         }
 284
 285         return should_stop;
 286 }
 287
 288 static int worker_thread(void *__cwq)
 289 {
 290         struct cpu_workqueue_struct *cwq = __cwq;
 291         DEFINE_WAIT(wait);
 292         struct k_sigaction sa;
 293         sigset_t blocked;
 294
 295         if (!cwq->wq->freezeable)
 296                 current->flags |= PF_NOFREEZE;
 297
 298         set_user_nice(current, -5);
 299
 300         /* Block and flush all signals */
 301         sigfillset(&blocked);
 302         sigprocmask(SIG_BLOCK, &blocked, NULL);
 303         flush_signals(current);
 304
 305         /*
 306          * We inherited MPOL_INTERLEAVE from the booting kernel.
 307          * Set MPOL_DEFAULT to insure node local allocations.
 308          */
 309         numa_default_policy();
 310
 311         /* SIG_IGN makes children autoreap: see do_notify_parent(). */
 312         sa.sa.sa_handler = SIG_IGN;
 313         sa.sa.sa_flags = 0;
 314         siginitset(&sa.sa.sa_mask, sigmask(SIGCHLD));
 315         do_sigaction(SIGCHLD, &sa, (struct k_sigaction *)0);
 316
 317         for (;;) {
 318                 if (cwq->wq->freezeable)
 319                         try_to_freeze();
 320
 321                 prepare_to_wait(&cwq->more_work, &wait, TASK_INTERRUPTIBLE);
 322                 if (!cwq->should_stop && list_empty(&cwq->worklist))
 323                         schedule();
 324                 finish_wait(&cwq->more_work, &wait);
 325
 326                 if (cwq_should_stop(cwq))
 327                         break;
 328
 329                 run_workqueue(cwq);
 330         }
 331
 332         return 0;
 333 }
 334
 335 struct wq_barrier {
 336         struct work_struct      work;
 337         struct completion       done;
 338 };
 339
 340 static void wq_barrier_func(struct work_struct *work)
 341 {
 342         struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
 343         complete(&barr->done);
 344 }
 345
 346 static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
 347                                         struct wq_barrier *barr, int tail)
 348 {
 349         INIT_WORK(&barr->work, wq_barrier_func);
 350         __set_bit(WORK_STRUCT_PENDING, work_data_bits(&barr->work));
 351
 352         init_completion(&barr->done);
 353
 354         insert_work(cwq, &barr->work, tail);
 355 }
 356
 357 static void flush_cpu_workqueue(struct cpu_workqueue_struct *cwq)
 358 {
 359         if (cwq->thread == current) {
 360                 /*
 361                  * Probably keventd trying to flush its own queue. So simply run
 362                  * it by hand rather than deadlocking.
 363                  */
 364                 run_workqueue(cwq);
 365         } else {
 366                 struct wq_barrier barr;
 367                 int active = 0;
 368
 369                 spin_lock_irq(&cwq->lock);
 370                 if (!list_empty(&cwq->worklist) || cwq->current_work != NULL) {
 371                         insert_wq_barrier(cwq, &barr, 1);
 372                         active = 1;
 373                 }
 374                 spin_unlock_irq(&cwq->lock);
 375
 376                 if (active)
 377                         wait_for_completion(&barr.done);
 378         }
 379 }
 380
 381 /**
 382  * flush_workqueue - ensure that any scheduled work has run to completion.
 383  * @wq: workqueue to flush
 384  *
 385  * Forces execution of the workqueue and blocks until its completion.
 386  * This is typically used in driver shutdown handlers.
 387  *
 388  * We sleep until all works which were queued on entry have been handled,
 389  * but we are not livelocked by new incoming ones.
 390  *
 391  * This function used to run the workqueues itself.  Now we just wait for the
 392  * helper threads to do it.
 393  */
 394 void fastcall flush_workqueue(struct workqueue_struct *wq)
 395 {
 396         const cpumask_t *cpu_map = wq_cpu_map(wq);
 397         int cpu;
 398
 399         might_sleep();
 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 static void wait_on_work(struct cpu_workqueue_struct *cwq,
 406                                 struct work_struct *work)
 407 {
 408         struct wq_barrier barr;
 409         int running = 0;
 410
 411         spin_lock_irq(&cwq->lock);
 412         if (unlikely(cwq->current_work == work)) {
 413                 insert_wq_barrier(cwq, &barr, 0);
 414                 running = 1;
 415         }
 416         spin_unlock_irq(&cwq->lock);
 417
 418         if (unlikely(running))
 419                 wait_for_completion(&barr.done);
 420 }
 421
 422 /**
 423  * flush_work - block until a work_struct's callback has terminated
 424  * @wq: the workqueue on which the work is queued
 425  * @work: the work which is to be flushed
 426  *
 427  * flush_work() will attempt to cancel the work if it is queued.  If the work's
 428  * callback appears to be running, flush_work() will block until it has
 429  * completed.
 430  *
 431  * flush_work() is designed to be used when the caller is tearing down data
 432  * structures which the callback function operates upon.  It is expected that,
 433  * prior to calling flush_work(), the caller has arranged for the work to not
 434  * be requeued.
 435  */
 436 void flush_work(struct workqueue_struct *wq, struct work_struct *work)
 437 {
 438         const cpumask_t *cpu_map = wq_cpu_map(wq);
 439         struct cpu_workqueue_struct *cwq;
 440         int cpu;
 441
 442         might_sleep();
 443
 444         cwq = get_wq_data(work);
 445         /* Was it ever queued ? */
 446         if (!cwq)
 447                 return;
 448
 449         /*
 450          * This work can't be re-queued, no need to re-check that
 451          * get_wq_data() is still the same when we take cwq->lock.
 452          */
 453         spin_lock_irq(&cwq->lock);
 454         list_del_init(&work->entry);
 455         work_clear_pending(work);
 456         spin_unlock_irq(&cwq->lock);
 457
 458         for_each_cpu_mask(cpu, *cpu_map)
 459                 wait_on_work(per_cpu_ptr(wq->cpu_wq, cpu), work);
 460 }
 461 EXPORT_SYMBOL_GPL(flush_work);
 462
 463
 464 static struct workqueue_struct *keventd_wq;
 465
 466 /**
 467  * schedule_work - put work task in global workqueue
 468  * @work: job to be done
 469  *
 470  * This puts a job in the kernel-global workqueue.
 471  */
 472 int fastcall schedule_work(struct work_struct *work)
 473 {
 474         return queue_work(keventd_wq, work);
 475 }
 476 EXPORT_SYMBOL(schedule_work);
 477
 478 /**
 479  * schedule_delayed_work - put work task in global workqueue after delay
 480  * @dwork: job to be done
 481  * @delay: number of jiffies to wait or 0 for immediate execution
 482  *
 483  * After waiting for a given time this puts a job in the kernel-global
 484  * workqueue.
 485  */
 486 int fastcall schedule_delayed_work(struct delayed_work *dwork,
 487                                         unsigned long delay)
 488 {
 489         timer_stats_timer_set_start_info(&dwork->timer);
 490         return queue_delayed_work(keventd_wq, dwork, delay);
 491 }
 492 EXPORT_SYMBOL(schedule_delayed_work);
 493
 494 /**
 495  * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
 496  * @cpu: cpu to use
 497  * @dwork: job to be done
 498  * @delay: number of jiffies to wait
 499  *
 500  * After waiting for a given time this puts a job in the kernel-global
 501  * workqueue on the specified CPU.
 502  */
 503 int schedule_delayed_work_on(int cpu,
 504                         struct delayed_work *dwork, unsigned long delay)
 505 {
 506         return queue_delayed_work_on(cpu, keventd_wq, dwork, delay);
 507 }
 508 EXPORT_SYMBOL(schedule_delayed_work_on);
 509
 510 /**
 511  * schedule_on_each_cpu - call a function on each online CPU from keventd
 512  * @func: the function to call
 513  *
 514  * Returns zero on success.
 515  * Returns -ve errno on failure.
 516  *
 517  * Appears to be racy against CPU hotplug.
 518  *
 519  * schedule_on_each_cpu() is very slow.
 520  */
 521 int schedule_on_each_cpu(work_func_t func)
 522 {
 523         int cpu;
 524         struct work_struct *works;
 525
 526         works = alloc_percpu(struct work_struct);
 527         if (!works)
 528                 return -ENOMEM;
 529
 530         preempt_disable();              /* CPU hotplug */
 531         for_each_online_cpu(cpu) {
 532                 struct work_struct *work = per_cpu_ptr(works, cpu);
 533
 534                 INIT_WORK(work, func);
 535                 set_bit(WORK_STRUCT_PENDING, work_data_bits(work));
 536                 __queue_work(per_cpu_ptr(keventd_wq->cpu_wq, cpu), work);
 537         }
 538         preempt_enable();
 539         flush_workqueue(keventd_wq);
 540         free_percpu(works);
 541         return 0;
 542 }
 543
 544 void flush_scheduled_work(void)
 545 {
 546         flush_workqueue(keventd_wq);
 547 }
 548 EXPORT_SYMBOL(flush_scheduled_work);
 549
 550 void flush_work_keventd(struct work_struct *work)
 551 {
 552         flush_work(keventd_wq, work);
 553 }
 554 EXPORT_SYMBOL(flush_work_keventd);
 555
 556 /**
 557  * cancel_rearming_delayed_work - kill off a delayed work whose handler rearms the delayed work.
 558  * @dwork: the delayed work struct
 559  *
 560  * Note that the work callback function may still be running on return from
 561  * cancel_delayed_work(). Run flush_workqueue() or flush_work() to wait on it.
 562  */
 563 void cancel_rearming_delayed_work(struct delayed_work *dwork)
 564 {
 565         struct cpu_workqueue_struct *cwq = get_wq_data(&dwork->work);
 566
 567         /* Was it ever queued ? */
 568         if (cwq != NULL) {
 569                 struct workqueue_struct *wq = cwq->wq;
 570
 571                 while (!cancel_delayed_work(dwork))
 572                         flush_workqueue(wq);
 573         }
 574 }
 575 EXPORT_SYMBOL(cancel_rearming_delayed_work);
 576
 577 /**
 578  * execute_in_process_context - reliably execute the routine with user context
 579  * @fn:         the function to execute
 580  * @ew:         guaranteed storage for the execute work structure (must
 581  *              be available when the work executes)
 582  *
 583  * Executes the function immediately if process context is available,
 584  * otherwise schedules the function for delayed execution.
 585  *
 586  * Returns:     0 - function was executed
 587  *              1 - function was scheduled for execution
 588  */
 589 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
 590 {
 591         if (!in_interrupt()) {
 592                 fn(&ew->work);
 593                 return 0;
 594         }
 595
 596         INIT_WORK(&ew->work, fn);
 597         schedule_work(&ew->work);
 598
 599         return 1;
 600 }
 601 EXPORT_SYMBOL_GPL(execute_in_process_context);
 602
 603 int keventd_up(void)
 604 {
 605         return keventd_wq != NULL;
 606 }
 607
 608 int current_is_keventd(void)
 609 {
 610         struct cpu_workqueue_struct *cwq;
 611         int cpu = smp_processor_id();   /* preempt-safe: keventd is per-cpu */
 612         int ret = 0;
 613
 614         BUG_ON(!keventd_wq);
 615
 616         cwq = per_cpu_ptr(keventd_wq->cpu_wq, cpu);
 617         if (current == cwq->thread)
 618                 ret = 1;
 619
 620         return ret;
 621
 622 }
 623
 624 static struct cpu_workqueue_struct *
 625 init_cpu_workqueue(struct workqueue_struct *wq, int cpu)
 626 {
 627         struct cpu_workqueue_struct *cwq = per_cpu_ptr(wq->cpu_wq, cpu);
 628
 629         cwq->wq = wq;
 630         spin_lock_init(&cwq->lock);
 631         INIT_LIST_HEAD(&cwq->worklist);
 632         init_waitqueue_head(&cwq->more_work);
 633
 634         return cwq;
 635 }
 636
 637 static int create_workqueue_thread(struct cpu_workqueue_struct *cwq, int cpu)
 638 {
 639         struct workqueue_struct *wq = cwq->wq;
 640         const char *fmt = is_single_threaded(wq) ? "%s" : "%s/%d";
 641         struct task_struct *p;
 642
 643         p = kthread_create(worker_thread, cwq, fmt, wq->name, cpu);
 644         /*
 645          * Nobody can add the work_struct to this cwq,
 646          *      if (caller is __create_workqueue)
 647          *              nobody should see this wq
 648          *      else // caller is CPU_UP_PREPARE
 649          *              cpu is not on cpu_online_map
 650          * so we can abort safely.
 651          */
 652         if (IS_ERR(p))
 653                 return PTR_ERR(p);
 654
 655         cwq->thread = p;
 656         cwq->should_stop = 0;
 657
 658         return 0;
 659 }
 660
 661 static void start_workqueue_thread(struct cpu_workqueue_struct *cwq, int cpu)
 662 {
 663         struct task_struct *p = cwq->thread;
 664
 665         if (p != NULL) {
 666                 if (cpu >= 0)
 667                         kthread_bind(p, cpu);
 668                 wake_up_process(p);
 669         }
 670 }
 671
 672 struct workqueue_struct *__create_workqueue(const char *name,
 673                                             int singlethread, int freezeable)
 674 {
 675         struct workqueue_struct *wq;
 676         struct cpu_workqueue_struct *cwq;
 677         int err = 0, cpu;
 678
 679         wq = kzalloc(sizeof(*wq), GFP_KERNEL);
 680         if (!wq)
 681                 return NULL;
 682
 683         wq->cpu_wq = alloc_percpu(struct cpu_workqueue_struct);
 684         if (!wq->cpu_wq) {
 685                 kfree(wq);
 686                 return NULL;
 687         }
 688
 689         wq->name = name;
 690         wq->singlethread = singlethread;
 691         wq->freezeable = freezeable;
 692         INIT_LIST_HEAD(&wq->list);
 693
 694         if (singlethread) {
 695                 cwq = init_cpu_workqueue(wq, singlethread_cpu);
 696                 err = create_workqueue_thread(cwq, singlethread_cpu);
 697                 start_workqueue_thread(cwq, -1);
 698         } else {
 699                 mutex_lock(&workqueue_mutex);
 700                 list_add(&wq->list, &workqueues);
 701
 702                 for_each_possible_cpu(cpu) {
 703                         cwq = init_cpu_workqueue(wq, cpu);
 704                         if (err || !cpu_online(cpu))
 705                                 continue;
 706                         err = create_workqueue_thread(cwq, cpu);
 707                         start_workqueue_thread(cwq, cpu);
 708                 }
 709                 mutex_unlock(&workqueue_mutex);
 710         }
 711
 712         if (err) {
 713                 destroy_workqueue(wq);
 714                 wq = NULL;
 715         }
 716         return wq;
 717 }
 718 EXPORT_SYMBOL_GPL(__create_workqueue);
 719
 720 static void cleanup_workqueue_thread(struct cpu_workqueue_struct *cwq, int cpu)
 721 {
 722         struct wq_barrier barr;
 723         int alive = 0;
 724
 725         spin_lock_irq(&cwq->lock);
 726         if (cwq->thread != NULL) {
 727                 insert_wq_barrier(cwq, &barr, 1);
 728                 cwq->should_stop = 1;
 729                 alive = 1;
 730         }
 731         spin_unlock_irq(&cwq->lock);
 732
 733         if (alive) {
 734                 wait_for_completion(&barr.done);
 735
 736                 while (unlikely(cwq->thread != NULL))
 737                         cpu_relax();
 738                 /*
 739                  * Wait until cwq->thread unlocks cwq->lock,
 740                  * it won't touch *cwq after that.
 741                  */
 742                 smp_rmb();
 743                 spin_unlock_wait(&cwq->lock);
 744         }
 745 }
 746
 747 /**
 748  * destroy_workqueue - safely terminate a workqueue
 749  * @wq: target workqueue
 750  *
 751  * Safely destroy a workqueue. All work currently pending will be done first.
 752  */
 753 void destroy_workqueue(struct workqueue_struct *wq)
 754 {
 755         const cpumask_t *cpu_map = wq_cpu_map(wq);
 756         struct cpu_workqueue_struct *cwq;
 757         int cpu;
 758
 759         mutex_lock(&workqueue_mutex);
 760         list_del(&wq->list);
 761         mutex_unlock(&workqueue_mutex);
 762
 763         for_each_cpu_mask(cpu, *cpu_map) {
 764                 cwq = per_cpu_ptr(wq->cpu_wq, cpu);
 765                 cleanup_workqueue_thread(cwq, cpu);
 766         }
 767
 768         free_percpu(wq->cpu_wq);
 769         kfree(wq);
 770 }
 771 EXPORT_SYMBOL_GPL(destroy_workqueue);
 772
 773 static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
 774                                                 unsigned long action,
 775                                                 void *hcpu)
 776 {
 777         unsigned int cpu = (unsigned long)hcpu;
 778         struct cpu_workqueue_struct *cwq;
 779         struct workqueue_struct *wq;
 780
 781         switch (action) {
 782         case CPU_LOCK_ACQUIRE:
 783                 mutex_lock(&workqueue_mutex);
 784                 return NOTIFY_OK;
 785
 786         case CPU_LOCK_RELEASE:
 787                 mutex_unlock(&workqueue_mutex);
 788                 return NOTIFY_OK;
 789
 790         case CPU_UP_PREPARE:
 791                 cpu_set(cpu, cpu_populated_map);
 792         }
 793
 794         list_for_each_entry(wq, &workqueues, list) {
 795                 cwq = per_cpu_ptr(wq->cpu_wq, cpu);
 796
 797                 switch (action) {
 798                 case CPU_UP_PREPARE:
 799                         if (!create_workqueue_thread(cwq, cpu))
 800                                 break;
 801                         printk(KERN_ERR "workqueue for %i failed\n", cpu);
 802                         return NOTIFY_BAD;
 803
 804                 case CPU_ONLINE:
 805                         start_workqueue_thread(cwq, cpu);
 806                         break;
 807
 808                 case CPU_UP_CANCELED:
 809                         start_workqueue_thread(cwq, -1);
 810                 case CPU_DEAD:
 811                         cleanup_workqueue_thread(cwq, cpu);
 812                         break;
 813                 }
 814         }
 815
 816         return NOTIFY_OK;
 817 }
 818
 819 void __init init_workqueues(void)
 820 {
 821         cpu_populated_map = cpu_online_map;
 822         singlethread_cpu = first_cpu(cpu_possible_map);
 823         cpu_singlethread_map = cpumask_of_cpu(singlethread_cpu);
 824         hotcpu_notifier(workqueue_cpu_callback, 0);
 825         keventd_wq = create_workqueue("events");
 826         BUG_ON(!keventd_wq);
 827 }