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
 293         if (!cwq->wq->freezeable)
 294                 current->flags |= PF_NOFREEZE;
 295
 296         set_user_nice(current, -5);
 297
 298         for (;;) {
 299                 prepare_to_wait(&cwq->more_work, &wait, TASK_INTERRUPTIBLE);
 300                 if (!freezing(current) && !cwq->should_stop
 301                     && list_empty(&cwq->worklist))
 302                         schedule();
 303                 finish_wait(&cwq->more_work, &wait);
 304
 305                 try_to_freeze();
 306
 307                 if (cwq_should_stop(cwq))
 308                         break;
 309
 310                 run_workqueue(cwq);
 311         }
 312
 313         return 0;
 314 }
 315
 316 struct wq_barrier {
 317         struct work_struct      work;
 318         struct completion       done;
 319 };
 320
 321 static void wq_barrier_func(struct work_struct *work)
 322 {
 323         struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
 324         complete(&barr->done);
 325 }
 326
 327 static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
 328                                         struct wq_barrier *barr, int tail)
 329 {
 330         INIT_WORK(&barr->work, wq_barrier_func);
 331         __set_bit(WORK_STRUCT_PENDING, work_data_bits(&barr->work));
 332
 333         init_completion(&barr->done);
 334
 335         insert_work(cwq, &barr->work, tail);
 336 }
 337
 338 static void flush_cpu_workqueue(struct cpu_workqueue_struct *cwq)
 339 {
 340         if (cwq->thread == current) {
 341                 /*
 342                  * Probably keventd trying to flush its own queue. So simply run
 343                  * it by hand rather than deadlocking.
 344                  */
 345                 run_workqueue(cwq);
 346         } else {
 347                 struct wq_barrier barr;
 348                 int active = 0;
 349
 350                 spin_lock_irq(&cwq->lock);
 351                 if (!list_empty(&cwq->worklist) || cwq->current_work != NULL) {
 352                         insert_wq_barrier(cwq, &barr, 1);
 353                         active = 1;
 354                 }
 355                 spin_unlock_irq(&cwq->lock);
 356
 357                 if (active)
 358                         wait_for_completion(&barr.done);
 359         }
 360 }
 361
 362 /**
 363  * flush_workqueue - ensure that any scheduled work has run to completion.
 364  * @wq: workqueue to flush
 365  *
 366  * Forces execution of the workqueue and blocks until its completion.
 367  * This is typically used in driver shutdown handlers.
 368  *
 369  * We sleep until all works which were queued on entry have been handled,
 370  * but we are not livelocked by new incoming ones.
 371  *
 372  * This function used to run the workqueues itself.  Now we just wait for the
 373  * helper threads to do it.
 374  */
 375 void fastcall flush_workqueue(struct workqueue_struct *wq)
 376 {
 377         const cpumask_t *cpu_map = wq_cpu_map(wq);
 378         int cpu;
 379
 380         might_sleep();
 381         for_each_cpu_mask(cpu, *cpu_map)
 382                 flush_cpu_workqueue(per_cpu_ptr(wq->cpu_wq, cpu));
 383 }
 384 EXPORT_SYMBOL_GPL(flush_workqueue);
 385
 386 static void wait_on_work(struct cpu_workqueue_struct *cwq,
 387                                 struct work_struct *work)
 388 {
 389         struct wq_barrier barr;
 390         int running = 0;
 391
 392         spin_lock_irq(&cwq->lock);
 393         if (unlikely(cwq->current_work == work)) {
 394                 insert_wq_barrier(cwq, &barr, 0);
 395                 running = 1;
 396         }
 397         spin_unlock_irq(&cwq->lock);
 398
 399         if (unlikely(running))
 400                 wait_for_completion(&barr.done);
 401 }
 402
 403 /**
 404  * cancel_work_sync - block until a work_struct's callback has terminated
 405  * @work: the work which is to be flushed
 406  *
 407  * cancel_work_sync() will attempt to cancel the work if it is queued. If the
 408  * work's callback appears to be running, cancel_work_sync() will block until
 409  * it has completed.
 410  *
 411  * cancel_work_sync() is designed to be used when the caller is tearing down
 412  * data structures which the callback function operates upon. It is expected
 413  * that, prior to calling cancel_work_sync(), the caller has arranged for the
 414  * work to not be requeued.
 415  */
 416 void cancel_work_sync(struct work_struct *work)
 417 {
 418         struct cpu_workqueue_struct *cwq;
 419         struct workqueue_struct *wq;
 420         const cpumask_t *cpu_map;
 421         int cpu;
 422
 423         might_sleep();
 424
 425         cwq = get_wq_data(work);
 426         /* Was it ever queued ? */
 427         if (!cwq)
 428                 return;
 429
 430         /*
 431          * This work can't be re-queued, no need to re-check that
 432          * get_wq_data() is still the same when we take cwq->lock.
 433          */
 434         spin_lock_irq(&cwq->lock);
 435         list_del_init(&work->entry);
 436         work_clear_pending(work);
 437         spin_unlock_irq(&cwq->lock);
 438
 439         wq = cwq->wq;
 440         cpu_map = wq_cpu_map(wq);
 441
 442         for_each_cpu_mask(cpu, *cpu_map)
 443                 wait_on_work(per_cpu_ptr(wq->cpu_wq, cpu), work);
 444 }
 445 EXPORT_SYMBOL_GPL(cancel_work_sync);
 446
 447
 448 static struct workqueue_struct *keventd_wq;
 449
 450 /**
 451  * schedule_work - put work task in global workqueue
 452  * @work: job to be done
 453  *
 454  * This puts a job in the kernel-global workqueue.
 455  */
 456 int fastcall schedule_work(struct work_struct *work)
 457 {
 458         return queue_work(keventd_wq, work);
 459 }
 460 EXPORT_SYMBOL(schedule_work);
 461
 462 /**
 463  * schedule_delayed_work - put work task in global workqueue after delay
 464  * @dwork: job to be done
 465  * @delay: number of jiffies to wait or 0 for immediate execution
 466  *
 467  * After waiting for a given time this puts a job in the kernel-global
 468  * workqueue.
 469  */
 470 int fastcall schedule_delayed_work(struct delayed_work *dwork,
 471                                         unsigned long delay)
 472 {
 473         timer_stats_timer_set_start_info(&dwork->timer);
 474         return queue_delayed_work(keventd_wq, dwork, delay);
 475 }
 476 EXPORT_SYMBOL(schedule_delayed_work);
 477
 478 /**
 479  * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
 480  * @cpu: cpu to use
 481  * @dwork: job to be done
 482  * @delay: number of jiffies to wait
 483  *
 484  * After waiting for a given time this puts a job in the kernel-global
 485  * workqueue on the specified CPU.
 486  */
 487 int schedule_delayed_work_on(int cpu,
 488                         struct delayed_work *dwork, unsigned long delay)
 489 {
 490         return queue_delayed_work_on(cpu, keventd_wq, dwork, delay);
 491 }
 492 EXPORT_SYMBOL(schedule_delayed_work_on);
 493
 494 /**
 495  * schedule_on_each_cpu - call a function on each online CPU from keventd
 496  * @func: the function to call
 497  *
 498  * Returns zero on success.
 499  * Returns -ve errno on failure.
 500  *
 501  * Appears to be racy against CPU hotplug.
 502  *
 503  * schedule_on_each_cpu() is very slow.
 504  */
 505 int schedule_on_each_cpu(work_func_t func)
 506 {
 507         int cpu;
 508         struct work_struct *works;
 509
 510         works = alloc_percpu(struct work_struct);
 511         if (!works)
 512                 return -ENOMEM;
 513
 514         preempt_disable();              /* CPU hotplug */
 515         for_each_online_cpu(cpu) {
 516                 struct work_struct *work = per_cpu_ptr(works, cpu);
 517
 518                 INIT_WORK(work, func);
 519                 set_bit(WORK_STRUCT_PENDING, work_data_bits(work));
 520                 __queue_work(per_cpu_ptr(keventd_wq->cpu_wq, cpu), work);
 521         }
 522         preempt_enable();
 523         flush_workqueue(keventd_wq);
 524         free_percpu(works);
 525         return 0;
 526 }
 527
 528 void flush_scheduled_work(void)
 529 {
 530         flush_workqueue(keventd_wq);
 531 }
 532 EXPORT_SYMBOL(flush_scheduled_work);
 533
 534 /**
 535  * cancel_rearming_delayed_work - kill off a delayed work whose handler rearms the delayed work.
 536  * @dwork: the delayed work struct
 537  *
 538  * Note that the work callback function may still be running on return from
 539  * cancel_delayed_work(). Run flush_workqueue() or cancel_work_sync() to wait
 540  * on it.
 541  */
 542 void cancel_rearming_delayed_work(struct delayed_work *dwork)
 543 {
 544         struct cpu_workqueue_struct *cwq = get_wq_data(&dwork->work);
 545
 546         /* Was it ever queued ? */
 547         if (cwq != NULL) {
 548                 struct workqueue_struct *wq = cwq->wq;
 549
 550                 while (!cancel_delayed_work(dwork))
 551                         flush_workqueue(wq);
 552         }
 553 }
 554 EXPORT_SYMBOL(cancel_rearming_delayed_work);
 555
 556 /**
 557  * execute_in_process_context - reliably execute the routine with user context
 558  * @fn:         the function to execute
 559  * @ew:         guaranteed storage for the execute work structure (must
 560  *              be available when the work executes)
 561  *
 562  * Executes the function immediately if process context is available,
 563  * otherwise schedules the function for delayed execution.
 564  *
 565  * Returns:     0 - function was executed
 566  *              1 - function was scheduled for execution
 567  */
 568 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
 569 {
 570         if (!in_interrupt()) {
 571                 fn(&ew->work);
 572                 return 0;
 573         }
 574
 575         INIT_WORK(&ew->work, fn);
 576         schedule_work(&ew->work);
 577
 578         return 1;
 579 }
 580 EXPORT_SYMBOL_GPL(execute_in_process_context);
 581
 582 int keventd_up(void)
 583 {
 584         return keventd_wq != NULL;
 585 }
 586
 587 int current_is_keventd(void)
 588 {
 589         struct cpu_workqueue_struct *cwq;
 590         int cpu = smp_processor_id();   /* preempt-safe: keventd is per-cpu */
 591         int ret = 0;
 592
 593         BUG_ON(!keventd_wq);
 594
 595         cwq = per_cpu_ptr(keventd_wq->cpu_wq, cpu);
 596         if (current == cwq->thread)
 597                 ret = 1;
 598
 599         return ret;
 600
 601 }
 602
 603 static struct cpu_workqueue_struct *
 604 init_cpu_workqueue(struct workqueue_struct *wq, int cpu)
 605 {
 606         struct cpu_workqueue_struct *cwq = per_cpu_ptr(wq->cpu_wq, cpu);
 607
 608         cwq->wq = wq;
 609         spin_lock_init(&cwq->lock);
 610         INIT_LIST_HEAD(&cwq->worklist);
 611         init_waitqueue_head(&cwq->more_work);
 612
 613         return cwq;
 614 }
 615
 616 static int create_workqueue_thread(struct cpu_workqueue_struct *cwq, int cpu)
 617 {
 618         struct workqueue_struct *wq = cwq->wq;
 619         const char *fmt = is_single_threaded(wq) ? "%s" : "%s/%d";
 620         struct task_struct *p;
 621
 622         p = kthread_create(worker_thread, cwq, fmt, wq->name, cpu);
 623         /*
 624          * Nobody can add the work_struct to this cwq,
 625          *      if (caller is __create_workqueue)
 626          *              nobody should see this wq
 627          *      else // caller is CPU_UP_PREPARE
 628          *              cpu is not on cpu_online_map
 629          * so we can abort safely.
 630          */
 631         if (IS_ERR(p))
 632                 return PTR_ERR(p);
 633
 634         cwq->thread = p;
 635         cwq->should_stop = 0;
 636
 637         return 0;
 638 }
 639
 640 static void start_workqueue_thread(struct cpu_workqueue_struct *cwq, int cpu)
 641 {
 642         struct task_struct *p = cwq->thread;
 643
 644         if (p != NULL) {
 645                 if (cpu >= 0)
 646                         kthread_bind(p, cpu);
 647                 wake_up_process(p);
 648         }
 649 }
 650
 651 struct workqueue_struct *__create_workqueue(const char *name,
 652                                             int singlethread, int freezeable)
 653 {
 654         struct workqueue_struct *wq;
 655         struct cpu_workqueue_struct *cwq;
 656         int err = 0, cpu;
 657
 658         wq = kzalloc(sizeof(*wq), GFP_KERNEL);
 659         if (!wq)
 660                 return NULL;
 661
 662         wq->cpu_wq = alloc_percpu(struct cpu_workqueue_struct);
 663         if (!wq->cpu_wq) {
 664                 kfree(wq);
 665                 return NULL;
 666         }
 667
 668         wq->name = name;
 669         wq->singlethread = singlethread;
 670         wq->freezeable = freezeable;
 671         INIT_LIST_HEAD(&wq->list);
 672
 673         if (singlethread) {
 674                 cwq = init_cpu_workqueue(wq, singlethread_cpu);
 675                 err = create_workqueue_thread(cwq, singlethread_cpu);
 676                 start_workqueue_thread(cwq, -1);
 677         } else {
 678                 mutex_lock(&workqueue_mutex);
 679                 list_add(&wq->list, &workqueues);
 680
 681                 for_each_possible_cpu(cpu) {
 682                         cwq = init_cpu_workqueue(wq, cpu);
 683                         if (err || !cpu_online(cpu))
 684                                 continue;
 685                         err = create_workqueue_thread(cwq, cpu);
 686                         start_workqueue_thread(cwq, cpu);
 687                 }
 688                 mutex_unlock(&workqueue_mutex);
 689         }
 690
 691         if (err) {
 692                 destroy_workqueue(wq);
 693                 wq = NULL;
 694         }
 695         return wq;
 696 }
 697 EXPORT_SYMBOL_GPL(__create_workqueue);
 698
 699 static void cleanup_workqueue_thread(struct cpu_workqueue_struct *cwq, int cpu)
 700 {
 701         struct wq_barrier barr;
 702         int alive = 0;
 703
 704         spin_lock_irq(&cwq->lock);
 705         if (cwq->thread != NULL) {
 706                 insert_wq_barrier(cwq, &barr, 1);
 707                 cwq->should_stop = 1;
 708                 alive = 1;
 709         }
 710         spin_unlock_irq(&cwq->lock);
 711
 712         if (alive) {
 713                 wait_for_completion(&barr.done);
 714
 715                 while (unlikely(cwq->thread != NULL))
 716                         cpu_relax();
 717                 /*
 718                  * Wait until cwq->thread unlocks cwq->lock,
 719                  * it won't touch *cwq after that.
 720                  */
 721                 smp_rmb();
 722                 spin_unlock_wait(&cwq->lock);
 723         }
 724 }
 725
 726 /**
 727  * destroy_workqueue - safely terminate a workqueue
 728  * @wq: target workqueue
 729  *
 730  * Safely destroy a workqueue. All work currently pending will be done first.
 731  */
 732 void destroy_workqueue(struct workqueue_struct *wq)
 733 {
 734         const cpumask_t *cpu_map = wq_cpu_map(wq);
 735         struct cpu_workqueue_struct *cwq;
 736         int cpu;
 737
 738         mutex_lock(&workqueue_mutex);
 739         list_del(&wq->list);
 740         mutex_unlock(&workqueue_mutex);
 741
 742         for_each_cpu_mask(cpu, *cpu_map) {
 743                 cwq = per_cpu_ptr(wq->cpu_wq, cpu);
 744                 cleanup_workqueue_thread(cwq, cpu);
 745         }
 746
 747         free_percpu(wq->cpu_wq);
 748         kfree(wq);
 749 }
 750 EXPORT_SYMBOL_GPL(destroy_workqueue);
 751
 752 static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
 753                                                 unsigned long action,
 754                                                 void *hcpu)
 755 {
 756         unsigned int cpu = (unsigned long)hcpu;
 757         struct cpu_workqueue_struct *cwq;
 758         struct workqueue_struct *wq;
 759
 760         switch (action) {
 761         case CPU_LOCK_ACQUIRE:
 762                 mutex_lock(&workqueue_mutex);
 763                 return NOTIFY_OK;
 764
 765         case CPU_LOCK_RELEASE:
 766                 mutex_unlock(&workqueue_mutex);
 767                 return NOTIFY_OK;
 768
 769         case CPU_UP_PREPARE:
 770                 cpu_set(cpu, cpu_populated_map);
 771         }
 772
 773         list_for_each_entry(wq, &workqueues, list) {
 774                 cwq = per_cpu_ptr(wq->cpu_wq, cpu);
 775
 776                 switch (action) {
 777                 case CPU_UP_PREPARE:
 778                         if (!create_workqueue_thread(cwq, cpu))
 779                                 break;
 780                         printk(KERN_ERR "workqueue for %i failed\n", cpu);
 781                         return NOTIFY_BAD;
 782
 783                 case CPU_ONLINE:
 784                         start_workqueue_thread(cwq, cpu);
 785                         break;
 786
 787                 case CPU_UP_CANCELED:
 788                         start_workqueue_thread(cwq, -1);
 789                 case CPU_DEAD:
 790                         cleanup_workqueue_thread(cwq, cpu);
 791                         break;
 792                 }
 793         }
 794
 795         return NOTIFY_OK;
 796 }
 797
 798 void __init init_workqueues(void)
 799 {
 800         cpu_populated_map = cpu_online_map;
 801         singlethread_cpu = first_cpu(cpu_possible_map);
 802         cpu_singlethread_map = cpumask_of_cpu(singlethread_cpu);
 803         hotcpu_notifier(workqueue_cpu_callback, 0);
 804         keventd_wq = create_workqueue("events");
 805         BUG_ON(!keventd_wq);
 806 }