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
 294         if (!cwq->wq->freezeable)
 295                 current->flags |= PF_NOFREEZE;
 296
 297         set_user_nice(current, -5);
 298         /*
 299          * We inherited MPOL_INTERLEAVE from the booting kernel.
 300          * Set MPOL_DEFAULT to insure node local allocations.
 301          */
 302         numa_default_policy();
 303
 304         /* SIG_IGN makes children autoreap: see do_notify_parent(). */
 305         sa.sa.sa_handler = SIG_IGN;
 306         sa.sa.sa_flags = 0;
 307         siginitset(&sa.sa.sa_mask, sigmask(SIGCHLD));
 308         do_sigaction(SIGCHLD, &sa, (struct k_sigaction *)0);
 309
 310         for (;;) {
 311                 prepare_to_wait(&cwq->more_work, &wait, TASK_INTERRUPTIBLE);
 312                 if (!freezing(current) && !cwq->should_stop
 313                     && list_empty(&cwq->worklist))
 314                         schedule();
 315                 finish_wait(&cwq->more_work, &wait);
 316
 317                 try_to_freeze();
 318
 319                 if (cwq_should_stop(cwq))
 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 void flush_cpu_workqueue(struct cpu_workqueue_struct *cwq)
 351 {
 352         if (cwq->thread == current) {
 353                 /*
 354                  * Probably keventd trying to flush its own queue. So simply run
 355                  * it by hand rather than deadlocking.
 356                  */
 357                 run_workqueue(cwq);
 358         } else {
 359                 struct wq_barrier barr;
 360                 int active = 0;
 361
 362                 spin_lock_irq(&cwq->lock);
 363                 if (!list_empty(&cwq->worklist) || cwq->current_work != NULL) {
 364                         insert_wq_barrier(cwq, &barr, 1);
 365                         active = 1;
 366                 }
 367                 spin_unlock_irq(&cwq->lock);
 368
 369                 if (active)
 370                         wait_for_completion(&barr.done);
 371         }
 372 }
 373
 374 /**
 375  * flush_workqueue - ensure that any scheduled work has run to completion.
 376  * @wq: workqueue to flush
 377  *
 378  * Forces execution of the workqueue and blocks until its completion.
 379  * This is typically used in driver shutdown handlers.
 380  *
 381  * We sleep until all works which were queued on entry have been handled,
 382  * but we are not livelocked by new incoming ones.
 383  *
 384  * This function used to run the workqueues itself.  Now we just wait for the
 385  * helper threads to do it.
 386  */
 387 void fastcall flush_workqueue(struct workqueue_struct *wq)
 388 {
 389         const cpumask_t *cpu_map = wq_cpu_map(wq);
 390         int cpu;
 391
 392         might_sleep();
 393         for_each_cpu_mask(cpu, *cpu_map)
 394                 flush_cpu_workqueue(per_cpu_ptr(wq->cpu_wq, cpu));
 395 }
 396 EXPORT_SYMBOL_GPL(flush_workqueue);
 397
 398 static void wait_on_work(struct cpu_workqueue_struct *cwq,
 399                                 struct work_struct *work)
 400 {
 401         struct wq_barrier barr;
 402         int running = 0;
 403
 404         spin_lock_irq(&cwq->lock);
 405         if (unlikely(cwq->current_work == work)) {
 406                 insert_wq_barrier(cwq, &barr, 0);
 407                 running = 1;
 408         }
 409         spin_unlock_irq(&cwq->lock);
 410
 411         if (unlikely(running))
 412                 wait_for_completion(&barr.done);
 413 }
 414
 415 /**
 416  * cancel_work_sync - block until a work_struct's callback has terminated
 417  * @work: the work which is to be flushed
 418  *
 419  * cancel_work_sync() will attempt to cancel the work if it is queued. If the
 420  * work's callback appears to be running, cancel_work_sync() will block until
 421  * it has completed.
 422  *
 423  * cancel_work_sync() is designed to be used when the caller is tearing down
 424  * data structures which the callback function operates upon. It is expected
 425  * that, prior to calling cancel_work_sync(), the caller has arranged for the
 426  * work to not be requeued.
 427  */
 428 void cancel_work_sync(struct work_struct *work)
 429 {
 430         struct cpu_workqueue_struct *cwq;
 431         struct workqueue_struct *wq;
 432         const cpumask_t *cpu_map;
 433         int cpu;
 434
 435         might_sleep();
 436
 437         cwq = get_wq_data(work);
 438         /* Was it ever queued ? */
 439         if (!cwq)
 440                 return;
 441
 442         /*
 443          * This work can't be re-queued, no need to re-check that
 444          * get_wq_data() is still the same when we take cwq->lock.
 445          */
 446         spin_lock_irq(&cwq->lock);
 447         list_del_init(&work->entry);
 448         work_clear_pending(work);
 449         spin_unlock_irq(&cwq->lock);
 450
 451         wq = cwq->wq;
 452         cpu_map = wq_cpu_map(wq);
 453
 454         for_each_cpu_mask(cpu, *cpu_map)
 455                 wait_on_work(per_cpu_ptr(wq->cpu_wq, cpu), work);
 456 }
 457 EXPORT_SYMBOL_GPL(cancel_work_sync);
 458
 459
 460 static struct workqueue_struct *keventd_wq;
 461
 462 /**
 463  * schedule_work - put work task in global workqueue
 464  * @work: job to be done
 465  *
 466  * This puts a job in the kernel-global workqueue.
 467  */
 468 int fastcall schedule_work(struct work_struct *work)
 469 {
 470         return queue_work(keventd_wq, work);
 471 }
 472 EXPORT_SYMBOL(schedule_work);
 473
 474 /**
 475  * schedule_delayed_work - put work task in global workqueue after delay
 476  * @dwork: job to be done
 477  * @delay: number of jiffies to wait or 0 for immediate execution
 478  *
 479  * After waiting for a given time this puts a job in the kernel-global
 480  * workqueue.
 481  */
 482 int fastcall schedule_delayed_work(struct delayed_work *dwork,
 483                                         unsigned long delay)
 484 {
 485         timer_stats_timer_set_start_info(&dwork->timer);
 486         return queue_delayed_work(keventd_wq, dwork, delay);
 487 }
 488 EXPORT_SYMBOL(schedule_delayed_work);
 489
 490 /**
 491  * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
 492  * @cpu: cpu to use
 493  * @dwork: job to be done
 494  * @delay: number of jiffies to wait
 495  *
 496  * After waiting for a given time this puts a job in the kernel-global
 497  * workqueue on the specified CPU.
 498  */
 499 int schedule_delayed_work_on(int cpu,
 500                         struct delayed_work *dwork, unsigned long delay)
 501 {
 502         return queue_delayed_work_on(cpu, keventd_wq, dwork, delay);
 503 }
 504 EXPORT_SYMBOL(schedule_delayed_work_on);
 505
 506 /**
 507  * schedule_on_each_cpu - call a function on each online CPU from keventd
 508  * @func: the function to call
 509  *
 510  * Returns zero on success.
 511  * Returns -ve errno on failure.
 512  *
 513  * Appears to be racy against CPU hotplug.
 514  *
 515  * schedule_on_each_cpu() is very slow.
 516  */
 517 int schedule_on_each_cpu(work_func_t func)
 518 {
 519         int cpu;
 520         struct work_struct *works;
 521
 522         works = alloc_percpu(struct work_struct);
 523         if (!works)
 524                 return -ENOMEM;
 525
 526         preempt_disable();              /* CPU hotplug */
 527         for_each_online_cpu(cpu) {
 528                 struct work_struct *work = per_cpu_ptr(works, cpu);
 529
 530                 INIT_WORK(work, func);
 531                 set_bit(WORK_STRUCT_PENDING, work_data_bits(work));
 532                 __queue_work(per_cpu_ptr(keventd_wq->cpu_wq, cpu), work);
 533         }
 534         preempt_enable();
 535         flush_workqueue(keventd_wq);
 536         free_percpu(works);
 537         return 0;
 538 }
 539
 540 void flush_scheduled_work(void)
 541 {
 542         flush_workqueue(keventd_wq);
 543 }
 544 EXPORT_SYMBOL(flush_scheduled_work);
 545
 546 /**
 547  * cancel_rearming_delayed_work - kill off a delayed work whose handler rearms the delayed work.
 548  * @dwork: the delayed work struct
 549  *
 550  * Note that the work callback function may still be running on return from
 551  * cancel_delayed_work(). Run flush_workqueue() or cancel_work_sync() to wait
 552  * on it.
 553  */
 554 void cancel_rearming_delayed_work(struct delayed_work *dwork)
 555 {
 556         struct cpu_workqueue_struct *cwq = get_wq_data(&dwork->work);
 557
 558         /* Was it ever queued ? */
 559         if (cwq != NULL) {
 560                 struct workqueue_struct *wq = cwq->wq;
 561
 562                 while (!cancel_delayed_work(dwork))
 563                         flush_workqueue(wq);
 564         }
 565 }
 566 EXPORT_SYMBOL(cancel_rearming_delayed_work);
 567
 568 /**
 569  * execute_in_process_context - reliably execute the routine with user context
 570  * @fn:         the function to execute
 571  * @ew:         guaranteed storage for the execute work structure (must
 572  *              be available when the work executes)
 573  *
 574  * Executes the function immediately if process context is available,
 575  * otherwise schedules the function for delayed execution.
 576  *
 577  * Returns:     0 - function was executed
 578  *              1 - function was scheduled for execution
 579  */
 580 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
 581 {
 582         if (!in_interrupt()) {
 583                 fn(&ew->work);
 584                 return 0;
 585         }
 586
 587         INIT_WORK(&ew->work, fn);
 588         schedule_work(&ew->work);
 589
 590         return 1;
 591 }
 592 EXPORT_SYMBOL_GPL(execute_in_process_context);
 593
 594 int keventd_up(void)
 595 {
 596         return keventd_wq != NULL;
 597 }
 598
 599 int current_is_keventd(void)
 600 {
 601         struct cpu_workqueue_struct *cwq;
 602         int cpu = smp_processor_id();   /* preempt-safe: keventd is per-cpu */
 603         int ret = 0;
 604
 605         BUG_ON(!keventd_wq);
 606
 607         cwq = per_cpu_ptr(keventd_wq->cpu_wq, cpu);
 608         if (current == cwq->thread)
 609                 ret = 1;
 610
 611         return ret;
 612
 613 }
 614
 615 static struct cpu_workqueue_struct *
 616 init_cpu_workqueue(struct workqueue_struct *wq, int cpu)
 617 {
 618         struct cpu_workqueue_struct *cwq = per_cpu_ptr(wq->cpu_wq, cpu);
 619
 620         cwq->wq = wq;
 621         spin_lock_init(&cwq->lock);
 622         INIT_LIST_HEAD(&cwq->worklist);
 623         init_waitqueue_head(&cwq->more_work);
 624
 625         return cwq;
 626 }
 627
 628 static int create_workqueue_thread(struct cpu_workqueue_struct *cwq, int cpu)
 629 {
 630         struct workqueue_struct *wq = cwq->wq;
 631         const char *fmt = is_single_threaded(wq) ? "%s" : "%s/%d";
 632         struct task_struct *p;
 633
 634         p = kthread_create(worker_thread, cwq, fmt, wq->name, cpu);
 635         /*
 636          * Nobody can add the work_struct to this cwq,
 637          *      if (caller is __create_workqueue)
 638          *              nobody should see this wq
 639          *      else // caller is CPU_UP_PREPARE
 640          *              cpu is not on cpu_online_map
 641          * so we can abort safely.
 642          */
 643         if (IS_ERR(p))
 644                 return PTR_ERR(p);
 645
 646         cwq->thread = p;
 647         cwq->should_stop = 0;
 648
 649         return 0;
 650 }
 651
 652 static void start_workqueue_thread(struct cpu_workqueue_struct *cwq, int cpu)
 653 {
 654         struct task_struct *p = cwq->thread;
 655
 656         if (p != NULL) {
 657                 if (cpu >= 0)
 658                         kthread_bind(p, cpu);
 659                 wake_up_process(p);
 660         }
 661 }
 662
 663 struct workqueue_struct *__create_workqueue(const char *name,
 664                                             int singlethread, int freezeable)
 665 {
 666         struct workqueue_struct *wq;
 667         struct cpu_workqueue_struct *cwq;
 668         int err = 0, cpu;
 669
 670         wq = kzalloc(sizeof(*wq), GFP_KERNEL);
 671         if (!wq)
 672                 return NULL;
 673
 674         wq->cpu_wq = alloc_percpu(struct cpu_workqueue_struct);
 675         if (!wq->cpu_wq) {
 676                 kfree(wq);
 677                 return NULL;
 678         }
 679
 680         wq->name = name;
 681         wq->singlethread = singlethread;
 682         wq->freezeable = freezeable;
 683         INIT_LIST_HEAD(&wq->list);
 684
 685         if (singlethread) {
 686                 cwq = init_cpu_workqueue(wq, singlethread_cpu);
 687                 err = create_workqueue_thread(cwq, singlethread_cpu);
 688                 start_workqueue_thread(cwq, -1);
 689         } else {
 690                 mutex_lock(&workqueue_mutex);
 691                 list_add(&wq->list, &workqueues);
 692
 693                 for_each_possible_cpu(cpu) {
 694                         cwq = init_cpu_workqueue(wq, cpu);
 695                         if (err || !cpu_online(cpu))
 696                                 continue;
 697                         err = create_workqueue_thread(cwq, cpu);
 698                         start_workqueue_thread(cwq, cpu);
 699                 }
 700                 mutex_unlock(&workqueue_mutex);
 701         }
 702
 703         if (err) {
 704                 destroy_workqueue(wq);
 705                 wq = NULL;
 706         }
 707         return wq;
 708 }
 709 EXPORT_SYMBOL_GPL(__create_workqueue);
 710
 711 static void cleanup_workqueue_thread(struct cpu_workqueue_struct *cwq, int cpu)
 712 {
 713         struct wq_barrier barr;
 714         int alive = 0;
 715
 716         spin_lock_irq(&cwq->lock);
 717         if (cwq->thread != NULL) {
 718                 insert_wq_barrier(cwq, &barr, 1);
 719                 cwq->should_stop = 1;
 720                 alive = 1;
 721         }
 722         spin_unlock_irq(&cwq->lock);
 723
 724         if (alive) {
 725                 wait_for_completion(&barr.done);
 726
 727                 while (unlikely(cwq->thread != NULL))
 728                         cpu_relax();
 729                 /*
 730                  * Wait until cwq->thread unlocks cwq->lock,
 731                  * it won't touch *cwq after that.
 732                  */
 733                 smp_rmb();
 734                 spin_unlock_wait(&cwq->lock);
 735         }
 736 }
 737
 738 /**
 739  * destroy_workqueue - safely terminate a workqueue
 740  * @wq: target workqueue
 741  *
 742  * Safely destroy a workqueue. All work currently pending will be done first.
 743  */
 744 void destroy_workqueue(struct workqueue_struct *wq)
 745 {
 746         const cpumask_t *cpu_map = wq_cpu_map(wq);
 747         struct cpu_workqueue_struct *cwq;
 748         int cpu;
 749
 750         mutex_lock(&workqueue_mutex);
 751         list_del(&wq->list);
 752         mutex_unlock(&workqueue_mutex);
 753
 754         for_each_cpu_mask(cpu, *cpu_map) {
 755                 cwq = per_cpu_ptr(wq->cpu_wq, cpu);
 756                 cleanup_workqueue_thread(cwq, cpu);
 757         }
 758
 759         free_percpu(wq->cpu_wq);
 760         kfree(wq);
 761 }
 762 EXPORT_SYMBOL_GPL(destroy_workqueue);
 763
 764 static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
 765                                                 unsigned long action,
 766                                                 void *hcpu)
 767 {
 768         unsigned int cpu = (unsigned long)hcpu;
 769         struct cpu_workqueue_struct *cwq;
 770         struct workqueue_struct *wq;
 771
 772         switch (action) {
 773         case CPU_LOCK_ACQUIRE:
 774                 mutex_lock(&workqueue_mutex);
 775                 return NOTIFY_OK;
 776
 777         case CPU_LOCK_RELEASE:
 778                 mutex_unlock(&workqueue_mutex);
 779                 return NOTIFY_OK;
 780
 781         case CPU_UP_PREPARE:
 782                 cpu_set(cpu, cpu_populated_map);
 783         }
 784
 785         list_for_each_entry(wq, &workqueues, list) {
 786                 cwq = per_cpu_ptr(wq->cpu_wq, cpu);
 787
 788                 switch (action) {
 789                 case CPU_UP_PREPARE:
 790                         if (!create_workqueue_thread(cwq, cpu))
 791                                 break;
 792                         printk(KERN_ERR "workqueue for %i failed\n", cpu);
 793                         return NOTIFY_BAD;
 794
 795                 case CPU_ONLINE:
 796                         start_workqueue_thread(cwq, cpu);
 797                         break;
 798
 799                 case CPU_UP_CANCELED:
 800                         start_workqueue_thread(cwq, -1);
 801                 case CPU_DEAD:
 802                         cleanup_workqueue_thread(cwq, cpu);
 803                         break;
 804                 }
 805         }
 806
 807         return NOTIFY_OK;
 808 }
 809
 810 void __init init_workqueues(void)
 811 {
 812         cpu_populated_map = cpu_online_map;
 813         singlethread_cpu = first_cpu(cpu_possible_map);
 814         cpu_singlethread_map = cpumask_of_cpu(singlethread_cpu);
 815         hotcpu_notifier(workqueue_cpu_callback, 0);
 816         keventd_wq = create_workqueue("events");
 817         BUG_ON(!keventd_wq);
 818 }