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         const char *name;
  62         struct list_head list;  /* Empty if single thread */
  63         int freezeable;         /* Freeze threads during suspend */
  64 };
  65
  66 /* All the per-cpu workqueues on the system, for hotplug cpu to add/remove
  67    threads to each one as cpus come/go. */
  68 static DEFINE_MUTEX(workqueue_mutex);
  69 static LIST_HEAD(workqueues);
  70
  71 static int singlethread_cpu __read_mostly;
  72 /* optimization, we could use cpu_possible_map */
  73 static cpumask_t cpu_populated_map __read_mostly;
  74
  75 /* If it's single threaded, it isn't in the list of workqueues. */
  76 static inline int is_single_threaded(struct workqueue_struct *wq)
  77 {
  78         return list_empty(&wq->list);
  79 }
  80
  81 /*
  82  * Set the workqueue on which a work item is to be run
  83  * - Must *only* be called if the pending flag is set
  84  */
  85 static inline void set_wq_data(struct work_struct *work, void *wq)
  86 {
  87         unsigned long new;
  88
  89         BUG_ON(!work_pending(work));
  90
  91         new = (unsigned long) wq | (1UL << WORK_STRUCT_PENDING);
  92         new |= WORK_STRUCT_FLAG_MASK & *work_data_bits(work);
  93         atomic_long_set(&work->data, new);
  94 }
  95
  96 static inline void *get_wq_data(struct work_struct *work)
  97 {
  98         return (void *) (atomic_long_read(&work->data) & WORK_STRUCT_WQ_DATA_MASK);
  99 }
 100
 101 static void insert_work(struct cpu_workqueue_struct *cwq,
 102                                 struct work_struct *work, int tail)
 103 {
 104         set_wq_data(work, cwq);
 105         if (tail)
 106                 list_add_tail(&work->entry, &cwq->worklist);
 107         else
 108                 list_add(&work->entry, &cwq->worklist);
 109         wake_up(&cwq->more_work);
 110 }
 111
 112 /* Preempt must be disabled. */
 113 static void __queue_work(struct cpu_workqueue_struct *cwq,
 114                          struct work_struct *work)
 115 {
 116         unsigned long flags;
 117
 118         spin_lock_irqsave(&cwq->lock, flags);
 119         insert_work(cwq, work, 1);
 120         spin_unlock_irqrestore(&cwq->lock, flags);
 121 }
 122
 123 /**
 124  * queue_work - queue work on a workqueue
 125  * @wq: workqueue to use
 126  * @work: work to queue
 127  *
 128  * Returns 0 if @work was already on a queue, non-zero otherwise.
 129  *
 130  * We queue the work to the CPU it was submitted, but there is no
 131  * guarantee that it will be processed by that CPU.
 132  */
 133 int fastcall queue_work(struct workqueue_struct *wq, struct work_struct *work)
 134 {
 135         int ret = 0, cpu = get_cpu();
 136
 137         if (!test_and_set_bit(WORK_STRUCT_PENDING, work_data_bits(work))) {
 138                 if (unlikely(is_single_threaded(wq)))
 139                         cpu = singlethread_cpu;
 140                 BUG_ON(!list_empty(&work->entry));
 141                 __queue_work(per_cpu_ptr(wq->cpu_wq, cpu), work);
 142                 ret = 1;
 143         }
 144         put_cpu();
 145         return ret;
 146 }
 147 EXPORT_SYMBOL_GPL(queue_work);
 148
 149 void delayed_work_timer_fn(unsigned long __data)
 150 {
 151         struct delayed_work *dwork = (struct delayed_work *)__data;
 152         struct workqueue_struct *wq = get_wq_data(&dwork->work);
 153         int cpu = smp_processor_id();
 154
 155         if (unlikely(is_single_threaded(wq)))
 156                 cpu = singlethread_cpu;
 157
 158         __queue_work(per_cpu_ptr(wq->cpu_wq, cpu), &dwork->work);
 159 }
 160
 161 /**
 162  * queue_delayed_work - queue work on a workqueue after delay
 163  * @wq: workqueue to use
 164  * @dwork: delayable work to queue
 165  * @delay: number of jiffies to wait before queueing
 166  *
 167  * Returns 0 if @work was already on a queue, non-zero otherwise.
 168  */
 169 int fastcall queue_delayed_work(struct workqueue_struct *wq,
 170                         struct delayed_work *dwork, unsigned long delay)
 171 {
 172         int ret = 0;
 173         struct timer_list *timer = &dwork->timer;
 174         struct work_struct *work = &dwork->work;
 175
 176         timer_stats_timer_set_start_info(timer);
 177         if (delay == 0)
 178                 return queue_work(wq, work);
 179
 180         if (!test_and_set_bit(WORK_STRUCT_PENDING, work_data_bits(work))) {
 181                 BUG_ON(timer_pending(timer));
 182                 BUG_ON(!list_empty(&work->entry));
 183
 184                 /* This stores wq for the moment, for the timer_fn */
 185                 set_wq_data(work, wq);
 186                 timer->expires = jiffies + delay;
 187                 timer->data = (unsigned long)dwork;
 188                 timer->function = delayed_work_timer_fn;
 189                 add_timer(timer);
 190                 ret = 1;
 191         }
 192         return ret;
 193 }
 194 EXPORT_SYMBOL_GPL(queue_delayed_work);
 195
 196 /**
 197  * queue_delayed_work_on - queue work on specific CPU after delay
 198  * @cpu: CPU number to execute work on
 199  * @wq: workqueue to use
 200  * @dwork: work to queue
 201  * @delay: number of jiffies to wait before queueing
 202  *
 203  * Returns 0 if @work was already on a queue, non-zero otherwise.
 204  */
 205 int queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
 206                         struct delayed_work *dwork, unsigned long delay)
 207 {
 208         int ret = 0;
 209         struct timer_list *timer = &dwork->timer;
 210         struct work_struct *work = &dwork->work;
 211
 212         if (!test_and_set_bit(WORK_STRUCT_PENDING, work_data_bits(work))) {
 213                 BUG_ON(timer_pending(timer));
 214                 BUG_ON(!list_empty(&work->entry));
 215
 216                 /* This stores wq for the moment, for the timer_fn */
 217                 set_wq_data(work, wq);
 218                 timer->expires = jiffies + delay;
 219                 timer->data = (unsigned long)dwork;
 220                 timer->function = delayed_work_timer_fn;
 221                 add_timer_on(timer, cpu);
 222                 ret = 1;
 223         }
 224         return ret;
 225 }
 226 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
 227
 228 static void run_workqueue(struct cpu_workqueue_struct *cwq)
 229 {
 230         spin_lock_irq(&cwq->lock);
 231         cwq->run_depth++;
 232         if (cwq->run_depth > 3) {
 233                 /* morton gets to eat his hat */
 234                 printk("%s: recursion depth exceeded: %d\n",
 235                         __FUNCTION__, cwq->run_depth);
 236                 dump_stack();
 237         }
 238         while (!list_empty(&cwq->worklist)) {
 239                 struct work_struct *work = list_entry(cwq->worklist.next,
 240                                                 struct work_struct, entry);
 241                 work_func_t f = work->func;
 242
 243                 cwq->current_work = work;
 244                 list_del_init(cwq->worklist.next);
 245                 spin_unlock_irq(&cwq->lock);
 246
 247                 BUG_ON(get_wq_data(work) != cwq);
 248                 if (!test_bit(WORK_STRUCT_NOAUTOREL, work_data_bits(work)))
 249                         work_release(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         might_sleep();
 397
 398         if (is_single_threaded(wq))
 399                 flush_cpu_workqueue(per_cpu_ptr(wq->cpu_wq, singlethread_cpu));
 400         else {
 401                 int cpu;
 402
 403                 for_each_cpu_mask(cpu, cpu_populated_map)
 404                         flush_cpu_workqueue(per_cpu_ptr(wq->cpu_wq, cpu));
 405         }
 406 }
 407 EXPORT_SYMBOL_GPL(flush_workqueue);
 408
 409 static void wait_on_work(struct cpu_workqueue_struct *cwq,
 410                                 struct work_struct *work)
 411 {
 412         struct wq_barrier barr;
 413         int running = 0;
 414
 415         spin_lock_irq(&cwq->lock);
 416         if (unlikely(cwq->current_work == work)) {
 417                 insert_wq_barrier(cwq, &barr, 0);
 418                 running = 1;
 419         }
 420         spin_unlock_irq(&cwq->lock);
 421
 422         if (unlikely(running))
 423                 wait_for_completion(&barr.done);
 424 }
 425
 426 /**
 427  * flush_work - block until a work_struct's callback has terminated
 428  * @wq: the workqueue on which the work is queued
 429  * @work: the work which is to be flushed
 430  *
 431  * flush_work() will attempt to cancel the work if it is queued.  If the work's
 432  * callback appears to be running, flush_work() will block until it has
 433  * completed.
 434  *
 435  * flush_work() is designed to be used when the caller is tearing down data
 436  * structures which the callback function operates upon.  It is expected that,
 437  * prior to calling flush_work(), the caller has arranged for the work to not
 438  * be requeued.
 439  */
 440 void flush_work(struct workqueue_struct *wq, struct work_struct *work)
 441 {
 442         struct cpu_workqueue_struct *cwq;
 443
 444         might_sleep();
 445
 446         cwq = get_wq_data(work);
 447         /* Was it ever queued ? */
 448         if (!cwq)
 449                 return;
 450
 451         /*
 452          * This work can't be re-queued, no need to re-check that
 453          * get_wq_data() is still the same when we take cwq->lock.
 454          */
 455         spin_lock_irq(&cwq->lock);
 456         list_del_init(&work->entry);
 457         work_release(work);
 458         spin_unlock_irq(&cwq->lock);
 459
 460         if (is_single_threaded(wq))
 461                 wait_on_work(per_cpu_ptr(wq->cpu_wq, singlethread_cpu), work);
 462         else {
 463                 int cpu;
 464
 465                 for_each_cpu_mask(cpu, cpu_populated_map)
 466                         wait_on_work(per_cpu_ptr(wq->cpu_wq, cpu), work);
 467         }
 468 }
 469 EXPORT_SYMBOL_GPL(flush_work);
 470
 471
 472 static struct workqueue_struct *keventd_wq;
 473
 474 /**
 475  * schedule_work - put work task in global workqueue
 476  * @work: job to be done
 477  *
 478  * This puts a job in the kernel-global workqueue.
 479  */
 480 int fastcall schedule_work(struct work_struct *work)
 481 {
 482         return queue_work(keventd_wq, work);
 483 }
 484 EXPORT_SYMBOL(schedule_work);
 485
 486 /**
 487  * schedule_delayed_work - put work task in global workqueue after delay
 488  * @dwork: job to be done
 489  * @delay: number of jiffies to wait or 0 for immediate execution
 490  *
 491  * After waiting for a given time this puts a job in the kernel-global
 492  * workqueue.
 493  */
 494 int fastcall schedule_delayed_work(struct delayed_work *dwork,
 495                                         unsigned long delay)
 496 {
 497         timer_stats_timer_set_start_info(&dwork->timer);
 498         return queue_delayed_work(keventd_wq, dwork, delay);
 499 }
 500 EXPORT_SYMBOL(schedule_delayed_work);
 501
 502 /**
 503  * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
 504  * @cpu: cpu to use
 505  * @dwork: job to be done
 506  * @delay: number of jiffies to wait
 507  *
 508  * After waiting for a given time this puts a job in the kernel-global
 509  * workqueue on the specified CPU.
 510  */
 511 int schedule_delayed_work_on(int cpu,
 512                         struct delayed_work *dwork, unsigned long delay)
 513 {
 514         return queue_delayed_work_on(cpu, keventd_wq, dwork, delay);
 515 }
 516 EXPORT_SYMBOL(schedule_delayed_work_on);
 517
 518 /**
 519  * schedule_on_each_cpu - call a function on each online CPU from keventd
 520  * @func: the function to call
 521  *
 522  * Returns zero on success.
 523  * Returns -ve errno on failure.
 524  *
 525  * Appears to be racy against CPU hotplug.
 526  *
 527  * schedule_on_each_cpu() is very slow.
 528  */
 529 int schedule_on_each_cpu(work_func_t func)
 530 {
 531         int cpu;
 532         struct work_struct *works;
 533
 534         works = alloc_percpu(struct work_struct);
 535         if (!works)
 536                 return -ENOMEM;
 537
 538         preempt_disable();              /* CPU hotplug */
 539         for_each_online_cpu(cpu) {
 540                 struct work_struct *work = per_cpu_ptr(works, cpu);
 541
 542                 INIT_WORK(work, func);
 543                 set_bit(WORK_STRUCT_PENDING, work_data_bits(work));
 544                 __queue_work(per_cpu_ptr(keventd_wq->cpu_wq, cpu), work);
 545         }
 546         preempt_enable();
 547         flush_workqueue(keventd_wq);
 548         free_percpu(works);
 549         return 0;
 550 }
 551
 552 void flush_scheduled_work(void)
 553 {
 554         flush_workqueue(keventd_wq);
 555 }
 556 EXPORT_SYMBOL(flush_scheduled_work);
 557
 558 void flush_work_keventd(struct work_struct *work)
 559 {
 560         flush_work(keventd_wq, work);
 561 }
 562 EXPORT_SYMBOL(flush_work_keventd);
 563
 564 /**
 565  * cancel_rearming_delayed_workqueue - reliably kill off a delayed work whose handler rearms the delayed work.
 566  * @wq:   the controlling workqueue structure
 567  * @dwork: the delayed work struct
 568  */
 569 void cancel_rearming_delayed_workqueue(struct workqueue_struct *wq,
 570                                        struct delayed_work *dwork)
 571 {
 572         /* Was it ever queued ? */
 573         if (!get_wq_data(&dwork->work))
 574                 return;
 575
 576         while (!cancel_delayed_work(dwork))
 577                 flush_workqueue(wq);
 578 }
 579 EXPORT_SYMBOL(cancel_rearming_delayed_workqueue);
 580
 581 /**
 582  * cancel_rearming_delayed_work - reliably kill off a delayed keventd work whose handler rearms the delayed work.
 583  * @dwork: the delayed work struct
 584  */
 585 void cancel_rearming_delayed_work(struct delayed_work *dwork)
 586 {
 587         cancel_rearming_delayed_workqueue(keventd_wq, dwork);
 588 }
 589 EXPORT_SYMBOL(cancel_rearming_delayed_work);
 590
 591 /**
 592  * execute_in_process_context - reliably execute the routine with user context
 593  * @fn:         the function to execute
 594  * @ew:         guaranteed storage for the execute work structure (must
 595  *              be available when the work executes)
 596  *
 597  * Executes the function immediately if process context is available,
 598  * otherwise schedules the function for delayed execution.
 599  *
 600  * Returns:     0 - function was executed
 601  *              1 - function was scheduled for execution
 602  */
 603 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
 604 {
 605         if (!in_interrupt()) {
 606                 fn(&ew->work);
 607                 return 0;
 608         }
 609
 610         INIT_WORK(&ew->work, fn);
 611         schedule_work(&ew->work);
 612
 613         return 1;
 614 }
 615 EXPORT_SYMBOL_GPL(execute_in_process_context);
 616
 617 int keventd_up(void)
 618 {
 619         return keventd_wq != NULL;
 620 }
 621
 622 int current_is_keventd(void)
 623 {
 624         struct cpu_workqueue_struct *cwq;
 625         int cpu = smp_processor_id();   /* preempt-safe: keventd is per-cpu */
 626         int ret = 0;
 627
 628         BUG_ON(!keventd_wq);
 629
 630         cwq = per_cpu_ptr(keventd_wq->cpu_wq, cpu);
 631         if (current == cwq->thread)
 632                 ret = 1;
 633
 634         return ret;
 635
 636 }
 637
 638 static struct cpu_workqueue_struct *
 639 init_cpu_workqueue(struct workqueue_struct *wq, int cpu)
 640 {
 641         struct cpu_workqueue_struct *cwq = per_cpu_ptr(wq->cpu_wq, cpu);
 642
 643         cwq->wq = wq;
 644         spin_lock_init(&cwq->lock);
 645         INIT_LIST_HEAD(&cwq->worklist);
 646         init_waitqueue_head(&cwq->more_work);
 647
 648         return cwq;
 649 }
 650
 651 static int create_workqueue_thread(struct cpu_workqueue_struct *cwq, int cpu)
 652 {
 653         struct workqueue_struct *wq = cwq->wq;
 654         const char *fmt = is_single_threaded(wq) ? "%s" : "%s/%d";
 655         struct task_struct *p;
 656
 657         p = kthread_create(worker_thread, cwq, fmt, wq->name, cpu);
 658         /*
 659          * Nobody can add the work_struct to this cwq,
 660          *      if (caller is __create_workqueue)
 661          *              nobody should see this wq
 662          *      else // caller is CPU_UP_PREPARE
 663          *              cpu is not on cpu_online_map
 664          * so we can abort safely.
 665          */
 666         if (IS_ERR(p))
 667                 return PTR_ERR(p);
 668
 669         cwq->thread = p;
 670         cwq->should_stop = 0;
 671         if (!is_single_threaded(wq))
 672                 kthread_bind(p, cpu);
 673
 674         if (is_single_threaded(wq) || cpu_online(cpu))
 675                 wake_up_process(p);
 676
 677         return 0;
 678 }
 679
 680 struct workqueue_struct *__create_workqueue(const char *name,
 681                                             int singlethread, int freezeable)
 682 {
 683         struct workqueue_struct *wq;
 684         struct cpu_workqueue_struct *cwq;
 685         int err = 0, cpu;
 686
 687         wq = kzalloc(sizeof(*wq), GFP_KERNEL);
 688         if (!wq)
 689                 return NULL;
 690
 691         wq->cpu_wq = alloc_percpu(struct cpu_workqueue_struct);
 692         if (!wq->cpu_wq) {
 693                 kfree(wq);
 694                 return NULL;
 695         }
 696
 697         wq->name = name;
 698         wq->freezeable = freezeable;
 699
 700         if (singlethread) {
 701                 INIT_LIST_HEAD(&wq->list);
 702                 cwq = init_cpu_workqueue(wq, singlethread_cpu);
 703                 err = create_workqueue_thread(cwq, singlethread_cpu);
 704         } else {
 705                 mutex_lock(&workqueue_mutex);
 706                 list_add(&wq->list, &workqueues);
 707
 708                 for_each_possible_cpu(cpu) {
 709                         cwq = init_cpu_workqueue(wq, cpu);
 710                         if (err || !cpu_online(cpu))
 711                                 continue;
 712                         err = create_workqueue_thread(cwq, cpu);
 713                 }
 714                 mutex_unlock(&workqueue_mutex);
 715         }
 716
 717         if (err) {
 718                 destroy_workqueue(wq);
 719                 wq = NULL;
 720         }
 721         return wq;
 722 }
 723 EXPORT_SYMBOL_GPL(__create_workqueue);
 724
 725 static void cleanup_workqueue_thread(struct cpu_workqueue_struct *cwq, int cpu)
 726 {
 727         struct wq_barrier barr;
 728         int alive = 0;
 729
 730         spin_lock_irq(&cwq->lock);
 731         if (cwq->thread != NULL) {
 732                 insert_wq_barrier(cwq, &barr, 1);
 733                 cwq->should_stop = 1;
 734                 alive = 1;
 735         }
 736         spin_unlock_irq(&cwq->lock);
 737
 738         if (alive) {
 739                 wait_for_completion(&barr.done);
 740
 741                 while (unlikely(cwq->thread != NULL))
 742                         cpu_relax();
 743                 /*
 744                  * Wait until cwq->thread unlocks cwq->lock,
 745                  * it won't touch *cwq after that.
 746                  */
 747                 smp_rmb();
 748                 spin_unlock_wait(&cwq->lock);
 749         }
 750 }
 751
 752 /**
 753  * destroy_workqueue - safely terminate a workqueue
 754  * @wq: target workqueue
 755  *
 756  * Safely destroy a workqueue. All work currently pending will be done first.
 757  */
 758 void destroy_workqueue(struct workqueue_struct *wq)
 759 {
 760         struct cpu_workqueue_struct *cwq;
 761
 762         if (is_single_threaded(wq)) {
 763                 cwq = per_cpu_ptr(wq->cpu_wq, singlethread_cpu);
 764                 cleanup_workqueue_thread(cwq, singlethread_cpu);
 765         } else {
 766                 int cpu;
 767
 768                 mutex_lock(&workqueue_mutex);
 769                 list_del(&wq->list);
 770                 mutex_unlock(&workqueue_mutex);
 771
 772                 for_each_cpu_mask(cpu, cpu_populated_map) {
 773                         cwq = per_cpu_ptr(wq->cpu_wq, cpu);
 774                         cleanup_workqueue_thread(cwq, cpu);
 775                 }
 776         }
 777
 778         free_percpu(wq->cpu_wq);
 779         kfree(wq);
 780 }
 781 EXPORT_SYMBOL_GPL(destroy_workqueue);
 782
 783 static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
 784                                                 unsigned long action,
 785                                                 void *hcpu)
 786 {
 787         unsigned int cpu = (unsigned long)hcpu;
 788         struct cpu_workqueue_struct *cwq;
 789         struct workqueue_struct *wq;
 790
 791         switch (action) {
 792         case CPU_LOCK_ACQUIRE:
 793                 mutex_lock(&workqueue_mutex);
 794                 return NOTIFY_OK;
 795
 796         case CPU_LOCK_RELEASE:
 797                 mutex_unlock(&workqueue_mutex);
 798                 return NOTIFY_OK;
 799
 800         case CPU_UP_PREPARE:
 801                 cpu_set(cpu, cpu_populated_map);
 802         }
 803
 804         list_for_each_entry(wq, &workqueues, list) {
 805                 cwq = per_cpu_ptr(wq->cpu_wq, cpu);
 806
 807                 switch (action) {
 808                 case CPU_UP_PREPARE:
 809                         if (!create_workqueue_thread(cwq, cpu))
 810                                 break;
 811                         printk(KERN_ERR "workqueue for %i failed\n", cpu);
 812                         return NOTIFY_BAD;
 813
 814                 case CPU_ONLINE:
 815                         wake_up_process(cwq->thread);
 816                         break;
 817
 818                 case CPU_UP_CANCELED:
 819                         if (cwq->thread)
 820                                 wake_up_process(cwq->thread);
 821                 case CPU_DEAD:
 822                         cleanup_workqueue_thread(cwq, cpu);
 823                         break;
 824                 }
 825         }
 826
 827         return NOTIFY_OK;
 828 }
 829
 830 void init_workqueues(void)
 831 {
 832         cpu_populated_map = cpu_online_map;
 833         singlethread_cpu = first_cpu(cpu_possible_map);
 834         hotcpu_notifier(workqueue_cpu_callback, 0);
 835         keventd_wq = create_workqueue("events");
 836         BUG_ON(!keventd_wq);
 837 }