Pileus Git - ~andy/linux/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
  32 /*
  33  * The per-CPU workqueue (if single thread, we always use the first
  34  * possible cpu).
  35  *
  36  * The sequence counters are for flush_scheduled_work().  It wants to wait
  37  * until until all currently-scheduled works are completed, but it doesn't
  38  * want to be livelocked by new, incoming ones.  So it waits until
  39  * remove_sequence is >= the insert_sequence which pertained when
  40  * flush_scheduled_work() was called.
  41  */
  42 struct cpu_workqueue_struct {
  43
  44         spinlock_t lock;
  45
  46         long remove_sequence;   /* Least-recently added (next to run) */
  47         long insert_sequence;   /* Next to add */
  48
  49         struct list_head worklist;
  50         wait_queue_head_t more_work;
  51         wait_queue_head_t work_done;
  52
  53         struct workqueue_struct *wq;
  54         task_t *thread;
  55
  56         int run_depth;          /* Detect run_workqueue() recursion depth */
  57 } ____cacheline_aligned;
  58
  59 /*
  60  * The externally visible workqueue abstraction is an array of
  61  * per-CPU workqueues:
  62  */
  63 struct workqueue_struct {
  64         struct cpu_workqueue_struct *cpu_wq;
  65         const char *name;
  66         struct list_head list;  /* Empty if single thread */
  67 };
  68
  69 /* All the per-cpu workqueues on the system, for hotplug cpu to add/remove
  70    threads to each one as cpus come/go. */
  71 static DEFINE_SPINLOCK(workqueue_lock);
  72 static LIST_HEAD(workqueues);
  73
  74 static int singlethread_cpu;
  75
  76 /* If it's single threaded, it isn't in the list of workqueues. */
  77 static inline int is_single_threaded(struct workqueue_struct *wq)
  78 {
  79         return list_empty(&wq->list);
  80 }
  81
  82 /* Preempt must be disabled. */
  83 static void __queue_work(struct cpu_workqueue_struct *cwq,
  84                          struct work_struct *work)
  85 {
  86         unsigned long flags;
  87
  88         spin_lock_irqsave(&cwq->lock, flags);
  89         work->wq_data = cwq;
  90         list_add_tail(&work->entry, &cwq->worklist);
  91         cwq->insert_sequence++;
  92         wake_up(&cwq->more_work);
  93         spin_unlock_irqrestore(&cwq->lock, flags);
  94 }
  95
  96 /*
  97  * Queue work on a workqueue. Return non-zero if it was successfully
  98  * added.
  99  *
 100  * We queue the work to the CPU it was submitted, but there is no
 101  * guarantee that it will be processed by that CPU.
 102  */
 103 int fastcall queue_work(struct workqueue_struct *wq, struct work_struct *work)
 104 {
 105         int ret = 0, cpu = get_cpu();
 106
 107         if (!test_and_set_bit(0, &work->pending)) {
 108                 if (unlikely(is_single_threaded(wq)))
 109                         cpu = singlethread_cpu;
 110                 BUG_ON(!list_empty(&work->entry));
 111                 __queue_work(per_cpu_ptr(wq->cpu_wq, cpu), work);
 112                 ret = 1;
 113         }
 114         put_cpu();
 115         return ret;
 116 }
 117
 118 static void delayed_work_timer_fn(unsigned long __data)
 119 {
 120         struct work_struct *work = (struct work_struct *)__data;
 121         struct workqueue_struct *wq = work->wq_data;
 122         int cpu = smp_processor_id();
 123
 124         if (unlikely(is_single_threaded(wq)))
 125                 cpu = singlethread_cpu;
 126
 127         __queue_work(per_cpu_ptr(wq->cpu_wq, cpu), work);
 128 }
 129
 130 int fastcall queue_delayed_work(struct workqueue_struct *wq,
 131                         struct work_struct *work, unsigned long delay)
 132 {
 133         int ret = 0;
 134         struct timer_list *timer = &work->timer;
 135
 136         if (!test_and_set_bit(0, &work->pending)) {
 137                 BUG_ON(timer_pending(timer));
 138                 BUG_ON(!list_empty(&work->entry));
 139
 140                 /* This stores wq for the moment, for the timer_fn */
 141                 work->wq_data = wq;
 142                 timer->expires = jiffies + delay;
 143                 timer->data = (unsigned long)work;
 144                 timer->function = delayed_work_timer_fn;
 145                 add_timer(timer);
 146                 ret = 1;
 147         }
 148         return ret;
 149 }
 150
 151 int queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
 152                         struct work_struct *work, unsigned long delay)
 153 {
 154         int ret = 0;
 155         struct timer_list *timer = &work->timer;
 156
 157         if (!test_and_set_bit(0, &work->pending)) {
 158                 BUG_ON(timer_pending(timer));
 159                 BUG_ON(!list_empty(&work->entry));
 160
 161                 /* This stores wq for the moment, for the timer_fn */
 162                 work->wq_data = wq;
 163                 timer->expires = jiffies + delay;
 164                 timer->data = (unsigned long)work;
 165                 timer->function = delayed_work_timer_fn;
 166                 add_timer_on(timer, cpu);
 167                 ret = 1;
 168         }
 169         return ret;
 170 }
 171
 172 static void run_workqueue(struct cpu_workqueue_struct *cwq)
 173 {
 174         unsigned long flags;
 175
 176         /*
 177          * Keep taking off work from the queue until
 178          * done.
 179          */
 180         spin_lock_irqsave(&cwq->lock, flags);
 181         cwq->run_depth++;
 182         if (cwq->run_depth > 3) {
 183                 /* morton gets to eat his hat */
 184                 printk("%s: recursion depth exceeded: %d\n",
 185                         __FUNCTION__, cwq->run_depth);
 186                 dump_stack();
 187         }
 188         while (!list_empty(&cwq->worklist)) {
 189                 struct work_struct *work = list_entry(cwq->worklist.next,
 190                                                 struct work_struct, entry);
 191                 void (*f) (void *) = work->func;
 192                 void *data = work->data;
 193
 194                 list_del_init(cwq->worklist.next);
 195                 spin_unlock_irqrestore(&cwq->lock, flags);
 196
 197                 BUG_ON(work->wq_data != cwq);
 198                 clear_bit(0, &work->pending);
 199                 f(data);
 200
 201                 spin_lock_irqsave(&cwq->lock, flags);
 202                 cwq->remove_sequence++;
 203                 wake_up(&cwq->work_done);
 204         }
 205         cwq->run_depth--;
 206         spin_unlock_irqrestore(&cwq->lock, flags);
 207 }
 208
 209 static int worker_thread(void *__cwq)
 210 {
 211         struct cpu_workqueue_struct *cwq = __cwq;
 212         DECLARE_WAITQUEUE(wait, current);
 213         struct k_sigaction sa;
 214         sigset_t blocked;
 215
 216         current->flags |= PF_NOFREEZE;
 217
 218         set_user_nice(current, -5);
 219
 220         /* Block and flush all signals */
 221         sigfillset(&blocked);
 222         sigprocmask(SIG_BLOCK, &blocked, NULL);
 223         flush_signals(current);
 224
 225         /* SIG_IGN makes children autoreap: see do_notify_parent(). */
 226         sa.sa.sa_handler = SIG_IGN;
 227         sa.sa.sa_flags = 0;
 228         siginitset(&sa.sa.sa_mask, sigmask(SIGCHLD));
 229         do_sigaction(SIGCHLD, &sa, (struct k_sigaction *)0);
 230
 231         set_current_state(TASK_INTERRUPTIBLE);
 232         while (!kthread_should_stop()) {
 233                 add_wait_queue(&cwq->more_work, &wait);
 234                 if (list_empty(&cwq->worklist))
 235                         schedule();
 236                 else
 237                         __set_current_state(TASK_RUNNING);
 238                 remove_wait_queue(&cwq->more_work, &wait);
 239
 240                 if (!list_empty(&cwq->worklist))
 241                         run_workqueue(cwq);
 242                 set_current_state(TASK_INTERRUPTIBLE);
 243         }
 244         __set_current_state(TASK_RUNNING);
 245         return 0;
 246 }
 247
 248 static void flush_cpu_workqueue(struct cpu_workqueue_struct *cwq)
 249 {
 250         if (cwq->thread == current) {
 251                 /*
 252                  * Probably keventd trying to flush its own queue. So simply run
 253                  * it by hand rather than deadlocking.
 254                  */
 255                 run_workqueue(cwq);
 256         } else {
 257                 DEFINE_WAIT(wait);
 258                 long sequence_needed;
 259
 260                 spin_lock_irq(&cwq->lock);
 261                 sequence_needed = cwq->insert_sequence;
 262
 263                 while (sequence_needed - cwq->remove_sequence > 0) {
 264                         prepare_to_wait(&cwq->work_done, &wait,
 265                                         TASK_UNINTERRUPTIBLE);
 266                         spin_unlock_irq(&cwq->lock);
 267                         schedule();
 268                         spin_lock_irq(&cwq->lock);
 269                 }
 270                 finish_wait(&cwq->work_done, &wait);
 271                 spin_unlock_irq(&cwq->lock);
 272         }
 273 }
 274
 275 /*
 276  * flush_workqueue - ensure that any scheduled work has run to completion.
 277  *
 278  * Forces execution of the workqueue and blocks until its completion.
 279  * This is typically used in driver shutdown handlers.
 280  *
 281  * This function will sample each workqueue's current insert_sequence number and
 282  * will sleep until the head sequence is greater than or equal to that.  This
 283  * means that we sleep until all works which were queued on entry have been
 284  * handled, but we are not livelocked by new incoming ones.
 285  *
 286  * This function used to run the workqueues itself.  Now we just wait for the
 287  * helper threads to do it.
 288  */
 289 void fastcall flush_workqueue(struct workqueue_struct *wq)
 290 {
 291         might_sleep();
 292
 293         if (is_single_threaded(wq)) {
 294                 /* Always use first cpu's area. */
 295                 flush_cpu_workqueue(per_cpu_ptr(wq->cpu_wq, singlethread_cpu));
 296         } else {
 297                 int cpu;
 298
 299                 lock_cpu_hotplug();
 300                 for_each_online_cpu(cpu)
 301                         flush_cpu_workqueue(per_cpu_ptr(wq->cpu_wq, cpu));
 302                 unlock_cpu_hotplug();
 303         }
 304 }
 305
 306 static struct task_struct *create_workqueue_thread(struct workqueue_struct *wq,
 307                                                    int cpu)
 308 {
 309         struct cpu_workqueue_struct *cwq = per_cpu_ptr(wq->cpu_wq, cpu);
 310         struct task_struct *p;
 311
 312         spin_lock_init(&cwq->lock);
 313         cwq->wq = wq;
 314         cwq->thread = NULL;
 315         cwq->insert_sequence = 0;
 316         cwq->remove_sequence = 0;
 317         INIT_LIST_HEAD(&cwq->worklist);
 318         init_waitqueue_head(&cwq->more_work);
 319         init_waitqueue_head(&cwq->work_done);
 320
 321         if (is_single_threaded(wq))
 322                 p = kthread_create(worker_thread, cwq, "%s", wq->name);
 323         else
 324                 p = kthread_create(worker_thread, cwq, "%s/%d", wq->name, cpu);
 325         if (IS_ERR(p))
 326                 return NULL;
 327         cwq->thread = p;
 328         return p;
 329 }
 330
 331 struct workqueue_struct *__create_workqueue(const char *name,
 332                                             int singlethread)
 333 {
 334         int cpu, destroy = 0;
 335         struct workqueue_struct *wq;
 336         struct task_struct *p;
 337
 338         wq = kzalloc(sizeof(*wq), GFP_KERNEL);
 339         if (!wq)
 340                 return NULL;
 341
 342         wq->cpu_wq = alloc_percpu(struct cpu_workqueue_struct);
 343         if (!wq->cpu_wq) {
 344                 kfree(wq);
 345                 return NULL;
 346         }
 347
 348         wq->name = name;
 349         /* We don't need the distraction of CPUs appearing and vanishing. */
 350         lock_cpu_hotplug();
 351         if (singlethread) {
 352                 INIT_LIST_HEAD(&wq->list);
 353                 p = create_workqueue_thread(wq, singlethread_cpu);
 354                 if (!p)
 355                         destroy = 1;
 356                 else
 357                         wake_up_process(p);
 358         } else {
 359                 spin_lock(&workqueue_lock);
 360                 list_add(&wq->list, &workqueues);
 361                 spin_unlock(&workqueue_lock);
 362                 for_each_online_cpu(cpu) {
 363                         p = create_workqueue_thread(wq, cpu);
 364                         if (p) {
 365                                 kthread_bind(p, cpu);
 366                                 wake_up_process(p);
 367                         } else
 368                                 destroy = 1;
 369                 }
 370         }
 371         unlock_cpu_hotplug();
 372
 373         /*
 374          * Was there any error during startup? If yes then clean up:
 375          */
 376         if (destroy) {
 377                 destroy_workqueue(wq);
 378                 wq = NULL;
 379         }
 380         return wq;
 381 }
 382
 383 static void cleanup_workqueue_thread(struct workqueue_struct *wq, int cpu)
 384 {
 385         struct cpu_workqueue_struct *cwq;
 386         unsigned long flags;
 387         struct task_struct *p;
 388
 389         cwq = per_cpu_ptr(wq->cpu_wq, cpu);
 390         spin_lock_irqsave(&cwq->lock, flags);
 391         p = cwq->thread;
 392         cwq->thread = NULL;
 393         spin_unlock_irqrestore(&cwq->lock, flags);
 394         if (p)
 395                 kthread_stop(p);
 396 }
 397
 398 void destroy_workqueue(struct workqueue_struct *wq)
 399 {
 400         int cpu;
 401
 402         flush_workqueue(wq);
 403
 404         /* We don't need the distraction of CPUs appearing and vanishing. */
 405         lock_cpu_hotplug();
 406         if (is_single_threaded(wq))
 407                 cleanup_workqueue_thread(wq, singlethread_cpu);
 408         else {
 409                 for_each_online_cpu(cpu)
 410                         cleanup_workqueue_thread(wq, cpu);
 411                 spin_lock(&workqueue_lock);
 412                 list_del(&wq->list);
 413                 spin_unlock(&workqueue_lock);
 414         }
 415         unlock_cpu_hotplug();
 416         free_percpu(wq->cpu_wq);
 417         kfree(wq);
 418 }
 419
 420 static struct workqueue_struct *keventd_wq;
 421
 422 int fastcall schedule_work(struct work_struct *work)
 423 {
 424         return queue_work(keventd_wq, work);
 425 }
 426
 427 int fastcall schedule_delayed_work(struct work_struct *work, unsigned long delay)
 428 {
 429         return queue_delayed_work(keventd_wq, work, delay);
 430 }
 431
 432 int schedule_delayed_work_on(int cpu,
 433                         struct work_struct *work, unsigned long delay)
 434 {
 435         return queue_delayed_work_on(cpu, keventd_wq, work, delay);
 436 }
 437
 438 /**
 439  * schedule_on_each_cpu - call a function on each online CPU from keventd
 440  * @func: the function to call
 441  * @info: a pointer to pass to func()
 442  *
 443  * Returns zero on success.
 444  * Returns -ve errno on failure.
 445  *
 446  * Appears to be racy against CPU hotplug.
 447  *
 448  * schedule_on_each_cpu() is very slow.
 449  */
 450 int schedule_on_each_cpu(void (*func)(void *info), void *info)
 451 {
 452         int cpu;
 453         struct work_struct *works;
 454
 455         works = alloc_percpu(struct work_struct);
 456         if (!works)
 457                 return -ENOMEM;
 458
 459         for_each_online_cpu(cpu) {
 460                 INIT_WORK(per_cpu_ptr(works, cpu), func, info);
 461                 __queue_work(per_cpu_ptr(keventd_wq->cpu_wq, cpu),
 462                                 per_cpu_ptr(works, cpu));
 463         }
 464         flush_workqueue(keventd_wq);
 465         free_percpu(works);
 466         return 0;
 467 }
 468
 469 void flush_scheduled_work(void)
 470 {
 471         flush_workqueue(keventd_wq);
 472 }
 473
 474 /**
 475  * cancel_rearming_delayed_workqueue - reliably kill off a delayed
 476  *                      work whose handler rearms the delayed work.
 477  * @wq:   the controlling workqueue structure
 478  * @work: the delayed work struct
 479  */
 480 void cancel_rearming_delayed_workqueue(struct workqueue_struct *wq,
 481                                        struct work_struct *work)
 482 {
 483         while (!cancel_delayed_work(work))
 484                 flush_workqueue(wq);
 485 }
 486 EXPORT_SYMBOL(cancel_rearming_delayed_workqueue);
 487
 488 /**
 489  * cancel_rearming_delayed_work - reliably kill off a delayed keventd
 490  *                      work whose handler rearms the delayed work.
 491  * @work: the delayed work struct
 492  */
 493 void cancel_rearming_delayed_work(struct work_struct *work)
 494 {
 495         cancel_rearming_delayed_workqueue(keventd_wq, work);
 496 }
 497 EXPORT_SYMBOL(cancel_rearming_delayed_work);
 498
 499 /**
 500  * execute_in_process_context - reliably execute the routine with user context
 501  * @fn:         the function to execute
 502  * @data:       data to pass to the function
 503  * @ew:         guaranteed storage for the execute work structure (must
 504  *              be available when the work executes)
 505  *
 506  * Executes the function immediately if process context is available,
 507  * otherwise schedules the function for delayed execution.
 508  *
 509  * Returns:     0 - function was executed
 510  *              1 - function was scheduled for execution
 511  */
 512 int execute_in_process_context(void (*fn)(void *data), void *data,
 513                                struct execute_work *ew)
 514 {
 515         if (!in_interrupt()) {
 516                 fn(data);
 517                 return 0;
 518         }
 519
 520         INIT_WORK(&ew->work, fn, data);
 521         schedule_work(&ew->work);
 522
 523         return 1;
 524 }
 525 EXPORT_SYMBOL_GPL(execute_in_process_context);
 526
 527 int keventd_up(void)
 528 {
 529         return keventd_wq != NULL;
 530 }
 531
 532 int current_is_keventd(void)
 533 {
 534         struct cpu_workqueue_struct *cwq;
 535         int cpu = smp_processor_id();   /* preempt-safe: keventd is per-cpu */
 536         int ret = 0;
 537
 538         BUG_ON(!keventd_wq);
 539
 540         cwq = per_cpu_ptr(keventd_wq->cpu_wq, cpu);
 541         if (current == cwq->thread)
 542                 ret = 1;
 543
 544         return ret;
 545
 546 }
 547
 548 #ifdef CONFIG_HOTPLUG_CPU
 549 /* Take the work from this (downed) CPU. */
 550 static void take_over_work(struct workqueue_struct *wq, unsigned int cpu)
 551 {
 552         struct cpu_workqueue_struct *cwq = per_cpu_ptr(wq->cpu_wq, cpu);
 553         struct list_head list;
 554         struct work_struct *work;
 555
 556         spin_lock_irq(&cwq->lock);
 557         list_replace_init(&cwq->worklist, &list);
 558
 559         while (!list_empty(&list)) {
 560                 printk("Taking work for %s\n", wq->name);
 561                 work = list_entry(list.next,struct work_struct,entry);
 562                 list_del(&work->entry);
 563                 __queue_work(per_cpu_ptr(wq->cpu_wq, smp_processor_id()), work);
 564         }
 565         spin_unlock_irq(&cwq->lock);
 566 }
 567
 568 /* We're holding the cpucontrol mutex here */
 569 static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
 570                                   unsigned long action,
 571                                   void *hcpu)
 572 {
 573         unsigned int hotcpu = (unsigned long)hcpu;
 574         struct workqueue_struct *wq;
 575
 576         switch (action) {
 577         case CPU_UP_PREPARE:
 578                 /* Create a new workqueue thread for it. */
 579                 list_for_each_entry(wq, &workqueues, list) {
 580                         if (!create_workqueue_thread(wq, hotcpu)) {
 581                                 printk("workqueue for %i failed\n", hotcpu);
 582                                 return NOTIFY_BAD;
 583                         }
 584                 }
 585                 break;
 586
 587         case CPU_ONLINE:
 588                 /* Kick off worker threads. */
 589                 list_for_each_entry(wq, &workqueues, list) {
 590                         struct cpu_workqueue_struct *cwq;
 591
 592                         cwq = per_cpu_ptr(wq->cpu_wq, hotcpu);
 593                         kthread_bind(cwq->thread, hotcpu);
 594                         wake_up_process(cwq->thread);
 595                 }
 596                 break;
 597
 598         case CPU_UP_CANCELED:
 599                 list_for_each_entry(wq, &workqueues, list) {
 600                         if (!per_cpu_ptr(wq->cpu_wq, hotcpu)->thread)
 601                                 continue;
 602                         /* Unbind so it can run. */
 603                         kthread_bind(per_cpu_ptr(wq->cpu_wq, hotcpu)->thread,
 604                                      any_online_cpu(cpu_online_map));
 605                         cleanup_workqueue_thread(wq, hotcpu);
 606                 }
 607                 break;
 608
 609         case CPU_DEAD:
 610                 list_for_each_entry(wq, &workqueues, list)
 611                         cleanup_workqueue_thread(wq, hotcpu);
 612                 list_for_each_entry(wq, &workqueues, list)
 613                         take_over_work(wq, hotcpu);
 614                 break;
 615         }
 616
 617         return NOTIFY_OK;
 618 }
 619 #endif
 620
 621 void init_workqueues(void)
 622 {
 623         singlethread_cpu = first_cpu(cpu_possible_map);
 624         hotcpu_notifier(workqueue_cpu_callback, 0);
 625         keventd_wq = create_workqueue("events");
 626         BUG_ON(!keventd_wq);
 627 }
 628
 629 EXPORT_SYMBOL_GPL(__create_workqueue);
 630 EXPORT_SYMBOL_GPL(queue_work);
 631 EXPORT_SYMBOL_GPL(queue_delayed_work);
 632 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
 633 EXPORT_SYMBOL_GPL(flush_workqueue);
 634 EXPORT_SYMBOL_GPL(destroy_workqueue);
 635
 636 EXPORT_SYMBOL(schedule_work);
 637 EXPORT_SYMBOL(schedule_delayed_work);
 638 EXPORT_SYMBOL(schedule_delayed_work_on);
 639 EXPORT_SYMBOL(flush_scheduled_work);