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 static void run_workqueue(struct cpu_workqueue_struct *cwq)
 152 {
 153         unsigned long flags;
 154
 155         /*
 156          * Keep taking off work from the queue until
 157          * done.
 158          */
 159         spin_lock_irqsave(&cwq->lock, flags);
 160         cwq->run_depth++;
 161         if (cwq->run_depth > 3) {
 162                 /* morton gets to eat his hat */
 163                 printk("%s: recursion depth exceeded: %d\n",
 164                         __FUNCTION__, cwq->run_depth);
 165                 dump_stack();
 166         }
 167         while (!list_empty(&cwq->worklist)) {
 168                 struct work_struct *work = list_entry(cwq->worklist.next,
 169                                                 struct work_struct, entry);
 170                 void (*f) (void *) = work->func;
 171                 void *data = work->data;
 172
 173                 list_del_init(cwq->worklist.next);
 174                 spin_unlock_irqrestore(&cwq->lock, flags);
 175
 176                 BUG_ON(work->wq_data != cwq);
 177                 clear_bit(0, &work->pending);
 178                 f(data);
 179
 180                 spin_lock_irqsave(&cwq->lock, flags);
 181                 cwq->remove_sequence++;
 182                 wake_up(&cwq->work_done);
 183         }
 184         cwq->run_depth--;
 185         spin_unlock_irqrestore(&cwq->lock, flags);
 186 }
 187
 188 static int worker_thread(void *__cwq)
 189 {
 190         struct cpu_workqueue_struct *cwq = __cwq;
 191         DECLARE_WAITQUEUE(wait, current);
 192         struct k_sigaction sa;
 193         sigset_t blocked;
 194
 195         current->flags |= PF_NOFREEZE;
 196
 197         set_user_nice(current, -5);
 198
 199         /* Block and flush all signals */
 200         sigfillset(&blocked);
 201         sigprocmask(SIG_BLOCK, &blocked, NULL);
 202         flush_signals(current);
 203
 204         /* SIG_IGN makes children autoreap: see do_notify_parent(). */
 205         sa.sa.sa_handler = SIG_IGN;
 206         sa.sa.sa_flags = 0;
 207         siginitset(&sa.sa.sa_mask, sigmask(SIGCHLD));
 208         do_sigaction(SIGCHLD, &sa, (struct k_sigaction *)0);
 209
 210         set_current_state(TASK_INTERRUPTIBLE);
 211         while (!kthread_should_stop()) {
 212                 add_wait_queue(&cwq->more_work, &wait);
 213                 if (list_empty(&cwq->worklist))
 214                         schedule();
 215                 else
 216                         __set_current_state(TASK_RUNNING);
 217                 remove_wait_queue(&cwq->more_work, &wait);
 218
 219                 if (!list_empty(&cwq->worklist))
 220                         run_workqueue(cwq);
 221                 set_current_state(TASK_INTERRUPTIBLE);
 222         }
 223         __set_current_state(TASK_RUNNING);
 224         return 0;
 225 }
 226
 227 static void flush_cpu_workqueue(struct cpu_workqueue_struct *cwq)
 228 {
 229         if (cwq->thread == current) {
 230                 /*
 231                  * Probably keventd trying to flush its own queue. So simply run
 232                  * it by hand rather than deadlocking.
 233                  */
 234                 run_workqueue(cwq);
 235         } else {
 236                 DEFINE_WAIT(wait);
 237                 long sequence_needed;
 238
 239                 spin_lock_irq(&cwq->lock);
 240                 sequence_needed = cwq->insert_sequence;
 241
 242                 while (sequence_needed - cwq->remove_sequence > 0) {
 243                         prepare_to_wait(&cwq->work_done, &wait,
 244                                         TASK_UNINTERRUPTIBLE);
 245                         spin_unlock_irq(&cwq->lock);
 246                         schedule();
 247                         spin_lock_irq(&cwq->lock);
 248                 }
 249                 finish_wait(&cwq->work_done, &wait);
 250                 spin_unlock_irq(&cwq->lock);
 251         }
 252 }
 253
 254 /*
 255  * flush_workqueue - ensure that any scheduled work has run to completion.
 256  *
 257  * Forces execution of the workqueue and blocks until its completion.
 258  * This is typically used in driver shutdown handlers.
 259  *
 260  * This function will sample each workqueue's current insert_sequence number and
 261  * will sleep until the head sequence is greater than or equal to that.  This
 262  * means that we sleep until all works which were queued on entry have been
 263  * handled, but we are not livelocked by new incoming ones.
 264  *
 265  * This function used to run the workqueues itself.  Now we just wait for the
 266  * helper threads to do it.
 267  */
 268 void fastcall flush_workqueue(struct workqueue_struct *wq)
 269 {
 270         might_sleep();
 271
 272         if (is_single_threaded(wq)) {
 273                 /* Always use first cpu's area. */
 274                 flush_cpu_workqueue(per_cpu_ptr(wq->cpu_wq, singlethread_cpu));
 275         } else {
 276                 int cpu;
 277
 278                 lock_cpu_hotplug();
 279                 for_each_online_cpu(cpu)
 280                         flush_cpu_workqueue(per_cpu_ptr(wq->cpu_wq, cpu));
 281                 unlock_cpu_hotplug();
 282         }
 283 }
 284
 285 static struct task_struct *create_workqueue_thread(struct workqueue_struct *wq,
 286                                                    int cpu)
 287 {
 288         struct cpu_workqueue_struct *cwq = per_cpu_ptr(wq->cpu_wq, cpu);
 289         struct task_struct *p;
 290
 291         spin_lock_init(&cwq->lock);
 292         cwq->wq = wq;
 293         cwq->thread = NULL;
 294         cwq->insert_sequence = 0;
 295         cwq->remove_sequence = 0;
 296         INIT_LIST_HEAD(&cwq->worklist);
 297         init_waitqueue_head(&cwq->more_work);
 298         init_waitqueue_head(&cwq->work_done);
 299
 300         if (is_single_threaded(wq))
 301                 p = kthread_create(worker_thread, cwq, "%s", wq->name);
 302         else
 303                 p = kthread_create(worker_thread, cwq, "%s/%d", wq->name, cpu);
 304         if (IS_ERR(p))
 305                 return NULL;
 306         cwq->thread = p;
 307         return p;
 308 }
 309
 310 struct workqueue_struct *__create_workqueue(const char *name,
 311                                             int singlethread)
 312 {
 313         int cpu, destroy = 0;
 314         struct workqueue_struct *wq;
 315         struct task_struct *p;
 316
 317         wq = kzalloc(sizeof(*wq), GFP_KERNEL);
 318         if (!wq)
 319                 return NULL;
 320
 321         wq->cpu_wq = alloc_percpu(struct cpu_workqueue_struct);
 322         if (!wq->cpu_wq) {
 323                 kfree(wq);
 324                 return NULL;
 325         }
 326
 327         wq->name = name;
 328         /* We don't need the distraction of CPUs appearing and vanishing. */
 329         lock_cpu_hotplug();
 330         if (singlethread) {
 331                 INIT_LIST_HEAD(&wq->list);
 332                 p = create_workqueue_thread(wq, singlethread_cpu);
 333                 if (!p)
 334                         destroy = 1;
 335                 else
 336                         wake_up_process(p);
 337         } else {
 338                 spin_lock(&workqueue_lock);
 339                 list_add(&wq->list, &workqueues);
 340                 spin_unlock(&workqueue_lock);
 341                 for_each_online_cpu(cpu) {
 342                         p = create_workqueue_thread(wq, cpu);
 343                         if (p) {
 344                                 kthread_bind(p, cpu);
 345                                 wake_up_process(p);
 346                         } else
 347                                 destroy = 1;
 348                 }
 349         }
 350         unlock_cpu_hotplug();
 351
 352         /*
 353          * Was there any error during startup? If yes then clean up:
 354          */
 355         if (destroy) {
 356                 destroy_workqueue(wq);
 357                 wq = NULL;
 358         }
 359         return wq;
 360 }
 361
 362 static void cleanup_workqueue_thread(struct workqueue_struct *wq, int cpu)
 363 {
 364         struct cpu_workqueue_struct *cwq;
 365         unsigned long flags;
 366         struct task_struct *p;
 367
 368         cwq = per_cpu_ptr(wq->cpu_wq, cpu);
 369         spin_lock_irqsave(&cwq->lock, flags);
 370         p = cwq->thread;
 371         cwq->thread = NULL;
 372         spin_unlock_irqrestore(&cwq->lock, flags);
 373         if (p)
 374                 kthread_stop(p);
 375 }
 376
 377 void destroy_workqueue(struct workqueue_struct *wq)
 378 {
 379         int cpu;
 380
 381         flush_workqueue(wq);
 382
 383         /* We don't need the distraction of CPUs appearing and vanishing. */
 384         lock_cpu_hotplug();
 385         if (is_single_threaded(wq))
 386                 cleanup_workqueue_thread(wq, singlethread_cpu);
 387         else {
 388                 for_each_online_cpu(cpu)
 389                         cleanup_workqueue_thread(wq, cpu);
 390                 spin_lock(&workqueue_lock);
 391                 list_del(&wq->list);
 392                 spin_unlock(&workqueue_lock);
 393         }
 394         unlock_cpu_hotplug();
 395         free_percpu(wq->cpu_wq);
 396         kfree(wq);
 397 }
 398
 399 static struct workqueue_struct *keventd_wq;
 400
 401 int fastcall schedule_work(struct work_struct *work)
 402 {
 403         return queue_work(keventd_wq, work);
 404 }
 405
 406 int fastcall schedule_delayed_work(struct work_struct *work, unsigned long delay)
 407 {
 408         return queue_delayed_work(keventd_wq, work, delay);
 409 }
 410
 411 int schedule_delayed_work_on(int cpu,
 412                         struct work_struct *work, unsigned long delay)
 413 {
 414         int ret = 0;
 415         struct timer_list *timer = &work->timer;
 416
 417         if (!test_and_set_bit(0, &work->pending)) {
 418                 BUG_ON(timer_pending(timer));
 419                 BUG_ON(!list_empty(&work->entry));
 420                 /* This stores keventd_wq for the moment, for the timer_fn */
 421                 work->wq_data = keventd_wq;
 422                 timer->expires = jiffies + delay;
 423                 timer->data = (unsigned long)work;
 424                 timer->function = delayed_work_timer_fn;
 425                 add_timer_on(timer, cpu);
 426                 ret = 1;
 427         }
 428         return ret;
 429 }
 430
 431 int schedule_on_each_cpu(void (*func) (void *info), void *info)
 432 {
 433         int cpu;
 434         struct work_struct *work;
 435
 436         work = kmalloc(NR_CPUS * sizeof(struct work_struct), GFP_KERNEL);
 437
 438         if (!work)
 439                 return -ENOMEM;
 440         for_each_online_cpu(cpu) {
 441                 INIT_WORK(work + cpu, func, info);
 442                 __queue_work(per_cpu_ptr(keventd_wq->cpu_wq, cpu),
 443                                 work + cpu);
 444         }
 445         flush_workqueue(keventd_wq);
 446         kfree(work);
 447         return 0;
 448 }
 449
 450 void flush_scheduled_work(void)
 451 {
 452         flush_workqueue(keventd_wq);
 453 }
 454
 455 /**
 456  * cancel_rearming_delayed_workqueue - reliably kill off a delayed
 457  *                      work whose handler rearms the delayed work.
 458  * @wq:   the controlling workqueue structure
 459  * @work: the delayed work struct
 460  */
 461 void cancel_rearming_delayed_workqueue(struct workqueue_struct *wq,
 462                                        struct work_struct *work)
 463 {
 464         while (!cancel_delayed_work(work))
 465                 flush_workqueue(wq);
 466 }
 467 EXPORT_SYMBOL(cancel_rearming_delayed_workqueue);
 468
 469 /**
 470  * cancel_rearming_delayed_work - reliably kill off a delayed keventd
 471  *                      work whose handler rearms the delayed work.
 472  * @work: the delayed work struct
 473  */
 474 void cancel_rearming_delayed_work(struct work_struct *work)
 475 {
 476         cancel_rearming_delayed_workqueue(keventd_wq, work);
 477 }
 478 EXPORT_SYMBOL(cancel_rearming_delayed_work);
 479
 480 /**
 481  * execute_in_process_context - reliably execute the routine with user context
 482  * @fn:         the function to execute
 483  * @data:       data to pass to the function
 484  * @ew:         guaranteed storage for the execute work structure (must
 485  *              be available when the work executes)
 486  *
 487  * Executes the function immediately if process context is available,
 488  * otherwise schedules the function for delayed execution.
 489  *
 490  * Returns:     0 - function was executed
 491  *              1 - function was scheduled for execution
 492  */
 493 int execute_in_process_context(void (*fn)(void *data), void *data,
 494                                struct execute_work *ew)
 495 {
 496         if (!in_interrupt()) {
 497                 fn(data);
 498                 return 0;
 499         }
 500
 501         INIT_WORK(&ew->work, fn, data);
 502         schedule_work(&ew->work);
 503
 504         return 1;
 505 }
 506 EXPORT_SYMBOL_GPL(execute_in_process_context);
 507
 508 int keventd_up(void)
 509 {
 510         return keventd_wq != NULL;
 511 }
 512
 513 int current_is_keventd(void)
 514 {
 515         struct cpu_workqueue_struct *cwq;
 516         int cpu = smp_processor_id();   /* preempt-safe: keventd is per-cpu */
 517         int ret = 0;
 518
 519         BUG_ON(!keventd_wq);
 520
 521         cwq = per_cpu_ptr(keventd_wq->cpu_wq, cpu);
 522         if (current == cwq->thread)
 523                 ret = 1;
 524
 525         return ret;
 526
 527 }
 528
 529 #ifdef CONFIG_HOTPLUG_CPU
 530 /* Take the work from this (downed) CPU. */
 531 static void take_over_work(struct workqueue_struct *wq, unsigned int cpu)
 532 {
 533         struct cpu_workqueue_struct *cwq = per_cpu_ptr(wq->cpu_wq, cpu);
 534         LIST_HEAD(list);
 535         struct work_struct *work;
 536
 537         spin_lock_irq(&cwq->lock);
 538         list_splice_init(&cwq->worklist, &list);
 539
 540         while (!list_empty(&list)) {
 541                 printk("Taking work for %s\n", wq->name);
 542                 work = list_entry(list.next,struct work_struct,entry);
 543                 list_del(&work->entry);
 544                 __queue_work(per_cpu_ptr(wq->cpu_wq, smp_processor_id()), work);
 545         }
 546         spin_unlock_irq(&cwq->lock);
 547 }
 548
 549 /* We're holding the cpucontrol mutex here */
 550 static int workqueue_cpu_callback(struct notifier_block *nfb,
 551                                   unsigned long action,
 552                                   void *hcpu)
 553 {
 554         unsigned int hotcpu = (unsigned long)hcpu;
 555         struct workqueue_struct *wq;
 556
 557         switch (action) {
 558         case CPU_UP_PREPARE:
 559                 /* Create a new workqueue thread for it. */
 560                 list_for_each_entry(wq, &workqueues, list) {
 561                         if (!create_workqueue_thread(wq, hotcpu)) {
 562                                 printk("workqueue for %i failed\n", hotcpu);
 563                                 return NOTIFY_BAD;
 564                         }
 565                 }
 566                 break;
 567
 568         case CPU_ONLINE:
 569                 /* Kick off worker threads. */
 570                 list_for_each_entry(wq, &workqueues, list) {
 571                         struct cpu_workqueue_struct *cwq;
 572
 573                         cwq = per_cpu_ptr(wq->cpu_wq, hotcpu);
 574                         kthread_bind(cwq->thread, hotcpu);
 575                         wake_up_process(cwq->thread);
 576                 }
 577                 break;
 578
 579         case CPU_UP_CANCELED:
 580                 list_for_each_entry(wq, &workqueues, list) {
 581                         /* Unbind so it can run. */
 582                         kthread_bind(per_cpu_ptr(wq->cpu_wq, hotcpu)->thread,
 583                                      any_online_cpu(cpu_online_map));
 584                         cleanup_workqueue_thread(wq, hotcpu);
 585                 }
 586                 break;
 587
 588         case CPU_DEAD:
 589                 list_for_each_entry(wq, &workqueues, list)
 590                         cleanup_workqueue_thread(wq, hotcpu);
 591                 list_for_each_entry(wq, &workqueues, list)
 592                         take_over_work(wq, hotcpu);
 593                 break;
 594         }
 595
 596         return NOTIFY_OK;
 597 }
 598 #endif
 599
 600 void init_workqueues(void)
 601 {
 602         singlethread_cpu = first_cpu(cpu_possible_map);
 603         hotcpu_notifier(workqueue_cpu_callback, 0);
 604         keventd_wq = create_workqueue("events");
 605         BUG_ON(!keventd_wq);
 606 }
 607
 608 EXPORT_SYMBOL_GPL(__create_workqueue);
 609 EXPORT_SYMBOL_GPL(queue_work);
 610 EXPORT_SYMBOL_GPL(queue_delayed_work);
 611 EXPORT_SYMBOL_GPL(flush_workqueue);
 612 EXPORT_SYMBOL_GPL(destroy_workqueue);
 613
 614 EXPORT_SYMBOL(schedule_work);
 615 EXPORT_SYMBOL(schedule_delayed_work);
 616 EXPORT_SYMBOL(schedule_delayed_work_on);
 617 EXPORT_SYMBOL(flush_scheduled_work);