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
  31 /*
  32  * The per-CPU workqueue (if single thread, we always use the first
  33  * possible cpu).
  34  *
  35  * The sequence counters are for flush_scheduled_work().  It wants to wait
  36  * until until all currently-scheduled works are completed, but it doesn't
  37  * want to be livelocked by new, incoming ones.  So it waits until
  38  * remove_sequence is >= the insert_sequence which pertained when
  39  * flush_scheduled_work() was called.
  40  */
  41 struct cpu_workqueue_struct {
  42
  43         spinlock_t lock;
  44
  45         long remove_sequence;   /* Least-recently added (next to run) */
  46         long insert_sequence;   /* Next to add */
  47
  48         struct list_head worklist;
  49         wait_queue_head_t more_work;
  50         wait_queue_head_t work_done;
  51
  52         struct workqueue_struct *wq;
  53         task_t *thread;
  54
  55         int run_depth;          /* Detect run_workqueue() recursion depth */
  56 } ____cacheline_aligned;
  57
  58 /*
  59  * The externally visible workqueue abstraction is an array of
  60  * per-CPU workqueues:
  61  */
  62 struct workqueue_struct {
  63         struct cpu_workqueue_struct *cpu_wq;
  64         const char *name;
  65         struct list_head list;  /* Empty if single thread */
  66 };
  67
  68 /* All the per-cpu workqueues on the system, for hotplug cpu to add/remove
  69    threads to each one as cpus come/go. */
  70 static DEFINE_SPINLOCK(workqueue_lock);
  71 static LIST_HEAD(workqueues);
  72
  73 static int singlethread_cpu;
  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 /* Preempt must be disabled. */
  82 static void __queue_work(struct cpu_workqueue_struct *cwq,
  83                          struct work_struct *work)
  84 {
  85         unsigned long flags;
  86
  87         spin_lock_irqsave(&cwq->lock, flags);
  88         work->wq_data = cwq;
  89         list_add_tail(&work->entry, &cwq->worklist);
  90         cwq->insert_sequence++;
  91         wake_up(&cwq->more_work);
  92         spin_unlock_irqrestore(&cwq->lock, flags);
  93 }
  94
  95 /*
  96  * Queue work on a workqueue. Return non-zero if it was successfully
  97  * added.
  98  *
  99  * We queue the work to the CPU it was submitted, but there is no
 100  * guarantee that it will be processed by that CPU.
 101  */
 102 int fastcall queue_work(struct workqueue_struct *wq, struct work_struct *work)
 103 {
 104         int ret = 0, cpu = get_cpu();
 105
 106         if (!test_and_set_bit(0, &work->pending)) {
 107                 if (unlikely(is_single_threaded(wq)))
 108                         cpu = singlethread_cpu;
 109                 BUG_ON(!list_empty(&work->entry));
 110                 __queue_work(per_cpu_ptr(wq->cpu_wq, cpu), work);
 111                 ret = 1;
 112         }
 113         put_cpu();
 114         return ret;
 115 }
 116
 117 static void delayed_work_timer_fn(unsigned long __data)
 118 {
 119         struct work_struct *work = (struct work_struct *)__data;
 120         struct workqueue_struct *wq = work->wq_data;
 121         int cpu = smp_processor_id();
 122
 123         if (unlikely(is_single_threaded(wq)))
 124                 cpu = singlethread_cpu;
 125
 126         __queue_work(per_cpu_ptr(wq->cpu_wq, cpu), work);
 127 }
 128
 129 int fastcall queue_delayed_work(struct workqueue_struct *wq,
 130                         struct work_struct *work, unsigned long delay)
 131 {
 132         int ret = 0;
 133         struct timer_list *timer = &work->timer;
 134
 135         if (!test_and_set_bit(0, &work->pending)) {
 136                 BUG_ON(timer_pending(timer));
 137                 BUG_ON(!list_empty(&work->entry));
 138
 139                 /* This stores wq for the moment, for the timer_fn */
 140                 work->wq_data = wq;
 141                 timer->expires = jiffies + delay;
 142                 timer->data = (unsigned long)work;
 143                 timer->function = delayed_work_timer_fn;
 144                 add_timer(timer);
 145                 ret = 1;
 146         }
 147         return ret;
 148 }
 149
 150 static void run_workqueue(struct cpu_workqueue_struct *cwq)
 151 {
 152         unsigned long flags;
 153
 154         /*
 155          * Keep taking off work from the queue until
 156          * done.
 157          */
 158         spin_lock_irqsave(&cwq->lock, flags);
 159         cwq->run_depth++;
 160         if (cwq->run_depth > 3) {
 161                 /* morton gets to eat his hat */
 162                 printk("%s: recursion depth exceeded: %d\n",
 163                         __FUNCTION__, cwq->run_depth);
 164                 dump_stack();
 165         }
 166         while (!list_empty(&cwq->worklist)) {
 167                 struct work_struct *work = list_entry(cwq->worklist.next,
 168                                                 struct work_struct, entry);
 169                 void (*f) (void *) = work->func;
 170                 void *data = work->data;
 171
 172                 list_del_init(cwq->worklist.next);
 173                 spin_unlock_irqrestore(&cwq->lock, flags);
 174
 175                 BUG_ON(work->wq_data != cwq);
 176                 clear_bit(0, &work->pending);
 177                 f(data);
 178
 179                 spin_lock_irqsave(&cwq->lock, flags);
 180                 cwq->remove_sequence++;
 181                 wake_up(&cwq->work_done);
 182         }
 183         cwq->run_depth--;
 184         spin_unlock_irqrestore(&cwq->lock, flags);
 185 }
 186
 187 static int worker_thread(void *__cwq)
 188 {
 189         struct cpu_workqueue_struct *cwq = __cwq;
 190         DECLARE_WAITQUEUE(wait, current);
 191         struct k_sigaction sa;
 192         sigset_t blocked;
 193
 194         current->flags |= PF_NOFREEZE;
 195
 196         set_user_nice(current, -5);
 197
 198         /* Block and flush all signals */
 199         sigfillset(&blocked);
 200         sigprocmask(SIG_BLOCK, &blocked, NULL);
 201         flush_signals(current);
 202
 203         /* SIG_IGN makes children autoreap: see do_notify_parent(). */
 204         sa.sa.sa_handler = SIG_IGN;
 205         sa.sa.sa_flags = 0;
 206         siginitset(&sa.sa.sa_mask, sigmask(SIGCHLD));
 207         do_sigaction(SIGCHLD, &sa, (struct k_sigaction *)0);
 208
 209         set_current_state(TASK_INTERRUPTIBLE);
 210         while (!kthread_should_stop()) {
 211                 add_wait_queue(&cwq->more_work, &wait);
 212                 if (list_empty(&cwq->worklist))
 213                         schedule();
 214                 else
 215                         __set_current_state(TASK_RUNNING);
 216                 remove_wait_queue(&cwq->more_work, &wait);
 217
 218                 if (!list_empty(&cwq->worklist))
 219                         run_workqueue(cwq);
 220                 set_current_state(TASK_INTERRUPTIBLE);
 221         }
 222         __set_current_state(TASK_RUNNING);
 223         return 0;
 224 }
 225
 226 static void flush_cpu_workqueue(struct cpu_workqueue_struct *cwq)
 227 {
 228         if (cwq->thread == current) {
 229                 /*
 230                  * Probably keventd trying to flush its own queue. So simply run
 231                  * it by hand rather than deadlocking.
 232                  */
 233                 run_workqueue(cwq);
 234         } else {
 235                 DEFINE_WAIT(wait);
 236                 long sequence_needed;
 237
 238                 spin_lock_irq(&cwq->lock);
 239                 sequence_needed = cwq->insert_sequence;
 240
 241                 while (sequence_needed - cwq->remove_sequence > 0) {
 242                         prepare_to_wait(&cwq->work_done, &wait,
 243                                         TASK_UNINTERRUPTIBLE);
 244                         spin_unlock_irq(&cwq->lock);
 245                         schedule();
 246                         spin_lock_irq(&cwq->lock);
 247                 }
 248                 finish_wait(&cwq->work_done, &wait);
 249                 spin_unlock_irq(&cwq->lock);
 250         }
 251 }
 252
 253 /*
 254  * flush_workqueue - ensure that any scheduled work has run to completion.
 255  *
 256  * Forces execution of the workqueue and blocks until its completion.
 257  * This is typically used in driver shutdown handlers.
 258  *
 259  * This function will sample each workqueue's current insert_sequence number and
 260  * will sleep until the head sequence is greater than or equal to that.  This
 261  * means that we sleep until all works which were queued on entry have been
 262  * handled, but we are not livelocked by new incoming ones.
 263  *
 264  * This function used to run the workqueues itself.  Now we just wait for the
 265  * helper threads to do it.
 266  */
 267 void fastcall flush_workqueue(struct workqueue_struct *wq)
 268 {
 269         might_sleep();
 270
 271         if (is_single_threaded(wq)) {
 272                 /* Always use first cpu's area. */
 273                 flush_cpu_workqueue(per_cpu_ptr(wq->cpu_wq, singlethread_cpu));
 274         } else {
 275                 int cpu;
 276
 277                 lock_cpu_hotplug();
 278                 for_each_online_cpu(cpu)
 279                         flush_cpu_workqueue(per_cpu_ptr(wq->cpu_wq, cpu));
 280                 unlock_cpu_hotplug();
 281         }
 282 }
 283
 284 static struct task_struct *create_workqueue_thread(struct workqueue_struct *wq,
 285                                                    int cpu)
 286 {
 287         struct cpu_workqueue_struct *cwq = per_cpu_ptr(wq->cpu_wq, cpu);
 288         struct task_struct *p;
 289
 290         spin_lock_init(&cwq->lock);
 291         cwq->wq = wq;
 292         cwq->thread = NULL;
 293         cwq->insert_sequence = 0;
 294         cwq->remove_sequence = 0;
 295         INIT_LIST_HEAD(&cwq->worklist);
 296         init_waitqueue_head(&cwq->more_work);
 297         init_waitqueue_head(&cwq->work_done);
 298
 299         if (is_single_threaded(wq))
 300                 p = kthread_create(worker_thread, cwq, "%s", wq->name);
 301         else
 302                 p = kthread_create(worker_thread, cwq, "%s/%d", wq->name, cpu);
 303         if (IS_ERR(p))
 304                 return NULL;
 305         cwq->thread = p;
 306         return p;
 307 }
 308
 309 struct workqueue_struct *__create_workqueue(const char *name,
 310                                             int singlethread)
 311 {
 312         int cpu, destroy = 0;
 313         struct workqueue_struct *wq;
 314         struct task_struct *p;
 315
 316         wq = kzalloc(sizeof(*wq), GFP_KERNEL);
 317         if (!wq)
 318                 return NULL;
 319
 320         wq->cpu_wq = alloc_percpu(struct cpu_workqueue_struct);
 321         if (!wq->cpu_wq) {
 322                 kfree(wq);
 323                 return NULL;
 324         }
 325
 326         wq->name = name;
 327         /* We don't need the distraction of CPUs appearing and vanishing. */
 328         lock_cpu_hotplug();
 329         if (singlethread) {
 330                 INIT_LIST_HEAD(&wq->list);
 331                 p = create_workqueue_thread(wq, singlethread_cpu);
 332                 if (!p)
 333                         destroy = 1;
 334                 else
 335                         wake_up_process(p);
 336         } else {
 337                 spin_lock(&workqueue_lock);
 338                 list_add(&wq->list, &workqueues);
 339                 spin_unlock(&workqueue_lock);
 340                 for_each_online_cpu(cpu) {
 341                         p = create_workqueue_thread(wq, cpu);
 342                         if (p) {
 343                                 kthread_bind(p, cpu);
 344                                 wake_up_process(p);
 345                         } else
 346                                 destroy = 1;
 347                 }
 348         }
 349         unlock_cpu_hotplug();
 350
 351         /*
 352          * Was there any error during startup? If yes then clean up:
 353          */
 354         if (destroy) {
 355                 destroy_workqueue(wq);
 356                 wq = NULL;
 357         }
 358         return wq;
 359 }
 360
 361 static void cleanup_workqueue_thread(struct workqueue_struct *wq, int cpu)
 362 {
 363         struct cpu_workqueue_struct *cwq;
 364         unsigned long flags;
 365         struct task_struct *p;
 366
 367         cwq = per_cpu_ptr(wq->cpu_wq, cpu);
 368         spin_lock_irqsave(&cwq->lock, flags);
 369         p = cwq->thread;
 370         cwq->thread = NULL;
 371         spin_unlock_irqrestore(&cwq->lock, flags);
 372         if (p)
 373                 kthread_stop(p);
 374 }
 375
 376 void destroy_workqueue(struct workqueue_struct *wq)
 377 {
 378         int cpu;
 379
 380         flush_workqueue(wq);
 381
 382         /* We don't need the distraction of CPUs appearing and vanishing. */
 383         lock_cpu_hotplug();
 384         if (is_single_threaded(wq))
 385                 cleanup_workqueue_thread(wq, singlethread_cpu);
 386         else {
 387                 for_each_online_cpu(cpu)
 388                         cleanup_workqueue_thread(wq, cpu);
 389                 spin_lock(&workqueue_lock);
 390                 list_del(&wq->list);
 391                 spin_unlock(&workqueue_lock);
 392         }
 393         unlock_cpu_hotplug();
 394         free_percpu(wq->cpu_wq);
 395         kfree(wq);
 396 }
 397
 398 static struct workqueue_struct *keventd_wq;
 399
 400 int fastcall schedule_work(struct work_struct *work)
 401 {
 402         return queue_work(keventd_wq, work);
 403 }
 404
 405 int fastcall schedule_delayed_work(struct work_struct *work, unsigned long delay)
 406 {
 407         return queue_delayed_work(keventd_wq, work, delay);
 408 }
 409
 410 int schedule_delayed_work_on(int cpu,
 411                         struct work_struct *work, unsigned long delay)
 412 {
 413         int ret = 0;
 414         struct timer_list *timer = &work->timer;
 415
 416         if (!test_and_set_bit(0, &work->pending)) {
 417                 BUG_ON(timer_pending(timer));
 418                 BUG_ON(!list_empty(&work->entry));
 419                 /* This stores keventd_wq for the moment, for the timer_fn */
 420                 work->wq_data = keventd_wq;
 421                 timer->expires = jiffies + delay;
 422                 timer->data = (unsigned long)work;
 423                 timer->function = delayed_work_timer_fn;
 424                 add_timer_on(timer, cpu);
 425                 ret = 1;
 426         }
 427         return ret;
 428 }
 429
 430 int schedule_on_each_cpu(void (*func) (void *info), void *info)
 431 {
 432         int cpu;
 433         struct work_struct *work;
 434
 435         work = kmalloc(NR_CPUS * sizeof(struct work_struct), GFP_KERNEL);
 436
 437         if (!work)
 438                 return -ENOMEM;
 439         for_each_online_cpu(cpu) {
 440                 INIT_WORK(work + cpu, func, info);
 441                 __queue_work(per_cpu_ptr(keventd_wq->cpu_wq, cpu),
 442                                 work + cpu);
 443         }
 444         flush_workqueue(keventd_wq);
 445         kfree(work);
 446         return 0;
 447 }
 448
 449 void flush_scheduled_work(void)
 450 {
 451         flush_workqueue(keventd_wq);
 452 }
 453
 454 /**
 455  * cancel_rearming_delayed_workqueue - reliably kill off a delayed
 456  *                      work whose handler rearms the delayed work.
 457  * @wq:   the controlling workqueue structure
 458  * @work: the delayed work struct
 459  */
 460 void cancel_rearming_delayed_workqueue(struct workqueue_struct *wq,
 461                                        struct work_struct *work)
 462 {
 463         while (!cancel_delayed_work(work))
 464                 flush_workqueue(wq);
 465 }
 466 EXPORT_SYMBOL(cancel_rearming_delayed_workqueue);
 467
 468 /**
 469  * cancel_rearming_delayed_work - reliably kill off a delayed keventd
 470  *                      work whose handler rearms the delayed work.
 471  * @work: the delayed work struct
 472  */
 473 void cancel_rearming_delayed_work(struct work_struct *work)
 474 {
 475         cancel_rearming_delayed_workqueue(keventd_wq, work);
 476 }
 477 EXPORT_SYMBOL(cancel_rearming_delayed_work);
 478
 479 int keventd_up(void)
 480 {
 481         return keventd_wq != NULL;
 482 }
 483
 484 int current_is_keventd(void)
 485 {
 486         struct cpu_workqueue_struct *cwq;
 487         int cpu = smp_processor_id();   /* preempt-safe: keventd is per-cpu */
 488         int ret = 0;
 489
 490         BUG_ON(!keventd_wq);
 491
 492         cwq = per_cpu_ptr(keventd_wq->cpu_wq, cpu);
 493         if (current == cwq->thread)
 494                 ret = 1;
 495
 496         return ret;
 497
 498 }
 499
 500 #ifdef CONFIG_HOTPLUG_CPU
 501 /* Take the work from this (downed) CPU. */
 502 static void take_over_work(struct workqueue_struct *wq, unsigned int cpu)
 503 {
 504         struct cpu_workqueue_struct *cwq = per_cpu_ptr(wq->cpu_wq, cpu);
 505         LIST_HEAD(list);
 506         struct work_struct *work;
 507
 508         spin_lock_irq(&cwq->lock);
 509         list_splice_init(&cwq->worklist, &list);
 510
 511         while (!list_empty(&list)) {
 512                 printk("Taking work for %s\n", wq->name);
 513                 work = list_entry(list.next,struct work_struct,entry);
 514                 list_del(&work->entry);
 515                 __queue_work(per_cpu_ptr(wq->cpu_wq, smp_processor_id()), work);
 516         }
 517         spin_unlock_irq(&cwq->lock);
 518 }
 519
 520 /* We're holding the cpucontrol mutex here */
 521 static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
 522                                   unsigned long action,
 523                                   void *hcpu)
 524 {
 525         unsigned int hotcpu = (unsigned long)hcpu;
 526         struct workqueue_struct *wq;
 527
 528         switch (action) {
 529         case CPU_UP_PREPARE:
 530                 /* Create a new workqueue thread for it. */
 531                 list_for_each_entry(wq, &workqueues, list) {
 532                         if (!create_workqueue_thread(wq, hotcpu)) {
 533                                 printk("workqueue for %i failed\n", hotcpu);
 534                                 return NOTIFY_BAD;
 535                         }
 536                 }
 537                 break;
 538
 539         case CPU_ONLINE:
 540                 /* Kick off worker threads. */
 541                 list_for_each_entry(wq, &workqueues, list) {
 542                         struct cpu_workqueue_struct *cwq;
 543
 544                         cwq = per_cpu_ptr(wq->cpu_wq, hotcpu);
 545                         kthread_bind(cwq->thread, hotcpu);
 546                         wake_up_process(cwq->thread);
 547                 }
 548                 break;
 549
 550         case CPU_UP_CANCELED:
 551                 list_for_each_entry(wq, &workqueues, list) {
 552                         /* Unbind so it can run. */
 553                         kthread_bind(per_cpu_ptr(wq->cpu_wq, hotcpu)->thread,
 554                                      any_online_cpu(cpu_online_map));
 555                         cleanup_workqueue_thread(wq, hotcpu);
 556                 }
 557                 break;
 558
 559         case CPU_DEAD:
 560                 list_for_each_entry(wq, &workqueues, list)
 561                         cleanup_workqueue_thread(wq, hotcpu);
 562                 list_for_each_entry(wq, &workqueues, list)
 563                         take_over_work(wq, hotcpu);
 564                 break;
 565         }
 566
 567         return NOTIFY_OK;
 568 }
 569 #endif
 570
 571 void init_workqueues(void)
 572 {
 573         singlethread_cpu = first_cpu(cpu_possible_map);
 574         hotcpu_notifier(workqueue_cpu_callback, 0);
 575         keventd_wq = create_workqueue("events");
 576         BUG_ON(!keventd_wq);
 577 }
 578
 579 EXPORT_SYMBOL_GPL(__create_workqueue);
 580 EXPORT_SYMBOL_GPL(queue_work);
 581 EXPORT_SYMBOL_GPL(queue_delayed_work);
 582 EXPORT_SYMBOL_GPL(flush_workqueue);
 583 EXPORT_SYMBOL_GPL(destroy_workqueue);
 584
 585 EXPORT_SYMBOL(schedule_work);
 586 EXPORT_SYMBOL(schedule_delayed_work);
 587 EXPORT_SYMBOL(schedule_delayed_work_on);
 588 EXPORT_SYMBOL(flush_scheduled_work);