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@redhat.com>
  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
  17 #include <linux/module.h>
  18 #include <linux/kernel.h>
  19 #include <linux/sched.h>
  20 #include <linux/init.h>
  21 #include <linux/signal.h>
  22 #include <linux/completion.h>
  23 #include <linux/workqueue.h>
  24 #include <linux/slab.h>
  25 #include <linux/cpu.h>
  26 #include <linux/notifier.h>
  27 #include <linux/kthread.h>
  28
  29 /*
  30  * The per-CPU workqueue (if single thread, we always use cpu 0's).
  31  *
  32  * The sequence counters are for flush_scheduled_work().  It wants to wait
  33  * until until all currently-scheduled works are completed, but it doesn't
  34  * want to be livelocked by new, incoming ones.  So it waits until
  35  * remove_sequence is >= the insert_sequence which pertained when
  36  * flush_scheduled_work() was called.
  37  */
  38 struct cpu_workqueue_struct {
  39
  40         spinlock_t lock;
  41
  42         long remove_sequence;   /* Least-recently added (next to run) */
  43         long insert_sequence;   /* Next to add */
  44
  45         struct list_head worklist;
  46         wait_queue_head_t more_work;
  47         wait_queue_head_t work_done;
  48
  49         struct workqueue_struct *wq;
  50         task_t *thread;
  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[NR_CPUS];
  61         const char *name;
  62         struct list_head list;  /* Empty if single thread */
  63 };
  64
  65 /* All the per-cpu workqueues on the system, for hotplug cpu to add/remove
  66    threads to each one as cpus come/go. */
  67 static spinlock_t workqueue_lock = SPIN_LOCK_UNLOCKED;
  68 static LIST_HEAD(workqueues);
  69
  70 /* If it's single threaded, it isn't in the list of workqueues. */
  71 static inline int is_single_threaded(struct workqueue_struct *wq)
  72 {
  73         return list_empty(&wq->list);
  74 }
  75
  76 /* Preempt must be disabled. */
  77 static void __queue_work(struct cpu_workqueue_struct *cwq,
  78                          struct work_struct *work)
  79 {
  80         unsigned long flags;
  81
  82         spin_lock_irqsave(&cwq->lock, flags);
  83         work->wq_data = cwq;
  84         list_add_tail(&work->entry, &cwq->worklist);
  85         cwq->insert_sequence++;
  86         wake_up(&cwq->more_work);
  87         spin_unlock_irqrestore(&cwq->lock, flags);
  88 }
  89
  90 /*
  91  * Queue work on a workqueue. Return non-zero if it was successfully
  92  * added.
  93  *
  94  * We queue the work to the CPU it was submitted, but there is no
  95  * guarantee that it will be processed by that CPU.
  96  */
  97 int fastcall queue_work(struct workqueue_struct *wq, struct work_struct *work)
  98 {
  99         int ret = 0, cpu = get_cpu();
 100
 101         if (!test_and_set_bit(0, &work->pending)) {
 102                 if (unlikely(is_single_threaded(wq)))
 103                         cpu = 0;
 104                 BUG_ON(!list_empty(&work->entry));
 105                 __queue_work(wq->cpu_wq + cpu, work);
 106                 ret = 1;
 107         }
 108         put_cpu();
 109         return ret;
 110 }
 111
 112 static void delayed_work_timer_fn(unsigned long __data)
 113 {
 114         struct work_struct *work = (struct work_struct *)__data;
 115         struct workqueue_struct *wq = work->wq_data;
 116         int cpu = smp_processor_id();
 117
 118         if (unlikely(is_single_threaded(wq)))
 119                 cpu = 0;
 120
 121         __queue_work(wq->cpu_wq + cpu, work);
 122 }
 123
 124 int fastcall queue_delayed_work(struct workqueue_struct *wq,
 125                         struct work_struct *work, unsigned long delay)
 126 {
 127         int ret = 0;
 128         struct timer_list *timer = &work->timer;
 129
 130         if (!test_and_set_bit(0, &work->pending)) {
 131                 BUG_ON(timer_pending(timer));
 132                 BUG_ON(!list_empty(&work->entry));
 133
 134                 /* This stores wq for the moment, for the timer_fn */
 135                 work->wq_data = wq;
 136                 timer->expires = jiffies + delay;
 137                 timer->data = (unsigned long)work;
 138                 timer->function = delayed_work_timer_fn;
 139                 add_timer(timer);
 140                 ret = 1;
 141         }
 142         return ret;
 143 }
 144
 145 static inline void run_workqueue(struct cpu_workqueue_struct *cwq)
 146 {
 147         unsigned long flags;
 148
 149         /*
 150          * Keep taking off work from the queue until
 151          * done.
 152          */
 153         spin_lock_irqsave(&cwq->lock, flags);
 154         cwq->run_depth++;
 155         if (cwq->run_depth > 3) {
 156                 /* morton gets to eat his hat */
 157                 printk("%s: recursion depth exceeded: %d\n",
 158                         __FUNCTION__, cwq->run_depth);
 159                 dump_stack();
 160         }
 161         while (!list_empty(&cwq->worklist)) {
 162                 struct work_struct *work = list_entry(cwq->worklist.next,
 163                                                 struct work_struct, entry);
 164                 void (*f) (void *) = work->func;
 165                 void *data = work->data;
 166
 167                 list_del_init(cwq->worklist.next);
 168                 spin_unlock_irqrestore(&cwq->lock, flags);
 169
 170                 BUG_ON(work->wq_data != cwq);
 171                 clear_bit(0, &work->pending);
 172                 f(data);
 173
 174                 spin_lock_irqsave(&cwq->lock, flags);
 175                 cwq->remove_sequence++;
 176                 wake_up(&cwq->work_done);
 177         }
 178         cwq->run_depth--;
 179         spin_unlock_irqrestore(&cwq->lock, flags);
 180 }
 181
 182 static int worker_thread(void *__cwq)
 183 {
 184         struct cpu_workqueue_struct *cwq = __cwq;
 185         DECLARE_WAITQUEUE(wait, current);
 186         struct k_sigaction sa;
 187         sigset_t blocked;
 188
 189         current->flags |= PF_NOFREEZE;
 190
 191         set_user_nice(current, -10);
 192
 193         /* Block and flush all signals */
 194         sigfillset(&blocked);
 195         sigprocmask(SIG_BLOCK, &blocked, NULL);
 196         flush_signals(current);
 197
 198         /* SIG_IGN makes children autoreap: see do_notify_parent(). */
 199         sa.sa.sa_handler = SIG_IGN;
 200         sa.sa.sa_flags = 0;
 201         siginitset(&sa.sa.sa_mask, sigmask(SIGCHLD));
 202         do_sigaction(SIGCHLD, &sa, (struct k_sigaction *)0);
 203
 204         while (!kthread_should_stop()) {
 205                 set_task_state(current, TASK_INTERRUPTIBLE);
 206
 207                 add_wait_queue(&cwq->more_work, &wait);
 208                 if (list_empty(&cwq->worklist))
 209                         schedule();
 210                 else
 211                         set_task_state(current, TASK_RUNNING);
 212                 remove_wait_queue(&cwq->more_work, &wait);
 213
 214                 if (!list_empty(&cwq->worklist))
 215                         run_workqueue(cwq);
 216         }
 217         return 0;
 218 }
 219
 220 /*
 221  * flush_workqueue - ensure that any scheduled work has run to completion.
 222  *
 223  * Forces execution of the workqueue and blocks until its completion.
 224  * This is typically used in driver shutdown handlers.
 225  *
 226  * This function will sample each workqueue's current insert_sequence number and
 227  * will sleep until the head sequence is greater than or equal to that.  This
 228  * means that we sleep until all works which were queued on entry have been
 229  * handled, but we are not livelocked by new incoming ones.
 230  *
 231  * This function used to run the workqueues itself.  Now we just wait for the
 232  * helper threads to do it.
 233  */
 234 void fastcall flush_workqueue(struct workqueue_struct *wq)
 235 {
 236         struct cpu_workqueue_struct *cwq;
 237         int cpu;
 238
 239         might_sleep();
 240
 241         lock_cpu_hotplug();
 242         for_each_online_cpu(cpu) {
 243                 DEFINE_WAIT(wait);
 244                 long sequence_needed;
 245
 246                 if (is_single_threaded(wq))
 247                         cwq = wq->cpu_wq + 0; /* Always use cpu 0's area. */
 248                 else
 249                         cwq = wq->cpu_wq + cpu;
 250
 251                 if (cwq->thread == current) {
 252                         /*
 253                          * Probably keventd trying to flush its own queue.
 254                          * So simply run it by hand rather than deadlocking.
 255                          */
 256                         run_workqueue(cwq);
 257                         continue;
 258                 }
 259                 spin_lock_irq(&cwq->lock);
 260                 sequence_needed = cwq->insert_sequence;
 261
 262                 while (sequence_needed - cwq->remove_sequence > 0) {
 263                         prepare_to_wait(&cwq->work_done, &wait,
 264                                         TASK_UNINTERRUPTIBLE);
 265                         spin_unlock_irq(&cwq->lock);
 266                         schedule();
 267                         spin_lock_irq(&cwq->lock);
 268                 }
 269                 finish_wait(&cwq->work_done, &wait);
 270                 spin_unlock_irq(&cwq->lock);
 271         }
 272         unlock_cpu_hotplug();
 273 }
 274
 275 static struct task_struct *create_workqueue_thread(struct workqueue_struct *wq,
 276                                                    int cpu)
 277 {
 278         struct cpu_workqueue_struct *cwq = wq->cpu_wq + cpu;
 279         struct task_struct *p;
 280
 281         spin_lock_init(&cwq->lock);
 282         cwq->wq = wq;
 283         cwq->thread = NULL;
 284         cwq->insert_sequence = 0;
 285         cwq->remove_sequence = 0;
 286         INIT_LIST_HEAD(&cwq->worklist);
 287         init_waitqueue_head(&cwq->more_work);
 288         init_waitqueue_head(&cwq->work_done);
 289
 290         if (is_single_threaded(wq))
 291                 p = kthread_create(worker_thread, cwq, "%s", wq->name);
 292         else
 293                 p = kthread_create(worker_thread, cwq, "%s/%d", wq->name, cpu);
 294         if (IS_ERR(p))
 295                 return NULL;
 296         cwq->thread = p;
 297         return p;
 298 }
 299
 300 struct workqueue_struct *__create_workqueue(const char *name,
 301                                             int singlethread)
 302 {
 303         int cpu, destroy = 0;
 304         struct workqueue_struct *wq;
 305         struct task_struct *p;
 306
 307         BUG_ON(strlen(name) > 10);
 308
 309         wq = kmalloc(sizeof(*wq), GFP_KERNEL);
 310         if (!wq)
 311                 return NULL;
 312         memset(wq, 0, sizeof(*wq));
 313
 314         wq->name = name;
 315         /* We don't need the distraction of CPUs appearing and vanishing. */
 316         lock_cpu_hotplug();
 317         if (singlethread) {
 318                 INIT_LIST_HEAD(&wq->list);
 319                 p = create_workqueue_thread(wq, 0);
 320                 if (!p)
 321                         destroy = 1;
 322                 else
 323                         wake_up_process(p);
 324         } else {
 325                 spin_lock(&workqueue_lock);
 326                 list_add(&wq->list, &workqueues);
 327                 spin_unlock_irq(&workqueue_lock);
 328                 for_each_online_cpu(cpu) {
 329                         p = create_workqueue_thread(wq, cpu);
 330                         if (p) {
 331                                 kthread_bind(p, cpu);
 332                                 wake_up_process(p);
 333                         } else
 334                                 destroy = 1;
 335                 }
 336         }
 337
 338         /*
 339          * Was there any error during startup? If yes then clean up:
 340          */
 341         if (destroy) {
 342                 destroy_workqueue(wq);
 343                 wq = NULL;
 344         }
 345         unlock_cpu_hotplug();
 346         return wq;
 347 }
 348
 349 static void cleanup_workqueue_thread(struct workqueue_struct *wq, int cpu)
 350 {
 351         struct cpu_workqueue_struct *cwq;
 352         unsigned long flags;
 353         struct task_struct *p;
 354
 355         cwq = wq->cpu_wq + cpu;
 356         spin_lock_irqsave(&cwq->lock, flags);
 357         p = cwq->thread;
 358         cwq->thread = NULL;
 359         spin_unlock_irqrestore(&cwq->lock, flags);
 360         if (p)
 361                 kthread_stop(p);
 362 }
 363
 364 void destroy_workqueue(struct workqueue_struct *wq)
 365 {
 366         int cpu;
 367
 368         flush_workqueue(wq);
 369
 370         /* We don't need the distraction of CPUs appearing and vanishing. */
 371         lock_cpu_hotplug();
 372         if (is_single_threaded(wq))
 373                 cleanup_workqueue_thread(wq, 0);
 374         else {
 375                 for_each_online_cpu(cpu)
 376                         cleanup_workqueue_thread(wq, cpu);
 377                 spin_lock(&workqueue_lock);
 378                 list_del(&wq->list);
 379                 spin_unlock_irq(&workqueue_lock);
 380         }
 381         unlock_cpu_hotplug();
 382         kfree(wq);
 383 }
 384
 385 static struct workqueue_struct *keventd_wq;
 386
 387 int fastcall schedule_work(struct work_struct *work)
 388 {
 389         return queue_work(keventd_wq, work);
 390 }
 391
 392 int fastcall schedule_delayed_work(struct work_struct *work, unsigned long delay)
 393 {
 394         return queue_delayed_work(keventd_wq, work, delay);
 395 }
 396
 397 void flush_scheduled_work(void)
 398 {
 399         flush_workqueue(keventd_wq);
 400 }
 401
 402 int keventd_up(void)
 403 {
 404         return keventd_wq != NULL;
 405 }
 406
 407 int current_is_keventd(void)
 408 {
 409         struct cpu_workqueue_struct *cwq;
 410         int cpu = smp_processor_id();   /* preempt-safe: keventd is per-cpu */
 411         int ret = 0;
 412
 413         BUG_ON(!keventd_wq);
 414
 415         cwq = keventd_wq->cpu_wq + cpu;
 416         if (current == cwq->thread)
 417                 ret = 1;
 418
 419         return ret;
 420
 421 }
 422
 423 #ifdef CONFIG_HOTPLUG_CPU
 424 /* Take the work from this (downed) CPU. */
 425 static void take_over_work(struct workqueue_struct *wq, unsigned int cpu)
 426 {
 427         struct cpu_workqueue_struct *cwq = wq->cpu_wq + cpu;
 428         LIST_HEAD(list);
 429         struct work_struct *work;
 430
 431         spin_lock_irq(&cwq->lock);
 432         list_splice_init(&cwq->worklist, &list);
 433
 434         while (!list_empty(&list)) {
 435                 printk("Taking work for %s\n", wq->name);
 436                 work = list_entry(list.next,struct work_struct,entry);
 437                 list_del(&work->entry);
 438                 __queue_work(wq->cpu_wq + smp_processor_id(), work);
 439         }
 440         spin_unlock_irq(&cwq->lock);
 441 }
 442
 443 /* We're holding the cpucontrol mutex here */
 444 static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
 445                                   unsigned long action,
 446                                   void *hcpu)
 447 {
 448         unsigned int hotcpu = (unsigned long)hcpu;
 449         struct workqueue_struct *wq;
 450
 451         switch (action) {
 452         case CPU_UP_PREPARE:
 453                 /* Create a new workqueue thread for it. */
 454                 list_for_each_entry(wq, &workqueues, list) {
 455                         if (create_workqueue_thread(wq, hotcpu) < 0) {
 456                                 printk("workqueue for %i failed\n", hotcpu);
 457                                 return NOTIFY_BAD;
 458                         }
 459                 }
 460                 break;
 461
 462         case CPU_ONLINE:
 463                 /* Kick off worker threads. */
 464                 list_for_each_entry(wq, &workqueues, list)
 465                         wake_up_process(wq->cpu_wq[hotcpu].thread);
 466                 break;
 467
 468         case CPU_UP_CANCELED:
 469                 list_for_each_entry(wq, &workqueues, list) {
 470                         /* Unbind so it can run. */
 471                         kthread_bind(wq->cpu_wq[hotcpu].thread,
 472                                      smp_processor_id());
 473                         cleanup_workqueue_thread(wq, hotcpu);
 474                 }
 475                 break;
 476
 477         case CPU_DEAD:
 478                 list_for_each_entry(wq, &workqueues, list)
 479                         cleanup_workqueue_thread(wq, hotcpu);
 480                 list_for_each_entry(wq, &workqueues, list)
 481                         take_over_work(wq, hotcpu);
 482                 break;
 483         }
 484
 485         return NOTIFY_OK;
 486 }
 487 #endif
 488
 489 void init_workqueues(void)
 490 {
 491         hotcpu_notifier(workqueue_cpu_callback, 0);
 492         keventd_wq = create_workqueue("events");
 493         BUG_ON(!keventd_wq);
 494 }
 495
 496 EXPORT_SYMBOL_GPL(__create_workqueue);
 497 EXPORT_SYMBOL_GPL(queue_work);
 498 EXPORT_SYMBOL_GPL(queue_delayed_work);
 499 EXPORT_SYMBOL_GPL(flush_workqueue);
 500 EXPORT_SYMBOL_GPL(destroy_workqueue);
 501
 502 EXPORT_SYMBOL(schedule_work);
 503 EXPORT_SYMBOL(schedule_delayed_work);
 504 EXPORT_SYMBOL(flush_scheduled_work);
 505