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
  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.
  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 #include <linux/lockdep.h>
  36 #define CREATE_TRACE_POINTS
  37 #include <trace/events/workqueue.h>
  38
  39 /*
  40  * The per-CPU workqueue (if single thread, we always use the first
  41  * possible cpu).
  42  */
  43 struct cpu_workqueue_struct {
  44
  45         spinlock_t lock;
  46
  47         struct list_head worklist;
  48         wait_queue_head_t more_work;
  49         struct work_struct *current_work;
  50
  51         struct workqueue_struct *wq;
  52         struct task_struct *thread;
  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         struct list_head list;
  62         const char *name;
  63         int singlethread;
  64         int freezeable;         /* Freeze threads during suspend */
  65         int rt;
  66 #ifdef CONFIG_LOCKDEP
  67         struct lockdep_map lockdep_map;
  68 #endif
  69 };
  70
  71 #ifdef CONFIG_DEBUG_OBJECTS_WORK
  72
  73 static struct debug_obj_descr work_debug_descr;
  74
  75 /*
  76  * fixup_init is called when:
  77  * - an active object is initialized
  78  */
  79 static int work_fixup_init(void *addr, enum debug_obj_state state)
  80 {
  81         struct work_struct *work = addr;
  82
  83         switch (state) {
  84         case ODEBUG_STATE_ACTIVE:
  85                 cancel_work_sync(work);
  86                 debug_object_init(work, &work_debug_descr);
  87                 return 1;
  88         default:
  89                 return 0;
  90         }
  91 }
  92
  93 /*
  94  * fixup_activate is called when:
  95  * - an active object is activated
  96  * - an unknown object is activated (might be a statically initialized object)
  97  */
  98 static int work_fixup_activate(void *addr, enum debug_obj_state state)
  99 {
 100         struct work_struct *work = addr;
 101
 102         switch (state) {
 103
 104         case ODEBUG_STATE_NOTAVAILABLE:
 105                 /*
 106                  * This is not really a fixup. The work struct was
 107                  * statically initialized. We just make sure that it
 108                  * is tracked in the object tracker.
 109                  */
 110                 if (test_bit(WORK_STRUCT_STATIC, work_data_bits(work))) {
 111                         debug_object_init(work, &work_debug_descr);
 112                         debug_object_activate(work, &work_debug_descr);
 113                         return 0;
 114                 }
 115                 WARN_ON_ONCE(1);
 116                 return 0;
 117
 118         case ODEBUG_STATE_ACTIVE:
 119                 WARN_ON(1);
 120
 121         default:
 122                 return 0;
 123         }
 124 }
 125
 126 /*
 127  * fixup_free is called when:
 128  * - an active object is freed
 129  */
 130 static int work_fixup_free(void *addr, enum debug_obj_state state)
 131 {
 132         struct work_struct *work = addr;
 133
 134         switch (state) {
 135         case ODEBUG_STATE_ACTIVE:
 136                 cancel_work_sync(work);
 137                 debug_object_free(work, &work_debug_descr);
 138                 return 1;
 139         default:
 140                 return 0;
 141         }
 142 }
 143
 144 static struct debug_obj_descr work_debug_descr = {
 145         .name           = "work_struct",
 146         .fixup_init     = work_fixup_init,
 147         .fixup_activate = work_fixup_activate,
 148         .fixup_free     = work_fixup_free,
 149 };
 150
 151 static inline void debug_work_activate(struct work_struct *work)
 152 {
 153         debug_object_activate(work, &work_debug_descr);
 154 }
 155
 156 static inline void debug_work_deactivate(struct work_struct *work)
 157 {
 158         debug_object_deactivate(work, &work_debug_descr);
 159 }
 160
 161 void __init_work(struct work_struct *work, int onstack)
 162 {
 163         if (onstack)
 164                 debug_object_init_on_stack(work, &work_debug_descr);
 165         else
 166                 debug_object_init(work, &work_debug_descr);
 167 }
 168 EXPORT_SYMBOL_GPL(__init_work);
 169
 170 void destroy_work_on_stack(struct work_struct *work)
 171 {
 172         debug_object_free(work, &work_debug_descr);
 173 }
 174 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
 175
 176 #else
 177 static inline void debug_work_activate(struct work_struct *work) { }
 178 static inline void debug_work_deactivate(struct work_struct *work) { }
 179 #endif
 180
 181 /* Serializes the accesses to the list of workqueues. */
 182 static DEFINE_SPINLOCK(workqueue_lock);
 183 static LIST_HEAD(workqueues);
 184
 185 static int singlethread_cpu __read_mostly;
 186 static const struct cpumask *cpu_singlethread_map __read_mostly;
 187 /*
 188  * _cpu_down() first removes CPU from cpu_online_map, then CPU_DEAD
 189  * flushes cwq->worklist. This means that flush_workqueue/wait_on_work
 190  * which comes in between can't use for_each_online_cpu(). We could
 191  * use cpu_possible_map, the cpumask below is more a documentation
 192  * than optimization.
 193  */
 194 static cpumask_var_t cpu_populated_map __read_mostly;
 195
 196 /* If it's single threaded, it isn't in the list of workqueues. */
 197 static inline int is_wq_single_threaded(struct workqueue_struct *wq)
 198 {
 199         return wq->singlethread;
 200 }
 201
 202 static const struct cpumask *wq_cpu_map(struct workqueue_struct *wq)
 203 {
 204         return is_wq_single_threaded(wq)
 205                 ? cpu_singlethread_map : cpu_populated_map;
 206 }
 207
 208 static
 209 struct cpu_workqueue_struct *wq_per_cpu(struct workqueue_struct *wq, int cpu)
 210 {
 211         if (unlikely(is_wq_single_threaded(wq)))
 212                 cpu = singlethread_cpu;
 213         return per_cpu_ptr(wq->cpu_wq, cpu);
 214 }
 215
 216 /*
 217  * Set the workqueue on which a work item is to be run
 218  * - Must *only* be called if the pending flag is set
 219  */
 220 static inline void set_wq_data(struct work_struct *work,
 221                                 struct cpu_workqueue_struct *cwq)
 222 {
 223         unsigned long new;
 224
 225         BUG_ON(!work_pending(work));
 226
 227         new = (unsigned long) cwq | (1UL << WORK_STRUCT_PENDING);
 228         new |= WORK_STRUCT_FLAG_MASK & *work_data_bits(work);
 229         atomic_long_set(&work->data, new);
 230 }
 231
 232 static inline
 233 struct cpu_workqueue_struct *get_wq_data(struct work_struct *work)
 234 {
 235         return (void *) (atomic_long_read(&work->data) & WORK_STRUCT_WQ_DATA_MASK);
 236 }
 237
 238 static void insert_work(struct cpu_workqueue_struct *cwq,
 239                         struct work_struct *work, struct list_head *head)
 240 {
 241         trace_workqueue_insertion(cwq->thread, work);
 242
 243         set_wq_data(work, cwq);
 244         /*
 245          * Ensure that we get the right work->data if we see the
 246          * result of list_add() below, see try_to_grab_pending().
 247          */
 248         smp_wmb();
 249         list_add_tail(&work->entry, head);
 250         wake_up(&cwq->more_work);
 251 }
 252
 253 static void __queue_work(struct cpu_workqueue_struct *cwq,
 254                          struct work_struct *work)
 255 {
 256         unsigned long flags;
 257
 258         debug_work_activate(work);
 259         spin_lock_irqsave(&cwq->lock, flags);
 260         insert_work(cwq, work, &cwq->worklist);
 261         spin_unlock_irqrestore(&cwq->lock, flags);
 262 }
 263
 264 /**
 265  * queue_work - queue work on a workqueue
 266  * @wq: workqueue to use
 267  * @work: work to queue
 268  *
 269  * Returns 0 if @work was already on a queue, non-zero otherwise.
 270  *
 271  * We queue the work to the CPU on which it was submitted, but if the CPU dies
 272  * it can be processed by another CPU.
 273  */
 274 int queue_work(struct workqueue_struct *wq, struct work_struct *work)
 275 {
 276         int ret;
 277
 278         ret = queue_work_on(get_cpu(), wq, work);
 279         put_cpu();
 280
 281         return ret;
 282 }
 283 EXPORT_SYMBOL_GPL(queue_work);
 284
 285 /**
 286  * queue_work_on - queue work on specific cpu
 287  * @cpu: CPU number to execute work on
 288  * @wq: workqueue to use
 289  * @work: work to queue
 290  *
 291  * Returns 0 if @work was already on a queue, non-zero otherwise.
 292  *
 293  * We queue the work to a specific CPU, the caller must ensure it
 294  * can't go away.
 295  */
 296 int
 297 queue_work_on(int cpu, struct workqueue_struct *wq, struct work_struct *work)
 298 {
 299         int ret = 0;
 300
 301         if (!test_and_set_bit(WORK_STRUCT_PENDING, work_data_bits(work))) {
 302                 BUG_ON(!list_empty(&work->entry));
 303                 __queue_work(wq_per_cpu(wq, cpu), work);
 304                 ret = 1;
 305         }
 306         return ret;
 307 }
 308 EXPORT_SYMBOL_GPL(queue_work_on);
 309
 310 static void delayed_work_timer_fn(unsigned long __data)
 311 {
 312         struct delayed_work *dwork = (struct delayed_work *)__data;
 313         struct cpu_workqueue_struct *cwq = get_wq_data(&dwork->work);
 314         struct workqueue_struct *wq = cwq->wq;
 315
 316         __queue_work(wq_per_cpu(wq, smp_processor_id()), &dwork->work);
 317 }
 318
 319 /**
 320  * queue_delayed_work - queue work on a workqueue after delay
 321  * @wq: workqueue to use
 322  * @dwork: delayable work to queue
 323  * @delay: number of jiffies to wait before queueing
 324  *
 325  * Returns 0 if @work was already on a queue, non-zero otherwise.
 326  */
 327 int queue_delayed_work(struct workqueue_struct *wq,
 328                         struct delayed_work *dwork, unsigned long delay)
 329 {
 330         if (delay == 0)
 331                 return queue_work(wq, &dwork->work);
 332
 333         return queue_delayed_work_on(-1, wq, dwork, delay);
 334 }
 335 EXPORT_SYMBOL_GPL(queue_delayed_work);
 336
 337 /**
 338  * queue_delayed_work_on - queue work on specific CPU after delay
 339  * @cpu: CPU number to execute work on
 340  * @wq: workqueue to use
 341  * @dwork: work to queue
 342  * @delay: number of jiffies to wait before queueing
 343  *
 344  * Returns 0 if @work was already on a queue, non-zero otherwise.
 345  */
 346 int queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
 347                         struct delayed_work *dwork, unsigned long delay)
 348 {
 349         int ret = 0;
 350         struct timer_list *timer = &dwork->timer;
 351         struct work_struct *work = &dwork->work;
 352
 353         if (!test_and_set_bit(WORK_STRUCT_PENDING, work_data_bits(work))) {
 354                 BUG_ON(timer_pending(timer));
 355                 BUG_ON(!list_empty(&work->entry));
 356
 357                 timer_stats_timer_set_start_info(&dwork->timer);
 358
 359                 /* This stores cwq for the moment, for the timer_fn */
 360                 set_wq_data(work, wq_per_cpu(wq, raw_smp_processor_id()));
 361                 timer->expires = jiffies + delay;
 362                 timer->data = (unsigned long)dwork;
 363                 timer->function = delayed_work_timer_fn;
 364
 365                 if (unlikely(cpu >= 0))
 366                         add_timer_on(timer, cpu);
 367                 else
 368                         add_timer(timer);
 369                 ret = 1;
 370         }
 371         return ret;
 372 }
 373 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
 374
 375 static void run_workqueue(struct cpu_workqueue_struct *cwq)
 376 {
 377         spin_lock_irq(&cwq->lock);
 378         while (!list_empty(&cwq->worklist)) {
 379                 struct work_struct *work = list_entry(cwq->worklist.next,
 380                                                 struct work_struct, entry);
 381                 work_func_t f = work->func;
 382 #ifdef CONFIG_LOCKDEP
 383                 /*
 384                  * It is permissible to free the struct work_struct
 385                  * from inside the function that is called from it,
 386                  * this we need to take into account for lockdep too.
 387                  * To avoid bogus "held lock freed" warnings as well
 388                  * as problems when looking into work->lockdep_map,
 389                  * make a copy and use that here.
 390                  */
 391                 struct lockdep_map lockdep_map = work->lockdep_map;
 392 #endif
 393                 trace_workqueue_execution(cwq->thread, work);
 394                 debug_work_deactivate(work);
 395                 cwq->current_work = work;
 396                 list_del_init(cwq->worklist.next);
 397                 spin_unlock_irq(&cwq->lock);
 398
 399                 BUG_ON(get_wq_data(work) != cwq);
 400                 work_clear_pending(work);
 401                 lock_map_acquire(&cwq->wq->lockdep_map);
 402                 lock_map_acquire(&lockdep_map);
 403                 f(work);
 404                 lock_map_release(&lockdep_map);
 405                 lock_map_release(&cwq->wq->lockdep_map);
 406
 407                 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
 408                         printk(KERN_ERR "BUG: workqueue leaked lock or atomic: "
 409                                         "%s/0x%08x/%d\n",
 410                                         current->comm, preempt_count(),
 411                                         task_pid_nr(current));
 412                         printk(KERN_ERR "    last function: ");
 413                         print_symbol("%s\n", (unsigned long)f);
 414                         debug_show_held_locks(current);
 415                         dump_stack();
 416                 }
 417
 418                 spin_lock_irq(&cwq->lock);
 419                 cwq->current_work = NULL;
 420         }
 421         spin_unlock_irq(&cwq->lock);
 422 }
 423
 424 static int worker_thread(void *__cwq)
 425 {
 426         struct cpu_workqueue_struct *cwq = __cwq;
 427         DEFINE_WAIT(wait);
 428
 429         if (cwq->wq->freezeable)
 430                 set_freezable();
 431
 432         for (;;) {
 433                 prepare_to_wait(&cwq->more_work, &wait, TASK_INTERRUPTIBLE);
 434                 if (!freezing(current) &&
 435                     !kthread_should_stop() &&
 436                     list_empty(&cwq->worklist))
 437                         schedule();
 438                 finish_wait(&cwq->more_work, &wait);
 439
 440                 try_to_freeze();
 441
 442                 if (kthread_should_stop())
 443                         break;
 444
 445                 run_workqueue(cwq);
 446         }
 447
 448         return 0;
 449 }
 450
 451 struct wq_barrier {
 452         struct work_struct      work;
 453         struct completion       done;
 454 };
 455
 456 static void wq_barrier_func(struct work_struct *work)
 457 {
 458         struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
 459         complete(&barr->done);
 460 }
 461
 462 static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
 463                         struct wq_barrier *barr, struct list_head *head)
 464 {
 465         /*
 466          * debugobject calls are safe here even with cwq->lock locked
 467          * as we know for sure that this will not trigger any of the
 468          * checks and call back into the fixup functions where we
 469          * might deadlock.
 470          */
 471         INIT_WORK_ON_STACK(&barr->work, wq_barrier_func);
 472         __set_bit(WORK_STRUCT_PENDING, work_data_bits(&barr->work));
 473
 474         init_completion(&barr->done);
 475
 476         debug_work_activate(&barr->work);
 477         insert_work(cwq, &barr->work, head);
 478 }
 479
 480 static int flush_cpu_workqueue(struct cpu_workqueue_struct *cwq)
 481 {
 482         int active = 0;
 483         struct wq_barrier barr;
 484
 485         WARN_ON(cwq->thread == current);
 486
 487         spin_lock_irq(&cwq->lock);
 488         if (!list_empty(&cwq->worklist) || cwq->current_work != NULL) {
 489                 insert_wq_barrier(cwq, &barr, &cwq->worklist);
 490                 active = 1;
 491         }
 492         spin_unlock_irq(&cwq->lock);
 493
 494         if (active) {
 495                 wait_for_completion(&barr.done);
 496                 destroy_work_on_stack(&barr.work);
 497         }
 498
 499         return active;
 500 }
 501
 502 /**
 503  * flush_workqueue - ensure that any scheduled work has run to completion.
 504  * @wq: workqueue to flush
 505  *
 506  * Forces execution of the workqueue and blocks until its completion.
 507  * This is typically used in driver shutdown handlers.
 508  *
 509  * We sleep until all works which were queued on entry have been handled,
 510  * but we are not livelocked by new incoming ones.
 511  *
 512  * This function used to run the workqueues itself.  Now we just wait for the
 513  * helper threads to do it.
 514  */
 515 void flush_workqueue(struct workqueue_struct *wq)
 516 {
 517         const struct cpumask *cpu_map = wq_cpu_map(wq);
 518         int cpu;
 519
 520         might_sleep();
 521         lock_map_acquire(&wq->lockdep_map);
 522         lock_map_release(&wq->lockdep_map);
 523         for_each_cpu(cpu, cpu_map)
 524                 flush_cpu_workqueue(per_cpu_ptr(wq->cpu_wq, cpu));
 525 }
 526 EXPORT_SYMBOL_GPL(flush_workqueue);
 527
 528 /**
 529  * flush_work - block until a work_struct's callback has terminated
 530  * @work: the work which is to be flushed
 531  *
 532  * Returns false if @work has already terminated.
 533  *
 534  * It is expected that, prior to calling flush_work(), the caller has
 535  * arranged for the work to not be requeued, otherwise it doesn't make
 536  * sense to use this function.
 537  */
 538 int flush_work(struct work_struct *work)
 539 {
 540         struct cpu_workqueue_struct *cwq;
 541         struct list_head *prev;
 542         struct wq_barrier barr;
 543
 544         might_sleep();
 545         cwq = get_wq_data(work);
 546         if (!cwq)
 547                 return 0;
 548
 549         lock_map_acquire(&cwq->wq->lockdep_map);
 550         lock_map_release(&cwq->wq->lockdep_map);
 551
 552         prev = NULL;
 553         spin_lock_irq(&cwq->lock);
 554         if (!list_empty(&work->entry)) {
 555                 /*
 556                  * See the comment near try_to_grab_pending()->smp_rmb().
 557                  * If it was re-queued under us we are not going to wait.
 558                  */
 559                 smp_rmb();
 560                 if (unlikely(cwq != get_wq_data(work)))
 561                         goto out;
 562                 prev = &work->entry;
 563         } else {
 564                 if (cwq->current_work != work)
 565                         goto out;
 566                 prev = &cwq->worklist;
 567         }
 568         insert_wq_barrier(cwq, &barr, prev->next);
 569 out:
 570         spin_unlock_irq(&cwq->lock);
 571         if (!prev)
 572                 return 0;
 573
 574         wait_for_completion(&barr.done);
 575         destroy_work_on_stack(&barr.work);
 576         return 1;
 577 }
 578 EXPORT_SYMBOL_GPL(flush_work);
 579
 580 /*
 581  * Upon a successful return (>= 0), the caller "owns" WORK_STRUCT_PENDING bit,
 582  * so this work can't be re-armed in any way.
 583  */
 584 static int try_to_grab_pending(struct work_struct *work)
 585 {
 586         struct cpu_workqueue_struct *cwq;
 587         int ret = -1;
 588
 589         if (!test_and_set_bit(WORK_STRUCT_PENDING, work_data_bits(work)))
 590                 return 0;
 591
 592         /*
 593          * The queueing is in progress, or it is already queued. Try to
 594          * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
 595          */
 596
 597         cwq = get_wq_data(work);
 598         if (!cwq)
 599                 return ret;
 600
 601         spin_lock_irq(&cwq->lock);
 602         if (!list_empty(&work->entry)) {
 603                 /*
 604                  * This work is queued, but perhaps we locked the wrong cwq.
 605                  * In that case we must see the new value after rmb(), see
 606                  * insert_work()->wmb().
 607                  */
 608                 smp_rmb();
 609                 if (cwq == get_wq_data(work)) {
 610                         debug_work_deactivate(work);
 611                         list_del_init(&work->entry);
 612                         ret = 1;
 613                 }
 614         }
 615         spin_unlock_irq(&cwq->lock);
 616
 617         return ret;
 618 }
 619
 620 static void wait_on_cpu_work(struct cpu_workqueue_struct *cwq,
 621                                 struct work_struct *work)
 622 {
 623         struct wq_barrier barr;
 624         int running = 0;
 625
 626         spin_lock_irq(&cwq->lock);
 627         if (unlikely(cwq->current_work == work)) {
 628                 insert_wq_barrier(cwq, &barr, cwq->worklist.next);
 629                 running = 1;
 630         }
 631         spin_unlock_irq(&cwq->lock);
 632
 633         if (unlikely(running)) {
 634                 wait_for_completion(&barr.done);
 635                 destroy_work_on_stack(&barr.work);
 636         }
 637 }
 638
 639 static void wait_on_work(struct work_struct *work)
 640 {
 641         struct cpu_workqueue_struct *cwq;
 642         struct workqueue_struct *wq;
 643         const struct cpumask *cpu_map;
 644         int cpu;
 645
 646         might_sleep();
 647
 648         lock_map_acquire(&work->lockdep_map);
 649         lock_map_release(&work->lockdep_map);
 650
 651         cwq = get_wq_data(work);
 652         if (!cwq)
 653                 return;
 654
 655         wq = cwq->wq;
 656         cpu_map = wq_cpu_map(wq);
 657
 658         for_each_cpu(cpu, cpu_map)
 659                 wait_on_cpu_work(per_cpu_ptr(wq->cpu_wq, cpu), work);
 660 }
 661
 662 static int __cancel_work_timer(struct work_struct *work,
 663                                 struct timer_list* timer)
 664 {
 665         int ret;
 666
 667         do {
 668                 ret = (timer && likely(del_timer(timer)));
 669                 if (!ret)
 670                         ret = try_to_grab_pending(work);
 671                 wait_on_work(work);
 672         } while (unlikely(ret < 0));
 673
 674         work_clear_pending(work);
 675         return ret;
 676 }
 677
 678 /**
 679  * cancel_work_sync - block until a work_struct's callback has terminated
 680  * @work: the work which is to be flushed
 681  *
 682  * Returns true if @work was pending.
 683  *
 684  * cancel_work_sync() will cancel the work if it is queued. If the work's
 685  * callback appears to be running, cancel_work_sync() will block until it
 686  * has completed.
 687  *
 688  * It is possible to use this function if the work re-queues itself. It can
 689  * cancel the work even if it migrates to another workqueue, however in that
 690  * case it only guarantees that work->func() has completed on the last queued
 691  * workqueue.
 692  *
 693  * cancel_work_sync(&delayed_work->work) should be used only if ->timer is not
 694  * pending, otherwise it goes into a busy-wait loop until the timer expires.
 695  *
 696  * The caller must ensure that workqueue_struct on which this work was last
 697  * queued can't be destroyed before this function returns.
 698  */
 699 int cancel_work_sync(struct work_struct *work)
 700 {
 701         return __cancel_work_timer(work, NULL);
 702 }
 703 EXPORT_SYMBOL_GPL(cancel_work_sync);
 704
 705 /**
 706  * cancel_delayed_work_sync - reliably kill off a delayed work.
 707  * @dwork: the delayed work struct
 708  *
 709  * Returns true if @dwork was pending.
 710  *
 711  * It is possible to use this function if @dwork rearms itself via queue_work()
 712  * or queue_delayed_work(). See also the comment for cancel_work_sync().
 713  */
 714 int cancel_delayed_work_sync(struct delayed_work *dwork)
 715 {
 716         return __cancel_work_timer(&dwork->work, &dwork->timer);
 717 }
 718 EXPORT_SYMBOL(cancel_delayed_work_sync);
 719
 720 static struct workqueue_struct *keventd_wq __read_mostly;
 721
 722 /**
 723  * schedule_work - put work task in global workqueue
 724  * @work: job to be done
 725  *
 726  * Returns zero if @work was already on the kernel-global workqueue and
 727  * non-zero otherwise.
 728  *
 729  * This puts a job in the kernel-global workqueue if it was not already
 730  * queued and leaves it in the same position on the kernel-global
 731  * workqueue otherwise.
 732  */
 733 int schedule_work(struct work_struct *work)
 734 {
 735         return queue_work(keventd_wq, work);
 736 }
 737 EXPORT_SYMBOL(schedule_work);
 738
 739 /*
 740  * schedule_work_on - put work task on a specific cpu
 741  * @cpu: cpu to put the work task on
 742  * @work: job to be done
 743  *
 744  * This puts a job on a specific cpu
 745  */
 746 int schedule_work_on(int cpu, struct work_struct *work)
 747 {
 748         return queue_work_on(cpu, keventd_wq, work);
 749 }
 750 EXPORT_SYMBOL(schedule_work_on);
 751
 752 /**
 753  * schedule_delayed_work - put work task in global workqueue after delay
 754  * @dwork: job to be done
 755  * @delay: number of jiffies to wait or 0 for immediate execution
 756  *
 757  * After waiting for a given time this puts a job in the kernel-global
 758  * workqueue.
 759  */
 760 int schedule_delayed_work(struct delayed_work *dwork,
 761                                         unsigned long delay)
 762 {
 763         return queue_delayed_work(keventd_wq, dwork, delay);
 764 }
 765 EXPORT_SYMBOL(schedule_delayed_work);
 766
 767 /**
 768  * flush_delayed_work - block until a dwork_struct's callback has terminated
 769  * @dwork: the delayed work which is to be flushed
 770  *
 771  * Any timeout is cancelled, and any pending work is run immediately.
 772  */
 773 void flush_delayed_work(struct delayed_work *dwork)
 774 {
 775         if (del_timer_sync(&dwork->timer)) {
 776                 struct cpu_workqueue_struct *cwq;
 777                 cwq = wq_per_cpu(get_wq_data(&dwork->work)->wq, get_cpu());
 778                 __queue_work(cwq, &dwork->work);
 779                 put_cpu();
 780         }
 781         flush_work(&dwork->work);
 782 }
 783 EXPORT_SYMBOL(flush_delayed_work);
 784
 785 /**
 786  * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
 787  * @cpu: cpu to use
 788  * @dwork: job to be done
 789  * @delay: number of jiffies to wait
 790  *
 791  * After waiting for a given time this puts a job in the kernel-global
 792  * workqueue on the specified CPU.
 793  */
 794 int schedule_delayed_work_on(int cpu,
 795                         struct delayed_work *dwork, unsigned long delay)
 796 {
 797         return queue_delayed_work_on(cpu, keventd_wq, dwork, delay);
 798 }
 799 EXPORT_SYMBOL(schedule_delayed_work_on);
 800
 801 /**
 802  * schedule_on_each_cpu - call a function on each online CPU from keventd
 803  * @func: the function to call
 804  *
 805  * Returns zero on success.
 806  * Returns -ve errno on failure.
 807  *
 808  * schedule_on_each_cpu() is very slow.
 809  */
 810 int schedule_on_each_cpu(work_func_t func)
 811 {
 812         int cpu;
 813         int orig = -1;
 814         struct work_struct *works;
 815
 816         works = alloc_percpu(struct work_struct);
 817         if (!works)
 818                 return -ENOMEM;
 819
 820         get_online_cpus();
 821
 822         /*
 823          * When running in keventd don't schedule a work item on
 824          * itself.  Can just call directly because the work queue is
 825          * already bound.  This also is faster.
 826          */
 827         if (current_is_keventd())
 828                 orig = raw_smp_processor_id();
 829
 830         for_each_online_cpu(cpu) {
 831                 struct work_struct *work = per_cpu_ptr(works, cpu);
 832
 833                 INIT_WORK(work, func);
 834                 if (cpu != orig)
 835                         schedule_work_on(cpu, work);
 836         }
 837         if (orig >= 0)
 838                 func(per_cpu_ptr(works, orig));
 839
 840         for_each_online_cpu(cpu)
 841                 flush_work(per_cpu_ptr(works, cpu));
 842
 843         put_online_cpus();
 844         free_percpu(works);
 845         return 0;
 846 }
 847
 848 void flush_scheduled_work(void)
 849 {
 850         flush_workqueue(keventd_wq);
 851 }
 852 EXPORT_SYMBOL(flush_scheduled_work);
 853
 854 /**
 855  * execute_in_process_context - reliably execute the routine with user context
 856  * @fn:         the function to execute
 857  * @ew:         guaranteed storage for the execute work structure (must
 858  *              be available when the work executes)
 859  *
 860  * Executes the function immediately if process context is available,
 861  * otherwise schedules the function for delayed execution.
 862  *
 863  * Returns:     0 - function was executed
 864  *              1 - function was scheduled for execution
 865  */
 866 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
 867 {
 868         if (!in_interrupt()) {
 869                 fn(&ew->work);
 870                 return 0;
 871         }
 872
 873         INIT_WORK(&ew->work, fn);
 874         schedule_work(&ew->work);
 875
 876         return 1;
 877 }
 878 EXPORT_SYMBOL_GPL(execute_in_process_context);
 879
 880 int keventd_up(void)
 881 {
 882         return keventd_wq != NULL;
 883 }
 884
 885 int current_is_keventd(void)
 886 {
 887         struct cpu_workqueue_struct *cwq;
 888         int cpu = raw_smp_processor_id(); /* preempt-safe: keventd is per-cpu */
 889         int ret = 0;
 890
 891         BUG_ON(!keventd_wq);
 892
 893         cwq = per_cpu_ptr(keventd_wq->cpu_wq, cpu);
 894         if (current == cwq->thread)
 895                 ret = 1;
 896
 897         return ret;
 898
 899 }
 900
 901 static struct cpu_workqueue_struct *
 902 init_cpu_workqueue(struct workqueue_struct *wq, int cpu)
 903 {
 904         struct cpu_workqueue_struct *cwq = per_cpu_ptr(wq->cpu_wq, cpu);
 905
 906         cwq->wq = wq;
 907         spin_lock_init(&cwq->lock);
 908         INIT_LIST_HEAD(&cwq->worklist);
 909         init_waitqueue_head(&cwq->more_work);
 910
 911         return cwq;
 912 }
 913
 914 static int create_workqueue_thread(struct cpu_workqueue_struct *cwq, int cpu)
 915 {
 916         struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 };
 917         struct workqueue_struct *wq = cwq->wq;
 918         const char *fmt = is_wq_single_threaded(wq) ? "%s" : "%s/%d";
 919         struct task_struct *p;
 920
 921         p = kthread_create(worker_thread, cwq, fmt, wq->name, cpu);
 922         /*
 923          * Nobody can add the work_struct to this cwq,
 924          *      if (caller is __create_workqueue)
 925          *              nobody should see this wq
 926          *      else // caller is CPU_UP_PREPARE
 927          *              cpu is not on cpu_online_map
 928          * so we can abort safely.
 929          */
 930         if (IS_ERR(p))
 931                 return PTR_ERR(p);
 932         if (cwq->wq->rt)
 933                 sched_setscheduler_nocheck(p, SCHED_FIFO, &param);
 934         cwq->thread = p;
 935
 936         trace_workqueue_creation(cwq->thread, cpu);
 937
 938         return 0;
 939 }
 940
 941 static void start_workqueue_thread(struct cpu_workqueue_struct *cwq, int cpu)
 942 {
 943         struct task_struct *p = cwq->thread;
 944
 945         if (p != NULL) {
 946                 if (cpu >= 0)
 947                         kthread_bind(p, cpu);
 948                 wake_up_process(p);
 949         }
 950 }
 951
 952 struct workqueue_struct *__create_workqueue_key(const char *name,
 953                                                 int singlethread,
 954                                                 int freezeable,
 955                                                 int rt,
 956                                                 struct lock_class_key *key,
 957                                                 const char *lock_name)
 958 {
 959         struct workqueue_struct *wq;
 960         struct cpu_workqueue_struct *cwq;
 961         int err = 0, cpu;
 962
 963         wq = kzalloc(sizeof(*wq), GFP_KERNEL);
 964         if (!wq)
 965                 return NULL;
 966
 967         wq->cpu_wq = alloc_percpu(struct cpu_workqueue_struct);
 968         if (!wq->cpu_wq) {
 969                 kfree(wq);
 970                 return NULL;
 971         }
 972
 973         wq->name = name;
 974         lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
 975         wq->singlethread = singlethread;
 976         wq->freezeable = freezeable;
 977         wq->rt = rt;
 978         INIT_LIST_HEAD(&wq->list);
 979
 980         if (singlethread) {
 981                 cwq = init_cpu_workqueue(wq, singlethread_cpu);
 982                 err = create_workqueue_thread(cwq, singlethread_cpu);
 983                 start_workqueue_thread(cwq, -1);
 984         } else {
 985                 cpu_maps_update_begin();
 986                 /*
 987                  * We must place this wq on list even if the code below fails.
 988                  * cpu_down(cpu) can remove cpu from cpu_populated_map before
 989                  * destroy_workqueue() takes the lock, in that case we leak
 990                  * cwq[cpu]->thread.
 991                  */
 992                 spin_lock(&workqueue_lock);
 993                 list_add(&wq->list, &workqueues);
 994                 spin_unlock(&workqueue_lock);
 995                 /*
 996                  * We must initialize cwqs for each possible cpu even if we
 997                  * are going to call destroy_workqueue() finally. Otherwise
 998                  * cpu_up() can hit the uninitialized cwq once we drop the
 999                  * lock.
1000                  */
1001                 for_each_possible_cpu(cpu) {
1002                         cwq = init_cpu_workqueue(wq, cpu);
1003                         if (err || !cpu_online(cpu))
1004                                 continue;
1005                         err = create_workqueue_thread(cwq, cpu);
1006                         start_workqueue_thread(cwq, cpu);
1007                 }
1008                 cpu_maps_update_done();
1009         }
1010
1011         if (err) {
1012                 destroy_workqueue(wq);
1013                 wq = NULL;
1014         }
1015         return wq;
1016 }
1017 EXPORT_SYMBOL_GPL(__create_workqueue_key);
1018
1019 static void cleanup_workqueue_thread(struct cpu_workqueue_struct *cwq)
1020 {
1021         /*
1022          * Our caller is either destroy_workqueue() or CPU_POST_DEAD,
1023          * cpu_add_remove_lock protects cwq->thread.
1024          */
1025         if (cwq->thread == NULL)
1026                 return;
1027
1028         lock_map_acquire(&cwq->wq->lockdep_map);
1029         lock_map_release(&cwq->wq->lockdep_map);
1030
1031         flush_cpu_workqueue(cwq);
1032         /*
1033          * If the caller is CPU_POST_DEAD and cwq->worklist was not empty,
1034          * a concurrent flush_workqueue() can insert a barrier after us.
1035          * However, in that case run_workqueue() won't return and check
1036          * kthread_should_stop() until it flushes all work_struct's.
1037          * When ->worklist becomes empty it is safe to exit because no
1038          * more work_structs can be queued on this cwq: flush_workqueue
1039          * checks list_empty(), and a "normal" queue_work() can't use
1040          * a dead CPU.
1041          */
1042         trace_workqueue_destruction(cwq->thread);
1043         kthread_stop(cwq->thread);
1044         cwq->thread = NULL;
1045 }
1046
1047 /**
1048  * destroy_workqueue - safely terminate a workqueue
1049  * @wq: target workqueue
1050  *
1051  * Safely destroy a workqueue. All work currently pending will be done first.
1052  */
1053 void destroy_workqueue(struct workqueue_struct *wq)
1054 {
1055         const struct cpumask *cpu_map = wq_cpu_map(wq);
1056         int cpu;
1057
1058         cpu_maps_update_begin();
1059         spin_lock(&workqueue_lock);
1060         list_del(&wq->list);
1061         spin_unlock(&workqueue_lock);
1062
1063         for_each_cpu(cpu, cpu_map)
1064                 cleanup_workqueue_thread(per_cpu_ptr(wq->cpu_wq, cpu));
1065         cpu_maps_update_done();
1066
1067         free_percpu(wq->cpu_wq);
1068         kfree(wq);
1069 }
1070 EXPORT_SYMBOL_GPL(destroy_workqueue);
1071
1072 static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
1073                                                 unsigned long action,
1074                                                 void *hcpu)
1075 {
1076         unsigned int cpu = (unsigned long)hcpu;
1077         struct cpu_workqueue_struct *cwq;
1078         struct workqueue_struct *wq;
1079         int ret = NOTIFY_OK;
1080
1081         action &= ~CPU_TASKS_FROZEN;
1082
1083         switch (action) {
1084         case CPU_UP_PREPARE:
1085                 cpumask_set_cpu(cpu, cpu_populated_map);
1086         }
1087 undo:
1088         list_for_each_entry(wq, &workqueues, list) {
1089                 cwq = per_cpu_ptr(wq->cpu_wq, cpu);
1090
1091                 switch (action) {
1092                 case CPU_UP_PREPARE:
1093                         if (!create_workqueue_thread(cwq, cpu))
1094                                 break;
1095                         printk(KERN_ERR "workqueue [%s] for %i failed\n",
1096                                 wq->name, cpu);
1097                         action = CPU_UP_CANCELED;
1098                         ret = NOTIFY_BAD;
1099                         goto undo;
1100
1101                 case CPU_ONLINE:
1102                         start_workqueue_thread(cwq, cpu);
1103                         break;
1104
1105                 case CPU_UP_CANCELED:
1106                         start_workqueue_thread(cwq, -1);
1107                 case CPU_POST_DEAD:
1108                         cleanup_workqueue_thread(cwq);
1109                         break;
1110                 }
1111         }
1112
1113         switch (action) {
1114         case CPU_UP_CANCELED:
1115         case CPU_POST_DEAD:
1116                 cpumask_clear_cpu(cpu, cpu_populated_map);
1117         }
1118
1119         return ret;
1120 }
1121
1122 #ifdef CONFIG_SMP
1123
1124 struct work_for_cpu {
1125         struct completion completion;
1126         long (*fn)(void *);
1127         void *arg;
1128         long ret;
1129 };
1130
1131 static int do_work_for_cpu(void *_wfc)
1132 {
1133         struct work_for_cpu *wfc = _wfc;
1134         wfc->ret = wfc->fn(wfc->arg);
1135         complete(&wfc->completion);
1136         return 0;
1137 }
1138
1139 /**
1140  * work_on_cpu - run a function in user context on a particular cpu
1141  * @cpu: the cpu to run on
1142  * @fn: the function to run
1143  * @arg: the function arg
1144  *
1145  * This will return the value @fn returns.
1146  * It is up to the caller to ensure that the cpu doesn't go offline.
1147  * The caller must not hold any locks which would prevent @fn from completing.
1148  */
1149 long work_on_cpu(unsigned int cpu, long (*fn)(void *), void *arg)
1150 {
1151         struct task_struct *sub_thread;
1152         struct work_for_cpu wfc = {
1153                 .completion = COMPLETION_INITIALIZER_ONSTACK(wfc.completion),
1154                 .fn = fn,
1155                 .arg = arg,
1156         };
1157
1158         sub_thread = kthread_create(do_work_for_cpu, &wfc, "work_for_cpu");
1159         if (IS_ERR(sub_thread))
1160                 return PTR_ERR(sub_thread);
1161         kthread_bind(sub_thread, cpu);
1162         wake_up_process(sub_thread);
1163         wait_for_completion(&wfc.completion);
1164         return wfc.ret;
1165 }
1166 EXPORT_SYMBOL_GPL(work_on_cpu);
1167 #endif /* CONFIG_SMP */
1168
1169 void __init init_workqueues(void)
1170 {
1171         alloc_cpumask_var(&cpu_populated_map, GFP_KERNEL);
1172
1173         cpumask_copy(cpu_populated_map, cpu_online_mask);
1174         singlethread_cpu = cpumask_first(cpu_possible_mask);
1175         cpu_singlethread_map = cpumask_of(singlethread_cpu);
1176         hotcpu_notifier(workqueue_cpu_callback, 0);
1177         keventd_wq = create_workqueue("events");
1178         BUG_ON(!keventd_wq);
1179 }