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