fs/btrfs/async-thread.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/kthread.h>
  20 #include <linux/slab.h>
  21 #include <linux/list.h>
  22 #include <linux/spinlock.h>
  23 #include <linux/freezer.h>
  24 #include "async-thread.h"
  25
  26 #define WORK_QUEUED_BIT 0
  27 #define WORK_DONE_BIT 1
  28 #define WORK_ORDER_DONE_BIT 2
  29 #define WORK_HIGH_PRIO_BIT 3
  30
  31 /*
  32  * container for the kthread task pointer and the list of pending work
  33  * One of these is allocated per thread.
  34  */
  35 struct btrfs_worker_thread {
  36         /* pool we belong to */
  37         struct btrfs_workers *workers;
  38
  39         /* list of struct btrfs_work that are waiting for service */
  40         struct list_head pending;
  41         struct list_head prio_pending;
  42
  43         /* list of worker threads from struct btrfs_workers */
  44         struct list_head worker_list;
  45
  46         /* kthread */
  47         struct task_struct *task;
  48
  49         /* number of things on the pending list */
  50         atomic_t num_pending;
  51
  52         /* reference counter for this struct */
  53         atomic_t refs;
  54
  55         unsigned long sequence;
  56
  57         /* protects the pending list. */
  58         spinlock_t lock;
  59
  60         /* set to non-zero when this thread is already awake and kicking */
  61         int working;
  62
  63         /* are we currently idle */
  64         int idle;
  65 };
  66
  67 static int __btrfs_start_workers(struct btrfs_workers *workers);
  68
  69 /*
  70  * btrfs_start_workers uses kthread_run, which can block waiting for memory
  71  * for a very long time.  It will actually throttle on page writeback,
  72  * and so it may not make progress until after our btrfs worker threads
  73  * process all of the pending work structs in their queue
  74  *
  75  * This means we can't use btrfs_start_workers from inside a btrfs worker
  76  * thread that is used as part of cleaning dirty memory, which pretty much
  77  * involves all of the worker threads.
  78  *
  79  * Instead we have a helper queue who never has more than one thread
  80  * where we scheduler thread start operations.  This worker_start struct
  81  * is used to contain the work and hold a pointer to the queue that needs
  82  * another worker.
  83  */
  84 struct worker_start {
  85         struct btrfs_work work;
  86         struct btrfs_workers *queue;
  87 };
  88
  89 static void start_new_worker_func(struct btrfs_work *work)
  90 {
  91         struct worker_start *start;
  92         start = container_of(work, struct worker_start, work);
  93         __btrfs_start_workers(start->queue);
  94         kfree(start);
  95 }
  96
  97 /*
  98  * helper function to move a thread onto the idle list after it
  99  * has finished some requests.
 100  */
 101 static void check_idle_worker(struct btrfs_worker_thread *worker)
 102 {
 103         if (!worker->idle && atomic_read(&worker->num_pending) <
 104             worker->workers->idle_thresh / 2) {
 105                 unsigned long flags;
 106                 spin_lock_irqsave(&worker->workers->lock, flags);
 107                 worker->idle = 1;
 108
 109                 /* the list may be empty if the worker is just starting */
 110                 if (!list_empty(&worker->worker_list)) {
 111                         list_move(&worker->worker_list,
 112                                  &worker->workers->idle_list);
 113                 }
 114                 spin_unlock_irqrestore(&worker->workers->lock, flags);
 115         }
 116 }
 117
 118 /*
 119  * helper function to move a thread off the idle list after new
 120  * pending work is added.
 121  */
 122 static void check_busy_worker(struct btrfs_worker_thread *worker)
 123 {
 124         if (worker->idle && atomic_read(&worker->num_pending) >=
 125             worker->workers->idle_thresh) {
 126                 unsigned long flags;
 127                 spin_lock_irqsave(&worker->workers->lock, flags);
 128                 worker->idle = 0;
 129
 130                 if (!list_empty(&worker->worker_list)) {
 131                         list_move_tail(&worker->worker_list,
 132                                       &worker->workers->worker_list);
 133                 }
 134                 spin_unlock_irqrestore(&worker->workers->lock, flags);
 135         }
 136 }
 137
 138 static void check_pending_worker_creates(struct btrfs_worker_thread *worker)
 139 {
 140         struct btrfs_workers *workers = worker->workers;
 141         struct worker_start *start;
 142         unsigned long flags;
 143
 144         rmb();
 145         if (!workers->atomic_start_pending)
 146                 return;
 147
 148         start = kzalloc(sizeof(*start), GFP_NOFS);
 149         if (!start)
 150                 return;
 151
 152         start->work.func = start_new_worker_func;
 153         start->queue = workers;
 154
 155         spin_lock_irqsave(&workers->lock, flags);
 156         if (!workers->atomic_start_pending)
 157                 goto out;
 158
 159         workers->atomic_start_pending = 0;
 160         if (workers->num_workers + workers->num_workers_starting >=
 161             workers->max_workers)
 162                 goto out;
 163
 164         workers->num_workers_starting += 1;
 165         spin_unlock_irqrestore(&workers->lock, flags);
 166         btrfs_queue_worker(workers->atomic_worker_start, &start->work);
 167         return;
 168
 169 out:
 170         kfree(start);
 171         spin_unlock_irqrestore(&workers->lock, flags);
 172 }
 173
 174 static noinline void run_ordered_completions(struct btrfs_workers *workers,
 175                                             struct btrfs_work *work)
 176 {
 177         if (!workers->ordered)
 178                 return;
 179
 180         set_bit(WORK_DONE_BIT, &work->flags);
 181
 182         spin_lock(&workers->order_lock);
 183
 184         while (1) {
 185                 if (!list_empty(&workers->prio_order_list)) {
 186                         work = list_entry(workers->prio_order_list.next,
 187                                           struct btrfs_work, order_list);
 188                 } else if (!list_empty(&workers->order_list)) {
 189                         work = list_entry(workers->order_list.next,
 190                                           struct btrfs_work, order_list);
 191                 } else {
 192                         break;
 193                 }
 194                 if (!test_bit(WORK_DONE_BIT, &work->flags))
 195                         break;
 196
 197                 /* we are going to call the ordered done function, but
 198                  * we leave the work item on the list as a barrier so
 199                  * that later work items that are done don't have their
 200                  * functions called before this one returns
 201                  */
 202                 if (test_and_set_bit(WORK_ORDER_DONE_BIT, &work->flags))
 203                         break;
 204
 205                 spin_unlock(&workers->order_lock);
 206
 207                 work->ordered_func(work);
 208
 209                 /* now take the lock again and drop our item from the list */
 210                 spin_lock(&workers->order_lock);
 211                 list_del(&work->order_list);
 212                 spin_unlock(&workers->order_lock);
 213
 214                 /*
 215                  * we don't want to call the ordered free functions
 216                  * with the lock held though
 217                  */
 218                 work->ordered_free(work);
 219                 spin_lock(&workers->order_lock);
 220         }
 221
 222         spin_unlock(&workers->order_lock);
 223 }
 224
 225 static void put_worker(struct btrfs_worker_thread *worker)
 226 {
 227         if (atomic_dec_and_test(&worker->refs))
 228                 kfree(worker);
 229 }
 230
 231 static int try_worker_shutdown(struct btrfs_worker_thread *worker)
 232 {
 233         int freeit = 0;
 234
 235         spin_lock_irq(&worker->lock);
 236         spin_lock(&worker->workers->lock);
 237         if (worker->workers->num_workers > 1 &&
 238             worker->idle &&
 239             !worker->working &&
 240             !list_empty(&worker->worker_list) &&
 241             list_empty(&worker->prio_pending) &&
 242             list_empty(&worker->pending) &&
 243             atomic_read(&worker->num_pending) == 0) {
 244                 freeit = 1;
 245                 list_del_init(&worker->worker_list);
 246                 worker->workers->num_workers--;
 247         }
 248         spin_unlock(&worker->workers->lock);
 249         spin_unlock_irq(&worker->lock);
 250
 251         if (freeit)
 252                 put_worker(worker);
 253         return freeit;
 254 }
 255
 256 static struct btrfs_work *get_next_work(struct btrfs_worker_thread *worker,
 257                                         struct list_head *prio_head,
 258                                         struct list_head *head)
 259 {
 260         struct btrfs_work *work = NULL;
 261         struct list_head *cur = NULL;
 262
 263         if(!list_empty(prio_head))
 264                 cur = prio_head->next;
 265
 266         smp_mb();
 267         if (!list_empty(&worker->prio_pending))
 268                 goto refill;
 269
 270         if (!list_empty(head))
 271                 cur = head->next;
 272
 273         if (cur)
 274                 goto out;
 275
 276 refill:
 277         spin_lock_irq(&worker->lock);
 278         list_splice_tail_init(&worker->prio_pending, prio_head);
 279         list_splice_tail_init(&worker->pending, head);
 280
 281         if (!list_empty(prio_head))
 282                 cur = prio_head->next;
 283         else if (!list_empty(head))
 284                 cur = head->next;
 285         spin_unlock_irq(&worker->lock);
 286
 287         if (!cur)
 288                 goto out_fail;
 289
 290 out:
 291         work = list_entry(cur, struct btrfs_work, list);
 292
 293 out_fail:
 294         return work;
 295 }
 296
 297 /*
 298  * main loop for servicing work items
 299  */
 300 static int worker_loop(void *arg)
 301 {
 302         struct btrfs_worker_thread *worker = arg;
 303         struct list_head head;
 304         struct list_head prio_head;
 305         struct btrfs_work *work;
 306
 307         INIT_LIST_HEAD(&head);
 308         INIT_LIST_HEAD(&prio_head);
 309
 310         do {
 311 again:
 312                 while (1) {
 313
 314
 315                         work = get_next_work(worker, &prio_head, &head);
 316                         if (!work)
 317                                 break;
 318
 319                         list_del(&work->list);
 320                         clear_bit(WORK_QUEUED_BIT, &work->flags);
 321
 322                         work->worker = worker;
 323
 324                         work->func(work);
 325
 326                         atomic_dec(&worker->num_pending);
 327                         /*
 328                          * unless this is an ordered work queue,
 329                          * 'work' was probably freed by func above.
 330                          */
 331                         run_ordered_completions(worker->workers, work);
 332
 333                         check_pending_worker_creates(worker);
 334                         cond_resched();
 335                 }
 336
 337                 spin_lock_irq(&worker->lock);
 338                 check_idle_worker(worker);
 339
 340                 if (freezing(current)) {
 341                         worker->working = 0;
 342                         spin_unlock_irq(&worker->lock);
 343                         try_to_freeze();
 344                 } else {
 345                         spin_unlock_irq(&worker->lock);
 346                         if (!kthread_should_stop()) {
 347                                 cpu_relax();
 348                                 /*
 349                                  * we've dropped the lock, did someone else
 350                                  * jump_in?
 351                                  */
 352                                 smp_mb();
 353                                 if (!list_empty(&worker->pending) ||
 354                                     !list_empty(&worker->prio_pending))
 355                                         continue;
 356
 357                                 /*
 358                                  * this short schedule allows more work to
 359                                  * come in without the queue functions
 360                                  * needing to go through wake_up_process()
 361                                  *
 362                                  * worker->working is still 1, so nobody
 363                                  * is going to try and wake us up
 364                                  */
 365                                 schedule_timeout(1);
 366                                 smp_mb();
 367                                 if (!list_empty(&worker->pending) ||
 368                                     !list_empty(&worker->prio_pending))
 369                                         continue;
 370
 371                                 if (kthread_should_stop())
 372                                         break;
 373
 374                                 /* still no more work?, sleep for real */
 375                                 spin_lock_irq(&worker->lock);
 376                                 set_current_state(TASK_INTERRUPTIBLE);
 377                                 if (!list_empty(&worker->pending) ||
 378                                     !list_empty(&worker->prio_pending)) {
 379                                         spin_unlock_irq(&worker->lock);
 380                                         set_current_state(TASK_RUNNING);
 381                                         goto again;
 382                                 }
 383
 384                                 /*
 385                                  * this makes sure we get a wakeup when someone
 386                                  * adds something new to the queue
 387                                  */
 388                                 worker->working = 0;
 389                                 spin_unlock_irq(&worker->lock);
 390
 391                                 if (!kthread_should_stop()) {
 392                                         schedule_timeout(HZ * 120);
 393                                         if (!worker->working &&
 394                                             try_worker_shutdown(worker)) {
 395                                                 return 0;
 396                                         }
 397                                 }
 398                         }
 399                         __set_current_state(TASK_RUNNING);
 400                 }
 401         } while (!kthread_should_stop());
 402         return 0;
 403 }
 404
 405 /*
 406  * this will wait for all the worker threads to shutdown
 407  */
 408 void btrfs_stop_workers(struct btrfs_workers *workers)
 409 {
 410         struct list_head *cur;
 411         struct btrfs_worker_thread *worker;
 412         int can_stop;
 413
 414         spin_lock_irq(&workers->lock);
 415         list_splice_init(&workers->idle_list, &workers->worker_list);
 416         while (!list_empty(&workers->worker_list)) {
 417                 cur = workers->worker_list.next;
 418                 worker = list_entry(cur, struct btrfs_worker_thread,
 419                                     worker_list);
 420
 421                 atomic_inc(&worker->refs);
 422                 workers->num_workers -= 1;
 423                 if (!list_empty(&worker->worker_list)) {
 424                         list_del_init(&worker->worker_list);
 425                         put_worker(worker);
 426                         can_stop = 1;
 427                 } else
 428                         can_stop = 0;
 429                 spin_unlock_irq(&workers->lock);
 430                 if (can_stop)
 431                         kthread_stop(worker->task);
 432                 spin_lock_irq(&workers->lock);
 433                 put_worker(worker);
 434         }
 435         spin_unlock_irq(&workers->lock);
 436 }
 437
 438 /*
 439  * simple init on struct btrfs_workers
 440  */
 441 void btrfs_init_workers(struct btrfs_workers *workers, char *name, int max,
 442                         struct btrfs_workers *async_helper)
 443 {
 444         workers->num_workers = 0;
 445         workers->num_workers_starting = 0;
 446         INIT_LIST_HEAD(&workers->worker_list);
 447         INIT_LIST_HEAD(&workers->idle_list);
 448         INIT_LIST_HEAD(&workers->order_list);
 449         INIT_LIST_HEAD(&workers->prio_order_list);
 450         spin_lock_init(&workers->lock);
 451         spin_lock_init(&workers->order_lock);
 452         workers->max_workers = max;
 453         workers->idle_thresh = 32;
 454         workers->name = name;
 455         workers->ordered = 0;
 456         workers->atomic_start_pending = 0;
 457         workers->atomic_worker_start = async_helper;
 458 }
 459
 460 /*
 461  * starts new worker threads.  This does not enforce the max worker
 462  * count in case you need to temporarily go past it.
 463  */
 464 static int __btrfs_start_workers(struct btrfs_workers *workers)
 465 {
 466         struct btrfs_worker_thread *worker;
 467         int ret = 0;
 468
 469         worker = kzalloc(sizeof(*worker), GFP_NOFS);
 470         if (!worker) {
 471                 ret = -ENOMEM;
 472                 goto fail;
 473         }
 474
 475         INIT_LIST_HEAD(&worker->pending);
 476         INIT_LIST_HEAD(&worker->prio_pending);
 477         INIT_LIST_HEAD(&worker->worker_list);
 478         spin_lock_init(&worker->lock);
 479
 480         atomic_set(&worker->num_pending, 0);
 481         atomic_set(&worker->refs, 1);
 482         worker->workers = workers;
 483         worker->task = kthread_run(worker_loop, worker,
 484                                    "btrfs-%s-%d", workers->name,
 485                                    workers->num_workers + 1);
 486         if (IS_ERR(worker->task)) {
 487                 ret = PTR_ERR(worker->task);
 488                 kfree(worker);
 489                 goto fail;
 490         }
 491         spin_lock_irq(&workers->lock);
 492         list_add_tail(&worker->worker_list, &workers->idle_list);
 493         worker->idle = 1;
 494         workers->num_workers++;
 495         workers->num_workers_starting--;
 496         WARN_ON(workers->num_workers_starting < 0);
 497         spin_unlock_irq(&workers->lock);
 498
 499         return 0;
 500 fail:
 501         spin_lock_irq(&workers->lock);
 502         workers->num_workers_starting--;
 503         spin_unlock_irq(&workers->lock);
 504         return ret;
 505 }
 506
 507 int btrfs_start_workers(struct btrfs_workers *workers)
 508 {
 509         spin_lock_irq(&workers->lock);
 510         workers->num_workers_starting++;
 511         spin_unlock_irq(&workers->lock);
 512         return __btrfs_start_workers(workers);
 513 }
 514
 515 /*
 516  * run through the list and find a worker thread that doesn't have a lot
 517  * to do right now.  This can return null if we aren't yet at the thread
 518  * count limit and all of the threads are busy.
 519  */
 520 static struct btrfs_worker_thread *next_worker(struct btrfs_workers *workers)
 521 {
 522         struct btrfs_worker_thread *worker;
 523         struct list_head *next;
 524         int enforce_min;
 525
 526         enforce_min = (workers->num_workers + workers->num_workers_starting) <
 527                 workers->max_workers;
 528
 529         /*
 530          * if we find an idle thread, don't move it to the end of the
 531          * idle list.  This improves the chance that the next submission
 532          * will reuse the same thread, and maybe catch it while it is still
 533          * working
 534          */
 535         if (!list_empty(&workers->idle_list)) {
 536                 next = workers->idle_list.next;
 537                 worker = list_entry(next, struct btrfs_worker_thread,
 538                                     worker_list);
 539                 return worker;
 540         }
 541         if (enforce_min || list_empty(&workers->worker_list))
 542                 return NULL;
 543
 544         /*
 545          * if we pick a busy task, move the task to the end of the list.
 546          * hopefully this will keep things somewhat evenly balanced.
 547          * Do the move in batches based on the sequence number.  This groups
 548          * requests submitted at roughly the same time onto the same worker.
 549          */
 550         next = workers->worker_list.next;
 551         worker = list_entry(next, struct btrfs_worker_thread, worker_list);
 552         worker->sequence++;
 553
 554         if (worker->sequence % workers->idle_thresh == 0)
 555                 list_move_tail(next, &workers->worker_list);
 556         return worker;
 557 }
 558
 559 /*
 560  * selects a worker thread to take the next job.  This will either find
 561  * an idle worker, start a new worker up to the max count, or just return
 562  * one of the existing busy workers.
 563  */
 564 static struct btrfs_worker_thread *find_worker(struct btrfs_workers *workers)
 565 {
 566         struct btrfs_worker_thread *worker;
 567         unsigned long flags;
 568         struct list_head *fallback;
 569         int ret;
 570
 571         spin_lock_irqsave(&workers->lock, flags);
 572 again:
 573         worker = next_worker(workers);
 574
 575         if (!worker) {
 576                 if (workers->num_workers + workers->num_workers_starting >=
 577                     workers->max_workers) {
 578                         goto fallback;
 579                 } else if (workers->atomic_worker_start) {
 580                         workers->atomic_start_pending = 1;
 581                         goto fallback;
 582                 } else {
 583                         workers->num_workers_starting++;
 584                         spin_unlock_irqrestore(&workers->lock, flags);
 585                         /* we're below the limit, start another worker */
 586                         ret = __btrfs_start_workers(workers);
 587                         spin_lock_irqsave(&workers->lock, flags);
 588                         if (ret)
 589                                 goto fallback;
 590                         goto again;
 591                 }
 592         }
 593         goto found;
 594
 595 fallback:
 596         fallback = NULL;
 597         /*
 598          * we have failed to find any workers, just
 599          * return the first one we can find.
 600          */
 601         if (!list_empty(&workers->worker_list))
 602                 fallback = workers->worker_list.next;
 603         if (!list_empty(&workers->idle_list))
 604                 fallback = workers->idle_list.next;
 605         BUG_ON(!fallback);
 606         worker = list_entry(fallback,
 607                   struct btrfs_worker_thread, worker_list);
 608 found:
 609         /*
 610          * this makes sure the worker doesn't exit before it is placed
 611          * onto a busy/idle list
 612          */
 613         atomic_inc(&worker->num_pending);
 614         spin_unlock_irqrestore(&workers->lock, flags);
 615         return worker;
 616 }
 617
 618 /*
 619  * btrfs_requeue_work just puts the work item back on the tail of the list
 620  * it was taken from.  It is intended for use with long running work functions
 621  * that make some progress and want to give the cpu up for others.
 622  */
 623 void btrfs_requeue_work(struct btrfs_work *work)
 624 {
 625         struct btrfs_worker_thread *worker = work->worker;
 626         unsigned long flags;
 627         int wake = 0;
 628
 629         if (test_and_set_bit(WORK_QUEUED_BIT, &work->flags))
 630                 return;
 631
 632         spin_lock_irqsave(&worker->lock, flags);
 633         if (test_bit(WORK_HIGH_PRIO_BIT, &work->flags))
 634                 list_add_tail(&work->list, &worker->prio_pending);
 635         else
 636                 list_add_tail(&work->list, &worker->pending);
 637         atomic_inc(&worker->num_pending);
 638
 639         /* by definition we're busy, take ourselves off the idle
 640          * list
 641          */
 642         if (worker->idle) {
 643                 spin_lock(&worker->workers->lock);
 644                 worker->idle = 0;
 645                 list_move_tail(&worker->worker_list,
 646                               &worker->workers->worker_list);
 647                 spin_unlock(&worker->workers->lock);
 648         }
 649         if (!worker->working) {
 650                 wake = 1;
 651                 worker->working = 1;
 652         }
 653
 654         if (wake)
 655                 wake_up_process(worker->task);
 656         spin_unlock_irqrestore(&worker->lock, flags);
 657 }
 658
 659 void btrfs_set_work_high_prio(struct btrfs_work *work)
 660 {
 661         set_bit(WORK_HIGH_PRIO_BIT, &work->flags);
 662 }
 663
 664 /*
 665  * places a struct btrfs_work into the pending queue of one of the kthreads
 666  */
 667 void btrfs_queue_worker(struct btrfs_workers *workers, struct btrfs_work *work)
 668 {
 669         struct btrfs_worker_thread *worker;
 670         unsigned long flags;
 671         int wake = 0;
 672
 673         /* don't requeue something already on a list */
 674         if (test_and_set_bit(WORK_QUEUED_BIT, &work->flags))
 675                 return;
 676
 677         worker = find_worker(workers);
 678         if (workers->ordered) {
 679                 /*
 680                  * you're not allowed to do ordered queues from an
 681                  * interrupt handler
 682                  */
 683                 spin_lock(&workers->order_lock);
 684                 if (test_bit(WORK_HIGH_PRIO_BIT, &work->flags)) {
 685                         list_add_tail(&work->order_list,
 686                                       &workers->prio_order_list);
 687                 } else {
 688                         list_add_tail(&work->order_list, &workers->order_list);
 689                 }
 690                 spin_unlock(&workers->order_lock);
 691         } else {
 692                 INIT_LIST_HEAD(&work->order_list);
 693         }
 694
 695         spin_lock_irqsave(&worker->lock, flags);
 696
 697         if (test_bit(WORK_HIGH_PRIO_BIT, &work->flags))
 698                 list_add_tail(&work->list, &worker->prio_pending);
 699         else
 700                 list_add_tail(&work->list, &worker->pending);
 701         check_busy_worker(worker);
 702
 703         /*
 704          * avoid calling into wake_up_process if this thread has already
 705          * been kicked
 706          */
 707         if (!worker->working)
 708                 wake = 1;
 709         worker->working = 1;
 710
 711         if (wake)
 712                 wake_up_process(worker->task);
 713         spin_unlock_irqrestore(&worker->lock, flags);
 714 }