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