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 int run_ordered_completions(struct btrfs_workers *workers,
 175                                             struct btrfs_work *work)
 176 {
 177         if (!workers->ordered)
 178                 return 0;
 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         return 0;
 224 }
 225
 226 static void put_worker(struct btrfs_worker_thread *worker)
 227 {
 228         if (atomic_dec_and_test(&worker->refs))
 229                 kfree(worker);
 230 }
 231
 232 static int try_worker_shutdown(struct btrfs_worker_thread *worker)
 233 {
 234         int freeit = 0;
 235
 236         spin_lock_irq(&worker->lock);
 237         spin_lock(&worker->workers->lock);
 238         if (worker->workers->num_workers > 1 &&
 239             worker->idle &&
 240             !worker->working &&
 241             !list_empty(&worker->worker_list) &&
 242             list_empty(&worker->prio_pending) &&
 243             list_empty(&worker->pending) &&
 244             atomic_read(&worker->num_pending) == 0) {
 245                 freeit = 1;
 246                 list_del_init(&worker->worker_list);
 247                 worker->workers->num_workers--;
 248         }
 249         spin_unlock(&worker->workers->lock);
 250         spin_unlock_irq(&worker->lock);
 251
 252         if (freeit)
 253                 put_worker(worker);
 254         return freeit;
 255 }
 256
 257 static struct btrfs_work *get_next_work(struct btrfs_worker_thread *worker,
 258                                         struct list_head *prio_head,
 259                                         struct list_head *head)
 260 {
 261         struct btrfs_work *work = NULL;
 262         struct list_head *cur = NULL;
 263
 264         if(!list_empty(prio_head))
 265                 cur = prio_head->next;
 266
 267         smp_mb();
 268         if (!list_empty(&worker->prio_pending))
 269                 goto refill;
 270
 271         if (!list_empty(head))
 272                 cur = head->next;
 273
 274         if (cur)
 275                 goto out;
 276
 277 refill:
 278         spin_lock_irq(&worker->lock);
 279         list_splice_tail_init(&worker->prio_pending, prio_head);
 280         list_splice_tail_init(&worker->pending, head);
 281
 282         if (!list_empty(prio_head))
 283                 cur = prio_head->next;
 284         else if (!list_empty(head))
 285                 cur = head->next;
 286         spin_unlock_irq(&worker->lock);
 287
 288         if (!cur)
 289                 goto out_fail;
 290
 291 out:
 292         work = list_entry(cur, struct btrfs_work, list);
 293
 294 out_fail:
 295         return work;
 296 }
 297
 298 /*
 299  * main loop for servicing work items
 300  */
 301 static int worker_loop(void *arg)
 302 {
 303         struct btrfs_worker_thread *worker = arg;
 304         struct list_head head;
 305         struct list_head prio_head;
 306         struct btrfs_work *work;
 307
 308         INIT_LIST_HEAD(&head);
 309         INIT_LIST_HEAD(&prio_head);
 310
 311         do {
 312 again:
 313                 while (1) {
 314
 315
 316                         work = get_next_work(worker, &prio_head, &head);
 317                         if (!work)
 318                                 break;
 319
 320                         list_del(&work->list);
 321                         clear_bit(WORK_QUEUED_BIT, &work->flags);
 322
 323                         work->worker = worker;
 324
 325                         work->func(work);
 326
 327                         atomic_dec(&worker->num_pending);
 328                         /*
 329                          * unless this is an ordered work queue,
 330                          * 'work' was probably freed by func above.
 331                          */
 332                         run_ordered_completions(worker->workers, work);
 333
 334                         check_pending_worker_creates(worker);
 335                         cond_resched();
 336                 }
 337
 338                 spin_lock_irq(&worker->lock);
 339                 check_idle_worker(worker);
 340
 341                 if (freezing(current)) {
 342                         worker->working = 0;
 343                         spin_unlock_irq(&worker->lock);
 344                         refrigerator();
 345                 } else {
 346                         spin_unlock_irq(&worker->lock);
 347                         if (!kthread_should_stop()) {
 348                                 cpu_relax();
 349                                 /*
 350                                  * we've dropped the lock, did someone else
 351                                  * jump_in?
 352                                  */
 353                                 smp_mb();
 354                                 if (!list_empty(&worker->pending) ||
 355                                     !list_empty(&worker->prio_pending))
 356                                         continue;
 357
 358                                 /*
 359                                  * this short schedule allows more work to
 360                                  * come in without the queue functions
 361                                  * needing to go through wake_up_process()
 362                                  *
 363                                  * worker->working is still 1, so nobody
 364                                  * is going to try and wake us up
 365                                  */
 366                                 schedule_timeout(1);
 367                                 smp_mb();
 368                                 if (!list_empty(&worker->pending) ||
 369                                     !list_empty(&worker->prio_pending))
 370                                         continue;
 371
 372                                 if (kthread_should_stop())
 373                                         break;
 374
 375                                 /* still no more work?, sleep for real */
 376                                 spin_lock_irq(&worker->lock);
 377                                 set_current_state(TASK_INTERRUPTIBLE);
 378                                 if (!list_empty(&worker->pending) ||
 379                                     !list_empty(&worker->prio_pending)) {
 380                                         spin_unlock_irq(&worker->lock);
 381                                         set_current_state(TASK_RUNNING);
 382                                         goto again;
 383                                 }
 384
 385                                 /*
 386                                  * this makes sure we get a wakeup when someone
 387                                  * adds something new to the queue
 388                                  */
 389                                 worker->working = 0;
 390                                 spin_unlock_irq(&worker->lock);
 391
 392                                 if (!kthread_should_stop()) {
 393                                         schedule_timeout(HZ * 120);
 394                                         if (!worker->working &&
 395                                             try_worker_shutdown(worker)) {
 396                                                 return 0;
 397                                         }
 398                                 }
 399                         }
 400                         __set_current_state(TASK_RUNNING);
 401                 }
 402         } while (!kthread_should_stop());
 403         return 0;
 404 }
 405
 406 /*
 407  * this will wait for all the worker threads to shutdown
 408  */
 409 int btrfs_stop_workers(struct btrfs_workers *workers)
 410 {
 411         struct list_head *cur;
 412         struct btrfs_worker_thread *worker;
 413         int can_stop;
 414
 415         spin_lock_irq(&workers->lock);
 416         list_splice_init(&workers->idle_list, &workers->worker_list);
 417         while (!list_empty(&workers->worker_list)) {
 418                 cur = workers->worker_list.next;
 419                 worker = list_entry(cur, struct btrfs_worker_thread,
 420                                     worker_list);
 421
 422                 atomic_inc(&worker->refs);
 423                 workers->num_workers -= 1;
 424                 if (!list_empty(&worker->worker_list)) {
 425                         list_del_init(&worker->worker_list);
 426                         put_worker(worker);
 427                         can_stop = 1;
 428                 } else
 429                         can_stop = 0;
 430                 spin_unlock_irq(&workers->lock);
 431                 if (can_stop)
 432                         kthread_stop(worker->task);
 433                 spin_lock_irq(&workers->lock);
 434                 put_worker(worker);
 435         }
 436         spin_unlock_irq(&workers->lock);
 437         return 0;
 438 }
 439
 440 /*
 441  * simple init on struct btrfs_workers
 442  */
 443 void btrfs_init_workers(struct btrfs_workers *workers, char *name, int max,
 444                         struct btrfs_workers *async_helper)
 445 {
 446         workers->num_workers = 0;
 447         workers->num_workers_starting = 0;
 448         INIT_LIST_HEAD(&workers->worker_list);
 449         INIT_LIST_HEAD(&workers->idle_list);
 450         INIT_LIST_HEAD(&workers->order_list);
 451         INIT_LIST_HEAD(&workers->prio_order_list);
 452         spin_lock_init(&workers->lock);
 453         spin_lock_init(&workers->order_lock);
 454         workers->max_workers = max;
 455         workers->idle_thresh = 32;
 456         workers->name = name;
 457         workers->ordered = 0;
 458         workers->atomic_start_pending = 0;
 459         workers->atomic_worker_start = async_helper;
 460 }
 461
 462 /*
 463  * starts new worker threads.  This does not enforce the max worker
 464  * count in case you need to temporarily go past it.
 465  */
 466 static int __btrfs_start_workers(struct btrfs_workers *workers)
 467 {
 468         struct btrfs_worker_thread *worker;
 469         int ret = 0;
 470
 471         worker = kzalloc(sizeof(*worker), GFP_NOFS);
 472         if (!worker) {
 473                 ret = -ENOMEM;
 474                 goto fail;
 475         }
 476
 477         INIT_LIST_HEAD(&worker->pending);
 478         INIT_LIST_HEAD(&worker->prio_pending);
 479         INIT_LIST_HEAD(&worker->worker_list);
 480         spin_lock_init(&worker->lock);
 481
 482         atomic_set(&worker->num_pending, 0);
 483         atomic_set(&worker->refs, 1);
 484         worker->workers = workers;
 485         worker->task = kthread_run(worker_loop, worker,
 486                                    "btrfs-%s-%d", workers->name,
 487                                    workers->num_workers + 1);
 488         if (IS_ERR(worker->task)) {
 489                 ret = PTR_ERR(worker->task);
 490                 kfree(worker);
 491                 goto fail;
 492         }
 493         spin_lock_irq(&workers->lock);
 494         list_add_tail(&worker->worker_list, &workers->idle_list);
 495         worker->idle = 1;
 496         workers->num_workers++;
 497         workers->num_workers_starting--;
 498         WARN_ON(workers->num_workers_starting < 0);
 499         spin_unlock_irq(&workers->lock);
 500
 501         return 0;
 502 fail:
 503         spin_lock_irq(&workers->lock);
 504         workers->num_workers_starting--;
 505         spin_unlock_irq(&workers->lock);
 506         return ret;
 507 }
 508
 509 int btrfs_start_workers(struct btrfs_workers *workers)
 510 {
 511         spin_lock_irq(&workers->lock);
 512         workers->num_workers_starting++;
 513         spin_unlock_irq(&workers->lock);
 514         return __btrfs_start_workers(workers);
 515 }
 516
 517 /*
 518  * run through the list and find a worker thread that doesn't have a lot
 519  * to do right now.  This can return null if we aren't yet at the thread
 520  * count limit and all of the threads are busy.
 521  */
 522 static struct btrfs_worker_thread *next_worker(struct btrfs_workers *workers)
 523 {
 524         struct btrfs_worker_thread *worker;
 525         struct list_head *next;
 526         int enforce_min;
 527
 528         enforce_min = (workers->num_workers + workers->num_workers_starting) <
 529                 workers->max_workers;
 530
 531         /*
 532          * if we find an idle thread, don't move it to the end of the
 533          * idle list.  This improves the chance that the next submission
 534          * will reuse the same thread, and maybe catch it while it is still
 535          * working
 536          */
 537         if (!list_empty(&workers->idle_list)) {
 538                 next = workers->idle_list.next;
 539                 worker = list_entry(next, struct btrfs_worker_thread,
 540                                     worker_list);
 541                 return worker;
 542         }
 543         if (enforce_min || list_empty(&workers->worker_list))
 544                 return NULL;
 545
 546         /*
 547          * if we pick a busy task, move the task to the end of the list.
 548          * hopefully this will keep things somewhat evenly balanced.
 549          * Do the move in batches based on the sequence number.  This groups
 550          * requests submitted at roughly the same time onto the same worker.
 551          */
 552         next = workers->worker_list.next;
 553         worker = list_entry(next, struct btrfs_worker_thread, worker_list);
 554         worker->sequence++;
 555
 556         if (worker->sequence % workers->idle_thresh == 0)
 557                 list_move_tail(next, &workers->worker_list);
 558         return worker;
 559 }
 560
 561 /*
 562  * selects a worker thread to take the next job.  This will either find
 563  * an idle worker, start a new worker up to the max count, or just return
 564  * one of the existing busy workers.
 565  */
 566 static struct btrfs_worker_thread *find_worker(struct btrfs_workers *workers)
 567 {
 568         struct btrfs_worker_thread *worker;
 569         unsigned long flags;
 570         struct list_head *fallback;
 571         int ret;
 572
 573         spin_lock_irqsave(&workers->lock, flags);
 574 again:
 575         worker = next_worker(workers);
 576
 577         if (!worker) {
 578                 if (workers->num_workers + workers->num_workers_starting >=
 579                     workers->max_workers) {
 580                         goto fallback;
 581                 } else if (workers->atomic_worker_start) {
 582                         workers->atomic_start_pending = 1;
 583                         goto fallback;
 584                 } else {
 585                         workers->num_workers_starting++;
 586                         spin_unlock_irqrestore(&workers->lock, flags);
 587                         /* we're below the limit, start another worker */
 588                         ret = __btrfs_start_workers(workers);
 589                         spin_lock_irqsave(&workers->lock, flags);
 590                         if (ret)
 591                                 goto fallback;
 592                         goto again;
 593                 }
 594         }
 595         goto found;
 596
 597 fallback:
 598         fallback = NULL;
 599         /*
 600          * we have failed to find any workers, just
 601          * return the first one we can find.
 602          */
 603         if (!list_empty(&workers->worker_list))
 604                 fallback = workers->worker_list.next;
 605         if (!list_empty(&workers->idle_list))
 606                 fallback = workers->idle_list.next;
 607         BUG_ON(!fallback);
 608         worker = list_entry(fallback,
 609                   struct btrfs_worker_thread, worker_list);
 610 found:
 611         /*
 612          * this makes sure the worker doesn't exit before it is placed
 613          * onto a busy/idle list
 614          */
 615         atomic_inc(&worker->num_pending);
 616         spin_unlock_irqrestore(&workers->lock, flags);
 617         return worker;
 618 }
 619
 620 /*
 621  * btrfs_requeue_work just puts the work item back on the tail of the list
 622  * it was taken from.  It is intended for use with long running work functions
 623  * that make some progress and want to give the cpu up for others.
 624  */
 625 int btrfs_requeue_work(struct btrfs_work *work)
 626 {
 627         struct btrfs_worker_thread *worker = work->worker;
 628         unsigned long flags;
 629         int wake = 0;
 630
 631         if (test_and_set_bit(WORK_QUEUED_BIT, &work->flags))
 632                 goto out;
 633
 634         spin_lock_irqsave(&worker->lock, flags);
 635         if (test_bit(WORK_HIGH_PRIO_BIT, &work->flags))
 636                 list_add_tail(&work->list, &worker->prio_pending);
 637         else
 638                 list_add_tail(&work->list, &worker->pending);
 639         atomic_inc(&worker->num_pending);
 640
 641         /* by definition we're busy, take ourselves off the idle
 642          * list
 643          */
 644         if (worker->idle) {
 645                 spin_lock(&worker->workers->lock);
 646                 worker->idle = 0;
 647                 list_move_tail(&worker->worker_list,
 648                               &worker->workers->worker_list);
 649                 spin_unlock(&worker->workers->lock);
 650         }
 651         if (!worker->working) {
 652                 wake = 1;
 653                 worker->working = 1;
 654         }
 655
 656         if (wake)
 657                 wake_up_process(worker->task);
 658         spin_unlock_irqrestore(&worker->lock, flags);
 659 out:
 660
 661         return 0;
 662 }
 663
 664 void btrfs_set_work_high_prio(struct btrfs_work *work)
 665 {
 666         set_bit(WORK_HIGH_PRIO_BIT, &work->flags);
 667 }
 668
 669 /*
 670  * places a struct btrfs_work into the pending queue of one of the kthreads
 671  */
 672 void btrfs_queue_worker(struct btrfs_workers *workers, struct btrfs_work *work)
 673 {
 674         struct btrfs_worker_thread *worker;
 675         unsigned long flags;
 676         int wake = 0;
 677
 678         /* don't requeue something already on a list */
 679         if (test_and_set_bit(WORK_QUEUED_BIT, &work->flags))
 680                 return;
 681
 682         worker = find_worker(workers);
 683         if (workers->ordered) {
 684                 /*
 685                  * you're not allowed to do ordered queues from an
 686                  * interrupt handler
 687                  */
 688                 spin_lock(&workers->order_lock);
 689                 if (test_bit(WORK_HIGH_PRIO_BIT, &work->flags)) {
 690                         list_add_tail(&work->order_list,
 691                                       &workers->prio_order_list);
 692                 } else {
 693                         list_add_tail(&work->order_list, &workers->order_list);
 694                 }
 695                 spin_unlock(&workers->order_lock);
 696         } else {
 697                 INIT_LIST_HEAD(&work->order_list);
 698         }
 699
 700         spin_lock_irqsave(&worker->lock, flags);
 701
 702         if (test_bit(WORK_HIGH_PRIO_BIT, &work->flags))
 703                 list_add_tail(&work->list, &worker->prio_pending);
 704         else
 705                 list_add_tail(&work->list, &worker->pending);
 706         check_busy_worker(worker);
 707
 708         /*
 709          * avoid calling into wake_up_process if this thread has already
 710          * been kicked
 711          */
 712         if (!worker->working)
 713                 wake = 1;
 714         worker->working = 1;
 715
 716         if (wake)
 717                 wake_up_process(worker->task);
 718         spin_unlock_irqrestore(&worker->lock, flags);
 719 }