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
  29 /*
  30  * container for the kthread task pointer and the list of pending work
  31  * One of these is allocated per thread.
  32  */
  33 struct btrfs_worker_thread {
  34         /* pool we belong to */
  35         struct btrfs_workers *workers;
  36
  37         /* list of struct btrfs_work that are waiting for service */
  38         struct list_head pending;
  39
  40         /* list of worker threads from struct btrfs_workers */
  41         struct list_head worker_list;
  42
  43         /* kthread */
  44         struct task_struct *task;
  45
  46         /* number of things on the pending list */
  47         atomic_t num_pending;
  48
  49         unsigned long sequence;
  50
  51         /* protects the pending list. */
  52         spinlock_t lock;
  53
  54         /* set to non-zero when this thread is already awake and kicking */
  55         int working;
  56
  57         /* are we currently idle */
  58         int idle;
  59 };
  60
  61 /*
  62  * helper function to move a thread onto the idle list after it
  63  * has finished some requests.
  64  */
  65 static void check_idle_worker(struct btrfs_worker_thread *worker)
  66 {
  67         if (!worker->idle && atomic_read(&worker->num_pending) <
  68             worker->workers->idle_thresh / 2) {
  69                 unsigned long flags;
  70                 spin_lock_irqsave(&worker->workers->lock, flags);
  71                 worker->idle = 1;
  72                 list_move(&worker->worker_list, &worker->workers->idle_list);
  73                 spin_unlock_irqrestore(&worker->workers->lock, flags);
  74         }
  75 }
  76
  77 /*
  78  * helper function to move a thread off the idle list after new
  79  * pending work is added.
  80  */
  81 static void check_busy_worker(struct btrfs_worker_thread *worker)
  82 {
  83         if (worker->idle && atomic_read(&worker->num_pending) >=
  84             worker->workers->idle_thresh) {
  85                 unsigned long flags;
  86                 spin_lock_irqsave(&worker->workers->lock, flags);
  87                 worker->idle = 0;
  88                 list_move_tail(&worker->worker_list,
  89                                &worker->workers->worker_list);
  90                 spin_unlock_irqrestore(&worker->workers->lock, flags);
  91         }
  92 }
  93
  94 static noinline int run_ordered_completions(struct btrfs_workers *workers,
  95                                             struct btrfs_work *work)
  96 {
  97         unsigned long flags;
  98
  99         if (!workers->ordered)
 100                 return 0;
 101
 102         set_bit(WORK_DONE_BIT, &work->flags);
 103
 104         spin_lock_irqsave(&workers->lock, flags);
 105
 106         while (!list_empty(&workers->order_list)) {
 107                 work = list_entry(workers->order_list.next,
 108                                   struct btrfs_work, order_list);
 109
 110                 if (!test_bit(WORK_DONE_BIT, &work->flags))
 111                         break;
 112
 113                 /* we are going to call the ordered done function, but
 114                  * we leave the work item on the list as a barrier so
 115                  * that later work items that are done don't have their
 116                  * functions called before this one returns
 117                  */
 118                 if (test_and_set_bit(WORK_ORDER_DONE_BIT, &work->flags))
 119                         break;
 120
 121                 spin_unlock_irqrestore(&workers->lock, flags);
 122
 123                 work->ordered_func(work);
 124
 125                 /* now take the lock again and call the freeing code */
 126                 spin_lock_irqsave(&workers->lock, flags);
 127                 list_del(&work->order_list);
 128                 work->ordered_free(work);
 129         }
 130
 131         spin_unlock_irqrestore(&workers->lock, flags);
 132         return 0;
 133 }
 134
 135 /*
 136  * main loop for servicing work items
 137  */
 138 static int worker_loop(void *arg)
 139 {
 140         struct btrfs_worker_thread *worker = arg;
 141         struct list_head *cur;
 142         struct btrfs_work *work;
 143         do {
 144                 spin_lock_irq(&worker->lock);
 145 again_locked:
 146                 while (!list_empty(&worker->pending)) {
 147                         cur = worker->pending.next;
 148                         work = list_entry(cur, struct btrfs_work, list);
 149                         list_del(&work->list);
 150                         clear_bit(WORK_QUEUED_BIT, &work->flags);
 151
 152                         work->worker = worker;
 153                         spin_unlock_irq(&worker->lock);
 154
 155                         work->func(work);
 156
 157                         atomic_dec(&worker->num_pending);
 158                         /*
 159                          * unless this is an ordered work queue,
 160                          * 'work' was probably freed by func above.
 161                          */
 162                         run_ordered_completions(worker->workers, work);
 163
 164                         spin_lock_irq(&worker->lock);
 165                         check_idle_worker(worker);
 166
 167                 }
 168                 if (freezing(current)) {
 169                         worker->working = 0;
 170                         spin_unlock_irq(&worker->lock);
 171                         refrigerator();
 172                 } else {
 173                         spin_unlock_irq(&worker->lock);
 174                         if (!kthread_should_stop()) {
 175                                 cpu_relax();
 176                                 /*
 177                                  * we've dropped the lock, did someone else
 178                                  * jump_in?
 179                                  */
 180                                 smp_mb();
 181                                 if (!list_empty(&worker->pending))
 182                                         continue;
 183
 184                                 /*
 185                                  * this short schedule allows more work to
 186                                  * come in without the queue functions
 187                                  * needing to go through wake_up_process()
 188                                  *
 189                                  * worker->working is still 1, so nobody
 190                                  * is going to try and wake us up
 191                                  */
 192                                 schedule_timeout(1);
 193                                 smp_mb();
 194                                 if (!list_empty(&worker->pending))
 195                                         continue;
 196
 197                                 if (kthread_should_stop())
 198                                         break;
 199
 200                                 /* still no more work?, sleep for real */
 201                                 spin_lock_irq(&worker->lock);
 202                                 set_current_state(TASK_INTERRUPTIBLE);
 203                                 if (!list_empty(&worker->pending))
 204                                         goto again_locked;
 205
 206                                 /*
 207                                  * this makes sure we get a wakeup when someone
 208                                  * adds something new to the queue
 209                                  */
 210                                 worker->working = 0;
 211                                 spin_unlock_irq(&worker->lock);
 212
 213                                 if (!kthread_should_stop())
 214                                         schedule();
 215                         }
 216                         __set_current_state(TASK_RUNNING);
 217                 }
 218         } while (!kthread_should_stop());
 219         return 0;
 220 }
 221
 222 /*
 223  * this will wait for all the worker threads to shutdown
 224  */
 225 int btrfs_stop_workers(struct btrfs_workers *workers)
 226 {
 227         struct list_head *cur;
 228         struct btrfs_worker_thread *worker;
 229
 230         list_splice_init(&workers->idle_list, &workers->worker_list);
 231         while (!list_empty(&workers->worker_list)) {
 232                 cur = workers->worker_list.next;
 233                 worker = list_entry(cur, struct btrfs_worker_thread,
 234                                     worker_list);
 235                 kthread_stop(worker->task);
 236                 list_del(&worker->worker_list);
 237                 kfree(worker);
 238         }
 239         return 0;
 240 }
 241
 242 /*
 243  * simple init on struct btrfs_workers
 244  */
 245 void btrfs_init_workers(struct btrfs_workers *workers, char *name, int max)
 246 {
 247         workers->num_workers = 0;
 248         INIT_LIST_HEAD(&workers->worker_list);
 249         INIT_LIST_HEAD(&workers->idle_list);
 250         INIT_LIST_HEAD(&workers->order_list);
 251         spin_lock_init(&workers->lock);
 252         workers->max_workers = max;
 253         workers->idle_thresh = 32;
 254         workers->name = name;
 255         workers->ordered = 0;
 256 }
 257
 258 /*
 259  * starts new worker threads.  This does not enforce the max worker
 260  * count in case you need to temporarily go past it.
 261  */
 262 int btrfs_start_workers(struct btrfs_workers *workers, int num_workers)
 263 {
 264         struct btrfs_worker_thread *worker;
 265         int ret = 0;
 266         int i;
 267
 268         for (i = 0; i < num_workers; i++) {
 269                 worker = kzalloc(sizeof(*worker), GFP_NOFS);
 270                 if (!worker) {
 271                         ret = -ENOMEM;
 272                         goto fail;
 273                 }
 274
 275                 INIT_LIST_HEAD(&worker->pending);
 276                 INIT_LIST_HEAD(&worker->worker_list);
 277                 spin_lock_init(&worker->lock);
 278                 atomic_set(&worker->num_pending, 0);
 279                 worker->task = kthread_run(worker_loop, worker,
 280                                            "btrfs-%s-%d", workers->name,
 281                                            workers->num_workers + i);
 282                 worker->workers = workers;
 283                 if (IS_ERR(worker->task)) {
 284                         kfree(worker);
 285                         ret = PTR_ERR(worker->task);
 286                         goto fail;
 287                 }
 288
 289                 spin_lock_irq(&workers->lock);
 290                 list_add_tail(&worker->worker_list, &workers->idle_list);
 291                 worker->idle = 1;
 292                 workers->num_workers++;
 293                 spin_unlock_irq(&workers->lock);
 294         }
 295         return 0;
 296 fail:
 297         btrfs_stop_workers(workers);
 298         return ret;
 299 }
 300
 301 /*
 302  * run through the list and find a worker thread that doesn't have a lot
 303  * to do right now.  This can return null if we aren't yet at the thread
 304  * count limit and all of the threads are busy.
 305  */
 306 static struct btrfs_worker_thread *next_worker(struct btrfs_workers *workers)
 307 {
 308         struct btrfs_worker_thread *worker;
 309         struct list_head *next;
 310         int enforce_min = workers->num_workers < workers->max_workers;
 311
 312         /*
 313          * if we find an idle thread, don't move it to the end of the
 314          * idle list.  This improves the chance that the next submission
 315          * will reuse the same thread, and maybe catch it while it is still
 316          * working
 317          */
 318         if (!list_empty(&workers->idle_list)) {
 319                 next = workers->idle_list.next;
 320                 worker = list_entry(next, struct btrfs_worker_thread,
 321                                     worker_list);
 322                 return worker;
 323         }
 324         if (enforce_min || list_empty(&workers->worker_list))
 325                 return NULL;
 326
 327         /*
 328          * if we pick a busy task, move the task to the end of the list.
 329          * hopefully this will keep things somewhat evenly balanced.
 330          * Do the move in batches based on the sequence number.  This groups
 331          * requests submitted at roughly the same time onto the same worker.
 332          */
 333         next = workers->worker_list.next;
 334         worker = list_entry(next, struct btrfs_worker_thread, worker_list);
 335         atomic_inc(&worker->num_pending);
 336         worker->sequence++;
 337
 338         if (worker->sequence % workers->idle_thresh == 0)
 339                 list_move_tail(next, &workers->worker_list);
 340         return worker;
 341 }
 342
 343 /*
 344  * selects a worker thread to take the next job.  This will either find
 345  * an idle worker, start a new worker up to the max count, or just return
 346  * one of the existing busy workers.
 347  */
 348 static struct btrfs_worker_thread *find_worker(struct btrfs_workers *workers)
 349 {
 350         struct btrfs_worker_thread *worker;
 351         unsigned long flags;
 352
 353 again:
 354         spin_lock_irqsave(&workers->lock, flags);
 355         worker = next_worker(workers);
 356         spin_unlock_irqrestore(&workers->lock, flags);
 357
 358         if (!worker) {
 359                 spin_lock_irqsave(&workers->lock, flags);
 360                 if (workers->num_workers >= workers->max_workers) {
 361                         struct list_head *fallback = NULL;
 362                         /*
 363                          * we have failed to find any workers, just
 364                          * return the force one
 365                          */
 366                         if (!list_empty(&workers->worker_list))
 367                                 fallback = workers->worker_list.next;
 368                         if (!list_empty(&workers->idle_list))
 369                                 fallback = workers->idle_list.next;
 370                         BUG_ON(!fallback);
 371                         worker = list_entry(fallback,
 372                                   struct btrfs_worker_thread, worker_list);
 373                         spin_unlock_irqrestore(&workers->lock, flags);
 374                 } else {
 375                         spin_unlock_irqrestore(&workers->lock, flags);
 376                         /* we're below the limit, start another worker */
 377                         btrfs_start_workers(workers, 1);
 378                         goto again;
 379                 }
 380         }
 381         return worker;
 382 }
 383
 384 /*
 385  * btrfs_requeue_work just puts the work item back on the tail of the list
 386  * it was taken from.  It is intended for use with long running work functions
 387  * that make some progress and want to give the cpu up for others.
 388  */
 389 int btrfs_requeue_work(struct btrfs_work *work)
 390 {
 391         struct btrfs_worker_thread *worker = work->worker;
 392         unsigned long flags;
 393         int wake = 0;
 394
 395         if (test_and_set_bit(WORK_QUEUED_BIT, &work->flags))
 396                 goto out;
 397
 398         spin_lock_irqsave(&worker->lock, flags);
 399         list_add_tail(&work->list, &worker->pending);
 400         atomic_inc(&worker->num_pending);
 401
 402         /* by definition we're busy, take ourselves off the idle
 403          * list
 404          */
 405         if (worker->idle) {
 406                 spin_lock_irqsave(&worker->workers->lock, flags);
 407                 worker->idle = 0;
 408                 list_move_tail(&worker->worker_list,
 409                                &worker->workers->worker_list);
 410                 spin_unlock_irqrestore(&worker->workers->lock, flags);
 411         }
 412         if (!worker->working) {
 413                 wake = 1;
 414                 worker->working = 1;
 415         }
 416
 417         spin_unlock_irqrestore(&worker->lock, flags);
 418         if (wake)
 419                 wake_up_process(worker->task);
 420 out:
 421
 422         return 0;
 423 }
 424
 425 /*
 426  * places a struct btrfs_work into the pending queue of one of the kthreads
 427  */
 428 int btrfs_queue_worker(struct btrfs_workers *workers, struct btrfs_work *work)
 429 {
 430         struct btrfs_worker_thread *worker;
 431         unsigned long flags;
 432         int wake = 0;
 433
 434         /* don't requeue something already on a list */
 435         if (test_and_set_bit(WORK_QUEUED_BIT, &work->flags))
 436                 goto out;
 437
 438         worker = find_worker(workers);
 439         if (workers->ordered) {
 440                 spin_lock_irqsave(&workers->lock, flags);
 441                 list_add_tail(&work->order_list, &workers->order_list);
 442                 spin_unlock_irqrestore(&workers->lock, flags);
 443         } else {
 444                 INIT_LIST_HEAD(&work->order_list);
 445         }
 446
 447         spin_lock_irqsave(&worker->lock, flags);
 448
 449         list_add_tail(&work->list, &worker->pending);
 450         atomic_inc(&worker->num_pending);
 451         check_busy_worker(worker);
 452
 453         /*
 454          * avoid calling into wake_up_process if this thread has already
 455          * been kicked
 456          */
 457         if (!worker->working)
 458                 wake = 1;
 459         worker->working = 1;
 460
 461         spin_unlock_irqrestore(&worker->lock, flags);
 462
 463         if (wake)
 464                 wake_up_process(worker->task);
 465 out:
 466         return 0;
 467 }