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/version.h>
  20 #include <linux/kthread.h>
  21 #include <linux/list.h>
  22 #include <linux/spinlock.h>
  23
  24 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,20)
  25 # include <linux/freezer.h>
  26 #else
  27 # include <linux/sched.h>
  28 #endif
  29
  30 #include "async-thread.h"
  31
  32 /*
  33  * container for the kthread task pointer and the list of pending work
  34  * One of these is allocated per thread.
  35  */
  36 struct btrfs_worker_thread {
  37         /* pool we belong to */
  38         struct btrfs_workers *workers;
  39
  40         /* list of struct btrfs_work that are waiting for service */
  41         struct list_head 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         /* protects the pending list. */
  53         spinlock_t lock;
  54
  55         /* set to non-zero when this thread is already awake and kicking */
  56         int working;
  57
  58         /* are we currently idle */
  59         int idle;
  60 };
  61
  62 /*
  63  * helper function to move a thread onto the idle list after it
  64  * has finished some requests.
  65  */
  66 static void check_idle_worker(struct btrfs_worker_thread *worker)
  67 {
  68         if (!worker->idle && atomic_read(&worker->num_pending) <
  69             worker->workers->idle_thresh / 2) {
  70                 unsigned long flags;
  71                 spin_lock_irqsave(&worker->workers->lock, flags);
  72                 worker->idle = 1;
  73                 list_move(&worker->worker_list, &worker->workers->idle_list);
  74                 spin_unlock_irqrestore(&worker->workers->lock, flags);
  75         }
  76 }
  77
  78 /*
  79  * helper function to move a thread off the idle list after new
  80  * pending work is added.
  81  */
  82 static void check_busy_worker(struct btrfs_worker_thread *worker)
  83 {
  84         if (worker->idle && atomic_read(&worker->num_pending) >=
  85             worker->workers->idle_thresh) {
  86                 unsigned long flags;
  87                 spin_lock_irqsave(&worker->workers->lock, flags);
  88                 worker->idle = 0;
  89                 list_move_tail(&worker->worker_list,
  90                                &worker->workers->worker_list);
  91                 spin_unlock_irqrestore(&worker->workers->lock, flags);
  92         }
  93 }
  94
  95 /*
  96  * main loop for servicing work items
  97  */
  98 static int worker_loop(void *arg)
  99 {
 100         struct btrfs_worker_thread *worker = arg;
 101         struct list_head *cur;
 102         struct btrfs_work *work;
 103         do {
 104                 spin_lock_irq(&worker->lock);
 105                 while(!list_empty(&worker->pending)) {
 106                         cur = worker->pending.next;
 107                         work = list_entry(cur, struct btrfs_work, list);
 108                         list_del(&work->list);
 109                         clear_bit(0, &work->flags);
 110
 111                         work->worker = worker;
 112                         spin_unlock_irq(&worker->lock);
 113
 114                         work->func(work);
 115
 116                         atomic_dec(&worker->num_pending);
 117                         spin_lock_irq(&worker->lock);
 118                         check_idle_worker(worker);
 119                 }
 120                 worker->working = 0;
 121                 if (freezing(current)) {
 122                         refrigerator();
 123                 } else {
 124                         set_current_state(TASK_INTERRUPTIBLE);
 125                         spin_unlock_irq(&worker->lock);
 126                         schedule();
 127                         __set_current_state(TASK_RUNNING);
 128                 }
 129         } while (!kthread_should_stop());
 130         return 0;
 131 }
 132
 133 /*
 134  * this will wait for all the worker threads to shutdown
 135  */
 136 int btrfs_stop_workers(struct btrfs_workers *workers)
 137 {
 138         struct list_head *cur;
 139         struct btrfs_worker_thread *worker;
 140
 141         list_splice_init(&workers->idle_list, &workers->worker_list);
 142         while(!list_empty(&workers->worker_list)) {
 143                 cur = workers->worker_list.next;
 144                 worker = list_entry(cur, struct btrfs_worker_thread,
 145                                     worker_list);
 146                 kthread_stop(worker->task);
 147                 list_del(&worker->worker_list);
 148                 kfree(worker);
 149         }
 150         return 0;
 151 }
 152
 153 /*
 154  * simple init on struct btrfs_workers
 155  */
 156 void btrfs_init_workers(struct btrfs_workers *workers, int max)
 157 {
 158         workers->num_workers = 0;
 159         INIT_LIST_HEAD(&workers->worker_list);
 160         INIT_LIST_HEAD(&workers->idle_list);
 161         spin_lock_init(&workers->lock);
 162         workers->max_workers = max;
 163         workers->idle_thresh = 32;
 164 }
 165
 166 /*
 167  * starts new worker threads.  This does not enforce the max worker
 168  * count in case you need to temporarily go past it.
 169  */
 170 int btrfs_start_workers(struct btrfs_workers *workers, int num_workers)
 171 {
 172         struct btrfs_worker_thread *worker;
 173         int ret = 0;
 174         int i;
 175
 176         for (i = 0; i < num_workers; i++) {
 177                 worker = kzalloc(sizeof(*worker), GFP_NOFS);
 178                 if (!worker) {
 179                         ret = -ENOMEM;
 180                         goto fail;
 181                 }
 182
 183                 INIT_LIST_HEAD(&worker->pending);
 184                 INIT_LIST_HEAD(&worker->worker_list);
 185                 spin_lock_init(&worker->lock);
 186                 atomic_set(&worker->num_pending, 0);
 187                 worker->task = kthread_run(worker_loop, worker, "btrfs");
 188                 worker->workers = workers;
 189                 if (IS_ERR(worker->task)) {
 190                         kfree(worker);
 191                         ret = PTR_ERR(worker->task);
 192                         goto fail;
 193                 }
 194
 195                 spin_lock_irq(&workers->lock);
 196                 list_add_tail(&worker->worker_list, &workers->idle_list);
 197                 workers->num_workers++;
 198                 spin_unlock_irq(&workers->lock);
 199         }
 200         return 0;
 201 fail:
 202         btrfs_stop_workers(workers);
 203         return ret;
 204 }
 205
 206 /*
 207  * run through the list and find a worker thread that doesn't have a lot
 208  * to do right now.  This can return null if we aren't yet at the thread
 209  * count limit and all of the threads are busy.
 210  */
 211 static struct btrfs_worker_thread *next_worker(struct btrfs_workers *workers)
 212 {
 213         struct btrfs_worker_thread *worker;
 214         struct list_head *next;
 215         int enforce_min = workers->num_workers < workers->max_workers;
 216
 217         /*
 218          * if we find an idle thread, don't move it to the end of the
 219          * idle list.  This improves the chance that the next submission
 220          * will reuse the same thread, and maybe catch it while it is still
 221          * working
 222          */
 223         if (!list_empty(&workers->idle_list)) {
 224                 next = workers->idle_list.next;
 225                 worker = list_entry(next, struct btrfs_worker_thread,
 226                                     worker_list);
 227                 return worker;
 228         }
 229         if (enforce_min || list_empty(&workers->worker_list))
 230                 return NULL;
 231
 232         /*
 233          * if we pick a busy task, move the task to the end of the list.
 234          * hopefully this will keep things somewhat evenly balanced
 235          */
 236         next = workers->worker_list.next;
 237         worker = list_entry(next, struct btrfs_worker_thread, worker_list);
 238         list_move_tail(next, &workers->worker_list);
 239         return worker;
 240 }
 241
 242 static struct btrfs_worker_thread *find_worker(struct btrfs_workers *workers)
 243 {
 244         struct btrfs_worker_thread *worker;
 245         unsigned long flags;
 246
 247 again:
 248         spin_lock_irqsave(&workers->lock, flags);
 249         worker = next_worker(workers);
 250         spin_unlock_irqrestore(&workers->lock, flags);
 251
 252         if (!worker) {
 253                 spin_lock_irqsave(&workers->lock, flags);
 254                 if (workers->num_workers >= workers->max_workers) {
 255                         struct list_head *fallback = NULL;
 256                         /*
 257                          * we have failed to find any workers, just
 258                          * return the force one
 259                          */
 260                         if (!list_empty(&workers->worker_list))
 261                                 fallback = workers->worker_list.next;
 262                         if (!list_empty(&workers->idle_list))
 263                                 fallback = workers->idle_list.next;
 264                         BUG_ON(!fallback);
 265                         worker = list_entry(fallback,
 266                                   struct btrfs_worker_thread, worker_list);
 267                         spin_unlock_irqrestore(&workers->lock, flags);
 268                 } else {
 269                         spin_unlock_irqrestore(&workers->lock, flags);
 270                         /* we're below the limit, start another worker */
 271                         btrfs_start_workers(workers, 1);
 272                         goto again;
 273                 }
 274         }
 275         return worker;
 276 }
 277
 278 /*
 279  * btrfs_requeue_work just puts the work item back on the tail of the list
 280  * it was taken from.  It is intended for use with long running work functions
 281  * that make some progress and want to give the cpu up for others.
 282  */
 283 int btrfs_requeue_work(struct btrfs_work *work)
 284 {
 285         struct btrfs_worker_thread *worker = work->worker;
 286         unsigned long flags;
 287
 288         if (test_and_set_bit(0, &work->flags))
 289                 goto out;
 290
 291         spin_lock_irqsave(&worker->lock, flags);
 292         atomic_inc(&worker->num_pending);
 293         list_add_tail(&work->list, &worker->pending);
 294         check_busy_worker(worker);
 295         spin_unlock_irqrestore(&worker->lock, flags);
 296 out:
 297         return 0;
 298 }
 299
 300 /*
 301  * places a struct btrfs_work into the pending queue of one of the kthreads
 302  */
 303 int btrfs_queue_worker(struct btrfs_workers *workers, struct btrfs_work *work)
 304 {
 305         struct btrfs_worker_thread *worker;
 306         unsigned long flags;
 307         int wake = 0;
 308
 309         /* don't requeue something already on a list */
 310         if (test_and_set_bit(0, &work->flags))
 311                 goto out;
 312
 313         worker = find_worker(workers);
 314
 315         spin_lock_irqsave(&worker->lock, flags);
 316         atomic_inc(&worker->num_pending);
 317         check_busy_worker(worker);
 318         list_add_tail(&work->list, &worker->pending);
 319
 320         /*
 321          * avoid calling into wake_up_process if this thread has already
 322          * been kicked
 323          */
 324         if (!worker->working)
 325                 wake = 1;
 326         worker->working = 1;
 327
 328         spin_unlock_irqrestore(&worker->lock, flags);
 329
 330         if (wake)
 331                 wake_up_process(worker->task);
 332 out:
 333         return 0;
 334 }