fs/btrfs/locking.c

   1 /*
   2  * Copyright (C) 2008 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 #include <linux/sched.h>
  19 #include <linux/gfp.h>
  20 #include <linux/pagemap.h>
  21 #include <linux/spinlock.h>
  22 #include <linux/page-flags.h>
  23 #include <asm/bug.h>
  24 #include "ctree.h"
  25 #include "extent_io.h"
  26 #include "locking.h"
  27
  28 static inline void spin_nested(struct extent_buffer *eb)
  29 {
  30         spin_lock(&eb->lock);
  31 }
  32
  33 /*
  34  * Setting a lock to blocking will drop the spinlock and set the
  35  * flag that forces other procs who want the lock to wait.  After
  36  * this you can safely schedule with the lock held.
  37  */
  38 void btrfs_set_lock_blocking(struct extent_buffer *eb)
  39 {
  40         if (!test_bit(EXTENT_BUFFER_BLOCKING, &eb->bflags)) {
  41                 set_bit(EXTENT_BUFFER_BLOCKING, &eb->bflags);
  42                 spin_unlock(&eb->lock);
  43         }
  44         /* exit with the spin lock released and the bit set */
  45 }
  46
  47 /*
  48  * clearing the blocking flag will take the spinlock again.
  49  * After this you can't safely schedule
  50  */
  51 void btrfs_clear_lock_blocking(struct extent_buffer *eb)
  52 {
  53         if (test_bit(EXTENT_BUFFER_BLOCKING, &eb->bflags)) {
  54                 spin_nested(eb);
  55                 clear_bit(EXTENT_BUFFER_BLOCKING, &eb->bflags);
  56                 smp_mb__after_clear_bit();
  57         }
  58         /* exit with the spin lock held */
  59 }
  60
  61 /*
  62  * unfortunately, many of the places that currently set a lock to blocking
  63  * don't end up blocking for every long, and often they don't block
  64  * at all.  For a dbench 50 run, if we don't spin one the blocking bit
  65  * at all, the context switch rate can jump up to 400,000/sec or more.
  66  *
  67  * So, we're still stuck with this crummy spin on the blocking bit,
  68  * at least until the most common causes of the short blocks
  69  * can be dealt with.
  70  */
  71 static int btrfs_spin_on_block(struct extent_buffer *eb)
  72 {
  73         int i;
  74         for (i = 0; i < 512; i++) {
  75                 cpu_relax();
  76                 if (!test_bit(EXTENT_BUFFER_BLOCKING, &eb->bflags))
  77                         return 1;
  78                 if (need_resched())
  79                         break;
  80         }
  81         return 0;
  82 }
  83
  84 /*
  85  * This is somewhat different from trylock.  It will take the
  86  * spinlock but if it finds the lock is set to blocking, it will
  87  * return without the lock held.
  88  *
  89  * returns 1 if it was able to take the lock and zero otherwise
  90  *
  91  * After this call, scheduling is not safe without first calling
  92  * btrfs_set_lock_blocking()
  93  */
  94 int btrfs_try_spin_lock(struct extent_buffer *eb)
  95 {
  96         int i;
  97
  98         spin_nested(eb);
  99         if (!test_bit(EXTENT_BUFFER_BLOCKING, &eb->bflags))
 100                 return 1;
 101         spin_unlock(&eb->lock);
 102
 103         /* spin for a bit on the BLOCKING flag */
 104         for (i = 0; i < 2; i++) {
 105                 if (!btrfs_spin_on_block(eb))
 106                         break;
 107
 108                 spin_nested(eb);
 109                 if (!test_bit(EXTENT_BUFFER_BLOCKING, &eb->bflags))
 110                         return 1;
 111                 spin_unlock(&eb->lock);
 112         }
 113         return 0;
 114 }
 115
 116 /*
 117  * the autoremove wake function will return 0 if it tried to wake up
 118  * a process that was already awake, which means that process won't
 119  * count as an exclusive wakeup.  The waitq code will continue waking
 120  * procs until it finds one that was actually sleeping.
 121  *
 122  * For btrfs, this isn't quite what we want.  We want a single proc
 123  * to be notified that the lock is ready for taking.  If that proc
 124  * already happen to be awake, great, it will loop around and try for
 125  * the lock.
 126  *
 127  * So, btrfs_wake_function always returns 1, even when the proc that we
 128  * tried to wake up was already awake.
 129  */
 130 static int btrfs_wake_function(wait_queue_t *wait, unsigned mode,
 131                                int sync, void *key)
 132 {
 133         autoremove_wake_function(wait, mode, sync, key);
 134         return 1;
 135 }
 136
 137 /*
 138  * returns with the extent buffer spinlocked.
 139  *
 140  * This will spin and/or wait as required to take the lock, and then
 141  * return with the spinlock held.
 142  *
 143  * After this call, scheduling is not safe without first calling
 144  * btrfs_set_lock_blocking()
 145  */
 146 int btrfs_tree_lock(struct extent_buffer *eb)
 147 {
 148         DEFINE_WAIT(wait);
 149         wait.func = btrfs_wake_function;
 150
 151         while(1) {
 152                 spin_nested(eb);
 153
 154                 /* nobody is blocking, exit with the spinlock held */
 155                 if (!test_bit(EXTENT_BUFFER_BLOCKING, &eb->bflags))
 156                         return 0;
 157
 158                 /*
 159                  * we have the spinlock, but the real owner is blocking.
 160                  * wait for them
 161                  */
 162                 spin_unlock(&eb->lock);
 163
 164                 /*
 165                  * spin for a bit, and if the blocking flag goes away,
 166                  * loop around
 167                  */
 168                 if (btrfs_spin_on_block(eb))
 169                         continue;
 170
 171                 prepare_to_wait_exclusive(&eb->lock_wq, &wait,
 172                                           TASK_UNINTERRUPTIBLE);
 173
 174                 if (test_bit(EXTENT_BUFFER_BLOCKING, &eb->bflags))
 175                         schedule();
 176
 177                 finish_wait(&eb->lock_wq, &wait);
 178         }
 179         return 0;
 180 }
 181
 182 /*
 183  * Very quick trylock, this does not spin or schedule.  It returns
 184  * 1 with the spinlock held if it was able to take the lock, or it
 185  * returns zero if it was unable to take the lock.
 186  *
 187  * After this call, scheduling is not safe without first calling
 188  * btrfs_set_lock_blocking()
 189  */
 190 int btrfs_try_tree_lock(struct extent_buffer *eb)
 191 {
 192         if (spin_trylock(&eb->lock)) {
 193                 if (test_bit(EXTENT_BUFFER_BLOCKING, &eb->bflags)) {
 194                         /*
 195                          * we've got the spinlock, but the real owner is
 196                          * blocking.  Drop the spinlock and return failure
 197                          */
 198                         spin_unlock(&eb->lock);
 199                         return 0;
 200                 }
 201                 return 1;
 202         }
 203         /* someone else has the spinlock giveup */
 204         return 0;
 205 }
 206
 207 int btrfs_tree_unlock(struct extent_buffer *eb)
 208 {
 209         /*
 210          * if we were a blocking owner, we don't have the spinlock held
 211          * just clear the bit and look for waiters
 212          */
 213         if (test_and_clear_bit(EXTENT_BUFFER_BLOCKING, &eb->bflags))
 214                 smp_mb__after_clear_bit();
 215         else
 216                 spin_unlock(&eb->lock);
 217
 218         if (waitqueue_active(&eb->lock_wq))
 219                 wake_up(&eb->lock_wq);
 220         return 0;
 221 }
 222
 223 int btrfs_tree_locked(struct extent_buffer *eb)
 224 {
 225         return test_bit(EXTENT_BUFFER_BLOCKING, &eb->bflags) ||
 226                         spin_is_locked(&eb->lock);
 227 }