return list_empty(&ai->ai_ail1_list);
}
-static void gfs2_ail1_start(struct gfs2_sbd *sdp, int flags)
+static void gfs2_ail1_start(struct gfs2_sbd *sdp)
{
struct list_head *head;
u64 sync_gen;
- struct list_head *first;
- struct gfs2_ail *first_ai, *ai, *tmp;
+ struct gfs2_ail *ai;
int done = 0;
gfs2_log_lock(sdp);
}
sync_gen = sdp->sd_ail_sync_gen++;
- first = head->prev;
- first_ai = list_entry(first, struct gfs2_ail, ai_list);
- first_ai->ai_sync_gen = sync_gen;
- gfs2_ail1_start_one(sdp, first_ai); /* This may drop log lock */
-
- if (flags & DIO_ALL)
- first = NULL;
-
while(!done) {
- if (first && (head->prev != first ||
- gfs2_ail1_empty_one(sdp, first_ai, 0)))
- break;
-
done = 1;
- list_for_each_entry_safe_reverse(ai, tmp, head, ai_list) {
+ list_for_each_entry_reverse(ai, head, ai_list) {
if (ai->ai_sync_gen >= sync_gen)
continue;
ai->ai_sync_gen = sync_gen;
* flush time, so we ensure that we have just enough free blocks at all
* times to avoid running out during a log flush.
*
+ * We no longer flush the log here, instead we wake up logd to do that
+ * for us. To avoid the thundering herd and to ensure that we deal fairly
+ * with queued waiters, we use an exclusive wait. This means that when we
+ * get woken with enough journal space to get our reservation, we need to
+ * wake the next waiter on the list.
+ *
* Returns: errno
*/
int gfs2_log_reserve(struct gfs2_sbd *sdp, unsigned int blks)
{
- unsigned int try = 0;
unsigned reserved_blks = 6 * (4096 / sdp->sd_vfs->s_blocksize);
+ unsigned wanted = blks + reserved_blks;
+ DEFINE_WAIT(wait);
+ int did_wait = 0;
+ unsigned int free_blocks;
if (gfs2_assert_warn(sdp, blks) ||
gfs2_assert_warn(sdp, blks <= sdp->sd_jdesc->jd_blocks))
return -EINVAL;
-
- mutex_lock(&sdp->sd_log_reserve_mutex);
- gfs2_log_lock(sdp);
- while(atomic_read(&sdp->sd_log_blks_free) <= (blks + reserved_blks)) {
- gfs2_log_unlock(sdp);
- gfs2_ail1_empty(sdp, 0);
- gfs2_log_flush(sdp, NULL);
-
- if (try++)
- gfs2_ail1_start(sdp, 0);
- gfs2_log_lock(sdp);
+retry:
+ free_blocks = atomic_read(&sdp->sd_log_blks_free);
+ if (unlikely(free_blocks <= wanted)) {
+ do {
+ prepare_to_wait_exclusive(&sdp->sd_log_waitq, &wait,
+ TASK_UNINTERRUPTIBLE);
+ wake_up(&sdp->sd_logd_waitq);
+ did_wait = 1;
+ if (atomic_read(&sdp->sd_log_blks_free) <= wanted)
+ io_schedule();
+ free_blocks = atomic_read(&sdp->sd_log_blks_free);
+ } while(free_blocks <= wanted);
+ finish_wait(&sdp->sd_log_waitq, &wait);
}
- atomic_sub(blks, &sdp->sd_log_blks_free);
+ if (atomic_cmpxchg(&sdp->sd_log_blks_free, free_blocks,
+ free_blocks - blks) != free_blocks)
+ goto retry;
trace_gfs2_log_blocks(sdp, -blks);
- gfs2_log_unlock(sdp);
- mutex_unlock(&sdp->sd_log_reserve_mutex);
+
+ /*
+ * If we waited, then so might others, wake them up _after_ we get
+ * our share of the log.
+ */
+ if (unlikely(did_wait))
+ wake_up(&sdp->sd_log_waitq);
down_read(&sdp->sd_log_flush_lock);
return 0;
}
-/**
- * gfs2_log_release - Release a given number of log blocks
- * @sdp: The GFS2 superblock
- * @blks: The number of blocks
- *
- */
-
-void gfs2_log_release(struct gfs2_sbd *sdp, unsigned int blks)
-{
-
- gfs2_log_lock(sdp);
- atomic_add(blks, &sdp->sd_log_blks_free);
- trace_gfs2_log_blocks(sdp, blks);
- gfs2_assert_withdraw(sdp,
- atomic_read(&sdp->sd_log_blks_free) <= sdp->sd_jdesc->jd_blocks);
- gfs2_log_unlock(sdp);
- up_read(&sdp->sd_log_flush_lock);
-}
-
static u64 log_bmap(struct gfs2_sbd *sdp, unsigned int lbn)
{
struct gfs2_journal_extent *je;
ail2_empty(sdp, new_tail);
- gfs2_log_lock(sdp);
atomic_add(dist, &sdp->sd_log_blks_free);
trace_gfs2_log_blocks(sdp, dist);
- gfs2_assert_withdraw(sdp, atomic_read(&sdp->sd_log_blks_free) <= sdp->sd_jdesc->jd_blocks);
- gfs2_log_unlock(sdp);
+ gfs2_assert_withdraw(sdp, atomic_read(&sdp->sd_log_blks_free) <=
+ sdp->sd_jdesc->jd_blocks);
sdp->sd_log_tail = new_tail;
}
* @sdp: the filesystem
* @tr: the transaction
*
+ * We wake up gfs2_logd if the number of pinned blocks exceed thresh1
+ * or the total number of used blocks (pinned blocks plus AIL blocks)
+ * is greater than thresh2.
+ *
+ * At mount time thresh1 is 1/3rd of journal size, thresh2 is 2/3rd of
+ * journal size.
+ *
* Returns: errno
*/
up_read(&sdp->sd_log_flush_lock);
- gfs2_log_lock(sdp);
- if (sdp->sd_log_num_buf > gfs2_tune_get(sdp, gt_incore_log_blocks))
- wake_up_process(sdp->sd_logd_process);
- gfs2_log_unlock(sdp);
+ if (atomic_read(&sdp->sd_log_pinned) > atomic_read(&sdp->sd_log_thresh1) ||
+ ((sdp->sd_jdesc->jd_blocks - atomic_read(&sdp->sd_log_blks_free)) >
+ atomic_read(&sdp->sd_log_thresh2)))
+ wake_up(&sdp->sd_logd_waitq);
}
/**
{
gfs2_log_flush(sdp, NULL);
for (;;) {
- gfs2_ail1_start(sdp, DIO_ALL);
+ gfs2_ail1_start(sdp);
if (gfs2_ail1_empty(sdp, DIO_ALL))
break;
msleep(10);
}
}
+static inline int gfs2_jrnl_flush_reqd(struct gfs2_sbd *sdp)
+{
+ return (atomic_read(&sdp->sd_log_pinned) >= atomic_read(&sdp->sd_log_thresh1));
+}
+
+static inline int gfs2_ail_flush_reqd(struct gfs2_sbd *sdp)
+{
+ unsigned int used_blocks = sdp->sd_jdesc->jd_blocks - atomic_read(&sdp->sd_log_blks_free);
+ return used_blocks >= atomic_read(&sdp->sd_log_thresh2);
+}
/**
* gfs2_logd - Update log tail as Active Items get flushed to in-place blocks
int gfs2_logd(void *data)
{
struct gfs2_sbd *sdp = data;
- unsigned long t;
- int need_flush;
+ unsigned long t = 1;
+ DEFINE_WAIT(wait);
+ unsigned preflush;
while (!kthread_should_stop()) {
- /* Advance the log tail */
- t = sdp->sd_log_flush_time +
- gfs2_tune_get(sdp, gt_log_flush_secs) * HZ;
+ preflush = atomic_read(&sdp->sd_log_pinned);
+ if (gfs2_jrnl_flush_reqd(sdp) || t == 0) {
+ gfs2_ail1_empty(sdp, DIO_ALL);
+ gfs2_log_flush(sdp, NULL);
+ gfs2_ail1_empty(sdp, DIO_ALL);
+ }
- gfs2_ail1_empty(sdp, DIO_ALL);
- gfs2_log_lock(sdp);
- need_flush = sdp->sd_log_num_buf > gfs2_tune_get(sdp, gt_incore_log_blocks);
- gfs2_log_unlock(sdp);
- if (need_flush || time_after_eq(jiffies, t)) {
+ if (gfs2_ail_flush_reqd(sdp)) {
+ gfs2_ail1_start(sdp);
+ io_schedule();
+ gfs2_ail1_empty(sdp, 0);
gfs2_log_flush(sdp, NULL);
- sdp->sd_log_flush_time = jiffies;
+ gfs2_ail1_empty(sdp, DIO_ALL);
}
+ wake_up(&sdp->sd_log_waitq);
t = gfs2_tune_get(sdp, gt_logd_secs) * HZ;
if (freezing(current))
refrigerator();
- schedule_timeout_interruptible(t);
+
+ do {
+ prepare_to_wait(&sdp->sd_logd_waitq, &wait,
+ TASK_UNINTERRUPTIBLE);
+ if (!gfs2_ail_flush_reqd(sdp) &&
+ !gfs2_jrnl_flush_reqd(sdp) &&
+ !kthread_should_stop())
+ t = schedule_timeout(t);
+ } while(t && !gfs2_ail_flush_reqd(sdp) &&
+ !gfs2_jrnl_flush_reqd(sdp) &&
+ !kthread_should_stop());
+ finish_wait(&sdp->sd_logd_waitq, &wait);
}
return 0;
{
spin_lock_init(>->gt_spin);
- gt->gt_incore_log_blocks = 1024;
- gt->gt_logd_secs = 1;
gt->gt_quota_simul_sync = 64;
gt->gt_quota_warn_period = 10;
gt->gt_quota_scale_num = 1;
spin_lock_init(&sdp->sd_trunc_lock);
spin_lock_init(&sdp->sd_log_lock);
-
+ atomic_set(&sdp->sd_log_pinned, 0);
INIT_LIST_HEAD(&sdp->sd_log_le_buf);
INIT_LIST_HEAD(&sdp->sd_log_le_revoke);
INIT_LIST_HEAD(&sdp->sd_log_le_rg);
INIT_LIST_HEAD(&sdp->sd_log_le_databuf);
INIT_LIST_HEAD(&sdp->sd_log_le_ordered);
- mutex_init(&sdp->sd_log_reserve_mutex);
+ init_waitqueue_head(&sdp->sd_log_waitq);
+ init_waitqueue_head(&sdp->sd_logd_waitq);
INIT_LIST_HEAD(&sdp->sd_ail1_list);
INIT_LIST_HEAD(&sdp->sd_ail2_list);
if (sdp->sd_args.ar_spectator) {
sdp->sd_jdesc = gfs2_jdesc_find(sdp, 0);
atomic_set(&sdp->sd_log_blks_free, sdp->sd_jdesc->jd_blocks);
+ atomic_set(&sdp->sd_log_thresh1, 2*sdp->sd_jdesc->jd_blocks/5);
+ atomic_set(&sdp->sd_log_thresh2, 4*sdp->sd_jdesc->jd_blocks/5);
} else {
if (sdp->sd_lockstruct.ls_jid >= gfs2_jindex_size(sdp)) {
fs_err(sdp, "can't mount journal #%u\n",
goto fail_jinode_gh;
}
atomic_set(&sdp->sd_log_blks_free, sdp->sd_jdesc->jd_blocks);
+ atomic_set(&sdp->sd_log_thresh1, 2*sdp->sd_jdesc->jd_blocks/5);
+ atomic_set(&sdp->sd_log_thresh2, 4*sdp->sd_jdesc->jd_blocks/5);
/* Map the extents for this journal's blocks */
map_journal_extents(sdp);
if (undo)
goto fail_quotad;
- sdp->sd_log_flush_time = jiffies;
-
p = kthread_run(gfs2_logd, sdp, "gfs2_logd");
error = IS_ERR(p);
if (error) {
GFS2_BASIC_BLOCK_SHIFT;
sdp->sd_fsb2bb = 1 << sdp->sd_fsb2bb_shift;
- sdp->sd_tune.gt_log_flush_secs = sdp->sd_args.ar_commit;
+ sdp->sd_tune.gt_logd_secs = sdp->sd_args.ar_commit;
sdp->sd_tune.gt_quota_quantum = sdp->sd_args.ar_quota_quantum;
if (sdp->sd_args.ar_statfs_quantum) {
sdp->sd_tune.gt_statfs_slow = 0;
memset(&args, 0, sizeof(args));
args.ar_quota = GFS2_QUOTA_DEFAULT;
args.ar_data = GFS2_DATA_DEFAULT;
- args.ar_commit = 60;
+ args.ar_commit = 30;
args.ar_statfs_quantum = 30;
args.ar_quota_quantum = 60;
args.ar_errors = GFS2_ERRORS_DEFAULT;