]> git.karo-electronics.de Git - karo-tx-linux.git/commitdiff
xfs: make AIL tail pushing independent of the grant lock
authorDave Chinner <dchinner@redhat.com>
Tue, 21 Dec 2010 01:09:20 +0000 (12:09 +1100)
committerDave Chinner <david@fromorbit.com>
Tue, 21 Dec 2010 01:09:20 +0000 (12:09 +1100)
The xlog_grant_push_ail() currently takes the grant lock internally to sample
the tail lsn, last sync lsn and the reserve grant head. Most of the callers
already hold the grant lock but have to drop it before calling
xlog_grant_push_ail(). This is a left over from when the AIL tail pushing was
done in line and hence xlog_grant_push_ail had to drop the grant lock. AIL push
is now done in another thread and hence we can safely hold the grant lock over
the entire xlog_grant_push_ail call.

Push the grant lock outside of xlog_grant_push_ail() to simplify the locking
and synchronisation needed for tail pushing.  This will reduce traffic on the
grant lock by itself, but this is only one step in preparing for the complete
removal of the grant lock.

While there, clean up the formatting of xlog_grant_push_ail() to match the
rest of the XFS code.

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
fs/xfs/xfs_log.c

index cc0504e0bb3b1f5efd4e7667642ab6c5bb722060..1e2020d5a8b64e850aa9d939e793c694621e09cb 100644 (file)
@@ -70,7 +70,7 @@ STATIC void xlog_state_want_sync(xlog_t       *log, xlog_in_core_t *iclog);
 /* local functions to manipulate grant head */
 STATIC int  xlog_grant_log_space(xlog_t                *log,
                                 xlog_ticket_t  *xtic);
-STATIC void xlog_grant_push_ail(xfs_mount_t    *mp,
+STATIC void xlog_grant_push_ail(struct log     *log,
                                int             need_bytes);
 STATIC void xlog_regrant_reserve_log_space(xlog_t       *log,
                                           xlog_ticket_t *ticket);
@@ -318,7 +318,9 @@ xfs_log_reserve(
 
                trace_xfs_log_reserve(log, internal_ticket);
 
-               xlog_grant_push_ail(mp, internal_ticket->t_unit_res);
+               spin_lock(&log->l_grant_lock);
+               xlog_grant_push_ail(log, internal_ticket->t_unit_res);
+               spin_unlock(&log->l_grant_lock);
                retval = xlog_regrant_write_log_space(log, internal_ticket);
        } else {
                /* may sleep if need to allocate more tickets */
@@ -332,9 +334,11 @@ xfs_log_reserve(
 
                trace_xfs_log_reserve(log, internal_ticket);
 
-               xlog_grant_push_ail(mp,
+               spin_lock(&log->l_grant_lock);
+               xlog_grant_push_ail(log,
                                    (internal_ticket->t_unit_res *
                                     internal_ticket->t_cnt));
+               spin_unlock(&log->l_grant_lock);
                retval = xlog_grant_log_space(log, internal_ticket);
        }
 
@@ -1185,59 +1189,58 @@ xlog_commit_record(
  * water mark.  In this manner, we would be creating a low water mark.
  */
 STATIC void
-xlog_grant_push_ail(xfs_mount_t        *mp,
-                   int         need_bytes)
+xlog_grant_push_ail(
+       struct log      *log,
+       int             need_bytes)
 {
-    xlog_t     *log = mp->m_log;       /* pointer to the log */
-    xfs_lsn_t  tail_lsn;               /* lsn of the log tail */
-    xfs_lsn_t  threshold_lsn = 0;      /* lsn we'd like to be at */
-    int                free_blocks;            /* free blocks left to write to */
-    int                free_bytes;             /* free bytes left to write to */
-    int                threshold_block;        /* block in lsn we'd like to be at */
-    int                threshold_cycle;        /* lsn cycle we'd like to be at */
-    int                free_threshold;
-
-    ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
-
-    spin_lock(&log->l_grant_lock);
-    free_bytes = xlog_space_left(log, &log->l_grant_reserve_head);
-    tail_lsn = log->l_tail_lsn;
-    free_blocks = BTOBBT(free_bytes);
-
-    /*
-     * Set the threshold for the minimum number of free blocks in the
-     * log to the maximum of what the caller needs, one quarter of the
-     * log, and 256 blocks.
-     */
-    free_threshold = BTOBB(need_bytes);
-    free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
-    free_threshold = MAX(free_threshold, 256);
-    if (free_blocks < free_threshold) {
+       xfs_lsn_t       threshold_lsn = 0;
+       xfs_lsn_t       tail_lsn;
+       int             free_blocks;
+       int             free_bytes;
+       int             threshold_block;
+       int             threshold_cycle;
+       int             free_threshold;
+
+       ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
+
+       tail_lsn = log->l_tail_lsn;
+       free_bytes = xlog_space_left(log, &log->l_grant_reserve_head);
+       free_blocks = BTOBBT(free_bytes);
+
+       /*
+        * Set the threshold for the minimum number of free blocks in the
+        * log to the maximum of what the caller needs, one quarter of the
+        * log, and 256 blocks.
+        */
+       free_threshold = BTOBB(need_bytes);
+       free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
+       free_threshold = MAX(free_threshold, 256);
+       if (free_blocks >= free_threshold)
+               return;
+
        threshold_block = BLOCK_LSN(tail_lsn) + free_threshold;
        threshold_cycle = CYCLE_LSN(tail_lsn);
        if (threshold_block >= log->l_logBBsize) {
-           threshold_block -= log->l_logBBsize;
-           threshold_cycle += 1;
+               threshold_block -= log->l_logBBsize;
+               threshold_cycle += 1;
        }
-       threshold_lsn = xlog_assign_lsn(threshold_cycle, threshold_block);
-
-       /* Don't pass in an lsn greater than the lsn of the last
+       threshold_lsn = xlog_assign_lsn(threshold_cycle,
+                                       threshold_block);
+       /*
+        * Don't pass in an lsn greater than the lsn of the last
         * log record known to be on disk.
         */
        if (XFS_LSN_CMP(threshold_lsn, log->l_last_sync_lsn) > 0)
-           threshold_lsn = log->l_last_sync_lsn;
-    }
-    spin_unlock(&log->l_grant_lock);
-
-    /*
-     * Get the transaction layer to kick the dirty buffers out to
-     * disk asynchronously. No point in trying to do this if
-     * the filesystem is shutting down.
-     */
-    if (threshold_lsn &&
-       !XLOG_FORCED_SHUTDOWN(log))
-           xfs_trans_ail_push(log->l_ailp, threshold_lsn);
-}      /* xlog_grant_push_ail */
+               threshold_lsn = log->l_last_sync_lsn;
+
+       /*
+        * Get the transaction layer to kick the dirty buffers out to
+        * disk asynchronously. No point in trying to do this if
+        * the filesystem is shutting down.
+        */
+       if (!XLOG_FORCED_SHUTDOWN(log))
+               xfs_trans_ail_push(log->l_ailp, threshold_lsn);
+}
 
 /*
  * The bdstrat callback function for log bufs. This gives us a central
@@ -2543,9 +2546,7 @@ redo:
 
                trace_xfs_log_grant_sleep2(log, tic);
 
-               spin_unlock(&log->l_grant_lock);
-               xlog_grant_push_ail(log->l_mp, need_bytes);
-               spin_lock(&log->l_grant_lock);
+               xlog_grant_push_ail(log, need_bytes);
 
                XFS_STATS_INC(xs_sleep_logspace);
                xlog_wait(&tic->t_wait, &log->l_grant_lock);
@@ -2641,9 +2642,7 @@ xlog_regrant_write_log_space(xlog_t          *log,
 
                        trace_xfs_log_regrant_write_sleep1(log, tic);
 
-                       spin_unlock(&log->l_grant_lock);
-                       xlog_grant_push_ail(log->l_mp, need_bytes);
-                       spin_lock(&log->l_grant_lock);
+                       xlog_grant_push_ail(log, need_bytes);
 
                        XFS_STATS_INC(xs_sleep_logspace);
                        xlog_wait(&tic->t_wait, &log->l_grant_lock);
@@ -2666,9 +2665,7 @@ redo:
        if (free_bytes < need_bytes) {
                if (list_empty(&tic->t_queue))
                        list_add_tail(&tic->t_queue, &log->l_writeq);
-               spin_unlock(&log->l_grant_lock);
-               xlog_grant_push_ail(log->l_mp, need_bytes);
-               spin_lock(&log->l_grant_lock);
+               xlog_grant_push_ail(log, need_bytes);
 
                XFS_STATS_INC(xs_sleep_logspace);
                trace_xfs_log_regrant_write_sleep2(log, tic);