[XFS] Use atomics for iclog reference counting
authorDavid Chinner <dgc@sgi.com>
Thu, 6 Mar 2008 02:44:14 +0000 (13:44 +1100)
committerLachlan McIlroy <lachlan@redback.melbourne.sgi.com>
Fri, 18 Apr 2008 01:38:10 +0000 (11:38 +1000)
Now that we update the log tail LSN less frequently on transaction
completion, we pass the contention straight to the global log state lock
(l_iclog_lock) during transaction completion.

We currently have to take this lock to decrement the iclog reference
count. there is a reference count on each iclog, so we need to take þhe
global lock for all refcount changes.

When large numbers of processes are all doing small trnasctions, the iclog
reference counts will be quite high, and the state change that absolutely
requires the l_iclog_lock is the except rather than the norm.

Change the reference counting on the iclogs to use atomic_inc/dec so that
we can use atomic_dec_and_lock during transaction completion and avoid the
need for grabbing the l_iclog_lock for every reference count decrement
except the one that matters - the last.

SGI-PV: 975671
SGI-Modid: xfs-linux-melb:xfs-kern:30505a

Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Tim Shimmin <tes@sgi.com>
Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>
fs/xfs/xfs_log.c
fs/xfs/xfs_log_priv.h

index 2e35077ff6b2a0700bbc19a6997f87cc12f369b8..1fa9809338954d05c460ce2562a91eaeee6e92e8 100644 (file)
@@ -675,7 +675,7 @@ xfs_log_unmount_write(xfs_mount_t *mp)
 
                spin_lock(&log->l_icloglock);
                iclog = log->l_iclog;
-               iclog->ic_refcnt++;
+               atomic_inc(&iclog->ic_refcnt);
                spin_unlock(&log->l_icloglock);
                xlog_state_want_sync(log, iclog);
                (void) xlog_state_release_iclog(log, iclog);
@@ -713,7 +713,7 @@ xfs_log_unmount_write(xfs_mount_t *mp)
                 */
                spin_lock(&log->l_icloglock);
                iclog = log->l_iclog;
-               iclog->ic_refcnt++;
+               atomic_inc(&iclog->ic_refcnt);
                spin_unlock(&log->l_icloglock);
 
                xlog_state_want_sync(log, iclog);
@@ -1405,7 +1405,7 @@ xlog_sync(xlog_t          *log,
        int             v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
 
        XFS_STATS_INC(xs_log_writes);
-       ASSERT(iclog->ic_refcnt == 0);
+       ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
 
        /* Add for LR header */
        count_init = log->l_iclog_hsize + iclog->ic_offset;
@@ -2309,7 +2309,7 @@ xlog_state_done_syncing(
 
        ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
               iclog->ic_state == XLOG_STATE_IOERROR);
-       ASSERT(iclog->ic_refcnt == 0);
+       ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
        ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
 
 
@@ -2391,7 +2391,7 @@ restart:
        ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
        head = &iclog->ic_header;
 
-       iclog->ic_refcnt++;                     /* prevents sync */
+       atomic_inc(&iclog->ic_refcnt);  /* prevents sync */
        log_offset = iclog->ic_offset;
 
        /* On the 1st write to an iclog, figure out lsn.  This works
@@ -2423,12 +2423,12 @@ restart:
                xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
 
                /* If I'm the only one writing to this iclog, sync it to disk */
-               if (iclog->ic_refcnt == 1) {
+               if (atomic_read(&iclog->ic_refcnt) == 1) {
                        spin_unlock(&log->l_icloglock);
                        if ((error = xlog_state_release_iclog(log, iclog)))
                                return error;
                } else {
-                       iclog->ic_refcnt--;
+                       atomic_dec(&iclog->ic_refcnt);
                        spin_unlock(&log->l_icloglock);
                }
                goto restart;
@@ -2819,18 +2819,21 @@ xlog_state_release_iclog(
 {
        int             sync = 0;       /* do we sync? */
 
-       spin_lock(&log->l_icloglock);
+       if (iclog->ic_state & XLOG_STATE_IOERROR)
+               return XFS_ERROR(EIO);
+
+       ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
+       if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
+               return 0;
+
        if (iclog->ic_state & XLOG_STATE_IOERROR) {
                spin_unlock(&log->l_icloglock);
                return XFS_ERROR(EIO);
        }
-
-       ASSERT(iclog->ic_refcnt > 0);
        ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
               iclog->ic_state == XLOG_STATE_WANT_SYNC);
 
-       if (--iclog->ic_refcnt == 0 &&
-           iclog->ic_state == XLOG_STATE_WANT_SYNC) {
+       if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
                /* update tail before writing to iclog */
                xlog_assign_tail_lsn(log->l_mp);
                sync++;
@@ -2950,7 +2953,8 @@ xlog_state_sync_all(xlog_t *log, uint flags, int *log_flushed)
                 * previous iclog and go to sleep.
                 */
                if (iclog->ic_state == XLOG_STATE_DIRTY ||
-                   (iclog->ic_refcnt == 0 && iclog->ic_offset == 0)) {
+                   (atomic_read(&iclog->ic_refcnt) == 0
+                    && iclog->ic_offset == 0)) {
                        iclog = iclog->ic_prev;
                        if (iclog->ic_state == XLOG_STATE_ACTIVE ||
                            iclog->ic_state == XLOG_STATE_DIRTY)
@@ -2958,14 +2962,14 @@ xlog_state_sync_all(xlog_t *log, uint flags, int *log_flushed)
                        else
                                goto maybe_sleep;
                } else {
-                       if (iclog->ic_refcnt == 0) {
+                       if (atomic_read(&iclog->ic_refcnt) == 0) {
                                /* We are the only one with access to this
                                 * iclog.  Flush it out now.  There should
                                 * be a roundoff of zero to show that someone
                                 * has already taken care of the roundoff from
                                 * the previous sync.
                                 */
-                               iclog->ic_refcnt++;
+                               atomic_inc(&iclog->ic_refcnt);
                                lsn = be64_to_cpu(iclog->ic_header.h_lsn);
                                xlog_state_switch_iclogs(log, iclog, 0);
                                spin_unlock(&log->l_icloglock);
@@ -3097,7 +3101,7 @@ try_again:
                        already_slept = 1;
                        goto try_again;
                } else {
-                       iclog->ic_refcnt++;
+                       atomic_inc(&iclog->ic_refcnt);
                        xlog_state_switch_iclogs(log, iclog, 0);
                        spin_unlock(&log->l_icloglock);
                        if (xlog_state_release_iclog(log, iclog))
index c6244cc733c09ab032ae572a6d1488bdc4685115..01c63db25a1d3071f9f855b8241cdafc185cd414 100644 (file)
@@ -339,7 +339,7 @@ typedef struct xlog_iclog_fields {
 #endif
        int                     ic_size;
        int                     ic_offset;
-       int                     ic_refcnt;
+       atomic_t                ic_refcnt;
        int                     ic_bwritecnt;
        ushort_t                ic_state;
        char                    *ic_datap;      /* pointer to iclog data */