[XFS] Move AIL pushing into it's own thread
authorDavid Chinner <dgc@sgi.com>
Tue, 5 Feb 2008 01:13:32 +0000 (12:13 +1100)
committerLachlan McIlroy <lachlan@redback.melbourne.sgi.com>
Thu, 7 Feb 2008 07:22:51 +0000 (18:22 +1100)
When many hundreds to thousands of threads all try to do simultaneous
transactions and the log is in a tail-pushing situation (i.e. full), we
can get multiple threads walking the AIL list and contending on the AIL
lock.

The AIL push is, in effect, a simple I/O dispatch algorithm complicated by
the ordering constraints placed on it by the transaction subsystem. It
really does not need multiple threads to push on it - even when only a
single CPU is pushing the AIL, it can push the I/O out far faster that
pretty much any disk subsystem can handle.

So, to avoid contention problems stemming from multiple list walkers, move
the list walk off into another thread and simply provide a "target" to
push to. When a thread requires a push, it sets the target and wakes the
push thread, then goes to sleep waiting for the required amount of space
to become available in the log.

This mechanism should also be a lot fairer under heavy load as the waiters
will queue in arrival order, rather than queuing in "who completed a push
first" order.

Also, by moving the pushing to a separate thread we can do more
effectively overload detection and prevention as we can keep context from
loop iteration to loop iteration. That is, we can push only part of the
list each loop and not have to loop back to the start of the list every
time we run. This should also help by reducing the number of items we try
to lock and/or push items that we cannot move.

Note that this patch is not intended to solve the inefficiencies in the
AIL structure and the associated issues with extremely large list
contents. That needs to be addresses separately; parallel access would
cause problems to any new structure as well, so I'm only aiming to isolate
the structure from unbounded parallelism here.

SGI-PV: 972759
SGI-Modid: xfs-linux-melb:xfs-kern:30371a

Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>
fs/xfs/linux-2.6/xfs_super.c
fs/xfs/xfs_log.c
fs/xfs/xfs_mount.c
fs/xfs/xfs_mount.h
fs/xfs/xfs_trans.h
fs/xfs/xfs_trans_ail.c
fs/xfs/xfs_trans_priv.h

index 70024a24692d4b1c8960e782d33360b426817566..156a1a7da16bc9cf9ddc7dc75ee12e30a7172b37 100644 (file)
@@ -51,6 +51,7 @@
 #include "xfs_vfsops.h"
 #include "xfs_version.h"
 #include "xfs_log_priv.h"
+#include "xfs_trans_priv.h"
 
 #include <linux/namei.h>
 #include <linux/init.h>
@@ -765,6 +766,64 @@ xfs_blkdev_issue_flush(
        blkdev_issue_flush(buftarg->bt_bdev, NULL);
 }
 
+/*
+ * XFS AIL push thread support
+ */
+void
+xfsaild_wakeup(
+       xfs_mount_t             *mp,
+       xfs_lsn_t               threshold_lsn)
+{
+       mp->m_ail.xa_target = threshold_lsn;
+       wake_up_process(mp->m_ail.xa_task);
+}
+
+int
+xfsaild(
+       void    *data)
+{
+       xfs_mount_t     *mp = (xfs_mount_t *)data;
+       xfs_lsn_t       last_pushed_lsn = 0;
+       long            tout = 0;
+
+       while (!kthread_should_stop()) {
+               if (tout)
+                       schedule_timeout_interruptible(msecs_to_jiffies(tout));
+               tout = 1000;
+
+               /* swsusp */
+               try_to_freeze();
+
+               ASSERT(mp->m_log);
+               if (XFS_FORCED_SHUTDOWN(mp))
+                       continue;
+
+               tout = xfsaild_push(mp, &last_pushed_lsn);
+       }
+
+       return 0;
+}      /* xfsaild */
+
+int
+xfsaild_start(
+       xfs_mount_t     *mp)
+{
+       mp->m_ail.xa_target = 0;
+       mp->m_ail.xa_task = kthread_run(xfsaild, mp, "xfsaild");
+       if (IS_ERR(mp->m_ail.xa_task))
+               return -PTR_ERR(mp->m_ail.xa_task);
+       return 0;
+}
+
+void
+xfsaild_stop(
+       xfs_mount_t     *mp)
+{
+       kthread_stop(mp->m_ail.xa_task);
+}
+
+
+
 STATIC struct inode *
 xfs_fs_alloc_inode(
        struct super_block      *sb)
index 4c86a26330a054b634d60145f16a22a1113fc87a..b3ac3805d3c4434ac13257846d326ca3c6c100af 100644 (file)
@@ -498,11 +498,14 @@ xfs_log_reserve(xfs_mount_t        *mp,
  * Return error or zero.
  */
 int
-xfs_log_mount(xfs_mount_t      *mp,
-             xfs_buftarg_t     *log_target,
-             xfs_daddr_t       blk_offset,
-             int               num_bblks)
+xfs_log_mount(
+       xfs_mount_t     *mp,
+       xfs_buftarg_t   *log_target,
+       xfs_daddr_t     blk_offset,
+       int             num_bblks)
 {
+       int             error;
+
        if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
                cmn_err(CE_NOTE, "XFS mounting filesystem %s", mp->m_fsname);
        else {
@@ -514,12 +517,22 @@ xfs_log_mount(xfs_mount_t *mp,
 
        mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
 
+       /*
+        * Initialize the AIL now we have a log.
+        */
+       spin_lock_init(&mp->m_ail_lock);
+       error = xfs_trans_ail_init(mp);
+       if (error) {
+               cmn_err(CE_WARN, "XFS: AIL initialisation failed: error %d", error);
+               goto error;
+       }
+
        /*
         * skip log recovery on a norecovery mount.  pretend it all
         * just worked.
         */
        if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
-               int             error, readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
+               int     readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
 
                if (readonly)
                        mp->m_flags &= ~XFS_MOUNT_RDONLY;
@@ -530,8 +543,7 @@ xfs_log_mount(xfs_mount_t   *mp,
                        mp->m_flags |= XFS_MOUNT_RDONLY;
                if (error) {
                        cmn_err(CE_WARN, "XFS: log mount/recovery failed: error %d", error);
-                       xlog_dealloc_log(mp->m_log);
-                       return error;
+                       goto error;
                }
        }
 
@@ -540,6 +552,9 @@ xfs_log_mount(xfs_mount_t   *mp,
 
        /* End mounting message in xfs_log_mount_finish */
        return 0;
+error:
+       xfs_log_unmount_dealloc(mp);
+       return error;
 }      /* xfs_log_mount */
 
 /*
@@ -722,10 +737,14 @@ xfs_log_unmount_write(xfs_mount_t *mp)
 
 /*
  * Deallocate log structures for unmount/relocation.
+ *
+ * We need to stop the aild from running before we destroy
+ * and deallocate the log as the aild references the log.
  */
 void
 xfs_log_unmount_dealloc(xfs_mount_t *mp)
 {
+       xfs_trans_ail_destroy(mp);
        xlog_dealloc_log(mp->m_log);
 }
 
index 766c01670b2ce15b74e71eece63c81df4ffe5c68..6409b376299594f4a90fd0e8b67b0f34656364da 100644 (file)
@@ -136,15 +136,9 @@ xfs_mount_init(void)
                mp->m_flags |= XFS_MOUNT_NO_PERCPU_SB;
        }
 
-       spin_lock_init(&mp->m_ail_lock);
        spin_lock_init(&mp->m_sb_lock);
        mutex_init(&mp->m_ilock);
        mutex_init(&mp->m_growlock);
-       /*
-        * Initialize the AIL.
-        */
-       xfs_trans_ail_init(mp);
-
        atomic_set(&mp->m_active_trans, 0);
 
        return mp;
index 15025e01300a5b1e5adfddf5cd411b658766d267..f7c620ec6e695eb338b4da9a21692bf5d0c59608 100644 (file)
@@ -219,12 +219,18 @@ extern void       xfs_icsb_sync_counters_flags(struct xfs_mount *, int);
 #define xfs_icsb_sync_counters_flags(mp, flags)        do { } while (0)
 #endif
 
+typedef struct xfs_ail {
+       xfs_ail_entry_t         xa_ail;
+       uint                    xa_gen;
+       struct task_struct      *xa_task;
+       xfs_lsn_t               xa_target;
+} xfs_ail_t;
+
 typedef struct xfs_mount {
        struct super_block      *m_super;
        xfs_tid_t               m_tid;          /* next unused tid for fs */
        spinlock_t              m_ail_lock;     /* fs AIL mutex */
-       xfs_ail_entry_t         m_ail;          /* fs active log item list */
-       uint                    m_ail_gen;      /* fs AIL generation count */
+       xfs_ail_t               m_ail;          /* fs active log item list */
        xfs_sb_t                m_sb;           /* copy of fs superblock */
        spinlock_t              m_sb_lock;      /* sb counter lock */
        struct xfs_buf          *m_sb_bp;       /* buffer for superblock */
index bea6e2b95a903d69011f014101c12a94d8c056ed..7f40628d85c70edc880bd4c5fc629460e77054b1 100644 (file)
@@ -992,8 +992,9 @@ int         _xfs_trans_commit(xfs_trans_t *,
                                  int *);
 #define xfs_trans_commit(tp, flags)    _xfs_trans_commit(tp, flags, NULL)
 void           xfs_trans_cancel(xfs_trans_t *, int);
-void           xfs_trans_ail_init(struct xfs_mount *);
-xfs_lsn_t      xfs_trans_push_ail(struct xfs_mount *, xfs_lsn_t);
+int            xfs_trans_ail_init(struct xfs_mount *);
+void           xfs_trans_ail_destroy(struct xfs_mount *);
+void           xfs_trans_push_ail(struct xfs_mount *, xfs_lsn_t);
 xfs_lsn_t      xfs_trans_tail_ail(struct xfs_mount *);
 void           xfs_trans_unlocked_item(struct xfs_mount *,
                                        xfs_log_item_t *);
index 2d3c297259c252bd17e268fc8ca7c05c3691d468..7ed7425478def6bafe5e4d9e5451fe000e1a92ef 100644 (file)
@@ -57,7 +57,7 @@ xfs_trans_tail_ail(
        xfs_log_item_t  *lip;
 
        spin_lock(&mp->m_ail_lock);
-       lip = xfs_ail_min(&(mp->m_ail));
+       lip = xfs_ail_min(&(mp->m_ail.xa_ail));
        if (lip == NULL) {
                lsn = (xfs_lsn_t)0;
        } else {
@@ -71,119 +71,185 @@ xfs_trans_tail_ail(
 /*
  * xfs_trans_push_ail
  *
- * This routine is called to move the tail of the AIL
- * forward.  It does this by trying to flush items in the AIL
- * whose lsns are below the given threshold_lsn.
+ * This routine is called to move the tail of the AIL forward.  It does this by
+ * trying to flush items in the AIL whose lsns are below the given
+ * threshold_lsn.
  *
- * The routine returns the lsn of the tail of the log.
+ * the push is run asynchronously in a separate thread, so we return the tail
+ * of the log right now instead of the tail after the push. This means we will
+ * either continue right away, or we will sleep waiting on the async thread to
+ * do it's work.
+ *
+ * We do this unlocked - we only need to know whether there is anything in the
+ * AIL at the time we are called. We don't need to access the contents of
+ * any of the objects, so the lock is not needed.
  */
-xfs_lsn_t
+void
 xfs_trans_push_ail(
        xfs_mount_t             *mp,
        xfs_lsn_t               threshold_lsn)
 {
-       xfs_lsn_t               lsn;
        xfs_log_item_t          *lip;
-       int                     gen;
-       int                     restarts;
-       int                     lock_result;
-       int                     flush_log;
 
-#define        XFS_TRANS_PUSH_AIL_RESTARTS     1000
+       lip = xfs_ail_min(&mp->m_ail.xa_ail);
+       if (lip && !XFS_FORCED_SHUTDOWN(mp)) {
+               if (XFS_LSN_CMP(threshold_lsn, mp->m_ail.xa_target) > 0)
+                       xfsaild_wakeup(mp, threshold_lsn);
+       }
+}
+
+/*
+ * Return the item in the AIL with the current lsn.
+ * Return the current tree generation number for use
+ * in calls to xfs_trans_next_ail().
+ */
+STATIC xfs_log_item_t *
+xfs_trans_first_push_ail(
+       xfs_mount_t     *mp,
+       int             *gen,
+       xfs_lsn_t       lsn)
+{
+       xfs_log_item_t  *lip;
+
+       lip = xfs_ail_min(&(mp->m_ail.xa_ail));
+       *gen = (int)mp->m_ail.xa_gen;
+       if (lsn == 0)
+               return lip;
+
+       while (lip && (XFS_LSN_CMP(lip->li_lsn, lsn) < 0))
+               lip = lip->li_ail.ail_forw;
+
+       return lip;
+}
+
+/*
+ * Function that does the work of pushing on the AIL
+ */
+long
+xfsaild_push(
+       xfs_mount_t     *mp,
+       xfs_lsn_t       *last_lsn)
+{
+       long            tout = 1000; /* milliseconds */
+       xfs_lsn_t       last_pushed_lsn = *last_lsn;
+       xfs_lsn_t       target =  mp->m_ail.xa_target;
+       xfs_lsn_t       lsn;
+       xfs_log_item_t  *lip;
+       int             gen;
+       int             restarts;
+       int             flush_log, count, stuck;
+
+#define        XFS_TRANS_PUSH_AIL_RESTARTS     10
 
        spin_lock(&mp->m_ail_lock);
-       lip = xfs_trans_first_ail(mp, &gen);
-       if (lip == NULL || XFS_FORCED_SHUTDOWN(mp)) {
+       lip = xfs_trans_first_push_ail(mp, &gen, *last_lsn);
+       if (!lip || XFS_FORCED_SHUTDOWN(mp)) {
                /*
-                * Just return if the AIL is empty.
+                * AIL is empty or our push has reached the end.
                 */
                spin_unlock(&mp->m_ail_lock);
-               return (xfs_lsn_t)0;
+               last_pushed_lsn = 0;
+               goto out;
        }
 
        XFS_STATS_INC(xs_push_ail);
 
        /*
         * While the item we are looking at is below the given threshold
-        * try to flush it out.  Make sure to limit the number of times
-        * we allow xfs_trans_next_ail() to restart scanning from the
-        * beginning of the list.  We'd like not to stop until we've at least
+        * try to flush it out. We'd like not to stop until we've at least
         * tried to push on everything in the AIL with an LSN less than
-        * the given threshold. However, we may give up before that if
-        * we realize that we've been holding the AIL lock for 'too long',
-        * blocking interrupts. Currently, too long is < 500us roughly.
+        * the given threshold.
+        *
+        * However, we will stop after a certain number of pushes and wait
+        * for a reduced timeout to fire before pushing further. This
+        * prevents use from spinning when we can't do anything or there is
+        * lots of contention on the AIL lists.
         */
-       flush_log = 0;
-       restarts = 0;
-       while (((restarts < XFS_TRANS_PUSH_AIL_RESTARTS) &&
-               (XFS_LSN_CMP(lip->li_lsn, threshold_lsn) < 0))) {
+       tout = 10;
+       lsn = lip->li_lsn;
+       flush_log = stuck = count = restarts = 0;
+       while ((XFS_LSN_CMP(lip->li_lsn, target) < 0)) {
+               int     lock_result;
                /*
-                * If we can lock the item without sleeping, unlock
-                * the AIL lock and flush the item.  Then re-grab the
-                * AIL lock so we can look for the next item on the
-                * AIL.  Since we unlock the AIL while we flush the
-                * item, the next routine may start over again at the
-                * the beginning of the list if anything has changed.
-                * That is what the generation count is for.
+                * If we can lock the item without sleeping, unlock the AIL
+                * lock and flush the item.  Then re-grab the AIL lock so we
+                * can look for the next item on the AIL. List changes are
+                * handled by the AIL lookup functions internally
                 *
-                * If we can't lock the item, either its holder will flush
-                * it or it is already being flushed or it is being relogged.
-                * In any of these case it is being taken care of and we
-                * can just skip to the next item in the list.
+                * If we can't lock the item, either its holder will flush it
+                * or it is already being flushed or it is being relogged.  In
+                * any of these case it is being taken care of and we can just
+                * skip to the next item in the list.
                 */
                lock_result = IOP_TRYLOCK(lip);
+               spin_unlock(&mp->m_ail_lock);
                switch (lock_result) {
-                     case XFS_ITEM_SUCCESS:
-                       spin_unlock(&mp->m_ail_lock);
+               case XFS_ITEM_SUCCESS:
                        XFS_STATS_INC(xs_push_ail_success);
                        IOP_PUSH(lip);
-                       spin_lock(&mp->m_ail_lock);
+                       last_pushed_lsn = lsn;
                        break;
 
-                     case XFS_ITEM_PUSHBUF:
-                       spin_unlock(&mp->m_ail_lock);
+               case XFS_ITEM_PUSHBUF:
                        XFS_STATS_INC(xs_push_ail_pushbuf);
-#ifdef XFSRACEDEBUG
-                       delay_for_intr();
-                       delay(300);
-#endif
-                       ASSERT(lip->li_ops->iop_pushbuf);
-                       ASSERT(lip);
                        IOP_PUSHBUF(lip);
-                       spin_lock(&mp->m_ail_lock);
+                       last_pushed_lsn = lsn;
                        break;
 
-                     case XFS_ITEM_PINNED:
+               case XFS_ITEM_PINNED:
                        XFS_STATS_INC(xs_push_ail_pinned);
+                       stuck++;
                        flush_log = 1;
                        break;
 
-                     case XFS_ITEM_LOCKED:
+               case XFS_ITEM_LOCKED:
                        XFS_STATS_INC(xs_push_ail_locked);
+                       last_pushed_lsn = lsn;
+                       stuck++;
                        break;
 
-                     case XFS_ITEM_FLUSHING:
+               case XFS_ITEM_FLUSHING:
                        XFS_STATS_INC(xs_push_ail_flushing);
+                       last_pushed_lsn = lsn;
+                       stuck++;
                        break;
 
-                     default:
+               default:
                        ASSERT(0);
                        break;
                }
 
-               lip = xfs_trans_next_ail(mp, lip, &gen, &restarts);
-               if (lip == NULL) {
+               spin_lock(&mp->m_ail_lock);
+               /* should we bother continuing? */
+               if (XFS_FORCED_SHUTDOWN(mp))
+                       break;
+               ASSERT(mp->m_log);
+
+               count++;
+
+               /*
+                * Are there too many items we can't do anything with?
+                * If we we are skipping too many items because we can't flush
+                * them or they are already being flushed, we back off and
+                * given them time to complete whatever operation is being
+                * done. i.e. remove pressure from the AIL while we can't make
+                * progress so traversals don't slow down further inserts and
+                * removals to/from the AIL.
+                *
+                * The value of 100 is an arbitrary magic number based on
+                * observation.
+                */
+               if (stuck > 100)
                        break;
-               }
-               if (XFS_FORCED_SHUTDOWN(mp)) {
-                       /*
-                        * Just return if we shut down during the last try.
-                        */
-                       spin_unlock(&mp->m_ail_lock);
-                       return (xfs_lsn_t)0;
-               }
 
+               lip = xfs_trans_next_ail(mp, lip, &gen, &restarts);
+               if (lip == NULL)
+                       break;
+               if (restarts > XFS_TRANS_PUSH_AIL_RESTARTS)
+                       break;
+               lsn = lip->li_lsn;
        }
+       spin_unlock(&mp->m_ail_lock);
 
        if (flush_log) {
                /*
@@ -191,22 +257,35 @@ xfs_trans_push_ail(
                 * push out the log so it will become unpinned and
                 * move forward in the AIL.
                 */
-               spin_unlock(&mp->m_ail_lock);
                XFS_STATS_INC(xs_push_ail_flush);
                xfs_log_force(mp, (xfs_lsn_t)0, XFS_LOG_FORCE);
-               spin_lock(&mp->m_ail_lock);
        }
 
-       lip = xfs_ail_min(&(mp->m_ail));
-       if (lip == NULL) {
-               lsn = (xfs_lsn_t)0;
-       } else {
-               lsn = lip->li_lsn;
+       /*
+        * We reached the target so wait a bit longer for I/O to complete and
+        * remove pushed items from the AIL before we start the next scan from
+        * the start of the AIL.
+        */
+       if ((XFS_LSN_CMP(lsn, target) >= 0)) {
+               tout += 20;
+               last_pushed_lsn = 0;
+       } else if ((restarts > XFS_TRANS_PUSH_AIL_RESTARTS) ||
+                  (count && ((stuck * 100) / count > 90))) {
+               /*
+                * Either there is a lot of contention on the AIL or we
+                * are stuck due to operations in progress. "Stuck" in this
+                * case is defined as >90% of the items we tried to push
+                * were stuck.
+                *
+                * Backoff a bit more to allow some I/O to complete before
+                * continuing from where we were.
+                */
+               tout += 10;
        }
-
-       spin_unlock(&mp->m_ail_lock);
-       return lsn;
-}      /* xfs_trans_push_ail */
+out:
+       *last_lsn = last_pushed_lsn;
+       return tout;
+}      /* xfsaild_push */
 
 
 /*
@@ -247,7 +326,7 @@ xfs_trans_unlocked_item(
         * the call to xfs_log_move_tail() doesn't do anything if there's
         * not enough free space to wake people up so we're safe calling it.
         */
-       min_lip = xfs_ail_min(&mp->m_ail);
+       min_lip = xfs_ail_min(&mp->m_ail.xa_ail);
 
        if (min_lip == lip)
                xfs_log_move_tail(mp, 1);
@@ -279,7 +358,7 @@ xfs_trans_update_ail(
        xfs_log_item_t          *dlip=NULL;
        xfs_log_item_t          *mlip;  /* ptr to minimum lip */
 
-       ailp = &(mp->m_ail);
+       ailp = &(mp->m_ail.xa_ail);
        mlip = xfs_ail_min(ailp);
 
        if (lip->li_flags & XFS_LI_IN_AIL) {
@@ -292,10 +371,10 @@ xfs_trans_update_ail(
        lip->li_lsn = lsn;
 
        xfs_ail_insert(ailp, lip);
-       mp->m_ail_gen++;
+       mp->m_ail.xa_gen++;
 
        if (mlip == dlip) {
-               mlip = xfs_ail_min(&(mp->m_ail));
+               mlip = xfs_ail_min(&(mp->m_ail.xa_ail));
                spin_unlock(&mp->m_ail_lock);
                xfs_log_move_tail(mp, mlip->li_lsn);
        } else {
@@ -330,7 +409,7 @@ xfs_trans_delete_ail(
        xfs_log_item_t          *mlip;
 
        if (lip->li_flags & XFS_LI_IN_AIL) {
-               ailp = &(mp->m_ail);
+               ailp = &(mp->m_ail.xa_ail);
                mlip = xfs_ail_min(ailp);
                dlip = xfs_ail_delete(ailp, lip);
                ASSERT(dlip == lip);
@@ -338,10 +417,10 @@ xfs_trans_delete_ail(
 
                lip->li_flags &= ~XFS_LI_IN_AIL;
                lip->li_lsn = 0;
-               mp->m_ail_gen++;
+               mp->m_ail.xa_gen++;
 
                if (mlip == dlip) {
-                       mlip = xfs_ail_min(&(mp->m_ail));
+                       mlip = xfs_ail_min(&(mp->m_ail.xa_ail));
                        spin_unlock(&mp->m_ail_lock);
                        xfs_log_move_tail(mp, (mlip ? mlip->li_lsn : 0));
                } else {
@@ -379,10 +458,10 @@ xfs_trans_first_ail(
 {
        xfs_log_item_t  *lip;
 
-       lip = xfs_ail_min(&(mp->m_ail));
-       *gen = (int)mp->m_ail_gen;
+       lip = xfs_ail_min(&(mp->m_ail.xa_ail));
+       *gen = (int)mp->m_ail.xa_gen;
 
-       return (lip);
+       return lip;
 }
 
 /*
@@ -402,11 +481,11 @@ xfs_trans_next_ail(
        xfs_log_item_t  *nlip;
 
        ASSERT(mp && lip && gen);
-       if (mp->m_ail_gen == *gen) {
-               nlip = xfs_ail_next(&(mp->m_ail), lip);
+       if (mp->m_ail.xa_gen == *gen) {
+               nlip = xfs_ail_next(&(mp->m_ail.xa_ail), lip);
        } else {
-               nlip = xfs_ail_min(&(mp->m_ail));
-               *gen = (int)mp->m_ail_gen;
+               nlip = xfs_ail_min(&(mp->m_ail).xa_ail);
+               *gen = (int)mp->m_ail.xa_gen;
                if (restarts != NULL) {
                        XFS_STATS_INC(xs_push_ail_restarts);
                        (*restarts)++;
@@ -431,12 +510,20 @@ xfs_trans_next_ail(
 /*
  * Initialize the doubly linked list to point only to itself.
  */
-void
+int
 xfs_trans_ail_init(
        xfs_mount_t     *mp)
 {
-       mp->m_ail.ail_forw = (xfs_log_item_t*)&(mp->m_ail);
-       mp->m_ail.ail_back = (xfs_log_item_t*)&(mp->m_ail);
+       mp->m_ail.xa_ail.ail_forw = (xfs_log_item_t*)&mp->m_ail.xa_ail;
+       mp->m_ail.xa_ail.ail_back = (xfs_log_item_t*)&mp->m_ail.xa_ail;
+       return xfsaild_start(mp);
+}
+
+void
+xfs_trans_ail_destroy(
+       xfs_mount_t     *mp)
+{
+       xfsaild_stop(mp);
 }
 
 /*
index c40f0512097ad92549dd6ee5be03f21be2db9375..3c748c456ed4f9a58c03babd86d9afd401ca8efe 100644 (file)
@@ -57,4 +57,12 @@ struct xfs_log_item  *xfs_trans_next_ail(struct xfs_mount *,
                                     struct xfs_log_item *, int *, int *);
 
 
+/*
+ * AIL push thread support
+ */
+long   xfsaild_push(struct xfs_mount *, xfs_lsn_t *);
+void   xfsaild_wakeup(struct xfs_mount *, xfs_lsn_t);
+int    xfsaild_start(struct xfs_mount *);
+void   xfsaild_stop(struct xfs_mount *);
+
 #endif /* __XFS_TRANS_PRIV_H__ */