xfs: split out iclog writing from xfs_trans_commit()
authorDave Chinner <dchinner@redhat.com>
Mon, 8 Mar 2010 00:28:28 +0000 (11:28 +1100)
committerAlex Elder <aelder@sgi.com>
Wed, 19 May 2010 14:58:08 +0000 (09:58 -0500)
Split the the part of xfs_trans_commit() that deals with writing the
transaction into the iclog into a separate function. This isolates the
physical commit process from the logical commit operation and makes
it easier to insert different transaction commit paths without affecting
the existing algorithm adversely.

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

index 6962f2b..e07b329 100644 (file)
@@ -47,8 +47,6 @@
 
 
 STATIC void    xfs_trans_apply_sb_deltas(xfs_trans_t *);
-STATIC uint    xfs_trans_count_vecs(xfs_trans_t *);
-STATIC void    xfs_trans_fill_vecs(xfs_trans_t *, xfs_log_iovec_t *);
 STATIC void    xfs_trans_uncommit(xfs_trans_t *, uint);
 STATIC void    xfs_trans_committed(xfs_trans_t *, int);
 STATIC void    xfs_trans_chunk_committed(xfs_log_item_chunk_t *, xfs_lsn_t, int);
@@ -764,94 +762,126 @@ xfs_trans_unreserve_and_mod_sb(
        }
 }
 
-
 /*
- * xfs_trans_commit
- *
- * Commit the given transaction to the log a/synchronously.
- *
- * XFS disk error handling mechanism is not based on a typical
- * transaction abort mechanism. Logically after the filesystem
- * gets marked 'SHUTDOWN', we can't let any new transactions
- * be durable - ie. committed to disk - because some metadata might
- * be inconsistent. In such cases, this returns an error, and the
- * caller may assume that all locked objects joined to the transaction
- * have already been unlocked as if the commit had succeeded.
- * Do not reference the transaction structure after this call.
+ * Total up the number of log iovecs needed to commit this
+ * transaction.  The transaction itself needs one for the
+ * transaction header.  Ask each dirty item in turn how many
+ * it needs to get the total.
  */
- /*ARGSUSED*/
-int
-_xfs_trans_commit(
-       xfs_trans_t     *tp,
-       uint            flags,
-       int             *log_flushed)
+static uint
+xfs_trans_count_vecs(
+       xfs_trans_t     *tp)
 {
-       xfs_log_iovec_t         *log_vector;
-       int                     nvec;
-       xfs_mount_t             *mp;
-       xfs_lsn_t               commit_lsn;
-       /* REFERENCED */
-       int                     error;
-       int                     log_flags;
-       int                     sync;
-#define        XFS_TRANS_LOGVEC_COUNT  16
-       xfs_log_iovec_t         log_vector_fast[XFS_TRANS_LOGVEC_COUNT];
-       struct xlog_in_core     *commit_iclog;
-       int                     shutdown;
+       int                     nvecs;
+       xfs_log_item_desc_t     *lidp;
 
-       commit_lsn = -1;
+       nvecs = 1;
+       lidp = xfs_trans_first_item(tp);
+       ASSERT(lidp != NULL);
 
-       /*
-        * Determine whether this commit is releasing a permanent
-        * log reservation or not.
+       /* In the non-debug case we need to start bailing out if we
+        * didn't find a log_item here, return zero and let trans_commit
+        * deal with it.
         */
-       if (flags & XFS_TRANS_RELEASE_LOG_RES) {
-               ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
-               log_flags = XFS_LOG_REL_PERM_RESERV;
-       } else {
-               log_flags = 0;
+       if (lidp == NULL)
+               return 0;
+
+       while (lidp != NULL) {
+               /*
+                * Skip items which aren't dirty in this transaction.
+                */
+               if (!(lidp->lid_flags & XFS_LID_DIRTY)) {
+                       lidp = xfs_trans_next_item(tp, lidp);
+                       continue;
+               }
+               lidp->lid_size = IOP_SIZE(lidp->lid_item);
+               nvecs += lidp->lid_size;
+               lidp = xfs_trans_next_item(tp, lidp);
        }
-       mp = tp->t_mountp;
+
+       return nvecs;
+}
+
+/*
+ * Fill in the vector with pointers to data to be logged
+ * by this transaction.  The transaction header takes
+ * the first vector, and then each dirty item takes the
+ * number of vectors it indicated it needed in xfs_trans_count_vecs().
+ *
+ * As each item fills in the entries it needs, also pin the item
+ * so that it cannot be flushed out until the log write completes.
+ */
+static void
+xfs_trans_fill_vecs(
+       struct xfs_trans        *tp,
+       struct xfs_log_iovec    *log_vector)
+{
+       xfs_log_item_desc_t     *lidp;
+       struct xfs_log_iovec    *vecp;
+       uint                    nitems;
 
        /*
-        * If there is nothing to be logged by the transaction,
-        * then unlock all of the items associated with the
-        * transaction and free the transaction structure.
-        * Also make sure to return any reserved blocks to
-        * the free pool.
+        * Skip over the entry for the transaction header, we'll
+        * fill that in at the end.
         */
-shut_us_down:
-       shutdown = XFS_FORCED_SHUTDOWN(mp) ? EIO : 0;
-       if (!(tp->t_flags & XFS_TRANS_DIRTY) || shutdown) {
-               xfs_trans_unreserve_and_mod_sb(tp);
+       vecp = log_vector + 1;
+
+       nitems = 0;
+       lidp = xfs_trans_first_item(tp);
+       ASSERT(lidp);
+       while (lidp) {
+               /* Skip items which aren't dirty in this transaction. */
+               if (!(lidp->lid_flags & XFS_LID_DIRTY)) {
+                       lidp = xfs_trans_next_item(tp, lidp);
+                       continue;
+               }
+
                /*
-                * It is indeed possible for the transaction to be
-                * not dirty but the dqinfo portion to be. All that
-                * means is that we have some (non-persistent) quota
-                * reservations that need to be unreserved.
+                * The item may be marked dirty but not log anything.  This can
+                * be used to get called when a transaction is committed.
                 */
-               xfs_trans_unreserve_and_mod_dquots(tp);
-               if (tp->t_ticket) {
-                       commit_lsn = xfs_log_done(mp, tp->t_ticket,
-                                                       NULL, log_flags);
-                       if (commit_lsn == -1 && !shutdown)
-                               shutdown = XFS_ERROR(EIO);
-               }
-               current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS);
-               xfs_trans_free_items(tp, shutdown? XFS_TRANS_ABORT : 0);
-               xfs_trans_free_busy(tp);
-               xfs_trans_free(tp);
-               XFS_STATS_INC(xs_trans_empty);
-               return (shutdown);
+               if (lidp->lid_size)
+                       nitems++;
+               IOP_FORMAT(lidp->lid_item, vecp);
+               vecp += lidp->lid_size;
+               IOP_PIN(lidp->lid_item);
+               lidp = xfs_trans_next_item(tp, lidp);
        }
-       ASSERT(tp->t_ticket != NULL);
 
        /*
-        * If we need to update the superblock, then do it now.
+        * Now that we've counted the number of items in this transaction, fill
+        * in the transaction header. Note that the transaction header does not
+        * have a log item.
         */
-       if (tp->t_flags & XFS_TRANS_SB_DIRTY)
-               xfs_trans_apply_sb_deltas(tp);
-       xfs_trans_apply_dquot_deltas(tp);
+       tp->t_header.th_magic = XFS_TRANS_HEADER_MAGIC;
+       tp->t_header.th_type = tp->t_type;
+       tp->t_header.th_num_items = nitems;
+       log_vector->i_addr = (xfs_caddr_t)&tp->t_header;
+       log_vector->i_len = sizeof(xfs_trans_header_t);
+       log_vector->i_type = XLOG_REG_TYPE_TRANSHDR;
+}
+
+/*
+ * Format the transaction direct to the iclog. This isolates the physical
+ * transaction commit operation from the logical operation and hence allows
+ * other methods to be introduced without affecting the existing commit path.
+ */
+static int
+xfs_trans_commit_iclog(
+       struct xfs_mount        *mp,
+       struct xfs_trans        *tp,
+       xfs_lsn_t               *commit_lsn,
+       int                     flags)
+{
+       int                     shutdown;
+       int                     error;
+       int                     log_flags = 0;
+       struct xlog_in_core     *commit_iclog;
+#define XFS_TRANS_LOGVEC_COUNT  16
+       struct xfs_log_iovec    log_vector_fast[XFS_TRANS_LOGVEC_COUNT];
+       struct xfs_log_iovec    *log_vector;
+       uint                    nvec;
+
 
        /*
         * Ask each log item how many log_vector entries it will
@@ -861,8 +891,7 @@ shut_us_down:
         */
        nvec = xfs_trans_count_vecs(tp);
        if (nvec == 0) {
-               xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
-               goto shut_us_down;
+               return ENOMEM;  /* triggers a shutdown! */
        } else if (nvec <= XFS_TRANS_LOGVEC_COUNT) {
                log_vector = log_vector_fast;
        } else {
@@ -877,6 +906,9 @@ shut_us_down:
         */
        xfs_trans_fill_vecs(tp, log_vector);
 
+       if (flags & XFS_TRANS_RELEASE_LOG_RES)
+               log_flags = XFS_LOG_REL_PERM_RESERV;
+
        error = xfs_log_write(mp, log_vector, nvec, tp->t_ticket, &(tp->t_lsn));
 
        /*
@@ -884,18 +916,17 @@ shut_us_down:
         * at any time after this call.  However, all the items associated
         * with the transaction are still locked and pinned in memory.
         */
-       commit_lsn = xfs_log_done(mp, tp->t_ticket, &commit_iclog, log_flags);
+       *commit_lsn = xfs_log_done(mp, tp->t_ticket, &commit_iclog, log_flags);
 
-       tp->t_commit_lsn = commit_lsn;
-       if (nvec > XFS_TRANS_LOGVEC_COUNT) {
+       tp->t_commit_lsn = *commit_lsn;
+       if (nvec > XFS_TRANS_LOGVEC_COUNT)
                kmem_free(log_vector);
-       }
 
        /*
         * If we got a log write error. Unpin the logitems that we
         * had pinned, clean up, free trans structure, and return error.
         */
-       if (error || commit_lsn == -1) {
+       if (error || *commit_lsn == -1) {
                current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS);
                xfs_trans_uncommit(tp, flags|XFS_TRANS_ABORT);
                return XFS_ERROR(EIO);
@@ -909,8 +940,6 @@ shut_us_down:
         */
        xfs_trans_unreserve_and_mod_sb(tp);
 
-       sync = tp->t_flags & XFS_TRANS_SYNC;
-
        /*
         * Tell the LM to call the transaction completion routine
         * when the log write with LSN commit_lsn completes (e.g.
@@ -953,7 +982,7 @@ shut_us_down:
         * the commit lsn of this transaction for dependency tracking
         * purposes.
         */
-       xfs_trans_unlock_items(tp, commit_lsn);
+       xfs_trans_unlock_items(tp, *commit_lsn);
 
        /*
         * If we detected a log error earlier, finish committing
@@ -973,7 +1002,92 @@ shut_us_down:
         * and the items are released we can finally allow the iclog to
         * go to disk.
         */
-       error = xfs_log_release_iclog(mp, commit_iclog);
+       return xfs_log_release_iclog(mp, commit_iclog);
+}
+
+
+/*
+ * xfs_trans_commit
+ *
+ * Commit the given transaction to the log a/synchronously.
+ *
+ * XFS disk error handling mechanism is not based on a typical
+ * transaction abort mechanism. Logically after the filesystem
+ * gets marked 'SHUTDOWN', we can't let any new transactions
+ * be durable - ie. committed to disk - because some metadata might
+ * be inconsistent. In such cases, this returns an error, and the
+ * caller may assume that all locked objects joined to the transaction
+ * have already been unlocked as if the commit had succeeded.
+ * Do not reference the transaction structure after this call.
+ */
+ /*ARGSUSED*/
+int
+_xfs_trans_commit(
+       xfs_trans_t     *tp,
+       uint            flags,
+       int             *log_flushed)
+{
+       xfs_mount_t             *mp = tp->t_mountp;
+       xfs_lsn_t               commit_lsn = -1;
+       int                     error;
+       int                     log_flags = 0;
+       int                     sync = tp->t_flags & XFS_TRANS_SYNC;
+       int                     shutdown;
+
+       /*
+        * Determine whether this commit is releasing a permanent
+        * log reservation or not.
+        */
+       if (flags & XFS_TRANS_RELEASE_LOG_RES) {
+               ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
+               log_flags = XFS_LOG_REL_PERM_RESERV;
+       }
+
+       /*
+        * If there is nothing to be logged by the transaction,
+        * then unlock all of the items associated with the
+        * transaction and free the transaction structure.
+        * Also make sure to return any reserved blocks to
+        * the free pool.
+        */
+shut_us_down:
+       shutdown = XFS_FORCED_SHUTDOWN(mp) ? EIO : 0;
+       if (!(tp->t_flags & XFS_TRANS_DIRTY) || shutdown) {
+               xfs_trans_unreserve_and_mod_sb(tp);
+               /*
+                * It is indeed possible for the transaction to be
+                * not dirty but the dqinfo portion to be. All that
+                * means is that we have some (non-persistent) quota
+                * reservations that need to be unreserved.
+                */
+               xfs_trans_unreserve_and_mod_dquots(tp);
+               if (tp->t_ticket) {
+                       commit_lsn = xfs_log_done(mp, tp->t_ticket,
+                                                       NULL, log_flags);
+                       if (commit_lsn == -1 && !shutdown)
+                               shutdown = XFS_ERROR(EIO);
+               }
+               current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS);
+               xfs_trans_free_items(tp, shutdown? XFS_TRANS_ABORT : 0);
+               xfs_trans_free_busy(tp);
+               xfs_trans_free(tp);
+               XFS_STATS_INC(xs_trans_empty);
+               return (shutdown);
+       }
+       ASSERT(tp->t_ticket != NULL);
+
+       /*
+        * If we need to update the superblock, then do it now.
+        */
+       if (tp->t_flags & XFS_TRANS_SB_DIRTY)
+               xfs_trans_apply_sb_deltas(tp);
+       xfs_trans_apply_dquot_deltas(tp);
+
+       error = xfs_trans_commit_iclog(mp, tp, &commit_lsn, flags);
+       if (error == ENOMEM) {
+               xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
+               goto shut_us_down;
+       }
 
        /*
         * If the transaction needs to be synchronous, then force the
@@ -992,47 +1106,6 @@ shut_us_down:
        return (error);
 }
 
-
-/*
- * Total up the number of log iovecs needed to commit this
- * transaction.  The transaction itself needs one for the
- * transaction header.  Ask each dirty item in turn how many
- * it needs to get the total.
- */
-STATIC uint
-xfs_trans_count_vecs(
-       xfs_trans_t     *tp)
-{
-       int                     nvecs;
-       xfs_log_item_desc_t     *lidp;
-
-       nvecs = 1;
-       lidp = xfs_trans_first_item(tp);
-       ASSERT(lidp != NULL);
-
-       /* In the non-debug case we need to start bailing out if we
-        * didn't find a log_item here, return zero and let trans_commit
-        * deal with it.
-        */
-       if (lidp == NULL)
-               return 0;
-
-       while (lidp != NULL) {
-               /*
-                * Skip items which aren't dirty in this transaction.
-                */
-               if (!(lidp->lid_flags & XFS_LID_DIRTY)) {
-                       lidp = xfs_trans_next_item(tp, lidp);
-                       continue;
-               }
-               lidp->lid_size = IOP_SIZE(lidp->lid_item);
-               nvecs += lidp->lid_size;
-               lidp = xfs_trans_next_item(tp, lidp);
-       }
-
-       return nvecs;
-}
-
 /*
  * Called from the trans_commit code when we notice that
  * the filesystem is in the middle of a forced shutdown.
@@ -1062,68 +1135,6 @@ xfs_trans_uncommit(
        xfs_trans_free(tp);
 }
 
-/*
- * Fill in the vector with pointers to data to be logged
- * by this transaction.  The transaction header takes
- * the first vector, and then each dirty item takes the
- * number of vectors it indicated it needed in xfs_trans_count_vecs().
- *
- * As each item fills in the entries it needs, also pin the item
- * so that it cannot be flushed out until the log write completes.
- */
-STATIC void
-xfs_trans_fill_vecs(
-       xfs_trans_t             *tp,
-       xfs_log_iovec_t         *log_vector)
-{
-       xfs_log_item_desc_t     *lidp;
-       xfs_log_iovec_t         *vecp;
-       uint                    nitems;
-
-       /*
-        * Skip over the entry for the transaction header, we'll
-        * fill that in at the end.
-        */
-       vecp = log_vector + 1;          /* pointer arithmetic */
-
-       nitems = 0;
-       lidp = xfs_trans_first_item(tp);
-       ASSERT(lidp != NULL);
-       while (lidp != NULL) {
-               /*
-                * Skip items which aren't dirty in this transaction.
-                */
-               if (!(lidp->lid_flags & XFS_LID_DIRTY)) {
-                       lidp = xfs_trans_next_item(tp, lidp);
-                       continue;
-               }
-               /*
-                * The item may be marked dirty but not log anything.
-                * This can be used to get called when a transaction
-                * is committed.
-                */
-               if (lidp->lid_size) {
-                       nitems++;
-               }
-               IOP_FORMAT(lidp->lid_item, vecp);
-               vecp += lidp->lid_size;         /* pointer arithmetic */
-               IOP_PIN(lidp->lid_item);
-               lidp = xfs_trans_next_item(tp, lidp);
-       }
-
-       /*
-        * Now that we've counted the number of items in this
-        * transaction, fill in the transaction header.
-        */
-       tp->t_header.th_magic = XFS_TRANS_HEADER_MAGIC;
-       tp->t_header.th_type = tp->t_type;
-       tp->t_header.th_num_items = nitems;
-       log_vector->i_addr = (xfs_caddr_t)&tp->t_header;
-       log_vector->i_len = sizeof(xfs_trans_header_t);
-       log_vector->i_type = XLOG_REG_TYPE_TRANSHDR;
-}
-
-
 /*
  * Unlock all of the transaction's items and free the transaction.
  * The transaction must not have modified any of its items, because