[XFS] Fix inode allocation latency

The log force added in xfs_iget_core() has been a performance issue since
it was introduced for tight loops that allocate then unlink a single file.
under heavy writeback, this can introduce unnecessary latency due tothe
log I/o getting stuck behind bulk data writes.

Fix this latency problem by avoinding the need for the log force by moving
the place we mark linux inode dirty to the transaction commit rather than
on transaction completion.

This also closes a potential hole in the sync code where a linux inode is
not dirty between the time it is modified and the time the log buffer has
been written to disk.

SGI-PV: 972753
SGI-Modid: xfs-linux-melb:xfs-kern:30007a

Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Christoph Hellwig <hch@infradead.org>
Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>

5 files changed, 23 insertions, 51 deletions

diff --git a/fs/xfs/linux-2.6/xfs_iops.c b/fs/xfs/linux-2.6/xfs_iops.c
index b5afcfcdc7d5..264b1e7dacf7 100644
--- a/fs/xfs/linux-2.6/xfs_iops.c
+++ b/fs/xfs/linux-2.6/xfs_iops.c
@@ -71,6 +71,22 @@ xfs_synchronize_atime(
 }
 
 /*
+ * If the linux inode exists, mark it dirty.
+ * Used when commiting a dirty inode into a transaction so that
+ * the inode will get written back by the linux code
+ */
+void
+xfs_mark_inode_dirty_sync(
+        xfs_inode_t     *ip)
+{
+        bhv_vnode_t     *vp;
+
+        vp = XFS_ITOV_NULL(ip);
+        if (vp)
+                mark_inode_dirty_sync(vn_to_inode(vp));
+}
+
+/*
  * Change the requested timestamp in the given inode.
  * We don't lock across timestamp updates, and we don't log them but
  * we do record the fact that there is dirty information in core.
diff --git a/fs/xfs/xfs_iget.c b/fs/xfs/xfs_iget.c
index eecc33d3751f..f01b07687faf 100644
--- a/fs/xfs/xfs_iget.c
+++ b/fs/xfs/xfs_iget.c
@@ -140,27 +140,9 @@ again:
                                 return ENOENT;
                         }
 
-                        /*
-                         * There may be transactions sitting in the
-                         * incore log buffers or being flushed to disk
-                         * at this time.  We can't clear the
-                         * XFS_IRECLAIMABLE flag until these
-                         * transactions have hit the disk, otherwise we
-                         * will void the guarantee the flag provides
-                         * xfs_iunpin()
-                         */
-                        if (xfs_ipincount(ip)) {
-                                read_unlock(&pag->pag_ici_lock);
-                                xfs_log_force(mp, 0,
-                                        XFS_LOG_FORCE|XFS_LOG_SYNC);
-                                XFS_STATS_INC(xs_ig_frecycle);
-                                goto again;
-                        }
-
                         xfs_itrace_exit_tag(ip, "xfs_iget.alloc");
 
                         XFS_STATS_INC(xs_ig_found);
-
                         xfs_iflags_clear(ip, XFS_IRECLAIMABLE);
                         read_unlock(&pag->pag_ici_lock);
 
diff --git a/fs/xfs/xfs_inode.c b/fs/xfs/xfs_inode.c
index 597e0ed4d2b6..805cab7b2770 100644
--- a/fs/xfs/xfs_inode.c
+++ b/fs/xfs/xfs_inode.c
@@ -2814,40 +2814,8 @@ xfs_iunpin(
 {
         ASSERT(atomic_read(&ip->i_pincount) > 0);
 
-        if (atomic_dec_and_lock(&ip->i_pincount, &ip->i_flags_lock)) {
-
-                /*
-                 * If the inode is currently being reclaimed, the link between
-                 * the bhv_vnode and the xfs_inode will be broken after the
-                 * XFS_IRECLAIM* flag is set. Hence, if these flags are not
-                 * set, then we can move forward and mark the linux inode dirty
-                 * knowing that it is still valid as it won't freed until after
-                 * the bhv_vnode<->xfs_inode link is broken in xfs_reclaim. The
-                 * i_flags_lock is used to synchronise the setting of the
-                 * XFS_IRECLAIM* flags and the breaking of the link, and so we
-                 * can execute atomically w.r.t to reclaim by holding this lock
-                 * here.
-                 *
-                 * However, we still need to issue the unpin wakeup call as the
-                 * inode reclaim may be blocked waiting for the inode to become
-                 * unpinned.
-                 */
-
-                if (!__xfs_iflags_test(ip, XFS_IRECLAIM|XFS_IRECLAIMABLE)) {
-                        bhv_vnode_t     *vp = XFS_ITOV_NULL(ip);
-                        struct inode *inode = NULL;
-
-                        BUG_ON(vp == NULL);
-                        inode = vn_to_inode(vp);
-                        BUG_ON(inode->i_state & I_CLEAR);
-
-                        /* make sync come back and flush this inode */
-                        if (!(inode->i_state & (I_NEW|I_FREEING)))
-                                mark_inode_dirty_sync(inode);
-                }
-                spin_unlock(&ip->i_flags_lock);
+        if (atomic_dec_and_test(&ip->i_pincount))
                 wake_up(&ip->i_ipin_wait);
-        }
 }
 
 /*
diff --git a/fs/xfs/xfs_inode.h b/fs/xfs/xfs_inode.h
index d8ed51e28cbb..bc869fd2f6ef 100644
--- a/fs/xfs/xfs_inode.h
+++ b/fs/xfs/xfs_inode.h
@@ -532,6 +532,7 @@ xfs_fsize_t	xfs_file_last_byte(xfs_inode_t *);
 void            xfs_lock_inodes(xfs_inode_t **, int, int, uint);
 
 void            xfs_synchronize_atime(xfs_inode_t *);
+void            xfs_mark_inode_dirty_sync(xfs_inode_t *);
 
 xfs_bmbt_rec_host_t *xfs_iext_get_ext(xfs_ifork_t *, xfs_extnum_t);
 void            xfs_iext_insert(xfs_ifork_t *, xfs_extnum_t, xfs_extnum_t,
diff --git a/fs/xfs/xfs_inode_item.c b/fs/xfs/xfs_inode_item.c
index e365b137ee4f..034ca7202295 100644
--- a/fs/xfs/xfs_inode_item.c
+++ b/fs/xfs/xfs_inode_item.c
@@ -274,6 +274,11 @@ xfs_inode_item_format(
          */
         xfs_synchronize_atime(ip);
 
+        /*
+         * make sure the linux inode is dirty
+         */
+        xfs_mark_inode_dirty_sync(ip);
+
         vecp->i_addr = (xfs_caddr_t)&ip->i_d;
         vecp->i_len  = sizeof(xfs_dinode_core_t);
         XLOG_VEC_SET_TYPE(vecp, XLOG_REG_TYPE_ICORE);