xfs: fix locking in xfs_iget_cache_hit
authorChristoph Hellwig <hch@infradead.org>
Mon, 17 Aug 2009 00:36:34 +0000 (20:36 -0400)
committerFelix Blyakher <felixb@sgi.com>
Mon, 17 Aug 2009 06:23:48 +0000 (01:23 -0500)
The locking in xfs_iget_cache_hit currently has numerous problems:

 - we clear the reclaim tag without i_flags_lock which protects
   modifications to it
 - we call inode_init_always which can sleep with pag_ici_lock
   held (this is oss.sgi.com BZ #819)
 - we acquire and drop i_flags_lock a lot and thus provide no
   consistency between the various flags we set/clear under it

This patch fixes all that with a major revamp of the locking in
the function.  The new version acquires i_flags_lock early and
only drops it once we need to call into inode_init_always or before
calling xfs_ilock.

This patch fixes a bug seen in the wild where we race modifying the
reclaim tag.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Felix Blyakher <felixb@sgi.com>
Reviewed-by: Eric Sandeen <sandeen@sandeen.net>
Signed-off-by: Felix Blyakher <felixb@sgi.com>
fs/xfs/linux-2.6/xfs_sync.c
fs/xfs/linux-2.6/xfs_sync.h
fs/xfs/xfs_iget.c

index b619d6b8ca4323afc02abab54d967f352d899e6a..98ef624d9baf65a2bb3503d0fb6c599fdbdfd02f 100644 (file)
@@ -708,6 +708,16 @@ xfs_reclaim_inode(
        return 0;
 }
 
+void
+__xfs_inode_set_reclaim_tag(
+       struct xfs_perag        *pag,
+       struct xfs_inode        *ip)
+{
+       radix_tree_tag_set(&pag->pag_ici_root,
+                          XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino),
+                          XFS_ICI_RECLAIM_TAG);
+}
+
 /*
  * We set the inode flag atomically with the radix tree tag.
  * Once we get tag lookups on the radix tree, this inode flag
@@ -722,8 +732,7 @@ xfs_inode_set_reclaim_tag(
 
        read_lock(&pag->pag_ici_lock);
        spin_lock(&ip->i_flags_lock);
-       radix_tree_tag_set(&pag->pag_ici_root,
-                       XFS_INO_TO_AGINO(mp, ip->i_ino), XFS_ICI_RECLAIM_TAG);
+       __xfs_inode_set_reclaim_tag(pag, ip);
        __xfs_iflags_set(ip, XFS_IRECLAIMABLE);
        spin_unlock(&ip->i_flags_lock);
        read_unlock(&pag->pag_ici_lock);
index 2a10301c99c79b8bf73e3d89a73f965ef580b23a..59120602588a370168282a3177a66c7a477bbb42 100644 (file)
@@ -48,6 +48,7 @@ int xfs_reclaim_inode(struct xfs_inode *ip, int locked, int sync_mode);
 int xfs_reclaim_inodes(struct xfs_mount *mp, int mode);
 
 void xfs_inode_set_reclaim_tag(struct xfs_inode *ip);
+void __xfs_inode_set_reclaim_tag(struct xfs_perag *pag, struct xfs_inode *ip);
 void xfs_inode_clear_reclaim_tag(struct xfs_inode *ip);
 void __xfs_inode_clear_reclaim_tag(struct xfs_mount *mp, struct xfs_perag *pag,
                                struct xfs_inode *ip);
index 34ec86923f7e354044fbe9ffa27a77126ce354cf..ecbf8b4d2e2e4d0f8ae4506b64a4b787dc84e5db 100644 (file)
@@ -191,80 +191,82 @@ xfs_iget_cache_hit(
        int                     flags,
        int                     lock_flags) __releases(pag->pag_ici_lock)
 {
+       struct inode            *inode = VFS_I(ip);
        struct xfs_mount        *mp = ip->i_mount;
-       int                     error = EAGAIN;
+       int                     error;
+
+       spin_lock(&ip->i_flags_lock);
 
        /*
-        * If INEW is set this inode is being set up
-        * If IRECLAIM is set this inode is being torn down
-        * Pause and try again.
+        * If we are racing with another cache hit that is currently
+        * instantiating this inode or currently recycling it out of
+        * reclaimabe state, wait for the initialisation to complete
+        * before continuing.
+        *
+        * XXX(hch): eventually we should do something equivalent to
+        *           wait_on_inode to wait for these flags to be cleared
+        *           instead of polling for it.
         */
-       if (xfs_iflags_test(ip, (XFS_INEW|XFS_IRECLAIM))) {
+       if (ip->i_flags & (XFS_INEW|XFS_IRECLAIM)) {
                XFS_STATS_INC(xs_ig_frecycle);
+               error = EAGAIN;
                goto out_error;
        }
 
-       /* If IRECLAIMABLE is set, we've torn down the vfs inode part */
-       if (xfs_iflags_test(ip, XFS_IRECLAIMABLE)) {
-
-               /*
-                * If lookup is racing with unlink, then we should return an
-                * error immediately so we don't remove it from the reclaim
-                * list and potentially leak the inode.
-                */
-               if ((ip->i_d.di_mode == 0) && !(flags & XFS_IGET_CREATE)) {
-                       error = ENOENT;
-                       goto out_error;
-               }
+       /*
+        * If lookup is racing with unlink return an error immediately.
+        */
+       if (ip->i_d.di_mode == 0 && !(flags & XFS_IGET_CREATE)) {
+               error = ENOENT;
+               goto out_error;
+       }
 
+       /*
+        * If IRECLAIMABLE is set, we've torn down the VFS inode already.
+        * Need to carefully get it back into useable state.
+        */
+       if (ip->i_flags & XFS_IRECLAIMABLE) {
                xfs_itrace_exit_tag(ip, "xfs_iget.alloc");
 
                /*
-                * We need to re-initialise the VFS inode as it has been
-                * 'freed' by the VFS. Do this here so we can deal with
-                * errors cleanly, then tag it so it can be set up correctly
-                * later.
+                * We need to set XFS_INEW atomically with clearing the
+                * reclaimable tag so that we do have an indicator of the
+                * inode still being initialized.
                 */
-               if (inode_init_always(mp->m_super, VFS_I(ip))) {
-                       error = ENOMEM;
-                       goto out_error;
-               }
+               ip->i_flags |= XFS_INEW;
+               ip->i_flags &= ~XFS_IRECLAIMABLE;
+               __xfs_inode_clear_reclaim_tag(mp, pag, ip);
 
-               /*
-                * We must set the XFS_INEW flag before clearing the
-                * XFS_IRECLAIMABLE flag so that if a racing lookup does
-                * not find the XFS_IRECLAIMABLE above but has the igrab()
-                * below succeed we can safely check XFS_INEW to detect
-                * that this inode is still being initialised.
-                */
-               xfs_iflags_set(ip, XFS_INEW);
-               xfs_iflags_clear(ip, XFS_IRECLAIMABLE);
+               spin_unlock(&ip->i_flags_lock);
+               read_unlock(&pag->pag_ici_lock);
 
-               /* clear the radix tree reclaim flag as well. */
-               __xfs_inode_clear_reclaim_tag(mp, pag, ip);
-       } else if (!igrab(VFS_I(ip))) {
+               error = -inode_init_always(mp->m_super, inode);
+               if (error) {
+                       /*
+                        * Re-initializing the inode failed, and we are in deep
+                        * trouble.  Try to re-add it to the reclaim list.
+                        */
+                       read_lock(&pag->pag_ici_lock);
+                       spin_lock(&ip->i_flags_lock);
+
+                       ip->i_flags &= ~XFS_INEW;
+                       ip->i_flags |= XFS_IRECLAIMABLE;
+                       __xfs_inode_set_reclaim_tag(pag, ip);
+                       goto out_error;
+               }
+               inode->i_state = I_LOCK|I_NEW;
+       } else {
                /* If the VFS inode is being torn down, pause and try again. */
-               XFS_STATS_INC(xs_ig_frecycle);
-               goto out_error;
-       } else if (xfs_iflags_test(ip, XFS_INEW)) {
-               /*
-                * We are racing with another cache hit that is
-                * currently recycling this inode out of the XFS_IRECLAIMABLE
-                * state. Wait for the initialisation to complete before
-                * continuing.
-                */
-               wait_on_inode(VFS_I(ip));
-       }
+               if (!igrab(inode)) {
+                       error = EAGAIN;
+                       goto out_error;
+               }
 
-       if (ip->i_d.di_mode == 0 && !(flags & XFS_IGET_CREATE)) {
-               error = ENOENT;
-               iput(VFS_I(ip));
-               goto out_error;
+               /* We've got a live one. */
+               spin_unlock(&ip->i_flags_lock);
+               read_unlock(&pag->pag_ici_lock);
        }
 
-       /* We've got a live one. */
-       read_unlock(&pag->pag_ici_lock);
-
        if (lock_flags != 0)
                xfs_ilock(ip, lock_flags);
 
@@ -274,6 +276,7 @@ xfs_iget_cache_hit(
        return 0;
 
 out_error:
+       spin_unlock(&ip->i_flags_lock);
        read_unlock(&pag->pag_ici_lock);
        return error;
 }