Skip to contentSkip to conversation

xfs: fix mmap_sem vs iolock lock order inversion in xfs_free_eofblocks

PATCH In review Christoph Hellwig 2 participants 4 messages Oct 19, 2009 – Nov 3, 2009 XFS filesystem

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 xfs: fix mmap_sem vs iolock lock order inversion in xfs_free_eofblocks

When xfs_free_eofblocks is called from ->release the VM might already hold the mmap_sem, but in the write path we take the iolock before taking the mmap_sem in the generic write code.

Switch xfs_free_eofblocks to only trylock the iolock if called from ->release and skip trimming the prellocated blocks in that case. We'll still free them later on the final iput.

fs/xfs/xfs_rw.h +0 -7
Index: xfs/fs/xfs/xfs_rw.h
===================================================================
@@ -37,13 +37,6 @@ xfs_fsb_to_db(struct xfs_inode *ip, xfs_
 }
 
 /*
- * Flags for xfs_free_eofblocks
- */
-#define XFS_FREE_EOF_LOCK	(1<<0)
-#define XFS_FREE_EOF_NOLOCK	(1<<1)
-
-
-/*
  * helper function to extract extent size hint from inode
  */
 STATIC_INLINE xfs_extlen_t
fs/xfs/xfs_vnodeops.c +26 -8
Index: xfs/fs/xfs/xfs_vnodeops.c
===================================================================
@@ -709,6 +709,11 @@ xfs_fsync(
 }
 
 /*
+ * Flags for xfs_free_eofblocks
+ */
+#define XFS_FREE_EOF_TRYLOCK	(1<<0)
+
+/*
  * This is called by xfs_inactive to free any blocks beyond eof
  * when the link count isn't zero and by xfs_dm_punch_hole() when
  * punching a hole to EOF.
@@ -726,7 +731,6 @@ xfs_free_eofblocks(
 	xfs_filblks_t	map_len;
 	int		nimaps;
 	xfs_bmbt_irec_t	imap;
-	int		use_iolock = (flags & XFS_FREE_EOF_LOCK);
 
 	/*
 	 * Figure out if there are any blocks beyond the end
@@ -768,14 +772,19 @@ xfs_free_eofblocks(
 		 * cache and we can't
 		 * do that within a transaction.
 		 */
-		if (use_iolock)
+		if (flags & XFS_FREE_EOF_TRYLOCK) {
+			if (!xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
+				xfs_trans_cancel(tp, 0);
+				return 0;
+			}
+		} else {
 			xfs_ilock(ip, XFS_IOLOCK_EXCL);
+		}
 		error = xfs_itruncate_start(ip, XFS_ITRUNC_DEFINITE,
 				    ip->i_size);
 		if (error) {
 			xfs_trans_cancel(tp, 0);
-			if (use_iolock)
-				xfs_iunlock(ip, XFS_IOLOCK_EXCL);
+			xfs_iunlock(ip, XFS_IOLOCK_EXCL);
 			return error;
 		}
 
@@ -812,8 +821,7 @@ xfs_free_eofblocks(
 			error = xfs_trans_commit(tp,
 						XFS_TRANS_RELEASE_LOG_RES);
 		}
-		xfs_iunlock(ip, (use_iolock ? (XFS_IOLOCK_EXCL|XFS_ILOCK_EXCL)
-					    : XFS_ILOCK_EXCL));
+		xfs_iunlock(ip, XFS_IOLOCK_EXCL|XFS_ILOCK_EXCL);
 	}
 	return error;
 }
@@ -1113,7 +1121,17 @@ xfs_release(
 		     (ip->i_df.if_flags & XFS_IFEXTENTS))  &&
 		    (!(ip->i_d.di_flags &
 				(XFS_DIFLAG_PREALLOC | XFS_DIFLAG_APPEND)))) {
-			error = xfs_free_eofblocks(mp, ip, XFS_FREE_EOF_LOCK);
+
+			/*
+			 * If we can't get the iolock just skip truncating
+			 * the blocks past EOF because we could deadlock
+			 * with the mmap_sem otherwise.  We'll get another
+			 * chance to drop them once the last reference to
+			 * the inode is dropped, so we'll never leak blocks
+			 * permanently.
+			 */
+			error = xfs_free_eofblocks(mp, ip,
+						   XFS_FREE_EOF_TRYLOCK);
 			if (error)
 				return error;
 		}
@@ -1184,7 +1202,7 @@ xfs_inactive(
 		     (!(ip->i_d.di_flags &
 				(XFS_DIFLAG_PREALLOC | XFS_DIFLAG_APPEND)) ||
 		      (ip->i_delayed_blks != 0)))) {
-			error = xfs_free_eofblocks(mp, ip, XFS_FREE_EOF_LOCK);
+			error = xfs_free_eofblocks(mp, ip, 0);
 			if (error)
 				return VN_INACTIVE_CACHE;
 		}

_______________________________________________
xfs mailing list
[email protected]
http://oss.sgi.com/mailman/listinfo/xfs

#1 lore raw

ping?

Quoted 114 lines · Christoph Hellwig

When xfs_free_eofblocks is called from ->release the VM might already hold the mmap_sem, but in the write path we take the iolock before taking the mmap_sem in the generic write code.

Switch xfs_free_eofblocks to only trylock the iolock if called from ->release and skip trimming the prellocated blocks in that case. We'll still free them later on the final iput.

Index: xfs/fs/xfs/xfs_rw.h
===================================================================

fs/xfs/xfs_rw.h +0 -7
@@ -37,13 +37,6 @@ xfs_fsb_to_db(struct xfs_inode *ip, xfs_
 }
 
 /*
- * Flags for xfs_free_eofblocks
- */
-#define XFS_FREE_EOF_LOCK	(1<<0)
-#define XFS_FREE_EOF_NOLOCK	(1<<1)
-
-
-/*
  * helper function to extract extent size hint from inode
  */
 STATIC_INLINE xfs_extlen_t

Index: xfs/fs/xfs/xfs_vnodeops.c
===================================================================

fs/xfs/xfs_vnodeops.c +26 -8
@@ -709,6 +709,11 @@ xfs_fsync(
 }
 
 /*
+ * Flags for xfs_free_eofblocks
+ */
+#define XFS_FREE_EOF_TRYLOCK	(1<<0)
+
+/*
  * This is called by xfs_inactive to free any blocks beyond eof
  * when the link count isn't zero and by xfs_dm_punch_hole() when
  * punching a hole to EOF.
@@ -726,7 +731,6 @@ xfs_free_eofblocks(
 	xfs_filblks_t	map_len;
 	int		nimaps;
 	xfs_bmbt_irec_t	imap;
-	int		use_iolock = (flags & XFS_FREE_EOF_LOCK);
 
 	/*
 	 * Figure out if there are any blocks beyond the end
@@ -768,14 +772,19 @@ xfs_free_eofblocks(
 		 * cache and we can't
 		 * do that within a transaction.
 		 */
-		if (use_iolock)
+		if (flags & XFS_FREE_EOF_TRYLOCK) {
+			if (!xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
+				xfs_trans_cancel(tp, 0);
+				return 0;
+			}
+		} else {
 			xfs_ilock(ip, XFS_IOLOCK_EXCL);
+		}
 		error = xfs_itruncate_start(ip, XFS_ITRUNC_DEFINITE,
 				    ip->i_size);
 		if (error) {
 			xfs_trans_cancel(tp, 0);
-			if (use_iolock)
-				xfs_iunlock(ip, XFS_IOLOCK_EXCL);
+			xfs_iunlock(ip, XFS_IOLOCK_EXCL);
 			return error;
 		}
 
@@ -812,8 +821,7 @@ xfs_free_eofblocks(
 			error = xfs_trans_commit(tp,
 						XFS_TRANS_RELEASE_LOG_RES);
 		}
-		xfs_iunlock(ip, (use_iolock ? (XFS_IOLOCK_EXCL|XFS_ILOCK_EXCL)
-					    : XFS_ILOCK_EXCL));
+		xfs_iunlock(ip, XFS_IOLOCK_EXCL|XFS_ILOCK_EXCL);
 	}
 	return error;
 }
@@ -1113,7 +1121,17 @@ xfs_release(
 		     (ip->i_df.if_flags & XFS_IFEXTENTS))  &&
 		    (!(ip->i_d.di_flags &
 				(XFS_DIFLAG_PREALLOC | XFS_DIFLAG_APPEND)))) {
-			error = xfs_free_eofblocks(mp, ip, XFS_FREE_EOF_LOCK);
+
+			/*
+			 * If we can't get the iolock just skip truncating
+			 * the blocks past EOF because we could deadlock
+			 * with the mmap_sem otherwise.  We'll get another
+			 * chance to drop them once the last reference to
+			 * the inode is dropped, so we'll never leak blocks
+			 * permanently.
+			 */
+			error = xfs_free_eofblocks(mp, ip,
+						   XFS_FREE_EOF_TRYLOCK);
 			if (error)
 				return error;
 		}
@@ -1184,7 +1202,7 @@ xfs_inactive(
 		     (!(ip->i_d.di_flags &
 				(XFS_DIFLAG_PREALLOC | XFS_DIFLAG_APPEND)) ||
 		      (ip->i_delayed_blks != 0)))) {
-			error = xfs_free_eofblocks(mp, ip, XFS_FREE_EOF_LOCK);
+			error = xfs_free_eofblocks(mp, ip, 0);
 			if (error)
 				return VN_INACTIVE_CACHE;
 		}
Quoted 4 lines

_______________________________________________
xfs mailing list
[email protected]
http://oss.sgi.com/mailman/listinfo/xfs

---end quoted text---

_______________________________________________
xfs mailing list
[email protected]
http://oss.sgi.com/mailman/listinfo/xfs

Reviewed-by on v1 1/1
#2 lore raw

RE: [PATCH] xfs: fix mmap_sem vs iolock lock order inversion inxfs_free_eofblocks

In reply to Christoph Hellwig
Quoted 9 lines · Christoph Hellwig

When xfs_free_eofblocks is called from ->release the VM might already hold the mmap_sem, but in the write path we take the iolock before taking the mmap_sem in the generic write code.

Switch xfs_free_eofblocks to only trylock the iolock if called from ->release and skip trimming the prellocated blocks in that case. We'll still free them later on the final iput.

I have a minor suggestion below, but it looks correct to me. I tried to get a better idea of what the conditions were where mmap_sem would be held by VM when ->release gets called, but didn't get to the bottom of that. If it is easily characterized you could mention it in comments.

Quoted 23 lines

Index: xfs/fs/xfs/xfs_rw.h
===================================================================

fs/xfs/xfs_rw.h +0 -7
@@ -37,13 +37,6 @@ xfs_fsb_to_db(struct xfs_inode *ip, xfs_
 }
 
 /*
- * Flags for xfs_free_eofblocks
- */
-#define XFS_FREE_EOF_LOCK	(1<<0)
-#define XFS_FREE_EOF_NOLOCK	(1<<1)
-
-
-/*
  * helper function to extract extent size hint from inode
  */
 STATIC_INLINE xfs_extlen_t

Index: xfs/fs/xfs/xfs_vnodeops.c
===================================================================

fs/xfs/xfs_vnodeops.c xfs_fsync( · line 714

Quoted 6 lines
fs/xfs/xfs_vnodeops.c +3 -0
@@ -709,6 +709,11 @@ xfs_fsync(
 }
 
 /*
+ * Flags for xfs_free_eofblocks
+ */
+#define XFS_FREE_EOF_TRYLOCK	(1<<0)

This is really a Boolean in the place it's used, rather than a flags mask. Unless you have plans to add other flags, maybe just don't bother defining this, and pass a zero/non-zero for a renamed argument to xfs_free_eofblocks(). (I mention this again below, in that context.)

Quoted 13 lines
fs/xfs/xfs_vnodeops.c +1 -1
+/*
  * This is called by xfs_inactive to free any blocks beyond eof
  * when the link count isn't zero and by xfs_dm_punch_hole() when
  * punching a hole to EOF.
@@ -726,7 +731,6 @@ xfs_free_eofblocks(
 	xfs_filblks_t	map_len;
 	int		nimaps;
 	xfs_bmbt_irec_t	imap;
-	int		use_iolock = (flags & XFS_FREE_EOF_LOCK);
 
 	/*
 	 * Figure out if there are any blocks beyond the end

fs/xfs/xfs_vnodeops.c xfs_free_eofblocks( · line 775

Quoted 5 lines
fs/xfs/xfs_vnodeops.c +1 -1
@@ -768,14 +772,19 @@ xfs_free_eofblocks(
 		 * cache and we can't
 		 * do that within a transaction.
 		 */
-		if (use_iolock)
+		if (flags & XFS_FREE_EOF_TRYLOCK) {

If you rename "flag" to "skip_ok" (or maybe "try_lock":

            if (skip_ok && ! xfs_ilock_nowait(...)) {
                      xfs_trans_cancel();
                      return 0;
            } else
                      xfs_ilock(...);

             if (try_lock)
Quoted 56 lines
fs/xfs/xfs_vnodeops.c +20 -6
+			if (!xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
+				xfs_trans_cancel(tp, 0);
+				return 0;
+			}
+		} else {
 			xfs_ilock(ip, XFS_IOLOCK_EXCL);
+		}
 		error = xfs_itruncate_start(ip, XFS_ITRUNC_DEFINITE,
 				    ip->i_size);
 		if (error) {
 			xfs_trans_cancel(tp, 0);
-			if (use_iolock)
-				xfs_iunlock(ip, XFS_IOLOCK_EXCL);
+			xfs_iunlock(ip, XFS_IOLOCK_EXCL);
 			return error;
 		}
 
@@ -812,8 +821,7 @@ xfs_free_eofblocks(
 			error = xfs_trans_commit(tp,
 						XFS_TRANS_RELEASE_LOG_RES);
 		}
-		xfs_iunlock(ip, (use_iolock ? (XFS_IOLOCK_EXCL|XFS_ILOCK_EXCL)
-					    : XFS_ILOCK_EXCL));
+		xfs_iunlock(ip, XFS_IOLOCK_EXCL|XFS_ILOCK_EXCL);
 	}
 	return error;
 }
@@ -1113,7 +1121,17 @@ xfs_release(
 		     (ip->i_df.if_flags & XFS_IFEXTENTS))  &&
 		    (!(ip->i_d.di_flags &
 				(XFS_DIFLAG_PREALLOC | XFS_DIFLAG_APPEND)))) {
-			error = xfs_free_eofblocks(mp, ip, XFS_FREE_EOF_LOCK);
+
+			/*
+			 * If we can't get the iolock just skip truncating
+			 * the blocks past EOF because we could deadlock
+			 * with the mmap_sem otherwise.  We'll get another
+			 * chance to drop them once the last reference to
+			 * the inode is dropped, so we'll never leak blocks
+			 * permanently.
+			 */
+			error = xfs_free_eofblocks(mp, ip,
+						   XFS_FREE_EOF_TRYLOCK);
 			if (error)
 				return error;
 		}
@@ -1184,7 +1202,7 @@ xfs_inactive(
 		     (!(ip->i_d.di_flags &
 				(XFS_DIFLAG_PREALLOC | XFS_DIFLAG_APPEND)) ||
 		      (ip->i_delayed_blks != 0)))) {
-			error = xfs_free_eofblocks(mp, ip, XFS_FREE_EOF_LOCK);
+			error = xfs_free_eofblocks(mp, ip, 0);
 			if (error)
 				return VN_INACTIVE_CACHE;
 		}
Quoted 4 lines

_______________________________________________
xfs mailing list
[email protected]
http://oss.sgi.com/mailman/listinfo/xfs

_______________________________________________
xfs mailing list
[email protected]
http://oss.sgi.com/mailman/listinfo/xfs

#3 lore raw

Re: [PATCH] xfs: fix mmap_sem vs iolock lock order inversion inxfs_free_eofblocks

Quoted 5 lines · Alex Elder

I have a minor suggestion below, but it looks correct to me. I tried to get a better idea of what the conditions were where mmap_sem would be held by VM when ->release gets called, but didn't get to the bottom of that. If it is easily characterized you could mention it in comments.

It's the VMA merging code. But IMHO that's too much of an implementation detail to put into the comment here. Commit message might be fine.

_______________________________________________
xfs mailing list
[email protected]
http://oss.sgi.com/mailman/listinfo/xfs