[PATCH] fix bmap-vs-truncate race

[Mikulas Patocka <mpatocka@redhat.com>]

Hi

I'm submitting this patch for 2.6.30 merge window.


FAT filesystem used down_read(&mapping->host->i_alloc_sem) to prevent 
a race between bmap and truncate. However, such race is present in all the
other filesystems --- it is generally assumed that blocks queried with
get_block won't disappear while get_block is in progress.

The race can be only triggered by root, non-privileged users can't use
bmap, so it is not a security issue (unless there is some program run
by root that bmaps users' files).

This patch fixes the race in a generic way, in all the filesystems. If some
filesystem employs its own locking and doesn't want to take i_alloc_sem
(I don't know about any, where taking i_alloc_sem could be problem),
let it use its own function and not generic_block_bmap.

Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>

---
 fs/buffer.c    |    8 ++++++++
 fs/fat/inode.c |    2 --
 2 files changed, 8 insertions(+), 2 deletions(-)

Index: linux-2.6.29-devel/fs/buffer.c
===================================================================
--- linux-2.6.29-devel.orig/fs/buffer.c	2009-03-29 04:48:38.000000000 +0200
+++ linux-2.6.29-devel/fs/buffer.c	2009-03-29 05:26:53.000000000 +0200
@@ -2963,7 +2963,15 @@ sector_t generic_block_bmap(struct addre
 	tmp.b_state = 0;
 	tmp.b_blocknr = 0;
 	tmp.b_size = 1 << inode->i_blkbits;
+
+	/*
+	 * Protect the inode from being truncated while get_block is
+	 * in progress.
+	 */
+	down_read(&mapping->host->i_alloc_sem);
 	get_block(inode, block, &tmp, 0);
+	up_read(&mapping->host->i_alloc_sem);
+
 	return tmp.b_blocknr;
 }
 
Index: linux-2.6.29-devel/fs/fat/inode.c
===================================================================
--- linux-2.6.29-devel.orig/fs/fat/inode.c	2009-03-29 04:46:35.000000000 +0200
+++ linux-2.6.29-devel/fs/fat/inode.c	2009-03-29 05:26:53.000000000 +0200
@@ -202,9 +202,7 @@ static sector_t _fat_bmap(struct address
 	sector_t blocknr;
 
 	/* fat_get_cluster() assumes the requested blocknr isn't truncated. */
-	down_read(&mapping->host->i_alloc_sem);
 	blocknr = generic_block_bmap(mapping, block, fat_get_block);
-	up_read(&mapping->host->i_alloc_sem);
 
 	return blocknr;
 }