Hi Ted,
On 11/25/2011 09:08 AM, Theodore Ts'o wrote:
> We use an separate flag in buffer head to determine whether the bitmap
> has been valid. This is distinct from it being uptodate, due to the
> uninit_bg feature. More details about the rationale for this flag can
> be found in commit 2ccb5fb9f1. We set this bitmap_uptodate bit before
> issuing the read request, so if another CPU attempts to load the same
> block or inode bitmap, since ext4_read_{block,inode}_bitmap() checks
> the bitmap_uptodate flag without locking the buffer head, hilarity
> ensues.
>
> This result of this bug is that occasionally a block or inode gets
> allocated twice, which gets noticed when the second user of the block
> gets deleted, or when an directory suddenly becomes a regular file or
> a symlink. I'm *really* surprised this doesn't happen more often; but
> in actual practice the fact that we tend to search for a zero bit in
> the bitmap without taking a lock, and then taking the block group lock
> and double checking to see if we actually got the allocation tends to
> protect us.
Sorry, but I don't get your meaning here.
In bitmap_uptodate, we check both the flag of BH_uptodate and
BH_BITMAP_UPTODATE. And in your patch below, we just move the set of
bitmap_uptodate after bh_uptodate. So I don't think the above scenario
would ever happen. Could you please explain it in more detail?
Thanks
Tao
>
> This bug was introduced in commit 2ccb5fb9f1, which dates back to
> January 2009 and 2.6.29. So this bug has been around for a *long*
> time. (We've seen it for over a year, but rarely enough that it we
> could never find a repro case so we could study it in controlled
> circumstances.)
>
> Google-Bug-Id: 2828254
> Signed-off-by: "Theodore Ts'o" <tytso@xxxxxxx>
> Cc: stable@xxxxxxxxxx
> ---
> fs/ext4/balloc.c | 12 ++++++------
> fs/ext4/ialloc.c | 12 ++++++------
> 2 files changed, 12 insertions(+), 12 deletions(-)
>
> diff --git a/fs/ext4/balloc.c b/fs/ext4/balloc.c
> index 12ccacd..4501aab 100644
> --- a/fs/ext4/balloc.c
> +++ b/fs/ext4/balloc.c
> @@ -372,7 +372,7 @@ ext4_read_block_bitmap(struct super_block *sb, ext4_group_t block_group)
> ext4_unlock_group(sb, block_group);
> if (buffer_uptodate(bh)) {
> /*
> - * if not uninit if bh is uptodate,
> + * if not uninit && bh is uptodate,
> * bitmap is also uptodate
> */
> set_bitmap_uptodate(bh);
> @@ -380,13 +380,12 @@ ext4_read_block_bitmap(struct super_block *sb, ext4_group_t block_group)
> return bh;
> }
> /*
> - * submit the buffer_head for read. We can
> - * safely mark the bitmap as uptodate now.
> - * We do it here so the bitmap uptodate bit
> - * get set with buffer lock held.
> + * submit the buffer_head for read. It's important that we
> + * *not* mark the bitmap up to date until the read is
> + * completed, since we check bitmap_update() above without
> + * locking the buffer for speed reasons.
> */
> trace_ext4_read_block_bitmap_load(sb, block_group);
> - set_bitmap_uptodate(bh);
> if (bh_submit_read(bh) < 0) {
> put_bh(bh);
> ext4_error(sb, "Cannot read block bitmap - "
> @@ -394,6 +393,7 @@ ext4_read_block_bitmap(struct super_block *sb, ext4_group_t block_group)
> block_group, bitmap_blk);
> return NULL;
> }
> + set_bitmap_uptodate(bh);
> ext4_valid_block_bitmap(sb, desc, block_group, bh);
> /*
> * file system mounted not to panic on error,
> diff --git a/fs/ext4/ialloc.c b/fs/ext4/ialloc.c
> index 00beb4f..6fbae6d 100644
> --- a/fs/ext4/ialloc.c
> +++ b/fs/ext4/ialloc.c
> @@ -139,7 +139,7 @@ ext4_read_inode_bitmap(struct super_block *sb, ext4_group_t block_group)
>
> if (buffer_uptodate(bh)) {
> /*
> - * if not uninit if bh is uptodate,
> + * if not uninit && bh is uptodate,
> * bitmap is also uptodate
> */
> set_bitmap_uptodate(bh);
> @@ -147,13 +147,12 @@ ext4_read_inode_bitmap(struct super_block *sb, ext4_group_t block_group)
> return bh;
> }
> /*
> - * submit the buffer_head for read. We can
> - * safely mark the bitmap as uptodate now.
> - * We do it here so the bitmap uptodate bit
> - * get set with buffer lock held.
> + * submit the buffer_head for read. It's important that we
> + * *not* mark the bitmap up to date until the read is
> + * completed, since we check bitmap_update() above without
> + * locking the buffer for speed reasons.
> */
> trace_ext4_load_inode_bitmap(sb, block_group);
> - set_bitmap_uptodate(bh);
> if (bh_submit_read(bh) < 0) {
> put_bh(bh);
> ext4_error(sb, "Cannot read inode bitmap - "
> @@ -161,6 +160,7 @@ ext4_read_inode_bitmap(struct super_block *sb, ext4_group_t block_group)
> block_group, bitmap_blk);
> return NULL;
> }
> + set_bitmap_uptodate(bh);
> return bh;
> }
>
--
To unsubscribe from this list: send the line "unsubscribe linux-ext4" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
