Let me improve English a bit:
On Fri 03-09-21 14:27:48, yangerkun wrote:
> We get a BUG as follow:
We hit the following bug:
>
> [52117.465187] ------------[ cut here ]------------
> [52117.465686] kernel BUG at fs/ext4/extents.c:1756!
> ...
> [52117.478306] Call Trace:
> [52117.478565] ext4_ext_shift_extents+0x3ee/0x710
> [52117.479020] ext4_fallocate+0x139c/0x1b40
> [52117.479405] ? __do_sys_newfstat+0x6b/0x80
> [52117.479805] vfs_fallocate+0x151/0x4b0
> [52117.480177] ksys_fallocate+0x4a/0xa0
> [52117.480533] __x64_sys_fallocate+0x22/0x30
> [52117.480930] do_syscall_64+0x35/0x80
> [52117.481277] entry_SYSCALL_64_after_hwframe+0x44/0xae
> [52117.481769] RIP: 0033:0x7fa062f855ca
>
> static int ext4_ext_try_to_merge_right(struct inode *inode,
> struct ext4_ext_path *path,
> struct ext4_extent *ex)
> {
> struct ext4_extent_header *eh;
> unsigned int depth, len;
> int merge_done = 0, unwritten;
>
> depth = ext_depth(inode);
> BUG_ON(path[depth].p_hdr == NULL); <=== trigger here
> eh = path[depth].p_hdr;
>
> Normally, we protect extent tree with i_data_sem, and once we really
> need drop i_data_sem, we should reload the ext4_ext_path array after we
> recatch i_data_sem since extent tree may has changed, the 'again' in
> ext4_ext_remove_space give us a sample. But the other case
> ext4_ext_shift_path_extents seems forget to do this(ext4_access_path may
> drop i_data_sem and recatch it with not enough credits), and will lead
> the upper BUG when there is a parallel extents split which will grow the
> extent tree.
Normally, the extent tree is protected by i_data_sem and if we drop
i_data_sem in ext4_datasem_ensure_credits(), we need to reload
ext4_ext_path array after reacquiring i_data_sem since the extent tree may
have changed. The 'again' label in ext4_ext_remove_space() is an example of
this. But ext4_ext_shift_path_extents() forgets to reload ext4_ext_path and
thus can cause the above mentioned BUG when there is a parallel extents
split which will grow the extent tree.
>
> Fix it by introduce the again in ext4_ext_shift_extents.
>
> Signed-off-by: yangerkun <yangerkun@xxxxxxxxxx>
> ---
> fs/ext4/extents.c | 14 +++++++++++++-
> 1 file changed, 13 insertions(+), 1 deletion(-)
>
> diff --git a/fs/ext4/extents.c b/fs/ext4/extents.c
> index a6fb0350f062..0aa14f6ca914 100644
> --- a/fs/ext4/extents.c
> +++ b/fs/ext4/extents.c
> @@ -5009,8 +5009,11 @@ ext4_ext_shift_path_extents(struct ext4_ext_path *path, ext4_lblk_t shift,
> restart_credits = ext4_writepage_trans_blocks(inode);
> err = ext4_datasem_ensure_credits(handle, inode, credits,
> restart_credits, 0);
> - if (err)
> + if (err) {
> + if (err > 0)
> + err = -EAGAIN;
> goto out;
> + }
Hum, I'd note that the previous patch actually broke
ext4_ext_shift_path_extents() which could return 1 after patch 2/3 and
probably confuse code upwards in the stack and now you fix it up in this
patch. Can you perhaps fixup the previous patch by changing the condition
to:
if (err < 0)
and then change it here?
>
> err = ext4_ext_get_access(handle, inode, path + depth);
> if (err)
> @@ -5084,6 +5087,7 @@ ext4_ext_shift_extents(struct inode *inode, handle_t *handle,
> int ret = 0, depth;
> struct ext4_extent *extent;
> ext4_lblk_t stop, *iterator, ex_start, ex_end;
> + ext4_lblk_t tmp = EXT_MAX_BLOCKS;
Can you perhaps name this more descriptively than 'tmp'? Something like
restart_lblk or something like that?
> /* Let path point to the last extent */
> path = ext4_find_extent(inode, EXT_MAX_BLOCKS - 1, NULL,
> @@ -5137,11 +5141,15 @@ ext4_ext_shift_extents(struct inode *inode, handle_t *handle,
> * till we reach stop. In case of right shift, iterator points to stop
> * and it is decreased till we reach start.
> */
> +again:
> if (SHIFT == SHIFT_LEFT)
> iterator = &start;
> else
> iterator = &stop;
>
> + if (tmp != EXT_MAX_BLOCKS)
> + *iterator = tmp;
> +
> /*
> * Its safe to start updating extents. Start and stop are unsigned, so
> * in case of right shift if extent with 0 block is reached, iterator
> @@ -5170,6 +5178,7 @@ ext4_ext_shift_extents(struct inode *inode, handle_t *handle,
> }
> }
>
> + tmp = *iterator;
> if (SHIFT == SHIFT_LEFT) {
> extent = EXT_LAST_EXTENT(path[depth].p_hdr);
> *iterator = le32_to_cpu(extent->ee_block) +
> @@ -5188,6 +5197,9 @@ ext4_ext_shift_extents(struct inode *inode, handle_t *handle,
> }
> ret = ext4_ext_shift_path_extents(path, shift, inode,
> handle, SHIFT);
> + /* iterator can be NULL which means we should break */
> + if (ret == -EAGAIN)
> + goto again;
Hum, but while we dropped i_data_sem, the extent depth may have increased
so we may need larger 'path' now?
Otherwise the patch looks good.
Honza
--
Jan Kara <jack@xxxxxxxx>
SUSE Labs, CR
