| Message ID | 20241231023423.656128-3-leo.lilong@huawei.com (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | xfs: fix two issues regarding mount failures | expand |
On Tue, Dec 31, 2024 at 10:34:23AM +0800, Long Li wrote: > When mounting an image containing a log with sb modifications that require > log replay, the mount process hang all the time and stack as follows: > > [root@localhost ~]# cat /proc/557/stack > [<0>] xfs_buftarg_wait+0x31/0x70 > [<0>] xfs_buftarg_drain+0x54/0x350 > [<0>] xfs_mountfs+0x66e/0xe80 > [<0>] xfs_fs_fill_super+0x7f1/0xec0 > [<0>] get_tree_bdev_flags+0x186/0x280 > [<0>] get_tree_bdev+0x18/0x30 > [<0>] xfs_fs_get_tree+0x1d/0x30 > [<0>] vfs_get_tree+0x2d/0x110 > [<0>] path_mount+0xb59/0xfc0 > [<0>] do_mount+0x92/0xc0 > [<0>] __x64_sys_mount+0xc2/0x160 > [<0>] x64_sys_call+0x2de4/0x45c0 > [<0>] do_syscall_64+0xa7/0x240 > [<0>] entry_SYSCALL_64_after_hwframe+0x76/0x7e > > During log recovery, while updating the in-memory superblock from the > primary SB buffer, if an error is encountered, such as superblock > corruption occurs or some other reasons, we will proceed to out_release > and release the xfs_buf. However, this is insufficient because the > xfs_buf's log item has already been initialized and the xfs_buf is held > by the buffer log item as follows, the xfs_buf will not be released, > causing the mount thread to hang. > > xlog_recover_do_primary_sb_buffer > xlog_recover_do_reg_buffer > xlog_recover_validate_buf_type > xfs_buf_item_init(bp, mp) > > The solution is straightforward: we simply need to allow it to be > handled by the normal buffer write process. The filesystem will be > shutdown before the submission of buffer_list in xlog_do_recovery_pass(), What shuts it down? If xlog_recover_do_primary_sb_buffer trips over something like "mp->m_sb.sb_rgcount < orig_rgcount" then we haven't shut anything down yet. Am I missing something? <confused> --D > ensuring the correct release of the xfs_buf. > > Fixes: 6a18765b54e2 ("xfs: update the file system geometry after recoverying superblock buffers") > Signed-off-by: Long Li <leo.lilong@huawei.com> > --- > fs/xfs/xfs_buf_item_recover.c | 3 ++- > 1 file changed, 2 insertions(+), 1 deletion(-) > > diff --git a/fs/xfs/xfs_buf_item_recover.c b/fs/xfs/xfs_buf_item_recover.c > index 3d0c6402cb36..ec2a42ef66ff 100644 > --- a/fs/xfs/xfs_buf_item_recover.c > +++ b/fs/xfs/xfs_buf_item_recover.c > @@ -1079,7 +1079,7 @@ xlog_recover_buf_commit_pass2( > error = xlog_recover_do_primary_sb_buffer(mp, item, bp, buf_f, > current_lsn); > if (error) > - goto out_release; > + goto out_writebuf; > > /* Update the rt superblock if we have one. */ > if (xfs_has_rtsb(mp) && mp->m_rtsb_bp) { > @@ -1096,6 +1096,7 @@ xlog_recover_buf_commit_pass2( > xlog_recover_do_reg_buffer(mp, item, bp, buf_f, current_lsn); > } > > +out_writebuf: > /* > * Perform delayed write on the buffer. Asynchronous writes will be > * slower when taking into account all the buffers to be flushed. > -- > 2.39.2 > >
On Mon, Jan 06, 2025 at 11:55:41AM -0800, Darrick J. Wong wrote: > On Tue, Dec 31, 2024 at 10:34:23AM +0800, Long Li wrote: > > When mounting an image containing a log with sb modifications that require > > log replay, the mount process hang all the time and stack as follows: > > > > [root@localhost ~]# cat /proc/557/stack > > [<0>] xfs_buftarg_wait+0x31/0x70 > > [<0>] xfs_buftarg_drain+0x54/0x350 > > [<0>] xfs_mountfs+0x66e/0xe80 > > [<0>] xfs_fs_fill_super+0x7f1/0xec0 > > [<0>] get_tree_bdev_flags+0x186/0x280 > > [<0>] get_tree_bdev+0x18/0x30 > > [<0>] xfs_fs_get_tree+0x1d/0x30 > > [<0>] vfs_get_tree+0x2d/0x110 > > [<0>] path_mount+0xb59/0xfc0 > > [<0>] do_mount+0x92/0xc0 > > [<0>] __x64_sys_mount+0xc2/0x160 > > [<0>] x64_sys_call+0x2de4/0x45c0 > > [<0>] do_syscall_64+0xa7/0x240 > > [<0>] entry_SYSCALL_64_after_hwframe+0x76/0x7e > > > > During log recovery, while updating the in-memory superblock from the > > primary SB buffer, if an error is encountered, such as superblock > > corruption occurs or some other reasons, we will proceed to out_release > > and release the xfs_buf. However, this is insufficient because the > > xfs_buf's log item has already been initialized and the xfs_buf is held > > by the buffer log item as follows, the xfs_buf will not be released, > > causing the mount thread to hang. > > > > xlog_recover_do_primary_sb_buffer > > xlog_recover_do_reg_buffer > > xlog_recover_validate_buf_type > > xfs_buf_item_init(bp, mp) > > > > The solution is straightforward: we simply need to allow it to be > > handled by the normal buffer write process. The filesystem will be > > shutdown before the submission of buffer_list in xlog_do_recovery_pass(), > > What shuts it down? If xlog_recover_do_primary_sb_buffer trips over > something like "mp->m_sb.sb_rgcount < orig_rgcount" then we haven't shut > anything down yet. Am I missing something? <confused> > > --D > Hi Darrick, Sorry for being unclear. I was referring to the shutdown in xlog_do_recovery_pass(). Here's the specific flow after the fix: xlog_do_recovery_pass error = xlog_recover_process xlog_recover_process_data xlog_recover_process_ophdr xlog_recovery_process_trans ... xlog_recover_buf_commit_pass2 error = xlog_recover_do_primary_sb_buffer //Encounter error and return if (error) goto out_writebuf ... out_writebuf: xfs_buf_delwri_queue(bp, buffer_list) //add bp to buffer_list return error ... if (!list_empty(&buffer_list)) if (error) xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR); //log shutdown first xfs_buf_delwri_submit(&buffer_list); __xfs_buf_submit if (bp->b_mount->m_log && xlog_is_shutdown(bp->b_mount->m_log)) xfs_buf_ioend_fail(bp) //release bp correctly It might be clearer to put this process into a commit message. Thanks, Long Li
On Tue, Jan 07, 2025 at 09:39:10PM +0800, Long Li wrote: > On Mon, Jan 06, 2025 at 11:55:41AM -0800, Darrick J. Wong wrote: > > On Tue, Dec 31, 2024 at 10:34:23AM +0800, Long Li wrote: > > > When mounting an image containing a log with sb modifications that require > > > log replay, the mount process hang all the time and stack as follows: > > > > > > [root@localhost ~]# cat /proc/557/stack > > > [<0>] xfs_buftarg_wait+0x31/0x70 > > > [<0>] xfs_buftarg_drain+0x54/0x350 > > > [<0>] xfs_mountfs+0x66e/0xe80 > > > [<0>] xfs_fs_fill_super+0x7f1/0xec0 > > > [<0>] get_tree_bdev_flags+0x186/0x280 > > > [<0>] get_tree_bdev+0x18/0x30 > > > [<0>] xfs_fs_get_tree+0x1d/0x30 > > > [<0>] vfs_get_tree+0x2d/0x110 > > > [<0>] path_mount+0xb59/0xfc0 > > > [<0>] do_mount+0x92/0xc0 > > > [<0>] __x64_sys_mount+0xc2/0x160 > > > [<0>] x64_sys_call+0x2de4/0x45c0 > > > [<0>] do_syscall_64+0xa7/0x240 > > > [<0>] entry_SYSCALL_64_after_hwframe+0x76/0x7e > > > > > > During log recovery, while updating the in-memory superblock from the > > > primary SB buffer, if an error is encountered, such as superblock > > > corruption occurs or some other reasons, we will proceed to out_release > > > and release the xfs_buf. However, this is insufficient because the > > > xfs_buf's log item has already been initialized and the xfs_buf is held > > > by the buffer log item as follows, the xfs_buf will not be released, > > > causing the mount thread to hang. > > > > > > xlog_recover_do_primary_sb_buffer > > > xlog_recover_do_reg_buffer > > > xlog_recover_validate_buf_type > > > xfs_buf_item_init(bp, mp) > > > > > > The solution is straightforward: we simply need to allow it to be > > > handled by the normal buffer write process. The filesystem will be > > > shutdown before the submission of buffer_list in xlog_do_recovery_pass(), > > > > What shuts it down? If xlog_recover_do_primary_sb_buffer trips over > > something like "mp->m_sb.sb_rgcount < orig_rgcount" then we haven't shut > > anything down yet. Am I missing something? <confused> > > > > --D > > > > Hi Darrick, > > Sorry for being unclear. I was referring to the shutdown in xlog_do_recovery_pass(). > Here's the specific flow after the fix: > > xlog_do_recovery_pass > error = xlog_recover_process > xlog_recover_process_data > xlog_recover_process_ophdr > xlog_recovery_process_trans > ... > xlog_recover_buf_commit_pass2 > error = xlog_recover_do_primary_sb_buffer > //Encounter error and return > if (error) > goto out_writebuf > ... > out_writebuf: > xfs_buf_delwri_queue(bp, buffer_list) //add bp to buffer_list > return error > ... > if (!list_empty(&buffer_list)) > if (error) > xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR); //log shutdown first > xfs_buf_delwri_submit(&buffer_list); > __xfs_buf_submit > if (bp->b_mount->m_log && xlog_is_shutdown(bp->b_mount->m_log)) > xfs_buf_ioend_fail(bp) //release bp correctly > > It might be clearer to put this process into a commit message. Yes please, put that in a code comment. That was too subtle for me to figure out. :/ --D > Thanks, > Long Li >
diff --git a/fs/xfs/xfs_buf_item_recover.c b/fs/xfs/xfs_buf_item_recover.c index 3d0c6402cb36..ec2a42ef66ff 100644 --- a/fs/xfs/xfs_buf_item_recover.c +++ b/fs/xfs/xfs_buf_item_recover.c @@ -1079,7 +1079,7 @@ xlog_recover_buf_commit_pass2( error = xlog_recover_do_primary_sb_buffer(mp, item, bp, buf_f, current_lsn); if (error) - goto out_release; + goto out_writebuf; /* Update the rt superblock if we have one. */ if (xfs_has_rtsb(mp) && mp->m_rtsb_bp) { @@ -1096,6 +1096,7 @@ xlog_recover_buf_commit_pass2( xlog_recover_do_reg_buffer(mp, item, bp, buf_f, current_lsn); } +out_writebuf: /* * Perform delayed write on the buffer. Asynchronous writes will be * slower when taking into account all the buffers to be flushed.
When mounting an image containing a log with sb modifications that require log replay, the mount process hang all the time and stack as follows: [root@localhost ~]# cat /proc/557/stack [<0>] xfs_buftarg_wait+0x31/0x70 [<0>] xfs_buftarg_drain+0x54/0x350 [<0>] xfs_mountfs+0x66e/0xe80 [<0>] xfs_fs_fill_super+0x7f1/0xec0 [<0>] get_tree_bdev_flags+0x186/0x280 [<0>] get_tree_bdev+0x18/0x30 [<0>] xfs_fs_get_tree+0x1d/0x30 [<0>] vfs_get_tree+0x2d/0x110 [<0>] path_mount+0xb59/0xfc0 [<0>] do_mount+0x92/0xc0 [<0>] __x64_sys_mount+0xc2/0x160 [<0>] x64_sys_call+0x2de4/0x45c0 [<0>] do_syscall_64+0xa7/0x240 [<0>] entry_SYSCALL_64_after_hwframe+0x76/0x7e During log recovery, while updating the in-memory superblock from the primary SB buffer, if an error is encountered, such as superblock corruption occurs or some other reasons, we will proceed to out_release and release the xfs_buf. However, this is insufficient because the xfs_buf's log item has already been initialized and the xfs_buf is held by the buffer log item as follows, the xfs_buf will not be released, causing the mount thread to hang. xlog_recover_do_primary_sb_buffer xlog_recover_do_reg_buffer xlog_recover_validate_buf_type xfs_buf_item_init(bp, mp) The solution is straightforward: we simply need to allow it to be handled by the normal buffer write process. The filesystem will be shutdown before the submission of buffer_list in xlog_do_recovery_pass(), ensuring the correct release of the xfs_buf. Fixes: 6a18765b54e2 ("xfs: update the file system geometry after recoverying superblock buffers") Signed-off-by: Long Li <leo.lilong@huawei.com> --- fs/xfs/xfs_buf_item_recover.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)