xfs: Fix deadlock between AGI and AGF when target_ip exists in xfs_rename()
authorkaixuxia <xiakaixu1987@gmail.com>
Tue, 12 Nov 2019 16:34:23 +0000 (08:34 -0800)
committerDarrick J. Wong <darrick.wong@oracle.com>
Wed, 13 Nov 2019 19:13:45 +0000 (11:13 -0800)
commit93597ae8dac0149b5c00b787cba6bf7ba213e666
tree81d81c6c0c66b66c379427ebf686adfb0ae4cb12
parent048a35d2f0b4cfeb24cbb7fe59e78124d8e7dc73
xfs: Fix deadlock between AGI and AGF when target_ip exists in xfs_rename()

When target_ip exists in xfs_rename(), the xfs_dir_replace() call may
need to hold the AGF lock to allocate more blocks, and then invoking
the xfs_droplink() call to hold AGI lock to drop target_ip onto the
unlinked list, so we get the lock order AGF->AGI. This would break the
ordering constraint on AGI and AGF locking - inode allocation locks
the AGI, then can allocate a new extent for new inodes, locking the
AGF after the AGI.

In this patch we check whether the replace operation need more
blocks firstly. If so, acquire the agi lock firstly to preserve
locking order(AGI/AGF). Actually, the locking order problem only
occurs when we are locking the AGI/AGF of the same AG. For multiple
AGs the AGI lock will be released after the transaction committed.

Signed-off-by: kaixuxia <kaixuxia@tencent.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
[darrick: reword the comment]
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
fs/xfs/libxfs/xfs_dir2.h
fs/xfs/libxfs/xfs_dir2_sf.c
fs/xfs/xfs_inode.c