summaryrefslogtreecommitdiff
path: root/fs/hfsplus
diff options
context:
space:
mode:
authorDarrick J. Wong <darrick.wong@oracle.com>2020-11-10 16:49:29 -0800
committerDarrick J. Wong <darrick.wong@oracle.com>2020-11-10 16:49:29 -0800
commit22843291efc986ce7722610073fcf85a39b4cb13 (patch)
tree151e95b483a9a4bb8cfec7610c43ba5cda607fef /fs/hfsplus
parent3650b228f83adda7e5ee532e2b90429c03f7b9ec (diff)
vfs: remove lockdep bogosity in __sb_start_write
__sb_start_write has some weird looking lockdep code that claims to exist to handle nested freeze locking requests from xfs. The code as written seems broken -- if we think we hold a read lock on any of the higher freeze levels (e.g. we hold SB_FREEZE_WRITE and are trying to lock SB_FREEZE_PAGEFAULT), it converts a blocking lock attempt into a trylock. However, it's not correct to downgrade a blocking lock attempt to a trylock unless the downgrading code or the callers are prepared to deal with that situation. Neither __sb_start_write nor its callers handle this at all. For example: sb_start_pagefault ignores the return value completely, with the result that if xfs_filemap_fault loses a race with a different thread trying to fsfreeze, it will proceed without pagefault freeze protection (thereby breaking locking rules) and then unlocks the pagefault freeze lock that it doesn't own on its way out (thereby corrupting the lock state), which leads to a system hang shortly afterwards. Normally, this won't happen because our ownership of a read lock on a higher freeze protection level blocks fsfreeze from grabbing a write lock on that higher level. *However*, if lockdep is offline, lock_is_held_type unconditionally returns 1, which means that percpu_rwsem_is_held returns 1, which means that __sb_start_write unconditionally converts blocking freeze lock attempts into trylocks, even when we *don't* hold anything that would block a fsfreeze. Apparently this all held together until 5.10-rc1, when bugs in lockdep caused lockdep to shut itself off early in an fstests run, and once fstests gets to the "race writes with freezer" tests, kaboom. This might explain the long trail of vanishingly infrequent livelocks in fstests after lockdep goes offline that I've never been able to diagnose. We could fix it by spinning on the trylock if wait==true, but AFAICT the locking works fine if lockdep is not built at all (and I didn't see any complaints running fstests overnight), so remove this snippet entirely. NOTE: Commit f4b554af9931 in 2015 created the current weird logic (which used to exist in a different form in commit 5accdf82ba25c from 2012) in __sb_start_write. XFS solved this whole problem in the late 2.6 era by creating a variant of transactions (XFS_TRANS_NO_WRITECOUNT) that don't grab intwrite freeze protection, thus making lockdep's solution unnecessary. The commit claims that Dave Chinner explained that the trylock hack + comment could be removed, but nobody ever did. Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com> Reviewed-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Jan Kara <jack@suse.cz>
Diffstat (limited to 'fs/hfsplus')
0 files changed, 0 insertions, 0 deletions