GitLab

__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 f4b554af in 2015 created the current weird logic (which
used to exist in a different form in commit 5accdf82

 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>