2023-10-07 20:36:43

by Matthew Wilcox

[permalink] [raw]
Subject: [PATCH v2 0/5] Remove the XFS mrlock

XFS has an mrlock wrapper around the rwsem which adds only the
functionality of knowing whether the rwsem is currently held in read
or write mode. Both regular rwsems and rt-rwsems know this, they just
don't expose it as an API. By adding that, we can remove the XFS mrlock
as well as improving the debug assertions for the mmap_lock when lockdep
is disabled.

v2: Add rwsem_assert_held() and rwsem_assert_held_write() instead of
augmenting the existing rwsem_is_locked() with rwsem_is_write_locked().
There's also an __rwsem_assert_held() and __rwsem_assert_held_write()
for the benefit of XFS when it's in a context where lockdep doesn't
know what's going on. It's still an improvement, so I hope those who
are looking for perfection can accept a mere improvement.

We can do more to replace uses of rwsem_is_locked(), and I have a few of
those in my tree, but let's focus on these three use cases for now and
we can trickle in other improvements through other maintainers after 6.7.

I'm sympathetic to "this will warn twice and dump much the same
information if you have lockdep enabled". Perhaps somebody has a
suggestion for not doing that?

Matthew Wilcox (Oracle) (5):
locking: Add rwsem_assert_held() and rwsem_assert_held_write()
mm: Use rwsem assertion macros for mmap_lock
xfs: Replace xfs_isilocked with xfs_assert_locked
xfs: Remove mrlock wrapper
fs: Add inode_assert_locked() and inode_assert_locked_excl()

fs/attr.c | 2 +-
fs/namei.c | 6 +--
fs/xfs/libxfs/xfs_attr.c | 2 +-
fs/xfs/libxfs/xfs_attr_remote.c | 2 +-
fs/xfs/libxfs/xfs_bmap.c | 19 ++++----
fs/xfs/libxfs/xfs_defer.c | 2 +-
fs/xfs/libxfs/xfs_inode_fork.c | 2 +-
fs/xfs/libxfs/xfs_rtbitmap.c | 2 +-
fs/xfs/libxfs/xfs_trans_inode.c | 6 +--
fs/xfs/mrlock.h | 78 ------------------------------
fs/xfs/scrub/readdir.c | 4 +-
fs/xfs/xfs_attr_list.c | 2 +-
fs/xfs/xfs_bmap_util.c | 10 ++--
fs/xfs/xfs_dir2_readdir.c | 2 +-
fs/xfs/xfs_dquot.c | 4 +-
fs/xfs/xfs_file.c | 4 +-
fs/xfs/xfs_inode.c | 86 ++++++++++++---------------------
fs/xfs/xfs_inode.h | 4 +-
fs/xfs/xfs_inode_item.c | 4 +-
fs/xfs/xfs_iops.c | 7 ++-
fs/xfs/xfs_linux.h | 2 +-
fs/xfs/xfs_qm.c | 10 ++--
fs/xfs/xfs_reflink.c | 2 +-
fs/xfs/xfs_rtalloc.c | 4 +-
fs/xfs/xfs_super.c | 4 +-
fs/xfs/xfs_symlink.c | 2 +-
fs/xfs/xfs_trans_dquot.c | 2 +-
include/linux/fs.h | 10 ++++
include/linux/mmap_lock.h | 10 ++--
include/linux/rwbase_rt.h | 9 +++-
include/linux/rwsem.h | 42 ++++++++++++++--
31 files changed, 142 insertions(+), 203 deletions(-)
delete mode 100644 fs/xfs/mrlock.h

--
2.40.1


2023-10-08 20:32:12

by Mateusz Guzik

[permalink] [raw]
Subject: Re: [PATCH v2 0/5] Remove the XFS mrlock

On 10/7/23, Matthew Wilcox (Oracle) <[email protected]> wrote:
> I'm sympathetic to "this will warn twice and dump much the same
> information if you have lockdep enabled". Perhaps somebody has a
> suggestion for not doing that?
>

Well the obvious idea is that lockdep could provide a macro indicating
what's up.

Then you would:
static inline void rwsem_assert_held(const struct rw_semaphore *sem)
{
if (lockdep_works)
lockdep_assert_held(sem);
else
__rwsem_assert_held(sem);
}

Am I missing something? If this is not feasible to achieve, then the
proposed routines need a comment justifying the state.

--
Mateusz Guzik <mjguzik gmail.com>