On Wed, May 17, 2023 at 2:34 PM Dave Chinner <david@xxxxxxxxxxxxx> wrote: > > On Wed, May 17, 2023 at 10:07:55AM +0300, Amir Goldstein wrote: > > On Wed, May 17, 2023 at 3:08 AM Dave Chinner <david@xxxxxxxxxxxxx> wrote: > > > > > > Hi folks, > > > > > > > Hi Dave, > > > > > The following patches are fixes for recently discovered problems. > > > I'd like to consider them all for a 6.4-rc merge, though really only > > > patch 2 is a necessary recent regression fix. > > > > > > The first patch addresses a long standing buffer UAF during shutdown > > > situations, where shutdown log item completions can race with CIL > > > abort and iclog log item completions. > > > > > > > Can you tell roughly how far back this UAF bug goes > > or is it long standing enough to be treated as "forever"? > > Longer than I cared to trace the history back. > > > > The second patch addresses a small performance regression from the > > > 6.3 allocator rewrite which is also a potential AGF-AGF deadlock > > > vector as it allows out-of-order locking. > > > > > > The third patch is a change needed by patch 4, which I split out for > > > clarity. By itself it does nothing. > > > > > > The fourth patch is a fix for a AGI->AGF->inode cluster buffer lock > > > ordering deadlock. This was introduced when we started pinning inode > > > cluster buffers in xfs_trans_log_inode() in 5.8 to fix long-standing > > > inode reclaim blocking issues, but I've only seen it in the wild > > > once on a system making heavy use of OVL and hence O_TMPFILE based > > > operations. > > > > Thank you for providing Fixes and this summary with backporing hints :) > > I don't think you're going to be able to backport the inode cluster > buffer lock fix. Nothing prior to 5.19 has the necessary > infrastructure or the iunlink log item code this latest fix builds > on top of. That was done to fix a whole class of relatively easy to > hit O_TMPFILE related AGI-AGF-inode cluster buffer deadlocks. This > fix avoids an entirely different class of inode cluster buffer > deadlocks using the infrastructure introduced in the 5.19 deadlock > fixes. > Even better. Less work for us ;) Thanks, Amir.
