Re: [PATCH 0/6] Overhaul memalloc_no*

From: Matthew Wilcox
Date: Thu Jun 25 2020 - 16:41:05 EST


On Thu, Jun 25, 2020 at 10:36:11PM +0200, Michal Hocko wrote:
> On Thu 25-06-20 11:48:32, Darrick J. Wong wrote:
> > On Thu, Jun 25, 2020 at 12:31:16PM +0100, Matthew Wilcox (Oracle) wrote:
> > > I want a memalloc_nowait like we have memalloc_noio and memalloc_nofs
> > > for an upcoming patch series, and Jens also wants it for non-blocking
> > > io_uring. It turns out we already have dm-bufio which could benefit
> > > from memalloc_nowait, so it may as well go into the tree now.
> > >
> > > The biggest problem is that we're basically out of PF_ flags, so we need
> > > to find somewhere else to store the PF_MEMALLOC_NOWAIT flag. It turns
> > > out the PF_ flags are really supposed to be used for flags which are
> > > accessed from other tasks, and the MEMALLOC flags are only going to
> > > be used by this task. So shuffling everything around frees up some PF
> > > flags and generally makes the world a better place.
> >
> > So, uh, how does this intersect with the patch "xfs: reintroduce
> > PF_FSTRANS for transaction reservation recursion protection" that
> > re-adds PF_TRANS because uh I guess we lost some subtlety or another at
> > some point?
>
> This is independent, really. It just relocates the NOFS flag. PF_TRANS
> is reintroduced for a different reason. When I have replaced the
> original PF_TRANS by PF_MEMALLOC_NOFS I didn't realized that xfs doesn't
> need only the NOFS semantic but also the transaction tracking so this
> cannot be a single bit only. So it has to be added back. But
> PF_MEMALLOC_NOFS needs to stay for the scoped NOFS semantic.

If XFS needs to track transactions, why doesn't it use
current->journal_info like btrfs/ceph/ext4/gfs2/nilfs2/reiserfs?