Re: [Cluster-devel] [PATCH v11 16/25] fs: Convert mpage_readpages to mpage_readahead

[Andreas Gruenbacher <agruenba@redhat.com>]

On Wed, Jun 17, 2020 at 4:22 AM Matthew Wilcox <willy@infradead.org> wrote:
> On Wed, Jun 17, 2020 at 02:57:14AM +0200, Andreas Grünbacher wrote:
> > Am Mi., 17. Juni 2020 um 02:33 Uhr schrieb Matthew Wilcox <willy@infradead.org>:
> > >
> > > On Wed, Jun 17, 2020 at 12:36:13AM +0200, Andreas Gruenbacher wrote:
> > > > Am Mi., 15. Apr. 2020 um 23:39 Uhr schrieb Matthew Wilcox <willy@infradead.org>:
> > > > > From: "Matthew Wilcox (Oracle)" <willy@infradead.org>
> > > > >
> > > > > Implement the new readahead aop and convert all callers (block_dev,
> > > > > exfat, ext2, fat, gfs2, hpfs, isofs, jfs, nilfs2, ocfs2, omfs, qnx6,
> > > > > reiserfs & udf).  The callers are all trivial except for GFS2 & OCFS2.
> > > >
> > > > This patch leads to an ABBA deadlock in xfstest generic/095 on gfs2.
> > > >
> > > > Our lock hierarchy is such that the inode cluster lock ("inode glock")
> > > > for an inode needs to be taken before any page locks in that inode's
> > > > address space.
> > >
> > > How does that work for ...
> > >
> > > writepage:              yes, unlocks (see below)
> > > readpage:               yes, unlocks
> > > invalidatepage:         yes
> > > releasepage:            yes
> > > freepage:               yes
> > > isolate_page:           yes
> > > migratepage:            yes (both)
> > > putback_page:           yes
> > > launder_page:           yes
> > > is_partially_uptodate:  yes
> > > error_remove_page:      yes
> > >
> > > Is there a reason that you don't take the glock in the higher level
> > > ops which are called before readhead gets called?  I'm looking at XFS,
> > > and it takes the xfs_ilock SHARED in xfs_file_buffered_aio_read()
> > > (called from xfs_file_read_iter).
> >
> > Right, the approach from the following thread might fix this:
> >
> > https://lore.kernel.org/linux-fsdevel/20191122235324.17245-1-agruenba@redhat.com/T/#t
>
> In general, I think this is a sound approach.
>
> Specifically, I think FAULT_FLAG_CACHED can go away.  map_pages()
> will bring in the pages which are in the page cache, so when we get to
> gfs2_fault(), we know there's a reason to acquire the glock.

We'd still be grabbing a glock while holding a dependent page lock.
Another process could be holding the glock and could try to grab the
same page lock (i.e., a concurrent writer), leading to the same kind
of deadlock.

Andreas