[Cluster-devel] [PATCH v7 5/5] gfs2: Fix iomap write page reclaim deadlock

[Darrick J. Wong]

On Tue, Apr 30, 2019 at 05:39:28PM +0200, Andreas Gruenbacher wrote:
> On Tue, 30 Apr 2019 at 17:33, Darrick J. Wong <darrick.wong at oracle.com> wrote:
> > On Tue, Apr 30, 2019 at 12:09:34AM +0200, Andreas Gruenbacher wrote:
> > > Since commit 64bc06bb32ee ("gfs2: iomap buffered write support"), gfs2 is doing
> > > buffered writes by starting a transaction in iomap_begin, writing a range of
> > > pages, and ending that transaction in iomap_end.  This approach suffers from
> > > two problems:
> > >
> > >   (1) Any allocations necessary for the write are done in iomap_begin, so when
> > >   the data aren't journaled, there is no need for keeping the transaction open
> > >   until iomap_end.
> > >
> > >   (2) Transactions keep the gfs2 log flush lock held.  When
> > >   iomap_file_buffered_write calls balance_dirty_pages, this can end up calling
> > >   gfs2_write_inode, which will try to flush the log.  This requires taking the
> > >   log flush lock which is already held, resulting in a deadlock.
> >
> > /me wonders how holding the log flush lock doesn't seriously limit
> > performance, but gfs2 isn't my fight so I'll set that aside and assume
> > that a patch S-o-B'd by both maintainers is ok. :)
> 
> This only affects inline and journaled data, not standard writes, so
> it's not quite as bad as it looks.

Ah, ok.

> > How should we merge this patch #5?  It doesn't touch fs/iomap.c itself,
> > so do you want me to pull it into the iomap branch along with the
> > previous four patches?  That would be fine with me (and easier than a
> > multi-tree merge mess)...
> 
> I'd prefer to get this merged via the gfs2 tree once the iomap fixes
> have been pulled.

Ok, I'll take the first four patches through the iomap branch and cc you
on the pull request.

--D

> 
> Thanks,
> Andreas