[Cluster-devel] gfs2 iomap dealock, IOMAP_F_UNBALANCED

[Christoph Hellwig <hch@lst.de>]

On Thu, Mar 21, 2019 at 02:13:04PM +0100, Andreas Gruenbacher wrote:
> Hi Christoph,
> 
> we need your help fixing a gfs2 deadlock involving iomap.  What's going
> on is the following:
> 
> * During iomap_file_buffered_write, gfs2_iomap_begin grabs the log flush
>   lock and keeps it until gfs2_iomap_end.  It currently always does that
>   even though there is no point other than for journaled data writes.
> 
> * iomap_file_buffered_write then calls balance_dirty_pages_ratelimited.
>   If that ends up calling gfs2_write_inode, gfs2 will try to grab the
>   log flush lock again and deadlock.

What is the exactly call chain?  balance_dirty_pages_ratelimited these
days doesn't start I/O, but just wakes up the flusher threads.  Or
do we have a issue where it is blocking on those threads?

Also why do you need to flush the log for background writeback in
->write_inode?

balance_dirty_pages_ratelimited is per definition not a data integrity
writeback, so there shouldn't be a good reason to flush the log
(which I assume the log flush log is for).  If we look gfs2_write_inode,
this seems to be the code:

	bool flush_all = (wbc->sync_mode == WB_SYNC_ALL || gfs2_is_jdata(ip));

        if (flush_all)
		gfs2_log_flush(GFS2_SB(inode), ip->i_gl,
			       GFS2_LOG_HEAD_FLUSH_NORMAL |
			       GFS2_LFC_WRITE_INODE);

But what is the requirement to do this in writeback context?  Can't
we move it out into another context instead?