[Cluster-devel] always fall back to buffered I/O after invalidation failures, was: Re: [PATCH 2/6] iomap: IOMAP_DIO_RWF_NO_STALE_PAGECACHE return if page invalidation fails

[Matthew Wilcox]

On Thu, Jul 09, 2020 at 09:09:26AM -0700, Darrick J. Wong wrote:
> On Thu, Jul 09, 2020 at 12:25:27PM +1000, Dave Chinner wrote:
> > iomap: Only invalidate page cache pages on direct IO writes
> > 
> > From: Dave Chinner <dchinner at redhat.com>
> > 
> > The historic requirement for XFS to invalidate cached pages on
> > direct IO reads has been lost in the twisty pages of history - it was
> > inherited from Irix, which implemented page cache invalidation on
> > read as a method of working around problems synchronising page
> > cache state with uncached IO.
> 
> Urk.
> 
> > XFS has carried this ever since. In the initial linux ports it was
> > necessary to get mmap and DIO to play "ok" together and not
> > immediately corrupt data. This was the state of play until the linux
> > kernel had infrastructure to track unwritten extents and synchronise
> > page faults with allocations and unwritten extent conversions
> > (->page_mkwrite infrastructure). IOws, the page cache invalidation
> > on DIO read was necessary to prevent trivial data corruptions. This
> > didn't solve all the problems, though.
> > 
> > There were peformance problems if we didn't invalidate the entire
> > page cache over the file on read - we couldn't easily determine if
> > the cached pages were over the range of the IO, and invalidation
> > required taking a serialising lock (i_mutex) on the inode. This
> > serialising lock was an issue for XFS, as it was the only exclusive
> > lock in the direct Io read path.
> > 
> > Hence if there were any cached pages, we'd just invalidate the
> > entire file in one go so that subsequent IOs didn't need to take the
> > serialising lock. This was a problem that prevented ranged
> > invalidation from being particularly useful for avoiding the
> > remaining coherency issues. This was solved with the conversion of
> > i_mutex to i_rwsem and the conversion of the XFS inode IO lock to
> > use i_rwsem. Hence we could now just do ranged invalidation and the
> > performance problem went away.
> > 
> > However, page cache invalidation was still needed to serialise
> > sub-page/sub-block zeroing via direct IO against buffered IO because
> > bufferhead state attached to the cached page could get out of whack
> > when direct IOs were issued.  We've removed bufferheads from the
> > XFS code, and we don't carry any extent state on the cached pages
> > anymore, and so this problem has gone away, too.
> > 
> > IOWs, it would appear that we don't have any good reason to be
> > invalidating the page cache on DIO reads anymore. Hence remove the
> > invalidation on read because it is unnecessary overhead,
> > not needed to maintain coherency between mmap/buffered access and
> > direct IO anymore, and prevents anyone from using direct IO reads
> > from intentionally invalidating the page cache of a file.
> > 
> > Signed-off-by: Dave Chinner <dchinner at redhat.com>
> > ---
> >  fs/iomap/direct-io.c | 33 +++++++++++++++++----------------
> >  1 file changed, 17 insertions(+), 16 deletions(-)
> > 
> > diff --git a/fs/iomap/direct-io.c b/fs/iomap/direct-io.c
> > index ec7b78e6feca..ef0059eb34b5 100644
> > --- a/fs/iomap/direct-io.c
> > +++ b/fs/iomap/direct-io.c
> > @@ -475,23 +475,24 @@ iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
> >  	if (ret)
> >  		goto out_free_dio;
> >  
> > -	/*
> > -	 * Try to invalidate cache pages for the range we're direct
> > -	 * writing.  If this invalidation fails, tough, the write will
> > -	 * still work, but racing two incompatible write paths is a
> > -	 * pretty crazy thing to do, so we don't support it 100%.
> 
> I always wondered about the repeated use of 'write' in this comment
> despite the lack of any sort of WRITE check logic.  Seems fine to me,
> let's throw it on the fstests pile and see what happens.
> 
> Reviewed-by: Darrick J. Wong <darrick.wong at oracle.com>

Reviewed-by: Matthew Wilcox (Oracle) <willy at infradead.org>

> --D
> 
> > -	 */
> > -	ret = invalidate_inode_pages2_range(mapping,
> > -			pos >> PAGE_SHIFT, end >> PAGE_SHIFT);
> > -	if (ret)
> > -		dio_warn_stale_pagecache(iocb->ki_filp);
> > -	ret = 0;
> > +	if (iov_iter_rw(iter) == WRITE) {
> > +		/*
> > +		 * Try to invalidate cache pages for the range we're direct
> > +		 * writing.  If this invalidation fails, tough, the write will
> > +		 * still work, but racing two incompatible write paths is a
> > +		 * pretty crazy thing to do, so we don't support it 100%.
> > +		 */
> > +		ret = invalidate_inode_pages2_range(mapping,
> > +				pos >> PAGE_SHIFT, end >> PAGE_SHIFT);
> > +		if (ret)
> > +			dio_warn_stale_pagecache(iocb->ki_filp);
> > +		ret = 0;
> >  
> > -	if (iov_iter_rw(iter) == WRITE && !wait_for_completion &&
> > -	    !inode->i_sb->s_dio_done_wq) {
> > -		ret = sb_init_dio_done_wq(inode->i_sb);
> > -		if (ret < 0)
> > -			goto out_free_dio;
> > +		if (!wait_for_completion &&
> > +		    !inode->i_sb->s_dio_done_wq) {
> > +			ret = sb_init_dio_done_wq(inode->i_sb);
> > +			if (ret < 0)
> > +				goto out_free_dio;
> >  	}
> >  
> >  	inode_dio_begin(inode);