splice: don't read more than available pipe space
authorDarrick J. Wong <darrick.wong@oracle.com>
Fri, 30 Nov 2018 18:37:49 +0000 (10:37 -0800)
committerDarrick J. Wong <darrick.wong@oracle.com>
Tue, 4 Dec 2018 16:50:49 +0000 (08:50 -0800)
commit17614445576b6af24e9cf36607c6448164719c96
treeb05471db334d90a6c2ba14693a6831d3544bce16
parent6744557b53a2b710ebce3736a5c27a0119685fcc
splice: don't read more than available pipe space

In commit 4721a601099, we tried to fix a problem wherein directio reads
into a splice pipe will bounce EFAULT/EAGAIN all the way out to
userspace by simulating a zero-byte short read.  This happens because
some directio read implementations (xfs) will call
bio_iov_iter_get_pages to grab pipe buffer pages and issue asynchronous
reads, but as soon as we run out of pipe buffers that _get_pages call
returns EFAULT, which the splice code translates to EAGAIN and bounces
out to userspace.

In that commit, the iomap code catches the EFAULT and simulates a
zero-byte read, but that causes assertion errors on regular splice reads
because xfs doesn't allow short directio reads.

The brokenness is compounded by splice_direct_to_actor immediately
bailing on do_splice_to returning <= 0 without ever calling ->actor
(which empties out the pipe), so if userspace calls back we'll EFAULT
again on the full pipe, and nothing ever gets copied.

Therefore, teach splice_direct_to_actor to clamp its requests to the
amount of free space in the pipe and remove the simulated short read
from the iomap directio code.

Fixes: 4721a601099 ("iomap: dio data corruption and spurious errors when pipes fill")
Reported-by: Murphy Zhou <jencce.kernel@gmail.com>
Ranted-by: Amir Goldstein <amir73il@gmail.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
fs/splice.c