aboutsummaryrefslogtreecommitdiffstats
path: root/fs/btrfs/inode.c
diff options
context:
space:
mode:
authorJosef Bacik <josef@toxicpanda.com>2022-08-19 11:53:39 -0400
committerDavid Sterba <dsterba@suse.com>2022-08-22 18:08:07 +0200
commit79d3d1d12e6f22c904195f9356069859e2595f00 (patch)
treef5e765449759728e886bdb80e1203b7518365048 /fs/btrfs/inode.c
parentbtrfs: don't merge pages into bio if their page offset is not contiguous (diff)
downloadlinux-dev-79d3d1d12e6f22c904195f9356069859e2595f00.tar.xz
linux-dev-79d3d1d12e6f22c904195f9356069859e2595f00.zip
btrfs: don't allow large NOWAIT direct reads
Dylan and Jens reported a problem where they had an io_uring test that was returning short reads, and bisected it to ee5b46a353af ("btrfs: increase direct io read size limit to 256 sectors"). The root cause is their test was doing larger reads via io_uring with NOWAIT and async. This was triggering a page fault during the direct read, however the first page was able to work just fine and thus we submitted a 4k read for a larger iocb. Btrfs allows for partial IO's in this case specifically because we don't allow page faults, and thus we'll attempt to do any io that we can, submit what we could, come back and fault in the rest of the range and try to do the remaining IO. However for !is_sync_kiocb() we'll call ->ki_complete() as soon as the partial dio is done, which is incorrect. In the sync case we can exit the iomap code, submit more io's, and return with the amount of IO we were able to complete successfully. We were always doing short reads in this case, but for NOWAIT we were getting saved by the fact that we were limiting direct reads to sectorsize, and if we were larger than that we would return EAGAIN. Fix the regression by simply returning EAGAIN in the NOWAIT case with larger reads, that way io_uring can retry and get the larger IO and have the fault logic handle everything properly. This still leaves the AIO short read case, but that existed before this change. The way to properly fix this would be to handle partial iocb completions, but that's a lot of work, for now deal with the regression in the most straightforward way possible. Reported-by: Dylan Yudaken <dylany@fb.com> Fixes: ee5b46a353af ("btrfs: increase direct io read size limit to 256 sectors") Reviewed-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: David Sterba <dsterba@suse.com>
Diffstat (limited to 'fs/btrfs/inode.c')
-rw-r--r--fs/btrfs/inode.c14
1 files changed, 14 insertions, 0 deletions
diff --git a/fs/btrfs/inode.c b/fs/btrfs/inode.c
index ecc5fa3343fc..bea0adf28735 100644
--- a/fs/btrfs/inode.c
+++ b/fs/btrfs/inode.c
@@ -7694,6 +7694,20 @@ static int btrfs_dio_iomap_begin(struct inode *inode, loff_t start,
bool unlock_extents = false;
/*
+ * We could potentially fault if we have a buffer > PAGE_SIZE, and if
+ * we're NOWAIT we may submit a bio for a partial range and return
+ * EIOCBQUEUED, which would result in an errant short read.
+ *
+ * The best way to handle this would be to allow for partial completions
+ * of iocb's, so we could submit the partial bio, return and fault in
+ * the rest of the pages, and then submit the io for the rest of the
+ * range. However we don't have that currently, so simply return
+ * -EAGAIN at this point so that the normal path is used.
+ */
+ if (!write && (flags & IOMAP_NOWAIT) && length > PAGE_SIZE)
+ return -EAGAIN;
+
+ /*
* Cap the size of reads to that usually seen in buffered I/O as we need
* to allocate a contiguous array for the checksums.
*/