From: Jens Axboe <axboe@xxxxxxxxx> Sent: Monday, January 16, 2023 1:06 PM
>
> If we're doing a large IO request which needs to be split into multiple
> bios for issue, then we can run into the same situation as the below
> marked commit fixes - parts will complete just fine, one or more parts
> will fail to allocate a request. This will result in a partially
> completed read or write request, where the caller gets EAGAIN even though
> parts of the IO completed just fine.
>
> Do the same for large bios as we do for splits - fail a NOWAIT request
> with EAGAIN. This isn't technically fixing an issue in the below marked
> patch, but for stable purposes, we should have either none of them or
> both.
>
> This depends on: 613b14884b85 ("block: handle bio_split_to_limits() NULL return")
>
> Cc: stable@xxxxxxxxxxxxxxx # 5.15+
> Fixes: 9cea62b2cbab ("block: don't allow splitting of a REQ_NOWAIT bio")
> Link: https://github.com/axboe/liburing/issues/766
> Reported-and-tested-by: Michael Kelley <mikelley@xxxxxxxxxxxxx>
> Signed-off-by: Jens Axboe <axboe@xxxxxxxxx>
>
> ---
>
> Since v1: catch this at submit time instead, since we can have various
> valid cases where the number of single page segments will not take a
> bio segment (page merging, huge pages).
>
> diff --git a/block/fops.c b/block/fops.c
> index 50d245e8c913..d2e6be4e3d1c 100644
> --- a/block/fops.c
> +++ b/block/fops.c
> @@ -221,6 +221,24 @@ static ssize_t __blkdev_direct_IO(struct kiocb *iocb, struct
> iov_iter *iter,
> bio_endio(bio);
> break;
> }
> + if (iocb->ki_flags & IOCB_NOWAIT) {
> + /*
> + * This is nonblocking IO, and we need to allocate
> + * another bio if we have data left to map. As we
> + * cannot guarantee that one of the sub bios will not
> + * fail getting issued FOR NOWAIT and as error results
> + * are coalesced across all of them, be safe and ask for
> + * a retry of this from blocking context.
> + */
> + if (unlikely(iov_iter_count(iter))) {
> + bio_release_pages(bio, false);
> + bio_clear_flag(bio, BIO_REFFED);
> + bio_put(bio);
> + blk_finish_plug(&plug);
> + return -EAGAIN;
> + }
> + bio->bi_opf |= REQ_NOWAIT;
> + }
>
> if (is_read) {
> if (dio->flags & DIO_SHOULD_DIRTY)
> @@ -228,9 +246,6 @@ static ssize_t __blkdev_direct_IO(struct kiocb *iocb, struct iov_iter *iter,
> } else {
> task_io_account_write(bio->bi_iter.bi_size);
> }
> - if (iocb->ki_flags & IOCB_NOWAIT)
> - bio->bi_opf |= REQ_NOWAIT;
> -
> dio->size += bio->bi_iter.bi_size;
> pos += bio->bi_iter.bi_size;
>
I've wrapped up my testing on this patch. All testing was via io_uring -- I did
not test other paths. Testing was against a combination of this patch and the
previous patch set for a similar problem. [1]
I tested with a simple test program to issue single I/Os, and verified the
expected paths were taken through the block layer and io_uring code for
various size I/Os, including over 1 Mbyte. No EAGAIN errors were seen.
This testing was with a 6.1 kernel.
Also tested the original app that surfaced the problem. It's a larger scale
workload using io_uring, and is where the problem was originally
encountered. That workload runs on a purpose-built 5.15 kernel, so I
backported both patches to 5.15 for this testing. All looks good.
No EAGAIN errors were seen.
Michael
[1] https://lore.kernel.org/linux-block/20230104160938.62636-1-axboe@xxxxxxxxx/
