loop: handle short DIO reads (f9de14bc) · Commits · 戴 / test

We ran into an issue with loop and btrfs, where btrfs would complain about checksum errors. It turns out that is because we don't handle short reads at all, we just zero fill the remainder. Worse than that, we don't handle the filling properly, which results in loop trying to advance a single bio by much more than its size, since it doesn't take chaining into account. Handle short reads appropriately, by simply retrying at the new correct offset. End the remainder of the request with EIO, if we get a 0 read. Fixes: ("block: loop: support DIO & AIO") Reviewed-by:

Ming Lei <ming.lei@redhat.com> Signed-off-by:

Jens Axboe <axboe@kernel.dk>

drivers/block/loop.c

+27 −7

Original line number	Diff line number	Diff line
		@@ -451,16 +451,36 @@ static int lo_req_flush(struct loop_device lo, struct request rq)
		static void lo_complete_rq(struct request *rq)
		{
		struct loop_cmd *cmd = blk_mq_rq_to_pdu(rq);
		blk_status_t ret = BLK_STS_OK;

		if (unlikely(req_op(rq) == REQ_OP_READ && cmd->use_aio &&
		cmd->ret >= 0 && cmd->ret < blk_rq_bytes(rq))) {
		if (!cmd->use_aio \|\| cmd->ret < 0 \|\| cmd->ret == blk_rq_bytes(rq) \|\|
		req_op(rq) != REQ_OP_READ) {
		if (cmd->ret < 0)
		ret = BLK_STS_IOERR;
		goto end_io;
		}

		/*
		* Short READ - if we got some data, advance our request and
		* retry it. If we got no data, end the rest with EIO.
		*/
		if (cmd->ret) {
		blk_update_request(rq, BLK_STS_OK, cmd->ret);
		cmd->ret = 0;
		blk_mq_requeue_request(rq, true);
		} else {
		if (cmd->use_aio) {
		struct bio *bio = rq->bio;

		bio_advance(bio, cmd->ret);
		while (bio) {
		zero_fill_bio(bio);
		bio = bio->bi_next;
		}
		}
		ret = BLK_STS_IOERR;
		end_io:
		blk_mq_end_request(rq, ret);
		}

		blk_mq_end_request(rq, cmd->ret < 0 ? BLK_STS_IOERR : BLK_STS_OK);
		}

		static void lo_rw_aio_do_completion(struct loop_cmd *cmd)

Admin message