test/uring_cmd_ublk.c: cover device buffer split

Force to split ublk-loop's device buffer into two parts:

1) the front part is fixed 4096 bytes

2) the remained bytes

Add two member IO requests for submitting IOs via the two buffers.

Then we can verify if buffer split works as expected.

Signed-off-by: Ming Lei <tom.leiming@gmail.com>

Author: ming1

test/uring_cmd_ublk.c

@@ -72,9 +72,10 @@ struct ublk_io {
 #define UBLKSRV_NEED_FETCH_RQ		(1UL << 0)
 #define UBLKSRV_NEED_COMMIT_RQ_COMP	(1UL << 1)
 #define UBLKSRV_IO_FREE			(1UL << 2)
-	unsigned int flags;
+	unsigned short flags;
+	unsigned short refs;		/* used by target code only */
 
-	unsigned int result;
+	int result;
 };
 
 struct ublk_tgt_ops {
@@ -1062,14 +1063,47 @@ static inline void ublk_uring_prep_rw_zc(struct ublk_queue *q,
 	__u64 off = iod->start_sector << 9;
 	struct io_uring_sqe *lead;
 	struct io_uring_sqe *mem;
+	struct ublk_io *io = &q->ios[tag];
+	const unsigned front_len = 4096;
 
+	io->refs = 1;
+	io->result = 0;
 	q->io_inflight++;
+
+	/* test buffer split */
+	if (len > front_len)
+		len = front_len;
+
 	ublk_get_sqe_pair(&q->ring, &lead, &mem);
 
 	io_uring_prep_grp_lead(lead, dev_fd, tag, q_id);
 	io_uring_prep_rw_group(op, mem, fd, 0, len, off);
 	io_uring_sqe_set_flags(mem, IOSQE_FIXED_FILE | IOSQE_GROUP_KBUF);
 	mem->user_data = build_user_data(tag, ublk_op, 0, 1);
+
+	len = (iod->nr_sectors << 9) - len;
+	if (len > 0) {
+		struct io_uring_sqe *mem2 = io_uring_get_sqe(&q->ring);
+
+		/* don't split buffer in case of running out of sqe */
+		if (!mem2) {
+			len = iod->nr_sectors << 9;
+			io_uring_prep_rw_group(op, mem, fd, 0, len, off);
+			return;
+		}
+
+		/*
+		 * The 1st member consumers buffer size of `front_size`,
+		 * and the 2nd member consumes the remained bytes
+		 */
+		mem->flags |= IOSQE_GROUP_LINK;
+		io_uring_prep_rw_group(op, mem2, fd, front_len, len,
+				off + front_len);
+		io_uring_sqe_set_flags(mem2, IOSQE_FIXED_FILE | IOSQE_GROUP_KBUF);
+		mem2->user_data = build_user_data(tag, ublk_op, 0, 1);
+		q->io_inflight += 1;
+		io->refs += 1;
+	}
 }
 
 static inline void ublk_uring_prep_flush(struct ublk_queue *q,
@@ -1078,7 +1112,10 @@ static inline void ublk_uring_prep_flush(struct ublk_queue *q,
 {
 	unsigned ublk_op = ublksrv_get_op(iod);
 	struct io_uring_sqe *sqe;
+	struct ublk_io *io = &q->ios[tag];
 
+	io->refs = 1;
+	io->result = 0;
 	sqe = io_uring_get_sqe(&q->ring);
 	io_uring_prep_sync_file_range(sqe, fd,
 			iod->nr_sectors << 9,
@@ -1148,10 +1185,21 @@ static void ublk_loop_io_done(struct ublk_queue *q, int tag,
 		const struct io_uring_cqe *cqe)
 {
 	int cqe_tag = user_data_to_tag(cqe->user_data);
+	struct ublk_io *io = &q->ios[tag];
 
 	assert(tag == cqe_tag);
-	ublk_complete_io(q, tag, cqe->res);
 	q->io_inflight--;
+
+	if (cqe->res >= 0) {
+		if (io->result >= 0)
+			io->result += cqe->res;
+	} else {
+		if (io->result >= 0)
+			io->result = cqe->res;
+	}
+
+	if (--io->refs == 0)
+		ublk_complete_io(q, tag, io->result);
 }
 
 static void ublk_loop_tgt_deinit(struct ublk_dev *dev)