Commit 5e91eabf authored by Jack Wang's avatar Jack Wang Committed by Jason Gunthorpe
Browse files

RDMA/rtrs-srv: Set minimal max_send_wr and max_recv_wr 

Currently rtrs when create_qp use a coarse numbers (bigger in general),
which leads to hardware create more resources which only waste memory with
no benefits.

For max_send_wr, we don't really need alway max_qp_wr size when creating
qp, reduce it to cq_size.

For max_recv_wr,  cq_size is enough.

With the patch when sess_queue_depth=128, per session (2 paths) memory
consumption reduced from 188 MB to 65MB

When always_invalidate is enabled, we need send more wr, so treat it
special.

Fixes: 9cb83748 ("RDMA/rtrs: server: main functionality")
Link: https://lore.kernel.org/r/20210614090337.29557-2-jinpu.wang@ionos.com


Signed-off-by: default avatarJack Wang <jinpu.wang@cloud.ionos.com>
Reviewed-by: default avatarMd Haris Iqbal <haris.iqbal@cloud.ionos.com>
Signed-off-by: default avatarGioh Kim <gi-oh.kim@ionos.com>
Reviewed-by: default avatarLeon Romanovsky <leonro@nvidia.com>
Signed-off-by: default avatarJason Gunthorpe <jgg@nvidia.com>
parent 915e4af5

drivers/infiniband/ulp/rtrs/rtrs-srv.c

+25 −13
Original line number Diff line number Diff line
@@ -1634,7 +1634,7 @@ static int create_con(struct rtrs_srv_sess *sess, 
	struct rtrs_sess *s = &sess->s;

	struct rtrs_srv_con *con;



	u32 cq_size, wr_queue_size;

	u32 cq_size, max_send_wr, max_recv_wr, wr_limit;

	int err, cq_vector;



	con = kzalloc(sizeof(*con), GFP_KERNEL);
@@ -1655,30 +1655,42 @@ static int create_con(struct rtrs_srv_sess *sess, 
		 * All receive and all send (each requiring invalidate)

		 * + 2 for drain and heartbeat

		 */

		wr_queue_size = SERVICE_CON_QUEUE_DEPTH * 3 + 2;

		cq_size = wr_queue_size;

		max_send_wr = SERVICE_CON_QUEUE_DEPTH * 2 + 2;

		max_recv_wr = SERVICE_CON_QUEUE_DEPTH + 2;

		cq_size = max_send_wr + max_recv_wr;

	} else {

		/*

		 * If we have all receive requests posted and

		 * all write requests posted and each read request

		 * requires an invalidate request + drain

		 * and qp gets into error state.

		 */

		cq_size = srv->queue_depth * 3 + 1;

		/*

		 * In theory we might have queue_depth * 32

		 * outstanding requests if an unsafe global key is used

		 * and we have queue_depth read requests each consisting

		 * of 32 different addresses. div 3 for mlx5.

		 */

		wr_queue_size = sess->s.dev->ib_dev->attrs.max_qp_wr / 3;

		wr_limit = sess->s.dev->ib_dev->attrs.max_qp_wr / 3;

		/* when always_invlaidate enalbed, we need linv+rinv+mr+imm */

		if (always_invalidate)

			max_send_wr =

				min_t(int, wr_limit,

				      srv->queue_depth * (1 + 4) + 1);

		else

			max_send_wr =

				min_t(int, wr_limit,

				      srv->queue_depth * (1 + 2) + 1);



		max_recv_wr = srv->queue_depth + 1;

		/*

		 * If we have all receive requests posted and

		 * all write requests posted and each read request

		 * requires an invalidate request + drain

		 * and qp gets into error state.

		 */

		cq_size = max_send_wr + max_recv_wr;

	}

	atomic_set(&con->sq_wr_avail, wr_queue_size);

	atomic_set(&con->sq_wr_avail, max_send_wr);

	cq_vector = rtrs_srv_get_next_cq_vector(sess);



	/* TODO: SOFTIRQ can be faster, but be careful with softirq context */

	err = rtrs_cq_qp_create(&sess->s, &con->c, 1, cq_vector, cq_size,

				 wr_queue_size, wr_queue_size,

				 max_send_wr, max_recv_wr,

				 IB_POLL_WORKQUEUE);

	if (err) {

		rtrs_err(s, "rtrs_cq_qp_create(), err: %d\n", err);