cxgb4/iw_cxgb4: use firmware ord/ird resource limits

Advertise a larger max read queue depth for qps, and gather the resource limits
from fw and use them to avoid exhaustinq all the resources.

Design:

cxgb4:

Obtain the max_ordird_qp and max_ird_adapter device params from FW
at init time and pass them up to the ULDs when they attach.  If these
parameters are not available, due to older firmware, then hard-code
the values based on the known values for older firmware.
iw_cxgb4:

Fix the c4iw_query_device() to report these correct values based on
adapter parameters.  ibv_query_device() will always return:

max_qp_rd_atom = max_qp_init_rd_atom = min(module_max, max_ordird_qp)
max_res_rd_atom = max_ird_adapter

Bump up the per qp max module option to 32, allowing it to be increased
by the user up to the device max of max_ordird_qp.  32 seems to be
sufficient to maximize throughput for streaming read benchmarks.

Fail connection setup if the negotiated IRD exhausts the available
adapter ird resources.  So the driver will track the amount of ird
resource in use and not send an RI_WR/INIT to FW that would reduce the
available ird resources below zero.

Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Hariprasad Shenai <hariprasad@chelsio.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

5 files changed, 117 insertions, 34 deletions

diff --git a/drivers/infiniband/hw/cxgb4/cm.c b/drivers/infiniband/hw/cxgb4/cm.c
index d62a0f9..df5bd3d 100644
--- a/drivers/infiniband/hw/cxgb4/cm.c
+++ b/drivers/infiniband/hw/cxgb4/cm.c
@@ -79,9 +79,10 @@ static int dack_mode = 1;
 module_param(dack_mode, int, 0644);
 MODULE_PARM_DESC(dack_mode, "Delayed ack mode (default=1)");
 
-int c4iw_max_read_depth = 8;
+uint c4iw_max_read_depth = 32;
 module_param(c4iw_max_read_depth, int, 0644);
-MODULE_PARM_DESC(c4iw_max_read_depth, "Per-connection max ORD/IRD (default=8)");
+MODULE_PARM_DESC(c4iw_max_read_depth,
+		 "Per-connection max ORD/IRD (default=32)");
 
 static int enable_tcp_timestamps;
 module_param(enable_tcp_timestamps, int, 0644);
@@ -813,6 +814,8 @@ static void send_mpa_req(struct c4iw_ep *ep, struct sk_buff *skb,
 	if (mpa_rev_to_use == 2) {
 		mpa->private_data_size = htons(ntohs(mpa->private_data_size) +
 					       sizeof (struct mpa_v2_conn_params));
+		PDBG("%s initiator ird %u ord %u\n", __func__, ep->ird,
+		     ep->ord);
 		mpa_v2_params.ird = htons((u16)ep->ird);
 		mpa_v2_params.ord = htons((u16)ep->ord);
 
@@ -1182,8 +1185,8 @@ static int connect_request_upcall(struct c4iw_ep *ep)
 			sizeof(struct mpa_v2_conn_params);
 	} else {
 		/* this means MPA_v1 is used. Send max supported */
-		event.ord = c4iw_max_read_depth;
-		event.ird = c4iw_max_read_depth;
+		event.ord = cur_max_read_depth(ep->com.dev);
+		event.ird = cur_max_read_depth(ep->com.dev);
 		event.private_data_len = ep->plen;
 		event.private_data = ep->mpa_pkt + sizeof(struct mpa_message);
 	}
@@ -1247,6 +1250,8 @@ static int update_rx_credits(struct c4iw_ep *ep, u32 credits)
 	return credits;
 }
 
+#define RELAXED_IRD_NEGOTIATION 1
+
 static int process_mpa_reply(struct c4iw_ep *ep, struct sk_buff *skb)
 {
 	struct mpa_message *mpa;
@@ -1358,17 +1363,33 @@ static int process_mpa_reply(struct c4iw_ep *ep, struct sk_buff *skb)
 				MPA_V2_IRD_ORD_MASK;
 			resp_ord = ntohs(mpa_v2_params->ord) &
 				MPA_V2_IRD_ORD_MASK;
+			PDBG("%s responder ird %u ord %u ep ird %u ord %u\n",
+			     __func__, resp_ird, resp_ord, ep->ird, ep->ord);
 
 			/*
 			 * This is a double-check. Ideally, below checks are
 			 * not required since ird/ord stuff has been taken
 			 * care of in c4iw_accept_cr
 			 */
-			if ((ep->ird < resp_ord) || (ep->ord > resp_ird)) {
+			if (ep->ird < resp_ord) {
+				if (RELAXED_IRD_NEGOTIATION && resp_ord <=
+				    ep->com.dev->rdev.lldi.max_ordird_qp)
+					ep->ird = resp_ord;
+				else
+					insuff_ird = 1;
+			} else if (ep->ird > resp_ord) {
+				ep->ird = resp_ord;
+			}
+			if (ep->ord > resp_ird) {
+				if (RELAXED_IRD_NEGOTIATION)
+					ep->ord = resp_ird;
+				else
+					insuff_ird = 1;
+			}
+			if (insuff_ird) {
 				err = -ENOMEM;
 				ep->ird = resp_ord;
 				ep->ord = resp_ird;
-				insuff_ird = 1;
 			}
 
 			if (ntohs(mpa_v2_params->ird) &
@@ -1571,6 +1592,8 @@ static void process_mpa_request(struct c4iw_ep *ep, struct sk_buff *skb)
 				MPA_V2_IRD_ORD_MASK;
 			ep->ord = ntohs(mpa_v2_params->ord) &
 				MPA_V2_IRD_ORD_MASK;
+			PDBG("%s initiator ird %u ord %u\n", __func__, ep->ird,
+			     ep->ord);
 			if (ntohs(mpa_v2_params->ird) & MPA_V2_PEER2PEER_MODEL)
 				if (peer2peer) {
 					if (ntohs(mpa_v2_params->ord) &
@@ -2724,8 +2747,8 @@ int c4iw_accept_cr(struct iw_cm_id *cm_id, struct iw_cm_conn_param *conn_param)
 	BUG_ON(!qp);
 
 	set_bit(ULP_ACCEPT, &ep->com.history);
-	if ((conn_param->ord > c4iw_max_read_depth) ||
-	    (conn_param->ird > c4iw_max_read_depth)) {
+	if ((conn_param->ord > cur_max_read_depth(ep->com.dev)) ||
+	    (conn_param->ird > cur_max_read_depth(ep->com.dev))) {
 		abort_connection(ep, NULL, GFP_KERNEL);
 		err = -EINVAL;
 		goto err;
@@ -2733,31 +2756,41 @@ int c4iw_accept_cr(struct iw_cm_id *cm_id, struct iw_cm_conn_param *conn_param)
 
 	if (ep->mpa_attr.version == 2 && ep->mpa_attr.enhanced_rdma_conn) {
 		if (conn_param->ord > ep->ird) {
-			ep->ird = conn_param->ird;
-			ep->ord = conn_param->ord;
-			send_mpa_reject(ep, conn_param->private_data,
-					conn_param->private_data_len);
-			abort_connection(ep, NULL, GFP_KERNEL);
-			err = -ENOMEM;
-			goto err;
+			if (RELAXED_IRD_NEGOTIATION) {
+				ep->ord = ep->ird;
+			} else {
+				ep->ird = conn_param->ird;
+				ep->ord = conn_param->ord;
+				send_mpa_reject(ep, conn_param->private_data,
+						conn_param->private_data_len);
+				abort_connection(ep, NULL, GFP_KERNEL);
+				err = -ENOMEM;
+				goto err;
+			}
 		}
-		if (conn_param->ird > ep->ord) {
-			if (!ep->ord)
-				conn_param->ird = 1;
-			else {
+		if (conn_param->ird < ep->ord) {
+			if (RELAXED_IRD_NEGOTIATION &&
+			    ep->ord <= h->rdev.lldi.max_ordird_qp) {
+				conn_param->ird = ep->ord;
+			} else {
 				abort_connection(ep, NULL, GFP_KERNEL);
 				err = -ENOMEM;
 				goto err;
 			}
 		}
-
 	}
 	ep->ird = conn_param->ird;
 	ep->ord = conn_param->ord;
 
-	if (ep->mpa_attr.version != 2)
+	if (ep->mpa_attr.version == 1) {
 		if (peer2peer && ep->ird == 0)
 			ep->ird = 1;
+	} else {
+		if (peer2peer &&
+		    (ep->mpa_attr.p2p_type != FW_RI_INIT_P2PTYPE_DISABLED) &&
+		    (p2p_type == FW_RI_INIT_P2PTYPE_READ_REQ) && ep->ord == 0)
+			ep->ird = 1;
+	}
 
 	PDBG("%s %d ird %d ord %d\n", __func__, __LINE__, ep->ird, ep->ord);
 
@@ -2796,6 +2829,7 @@ int c4iw_accept_cr(struct iw_cm_id *cm_id, struct iw_cm_conn_param *conn_param)
 	return 0;
 err1:
 	ep->com.cm_id = NULL;
+	abort_connection(ep, NULL, GFP_KERNEL);
 	cm_id->rem_ref(cm_id);
 err:
 	mutex_unlock(&ep->com.mutex);
@@ -2879,8 +2913,8 @@ int c4iw_connect(struct iw_cm_id *cm_id, struct iw_cm_conn_param *conn_param)
 	int iptype;
 	int iwpm_err = 0;
 
-	if ((conn_param->ord > c4iw_max_read_depth) ||
-	    (conn_param->ird > c4iw_max_read_depth)) {
+	if ((conn_param->ord > cur_max_read_depth(dev)) ||
+	    (conn_param->ird > cur_max_read_depth(dev))) {
 		err = -EINVAL;
 		goto out;
 	}
diff --git a/drivers/infiniband/hw/cxgb4/device.c b/drivers/infiniband/hw/cxgb4/device.c
index 88291ef..e76358e 100644
--- a/drivers/infiniband/hw/cxgb4/device.c
+++ b/drivers/infiniband/hw/cxgb4/device.c
@@ -348,6 +348,7 @@ static int stats_show(struct seq_file *seq, void *v)
 		   dev->rdev.stats.act_ofld_conn_fails);
 	seq_printf(seq, "PAS_OFLD_CONN_FAILS: %10llu\n",
 		   dev->rdev.stats.pas_ofld_conn_fails);
+	seq_printf(seq, "AVAILABLE IRD: %10u\n", dev->avail_ird);
 	return 0;
 }
 
@@ -839,6 +840,7 @@ static struct c4iw_dev *c4iw_alloc(const struct cxgb4_lld_info *infop)
 	mutex_init(&devp->rdev.stats.lock);
 	mutex_init(&devp->db_mutex);
 	INIT_LIST_HEAD(&devp->db_fc_list);
+	devp->avail_ird = devp->rdev.lldi.max_ird_adapter;
 
 	if (c4iw_debugfs_root) {
 		devp->debugfs_root = debugfs_create_dir(
diff --git a/drivers/infiniband/hw/cxgb4/iw_cxgb4.h b/drivers/infiniband/hw/cxgb4/iw_cxgb4.h
index 9b9754c..75541cb 100644
--- a/drivers/infiniband/hw/cxgb4/iw_cxgb4.h
+++ b/drivers/infiniband/hw/cxgb4/iw_cxgb4.h
@@ -249,6 +249,7 @@ struct c4iw_dev {
 	struct idr atid_idr;
 	struct idr stid_idr;
 	struct list_head db_fc_list;
+	u32 avail_ird;
 };
 
 static inline struct c4iw_dev *to_c4iw_dev(struct ib_device *ibdev)
@@ -330,6 +331,13 @@ static inline void remove_handle_nolock(struct c4iw_dev *rhp,
 	_remove_handle(rhp, idr, id, 0);
 }
 
+extern uint c4iw_max_read_depth;
+
+static inline int cur_max_read_depth(struct c4iw_dev *dev)
+{
+	return min(dev->rdev.lldi.max_ordird_qp, c4iw_max_read_depth);
+}
+
 struct c4iw_pd {
 	struct ib_pd ibpd;
 	u32 pdid;
@@ -1003,7 +1011,6 @@ void c4iw_ev_dispatch(struct c4iw_dev *dev, struct t4_cqe *err_cqe);
 
 extern struct cxgb4_client t4c_client;
 extern c4iw_handler_func c4iw_handlers[NUM_CPL_CMDS];
-extern int c4iw_max_read_depth;
 extern int db_fc_threshold;
 extern int db_coalescing_threshold;
 extern int use_dsgl;
diff --git a/drivers/infiniband/hw/cxgb4/provider.c b/drivers/infiniband/hw/cxgb4/provider.c
index 1d41b92..67c4a69 100644
--- a/drivers/infiniband/hw/cxgb4/provider.c
+++ b/drivers/infiniband/hw/cxgb4/provider.c
@@ -322,8 +322,10 @@ static int c4iw_query_device(struct ib_device *ibdev,
 	props->max_qp_wr = dev->rdev.hw_queue.t4_max_qp_depth;
 	props->max_sge = T4_MAX_RECV_SGE;
 	props->max_sge_rd = 1;
-	props->max_qp_rd_atom = c4iw_max_read_depth;
-	props->max_qp_init_rd_atom = c4iw_max_read_depth;
+	props->max_res_rd_atom = dev->rdev.lldi.max_ird_adapter;
+	props->max_qp_rd_atom = min(dev->rdev.lldi.max_ordird_qp,
+				    c4iw_max_read_depth);
+	props->max_qp_init_rd_atom = props->max_qp_rd_atom;
 	props->max_cq = T4_MAX_NUM_CQ;
 	props->max_cqe = dev->rdev.hw_queue.t4_max_cq_depth;
 	props->max_mr = c4iw_num_stags(&dev->rdev);
diff --git a/drivers/infiniband/hw/cxgb4/qp.c b/drivers/infiniband/hw/cxgb4/qp.c
index 6f74e0e..0de3cf6 100644
--- a/drivers/infiniband/hw/cxgb4/qp.c
+++ b/drivers/infiniband/hw/cxgb4/qp.c
@@ -58,6 +58,31 @@ static int max_fr_immd = T4_MAX_FR_IMMD;
 module_param(max_fr_immd, int, 0644);
 MODULE_PARM_DESC(max_fr_immd, "fastreg threshold for using DSGL instead of immedate");
 
+static int alloc_ird(struct c4iw_dev *dev, u32 ird)
+{
+	int ret = 0;
+
+	spin_lock_irq(&dev->lock);
+	if (ird <= dev->avail_ird)
+		dev->avail_ird -= ird;
+	else
+		ret = -ENOMEM;
+	spin_unlock_irq(&dev->lock);
+
+	if (ret)
+		dev_warn(&dev->rdev.lldi.pdev->dev,
+			 "device IRD resources exhausted\n");
+
+	return ret;
+}
+
+static void free_ird(struct c4iw_dev *dev, int ird)
+{
+	spin_lock_irq(&dev->lock);
+	dev->avail_ird += ird;
+	spin_unlock_irq(&dev->lock);
+}
+
 static void set_state(struct c4iw_qp *qhp, enum c4iw_qp_state state)
 {
 	unsigned long flag;
@@ -1204,12 +1229,20 @@ static int rdma_init(struct c4iw_dev *rhp, struct c4iw_qp *qhp)
 	int ret;
 	struct sk_buff *skb;
 
-	PDBG("%s qhp %p qid 0x%x tid %u\n", __func__, qhp, qhp->wq.sq.qid,
-	     qhp->ep->hwtid);
+	PDBG("%s qhp %p qid 0x%x tid %u ird %u ord %u\n", __func__, qhp,
+	     qhp->wq.sq.qid, qhp->ep->hwtid, qhp->ep->ird, qhp->ep->ord);
 
 	skb = alloc_skb(sizeof *wqe, GFP_KERNEL);
-	if (!skb)
-		return -ENOMEM;
+	if (!skb) {
+		ret = -ENOMEM;
+		goto out;
+	}
+	ret = alloc_ird(rhp, qhp->attr.max_ird);
+	if (ret) {
+		qhp->attr.max_ird = 0;
+		kfree_skb(skb);
+		goto out;
+	}
 	set_wr_txq(skb, CPL_PRIORITY_DATA, qhp->ep->txq_idx);
 
 	wqe = (struct fw_ri_wr *)__skb_put(skb, sizeof(*wqe));
@@ -1260,10 +1293,14 @@ static int rdma_init(struct c4iw_dev *rhp, struct c4iw_qp *qhp)
 
 	ret = c4iw_ofld_send(&rhp->rdev, skb);
 	if (ret)
-		goto out;
+		goto err1;
 
 	ret = c4iw_wait_for_reply(&rhp->rdev, &qhp->ep->com.wr_wait,
 				  qhp->ep->hwtid, qhp->wq.sq.qid, __func__);
+	if (!ret)
+		goto out;
+err1:
+	free_ird(rhp, qhp->attr.max_ird);
 out:
 	PDBG("%s ret %d\n", __func__, ret);
 	return ret;
@@ -1308,7 +1345,7 @@ int c4iw_modify_qp(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
 			newattr.max_ord = attrs->max_ord;
 		}
 		if (mask & C4IW_QP_ATTR_MAX_IRD) {
-			if (attrs->max_ird > c4iw_max_read_depth) {
+			if (attrs->max_ird > cur_max_read_depth(rhp)) {
 				ret = -EINVAL;
 				goto out;
 			}
@@ -1531,6 +1568,7 @@ int c4iw_destroy_qp(struct ib_qp *ib_qp)
 	if (!list_empty(&qhp->db_fc_entry))
 		list_del_init(&qhp->db_fc_entry);
 	spin_unlock_irq(&rhp->lock);
+	free_ird(rhp, qhp->attr.max_ird);
 
 	ucontext = ib_qp->uobject ?
 		   to_c4iw_ucontext(ib_qp->uobject->context) : NULL;
@@ -1621,8 +1659,8 @@ struct ib_qp *c4iw_create_qp(struct ib_pd *pd, struct ib_qp_init_attr *attrs,
 	qhp->attr.enable_rdma_read = 1;
 	qhp->attr.enable_rdma_write = 1;
 	qhp->attr.enable_bind = 1;
-	qhp->attr.max_ord = 1;
-	qhp->attr.max_ird = 1;
+	qhp->attr.max_ord = 0;
+	qhp->attr.max_ird = 0;
 	qhp->sq_sig_all = attrs->sq_sig_type == IB_SIGNAL_ALL_WR;
 	spin_lock_init(&qhp->lock);
 	mutex_init(&qhp->mutex);