linux-rdma.vger.kernel.org archive mirror
 help / color / mirror / Atom feed
From: Bob Pearson <rpearsonhpe@gmail.com>
To: jgg@nvidia.com, zyjzyj2000@gmail.com, linux-rdma@vger.kernel.org
Cc: Bob Pearson <rpearsonhpe@gmail.com>
Subject: [PATCH v2 2/2] Providers/rxe: Replace AV by AH for UD sends
Date: Sun, 18 Jul 2021 16:28:43 -0500	[thread overview]
Message-ID: <20210718212842.21559-3-rpearsonhpe@gmail.com> (raw)
In-Reply-To: <20210718212842.21559-1-rpearsonhpe@gmail.com>

Currently the rxe provider computes an AV for each AH created. These
are then passed to the kernel in each UD send WQE. This rxe provider
patch matches a set of kernel patches that creates an index for the
pool of AHs which are passed back to the rxe provider which in turn
returns them in the UD send WQEs instead of AVs. This falls back to
the older AV based WQE if either the rxe provider or the kernel does
not have these patches applied.

Signed-off-by: Bob Pearson <rpearsonhpe@gmail.com>
---
 providers/rxe/rxe-abi.h |   2 +
 providers/rxe/rxe.c     | 117 ++++++++++++++++++++++++++--------------
 providers/rxe/rxe.h     |   8 ++-
 3 files changed, 85 insertions(+), 42 deletions(-)

diff --git a/providers/rxe/rxe-abi.h b/providers/rxe/rxe-abi.h
index 0689c3f4..020201a9 100644
--- a/providers/rxe/rxe-abi.h
+++ b/providers/rxe/rxe-abi.h
@@ -39,6 +39,8 @@
 #include <rdma/rdma_user_rxe.h>
 #include <kernel-abi/rdma_user_rxe.h>
 
+DECLARE_DRV_CMD(urxe_create_ah, IB_USER_VERBS_CMD_CREATE_AH,
+		empty, rxe_create_ah_resp);
 DECLARE_DRV_CMD(urxe_create_cq, IB_USER_VERBS_CMD_CREATE_CQ,
 		empty, rxe_create_cq_resp);
 DECLARE_DRV_CMD(urxe_create_cq_ex, IB_USER_VERBS_EX_CMD_CREATE_CQ,
diff --git a/providers/rxe/rxe.c b/providers/rxe/rxe.c
index d19aa9df..b34fb488 100644
--- a/providers/rxe/rxe.c
+++ b/providers/rxe/rxe.c
@@ -975,16 +975,20 @@ static void wr_set_ud_addr(struct ibv_qp_ex *ibqp, struct ibv_ah *ibah,
 			   uint32_t remote_qpn, uint32_t remote_qkey)
 {
 	struct rxe_qp *qp = container_of(ibqp, struct rxe_qp, vqp.qp_ex);
-	struct rxe_ah *ah = container_of(ibah, struct rxe_ah, ibv_ah);
+	struct rxe_ah *ah = to_rah(ibah);
 	struct rxe_send_wqe *wqe = addr_from_index(qp->sq.queue,
 						   qp->cur_index - 1);
 
 	if (qp->err)
 		return;
 
-	memcpy(&wqe->wr.wr.ud.av, &ah->av, sizeof(ah->av));
 	wqe->wr.wr.ud.remote_qpn = remote_qpn;
 	wqe->wr.wr.ud.remote_qkey = remote_qkey;
+	wqe->wr.wr.ud.ah_num = ah->ah_num;
+
+	if (!ah->ah_num)
+		/* old kernels only */
+		memcpy(&wqe->wr.wr.ud.av, &ah->av, sizeof(ah->av));
 }
 
 static void wr_set_inline_data(struct ibv_qp_ex *ibqp, void *addr,
@@ -1404,7 +1408,8 @@ static int validate_send_wr(struct rxe_qp *qp, struct ibv_send_wr *ibwr,
 	return 0;
 }
 
-static void convert_send_wr(struct rxe_send_wr *kwr, struct ibv_send_wr *uwr)
+static void convert_send_wr(struct rxe_qp *qp, struct rxe_send_wr *kwr,
+					struct ibv_send_wr *uwr)
 {
 	struct ibv_mw *ibmw;
 	struct ibv_mr *ibmr;
@@ -1427,8 +1432,13 @@ static void convert_send_wr(struct rxe_send_wr *kwr, struct ibv_send_wr *uwr)
 
 	case IBV_WR_SEND:
 	case IBV_WR_SEND_WITH_IMM:
-		kwr->wr.ud.remote_qpn		= uwr->wr.ud.remote_qpn;
-		kwr->wr.ud.remote_qkey		= uwr->wr.ud.remote_qkey;
+		if (qp_type(qp) == IBV_QPT_UD) {
+			struct rxe_ah *ah = to_rah(uwr->wr.ud.ah);
+
+			kwr->wr.ud.remote_qpn	= uwr->wr.ud.remote_qpn;
+			kwr->wr.ud.remote_qkey	= uwr->wr.ud.remote_qkey;
+			kwr->wr.ud.ah_num	= ah->ah_num;
+		}
 		break;
 
 	case IBV_WR_ATOMIC_CMP_AND_SWP:
@@ -1464,11 +1474,15 @@ static int init_send_wqe(struct rxe_qp *qp, struct rxe_wq *sq,
 	int i;
 	unsigned int opcode = ibwr->opcode;
 
-	convert_send_wr(&wqe->wr, ibwr);
+	convert_send_wr(qp, &wqe->wr, ibwr);
 
-	if (qp_type(qp) == IBV_QPT_UD)
-		memcpy(&wqe->wr.wr.ud.av, &to_rah(ibwr->wr.ud.ah)->av,
-		       sizeof(struct rxe_av));
+	if (qp_type(qp) == IBV_QPT_UD) {
+		struct rxe_ah *ah = to_rah(ibwr->wr.ud.ah);
+
+		if (!ah->ah_num)
+			/* old kernels only */
+			memcpy(&wqe->wr.wr.ud.av, &ah->av, sizeof(struct rxe_av));
+	}
 
 	if (ibwr->send_flags & IBV_SEND_INLINE) {
 		uint8_t *inline_data = wqe->dma.inline_data;
@@ -1645,59 +1659,82 @@ static inline int rdma_gid2ip(sockaddr_union_t *out, union ibv_gid *gid)
 	return 0;
 }
 
-static struct ibv_ah *rxe_create_ah(struct ibv_pd *pd, struct ibv_ah_attr *attr)
+static int rxe_create_av(struct rxe_ah *ah, struct ibv_pd *pd,
+					struct ibv_ah_attr *attr)
 {
-	int err;
-	struct rxe_ah *ah;
-	struct rxe_av *av;
+	struct rxe_av *av = &ah->av;
 	union ibv_gid sgid;
-	struct ib_uverbs_create_ah_resp resp;
-
-	err = ibv_query_gid(pd->context, attr->port_num, attr->grh.sgid_index,
-			    &sgid);
-	if (err) {
-		fprintf(stderr, "rxe: Failed to query sgid.\n");
-		return NULL;
-	}
+	int ret;
 
-	ah = malloc(sizeof(*ah));
-	if (ah == NULL)
-		return NULL;
+	ret = ibv_query_gid(pd->context, attr->port_num,
+			attr->grh.sgid_index, &sgid);
+	if (ret)
+		return ret;
 
-	av = &ah->av;
 	av->port_num = attr->port_num;
 	memcpy(&av->grh, &attr->grh, sizeof(attr->grh));
-	av->network_type =
-		ipv6_addr_v4mapped((struct in6_addr *)attr->grh.dgid.raw) ?
-		RXE_NETWORK_TYPE_IPV4 : RXE_NETWORK_TYPE_IPV6;
+
+	ret = ipv6_addr_v4mapped((struct in6_addr *)attr->grh.dgid.raw);
+	av->network_type = ret ? RXE_NETWORK_TYPE_IPV4 :
+				 RXE_NETWORK_TYPE_IPV6;
 
 	rdma_gid2ip(&av->sgid_addr, &sgid);
 	rdma_gid2ip(&av->dgid_addr, &attr->grh.dgid);
-	if (ibv_resolve_eth_l2_from_gid(pd->context, attr, av->dmac, NULL)) {
-		free(ah);
-		return NULL;
-	}
 
-	memset(&resp, 0, sizeof(resp));
-	if (ibv_cmd_create_ah(pd, &ah->ibv_ah, attr, &resp, sizeof(resp))) {
-		free(ah);
+	ret = ibv_resolve_eth_l2_from_gid(pd->context, attr,
+			av->dmac, NULL);
+
+	return ret;
+}
+
+/*
+ * Newer kernels will return a non-zero AH index in resp.ah_num
+ * which can be returned in UD send WQEs.
+ * Older kernels will leave ah_num == 0. For these create an AV and use
+ * in UD send WQEs.
+ */
+static struct ibv_ah *rxe_create_ah(struct ibv_pd *pd,
+					struct ibv_ah_attr *attr)
+{
+	struct rxe_ah *ah;
+	struct urxe_create_ah_resp resp = {};
+	int ret;
+
+	ah = calloc(1, sizeof(*ah));
+	if (!ah)
 		return NULL;
+
+	ret = ibv_cmd_create_ah(pd, &ah->ibv_ah, attr,
+					&resp.ibv_resp, sizeof(resp));
+	if (ret)
+		goto err_free;
+
+	ah->ah_num = resp.ah_num;
+
+	if (!ah->ah_num) {
+		/* old kernels only */
+		ret = rxe_create_av(ah, pd, attr);
+		if (ret)
+			goto err_free;
 	}
 
 	return &ah->ibv_ah;
+
+err_free:
+	free(ah);
+	return NULL;
 }
 
 static int rxe_destroy_ah(struct ibv_ah *ibah)
 {
-	int ret;
 	struct rxe_ah *ah = to_rah(ibah);
+	int ret;
 
 	ret = ibv_cmd_destroy_ah(&ah->ibv_ah);
-	if (ret)
-		return ret;
+	if (!ret)
+		free(ah);
 
-	free(ah);
-	return 0;
+	return ret;
 }
 
 static const struct verbs_context_ops rxe_ctx_ops = {
diff --git a/providers/rxe/rxe.h b/providers/rxe/rxe.h
index ddb8fe89..6882d9c7 100644
--- a/providers/rxe/rxe.h
+++ b/providers/rxe/rxe.h
@@ -67,6 +67,7 @@ struct rxe_cq {
 struct rxe_ah {
 	struct ibv_ah		ibv_ah;
 	struct rxe_av		av;
+	int			ah_num;
 };
 
 struct rxe_wq {
@@ -89,8 +90,6 @@ struct rxe_qp {
 	int			err;
 };
 
-#define qp_type(qp)		((qp)->vqp.qp.qp_type)
-
 struct rxe_srq {
 	struct ibv_srq		ibv_srq;
 	struct mminfo		mmap_info;
@@ -130,4 +129,9 @@ static inline struct rxe_ah *to_rah(struct ibv_ah *ibah)
 	return to_rxxx(ah, ah);
 }
 
+static inline enum ibv_qp_type qp_type(struct rxe_qp *qp)
+{
+	return qp->vqp.qp.qp_type;
+}
+
 #endif /* RXE_H */
-- 
2.30.2


      parent reply	other threads:[~2021-07-18 21:29 UTC|newest]

Thread overview: 3+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2021-07-18 21:28 [PATCH v2 0/2] Providers/rxe: Replace AV by AH for UD sends Bob Pearson
2021-07-18 21:28 ` [PATCH v2 1/2] Update kernel headers Bob Pearson
2021-07-18 21:28 ` Bob Pearson [this message]

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=20210718212842.21559-3-rpearsonhpe@gmail.com \
    --to=rpearsonhpe@gmail.com \
    --cc=jgg@nvidia.com \
    --cc=linux-rdma@vger.kernel.org \
    --cc=zyjzyj2000@gmail.com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).