rdma-core/0008-libhns-Add-support-for-lock-free-QP.patch

From f5f54bf889825da254b2a5df859da1c471a40314 Mon Sep 17 00:00:00 2001
From: zzry <1245464216@qq.com>
Date: Fri, 8 Mar 2024 15:56:09 +0800
Subject: [PATCH 08/10] libhns: Add support for lock-free QP

driver inclusion
category: feature
bugzilla: https://gitee.com/src-openeuler/rdma-core/issues/I97WST

------------------------------------------------------------------

Drop QP locks when associated to a PAD holding a TD.

Signed-off-by: Yixing Liu <liuyixing1@huawei.com>
Signed-off-by: Junxian Huang <huangjunxian6@hisilicon.com>
---
 providers/hns/hns_roce_u.h       |  2 +-
 providers/hns/hns_roce_u_hw_v2.c | 26 ++++++++++-------
 providers/hns/hns_roce_u_verbs.c | 49 +++++++++++++++++++++++++++++---
 3 files changed, 61 insertions(+), 16 deletions(-)

diff --git a/providers/hns/hns_roce_u.h b/providers/hns/hns_roce_u.h
index 5d3f480..5732e39 100644
--- a/providers/hns/hns_roce_u.h
+++ b/providers/hns/hns_roce_u.h
@@ -305,7 +305,7 @@ struct hns_roce_srq {
 
 struct hns_roce_wq {
 	unsigned long			*wrid;
-	pthread_spinlock_t		lock;
+	struct hns_roce_spinlock	hr_lock;
 	unsigned int			wqe_cnt;
 	int				max_post;
 	unsigned int			head;
diff --git a/providers/hns/hns_roce_u_hw_v2.c b/providers/hns/hns_roce_u_hw_v2.c
index dd13049..90a76e2 100644
--- a/providers/hns/hns_roce_u_hw_v2.c
+++ b/providers/hns/hns_roce_u_hw_v2.c
@@ -1270,7 +1270,7 @@ int hns_roce_u_v2_post_send(struct ibv_qp *ibvqp, struct ibv_send_wr *wr,
 		return ret;
 	}
 
-	pthread_spin_lock(&qp->sq.lock);
+	hns_roce_spin_lock(&qp->sq.hr_lock);
 
 	sge_info.start_idx = qp->next_sge; /* start index of extend sge */
 
@@ -1331,7 +1331,7 @@ out:
 			*(qp->sdb) = qp->sq.head & 0xffff;
 	}
 
-	pthread_spin_unlock(&qp->sq.lock);
+	hns_roce_spin_unlock(&qp->sq.hr_lock);
 
 	if (ibvqp->state == IBV_QPS_ERR) {
 		attr.qp_state = IBV_QPS_ERR;
@@ -1420,7 +1420,7 @@ static int hns_roce_u_v2_post_recv(struct ibv_qp *ibvqp, struct ibv_recv_wr *wr,
 		return ret;
 	}
 
-	pthread_spin_lock(&qp->rq.lock);
+	hns_roce_spin_lock(&qp->rq.hr_lock);
 
 	max_sge = qp->rq.max_gs - qp->rq.rsv_sge;
 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
@@ -1454,7 +1454,7 @@ out:
 			hns_roce_update_rq_db(ctx, ibvqp->qp_num, qp->rq.head);
 	}
 
-	pthread_spin_unlock(&qp->rq.lock);
+	hns_roce_spin_unlock(&qp->rq.hr_lock);
 
 	if (ibvqp->state == IBV_QPS_ERR) {
 		attr.qp_state = IBV_QPS_ERR;
@@ -1549,8 +1549,8 @@ static int hns_roce_u_v2_modify_qp(struct ibv_qp *qp, struct ibv_qp_attr *attr,
 	int ret;
 
 	if ((attr_mask & IBV_QP_STATE) && (attr->qp_state == IBV_QPS_ERR)) {
-		pthread_spin_lock(&hr_qp->sq.lock);
-		pthread_spin_lock(&hr_qp->rq.lock);
+		hns_roce_spin_lock(&hr_qp->sq.hr_lock);
+		hns_roce_spin_lock(&hr_qp->rq.hr_lock);
 		flag = true;
 	}
 
@@ -1561,8 +1561,8 @@ static int hns_roce_u_v2_modify_qp(struct ibv_qp *qp, struct ibv_qp_attr *attr,
 	if (flag) {
 		if (!ret)
 			qp->state = IBV_QPS_ERR;
-		pthread_spin_unlock(&hr_qp->rq.lock);
-		pthread_spin_unlock(&hr_qp->sq.lock);
+		hns_roce_spin_unlock(&hr_qp->sq.hr_lock);
+		hns_roce_spin_unlock(&hr_qp->rq.hr_lock);
 	}
 
 	if (ret)
@@ -1640,6 +1640,7 @@ static void hns_roce_unlock_cqs(struct ibv_qp *qp)
 static int hns_roce_u_v2_destroy_qp(struct ibv_qp *ibqp)
 {
 	struct hns_roce_context *ctx = to_hr_ctx(ibqp->context);
+	struct hns_roce_pad *pad = to_hr_pad(ibqp->pd);
 	struct hns_roce_qp *qp = to_hr_qp(ibqp);
 	int ret;
 
@@ -1666,6 +1667,9 @@ static int hns_roce_u_v2_destroy_qp(struct ibv_qp *ibqp)
 
 	hns_roce_free_qp_buf(qp, ctx);
 
+	if (pad)
+		atomic_fetch_sub(&pad->pd.refcount, 1);
+
 	free(qp);
 
 	return ret;
@@ -2555,7 +2559,7 @@ static void wr_start(struct ibv_qp_ex *ibv_qp)
 		return;
 	}
 
-	pthread_spin_lock(&qp->sq.lock);
+	hns_roce_spin_lock(&qp->sq.hr_lock);
 	qp->sge_info.start_idx = qp->next_sge;
 	qp->rb_sq_head = qp->sq.head;
 	qp->err = 0;
@@ -2588,7 +2592,7 @@ static int wr_complete(struct ibv_qp_ex *ibv_qp)
 	}
 
 out:
-	pthread_spin_unlock(&qp->sq.lock);
+	hns_roce_spin_unlock(&qp->sq.hr_lock);
 	if (ibv_qp->qp_base.state == IBV_QPS_ERR) {
 		attr.qp_state = IBV_QPS_ERR;
 		hns_roce_u_v2_modify_qp(&ibv_qp->qp_base, &attr, IBV_QP_STATE);
@@ -2603,7 +2607,7 @@ static void wr_abort(struct ibv_qp_ex *ibv_qp)
 
 	qp->sq.head = qp->rb_sq_head;
 
-	pthread_spin_unlock(&qp->sq.lock);
+	hns_roce_spin_unlock(&qp->sq.hr_lock);
 }
 
 enum {
diff --git a/providers/hns/hns_roce_u_verbs.c b/providers/hns/hns_roce_u_verbs.c
index ecf8666..d503031 100644
--- a/providers/hns/hns_roce_u_verbs.c
+++ b/providers/hns/hns_roce_u_verbs.c
@@ -1033,6 +1033,41 @@ static int verify_qp_create_attr(struct hns_roce_context *ctx,
 	return verify_qp_create_cap(ctx, attr);
 }
 
+static int hns_roce_qp_spinlock_init(struct hns_roce_context *ctx,
+				     struct ibv_qp_init_attr_ex *attr,
+				     struct hns_roce_qp *qp)
+{
+	bool sq_need_lock;
+	bool rq_need_lock;
+	int ret;
+
+	sq_need_lock = hns_roce_whether_need_lock(attr->pd);
+	if (!sq_need_lock)
+		verbs_info(&ctx->ibv_ctx, "configure sq as no lock.\n");
+
+	rq_need_lock = hns_roce_whether_need_lock(attr->pd);
+	if (!rq_need_lock)
+		verbs_info(&ctx->ibv_ctx, "configure rq as no lock.\n");
+
+	ret = hns_roce_spinlock_init(&qp->sq.hr_lock, sq_need_lock);
+	if (ret)
+		return ret;
+
+	ret = hns_roce_spinlock_init(&qp->rq.hr_lock, rq_need_lock);
+	if (ret) {
+		hns_roce_spinlock_destroy(&qp->sq.hr_lock);
+		return ret;
+	}
+
+	return 0;
+}
+
+void hns_roce_qp_spinlock_destroy(struct hns_roce_qp *qp)
+{
+	hns_roce_spinlock_destroy(&qp->rq.hr_lock);
+	hns_roce_spinlock_destroy(&qp->sq.hr_lock);
+}
+
 static int alloc_recv_rinl_buf(uint32_t max_sge,
 			       struct hns_roce_rinl_buf *rinl_buf)
 {
@@ -1435,10 +1470,6 @@ static int hns_roce_alloc_qp_buf(struct ibv_qp_init_attr_ex *attr,
 {
 	int ret;
 
-	if (pthread_spin_init(&qp->sq.lock, PTHREAD_PROCESS_PRIVATE) ||
-	    pthread_spin_init(&qp->rq.lock, PTHREAD_PROCESS_PRIVATE))
-		return -ENOMEM;
-
 	ret = qp_alloc_wqe(&attr->cap, qp, ctx);
 	if (ret)
 		return ret;
@@ -1466,6 +1497,7 @@ static struct ibv_qp *create_qp(struct ibv_context *ibv_ctx,
 				struct hnsdv_qp_init_attr *hns_attr)
 {
 	struct hns_roce_context *context = to_hr_ctx(ibv_ctx);
+	struct hns_roce_pad *pad = to_hr_pad(attr->pd);
 	struct hns_roce_qp *qp;
 	uint64_t dwqe_mmap_key;
 	int ret;
@@ -1482,6 +1514,13 @@ static struct ibv_qp *create_qp(struct ibv_context *ibv_ctx,
 
 	hns_roce_set_qp_params(attr, qp, context);
 
+	if (pad)
+		atomic_fetch_add(&pad->pd.refcount, 1);
+
+	ret = hns_roce_qp_spinlock_init(context, attr, qp);
+	if (ret)
+		goto err_spinlock;
+
 	ret = hns_roce_alloc_qp_buf(attr, qp, context);
 	if (ret)
 		goto err_buf;
@@ -1515,6 +1554,8 @@ err_ops:
 err_cmd:
 	hns_roce_free_qp_buf(qp, context);
 err_buf:
+	hns_roce_qp_spinlock_destroy(qp);
+err_spinlock:
 	free(qp);
 err:
 	if (ret < 0)
-- 
2.33.0