Skip to contentSkip to conversation

RDMA/hns: Several fixes and cleanups of RQ/SRQ

PATCH Applied Weihang Li 2 participants 14 messages Jan 30, 2021 – Feb 8, 2021

Subsystems (2)
lore
Share & tools
lore

Weihang Li

author
#0 lore raw

cover RDMA/hns: Several fixes and cleanups of RQ/SRQ

There are some issues when using SRQ on HIP08/HIP09, the first part of this series is some fixes on them.

In addition, the codes about RQ/SRQ including the creation and post recv flow are a bit hard to understand, they need to be refactored.

Lang Cheng (2):
RDMA/hns: Allocate one more recv SGE for HIP08
RDMA/hns: Use new interfaces to write SRQC

Wenpeng Liang (8):
RDMA/hns: Bugfix for checking whether the srq is full when post wr RDMA/hns: Force srq_limit to 0 when creating SRQ RDMA/hns: Fixed wrong judgments in the goto branch RDMA/hns: Remove the reserved WQE of SRQ
RDMA/hns: Refactor hns_roce_create_srq()
RDMA/hns: Refactor code about SRQ Context
RDMA/hns: Refactor hns_roce_v2_post_srq_recv() RDMA/hns: Add verification of QP type when post_recv

Xi Wang (2):
RDMA/hns: Refactor post recv flow
RDMA/hns: Clear remaining unused sges when post_recv

 drivers/infiniband/hw/hns/hns_roce_device.h |  16 +-
 drivers/infiniband/hw/hns/hns_roce_hw_v2.c  | 407 +++++++++++++++-------------
 drivers/infiniband/hw/hns/hns_roce_hw_v2.h  |  72 +++--
 drivers/infiniband/hw/hns/hns_roce_main.c   |   3 +-
 drivers/infiniband/hw/hns/hns_roce_qp.c     |  37 ++-
 drivers/infiniband/hw/hns/hns_roce_srq.c    | 329 ++++++++++++----------
 6 files changed, 510 insertions(+), 354 deletions(-)
Signature · 1 line
2.8.1
author
#1 lore raw

PATCH 1/12 RDMA/hns: Allocate one more recv SGE for HIP08

From: Lang Cheng <[email protected]>

The RQ/SRQ of HIP08 needs one special sge to stop receive reliably. So the driver needs to allocate at least one SGE when creating RQ/SRQ and ensure that at least one SGE is filled with the special value during post_recv.

Besides, the kernel driver should only do this for kernel ULP. For userspace ULP, the userspace driver will allocate the reserved SGE in buffer, and the kernel driver just needs to pin the corresponding size of memory based on the userspace driver's requirements.

---
 drivers/infiniband/hw/hns/hns_roce_device.h |  4 +++
 drivers/infiniband/hw/hns/hns_roce_hw_v2.c  | 28 +++++++++++++-------
 drivers/infiniband/hw/hns/hns_roce_hw_v2.h  |  3 ++-
 drivers/infiniband/hw/hns/hns_roce_qp.c     | 37 +++++++++++++++++++++-----
 drivers/infiniband/hw/hns/hns_roce_srq.c    | 40 ++++++++++++++++++++++++++---
 5 files changed, 93 insertions(+), 19 deletions(-)
drivers/infiniband/hw/hns/hns_roce_device.h +4 -0
index f62851f..72961e4 100644
@@ -65,6 +65,8 @@
 #define HNS_ROCE_CQE_WCMD_EMPTY_BIT		0x2
 #define HNS_ROCE_MIN_CQE_CNT			16
 
+#define HNS_ROCE_RESERVED_SGE			1
+
 #define HNS_ROCE_MAX_IRQ_NUM			128
 
 #define HNS_ROCE_SGE_IN_WQE			2
@@ -395,6 +397,7 @@ struct hns_roce_wq {
 	spinlock_t	lock;
 	u32		wqe_cnt;  /* WQE num */
 	u32		max_gs;
+	u32		rsv_sge;
 	int		offset;
 	int		wqe_shift;	/* WQE size */
 	u32		head;
@@ -498,6 +501,7 @@ struct hns_roce_srq {
 	unsigned long		srqn;
 	u32			wqe_cnt;
 	int			max_gs;
+	u32			rsv_sge;
 	int			wqe_shift;
 	void __iomem		*db_reg_l;
 
drivers/infiniband/hw/hns/hns_roce_hw_v2.c +19 -9
index a5bbfb1..2245d25 100644
@@ -741,6 +741,7 @@ static int hns_roce_v2_post_recv(struct ib_qp *ibqp,
 	unsigned long flags;
 	void *wqe = NULL;
 	u32 wqe_idx;
+	u32 max_sge;
 	int nreq;
 	int ret;
 	int i;
@@ -754,6 +755,7 @@ static int hns_roce_v2_post_recv(struct ib_qp *ibqp,
 		goto out;
 	}
 
+	max_sge = hr_qp->rq.max_gs - hr_qp->rq.rsv_sge;
 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
 		if (unlikely(hns_roce_wq_overflow(&hr_qp->rq, nreq,
 						  hr_qp->ibqp.recv_cq))) {
@@ -764,9 +766,9 @@ static int hns_roce_v2_post_recv(struct ib_qp *ibqp,
 
 		wqe_idx = (hr_qp->rq.head + nreq) & (hr_qp->rq.wqe_cnt - 1);
 
-		if (unlikely(wr->num_sge > hr_qp->rq.max_gs)) {
+		if (unlikely(wr->num_sge > max_sge)) {
 			ibdev_err(ibdev, "num_sge = %d >= max_sge = %u.\n",
-				  wr->num_sge, hr_qp->rq.max_gs);
+				  wr->num_sge, max_sge);
 			ret = -EINVAL;
 			*bad_wr = wr;
 			goto out;
@@ -781,9 +783,10 @@ static int hns_roce_v2_post_recv(struct ib_qp *ibqp,
 			dseg++;
 		}
 
-		if (wr->num_sge < hr_qp->rq.max_gs) {
+		if (hr_qp->rq.rsv_sge) {
 			dseg->lkey = cpu_to_le32(HNS_ROCE_INVALID_LKEY);
 			dseg->addr = 0;
+			dseg->len = cpu_to_le32(HNS_ROCE_INVALID_SGE_LENGTH);
 		}
 
 		/* rq support inline data */
@@ -879,6 +882,7 @@ static int hns_roce_v2_post_srq_recv(struct ib_srq *ibsrq,
 	__le32 *srq_idx;
 	int ret = 0;
 	int wqe_idx;
+	u32 max_sge;
 	void *wqe;
 	int nreq;
 	int i;
@@ -886,9 +890,13 @@ static int hns_roce_v2_post_srq_recv(struct ib_srq *ibsrq,
 	spin_lock_irqsave(&srq->lock, flags);
 
 	ind = srq->head & (srq->wqe_cnt - 1);
+	max_sge = srq->max_gs - srq->rsv_sge;
 
 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
-		if (unlikely(wr->num_sge >= srq->max_gs)) {
+		if (unlikely(wr->num_sge > max_sge)) {
+			ibdev_err(&hr_dev->ib_dev,
+				  "srq: num_sge = %d, max_sge = %u.\n",
+				  wr->num_sge, max_sge);
 			ret = -EINVAL;
 			*bad_wr = wr;
 			break;
@@ -916,9 +924,9 @@ static int hns_roce_v2_post_srq_recv(struct ib_srq *ibsrq,
 			dseg[i].addr = cpu_to_le64(wr->sg_list[i].addr);
 		}
 
-		if (wr->num_sge < srq->max_gs) {
-			dseg[i].len = 0;
-			dseg[i].lkey = cpu_to_le32(0x100);
+		if (srq->rsv_sge) {
+			dseg[i].len = cpu_to_le32(HNS_ROCE_INVALID_SGE_LENGTH);
+			dseg[i].lkey = cpu_to_le32(HNS_ROCE_INVALID_LKEY);
 			dseg[i].addr = 0;
 		}
 
@@ -1999,10 +2007,12 @@ static int hns_roce_query_pf_caps(struct hns_roce_dev *hr_dev)
 	caps->max_sq_sg		     = le16_to_cpu(resp_a->max_sq_sg);
 	caps->max_sq_inline	     = le16_to_cpu(resp_a->max_sq_inline);
 	caps->max_rq_sg		     = le16_to_cpu(resp_a->max_rq_sg);
+	caps->max_rq_sg = roundup_pow_of_two(caps->max_rq_sg);
 	caps->max_extend_sg	     = le32_to_cpu(resp_a->max_extend_sg);
 	caps->num_qpc_timer	     = le16_to_cpu(resp_a->num_qpc_timer);
 	caps->num_cqc_timer	     = le16_to_cpu(resp_a->num_cqc_timer);
 	caps->max_srq_sges	     = le16_to_cpu(resp_a->max_srq_sges);
+	caps->max_srq_sges = roundup_pow_of_two(caps->max_srq_sges);
 	caps->num_aeq_vectors	     = resp_a->num_aeq_vectors;
 	caps->num_other_vectors	     = resp_a->num_other_vectors;
 	caps->max_sq_desc_sz	     = resp_a->max_sq_desc_sz;
@@ -5071,7 +5081,7 @@ static int hns_roce_v2_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr,
 done:
 	qp_attr->cur_qp_state = qp_attr->qp_state;
 	qp_attr->cap.max_recv_wr = hr_qp->rq.wqe_cnt;
-	qp_attr->cap.max_recv_sge = hr_qp->rq.max_gs;
+	qp_attr->cap.max_recv_sge = hr_qp->rq.max_gs - hr_qp->rq.rsv_sge;
 
 	if (!ibqp->uobject) {
 		qp_attr->cap.max_send_wr = hr_qp->sq.wqe_cnt;
@@ -5383,7 +5393,7 @@ static int hns_roce_v2_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr)
 
 	attr->srq_limit = limit_wl;
 	attr->max_wr = srq->wqe_cnt - 1;
-	attr->max_sge = srq->max_gs;
+	attr->max_sge = srq->max_gs - srq->rsv_sge;
 
 out:
 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
drivers/infiniband/hw/hns/hns_roce_hw_v2.h +2 -1
index 69bc072..cd9abdd 100644
@@ -96,7 +96,8 @@
 #define HNS_ROCE_V2_CQC_TIMER_ENTRY_SZ		PAGE_SIZE
 #define HNS_ROCE_V2_PAGE_SIZE_SUPPORTED		0xFFFFF000
 #define HNS_ROCE_V2_MAX_INNER_MTPT_NUM		2
-#define HNS_ROCE_INVALID_LKEY			0x100
+#define HNS_ROCE_INVALID_LKEY			0x0
+#define HNS_ROCE_INVALID_SGE_LENGTH		0x80000000
 #define HNS_ROCE_CMQ_TX_TIMEOUT			30000
 #define HNS_ROCE_V2_UC_RC_SGE_NUM_IN_WQE	2
 #define HNS_ROCE_V2_RSV_QPS			8
drivers/infiniband/hw/hns/hns_roce_qp.c +31 -6
index 9988ca9..8af411f 100644
@@ -413,9 +413,32 @@ static void free_qpn(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
 	spin_unlock(&hr_dev->qp_table.bank_lock);
 }
 
+static u32 proc_rq_sge(struct hns_roce_dev *dev, struct hns_roce_qp *hr_qp,
+		       bool user)
+{
+	u32 max_sge = dev->caps.max_rq_sg;
+
+	if (dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
+		return max_sge;
+
+	/* Reserve SGEs only for HIP08 in kernel; The userspace driver will
+	 * calculate number of max_sge with reserved SGEs when allocating wqe
+	 * buf, so there is no need to do this again in kernel. But the number
+	 * may exceed the capacity of SGEs recorded in the firmware, so the
+	 * kernel driver should just adapt the value accordingly.
+	 */
+	if (user)
+		max_sge = roundup_pow_of_two(max_sge + 1);
+	else
+		hr_qp->rq.rsv_sge = 1;
+
+	return max_sge;
+}
+
 static int set_rq_size(struct hns_roce_dev *hr_dev, struct ib_qp_cap *cap,
-		       struct hns_roce_qp *hr_qp, int has_rq)
+		       struct hns_roce_qp *hr_qp, int has_rq, bool user)
 {
+	u32 max_sge = proc_rq_sge(hr_dev, hr_qp, user);
 	u32 cnt;
 
 	/* If srq exist, set zero for relative number of rq */
@@ -431,8 +454,9 @@ static int set_rq_size(struct hns_roce_dev *hr_dev, struct ib_qp_cap *cap,
 
 	/* Check the validity of QP support capacity */
 	if (!cap->max_recv_wr || cap->max_recv_wr > hr_dev->caps.max_wqes ||
-	    cap->max_recv_sge > hr_dev->caps.max_rq_sg) {
-		ibdev_err(&hr_dev->ib_dev, "RQ config error, depth=%u, sge=%d\n",
+	    cap->max_recv_sge > max_sge) {
+		ibdev_err(&hr_dev->ib_dev,
+			  "RQ config error, depth = %u, sge = %u\n",
 			  cap->max_recv_wr, cap->max_recv_sge);
 		return -EINVAL;
 	}
@@ -444,7 +468,8 @@ static int set_rq_size(struct hns_roce_dev *hr_dev, struct ib_qp_cap *cap,
 		return -EINVAL;
 	}
 
-	hr_qp->rq.max_gs = roundup_pow_of_two(max(1U, cap->max_recv_sge));
+	hr_qp->rq.max_gs = roundup_pow_of_two(max(1U, cap->max_recv_sge) +
+					      hr_qp->rq.rsv_sge);
 
 	if (hr_dev->caps.max_rq_sg <= HNS_ROCE_SGE_IN_WQE)
 		hr_qp->rq.wqe_shift = ilog2(hr_dev->caps.max_rq_desc_sz);
@@ -459,7 +484,7 @@ static int set_rq_size(struct hns_roce_dev *hr_dev, struct ib_qp_cap *cap,
 		hr_qp->rq_inl_buf.wqe_cnt = 0;
 
 	cap->max_recv_wr = cnt;
-	cap->max_recv_sge = hr_qp->rq.max_gs;
+	cap->max_recv_sge = hr_qp->rq.max_gs - hr_qp->rq.rsv_sge;
 
 	return 0;
 }
@@ -918,7 +943,7 @@ static int set_qp_param(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
 		hr_qp->sq_signal_bits = IB_SIGNAL_REQ_WR;
 
 	ret = set_rq_size(hr_dev, &init_attr->cap, hr_qp,
-			  hns_roce_qp_has_rq(init_attr));
+			  hns_roce_qp_has_rq(init_attr), !!udata);
 	if (ret) {
 		ibdev_err(ibdev, "failed to set user RQ size, ret = %d.\n",
 			  ret);
drivers/infiniband/hw/hns/hns_roce_srq.c +37 -3
index 9403828..1be6812 100644
@@ -3,6 +3,7 @@
  * Copyright (c) 2018 Hisilicon Limited.
  */
 
+#include <linux/pci.h>
 #include <rdma/ib_umem.h>
 #include "hns_roce_device.h"
 #include "hns_roce_cmd.h"
@@ -277,6 +278,28 @@ static void free_srq_wrid(struct hns_roce_srq *srq)
 	srq->wrid = NULL;
 }
 
+static u32 proc_srq_sge(struct hns_roce_dev *dev, struct hns_roce_srq *hr_srq,
+			bool user)
+{
+	u32 max_sge = dev->caps.max_srq_sges;
+
+	if (dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
+		return max_sge;
+
+	/* Reserve SGEs only for HIP08 in kernel; The userspace driver will
+	 * calculate number of max_sge with reserved SGEs when allocating wqe
+	 * buf, so there is no need to do this again in kernel. But the number
+	 * may exceed the capacity of SGEs recorded in the firmware, so the
+	 * kernel driver should just adapt the value accordingly.
+	 */
+	if (user)
+		max_sge = roundup_pow_of_two(max_sge + 1);
+	else
+		hr_srq->rsv_sge = 1;
+
+	return max_sge;
+}
+
 int hns_roce_create_srq(struct ib_srq *ib_srq,
 			struct ib_srq_init_attr *init_attr,
 			struct ib_udata *udata)
@@ -286,6 +309,7 @@ int hns_roce_create_srq(struct ib_srq *ib_srq,
 	struct hns_roce_srq *srq = to_hr_srq(ib_srq);
 	struct ib_device *ibdev = &hr_dev->ib_dev;
 	struct hns_roce_ib_create_srq ucmd = {};
+	u32 max_sge;
 	int ret;
 	u32 cqn;
 
@@ -293,16 +317,24 @@ int hns_roce_create_srq(struct ib_srq *ib_srq,
 	    init_attr->srq_type != IB_SRQT_XRC)
 		return -EOPNOTSUPP;
 
-	/* Check the actual SRQ wqe and SRQ sge num */
+	max_sge = proc_srq_sge(hr_dev, srq, !!udata);
+
 	if (init_attr->attr.max_wr >= hr_dev->caps.max_srq_wrs ||
-	    init_attr->attr.max_sge > hr_dev->caps.max_srq_sges)
+	    init_attr->attr.max_sge > max_sge) {
+		ibdev_err(&hr_dev->ib_dev,
+			  "SRQ config error, depth = %u, sge = %d\n",
+			  init_attr->attr.max_wr, init_attr->attr.max_sge);
 		return -EINVAL;
+	}
 
 	mutex_init(&srq->mutex);
 	spin_lock_init(&srq->lock);
 
 	srq->wqe_cnt = roundup_pow_of_two(init_attr->attr.max_wr + 1);
-	srq->max_gs = init_attr->attr.max_sge;
+	srq->max_gs =
+		roundup_pow_of_two(init_attr->attr.max_sge + srq->rsv_sge);
+	init_attr->attr.max_wr = srq->wqe_cnt;
+	init_attr->attr.max_sge = srq->max_gs;
 
 	if (udata) {
 		ret = ib_copy_from_udata(&ucmd, udata,
@@ -349,6 +381,8 @@ int hns_roce_create_srq(struct ib_srq *ib_srq,
 
 	srq->event = hns_roce_ib_srq_event;
 	resp.srqn = srq->srqn;
+	srq->max_gs = init_attr->attr.max_sge;
+	init_attr->attr.max_sge = srq->max_gs - srq->rsv_sge;
 
 	if (udata) {
 		ret = ib_copy_to_udata(udata, &resp,
Signature · 1 line
2.8.1
author
#2 lore raw

PATCH 2/12 RDMA/hns: Bugfix for checking whether the srq is full when post wr

In reply to Weihang Li

From: Wenpeng Liang <[email protected]>

If a user posts WR by wr_list, the head pointer of idx_queue won't be updated until all wqes are filled, so the judgment of whether head equals to tail will get a wrong result. Fix above issue and move the head and tail pointer from the srq structure into the idx_queue structure. After idx_queue is filled with wqe idx, the head pointer of it will increase.

Fixes c7bcb13442e1 ("RDMA/hns: Add SRQ support for hip08 kernel mode")
---
 drivers/infiniband/hw/hns/hns_roce_device.h |  4 ++--
 drivers/infiniband/hw/hns/hns_roce_hw_v2.c  | 19 ++++++++++++++-----
 drivers/infiniband/hw/hns/hns_roce_srq.c    |  5 +++--
 3 files changed, 19 insertions(+), 9 deletions(-)
drivers/infiniband/hw/hns/hns_roce_device.h +2 -2
index 72961e4..916e031 100644
@@ -494,6 +494,8 @@ struct hns_roce_idx_que {
 	struct hns_roce_mtr		mtr;
 	int				entry_shift;
 	unsigned long			*bitmap;
+	u32				head;
+	u32				tail;
 };
 
 struct hns_roce_srq {
@@ -513,8 +515,6 @@ struct hns_roce_srq {
 	u64		       *wrid;
 	struct hns_roce_idx_que idx_que;
 	spinlock_t		lock;
-	u16			head;
-	u16			tail;
 	struct mutex		mutex;
 	void (*event)(struct hns_roce_srq *srq, enum hns_roce_event event);
 };
drivers/infiniband/hw/hns/hns_roce_hw_v2.c +14 -5
index 2245d25..b13775a 100644
@@ -849,11 +849,20 @@ static void hns_roce_free_srq_wqe(struct hns_roce_srq *srq, int wqe_index)
 	spin_lock(&srq->lock);
 
 	bitmap_clear(srq->idx_que.bitmap, wqe_index, 1);
-	srq->tail++;
+	srq->idx_que.tail++;
 
 	spin_unlock(&srq->lock);
 }
 
+int hns_roce_srqwq_overflow(struct hns_roce_srq *srq, int nreq)
+{
+	struct hns_roce_idx_que *idx_que = &srq->idx_que;
+	unsigned int cur;
+
+	cur = idx_que->head - idx_que->tail;
+	return cur + nreq >= srq->wqe_cnt - 1;
+}
+
 static int find_empty_entry(struct hns_roce_idx_que *idx_que,
 			    unsigned long size)
 {
@@ -889,7 +898,7 @@ static int hns_roce_v2_post_srq_recv(struct ib_srq *ibsrq,
 
 	spin_lock_irqsave(&srq->lock, flags);
 
-	ind = srq->head & (srq->wqe_cnt - 1);
+	ind = srq->idx_que.head & (srq->wqe_cnt - 1);
 	max_sge = srq->max_gs - srq->rsv_sge;
 
 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
@@ -902,7 +911,7 @@ static int hns_roce_v2_post_srq_recv(struct ib_srq *ibsrq,
 			break;
 		}
 
-		if (unlikely(srq->head == srq->tail)) {
+		if (unlikely(hns_roce_srqwq_overflow(srq, nreq))) {
 			ret = -ENOMEM;
 			*bad_wr = wr;
 			break;
@@ -938,7 +947,7 @@ static int hns_roce_v2_post_srq_recv(struct ib_srq *ibsrq,
 	}
 
 	if (likely(nreq)) {
-		srq->head += nreq;
+		srq->idx_que.head += nreq;
 
 		/*
 		 * Make sure that descriptors are written before
@@ -950,7 +959,7 @@ static int hns_roce_v2_post_srq_recv(struct ib_srq *ibsrq,
 			cpu_to_le32(HNS_ROCE_V2_SRQ_DB << V2_DB_BYTE_4_CMD_S |
 				    (srq->srqn & V2_DB_BYTE_4_TAG_M));
 		srq_db.parameter =
-			cpu_to_le32(srq->head & V2_DB_PARAMETER_IDX_M);
+			cpu_to_le32(srq->idx_que.head & V2_DB_PARAMETER_IDX_M);
 
 		hns_roce_write64(hr_dev, (__le32 *)&srq_db, srq->db_reg_l);
 	}
drivers/infiniband/hw/hns/hns_roce_srq.c +3 -2
index 1be6812..e622fd1d 100644
@@ -245,6 +245,9 @@ static int alloc_srq_idx(struct hns_roce_dev *hr_dev, struct hns_roce_srq *srq,
 		}
 	}
 
+	idx_que->head = 0;
+	idx_que->tail = 0;
+
 	return 0;
 err_idx_mtr:
 	hns_roce_mtr_destroy(hr_dev, &idx_que->mtr);
@@ -263,8 +266,6 @@ static void free_srq_idx(struct hns_roce_dev *hr_dev, struct hns_roce_srq *srq)
 
 static int alloc_srq_wrid(struct hns_roce_dev *hr_dev, struct hns_roce_srq *srq)
 {
-	srq->head = 0;
-	srq->tail = srq->wqe_cnt - 1;
 	srq->wrid = kvmalloc_array(srq->wqe_cnt, sizeof(u64), GFP_KERNEL);
 	if (!srq->wrid)
 		return -ENOMEM;
Signature · 1 line
2.8.1
author
#3 lore raw

PATCH 3/12 RDMA/hns: Force srq_limit to 0 when creating SRQ

In reply to Weihang Li

From: Wenpeng Liang <[email protected]>

According to the IB Specification, srq_limit shouldn't be configured during SRQ creation. If a user set srq_limit at this time, the driver should forced it to zero, or the result of creating SRQ will conflict with the result of querying SRQ.

Fixes c7bcb13442e1 ("RDMA/hns: Add SRQ support for hip08 kernel mode")
---
 drivers/infiniband/hw/hns/hns_roce_srq.c | 1 +
 1 file changed, 1 insertion(+)
drivers/infiniband/hw/hns/hns_roce_srq.c +1 -0
index e622fd1d..47e66fe 100644
@@ -336,6 +336,7 @@ int hns_roce_create_srq(struct ib_srq *ib_srq,
 		roundup_pow_of_two(init_attr->attr.max_sge + srq->rsv_sge);
 	init_attr->attr.max_wr = srq->wqe_cnt;
 	init_attr->attr.max_sge = srq->max_gs;
+	init_attr->attr.srq_limit = 0;
 
 	if (udata) {
 		ret = ib_copy_from_udata(&ucmd, udata,
Signature · 1 line
2.8.1
author
#4 lore raw

PATCH 4/12 RDMA/hns: Fixed wrong judgments in the goto branch

In reply to Weihang Li

From: Wenpeng Liang <[email protected]>

When an error occurs, the qp_table must be cleared, regardless of whether the SRQ feature is enabled.

Fixes 5c1f167af112 ("RDMA/hns: Init SRQ table for hip08")
---
 drivers/infiniband/hw/hns/hns_roce_main.c | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)
drivers/infiniband/hw/hns/hns_roce_main.c +1 -2
index 2b78b1f..7978220 100644
@@ -768,8 +768,7 @@ static int hns_roce_setup_hca(struct hns_roce_dev *hr_dev)
 	return 0;
 
 err_qp_table_free:
-	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_SRQ)
-		hns_roce_cleanup_qp_table(hr_dev);
+	hns_roce_cleanup_qp_table(hr_dev);
 
 err_cq_table_free:
 	hns_roce_cleanup_cq_table(hr_dev);
Signature · 1 line
2.8.1
author
#5 lore raw

PATCH 5/12 RDMA/hns: Remove the reserved WQE of SRQ

In reply to Weihang Li

From: Wenpeng Liang <[email protected]>

Each SRQs contain an reserved WQE, it is inappropriate and should be removed.

Fixes c7bcb13442e1 ("RDMA/hns: Add SRQ support for hip08 kernel mode")
---
 drivers/infiniband/hw/hns/hns_roce_device.h | 1 +
 drivers/infiniband/hw/hns/hns_roce_hw_v2.c  | 6 +++---
 drivers/infiniband/hw/hns/hns_roce_srq.c    | 6 ++++--
 3 files changed, 8 insertions(+), 5 deletions(-)
drivers/infiniband/hw/hns/hns_roce_device.h +1 -0
index 916e031..d6a846b 100644
@@ -54,6 +54,7 @@
 /* Hardware specification only for v1 engine */
 #define HNS_ROCE_MIN_CQE_NUM			0x40
 #define HNS_ROCE_MIN_WQE_NUM			0x20
+#define HNS_ROCE_MIN_SRQ_WQE_NUM		1
 
 /* Hardware specification only for v1 engine */
 #define HNS_ROCE_MAX_INNER_MTPT_NUM		0x7
drivers/infiniband/hw/hns/hns_roce_hw_v2.c +3 -3
index b13775a..49a8456 100644
@@ -860,7 +860,7 @@ int hns_roce_srqwq_overflow(struct hns_roce_srq *srq, int nreq)
 	unsigned int cur;
 
 	cur = idx_que->head - idx_que->tail;
-	return cur + nreq >= srq->wqe_cnt - 1;
+	return cur + nreq >= srq->wqe_cnt;
 }
 
 static int find_empty_entry(struct hns_roce_idx_que *idx_que,
@@ -5338,7 +5338,7 @@ static int hns_roce_v2_modify_srq(struct ib_srq *ibsrq,
 		return -EINVAL;
 
 	if (srq_attr_mask & IB_SRQ_LIMIT) {
-		if (srq_attr->srq_limit >= srq->wqe_cnt)
+		if (srq_attr->srq_limit > srq->wqe_cnt)
 			return -EINVAL;
 
 		mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
@@ -5401,7 +5401,7 @@ static int hns_roce_v2_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr)
 				  SRQC_BYTE_8_SRQ_LIMIT_WL_S);
 
 	attr->srq_limit = limit_wl;
-	attr->max_wr = srq->wqe_cnt - 1;
+	attr->max_wr = srq->wqe_cnt;
 	attr->max_sge = srq->max_gs - srq->rsv_sge;
 
 out:
drivers/infiniband/hw/hns/hns_roce_srq.c +4 -2
index 47e66fe..5d20b30 100644
@@ -320,7 +320,7 @@ int hns_roce_create_srq(struct ib_srq *ib_srq,
 
 	max_sge = proc_srq_sge(hr_dev, srq, !!udata);
 
-	if (init_attr->attr.max_wr >= hr_dev->caps.max_srq_wrs ||
+	if (init_attr->attr.max_wr > hr_dev->caps.max_srq_wrs ||
 	    init_attr->attr.max_sge > max_sge) {
 		ibdev_err(&hr_dev->ib_dev,
 			  "SRQ config error, depth = %u, sge = %d\n",
@@ -331,7 +331,9 @@ int hns_roce_create_srq(struct ib_srq *ib_srq,
 	mutex_init(&srq->mutex);
 	spin_lock_init(&srq->lock);
 
-	srq->wqe_cnt = roundup_pow_of_two(init_attr->attr.max_wr + 1);
+	init_attr->attr.max_wr = max_t(u32, init_attr->attr.max_wr,
+				       HNS_ROCE_MIN_SRQ_WQE_NUM);
+	srq->wqe_cnt = roundup_pow_of_two(init_attr->attr.max_wr);
 	srq->max_gs =
 		roundup_pow_of_two(init_attr->attr.max_sge + srq->rsv_sge);
 	init_attr->attr.max_wr = srq->wqe_cnt;
Signature · 1 line
2.8.1
author
#6 lore raw

PATCH 6/12 RDMA/hns: Refactor hns_roce_create_srq()

In reply to Weihang Li

From: Wenpeng Liang <[email protected]>

Split the SRQ creation process into multiple steps and encapsulate them into functions.

---
 drivers/infiniband/hw/hns/hns_roce_device.h |   5 +-
 drivers/infiniband/hw/hns/hns_roce_hw_v2.c  |  12 +-
 drivers/infiniband/hw/hns/hns_roce_srq.c    | 269 ++++++++++++++++------------
 3 files changed, 161 insertions(+), 125 deletions(-)
drivers/infiniband/hw/hns/hns_roce_device.h +3 -2
index d6a846b..b325b9c 100644
@@ -506,6 +506,7 @@ struct hns_roce_srq {
 	int			max_gs;
 	u32			rsv_sge;
 	int			wqe_shift;
+	u32			cqn;
 	void __iomem		*db_reg_l;
 
 	atomic_t		refcount;
@@ -953,8 +954,8 @@ struct hns_roce_hw {
 	int (*init_eq)(struct hns_roce_dev *hr_dev);
 	void (*cleanup_eq)(struct hns_roce_dev *hr_dev);
 	void (*write_srqc)(struct hns_roce_dev *hr_dev,
-			   struct hns_roce_srq *srq, u32 pdn, u16 xrcd, u32 cqn,
-			   void *mb_buf, u64 *mtts_wqe, u64 *mtts_idx,
+			   struct hns_roce_srq *srq, void *mb_buf,
+			   u64 *mtts_wqe, u64 *mtts_idx,
 			   dma_addr_t dma_handle_wqe,
 			   dma_addr_t dma_handle_idx);
 	int (*modify_srq)(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr,
drivers/infiniband/hw/hns/hns_roce_hw_v2.c +6 -6
index 49a8456..ec2d64c 100644
@@ -5228,9 +5228,9 @@ static int hns_roce_v2_qp_flow_control_init(struct hns_roce_dev *hr_dev,
 }
 
 static void hns_roce_v2_write_srqc(struct hns_roce_dev *hr_dev,
-				   struct hns_roce_srq *srq, u32 pdn, u16 xrcd,
-				   u32 cqn, void *mb_buf, u64 *mtts_wqe,
-				   u64 *mtts_idx, dma_addr_t dma_handle_wqe,
+				   struct hns_roce_srq *srq, void *mb_buf,
+				   u64 *mtts_wqe, u64 *mtts_idx,
+				   dma_addr_t dma_handle_wqe,
 				   dma_addr_t dma_handle_idx)
 {
 	struct hns_roce_srq_context *srq_context;
@@ -5257,7 +5257,7 @@ static void hns_roce_v2_write_srqc(struct hns_roce_dev *hr_dev,
 		       SRQC_BYTE_8_SRQ_LIMIT_WL_S, 0);
 
 	roce_set_field(srq_context->byte_12_xrcd, SRQC_BYTE_12_SRQ_XRCD_M,
-		       SRQC_BYTE_12_SRQ_XRCD_S, xrcd);
+		       SRQC_BYTE_12_SRQ_XRCD_S, 0);
 
 	srq_context->wqe_bt_ba = cpu_to_le32((u32)(dma_handle_wqe >> 3));
 
@@ -5267,7 +5267,7 @@ static void hns_roce_v2_write_srqc(struct hns_roce_dev *hr_dev,
 		       dma_handle_wqe >> 35);
 
 	roce_set_field(srq_context->byte_28_rqws_pd, SRQC_BYTE_28_PD_M,
-		       SRQC_BYTE_28_PD_S, pdn);
+		       SRQC_BYTE_28_PD_S, to_hr_pd(srq->ibsrq.pd)->pdn);
 	roce_set_field(srq_context->byte_28_rqws_pd, SRQC_BYTE_28_RQWS_M,
 		       SRQC_BYTE_28_RQWS_S, srq->max_gs <= 0 ? 0 :
 		       fls(srq->max_gs - 1));
@@ -5307,7 +5307,7 @@ static void hns_roce_v2_write_srqc(struct hns_roce_dev *hr_dev,
 		       upper_32_bits(to_hr_hw_page_addr(mtts_idx[1])));
 	roce_set_field(srq_context->byte_56_xrc_cqn,
 		       SRQC_BYTE_56_SRQ_XRC_CQN_M, SRQC_BYTE_56_SRQ_XRC_CQN_S,
-		       cqn);
+		       srq->cqn);
 	roce_set_field(srq_context->byte_56_xrc_cqn,
 		       SRQC_BYTE_56_SRQ_WQE_BA_PG_SZ_M,
 		       SRQC_BYTE_56_SRQ_WQE_BA_PG_SZ_S,
drivers/infiniband/hw/hns/hns_roce_srq.c +152 -117

374 lines not on this page · raw message

Signature · 1 line
2.8.1
author
#7 lore raw

PATCH 7/12 RDMA/hns: Refactor code about SRQ Context

In reply to Weihang Li

From: Wenpeng Liang <[email protected]>

Reduce parameter numbers of write_srqc() and move some related code into it from alloc_srqc().

---
 drivers/infiniband/hw/hns/hns_roce_device.h |  6 +---
 drivers/infiniband/hw/hns/hns_roce_hw_v2.c  | 33 ++++++++++++++++++----
 drivers/infiniband/hw/hns/hns_roce_srq.c    | 44 +++++++++--------------------
 3 files changed, 42 insertions(+), 41 deletions(-)
drivers/infiniband/hw/hns/hns_roce_device.h +1 -5
index b325b9c..d51641a 100644
@@ -953,11 +953,7 @@ struct hns_roce_hw {
 	int (*modify_cq)(struct ib_cq *cq, u16 cq_count, u16 cq_period);
 	int (*init_eq)(struct hns_roce_dev *hr_dev);
 	void (*cleanup_eq)(struct hns_roce_dev *hr_dev);
-	void (*write_srqc)(struct hns_roce_dev *hr_dev,
-			   struct hns_roce_srq *srq, void *mb_buf,
-			   u64 *mtts_wqe, u64 *mtts_idx,
-			   dma_addr_t dma_handle_wqe,
-			   dma_addr_t dma_handle_idx);
+	int (*write_srqc)(struct hns_roce_srq *srq, void *mb_buf);
 	int (*modify_srq)(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr,
 		       enum ib_srq_attr_mask srq_attr_mask,
 		       struct ib_udata *udata);
drivers/infiniband/hw/hns/hns_roce_hw_v2.c +28 -5
index ec2d64c..dd5f7b5 100644
@@ -5227,17 +5227,38 @@ static int hns_roce_v2_qp_flow_control_init(struct hns_roce_dev *hr_dev,
 	return ret;
 }
 
-static void hns_roce_v2_write_srqc(struct hns_roce_dev *hr_dev,
-				   struct hns_roce_srq *srq, void *mb_buf,
-				   u64 *mtts_wqe, u64 *mtts_idx,
-				   dma_addr_t dma_handle_wqe,
-				   dma_addr_t dma_handle_idx)
+static int hns_roce_v2_write_srqc(struct hns_roce_srq *srq, void *mb_buf)
 {
+	struct ib_device *ibdev = srq->ibsrq.device;
+	struct hns_roce_dev *hr_dev = to_hr_dev(ibdev);
 	struct hns_roce_srq_context *srq_context;
+	u64 mtts_wqe[MTT_MIN_COUNT] = {};
+	u64 mtts_idx[MTT_MIN_COUNT] = {};
+	dma_addr_t dma_handle_wqe = 0;
+	dma_addr_t dma_handle_idx = 0;
+	int ret;
 
 	srq_context = mb_buf;
 	memset(srq_context, 0, sizeof(*srq_context));
 
+	/* Get the physical address of srq buf */
+	ret = hns_roce_mtr_find(hr_dev, &srq->buf_mtr, 0, mtts_wqe,
+				ARRAY_SIZE(mtts_wqe), &dma_handle_wqe);
+	if (ret < 1) {
+		ibdev_err(ibdev, "failed to find mtr for SRQ WQE, ret = %d.\n",
+			  ret);
+		return -ENOBUFS;
+	}
+
+	/* Get physical address of idx que buf */
+	ret = hns_roce_mtr_find(hr_dev, &srq->idx_que.mtr, 0, mtts_idx,
+				ARRAY_SIZE(mtts_idx), &dma_handle_idx);
+	if (ret < 1) {
+		ibdev_err(ibdev, "failed to find mtr for SRQ idx, ret = %d.\n",
+			  ret);
+		return -ENOBUFS;
+	}
+
 	roce_set_field(srq_context->byte_4_srqn_srqst, SRQC_BYTE_4_SRQ_ST_M,
 		       SRQC_BYTE_4_SRQ_ST_S, 1);
 
@@ -5319,6 +5340,8 @@ static void hns_roce_v2_write_srqc(struct hns_roce_dev *hr_dev,
 
 	roce_set_bit(srq_context->db_record_addr_record_en,
 		     SRQC_BYTE_60_SRQ_RECORD_EN_S, 0);
+
+	return 0;
 }
 
 static int hns_roce_v2_modify_srq(struct ib_srq *ibsrq,
drivers/infiniband/hw/hns/hns_roce_srq.c +13 -31

78 lines not on this page · raw message

Signature · 1 line
2.8.1
author
#8 lore raw

PATCH 8/12 RDMA/hns: Use new interfaces to write SRQC

In reply to Weihang Li

From: Lang Cheng <[email protected]>

Use new register operation interfaces to simplify the process of write SRQ Context.

---
 drivers/infiniband/hw/hns/hns_roce_hw_v2.c | 160 ++++++++++++-----------------
 drivers/infiniband/hw/hns/hns_roce_hw_v2.h |  69 ++++++++++---
 2 files changed, 118 insertions(+), 111 deletions(-)
drivers/infiniband/hw/hns/hns_roce_hw_v2.c +65 -95
index dd5f7b5..105019c5 100644
@@ -5227,19 +5227,59 @@ static int hns_roce_v2_qp_flow_control_init(struct hns_roce_dev *hr_dev,
 	return ret;
 }
 
+static int hns_roce_v2_write_srqc_index_queue(struct hns_roce_srq *srq,
+					      struct hns_roce_srq_context *ctx)
+{
+	struct hns_roce_idx_que *idx_que = &srq->idx_que;
+	struct ib_device *ibdev = srq->ibsrq.device;
+	struct hns_roce_dev *hr_dev = to_hr_dev(ibdev);
+	u64 mtts_idx[MTT_MIN_COUNT] = {};
+	dma_addr_t dma_handle_idx = 0;
+	int ret;
+
+	/* Get physical address of idx que buf */
+	ret = hns_roce_mtr_find(hr_dev, &idx_que->mtr, 0, mtts_idx,
+				ARRAY_SIZE(mtts_idx), &dma_handle_idx);
+	if (ret < 1) {
+		ibdev_err(ibdev, "failed to find mtr for SRQ idx, ret = %d.\n",
+			  ret);
+		return -ENOBUFS;
+	}
+
+	hr_reg_write(ctx, SRQC_IDX_HOP_NUM,
+		     to_hr_hem_hopnum(hr_dev->caps.idx_hop_num, srq->wqe_cnt));
+
+	hr_reg_write(ctx, SRQC_IDX_BT_BA_L, dma_handle_idx >> 3);
+	hr_reg_write(ctx, SRQC_IDX_BT_BA_H, upper_32_bits(dma_handle_idx >> 3));
+
+	hr_reg_write(ctx, SRQC_IDX_BA_PG_SZ,
+		     to_hr_hw_page_shift(idx_que->mtr.hem_cfg.ba_pg_shift));
+	hr_reg_write(ctx, SRQC_IDX_BUF_PG_SZ,
+		     to_hr_hw_page_shift(idx_que->mtr.hem_cfg.buf_pg_shift));
+
+	hr_reg_write(ctx, SRQC_IDX_CUR_BLK_ADDR_L,
+		     to_hr_hw_page_addr(mtts_idx[0]));
+	hr_reg_write(ctx, SRQC_IDX_CUR_BLK_ADDR_H,
+		     upper_32_bits(to_hr_hw_page_addr(mtts_idx[0])));
+
+	hr_reg_write(ctx, SRQC_IDX_NXT_BLK_ADDR_L,
+		     to_hr_hw_page_addr(mtts_idx[1]));
+	hr_reg_write(ctx, SRQC_IDX_NXT_BLK_ADDR_H,
+		     upper_32_bits(to_hr_hw_page_addr(mtts_idx[1])));
+
+	return 0;
+}
+
 static int hns_roce_v2_write_srqc(struct hns_roce_srq *srq, void *mb_buf)
 {
 	struct ib_device *ibdev = srq->ibsrq.device;
 	struct hns_roce_dev *hr_dev = to_hr_dev(ibdev);
-	struct hns_roce_srq_context *srq_context;
+	struct hns_roce_srq_context *ctx = mb_buf;
 	u64 mtts_wqe[MTT_MIN_COUNT] = {};
-	u64 mtts_idx[MTT_MIN_COUNT] = {};
 	dma_addr_t dma_handle_wqe = 0;
-	dma_addr_t dma_handle_idx = 0;
 	int ret;
 
-	srq_context = mb_buf;
-	memset(srq_context, 0, sizeof(*srq_context));
+	memset(ctx, 0, sizeof(*ctx));
 
 	/* Get the physical address of srq buf */
 	ret = hns_roce_mtr_find(hr_dev, &srq->buf_mtr, 0, mtts_wqe,
@@ -5250,98 +5290,28 @@ static int hns_roce_v2_write_srqc(struct hns_roce_srq *srq, void *mb_buf)
 		return -ENOBUFS;
 	}
 
-	/* Get physical address of idx que buf */
-	ret = hns_roce_mtr_find(hr_dev, &srq->idx_que.mtr, 0, mtts_idx,
-				ARRAY_SIZE(mtts_idx), &dma_handle_idx);
-	if (ret < 1) {
-		ibdev_err(ibdev, "failed to find mtr for SRQ idx, ret = %d.\n",
-			  ret);
-		return -ENOBUFS;
-	}
+	hr_reg_write(ctx, SRQC_SRQ_ST, 1);
+	hr_reg_write(ctx, SRQC_PD, to_hr_pd(srq->ibsrq.pd)->pdn);
+	hr_reg_write(ctx, SRQC_SRQN, srq->srqn);
+	hr_reg_write(ctx, SRQC_XRCD, 0);
+	hr_reg_write(ctx, SRQC_XRC_CQN, srq->cqn);
+	hr_reg_write(ctx, SRQC_SHIFT, ilog2(srq->wqe_cnt));
+	hr_reg_write(ctx, SRQC_RQWS,
+		     srq->max_gs <= 0 ? 0 : fls(srq->max_gs - 1));
 
-	roce_set_field(srq_context->byte_4_srqn_srqst, SRQC_BYTE_4_SRQ_ST_M,
-		       SRQC_BYTE_4_SRQ_ST_S, 1);
-
-	roce_set_field(srq_context->byte_4_srqn_srqst,
-		       SRQC_BYTE_4_SRQ_WQE_HOP_NUM_M,
-		       SRQC_BYTE_4_SRQ_WQE_HOP_NUM_S,
-		       to_hr_hem_hopnum(hr_dev->caps.srqwqe_hop_num,
-					srq->wqe_cnt));
-	roce_set_field(srq_context->byte_4_srqn_srqst,
-		       SRQC_BYTE_4_SRQ_SHIFT_M, SRQC_BYTE_4_SRQ_SHIFT_S,
-		       ilog2(srq->wqe_cnt));
-
-	roce_set_field(srq_context->byte_4_srqn_srqst, SRQC_BYTE_4_SRQN_M,
-		       SRQC_BYTE_4_SRQN_S, srq->srqn);
-
-	roce_set_field(srq_context->byte_8_limit_wl, SRQC_BYTE_8_SRQ_LIMIT_WL_M,
-		       SRQC_BYTE_8_SRQ_LIMIT_WL_S, 0);
-
-	roce_set_field(srq_context->byte_12_xrcd, SRQC_BYTE_12_SRQ_XRCD_M,
-		       SRQC_BYTE_12_SRQ_XRCD_S, 0);
-
-	srq_context->wqe_bt_ba = cpu_to_le32((u32)(dma_handle_wqe >> 3));
-
-	roce_set_field(srq_context->byte_24_wqe_bt_ba,
-		       SRQC_BYTE_24_SRQ_WQE_BT_BA_M,
-		       SRQC_BYTE_24_SRQ_WQE_BT_BA_S,
-		       dma_handle_wqe >> 35);
-
-	roce_set_field(srq_context->byte_28_rqws_pd, SRQC_BYTE_28_PD_M,
-		       SRQC_BYTE_28_PD_S, to_hr_pd(srq->ibsrq.pd)->pdn);
-	roce_set_field(srq_context->byte_28_rqws_pd, SRQC_BYTE_28_RQWS_M,
-		       SRQC_BYTE_28_RQWS_S, srq->max_gs <= 0 ? 0 :
-		       fls(srq->max_gs - 1));
-
-	srq_context->idx_bt_ba = cpu_to_le32(dma_handle_idx >> 3);
-	roce_set_field(srq_context->rsv_idx_bt_ba,
-		       SRQC_BYTE_36_SRQ_IDX_BT_BA_M,
-		       SRQC_BYTE_36_SRQ_IDX_BT_BA_S,
-		       dma_handle_idx >> 35);
-
-	srq_context->idx_cur_blk_addr =
-		cpu_to_le32(to_hr_hw_page_addr(mtts_idx[0]));
-	roce_set_field(srq_context->byte_44_idxbufpgsz_addr,
-		       SRQC_BYTE_44_SRQ_IDX_CUR_BLK_ADDR_M,
-		       SRQC_BYTE_44_SRQ_IDX_CUR_BLK_ADDR_S,
-		       upper_32_bits(to_hr_hw_page_addr(mtts_idx[0])));
-	roce_set_field(srq_context->byte_44_idxbufpgsz_addr,
-		       SRQC_BYTE_44_SRQ_IDX_HOP_NUM_M,
-		       SRQC_BYTE_44_SRQ_IDX_HOP_NUM_S,
-		       to_hr_hem_hopnum(hr_dev->caps.idx_hop_num,
-					srq->wqe_cnt));
-
-	roce_set_field(srq_context->byte_44_idxbufpgsz_addr,
-		       SRQC_BYTE_44_SRQ_IDX_BA_PG_SZ_M,
-		       SRQC_BYTE_44_SRQ_IDX_BA_PG_SZ_S,
-		to_hr_hw_page_shift(srq->idx_que.mtr.hem_cfg.ba_pg_shift));
-	roce_set_field(srq_context->byte_44_idxbufpgsz_addr,
-		       SRQC_BYTE_44_SRQ_IDX_BUF_PG_SZ_M,
-		       SRQC_BYTE_44_SRQ_IDX_BUF_PG_SZ_S,
-		to_hr_hw_page_shift(srq->idx_que.mtr.hem_cfg.buf_pg_shift));
-
-	srq_context->idx_nxt_blk_addr =
-				cpu_to_le32(to_hr_hw_page_addr(mtts_idx[1]));
-	roce_set_field(srq_context->rsv_idxnxtblkaddr,
-		       SRQC_BYTE_52_SRQ_IDX_NXT_BLK_ADDR_M,
-		       SRQC_BYTE_52_SRQ_IDX_NXT_BLK_ADDR_S,
-		       upper_32_bits(to_hr_hw_page_addr(mtts_idx[1])));
-	roce_set_field(srq_context->byte_56_xrc_cqn,
-		       SRQC_BYTE_56_SRQ_XRC_CQN_M, SRQC_BYTE_56_SRQ_XRC_CQN_S,
-		       srq->cqn);
-	roce_set_field(srq_context->byte_56_xrc_cqn,
-		       SRQC_BYTE_56_SRQ_WQE_BA_PG_SZ_M,
-		       SRQC_BYTE_56_SRQ_WQE_BA_PG_SZ_S,
-		       to_hr_hw_page_shift(srq->buf_mtr.hem_cfg.ba_pg_shift));
-	roce_set_field(srq_context->byte_56_xrc_cqn,
-		       SRQC_BYTE_56_SRQ_WQE_BUF_PG_SZ_M,
-		       SRQC_BYTE_56_SRQ_WQE_BUF_PG_SZ_S,
-		       to_hr_hw_page_shift(srq->buf_mtr.hem_cfg.buf_pg_shift));
-
-	roce_set_bit(srq_context->db_record_addr_record_en,
-		     SRQC_BYTE_60_SRQ_RECORD_EN_S, 0);
+	hr_reg_write(ctx, SRQC_WQE_HOP_NUM,
+		     to_hr_hem_hopnum(hr_dev->caps.srqwqe_hop_num,
+				      srq->wqe_cnt));
 
-	return 0;
+	hr_reg_write(ctx, SRQC_WQE_BT_BA_L, dma_handle_wqe >> 3);
+	hr_reg_write(ctx, SRQC_WQE_BT_BA_H, upper_32_bits(dma_handle_wqe >> 3));
+
+	hr_reg_write(ctx, SRQC_WQE_BA_PG_SZ,
+		     to_hr_hw_page_shift(srq->buf_mtr.hem_cfg.ba_pg_shift));
+	hr_reg_write(ctx, SRQC_WQE_BUF_PG_SZ,
+		     to_hr_hw_page_shift(srq->buf_mtr.hem_cfg.buf_pg_shift));
+
+	return hns_roce_v2_write_srqc_index_queue(srq, ctx);
 }
 
 static int hns_roce_v2_modify_srq(struct ib_srq *ibsrq,
drivers/infiniband/hw/hns/hns_roce_hw_v2.h +53 -16

77 lines not on this page · raw message

Signature · 1 line
2.8.1
author
#9 lore raw

PATCH 9/12 RDMA/hns: Refactor post recv flow

In reply to Weihang Li

From: Xi Wang <[email protected]>

Refactor post recv flow by removing unnecessary checking and removing duplicated code.

---
 drivers/infiniband/hw/hns/hns_roce_hw_v2.c | 81 +++++++++++++++---------------
 1 file changed, 41 insertions(+), 40 deletions(-)
drivers/infiniband/hw/hns/hns_roce_hw_v2.c +41 -40
index 105019c5..1f70422 100644
@@ -48,8 +48,8 @@
 #include "hns_roce_hem.h"
 #include "hns_roce_hw_v2.h"
 
-static void set_data_seg_v2(struct hns_roce_v2_wqe_data_seg *dseg,
-			    struct ib_sge *sg)
+static inline void set_data_seg_v2(struct hns_roce_v2_wqe_data_seg *dseg,
+				   struct ib_sge *sg)
 {
 	dseg->lkey = cpu_to_le32(sg->lkey);
 	dseg->addr = cpu_to_le64(sg->addr);
@@ -729,6 +729,40 @@ static int check_recv_valid(struct hns_roce_dev *hr_dev,
 	return 0;
 }
 
+static void fill_rq_wqe(struct hns_roce_qp *hr_qp, const struct ib_recv_wr *wr,
+			u32 wqe_idx)
+{
+	struct hns_roce_v2_wqe_data_seg *dseg;
+	struct hns_roce_rinl_sge *sge_list;
+	void *wqe = NULL;
+	int i;
+
+	wqe = hns_roce_get_recv_wqe(hr_qp, wqe_idx);
+	dseg = (struct hns_roce_v2_wqe_data_seg *)wqe;
+	for (i = 0; i < wr->num_sge; i++) {
+		if (!wr->sg_list[i].length)
+			continue;
+		set_data_seg_v2(dseg, wr->sg_list + i);
+		dseg++;
+	}
+
+	if (hr_qp->rq.rsv_sge) {
+		dseg->lkey = cpu_to_le32(HNS_ROCE_INVALID_LKEY);
+		dseg->addr = 0;
+		dseg->len = cpu_to_le32(HNS_ROCE_INVALID_SGE_LENGTH);
+	}
+
+	/* rq support inline data */
+	if (hr_qp->rq_inl_buf.wqe_cnt) {
+		sge_list = hr_qp->rq_inl_buf.wqe_list[wqe_idx].sg_list;
+		hr_qp->rq_inl_buf.wqe_list[wqe_idx].sge_cnt = (u32)wr->num_sge;
+		for (i = 0; i < wr->num_sge; i++) {
+			sge_list[i].addr = (void *)(u64)wr->sg_list[i].addr;
+			sge_list[i].len = wr->sg_list[i].length;
+		}
+	}
+}
+
 static int hns_roce_v2_post_recv(struct ib_qp *ibqp,
 				 const struct ib_recv_wr *wr,
 				 const struct ib_recv_wr **bad_wr)
@@ -736,15 +770,9 @@ static int hns_roce_v2_post_recv(struct ib_qp *ibqp,
 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
 	struct ib_device *ibdev = &hr_dev->ib_dev;
-	struct hns_roce_v2_wqe_data_seg *dseg;
-	struct hns_roce_rinl_sge *sge_list;
+	u32 wqe_idx, nreq, max_sge;
 	unsigned long flags;
-	void *wqe = NULL;
-	u32 wqe_idx;
-	u32 max_sge;
-	int nreq;
 	int ret;
-	int i;
 
 	spin_lock_irqsave(&hr_qp->rq.lock, flags);
 
@@ -764,8 +792,6 @@ static int hns_roce_v2_post_recv(struct ib_qp *ibqp,
 			goto out;
 		}
 
-		wqe_idx = (hr_qp->rq.head + nreq) & (hr_qp->rq.wqe_cnt - 1);
-
 		if (unlikely(wr->num_sge > max_sge)) {
 			ibdev_err(ibdev, "num_sge = %d >= max_sge = %u.\n",
 				  wr->num_sge, max_sge);
@@ -774,32 +800,8 @@ static int hns_roce_v2_post_recv(struct ib_qp *ibqp,
 			goto out;
 		}
 
-		wqe = hns_roce_get_recv_wqe(hr_qp, wqe_idx);
-		dseg = (struct hns_roce_v2_wqe_data_seg *)wqe;
-		for (i = 0; i < wr->num_sge; i++) {
-			if (!wr->sg_list[i].length)
-				continue;
-			set_data_seg_v2(dseg, wr->sg_list + i);
-			dseg++;
-		}
-
-		if (hr_qp->rq.rsv_sge) {
-			dseg->lkey = cpu_to_le32(HNS_ROCE_INVALID_LKEY);
-			dseg->addr = 0;
-			dseg->len = cpu_to_le32(HNS_ROCE_INVALID_SGE_LENGTH);
-		}
-
-		/* rq support inline data */
-		if (hr_qp->rq_inl_buf.wqe_cnt) {
-			sge_list = hr_qp->rq_inl_buf.wqe_list[wqe_idx].sg_list;
-			hr_qp->rq_inl_buf.wqe_list[wqe_idx].sge_cnt =
-							       (u32)wr->num_sge;
-			for (i = 0; i < wr->num_sge; i++) {
-				sge_list[i].addr =
-					       (void *)(u64)wr->sg_list[i].addr;
-				sge_list[i].len = wr->sg_list[i].length;
-			}
-		}
+		wqe_idx = (hr_qp->rq.head + nreq) & (hr_qp->rq.wqe_cnt - 1);
+		fill_rq_wqe(hr_qp, wr, wqe_idx);
 
 		hr_qp->rq.wrid[wqe_idx] = wr->wr_id;
 	}
@@ -928,9 +930,8 @@ static int hns_roce_v2_post_srq_recv(struct ib_srq *ibsrq,
 		dseg = (struct hns_roce_v2_wqe_data_seg *)wqe;
 
 		for (i = 0; i < wr->num_sge; ++i) {
-			dseg[i].len = cpu_to_le32(wr->sg_list[i].length);
-			dseg[i].lkey = cpu_to_le32(wr->sg_list[i].lkey);
-			dseg[i].addr = cpu_to_le64(wr->sg_list[i].addr);
+			set_data_seg_v2(dseg, wr->sg_list + i);
+			dseg++;
 		}
 
 		if (srq->rsv_sge) {
Signature · 1 line
2.8.1
author
#10 lore raw

PATCH 10/12 RDMA/hns: Clear remaining unused sges when post_recv

In reply to Weihang Li

From: Xi Wang <[email protected]>

The HIP09 requires the driver to clear the unused data segments in wqe buffer to make the hns ROCEE stop reading the remaining invalid sges for RQ.

---
 drivers/infiniband/hw/hns/hns_roce_hw_v2.c | 99 ++++++++++++++----------------
 1 file changed, 47 insertions(+), 52 deletions(-)
drivers/infiniband/hw/hns/hns_roce_hw_v2.c +47 -52
index 1f70422..7eba9b5 100644
@@ -729,28 +729,42 @@ static int check_recv_valid(struct hns_roce_dev *hr_dev,
 	return 0;
 }
 
-static void fill_rq_wqe(struct hns_roce_qp *hr_qp, const struct ib_recv_wr *wr,
-			u32 wqe_idx)
+static void fill_recv_sge_to_wqe(const struct ib_recv_wr *wr, void *wqe,
+				 u32 max_sge, bool rsv)
 {
-	struct hns_roce_v2_wqe_data_seg *dseg;
-	struct hns_roce_rinl_sge *sge_list;
-	void *wqe = NULL;
-	int i;
+	struct hns_roce_v2_wqe_data_seg *dseg = wqe;
+	u32 i, cnt;
 
-	wqe = hns_roce_get_recv_wqe(hr_qp, wqe_idx);
-	dseg = (struct hns_roce_v2_wqe_data_seg *)wqe;
-	for (i = 0; i < wr->num_sge; i++) {
+	for (i = 0, cnt = 0; i < wr->num_sge; i++) {
+		/* Skip zero-length sge */
 		if (!wr->sg_list[i].length)
 			continue;
-		set_data_seg_v2(dseg, wr->sg_list + i);
-		dseg++;
+		set_data_seg_v2(dseg + cnt, wr->sg_list + i);
+		cnt++;
 	}
 
-	if (hr_qp->rq.rsv_sge) {
-		dseg->lkey = cpu_to_le32(HNS_ROCE_INVALID_LKEY);
-		dseg->addr = 0;
-		dseg->len = cpu_to_le32(HNS_ROCE_INVALID_SGE_LENGTH);
+	/* Fill a reserved sge to make hw stop reading remaining segments */
+	if (rsv) {
+		dseg[cnt].lkey = cpu_to_le32(HNS_ROCE_INVALID_LKEY);
+		dseg[cnt].addr = 0;
+		dseg[cnt].len = cpu_to_le32(HNS_ROCE_INVALID_SGE_LENGTH);
+	} else {
+		/* Clear remaining segments to make ROCEE ignore sges */
+		if (cnt < max_sge)
+			memset(dseg + cnt, 0,
+			       (max_sge - cnt) * HNS_ROCE_SGE_SIZE);
 	}
+}
+
+static void fill_rq_wqe(struct hns_roce_qp *hr_qp, const struct ib_recv_wr *wr,
+			u32 wqe_idx, u32 max_sge)
+{
+	struct hns_roce_rinl_sge *sge_list;
+	void *wqe = NULL;
+	u32 i;
+
+	wqe = hns_roce_get_recv_wqe(hr_qp, wqe_idx);
+	fill_recv_sge_to_wqe(wr, wqe, max_sge, hr_qp->rq.rsv_sge);
 
 	/* rq support inline data */
 	if (hr_qp->rq_inl_buf.wqe_cnt) {
@@ -801,8 +815,7 @@ static int hns_roce_v2_post_recv(struct ib_qp *ibqp,
 		}
 
 		wqe_idx = (hr_qp->rq.head + nreq) & (hr_qp->rq.wqe_cnt - 1);
-		fill_rq_wqe(hr_qp, wr, wqe_idx);
-
+		fill_rq_wqe(hr_qp, wr, wqe_idx, max_sge);
 		hr_qp->rq.wrid[wqe_idx] = wr->wr_id;
 	}
 
@@ -834,18 +847,18 @@ static int hns_roce_v2_post_recv(struct ib_qp *ibqp,
 	return ret;
 }
 
-static void *get_srq_wqe(struct hns_roce_srq *srq, int n)
+static void *get_srq_wqe_buf(struct hns_roce_srq *srq, u32 n)
 {
 	return hns_roce_buf_offset(srq->buf_mtr.kmem, n << srq->wqe_shift);
 }
 
-static void *get_idx_buf(struct hns_roce_idx_que *idx_que, unsigned int n)
+static void *get_idx_buf(struct hns_roce_idx_que *idx_que, u32 n)
 {
 	return hns_roce_buf_offset(idx_que->mtr.kmem,
 				   n << idx_que->entry_shift);
 }
 
-static void hns_roce_free_srq_wqe(struct hns_roce_srq *srq, int wqe_index)
+static void hns_roce_free_srq_wqe(struct hns_roce_srq *srq, u32 wqe_index)
 {
 	/* always called with interrupts disabled. */
 	spin_lock(&srq->lock);
@@ -856,7 +869,7 @@ static void hns_roce_free_srq_wqe(struct hns_roce_srq *srq, int wqe_index)
 	spin_unlock(&srq->lock);
 }
 
-int hns_roce_srqwq_overflow(struct hns_roce_srq *srq, int nreq)
+int hns_roce_srqwq_overflow(struct hns_roce_srq *srq, u32 nreq)
 {
 	struct hns_roce_idx_que *idx_que = &srq->idx_que;
 	unsigned int cur;
@@ -865,19 +878,18 @@ int hns_roce_srqwq_overflow(struct hns_roce_srq *srq, int nreq)
 	return cur + nreq >= srq->wqe_cnt;
 }
 
-static int find_empty_entry(struct hns_roce_idx_que *idx_que,
-			    unsigned long size)
+static int get_srq_wqe_idx(struct hns_roce_srq *srq, u32 *wqe_idx)
 {
-	int wqe_idx;
+	struct hns_roce_idx_que *idx_que = &srq->idx_que;
+	u32 pos;
 
-	if (unlikely(bitmap_full(idx_que->bitmap, size)))
+	pos = find_first_zero_bit(idx_que->bitmap, srq->wqe_cnt);
+	if (unlikely(pos == srq->wqe_cnt))
 		return -ENOSPC;
 
-	wqe_idx = find_first_zero_bit(idx_que->bitmap, size);
-
-	bitmap_set(idx_que->bitmap, wqe_idx, 1);
-
-	return wqe_idx;
+	bitmap_set(idx_que->bitmap, pos, 1);
+	*wqe_idx = pos;
+	return 0;
 }
 
 static int hns_roce_v2_post_srq_recv(struct ib_srq *ibsrq,
@@ -886,17 +898,12 @@ static int hns_roce_v2_post_srq_recv(struct ib_srq *ibsrq,
 {
 	struct hns_roce_dev *hr_dev = to_hr_dev(ibsrq->device);
 	struct hns_roce_srq *srq = to_hr_srq(ibsrq);
-	struct hns_roce_v2_wqe_data_seg *dseg;
+	u32 wqe_idx, ind, nreq, max_sge;
 	struct hns_roce_v2_db srq_db;
 	unsigned long flags;
-	unsigned int ind;
 	__le32 *srq_idx;
 	int ret = 0;
-	int wqe_idx;
-	u32 max_sge;
 	void *wqe;
-	int nreq;
-	int i;
 
 	spin_lock_irqsave(&srq->lock, flags);
 
@@ -919,26 +926,14 @@ static int hns_roce_v2_post_srq_recv(struct ib_srq *ibsrq,
 			break;
 		}
 
-		wqe_idx = find_empty_entry(&srq->idx_que, srq->wqe_cnt);
-		if (unlikely(wqe_idx < 0)) {
-			ret = -ENOMEM;
+		ret = get_srq_wqe_idx(srq, &wqe_idx);
+		if (unlikely(ret)) {
 			*bad_wr = wr;
 			break;
 		}
 
-		wqe = get_srq_wqe(srq, wqe_idx);
-		dseg = (struct hns_roce_v2_wqe_data_seg *)wqe;
-
-		for (i = 0; i < wr->num_sge; ++i) {
-			set_data_seg_v2(dseg, wr->sg_list + i);
-			dseg++;
-		}
-
-		if (srq->rsv_sge) {
-			dseg[i].len = cpu_to_le32(HNS_ROCE_INVALID_SGE_LENGTH);
-			dseg[i].lkey = cpu_to_le32(HNS_ROCE_INVALID_LKEY);
-			dseg[i].addr = 0;
-		}
+		wqe = get_srq_wqe_buf(srq, wqe_idx);
+		fill_recv_sge_to_wqe(wr, wqe, max_sge, srq->rsv_sge);
 
 		srq_idx = get_idx_buf(&srq->idx_que, ind);
 		*srq_idx = cpu_to_le32(wqe_idx);
Signature · 1 line
2.8.1
author
#11 lore raw

PATCH 11/12 RDMA/hns: Refactor hns_roce_v2_post_srq_recv()

In reply to Weihang Li

From: Wenpeng Liang <[email protected]>

The SRQ in the hns driver consists of the following four parts:
* wqe buf: the buffer to store WQE.
* wqe_idx buf: the cqe of SRQ may be not generated in the order of wqe, so the wqe_idx corresponding to the idle WQE needs to be pushed into the index queue which is a FIFO, then it instructs the hardware to obtain the corresponding WQE.
* bitmap: bitmap is used to generate and release wqe_idx. When the user has a new WR, the driver finds the idx of the idle wqe in bitmap. When the CQE of wqe is generated, the driver will release the idx.
* wr_id buf: wr_id buf is used to store the user's wr_id, then return it to the user when poll_cq verb is invoked.

The process of post SRQ recv is refactored to make preceding code clearer.

---
 drivers/infiniband/hw/hns/hns_roce_hw_v2.c | 69 +++++++++++++++++++-----------
 1 file changed, 43 insertions(+), 26 deletions(-)
drivers/infiniband/hw/hns/hns_roce_hw_v2.c +43 -26
index 7eba9b5..d5a63e4 100644
@@ -869,13 +869,32 @@ static void hns_roce_free_srq_wqe(struct hns_roce_srq *srq, u32 wqe_index)
 	spin_unlock(&srq->lock);
 }
 
-int hns_roce_srqwq_overflow(struct hns_roce_srq *srq, u32 nreq)
+static int hns_roce_srqwq_overflow(struct hns_roce_srq *srq)
 {
 	struct hns_roce_idx_que *idx_que = &srq->idx_que;
-	unsigned int cur;
 
-	cur = idx_que->head - idx_que->tail;
-	return cur + nreq >= srq->wqe_cnt;
+	return idx_que->head - idx_que->tail >= srq->wqe_cnt;
+}
+
+static int check_post_srq_valid(struct hns_roce_srq *srq, u32 max_sge,
+				const struct ib_recv_wr *wr)
+{
+	struct ib_device *ib_dev = srq->ibsrq.device;
+
+	if (unlikely(wr->num_sge > max_sge)) {
+		ibdev_err(ib_dev,
+			  "failed to check sge, wr->num_sge = %d, max_sge = %u.\n",
+			  wr->num_sge, max_sge);
+		return -EINVAL;
+	}
+
+	if (unlikely(hns_roce_srqwq_overflow(srq))) {
+		ibdev_err(ib_dev,
+			  "failed to check srqwq status, srqwq is full.\n");
+		return -ENOMEM;
+	}
+
+	return 0;
 }
 
 static int get_srq_wqe_idx(struct hns_roce_srq *srq, u32 *wqe_idx)
@@ -892,36 +911,40 @@ static int get_srq_wqe_idx(struct hns_roce_srq *srq, u32 *wqe_idx)
 	return 0;
 }
 
+static void fill_wqe_idx(struct hns_roce_srq *srq, unsigned int wqe_idx)
+{
+	struct hns_roce_idx_que *idx_que = &srq->idx_que;
+	unsigned int head;
+	__le32 *buf;
+
+	head = idx_que->head & (srq->wqe_cnt - 1);
+
+	buf = get_idx_buf(idx_que, head);
+	*buf = cpu_to_le32(wqe_idx);
+
+	idx_que->head++;
+}
+
 static int hns_roce_v2_post_srq_recv(struct ib_srq *ibsrq,
 				     const struct ib_recv_wr *wr,
 				     const struct ib_recv_wr **bad_wr)
 {
 	struct hns_roce_dev *hr_dev = to_hr_dev(ibsrq->device);
 	struct hns_roce_srq *srq = to_hr_srq(ibsrq);
-	u32 wqe_idx, ind, nreq, max_sge;
 	struct hns_roce_v2_db srq_db;
 	unsigned long flags;
-	__le32 *srq_idx;
 	int ret = 0;
+	u32 max_sge;
+	u32 wqe_idx;
 	void *wqe;
+	u32 nreq;
 
 	spin_lock_irqsave(&srq->lock, flags);
 
-	ind = srq->idx_que.head & (srq->wqe_cnt - 1);
 	max_sge = srq->max_gs - srq->rsv_sge;
-
 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
-		if (unlikely(wr->num_sge > max_sge)) {
-			ibdev_err(&hr_dev->ib_dev,
-				  "srq: num_sge = %d, max_sge = %u.\n",
-				  wr->num_sge, max_sge);
-			ret = -EINVAL;
-			*bad_wr = wr;
-			break;
-		}
-
-		if (unlikely(hns_roce_srqwq_overflow(srq, nreq))) {
-			ret = -ENOMEM;
+		ret = check_post_srq_valid(srq, max_sge, wr);
+		if (ret) {
 			*bad_wr = wr;
 			break;
 		}
@@ -934,17 +957,11 @@ static int hns_roce_v2_post_srq_recv(struct ib_srq *ibsrq,
 
 		wqe = get_srq_wqe_buf(srq, wqe_idx);
 		fill_recv_sge_to_wqe(wr, wqe, max_sge, srq->rsv_sge);
-
-		srq_idx = get_idx_buf(&srq->idx_que, ind);
-		*srq_idx = cpu_to_le32(wqe_idx);
-
+		fill_wqe_idx(srq, wqe_idx);
 		srq->wrid[wqe_idx] = wr->wr_id;
-		ind = (ind + 1) & (srq->wqe_cnt - 1);
 	}
 
 	if (likely(nreq)) {
-		srq->idx_que.head += nreq;
-
 		/*
 		 * Make sure that descriptors are written before
 		 * doorbell record.
Signature · 1 line
2.8.1
author
#12 lore raw

PATCH 12/12 RDMA/hns: Add verification of QP type when post_recv

In reply to Weihang Li

From: Wenpeng Liang <[email protected]>

The post_recv only supports QP types of RC, GSI and UD.

---
 drivers/infiniband/hw/hns/hns_roce_hw_v2.c | 14 +++++++++++++-
 1 file changed, 13 insertions(+), 1 deletion(-)
drivers/infiniband/hw/hns/hns_roce_hw_v2.c +13 -1
index d5a63e4..3adb77d7 100644
@@ -721,9 +721,21 @@ static int hns_roce_v2_post_send(struct ib_qp *ibqp,
 static int check_recv_valid(struct hns_roce_dev *hr_dev,
 			    struct hns_roce_qp *hr_qp)
 {
+	struct ib_device *ibdev = &hr_dev->ib_dev;
+	struct ib_qp *ibqp = &hr_qp->ibqp;
+
+	if (unlikely(ibqp->qp_type != IB_QPT_RC &&
+		     ibqp->qp_type != IB_QPT_GSI &&
+		     ibqp->qp_type != IB_QPT_UD)) {
+		ibdev_err(ibdev, "unsupported qp type, qp_type = %d.\n",
+			  ibqp->qp_type);
+		return -EOPNOTSUPP;
+	}
+
 	if (unlikely(hr_dev->state >= HNS_ROCE_DEVICE_STATE_RST_DOWN))
 		return -EIO;
-	else if (hr_qp->state == IB_QPS_RESET)
+
+	if (hr_qp->state == IB_QPS_RESET)
 		return -EINVAL;
 
 	return 0;
Signature · 1 line
2.8.1
maintainer Applied
#13 lore raw
In reply to Weihang Li
Quoted 20 lines · Weihang Li

There are some issues when using SRQ on HIP08/HIP09, the first part of this series is some fixes on them.

In addition, the codes about RQ/SRQ including the creation and post recv flow are a bit hard to understand, they need to be refactored.

Lang Cheng (2):
RDMA/hns: Allocate one more recv SGE for HIP08
RDMA/hns: Use new interfaces to write SRQC

Wenpeng Liang (8):
RDMA/hns: Bugfix for checking whether the srq is full when post wr RDMA/hns: Force srq_limit to 0 when creating SRQ RDMA/hns: Fixed wrong judgments in the goto branch RDMA/hns: Remove the reserved WQE of SRQ
RDMA/hns: Refactor hns_roce_create_srq()
RDMA/hns: Refactor code about SRQ Context
RDMA/hns: Refactor hns_roce_v2_post_srq_recv() RDMA/hns: Add verification of QP type when post_recv

Quoted 3 lines

Xi Wang (2):
RDMA/hns: Refactor post recv flow
RDMA/hns: Clear remaining unused sges when post_recv

Applied to for-next, thanks

Jason