[PATCH for-next v2 0/3] RDMA/erdma: Add atomic operations support

[PATCH for-next v2 0/3] RDMA/erdma: Add atomic operations support

[Cheng Xu]

Hi,

This series introcuces atomic operations support for erdma driver:
- #1 extends access rights field in FRMR and REG MR commands to support
  IB_ACCESS_REMOTE_ATOMIC.
- #2 gets atomic capacity from hardware, and reports the cap to core.
- #3 implements the IO-path of atomic WR processing.

v2:
- Fix static analyzer check errors.

Thanks,
Cheng Xu

Cheng Xu (3):
  RDMA/erdma: Extend access right field of FRMR and REG MR to support
    atomic
  RDMA/erdma: Report atomic capacity when hardware supports atomic
    feature
  RDMA/erdma: Implement atomic operations support

 drivers/infiniband/hw/erdma/erdma.h       |  1 +
 drivers/infiniband/hw/erdma/erdma_cq.c    |  2 ++
 drivers/infiniband/hw/erdma/erdma_hw.h    | 29 +++++++++++----
 drivers/infiniband/hw/erdma/erdma_main.c  |  1 +
 drivers/infiniband/hw/erdma/erdma_qp.c    | 43 +++++++++++++++++++----
 drivers/infiniband/hw/erdma/erdma_verbs.c |  7 ++--
 drivers/infiniband/hw/erdma/erdma_verbs.h | 12 ++++---
 7 files changed, 74 insertions(+), 21 deletions(-)

-- 
2.27.0

[PATCH for-next v2 1/3] RDMA/erdma: Extend access right field of FRMR and REG MR to support atomic

[Cheng Xu]

To support atomic operations, IB_ACCESS_REMOTE_ATOMIC right should be
passed to hardware for permission check. Since "access mode" field in FRMR
SQE and RegMr command is never used by hw, we remove the "access mode"
field, so that we can then have enough space to extend access fields.

Signed-off-by: Cheng Xu <chengyou@linux.alibaba.com>
---
 drivers/infiniband/hw/erdma/erdma_hw.h    |  6 ++----
 drivers/infiniband/hw/erdma/erdma_qp.c    |  3 +--
 drivers/infiniband/hw/erdma/erdma_verbs.c |  3 +--
 drivers/infiniband/hw/erdma/erdma_verbs.h | 12 +++++++-----
 4 files changed, 11 insertions(+), 13 deletions(-)

diff --git a/drivers/infiniband/hw/erdma/erdma_hw.h b/drivers/infiniband/hw/erdma/erdma_hw.h
index e788887732e1..2a9a4c73d52c 100644
--- a/drivers/infiniband/hw/erdma/erdma_hw.h
+++ b/drivers/infiniband/hw/erdma/erdma_hw.h
@@ -224,8 +224,7 @@ struct erdma_cmdq_create_cq_req {
 /* regmr cfg1 */
 #define ERDMA_CMD_REGMR_PD_MASK GENMASK(31, 12)
 #define ERDMA_CMD_REGMR_TYPE_MASK GENMASK(7, 6)
-#define ERDMA_CMD_REGMR_RIGHT_MASK GENMASK(5, 2)
-#define ERDMA_CMD_REGMR_ACC_MODE_MASK GENMASK(1, 0)
+#define ERDMA_CMD_REGMR_RIGHT_MASK GENMASK(5, 1)
 
 /* regmr cfg2 */
 #define ERDMA_CMD_REGMR_PAGESIZE_MASK GENMASK(31, 27)
@@ -370,8 +369,7 @@ struct erdma_rqe {
 #define ERDMA_SQE_HDR_WQEBB_INDEX_MASK GENMASK_ULL(15, 0)
 
 /* REG MR attrs */
-#define ERDMA_SQE_MR_MODE_MASK GENMASK(1, 0)
-#define ERDMA_SQE_MR_ACCESS_MASK GENMASK(5, 2)
+#define ERDMA_SQE_MR_ACCESS_MASK GENMASK(5, 1)
 #define ERDMA_SQE_MR_MTT_TYPE_MASK GENMASK(7, 6)
 #define ERDMA_SQE_MR_MTT_CNT_MASK GENMASK(31, 12)
 
diff --git a/drivers/infiniband/hw/erdma/erdma_qp.c b/drivers/infiniband/hw/erdma/erdma_qp.c
index 5fe1a339a435..c7f343173cb9 100644
--- a/drivers/infiniband/hw/erdma/erdma_qp.c
+++ b/drivers/infiniband/hw/erdma/erdma_qp.c
@@ -397,8 +397,7 @@ static int erdma_push_one_sqe(struct erdma_qp *qp, u16 *pi,
 		regmr_sge->addr = cpu_to_le64(mr->ibmr.iova);
 		regmr_sge->length = cpu_to_le32(mr->ibmr.length);
 		regmr_sge->stag = cpu_to_le32(reg_wr(send_wr)->key);
-		attrs = FIELD_PREP(ERDMA_SQE_MR_MODE_MASK, 0) |
-			FIELD_PREP(ERDMA_SQE_MR_ACCESS_MASK, mr->access) |
+		attrs = FIELD_PREP(ERDMA_SQE_MR_ACCESS_MASK, mr->access) |
 			FIELD_PREP(ERDMA_SQE_MR_MTT_CNT_MASK,
 				   mr->mem.mtt_nents);
 
diff --git a/drivers/infiniband/hw/erdma/erdma_verbs.c b/drivers/infiniband/hw/erdma/erdma_verbs.c
index 62be98e2b941..f3bf87f17527 100644
--- a/drivers/infiniband/hw/erdma/erdma_verbs.c
+++ b/drivers/infiniband/hw/erdma/erdma_verbs.c
@@ -118,8 +118,7 @@ static int regmr_cmd(struct erdma_dev *dev, struct erdma_mr *mr)
 		   FIELD_PREP(ERDMA_CMD_MR_MPT_IDX_MASK, mr->ibmr.lkey >> 8);
 	req.cfg1 = FIELD_PREP(ERDMA_CMD_REGMR_PD_MASK, pd->pdn) |
 		   FIELD_PREP(ERDMA_CMD_REGMR_TYPE_MASK, mr->type) |
-		   FIELD_PREP(ERDMA_CMD_REGMR_RIGHT_MASK, mr->access) |
-		   FIELD_PREP(ERDMA_CMD_REGMR_ACC_MODE_MASK, 0);
+		   FIELD_PREP(ERDMA_CMD_REGMR_RIGHT_MASK, mr->access);
 	req.cfg2 = FIELD_PREP(ERDMA_CMD_REGMR_PAGESIZE_MASK,
 			      ilog2(mr->mem.page_size)) |
 		   FIELD_PREP(ERDMA_CMD_REGMR_MTT_TYPE_MASK, mr->mem.mtt_type) |
diff --git a/drivers/infiniband/hw/erdma/erdma_verbs.h b/drivers/infiniband/hw/erdma/erdma_verbs.h
index ab6380635e9e..a5574f0252bb 100644
--- a/drivers/infiniband/hw/erdma/erdma_verbs.h
+++ b/drivers/infiniband/hw/erdma/erdma_verbs.h
@@ -71,16 +71,18 @@ struct erdma_pd {
 #define ERDMA_MR_INLINE_MTT 0
 #define ERDMA_MR_INDIRECT_MTT 1
 
-#define ERDMA_MR_ACC_LR BIT(0)
-#define ERDMA_MR_ACC_LW BIT(1)
-#define ERDMA_MR_ACC_RR BIT(2)
-#define ERDMA_MR_ACC_RW BIT(3)
+#define ERDMA_MR_ACC_RA BIT(0)
+#define ERDMA_MR_ACC_LR BIT(1)
+#define ERDMA_MR_ACC_LW BIT(2)
+#define ERDMA_MR_ACC_RR BIT(3)
+#define ERDMA_MR_ACC_RW BIT(4)
 
 static inline u8 to_erdma_access_flags(int access)
 {
 	return (access & IB_ACCESS_REMOTE_READ ? ERDMA_MR_ACC_RR : 0) |
 	       (access & IB_ACCESS_LOCAL_WRITE ? ERDMA_MR_ACC_LW : 0) |
-	       (access & IB_ACCESS_REMOTE_WRITE ? ERDMA_MR_ACC_RW : 0);
+	       (access & IB_ACCESS_REMOTE_WRITE ? ERDMA_MR_ACC_RW : 0) |
+	       (access & IB_ACCESS_REMOTE_ATOMIC ? ERDMA_MR_ACC_RA : 0);
 }
 
 struct erdma_mem {
-- 
2.27.0

[PATCH for-next v2 2/3] RDMA/erdma: Report atomic capacity when hardware supports atomic feature

[Cheng Xu]

Introduce "capacity flags" field at where hardware put all zeros originally
in "query device" response. Using this field, hardware can report atomic
feature if supports.

Signed-off-by: Cheng Xu <chengyou@linux.alibaba.com>
---
 drivers/infiniband/hw/erdma/erdma.h       | 1 +
 drivers/infiniband/hw/erdma/erdma_hw.h    | 5 +++++
 drivers/infiniband/hw/erdma/erdma_main.c  | 1 +
 drivers/infiniband/hw/erdma/erdma_verbs.c | 4 ++++
 4 files changed, 11 insertions(+)

diff --git a/drivers/infiniband/hw/erdma/erdma.h b/drivers/infiniband/hw/erdma/erdma.h
index 730783fbc894..bb23d897c710 100644
--- a/drivers/infiniband/hw/erdma/erdma.h
+++ b/drivers/infiniband/hw/erdma/erdma.h
@@ -124,6 +124,7 @@ struct erdma_devattr {
 	u32 fw_version;
 
 	unsigned char peer_addr[ETH_ALEN];
+	unsigned long cap_flags;
 
 	int numa_node;
 	enum erdma_cc_alg cc;
diff --git a/drivers/infiniband/hw/erdma/erdma_hw.h b/drivers/infiniband/hw/erdma/erdma_hw.h
index 2a9a4c73d52c..808e7ee56d93 100644
--- a/drivers/infiniband/hw/erdma/erdma_hw.h
+++ b/drivers/infiniband/hw/erdma/erdma_hw.h
@@ -303,6 +303,7 @@ struct erdma_cmdq_destroy_qp_req {
 
 /* cap qword 0 definition */
 #define ERDMA_CMD_DEV_CAP_MAX_CQE_MASK GENMASK_ULL(47, 40)
+#define ERDMA_CMD_DEV_CAP_FLAGS_MASK GENMASK_ULL(31, 24)
 #define ERDMA_CMD_DEV_CAP_MAX_RECV_WR_MASK GENMASK_ULL(23, 16)
 #define ERDMA_CMD_DEV_CAP_MAX_MR_SIZE_MASK GENMASK_ULL(7, 0)
 
@@ -314,6 +315,10 @@ struct erdma_cmdq_destroy_qp_req {
 
 #define ERDMA_NQP_PER_QBLOCK 1024
 
+enum {
+	ERDMA_DEV_CAP_FLAGS_ATOMIC = 1 << 7,
+};
+
 #define ERDMA_CMD_INFO0_FW_VER_MASK GENMASK_ULL(31, 0)
 
 /* CQE hdr */
diff --git a/drivers/infiniband/hw/erdma/erdma_main.c b/drivers/infiniband/hw/erdma/erdma_main.c
index 49778bb294ae..e44b06fea595 100644
--- a/drivers/infiniband/hw/erdma/erdma_main.c
+++ b/drivers/infiniband/hw/erdma/erdma_main.c
@@ -374,6 +374,7 @@ static int erdma_dev_attrs_init(struct erdma_dev *dev)
 	dev->attrs.max_qp = ERDMA_NQP_PER_QBLOCK * ERDMA_GET_CAP(QBLOCK, cap1);
 	dev->attrs.max_mr = dev->attrs.max_qp << 1;
 	dev->attrs.max_cq = dev->attrs.max_qp << 1;
+	dev->attrs.cap_flags = ERDMA_GET_CAP(FLAGS, cap0);
 
 	dev->attrs.max_send_wr = ERDMA_MAX_SEND_WR;
 	dev->attrs.max_ord = ERDMA_MAX_ORD;
diff --git a/drivers/infiniband/hw/erdma/erdma_verbs.c b/drivers/infiniband/hw/erdma/erdma_verbs.c
index f3bf87f17527..d843ce1f35f3 100644
--- a/drivers/infiniband/hw/erdma/erdma_verbs.c
+++ b/drivers/infiniband/hw/erdma/erdma_verbs.c
@@ -288,6 +288,10 @@ int erdma_query_device(struct ib_device *ibdev, struct ib_device_attr *attr,
 	attr->max_mw = dev->attrs.max_mw;
 	attr->max_fast_reg_page_list_len = ERDMA_MAX_FRMR_PA;
 	attr->page_size_cap = ERDMA_PAGE_SIZE_SUPPORT;
+
+	if (dev->attrs.cap_flags & ERDMA_DEV_CAP_FLAGS_ATOMIC)
+		attr->atomic_cap = IB_ATOMIC_GLOB;
+
 	attr->fw_ver = dev->attrs.fw_version;
 
 	if (dev->netdev)
-- 
2.27.0

[PATCH for-next v2 3/3] RDMA/erdma: Implement atomic operations support

[Cheng Xu]

Add atomic operations support in post_send and poll_cq implementation.
Also, rename 'laddr' and 'lkey' in struct erdma_sge to 'addr' and 'key',
because this structure is used for both local and remote SGEs.

Signed-off-by: Cheng Xu <chengyou@linux.alibaba.com>
---
v1->v2:
- Fix static analyzer check errors.
---
 drivers/infiniband/hw/erdma/erdma_cq.c |  2 ++
 drivers/infiniband/hw/erdma/erdma_hw.h | 18 ++++++++++--
 drivers/infiniband/hw/erdma/erdma_qp.c | 40 ++++++++++++++++++++++----
 3 files changed, 52 insertions(+), 8 deletions(-)

diff --git a/drivers/infiniband/hw/erdma/erdma_cq.c b/drivers/infiniband/hw/erdma/erdma_cq.c
index 58e0dc5c75d1..cabd8678b355 100644
--- a/drivers/infiniband/hw/erdma/erdma_cq.c
+++ b/drivers/infiniband/hw/erdma/erdma_cq.c
@@ -64,6 +64,8 @@ static const enum ib_wc_opcode wc_mapping_table[ERDMA_NUM_OPCODES] = {
 	[ERDMA_OP_REG_MR] = IB_WC_REG_MR,
 	[ERDMA_OP_LOCAL_INV] = IB_WC_LOCAL_INV,
 	[ERDMA_OP_READ_WITH_INV] = IB_WC_RDMA_READ,
+	[ERDMA_OP_ATOMIC_CAS] = IB_WC_COMP_SWAP,
+	[ERDMA_OP_ATOMIC_FAD] = IB_WC_FETCH_ADD,
 };
 
 static const struct {
diff --git a/drivers/infiniband/hw/erdma/erdma_hw.h b/drivers/infiniband/hw/erdma/erdma_hw.h
index 808e7ee56d93..1b2e2b70678f 100644
--- a/drivers/infiniband/hw/erdma/erdma_hw.h
+++ b/drivers/infiniband/hw/erdma/erdma_hw.h
@@ -344,9 +344,9 @@ struct erdma_cqe {
 };
 
 struct erdma_sge {
-	__aligned_le64 laddr;
+	__aligned_le64 addr;
 	__le32 length;
-	__le32 lkey;
+	__le32 key;
 };
 
 /* Receive Queue Element */
@@ -413,6 +413,16 @@ struct erdma_readreq_sqe {
 	__le32 rsvd;
 };
 
+struct erdma_atomic_sqe {
+	__le64 hdr;
+	__le64 rsvd;
+	__le64 fetchadd_swap_data;
+	__le64 cmp_data;
+
+	struct erdma_sge remote;
+	struct erdma_sge sgl;
+};
+
 struct erdma_reg_mr_sqe {
 	__le64 hdr;
 	__le64 addr;
@@ -472,7 +482,9 @@ enum erdma_opcode {
 	ERDMA_OP_REG_MR = 14,
 	ERDMA_OP_LOCAL_INV = 15,
 	ERDMA_OP_READ_WITH_INV = 16,
-	ERDMA_NUM_OPCODES = 17,
+	ERDMA_OP_ATOMIC_CAS = 17,
+	ERDMA_OP_ATOMIC_FAD = 18,
+	ERDMA_NUM_OPCODES = 19,
 	ERDMA_OP_INVALID = ERDMA_NUM_OPCODES + 1
 };
 
diff --git a/drivers/infiniband/hw/erdma/erdma_qp.c b/drivers/infiniband/hw/erdma/erdma_qp.c
index c7f343173cb9..521e97258de7 100644
--- a/drivers/infiniband/hw/erdma/erdma_qp.c
+++ b/drivers/infiniband/hw/erdma/erdma_qp.c
@@ -285,15 +285,16 @@ static int erdma_push_one_sqe(struct erdma_qp *qp, u16 *pi,
 	u32 wqe_size, wqebb_cnt, hw_op, flags, sgl_offset;
 	u32 idx = *pi & (qp->attrs.sq_size - 1);
 	enum ib_wr_opcode op = send_wr->opcode;
+	struct erdma_atomic_sqe *atomic_sqe;
 	struct erdma_readreq_sqe *read_sqe;
 	struct erdma_reg_mr_sqe *regmr_sge;
 	struct erdma_write_sqe *write_sqe;
 	struct erdma_send_sqe *send_sqe;
 	struct ib_rdma_wr *rdma_wr;
-	struct erdma_mr *mr;
+	struct erdma_sge *sge;
 	__le32 *length_field;
+	struct erdma_mr *mr;
 	u64 wqe_hdr, *entry;
-	struct ib_sge *sge;
 	u32 attrs;
 	int ret;
 
@@ -360,9 +361,9 @@ static int erdma_push_one_sqe(struct erdma_qp *qp, u16 *pi,
 
 		sge = get_queue_entry(qp->kern_qp.sq_buf, idx + 1,
 				      qp->attrs.sq_size, SQEBB_SHIFT);
-		sge->addr = rdma_wr->remote_addr;
-		sge->lkey = rdma_wr->rkey;
-		sge->length = send_wr->sg_list[0].length;
+		sge->addr = cpu_to_le64(rdma_wr->remote_addr);
+		sge->key = cpu_to_le32(rdma_wr->rkey);
+		sge->length = cpu_to_le32(send_wr->sg_list[0].length);
 		wqe_size = sizeof(struct erdma_readreq_sqe) +
 			   send_wr->num_sge * sizeof(struct ib_sge);
 
@@ -423,6 +424,35 @@ static int erdma_push_one_sqe(struct erdma_qp *qp, u16 *pi,
 		regmr_sge->stag = cpu_to_le32(send_wr->ex.invalidate_rkey);
 		wqe_size = sizeof(struct erdma_reg_mr_sqe);
 		goto out;
+	case IB_WR_ATOMIC_CMP_AND_SWP:
+	case IB_WR_ATOMIC_FETCH_AND_ADD:
+		atomic_sqe = (struct erdma_atomic_sqe *)entry;
+		if (op == IB_WR_ATOMIC_CMP_AND_SWP) {
+			wqe_hdr |= FIELD_PREP(ERDMA_SQE_HDR_OPCODE_MASK,
+					      ERDMA_OP_ATOMIC_CAS);
+			atomic_sqe->fetchadd_swap_data =
+				cpu_to_le64(atomic_wr(send_wr)->swap);
+			atomic_sqe->cmp_data =
+				cpu_to_le64(atomic_wr(send_wr)->compare_add);
+		} else {
+			wqe_hdr |= FIELD_PREP(ERDMA_SQE_HDR_OPCODE_MASK,
+					      ERDMA_OP_ATOMIC_FAD);
+			atomic_sqe->fetchadd_swap_data =
+				cpu_to_le64(atomic_wr(send_wr)->compare_add);
+		}
+
+		sge = get_queue_entry(qp->kern_qp.sq_buf, idx + 1,
+				      qp->attrs.sq_size, SQEBB_SHIFT);
+		sge->addr = cpu_to_le64(atomic_wr(send_wr)->remote_addr);
+		sge->key = cpu_to_le32(atomic_wr(send_wr)->rkey);
+		sge++;
+
+		sge->addr = cpu_to_le64(send_wr->sg_list[0].addr);
+		sge->key = cpu_to_le32(send_wr->sg_list[0].lkey);
+		sge->length = cpu_to_le32(send_wr->sg_list[0].length);
+
+		wqe_size = sizeof(*atomic_sqe);
+		goto out;
 	default:
 		return -EOPNOTSUPP;
 	}
-- 
2.27.0

Re: [PATCH for-next v2 0/3] RDMA/erdma: Add atomic operations support

[Leon Romanovsky]

On Mon, 7 Nov 2022 10:18:42 +0800, Cheng Xu wrote:
> This series introcuces atomic operations support for erdma driver:
> - #1 extends access rights field in FRMR and REG MR commands to support
>   IB_ACCESS_REMOTE_ATOMIC.
> - #2 gets atomic capacity from hardware, and reports the cap to core.
> - #3 implements the IO-path of atomic WR processing.
> 
> v2:
> - Fix static analyzer check errors.
> 
> [...]

Applied, thanks!

[1/3] RDMA/erdma: Extend access right field of FRMR and REG MR to support atomic
      https://git.kernel.org/rdma/rdma/c/ece43fad220ba0
[2/3] RDMA/erdma: Report atomic capacity when hardware supports atomic feature
      https://git.kernel.org/rdma/rdma/c/71c6925f280ae8
[3/3] RDMA/erdma: Implement atomic operations support
      https://git.kernel.org/rdma/rdma/c/0ca9c2e2844aa2

Best regards,
-- 
Leon Romanovsky <leon@kernel.org>