packages/rdma-core
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Bugfix for sge num and support inline feature

Browse Source 
Fix sge num bug, add compatibility for rq inline, support cqe inline

Signed-off-by: Luoyouming <luoyouming@huawei.com>
This commit is contained in:
Luoyouming 2022-10-31 14:17:19 +08:00 committed by Chengchang Tang
parent 1a0fe62ace
commit 0759ff83eb
11 changed files with 1138 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

45
0001-libhns-Use-a-constant-instead-of-sizeof-operation.patch Normal file
View File

@ -0,0 +1,45 @@
From 0de1678211e710c2cd33e3aea98b1271cae9bd98 Mon Sep 17 00:00:00 2001
From: Luoyouming <luoyouming@huawei.com>
Date: Tue, 20 Sep 2022 11:47:45 +0800
Subject: [PATCH v4 01/10] libhns: Use a constant instead of sizeof operation
The sge size is known to be constant, so it's unnecessary to use sizeof to
calculate.
Signed-off-by: Luoyouming <luoyouming@huawei.com>
Reviewed-by: Yangyang Li <liyangyang20@huawei.com>
---
providers/hns/hns_roce_u_hw_v2.c | 5 ++---
1 file changed, 2 insertions(+), 3 deletions(-)
diff --git a/providers/hns/hns_roce_u_hw_v2.c b/providers/hns/hns_roce_u_hw_v2.c
index 0169250..d9ea18e 100644
--- a/providers/hns/hns_roce_u_hw_v2.c
+++ b/providers/hns/hns_roce_u_hw_v2.c
@@ -847,13 +847,12 @@ static int fill_ext_sge_inl_data(struct hns_roce_qp *qp,
uint32_t num_buf,
enum hns_roce_wr_buf_type buf_type)
{
- unsigned int sge_sz = sizeof(struct hns_roce_v2_wqe_data_seg);
unsigned int sge_mask = qp->ex_sge.sge_cnt - 1;
void *dst_addr, *src_addr, *tail_bound_addr;
uint32_t src_len, tail_len;
int i;
- if (sge_info->total_len > qp->sq.max_gs * sge_sz)
+ if (sge_info->total_len > qp->sq.max_gs * HNS_ROCE_SGE_SIZE)
return EINVAL;
dst_addr = get_send_sge_ex(qp, sge_info->start_idx & sge_mask);
@@ -880,7 +879,7 @@ static int fill_ext_sge_inl_data(struct hns_roce_qp *qp,
}
}
- sge_info->valid_num = DIV_ROUND_UP(sge_info->total_len, sge_sz);
+ sge_info->valid_num = DIV_ROUND_UP(sge_info->total_len, HNS_ROCE_SGE_SIZE);
sge_info->start_idx += sge_info->valid_num;
return 0;
--
2.30.0

53
0002-libhns-Fix-ext_sge-num-error-when-post-send.patch Normal file
View File

@ -0,0 +1,53 @@
From a57d5dfbc2701b9d0c47eb70a1bb82b16170a7d2 Mon Sep 17 00:00:00 2001
From: Luoyouming <luoyouming@huawei.com>
Date: Tue, 20 Sep 2022 11:53:18 +0800
Subject: [PATCH v4 02/10] libhns: Fix ext_sge num error when post send
The max_gs is the sum of extended sge and standard sge. In function
fill_ext_sge_inl_data, max_gs does not subtract the number of extended
sges, but is directly used to calculate the size of extended sges.
Fixes:b7814b7b9715("libhns: Support inline data in extented sge space for RC")
Signed-off-by: Luoyouming <luoyouming@huawei.com>
Reviewed-by: Yangyang Li <liyangyang20@huawei.com>
---
providers/hns/hns_roce_u_hw_v2.c | 13 ++++++++++++-
1 file changed, 12 insertions(+), 1 deletion(-)
diff --git a/providers/hns/hns_roce_u_hw_v2.c b/providers/hns/hns_roce_u_hw_v2.c
index d9ea18e..bb4298f 100644
--- a/providers/hns/hns_roce_u_hw_v2.c
+++ b/providers/hns/hns_roce_u_hw_v2.c
@@ -841,6 +841,14 @@ static void get_src_buf_info(void **src_addr, uint32_t *src_len,
}
}
+static unsigned int get_std_sge_num(struct hns_roce_qp *qp)
+{
+ if (qp->verbs_qp.qp.qp_type == IBV_QPT_UD)
+ return 0;
+
+ return HNS_ROCE_SGE_IN_WQE;
+}
+
static int fill_ext_sge_inl_data(struct hns_roce_qp *qp,
struct hns_roce_sge_info *sge_info,
const void *buf_list,
@@ -850,9 +858,12 @@ static int fill_ext_sge_inl_data(struct hns_roce_qp *qp,
unsigned int sge_mask = qp->ex_sge.sge_cnt - 1;
void *dst_addr, *src_addr, *tail_bound_addr;
uint32_t src_len, tail_len;
+ unsigned int std_sge_num;
int i;
- if (sge_info->total_len > qp->sq.max_gs * HNS_ROCE_SGE_SIZE)
+ std_sge_num = get_std_sge_num(qp);
+ if (sge_info->total_len >
+ (qp->sq.max_gs - std_sge_num) * HNS_ROCE_SGE_SIZE)
return EINVAL;
dst_addr = get_send_sge_ex(qp, sge_info->start_idx & sge_mask);
--
2.30.0

46
0003-Update-kernel-headers.patch Normal file
View File

@ -0,0 +1,46 @@
From f20816984da80e2fe9a82b3b330f85150763243e Mon Sep 17 00:00:00 2001
From: Luoyouming <luoyouming@huawei.com>
Date: Mon, 26 Sep 2022 11:14:05 +0800
Subject: [PATCH v4 03/10] Update kernel headers
To commit ?? ("RDMA/hns: Fix the problem of sge nums").
Signed-off-by: Luoyouming <luoyouming@huawei.com>
---
kernel-headers/rdma/hns-abi.h | 15 +++++++++++++++
1 file changed, 15 insertions(+)
diff --git a/kernel-headers/rdma/hns-abi.h b/kernel-headers/rdma/hns-abi.h
index f6fde06..fc83dfb 100644
--- a/kernel-headers/rdma/hns-abi.h
+++ b/kernel-headers/rdma/hns-abi.h
@@ -85,11 +85,26 @@ struct hns_roce_ib_create_qp_resp {
__aligned_u64 dwqe_mmap_key;
};
+enum {
+ HNS_ROCE_EXSGE_FLAGS = 1 << 0,
+};
+
+enum {
+ HNS_ROCE_RSP_EXSGE_FLAGS = 1 << 0,
+};
+
struct hns_roce_ib_alloc_ucontext_resp {
__u32 qp_tab_size;
__u32 cqe_size;
__u32 srq_tab_size;
__u32 reserved;
+ __u32 config;
+ __u32 max_inline_data;
+};
+
+struct hns_roce_ib_alloc_ucontext {
+ __u32 config;
+ __u32 reserved;
};
struct hns_roce_ib_alloc_pd_resp {
--
2.30.0

254
0004-libhns-Fix-the-problem-of-sge-nums.patch Normal file
View File

@ -0,0 +1,254 @@
From 448d82b2c62f09f1dd9c8045d34623dedef1c111 Mon Sep 17 00:00:00 2001
From: Luoyouming <luoyouming@huawei.com>
Date: Fri, 19 Nov 2021 20:21:21 +0800
Subject: [PATCH v4 04/10] libhns: Fix the problem of sge nums
Currently, the driver only uses max_send_sge to initialize sge num
when creating_qp. So, in the sq inline scenario, the driver may not
has enough sge to send data. For example, if max_send_sge is 16 and
max_inline_data is 1024, the driver needs 1024/16=64 sge to send data.
Therefore, the calculation method of sge num is modified to take the
maximum value of max_send_sge and max_inline_data/16 to solve this
problem.
Fixes:11c81d0e3a98("libhns: Refactor process of setting extended sge")
Fixes:b7814b7b9715("libhns: Support inline data in extented sge space for RC")
Signed-off-by: Luoyouming <luoyouming@huawei.com>
Reviewed-by: Yangyang Li <liyangyang20@huawei.com>
---
providers/hns/hns_roce_u.c | 9 +++-
providers/hns/hns_roce_u.h | 3 ++
providers/hns/hns_roce_u_abi.h | 2 +-
providers/hns/hns_roce_u_hw_v2.c | 13 +----
providers/hns/hns_roce_u_verbs.c | 84 ++++++++++++++++++++++++--------
5 files changed, 77 insertions(+), 34 deletions(-)
diff --git a/providers/hns/hns_roce_u.c b/providers/hns/hns_roce_u.c
index a46ceb9..1bd5bb1 100644
--- a/providers/hns/hns_roce_u.c
+++ b/providers/hns/hns_roce_u.c
@@ -103,9 +103,9 @@ static struct verbs_context *hns_roce_alloc_context(struct ibv_device *ibdev,
{
struct hns_roce_device *hr_dev = to_hr_dev(ibdev);
struct hns_roce_alloc_ucontext_resp resp = {};
+ struct hns_roce_alloc_ucontext cmd = {};
struct ibv_device_attr dev_attrs;
struct hns_roce_context *context;
- struct ibv_get_context cmd;
int i;
context = verbs_init_and_alloc_context(ibdev, cmd_fd, context, ibv_ctx,
@@ -113,7 +113,8 @@ static struct verbs_context *hns_roce_alloc_context(struct ibv_device *ibdev,
if (!context)
return NULL;
- if (ibv_cmd_get_context(&context->ibv_ctx, &cmd, sizeof(cmd),
+ cmd.config |= HNS_ROCE_EXSGE_FLAGS;
+ if (ibv_cmd_get_context(&context->ibv_ctx, &cmd.ibv_cmd, sizeof(cmd),
&resp.ibv_resp, sizeof(resp)))
goto err_free;
@@ -124,6 +125,10 @@ static struct verbs_context *hns_roce_alloc_context(struct ibv_device *ibdev,
else
context->cqe_size = HNS_ROCE_V3_CQE_SIZE;
+ context->config = resp.config;
+ if (resp.config & HNS_ROCE_RSP_EXSGE_FLAGS)
+ context->max_inline_data = resp.max_inline_data;
+
context->qp_table_shift = calc_table_shift(resp.qp_tab_size,
HNS_ROCE_QP_TABLE_BITS);
context->qp_table_mask = (1 << context->qp_table_shift) - 1;
diff --git a/providers/hns/hns_roce_u.h b/providers/hns/hns_roce_u.h
index 5d90634..5388f9c 100644
--- a/providers/hns/hns_roce_u.h
+++ b/providers/hns/hns_roce_u.h
@@ -213,6 +213,8 @@ struct hns_roce_context {
unsigned int max_srq_sge;
int max_cqe;
unsigned int cqe_size;
+ uint32_t config;
+ unsigned int max_inline_data;
};
struct hns_roce_pd {
@@ -267,6 +269,7 @@ struct hns_roce_wq {
unsigned int head;
unsigned int tail;
unsigned int max_gs;
+ unsigned int ext_sge_cnt;
unsigned int rsv_sge;
unsigned int wqe_shift;
unsigned int shift; /* wq size is 2^shift */
diff --git a/providers/hns/hns_roce_u_abi.h b/providers/hns/hns_roce_u_abi.h
index 333f977..2753d30 100644
--- a/providers/hns/hns_roce_u_abi.h
+++ b/providers/hns/hns_roce_u_abi.h
@@ -47,7 +47,7 @@ DECLARE_DRV_CMD(hns_roce_create_cq_ex, IB_USER_VERBS_EX_CMD_CREATE_CQ,
hns_roce_ib_create_cq, hns_roce_ib_create_cq_resp);
DECLARE_DRV_CMD(hns_roce_alloc_ucontext, IB_USER_VERBS_CMD_GET_CONTEXT,
- empty, hns_roce_ib_alloc_ucontext_resp);
+ hns_roce_ib_alloc_ucontext, hns_roce_ib_alloc_ucontext_resp);
DECLARE_DRV_CMD(hns_roce_create_qp, IB_USER_VERBS_CMD_CREATE_QP,
hns_roce_ib_create_qp, hns_roce_ib_create_qp_resp);
diff --git a/providers/hns/hns_roce_u_hw_v2.c b/providers/hns/hns_roce_u_hw_v2.c
index bb4298f..ebe68bc 100644
--- a/providers/hns/hns_roce_u_hw_v2.c
+++ b/providers/hns/hns_roce_u_hw_v2.c
@@ -841,14 +841,6 @@ static void get_src_buf_info(void **src_addr, uint32_t *src_len,
}
}
-static unsigned int get_std_sge_num(struct hns_roce_qp *qp)
-{
- if (qp->verbs_qp.qp.qp_type == IBV_QPT_UD)
- return 0;
-
- return HNS_ROCE_SGE_IN_WQE;
-}
-
static int fill_ext_sge_inl_data(struct hns_roce_qp *qp,
struct hns_roce_sge_info *sge_info,
const void *buf_list,
@@ -858,12 +850,9 @@ static int fill_ext_sge_inl_data(struct hns_roce_qp *qp,
unsigned int sge_mask = qp->ex_sge.sge_cnt - 1;
void *dst_addr, *src_addr, *tail_bound_addr;
uint32_t src_len, tail_len;
- unsigned int std_sge_num;
int i;
- std_sge_num = get_std_sge_num(qp);
- if (sge_info->total_len >
- (qp->sq.max_gs - std_sge_num) * HNS_ROCE_SGE_SIZE)
+ if (sge_info->total_len > qp->sq.ext_sge_cnt * HNS_ROCE_SGE_SIZE)
return EINVAL;
dst_addr = get_send_sge_ex(qp, sge_info->start_idx & sge_mask);
diff --git a/providers/hns/hns_roce_u_verbs.c b/providers/hns/hns_roce_u_verbs.c
index ba7f2ae..851b145 100644
--- a/providers/hns/hns_roce_u_verbs.c
+++ b/providers/hns/hns_roce_u_verbs.c
@@ -978,41 +978,88 @@ err_alloc:
return -ENOMEM;
}
-static unsigned int get_wqe_ext_sge_cnt(struct hns_roce_qp *qp)
+/**
+ * Calculated sge num according to attr's max_send_sge
+ */
+static unsigned int get_sge_num_from_max_send_sge(bool is_ud,
+ uint32_t max_send_sge)
{
- if (qp->verbs_qp.qp.qp_type == IBV_QPT_UD)
- return qp->sq.max_gs;
+ unsigned int std_sge_num;
+ unsigned int min_sge;
- if (qp->sq.max_gs > HNS_ROCE_SGE_IN_WQE)
- return qp->sq.max_gs - HNS_ROCE_SGE_IN_WQE;
+ std_sge_num = is_ud ? 0 : HNS_ROCE_SGE_IN_WQE;
+ min_sge = is_ud ? 1 : 0;
+ return max_send_sge > std_sge_num ? (max_send_sge - std_sge_num) :
+ min_sge;
+}
- return 0;
+/**
+ * Calculated sge num according to attr's max_inline_data
+ */
+static unsigned int get_sge_num_from_max_inl_data(bool is_ud,
+ uint32_t max_inline_data)
+{
+ unsigned int inline_sge = 0;
+
+ inline_sge = max_inline_data / HNS_ROCE_SGE_SIZE;
+ /*
+ * if max_inline_data less than
+ * HNS_ROCE_SGE_IN_WQE * HNS_ROCE_SGE_SIZE,
+ * In addition to ud's mode, no need to extend sge.
+ */
+ if (!is_ud && (inline_sge <= HNS_ROCE_SGE_IN_WQE))
+ inline_sge = 0;
+
+ return inline_sge;
}
-static void set_ext_sge_param(struct hns_roce_device *hr_dev,
+static void set_ext_sge_param(struct hns_roce_context *ctx,
struct ibv_qp_init_attr_ex *attr,
struct hns_roce_qp *qp, unsigned int wr_cnt)
{
+ bool is_ud = (qp->verbs_qp.qp.qp_type == IBV_QPT_UD);
+ unsigned int ext_wqe_sge_cnt;
+ unsigned int inline_ext_sge;
unsigned int total_sge_cnt;
- unsigned int wqe_sge_cnt;
+ unsigned int std_sge_num;
qp->ex_sge.sge_shift = HNS_ROCE_SGE_SHIFT;
-
- qp->sq.max_gs = attr->cap.max_send_sge;
-
- wqe_sge_cnt = get_wqe_ext_sge_cnt(qp);
+ std_sge_num = is_ud ? 0 : HNS_ROCE_SGE_IN_WQE;
+ ext_wqe_sge_cnt = get_sge_num_from_max_send_sge(is_ud,
+ attr->cap.max_send_sge);
+
+ if (ctx->config & HNS_ROCE_RSP_EXSGE_FLAGS) {
+ attr->cap.max_inline_data = min_t(uint32_t, roundup_pow_of_two(
+ attr->cap.max_inline_data),
+ ctx->max_inline_data);
+
+ inline_ext_sge = max(ext_wqe_sge_cnt,
+ get_sge_num_from_max_inl_data(is_ud,
+ attr->cap.max_inline_data));
+ qp->sq.ext_sge_cnt = inline_ext_sge ?
+ roundup_pow_of_two(inline_ext_sge) : 0;
+ qp->sq.max_gs = min((qp->sq.ext_sge_cnt + std_sge_num),
+ ctx->max_sge);
+
+ ext_wqe_sge_cnt = qp->sq.ext_sge_cnt;
+ } else {
+ qp->sq.max_gs = max(1U, attr->cap.max_send_sge);
+ qp->sq.max_gs = min(qp->sq.max_gs, ctx->max_sge);
+ qp->sq.ext_sge_cnt = qp->sq.max_gs;
+ }
/* If the number of extended sge is not zero, they MUST use the
* space of HNS_HW_PAGE_SIZE at least.
*/
- if (wqe_sge_cnt) {
- total_sge_cnt = roundup_pow_of_two(wr_cnt * wqe_sge_cnt);
- qp->ex_sge.sge_cnt =
- max(total_sge_cnt,
- (unsigned int)HNS_HW_PAGE_SIZE / HNS_ROCE_SGE_SIZE);
+ if (ext_wqe_sge_cnt) {
+ total_sge_cnt = roundup_pow_of_two(wr_cnt * ext_wqe_sge_cnt);
+ qp->ex_sge.sge_cnt = max(total_sge_cnt,
+ (unsigned int)HNS_HW_PAGE_SIZE /
+ HNS_ROCE_SGE_SIZE);
}
}
+
static void hns_roce_set_qp_params(struct ibv_qp_init_attr_ex *attr,
struct hns_roce_qp *qp,
struct hns_roce_context *ctx)
@@ -1044,10 +1091,9 @@ static void hns_roce_set_qp_params(struct ibv_qp_init_attr_ex *attr,
qp->sq.wqe_cnt = cnt;
qp->sq.shift = hr_ilog32(cnt);
- set_ext_sge_param(hr_dev, attr, qp, cnt);
+ set_ext_sge_param(ctx, attr, qp, cnt);
qp->sq.max_post = min(ctx->max_qp_wr, cnt);
- qp->sq.max_gs = min(ctx->max_sge, qp->sq.max_gs);
qp->sq_signal_bits = attr->sq_sig_all ? 0 : 1;
--
2.30.0

33
0005-Update-kernel-headers.patch Normal file
View File

@ -0,0 +1,33 @@
From 542b54285dbaebbe0b5eb3279134b02484d7329d Mon Sep 17 00:00:00 2001
From: Luoyouming <luoyouming@huawei.com>
Date: Mon, 10 Oct 2022 21:21:55 +0800
Subject: [PATCH v4 05/10] Update kernel headers
To commit ?? ("RDMA/hns: Remove enable rq inline in kernel and add
compatibility handling").
Signed-off-by: Luoyouming <luoyouming@huawei.com>
---
kernel-headers/rdma/hns-abi.h | 2 ++
1 file changed, 2 insertions(+)
diff --git a/kernel-headers/rdma/hns-abi.h b/kernel-headers/rdma/hns-abi.h
index fc83dfb..c70465d 100644
--- a/kernel-headers/rdma/hns-abi.h
+++ b/kernel-headers/rdma/hns-abi.h
@@ -87,10 +87,12 @@ struct hns_roce_ib_create_qp_resp {
enum {
HNS_ROCE_EXSGE_FLAGS = 1 << 0,
+ HNS_ROCE_RQ_INLINE_FLAGS = 1 << 1,
};
enum {
HNS_ROCE_RSP_EXSGE_FLAGS = 1 << 0,
+ HNS_ROCE_RSP_RQ_INLINE_FLAGS = 1 << 1,
};
struct hns_roce_ib_alloc_ucontext_resp {
--
2.30.0

46
0006-libhns-Add-compatibility-handling-for-rq-inline.patch Normal file
View File

@ -0,0 +1,46 @@
From 996bca51e2063dc790286cbc894e2c438f499441 Mon Sep 17 00:00:00 2001
From: Luoyouming <luoyouming@huawei.com>
Date: Mon, 10 Oct 2022 21:49:35 +0800
Subject: [PATCH v4 06/10] libhns: Add compatibility handling for rq inline
Add compatibility processing between different user space
and kernel space.
Signed-off-by: Luoyouming <luoyouming@huawei.com>
Reviewed-by: Yangyang Li <liyangyang20@huawei.com>
---
providers/hns/hns_roce_u.c | 2 +-
providers/hns/hns_roce_u_verbs.c | 4 +++-
2 files changed, 4 insertions(+), 2 deletions(-)
diff --git a/providers/hns/hns_roce_u.c b/providers/hns/hns_roce_u.c
index 1bd5bb1..6c9aefa 100644
--- a/providers/hns/hns_roce_u.c
+++ b/providers/hns/hns_roce_u.c
@@ -113,7 +113,7 @@ static struct verbs_context *hns_roce_alloc_context(struct ibv_device *ibdev,
if (!context)
return NULL;
- cmd.config |= HNS_ROCE_EXSGE_FLAGS;
+ cmd.config |= HNS_ROCE_EXSGE_FLAGS | HNS_ROCE_RQ_INLINE_FLAGS;
if (ibv_cmd_get_context(&context->ibv_ctx, &cmd.ibv_cmd, sizeof(cmd),
&resp.ibv_resp, sizeof(resp)))
goto err_free;
diff --git a/providers/hns/hns_roce_u_verbs.c b/providers/hns/hns_roce_u_verbs.c
index 851b145..3e9a306 100644
--- a/providers/hns/hns_roce_u_verbs.c
+++ b/providers/hns/hns_roce_u_verbs.c
@@ -1079,7 +1079,9 @@ static void hns_roce_set_qp_params(struct ibv_qp_init_attr_ex *attr,
cnt = roundup_pow_of_two(attr->cap.max_recv_wr);
qp->rq.wqe_cnt = cnt;
qp->rq.shift = hr_ilog32(cnt);
- qp->rq_rinl_buf.wqe_cnt = cnt;
+ qp->rq_rinl_buf.wqe_cnt = 0;
+ if (ctx->config & HNS_ROCE_RSP_RQ_INLINE_FLAGS)
+ qp->rq_rinl_buf.wqe_cnt = cnt;
attr->cap.max_recv_wr = qp->rq.wqe_cnt;
attr->cap.max_recv_sge = qp->rq.max_gs;
--
2.30.0

301
0007-libhns-Refactor-rq-inline.patch Normal file
View File

@ -0,0 +1,301 @@
From 9e5f5d39757a5479a1a4e1170978d2e09acb995b Mon Sep 17 00:00:00 2001
From: Luoyouming <luoyouming@huawei.com>
Date: Fri, 9 Sep 2022 17:42:38 +0800
Subject: [PATCH v4 07/10] libhns: Refactor rq inline
The ibv_sge struct is enough, there is no need to customize the
hns_roce_rinl_sge struct. Refactored structures and functions
for reuse in cqe inline (rq, srq scenarios).
Signed-off-by: Luoyouming <luoyouming@huawei.com>
---
providers/hns/hns_roce_u.h | 7 +--
providers/hns/hns_roce_u_hw_v2.c | 103 +++++++++++++++----------------
providers/hns/hns_roce_u_verbs.c | 46 +++++++-------
3 files changed, 75 insertions(+), 81 deletions(-)
diff --git a/providers/hns/hns_roce_u.h b/providers/hns/hns_roce_u.h
index 5388f9c..57ebe55 100644
--- a/providers/hns/hns_roce_u.h
+++ b/providers/hns/hns_roce_u.h
@@ -290,13 +290,8 @@ struct hns_roce_sge_ex {
unsigned int sge_shift;
};
-struct hns_roce_rinl_sge {
- void *addr;
- unsigned int len;
-};
-
struct hns_roce_rinl_wqe {
- struct hns_roce_rinl_sge *sg_list;
+ struct ibv_sge *sg_list;
unsigned int sge_cnt;
};
diff --git a/providers/hns/hns_roce_u_hw_v2.c b/providers/hns/hns_roce_u_hw_v2.c
index ebe68bc..73acc9e 100644
--- a/providers/hns/hns_roce_u_hw_v2.c
+++ b/providers/hns/hns_roce_u_hw_v2.c
@@ -34,6 +34,7 @@
#include <stdio.h>
#include <string.h>
#include <sys/mman.h>
+#include <ccan/minmax.h>
#include "hns_roce_u.h"
#include "hns_roce_u_db.h"
#include "hns_roce_u_hw_v2.h"
@@ -417,46 +418,42 @@ static void get_opcode_for_resp(struct hns_roce_v2_cqe *cqe, struct ibv_wc *wc,
wc->opcode = wc_rcv_op_map[opcode];
}
-static int handle_recv_inl_wqe(struct hns_roce_v2_cqe *cqe, struct ibv_wc *wc,
- struct hns_roce_qp **cur_qp, uint32_t opcode)
+static void handle_recv_inl_data(struct hns_roce_v2_cqe *cqe,
+ struct hns_roce_rinl_buf *rinl_buf,
+ uint32_t wr_cnt, uint8_t *buf)
{
- if (((*cur_qp)->verbs_qp.qp.qp_type == IBV_QPT_RC) &&
- (opcode == HNS_ROCE_RECV_OP_SEND ||
- opcode == HNS_ROCE_RECV_OP_SEND_WITH_IMM ||
- opcode == HNS_ROCE_RECV_OP_SEND_WITH_INV) &&
- hr_reg_read(cqe, CQE_RQ_INLINE)) {
- struct hns_roce_rinl_sge *sge_list;
- uint32_t wr_num, wr_cnt, sge_num, data_len;
- uint8_t *wqe_buf;
- uint32_t sge_cnt, size;
+ struct ibv_sge *sge_list;
+ uint32_t sge_num, data_len;
+ uint32_t sge_cnt, size;
- wr_num = hr_reg_read(cqe, CQE_WQE_IDX);
- wr_cnt = wr_num & ((*cur_qp)->rq.wqe_cnt - 1);
+ sge_list = rinl_buf->wqe_list[wr_cnt].sg_list;
+ sge_num = rinl_buf->wqe_list[wr_cnt].sge_cnt;
- sge_list = (*cur_qp)->rq_rinl_buf.wqe_list[wr_cnt].sg_list;
- sge_num = (*cur_qp)->rq_rinl_buf.wqe_list[wr_cnt].sge_cnt;
- wqe_buf = (uint8_t *)get_recv_wqe_v2(*cur_qp, wr_cnt);
+ data_len = le32toh(cqe->byte_cnt);
- data_len = wc->byte_len;
+ for (sge_cnt = 0; (sge_cnt < sge_num) && (data_len); sge_cnt++) {
+ size = min(sge_list[sge_cnt].length, data_len);
- for (sge_cnt = 0; (sge_cnt < sge_num) && (data_len);
- sge_cnt++) {
- size = sge_list[sge_cnt].len < data_len ?
- sge_list[sge_cnt].len : data_len;
+ memcpy((void *)(uintptr_t)sge_list[sge_cnt].addr, (void *)buf, size);
+ data_len -= size;
+ buf += size;
+ }
- memcpy((void *)sge_list[sge_cnt].addr,
- (void *)wqe_buf, size);
- data_len -= size;
- wqe_buf += size;
- }
+ if (data_len)
+ hr_reg_write(cqe, CQE_STATUS, HNS_ROCE_V2_CQE_LOCAL_LENGTH_ERR);
- if (data_len) {
- wc->status = IBV_WC_LOC_LEN_ERR;
- return V2_CQ_POLL_ERR;
- }
- }
+}
- return V2_CQ_OK;
+static void handle_recv_rq_inl(struct hns_roce_v2_cqe *cqe,
+ struct hns_roce_qp *cur_qp)
+{
+ uint8_t *wqe_buf;
+ uint32_t wr_num;
+
+ wr_num = hr_reg_read(cqe, CQE_WQE_IDX) & (cur_qp->rq.wqe_cnt - 1);
+
+ wqe_buf = (uint8_t *)get_recv_wqe_v2(cur_qp, wr_num);
+ handle_recv_inl_data(cqe, &(cur_qp->rq_rinl_buf), wr_num, wqe_buf);
}
static void parse_for_ud_qp(struct hns_roce_v2_cqe *cqe, struct ibv_wc *wc)
@@ -479,10 +476,9 @@ static void parse_cqe_for_srq(struct hns_roce_v2_cqe *cqe, struct ibv_wc *wc,
}
static int parse_cqe_for_resp(struct hns_roce_v2_cqe *cqe, struct ibv_wc *wc,
- struct hns_roce_qp *hr_qp, uint8_t opcode)
+ struct hns_roce_qp *hr_qp)
{
struct hns_roce_wq *wq;
- int ret;
wq = &hr_qp->rq;
wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
@@ -491,12 +487,8 @@ static int parse_cqe_for_resp(struct hns_roce_v2_cqe *cqe, struct ibv_wc *wc,
if (hr_qp->verbs_qp.qp.qp_type == IBV_QPT_UD)
parse_for_ud_qp(cqe, wc);
- ret = handle_recv_inl_wqe(cqe, wc, &hr_qp, opcode);
- if (ret) {
- verbs_err(verbs_get_ctx(hr_qp->verbs_qp.qp.context),
- PFX "failed to handle recv inline wqe!\n");
- return ret;
- }
+ if (hr_reg_read(cqe, CQE_RQ_INLINE))
+ handle_recv_rq_inl(cqe, hr_qp);
return 0;
}
@@ -626,7 +618,7 @@ static int parse_cqe_for_cq(struct hns_roce_context *ctx, struct hns_roce_cq *cq
if (srq)
parse_cqe_for_srq(cqe, wc, srq);
else
- parse_cqe_for_resp(cqe, wc, cur_qp, opcode);
+ parse_cqe_for_resp(cqe, wc, cur_qp);
}
return 0;
@@ -1355,26 +1347,31 @@ static void fill_recv_sge_to_wqe(struct ibv_recv_wr *wr, void *wqe,
}
}
+static void fill_recv_inl_buf(struct hns_roce_rinl_buf *rinl_buf,
+ unsigned int wqe_idx, struct ibv_recv_wr *wr)
+{
+ struct ibv_sge *sge_list;
+ unsigned int i;
+
+ if (!rinl_buf->wqe_cnt)
+ return;
+
+ sge_list = rinl_buf->wqe_list[wqe_idx].sg_list;
+ rinl_buf->wqe_list[wqe_idx].sge_cnt = (unsigned int)wr->num_sge;
+ for (i = 0; i < wr->num_sge; i++)
+ memcpy((void *)&sge_list[i], (void *)&wr->sg_list[i],
+ sizeof(struct ibv_sge));
+}
+
static void fill_rq_wqe(struct hns_roce_qp *qp, struct ibv_recv_wr *wr,
unsigned int wqe_idx, unsigned int max_sge)
{
- struct hns_roce_rinl_sge *sge_list;
- unsigned int i;
void *wqe;
wqe = get_recv_wqe_v2(qp, wqe_idx);
fill_recv_sge_to_wqe(wr, wqe, max_sge, qp->rq.rsv_sge);
- if (!qp->rq_rinl_buf.wqe_cnt)
- return;
-
- /* QP support receive inline wqe */
- sge_list = qp->rq_rinl_buf.wqe_list[wqe_idx].sg_list;
- qp->rq_rinl_buf.wqe_list[wqe_idx].sge_cnt = (unsigned int)wr->num_sge;
- for (i = 0; i < wr->num_sge; i++) {
- sge_list[i].addr = (void *)(uintptr_t)wr->sg_list[i].addr;
- sge_list[i].len = wr->sg_list[i].length;
- }
+ fill_recv_inl_buf(&qp->rq_rinl_buf, wqe_idx, wr);
}
static int hns_roce_u_v2_post_recv(struct ibv_qp *ibvqp, struct ibv_recv_wr *wr,
diff --git a/providers/hns/hns_roce_u_verbs.c b/providers/hns/hns_roce_u_verbs.c
index 3e9a306..1d661dd 100644
--- a/providers/hns/hns_roce_u_verbs.c
+++ b/providers/hns/hns_roce_u_verbs.c
@@ -855,43 +855,45 @@ static int verify_qp_create_attr(struct hns_roce_context *ctx,
return verify_qp_create_cap(ctx, attr);
}
-static int qp_alloc_recv_inl_buf(struct ibv_qp_cap *cap,
- struct hns_roce_qp *qp)
+static int alloc_recv_rinl_buf(uint32_t max_sge,
+ struct hns_roce_rinl_buf *rinl_buf)
{
unsigned int cnt;
int i;
- cnt = qp->rq_rinl_buf.wqe_cnt;
- qp->rq_rinl_buf.wqe_list = calloc(cnt,
- sizeof(struct hns_roce_rinl_wqe));
- if (!qp->rq_rinl_buf.wqe_list)
+ cnt = rinl_buf->wqe_cnt;
+ rinl_buf->wqe_list = calloc(cnt,
+ sizeof(struct hns_roce_rinl_wqe));
+ if (!rinl_buf->wqe_list)
return ENOMEM;
- qp->rq_rinl_buf.wqe_list[0].sg_list = calloc(cnt * cap->max_recv_sge,
- sizeof(struct hns_roce_rinl_sge));
- if (!qp->rq_rinl_buf.wqe_list[0].sg_list)
+ rinl_buf->wqe_list[0].sg_list = calloc(cnt * max_sge,
+ sizeof(struct ibv_sge));
+ if (!rinl_buf->wqe_list[0].sg_list) {
+ free(rinl_buf->wqe_list);
return ENOMEM;
+ }
for (i = 0; i < cnt; i++) {
- int wqe_size = i * cap->max_recv_sge;
+ int wqe_size = i * max_sge;
- qp->rq_rinl_buf.wqe_list[i].sg_list =
- &(qp->rq_rinl_buf.wqe_list[0].sg_list[wqe_size]);
+ rinl_buf->wqe_list[i].sg_list =
+ &(rinl_buf->wqe_list[0].sg_list[wqe_size]);
}
return 0;
}
-static void qp_free_recv_inl_buf(struct hns_roce_qp *qp)
+static void free_recv_rinl_buf(struct hns_roce_rinl_buf *rinl_buf)
{
- if (qp->rq_rinl_buf.wqe_list) {
- if (qp->rq_rinl_buf.wqe_list[0].sg_list) {
- free(qp->rq_rinl_buf.wqe_list[0].sg_list);
- qp->rq_rinl_buf.wqe_list[0].sg_list = NULL;
+ if (rinl_buf->wqe_list) {
+ if (rinl_buf->wqe_list[0].sg_list) {
+ free(rinl_buf->wqe_list[0].sg_list);
+ rinl_buf->wqe_list[0].sg_list = NULL;
}
- free(qp->rq_rinl_buf.wqe_list);
- qp->rq_rinl_buf.wqe_list = NULL;
+ free(rinl_buf->wqe_list);
+ rinl_buf->wqe_list = NULL;
}
}
@@ -930,7 +932,7 @@ static int calc_qp_buff_size(struct hns_roce_device *hr_dev,
static void qp_free_wqe(struct hns_roce_qp *qp)
{
- qp_free_recv_inl_buf(qp);
+ free_recv_rinl_buf(&qp->rq_rinl_buf);
if (qp->sq.wqe_cnt)
free(qp->sq.wrid);
@@ -958,7 +960,7 @@ static int qp_alloc_wqe(struct ibv_qp_cap *cap, struct hns_roce_qp *qp,
}
if (qp->rq_rinl_buf.wqe_cnt) {
- if (qp_alloc_recv_inl_buf(cap, qp))
+ if (alloc_recv_rinl_buf(cap->max_recv_sge, &qp->rq_rinl_buf))
goto err_alloc;
}
@@ -968,7 +970,7 @@ static int qp_alloc_wqe(struct ibv_qp_cap *cap, struct hns_roce_qp *qp,
return 0;
err_alloc:
- qp_free_recv_inl_buf(qp);
+ free_recv_rinl_buf(&qp->rq_rinl_buf);
if (qp->rq.wrid)
free(qp->rq.wrid);
--
2.30.0

46
0008-libhns-RQ-inline-support-wc_x_poll_cq-interface.patch Normal file
View File

@ -0,0 +1,46 @@
From 22beeec9a0d8272fc4db60275ee4eee890068102 Mon Sep 17 00:00:00 2001
From: Luoyouming <luoyouming@huawei.com>
Date: Tue, 11 Oct 2022 10:21:24 +0800
Subject: [PATCH v4 08/10] libhns: RQ inline support wc_x_poll_cq interface
RQ inline support user use wc_x_poll_cq get data.
Signed-off-by: Luoyouming <luoyouming@huawei.com>
Reviewed-by: Yangyang Li <liyangyang20@huawei.com>
---
providers/hns/hns_roce_u_hw_v2.c | 9 +++++++--
1 file changed, 7 insertions(+), 2 deletions(-)
diff --git a/providers/hns/hns_roce_u_hw_v2.c b/providers/hns/hns_roce_u_hw_v2.c
index b37ea92..25d8861 100644
--- a/providers/hns/hns_roce_u_hw_v2.c
+++ b/providers/hns/hns_roce_u_hw_v2.c
@@ -561,10 +561,15 @@ static void cqe_proc_srq(struct hns_roce_srq *srq, uint32_t wqe_idx,
hns_roce_free_srq_wqe(srq, wqe_idx);
}
-static void cqe_proc_rq(struct hns_roce_wq *wq, struct hns_roce_cq *cq)
+static void cqe_proc_rq(struct hns_roce_qp *hr_qp, struct hns_roce_cq *cq)
{
+ struct hns_roce_wq *wq = &hr_qp->rq;
+
cq->verbs_cq.cq_ex.wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
++wq->tail;
+
+ if (hr_reg_read(cq->cqe, CQE_RQ_INLINE))
+ handle_recv_rq_inl(cq->cqe, hr_qp);
}
static int cqe_proc_wq(struct hns_roce_context *ctx, struct hns_roce_qp *qp,
@@ -584,7 +589,7 @@ static int cqe_proc_wq(struct hns_roce_context *ctx, struct hns_roce_qp *qp,
if (srq)
cqe_proc_srq(srq, wqe_idx, cq);
else
- cqe_proc_rq(&qp->rq, cq);
+ cqe_proc_rq(qp, cq);
}
return 0;
--
2.30.0

33
0009-Update-kernel-headers.patch Normal file
View File

@ -0,0 +1,33 @@
From 14cee9bd8ab06104b9f9a0326b8d17a5bf8ee647 Mon Sep 17 00:00:00 2001
From: Luoyouming <luoyouming@huawei.com>
Date: Tue, 11 Oct 2022 10:50:36 +0800
Subject: [PATCH v4 09/10] Update kernel headers
To commit ?? ("RDMA/hns: Support cqe inline in user space").
Signed-off-by: Luoyouming <luoyouming@huawei.com>
---
kernel-headers/rdma/hns-abi.h | 2 ++
1 file changed, 2 insertions(+)
diff --git a/kernel-headers/rdma/hns-abi.h b/kernel-headers/rdma/hns-abi.h
index c70465d..41738b8 100644
--- a/kernel-headers/rdma/hns-abi.h
+++ b/kernel-headers/rdma/hns-abi.h
@@ -88,11 +88,13 @@ struct hns_roce_ib_create_qp_resp {
enum {
HNS_ROCE_EXSGE_FLAGS = 1 << 0,
HNS_ROCE_RQ_INLINE_FLAGS = 1 << 1,
+ HNS_ROCE_CQE_INLINE_FLAGS = 1 << 2,
};
enum {
HNS_ROCE_RSP_EXSGE_FLAGS = 1 << 0,
HNS_ROCE_RSP_RQ_INLINE_FLAGS = 1 << 1,
+ HNS_ROCE_RSP_CQE_INLINE_FLAGS = 1 << 2,
};
struct hns_roce_ib_alloc_ucontext_resp {
--
2.30.0

263
0010-libhns-Support-cqe-inline.patch Normal file
View File

@ -0,0 +1,263 @@
From 71eb90581a338242a26123790e5f24df90327465 Mon Sep 17 00:00:00 2001
From: Luoyouming <luoyouming@huawei.com>
Date: Thu, 11 Aug 2022 20:50:54 +0800
Subject: [PATCH v4 10/10] libhns: Support cqe inline
When rq or srq recv data less than or equal to 32 byte in size, roce driver
support get data from cqe.
Signed-off-by: Luoyouming <luoyouming@huawei.com>
Reviewed-by: Yangyang Li <liyangyang20@huawei.com>
---
providers/hns/hns_roce_u.c | 3 ++-
providers/hns/hns_roce_u.h | 21 +++++++++--------
providers/hns/hns_roce_u_hw_v2.c | 39 ++++++++++++++++++++++++++++++--
providers/hns/hns_roce_u_hw_v2.h | 4 ++--
providers/hns/hns_roce_u_verbs.c | 25 ++++++++++++++++++--
5 files changed, 75 insertions(+), 17 deletions(-)
diff --git a/providers/hns/hns_roce_u.c b/providers/hns/hns_roce_u.c
index 6c9aefa..266e73e 100644
--- a/providers/hns/hns_roce_u.c
+++ b/providers/hns/hns_roce_u.c
@@ -113,7 +113,8 @@ static struct verbs_context *hns_roce_alloc_context(struct ibv_device *ibdev,
if (!context)
return NULL;
- cmd.config |= HNS_ROCE_EXSGE_FLAGS | HNS_ROCE_RQ_INLINE_FLAGS;
+ cmd.config |= HNS_ROCE_EXSGE_FLAGS | HNS_ROCE_RQ_INLINE_FLAGS |
+ HNS_ROCE_CQE_INLINE_FLAGS;
if (ibv_cmd_get_context(&context->ibv_ctx, &cmd.ibv_cmd, sizeof(cmd),
&resp.ibv_resp, sizeof(resp)))
goto err_free;
diff --git a/providers/hns/hns_roce_u.h b/providers/hns/hns_roce_u.h
index 57ebe55..6b64cd0 100644
--- a/providers/hns/hns_roce_u.h
+++ b/providers/hns/hns_roce_u.h
@@ -246,10 +246,21 @@ struct hns_roce_idx_que {
unsigned int tail;
};
+struct hns_roce_rinl_wqe {
+ struct ibv_sge *sg_list;
+ unsigned int sge_cnt;
+};
+
+struct hns_roce_rinl_buf {
+ struct hns_roce_rinl_wqe *wqe_list;
+ unsigned int wqe_cnt;
+};
+
struct hns_roce_srq {
struct verbs_srq verbs_srq;
struct hns_roce_idx_que idx_que;
struct hns_roce_buf wqe_buf;
+ struct hns_roce_rinl_buf srq_rinl_buf;
pthread_spinlock_t lock;
unsigned long *wrid;
unsigned int srqn;
@@ -290,16 +301,6 @@ struct hns_roce_sge_ex {
unsigned int sge_shift;
};
-struct hns_roce_rinl_wqe {
- struct ibv_sge *sg_list;
- unsigned int sge_cnt;
-};
-
-struct hns_roce_rinl_buf {
- struct hns_roce_rinl_wqe *wqe_list;
- unsigned int wqe_cnt;
-};
-
struct hns_roce_qp {
struct verbs_qp verbs_qp;
struct hns_roce_buf buf;
diff --git a/providers/hns/hns_roce_u_hw_v2.c b/providers/hns/hns_roce_u_hw_v2.c
index 25d8861..7063b26 100644
--- a/providers/hns/hns_roce_u_hw_v2.c
+++ b/providers/hns/hns_roce_u_hw_v2.c
@@ -444,6 +444,28 @@ static void handle_recv_inl_data(struct hns_roce_v2_cqe *cqe,
}
+static void handle_recv_cqe_inl_from_rq(struct hns_roce_v2_cqe *cqe,
+ struct hns_roce_qp *cur_qp)
+{
+ uint32_t wr_num;
+
+ wr_num = hr_reg_read(cqe, CQE_WQE_IDX) & (cur_qp->rq.wqe_cnt - 1);
+
+ handle_recv_inl_data(cqe, &(cur_qp->rq_rinl_buf), wr_num,
+ (uint8_t *)cqe->payload);
+}
+
+static void handle_recv_cqe_inl_from_srq(struct hns_roce_v2_cqe *cqe,
+ struct hns_roce_srq *srq)
+{
+ uint32_t wr_num;
+
+ wr_num = hr_reg_read(cqe, CQE_WQE_IDX) & (srq->wqe_cnt - 1);
+
+ handle_recv_inl_data(cqe, &(srq->srq_rinl_buf), wr_num,
+ (uint8_t *)cqe->payload);
+}
+
static void handle_recv_rq_inl(struct hns_roce_v2_cqe *cqe,
struct hns_roce_qp *cur_qp)
{
@@ -473,6 +495,9 @@ static void parse_cqe_for_srq(struct hns_roce_v2_cqe *cqe, struct ibv_wc *wc,
wqe_idx = hr_reg_read(cqe, CQE_WQE_IDX);
wc->wr_id = srq->wrid[wqe_idx & (srq->wqe_cnt - 1)];
hns_roce_free_srq_wqe(srq, wqe_idx);
+
+ if (hr_reg_read(cqe, CQE_CQE_INLINE))
+ handle_recv_cqe_inl_from_srq(cqe, srq);
}
static int parse_cqe_for_resp(struct hns_roce_v2_cqe *cqe, struct ibv_wc *wc,
@@ -487,7 +512,9 @@ static int parse_cqe_for_resp(struct hns_roce_v2_cqe *cqe, struct ibv_wc *wc,
if (hr_qp->verbs_qp.qp.qp_type == IBV_QPT_UD)
parse_for_ud_qp(cqe, wc);
- if (hr_reg_read(cqe, CQE_RQ_INLINE))
+ if (hr_reg_read(cqe, CQE_CQE_INLINE))
+ handle_recv_cqe_inl_from_rq(cqe, hr_qp);
+ else if (hr_reg_read(cqe, CQE_RQ_INLINE))
handle_recv_rq_inl(cqe, hr_qp);
return 0;
@@ -559,6 +586,9 @@ static void cqe_proc_srq(struct hns_roce_srq *srq, uint32_t wqe_idx,
{
cq->verbs_cq.cq_ex.wr_id = srq->wrid[wqe_idx & (srq->wqe_cnt - 1)];
hns_roce_free_srq_wqe(srq, wqe_idx);
+
+ if (hr_reg_read(cq->cqe, CQE_CQE_INLINE))
+ handle_recv_cqe_inl_from_srq(cq->cqe, srq);
}
static void cqe_proc_rq(struct hns_roce_qp *hr_qp, struct hns_roce_cq *cq)
@@ -568,7 +598,9 @@ static void cqe_proc_rq(struct hns_roce_qp *hr_qp, struct hns_roce_cq *cq)
cq->verbs_cq.cq_ex.wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
++wq->tail;
- if (hr_reg_read(cq->cqe, CQE_RQ_INLINE))
+ if (hr_reg_read(cq->cqe, CQE_CQE_INLINE))
+ handle_recv_cqe_inl_from_rq(cq->cqe, hr_qp);
+ else if (hr_reg_read(cq->cqe, CQE_RQ_INLINE))
handle_recv_rq_inl(cq->cqe, hr_qp);
}
@@ -1725,6 +1757,9 @@ static int hns_roce_u_v2_post_srq_recv(struct ibv_srq *ib_srq,
wqe = get_srq_wqe(srq, wqe_idx);
fill_recv_sge_to_wqe(wr, wqe, max_sge, srq->rsv_sge);
+
+ fill_recv_inl_buf(&srq->srq_rinl_buf, wqe_idx, wr);
+
fill_wqe_idx(srq, wqe_idx);
srq->wrid[wqe_idx] = wr->wr_id;
diff --git a/providers/hns/hns_roce_u_hw_v2.h b/providers/hns/hns_roce_u_hw_v2.h
index 098dbdf..d71c695 100644
--- a/providers/hns/hns_roce_u_hw_v2.h
+++ b/providers/hns/hns_roce_u_hw_v2.h
@@ -157,7 +157,7 @@ struct hns_roce_v2_cqe {
__le32 smac;
__le32 byte_28;
__le32 byte_32;
- __le32 rsv[8];
+ __le32 payload[8];
};
#define CQE_FIELD_LOC(h, l) FIELD_LOC(struct hns_roce_v2_cqe, h, l)
@@ -170,7 +170,7 @@ struct hns_roce_v2_cqe {
#define CQE_WQE_IDX CQE_FIELD_LOC(31, 16)
#define CQE_RKEY_IMMTDATA CQE_FIELD_LOC(63, 32)
#define CQE_XRC_SRQN CQE_FIELD_LOC(87, 64)
-#define CQE_RSV0 CQE_FIELD_LOC(95, 88)
+#define CQE_CQE_INLINE CQE_FIELD_LOC(89, 88)
#define CQE_LCL_QPN CQE_FIELD_LOC(119, 96)
#define CQE_SUB_STATUS CQE_FIELD_LOC(127, 120)
#define CQE_BYTE_CNT CQE_FIELD_LOC(159, 128)
diff --git a/providers/hns/hns_roce_u_verbs.c b/providers/hns/hns_roce_u_verbs.c
index 1d661dd..cff9d1d 100644
--- a/providers/hns/hns_roce_u_verbs.c
+++ b/providers/hns/hns_roce_u_verbs.c
@@ -522,6 +522,8 @@ static int verify_srq_create_attr(struct hns_roce_context *context,
static void set_srq_param(struct ibv_context *context, struct hns_roce_srq *srq,
struct ibv_srq_init_attr_ex *attr)
{
+ struct hns_roce_context *ctx = to_hr_ctx(context);
+
if (to_hr_dev(context->device)->hw_version == HNS_ROCE_HW_VER2)
srq->rsv_sge = 1;
@@ -531,6 +533,10 @@ static void set_srq_param(struct ibv_context *context, struct hns_roce_srq *srq,
srq->max_gs));
attr->attr.max_sge = srq->max_gs;
attr->attr.srq_limit = 0;
+
+ srq->srq_rinl_buf.wqe_cnt = 0;
+ if (ctx->config & HNS_ROCE_RSP_CQE_INLINE_FLAGS)
+ srq->srq_rinl_buf.wqe_cnt = srq->wqe_cnt;
}
static int alloc_srq_idx_que(struct hns_roce_srq *srq)
@@ -570,6 +576,11 @@ static int alloc_srq_wqe_buf(struct hns_roce_srq *srq)
return hns_roce_alloc_buf(&srq->wqe_buf, buf_size, HNS_HW_PAGE_SIZE);
}
+static int alloc_recv_rinl_buf(uint32_t max_sge,
+ struct hns_roce_rinl_buf *rinl_buf);
+
+static void free_recv_rinl_buf(struct hns_roce_rinl_buf *rinl_buf);
+
static int alloc_srq_buf(struct hns_roce_srq *srq)
{
int ret;
@@ -582,14 +593,22 @@ static int alloc_srq_buf(struct hns_roce_srq *srq)
if (ret)
goto err_idx_que;
+ if (srq->srq_rinl_buf.wqe_cnt) {
+ ret = alloc_recv_rinl_buf(srq->max_gs, &srq->srq_rinl_buf);
+ if (ret)
+ goto err_wqe_buf;
+ }
+
srq->wrid = calloc(srq->wqe_cnt, sizeof(*srq->wrid));
if (!srq->wrid) {
ret = -ENOMEM;
- goto err_wqe_buf;
+ goto err_inl_buf;
}
return 0;
+err_inl_buf:
+ free_recv_rinl_buf(&srq->srq_rinl_buf);
err_wqe_buf:
hns_roce_free_buf(&srq->wqe_buf);
err_idx_que:
@@ -603,6 +622,7 @@ static void free_srq_buf(struct hns_roce_srq *srq)
{
free(srq->wrid);
hns_roce_free_buf(&srq->wqe_buf);
+ free_recv_rinl_buf(&srq->srq_rinl_buf);
hns_roce_free_buf(&srq->idx_que.buf);
free(srq->idx_que.bitmap);
}
@@ -1082,7 +1102,8 @@ static void hns_roce_set_qp_params(struct ibv_qp_init_attr_ex *attr,
qp->rq.wqe_cnt = cnt;
qp->rq.shift = hr_ilog32(cnt);
qp->rq_rinl_buf.wqe_cnt = 0;
- if (ctx->config & HNS_ROCE_RSP_RQ_INLINE_FLAGS)
+ if (ctx->config & (HNS_ROCE_RSP_RQ_INLINE_FLAGS |
+ HNS_ROCE_RSP_CQE_INLINE_FLAGS))
qp->rq_rinl_buf.wqe_cnt = cnt;
attr->cap.max_recv_wr = qp->rq.wqe_cnt;
--
2.30.0

19
rdma-core.spec
View File

@ -1,11 +1,22 @@
Name: rdma-core
Version: 41.0
Release: 1
Release: 2
Summary: RDMA core userspace libraries and daemons
License: GPLv2 or BSD
Url: https://github.com/linux-rdma/rdma-core
Source: https://github.com/linux-rdma/rdma-core/releases/download/v%{version}/%{name}-%{version}.tar.gz
Patch0: 0001-libhns-Use-a-constant-instead-of-sizeof-operation.patch
Patch1: 0002-libhns-Fix-ext_sge-num-error-when-post-send.patch
Patch2: 0003-Update-kernel-headers.patch
Patch3: 0004-libhns-Fix-the-problem-of-sge-nums.patch
Patch4: 0005-Update-kernel-headers.patch
Patch5: 0006-libhns-Add-compatibility-handling-for-rq-inline.patch
Patch6: 0007-libhns-Refactor-rq-inline.patch
Patch7: 0008-libhns-RQ-inline-support-wc_x_poll_cq-interface.patch
Patch8: 0009-Update-kernel-headers.patch
Patch9: 0010-libhns-Support-cqe-inline.patch
BuildRequires: binutils cmake >= 2.8.11 gcc libudev-devel pkgconfig pkgconfig(libnl-3.0)
BuildRequires: pkgconfig(libnl-route-3.0) valgrind-devel systemd systemd-devel
BuildRequires: python3-devel python3-Cython python3 python3-docutils perl-generators
@ -249,6 +260,12 @@ fi
%{_mandir}/*
%changelog
* Sat Oct 08 2022 luoyouming <luoyouming@huawei.com> - 41.0-2
- Type: requirement
- ID: NA
- SUG: NA
- DESC: Support rq inline and cqe inline
* Mon Jul 25 2022 tangchengchang <tangchengchang@huawei.com> - 41.0-1
- Type: requirement
- ID: NA