* [PATCH rdma-core 1/5] libhns: Add support of handling AH for hip08
2019-09-10 12:20 [PATCH rdma-core 0/5] Support UD on hip08 Weihang Li
@ 2019-09-10 12:20 ` Weihang Li
2019-09-10 12:20 ` [PATCH rdma-core 2/5] libhns: Bugfix for wqe idx calc of post verbs Weihang Li
` (3 subsequent siblings)
4 siblings, 0 replies; 10+ messages in thread
From: Weihang Li @ 2019-09-10 12:20 UTC (permalink / raw)
To: dledford, jgg; +Cc: linux-rdma, linuxarm
From: Lijun Ou <oulijun@huawei.com>
This patch achieves two verbs create_ah and destroy_ah to support
allocation and destruction of Address Handle.
Signed-off-by: Lijun Ou <oulijun@huawei.com>
Signed-off-by: Weihang Li <liweihang@hisilicon.com>
---
| 7 ++++++
providers/hns/hns_roce_u.c | 2 ++
providers/hns/hns_roce_u.h | 30 +++++++++++++++++++++++
providers/hns/hns_roce_u_abi.h | 3 ++-
providers/hns/hns_roce_u_verbs.c | 53 ++++++++++++++++++++++++++++++++++++++++
5 files changed, 94 insertions(+), 1 deletion(-)
--git a/kernel-headers/rdma/hns-abi.h b/kernel-headers/rdma/hns-abi.h
index eb76b38..5f9b178 100644
--- a/kernel-headers/rdma/hns-abi.h
+++ b/kernel-headers/rdma/hns-abi.h
@@ -80,4 +80,11 @@ struct hns_roce_ib_alloc_pd_resp {
__u32 pdn;
};
+struct hns_roce_ib_create_ah_resp {
+ __u8 dmac[6];
+ __u16 vlan_id;
+ __u8 vlan_en;
+ __u8 reserved[7];
+};
+
#endif /* HNS_ABI_USER_H */
diff --git a/providers/hns/hns_roce_u.c b/providers/hns/hns_roce_u.c
index 5872599..8ba41de 100644
--- a/providers/hns/hns_roce_u.c
+++ b/providers/hns/hns_roce_u.c
@@ -81,6 +81,8 @@ static const struct verbs_context_ops hns_common_ops = {
.modify_srq = hns_roce_u_modify_srq,
.query_srq = hns_roce_u_query_srq,
.destroy_srq = hns_roce_u_destroy_srq,
+ .create_ah = hns_roce_u_create_ah,
+ .destroy_ah = hns_roce_u_destroy_ah,
};
static struct verbs_context *hns_roce_alloc_context(struct ibv_device *ibdev,
diff --git a/providers/hns/hns_roce_u.h b/providers/hns/hns_roce_u.h
index 23e0f13..45472fe 100644
--- a/providers/hns/hns_roce_u.h
+++ b/providers/hns/hns_roce_u.h
@@ -68,6 +68,8 @@
#define HNS_ROCE_TPTR_OFFSET 0x1000
#define HNS_ROCE_STATIC_RATE 3 /* Gbps */
+#define ETH_ALEN 6
+
#define roce_get_field(origin, mask, shift) \
(((le32toh(origin)) & (mask)) >> (shift))
@@ -244,6 +246,25 @@ struct hns_roce_qp {
unsigned long flags;
};
+struct hns_roce_av {
+ uint8_t port;
+ uint8_t gid_index;
+ uint8_t static_rate;
+ uint8_t hop_limit;
+ uint32_t flowlabel;
+ uint8_t sl;
+ uint8_t tclass;
+ uint8_t dgid[HNS_ROCE_GID_SIZE];
+ uint8_t mac[ETH_ALEN];
+ uint16_t vlan_id;
+ uint8_t vlan_en;
+};
+
+struct hns_roce_ah {
+ struct ibv_ah ibv_ah;
+ struct hns_roce_av av;
+};
+
struct hns_roce_u_hw {
uint32_t hw_version;
struct verbs_context_ops hw_ops;
@@ -280,6 +301,11 @@ static inline struct hns_roce_qp *to_hr_qp(struct ibv_qp *ibv_qp)
return container_of(ibv_qp, struct hns_roce_qp, ibv_qp);
}
+static inline struct hns_roce_ah *to_hr_ah(struct ibv_ah *ibv_ah)
+{
+ return container_of(ibv_ah, struct hns_roce_ah, ibv_ah);
+}
+
int hns_roce_u_query_device(struct ibv_context *context,
struct ibv_device_attr *attr);
int hns_roce_u_query_port(struct ibv_context *context, uint8_t port,
@@ -319,6 +345,10 @@ struct ibv_qp *hns_roce_u_create_qp(struct ibv_pd *pd,
int hns_roce_u_query_qp(struct ibv_qp *ibqp, struct ibv_qp_attr *attr,
int attr_mask, struct ibv_qp_init_attr *init_attr);
+struct ibv_ah *hns_roce_u_create_ah(struct ibv_pd *pd,
+ struct ibv_ah_attr *attr);
+int hns_roce_u_destroy_ah(struct ibv_ah *ah);
+
int hns_roce_alloc_buf(struct hns_roce_buf *buf, unsigned int size,
int page_size);
void hns_roce_free_buf(struct hns_roce_buf *buf);
diff --git a/providers/hns/hns_roce_u_abi.h b/providers/hns/hns_roce_u_abi.h
index 79fd7dd..029535e 100644
--- a/providers/hns/hns_roce_u_abi.h
+++ b/providers/hns/hns_roce_u_abi.h
@@ -48,5 +48,6 @@ DECLARE_DRV_CMD(hns_roce_alloc_ucontext, IB_USER_VERBS_CMD_GET_CONTEXT,
DECLARE_DRV_CMD(hns_roce_create_srq, IB_USER_VERBS_CMD_CREATE_SRQ,
hns_roce_ib_create_srq, hns_roce_ib_create_srq_resp);
-
+DECLARE_DRV_CMD(hns_roce_create_ah, IB_USER_VERBS_CMD_CREATE_AH, empty,
+ hns_roce_ib_create_ah_resp);
#endif /* _HNS_ROCE_U_ABI_H */
diff --git a/providers/hns/hns_roce_u_verbs.c b/providers/hns/hns_roce_u_verbs.c
index 9d222c0..0d10b1c 100644
--- a/providers/hns/hns_roce_u_verbs.c
+++ b/providers/hns/hns_roce_u_verbs.c
@@ -38,6 +38,7 @@
#include <sys/mman.h>
#include <ccan/ilog.h>
#include <ccan/minmax.h>
+#include <ccan/array_size.h>
#include <util/util.h>
#include "hns_roce_u.h"
#include "hns_roce_u_abi.h"
@@ -952,3 +953,55 @@ int hns_roce_u_query_qp(struct ibv_qp *ibqp, struct ibv_qp_attr *attr,
return ret;
}
+
+struct ibv_ah *hns_roce_u_create_ah(struct ibv_pd *pd, struct ibv_ah_attr *attr)
+{
+ struct hns_roce_create_ah_resp resp = {};
+ struct hns_roce_ah *ah;
+ int status;
+
+ ah = calloc(1, sizeof(*ah));
+ if (!ah)
+ return NULL;
+
+ ah->av.port = attr->port_num;
+ ah->av.sl = attr->sl;
+
+ if (attr->static_rate)
+ ah->av.static_rate = IBV_RATE_10_GBPS;
+
+ if (attr->is_global) {
+ ah->av.gid_index = attr->grh.sgid_index;
+ ah->av.hop_limit = attr->grh.hop_limit;
+ ah->av.tclass = attr->grh.traffic_class;
+ ah->av.flowlabel = attr->grh.flow_label;
+
+ memcpy(ah->av.dgid, attr->grh.dgid.raw,
+ ARRAY_SIZE(ah->av.dgid));
+ }
+
+ status = ibv_cmd_create_ah(pd, &ah->ibv_ah, attr, &resp.ibv_resp,
+ sizeof(resp));
+ if (status) {
+ free(ah);
+ return NULL;
+ }
+
+ memcpy(ah->av.mac, resp.dmac, ETH_ALEN);
+ ah->av.vlan_id = resp.vlan_id;
+ ah->av.vlan_en = resp.vlan_en;
+ return &ah->ibv_ah;
+}
+
+int hns_roce_u_destroy_ah(struct ibv_ah *ah)
+{
+ int ret;
+
+ ret = ibv_cmd_destroy_ah(ah);
+ if (ret)
+ return ret;
+
+ free(to_hr_ah(ah));
+
+ return 0;
+}
--
2.8.1
^ permalink raw reply related [flat|nested] 10+ messages in thread* [PATCH rdma-core 2/5] libhns: Bugfix for wqe idx calc of post verbs
2019-09-10 12:20 [PATCH rdma-core 0/5] Support UD on hip08 Weihang Li
2019-09-10 12:20 ` [PATCH rdma-core 1/5] libhns: Add support of handling AH for hip08 Weihang Li
@ 2019-09-10 12:20 ` Weihang Li
2019-09-10 12:20 ` [PATCH rdma-core 3/5] libhns: Refactor for post send Weihang Li
` (2 subsequent siblings)
4 siblings, 0 replies; 10+ messages in thread
From: Weihang Li @ 2019-09-10 12:20 UTC (permalink / raw)
To: dledford, jgg; +Cc: linux-rdma, linuxarm
From: Yixian Liu <liuyixian@huawei.com>
Current wqe idx is calculated before checking wq overflow, this patch
fixes it and simplifies the usage of it.
Signed-off-by: Yixian Liu <liuyixian@huawei.com>
Signed-off-by: Weihang Li <liweihang@hisilicon.com>
---
providers/hns/hns_roce_u_hw_v1.c | 27 ++++++++-----------
providers/hns/hns_roce_u_hw_v2.c | 56 +++++++++++++++++++---------------------
2 files changed, 37 insertions(+), 46 deletions(-)
diff --git a/providers/hns/hns_roce_u_hw_v1.c b/providers/hns/hns_roce_u_hw_v1.c
index fceb57a..6d3aba5 100644
--- a/providers/hns/hns_roce_u_hw_v1.c
+++ b/providers/hns/hns_roce_u_hw_v1.c
@@ -462,7 +462,6 @@ static int hns_roce_u_v1_arm_cq(struct ibv_cq *ibvcq, int solicited)
static int hns_roce_u_v1_post_send(struct ibv_qp *ibvqp, struct ibv_send_wr *wr,
struct ibv_send_wr **bad_wr)
{
- unsigned int ind;
void *wqe;
int nreq;
int ps_opcode, i;
@@ -471,12 +470,10 @@ static int hns_roce_u_v1_post_send(struct ibv_qp *ibvqp, struct ibv_send_wr *wr,
struct hns_roce_wqe_data_seg *dseg = NULL;
struct hns_roce_qp *qp = to_hr_qp(ibvqp);
struct hns_roce_context *ctx = to_hr_ctx(ibvqp->context);
+ unsigned int wqe_idx;
pthread_spin_lock(&qp->sq.lock);
- /* check that state is OK to post send */
- ind = qp->sq.head;
-
for (nreq = 0; wr; ++nreq, wr = wr->next) {
if (hns_roce_wq_overflow(&qp->sq, nreq,
to_hr_cq(qp->ibv_qp.send_cq))) {
@@ -484,6 +481,9 @@ static int hns_roce_u_v1_post_send(struct ibv_qp *ibvqp, struct ibv_send_wr *wr,
*bad_wr = wr;
goto out;
}
+
+ wqe_idx = (qp->sq.head + nreq) & (qp->sq.wqe_cnt - 1);
+
if (wr->num_sge > qp->sq.max_gs) {
ret = -1;
*bad_wr = wr;
@@ -492,10 +492,10 @@ static int hns_roce_u_v1_post_send(struct ibv_qp *ibvqp, struct ibv_send_wr *wr,
goto out;
}
- ctrl = wqe = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
+ ctrl = wqe = get_send_wqe(qp, wqe_idx);
memset(ctrl, 0, sizeof(struct hns_roce_wqe_ctrl_seg));
- qp->sq.wrid[ind & (qp->sq.wqe_cnt - 1)] = wr->wr_id;
+ qp->sq.wrid[wqe_idx] = wr->wr_id;
for (i = 0; i < wr->num_sge; i++)
ctrl->msg_length = htole32(le32toh(ctrl->msg_length) +
wr->sg_list[i].length);
@@ -578,8 +578,6 @@ static int hns_roce_u_v1_post_send(struct ibv_qp *ibvqp, struct ibv_send_wr *wr,
ctrl->flag |=
htole32(wr->num_sge << HNS_ROCE_WQE_SGE_NUM_BIT);
}
-
- ind++;
}
out:
@@ -745,17 +743,14 @@ static int hns_roce_u_v1_post_recv(struct ibv_qp *ibvqp, struct ibv_recv_wr *wr,
{
int ret = 0;
int nreq;
- int ind;
struct ibv_sge *sg;
struct hns_roce_rc_rq_wqe *rq_wqe;
struct hns_roce_qp *qp = to_hr_qp(ibvqp);
struct hns_roce_context *ctx = to_hr_ctx(ibvqp->context);
+ unsigned int wqe_idx;
pthread_spin_lock(&qp->rq.lock);
- /* check that state is OK to post receive */
- ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
-
for (nreq = 0; wr; ++nreq, wr = wr->next) {
if (hns_roce_wq_overflow(&qp->rq, nreq,
to_hr_cq(qp->ibv_qp.recv_cq))) {
@@ -764,13 +759,15 @@ static int hns_roce_u_v1_post_recv(struct ibv_qp *ibvqp, struct ibv_recv_wr *wr,
goto out;
}
+ wqe_idx = (qp->rq.head + nreq) & (qp->rq.wqe_cnt - 1);
+
if (wr->num_sge > qp->rq.max_gs) {
ret = -1;
*bad_wr = wr;
goto out;
}
- rq_wqe = get_recv_wqe(qp, ind);
+ rq_wqe = get_recv_wqe(qp, wqe_idx);
if (wr->num_sge > HNS_ROCE_RC_RQ_WQE_MAX_SGE_NUM) {
ret = -1;
*bad_wr = wr;
@@ -811,9 +808,7 @@ static int hns_roce_u_v1_post_recv(struct ibv_qp *ibvqp, struct ibv_recv_wr *wr,
HNS_ROCE_RC_RQ_WQE_MAX_SGE_NUM - 2);
}
- qp->rq.wrid[ind] = wr->wr_id;
-
- ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
+ qp->rq.wrid[wqe_idx] = wr->wr_id;
}
out:
diff --git a/providers/hns/hns_roce_u_hw_v2.c b/providers/hns/hns_roce_u_hw_v2.c
index 931f59d..439e547 100644
--- a/providers/hns/hns_roce_u_hw_v2.c
+++ b/providers/hns/hns_roce_u_hw_v2.c
@@ -612,27 +612,26 @@ static int hns_roce_u_v2_arm_cq(struct ibv_cq *ibvcq, int solicited)
int hns_roce_u_v2_post_send(struct ibv_qp *ibvqp, struct ibv_send_wr *wr,
struct ibv_send_wr **bad_wr)
{
- unsigned int ind_sge;
- unsigned int ind;
- int nreq;
- void *wqe;
- int ret = 0;
- struct hns_roce_qp *qp = to_hr_qp(ibvqp);
struct hns_roce_context *ctx = to_hr_ctx(ibvqp->context);
- struct hns_roce_rc_sq_wqe *rc_sq_wqe;
+ struct hns_roce_qp *qp = to_hr_qp(ibvqp);
struct hns_roce_v2_wqe_data_seg *dseg;
+ struct hns_roce_rc_sq_wqe *rc_sq_wqe;
struct ibv_qp_attr attr;
+ unsigned int wqe_idx;
+ unsigned int sge_idx;
int valid_num_sge;
int attr_mask;
+ int ret = 0;
+ void *wqe;
+ int nreq;
int j;
int i;
pthread_spin_lock(&qp->sq.lock);
- /* check that state is OK to post send */
- ind = qp->sq.head;
- ind_sge = qp->next_sge;
+ sge_idx = qp->next_sge;
+ /* check that state is OK to post send */
if (ibvqp->state == IBV_QPS_RESET || ibvqp->state == IBV_QPS_INIT ||
ibvqp->state == IBV_QPS_RTR) {
pthread_spin_unlock(&qp->sq.lock);
@@ -648,18 +647,19 @@ int hns_roce_u_v2_post_send(struct ibv_qp *ibvqp, struct ibv_send_wr *wr,
goto out;
}
+ wqe_idx = (qp->sq.head + nreq) & (qp->sq.wqe_cnt - 1);
+
if (wr->num_sge > qp->sq.max_gs) {
ret = EINVAL;
*bad_wr = wr;
goto out;
}
- wqe = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
+ wqe = get_send_wqe(qp, wqe_idx);
rc_sq_wqe = wqe;
memset(rc_sq_wqe, 0, sizeof(struct hns_roce_rc_sq_wqe));
-
- qp->sq.wrid[ind & (qp->sq.wqe_cnt - 1)] = wr->wr_id;
+ qp->sq.wrid[wqe_idx] = wr->wr_id;
valid_num_sge = wr->num_sge;
j = 0;
@@ -880,7 +880,7 @@ int hns_roce_u_v2_post_send(struct ibv_qp *ibvqp, struct ibv_send_wr *wr,
roce_set_field(rc_sq_wqe->byte_20,
RC_SQ_WQE_BYTE_20_MSG_START_SGE_IDX_M,
RC_SQ_WQE_BYTE_20_MSG_START_SGE_IDX_S,
- ind_sge & (qp->sge.sge_cnt - 1));
+ sge_idx & (qp->sge.sge_cnt - 1));
for (i = 0; i < wr->num_sge && j < 2; i++)
if (likely(wr->sg_list[i].length)) {
@@ -893,17 +893,15 @@ int hns_roce_u_v2_post_send(struct ibv_qp *ibvqp, struct ibv_send_wr *wr,
for (; i < wr->num_sge; i++) {
if (likely(wr->sg_list[i].length)) {
dseg = get_send_sge_ex(qp,
- ind_sge &
+ sge_idx &
(qp->sge.sge_cnt - 1));
set_data_seg_v2(dseg,
wr->sg_list + i);
- ind_sge++;
+ sge_idx++;
}
}
}
}
-
- ind++;
}
out:
@@ -916,7 +914,7 @@ out:
if (qp->flags & HNS_ROCE_SUPPORT_SQ_RECORD_DB)
*(qp->sdb) = qp->sq.head & 0xffff;
- qp->next_sge = ind_sge;
+ qp->next_sge = sge_idx;
}
pthread_spin_unlock(&qp->sq.lock);
@@ -936,23 +934,21 @@ out:
static int hns_roce_u_v2_post_recv(struct ibv_qp *ibvqp, struct ibv_recv_wr *wr,
struct ibv_recv_wr **bad_wr)
{
- int ret = 0;
- int nreq;
- int ind;
struct hns_roce_qp *qp = to_hr_qp(ibvqp);
struct hns_roce_context *ctx = to_hr_ctx(ibvqp->context);
struct hns_roce_v2_wqe_data_seg *dseg;
struct hns_roce_rinl_sge *sge_list;
struct ibv_qp_attr attr;
int attr_mask;
+ int ret = 0;
+ int wqe_idx;
void *wqe;
+ int nreq;
int i;
pthread_spin_lock(&qp->rq.lock);
/* check that state is OK to post receive */
- ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
-
if (ibvqp->state == IBV_QPS_RESET) {
pthread_spin_unlock(&qp->rq.lock);
*bad_wr = wr;
@@ -967,13 +963,15 @@ static int hns_roce_u_v2_post_recv(struct ibv_qp *ibvqp, struct ibv_recv_wr *wr,
goto out;
}
+ wqe_idx = (qp->rq.head + nreq) & (qp->rq.wqe_cnt - 1);
+
if (wr->num_sge > qp->rq.max_gs) {
ret = -EINVAL;
*bad_wr = wr;
goto out;
}
- wqe = get_recv_wqe_v2(qp, ind);
+ wqe = get_recv_wqe_v2(qp, wqe_idx);
if (!wqe) {
ret = -EINVAL;
*bad_wr = wr;
@@ -995,8 +993,8 @@ static int hns_roce_u_v2_post_recv(struct ibv_qp *ibvqp, struct ibv_recv_wr *wr,
}
/* QP support receive inline wqe */
- sge_list = qp->rq_rinl_buf.wqe_list[ind].sg_list;
- qp->rq_rinl_buf.wqe_list[ind].sge_cnt =
+ 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++) {
@@ -1005,9 +1003,7 @@ static int hns_roce_u_v2_post_recv(struct ibv_qp *ibvqp, struct ibv_recv_wr *wr,
sge_list[i].len = wr->sg_list[i].length;
}
- qp->rq.wrid[ind] = wr->wr_id;
-
- ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
+ qp->rq.wrid[wqe_idx] = wr->wr_id;
}
out:
--
2.8.1
^ permalink raw reply related [flat|nested] 10+ messages in thread* [PATCH rdma-core 3/5] libhns: Refactor for post send
2019-09-10 12:20 [PATCH rdma-core 0/5] Support UD on hip08 Weihang Li
2019-09-10 12:20 ` [PATCH rdma-core 1/5] libhns: Add support of handling AH for hip08 Weihang Li
2019-09-10 12:20 ` [PATCH rdma-core 2/5] libhns: Bugfix for wqe idx calc of post verbs Weihang Li
@ 2019-09-10 12:20 ` Weihang Li
2019-09-11 7:35 ` Leon Romanovsky
2019-09-10 12:20 ` [PATCH rdma-core 4/5] libhns: Add UD support for hip08 in user mode Weihang Li
2019-09-10 12:20 ` [PATCH rdma-core 5/5] libhns: Support configuring loopback mode by user Weihang Li
4 siblings, 1 reply; 10+ messages in thread
From: Weihang Li @ 2019-09-10 12:20 UTC (permalink / raw)
To: dledford, jgg; +Cc: linux-rdma, linuxarm
From: Yixian Liu <liuyixian@huawei.com>
This patch refactors the interface of hns_roce_u_v2_post_send, which
is now very complicated. We reduce the complexity with following points:
1. Separate RC server into a function.
2. Simplify and separate the process of sge.
3. Keep the logic and consistence of all operations.
Signed-off-by: Yixian Liu <liuyixian@huawei.com>
Signed-off-by: Weihang Li <liweihang@hisilicon.com>
---
providers/hns/hns_roce_u.h | 7 +
providers/hns/hns_roce_u_hw_v2.c | 427 ++++++++++++++++-----------------------
2 files changed, 186 insertions(+), 248 deletions(-)
diff --git a/providers/hns/hns_roce_u.h b/providers/hns/hns_roce_u.h
index 45472fe..9d8f72e 100644
--- a/providers/hns/hns_roce_u.h
+++ b/providers/hns/hns_roce_u.h
@@ -206,6 +206,13 @@ struct hns_roce_wq {
int offset;
};
+/* record the result of sge process */
+struct hns_roce_sge_info {
+ unsigned int valid_num; /* sge length is not 0 */
+ unsigned int start_idx; /* start position of extend sge */
+ unsigned int total_len; /* total length of valid sges */
+};
+
struct hns_roce_sge_ex {
int offset;
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 439e547..21a6ee0 100644
--- a/providers/hns/hns_roce_u_hw_v2.c
+++ b/providers/hns/hns_roce_u_hw_v2.c
@@ -609,28 +609,188 @@ static int hns_roce_u_v2_arm_cq(struct ibv_cq *ibvcq, int solicited)
return 0;
}
+static void set_sge(struct hns_roce_v2_wqe_data_seg *dseg,
+ struct hns_roce_qp *qp, struct ibv_send_wr *wr,
+ struct hns_roce_sge_info *sge_info)
+{
+ int i;
+
+ for (i = 0; i < wr->num_sge; i++)
+ if (likely(wr->sg_list[i].length)) {
+ sge_info->total_len += wr->sg_list[i].length;
+ sge_info->valid_num++;
+
+ /* No inner sge in UD wqe */
+ if (sge_info->valid_num <= HNS_ROCE_SGE_IN_WQE &&
+ qp->ibv_qp.qp_type != IBV_QPT_UD) {
+ set_data_seg_v2(dseg, wr->sg_list + i);
+ dseg++;
+ } else {
+ dseg = get_send_sge_ex(qp, sge_info->start_idx &
+ (qp->sge.sge_cnt - 1));
+ set_data_seg_v2(dseg, wr->sg_list + i);
+ sge_info->start_idx++;
+ }
+ }
+}
+
+static int set_rc_wqe(void *wqe, struct hns_roce_qp *qp, struct ibv_send_wr *wr,
+ int nreq, struct hns_roce_sge_info *sge_info)
+{
+ struct hns_roce_rc_sq_wqe *rc_sq_wqe = wqe;
+ struct hns_roce_v2_wqe_data_seg *dseg;
+ int hr_op;
+ int i;
+
+ memset(rc_sq_wqe, 0, sizeof(struct hns_roce_rc_sq_wqe));
+
+ switch (wr->opcode) {
+ case IBV_WR_RDMA_READ:
+ hr_op = HNS_ROCE_WQE_OP_RDMA_READ;
+ rc_sq_wqe->va = htole64(wr->wr.rdma.remote_addr);
+ rc_sq_wqe->rkey = htole32(wr->wr.rdma.rkey);
+ break;
+ case IBV_WR_RDMA_WRITE:
+ hr_op = HNS_ROCE_WQE_OP_RDMA_WRITE;
+ rc_sq_wqe->va = htole64(wr->wr.rdma.remote_addr);
+ rc_sq_wqe->rkey = htole32(wr->wr.rdma.rkey);
+ break;
+ case IBV_WR_RDMA_WRITE_WITH_IMM:
+ hr_op = HNS_ROCE_WQE_OP_RDMA_WRITE_WITH_IMM;
+ rc_sq_wqe->va = htole64(wr->wr.rdma.remote_addr);
+ rc_sq_wqe->rkey = htole32(wr->wr.rdma.rkey);
+ rc_sq_wqe->immtdata = htole32(be32toh(wr->imm_data));
+ break;
+ case IBV_WR_SEND:
+ hr_op = HNS_ROCE_WQE_OP_SEND;
+ break;
+ case IBV_WR_SEND_WITH_INV:
+ hr_op = HNS_ROCE_WQE_OP_SEND_WITH_INV;
+ rc_sq_wqe->inv_key = htole32(wr->invalidate_rkey);
+ break;
+ case IBV_WR_SEND_WITH_IMM:
+ hr_op = HNS_ROCE_WQE_OP_SEND_WITH_IMM;
+ rc_sq_wqe->immtdata = htole32(be32toh(wr->imm_data));
+ break;
+ case IBV_WR_LOCAL_INV:
+ hr_op = HNS_ROCE_WQE_OP_LOCAL_INV;
+ roce_set_bit(rc_sq_wqe->byte_4, RC_SQ_WQE_BYTE_4_SO_S, 1);
+ rc_sq_wqe->inv_key = htole32(wr->invalidate_rkey);
+ break;
+ case IBV_WR_BIND_MW:
+ hr_op = HNS_ROCE_WQE_OP_BIND_MW_TYPE;
+ roce_set_bit(rc_sq_wqe->byte_4, RC_SQ_WQE_BYTE_4_MW_TYPE_S,
+ wr->bind_mw.mw->type - 1);
+ roce_set_bit(rc_sq_wqe->byte_4, RC_SQ_WQE_BYTE_4_ATOMIC_S,
+ (wr->bind_mw.bind_info.mw_access_flags &
+ IBV_ACCESS_REMOTE_ATOMIC) ? 1 : 0);
+ roce_set_bit(rc_sq_wqe->byte_4, RC_SQ_WQE_BYTE_4_RDMA_READ_S,
+ (wr->bind_mw.bind_info.mw_access_flags &
+ IBV_ACCESS_REMOTE_READ) ? 1 : 0);
+ roce_set_bit(rc_sq_wqe->byte_4, RC_SQ_WQE_BYTE_4_RDMA_WRITE_S,
+ (wr->bind_mw.bind_info.mw_access_flags &
+ IBV_ACCESS_REMOTE_WRITE) ? 1 : 0);
+ rc_sq_wqe->new_rkey = htole32(wr->bind_mw.rkey);
+ rc_sq_wqe->byte_16 =
+ htole32(wr->bind_mw.bind_info.length & 0xffffffff);
+ rc_sq_wqe->byte_20 =
+ htole32(wr->bind_mw.bind_info.length >> 32);
+ rc_sq_wqe->rkey = htole32(wr->bind_mw.bind_info.mr->rkey);
+ rc_sq_wqe->va = htole64(wr->bind_mw.bind_info.addr);
+ break;
+ case IBV_WR_ATOMIC_CMP_AND_SWP:
+ hr_op = HNS_ROCE_WQE_OP_ATOMIC_COM_AND_SWAP;
+ rc_sq_wqe->rkey = htole32(wr->wr.atomic.rkey);
+ rc_sq_wqe->va = htole64(wr->wr.atomic.remote_addr);
+ break;
+ case IBV_WR_ATOMIC_FETCH_AND_ADD:
+ hr_op = HNS_ROCE_WQE_OP_ATOMIC_FETCH_AND_ADD;
+ rc_sq_wqe->rkey = htole32(wr->wr.atomic.rkey);
+ rc_sq_wqe->va = htole64(wr->wr.atomic.remote_addr);
+ break;
+ default:
+ hr_op = HNS_ROCE_WQE_OP_MASK;
+ printf("Not supported transport opcode %d\n", wr->opcode);
+ return -EINVAL;
+ }
+
+ roce_set_field(rc_sq_wqe->byte_4, RC_SQ_WQE_BYTE_4_OPCODE_M,
+ RC_SQ_WQE_BYTE_4_OPCODE_S, hr_op);
+
+ roce_set_bit(rc_sq_wqe->byte_4, RC_SQ_WQE_BYTE_4_CQE_S,
+ (wr->send_flags & IBV_SEND_SIGNALED) ? 1 : 0);
+
+ roce_set_bit(rc_sq_wqe->byte_4, RC_SQ_WQE_BYTE_4_FENCE_S,
+ (wr->send_flags & IBV_SEND_FENCE) ? 1 : 0);
+
+ roce_set_bit(rc_sq_wqe->byte_4, RC_SQ_WQE_BYTE_4_SE_S,
+ (wr->send_flags & IBV_SEND_SOLICITED) ? 1 : 0);
+
+ roce_set_bit(rc_sq_wqe->byte_4, RC_SQ_WQE_BYTE_4_OWNER_S,
+ ~(((qp->sq.head + nreq) >> qp->sq.shift) & 0x1));
+
+ roce_set_field(rc_sq_wqe->byte_20,
+ RC_SQ_WQE_BYTE_20_MSG_START_SGE_IDX_M,
+ RC_SQ_WQE_BYTE_20_MSG_START_SGE_IDX_S,
+ sge_info->start_idx & (qp->sge.sge_cnt - 1));
+
+ if (wr->opcode == IBV_WR_BIND_MW)
+ return 0;
+
+ wqe += sizeof(struct hns_roce_rc_sq_wqe);
+ dseg = wqe;
+
+ set_sge(dseg, qp, wr, sge_info);
+
+ rc_sq_wqe->msg_len = htole32(sge_info->total_len);
+
+ roce_set_field(rc_sq_wqe->byte_16, RC_SQ_WQE_BYTE_16_SGE_NUM_M,
+ RC_SQ_WQE_BYTE_16_SGE_NUM_S, sge_info->valid_num);
+
+ if (wr->opcode == IBV_WR_ATOMIC_FETCH_AND_ADD ||
+ wr->opcode == IBV_WR_ATOMIC_CMP_AND_SWP) {
+ set_atomic_seg((struct hns_roce_wqe_atomic_seg *)dseg, wr);
+ return 0;
+ }
+
+ if (wr->send_flags & IBV_SEND_INLINE) {
+ if (wr->opcode == IBV_WR_RDMA_READ) {
+ printf("Not supported inline data!\n");
+ return -EINVAL;
+ }
+
+ if (sge_info->total_len > qp->max_inline_data) {
+ printf("Failed to inline, data len=%d, max inline len=%d!\n",
+ sge_info->total_len, qp->max_inline_data);
+ return -EINVAL;
+ }
+
+ for (i = 0; i < wr->num_sge; i++) {
+ memcpy(dseg, (void *)(uintptr_t)(wr->sg_list[i].addr),
+ wr->sg_list[i].length);
+ dseg = dseg + wr->sg_list[i].length;
+ }
+ roce_set_bit(rc_sq_wqe->byte_4, RC_SQ_WQE_BYTE_4_INLINE_S, 1);
+ }
+
+ return 0;
+}
+
int hns_roce_u_v2_post_send(struct ibv_qp *ibvqp, struct ibv_send_wr *wr,
struct ibv_send_wr **bad_wr)
{
struct hns_roce_context *ctx = to_hr_ctx(ibvqp->context);
struct hns_roce_qp *qp = to_hr_qp(ibvqp);
- struct hns_roce_v2_wqe_data_seg *dseg;
- struct hns_roce_rc_sq_wqe *rc_sq_wqe;
+ struct hns_roce_sge_info sge_info = {};
struct ibv_qp_attr attr;
unsigned int wqe_idx;
- unsigned int sge_idx;
- int valid_num_sge;
int attr_mask;
int ret = 0;
void *wqe;
int nreq;
- int j;
- int i;
pthread_spin_lock(&qp->sq.lock);
- sge_idx = qp->next_sge;
-
/* check that state is OK to post send */
if (ibvqp->state == IBV_QPS_RESET || ibvqp->state == IBV_QPS_INIT ||
ibvqp->state == IBV_QPS_RTR) {
@@ -639,6 +799,8 @@ int hns_roce_u_v2_post_send(struct ibv_qp *ibvqp, struct ibv_send_wr *wr,
return EINVAL;
}
+ sge_info.start_idx = qp->next_sge; /* start index of extend sge */
+
for (nreq = 0; wr; ++nreq, wr = wr->next) {
if (hns_roce_v2_wq_overflow(&qp->sq, nreq,
to_hr_cq(qp->ibv_qp.send_cq))) {
@@ -647,274 +809,43 @@ int hns_roce_u_v2_post_send(struct ibv_qp *ibvqp, struct ibv_send_wr *wr,
goto out;
}
- wqe_idx = (qp->sq.head + nreq) & (qp->sq.wqe_cnt - 1);
-
if (wr->num_sge > qp->sq.max_gs) {
ret = EINVAL;
*bad_wr = wr;
goto out;
}
+ wqe_idx = (qp->sq.head + nreq) & (qp->sq.wqe_cnt - 1);
wqe = get_send_wqe(qp, wqe_idx);
- rc_sq_wqe = wqe;
-
- memset(rc_sq_wqe, 0, sizeof(struct hns_roce_rc_sq_wqe));
qp->sq.wrid[wqe_idx] = wr->wr_id;
- valid_num_sge = wr->num_sge;
- j = 0;
-
- for (i = 0; i < wr->num_sge; i++) {
- if (unlikely(!wr->sg_list[i].length))
- valid_num_sge--;
-
- rc_sq_wqe->msg_len =
- htole32(le32toh(rc_sq_wqe->msg_len) +
- wr->sg_list[i].length);
- }
-
- if (wr->opcode == IBV_WR_SEND_WITH_IMM ||
- wr->opcode == IBV_WR_RDMA_WRITE_WITH_IMM)
- rc_sq_wqe->immtdata = htole32(be32toh(wr->imm_data));
-
- roce_set_field(rc_sq_wqe->byte_16, RC_SQ_WQE_BYTE_16_SGE_NUM_M,
- RC_SQ_WQE_BYTE_16_SGE_NUM_S, valid_num_sge);
-
- roce_set_field(rc_sq_wqe->byte_20,
- RC_SQ_WQE_BYTE_20_MSG_START_SGE_IDX_S,
- RC_SQ_WQE_BYTE_20_MSG_START_SGE_IDX_S,
- 0);
-
- roce_set_bit(rc_sq_wqe->byte_4, RC_SQ_WQE_BYTE_4_CQE_S,
- (wr->send_flags & IBV_SEND_SIGNALED) ? 1 : 0);
-
- /* Set fence attr */
- roce_set_bit(rc_sq_wqe->byte_4, RC_SQ_WQE_BYTE_4_FENCE_S,
- (wr->send_flags & IBV_SEND_FENCE) ? 1 : 0);
-
- /* Set solicited attr */
- roce_set_bit(rc_sq_wqe->byte_4, RC_SQ_WQE_BYTE_4_SE_S,
- (wr->send_flags & IBV_SEND_SOLICITED) ? 1 : 0);
-
- roce_set_bit(rc_sq_wqe->byte_4, RC_SQ_WQE_BYTE_4_OWNER_S,
- ~(((qp->sq.head + nreq) >> qp->sq.shift) & 0x1));
-
- wqe += sizeof(struct hns_roce_rc_sq_wqe);
- /* set remote addr segment */
switch (ibvqp->qp_type) {
case IBV_QPT_RC:
- switch (wr->opcode) {
- case IBV_WR_RDMA_READ:
- roce_set_field(rc_sq_wqe->byte_4,
- RC_SQ_WQE_BYTE_4_OPCODE_M,
- RC_SQ_WQE_BYTE_4_OPCODE_S,
- HNS_ROCE_WQE_OP_RDMA_READ);
- rc_sq_wqe->va =
- htole64(wr->wr.rdma.remote_addr);
- rc_sq_wqe->rkey = htole32(wr->wr.rdma.rkey);
- break;
-
- case IBV_WR_RDMA_WRITE:
- roce_set_field(rc_sq_wqe->byte_4,
- RC_SQ_WQE_BYTE_4_OPCODE_M,
- RC_SQ_WQE_BYTE_4_OPCODE_S,
- HNS_ROCE_WQE_OP_RDMA_WRITE);
- rc_sq_wqe->va =
- htole64(wr->wr.rdma.remote_addr);
- rc_sq_wqe->rkey = htole32(wr->wr.rdma.rkey);
- break;
-
- case IBV_WR_RDMA_WRITE_WITH_IMM:
- roce_set_field(rc_sq_wqe->byte_4,
- RC_SQ_WQE_BYTE_4_OPCODE_M,
- RC_SQ_WQE_BYTE_4_OPCODE_S,
- HNS_ROCE_WQE_OP_RDMA_WRITE_WITH_IMM);
- rc_sq_wqe->va =
- htole64(wr->wr.rdma.remote_addr);
- rc_sq_wqe->rkey = htole32(wr->wr.rdma.rkey);
- break;
-
- case IBV_WR_SEND:
- roce_set_field(rc_sq_wqe->byte_4,
- RC_SQ_WQE_BYTE_4_OPCODE_M,
- RC_SQ_WQE_BYTE_4_OPCODE_S,
- HNS_ROCE_WQE_OP_SEND);
- break;
- case IBV_WR_SEND_WITH_INV:
- roce_set_field(rc_sq_wqe->byte_4,
- RC_SQ_WQE_BYTE_4_OPCODE_M,
- RC_SQ_WQE_BYTE_4_OPCODE_S,
- HNS_ROCE_WQE_OP_SEND_WITH_INV);
- rc_sq_wqe->inv_key =
- htole32(wr->invalidate_rkey);
- break;
- case IBV_WR_SEND_WITH_IMM:
- roce_set_field(rc_sq_wqe->byte_4,
- RC_SQ_WQE_BYTE_4_OPCODE_M,
- RC_SQ_WQE_BYTE_4_OPCODE_S,
- HNS_ROCE_WQE_OP_SEND_WITH_IMM);
- break;
-
- case IBV_WR_LOCAL_INV:
- roce_set_field(rc_sq_wqe->byte_4,
- RC_SQ_WQE_BYTE_4_OPCODE_M,
- RC_SQ_WQE_BYTE_4_OPCODE_S,
- HNS_ROCE_WQE_OP_LOCAL_INV);
- roce_set_bit(rc_sq_wqe->byte_4,
- RC_SQ_WQE_BYTE_4_SO_S, 1);
- rc_sq_wqe->inv_key =
- htole32(wr->invalidate_rkey);
- break;
- case IBV_WR_ATOMIC_CMP_AND_SWP:
- roce_set_field(rc_sq_wqe->byte_4,
- RC_SQ_WQE_BYTE_4_OPCODE_M,
- RC_SQ_WQE_BYTE_4_OPCODE_S,
- HNS_ROCE_WQE_OP_ATOMIC_COM_AND_SWAP);
- rc_sq_wqe->rkey = htole32(wr->wr.atomic.rkey);
- rc_sq_wqe->va =
- htole64(wr->wr.atomic.remote_addr);
- break;
-
- case IBV_WR_ATOMIC_FETCH_AND_ADD:
- roce_set_field(rc_sq_wqe->byte_4,
- RC_SQ_WQE_BYTE_4_OPCODE_M,
- RC_SQ_WQE_BYTE_4_OPCODE_S,
- HNS_ROCE_WQE_OP_ATOMIC_FETCH_AND_ADD);
- rc_sq_wqe->rkey = htole32(wr->wr.atomic.rkey);
- rc_sq_wqe->va =
- htole64(wr->wr.atomic.remote_addr);
- break;
-
- case IBV_WR_BIND_MW:
- roce_set_field(rc_sq_wqe->byte_4,
- RC_SQ_WQE_BYTE_4_OPCODE_M,
- RC_SQ_WQE_BYTE_4_OPCODE_S,
- HNS_ROCE_WQE_OP_BIND_MW_TYPE);
- roce_set_bit(rc_sq_wqe->byte_4,
- RC_SQ_WQE_BYTE_4_MW_TYPE_S,
- wr->bind_mw.mw->type - 1);
- roce_set_bit(rc_sq_wqe->byte_4,
- RC_SQ_WQE_BYTE_4_ATOMIC_S,
- wr->bind_mw.bind_info.mw_access_flags &
- IBV_ACCESS_REMOTE_ATOMIC ? 1 : 0);
- roce_set_bit(rc_sq_wqe->byte_4,
- RC_SQ_WQE_BYTE_4_RDMA_READ_S,
- wr->bind_mw.bind_info.mw_access_flags &
- IBV_ACCESS_REMOTE_READ ? 1 : 0);
- roce_set_bit(rc_sq_wqe->byte_4,
- RC_SQ_WQE_BYTE_4_RDMA_WRITE_S,
- wr->bind_mw.bind_info.mw_access_flags &
- IBV_ACCESS_REMOTE_WRITE ? 1 : 0);
-
- rc_sq_wqe->new_rkey = htole32(wr->bind_mw.rkey);
- rc_sq_wqe->byte_16 =
- htole32(wr->bind_mw.bind_info.length &
- 0xffffffff);
- rc_sq_wqe->byte_20 =
- htole32(wr->bind_mw.bind_info.length >>
- 32);
- rc_sq_wqe->rkey =
- htole32(wr->bind_mw.bind_info.mr->rkey);
- rc_sq_wqe->va =
- htole64(wr->bind_mw.bind_info.addr);
- break;
-
- default:
- roce_set_field(rc_sq_wqe->byte_4,
- RC_SQ_WQE_BYTE_4_OPCODE_M,
- RC_SQ_WQE_BYTE_4_OPCODE_S,
- HNS_ROCE_WQE_OP_MASK);
- printf("Not supported transport opcode %d\n",
- wr->opcode);
- break;
+ ret = set_rc_wqe(wqe, qp, wr, nreq, &sge_info);
+ if (ret) {
+ *bad_wr = wr;
+ goto out;
}
-
break;
case IBV_QPT_UC:
case IBV_QPT_UD:
default:
- break;
- }
-
- dseg = wqe;
- if (wr->opcode == IBV_WR_ATOMIC_FETCH_AND_ADD ||
- wr->opcode == IBV_WR_ATOMIC_CMP_AND_SWP) {
- set_data_seg_v2(dseg, wr->sg_list);
- wqe += sizeof(struct hns_roce_v2_wqe_data_seg);
- set_atomic_seg(wqe, wr);
- } else if (wr->send_flags & IBV_SEND_INLINE && valid_num_sge) {
- if (le32toh(rc_sq_wqe->msg_len) > qp->max_inline_data) {
- ret = EINVAL;
- *bad_wr = wr;
- printf("data len=%d, send_flags = 0x%x!\r\n",
- rc_sq_wqe->msg_len, wr->send_flags);
- goto out;
- }
-
- if (wr->opcode == IBV_WR_RDMA_READ) {
- ret = EINVAL;
- *bad_wr = wr;
- printf("Not supported inline data!\n");
- goto out;
- }
-
- for (i = 0; i < wr->num_sge; i++) {
- memcpy(wqe,
- ((void *) (uintptr_t) wr->sg_list[i].addr),
- wr->sg_list[i].length);
- wqe = wqe + wr->sg_list[i].length;
- }
-
- roce_set_bit(rc_sq_wqe->byte_4,
- RC_SQ_WQE_BYTE_4_INLINE_S, 1);
- } else {
- /* set sge */
- if (valid_num_sge <= HNS_ROCE_SGE_IN_WQE) {
- for (i = 0; i < wr->num_sge; i++)
- if (likely(wr->sg_list[i].length)) {
- set_data_seg_v2(dseg,
- wr->sg_list + i);
- dseg++;
- }
- } else {
- roce_set_field(rc_sq_wqe->byte_20,
- RC_SQ_WQE_BYTE_20_MSG_START_SGE_IDX_M,
- RC_SQ_WQE_BYTE_20_MSG_START_SGE_IDX_S,
- sge_idx & (qp->sge.sge_cnt - 1));
-
- for (i = 0; i < wr->num_sge && j < 2; i++)
- if (likely(wr->sg_list[i].length)) {
- set_data_seg_v2(dseg,
- wr->sg_list + i);
- dseg++;
- j++;
- }
-
- for (; i < wr->num_sge; i++) {
- if (likely(wr->sg_list[i].length)) {
- dseg = get_send_sge_ex(qp,
- sge_idx &
- (qp->sge.sge_cnt - 1));
- set_data_seg_v2(dseg,
- wr->sg_list + i);
- sge_idx++;
- }
- }
- }
+ ret = -EINVAL;
+ *bad_wr = wr;
+ goto out;
}
}
out:
if (likely(nreq)) {
qp->sq.head += nreq;
+ qp->next_sge = sge_info.start_idx;
hns_roce_update_sq_db(ctx, qp->ibv_qp.qp_num, qp->sl,
qp->sq.head & ((qp->sq.wqe_cnt << 1) - 1));
if (qp->flags & HNS_ROCE_SUPPORT_SQ_RECORD_DB)
*(qp->sdb) = qp->sq.head & 0xffff;
-
- qp->next_sge = sge_idx;
}
pthread_spin_unlock(&qp->sq.lock);
--
2.8.1
^ permalink raw reply related [flat|nested] 10+ messages in thread* Re: [PATCH rdma-core 3/5] libhns: Refactor for post send
2019-09-10 12:20 ` [PATCH rdma-core 3/5] libhns: Refactor for post send Weihang Li
@ 2019-09-11 7:35 ` Leon Romanovsky
2019-09-11 7:50 ` liweihang
0 siblings, 1 reply; 10+ messages in thread
From: Leon Romanovsky @ 2019-09-11 7:35 UTC (permalink / raw)
To: Weihang Li; +Cc: dledford, jgg, linux-rdma, linuxarm
On Tue, Sep 10, 2019 at 08:20:50PM +0800, Weihang Li wrote:
> From: Yixian Liu <liuyixian@huawei.com>
>
> This patch refactors the interface of hns_roce_u_v2_post_send, which
> is now very complicated. We reduce the complexity with following points:
> 1. Separate RC server into a function.
> 2. Simplify and separate the process of sge.
> 3. Keep the logic and consistence of all operations.
>
> Signed-off-by: Yixian Liu <liuyixian@huawei.com>
> Signed-off-by: Weihang Li <liweihang@hisilicon.com>
> ---
> providers/hns/hns_roce_u.h | 7 +
> providers/hns/hns_roce_u_hw_v2.c | 427 ++++++++++++++++-----------------------
> 2 files changed, 186 insertions(+), 248 deletions(-)
>
No printf() calls in the providers code, please.
Thanks
^ permalink raw reply [flat|nested] 10+ messages in thread
* RE: [PATCH rdma-core 3/5] libhns: Refactor for post send
2019-09-11 7:35 ` Leon Romanovsky
@ 2019-09-11 7:50 ` liweihang
0 siblings, 0 replies; 10+ messages in thread
From: liweihang @ 2019-09-11 7:50 UTC (permalink / raw)
To: Leon Romanovsky; +Cc: dledford, jgg, linux-rdma, Linuxarm
> -----Original Message-----
> From: Leon Romanovsky [mailto:leon@kernel.org]
> Sent: Wednesday, September 11, 2019 3:36 PM
> To: liweihang <liweihang@hisilicon.com>
> Cc: dledford@redhat.com; jgg@ziepe.ca; linux-rdma@vger.kernel.org;
> Linuxarm <linuxarm@huawei.com>
> Subject: Re: [PATCH rdma-core 3/5] libhns: Refactor for post send
>
> On Tue, Sep 10, 2019 at 08:20:50PM +0800, Weihang Li wrote:
> > From: Yixian Liu <liuyixian@huawei.com>
> >
> > This patch refactors the interface of hns_roce_u_v2_post_send, which
> > is now very complicated. We reduce the complexity with following points:
> > 1. Separate RC server into a function.
> > 2. Simplify and separate the process of sge.
> > 3. Keep the logic and consistence of all operations.
> >
> > Signed-off-by: Yixian Liu <liuyixian@huawei.com>
> > Signed-off-by: Weihang Li <liweihang@hisilicon.com>
> > ---
> > providers/hns/hns_roce_u.h | 7 +
> > providers/hns/hns_roce_u_hw_v2.c | 427
> > ++++++++++++++++-----------------------
> > 2 files changed, 186 insertions(+), 248 deletions(-)
> >
>
> No printf() calls in the providers code, please.
>
> Thanks
OK, will modify them, thank you.
^ permalink raw reply [flat|nested] 10+ messages in thread
* [PATCH rdma-core 4/5] libhns: Add UD support for hip08 in user mode
2019-09-10 12:20 [PATCH rdma-core 0/5] Support UD on hip08 Weihang Li
` (2 preceding siblings ...)
2019-09-10 12:20 ` [PATCH rdma-core 3/5] libhns: Refactor for post send Weihang Li
@ 2019-09-10 12:20 ` Weihang Li
2019-09-10 12:20 ` [PATCH rdma-core 5/5] libhns: Support configuring loopback mode by user Weihang Li
4 siblings, 0 replies; 10+ messages in thread
From: Weihang Li @ 2019-09-10 12:20 UTC (permalink / raw)
To: dledford, jgg; +Cc: linux-rdma, linuxarm
From: Lijun Ou <oulijun@huawei.com>
User Application can use Unreliable Datagram on hip08 now.
Signed-off-by: Lijun Ou <oulijun@huawei.com>
Signed-off-by: Weihang Li <liweihang@hisilicon.com>
---
providers/hns/hns_roce_u_hw_v2.c | 90 ++++++++++++++++++++++++++++++++++++++-
providers/hns/hns_roce_u_hw_v2.h | 91 ++++++++++++++++++++++++++++++++++++++++
providers/hns/hns_roce_u_verbs.c | 4 ++
3 files changed, 184 insertions(+), 1 deletion(-)
diff --git a/providers/hns/hns_roce_u_hw_v2.c b/providers/hns/hns_roce_u_hw_v2.c
index 21a6ee0..dd7545f 100644
--- a/providers/hns/hns_roce_u_hw_v2.c
+++ b/providers/hns/hns_roce_u_hw_v2.c
@@ -634,6 +634,92 @@ static void set_sge(struct hns_roce_v2_wqe_data_seg *dseg,
}
}
+static void set_ud_wqe(void *wqe, struct hns_roce_qp *qp,
+ struct ibv_send_wr *wr, int nreq,
+ struct hns_roce_sge_info *sge_info)
+{
+ struct hns_roce_ah *ah = to_hr_ah(wr->wr.ud.ah);
+ struct hns_roce_ud_sq_wqe *ud_sq_wqe = wqe;
+
+ memset(ud_sq_wqe, 0, sizeof(*ud_sq_wqe));
+
+ roce_set_field(ud_sq_wqe->rsv_opcode, UD_SQ_WQE_OPCODE_M,
+ UD_SQ_WQE_OPCODE_S, HNS_ROCE_WQE_OP_SEND);
+ roce_set_bit(ud_sq_wqe->rsv_opcode, UD_SQ_WQE_CQE_S,
+ (wr->send_flags & IBV_SEND_SIGNALED) ? 1 : 0);
+ roce_set_bit(ud_sq_wqe->rsv_opcode, UD_SQ_WQE_SE_S,
+ (wr->send_flags & IBV_SEND_SOLICITED) ? 1 : 0);
+ roce_set_bit(ud_sq_wqe->rsv_opcode, UD_SQ_WQE_OWNER_S,
+ ~(((qp->sq.head + nreq) >> qp->sq.shift) & 0x1));
+
+ ud_sq_wqe->immtdata = htole32(be32toh(wr->imm_data));
+
+ roce_set_field(ud_sq_wqe->sge_num_pd, UD_SQ_WQE_PD_M,
+ UD_SQ_WQE_PD_S, to_hr_pd(qp->ibv_qp.pd)->pdn);
+
+ roce_set_field(ud_sq_wqe->rsv_msg_start_sge_idx,
+ UD_SQ_WQE_MSG_START_SGE_IDX_M,
+ UD_SQ_WQE_MSG_START_SGE_IDX_S,
+ sge_info->start_idx & (qp->sge.sge_cnt - 1));
+
+ roce_set_field(ud_sq_wqe->udpspn_rsv, UD_SQ_WQE_UDP_SPN_M,
+ UD_SQ_WQE_UDP_SPN_S, 0);
+
+ ud_sq_wqe->qkey = htole32(wr->wr.ud.remote_qkey);
+
+ roce_set_field(ud_sq_wqe->rsv_dqpn, UD_SQ_WQE_DQPN_M,
+ UD_SQ_WQE_DQPN_S, wr->wr.ud.remote_qpn);
+
+ roce_set_field(ud_sq_wqe->tclass_vlan, UD_SQ_WQE_VLAN_M,
+ UD_SQ_WQE_VLAN_S, ah->av.vlan_id);
+
+ roce_set_field(ud_sq_wqe->tclass_vlan, UD_SQ_WQE_TCLASS_M,
+ UD_SQ_WQE_TCLASS_S, ah->av.tclass);
+
+ roce_set_field(ud_sq_wqe->tclass_vlan, UD_SQ_WQE_HOPLIMIT_M,
+ UD_SQ_WQE_HOPLIMIT_S, ah->av.hop_limit);
+
+ roce_set_field(ud_sq_wqe->lbi_flow_label, UD_SQ_WQE_FLOW_LABEL_M,
+ UD_SQ_WQE_FLOW_LABEL_S, ah->av.flowlabel);
+
+ roce_set_bit(ud_sq_wqe->lbi_flow_label, UD_SQ_WQE_VLAN_EN_S,
+ ah->av.vlan_en ? 1 : 0);
+
+ roce_set_bit(ud_sq_wqe->lbi_flow_label, UD_SQ_WQE_LBI_S, 0);
+
+ roce_set_field(ud_sq_wqe->lbi_flow_label, UD_SQ_WQE_SL_M,
+ UD_SQ_WQE_SL_S, ah->av.sl);
+
+ roce_set_field(ud_sq_wqe->lbi_flow_label, UD_SQ_WQE_PORTN_M,
+ UD_SQ_WQE_PORTN_S, qp->ibv_qp.qp_num);
+
+ roce_set_field(ud_sq_wqe->dmac, UD_SQ_WQE_DMAC_0_M,
+ UD_SQ_WQE_DMAC_0_S, ah->av.mac[0]);
+ roce_set_field(ud_sq_wqe->dmac, UD_SQ_WQE_DMAC_1_M,
+ UD_SQ_WQE_DMAC_1_S, ah->av.mac[1]);
+ roce_set_field(ud_sq_wqe->dmac, UD_SQ_WQE_DMAC_2_M,
+ UD_SQ_WQE_DMAC_2_S, ah->av.mac[2]);
+ roce_set_field(ud_sq_wqe->dmac, UD_SQ_WQE_DMAC_3_M,
+ UD_SQ_WQE_DMAC_3_S, ah->av.mac[3]);
+ roce_set_field(ud_sq_wqe->smac_index_dmac,
+ UD_SQ_WQE_DMAC_4_M, UD_SQ_WQE_DMAC_4_S,
+ to_hr_ah(wr->wr.ud.ah)->av.mac[4]);
+ roce_set_field(ud_sq_wqe->smac_index_dmac,
+ UD_SQ_WQE_DMAC_5_M, UD_SQ_WQE_DMAC_5_S,
+ to_hr_ah(wr->wr.ud.ah)->av.mac[5]);
+ roce_set_field(ud_sq_wqe->smac_index_dmac, UD_SQ_WQE_SGID_IDX_M,
+ UD_SQ_WQE_SGID_IDX_S, ah->av.gid_index);
+
+ memcpy(ud_sq_wqe->dgid, ah->av.dgid, HNS_ROCE_GID_SIZE);
+
+ set_sge(NULL, qp, wr, sge_info);
+
+ ud_sq_wqe->msg_len = htole32(sge_info->total_len);
+
+ roce_set_field(ud_sq_wqe->sge_num_pd, UD_SQ_WQE_SGE_NUM_M,
+ UD_SQ_WQE_SGE_NUM_S, sge_info->valid_num);
+}
+
static int set_rc_wqe(void *wqe, struct hns_roce_qp *qp, struct ibv_send_wr *wr,
int nreq, struct hns_roce_sge_info *sge_info)
{
@@ -827,8 +913,10 @@ int hns_roce_u_v2_post_send(struct ibv_qp *ibvqp, struct ibv_send_wr *wr,
goto out;
}
break;
- case IBV_QPT_UC:
case IBV_QPT_UD:
+ set_ud_wqe(wqe, qp, wr, nreq, &sge_info);
+ break;
+ case IBV_QPT_UC:
default:
ret = -EINVAL;
*bad_wr = wr;
diff --git a/providers/hns/hns_roce_u_hw_v2.h b/providers/hns/hns_roce_u_hw_v2.h
index 3eca1fd..84cf6c4 100644
--- a/providers/hns/hns_roce_u_hw_v2.h
+++ b/providers/hns/hns_roce_u_hw_v2.h
@@ -291,4 +291,95 @@ struct hns_roce_wqe_atomic_seg {
int hns_roce_u_v2_post_send(struct ibv_qp *ibvqp, struct ibv_send_wr *wr,
struct ibv_send_wr **bad_wr);
+struct hns_roce_ud_sq_wqe {
+ __le32 rsv_opcode;
+ __le32 msg_len;
+ __le32 immtdata;
+ __le32 sge_num_pd;
+ __le32 rsv_msg_start_sge_idx;
+ __le32 udpspn_rsv;
+ __le32 qkey;
+ __le32 rsv_dqpn;
+ __le32 tclass_vlan;
+ __le32 lbi_flow_label;
+ __le32 dmac;
+ __le32 smac_index_dmac;
+ uint8_t dgid[HNS_ROCE_GID_SIZE];
+};
+
+#define UD_SQ_WQE_OPCODE_S 0
+#define UD_SQ_WQE_OPCODE_M (((1UL << 5) - 1) << UD_SQ_WQE_OPCODE_S)
+
+#define UD_SQ_WQE_OWNER_S 7
+
+#define UD_SQ_WQE_CQE_S 8
+
+#define UD_SQ_WQE_SE_S 11
+
+#define UD_SQ_WQE_PD_S 0
+#define UD_SQ_WQE_PD_M (((1UL << 24) - 1) << UD_SQ_WQE_PD_S)
+
+#define UD_SQ_WQE_SGE_NUM_S 24
+#define UD_SQ_WQE_SGE_NUM_M (((1UL << 8) - 1) << UD_SQ_WQE_SGE_NUM_S)
+
+#define UD_SQ_WQE_MSG_START_SGE_IDX_S 0
+#define UD_SQ_WQE_MSG_START_SGE_IDX_M \
+ (((1UL << 24) - 1) << UD_SQ_WQE_MSG_START_SGE_IDX_S)
+
+#define UD_SQ_WQE_UDP_SPN_S 16
+#define UD_SQ_WQE_UDP_SPN_M \
+ (((1UL << 16) - 1) << UD_SQ_WQE_UDP_SPN_S)
+
+#define UD_SQ_WQE_DQPN_S 0
+#define UD_SQ_WQE_DQPN_M (((1UL << 24) - 1) << UD_SQ_WQE_DQPN_S)
+
+#define UD_SQ_WQE_VLAN_S 0
+#define UD_SQ_WQE_VLAN_M (((1UL << 16) - 1) << UD_SQ_WQE_VLAN_S)
+
+#define UD_SQ_WQE_HOPLIMIT_S 16
+#define UD_SQ_WQE_HOPLIMIT_M (((1UL << 8) - 1) << UD_SQ_WQE_HOPLIMIT_S)
+
+#define UD_SQ_WQE_TCLASS_S 24
+#define UD_SQ_WQE_TCLASS_M (((1UL << 8) - 1) << UD_SQ_WQE_TCLASS_S)
+
+#define UD_SQ_WQE_FLOW_LABEL_S 0
+#define UD_SQ_WQE_FLOW_LABEL_M (((1UL << 20) - 1) << UD_SQ_WQE_FLOW_LABEL_S)
+
+#define UD_SQ_WQE_SL_S 20
+#define UD_SQ_WQE_SL_M (((1UL << 4) - 1) << UD_SQ_WQE_SL_S)
+
+#define UD_SQ_WQE_PORTN_S 24
+#define UD_SQ_WQE_PORTN_M (((1UL << 3) - 1) << UD_SQ_WQE_PORTN_S)
+
+#define UD_SQ_WQE_VLAN_EN_S 30
+
+#define UD_SQ_WQE_LBI_S 31
+
+#define UD_SQ_WQE_PORTN_S 24
+#define UD_SQ_WQE_PORTN_M (((1UL << 3) - 1) << UD_SQ_WQE_PORTN_S)
+
+#define UD_SQ_WQE_DMAC_0_S 0
+#define UD_SQ_WQE_DMAC_0_M (((1UL << 8) - 1) << UD_SQ_WQE_DMAC_0_S)
+
+#define UD_SQ_WQE_DMAC_1_S 8
+#define UD_SQ_WQE_DMAC_1_M (((1UL << 8) - 1) << UD_SQ_WQE_DMAC_1_S)
+
+#define UD_SQ_WQE_DMAC_2_S 16
+#define UD_SQ_WQE_DMAC_2_M (((1UL << 8) - 1) << UD_SQ_WQE_DMAC_2_S)
+
+#define UD_SQ_WQE_DMAC_3_S 24
+#define UD_SQ_WQE_DMAC_3_M (((1UL << 8) - 1) << UD_SQ_WQE_DMAC_3_S)
+
+#define UD_SQ_WQE_DMAC_4_S 0
+#define UD_SQ_WQE_DMAC_4_M (((1UL << 8) - 1) << UD_SQ_WQE_DMAC_4_S)
+
+#define UD_SQ_WQE_DMAC_5_S 8
+#define UD_SQ_WQE_DMAC_5_M (((1UL << 8) - 1) << UD_SQ_WQE_DMAC_5_S)
+
+#define UD_SQ_WQE_SGID_IDX_S 16
+#define UD_SQ_WQE_SGID_IDX_M (((1UL << 8) - 1) << UD_SQ_WQE_SGID_IDX_S)
+
+#define UD_SQ_WQE_SMAC_IDX_S 24
+#define UD_SQ_WQE_SMAC_IDX_M (((1UL << 8) - 1) << UD_SQ_WQE_SMAC_IDX_S)
+
#endif /* _HNS_ROCE_U_HW_V2_H */
diff --git a/providers/hns/hns_roce_u_verbs.c b/providers/hns/hns_roce_u_verbs.c
index 0d10b1c..5255fff 100644
--- a/providers/hns/hns_roce_u_verbs.c
+++ b/providers/hns/hns_roce_u_verbs.c
@@ -767,6 +767,10 @@ static void hns_roce_set_qp_params(struct ibv_pd *pd,
}
}
+ if (attr->qp_type == IBV_QPT_UD)
+ qp->sge.sge_cnt = roundup_pow_of_two(qp->sq.wqe_cnt *
+ qp->sq.max_gs);
+
/* limit by the context queried during alloc context */
qp->rq.max_post = min(ctx->max_qp_wr, qp->rq.wqe_cnt);
qp->sq.max_post = min(ctx->max_qp_wr, qp->sq.wqe_cnt);
--
2.8.1
^ permalink raw reply related [flat|nested] 10+ messages in thread* [PATCH rdma-core 5/5] libhns: Support configuring loopback mode by user
2019-09-10 12:20 [PATCH rdma-core 0/5] Support UD on hip08 Weihang Li
` (3 preceding siblings ...)
2019-09-10 12:20 ` [PATCH rdma-core 4/5] libhns: Add UD support for hip08 in user mode Weihang Li
@ 2019-09-10 12:20 ` Weihang Li
2019-09-11 7:42 ` Leon Romanovsky
4 siblings, 1 reply; 10+ messages in thread
From: Weihang Li @ 2019-09-10 12:20 UTC (permalink / raw)
To: dledford, jgg; +Cc: linux-rdma, linuxarm
User can configure whether hardware working on loopback mode or not
by export an environment variable "HNS_ROCE_LOOPBACK".
Signed-off-by: Weihang Li <liweihang@hisilicon.com>
---
providers/hns/hns_roce_u.c | 15 +++++++++++++++
providers/hns/hns_roce_u.h | 2 ++
providers/hns/hns_roce_u_hw_v2.c | 3 ++-
providers/hns/hns_roce_u_verbs.c | 1 +
4 files changed, 20 insertions(+), 1 deletion(-)
diff --git a/providers/hns/hns_roce_u.c b/providers/hns/hns_roce_u.c
index 8ba41de..f68c84b 100644
--- a/providers/hns/hns_roce_u.c
+++ b/providers/hns/hns_roce_u.c
@@ -40,6 +40,8 @@
#include "hns_roce_u.h"
#include "hns_roce_u_abi.h"
+bool loopback;
+
#define HID_LEN 15
#define DEV_MATCH_LEN 128
@@ -85,6 +87,17 @@ static const struct verbs_context_ops hns_common_ops = {
.destroy_ah = hns_roce_u_destroy_ah,
};
+static bool get_param_config(const char *var)
+{
+ char *env;
+
+ env = getenv(var);
+ if (env)
+ return strcmp(env, "true") ? true : false;
+
+ return false;
+}
+
static struct verbs_context *hns_roce_alloc_context(struct ibv_device *ibdev,
int cmd_fd,
void *private_data)
@@ -105,6 +118,8 @@ static struct verbs_context *hns_roce_alloc_context(struct ibv_device *ibdev,
&resp.ibv_resp, sizeof(resp)))
goto err_free;
+ loopback = get_param_config("HNS_ROCE_LOOPBACK");
+
context->num_qps = resp.qp_tab_size;
context->qp_table_shift = ffs(context->num_qps) - 1 -
HNS_ROCE_QP_TABLE_BITS;
diff --git a/providers/hns/hns_roce_u.h b/providers/hns/hns_roce_u.h
index 9d8f72e..624bdaa 100644
--- a/providers/hns/hns_roce_u.h
+++ b/providers/hns/hns_roce_u.h
@@ -43,6 +43,7 @@
#include <ccan/bitmap.h>
#include <ccan/container_of.h>
+extern bool loopback;
#define HNS_ROCE_HW_VER1 ('h' << 24 | 'i' << 16 | '0' << 8 | '6')
#define HNS_ROCE_HW_VER2 ('h' << 24 | 'i' << 16 | '0' << 8 | '8')
@@ -265,6 +266,7 @@ struct hns_roce_av {
uint8_t mac[ETH_ALEN];
uint16_t vlan_id;
uint8_t vlan_en;
+ uint8_t loopback;
};
struct hns_roce_ah {
diff --git a/providers/hns/hns_roce_u_hw_v2.c b/providers/hns/hns_roce_u_hw_v2.c
index dd7545f..62e8561 100644
--- a/providers/hns/hns_roce_u_hw_v2.c
+++ b/providers/hns/hns_roce_u_hw_v2.c
@@ -685,7 +685,8 @@ static void set_ud_wqe(void *wqe, struct hns_roce_qp *qp,
roce_set_bit(ud_sq_wqe->lbi_flow_label, UD_SQ_WQE_VLAN_EN_S,
ah->av.vlan_en ? 1 : 0);
- roce_set_bit(ud_sq_wqe->lbi_flow_label, UD_SQ_WQE_LBI_S, 0);
+ roce_set_bit(ud_sq_wqe->lbi_flow_label, UD_SQ_WQE_LBI_S,
+ ah->av.loopback);
roce_set_field(ud_sq_wqe->lbi_flow_label, UD_SQ_WQE_SL_M,
UD_SQ_WQE_SL_S, ah->av.sl);
diff --git a/providers/hns/hns_roce_u_verbs.c b/providers/hns/hns_roce_u_verbs.c
index 5255fff..ff24ae3 100644
--- a/providers/hns/hns_roce_u_verbs.c
+++ b/providers/hns/hns_roce_u_verbs.c
@@ -970,6 +970,7 @@ struct ibv_ah *hns_roce_u_create_ah(struct ibv_pd *pd, struct ibv_ah_attr *attr)
ah->av.port = attr->port_num;
ah->av.sl = attr->sl;
+ ah->av.loopback = loopback ? 1 : 0;
if (attr->static_rate)
ah->av.static_rate = IBV_RATE_10_GBPS;
--
2.8.1
^ permalink raw reply related [flat|nested] 10+ messages in thread* Re: [PATCH rdma-core 5/5] libhns: Support configuring loopback mode by user
2019-09-10 12:20 ` [PATCH rdma-core 5/5] libhns: Support configuring loopback mode by user Weihang Li
@ 2019-09-11 7:42 ` Leon Romanovsky
2019-09-11 12:40 ` liweihang
0 siblings, 1 reply; 10+ messages in thread
From: Leon Romanovsky @ 2019-09-11 7:42 UTC (permalink / raw)
To: Weihang Li; +Cc: dledford, jgg, linux-rdma, linuxarm
On Tue, Sep 10, 2019 at 08:20:52PM +0800, Weihang Li wrote:
> User can configure whether hardware working on loopback mode or not
> by export an environment variable "HNS_ROCE_LOOPBACK".
It is definitely wrong interface to configure behaviour of application.
Environment variables make sense if you need to change library
behaviour.
Thanks
^ permalink raw reply [flat|nested] 10+ messages in thread
* RE: [PATCH rdma-core 5/5] libhns: Support configuring loopback mode by user
2019-09-11 7:42 ` Leon Romanovsky
@ 2019-09-11 12:40 ` liweihang
0 siblings, 0 replies; 10+ messages in thread
From: liweihang @ 2019-09-11 12:40 UTC (permalink / raw)
To: Leon Romanovsky; +Cc: dledford, jgg, linux-rdma, Linuxarm
> -----Original Message-----
> From: Leon Romanovsky [mailto:leon@kernel.org]
> Sent: Wednesday, September 11, 2019 3:42 PM
> To: liweihang <liweihang@hisilicon.com>
> Cc: dledford@redhat.com; jgg@ziepe.ca; linux-rdma@vger.kernel.org;
> Linuxarm <linuxarm@huawei.com>
> Subject: Re: [PATCH rdma-core 5/5] libhns: Support configuring loopback
> mode by user
>
> On Tue, Sep 10, 2019 at 08:20:52PM +0800, Weihang Li wrote:
> > User can configure whether hardware working on loopback mode or not by
> > export an environment variable "HNS_ROCE_LOOPBACK".
>
> It is definitely wrong interface to configure behaviour of application.
> Environment variables make sense if you need to change library behaviour.
>
> Thanks
Hi Leon,
Could you please give some advices on how to get configurations from users?
Thanks,
Weihang
^ permalink raw reply [flat|nested] 10+ messages in thread