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

Update to 41.0

Browse Source 
Update rdma-core version from 35.1 to 41.0.

Version 41.0 is the latest version in community until
2022/7/27. It includes some new bugfixes and new features,
we choose this version to facilitate future development.

The patches added to this repo has already included in the
new version, so remove them.

Signed-off-by: Chengchang Tang <tangchengchang@huawei.com>
This commit is contained in:
Chengchang Tang 2022-07-27 18:14:12 +08:00 committed by Chengchang Tang
parent 9908c135be
commit 060786cc7e
20 changed files with 8 additions and 1942 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

41
0001-Update-kernel-headers.patch
View File

@ -1,41 +0,0 @@
From 693d55e80976217215844258e5b78bc115382689 Mon Sep 17 00:00:00 2001
From: Guofeng Yue <yueguofeng@hisilicon.com>
Date: Mon, 10 Jan 2022 10:44:23 +0800
Subject: [PATCH 1/8] Update kernel headers
To commit 62c4d8878d13 ("RDMA/hns: Remove support for HIP06").
Signed-off-by: Guofeng Yue <yueguofeng@hisilicon.com>
---
kernel-headers/rdma/hns-abi.h | 6 ++++--
1 file changed, 4 insertions(+), 2 deletions(-)
diff --git a/kernel-headers/rdma/hns-abi.h b/kernel-headers/rdma/hns-abi.h
index 42b17765..abfd36e2 100644
--- a/kernel-headers/rdma/hns-abi.h
+++ b/kernel-headers/rdma/hns-abi.h
@@ -77,17 +77,19 @@ enum hns_roce_qp_cap_flags {
HNS_ROCE_QP_CAP_RQ_RECORD_DB = 1 << 0,
HNS_ROCE_QP_CAP_SQ_RECORD_DB = 1 << 1,
HNS_ROCE_QP_CAP_OWNER_DB = 1 << 2,
+ HNS_ROCE_QP_CAP_DIRECT_WQE = 1 << 5,
};
struct hns_roce_ib_create_qp_resp {
__aligned_u64 cap_flags;
+ __aligned_u64 dwqe_mmap_key;
};
struct hns_roce_ib_alloc_ucontext_resp {
__u32 qp_tab_size;
__u32 cqe_size;
- __u32 srq_tab_size;
- __u32 reserved;
+ __u32 srq_tab_size;
+ __u32 reserved;
};
struct hns_roce_ib_alloc_pd_resp {
--
2.33.0

120
0002-libhns-Fix-the-ownership-of-the-head-tail-pointer-of.patch
View File

@ -1,120 +0,0 @@
From 08ec3c43bf9710fdf3ca664f7cd63436e67339d7 Mon Sep 17 00:00:00 2001
From: Wenpeng Liang <liangwenpeng@huawei.com>
Date: Tue, 11 May 2021 19:06:34 +0800
Subject: [PATCH 2/8] libhns: Fix the ownership of the head/tail pointer of SRQ
WQE
The CQE of SRQ is not generated in the order of wqe, so the wqe_idx
corresponding to the idle WQE should be placed in a FIFO, then the hardware
will be instructed to obtain the corresponding WQE. Therefore, the WQ
of SRQ has no concept of head pointer and tail pointer, but the queue of
wqe_idx does.
Signed-off-by: Wenpeng Liang <liangwenpeng@huawei.com>
Signed-off-by: Weihang Li <liweihang@huawei.com>
---
providers/hns/hns_roce_u.h | 4 ++--
providers/hns/hns_roce_u_hw_v2.c | 12 ++++++------
providers/hns/hns_roce_u_verbs.c | 6 +++---
3 files changed, 11 insertions(+), 11 deletions(-)
diff --git a/providers/hns/hns_roce_u.h b/providers/hns/hns_roce_u.h
index 8f805dd1..b3f48113 100644
--- a/providers/hns/hns_roce_u.h
+++ b/providers/hns/hns_roce_u.h
@@ -205,6 +205,8 @@ struct hns_roce_idx_que {
int entry_shift;
unsigned long *bitmap;
int bitmap_cnt;
+ unsigned int head;
+ unsigned int tail;
};
struct hns_roce_srq {
@@ -217,8 +219,6 @@ struct hns_roce_srq {
unsigned int max_gs;
unsigned int rsv_sge;
unsigned int wqe_shift;
- int head;
- int tail;
unsigned int *db;
unsigned short counter;
struct hns_roce_idx_que idx_que;
diff --git a/providers/hns/hns_roce_u_hw_v2.c b/providers/hns/hns_roce_u_hw_v2.c
index 4988943a..f947dbd7 100644
--- a/providers/hns/hns_roce_u_hw_v2.c
+++ b/providers/hns/hns_roce_u_hw_v2.c
@@ -262,7 +262,7 @@ static void hns_roce_free_srq_wqe(struct hns_roce_srq *srq, uint16_t ind)
bitmap_num = ind / BIT_CNT_PER_LONG;
bit_num = ind % BIT_CNT_PER_LONG;
srq->idx_que.bitmap[bitmap_num] |= (1ULL << bit_num);
- srq->tail++;
+ srq->idx_que.tail++;
pthread_spin_unlock(&srq->lock);
}
@@ -1564,7 +1564,7 @@ static int hns_roce_u_v2_post_srq_recv(struct ibv_srq *ib_srq,
pthread_spin_lock(&srq->lock);
/* current idx of srqwq */
- ind = srq->head & (srq->wqe_cnt - 1);
+ ind = srq->idx_que.head & (srq->wqe_cnt - 1);
max_sge = srq->max_gs - srq->rsv_sge;
for (nreq = 0; wr; ++nreq, wr = wr->next) {
@@ -1574,7 +1574,7 @@ static int hns_roce_u_v2_post_srq_recv(struct ibv_srq *ib_srq,
break;
}
- if (srq->head == srq->tail) {
+ if (srq->idx_que.head == srq->idx_que.tail) {
ret = -ENOMEM;
*bad_wr = wr;
break;
@@ -1607,7 +1607,7 @@ static int hns_roce_u_v2_post_srq_recv(struct ibv_srq *ib_srq,
}
if (nreq) {
- srq->head += nreq;
+ srq->idx_que.head += nreq;
/*
* Make sure that descriptors are written before
@@ -1617,8 +1617,8 @@ static int hns_roce_u_v2_post_srq_recv(struct ibv_srq *ib_srq,
srq_db.byte_4 = htole32(HNS_ROCE_V2_SRQ_DB << DB_BYTE_4_CMD_S |
srq->srqn);
- srq_db.parameter =
- htole32(srq->head & DB_PARAM_SRQ_PRODUCER_COUNTER_M);
+ srq_db.parameter = htole32(srq->idx_que.head &
+ DB_PARAM_SRQ_PRODUCER_COUNTER_M);
hns_roce_write64((uint32_t *)&srq_db, ctx,
ROCEE_VF_DB_CFG0_OFFSET);
diff --git a/providers/hns/hns_roce_u_verbs.c b/providers/hns/hns_roce_u_verbs.c
index 30ab072a..9b4934b9 100644
--- a/providers/hns/hns_roce_u_verbs.c
+++ b/providers/hns/hns_roce_u_verbs.c
@@ -491,6 +491,9 @@ static int hns_roce_create_idx_que(struct hns_roce_srq *srq)
for (i = 0; i < idx_que->bitmap_cnt; ++i)
idx_que->bitmap[i] = ~(0UL);
+ idx_que->head = 0;
+ idx_que->tail = srq->wqe_cnt - 1;
+
return 0;
}
@@ -512,9 +515,6 @@ static int hns_roce_alloc_srq_buf(struct hns_roce_srq *srq)
return ENOMEM;
}
- srq->head = 0;
- srq->tail = srq->wqe_cnt - 1;
-
return 0;
}
--
2.33.0

180
0003-libhns-Fix-wrong-data-type-when-writing-doorbell.patch
View File

@ -1,180 +0,0 @@
From 9cc4c4b8d31b35428859ef626d4428fc393aace4 Mon Sep 17 00:00:00 2001
From: Lang Cheng <chenglang@huawei.com>
Date: Thu, 11 Nov 2021 21:08:35 +0800
Subject: [PATCH 3/8] libhns: Fix wrong data type when writing doorbell
The DB data is a __le32[] value instead of uint32_t[], and the DB register
should be written with a little-endian data instead of uint64_t.
Fixes: 1523fbb1ea8e ("libhns: Add verbs of cq support")
Signed-off-by: Lang Cheng <chenglang@huawei.com>
Signed-off-by: Yixing Liu <liuyixing1@huawei.com>
Signed-off-by: Wenpeng Liang <liangwenpeng@huawei.com>
---
providers/hns/hns_roce_u_db.h | 14 ++++----------
providers/hns/hns_roce_u_hw_v1.c | 17 +++++++++--------
providers/hns/hns_roce_u_hw_v2.c | 23 ++++++++++++-----------
3 files changed, 25 insertions(+), 29 deletions(-)
diff --git a/providers/hns/hns_roce_u_db.h b/providers/hns/hns_roce_u_db.h
index b44e64d4..13df9b52 100644
--- a/providers/hns/hns_roce_u_db.h
+++ b/providers/hns/hns_roce_u_db.h
@@ -32,23 +32,17 @@
#include <linux/types.h>
+#include <util/mmio.h>
#include "hns_roce_u.h"
#ifndef _HNS_ROCE_U_DB_H
#define _HNS_ROCE_U_DB_H
-#if __BYTE_ORDER == __LITTLE_ENDIAN
-#define HNS_ROCE_PAIR_TO_64(val) ((uint64_t) val[1] << 32 | val[0])
-#elif __BYTE_ORDER == __BIG_ENDIAN
-#define HNS_ROCE_PAIR_TO_64(val) ((uint64_t) val[0] << 32 | val[1])
-#else
-#error __BYTE_ORDER not defined
-#endif
+#define HNS_ROCE_WORD_NUM 2
-static inline void hns_roce_write64(uint32_t val[2],
- struct hns_roce_context *ctx, int offset)
+static inline void hns_roce_write64(void *dest, __le32 val[HNS_ROCE_WORD_NUM])
{
- *(volatile uint64_t *) (ctx->uar + offset) = HNS_ROCE_PAIR_TO_64(val);
+ mmio_write64_le(dest, *(__le64 *)val);
}
void *hns_roce_alloc_db(struct hns_roce_context *ctx,
diff --git a/providers/hns/hns_roce_u_hw_v1.c b/providers/hns/hns_roce_u_hw_v1.c
index 8f0a71aa..14ee4817 100644
--- a/providers/hns/hns_roce_u_hw_v1.c
+++ b/providers/hns/hns_roce_u_hw_v1.c
@@ -65,7 +65,7 @@ static void hns_roce_update_rq_head(struct hns_roce_context *ctx,
udma_to_device_barrier();
- hns_roce_write64((uint32_t *)&rq_db, ctx, ROCEE_DB_OTHERS_L_0_REG);
+ hns_roce_write64(ctx->uar + ROCEE_DB_OTHERS_L_0_REG, (__le32 *)&rq_db);
}
static void hns_roce_update_sq_head(struct hns_roce_context *ctx,
@@ -84,7 +84,7 @@ static void hns_roce_update_sq_head(struct hns_roce_context *ctx,
udma_to_device_barrier();
- hns_roce_write64((uint32_t *)&sq_db, ctx, ROCEE_DB_SQ_L_0_REG);
+ hns_roce_write64(ctx->uar + ROCEE_DB_SQ_L_0_REG, (__le32 *)&sq_db);
}
static void hns_roce_update_cq_cons_index(struct hns_roce_context *ctx,
@@ -102,7 +102,7 @@ static void hns_roce_update_cq_cons_index(struct hns_roce_context *ctx,
CQ_DB_U32_4_CONS_IDX_S,
cq->cons_index & ((cq->cq_depth << 1) - 1));
- hns_roce_write64((uint32_t *)&cq_db, ctx, ROCEE_DB_OTHERS_L_0_REG);
+ hns_roce_write64(ctx->uar + ROCEE_DB_OTHERS_L_0_REG, (__le32 *)&cq_db);
}
static void hns_roce_handle_error_cqe(struct hns_roce_cqe *cqe,
@@ -422,10 +422,11 @@ static int hns_roce_u_v1_poll_cq(struct ibv_cq *ibvcq, int ne,
*/
static int hns_roce_u_v1_arm_cq(struct ibv_cq *ibvcq, int solicited)
{
- uint32_t ci;
- uint32_t solicited_flag;
- struct hns_roce_cq_db cq_db = {};
+ struct hns_roce_context *ctx = to_hr_ctx(ibvcq->context);
struct hns_roce_cq *cq = to_hr_cq(ibvcq);
+ struct hns_roce_cq_db cq_db = {};
+ uint32_t solicited_flag;
+ uint32_t ci;
ci = cq->cons_index & ((cq->cq_depth << 1) - 1);
solicited_flag = solicited ? HNS_ROCE_CQ_DB_REQ_SOL :
@@ -441,8 +442,8 @@ static int hns_roce_u_v1_arm_cq(struct ibv_cq *ibvcq, int solicited)
roce_set_field(cq_db.u32_4, CQ_DB_U32_4_CONS_IDX_M,
CQ_DB_U32_4_CONS_IDX_S, ci);
- hns_roce_write64((uint32_t *)&cq_db, to_hr_ctx(ibvcq->context),
- ROCEE_DB_OTHERS_L_0_REG);
+ hns_roce_write64(ctx->uar + ROCEE_DB_OTHERS_L_0_REG, (__le32 *)&cq_db);
+
return 0;
}
diff --git a/providers/hns/hns_roce_u_hw_v2.c b/providers/hns/hns_roce_u_hw_v2.c
index f947dbd7..efd949f4 100644
--- a/providers/hns/hns_roce_u_hw_v2.c
+++ b/providers/hns/hns_roce_u_hw_v2.c
@@ -293,7 +293,7 @@ static void hns_roce_update_rq_db(struct hns_roce_context *ctx,
HNS_ROCE_V2_RQ_DB);
rq_db.parameter = htole32(rq_head);
- hns_roce_write64((uint32_t *)&rq_db, ctx, ROCEE_VF_DB_CFG0_OFFSET);
+ hns_roce_write64(ctx->uar + ROCEE_VF_DB_CFG0_OFFSET, (__le32 *)&rq_db);
}
static void hns_roce_update_sq_db(struct hns_roce_context *ctx,
@@ -308,7 +308,7 @@ static void hns_roce_update_sq_db(struct hns_roce_context *ctx,
sq_db.parameter = htole32(sq_head);
roce_set_field(sq_db.parameter, DB_PARAM_SL_M, DB_PARAM_SL_S, sl);
- hns_roce_write64((uint32_t *)&sq_db, ctx, ROCEE_VF_DB_CFG0_OFFSET);
+ hns_roce_write64(ctx->uar + ROCEE_VF_DB_CFG0_OFFSET, (__le32 *)&sq_db);
}
static void hns_roce_v2_update_cq_cons_index(struct hns_roce_context *ctx,
@@ -325,7 +325,7 @@ static void hns_roce_v2_update_cq_cons_index(struct hns_roce_context *ctx,
roce_set_field(cq_db.parameter, DB_PARAM_CQ_CMD_SN_M,
DB_PARAM_CQ_CMD_SN_S, 1);
- hns_roce_write64((uint32_t *)&cq_db, ctx, ROCEE_VF_DB_CFG0_OFFSET);
+ hns_roce_write64(ctx->uar + ROCEE_VF_DB_CFG0_OFFSET, (__le32 *)&cq_db);
}
static struct hns_roce_qp *hns_roce_v2_find_qp(struct hns_roce_context *ctx,
@@ -659,11 +659,12 @@ static int hns_roce_u_v2_poll_cq(struct ibv_cq *ibvcq, int ne,
static int hns_roce_u_v2_arm_cq(struct ibv_cq *ibvcq, int solicited)
{
- uint32_t ci;
- uint32_t cmd_sn;
- uint32_t solicited_flag;
- struct hns_roce_db cq_db = {};
+ struct hns_roce_context *ctx = to_hr_ctx(ibvcq->context);
struct hns_roce_cq *cq = to_hr_cq(ibvcq);
+ struct hns_roce_db cq_db = {};
+ uint32_t solicited_flag;
+ uint32_t cmd_sn;
+ uint32_t ci;
ci = cq->cons_index & ((cq->cq_depth << 1) - 1);
cmd_sn = cq->arm_sn & HNS_ROCE_CMDSN_MASK;
@@ -681,8 +682,8 @@ static int hns_roce_u_v2_arm_cq(struct ibv_cq *ibvcq, int solicited)
DB_PARAM_CQ_CMD_SN_S, cmd_sn);
roce_set_bit(cq_db.parameter, DB_PARAM_CQ_NOTIFY_S, solicited_flag);
- hns_roce_write64((uint32_t *)&cq_db, to_hr_ctx(ibvcq->context),
- ROCEE_VF_DB_CFG0_OFFSET);
+ hns_roce_write64(ctx->uar + ROCEE_VF_DB_CFG0_OFFSET, (__le32 *)&cq_db);
+
return 0;
}
@@ -1620,8 +1621,8 @@ static int hns_roce_u_v2_post_srq_recv(struct ibv_srq *ib_srq,
srq_db.parameter = htole32(srq->idx_que.head &
DB_PARAM_SRQ_PRODUCER_COUNTER_M);
- hns_roce_write64((uint32_t *)&srq_db, ctx,
- ROCEE_VF_DB_CFG0_OFFSET);
+ hns_roce_write64(ctx->uar + ROCEE_VF_DB_CFG0_OFFSET,
+ (__le32 *)&srq_db);
}
pthread_spin_unlock(&srq->lock);
--
2.33.0

43
0004-libhns-Remove-unsupported-QP-type.patch
View File

@ -1,43 +0,0 @@
From 60d82566fc94b11280be26733bc306e6af3d2697 Mon Sep 17 00:00:00 2001
From: Wenpeng Liang <liangwenpeng@huawei.com>
Date: Tue, 9 Nov 2021 20:40:58 +0800
Subject: [PATCH 4/8] libhns: Remove unsupported QP type
Currently, user space does not support UC type QP.
Signed-off-by: Wenpeng Liang <liangwenpeng@huawei.com>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
---
providers/hns/hns_roce_u_hw_v1.c | 1 -
providers/hns/hns_roce_u_hw_v2.c | 3 +--
2 files changed, 1 insertion(+), 3 deletions(-)
diff --git a/providers/hns/hns_roce_u_hw_v1.c b/providers/hns/hns_roce_u_hw_v1.c
index 14ee4817..279c9b0f 100644
--- a/providers/hns/hns_roce_u_hw_v1.c
+++ b/providers/hns/hns_roce_u_hw_v1.c
@@ -532,7 +532,6 @@ static int hns_roce_u_v1_post_send(struct ibv_qp *ibvqp, struct ibv_send_wr *wr,
ctrl->flag |= htole32(ps_opcode);
wqe += sizeof(struct hns_roce_wqe_raddr_seg);
break;
- case IBV_QPT_UC:
case IBV_QPT_UD:
default:
break;
diff --git a/providers/hns/hns_roce_u_hw_v2.c b/providers/hns/hns_roce_u_hw_v2.c
index efd949f4..c62f74b5 100644
--- a/providers/hns/hns_roce_u_hw_v2.c
+++ b/providers/hns/hns_roce_u_hw_v2.c
@@ -460,8 +460,7 @@ static int hns_roce_handle_recv_inl_wqe(struct hns_roce_v2_cqe *cqe,
struct hns_roce_qp **cur_qp,
struct ibv_wc *wc, uint32_t opcode)
{
- if (((*cur_qp)->verbs_qp.qp.qp_type == IBV_QPT_RC ||
- (*cur_qp)->verbs_qp.qp.qp_type == IBV_QPT_UC) &&
+ 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) &&
--
2.33.0

67
0005-libhns-Avoid-using-WQE-indexes-that-exceed-the-SRQ-s.patch
View File

@ -1,67 +0,0 @@
From e460a4208d1821b1477e621ad5a7b72068e844f9 Mon Sep 17 00:00:00 2001
From: Wenpeng Liang <liangwenpeng@huawei.com>
Date: Tue, 11 May 2021 19:06:32 +0800
Subject: [PATCH 5/8] libhns: Avoid using WQE indexes that exceed the SRQ size
The index of SRQ WQE got from bitmap may be greater than the capability,
so a check for that should be added.
Signed-off-by: Wenpeng Liang <liangwenpeng@huawei.com>
Signed-off-by: Weihang Li <liweihang@huawei.com>
---
providers/hns/hns_roce_u_hw_v2.c | 20 ++++++++++++++------
1 file changed, 14 insertions(+), 6 deletions(-)
diff --git a/providers/hns/hns_roce_u_hw_v2.c b/providers/hns/hns_roce_u_hw_v2.c
index c62f74b5..1169b64b 100644
--- a/providers/hns/hns_roce_u_hw_v2.c
+++ b/providers/hns/hns_roce_u_hw_v2.c
@@ -1527,8 +1527,9 @@ static int hns_roce_u_v2_destroy_qp(struct ibv_qp *ibqp)
return ret;
}
-static int find_empty_entry(struct hns_roce_idx_que *idx_que)
+static int get_wqe_idx(struct hns_roce_srq *srq, int *wqe_idx)
{
+ struct hns_roce_idx_que *idx_que = &srq->idx_que;
int bit_num;
int i;
@@ -1536,12 +1537,20 @@ static int find_empty_entry(struct hns_roce_idx_que *idx_que)
for (i = 0; i < idx_que->bitmap_cnt && idx_que->bitmap[i] == 0; ++i)
;
if (i == idx_que->bitmap_cnt)
- return ENOMEM;
+ return -ENOMEM;
bit_num = ffsl(idx_que->bitmap[i]);
idx_que->bitmap[i] &= ~(1ULL << (bit_num - 1));
- return i * BIT_CNT_PER_LONG + (bit_num - 1);
+ *wqe_idx = i * BIT_CNT_PER_LONG + (bit_num - 1);
+
+ /* If wqe_cnt is less than BIT_CNT_PER_LONG, wqe_idx may be greater
+ * than wqe_cnt.
+ */
+ if (*wqe_idx >= srq->wqe_cnt)
+ return -ENOMEM;
+
+ return 0;
}
static int hns_roce_u_v2_post_srq_recv(struct ibv_srq *ib_srq,
@@ -1580,9 +1589,8 @@ static int hns_roce_u_v2_post_srq_recv(struct ibv_srq *ib_srq,
break;
}
- wqe_idx = find_empty_entry(&srq->idx_que);
- if (wqe_idx < 0 || wqe_idx >= srq->wqe_cnt) {
- ret = -ENOMEM;
+ ret = get_wqe_idx(srq, &wqe_idx);
+ if (ret) {
*bad_wr = wr;
break;
}
--
2.33.0

33
0006-libhns-Don-t-create-RQ-for-a-QP-that-associated-with.patch
View File

@ -1,33 +0,0 @@
From 91034654bdb2fd6e1fce81b4c1aea41bb4b6bf98 Mon Sep 17 00:00:00 2001
From: Wenpeng Liang <liangwenpeng@huawei.com>
Date: Tue, 11 May 2021 19:06:33 +0800
Subject: [PATCH 6/8] libhns: Don't create RQ for a QP that associated with a
SRQ
If a QP is associated with a SRQ, it's RQ should not be created.
Signed-off-by: Wenpeng Liang <liangwenpeng@huawei.com>
Signed-off-by: Weihang Li <liweihang@huawei.com>
---
providers/hns/hns_roce_u_verbs.c | 5 +++++
1 file changed, 5 insertions(+)
diff --git a/providers/hns/hns_roce_u_verbs.c b/providers/hns/hns_roce_u_verbs.c
index 9b4934b9..125858d2 100644
--- a/providers/hns/hns_roce_u_verbs.c
+++ b/providers/hns/hns_roce_u_verbs.c
@@ -760,6 +760,11 @@ static int verify_qp_create_cap(struct hns_roce_context *ctx,
cap->max_recv_sge > ctx->max_sge)
return -EINVAL;
+ if (attr->srq) {
+ cap->max_recv_wr = 0;
+ cap->max_recv_sge = 0;
+ }
+
min_wqe_num = hr_dev->hw_version == HNS_ROCE_HW_VER1 ?
HNS_ROCE_V1_MIN_WQE_NUM : HNS_ROCE_V2_MIN_WQE_NUM;
--
2.33.0

368
0007-libhns-Add-support-for-direct-wqe.patch
View File

@ -1,368 +0,0 @@
From 64c66455fef1c908cc8f06a2b71aa2fd71806218 Mon Sep 17 00:00:00 2001
From: Yixing Liu <liuyixing1@huawei.com>
Date: Wed, 15 Dec 2021 16:42:30 +0800
Subject: [PATCH 7/8] libhns: Add support for direct wqe
The current write wqe mechanism is to write to DDR first, and then notify
the hardware through doorbell to read the data. Direct wqe is a mechanism
to fill wqe directly into the hardware. In the case of light load, the wqe
will be filled into pcie bar space of the hardware, this will reduce one
memory access operation and therefore reduce the latency. SIMD instructions
allows cpu to write the 512 bits at one time to device memory, thus it can
be used for posting direct wqe.
The process of post send of HIP08/09:
+-----------+
| post send |
+-----+-----+
|
+-----+-----+
| write WQE |
+-----+-----+
|
| udma_to_device_barrier()
|
+-----+-----+ Y +-----------+ N
| HIP09 ? +------+ multi WR ?+-------------+
+-----+-----+ +-----+-----+ |
| N | Y |
+-----+-----+ +-----+-----+ +--------+--------+
| ring DB | | ring DB | |direct WQE (ST4) |
+-----------+ +-----------+ +-----------------+
Signed-off-by: Yixing Liu <liuyixing1@huawei.com>
Signed-off-by: Lang Cheng <chenglang@huawei.com>
Signed-off-by: Wenpeng Liang <liangwenpeng@huawei.com>
---
providers/hns/hns_roce_u.h | 5 +++-
providers/hns/hns_roce_u_hw_v2.c | 43 ++++++++++++++++++++++++++------
providers/hns/hns_roce_u_hw_v2.h | 31 +++++++++++++----------
providers/hns/hns_roce_u_verbs.c | 26 +++++++++++++++++--
util/mmio.h | 27 +++++++++++++++++++-
5 files changed, 107 insertions(+), 25 deletions(-)
diff --git a/providers/hns/hns_roce_u.h b/providers/hns/hns_roce_u.h
index b3f48113..37711363 100644
--- a/providers/hns/hns_roce_u.h
+++ b/providers/hns/hns_roce_u.h
@@ -80,6 +80,8 @@
#define INVALID_SGE_LENGTH 0x80000000
+#define HNS_ROCE_DWQE_PAGE_SIZE 65536
+
#define HNS_ROCE_ADDRESS_MASK 0xFFFFFFFF
#define HNS_ROCE_ADDRESS_SHIFT 32
@@ -279,13 +281,14 @@ struct hns_roce_qp {
struct hns_roce_sge_ex ex_sge;
unsigned int next_sge;
int port_num;
- int sl;
+ uint8_t sl;
unsigned int qkey;
enum ibv_mtu path_mtu;
struct hns_roce_rinl_buf rq_rinl_buf;
unsigned long flags;
int refcnt; /* specially used for XRC */
+ void *dwqe_page;
};
struct hns_roce_av {
diff --git a/providers/hns/hns_roce_u_hw_v2.c b/providers/hns/hns_roce_u_hw_v2.c
index 1169b64b..f102fd61 100644
--- a/providers/hns/hns_roce_u_hw_v2.c
+++ b/providers/hns/hns_roce_u_hw_v2.c
@@ -33,6 +33,7 @@
#define _GNU_SOURCE
#include <stdio.h>
#include <string.h>
+#include <sys/mman.h>
#include "hns_roce_u.h"
#include "hns_roce_u_db.h"
#include "hns_roce_u_hw_v2.h"
@@ -297,20 +298,40 @@ static void hns_roce_update_rq_db(struct hns_roce_context *ctx,
}
static void hns_roce_update_sq_db(struct hns_roce_context *ctx,
- unsigned int qpn, unsigned int sl,
- unsigned int sq_head)
+ struct hns_roce_qp *qp)
{
struct hns_roce_db sq_db = {};
- sq_db.byte_4 = htole32(qpn);
+ sq_db.byte_4 = htole32(qp->verbs_qp.qp.qp_num);
roce_set_field(sq_db.byte_4, DB_BYTE_4_CMD_M, DB_BYTE_4_CMD_S,
HNS_ROCE_V2_SQ_DB);
- sq_db.parameter = htole32(sq_head);
- roce_set_field(sq_db.parameter, DB_PARAM_SL_M, DB_PARAM_SL_S, sl);
+ sq_db.parameter = htole32(qp->sq.head);
+ roce_set_field(sq_db.parameter, DB_PARAM_SL_M, DB_PARAM_SL_S, qp->sl);
hns_roce_write64(ctx->uar + ROCEE_VF_DB_CFG0_OFFSET, (__le32 *)&sq_db);
}
+static void hns_roce_write512(uint64_t *dest, uint64_t *val)
+{
+ mmio_memcpy_x64(dest, val, sizeof(struct hns_roce_rc_sq_wqe));
+}
+
+static void hns_roce_write_dwqe(struct hns_roce_qp *qp, void *wqe)
+{
+ struct hns_roce_rc_sq_wqe *rc_sq_wqe = wqe;
+
+ /* All kinds of DirectWQE have the same header field layout */
+ roce_set_bit(rc_sq_wqe->byte_4, RC_SQ_WQE_BYTE_4_FLAG_S, 1);
+ roce_set_field(rc_sq_wqe->byte_4, RC_SQ_WQE_BYTE_4_DB_SL_L_M,
+ RC_SQ_WQE_BYTE_4_DB_SL_L_S, qp->sl);
+ roce_set_field(rc_sq_wqe->byte_4, RC_SQ_WQE_BYTE_4_DB_SL_H_M,
+ RC_SQ_WQE_BYTE_4_DB_SL_H_S, qp->sl >> HNS_ROCE_SL_SHIFT);
+ roce_set_field(rc_sq_wqe->byte_4, RC_SQ_WQE_BYTE_4_WQE_INDEX_M,
+ RC_SQ_WQE_BYTE_4_WQE_INDEX_S, qp->sq.head);
+
+ hns_roce_write512(qp->dwqe_page, wqe);
+}
+
static void hns_roce_v2_update_cq_cons_index(struct hns_roce_context *ctx,
struct hns_roce_cq *cq)
{
@@ -339,8 +360,7 @@ static struct hns_roce_qp *hns_roce_v2_find_qp(struct hns_roce_context *ctx,
return NULL;
}
-static void hns_roce_v2_clear_qp(struct hns_roce_context *ctx,
- struct hns_roce_qp *qp)
+void hns_roce_v2_clear_qp(struct hns_roce_context *ctx, struct hns_roce_qp *qp)
{
uint32_t qpn = qp->verbs_qp.qp.qp_num;
uint32_t tind = (qpn & (ctx->num_qps - 1)) >> ctx->qp_table_shift;
@@ -1196,6 +1216,7 @@ int hns_roce_u_v2_post_send(struct ibv_qp *ibvqp, struct ibv_send_wr *wr,
break;
case IBV_QPT_UD:
ret = set_ud_wqe(wqe, qp, wr, nreq, &sge_info);
+ qp->sl = to_hr_ah(wr->wr.ud.ah)->av.sl;
break;
default:
ret = EINVAL;
@@ -1214,7 +1235,10 @@ out:
udma_to_device_barrier();
- hns_roce_update_sq_db(ctx, ibvqp->qp_num, qp->sl, qp->sq.head);
+ if (nreq == 1 && (qp->flags & HNS_ROCE_QP_CAP_DIRECT_WQE))
+ hns_roce_write_dwqe(qp, wqe);
+ else
+ hns_roce_update_sq_db(ctx, qp);
if (qp->flags & HNS_ROCE_QP_CAP_SQ_RECORD_DB)
*(qp->sdb) = qp->sq.head & 0xffff;
@@ -1506,6 +1530,9 @@ static int hns_roce_u_v2_destroy_qp(struct ibv_qp *ibqp)
if (ret)
return ret;
+ if (qp->flags & HNS_ROCE_QP_CAP_DIRECT_WQE)
+ munmap(qp->dwqe_page, HNS_ROCE_DWQE_PAGE_SIZE);
+
hns_roce_v2_clear_qp(ctx, qp);
hns_roce_lock_cqs(ibqp);
diff --git a/providers/hns/hns_roce_u_hw_v2.h b/providers/hns/hns_roce_u_hw_v2.h
index c13d82e3..af72cd70 100644
--- a/providers/hns/hns_roce_u_hw_v2.h
+++ b/providers/hns/hns_roce_u_hw_v2.h
@@ -40,6 +40,8 @@
#define HNS_ROCE_CMDSN_MASK 0x3
+#define HNS_ROCE_SL_SHIFT 2
+
/* V2 REG DEFINITION */
#define ROCEE_VF_DB_CFG0_OFFSET 0x0230
@@ -133,6 +135,8 @@ struct hns_roce_db {
#define DB_BYTE_4_CMD_S 24
#define DB_BYTE_4_CMD_M GENMASK(27, 24)
+#define DB_BYTE_4_FLAG_S 31
+
#define DB_PARAM_SRQ_PRODUCER_COUNTER_S 0
#define DB_PARAM_SRQ_PRODUCER_COUNTER_M GENMASK(15, 0)
@@ -216,8 +220,16 @@ struct hns_roce_rc_sq_wqe {
};
#define RC_SQ_WQE_BYTE_4_OPCODE_S 0
-#define RC_SQ_WQE_BYTE_4_OPCODE_M \
- (((1UL << 5) - 1) << RC_SQ_WQE_BYTE_4_OPCODE_S)
+#define RC_SQ_WQE_BYTE_4_OPCODE_M GENMASK(4, 0)
+
+#define RC_SQ_WQE_BYTE_4_DB_SL_L_S 5
+#define RC_SQ_WQE_BYTE_4_DB_SL_L_M GENMASK(6, 5)
+
+#define RC_SQ_WQE_BYTE_4_DB_SL_H_S 13
+#define RC_SQ_WQE_BYTE_4_DB_SL_H_M GENMASK(14, 13)
+
+#define RC_SQ_WQE_BYTE_4_WQE_INDEX_S 15
+#define RC_SQ_WQE_BYTE_4_WQE_INDEX_M GENMASK(30, 15)
#define RC_SQ_WQE_BYTE_4_OWNER_S 7
@@ -239,6 +251,8 @@ struct hns_roce_rc_sq_wqe {
#define RC_SQ_WQE_BYTE_4_RDMA_WRITE_S 22
+#define RC_SQ_WQE_BYTE_4_FLAG_S 31
+
#define RC_SQ_WQE_BYTE_16_XRC_SRQN_S 0
#define RC_SQ_WQE_BYTE_16_XRC_SRQN_M \
(((1UL << 24) - 1) << RC_SQ_WQE_BYTE_16_XRC_SRQN_S)
@@ -311,23 +325,12 @@ struct hns_roce_ud_sq_wqe {
#define UD_SQ_WQE_OPCODE_S 0
#define UD_SQ_WQE_OPCODE_M GENMASK(4, 0)
-#define UD_SQ_WQE_DB_SL_L_S 5
-#define UD_SQ_WQE_DB_SL_L_M GENMASK(6, 5)
-
-#define UD_SQ_WQE_DB_SL_H_S 13
-#define UD_SQ_WQE_DB_SL_H_M GENMASK(14, 13)
-
-#define UD_SQ_WQE_INDEX_S 15
-#define UD_SQ_WQE_INDEX_M GENMASK(30, 15)
-
#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_FLAG_S 31
-
#define UD_SQ_WQE_PD_S 0
#define UD_SQ_WQE_PD_M GENMASK(23, 0)
@@ -376,4 +379,6 @@ struct hns_roce_ud_sq_wqe {
#define MAX_SERVICE_LEVEL 0x7
+void hns_roce_v2_clear_qp(struct hns_roce_context *ctx, struct hns_roce_qp *qp);
+
#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 125858d2..fc902815 100644
--- a/providers/hns/hns_roce_u_verbs.c
+++ b/providers/hns/hns_roce_u_verbs.c
@@ -1076,7 +1076,8 @@ static int hns_roce_store_qp(struct hns_roce_context *ctx,
static int qp_exec_create_cmd(struct ibv_qp_init_attr_ex *attr,
struct hns_roce_qp *qp,
- struct hns_roce_context *ctx)
+ struct hns_roce_context *ctx,
+ uint64_t *dwqe_mmap_key)
{
struct hns_roce_create_qp_ex_resp resp_ex = {};
struct hns_roce_create_qp_ex cmd_ex = {};
@@ -1093,6 +1094,7 @@ static int qp_exec_create_cmd(struct ibv_qp_init_attr_ex *attr,
&resp_ex.ibv_resp, sizeof(resp_ex));
qp->flags = resp_ex.drv_payload.cap_flags;
+ *dwqe_mmap_key = resp_ex.drv_payload.dwqe_mmap_key;
return ret;
}
@@ -1144,11 +1146,23 @@ static int hns_roce_alloc_qp_buf(struct ibv_qp_init_attr_ex *attr,
return ret;
}
+static int mmap_dwqe(struct ibv_context *ibv_ctx, struct hns_roce_qp *qp,
+ uint64_t dwqe_mmap_key)
+{
+ qp->dwqe_page = mmap(NULL, HNS_ROCE_DWQE_PAGE_SIZE, PROT_WRITE,
+ MAP_SHARED, ibv_ctx->cmd_fd, dwqe_mmap_key);
+ if (qp->dwqe_page == MAP_FAILED)
+ return -EINVAL;
+
+ return 0;
+}
+
static struct ibv_qp *create_qp(struct ibv_context *ibv_ctx,
struct ibv_qp_init_attr_ex *attr)
{
struct hns_roce_context *context = to_hr_ctx(ibv_ctx);
struct hns_roce_qp *qp;
+ uint64_t dwqe_mmap_key;
int ret;
ret = verify_qp_create_attr(context, attr);
@@ -1167,7 +1181,7 @@ static struct ibv_qp *create_qp(struct ibv_context *ibv_ctx,
if (ret)
goto err_buf;
- ret = qp_exec_create_cmd(attr, qp, context);
+ ret = qp_exec_create_cmd(attr, qp, context, &dwqe_mmap_key);
if (ret)
goto err_cmd;
@@ -1175,10 +1189,18 @@ static struct ibv_qp *create_qp(struct ibv_context *ibv_ctx,
if (ret)
goto err_store;
+ if (qp->flags & HNS_ROCE_QP_CAP_DIRECT_WQE) {
+ ret = mmap_dwqe(ibv_ctx, qp, dwqe_mmap_key);
+ if (ret)
+ goto err_dwqe;
+ }
+
qp_setup_config(attr, qp, context);
return &qp->verbs_qp.qp;
+err_dwqe:
+ hns_roce_v2_clear_qp(context, qp);
err_store:
ibv_cmd_destroy_qp(&qp->verbs_qp.qp);
err_cmd:
diff --git a/util/mmio.h b/util/mmio.h
index 101af9dd..01d1455e 100644
--- a/util/mmio.h
+++ b/util/mmio.h
@@ -210,8 +210,33 @@ static inline void mmio_memcpy_x64(void *dest, const void *src, size_t bytecnt)
{
s390_mmio_write(dest, src, bytecnt);
}
-#else
+#elif defined(__aarch64__) || defined(__arm__)
+#include <arm_neon.h>
+
+static inline void _mmio_memcpy_x64_64b(void *dest, const void *src)
+{
+ vst4q_u64(dest, vld4q_u64(src));
+}
+
+static inline void _mmio_memcpy_x64(void *dest, const void *src, size_t bytecnt)
+{
+ do {
+ _mmio_memcpy_x64_64b(dest, src);
+ bytecnt -= sizeof(uint64x2x4_t);
+ src += sizeof(uint64x2x4_t);
+ } while (bytecnt > 0);
+}
+
+#define mmio_memcpy_x64(dest, src, bytecount) \
+ ({ \
+ if (__builtin_constant_p((bytecount) == 64)) \
+ _mmio_memcpy_x64_64b((dest), (src)); \
+ else \
+ _mmio_memcpy_x64((dest), (src), (bytecount)); \
+ })
+
+#else
/* Transfer is some multiple of 64 bytes */
static inline void mmio_memcpy_x64(void *dest, const void *src, size_t bytecnt)
{
--
2.33.0

70
0008-libhns-Use-new-SQ-doorbell-register-for-HIP09.patch
View File

@ -1,70 +0,0 @@
From 608c142e7cbac2a6c02071022fe87b081a6ddc4f Mon Sep 17 00:00:00 2001
From: Yixing Liu <liuyixing1@huawei.com>
Date: Tue, 21 Dec 2021 21:38:08 +0800
Subject: [PATCH 8/8] libhns: Use new SQ doorbell register for HIP09
HIP09 set a new BAR space for SQ doorbell. Each SQ doorbell has an
independent BAR space and the size is 64KB. SQ doorbell share
the same BAR space with direct WQE.
Signed-off-by: Yixing Liu <liuyixing1@huawei.com>
Signed-off-by: Wenpeng Liang <liangwenpeng@huawei.com>
---
providers/hns/hns_roce_u.h | 1 +
providers/hns/hns_roce_u_hw_v2.c | 4 ++--
providers/hns/hns_roce_u_verbs.c | 5 +++++
3 files changed, 8 insertions(+), 2 deletions(-)
diff --git a/providers/hns/hns_roce_u.h b/providers/hns/hns_roce_u.h
index 37711363..460363b7 100644
--- a/providers/hns/hns_roce_u.h
+++ b/providers/hns/hns_roce_u.h
@@ -238,6 +238,7 @@ struct hns_roce_wq {
unsigned int wqe_shift;
unsigned int shift; /* wq size is 2^shift */
int offset;
+ void *db_reg;
};
/* record the result of sge process */
diff --git a/providers/hns/hns_roce_u_hw_v2.c b/providers/hns/hns_roce_u_hw_v2.c
index f102fd61..9cbc0aac 100644
--- a/providers/hns/hns_roce_u_hw_v2.c
+++ b/providers/hns/hns_roce_u_hw_v2.c
@@ -308,7 +308,7 @@ static void hns_roce_update_sq_db(struct hns_roce_context *ctx,
sq_db.parameter = htole32(qp->sq.head);
roce_set_field(sq_db.parameter, DB_PARAM_SL_M, DB_PARAM_SL_S, qp->sl);
- hns_roce_write64(ctx->uar + ROCEE_VF_DB_CFG0_OFFSET, (__le32 *)&sq_db);
+ hns_roce_write64(qp->sq.db_reg, (__le32 *)&sq_db);
}
static void hns_roce_write512(uint64_t *dest, uint64_t *val)
@@ -329,7 +329,7 @@ static void hns_roce_write_dwqe(struct hns_roce_qp *qp, void *wqe)
roce_set_field(rc_sq_wqe->byte_4, RC_SQ_WQE_BYTE_4_WQE_INDEX_M,
RC_SQ_WQE_BYTE_4_WQE_INDEX_S, qp->sq.head);
- hns_roce_write512(qp->dwqe_page, wqe);
+ hns_roce_write512(qp->sq.db_reg, wqe);
}
static void hns_roce_v2_update_cq_cons_index(struct hns_roce_context *ctx,
diff --git a/providers/hns/hns_roce_u_verbs.c b/providers/hns/hns_roce_u_verbs.c
index fc902815..c5022c83 100644
--- a/providers/hns/hns_roce_u_verbs.c
+++ b/providers/hns/hns_roce_u_verbs.c
@@ -1117,6 +1117,11 @@ static void qp_setup_config(struct ibv_qp_init_attr_ex *attr,
}
qp->max_inline_data = attr->cap.max_inline_data;
+
+ if (qp->flags & HNS_ROCE_QP_CAP_DIRECT_WQE)
+ qp->sq.db_reg = qp->dwqe_page;
+ else
+ qp->sq.db_reg = ctx->uar + ROCEE_VF_DB_CFG0_OFFSET;
}
void hns_roce_free_qp_buf(struct hns_roce_qp *qp, struct hns_roce_context *ctx)
--
2.33.0

70
0009-libhns-Bugfix-for-checking-whether-the-SRQ-is-full-w.patch
View File

@ -1,70 +0,0 @@
From 5cc1a047c4d71ced86b0f71f66adf12475a3c788 Mon Sep 17 00:00:00 2001
From: Wenpeng Liang <liangwenpeng@huawei.com>
Date: Tue, 11 May 2021 19:06:35 +0800
Subject: libhns: Bugfix for checking whether the SRQ is full when posting WR
If the user post a list of WRs, the head in the for loop is not updated in
time, and the judgment of if (head == tail) becomes invalid.
Signed-off-by: Wenpeng Liang <liangwenpeng@huawei.com>
Signed-off-by: Weihang Li <liweihang@huawei.com>
---
providers/hns/hns_roce_u_hw_v2.c | 17 +++++++++++++----
providers/hns/hns_roce_u_verbs.c | 2 +-
2 files changed, 14 insertions(+), 5 deletions(-)
diff --git a/providers/hns/hns_roce_u_hw_v2.c b/providers/hns/hns_roce_u_hw_v2.c
index 82124082..0c15bdbe 100644
--- a/providers/hns/hns_roce_u_hw_v2.c
+++ b/providers/hns/hns_roce_u_hw_v2.c
@@ -1527,6 +1527,15 @@ static int hns_roce_u_v2_destroy_qp(struct ibv_qp *ibqp)
return ret;
}
+static int hns_roce_v2_srqwq_overflow(struct hns_roce_srq *srq)
+{
+ struct hns_roce_idx_que *idx_que = &srq->idx_que;
+ unsigned int cur;
+
+ cur = idx_que->head - idx_que->tail;
+ return cur >= srq->wqe_cnt - 1;
+}
+
static int get_wqe_idx(struct hns_roce_srq *srq, int *wqe_idx)
{
struct hns_roce_idx_que *idx_que = &srq->idx_que;
@@ -1577,14 +1586,14 @@ static int hns_roce_u_v2_post_srq_recv(struct ibv_srq *ib_srq,
max_sge = srq->max_gs - srq->rsv_sge;
for (nreq = 0; wr; ++nreq, wr = wr->next) {
- if (wr->num_sge > max_sge) {
- ret = -EINVAL;
+ if (hns_roce_v2_srqwq_overflow(srq)) {
+ ret = -ENOMEM;
*bad_wr = wr;
break;
}
- if (srq->idx_que.head == srq->idx_que.tail) {
- ret = -ENOMEM;
+ if (wr->num_sge > max_sge) {
+ ret = -EINVAL;
*bad_wr = wr;
break;
}
diff --git a/providers/hns/hns_roce_u_verbs.c b/providers/hns/hns_roce_u_verbs.c
index 3abf7b48..dace35fd 100644
--- a/providers/hns/hns_roce_u_verbs.c
+++ b/providers/hns/hns_roce_u_verbs.c
@@ -492,7 +492,7 @@ static int hns_roce_create_idx_que(struct hns_roce_srq *srq)
idx_que->bitmap[i] = ~(0UL);
idx_que->head = 0;
- idx_que->tail = srq->wqe_cnt - 1;
+ idx_que->tail = 0;
return 0;
}
--
2.30.0

30
0010-libhns-Allow-users-to-create-a-0-depth-SRQs.patch
View File

@ -1,30 +0,0 @@
From a79800afbbc48e5c5274bf3fc0e890705b3a596d Mon Sep 17 00:00:00 2001
From: Wenpeng Liang <liangwenpeng@huawei.com>
Date: Tue, 11 May 2021 19:06:36 +0800
Subject: libhns: Allow users to create a 0-depth SRQs
Users is allowed to create 0-depth SRQs, so the judgement about whether
max_wr is zero should be removed.
Signed-off-by: Wenpeng Liang <liangwenpeng@huawei.com>
Signed-off-by: Weihang Li <liweihang@huawei.com>
---
providers/hns/hns_roce_u_verbs.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/providers/hns/hns_roce_u_verbs.c b/providers/hns/hns_roce_u_verbs.c
index dace35fd..2d1a6de3 100644
--- a/providers/hns/hns_roce_u_verbs.c
+++ b/providers/hns/hns_roce_u_verbs.c
@@ -525,7 +525,7 @@ static int hns_roce_verify_srq(struct hns_roce_context *context,
init_attr->srq_type != IBV_SRQT_XRC)
return -EINVAL;
- if (!init_attr->attr.max_wr || !init_attr->attr.max_sge ||
+ if (!init_attr->attr.max_sge ||
init_attr->attr.max_wr > context->max_srq_wr ||
init_attr->attr.max_sge > context->max_srq_sge)
return -EINVAL;
--
2.30.0

176
0011-libhns-Refactor-the-process-of-post_srq_recv.patch
View File

@ -1,176 +0,0 @@
From f46d1f312984bdb372d2f86ac7dd7c2dcaa8c721 Mon Sep 17 00:00:00 2001
From: Wenpeng Liang <liangwenpeng@huawei.com>
Date: Tue, 11 May 2021 19:06:37 +0800
Subject: libhns: Refactor the process of post_srq_recv
SRQ is a shared queue, it mainly consists of four parts:
1. wqe buf: wqe buf is used to store wqe data.
2. wqe_idx buf: the cqe of SRQ is not generated in the order of wqe, so
the wqe_idx corresponding to the idle WQE needs to be placed in an FIFO
queue, it can instruct the hardware to obtain the corresponding WQE.
3.bitmap: bitmap is used to generate and release wqe_idx. When the user
has a new WR, the driver finds the idx of the idle wqe in bitmap. When the
CQE of wqe is generated, the driver releases the idx.
4. wr_id buf: wr_id buf is used to store the user's wr_id, then return it
to the user when ibv_poll_cq() is invoked.
After refactor, the functions of the four parts are more clearer.
Signed-off-by: Wenpeng Liang <liangwenpeng@huawei.com>
Signed-off-by: Weihang Li <liweihang@huawei.com>
---
providers/hns/hns_roce_u_hw_v2.c | 95 +++++++++++++++++++-------------
1 file changed, 57 insertions(+), 38 deletions(-)
diff --git a/providers/hns/hns_roce_u_hw_v2.c b/providers/hns/hns_roce_u_hw_v2.c
index 0c15bdbe..b622eaef 100644
--- a/providers/hns/hns_roce_u_hw_v2.c
+++ b/providers/hns/hns_roce_u_hw_v2.c
@@ -242,7 +242,7 @@ static void *get_send_sge_ex(struct hns_roce_qp *qp, unsigned int n)
return qp->buf.buf + qp->ex_sge.offset + (n << qp->ex_sge.sge_shift);
}
-static void *get_srq_wqe(struct hns_roce_srq *srq, int n)
+static void *get_srq_wqe(struct hns_roce_srq *srq, unsigned int n)
{
return srq->buf.buf + (n << srq->wqe_shift);
}
@@ -1536,7 +1536,21 @@ static int hns_roce_v2_srqwq_overflow(struct hns_roce_srq *srq)
return cur >= srq->wqe_cnt - 1;
}
-static int get_wqe_idx(struct hns_roce_srq *srq, int *wqe_idx)
+static int check_post_srq_valid(struct hns_roce_srq *srq,
+ struct ibv_recv_wr *wr)
+{
+ unsigned int max_sge = srq->max_gs - srq->rsv_sge;
+
+ if (hns_roce_v2_srqwq_overflow(srq))
+ return -ENOMEM;
+
+ if (wr->num_sge > max_sge)
+ return -EINVAL;
+
+ return 0;
+}
+
+static int get_wqe_idx(struct hns_roce_srq *srq, unsigned int *wqe_idx)
{
struct hns_roce_idx_que *idx_que = &srq->idx_que;
int bit_num;
@@ -1562,38 +1576,58 @@ static int get_wqe_idx(struct hns_roce_srq *srq, int *wqe_idx)
return 0;
}
+static void fill_srq_wqe(struct hns_roce_srq *srq, unsigned int wqe_idx,
+ struct ibv_recv_wr *wr)
+{
+ struct hns_roce_v2_wqe_data_seg *dseg;
+ int i;
+
+ dseg = get_srq_wqe(srq, wqe_idx);
+
+ for (i = 0; i < wr->num_sge; ++i) {
+ dseg[i].len = htole32(wr->sg_list[i].length);
+ dseg[i].lkey = htole32(wr->sg_list[i].lkey);
+ dseg[i].addr = htole64(wr->sg_list[i].addr);
+ }
+
+ /* hw stop reading when identify the last one */
+ if (srq->rsv_sge) {
+ dseg[i].len = htole32(INVALID_SGE_LENGTH);
+ dseg[i].lkey = htole32(0x0);
+ dseg[i].addr = 0;
+ }
+}
+
+static void fill_wqe_idx(struct hns_roce_srq *srq, unsigned int wqe_idx)
+{
+ struct hns_roce_idx_que *idx_que = &srq->idx_que;
+ unsigned int head;
+ __le32 *idx_buf;
+
+ head = idx_que->head & (srq->wqe_cnt - 1);
+
+ idx_buf = get_idx_buf(idx_que, head);
+ *idx_buf = htole32(wqe_idx);
+
+ idx_que->head++;
+}
+
static int hns_roce_u_v2_post_srq_recv(struct ibv_srq *ib_srq,
struct ibv_recv_wr *wr,
struct ibv_recv_wr **bad_wr)
{
struct hns_roce_context *ctx = to_hr_ctx(ib_srq->context);
struct hns_roce_srq *srq = to_hr_srq(ib_srq);
- struct hns_roce_v2_wqe_data_seg *dseg;
struct hns_roce_db srq_db;
- unsigned int max_sge;
- __le32 *srq_idx;
+ unsigned int wqe_idx;
int ret = 0;
- int wqe_idx;
- void *wqe;
int nreq;
- int ind;
- int i;
pthread_spin_lock(&srq->lock);
- /* current idx of srqwq */
- ind = srq->idx_que.head & (srq->wqe_cnt - 1);
-
- max_sge = srq->max_gs - srq->rsv_sge;
for (nreq = 0; wr; ++nreq, wr = wr->next) {
- if (hns_roce_v2_srqwq_overflow(srq)) {
- ret = -ENOMEM;
- *bad_wr = wr;
- break;
- }
-
- if (wr->num_sge > max_sge) {
- ret = -EINVAL;
+ ret = check_post_srq_valid(srq, wr);
+ if (ret) {
*bad_wr = wr;
break;
}
@@ -1604,28 +1638,13 @@ static int hns_roce_u_v2_post_srq_recv(struct ibv_srq *ib_srq,
break;
}
- wqe = get_srq_wqe(srq, wqe_idx);
- dseg = (struct hns_roce_v2_wqe_data_seg *)wqe;
-
- for (i = 0; i < wr->num_sge; ++i) {
- set_data_seg_v2(dseg, wr->sg_list + i);
- dseg++;
- }
-
- /* hw stop reading when identify the last one */
- if (srq->rsv_sge)
- set_ending_data_seg(dseg);
-
- srq_idx = (__le32 *)get_idx_buf(&srq->idx_que, ind);
- *srq_idx = htole32(wqe_idx);
+ fill_srq_wqe(srq, wqe_idx, wr);
+ fill_wqe_idx(srq, wqe_idx);
srq->wrid[wqe_idx] = wr->wr_id;
- ind = (ind + 1) & (srq->wqe_cnt - 1);
}
if (nreq) {
- srq->idx_que.head += nreq;
-
/*
* Make sure that descriptors are written before
* we write doorbell record.
--
2.30.0

33
0012-libhns-Set-srqlimit-to-0-when-creating-SRQ.patch
View File

@ -1,33 +0,0 @@
From a18b0ee409d3382aa556b8f06a6cd6bfbef3f5c8 Mon Sep 17 00:00:00 2001
From: Wenpeng Liang <liangwenpeng@huawei.com>
Date: Tue, 11 May 2021 19:06:38 +0800
Subject: libhns: Set srqlimit to 0 when creating SRQ
According to the IB specification, the srq_limt parameter should not be
configured when creating srq. But the libhns does not set attr.srq_limit
to 0 currently. As a result, when attr.srq_limit provided by the user is
not 0, the value of attr.srq_limit returned to the user will be different
from that obtained by ibv_query_srq(). Therefore, the driver should set
attr.srq_limit to 0 when creating SRQ.
Signed-off-by: Wenpeng Liang <liangwenpeng@huawei.com>
Signed-off-by: Weihang Li <liweihang@huawei.com>
---
providers/hns/hns_roce_u_verbs.c | 1 +
1 file changed, 1 insertion(+)
diff --git a/providers/hns/hns_roce_u_verbs.c b/providers/hns/hns_roce_u_verbs.c
index 2d1a6de3..107da753 100644
--- a/providers/hns/hns_roce_u_verbs.c
+++ b/providers/hns/hns_roce_u_verbs.c
@@ -580,6 +580,7 @@ static struct ibv_srq *create_srq(struct ibv_context *context,
srq->wqe_cnt = roundup_pow_of_two(attr->max_wr + 1);
srq->max_gs = roundup_pow_of_two(attr->max_sge + srq->rsv_sge);
attr->max_sge = srq->max_gs;
+ attr->srq_limit = 0;
ret = hns_roce_create_idx_que(srq);
if (ret)
--
2.30.0

367
0013-libhns-Refactor-the-process-of-create_srq.patch
View File

@ -1,367 +0,0 @@
From b914c76318f5b95e3157c3cbf1ccb49ec6d27635 Mon Sep 17 00:00:00 2001
From: Wenpeng Liang <liangwenpeng@huawei.com>
Date: Tue, 11 May 2021 19:06:39 +0800
Subject: libhns: Refactor the process of create_srq
Reorganize create_srq() as several sub-functions to make the process
clearer.
Signed-off-by: Wenpeng Liang <liangwenpeng@huawei.com>
Signed-off-by: Weihang Li <liweihang@huawei.com>
---
providers/hns/hns_roce_u.h | 7 +-
providers/hns/hns_roce_u_hw_v2.c | 2 +-
providers/hns/hns_roce_u_verbs.c | 178 ++++++++++++++++++-------------
3 files changed, 105 insertions(+), 82 deletions(-)
diff --git a/providers/hns/hns_roce_u.h b/providers/hns/hns_roce_u.h
index b3f48113..a437727c 100644
--- a/providers/hns/hns_roce_u.h
+++ b/providers/hns/hns_roce_u.h
@@ -211,7 +211,8 @@ struct hns_roce_idx_que {
struct hns_roce_srq {
struct verbs_srq verbs_srq;
- struct hns_roce_buf buf;
+ struct hns_roce_idx_que idx_que;
+ struct hns_roce_buf wqe_buf;
pthread_spinlock_t lock;
unsigned long *wrid;
unsigned int srqn;
@@ -221,7 +222,6 @@ struct hns_roce_srq {
unsigned int wqe_shift;
unsigned int *db;
unsigned short counter;
- struct hns_roce_idx_que idx_que;
};
struct hns_roce_wq {
@@ -343,8 +343,7 @@ static inline struct hns_roce_cq *to_hr_cq(struct ibv_cq *ibv_cq)
static inline struct hns_roce_srq *to_hr_srq(struct ibv_srq *ibv_srq)
{
- return container_of(container_of(ibv_srq, struct verbs_srq, srq),
- struct hns_roce_srq, verbs_srq);
+ return container_of(ibv_srq, struct hns_roce_srq, verbs_srq.srq);
}
static inline struct hns_roce_qp *to_hr_qp(struct ibv_qp *ibv_qp)
diff --git a/providers/hns/hns_roce_u_hw_v2.c b/providers/hns/hns_roce_u_hw_v2.c
index b622eaef..d4e7e4f9 100644
--- a/providers/hns/hns_roce_u_hw_v2.c
+++ b/providers/hns/hns_roce_u_hw_v2.c
@@ -244,7 +244,7 @@ static void *get_send_sge_ex(struct hns_roce_qp *qp, unsigned int n)
static void *get_srq_wqe(struct hns_roce_srq *srq, unsigned int n)
{
- return srq->buf.buf + (n << srq->wqe_shift);
+ return srq->wqe_buf.buf + (n << srq->wqe_shift);
}
static void *get_idx_buf(struct hns_roce_idx_que *idx_que, int n)
diff --git a/providers/hns/hns_roce_u_verbs.c b/providers/hns/hns_roce_u_verbs.c
index 107da753..75b9e530 100644
--- a/providers/hns/hns_roce_u_verbs.c
+++ b/providers/hns/hns_roce_u_verbs.c
@@ -432,17 +432,23 @@ static int hns_roce_store_srq(struct hns_roce_context *ctx,
uint32_t tind = (srq->srqn & (ctx->num_srqs - 1)) >>
ctx->srq_table_shift;
+ pthread_mutex_lock(&ctx->srq_table_mutex);
+
if (!ctx->srq_table[tind].refcnt) {
ctx->srq_table[tind].table =
calloc(ctx->srq_table_mask + 1,
sizeof(struct hns_roce_srq *));
- if (!ctx->srq_table[tind].table)
+ if (!ctx->srq_table[tind].table) {
+ pthread_mutex_unlock(&ctx->srq_table_mutex);
return -ENOMEM;
+ }
}
++ctx->srq_table[tind].refcnt;
ctx->srq_table[tind].table[srq->srqn & ctx->srq_table_mask] = srq;
+ pthread_mutex_unlock(&ctx->srq_table_mutex);
+
return 0;
}
@@ -461,13 +467,46 @@ static void hns_roce_clear_srq(struct hns_roce_context *ctx, uint32_t srqn)
{
uint32_t tind = (srqn & (ctx->num_srqs - 1)) >> ctx->srq_table_shift;
+ pthread_mutex_lock(&ctx->srq_table_mutex);
+
if (!--ctx->srq_table[tind].refcnt)
free(ctx->srq_table[tind].table);
else
ctx->srq_table[tind].table[srqn & ctx->srq_table_mask] = NULL;
+
+ pthread_mutex_unlock(&ctx->srq_table_mutex);
+}
+
+static int verify_srq_create_attr(struct hns_roce_context *context,
+ struct ibv_srq_init_attr_ex *attr)
+{
+ if (attr->srq_type != IBV_SRQT_BASIC &&
+ attr->srq_type != IBV_SRQT_XRC)
+ return -EINVAL;
+
+ if (!attr->attr.max_sge ||
+ attr->attr.max_wr > context->max_srq_wr ||
+ attr->attr.max_sge > context->max_srq_sge)
+ return -EINVAL;
+
+ return 0;
+}
+
+static void set_srq_param(struct ibv_context *context, struct hns_roce_srq *srq,
+ struct ibv_srq_init_attr_ex *attr)
+{
+ if (to_hr_dev(context->device)->hw_version == HNS_ROCE_HW_VER2)
+ srq->rsv_sge = 1;
+
+ srq->wqe_cnt = roundup_pow_of_two(attr->attr.max_wr + 1);
+ srq->max_gs = roundup_pow_of_two(attr->attr.max_sge + srq->rsv_sge);
+ srq->wqe_shift = hr_ilog32(roundup_pow_of_two(HNS_ROCE_SGE_SIZE *
+ srq->max_gs));
+ attr->attr.max_sge = srq->max_gs;
+ attr->attr.srq_limit = 0;
}
-static int hns_roce_create_idx_que(struct hns_roce_srq *srq)
+static int alloc_srq_idx_que(struct hns_roce_srq *srq)
{
struct hns_roce_idx_que *idx_que = &srq->idx_que;
unsigned int buf_size;
@@ -478,13 +517,13 @@ static int hns_roce_create_idx_que(struct hns_roce_srq *srq)
BIT_CNT_PER_LONG;
idx_que->bitmap = calloc(idx_que->bitmap_cnt, sizeof(unsigned long));
if (!idx_que->bitmap)
- return ENOMEM;
+ return -ENOMEM;
buf_size = to_hr_hem_entries_size(srq->wqe_cnt, idx_que->entry_shift);
if (hns_roce_alloc_buf(&idx_que->buf, buf_size, HNS_HW_PAGE_SIZE)) {
free(idx_que->bitmap);
idx_que->bitmap = NULL;
- return ENOMEM;
+ return -ENOMEM;
}
/* init the idx_que bitmap */
@@ -497,40 +536,48 @@ static int hns_roce_create_idx_que(struct hns_roce_srq *srq)
return 0;
}
-static int hns_roce_alloc_srq_buf(struct hns_roce_srq *srq)
+static int alloc_srq_wqe_buf(struct hns_roce_srq *srq)
{
- int srq_buf_size;
+ int buf_size = to_hr_hem_entries_size(srq->wqe_cnt, srq->wqe_shift);
- srq->wrid = calloc(srq->wqe_cnt, sizeof(unsigned long));
- if (!srq->wrid)
- return ENOMEM;
+ return hns_roce_alloc_buf(&srq->wqe_buf, buf_size, HNS_HW_PAGE_SIZE);
+}
- srq->wqe_shift = hr_ilog32(roundup_pow_of_two(HNS_ROCE_SGE_SIZE *
- srq->max_gs));
- srq_buf_size = to_hr_hem_entries_size(srq->wqe_cnt, srq->wqe_shift);
+static int alloc_srq_buf(struct hns_roce_srq *srq)
+{
+ int ret;
- /* allocate srq wqe buf */
- if (hns_roce_alloc_buf(&srq->buf, srq_buf_size, HNS_HW_PAGE_SIZE)) {
- free(srq->wrid);
- return ENOMEM;
+ ret = alloc_srq_idx_que(srq);
+ if (ret)
+ return ret;
+
+ ret = alloc_srq_wqe_buf(srq);
+ if (ret)
+ goto err_idx_que;
+
+ srq->wrid = calloc(srq->wqe_cnt, sizeof(*srq->wrid));
+ if (!srq->wrid) {
+ ret = -ENOMEM;
+ goto err_wqe_buf;
}
return 0;
-}
-static int hns_roce_verify_srq(struct hns_roce_context *context,
- struct ibv_srq_init_attr_ex *init_attr)
-{
- if (init_attr->srq_type != IBV_SRQT_BASIC &&
- init_attr->srq_type != IBV_SRQT_XRC)
- return -EINVAL;
+err_wqe_buf:
+ hns_roce_free_buf(&srq->wqe_buf);
+err_idx_que:
+ hns_roce_free_buf(&srq->idx_que.buf);
+ free(srq->idx_que.bitmap);
- if (!init_attr->attr.max_sge ||
- init_attr->attr.max_wr > context->max_srq_wr ||
- init_attr->attr.max_sge > context->max_srq_sge)
- return -EINVAL;
+ return ret;
+}
- return 0;
+static void free_srq_buf(struct hns_roce_srq *srq)
+{
+ free(srq->wrid);
+ hns_roce_free_buf(&srq->wqe_buf);
+ hns_roce_free_buf(&srq->idx_que.buf);
+ free(srq->idx_que.bitmap);
}
static int exec_srq_create_cmd(struct ibv_context *context,
@@ -541,7 +588,7 @@ static int exec_srq_create_cmd(struct ibv_context *context,
struct hns_roce_create_srq_ex cmd_ex = {};
int ret;
- cmd_ex.buf_addr = (uintptr_t)srq->buf.buf;
+ cmd_ex.buf_addr = (uintptr_t)srq->wqe_buf.buf;
cmd_ex.que_addr = (uintptr_t)srq->idx_que.buf.buf;
cmd_ex.db_addr = (uintptr_t)srq->db;
@@ -559,57 +606,44 @@ static int exec_srq_create_cmd(struct ibv_context *context,
static struct ibv_srq *create_srq(struct ibv_context *context,
struct ibv_srq_init_attr_ex *init_attr)
{
- struct hns_roce_context *ctx = to_hr_ctx(context);
- struct ibv_srq_attr *attr = &init_attr->attr;
+ struct hns_roce_context *hr_ctx = to_hr_ctx(context);
struct hns_roce_srq *srq;
int ret;
- if (hns_roce_verify_srq(ctx, init_attr))
- return NULL;
+ ret = verify_srq_create_attr(hr_ctx, init_attr);
+ if (ret)
+ goto err;
srq = calloc(1, sizeof(*srq));
- if (!srq)
- return NULL;
+ if (!srq) {
+ ret = -ENOMEM;
+ goto err;
+ }
if (pthread_spin_init(&srq->lock, PTHREAD_PROCESS_PRIVATE))
goto err_free_srq;
- if (to_hr_dev(context->device)->hw_version == HNS_ROCE_HW_VER2)
- srq->rsv_sge = 1;
-
- srq->wqe_cnt = roundup_pow_of_two(attr->max_wr + 1);
- srq->max_gs = roundup_pow_of_two(attr->max_sge + srq->rsv_sge);
- attr->max_sge = srq->max_gs;
- attr->srq_limit = 0;
-
- ret = hns_roce_create_idx_que(srq);
- if (ret)
+ set_srq_param(context, srq, init_attr);
+ if (alloc_srq_buf(srq))
goto err_free_srq;
- ret = hns_roce_alloc_srq_buf(srq);
- if (ret)
- goto err_idx_que;
-
- srq->db = hns_roce_alloc_db(ctx, HNS_ROCE_QP_TYPE_DB);
+ srq->db = hns_roce_alloc_db(hr_ctx, HNS_ROCE_QP_TYPE_DB);
if (!srq->db)
goto err_srq_buf;
- *(srq->db) = 0;
-
- pthread_mutex_lock(&ctx->srq_table_mutex);
+ *srq->db = 0;
ret = exec_srq_create_cmd(context, srq, init_attr);
if (ret)
goto err_srq_db;
- ret = hns_roce_store_srq(ctx, srq);
+ ret = hns_roce_store_srq(hr_ctx, srq);
if (ret)
goto err_destroy_srq;
- pthread_mutex_unlock(&ctx->srq_table_mutex);
-
- srq->max_gs = attr->max_sge;
- attr->max_sge = min(attr->max_sge - srq->rsv_sge, ctx->max_srq_sge);
+ srq->max_gs = init_attr->attr.max_sge;
+ init_attr->attr.max_sge =
+ min(init_attr->attr.max_sge - srq->rsv_sge, hr_ctx->max_srq_sge);
return &srq->verbs_srq.srq;
@@ -617,20 +651,19 @@ err_destroy_srq:
ibv_cmd_destroy_srq(&srq->verbs_srq.srq);
err_srq_db:
- pthread_mutex_unlock(&ctx->srq_table_mutex);
- hns_roce_free_db(ctx, srq->db, HNS_ROCE_QP_TYPE_DB);
+ hns_roce_free_db(hr_ctx, srq->db, HNS_ROCE_QP_TYPE_DB);
err_srq_buf:
- free(srq->wrid);
- hns_roce_free_buf(&srq->buf);
-
-err_idx_que:
- free(srq->idx_que.bitmap);
- hns_roce_free_buf(&srq->idx_que.buf);
+ free_srq_buf(srq);
err_free_srq:
free(srq);
+err:
+ if (ret < 0)
+ ret = -ret;
+
+ errno = ret;
return NULL;
}
@@ -690,23 +723,14 @@ int hns_roce_u_destroy_srq(struct ibv_srq *ibv_srq)
struct hns_roce_srq *srq = to_hr_srq(ibv_srq);
int ret;
- pthread_mutex_lock(&ctx->srq_table_mutex);
-
ret = ibv_cmd_destroy_srq(ibv_srq);
- if (ret) {
- pthread_mutex_unlock(&ctx->srq_table_mutex);
+ if (ret)
return ret;
- }
hns_roce_clear_srq(ctx, srq->srqn);
- pthread_mutex_unlock(&ctx->srq_table_mutex);
-
hns_roce_free_db(ctx, srq->db, HNS_ROCE_QP_TYPE_DB);
- hns_roce_free_buf(&srq->buf);
- free(srq->wrid);
- hns_roce_free_buf(&srq->idx_que.buf);
- free(srq->idx_que.bitmap);
+ free_srq_buf(srq);
free(srq);
return 0;
--
2.30.0

69
0014-libhns-Remove-the-reserved-wqe-of-SRQ.patch
View File

@ -1,69 +0,0 @@
From d68ac72a8e4f2cf9754d3fcbbb8ff2a03e514c2f Mon Sep 17 00:00:00 2001
From: Wenpeng Liang <liangwenpeng@huawei.com>
Date: Tue, 11 May 2021 19:06:40 +0800
Subject: libhns: Remove the reserved wqe of SRQ
There is an unreasonable reserved WQE in SRQ, it should be removed.
Signed-off-by: Wenpeng Liang <liangwenpeng@huawei.com>
Signed-off-by: Weihang Li <liweihang@huawei.com>
---
providers/hns/hns_roce_u.h | 1 +
providers/hns/hns_roce_u_hw_v2.c | 4 +---
providers/hns/hns_roce_u_verbs.c | 5 ++++-
3 files changed, 6 insertions(+), 4 deletions(-)
diff --git a/providers/hns/hns_roce_u.h b/providers/hns/hns_roce_u.h
index a437727c..0d7abd81 100644
--- a/providers/hns/hns_roce_u.h
+++ b/providers/hns/hns_roce_u.h
@@ -64,6 +64,7 @@
#define HNS_ROCE_MIN_CQE_NUM 0x40
#define HNS_ROCE_V1_MIN_WQE_NUM 0x20
#define HNS_ROCE_V2_MIN_WQE_NUM 0x40
+#define HNS_ROCE_MIN_SRQ_WQE_NUM 1
#define HNS_ROCE_CQE_SIZE 0x20
#define HNS_ROCE_V3_CQE_SIZE 0x40
diff --git a/providers/hns/hns_roce_u_hw_v2.c b/providers/hns/hns_roce_u_hw_v2.c
index d4e7e4f9..2fb6cdaf 100644
--- a/providers/hns/hns_roce_u_hw_v2.c
+++ b/providers/hns/hns_roce_u_hw_v2.c
@@ -1530,10 +1530,8 @@ static int hns_roce_u_v2_destroy_qp(struct ibv_qp *ibqp)
static int hns_roce_v2_srqwq_overflow(struct hns_roce_srq *srq)
{
struct hns_roce_idx_que *idx_que = &srq->idx_que;
- unsigned int cur;
- cur = idx_que->head - idx_que->tail;
- return cur >= srq->wqe_cnt - 1;
+ return idx_que->head - idx_que->tail >= srq->wqe_cnt;
}
static int check_post_srq_valid(struct hns_roce_srq *srq,
diff --git a/providers/hns/hns_roce_u_verbs.c b/providers/hns/hns_roce_u_verbs.c
index 75b9e530..4847639b 100644
--- a/providers/hns/hns_roce_u_verbs.c
+++ b/providers/hns/hns_roce_u_verbs.c
@@ -489,6 +489,9 @@ static int verify_srq_create_attr(struct hns_roce_context *context,
attr->attr.max_sge > context->max_srq_sge)
return -EINVAL;
+ attr->attr.max_wr = max_t(uint32_t, attr->attr.max_wr,
+ HNS_ROCE_MIN_SRQ_WQE_NUM);
+
return 0;
}
@@ -498,7 +501,7 @@ static void set_srq_param(struct ibv_context *context, struct hns_roce_srq *srq,
if (to_hr_dev(context->device)->hw_version == HNS_ROCE_HW_VER2)
srq->rsv_sge = 1;
- srq->wqe_cnt = roundup_pow_of_two(attr->attr.max_wr + 1);
+ srq->wqe_cnt = roundup_pow_of_two(attr->attr.max_wr);
srq->max_gs = roundup_pow_of_two(attr->attr.max_sge + srq->rsv_sge);
srq->wqe_shift = hr_ilog32(roundup_pow_of_two(HNS_ROCE_SGE_SIZE *
srq->max_gs));
--
2.30.0

89
0015-libhns-Refactor-process-of-setting-extended-sge.patch
View File

@ -1,89 +0,0 @@
From 11c81d0e3a987f95b74e03b5e592a45029302f1d Mon Sep 17 00:00:00 2001
From: Weihang Li <liweihang@huawei.com>
Date: Fri, 14 May 2021 10:02:56 +0800
Subject: libhns: Refactor process of setting extended sge
Refactor and encapsulate the parts of getting number of extended sge a WQE
can use to make it easier to understand.
Signed-off-by: Weihang Li <liweihang@huawei.com>
---
providers/hns/hns_roce_u_verbs.c | 45 ++++++++++++++++++++------------
1 file changed, 29 insertions(+), 16 deletions(-)
diff --git a/providers/hns/hns_roce_u_verbs.c b/providers/hns/hns_roce_u_verbs.c
index 30ab072a..a8508fc5 100644
--- a/providers/hns/hns_roce_u_verbs.c
+++ b/providers/hns/hns_roce_u_verbs.c
@@ -920,31 +920,44 @@ err_alloc:
return -ENOMEM;
}
-static void set_extend_sge_param(struct hns_roce_device *hr_dev,
- struct ibv_qp_init_attr_ex *attr,
- struct hns_roce_qp *qp, unsigned int wr_cnt)
+static unsigned int get_wqe_ext_sge_cnt(struct hns_roce_qp *qp)
{
- int cnt = 0;
+ if (qp->verbs_qp.qp.qp_type == IBV_QPT_UD)
+ return qp->sq.max_gs;
+
+ if (qp->sq.max_gs > HNS_ROCE_SGE_IN_WQE)
+ return qp->sq.max_gs - HNS_ROCE_SGE_IN_WQE;
+
+ return 0;
+}
+
+static void set_ext_sge_param(struct hns_roce_device *hr_dev,
+ struct ibv_qp_init_attr_ex *attr,
+ struct hns_roce_qp *qp, unsigned int wr_cnt)
+{
+ unsigned int total_sge_cnt;
+ unsigned int wqe_sge_cnt;
+
+ qp->ex_sge.sge_shift = HNS_ROCE_SGE_SHIFT;
if (hr_dev->hw_version == HNS_ROCE_HW_VER1) {
qp->sq.max_gs = HNS_ROCE_SGE_IN_WQE;
- } else {
- qp->sq.max_gs = attr->cap.max_send_sge;
- if (attr->qp_type == IBV_QPT_UD)
- cnt = roundup_pow_of_two(wr_cnt * qp->sq.max_gs);
- else if (qp->sq.max_gs > HNS_ROCE_SGE_IN_WQE)
- cnt = roundup_pow_of_two(wr_cnt *
- (qp->sq.max_gs -
- HNS_ROCE_SGE_IN_WQE));
+ return;
}
- 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);
/* If the number of extended sge is not zero, they MUST use the
* space of HNS_HW_PAGE_SIZE at least.
*/
- qp->ex_sge.sge_cnt = cnt ?
- max(cnt, HNS_HW_PAGE_SIZE / HNS_ROCE_SGE_SIZE) : 0;
+ 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);
+ }
}
static void hns_roce_set_qp_params(struct ibv_qp_init_attr_ex *attr,
@@ -988,7 +1001,7 @@ 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_extend_sge_param(hr_dev, attr, qp, cnt);
+ set_ext_sge_param(hr_dev, 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);
--
2.30.0

139
0016-libhns-Optimize-set_sge-process.patch
View File

@ -1,139 +0,0 @@
From 3507f87f776043acd238d7c0c41cc3511f186d08 Mon Sep 17 00:00:00 2001
From: Lang Cheng <chenglang@huawei.com>
Date: Fri, 14 May 2021 10:02:57 +0800
Subject: libhns: Optimize set_sge process
Use local variables to avoid frequent ldr/str operations. And because UD's
process of setting sge is more simple then RC, set_sge() can be splited
into two functions for compiler optimization.
Signed-off-by: Lang Cheng <chenglang@huawei.com>
Signed-off-by: Weihang Li <liweihang@huawei.com>
---
providers/hns/hns_roce_u_hw_v2.c | 83 +++++++++++++++++++++++---------
1 file changed, 61 insertions(+), 22 deletions(-)
diff --git a/providers/hns/hns_roce_u_hw_v2.c b/providers/hns/hns_roce_u_hw_v2.c
index 4988943a..dc79a6f8 100644
--- a/providers/hns/hns_roce_u_hw_v2.c
+++ b/providers/hns/hns_roce_u_hw_v2.c
@@ -701,39 +701,78 @@ static int check_qp_send(struct ibv_qp *qp, struct hns_roce_context *ctx)
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)
+static void set_rc_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)
{
+ uint32_t mask = qp->ex_sge.sge_cnt - 1;
+ uint32_t index = sge_info->start_idx;
+ struct ibv_sge *sge = wr->sg_list;
+ uint32_t len = 0;
+ uint32_t cnt = 0;
+ int flag;
int i;
- sge_info->valid_num = 0;
- sge_info->total_len = 0;
+ flag = (wr->send_flags & IBV_SEND_INLINE &&
+ wr->opcode != IBV_WR_ATOMIC_FETCH_AND_ADD &&
+ wr->opcode != IBV_WR_ATOMIC_CMP_AND_SWP);
- for (i = 0; i < wr->num_sge; i++) {
- if (unlikely(!wr->sg_list[i].length))
+ for (i = 0; i < wr->num_sge; i++, sge++) {
+ if (unlikely(!sge->length))
continue;
- sge_info->total_len += wr->sg_list[i].length;
- sge_info->valid_num++;
+ len += sge->length;
+ cnt++;
- if (wr->send_flags & IBV_SEND_INLINE &&
- wr->opcode != IBV_WR_ATOMIC_FETCH_AND_ADD &&
- wr->opcode != IBV_WR_ATOMIC_CMP_AND_SWP)
+ if (flag)
continue;
- /* No inner sge in UD wqe */
- if (sge_info->valid_num <= HNS_ROCE_SGE_IN_WQE &&
- qp->verbs_qp.qp.qp_type != IBV_QPT_UD) {
- set_data_seg_v2(dseg, wr->sg_list + i);
+ if (cnt <= HNS_ROCE_SGE_IN_WQE) {
+ set_data_seg_v2(dseg, sge);
dseg++;
} else {
- dseg = get_send_sge_ex(qp, sge_info->start_idx &
- (qp->ex_sge.sge_cnt - 1));
- set_data_seg_v2(dseg, wr->sg_list + i);
- sge_info->start_idx++;
+ dseg = get_send_sge_ex(qp, index & mask);
+ set_data_seg_v2(dseg, sge);
+ index++;
}
}
+
+ sge_info->start_idx = index;
+ sge_info->valid_num = cnt;
+ sge_info->total_len = len;
+}
+
+static void set_ud_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 flag = wr->send_flags & IBV_SEND_INLINE;
+ uint32_t mask = qp->ex_sge.sge_cnt - 1;
+ uint32_t index = sge_info->start_idx;
+ struct ibv_sge *sge = wr->sg_list;
+ uint32_t len = 0;
+ uint32_t cnt = 0;
+ int i;
+
+ for (i = 0; i < wr->num_sge; i++, sge++) {
+ if (unlikely(!sge->length))
+ continue;
+
+ len += sge->length;
+ cnt++;
+
+ if (flag)
+ continue;
+
+ /* No inner sge in UD wqe */
+ dseg = get_send_sge_ex(qp, index & mask);
+ set_data_seg_v2(dseg, sge);
+ index++;
+ }
+
+ sge_info->start_idx = index;
+ sge_info->valid_num = cnt;
+ sge_info->total_len = len;
}
static int fill_ext_sge_inl_data(struct hns_roce_qp *qp,
@@ -910,7 +949,7 @@ static int fill_ud_data_seg(struct hns_roce_ud_sq_wqe *ud_sq_wqe,
UD_SQ_WQE_MSG_START_SGE_IDX_S,
sge_info->start_idx & (qp->ex_sge.sge_cnt - 1));
- set_sge((struct hns_roce_v2_wqe_data_seg *)ud_sq_wqe, qp, wr, sge_info);
+ set_ud_sge((struct hns_roce_v2_wqe_data_seg *)ud_sq_wqe, qp, wr, sge_info);
ud_sq_wqe->msg_len = htole32(sge_info->total_len);
@@ -1111,7 +1150,7 @@ static int set_rc_wqe(void *wqe, struct hns_roce_qp *qp, struct ibv_send_wr *wr,
wqe += sizeof(struct hns_roce_rc_sq_wqe);
dseg = wqe;
- set_sge(dseg, qp, wr, sge_info);
+ set_rc_sge(dseg, qp, wr, sge_info);
rc_sq_wqe->msg_len = htole32(sge_info->total_len);
--
2.30.0

27
backport-fixbug-increase-maximum-number-of-cpus-rdma.patch
View File

@ -1,27 +0,0 @@
From c381cfa26ba6163b9cc51212702e64bf1d83f838 Mon Sep 17 00:00:00 2001
From: swimlessbird <52704385+swimlessbird@users.noreply.github.com>
Date: Fri, 17 Sep 2021 14:35:05 +0800
Subject: [PATCH] ibdiags: Increase maximum number of CPUs
In modern systems, the old limit (8) is small enough, so increase
to something larger (256).
Signed-off-by: Suwan Sun <swimlessbird@gmail.com>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
---
infiniband-diags/ibsysstat.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/infiniband-diags/ibsysstat.c b/infiniband-diags/ibsysstat.c
index 6ff7ca0c4..73972d039 100644
--- a/infiniband-diags/ibsysstat.c
+++ b/infiniband-diags/ibsysstat.c
@@ -41,7 +41,7 @@
#include "ibdiag_common.h"
-#define MAX_CPUS 8
+#define MAX_CPUS 256
static struct ibmad_port *srcport;

BIN
rdma-core-35.1.tar.gz
View File

Binary file not shown.

BIN
rdma-core-41.0.tar.gz Normal file
View File

Binary file not shown.

28
rdma-core.spec
View File

@ -1,29 +1,11 @@
Name: rdma-core Name: rdma-core
Version: 35.1 Version: 41.0
Release: 3 Release: 1
Summary: RDMA core userspace libraries and daemons Summary: RDMA core userspace libraries and daemons
License: GPLv2 or BSD License: GPLv2 or BSD
Url: https://github.com/linux-rdma/rdma-core Url: https://github.com/linux-rdma/rdma-core
Source: https://github.com/linux-rdma/rdma-core/releases/download/v%{version}/%{name}-%{version}.tar.gz Source: https://github.com/linux-rdma/rdma-core/releases/download/v%{version}/%{name}-%{version}.tar.gz
Patch0: backport-fixbug-increase-maximum-number-of-cpus-rdma.patch
Patch1: 0001-Update-kernel-headers.patch
Patch2: 0002-libhns-Fix-the-ownership-of-the-head-tail-pointer-of.patch
Patch3: 0003-libhns-Fix-wrong-data-type-when-writing-doorbell.patch
Patch4: 0004-libhns-Remove-unsupported-QP-type.patch
Patch5: 0005-libhns-Avoid-using-WQE-indexes-that-exceed-the-SRQ-s.patch
Patch6: 0006-libhns-Don-t-create-RQ-for-a-QP-that-associated-with.patch
Patch7: 0007-libhns-Add-support-for-direct-wqe.patch
Patch8: 0008-libhns-Use-new-SQ-doorbell-register-for-HIP09.patch
Patch9: 0009-libhns-Bugfix-for-checking-whether-the-SRQ-is-full-w.patch
Patch10: 0010-libhns-Allow-users-to-create-a-0-depth-SRQs.patch
Patch11: 0011-libhns-Refactor-the-process-of-post_srq_recv.patch
Patch12: 0012-libhns-Set-srqlimit-to-0-when-creating-SRQ.patch
Patch13: 0013-libhns-Refactor-the-process-of-create_srq.patch
Patch14: 0014-libhns-Remove-the-reserved-wqe-of-SRQ.patch
Patch15: 0015-libhns-Refactor-process-of-setting-extended-sge.patch
Patch16: 0016-libhns-Optimize-set_sge-process.patch
BuildRequires: binutils cmake >= 2.8.11 gcc libudev-devel pkgconfig pkgconfig(libnl-3.0) 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: pkgconfig(libnl-route-3.0) valgrind-devel systemd systemd-devel
BuildRequires: python3-devel python3-Cython python3 python3-docutils perl-generators BuildRequires: python3-devel python3-Cython python3 python3-docutils perl-generators
@ -267,6 +249,12 @@ fi
%{_mandir}/* %{_mandir}/*
%changelog %changelog
* Mon Jul 25 2022 tangchengchang <tangchengchang@huawei.com> - 41.0-1
- Type: requirement
- ID: NA
- SUG: NA
- DESC: update to 41.0
* Mon Jul 11 2022 luozhengfeng <luozhengfeng@h-partners.com> - 35.1-3 * Mon Jul 11 2022 luozhengfeng <luozhengfeng@h-partners.com> - 35.1-3
- Type: bugfix - Type: bugfix
- ID: NA - ID: NA