- High-performance-network - mailweb.openeuler.org

[PATCH] libhns: Fix missing fields for SRQ WC
by Chengchang Tang 17 Jan '25

17 Jan '25

From: Xinghai Cen <cenxinghai(a)h-partners.com> mainline inclusion: libhns: Fix missing fields for SRQ WC Modify the information of some patch Fixes Signed-off-by: Xinghai Cen <cenxinghai(a)h-partners.com> --- ...x-memory-leakage-when-DCA-is-enabled.patch | 7 +- ...ump-during-QP-destruction-when-send_.patch | 4 +- ...libhns-Fix-missing-fields-for-SRQ-WC.patch | 82 +++++++++++++++++++ rdma-core.spec | 9 +- 4 files changed, 95 insertions(+), 7 deletions(-) create mode 100644 0043-libhns-Fix-missing-fields-for-SRQ-WC.patch diff --git a/0040-libhns-Fix-memory-leakage-when-DCA-is-enabled.patch b/0040-libhns-Fix-memory-leakage-when-DCA-is-enabled.patch index e1763f7..f046264 100644 --- a/0040-libhns-Fix-memory-leakage-when-DCA-is-enabled.patch +++ b/0040-libhns-Fix-memory-leakage-when-DCA-is-enabled.patch @@ -1,4 +1,4 @@ -From edaf09dbfc7203ea68becfcb56eecf4af31ba555 Mon Sep 17 00:00:00 2001 +From af20dd32df73ff72d35a430a5fb87ac42d70cdf4 Mon Sep 17 00:00:00 2001 From: wenglianfa <wenglianfa(a)huawei.com> Date: Thu, 25 Jul 2024 11:06:01 +0800 Subject: [PATCH] libhns: Fix memory leakage when DCA is enabled @@ -13,8 +13,7 @@ After DCA is enabled and a QP is created, the memory block applied for DCA is not free when the QP is destroyed. Here fix it. -Fixes: 41e39ab792c8 ("[BigDipperV3R9,NeZha][ROCE] libhns: Add support for at taching QP's WQE buffer") - +Fixes: 2783884a97e7 ("libhns: Add support for attaching QP's WQE buffer") Signed-off-by: wenglianfa <wenglianfa(a)huawei.com> Signed-off-by: Xinghai Cen <cenxinghai(a)h-partners.com> --- @@ -36,5 +35,5 @@ index e30880c..c733b21 100644 static int qp_alloc_wqe(struct ibv_qp_init_attr_ex *attr, -- -2.25.1 +2.33.0 diff --git a/0041-libhns-Fix-coredump-during-QP-destruction-when-send_.patch b/0041-libhns-Fix-coredump-during-QP-destruction-when-send_.patch index b3ee84f..2e9f546 100644 --- a/0041-libhns-Fix-coredump-during-QP-destruction-when-send_.patch +++ b/0041-libhns-Fix-coredump-during-QP-destruction-when-send_.patch @@ -1,4 +1,4 @@ -From 263479c6fb4712528ccae276960ec94fd77afc51 Mon Sep 17 00:00:00 2001 +From 83784fc2538d24f3f06f023c21cc045d5b7f44ce Mon Sep 17 00:00:00 2001 From: Yuyu Li <liyuyu6(a)huawei.com> Date: Mon, 25 Nov 2024 16:13:48 +0800 Subject: [PATCH] libhns: Fix coredump during QP destruction when send_cq == @@ -24,7 +24,7 @@ coredump info: 0x0000ffff8feae39c in __ibv_create_qp_1_1 0x0000000000401420 in test_ctrl_path -Fixes: 95e05809d2d2 ("[BigDipperV3R9,NeZha][ROCE] libhns: Support reporting wc as software mode") +Fixes: 5494e44cf97e ("Support reporting wc as software mode.") Signed-off-by: Yuyu Li <liyuyu6(a)huawei.com> Signed-off-by: Xinghai Cen <cenxinghai(a)h-partners.com> --- diff --git a/0043-libhns-Fix-missing-fields-for-SRQ-WC.patch b/0043-libhns-Fix-missing-fields-for-SRQ-WC.patch new file mode 100644 index 0000000..4058094 --- /dev/null +++ b/0043-libhns-Fix-missing-fields-for-SRQ-WC.patch @@ -0,0 +1,82 @@ +From b52618371517527ce8ea4b8f5bd2571c7f69a2ba Mon Sep 17 00:00:00 2001 +From: wenglianfa <wenglianfa(a)huawei.com> +Date: Wed, 15 Jan 2025 15:55:29 +0800 +Subject: [PATCH] libhns: Fix missing fields for SRQ WC + +mainline inclusion +from mainline-master +commit 65a7ce99cf4bfd6748346206f546e51c0a82c993 +category: bugfix +bugzilla: https://gitee.com/src-openeuler/rdma-core/issues/IBIEA4 +CVE: NA +Reference: https://github.com/linux-rdma/rdma-core/pull/1543/commits/65a7ce99cf4bfd674… + +---------------------------------------------------------------------- + +The sl and src_qpn fields in recv-WC are not filled when the QP is UD +and has an SRQ. Here fix it. + +In addition, UD QP does not support RQ INLINE and CQE INLINE features. +Reorder the related if-else statements to reduce the number of +conditional checks in IO path. + +Fixes: 061f7e1757ca ("libhns: Refactor the poll one interface") +Signed-off-by: wenglianfa <wenglianfa(a)huawei.com> +Signed-off-by: Junxian Huang <huangjunxian6(a)hisilicon.com> +Signed-off-by: Xinghai Cen <cenxinghai(a)h-partners.com> +--- + providers/hns/hns_roce_u_hw_v2.c | 13 ++++++++----- + 1 file changed, 8 insertions(+), 5 deletions(-) + +diff --git a/providers/hns/hns_roce_u_hw_v2.c b/providers/hns/hns_roce_u_hw_v2.c +index 0628646..aadea7a 100644 +--- a/providers/hns/hns_roce_u_hw_v2.c ++++ b/providers/hns/hns_roce_u_hw_v2.c +@@ -519,7 +519,8 @@ static void parse_for_ud_qp(struct hns_roce_v2_cqe *cqe, struct ibv_wc *wc) + } + + static void parse_cqe_for_srq(struct hns_roce_v2_cqe *cqe, struct ibv_wc *wc, +- struct hns_roce_srq *srq) ++ struct hns_roce_srq *srq, ++ struct hns_roce_qp *hr_qp) + { + uint32_t wqe_idx; + +@@ -529,6 +530,8 @@ static void parse_cqe_for_srq(struct hns_roce_v2_cqe *cqe, struct ibv_wc *wc, + + if (hr_reg_read(cqe, CQE_CQE_INLINE)) + handle_recv_cqe_inl_from_srq(cqe, srq); ++ else if (hr_qp->verbs_qp.qp.qp_type == IBV_QPT_UD) ++ parse_for_ud_qp(cqe, wc); + } + + static void parse_cqe_for_resp(struct hns_roce_v2_cqe *cqe, struct ibv_wc *wc, +@@ -540,13 +543,13 @@ static void parse_cqe_for_resp(struct hns_roce_v2_cqe *cqe, struct ibv_wc *wc, + wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)]; + ++wq->tail; + +- if (hr_qp->verbs_qp.qp.qp_type == IBV_QPT_UD) +- parse_for_ud_qp(cqe, wc); +- + if (hr_reg_read(cqe, CQE_CQE_INLINE)) + handle_recv_cqe_inl_from_rq(cqe, hr_qp); + else if (hr_reg_read(cqe, CQE_RQ_INLINE)) + handle_recv_rq_inl(cqe, hr_qp); ++ else if (hr_qp->verbs_qp.qp.qp_type == IBV_QPT_UD) ++ parse_for_ud_qp(cqe, wc); ++ + } + + static void parse_cqe_for_req(struct hns_roce_v2_cqe *cqe, struct ibv_wc *wc, +@@ -753,7 +756,7 @@ static int parse_cqe_for_cq(struct hns_roce_context *ctx, struct hns_roce_cq *cq + return V2_CQ_POLL_ERR; + + if (srq) +- parse_cqe_for_srq(cqe, wc, srq); ++ parse_cqe_for_srq(cqe, wc, srq, cur_qp); + else + parse_cqe_for_resp(cqe, wc, cur_qp); + } +-- +2.33.0 + diff --git a/rdma-core.spec b/rdma-core.spec index 035faf7..8fd61b8 100644 --- a/rdma-core.spec +++ b/rdma-core.spec @@ -1,6 +1,6 @@ Name: rdma-core Version: 50.0 -Release: 20 +Release: 21 Summary: RDMA core userspace libraries and daemons License: GPL-2.0-only OR BSD-2-Clause AND BSD-3-Clause Url: https://github.com/linux-rdma/rdma-core @@ -48,6 +48,7 @@ patch39: 0039-libhns-Fix-the-exception-branch-of-wr_start-is-not-l.patch patch40: 0040-libhns-Fix-memory-leakage-when-DCA-is-enabled.patch patch41: 0041-libhns-Fix-coredump-during-QP-destruction-when-send_.patch patch42: 0042-libhns-Add-error-logs-to-help-diagnosis.patch +patch43: 0043-libhns-Fix-missing-fields-for-SRQ-WC.patch BuildRequires: binutils cmake >= 2.8.11 gcc libudev-devel pkgconfig pkgconfig(libnl-3.0) BuildRequires: pkgconfig(libnl-route-3.0) systemd systemd-devel @@ -623,6 +624,12 @@ fi %doc %{_docdir}/%{name}-%{version}/70-persistent-ipoib.rules %changelog +* Fri Jan 17 2025 Xinghai Cen <cenxinghai(a)h-partners.com> - 50.0-21 +- Type: bugfix +- ID: NA +- SUG: NA +- DESC: Fix missing fields for SRQ WC + * Wed Jan 08 2025 Funda Wang <fundawang(a)yeah.net> - 50.0-20 - Type: bugfix - ID: NA -- 2.33.0

1 0

libhns: Fix missing fields for SRQ WC
by Junxian Huang 15 Jan '25

15 Jan '25

From: wenglianfa <wenglianfa(a)huawei.com> The sl and src_qpn fields in recv-WC are not filled when the QP is UD and has an SRQ. Here fix it. In addition, UD QP does not support RQ INLINE and CQE INLINE features. Reorder the related if-else statements to reduce the number of conditional checks in IO path. Fixes: 061f7e1757ca ("libhns: Refactor the poll one interface") Signed-off-by: wenglianfa <wenglianfa(a)huawei.com> Signed-off-by: Junxian Huang <huangjunxian6(a)hisilicon.com> --- providers/hns/hns_roce_u_hw_v2.c | 13 ++++++++----- 1 file changed, 8 insertions(+), 5 deletions(-) diff --git a/providers/hns/hns_roce_u_hw_v2.c b/providers/hns/hns_roce_u_hw_v2.c index 1d5c9a6a5..f46337226 100644 --- a/providers/hns/hns_roce_u_hw_v2.c +++ b/providers/hns/hns_roce_u_hw_v2.c @@ -510,12 +510,15 @@ static void parse_for_ud_qp(struct hns_roce_v2_cqe *cqe, struct ibv_wc *wc) } static void parse_cqe_for_srq(struct hns_roce_v2_cqe *cqe, struct ibv_wc *wc, - struct hns_roce_srq *srq) + struct hns_roce_srq *srq, + struct hns_roce_qp *hr_qp) { uint32_t wqe_idx; if (hr_reg_read(cqe, CQE_CQE_INLINE)) handle_recv_cqe_inl_from_srq(cqe, srq); + else if (hr_qp->verbs_qp.qp.qp_type == IBV_QPT_UD) + parse_for_ud_qp(cqe, wc); wqe_idx = hr_reg_read(cqe, CQE_WQE_IDX); wc->wr_id = srq->wrid[wqe_idx & (srq->wqe_cnt - 1)]; @@ -531,13 +534,13 @@ static void parse_cqe_for_resp(struct hns_roce_v2_cqe *cqe, struct ibv_wc *wc, wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)]; ++wq->tail; - if (hr_qp->verbs_qp.qp.qp_type == IBV_QPT_UD) - parse_for_ud_qp(cqe, wc); - if (hr_reg_read(cqe, CQE_CQE_INLINE)) handle_recv_cqe_inl_from_rq(cqe, hr_qp); else if (hr_reg_read(cqe, CQE_RQ_INLINE)) handle_recv_rq_inl(cqe, hr_qp); + else if (hr_qp->verbs_qp.qp.qp_type == IBV_QPT_UD) + parse_for_ud_qp(cqe, wc); + } static void parse_cqe_for_req(struct hns_roce_v2_cqe *cqe, struct ibv_wc *wc, @@ -669,7 +672,7 @@ static int parse_cqe_for_cq(struct hns_roce_context *ctx, struct hns_roce_cq *cq return V2_CQ_POLL_ERR; if (srq) - parse_cqe_for_srq(cqe, wc, srq); + parse_cqe_for_srq(cqe, wc, srq, cur_qp); else parse_cqe_for_resp(cqe, wc, cur_qp); } -- 2.33.0

1 0

[PATCH] Perftest: modify --source_ip to --bind_source_ip to fix init connection establishment with specific interface
by Chengchang Tang 10 Jan '25

10 Jan '25

From: Guofeng Yue <yueguofeng(a)h-partners.com> mainline inclusion commit ba4580a6c4f16ab7791e4d809eda91d586c1f04f category: bugfix bugzilla: https://gitee.com/src-openeuler/perftest/issues/IBGXBU Reference: https://github.com/linux-rdma/perftest/commit/8ff29c1603215c012a886284b7184… ---------------------------- When there are several network interface with different subnet address, perftest tools will always choose default route even I add --source_ip option to ask perftest to bind an interface. I found that there are two options that use the same name "--source_ip". Therefore, change --source_ip to --bind_source_ip to fix init connection establishment with specific interface. Signed-off-by: Guofeng Yue <yueguofeng(a)h-partners.com> --- ...-to-bind_sounce_ip-to-fix-init-conne.patch | 35 +++++++++++++++++++ perftest.spec | 9 ++++- 2 files changed, 43 insertions(+), 1 deletion(-) create mode 100644 0015-modify-source_ip-to-bind_sounce_ip-to-fix-init-conne.patch diff --git a/0015-modify-source_ip-to-bind_sounce_ip-to-fix-init-conne.patch b/0015-modify-source_ip-to-bind_sounce_ip-to-fix-init-conne.patch new file mode 100644 index 0000000..95e5c41 --- /dev/null +++ b/0015-modify-source_ip-to-bind_sounce_ip-to-fix-init-conne.patch @@ -0,0 +1,35 @@ +From ee491fb51e068f1ff3d0277a3a3b33ee0bb38779 Mon Sep 17 00:00:00 2001 +From: "Huai-En, Tseng" <huaien.tseng(a)shopee.com> +Date: Thu, 18 May 2023 11:31:31 +0800 +Subject: [PATCH] modify --source_ip to --bind_sounce_ip to fix init connection + establishment with specific interface + +--- + src/perftest_parameters.c | 4 ++-- + 1 file changed, 2 insertions(+), 2 deletions(-) + +diff --git a/src/perftest_parameters.c b/src/perftest_parameters.c +index 0a18b2a..73907ca 100755 +--- a/src/perftest_parameters.c ++++ b/src/perftest_parameters.c +@@ -476,7 +476,7 @@ static void usage(const char *argv0, VerbType verb, TestType tst, int connection + + // please note it is a different source_ip from raw_ethernet case + if (connection_type != RawEth) { +- printf(" --source_ip "); ++ printf(" --bind_source_ip "); + printf(" Source IP of the interface used for connection establishment. By default taken from routing table.\n"); + } + +@@ -2386,7 +2386,7 @@ int parser(struct perftest_parameters *user_param,char *argv[], int argc) + #ifdef HAVE_HNSDV + { .name = "congest_type", .has_arg = 1, .flag = &congest_type_flag, .val = 1}, + #endif +- {.name = "source_ip", .has_arg = 1, .flag = &source_ip_flag, .val = 1}, ++ {.name = "bind_source_ip", .has_arg = 1, .flag = &source_ip_flag, .val = 1}, + {0} + }; + c = getopt_long(argc,argv,"w:y:p:d:i:m:s:n:t:u:S:x:c:q:I:o:M:r:Q:A:l:D:f:B:T:L:E:J:j:K:k:X:W:aFegzRvhbNVCHUOZP",long_options,NULL); +-- +2.33.0 + diff --git a/perftest.spec b/perftest.spec index 737c993..6f48fac 100644 --- a/perftest.spec +++ b/perftest.spec @@ -1,6 +1,6 @@ Name: perftest Version: 4.5 -Release: 11 +Release: 12 License: GPLv2 or BSD Summary: RDMA Performance Testing Tools Url: https://github.com/linux-rdma/perftest @@ -20,6 +20,7 @@ Patch11: 0011-Perftest-Fix-rx_depth-check-for-XRC.patch Patch12: 0012-Perftest-Add-support-for-TD-lock-free-mode.patch Patch13: 0013-Perftest-Fix-TD-lock-free-mode-not-working-for-QP.patch Patch14: 0014-Perftest-Fix-failure-in-creating-cq-when-create-cq-e.patch +Patch15: 0015-modify-source_ip-to-bind_sounce_ip-to-fix-init-conne.patch BuildRequires: automake gcc libibverbs-devel >= 1.2.0 librdmacm-devel >= 1.0.21 libibumad-devel >= 1.3.10.2 BuildRequires: pciutils-devel libibverbs librdmacm libibumad @@ -47,6 +48,12 @@ done %_bindir/* %changelog +* Thu Jun 10 2025 Guofeng Yue <yueguofeng(a)h-partners.com> - 4.5-12 +- Type: bugfix +- ID: NA +- SUG: NA +- DESC: modify --source_ip to --bind_source_ip to fix init connection establishment with specific interface + * Mon Sep 2 2024 Xinghai Cen <cenxinghai(a)h-partners.com> - 4.5-11 - Type: bugfix - ID: NA -- 2.33.0

1 0

[PATCH] libhns: Add error logs to help diagnosis
by Chengchang Tang 03 Jan '25

03 Jan '25

From: Xinghai Cen <cenxinghai(a)h-partners.com> Add error logs to help diagnosis. Signed-off-by: Xinghai Cen <cenxinghai(a)h-partners.com> --- ...hns-Add-error-logs-to-help-diagnosis.patch | 242 ++++++++++++++++++ rdma-core.spec | 9 +- 2 files changed, 250 insertions(+), 1 deletion(-) create mode 100644 0042-libhns-Add-error-logs-to-help-diagnosis.patch diff --git a/0042-libhns-Add-error-logs-to-help-diagnosis.patch b/0042-libhns-Add-error-logs-to-help-diagnosis.patch new file mode 100644 index 0000000..9d880ea --- /dev/null +++ b/0042-libhns-Add-error-logs-to-help-diagnosis.patch @@ -0,0 +1,242 @@ +From 60c45b5f7c2cd0c2e7139d472406f071f327bb91 Mon Sep 17 00:00:00 2001 +From: Junxian Huang <huangjunxian6(a)hisilicon.com> +Date: Fri, 27 Dec 2024 14:02:29 +0800 +Subject: [PATCH] libhns: Add error logs to help diagnosis + +mainline inclusion +from mainline-master +commit 7849f1b17f89b8baa0065adaf9cd04204698ea82 +category: feature +bugzilla: https://gitee.com/src-openeuler/rdma-core/issues/IBFGPH +CVE: NA + +Reference: https://github.com/linux-rdma/rdma-core/pull/1533/commits/7849f1b17f89b8baa… + +---------------------------------------------------------------------- + +Add error logs to help diagnosis. + +Signed-off-by: Junxian Huang <huangjunxian6(a)hisilicon.com> +--- + providers/hns/hns_roce_u.c | 4 +- + providers/hns/hns_roce_u_hw_v2.c | 3 ++ + providers/hns/hns_roce_u_verbs.c | 79 ++++++++++++++++++++++++++------ + 3 files changed, 70 insertions(+), 16 deletions(-) + +diff --git a/providers/hns/hns_roce_u.c b/providers/hns/hns_roce_u.c +index e219b9e..ec995e7 100644 +--- a/providers/hns/hns_roce_u.c ++++ b/providers/hns/hns_roce_u.c +@@ -424,8 +424,10 @@ static struct verbs_context *hns_roce_alloc_context(struct ibv_device *ibdev, + + context->uar = mmap(NULL, hr_dev->page_size, PROT_READ | PROT_WRITE, + MAP_SHARED, cmd_fd, 0); +- if (context->uar == MAP_FAILED) ++ if (context->uar == MAP_FAILED) { ++ verbs_err(&context->ibv_ctx, "failed to mmap uar page.\n"); + goto err_set_attr; ++ } + + if (init_dca_context(context, cmd_fd, + &resp, ctx_attr, hr_dev->page_size)) +diff --git a/providers/hns/hns_roce_u_hw_v2.c b/providers/hns/hns_roce_u_hw_v2.c +index c746e03..0628646 100644 +--- a/providers/hns/hns_roce_u_hw_v2.c ++++ b/providers/hns/hns_roce_u_hw_v2.c +@@ -3057,6 +3057,9 @@ static int fill_send_wr_ops(const struct ibv_qp_init_attr_ex *attr, + fill_send_wr_ops_ud(qp_ex); + break; + default: ++ verbs_err(verbs_get_ctx(qp_ex->qp_base.context), ++ "QP type %d not supported for qp_ex send ops.\n", ++ attr->qp_type); + return -EOPNOTSUPP; + } + +diff --git a/providers/hns/hns_roce_u_verbs.c b/providers/hns/hns_roce_u_verbs.c +index c733b21..e9acfab 100644 +--- a/providers/hns/hns_roce_u_verbs.c ++++ b/providers/hns/hns_roce_u_verbs.c +@@ -422,8 +422,11 @@ static int verify_cq_create_attr(struct ibv_cq_init_attr_ex *attr, + { + struct hns_roce_pad *pad = to_hr_pad(attr->parent_domain); + +- if (!attr->cqe || attr->cqe > context->max_cqe) ++ if (!attr->cqe || attr->cqe > context->max_cqe) { ++ verbs_err(&context->ibv_ctx, "unsupported cq depth %u.\n", ++ attr->cqe); + return EINVAL; ++ } + + if (!check_comp_mask(attr->comp_mask, CREATE_CQ_SUPPORTED_COMP_MASK)) { + verbs_err(&context->ibv_ctx, "unsupported cq comps 0x%x\n", +@@ -431,8 +434,11 @@ static int verify_cq_create_attr(struct ibv_cq_init_attr_ex *attr, + return EOPNOTSUPP; + } + +- if (!check_comp_mask(attr->wc_flags, CREATE_CQ_SUPPORTED_WC_FLAGS)) ++ if (!check_comp_mask(attr->wc_flags, CREATE_CQ_SUPPORTED_WC_FLAGS)) { ++ verbs_err(&context->ibv_ctx, "unsupported wc flags 0x%llx.\n", ++ attr->wc_flags); + return EOPNOTSUPP; ++ } + + if (attr->comp_mask & IBV_CQ_INIT_ATTR_MASK_PD) { + if (!pad) { +@@ -504,8 +510,11 @@ static int exec_cq_create_cmd(struct ibv_context *context, + ret = ibv_cmd_create_cq_ex(context, attr, &cq->verbs_cq, + &cmd_ex.ibv_cmd, sizeof(cmd_ex), + &resp_ex.ibv_resp, sizeof(resp_ex), 0); +- if (ret) ++ if (ret) { ++ verbs_err(verbs_get_ctx(context), ++ "failed to exec create cq cmd, ret = %d.\n", ret); + return ret; ++ } + + cq->cqn = resp_drv->cqn; + cq->flags = resp_drv->cap_flags; +@@ -724,13 +733,20 @@ 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) ++ attr->srq_type != IBV_SRQT_XRC) { ++ verbs_err(&context->ibv_ctx, ++ "unsupported srq type, type = %d.\n", attr->srq_type); + return -EINVAL; ++ } + + if (!attr->attr.max_sge || + attr->attr.max_wr > context->max_srq_wr || +- attr->attr.max_sge > context->max_srq_sge) ++ attr->attr.max_sge > context->max_srq_sge) { ++ verbs_err(&context->ibv_ctx, ++ "invalid srq attr size, max_wr = %u, max_sge = %u.\n", ++ attr->attr.max_wr, attr->attr.max_sge); + return -EINVAL; ++ } + + attr->attr.max_wr = max_t(uint32_t, attr->attr.max_wr, + HNS_ROCE_MIN_SRQ_WQE_NUM); +@@ -862,8 +878,12 @@ static int exec_srq_create_cmd(struct ibv_context *context, + ret = ibv_cmd_create_srq_ex(context, &srq->verbs_srq, init_attr, + &cmd_ex.ibv_cmd, sizeof(cmd_ex), + &resp_ex.ibv_resp, sizeof(resp_ex)); +- if (ret) ++ if (ret) { ++ verbs_err(verbs_get_ctx(context), ++ "failed to exec create srq cmd, ret = %d.\n", ++ ret); + return ret; ++ } + + srq->srqn = resp_ex.srqn; + srq->cap_flags = resp_ex.cap_flags; +@@ -1086,9 +1106,12 @@ static int check_qp_create_mask(struct hns_roce_context *ctx, + struct ibv_qp_init_attr_ex *attr) + { + struct hns_roce_device *hr_dev = to_hr_dev(ctx->ibv_ctx.context.device); ++ int ret = 0; + +- if (!check_comp_mask(attr->comp_mask, CREATE_QP_SUP_COMP_MASK)) +- return EOPNOTSUPP; ++ if (!check_comp_mask(attr->comp_mask, CREATE_QP_SUP_COMP_MASK)) { ++ ret = EOPNOTSUPP; ++ goto out; ++ } + + if (attr->comp_mask & IBV_QP_INIT_ATTR_SEND_OPS_FLAGS && + !check_comp_mask(attr->send_ops_flags, SEND_OPS_FLAG_MASK)) +@@ -1102,17 +1125,21 @@ static int check_qp_create_mask(struct hns_roce_context *ctx, + case IBV_QPT_RC: + case IBV_QPT_XRC_SEND: + if (!(attr->comp_mask & IBV_QP_INIT_ATTR_PD)) +- return EINVAL; ++ ret = EINVAL; + break; + case IBV_QPT_XRC_RECV: + if (!(attr->comp_mask & IBV_QP_INIT_ATTR_XRCD)) +- return EINVAL; ++ ret = EINVAL; + break; + default: + return EOPNOTSUPP; + } + +- return 0; ++out: ++ if (ret) ++ verbs_err(&ctx->ibv_ctx, "invalid comp_mask 0x%x.\n", ++ attr->comp_mask); ++ return ret; + } + + static int hns_roce_qp_has_rq(struct ibv_qp_init_attr_ex *attr) +@@ -1137,8 +1164,13 @@ static int verify_qp_create_cap(struct hns_roce_context *ctx, + if (cap->max_send_wr > ctx->max_qp_wr || + cap->max_recv_wr > ctx->max_qp_wr || + cap->max_send_sge > ctx->max_sge || +- cap->max_recv_sge > ctx->max_sge) ++ cap->max_recv_sge > ctx->max_sge) { ++ verbs_err(&ctx->ibv_ctx, ++ "invalid qp cap size, max_send/recv_wr = {%u, %u}, max_send/recv_sge = {%u, %u}.\n", ++ cap->max_send_wr, cap->max_recv_wr, ++ cap->max_send_sge, cap->max_recv_sge); + return -EINVAL; ++ } + + has_rq = hns_roce_qp_has_rq(attr); + if (!has_rq) { +@@ -1147,12 +1179,20 @@ static int verify_qp_create_cap(struct hns_roce_context *ctx, + } + + min_wqe_num = HNS_ROCE_V2_MIN_WQE_NUM; +- if (cap->max_send_wr < min_wqe_num) ++ if (cap->max_send_wr < min_wqe_num) { ++ verbs_debug(&ctx->ibv_ctx, ++ "change sq depth from %u to minimum %u.\n", ++ cap->max_send_wr, min_wqe_num); + cap->max_send_wr = min_wqe_num; ++ } + + if (cap->max_recv_wr) { +- if (cap->max_recv_wr < min_wqe_num) ++ if (cap->max_recv_wr < min_wqe_num) { ++ verbs_debug(&ctx->ibv_ctx, ++ "change rq depth from %u to minimum %u.\n", ++ cap->max_recv_wr, min_wqe_num); + cap->max_recv_wr = min_wqe_num; ++ } + + if (!cap->max_recv_sge) + return -EINVAL; +@@ -1646,6 +1686,11 @@ static int qp_exec_create_cmd(struct ibv_qp_init_attr_ex *attr, + ret = ibv_cmd_create_qp_ex2(&ctx->ibv_ctx.context, &qp->verbs_qp, attr, + &cmd_ex.ibv_cmd, sizeof(cmd_ex), + &resp_ex.ibv_resp, sizeof(resp_ex)); ++ if (ret) { ++ verbs_err(&ctx->ibv_ctx, ++ "failed to exec create qp cmd, ret = %d.\n", ret); ++ return ret; ++ } + + qp->flags = resp_ex.drv_payload.cap_flags; + *dwqe_mmap_key = resp_ex.drv_payload.dwqe_mmap_key; +@@ -1707,8 +1752,12 @@ static int mmap_dwqe(struct ibv_context *ibv_ctx, struct hns_roce_qp *qp, + { + 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) ++ if (qp->dwqe_page == MAP_FAILED) { ++ verbs_err(verbs_get_ctx(ibv_ctx), ++ "failed to mmap direct wqe page, QPN = %u.\n", ++ qp->verbs_qp.qp.qp_num); + return -EINVAL; ++ } + + return 0; + } +-- +2.33.0 + diff --git a/rdma-core.spec b/rdma-core.spec index 121f2f5..db8e113 100644 --- a/rdma-core.spec +++ b/rdma-core.spec @@ -1,6 +1,6 @@ Name: rdma-core Version: 50.0 -Release: 17 +Release: 18 Summary: RDMA core userspace libraries and daemons License: GPL-2.0-only OR BSD-2-Clause AND BSD-3-Clause Url: https://github.com/linux-rdma/rdma-core @@ -47,6 +47,7 @@ patch38: 0038-libhns-Fix-reference-to-uninitialized-cq-pointer.patch patch39: 0039-libhns-Fix-the-exception-branch-of-wr_start-is-not-l.patch patch40: 0040-libhns-Fix-memory-leakage-when-DCA-is-enabled.patch patch41: 0041-libhns-Fix-coredump-during-QP-destruction-when-send_.patch +patch42: 0042-libhns-Add-error-logs-to-help-diagnosis.patch BuildRequires: binutils cmake >= 2.8.11 gcc libudev-devel pkgconfig pkgconfig(libnl-3.0) BuildRequires: pkgconfig(libnl-route-3.0) systemd systemd-devel @@ -626,6 +627,12 @@ fi %doc %{_docdir}/%{name}-%{version}/70-persistent-ipoib.rules %changelog +* Fri Jan 3 2025 Xinghai Cen <cenxinghai(a)h-partners.com> - 50.0-18 +- Type: requirement +- ID: NA +- SUG: NA +- DESC: Add error logs to help diagnosis + * Thu Nov 28 2024 Xinghai Cen <cenxinghai(a)h-partners.com> - 50.0-17 - Type: bugfix - ID: NA -- 2.33.0

1 0

[PATCH] libhns: Fixed two bugs in libhns
by Chengchang Tang 03 Jan '25

03 Jan '25

From: Xinghai Cen <cenxinghai(a)h-partners.com> Fixed two bugs in libhns: libhns: Fix bypassed vendor check in hnsdv_query_device() libhns: Fix coredump during QP destruction when send_cq： == recv_cq Signed-off-by: Xinghai Cen <cenxinghai(a)h-partners.com> --- ...sed-vendor-check-in-hnsdv_query_devi.patch | 40 ++++++++++++++ ...ump-during-QP-destruction-when-send_.patch | 53 +++++++++++++++++++ rdma-core.spec | 10 +++- 3 files changed, 102 insertions(+), 1 deletion(-) create mode 100644 0089-libhns-Fix-bypassed-vendor-check-in-hnsdv_query_devi.patch create mode 100644 0090-libhns-Fix-coredump-during-QP-destruction-when-send_.patch diff --git a/0089-libhns-Fix-bypassed-vendor-check-in-hnsdv_query_devi.patch b/0089-libhns-Fix-bypassed-vendor-check-in-hnsdv_query_devi.patch new file mode 100644 index 0000000..14f2edc --- /dev/null +++ b/0089-libhns-Fix-bypassed-vendor-check-in-hnsdv_query_devi.patch @@ -0,0 +1,40 @@ +From 485cddd47c83d6f229450b28d55d8e07f60ddcc0 Mon Sep 17 00:00:00 2001 +From: Yuyu Li <liyuyu6(a)huawei.com> +Date: Thu, 21 Nov 2024 21:37:15 +0800 +Subject: [PATCH] libhns: Fix bypassed vendor check in hnsdv_query_device() + +driver inclusion +category: bugfix +bugzilla: https://gitee.com/src-openeuler/rdma-core/issues/IBF87T + +-------------------------------------------------------------------------- + +The device vendor check is actually bypassed currently due +to the wrong if-condition. It should be a '||' statement. + +Fixes: 19e1eabc154f ("libhns: Add input parameter check for hnsdv_query_device()") +Signed-off-by: Yuyu Li <liyuyu6(a)huawei.com> +--- + providers/hns/hns_roce_u_verbs.c | 4 ++-- + 1 file changed, 2 insertions(+), 2 deletions(-) + +diff --git a/providers/hns/hns_roce_u_verbs.c b/providers/hns/hns_roce_u_verbs.c +index 090efbf..a6afce2 100644 +--- a/providers/hns/hns_roce_u_verbs.c ++++ b/providers/hns/hns_roce_u_verbs.c +@@ -128,10 +128,10 @@ int hnsdv_query_device(struct ibv_context *context, + struct hns_roce_context *ctx = context ? to_hr_ctx(context) : NULL; + struct hns_roce_device *hr_dev; + +- if (!ctx || !attrs_out) ++ if (!ctx || !context->device || !attrs_out) + return EINVAL; + +- if (!context->device && !is_hns_dev(context->device)) { ++ if (!is_hns_dev(context->device)) { + verbs_err(verbs_get_ctx(context), "not a HNS RoCE device!\n"); + return EOPNOTSUPP; + } +-- +2.33.0 + diff --git a/0090-libhns-Fix-coredump-during-QP-destruction-when-send_.patch b/0090-libhns-Fix-coredump-during-QP-destruction-when-send_.patch new file mode 100644 index 0000000..4909bd3 --- /dev/null +++ b/0090-libhns-Fix-coredump-during-QP-destruction-when-send_.patch @@ -0,0 +1,53 @@ +From ad5055f9b32ab0915803575385985fb10a29434a Mon Sep 17 00:00:00 2001 +From: Yuyu Li <liyuyu6(a)huawei.com> +Date: Mon, 25 Nov 2024 15:42:16 +0800 +Subject: [PATCH] libhns: Fix coredump during QP destruction when send_cq + == recv_cq + +driver inclusion +category: bugfix +bugzilla: https://gitee.com/src-openeuler/rdma-core/issues/IBF87T + +-------------------------------------------------------------------------- + +If the specified send CQ and recv CQ are both +the same CQ, the QP node in SCQ is not deleted. +which causes a segfault to occur when recreating +the QP. Here fix it. + +coredump info: +0x0000ffff8fbc37d4 in list_add_before_ +0x0000ffff8fbc381c in list_add_tail_ +0x0000ffff8fbc9d9c in add_qp_to_cq_list +0x0000ffff8fbca008 in create_qp +0x0000ffff8fbca110 in hns_roce_u_create_qp +0x0000ffff8feae39c in __ibv_create_qp_1_1 +0x0000000000401420 in test_ctrl_path + +Fixes: 5bebdb5ba77b ("libhns: Support reporting wc as software mode") +Signed-off-by: Yuyu Li <liyuyu6(a)huawei.com> +--- + providers/hns/hns_roce_u_hw_v2.c | 7 ++++--- + 1 file changed, 4 insertions(+), 3 deletions(-) + +diff --git a/providers/hns/hns_roce_u_hw_v2.c b/providers/hns/hns_roce_u_hw_v2.c +index 8f071e1..48a7566 100644 +--- a/providers/hns/hns_roce_u_hw_v2.c ++++ b/providers/hns/hns_roce_u_hw_v2.c +@@ -2033,9 +2033,10 @@ static int hns_roce_u_v2_destroy_qp(struct ibv_qp *ibqp) + list_del(&qp->rcq_node); + } + +- if (ibqp->send_cq && ibqp->send_cq != ibqp->recv_cq) { +- __hns_roce_v2_cq_clean(to_hr_cq(ibqp->send_cq), ibqp->qp_num, +- NULL); ++ if (ibqp->send_cq) { ++ if (ibqp->send_cq != ibqp->recv_cq) ++ __hns_roce_v2_cq_clean(to_hr_cq(ibqp->send_cq), ibqp->qp_num, ++ NULL); + list_del(&qp->scq_node); + } + +-- +2.33.0 + diff --git a/rdma-core.spec b/rdma-core.spec index 229810a..5e7be99 100644 --- a/rdma-core.spec +++ b/rdma-core.spec @@ -1,6 +1,6 @@ Name: rdma-core Version: 41.0 -Release: 28 +Release: 29 Summary: RDMA core userspace libraries and daemons License: GPLv2 or BSD Url: https://github.com/linux-rdma/rdma-core @@ -94,6 +94,8 @@ patch85: 0085-libhns-Fix-memory-leakage-when-DCA-is-enabled.patch patch86: 0086-libhns-Fix-the-exception-branch-of-wr_start-is-not-l.patch patch87: 0087-libhns-Fix-out-of-order-issue-of-requester-when-sett.patch patch88: 0088-libhns-Fix-reference-to-uninitialized-cq-pointer.patch +patch89: 0089-libhns-Fix-bypassed-vendor-check-in-hnsdv_query_devi.patch +patch90: 0090-libhns-Fix-coredump-during-QP-destruction-when-send_.patch BuildRequires: binutils cmake >= 2.8.11 gcc libudev-devel pkgconfig pkgconfig(libnl-3.0) BuildRequires: pkgconfig(libnl-route-3.0) valgrind-devel systemd systemd-devel @@ -341,6 +343,12 @@ fi %{_mandir}/* %changelog +* Thu Jan 2 2025 Xinghai Cen <cenxinghai(a)h-partners.com> - 41.0-29 +- Type: bugfix +- ID: NA +- SUG: NA +- DESC: Fixed two bugs in libhns + * Thu Nov 21 2024 Wentao Hu <huwentao19(a)h-partners.com> - 41.0-28 - Type: bugfix - ID: NA -- 2.33.0

1 0

[PATCH v1.1 rdma-core 0/3] Misc patches
by Junxian Huang 27 Dec '24

27 Dec '24

Junxian Huang (3): MAINTAINERS: Update hns maintainers README: Update the name of hns kernel module libhns: Add error logs to help diagnosis MAINTAINERS | 6 +-- README.md | 2 +- providers/hns/hns_roce_u.c | 5 +- providers/hns/hns_roce_u_hw_v2.c | 3 ++ providers/hns/hns_roce_u_verbs.c | 79 ++++++++++++++++++++++++++------ 5 files changed, 74 insertions(+), 21 deletions(-) -- 2.33.0

1 3

[PATCH] libhns: Fix coredump during QP destruction when send_cq == recv_cq
by Chengchang Tang 28 Nov '24

28 Nov '24

From: Xinghai Cen <cenxinghai(a)h-partners.com> driver inclusion category: feature bugzilla: https://gitee.com/src-openeuler/rdma-core/issues/IB7JZL ------------------------------------------------------------------ If the specified send CQ and recv CQ are both the same CQ, the QP node in SCQ is not deleted. which causes a segfault to occur when recreating the QP. Here fix it. Signed-off-by: Xinghai Cen <cenxinghai(a)h-partners.com> --- ...ump-during-QP-destruction-when-send_.patch | 54 +++++++++++++++++++ rdma-core.spec | 9 +++- 2 files changed, 62 insertions(+), 1 deletion(-) create mode 100644 0041-libhns-Fix-coredump-during-QP-destruction-when-send_.patch diff --git a/0041-libhns-Fix-coredump-during-QP-destruction-when-send_.patch b/0041-libhns-Fix-coredump-during-QP-destruction-when-send_.patch new file mode 100644 index 0000000..b3ee84f --- /dev/null +++ b/0041-libhns-Fix-coredump-during-QP-destruction-when-send_.patch @@ -0,0 +1,54 @@ +From 263479c6fb4712528ccae276960ec94fd77afc51 Mon Sep 17 00:00:00 2001 +From: Yuyu Li <liyuyu6(a)huawei.com> +Date: Mon, 25 Nov 2024 16:13:48 +0800 +Subject: [PATCH] libhns: Fix coredump during QP destruction when send_cq == + recv_cq + +driver inclusion +category: feature +bugzilla: https://gitee.com/src-openeuler/rdma-core/issues/IB7JZL + +------------------------------------------------------------------ + +If the specified send CQ and recv CQ are both +the same CQ, the QP node in SCQ is not deleted. +which causes a segfault to occur when recreating +the QP. Here fix it. + +coredump info: +0x0000ffff8fbc37d4 in list_add_before_ +0x0000ffff8fbc381c in list_add_tail_ +0x0000ffff8fbc9d9c in add_qp_to_cq_list +0x0000ffff8fbca008 in create_qp +0x0000ffff8fbca110 in hns_roce_u_create_qp +0x0000ffff8feae39c in __ibv_create_qp_1_1 +0x0000000000401420 in test_ctrl_path + +Fixes: 95e05809d2d2 ("[BigDipperV3R9,NeZha][ROCE] libhns: Support reporting wc as software mode") +Signed-off-by: Yuyu Li <liyuyu6(a)huawei.com> +Signed-off-by: Xinghai Cen <cenxinghai(a)h-partners.com> +--- + providers/hns/hns_roce_u_hw_v2.c | 7 ++++--- + 1 file changed, 4 insertions(+), 3 deletions(-) + +diff --git a/providers/hns/hns_roce_u_hw_v2.c b/providers/hns/hns_roce_u_hw_v2.c +index e4232ea..c746e03 100644 +--- a/providers/hns/hns_roce_u_hw_v2.c ++++ b/providers/hns/hns_roce_u_hw_v2.c +@@ -2006,9 +2006,10 @@ static int hns_roce_u_v2_destroy_qp(struct ibv_qp *ibqp) + list_del(&qp->rcq_node); + } + +- if (ibqp->send_cq && ibqp->send_cq != ibqp->recv_cq) { +- __hns_roce_v2_cq_clean(to_hr_cq(ibqp->send_cq), ibqp->qp_num, +- NULL); ++ if (ibqp->send_cq) { ++ if (ibqp->send_cq != ibqp->recv_cq) ++ __hns_roce_v2_cq_clean(to_hr_cq(ibqp->send_cq), ibqp->qp_num, ++ NULL); + list_del(&qp->scq_node); + } + +-- +2.33.0 + diff --git a/rdma-core.spec b/rdma-core.spec index fb2c277..121f2f5 100644 --- a/rdma-core.spec +++ b/rdma-core.spec @@ -1,6 +1,6 @@ Name: rdma-core Version: 50.0 -Release: 16 +Release: 17 Summary: RDMA core userspace libraries and daemons License: GPL-2.0-only OR BSD-2-Clause AND BSD-3-Clause Url: https://github.com/linux-rdma/rdma-core @@ -46,6 +46,7 @@ patch37: 0037-libhns-Fix-out-of-order-issue-of-requester-when-sett.patch patch38: 0038-libhns-Fix-reference-to-uninitialized-cq-pointer.patch patch39: 0039-libhns-Fix-the-exception-branch-of-wr_start-is-not-l.patch patch40: 0040-libhns-Fix-memory-leakage-when-DCA-is-enabled.patch +patch41: 0041-libhns-Fix-coredump-during-QP-destruction-when-send_.patch BuildRequires: binutils cmake >= 2.8.11 gcc libudev-devel pkgconfig pkgconfig(libnl-3.0) BuildRequires: pkgconfig(libnl-route-3.0) systemd systemd-devel @@ -625,6 +626,12 @@ fi %doc %{_docdir}/%{name}-%{version}/70-persistent-ipoib.rules %changelog +* Thu Nov 28 2024 Xinghai Cen <cenxinghai(a)h-partners.com> - 50.0-17 +- Type: bugfix +- ID: NA +- SUG: NA +- DESC: Fix coredump during QP destruction when send_cq == recv_cq + * Mon Nov 25 2024 Xinghai Cen <cenxinghai(a)h-partners.com> - 50.0-16 - Type: bugfix - ID: NA -- 2.33.0

1 0

[PATCH OLK-6.6] RDMA/hns: Fix bonding failure due to wrong reset_state
by Chengchang Tang 22 Nov '24

22 Nov '24

From: Junxian Huang <huangjunxian6(a)hisilicon.com> driver inclusion category: feature bugzilla: https://gitee.com/openeuler/kernel/issues/IB3K00 ---------------------------------------------------------------------- When roce driver is removed during reset, the reset flow of roce may not be fully completed. This may lead to the reset_state of roce handler stored in nic driver remaining in a middle state, such as HNS_ROCE_STATE_RST_DOWN or HNS_ROCE_STATE_RST_UNINIT. The reset_state won't be cleared even if roce driver is re-inited. This cause that roce bonding which currently relies on reset_state fails in this case. Replace the reset detection for bonding with nic APIs (.ae_dev_resetting() and .get_hw_reset_stat()), just like the reset detection elsewhere in roce driver. Fixes: 26d71e7c13cf ("RDMA/hns: Fix the concurrency error between bond and reset.") Signed-off-by: Junxian Huang <huangjunxian6(a)hisilicon.com> --- drivers/infiniband/hw/hns/hns_roce_bond.c | 9 ++++++--- 1 file changed, 6 insertions(+), 3 deletions(-) diff --git a/drivers/infiniband/hw/hns/hns_roce_bond.c b/drivers/infiniband/hw/hns/hns_roce_bond.c index 7adae8990acd..f76667335189 100644 --- a/drivers/infiniband/hw/hns/hns_roce_bond.c +++ b/drivers/infiniband/hw/hns/hns_roce_bond.c @@ -543,6 +543,7 @@ static void hns_roce_do_bond(struct hns_roce_bond_group *bond_grp) bool is_bond_slave_in_reset(struct hns_roce_bond_group *bond_grp) { + const struct hnae3_ae_ops *ops; struct hnae3_handle *handle; struct net_device *net_dev; int i; @@ -550,9 +551,11 @@ bool is_bond_slave_in_reset(struct hns_roce_bond_group *bond_grp) for (i = 0; i < ROCE_BOND_FUNC_MAX; i++) { net_dev = bond_grp->bond_func_info[i].net_dev; handle = bond_grp->bond_func_info[i].handle; - if (net_dev && handle && - handle->rinfo.reset_state != HNS_ROCE_STATE_NON_RST && - handle->rinfo.reset_state != HNS_ROCE_STATE_RST_INITED) + if (!net_dev || !handle) + continue; + ops = handle->ae_algo->ops; + if (ops->ae_dev_resetting(handle) || + ops->get_hw_reset_stat(handle)) return true; } -- 2.33.0

1 0

[PATCH] libhns: Fixes several bugs for libhns
by Chengchang Tang 22 Nov '24

22 Nov '24

From: Xinghai Cen <cenxinghai(a)h-partners.com> Fixes several bugs for libhns: libhns: Fix memory leakage when DCA is enabled libhns: Fix the exception branch of wr_start() is not locked Signed-off-by: Xinghai Cen <cenxinghai(a)h-partners.com> --- ...xception-branch-of-wr_start-is-not-l.patch | 42 +++++++++++++++++++ ...x-memory-leakage-when-DCA-is-enabled.patch | 41 ++++++++++++++++++ rdma-core.spec | 10 ++++- 3 files changed, 92 insertions(+), 1 deletion(-) create mode 100644 0039-libhns-Fix-the-exception-branch-of-wr_start-is-not-l.patch create mode 100644 0040-libhns-Fix-memory-leakage-when-DCA-is-enabled.patch diff --git a/0039-libhns-Fix-the-exception-branch-of-wr_start-is-not-l.patch b/0039-libhns-Fix-the-exception-branch-of-wr_start-is-not-l.patch new file mode 100644 index 0000000..a5fc588 --- /dev/null +++ b/0039-libhns-Fix-the-exception-branch-of-wr_start-is-not-l.patch @@ -0,0 +1,42 @@ +From 7acbe6f79bbfe32f207800173e5b2b0c13ef23d9 Mon Sep 17 00:00:00 2001 +From: wenglianfa <wenglianfa(a)huawei.com> +Date: Wed, 12 Jun 2024 17:11:13 +0800 +Subject: [PATCH 39/40] libhns: Fix the exception branch of wr_start() is not + locked + +driver inclusion +category: feature +bugzilla: https://gitee.com/src-openeuler/rdma-core/issues/IB66RT + +------------------------------------------------------------------ + +The provider should provide locking to ensure that ibv_wr_start() +and ibv_wr_complete()/abort() form a per-QP critical section +where no other threads can enter. + +The exception branch of wr_start() is not locked, fix it here. +Because check_qp_send () does not require lock protection, +hns_roce_spin_lock () is placed after check_qp_send (). + +Fixes: 36446a56eea5 ("libhns: Extended QP supports the new post send mechanism") +Signed-off-by: wenglianfa <wenglianfa(a)huawei.com> +Signed-off-by: Xinghai Cen <cenxinghai(a)h-partners.com> +--- + providers/hns/hns_roce_u_hw_v2.c | 1 + + 1 file changed, 1 insertion(+) + +diff --git a/providers/hns/hns_roce_u_hw_v2.c b/providers/hns/hns_roce_u_hw_v2.c +index 465ef1e..e4232ea 100644 +--- a/providers/hns/hns_roce_u_hw_v2.c ++++ b/providers/hns/hns_roce_u_hw_v2.c +@@ -2930,6 +2930,7 @@ static void wr_start(struct ibv_qp_ex *ibv_qp) + + ret = check_qp_send(qp, ctx); + if (ret) { ++ hns_roce_spin_lock(&qp->sq.hr_lock); + qp->err = ret; + return; + } +-- +2.33.0 + diff --git a/0040-libhns-Fix-memory-leakage-when-DCA-is-enabled.patch b/0040-libhns-Fix-memory-leakage-when-DCA-is-enabled.patch new file mode 100644 index 0000000..6f5a9c6 --- /dev/null +++ b/0040-libhns-Fix-memory-leakage-when-DCA-is-enabled.patch @@ -0,0 +1,41 @@ +From 1f5569fa17fc65391b504ca8118c3089e598dbc1 Mon Sep 17 00:00:00 2001 +From: wenglianfa <wenglianfa(a)huawei.com> +Date: Thu, 25 Jul 2024 11:06:01 +0800 +Subject: [PATCH 40/40] libhns: Fix memory leakage when DCA is enabled + +driver inclusion +category: feature +bugzilla: https://gitee.com/src-openeuler/rdma-core/issues/IB66RT + +------------------------------------------------------------------ + +After DCA is enabled and a QP is created, the memory block +applied for DCA is not free when the QP is destroyed. Here +fix it. + +Fixes: 41e39ab792c8 ("[BigDipperV3R9,NeZha][ROCE] libhns: Add support for attaching QP's WQE buffer") +Signed-off-by: wenglianfa <wenglianfa(a)huawei.com> +Signed-off-by: Xinghai Cen <cenxinghai(a)h-partners.com> +--- + providers/hns/hns_roce_u_verbs.c | 5 ++++- + 1 file changed, 4 insertions(+), 1 deletion(-) + +diff --git a/providers/hns/hns_roce_u_verbs.c b/providers/hns/hns_roce_u_verbs.c +index e30880c..154e800 100644 +--- a/providers/hns/hns_roce_u_verbs.c ++++ b/providers/hns/hns_roce_u_verbs.c +@@ -1357,7 +1357,10 @@ static void qp_free_wqe(struct hns_roce_qp *qp) + + if (qp->rq.wqe_cnt) + free(qp->rq.wrid); +- hns_roce_free_buf(&qp->buf); ++ if (qp->dca_wqe.bufs) ++ free(qp->dca_wqe.bufs); ++ else ++ hns_roce_free_buf(&qp->buf); + } + + static int qp_alloc_wqe(struct ibv_qp_init_attr_ex *attr, +-- +2.33.0 + diff --git a/rdma-core.spec b/rdma-core.spec index fd956d5..2f4ebe5 100644 --- a/rdma-core.spec +++ b/rdma-core.spec @@ -1,6 +1,6 @@ Name: rdma-core Version: 50.0 -Release: 15 +Release: 16 Summary: RDMA core userspace libraries and daemons License: GPL-2.0-only OR BSD-2-Clause AND BSD-3-Clause Url: https://github.com/linux-rdma/rdma-core @@ -44,6 +44,8 @@ patch35: 0035-Fix-the-stride-calculation-for-MSN-PSN-area.patch patch36: 0036-add-ZTE-Dinghai-rdma-driver.patch patch37: 0037-libhns-Fix-out-of-order-issue-of-requester-when-sett.patch patch38: 0038-libhns-Fix-reference-to-uninitialized-cq-pointer.patch +patch39: 0039-libhns-Fix-the-exception-branch-of-wr_start-is-not-l.patch +patch40: 0040-libhns-Fix-memory-leakage-when-DCA-is-enabled.patch BuildRequires: binutils cmake >= 2.8.11 gcc libudev-devel pkgconfig pkgconfig(libnl-3.0) BuildRequires: pkgconfig(libnl-route-3.0) systemd systemd-devel @@ -623,6 +625,12 @@ fi %doc %{_docdir}/%{name}-%{version}/70-persistent-ipoib.rules %changelog +* Thu Nov 21 2024 Xinghai Cen <cenxinghai(a)h-partners.com> - 50.0-16 +- Type: bugfix +- ID: NA +- SUG: NA +- DESC: Fixes several bugs for libhns + * Fri Nov 15 2024 Xinghai Cen <cenxinghai(a)h-partners.com> - 50.0-15 - Type: bugfix - ID: NA -- 2.33.0

1 0

[PATCH OLK-6.6 00/13] Some patches for RDMA/hns
by Chengchang Tang 20 Nov '24

20 Nov '24

From: Xinghai Cen <cenxinghai(a)h-partners.com> Some patches for RDMA/hns Chengchang Tang (6): RDMA/hns: Fix HW UAF when destroy context timeout RDMA/hns: Fix integer overflow in calc_loading_percent() RDMA/hns: Fix possible RAS when DCA is not attached RDMA/hns: Fix a meaningless loop in active_dca_pages_proc() RDMA/hns: Fix list_*_careful() not being used in pairs RDMA/hns: Use one CQ bank per context Junxian Huang (5): RDMA/hns: Fix wrong output of sysfs scc pram when configuration failed RDMA/hns: Fix concurrency between sysfs store and FW configuration of scc params RDMA/hns: Fix mixed use of u32 and __le32 in sysfs RDMA/hns: Fix dereference of noderef expression RDMA/hns: Fix "Should it be static?" warnings wenglianfa (2): RDMA/hns: Fix the modification of max_send_sge RDMA/hns: Fix RoCEE hang when multiple QP banks use EXT_SGE EXT_SGE drivers/infiniband/hw/hns/hns_roce_bond.c | 8 ++- drivers/infiniband/hw/hns/hns_roce_cq.c | 73 +++++++++++++++++-- drivers/infiniband/hw/hns/hns_roce_dca.c | 60 ++++++++++++++-- drivers/infiniband/hw/hns/hns_roce_dca.h | 2 + drivers/infiniband/hw/hns/hns_roce_debugfs.c | 10 ++- drivers/infiniband/hw/hns/hns_roce_device.h | 17 ++++- drivers/infiniband/hw/hns/hns_roce_hw_v2.c | 26 +++++-- drivers/infiniband/hw/hns/hns_roce_main.c | 20 ++++++ drivers/infiniband/hw/hns/hns_roce_mr.c | 6 +- drivers/infiniband/hw/hns/hns_roce_qp.c | 74 ++++++++++++++++---- drivers/infiniband/hw/hns/hns_roce_srq.c | 6 +- drivers/infiniband/hw/hns/hns_roce_sysfs.c | 41 +++++++---- 12 files changed, 288 insertions(+), 55 deletions(-) -- 2.33.0

1 13

[PATCH OLK-6.6] RDMA/hns: Fix different dgids mapping to the same dip_idx
by Chengchang Tang 20 Nov '24

20 Nov '24

From: Feng Fang <fangfeng4(a)huawei.com> maillist inclusion commit faa62440a5772b40bb7d78bf9e29556a82ecf153 category: bugfix bugzilla: https://gitee.com/openeuler/kernel/issues/IB4OOG Reference: https://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma.git/commit/?id=fa… --------------------------------------------------------------------- DIP algorithm requires a one-to-one mapping between dgid and dip_idx. Currently a queue 'spare_idx' is used to store QPN of QPs that use DIP algorithm. For a new dgid, use a QPN from spare_idx as dip_idx. This method lacks a mechanism for deduplicating QPN, which may result in different dgids sharing the same dip_idx and break the one-to-one mapping requirement. This patch replaces spare_idx with xarray and introduces a refcnt of a dip_idx to indicate the number of QPs that using this dip_idx. The state machine for dip_idx management is implemented as: * The entry at an index in xarray is empty -- This indicates that the corresponding dip_idx hasn't been created. * The entry at an index in xarray is not empty but with 0 refcnt -- This indicates that the corresponding dip_idx has been created but not used as dip_idx yet. * The entry at an index in xarray is not empty and with non-0 refcnt -- This indicates that the corresponding dip_idx is being used by refcnt number of DIP QPs. Fixes: eb653eda1e91 ("RDMA/hns: Bugfix for incorrect association between dip_idx and dgid") Fixes: f91696f2f053 ("RDMA/hns: Support congestion control type selection according to the FW") Signed-off-by: Feng Fang <fangfeng4(a)huawei.com> Signed-off-by: Junxian Huang <huangjunxian6(a)hisilicon.com> Link: https://patch.msgid.link/20241112055553.3681129-1-huangjunxian6@hisilicon.c… Signed-off-by: Leon Romanovsky <leon(a)kernel.org> Signed-off-by: Xinghai Cen <cenxinghai(a)h-partners.com> --- drivers/infiniband/hw/hns/hns_roce_device.h | 11 +-- drivers/infiniband/hw/hns/hns_roce_hw_v2.c | 96 +++++++++++++++------ drivers/infiniband/hw/hns/hns_roce_hw_v2.h | 2 +- drivers/infiniband/hw/hns/hns_roce_main.c | 2 - drivers/infiniband/hw/hns/hns_roce_qp.c | 8 +- 5 files changed, 75 insertions(+), 44 deletions(-) diff --git a/drivers/infiniband/hw/hns/hns_roce_device.h b/drivers/infiniband/hw/hns/hns_roce_device.h index fdc1fe5e6a81..e3303cc3584a 100644 --- a/drivers/infiniband/hw/hns/hns_roce_device.h +++ b/drivers/infiniband/hw/hns/hns_roce_device.h @@ -570,12 +570,6 @@ struct hns_roce_bank { u32 next; /* Next ID to allocate. */ }; -struct hns_roce_idx_table { - u32 *spare_idx; - u32 head; - u32 tail; -}; - struct hns_roce_qp_table { struct hns_roce_hem_table qp_table; struct hns_roce_hem_table irrl_table; @@ -584,7 +578,7 @@ struct hns_roce_qp_table { struct mutex scc_mutex; struct hns_roce_bank bank[HNS_ROCE_QP_BANK_NUM]; struct mutex bank_mutex; - struct hns_roce_idx_table idx_table; + struct xarray dip_xa; }; struct hns_roce_cq_table { @@ -742,6 +736,7 @@ struct hns_roce_qp { bool delayed_destroy_flag; struct hns_roce_mtr_node *mtr_node; spinlock_t flush_lock; + struct hns_roce_dip *dip; }; struct hns_roce_ib_iboe { @@ -1102,8 +1097,6 @@ struct hns_roce_dev { enum hns_roce_device_state state; struct list_head qp_list; /* list of all qps on this dev */ spinlock_t qp_list_lock; /* protect qp_list */ - struct list_head dip_list; /* list of all dest ips on this dev */ - spinlock_t dip_list_lock; /* protect dip_list */ struct list_head pgdir_list; struct mutex pgdir_mutex; diff --git a/drivers/infiniband/hw/hns/hns_roce_hw_v2.c b/drivers/infiniband/hw/hns/hns_roce_hw_v2.c index 3f254ac48b42..88bd75c5743e 100644 --- a/drivers/infiniband/hw/hns/hns_roce_hw_v2.c +++ b/drivers/infiniband/hw/hns/hns_roce_hw_v2.c @@ -2720,20 +2720,19 @@ static void hns_roce_free_link_table(struct hns_roce_dev *hr_dev) free_link_table_buf(hr_dev, &priv->ext_llm); } -static void free_dip_list(struct hns_roce_dev *hr_dev) +static void free_dip_entry(struct hns_roce_dev *hr_dev) { struct hns_roce_dip *hr_dip; - struct hns_roce_dip *tmp; - unsigned long flags; + unsigned long idx; - spin_lock_irqsave(&hr_dev->dip_list_lock, flags); + xa_lock(&hr_dev->qp_table.dip_xa); - list_for_each_entry_safe(hr_dip, tmp, &hr_dev->dip_list, node) { - list_del(&hr_dip->node); + xa_for_each(&hr_dev->qp_table.dip_xa, idx, hr_dip) { + __xa_erase(&hr_dev->qp_table.dip_xa, hr_dip->dip_idx); kfree(hr_dip); } - spin_unlock_irqrestore(&hr_dev->dip_list_lock, flags); + xa_unlock(&hr_dev->qp_table.dip_xa); } static int hns_roce_v2_get_reset_page(struct hns_roce_dev *hr_dev) @@ -3182,7 +3181,7 @@ static void hns_roce_v2_exit(struct hns_roce_dev *hr_dev) hns_roce_v2_put_reset_page(hr_dev); if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP09) - free_dip_list(hr_dev); + free_dip_entry(hr_dev); } static inline void mbox_desc_init(struct hns_roce_post_mbox *mb, @@ -5042,26 +5041,49 @@ static int modify_qp_rtr_to_rts(struct ib_qp *ibqp, int attr_mask, return 0; } +static int alloc_dip_entry(struct xarray *dip_xa, u32 qpn) +{ + struct hns_roce_dip *hr_dip; + int ret; + + hr_dip = xa_load(dip_xa, qpn); + if (hr_dip) + return 0; + + hr_dip = kzalloc(sizeof(*hr_dip), GFP_KERNEL); + if (!hr_dip) + return -ENOMEM; + + ret = xa_err(xa_store(dip_xa, qpn, hr_dip, GFP_KERNEL)); + if (ret) + kfree(hr_dip); + + return ret; +} + static int get_dip_ctx_idx(struct ib_qp *ibqp, const struct ib_qp_attr *attr, u32 *dip_idx) { const struct ib_global_route *grh = rdma_ah_read_grh(&attr->ah_attr); struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device); - u32 *spare_idx = hr_dev->qp_table.idx_table.spare_idx; - u32 *head = &hr_dev->qp_table.idx_table.head; - u32 *tail = &hr_dev->qp_table.idx_table.tail; + struct xarray *dip_xa = &hr_dev->qp_table.dip_xa; + struct hns_roce_qp *hr_qp = to_hr_qp(ibqp); struct hns_roce_dip *hr_dip; - unsigned long flags; + unsigned long idx; int ret = 0; - spin_lock_irqsave(&hr_dev->dip_list_lock, flags); + ret = alloc_dip_entry(dip_xa, ibqp->qp_num); + if (ret) + return ret; - spare_idx[*tail] = ibqp->qp_num; - *tail = (*tail == hr_dev->caps.num_qps - 1) ? 0 : (*tail + 1); + xa_lock(dip_xa); - list_for_each_entry(hr_dip, &hr_dev->dip_list, node) { - if (!memcmp(grh->dgid.raw, hr_dip->dgid, GID_LEN_V2)) { + xa_for_each(dip_xa, idx, hr_dip) { + if (hr_dip->qp_cnt && + !memcmp(grh->dgid.raw, hr_dip->dgid, GID_LEN_V2)) { *dip_idx = hr_dip->dip_idx; + hr_dip->qp_cnt++; + hr_qp->dip = hr_dip; goto out; } } @@ -5069,19 +5091,24 @@ static int get_dip_ctx_idx(struct ib_qp *ibqp, const struct ib_qp_attr *attr, /* If no dgid is found, a new dip and a mapping between dgid and * dip_idx will be created. */ - hr_dip = kzalloc(sizeof(*hr_dip), GFP_ATOMIC); - if (!hr_dip) { - ret = -ENOMEM; - goto out; + xa_for_each(dip_xa, idx, hr_dip) { + if (hr_dip->qp_cnt) + continue; + + *dip_idx = idx; + memcpy(hr_dip->dgid, grh->dgid.raw, sizeof(grh->dgid.raw)); + hr_dip->dip_idx = idx; + hr_dip->qp_cnt++; + hr_qp->dip = hr_dip; + break; } - memcpy(hr_dip->dgid, grh->dgid.raw, sizeof(grh->dgid.raw)); - hr_dip->dip_idx = *dip_idx = spare_idx[*head]; - *head = (*head == hr_dev->caps.num_qps - 1) ? 0 : (*head + 1); - list_add_tail(&hr_dip->node, &hr_dev->dip_list); + /* This should never happen. */ + if (WARN_ON_ONCE(!hr_qp->dip)) + ret = -ENOSPC; out: - spin_unlock_irqrestore(&hr_dev->dip_list_lock, flags); + xa_unlock(dip_xa); return ret; } @@ -6005,6 +6032,20 @@ static int hns_roce_v2_destroy_qp_common(struct hns_roce_dev *hr_dev, return ret; } +static void put_dip_ctx_idx(struct hns_roce_dev *hr_dev, + struct hns_roce_qp *hr_qp) +{ + struct hns_roce_dip *hr_dip = hr_qp->dip; + + xa_lock(&hr_dev->qp_table.dip_xa); + + hr_dip->qp_cnt--; + if (!hr_dip->qp_cnt) + memset(hr_dip->dgid, 0, GID_LEN_V2); + + xa_unlock(&hr_dev->qp_table.dip_xa); +} + int hns_roce_v2_destroy_qp(struct ib_qp *ibqp, struct ib_udata *udata) { struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device); @@ -6018,6 +6059,9 @@ int hns_roce_v2_destroy_qp(struct ib_qp *ibqp, struct ib_udata *udata) spin_unlock_irqrestore(&hr_qp->flush_lock, flags); flush_work(&hr_qp->flush_work.work); + if (hr_qp->cong_type == CONG_TYPE_DIP) + put_dip_ctx_idx(hr_dev, hr_qp); + ret = hns_roce_v2_destroy_qp_common(hr_dev, hr_qp, udata); if (ret) ibdev_err_ratelimited(&hr_dev->ib_dev, diff --git a/drivers/infiniband/hw/hns/hns_roce_hw_v2.h b/drivers/infiniband/hw/hns/hns_roce_hw_v2.h index 201902dff611..f43444be0308 100644 --- a/drivers/infiniband/hw/hns/hns_roce_hw_v2.h +++ b/drivers/infiniband/hw/hns/hns_roce_hw_v2.h @@ -1366,7 +1366,7 @@ struct hns_roce_v2_priv { struct hns_roce_dip { u8 dgid[GID_LEN_V2]; u32 dip_idx; - struct list_head node; /* all dips are on a list */ + u32 qp_cnt; }; struct fmea_ram_ecc { diff --git a/drivers/infiniband/hw/hns/hns_roce_main.c b/drivers/infiniband/hw/hns/hns_roce_main.c index d488c3d5986f..520112bd43d5 100644 --- a/drivers/infiniband/hw/hns/hns_roce_main.c +++ b/drivers/infiniband/hw/hns/hns_roce_main.c @@ -1250,8 +1250,6 @@ static int hns_roce_setup_hca(struct hns_roce_dev *hr_dev) INIT_LIST_HEAD(&hr_dev->qp_list); spin_lock_init(&hr_dev->qp_list_lock); - INIT_LIST_HEAD(&hr_dev->dip_list); - spin_lock_init(&hr_dev->dip_list_lock); INIT_LIST_HEAD(&hr_dev->uctx_list); mutex_init(&hr_dev->uctx_list_mutex); diff --git a/drivers/infiniband/hw/hns/hns_roce_qp.c b/drivers/infiniband/hw/hns/hns_roce_qp.c index 5ed2647567aa..77ec0c8678d3 100644 --- a/drivers/infiniband/hw/hns/hns_roce_qp.c +++ b/drivers/infiniband/hw/hns/hns_roce_qp.c @@ -1701,14 +1701,10 @@ int hns_roce_init_qp_table(struct hns_roce_dev *hr_dev) unsigned int reserved_from_bot; unsigned int i; - qp_table->idx_table.spare_idx = kcalloc(hr_dev->caps.num_qps, - sizeof(u32), GFP_KERNEL); - if (!qp_table->idx_table.spare_idx) - return -ENOMEM; - mutex_init(&qp_table->scc_mutex); mutex_init(&qp_table->bank_mutex); xa_init(&hr_dev->qp_table_xa); + xa_init(&qp_table->dip_xa); reserved_from_bot = hr_dev->caps.reserved_qps; @@ -1733,7 +1729,7 @@ void hns_roce_cleanup_qp_table(struct hns_roce_dev *hr_dev) for (i = 0; i < HNS_ROCE_QP_BANK_NUM; i++) ida_destroy(&hr_dev->qp_table.bank[i].ida); + xa_destroy(&hr_dev->qp_table.dip_xa); mutex_destroy(&hr_dev->qp_table.bank_mutex); mutex_destroy(&hr_dev->qp_table.scc_mutex); - kfree(hr_dev->qp_table.idx_table.spare_idx); } -- 2.33.0

1 0

[PATCH OLK-6.6] RDMA/hns: Fix different dgids mapping to the same dip_idx
by Chengchang Tang 19 Nov '24

19 Nov '24

From: Feng Fang <fangfeng4(a)huawei.com> mainline inclusion from mainline-v6.12-rc2 commit faa62440a5772b40bb7d78bf9e29556a82ecf153 category: bugfix bugzilla: https://gitee.com/openeuler/kernel/issues/IB4OOG CVE: NA Reference: https://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma.git/commit/?id=fa… --------------------------------------------------------------------- DIP algorithm requires a one-to-one mapping between dgid and dip_idx. Currently a queue 'spare_idx' is used to store QPN of QPs that use DIP algorithm. For a new dgid, use a QPN from spare_idx as dip_idx. This method lacks a mechanism for deduplicating QPN, which may result in different dgids sharing the same dip_idx and break the one-to-one mapping requirement. This patch replaces spare_idx with xarray and introduces a refcnt of a dip_idx to indicate the number of QPs that using this dip_idx. The state machine for dip_idx management is implemented as: * The entry at an index in xarray is empty -- This indicates that the corresponding dip_idx hasn't been created. * The entry at an index in xarray is not empty but with 0 refcnt -- This indicates that the corresponding dip_idx has been created but not used as dip_idx yet. * The entry at an index in xarray is not empty and with non-0 refcnt -- This indicates that the corresponding dip_idx is being used by refcnt number of DIP QPs. Fixes: eb653eda1e91 ("RDMA/hns: Bugfix for incorrect association between dip_idx and dgid") Fixes: f91696f2f053 ("RDMA/hns: Support congestion control type selection according to the FW") Signed-off-by: Feng Fang <fangfeng4(a)huawei.com> Signed-off-by: Junxian Huang <huangjunxian6(a)hisilicon.com> Link: https://patch.msgid.link/20241112055553.3681129-1-huangjunxian6@hisilicon.c… Signed-off-by: Leon Romanovsky <leon(a)kernel.org> --- drivers/infiniband/hw/hns/hns_roce_device.h | 11 +-- drivers/infiniband/hw/hns/hns_roce_hw_v2.c | 96 +++++++++++++++------ drivers/infiniband/hw/hns/hns_roce_hw_v2.h | 2 +- drivers/infiniband/hw/hns/hns_roce_main.c | 2 - drivers/infiniband/hw/hns/hns_roce_qp.c | 8 +- 5 files changed, 75 insertions(+), 44 deletions(-) diff --git a/drivers/infiniband/hw/hns/hns_roce_device.h b/drivers/infiniband/hw/hns/hns_roce_device.h index fdc1fe5e6a81..e3303cc3584a 100644 --- a/drivers/infiniband/hw/hns/hns_roce_device.h +++ b/drivers/infiniband/hw/hns/hns_roce_device.h @@ -570,12 +570,6 @@ struct hns_roce_bank { u32 next; /* Next ID to allocate. */ }; -struct hns_roce_idx_table { - u32 *spare_idx; - u32 head; - u32 tail; -}; - struct hns_roce_qp_table { struct hns_roce_hem_table qp_table; struct hns_roce_hem_table irrl_table; @@ -584,7 +578,7 @@ struct hns_roce_qp_table { struct mutex scc_mutex; struct hns_roce_bank bank[HNS_ROCE_QP_BANK_NUM]; struct mutex bank_mutex; - struct hns_roce_idx_table idx_table; + struct xarray dip_xa; }; struct hns_roce_cq_table { @@ -742,6 +736,7 @@ struct hns_roce_qp { bool delayed_destroy_flag; struct hns_roce_mtr_node *mtr_node; spinlock_t flush_lock; + struct hns_roce_dip *dip; }; struct hns_roce_ib_iboe { @@ -1102,8 +1097,6 @@ struct hns_roce_dev { enum hns_roce_device_state state; struct list_head qp_list; /* list of all qps on this dev */ spinlock_t qp_list_lock; /* protect qp_list */ - struct list_head dip_list; /* list of all dest ips on this dev */ - spinlock_t dip_list_lock; /* protect dip_list */ struct list_head pgdir_list; struct mutex pgdir_mutex; diff --git a/drivers/infiniband/hw/hns/hns_roce_hw_v2.c b/drivers/infiniband/hw/hns/hns_roce_hw_v2.c index 3f254ac48b42..88bd75c5743e 100644 --- a/drivers/infiniband/hw/hns/hns_roce_hw_v2.c +++ b/drivers/infiniband/hw/hns/hns_roce_hw_v2.c @@ -2720,20 +2720,19 @@ static void hns_roce_free_link_table(struct hns_roce_dev *hr_dev) free_link_table_buf(hr_dev, &priv->ext_llm); } -static void free_dip_list(struct hns_roce_dev *hr_dev) +static void free_dip_entry(struct hns_roce_dev *hr_dev) { struct hns_roce_dip *hr_dip; - struct hns_roce_dip *tmp; - unsigned long flags; + unsigned long idx; - spin_lock_irqsave(&hr_dev->dip_list_lock, flags); + xa_lock(&hr_dev->qp_table.dip_xa); - list_for_each_entry_safe(hr_dip, tmp, &hr_dev->dip_list, node) { - list_del(&hr_dip->node); + xa_for_each(&hr_dev->qp_table.dip_xa, idx, hr_dip) { + __xa_erase(&hr_dev->qp_table.dip_xa, hr_dip->dip_idx); kfree(hr_dip); } - spin_unlock_irqrestore(&hr_dev->dip_list_lock, flags); + xa_unlock(&hr_dev->qp_table.dip_xa); } static int hns_roce_v2_get_reset_page(struct hns_roce_dev *hr_dev) @@ -3182,7 +3181,7 @@ static void hns_roce_v2_exit(struct hns_roce_dev *hr_dev) hns_roce_v2_put_reset_page(hr_dev); if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP09) - free_dip_list(hr_dev); + free_dip_entry(hr_dev); } static inline void mbox_desc_init(struct hns_roce_post_mbox *mb, @@ -5042,26 +5041,49 @@ static int modify_qp_rtr_to_rts(struct ib_qp *ibqp, int attr_mask, return 0; } +static int alloc_dip_entry(struct xarray *dip_xa, u32 qpn) +{ + struct hns_roce_dip *hr_dip; + int ret; + + hr_dip = xa_load(dip_xa, qpn); + if (hr_dip) + return 0; + + hr_dip = kzalloc(sizeof(*hr_dip), GFP_KERNEL); + if (!hr_dip) + return -ENOMEM; + + ret = xa_err(xa_store(dip_xa, qpn, hr_dip, GFP_KERNEL)); + if (ret) + kfree(hr_dip); + + return ret; +} + static int get_dip_ctx_idx(struct ib_qp *ibqp, const struct ib_qp_attr *attr, u32 *dip_idx) { const struct ib_global_route *grh = rdma_ah_read_grh(&attr->ah_attr); struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device); - u32 *spare_idx = hr_dev->qp_table.idx_table.spare_idx; - u32 *head = &hr_dev->qp_table.idx_table.head; - u32 *tail = &hr_dev->qp_table.idx_table.tail; + struct xarray *dip_xa = &hr_dev->qp_table.dip_xa; + struct hns_roce_qp *hr_qp = to_hr_qp(ibqp); struct hns_roce_dip *hr_dip; - unsigned long flags; + unsigned long idx; int ret = 0; - spin_lock_irqsave(&hr_dev->dip_list_lock, flags); + ret = alloc_dip_entry(dip_xa, ibqp->qp_num); + if (ret) + return ret; - spare_idx[*tail] = ibqp->qp_num; - *tail = (*tail == hr_dev->caps.num_qps - 1) ? 0 : (*tail + 1); + xa_lock(dip_xa); - list_for_each_entry(hr_dip, &hr_dev->dip_list, node) { - if (!memcmp(grh->dgid.raw, hr_dip->dgid, GID_LEN_V2)) { + xa_for_each(dip_xa, idx, hr_dip) { + if (hr_dip->qp_cnt && + !memcmp(grh->dgid.raw, hr_dip->dgid, GID_LEN_V2)) { *dip_idx = hr_dip->dip_idx; + hr_dip->qp_cnt++; + hr_qp->dip = hr_dip; goto out; } } @@ -5069,19 +5091,24 @@ static int get_dip_ctx_idx(struct ib_qp *ibqp, const struct ib_qp_attr *attr, /* If no dgid is found, a new dip and a mapping between dgid and * dip_idx will be created. */ - hr_dip = kzalloc(sizeof(*hr_dip), GFP_ATOMIC); - if (!hr_dip) { - ret = -ENOMEM; - goto out; + xa_for_each(dip_xa, idx, hr_dip) { + if (hr_dip->qp_cnt) + continue; + + *dip_idx = idx; + memcpy(hr_dip->dgid, grh->dgid.raw, sizeof(grh->dgid.raw)); + hr_dip->dip_idx = idx; + hr_dip->qp_cnt++; + hr_qp->dip = hr_dip; + break; } - memcpy(hr_dip->dgid, grh->dgid.raw, sizeof(grh->dgid.raw)); - hr_dip->dip_idx = *dip_idx = spare_idx[*head]; - *head = (*head == hr_dev->caps.num_qps - 1) ? 0 : (*head + 1); - list_add_tail(&hr_dip->node, &hr_dev->dip_list); + /* This should never happen. */ + if (WARN_ON_ONCE(!hr_qp->dip)) + ret = -ENOSPC; out: - spin_unlock_irqrestore(&hr_dev->dip_list_lock, flags); + xa_unlock(dip_xa); return ret; } @@ -6005,6 +6032,20 @@ static int hns_roce_v2_destroy_qp_common(struct hns_roce_dev *hr_dev, return ret; } +static void put_dip_ctx_idx(struct hns_roce_dev *hr_dev, + struct hns_roce_qp *hr_qp) +{ + struct hns_roce_dip *hr_dip = hr_qp->dip; + + xa_lock(&hr_dev->qp_table.dip_xa); + + hr_dip->qp_cnt--; + if (!hr_dip->qp_cnt) + memset(hr_dip->dgid, 0, GID_LEN_V2); + + xa_unlock(&hr_dev->qp_table.dip_xa); +} + int hns_roce_v2_destroy_qp(struct ib_qp *ibqp, struct ib_udata *udata) { struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device); @@ -6018,6 +6059,9 @@ int hns_roce_v2_destroy_qp(struct ib_qp *ibqp, struct ib_udata *udata) spin_unlock_irqrestore(&hr_qp->flush_lock, flags); flush_work(&hr_qp->flush_work.work); + if (hr_qp->cong_type == CONG_TYPE_DIP) + put_dip_ctx_idx(hr_dev, hr_qp); + ret = hns_roce_v2_destroy_qp_common(hr_dev, hr_qp, udata); if (ret) ibdev_err_ratelimited(&hr_dev->ib_dev, diff --git a/drivers/infiniband/hw/hns/hns_roce_hw_v2.h b/drivers/infiniband/hw/hns/hns_roce_hw_v2.h index 201902dff611..f43444be0308 100644 --- a/drivers/infiniband/hw/hns/hns_roce_hw_v2.h +++ b/drivers/infiniband/hw/hns/hns_roce_hw_v2.h @@ -1366,7 +1366,7 @@ struct hns_roce_v2_priv { struct hns_roce_dip { u8 dgid[GID_LEN_V2]; u32 dip_idx; - struct list_head node; /* all dips are on a list */ + u32 qp_cnt; }; struct fmea_ram_ecc { diff --git a/drivers/infiniband/hw/hns/hns_roce_main.c b/drivers/infiniband/hw/hns/hns_roce_main.c index d488c3d5986f..520112bd43d5 100644 --- a/drivers/infiniband/hw/hns/hns_roce_main.c +++ b/drivers/infiniband/hw/hns/hns_roce_main.c @@ -1250,8 +1250,6 @@ static int hns_roce_setup_hca(struct hns_roce_dev *hr_dev) INIT_LIST_HEAD(&hr_dev->qp_list); spin_lock_init(&hr_dev->qp_list_lock); - INIT_LIST_HEAD(&hr_dev->dip_list); - spin_lock_init(&hr_dev->dip_list_lock); INIT_LIST_HEAD(&hr_dev->uctx_list); mutex_init(&hr_dev->uctx_list_mutex); diff --git a/drivers/infiniband/hw/hns/hns_roce_qp.c b/drivers/infiniband/hw/hns/hns_roce_qp.c index 5ed2647567aa..77ec0c8678d3 100644 --- a/drivers/infiniband/hw/hns/hns_roce_qp.c +++ b/drivers/infiniband/hw/hns/hns_roce_qp.c @@ -1701,14 +1701,10 @@ int hns_roce_init_qp_table(struct hns_roce_dev *hr_dev) unsigned int reserved_from_bot; unsigned int i; - qp_table->idx_table.spare_idx = kcalloc(hr_dev->caps.num_qps, - sizeof(u32), GFP_KERNEL); - if (!qp_table->idx_table.spare_idx) - return -ENOMEM; - mutex_init(&qp_table->scc_mutex); mutex_init(&qp_table->bank_mutex); xa_init(&hr_dev->qp_table_xa); + xa_init(&qp_table->dip_xa); reserved_from_bot = hr_dev->caps.reserved_qps; @@ -1733,7 +1729,7 @@ void hns_roce_cleanup_qp_table(struct hns_roce_dev *hr_dev) for (i = 0; i < HNS_ROCE_QP_BANK_NUM; i++) ida_destroy(&hr_dev->qp_table.bank[i].ida); + xa_destroy(&hr_dev->qp_table.dip_xa); mutex_destroy(&hr_dev->qp_table.bank_mutex); mutex_destroy(&hr_dev->qp_table.scc_mutex); - kfree(hr_dev->qp_table.idx_table.spare_idx); } -- 2.33.0

1 0

高性能网络sig会议
by openEuler conference 18 Nov '24

18 Nov '24

您好！ sig-high-performance-network 邀请您参加 2024-11-18 14:30 召开的WeLink会议(自动录制) 会议主题：高性能网络sig会议会议链接：https://meeting.huaweicloud.com:36443/#/j/982559975 会议纪要：https://etherpad.openeuler.org/p/sig-high-performance-network-meetings 更多资讯尽在：https://www.openeuler.org/zh/ Hello! sig-high-performance-network invites you to attend the WeLink conference(auto recording) will be held at 2024-11-18 14:30, The subject of the conference is 高性能网络sig会议, You can join the meeting at https://meeting.huaweicloud.com:36443/#/j/982559975. Add topics at https://etherpad.openeuler.org/p/sig-high-performance-network-meetings. More information: https://www.openeuler.org/en/

1 0

[PATCH] libhns: Two bugfixes in post_send flow
by Chengchang Tang 15 Nov '24

15 Nov '24

From: Xinghai Cen <cenxinghai(a)h-partners.com> Two bugfixes in post_send flow: libhns: Fix out-of-order issue of requester when setting FENCE libhns: Fix reference to uninitialized cq pointer Signed-off-by: Xinghai Cen <cenxinghai(a)h-partners.com> --- ...f-order-issue-of-requester-when-sett.patch | 50 ++++++++++++++ ...eference-to-uninitialized-cq-pointer.patch | 68 +++++++++++++++++++ rdma-core.spec | 10 ++- 3 files changed, 127 insertions(+), 1 deletion(-) create mode 100644 0037-libhns-Fix-out-of-order-issue-of-requester-when-sett.patch create mode 100644 0038-libhns-Fix-reference-to-uninitialized-cq-pointer.patch diff --git a/0037-libhns-Fix-out-of-order-issue-of-requester-when-sett.patch b/0037-libhns-Fix-out-of-order-issue-of-requester-when-sett.patch new file mode 100644 index 0000000..0996c7f --- /dev/null +++ b/0037-libhns-Fix-out-of-order-issue-of-requester-when-sett.patch @@ -0,0 +1,50 @@ +From 23c00575e9dc2b61f2ac429f9d039cb7e2e6deb2 Mon Sep 17 00:00:00 2001 +From: Junxian Huang <huangjunxian6(a)hisilicon.com> +Date: Fri, 8 Nov 2024 17:04:09 +0800 +Subject: [PATCH] libhns: Fix out-of-order issue of requester when setting + FENCE + +mainline inclusion +from mainline-master +commit c4119911c212aaa552c9cb928fba0a696640c9b5 +category: bugfix +bugzilla: https://gitee.com/openeuler/kernel/issues/IB3ZHQ +CVE: NA +Reference: https://github.com/linux-rdma/rdma-core/pull/1513/commits/c4119911c212aaa55… + +---------------------------------------------------------------------- + +The FENCE indicator in hns WQE doesn't ensure that response data from +a previous Read/Atomic operation has been written to the requester's +memory before the subsequent Send/Write operation is processed. This +may result in the subsequent Send/Write operation accessing the original +data in memory instead of the expected response data. + +Unlike FENCE, the SO (Strong Order) indicator blocks the subsequent +operation until the previous response data is written to memory and a +bresp is returned. Set the SO indicator instead of FENCE to maintain +strict order. + +Fixes: cbdf5e32a855 ("libhns: Reimplement verbs of post_send and post_recv for hip08 RoCE") +Signed-off-by: Junxian Huang <huangjunxian6(a)hisilicon.com> +Signed-off-by: Xinghai Cen <cenxinghai(a)h-partners.com> +--- + providers/hns/hns_roce_u_hw_v2.c | 2 +- + 1 file changed, 1 insertion(+), 1 deletion(-) + +diff --git a/providers/hns/hns_roce_u_hw_v2.c b/providers/hns/hns_roce_u_hw_v2.c +index 9371150..2debcb3 100644 +--- a/providers/hns/hns_roce_u_hw_v2.c ++++ b/providers/hns/hns_roce_u_hw_v2.c +@@ -1527,7 +1527,7 @@ static int set_rc_wqe(void *wqe, struct hns_roce_qp *qp, struct ibv_send_wr *wr, + + hr_reg_write_bool(wqe, RCWQE_CQE, + !!(wr->send_flags & IBV_SEND_SIGNALED)); +- hr_reg_write_bool(wqe, RCWQE_FENCE, ++ hr_reg_write_bool(wqe, RCWQE_SO, + !!(wr->send_flags & IBV_SEND_FENCE)); + hr_reg_write_bool(wqe, RCWQE_SE, + !!(wr->send_flags & IBV_SEND_SOLICITED)); +-- +2.33.0 + diff --git a/0038-libhns-Fix-reference-to-uninitialized-cq-pointer.patch b/0038-libhns-Fix-reference-to-uninitialized-cq-pointer.patch new file mode 100644 index 0000000..e45d61a --- /dev/null +++ b/0038-libhns-Fix-reference-to-uninitialized-cq-pointer.patch @@ -0,0 +1,68 @@ +From e941cc407d89595d1affd5211e1daf34786c7641 Mon Sep 17 00:00:00 2001 +From: Chengchang Tang <tangchengchang(a)huawei.com> +Date: Fri, 8 Nov 2024 17:04:08 +0800 +Subject: [PATCH] libhns: Fix reference to uninitialized cq pointer +MIME-Version: 1.0 +Content-Type: text/plain; charset=UTF-8 +Content-Transfer-Encoding: 8bit + +mainline inclusion +from mainline-master +commit 18e3117cdd161a3f40b8a917f24cfb5227a1d75a +category: bugfix +bugzilla: https://gitee.com/openeuler/kernel/issues/IB3ZHQ +CVE: NA +Reference: https://github.com/linux-rdma/rdma-core/pull/1513/commits/18e3117cdd161a3f4… + +---------------------------------------------------------------------- + +For QPs which do not have an SQ, such as XRC TGT，the send_cq +pointer will not be initailized. Since the supported max_gs +will be 0 in this case, check it and return before referencing +the send_cq pointer. + +Fixes: cbdf5e32a855 ("libhns: Reimplement verbs of post_send and post_recv for hip08 RoCE") +Signed-off-by: Chengchang Tang <tangchengchang(a)huawei.com> +Signed-off-by: Junxian Huang <huangjunxian6(a)hisilicon.com> +Signed-off-by: Xinghai Cen <cenxinghai(a)h-partners.com> +--- + providers/hns/hns_roce_u_hw_v2.c | 12 ++++++------ + 1 file changed, 6 insertions(+), 6 deletions(-) + +diff --git a/providers/hns/hns_roce_u_hw_v2.c b/providers/hns/hns_roce_u_hw_v2.c +index 2debcb3..465ef1e 100644 +--- a/providers/hns/hns_roce_u_hw_v2.c ++++ b/providers/hns/hns_roce_u_hw_v2.c +@@ -1579,7 +1579,7 @@ int hns_roce_u_v2_post_send(struct ibv_qp *ibvqp, struct ibv_send_wr *wr, + struct hns_roce_context *ctx = to_hr_ctx(ibvqp->context); + struct hns_roce_qp *qp = to_hr_qp(ibvqp); + struct hns_roce_sge_info sge_info = {}; +- struct hns_roce_rc_sq_wqe *wqe; ++ struct hns_roce_rc_sq_wqe *wqe = NULL; + struct ibv_qp_attr attr = {}; + unsigned int wqe_idx, nreq; + int ret; +@@ -1595,15 +1595,15 @@ int hns_roce_u_v2_post_send(struct ibv_qp *ibvqp, struct ibv_send_wr *wr, + 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->verbs_qp.qp.send_cq))) { +- ret = ENOMEM; ++ if (wr->num_sge > (int)qp->sq.max_gs) { ++ ret = qp->sq.max_gs > 0 ? EINVAL : EOPNOTSUPP; + *bad_wr = wr; + goto out; + } + +- if (wr->num_sge > qp->sq.max_gs) { +- ret = EINVAL; ++ if (hns_roce_v2_wq_overflow(&qp->sq, nreq, ++ to_hr_cq(qp->verbs_qp.qp.send_cq))) { ++ ret = ENOMEM; + *bad_wr = wr; + goto out; + } +-- +2.33.0 + diff --git a/rdma-core.spec b/rdma-core.spec index 21a690e..4135ada 100644 --- a/rdma-core.spec +++ b/rdma-core.spec @@ -1,6 +1,6 @@ Name: rdma-core Version: 50.0 -Release: 13 +Release: 14 Summary: RDMA core userspace libraries and daemons License: GPLv2 or BSD Url: https://github.com/linux-rdma/rdma-core @@ -42,6 +42,8 @@ patch33: 0033-libhns-Fix-missing-flag-when-creating-qp-by-hnsdv_cr.patch patch34: 0034-librdmacm-Fix-an-overflow-bug-in-qsort-comparison-function.patch patch35: 0035-Fix-the-stride-calculation-for-MSN-PSN-area.patch patch36: 0036-add-ZTE-Dinghai-rdma-driver.patch +patch37: 0037-libhns-Fix-out-of-order-issue-of-requester-when-sett.patch +patch38: 0038-libhns-Fix-reference-to-uninitialized-cq-pointer.patch BuildRequires: binutils cmake >= 2.8.11 gcc libudev-devel pkgconfig pkgconfig(libnl-3.0) BuildRequires: pkgconfig(libnl-route-3.0) systemd systemd-devel @@ -627,6 +629,12 @@ fi %doc %{_docdir}/%{name}-%{version}/tag_matching.md %doc %{_docdir}/%{name}-%{version}/70-persistent-ipoib.rules +* Fri Nov 15 2024 Xinghai Cen <cenxinghai(a)h-partners.com> - 50.0-14 +- Type: bugfix +- ID: NA +- SUG: NA +- DESC: Two bugfixes in post_send flow + %changelog * Sat Aug 31 2024 Li Fuyan <li.fuyan(a)zte.com.cn> - 50.0-13 - Type: requirement -- 2.33.0

1 0

高性能网络sig会议
by openEuler conference 15 Nov '24

15 Nov '24

您好！ sig-high-performance-network 邀请您参加 2024-11-15 14:00 召开的WeLink会议(自动录制) 会议主题：高性能网络sig会议会议链接：https://meeting.huaweicloud.com:36443/#/j/965651261 会议纪要：https://etherpad.openeuler.org/p/sig-high-performance-network-meetings 更多资讯尽在：https://www.openeuler.org/zh/ Hello! sig-high-performance-network invites you to attend the WeLink conference(auto recording) will be held at 2024-11-15 14:00, The subject of the conference is 高性能网络sig会议, You can join the meeting at https://meeting.huaweicloud.com:36443/#/j/965651261. Add topics at https://etherpad.openeuler.org/p/sig-high-performance-network-meetings. More information: https://www.openeuler.org/en/

1 0

[PATCH rdma-core 0/2] libhns: Two bugfixes in post_send flow
by Junxian Huang 08 Nov '24

08 Nov '24

Two hns bugfixes in post_send flow Chengchang Tang (1): libhns: Fix reference to uninitialized cq pointer Junxian Huang (1): libhns: Fix out-of-order issue of requester when setting FENCE providers/hns/hns_roce_u_hw_v2.c | 12 ++++++------ 1 file changed, 6 insertions(+), 6 deletions(-) -- 2.33.0

1 2

高性能网络sig例会
by openEuler conference 06 Nov '24

06 Nov '24

您好！ sig-high-performance-network 邀请您参加 2024-11-08 11:00 召开的WeLink会议(自动录制) 会议主题：高性能网络sig例会会议链接：https://meeting.huaweicloud.com:36443/#/j/960933891 会议纪要：https://etherpad.openeuler.org/p/sig-high-performance-network-meetings 更多资讯尽在：https://www.openeuler.org/zh/ Hello! sig-high-performance-network invites you to attend the WeLink conference(auto recording) will be held at 2024-11-08 11:00, The subject of the conference is 高性能网络sig例会, You can join the meeting at https://meeting.huaweicloud.com:36443/#/j/960933891. Add topics at https://etherpad.openeuler.org/p/sig-high-performance-network-meetings. More information: https://www.openeuler.org/en/

1 0

高性能网络sig会议
by openEuler conference 23 Oct '24

23 Oct '24

您好！ sig-high-performance-network 邀请您参加 2024-10-25 11:00 召开的WeLink会议会议主题：高性能网络sig会议会议链接：https://meeting.huaweicloud.com:36443/#/j/982220702 会议纪要：https://etherpad.openeuler.org/p/sig-high-performance-network-meetings 更多资讯尽在：https://www.openeuler.org/zh/ Hello! sig-high-performance-network invites you to attend the WeLink conference will be held at 2024-10-25 11:00, The subject of the conference is 高性能网络sig会议, You can join the meeting at https://meeting.huaweicloud.com:36443/#/j/982220702. Add topics at https://etherpad.openeuler.org/p/sig-high-performance-network-meetings. More information: https://www.openeuler.org/en/

1 0

Update kernel headers
by Junxian Huang 14 Oct '24

14 Oct '24

To commit: ?? ("RDMA/hns: Support mmapping reset state to userspace"). Signed-off-by: Junxian Huang <huangjunxian6(a)hisilicon.com> --- kernel-headers/rdma/hns-abi.h | 6 ++++++ 1 file changed, 6 insertions(+) diff --git a/kernel-headers/rdma/hns-abi.h b/kernel-headers/rdma/hns-abi.h index 94e861870..065eb2e0a 100644 --- a/kernel-headers/rdma/hns-abi.h +++ b/kernel-headers/rdma/hns-abi.h @@ -136,6 +136,7 @@ struct hns_roce_ib_alloc_ucontext_resp { __u32 max_inline_data; __u8 congest_type; __u8 reserved0[7]; + __aligned_u64 reset_mmap_key; }; struct hns_roce_ib_alloc_ucontext { @@ -153,4 +154,9 @@ struct hns_roce_ib_create_ah_resp { __u8 tc_mode; }; +struct hns_roce_reset_state { + __u32 hw_ready; + __u32 reserved; +}; + #endif /* HNS_ABI_USER_H */ -- 2.33.0

1 3

Re: rdma-core PR流程rpm差异分析已完成，待下一步处理
by li.fuyan＠zte.com.cn 27 Sep '24

27 Sep '24

你好，帮忙看下新的rpm差异分析是否还有问题。 https://gitee.com/src-openeuler/rdma-core/pulls/165 感谢！李富艳软件开发工程师 ICF六部/智算云底座研发中心/无线及算力研究院 ICF Dept. VI/Intelligent Computing and Cloud Foundation R&D Center/Wireless and Computing Product R&D Institute 中兴通讯股份有限公司西安市长安区西沣路五星段9号中兴通讯1E-401 邮编: 710114 T: +86 029 xxxxxxxx M: +86 15332482266 E: li.fuyan(a)zte.com.cn www.zte.com.cn Original From: 李富艳00122684 To: tangchengchang <tangchengchang(a)huawei.com>; Cc: high-performance-network(a)openeuler.org <high-performance-network(a)openeuler.org>;dev(a)openeuler.org <dev(a)openeuler.org>; Date: 2024年09月19日 14:49 Subject: rdma-core PR流程rpm差异分析已完成，待下一步处理您好，下面PR链接已经在9月11日针对RPM构建的差异完成了分析，下一步处理是不是到您那边了，请帮忙看一下。 https://gitee.com/src-openeuler/rdma-core/pulls/165 感谢！李富艳软件开发工程师 ICF六部/智算云底座研发中心/无线及算力研究院 ICF Dept. VI/Intelligent Computing and Cloud Foundation R&D Center/Wireless and Computing Product R&D Institute 中兴通讯股份有限公司西安市长安区西沣路五星段9号中兴通讯1E-401 邮编: 710114 T: +86 029 xxxxxxxx M: +86 15332482266 E: li.fuyan(a)zte.com.cn www.zte.com.cn

1 0

rdma-core PR流程rpm差异分析已完成，待下一步处理
by li.fuyan＠zte.com.cn 19 Sep '24

19 Sep '24

您好，下面PR链接已经在9月11日针对RPM构建的差异完成了分析，下一步处理是不是到您那边了，请帮忙看一下。 https://gitee.com/src-openeuler/rdma-core/pulls/165 感谢！李富艳软件开发工程师 ICF六部/智算云底座研发中心/无线及算力研究院 ICF Dept. VI/Intelligent Computing and Cloud Foundation R&D Center/Wireless and Computing Product R&D Institute 中兴通讯股份有限公司西安市长安区西沣路五星段9号中兴通讯1E-401 邮编: 710114 T: +86 029 xxxxxxxx M: +86 15332482266 E: li.fuyan(a)zte.com.cn www.zte.com.cn

1 0

关于rdma-core PR状态
by li.fuyan＠zte.com.cn 05 Sep '24

05 Sep '24

您好，请帮忙看一下我在高性能网络SIG下提交的关于“ZTE Dinghai rdma-core驱动合入”的PR是什么状态，我在页面上看不到构建信息了。 https://gitee.com/src-openeuler/rdma-core/pulls/165 李富艳中兴通讯股份有限公司 T: +86 029 xxxxxxxx M: +86 15332482266 E: li.fuyan(a)zte.com.cn www.zte.com.cn

1 0

Re: 关于欧拉社区合入rdma-core PR时门禁系统构建问题
by li.fuyan＠zte.com.cn 03 Sep '24

03 Sep '24

你好，帮忙看一下如下构建信息，我从日志里看不出来是哪里有问题，需要你的帮助。 PR链接: https://gitee.com/src-openeuler/rdma-core/pulls/165 构建链接： https://openeulerjenkins.osinfra.cn/job/multiarch/job/src-openeuler/job/x86… 构建错误日志： 3771 2024-09-03 17:22:56 patching file redhat/rdma-core.spec 3772 2024-09-03 17:22:56 Hunk #1 succeeded at 176 (offset 2 lines). 3773 2024-09-03 17:22:56 Hunk #2 succeeded at 204 (offset 2 lines). 3774 2024-09-03 17:22:56 Hunk #3 succeeded at 586 (offset 3 lines). 3775 2024-09-03 17:22:56 error: Bad exit status from /var/tmp/rpm-tmp.43Oe9Z (%prep) 3776 2024-09-03 17:22:56 Bad exit status from /var/tmp/rpm-tmp.43Oe9Z (%prep) 3777 2024-09-03 17:22:56 3778 2024-09-03 17:22:56 RPM build errors: 3779 2024-09-03 17:22:56 failed to build rpms with rdma-core.spec 看起来像是redhat/rdma-core.spec文件有异常，但是从如下修改内容来看，没有看到明显的错误，请帮忙看一下。李富艳软件开发工程师 ICF六部/智算云底座研发中心/无线及算力研究院 ICF Dept. VI/Intelligent Computing and Cloud Foundation R&D Center/Wireless and Computing Product R&D Institute 中兴通讯股份有限公司西安市长安区西沣路五星段9号中兴通讯1E-401 邮编: 710114 T: +86 029 xxxxxxxx M: +86 15332482266 E: li.fuyan(a)zte.com.cn www.zte.com.cn Original From: 李富艳00122684 To: tangchengchang <tangchengchang(a)huawei.com>;chenshaowei3(a)h-partners.com <chenshaowei3(a)h-partners.com>;wanghuan158(a)h-partners.com <wanghuan158(a)h-partners.com>; Cc: wanghuan158(a)h-partners.com <wanghuan158(a)h-partners.com>; Date: 2024年08月31日 14:07 Subject: 关于欧拉社区合入PR时门禁系统构建问题 tangchegnchang 你好，已经根据你之前提交的评审意见修改了ZTE Dinghai rdma-core的代码，但是由于操作问题，修改后的代码提交了新的PR，链接如下： https://gitee.com/src-openeuler/rdma-core/pulls/165 门禁系统当前状态为check_build失败，查看build日志，我发现提示信息如下红色标记，changelog部分错误看起来是基础内容，不是本次合入的，请你帮忙看一下如何解决这种问题。 https://eulermaker.compass-ci.openeuler.openatom.cn/package/build-record?os… 2024-08-31 12:59:19 Hunk #1 succeeded at 245 (offset 5 lines). 3772 2024-08-31 12:59:19 3773 2024-08-31 12:59:19 RPM build errors: 3774 2024-08-31 12:59:19 error: Bad exit status from /var/tmp/rpm-tmp.ikn0I8 (%prep) 3775 2024-08-31 12:59:19 bad date in %changelog: Wed May 29 zhangyaqi <zhangyaqi(a)kylinos.cn> - 50.0-11 3776 2024-08-31 12:59:19 Bad exit status from /var/tmp/rpm-tmp.ikn0I8 (%prep) 3777 2024-08-31 12:59:19 failed to build rpms with rdma-core.spec 另外，能否麻烦你加我微信或者你提供下你的电话，方便后续沟通。我的电话：15332482266 我的微信二维码如下，方便的话请你加一下我李富艳软件开发工程师 ICF六部/智算云底座研发中心/无线及算力研究院 ICF Dept. VI/Intelligent Computing and Cloud Foundation R&D Center/Wireless and Computing Product R&D Institute 中兴通讯股份有限公司西安市长安区西沣路五星段9号中兴通讯1E-401 邮编: 710114 T: +86 029 xxxxxxxx M: +86 15332482266 E: li.fuyan(a)zte.com.cn www.zte.com.cn

1 0

Perftest: Fix failure in creating cq when create_cq_ex is not supported by providers
by Junxian Huang 30 Aug '24

30 Aug '24

The previous TD lock-free patch introduced ibv_create_cq_ex() to perftest, but it breaks the provider devices which do not support this API yet. For these devices calling this API leads to an errno of EOPNOTSUPP. So add a check of errno, and if it is EOPNOTSUPP, use ibv_create_cq() as a fallback. Fixes: b6f957f6bc6c ("Perftest: Add support for TD lock-free mode") Signed-off-by: Junxian Huang <huangjunxian6(a)hisilicon.com> --- src/perftest_resources.c | 7 +++++-- 1 file changed, 5 insertions(+), 2 deletions(-) diff --git a/src/perftest_resources.c b/src/perftest_resources.c index 2653207..814ffdb 100755 --- a/src/perftest_resources.c +++ b/src/perftest_resources.c @@ -1547,6 +1547,8 @@ int create_reg_cqs(struct pingpong_context *ctx, #endif ctx->send_cq = ibv_cq_ex_to_cq(ibv_create_cq_ex(ctx->context, &send_cq_attr)); if (!ctx->send_cq) { + if (!user_param->no_lock && errno == EOPNOTSUPP) + goto cq_ex_not_supported; fprintf(stderr, "Couldn't create CQ\n"); return FAILURE; } @@ -1571,7 +1573,9 @@ int create_reg_cqs(struct pingpong_context *ctx, } } return SUCCESS; -#else + +cq_ex_not_supported: +#endif ctx->send_cq = ibv_create_cq(ctx->context,tx_buffer_depth * user_param->num_of_qps, NULL, ctx->send_channel, user_param->eq_num); if (!ctx->send_cq) { @@ -1589,7 +1593,6 @@ int create_reg_cqs(struct pingpong_context *ctx, } return SUCCESS; -#endif } /****************************************************************************** -- 2.33.0

1 0

Perftest: Fix TD lock-free mode not working for QP
by Junxian Huang 19 Aug '24

19 Aug '24

From: Guofeng Yue <yueguofeng(a)h-partners.com> When creating QP in TD lock-free mode, set attr_ex.pd with ctx->pad instead of ctx->pd, otherwise the lock-free won't work. Fixes: 38d9fcfc3580 ("Perftest: Add support for TD lock-free mode") Signed-off-by: Guofeng Yue <yueguofeng(a)h-partners.com> Signed-off-by: Junxian Huang <huangjunxian6(a)hisilicon.com> --- src/perftest_resources.c | 6 ++++++ 1 file changed, 6 insertions(+) diff --git a/src/perftest_resources.c b/src/perftest_resources.c index 40b7ddf..2ea07fb 100755 --- a/src/perftest_resources.c +++ b/src/perftest_resources.c @@ -2384,7 +2384,13 @@ struct ibv_qp* ctx_qp_create(struct pingpong_context *ctx, else if (opcode == IBV_WR_RDMA_READ) attr_ex.send_ops_flags |= IBV_QP_EX_WITH_RDMA_READ; } + + #ifdef HAVE_TD_API + attr_ex.pd = user_param->no_lock ? ctx->pad : ctx->pd; + #else attr_ex.pd = ctx->pd; + #endif + attr_ex.comp_mask |= IBV_QP_INIT_ATTR_SEND_OPS_FLAGS | IBV_QP_INIT_ATTR_PD; attr_ex.send_cq = attr.send_cq; attr_ex.recv_cq = attr.recv_cq; -- 2.33.0

1 0