From: Xinghai Cen cenxinghai@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@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@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@huawei.com +Signed-off-by: Xinghai Cen cenxinghai@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@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@huawei.com +Signed-off-by: Xinghai Cen cenxinghai@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@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@h-partners.com - 50.0-15 - Type: bugfix - ID: NA -- 2.33.0