pool/rdma-core
SHA256
2
0
Fork 1
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Accepting request 540247 from home:NMoreyChaisemartin:branches:science:HPC

Browse Source 
- Backport upstream patches
  * bnxt_re-lib-fix-the-memory-barrier-call-during-poll-cq.patch
    Fix memory ordering issue
  * bnxt_re-lib-increment-psn-in-case-of-0-length-packets.patch
    Fix PSN getting out of sync when sending 0 length packet
  * verbs-Do-not-block-QP-attr_masks-used-by-older-kernels.patch
    Fix bits detection to allow RDMA CM to work on older kernsl
- Refresh older patches to include commit logs:
  * ibacm-Incorrect-list-used-for-subnet-list-causes-a-segfault.patch
  * ibacm-Incorrect-usage-of-BE-byte-order-of-MLID-attach-detach_mcast.patch
  * libibumad-umad.c-In-get_port-ignore-sysfs-rate-file-errors.patch
  * libqedr-fix-inline-data-copy.patch

OBS-URL: https://build.opensuse.org/request/show/540247
OBS-URL: https://build.opensuse.org/package/show/science:HPC/rdma-core?expand=0&rev=59
This commit is contained in:
Nicolas Morey-Chaisemartin 2017-11-09 15:20:02 +00:00 committed by Git OBS Bridge
parent f33a17a320
commit 6c79bc4c77
9 changed files with 240 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

33
bnxt_re-lib-fix-the-memory-barrier-call-during-poll-cq.patch Normal file
View File

@ -0,0 +1,33 @@
commit e3609c2c453fc3198a759db17d660f04781f8583
Author: Devesh Sharma <devesh.sharma@broadcom.com>
Date: Thu Nov 9 02:31:58 2017 -0500
bnxt_re/lib: fix the memory barrier call during poll-cq
Putting a read barrier before issuing a read on valid bit
is incorrect. When checking for the validity of CQE in the
CQ buffer the code must wait for the read-barrier to finish
after issuing a read operation on CQE valid bit.
Signed-off-by: Devesh Sharma <devesh.sharma@broadcom.com>
diff --git providers/bnxt_re/main.h providers/bnxt_re/main.h
index 9688fec6..82c89484 100644
--- providers/bnxt_re/main.h
+++ providers/bnxt_re/main.h
@@ -366,9 +366,13 @@ static inline uint8_t bnxt_re_to_ibv_wc_status(uint8_t bnxt_wcst,
static inline uint8_t bnxt_re_is_cqe_valid(struct bnxt_re_cq *cq,
struct bnxt_re_bcqe *hdr)
{
+ uint8_t valid = 0;
+
+ valid = ((le32toh(hdr->flg_st_typ_ph) &
+ BNXT_RE_BCQE_PH_MASK) == cq->phase);
udma_from_device_barrier();
- return ((le32toh(hdr->flg_st_typ_ph) &
- BNXT_RE_BCQE_PH_MASK) == cq->phase);
+
+ return valid;
}
static inline void bnxt_re_change_cq_phase(struct bnxt_re_cq *cq)

29
bnxt_re-lib-increment-psn-in-case-of-0-length-packets.patch Normal file
View File

@ -0,0 +1,29 @@
commit 093dadc78984c9e44ea1c86ca4527a592a42417e
Author: Devesh Sharma <devesh.sharma@broadcom.com>
Date: Thu Nov 9 02:42:29 2017 -0500
bnxt_re/lib: increment psn in case of 0 length packets
If application posts a 0 length packet, post send routine
is skipping to increment the psn number. This will cause
PSN number to go out of sync and eventually connection would
terminate due to sequence error.
post_send routine must increment the psn number by 1 even
for zero length packets.
Signed-off-by: Devesh Sharma <devesh.sharma@broadcom.com>
diff --git providers/bnxt_re/verbs.c providers/bnxt_re/verbs.c
index 4d9b044a..9d4e02bb 100644
--- providers/bnxt_re/verbs.c
+++ providers/bnxt_re/verbs.c
@@ -1048,6 +1048,8 @@ static void bnxt_re_fill_psns(struct bnxt_re_qp *qp, struct bnxt_re_psns *psns,
pkt_cnt = (len / qp->mtu);
if (len % qp->mtu)
pkt_cnt++;
+ if (len == 0)
+ pkt_cnt = 1;
nxt_psn = ((qp->sq_psn + pkt_cnt) & BNXT_RE_PSNS_NPSN_MASK);
psns->flg_npsn = htole32(nxt_psn);
qp->sq_psn = nxt_psn;

34
ibacm-Incorrect-list-used-for-subnet-list-causes-a-segfault.patch
View File

@ -1,3 +1,37 @@
commit e733db47760bcdb59d8296fecc4e06c7d451712b
Author: Michael J. Ruhl <michael.j.ruhl@intel.com>
Date: Mon Oct 16 10:24:40 2017 -0400
ibacm: Incorrect list used for subnet list causes a segfault
Setting the provider keyword in the ibacm_opts.cfg file to something
other than 'default' will cause ibacm to segfault on startup:
ibacm[32739]: segfault at 302e40 ip 000000000040432d
sp 00007ffe4039e1c0 error 4 in ibacm[400000+c000]
To re-produce the segfault, change the provider keyword in
ibacm_opts.cfg from:
provider ibacmp default
to:
provider ibacmp 0xFE80000000000000
When adding subnets to a provider subnet list, the incorrect list is
used. The list used is the provider_list (list of all providers)
rather than the (specific) provider subnet list.
This corrupts the provider_list, and causes ibacm to crash with the
above segfault.
Use the correct list when adding subnet information to a provider.
Fixes: 26e05f8304a506 ("ibacm: use ccan/list.h")
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Michael J. Ruhl <michael.j.ruhl@intel.com>
diff --git ibacm/src/acm.c ibacm/src/acm.c
index 367a43fe..9367fe89 100644
--- ibacm/src/acm.c

28
ibacm-Incorrect-usage-of-BE-byte-order-of-MLID-attach-detach_mcast.patch
View File

@ -1,3 +1,31 @@
commit ed6a98208e04be9401eb727b57e78a29eecbe660
Author: Michael J. Ruhl <michael.j.ruhl@intel.com>
Date: Mon Oct 16 10:24:21 2017 -0400
ibacm: Incorrect usage of BE byte order of MLID attach/detach_mcast()
The MLID value passed to ibv_attach/detach_mcast() must be in host
byte order.
acmp.c incorrectly uses the big endian format when doing a multicast
attach/detach (join). Multicast packets are used to do name resolution
by the libibacmp library.
There are two possible results because of this issue.
If a kernel has commit 8561eae60ff9, the attach will fail with an
EINVAL. ibacm will log this as a failure during the multicast join.
If a kernel does not have commit 8561eae60ff9, the attach will
complete successfully. Packets sent to this address will be dropped
because the packet dlid value and the multicast address information
given by the attach will not match.
Update MLID usage to use the correct byte order.
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Michael J. Ruhl <michael.j.ruhl@intel.com>
diff --git ibacm/prov/acmp/src/acmp.c ibacm/prov/acmp/src/acmp.c
index aa784166..78d9a295 100644
--- ibacm/prov/acmp/src/acmp.c

19
libibumad-umad.c-In-get_port-ignore-sysfs-rate-file-errors.patch
View File

@ -1,3 +1,22 @@
commit 1da6cfac736285ac822b1f48aec327ce42822356
Author: Hal Rosenstock <hal@mellanox.com>
Date: Tue Oct 31 14:05:42 2017 +0200
libibumad/umad.c: In get_port, ignore sysfs rate file errors
This can cause ibpanic in ibstat when width is not set properly
as can occur when QSPF is not plugged into port.
ibpanic: [7851] main: stat of IB device 'mlx5_1' failed: Invalid argument
It's caused by kernel sysfs.c:show_rate returning -EINVAL
and that error being treated as failure in umad.c:get_port.
With this change, Rate is displayed as 0 with ibstat for this scenario.
Tested by: Ghazale Hosseinabadi <ghazale.hosseinabadi@oracle.com>
Signed-off-by: Hal Rosenstock <hal@mellanox.com>
diff --git libibumad/umad.c libibumad/umad.c
index 535256e2..dcb2c680 100644
--- libibumad/umad.c

10
libqedr-fix-inline-data-copy.patch
View File

@ -1,3 +1,13 @@
commit 774d9d7a17606ffc37bb43b87130c4f177de8e25
Author: Ram Amrani <Ram.Amrani@cavium.com>
Date: Mon Oct 30 12:32:35 2017 +0200
libqedr: fix inline data copy
Advance target pointer during data copy.
Signed-off-by: Ram Amrani <Ram.Amrani@cavium.com>
diff --git providers/qedr/qelr_verbs.c providers/qedr/qelr_verbs.c
index 4a74c50e..7db0fb32 100644
--- providers/qedr/qelr_verbs.c

17
rdma-core.changes
View File

@ -1,3 +1,20 @@
-------------------------------------------------------------------
Thu Nov 9 13:34:40 UTC 2017 - nmoreychaisemartin@suse.com
- Backport upstream patches
* bnxt_re-lib-fix-the-memory-barrier-call-during-poll-cq.patch
Fix memory ordering issue
* bnxt_re-lib-increment-psn-in-case-of-0-length-packets.patch
Fix PSN getting out of sync when sending 0 length packet
* verbs-Do-not-block-QP-attr_masks-used-by-older-kernels.patch
Fix bits detection to allow RDMA CM to work on older kernsl
- Refresh older patches to include commit logs:
* ibacm-Incorrect-list-used-for-subnet-list-causes-a-segfault.patch
* ibacm-Incorrect-usage-of-BE-byte-order-of-MLID-attach-detach_mcast.patch
* libibumad-umad.c-In-get_port-ignore-sysfs-rate-file-errors.patch
* libqedr-fix-inline-data-copy.patch
-------------------------------------------------------------------
Wed Nov 8 14:05:44 UTC 2017 - nmoreychaisemartin@suse.com

6
rdma-core.spec
View File

@ -54,6 +54,9 @@ Patch0: ibacm-Incorrect-usage-of-BE-byte-order-of-MLID-attach-detach_mca
Patch1: ibacm-Incorrect-list-used-for-subnet-list-causes-a-segfault.patch
Patch2: libqedr-fix-inline-data-copy.patch
Patch3: libibumad-umad.c-In-get_port-ignore-sysfs-rate-file-errors.patch
Patch4: bnxt_re-lib-fix-the-memory-barrier-call-during-poll-cq.patch
Patch5: bnxt_re-lib-increment-psn-in-case-of-0-length-packets.patch
Patch6: verbs-Do-not-block-QP-attr_masks-used-by-older-kernels.patch
BuildRequires: binutils
BuildRequires: cmake >= 2.8.11
BuildRequires: gcc
@ -319,6 +322,9 @@ on those changes.
%patch1
%patch2
%patch3
%patch4
%patch5
%patch6
%build

64
verbs-Do-not-block-QP-attr_masks-used-by-older-kernels.patch Normal file
View File

@ -0,0 +1,64 @@
commit a5d774e8d247ebcdabe3e450abadfb659d2ae47a
Author: Jason Gunthorpe <jgunthorpe@ziepe.ca>
Date: Tue Nov 7 10:30:08 2017 -0700
verbs: Do not block QP attr_masks used by older kernels
At least the RDMA CM in 4.3 will give user space structs with these bits
set. Those kernels require user space to pass those set bits back to the
kernel. Blocking them causes RDMA CM to be unusable on older kernels.
Fixes: 3ca7a1031486 ("ibverbs: Add support for packet pacing")
Reported-by: Ram Amrani <Ram.Amrani@cavium.com>
Signed-off-by: Jason Gunthorpe <jgunthorpe@ziepe.ca>
diff --git libibverbs/cmd.c libibverbs/cmd.c
index 9d2140fb..77aba334 100644
--- libibverbs/cmd.c
+++ libibverbs/cmd.c
@@ -1154,10 +1154,10 @@ int ibv_cmd_query_qp(struct ibv_qp *qp, struct ibv_qp_attr *attr,
struct ibv_query_qp_resp resp;
/*
- * Masks over IBV_QP_DEST_QPN are not supported by
- * that not extended command.
+ * Starting with IBV_QP_RATE_LIMIT the attribute must go through the
+ * _ex path.
*/
- if (attr_mask & ~((IBV_QP_DEST_QPN << 1) - 1))
+ if (attr_mask & ~(IBV_QP_RATE_LIMIT - 1))
return EOPNOTSUPP;
IBV_INIT_CMD_RESP(cmd, cmd_size, QUERY_QP, &resp, sizeof resp);
@@ -1324,10 +1324,10 @@ int ibv_cmd_modify_qp(struct ibv_qp *qp, struct ibv_qp_attr *attr,
struct ibv_modify_qp *cmd, size_t cmd_size)
{
/*
- * Masks over IBV_QP_DEST_QPN are only supported by
- * ibv_cmd_modify_qp_ex.
+ * Starting with IBV_QP_RATE_LIMIT the attribute must go through the
+ * _ex path.
*/
- if (attr_mask & ~((IBV_QP_DEST_QPN << 1) - 1))
+ if (attr_mask & ~(IBV_QP_RATE_LIMIT - 1))
return EOPNOTSUPP;
IBV_INIT_CMD(cmd, cmd_size, MODIFY_QP);
diff --git libibverbs/verbs.h libibverbs/verbs.h
index 8cdf8ab5..a6a9441d 100644
--- libibverbs/verbs.h
+++ libibverbs/verbs.h
@@ -861,6 +861,13 @@ enum ibv_qp_attr_mask {
IBV_QP_PATH_MIG_STATE = 1 << 18,
IBV_QP_CAP = 1 << 19,
IBV_QP_DEST_QPN = 1 << 20,
+ /* These bits were supported on older kernels, but never exposed from
+ libibverbs:
+ _IBV_QP_SMAC = 1 << 21,
+ _IBV_QP_ALT_SMAC = 1 << 22,
+ _IBV_QP_VID = 1 << 23,
+ _IBV_QP_ALT_VID = 1 << 24,
+ */
IBV_QP_RATE_LIMIT = 1 << 25,
};