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

Accepting request 558102 from home:NMoreyChaisemartin:branches:science:HPC

Browse Source 
- Add srp_daemon-handle-SM-lid-change.patch and
  srp_daemon-fix-CQ-handling.patch to fix issue
  with SRP daemon not handling SM changes (bsc#1072884)

OBS-URL: https://build.opensuse.org/request/show/558102
OBS-URL: https://build.opensuse.org/package/show/science:HPC/rdma-core?expand=0&rev=68
This commit is contained in:
Nicolas Morey-Chaisemartin 2017-12-18 13:33:53 +00:00 committed by Git OBS Bridge
parent 060b3d871d
commit 5d8ed532f3
4 changed files with 216 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
rdma-core.changes
View File

@ -1,3 +1,10 @@
-------------------------------------------------------------------
Fri Dec 15 08:28:17 UTC 2017 - nmoreychaisemartin@suse.com
- Add srp_daemon-handle-SM-lid-change.patch and
srp_daemon-fix-CQ-handling.patch to fix issue
with SRP daemon not handling SM changes (bsc#1072884)
-------------------------------------------------------------------
Wed Dec 6 08:06:46 UTC 2017 - nmoreychaisemartin@suse.com

4
rdma-core.spec
View File

@ -50,6 +50,8 @@ Group: Productivity/Networking/Other
Url: https://github.com/linux-rdma/rdma-core
Source: rdma-core-%{version}%{git_ver}.tar.gz
Source1: baselibs.conf
Patch1: srp_daemon-handle-SM-lid-change.patch
Patch2: srp_daemon-fix-CQ-handling.patch
BuildRequires: binutils
BuildRequires: cmake >= 2.8.11
BuildRequires: gcc
@ -325,6 +327,8 @@ on those changes.
%prep
%setup -q -n %{name}-%{version}%{git_ver}
%patch1
%patch2
%build

105
srp_daemon-fix-CQ-handling.patch Normal file
View File

@ -0,0 +1,105 @@
commit c1c584c34d249987d7c36ff061bc5f2eedec38fe
Author: Nicolas Morey-Chaisemartin <NMoreyChaisemartin@suse.com>
Date: Mon Dec 11 15:37:28 2017 +0100
srp_daemon: fix CQ handling
SM traps are polled through poll_cq which waited for a CQ event
before polling the CQ itself.
However it may happens that multiple completions are attached
to a single event. As stated by the ibv_get_cq_event man page,
it is required to poll the the CQ to get those completions
after the call to ibv_req_notify_cq.
As completions need to be handled one by one in an outer function,
start by polling the CQ and return the completion (if any) before
waiting for the next completion event.
This will allow emptying all pending completions, through multiple calls
to poll_cq, before waiting for a new event.
The buggy use case seems to appear when the master SM is switched multiple
times between two nodes. As the number of ping-pong between the SMs increases,
the number of traps sent to notify that the SM just became master increases
too. This causes burst of completions linked to a single event.
Note that the race condition is also possible in other scenario.
Signed-off-by: Nicolas Morey-Chaisemartin <NMoreyChaisemartin@suse.com>
Cc: stable@linux-rdma.org # v14, v15, v16
diff --git srp_daemon/srp_handle_traps.c srp_daemon/srp_handle_traps.c
index 25f2b9ab..77a47db3 100644
--- srp_daemon/srp_handle_traps.c
+++ srp_daemon/srp_handle_traps.c
@@ -496,6 +496,34 @@ static int stop_threads(struct sync_resources *sync_res)
return result;
}
+/*****************************************************************************
+* Function: poll_cq_once
+* Poll a CQ once.
+* Returns the number of completion polled (0 or 1).
+* Returns a negative value on error.
+*****************************************************************************/
+static int poll_cq_once(struct sync_resources *sync_res, struct ibv_cq *cq,
+ struct ibv_wc *wc)
+{
+ int ret;
+
+ ret = ibv_poll_cq(cq, 1, wc);
+ if (ret < 0) {
+ pr_err("poll CQ failed\n");
+ return ret;
+ }
+
+ if (ret > 0 && wc->status != IBV_WC_SUCCESS) {
+ if (!stop_threads(sync_res))
+ pr_err("got bad completion with status: 0x%x\n",
+ wc->status);
+ return -ret;
+ }
+
+ return ret;
+}
+
+
static int poll_cq(struct sync_resources *sync_res, struct ibv_cq *cq,
struct ibv_wc *wc, struct ibv_comp_channel *channel)
{
@@ -504,6 +532,16 @@ static int poll_cq(struct sync_resources *sync_res, struct ibv_cq *cq,
void *ev_ctx;
if (channel) {
+ /* There may be extra completions that
+ * were associated to the previous event.
+ * Only poll for the first one. If there are more than one,
+ * they will be handled by later call to poll_cq */
+ ret = poll_cq_once(sync_res, cq, wc);
+ /* return directly if there was an error or
+ * 1 completion polled */
+ if (ret)
+ return ret;
+
if (ibv_get_cq_event(channel, &ev_cq, &ev_ctx)) {
pr_err("Failed to get cq_event\n");
return -1;
@@ -524,18 +562,9 @@ static int poll_cq(struct sync_resources *sync_res, struct ibv_cq *cq,
}
do {
- ret = ibv_poll_cq(cq, 1, wc);
- if (ret < 0) {
- pr_err("poll CQ failed\n");
+ ret = poll_cq_once(sync_res, cq, wc);
+ if (ret < 0)
return ret;
- }
-
- if (ret > 0 && wc->status != IBV_WC_SUCCESS) {
- if (!stop_threads(sync_res))
- pr_err("got bad completion with status: 0x%x\n",
- wc->status);
- return -ret;
- }
if (ret == 0 && channel) {
pr_err("Weird poll returned no cqe after CQ event\n");

100
srp_daemon-handle-SM-lid-change.patch Normal file
View File

@ -0,0 +1,100 @@
commit 2fbc501061218e7df8e37bb2df6db73e00005e9b
Author: Nicolas Morey-Chaisemartin <NMoreyChaisemartin@suse.com>
Date: Mon Dec 4 15:15:55 2017 +0100
srp_daemon: handle SM lid change
When srp_daemon was running and the master SM host changes,
srp_daemon output these errors at every scan:
srp_daemon[25394]: No response to inform info registration
srp_daemon[25394]: Fail to register to traps, maybe there is no opensm
running on fabric or IB port is down
This was introduced by commit 4952e5f Fix a memory leak.
A side effect of this patch was that create_ah was only called when the
port lid changes. Which meant register_to_traps used an older, obsolete,
version of sm_lid and failed to connect to it.
This patch fixes this behaviour by checking for both local lid changes and
SM lid changes, and calling create_ah on any of these events.
Fixes: 4952e5f7 (Fix a memory leak)
Signed-off-by: Nicolas Morey-Chaisemartin <NMoreyChaisemartin@suse.com>
Cc: stable@linux-rdma.org # v14, v15, v16
diff --git srp_daemon/srp_daemon.c srp_daemon/srp_daemon.c
index 2465ccd9..36df5c3b 100644
--- srp_daemon/srp_daemon.c
+++ srp_daemon/srp_daemon.c
@@ -1103,7 +1103,7 @@ static int get_shared_pkeys(struct resources *res,
int i, num_pkeys = 0;
uint16_t pkey;
uint16_t local_port_lid = get_port_lid(res->ud_res->ib_ctx,
- config->port_num);
+ config->port_num, NULL);
in_mad_buf = malloc(sizeof(struct ib_user_mad) +
node_table_response_size);
@@ -2092,7 +2092,7 @@ int main(int argc, char *argv[])
{
int ret;
struct resources *res;
- uint16_t lid;
+ uint16_t lid, sm_lid;
uint16_t pkey;
union umad_gid gid;
struct target_details *target;
@@ -2196,8 +2196,10 @@ catas_start:
pr_debug("Starting a recalculation\n");
port_lid = get_port_lid(res->ud_res->ib_ctx,
- config->port_num);
- if (port_lid != res->ud_res->port_attr.lid) {
+ config->port_num, &sm_lid);
+ if (port_lid != res->ud_res->port_attr.lid ||
+ sm_lid != res->ud_res->port_attr.sm_lid) {
+
if (res->ud_res->ah) {
ibv_destroy_ah(res->ud_res->ah);
res->ud_res->ah = NULL;
diff --git srp_daemon/srp_daemon.h srp_daemon/srp_daemon.h
index 5d268ed3..864b3d42 100644
--- srp_daemon/srp_daemon.h
+++ srp_daemon/srp_daemon.h
@@ -299,7 +299,7 @@ void *run_thread_listen_to_events(void *res_in);
int get_node(struct umad_resources *umad_res, uint16_t dlid, uint64_t *guid);
int create_trap_resources(struct ud_resources *ud_res);
int register_to_traps(struct resources *res, int subscribe);
-uint16_t get_port_lid(struct ibv_context *ib_ctx, int port_num);
+uint16_t get_port_lid(struct ibv_context *ib_ctx, int port_num, uint16_t *sm_lid);
int create_ah(struct ud_resources *ud_res);
void push_gid_to_list(struct sync_resources *res, union umad_gid *gid,
uint16_t pkey);
diff --git srp_daemon/srp_handle_traps.c srp_daemon/srp_handle_traps.c
index 6d94634e..25f2b9ab 100644
--- srp_daemon/srp_handle_traps.c
+++ srp_daemon/srp_handle_traps.c
@@ -340,12 +340,20 @@ int ud_resources_create(struct ud_resources *res)
return 0;
}
-uint16_t get_port_lid(struct ibv_context *ib_ctx, int port_num)
+uint16_t get_port_lid(struct ibv_context *ib_ctx, int port_num, uint16_t *sm_lid)
{
struct ibv_port_attr port_attr;
+ int ret;
+
+ ret = ibv_query_port(ib_ctx, port_num, &port_attr);
- return ibv_query_port(ib_ctx, port_num, &port_attr) == 0 ?
- port_attr.lid : 0;
+ if (!ret) {
+ if (sm_lid)
+ *sm_lid = port_attr.sm_lid;
+ return port_attr.lid;
+ }
+
+ return 0;
}
int create_ah(struct ud_resources *ud_res)