On Thu, May 12, 2016 at 14:55:16 +0800, Qiaowei Ren wrote:
Some Intel processor families (e.g. the Intel Xeon processor E5 v3
family) introduced some RDT (Resource Director Technology) features
to monitor or control shared resource. Among these features, MBM
(Memory Bandwidth Monitoring), which is build on the CMT (Cache
Monitoring Technology) infrastructure, provides OS/VMM a way to
monitor bandwidth from one level of cache to another.
With current perf framework, this patch adds support to perf event
for MBM.
Signed-off-by: Qiaowei Ren <qiaowei.ren(a)intel.com>
---
include/libvirt/libvirt-domain.h | 18 ++++++++++++
src/qemu/qemu_driver.c | 35 +++++++++++++++++++----
src/util/virperf.c | 60 ++++++++++++++++++++++++----------------
src/util/virperf.h | 2 ++
4 files changed, 85 insertions(+), 30 deletions(-)
diff --git a/include/libvirt/libvirt-domain.h b/include/libvirt/libvirt-domain.h
index 160f20f..e4594f0 100644
--- a/include/libvirt/libvirt-domain.h
+++ b/include/libvirt/libvirt-domain.h
@@ -1904,6 +1904,24 @@ void virDomainStatsRecordListFree(virDomainStatsRecordPtr
*stats);
*/
# define VIR_PERF_PARAM_CMT "cmt"
+/**
+ * VIR_PERF_PARAM_MBMT:
+ *
+ * Macro for typed parameter name that represents MBMT perf event
+ * which can be used to monitor total system bandwidth (bytes/s)
+ * from one level of cache to another.
+ */
+# define VIR_PERF_PARAM_MBMT "mbmt"
+
+/**
+ * VIR_PERF_PARAM_MBML:
+ *
+ * Macro for typed parameter name that represents MBML perf event
+ * which can be used to monitor the amount of data (bytes/s) sent
+ * through the memory controller on the socket.
+ */
+# define VIR_PERF_PARAM_MBML "mbml"
So I was a little bit mistaken what this is documenting. So these
document the arguments for virDomainSetPerfEvents where the perf stuff
is enabled.
I was actually thinking this is documenting the values that are returned
via the virConnectGetAllDomainStats/virDomainListGetStats APIs. I've
just noticed that the perf events are not documented at all there.
This should be fixed also for the CMT event. This part of the
documentation should then state which fields in the *Stats APIs the perf
events correspond to.
In the documentation for the fields returned bulk Stats API the
individual fields should be marked by which event they are produced.
+
int virDomainGetPerfEvents(virDomainPtr dom,
virTypedParameterPtr *params,
int *nparams,
diff --git a/src/qemu/qemu_driver.c b/src/qemu/qemu_driver.c
index c4c4968..a3bd7ec 100644
--- a/src/qemu/qemu_driver.c
+++ b/src/qemu/qemu_driver.c
@@ -10051,6 +10051,8 @@ qemuDomainSetPerfEvents(virDomainPtr dom,
if (virTypedParamsValidate(params, nparams,
VIR_PERF_PARAM_CMT, VIR_TYPED_PARAM_BOOLEAN,
+ VIR_PERF_PARAM_MBMT, VIR_TYPED_PARAM_BOOLEAN,
+ VIR_PERF_PARAM_MBML, VIR_TYPED_PARAM_BOOLEAN,
NULL) < 0)
return -1;
@@ -19495,20 +19497,38 @@ qemuDomainGetStatsBlock(virQEMUDriverPtr driver,
#undef QEMU_ADD_COUNT_PARAM
static int
-qemuDomainGetStatsPerfCmt(virPerfPtr perf,
+qemuDomainGetStatsPerfRdt(virPerfPtr perf,
+ virPerfEventType type,
virDomainStatsRecordPtr record,
int *maxparams)
{
- uint64_t cache = 0;
+ uint64_t value = 0;
- if (virPerfReadEvent(perf, VIR_PERF_EVENT_CMT, &cache) < 0)
+ if (virPerfReadEvent(perf, type, &value) < 0)
return -1;
- if (virTypedParamsAddULLong(&record->params,
+ if (type == VIR_PERF_EVENT_CMT &&
+ virTypedParamsAddULLong(&record->params,
&record->nparams,
maxparams,
"perf.cache",
- cache) < 0)
+ value) < 0)
+ return -1;
+
+ if (type == VIR_PERF_EVENT_MBMT &&
+ virTypedParamsAddULLong(&record->params,
+ &record->nparams,
+ maxparams,
+ "perf.total_bytes",
If you pass this string into this function you won't need to special
case all the different types.
Also by naming the fields similarly to the name of the feature
(virPerfEvent enum) you could entirely avoid the need to pass the name
of the value as it could be generated by the virPerfEventTypeToString.
(concatenate it with 'perf.' to get 'perf.mbmt', 'perf.cmt' and
'perf.mbml').
Unfortunately the implementation of the CMT event already used a
different name for the stats entry.
Since this actually was never documented I would be in favor of changing
perf.cache to perf.cmt, so that the macros can be used. I cc'd danpb who
reviewed and pushed the first impl for another opinion.
[...]
diff --git a/src/util/virperf.c b/src/util/virperf.c
index bd65587..826d36a 100644
--- a/src/util/virperf.c
+++ b/src/util/virperf.c
[...]
@@ -127,32 +127,42 @@ virPerfCmtEnable(virPerfEventPtr event,
if (virStrToLong_ui(buf, NULL, 10, &event_type) < 0) {
virReportSystemError(errno, "%s",
- _("failed to get cmt event type"));
+ _("failed to get rdt event type"));
[...]
@@ -162,7 +172,7 @@ virPerfCmtEnable(virPerfEventPtr event,
if (ioctl(event->fd, PERF_EVENT_IOC_ENABLE) < 0) {
virReportSystemError(errno, "%s",
- _("Unable to enable perf event for CMT"));
+ _("Unable to enable perf event for RDT"));
Both error messages could keep the name of the event by using
virPerfEventTypeToString.
Peter