-----Original Message----- From: Daniel P. Berrange [mailto:berrange@redhat.com] Sent: Monday, July 20, 2015 5:32 PM To: Ren, Qiaowei Cc: libvir-list(a)redhat.com Subject: Re: [libvirt] [PATCH 2/3] Qemu: add CMT support On Sun, Jul 05, 2015 at 07:43:43PM +0800, Qiaowei Ren wrote: > One RFC in > https://www.redhat.com/archives/libvir-list/2015-June/msg01509.html > > CMT (Cache Monitoring Technology) can be used to measure the usage of > cache by VM running on the host. This patch will extend the bulk stats > API (virDomainListGetStats) to add this field. Applications based on > libvirt can use this API to achieve cache usage of VM. Because CMT > implementation in Linux kernel is based on perf mechanism, this patch > will enable perf event for CMT when VM is created and disable it when > VM is destroyed. > > Signed-off-by: Qiaowei Ren <qiaowei.ren(a)intel.com&gt; > diff --git a/src/qemu/qemu_driver.c b/src/qemu/qemu_driver.c index > 4cfae03..8c678c9 100644 > --- a/src/qemu/qemu_driver.c > +++ b/src/qemu/qemu_driver.c > @@ -19320,6 +19320,53 @@ qemuDomainGetStatsBlock(virQEMUDriverPtr > driver, > > #undef QEMU_ADD_COUNT_PARAM > > +static int > +qemuDomainGetStatsCache(virQEMUDriverPtr driver ATTRIBUTE_UNUSED, > + virDomainObjPtr dom, > + virDomainStatsRecordPtr record, > + int *maxparams, > + unsigned int privflags ATTRIBUTE_UNUSED) So this is a method that is used to collect per-domain information > +{ > + qemuDomainObjPrivatePtr priv = dom->privateData; > + FILE *fd; > + unsigned long long cache = 0; > + int scaling_factor = 0; > + > + if (priv->cmt_fd <= 0) > + return -1; > + > + if (read(priv->cmt_fd, &cache, sizeof(uint64_t)) < 0) { > + virReportSystemError(errno, "%s", > + _("Unable to read cache data")); > + return -1; > + } > + > + fd = fopen("/sys/devices/intel_cqm/events/llc_occupancy.scale", "r"); > + if (!fd) { > + virReportSystemError(errno, "%s", > + _("Unable to open CMT scale file")); > + return -1; > + } > + if (fscanf(fd, "%d", &scaling_factor) != 1) { > + virReportSystemError(errno, "%s", > + _("Unable to read CMT scale file")); > + VIR_FORCE_FCLOSE(fd); > + return -1; > + } > + VIR_FORCE_FCLOSE(fd); But this data you are reading is global to the entire host.

> + > + cache *= scaling_factor; > + > + if (virTypedParamsAddULLong(&record->params, > + &record->nparams, > + maxparams, > + "cache.current", > + cache) < 0) > + return -1; > + > + return 0; > +} > + > diff --git a/src/qemu/qemu_process.c b/src/qemu/qemu_process.c index > ba84182..00b889d 100644 > --- a/src/qemu/qemu_process.c > +++ b/src/qemu/qemu_process.c > +/* > + * Enable CMT(Cache Monitoring Technology) to measure the usage of > + * cache by VM running on the node. > + * > + * Because the hypervisor implement CMT support basedon perf > +mechanism, > + * we should enable perf event for CMT. The function 'sys_erf_event_open' > + * is perf syscall wrapper. > + */ > +#ifdef __linux__ > +static long sys_perf_event_open(struct perf_event_attr *hw_event, > + pid_t pid, int cpu, int group_fd, > + unsigned long flags) { > + return syscall(__NR_perf_event_open, hw_event, pid, cpu, > + group_fd, flags); } static int > +qemuCmtEnable(virDomainObjPtr vm) { > + qemuDomainObjPrivatePtr priv = vm->privateData; > + struct perf_event_attr cmt_attr; > + int event_type; > + FILE *fp; > + > + fp = fopen("/sys/devices/intel_cqm/type", "r"); > + if (!fp) { > + virReportSystemError(errno, "%s", > + _("CMT is not available on this host")); > + return -1; > + } > + if (fscanf(fp, "%d", &event_type) != 1) { > + virReportSystemError(errno, "%s", > + _("Unable to read event type file.")); > + VIR_FORCE_FCLOSE(fp); > + return -1; > + } > + VIR_FORCE_FCLOSE(fp); > + > + memset(&cmt_attr, 0, sizeof(struct perf_event_attr)); > + cmt_attr.size = sizeof(struct perf_event_attr); > + cmt_attr.type = event_type; > + cmt_attr.config = 1; > + cmt_attr.inherit = 1; > + cmt_attr.disabled = 1; > + cmt_attr.enable_on_exec = 0; > + > + priv->cmt_fd = sys_perf_event_open(&cmt_attr, vm->pid, -1, -1, 0); > + if (priv->cmt_fd < 0) { > + virReportSystemError(errno, > + _("Unable to open perf type=%d for pid=%d"), > + event_type, vm->pid); > + return -1; > + } > + > + if (ioctl(priv->cmt_fd, PERF_EVENT_IOC_ENABLE) < 0) { > + virReportSystemError(errno, "%s", > + _("Unable to enable perf event for CMT")); > + return -1; > + } > + > + return 0; > +} > +#else > +static int qemuCmtEnable(virDomainObjPtr vm) { > + virReportUnsupportedError(); > + return -1; > +} > +#endif > + > int qemuProcessStart(virConnectPtr conn, > virQEMUDriverPtr driver, > virDomainObjPtr vm, @@ -4954,6 +5026,11 @@ int > qemuProcessStart(virConnectPtr conn, > if (virDomainSaveStatus(driver->xmlopt, cfg->stateDir, vm) < 0) > goto cleanup; > > + VIR_DEBUG("Setting CMT perf counter"); > + if (qemuCmtEnable(vm) < 0) > + virReportSystemError(errno, "%s", > + _("CMT is not available on this host")); > + > /* finally we can call the 'started' hook script if any */ > if (virHookPresent(VIR_HOOK_DRIVER_QEMU)) { > char *xml = qemuDomainDefFormatXML(driver, vm->def, 0); @@ > -5122,6 +5199,15 @@ void qemuProcessStop(virQEMUDriverPtr driver, > virPortAllocatorRelease(driver->migrationPorts, priv->nbdPort); > priv->nbdPort = 0; > > + /* Disable CMT */ > + if (priv->cmt_fd > 0) { You can't rely on keeping an open file descriptor for the guest because libvirtd may be restarted.

> + if (ioctl(priv->cmt_fd, PERF_EVENT_IOC_DISABLE) < 0) { > + virReportSystemError(errno, "%s", > + _("Unable to disable perf event for CMT")); > + } > + VIR_FORCE_CLOSE(priv->cmt_fd); > + } > + > if (priv->agent) { > qemuAgentClose(priv->agent); > priv->agent = NULL; Conceptually I think this approach to implementation is flawed. While you are turning on/off the perf events for each QEMU process, the data collection does not distinguish data from each QEMU process - the data reported is host wide. So this really doesn't make much sense IMHO.

I'm also wondering whether this is really going to be sufficiently useful on its own. CPUs have countless other performance counters that I would imagine apps/admins will want to read in order to analyse QEMU performance, beyond this new CMT feature. The domain stats API won't really scale up to dealing with arbitrary perf event counter reporting so I'm not much of a fan of just special casing CMT in this way. IOW, if we want to support host performance analysis in libvirt, then we probably want to design an set of APIs specifically for this purpose, but I could well see us saying that this is out of scope for libvirt and apps shoud just use the linux perf interfaces directly.

On Mon, Jul 20, 2015 at 01:50:54PM +0000, Ren, Qiaowei wrote: > > Daniel P. Berrange wrote on Jul 20, 2015 17:32: >> On Sun, Jul 05, 2015 at 07:43:43PM +0800, Qiaowei Ren wrote: >>> One RFC in >>> https://www.redhat.com/archives/libvir-list/2015-June/msg01509.htm >>> l >>> >>> CMT (Cache Monitoring Technology) can be used to measure the usage >>> of cache by VM running on the host. This patch will extend the >>> bulk stats API (virDomainListGetStats) to add this field. >>> Applications based on libvirt can use this API to achieve cache >>> usage of VM. Because CMT implementation in Linux kernel is based >>> on perf mechanism, this patch will enable perf event for CMT when >>> VM is created and disable it when VM is destroyed. >>> >>> Signed-off-by: Qiaowei Ren <qiaowei.ren(a)intel.com&gt; >>> >>> diff --git a/src/qemu/qemu_driver.c b/src/qemu/qemu_driver.c index >>> 4cfae03..8c678c9 100644 --- a/src/qemu/qemu_driver.c +++ >>> b/src/qemu/qemu_driver.c @@ -19320,6 +19320,53 @@ >>> qemuDomainGetStatsBlock(virQEMUDriverPtr driver, >>> >>> #undef QEMU_ADD_COUNT_PARAM >>> +static int +qemuDomainGetStatsCache(virQEMUDriverPtr driver >>> ATTRIBUTE_UNUSED, + virDomainObjPtr dom, + >>> virDomainStatsRecordPtr record, + >>> int *maxparams, + unsigned int privflags >>> ATTRIBUTE_UNUSED) >> >> So this is a method that is used to collect per-domain information >> >>> +{ >>> + qemuDomainObjPrivatePtr priv = dom->privateData; >>> + FILE *fd; >>> + unsigned long long cache = 0; >>> + int scaling_factor = 0; >>> + >>> + if (priv->cmt_fd <= 0) >>> + return -1; >>> + >>> + if (read(priv->cmt_fd, &cache, sizeof(uint64_t)) < 0) { >>> + virReportSystemError(errno, "%s", >>> + _("Unable to read cache data")); >>> + return -1; >>> + } >>> + >>> + fd = fopen("/sys/devices/intel_cqm/events/llc_occupancy.scale", "r"); >>> + if (!fd) { >>> + virReportSystemError(errno, "%s", >>> + _("Unable to open CMT scale file")); >>> + return -1; >>> + } >>> + if (fscanf(fd, "%d", &scaling_factor) != 1) { >>> + virReportSystemError(errno, "%s", >>> + _("Unable to read CMT scale file")); >>> + VIR_FORCE_FCLOSE(fd); >>> + return -1; >>> + } >>> + VIR_FORCE_FCLOSE(fd); >> >> But this data you are reading is global to the entire host. >> > > In fact this data is only for per-domain. When the perf syscall is > called to enable perf event for domain, the pid of that domain is > passed. Ah, I see - you rely on the open file descriptor to be associated with the VM pid. >>> -5122,6 +5199,15 @@ void qemuProcessStop(virQEMUDriverPtr driver, >>> virPortAllocatorRelease(driver->migrationPorts, priv->nbdPort); >>> priv->nbdPort = 0; >>> + /* Disable CMT */ >>> + if (priv->cmt_fd > 0) { >> >> You can't rely on keeping an open file descriptor for the guest >> because libvirtd may be restarted. >> > > Sorry, I don't really get the meaning of this. You mean that when > libvirtd is restarted, those resource which the domain opened should > be closed, right? No, when libvirtd is restarted, the domains must all continuing running without loss of state. You open the FD when starting the guest, then libvirtd is restarted, now someone wants to query the perf data. The perf FD will not be open anymore because libvirtd was restarted. At least you'd need to re-open the file descriptor when libvirtd starts up again, for any running guest. I'm not really convinced we want to keep the perf file descriptors open for all the domains for the entire time they are running. Should really only open them when we actually want to read the collected data.

> >>> + if (ioctl(priv->cmt_fd, PERF_EVENT_IOC_DISABLE) < 0) { >>> + virReportSystemError(errno, "%s", >>> + _("Unable to disable perf event for CMT")); >>> + } >>> + VIR_FORCE_CLOSE(priv->cmt_fd); >>> + } >>> + >>> if (priv->agent) { >>> qemuAgentClose(priv->agent); >>> priv->agent = NULL; >> >> >> Conceptually I think this approach to implementation is flawed. While >> you are turning on/off the perf events for each QEMU process, the data >> collection does not distinguish data from each QEMU process - the data >> reported is host wide. So this really doesn't make much sense IMHO. >> > > As mentioned above, the data reported is only for domain. > >> I'm also wondering whether this is really going to be sufficiently >> useful on its own. CPUs have countless other performance counters that >> I would imagine apps/admins will want to read in order to analyse QEMU >> performance, beyond this new CMT feature. The domain stats API won't >> really scale up to dealing with arbitrary perf event counter reporting >> so I'm not much of a fan of just special casing CMT in this way. >> >> IOW, if we want to support host performance analysis in libvirt, >> then we probably want to design an set of APIs specifically for this >> purpose, but I could well see us saying that this is out of scope >> for libvirt and apps shoud just use the linux perf interfaces directly. >> > > Yes. I can get what you mean. Maybe libvirt doesn't have to be > responsible for supporting host performance. > > But I guess cache usage should be important for each domain, if those > apps based on libvirt can achieve this information they will be able > to better check and confirm the domain works normally, like the stats > of cpu/memory/block/... which have be supported in libvirt now. Do you > think so? I'm not saying cache usage is unimportant. There are quite a lot of other hardware event counters in modern CPUs though, so I'm asking why we should add just this new special intel event, and not any of the other existing performance counters that are useful in diagnosing performance issues.

Also, I'm thinking that QEMU has many different threads - VCPU threads, I/O threads, emulator threads and so on. I could see users want to have distinct profiling for these different functional areas of QEMU too instead of just whole- QEMU granularity.

-----Original Message----- From: libvir-list-bounces(a)redhat.com [mailto:libvir-list-bounces@redhat.com] On Behalf Of Ren, Qiaowei Sent: Monday, August 10, 2015 9:06 AM To: 'Daniel P. Berrange' Cc: &#39;libvir-list(a)redhat.com&#39; Subject: Re: [libvirt] [PATCH 2/3] Qemu: add CMT support > -----Original Message----- > From: Ren, Qiaowei > Sent: Tuesday, July 21, 2015 4:00 PM > To: Daniel P. Berrange > Cc: libvir-list(a)redhat.com > Subject: RE: [libvirt] [PATCH 2/3] Qemu: add CMT support > > On Jul 20, 2015 22:34, Daniel P. Berrange wrote: > > On Mon, Jul 20, 2015 at 01:50:54PM +0000, Ren, Qiaowei wrote: > >> > >> Daniel P. Berrange wrote on Jul 20, 2015 17:32: > >>> On Sun, Jul 05, 2015 at 07:43:43PM +0800, Qiaowei Ren wrote: > >>>> One RFC in > >>>> https://www.redhat.com/archives/libvir-list/2015-June/msg01509.ht > >>>> m > >>>> l > >>>> > >>>> CMT (Cache Monitoring Technology) can be used to measure the > >>>> usage of cache by VM running on the host. This patch will extend > >>>> the bulk stats API (virDomainListGetStats) to add this field. > >>>> Applications based on libvirt can use this API to achieve cache > >>>> usage of VM. Because CMT implementation in Linux kernel is based > >>>> on perf mechanism, this patch will enable perf event for CMT when > >>>> VM is created and disable it when VM is destroyed. > >>>> > >>>> Signed-off-by: Qiaowei Ren <qiaowei.ren(a)intel.com&gt; > >>>> > >>>> diff --git a/src/qemu/qemu_driver.c b/src/qemu/qemu_driver.c > >>>> index > >>>> 4cfae03..8c678c9 100644 --- a/src/qemu/qemu_driver.c +++ > >>>> b/src/qemu/qemu_driver.c @@ -19320,6 +19320,53 @@ > >>>> qemuDomainGetStatsBlock(virQEMUDriverPtr driver, > >>>> > >>>> #undef QEMU_ADD_COUNT_PARAM > >>>> +static int +qemuDomainGetStatsCache(virQEMUDriverPtr driver > >>>> ATTRIBUTE_UNUSED, + virDomainObjPtr dom, + > >>>> virDomainStatsRecordPtr record, + > >>>> int *maxparams, + unsigned int privflags > >>>> ATTRIBUTE_UNUSED) > >>> > >>> So this is a method that is used to collect per-domain information > >>> > >>>> +{ > >>>> + qemuDomainObjPrivatePtr priv = dom->privateData; > >>>> + FILE *fd; > >>>> + unsigned long long cache = 0; > >>>> + int scaling_factor = 0; > >>>> + > >>>> + if (priv->cmt_fd <= 0) > >>>> + return -1; > >>>> + > >>>> + if (read(priv->cmt_fd, &cache, sizeof(uint64_t)) < 0) { > >>>> + virReportSystemError(errno, "%s", > >>>> + _("Unable to read cache data")); > >>>> + return -1; > >>>> + } > >>>> + > >>>> + fd = fopen("/sys/devices/intel_cqm/events/llc_occupancy.scale", "r"); > >>>> + if (!fd) { > >>>> + virReportSystemError(errno, "%s", > >>>> + _("Unable to open CMT scale file")); > >>>> + return -1; > >>>> + } > >>>> + if (fscanf(fd, "%d", &scaling_factor) != 1) { > >>>> + virReportSystemError(errno, "%s", > >>>> + _("Unable to read CMT scale file")); > >>>> + VIR_FORCE_FCLOSE(fd); > >>>> + return -1; > >>>> + } > >>>> + VIR_FORCE_FCLOSE(fd); > >>> > >>> But this data you are reading is global to the entire host. > >>> > >> > >> In fact this data is only for per-domain. When the perf syscall is > >> called to enable perf event for domain, the pid of that domain is > >> passed. > > > > Ah, I see - you rely on the open file descriptor to be associated with the VM pid. > > > > > >>>> -5122,6 +5199,15 @@ void qemuProcessStop(virQEMUDriverPtr driver, > >>>> virPortAllocatorRelease(driver->migrationPorts, priv->nbdPort); > >>>> priv->nbdPort = 0; > >>>> + /* Disable CMT */ > >>>> + if (priv->cmt_fd > 0) { > >>> > >>> You can't rely on keeping an open file descriptor for the guest > >>> because libvirtd may be restarted. > >>> > >> > >> Sorry, I don't really get the meaning of this. You mean that when > >> libvirtd is restarted, those resource which the domain opened > >> should be closed, right? > > > > No, when libvirtd is restarted, the domains must all continuing > > running without loss of state. You open the FD when starting the > > guest, then libvirtd is restarted, now someone wants to query the > > perf data. The perf FD will not be open anymore because libvirtd was > > restarted. At least you'd need to re-open the file descriptor when > > libvirtd starts up again, for any running guest. I'm not really > > convinced we want to keep the perf file descriptors open for all the > > domains for the entire time they are running. Should really only > > open them > when we actually want to read the collected data. > > > > Got it! Should open/disable them when read the data. > > >> > >>>> + if (ioctl(priv->cmt_fd, PERF_EVENT_IOC_DISABLE) < 0) { > >>>> + virReportSystemError(errno, "%s", > >>>> + _("Unable to disable perf event for CMT")); > >>>> + } > >>>> + VIR_FORCE_CLOSE(priv->cmt_fd); > >>>> + } > >>>> + > >>>> if (priv->agent) { > >>>> qemuAgentClose(priv->agent); > >>>> priv->agent = NULL; > >>> > >>> > >>> Conceptually I think this approach to implementation is flawed. > >>> While you are turning on/off the perf events for each QEMU > >>> process, the data collection does not distinguish data from each > >>> QEMU process > >>> - the data reported is host wide. So this really doesn't make much > >>> sense > IMHO. > >>> > >> > >> As mentioned above, the data reported is only for domain. > >> > >>> I'm also wondering whether this is really going to be sufficiently > >>> useful on its own. CPUs have countless other performance counters > >>> that I would imagine apps/admins will want to read in order to > >>> analyse QEMU performance, beyond this new CMT feature. The domain > >>> stats API won't really scale up to dealing with arbitrary perf > >>> event counter reporting so I'm not much of a fan of just special > >>> casing CMT in this > way. > >>> > >>> IOW, if we want to support host performance analysis in libvirt, > >>> then we probably want to design an set of APIs specifically for > >>> this purpose, but I could well see us saying that this is out of > >>> scope for libvirt and apps shoud just use the linux perf interfaces directly. > >>> > >> > >> Yes. I can get what you mean. Maybe libvirt doesn't have to be > >> responsible for supporting host performance. > >> > >> But I guess cache usage should be important for each domain, if > >> those apps based on libvirt can achieve this information they will > >> be able to better check and confirm the domain works normally, like > >> the stats of cpu/memory/block/... which have be supported in > >> libvirt now. Do you think so? > > > > I'm not saying cache usage is unimportant. There are quite a lot of > > other hardware event counters in modern CPUs though, so I'm asking > > why we should add just this new special intel event, and not any of > > the other existing performance counters that are useful in > > diagnosing > performance issues. > > Ah, I guess you mean that the best way is providing a set of utility > method for perf operation, like cgroup support. Then we can use these > interface to support a lot of necessary perf counters. > > If we have to do so, maybe I can try to firstly implement such methods > and then you can help review them. > > > Also, I'm thinking that QEMU has many different threads - VCPU > > threads, I/O threads, emulator threads and so on. I could see users > > want to have distinct profiling for these different functional areas > > of QEMU too instead of just whole- QEMU granularity. > > > > Yes. I believe such features should be useful for some users. I am > currently working on adding some features like CMT into OpenStack, I > only know profiling for VCPU/ IO / emulator threads should be not > necessary for OpenStack. ^-^ > Hi Daniel, what do you think about it now? If we have to add CMT support in Nova, do you think what is the best way? Can we directly use linux perf interface in OpenStack?