[libvirt] [PATCH 03/11] qemu: Implement the ability to return IOThread stats
Michal Privoznik
mprivozn at redhat.com
Fri Oct 19 11:06:38 UTC 2018
On 10/07/2018 03:00 PM, John Ferlan wrote:
> Process the IOThreads polling stats if available. Generate the
> output params record to be returned to the caller with the three
> values - poll-max-ns, poll-grow, and poll-shrink.
>
> Signed-off-by: John Ferlan <jferlan at redhat.com>
> ---
> include/libvirt/libvirt-domain.h | 1 +
> src/libvirt-domain.c | 38 ++++++++++++++++
> src/qemu/qemu_driver.c | 78 ++++++++++++++++++++++++++++++++
> 3 files changed, 117 insertions(+)
>
> diff --git a/include/libvirt/libvirt-domain.h b/include/libvirt/libvirt-domain.h
> index fdd2d6b8ea..58fd4bc10c 100644
> --- a/include/libvirt/libvirt-domain.h
> +++ b/include/libvirt/libvirt-domain.h
> @@ -2048,6 +2048,7 @@ typedef enum {
> VIR_DOMAIN_STATS_INTERFACE = (1 << 4), /* return domain interfaces info */
> VIR_DOMAIN_STATS_BLOCK = (1 << 5), /* return domain block info */
> VIR_DOMAIN_STATS_PERF = (1 << 6), /* return domain perf event info */
> + VIR_DOMAIN_STATS_IOTHREAD = (1 << 7), /* return iothread poll info */
> } virDomainStatsTypes;
>
> typedef enum {
> diff --git a/src/libvirt-domain.c b/src/libvirt-domain.c
> index 7690339521..9fda56d660 100644
> --- a/src/libvirt-domain.c
> +++ b/src/libvirt-domain.c
> @@ -11499,6 +11499,44 @@ virConnectGetDomainCapabilities(virConnectPtr conn,
> * long long. It is produced by the
> * emulation_faults perf event
> *
> + * VIR_DOMAIN_STATS_IOTHREAD:
> + * Return IOThread statistics if available. IOThread polling is a
> + * timing mechanism that allows the hypervisor to generate a longer
> + * period of time in which the guest will perform operations on the
> + * CPU being used by the IOThread. The higher the value for poll-max-ns
> + * the longer the guest will keep the CPU. This may affect other host
> + * threads using the CPU. The poll-grow and poll-shrink values allow
> + * the hypervisor to generate a mechanism to add or remove polling time
> + * within the confines of 0 and poll-max-ns. For QEMU, the poll-grow is
> + * multiplied by the polling interval, while poll-shrink is used as a
> + * divisor. When not provided, QEMU may double the polling time until
> + * poll-max-ns is reached. When poll-shrink is 0 (zero) QEMU may reset
> + * the polling interval to 0 until it finds its "sweet spot". Setting
> + * poll-grow too large may cause frequent fluctution of the time; however,
> + * this can be tempered by a high poll-shrink to reduce the polling
> + * interval. For example, a poll-grow of 3 will triple the polling time
> + * which could quickly exceed poll-max-ns; however, a poll-shrink of
> + * 10 would cut that polling time more gradually.
> + *
> + * The typed parameter keys are in this format:
> + *
> + * "iothread.cnt" - maximum number of IOThreads in the subsequent list
> + * as unsigned int. Each IOThread in the list will
> + * will use it's iothread_id value as the <id>. There
> + * may be fewer <id> entries than the iothread.cnt
> + * value if the polling values are not supported.
> + * "iothread.<id>.poll-max-ns" - maximum polling time in ns as an unsigned
> + * long long. A 0 (zero) means polling is
> + * disabled.
> + * "iothread.<id>.poll-grow" - polling time factor as an unsigned int.
> + * A 0 (zero) indicates to allow the underlying
> + * hypervisor to choose how to grow the
> + * polling time.
> + * "iothread.<id>.poll-shrink" - polling time divisor as an unsigned int.
> + * A 0 (zero) indicates to allow the underlying
> + * hypervisor to choose how to shrink the
> + * polling time.
> + *
> * Note that entire stats groups or individual stat fields may be missing from
> * the output in case they are not supported by the given hypervisor, are not
> * applicable for the current state of the guest domain, or their retrieval
> diff --git a/src/qemu/qemu_driver.c b/src/qemu/qemu_driver.c
> index e0edb43557..ff87865fe6 100644
> --- a/src/qemu/qemu_driver.c
> +++ b/src/qemu/qemu_driver.c
> @@ -20431,6 +20431,83 @@ qemuDomainGetStatsBlock(virQEMUDriverPtr driver,
>
> #undef QEMU_ADD_NAME_PARAM
>
> +#define QEMU_ADD_IOTHREAD_PARAM_UI(record, maxparams, id, name, value) \
> + do { \
> + char param_name[VIR_TYPED_PARAM_FIELD_LENGTH]; \
> + snprintf(param_name, VIR_TYPED_PARAM_FIELD_LENGTH, \
> + "block.%u.%s", id, name); \
s/block/iothread/
> + if (virTypedParamsAddUInt(&(record)->params, \
> + &(record)->nparams, \
> + maxparams, \
> + param_name, \
> + value) < 0) \
> + goto cleanup; \
> + } while (0)
> +
> +#define QEMU_ADD_IOTHREAD_PARAM_ULL(record, maxparams, id, name, value) \
> +do { \
> + char param_name[VIR_TYPED_PARAM_FIELD_LENGTH]; \
> + snprintf(param_name, VIR_TYPED_PARAM_FIELD_LENGTH, \
> + "iothread.%u.%s", id, name); \
> + if (virTypedParamsAddULLong(&(record)->params, \
> + &(record)->nparams, \
> + maxparams, \
> + param_name, \
> + value) < 0) \
> + goto cleanup; \
> +} while (0)
> +
> +static int
> +qemuDomainGetStatsIOThread(virQEMUDriverPtr driver,
> + virDomainObjPtr dom,
> + virDomainStatsRecordPtr record,
> + int *maxparams,
> + unsigned int privflags ATTRIBUTE_UNUSED)
> +{
> + size_t i;
> + qemuMonitorIOThreadInfoPtr *iothreads = NULL;
> + int niothreads;
> + int ret = -1;
This MUST have the check for job - it is talking to a monitor and
therefore @dom is locked and unlocked meanwhile. You can take a look at
qemuDomainGetStatsBlock() what the check should look like.
> +
> + if ((niothreads = qemuDomainGetIOThreadsMon(driver, dom, &iothreads)) < 0)
> + return -1;
Worse, since you moved acquiring the job here, this will fail miserably
when running all stats for a domain. Because the code works like this:
1) it calls qemuDomainGetStatsNeedMonitor() to figure out if monitor is
needed and if it is a job is started,
2) it iterate over all stats callbacks,
3) this function is reached which tries to set the job again.
> +
> + if (niothreads == 0)
> + return 0;
> +
> + QEMU_ADD_COUNT_PARAM(record, maxparams, "iothread", niothreads);
> +
> + for (i = 0; i < niothreads; i++) {
> + if (iothreads[i]->poll_valid) {
> + QEMU_ADD_IOTHREAD_PARAM_ULL(record, maxparams,
> + iothreads[i]->iothread_id,
> + "poll-max-ns",
> + iothreads[i]->poll_max_ns);
> + QEMU_ADD_IOTHREAD_PARAM_UI(record, maxparams,
> + iothreads[i]->iothread_id,
> + "poll-grow",
> + iothreads[i]->poll_grow);
> + QEMU_ADD_IOTHREAD_PARAM_UI(record, maxparams,
> + iothreads[i]->iothread_id,
> + "poll-shrink",
> + iothreads[i]->poll_shrink);
> + }
> + }
> +
> + ret = 0;
> +
> + cleanup:
> + for (i = 0; i < niothreads; i++)
> + VIR_FREE(iothreads[i]);
> + VIR_FREE(iothreads);
> +
> + return ret;
> +}
> +
> +#undef QEMU_ADD_IOTHREAD_PARAM_UI
> +
> +#undef QEMU_ADD_IOTHREAD_PARAM_ULL
> +
> #undef QEMU_ADD_COUNT_PARAM
>
> static int
> @@ -20505,6 +20582,7 @@ static struct qemuDomainGetStatsWorker qemuDomainGetStatsWorkers[] = {
> { qemuDomainGetStatsInterface, VIR_DOMAIN_STATS_INTERFACE, false },
> { qemuDomainGetStatsBlock, VIR_DOMAIN_STATS_BLOCK, true },
> { qemuDomainGetStatsPerf, VIR_DOMAIN_STATS_PERF, false },
> + { qemuDomainGetStatsIOThread, VIR_DOMAIN_STATS_IOTHREAD, false },
s/false/true/ The function you're calling does require access to the
monitor.
> { NULL, 0, false }
> };
>
>
Michal
More information about the libvir-list
mailing list