Hi
On Fri, Jul 1, 2022 at 7:11 AM zhenwei pi <pizhen...@bytedance.com> wrote:
> A vCPU thread always reaches 100% utilization when:
> - guest uses idle=poll
> - disable HLT vm-exit
> - enable MWAIT
>
> Add new guest agent command 'guest-get-cpustats' to get guest CPU
> statistics, we can know the guest workload and how busy the CPU is.
>
> Signed-off-by: zhenwei pi <pizhen...@bytedance.com>
> ---
> qga/commands-posix.c | 72 ++++++++++++++++++++++++++++++++++++++++++++
> qga/commands-win32.c | 6 ++++
> qga/qapi-schema.json | 49 ++++++++++++++++++++++++++++++
> 3 files changed, 127 insertions(+)
>
> diff --git a/qga/commands-posix.c b/qga/commands-posix.c
> index 0469dc409d..2847023876 100644
> --- a/qga/commands-posix.c
> +++ b/qga/commands-posix.c
> @@ -2893,6 +2893,73 @@ GuestDiskStatsInfoList
> *qmp_guest_get_diskstats(Error **errp)
> return guest_get_diskstats(errp);
> }
>
> +GuestCpuStatsList *qmp_guest_get_cpustats(Error **errp)
> +{
> + GuestCpuStatsList *head = NULL, **tail = &head;
> + const char *cpustats = "/proc/stat";
> + FILE *fp;
> + size_t n;
> + char *line = NULL;
> +
> + fp = fopen(cpustats, "r");
> + if (fp == NULL) {
> + error_setg_errno(errp, errno, "open(\"%s\")", cpustats);
> + return NULL;
> + }
> +
> + while (getline(&line, &n, fp) != -1) {
> + GuestCpuStats *cpustat = NULL;
> + int i;
> + unsigned long user, system, idle, iowait, irq, softirq, steal,
> guest;
> + unsigned long nice, guest_nice;
> + char name[64];
> +
> + i = sscanf(line, "%s %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu",
> + name, &user, &nice, &system, &idle, &iowait, &irq,
> &softirq,
> + &steal, &guest, &guest_nice);
> +
> + /* drop "cpu 1 2 3 ...", get "cpuX 1 2 3 ..." only */
> + if (strncmp(name, "cpu", 3) || (name[3] == '\0')) {
>
For extra safety, check !name as well
> + continue;
> + }
> +
>
if i < 5, I guess you should warn and continue
> + cpustat = g_new0(GuestCpuStats, 1);
> + cpustat->cpu = atoi(&name[3]);
> + cpustat->has_user = true;
> + cpustat->user = user * 10;
>
proc(5) says that the value is given "in units of USER_HZ (1/100ths of a
second on most architectures, use sysconf(_SC_CLK_TCK) to obtain the right
value)", so we should adjust this code if we want to return correctly in ms.
> + cpustat->has_system = true;
> + cpustat->system = system * 10;
> + cpustat->has_idle = true;
> + cpustat->idle = idle * 10;
> +
> + /* Linux version >= 2.6 */
>
That's pretty old now (2003), not sure anyone would care about that
comment, but np ;)
> + if (i > 5) {
> + cpustat->has_iowait = true;
> + cpustat->iowait = iowait * 10;
> + cpustat->has_irq = true;
> + cpustat->irq = irq * 10;
> + cpustat->has_softirq = true;
> + cpustat->softirq = softirq * 10;
> + }
> +
> + if (i > 8) {
> + cpustat->has_steal = true;
> + cpustat->steal = steal * 10;
> + }
> +
> + if (i > 9) {
> + cpustat->has_guest = true;
> + cpustat->guest = guest * 10;
> + }
> +
> + QAPI_LIST_APPEND(tail, cpustat);
> + }
> +
> + free(line);
> + fclose(fp);
> + return head;
> +}
> +
> #else /* defined(__linux__) */
>
> void qmp_guest_suspend_disk(Error **errp)
> @@ -3247,6 +3314,11 @@ GuestDiskStatsInfoList
> *qmp_guest_get_diskstats(Error **errp)
> return NULL;
> }
>
> +GuestCpuStatsList *qmp_guest_get_cpustats(Error **errp)
> +{
> + error_setg(errp, QERR_UNSUPPORTED);
> + return NULL;
> +}
>
> #endif /* CONFIG_FSFREEZE */
>
> diff --git a/qga/commands-win32.c b/qga/commands-win32.c
> index 36f94c0f9c..7ed7664715 100644
> --- a/qga/commands-win32.c
> +++ b/qga/commands-win32.c
> @@ -2543,3 +2543,9 @@ GuestDiskStatsInfoList
> *qmp_guest_get_diskstats(Error **errp)
> error_setg(errp, QERR_UNSUPPORTED);
> return NULL;
> }
> +
> +GuestCpuStatsList *qmp_guest_get_cpustats(Error **errp)
> +{
> + error_setg(errp, QERR_UNSUPPORTED);
> + return NULL;
> +}
> diff --git a/qga/qapi-schema.json b/qga/qapi-schema.json
> index 9fa20e791b..4859c887b2 100644
> --- a/qga/qapi-schema.json
> +++ b/qga/qapi-schema.json
> @@ -1576,3 +1576,52 @@
> { 'command': 'guest-get-diskstats',
> 'returns': ['GuestDiskStatsInfo']
> }
> +
> +##
> +# @GuestCpuStats:
> +#
> +# Get statistics of each CPU in millisecond.
> +#
> +# @cpu: CPU index in guest OS
> +#
> +# @user: CPU time of user mode
>
"Time spent in user mode." is more understandable (from man proc(5))
Same for the other descriptions.
+#
> +# @system: CPU time of system mode
> +#
> +# @idle: CPU time of idle state
> +#
> +# @iowait: CPU time waiting IO
> +#
> +# @irq: CPU time of hardware interrupt
> +#
> +# @softirq: CPU time of soft interrupt
> +#
> +# @steal: CPU time stolen by host
> +#
> +# @guest: CPU time of running guest mode
>
Why not "guest_nice" ?
Do we expect this struct to be equally meaningful for other OSes?
Otherwise, I would suggest to make a "linux" variant, perhaps.
+#
> +# Since: 7.1
> +##
> +{ 'struct': 'GuestCpuStats',
> + 'data': {'cpu': 'int',
> + '*user': 'uint64',
> + '*system': 'uint64',
> + '*idle': 'uint64',
> + '*iowait': 'uint64',
> + '*irq': 'uint64',
> + '*softirq': 'uint64',
> + '*steal': 'uint64',
> + '*guest': 'uint64'
> + } }
> +
> +##
> +# @guest-get-cpustats:
> +#
> +# Retrieve information about CPU stats.
> +# Returns: List of CPU stats of guest.
> +#
> +# Since: 7.1
> +##
> +{ 'command': 'guest-get-cpustats',
> + 'returns': ['GuestCpuStats']
> +}
> --
> 2.20.1
>
>
>
--
Marc-André Lureau
