Re: [PATCH v4 6/6] powerpc/kvm-hv-pmu: Add perf-events for Hostwide counters

> On 11 Mar 2025, at 3:02=E2=80=AFPM, Vaibhav Jain =
<[email protected]> wrote:
>=20
> Athira Rajeev <[email protected]> writes:
>=20
>>> On 24 Feb 2025, at 6:45=E2=80=AFPM, Vaibhav Jain =
<[email protected]> wrote:
>>>=20
>>> Update 'kvm-hv-pmu.c' to add five new perf-events mapped to the five
>>> Hostwide counters. Since these newly introduced perf events are at =
system
>>> wide scope and can be read from any L1-Lpar CPU, 'kvmppc_pmu' scope =
and
>>> capabilities are updated appropriately.
>>>=20
>>> Also introduce two new helpers. First is kvmppc_update_l0_stats() =
that uses
>>> the infrastructure introduced in previous patches to issues the
>>> H_GUEST_GET_STATE hcall L0-PowerVM to fetch guest-state-buffer =
holding the
>>> latest values of these counters which is then parsed and 'l0_stats'
>>> variable updated.
>>>=20
>>> Second helper is kvmppc_pmu_event_update() which is called from
>>> 'kvmppv_pmu' callbacks and uses kvmppc_update_l0_stats() to update
>>> 'l0_stats' and the update the 'struct perf_event's event-counter.
>>>=20
>>> Some minor updates to kvmppc_pmu_{add, del, read}() to remove some =
debug
>>> scaffolding code.
>>>=20
>>> Signed-off-by: Vaibhav Jain <[email protected]>
>>> ---
>>> Changelog
>>>=20
>>> v3->v4:
>>> * Minor tweaks to patch description and code as its now being built =
as a
>>> separate kernel module.
>>>=20
>>> v2->v3:
>>> None
>>>=20
>>> v1->v2:
>>> None
>>> ---
>>> arch/powerpc/perf/kvm-hv-pmu.c | 92 =
+++++++++++++++++++++++++++++++++-
>>> 1 file changed, 91 insertions(+), 1 deletion(-)
>>>=20
>>> diff --git a/arch/powerpc/perf/kvm-hv-pmu.c =
b/arch/powerpc/perf/kvm-hv-pmu.c
>>> index ed371454f7b5..274459bb32d6 100644
>>> --- a/arch/powerpc/perf/kvm-hv-pmu.c
>>> +++ b/arch/powerpc/perf/kvm-hv-pmu.c
>>> @@ -30,6 +30,11 @@
>>> #include "asm/guest-state-buffer.h"
>>>=20
>>> enum kvmppc_pmu_eventid {
>>> + KVMPPC_EVENT_HOST_HEAP,
>>> + KVMPPC_EVENT_HOST_HEAP_MAX,
>>> + KVMPPC_EVENT_HOST_PGTABLE,
>>> + KVMPPC_EVENT_HOST_PGTABLE_MAX,
>>> + KVMPPC_EVENT_HOST_PGTABLE_RECLAIM,
>>> KVMPPC_EVENT_MAX,
>>> };
>>>=20
>>> @@ -61,8 +66,14 @@ static DEFINE_SPINLOCK(lock_l0_stats);
>>> /* GSB related structs needed to talk to L0 */
>>> static struct kvmppc_gs_msg *gsm_l0_stats;
>>> static struct kvmppc_gs_buff *gsb_l0_stats;
>>> +static struct kvmppc_gs_parser gsp_l0_stats;
>>>=20
>>> static struct attribute *kvmppc_pmu_events_attr[] =3D {
>>> + KVMPPC_PMU_EVENT_ATTR(host_heap, KVMPPC_EVENT_HOST_HEAP),
>>> + KVMPPC_PMU_EVENT_ATTR(host_heap_max, KVMPPC_EVENT_HOST_HEAP_MAX),
>>> + KVMPPC_PMU_EVENT_ATTR(host_pagetable, KVMPPC_EVENT_HOST_PGTABLE),
>>> + KVMPPC_PMU_EVENT_ATTR(host_pagetable_max, =
KVMPPC_EVENT_HOST_PGTABLE_MAX),
>>> + KVMPPC_PMU_EVENT_ATTR(host_pagetable_reclaim, =
KVMPPC_EVENT_HOST_PGTABLE_RECLAIM),
>>> NULL,
>>> };
>>>=20
>>> @@ -71,7 +82,7 @@ static const struct attribute_group =
kvmppc_pmu_events_group =3D {
>>> .attrs =3D kvmppc_pmu_events_attr,
>>> };
>>>=20
>>> -PMU_FORMAT_ATTR(event, "config:0");
>>> +PMU_FORMAT_ATTR(event, "config:0-5");
>>> static struct attribute *kvmppc_pmu_format_attr[] =3D {
>>> &format_attr_event.attr,
>>> NULL,
>>> @@ -88,6 +99,79 @@ static const struct attribute_group =
*kvmppc_pmu_attr_groups[] =3D {
>>> NULL,
>>> };
>>>=20
>>> +/*
>>> + * Issue the hcall to get the L0-host stats.
>>> + * Should be called with l0-stat lock held
>>> + */
>>> +static int kvmppc_update_l0_stats(void)
>>> +{
>>> + int rc;
>>> +
>>> + /* With HOST_WIDE flags guestid and vcpuid will be ignored */
>>> + rc =3D kvmppc_gsb_recv(gsb_l0_stats, KVMPPC_GS_FLAGS_HOST_WIDE);
>>> + if (rc)
>>> + goto out;
>>> +
>>> + /* Parse the guest state buffer is successful */
>>> + rc =3D kvmppc_gse_parse(&gsp_l0_stats, gsb_l0_stats);
>>> + if (rc)
>>> + goto out;
>>> +
>>> + /* Update the l0 returned stats*/
>>> + memset(&l0_stats, 0, sizeof(l0_stats));
>>> + rc =3D kvmppc_gsm_refresh_info(gsm_l0_stats, gsb_l0_stats);
>>> +
>>> +out:
>>> + return rc;
>>> +}
>>> +
>>> +/* Update the value of the given perf_event */
>>> +static int kvmppc_pmu_event_update(struct perf_event *event)
>>> +{
>>> + int rc;
>>> + u64 curr_val, prev_val;
>>> + unsigned long flags;
>>> + unsigned int config =3D event->attr.config;
>>> +
>>> + /* Ensure no one else is modifying the l0_stats */
>>> + spin_lock_irqsave(&lock_l0_stats, flags);
>>> +
>>> + rc =3D kvmppc_update_l0_stats();
>>> + if (!rc) {
>>> + switch (config) {
>>> + case KVMPPC_EVENT_HOST_HEAP:
>>> + curr_val =3D l0_stats.guest_heap;
>>> + break;
>>> + case KVMPPC_EVENT_HOST_HEAP_MAX:
>>> + curr_val =3D l0_stats.guest_heap_max;
>>> + break;
>>> + case KVMPPC_EVENT_HOST_PGTABLE:
>>> + curr_val =3D l0_stats.guest_pgtable_size;
>>> + break;
>>> + case KVMPPC_EVENT_HOST_PGTABLE_MAX:
>>> + curr_val =3D l0_stats.guest_pgtable_size_max;
>>> + break;
>>> + case KVMPPC_EVENT_HOST_PGTABLE_RECLAIM:
>>> + curr_val =3D l0_stats.guest_pgtable_reclaim;
>>> + break;
>>> + default:
>>> + rc =3D -ENOENT;
>>> + break;
>>> + }
>>> + }
>>> +
>>> + spin_unlock_irqrestore(&lock_l0_stats, flags);
>>> +
>>> + /* If no error than update the perf event */
>>> + if (!rc) {
>>> + prev_val =3D local64_xchg(&event->hw.prev_count, curr_val);
>>> + if (curr_val > prev_val)
>>> + local64_add(curr_val - prev_val, &event->count);
>>> + }
>>> +
>>> + return rc;
>>> +}
>>> +
>>> static int kvmppc_pmu_event_init(struct perf_event *event)
>>> {
>>> unsigned int config =3D event->attr.config;
>>> @@ -110,15 +194,19 @@ static int kvmppc_pmu_event_init(struct =
perf_event *event)
>>>=20
>>> static void kvmppc_pmu_del(struct perf_event *event, int flags)
>>> {
>>> + /* Do nothing */
>>> }
>>=20
>> If we don=E2=80=99t read the counter stats in =E2=80=9Cdel=E2=80=9D =
call back, we will loose the final count getting updated, right ?
>> Del callback needs to call kvmppc_pmu_read. Can you check the =
difference in count stats by calling kvmppc_pmu_read here ?
>>=20
>=20
> Yes, agreed. Will address this in next version of the patch series
>=20

Sure thanks !

Athira
>> Thanks
>> Athira
>>=20
>>>=20
>>> static int kvmppc_pmu_add(struct perf_event *event, int flags)
>>> {
>>> + if (flags & PERF_EF_START)
>>> + return kvmppc_pmu_event_update(event);
>>> return 0;
>>> }
>>>=20
>>> static void kvmppc_pmu_read(struct perf_event *event)
>>> {
>>> + kvmppc_pmu_event_update(event);
>>> }
>>>=20
>>> /* Return the size of the needed guest state buffer */
>>> @@ -302,6 +390,8 @@ static struct pmu kvmppc_pmu =3D {
>>> .read =3D kvmppc_pmu_read,
>>> .attr_groups =3D kvmppc_pmu_attr_groups,
>>> .type =3D -1,
>>> + .scope =3D PERF_PMU_SCOPE_SYS_WIDE,
>>> + .capabilities =3D PERF_PMU_CAP_NO_EXCLUDE | =
PERF_PMU_CAP_NO_INTERRUPT,
>>> };
>>>=20
>>> static int __init kvmppc_register_pmu(void)
>>> --=20
>>> 2.48.1
>>>=20
>>>=20
>>>=20
>>=20
>>=20
>=20
> --=20
> Cheers
> ~ Vaibhav