Re: [PATCH v4 04/38] perf: Add a EVENT_GUEST flag

[Namhyung Kim <namhyung@kernel.org>]

On Mon, Mar 24, 2025 at 05:30:44PM +0000, Mingwei Zhang wrote:
> From: Kan Liang <kan.liang@linux.intel.com>
> 
> Current perf doesn't explicitly schedule out all exclude_guest events
> while the guest is running. There is no problem with the current
> emulated vPMU. Because perf owns all the PMU counters. It can mask the
> counter which is assigned to an exclude_guest event when a guest is
> running (Intel way), or set the corresponding HOSTONLY bit in evsentsel
> (AMD way). The counter doesn't count when a guest is running.
> 
> However, either way doesn't work with the introduced passthrough vPMU.
> A guest owns all the PMU counters when it's running. The host should not
> mask any counters. The counter may be used by the guest. The evsentsel
> may be overwritten.
> 
> Perf should explicitly schedule out all exclude_guest events to release
> the PMU resources when entering a guest, and resume the counting when
> exiting the guest.
> 
> It's possible that an exclude_guest event is created when a guest is
> running. The new event should not be scheduled in as well.
> 
> The ctx time is shared among different PMUs. The time cannot be stopped
> when a guest is running. It is required to calculate the time for events
> from other PMUs, e.g., uncore events. Add timeguest to track the guest
> run time. For an exclude_guest event, the elapsed time equals
> the ctx time - guest time.
> Cgroup has dedicated times. Use the same method to deduct the guest time
> from the cgroup time as well.
> 
> Co-developed-by: Peter Zijlstra (Intel) <peterz@infradead.org>
> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
> Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
> Signed-off-by: Mingwei Zhang <mizhang@google.com>
> ---
>  include/linux/perf_event.h |   6 ++
>  kernel/events/core.c       | 209 +++++++++++++++++++++++++++++--------
>  2 files changed, 169 insertions(+), 46 deletions(-)
> 
> diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h
> index a2fd1bdc955c..7bda1e20be12 100644
> --- a/include/linux/perf_event.h
> +++ b/include/linux/perf_event.h
> @@ -999,6 +999,11 @@ struct perf_event_context {
>  	 */
>  	struct perf_time_ctx		time;
>  
> +	/*
> +	 * Context clock, runs when in the guest mode.
> +	 */
> +	struct perf_time_ctx		timeguest;
> +
>  	/*
>  	 * These fields let us detect when two contexts have both
>  	 * been cloned (inherited) from a common ancestor.
> @@ -1089,6 +1094,7 @@ struct bpf_perf_event_data_kern {
>   */
>  struct perf_cgroup_info {
>  	struct perf_time_ctx		time;
> +	struct perf_time_ctx		timeguest;
>  	int				active;
>  };
>  
> diff --git a/kernel/events/core.c b/kernel/events/core.c
> index e38c8b5e8086..7a2115b2c5c1 100644
> --- a/kernel/events/core.c
> +++ b/kernel/events/core.c
> @@ -163,7 +163,8 @@ enum event_type_t {
>  	/* see ctx_resched() for details */
>  	EVENT_CPU	= 0x10,
>  	EVENT_CGROUP	= 0x20,
> -	EVENT_FLAGS	= EVENT_CGROUP,
> +	EVENT_GUEST	= 0x40,
> +	EVENT_FLAGS	= EVENT_CGROUP | EVENT_GUEST,
>  	/* compound helpers */
>  	EVENT_ALL         = EVENT_FLEXIBLE | EVENT_PINNED,
>  	EVENT_TIME_FROZEN = EVENT_TIME | EVENT_FROZEN,
> @@ -435,6 +436,7 @@ static atomic_t nr_include_guest_events __read_mostly;
>  
>  static atomic_t nr_mediated_pmu_vms;
>  static DEFINE_MUTEX(perf_mediated_pmu_mutex);
> +static DEFINE_PER_CPU(bool, perf_in_guest);
>  
>  /* !exclude_guest event of PMU with PERF_PMU_CAP_MEDIATED_VPMU */
>  static inline bool is_include_guest_event(struct perf_event *event)
> @@ -738,6 +740,9 @@ static bool perf_skip_pmu_ctx(struct perf_event_pmu_context *pmu_ctx,
>  {
>  	if ((event_type & EVENT_CGROUP) && !pmu_ctx->nr_cgroups)
>  		return true;
> +	if ((event_type & EVENT_GUEST) &&
> +	    !(pmu_ctx->pmu->capabilities & PERF_PMU_CAP_MEDIATED_VPMU))
> +		return true;
>  	return false;
>  }
>  
> @@ -788,6 +793,39 @@ static inline void update_perf_time_ctx(struct perf_time_ctx *time, u64 now, boo
>  	WRITE_ONCE(time->offset, time->time - time->stamp);
>  }
>  
> +static_assert(offsetof(struct perf_event_context, timeguest) -
> +	      offsetof(struct perf_event_context, time) ==
> +	      sizeof(struct perf_time_ctx));
> +
> +#define T_TOTAL		0
> +#define T_GUEST		1
> +
> +static inline u64 __perf_event_time_ctx(struct perf_event *event,
> +					struct perf_time_ctx *times)
> +{
> +	u64 time = times[T_TOTAL].time;
> +
> +	if (event->attr.exclude_guest)
> +		time -= times[T_GUEST].time;
> +
> +	return time;
> +}
> +
> +static inline u64 __perf_event_time_ctx_now(struct perf_event *event,
> +					    struct perf_time_ctx *times,
> +					    u64 now)
> +{
> +	if (event->attr.exclude_guest && __this_cpu_read(perf_in_guest)) {
> +		/*
> +		 * (now + times[total].offset) - (now + times[guest].offset) :=
> +		 * times[total].offset - times[guest].offset
> +		 */
> +		return READ_ONCE(times[T_TOTAL].offset) - READ_ONCE(times[T_GUEST].offset);

So this will remove both time_enabled and time_running.  I think it's
fine as the events are not multiplexed, but some curious users may
wonder why the time_enabled is less than expected. :)


> +	}
> +
> +	return now + READ_ONCE(times[T_TOTAL].offset);
> +}
> +

[SNIP] 
> @@ -6285,23 +6400,25 @@ void perf_event_update_userpage(struct perf_event *event)
>  	if (!rb)
>  		goto unlock;
>  
> -	/*
> -	 * compute total_time_enabled, total_time_running
> -	 * based on snapshot values taken when the event
> -	 * was last scheduled in.
> -	 *
> -	 * we cannot simply called update_context_time()
> -	 * because of locking issue as we can be called in
> -	 * NMI context
> -	 */
> -	calc_timer_values(event, &now, &enabled, &running);
> -
> -	userpg = rb->user_page;
>  	/*
>  	 * Disable preemption to guarantee consistent time stamps are stored to
>  	 * the user page.
>  	 */
>  	preempt_disable();
> +
> +	/*
> +	 * compute total_time_enabled, total_time_running
> +	 * based on snapshot values taken when the event
> +	 * was last scheduled in.
> +	 *
> +	 * we cannot simply called update_context_time()

s/called/call.  I know you just moved the code though. :)

Thanks,
Namhyung


> +	 * because of locking issue as we can be called in
> +	 * NMI context
> +	 */
> +	calc_timer_values(event, &now, &enabled, &running);
> +
> +	userpg = rb->user_page;
> +
>  	++userpg->lock;
>  	barrier();
>  	userpg->index = perf_event_index(event);
> -- 
> 2.49.0.395.g12beb8f557-goog
>