Re: [PATCH v2] perf: Allocate non-contiguous AUX pages by default

[James Clark <james.clark@linaro.org>]

On 29/04/2025 10:31 pm, Yabin Cui wrote:
> perf always allocates contiguous AUX pages based on aux_watermark.
> However, this contiguous allocation doesn't benefit all PMUs. For
> instance, ARM SPE and TRBE operate with virtual pages, and Coresight
> ETR allocates a separate buffer. For these PMUs, allocating contiguous
> AUX pages unnecessarily exacerbates memory fragmentation. This
> fragmentation can prevent their use on long-running devices.
> 
> This patch modifies the perf driver to allocate non-contiguous AUX
> pages by default. For PMUs that can benefit from contiguous pages (
> Intel PT and BTS), a new PMU capability, PERF_PMU_CAP_AUX_PREFER_LARGE
> is introduced to maintain their existing behavior.
> 
> Signed-off-by: Yabin Cui <yabinc@google.com>
> ---
> Changes since v1:
> In v1, default is preferring contiguous pages, and add a flag to
> allocate non-contiguous pages. In v2, default is allocating
> non-contiguous pages, and add a flag to prefer contiguous pages.
> 
> v1 patchset:
> perf,coresight: Reduce fragmentation with non-contiguous AUX pages for
> cs_etm
> 
>   arch/x86/events/intel/bts.c |  3 ++-
>   arch/x86/events/intel/pt.c  |  3 ++-
>   include/linux/perf_event.h  |  1 +
>   kernel/events/ring_buffer.c | 18 +++++++++++-------
>   4 files changed, 16 insertions(+), 9 deletions(-)
> 
> diff --git a/arch/x86/events/intel/bts.c b/arch/x86/events/intel/bts.c
> index a95e6c91c4d7..9129f00e4b9f 100644
> --- a/arch/x86/events/intel/bts.c
> +++ b/arch/x86/events/intel/bts.c
> @@ -625,7 +625,8 @@ static __init int bts_init(void)
>   		return -ENOMEM;
>   
>   	bts_pmu.capabilities	= PERF_PMU_CAP_AUX_NO_SG | PERF_PMU_CAP_ITRACE |
> -				  PERF_PMU_CAP_EXCLUSIVE;
> +				  PERF_PMU_CAP_EXCLUSIVE |
> +				  PERF_PMU_CAP_AUX_PREFER_LARGE;
>   	bts_pmu.task_ctx_nr	= perf_sw_context;
>   	bts_pmu.event_init	= bts_event_init;
>   	bts_pmu.add		= bts_event_add;
> diff --git a/arch/x86/events/intel/pt.c b/arch/x86/events/intel/pt.c
> index fa37565f6418..37179e813b8c 100644
> --- a/arch/x86/events/intel/pt.c
> +++ b/arch/x86/events/intel/pt.c
> @@ -1866,7 +1866,8 @@ static __init int pt_init(void)
>   
>   	pt_pmu.pmu.capabilities		|= PERF_PMU_CAP_EXCLUSIVE |
>   					   PERF_PMU_CAP_ITRACE |
> -					   PERF_PMU_CAP_AUX_PAUSE;
> +					   PERF_PMU_CAP_AUX_PAUSE |
> +					   PERF_PMU_CAP_AUX_PREFER_LARGE;
>   	pt_pmu.pmu.attr_groups		 = pt_attr_groups;
>   	pt_pmu.pmu.task_ctx_nr		 = perf_sw_context;
>   	pt_pmu.pmu.event_init		 = pt_event_init;
> diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h
> index 0069ba6866a4..56d77348c511 100644
> --- a/include/linux/perf_event.h
> +++ b/include/linux/perf_event.h
> @@ -301,6 +301,7 @@ struct perf_event_pmu_context;
>   #define PERF_PMU_CAP_AUX_OUTPUT			0x0080
>   #define PERF_PMU_CAP_EXTENDED_HW_TYPE		0x0100
>   #define PERF_PMU_CAP_AUX_PAUSE			0x0200
> +#define PERF_PMU_CAP_AUX_PREFER_LARGE		0x0400
>   
>   /**
>    * pmu::scope
> diff --git a/kernel/events/ring_buffer.c b/kernel/events/ring_buffer.c
> index 5130b119d0ae..d76249ce4f17 100644
> --- a/kernel/events/ring_buffer.c
> +++ b/kernel/events/ring_buffer.c
> @@ -679,7 +679,7 @@ int rb_alloc_aux(struct perf_buffer *rb, struct perf_event *event,
>   {
>   	bool overwrite = !(flags & RING_BUFFER_WRITABLE);
>   	int node = (event->cpu == -1) ? -1 : cpu_to_node(event->cpu);
> -	int ret = -ENOMEM, max_order;
> +	int ret = -ENOMEM, max_order = 0;
>   
>   	if (!has_aux(event))
>   		return -EOPNOTSUPP;
> @@ -689,8 +689,8 @@ int rb_alloc_aux(struct perf_buffer *rb, struct perf_event *event,
>   
>   	if (!overwrite) {
>   		/*
> -		 * Watermark defaults to half the buffer, and so does the
> -		 * max_order, to aid PMU drivers in double buffering.
> +		 * Watermark defaults to half the buffer, to aid PMU drivers
> +		 * in double buffering.
>   		 */
>   		if (!watermark)
>   			watermark = min_t(unsigned long,
> @@ -698,16 +698,20 @@ int rb_alloc_aux(struct perf_buffer *rb, struct perf_event *event,
>   					  (unsigned long)nr_pages << (PAGE_SHIFT - 1));
>   
>   		/*
> -		 * Use aux_watermark as the basis for chunking to
> +		 * For PMUs that prefer large contiguous buffers,
> +		 * use aux_watermark as the basis for chunking to
>   		 * help PMU drivers honor the watermark.
>   		 */
> -		max_order = get_order(watermark);
> +		if (event->pmu->capabilities & PERF_PMU_CAP_AUX_PREFER_LARGE)
> +			max_order = get_order(watermark);
>   	} else {
>   		/*
> -		 * We need to start with the max_order that fits in nr_pages,
> +		 * For PMUs that prefer large contiguous buffers,
> +		 * we need to start with the max_order that fits in nr_pages,
>   		 * not the other way around, hence ilog2() and not get_order.
>   		 */
> -		max_order = ilog2(nr_pages);
> +		if (event->pmu->capabilities & PERF_PMU_CAP_AUX_PREFER_LARGE)
> +			max_order = ilog2(nr_pages);

Doesn't this one need to be 'PERF_PMU_CAP_AUX_PREFER_LARGE | 
PERF_PMU_CAP_AUX_NO_SG', otherwise the NO_SG test further down doesn't 
work for devices that only have NO_SG and not PREFER_LARGE.

NO_SG implies PREFER_LARGE behavior, except that NO_SG additionally hard 
fails if it can't do it in one alloc. But I think you shouldn't have to 
set them both to get the correct behavior.