Re: [PATCH v3] perf: Allocate non-contiguous AUX pages by default

[Anshuman Khandual <anshuman.khandual@arm.com>]

On 5/2/25 23:00, Yabin Cui wrote:
> On Fri, May 2, 2025 at 3:51 AM Anshuman Khandual
> <anshuman.khandual@arm.com> wrote:
>>
>> On 5/2/25 01:05, 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 be memory-friendly by default,
>>> by allocating non-contiguous AUX pages. For PMUs requiring contiguous
>>> pages (Intel BTS and some Intel PT), the existing
>>> PERF_PMU_CAP_AUX_NO_SG capability can be used. For PMUs that don't
>>> require but can benefit from contiguous pages (some Intel PT), a new
>>> capability, PERF_PMU_CAP_AUX_PREFER_LARGE, is added to maintain their
>>> existing behavior.
>>>
>>> Signed-off-by: Yabin Cui <yabinc@google.com>
>>> ---
>>> Changes since v2:
>>> Let NO_SG imply PREFER_LARGE. So PMUs don't need to set both flags.
>>> Then the only place needing PREFER_LARGE is intel/pt.c.
>>>
>>> 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/pt.c  |  2 ++
>>>  include/linux/perf_event.h  |  1 +
>>>  kernel/events/ring_buffer.c | 20 +++++++++++++-------
>>>  3 files changed, 16 insertions(+), 7 deletions(-)
>>>
>>> diff --git a/arch/x86/events/intel/pt.c b/arch/x86/events/intel/pt.c
>>> index fa37565f6418..25ead919fc48 100644
>>> --- a/arch/x86/events/intel/pt.c
>>> +++ b/arch/x86/events/intel/pt.c
>>> @@ -1863,6 +1863,8 @@ static __init int pt_init(void)
>>>
>>>       if (!intel_pt_validate_hw_cap(PT_CAP_topa_multiple_entries))
>>>               pt_pmu.pmu.capabilities = PERF_PMU_CAP_AUX_NO_SG;
>>> +     else
>>> +             pt_pmu.pmu.capabilities = PERF_PMU_CAP_AUX_PREFER_LARGE;
>>>
>>>       pt_pmu.pmu.capabilities         |= PERF_PMU_CAP_EXCLUSIVE |
>>>                                          PERF_PMU_CAP_ITRACE |
>>
>> Why this PMU has PERF_PMU_CAP_AUX_PREFER_LARGE fallback option but
>> not the other PMU in arch/x86/events/intel/bts.c even though both
>> had PERF_PMU_CAP_AUX_NO_SG ?
> 
> Because Intel BTS always use NO_SG, while in some cases Intel PT
> doesn't use NO_SG.

Makes sense.

>>
>>> 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..4d2f1c95673e 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;
>>
>> 0 order is now the default allocation granularity. This might benefit
>> from a comment above explaining that max_order could change only with
>> PERF_PMU_CAP_AUX_NO_SG or PERF_PMU_CAP_AUX_PREFER_LARGE PMU flags etc.
>>
> Will add the comment in the next respin.
>>>
>>>       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 +698,22 @@ 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 need or 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_NO_SG |
>>> +                                             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 need or 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_NO_SG |
>>> +                                             PERF_PMU_CAP_AUX_PREFER_LARGE))
>>> +                     max_order = ilog2(nr_pages);
>>>               watermark = 0;
>>>       }
>>>
>>
>> Although not really sure, could event->pmu->capabilities check against the ORed
>> PMU flags PERF_PMU_CAP_AUX_NO_SG and PERF_PMU_CAP_AUX_PREFER_LARGE be contained
>> in a helper pmu_prefers_cont_alloc(struct *pmu ...) or something similar ?
> 
> Sure, but I feel it's not very worthwhile. Maybe add a local variable
> use_contiguous_pages? It can also work as another comment near
> max_order.

Probably that will be better.