Re: [PATCH] mm/lruvec: trace LRU add drains and drain-all queuing

[JP Kobryn <jp.kobryn@linux.dev>]

On 6/9/26 5:07 PM, JP Kobryn wrote:
> On 6/9/26 12:44 AM, Barry Song wrote:
>> On Tue, Jun 9, 2026 at 12:12 PM JP Kobryn <jp.kobryn@linux.dev> wrote:
>>>
>>> LRU add batches can be drained before they reach capacity. This can be a
>>> source of LRU lock contention, but it is not currently possible to
>>> attribute these drains to callers with existing tracepoints.
>>>
>>> Add mm_lru_add_drain to report the CPU and lru_add batch count when an
>>> lru_add batch is drained. This allows tracing to distinguish full drains
>>> from partial drains and attribute them to the calling stack.
>>>
>>> Add mm_lru_drain_all_queue to report when lru_add_drain_all() queues
>>> per-CPU drain work. This captures the requester stack and target CPU for
>>> remote drain work. The event is named as a drain-all queue event because
>>> the queued work can be needed for batches other than lru_add.
>>>
>>> Signed-off-by: JP Kobryn <jp.kobryn@linux.dev>
>>> ---
>>>  include/trace/events/pagemap.h | 40 ++++++++++++++++++++++++++++++++++
>>>  mm/swap.c                      |  6 ++++-
>>>  2 files changed, 45 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/include/trace/events/pagemap.h b/include/trace/events/pagemap.h
>>> index 171524d3526d..ea8fc46bedb0 100644
>>> --- a/include/trace/events/pagemap.h
>>> +++ b/include/trace/events/pagemap.h
>>> @@ -77,6 +77,46 @@ TRACE_EVENT(mm_lru_activate,
>>>         TP_printk("folio=%p pfn=0x%lx", __entry->folio, __entry->pfn)
>>>  );
>>>
>>> +TRACE_EVENT(mm_lru_add_drain,
>>> +
>>> +       TP_PROTO(int cpu, unsigned int nr),
>>> +
>>> +       TP_ARGS(cpu, nr),
>>> +
>>> +       TP_STRUCT__entry(
>>> +               __field(int,            cpu     )
>>> +               __field(unsigned int,   nr      )
>>> +       ),
>>> +
>>> +       TP_fast_assign(
>>> +               __entry->cpu    = cpu;
>>> +               __entry->nr     = nr;
>>> +       ),
>>> +
>>> +       TP_printk("cpu=%d nr=%u", __entry->cpu, __entry->nr)
>>> +);
>>> +
>>> +TRACE_EVENT(mm_lru_drain_all_queue,
>>> +
>>> +       TP_PROTO(int target_cpu, bool force_all_cpus),
>>> +
>>> +       TP_ARGS(target_cpu, force_all_cpus),
>>> +
>>> +       TP_STRUCT__entry(
>>> +               __field(int,    target_cpu      )
>>> +               __field(bool,   force_all_cpus  )
>>> +       ),
>>> +
>>> +       TP_fast_assign(
>>> +               __entry->target_cpu     = target_cpu;
>>> +               __entry->force_all_cpus = force_all_cpus;
>>> +       ),
>>> +
>>> +       TP_printk("target_cpu=%d force_all_cpus=%s",
>>> +               __entry->target_cpu,
>>> +               __entry->force_all_cpus ? "true" : "false")
>>> +);
>>> +
>>>  #endif /* _TRACE_PAGEMAP_H */
>>>
>>>  /* This part must be outside protection */
>>> diff --git a/mm/swap.c b/mm/swap.c
>>> index 588f50d8f1a8..c385b93582eb 100644
>>> --- a/mm/swap.c
>>> +++ b/mm/swap.c
>>> @@ -694,9 +694,12 @@ void lru_add_drain_cpu(int cpu)
>>>  {
>>>         struct cpu_fbatches *fbatches = &per_cpu(cpu_fbatches, cpu);
>>>         struct folio_batch *fbatch = &fbatches->lru_add;
>>> +       unsigned int nr_folios_add = folio_batch_count(fbatch);
>>>
>>> -       if (folio_batch_count(fbatch))
>>> +       if (nr_folios_add) {
>>>                 folio_batch_move_lru(fbatch, lru_add);
>>> +               trace_mm_lru_add_drain(cpu, nr_folios_add);
>>> +       }
>>>
>>>         fbatch = &fbatches->lru_move_tail;
>>>         /* Disabling interrupts below acts as a compiler barrier. */
>>> @@ -928,6 +931,7 @@ static inline void __lru_add_drain_all(bool force_all_cpus)
>>>                 if (cpu_needs_drain(cpu)) {
>>>                         INIT_WORK(work, lru_add_drain_per_cpu);
>>>                         queue_work_on(cpu, mm_percpu_wq, work);
>>> +                       trace_mm_lru_drain_all_queue(cpu, force_all_cpus);
>>
>> Do you need tracing on each CPU individually, or is tracing the
>> entire __lru_add_drain_all() invocation sufficient?
> 
> I think the latter would be fine. The remote work will invoke the
> mm_lru_add_drain tracepoint, which will show up as kworker stacks. Since
> the event already has the CPU, we could see where queued drains actually
> ran.

Actually if it's just a single invocation and the only event data is the
force flag, a tracepoint may not even be needed. Other probes can be
installed on function invocation and read the single argument. I can
drop this from v2 and keep the single mm_lru_add_drain tracepoint.