On Wed, Sep 9, 2026 at 8:00 PM Ridong Chen <[email protected]> wrote:
>
>
>
> On 9/9/2026 2:40 PM, Barry Song wrote:
> > On Mon, Sep 7, 2026 at 12:25 PM Ridong Chen <[email protected]> wrote:
> >>
> >> From: Ridong Chen <[email protected]>
> >>
> >> MGLRU's scan_folios() and evict_folios() emit the classic-LRU
> >> tracepoints trace_mm_vmscan_lru_isolate() and
> >> trace_mm_vmscan_lru_shrink_inactive(). Those predate MGLRU and are
> >> indistinguishable from the classic-LRU path: they carry no generation,
> >> sequence, memcg or swappiness context, so a trace of an MGLRU run cannot
> >> tell which memcg a given scan/evict belongs to, nor how far reclaim has
> >> progressed through the generations.
> >>
> >> Add mm_mglru_isolate_folios, emitted once per isolate_folios() call with
> >> the memcg id, the type actually scanned, the effective swappiness, the
> >> scanned/isolated counts, and the anon/file min_seq and max_seq. The
> >> min_seq/max_seq triplet ties each isolation to the generation layout it
> >> ran against, which the classic-LRU tracepoints cannot express.
> >>
> >> This is emitted at the isolate_folios() layer, which is MGLRU-specific
> >> and has no classic-LRU counterpart, so it neither changes nor duplicates
> >> the existing scan/evict tracepoints.
> >>
> >
> > I have no objection to MGLRU having some tracepoints. However, one
> > major concern is that the code is changing rapidly, so the tracepoints
> > may not be stable. We would need to maintain them as the code evolves.
> >
>
> Indeed. GEN-LRU has been available for several years now, and many vendors are
> already using it, so adding tracepoints is a necessary step.
>
> >> Assisted-by: Claude:claude-opus-4-8
> >> Signed-off-by: Ridong Chen <[email protected]>
> >> ---
> >>   include/trace/events/vmscan.h | 47 +++++++++++++++++++++++++++++++++++
> >>   mm/vmscan.c                   |  9 ++++++-
> >>   2 files changed, 55 insertions(+), 1 deletion(-)
> >>
> > [...]
> >> diff --git a/mm/vmscan.c b/mm/vmscan.c
> >> index 8409ea4bbf37..771fe6827939 100644
> >> --- a/mm/vmscan.c
> >> +++ b/mm/vmscan.c
> > [...]
> >> @@ -4876,6 +4877,12 @@ static int isolate_folios(unsigned long nr_to_scan, 
> >> struct lruvec *lruvec,
> >>                  goto retry;
> >>          }
> >>
> >> +done:
> >> +       trace_mm_mglru_isolate_folios(mem_cgroup_id(lruvec_memcg(lruvec)),
> >> +                                     type, swappiness, total_scanned, 
> >> *isolated,
> >> +                                     lrugen->min_seq[LRU_GEN_ANON],
> >> +                                     lrugen->min_seq[LRU_GEN_FILE],
> >> +                                     lrugen->max_seq);
> >
> > Maybe `scan_folios()` would be a better place for this, as it can more
> > directly reflect what we're doing for each type.
> >
>
> scan_folios() already contains the trace_mm_vmscan_lru_isolate
> tracepoint(traditional tracepoint), but it lacks any GEN-LRU-specific
> information. Adding another tracepoint there would be redundant. Therefore, I
> placed the new tracepoint in isolate_folios(), where it can capture GEN-LRU's
> specific details. Additionally, it can indicate whether a fallback to the 
> other
> type has occurred.

if `trace_mm_vmscan_lru_isolate` is located in `scan_folios()`, that
would be strong evidence that `scan_folios()` is a better place for
adding a new tracepoint for LRU gen.

There is no need to avoid this location just because we already have a
tracepoint here. In fact, I think this is evidence that it is a better
place, not the opposite.

BTW, `isolate_folios()` might be deleted soon, depending on how we
improve MGLRU and on some ongoing work from Kairui. But `scan_folios()`
should be long-lived, as we need to scan folios in any case. Therefore,
adding tracepoints to `isolate_folios()` doesn't seem to be related to
the stable ABI concern mentioned in your cover letter.

Best Regards
Barry

Reply via email to