Re: [PATCH] mm/lruvec: preemptively free dead folios during lru_add drain

["JP Kobryn (Meta)" <jp.kobryn@linux.dev>]

On 4/23/26 4:53 PM, Barry Song wrote:
> On Fri, Apr 24, 2026 at 7:46 AM Shakeel Butt <shakeel.butt@linux.dev> wrote:
>>
>> On Fri, Apr 24, 2026 at 07:22:30AM +0800, Barry Song wrote:
>>> On Fri, Apr 24, 2026 at 12:43 AM JP Kobryn (Meta) <jp.kobryn@linux.dev> wrote:
>>>>
>>>> Of all observable lruvec lock contention in our fleet, we find that ~24%
>>>> occurs when dead folios are present in lru_add batches at drain time. This
>>>> is wasteful in the sense that the folio is added to the LRU just to be
>>>> immediately removed via folios_put_refs(), incurring two unnecessary lock
>>>> acquisitions.
>>>>
>>>> Eliminate this overhead by preemptively cleaning up dead folios before they
>>>> make it into the LRU. Use folio_ref_freeze() to filter folios whose only
>>>> remaining refcount is the batch ref. When dead folios are found, move them
>>>> off the add batch and onto a temporary batch to be freed.
>>>>
>>>> During A/B testing on one of our prod instagram workloads (high-frequency
>>>> short-lived requests), the patch intercepted almost all dead folios before
>>>> they entered the LRU. Data collected using the mm_lru_insertion tracepoint
>>>> shows the effectiveness of the patch:
>>>>
>>>> Per-host LRU add averages at 95% CPU load
>>>> (60 hosts each side, 3 x 60s intervals)
>>>>
>>>>              dead folios/min  total folios/min   dead %
>>>> unpatched:        1,297,785        19,341,986  6.7097%
>>>> patched:                 14        19,039,996  0.0001%
>>>>
>>>> Within this workload, we save ~2.6M lock acquisitions per minute per host
>>>> as a result.
>>>>
>>>> System-wide memory stats improved on the patched side also at 95% CPU load:
>>>>   - direct reclaim scanning reduced 7%
>>>>   - allocation stalls reduced 5.2%
>>>>   - compaction stalls reduced 12.3%
>>>>   - page frees reduced 4.9%
>>>>
>>>> No regressions were observed in requests served per second or request tail
>>>> latency (p99). Both metrics showed directional improvement at higher CPU
>>>> utilization (comparing 85% to 95%).
>>>>
>>>> Signed-off-by: JP Kobryn (Meta) <jp.kobryn@linux.dev>
>>>
>>> Hi JP,
>>> I’m seeing a large number of "BAD page" bugs.
>>> Not sure if it’s related, but reverting this patch
>>> seems to fix the issue.

It seems this was missed since classic LRU was used in testing.

>>>
>>> [ 2869.365978] BUG: Bad page state in process uname  pfn:3a5417
>>> [ 2869.365981] page: refcount:0 mapcount:0 mapping:0000000000000000
>>> index:0x724884c20 pfn:0x3a5417
>>> [ 2869.365983] flags:
>>> 0x17ffffc0020908(uptodate|active|owner_2|swapbacked|node=0|zone=2|lastcpupid=0x1fffff)
>>
>> Hi Barry, are you using MGLRU? It seems like MGLRU set active flag in
>> folio_add_lru().
> 
> Yes. If you are referring to this set_active, I think it is
> incorrect, so I have fixed it here and am waiting for review:
> 
> https://lore.kernel.org/linux-mm/20260418120233.7162-1-baohua@kernel.org/
> 
>>
>> JP, we need to clean active flag but let's check what else can be set before
>> folio_add_lru().

Looks like only active is the problem. If we start manually clearing
flags it starts to feel messy. I get that some fix is needed though. I
don't see this patch in mm-new yet so maybe we can hold off on merging
there to avoid the MGLRU case. But if Barry's patch is accepted, could
we re-apply?

Let me know if you're thinking there are any implications beyond the
active flag.