Re: [PATCH v2] mm: fix the inaccurate memory statistics issue for users

[Baolin Wang <baolin.wang@linux.alibaba.com>]

On 2025/6/9 15:35, Michal Hocko wrote:
> On Mon 09-06-25 10:57:41, Ritesh Harjani wrote:
>> Baolin Wang <baolin.wang@linux.alibaba.com> writes:
>>
>>> On some large machines with a high number of CPUs running a 64K pagesize
>>> kernel, we found that the 'RES' field is always 0 displayed by the top
>>> command for some processes, which will cause a lot of confusion for users.
>>>
>>>      PID USER      PR  NI    VIRT    RES    SHR S  %CPU  %MEM     TIME+ COMMAND
>>>   875525 root      20   0   12480      0      0 R   0.3   0.0   0:00.08 top
>>>        1 root      20   0  172800      0      0 S   0.0   0.0   0:04.52 systemd
>>>
>>> The main reason is that the batch size of the percpu counter is quite large
>>> on these machines, caching a significant percpu value, since converting mm's
>>> rss stats into percpu_counter by commit f1a7941243c1 ("mm: convert mm's rss
>>> stats into percpu_counter"). Intuitively, the batch number should be optimized,
>>> but on some paths, performance may take precedence over statistical accuracy.
>>> Therefore, introducing a new interface to add the percpu statistical count
>>> and display it to users, which can remove the confusion. In addition, this
>>> change is not expected to be on a performance-critical path, so the modification
>>> should be acceptable.
>>>
>>> In addition, the 'mm->rss_stat' is updated by using add_mm_counter() and
>>> dec/inc_mm_counter(), which are all wrappers around percpu_counter_add_batch().
>>> In percpu_counter_add_batch(), there is percpu batch caching to avoid 'fbc->lock'
>>> contention. This patch changes task_mem() and task_statm() to get the accurate
>>> mm counters under the 'fbc->lock', but this should not exacerbate kernel
>>> 'mm->rss_stat' lock contention due to the percpu batch caching of the mm
>>> counters. The following test also confirm the theoretical analysis.
>>>
>>> I run the stress-ng that stresses anon page faults in 32 threads on my 32 cores
>>> machine, while simultaneously running a script that starts 32 threads to
>>> busy-loop pread each stress-ng thread's /proc/pid/status interface. From the
>>> following data, I did not observe any obvious impact of this patch on the
>>> stress-ng tests.
>>>
>>> w/o patch:
>>> stress-ng: info:  [6848]          4,399,219,085,152 CPU Cycles          67.327 B/sec
>>> stress-ng: info:  [6848]          1,616,524,844,832 Instructions          24.740 B/sec (0.367 instr. per cycle)
>>> stress-ng: info:  [6848]          39,529,792 Page Faults Total           0.605 M/sec
>>> stress-ng: info:  [6848]          39,529,792 Page Faults Minor           0.605 M/sec
>>>
>>> w/patch:
>>> stress-ng: info:  [2485]          4,462,440,381,856 CPU Cycles          68.382 B/sec
>>> stress-ng: info:  [2485]          1,615,101,503,296 Instructions          24.750 B/sec (0.362 instr. per cycle)
>>> stress-ng: info:  [2485]          39,439,232 Page Faults Total           0.604 M/sec
>>> stress-ng: info:  [2485]          39,439,232 Page Faults Minor           0.604 M/sec
>>>
>>> Tested-by Donet Tom <donettom@linux.ibm.com>
>>> Reviewed-by: Aboorva Devarajan <aboorvad@linux.ibm.com>
>>> Tested-by: Aboorva Devarajan <aboorvad@linux.ibm.com>
>>> Acked-by: Shakeel Butt <shakeel.butt@linux.dev>
>>> Acked-by: SeongJae Park <sj@kernel.org>
>>> Acked-by: Michal Hocko <mhocko@suse.com>
>>> Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
>>> ---
>>> Changes from v1:
>>>   - Update the commit message to add some measurements.
>>>   - Add acked tag from Michal. Thanks.
>>>   - Drop the Fixes tag.
>>
>> Any reason why we dropped the Fixes tag? I see there were a series of
>> discussion on v1 and it got concluded that the fix was correct, then why
>> drop the fixes tag?
> 
> This seems more like an improvement than a bug fix.

Yes. I don't have a strong opinion on this, but we (Alibaba) will 
backport it manually, because some of user-space monitoring tools depend 
on these statistics.