Re: [PATCH bpf-next v2 8/8] bpf: add a selftest for cgroup hierarchical stats collection

[Yonghong Song <yhs@fb.com>]

On 6/28/22 12:14 AM, Yosry Ahmed wrote:
> On Mon, Jun 27, 2022 at 11:47 PM Yosry Ahmed <yosryahmed@google.com> wrote:
>>
>> On Mon, Jun 27, 2022 at 11:14 PM Yonghong Song <yhs@fb.com> wrote:
>>>
>>>
>>>
>>> On 6/10/22 12:44 PM, Yosry Ahmed wrote:
>>>> Add a selftest that tests the whole workflow for collecting,
>>>> aggregating (flushing), and displaying cgroup hierarchical stats.
>>>>
>>>> TL;DR:
>>>> - Whenever reclaim happens, vmscan_start and vmscan_end update
>>>>     per-cgroup percpu readings, and tell rstat which (cgroup, cpu) pairs
>>>>     have updates.
>>>> - When userspace tries to read the stats, vmscan_dump calls rstat to flush
>>>>     the stats, and outputs the stats in text format to userspace (similar
>>>>     to cgroupfs stats).
>>>> - rstat calls vmscan_flush once for every (cgroup, cpu) pair that has
>>>>     updates, vmscan_flush aggregates cpu readings and propagates updates
>>>>     to parents.
>>>>
>>>> Detailed explanation:
>>>> - The test loads tracing bpf programs, vmscan_start and vmscan_end, to
>>>>     measure the latency of cgroup reclaim. Per-cgroup ratings are stored in
>>>>     percpu maps for efficiency. When a cgroup reading is updated on a cpu,
>>>>     cgroup_rstat_updated(cgroup, cpu) is called to add the cgroup to the
>>>>     rstat updated tree on that cpu.
>>>>
>>>> - A cgroup_iter program, vmscan_dump, is loaded and pinned to a file, for
>>>>     each cgroup. Reading this file invokes the program, which calls
>>>>     cgroup_rstat_flush(cgroup) to ask rstat to propagate the updates for all
>>>>     cpus and cgroups that have updates in this cgroup's subtree. Afterwards,
>>>>     the stats are exposed to the user. vmscan_dump returns 1 to terminate
>>>>     iteration early, so that we only expose stats for one cgroup per read.
>>>>
>>>> - An ftrace program, vmscan_flush, is also loaded and attached to
>>>>     bpf_rstat_flush. When rstat flushing is ongoing, vmscan_flush is invoked
>>>>     once for each (cgroup, cpu) pair that has updates. cgroups are popped
>>>>     from the rstat tree in a bottom-up fashion, so calls will always be
>>>>     made for cgroups that have updates before their parents. The program
>>>>     aggregates percpu readings to a total per-cgroup reading, and also
>>>>     propagates them to the parent cgroup. After rstat flushing is over, all
>>>>     cgroups will have correct updated hierarchical readings (including all
>>>>     cpus and all their descendants).
>>>>
>>>> Signed-off-by: Yosry Ahmed <yosryahmed@google.com>
>>>
>>> There are a selftest failure with test:
>>>
>>> get_cgroup_vmscan_delay:PASS:output format 0 nsec
>>> get_cgroup_vmscan_delay:PASS:cgroup_id 0 nsec
>>> get_cgroup_vmscan_delay:PASS:vmscan_reading 0 nsec
>>> get_cgroup_vmscan_delay:PASS:read cgroup_iter 0 nsec
>>> get_cgroup_vmscan_delay:PASS:output format 0 nsec
>>> get_cgroup_vmscan_delay:PASS:cgroup_id 0 nsec
>>> get_cgroup_vmscan_delay:FAIL:vmscan_reading unexpected vmscan_reading:
>>> actual 0 <= expected 0
>>> check_vmscan_stats:FAIL:child1_vmscan unexpected child1_vmscan: actual
>>> 781874 != expected 382092
>>> check_vmscan_stats:FAIL:child2_vmscan unexpected child2_vmscan: actual
>>> -1 != expected -2
>>> check_vmscan_stats:FAIL:test_vmscan unexpected test_vmscan: actual
>>> 781874 != expected 781873
>>> check_vmscan_stats:FAIL:root_vmscan unexpected root_vmscan: actual 0 <
>>> expected 781874
>>> destroy_progs:PASS:remove cgroup_iter pin 0 nsec
>>> destroy_progs:PASS:remove cgroup_iter pin 0 nsec
>>> destroy_progs:PASS:remove cgroup_iter pin 0 nsec
>>> destroy_progs:PASS:remove cgroup_iter pin 0 nsec
>>> destroy_progs:PASS:remove cgroup_iter pin 0 nsec
>>> destroy_progs:PASS:remove cgroup_iter pin 0 nsec
>>> destroy_progs:PASS:remove cgroup_iter pin 0 nsec
>>> destroy_progs:PASS:remove cgroup_iter root pin 0 nsec
>>> cleanup_bpffs:PASS:rmdir /sys/fs/bpf/vmscan/ 0 nsec
>>> #33      cgroup_hierarchical_stats:FAIL
>>>
>>
>> The test is passing on my setup. I am trying to figure out if there is
>> something outside the setup done by the test that can cause the test
>> to fail.
>>
> 
> I can't reproduce the failure on my machine. It seems like for some
> reason reclaim is not invoked in one of the test cgroups which results
> in the expected stats not being there. I have a few suspicions as to
> what might cause this but I am not sure.
> 
> If you have the capacity, do you mind re-running the test with the
> attached diff1.patch? (and maybe diff2.patch if that fails, this will
> cause OOMs in the test cgroup, you might see some process killed
> warnings).

The patch doesn't help. Still failed.

get_cgroup_vmscan_delay:PASS:cgroup_id 0 nsec
get_cgroup_vmscan_delay:FAIL:vmscan_reading unexpected vmscan_reading: 
actual 0 <= expected 0
check_vmscan_stats:FAIL:child1_vmscan unexpected child1_vmscan: actual 
676612 != expected 339142
check_vmscan_stats:FAIL:child2_vmscan unexpected child2_vmscan: actual 
-1 != expected -2
check_vmscan_stats:FAIL:test_vmscan unexpected test_vmscan: actual 
676612 != expected 676611
check_vmscan_stats:FAIL:root_vmscan unexpected root_vmscan: actual 0 < 
expected 676612
destroy_progs:PASS:remove cgroup_iter pin 0 nsec
destroy_progs:PASS:remove cgroup_iter pin 0 nsec


> Thanks!
> 
> 
>>>
>>> Also an existing test also failed.
[...]