On Mon, Jul 11, 2022 at 8:55 PM Yosry Ahmed <yosryahmed@xxxxxxxxxx> wrote:
>
> On Sun, Jul 10, 2022 at 5:51 PM Yonghong Song <yhs@xxxxxx> wrote:
> >
> >
> >
> > On 7/10/22 5:26 PM, Yonghong Song wrote:
> > >
> > >
> > > On 7/8/22 5:04 PM, Yosry Ahmed wrote:
> > >> Add a selftest that tests the whole workflow for collecting,
> > >> aggregating (flushing), and displaying cgroup hierarchical stats.
> > >>
> > >> TL;DR:
> > >> - Userspace program creates a cgroup hierarchy and induces memcg reclaim
> > >> in parts of it.
> > >> - 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.
> > >> - Userspace program makes sure the stats are aggregated and read
> > >> correctly.
> > >>
> > >> Detailed explanation:
> > >> - The test loads tracing bpf programs, vmscan_start and vmscan_end, to
> > >> measure the latency of cgroup reclaim. Per-cgroup readings 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).
> > >>
> > >> - Finally, the test creates a cgroup hierarchy and induces memcg reclaim
> > >> in parts of it, and makes sure that the stats collection, aggregation,
> > >> and reading workflow works as expected.
> > >>
> > >> Signed-off-by: Yosry Ahmed <yosryahmed@xxxxxxxxxx>
> > >> ---
> > >> .../prog_tests/cgroup_hierarchical_stats.c | 362 ++++++++++++++++++
> > >> .../bpf/progs/cgroup_hierarchical_stats.c | 235 ++++++++++++
> > >> 2 files changed, 597 insertions(+)
> > >> create mode 100644
> > >> tools/testing/selftests/bpf/prog_tests/cgroup_hierarchical_stats.c
> > >> create mode 100644
> > >> tools/testing/selftests/bpf/progs/cgroup_hierarchical_stats.c
> > >>
> > > [...]
> > >> +
> > >> +static unsigned long long get_cgroup_vmscan_delay(unsigned long long
> > >> cgroup_id,
> > >> + const char *file_name)
> > >> +{
> > >> + char buf[128], path[128];
> > >> + unsigned long long vmscan = 0, id = 0;
> > >> + int err;
> > >> +
> > >> + /* For every cgroup, read the file generated by cgroup_iter */
> > >> + snprintf(path, 128, "%s%s", BPFFS_VMSCAN, file_name);
> > >> + err = read_from_file(path, buf, 128);
> > >> + if (!ASSERT_OK(err, "read cgroup_iter"))
> > >> + return 0;
> > >> +
> > >> + /* Check the output file formatting */
> > >> + ASSERT_EQ(sscanf(buf, "cg_id: %llu, total_vmscan_delay: %llu\n",
> > >> + &id, &vmscan), 2, "output format");
> > >> +
> > >> + /* Check that the cgroup_id is displayed correctly */
> > >> + ASSERT_EQ(id, cgroup_id, "cgroup_id");
> > >> + /* Check that the vmscan reading is non-zero */
> > >> + ASSERT_GT(vmscan, 0, "vmscan_reading");
> > >> + return vmscan;
> > >> +}
> > >> +
> > >> +static void check_vmscan_stats(void)
> > >> +{
> > >> + int i;
> > >> + unsigned long long vmscan_readings[N_CGROUPS], vmscan_root;
> > >> +
> > >> + for (i = 0; i < N_CGROUPS; i++)
> > >> + vmscan_readings[i] = get_cgroup_vmscan_delay(cgroups[i].id,
> > >> + cgroups[i].name);
> > >> +
> > >> + /* Read stats for root too */
> > >> + vmscan_root = get_cgroup_vmscan_delay(CG_ROOT_ID, CG_ROOT_NAME);
> > >> +
> > >> + /* Check that child1 == child1_1 + child1_2 */
> > >> + ASSERT_EQ(vmscan_readings[1], vmscan_readings[3] +
> > >> vmscan_readings[4],
> > >> + "child1_vmscan");
> > >> + /* Check that child2 == child2_1 + child2_2 */
> > >> + ASSERT_EQ(vmscan_readings[2], vmscan_readings[5] +
> > >> vmscan_readings[6],
> > >> + "child2_vmscan");
> > >> + /* Check that test == child1 + child2 */
> > >> + ASSERT_EQ(vmscan_readings[0], vmscan_readings[1] +
> > >> vmscan_readings[2],
> > >> + "test_vmscan");
> > >> + /* Check that root >= test */
> > >> + ASSERT_GE(vmscan_root, vmscan_readings[1], "root_vmscan");
> > >
> > > I still get a test failure with
> > >
> > > 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 0
> > > != expected -2
> > > check_vmscan_stats:FAIL:child2_vmscan unexpected child2_vmscan: actual 0
> > > != expected -2
> > > check_vmscan_stats:PASS:test_vmscan 0 nsec
> > > check_vmscan_stats:PASS:root_vmscan 0 nsec
> > >
> > > I added 'dump_stack()' in function try_to_free_mem_cgroup_pages()
> > > and run this test (#33) and didn't get any stacktrace.
> > > But I do get stacktraces due to other operations like
> > > try_to_free_mem_cgroup_pages+0x1fd [kernel]
> > > try_to_free_mem_cgroup_pages+0x1fd [kernel]
> > > memory_reclaim_write+0x88 [kernel]
> > > cgroup_file_write+0x88 [kernel]
> > > kernfs_fop_write_iter+0xd0 [kernel]
> > > vfs_write+0x2c4 [kernel]
> > > __x64_sys_write+0x60 [kernel]
> > > do_syscall_64+0x2d [kernel]
> > > entry_SYSCALL_64_after_hwframe+0x44 [kernel]
> > >
> > > If you can show me the stacktrace about how
> > > try_to_free_mem_cgroup_pages() is triggered in your setup, I can
> > > help debug this problem in my environment.
> >
> > BTW, CI also reported the test failure.
> > https://github.com/kernel-patches/bpf/pull/3284
> >
> > For example, with gcc built kernel,
> > https://github.com/kernel-patches/bpf/runs/7272407890?check_suite_focus=true
> >
> > The error:
> >
> > get_cgroup_vmscan_delay:PASS:cgroup_id 0 nsec
> > get_cgroup_vmscan_delay:PASS:vmscan_reading 0 nsec
> > check_vmscan_stats:FAIL:child1_vmscan unexpected child1_vmscan:
> > actual 28390910 != expected 28390909
> > check_vmscan_stats:FAIL:child2_vmscan unexpected child2_vmscan:
> > actual 0 != expected -2
> > check_vmscan_stats:PASS:test_vmscan 0 nsec
> > check_vmscan_stats:PASS:root_vmscan 0 nsec
> >
>
> Hey Yonghong,
>
> Thanks for helping us debug this failure. I can reproduce the CI
> failure in my enviornment, but this failure is actually different from
> the failure in your environment. In your environment it looks like no
> stats are gathered for all cgroups (either no reclaim happening or bpf
> progs not being run). In the CI and in my environment, only one cgroup
> observes this behavior.
>
> The thing is, I was able to reproduce the problem only when I ran all
> test_progs. When I run the selftest alone (test_progs -t
> cgroup_hierarchical_stats), it consistently passes, which is
> interesting.
I think I figured this one out (the CI failure). I set max_entries for
the maps in the test to 10, because I have 1 entry per-cgroup, and I
have less than 10 cgroups. When I run the test with other tests I
*think* there are other cgroups that are being created, so the number
exceeds 10, and some of the entries for the test cgroups cannot be
created. I saw a lot of "failed to create entry for cgroup.." message
in the bpf trace produced by my test, and the error turned out to be
-E2BIG. I increased max_entries to 100 and it seems to be consistently
passing when run with all the other tests, using both test_progs and
test_progs-no_alu32.
Please find a diff attached fixing this problem and a few other nits:
- Return meaningful exit codes from the reclaimer() child process and
check them in induce_vmscan().
- Make buf and path variables static in get_cgroup_vmscan_delay()
- Print error code in bpf trace when we fail to create a bpf map entry.
- Print 0 instead of -1 when we can't find a map entry, to avoid
underflowing the unsigned counters in the test.
Let me know if this diff works or not, and if I need to send a new
version with the diff or not. Also let me know if this fixes the
failures that you have been seeing locally (which looked different
from the CI failures).
Thanks!
>
> Anyway, one failure at a time :) I am working on debugging the CI
> failure (that occurs only when all tests are run), then we'll see if
> fixing that fixes the problem in our environment as well.
>
> If you have any pointers about why a test would consistently pass
> alone and consistently fail with others that would be good. Otherwise,
> I will keep you updated with any findings I reach.
>
> Thanks again!
>
> > >
> > >> +}
> > >> +
> > >> +static int setup_cgroup_iter(struct cgroup_hierarchical_stats *obj,
> > >> int cgroup_fd,
> > > [...]
Attachment:
selftest_fix.patch
Description: Binary data
