On 4/30/2019 5:08 AM, Peter Zijlstra wrote:
On Mon, Apr 29, 2019 at 04:02:33PM -0700, Ian Rogers wrote:
This is very interesting. How does the code handle cgroup hierarchies?
For example, if we have:

cgroup0 is the cgroup root
cgroup1 whose parent is cgroup0
cgroup2 whose parent is cgroup1

we have task0 running in cgroup0, task1 in cgroup1, task2 in cgroup2
and then a perf command line like:
perf stat -e cycles,cycles,cycles -G cgroup0,cgroup1,cgroup2 --no-merge sleep 10

we expected 3 cycles counts:
  - for cgroup0 including task2, task1 and task0
  - for cgroup1 including task2 and task1
  - for cgroup2 just including task2

It looks as though:
+       if (next && (next->cpu == event->cpu) && (next->cgrp_id ==
event->cgrp_id))

will mean that events will only consider cgroups that directly match
the cgroup of the event. Ie we'd get 3 cycles counts of:
  - for cgroup0 just task0
  - for cgroup1 just task1
  - for cgroup2 just task2
Yeah, I think you're right; the proposed code doesn't capture the
hierarchy thing at all.


The hierarchies is handled in the next patch as below.

But I once thought we only need to handle directly match. So it will be return immediately once a match found.
I believe we can fix it by simply remove the "return 0".

+static int cgroup_visit_groups_merge(struct perf_event_groups *groups, int cpu,
+                                    int (*func)(struct perf_event *, void *, 
int (*)(struct perf_event *)),
+                                    void *data)
+{
+       struct sched_in_data *sid = data;
+       struct cgroup_subsys_state *css;
+       struct perf_cgroup *cgrp;
+       struct perf_event *evt;
+       u64 cgrp_id;
+
+       for (css = &sid->cpuctx->cgrp->css; css; css = css->parent) {
+               /* root cgroup doesn't have events */
+               if (css->id == 1)
+                       return 0;
+
+               cgrp = container_of(css, struct perf_cgroup, css);
+               cgrp_id = *this_cpu_ptr(cgrp->cgrp_id);
+               /* Only visit groups when the cgroup has events */
+               if (cgrp_id) {
+                       evt = perf_event_groups_first_cgroup(groups, cpu, 
cgrp_id);
+                       while (evt) {
+                               if (func(evt, (void *)sid, pmu_filter_match))
+                                       break;
+                               evt = perf_event_groups_next_cgroup(evt);
+                       }
+                       return 0;               <--- need to remove for 
hierarchies
+               }
+       }
+
+       return 0;
+}

Thanks,
Kan



Reply via email to