perf/cgroup: Grow per perf_cpu_context heap storage

Allow the per-CPU min heap storage to have sufficient space for per-cgroup
iterators.

Based-on-work-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Link: https://lkml.kernel.org/r/20200214075133.181299-6-irogers@google.com

diff --git a/kernel/events/core.c b/kernel/events/core.c
index 7529e76a2e36a6a66a474f7f22b7cdcb233232b8..8065949a865e349f3077c6a030f65b286611983a 100644 (file)
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -892,6 +892,47 @@ static inline void perf_cgroup_sched_in(struct task_struct *prev,
        rcu_read_unlock();
 }
 
+static int perf_cgroup_ensure_storage(struct perf_event *event,
+                               struct cgroup_subsys_state *css)
+{
+       struct perf_cpu_context *cpuctx;
+       struct perf_event **storage;
+       int cpu, heap_size, ret = 0;
+
+       /*
+        * Allow storage to have sufficent space for an iterator for each
+        * possibly nested cgroup plus an iterator for events with no cgroup.
+        */
+       for (heap_size = 1; css; css = css->parent)
+               heap_size++;
+
+       for_each_possible_cpu(cpu) {
+               cpuctx = per_cpu_ptr(event->pmu->pmu_cpu_context, cpu);
+               if (heap_size <= cpuctx->heap_size)
+                       continue;
+
+               storage = kmalloc_node(heap_size * sizeof(struct perf_event *),
+                                      GFP_KERNEL, cpu_to_node(cpu));
+               if (!storage) {
+                       ret = -ENOMEM;
+                       break;
+               }
+
+               raw_spin_lock_irq(&cpuctx->ctx.lock);
+               if (cpuctx->heap_size < heap_size) {
+                       swap(cpuctx->heap, storage);
+                       if (storage == cpuctx->heap_default)
+                               storage = NULL;
+                       cpuctx->heap_size = heap_size;
+               }
+               raw_spin_unlock_irq(&cpuctx->ctx.lock);
+
+               kfree(storage);
+       }
+
+       return ret;
+}
+
 static inline int perf_cgroup_connect(int fd, struct perf_event *event,
                                      struct perf_event_attr *attr,
                                      struct perf_event *group_leader)
@@ -911,6 +952,10 @@ static inline int perf_cgroup_connect(int fd, struct perf_event *event,
                goto out;
        }
 
+       ret = perf_cgroup_ensure_storage(event, css);
+       if (ret)
+               goto out;
+
        cgrp = container_of(css, struct perf_cgroup, css);
        event->cgrp = cgrp;
 
@@ -3440,6 +3485,8 @@ static noinline int visit_groups_merge(struct perf_cpu_context *cpuctx,
                        .nr = 0,
                        .size = cpuctx->heap_size,
                };
+
+               lockdep_assert_held(&cpuctx->ctx.lock);
        } else {
                event_heap = (struct min_heap){
                        .data = itrs,