perf: core: Avoid race condition when releasing perf-events

The function, perf_event_release_kernel(), that's used to free the
perf events is dependent on the CPU associated to this event be
online. The function checks at the beginning: if the CPU is offline,
put it in a zombie list and return immediately. Else, proceed and
make a cross-cpu call (from perf_remove_from_context()) to complete
the functionality.

However, there's a potential chance of a race if the CPU went
offline between the initial check and the cross-cpu call. The
cross-cpu call deletes the event from the context's list, but
if the CPU is offline, this deletion doesn't happen. Later
the event is freed irrespective of this failure and the event
still exists in the list. Now, when the list is traversed, it
would try to access the memory which is freed, resulting in a
memory abort.

As a result, before calling perf_event_release_kernel(), capture
the perf's pmus_mutex lock to prevent the CPU from going offline
during the operation.

Change-Id: I20241639ea9a8dc87e5a88cf81e940b3d6cb773c
Signed-off-by: Raghavendra Rao Ananta <rananta@codeaurora.org>