Messages in this thread | | | Subject | Re: [PATCH v2] tracing: Protect the buffer from recursion in perf | From | Peter Zijlstra <> | Date | Tue, 10 Nov 2009 11:27:42 +0100 |
| |
On Fri, 2009-11-06 at 04:13 +0100, Frederic Weisbecker wrote: > While tracing using events with perf, if one enables the > lockdep:lock_acquire event, it will infect every other perf trace > events. > > Basically, you can enable whatever set of trace events through perf > but if this event is part of the set, the only result we can get is a > long list of lock_acquire events of rcu read lock, and only that. > > This is because of a recursion inside perf. > > 1) When a trace event is triggered, it will fill a per cpu buffer and > submit it to perf. > 2) Perf will commit this event but will also protect some data using > rcu_read_lock > 3) A recursion appears: rcu_read_lock triggers a lock_acquire event that > will fill the per cpu event and then submit the buffer to perf. > 4) Perf detects a recursion and ignore it > 5) Perf continue its work on the previous event, but its buffer has been > overwritten by the lock_acquire event, it has then been turned into > a lock_acquire event of rcu read lock > > Such scenario also happens with lock_release with rcu_read_unlock(). > > We could turn the rcu_read_lock() into __rcu_read_lock() to drop the > lock debugging from perf fast path, but that would make us lose > the rcu debugging and that doesn't prevent from other possible kind of > recursion from perf in the future. > > This patch adds a recursion protection based on a counter on the perf > trace per cpu buffers to solve the problem.
There already is recursion protection in kernel/perf_event.c:perf_swevent_recursion_context() and thereabouts. Could you not fix this by widening its scope?
| |