This seems to work, in contrast to everything I tried yesterday. Not
exactly a thing of beauty, but at least it's an option, so...
perf: fix perf sched record forkbomb deadlock
perf sched record can deadlock a box should the holder of handle->data->lock
take an interrupt, and then attempt to acquire an rq lock held by a CPU trying
to acquire the same lock. Disable interrupts.
Signed-off-by: Mike Galbraith <efault@gmx.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Reported-by: Li Zefan <lizf@cn.fujitsu.com>
LKML-Reference: <new-submission>
diff --git a/kernel/perf_event.c b/kernel/perf_event.c
index 574ee58..2ba2e9f 100644
--- a/kernel/perf_event.c
+++ b/kernel/perf_event.c
@@ -3378,15 +3378,23 @@ static void perf_event_task_output(struct perf_event *event,
struct perf_task_event *task_event)
{
struct perf_output_handle handle;
- int size;
struct task_struct *task = task_event->task;
- int ret;
+ unsigned long flags;
+ int size, ret;
+
+ /*
+ * If this CPU attempts to acquire an rq lock held by a CPU spinning
+ * in perf_output_lock() from interrupt context, it's game over.
+ */
+ local_irq_save(flags);
size = task_event->event_id.header.size;
ret = perf_output_begin(&handle, event, size, 0, 0);
- if (ret)
+ if (ret) {
+ local_irq_restore(flags);
return;
+ }
task_event->event_id.pid = perf_event_pid(event, task);
task_event->event_id.ppid = perf_event_pid(event, current);
@@ -3397,6 +3405,7 @@ static void perf_event_task_output(struct perf_event *event,
perf_output_put(&handle, task_event->event_id);
perf_output_end(&handle);
+ local_irq_restore(flags);
}
static int perf_event_task_match(struct perf_event *event)
--
