Re: deadlock in perf_ioctl

[Peter Zijlstra <peterz@infradead.org>]

On Mon, Nov 30, 2015 at 10:20:49AM +0100, Dmitry Vyukov wrote:
> =============================================
> [ INFO: possible recursive locking detected ]
> 4.4.0-rc1+ #129 Not tainted
> ---------------------------------------------
> a.out/6283 is trying to acquire lock:
>  (&ctx->lock){-.....}, at: [<ffffffff815072ce>]
> __perf_event_period+0x8e/0x4b0 kernel/events/core.c:4156
> 
> but task is already holding lock:
>  (&ctx->lock){-.....}, at: [<     inline     >] perf_event_period
> kernel/events/core.c:4212
>  (&ctx->lock){-.....}, at: [<     inline     >] _perf_ioctl
> kernel/events/core.c:4266
>  (&ctx->lock){-.....}, at: [<ffffffff8152331c>] perf_ioctl+0x7bc/0xcc0
> kernel/events/core.c:4320
> 

Indeed so. I suppose the below should fix this, I'll go try in a bit.

---
Subject: perf: Fix PERF_EVENT_IOC_PERIOD deadlock

Dmitry reported a fairly silly recursive lock deadlock for
PERF_EVENT_IOC_PERIOD, fix this by explicitly doing the inactive part of
__perf_event_period() instead of calling that function.

Reported-by: Dmitry Vyukov <dvyukov@google.com>
Fixes: c7999c6f3fed ("perf: Fix PERF_EVENT_IOC_PERIOD migration race")
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
---
 kernel/events/core.c | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)

diff --git a/kernel/events/core.c b/kernel/events/core.c
index 36babfd20648..a991e12e4d0d 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -4216,7 +4216,14 @@ static int perf_event_period(struct perf_event *event, u64 __user *arg)
 		goto retry;
 	}
 
-	__perf_event_period(&pe);
+	if (event->attr.freq) {
+		event->attr.sample_freq = value;
+	} else {
+		event->attr.sample_period = value;
+		event->hw.sample_period = value;
+	}
+
+	local64_set(&event->hw.period_left, 0);
 	raw_spin_unlock_irq(&ctx->lock);
 
 	return 0;