Re: perf: perf_fuzzer quickly locks up on 4.15-rc7

[Peter Zijlstra]

On Tue, Jan 09, 2018 at 04:12:53PM +0100, Peter Zijlstra wrote:

> In any case, I found yet another lockdep splat, trying to figure out wth
> to do about that.

An of course, testing this one yields yet another lockdep splat..
onwards to #3 :/

---
Subject: perf: Fix another perf,trace,cpuhp lock inversion
From: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
Date: Tue Jan 9 17:07:59 CET 2018

Lockdep complained about:

        perf_trace_init()
#0        mutex_lock(&event_mutex)
          perf_trace_event_init()
            perf_trace_event_reg()
              tp_event->class->reg() := tracepoint_probe_register
#1              mutex_lock(&tracepoints_mutex)
                  trace_point_add_func()
#2                  static_key_enable()



#2	do_cpu_up()
	  perf_event_init_cpu()
#3	    mutex_lock(&pmus_lock)
#4	    mutex_lock(&ctx->mutex)


	perf_ioctl()
#4	  ctx = perf_event_ctx_lock()
	  _perf_iotcl()
	    ftrace_profile_set_filter()
#0	      mutex_lock(&event_mutex)


Fudge it for now by noting that the tracepoint state does not depend
on the event <-> context relation. Ugly though :/

Cc: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
Cc: Steven Rostedt <rostedt@xxxxxxxxxxx>
Signed-off-by: Peter Zijlstra (Intel) <peterz@xxxxxxxxxxxxx>
---
 kernel/events/core.c |   26 ++++++++++++++++++++++++--
 1 file changed, 24 insertions(+), 2 deletions(-)

--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -8525,6 +8525,29 @@ perf_event_set_addr_filter(struct perf_e
 	return ret;
 }
 
+static int
+perf_tracepoint_set_filter(struct perf_event *event, char *filter_str)
+{
+	struct perf_event_context *ctx = event->ctx;
+	int ret;
+
+	/*
+	 * Beware, here be dragons!!
+	 *
+	 * the tracepoint muck will deadlock against ctx->mutex, but the tracepoint
+	 * stuff does not actually need it. So temporarily drop ctx->mutex. As per
+	 * perf_event_ctx_lock() we already have a reference on ctx.
+	 *
+	 * This can result in event getting moved to a different ctx, but that
+	 * does not affect the tracepoint state.
+	 */
+	mutex_unlock(&ctx->mutex);
+	ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
+	mutex_lock(&ctx->mutex);
+
+	return ret;
+}
+
 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
 {
 	char *filter_str;
@@ -8541,8 +8564,7 @@ static int perf_event_set_filter(struct
 
 	if (IS_ENABLED(CONFIG_EVENT_TRACING) &&
 	    event->attr.type == PERF_TYPE_TRACEPOINT)
-		ret = ftrace_profile_set_filter(event, event->attr.config,
-						filter_str);
+		ret = perf_tracepoint_set_filter(event, filter_str);
 	else if (has_addr_filter(event))
 		ret = perf_event_set_addr_filter(event, filter_str);