Donate to e Foundation | Murena handsets with /e/OS | Own a part of Murena! Learn more

Commit 85dc6002 authored by Peter Zijlstra's avatar Peter Zijlstra Committed by Ingo Molnar
Browse files

perf/x86/amd/ibs: Fix pmu::stop() nesting



Patch 5a50f529 ("perf/x86/ibs: Fix race with IBS_STARTING state")
closed a big hole while opening another, smaller hole.

Signed-off-by: default avatarPeter Zijlstra (Intel) <peterz@infradead.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vince Weaver <vincent.weaver@maine.edu>
Fixes: 5a50f529 ("perf/x86/ibs: Fix race with IBS_STARTING state")
Signed-off-by: default avatarIngo Molnar <mingo@kernel.org>
parent 201c2f85
Loading
Loading
Loading
Loading
+45 −7
Original line number Diff line number Diff line
@@ -28,10 +28,46 @@ static u32 ibs_caps;
#define IBS_FETCH_CONFIG_MASK	(IBS_FETCH_RAND_EN | IBS_FETCH_MAX_CNT)
#define IBS_OP_CONFIG_MASK	IBS_OP_MAX_CNT


/*
 * IBS states:
 *
 * ENABLED; tracks the pmu::add(), pmu::del() state, when set the counter is taken
 * and any further add()s must fail.
 *
 * STARTED/STOPPING/STOPPED; deal with pmu::start(), pmu::stop() state but are
 * complicated by the fact that the IBS hardware can send late NMIs (ie. after
 * we've cleared the EN bit).
 *
 * In order to consume these late NMIs we have the STOPPED state, any NMI that
 * happens after we've cleared the EN state will clear this bit and report the
 * NMI handled (this is fundamentally racy in the face or multiple NMI sources,
 * someone else can consume our BIT and our NMI will go unhandled).
 *
 * And since we cannot set/clear this separate bit together with the EN bit,
 * there are races; if we cleared STARTED early, an NMI could land in
 * between clearing STARTED and clearing the EN bit (in fact multiple NMIs
 * could happen if the period is small enough), and consume our STOPPED bit
 * and trigger streams of unhandled NMIs.
 *
 * If, however, we clear STARTED late, an NMI can hit between clearing the
 * EN bit and clearing STARTED, still see STARTED set and process the event.
 * If this event will have the VALID bit clear, we bail properly, but this
 * is not a given. With VALID set we can end up calling pmu::stop() again
 * (the throttle logic) and trigger the WARNs in there.
 *
 * So what we do is set STOPPING before clearing EN to avoid the pmu::stop()
 * nesting, and clear STARTED late, so that we have a well defined state over
 * the clearing of the EN bit.
 *
 * XXX: we could probably be using !atomic bitops for all this.
 */

enum ibs_states {
	IBS_ENABLED	= 0,
	IBS_STARTED	= 1,
	IBS_STOPPING	= 2,
	IBS_STOPPED	= 3,

	IBS_MAX_STATES,
};
@@ -377,9 +413,8 @@ static void perf_ibs_start(struct perf_event *event, int flags)

	perf_ibs_set_period(perf_ibs, hwc, &period);
	/*
	 * Set STARTED before enabling the hardware, such that
	 * a subsequent NMI must observe it. Then clear STOPPING
	 * such that we don't consume NMIs by accident.
	 * Set STARTED before enabling the hardware, such that a subsequent NMI
	 * must observe it.
	 */
	set_bit(IBS_STARTED,    pcpu->state);
	clear_bit(IBS_STOPPING, pcpu->state);
@@ -396,6 +431,9 @@ static void perf_ibs_stop(struct perf_event *event, int flags)
	u64 config;
	int stopping;

	if (test_and_set_bit(IBS_STOPPING, pcpu->state))
		return;

	stopping = test_bit(IBS_STARTED, pcpu->state);

	if (!stopping && (hwc->state & PERF_HES_UPTODATE))
@@ -405,12 +443,12 @@ static void perf_ibs_stop(struct perf_event *event, int flags)

	if (stopping) {
		/*
		 * Set STOPPING before disabling the hardware, such that it
		 * Set STOPPED before disabling the hardware, such that it
		 * must be visible to NMIs the moment we clear the EN bit,
		 * at which point we can generate an !VALID sample which
		 * we need to consume.
		 */
		set_bit(IBS_STOPPING, pcpu->state);
		set_bit(IBS_STOPPED, pcpu->state);
		perf_ibs_disable_event(perf_ibs, hwc, config);
		/*
		 * Clear STARTED after disabling the hardware; if it were
@@ -556,7 +594,7 @@ static int perf_ibs_handle_irq(struct perf_ibs *perf_ibs, struct pt_regs *iregs)
		 * with samples that even have the valid bit cleared.
		 * Mark all this NMIs as handled.
		 */
		if (test_and_clear_bit(IBS_STOPPING, pcpu->state))
		if (test_and_clear_bit(IBS_STOPPED, pcpu->state))
			return 1;

		return 0;