[PATCH] perf/x86/amd/uncore: Add group validation

[Sandipan Das]

The amd_uncore driver currently does not validate event groups and
allows creation of groups with more events than the number of available
hardware counters. Because of this, pmu->event_init() succeeds but
counter assignment fails later in pmu->add() which returns -EBUSY once
all counters are exhausted.

Address this by introducing group validation in the pmu->event_init()
path. Since the uncore PMUs have no per-event constraints and all
counters of a PMU are interchangeable, validation is reduced to just
counting the group members that target a PMU and ensuring that they fit
within the available set of counters.

Signed-off-by: Sandipan Das <sandipan.das@xxxxxxx>
---
 arch/x86/events/amd/uncore.c | 32 ++++++++++++++++++++++++++++++++
 1 file changed, 32 insertions(+)

diff --git a/arch/x86/events/amd/uncore.c b/arch/x86/events/amd/uncore.c
index dbc00b6dd69e..9715cfb8cce3 100644
--- a/arch/x86/events/amd/uncore.c
+++ b/arch/x86/events/amd/uncore.c
@@ -265,6 +265,29 @@ static void amd_uncore_del(struct perf_event *event, int flags)
 	hwc->idx = -1;
 }
 
+static bool amd_uncore_group_valid(struct perf_event *event)
+{
+	struct amd_uncore_pmu *pmu = event_to_amd_uncore_pmu(event);
+	struct perf_event *leader = event->group_leader;
+	struct perf_event *sibling;
+	int counters = 0;
+
+	/* Ignore software events as they do not occupy hardware counters */
+	if (leader->pmu == event->pmu && !is_software_event(leader))
+		counters++;
+
+	for_each_sibling_event(sibling, leader) {
+		if (sibling->pmu == event->pmu && !is_software_event(sibling))
+			counters++;
+	}
+
+	/*
+	 * When pmu->event_init() is called, the event is yet to be linked to
+	 * its leader's sibling list, so it is counted separately
+	 */
+	return (counters + 1) <= pmu->num_counters;
+}
+
 static int amd_uncore_event_init(struct perf_event *event)
 {
 	struct amd_uncore_pmu *pmu;
@@ -282,6 +305,15 @@ static int amd_uncore_event_init(struct perf_event *event)
 	if (!ctx)
 		return -ENODEV;
 
+	/*
+	 * Ensure that all events in a group can be scheduled together so that
+	 * a failure can be reported at perf_event_open() time rather than
+	 * silently at pmu->add() time when no free counter is found
+	 */
+	if (event->group_leader != event &&
+	    !amd_uncore_group_valid(event))
+		return -EINVAL;
+
 	/*
 	 * NB and Last level cache counters (MSRs) are shared across all cores
 	 * that share the same NB / Last level cache.  On family 16h and below,
-- 
2.53.0