[PATCH v2 1/2] perf/arm-cmn: Don't schedule events on a CPU which no longer owns the PMU

[Haris Okanovic]

perf_event_open() can latch cmn->cpu and then install the event after a
migration has already moved the PMU. The driver has no locking -- it
relies on all events living in one CPU's context -- so such an event
races the owning CPU and can corrupt the shared DTC and DTM state,
giving wrong counts. arm_cmn_event_add() now rejects an event which is
not on the owning CPU.

arm_cmn_migrate() now claims cmn->cpu before migrating, so that events
it reinstalls via perf_pmu_migrate_context() still pass that check.

Signed-off-by: Haris Okanovic <harisokn@xxxxxxxxxx>
---
 drivers/perf/arm-cmn.c | 9 ++++++---
 1 file changed, 6 insertions(+), 3 deletions(-)

diff --git a/drivers/perf/arm-cmn.c b/drivers/perf/arm-cmn.c
index 5378fba916cf5..a5593e5821925 100644
--- a/drivers/perf/arm-cmn.c
+++ b/drivers/perf/arm-cmn.c
@@ -2112,6 +2112,9 @@ static int arm_cmn_event_add(struct perf_event *event, int flags)
 	enum cmn_node_type type = CMN_EVENT_TYPE(event);
 	unsigned int input_sel, i = 0;
 
+	if (cmn->cpu != smp_processor_id())
+		return -ENOENT;
+
 	if (type == CMN_TYPE_DTC) {
 		while (cmn->dtc[i].cycles)
 			if (++i == cmn->num_dtcs)
@@ -2245,12 +2248,12 @@ static int arm_cmn_commit_txn(struct pmu *pmu)
 
 static void arm_cmn_migrate(struct arm_cmn *cmn, unsigned int cpu)
 {
-	unsigned int i;
+	unsigned int i, old = cmn->cpu;
 
-	perf_pmu_migrate_context(&cmn->pmu, cmn->cpu, cpu);
+	cmn->cpu = cpu;
+	perf_pmu_migrate_context(&cmn->pmu, old, cpu);
 	for (i = 0; i < cmn->num_dtcs; i++)
 		irq_set_affinity(cmn->dtc[i].irq, cpumask_of(cpu));
-	cmn->cpu = cpu;
 }
 
 static int arm_cmn_pmu_online_cpu(unsigned int cpu, struct hlist_node *cpuhp_node)
-- 
Haris Okanovic
AWS Graviton