perf/core: Detach 'struct perf_cpu_pmu_context' and 'struct pmu' lifetimes

In prepration for being able to unregister a PMU with existing events,
it becomes important to detach struct perf_cpu_pmu_context lifetimes
from that of struct pmu.

Notably struct perf_cpu_pmu_context embeds a struct perf_event_pmu_context
that can stay referenced until the last event goes.

Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Ravi Bangoria <ravi.bangoria@amd.com>
Link: https://lore.kernel.org/r/20241104135518.760214287@infradead.org

Author: Peter Zijlstra

include/linux/perf_event.h

@@ -343,7 +343,7 @@ struct pmu {
 	 */
 	unsigned int			scope;
 
-	struct perf_cpu_pmu_context __percpu *cpu_pmu_context;
+	struct perf_cpu_pmu_context __percpu **cpu_pmu_context;
 	atomic_t			exclusive_cnt; /* < 0: cpu; > 0: tsk */
 	int				task_ctx_nr;
 	int				hrtimer_interval_ms;
@@ -922,7 +922,7 @@ struct perf_event_pmu_context {
 	struct list_head		pinned_active;
 	struct list_head		flexible_active;
 
-	/* Used to avoid freeing per-cpu perf_event_pmu_context */
+	/* Used to identify the per-cpu perf_event_pmu_context */
 	unsigned int			embedded : 1;
 
 	unsigned int			nr_events;

kernel/events/core.c

@@ -1219,7 +1219,7 @@ static int perf_mux_hrtimer_restart_ipi(void *arg)
 
 static __always_inline struct perf_cpu_pmu_context *this_cpc(struct pmu *pmu)
 {
-	return this_cpu_ptr(pmu->cpu_pmu_context);
+	return *this_cpu_ptr(pmu->cpu_pmu_context);
 }
 
 void perf_pmu_disable(struct pmu *pmu)
@@ -5007,11 +5007,14 @@ find_get_pmu_context(struct pmu *pmu, struct perf_event_context *ctx,
 		 */
 		struct perf_cpu_pmu_context *cpc;
 
-		cpc = per_cpu_ptr(pmu->cpu_pmu_context, event->cpu);
+		cpc = *per_cpu_ptr(pmu->cpu_pmu_context, event->cpu);
 		epc = &cpc->epc;
 		raw_spin_lock_irq(&ctx->lock);
 		if (!epc->ctx) {
-			atomic_set(&epc->refcount, 1);
+			/*
+			 * One extra reference for the pmu; see perf_pmu_free().
+			 */
+			atomic_set(&epc->refcount, 2);
 			epc->embedded = 1;
 			list_add(&epc->pmu_ctx_entry, &ctx->pmu_ctx_list);
 			epc->ctx = ctx;
@@ -5087,6 +5090,15 @@ static void get_pmu_ctx(struct perf_event_pmu_context *epc)
 	WARN_ON_ONCE(!atomic_inc_not_zero(&epc->refcount));
 }
 
+static void free_cpc_rcu(struct rcu_head *head)
+{
+	struct perf_cpu_pmu_context *cpc =
+		container_of(head, typeof(*cpc), epc.rcu_head);
+
+	kfree(cpc->epc.task_ctx_data);
+	kfree(cpc);
+}
+
 static void free_epc_rcu(struct rcu_head *head)
 {
 	struct perf_event_pmu_context *epc = container_of(head, typeof(*epc), rcu_head);
@@ -5121,8 +5133,10 @@ static void put_pmu_ctx(struct perf_event_pmu_context *epc)
 
 	raw_spin_unlock_irqrestore(&ctx->lock, flags);
 
-	if (epc->embedded)
+	if (epc->embedded) {
+		call_rcu(&epc->rcu_head, free_cpc_rcu);
 		return;
+	}
 
 	call_rcu(&epc->rcu_head, free_epc_rcu);
 }
@@ -11752,7 +11766,7 @@ perf_event_mux_interval_ms_store(struct device *dev,
 	cpus_read_lock();
 	for_each_online_cpu(cpu) {
 		struct perf_cpu_pmu_context *cpc;
-		cpc = per_cpu_ptr(pmu->cpu_pmu_context, cpu);
+		cpc = *per_cpu_ptr(pmu->cpu_pmu_context, cpu);
 		cpc->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer);
 
 		cpu_function_call(cpu, perf_mux_hrtimer_restart_ipi, cpc);
@@ -11925,7 +11939,25 @@ static void perf_pmu_free(struct pmu *pmu)
 		device_del(pmu->dev);
 		put_device(pmu->dev);
 	}
-	free_percpu(pmu->cpu_pmu_context);
+
+	if (pmu->cpu_pmu_context) {
+		int cpu;
+
+		for_each_possible_cpu(cpu) {
+			struct perf_cpu_pmu_context *cpc;
+
+			cpc = *per_cpu_ptr(pmu->cpu_pmu_context, cpu);
+			if (!cpc)
+				continue;
+			if (cpc->epc.embedded) {
+				/* refcount managed */
+				put_pmu_ctx(&cpc->epc);
+				continue;
+			}
+			kfree(cpc);
+		}
+		free_percpu(pmu->cpu_pmu_context);
+	}
 }
 
 DEFINE_FREE(pmu_unregister, struct pmu *, if (_T) perf_pmu_free(_T))
@@ -11964,14 +11996,20 @@ int perf_pmu_register(struct pmu *_pmu, const char *name, int type)
 			return ret;
 	}
 
-	pmu->cpu_pmu_context = alloc_percpu(struct perf_cpu_pmu_context);
+	pmu->cpu_pmu_context = alloc_percpu(struct perf_cpu_pmu_context *);
 	if (!pmu->cpu_pmu_context)
 		return -ENOMEM;
 
 	for_each_possible_cpu(cpu) {
-		struct perf_cpu_pmu_context *cpc;
+		struct perf_cpu_pmu_context *cpc =
+			kmalloc_node(sizeof(struct perf_cpu_pmu_context),
+				     GFP_KERNEL | __GFP_ZERO,
+				     cpu_to_node(cpu));
+
+		if (!cpc)
+			return -ENOMEM;
 
-		cpc = per_cpu_ptr(pmu->cpu_pmu_context, cpu);
+		*per_cpu_ptr(pmu->cpu_pmu_context, cpu) = cpc;
 		__perf_init_event_pmu_context(&cpc->epc, pmu);
 		__perf_mux_hrtimer_init(cpc, cpu);
 	}