Skip to contentSkip to conversation

perf/core: Remove group siblings before the leader in move_group

PATCH New Jun Yang 2 participants 2 messages Oct 22 perf events

lore
Share & tools
lore

Jun Yang

author
#0 lore raw

PATCH 1/1 perf/core: Remove group siblings before the leader in move_group

From: Jun Yang <[email protected]>

During group migration, perf_event_open() removes the group leader before its siblings. The context's active lists contain only group leaders, so a concurrent context switch can no longer reach the remaining active siblings after the leader has been removed.

The context can then become inactive while those siblings are still active. Removing a sibling in this state takes the inactive-context path and calls pmu->del() on the calling CPU rather than the CPU where the event is running.

Remove the siblings first so that a concurrent context switch can still schedule them out through the leader.

Fixes bd2756811766 ("perf: Rewrite core context handling")
Assisted-by tencentos-corvus-ai:kimi-k3

---
A KASAN reproducer for this issue is available if requested.

 kernel/events/core.c | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)
kernel/events/core.c +4 -3
index fe33fe1..5bf8053 100644
@@ -14230,14 +14230,15 @@ SYSCALL_DEFINE5(perf_event_open,
 	 */
 
 	if (move_group) {
-		perf_remove_from_context(group_leader, 0);
-		put_pmu_ctx(group_leader->pmu_ctx);
-
+		/* Keep siblings reachable by group_sched_out(). */
 		for_each_sibling_event(sibling, group_leader) {
 			perf_remove_from_context(sibling, 0);
 			put_pmu_ctx(sibling->pmu_ctx);
 		}
 
+		perf_remove_from_context(group_leader, 0);
+		put_pmu_ctx(group_leader->pmu_ctx);
+
 		/*
 		 * Install the group siblings before the group leader.
 		 *
Signature · 1 line
2.54.0 (Apple Git-157)
#1 lore raw
Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider: - [Critical] perf/core: use-after-free of sibling->pmu_ctx in perf_event_open() --

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
- [Critical] perf/core: use-after-free of sibling->pmu_ctx in perf_event_open() --

commit dae4cdd7da7aed3ce65185b981d41f47553d78fb
Author: Jun Yang <[email protected]>

perf/core: Remove group siblings before the leader in move_group

This commit changes the group migration path in perf_event_open() to remove sibling events from the context before removing the group leader, so that a concurrent context switch can still schedule active siblings out through the leader.

Quoted 9 lines
kernel/events/core.c +0 -1
index c180b174bac7c..d16178f3f53e6 100644
@@ -14335,14 +14335,15 @@ SYSCALL_DEFINE5(perf_event_open,
 	 */
 
 	if (move_group) {
-		perf_remove_from_context(group_leader, 0);

kernel/events/core.c SYSCALL_DEFINE5(perf_event_open, · line 14345

Quoted 10 lines
kernel/events/core.c +3 -2
-		put_pmu_ctx(group_leader->pmu_ctx);
-
+		/* Keep siblings reachable by group_sched_out(). */
 		for_each_sibling_event(sibling, group_leader) {
 			perf_remove_from_context(sibling, 0);
 			put_pmu_ctx(sibling->pmu_ctx);
 		}
 
+		perf_remove_from_context(group_leader, 0);
+		put_pmu_ctx(group_leader->pmu_ctx);

[Severity: Critical]
Can calling put_pmu_ctx(sibling->pmu_ctx) before removing group_leader cause a use-after-free of sibling->pmu_ctx during a concurrent context switch?

In a pure software group, a software sibling backed by a different software PMU from group_leader (for example, perf_task_clock or perf_tracepoint paired with a perf_swevent leader) keeps its own PMU because in_software_context(group_leader) is true, and gets a separate perf_event_pmu_context via find_get_pmu_context():

kernel/events/core.c:SYSCALL_DEFINE5(perf_event_open) {
    ...
    if (is_software_event(event) &&
        !in_software_context(group_leader)) {
        pmu = group_leader->pmu_ctx->pmu;
    }
    ...
    pmu_ctx = find_get_pmu_context(pmu, ctx, event);
    ...
    event->pmu_ctx = pmu_ctx;
}

When migrating that group to a hardware context (move_group == 1), perf_remove_from_context(sibling, 0) leaves sibling on group_leader->sibling_list, and put_pmu_ctx(sibling->pmu_ctx) is now called while group_leader is still active in ctx and before sibling->pmu_ctx is reassigned to the new pmu_ctx.

If sibling holds the sole reference to its perf_event_pmu_context, put_pmu_ctx() drops epc->refcount to zero and queues free_epc_rcu() via call_rcu(). Because perf_event_open() runs in preemptible process context holding only pmus_srcu and ctx->mutex, and subsequent perf_remove_from_context() calls invoke event_function_call() -> task_function_call(), an RCU grace period can elapse and free sibling->pmu_ctx before perf_remove_from_context(group_leader, 0) completes.

If a concurrent context switch or reschedule on the target task's CPU invokes group_sched_out() -> event_sched_out() or group_sched_in() -> event_sched_in() for group_leader, both functions walk group_leader->sibling_list and dereference event->pmu_ctx->pmu via this_cpc(epc->pmu) before checking event->state:

kernel/events/core.c:event_sched_out() {
    struct perf_event_pmu_context *epc = event->pmu_ctx;
    struct perf_cpu_pmu_context *cpc = this_cpc(epc->pmu);
    enum perf_event_state state = PERF_EVENT_STATE_INACTIVE;
    ...
    if (event->state != PERF_EVENT_STATE_ACTIVE)
        return;
}
kernel/events/core.c:event_sched_in() {
    struct perf_event_pmu_context *epc = event->pmu_ctx;
    struct perf_cpu_pmu_context *cpc = this_cpc(epc->pmu);
    int ret = 0;
    ...
    if (event->state <= PERF_EVENT_STATE_OFF)
        return 0;
}

Should put_pmu_ctx(sibling->pmu_ctx) be deferred until after group_leader has been removed from ctx?

Quoted 4 lines
kernel/events/core.c +1 -0
+
 		/*
 		 * Install the group siblings before the group leader.
 		 *
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1