[PATCH] perf/x86: Reject pair-counter groups with no GP budget

[PATCH] perf/x86: Reject pair-counter groups with no GP budget

[Zhao Gongyi]

AMD large-increment events reserve an extra GP counter. A syzkaller
reproducer opens a perf event group until the adjusted GP budget is
exhausted. x86_schedule_events() then WARNs on the unschedulable group.

Treat an exhausted budget as a scheduling failure and return -EINVAL
through the existing cleanup path.

Fixes: 5738891229a2 ("perf/x86/amd: Add support for Large Increment per Cycle Events")
Cc: stable@vger.kernel.org
Signed-off-by: Zhao Gongyi <zhaogongyi@bytedance.com>
---
 arch/x86/events/core.c | 8 +++++---
 1 file changed, 5 insertions(+), 3 deletions(-)

diff --git a/arch/x86/events/core.c b/arch/x86/events/core.c
index 4e344fb1a..20ba7b5b0 100644
--- a/arch/x86/events/core.c
+++ b/arch/x86/events/core.c
@@ -1246,11 +1246,13 @@ int x86_schedule_events(struct cpu_hw_events *cpuc, int n, int *assign)
 		 */
 		if (x86_pmu.flags & PMU_FL_PAIR) {
 			gpmax -= cpuc->n_pair;
-			WARN_ON(gpmax <= 0);
+			if (gpmax <= 0)
+				unsched = 1;
 		}
 
-		unsched = perf_assign_events(cpuc->event_constraint, n, wmin,
-					     wmax, gpmax, assign);
+		if (!unsched)
+			unsched = perf_assign_events(cpuc->event_constraint, n, wmin,
+						     wmax, gpmax, assign);
 	}
 
 	/*

base-commit: 59f1398140869432c7171b9a2a279ca315467c4e
-- 
2.39.5 (Apple Git-154)

Re: [PATCH] perf/x86: Reject pair-counter groups with no GP budget

[Hang Zhengyang]

On Mon, 09 Nov 2026 19:03:56 +0800, Zhao Gongyi wrote:
> AMD large-increment events reserve an extra GP counter. A syzkaller
> reproducer opens a perf event group until the adjusted GP budget is
> exhausted. x86_schedule_events() then WARNs on the unschedulable group.

Tested on AMD Zen 3 (KVM guest, tip perf/urgent): a group of six
fp_ret_sse_avx_ops.all events hits the WARN without this patch and not
with it; smaller groups still schedule. Without your n_pair fix the WARN
fires only once per CPU.

Tested-by: Hang Zhengyang <hang.zhengyang1010@gmail.com>