[PATCH] perf/x86/amd: Avoid double-decrementing n_pair on failed transactions

[PATCH] perf/x86/amd: Avoid double-decrementing n_pair on failed transactions

[Zhao Gongyi]

On a failed PERF_PMU_TXN_ADD commit, x86_schedule_events() releases
constraints for newly added pair events. The AMD put callback decrements
n_pair, and x86_pmu_cancel_txn() then subtracts n_txn_pair for the same
events. This undercounts active pairs and overstates the GP counter budget
for later scheduling.

Leave transaction rollback to x86_pmu_cancel_txn(). Keep the AMD
decrement for constraint releases outside transactions.

Fixes: 871a93b0aad6 ("perf/x86: Fix n_pair for cancelled txn")
Signed-off-by: Zhao Gongyi <zhaogongyi@bytedance.com>
---
 arch/x86/events/amd/core.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/arch/x86/events/amd/core.c b/arch/x86/events/amd/core.c
index 49b6b8fce..3363cfd3b 100644
--- a/arch/x86/events/amd/core.c
+++ b/arch/x86/events/amd/core.c
@@ -1251,7 +1251,8 @@ static void amd_put_event_constraints_f17h(struct cpu_hw_events *cpuc,
 {
 	struct hw_perf_event *hwc = &event->hw;
 
-	if (is_counter_pair(hwc))
+	if (is_counter_pair(hwc) &&
+	    !(cpuc->txn_flags & PERF_PMU_TXN_ADD))
 		--cpuc->n_pair;
 }
 

base-commit: f5a78ecccd98a04c47cea1cd63c2a2fdce8a4b48
-- 
2.39.5 (Apple Git-154)

Re: [PATCH] perf/x86/amd: Avoid double-decrementing n_pair on failed transactions

[Hang Zhengyang]

On Mon,  9 Nov 2026 19:19:09 +0800, Zhao Gongyi wrote:
> On a failed PERF_PMU_TXN_ADD commit, x86_schedule_events() releases
> constraints for newly added pair events. The AMD put callback decrements
> n_pair, and x86_pmu_cancel_txn() then subtracts n_txn_pair for the same
> events.

Reproduced in a KVM guest on a Zen 3 host (Ryzen 7 PRO 5850U), kernel
7.3-rc6, with a trace_printk() in x86_schedule_events() and
x86_pmu_cancel_txn().  With four pinned cycles events on CPU 0, each
failed attempt to add a pinned '{fp_ret_sse_avx_ops.all,cycles}' group
leaves n_pair one lower: -1, -2, -3, so gpmax goes 5, 6, 7 on 6
counters.  With this patch n_pair is back to 0 after each attempt and
gpmax stays 5.

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

To reproduce:

  perf stat -C 0 -e '{cycles,cycles,cycles,cycles}:D' -- sleep 4 &
  # three times:
  perf stat -C 0 -e '{fp_ret_sse_avx_ops.all,cycles}:D' -- taskset -c 0 <workload>

(Analysis and testing done with an AI assistant; I reproduced the
n_pair drift.)