From: Rabin Vincent <[email protected]>
mipsxx_pmu_handle_shared_irq() calls irq_work_run() while holding the
pmuint_rwlock for read. irq_work_run() can, via perf_pending_event(),
call try_to_wake_up() which can try to take rq->lock.
However, perf can also call perf_pmu_enable() (and thus take the
pmuint_rwlock for write) while holding the rq->lock, from
finish_task_switch() via perf_event_context_sched_in().
This leads to an ABBA deadlock:
PID: 3855 TASK: 8f7ce288 CPU: 2 COMMAND: "process"
#0 [89c39ac8] __delay at 803b5be4
#1 [89c39ac8] do_raw_spin_lock at 8008fdcc
#2 [89c39af8] try_to_wake_up at 8006e47c
#3 [89c39b38] pollwake at 8018eab0
#4 [89c39b68] __wake_up_common at 800879f4
#5 [89c39b98] __wake_up at 800880e4
#6 [89c39bc8] perf_event_wakeup at 8012109c
#7 [89c39be8] perf_pending_event at 80121184
#8 [89c39c08] irq_work_run_list at 801151f0
#9 [89c39c38] irq_work_run at 80115274
#10 [89c39c50] mipsxx_pmu_handle_shared_irq at 8002cc7c
PID: 1481 TASK: 8eaac6a8 CPU: 3 COMMAND: "process"
#0 [8de7f900] do_raw_write_lock at 800900e0
#1 [8de7f918] perf_event_context_sched_in at 80122310
#2 [8de7f938] __perf_event_task_sched_in at 80122608
#3 [8de7f958] finish_task_switch at 8006b8a4
#4 [8de7f998] __schedule at 805e4dc4
#5 [8de7f9f8] schedule at 805e5558
#6 [8de7fa10] schedule_hrtimeout_range_clock at 805e9984
#7 [8de7fa70] poll_schedule_timeout at 8018e8f8
#8 [8de7fa88] do_select at 8018f338
#9 [8de7fd88] core_sys_select at 8018f5cc
#10 [8de7fee0] sys_select at 8018f854
#11 [8de7ff28] syscall_common at 80028fc8
The lock seems to be there to protect the hardware counters so there is
no need to hold it across irq_work_run().
Signed-off-by: Rabin Vincent <[email protected]>
---
arch/mips/kernel/perf_event_mipsxx.c | 9 +++++----
1 file changed, 5 insertions(+), 4 deletions(-)
diff --git a/arch/mips/kernel/perf_event_mipsxx.c b/arch/mips/kernel/perf_event_mipsxx.c
index 8c35b31..9452b02 100644
--- a/arch/mips/kernel/perf_event_mipsxx.c
+++ b/arch/mips/kernel/perf_event_mipsxx.c
@@ -1446,6 +1446,11 @@ static int mipsxx_pmu_handle_shared_irq(void)
HANDLE_COUNTER(0)
}
+#ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS
+ read_unlock(&pmuint_rwlock);
+#endif
+ resume_local_counters();
+
/*
* Do all the work for the pending perf events. We can do this
* in here because the performance counter interrupt is a regular
@@ -1454,10 +1459,6 @@ static int mipsxx_pmu_handle_shared_irq(void)
if (handled == IRQ_HANDLED)
irq_work_run();
-#ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS
- read_unlock(&pmuint_rwlock);
-#endif
- resume_local_counters();
return handled;
}
--
2.7.0
On Wed, Apr 05, 2017 at 03:14:08PM +0200, Rabin Vincent wrote:
That lock is disgusting... but yes patch looks about right.
I'll leave it to the MIPS people though.
> ---
> arch/mips/kernel/perf_event_mipsxx.c | 9 +++++----
> 1 file changed, 5 insertions(+), 4 deletions(-)
>
> diff --git a/arch/mips/kernel/perf_event_mipsxx.c b/arch/mips/kernel/perf_event_mipsxx.c
> index 8c35b31..9452b02 100644
> --- a/arch/mips/kernel/perf_event_mipsxx.c
> +++ b/arch/mips/kernel/perf_event_mipsxx.c
> @@ -1446,6 +1446,11 @@ static int mipsxx_pmu_handle_shared_irq(void)
> HANDLE_COUNTER(0)
> }
>
> +#ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS
> + read_unlock(&pmuint_rwlock);
> +#endif
> + resume_local_counters();
> +
> /*
> * Do all the work for the pending perf events. We can do this
> * in here because the performance counter interrupt is a regular
> @@ -1454,10 +1459,6 @@ static int mipsxx_pmu_handle_shared_irq(void)
> if (handled == IRQ_HANDLED)
> irq_work_run();
>
> -#ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS
> - read_unlock(&pmuint_rwlock);
> -#endif
> - resume_local_counters();
> return handled;
> }
>
> --
> 2.7.0
>