perf: RISC-V: Remove PERF_HES_STOPPED flag checking in riscv_pmu_start()

Commit Message

Eric Lin July 10, 2023, 3:43 p.m. UTC

Since commit 096b52fd2bb4 ("perf: RISC-V: throttle perf events") the
perf_sample_event_took() function was added to report time spent in
overflow interrupts. If the interrupt takes too long, the perf framework
will lower the sysctl_perf_event_sample_rate and max_samples_per_tick.
When hwc->interrupts is larger than max_samples_per_tick, the
hwc->interrupts will be set to MAX_INTERRUPTS, and events will be
throttled within the __perf_event_account_interrupt() function.

However, the RISC-V PMU driver doesn't call riscv_pmu_stop() to update the
PERF_HES_STOPPED flag after perf_event_overflow() in pmu_sbi_ovf_handler()
function to avoid throttling. When the perf framework unthrottled the event
in the timer interrupt handler, it triggers riscv_pmu_start() function
and causes a WARN_ON_ONCE() warning, as shown below:

 ------------[ cut here ]------------
 WARNING: CPU: 0 PID: 240 at drivers/perf/riscv_pmu.c:184 riscv_pmu_start+0x7c/0x8e
 Modules linked in:
 CPU: 0 PID: 240 Comm: ls Not tainted 6.4-rc4-g19d0788e9ef2 #1
 Hardware name: SiFive (DT)
 epc : riscv_pmu_start+0x7c/0x8e
  ra : riscv_pmu_start+0x28/0x8e
 epc : ffffffff80aef864 ra : ffffffff80aef810 sp : ffff8f80004db6f0
  gp : ffffffff81c83750 tp : ffffaf80069f9bc0 t0 : ffff8f80004db6c0
  t1 : 0000000000000000 t2 : 000000000000001f s0 : ffff8f80004db720
  s1 : ffffaf8008ca1068 a0 : 0000ffffffffffff a1 : 0000000000000000
  a2 : 0000000000000001 a3 : 0000000000000870 a4 : 0000000000000000
  a5 : 0000000000000000 a6 : 0000000000000840 a7 : 0000000000000030
  s2 : 0000000000000000 s3 : ffffaf8005165800 s4 : ffffaf800424da00
  s5 : ffffffffffffffff s6 : ffffffff81cc7590 s7 : 0000000000000000
  s8 : 0000000000000006 s9 : 0000000000000001 s10: ffffaf807efbc340
  s11: ffffaf807efbbf00 t3 : ffffaf8006a16028 t4 : 00000000dbfbb796
  t5 : 0000000700000000 t6 : ffffaf8005269870
 status: 0000000200000100 badaddr: 0000000000000000 cause: 0000000000000003
 [<ffffffff80aef864>] riscv_pmu_start+0x7c/0x8e
 [<ffffffff80185b56>] perf_adjust_freq_unthr_context+0x15e/0x174
 [<ffffffff80188642>] perf_event_task_tick+0x88/0x9c
 [<ffffffff800626a8>] scheduler_tick+0xfe/0x27c
 [<ffffffff800b5640>] update_process_times+0x9a/0xba
 [<ffffffff800c5bd4>] tick_sched_handle+0x32/0x66
 [<ffffffff800c5e0c>] tick_sched_timer+0x64/0xb0
 [<ffffffff800b5e50>] __hrtimer_run_queues+0x156/0x2f4
 [<ffffffff800b6bdc>] hrtimer_interrupt+0xe2/0x1fe
 [<ffffffff80acc9e8>] riscv_timer_interrupt+0x38/0x42
 [<ffffffff80090a16>] handle_percpu_devid_irq+0x90/0x1d2
 [<ffffffff8008a9f4>] generic_handle_domain_irq+0x28/0x36

After referring other PMU drivers like Arm, Loongarch, Csky, and Mips,
they don't call *_pmu_stop() to update with PERF_HES_STOPPED flag
after perf_event_overflow() function nor do they add PERF_HES_STOPPED
flag checking in *_pmu_start() which don't cause this warning.

Thus, it's recommended to remove this unnecessary check in
riscv_pmu_start() function to prevent this warning.

Signed-off-by: Eric Lin <[e-mail]>
---
 drivers/perf/riscv_pmu.c | 3 ---
 1 file changed, 3 deletions(-)

Comments

Palmer Dabbelt July 12, 2023, 3:52 p.m. UTC | #1

On Mon, 10 Jul 2023 15:43:28 +0000, Eric Lin wrote:
> Since commit 096b52fd2bb4 ("perf: RISC-V: throttle perf events") the
> perf_sample_event_took() function was added to report time spent in
> overflow interrupts. If the interrupt takes too long, the perf framework
> will lower the sysctl_perf_event_sample_rate and max_samples_per_tick.
> When hwc->interrupts is larger than max_samples_per_tick, the
> hwc->interrupts will be set to MAX_INTERRUPTS, and events will be
> throttled within the __perf_event_account_interrupt() function.
> 
> [...]

Applied, thanks!

[1/1] perf: RISC-V: Remove PERF_HES_STOPPED flag checking in riscv_pmu_start()
      https://git.kernel.org/palmer/c/66843b14fb71

Best regards,

[e-mail] July 12, 2023, 4 p.m. UTC | #2

Hello:

This patch was applied to riscv/linux.git (fixes)
by Palmer Dabbelt <[e-mail]>:

On Mon, 10 Jul 2023 15:43:28 +0000 you wrote:
> Since commit 096b52fd2bb4 ("perf: RISC-V: throttle perf events") the
> perf_sample_event_took() function was added to report time spent in
> overflow interrupts. If the interrupt takes too long, the perf framework
> will lower the sysctl_perf_event_sample_rate and max_samples_per_tick.
> When hwc->interrupts is larger than max_samples_per_tick, the
> hwc->interrupts will be set to MAX_INTERRUPTS, and events will be
> throttled within the __perf_event_account_interrupt() function.
> 
> [...]

Here is the summary with links:
  - perf: RISC-V: Remove PERF_HES_STOPPED flag checking in riscv_pmu_start()
    https://git.kernel.org/riscv/c/66843b14fb71

You are awesome, thank you!

Patch

diff --git a/drivers/perf/riscv_pmu.c b/drivers/perf/riscv_pmu.c
index ebca5eab9c9b..56897d4d4fd3 100644
--- a/drivers/perf/riscv_pmu.c
+++ b/drivers/perf/riscv_pmu.c
@@ -181,9 +181,6 @@  void riscv_pmu_start(struct perf_event *event, int flags)
 	uint64_t max_period = riscv_pmu_ctr_get_width_mask(event);
 	u64 init_val;
 
-	if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
-		return;
-
 	if (flags & PERF_EF_RELOAD)
 		WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));