drivers/clocksource/arm_arch_timer: Tighten Allwinner arch timer workaround

As we know, the Allwinner A64 SoC has a buggy RCU time unit. The workaround named UNKNOWN1 was not sufficient to cover some more buggy bunches of this SoC. This workaround diminish the mask to 8 bits instead of 9.

An example run of timer test tool https://github.com/smaeul/timer-tools
on PinePhone device (owns the A64 SoC) gives following result on a non patched kernel (cut off):

Running parallel counter test...
0: Failed after 5507 reads (0.003578 s)
0: 0x0000000c8272cbf1 -> 0x0000000c8272ccff -> 0x0000000c8272cc0e ( 0.011 ms)
2: Failed after 14518 reads (0.009248 s)
2: 0x0000000c827513f1 -> 0x0000000c82751300 -> 0x0000000c8275140e ( -0.010 ms)
3: Failed after 14112 reads (0.008730 s)
3: 0x0000000c8274f3f2 -> 0x0000000c8274f300 -> 0x0000000c8274f40d ( -0.010 ms)
1: Failed after 12030 reads (0.008409 s)

1: 0x0000000c8274abf1 -> 0x0000000c8274acff -> 0x0000000c8274ac0f (     0.011 ms)
1: 0x0000000c827759f2 -> 0x0000000c82775aff -> 0x0000000c82775a0e (     0.011 ms)
0: 0x0000000c8277a9f2 -> 0x0000000c8277aaff -> 0x0000000c8277aa0d (     0.011 ms)

2: 0x0000000c8278f3f1 -> 0x0000000c8278f300 -> 0x0000000c8278f40e ( -0.010 ms)
0: 0x0000000c82785ff2 -> 0x0000000c82784300 -> 0x0000000c8278600d ( -0.309 ms)

After the proposed patch applied the test runs correctly (~2 hours of testing with a tool above without fails)

---
 drivers/clocksource/arm_arch_timer.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

index a7ff77550e17..3019faa263f5 100644
@@ -371,7 +371,7 @@ static u64 notrace arm64_858921_read_cntvct_el0(void)
 	do {								\
 		_val = read_sysreg(reg);				\
 		_retries--;						\
-	} while (((_val + 1) & GENMASK(8, 0)) <= 1 && _retries);	\
+	} while (((_val + 1) & GENMASK(7, 0)) <= 1 && _retries);	\
 									\
 	WARN_ON_ONCE(!_retries);					\
 	_val;								\

2.25.1

_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

What is RCU?

2 hours seems like an incredibly small amount of time given that the existing workaround was believed to be correct. Run it continuously for a couple of weeks on several different machines with varying workloads and less us know the outcome.

Thanks,

	M.

Without deviation from the norm, progress is not possible.

El Thu, 10 Nov 2022 08:31:21 +0000
Marc Zyngier <[email protected]> escribió:

MZ> >
MZ> > As we know, the Allwinner A64 SoC has a buggy RCU time unit. The MZ>
MZ> What is RCU?

I think I called it wrong. Anyway I mean the HR timer of A64.

MZ>
MZ> > workaround named UNKNOWN1 was not sufficient to cover some more buggy MZ> > bunches of this SoC. This workaround diminish the mask to 8 bits instead MZ> > of 9.
MZ> >
MZ> > An example run of timer test tool https://github.com/smaeul/timer-tools
MZ> > on PinePhone device (owns the A64 SoC) gives following result on a non MZ> > patched kernel (cut off):
MZ> >
MZ> > Running parallel counter test...
MZ> > 0: Failed after 5507 reads (0.003578 s)
MZ> > 0: 0x0000000c8272cbf1 -> 0x0000000c8272ccff -> 0x0000000c8272cc0e ( 0.011 ms)

[......]

MZ> > After the proposed patch applied the test runs
MZ> > correctly (~2 hours of testing with a tool above without fails) MZ>
MZ> 2 hours seems like an incredibly small amount of time given that the MZ> existing workaround was believed to be correct. Run it continuously MZ> for a couple of weeks on several different machines with varying MZ> workloads and less us know the outcome.

The only A64 machine I own is the Pinephone. First time I did this patch ~9 month ago (on behalf os Samuel). Before it the system suffered hangs every 15-20 minutes and backward time jumps ~1 time a day. Since applying of this patch none of above occurred.
To be honest, I never did long tests (weeks). I will put the device on a probe for some weeks and let you know then.

Best regards,
Ilya

запрягай медленно,
едь быстро!