ERRATUM_858921 is broken on 5.15 kernel

[Yogesh Lal]

On 1/5/2023 7:42 PM, Mark Rutland wrote:
> On Thu, Jan 05, 2023 at 07:03:48PM +0530, Yogesh Lal wrote:
>> Hi,
>>
>> We are observing issue on A73 core where ERRATUM_858921 is broken.
> Do you *only* see this issue on v5.15.y, or is mainline (e.g. v6.2-rc2) also
> broken?

Checked the code path and looks like its broken on mainline also.

> I don't see any fix that fits your exact description below, but I do see that
> we've made a bunch of changes in this area since.
>
>> On 5.15 kernel arch_timer_enable_workaround is set by reading
>> arm64_858921_read_cntpct_el0 and arm64_858921_read_cntvct_el0 during timer
>> register using following path.
>>
>> arch_timer_enable_workaround->atomic_set(&timer_unstable_counter_workaround_in_use,
>> 1);
>>
>> [code snap]
>> 564 static
>> 565 void arch_timer_enable_workaround(const struct
>> arch_timer_erratum_workaround *wa,
>> 566                               bool local)
>> 567 {
>> 568     int i;
>> 569
>> 570     if (local) {
>> 571 __this_cpu_write(timer_unstable_counter_workaround, wa);
>> 572     } else {
>> 573             for_each_possible_cpu(i)
>> 574                     per_cpu(timer_unstable_counter_workaround, i) = wa;
>> 575     }
>> 576
>> 577     if (wa->read_cntvct_el0 || wa->read_cntpct_el0)
>> 578 atomic_set(&timer_unstable_counter_workaround_in_use, 1);
>>
>>
>> and based on above workaround enablement , appropriate function to get
>> counter is used.
>>
>> 1008 static void __init arch_counter_register(unsigned type)
>> 1009 {
>> 1010     u64 start_count;
>> 1011
>> 1012     /* Register the CP15 based counter if we have one */
>> 1013     if (type & ARCH_TIMER_TYPE_CP15) {
>> 1014         u64 (*rd)(void);
>> 1015
>> 1016         if ((IS_ENABLED(CONFIG_ARM64) && !is_hyp_mode_available()) ||
>> 1017             arch_timer_uses_ppi == ARCH_TIMER_VIRT_PPI) {
>> 1018             if (arch_timer_counter_has_wa())
>> 1019                 rd = arch_counter_get_cntvct_stable;
>> 1020             else
>> 1021                 rd = arch_counter_get_cntvct;
>> 1022         } else {
>> 1023             if (arch_timer_counter_has_wa())
>> 1024                 rd = arch_counter_get_cntpct_stable;
>> 1025             else
>> 1026                 rd = arch_counter_get_cntpct;
>> 1027         }
>> [snap]
>> 1043     /* 56 bits minimum, so we assume worst case rollover */
>> 1044     sched_clock_register(arch_timer_read_counter, 56, arch_timer_rate);
>>
>>
>> As our boot cores are not impacted by errata sched_clock_register() will
>> register !arch_timer_counter_has_wa() callback.
> It would be helpful to mention this fact (that the system is big.LITTLE, and
> the boot cores are not Cortex-A73) earlier in the report.
will take care
>
>> Now when errata impacted core boots up and sched_clock_register already
>> register will !arch_timer_counter_has_wa() path.
>> As sched_clock_register is not per_cpu bases so arch_timer_read_counter will
>> always point to !arch_timer_counter_has_wa() function calls.
> Hmm... yes, AFAICT this cannot work unless the affected CPUs are up before we
> probe, and it doesn't make much sense for arch_counter_register() to look at
> arch_timer_counter_has_wa() since it can be called before all CPUs are up.
>
>> Looks like this bug is side effect of following patch:
>>
>> commit 0ea415390cd345b7d09e8c9ebd4b68adfe873043
>> Author: Marc Zyngier <marc.zyngier at arm.com>
>> Date:   Mon Apr 8 16:49:07 2019 +0100
>>
>>      clocksource/arm_arch_timer: Use arch_timer_read_counter to access stable
>> counters
>>
>>      Instead of always going via arch_counter_get_cntvct_stable to access the
>>      counter workaround, let's have arch_timer_read_counter point to the
>>      right method.
>>
>>      For that, we need to track whether any CPU in the system has a
>>      workaround for the counter. This is done by having an atomic variable
>>      tracking this.
>>
>>      Acked-by: Mark Rutland <mark.rutland at arm.com>
>>      Signed-off-by: Marc Zyngier <marc.zyngier at arm.com>
>>      Signed-off-by: Will Deacon <will.deacon at arm.com>
>>
> Yeah, that does look to be broken, but I think there are futher issues anyway
> (e.g. late onlining).
>
> AFAICT we need to detect this *stupidly early* in the CPU bringup path in order
> to handle this safely, which is quite painful.
>
> What a great.
>
> Thanks,
> Mark.