ERRATUM_858921 is broken on 5.15 kernel

[Yogesh Lal]

Hi,

We are observing issue on A73 core where ERRATUM_858921 is broken.

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.

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.


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>


Thanks

Yogesh Lal