Re: [PATCH 0/3] arm64: WFxT fixes, take #2

Hi Marc,

On 2/26/26 08:22, Marc Zyngier wrote:

After my previous WFxT fix went in 7.0-rc1 as 29cc0f3aa7c64 ("arm64:
Force the use of CNTVCT_EL0 in __delay()"), Ben reported that it isn't
playing nice with preemption, due to the use of the timer workaround
percpu variable. This series tries to address this, and propose a
hopefully better alternative.

That alternative is in the form of a new "virtual counter" accessor,
available in the same way arch_timer_read_counter() is, except that it
is guaranteed to be the virtual counter. This helper (a function
pointer, really) is updated on each CPU boot in the same manner the
counter accessor is.

This is then plugged into the __delay() helper, providing the expected
guarantees (and resulting in a much nicer code gen).

Patches on top of -rc1.

Marc Zyngier (3):
  arm64: Fix sampling the "stable" virtual counter in preemptible
    section
  clocksource/drivers/arm_arch_timer: Expose a direct accessor for the
    virtual counter
  arm64: Convert __delay_cycles() to arch_timer_read_vcounter()

 arch/arm64/lib/delay.c               | 5 ++++-
 drivers/clocksource/arm_arch_timer.c | 5 +++++
 include/clocksource/arm_arch_timer.h | 1 +
 3 files changed, 10 insertions(+), 1 deletion(-)

I checked with just the first patch applied and with all the patchs
applied. In both case I no longer see the preemption warnings on the s/w
model FVP Base C with CONFIG_DEBUG_PREEMPT.

Tested-by: Ben Horgan <ben.horgan@arm.com>

Thanks,

Ben