[PATCH] arm64: Force the use of CNTVCT_EL0 in __delay()
From: Marc Zyngier <maz@kernel.org>
Date: 2026-02-13 14:16:24
Also in:
kvmarm, stable
Subsystem:
arm64 port (aarch64 architecture), the rest · Maintainers:
Catalin Marinas, Will Deacon, Linus Torvalds
Quentin forwards a report from Hyesoo Yu, describing an interesting
problem with the use of WFxT in __delay() when a vcpu is loaded and
that KVM is *not* in VHE mode (either nVHE or hVHE).
In this case, CNTVOFF_EL2 is set to a non-zero value to reflect the
state of the guest virtual counter. At the same time, __delay() is
using get_cycles() to read the counter value, which is indirected to
reading CNTPCT_EL0.
The core of the issue is that WFxT is using the *virtual* counter,
while the kernel is using the physical counter, and that the offset
introduces a really bad discrepancy between the two.
Fix this by forcing the use of CNTVCT_EL0, making __delay() consistent
irrespective of the value of CNTVOFF_EL2.
Reported-by: Hyesoo Yu <redacted>
Reported-by: Quentin Perret <redacted>
Reviewed-by: Quentin Perret <redacted>
Fixes: 7d26b0516a0df ("arm64: Use WFxT for __delay() when possible")
Signed-off-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/ktosachvft2cgqd5qkukn275ugmhy6xrhxur4zqpdxlfr3qh5h@o3zrfnsq63od (local)
Cc: stable@vger.kernel.org
---
arch/arm64/lib/delay.c | 19 +++++++++++++++----
1 file changed, 15 insertions(+), 4 deletions(-)
diff --git a/arch/arm64/lib/delay.c b/arch/arm64/lib/delay.c
index cb2062e7e2340..d02341303899e 100644
--- a/arch/arm64/lib/delay.c
+++ b/arch/arm64/lib/delay.c@@ -23,9 +23,20 @@ static inline unsigned long xloops_to_cycles(unsigned long xloops) return (xloops * loops_per_jiffy * HZ) >> 32; } +/* + * Force the use of CNTVCT_EL0 in order to have the same base as WFxT. + * This avoids some annoying issues when CNTVOFF_EL2 is not reset 0 on a + * KVM host running at EL1 until we do a vcpu_put() on the vcpu. When + * running at EL2, the effective offset is always 0. + * + * Note that userspace cannot change the offset behind our back either, + * as the vcpu mutex is held as long as KVM_RUN is in progress. + */ +#define __delay_cycles() __arch_counter_get_cntvct_stable() + void __delay(unsigned long cycles) { - cycles_t start = get_cycles(); + cycles_t start = __delay_cycles(); if (alternative_has_cap_unlikely(ARM64_HAS_WFXT)) { u64 end = start + cycles;
@@ -35,17 +46,17 @@ void __delay(unsigned long cycles) * early, use a WFET loop to complete the delay. */ wfit(end); - while ((get_cycles() - start) < cycles) + while ((__delay_cycles() - start) < cycles) wfet(end); } else if (arch_timer_evtstrm_available()) { const cycles_t timer_evt_period = USECS_TO_CYCLES(ARCH_TIMER_EVT_STREAM_PERIOD_US); - while ((get_cycles() - start + timer_evt_period) < cycles) + while ((__delay_cycles() - start + timer_evt_period) < cycles) wfe(); } - while ((get_cycles() - start) < cycles) + while ((__delay_cycles() - start) < cycles) cpu_relax(); } EXPORT_SYMBOL(__delay);
--
2.47.3