> On 13. May 2026, at 04:25, Scott J. Goldman <[email protected]> wrote:
>
> On Tue May 12, 2026 at 9:21 PM EDT, Mohamed Mediouni wrote:
>>
>>> On 27. Apr 2026, at 21:55, Scott J. Goldman <[email protected]> wrote:
>>>
>>> Commit b5f8f77271 ("accel/hvf: Implement WFI without using pselect()")
>>> changed hvf_wfi() from blocking the vCPU thread with pselect() to
>>> returning EXCP_HLT, intending QEMU's main event loop to handle the
>>> idle wait. However, cpu->halted was never set, so cpu_thread_is_idle()
>>> always returns false and the vCPU thread spins at 100% CPU per core
>>> while the guest is idle.
>>>
>>> Fix this by:
>>>
>>> 1. Setting cpu->halted = 1 in hvf_wfi() so the vCPU thread sleeps on
>>> halt_cond in qemu_process_cpu_events().
>>>
>>> 2. Arming a per-vCPU QEMU_CLOCK_VIRTUAL timer to fire when the guest's
>>> virtual timer (CNTV_CVAL_EL0) would expire. This is necessary
>>> because HVF only delivers HV_EXIT_REASON_VTIMER_ACTIVATED during
>>> hv_vcpu_run(), which is not called while the CPU is halted. The
>>> timer callback mirrors the VTIMER_ACTIVATED handler: it raises the
>>> vtimer IRQ through the GIC and marks vtimer_masked, causing the
>>> interrupt delivery chain to wake the vCPU via qemu_cpu_kick().
>>>
>>> 3. Clearing cpu->halted in hvf_arch_vcpu_exec() when cpu_has_work()
>>> indicates a pending interrupt, and cancelling the WFI timer.
>>>
>>> 4. Re-arming the WFI timer from hvf_vm_state_change() on the resume
>>> transition for any halted vCPU, since the QEMUTimer is per-instance
>>> state and is not migrated. After cpu_synchronize_all_states() the
>>> migrated vtimer state is mirrored in env, so we can read CNTV_CTL
>>> and CNTV_CVAL from there. If the vtimer has already expired by the
>>> time the destination resumes, hvf_wfi_timer_cb() is invoked
>>> directly so the halted vCPU is woken up.
>>>
>>> Fixes: b5f8f77271 ("accel/hvf: Implement WFI without using pselect()")
>>> Signed-off-by: Scott J. Goldman <[email protected]>
>>> ---
>>> Changes since v2:
>>> - Use QEMU_CLOCK_VIRTUAL instead of QEMU_CLOCK_HOST so the timer
>>> pauses with the VM and a halted vCPU isn't woken (or its IRQ
>>> raised) while the user has stopped the guest. (Peter)
>>> - Convert vtimer ticks to nanoseconds with muldiv64() to avoid
>>> intermediate overflow. (Peter)
>>> - Re-arm the WFI timer from hvf_vm_state_change() on the resume
>>> transition so a halted vCPU on the migration destination is
>>> woken when its vtimer expires (the QEMUTimer is per-instance
>>> state and isn't migrated). (Peter)
>>> v2:
>>> https://lore.kernel.org/qemu-devel/[email protected]/
>>> v1:
>>> https://lore.kernel.org/qemu-devel/[email protected]/
>>
>> For QEMU 11.0 (for backporting to stable):
>>
>> Reviewed-by: Mohamed Mediouni <[email protected]>
>>
>> For QEMU 11.1:
>>
>> Adding some checks for !hvf_irqchip_in_kernel() needed
>> but can do them on my side if you prefer.
>>
>> Looks ready apart from that bit.
>
> Hi Mohamed,
>
> TFTR - I sent a follow-up patch that gates the wfi timer as you requested.
> For applying this v3 version to the stable tree, is there anyone else I
> should ping? I'm not very familiar with the process.
>
> Thanks!
> -sjg
>
>
Hi,
v3 of this patch is for QEMU 11.0 qemu-stable and v4 is for QEMU 11.1.
A catch is that this isn’t in master yet though.
This regression doesn’t apply to QEMU 10.2 or earlier.
Could this regression fix be included into QEMU 11.0.1?
(with qemu-stable@ in cc)
