Re: [PATCH v2 1/2] KVM: x86: Add KVM_[GS]ET_CLOCK_GUEST for accurate KVM clock migration

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]


On 10 April 2024 11:29:13 BST, Paul Durrant <xadimgnik@xxxxxxxxx> wrote:
>On 10/04/2024 10:52, Jack Allister wrote:
>> +	 * It's possible that this vCPU doesn't have a HVCLOCK configured
>> +	 * but the other vCPUs may. If this is the case calculate based
>> +	 * upon the time gathered in the seqcount but do not update the
>> +	 * vCPU specific PVTI. If we have one, then use that.
>Given this is a per-vCPU ioctl, why not fail in the case the vCPU doesn't have HVCLOCK configured? Or is your intention that a GET/SET should always work if TSC is stable?

It definitely needs to work for SET even when the vCPU hasn't been run yet (and doesn't have a hvclock in vcpu->arch.hv_clock).

I think it should ideally work for GET too. I did try arguing that if the vCPU hasn't set up its pvclock then why would it care if it's inaccurate? But there's a pathological case of AMP where one vCPU is dedicated to an RTOS or something, and only the *other* vCPUs bring up their pvclock.

This of course brings you to the question of why we have it as a per-vCPU ioctl at all? It only needs to be done *once* to get/set the KVM-wide clock
 And a function of *this* vCPU's TSC. And the point is that if we're in use_master_clock mode, that's consistent across *all* vCPUs. There would be a bunch of additional complexity in making it a VM ioctl though, especially around the question of what to do if userspace tries to restore it when there *aren't* any vCPUs yet. So we didn't do that.

[Index of Archives]     [KVM ARM]     [KVM ia64]     [KVM ppc]     [Virtualization Tools]     [Spice Development]     [Libvirt]     [Libvirt Users]     [Linux USB Devel]     [Linux Audio Users]     [Yosemite Questions]     [Linux Kernel]     [Linux SCSI]     [XFree86]

  Powered by Linux