On 26.01.2026 10:47, Oleksii Kurochko wrote:
>
> On 1/26/26 9:36 AM, Jan Beulich wrote:
>> On 24.01.2026 23:44, Teddy Astie wrote:
>>> Le 22/01/2026 à 17:49, Oleksii Kurochko a écrit :
>>>> +static void vcpu_csr_init(struct vcpu *v)
>>>> +{
>>>> + register_t hstateen0;
>>>> +
>>>> + csr_write(CSR_HEDELEG, HEDELEG_DEFAULT);
>>>> + v->arch.hedeleg = csr_read(CSR_HEDELEG);
>>>> +
>>>> + vcpu_guest_cpu_user_regs(v)->hstatus = HSTATUS_SPV | HSTATUS_SPVP;
>>>> +
>>>> + csr_write(CSR_HIDELEG, HIDELEG_DEFAULT);
>>>> + v->arch.hideleg = csr_read(CSR_HIDELEG);
>>>> +
>>> To me, that feels odd to set machine CSR here. Especially if (I guess)
>>> that we would update them anyway during context switches.
>>>
>>> I think it would be better to have :
>>> - vcpu_csr_init -> sets initial state CSR in vcpu structure, doesn't
>>> touch machine CSR
>>> - context switching logic -> loads vcpu-specific machine CSR from vcpu
>>> structure
>>>
>>> We would have to make a context switch to enter the vcpu for the first
>>> time; but that's to be expected.
>>>
>>> With my proposal, we would also want to move the vcpu_csr_init()
>>> invocation to the place we initialize the vcpu_arch structure rather
>>> than the first time we schedule inside that vcpu.
>> Aiui the writes were put here to be able to cope with r/o-zero bits. Yet
>> indeed it can't be cone like this. While it may work for domains created
>> during boot, these CSRs would be changed under the feet of a running vCPU
>> when done this way for domain creation later at runtime.
>
> Why these CSRs would want to be changed when we have already running vCPU?
>
> Even they will be changed what is an issue, they will be sync-ed during
> context switch.
Which is too late (plus on ctxt-switch-out I assume they'll be synced from
HW to internal structures). The problem here is that a vCPU, having invoked
a hypercall resulting in a new vCPU (for another domain) to be created, will
observe a change of some of the CSRs controlling its own behavior.
>> Instead, as I think I had also suggested earlier on, the r/o-zero-ness of
>> bits in particular CSRs wants determining once during boot, and then that
>> mask wants using when setting up vCPU-s.
>
> Somewhere even before create_domUs() will be called?
Yes, before any domains (and in particular vCPU-s) are created. The idle
domain may be an exception here, and system domains (not having vCPU-s) may
also be unaffected and hence not depend on particular ordering.
Jan
