Re: [RFC PATCH v2 3/5] x86: Track vcpu context switches and introduce needs_tlb_flush field

[Ross Lagerwall <ross.lagerwall@xxxxxxxxxx>]

On 7/31/26 3:46 PM, Teddy Astie wrote:
> Introduce needs_tlb_flush that indicate that the vCPU related TLB
> needs to be flushed before entering this domain. This is intended
> to be used later for using the same ASID for all vCPU of a domain.
>
> Also track vCPU context switch to determine :
>   - per vcpu "latest_cpu" that tracks which pCPU last ran this vCPU,
>     this is used to know whether our current TLB (of our pCPU) state
>     is stale.
>     Schedule a TLB flush if the vCPU ran on another pCPU previously.
>   - per pCPU "latest_vcpu" (per domain) that tracks which vCPU the
>     TLB+ASID/VPID is holding onto.
>     Schedule a TLB flush if this pCPU hasn't ran this vCPU previously
>     (which can happen if we context-switch multiples vCPUs of a same
>     domain on a same pCPU).
>
> If ASID use is disabled, unconditionnaly perform a TLB flush.
>
> Signed-off-by: Teddy Astie <teddy.astie@xxxxxxxxxx>
> ---
> v2: Merge several patches into one, move logic to context_switch(),
>      move fields to arch_vcpu/arch_domain, consider !asid_enabled case.
> ---
>   xen/arch/x86/domain.c             | 26 ++++++++++++++++++++++++++
>   xen/arch/x86/include/asm/domain.h |  6 ++++++
>   2 files changed, 32 insertions(+)
>
> diff --git a/xen/arch/x86/domain.c b/xen/arch/x86/domain.c
> index 996b50af7a..27155546ca 100644
> --- a/xen/arch/x86/domain.c
> +++ b/xen/arch/x86/domain.c
> @@ -38,6 +38,7 @@
>   #include <xen/smp.h>
>   #include <xen/softirq.h>
>   #include <xen/wait.h>
> +#include <xen/xvmalloc.h>
>   
>   #include <asm/amd.h>
>   #include <asm/cpu-policy.h>
> @@ -874,6 +875,13 @@ int arch_domain_create(struct domain *d,
>   
>       spec_ctrl_init_domain(d);
>   
> +    rc = -ENOMEM;
> +    d->arch.latest_vcpu = xvmalloc_array(int, nr_cpu_ids);
Does this work with pCPU hotplug or do you need to use NR_CPUS?

> +    if ( !d->arch.latest_vcpu )
> +        goto fail;
> +    for (unsigned int i = 0; i < nr_cpu_ids; i++)
> +        d->arch.latest_vcpu[i] = -1;
> +
The other arches define INVALID_VCPU_ID to be MAX_VIRT_CPUS. Not sure if it is
worth doing the same for x86?

Ross