Re: [Xen-devel] [PATCH v3 15/16] x86/VPMU: NMI-based VPMU support

["Jan Beulich" <JBeulich@xxxxxxxx>]

>>> On 06.01.14 at 20:24, Boris Ostrovsky <boris.ostrovsky@xxxxxxxxxx> wrote:
> Add support for using NMIs as PMU interrupts

This needs significant extension, detailing how all the handling
involved is safe in NMI context.

> @@ -202,10 +227,14 @@ int vpmu_do_interrupt(struct cpu_user_regs *regs)
>              struct segment_register cs;
>  
>              gregs = guest_cpu_user_regs();
> -            hvm_get_segment_register(current, x86_seg_cs, &cs);
>  
>              memcpy(p, gregs, sizeof(struct cpu_user_regs));
> -            ((struct cpu_user_regs *)p)->cs = cs.attr.fields.dpl;
> +            /* This is unsafe in NMI context, we'll do it in softint handler 
> */
> +            if ( vpmu_apic_vector != APIC_DM_NMI )

Even if correct at present, I'd strongly suggest against != here (and
below), in favor of &.

> +/* Process the softirq set by PMU NMI handler */
> +static void pmu_softnmi(void)
> +{
> +    struct cpu_user_regs *regs;
> +    struct vcpu *v, *sampled = per_cpu(sampled_vcpu, smp_processor_id());
> +
> +    if ( vpmu_mode & XENPMU_MODE_PRIV ||
> +         sampled->domain->domain_id >= DOMID_FIRST_RESERVED )
> +        v = dom0->vcpu[smp_processor_id()];
> +    else
> +        v = sampled;
> +
> +    regs = &v->arch.vpmu.xenpmu_data->pmu.regs;
> +    if ( is_hvm_domain(sampled->domain) )
> +    {
> +        struct segment_register cs;
> +
> +        hvm_get_segment_register(sampled, x86_seg_cs, &cs);
> +        regs->cs = cs.attr.fields.dpl;
> +    }
> +
> +    send_guest_vcpu_virq(v, VIRQ_XENPMU);
> +}

Perhaps I should have asked this on an earlier patch already:
How is this supposed to work for a 32-bit HVM guest?
struct cpu_user_regs is clearly different for it than what the
hypervisor or a 64-bit HVM guest would use.

Jan


_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel