Re: [PATCH Part1 v5 32/38] x86/sev: enable SEV-SNP-validated CPUID in #VC handlers

[Michael Roth]

On Fri, Aug 27, 2021 at 05:18:49PM +0200, Borislav Petkov wrote:
> On Fri, Aug 20, 2021 at 10:19:27AM -0500, Brijesh Singh wrote:
> > From: Michael Roth <michael.roth@xxxxxxx>
> > 
> > This adds support for utilizing the SEV-SNP-validated CPUID table in
> 
> s/This adds support for utilizing/Utilize/
> 
> Yap, it can really be that simple. :)
> 
> > the various #VC handler routines used throughout boot/run-time. Mostly
> > this is handled by re-using the CPUID lookup code introduced earlier
> > for the boot/compressed kernel, but at various stages of boot some work
> > needs to be done to ensure the CPUID table is set up and remains
> > accessible throughout. The following init routines are introduced to
> > handle this:
> 
> Do not talk about what your patch does - that should hopefully be
> visible in the diff itself. Rather, talk about *why* you're doing what
> you're doing.

I'll get this cleaned up.

> 
> > sev_snp_cpuid_init():
> 
> This one is not really introduced - it is already there.
> 
> <snip all the complex rest>
> 
> So this patch is making my head spin. It seems we're dancing a lot of
> dance just to have our CPUID page present at all times. Which begs the
> question: do we need it during the whole lifetime of the guest?
> 
> Regardless, I think this can be simplified by orders of
> magnitude if we allocated statically 4K for that CPUID page in
> arch/x86/boot/compressed/mem_encrypt.S, copied the supplied CPUID page
> from the firmware to it and from now on, work with our own copy.

That makes sense. I was thinking it was safer to work with the FW page
since it would be less susceptible to something like a buffer overflow
modifying the CPUID table, but __ro_after_init seems like it would
provide similar protections. And yes, would definitely be great to avoid
the need for so many [re-]init routines.

> 
> You probably would need to still remap it for kernel proper but it would
> get rid of all that crazy in this patch here.
> 
> Hmmm?

If the memory is allocated in boot/compressed/mem_encrypt.S, wouldn't
kernel proper still need to create a static buffer for its copy? And if
not, wouldn't boot compressed still need a way to pass the PA of this
buffer? That seems like it would need to be done via boot_params. It
seems like it would also need to be marked as reserved as well since
kernel proper could no longer rely on the EFI map to handle it.

I've been testing a similar approach based on your suggestion that seems
to work out pretty well, but there's still some ugliness due to the
fixup_pointer() stuff that's needed early during snp_cpuid_init() in
kernel proper, which results in the need for 2 init routines there. Not
sure if there's a better way to handle it, but it's a lot better than 4
init routines at least, and with this there is no longer any need to
store the address/size of the FW page:

in arch/x86/kernel/sev-shared.c:

/* Firmware-enforced limit on CPUID table entries */
#define SNP_CPUID_COUNT_MAX 64

struct sev_snp_cpuid_info {
        u32 count;
        u32 __reserved1;
        u64 __reserved2;
        struct sev_snp_cpuid_fn fn[SNP_CPUID_COUNT_MAX];
} __packed;

static struct snp_cpuid_info cpuid_info_copy __ro_after_init;
static const struct snp_cpuid_info *cpuid_info __ro_after_init;
static int sev_snp_cpuid_enabled __ro_after_init;

/*
 * Initial set up of CPUID table when running identity-mapped.
 */
#ifdef __BOOT_COMPRESSED
void sev_snp_cpuid_init(struct boot_params *bp)
#else
void __init sev_snp_cpuid_init(struct boot_params *bp, unsigned long physaddr)
#endif
{
        const struct sev_snp_cpuid_info *cpuid_info_fw;

        cpuid_info_fw = snp_probe_cpuid_info(bp);
        if (!cpuid_info_fw)
                return;

#ifdef __BOOT_COMPRESSED
        cpuid_info2 = &cpuid_info_copy;
#else
        /* Kernel proper calls this while pointer fixups are still needed. */
        cpuid_info2 = (const struct sev_snp_cpuid_info *)
                       ((void *)&cpuid_info_copy - (void *)_text + physaddr);
#endif
        memcpy((struct sev_snp_cpuid_info *)cpuid_info2, cpuid_info_fw,
               sizeof(*cpuid_info2));

        sev_snp_cpuid_enabled = 1;
}

#ifndef __BOOT_COMPRESSED
/*
 * This is called after the switch to virtual kernel addresses. At this
 * point pointer fixups are no longer needed, and the virtual address of
 * the CPUID info buffer has changed, so re-initialize the pointer.
 */
void __init sev_snp_cpuid_init_virtual(void)
{
        /*
         * sev_snp_cpuid_init() already did the initial parsing of bootparams
         * and initial setup. If that didn't enable the feature then don't try
         * to enable it here.
         */
        if (!sev_snp_cpuid_active())
                return;

        /*
         * Either boot_params/EFI advertised the feature even though SNP isn't
         * enabled, or something else went wrong. Bail out.
         */
        if (!sev_feature_enabled(SEV_SNP))
                sev_es_terminate(1, GHCB_TERM_CPUID);

        cpuid_info = &cpuid_info_copy;
}
#endif

Then the rest of the code just accesses cpuid_info directly as it does now.

Would that be a reasonable approach for v6?

> 
> -- 
> Regards/Gruss,
>     Boris.
> 
> https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Fpeople.kernel.org%2Ftglx%2Fnotes-about-netiquette&amp;data=04%7C01%7Cmichael.roth%40amd.com%7C464de4ea70544dc32ba108d9696de67a%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637656743008963071%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&amp;sdata=wYK2a%2FuHgw0%2FQJsCVCIpoaJxdT0XASMYUfHvmDcowXg%3D&amp;reserved=0