Re: [PATCH v11 18/35] KVM: SEV: Add KVM_SEV_SNP_LAUNCH_UPDATE command

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

 



On Wed, Feb 7, 2024 at 3:43 AM Sean Christopherson <seanjc@xxxxxxxxxx> wrote:
> > > Userspace owns whether a page is PRIVATE or SHARED, full stop.  If KVM can't
> > > honor that, then we need to come up with better uAPI.
> >
> > Can you explain more verbosely what you mean?
>
> As proposed, snp_launch_update_gfn_handler() doesn't verify the state of the
> gfns' attributes.  But that's a minor problem and probably not a sticking point.
>
> My overarching complaint is that the code is to be wildly unsafe, or at the very
> least brittle.  Without guest_memfd's knowledge, and without holding any locks
> beyond kvm->lock, it
>
>  1) checks if a pfn is shared in the RMP
>  2) copies data to the page
>  3) converts the page to private in the RMP
>  4) does PSP stuff
>  5) on failure, converts the page back to shared in RMP
>  6) conditionally on failure, writes to the page via a gfn
>
> I'm not at all confident that 1-4 isn't riddled with TOCTOU bugs, and that's
> before KVM gains support for intrahost migration, i.e. before KVM allows multiple
> VM instances to bind to a single guest_memfd.

Absolutely.

> But I _think_ we mostly sorted this out at PUCK.  IIRC, the plan is to have guest_memfd
> provide (kernel) APIs to allow arch/vendor code to initialize a guest_memfd range.
> That will give guest_memfd complete control over the state of a given page, will
> allow guest_memfd to take the appropriate locks, and if we're lucky, will be reusable
> by other CoCo flavors beyond SNP.

Ok, either way it's clear that guest_memfd needs to be in charge.
Whether it's MEMORY_ENCRYPT_OP that calls into guest_memfd or vice
versa, that only matters so much.

> > > Yes, I am specifically complaining about writing guest memory on failure, which is
> > > all kinds of weird.
> >
> > It is weird but I am not sure if you are complaining about firmware
> > behavior or something else.
>
> This proposed KVM code:
>
> +                               host_rmp_make_shared(pfns[i], PG_LEVEL_4K, true);
> +
> +                               ret = kvm_write_guest_page(kvm, gfn, kvaddr, 0, PAGE_SIZE);
> +                               if (ret)
> +                                       pr_err("Failed to write CPUID page back to userspace, ret: 0x%x\n",
> +                                              ret);
>
>
> I have no objection to propagating error/debug information back to userspace,
> but it needs to be routed through the source page (or I suppose some dedicated
> error page, but that seems like overkill).  Shoving the error information into
> guest memory is gross.

Yes, but it should be just a consequence of not actually using
start_gfn. Having to copy back remains weird, but what can you do.

Paolo






[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