Date:   Mon, 21 Aug 2023 12:25:47 -0700
From:   Steve Rutherford <srutherford@...gle.com>
To:     Tom Lendacky <thomas.lendacky@....com>
Cc:     Borislav Petkov <bp@...en8.de>,
        Thomas Gleixner <tglx@...utronix.de>,
        Paolo Bonzini <pbonzini@...hat.com>,
        Wanpeng Li <wanpengli@...cent.com>,
        Vitaly Kuznetsov <vkuznets@...hat.com>,
        Ingo Molnar <mingo@...hat.com>,
        Dave Hansen <dave.hansen@...ux.intel.com>, x86@...nel.org,
        "H . Peter Anvin" <hpa@...or.com>, kvm@...r.kernel.org,
        linux-kernel@...r.kernel.org, David.Kaplan@....com,
        jacobhxu@...gle.com, patelsvishal@...gle.com, bhillier@...gle.com
Subject: Re: [PATCH] x86/sev: Make early_set_memory_decrypted() calls page aligned

On Mon, Aug 21, 2023 at 11:54 AM Tom Lendacky <thomas.lendacky@....com> wrote:
>
> On 8/21/23 13:15, Steve Rutherford wrote:
> > On Mon, Aug 21, 2023 at 6:10 AM Tom Lendacky <thomas.lendacky@....com> wrote:
> >>
> >> On 8/18/23 18:34, Steve Rutherford wrote:
> >>> early_set_memory_decrypted() assumes its parameters are page aligned.
> >>> Non-page aligned calls result in additional pages being marked as
> >>> decrypted via the encryption status hypercall, which results in
> >>> consistent corruption of pages during live migration. Live
> >>> migration requires accurate encryption status information to avoid
> >>> migrating pages from the wrong perspective.
> >>
> >> Hmmm... I'm not sure this is the proper fix. The code is actually doing
> >> the right thing from a encyrption/decryption point of view by checking the
> >> c-bit for the PTE associated with the virtual address and the size
> >> (possibly crossing page boundaries).
> >>
> >> I think the problem is on the call to early_set_mem_enc_dec_hypercall()
> >> where it doesn't take into account the possible crossing of page
> >> boundaries and so can under-count the number of pages, right?
> >
> > Right now, if you request decryption of e.g. a non-page aligned 0x40
> > byte structure, it rounds the 0x40 bytes up to one page, and then
> > hypercalls to mark both the page it's on and the subsequent page as
> > decrypted (since the rounding stretches the structure onto the next
> > page spuriously). The arithmetic in the combination of
> > early_set_memory_enc_dec() and early_set_mem_enc_dec_hypercall() are
> > correct if they are called with page aligned vaddrs (non-page-aligned
> > sizes are fine iiuc).
>
> Ah, right, correct. It is still related to how the page count is
> calculated for the hypercall, though, right? The encryption/decryption
> operations function properly.

Yep! It's just the hypercall that behaves poorly in this situation.
>
> If another caller of early_set_memory_decrypted() gets added, it would
> need to know to do the same thing. So I just wonder if this wouldn't be
> better fixed in early_set_memory_enc_dec() by using a page aligned address
> and proper number of pages when calling early_set_mem_enc_dec_hypercall()
> or in early_set_mem_enc_dec_hypercall() where it would take a size
> argument instead of a page count and does the proper work to get a page
> aligned address and proper page count.
>
> Also, if it is the hypercall that is causing the issue, should the Fixes
> tag be 064ce6c550a0 ("mm: x86: Invoke hypercall when page encryption
> status is changed") since the problem is around the hypercall.

Fair question. I was torn about where to point this, since either
fixing up the value inside early_set_memory_enc_dec() or fixing up the
per-cpu callers is correct. The non-early version
(__set_memory_enc_pgtable()) calls WARN_ONCE for misaligned addresses
under the hood, so I thought the early version should have the same
contract (though, obviously, this lacks the actual WARN_ONCE). I can
re-upload with a WARN_ONCE or with the masking moved into
early_set_memory_enc_dec().
Thanks,
Steve

>
> Thanks,
> Tom
>
> >
> > Thanks,
> > Steve
> >>
> >> Thanks,
> >> Tom
> >>
> >>>
> >>> Fixes: 4716276184ec ("X86/KVM: Decrypt shared per-cpu variables when SEV is active")
> >>> Signed-off-by: Steve Rutherford <srutherford@...gle.com>
> >>> ---
> >>>    arch/x86/kernel/kvm.c | 14 +++++++++++++-
> >>>    1 file changed, 13 insertions(+), 1 deletion(-)
> >>>
> >>> diff --git a/arch/x86/kernel/kvm.c b/arch/x86/kernel/kvm.c
> >>> index 6a36db4f79fd..a0c072d3103c 100644
> >>> --- a/arch/x86/kernel/kvm.c
> >>> +++ b/arch/x86/kernel/kvm.c
> >>> @@ -419,7 +419,14 @@ static u64 kvm_steal_clock(int cpu)
> >>>
> >>>    static inline void __set_percpu_decrypted(void *ptr, unsigned long size)
> >>>    {
> >>> -     early_set_memory_decrypted((unsigned long) ptr, size);
> >>> +     /*
> >>> +      * early_set_memory_decrypted() requires page aligned parameters, but
> >>> +      * this function needs to handle ptrs offset into a page.
> >>> +      */
> >>> +     unsigned long start = PAGE_ALIGN_DOWN((unsigned long) ptr);
> >>> +     unsigned long end = (unsigned long) ptr + size;
> >>> +
> >>> +     early_set_memory_decrypted(start, end - start);
> >>>    }
> >>>
> >>>    /*
> >>> @@ -438,6 +445,11 @@ static void __init sev_map_percpu_data(void)
> >>>                return;
> >>>
> >>>        for_each_possible_cpu(cpu) {
> >>> +             /*
> >>> +              * Calling __set_percpu_decrypted() for each per-cpu variable is
> >>> +              * inefficent, since it may decrypt the same page multiple times.
> >>> +              * That said, it avoids the need for more complicated logic.
> >>> +              */
> >>>                __set_percpu_decrypted(&per_cpu(apf_reason, cpu), sizeof(apf_reason));
> >>>                __set_percpu_decrypted(&per_cpu(steal_time, cpu), sizeof(steal_time));
> >>>                __set_percpu_decrypted(&per_cpu(kvm_apic_eoi, cpu), sizeof(kvm_apic_eoi));

Powered by blists - more mailing lists