On 8/21/23 14:25, Steve Rutherford wrote:
> On Mon, Aug 21, 2023 at 11:54 AM Tom Lendacky <[email protected]> wrote:
>>
>> On 8/21/23 13:15, Steve Rutherford wrote:
>>> On Mon, Aug 21, 2023 at 6:10 AM Tom Lendacky <[email protected]> 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.
Ok, cool.
>>
>> 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().
I like the fix for the hypercall being in early_set_memory_enc_dec(). This
way the behavior doesn't change for existing callers and doesn't require
adding a WARN.
Thanks,
Tom
> 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 <[email protected]>
>>>>> ---
>>>>> 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));