RE: [PATCHv3 0/3] x86/tdx: Fix one more load_unaligned_zeropad() issue

[Michael Kelley (LINUX)]

From: Kirill A. Shutemov <kirill.shutemov@xxxxxxxxxxxxxxx> Sent: Tuesday, June 6, 2023 2:56 AM
> 
> During review of TDX guests on Hyper-V patchset Dave pointed to the
> potential race between changing page private/shared status and
> load_unaligned_zeropad().
> 
> Fix the issue.
> 
> v3:
>  - Fix grammar;
>  - Add Sathya's Reviewed-bys;
> v2:
>  - Add more info in commit message of the first patch.
>  - Move enc_status_change_finish_noop() into a separate patch.
>  - Fix typo in commit message and comment.
> 
> Kirill A. Shutemov (3):
>   x86/mm: Allow guest.enc_status_change_prepare() to fail
>   x86/tdx: Fix race between set_memory_encrypted() and
>     load_unaligned_zeropad()
>   x86/mm: Fix enc_status_change_finish_noop()
> 
>  arch/x86/coco/tdx/tdx.c         | 64 +++++++++++++++++++++++++++++++--
>  arch/x86/include/asm/x86_init.h |  2 +-
>  arch/x86/kernel/x86_init.c      |  4 +--
>  arch/x86/mm/mem_encrypt_amd.c   |  4 ++-
>  arch/x86/mm/pat/set_memory.c    |  3 +-
>  5 files changed, 69 insertions(+), 8 deletions(-)
> 
> --
> 2.39.3

These fixes notwithstanding, load_unaligned_zeropad() is not handled
properly in a TDX VM.  The problem was introduced with commit
c4e34dd99f2e, which moved the fixup code to function
ex_handler_zeropad().  This new function does a verification against
fault_addr, and the verification always fails because fault_addr is zero.
The call sequence is:

exc_virtualization_exception()
ve_raise_fault()
gp_try_fixup_and_notify()  <-- always passes 0 as fault_addr
fixup_exception()
ex_handler_zeropad()

The validation of fault_addr could probably be removed since
such validation wasn't there prior to c4e34dd99f2e.  But before
going down that path, I want to propose a different top-level
solution to the interaction between load_unaligned_zeropad()
and CoCo VM page transitions between private and shared.

When a page is transitioning, the caller can and should ensure
that it is not being accessed during the transition.  But we have
the load_unaligned_zeropad() wildcard.   So do the following for
the transition sequence in __set_memory_enc_pgtable():

1.  Remove aliasing mappings
2.  Remove the PRESENT bit from the PTEs of all transitioning pages
3.  Flush the TLB globally
4.  Flush the data cache if needed
5.  Set/clear the encryption attribute as appropriate
6.  Notify the hypervisor of the page status change
7.  Add back the PRESENT bit

With this approach, load_unaligned_zeropad() just takes the
normal page-fault-based fixup path if it touches a page that is
transitioning.  As a result, load_unaligned_zeropad() and CoCo
VM page transitioning are completely decoupled.  We don't
need to handle architecture-specific CoCo VM exceptions and
fix things up.

I've posted an RFC PATCH that implements this approach [1],
and tested on TDX VMs and SEV-SNP VMs in vTOM mode.
The RFC PATCH has more details on the benefits and
implications.  Follow-up discussion should probably be done
on that email thread.

Michael

[1] https://lore.kernel.org/lkml/1688661719-60329-1-git-send-email-mikelley@xxxxxxxxxxxxx/T/#u