Re: [PATCH v7 4/5] efi: Load efi_secret module if EFI secret area is populated

[Dov Murik <dovmurik@linux.ibm.com>]

On 03/02/2022 8:16, Gerd Hoffmann wrote:
>   Hi,
> 
>>> I think the module should fail noisily.  See above for autoload.  In
>>> case the module is loaded (either manually by the admin, or because
>>> efi.coco_secret != EFI_INVALID_TABLE_ADDR) and it can't actually load
>>> the secrets we want know why ...
>>
>> Note that the AmdSev build of OVMF always publishes
>> LINUX_EFI_COCO_SECRET_TABLE_GUID in the EFI table.  Even when
>> LAUNCH_SECRET was not executed.  In such cases the secret area will be
>> empty.
> 
> Hmm, ok.  Why?  I assume the secret area is filled by the host and ovmf
> doesn't even look at it?
> 

Exactly.  OVMF just reserves this area, and puts its address+size in the
EFI config table.  It doesn't care about its format and usage.

There are currently two "users" for the actual data in this memory area:

1. grub's efisecret module (which reads the disk passphrase from an
entry in the secret area)
2. linux's efi_secret module (which we're discussing here)



>> If we keep only the 'efi.coco_secret != EFI_INVALID_TABLE_ADDR' check,
>> we'll get errors from efi_secret for every VM launch that doesn't
>> undergo LAUNCH_SECRET.  I don't think that's good.
> 
> Well, if that is a common case the module could either print nothing or
> log KERN_INFO level instead of KERN_ERROR.
> 

What if the user doesn't inject a secret and doesn't include the
efi_secret module at all in the initrd?  request_module("efi_secret")
will fail.

I can ignore the error code of request_module("efi_secret") but that
feels bad.



>> If we *do* want to check that the area starts with
>> EFI_SECRET_TABLE_HEADER_GUID (like I think we should), we need all the
>> checks before that, like checking that the area is big enough, and that
>> all the memremap()s succeed -- before actually comparing the
>> header_guid.  The checks are basically prerequisites for calling
>> efi_guidcmp() safely.
> 
> It is still not fully clear to me why you want do that check twice.
> 

I want to load the module only if secrets were injected by the Guest
Owner.

Again, I'm open to ideas on how to de-duplicate these early checks, if
that's important.


-Dov