Re: [Patch v4 05/13] init: Call mem_encrypt_init() after Hyper-V hypercall init is done

From: Sathyanarayanan Kuppuswamy
Date: Tue Dec 06 2022 - 15:30:32 EST




On 12/6/22 12:13 PM, Michael Kelley (LINUX) wrote:
> From: Sathyanarayanan Kuppuswamy <sathyanarayanan.kuppuswamy@xxxxxxxxxxxxxxx>
>>
>>
>>
>> On 12/1/22 7:30 PM, Michael Kelley wrote:
>>> Full Hyper-V initialization, including support for hypercalls, is done
>>> as an apic_post_init callback via late_time_init(). mem_encrypt_init()
>>> needs to make hypercalls when it marks swiotlb memory as decrypted.
>>> But mem_encrypt_init() is currently called a few lines before
>>> late_time_init(), so the hypercalls don't work.
>>
>> Did you consider moving hyper-v hypercall initialization before
>> mem_encrypt_init(). Is there any dependency issue?
>
> Hyper-V initialization has historically been done using the callbacks
> that exist in the x86 initialization paths, rather than adding explicit
> Hyper-V init calls. As noted above, the full Hyper-V init is done on
> the apic_post_init callback via late_time_init(). Conceivably we could
> add an explicit call to do the Hyper-V init, but I think there's still a
> problem in putting that Hyper-V init prior to the current location of
> mem_encrypt_init(). I'd have to go check the history, but I think the
> Hyper-V init needs to happen after the APIC is initialized.

Ok. If there is a dependency or complexity issue, I recommend adding that
detail in the commit log.

>
> It seems like moving mem_encrypt_init() slightly later is the cleaner
> long-term solution. Are you aware of a likely problem arising in the
> future with moving mem_encrypt_init()?

I did not investigate in depth, but there appears to be no problem with
moving mem_encrypt_init(). But my point is, if it is possible to fix this
easily by changing Hyper-v specific initialization, we should consider it
first before considering moving the common mem_encrypt_init() function.



>
> Michael
>
>>
>>>
>>> Fix this by moving mem_encrypt_init() after late_time_init() and
>>> related clock initializations. The intervening initializations don't
>>> do any I/O that requires the swiotlb, so moving mem_encrypt_init()
>>> slightly later has no impact.
>>>

--
Sathyanarayanan Kuppuswamy
Linux Kernel Developer