Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Message-ID: <3fd38a32-863d-36a8-d5b3-3b3f181d0e56@redhat.com>
Date:   Wed, 28 Sep 2022 12:04:42 +0200
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:102.0) Gecko/20100101
 Thunderbird/102.2.1
Content-Language: en-US
To:     Sean Christopherson <seanjc@google.com>
Cc:     linux-kernel@vger.kernel.org, kvm@vger.kernel.org
References: <20220926165112.603078-1-pbonzini@redhat.com>
 <YzMt24/14n1BVdnI@google.com>
From:   Paolo Bonzini <pbonzini@redhat.com>
Subject: Re: [PATCH] KVM: x86: disable on 32-bit unless CONFIG_BROKEN
In-Reply-To: <YzMt24/14n1BVdnI@google.com>
Content-Type: text/plain; charset=UTF-8; format=flowed
Content-Transfer-Encoding: 7bit
Precedence: bulk

On 9/27/22 19:07, Sean Christopherson wrote:
> On Mon, Sep 26, 2022, Paolo Bonzini wrote:
>> 32-bit KVM has extra complications in the code due to:
>>
>> - different ways to write 64-bit values in VMCS
>>
>> - different handling of DS and ES selectors as well as FS/GS bases
>>
>> - lack of CR8 and EFER
>>
>> - lack of XFD
>>
> 
> More for the list:
> 
>    - SVM is effectively restricted to PAE kernels due to NX requirements

True, but I'm not sure how it complicates the code?  Do you mean having 
to kmap/kunmap, and if so are you thinking of making KVM depend on !HIGHMEM?

>> - impossibility of writing 64-bit PTEs atomically
> 
> It's not impossible, just ugly.  KVM could use CMPXCHG8B to do all of the accesses
> for the TDP MMU, including the non-atomic reads and writes.

Ok, rephrased:

==========
Breakage in 32-bit KVM is somewhat rare, but it did happen and
developers themselves found out a few months later.  This means that it
is very unlikely that it has any users.  32-bit processors with
virtualization extensions are a historical curiosity; 32-bit userspace
can only deal with small guests due to limited address space.

Hence, it makes sense to restrict x86 KVM to 64-bit hosts and kernels.
This removes all the conditional code to deal with the above
differences, and it also enables unconditional use of the TDP MMU:
making it work on 32-bit systems would require contortions using
CMPXCHG8B to write 64-bit PTEs atomically, and they are simply not
worth it.
==========

>> The last is the big one, because it prevents from using the TDP MMU
>> unconditionally.
> 
> As above, if the TDP MMU really is the sticking point, that's solvable.

Yeah, but until now the maintainability cost of keeping 32-bit on life 
support was small.  Using CMPXCHG8B in the TDP MMU is possibly worse 
than having two MMUs for the two-dimensional paging case, therefore 
dropping the old x86 direct MMU for me is what tips the balance in favor 
of removal.

(Once David submits his MMU callbacks work, I still want to see if it's 
possible to preserve the "old" x86 direct MMU, for example for use in 
pKVM; however, it would not be in arch/x86/).

Paolo

> The real justification for deprecating 32-bit KVM is that, outside of KVM developers,
> literally no one uses 32-bit KVM.  I.e. any amount of effort that is required to
> continue supporting 32-bit kernels is a complete waste of resources.
>