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: <67222fe0-7bf0-ec7a-0791-a4d48391a15e@redhat.com>
Date:   Fri, 29 Apr 2022 16:45:03 +0200
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:91.0) Gecko/20100101
 Thunderbird/91.8.0
Content-Language: en-US
To:     Sean Christopherson <seanjc@google.com>
Cc:     Mingwei Zhang <mizhang@google.com>,
        Vitaly Kuznetsov <vkuznets@redhat.com>,
        Wanpeng Li <wanpengli@tencent.com>,
        Jim Mattson <jmattson@google.com>,
        Joerg Roedel <joro@8bytes.org>, kvm@vger.kernel.org,
        linux-kernel@vger.kernel.org
References: <20220429031757.2042406-1-mizhang@google.com>
 <4b0936bf-fd3e-950a-81af-fd393475553f@redhat.com>
 <Ymv3vwBEgCH0CMPH@google.com>
From:   Paolo Bonzini <pbonzini@redhat.com>
Subject: Re: [PATCH] KVM: x86/mmu: fix potential races when walking host page
 table
In-Reply-To: <Ymv3vwBEgCH0CMPH@google.com>
Content-Type: text/plain; charset=UTF-8; format=flowed
Content-Transfer-Encoding: 7bit
Precedence: bulk

On 4/29/22 16:35, Sean Christopherson wrote:
> On Fri, Apr 29, 2022, Paolo Bonzini wrote:
>>> +out:
>>> +	local_irq_restore(flags);
>>> +	return level;
>>> +}
>>> +EXPORT_SYMBOL_GPL(kvm_lookup_address_level_in_mm);
>>
>> Exporting is not needed.
>>
>> Thanks for writing the walk code though.  I'll adapt it and integrate the
>> patch.
> 
> But why are we fixing this only in KVM?  I liked the idea of stealing perf's
> implementation because it was a seemlingly perfect fit and wouldn't introduce
> new code (ignoring wrappers, etc...).
> 
> We _know_ that at least one subsystem is misusing lookup_address_in_pgd() and
> given that its wrappers are exported, I highly doubt KVM is the only offender.
> It really feels like we're passing the buck here by burying the fix in KVM.

There are two ways to do it:

* having a generic function in mm/.  The main issue there is the lack of 
a PG_LEVEL_{P4D,PUD,PMD,PTE} enum at the mm/ level.  We could use 
(ctz(x) - 12) / 9 to go from size to level, but it's ugly and there 
could be architectures with heterogeneous page table sizes.

* having a generic function in arch/x86/.  In this case KVM seems to be 
the odd one that doesn't need the PTE.  For example vc_slow_virt_to_phys 
needs the PTE, and needs the size rather than the "level" per se.

So for now I punted, while keeping open the door for moving code from 
arch/x86/kvm/ to mm/ if anyone else (even other KVM ports) need the same 
logic.

Paolo