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: <40486dbb-9f19-6fa6-d46d-99d2b033883d@redhat.com>
Date:   Tue, 6 Sep 2022 16:44:20 +0200
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:102.0) Gecko/20100101
 Thunderbird/102.2.0
Content-Language: en-US
To:     Jason Gunthorpe <jgg@nvidia.com>
Cc:     John Hubbard <jhubbard@nvidia.com>, Yang Shi <shy828301@gmail.com>,
        peterx@redhat.com, kirill.shutemov@linux.intel.com,
        hughd@google.com, akpm@linux-foundation.org, linux-mm@kvack.org,
        linux-kernel@vger.kernel.org
References: <20220901222707.477402-1-shy828301@gmail.com>
 <e6ad1084-c301-9f11-1fa7-7614bf859aaf@nvidia.com>
 <a969abc5-1ad0-4073-a1f9-82f0431a0104@redhat.com>
 <YxdPi2E63aO0Dgyd@nvidia.com>
 <4516a349-49cb-fd7b-176a-f1a9479906d9@redhat.com>
 <YxdZlCly2ad1rtcI@nvidia.com>
From:   David Hildenbrand <david@redhat.com>
Organization: Red Hat
Subject: Re: [PATCH] mm: gup: fix the fast GUP race against THP collapse
In-Reply-To: <YxdZlCly2ad1rtcI@nvidia.com>
Content-Type: text/plain; charset=UTF-8; format=flowed
Content-Transfer-Encoding: 7bit
Precedence: bulk

On 06.09.22 16:30, Jason Gunthorpe wrote:
> On Tue, Sep 06, 2022 at 03:57:30PM +0200, David Hildenbrand wrote:
> 
>>> READ_ONCE primarily is a marker that the data being read is unstable
>>> and that the compiler must avoid all instability when reading it. eg
>>> in this case the compiler could insanely double read the value, even
>>> though the 'if' requires only a single read. This would result in
>>> corrupt calculation.
>>
>> As we have a full memory barrier + compile barrier, the compiler might
>> indeed do double reads and all that stuff. BUT, it has to re-read after we
>> incremented the refcount, and IMHO that's the important part to detect the
>> change.
> 
> Yes, it is important, but it is not the only important part.
> 
> The compiler still has to exectute "if (*a != b)" *correctly*.
> 
> This is what READ_ONCE is for. It doesn't set order, it doesn't
> implement a barrier, it tells the compiler that '*a' is unstable data
> and the compiler cannot make assumptions based on the idea that
> reading '*a' multiple times will always return the same value.
> 
> If the compiler makes those assumptions then maybe even though 'if (*a
> != b)' is the reality, it could mis-compute '*a == b'. You enter into
> undefined behavior here.
> 
> Though it is all very unlikely, the general memory model standard is
> to annotate with READ_ONCE.

The only thing I could see going wrong in the comparison once the stars 
alingn would be something like the following:

if (*a != b)

implemented as

if ((*a).lower != b.lower && (*a).higher != b.higher)


This could only go wrong if we have more than one change such that:

Original:

*a = 0x00000000ffffffffull;


First modification:
*a = 0xffffffffffffffffull;

Second modification:
*a = 0x00000000eeeeeeeeull;


If we race with both modifications, we could see that ffffffff matches, 
and could see that 00000000 matches as well.


So I agree that we should change it, but not necessarily as an urgent 
fix and not necessarily in this patch. It's best to adjust all gup_* 
functions in one patch.

... I do wonder if we want to reuse ptep_get_lockless() instead of the 
READ_ONCE(). CONFIG_GUP_GET_PTE_LOW_HIGH is confusing.

-- 
Thanks,

David / dhildenb