Re: [PATCH] /dev/zero: make private mapping full anonymous mapping

[David Hildenbrand <david@redhat.com>]

On 14.01.25 18:20, Yang Shi wrote:
> 
> 
> 
> On 1/14/25 9:02 AM, David Hildenbrand wrote:
>> On 14.01.25 16:06, David Hildenbrand wrote:
>>> On 14.01.25 15:52, Lorenzo Stoakes wrote:
>>>> On Tue, Jan 14, 2025 at 02:01:32PM +0100, David Hildenbrand wrote:
>>>>> On 13.01.25 23:30, Yang Shi wrote:
>>>>>> When creating private mapping for /dev/zero, the driver makes it an
>>>>>> anonymous mapping by calling set_vma_anonymous().  But it just sets
>>>>>> vm_ops to NULL, vm_file is still valid and vm_pgoff is also file
>>>>>> offset.
>>>>>>
>>>>>> This is a special case and the VMA doesn't look like either
>>>>>> anonymous VMA
>>>>>> or file VMA.  It confused other kernel subsystem, for example,
>>>>>> khugepaged [1].
>>>>>>
>>>>>> It seems pointless to keep such special case.  Making private
>>>>>> /dev/zero>
>>>>> mapping a full anonymous mapping doesn't change the semantic of
>>>>>> /dev/zero either.
>>>>>>
>>>>>> The user visible effect is the mapping entry shown in
>>>>>> /proc/<PID>/smaps
>>>>>> and /proc/<PID>/maps.
>>>>>>
>>>>>> Before the change:
>>>>>> ffffb7190000-ffffb7590000 rw-p 00001000 00:06
>>>>>> 8                          /dev/zero
>>>>>>
>>>>>> After the change:
>>>>>> ffffb6130000-ffffb6530000 rw-p 00000000 00:00 0
>>>>>>
>>>>>
>>>>> Hm, not sure about this. It's actually quite consistent to have
>>>>> that output
>>>>> in smaps the way it is. You mapped a file at an offset, and it
>>>>> behaves like
>>>>> an anonymous mapping apart from that.
>>>>>
>>>>> Not sure if the buggy khugepaged thing is a good indicator to
>>>>> warrant this
>>>>> change.
>>>>
>>>> Yeah, this is a user-facing fundamental change that hides
>>>> information and
>>>> defies expectation so I mean - it's a no go really isn't it?
>>>>
>>>> I'd rather we _not_ make this anon though, because isn't life confusing
>>>> enough David? I thought it was bad enough with 'anon, file and lol
>>>> shmem'
>>>> but 'lol lol also /dev/zero' is enough to make me want to frolick in
>>>> the
>>>> fields...
>>>
>>> I recall there are users that rely on this memory to get the shared
>>> zeropage on reads etc (in comparison to shmem!), so I better not ...
>>> mess with this *at all* :)
>>
>> Heh, and I recall reading something about odd behavior of /dev/zero
>> and some interesting history [1].
>>
>> "
>> Unlike /dev/null, /dev/zero may be used as a source, not only as a
>> sink for data. All write operations to /dev/zero succeed with no other
>> effects. However, /dev/null is more commonly used for this purpose.
>>
>> When /dev/zero is memory-mapped, e.g., with mmap, to the virtual
>> address space, it is equivalent to using anonymous memory; i.e. memory
>> not connected to any file.
>> "
>>
>> "equivalent to using anonymous memory" is interesting.
> 
> For private mapping. Shared mapping is equivalent to shmem.

"shared anonymous memory", yes.

-- 
Cheers,

David / dhildenb