Re: [PATCH v2] mmap.2: MAP_FIXED updated documentation

[John Hubbard]

On 12/04/2017 11:08 PM, Michal Hocko wrote:
> On Mon 04-12-17 18:52:27, John Hubbard wrote:
>> On 12/04/2017 03:31 AM, Mike Rapoport wrote:
>>> On Sun, Dec 03, 2017 at 06:14:11PM -0800, john.hubbard@xxxxxxxxx wrote:
>>>> From: John Hubbard <jhubbard@xxxxxxxxxx>
>>>>
>> [...]
>>>> +.IP
>>>> +Given the above limitations, one of the very few ways to use this option
>>>> +safely is: mmap() a region, without specifying MAP_FIXED. Then, within that
>>>> +region, call mmap(MAP_FIXED) to suballocate regions. This avoids both the
>>>> +portability problem (because the first mmap call lets the kernel pick the
>>>> +address), and the address space corruption problem (because the region being
>>>> +overwritten is already owned by the calling thread).
>>>
>>> Maybe "address space corruption problem caused by implicit calls to mmap"?
>>> The region allocated with the first mmap is not exactly owned by the
>>> thread and a multi-thread application can still corrupt its memory if
>>> different threads use mmap(MAP_FIXED) for overlapping regions.
>>>
>>> My 2 cents.
>>>
>>
>> Hi Mike,
>>
>> Yes, thanks for picking through this, and I agree that the above is misleading.
>> It should definitely not use the word "owned" at all. Re-doing the whole 
>> paragraph in order to make it all fit together nicely, I get this:
>>
>> "Given the above limitations, one of the very few ways to use this option
>> safely is: mmap() an enclosing region, without specifying MAP_FIXED.
>> Then, within that region, call mmap(MAP_FIXED) to suballocate regions
>> within the enclosing region. This avoids both the portability problem 
>> (because the first mmap call lets the kernel pick the address), and the 
>> address space corruption problem (because implicit calls to mmap will 
>> not affect the already-mapped enclosing region)."
>>
>> ...how's that sound to you? I'll post a v3 soon with this.
> 
> It sounds to me you are trying to tell way to much while actually being
> a bit misleading. Even sub-range MAP_FIXED is not multi-thread safe.
> 
> Really the more corner cases you will try to cover the worse the end
> result will end up. I would just try to be simple here and mention the
> address space corruption issues you've had earlier and be done with it.
> Maybe add a note that some architectures might need a special alignement
> and fail if it is not the case but nothing really specific.
> 

Sure, I can drop the "how to use this safely" section.  It seemed like a good
idea at the time... :)

thanks,
John Hubbard
NVIDIA