Re: [PATCH v6 3/5] mm: introduce mmap3 for safely defining new mmap flags

[Helge Deller]

On 26.08.2017 17:15, Dan Williams wrote:
> On Sat, Aug 26, 2017 at 12:40 AM, Helge Deller <deller@xxxxxx> wrote:
>> * Dan Williams <dan.j.williams@xxxxxxxxx>:
>>> On Fri, Aug 25, 2017 at 9:19 AM, Helge Deller <deller@xxxxxx> wrote:
>>>> On 25.08.2017 18:16, Kirill A. Shutemov wrote:
>>>>> On Fri, Aug 25, 2017 at 09:02:36AM -0700, Christoph Hellwig wrote:
>>>>>> On Fri, Aug 25, 2017 at 06:58:03PM +0300, Kirill A. Shutemov wrote:
>>>>>>> Not all archs are ready for this:
>>>>>>>
>>>>>>> arch/parisc/include/uapi/asm/mman.h:#define MAP_TYPE    0x03            /* Mask for type of mapping */
>>>>>>> arch/parisc/include/uapi/asm/mman.h:#define MAP_FIXED   0x04            /* Interpret addr exactly */
>>>>>>
>>>>>> I'd be happy to say that we should not care about parisc for
>>>>>> persistent memory.  We'll just have to find a way to exclude
>>>>>> parisc without making life too ugly.
>>>>>
>>>>> I don't think creapling mmap() interface for one arch is the right way to
>>>>> go. I think the interface should be universal.
>>>>>
>>>>> I may imagine MAP_DIRECT can be useful not only for persistent memory.
>>>>> For tmpfs instead of mlock()?
>>>>
>>>> On parisc we have
>>>> #define MAP_SHARED      0x01            /* Share changes */
>>>> #define MAP_PRIVATE     0x02            /* Changes are private */
>>>> #define MAP_TYPE        0x03            /* Mask for type of mapping */
>>>> #define MAP_FIXED       0x04            /* Interpret addr exactly */
>>>> #define MAP_ANONYMOUS   0x10            /* don't use a file */
>>>>
>>>> So, if you need a MAP_DIRECT, wouldn't e.g.
>>>> #define MAP_DIRECT      0x08
>>>> be possible (for parisc, and others 0x04).
>>>> And if MAP_TYPE needs to include this flag on parisc:
>>>> #define MAP_TYPE        (0x03 | 0x08)  /* Mask for type of mapping */
>>>
>>> The problem here is that to support new the mmap flags the arch needs
>>> to find a flag that is guaranteed to fail on older kernels. Defining
>>> MAP_DIRECT to 0x8 on parisc doesn't work because it will simply be
>>> ignored on older parisc kernels.
>>>
>>> However, it's already the case that several archs have their own
>>> sys_mmap entry points. Those archs that can't follow the common scheme
>>> (only parsic it seems) will need to add a new mmap syscall. I think
>>> that's a reasonable tradeoff to allow every other architecture to add
>>> this support with their existing mmap syscall paths.
>>
>> I don't want other architectures to suffer just because of parisc.
>> But adding a new syscall just for usage on parisc won't work either,
>> because nobody will add code to call it then.
> 
> I don't understand this comment, if / when parisc gets around to
> adding pmem and dax support why wouldn't libc grow support for the new
> parisc mmap variant? Also, it's not just MAP_DIRECT you would also
> need space for a MAP_SYNC flag.
> 
>>> That means MAP_DIRECT should be defined to MAP_TYPE on parisc until it
>>> later defines an opt-in mechanism to a new syscall that honors
>>> MAP_DIRECT as a valid flag.
>>
>> I'd instead propose to to introduce an ABI breakage for parisc users
>> (which aren't many). Most parisc users update their kernel regularily
>> anyway, because we fixed so many bugs in the latest kernel.
>>
>> With the following patch pushed down to the stable kernel series,
>> MAP_DIRECT will fail as expected on those kernels, while we can
>> keep parisc up with current developments regarding MAP_DIRECT.
> 
> The whole point is to avoid an ABI regression and the chance for false
> positive results. We're immediately stuck if some application was
> expecting 0x8 to be ignored, or conversely an application that
> absolutely needs to rely on MAP_SYNC/MAP_DIRECT semantics assumes the
> wrong result on a parisc kernel where they are ignored.
> 
> I have not seen any patches for parisc pmem+dax enabling so it seems
> too early to worry about these "last mile" enabling features of
> MAP_DIRECT and MAP_SYNC. In particular parisc doesn't appear to have
> ARCH_ENABLE_MEMORY_HOTPLUG, so as far as I can see it can't yet
> support the ZONE_DEVICE scheme that is a pre-requisite for MAP_DIRECT.

I see, but then it's probably best to not to define any MAP_DIRECT or 
MAP_SYNC at all in the headers of those arches which don't support
pmem+dax (parisc, m68k, alpha, and probably quite some others).
That way applications can detect at configure time if the platform
supports that, and can leave out the functionality completely.

Helge