Re: [PATCH 3/3] mm: make pagoff_t type 64-bit

[Arnd Bergmann]

On Friday, December 16, 2016 4:25:14 PM CET Yury Norov wrote:
> On Sun, Dec 11, 2016 at 03:59:01PM +0100, Arnd Bergmann wrote:
> > On Sunday, December 11, 2016 6:26:42 PM CET Yury Norov wrote:
> > > Also fix related interfaces
> > > 
> > > Signed-off-by: Yury Norov <ynorov@xxxxxxxxxxxxxxxxxx>
> > 
> > Thanks Yury for the demonstration. I think this would put the nail
> > in the coffin of the idea of mmap64 even for Pavel, who didn't
> > seem convinced already.
> > 
> > Changing all those interfaces and structure, struct page in particular,
> > is clearly too costly for any advantage we might have otherwise
> > gained.
> > 
> >       Arnd
> 
> To be complete, we have 3 options:
>  1 leave things as is. 32-bit architectures will have no option to
>    mmap big offsets, and no one cares - as usual.
>  2 add mmap64() for compat arches only. This way we don't need patch
>    3, and arches like aarch32 or aarch64/ilp32 will enjoy true 64-bit
>    offsets.
>  3 introduce CONFIG_64_BIT_PGOFF_T, and let Pavel enable it if he has
>    to work with big files on 32-bit arches.
> 
> The most realistic approach for me is 1 because I never heard about
> 64-bit pgoff_t requests, except Pavel's one. Thinking about
> aarch64/ilp32, we probably need second approach. This is only 2 simple
> patches that are already there, and one patch in glibc. It will let
> 32-bit software work in 64-bit environment more smoothly. Cavium
> people should be completely satisfied with 2.

Agreed: If there is a serious request from Cavium or Huawei (which
are also very interested in this feature) and a specific use case,
we can still do 2 easily.

> Third is more looking like research exercise than something we need
> in practice.

Right.

> The only thing that makes me sad is that we proudly declare 64-bit
> off_t in new 32-bit ABIs but in fact we lie, at least in this
> specific case. We should add corresponding checks on glibc side at
> least. It's also simple.

Well, the only thing we are really saying there is that we support
more than 32-bit, and that the ABI uses 64-bit. Actually doing 64-bit
offsets within (very sparse) files probably also fails on 64-bit
architectures, at least on some file systems.

	Arnd