Re: [RFC PATCH] mm: shm: round up tmpfs size to huge page size when huge=always

[Yang Shi]

On 10/9/17 10:26 AM, Michal Hocko wrote:
> On Tue 10-10-17 00:43:31, Yang Shi wrote:
> > On 10/8/17 11:48 PM, Michal Hocko wrote:
> > > On Sun 08-10-17 15:56:51, Kirill A. Shutemov wrote:
> > > > On Sat, Oct 07, 2017 at 04:22:10AM +0800, Yang Shi wrote:
> > > > > When passing "huge=always" option for mounting tmpfs, THP is supposed to
> > > > > be allocated all the time when it can fit, but when the available space is
> > > > > smaller than the size of THP (2MB on x86), shmem fault handler still tries
> > > > > to allocate huge page every time, then fallback to regular 4K page
> > > > > allocation, i.e.:
> > > > >
> > > > > 	# mount -t tmpfs -o huge,size=3000k tmpfs /tmp
> > > > > 	# dd if=/dev/zero of=/tmp/test bs=1k count=2048
> > > > > 	# dd if=/dev/zero of=/tmp/test1 bs=1k count=2048
> > > > >
> > > > > The last dd command will handle 952 times page fault handler, then exit
> > > > > with -ENOSPC.
> > > > >
> > > > > Rounding up tmpfs size to THP size in order to use THP with "always"
> > > > > more efficiently. And, it will not wast too much memory (just allocate
> > > > > 511 extra pages in worst case).
> > > >
> > > > Hm. I don't think it's good idea to silently increase size of fs.
> > >
> > > Agreed!
> > >
> > > > Maybe better just refuse to mount with huge=always for too small fs?
> > >
> > > We cannot we simply have the remaining page !THP? What is the actual
>
> ups s@We@Why@
>
> > > problem?
> >
> > The remaining pages can be !THP, it will fall back to regular 4k pages when
> > the available space is less than THP size.
> >
> > I just wonder it sounds not make sense to *not* mount tmpfs with THP size
> > alignment when "huge=always" is passed.
>
> yes failure seems overly excessive reaction to me.
>
> > I guess someone would like to assume all allocation in tmpfs with
> > "huge=always" should be THP.
>
> Nobody can assume that because THP pages can be broken up at any point
> in time. We have hugetlb to provide a guarantee
>
> > But, they might not be fully aware of in some
> > corner cases THP might be not used, for example, the remaining space is less
> > then THP size, then some unexpected performance degrade might be perceived.
> >
> > So, why not we do the mount correctly at the first place. It could be
> > delegated to the administrator, but it should be better to give some hint
> > from kernel side.
>
> Because we are not trying to be more clever than the user. I still do
> not see what is the actual problem you are trying to fix to be honest.

Just try to provide a warning or hint to the users that it'd better to 
mount tmpfs with THP size aligned when "huge=always" is passed to avoid 
some unexpected performance degrade.

Resizing or failure to mount might be overkill, documenting it might be 
good enough.

Thanks,
Yang

>