[PATCH v3 3/5] arm64: mm: set the contiguous bit for kernel mappings where appropriate

[Catalin Marinas]

On Thu, Oct 13, 2016 at 05:57:33PM +0100, Ard Biesheuvel wrote:
> On 13 October 2016 at 17:28, Catalin Marinas <catalin.marinas at arm.com> wrote:
> > On Wed, Oct 12, 2016 at 12:23:43PM +0100, Ard Biesheuvel wrote:
> >> Now that we no longer allow live kernel PMDs to be split, it is safe to
> >> start using the contiguous bit for kernel mappings. So set the contiguous
> >> bit in the kernel page mappings for regions whose size and alignment are
> >> suitable for this.
> >>
> >> This enables the following contiguous range sizes for the virtual mapping
> >> of the kernel image, and for the linear mapping:
> >>
> >>           granule size |  cont PTE  |  cont PMD  |
> >>           -------------+------------+------------+
> >>                4 KB    |    64 KB   |   32 MB    |
> >>               16 KB    |     2 MB   |    1 GB*   |
> >>               64 KB    |     2 MB   |   16 GB*   |
> >>
> >> * only when built for 3 or more levels of translation
> >
> > I assume the limitation to have contiguous PMD only with 3 or move
> > levels is because of the way p*d folding was implemented in the kernel.
> > With nopmd, looping over pmds is done in __create_pgd_mapping() rather
> > than alloc_init_pmd().
> >
> > A potential solution would be to replicate the contiguous pmd code to
> > the pud and pgd level, though we probably won't benefit from any
> > contiguous entries at higher level (when more than 2 levels).
> > Alternatively, with an #ifdef __PGTABLE_PMD_FOLDED, we could set the
> > PMD_CONT in prot in __create_pgd_mapping() directly (if the right
> > addr/phys alignment).
> 
> Indeed. See the next patch :-)

I got there eventually ;).

> > Anyway, it's probably not worth the effort given that 42-bit VA with 64K
> > pages is becoming a less likely configuration (36-bit VA with 16K pages
> > is even less likely, also depending on EXPERT).
> 
> This is the reason I put it in a separate patch: this one contains the
> most useful combinations, and the next patch adds the missing ones,
> but clutters up the code significantly. I'm perfectly happy to drop 4
> and 5 if you don't think it is worth the trouble.

I'll have a look at patch 4 first.

Both 64KB contiguous pmd and 4K contiguous pud give us a 16GB range
which (AFAIK) is less likely to be optimised in hardware.

-- 
Catalin