On 7 January 2016 at 14:55, Mark Rutland [off-list ref] wrote:

On Wed, Dec 30, 2015 at 04:26:02PM +0100, Ard Biesheuvel wrote:

quoted

Before introducing new statically allocated page tables and increasing
their alignment in subsequent patches, update the reservation logic
so that only pages that are in actual use end up as reserved with
memblock.

Could you add something to the commit message about what this will gain
us (i.e. which pages we don't have to reserve)? It's not immediately
obvious why we'd have page tables we wouldn't want to reserve.

OK. In the original series, I also aligned the pgdir section to a log2
upper bound of its size, but that is not necessary anymore with your
changes. So the original goal was to avoid reserving the alignment
padding as well as the pgdirs that end up unused

From the looks of the next patch we won't have redundant levels of
fixmap table for a given configuration, so I guess we're catering for
the case the fixmap shares a pgd/pud/pmd entry with the image mapping?

Does that happen? If so that would invalidate the assumption I make when
copying the fixmap over in [1] (see map_kernel).

It is a lot less likely to happen now that I moved the kernel to the
start of the vmalloc area rather than right below PAGE_OFFSET. But in
general, it seems sensible to only populate entries after confirming
that they are in fact vacant.

To handle that either we need some special logic to copy over the
relevant bits for the fixmap (as with kasan_copy_shadow), or we need to
avoid sharing a pgd entry.

Thoughts?

Yes, I have added that to my v3 version of the vmalloc base move patch here

https://git.linaro.org/people/ard.biesheuvel/linux-arm.git/commitdiff/0beef2c1a6bfc90cc116a6ba1b24f2ba35e7e5f6

but I think 16k/4 levels is the only config affected when the kernel
is always in the lower half of the vmalloc area. That also implies
that the fixmap pgd is either always shared, or never, depending on
the build time config, so I could probably simplify that part
somewhat.

-- 
Ard.