On Tue, Feb 27, 2024, at 12:12, Geert Uytterhoeven wrote:
On Tue, Feb 27, 2024 at 11:59 AM Arnd Bergmann <arnd@xxxxxxxx> wrote:
On Tue, Feb 27, 2024, at 09:54, Geert Uytterhoeven wrote:
I was a bit unsure about how to best do this since there
is not really a need for a fixed page size on nommu kernels,
whereas the three MMU configs clearly tie the page size to
the MMU rather than the platform.
There should be no reason for coldfire to have a different
page size from dragonball if neither of them actually uses
hardware pages, so one of them could be changed later.
Indeed, in theory, PAGE_SIZE doesn't matter for nommu, but the concept
of pages is used all over the place in Linux.
I'm mostly worried about some Coldfire code relying on the actual value
of PAGE_SIZE in some other context. e.g. for configuring non-cacheable
regions.
Right, any change here would have to be carefully tested. I would
expect that a 4K page size would reduce memory consumption even on
NOMMU systems that should have the same tradeoffs for representing
files in the page cache and in mem_map[].
And does this impact running nommu binaries on a system with MMU?
I.e. if nommu binaries were built with a 4 KiB PAGE_SIZE, do they
still run on MMU systems with an 8 KiB PAGE_SIZE (coldfire and sun3),
or are there some subtleties to take into account?
As far as I understand, binaries have to be built and linked for
the largest page size they can run on, so running them on a kernel
with smaller page size usually works.
One notable exception is sys_mmap2(), which on most architectures
takes units of 4KiB but on m68k is actually written to take
PAGE_SIZE units. As Al pointed out in f8b7256096a2 ("Unify
sys_mmap*"), it has always been wrong on sun3, presumably
because users of that predate modern glibc. Running coldfire
nommu binaries on coldfire mmu kernels would run into the same
bug if either of them changes PAGE_SIZE. If you can run
coldfire nommu binaries on classic m68k, that is already
broken in the same way.
Arnd
