[PATCH v5 1/4] riscv: Move kernel mapping to vmalloc zone

[Arnd Bergmann]

On Fri, Jul 24, 2020 at 12:34 AM Benjamin Herrenschmidt
<benh at kernel.crashing.org> wrote:
> On Thu, 2020-07-23 at 01:21 -0400, Alex Ghiti wrote:
> > > works fine with huge pages, what is your problem there ? You rely on
> > > punching small-page size holes in there ?
> > >
> >
> > ARCH_HAS_STRICT_KERNEL_RWX prevents the use of a hugepage for the kernel
> > mapping in the direct mapping as it sets different permissions to
> > different part of the kernel (data, text..etc).
>
> Ah ok, that can be solved in a couple of ways...
>
> One is to use the linker script to ensure those sections are linked
> HUGE_PAGE_SIZE appart and moved appropriately by early boot code. One
> is to selectively degrade just those huge pages.
>
> I'm not familiar with the RiscV MMU (I should probably go have a look)
> but if it's a classic radix tree with huge pages at PUD/PMD level, then
> you could just degrade the one(s) that cross those boundaries.

That would work, but if the system can otherwise use 1GB-sized pages,
that might mean degrading the first gigabyte into a mix of 2MB pages
and 4KB pages.

If the kernel is in vmalloc space and vmap is able to use 2MB pages
for contiguous chunks of the mapping, you get a somewhat better
TLB usage.

However, this also means that a writable mapping exists in the
linear mapping for any executable part of the kernel (.text in
both vmlinux and modules). Do we have that on other architectures
as well, or is this something that ought to be prevented with
STRICT_KERNEL_RWX/STRICT_MODULE_RWX?

     Arnd