Re: [PATCH v5] arch/ia64: Restore arch-specific pgd_offset_k implementation

[Jessica Clarke <jrtc27@jrtc27.com>]

On 13 Aug 2020, at 22:36, Matthew Wilcox <willy@infradead.org> wrote:
> 
> On Thu, Aug 13, 2020 at 11:55:21PM +0300, Mike Rapoport wrote:
>> +/*
>> + * In the kernel's mapped region we know everything is in region number 5, so
>> + * as an optimisation its PGD already points to the area for that region.
> 
> Is it actually an optimisation?  Are there any benchmarks where this
> makes any difference whatsoever?  Or should the comment be closer to
> 
> /* ia64 is gratuitously different.  Cope with it. */
> 
> I might suggest that pgd be changed to be the same as every other pgd
> in the system, and every other architecture.

IA-64 is just weird. There's an optional hardware page table walker,
and two different options for how to use it within that, short-format
and long-format. My understanding is that Linux uses the short-format,
and that means that the VHPTs (Virtual Hashed Page Table) are
per-region. My assumption therefore is that the kernel's PGD is a
completely separate allocation from the userspace ones, and it's not
just a case of "moving the pointer back to the start" but that "that is
all there is", so unifying the two would be a significant undertaking
that avoids one special case at the cost of invasive refactoring and a
performance hit by duplication and/or unused space. My knowledge of
IA-64 translation is pretty limited, that's about all I know (and this
likely has mistakes in it), but I know enough to avoid finding out too
much more. Never make the mistake of assuming IA-64 can be made to look
like any other architecture :)

Jess