[Intel-gfx] [PATCH 2/4] drm/i915/gem: Sync the vmap PTEs upon construction

[Chris Wilson]

Quoting Linus Torvalds (2020-08-21 13:41:03)
> On Fri, Aug 21, 2020 at 1:50 AM Chris Wilson <chris at chris-wilson.co.uk> wrote:
> >
> > Since synchronising the PTE after assignment is a manual step, use the
> > newly exported interface to flush the PTE after assigning via
> > alloc_vm_area().
> 
> This commit message doesn't make much sense to me.
> 
> Are you talking about synchronizing the page directory structure
> across processes after possibly creating new kernel page tables?
> 
> Because that has nothing to do with the PTE. It's all about making
> sure the _upper_ layers of the page directories are populated
> everywhere..
> 
> The name seems off to me too - what are you "flushing"? (And yes, I
> know about the flush_cache_vmap(), but that looks just bogus, since
> any non-mapped area shouldn't have any virtual caches to begin with,
> so I suspect that is just the crazy architectures being confused -
> flush_cache_vmap() is a no-op on any sane architecture - and powerpc
> that mis-uses it for other things).

I was trying to mimic map_kernel_range() which does the
arch_sync_kernel_mappings and flush_cache_vmap on top of the
apply_to_page_range performed by alloc_vm_area, because buried away in
there, on x86-32, is a set_pmd(). Since map_kernel_range() wrapped
map_kernel_range_noflush(), flush seemed like the right verb.

Joerg pointed out that the sync belonged to __apply_to_page_range and
fixed it in situ.
-Chris