[Intel-gfx] [PATCH 2/2] drm/i915: Use remap_pfn_range() to prefault all PTE in a single pass

[Daniel Vetter]

On Wed, Jun 11, 2014 at 02:17:21PM -0700, Volkin, Bradley D wrote:
> On Tue, Jun 10, 2014 at 04:14:41AM -0700, Chris Wilson wrote:
> > On an Ivybridge i7-3720qm with 1600MHz DDR3, with 32 fences,
> > Upload rate for 2 linear surfaces:  8134MiB/s -> 8154MiB/s
> > Upload rate for 2 tiled surfaces:   8625MiB/s -> 8632MiB/s
> > Upload rate for 4 linear surfaces:  8127MiB/s -> 8134MiB/s
> > Upload rate for 4 tiled surfaces:   8602MiB/s -> 8629MiB/s
> > Upload rate for 8 linear surfaces:  8124MiB/s -> 8137MiB/s
> > Upload rate for 8 tiled surfaces:   8603MiB/s -> 8624MiB/s
> > Upload rate for 16 linear surfaces: 8123MiB/s -> 8128MiB/s
> > Upload rate for 16 tiled surfaces:  8606MiB/s -> 8618MiB/s
> > Upload rate for 32 linear surfaces: 8121MiB/s -> 8128MiB/s
> > Upload rate for 32 tiled surfaces:  8605MiB/s -> 8614MiB/s
> > Upload rate for 64 linear surfaces: 8121MiB/s -> 8127MiB/s
> > Upload rate for 64 tiled surfaces:  3017MiB/s -> 5127MiB/s
> > 
> > Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> 
> The translation from vm_insert_pfn to remap_pfn_range looks correct. I don't know
> these APIs particularly well though. I wonder if there's any reason it would be
> unsafe to call remap_pfn_range from .fault() since it seems to only be used in
> .mmap() handlers in other places. Reading their implementations, nothing jumped
> out, so I'll say

So apparently wasn't quite the same semantics and remap_pfn_range doesn't
handle concurrent operations too well. Can you please also review Chris'
fixup patch for that?

Thanks, Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch