[Intel-gfx] [PATCH i-g-t] tests/gem_mmap_gtt(huge_bo): Use pread more to verify domain transfer.

[Joonas Lahtinen]

On ke, 2015-05-06 at 14:29 +0100, Chris Wilson wrote:
> On Wed, May 06, 2015 at 04:22:23PM +0300, Joonas Lahtinen wrote:
> > On ke, 2015-05-06 at 14:15 +0100, Chris Wilson wrote:
> > > On Wed, May 06, 2015 at 03:31:10PM +0300, Joonas Lahtinen wrote:
> > > > 
> > > > Use pread right after moving out of CPU domain to verify all
> > > > writes flushed correctly.
> > > 
> > > Wrong test, see gem_pread. Adding extra work may have the unintended
> > > side-effects of masking bugs. If you want to mix access modes, call it a
> > > new test, and that is handled by gem_concurrent_blt (which once again is
> > > difting far beyond the original name).
> > 
> > This is specifically for the huge_bo test, where the kernel side
> > handling is expected to be different than for the others, so it stresses
> > the code path related to huge BO's.
> 
> See gem_pread. Your huge_bo merely classifies as big...
>  
> > Better wording to add to it:
> > "Make sure that the introduction of the partial GTT mappings which are
> > expected to be in use when mapping an object bigger than the aperture
> > size do not corrupt the contents of the backing storage"
> 
> But mixing pread and set-domain has strictly undefined semantics in
> terms of what domain we expect the buffer to be in afterwards i.e. to be
> correct you would have to call set-domain again after pread.
> 

Ah, ok.

> And you really should add huge_bo testing to gem_concurrent_blt.

Yeah, if the above is true, definitely. Will do so. Thanks for the
directions!

Regards, Joonas

> -Chris
>