[Intel-gfx] [RFC PATCH i-g-t v4 4/4] tests/gem_ctx_shared: Align objects using minimum GTT alignment

Janusz Krzysztofik janusz.krzysztofik at linux.intel.com
Mon Nov 4 09:39:32 UTC 2019


Hi Chris,

On Friday, November 1, 2019 10:42:18 AM CET Chris Wilson wrote:
> Quoting Janusz Krzysztofik (2019-10-31 15:28:57)
> > exec-shared-gtt-* subtests use hardcoded values for object size and
> > softpin offset, based on 4kB GTT alignment assumption.  That may result
> > in those subtests failing when run on future backing stores with
> > possibly larger minimum page sizes.
> > 
> > Replace hardcoded constants with values calculated from minimum GTT
> > alignment of actual backing store the test is running on.
> > 
> > v2: Update helper name, use 'minimum GTT alignment' term across the
> >     change, adjust variable name.
> > 
> > Signed-off-by: Janusz Krzysztofik <janusz.krzysztofik at linux.intel.com>
> > Cc: Chris Wilson <chris at chris-wilson.co.uk>
> > ---
> >  tests/i915/gem_ctx_shared.c | 6 ++++--
> >  1 file changed, 4 insertions(+), 2 deletions(-)
> > 
> > diff --git a/tests/i915/gem_ctx_shared.c b/tests/i915/gem_ctx_shared.c
> > index 6d8cbcce..1e9c7f78 100644
> > --- a/tests/i915/gem_ctx_shared.c
> > +++ b/tests/i915/gem_ctx_shared.c
> > @@ -195,6 +195,7 @@ static void exec_shared_gtt(int i915, unsigned int 
ring)
> >         uint32_t scratch, *s;
> >         uint32_t batch, cs[16];
> >         uint64_t offset;
> > +       uint64_t alignment;
> >         int i;
> >  
> >         gem_require_ring(i915, ring);
> > @@ -203,7 +204,8 @@ static void exec_shared_gtt(int i915, unsigned int 
ring)
> >         clone = gem_context_clone(i915, 0, I915_CONTEXT_CLONE_VM, 0);
> >  
> >         /* Find a hole big enough for both objects later */
> > -       scratch = gem_create(i915, 16384);
> > +       alignment = 2 * gem_gtt_min_alignment(i915);

alignment = one page for an object + one page for stride

> > +       scratch = gem_create(i915, 2 * alignment);

space reserved = 2 * (one page for an object + one page for stride)

> >         gem_write(i915, scratch, 0, &bbe, sizeof(bbe));
> >         obj.handle = scratch;
> >         gem_execbuf(i915, &execbuf);
> > @@ -246,7 +248,7 @@ static void exec_shared_gtt(int i915, unsigned int 
ring)
> >         gem_write(i915, batch, 0, cs, sizeof(cs));
> >  
> >         obj.handle = batch;
> > -       obj.offset += 8192; /* make sure we don't cause an eviction! */
> > +       obj.offset += alignment; /* make sure we don't cause an eviction! 
*/

offset += (one page for an object + one page for stride)

> It's 'stride' here. It's leaving a guard page in between, just in case
> page coloring demands it.

Assuming I correctly understood the intentions here but my implementation 
is not readable clearly enough, how do you think this should be fixed?  How 
about a comment '/* one page for an object + one page for stride */' above the 
line where the 'alignment' variable is assigned the value of 2 * minimum GTT 
alignment?

Thanks,
Janusz

> -Chris
> 






More information about the Intel-gfx mailing list