[Beignet] [PATCH] intel: Check that we can reserve the zero-offset

['Chris Wilson']

On Sun, Nov 20, 2016 at 12:14:50PM +0000, Song, Ruiling wrote:
> 
> 
> > -----Original Message-----
> > From: Beignet [mailto:beignet-bounces at lists.freedesktop.org] On Behalf Of
> > Chris Wilson
> > Sent: Monday, November 14, 2016 9:35 PM
> > To: beignet at lists.freedesktop.org
> > Cc: Chris Wilson <chris at chris-wilson.co.uk>
> > Subject: [Beignet] [PATCH] intel: Check that we can reserve the zero-offset
> > 
> > commit ff57cee0519d ("ocl20/runtime: take the first 64KB page table
> > entries") tries to allocate a bo at 0 offset, but failed to take into
> > account that something may already be allocated there that it is not
> > allowed to evict (particularly when not using full-ppgtt separation).
> > Failure to do so causes all execution to subsequentally fail with
> > "drm_intel_gem_bo_context_exec() failed: Device or resource busy"
> > 
> > Reported-by: Kenneth Johansson <ken at kenjo.org>
> > Fixes: https://bugs.freedesktop.org/show_bug.cgi?id=98647
> > Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> > ---
> >  src/intel/intel_driver.c | 12 ++++++++----
> >  1 file changed, 8 insertions(+), 4 deletions(-)
> > 
> > diff --git a/src/intel/intel_driver.c b/src/intel/intel_driver.c
> > index 363e018..a6074ae 100644
> > --- a/src/intel/intel_driver.c
> > +++ b/src/intel/intel_driver.c
> > @@ -143,10 +143,14 @@ intel_driver_context_init(intel_driver_t *driver)
> >  #ifdef HAS_BO_SET_SOFTPIN
> >    drm_intel_bo *bo = dri_bo_alloc(driver->bufmgr, "null_bo", 64*1024, 4096);
> >    drm_intel_bo_set_softpin_offset(bo, 0);
> > -  // don't reuse it, that would make two bo trying to bind to same address,
> > -  // which is un-reasonable.
> > -  drm_intel_bo_disable_reuse(bo);
> Please don't remove the " drm_intel_bo_disable_reuse()". when the bo is reused, the soft-pinned flag is not cleared in libdrm.
> So when it is reused as an ordinary bo(not soft-pinned), it will again ask the kmd to bind it to the old zero offset which is not as expected.

Then fix libdrm. Better yet, just don't use libdrm_intel. But that is a
really nasty bug that should have been filed and fixed long ago.

> > -  driver->null_bo = bo;
> > +  drm_intel_bo_map(bo, true);
> > +  *(uint32_t *)bo->virtual = MI_BATCH_BUFFER_END;
> > +  if (drm_intel_gem_bo_context_exec(bo, driver->ctx, 0, 0) == 0) {
> > +    drm_intel_bo_map(bo, true);
> > +    driver->null_bo = memset(bo->virtual, 0, 64*1024);
> And the null_bo should points to the 'bo', but you let it points to bo->virtual, any reason?
> I notice you mapped the bo twice without unmap. Do we need unmap() here?
> I am sorry for the late reply.

No, thats just me trying to prevent an information leak at the same
time.
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre