Adreno crash on i.MX53 running 5.3-rc6

[Rob Clark]

On Tue, Sep 3, 2019 at 12:31 PM Fabio Estevam <festevam at gmail.com> wrote:
>
> Hi Jonathan,
>
> On Tue, Sep 3, 2019 at 4:25 PM Jonathan Marek <jonathan at marek.ca> wrote:
> >
> > Hi,
> >
> > I tried this and it works with patches 4+5 from Rob's series and
> > changing gpummu to use sg_phys(sg) instead of sg->dma_address
> > (dma_address isn't set now that dma_map_sg isn't used).
>
> Thanks for testing it. I haven't had a chance to test it yet.
>
> Rob,
>
> I assume your series is targeted to 5.4, correct?

maybe, although Christoph Hellwig didn't seem like a big fan of
exposing cache ops, and would rather add a new allocation API for
uncached pages.. so I'm not entirely sure what the way forward will
be.

In the mean time, it is a bit ugly, but I guess something like this should work:

--------------------
diff --git a/drivers/gpu/drm/msm/msm_gem.c b/drivers/gpu/drm/msm/msm_gem.c
index 7263f4373f07..5a6a79fbc9d6 100644
--- a/drivers/gpu/drm/msm/msm_gem.c
+++ b/drivers/gpu/drm/msm/msm_gem.c
@@ -52,7 +52,7 @@ static void sync_for_device(struct msm_gem_object *msm_obj)
 {
     struct device *dev = msm_obj->base.dev->dev;

-    if (get_dma_ops(dev)) {
+    if (get_dma_ops(dev) && IS_ENABLED(CONFIG_ARM64)) {
         dma_sync_sg_for_device(dev, msm_obj->sgt->sgl,
             msm_obj->sgt->nents, DMA_BIDIRECTIONAL);
     } else {
@@ -65,7 +65,7 @@ static void sync_for_cpu(struct msm_gem_object *msm_obj)
 {
     struct device *dev = msm_obj->base.dev->dev;

-    if (get_dma_ops(dev)) {
+    if (get_dma_ops(dev) && IS_ENABLED(CONFIG_ARM64)) {
         dma_sync_sg_for_cpu(dev, msm_obj->sgt->sgl,
             msm_obj->sgt->nents, DMA_BIDIRECTIONAL);
     } else {
--------------------

BR,
-R

> If this is the case, what we should do about the i.MX5 regression on 5.3?
>
> Would a revert of the two commits be acceptable in 5.3 in order to
> avoid the regression?
>
> Please advise.
>
> Thanks