Re: [PATCH v2 02/13] drm/prime: Provide default ::{begin,end}_cpu_access() implementations

[Boris Brezillon <boris.brezillon@collabora.com>]

On Fri, 10 Oct 2025 16:34:14 +0100
Steven Price <steven.price@arm.com> wrote:

> On 10/10/2025 16:03, Boris Brezillon wrote:
> > On Fri, 10 Oct 2025 15:11:54 +0100
> > Steven Price <steven.price@arm.com> wrote:
> >   
> >> On 10/10/2025 11:11, Boris Brezillon wrote:  
> >>> Hook-up drm_gem_dmabuf_{begin,end}_cpu_access() to drm_gem_sync() so
> >>> that drivers relying on the default prime_dmabuf_ops can still have
> >>> a way to prepare for CPU accesses from outside the UMD.
> >>>
> >>> v2:
> >>> - New commit
> >>>
> >>> Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
> >>> ---
> >>>  drivers/gpu/drm/drm_prime.c | 36 ++++++++++++++++++++++++++++++++++++
> >>>  include/drm/drm_prime.h     |  5 +++++
> >>>  2 files changed, 41 insertions(+)
> >>>
> >>> diff --git a/drivers/gpu/drm/drm_prime.c b/drivers/gpu/drm/drm_prime.c
> >>> index 43a10b4af43a..03c09f9ab129 100644
> >>> --- a/drivers/gpu/drm/drm_prime.c
> >>> +++ b/drivers/gpu/drm/drm_prime.c
> >>> @@ -823,6 +823,40 @@ int drm_gem_dmabuf_mmap(struct dma_buf *dma_buf, struct vm_area_struct *vma)
> >>>  }
> >>>  EXPORT_SYMBOL(drm_gem_dmabuf_mmap);
> >>>  
> >>> +int drm_gem_dmabuf_begin_cpu_access(struct dma_buf *dma_buf,
> >>> +				    enum dma_data_direction direction)
> >>> +{
> >>> +	struct drm_gem_object *obj = dma_buf->priv;
> >>> +	enum drm_gem_object_access_flags access = DRM_GEM_OBJECT_CPU_ACCESS;
> >>> +
> >>> +	if (direction == DMA_FROM_DEVICE)
> >>> +		access |= DRM_GEM_OBJECT_READ_ACCESS;
> >>> +	else if (direction == DMA_BIDIRECTIONAL)
> >>> +		access |= DRM_GEM_OBJECT_RW_ACCESS;
> >>> +	else
> >>> +		return -EINVAL;
> >>> +
> >>> +	return drm_gem_sync(obj, 0, obj->size, access);
> >>> +}
> >>> +EXPORT_SYMBOL(drm_gem_dmabuf_begin_cpu_access);
> >>> +
> >>> +int drm_gem_dmabuf_end_cpu_access(struct dma_buf *dma_buf,
> >>> +				  enum dma_data_direction direction)
> >>> +{
> >>> +	struct drm_gem_object *obj = dma_buf->priv;
> >>> +	enum drm_gem_object_access_flags access = DRM_GEM_OBJECT_DEV_ACCESS;
> >>> +
> >>> +	if (direction == DMA_TO_DEVICE)
> >>> +		access |= DRM_GEM_OBJECT_READ_ACCESS;
> >>> +	else if (direction == DMA_BIDIRECTIONAL)
> >>> +		access |= DRM_GEM_OBJECT_RW_ACCESS;
> >>> +	else
> >>> +		return -EINVAL;
> >>> +
> >>> +	return drm_gem_sync(obj, 0, obj->size, access);
> >>> +}
> >>> +EXPORT_SYMBOL(drm_gem_dmabuf_end_cpu_access);    
> >>
> >> I feel I must be missing something, but why does
> >> drm_gem_dmabuf_begin_cpu_access() reject DMA_TO_DEVICE and
> >> drm_gem_dmabuf_end_cpu_access() reject DMA_FROM_DEVICE?  
> > 
> > Not really sure what it means to prepare for dev access and synchronize
> > with what the device might have changed in memory. Sounds like device  
> > -> device synchronization, which is not what this API is for.  
> > 
> > Similarly preparing for CPU access with TO_DEVICE (AKA forcing previous
> > CPU changes to be visible to the device) doesn't make sense either.  
> 
> Well those operations might well be no-ops, but we shouldn't return an
> error just because of that.

Fair enough.

> 
> >>
> >> My understanding is that these begin/end calls should be bracketing the
> >> operation and the same direction should be specified for each.  
> > 
> > If [1] is correct and the begin/end_cpu_access() is based on the
> > dma_sync_ semantics, nope, that's not how it's supposed to work. The
> > way I see it, it just expresses the cache operations you want to take
> > place around your CPU access.  
> 
> Well that function completely ignored the direction flag. Looking at a
> caller of the APIs it seems very common to bracket the work with the
> same direction - admittedly they are all DMA_FROM_DEVICE examples I've
> found. E.g. a random FB driver:
> 
> https://elixir.bootlin.com/linux/v6.17.1/source/drivers/gpu/drm/solomon/ssd130x.c#L1026
> 
> > 	drm_gem_fb_end_cpu_access(fb, DMA_FROM_DEVICE);  
> 
> I have to admit I don't really know what the point of this is - and the
> error return isn't used beyond a drm_err(). But we don't want to spam
> the logs with errors!

Sure, I can return zero instead of an error.

> 
> > If you read data from the CPU, you want dir=FROM_DEVICE in your
> > begin_cpu_access(), so that the CPU caches are invalidated. If you
> > write from the CPU, you want dir=TO_DEVICE in your end_cpu_access. If
> > you know you will be reading again soon, you might want to pass
> > dir=BIDIR in your end_cpu_access(), though I'm not too sure what's the
> > benefit of that to be honest.
> > 
> > [1]https://elixir.bootlin.com/linux/v6.17.1/source/drivers/gpu/drm/tegra/gem.c#L684  
> 
> The documentation[2] states:
> 
> >  * - Fallback operations in the kernel, for example when a device is connected
> >  *   over USB and the kernel needs to shuffle the data around first before
> >  *   sending it away. Cache coherency is handled by bracketing any transactions
> >  *   with calls to dma_buf_begin_cpu_access() and dma_buf_end_cpu_access()
> >  *   access.  
> 
> So I guess the purpose is where it's not "cache coherency" as such, but
> actually having to shuffle the data between different memories.

The fact it's using the same flags that are used for dma_sync_ ops
makes it so confusing :face_palm:. I wish these were separate R/W
flags instead encoding what the CPU intends to do, and that the API
would make it clear that the same flags should be passed on both begin
and end. Anyway, I guess returning 0 instead of an error fixes the
problem, so I'll do that.