[PATCH v4 4/7] drm/gpuvm: Add a helper to check if two VA can be merged

[Boris Brezillon]

On Mon, 07 Jul 2025 21:00:54 +0200
"Danilo Krummrich" <dakr at kernel.org> wrote:

> On Mon Jul 7, 2025 at 7:04 PM CEST, Caterina Shablia wrote:
> > diff --git a/drivers/gpu/drm/drm_gpuvm.c b/drivers/gpu/drm/drm_gpuvm.c
> > index 05978c5c38b1..dc3c2f906400 100644
> > --- a/drivers/gpu/drm/drm_gpuvm.c
> > +++ b/drivers/gpu/drm/drm_gpuvm.c
> > @@ -2098,12 +2098,48 @@ op_unmap_cb(const struct drm_gpuvm_ops *fn, void *priv,
> >  	return fn->sm_step_unmap(&op, priv);
> >  }
> >  
> > +static bool can_merge(struct drm_gpuvm *gpuvm, const struct drm_gpuva *a,
> > +		      const struct drm_gpuva *b)
> > +{
> > +	/* Only GEM-based mappings can be merged, and they must point to
> > +	 * the same GEM object.
> > +	 */
> > +	if (a->gem.obj != b->gem.obj || !a->gem.obj)
> > +		return false;
> > +
> > +	/* Let's keep things simple for now and force all flags to match. */
> > +	if (a->flags != b->flags)
> > +		return false;
> > +
> > +	/* Order VAs for the rest of the checks. */
> > +	if (a->va.addr > b->va.addr)
> > +		swap(a, b);
> > +
> > +	/* We assume the caller already checked that VAs overlap or are
> > +	 * contiguous.
> > +	 */
> > +	if (drm_WARN_ON(gpuvm->drm, b->va.addr > a->va.addr + a->va.range))
> > +		return false;
> > +
> > +	/* We intentionally ignore u64 underflows because all we care about
> > +	 * here is whether the VA diff matches the GEM offset diff.
> > +	 */
> > +	return b->va.addr - a->va.addr == b->gem.offset - a->gem.offset;
> > +}
> > +
> >  static int
> >  __drm_gpuvm_sm_map(struct drm_gpuvm *gpuvm,
> >  		   const struct drm_gpuvm_ops *ops, void *priv,
> >  		   const struct drm_gpuvm_map_req *req)
> >  {
> >  	struct drm_gpuva *va, *next;
> > +	struct drm_gpuva reqva = {
> > +		.va.addr = req->va.addr,
> > +		.va.range = req->va.range,
> > +		.gem.offset = req->gem.offset,
> > +		.gem.obj = req->gem.obj,
> > +		.flags = req->flags,  
> 
> Huh? Where does req->flags come from? I don't remember that this flag exists in
> struct drm_gpuvm_map_req in the preceding patch?

Oops, I re-ordered commits, and forgot to verify that the series was
bisectable. This should be part of patch 4 actually.

> 
> > +	};
> >  	u64 req_end = req->va.addr + req->va.range;
> >  	int ret;
> >  
> > @@ -2116,12 +2152,9 @@ __drm_gpuvm_sm_map(struct drm_gpuvm *gpuvm,
> >  		u64 addr = va->va.addr;
> >  		u64 range = va->va.range;
> >  		u64 end = addr + range;
> > -		bool merge = !!va->gem.obj;
> > +		bool merge = can_merge(gpuvm, va, &reqva);  
> 
> I know you want to do the swap() trick above, but I don't like creating a
> temporary struct drm_gpuva with all the other uninitialized fields.

I mean, I could do it the other way around (gpuva -> op_map), but it
means doing it on each va with cross.

> 
> If you really want this, can we please limit the scope? Maybe the following
> helper:
> 
> 	static bool can_merge(struct drm_gpuvm *gpuvm,
> 			      const struct drm_gpuva *va,
> 			      struct drm_gpuvm_map_req *req)
> 	{
> 		struct drm_gpuva reqva = { ... };
> 		return __can_merge(gpuvm, va, reqva);

It's a bit of a shame though, because then this reqva is
initialized every time can_merge() is called, instead of once at the
beginning of an sm_map() operation. But maybe the compiler is smart
enough to see through it when inlining (assuming it actually inlines
the check).

> 	}