drm/msm: fix msm_gem_vma_new() allocations for managed GPUVMs

[PATCH] drm/msm: fix msm_gem_vma_new() allocations for managed GPUVMs

Posted by Dmitry Baryshkov 1 month, 1 week ago

Since commit 3309323241fb ("drm/gpuvm: Kill drm_gpuva_init()") MSM
driver fails to init, failing with "[drm:msm_gpu_init] *ERROR* could not
allocate memptrs: -22" errors. The mentioned commit reworked the
function, but didn't take into account that op_map is initialized at the
top of the function, while ranges might change if GPUVM is managed by
the kernel.

Move op_mode initialization after finalizing all addresses and right
before the drm_gpuva_init_from_op() call.

Reported-by: Danct12 <danct12@disroot.org>
Fixes: 3309323241fb ("drm/gpuvm: Kill drm_gpuva_init()")
Suggested-by: Rob Clark <robin.clark@oss.qualcomm.com>
Signed-off-by: Dmitry Baryshkov <dmitry.baryshkov@oss.qualcomm.com>
---
 drivers/gpu/drm/msm/msm_gem_vma.c | 13 +++++++------
 1 file changed, 7 insertions(+), 6 deletions(-)

diff --git a/drivers/gpu/drm/msm/msm_gem_vma.c b/drivers/gpu/drm/msm/msm_gem_vma.c
index 3f440bc1f7106f3b0091f037611d0b433e5e2c18..6df6b7c0984da57fe64de41fa54f7dea0a324c74 100644
--- a/drivers/gpu/drm/msm/msm_gem_vma.c
+++ b/drivers/gpu/drm/msm/msm_gem_vma.c
@@ -368,12 +368,6 @@ struct drm_gpuva *
 msm_gem_vma_new(struct drm_gpuvm *gpuvm, struct drm_gem_object *obj,
 		u64 offset, u64 range_start, u64 range_end)
 {
-	struct drm_gpuva_op_map op_map = {
-		.va.addr = range_start,
-		.va.range = range_end - range_start,
-		.gem.obj = obj,
-		.gem.offset = offset,
-	};
 	struct msm_gem_vm *vm = to_msm_vm(gpuvm);
 	struct drm_gpuvm_bo *vm_bo;
 	struct msm_gem_vma *vma;
@@ -402,6 +396,13 @@ msm_gem_vma_new(struct drm_gpuvm *gpuvm, struct drm_gem_object *obj,
 	if (obj)
 		GEM_WARN_ON((range_end - range_start) > obj->size);
 
+	struct drm_gpuva_op_map op_map = {
+		.va.addr = range_start,
+		.va.range = range_end - range_start,
+		.gem.obj = obj,
+		.gem.offset = offset,
+	};
+
 	drm_gpuva_init_from_op(&vma->base, &op_map);
 	vma->mapped = false;
 

---
base-commit: 0f4c93f7eb861acab537dbe94441817a270537bf
change-id: 20250823-msm-fix-gpuvm-init-520d87ebcf26

Best regards,
-- 
With best wishes
Dmitry

Re: [PATCH] drm/msm: fix msm_gem_vma_new() allocations for managed GPUVMs

Posted by Boris Brezillon 1 month, 1 week ago

On Sat, 23 Aug 2025 03:12:00 +0300
Dmitry Baryshkov <dmitry.baryshkov@oss.qualcomm.com> wrote:

> Since commit 3309323241fb ("drm/gpuvm: Kill drm_gpuva_init()") MSM
> driver fails to init, failing with "[drm:msm_gpu_init] *ERROR* could not
> allocate memptrs: -22" errors. The mentioned commit reworked the
> function, but didn't take into account that op_map is initialized at the
> top of the function, while ranges might change if GPUVM is managed by
> the kernel.
> 
> Move op_mode initialization after finalizing all addresses and right
> before the drm_gpuva_init_from_op() call.
> 
> Reported-by: Danct12 <danct12@disroot.org>
> Fixes: 3309323241fb ("drm/gpuvm: Kill drm_gpuva_init()")
> Suggested-by: Rob Clark <robin.clark@oss.qualcomm.com>
> Signed-off-by: Dmitry Baryshkov <dmitry.baryshkov@oss.qualcomm.com>
> ---
>  drivers/gpu/drm/msm/msm_gem_vma.c | 13 +++++++------
>  1 file changed, 7 insertions(+), 6 deletions(-)
> 
> diff --git a/drivers/gpu/drm/msm/msm_gem_vma.c b/drivers/gpu/drm/msm/msm_gem_vma.c
> index 3f440bc1f7106f3b0091f037611d0b433e5e2c18..6df6b7c0984da57fe64de41fa54f7dea0a324c74 100644
> --- a/drivers/gpu/drm/msm/msm_gem_vma.c
> +++ b/drivers/gpu/drm/msm/msm_gem_vma.c
> @@ -368,12 +368,6 @@ struct drm_gpuva *
>  msm_gem_vma_new(struct drm_gpuvm *gpuvm, struct drm_gem_object *obj,
>  		u64 offset, u64 range_start, u64 range_end)
>  {
> -	struct drm_gpuva_op_map op_map = {
> -		.va.addr = range_start,
> -		.va.range = range_end - range_start,
> -		.gem.obj = obj,
> -		.gem.offset = offset,
> -	};
>  	struct msm_gem_vm *vm = to_msm_vm(gpuvm);
>  	struct drm_gpuvm_bo *vm_bo;
>  	struct msm_gem_vma *vma;
> @@ -402,6 +396,13 @@ msm_gem_vma_new(struct drm_gpuvm *gpuvm, struct drm_gem_object *obj,
>  	if (obj)
>  		GEM_WARN_ON((range_end - range_start) > obj->size);
>  
> +	struct drm_gpuva_op_map op_map = {
> +		.va.addr = range_start,
> +		.va.range = range_end - range_start,
> +		.gem.obj = obj,
> +		.gem.offset = offset,
> +	};

OOC, are we now allowed to declare local variables in the middle of a
code block in kernel code? I must admit that's not something I tried
doing recently, but I've had gcc warnings in the past because of that.

> +
>  	drm_gpuva_init_from_op(&vma->base, &op_map);
>  	vma->mapped = false;
>  
> 
> ---
> base-commit: 0f4c93f7eb861acab537dbe94441817a270537bf
> change-id: 20250823-msm-fix-gpuvm-init-520d87ebcf26
> 
> Best regards,

Re: [PATCH] drm/msm: fix msm_gem_vma_new() allocations for managed GPUVMs

Posted by Rob Clark 1 month, 1 week ago

On Mon, Aug 25, 2025 at 12:09 AM Boris Brezillon
<boris.brezillon@collabora.com> wrote:
>
> On Sat, 23 Aug 2025 03:12:00 +0300
> Dmitry Baryshkov <dmitry.baryshkov@oss.qualcomm.com> wrote:
>
> > Since commit 3309323241fb ("drm/gpuvm: Kill drm_gpuva_init()") MSM
> > driver fails to init, failing with "[drm:msm_gpu_init] *ERROR* could not
> > allocate memptrs: -22" errors. The mentioned commit reworked the
> > function, but didn't take into account that op_map is initialized at the
> > top of the function, while ranges might change if GPUVM is managed by
> > the kernel.
> >
> > Move op_mode initialization after finalizing all addresses and right
> > before the drm_gpuva_init_from_op() call.
> >
> > Reported-by: Danct12 <danct12@disroot.org>
> > Fixes: 3309323241fb ("drm/gpuvm: Kill drm_gpuva_init()")
> > Suggested-by: Rob Clark <robin.clark@oss.qualcomm.com>
> > Signed-off-by: Dmitry Baryshkov <dmitry.baryshkov@oss.qualcomm.com>
> > ---
> >  drivers/gpu/drm/msm/msm_gem_vma.c | 13 +++++++------
> >  1 file changed, 7 insertions(+), 6 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/msm/msm_gem_vma.c b/drivers/gpu/drm/msm/msm_gem_vma.c
> > index 3f440bc1f7106f3b0091f037611d0b433e5e2c18..6df6b7c0984da57fe64de41fa54f7dea0a324c74 100644
> > --- a/drivers/gpu/drm/msm/msm_gem_vma.c
> > +++ b/drivers/gpu/drm/msm/msm_gem_vma.c
> > @@ -368,12 +368,6 @@ struct drm_gpuva *
> >  msm_gem_vma_new(struct drm_gpuvm *gpuvm, struct drm_gem_object *obj,
> >               u64 offset, u64 range_start, u64 range_end)
> >  {
> > -     struct drm_gpuva_op_map op_map = {
> > -             .va.addr = range_start,
> > -             .va.range = range_end - range_start,
> > -             .gem.obj = obj,
> > -             .gem.offset = offset,
> > -     };
> >       struct msm_gem_vm *vm = to_msm_vm(gpuvm);
> >       struct drm_gpuvm_bo *vm_bo;
> >       struct msm_gem_vma *vma;
> > @@ -402,6 +396,13 @@ msm_gem_vma_new(struct drm_gpuvm *gpuvm, struct drm_gem_object *obj,
> >       if (obj)
> >               GEM_WARN_ON((range_end - range_start) > obj->size);
> >
> > +     struct drm_gpuva_op_map op_map = {
> > +             .va.addr = range_start,
> > +             .va.range = range_end - range_start,
> > +             .gem.obj = obj,
> > +             .gem.offset = offset,
> > +     };
>
> OOC, are we now allowed to declare local variables in the middle of a
> code block in kernel code? I must admit that's not something I tried
> doing recently, but I've had gcc warnings in the past because of that.

yes.. I try not to go overboard with it, but this is a case where it
seems like the cleanest solution

BR,
-R

> > +
> >       drm_gpuva_init_from_op(&vma->base, &op_map);
> >       vma->mapped = false;
> >
> >
> > ---
> > base-commit: 0f4c93f7eb861acab537dbe94441817a270537bf
> > change-id: 20250823-msm-fix-gpuvm-init-520d87ebcf26
> >
> > Best regards,
>

Re: [PATCH] drm/msm: fix msm_gem_vma_new() allocations for managed GPUVMs

Posted by Dmitry Baryshkov 1 month, 1 week ago

On Sat, 23 Aug 2025 03:12:00 +0300, Dmitry Baryshkov wrote:
> Since commit 3309323241fb ("drm/gpuvm: Kill drm_gpuva_init()") MSM
> driver fails to init, failing with "[drm:msm_gpu_init] *ERROR* could not
> allocate memptrs: -22" errors. The mentioned commit reworked the
> function, but didn't take into account that op_map is initialized at the
> top of the function, while ranges might change if GPUVM is managed by
> the kernel.
> 
> [...]

Applied to drm-misc-next, thanks!

[1/1] drm/msm: fix msm_gem_vma_new() allocations for managed GPUVMs
      commit: efe927b9702643a1d80472664c2642f0304cb608

Best regards,
-- 
With best wishes
Dmitry

Re: [PATCH] drm/msm: fix msm_gem_vma_new() allocations for managed GPUVMs

Posted by Rob Clark 1 month, 1 week ago

On Fri, Aug 22, 2025 at 5:12 PM Dmitry Baryshkov
<dmitry.baryshkov@oss.qualcomm.com> wrote:
>
> Since commit 3309323241fb ("drm/gpuvm: Kill drm_gpuva_init()") MSM
> driver fails to init, failing with "[drm:msm_gpu_init] *ERROR* could not
> allocate memptrs: -22" errors. The mentioned commit reworked the
> function, but didn't take into account that op_map is initialized at the
> top of the function, while ranges might change if GPUVM is managed by
> the kernel.
>
> Move op_mode initialization after finalizing all addresses and right
> before the drm_gpuva_init_from_op() call.
>
> Reported-by: Danct12 <danct12@disroot.org>
> Fixes: 3309323241fb ("drm/gpuvm: Kill drm_gpuva_init()")
> Suggested-by: Rob Clark <robin.clark@oss.qualcomm.com>
> Signed-off-by: Dmitry Baryshkov <dmitry.baryshkov@oss.qualcomm.com>
> ---
>  drivers/gpu/drm/msm/msm_gem_vma.c | 13 +++++++------
>  1 file changed, 7 insertions(+), 6 deletions(-)
>
> diff --git a/drivers/gpu/drm/msm/msm_gem_vma.c b/drivers/gpu/drm/msm/msm_gem_vma.c
> index 3f440bc1f7106f3b0091f037611d0b433e5e2c18..6df6b7c0984da57fe64de41fa54f7dea0a324c74 100644
> --- a/drivers/gpu/drm/msm/msm_gem_vma.c
> +++ b/drivers/gpu/drm/msm/msm_gem_vma.c
> @@ -368,12 +368,6 @@ struct drm_gpuva *
>  msm_gem_vma_new(struct drm_gpuvm *gpuvm, struct drm_gem_object *obj,
>                 u64 offset, u64 range_start, u64 range_end)
>  {
> -       struct drm_gpuva_op_map op_map = {
> -               .va.addr = range_start,
> -               .va.range = range_end - range_start,
> -               .gem.obj = obj,
> -               .gem.offset = offset,
> -       };
>         struct msm_gem_vm *vm = to_msm_vm(gpuvm);
>         struct drm_gpuvm_bo *vm_bo;
>         struct msm_gem_vma *vma;
> @@ -402,6 +396,13 @@ msm_gem_vma_new(struct drm_gpuvm *gpuvm, struct drm_gem_object *obj,
>         if (obj)
>                 GEM_WARN_ON((range_end - range_start) > obj->size);
>
> +       struct drm_gpuva_op_map op_map = {
> +               .va.addr = range_start,
> +               .va.range = range_end - range_start,
> +               .gem.obj = obj,
> +               .gem.offset = offset,
> +       };

Thanks,

Reviewed-by: Rob Clark <rob.clark@oss.qualcomm.com>

> +
>         drm_gpuva_init_from_op(&vma->base, &op_map);
>         vma->mapped = false;
>
>
> ---
> base-commit: 0f4c93f7eb861acab537dbe94441817a270537bf
> change-id: 20250823-msm-fix-gpuvm-init-520d87ebcf26
>
> Best regards,
> --
> With best wishes
> Dmitry
>

Re: [PATCH] drm/msm: fix msm_gem_vma_new() allocations for managed GPUVMs

Posted by Dmitry Baryshkov 1 month, 1 week ago

On Fri, Aug 22, 2025 at 05:14:02PM -0700, Rob Clark wrote:
> On Fri, Aug 22, 2025 at 5:12 PM Dmitry Baryshkov
> <dmitry.baryshkov@oss.qualcomm.com> wrote:
> >
> > Since commit 3309323241fb ("drm/gpuvm: Kill drm_gpuva_init()") MSM
> > driver fails to init, failing with "[drm:msm_gpu_init] *ERROR* could not
> > allocate memptrs: -22" errors. The mentioned commit reworked the
> > function, but didn't take into account that op_map is initialized at the
> > top of the function, while ranges might change if GPUVM is managed by
> > the kernel.
> >
> > Move op_mode initialization after finalizing all addresses and right
> > before the drm_gpuva_init_from_op() call.
> >
> > Reported-by: Danct12 <danct12@disroot.org>
> > Fixes: 3309323241fb ("drm/gpuvm: Kill drm_gpuva_init()")
> > Suggested-by: Rob Clark <robin.clark@oss.qualcomm.com>
> > Signed-off-by: Dmitry Baryshkov <dmitry.baryshkov@oss.qualcomm.com>
> > ---
> >  drivers/gpu/drm/msm/msm_gem_vma.c | 13 +++++++------
> >  1 file changed, 7 insertions(+), 6 deletions(-)
> >
> 
> Thanks,
> 
> Reviewed-by: Rob Clark <rob.clark@oss.qualcomm.com>

I'll also need Acked-by to merge through drm-misc-next, where the
offending commit is.


-- 
With best wishes
Dmitry

Re: [PATCH] drm/msm: fix msm_gem_vma_new() allocations for managed GPUVMs

Posted by Rob Clark 1 month, 1 week ago

On Sat, Aug 23, 2025 at 6:50 AM Dmitry Baryshkov
<dmitry.baryshkov@oss.qualcomm.com> wrote:
>
> On Fri, Aug 22, 2025 at 05:14:02PM -0700, Rob Clark wrote:
> > On Fri, Aug 22, 2025 at 5:12 PM Dmitry Baryshkov
> > <dmitry.baryshkov@oss.qualcomm.com> wrote:
> > >
> > > Since commit 3309323241fb ("drm/gpuvm: Kill drm_gpuva_init()") MSM
> > > driver fails to init, failing with "[drm:msm_gpu_init] *ERROR* could not
> > > allocate memptrs: -22" errors. The mentioned commit reworked the
> > > function, but didn't take into account that op_map is initialized at the
> > > top of the function, while ranges might change if GPUVM is managed by
> > > the kernel.
> > >
> > > Move op_mode initialization after finalizing all addresses and right
> > > before the drm_gpuva_init_from_op() call.
> > >
> > > Reported-by: Danct12 <danct12@disroot.org>
> > > Fixes: 3309323241fb ("drm/gpuvm: Kill drm_gpuva_init()")
> > > Suggested-by: Rob Clark <robin.clark@oss.qualcomm.com>
> > > Signed-off-by: Dmitry Baryshkov <dmitry.baryshkov@oss.qualcomm.com>
> > > ---
> > >  drivers/gpu/drm/msm/msm_gem_vma.c | 13 +++++++------
> > >  1 file changed, 7 insertions(+), 6 deletions(-)
> > >
> >
> > Thanks,
> >
> > Reviewed-by: Rob Clark <rob.clark@oss.qualcomm.com>
>
> I'll also need Acked-by to merge through drm-misc-next, where the
> offending commit is.
>

Acked-by: Rob Clark <rob.clark@oss.qualcomm.com>

>
> --
> With best wishes
> Dmitry