drivers/gpu/drm/msm/msm_drv.c | 10 ++++++++++ 1 file changed, 10 insertions(+)
The MSM GBM backend uses MSM_GEM_NEW to allocate GEM buffers from the
KMS driver, imports them into the GPU driver (msm or kgsl) and then
uses them for rendering / blending. Commit 98f11fd1cf92 ("drm/msm: Take
the ioctls away from the KMS-only driver") dropped all IOCTLs from the
MSM KMS devices, pointing out the need to use dumb buffers, however dumb
buffers should not be used by the GPU for rendering. Restore GEM-related
IOCTLs for the KMS devices.
Fixes: 98f11fd1cf92 ("drm/msm: Take the ioctls away from the KMS-only driver")
Signed-off-by: Dmitry Baryshkov <dmitry.baryshkov@oss.qualcomm.com>
---
Note, here I assume that dumb buffers generally should not be used for
rendering. That doesn't seem to be complete truth as Mesa kmsro on MSM
devices uses DRM_IOCTL_MODE_CREATE_DUMB to create buffers for resources.
---
drivers/gpu/drm/msm/msm_drv.c | 10 ++++++++++
1 file changed, 10 insertions(+)
diff --git a/drivers/gpu/drm/msm/msm_drv.c b/drivers/gpu/drm/msm/msm_drv.c
index ed2a61c66ac9..f82c9a36a166 100644
--- a/drivers/gpu/drm/msm/msm_drv.c
+++ b/drivers/gpu/drm/msm/msm_drv.c
@@ -799,6 +799,14 @@ static const struct drm_ioctl_desc msm_ioctls[] = {
DRM_IOCTL_DEF_DRV(MSM_VM_BIND, msm_ioctl_vm_bind, DRM_RENDER_ALLOW),
};
+static const struct drm_ioctl_desc msm_kms_ioctls[] = {
+ DRM_IOCTL_DEF_DRV(MSM_GEM_NEW, msm_ioctl_gem_new, DRM_RENDER_ALLOW),
+ DRM_IOCTL_DEF_DRV(MSM_GEM_INFO, msm_ioctl_gem_info, DRM_RENDER_ALLOW),
+ DRM_IOCTL_DEF_DRV(MSM_GEM_CPU_PREP, msm_ioctl_gem_cpu_prep, DRM_RENDER_ALLOW),
+ DRM_IOCTL_DEF_DRV(MSM_GEM_CPU_FINI, msm_ioctl_gem_cpu_fini, DRM_RENDER_ALLOW),
+ DRM_IOCTL_DEF_DRV(MSM_GEM_MADVISE, msm_ioctl_gem_madvise, DRM_RENDER_ALLOW),
+};
+
static void msm_show_fdinfo(struct drm_printer *p, struct drm_file *file)
{
struct drm_device *dev = file->minor->dev;
@@ -868,6 +876,8 @@ static const struct drm_driver msm_kms_driver = {
#endif
MSM_FBDEV_DRIVER_OPS,
.show_fdinfo = msm_show_fdinfo,
+ .ioctls = msm_kms_ioctls,
+ .num_ioctls = ARRAY_SIZE(msm_kms_ioctls),
.fops = &fops,
.name = "msm-kms",
.desc = "MSM Snapdragon DRM",
---
base-commit: ac47870fd795549f03d57e0879fc730c79119f4b
change-id: 20260304-msm-restore-ioctls-2a1792311a09
Best regards,
--
With best wishes
DmitryOn Wed, Mar 4, 2026 at 5:34 AM Dmitry Baryshkov
<dmitry.baryshkov@oss.qualcomm.com> wrote:
>
> The MSM GBM backend uses MSM_GEM_NEW to allocate GEM buffers from the
> KMS driver, imports them into the GPU driver (msm or kgsl) and then
> uses them for rendering / blending. Commit 98f11fd1cf92 ("drm/msm: Take
> the ioctls away from the KMS-only driver") dropped all IOCTLs from the
> MSM KMS devices, pointing out the need to use dumb buffers, however dumb
> buffers should not be used by the GPU for rendering. Restore GEM-related
> IOCTLs for the KMS devices.
We shouldn't be exposing DRIVER_RENDER on KMS-only nodes. If msm gbm
is allocating from a primary node, that is problematic, since there is
no guarantee that anything but the compositor has access to the
primary node.
BR,
-R
>
> Fixes: 98f11fd1cf92 ("drm/msm: Take the ioctls away from the KMS-only driver")
> Signed-off-by: Dmitry Baryshkov <dmitry.baryshkov@oss.qualcomm.com>
> ---
> Note, here I assume that dumb buffers generally should not be used for
> rendering. That doesn't seem to be complete truth as Mesa kmsro on MSM
> devices uses DRM_IOCTL_MODE_CREATE_DUMB to create buffers for resources.
> ---
> drivers/gpu/drm/msm/msm_drv.c | 10 ++++++++++
> 1 file changed, 10 insertions(+)
>
> diff --git a/drivers/gpu/drm/msm/msm_drv.c b/drivers/gpu/drm/msm/msm_drv.c
> index ed2a61c66ac9..f82c9a36a166 100644
> --- a/drivers/gpu/drm/msm/msm_drv.c
> +++ b/drivers/gpu/drm/msm/msm_drv.c
> @@ -799,6 +799,14 @@ static const struct drm_ioctl_desc msm_ioctls[] = {
> DRM_IOCTL_DEF_DRV(MSM_VM_BIND, msm_ioctl_vm_bind, DRM_RENDER_ALLOW),
> };
>
> +static const struct drm_ioctl_desc msm_kms_ioctls[] = {
> + DRM_IOCTL_DEF_DRV(MSM_GEM_NEW, msm_ioctl_gem_new, DRM_RENDER_ALLOW),
> + DRM_IOCTL_DEF_DRV(MSM_GEM_INFO, msm_ioctl_gem_info, DRM_RENDER_ALLOW),
> + DRM_IOCTL_DEF_DRV(MSM_GEM_CPU_PREP, msm_ioctl_gem_cpu_prep, DRM_RENDER_ALLOW),
> + DRM_IOCTL_DEF_DRV(MSM_GEM_CPU_FINI, msm_ioctl_gem_cpu_fini, DRM_RENDER_ALLOW),
> + DRM_IOCTL_DEF_DRV(MSM_GEM_MADVISE, msm_ioctl_gem_madvise, DRM_RENDER_ALLOW),
> +};
> +
> static void msm_show_fdinfo(struct drm_printer *p, struct drm_file *file)
> {
> struct drm_device *dev = file->minor->dev;
> @@ -868,6 +876,8 @@ static const struct drm_driver msm_kms_driver = {
> #endif
> MSM_FBDEV_DRIVER_OPS,
> .show_fdinfo = msm_show_fdinfo,
> + .ioctls = msm_kms_ioctls,
> + .num_ioctls = ARRAY_SIZE(msm_kms_ioctls),
> .fops = &fops,
> .name = "msm-kms",
> .desc = "MSM Snapdragon DRM",
>
> ---
> base-commit: ac47870fd795549f03d57e0879fc730c79119f4b
> change-id: 20260304-msm-restore-ioctls-2a1792311a09
>
> Best regards,
> --
> With best wishes
> Dmitry
>
On Wed, Mar 04, 2026 at 07:57:34AM -0800, Rob Clark wrote:
> On Wed, Mar 4, 2026 at 5:34 AM Dmitry Baryshkov
> <dmitry.baryshkov@oss.qualcomm.com> wrote:
> >
> > The MSM GBM backend uses MSM_GEM_NEW to allocate GEM buffers from the
> > KMS driver, imports them into the GPU driver (msm or kgsl) and then
> > uses them for rendering / blending. Commit 98f11fd1cf92 ("drm/msm: Take
> > the ioctls away from the KMS-only driver") dropped all IOCTLs from the
> > MSM KMS devices, pointing out the need to use dumb buffers, however dumb
> > buffers should not be used by the GPU for rendering. Restore GEM-related
> > IOCTLs for the KMS devices.
>
> We shouldn't be exposing DRIVER_RENDER on KMS-only nodes. If msm gbm
> is allocating from a primary node, that is problematic, since there is
> no guarantee that anything but the compositor has access to the
> primary node.
We are not exposing DRIVER_RENDER for KMS-only nodes. I'll check with
the correponding colleagues, what exactly is their use case, if the
allocation is happening in the compositor or in the app. In the latter
case, I agree, the buffers should be allocated from the GPU.
>
> BR,
> -R
>
> >
> > Fixes: 98f11fd1cf92 ("drm/msm: Take the ioctls away from the KMS-only driver")
> > Signed-off-by: Dmitry Baryshkov <dmitry.baryshkov@oss.qualcomm.com>
> > ---
> > Note, here I assume that dumb buffers generally should not be used for
> > rendering. That doesn't seem to be complete truth as Mesa kmsro on MSM
> > devices uses DRM_IOCTL_MODE_CREATE_DUMB to create buffers for resources.
> > ---
> > drivers/gpu/drm/msm/msm_drv.c | 10 ++++++++++
> > 1 file changed, 10 insertions(+)
> >
> > diff --git a/drivers/gpu/drm/msm/msm_drv.c b/drivers/gpu/drm/msm/msm_drv.c
> > index ed2a61c66ac9..f82c9a36a166 100644
> > --- a/drivers/gpu/drm/msm/msm_drv.c
> > +++ b/drivers/gpu/drm/msm/msm_drv.c
> > @@ -799,6 +799,14 @@ static const struct drm_ioctl_desc msm_ioctls[] = {
> > DRM_IOCTL_DEF_DRV(MSM_VM_BIND, msm_ioctl_vm_bind, DRM_RENDER_ALLOW),
> > };
> >
> > +static const struct drm_ioctl_desc msm_kms_ioctls[] = {
> > + DRM_IOCTL_DEF_DRV(MSM_GEM_NEW, msm_ioctl_gem_new, DRM_RENDER_ALLOW),
> > + DRM_IOCTL_DEF_DRV(MSM_GEM_INFO, msm_ioctl_gem_info, DRM_RENDER_ALLOW),
> > + DRM_IOCTL_DEF_DRV(MSM_GEM_CPU_PREP, msm_ioctl_gem_cpu_prep, DRM_RENDER_ALLOW),
> > + DRM_IOCTL_DEF_DRV(MSM_GEM_CPU_FINI, msm_ioctl_gem_cpu_fini, DRM_RENDER_ALLOW),
> > + DRM_IOCTL_DEF_DRV(MSM_GEM_MADVISE, msm_ioctl_gem_madvise, DRM_RENDER_ALLOW),
> > +};
> > +
> > static void msm_show_fdinfo(struct drm_printer *p, struct drm_file *file)
> > {
> > struct drm_device *dev = file->minor->dev;
> > @@ -868,6 +876,8 @@ static const struct drm_driver msm_kms_driver = {
> > #endif
> > MSM_FBDEV_DRIVER_OPS,
> > .show_fdinfo = msm_show_fdinfo,
> > + .ioctls = msm_kms_ioctls,
> > + .num_ioctls = ARRAY_SIZE(msm_kms_ioctls),
> > .fops = &fops,
> > .name = "msm-kms",
> > .desc = "MSM Snapdragon DRM",
> >
> > ---
> > base-commit: ac47870fd795549f03d57e0879fc730c79119f4b
> > change-id: 20260304-msm-restore-ioctls-2a1792311a09
> >
> > Best regards,
> > --
> > With best wishes
> > Dmitry
> >
--
With best wishes
Dmitry
On Wed, Mar 4, 2026 at 5:34 AM Dmitry Baryshkov
<dmitry.baryshkov@oss.qualcomm.com> wrote:
>
> The MSM GBM backend uses MSM_GEM_NEW to allocate GEM buffers from the
> KMS driver, imports them into the GPU driver (msm or kgsl) and then
> uses them for rendering / blending. Commit 98f11fd1cf92 ("drm/msm: Take
> the ioctls away from the KMS-only driver") dropped all IOCTLs from the
> MSM KMS devices, pointing out the need to use dumb buffers, however dumb
> buffers should not be used by the GPU for rendering. Restore GEM-related
> IOCTLs for the KMS devices.
>
> Fixes: 98f11fd1cf92 ("drm/msm: Take the ioctls away from the KMS-only driver")
> Signed-off-by: Dmitry Baryshkov <dmitry.baryshkov@oss.qualcomm.com>
> ---
> Note, here I assume that dumb buffers generally should not be used for
> rendering. That doesn't seem to be complete truth as Mesa kmsro on MSM
> devices uses DRM_IOCTL_MODE_CREATE_DUMB to create buffers for resources.
That is problematic in kmsro.. (but also unsure to what degree kmsro
ever got used "in production".. the x86 drivers don't use it. Android
and chromeos didn't use it. Etc.)
BR,
-R
> ---
> drivers/gpu/drm/msm/msm_drv.c | 10 ++++++++++
> 1 file changed, 10 insertions(+)
>
> diff --git a/drivers/gpu/drm/msm/msm_drv.c b/drivers/gpu/drm/msm/msm_drv.c
> index ed2a61c66ac9..f82c9a36a166 100644
> --- a/drivers/gpu/drm/msm/msm_drv.c
> +++ b/drivers/gpu/drm/msm/msm_drv.c
> @@ -799,6 +799,14 @@ static const struct drm_ioctl_desc msm_ioctls[] = {
> DRM_IOCTL_DEF_DRV(MSM_VM_BIND, msm_ioctl_vm_bind, DRM_RENDER_ALLOW),
> };
>
> +static const struct drm_ioctl_desc msm_kms_ioctls[] = {
> + DRM_IOCTL_DEF_DRV(MSM_GEM_NEW, msm_ioctl_gem_new, DRM_RENDER_ALLOW),
> + DRM_IOCTL_DEF_DRV(MSM_GEM_INFO, msm_ioctl_gem_info, DRM_RENDER_ALLOW),
> + DRM_IOCTL_DEF_DRV(MSM_GEM_CPU_PREP, msm_ioctl_gem_cpu_prep, DRM_RENDER_ALLOW),
> + DRM_IOCTL_DEF_DRV(MSM_GEM_CPU_FINI, msm_ioctl_gem_cpu_fini, DRM_RENDER_ALLOW),
> + DRM_IOCTL_DEF_DRV(MSM_GEM_MADVISE, msm_ioctl_gem_madvise, DRM_RENDER_ALLOW),
> +};
> +
> static void msm_show_fdinfo(struct drm_printer *p, struct drm_file *file)
> {
> struct drm_device *dev = file->minor->dev;
> @@ -868,6 +876,8 @@ static const struct drm_driver msm_kms_driver = {
> #endif
> MSM_FBDEV_DRIVER_OPS,
> .show_fdinfo = msm_show_fdinfo,
> + .ioctls = msm_kms_ioctls,
> + .num_ioctls = ARRAY_SIZE(msm_kms_ioctls),
> .fops = &fops,
> .name = "msm-kms",
> .desc = "MSM Snapdragon DRM",
>
> ---
> base-commit: ac47870fd795549f03d57e0879fc730c79119f4b
> change-id: 20260304-msm-restore-ioctls-2a1792311a09
>
> Best regards,
> --
> With best wishes
> Dmitry
>
On Wed, Mar 4, 2026 at 6:57 AM Rob Clark <rob.clark@oss.qualcomm.com> wrote:
>
> On Wed, Mar 4, 2026 at 5:34 AM Dmitry Baryshkov
> <dmitry.baryshkov@oss.qualcomm.com> wrote:
> >
> > The MSM GBM backend uses MSM_GEM_NEW to allocate GEM buffers from the
> > KMS driver, imports them into the GPU driver (msm or kgsl) and then
> > uses them for rendering / blending. Commit 98f11fd1cf92 ("drm/msm: Take
> > the ioctls away from the KMS-only driver") dropped all IOCTLs from the
> > MSM KMS devices, pointing out the need to use dumb buffers, however dumb
> > buffers should not be used by the GPU for rendering. Restore GEM-related
> > IOCTLs for the KMS devices.
> >
> > Fixes: 98f11fd1cf92 ("drm/msm: Take the ioctls away from the KMS-only driver")
> > Signed-off-by: Dmitry Baryshkov <dmitry.baryshkov@oss.qualcomm.com>
> > ---
> > Note, here I assume that dumb buffers generally should not be used for
> > rendering. That doesn't seem to be complete truth as Mesa kmsro on MSM
> > devices uses DRM_IOCTL_MODE_CREATE_DUMB to create buffers for resources.
>
> That is problematic in kmsro.. (but also unsure to what degree kmsro
> ever got used "in production".. the x86 drivers don't use it. Android
> and chromeos didn't use it. Etc.)
(also, allocate from the gpu render node)
BR,
-R
> BR,
> -R
>
> > ---
> > drivers/gpu/drm/msm/msm_drv.c | 10 ++++++++++
> > 1 file changed, 10 insertions(+)
> >
> > diff --git a/drivers/gpu/drm/msm/msm_drv.c b/drivers/gpu/drm/msm/msm_drv.c
> > index ed2a61c66ac9..f82c9a36a166 100644
> > --- a/drivers/gpu/drm/msm/msm_drv.c
> > +++ b/drivers/gpu/drm/msm/msm_drv.c
> > @@ -799,6 +799,14 @@ static const struct drm_ioctl_desc msm_ioctls[] = {
> > DRM_IOCTL_DEF_DRV(MSM_VM_BIND, msm_ioctl_vm_bind, DRM_RENDER_ALLOW),
> > };
> >
> > +static const struct drm_ioctl_desc msm_kms_ioctls[] = {
> > + DRM_IOCTL_DEF_DRV(MSM_GEM_NEW, msm_ioctl_gem_new, DRM_RENDER_ALLOW),
> > + DRM_IOCTL_DEF_DRV(MSM_GEM_INFO, msm_ioctl_gem_info, DRM_RENDER_ALLOW),
> > + DRM_IOCTL_DEF_DRV(MSM_GEM_CPU_PREP, msm_ioctl_gem_cpu_prep, DRM_RENDER_ALLOW),
> > + DRM_IOCTL_DEF_DRV(MSM_GEM_CPU_FINI, msm_ioctl_gem_cpu_fini, DRM_RENDER_ALLOW),
> > + DRM_IOCTL_DEF_DRV(MSM_GEM_MADVISE, msm_ioctl_gem_madvise, DRM_RENDER_ALLOW),
> > +};
> > +
> > static void msm_show_fdinfo(struct drm_printer *p, struct drm_file *file)
> > {
> > struct drm_device *dev = file->minor->dev;
> > @@ -868,6 +876,8 @@ static const struct drm_driver msm_kms_driver = {
> > #endif
> > MSM_FBDEV_DRIVER_OPS,
> > .show_fdinfo = msm_show_fdinfo,
> > + .ioctls = msm_kms_ioctls,
> > + .num_ioctls = ARRAY_SIZE(msm_kms_ioctls),
> > .fops = &fops,
> > .name = "msm-kms",
> > .desc = "MSM Snapdragon DRM",
> >
> > ---
> > base-commit: ac47870fd795549f03d57e0879fc730c79119f4b
> > change-id: 20260304-msm-restore-ioctls-2a1792311a09
> >
> > Best regards,
> > --
> > With best wishes
> > Dmitry
> >
On Wed, Mar 04, 2026 at 06:59:42AM -0800, Rob Clark wrote:
> On Wed, Mar 4, 2026 at 6:57 AM Rob Clark <rob.clark@oss.qualcomm.com> wrote:
> >
> > On Wed, Mar 4, 2026 at 5:34 AM Dmitry Baryshkov
> > <dmitry.baryshkov@oss.qualcomm.com> wrote:
> > >
> > > The MSM GBM backend uses MSM_GEM_NEW to allocate GEM buffers from the
> > > KMS driver, imports them into the GPU driver (msm or kgsl) and then
> > > uses them for rendering / blending. Commit 98f11fd1cf92 ("drm/msm: Take
> > > the ioctls away from the KMS-only driver") dropped all IOCTLs from the
> > > MSM KMS devices, pointing out the need to use dumb buffers, however dumb
> > > buffers should not be used by the GPU for rendering. Restore GEM-related
> > > IOCTLs for the KMS devices.
> > >
> > > Fixes: 98f11fd1cf92 ("drm/msm: Take the ioctls away from the KMS-only driver")
> > > Signed-off-by: Dmitry Baryshkov <dmitry.baryshkov@oss.qualcomm.com>
> > > ---
> > > Note, here I assume that dumb buffers generally should not be used for
> > > rendering. That doesn't seem to be complete truth as Mesa kmsro on MSM
> > > devices uses DRM_IOCTL_MODE_CREATE_DUMB to create buffers for resources.
> >
> > That is problematic in kmsro.. (but also unsure to what degree kmsro
> > ever got used "in production".. the x86 drivers don't use it. Android
> > and chromeos didn't use it. Etc.)
>
> (also, allocate from the gpu render node)
I think Asahi and VC4 allocate buffers from the GPU node and then import
them on the render side, but unfortunately iMX IPU driver doesn't seem
to support PRIME_FD_TO_HANDLE.
--
With best wishes
Dmitry
On Wed, Mar 4, 2026 at 5:03 PM Dmitry Baryshkov
<dmitry.baryshkov@oss.qualcomm.com> wrote:
>
> On Wed, Mar 04, 2026 at 06:59:42AM -0800, Rob Clark wrote:
> > On Wed, Mar 4, 2026 at 6:57 AM Rob Clark <rob.clark@oss.qualcomm.com> wrote:
> > >
> > > On Wed, Mar 4, 2026 at 5:34 AM Dmitry Baryshkov
> > > <dmitry.baryshkov@oss.qualcomm.com> wrote:
> > > >
> > > > The MSM GBM backend uses MSM_GEM_NEW to allocate GEM buffers from the
> > > > KMS driver, imports them into the GPU driver (msm or kgsl) and then
> > > > uses them for rendering / blending. Commit 98f11fd1cf92 ("drm/msm: Take
> > > > the ioctls away from the KMS-only driver") dropped all IOCTLs from the
> > > > MSM KMS devices, pointing out the need to use dumb buffers, however dumb
> > > > buffers should not be used by the GPU for rendering. Restore GEM-related
> > > > IOCTLs for the KMS devices.
> > > >
> > > > Fixes: 98f11fd1cf92 ("drm/msm: Take the ioctls away from the KMS-only driver")
> > > > Signed-off-by: Dmitry Baryshkov <dmitry.baryshkov@oss.qualcomm.com>
> > > > ---
> > > > Note, here I assume that dumb buffers generally should not be used for
> > > > rendering. That doesn't seem to be complete truth as Mesa kmsro on MSM
> > > > devices uses DRM_IOCTL_MODE_CREATE_DUMB to create buffers for resources.
> > >
> > > That is problematic in kmsro.. (but also unsure to what degree kmsro
> > > ever got used "in production".. the x86 drivers don't use it. Android
> > > and chromeos didn't use it. Etc.)
> >
> > (also, allocate from the gpu render node)
>
> I think Asahi and VC4 allocate buffers from the GPU node and then import
> them on the render side, but unfortunately iMX IPU driver doesn't seem
> to support PRIME_FD_TO_HANDLE.
I guess my general suggestion would be to allocate from GPU and import
to display first, and if that fails then try the other way around if
that fails. For kmsro we could handle this by providing our own
ro->create_for_resource().
BR,
-R
© 2016 - 2026 Red Hat, Inc.