Since it's inception in 2012 it has been understood that the DRM GEM CMA
helpers do not depend on CMA as the backend allocator. In fact the first
bug fix to ensure the cma-helpers work correctly with an IOMMU backend
appeared in 2014. However currently the documentation for
drm_gem_cma_create() talks about "a contiguous chunk of memory" without
making clear which address space it will be a contiguous part of.
Additionally the CMA introduction is actively misleading because it only
contemplates the CMA backend.
This matters because when the device accesses the bus through an IOMMU
(and don't use the CMA backend) then the allocated memory is contiguous
only in the IOVA space. This is a significant difference compared to the
CMA backend and the behaviour can be a surprise even to someone who does
a reasonable level of code browsing (but doesn't find all the relevant
function pointers ;-) ).
Improve the kernel doc comments accordingly.
Signed-off-by: Daniel Thompson <[email protected]>
Reviewed-by: Lucas Stach <[email protected]>
---
Notes:
Am I Cc:ing the correct reviewers/maintainers with this patch? There
has been no negative feedback but I've been rebasing and re-posting it
for three kernel cycles now. Do I need to queue it somewhere special or
get it in front of someone specific?
Either way...
This RESEND is unaltered (except for collecting tags) and is rebased on
v5.19-rc1.
RESEND for v5.18-rc3
- Unaltered but rebased on v5.18-rc3
Changes in v3:
- Rebased on v5.17-rc2
- Minor improvements to wording.
Changes in v2:
- Oops. I did a final proof read and accidentally committed these
changes as a seperate patch. This means that v1 contains only
one tenth of the actual patch. This is fixed in v2. Many apologies
for the noise!
drivers/gpu/drm/drm_gem_cma_helper.c | 39 +++++++++++++++++++++-------
1 file changed, 29 insertions(+), 10 deletions(-)
diff --git a/drivers/gpu/drm/drm_gem_cma_helper.c b/drivers/gpu/drm/drm_gem_cma_helper.c
index f36734c2c9e1..42abee9a0f4f 100644
--- a/drivers/gpu/drm/drm_gem_cma_helper.c
+++ b/drivers/gpu/drm/drm_gem_cma_helper.c
@@ -26,12 +26,22 @@
/**
* DOC: cma helpers
*
- * The Contiguous Memory Allocator reserves a pool of memory at early boot
- * that is used to service requests for large blocks of contiguous memory.
+ * The DRM GEM/CMA helpers are a means to provide buffer objects that are
+ * presented to the device as a contiguous chunk of memory. This is useful
+ * for devices that do not support scatter-gather DMA (either directly or
+ * by using an intimately attached IOMMU).
*
- * The DRM GEM/CMA helpers use this allocator as a means to provide buffer
- * objects that are physically contiguous in memory. This is useful for
- * display drivers that are unable to map scattered buffers via an IOMMU.
+ * Despite the name, the DRM GEM/CMA helpers are not hardwired to use the
+ * Contiguous Memory Allocator (CMA).
+ *
+ * For devices that access the memory bus through an (external) IOMMU then
+ * the buffer objects are allocated using a traditional page-based
+ * allocator and may be scattered through physical memory. However they
+ * are contiguous in the IOVA space so appear contiguous to devices using
+ * them.
+ *
+ * For other devices then the helpers rely on CMA to provide buffer
+ * objects that are physically contiguous in memory.
*
* For GEM callback helpers in struct &drm_gem_object functions, see likewise
* named functions with an _object_ infix (e.g., drm_gem_cma_object_vmap() wraps
@@ -111,8 +121,14 @@ __drm_gem_cma_create(struct drm_device *drm, size_t size, bool private)
* @drm: DRM device
* @size: size of the object to allocate
*
- * This function creates a CMA GEM object and allocates a contiguous chunk of
- * memory as backing store.
+ * This function creates a CMA GEM object and allocates memory as backing store.
+ * The allocated memory will occupy a contiguous chunk of bus address space.
+ *
+ * For devices that are directly connected to the memory bus then the allocated
+ * memory will be physically contiguous. For devices that access through an
+ * IOMMU, then the allocated memory is not expected to be physically contiguous
+ * because having contiguous IOVAs is sufficient to meet a devices DMA
+ * requirements.
*
* Returns:
* A struct drm_gem_cma_object * on success or an ERR_PTR()-encoded negative
@@ -162,9 +178,12 @@ EXPORT_SYMBOL_GPL(drm_gem_cma_create);
* @size: size of the object to allocate
* @handle: return location for the GEM handle
*
- * This function creates a CMA GEM object, allocating a physically contiguous
- * chunk of memory as backing store. The GEM object is then added to the list
- * of object associated with the given file and a handle to it is returned.
+ * This function creates a CMA GEM object, allocating a chunk of memory as
+ * backing store. The GEM object is then added to the list of object associated
+ * with the given file and a handle to it is returned.
+ *
+ * The allocated memory will occupy a contiguous chunk of bus address space.
+ * See drm_gem_cma_create() for more details.
*
* Returns:
* A struct drm_gem_cma_object * on success or an ERR_PTR()-encoded negative
base-commit: f2906aa863381afb0015a9eb7fefad885d4e5a56
--
2.35.1
On Wed, Jun 08, 2022 at 02:58:21PM +0100, Daniel Thompson wrote:
> Since it's inception in 2012 it has been understood that the DRM GEM CMA
> helpers do not depend on CMA as the backend allocator. In fact the first
> bug fix to ensure the cma-helpers work correctly with an IOMMU backend
> appeared in 2014. However currently the documentation for
> drm_gem_cma_create() talks about "a contiguous chunk of memory" without
> making clear which address space it will be a contiguous part of.
> Additionally the CMA introduction is actively misleading because it only
> contemplates the CMA backend.
>
> This matters because when the device accesses the bus through an IOMMU
> (and don't use the CMA backend) then the allocated memory is contiguous
> only in the IOVA space. This is a significant difference compared to the
> CMA backend and the behaviour can be a surprise even to someone who does
> a reasonable level of code browsing (but doesn't find all the relevant
> function pointers ;-) ).
>
> Improve the kernel doc comments accordingly.
>
> Signed-off-by: Daniel Thompson <[email protected]>
> Reviewed-by: Lucas Stach <[email protected]>
> ---
>
> Notes:
> Am I Cc:ing the correct reviewers/maintainers with this patch? There
> has been no negative feedback but I've been rebasing and re-posting it
> for three kernel cycles now. Do I need to queue it somewhere special or
> get it in front of someone specific?
Occasionally stuff falls through a few too many cracks, that's all. We
have tons of committers for drm-misc (and Lucas is one of them), but
sometimes they shy away from pushing themselves and others see the r-b and
assume it's already handled, and then it doesn't move :-/
Anyway I pushed it now, thanks for your patch.
-Daniel
>
> Either way...
>
> This RESEND is unaltered (except for collecting tags) and is rebased on
> v5.19-rc1.
>
> RESEND for v5.18-rc3
> - Unaltered but rebased on v5.18-rc3
>
> Changes in v3:
> - Rebased on v5.17-rc2
> - Minor improvements to wording.
>
> Changes in v2:
> - Oops. I did a final proof read and accidentally committed these
> changes as a seperate patch. This means that v1 contains only
> one tenth of the actual patch. This is fixed in v2. Many apologies
> for the noise!
>
> drivers/gpu/drm/drm_gem_cma_helper.c | 39 +++++++++++++++++++++-------
> 1 file changed, 29 insertions(+), 10 deletions(-)
>
> diff --git a/drivers/gpu/drm/drm_gem_cma_helper.c b/drivers/gpu/drm/drm_gem_cma_helper.c
> index f36734c2c9e1..42abee9a0f4f 100644
> --- a/drivers/gpu/drm/drm_gem_cma_helper.c
> +++ b/drivers/gpu/drm/drm_gem_cma_helper.c
> @@ -26,12 +26,22 @@
> /**
> * DOC: cma helpers
> *
> - * The Contiguous Memory Allocator reserves a pool of memory at early boot
> - * that is used to service requests for large blocks of contiguous memory.
> + * The DRM GEM/CMA helpers are a means to provide buffer objects that are
> + * presented to the device as a contiguous chunk of memory. This is useful
> + * for devices that do not support scatter-gather DMA (either directly or
> + * by using an intimately attached IOMMU).
> *
> - * The DRM GEM/CMA helpers use this allocator as a means to provide buffer
> - * objects that are physically contiguous in memory. This is useful for
> - * display drivers that are unable to map scattered buffers via an IOMMU.
> + * Despite the name, the DRM GEM/CMA helpers are not hardwired to use the
> + * Contiguous Memory Allocator (CMA).
> + *
> + * For devices that access the memory bus through an (external) IOMMU then
> + * the buffer objects are allocated using a traditional page-based
> + * allocator and may be scattered through physical memory. However they
> + * are contiguous in the IOVA space so appear contiguous to devices using
> + * them.
> + *
> + * For other devices then the helpers rely on CMA to provide buffer
> + * objects that are physically contiguous in memory.
> *
> * For GEM callback helpers in struct &drm_gem_object functions, see likewise
> * named functions with an _object_ infix (e.g., drm_gem_cma_object_vmap() wraps
> @@ -111,8 +121,14 @@ __drm_gem_cma_create(struct drm_device *drm, size_t size, bool private)
> * @drm: DRM device
> * @size: size of the object to allocate
> *
> - * This function creates a CMA GEM object and allocates a contiguous chunk of
> - * memory as backing store.
> + * This function creates a CMA GEM object and allocates memory as backing store.
> + * The allocated memory will occupy a contiguous chunk of bus address space.
> + *
> + * For devices that are directly connected to the memory bus then the allocated
> + * memory will be physically contiguous. For devices that access through an
> + * IOMMU, then the allocated memory is not expected to be physically contiguous
> + * because having contiguous IOVAs is sufficient to meet a devices DMA
> + * requirements.
> *
> * Returns:
> * A struct drm_gem_cma_object * on success or an ERR_PTR()-encoded negative
> @@ -162,9 +178,12 @@ EXPORT_SYMBOL_GPL(drm_gem_cma_create);
> * @size: size of the object to allocate
> * @handle: return location for the GEM handle
> *
> - * This function creates a CMA GEM object, allocating a physically contiguous
> - * chunk of memory as backing store. The GEM object is then added to the list
> - * of object associated with the given file and a handle to it is returned.
> + * This function creates a CMA GEM object, allocating a chunk of memory as
> + * backing store. The GEM object is then added to the list of object associated
> + * with the given file and a handle to it is returned.
> + *
> + * The allocated memory will occupy a contiguous chunk of bus address space.
> + * See drm_gem_cma_create() for more details.
> *
> * Returns:
> * A struct drm_gem_cma_object * on success or an ERR_PTR()-encoded negative
>
> base-commit: f2906aa863381afb0015a9eb7fefad885d4e5a56
> --
> 2.35.1
>
--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
On Wed, Jun 08, 2022 at 05:37:50PM +0200, Daniel Vetter wrote:
> On Wed, Jun 08, 2022 at 02:58:21PM +0100, Daniel Thompson wrote:
> > Since it's inception in 2012 it has been understood that the DRM GEM CMA
> > helpers do not depend on CMA as the backend allocator. In fact the first
> > bug fix to ensure the cma-helpers work correctly with an IOMMU backend
> > appeared in 2014. However currently the documentation for
> > drm_gem_cma_create() talks about "a contiguous chunk of memory" without
> > making clear which address space it will be a contiguous part of.
> > Additionally the CMA introduction is actively misleading because it only
> > contemplates the CMA backend.
> >
> > This matters because when the device accesses the bus through an IOMMU
> > (and don't use the CMA backend) then the allocated memory is contiguous
> > only in the IOVA space. This is a significant difference compared to the
> > CMA backend and the behaviour can be a surprise even to someone who does
> > a reasonable level of code browsing (but doesn't find all the relevant
> > function pointers ;-) ).
> >
> > Improve the kernel doc comments accordingly.
> >
> > Signed-off-by: Daniel Thompson <[email protected]>
> > Reviewed-by: Lucas Stach <[email protected]>
> > ---
> >
> > Notes:
> > Am I Cc:ing the correct reviewers/maintainers with this patch? There
> > has been no negative feedback but I've been rebasing and re-posting it
> > for three kernel cycles now. Do I need to queue it somewhere special or
> > get it in front of someone specific?
>
> Occasionally stuff falls through a few too many cracks, that's all. We
> have tons of committers for drm-misc (and Lucas is one of them), but
> sometimes they shy away from pushing themselves and others see the r-b and
> assume it's already handled, and then it doesn't move :-/
No worries. Arguably I should have asked this question a little earlier
anyway.
Thanks for pushing it.
Daniel.