Set vm_pgoff to 0 after using it to look up the GEM node, before passing
it on rockchip_gem_mmap_buf() where the offset must be from the start of
the buffer.
Passing in the fake offset currently works because the
dma_mmap_attrs implementation that is used for this device,
arm_iommu_mmap_attrs, ignores the offset completely.
Signed-off-by: Ørjan Eide <[email protected]>
---
drivers/gpu/drm/rockchip/rockchip_drm_gem.c | 5 +++++
1 file changed, 5 insertions(+)
diff --git a/drivers/gpu/drm/rockchip/rockchip_drm_gem.c b/drivers/gpu/drm/rockchip/rockchip_drm_gem.c
index 7ca8799e..69f01c3 100644
--- a/drivers/gpu/drm/rockchip/rockchip_drm_gem.c
+++ b/drivers/gpu/drm/rockchip/rockchip_drm_gem.c
@@ -94,6 +94,11 @@ int rockchip_gem_mmap(struct file *filp, struct vm_area_struct *vma)
return -EACCES;
}
+ /* Set vm_pgoff (used as a fake buffer offset by DRM) to 0 and map the
+ * whole buffer from the start.
+ */
+ vma->vm_pgoff = 0;
+
obj = container_of(node, struct drm_gem_object, vma_node);
ret = rockchip_gem_mmap_buf(obj, vma);
--
1.9.1
arm_iommu_mmap_attrs() ignores the page offset in vma->vm_pgoff that was
specified in the mmap call. Any user of dma_mmap_* will get an
unexpected result when the device uses the IOMMU DMA ops.
Some DRM driver that use dma_mmap_* seems to depend on it ignoring
vma->vm_pgoff. Both the Samsung Exynos DRM driver before v3.18, and the
current Rockchip DRM driver, use the page offset to pass GEM cookies
that are used to look up the GEM buffers and then call dma_mmap_attrs()
with the cookie value still in vma->vm_pgoff.
rockchip_gem_mmap_buf() in drivers/gpu/drm/rockchip/rockchip_drm_gem.c
still does this. It should be fixed before fixing
arm_iommu_mmap_attrs().
Signed-off-by: Ørjan Eide <[email protected]>
---
arch/arm/mm/dma-mapping.c | 8 ++++++++
1 file changed, 8 insertions(+)
diff --git a/arch/arm/mm/dma-mapping.c b/arch/arm/mm/dma-mapping.c
index c274476..07d571b 100644
--- a/arch/arm/mm/dma-mapping.c
+++ b/arch/arm/mm/dma-mapping.c
@@ -1364,6 +1364,9 @@ static int arm_iommu_mmap_attrs(struct device *dev, struct vm_area_struct *vma,
{
unsigned long uaddr = vma->vm_start;
unsigned long usize = vma->vm_end - vma->vm_start;
+ unsigned long nr_vma_pages = usize >> PAGE_SHIFT;
+ unsigned long nr_pages = PAGE_ALIGN(size) >> PAGE_SHIFT;
+ unsigned long off = vma->vm_pgoff;
struct page **pages = __iommu_get_pages(cpu_addr, attrs);
vma->vm_page_prot = __get_dma_pgprot(attrs, vma->vm_page_prot);
@@ -1371,6 +1374,11 @@ static int arm_iommu_mmap_attrs(struct device *dev, struct vm_area_struct *vma,
if (!pages)
return -ENXIO;
+ if (off >= nr_pages || nr_vma_pages > (nr_pages - off))
+ return -ENXIO;
+
+ pages += off;
+
do {
int ret = vm_insert_page(vma, uaddr, *pages++);
if (ret) {
--
1.9.1
Am Donnerstag, 16. April 2015, 16:41:51 schrieb ?rjan Eide:
> Set vm_pgoff to 0 after using it to look up the GEM node, before passing
> it on rockchip_gem_mmap_buf() where the offset must be from the start of
> the buffer.
>
> Passing in the fake offset currently works because the
> dma_mmap_attrs implementation that is used for this device,
> arm_iommu_mmap_attrs, ignores the offset completely.
>
> Signed-off-by: ?rjan Eide <[email protected]>
both patches on a rk3288-veyron-pinky
Tested-by: Heiko Stuebner <[email protected]>
Through which tree do you want to take these patches? I guess the rockchip-drm
related patch should go through the tree that will take the dma-mapping patch,
so you'll probably need an "Ack" from Mark Yao (Cc'ed).
Heiko
> ---
> drivers/gpu/drm/rockchip/rockchip_drm_gem.c | 5 +++++
> 1 file changed, 5 insertions(+)
>
> diff --git a/drivers/gpu/drm/rockchip/rockchip_drm_gem.c
> b/drivers/gpu/drm/rockchip/rockchip_drm_gem.c index 7ca8799e..69f01c3
> 100644
> --- a/drivers/gpu/drm/rockchip/rockchip_drm_gem.c
> +++ b/drivers/gpu/drm/rockchip/rockchip_drm_gem.c
> @@ -94,6 +94,11 @@ int rockchip_gem_mmap(struct file *filp, struct
> vm_area_struct *vma) return -EACCES;
> }
>
> + /* Set vm_pgoff (used as a fake buffer offset by DRM) to 0 and map the
> + * whole buffer from the start.
> + */
> + vma->vm_pgoff = 0;
> +
> obj = container_of(node, struct drm_gem_object, vma_node);
> ret = rockchip_gem_mmap_buf(obj, vma);
Hi Ørjan,
On Thu, Apr 16, 2015 at 10:41 PM, Ørjan Eide <[email protected]> wrote:
> arm_iommu_mmap_attrs() ignores the page offset in vma->vm_pgoff that was
> specified in the mmap call. Any user of dma_mmap_* will get an
> unexpected result when the device uses the IOMMU DMA ops.
>
> Some DRM driver that use dma_mmap_* seems to depend on it ignoring
> vma->vm_pgoff. Both the Samsung Exynos DRM driver before v3.18, and the
> current Rockchip DRM driver, use the page offset to pass GEM cookies
> that are used to look up the GEM buffers and then call dma_mmap_attrs()
> with the cookie value still in vma->vm_pgoff.
>
> rockchip_gem_mmap_buf() in drivers/gpu/drm/rockchip/rockchip_drm_gem.c
> still does this. It should be fixed before fixing
> arm_iommu_mmap_attrs().
>
> Signed-off-by: Ørjan Eide <[email protected]>
Reviewed-by: Daniel Kurtz <[email protected]>
I tested this patch on my RK3288 board with the corresponding
rockchip_drm patch [0], and it works. So, also:
Tested-by: Daniel Kurtz <[email protected]>
[0] https://patchwork.kernel.org/patch/6226591/
> ---
> arch/arm/mm/dma-mapping.c | 8 ++++++++
> 1 file changed, 8 insertions(+)
>
> diff --git a/arch/arm/mm/dma-mapping.c b/arch/arm/mm/dma-mapping.c
> index c274476..07d571b 100644
> --- a/arch/arm/mm/dma-mapping.c
> +++ b/arch/arm/mm/dma-mapping.c
> @@ -1364,6 +1364,9 @@ static int arm_iommu_mmap_attrs(struct device *dev, struct vm_area_struct *vma,
> {
> unsigned long uaddr = vma->vm_start;
> unsigned long usize = vma->vm_end - vma->vm_start;
> + unsigned long nr_vma_pages = usize >> PAGE_SHIFT;
> + unsigned long nr_pages = PAGE_ALIGN(size) >> PAGE_SHIFT;
> + unsigned long off = vma->vm_pgoff;
> struct page **pages = __iommu_get_pages(cpu_addr, attrs);
>
> vma->vm_page_prot = __get_dma_pgprot(attrs, vma->vm_page_prot);
> @@ -1371,6 +1374,11 @@ static int arm_iommu_mmap_attrs(struct device *dev, struct vm_area_struct *vma,
> if (!pages)
> return -ENXIO;
>
> + if (off >= nr_pages || nr_vma_pages > (nr_pages - off))
> + return -ENXIO;
> +
> + pages += off;
> +
> do {
> int ret = vm_insert_page(vma, uaddr, *pages++);
> if (ret) {
> --
> 1.9.1
>
Hi Ørjan,
On Thu, Apr 16, 2015 at 10:41 PM, Ørjan Eide <[email protected]> wrote:
>
> Set vm_pgoff to 0 after using it to look up the GEM node, before passing
> it on rockchip_gem_mmap_buf() where the offset must be from the start of
> the buffer.
>
> Passing in the fake offset currently works because the
> dma_mmap_attrs implementation that is used for this device,
> arm_iommu_mmap_attrs, ignores the offset completely.
>
> Signed-off-by: Ørjan Eide <[email protected]>
>
> ---
> drivers/gpu/drm/rockchip/rockchip_drm_gem.c | 5 +++++
> 1 file changed, 5 insertions(+)
>
> diff --git a/drivers/gpu/drm/rockchip/rockchip_drm_gem.c b/drivers/gpu/drm/rockchip/rockchip_drm_gem.c
> index 7ca8799e..69f01c3 100644
> --- a/drivers/gpu/drm/rockchip/rockchip_drm_gem.c
> +++ b/drivers/gpu/drm/rockchip/rockchip_drm_gem.c
> @@ -94,6 +94,11 @@ int rockchip_gem_mmap(struct file *filp, struct vm_area_struct *vma)
> return -EACCES;
> }
>
> + /* Set vm_pgoff (used as a fake buffer offset by DRM) to 0 and map the
> + * whole buffer from the start.
> + */
One very tiny nit. According to [0], multi-line comments start are
supposed to start with a single "/*"
[0] https://www.kernel.org/doc/Documentation/CodingStyle
Other than that, this patch is
Reviewed-by: Daniel Kurtz <[email protected]>
Thanks for sending this up!
>
> + vma->vm_pgoff = 0;
> +
> obj = container_of(node, struct drm_gem_object, vma_node);
> ret = rockchip_gem_mmap_buf(obj, vma);
>
> --
> 1.9.1
>
On 2015年04月19日 00:55, Heiko Stübner wrote:
> Am Donnerstag, 16. April 2015, 16:41:51 schrieb Ørjan Eide:
>> Set vm_pgoff to 0 after using it to look up the GEM node, before passing
>> it on rockchip_gem_mmap_buf() where the offset must be from the start of
>> the buffer.
>>
>> Passing in the fake offset currently works because the
>> dma_mmap_attrs implementation that is used for this device,
>> arm_iommu_mmap_attrs, ignores the offset completely.
>>
>> Signed-off-by: Ørjan Eide <[email protected]>
> both patches on a rk3288-veyron-pinky
>
> Tested-by: Heiko Stuebner <[email protected]>
>
> Through which tree do you want to take these patches? I guess the rockchip-drm
> related patch should go through the tree that will take the dma-mapping patch,
> so you'll probably need an "Ack" from Mark Yao (Cc'ed).
>
>
> Heiko
>
>> ---
>> drivers/gpu/drm/rockchip/rockchip_drm_gem.c | 5 +++++
>> 1 file changed, 5 insertions(+)
>>
>> diff --git a/drivers/gpu/drm/rockchip/rockchip_drm_gem.c
>> b/drivers/gpu/drm/rockchip/rockchip_drm_gem.c index 7ca8799e..69f01c3
>> 100644
>> --- a/drivers/gpu/drm/rockchip/rockchip_drm_gem.c
>> +++ b/drivers/gpu/drm/rockchip/rockchip_drm_gem.c
>> @@ -94,6 +94,11 @@ int rockchip_gem_mmap(struct file *filp, struct
>> vm_area_struct *vma) return -EACCES;
>> }
>>
>> + /* Set vm_pgoff (used as a fake buffer offset by DRM) to 0 and map the
>> + * whole buffer from the start.
>> + */
>> + vma->vm_pgoff = 0;
>> +
>> obj = container_of(node, struct drm_gem_object, vma_node);
>> ret = rockchip_gem_mmap_buf(obj, vma);
>
>
>
Thanks for this fix,
Acked-by: Mark Yao <[email protected]>
--
Ｍark
On 2015年04月20日 13:34, Mark yao wrote:
> On 2015年04月19日 00:55, Heiko Stübner wrote:
>> Am Donnerstag, 16. April 2015, 16:41:51 schrieb Ørjan Eide:
>>> Set vm_pgoff to 0 after using it to look up the GEM node, before
>>> passing
>>> it on rockchip_gem_mmap_buf() where the offset must be from the
>>> start of
>>> the buffer.
>>>
>>> Passing in the fake offset currently works because the
>>> dma_mmap_attrs implementation that is used for this device,
>>> arm_iommu_mmap_attrs, ignores the offset completely.
>>>
>>> Signed-off-by: Ørjan Eide <[email protected]>
>> both patches on a rk3288-veyron-pinky
>>
>> Tested-by: Heiko Stuebner <[email protected]>
>>
>> Through which tree do you want to take these patches? I guess the
>> rockchip-drm
>> related patch should go through the tree that will take the
>> dma-mapping patch,
>> so you'll probably need an "Ack" from Mark Yao (Cc'ed).
>>
>>
>> Heiko
>>
>>> ---
>>> drivers/gpu/drm/rockchip/rockchip_drm_gem.c | 5 +++++
>>> 1 file changed, 5 insertions(+)
>>>
>>> diff --git a/drivers/gpu/drm/rockchip/rockchip_drm_gem.c
>>> b/drivers/gpu/drm/rockchip/rockchip_drm_gem.c index 7ca8799e..69f01c3
>>> 100644
>>> --- a/drivers/gpu/drm/rockchip/rockchip_drm_gem.c
>>> +++ b/drivers/gpu/drm/rockchip/rockchip_drm_gem.c
>>> @@ -94,6 +94,11 @@ int rockchip_gem_mmap(struct file *filp, struct
>>> vm_area_struct *vma) return -EACCES;
>>> }
>>>
>>> + /* Set vm_pgoff (used as a fake buffer offset by DRM) to 0 and
>>> map the
>>> + * whole buffer from the start.
>>> + */
>>> + vma->vm_pgoff = 0;
>>> +
>>> obj = container_of(node, struct drm_gem_object, vma_node);
>>> ret = rockchip_gem_mmap_buf(obj, vma);
>>
>>
>>
> Thanks for this fix,
> Acked-by: Mark Yao <[email protected]>
>
I met this problem when work with gem non-iommu path, set vma->vm_pgoff
= 0 solved it. :-)
--
Ｍark
On Sun, Apr 19, 2015 at 12:55 AM, Heiko Stübner <[email protected]> wrote:
>
> Am Donnerstag, 16. April 2015, 16:41:51 schrieb Ørjan Eide:
> > Set vm_pgoff to 0 after using it to look up the GEM node, before passing
> > it on rockchip_gem_mmap_buf() where the offset must be from the start of
> > the buffer.
> >
> > Passing in the fake offset currently works because the
> > dma_mmap_attrs implementation that is used for this device,
> > arm_iommu_mmap_attrs, ignores the offset completely.
> >
> > Signed-off-by: Ørjan Eide <[email protected]>
>
> both patches on a rk3288-veyron-pinky
>
> Tested-by: Heiko Stuebner <[email protected]>
>
> Through which tree do you want to take these patches? I guess the rockchip-drm
> related patch should go through the tree that will take the dma-mapping patch,
> so you'll probably need an "Ack" from Mark Yao (Cc'ed).
As far as I can tell, these two patches ([0] & [1]) were never picked up.
Russell, can you pick both of them up in your tree?
[0] https://patchwork.kernel.org/patch/6226591/
[1] https://patchwork.kernel.org/patch/6226581/
-Dan
> Heiko
>
> > ---
> > drivers/gpu/drm/rockchip/rockchip_drm_gem.c | 5 +++++
> > 1 file changed, 5 insertions(+)
> >
> > diff --git a/drivers/gpu/drm/rockchip/rockchip_drm_gem.c
> > b/drivers/gpu/drm/rockchip/rockchip_drm_gem.c index 7ca8799e..69f01c3
> > 100644
> > --- a/drivers/gpu/drm/rockchip/rockchip_drm_gem.c
> > +++ b/drivers/gpu/drm/rockchip/rockchip_drm_gem.c
> > @@ -94,6 +94,11 @@ int rockchip_gem_mmap(struct file *filp, struct
> > vm_area_struct *vma) return -EACCES;
> > }
> >
> > + /* Set vm_pgoff (used as a fake buffer offset by DRM) to 0 and map the
> > + * whole buffer from the start.
> > + */
> > + vma->vm_pgoff = 0;
> > +
> > obj = container_of(node, struct drm_gem_object, vma_node);
> > ret = rockchip_gem_mmap_buf(obj, vma);
>