With commit ce88313069c36eef80f21fd7 ("arch/sh: make the DMA mapping
operations observe dev->dma_pfn_offset") the generic DMA allocation
function on which the SH 'dma_alloc_coherent()' function relies on,
accesses the 'dma_pfn_offset' field of struct device.
Unfortunately the 'dma_generic_alloc_coherent()' function is called from
several places with a NULL struct device argument, halting the CPU
during the boot process.
This patch fixes the issue by protecting access to dev->dma_pfn_offset,
with a trivial check for validity. It also passes a valid 'struct device'
in the 'platform_resource_setup_memory()' function which is the main user
of 'dma_alloc_coherent()', and inserts a WARN_ON() check to remind to future
(and existing) bogus users of this function to provide a valid 'struct device'
whenever possible.
Fixes: ce88313069c36eef80f21fd7 ("arch/sh: make the DMA mapping operations observe dev->dma_pfn_offset")
Signed-off-by: Jacopo Mondi <[email protected]>
Reviewed-by: Geert Uytterhoeven <[email protected]>
Reviewed-by: Thomas Petazzoni <[email protected]>
---
v2 -> v3:
- remove (now) useless parenthesis around pfn assignement as suggested
by Sergei
- Add changelog to the patch, which I forgot in v2
v1 -> v2:
- Move WARN_ON() closer to dev validity check as suggested by Geert
---
arch/sh/mm/consistent.c | 11 ++++++++---
1 file changed, 8 insertions(+), 3 deletions(-)
diff --git a/arch/sh/mm/consistent.c b/arch/sh/mm/consistent.c
index 8ce9869..f1b4469 100644
--- a/arch/sh/mm/consistent.c
+++ b/arch/sh/mm/consistent.c
@@ -59,7 +59,9 @@ void *dma_generic_alloc_coherent(struct device *dev, size_t size,
split_page(pfn_to_page(virt_to_phys(ret) >> PAGE_SHIFT), order);
- *dma_handle = virt_to_phys(ret) - PFN_PHYS(dev->dma_pfn_offset);
+ *dma_handle = virt_to_phys(ret);
+ if (!WARN_ON(!dev))
+ *dma_handle -= PFN_PHYS(dev->dma_pfn_offset);
return ret_nocache;
}
@@ -69,9 +71,12 @@ void dma_generic_free_coherent(struct device *dev, size_t size,
unsigned long attrs)
{
int order = get_order(size);
- unsigned long pfn = (dma_handle >> PAGE_SHIFT) + dev->dma_pfn_offset;
+ unsigned long pfn = dma_handle >> PAGE_SHIFT;
int k;
+ if (!WARN_ON(!dev))
+ pfn += dev->dma_pfn_offset;
+
for (k = 0; k < (1 << order); k++)
__free_pages(pfn_to_page(pfn + k), 0);
@@ -143,7 +148,7 @@ int __init platform_resource_setup_memory(struct platform_device *pdev,
if (!memsize)
return 0;
- buf = dma_alloc_coherent(NULL, memsize, &dma_handle, GFP_KERNEL);
+ buf = dma_alloc_coherent(&pdev->dev, memsize, &dma_handle, GFP_KERNEL);
if (!buf) {
pr_warning("%s: unable to allocate memory\n", name);
return -ENOMEM;
--
2.7.4
Hi again Christoph,
The gentle ping actually applies to this version of the patch.
Sorry for the confusion.
Thanks
j
On Wed, Apr 18, 2018 at 11:42:35AM +0200, Jacopo Mondi wrote:
> With commit ce88313069c36eef80f21fd7 ("arch/sh: make the DMA mapping
> operations observe dev->dma_pfn_offset") the generic DMA allocation
> function on which the SH 'dma_alloc_coherent()' function relies on,
> accesses the 'dma_pfn_offset' field of struct device.
>
> Unfortunately the 'dma_generic_alloc_coherent()' function is called from
> several places with a NULL struct device argument, halting the CPU
> during the boot process.
>
> This patch fixes the issue by protecting access to dev->dma_pfn_offset,
> with a trivial check for validity. It also passes a valid 'struct device'
> in the 'platform_resource_setup_memory()' function which is the main user
> of 'dma_alloc_coherent()', and inserts a WARN_ON() check to remind to future
> (and existing) bogus users of this function to provide a valid 'struct device'
> whenever possible.
>
> Fixes: ce88313069c36eef80f21fd7 ("arch/sh: make the DMA mapping operations observe dev->dma_pfn_offset")
> Signed-off-by: Jacopo Mondi <[email protected]>
> Reviewed-by: Geert Uytterhoeven <[email protected]>
> Reviewed-by: Thomas Petazzoni <[email protected]>
>
> ---
> v2 -> v3:
> - remove (now) useless parenthesis around pfn assignement as suggested
> by Sergei
> - Add changelog to the patch, which I forgot in v2
>
> v1 -> v2:
> - Move WARN_ON() closer to dev validity check as suggested by Geert
>
> ---
> arch/sh/mm/consistent.c | 11 ++++++++---
> 1 file changed, 8 insertions(+), 3 deletions(-)
>
> diff --git a/arch/sh/mm/consistent.c b/arch/sh/mm/consistent.c
> index 8ce9869..f1b4469 100644
> --- a/arch/sh/mm/consistent.c
> +++ b/arch/sh/mm/consistent.c
> @@ -59,7 +59,9 @@ void *dma_generic_alloc_coherent(struct device *dev, size_t size,
>
> split_page(pfn_to_page(virt_to_phys(ret) >> PAGE_SHIFT), order);
>
> - *dma_handle = virt_to_phys(ret) - PFN_PHYS(dev->dma_pfn_offset);
> + *dma_handle = virt_to_phys(ret);
> + if (!WARN_ON(!dev))
> + *dma_handle -= PFN_PHYS(dev->dma_pfn_offset);
>
> return ret_nocache;
> }
> @@ -69,9 +71,12 @@ void dma_generic_free_coherent(struct device *dev, size_t size,
> unsigned long attrs)
> {
> int order = get_order(size);
> - unsigned long pfn = (dma_handle >> PAGE_SHIFT) + dev->dma_pfn_offset;
> + unsigned long pfn = dma_handle >> PAGE_SHIFT;
> int k;
>
> + if (!WARN_ON(!dev))
> + pfn += dev->dma_pfn_offset;
> +
> for (k = 0; k < (1 << order); k++)
> __free_pages(pfn_to_page(pfn + k), 0);
>
> @@ -143,7 +148,7 @@ int __init platform_resource_setup_memory(struct platform_device *pdev,
> if (!memsize)
> return 0;
>
> - buf = dma_alloc_coherent(NULL, memsize, &dma_handle, GFP_KERNEL);
> + buf = dma_alloc_coherent(&pdev->dev, memsize, &dma_handle, GFP_KERNEL);
> if (!buf) {
> pr_warning("%s: unable to allocate memory\n", name);
> return -ENOMEM;
> --
> 2.7.4
>
On Wed, May 02, 2018 at 09:46:31AM +0200, jacopo mondi wrote:
> Hi again Christoph,
>
> The gentle ping actually applies to this version of the patch.
>
> Sorry for the confusion.
I'd expect this to go in through the sh tree, but if the sh maintainers
are fine with it I can take it through the dma-mapping tree.
Hi Christoph,
On Wed, May 02, 2018 at 05:39:09AM -0700, Christoph Hellwig wrote:
> On Wed, May 02, 2018 at 09:46:31AM +0200, jacopo mondi wrote:
> > Hi again Christoph,
> >
> > The gentle ping actually applies to this version of the patch.
> >
> > Sorry for the confusion.
>
> I'd expect this to go in through the sh tree, but if the sh maintainers
> are fine with it I can take it through the dma-mapping tree.
Oh, that's fine then. I'll let Rich comment if he plans to send a pull
requets for v4.17 fixes or not.
Thanks
j
On Thu, May 03, 2018 at 09:15:39AM +0200, jacopo mondi wrote:
> Hi Christoph,
>
> On Wed, May 02, 2018 at 05:39:09AM -0700, Christoph Hellwig wrote:
> > On Wed, May 02, 2018 at 09:46:31AM +0200, jacopo mondi wrote:
> > > Hi again Christoph,
> > >
> > > The gentle ping actually applies to this version of the patch.
> > >
> > > Sorry for the confusion.
> >
> > I'd expect this to go in through the sh tree, but if the sh maintainers
> > are fine with it I can take it through the dma-mapping tree.
>
> Oh, that's fine then. I'll let Rich comment if he plans to send a pull
> requets for v4.17 fixes or not.
OK, I'll prepare a pull request with this and another fix I just made.
Rich