mm: export cma alloc and release
diff mbox series

Message ID 20191002212257.196849-1-salyzyn@android.com
State New
Headers show
Series
  • mm: export cma alloc and release
Related show

Commit Message

Mark Salyzyn Oct. 2, 2019, 9:22 p.m. UTC
Some drivers can not be turned into a module without cma_alloc and
cma_release exported.  Examples include ion, and we also found some
out of tree infiniband and camera drivers.

Signed-off-by: Mark Salyzyn <salyzyn@android.com>
Cc: kernel-team@android.com
Cc: linux-kernel@vger.kernel.org
---
 mm/cma.c | 2 ++
 1 file changed, 2 insertions(+)

Comments

Catalin Marinas Oct. 3, 2019, 8:55 a.m. UTC | #1
On Wed, Oct 02, 2019 at 02:22:48PM -0700, Mark Salyzyn wrote:
> Some drivers can not be turned into a module without cma_alloc and
> cma_release exported.  Examples include ion, and we also found some
> out of tree infiniband and camera drivers.
> 
> Signed-off-by: Mark Salyzyn <salyzyn@android.com>
> Cc: kernel-team@android.com
> Cc: linux-kernel@vger.kernel.org
> ---
>  mm/cma.c | 2 ++
>  1 file changed, 2 insertions(+)
> 
> diff --git a/mm/cma.c b/mm/cma.c
> index 7fe0b8356775..65d830eea3b1 100644
> --- a/mm/cma.c
> +++ b/mm/cma.c
> @@ -500,6 +500,7 @@ struct page *cma_alloc(struct cma *cma, size_t count, unsigned int align,
>  	pr_debug("%s(): returned %p\n", __func__, page);
>  	return page;
>  }
> +EXPORT_SYMBOL_GPL(cma_alloc);
>  
>  /**
>   * cma_release() - release allocated pages
> @@ -533,6 +534,7 @@ bool cma_release(struct cma *cma, const struct page *pages, unsigned int count)
>  
>  	return true;
>  }
> +EXPORT_SYMBOL_GPL(cma_release);

Aren't drivers supposed to use the DMA API for such allocations rather
than invoking cma_*() directly?
Christoph Hellwig Oct. 5, 2019, 8:37 a.m. UTC | #2
On Thu, Oct 03, 2019 at 09:55:28AM +0100, Catalin Marinas wrote:
> Aren't drivers supposed to use the DMA API for such allocations rather
> than invoking cma_*() directly?

Yes, they are.
Mark Salyzyn Oct. 7, 2019, 4:50 p.m. UTC | #3
On 10/5/19 1:37 AM, Christoph Hellwig wrote:
> On Thu, Oct 03, 2019 at 09:55:28AM +0100, Catalin Marinas wrote:
>> Aren't drivers supposed to use the DMA API for such allocations rather
>> than invoking cma_*() directly?
> Yes, they are.

We have an engineer assigned to rewriting the ion memory driver to use 
dma_buf interfaces. Hopefully that effort will solve the problem of 
requiring these interfaces to be exported so that that driver (and 
others) can be modularized.

Thanks for the reviews, drop this patch from the list and we will 
regroup, and accept that standing code in the kernel can not be 
modularized for the moment.

Sincerely -- Mark Salyzyn
Christoph Hellwig Oct. 7, 2019, 4:53 p.m. UTC | #4
On Mon, Oct 07, 2019 at 09:50:31AM -0700, Mark Salyzyn wrote:
> On 10/5/19 1:37 AM, Christoph Hellwig wrote:
>> On Thu, Oct 03, 2019 at 09:55:28AM +0100, Catalin Marinas wrote:
>>> Aren't drivers supposed to use the DMA API for such allocations rather
>>> than invoking cma_*() directly?
>> Yes, they are.
>
> We have an engineer assigned to rewriting the ion memory driver to use 
> dma_buf interfaces. Hopefully that effort will solve the problem of 
> requiring these interfaces to be exported so that that driver (and others) 
> can be modularized.
>
> Thanks for the reviews, drop this patch from the list and we will regroup, 
> and accept that standing code in the kernel can not be modularized for the 
> moment.

How is that different from the "DMA-BUF Heaps (destaging ION)" series
floating around?
Mark Salyzyn Oct. 7, 2019, 5:06 p.m. UTC | #5
On 10/7/19 9:53 AM, Christoph Hellwig wrote:
> On Mon, Oct 07, 2019 at 09:50:31AM -0700, Mark Salyzyn wrote:
>> On 10/5/19 1:37 AM, Christoph Hellwig wrote:
>>> On Thu, Oct 03, 2019 at 09:55:28AM +0100, Catalin Marinas wrote:
>>>> Aren't drivers supposed to use the DMA API for such allocations rather
>>>> than invoking cma_*() directly?
>>> Yes, they are.
>> We have an engineer assigned to rewriting the ion memory driver to use
>> dma_buf interfaces. Hopefully that effort will solve the problem of
>> requiring these interfaces to be exported so that that driver (and others)
>> can be modularized.
>>
>> Thanks for the reviews, drop this patch from the list and we will regroup,
>> and accept that standing code in the kernel can not be modularized for the
>> moment.
> How is that different from the "DMA-BUF Heaps (destaging ION)" series
> floating around?

IDK, I am asking around because I am only superficially aware of that 
effort. Please view this as the left hand does not know what the right 
hand is doing. My issue was with a series of out-of-tree drivers that 
use the calls, and noted that currently the ion driver is using them as 
well, as rationalization for a 'user' for the export. I am pushing back 
on those out-of-tree drivers to switch their CMA interfaces to a modern 
approach as a result of these reviews; the result of this review had a 
positive effect because I was considering exporting them in the Android 
distro as a minimum, and now I am soundly rejecting that approach.

Sincerely -- Mark Salyzyn

Patch
diff mbox series

diff --git a/mm/cma.c b/mm/cma.c
index 7fe0b8356775..65d830eea3b1 100644
--- a/mm/cma.c
+++ b/mm/cma.c
@@ -500,6 +500,7 @@  struct page *cma_alloc(struct cma *cma, size_t count, unsigned int align,
 	pr_debug("%s(): returned %p\n", __func__, page);
 	return page;
 }
+EXPORT_SYMBOL_GPL(cma_alloc);
 
 /**
  * cma_release() - release allocated pages
@@ -533,6 +534,7 @@  bool cma_release(struct cma *cma, const struct page *pages, unsigned int count)
 
 	return true;
 }
+EXPORT_SYMBOL_GPL(cma_release);
 
 int cma_for_each_area(int (*it)(struct cma *cma, void *data), void *data)
 {