Use dma_get_sgtable rather than dma_common_get_sgtable so a device's
dma_ops aren't bypassed. This is essential in situations where a device
uses an IOMMU and the physical memory is not contiguous (as the common
function assumes).
Signed-off-by: Jon Medhurst <[email protected]>
---
This also begs the question as to what happens if the memory region _is_
contiguous but is in highmem or an ioremapped region. Should a device
always provide dma_ops for that case? Because I believe the current
implementation of dma_common_get_sgtable won't work for those as it uses
virt_to_page.
I see that this point has been raised before [1] by Zeng Tao, and I
myself have been given a different fix to apply to a Linaro kernel tree.
However, both solutions looked wrong to me as they treat a dma_addr_t as
a physical address, so should at least be using dma_to_phys.
So, should we fix dma_common_get_sgtable or mandate that the device
has dma_ops? The latter seems to be implied by the commit message which
introduced the function:
This patch provides a generic implementation based on
virt_to_page() call. Architectures which require more
sophisticated translation might provide their own get_sgtable()
methods.
Note, I don't have a system where any of this code is used to test
things, and have never looked at this area before yesterday, so I may
have misunderstood what’s going on in the code.
[1] https://lkml.org/lkml/2014/12/1/584
drivers/staging/android/ion/ion_cma_heap.c | 3 +--
1 file changed, 1 insertion(+), 2 deletions(-)
diff --git a/drivers/staging/android/ion/ion_cma_heap.c b/drivers/staging/android/ion/ion_cma_heap.c
index f4211f1..86b91fd 100644
--- a/drivers/staging/android/ion/ion_cma_heap.c
+++ b/drivers/staging/android/ion/ion_cma_heap.c
@@ -73,8 +73,7 @@ static int ion_cma_allocate(struct ion_heap *heap, struct ion_buffer *buffer,
if (!info->table)
goto free_mem;
- if (dma_common_get_sgtable
- (dev, info->table, info->cpu_addr, info->handle, len))
+ if (dma_get_sgtable(dev, info->table, info->cpu_addr, info->handle, len))
goto free_table;
/* keep this for memory release */
buffer->priv_virt = info;
--
2.1.4
Hi Tixy,
On 17/07/15 12:01, Jon Medhurst (Tixy) wrote:
> Use dma_get_sgtable rather than dma_common_get_sgtable so a device's
> dma_ops aren't bypassed. This is essential in situations where a device
> uses an IOMMU and the physical memory is not contiguous (as the common
> function assumes).
>
> Signed-off-by: Jon Medhurst <[email protected]>
The lack of obvious users of this code makes it hard to tell if "dev"
here is always the right, real, device pointer and never null or some
dummy device with the wrong dma_ops, but the rest of the calls in this
file are to the proper DMA API interface so at least this patch
definitely makes things less wrong in that respect.
Reviewed-by: Robin Murphy <[email protected]>
> ---
>
> This also begs the question as to what happens if the memory region _is_
> contiguous but is in highmem or an ioremapped region. Should a device
> always provide dma_ops for that case? Because I believe the current
> implementation of dma_common_get_sgtable won't work for those as it uses
> virt_to_page.
>
> I see that this point has been raised before [1] by Zeng Tao, and I
> myself have been given a different fix to apply to a Linaro kernel tree.
> However, both solutions looked wrong to me as they treat a dma_addr_t as
> a physical address, so should at least be using dma_to_phys.
> So, should we fix dma_common_get_sgtable or mandate that the device
> has dma_ops? The latter seems to be implied by the commit message which
> introduced the function:
>
> This patch provides a generic implementation based on
> virt_to_page() call. Architectures which require more
> sophisticated translation might provide their own get_sgtable()
> methods.
Given that we're largely here due to having poked this on arm64 systems,
I'm inclined to think that implementing our own get_sgtable as per
arch/arm is the right course of action. Since a lot of architectures
using dma_common_get_sgtable don't even implement dma_to_phys, I don't
think it would be right to try complicating the common code for a case
that seems to be all but common. I can spin an arm64 patch if you like.
Robin.
> Note, I don't have a system where any of this code is used to test
> things, and have never looked at this area before yesterday, so I may
> have misunderstood what’s going on in the code.
>
> [1] https://lkml.org/lkml/2014/12/1/584
>
> drivers/staging/android/ion/ion_cma_heap.c | 3 +--
> 1 file changed, 1 insertion(+), 2 deletions(-)
>
> diff --git a/drivers/staging/android/ion/ion_cma_heap.c b/drivers/staging/android/ion/ion_cma_heap.c
> index f4211f1..86b91fd 100644
> --- a/drivers/staging/android/ion/ion_cma_heap.c
> +++ b/drivers/staging/android/ion/ion_cma_heap.c
> @@ -73,8 +73,7 @@ static int ion_cma_allocate(struct ion_heap *heap, struct ion_buffer *buffer,
> if (!info->table)
> goto free_mem;
>
> - if (dma_common_get_sgtable
> - (dev, info->table, info->cpu_addr, info->handle, len))
> + if (dma_get_sgtable(dev, info->table, info->cpu_addr, info->handle, len))
> goto free_table;
> /* keep this for memory release */
> buffer->priv_virt = info;
>
On Fri, 2015-07-17 at 16:21 +0100, Robin Murphy wrote:
> This also begs the question as to what happens if the memory region _is_
> > contiguous but is in highmem or an ioremapped region. Should a device
> > always provide dma_ops for that case? Because I believe the current
> > implementation of dma_common_get_sgtable won't work for those as it uses
> > virt_to_page.
> >
> > I see that this point has been raised before [1] by Zeng Tao, and I
> > myself have been given a different fix to apply to a Linaro kernel tree.
> > However, both solutions looked wrong to me as they treat a dma_addr_t as
> > a physical address, so should at least be using dma_to_phys.
> > So, should we fix dma_common_get_sgtable or mandate that the device
> > has dma_ops? The latter seems to be implied by the commit message which
> > introduced the function:
> >
> > This patch provides a generic implementation based on
> > virt_to_page() call. Architectures which require more
> > sophisticated translation might provide their own get_sgtable()
> > methods.
>
> Given that we're largely here due to having poked this on arm64 systems,
> I'm inclined to think that implementing our own get_sgtable as per
> arch/arm is the right course of action. Since a lot of architectures
> using dma_common_get_sgtable don't even implement dma_to_phys,
I had another check and that seems to be true.
> I don't
> think it would be right to try complicating the common code for a case
> that seems to be all but common.
I'm inclined to agree, however I'm rather new to this area.
> I can spin an arm64 patch if you like.
That would be good. Especially as from what I see on the arm kernel
lists you are already working in that area... And my inbox has just
pinged with that patch from you, so I'll add a reference here [2] so
people coming across this thread can find it easily.
For 32-bit arm my $subject patch should fix ION as that already has the
DMA ops.
--
Tixy
[1] https://lkml.org/lkml/2014/12/1/584
[2] http://lists.infradead.org/pipermail/linux-arm-kernel/2015-July/357561.html
On 07/17/2015 08:21 AM, Robin Murphy wrote:
> Hi Tixy,
>
> On 17/07/15 12:01, Jon Medhurst (Tixy) wrote:
>> Use dma_get_sgtable rather than dma_common_get_sgtable so a device's
>> dma_ops aren't bypassed. This is essential in situations where a device
>> uses an IOMMU and the physical memory is not contiguous (as the common
>> function assumes).
>>
>> Signed-off-by: Jon Medhurst <[email protected]>
>
> The lack of obvious users of this code makes it hard to tell if "dev"
> hereis always the right, real, device pointer and never null or some
> dummy device with the wrong dma_ops, but the rest of the calls in this
> file are to the proper DMA API interface so at least this patch definitely
> makes things less wrong in that respect.
>
Ion currently lacks any standard way to set up heaps and associate a device
with a heap. This means it's basically a free for all for what devices get
associated (getting something mainlined might help...). I agree that using
the proper DMA APIs is a step in the right direction.
> Reviewed-by: Robin Murphy <[email protected]>
>
>> ---
>>
>> This also begs the question as to what happens if the memory region _is_
>> contiguous but is in highmem or an ioremapped region. Should a device
>> always provide dma_ops for that case? Because I believe the current
>> implementation of dma_common_get_sgtable won't work for those as it uses
>> virt_to_page.
>>
>> I see that this point has been raised before [1] by Zeng Tao, and I
>> myself have been given a different fix to apply to a Linaro kernel tree.
>> However, both solutions looked wrong to me as they treat a dma_addr_t as
>> a physical address, so should at least be using dma_to_phys.
>> So, should we fix dma_common_get_sgtable or mandate that the device
>> has dma_ops? The latter seems to be implied by the commit message which
>> introduced the function:
>>
>> This patch provides a generic implementation based on
>> virt_to_page() call. Architectures which require more
>> sophisticated translation might provide their own get_sgtable()
>> methods.
>
> Given that we're largely here due to having poked this on arm64 systems,
> I'm inclined to think that implementing our own get_sgtable as per arch/arm
> is the right course of action. Since a lot of architectures using
> dma_common_get_sgtable don't even implement dma_to_phys, I don't think it
> would be right to try complicating the common code for a case that seems to
> be all but common. I can spin an arm64 patch if you like.
>
This would be hit on any system that has non-coherent DMA or highmem. I'm
not sure I agree this isn't a common case. How many of the other
architectures are actually using the dma_get_sgtable and would have the
potential to find a problem?
Thanks,
Laura
Hi Laura,
On 17/07/15 17:50, Laura Abbott wrote:
> On 07/17/2015 08:21 AM, Robin Murphy wrote:
>> Hi Tixy,
>>
>> On 17/07/15 12:01, Jon Medhurst (Tixy) wrote:
>>> Use dma_get_sgtable rather than dma_common_get_sgtable so a device's
>>> dma_ops aren't bypassed. This is essential in situations where a device
>>> uses an IOMMU and the physical memory is not contiguous (as the common
>>> function assumes).
>>>
>>> Signed-off-by: Jon Medhurst <[email protected]>
>>
>> The lack of obvious users of this code makes it hard to tell if "dev"
>> hereis always the right, real, device pointer and never null or some
>> dummy device with the wrong dma_ops, but the rest of the calls in this
>> file are to the proper DMA API interface so at least this patch definitely
>> makes things less wrong in that respect.
>>
>
>
> Ion currently lacks any standard way to set up heaps and associate a device
> with a heap. This means it's basically a free for all for what devices get
> associated (getting something mainlined might help...). I agree that using
> the proper DMA APIs is a step in the right direction.
I suspected as much, thanks for the confirmation.
>
>> Reviewed-by: Robin Murphy <[email protected]>
>>
>>> ---
>>>
>>> This also begs the question as to what happens if the memory region _is_
>>> contiguous but is in highmem or an ioremapped region. Should a device
>>> always provide dma_ops for that case? Because I believe the current
>>> implementation of dma_common_get_sgtable won't work for those as it uses
>>> virt_to_page.
>>>
>>> I see that this point has been raised before [1] by Zeng Tao, and I
>>> myself have been given a different fix to apply to a Linaro kernel tree.
>>> However, both solutions looked wrong to me as they treat a dma_addr_t as
>>> a physical address, so should at least be using dma_to_phys.
>>> So, should we fix dma_common_get_sgtable or mandate that the device
>>> has dma_ops? The latter seems to be implied by the commit message which
>>> introduced the function:
>>>
>>> This patch provides a generic implementation based on
>>> virt_to_page() call. Architectures which require more
>>> sophisticated translation might provide their own get_sgtable()
>>> methods.
>>
>> Given that we're largely here due to having poked this on arm64 systems,
>> I'm inclined to think that implementing our own get_sgtable as per arch/arm
>> is the right course of action. Since a lot of architectures using
>> dma_common_get_sgtable don't even implement dma_to_phys, I don't think it
>> would be right to try complicating the common code for a case that seems to
>> be all but common. I can spin an arm64 patch if you like.
>>
>
> This would be hit on any system that has non-coherent DMA or highmem. I'm
> not sure I agree this isn't a common case. How many of the other
> architectures are actually using the dma_get_sgtable and would have the
> potential to find a problem?
This appears to be pretty much exclusively a graphics/video thing.
Surveying in-tree callers (other than Ion) gives DRM, V4L, and a couple
of specific ARM SoC drivers - my hunch is that none of those see much
action on the likes of Blackfin and 68k.
That said, going through the git logs, the primary purpose of
dma_common_get_sgtable would appear to be not breaking allmodconfig
builds on architectures other than ARM. Thus I'm not really sure which
is the least worst option - having "common" code which doesn't actually
represent the common use case, or adding bogus dma_to_phys definitions
to loads of architectures that don't even have proper DMA mapping
implementations for the sake of some code they don't even use...
Robin.
>
> Thanks,
> Laura
>