Nouveau - Jul 2014 - [PATCH v4 2/6] drm/nouveau: map pages using DMA API on platform devices

On Fri, Jul 11, 2014 at 12:35 PM, Alexandre Courbot <acourbot at
nvidia.com> wrote:
> On 07/10/2014 09:58 PM, Daniel Vetter wrote:
>>
>> On Tue, Jul 08, 2014 at 05:25:57PM +0900, Alexandre Courbot wrote:
>>>
>>> page_to_phys() is not the correct way to obtain the DMA address of
a
>>> buffer on a non-PCI system. Use the DMA API functions for this,
which
>>> are portable and will allow us to use other DMA API functions for
>>> buffer synchronization.
>>>
>>> Signed-off-by: Alexandre Courbot <acourbot at nvidia.com>
>>> ---
>>>   drivers/gpu/drm/nouveau/core/engine/device/base.c | 8 +++++++-
>>>   1 file changed, 7 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/drivers/gpu/drm/nouveau/core/engine/device/base.c
>>> b/drivers/gpu/drm/nouveau/core/engine/device/base.c
>>> index 18c8c7245b73..e4e9e64988fe 100644
>>> --- a/drivers/gpu/drm/nouveau/core/engine/device/base.c
>>> +++ b/drivers/gpu/drm/nouveau/core/engine/device/base.c
>>> @@ -489,7 +489,10 @@ nv_device_map_page(struct nouveau_device
*device,
>>> struct page *page)
>>>                 if (pci_dma_mapping_error(device->pdev, ret))
>>>                         ret = 0;
>>>         } else {
>>> -               ret = page_to_phys(page);
>>> +               ret =
dma_map_page(&device->platformdev->dev, page, 0,
>>> +                                  PAGE_SIZE, DMA_BIDIRECTIONAL);
>>> +               if
(dma_mapping_error(&device->platformdev->dev, ret))
>>> +                       ret = 0;
>>>         }
>>>
>>>         return ret;
>>> @@ -501,6 +504,9 @@ nv_device_unmap_page(struct nouveau_device
*device,
>>> dma_addr_t addr)
>>>         if (nv_device_is_pci(device))
>>>                 pci_unmap_page(device->pdev, addr, PAGE_SIZE,
>>>                                PCI_DMA_BIDIRECTIONAL);
>>
>>
>> pci_map/unmap alias to dma_unmap/map when called on the underlying
struct
>> device embedded in pci_device (like for platform drivers). Dunno
whether
>> it's worth to track a pointer to the struct device directly and
always
>> call dma_unmap/map.
>
>
> Isn't it (theoretically) possible to have a platform that does not use
the
> DMA API for its PCI implementation and thus requires the pci_* functions to
> be called? I could not find such a case in -next, which suggests that all
> PCI platforms have been converted to the DMA API already and that we could
> indeed refactor this to always use the DMA functions.
>
> But at the same time the way we use APIs should not be directed by their
> implementation, but by their intent - and unless the PCI API has been
> deprecated in some way (something I am not aware of), the rule is still
that
> you should use it on a PCI device.
>
>
>>
>> Just drive-by comment since I'm interested in how you solve this -
i915
>> has similar fun with buffer sharing and coherent and non-coherent
>> platforms. Although we don't have fun with pci and non-pci based
>> platforms.
>
>
> Yeah, I am not familiar with i915 but it seems like we are on a similar
boat
> here (excepted ARM is more constrained as to its memory mappings). The
> strategy in this series is, map buffers used by user-space cached and
> explicitly synchronize them (since the ownership transition from user to
GPU
> is always clearly performed by syscalls), and use coherent mappings for
> buffers used by the kernel which are accessed more randomly. This has
solved
> all our coherency issues and resulted in the best performance so far.
I wonder if we might want to use unsnooped cached mappings of pages on
non-ARM platforms also, to avoid the overhead of the cache snooping?
>
>
> _______________________________________________
> dri-devel mailing list
> dri-devel at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/dri-devel

On 07/11/2014 11:50 AM, Ben Skeggs wrote:
> On Fri, Jul 11, 2014 at 12:35 PM, Alexandre Courbot <acourbot at
nvidia.com> wrote:
>> On 07/10/2014 09:58 PM, Daniel Vetter wrote:
>>>
>>> On Tue, Jul 08, 2014 at 05:25:57PM +0900, Alexandre Courbot wrote:
>>>>
>>>> page_to_phys() is not the correct way to obtain the DMA address
of a
>>>> buffer on a non-PCI system. Use the DMA API functions for this,
which
>>>> are portable and will allow us to use other DMA API functions
for
>>>> buffer synchronization.
>>>>
>>>> Signed-off-by: Alexandre Courbot <acourbot at nvidia.com>
>>>> ---
>>>>    drivers/gpu/drm/nouveau/core/engine/device/base.c | 8
+++++++-
>>>>    1 file changed, 7 insertions(+), 1 deletion(-)
>>>>
>>>> diff --git a/drivers/gpu/drm/nouveau/core/engine/device/base.c
>>>> b/drivers/gpu/drm/nouveau/core/engine/device/base.c
>>>> index 18c8c7245b73..e4e9e64988fe 100644
>>>> --- a/drivers/gpu/drm/nouveau/core/engine/device/base.c
>>>> +++ b/drivers/gpu/drm/nouveau/core/engine/device/base.c
>>>> @@ -489,7 +489,10 @@ nv_device_map_page(struct nouveau_device
*device,
>>>> struct page *page)
>>>>                  if (pci_dma_mapping_error(device->pdev,
ret))
>>>>                          ret = 0;
>>>>          } else {
>>>> -               ret = page_to_phys(page);
>>>> +               ret =
dma_map_page(&device->platformdev->dev, page, 0,
>>>> +                                  PAGE_SIZE,
DMA_BIDIRECTIONAL);
>>>> +               if
(dma_mapping_error(&device->platformdev->dev, ret))
>>>> +                       ret = 0;
>>>>          }
>>>>
>>>>          return ret;
>>>> @@ -501,6 +504,9 @@ nv_device_unmap_page(struct nouveau_device
*device,
>>>> dma_addr_t addr)
>>>>          if (nv_device_is_pci(device))
>>>>                  pci_unmap_page(device->pdev, addr,
PAGE_SIZE,
>>>>                                 PCI_DMA_BIDIRECTIONAL);
>>>
>>>
>>> pci_map/unmap alias to dma_unmap/map when called on the underlying
struct
>>> device embedded in pci_device (like for platform drivers). Dunno
whether
>>> it's worth to track a pointer to the struct device directly and
always
>>> call dma_unmap/map.
>>
>>
>> Isn't it (theoretically) possible to have a platform that does not
use the
>> DMA API for its PCI implementation and thus requires the pci_*
functions to
>> be called? I could not find such a case in -next, which suggests that
all
>> PCI platforms have been converted to the DMA API already and that we
could
>> indeed refactor this to always use the DMA functions.
>>
>> But at the same time the way we use APIs should not be directed by
their
>> implementation, but by their intent - and unless the PCI API has been
>> deprecated in some way (something I am not aware of), the rule is still
that
>> you should use it on a PCI device.
>>
>>
>>>
>>> Just drive-by comment since I'm interested in how you solve
this - i915
>>> has similar fun with buffer sharing and coherent and non-coherent
>>> platforms. Although we don't have fun with pci and non-pci
based
>>> platforms.
>>
>>
>> Yeah, I am not familiar with i915 but it seems like we are on a similar
boat
>> here (excepted ARM is more constrained as to its memory mappings). The
>> strategy in this series is, map buffers used by user-space cached and
>> explicitly synchronize them (since the ownership transition from user
to GPU
>> is always clearly performed by syscalls), and use coherent mappings for
>> buffers used by the kernel which are accessed more randomly. This has
solved
>> all our coherency issues and resulted in the best performance so far.
> I wonder if we might want to use unsnooped cached mappings of pages on
> non-ARM platforms also, to avoid the overhead of the cache snooping?
You might want to indeed, now that coherency is guaranteed by the sync 
functions originally introduced by Lucas. The only issue I could see is 
that they always invalidate the full buffer whereas bus snooping only 
affects pages that are actually touched. Someone would need to try this 
on a desktop machine and see how it affects performance.

I'd be all for it though, since it would also allow us to get rid of 
this ungraceful nv_device_is_cpu_coherent() function and result in 
simplifying nouveau_bo.c a bit.

Am Freitag, den 11.07.2014, 11:57 +0900 schrieb Alexandre Courbot:
[...]
> >> Yeah, I am not familiar with i915 but it seems like we are on a
similar boat
> >> here (excepted ARM is more constrained as to its memory mappings).
The
> >> strategy in this series is, map buffers used by user-space cached
and
> >> explicitly synchronize them (since the ownership transition from
user to GPU
> >> is always clearly performed by syscalls), and use coherent
mappings for
> >> buffers used by the kernel which are accessed more randomly. This
has solved
> >> all our coherency issues and resulted in the best performance so
far.
> > I wonder if we might want to use unsnooped cached mappings of pages on
> > non-ARM platforms also, to avoid the overhead of the cache snooping?
> 
> You might want to indeed, now that coherency is guaranteed by the sync 
> functions originally introduced by Lucas. The only issue I could see is 
> that they always invalidate the full buffer whereas bus snooping only 
> affects pages that are actually touched. Someone would need to try this 
> on a desktop machine and see how it affects performance.
> 
> I'd be all for it though, since it would also allow us to get rid of 
> this ungraceful nv_device_is_cpu_coherent() function and result in 
> simplifying nouveau_bo.c a bit.
This will need some testing to get hard numbers, but I suspect that
invalidating the whole buffer isn't to bad as the prefetch machinery
works very well with the access patterns we see in graphics drivers.

Flushing out the whole buffer should be even less problematic, as it
will only flush out dirty lines that would need to be flushed on GPU
read snooping anyways.

In the long run we might want a separate cpu prepare/finish ioctl where
we can indicate the area of interest. This might help to avoid some of
the invalidate overhead especially for userspace suballocated buffers.

Regards,
Lucas

-- 
Pengutronix e.K.             | Lucas Stach                 |
Industrial Linux Solutions   | http://www.pengutronix.de/  |