Re: [PATCH] PCI: endpoint: Add DMA to Linux PCI EP Framework

[Kishon Vijay Abraham I]

Hi Vinod,

On 31/05/19 12:02 PM, Vinod Koul wrote:
> On 31-05-19, 10:50, Kishon Vijay Abraham I wrote:
>> Hi Vinod,
>>
>> On 31/05/19 10:37 AM, Vinod Koul wrote:
>>> Hi Kishon,
>>>
>>> On 30-05-19, 11:16, Kishon Vijay Abraham I wrote:
>>>> +Vinod Koul
>>>>
>>>> Hi,
>>>>
>>>> On 30/05/19 4:07 AM, Alan Mikhak wrote:
>>>>> On Mon, May 27, 2019 at 2:09 AM Gustavo Pimentel
>>>>> <Gustavo.Pimentel@xxxxxxxxxxxx> wrote:
>>>>>>
>>>>>> On Fri, May 24, 2019 at 20:42:43, Alan Mikhak <alan.mikhak@xxxxxxxxxx>
>>>>>> wrote:
>>>>>>
>>>>>> Hi Alan,
>>>>>>
>>>>>>> On Fri, May 24, 2019 at 1:59 AM Gustavo Pimentel
>>>>>>> <Gustavo.Pimentel@xxxxxxxxxxxx> wrote:
>>>>>>>>
>>>>>>>> Hi Alan,
>>>>>>>>
>>>>>>>> This patch implementation is very HW implementation dependent and
>>>>>>>> requires the DMA to exposed through PCIe BARs, which aren't always the
>>>>>>>> case. Besides, you are defining some control bits on
>>>>>>>> include/linux/pci-epc.h that may not have any meaning to other types of
>>>>>>>> DMA.
>>>>>>>>
>>>>>>>> I don't think this was what Kishon had in mind when he developed the
>>>>>>>> pcitest, but let see what Kishon was to say about it.
>>>>>>>>
>>>>>>>> I've developed a DMA driver for DWC PCI using Linux Kernel DMAengine API
>>>>>>>> and which I submitted some days ago.
>>>>>>>> By having a DMA driver which implemented using DMAengine API, means the
>>>>>>>> pcitest can use the DMAengine client API, which will be completely
>>>>>>>> generic to any other DMA implementation.
>>>>
>>>> right, my initial thought process was to use only dmaengine APIs in
>>>> pci-epf-test so that the system DMA or DMA within the PCIe controller can be
>>>> used transparently. But can we register DMA within the PCIe controller to the
>>>> DMA subsystem? AFAIK only system DMA should register with the DMA subsystem.
>>>> (ADMA in SDHCI doesn't use dmaengine). Vinod Koul can confirm.
>>>
>>> So would this DMA be dedicated for PCI and all PCI devices on the bus?
>>
>> Yes, this DMA will be used only by PCI ($patch is w.r.t PCIe device mode. So
>> all endpoint functions both physical and virtual functions will use the DMA in
>> the controller).
>>> If so I do not see a reason why this cannot be using dmaengine. The use
>>
>> Thanks for clarifying. I was under the impression any DMA within a peripheral
>> controller shouldn't use DMAengine.
> 
> That is indeed a correct assumption. The dmaengine helps in cases where
> we have a dma controller with multiple users, for a single user case it
> might be overhead to setup dma driver and then use it thru framework.
> 
> Someone needs to see the benefit and cost of using the framework and
> decide.

The DMA within the endpoint controller can indeed be used by multiple users for
e.g in the case of multi function EP devices or SR-IOV devices, all the
function drivers can use the DMA in the endpoint controller.

I think it makes sense to use dmaengine for DMA within the endpoint controller.
> 
>>> case would be memcpy for DMA right or mem to device (vice versa) transfers?
>>
>> The device is memory mapped so it would be only memcopy.
>>>
>>> Btw many driver in sdhci do use dmaengine APIs and yes we are missing
>>> support in framework than individual drivers
>>
>> I think dmaengine APIs is used only when the platform uses system DMA and not
>> ADMA within the SDHCI controller. IOW there is no dma_async_device_register()
>> to register ADMA in SDHCI with DMA subsystem.
> 
> We are looking it from the different point of view. You are looking for
> dmaengine drivers in that (which would be in drivers/dma/) and I am
> pointing to users of dmaengine in that.
> 
> So the users in mmc would be ones using dmaengine APIs:
> $git grep -l dmaengine_prep_* drivers/mmc/
> 
> which tells me 17 drivers!

right. For the endpoint case, drivers/pci/controller should register with the
dmaengine i.e if the controller has aN embedded DMA (I think it should be okay
to keep that in drivers/pci/controller itself instead of drivers/dma) and
drivers/pci/endpoint/functions/ should use dmaengine API's (Depending on the
platform, this will either use system DMA or DMA within the PCI controller).

Thanks
Kishon