idxd: shared workqueues and dmaengine

idxd: shared workqueues and dmaengine

[Kevin Krakauer]

Hey, I'm trying to use idxd inside the kernel via dmaengine. I'm
attempting to get shared work queues (SWQs) via dma_find_channel(),
which stubbornly returns NULL.

After a lot of printk-ing, I finally ran across DMA_PRIVATE. I also
found [1] and saw that DMA_PRIVATE is also set in the idxd driver since
c06e424be5f51844 ("dmaengine: idxd: set DMA channel to be private").

It seems like this forces use of dma_request_channel(). But I believe
the idea behind SWQs is that they don't have to be exclusively owned.

Is there a way to use SWQs via dma_find_channel()? dma_request_channel()
seems like overkill, as the point of SWQs is to not be exclusive. Also a
good chance I'm missing something basic here :)

Thanks!
Kevin

1 - https://lore.kernel.org/dmaengine/9714cb6c-8b37-4fc6-bb09-a1fb4a5bb1b8@intel.com/#R

Re: idxd: shared workqueues and dmaengine

[Dave Jiang]

On 3/26/25 5:05 PM, Kevin Krakauer wrote:
> Hey, I'm trying to use idxd inside the kernel via dmaengine. I'm
> attempting to get shared work queues (SWQs) via dma_find_channel(),
> which stubbornly returns NULL.
> 
> After a lot of printk-ing, I finally ran across DMA_PRIVATE. I also
> found [1] and saw that DMA_PRIVATE is also set in the idxd driver since
> c06e424be5f51844 ("dmaengine: idxd: set DMA channel to be private").
> 
> It seems like this forces use of dma_request_channel(). But I believe
> the idea behind SWQs is that they don't have to be exclusively owned.
> 
> Is there a way to use SWQs via dma_find_channel()? dma_request_channel()
> seems like overkill, as the point of SWQs is to not be exclusive. Also a
> good chance I'm missing something basic here :)

+Vinicius, current maintainer of the driver.

So I think maybe a new sysfs attribute can be added to wq to allow the user to choose whether to make the channel DMA_PRIVATE or not. By default it can be set to DMA_PRIVATE to keep legacy behavior. I think that should solve the issue you are encountering?

> 
> Thanks!
> Kevin
> 
> 1 - https://lore.kernel.org/dmaengine/9714cb6c-8b37-4fc6-bb09-a1fb4a5bb1b8@intel.com/#R

Re: idxd: shared workqueues and dmaengine

[Kevin Krakauer]

On Wed, Mar 26, 2025 at 05:15:37PM -0700, Dave Jiang wrote:
> So I think maybe a new sysfs attribute can be added to wq to allow the user to choose whether to make the channel DMA_PRIVATE or not. By default it can be set to DMA_PRIVATE to keep legacy behavior. I think that should solve the issue you are encountering?

Yep, that would do it. Thanks for helping me understand -- I was
wondering whether I was missing something.

Easy for me to workaround for now, or maybe I'll just add the sysfs
attribute and send it to the list. Thanks!

Re: idxd: shared workqueues and dmaengine

[Dave Jiang]

On 3/27/25 11:56 AM, Kevin Krakauer wrote:
> On Wed, Mar 26, 2025 at 05:15:37PM -0700, Dave Jiang wrote:
>> So I think maybe a new sysfs attribute can be added to wq to allow the user to choose whether to make the channel DMA_PRIVATE or not. By default it can be set to DMA_PRIVATE to keep legacy behavior. I think that should solve the issue you are encountering?
> 
> Yep, that would do it. Thanks for helping me understand -- I was
> wondering whether I was missing something.
> 
> Easy for me to workaround for now, or maybe I'll just add the sysfs
> attribute and send it to the list. Thanks!

Patches are always welcome. I'm sure Vinicius would appreciate it.