On Thu, Mar 28, 2024, at 06:55, Vinod Koul wrote:
> On 27-03-24, 16:03, Allen Pais wrote:
>> The only generic interface to execute asynchronously in the BH context is
>> tasklet; however, it's marked deprecated and has some design flaws. To
>> replace tasklets, BH workqueue support was recently added. A BH workqueue
>> behaves similarly to regular workqueues except that the queued work items
>> are executed in the BH context.
>
> Thanks for conversion, am happy with BH alternative as it helps in
> dmaengine where we need shortest possible time between tasklet and
> interrupt handling to maximize dma performance
I still feel that we want something different for dmaengine,
at least in the long run. As we have discussed in the past,
the tasklet context in these drivers is what the callbacks
from the dma client device is run in, and a lot of these probably
want something other than tasklet context, e.g. just call
complete() on a client-provided completion structure.
Instead of open-coding the use of the system_bh_wq in each
dmaengine, how about we start with a custom WQ_BH
specifically for the dmaengine subsystem and wrap them
inside of another interface.
Since almost every driver associates the tasklet with the
dma_chan, we could go one step further and add the
work_queue structure directly into struct dma_chan,
with the wrapper operating on the dma_chan rather than
the work_queue.
Arnd
