Re: [RFC 1/1] drivers/dma/*: replace tasklets with workqueue

[Vinod Koul <vkoul@kernel.org>]

On 27-05-22, 12:59, Arnd Bergmann wrote:
> On Fri, May 27, 2022 at 10:06 AM Vinod Koul <vkoul@kernel.org> wrote:
> > On 25-05-22, 13:03, Arnd Bergmann wrote:
> > > What might work better in the case of the dmaengine API would
> > > be an approach like:
> > >
> > > 1. add helper functions to call the callback functions from a
> > >     tasklet locally defined in drivers/dma/dmaengine.c to allow
> > >     deferring it from hardirq context
> > >
> > > 2. Change all  tasklets that are not part of the callback
> > >     mechanism to work queue functions, I only see
> > >     xilinx_dpdma_chan_err_task in the patch, but there
> > >     may be more
> > >
> > > 3. change all drivers to move their custom tasklets back into
> > >     hardirq context and instead call the new helper for deferring
> > >     the callback.
> > >
> > > 4. Extend the dmaengine callback API to let slave drivers
> > >     pick hardirq, tasklet or task context for the callback.
> > >     task context can mean either a workqueue, or a threaded
> > >     IRQ here, with the default remaining the tasklet version.
> >
> > That does sound a good idea, but I dont know who will use the workqueue
> > or a threaded context here, it might be that most would default to
> > hardirq or tasklet context for obvious reasons...
> 
> If the idea is to remove tasklets from the kernel for good, then the
> choice is only between workqueue and hardirq at this point. The
> workqueue version is the one that would make sense for any driver
> that just defers execution from the callback down into task context.
> If that gets called in task context already, the driver can be simpler.
> 
> I took a brief look at the roughly 150 slave drivers, and it does
> seem like very few of them actually want task context:
> 
> * Over Half the drivers just do a complete(), which could
>   probably be pulled into the dmaengine layer and done from
>   hardirq, avoiding the callback entirely
> 
> * A lot of the remaining drivers have interrupts disabled for
>   the entire callback, which means they might as well use
>   hardirqs, regardless of what they want
> 
> * drivers/crypto/* and drivers/mmc/* tend to call another tasklet
>   to do the real work.
> 
> * drivers/ata/sata_dwc_460ex.c and drivers/ntb/ntb_transport.c
>    probably want task context
> 
> * Some drivers like sound/soc/sh/siu_pcm.c start a new DMA
>   from the callback. Is that allowed from hardirq?
> 
> If we do the first three steps above, and then add a 'struct
> completion' pointer to dma_async_tx_descriptor as an alternative
> to the callback, that would already reduce the number of drivers
> that end up in a tasklet significantly and should be completely
> safe.

That is a good idea, lot of drivers are waiting for completion which can
be signalled from hardirq, this would also reduce the hops we have and
help improve latency a bit. On the downside, some controllers provide
error information, which would need to be dealt with.

I will prototype this on Qcom boards I have...

> 
> Unfortunately we can't just move the rest into hardirq
> context because that breaks anything using spin_lock_bh
> to protect against concurrent execution of the tasklet.
> 
> A possible alternative might be to then replace the global
> dmaengine tasklet with a custom softirq. Obviously those
> are not so hot either,  but dmaengine could be considered
> special enough to fit in the same category as net_rx/tx
> and block with their global softirqs.

Yes that would be a very reasonable mechanism, thanks for the
suggestions.

-- 
~Vinod