[RFC] Is lib/irq_poll still considered useful?

[RFC] Is lib/irq_poll still considered useful?

[Sebastian Andrzej Siewior]

I was looking at irq_poll and for missing scheduling points.
It raised the question why are there 7 driver still using irq_poll and
not moved on to something else like threaded interrupts or kworker.

There is:
- Infiband can complete direct, irq_poll and kworker.
- be2iscsi only irq_poll.
- cxlflash only irq_poll. Not sure how IRQs are acked.
- ipr direct or irq_poll, can be configured. Now sure how IRQs are acked.
- lpfc kworker and/or irq_poll. Not sure all invocations are from
  interrupts like context [0].
- megaraid irq_poll. Not sure all invocations are from interrupts like
  context [0].
- mpt3sas irq_poll or io_uring io poll. Not sure all invocations are
  from interrupts like context [0].

[0] If irq_poll_sched() is not used from an interrupt (as in interrupt
service routine, timer handler, tasklet (not encouraging just noticed))
but from task context (as in kworker for instance) then irq-poll handler
will not be invoked right away. Instead it will be delayed to random
point in time until an interrupt fires or something down the stack
performs a pending softirq check.
Waking ksoftirqd itself isn't helping much since it will set a
NEED_RESCHED bit in the task_struct which local_irq_restore() isn't
testing for. So the scheduling event is delayed until spin_unlock() for
instance.

Is there a reason for the remaining user of irq_poll to keep using it?

Sebastian

Re: [RFC] Is lib/irq_poll still considered useful?

[Christoph Hellwig]

On Thu, Sep 30, 2021 at 12:56:05PM +0200, Sebastian Andrzej Siewior wrote:
> Is there a reason for the remaining user of irq_poll to keep using it?

At least for RDMA there are workloads where the latency difference
matters.  That's why we added both the irq_poll and workqueue mode
to thew new CQ API a few years ago.

Re: [RFC] Is lib/irq_poll still considered useful?

[Sebastian Andrzej Siewior]

On 2021-10-01 05:24:49 [+0100], Christoph Hellwig wrote:
> On Thu, Sep 30, 2021 at 12:56:05PM +0200, Sebastian Andrzej Siewior wrote:
> > Is there a reason for the remaining user of irq_poll to keep using it?
> 
> At least for RDMA there are workloads where the latency difference
> matters.  That's why we added both the irq_poll and workqueue mode
> to thew new CQ API a few years ago.

Would it work for them to move to threaded interrupts or is the NAPI
like behaviour (delay after a while to the next jiffy) the killer
feature?

Sebastian