Re: [RFC PATCH 29/30] softirq: Make softirq processing softinterruptible

[Pavan Kondeti]

Hi Frederic,

On Wed, Oct 17, 2018 at 02:26:02AM +0200, Frederic Weisbecker wrote:
> Hi Pavan,
> 
> On Tue, Oct 16, 2018 at 09:45:52AM +0530, Pavan Kondeti wrote:
> > Hi Frederic,
> > 
> > On Thu, Oct 11, 2018 at 01:12:16AM +0200, Frederic Weisbecker wrote:
> > > From: Frederic Weisbecker <fweisbec@xxxxxxxxx>
> > > 
> > > Make do_softirq() re-entrant and allow a vector, being either processed
> > > or disabled, to be interrupted by another vector. This way a vector
> > > won't be able to monopolize the CPU for a long while at the expense of
> > > the others that may rely on some predictable latency, especially on
> > > softirq disabled sections that used to disable all vectors.
> > > 
> > I understand that a long running softirq can be preempted/interrupted by
> > other softirqs which is not possible today. I have few questions on your
> > patches.
> > 
> > (1) When softirq processing is pushed to ksoftirqd, then the long running
> > softirq can still block other softirqs (not in SOFTIRQ_NOW_MASK) for a while.
> > correct?
> 
> No, Ksoftirqd is treated the same as IRQ tail processing here: a vector can
> interrupt another. So for example, a NET_RX softirq running in Ksoftirqd can
> be interrupted by a TIMER softirq running in hardirq tail.
> 
When ksoftirqd is running, we are only allowing softirqs in SOFTIRQ_NOW_MASK
to run after serving an interrupt. So I don't see how TIMER which is not
in SOFTIRQ_NOW_MASK can interrupt a NET_RX softirq running in ksoftirqd
context.

> > 
> > (2) When softirqs processing happens asynchronously, a particular softirq
> > like TASKLET can keep interrupting an already running softirq like TIMER/NET_RX,
> > correct? In worse case scenario, a long running softirq like NET_RX interrupt
> > a TIMER softirq. But I guess this is something expected with this. i.e
> > each softirq is independent and whichever comes recent gets to interrupt the
> > previously running softirqs.
> 
> Exactly, and that's inherent with interrupts in general. The only way to work
> around that is to thread each vector independantly but that's a whole different
> dimension :-)
> 
Right.

Assigning a thread for each vector also may not solve this problem because
preemption would be disabled while a softirq vector is running in its own
thread.

I guess there is no hard priorities among softirq vectors. Earlier
it was like first come first serve, now it is not. If we had priorities
defined, (don't know how :-)) we could disable the lower prio vectors while a
higher prio vector is being handled. This way we could gaurantee that TIMER
softirq or HI-TASKLET won't be starved while a long running softirq like
NET_RX/NET_TX/RCU is running.

Thanks,
Pavan
-- 
Qualcomm India Private Limited, on behalf of Qualcomm Innovation Center, Inc.
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a Linux Foundation Collaborative Project.