Re: [PATCH 00/37] softirq: Per vector masking v3

From: Frederic Weisbecker
Date: Fri Mar 08 2019 - 10:30:29 EST

On Fri, Mar 01, 2019 at 08:51:38AM -0800, Linus Torvalds wrote:
> On Thu, Feb 28, 2019 at 7:45 PM Frederic Weisbecker <frederic@xxxxxxxxxx> wrote:
> >
> > Numbers are indeed missing. In fact this patchset mostly just brings an
> > infrastructure. We have yet to pinpoint the most latency-inducing
> > softirq disabled sites and make them disable only the vectors that
> > are involved in a given lock.
> Note that I think we pretty much know that already: the people who
> have had issues have never actually really had issues with the actual
> "disable softirq" paths, they've all been about actually *running* the
> softirq's (and that in turn being a latency issue for running _other_
> softirqs, and in particular for delaying them into softirqd).
> Now, it may well be that yes, we'll have "block softirqs" code that
> has issues too, but it's absolutely been swamped by the latencies for
> actually running them so far.
> Note that this is all really fairly independent of the whole masking
> logic. Yes, the masking logic comes into play too (allowing you to run
> a subset of softirq's at a time), but on the whole the complaints I've
> seen have not been "the networking softirq takes so long that it
> delays USB tasklet handling", but they have been along the lines of
> "the networking softirq gets invoked so often that it then floods the
> system and triggers softirqd, and _that_ then makes tasklet handling
> latency go up insanely".
> See the difference? Not the latency of softirq's disabled, but the
> latency of one group of softirqs causing problems for another when
> they all get batched together (and soft-scheduled to another context
> together).

I see, so that's an entirely different problem that vector
soft-interruptibility can't fix, at least not alone.

The only solution I can imagine is to have a seperate pending mask
for normal softirq processing and ksoftirqd, so that only vectors
that have been enqueued for threaded processing are delayed.

I can work on that first, but I really need to be able to reproduce
an example of the issue. The USB capture thing seems to be one the best.
Let's browse some history to see if I can find some details on the
relevant scenario.