Re: [PATCH] usb: don't offload isochronous urb completions to ksoftirq

From: Mikulas Patocka
Date: Thu Jun 14 2018 - 18:23:19 EST




On Wed, 13 Jun 2018, Steven Rostedt wrote:

> On Wed, 13 Jun 2018 15:30:31 -0400 (EDT)
> Mikulas Patocka <mpatocka@xxxxxxxxxx> wrote:
>
> > On Wed, 13 Jun 2018, Alan Stern wrote:
> >
> > > [Steve: Sorry for dumping you into the middle of this discussion.
> > > Please see especially the last two paragraphs below. Mikulas is
> > > getting dropouts with USB audio because part of the processing uses a
> > > tasklet.]
> >
> > The problem is this:
> >
> > I have a single core machine with a usb2 sound card. When I increase the
> > priority of a music player, the audio starts skipping.
> >
> > The reason for the skipping is that the ehci usb driver is offloading urb
> > callbacks using tasklet_hi_schedule, the callbacks end up being offloaded
> > to the ksoftirqd thread (that has priority 0), the music player with
> > elevated priority preempts ksoftirqd and causes delays in the urb
> > callbacks.
> >
> > Is this some deficiency in the softirq subsystem? (should we perhaps treat
> > tasklet_hi specially and not offload it as much as the others?) Or should
> > the ehci driver be fixed not to use tasklets?
> >
>
> What we do for softirqs in the RT patch is to have whoever raised the
> softirq run the softirq. If local_bh_disabled() is active (bh is
> disabled) then a bit is set in the current task struct, where when
> local_bh_enable() is called, it will then execute the softirqs that it
> raised while bh was disabled.
>
> Perhaps try out the PREEMPT_RT patch and see if the problem goes away.

I tried the realtime patch with CONFIG_PREEMPT_RT_FULL and it plays well.

> Hopefully this softirq work may make it into the kernel soon. We could
> even enabled it without full PREEMPT_RT.

I don't think it's so easy. The kernel 2.4 and below did this. And the
problem was that if there's a storm of network packets, the softirq code
would cause lockup of the whole machine. In order to not lockup the
machine - somewhere in the 2.4 cycle - the ksoftirqd thread was
introduced.

If you have CONFIG_PREEMPT_RT_FULL and you call softirqs in the interrupt
thread, you could only stall the interrupt thread. If you do the same
thing without CONFIG_PREEMPT_RT_FULL, you stall the whole CPU.

Mikulas

> https://mirrors.edge.kernel.org/pub/linux/kernel/projects/rt/4.16/patch-4.16.12-rt5.patch.xz
>
> -- Steve
>