Re: High interrupt latency with low power idle mode on i.MX6

From: Russell King - ARM Linux admin
Date: Wed May 27 2020 - 09:24:25 EST

On Wed, May 27, 2020 at 12:50:01PM +0000, Schrempf Frieder wrote:
> On 27.05.20 13:53, Russell King - ARM Linux admin wrote:
> > On Wed, May 27, 2020 at 10:39:12AM +0000, Schrempf Frieder wrote:
> >> Hi,
> >>
> >> on our i.MX6UL/ULL boards running mainline kernels, we see an issue with
> >> RS485 collisions on the bus. These are caused by the resetting of the
> >> RTS signal being delayed after each transmission. The TXDC interrupt
> >> takes several milliseconds to trigger and the slave on the bus already
> >> starts to send a reply in the meantime.
> >>
> >> We found out that these delays only happen when the CPU is in "low power
> >> idle" mode (ARM power off). When we disable cpuidle state 2 or put some
> >> background load on the CPU everything works fine and the delays are gone.
> >>
> >> echo 1 > /sys/devices/system/cpu/cpu0/cpuidle/state2/disable
> >>
> >> It seems like also other interfaces (I2C, etc.) might be affected by
> >> these increased latencies, we haven't investigated this more closely,
> >> though.
> >>
> >> We currently apply a patch to our kernel, that disables low power idle
> >> mode by default, but I'm wondering if there's a way to fix this
> >> properly? Any ideas?
> >
> > Let's examine a basic fact about power management:
> >
> > The deeper PM modes that the system enters, the higher the latency to
> > resume operation.
> >
> > So, I'm not surprised that you have higher latency when you allow the
> > system to enter lower power modes. Does that mean that the kernel
> > should not permit entering lower power modes - no, it's policy and
> > application dependent.
> >
> > If the hardware is designed to use software to manage the RTS signal
> > to control the RS485 receiver, then I'm afraid that your report really
> > does not surprise me - throwing that at software to manage is a really
> > stupid idea, but it seems lots of people do this. I've held this view
> > since I worked on a safety critical system that used RS485 back in the
> > 1990s (London Underground Jubilee Line Extension public address system.)
> >
> > So, what we have here is several things that come together to create a
> > problem:
> >
> > 1) higher power savings produce higher latency to resume from
> > 2) lack of hardware support for RS485 half duplex communication needing
> > software support
> > 3) an application that makes use of RS485 half duplex communication
> > without disabling the higher latency power saving modes
> >
> > The question is, who should disable those higher latency power saving
> > modes - the kernel, or userspace?
> >
> > The kernel knows whether it needs to provide software control of the
> > RTS signal or not, but the kernel does not know the maximum permissible
> > latency (which is application specific.) So, the kernel doesn't have
> > all the information it needs. However, there is a QoS subsystem which
> > may help you.
> >
> > There's also tweaks available via
> > /sys/devices/system/cpu/cpu*/power/pm_qos_resume_latency_us
> >
> > which can be poked to configure the latency that is required, and will
> > prevent the deeper PM states being entered.
> Thanks for the detailed explanation. This all makes perfect sense to me.
> I will keep in mind that we need to consider this aspect of power saving
> vs. latency when designing systems and also that we need to provide the
> information for the kernel to decide which of the two is more important.
> Also thanks for pointing out the QoS subsystem. I'm not quite sure if it
> would work for us to use pm_qos_resume_latency_us in our specific case.
> The actual latency we observe is something like 2 to 3 milliseconds
> longer with low power idle than without, but the exit_latency for low
> power idle specified in the cpuidle driver is only 300 us.

I wonder whether the exit latencies are correct in that case.
>From the comments, it seems 80us is allowed for the software overhead
of entering/leaving the idle state vs 220us for the hardware.
It may be a good idea for someone to add some tracing points in there
to try and measure the minimum software latencies.

> So as far as I can see with this difference even if we would set
> pm_qos_resume_latency_us to 1000 us (which should be fast enough for the
> RS485 to work properly), the low power idle wouldn't be disabled.
> It's rather this discrepancy between the latency set in the driver and
> what we see in reality which makes me wonder if there's something I'm
> missing.

It's possible that there's something missing from the kernel's
estimation of the latency required for entering / exiting those

There is an amount of cache flushing that is required when entering
those lower states, and I wonder if that has been accounted for.

RMK's Patch system:
FTTC for 0.8m (est. 1762m) line in suburbia: sync at 13.1Mbps down 424kbps up