Hello Michal,

On Sun, Apr 25, 2021 at 01:07:14PM +0200, Michal Suchánek wrote:

On Sat, Apr 24, 2021 at 01:07:16PM +0530, Vaidyanathan Srinivasan wrote:

quoted

* Michal Such?nek [off-list ref] [2021-04-23 20:42:16]:

quoted

On Fri, Apr 23, 2021 at 11:59:30PM +0530, Vaidyanathan Srinivasan wrote:

quoted

* Michal Such?nek [off-list ref] [2021-04-23 19:45:05]:

quoted

On Fri, Apr 23, 2021 at 09:29:39PM +0530, Vaidyanathan Srinivasan wrote:

quoted

* Michal Such?nek [off-list ref] [2021-04-23 09:35:51]:

quoted

On Thu, Apr 22, 2021 at 08:37:29PM +0530, Gautham R. Shenoy wrote:

quoted

From: "Gautham R. Shenoy" <redacted>

Commit d947fb4c965c ("cpuidle: pseries: Fixup exit latency for
CEDE(0)") sets the exit latency of CEDE(0) based on the latency values
of the Extended CEDE states advertised by the platform

On some of the POWER9 LPARs, the older firmwares advertise a very low
value of 2us for CEDE1 exit latency on a Dedicated LPAR. However the

Can you be more specific about 'older firmwares'?

Hi Michal,

This is POWER9 vs POWER10 difference, not really an obsolete FW.  The
key idea behind the original patch was to make the H_CEDE latency and
hence target residency come from firmware instead of being decided by
the kernel.  The advantage is such that, different type of systems in
POWER10 generation can adjust this value and have an optimal H_CEDE
entry criteria which balances good single thread performance and
wakeup latency.  Further we can have additional H_CEDE state to feed
into the cpuidle.  

So all POWER9 machines are affected by the firmware bug where firmware
reports CEDE1 exit latency of 2us and the real latency is 5us which
causes the kernel to prefer CEDE1 too much when relying on the values
supplied by the firmware. It is not about 'older firmware'.

Correct.  All POWER9 systems running Linux as guest LPARs will see
extra usage of CEDE idle state, but not baremetal (PowerNV).

The correct definition of the bug or miss-match in expectation is that
firmware reports wakeup latency from a core/thread wakeup timing, but
not end-to-end time from sending a wakeup event like an IPI using
H_calls and receiving the events on the target.  Practically there are
few extra micro-seconds needed after deciding to wakeup a target
core/thread to getting the target to start executing instructions
within the LPAR instance.

Thanks for the detailed explanation.

Maybe just adding a few microseconds to the reported time would be a
more reasonable workaround than using a blanket fixed value then.

Yes, that is an option.  But that may only reduce the difference
between existing kernel and new kernel unless we make it the same
number.  Further we are fixing this in P10 and hence we will have to
add "if(P9) do the compensation" and otherwise take it as is.  That
would not be elegant.  Given that our goal for P9 platform is to not
introduce changes in H_CEDE entry behaviour, we arrived at this
approach (this small patch) and this also makes it easy to backport to
various distro products.

I don't see how this is more elegent.

The current patch is

if(p9)
	use fixed value

the suggested patch is

if(p9)
	apply compensation

We could do that, however, from the recent measurements the default
value is closer to the latency value measured using an IPI.

As Vaidy described earlier, on POWER9 and prior platforms, the wakeup
latency advertized by the PHYP hypervisor corresponds to the latency
required to wakeup from the underlying hardware idle state (Nap in
POWER8 and stop0/1/2 on POWER9) into the hypervisor. That's 2us on
POWER9.

We need to apply two kinds of compensation,

1. Compensation for the time taken to transition the CPU from the
   Hypervisor into the LPAR post wakeup from platform idle state

2. Compensation for the time taken to send the IPI from the source CPU
   (waker) to the idle target CPU (wakee).

1. can be measured via timer idle test (I am using Pratik's
cpuidle self-test posted here
https://lore.kernel.org/lkml/20210412074309.38484-1-psampat@linux.ibm.com/ (local))

We queue a timer, say for 1ms, and enter the CEDE state. When the
timer fires, in the timer handler we compute how much extra timer over
the expected 1ms have we consumed. This is what it looks like on
POWER9 LPAR

CEDE latency measured using a timer (numbers in ns)
===================================================================
N       Min      Median   Avg       90%ile  99%ile    Max    Stddev
400     2601     5677     5668.74    5917    6413     9299   455.01

If we consider the avg and the 99th %ile values, it takes on an avg
about somewhere between 3.5-4.5 us to transition from the Hypervisor
to the guest VCPU after the CPU has woken up from the idle state. 

1. and 2. combined can be determined by an IPI latency test (from the
same self-test linked above). We send an IPI to an idle CPU and in the
handler compute the time difference between when the IPI was sent and
when the handler ran. We see the following numbers on POWER9 LPAR.

CEDE latency measured using an IPI (numbers in us)
==================================================
N       Min      Median   Avg       90%ile  99%ile    Max    Stddev
400     711      7564     7369.43   8559    9514      9698   1200.01

Thus considering the avg and the 99th percentile this compensation
would be 5.4-7.5us.

Suppose, we consider the compensation corresponding to the 99th
percentile latency value measured using the IPI, the compensation will
be 7.5us, which will take the total CEDE latency to 9.5us.

This is in the ballpark of the default value of 10us which we obtain
if we do

if (!p10)
   use default hardcoded value;

That is either will add one branch for the affected platform.

Since POWER10 onwards, the latency value advertized by the hypervisor
will be the latency as observed by the LPAR VCPU, any new code that we
will be adding will only be applicable for POWER9. We can get the same
effect by using the default value.

Given this, if you feel that it might still be worth pursuing the
compensation approach, I will send out a patch for that.

But I understand if you do not have confidence that the compensation is
the same in all cases and do not have the opportunity to measure it it
may be simpler to apply one very conservative adjustment.

Thanks

Michal

--
Thanks and Regards
gautham.