linux-kernel - Re: [PATCHv2 3/3] cpuidle: teo: Don't count non-existent intercepts

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [day] [month] [year] [list]

Date: Fri, 28 Jun 2024 10:33:36 +0100
From: Christian Loehle <christian.loehle@....com>
To: Dietmar Eggemann <dietmar.eggemann@....com>, linux-pm@...r.kernel.org,
 linux-kernel@...r.kernel.org, rafael@...nel.org
Cc: vincent.guittot@...aro.org, qyousef@...alina.io, peterz@...radead.org,
 daniel.lezcano@...aro.org, ulf.hansson@...aro.org, anna-maria@...utronix.de,
 kajetan.puchalski@....com, lukasz.luba@....com
Subject: Re: [PATCHv2 3/3] cpuidle: teo: Don't count non-existent intercepts

On 6/26/24 14:09, Dietmar Eggemann wrote:
> On 11/06/2024 13:24, Christian Loehle wrote:
>> When bailing out early, teo will not query the sleep length anymore
>> since commit 6da8f9ba5a87 ("cpuidle: teo:
>> Skip tick_nohz_get_sleep_length() call in some cases") with an
>> expected sleep_length_ns value of KTIME_MAX.
>> This lead to state0 accumulating lots of 'intercepts' because
>> the actually measured sleep length was < KTIME_MAX, so count KTIME_MAX
>> as a hit (we have to count them as something otherwise we are stuck)
>> and therefore teo taking too long to recover from intercept-heavy
>> scenarios.
>>
>> Fundamentally we can only do one of the two:
>> 1. Skip sleep_length_ns query when we think intercept is likely
> 
> Isn't this what we do right now? Skip tick_nohz_get_sleep_length() for
> state0 and set cpu_data->sleep_length_ns to KTIME_MAX in teo_select()
> and then count this as an 'intercept' in teo_update().

Yes it is, I'll state that more explicitly in the future, it was
kind of implied by "we can only do one of the two" and "This patch
chooses the latter".

> 
>> 2. Have accurate data if sleep_length_ns is actually intercepted when
>> we believe it is currently intercepted.
>>
>> This patch chooses the latter as I've found the additional time it
>> takes to query the sleep length to be negligible and the variants of
>> option 1 (count all unknowns as misses or count all unknown as hits)
>> had significant regressions (as misses had lots of too shallow idle
>> state selections and as hits had terrible performance in
>> intercept-heavy workloads).
> 
> IMHO, you do 2 things here:
> 
> (1) Set 'cpu_data->sleep_length_ns != KTIME_MAX' for '!idx &&
>     prev_intercept_idx' in teo_select().

This is just the technical way of saying "We query the sleep length
even though teo is in an intercept scenario (and don't use the sleep
length for determining the current idle state selection)".

> 
> (2) Force an update with 'cpu_data->sleep_length_ns == KTIME_MAX' to be
>     counted as a 'hit' rather an 'intercept' in teo_update().

I assume this is the one you have an issue with and you're correct.
This isn't quite obvious from the text.
Furthermore the KTIME_MAX won't be set for intercept-bailouts anyway
with (1) and thus doesn't become that important in terms of resulting
behavior. I still think (2) is the semantically correct thing to do,
but I'll drop it for the fixes series and add it to the remaining
'improvements' pile.
Thanks!

> 
> Can't really see how this matches the explanatory text exactly.
> [SNIP]