Message-ID: <CAJZ5v0gPuo1pwDdYjSKS5zuwCk2jN0uPccFhu-nKnpiK1fxtBw@mail.gmail.com>
Date:   Mon, 29 Mar 2021 14:44:49 +0200
From:   "Rafael J. Wysocki" <rafael@...nel.org>
To:     "Zhou Ti (x2019cwm)" <x2019cwm@...x.ca>
Cc:     "Rafael J. Wysocki" <rafael@...nel.org>,
        "Rafael J. Wysocki" <rjw@...ysocki.net>,
        Frederic Weisbecker <frederic@...nel.org>,
        LKML <linux-kernel@...r.kernel.org>,
        "Rafael J. Wysocki" <rafael.j.wysocki@...el.com>,
        Peter Zijlstra <peterz@...radead.org>,
        Thomas Gleixner <tglx@...utronix.de>,
        Yunfeng Ye <yeyunfeng@...wei.com>,
        "Paul E . McKenney" <paulmck@...nel.org>,
        Marcelo Tosatti <mtosatti@...hat.com>,
        Ingo Molnar <mingo@...nel.org>,
        Linux PM <linux-pm@...r.kernel.org>
Subject: Re: 回复: [PATCH 01/10] tick/nohz: Prevent tick_nohz_get_sleep_length() from returning negative value

On Fri, Mar 26, 2021 at 11:53 PM Zhou Ti (x2019cwm) <x2019cwm@...x.ca> wrote:
>
> On Fri, 26 Mar 2021 19:54:26 +0100, Rafael J. Wysocki wrote:
> > On Fri, Mar 26, 2021 at 6:53 PM Zhou Ti (x2019cwm) <x2019cwm@...x.ca> wrote:
> > >
> > > On Fri, 26 Mar 2021 18:01:47 +0100, Rafael J. Wysocki wrote:
> > > > On Thu, Mar 25, 2021 at 9:37 PM Zhou Ti (x2019cwm) <x2019cwm@...x.ca> wrote:
> > > > >
> > > > > On March 25, 2021 15:50, Rafael J. Wysocki wrote:
> > > > > > On Thu, Mar 25, 2021 at 8:18 PM Zhou Ti (x2019cwm) <x2019cwm@...x.ca> wrote:
> > > > > > >
> > > > > > > On March 25, 2021 14:56, Rafael J. Wysocki wrote:
> > > > > > > > On Thursday, March 25, 2021 2:14:00 PM CET Frederic Weisbecker wrote:
> > > > > > > > > On Tue, Mar 16, 2021 at 04:08:08PM +0000, Zhou Ti (x2019cwm) wrote:
> > > > > > > > > > But I don't think it's a good idea to handle this in callers, because logically the function shouldn't return negative values. Returning 0 directly would allow idle governors to get another chance to select again.
> > > > > > > > >
> > > > > > > > > Hmm, I'm going to leave the last word to Rafael since cpuidle are the only
> > > > > > > > > callers of this. In any case we need to fix it.
> > > > > > > >
> > > > > > > > Yes, we do.
> > > > > > > >
> > > > > > > > So I said that I preferred to address this in the callers and the reason why
> > > > > > > > is because, for example, for the teo governor it would be a matter of using
> > > > > > > > a different data type to store the tick_nohz_get_sleep_length() return value,
> > > > > > > > like in the (untested) patch below.
> > > > > > > >
> > > > > > > > So at least in this case there is no need to add any new branches anywhere.
> > > > > > > >
> > > > > > > > I'm still not sure about menu, because it is more complicated, but even if
> > > > > > > > that one needs an extra branch, that is a win already.
> > > > > > >
> > > > > > > I would like to point out the potential trouble that fixing this issue in the
> > > > > > > callers could cause.
> > > > > > >
> > > > > > > 1. This function is called multiple times in menu governor and TEO
> > > > > > > governor.
> > > > > >
> > > > > > What do you mean by "multiple times"?
> > > > > >
> > > > > > Each of the governors calls it once per cycle and its previous return
> > > > > > value is not used in the next cycle at least in teo.
> > > > >
> > > > > I remember a governor called this function twice in a cycle, I guess I remember
> > > > > wrong.
> > > >
> > > > That obviously depends on the governor, but both teo and menu call it
> > > > once per cycle.
> > > >
> > > > > > > I'm not sure that receiving results using signed integers is enough
> > > > > > > to solve all the problems, in the worst case it may require increasing
> > > > > > > the logical complexity of the code.
> > > > > >
> > > > > > That is a valid concern, so it is a tradeoff between increasing the
> > > > > > logical complexity of the code and adding branches to it.
> > > > > >
> > > > > > > 2. This function is important for developing idle governor.
> > > > > > > If the problem is not fixed in the function itself, then this potential
> > > > > > > pitfall should be explicitly stated in the documentation.
> > > > > >
> > > > > > That I can agree with.
> > > > > >
> > > > > > > This is because
> > > > > > > it is difficult to predict from the definition and naming of the function
> > > > > > > that it might return a negative number. I actually discovered this anomaly
> > > > > > > when I was doing data analysis on my own idle governor. For some idle control
> > > > > > > algorithms, this exception return could lead to serious consequences,
> > > > > > > because negative return logically won't happen.
> > > > > >
> > > > > > Well, it's a matter of how to take the possible negative return value
> > > > > > into account so it does not affect the result of the computations.
> > > > >
> > > > > I think it is challenging for some algorithms to take negative return values
> > > > > into account properly. For TEO (and even menu), it is possible to
> > > > > solve the problem by just changing the way the data is received is because the
> > > > > learning mechanism for both algorithms is simple.
> > > >
> > > > Of course this depends on the governor.
> > > >
> > > > > One of the interesting things about the CPUIdle subsystem is that it is well
> > > > > suited to introduce machine learning and probabilistic statistical methods.
> > > >
> > > > You need to remember that the governor code runs in the idle loop
> > > > context which is expected to be reasonably fast.
> > > >
> > > > That's why we are worrying about individual branches here.
> > > >
> > > > > This means that many of the more complex and data-sensitive algorithms can
> > > > > potentially be explored. In the best case we will still need to add additional
> > > > > code complexity to a new algorithm.
> > > >
> > > > So I'm not sure what the problem with adding an upfront negative value
> > > > check to the governor is.
> > > >
> > > > > It would reduce a lot of unnecessary considerations (for example, highlight
> > > > > this shortcoming in the documentation) if we could ensure that this function
> > > > > would work as it is logically defined. But I don't really understand
> > > > > how much of a burden adding an extra branch would impose, so I don't know if
> > > > > this tradeoff is worth it.
> > > >
> > > > It ultimately depends on the governor, which is why I think that the
> > > > negative value check should be done by the governor, if needed, and
> > > > not by the function called by it, because in the latter case the check
> > > > may be redundant and we end up with an extra branch (or two branches
> > > > in this particular case) for no good reason whatsoever.
> > > >
> > > > Yes, there are governors which simply can do the negative value check
> > > > upfront right after calling that function and ensure that they will
> > > > not deal with negative values going forward.  This is probably what
> > > > I'll do in the menu case.
> > > >
> > > > However, if the governor is simple enough and it can avoid doing the
> > > > explicit negative value check, I don't see a reason to do that check
> > > > elsewhere "just in case".
> > >
> > > Makes sense. I will submit my patch to fix this issue in menu and TEO.
> >
> > Well, I have patches for that already and they are not
> > super-straightforward.  Though If you want to try to fix this
> > yourself, I'll wait for your submission.
>
> Thanks! I really like this subsystem, so I hope to contribute a little.
> I still have some questions.
>
> For TEO governor:
>     Even if we change some datatypes as your patch did before, some explicit
>     type conversions still need to be added to prevent wrong results.

That's true.  That patch was way incomplete and I would start with
changing the data type of exit_latency_ns and target_residency_ns in
struct cpuidle_state to s64 for this reason.

>     For example:
>         line 276 (because 1u > -1 will be false)
>         line 329
>         line 422
>
>     Since some of them are in a loop, does the overhead caused by the
>     type conversions worth it? or do I need to do some pre-processing to avoid
>     duplicate conversions (which may cause additional space overhead) ?

Well, it looks like it may be better if I send my patches after all,
because it will take less time overall than me explaining here what I
would do.  Let me do that.

> For menu governor:
>     If we simply change the datatypes, the conversions required are even more.

I would add a negative return value check to menu as I said before.

>     For example:
>         lines 133-142
>         line 202
>         line 289
>         lines 302-309
>         line 320
>         lines 327-328
>         line 347
>
>     If we just do exception handling on next_timer_ns and delta_next,
>     that still requires two additional branches.

Not really AFAICS, but again let me send a patch to illustrate my point.

Powered by blists - more mailing lists