Received-SPF: pass (google.com: domain of linux-kernel+bounces-154437-linux.lists.archive=gmail.com@vger.kernel.org designates 2604:1380:40f1:3f00::1 as permitted sender) client-ip=2604:1380:40f1:3f00::1;
Precedence: bulk
MIME-Version: 1.0
References: <20240422082238.5784-1-xuewen.yan@unisoc.com> <CAB8ipk-LhSuMsp0DdzjEJN-4XEBT1_ri6BPH_nvxSgFZONMT2Q@mail.gmail.com>
 <20240422094157.GA34453@noisy.programming.kicks-ass.net> <CAB8ipk9M1wLXV7MWMBYGMhKbY71QqwkvchiENeWTg7JERe6kZw@mail.gmail.com>
 <20240422111744.GO30852@noisy.programming.kicks-ass.net> <CAB8ipk8ZBSnJfw9Ow9My-VXG1TU6DVY7mOL_i0Ejdd7GUZCLWA@mail.gmail.com>
 <20240422155937.GP30852@noisy.programming.kicks-ass.net>
In-Reply-To: <20240422155937.GP30852@noisy.programming.kicks-ass.net>
From: Xuewen Yan <xuewen.yan94@gmail.com>
Date: Tue, 23 Apr 2024 11:05:20 +0800
Message-ID: <CAB8ipk-WYVLsRi1_wnm2vF3M4RDnV1cSomb4MkJ83ef2GY0gzg@mail.gmail.com>
Subject: Re: [PATCH v2] sched/eevdf: Prevent vlag from going out of bounds
 when reweight_eevdf
To: Peter Zijlstra <peterz@infradead.org>
Cc: Xuewen Yan <xuewen.yan@unisoc.com>, mingo@redhat.com, juri.lelli@redhat.com, 
	vincent.guittot@linaro.org, dietmar.eggemann@arm.com, rostedt@goodmis.org, 
	bsegall@google.com, mgorman@suse.de, bristot@redhat.com, vschneid@redhat.com, 
	yu.c.chen@intel.com, ke.wang@unisoc.com, di.shen@unisoc.com, 
	linux-kernel@vger.kernel.org
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable

On Mon, Apr 22, 2024 at 11:59=E2=80=AFPM Peter Zijlstra <peterz@infradead.o=
rg> wrote:
>
> On Mon, Apr 22, 2024 at 09:12:12PM +0800, Xuewen Yan wrote:
>
> > By adding a log to observe weight changes in reweight_entity, I found
> > that calc_group_shares() often causes new_weight to become very small:
>
> Yes, cgroups do that. But over-all that should not matter no?
>
> Specifically, the whole re-weight thing turns into a series like:
>
>             w_0   w_1         w_n-1   w_0
>         S =3D --- * --- * ... * ----- =3D ---
>             w_1   w_2          w_n    w_n
>
> Where S is our ultimate scale factor.
>
> So even if w_m (0 < m < n) is 2, it completely disappears. But yes, it
> will create a big term, which is why the initial vlag should be limited.

Okay, I understand what you mean. Even if the weight during dequeue is
very small, the weight will be eliminated during enqueue.
In this case, the necessity of the !on_rq case does not seem to be
very important.

On the other hand, the following case:
place_entity()
{
..
 5244                 load =3D cfs_rq->avg_load;
 5245                 if (curr && curr->on_rq)
 5246                         load +=3D scale_load_down(curr->load.weight);
 5247
 5248                 lag *=3D load + scale_load_down(se->load.weight);
 5249                 if (WARN_ON_ONCE(!load))
 5250                         load =3D 1;
 5251                 lag =3D div_s64(lag, load);<<<<
..
}
reweight_eevdf()
{
..
                 if (avruntime !=3D se->vruntime) {
 3770                 vlag =3D entity_lag(avruntime, se);
 3771                 vlag =3D div_s64(vlag * old_weight, weight); <<<<
 3772                 se->vruntime =3D avruntime - vlag;
 3773         }
....
}

There is no need to clamp the above two positions because these two
calculations will not theoretically cause s64 overflow?

Thanks!

>
> Notably, nice should not exceed 88761*1024 / 2, but I'm not sure I
> remember the limits (if there are any on the cgrou pmuck).
>
> But if roughly 27 bits go to weight, then vlag should not exceed 36,
> which should be well within the slice limit iirc.
>
> Also, as said before, due to integer division being truncating, the
> actual S should be smaller than the expected S due to error
> accumulation.
>
> Anyway, the things to verify are:
>
>  - the S series is complete -- missing terms will mess things up right
>    quick;
>
>  - the limits on both the weight and vlag part, their sum exceeding
>    63bit (plut 1 for sign) will also mess things up.