From mboxrd@z Thu Jan 1 00:00:00 1970 From: Roman Gushchin Subject: Re: [PATCH 1/3] mm: memcontrol: fix memory.low proportional distribution Date: Mon, 16 Dec 2019 19:11:35 +0000 Message-ID: <20191216191131.GB3760@localhost.localdomain> References: <20191213192158.188939-1-hannes@cmpxchg.org> <20191213192158.188939-2-hannes@cmpxchg.org> <20191213204026.GA6830@localhost.localdomain> <20191216182518.GA209920@cmpxchg.org> Mime-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: quoted-printable Return-path: DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=fb.com; h=from : to : cc : subject : date : message-id : references : in-reply-to : content-type : content-id : content-transfer-encoding : mime-version; s=facebook; bh=DoO0l2psZNOhC8HxuixcokzTJJOPI9PvHeEkLcoSny0=; b=oLbpTEHFDy3rPtVaB8bd/l8p/Ewj+8vSrbjs7kHz0BTfiBos2/Xkmo6eIHpeh6v5PZ9L CryZT4Si7bgf9iurVaEJFW4ThZqHFBs5L1fon7QdMz8WewxHppyzRIuMiqRdHTWbabnR pXK3092f1wxOiCY5tIqKv0sRjCLOWMirQI0= DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=fb.onmicrosoft.com; s=selector2-fb-onmicrosoft-com; h=From:Date:Subject:Message-ID:Content-Type:MIME-Version:X-MS-Exchange-SenderADCheck; bh=DoO0l2psZNOhC8HxuixcokzTJJOPI9PvHeEkLcoSny0=; b=HuysLj5lx6wk4Oocgt957LX8/cV0rnDcvkqYndPaZniVSxMHyz25VJ3pM81Tkz7+u/zxwf1Sa1IPM7iIHNcHk4mphhdz6B/WPpYNIYE+4gPqvYz6hxHPQc8wfmVeUeGwjrTlVf+2K2/7U7ZSYb6Bykido6u3KKm6LmSFi2fAvW8= In-Reply-To: <20191216182518.GA209920@cmpxchg.org> Content-Language: en-US Content-ID: <080A4AC9D9F95A44A81CD56E29FE6C02@namprd15.prod.outlook.com> Sender: linux-kernel-owner@vger.kernel.org List-ID: To: Johannes Weiner Cc: Andrew Morton , Michal Hocko , Tejun Heo , "linux-mm@kvack.org" , "cgroups@vger.kernel.org" , "linux-kernel@vger.kernel.org" , Kernel Team On Mon, Dec 16, 2019 at 01:25:18PM -0500, Johannes Weiner wrote: > On Fri, Dec 13, 2019 at 08:40:31PM +0000, Roman Gushchin wrote: > > On Fri, Dec 13, 2019 at 02:21:56PM -0500, Johannes Weiner wrote: > > > When memory.low is overcommitted - i.e. the children claim more > > > protection than their shared ancestor grants them - the allowance is > > > distributed in proportion to each siblings's utilized protection: > > >=20 > > > low_usage =3D min(low, usage) > > > elow =3D parent_elow * (low_usage / siblings_low_usage) > > >=20 > > > However, siblings_low_usage is not the sum of all low_usages. It sums > > > up the usages of *only those cgroups that are within their memory.low= * > > > That means that low_usage can be *bigger* than siblings_low_usage, an= d > > > consequently the total protection afforded to the children can be > > > bigger than what the ancestor grants the subtree. > > >=20 > > > Consider three groups where two are in excess of their protection: > > >=20 > > > A/memory.low =3D 10G > > > A/A1/memory.low =3D 10G, A/memory.current =3D 20G > > > A/A2/memory.low =3D 10G, B/memory.current =3D 20G > > > A/A3/memory.low =3D 10G, C/memory.current =3D 8G > > >=20 > > > siblings_low_usage =3D 8G (only A3 contributes) > > > A1/elow =3D parent_elow(10G) * low_usage(20G) / siblings_low_usage(= 8G) =3D 25G > > >=20 > > > The 25G are then capped to A1's own memory.low setting, i.e. 10G. The > > > same is true for A2. And A3 would also receive 10G. The combined > > > protection of A1, A2 and A3 is 30G, when A limits the tree to 10G. > > >=20 > > > What does this mean in practice? A1 and A2 would still be in excess o= f > > > their 10G allowance and would be reclaimed, whereas A3 would not. As > > > they eventually drop below their protection setting, they would be > > > counted in siblings_low_usage again and the error would right itself. > > >=20 > > > When reclaim is applied in a binary fashion - cgroup is reclaimed whe= n > > > it's above its protection, otherwise it's skipped - this could work > > > actually work out just fine - although it's not quite clear to me why > > > we'd introduce this error in the first place. > >=20 > > This complication is not simple an error, it protects cgroups under > > their low limits if there is unprotected memory. > >=20 > > So, here is an example: > >=20 > > A A/memory.low =3D 2G, A/memory.current =3D 4G > > / \ > > B C B/memory.low =3D 3G B/memory.current =3D 2G > > C/memory.low =3D 1G C/memory.current =3D 2G > >=20 > > as now: > >=20 > > B/elow =3D 2G * 2G / 2G =3D 2G =3D=3D B/memory.current > > C/elow =3D 2G * 1G / 2G =3D 1G < C/memory.current > >=20 > > with this fix: > >=20 > > B/elow =3D 2G * 2G / 3G =3D 4/3 G < B/memory.current > > C/elow =3D 2G * 1G / 3G =3D 2/3 G < C/memory.current > >=20 > > So in other words, currently B won't be scanned at all, because > > there is 1G of unprotected memory in C. With your patch both B and C > > will be scanned. >=20 > Looking at the B and C numbers alone: C is bigger than what it claims > for protection and B is smaller than what it claims for protection. >=20 > However, A doesn't provide 4G to its children. It provides 2G to be > distributed between the two. So how can B claim 3G and be exempted > from reclaim? First, what if the memory pressure comes from memory.high/max set on A? Second, it's up to semantics we define. Looking at it from the other side: there is clearly 1G of memory in C which is not protected no matter what. B wants it's memory to be fully protected, but it's limited by the competit= ion on the parent level. Now we try to satisfy B's requirements until we can. Should we treat B and C equally from scratch? I think both approaches is acceptable, but if we're switching from one opti= on to another, let's make it clear. >=20 > But more importantly, it isn't in either case! The end result is the > same in both implementations. Because as soon as C is reclaimed down > to below 1G, A is still in excess of its memory.low (because it's > overcommitted!), and they both will be reclaimed proportionally. I do not disagree: the introduction of the proportional reclaim made this complication (partially?) obsolete. But originally it was required to make target distribution correct. >=20 > From the example in the current code: >=20 > * For example, if there are memcgs A, A/B, A/C, A/D and A/E: > * > * A A/memory.low =3D 2G, A/memory.current =3D 6G > * //\\ > * BC DE B/memory.low =3D 3G B/memory.current =3D 2G > * C/memory.low =3D 1G C/memory.current =3D 2G > * D/memory.low =3D 0 D/memory.current =3D 2G > * E/memory.low =3D 10G E/memory.current =3D 0 > * > * and the memory pressure is applied, the following memory distribution > * is expected (approximately): > * > * A/memory.current =3D 2G > * > * B/memory.current =3D 1.3G > * C/memory.current =3D 0.6G > * D/memory.current =3D 0 > * E/memory.current =3D 0 >=20 > Even though B starts out within whatever it claims to be its > protection, A is overcommitted and so B and C converge on their > proportional share of the parent's allowance. >=20 > So to go back to the example chosen above: >=20 > > A A/memory.low =3D 2G, A/memory.current =3D 4G > > / \ > > B C B/memory.low =3D 3G B/memory.current =3D 2G > > C/memory.low =3D 1G C/memory.current =3D 2G >=20 > With either implementation we'd expect the distribution to be about > 1.5G and 0.5G for B and C, respectively. >=20 > And they'd have to be, too. Otherwise the semantics would be > completely unpredictable to anyone trying to configure this. >=20 > So I think mixing proportional distribution with absolute thresholds > like this makes the implementation unnecessarily hard to reason > about. It's also clearly buggy as pointed out in the changelog. >=20 > > > However, since > > > 1bc63fb1272b ("mm, memcg: make scan aggression always exclude > > > protection"), reclaim pressure is scaled to how much a cgroup is abov= e > > > its protection. As a result this calculation error unduly skews > > > pressure away from A1 and A2 toward the rest of the system. > >=20 > > It could be that with 1bc63fb1272b the target memory distribution > > will be fine. However the patch will change the memory pressure in B an= d C > > (in the example above). Maybe it's ok, but at least it should be discus= sed > > and documented. >=20 > I'll try to improve the changelog based on this, thanks for filling in > the original motivation. But I do think it's a change we want to make. Absolutely, I'm not against the change. I just want to make sure we will put all details into the changelog. Thanks!