Date:   Tue, 20 Sep 2022 20:42:36 +0800
From:   Yafang Shao <laoar.shao@...il.com>
To:     Roman Gushchin <roman.gushchin@...ux.dev>
Cc:     Tejun Heo <tj@...nel.org>, Alexei Starovoitov <ast@...nel.org>,
        Daniel Borkmann <daniel@...earbox.net>,
        Andrii Nakryiko <andrii@...nel.org>, Martin Lau <kafai@...com>,
        Song Liu <songliubraving@...com>, Yonghong Song <yhs@...com>,
        john fastabend <john.fastabend@...il.com>,
        KP Singh <kpsingh@...nel.org>,
        Stanislav Fomichev <sdf@...gle.com>,
        Hao Luo <haoluo@...gle.com>, Jiri Olsa <jolsa@...nel.org>,
        Johannes Weiner <hannes@...xchg.org>,
        Michal Hocko <mhocko@...nel.org>,
        Shakeel Butt <shakeelb@...gle.com>,
        Muchun Song <songmuchun@...edance.com>,
        Andrew Morton <akpm@...ux-foundation.org>,
        Zefan Li <lizefan.x@...edance.com>,
        Cgroups <cgroups@...r.kernel.org>,
        netdev <netdev@...r.kernel.org>, bpf <bpf@...r.kernel.org>,
        Linux MM <linux-mm@...ck.org>
Subject: Re: [PATCH bpf-next v3 00/13] bpf: Introduce selectable memcg for bpf map

On Tue, Sep 20, 2022 at 10:40 AM Roman Gushchin
<roman.gushchin@...ux.dev> wrote:
>
> On Sun, Sep 18, 2022 at 11:44:48AM +0800, Yafang Shao wrote:
> > On Sat, Sep 17, 2022 at 12:53 AM Roman Gushchin
> > <roman.gushchin@...ux.dev> wrote:
> > >
> > > On Tue, Sep 13, 2022 at 02:15:20PM +0800, Yafang Shao wrote:
> > > > On Fri, Sep 9, 2022 at 12:13 AM Roman Gushchin <roman.gushchin@...ux.dev> wrote:
> > > > >
> > > > > On Thu, Sep 08, 2022 at 10:37:02AM +0800, Yafang Shao wrote:
> > > > > > On Thu, Sep 8, 2022 at 6:29 AM Roman Gushchin <roman.gushchin@...ux.dev> wrote:
> > > > > > >
> > > > > > > On Wed, Sep 07, 2022 at 05:43:31AM -1000, Tejun Heo wrote:
> > > > > > > > Hello,
> > > > > > > >
> > > > > > > > On Fri, Sep 02, 2022 at 02:29:50AM +0000, Yafang Shao wrote:
> > > > > > > > ...
> > > > > > > > > This patchset tries to resolve the above two issues by introducing a
> > > > > > > > > selectable memcg to limit the bpf memory. Currently we only allow to
> > > > > > > > > select its ancestor to avoid breaking the memcg hierarchy further.
> > > > > > > > > Possible use cases of the selectable memcg as follows,
> > > > > > > >
> > > > > > > > As discussed in the following thread, there are clear downsides to an
> > > > > > > > interface which requires the users to specify the cgroups directly.
> > > > > > > >
> > > > > > > >  https://lkml.kernel.org/r/YwNold0GMOappUxc@slm.duckdns.org
> > > > > > > >
> > > > > > > > So, I don't really think this is an interface we wanna go for. I was hoping
> > > > > > > > to hear more from memcg folks in the above thread. Maybe ping them in that
> > > > > > > > thread and continue there?
> > > > > > >
> > > > > >
> > > > > > Hi Roman,
> > > > > >
> > > > > > > As I said previously, I don't like it, because it's an attempt to solve a non
> > > > > > > bpf-specific problem in a bpf-specific way.
> > > > > > >
> > > > > >
> > > > > > Why do you still insist that bpf_map->memcg is not a bpf-specific
> > > > > > issue after so many discussions?
> > > > > > Do you charge the bpf-map's memory the same way as you charge the page
> > > > > > caches or slabs ?
> > > > > > No, you don't. You charge it in a bpf-specific way.
> > > > >
> > > >
> > > > Hi Roman,
> > > >
> > > > Sorry for the late response.
> > > > I've been on vacation in the past few days.
> > > >
> > > > > The only difference is that we charge the cgroup of the processes who
> > > > > created a map, not a process who is doing a specific allocation.
> > > >
> > > > This means the bpf-map can be indepent of process, IOW, the memcg of
> > > > bpf-map can be indepent of the memcg of the processes.
> > > > This is the fundamental difference between bpf-map and page caches, then...
> > > >
> > > > > Your patchset doesn't change this.
> > > >
> > > > We can make this behavior reasonable by introducing an independent
> > > > memcg, as what I did in the previous version.
> > > >
> > > > > There are pros and cons with this approach, we've discussed it back
> > > > > to the times when bpf memcg accounting was developed. If you want
> > > > > to revisit this, it's maybe possible (given there is a really strong and likely
> > > > > new motivation appears), but I haven't seen any complaints yet except from you.
> > > > >
> > > >
> > > > memcg-base bpf accounting is a new feature, which may not be used widely.
> > > >
> > > > > >
> > > > > > > Yes, memory cgroups are not great for accounting of shared resources, it's well
> > > > > > > known. This patchset looks like an attempt to "fix" it specifically for bpf maps
> > > > > > > in a particular cgroup setup. Honestly, I don't think it's worth the added
> > > > > > > complexity. Especially because a similar behaviour can be achieved simple
> > > > > > > by placing the task which creates the map into the desired cgroup.
> > > > > >
> > > > > > Are you serious ?
> > > > > > Have you ever read the cgroup doc? Which clearly describe the "No
> > > > > > Internal Process Constraint".[1]
> > > > > > Obviously you can't place the task in the desired cgroup, i.e. the parent memcg.
> > > > >
> > > > > But you can place it into another leaf cgroup. You can delete this leaf cgroup
> > > > > and your memcg will get reparented. You can attach this process and create
> > > > > a bpf map to the parent cgroup before it gets child cgroups.
> > > >
> > > > If the process doesn't exit after it created bpf-map, we have to
> > > > migrate it around memcgs....
> > > > The complexity in deployment can introduce unexpected issues easily.
> > > >
> > > > > You can revisit the idea of shared bpf maps and outlive specific cgroups.
> > > > > Lof of options.
> > > > >
> > > > > >
> > > > > > [1] https://www.kernel.org/doc/Documentation/cgroup-v2.txt
> > > > > >
> > > > > > > Beatiful? Not. Neither is the proposed solution.
> > > > > > >
> > > > > >
> > > > > > Is it really hard to admit a fault?
> > > > >
> > > > > Yafang, you posted several versions and so far I haven't seen much of support
> > > > > or excitement from anyone (please, fix me if I'm wrong). It's not like I'm
> > > > > nacking a patchset with many acks, reviews and supporters.
> > > > >
> > > > > Still think you're solving an important problem in a reasonable way?
> > > > > It seems like not many are convinced yet. I'd recommend to focus on this instead
> > > > > of blaming me.
> > > > >
> > > >
> > > > The best way so far is to introduce specific memcg for specific resources.
> > > > Because not only the process owns its memcg, but also specific
> > > > resources own their memcgs, for example bpf-map, or socket.
> > > >
> > > > struct bpf_map {                                 <<<< memcg owner
> > > >     struct memcg_cgroup *memcg;
> > > > };
> > > >
> > > > struct sock {                                       <<<< memcg owner
> > > >     struct mem_cgroup *sk_memcg;
> > > > };
> > > >
> > > > These resources already have their own memcgs, so we should make this
> > > > behavior formal.
> > > >
> > > > The selectable memcg is just a variant of 'echo ${proc} > cgroup.procs'.
> > >
> > > This is a fundamental change: cgroups were always hierarchical groups
> > > of processes/threads. You're basically suggesting to extend it to
> > > hierarchical groups of processes and some other objects (what's a good
> > > definition?).
> >
> > Kind of, but not exactly.
> > We can do it without breaking the cgroup hierarchy. Under current
> > cgroup hierarchy, the user can only echo processes/threads into a
> > cgroup, that won't be changed in the future. The specific resources
> > are not exposed to the user, the user can only control these specific
> > resources by controlling their associated processes/threads.
> > For example,
> >
> >                 Memcg-A
> >                        |---- Memcg-A1
> >                        |---- Memcg-A2
> >
> > We can introduce a new file memory.owner into each memcg. Each bit of
> > memory.owner represents a specific resources,
> >
> >  memory.owner: | bit31 | bitN | ... | bit1 | bit0 |
> >                                          |               |
> > |------ bit0: bpf memory
> >                                          |
> > |-------------- bit1: socket memory
> >                                          |
> >                                          |---------------------------
> > bitN: a specific resource
> >
> > There won't be too many specific resources which have to own their
> > memcgs, so I think 32bits is enough.
> >
> >                 Memcg-A : memory.owner == 0x1
> >                        |---- Memcg-A1 : memory.owner == 0
> >                        |---- Memcg-A2 : memory.owner == 0x1
> >
> > Then the bpf created by processes in Memcg-A1 will be charged into
> > Memcg-A directly without charging into Memcg-A1.
> > But the bpf created by processes in Memcg-A2 will be charged into
> > Memcg-A2 as its memory.owner is 0x1.
> > That said, these specific resources are not fully independent of
> > process, while they are still associated with the processes which
> > create them.
> > Luckily memory.move_charge_at_immigrate is disabled in cgroup2, so we
> > don't need to care about the possible migration issue.
> >
> > I think we may also apply it to shared page caches.  For example,
> >       struct inode {
> >           struct mem_cgroup *memcg;          <<<< add a new member
> >       };
> >
> > We define struct inode as a memcg owner, and use scope-based charge to
> > charge its pages into inode->memcg.
> > And then put all memcgs which shared these resources under the same
> > parent. The page caches of this inode will be charged into the parent
> > directly.
>
> Ok, so it's something like premature selective reparenting.
>

Right. I think it  may be a good way to handle the resources which may
outlive the process.

> > The shared page cache is more complicated than bpf memory, so I'm not
> > quite sure if it can apply to shared page cache, but it can work well
> > for bpf memory.
>
> Yeah, this is the problem. It feels like it's a problem very specific
> to bpf maps and an exact way you use them. I don't think you can successfully
> advocate for changes of these calibre without a more generic problem. I might
> be wrong.
>

What is your concern about this method? Are there any potential issues?

> >
> > Regarding the observability, we can introduce a specific item into
> > memory.stat for this specific memory. For example a new item 'bpf' for
> > bpf memory.
> > That can be accounted/unaccounted for in the same way as we do in
> > set_active_memcg(). for example,
> >
> >     struct task_struct {
> >         struct mem_cgroup  *active_memcg;
> >         int                             active_memcg_item;   <<<<
> > introduce a new member
> >     };
> >
> >     bpf_memory_alloc()
> >     {
> >              old_memcg = set_active_memcg(memcg);
> >              old_item = set_active_memcg_item(MEMCG_BPF);
>
> I thought about something like this but for a different purpose:
> to track the amount of memory consumed by bpf.
>

Right, we can use it to track bpf memory consumption.

> >              alloc();
> >              set_active_memcg_item(old_item);
> >              set_active_memcg(old_memcg);
> >     }
> >
> >     bpf_memory_free()
> >     {
> >              old = set_active_memcg_item(MEMCG_BPF);
> >              free();
> >              set_active_memcg_item(old);
> >     }
>
> But the problem is that we shoud very carefully mark all allocations and
> releases, which is very error-prone. Interfaces which don't require annotating
> releases are generally better, but require additional memory.
>

If we don't annotate the releases, we have to add something into the
struct page, which may not be worth it.
It is clear how the bpf memory is allocated and freed, so I think we
can start it with bpf memory.
If in the future we can figure out a lightweight way to avoid
annotating the releases, then we can remove the annotations in the bpf
memory releases.

-- 
Regards
Yafang

Powered by blists - more mailing lists