Messages in this thread | | | Date | Wed, 21 Sep 2011 15:59:55 -0300 | From | Glauber Costa <> | Subject | Re: [PATCH v3 2/7] socket: initial cgroup code. |
| |
On 09/21/2011 03:47 PM, Greg Thelen wrote: > On Sun, Sep 18, 2011 at 5:56 PM, Glauber Costa<glommer@parallels.com> wrote: >> We aim to control the amount of kernel memory pinned at any >> time by tcp sockets. To lay the foundations for this work, >> this patch adds a pointer to the kmem_cgroup to the socket >> structure. >> >> Signed-off-by: Glauber Costa<glommer@parallels.com> >> CC: David S. Miller<davem@davemloft.net> >> CC: Hiroyouki Kamezawa<kamezawa.hiroyu@jp.fujitsu.com> >> CC: Eric W. Biederman<ebiederm@xmission.com> > ... >> +void sock_update_memcg(struct sock *sk) >> +{ >> + /* right now a socket spends its whole life in the same cgroup */ >> + BUG_ON(sk->sk_cgrp); >> + >> + rcu_read_lock(); >> + sk->sk_cgrp = mem_cgroup_from_task(current); >> + >> + /* >> + * We don't need to protect against anything task-related, because >> + * we are basically stuck with the sock pointer that won't change, >> + * even if the task that originated the socket changes cgroups. >> + * >> + * What we do have to guarantee, is that the chain leading us to >> + * the top level won't change under our noses. Incrementing the >> + * reference count via cgroup_exclude_rmdir guarantees that. >> + */ >> + cgroup_exclude_rmdir(mem_cgroup_css(sk->sk_cgrp)); > > This grabs a css_get() reference, which prevents rmdir (will return > -EBUSY). Yes.
How long is this reference held? For the socket lifetime.
> I wonder about the case > where a process creates a socket in memcg M1 and later is moved into > memcg M2. At that point an admin would expect to be able to 'rmdir > M1'. I think this rmdir would return -EBUSY and I suspect it would be > difficult for the admin to understand why the rmdir of M1 failed. It > seems that to rmdir a memcg, an admin would have to kill all processes > that allocated sockets while in M1. Such processes may not still be > in M1. > >> + rcu_read_unlock(); >> +} I agree. But also, don't see too much ways around it without implementing full task migration.
Right now I am working under the assumption that tasks are long lived inside the cgroup. Migration potentially introduces some nasty locking problems in the mem_schedule path.
Also, unless I am missing something, the memcg already has the policy of not carrying charges around, probably because of this very same complexity.
True that at least it won't EBUSY you... But I think this is at least a way to guarantee that the cgroup under our nose won't disappear in the middle of our allocations.
| |