Re: cgroup pointed by sock is leaked on mode switch

[Zefan Li <lizefan@huawei.com>]

On 2020/5/6 9:50, Yang Yingliang wrotee:
> +cc lizefan@huawei.com
> 
> On 2020/5/6 0:06, Tejun Heo wrote:
>> Hello, Yang.
>>
>> On Sat, May 02, 2020 at 06:27:21PM +0800, Yang Yingliang wrote:
>>> I find the number nr_dying_descendants is increasing:
>>> linux-dVpNUK:~ # find /sys/fs/cgroup/ -name cgroup.stat -exec grep
>>> '^nr_dying_descendants [^0]'  {} +
>>> /sys/fs/cgroup/unified/cgroup.stat:nr_dying_descendants 80
>>> /sys/fs/cgroup/unified/system.slice/cgroup.stat:nr_dying_descendants 1
>>> /sys/fs/cgroup/unified/system.slice/system-hostos.slice/cgroup.stat:nr_dying_descendants
>>> 1
>>> /sys/fs/cgroup/unified/lxc/cgroup.stat:nr_dying_descendants 79
>>> /sys/fs/cgroup/unified/lxc/5f1fdb8c54fa40c3e599613dab6e4815058b76ebada8a27bc1fe80c0d4801764/cgroup.stat:nr_dying_descendants
>>> 78
>>> /sys/fs/cgroup/unified/lxc/5f1fdb8c54fa40c3e599613dab6e4815058b76ebada8a27bc1fe80c0d4801764/system.slice/cgroup.stat:nr_dying_descendants
>>> 78
>> Those numbers are nowhere close to causing oom issues. There are some
>> aspects of page and other cache draining which is being improved but unless
>> you're seeing numbers multiple orders of magnitude higher, this isn't the
>> source of your problem.
>>
>>> The situation is as same as the commit bd1060a1d671 ("sock, cgroup: add
>>> sock->sk_cgroup") describes.
>>> "On mode switch, cgroup references which are already being pointed to by
>>> socks may be leaked."
>> I'm doubtful that you're hitting that issue. Mode switching means memcg
>> being switched between cgroup1 and cgroup2 hierarchies, which is unlikely to
>> be what's happening when you're launching docker containers.
>>
>> The first step would be identifying where memory is going and finding out
>> whether memcg is actually being switched between cgroup1 and 2 - look at the
>> hierarchy number in /proc/cgroups, if that's switching between 0 and
>> someting not zero, it is switching.
>>

I think there's a bug here which can lead to unlimited memory leak.
This should reproduce the bug:

    # mount -t cgroup -o netprio xxx /cgroup/netprio
    # mkdir /cgroup/netprio/xxx
    # echo PID > /cgroup/netprio/xxx/tasks
    /* this PID process starts to do some network thing and then exits */
    # rmdir /cgroup/netprio/xxx
    /* now this cgroup will never be freed */

Look at the code:

static inline void sock_update_netprioidx(struct sock_cgroup_data *skcd)
{
	...
	sock_cgroup_set_prioidx(skcd, task_netprioidx(current));
}

static inline void sock_cgroup_set_prioidx(struct sock_cgroup_data *skcd,
					u16 prioidx)
{
	...
	if (sock_cgroup_prioidx(&skcd_buf) == prioidx)
		return ;
	...
	skcd_buf.prioidx = prioidx;
	WRITE_ONCE(skcd->val, skcd_buf.val);
}

task_netprioidx() will be the cgrp id of xxx which is not 1, but
sock_cgroup_prioidx(&skcd_buf) is 1 because it thought it's in v2 mode.
Now we have a memory leak.

I think the eastest fix is to do the mode switch here:

diff --git a/net/core/netprio_cgroup.c b/net/core/netprio_cgroup.c
index b905747..2397866 100644
--- a/net/core/netprio_cgroup.c
+++ b/net/core/netprio_cgroup.c
@@ -240,6 +240,8 @@ static void net_prio_attach(struct cgroup_taskset *tset)
         struct task_struct *p;
         struct cgroup_subsys_state *css;

+       cgroup_sk_alloc_disable();
+
         cgroup_taskset_for_each(p, css, tset) {
                 void *v = (void *)(unsigned long)css->cgroup->id;