[Devel] [PATCH 2/2 v3] ve/cgroup: Added pointers to owning ve to root cgroups

[Kirill Tkhai]

On 24.03.2020 13:46, Valeriy Vdovin wrote:
> Follow-up patch to per-cgroup release_agent property. release_agent
> notifications are spawned from a special kthread, running under ve0. But
> newly spawned tasks should run under their own ve context. Easy way to
> pass this information to a spawning thread is by adding 've_owner' field
> to a root cgroup. At notification any cgroup can be walked upwards to
> it's root and get ve_owner from there. ve_owner is initialized at container
> start, that's when we know all cgroup ve roots for free. For ve0 roots
> we don't need ve_owner field set, because we can always get pointer to ve0.
> Only place where we can detch ve_owner knowledge from root ve cgroups is
> at cgroup_exit for ve->init_task, because that's that's the last time, when
> ve->init_task is related to it's ve root cgroups (see cgroup_exit comments).
> 
> https://jira.sw.ru/browse/PSBM-83887
> Signed-off-by: Valeriy Vdovin <valeriy.vdovin at virtuozzo.com>
> ---
>  include/linux/cgroup.h |  3 +++
>  kernel/cgroup.c        | 45 +++++++++++++++++++++++++++++++++++++++++++++
>  2 files changed, 48 insertions(+)
> 
> diff --git a/include/linux/cgroup.h b/include/linux/cgroup.h
> index 5e289fc..34fc017 100644
> --- a/include/linux/cgroup.h
> +++ b/include/linux/cgroup.h
> @@ -291,6 +291,9 @@ struct cgroup {
>  	struct simple_xattrs xattrs;
>  	u64 subgroups_limit;
>  
> +	/* ve_owner, responsible for running release agent. */
> +	struct ve_struct *ve_owner;
> +
>  	/*
>  	 * Per-cgroup path to release agent binary for release
>  	 * notifications.
> diff --git a/kernel/cgroup.c b/kernel/cgroup.c
> index d1d4605..2500095 100644
> --- a/kernel/cgroup.c
> +++ b/kernel/cgroup.c
> @@ -4352,6 +4352,7 @@ int cgroup_mark_ve_root(struct ve_struct *ve)
>  	mutex_lock(&cgroup_mutex);
>  	for_each_active_root(root) {
>  		cgrp = task_cgroup_from_root(ve->init_task, root);
> +		cgrp->ve_owner = ve;
>  		set_bit(CGRP_VE_ROOT, &cgrp->flags);
>  		err = cgroup_add_file_on_mark_ve(cgrp);
>  		if (err)
> @@ -4363,6 +4364,20 @@ int cgroup_mark_ve_root(struct ve_struct *ve)
>  	return err;
>  }
>  
> +static void cgroup_unbind_roots_from_ve(struct ve_struct *ve)
> +{
> +	struct cgroup *cgrp;
> +	struct cgroupfs_root *root;
> +
> +	mutex_lock(&cgroup_mutex);
> +	for_each_active_root(root) {
> +		cgrp = task_cgroup_from_root(ve->init_task, root);
> +		BUG_ON(!cgrp->ve_owner);
> +		cgrp->ve_owner = NULL;
> +	}
> +	mutex_unlock(&cgroup_mutex);
> +}
> +
>  struct cgroup *cgroup_get_ve_root(struct cgroup *cgrp)
>  {
>  	struct cgroup *ve_root = NULL;
> @@ -5401,6 +5416,14 @@ void cgroup_exit(struct task_struct *tsk, int run_callbacks)
>  	int i;
>  
>  	/*
> +	 * Detect that the task in question is ve->init_task
> +	 * if so, unbind all cgroups that want to be released under
> +	 * this ve.
> +	 */
> +	if (!ve_is_super(tsk->task_ve) && tsk == tsk->task_ve->init_task)
> +		cgroup_unbind_roots_from_ve(tsk->task_ve);
> +
> +	/*
>  	 * Unlink from the css_set task list if necessary.
>  	 * Optimistically check cg_list before taking
>  	 * css_set_lock
> @@ -5493,6 +5516,7 @@ static void cgroup_release_agent(struct work_struct *work)
>  	raw_spin_lock(&release_list_lock);
>  	while (!list_empty(&release_list)) {
>  		char *argv[3], *envp[3];
> +		struct ve_struct *ve = NULL;
>  		int i, err;
>  		const char *release_agent;
>  		char *pathbuf = NULL, *agentbuf = NULL;
> @@ -5507,8 +5531,24 @@ static void cgroup_release_agent(struct work_struct *work)
>  			goto continue_free;
>  		if (__cgroup_path(cgrp, pathbuf, PAGE_SIZE, true) < 0)
>  			goto continue_free;
> +
> +		/*
> +		 * Deduce under which VE we are going to spawn nofitier
> +		 * binary. Non-root VE has it's VE-local root cgroup,
> +		 * marked with VE_ROOT flag. It has non-NULL ve_owner
> +		 * set during cgroup_mark_ve_root.
> +		 * VE0 root cgroup does not need ve_owner field, because
> +		 * it's ve is ve0. Non-VE root cgroup does not have a parent.
> +		 */
>  		root_cgrp = cgroup_get_local_root(cgrp);
> +		if (root_cgrp->parent == NULL)
> +			ve = &ve0;
> +		else
> +			ve = root_cgrp->ve_owner;
> +		if (!ve)
> +			goto continue_free;
>  		rcu_read_lock();
> +
>  		release_agent = cgroup_release_agent_path(root_cgrp);
>  		if (release_agent)
>  			agentbuf = kstrdup(release_agent, GFP_KERNEL);
> @@ -5531,7 +5571,12 @@ static void cgroup_release_agent(struct work_struct *work)
>  		 * since the exec could involve hitting disk and hence
>  		 * be a slow process */
>  		mutex_unlock(&cgroup_mutex);
> +#ifdef CONFIG_VE
> +		err = call_usermodehelper_fns_ve(ve, argv[0], argv,
> +			envp, UMH_WAIT_EXEC, NULL, NULL, NULL);
> +#else
>  		err = call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
> +#endif
>  		if (err < 0)
>  			pr_warn_ratelimited("cgroup release_agent "
>  					    "%s %s failed: %d\n",

Is there a race like the below?

[cpu 0]                                                    [cpu1]
cgroup_release_agent()
  mutex_lock(&cgroup_mutex)
  ve = root_cgrp->ve_owner
  mutex_unlock(&cgroup_mutex)

	                                                   cgroup_unbind_roots_from_ve(ve)
                                                             cgroup->ve_owner = NULL;
                                                           css_put(ve)  <--- final put


  call_usermodehelper_fns_ve(ve)
    get_ve(ve)   <--- gets zero counter