On Tue, Sep 12, 2023 at 4:24 AM Tejun Heo <tj@xxxxxxxxxx> wrote:
>
> On Sun, Sep 10, 2023 at 11:17:48AM +0800, Yafang Shao wrote:
> > To acquire the cgroup_id, we can resort to open coding, as exemplified below:
> >
> > task = bpf_get_current_task_btf();
> > cgroups = task->cgroups;
> > cgroup = cgroups->subsys[cpu_cgrp_id]->cgroup;
> > key = cgroup->kn->id;
>
> You can't hardcode it to a specific controller tree like that. You either
> stick with fd based interface or need also add something to identify the
> specifc cgroup1 tree.
As pointed out by Alexei, I think we can introduce some
cgroup_id-based kfuncs which can work for both cgroup1 and cgroup2.
Something as follows (untested),
__bpf_kfunc u64 bpf_current_cgroup_id_from_subsys(int subsys)
{
struct cgroup *cgroup;
cgroup = task_cgroup(current, subsys);
if (!cgroup)
return 0;
return cgroup_id(cgroup);
}
__bpf_kfunc struct cgroup *bpf_cgroup_from_subsys_id(u64 cgid, int subsys)
{
struct cgroup_subsys_state *css = init_css_set.subsys[subsys];
struct cgroup *subsys_root = css->cgroup;
// We should introduce a new helper cgroup_get_from_subsys_id()
// in the cgroup subsystem.
return cgroup_get_from_subsys_id(subsys_root, cgid);
}
And change task_under_cgroup_hierarchy() as follows,
static inline bool task_under_cgroup_hierarchy(struct task_struct *task,
struct cgroup *ancestor)
{
struct css_set *cset = task_css_set(task);
-
- return cgroup_is_descendant(cset->dfl_cgrp, ancestor);
+ struct cgroup *cgrp;
+ bool ret = false;
+
+ if (ancestor->root == &cgrp_dfl_root)
+ return cgroup_is_descendant(cset->dfl_cgrp, ancestor);
+
+ cgroup_lock();
+ spin_lock_irq(&css_set_lock);
+ cgrp = task_cgroup_from_root(task, ancestor->root);
+ if (cgrp && cgroup_is_descendant(cgrp, ancestor))
+ ret = true;
+ spin_unlock_irq(&css_set_lock);
+ cgroup_unlock();
+ return ret;
}
With the above changes, I think it can meet most use cases with BPF on cgroup1.
What do you think ?
--
Regards
Yafang
