On Tue, Oct 31, 2023 at 7:41 PM Chuyi Zhou <zhouchuyi@xxxxxxxxxxxxx> wrote:
>
> Hello,
>
> 在 2023/11/1 06:06, Alexei Starovoitov 写道:
> > On Tue, Oct 31, 2023 at 4:38 AM Chuyi Zhou <zhouchuyi@xxxxxxxxxxxxx> wrote:
> >>
> >>
> >> So, maybe another possible solution is:
> >>
> >> diff --git a/kernel/bpf/cgroup_iter.c b/kernel/bpf/cgroup_iter.c
> >> index 209e5135f9fb..72a6778e3fba 100644
> >> --- a/kernel/bpf/cgroup_iter.c
> >> +++ b/kernel/bpf/cgroup_iter.c
> >> @@ -282,7 +282,7 @@ static struct bpf_iter_reg bpf_cgroup_reg_info = {
> >> .ctx_arg_info_size = 1,
> >> .ctx_arg_info = {
> >> { offsetof(struct bpf_iter__cgroup, cgroup),
> >> - PTR_TO_BTF_ID_OR_NULL },
> >> + PTR_TO_BTF_ID_OR_NULL | MEM_RCU },
> >> },
> >> .seq_info = &cgroup_iter_seq_info,
> >> };
> >> diff --git a/kernel/bpf/task_iter.c b/kernel/bpf/task_iter.c
> >> index 59e747938bdb..4fd3f734dffd 100644
> >> --- a/kernel/bpf/task_iter.c
> >> +++ b/kernel/bpf/task_iter.c
> >> @@ -706,7 +706,7 @@ static struct bpf_iter_reg task_reg_info = {
> >> .ctx_arg_info_size = 1,
> >> .ctx_arg_info = {
> >> { offsetof(struct bpf_iter__task, task),
> >> - PTR_TO_BTF_ID_OR_NULL },
> >> + PTR_TO_BTF_ID_OR_NULL | PTR_TRUSTED },
> >
> > Yep. That looks good.
> > bpf_cgroup_reg_info -> cgroup is probably PTR_TRUSTED too.
> > Not sure... why did you go with MEM_RCU there ?
>
> hmm...
>
> That is because in our previous discussion, you suggested we'd better
> add BTF_TYPE_SAFE_RCU_OR_NULL(struct bpf_iter__cgroup) {...}
I mentioned that because we don't have
BTF_TYPE_SAFE_TRUSTED_OR_NULL.
and cgroup pointer can be NULL, but since you found a cleaner way
we can do PTR_TO_BTF_ID_OR_NULL | PTR_TRUSTED.
> I didn't think too much about it. But I noticed that we only use
> cgroup_mutex to protect the iteration in cgroup_iter.c.
>
> Looking at cgroup_kn_lock_live() in kernel/cgroup/cgroup.c, it would use
> cgroup_tryget()/cgroup_is_dead() to check whether the cgrp is 'dead'.
> cgroup_tryget() seems is equal to bpf_cgroup_acquire(). So, maybe let's
> return a 'rcu' cgrp pointer. If BPF Prog want stronger guarantee of
> cgrp, just use bpf_cgroup_acquire().
and that would be misleading.
MEM_RCU means that the pointer is valid, but it could have refcnt == 0,
while PTR_TRUSTED means that it's good to use as-is.
Here cgroup pointer is trusted. It's not a dead cgroup.
See kernel/bpf/cgroup_iter.c __cgroup_iter_seq_show().
bpf prog doesn't need to call bpf_cgroup_acquire.
