Re: [PATCH bpf-next v2 5/6] bpf: teach the verifier to enforce css_iter and process_iter in RCU CS

[Chuyi Zhou]

Hello.

在 2023/9/12 15:01, Chuyi Zhou 写道:
> css_iter and process_iter should be used in rcu section. Specifically, in
> sleepable progs explicit bpf_rcu_read_lock() is needed before use these
> iters. In normal bpf progs that have implicit rcu_read_lock(), it's OK to
> use them directly.
>
> This patch checks whether we are in rcu cs before we want to invoke
> bpf_iter_process_new and bpf_iter_css_{pre, post}_new in
> mark_stack_slots_iter(). If the rcu protection is guaranteed, we would
> let st->type = PTR_TO_STACK | MEM_RCU. is_iter_reg_valid_init() will
> reject if reg->type is UNTRUSTED.

I use the following BPF Prog to test this patch:

SEC("?fentry.s/" SYS_PREFIX "sys_getpgid")
int iter_task_for_each_sleep(void *ctx)
{
	struct task_struct *task;
	struct task_struct *cur_task = bpf_get_current_task_btf();

	if (cur_task->pid != target_pid)
		return 0;
	bpf_rcu_read_lock();
	bpf_for_each(process, task) {
		bpf_rcu_read_unlock();
		if (task->pid == target_pid)
			process_cnt += 1;
		bpf_rcu_read_lock();
	}
	bpf_rcu_read_unlock();
	return 0;
}

Unfortunately, we can pass the verifier.

Then I add some printk-messages before setting/clearing state to help debug:

diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
index d151e6b43a5f..35f3fa9471a9 100644
--- a/kernel/bpf/verifier.c
+++ b/kernel/bpf/verifier.c
@@ -1200,7 +1200,7 @@ static int mark_stack_slots_iter(struct 
bpf_verifier_env *env,
                __mark_reg_known_zero(st);
                st->type = PTR_TO_STACK; /* we don't have dedicated reg 
type */
                if (is_iter_need_rcu(meta)) {
+                       printk("mark reg_addr : %px", st);
                        if (in_rcu_cs(env))
                                st->type |= MEM_RCU;
                        else
@@ -11472,8 +11472,8 @@ static int check_kfunc_call(struct 
bpf_verifier_env *env, struct bpf_insn *insn,
                        return -EINVAL;
                } else if (rcu_unlock) {
                        bpf_for_each_reg_in_vstate(env->cur_state, 
state, reg, ({
+                               printk("clear reg_addr : %px MEM_RCU : 
%d PTR_UNTRUSTED : %d\n ", reg, reg->type & MEM_RCU, reg->type & 
PTR_UNTRUSTED);
                                if (reg->type & MEM_RCU) {
-                                       printk("clear reg addr : %lld", 
reg);
                                        reg->type &= ~(MEM_RCU | 
PTR_MAYBE_NULL);
                                        reg->type |= PTR_UNTRUSTED;
                                }


The demsg log:

[  393.705324] mark reg_addr : ffff88814e40e200

[  393.706883] clear reg_addr : ffff88814d5f8000 MEM_RCU : 0 
PTR_UNTRUSTED : 0

[  393.707353] clear reg_addr : ffff88814d5f8078 MEM_RCU : 0 
PTR_UNTRUSTED : 0

[  393.708099] clear reg_addr : ffff88814d5f80f0 MEM_RCU : 0 
PTR_UNTRUSTED : 0
....
....

I didn't see ffff88814e40e200 is cleared as expected because 
bpf_for_each_reg_in_vstate didn't find it.

It seems when we are doing bpf_read_unlock() in the middle of iteration 
and want to clearing state through bpf_for_each_reg_in_vstate, we can 
not find the previous reg which we marked MEM_RCU/PTR_UNTRUSTED in 
mark_stack_slots_iter().

I thought maybe the correct answer here is operating the *iter_reg* 
parameter in mark_stack_slots_iter() direcly so we can find it in 
bpf_for_each_reg_in_vstate.

diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
index 6a6827ba7a18..53330ddf2b3c 100644
--- a/kernel/bpf/verifier.c
+++ b/kernel/bpf/verifier.c
@@ -1218,6 +1218,12 @@ static int mark_stack_slots_iter(struct 
bpf_verifier_env *env,
                mark_stack_slot_scratched(env, spi - i);
        }

+       if (is_iter_need_rcu(meta)) {
+               if (in_rcu_cs(env))
+                       reg->type |= MEM_RCU;
+               else
+                       reg->type |= PTR_UNTRUSTED;
+       }
        return 0;
 }

@@ -1307,7 +1315,8 @@ static bool is_iter_reg_valid_init(struct 
bpf_verifier_env *env, struct bpf_reg_
                        if (slot->slot_type[j] != STACK_ITER)
                                Kumarreturn false;
        }
-
+       if (reg->type & PTR_UNTRUSTED)
+               return false;
        return true;
 }

However, it did not work either. The reason it didn't work is the state 
of iter_reg will be cleared implicitly before the 
is_iter_reg_valid_init() even we don't call bpf_rcu_unlock.

It would be appreciate if you could give some suggestion. Maby it worthy 
to try the solution proposed by Kumar?[1]

[1] 
https://lore.kernel.org/lkml/CAP01T77cWxWNwq5HLr+Woiu7k4-P3QQfJWX1OeQJUkxW3=P4bA@xxxxxxxxxxxxxx/#t