Re: [PATCH bpf-next v5 2/5] selftests/bpf: Add tests for open-coded style process file iterator

From: Juntong Deng
Date: Wed Dec 11 2024 - 16:28:01 EST


On 2024/12/10 18:51, Alexei Starovoitov wrote:
On Tue, Dec 10, 2024 at 8:23 AM Juntong Deng <juntong.deng@xxxxxxxxxxx> wrote:

+SEC("fentry/" SYS_PREFIX "sys_nanosleep")
+int test_bpf_iter_task_file(void *ctx)
+{
+ struct bpf_iter_task_file task_file_it;
+ struct bpf_iter_task_file_item *item;
+ struct task_struct *task;
+
+ task = bpf_get_current_task_btf();
+ if (task->parent->pid != parent_pid)
+ return 0;
+
+ count++;
+
+ bpf_rcu_read_lock();

What does the RCU read lock do here exactly?


Thanks for your reply.

This is used to solve the problem previously discussed in v3 [0].

Task ref may be released during iteration.

[0]:
https://lore.kernel.org/bpf/CAADnVQ+0LUXxmfm1YgyGDz=cciy3+dGGM-Zysq84fpAdaB74Qw@xxxxxxxxxxxxxx/

I think you misunderstood my comment.

"If this object _was_ RCU protected ..."

Adding rcu_read_lock doesn't make 'task' pointer RCU protected.
That's not how RCU works.

So patch 1 doing:

item->task = task;

is not correct.

See bpf_iter_task_vma_new(). It's doing:
kit->data->task = get_task_struct(task);
to make sure task stays valid while iterating.

pw-bot: cr

Thanks for your reply.

Sorry for the misunderstanding.

I will fix it in the next version.