On Thu, Feb 16, 2023 at 5:55 AM Hou Tao <houtao@xxxxxxxxxxxxxxx> wrote:
>
> Beside BPF_REUSE_AFTER_RCU_GP, is BPF_FREE_AFTER_RCU_GP a feasible solution ?
The idea is for bpf_mem_free to wait normal RCU GP before adding
the elements back to the free list and free the elem to global kernel memory
only after both rcu and rcu_tasks_trace GPs as it's doing now.
> Its downside is that it will enforce sleep-able program to use
> bpf_rcu_read_{lock,unlock}() to access these returned pointers ?
sleepable can access elems without kptrs/spin_locks
even when not using rcu_read_lock, since it's safe, but there is uaf.
Some progs might be fine with it.
When sleepable needs to avoid uaf they will use bpf_rcu_read_lock.
