On Fri, Jul 17, 2020 at 01:06:07AM +0200, Daniel Borkmann wrote:
> > + ret = bpf_arch_text_poke(poke->tailcall_bypass,
> > + BPF_MOD_JUMP,
> > + NULL, bypass_addr);
> > + BUG_ON(ret < 0 && ret != -EINVAL);
> > + /* let other CPUs finish the execution of program
> > + * so that it will not possible to expose them
> > + * to invalid nop, stack unwind, nop state
> > + */
> > + synchronize_rcu();
>
> Very heavyweight that we need to potentially call this /multiple/ times for just a
> /single/ map update under poke mutex even ... but agree it's needed here to avoid
> racing. :(
Yeah. I wasn't clear with my suggestion earlier.
I meant to say that synchronize_rcu() can be done between two loops.
list_for_each_entry(elem, &aux->poke_progs, list)
for (i = 0; i < elem->aux->size_poke_tab; i++)
bpf_arch_text_poke(poke->tailcall_bypass, ...
synchronize_rcu();
list_for_each_entry(elem, &aux->poke_progs, list)
for (i = 0; i < elem->aux->size_poke_tab; i++)
bpf_arch_text_poke(poke->poke->tailcall_target, ...
Not sure how much better it will be though.
text_poke is heavy.
I think it's heavier than synchronize_rcu().
Long term we can do batch of text_poke-s.
I'm actually fine with above approach of synchronize_rcu() without splitting the loop.
This kind of optimizations can be done later as a follow up.
I'd really like to land this stuff in this bpf-next cycle.
It's a big improvement to tail_calls and bpf2bpf calls.
