Re: [PATCH RFC bpf-next v1 1/2] bpf: Fix deadlock for bpf_timer's spinlock

[Kumar Kartikeya Dwivedi <memxor@gmail.com>]

On Mon, Nov 07, 2022 at 09:15:26AM IST, Alexei Starovoitov wrote:
> On Sun, Nov 6, 2022 at 6:34 PM Alexei Starovoitov
> <alexei.starovoitov@gmail.com> wrote:
> >
> > >
> > > For bpf_timer, it would mean moving drop_prog_refcnt outside spin lock critical
> > > section. hrtimer_cancel is already done after the unlock. For bpf_list_head, it
>
> > > would mean swapping out the list_head and then draining it outside the lock.
> >
> > That also works.
> > drop_prog_refcnt() can be moved after unlock.
> > Don't see any race.
>
> I mean not the whole function obviously.
> Instead of
> static void drop_prog_refcnt(struct bpf_hrtimer *t)
> it can become
> static struct bpf_prog *drop_prog_refcnt(struct bpf_hrtimer *t)
> t->prog and callback_fn should only be manipulated
> under lock.
> bpf_prog_put itself can happen after unlock.

Right, both t->prog and t->callback_fn need to be set to NULL under the lock.

I will send out the bpf_timer change separately. For now, I moved list draining
out of the lock in my series and removed the check on BPF_PROG_TYPE_TRACING, and
posted it.