On Mon, Apr 10, 2023 at 1:25 AM Steven Rostedt <rostedt@xxxxxxxxxxx> wrote:
>
> On Sun, 9 Apr 2023 20:45:39 +0800
> Yafang Shao <laoar.shao@xxxxxxxxx> wrote:
> >
> > I didn't notice preempt_{disable,enable}_notrace before :(
> > Seems that is a better solution. I have verified that it really works.
>
> Great.
>
> >
> > BTW, why can we attach fentry to preempt_count_sub(), but can't attach
> > it to migrate_disable() ?
> > migrate_disable() isn't hidden from ftrace either...
>
> You can't?
>
> # trace-cmd -p function -l migrate_disable
> # trace-cmd show
> <idle>-0 [000] ..s2. 153664.937829: migrate_disable <-bpf_prog_run_clear_cb
> <idle>-0 [000] ..s2. 153664.937834: migrate_disable <-bpf_prog_run_clear_cb
> <idle>-0 [000] ..s2. 153664.937835: migrate_disable <-bpf_prog_run_clear_cb
> <idle>-0 [000] ..s2. 153665.769555: migrate_disable <-bpf_prog_run_clear_cb
> <idle>-0 [000] ..s2. 153665.772109: migrate_disable <-bpf_prog_run_clear_cb
>
> I think bpf prevents it. Perhaps that's another solution:
>
> See 35e3815fa8102 ("bpf: Add deny list of btf ids check for tracing programs")
>
Right, it is because of the BTF_ID(). I think we should introduce
migrate_{disable,enable}_notrace instead. Because actually we only
don't want it to attach migrate_{disable,enable} used in bpf prog
enter and exit, while it is safe to attach it at other places.
--
Regards
Yafang
