Re: [RFC PATCH bpf-next v2 2/2] bpf: Pass external callchain entry to get_perf_callchain

[Andrii Nakryiko <andrii.nakryiko@gmail.com>]

On Tue, Oct 14, 2025 at 8:02 AM Alexei Starovoitov
<alexei.starovoitov@gmail.com> wrote:
>
> On Tue, Oct 14, 2025 at 5:14 AM Jiri Olsa <olsajiri@gmail.com> wrote:
> >
> > On Tue, Oct 14, 2025 at 06:01:28PM +0800, Tao Chen wrote:
> > > As Alexei noted, get_perf_callchain() return values may be reused
> > > if a task is preempted after the BPF program enters migrate disable
> > > mode. Drawing on the per-cpu design of bpf_perf_callchain_entries,
> > > stack-allocated memory of bpf_perf_callchain_entry is used here.
> > >
> > > Signed-off-by: Tao Chen <chen.dylane@linux.dev>
> > > ---
> > >  kernel/bpf/stackmap.c | 19 +++++++++++--------
> > >  1 file changed, 11 insertions(+), 8 deletions(-)
> > >
> > > diff --git a/kernel/bpf/stackmap.c b/kernel/bpf/stackmap.c
> > > index 94e46b7f340..acd72c021c0 100644
> > > --- a/kernel/bpf/stackmap.c
> > > +++ b/kernel/bpf/stackmap.c
> > > @@ -31,6 +31,11 @@ struct bpf_stack_map {
> > >       struct stack_map_bucket *buckets[] __counted_by(n_buckets);
> > >  };
> > >
> > > +struct bpf_perf_callchain_entry {
> > > +     u64 nr;
> > > +     u64 ip[PERF_MAX_STACK_DEPTH];
> > > +};
> > > +

we shouldn't introduce another type, there is perf_callchain_entry in
linux/perf_event.h, what's the problem with using that?

> > >  static inline bool stack_map_use_build_id(struct bpf_map *map)
> > >  {
> > >       return (map->map_flags & BPF_F_STACK_BUILD_ID);
> > > @@ -305,6 +310,7 @@ BPF_CALL_3(bpf_get_stackid, struct pt_regs *, regs, struct bpf_map *, map,
> > >       bool user = flags & BPF_F_USER_STACK;
> > >       struct perf_callchain_entry *trace;
> > >       bool kernel = !user;
> > > +     struct bpf_perf_callchain_entry entry = { 0 };
> >
> > so IIUC having entries on stack we do not need to do preempt_disable
> > you had in the previous version, right?
> >
> > I saw Andrii's justification to have this on the stack, I think it's
> > fine, but does it have to be initialized? it seems that only used
> > entries are copied to map
>
> No. We're not adding 1k stack consumption.

Right, and I thought we concluded as much last time, so it's a bit
surprising to see this in this patch.

Tao, you should go with 3 entries per CPU used in a stack-like
fashion. And then passing that entry into get_perf_callchain() (to
avoid one extra copy).

>
> pw-bot: cr