Re: [PATCH bpf-next] bpf: Restrict attachment of bpf program to some tracepoints

[Jiri Olsa <olsajiri@gmail.com>]

On Mon, Dec 05, 2022 at 08:00:16PM -0800, Namhyung Kim wrote:
> On Mon, Dec 5, 2022 at 4:28 AM Jiri Olsa <olsajiri@gmail.com> wrote:
> >
> > On Sat, Dec 03, 2022 at 09:58:34AM -0800, Namhyung Kim wrote:
> > > On Wed, Nov 30, 2022 at 03:29:39PM -0800, Andrii Nakryiko wrote:
> > > > On Fri, Nov 25, 2022 at 1:35 AM Jiri Olsa <olsajiri@gmail.com> wrote:
> > > > >
> > > > > On Thu, Nov 24, 2022 at 09:17:22AM -0800, Alexei Starovoitov wrote:
> > > > > > On Thu, Nov 24, 2022 at 1:42 AM Jiri Olsa <olsajiri@gmail.com> wrote:
> > > > > > >
> > > > > > > On Thu, Nov 24, 2022 at 01:41:23AM +0100, Daniel Borkmann wrote:
> > > > > > > > On 11/21/22 10:31 PM, Jiri Olsa wrote:
> > > > > > > > > We hit following issues [1] [2] when we attach bpf program that calls
> > > > > > > > > bpf_trace_printk helper to the contention_begin tracepoint.
> > > > > > > > >
> > > > > > > > > As described in [3] with multiple bpf programs that call bpf_trace_printk
> > > > > > > > > helper attached to the contention_begin might result in exhaustion of
> > > > > > > > > printk buffer or cause a deadlock [2].
> > > > > > > > >
> > > > > > > > > There's also another possible deadlock when multiple bpf programs attach
> > > > > > > > > to bpf_trace_printk tracepoint and call one of the printk bpf helpers.
> > > > > > > > >
> > > > > > > > > This change denies the attachment of bpf program to contention_begin
> > > > > > > > > and bpf_trace_printk tracepoints if the bpf program calls one of the
> > > > > > > > > printk bpf helpers.
> > > > > > > > >
> > > > > > > > > Adding also verifier check for tb_btf programs, so this can be cought
> > > > > > > > > in program loading time with error message like:
> > > > > > > > >
> > > > > > > > >    Can't attach program with bpf_trace_printk#6 helper to contention_begin tracepoint.
> > > > > > > > >
> > > > > > > > > [1] https://lore.kernel.org/bpf/CACkBjsakT_yWxnSWr4r-0TpPvbKm9-OBmVUhJb7hV3hY8fdCkw@mail.gmail.com/
> > > > > > > > > [2] https://lore.kernel.org/bpf/CACkBjsaCsTovQHFfkqJKto6S4Z8d02ud1D7MPESrHa1cVNNTrw@mail.gmail.com/
> > > > > > > > > [3] https://lore.kernel.org/bpf/Y2j6ivTwFmA0FtvY@krava/
> > > > > > > > >
> > > > > > > > > Reported-by: Hao Sun <sunhao.th@gmail.com>
> > > > > > > > > Suggested-by: Alexei Starovoitov <ast@kernel.org>
> > > > > > > > > Signed-off-by: Jiri Olsa <jolsa@kernel.org>
> > > > > > > > > ---
> > > > > > > > >   include/linux/bpf.h          |  1 +
> > > > > > > > >   include/linux/bpf_verifier.h |  2 ++
> > > > > > > > >   kernel/bpf/syscall.c         |  3 +++
> > > > > > > > >   kernel/bpf/verifier.c        | 46 ++++++++++++++++++++++++++++++++++++
> > > > > > > > >   4 files changed, 52 insertions(+)
> > > > > > > > >
> > > > > > > > > diff --git a/include/linux/bpf.h b/include/linux/bpf.h
> > > > > > > > > index c9eafa67f2a2..3ccabede0f50 100644
> > > > > > > > > --- a/include/linux/bpf.h
> > > > > > > > > +++ b/include/linux/bpf.h
> > > > > > > > > @@ -1319,6 +1319,7 @@ struct bpf_prog {
> > > > > > > > >                             enforce_expected_attach_type:1, /* Enforce expected_attach_type checking at attach time */
> > > > > > > > >                             call_get_stack:1, /* Do we call bpf_get_stack() or bpf_get_stackid() */
> > > > > > > > >                             call_get_func_ip:1, /* Do we call get_func_ip() */
> > > > > > > > > +                           call_printk:1, /* Do we call trace_printk/trace_vprintk  */
> > > > > > > > >                             tstamp_type_access:1; /* Accessed __sk_buff->tstamp_type */
> > > > > > > > >     enum bpf_prog_type      type;           /* Type of BPF program */
> > > > > > > > >     enum bpf_attach_type    expected_attach_type; /* For some prog types */
> > > > > > > > > diff --git a/include/linux/bpf_verifier.h b/include/linux/bpf_verifier.h
> > > > > > > > > index 545152ac136c..7118c2fda59d 100644
> > > > > > > > > --- a/include/linux/bpf_verifier.h
> > > > > > > > > +++ b/include/linux/bpf_verifier.h
> > > > > > > > > @@ -618,6 +618,8 @@ bool is_dynptr_type_expected(struct bpf_verifier_env *env,
> > > > > > > > >                          struct bpf_reg_state *reg,
> > > > > > > > >                          enum bpf_arg_type arg_type);
> > > > > > > > > +int bpf_check_tp_printk_denylist(const char *name, struct bpf_prog *prog);
> > > > > > > > > +
> > > > > > > > >   /* this lives here instead of in bpf.h because it needs to dereference tgt_prog */
> > > > > > > > >   static inline u64 bpf_trampoline_compute_key(const struct bpf_prog *tgt_prog,
> > > > > > > > >                                          struct btf *btf, u32 btf_id)
> > > > > > > > > diff --git a/kernel/bpf/syscall.c b/kernel/bpf/syscall.c
> > > > > > > > > index 35972afb6850..9a69bda7d62b 100644
> > > > > > > > > --- a/kernel/bpf/syscall.c
> > > > > > > > > +++ b/kernel/bpf/syscall.c
> > > > > > > > > @@ -3329,6 +3329,9 @@ static int bpf_raw_tp_link_attach(struct bpf_prog *prog,
> > > > > > > > >             return -EINVAL;
> > > > > > > > >     }
> > > > > > > > > +   if (bpf_check_tp_printk_denylist(tp_name, prog))
> > > > > > > > > +           return -EACCES;
> > > > > > > > > +
> > > > > > > > >     btp = bpf_get_raw_tracepoint(tp_name);
> > > > > > > > >     if (!btp)
> > > > > > > > >             return -ENOENT;
> > > > > > > > > diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
> > > > > > > > > index f07bec227fef..b662bc851e1c 100644
> > > > > > > > > --- a/kernel/bpf/verifier.c
> > > > > > > > > +++ b/kernel/bpf/verifier.c
> > > > > > > > > @@ -7472,6 +7472,47 @@ static void update_loop_inline_state(struct bpf_verifier_env *env, u32 subprogno
> > > > > > > > >                              state->callback_subprogno == subprogno);
> > > > > > > > >   }
> > > > > > > > > +int bpf_check_tp_printk_denylist(const char *name, struct bpf_prog *prog)
> > > > > > > > > +{
> > > > > > > > > +   static const char * const denylist[] = {
> > > > > > > > > +           "contention_begin",
> > > > > > > > > +           "bpf_trace_printk",
> > > > > > > > > +   };
> > > > > > > > > +   int i;
> > > > > > > > > +
> > > > > > > > > +   /* Do not allow attachment to denylist[] tracepoints,
> > > > > > > > > +    * if the program calls some of the printk helpers,
> > > > > > > > > +    * because there's possibility of deadlock.
> > > > > > > > > +    */
> > > > > > > >
> > > > > > > > What if that prog doesn't but tail calls into another one which calls printk helpers?
> > > > > > >
> > > > > > > right, I'll deny that for all BPF_PROG_TYPE_RAW_TRACEPOINT* programs,
> > > > > > > because I don't see easy way to check on that
> > > > > > >
> > > > > > > we can leave printk check for tracing BPF_TRACE_RAW_TP programs,
> > > > > > > because verifier known the exact tracepoint already
> > > > > >
> > > > > > This is all fragile and merely a stop gap.
> > > > > > Doesn't sound that the issue is limited to bpf_trace_printk
> > > > >
> > > > > hm, I don't have a better idea how to fix that.. I can't deny
> > > > > contention_begin completely, because we use it in perf via
> > > > > tp_btf/contention_begin (perf lock contention) and I don't
> > > > > think there's another way for perf to do that
> > > > >
> > > > > fwiw the last version below denies BPF_PROG_TYPE_RAW_TRACEPOINT
> > > > > programs completely and tracing BPF_TRACE_RAW_TP with printks
> > > > >
> > > >
> > > > I think disabling bpf_trace_printk() tracepoint for any BPF program is
> > > > totally fine. This tracepoint was never intended to be attached to.
> > > >
> > > > But as for the general bpf_trace_printk() deadlocking. Should we
> > > > discuss how to make it not deadlock instead of starting to denylist
> > > > things left and right?
> > > >
> > > > Do I understand that we take trace_printk_lock only to protect that
> > > > static char buf[]? Can we just make this buf per-CPU and do a trylock
> > > > instead? We'll only fail to bpf_trace_printk() something if we have
> > > > nested BPF programs (rare) or NMI (also rare).
> > > >
> > > > And it's a printk(), it's never mission-critical, so if we drop some
> > > > message in rare case it's totally fine.
> > >
> > > What about contention_begin?  I wonder if we can disallow recursions
> > > for those in the deny list like using bpf_prog_active..
> >
> > I was testing change below which allows to check recursion just
> > for contention_begin tracepoint
> >
> > for the reported issue we might be ok with the change that Andrii
> > suggested, but we could have the change below as extra precaution
> 
> Looks ok to me.  But it seems it'd add the recursion check to every

hm, it should allocate recursion variable just for the contention_begin
tracepoint, rest should see NULL pointer

> tracepoint.  Can we just change the affected tracepoints only by
> using a kind of wrapped btp->bpf_func with some macro magic? ;-)

I tried that and the only other ways I found are:

  - add something like TRACE_EVENT_FLAGS macro and have __init call
    for specific tracepoint that sets the flag

  - add extra new 'bpf_func' that checks the re-entry, but that'd mean
    around 1000 extra mostly unused small functions

> 
> >
> > ---
> 
> [SNIP]
> > diff --git a/kernel/trace/bpf_trace.c b/kernel/trace/bpf_trace.c
> > index 3bbd3f0c810c..d27b7dc77894 100644
> > --- a/kernel/trace/bpf_trace.c
> > +++ b/kernel/trace/bpf_trace.c
> > @@ -2252,9 +2252,8 @@ void bpf_put_raw_tracepoint(struct bpf_raw_event_map *btp)
> >  }
> >
> >  static __always_inline
> > -void __bpf_trace_run(struct bpf_prog *prog, u64 *args)
> > +void __bpf_trace_prog_run(struct bpf_prog *prog, u64 *args)
> >  {
> > -       cant_sleep();
> >         if (unlikely(this_cpu_inc_return(*(prog->active)) != 1)) {
> >                 bpf_prog_inc_misses_counter(prog);
> >                 goto out;
> > @@ -2266,6 +2265,22 @@ void __bpf_trace_run(struct bpf_prog *prog, u64 *args)
> >         this_cpu_dec(*(prog->active));
> >  }
> >
> > +static __always_inline
> > +void __bpf_trace_run(struct bpf_raw_event_data *data, u64 *args)
> > +{
> > +       struct bpf_prog *prog = data->prog;
> > +
> > +       cant_sleep();
> > +       if (unlikely(!data->recursion))
> 
> likely ?

right, thanks

jirka