Re: [PATCH bpf v2] bpf: guard uprobes against private-stack corruption

On Tue, Aug 18, 2026 at 1:33 PM Christian Simon <[email protected]> wrote:
>
> Eligible BPF programs use one private stack per program and CPU. Both
> bpf_prog_run_array_uprobe() and uprobe_prog_run() use migrate_disable()
> to keep an invocation on one CPU, but another task can still preempt it
> and run the same program on that CPU. The second invocation then reuses
> and can overwrite the first invocation's private stack.
>
> Protect each real program invocation with the existing per-program
> recursion context. When the program is already active on this CPU,
> account for the missed invocation and skip it. Return zero when skipping
> an uprobe-multi invocation so session handling does not suppress its
> return probe.
>
> Skip dummy_bpf_prog in the classic array path before acquiring the
> recursion context because its active pointer is NULL.
>
> Add a regression test that pins two threads to one CPU and overlaps
> classic and multi uprobe invocations while preserving a sentinel in a
> private stack frame. Without the guards, the second invocation executes
> and corrupts the first invocation's sentinel.
>
> Fixes: 7d1cd70d4b16 ("bpf, x86: Support private stack in jit")
> Fixes: 6c17a882d380 ("bpf, arm64: JIT support for private stack")
> Closes: https://github.com/open-telemetry/opentelemetry-ebpf-instrumentation/issues/3056
> Cc: [email protected]
> Signed-off-by: Christian Simon <[email protected]>
> ---
> Changes in v2:
> - Address review comments from sashiko-bot
>   - Guard the uprobe-multi path as well.
>   - Remove const from correct line
> - Add a regression selftest for uprobe-classic/multi paths.
> - Add the arm64 Fixes tag.
>
>  include/linux/bpf.h                           |  33 +++--
>  kernel/trace/bpf_trace.c                      |  11 +-
>  .../bpf/prog_tests/uprobe_private_stack.c     | 118 ++++++++++++++++++
>  .../bpf/progs/uprobe_private_stack.c          |  54 ++++++++
>  4 files changed, 206 insertions(+), 10 deletions(-)
>  create mode 100644 tools/testing/selftests/bpf/prog_tests/uprobe_private_stack.c
>  create mode 100644 tools/testing/selftests/bpf/progs/uprobe_private_stack.c
>
> diff --git a/include/linux/bpf.h b/include/linux/bpf.h
> index 7719f6528445..a94fc9898ece 100644
> --- a/include/linux/bpf.h
> +++ b/include/linux/bpf.h
> @@ -2572,6 +2572,14 @@ static inline void bpf_reset_run_ctx(struct bpf_run_ctx *old_ctx)
>
>  typedef u32 (*bpf_prog_run_fn)(const struct bpf_prog *prog, const void *ctx);
>
> +#ifdef CONFIG_BPF_SYSCALL
> +void notrace bpf_prog_inc_misses_counter(struct bpf_prog *prog);
> +#else
> +static inline void bpf_prog_inc_misses_counter(struct bpf_prog *prog)
> +{
> +}
> +#endif
> +
>  static __always_inline u32
>  bpf_prog_run_array(const struct bpf_prog_array *array,
>                    const void *ctx, bpf_prog_run_fn run_prog)
> @@ -2617,7 +2625,7 @@ bpf_prog_run_array_uprobe(const struct bpf_prog_array *array,
>                           const void *ctx, bpf_prog_run_fn run_prog)
>  {
>         const struct bpf_prog_array_item *item;
> -       const struct bpf_prog *prog;
> +       struct bpf_prog *prog;
>         struct bpf_run_ctx *old_run_ctx;
>         struct bpf_trace_run_ctx run_ctx;
>         u32 ret = 1;
> @@ -2635,15 +2643,30 @@ bpf_prog_run_array_uprobe(const struct bpf_prog_array *array,
>         old_run_ctx = bpf_set_run_ctx(&run_ctx.run_ctx);
>         item = &array->items[0];
>         while ((prog = READ_ONCE(item->prog))) {
> +               /* dummy_bpf_prog has no recursion state. */
> +               if (unlikely(!prog->len)) {
> +                       item++;
> +                       continue;
> +               }
> +
> +               if (unlikely(!bpf_prog_get_recursion_context(prog))) {
> +                       bpf_prog_inc_misses_counter(prog);
> +                       bpf_prog_put_recursion_context(prog);
> +                       item++;
> +                       continue;
> +               }
> +

I think it's unacceptable to skip sleepable uprobe execution just
because there is the same BPF program attached to a *different* uprobe
(and all due to a private stack that no one asked for or needs for
uprobes, really).

As a short-term fix, we should probably disable private stack for
sleepable uprobe/kprobe program (and tracepoint/raw_tracepoint), and
think how we can make private stack less per-CPU dependent.

pw-bot: cr


>                 if (!prog->sleepable)
>                         rcu_read_lock();
>
>                 run_ctx.bpf_cookie = item->bpf_cookie;
>                 ret &= run_prog(prog, ctx);
> -               item++;
>
>                 if (!prog->sleepable)
>                         rcu_read_unlock();
> +
> +               bpf_prog_put_recursion_context(prog);
> +               item++;
>         }
>         bpf_reset_run_ctx(old_run_ctx);
>         migrate_enable();
> @@ -3208,8 +3231,6 @@ static inline bool has_current_bpf_ctx(void)
>         return !!current->bpf_ctx;
>  }
>
> -void notrace bpf_prog_inc_misses_counter(struct bpf_prog *prog);
> -
>  void bpf_dynptr_init(struct bpf_dynptr_kern *ptr, void *data,
>                      enum bpf_dynptr_type type, u32 offset, u32 size);
>  void bpf_dynptr_set_null(struct bpf_dynptr_kern *ptr);

[...]