On 7/27/26 7:28 AM, Leon Hwang wrote:
> The BPF tracing multi link updates several direct-call sites through one
> ftrace_ops. Its implementation is therefore gated by
> HAVE_SINGLE_FTRACE_DIRECT_OPS in addition to
> DYNAMIC_FTRACE_WITH_DIRECT_CALLS.
> 
> Select HAVE_SINGLE_FTRACE_DIRECT_OPS whenever arm64 enables dynamic ftrace
> direct calls. This enables BPF tracing multi links on arm64. Also
> generalize the unreachable-trampoline comment because the single-ops path
> does not use ops->direct_call.

Hi Leon,

I don't think this change can land as is yet. The series doesn't even
apply cleanly to bpf-next, but that's minor.

More importantly, it depends on Jose's series [1], which is not in the
mainline yet. And there Mark has raised performance concerns [2] and
the discussion still seems to be open.

[1] 
https://lore.kernel.org/all/[email protected]/
[2] https://lore.kernel.org/all/amjnf5gz0xP5PTSB@J2N7QTR9R3/

> 
> Assisted-by: Codex:gpt-5.6-sol
> Signed-off-by: Leon Hwang <[email protected]>
> ---
>  arch/arm64/Kconfig         | 2 ++
>  arch/arm64/kernel/ftrace.c | 3 +--
>  2 files changed, 3 insertions(+), 2 deletions(-)
> 
> diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig
> index 0de419ed780f..c98dca76859b 100644
> --- a/arch/arm64/Kconfig
> +++ b/arch/arm64/Kconfig
> @@ -188,6 +188,8 @@ config ARM64
>                   CLANG_SUPPORTS_DYNAMIC_FTRACE_WITH_ARGS)
>       select HAVE_DYNAMIC_FTRACE_WITH_DIRECT_CALLS \
>               if DYNAMIC_FTRACE_WITH_ARGS
> +     select HAVE_SINGLE_FTRACE_DIRECT_OPS \
> +             if DYNAMIC_FTRACE_WITH_DIRECT_CALLS\

The select is only conditional on DYNAMIC_FTRACE_WITH_DIRECT_CALLS, so
it can be set along with HAVE_DYNAMIC_FTRACE_WITH_CALL_OPS. And AFAIU
this would make the fast path effectively dead: every BPF direct call
routed through ftrace_caller now goes through the slow path.

As Mark noted in the other thread, on arm64 trampolines come from
EXECMEM_BPF, so they always land out of BL range (chance of landing in
range is 256M/terabytes).

I vibe-slop-coded a benchmark and ran it on a Neoverse V2 machine, and
toggling your config change seems to be causing a 1.3x regression in
the tracing overhead:

  do-nothing fentry (r0=0; exit) on __arm64_sys_getpid, 20 M calls, min-of-N, 
several boots. Results:

              untraced (base)   traced        overhead
  Kernel A    ~128 ns/call      ~145.5 ns     ~17.6 ns   (fast path: br x17)
  Kernel B    ~127 ns/call      ~150.4 ns     ~23.4 ns   (slow path: save regs 
+ call_direct_funcs + hash)

This confirms Jiri's suspicion.

However my understanding is the regression should mostly disappear in
case some version of in-range trampoline allocation on arm64 lands.

So, I think the landing sequence should be something like follows:
  * in-range BPF-trampoline allocation that Jose proposed [3]
  * then HAVE_SINGLE_FTRACE_DIRECT_OPS selection

After all of that reaches mainline, then a selftest patch can go
through the bpf-next.

[3] https://lore.kernel.org/all/[email protected]/

>       select HAVE_DYNAMIC_FTRACE_WITH_CALL_OPS \
>               if (DYNAMIC_FTRACE_WITH_ARGS && !CFI && \
>                   (CC_IS_CLANG || !CC_OPTIMIZE_FOR_SIZE))
> diff --git a/arch/arm64/kernel/ftrace.c b/arch/arm64/kernel/ftrace.c
> index e1a3c0b3a051..56ba72a87dfa 100644
> --- a/arch/arm64/kernel/ftrace.c
> +++ b/arch/arm64/kernel/ftrace.c
> @@ -301,8 +301,7 @@ static bool ftrace_find_callable_addr(struct dyn_ftrace 
> *rec,
>  
>       /*
>        * If a custom trampoline is unreachable, rely on the ftrace_caller
> -      * trampoline which knows how to indirectly reach that trampoline
> -      * through ops->direct_call.
> +      * trampoline which knows how to indirectly reach that trampoline.
>        */
>       if (*addr != FTRACE_ADDR && !reachable_by_bl(*addr, pc))
>               *addr = FTRACE_ADDR;


Reply via email to