Re: [TEST FAILURE] bpf: s390: missed/kprobe_recursion

[Jiri Olsa <olsajiri@gmail.com>]

On Sun, Jan 26, 2025 at 11:40:05PM +0900, Masami Hiramatsu wrote:
> On Fri, 24 Jan 2025 16:41:38 +0100
> Jiri Olsa <olsajiri@gmail.com> wrote:
> 
> > On Fri, Jan 24, 2025 at 12:23:35PM +0100, Jiri Olsa wrote:
> > > On Thu, Jan 23, 2025 at 02:32:38PM -0800, Martin KaFai Lau wrote:
> > > > Hi Jiri,
> > > > 
> > > > The "missed/kprobe_recursion" fails consistently on s390. It seems to start
> > > > failing after the recent bpf and bpf-next tree ffwd.
> > > > 
> > > > An example:
> > > > https://github.com/kernel-patches/bpf/actions/runs/12934431612/job/36076956920
> > > > 
> > > > Can you help to take a look?
> > > > 
> > > > afaict, it only happens on s390 so far, so cc IIya if there is any recent
> > > > change that may ring the bell.
> > > 
> > > hi,
> > > I need to check more but I wonder it's the:
> > >   7495e179b478 s390/tracing: Enable HAVE_FTRACE_GRAPH_FUNC
> > > 
> > > which seems to add recursion check and bail out before we have
> > > a chance to trigger it in bpf code
> > 
> > so the test attaches bpf program test1 to bpf_fentry_test1 via kprobe.multi
> > 
> > 	SEC("kprobe.multi/bpf_fentry_test1")
> > 	int test1(struct pt_regs *ctx)
> > 	{
> > 		bpf_kfunc_common_test();
> > 		return 0;
> > 	}
> > 
> > and several other programs are attached to bpf_kfunc_common_test function
> > 
> > 
> > I can't test this on s390, but looks like following is happening:
> > 
> > kprobe.multi uses fprobe, so the test kernel path goes:
> > 
> >     bpf_fentry_test1
> >       ftrace_graph_func
> >         function_graph_enter_regs
> > 	   fprobe_entry
> > 	     kprobe_multi_link_prog_run
> > 	       test1 (bpf program)
> > 	         bpf_kfunc_common_test
> > 		   kprobe_ftrace_handler
> > 		     kprobe_perf_func
> > 		       trace_call_bpf
> > 		         -> bpf_prog_active check fails, missed count is incremented
> > 
> > 
> > kprobe_ftrace_handler calls/takes ftrace_test_recursion_trylock (ftrace recursion lock)
> > 
> > but s390 now calls/takes ftrace_test_recursion_trylock already in ftrace_graph_func,
> > so s390 stops at kprobe_ftrace_handler and does not get to trace_call_bpf to increment
> > prog->missed counters
> 
> Oops, good catch! I missed to remove it from s390. We've already moved it
> in function_graph_enter_regs().
> 
> 
> > 
> > adding Sven, Masami, any idea?
> > 
> > if the ftrace_test_recursion_trylock is needed ftrace_graph_func on s390, then
> > I think we will need to fix our test to skip s390 arch
> 
> Yes. Please try this patch;
> 
> 
> From 12fcda79d0b1082449d5f7cfb8039b0237cf246d Mon Sep 17 00:00:00 2001
> From: "Masami Hiramatsu (Google)" <mhiramat@kernel.org>
> Date: Sun, 26 Jan 2025 23:38:59 +0900
> Subject: [PATCH] s390: fgraph: Fix to remove ftrace_test_recursion_trylock()
> 
> Fix to remove ftrace_test_recursion_trylock() from ftrace_graph_func()
> because commit d576aec24df9 ("fgraph: Get ftrace recursion lock in
> function_graph_enter") has been moved it to function_graph_enter_regs()
> already.
> 
> Reported-by: Jiri Olsa <olsajiri@gmail.com>
> Fixes: d576aec24df9 ("fgraph: Get ftrace recursion lock in function_graph_enter")
> Signed-off-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>

great, ci is passing with this fix

Tested-by: Jiri Olsa <jolsa@kernel.org>

thanks,
jirka


> ---
>  arch/s390/kernel/ftrace.c | 5 -----
>  1 file changed, 5 deletions(-)
> 
> diff --git a/arch/s390/kernel/ftrace.c b/arch/s390/kernel/ftrace.c
> index c0b2c97efefb..63ba6306632e 100644
> --- a/arch/s390/kernel/ftrace.c
> +++ b/arch/s390/kernel/ftrace.c
> @@ -266,18 +266,13 @@ void ftrace_graph_func(unsigned long ip, unsigned long parent_ip,
>  		       struct ftrace_ops *op, struct ftrace_regs *fregs)
>  {
>  	unsigned long *parent = &arch_ftrace_regs(fregs)->regs.gprs[14];
> -	int bit;
>  
>  	if (unlikely(ftrace_graph_is_dead()))
>  		return;
>  	if (unlikely(atomic_read(&current->tracing_graph_pause)))
>  		return;
> -	bit = ftrace_test_recursion_trylock(ip, *parent);
> -	if (bit < 0)
> -		return;
>  	if (!function_graph_enter_regs(*parent, ip, 0, parent, fregs))
>  		*parent = (unsigned long)&return_to_handler;
> -	ftrace_test_recursion_unlock(bit);
>  }
>  
>  #endif /* CONFIG_FUNCTION_GRAPH_TRACER */
> -- 
> 2.43.0
> 
> Thank you,
> 
> -- 
> Masami Hiramatsu (Google) <mhiramat@kernel.org>