Re: [PATCH v4 1/2] function_graph: Support recording and printing the return value of function

[Donglin Peng]

On 2023/3/16 20:39, Masami Hiramatsu (Google) wrote:
> On Thu, 16 Mar 2023 11:18:23 +0800
> Donglin Peng <pengdonglin@xxxxxxxxxxxxxx> wrote:
>
> > On 2023/3/15 22:57, Peter Zijlstra wrote:
> > > On Wed, Mar 15, 2023 at 10:13:48AM -0400, Steven Rostedt wrote:
> > > > On Wed, 15 Mar 2023 14:49:11 +0100
> > > > Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> > > >
> > > > > > diff --git a/arch/x86/kernel/ftrace_64.S b/arch/x86/kernel/ftrace_64.S
> > > > > > index 1265ad519249..35ac9c58dc77 100644
> > > > > > --- a/arch/x86/kernel/ftrace_64.S
> > > > > > +++ b/arch/x86/kernel/ftrace_64.S
> > > > > > @@ -348,6 +348,10 @@ SYM_CODE_START(return_to_handler)
> > > > > >    	movq %rax, (%rsp)
> > > > > >    	movq %rdx, 8(%rsp)
> > > > > >    	movq %rbp, %rdi
> > > > > > +#ifdef CONFIG_FUNCTION_GRAPH_RETVAL
> > > > > > +	/* Pass the function return value to ftrace_return_to_handler */
> > > > > > +	movq %rax, %rsi
> > > > > > +#endif
> > > > > >    
> > > > > >    	call ftrace_return_to_handler
> > > > >
> > > > > What about the case of double register return values (when the value
> > > > > is returned in the A,D pair) ?
> > > >
> > > > Is there anything that does that in 64 bit kernels?
> > >
> > > Note sure; but I have a patch series that introduces cmpxchg128 and
> > > friends. Most of the actual functions are __always_inline, but still,
> > > the moment a compiler decides to break out a subfunction on a u128
> > > boundary we're in luck.
> > I have reviewed the kretprobe implementation and noticed that $retval
> > only retrieves the value of pt_regs.ax, which is an unsigned long data
> > type. I wrote a demo and tested it on an x86 machine, and found that
> > $retval only shows the least significant 32 bits of retval.Therefore,I
> > think it can be consistent with kretprobe.
>
> Yeah, because it doesn't support 128bit type (u128/long long).
> It is possible to support it, but I think we can just document this
> as a limitation for fgraph return value. It is just for quickly trace
> most of the return values.
Yeah,I will add this limitation to the documents in v5.

> Thank you,
>
> > static noinline unsigned long long test_retval_func(void)
> > {
> > 	unsigned long long value = 0x1234567887654321;
> > 	return value;
> > }
> >
> > add a kretprobe event:
> > echo 'r:myretprobe test_retval_func $retval:x64' > kprobe_events
> >
> > the trace log:
> > myretprobe: (retval_open+0x1c/0x2c [test_retval] <- test_retval_func)
> > arg1=0x87654321