Re: [PATCH v3] tracing: precise log info for kretprobe addr err

[Steven Rostedt]

On Mon, 25 Jan 2021 19:19:27 +0100
Oleg Nesterov <oleg@xxxxxxxxxx> wrote:

> On 01/26, Jianlin Lv wrote:
> >
> > When trying to create kretprobe with the wrong function symbol in tracefs;
> > The error is triggered in the register_trace_kprobe() and recorded as
> > FAIL_REG_PROBE issue,
> >
> > Example:
> >   $ cd /sys/kernel/debug/tracing
> >   $ echo 'r:myprobe ERROR_SYMBOL_XXX ret=%x0' >> kprobe_events
> >     bash: echo: write error: Invalid argument
> >   $ cat error_log
> >     [142797.347877] trace_kprobe: error: Failed to register probe event
> >     Command: r:myprobe ERROR_SYMBOL_XXX ret=%x0
> >                        ^
> >
> > This error can be detected in the parameter parsing stage, the effect of
> > applying this patch is as follows:
> >
> >   $ echo 'r:myprobe ERROR_SYMBOL_XXX ret=%x0' >> kprobe_events
> >     bash: echo: write error: Invalid argument
> >   $ cat error_log
> >     [415.89]trace_kprobe: error: Retprobe address must be an function entry
> >     Command: r:myprobe ERROR_SYMBOL_XXX ret=%x0  
> 
> IOW, the "offset != 0" check removed by this patch is obviously wrong, right?
> 
> Agreed, but...
> 
> > --- a/kernel/trace/trace_kprobe.c
> > +++ b/kernel/trace/trace_kprobe.c
> > @@ -830,7 +830,7 @@ static int trace_kprobe_create(int argc, const char *argv[])
> >  			flags |= TPARG_FL_RETURN;
> >  		if (kprobe_on_func_entry(NULL, symbol, offset))
> >  			flags |= TPARG_FL_FENTRY;
> > -		if (offset && is_return && !(flags & TPARG_FL_FENTRY)) {
> > +		if (!strchr(symbol, ':') && is_return && !(flags & TPARG_FL_FENTRY)) {  
> 
> but why did you add the strchr(':') check instead?
> 
> I was really puzzled until I found the this email from Masami:
> https://lore.kernel.org/lkml/20210120131406.5a992c1e434681750a0cd5d4@xxxxxxxxxx/
> 
> So I leave this to you and Masami, but perhaps you can document this check at
> least in the changelog?
> 

No, you are correct. That needs to be documented in the code.

I was about to comment that the check requires a comment ;-)

Jianlin,

Care to send a v4 of the patch with a comment to why we are checking the
symbol for ':'.

Thanks!

-- Steve