[PATCH v4 2/3] perf: kretprobes: offset from reloc_sym if kernel supports it

[Naveen N. Rao]

On 2017/03/04 01:34PM, Masami Hiramatsu wrote:
> On Sat, 4 Mar 2017 11:35:51 +0900
> Masami Hiramatsu <mhiramat at kernel.org> wrote:
> 
> > On Sat, 4 Mar 2017 09:49:11 +0900
> > Masami Hiramatsu <mhiramat at kernel.org> wrote:
> > 
> > > On Thu,  2 Mar 2017 23:25:06 +0530
> > > "Naveen N. Rao" <naveen.n.rao at linux.vnet.ibm.com> wrote:
> > > 
> > > > We indicate support for accepting sym+offset with kretprobes through a
> > > > line in ftrace README. Parse the same to identify support and choose the
> > > > appropriate format for kprobe_events.
> > > 
> > > Could you give us an example of this change here? :)
> > > for example, comment of commit 613f050d68a8 .
> > > 
> > > I think the code is OK, but we need actual example of result.
> > 
> > Hi Naveen,
> > 
> > I've tried following commands
> > 
> > $ grep "[Tt] user_read$" /proc/kallsyms  
> > 0000000000000000 T user_read
> > 0000000000000000 t user_read
> > $ sudo ./perf probe -D user_read%return
> > r:probe/user_read _text+3539616
> > r:probe/user_read_1 _text+3653408
> > 
> > OK, looks good. However, when I set the retprobes, I got an error.
> > 
> > $ sudo ./perf probe -a user_read%return
> > Failed to write event: Invalid argument
> >   Error: Failed to add events.
> > 
> > And kernel rejected that.
> > 
> > $ dmesg -k | tail -n 1
> > [  850.315068] Given offset is not valid for return probe.
> > 
> > Hmm, curious..
> 
> Ah, I see.
> 
> static int create_trace_kprobe(int argc, char **argv)
> ...
>         } else {
>                 /* a symbol specified */
>                 symbol = argv[1];
>                 /* TODO: support .init module functions */
>                 ret = traceprobe_split_symbol_offset(symbol, &offset);
>                 if (ret) {
>                         pr_info("Failed to parse symbol.\n");
>                         return ret;
>                 }
>                 if (offset && is_return &&
>                     !arch_function_offset_within_entry(offset)) {
>                         pr_info("Given offset is not valid for return probe.\n");
>                         return -EINVAL;
>                 }
>         }
> 
> So, actually, traceprobe_split_symbol_offset() just split out symbol
> and offset from symbol string (e.g. "_text+3539616").
> So, you should use kallsyms_lookup_size_offset() here again to check
> offset.

Ah, nice catch! I should have tested Steven's patch...

> 
> Please try attached patch (I've already tested on x86-64).
> 
> $ sudo ./perf probe -a user_read%return
> Added new events:
>   probe:user_read      (on user_read%return)
>   probe:user_read_1    (on user_read%return)
> 
> You can now use it in all perf tools, such as:
> 
> 	perf record -e probe:user_read_1 -aR sleep 1
> 
> $ sudo ./perf probe -l
>   probe:user_read      (on user_read%return at security/keys/user_defined.c)
>   probe:user_read_1    (on user_read%return at selinux/ss/policydb.c)
> $ sudo cat /sys/kernel/debug/kprobes/list
> ffffffff9637bf70  r  user_read+0x0    [DISABLED][FTRACE]
> ffffffff963602f0  r  user_read+0x0    [DISABLED][FTRACE]

Thanks, I will test this.


- Naveen