Re: Part 2: perf test case probe libc fails with latest Fedora34 glibc update

[Masami Hiramatsu <mhiramat@kernel.org>]

Hi,

On Thu, 1 Jul 2021 01:02:08 +0900
Masami Hiramatsu <mhiramat@kernel.org> wrote:

> Hi Thomas,
> 
> OK, I finally found 2 root causes for this issue.
> 
> 1. debuginfo__new() forgot to call dso__load(). Thus, the dso is
> just initialized, but doesn't load the build-id. Then the build-id
> based debuginfo file is not searched. It must use map instead of dso.

I made a fix for this issue using dso__set_build_id() because
using dso__load() which loads symbols just for build-id seems sledgehammer.
So now I can see the perf probe opens build-id based debuginfo correctly.

However, this will not fix your issue because;

> 2. Not sure why, but /usr/lib64/libc-2.33.so is shown as "not stripped"
> binary. Then the elfutils is fooled by the information and open it
> as a debuginfo file. I think probe-finder.c also need to check the
> given file has DIE tree (debuginfo) or not by itself.

this part is not correct. perf-probe is using elfutils, which has
its own debuginfo search algorithm. Thus with or without the above
fix, perf-probe finally opens the correct debuginfo.
(from this reason, I can not show what has been fixed by the command
 output in commit message without debug print.)

And with decoding the debuginfo, I finally understand what was wrong.
The problem is that the "inet_pton" is a dynamic symbol, which does *NOT*
exist in the debuginfo. There is "__inet_pton" instead of "inet_pton".

-----
/* Like __inet_pton_length, but use strlen (SRC) as the length of
   SRC.  */
int
__inet_pton (int af, const char *src, void *dst)
{
  return __inet_pton_length (af, src, strlen (src), dst);
}
libc_hidden_def (__inet_pton)
weak_alias (__inet_pton, inet_pton)
libc_hidden_weak (inet_pton)
-----

Thus, if you pass the __inet_pton to perf-probe, it works.

-----
$ ./perf probe -x /lib64/libc.so.6 -L __inet_pton
<__inet_pton@/usr/src/debug/glibc-2.33-18.fc34.aarch64/resolv/inet_pton.c:0>
      0  __inet_pton (int af, const char *src, void *dst)
         {
           return __inet_pton_length (af, src, strlen (src), dst);
         }
-----

And -F (--functions) doesn't work yet.
-----
./perf probe  -x /usr/lib64/libc-2.33.so -F
@plt
@plt
calloc@plt
free@plt
malloc@plt
memalign@plt
realloc@plt
-----

I think we should remove @plt, and decode ".dynsym" in addition to ".symtab"
by map object for this issue at first. Then we may be able to get the address
of "inet_pton" from map.

Thank you,

-- 
Masami Hiramatsu <mhiramat@kernel.org>