Re: PR3498: more efficient dwfl searching

[Wenji Huang <wenji dot huang at oracle dot com>]

Josh Stone wrote:
> On 07/15/2009 02:18 AM, Wenji Huang wrote:
> > Wenji Huang wrote:
> > > I found one regression since this commit.
> > >
> > > $ stap -l nfsd.create*
> > > Segmentation fault
> > > [78375.044067] <6>stap[24911]: segfault at 0 ip 0000000000480efb sp 
> > > 00007ffffe706400 error 4 in stap[400000+142000]
>
> Works for me -- can you provide more details about your configuration?

Related config.
CONFIG_NFSD=y
# CONFIG_NFSD_V3 is not set

Another test case
$ stap -p2 -e 'probe a2=module("no module").function("no_func")? {} 
probe a1=kernel.function("sys_read")?, module("no 
module").function("a")? {} probe a* {}'
Segmentation fault

This can be reproduced on self-built kernel x86/x86_64, elfutils 
0.141/0.137. But it's strange that it works well on FC11, the 
distributed kernel. And output
Missing separate debuginfos, use: debuginfo-install 
kernel-PAE-2.6.29.4-167.fc11.i686
> It should not be possible for dw to come out NULL, so I would rather not
> paper over this.  Can you root cause how this happened?

Did deep investigation, Seems the culprit is the segment in 
dwflpp::setup_kernel.
...
  if (! offline_search_match_p)
    {
      if (debuginfo_needed) {
        // Suggest a likely kernel dir to find debuginfo rpm for
        string dir = string("/lib/modules/" + sess.kernel_release );
        find_debug_rpms(sess, dir.c_str());
      }
      throw semantic_error (string("missing ") + sess.architecture +
                            string(" kernel/module debuginfo under '") +
                            sess.kernel_build_tree + string("'"));
    }
...
For example, "no module" doesn't exit, there will be error thrown and
dwflpp::dwflpp{setup_kernel} constructor procedure can't be completely 
finished. But stap will tolerate such kind of error for optional probe 
and regard the structure is well formatted. So the searched dw from 
get_kern_dw isn't in good shape.

Regards,
Wenji