[PATCH] nptl: Export libthread_db-used symbols under GLIBC_PRIVATE

[Simon Marchi]

On 2021-06-27 6:25 p.m., Carlos O'Donell wrote:> On 6/20/21 8:38 AM, Simon Marchi via Libc-alpha wrote:
>> On 2021-06-17 3:23 p.m., Florian Weimer via Gdb wrote:
>>> Kevin, this *almost* gives us the perfect debugging experience with your
>>> patched GDB in Fedora 35, according to my preliminary tests.  A build
>>> with the patch is running here:
>>>
>>>   <https://koji.fedoraproject.org/koji/buildinfo?buildID=1773177>
>>>
>>> However, it seems that GDB still needs pthread_create in the .symtab to
>>> trigger loading of libpthread_db.  A fully stripped libc.so.6 without
>>> .symtab does not trigger loading in my experiments.  (The other symbols
>>> we preserve for valgrind's sake and are immaterial to GDB.)  In other
>>> words,
>>
>> Stupid question, but: since pthread_create is a function exposed by the
>> libc.so shared library, won't there still be an entry for it in .dynsym?
> 
> No question is stupid. Thank you for asking.
> 
> Yes, there *must* be a .dynsym entry for pthread_create in libc.so.6 in
> order for the dynamic loader to bind the reference to the definition.
> 
>> And if so, shouldn't GDB see it?
> 
> It should. Today gdb is only using the internal lookup_minimal_symbol()
> API in order to trigger libthread_db loading, and it's not clear to me
> if that will load *all* of the dynamic symbols during gdb startup.
> 
> Someone would need to debug the minsyms.c API to see if something isn't
> working as expected.
> 
> gdb/minsyms.c:
> 
> 1267    When files contain multiple sources of symbol information, it is
> 1268    possible for the minimal symbol table to contain many duplicate entries.
> 1269    As an example, SVR4 systems use ELF formatted object files, which
> 1270    usually contain at least two different types of symbol tables (a
> 1271    standard ELF one and a smaller dynamic linking table), as well as
> 1272    DWARF debugging information for files compiled with -g.
> 
> So it sounds like the data is loadable, and mergeable, but it appears that
> it's not all there at early startup in gdb.

My intuition is that the .dynsym symbols are normally loaded and result
in minimal symbols in GDB:

    https://gitlab.com/gnutools/gdb/-/blob/master/gdb/elfread.c#L1102-1128

But there may be some filtering that prevents the symbols like
pthread_create from making their way into the minimal symbols database.

I'll have to try using the container that Florian pointed to
(registry.fedoraproject.org/fedora:rawhide).

Simon