[PATCH v6 2/2] Extend struct r_debug to support multiple namespaces

[H.J. Lu]

On Mon, Sep 6, 2021 at 7:24 AM Florian Weimer <fweimer@redhat.com> wrote:
>
> * H. J. Lu:
>
> > On Mon, Sep 6, 2021 at 2:39 AM Florian Weimer <fweimer@redhat.com> wrote:
> >>
> >> * H. J. Lu:
> >>
> >> > +* The r_version update in the debugger interface makes the glibc binary
> >> > +  incompatible with GDB binaries built without the following commits:
> >> > +
> >> > +  c0154a4a21a gdb: Don't assume r_ldsomap when r_version > 1 on Linux
> >> > +  4eb629d50d4 gdbserver: Check r_version < 1 for Linux debugger interface
> >>
> >> Does this incompatibility happen even if audit modules and dlmopen are
> >> not used?
> >
> > Yes.
>
> Why?  Can't we keep r_version at 1 in this case?

r_version is checked in GDB whenever DT_DEBUG is used to access
the rtld debug interface, independent of audit modules and dlmopen.

We can bump r_version only if a non-default namespace is used.

> >> This puts the assember output from the compiler through the
> >> preprocessor.  That seems to be brittle.  I think you would have to
> >> preprocess the manually written fragment separately.
> >>
> >> However, I think we are overdesigning things here.  The following in
> >> dl-debug-symbols-gen.c should work (and the file should have a different
> >> name then):
> >>
> >> /* Alias _r_debug to a prefix of _r_debug_extended.  */
> >> asm (".set _r_debug, _r_debug_extended\n\t"
> >>      ".type _r_debug, %object\n\t"
> >>      ".symver _r_debug_extended, _r_debug@@" FIRST_VERSION_ld__r_debug_STRING);
> >> #if __WORDSIZE == 64
> >> _Static_assert (sizeof (struct r_debug) == 40, "sizeof (struct r_debug)");
> >> asm (".size _r_debug, 40");
> >> #else
> >> _Static_assert (sizeof (struct r_debug) == 20, "sizeof (struct r_debug)");
> >> asm (".size _r_debug, 20");
> >> #endif
> >>
> >> It's not exactly pretty, but at least it's obvious what is going on.
> >> (Extended asm with input operands is not supported outside of functions.)
> >
> > This was the first thing I tried and it didn't work:
> >
> > [hjl@gnu-cfl-2 tmp]$ cat foo.s
> > .set  _r_debug, _r_debug_extended
> > .globl _r_debug
> > .type _r_debug, %object
> > .size _r_debug, 40
> > .data
> > .type _r_debug_extended, %object
> > .size _r_debug_extended, 48
> > .globl _r_debug_extended
> > _r_debug_extended:
> > .zero 48
> > [hjl@gnu-cfl-2 tmp]$ gcc -c foo.s
> > [hjl@gnu-cfl-2 tmp]$ readelf -sW foo.o | grep _r_debug
> >      1: 0000000000000000    48 OBJECT  GLOBAL DEFAULT    2 _r_debug
> >      2: 0000000000000000    48 OBJECT  GLOBAL DEFAULT    2 _r_debug_extended
> > [hjl@gnu-cfl-2 tmp]$
>
> Huh.  Does this mean this depends on the symbol definition order in the
> assembler file?
>
> I really hate the post-processing of compiler output.  This isn't GHC. 8->
>
> Can we write a portable assembler file instead?
>
> Nick Clifton has written down some guidelines:
>
>   Tips for writing portable assembler with GNU Assembler (GAS)
>   <https://developers.redhat.com/blog/2021/02/26/tips-for-writing-portable-assembler-with-gnu-assembler-gas>
>
> There's no initializer, so all we need to know is size and alignment.

Yes, I can do that.

> >> Is this necessary?  It makes concurrent access to the list harder and
> >
> > When _dl_close_worker is called, it holds GL(dl_load_lock).  Why does
> > this change make concurrent access harder?
>
> Something else might want to read the list directly, by starting with
> DT_DEBUG.

I can add _dl_debug_get and use _dl_debug_initialize only for initialization.
Will it work?


-- 
H.J.