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

[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.

> > +ifeq ($(build-shared),yes)
> > +# dl-debug-compat-symbols.os and dl-debug-compat-symbols.o are assembly
> > +# codes included in dl-debug-symbols.S.
> > +generated += dl-debug-compat-symbols.os dl-debug-compat-symbols.o
> > +
> > +libof-dl-debug-compat-symbols = rtld
> > +
> > +$(objpfx)dl-debug-compat-symbols.os: dl-debug-symbols-gen.c
> > +     $(compile-command.c) -S
> > +
> > +$(objpfx)dl-debug-compat-symbols.o: dl-debug-symbols-gen.c
> > +     $(compile-command.c) -S
> > +
> > +$(objpfx)dl-debug-symbols.os: $(objpfx)dl-debug-compat-symbols.os
> > +$(objpfx)dl-debug-symbols.o: $(objpfx)dl-debug-compat-symbols.o
> > +endif
>
> 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]$

> And it's not that the size of struct _r_debug is going to change.
>
> > diff --git a/elf/dl-close.c b/elf/dl-close.c
> > index f39001cab9..43873d2543 100644
> > --- a/elf/dl-close.c
> > +++ b/elf/dl-close.c
> > @@ -709,6 +709,27 @@ _dl_close_worker (struct link_map *map, bool force)
> >             assert (nsid != LM_ID_BASE);
> >             ns->_ns_loaded = imap->l_next;
> >
> > +           if (ns->_ns_loaded == NULL)
> > +             {
> > +               /* Remove the empty namespace from the namespace linked
> > +                  list.  */
> > +               struct r_debug_extended **pp, *p;
> > +
> > +               for (pp = &_r_debug_extended.r_next;
> > +                    (p = *pp) != NULL;
> > +                    pp = &p->r_next)
> > +                 if (p == &ns->_ns_debug)
> > +                   {
> > +                     /* Remove the empty namespace.  */
> > +                     *pp = p->r_next;
> > +
> > +                     /* Clear r_version to indicate that it is
> > +                        unused.  */
> > +                     p->base.r_version = 0;
> > +                     break;
> > +                   }
> > +             }
>
> 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?

> does not save any memory.

_dl_debug_initialize can be called multiple times with the same
namespace.  Only the first time we need to initialize the namespace.
I use base.r_version == 0 to check if I need to initialize the namespace
and put it on the linked list.

>
> > diff --git a/elf/rtld-debugger-interface.txt b/elf/rtld-debugger-interface.txt
> > index 61bc99e4b0..f6aaa28706 100644
> > --- a/elf/rtld-debugger-interface.txt
> > +++ b/elf/rtld-debugger-interface.txt
> > @@ -9,6 +9,9 @@ structure can be found.
> >
> >  The r_debug structure contains (amongst others) the following fields:
> >
> > +  int r_version:
> > +    Version number for this protocol.  It should be greater than 0.
> > +
>
> It seems to me that r_version starts out as 0 and is initialized later.
> Maybe this should be described here, and tested in the test case.

r_version is never zero (initialized by _dl_debug_initialize) when the
namespace is in use.  The unused namespace isn't accessible from
user programs.

> Thanks,
> Florian
>

Thanks.

-- 
H.J.