Re: BUG: rseq selftests and librseq vs. glibc fail

[Sean Christopherson]

On Tue, Aug 19, 2025, Thomas Gleixner wrote:
> On Mon, Aug 18 2025 at 13:27, Sean Christopherson wrote:
> > On Mon, Aug 18, 2025, Florian Weimer wrote:
> >> You need both (extern and weak) to get a weak symbol reference instead
> >> of a weak symbol definition.  You still need to check &__rseq_offset, of
> >> course.
> >
> > Ooh, you're saying add "extern" to the existing __weak symbol, not replace it.
> > Huh, TIL weak symbol references are a thing.
> >
> > This works with static and dynamic linking, with and without an rseq-aware glibc.
> >
> > Thomas, does this fix the problem you were seeing?
> >
> > diff --git a/tools/testing/selftests/rseq/rseq.c b/tools/testing/selftests/rseq/rseq.c
> > index 663a9cef1952..d17ded120d48 100644
> > --- a/tools/testing/selftests/rseq/rseq.c
> > +++ b/tools/testing/selftests/rseq/rseq.c
> > @@ -40,9 +40,9 @@
> >   * Define weak versions to play nice with binaries that are statically linked
> >   * against a libc that doesn't support registering its own rseq.
> >   */
> > -__weak ptrdiff_t __rseq_offset;
> > -__weak unsigned int __rseq_size;
> > -__weak unsigned int __rseq_flags;
> > +extern __weak ptrdiff_t __rseq_offset;
> > +extern __weak unsigned int __rseq_size;
> > +extern __weak unsigned int __rseq_flags;
> >  
> >  static const ptrdiff_t *libc_rseq_offset_p = &__rseq_offset;
> >  static const unsigned int *libc_rseq_size_p = &__rseq_size;
> > @@ -209,7 +209,7 @@ void rseq_init(void)
> >          * libc not having registered a restartable sequence.  Try to find the
> >          * symbols if that's the case.
> >          */
> > -       if (!*libc_rseq_size_p) {
> > +       if (!libc_rseq_offset_p || !*libc_rseq_size_p) {

Doh, I meant to check libc_rseq_size_p for NULL, i.e.

	if (!libc_rseq_size_p || !*libc_rseq_size_p) {

> 
> If I make that:
> 
> +       if (!*libc_rseq_offset_p || !*libc_rseq_size_p) {
> 
> then it makes sense and actually works. The pointer can hardly be NULL,
> even when statically linked, no?

IIUC, it is indeed the pointers that are set to NULL/0, because for unresolved
symbols, the symbol itself, not its value, is set to '0'.  Which makes sense,
because if there is no symbol, then it can't have a value.

I.e. the address of the symbol is '0', and its value is undefined.

E.g. statically linking this against glibc without rseq support:

diff --git a/tools/testing/selftests/rseq/rseq.c b/tools/testing/selftests/rseq/rseq.c
index 663a9cef1952..959bdcb32e96 100644
--- a/tools/testing/selftests/rseq/rseq.c
+++ b/tools/testing/selftests/rseq/rseq.c
@@ -40,9 +40,9 @@
  * Define weak versions to play nice with binaries that are statically linked
  * against a libc that doesn't support registering its own rseq.
  */
-__weak ptrdiff_t __rseq_offset;
-__weak unsigned int __rseq_size;
-__weak unsigned int __rseq_flags;
+extern __weak ptrdiff_t __rseq_offset;
+extern __weak unsigned int __rseq_size;
+extern __weak unsigned int __rseq_flags;
 
 static const ptrdiff_t *libc_rseq_offset_p = &__rseq_offset;
 static const unsigned int *libc_rseq_size_p = &__rseq_size;
@@ -209,7 +209,12 @@ void rseq_init(void)
         * libc not having registered a restartable sequence.  Try to find the
         * symbols if that's the case.
         */
-       if (!*libc_rseq_size_p) {
+       printf("libc_rseq_offset_p = %lx (%lx), libc_rseq_size_p = %lx (%lx)\n",
+              (unsigned long)libc_rseq_offset_p, (unsigned long)libc_rseq_size_p,
+              (unsigned long)&__rseq_offset, (unsigned long)&__rseq_size);
+       printf("__rseq_size = %u\n", __rseq_size);
+
+       if (!libc_rseq_size_p || !*libc_rseq_size_p) {
                libc_rseq_offset_p = dlsym(RTLD_NEXT, "__rseq_offset");
                libc_rseq_size_p = dlsym(RTLD_NEXT, "__rseq_size");
                libc_rseq_flags_p = dlsym(RTLD_NEXT, "__rseq_flags");

Generates this output:

  $ ./rseq_test 
  libc_rseq_offset_p = 0 (0), libc_rseq_size_p = 0 (0)
  Segmentation fault

Because trying to dereference __rseq_size hits NULL/0.