Linux: Make __rseq_size useful for feature detection (bug 31965)

The __rseq_size value is now the active area of struct rseq
(so 20 initially), not the full struct size including padding
at the end (32 initially).

Update misc/tst-rseq to print some additional diagnostics.

Reviewed-by: Michael Jeanson <mjeanson@efficios.com>
Reviewed-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
(cherry picked from commit 2e456ccf0c34a056e3ccafac4a0c7effef14d918)

5 files changed, 43 insertions, 12 deletions

diff --git a/NEWS b/NEWS
index 0f0ebce3f0..f6ae9e2337 100644
--- a/NEWS
+++ b/NEWS
@@ -14,6 +14,11 @@ Major new features:
   configured on the current host i.e. as-if you had not passed
   AI_ADDRCONFIG to getaddrinfo calls.
 
+Deprecated and removed features, and other changes affecting compatibility:
+
+* __rseq_size now denotes the size of the active rseq area (20 bytes
+  initially), not the size of struct rseq (32 bytes initially).
+
 Security related changes:
 
   CVE-2022-39046: When the syslog function is passed a crafted input
@@ -92,6 +97,7 @@ The following bugs are resolved with this release:
   [30843] potential use-after-free in getcanonname (CVE-2023-4806)
   [31184] FAIL: elf/tst-tlsgap
   [31185] Incorrect thread point access in _dl_tlsdesc_undefweak and _dl_tlsdesc_dynamic
+  [31965] rseq extension mechanism does not work as intended
 
 Version 2.36
 
diff --git a/manual/threads.texi b/manual/threads.texi
index 48fd562923..fcf9e2bba5 100644
--- a/manual/threads.texi
+++ b/manual/threads.texi
@@ -1020,8 +1020,12 @@ This variable is either zero (if restartable sequence registration
 failed or has been disabled) or the size of the restartable sequence
 registration.  This can be different from the size of @code{struct rseq}
 if the kernel has extended the size of the registration.  If
-registration is successful, @code{__rseq_size} is at least 32 (the
-initial size of @code{struct rseq}).
+registration is successful, @code{__rseq_size} is at least 20 (the
+initially active size of @code{struct rseq}).
+
+Previous versions of @theglibc{} set this to 32 even if the kernel only
+supported the initial area of 20 bytes because the value included unused
+padding at the end of the restartable sequence area.
 @end deftypevar
 
 @deftypevar {unsigned int} __rseq_flags
diff --git a/sysdeps/nptl/dl-tls_init_tp.c b/sysdeps/nptl/dl-tls_init_tp.c
index 1e451a9dc1..662bc0158d 100644
--- a/sysdeps/nptl/dl-tls_init_tp.c
+++ b/sysdeps/nptl/dl-tls_init_tp.c
@@ -46,10 +46,6 @@ rtld_mutex_dummy (pthread_mutex_t *lock)
 
 const unsigned int __rseq_flags;
 
-/* The variables are in .data.relro but are not yet write-protected.  */
-extern unsigned int _rseq_size attribute_hidden;
-extern ptrdiff_t _rseq_offset attribute_hidden;
-
 void
 __tls_pre_init_tp (void)
 {
@@ -108,9 +104,7 @@ __tls_init_tp (void)
     do_rseq = TUNABLE_GET (rseq, int, NULL);
 #endif
     if (rseq_register_current_thread (pd, do_rseq))
-      {
-        _rseq_size = sizeof (pd->rseq_area);
-      }
+      _rseq_size = RSEQ_AREA_SIZE_INITIAL_USED;
 
 #ifdef RSEQ_SIG
     /* This should be a compile-time constant, but the current
diff --git a/sysdeps/unix/sysv/linux/rseq-internal.h b/sysdeps/unix/sysv/linux/rseq-internal.h
index 210f3ec566..f08a70dfc4 100644
--- a/sysdeps/unix/sysv/linux/rseq-internal.h
+++ b/sysdeps/unix/sysv/linux/rseq-internal.h
@@ -25,15 +25,34 @@
 #include <stdio.h>
 #include <sys/rseq.h>
 
+/* 32 is the initially required value for the area size.  The
+   actually used rseq size may be less (20 bytes initially).  */
+#define RSEQ_AREA_SIZE_INITIAL 32
+#define RSEQ_AREA_SIZE_INITIAL_USED 20
+
+/* The variables are in .data.relro but are not yet write-protected.  */
+extern unsigned int _rseq_size attribute_hidden;
+extern ptrdiff_t _rseq_offset attribute_hidden;
+
 #ifdef RSEQ_SIG
 static inline bool
 rseq_register_current_thread (struct pthread *self, bool do_rseq)
 {
   if (do_rseq)
     {
+      unsigned int size;
+#if IS_IN (rtld)
+      /* Use the hidden symbol in ld.so.  */
+      size = _rseq_size;
+#else
+      size = __rseq_size;
+#endif
+      if (size < RSEQ_AREA_SIZE_INITIAL)
+        /* The initial implementation used only 20 bytes out of 32,
+           but still expected size 32.  */
+        size = RSEQ_AREA_SIZE_INITIAL;
       int ret = INTERNAL_SYSCALL_CALL (rseq, &self->rseq_area,
-                                       sizeof (self->rseq_area),
-                                       0, RSEQ_SIG);
+                                       size, 0, RSEQ_SIG);
       if (!INTERNAL_SYSCALL_ERROR_P (ret))
         return true;
     }
diff --git a/sysdeps/unix/sysv/linux/tst-rseq.c b/sysdeps/unix/sysv/linux/tst-rseq.c
index fa6a89541f..613593f7f9 100644
--- a/sysdeps/unix/sysv/linux/tst-rseq.c
+++ b/sysdeps/unix/sysv/linux/tst-rseq.c
@@ -29,6 +29,7 @@
 # include <stdlib.h>
 # include <string.h>
 # include <syscall.h>
+# include <sys/auxv.h>
 # include <thread_pointer.h>
 # include <tls.h>
 # include "tst-rseq.h"
@@ -42,7 +43,8 @@ do_rseq_main_test (void)
   TEST_COMPARE (__rseq_flags, 0);
   TEST_VERIFY ((char *) __thread_pointer () + __rseq_offset
                == (char *) &pd->rseq_area);
-  TEST_COMPARE (__rseq_size, sizeof (pd->rseq_area));
+  /* The current implementation only supports the initial size.  */
+  TEST_COMPARE (__rseq_size, 20);
 }
 
 static void
@@ -52,6 +54,12 @@ do_rseq_test (void)
     {
       FAIL_UNSUPPORTED ("kernel does not support rseq, skipping test");
     }
+  printf ("info: __rseq_size: %u\n", __rseq_size);
+  printf ("info: __rseq_offset: %td\n", __rseq_offset);
+  printf ("info: __rseq_flags: %u\n", __rseq_flags);
+  printf ("info: getauxval (AT_RSEQ_FEATURE_SIZE): %ld\n",
+          getauxval (AT_RSEQ_FEATURE_SIZE));
+  printf ("info: getauxval (AT_RSEQ_ALIGN): %ld\n", getauxval (AT_RSEQ_ALIGN));
   do_rseq_main_test ();
 }
 #else /* RSEQ_SIG */