Add SystemTap static probes to the runtime linker. [BZ #14298]

7 files changed, 173 insertions, 5 deletions

diff --git a/ChangeLog b/ChangeLog
index 91826c2a0b..9702e50d08 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,3 +1,22 @@
+2012-07-27  Gary Benson  <gbenson@redhat.com>
+
+	[BZ #14298]
+	* elf/rtld.c: Include <stap-probe.h>.
+	(dl_main): Added static probes "init_start" and "init_complete".
+	* elf/dl-load.c: Include <stap-probe.h>.
+	(lose): Take new parameter "nsid".
+	Added static probe "map_failed".
+	(_dl_map_object_from_fd): Pass namespace id to lose.
+	Added static probe "map_start".
+	(open_verify): Pass namespace id to lose.
+	* elf/dl-open.c: Include <stap-probe.h>.
+	(dl_open_worker) Added static probes "map_complete", "reloc_start"
+	and "reloc_complete".
+	* elf/dl-close.c: Include <stap-probe.h>.
+	(_dl_close_worker): Added static probes "unmap_start" and
+	"unmap_complete".
+	* elf/rtld-debugger-interface.txt: New file documenting the above.
+
 2012-07-26  Roland McGrath  <roland@hack.frob.com>
 
 	* sunrpc/rpc_hout.c (pdeclaration): Call f_print with a "%s" format
diff --git a/NEWS b/NEWS
index d7084a02db..3c12922faf 100644
--- a/NEWS
+++ b/NEWS
@@ -9,8 +9,8 @@ Version 2.17
 
 * The following bugs are resolved with this release:
 
-  6778, 6808, 14042, 14151, 14154, 14157, 14173, 14283, 14328, 14331, 14337,
-  14347, 14349
+  6778, 6808, 14042, 14151, 14154, 14157, 14173, 14283, 14298, 14328, 14331,
+  14337, 14347, 14349
 
 * Support for STT_GNU_IFUNC symbols added for s390 and s390x.
   Optimized versions of memcpy, memset, and memcmp added for System z10 and
@@ -21,6 +21,9 @@ Version 2.17
   returning NULL if running in a SUID/SGID process.  This function replaces
   the internal function __secure_getenv.
 
+* SystemTap static probes have been added into the dynamic linker.
+  Implemented by Gary Benson.
+
 
 Version 2.16
 
diff --git a/elf/dl-close.c b/elf/dl-close.c
index a250ea5e3e..45b218787b 100644
--- a/elf/dl-close.c
+++ b/elf/dl-close.c
@@ -31,6 +31,7 @@
 #include <sys/mman.h>
 #include <sysdep-cancel.h>
 #include <tls.h>
+#include <stap-probe.h>
 
 
 /* Type of the constructor functions.  */
@@ -468,6 +469,7 @@ _dl_close_worker (struct link_map *map)
   struct r_debug *r = _dl_debug_initialize (0, nsid);
   r->r_state = RT_DELETE;
   _dl_debug_state ();
+  LIBC_PROBE (unmap_start, 2, nsid, r);
 
   if (unload_global)
     {
@@ -737,6 +739,7 @@ _dl_close_worker (struct link_map *map)
   /* Notify the debugger those objects are finalized and gone.  */
   r->r_state = RT_CONSISTENT;
   _dl_debug_state ();
+  LIBC_PROBE (unmap_complete, 2, nsid, r);
 
   /* Recheck if we need to retry, release the lock.  */
  out:
diff --git a/elf/dl-load.c b/elf/dl-load.c
index fe83f87eb9..43e1269c43 100644
--- a/elf/dl-load.c
+++ b/elf/dl-load.c
@@ -35,6 +35,7 @@
 #include <stackinfo.h>
 #include <caller.h>
 #include <sysdep.h>
+#include <stap-probe.h>
 
 #include <dl-dst.h>
 
@@ -882,7 +883,7 @@ _dl_init_paths (const char *llp)
 static void
 __attribute__ ((noreturn, noinline))
 lose (int code, int fd, const char *name, char *realname, struct link_map *l,
-      const char *msg, struct r_debug *r)
+      const char *msg, struct r_debug *r, Lmid_t nsid)
 {
   /* The file might already be closed.  */
   if (fd != -1)
@@ -896,6 +897,7 @@ lose (int code, int fd, const char *name, char *realname, struct link_map *l,
     {
       r->r_state = RT_CONSISTENT;
       _dl_debug_state ();
+      LIBC_PROBE (map_failed, 2, nsid, r);
     }
 
   _dl_signal_error (code, name, NULL, msg);
@@ -934,7 +936,7 @@ _dl_map_object_from_fd (const char *name, int fd, struct filebuf *fbp,
       errval = errno;
     call_lose:
       lose (errval, fd, name, realname, l, errstring,
-	    make_consistent ? r : NULL);
+	    make_consistent ? r : NULL, nsid);
     }
 
   /* Look again to see if the real name matched another already loaded.  */
@@ -1041,6 +1043,7 @@ _dl_map_object_from_fd (const char *name, int fd, struct filebuf *fbp,
 	 linking has not been used before.  */
       r->r_state = RT_ADD;
       _dl_debug_state ();
+      LIBC_PROBE (map_start, 2, nsid, r);
       make_consistent = true;
     }
   else
@@ -1736,7 +1739,7 @@ open_verify (const char *name, struct filebuf *fbp, struct link_map *loader,
 	      name = strdupa (realname);
 	      free (realname);
 	    }
-	  lose (errval, fd, name, NULL, NULL, errstring, NULL);
+	  lose (errval, fd, name, NULL, NULL, errstring, NULL, 0);
 	}
 
       /* See whether the ELF header is what we expect.  */
diff --git a/elf/dl-open.c b/elf/dl-open.c
index e2780a491b..00781afcf7 100644
--- a/elf/dl-open.c
+++ b/elf/dl-open.c
@@ -32,6 +32,7 @@
 #include <caller.h>
 #include <sysdep-cancel.h>
 #include <tls.h>
+#include <stap-probe.h>
 
 #include <dl-dst.h>
 
@@ -291,6 +292,7 @@ dl_open_worker (void *a)
   struct r_debug *r = _dl_debug_initialize (0, args->nsid);
   r->r_state = RT_CONSISTENT;
   _dl_debug_state ();
+  LIBC_PROBE (map_complete, 3, args->nsid, r, new);
 
   /* Print scope information.  */
   if (__builtin_expect (GLRO(dl_debug_mask) & DL_DEBUG_SCOPES, 0))
@@ -376,10 +378,19 @@ dl_open_worker (void *a)
 	}
     }
 
+  int relocation_in_progress = 0;
+
   for (size_t i = nmaps; i-- > 0; )
     {
       l = maps[i];
 
+      if (! relocation_in_progress)
+	{
+	  /* Notify the debugger that relocations are about to happen.  */
+	  LIBC_PROBE (reloc_start, 2, args->nsid, r);
+	  relocation_in_progress = 1;
+	}
+
 #ifdef SHARED
       if (__builtin_expect (GLRO(dl_profile) != NULL, 0))
 	{
@@ -544,6 +555,10 @@ cannot load any more object with static TLS"));
 	}
     }
 
+  /* Notify the debugger all new objects have been relocated.  */
+  if (relocation_in_progress)
+    LIBC_PROBE (reloc_complete, 3, args->nsid, r, new);
+
   /* Run the initializer functions of new objects.  */
   _dl_init (new, args->argc, args->argv, args->env);
 
diff --git a/elf/rtld-debugger-interface.txt b/elf/rtld-debugger-interface.txt
new file mode 100644
index 0000000000..61bc99e4b0
--- /dev/null
+++ b/elf/rtld-debugger-interface.txt
@@ -0,0 +1,122 @@
+Standard debugger interface
+===========================
+
+The run-time linker exposes a rendezvous structure to allow debuggers
+to interface with it.  This structure, r_debug, is defined in link.h.
+If the executable's dynamic section has a DT_DEBUG element, the
+run-time linker sets that element's value to the address where this
+structure can be found.
+
+The r_debug structure contains (amongst others) the following fields:
+
+  struct link_map *r_map:
+    A linked list of loaded objects.
+
+  enum { RT_CONSISTENT, RT_ADD, RT_DELETE } r_state:
+    The current state of the r_map list.  RT_CONSISTENT means that r_map
+    is not currently being modified and may safely be inspected.  RT_ADD
+    means that an object is being added to r_map, and that the list is
+    not guaranteed to be consistent.  Likewise RT_DELETE means that an
+    object is being removed from the list.
+
+  ElfW(Addr) r_brk:
+    The address of a function internal to the run-time linker which is
+    called whenever r_state is changed.  The debugger should set a
+    breakpoint at this address if it wants to notice mapping changes.
+
+This protocol is widely supported, but somewhat limited in that it
+has no provision to provide access to multiple namespaces, and that
+the notifications (via r_brk) only refer to changes to r_map--the
+debugger is notified that a new object has been added, for instance,
+but there is no way for the debugger to discover whether any of the
+objects in the link-map have been relocated or not.
+
+
+Probe-based debugger interface
+==============================
+
+Systemtap is a dynamic tracing/instrumenting tool available on Linux.
+Probes that are not fired at run time have close to zero overhead.
+glibc contains a number of probes that debuggers can set breakpoints
+on in order to notice certain events.
+
+All rtld probes have the following arguments:
+
+  arg1: Lmid_t lmid:
+    The link-map ID of the link-map list that the object was loaded
+    into.  This will be LM_ID_BASE for the application's main link-map
+    list, or some other value for different namespaces.
+
+  arg2: struct r_debug *r_debug:
+    A pointer to the r_debug structure containing the link-map list
+    that the object was loaded into.  This will be the value stored in
+    DT_DEBUG for the application's main link-map list, or some other
+    value for different namespaces.
+
+map_complete and reloc_complete may have the following additional
+argument:
+
+  arg3: struct link_map *new:
+    A pointer which, if not NULL, points to the entry in the specified
+    r_debug structure's link-map list corresponding to the first new
+    object to have been mapped or relocated, with new->l_next pointing
+    to the link-map of the next new object to have been mapped or
+    relocated, and so on.  Note that because `new' is an entry in a
+    larger list, new->l_prev (if not NULL) will point to what was the
+    last link-map in the link-map list prior to the new objects being
+    mapped or relocated.
+
+The following probes are available:
+
+  init_start:
+    This is called once, when the linker is about to fill in the main
+    r_debug structure at application startup.  init_start always has
+    lmid set to LM_ID_BASE and r_debug set to the value stored in
+    DT_DEBUG.  r_debug is not guaranteed to be consistent until
+    init_complete is fired.
+
+  init_complete:
+    This is called once, when the linker has filled in the main
+    r_debug structure at application startup. init_complete always
+    has lmid set to LM_ID_BASE and r_debug set to the value stored
+    in DT_DEBUG.  The r_debug structure is consistent and may be
+    inspected, and all objects in the link-map are guaranteed to
+    have been relocated.
+
+  map_start:
+    The linker is about to map new objects into the specified
+    namespace.  The namespace's r_debug structure is not guaranteed
+    to be consistent until a corresponding map_complete is fired.
+
+  map_complete:
+    The linker has finished mapping new objects into the specified
+    namespace.  The namespace's r_debug structure is consistent and
+    may be inspected, although objects in the namespace's link-map
+    are not guaranteed to have been relocated.
+
+  map_failed:
+    The linker failed while attempting to map new objects into
+    the specified namespace.  The namespace's r_debug structure
+    is consistent and may be inspected.
+
+  reloc_start:
+    The linker is about to relocate all unrelocated objects in the
+    specified namespace.  The namespace's r_debug structure is not
+    guaranteed to be consistent until a corresponding reloc_complete
+    is fired.
+
+  reloc_complete:
+    The linker has relocated all objects in the specified namespace.
+    The namespace's r_debug structure is consistent and may be
+    inspected, and all objects in the namespace's link-map are
+    guaranteed to have been relocated.
+
+  unmap_start:
+    The linker is about to remove objects from the specified
+    namespace.  The namespace's r_debug structure is not guaranteed to
+    be consistent until a corresponding unmap_complete is fired.
+
+  unmap_complete:
+    The linker has finished removing objects into the specified
+    namespace.  The namespace's r_debug structure is consistent and
+    may be inspected.
diff --git a/elf/rtld.c b/elf/rtld.c
index 6bcf224c49..06c4220c03 100644
--- a/elf/rtld.c
+++ b/elf/rtld.c
@@ -39,6 +39,7 @@
 #include <dl-osinfo.h>
 #include <dl-procinfo.h>
 #include <tls.h>
+#include <stap-probe.h>
 #include <stackinfo.h>
 
 #include <assert.h>
@@ -1683,6 +1684,7 @@ ERROR: ld.so: object '%s' cannot be loaded as audit interface: %s; ignored.\n",
   /* We start adding objects.  */
   r->r_state = RT_ADD;
   _dl_debug_state ();
+  LIBC_PROBE (init_start, 2, LM_ID_BASE, r);
 
   /* Auditing checkpoint: we are ready to signal that the initial map
      is being constructed.  */
@@ -2402,6 +2404,7 @@ ERROR: ld.so: object '%s' cannot be loaded as audit interface: %s; ignored.\n",
   r = _dl_debug_initialize (0, LM_ID_BASE);
   r->r_state = RT_CONSISTENT;
   _dl_debug_state ();
+  LIBC_PROBE (init_complete, 2, LM_ID_BASE, r);
 
 #ifndef MAP_COPY
   /* We must munmap() the cache file.  */