r16111: Patch from Björn JACKE <samba@j3e.de>.

This fixes a problem where the clock definition for clock_gettime() is
present at compile time, but is not available on the running system. In
this case, we fall back to less-preferred clocks until we find one that
we can use.

source/aclocal.m4

@@ -839,10 +839,17 @@ dnl Test whether the current LIBS results in libpthread being present.
 dnl Execute the corresponding user action list.
 AC_DEFUN([SMB_IS_LIBPTHREAD_LINKED],
 [
+    AC_MSG_CHECKING(if libpthread is linked)
     AC_TRY_LINK([],
 	[return pthread_create(0, 0, 0, 0);],
-	[$1],
+	[
-	[$2])
+	    AC_MSG_RESULT(yes)
+	    $1
+	],
+	[
+	    AC_MSG_RESULT(no)
+	    $2
+	])
 ])
 
 dnl SMB_REMOVE_LIB(lib)
@@ -943,3 +950,34 @@ void main(void) {
     fi
 
 ])
+
+dnl SMB_CHECK_CLOCK_ID(clockid)
+dnl Test whether the specified clock_gettime clock ID is available. If it
+dnl is, we define HAVE_clockid
+AC_DEFUN([SMB_CHECK_CLOCK_ID],
+[
+    AC_MSG_CHECKING(for $1)
+    AC_TRY_LINK([
+#if TIME_WITH_SYS_TIME
+# include <sys/time.h>
+# include <time.h>
+#else
+# if HAVE_SYS_TIME_H
+#  include <sys/time.h>
+# else
+#  include <time.h>
+# endif
+#endif
+    ],
+    [
+clockid_t clk = $1;
+    ],
+    [
+	AC_MSG_RESULT(yes)
+	AC_DEFINE(HAVE_$1, 1,
+	    [Whether the clock_gettime clock ID $1 is available])
+    ],
+    [
+	AC_MSG_RESULT(no)
+    ])
+])

source/configure.in

@@ -2053,8 +2053,13 @@ if test x"$samba_cv_WITH_PROFILE" = x"yes"; then
 	    [
 		SMB_IS_LIBPTHREAD_LINKED(
 			[ SMB_REMOVELIB(rt) ],
-			[ AC_DEFINE(HAVE_CLOCK_GETTIME, 1,
+			[
-			    [Whether clock_gettime is available]) ])
+			    AC_DEFINE(HAVE_CLOCK_GETTIME, 1,
+				[Whether clock_gettime is available])
+			    SMB_CHECK_CLOCK_ID(CLOCK_MONOTONIC)
+			    SMB_CHECK_CLOCK_ID(CLOCK_PROCESS_CPUTIME_ID)
+			    SMB_CHECK_CLOCK_ID(CLOCK_REALTIME)
+			])
 	    ])
 
 fi

source/profile/profile.c

@@ -30,6 +30,7 @@ static int shm_id;
 static BOOL read_only;
 #if defined(HAVE_CLOCK_GETTIME)
 clockid_t __profile_clock;
+BOOL have_profiling_clock = False;
 #endif
 #endif
 
@@ -62,6 +63,19 @@ void profile_message(int msg_type, struct process_id src, void *buf, size_t len)
 			 (int)procid_to_pid(&src)));
 		break;
 	case 2:		/* turn on complete profiling */
+
+#if defined(HAVE_CLOCK_GETTIME)
+		if (!have_profiling_clock) {
+			do_profile_flag = True;
+			do_profile_times = False;
+			DEBUG(1,("INFO: Profiling counts turned ON from "
+				"pid %d\n", (int)procid_to_pid(&src)));
+			DEBUGADD(1,("INFO: Profiling times disabled "
+				"due to lack of a suitable clock\n"));
+			break;
+		}
+#endif
+
 		do_profile_flag = True;
 		do_profile_times = True;
 		DEBUG(1,("INFO: Full profiling turned ON from pid %d\n",
@@ -100,28 +114,74 @@ void reqprofile_message(int msg_type, struct process_id src,
   open the profiling shared memory area
   ******************************************************************/
 #ifdef WITH_PROFILE
+
+#ifdef HAVE_CLOCK_GETTIME
+
+/* Find a clock. Just because the definition for a particular clock ID is
+ * present doesn't mean the system actually supports it.
+ */
+static void init_clock_gettime(void)
+{
+	struct timespec ts;
+
+	have_profiling_clock = False;
+
+#ifdef HAVE_CLOCK_PROCESS_CPUTIME_ID
+	/* CLOCK_PROCESS_CPUTIME_ID is sufficiently fast that the
+	 * always profiling times is plausible. Unfortunately on Linux
+	 * it is only accurate if we can guarantee we will not be scheduled
+	 * scheduled onto a different CPU between samples. Until there is
+	 * some way to set processor affinity, we can only use this on
+	 * uniprocessors.
+	 */
+	if (!this_is_smp()) {
+	    if (clock_gettime(CLOCK_PROCESS_CPUTIME_ID, &ts) == 0) {
+		    DEBUG(10, ("Using CLOCK_PROCESS_CPUTIME_ID "
+				"for profile_clock\n"));
+		    __profile_clock = CLOCK_PROCESS_CPUTIME_ID;
+		    have_profiling_clock = True;
+	    }
+	}
+#endif
+
+#ifdef HAVE_CLOCK_MONOTONIC
+	if (!have_profiling_clock &&
+	    clock_gettime(CLOCK_MONOTONIC, &ts) == 0) {
+		DEBUG(10, ("Using CLOCK_MONOTONIC for profile_clock\n"));
+		__profile_clock = CLOCK_MONOTONIC;
+		have_profiling_clock = True;
+		return;
+	}
+#endif
+
+#ifdef HAVE_CLOCK_REALTIME
+	/* POSIX says that CLOCK_REALTIME should be defined everywhere
+	 * where we have clock_gettime...
+	 */
+	if (!have_profiling_clock &&
+	    clock_gettime(CLOCK_REALTIME, &ts) == 0) {
+		__profile_clock = CLOCK_REALTIME;
+		have_profiling_clock = True;
+	}
+
+	SMB_WARN(__profile_clock == CLOCK_REALTIME,
+		("Using (slow) CLOCK_REALTIME for profile_clock"));
+#endif
+
+	SMB_WARN(have_profiling_clock == False,
+		("could not find a working clock for profiling"));
+	return have_profiling_clock;
+}
+#endif
+
 BOOL profile_setup(BOOL rdonly)
 {
 	struct shmid_ds shm_ds;
 
 	read_only = rdonly;
 
-#if defined(HAVE_CLOCK_GETTIME)
+#ifdef HAVE_CLOCK_GETTIME
-	if (this_is_smp()) {
+	init_clock_gettime();
-		/* This is faster that gettimeofday, but not fast enough to
-		 * leave it enabled in production.
-		 */
-		__profile_clock = CLOCK_MONOTONIC;
-	} else {
-		/* CLOCK_PROCESS_CPUTIME_ID is sufficiently fast that the
-		 * always profiling times is plausible. Unfortunately it is
-		 * only accurate if we can guarantee we will not be scheduled
-		 * onto a different CPU between samples. Until there is some
-		 * way to set processor affinity, we can only use this on
-		 * uniprocessors.
-		 */
-		__profile_clock = CLOCK_PROCESS_CPUTIME_ID;
-	}
 #endif
 
  again: