hwclock: improve, and then disable clever sync code: it's bloat

...and hardware is too stupid to benefit from it anyway

function                                             old     new   delta
hwclock_main                                         439     319    -120

Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>

util-linux/hwclock.c

@@ -109,10 +109,53 @@ static void to_sys_clock(const char **pp_rtcname, int utc)
 
 static void from_sys_clock(const char **pp_rtcname, int utc)
 {
-#define TWEAK_USEC 200
+#if 1
+	struct timeval tv;
+	struct tm tm_time;
+	int rtc;
+
+	rtc = rtc_xopen(pp_rtcname, O_WRONLY);
+	gettimeofday(&tv, NULL);
+	/* Prepare tm_time */
+	if (sizeof(time_t) == sizeof(tv.tv_sec)) {
+		if (utc)
+			gmtime_r((time_t*)&tv.tv_sec, &tm_time);
+		else
+			localtime_r((time_t*)&tv.tv_sec, &tm_time);
+	} else {
+		time_t t = tv.tv_sec;
+		if (utc)
+			gmtime_r(&t, &tm_time);
+		else
+			localtime_r(&t, &tm_time);
+	}
+#else
+/* Bloated code which tries to set hw clock with better precision.
+ * On x86, even though code does set hw clock within <1ms of exact
+ * whole seconds, apparently hw clock (at least on some machines)
+ * doesn't reset internal fractional seconds to 0,
+ * making all this a pointless excercise.
+ */
+	/* If we see that we are N usec away from whole second,
+	 * we'll sleep for N-ADJ usecs. ADJ corrects for the fact
+	 * that CPU is not infinitely fast.
+	 * On infinitely fast CPU, next wakeup would be
+	 * on (exactly_next_whole_second - ADJ). On real CPUs,
+	 * this difference between current time and whole second
+	 * is less than ADJ (assuming system isn't heavily loaded).
+	 */
+	/* Small value of 256us gives very precise sync for 2+ GHz CPUs.
+	 * Slower CPUs will fail to sync and will go to bigger
+	 * ADJ values. qemu-emulated armv4tl with ~100 MHz
+	 * performance ends up using ADJ ~= 4*1024 and it takes
+	 * 2+ secs (2 tries with successively larger ADJ)
+	 * to sync. Even straced one on the same qemu (very slow)
+	 * takes only 4 tries.
+	 */
+#define TWEAK_USEC 256
+	unsigned adj = TWEAK_USEC;
 	struct tm tm_time;
 	struct timeval tv;
-	unsigned adj = TWEAK_USEC;
 	int rtc = rtc_xopen(pp_rtcname, O_WRONLY);
 
 	/* Try to catch the moment when whole second is close */
@@ -124,55 +167,64 @@ static void from_sys_clock(const char **pp_rtcname, int utc)
 
 		t = tv.tv_sec;
 		rem_usec = 1000000 - tv.tv_usec;
-		if (rem_usec < 1024) {
+		if (rem_usec < adj) {
-			/* Less than 1ms to next second. Good enough */
+			/* Close enough */
  small_rem:
 			t++;
 		}
 
-		/* Prepare tm */
+		/* Prepare tm_time from t */
 		if (utc)
 			gmtime_r(&t, &tm_time); /* may read /etc/xxx (it takes time) */
 		else
 			localtime_r(&t, &tm_time); /* same */
-		tm_time.tm_isdst = 0;
+
+		if (adj >= 32*1024) {
+			break; /* 32 ms diff and still no luck?? give up trying to sync */
+		}
 
 		/* gmtime/localtime took some time, re-get cur time */
 		gettimeofday(&tv, NULL);
 
-		if (tv.tv_sec < t /* may happen if rem_usec was < 1024 */
+		if (tv.tv_sec < t /* we are still in old second */
-		 || (tv.tv_sec == t && tv.tv_usec < 1024)
+		 || (tv.tv_sec == t && tv.tv_usec < adj) /* not too far into next second */
 		) {
-			/* We are not too far into next second. Good. */
+			break; /* good, we are in sync! */
-			break;
 		}
-		adj += 32; /* 2^(10-5) = 2^5 = 32 iterations max */
+
-		if (adj >= 1024) {
+		rem_usec = 1000000 - tv.tv_usec;
-			/* Give up trying to sync */
+		if (rem_usec < adj) {
-			break;
+			t = tv.tv_sec;
+			goto small_rem; /* already close to next sec, don't sleep */
 		}
 
 		/* Try to sync up by sleeping */
-		rem_usec = 1000000 - tv.tv_usec;
-		if (rem_usec < 1024) {
-			goto small_rem; /* already close, don't sleep */
-		}
-		/* Need to sleep.
-		 * Note that small adj on slow processors can make us
-		 * to always overshoot tv.tv_usec < 1024 check on next
-		 * iteration. That's why adj is increased on each iteration.
-		 * This also allows it to be reused as a loop limiter.
-		 */
 		usleep(rem_usec - adj);
-	}
 
-	xioctl(rtc, RTC_SET_TIME, &tm_time);
+		/* Jump to 1ms diff, then increase fast (x2): EVERY loop
-
+		 * takes ~1 sec, people won't like slowly converging code here!
-	/* Debug aid to find "good" TWEAK_USEC.
-	 * Look for a value which makes tv_usec close to 999999 or 0.
-	 * for 2.20GHz Intel Core 2: TWEAK_USEC ~= 200
 		 */
-	//bb_error_msg("tv.tv_usec:%d adj:%d", (int)tv.tv_usec, adj);
+	//bb_error_msg("adj:%d tv.tv_usec:%d", adj, (int)tv.tv_usec);
+		if (adj < 512)
+			adj = 512;
+		/* ... and if last "overshoot" does not look insanely big,
+		 * just use it as adj increment. This makes convergence faster.
+		 */
+		if (tv.tv_usec < adj * 8) {
+			adj += tv.tv_usec;
+			continue;
+		}
+		adj *= 2;
+	}
+	/* Debug aid to find "optimal" TWEAK_USEC with nearly exact sync.
+	 * Look for a value which makes tv_usec close to 999999 or 0.
+	 * For 2.20GHz Intel Core 2: optimal TWEAK_USEC ~= 200
+	 */
+	//bb_error_msg("tv.tv_usec:%d", (int)tv.tv_usec);
+#endif
+
+	tm_time.tm_isdst = 0;
+	xioctl(rtc, RTC_SET_TIME, &tm_time);
 
 	if (ENABLE_FEATURE_CLEAN_UP)
 		close(rtc);