>Number:         3944
>Category:       bin
>Synopsis:       <ntptime -c results off by factor of ~4>
>Confidential:   yes
>Severity:       serious
>Priority:       medium
>Responsible:    bin-bug-people (Utility Bug People)
>State:          open
>Class:          sw-bug
>Submitter-Id:   net
>Arrival-Date:   Wed Aug  6 02:20:00 1997
>Last-Modified:
>Originator:     
>Organization:
	
>Release:        <NetBSD-current source date>
>Environment:
	
System: NetBSD Cup.DSG.Stanford.EDU 1.2G NetBSD 1.2G (PLACEBO) #0: Wed Jul 30 10:31:00 PDT 1997 jonathan@Cup.DSG.Stanford.EDU:/usr/src/sys/arch/i386/compile/PLACEBO i386


[Mailed to xntpd maintainers and submitted to the NetBSD bug database.]


>Description:

    ntptime -c in xntp-3.5-90 produces an estimates of  syscall overhead
    (time between repeated calls to ntp_gettime) that is 2x-4x
    the true cost.

    This does not appear to  be fixed in 3.5-91.


>How-To-Repeat:

    Build ntptime from 3-5.90 (or as modified from the NetBSD source tree.)
    run ntptime -c.

    Compare to xntpd 3.5f ntptime -c on the same hardware and OS.

    A Decstation 5000/240 with reports 43 usecs and 16 usecs,
    for  3-5.90 and 3.5f,  respectively (either Ultrix or NetBSD).
    An Intel Pentium/133 with NetBSD reports 12 and 6, respectively.

    This rather defeats the raison d'e\^tre of ntptime -c :).

	
>Fix:

    The ntp source code uses sigsetjmp() to establish a signal hanlder to
    gracefully handle the case of calling ntp_gettime() on a kernel that
    doesn't have the in-kernel PLL code.

    UTSL shows that ntptime is establishing such a signal handler for
    every iteration of the ntptime -c loop, rather than just once,
    as the 3.5f code did.

    (Since the loop breaks if pll_control indicates an error, we can
     only trigger the handler once here anyway.)

    The fix is obvious: move the sigsetjmp() outside the loop.
    I can't figure out the house indentation style for the xntp code, sorry.



*** ntptime.c.DIST	Tue Jun 24 04:30:03 1997
--- ntptime.c	Wed Aug  6 01:39:14 1997
***************
*** 197,220 ****
  #endif /* BADCALL */
  
    if (cost) {
-     for (c = 0; c < sizeof times / sizeof times[0]; c++)
-       {
  #ifdef SIGSYS
  	if (sigsetjmp(env, 1) == 0)
  	  {
  #endif
  	    status = ntp_gettime(&ntv);
  	    if ((status < 0) && (errno == ENOSYS))
  	      {
  		--pll_control;
  	      }
- #ifdef SIGSYS
- 	  }
- #endif
  	if (pll_control < 0)
  	  break;
  	times[c] = ntv.time.tv_usec;
        }
      if (pll_control >= 0) {
        printf("[ us %06d:", times[0]);
        for (c = 1; c < sizeof times / sizeof times[0]; c++)
--- 197,221 ----
  #endif /* BADCALL */
  
    if (cost) {
  #ifdef SIGSYS
       if (sigsetjmp(env, 1) == 0)
         {
  #endif
+         for (c = 0; c < sizeof times / sizeof times[0]; c++)
+           {
  	    status = ntp_gettime(&ntv);
  	    if ((status < 0) && (errno == ENOSYS))
  	      {
  		--pll_control;
  	      }
  	if (pll_control < 0)
  	  break;
  	times[c] = ntv.time.tv_usec;
        }
+ #ifdef SIGSYS
+       }
+ #endif
+ 
      if (pll_control >= 0) {
        printf("[ us %06d:", times[0]);
        for (c = 1; c < sizeof times / sizeof times[0]; c++)
>Audit-Trail:
>Unformatted: