[B.A.T.M.A.N.] times wrapper

[Nathan Wharton <naw@greptar.com>]

[-- Attachment #1: Type: text/plain, Size: 904 bytes --]

When I first started meshing, I used olsrd.  There was a problem on
our platform that caused olsrd packets to stop going out after the
board had been up about 5 minutes.  The outage lasted about 40
seconds, then came back.  It turns out that there is a problem with
the times() call that will return -1 for the 4096 jiffies before wrap
around.  This caused no apparent change in time during that duration,
so no packets were sent.  Also, for some reason on our platform,
times() starts out about 5 minutes before wrap around.

I found a work around elsewhere (and I leave credit in the patch) for
some other software that was failing every 400 something days.  I
applied this to olsrd, and the problem went away.

When I switched to batman, I saw that times() is used in it as well,
so I made the patch attached.  So, if anyone sees packets fail to come
from batman for 40 seconds, they might need this.

[-- Attachment #2: 001-times.patch --]
[-- Type: application/octet-stream, Size: 2359 bytes --]

diff -ruN batmand-r1178.orig/posix/posix.c batmand-r1178/posix/posix.c
--- batmand-r1178.orig/posix/posix.c	2008-12-29 09:35:38.000000000 -0600
+++ batmand-r1178/posix/posix.c	2009-01-02 13:31:27.000000000 -0600
@@ -49,11 +49,49 @@
 
 static pthread_mutex_t batman_clock_mutex = PTHREAD_MUTEX_INITIALIZER;
 
+static struct tms dummy_tms_struct;
+static clock_t
+times_wrapper(void) /* Make times(2) behave rationally on Linux */
+{
+       int             save_errno = errno;
+       clock_t ret;
+
+       /*
+        * times(2) really returns an unsigned value ...
+        *
+        * We don't check to see if we got back the error value (-1), because
+        * the only possibility for an error would be if the address of
+        * dummy_tms_struct was invalid.  Since it's a
+        * compiler-generated address, we assume that errors are impossible.
+        * And, unfortunately, it is quite possible for the correct return
+        * from times(2) to be exactly (clock_t)-1.  Sigh...
+        *
+        */
+       errno   = 0;
+       ret     = times(&dummy_tms_struct);
+
+/*
+ *     This is to work around a bug in the system call interface
+ *     for times(2) found in glibc on Linux (and maybe elsewhere)
+ *     It changes the return values from -1 to -4096 all into
+ *     -1 and then dumps the -(return value) into errno.
+ *
+ *     This totally bizarre behavior seems to be widespread in
+ *     versions of Linux and glibc.
+ *
+ *     Many thanks to Wolfgang Dumhs <wolfgang.dumhs (at) gmx.at>
+ *     for finding and documenting this bizarre behavior.
+ */
+       if (errno != 0) {
+               ret = (clock_t) (-errno);
+       }
+       errno = save_errno;
+       return ret;
+}
 
 static void update_internal_clock(void)
 {
-	struct tms tp;
-	clock_t current_clock_tick = times(&tp);
+	clock_t current_clock_tick = times_wrapper();
 
 	batman_clock_ticks += (current_clock_tick - last_clock_tick);
 	last_clock_tick = current_clock_tick;
@@ -553,7 +591,6 @@
 int main( int argc, char *argv[] ) {
 
 	int8_t res;
-	struct tms tp;
 
 	/* check if user is root */
 	if ( ( getuid() ) || ( getgid() ) ) {
@@ -575,7 +612,7 @@
 
 	/* save start value */
 	system_tick = (float)sysconf(_SC_CLK_TCK);
-	last_clock_tick = times(&tp);
+	last_clock_tick = times_wrapper();
 	update_internal_clock();
 
 	apply_init_args( argc, argv );