Index | Thread | Search

From:
Claudio Jeker <cjeker@diehard.n-r-g.com>
Subject:
Re: bgpd: switch to milisecond timestamps
To:
Theo Buehler <tb@theobuehler.org>
Cc:
tech@openbsd.org
Date:
Thu, 13 Feb 2025 10:49:51 +0100

Download raw body.

Thread
On Thu, Feb 13, 2025 at 09:09:04AM +0100, Theo Buehler wrote:
> On Wed, Feb 12, 2025 at 03:27:21PM +0100, Claudio Jeker wrote:
> > This moves bgpd to use milisecond for its internal time (getmonotime())
> > now nothing really needs this right now but it may come in handy at some
> > point. At least it will be possible to have a timer_set_msec() function
> > and run timers with sub second resolution.
> > I think using milisecond is good enough for now.
> 
> While superficially relatively simple and smart, I'm not entirely
> convinced by this approach.
> 
> It feels a bit odd (and dangerous) to store milliseconds in variables of
> type time_t, which is defined to be a type holding seconds. Are you
> planning on changing that in a later pass? Of course C won't help much
> with avoiding mistakes coming from type confusion mixing s and ms, at
> least as long as you keep using an integer type for millis.

We can switch time_t to a different type in a follow up. It is one bit I
dislike as well. Using a 64bit msec type will help with -portable on bad
systems. I may make a typedef for this but that still does not prevent
type confusion that the compile is able to see (unless we wrap it into a
struct but then everything gets harder).
 
> The second concern I have is that by multiplying tv_sec by 1000 you
> assume that CLOCK_MONOTONIC returns a reasonably small value so that the
> time_t won't overflow. That's probably safe enough, at least as long as
> you ensure in portable that time_t is 64 bit, but still, you're making
> an assumption about something unspecified (the absolute value of seconds
> in the monotonic clock).

Yeah, shitty systems are shitty. Not sure if it is worth trying to fix
them up. Now defining and using a 64bit type may help with that.
 
> > bgpctl still remains stupid and only prints seconds. Adjusting that is a
> > bit of a pain since we have various places that want seconds. So it is
> > easier to keep the output the same and ignore the extra precision.
> > We could improve fmt_monotime() to show the msec if the value is less than
> > an hour...
> 
> I think it is fine.

:)
 
> > Should I wrap the magic 1000 into some define?
> 
> Up to you. Not sure that would help much.
 
I think I will introduce some extra helper functions so that most of that
is wrapped inside timer.c

My worry is that at some point the need for microsecond resolution will
show up...

> The above concerns aside, the diff reads fine. Maybe I'm overthinking
> this. I also don't have to offer a simple and clean approach that would
> avoid these concerns.

Let me do another round on this...
 
> > -- 
> > :wq Claudio
> > 
> > Index: bgpctl/bgpctl.c
> > ===================================================================
> > RCS file: /cvs/src/usr.sbin/bgpctl/bgpctl.c,v
> > diff -u -p -r1.315 bgpctl.c
> > --- bgpctl/bgpctl.c	31 Jan 2025 20:07:53 -0000	1.315
> > +++ bgpctl/bgpctl.c	12 Feb 2025 13:54:02 -0000
> > @@ -592,6 +592,7 @@ get_monotime(time_t t)
> >  {
> >  	struct timespec ts;
> >  
> > +	t /= 1000;
> >  	if (t == 0)
> >  		return -1;
> >  	if (clock_gettime(CLOCK_MONOTONIC, &ts) != 0)
> > Index: bgpd/bgpd.c
> > ===================================================================
> > RCS file: /cvs/src/usr.sbin/bgpd/bgpd.c,v
> > diff -u -p -r1.280 bgpd.c
> > --- bgpd/bgpd.c	3 Dec 2024 13:46:53 -0000	1.280
> > +++ bgpd/bgpd.c	12 Feb 2025 13:31:19 -0000
> > @@ -118,7 +118,7 @@ usage(void)
> >  #define PFD_SOCK_ROUTE		3
> >  #define PFD_SOCK_PFKEY		4
> >  #define PFD_CONNECT_START	5
> > -#define MAX_TIMEOUT		3600
> > +#define MAX_TIMEOUT		(3600 * 1000)
> >  
> >  int	 cmd_opts;
> >  
> > @@ -359,7 +359,7 @@ BROKEN	if (pledge("stdio rpath wpath cpa
> >  
> >  		if (timeout < 0 || timeout > MAX_TIMEOUT)
> >  			timeout = MAX_TIMEOUT;
> > -		if (poll(pfd, npfd, timeout * 1000) == -1) {
> > +		if (poll(pfd, npfd, timeout) == -1) {
> >  			if (errno != EINTR) {
> >  				log_warn("poll error");
> >  				quit = 1;
> > Index: bgpd/mrt.c
> > ===================================================================
> > RCS file: /cvs/src/usr.sbin/bgpd/mrt.c,v
> > diff -u -p -r1.124 mrt.c
> > --- bgpd/mrt.c	31 Jan 2025 20:07:18 -0000	1.124
> > +++ bgpd/mrt.c	12 Feb 2025 13:38:43 -0000
> > @@ -416,7 +416,7 @@ mrt_dump_entry_mp(struct mrt *mrt, struc
> >  		goto fail;
> >  	/* originated timestamp */
> >  	if (ibuf_add_n32(h2buf, time(NULL) - (getmonotime() -
> > -	    p->lastchange)) == -1)
> > +	    p->lastchange) / 1000) == -1)
> >  		goto fail;
> >  
> >  	n = prefix_nexthop(p);
> > @@ -578,7 +578,7 @@ mrt_dump_entry(struct mrt *mrt, struct p
> >  		goto fail;
> >  	/* originated timestamp */
> >  	if (ibuf_add_n32(hbuf, time(NULL) - (getmonotime() -
> > -	    p->lastchange)) == -1)
> > +	    p->lastchange) / 1000) == -1)
> >  		goto fail;
> >  	switch (p->pt->aid) {
> >  	case AID_INET:
> > @@ -654,7 +654,7 @@ mrt_dump_entry_v2_rib(struct rib_entry *
> >  			goto fail;
> >  		/* originated timestamp */
> >  		if (ibuf_add_n32(buf, time(NULL) - (getmonotime() -
> > -		    p->lastchange)) == -1)
> > +		    p->lastchange) / 1000) == -1)
> >  			goto fail;
> >  
> >  		/* RFC8050: path-id if add-path is used */
> > Index: bgpd/rde_peer.c
> > ===================================================================
> > RCS file: /cvs/src/usr.sbin/bgpd/rde_peer.c,v
> > diff -u -p -r1.46 rde_peer.c
> > --- bgpd/rde_peer.c	27 Jan 2025 15:22:11 -0000	1.46
> > +++ bgpd/rde_peer.c	12 Feb 2025 13:36:56 -0000
> > @@ -509,8 +509,10 @@ peer_stale(struct rde_peer *peer, uint8_
> >  		peer_flush(peer, aid, 0);
> >  
> >  	/* make sure new prefixes start on a higher timestamp */
> > -	while (now >= getmonotime())
> > -		sleep(1);
> > +	while (now >= getmonotime()) {
> > +		struct timespec ts = { .tv_nsec = 1000 * 1000 };
> > +		nanosleep(&ts, NULL);
> > +	}
> >  }
> >  
> >  /*
> > @@ -628,8 +630,10 @@ peer_begin_rrefresh(struct rde_peer *pee
> >  	peer->staletime[aid] = now = getmonotime();
> >  
> >  	/* make sure new prefixes start on a higher timestamp */
> > -	while (now >= getmonotime())
> > -		sleep(1);
> > +	while (now >= getmonotime()) {
> > +		struct timespec ts = { .tv_nsec = 1000 * 1000 };
> > +		nanosleep(&ts, NULL);
> > +	}
> >  }
> >  
> >  void
> > Index: bgpd/rtr.c
> > ===================================================================
> > RCS file: /cvs/src/usr.sbin/bgpd/rtr.c,v
> > diff -u -p -r1.29 rtr.c
> > --- bgpd/rtr.c	2 Dec 2024 15:13:57 -0000	1.29
> > +++ bgpd/rtr.c	12 Feb 2025 13:29:35 -0000
> > @@ -254,7 +254,7 @@ rtr_main(int debug, int verbose)
> >  		i = PFD_PIPE_COUNT;
> >  		i += rtr_poll_events(pfd + i, pfd_elms - i, &timeout);
> >  
> > -		if (poll(pfd, i, timeout * 1000) == -1) {
> > +		if (poll(pfd, i, timeout) == -1) {
> >  			if (errno == EINTR)
> >  				continue;
> >  			fatal("poll error");
> > Index: bgpd/session.c
> > ===================================================================
> > RCS file: /cvs/src/usr.sbin/bgpd/session.c,v
> > diff -u -p -r1.514 session.c
> > --- bgpd/session.c	12 Feb 2025 13:10:13 -0000	1.514
> > +++ bgpd/session.c	12 Feb 2025 14:19:13 -0000
> > @@ -54,6 +54,9 @@
> >  #define PFD_SOCK_RCTL		4
> >  #define PFD_LISTENERS_START	5
> >  
> > +#define MAX_TIMEOUT		(240 * 1000)
> > +#define PAUSEACCEPT_TIMEOUT	(1 * 1000)
> > +
> >  void	session_sighdlr(int);
> >  int	setup_listeners(u_int *);
> >  void	init_peer(struct peer *);
> > @@ -359,7 +362,7 @@ session_main(int debug, int verbose)
> >  			i++;
> >  		}
> >  		idx_listeners = i;
> > -		timeout = 240;	/* loop every 240s at least */
> > +		timeout = MAX_TIMEOUT;
> >  
> >  		now = getmonotime();
> >  		RB_FOREACH(p, peer_head, &conf->peers) {
> > @@ -460,11 +463,11 @@ session_main(int debug, int verbose)
> >  		if (i > pfd_elms)
> >  			fatalx("poll pfd overflow");
> >  
> > -		if (pauseaccept && timeout > 1)
> > -			timeout = 1;
> > +		if (pauseaccept && timeout > PAUSEACCEPT_TIMEOUT)
> > +			timeout = PAUSEACCEPT_TIMEOUT;
> >  		if (timeout < 0)
> >  			timeout = 0;
> > -		if (poll(pfd, i, timeout * 1000) == -1) {
> > +		if (poll(pfd, i, timeout) == -1) {
> >  			if (errno == EINTR)
> >  				continue;
> >  			fatal("poll error");
> > @@ -474,7 +477,8 @@ session_main(int debug, int verbose)
> >  		 * If we previously saw fd exhaustion, we stop accept()
> >  		 * for 1 second to throttle the accept() loop.
> >  		 */
> > -		if (pauseaccept && getmonotime() > pauseaccept + 1)
> > +		if (pauseaccept &&
> > +		    getmonotime() > pauseaccept + PAUSEACCEPT_TIMEOUT)
> >  			pauseaccept = 0;
> >  
> >  		if (handle_pollfd(&pfd[PFD_PIPE_MAIN], ibuf_main) == -1) {
> > Index: bgpd/timer.c
> > ===================================================================
> > RCS file: /cvs/src/usr.sbin/bgpd/timer.c,v
> > diff -u -p -r1.19 timer.c
> > --- bgpd/timer.c	11 Dec 2020 12:00:01 -0000	1.19
> > +++ bgpd/timer.c	12 Feb 2025 13:29:50 -0000
> > @@ -33,7 +33,7 @@ getmonotime(void)
> >  	if (clock_gettime(CLOCK_MONOTONIC, &ts) != 0)
> >  		fatal("clock_gettime");
> >  
> > -	return (ts.tv_sec);
> > +	return (ts.tv_sec * 1000 + ts.tv_nsec / (1000 * 1000));
> >  }
> >  
> >  struct timer *
> > @@ -93,12 +93,12 @@ timer_set(struct timer_head *th, enum Ti
> >  			fatal("timer_set");
> >  		t->type = timer;
> >  	} else {
> > -		if (t->val == getmonotime() + (time_t)offset)
> > +		if (t->val == getmonotime() + (time_t)offset * 1000)
> >  			return;
> >  		TAILQ_REMOVE(th, t, entry);
> >  	}
> >  
> > -	t->val = getmonotime() + offset;
> > +	t->val = getmonotime() + (time_t)offset * 1000;
> >  
> >  	TAILQ_FOREACH(next, th, entry)
> >  		if (next->val == 0 || next->val > t->val)
> > 
> 

-- 
:wq Claudio