Re: [9fans] fossil periodic thread does zero sleep()

[cinap_lenrek@gmx.de]

[-- Attachment #1: Type: text/plain, Size: 1037 bytes --]

This is actualy very interesting. Riped the parts from periodic.c
to play a with the code to see how it reacts to some changes.

The code below reproduces the problem:

sleep()+0x7 /sys/src/libc/9syscall/sleep.s:5
periodicThread(msec=0x3e8)+0xb2 /tmp/a.c:21
	ct=0x47a68e5b
	t=0x47e50e4d
	ts=0x0
main()+0x10 /tmp/a.c:32
_main+0x31 /sys/src/libc/386/main9.s:16


The zerosleeps go away if one uncomments the foo print. It also
goes away if one makes the sleep one milli second longer by
changing ts to ts+1.

I would love if anybody gives a good explaination of this bug
and how to fix it :-)

#include <u.h>
#include <libc.h>

static void
periodicThread(int msec)
{
	double t, ct, ts;

	ct = nsec()*1e-6;
	t = ct + msec;

	for(;;){
		/* skip missed */
		while(t <= ct)
			t += msec;

		ts = t - ct;
		if(ts > 1000)
			ts = 1000;
		sleep(ts);
		ct = nsec()*1e-6;
		if(t <= ct){
//			print("foo!\n");
			t += msec;
		}
	}
}

void
main(int argc, char *argv[])
{
	periodicThread(1000);
}

[-- Attachment #2: Type: message/rfc822, Size: 2176 bytes --]

From: erik quanstrom <quanstro@quanstro.net>
To: 9fans@9fans.net
Subject: Re: [9fans] fossil periodic thread does zero sleep()
Date: Sun, 22 Mar 2009 19:41:24 -0400
Message-ID: <959ad71b3d677ea98c0e6b2ef28db2dd@quanstro.net>

> Can anybody explain this fossil behaviour from the periodic code?

what if 0 < t - ct < 1?

wouldn't it be a good idea to replace the doubles with vlongs?
and replace *1e-6 with /1000000?  that would make 0 < t - ct < 1
impossible.

- erik