xref: /csrg-svn/sys/kern/kern_clock.c (revision 2872)
1*2872Swnj /*	kern_clock.c	4.15	03/02/81	*/
29Sbill 
39Sbill #include "../h/param.h"
49Sbill #include "../h/systm.h"
5329Sbill #include "../h/dk.h"
62768Swnj #include "../h/callout.h"
79Sbill #include "../h/seg.h"
89Sbill #include "../h/dir.h"
99Sbill #include "../h/user.h"
109Sbill #include "../h/proc.h"
119Sbill #include "../h/reg.h"
129Sbill #include "../h/psl.h"
139Sbill #include "../h/vm.h"
149Sbill #include "../h/buf.h"
159Sbill #include "../h/text.h"
16877Sbill #include "../h/vlimit.h"
17877Sbill #include "../h/mtpr.h"
18877Sbill #include "../h/clock.h"
192689Swnj #include "../h/cpu.h"
209Sbill 
211943Swnj #include "dh.h"
221943Swnj #include "dz.h"
231559Sbill 
249Sbill #define	SCHMAG	9/10
259Sbill 
269Sbill 
279Sbill /*
282442Swnj  * Hardclock is called straight from
299Sbill  * the real time clock interrupt.
302442Swnj  * We limit the work we do at real clock interrupt time to:
312442Swnj  *	reloading clock
322442Swnj  *	decrementing time to callouts
332442Swnj  *	recording cpu time usage
342450Swnj  *	modifying priority of current process
352442Swnj  *	arrange for soft clock interrupt
362442Swnj  *	kernel pc profiling
379Sbill  *
382442Swnj  * At softclock interrupt time we:
399Sbill  *	implement callouts
409Sbill  *	maintain date
419Sbill  *	lightning bolt wakeup (every second)
429Sbill  *	alarm clock signals
439Sbill  *	jab the scheduler
442442Swnj  *
452442Swnj  * On the vax softclock interrupts are implemented by
462442Swnj  * software interrupts.  Note that we may have multiple softclock
472442Swnj  * interrupts compressed into one (due to excessive interrupt load),
482442Swnj  * but that hardclock interrupts should never be lost.
499Sbill  */
509Sbill 
512609Swnj /*ARGSUSED*/
522442Swnj hardclock(pc, ps)
532450Swnj 	caddr_t pc;
549Sbill {
552768Swnj 	register struct callout *p1;
569Sbill 	register struct proc *pp;
572442Swnj 	register int s, cpstate;
589Sbill 
599Sbill 	/*
609Sbill 	 * reprime clock
619Sbill 	 */
629Sbill 	clkreld();
639Sbill 
649Sbill 	/*
652442Swnj 	 * update callout times
669Sbill 	 */
679Sbill 	if(callout[0].c_func == NULL)
689Sbill 		goto out;
692442Swnj 	p1 = &callout[0];
702442Swnj 	while(p1->c_time<=0 && p1->c_func!=NULL)
712442Swnj 		p1++;
722442Swnj 	p1->c_time--;
739Sbill out:
74138Sbill 
75138Sbill 	/*
762442Swnj 	 * Maintain iostat and per-process cpu statistics
77138Sbill 	 */
789Sbill 	if (!noproc) {
799Sbill 		s = u.u_procp->p_rssize;
809Sbill 		u.u_vm.vm_idsrss += s;
819Sbill 		if (u.u_procp->p_textp) {
829Sbill 			register int xrss = u.u_procp->p_textp->x_rssize;
839Sbill 
849Sbill 			s += xrss;
859Sbill 			u.u_vm.vm_ixrss += xrss;
869Sbill 		}
879Sbill 		if (s > u.u_vm.vm_maxrss)
889Sbill 			u.u_vm.vm_maxrss = s;
892768Swnj 		if ((u.u_vm.vm_utime+u.u_vm.vm_stime+1)/hz > u.u_limit[LIM_CPU]) {
90375Sbill 			psignal(u.u_procp, SIGXCPU);
91375Sbill 			if (u.u_limit[LIM_CPU] < INFINITY - 5)
92375Sbill 				u.u_limit[LIM_CPU] += 5;
93375Sbill 		}
949Sbill 	}
959Sbill 	if (USERMODE(ps)) {
969Sbill 		u.u_vm.vm_utime++;
979Sbill 		if(u.u_procp->p_nice > NZERO)
98305Sbill 			cpstate = CP_NICE;
99305Sbill 		else
100305Sbill 			cpstate = CP_USER;
1019Sbill 	} else {
102305Sbill 		cpstate = CP_SYS;
1039Sbill 		if (noproc)
104305Sbill 			cpstate = CP_IDLE;
1059Sbill 		else
1069Sbill 			u.u_vm.vm_stime++;
1079Sbill 	}
1081408Sbill 	cp_time[cpstate]++;
1092442Swnj 	for (s = 0; s < DK_NDRIVE; s++)
1102442Swnj 		if (dk_busy&(1<<s))
1112442Swnj 			dk_time[s]++;
1129Sbill 	if (!noproc) {
1139Sbill 		pp = u.u_procp;
1141399Sbill 		pp->p_cpticks++;
1159Sbill 		if(++pp->p_cpu == 0)
1169Sbill 			pp->p_cpu--;
1179Sbill 		if(pp->p_cpu % 16 == 0) {
118125Sbill 			(void) setpri(pp);
1199Sbill 			if (pp->p_pri >= PUSER)
1209Sbill 				pp->p_pri = pp->p_usrpri;
1219Sbill 		}
1229Sbill 	}
1239Sbill 	++lbolt;
1242689Swnj #if VAX780
125*2872Swnj 	if (cpu == VAX_780 && panicstr == 0 && !BASEPRI(ps))
1262442Swnj 		unhang();
1272442Swnj #endif
1282442Swnj 	setsoftclock();
1292442Swnj }
1302442Swnj 
1312442Swnj /*
1322442Swnj  * Constant for decay filter for cpu usage.
1332442Swnj  */
1342442Swnj double	ccpu = 0.95122942450071400909;		/* exp(-1/20) */
1352442Swnj 
1362442Swnj /*
1372442Swnj  * Software clock interrupt.
1382442Swnj  * This routine is blocked by spl1(),
1392442Swnj  * which doesn't block device interrupts!
1402442Swnj  */
1412609Swnj /*ARGSUSED*/
1422442Swnj softclock(pc, ps)
1432450Swnj 	caddr_t pc;
1442442Swnj {
1452768Swnj 	register struct callout *p1, *p2;
1462442Swnj 	register struct proc *pp;
1472442Swnj 	register int a, s;
1482442Swnj 
1492442Swnj 	/*
150*2872Swnj 	 * Perform callouts (but not after panic's!)
1512442Swnj 	 */
152*2872Swnj 	if (panicstr == 0 && callout[0].c_time <= 0) {
1532442Swnj 		p1 = &callout[0];
154*2872Swnj 		while (p1->c_func != 0 && p1->c_time <= 0) {
1552442Swnj 			(*p1->c_func)(p1->c_arg);
1562442Swnj 			p1++;
1572442Swnj 		}
1582442Swnj 		p2 = &callout[0];
159*2872Swnj 		while (p2->c_func = p1->c_func) {
1602442Swnj 			p2->c_time = p1->c_time;
1612442Swnj 			p2->c_arg = p1->c_arg;
1622442Swnj 			p1++;
1632442Swnj 			p2++;
1642442Swnj 		}
1652442Swnj 	}
1662442Swnj 
1672442Swnj 	/*
1682442Swnj 	 * Drain silos.
1692442Swnj 	 */
1702647Swnj #if NDH > 0
1712442Swnj 	s = spl5(); dhtimer(); splx(s);
1722442Swnj #endif
1732647Swnj #if NDZ > 0
1742442Swnj 	s = spl5(); dztimer(); splx(s);
1752442Swnj #endif
1762442Swnj 
1772442Swnj 	/*
1782450Swnj 	 * If idling and processes are waiting to swap in,
1792450Swnj 	 * check on them.
1802450Swnj 	 */
1812450Swnj 	if (noproc && runin) {
1822450Swnj 		runin = 0;
1832450Swnj 		wakeup((caddr_t)&runin);
1842450Swnj 	}
1852450Swnj 
1862450Swnj 	/*
1872442Swnj 	 * Run paging daemon and reschedule every 1/4 sec.
1882442Swnj 	 */
1892768Swnj 	if (lbolt % (hz/4) == 0) {
1909Sbill 		vmpago();
1919Sbill 		runrun++;
1922442Swnj 		aston();
1939Sbill 	}
1942442Swnj 
1952442Swnj 	/*
1962442Swnj 	 * Lightning bolt every second:
1972442Swnj 	 *	sleep timeouts
1982442Swnj 	 *	process priority recomputation
1992442Swnj 	 *	process %cpu averaging
2002442Swnj 	 *	virtual memory metering
2012442Swnj 	 *	kick swapper if processes want in
2022442Swnj 	 */
2032768Swnj 	if (lbolt >= hz) {
204*2872Swnj 		/*
205*2872Swnj 		 * This doesn't mean much since we run at
206*2872Swnj 		 * software interrupt time... if hardclock()
207*2872Swnj 		 * calls softclock() directly, it prevents
208*2872Swnj 		 * this code from running when the priority
209*2872Swnj 		 * was raised when the clock interrupt occurred.
210*2872Swnj 		 */
2119Sbill 		if (BASEPRI(ps))
2129Sbill 			return;
213*2872Swnj 
214*2872Swnj 		/*
215*2872Swnj 		 * If we didn't run a few times because of
216*2872Swnj 		 * long blockage at high ipl, we don't
217*2872Swnj 		 * really want to run this code several times,
218*2872Swnj 		 * so squish out all multiples of hz here.
219*2872Swnj 		 */
220*2872Swnj 		time += lbolt / hz;
221*2872Swnj 		lbolt %= hz;
222*2872Swnj 
223*2872Swnj 		/*
224*2872Swnj 		 * Wakeup lightning bolt sleepers.
225*2872Swnj 		 * Processes sleep on lbolt to wait
226*2872Swnj 		 * for short amounts of time (e.g. 1 second).
227*2872Swnj 		 */
2289Sbill 		wakeup((caddr_t)&lbolt);
229*2872Swnj 
230*2872Swnj 		/*
231*2872Swnj 		 * Recompute process priority and process
232*2872Swnj 		 * sleep() system calls as well as internal
233*2872Swnj 		 * sleeps with timeouts (tsleep() kernel routine).
234*2872Swnj 		 */
235*2872Swnj 		for (pp = proc; pp < procNPROC; pp++)
236928Sbill 		if (pp->p_stat && pp->p_stat!=SZOMB) {
237*2872Swnj 			/*
238*2872Swnj 			 * Increase resident time, to max of 127 seconds
239*2872Swnj 			 * (it is kept in a character.)  For
240*2872Swnj 			 * loaded processes this is time in core; for
241*2872Swnj 			 * swapped processes, this is time on drum.
242*2872Swnj 			 */
243*2872Swnj 			if (pp->p_time != 127)
2449Sbill 				pp->p_time++;
245*2872Swnj 			/*
246*2872Swnj 			 * If process has clock counting down, and it
247*2872Swnj 			 * expires, set it running (if this is a tsleep()),
248*2872Swnj 			 * or give it an SIGALRM (if the user process
249*2872Swnj 			 * is using alarm signals.
250*2872Swnj 			 */
251*2872Swnj 			if (pp->p_clktim && --pp->p_clktim == 0)
252*2872Swnj 				if (pp->p_flag & STIMO) {
253*2872Swnj 					s = spl6();
254*2872Swnj 					switch (pp->p_stat) {
255204Sbill 
256*2872Swnj 					case SSLEEP:
257*2872Swnj 						setrun(pp);
258*2872Swnj 						break;
259204Sbill 
260*2872Swnj 					case SSTOP:
261*2872Swnj 						unsleep(pp);
262*2872Swnj 						break;
263*2872Swnj 					}
264*2872Swnj 					pp->p_flag &= ~STIMO;
265*2872Swnj 					splx(s);
266*2872Swnj 				} else
267*2872Swnj 					psignal(pp, SIGALRM);
268*2872Swnj 			/*
269*2872Swnj 			 * If process is blocked, increment computed
270*2872Swnj 			 * time blocked.  This is used in swap scheduling.
271*2872Swnj 			 */
272*2872Swnj 			if (pp->p_stat==SSLEEP || pp->p_stat==SSTOP)
2739Sbill 				if (pp->p_slptime != 127)
2749Sbill 					pp->p_slptime++;
275*2872Swnj 			/*
276*2872Swnj 			 * Update digital filter estimation of process
277*2872Swnj 			 * cpu utilization for loaded processes.
278*2872Swnj 			 */
2791399Sbill 			if (pp->p_flag&SLOAD)
2801399Sbill 				pp->p_pctcpu = ccpu * pp->p_pctcpu +
2812768Swnj 				    (1.0 - ccpu) * (pp->p_cpticks/(float)hz);
282*2872Swnj 			/*
283*2872Swnj 			 * Recompute process priority.  The number p_cpu
284*2872Swnj 			 * is a weighted estimate of cpu time consumed.
285*2872Swnj 			 * A process which consumes cpu time has this
286*2872Swnj 			 * increase regularly.  We here decrease it by
287*2872Swnj 			 * a fraction (SCHMAG is 90%), giving a digital
288*2872Swnj 			 * decay filter which damps out in about 10 seconds.
289*2872Swnj 			 *
290*2872Swnj 			 * If a process is niced, then the nice directly
291*2872Swnj 			 * affects the new priority.  The final priority
292*2872Swnj 			 * is in the range 0 to 255, to fit in a character.
293*2872Swnj 			 */
2941399Sbill 			pp->p_cpticks = 0;
2959Sbill 			a = (pp->p_cpu & 0377)*SCHMAG + pp->p_nice - NZERO;
296*2872Swnj 			if (a < 0)
2979Sbill 				a = 0;
298*2872Swnj 			if (a > 255)
2999Sbill 				a = 255;
3009Sbill 			pp->p_cpu = a;
301125Sbill 			(void) setpri(pp);
302*2872Swnj 			/*
303*2872Swnj 			 * Now have computed new process priority
304*2872Swnj 			 * in p->p_usrpri.  Carefully change p->p_pri.
305*2872Swnj 			 * A process is on a run queue associated with
306*2872Swnj 			 * this priority, so we must block out process
307*2872Swnj 			 * state changes during the transition.
308*2872Swnj 			 */
3099Sbill 			s = spl6();
310*2872Swnj 			if (pp->p_pri >= PUSER) {
3119Sbill 				if ((pp != u.u_procp || noproc) &&
3129Sbill 				    pp->p_stat == SRUN &&
3139Sbill 				    (pp->p_flag & SLOAD) &&
3149Sbill 				    pp->p_pri != pp->p_usrpri) {
3159Sbill 					remrq(pp);
3169Sbill 					pp->p_pri = pp->p_usrpri;
3179Sbill 					setrq(pp);
3189Sbill 				} else
3199Sbill 					pp->p_pri = pp->p_usrpri;
3209Sbill 			}
3219Sbill 			splx(s);
3229Sbill 		}
323*2872Swnj 
324*2872Swnj 		/*
325*2872Swnj 		 * Perform virtual memory metering.
326*2872Swnj 		 */
3279Sbill 		vmmeter();
328*2872Swnj 
329*2872Swnj 		/*
330*2872Swnj 		 * If the swap process is trying to bring
331*2872Swnj 		 * a process in, have it look again to see
332*2872Swnj 		 * if it is possible now.
333*2872Swnj 		 */
334*2872Swnj 		if (runin!=0) {
3359Sbill 			runin = 0;
3369Sbill 			wakeup((caddr_t)&runin);
3379Sbill 		}
338*2872Swnj 
3399Sbill 		/*
3409Sbill 		 * If there are pages that have been cleaned,
3419Sbill 		 * jolt the pageout daemon to process them.
3429Sbill 		 * We do this here so that these pages will be
3439Sbill 		 * freed if there is an abundance of memory and the
3449Sbill 		 * daemon would not be awakened otherwise.
3459Sbill 		 */
3469Sbill 		if (bclnlist != NULL)
3479Sbill 			wakeup((caddr_t)&proc[2]);
348*2872Swnj 
349*2872Swnj 		/*
350*2872Swnj 		 * If the trap occurred from usermode,
351*2872Swnj 		 * then check to see if it has now been
352*2872Swnj 		 * running more than 10 minutes of user time
353*2872Swnj 		 * and should thus run with reduced priority
354*2872Swnj 		 * to give other processes a chance.
355*2872Swnj 		 */
3569Sbill 		if (USERMODE(ps)) {
3579Sbill 			pp = u.u_procp;
358*2872Swnj 			if (pp->p_uid && pp->p_nice == NZERO &&
359*2872Swnj 			    u.u_vm.vm_utime > 600 * hz)
360*2872Swnj 				pp->p_nice = NZERO+4;
361125Sbill 			(void) setpri(pp);
3629Sbill 			pp->p_pri = pp->p_usrpri;
3639Sbill 		}
3649Sbill 	}
365*2872Swnj 	/*
366*2872Swnj 	 * If trapped user-mode, give it a profiling tick.
367*2872Swnj 	 */
3682442Swnj 	if (USERMODE(ps) && u.u_prof.pr_scale) {
3692442Swnj 		u.u_procp->p_flag |= SOWEUPC;
3702442Swnj 		aston();
3719Sbill 	}
3729Sbill }
3739Sbill 
3749Sbill /*
3759Sbill  * timeout is called to arrange that
3762768Swnj  * fun(arg) is called in tim/hz seconds.
3779Sbill  * An entry is sorted into the callout
3789Sbill  * structure. The time in each structure
3792768Swnj  * entry is the number of hz's more
3809Sbill  * than the previous entry.
3819Sbill  * In this way, decrementing the
3829Sbill  * first entry has the effect of
3839Sbill  * updating all entries.
3849Sbill  *
3859Sbill  * The panic is there because there is nothing
3869Sbill  * intelligent to be done if an entry won't fit.
3879Sbill  */
3889Sbill timeout(fun, arg, tim)
3892450Swnj 	int (*fun)();
3902450Swnj 	caddr_t arg;
3919Sbill {
3922768Swnj 	register struct callout *p1, *p2, *p3;
3939Sbill 	register int t;
3949Sbill 	int s;
3959Sbill 
3969Sbill 	t = tim;
3979Sbill 	p1 = &callout[0];
3989Sbill 	s = spl7();
3999Sbill 	while(p1->c_func != 0 && p1->c_time <= t) {
4009Sbill 		t -= p1->c_time;
4019Sbill 		p1++;
4029Sbill 	}
4039Sbill 	p1->c_time -= t;
4049Sbill 	p2 = p1;
4052768Swnj 	p3 = callout+(ncallout-2);
4062430Swnj 	while(p2->c_func != 0) {
4072430Swnj 		if (p2 >= p3)
4082442Swnj 			panic("timeout");
4099Sbill 		p2++;
4102430Swnj 	}
4119Sbill 	while(p2 >= p1) {
4129Sbill 		(p2+1)->c_time = p2->c_time;
4139Sbill 		(p2+1)->c_func = p2->c_func;
4149Sbill 		(p2+1)->c_arg = p2->c_arg;
4159Sbill 		p2--;
4169Sbill 	}
4179Sbill 	p1->c_time = t;
4189Sbill 	p1->c_func = fun;
4199Sbill 	p1->c_arg = arg;
4209Sbill 	splx(s);
4219Sbill }
422