vm_glue.c

open bsd vm device design

 *	2. If not enough memory, wake the pageout daemon and let it
 *	   clear some space.
 */
void
scheduler()
{
	register struct proc *p;
	register int pri;
	struct proc *pp;
	int ppri;
	vm_offset_t addr;
	vm_size_t size;

loop:
#ifdef DEBUG
	while (!enableswap)
		tsleep((caddr_t)&proc0, PVM, "noswap", 0);
#endif
	pp = NULL;
	ppri = INT_MIN;
	for (p = allproc.lh_first; p != 0; p = p->p_list.le_next) {
		if (p->p_stat == SRUN && (p->p_flag & P_INMEM) == 0) {
			/* XXX should also penalize based on vm_swrss */
			pri = p->p_swtime + p->p_slptime - p->p_nice * 8;
			if (pri > ppri) {
				pp = p;
				ppri = pri;
			}
		}
	}
#ifdef DEBUG
	if (swapdebug & SDB_FOLLOW)
		printf("scheduler: running, procp %x pri %d\n", pp, ppri);
#endif
	/*
	 * Nothing to do, back to sleep
	 */
	if ((p = pp) == NULL) {
		tsleep((caddr_t)&proc0, PVM, "scheduler", 0);
		goto loop;
	}

	/*
	 * We would like to bring someone in.
	 * This part is really bogus cuz we could deadlock on memory
	 * despite our feeble check.
	 * XXX should require at least vm_swrss / 2
	 */
	size = round_page(ctob(UPAGES));
	addr = (vm_offset_t) p->p_addr;
	if (cnt.v_free_count > atop(size)) {
#ifdef DEBUG
		if (swapdebug & SDB_SWAPIN)
			printf("swapin: pid %d(%s)@%x, pri %d free %d\n",
			       p->p_pid, p->p_comm, p->p_addr,
			       ppri, cnt.v_free_count);
#endif
		vm_map_pageable(kernel_map, addr, addr+size, FALSE);
		/*
		 * Some architectures need to be notified when the
		 * user area has moved to new physical page(s) (e.g.
		 * see pmax/pmax/vm_machdep.c).
		 */
		cpu_swapin(p);
		(void) splstatclock();
		if (p->p_stat == SRUN)
			setrunqueue(p);
		p->p_flag |= P_INMEM;
		(void) spl0();
		p->p_swtime = 0;
		goto loop;
	}
	/*
	 * Not enough memory, jab the pageout daemon and wait til the
	 * coast is clear.
	 */
#ifdef DEBUG
	if (swapdebug & SDB_FOLLOW)
		printf("scheduler: no room for pid %d(%s), free %d\n",
		       p->p_pid, p->p_comm, cnt.v_free_count);
#endif
	(void) splhigh();
	VM_WAIT;
	(void) spl0();
#ifdef DEBUG
	if (swapdebug & SDB_FOLLOW)
		printf("scheduler: room again, free %d\n", cnt.v_free_count);
#endif
	goto loop;
}

#define	swappable(p)							\
	(((p)->p_flag &							\
	    (P_SYSTEM | P_INMEM | P_NOSWAP | P_WEXIT | P_PHYSIO)) == P_INMEM)

/*
 * Swapout is driven by the pageout daemon.  Very simple, we find eligible
 * procs and unwire their u-areas.  We try to always "swap" at least one
 * process in case we need the room for a swapin.
 * If any procs have been sleeping/stopped for at least maxslp seconds,
 * they are swapped.  Else, we swap the longest-sleeping or stopped process,
 * if any, otherwise the longest-resident process.
 */
void
swapout_threads()
{
	register struct proc *p;
	struct proc *outp, *outp2;
	int outpri, outpri2;
	int didswap = 0;
	extern int maxslp;

#ifdef DEBUG
	if (!enableswap)
		return;
#endif
	outp = outp2 = NULL;
	outpri = outpri2 = 0;
	for (p = allproc.lh_first; p != 0; p = p->p_list.le_next) {
		if (!swappable(p))
			continue;
		switch (p->p_stat) {
		case SRUN:
			if (p->p_swtime > outpri2) {
				outp2 = p;
				outpri2 = p->p_swtime;
			}
			continue;
			
		case SSLEEP:
		case SSTOP:
			if (p->p_slptime >= maxslp) {
				swapout(p);
				didswap++;
			} else if (p->p_slptime > outpri) {
				outp = p;
				outpri = p->p_slptime;
			}
			continue;
		}
	}
	/*
	 * If we didn't get rid of any real duds, toss out the next most
	 * likely sleeping/stopped or running candidate.  We only do this
	 * if we are real low on memory since we don't gain much by doing
	 * it (UPAGES pages).
	 */
	if (didswap == 0 &&
	    cnt.v_free_count <= atop(round_page(ctob(UPAGES)))) {
		if ((p = outp) == 0)
			p = outp2;
#ifdef DEBUG
		if (swapdebug & SDB_SWAPOUT)
			printf("swapout_threads: no duds, try procp %x\n", p);
#endif
		if (p)
			swapout(p);
	}
}

void
swapout(p)
	register struct proc *p;
{
	vm_offset_t addr;
	vm_size_t size;

#ifdef DEBUG
	if (swapdebug & SDB_SWAPOUT)
		printf("swapout: pid %d(%s)@%x, stat %x pri %d free %d\n",
		       p->p_pid, p->p_comm, p->p_addr, p->p_stat,
		       p->p_slptime, cnt.v_free_count);
#endif
	size = round_page(ctob(UPAGES));
	addr = (vm_offset_t) p->p_addr;
#if defined(hp300) || defined(luna68k)
	/*
	 * Ugh!  u-area is double mapped to a fixed address behind the
	 * back of the VM system and accesses are usually through that
	 * address rather than the per-process address.  Hence reference
	 * and modify information are recorded at the fixed address and
	 * lost at context switch time.  We assume the u-struct and
	 * kernel stack are always accessed/modified and force it to be so.
	 */
	{
		register int i;
		volatile long tmp;

		for (i = 0; i < UPAGES; i++) {
			tmp = *(long *)addr; *(long *)addr = tmp;
			addr += NBPG;
		}
		addr = (vm_offset_t) p->p_addr;
	}
#endif
#ifdef mips
	/*
	 * Be sure to save the floating point coprocessor state before
	 * paging out the u-struct.
	 */
	{
		extern struct proc *machFPCurProcPtr;

		if (p == machFPCurProcPtr) {
			MachSaveCurFPState(p);
			machFPCurProcPtr = (struct proc *)0;
		}
	}
#endif
#ifndef	i386 /* temporary measure till we find spontaineous unwire of kstack */
	vm_map_pageable(kernel_map, addr, addr+size, TRUE);
	pmap_collect(vm_map_pmap(&p->p_vmspace->vm_map));
#endif
	(void) splhigh();
	p->p_flag &= ~P_INMEM;
	if (p->p_stat == SRUN)
		remrq(p);
	(void) spl0();
	p->p_swtime = 0;
}

/*
 * The rest of these routines fake thread handling
 */

void
assert_wait(event, ruptible)
	void *event;
	boolean_t ruptible;
{
#ifdef lint
	ruptible++;
#endif
	curproc->p_thread = event;
}

void
thread_block()
{
	int s = splhigh();

	if (curproc->p_thread)
		tsleep(curproc->p_thread, PVM, "thrd_block", 0);
	splx(s);
}

void
thread_sleep(event, lock, ruptible)
	void *event;
	simple_lock_t lock;
	boolean_t ruptible;
{
	int s = splhigh();

#ifdef lint
	ruptible++;
#endif
	curproc->p_thread = event;
	simple_unlock(lock);
	if (curproc->p_thread)
		tsleep(event, PVM, "thrd_sleep", 0);
	splx(s);
}

void
thread_wakeup(event)
	void *event;
{
	int s = splhigh();

	wakeup(event);
	splx(s);
}

/*
 * DEBUG stuff
 */

int indent = 0;

#include <machine/stdarg.h>		/* see subr_prf.c */

/*ARGSUSED2*/
void
#if __STDC__
iprintf(const char *fmt, ...)
#else
iprintf(fmt /* , va_alist */)
	char *fmt;
	/* va_dcl */
#endif
{
	register int i;
	va_list ap;

	for (i = indent; i >= 8; i -= 8)
		printf("\t");
	while (--i >= 0)
		printf(" ");
	va_start(ap, fmt);
	printf("%r", fmt, ap);
	va_end(ap);
}