xref: /freebsd/sys/vm/vm_glue.c (revision 77b7cdf1999ee965ad494fddd184b18f532ac91a)
1 /*
2  * Copyright (c) 1991, 1993
3  *	The Regents of the University of California.  All rights reserved.
4  *
5  * This code is derived from software contributed to Berkeley by
6  * The Mach Operating System project at Carnegie-Mellon University.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 3. All advertising materials mentioning features or use of this software
17  *    must display the following acknowledgement:
18  *	This product includes software developed by the University of
19  *	California, Berkeley and its contributors.
20  * 4. Neither the name of the University nor the names of its contributors
21  *    may be used to endorse or promote products derived from this software
22  *    without specific prior written permission.
23  *
24  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34  * SUCH DAMAGE.
35  *
36  *	from: @(#)vm_glue.c	8.6 (Berkeley) 1/5/94
37  *
38  *
39  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
40  * All rights reserved.
41  *
42  * Permission to use, copy, modify and distribute this software and
43  * its documentation is hereby granted, provided that both the copyright
44  * notice and this permission notice appear in all copies of the
45  * software, derivative works or modified versions, and any portions
46  * thereof, and that both notices appear in supporting documentation.
47  *
48  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
49  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
50  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
51  *
52  * Carnegie Mellon requests users of this software to return to
53  *
54  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
55  *  School of Computer Science
56  *  Carnegie Mellon University
57  *  Pittsburgh PA 15213-3890
58  *
59  * any improvements or extensions that they make and grant Carnegie the
60  * rights to redistribute these changes.
61  *
62  * $FreeBSD$
63  */
64 
65 #include "opt_vm.h"
66 
67 #include <sys/param.h>
68 #include <sys/systm.h>
69 #include <sys/limits.h>
70 #include <sys/lock.h>
71 #include <sys/mutex.h>
72 #include <sys/proc.h>
73 #include <sys/resourcevar.h>
74 #include <sys/shm.h>
75 #include <sys/vmmeter.h>
76 #include <sys/sx.h>
77 #include <sys/sysctl.h>
78 
79 #include <sys/kernel.h>
80 #include <sys/ktr.h>
81 #include <sys/unistd.h>
82 
83 #include <vm/vm.h>
84 #include <vm/vm_param.h>
85 #include <vm/pmap.h>
86 #include <vm/vm_map.h>
87 #include <vm/vm_page.h>
88 #include <vm/vm_pageout.h>
89 #include <vm/vm_object.h>
90 #include <vm/vm_kern.h>
91 #include <vm/vm_extern.h>
92 #include <vm/vm_pager.h>
93 #include <vm/swap_pager.h>
94 
95 #include <sys/user.h>
96 
97 extern int maxslp;
98 
99 /*
100  * System initialization
101  *
102  * Note: proc0 from proc.h
103  */
104 static void vm_init_limits(void *);
105 SYSINIT(vm_limits, SI_SUB_VM_CONF, SI_ORDER_FIRST, vm_init_limits, &proc0)
106 
107 /*
108  * THIS MUST BE THE LAST INITIALIZATION ITEM!!!
109  *
110  * Note: run scheduling should be divorced from the vm system.
111  */
112 static void scheduler(void *);
113 SYSINIT(scheduler, SI_SUB_RUN_SCHEDULER, SI_ORDER_FIRST, scheduler, NULL)
114 
115 #ifndef NO_SWAPPING
116 static void swapout(struct proc *);
117 static void vm_proc_swapin(struct proc *p);
118 static void vm_proc_swapout(struct proc *p);
119 #endif
120 
121 /*
122  * MPSAFE
123  *
124  * WARNING!  This code calls vm_map_check_protection() which only checks
125  * the associated vm_map_entry range.  It does not determine whether the
126  * contents of the memory is actually readable or writable.  In most cases
127  * just checking the vm_map_entry is sufficient within the kernel's address
128  * space.
129  */
130 int
131 kernacc(addr, len, rw)
132 	void *addr;
133 	int len, rw;
134 {
135 	boolean_t rv;
136 	vm_offset_t saddr, eaddr;
137 	vm_prot_t prot;
138 
139 	KASSERT((rw & ~VM_PROT_ALL) == 0,
140 	    ("illegal ``rw'' argument to kernacc (%x)\n", rw));
141 	prot = rw;
142 	saddr = trunc_page((vm_offset_t)addr);
143 	eaddr = round_page((vm_offset_t)addr + len);
144 	rv = vm_map_check_protection(kernel_map, saddr, eaddr, prot);
145 	return (rv == TRUE);
146 }
147 
148 /*
149  * MPSAFE
150  *
151  * WARNING!  This code calls vm_map_check_protection() which only checks
152  * the associated vm_map_entry range.  It does not determine whether the
153  * contents of the memory is actually readable or writable.  vmapbuf(),
154  * vm_fault_quick(), or copyin()/copout()/su*()/fu*() functions should be
155  * used in conjuction with this call.
156  */
157 int
158 useracc(addr, len, rw)
159 	void *addr;
160 	int len, rw;
161 {
162 	boolean_t rv;
163 	vm_prot_t prot;
164 	vm_map_t map;
165 
166 	KASSERT((rw & ~VM_PROT_ALL) == 0,
167 	    ("illegal ``rw'' argument to useracc (%x)\n", rw));
168 	prot = rw;
169 	map = &curproc->p_vmspace->vm_map;
170 	if ((vm_offset_t)addr + len > vm_map_max(map) ||
171 	    (vm_offset_t)addr + len < (vm_offset_t)addr) {
172 		return (FALSE);
173 	}
174 	rv = vm_map_check_protection(map, trunc_page((vm_offset_t)addr),
175 	    round_page((vm_offset_t)addr + len), prot);
176 	return (rv == TRUE);
177 }
178 
179 /*
180  * MPSAFE
181  */
182 void
183 vslock(addr, len)
184 	void *addr;
185 	u_int len;
186 {
187 
188 	vm_map_wire(&curproc->p_vmspace->vm_map, trunc_page((vm_offset_t)addr),
189 	    round_page((vm_offset_t)addr + len), FALSE);
190 }
191 
192 /*
193  * MPSAFE
194  */
195 void
196 vsunlock(addr, len)
197 	void *addr;
198 	u_int len;
199 {
200 
201 	vm_map_unwire(&curproc->p_vmspace->vm_map,
202 	    trunc_page((vm_offset_t)addr),
203 	    round_page((vm_offset_t)addr + len), FALSE);
204 }
205 
206 /*
207  * Create the U area for a new process.
208  * This routine directly affects the fork perf for a process.
209  */
210 void
211 vm_proc_new(struct proc *p)
212 {
213 	vm_page_t ma[UAREA_PAGES];
214 	vm_object_t upobj;
215 	vm_offset_t up;
216 	vm_page_t m;
217 	u_int i;
218 
219 	/*
220 	 * Allocate object for the upage.
221 	 */
222 	upobj = vm_object_allocate(OBJT_DEFAULT, UAREA_PAGES);
223 	p->p_upages_obj = upobj;
224 
225 	/*
226 	 * Get a kernel virtual address for the U area for this process.
227 	 */
228 	up = kmem_alloc_nofault(kernel_map, UAREA_PAGES * PAGE_SIZE);
229 	if (up == 0)
230 		panic("vm_proc_new: upage allocation failed");
231 	p->p_uarea = (struct user *)up;
232 
233 	for (i = 0; i < UAREA_PAGES; i++) {
234 		/*
235 		 * Get a uarea page.
236 		 */
237 		m = vm_page_grab(upobj, i,
238 		    VM_ALLOC_NORMAL | VM_ALLOC_RETRY | VM_ALLOC_WIRED);
239 		ma[i] = m;
240 
241 		vm_page_lock_queues();
242 		vm_page_wakeup(m);
243 		vm_page_flag_clear(m, PG_ZERO);
244 		m->valid = VM_PAGE_BITS_ALL;
245 		vm_page_unlock_queues();
246 	}
247 
248 	/*
249 	 * Enter the pages into the kernel address space.
250 	 */
251 	pmap_qenter(up, ma, UAREA_PAGES);
252 }
253 
254 /*
255  * Dispose the U area for a process that has exited.
256  * This routine directly impacts the exit perf of a process.
257  * XXX proc_zone is marked UMA_ZONE_NOFREE, so this should never be called.
258  */
259 void
260 vm_proc_dispose(struct proc *p)
261 {
262 	vm_object_t upobj;
263 	vm_offset_t up;
264 	vm_page_t m;
265 
266 	upobj = p->p_upages_obj;
267 	VM_OBJECT_LOCK(upobj);
268 	if (upobj->resident_page_count != UAREA_PAGES)
269 		panic("vm_proc_dispose: incorrect number of pages in upobj");
270 	vm_page_lock_queues();
271 	while ((m = TAILQ_FIRST(&upobj->memq)) != NULL) {
272 		vm_page_busy(m);
273 		vm_page_unwire(m, 0);
274 		vm_page_free(m);
275 	}
276 	vm_page_unlock_queues();
277 	VM_OBJECT_UNLOCK(upobj);
278 	up = (vm_offset_t)p->p_uarea;
279 	pmap_qremove(up, UAREA_PAGES);
280 	kmem_free(kernel_map, up, UAREA_PAGES * PAGE_SIZE);
281 	vm_object_deallocate(upobj);
282 }
283 
284 #ifndef NO_SWAPPING
285 /*
286  * Allow the U area for a process to be prejudicially paged out.
287  */
288 static void
289 vm_proc_swapout(struct proc *p)
290 {
291 	vm_object_t upobj;
292 	vm_offset_t up;
293 	vm_page_t m;
294 
295 	upobj = p->p_upages_obj;
296 	VM_OBJECT_LOCK(upobj);
297 	if (upobj->resident_page_count != UAREA_PAGES)
298 		panic("vm_proc_dispose: incorrect number of pages in upobj");
299 	vm_page_lock_queues();
300 	TAILQ_FOREACH(m, &upobj->memq, listq) {
301 		vm_page_dirty(m);
302 		vm_page_unwire(m, 0);
303 	}
304 	vm_page_unlock_queues();
305 	VM_OBJECT_UNLOCK(upobj);
306 	up = (vm_offset_t)p->p_uarea;
307 	pmap_qremove(up, UAREA_PAGES);
308 }
309 
310 /*
311  * Bring the U area for a specified process back in.
312  */
313 static void
314 vm_proc_swapin(struct proc *p)
315 {
316 	vm_page_t ma[UAREA_PAGES];
317 	vm_object_t upobj;
318 	vm_offset_t up;
319 	vm_page_t m;
320 	int rv;
321 	int i;
322 
323 	upobj = p->p_upages_obj;
324 	for (i = 0; i < UAREA_PAGES; i++) {
325 		m = vm_page_grab(upobj, i, VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
326 		if (m->valid != VM_PAGE_BITS_ALL) {
327 			rv = vm_pager_get_pages(upobj, &m, 1, 0);
328 			if (rv != VM_PAGER_OK)
329 				panic("vm_proc_swapin: cannot get upage");
330 		}
331 		ma[i] = m;
332 	}
333 	VM_OBJECT_LOCK(upobj);
334 	if (upobj->resident_page_count != UAREA_PAGES)
335 		panic("vm_proc_swapin: lost pages from upobj");
336 	vm_page_lock_queues();
337 	TAILQ_FOREACH(m, &upobj->memq, listq) {
338 		m->valid = VM_PAGE_BITS_ALL;
339 		vm_page_wire(m);
340 		vm_page_wakeup(m);
341 	}
342 	vm_page_unlock_queues();
343 	VM_OBJECT_UNLOCK(upobj);
344 	up = (vm_offset_t)p->p_uarea;
345 	pmap_qenter(up, ma, UAREA_PAGES);
346 }
347 
348 /*
349  * Swap in the UAREAs of all processes swapped out to the given device.
350  * The pages in the UAREA are marked dirty and their swap metadata is freed.
351  */
352 void
353 vm_proc_swapin_all(int devidx)
354 {
355 	struct proc *p;
356 	vm_object_t object;
357 	vm_page_t m;
358 
359 retry:
360 	sx_slock(&allproc_lock);
361 	FOREACH_PROC_IN_SYSTEM(p) {
362 		PROC_LOCK(p);
363 		object = p->p_upages_obj;
364 		if (object != NULL) {
365 			VM_OBJECT_LOCK(object);
366 			if (swap_pager_isswapped(object, devidx)) {
367 				VM_OBJECT_UNLOCK(object);
368 				sx_sunlock(&allproc_lock);
369 				faultin(p);
370 				PROC_UNLOCK(p);
371 				VM_OBJECT_LOCK(object);
372 				vm_page_lock_queues();
373 				TAILQ_FOREACH(m, &object->memq, listq)
374 					vm_page_dirty(m);
375 				vm_page_unlock_queues();
376 				swap_pager_freespace(object, 0,
377 				    object->un_pager.swp.swp_bcount);
378 				VM_OBJECT_UNLOCK(object);
379 				goto retry;
380 			}
381 			VM_OBJECT_UNLOCK(object);
382 		}
383 		PROC_UNLOCK(p);
384 	}
385 	sx_sunlock(&allproc_lock);
386 }
387 #endif
388 
389 /*
390  * Implement fork's actions on an address space.
391  * Here we arrange for the address space to be copied or referenced,
392  * allocate a user struct (pcb and kernel stack), then call the
393  * machine-dependent layer to fill those in and make the new process
394  * ready to run.  The new process is set up so that it returns directly
395  * to user mode to avoid stack copying and relocation problems.
396  */
397 void
398 vm_forkproc(td, p2, td2, flags)
399 	struct thread *td;
400 	struct proc *p2;
401 	struct thread *td2;
402 	int flags;
403 {
404 	struct proc *p1 = td->td_proc;
405 	struct user *up;
406 
407 	GIANT_REQUIRED;
408 
409 	if ((flags & RFPROC) == 0) {
410 		/*
411 		 * Divorce the memory, if it is shared, essentially
412 		 * this changes shared memory amongst threads, into
413 		 * COW locally.
414 		 */
415 		if ((flags & RFMEM) == 0) {
416 			if (p1->p_vmspace->vm_refcnt > 1) {
417 				vmspace_unshare(p1);
418 			}
419 		}
420 		cpu_fork(td, p2, td2, flags);
421 		return;
422 	}
423 
424 	if (flags & RFMEM) {
425 		p2->p_vmspace = p1->p_vmspace;
426 		p1->p_vmspace->vm_refcnt++;
427 	}
428 
429 	while (vm_page_count_severe()) {
430 		VM_WAIT;
431 	}
432 
433 	if ((flags & RFMEM) == 0) {
434 		p2->p_vmspace = vmspace_fork(p1->p_vmspace);
435 
436 		pmap_pinit2(vmspace_pmap(p2->p_vmspace));
437 
438 		if (p1->p_vmspace->vm_shm)
439 			shmfork(p1, p2);
440 	}
441 
442 	/* XXXKSE this is unsatisfactory but should be adequate */
443 	up = p2->p_uarea;
444 
445 	/*
446 	 * p_stats currently points at fields in the user struct
447 	 * but not at &u, instead at p_addr. Copy parts of
448 	 * p_stats; zero the rest of p_stats (statistics).
449 	 *
450 	 * If procsig->ps_refcnt is 1 and p2->p_sigacts is NULL we dont' need
451 	 * to share sigacts, so we use the up->u_sigacts.
452 	 */
453 	p2->p_stats = &up->u_stats;
454 	if (p2->p_sigacts == NULL) {
455 		if (p2->p_procsig->ps_refcnt != 1)
456 			printf ("PID:%d NULL sigacts with refcnt not 1!\n",p2->p_pid);
457 		p2->p_sigacts = &up->u_sigacts;
458 		up->u_sigacts = *p1->p_sigacts;
459 	}
460 
461 	bzero(&up->u_stats.pstat_startzero,
462 	    (unsigned) ((caddr_t) &up->u_stats.pstat_endzero -
463 		(caddr_t) &up->u_stats.pstat_startzero));
464 	bcopy(&p1->p_stats->pstat_startcopy, &up->u_stats.pstat_startcopy,
465 	    ((caddr_t) &up->u_stats.pstat_endcopy -
466 		(caddr_t) &up->u_stats.pstat_startcopy));
467 
468 
469 	/*
470 	 * cpu_fork will copy and update the pcb, set up the kernel stack,
471 	 * and make the child ready to run.
472 	 */
473 	cpu_fork(td, p2, td2, flags);
474 }
475 
476 /*
477  * Called after process has been wait(2)'ed apon and is being reaped.
478  * The idea is to reclaim resources that we could not reclaim while
479  * the process was still executing.
480  */
481 void
482 vm_waitproc(p)
483 	struct proc *p;
484 {
485 
486 	GIANT_REQUIRED;
487 	cpu_wait(p);
488 	vmspace_exitfree(p);		/* and clean-out the vmspace */
489 }
490 
491 /*
492  * Set default limits for VM system.
493  * Called for proc 0, and then inherited by all others.
494  *
495  * XXX should probably act directly on proc0.
496  */
497 static void
498 vm_init_limits(udata)
499 	void *udata;
500 {
501 	struct proc *p = udata;
502 	int rss_limit;
503 
504 	/*
505 	 * Set up the initial limits on process VM. Set the maximum resident
506 	 * set size to be half of (reasonably) available memory.  Since this
507 	 * is a soft limit, it comes into effect only when the system is out
508 	 * of memory - half of main memory helps to favor smaller processes,
509 	 * and reduces thrashing of the object cache.
510 	 */
511 	p->p_rlimit[RLIMIT_STACK].rlim_cur = dflssiz;
512 	p->p_rlimit[RLIMIT_STACK].rlim_max = maxssiz;
513 	p->p_rlimit[RLIMIT_DATA].rlim_cur = dfldsiz;
514 	p->p_rlimit[RLIMIT_DATA].rlim_max = maxdsiz;
515 	/* limit the limit to no less than 2MB */
516 	rss_limit = max(cnt.v_free_count, 512);
517 	p->p_rlimit[RLIMIT_RSS].rlim_cur = ptoa(rss_limit);
518 	p->p_rlimit[RLIMIT_RSS].rlim_max = RLIM_INFINITY;
519 }
520 
521 void
522 faultin(p)
523 	struct proc *p;
524 {
525 #ifdef NO_SWAPPING
526 
527 	PROC_LOCK_ASSERT(p, MA_OWNED);
528 	if ((p->p_sflag & PS_INMEM) == 0)
529 		panic("faultin: proc swapped out with NO_SWAPPING!");
530 #else /* !NO_SWAPPING */
531 	struct thread *td;
532 
533 	GIANT_REQUIRED;
534 	PROC_LOCK_ASSERT(p, MA_OWNED);
535 	/*
536 	 * If another process is swapping in this process,
537 	 * just wait until it finishes.
538 	 */
539 	if (p->p_sflag & PS_SWAPPINGIN)
540 		msleep(&p->p_sflag, &p->p_mtx, PVM, "faultin", 0);
541 	else if ((p->p_sflag & PS_INMEM) == 0) {
542 		/*
543 		 * Don't let another thread swap process p out while we are
544 		 * busy swapping it in.
545 		 */
546 		++p->p_lock;
547 		mtx_lock_spin(&sched_lock);
548 		p->p_sflag |= PS_SWAPPINGIN;
549 		mtx_unlock_spin(&sched_lock);
550 		PROC_UNLOCK(p);
551 
552 		vm_proc_swapin(p);
553 		FOREACH_THREAD_IN_PROC(p, td)
554 			pmap_swapin_thread(td);
555 
556 		PROC_LOCK(p);
557 		mtx_lock_spin(&sched_lock);
558 		p->p_sflag &= ~PS_SWAPPINGIN;
559 		p->p_sflag |= PS_INMEM;
560 		FOREACH_THREAD_IN_PROC(p, td) {
561 			TD_CLR_SWAPPED(td);
562 			if (TD_CAN_RUN(td))
563 				setrunnable(td);
564 		}
565 		mtx_unlock_spin(&sched_lock);
566 
567 		wakeup(&p->p_sflag);
568 
569 		/* Allow other threads to swap p out now. */
570 		--p->p_lock;
571 	}
572 #endif /* NO_SWAPPING */
573 }
574 
575 /*
576  * This swapin algorithm attempts to swap-in processes only if there
577  * is enough space for them.  Of course, if a process waits for a long
578  * time, it will be swapped in anyway.
579  *
580  *  XXXKSE - process with the thread with highest priority counts..
581  *
582  * Giant is still held at this point, to be released in tsleep.
583  */
584 /* ARGSUSED*/
585 static void
586 scheduler(dummy)
587 	void *dummy;
588 {
589 	struct proc *p;
590 	struct thread *td;
591 	int pri;
592 	struct proc *pp;
593 	int ppri;
594 
595 	mtx_assert(&Giant, MA_OWNED | MA_NOTRECURSED);
596 	/* GIANT_REQUIRED */
597 
598 loop:
599 	if (vm_page_count_min()) {
600 		VM_WAIT;
601 		goto loop;
602 	}
603 
604 	pp = NULL;
605 	ppri = INT_MIN;
606 	sx_slock(&allproc_lock);
607 	FOREACH_PROC_IN_SYSTEM(p) {
608 		struct ksegrp *kg;
609 		if (p->p_sflag & (PS_INMEM | PS_SWAPPINGOUT | PS_SWAPPINGIN)) {
610 			continue;
611 		}
612 		mtx_lock_spin(&sched_lock);
613 		FOREACH_THREAD_IN_PROC(p, td) {
614 			/*
615 			 * An otherwise runnable thread of a process
616 			 * swapped out has only the TDI_SWAPPED bit set.
617 			 *
618 			 */
619 			if (td->td_inhibitors == TDI_SWAPPED) {
620 				kg = td->td_ksegrp;
621 				pri = p->p_swtime + kg->kg_slptime;
622 				if ((p->p_sflag & PS_SWAPINREQ) == 0) {
623 					pri -= kg->kg_nice * 8;
624 				}
625 
626 				/*
627 				 * if this ksegrp is higher priority
628 				 * and there is enough space, then select
629 				 * this process instead of the previous
630 				 * selection.
631 				 */
632 				if (pri > ppri) {
633 					pp = p;
634 					ppri = pri;
635 				}
636 			}
637 		}
638 		mtx_unlock_spin(&sched_lock);
639 	}
640 	sx_sunlock(&allproc_lock);
641 
642 	/*
643 	 * Nothing to do, back to sleep.
644 	 */
645 	if ((p = pp) == NULL) {
646 		tsleep(&proc0, PVM, "sched", maxslp * hz / 2);
647 		goto loop;
648 	}
649 	PROC_LOCK(p);
650 
651 	/*
652 	 * Another process may be bringing or may have already
653 	 * brought this process in while we traverse all threads.
654 	 * Or, this process may even be being swapped out again.
655 	 */
656 	if (p->p_sflag & (PS_INMEM | PS_SWAPPINGOUT | PS_SWAPPINGIN)) {
657 		PROC_UNLOCK(p);
658 		goto loop;
659 	}
660 
661 	mtx_lock_spin(&sched_lock);
662 	p->p_sflag &= ~PS_SWAPINREQ;
663 	mtx_unlock_spin(&sched_lock);
664 
665 	/*
666 	 * We would like to bring someone in. (only if there is space).
667 	 * [What checks the space? ]
668 	 */
669 	faultin(p);
670 	PROC_UNLOCK(p);
671 	mtx_lock_spin(&sched_lock);
672 	p->p_swtime = 0;
673 	mtx_unlock_spin(&sched_lock);
674 	goto loop;
675 }
676 
677 #ifndef NO_SWAPPING
678 
679 /*
680  * Swap_idle_threshold1 is the guaranteed swapped in time for a process
681  */
682 static int swap_idle_threshold1 = 2;
683 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold1, CTLFLAG_RW,
684     &swap_idle_threshold1, 0, "Guaranteed swapped in time for a process");
685 
686 /*
687  * Swap_idle_threshold2 is the time that a process can be idle before
688  * it will be swapped out, if idle swapping is enabled.
689  */
690 static int swap_idle_threshold2 = 10;
691 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold2, CTLFLAG_RW,
692     &swap_idle_threshold2, 0, "Time before a process will be swapped out");
693 
694 /*
695  * Swapout is driven by the pageout daemon.  Very simple, we find eligible
696  * procs and unwire their u-areas.  We try to always "swap" at least one
697  * process in case we need the room for a swapin.
698  * If any procs have been sleeping/stopped for at least maxslp seconds,
699  * they are swapped.  Else, we swap the longest-sleeping or stopped process,
700  * if any, otherwise the longest-resident process.
701  */
702 void
703 swapout_procs(action)
704 int action;
705 {
706 	struct proc *p;
707 	struct thread *td;
708 	struct ksegrp *kg;
709 	struct proc *outp, *outp2;
710 	int outpri, outpri2;
711 	int didswap = 0;
712 
713 	GIANT_REQUIRED;
714 
715 	outp = outp2 = NULL;
716 	outpri = outpri2 = INT_MIN;
717 retry:
718 	sx_slock(&allproc_lock);
719 	FOREACH_PROC_IN_SYSTEM(p) {
720 		struct vmspace *vm;
721 		int minslptime = 100000;
722 
723 		/*
724 		 * Watch out for a process in
725 		 * creation.  It may have no
726 		 * address space or lock yet.
727 		 */
728 		mtx_lock_spin(&sched_lock);
729 		if (p->p_state == PRS_NEW) {
730 			mtx_unlock_spin(&sched_lock);
731 			continue;
732 		}
733 		mtx_unlock_spin(&sched_lock);
734 
735 		/*
736 		 * An aio daemon switches its
737 		 * address space while running.
738 		 * Perform a quick check whether
739 		 * a process has P_SYSTEM.
740 		 */
741 		if ((p->p_flag & P_SYSTEM) != 0)
742 			continue;
743 
744 		/*
745 		 * Do not swapout a process that
746 		 * is waiting for VM data
747 		 * structures as there is a possible
748 		 * deadlock.  Test this first as
749 		 * this may block.
750 		 *
751 		 * Lock the map until swapout
752 		 * finishes, or a thread of this
753 		 * process may attempt to alter
754 		 * the map.
755 		 */
756 		PROC_LOCK(p);
757 		vm = p->p_vmspace;
758 		KASSERT(vm != NULL,
759 			("swapout_procs: a process has no address space"));
760 		++vm->vm_refcnt;
761 		PROC_UNLOCK(p);
762 		if (!vm_map_trylock(&vm->vm_map))
763 			goto nextproc1;
764 
765 		PROC_LOCK(p);
766 		if (p->p_lock != 0 ||
767 		    (p->p_flag & (P_STOPPED_SINGLE|P_TRACED|P_SYSTEM|P_WEXIT)
768 		    ) != 0) {
769 			goto nextproc2;
770 		}
771 		/*
772 		 * only aiod changes vmspace, however it will be
773 		 * skipped because of the if statement above checking
774 		 * for P_SYSTEM
775 		 */
776 		if ((p->p_sflag & (PS_INMEM|PS_SWAPPINGOUT|PS_SWAPPINGIN)) != PS_INMEM)
777 			goto nextproc2;
778 
779 		switch (p->p_state) {
780 		default:
781 			/* Don't swap out processes in any sort
782 			 * of 'special' state. */
783 			break;
784 
785 		case PRS_NORMAL:
786 			mtx_lock_spin(&sched_lock);
787 			/*
788 			 * do not swapout a realtime process
789 			 * Check all the thread groups..
790 			 */
791 			FOREACH_KSEGRP_IN_PROC(p, kg) {
792 				if (PRI_IS_REALTIME(kg->kg_pri_class))
793 					goto nextproc;
794 
795 				/*
796 				 * Guarantee swap_idle_threshold1
797 				 * time in memory.
798 				 */
799 				if (kg->kg_slptime < swap_idle_threshold1)
800 					goto nextproc;
801 
802 				/*
803 				 * Do not swapout a process if it is
804 				 * waiting on a critical event of some
805 				 * kind or there is a thread whose
806 				 * pageable memory may be accessed.
807 				 *
808 				 * This could be refined to support
809 				 * swapping out a thread.
810 				 */
811 				FOREACH_THREAD_IN_GROUP(kg, td) {
812 					if ((td->td_priority) < PSOCK ||
813 					    !thread_safetoswapout(td))
814 						goto nextproc;
815 				}
816 				/*
817 				 * If the system is under memory stress,
818 				 * or if we are swapping
819 				 * idle processes >= swap_idle_threshold2,
820 				 * then swap the process out.
821 				 */
822 				if (((action & VM_SWAP_NORMAL) == 0) &&
823 				    (((action & VM_SWAP_IDLE) == 0) ||
824 				    (kg->kg_slptime < swap_idle_threshold2)))
825 					goto nextproc;
826 
827 				if (minslptime > kg->kg_slptime)
828 					minslptime = kg->kg_slptime;
829 			}
830 
831 			/*
832 			 * If the process has been asleep for awhile and had
833 			 * most of its pages taken away already, swap it out.
834 			 */
835 			if ((action & VM_SWAP_NORMAL) ||
836 				((action & VM_SWAP_IDLE) &&
837 				 (minslptime > swap_idle_threshold2))) {
838 				swapout(p);
839 				didswap++;
840 				mtx_unlock_spin(&sched_lock);
841 				PROC_UNLOCK(p);
842 				vm_map_unlock(&vm->vm_map);
843 				vmspace_free(vm);
844 				sx_sunlock(&allproc_lock);
845 				goto retry;
846 			}
847 nextproc:
848 			mtx_unlock_spin(&sched_lock);
849 		}
850 nextproc2:
851 		PROC_UNLOCK(p);
852 		vm_map_unlock(&vm->vm_map);
853 nextproc1:
854 		vmspace_free(vm);
855 		continue;
856 	}
857 	sx_sunlock(&allproc_lock);
858 	/*
859 	 * If we swapped something out, and another process needed memory,
860 	 * then wakeup the sched process.
861 	 */
862 	if (didswap)
863 		wakeup(&proc0);
864 }
865 
866 static void
867 swapout(p)
868 	struct proc *p;
869 {
870 	struct thread *td;
871 
872 	PROC_LOCK_ASSERT(p, MA_OWNED);
873 	mtx_assert(&sched_lock, MA_OWNED | MA_NOTRECURSED);
874 #if defined(SWAP_DEBUG)
875 	printf("swapping out %d\n", p->p_pid);
876 #endif
877 
878 	/*
879 	 * The states of this process and its threads may have changed
880 	 * by now.  Assuming that there is only one pageout daemon thread,
881 	 * this process should still be in memory.
882 	 */
883 	KASSERT((p->p_sflag & (PS_INMEM|PS_SWAPPINGOUT|PS_SWAPPINGIN)) == PS_INMEM,
884 		("swapout: lost a swapout race?"));
885 
886 #if defined(INVARIANTS)
887 	/*
888 	 * Make sure that all threads are safe to be swapped out.
889 	 *
890 	 * Alternatively, we could swap out only safe threads.
891 	 */
892 	FOREACH_THREAD_IN_PROC(p, td) {
893 		KASSERT(thread_safetoswapout(td),
894 			("swapout: there is a thread not safe for swapout"));
895 	}
896 #endif /* INVARIANTS */
897 
898 	++p->p_stats->p_ru.ru_nswap;
899 	/*
900 	 * remember the process resident count
901 	 */
902 	p->p_vmspace->vm_swrss = vmspace_resident_count(p->p_vmspace);
903 
904 	p->p_sflag &= ~PS_INMEM;
905 	p->p_sflag |= PS_SWAPPINGOUT;
906 	PROC_UNLOCK(p);
907 	FOREACH_THREAD_IN_PROC(p, td)
908 		TD_SET_SWAPPED(td);
909 	mtx_unlock_spin(&sched_lock);
910 
911 	vm_proc_swapout(p);
912 	FOREACH_THREAD_IN_PROC(p, td)
913 		pmap_swapout_thread(td);
914 
915 	PROC_LOCK(p);
916 	mtx_lock_spin(&sched_lock);
917 	p->p_sflag &= ~PS_SWAPPINGOUT;
918 	p->p_swtime = 0;
919 }
920 #endif /* !NO_SWAPPING */
921