xref: /freebsd/sys/vm/vm_glue.c (revision 5521ff5a4d1929056e7ffc982fac3341ca54df7c)
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_rlimit.h"
66 #include "opt_vm.h"
67 
68 #include <sys/param.h>
69 #include <sys/systm.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 <machine/limits.h>
84 
85 #include <vm/vm.h>
86 #include <vm/vm_param.h>
87 #include <vm/pmap.h>
88 #include <vm/vm_map.h>
89 #include <vm/vm_page.h>
90 #include <vm/vm_pageout.h>
91 #include <vm/vm_kern.h>
92 #include <vm/vm_extern.h>
93 
94 #include <sys/user.h>
95 
96 extern int maxslp;
97 
98 /*
99  * System initialization
100  *
101  * Note: proc0 from proc.h
102  */
103 
104 static void vm_init_limits __P((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 __P((void *));
113 SYSINIT(scheduler, SI_SUB_RUN_SCHEDULER, SI_ORDER_FIRST, scheduler, NULL)
114 
115 
116 static void swapout __P((struct proc *));
117 
118 int
119 kernacc(addr, len, rw)
120 	caddr_t addr;
121 	int len, rw;
122 {
123 	boolean_t rv;
124 	vm_offset_t saddr, eaddr;
125 	vm_prot_t prot;
126 
127 	KASSERT((rw & (~VM_PROT_ALL)) == 0,
128 	    ("illegal ``rw'' argument to kernacc (%x)\n", rw));
129 	prot = rw;
130 	saddr = trunc_page((vm_offset_t)addr);
131 	eaddr = round_page((vm_offset_t)addr + len);
132 	vm_map_lock_read(kernel_map);
133 	rv = vm_map_check_protection(kernel_map, saddr, eaddr, prot);
134 	vm_map_unlock_read(kernel_map);
135 	return (rv == TRUE);
136 }
137 
138 int
139 useracc(addr, len, rw)
140 	caddr_t addr;
141 	int len, rw;
142 {
143 	boolean_t rv;
144 	vm_prot_t prot;
145 	vm_map_t map;
146 	vm_map_entry_t save_hint;
147 
148 	GIANT_REQUIRED;
149 
150 	KASSERT((rw & (~VM_PROT_ALL)) == 0,
151 	    ("illegal ``rw'' argument to useracc (%x)\n", rw));
152 	prot = rw;
153 	/*
154 	 * XXX - check separately to disallow access to user area and user
155 	 * page tables - they are in the map.
156 	 *
157 	 * XXX - VM_MAXUSER_ADDRESS is an end address, not a max.  It was once
158 	 * only used (as an end address) in trap.c.  Use it as an end address
159 	 * here too.  This bogusness has spread.  I just fixed where it was
160 	 * used as a max in vm_mmap.c.
161 	 */
162 	if ((vm_offset_t) addr + len > /* XXX */ VM_MAXUSER_ADDRESS
163 	    || (vm_offset_t) addr + len < (vm_offset_t) addr) {
164 		return (FALSE);
165 	}
166 	map = &curproc->p_vmspace->vm_map;
167 	vm_map_lock_read(map);
168 	/*
169 	 * We save the map hint, and restore it.  Useracc appears to distort
170 	 * the map hint unnecessarily.
171 	 */
172 	save_hint = map->hint;
173 	rv = vm_map_check_protection(map,
174 	    trunc_page((vm_offset_t)addr), round_page((vm_offset_t)addr + len), prot);
175 	map->hint = save_hint;
176 	vm_map_unlock_read(map);
177 
178 	return (rv == TRUE);
179 }
180 
181 void
182 vslock(addr, len)
183 	caddr_t addr;
184 	u_int len;
185 {
186 	GIANT_REQUIRED;
187 	vm_map_pageable(&curproc->p_vmspace->vm_map,
188 	    trunc_page((vm_offset_t)addr),
189 	    round_page((vm_offset_t)addr + len), FALSE);
190 }
191 
192 void
193 vsunlock(addr, len)
194 	caddr_t addr;
195 	u_int len;
196 {
197 	GIANT_REQUIRED;
198 	vm_map_pageable(&curproc->p_vmspace->vm_map,
199 	    trunc_page((vm_offset_t)addr),
200 	    round_page((vm_offset_t)addr + len), TRUE);
201 }
202 
203 /*
204  * Implement fork's actions on an address space.
205  * Here we arrange for the address space to be copied or referenced,
206  * allocate a user struct (pcb and kernel stack), then call the
207  * machine-dependent layer to fill those in and make the new process
208  * ready to run.  The new process is set up so that it returns directly
209  * to user mode to avoid stack copying and relocation problems.
210  */
211 void
212 vm_fork(p1, p2, flags)
213 	struct proc *p1, *p2;
214 	int flags;
215 {
216 	struct user *up;
217 
218 	GIANT_REQUIRED;
219 
220 	if ((flags & RFPROC) == 0) {
221 		/*
222 		 * Divorce the memory, if it is shared, essentially
223 		 * this changes shared memory amongst threads, into
224 		 * COW locally.
225 		 */
226 		if ((flags & RFMEM) == 0) {
227 			if (p1->p_vmspace->vm_refcnt > 1) {
228 				vmspace_unshare(p1);
229 			}
230 		}
231 		cpu_fork(p1, p2, flags);
232 		return;
233 	}
234 
235 	if (flags & RFMEM) {
236 		p2->p_vmspace = p1->p_vmspace;
237 		p1->p_vmspace->vm_refcnt++;
238 	}
239 
240 	while (vm_page_count_severe()) {
241 		VM_WAIT;
242 	}
243 
244 	if ((flags & RFMEM) == 0) {
245 		p2->p_vmspace = vmspace_fork(p1->p_vmspace);
246 
247 		pmap_pinit2(vmspace_pmap(p2->p_vmspace));
248 
249 		if (p1->p_vmspace->vm_shm)
250 			shmfork(p1, p2);
251 	}
252 
253 	pmap_new_proc(p2);
254 
255 	up = p2->p_addr;
256 
257 	/*
258 	 * p_stats currently points at fields in the user struct
259 	 * but not at &u, instead at p_addr. Copy parts of
260 	 * p_stats; zero the rest of p_stats (statistics).
261 	 *
262 	 * If procsig->ps_refcnt is 1 and p2->p_sigacts is NULL we dont' need
263 	 * to share sigacts, so we use the up->u_sigacts.
264 	 */
265 	p2->p_stats = &up->u_stats;
266 	if (p2->p_sigacts == NULL) {
267 		if (p2->p_procsig->ps_refcnt != 1)
268 			printf ("PID:%d NULL sigacts with refcnt not 1!\n",p2->p_pid);
269 		p2->p_sigacts = &up->u_sigacts;
270 		up->u_sigacts = *p1->p_sigacts;
271 	}
272 
273 	bzero(&up->u_stats.pstat_startzero,
274 	    (unsigned) ((caddr_t) &up->u_stats.pstat_endzero -
275 		(caddr_t) &up->u_stats.pstat_startzero));
276 	bcopy(&p1->p_stats->pstat_startcopy, &up->u_stats.pstat_startcopy,
277 	    ((caddr_t) &up->u_stats.pstat_endcopy -
278 		(caddr_t) &up->u_stats.pstat_startcopy));
279 
280 
281 	/*
282 	 * cpu_fork will copy and update the pcb, set up the kernel stack,
283 	 * and make the child ready to run.
284 	 */
285 	cpu_fork(p1, p2, flags);
286 }
287 
288 /*
289  * Set default limits for VM system.
290  * Called for proc 0, and then inherited by all others.
291  *
292  * XXX should probably act directly on proc0.
293  */
294 static void
295 vm_init_limits(udata)
296 	void *udata;
297 {
298 	struct proc *p = udata;
299 	int rss_limit;
300 
301 	/*
302 	 * Set up the initial limits on process VM. Set the maximum resident
303 	 * set size to be half of (reasonably) available memory.  Since this
304 	 * is a soft limit, it comes into effect only when the system is out
305 	 * of memory - half of main memory helps to favor smaller processes,
306 	 * and reduces thrashing of the object cache.
307 	 */
308 	p->p_rlimit[RLIMIT_STACK].rlim_cur = DFLSSIZ;
309 	p->p_rlimit[RLIMIT_STACK].rlim_max = MAXSSIZ;
310 	p->p_rlimit[RLIMIT_DATA].rlim_cur = DFLDSIZ;
311 	p->p_rlimit[RLIMIT_DATA].rlim_max = MAXDSIZ;
312 	/* limit the limit to no less than 2MB */
313 	rss_limit = max(cnt.v_free_count, 512);
314 	p->p_rlimit[RLIMIT_RSS].rlim_cur = ptoa(rss_limit);
315 	p->p_rlimit[RLIMIT_RSS].rlim_max = RLIM_INFINITY;
316 }
317 
318 /*
319  * Must be called with the proc struc mutex held.
320  */
321 void
322 faultin(p)
323 	struct proc *p;
324 {
325 	GIANT_REQUIRED;
326 
327 	PROC_LOCK_ASSERT(p, MA_OWNED);
328 	mtx_lock_spin(&sched_lock);
329 	if ((p->p_sflag & PS_INMEM) == 0) {
330 		++p->p_lock;
331 		mtx_unlock_spin(&sched_lock);
332 		PROC_UNLOCK(p);
333 
334 		pmap_swapin_proc(p);
335 
336 		PROC_LOCK(p);
337 		mtx_lock_spin(&sched_lock);
338 		if (p->p_stat == SRUN) {
339 			setrunqueue(p);
340 		}
341 
342 		p->p_sflag |= PS_INMEM;
343 
344 		/* undo the effect of setting SLOCK above */
345 		--p->p_lock;
346 	}
347 	mtx_unlock_spin(&sched_lock);
348 }
349 
350 /*
351  * This swapin algorithm attempts to swap-in processes only if there
352  * is enough space for them.  Of course, if a process waits for a long
353  * time, it will be swapped in anyway.
354  *
355  * Giant is still held at this point, to be released in tsleep.
356  */
357 /* ARGSUSED*/
358 static void
359 scheduler(dummy)
360 	void *dummy;
361 {
362 	struct proc *p;
363 	int pri;
364 	struct proc *pp;
365 	int ppri;
366 
367 	mtx_assert(&Giant, MA_OWNED | MA_NOTRECURSED);
368 	/* GIANT_REQUIRED */
369 
370 loop:
371 	if (vm_page_count_min()) {
372 		VM_WAIT;
373 		goto loop;
374 	}
375 
376 	pp = NULL;
377 	ppri = INT_MIN;
378 	sx_slock(&allproc_lock);
379 	LIST_FOREACH(p, &allproc, p_list) {
380 		mtx_lock_spin(&sched_lock);
381 		if (p->p_stat == SRUN &&
382 			(p->p_sflag & (PS_INMEM | PS_SWAPPING)) == 0) {
383 
384 			pri = p->p_swtime + p->p_slptime;
385 			if ((p->p_sflag & PS_SWAPINREQ) == 0) {
386 				pri -= p->p_nice * 8;
387 			}
388 
389 			/*
390 			 * if this process is higher priority and there is
391 			 * enough space, then select this process instead of
392 			 * the previous selection.
393 			 */
394 			if (pri > ppri) {
395 				pp = p;
396 				ppri = pri;
397 			}
398 		}
399 		mtx_unlock_spin(&sched_lock);
400 	}
401 	sx_sunlock(&allproc_lock);
402 
403 	/*
404 	 * Nothing to do, back to sleep.
405 	 */
406 	if ((p = pp) == NULL) {
407 		tsleep(&proc0, PVM, "sched", maxslp * hz / 2);
408 		goto loop;
409 	}
410 	mtx_lock_spin(&sched_lock);
411 	p->p_sflag &= ~PS_SWAPINREQ;
412 	mtx_unlock_spin(&sched_lock);
413 
414 	/*
415 	 * We would like to bring someone in. (only if there is space).
416 	 */
417 	PROC_LOCK(p);
418 	faultin(p);
419 	PROC_UNLOCK(p);
420 	mtx_lock_spin(&sched_lock);
421 	p->p_swtime = 0;
422 	mtx_unlock_spin(&sched_lock);
423 	goto loop;
424 }
425 
426 #ifndef NO_SWAPPING
427 
428 /*
429  * Swap_idle_threshold1 is the guaranteed swapped in time for a process
430  */
431 static int swap_idle_threshold1 = 2;
432 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold1,
433 	CTLFLAG_RW, &swap_idle_threshold1, 0, "");
434 
435 /*
436  * Swap_idle_threshold2 is the time that a process can be idle before
437  * it will be swapped out, if idle swapping is enabled.
438  */
439 static int swap_idle_threshold2 = 10;
440 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold2,
441 	CTLFLAG_RW, &swap_idle_threshold2, 0, "");
442 
443 /*
444  * Swapout is driven by the pageout daemon.  Very simple, we find eligible
445  * procs and unwire their u-areas.  We try to always "swap" at least one
446  * process in case we need the room for a swapin.
447  * If any procs have been sleeping/stopped for at least maxslp seconds,
448  * they are swapped.  Else, we swap the longest-sleeping or stopped process,
449  * if any, otherwise the longest-resident process.
450  */
451 void
452 swapout_procs(action)
453 int action;
454 {
455 	struct proc *p;
456 	struct proc *outp, *outp2;
457 	int outpri, outpri2;
458 	int didswap = 0;
459 
460 	GIANT_REQUIRED;
461 
462 	outp = outp2 = NULL;
463 	outpri = outpri2 = INT_MIN;
464 retry:
465 	sx_slock(&allproc_lock);
466 	LIST_FOREACH(p, &allproc, p_list) {
467 		struct vmspace *vm;
468 
469 		PROC_LOCK(p);
470 		if (p->p_lock != 0 ||
471 		    (p->p_flag & (P_TRACED|P_SYSTEM|P_WEXIT)) != 0) {
472 			PROC_UNLOCK(p);
473 			continue;
474 		}
475 		/*
476 		 * only aiod changes vmspace, however it will be
477 		 * skipped because of the if statement above checking
478 		 * for P_SYSTEM
479 		 */
480 		vm = p->p_vmspace;
481 		mtx_lock_spin(&sched_lock);
482 		if ((p->p_sflag & (PS_INMEM|PS_SWAPPING)) != PS_INMEM) {
483 			mtx_unlock_spin(&sched_lock);
484 			PROC_UNLOCK(p);
485 			continue;
486 		}
487 
488 		switch (p->p_stat) {
489 		default:
490 			mtx_unlock_spin(&sched_lock);
491 			PROC_UNLOCK(p);
492 			continue;
493 
494 		case SSLEEP:
495 		case SSTOP:
496 			/*
497 			 * do not swapout a realtime process
498 			 */
499 			if (PRI_IS_REALTIME(p->p_pri.pri_class)) {
500 				mtx_unlock_spin(&sched_lock);
501 				PROC_UNLOCK(p);
502 				continue;
503 			}
504 
505 			/*
506 			 * Do not swapout a process waiting on a critical
507 			 * event of some kind.  Also guarantee swap_idle_threshold1
508 			 * time in memory.
509 			 */
510 			if (((p->p_pri.pri_level) < PSOCK) ||
511 				(p->p_slptime < swap_idle_threshold1)) {
512 				mtx_unlock_spin(&sched_lock);
513 				PROC_UNLOCK(p);
514 				continue;
515 			}
516 
517 			/*
518 			 * If the system is under memory stress, or if we are swapping
519 			 * idle processes >= swap_idle_threshold2, then swap the process
520 			 * out.
521 			 */
522 			if (((action & VM_SWAP_NORMAL) == 0) &&
523 				(((action & VM_SWAP_IDLE) == 0) ||
524 				  (p->p_slptime < swap_idle_threshold2))) {
525 				mtx_unlock_spin(&sched_lock);
526 				PROC_UNLOCK(p);
527 				continue;
528 			}
529 			mtx_unlock_spin(&sched_lock);
530 
531 			++vm->vm_refcnt;
532 			/*
533 			 * do not swapout a process that is waiting for VM
534 			 * data structures there is a possible deadlock.
535 			 */
536 			if (lockmgr(&vm->vm_map.lock,
537 					LK_EXCLUSIVE | LK_NOWAIT,
538 					NULL, curproc)) {
539 				vmspace_free(vm);
540 				PROC_UNLOCK(p);
541 				continue;
542 			}
543 			vm_map_unlock(&vm->vm_map);
544 			/*
545 			 * If the process has been asleep for awhile and had
546 			 * most of its pages taken away already, swap it out.
547 			 */
548 			if ((action & VM_SWAP_NORMAL) ||
549 				((action & VM_SWAP_IDLE) &&
550 				 (p->p_slptime > swap_idle_threshold2))) {
551 				sx_sunlock(&allproc_lock);
552 				swapout(p);
553 				vmspace_free(vm);
554 				didswap++;
555 				goto retry;
556 			}
557 			PROC_UNLOCK(p);
558 			vmspace_free(vm);
559 		}
560 	}
561 	sx_sunlock(&allproc_lock);
562 	/*
563 	 * If we swapped something out, and another process needed memory,
564 	 * then wakeup the sched process.
565 	 */
566 	if (didswap)
567 		wakeup(&proc0);
568 }
569 
570 static void
571 swapout(p)
572 	struct proc *p;
573 {
574 
575 	PROC_LOCK_ASSERT(p, MA_OWNED);
576 #if defined(SWAP_DEBUG)
577 	printf("swapping out %d\n", p->p_pid);
578 #endif
579 	++p->p_stats->p_ru.ru_nswap;
580 	/*
581 	 * remember the process resident count
582 	 */
583 	p->p_vmspace->vm_swrss = vmspace_resident_count(p->p_vmspace);
584 
585 	mtx_lock_spin(&sched_lock);
586 	p->p_sflag &= ~PS_INMEM;
587 	p->p_sflag |= PS_SWAPPING;
588 	PROC_UNLOCK_NOSWITCH(p);
589 	if (p->p_stat == SRUN)
590 		remrunqueue(p);
591 	mtx_unlock_spin(&sched_lock);
592 
593 	pmap_swapout_proc(p);
594 
595 	mtx_lock_spin(&sched_lock);
596 	p->p_sflag &= ~PS_SWAPPING;
597 	p->p_swtime = 0;
598 	mtx_unlock_spin(&sched_lock);
599 }
600 #endif /* !NO_SWAPPING */
601