xref: /freebsd/sys/vm/vm_glue.c (revision 7f3dea244c40159a41ab22da77a434d7c5b5e85a)
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  * $Id: vm_glue.c,v 1.87 1999/07/21 18:02:26 alc Exp $
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/proc.h>
71 #include <sys/resourcevar.h>
72 #include <sys/buf.h>
73 #include <sys/shm.h>
74 #include <sys/vmmeter.h>
75 #include <sys/sysctl.h>
76 
77 #include <sys/kernel.h>
78 #include <sys/unistd.h>
79 
80 #include <machine/limits.h>
81 
82 #include <vm/vm.h>
83 #include <vm/vm_param.h>
84 #include <vm/vm_prot.h>
85 #include <sys/lock.h>
86 #include <vm/pmap.h>
87 #include <vm/vm_map.h>
88 #include <vm/vm_page.h>
89 #include <vm/vm_pageout.h>
90 #include <vm/vm_kern.h>
91 #include <vm/vm_extern.h>
92 
93 #include <sys/user.h>
94 
95 /*
96  * System initialization
97  *
98  * Note: proc0 from proc.h
99  */
100 
101 static void vm_init_limits __P((void *));
102 SYSINIT(vm_limits, SI_SUB_VM_CONF, SI_ORDER_FIRST, vm_init_limits, &proc0)
103 
104 /*
105  * THIS MUST BE THE LAST INITIALIZATION ITEM!!!
106  *
107  * Note: run scheduling should be divorced from the vm system.
108  */
109 static void scheduler __P((void *));
110 SYSINIT(scheduler, SI_SUB_RUN_SCHEDULER, SI_ORDER_FIRST, scheduler, NULL)
111 
112 
113 static void swapout __P((struct proc *));
114 
115 int
116 kernacc(addr, len, rw)
117 	caddr_t addr;
118 	int len, rw;
119 {
120 	boolean_t rv;
121 	vm_offset_t saddr, eaddr;
122 	vm_prot_t prot = rw == B_READ ? VM_PROT_READ : VM_PROT_WRITE;
123 
124 	saddr = trunc_page((vm_offset_t)addr);
125 	eaddr = round_page((vm_offset_t)addr + len);
126 	vm_map_lock_read(kernel_map);
127 	rv = vm_map_check_protection(kernel_map, saddr, eaddr, prot);
128 	vm_map_unlock_read(kernel_map);
129 	return (rv == TRUE);
130 }
131 
132 int
133 useracc(addr, len, rw)
134 	caddr_t addr;
135 	int len, rw;
136 {
137 	boolean_t rv;
138 	vm_prot_t prot = rw == B_READ ? VM_PROT_READ : VM_PROT_WRITE;
139 	vm_map_t map;
140 	vm_map_entry_t save_hint;
141 
142 	/*
143 	 * XXX - check separately to disallow access to user area and user
144 	 * page tables - they are in the map.
145 	 *
146 	 * XXX - VM_MAXUSER_ADDRESS is an end address, not a max.  It was once
147 	 * only used (as an end address) in trap.c.  Use it as an end address
148 	 * here too.  This bogusness has spread.  I just fixed where it was
149 	 * used as a max in vm_mmap.c.
150 	 */
151 	if ((vm_offset_t) addr + len > /* XXX */ VM_MAXUSER_ADDRESS
152 	    || (vm_offset_t) addr + len < (vm_offset_t) addr) {
153 		return (FALSE);
154 	}
155 	map = &curproc->p_vmspace->vm_map;
156 	vm_map_lock_read(map);
157 	/*
158 	 * We save the map hint, and restore it.  Useracc appears to distort
159 	 * the map hint unnecessarily.
160 	 */
161 	save_hint = map->hint;
162 	rv = vm_map_check_protection(map,
163 	    trunc_page((vm_offset_t)addr), round_page((vm_offset_t)addr + len), prot);
164 	map->hint = save_hint;
165 	vm_map_unlock_read(map);
166 
167 	return (rv == TRUE);
168 }
169 
170 void
171 vslock(addr, len)
172 	caddr_t addr;
173 	u_int len;
174 {
175 	vm_map_pageable(&curproc->p_vmspace->vm_map, trunc_page((vm_offset_t)addr),
176 	    round_page((vm_offset_t)addr + len), FALSE);
177 }
178 
179 void
180 vsunlock(addr, len, dirtied)
181 	caddr_t addr;
182 	u_int len;
183 	int dirtied;
184 {
185 #ifdef	lint
186 	dirtied++;
187 #endif	/* lint */
188 	vm_map_pageable(&curproc->p_vmspace->vm_map, trunc_page((vm_offset_t)addr),
189 	    round_page((vm_offset_t)addr + len), TRUE);
190 }
191 
192 /*
193  * Implement fork's actions on an address space.
194  * Here we arrange for the address space to be copied or referenced,
195  * allocate a user struct (pcb and kernel stack), then call the
196  * machine-dependent layer to fill those in and make the new process
197  * ready to run.  The new process is set up so that it returns directly
198  * to user mode to avoid stack copying and relocation problems.
199  */
200 void
201 vm_fork(p1, p2, flags)
202 	register struct proc *p1, *p2;
203 	int flags;
204 {
205 	register struct user *up;
206 
207 	if (flags & RFMEM) {
208 		p2->p_vmspace = p1->p_vmspace;
209 		p1->p_vmspace->vm_refcnt++;
210 	}
211 
212 	/*
213 	 * Great, so we have a memory-heavy process and the
214 	 * entire machine comes to a screaching halt because
215 	 * nobody can fork/exec anything.  What we really need
216 	 * to do is fix the process swapper so it swaps out the right
217 	 * processes.
218 	 */
219 #if 0
220 	while ((cnt.v_free_count + cnt.v_cache_count) < cnt.v_free_min) {
221 		vm_pageout_deficit += (UPAGES + VM_INITIAL_PAGEIN);
222 		VM_WAIT;
223 	}
224 #endif
225 
226 	if ((flags & RFMEM) == 0) {
227 		p2->p_vmspace = vmspace_fork(p1->p_vmspace);
228 
229 		pmap_pinit2(vmspace_pmap(p2->p_vmspace));
230 
231 		if (p1->p_vmspace->vm_shm)
232 			shmfork(p1, p2);
233 	}
234 
235 	pmap_new_proc(p2);
236 
237 	up = p2->p_addr;
238 
239 	/*
240 	 * p_stats currently points at fields in the user struct
241 	 * but not at &u, instead at p_addr. Copy parts of
242 	 * p_stats; zero the rest of p_stats (statistics).
243 	 *
244 	 * If procsig->ps_refcnt is 1 and p2->p_sigacts is NULL we dont' need
245 	 * to share sigacts, so we use the up->u_sigacts.
246 	 */
247 	p2->p_stats = &up->u_stats;
248 	if (p2->p_sigacts == NULL) {
249 		if (p2->p_procsig->ps_refcnt != 1)
250 			printf ("PID:%d NULL sigacts with refcnt not 1!\n",p2->p_pid);
251 		p2->p_sigacts = &up->u_sigacts;
252 		up->u_sigacts = *p1->p_sigacts;
253 	}
254 
255 	bzero(&up->u_stats.pstat_startzero,
256 	    (unsigned) ((caddr_t) &up->u_stats.pstat_endzero -
257 		(caddr_t) &up->u_stats.pstat_startzero));
258 	bcopy(&p1->p_stats->pstat_startcopy, &up->u_stats.pstat_startcopy,
259 	    ((caddr_t) &up->u_stats.pstat_endcopy -
260 		(caddr_t) &up->u_stats.pstat_startcopy));
261 
262 
263 	/*
264 	 * cpu_fork will copy and update the pcb, set up the kernel stack,
265 	 * and make the child ready to run.
266 	 */
267 	cpu_fork(p1, p2);
268 }
269 
270 /*
271  * Set default limits for VM system.
272  * Called for proc 0, and then inherited by all others.
273  *
274  * XXX should probably act directly on proc0.
275  */
276 static void
277 vm_init_limits(udata)
278 	void *udata;
279 {
280 	register struct proc *p = udata;
281 	int rss_limit;
282 
283 	/*
284 	 * Set up the initial limits on process VM. Set the maximum resident
285 	 * set size to be half of (reasonably) available memory.  Since this
286 	 * is a soft limit, it comes into effect only when the system is out
287 	 * of memory - half of main memory helps to favor smaller processes,
288 	 * and reduces thrashing of the object cache.
289 	 */
290 	p->p_rlimit[RLIMIT_STACK].rlim_cur = DFLSSIZ;
291 	p->p_rlimit[RLIMIT_STACK].rlim_max = MAXSSIZ;
292 	p->p_rlimit[RLIMIT_DATA].rlim_cur = DFLDSIZ;
293 	p->p_rlimit[RLIMIT_DATA].rlim_max = MAXDSIZ;
294 	/* limit the limit to no less than 2MB */
295 	rss_limit = max(cnt.v_free_count, 512);
296 	p->p_rlimit[RLIMIT_RSS].rlim_cur = ptoa(rss_limit);
297 	p->p_rlimit[RLIMIT_RSS].rlim_max = RLIM_INFINITY;
298 }
299 
300 void
301 faultin(p)
302 	struct proc *p;
303 {
304 	int s;
305 
306 	if ((p->p_flag & P_INMEM) == 0) {
307 
308 		++p->p_lock;
309 
310 		pmap_swapin_proc(p);
311 
312 		s = splhigh();
313 
314 		if (p->p_stat == SRUN)
315 			setrunqueue(p);
316 
317 		p->p_flag |= P_INMEM;
318 
319 		/* undo the effect of setting SLOCK above */
320 		--p->p_lock;
321 		splx(s);
322 
323 	}
324 }
325 
326 /*
327  * This swapin algorithm attempts to swap-in processes only if there
328  * is enough space for them.  Of course, if a process waits for a long
329  * time, it will be swapped in anyway.
330  */
331 /* ARGSUSED*/
332 static void
333 scheduler(dummy)
334 	void *dummy;
335 {
336 	register struct proc *p;
337 	register int pri;
338 	struct proc *pp;
339 	int ppri;
340 
341 loop:
342 	while ((cnt.v_free_count + cnt.v_cache_count) < cnt.v_free_min) {
343 		VM_WAIT;
344 	}
345 
346 	pp = NULL;
347 	ppri = INT_MIN;
348 	for (p = allproc.lh_first; p != 0; p = p->p_list.le_next) {
349 		if (p->p_stat == SRUN &&
350 			(p->p_flag & (P_INMEM | P_SWAPPING)) == 0) {
351 
352 			pri = p->p_swtime + p->p_slptime;
353 			if ((p->p_flag & P_SWAPINREQ) == 0) {
354 				pri -= p->p_nice * 8;
355 			}
356 
357 			/*
358 			 * if this process is higher priority and there is
359 			 * enough space, then select this process instead of
360 			 * the previous selection.
361 			 */
362 			if (pri > ppri) {
363 				pp = p;
364 				ppri = pri;
365 			}
366 		}
367 	}
368 
369 	/*
370 	 * Nothing to do, back to sleep.
371 	 */
372 	if ((p = pp) == NULL) {
373 		tsleep(&proc0, PVM, "sched", 0);
374 		goto loop;
375 	}
376 	p->p_flag &= ~P_SWAPINREQ;
377 
378 	/*
379 	 * We would like to bring someone in. (only if there is space).
380 	 */
381 	faultin(p);
382 	p->p_swtime = 0;
383 	goto loop;
384 }
385 
386 #ifndef NO_SWAPPING
387 
388 #define	swappable(p) \
389 	(((p)->p_lock == 0) && \
390 		((p)->p_flag & (P_TRACED|P_SYSTEM|P_INMEM|P_WEXIT|P_SWAPPING)) == P_INMEM)
391 
392 
393 /*
394  * Swap_idle_threshold1 is the guaranteed swapped in time for a process
395  */
396 static int swap_idle_threshold1 = 2;
397 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold1,
398 	CTLFLAG_RW, &swap_idle_threshold1, 0, "");
399 
400 /*
401  * Swap_idle_threshold2 is the time that a process can be idle before
402  * it will be swapped out, if idle swapping is enabled.
403  */
404 static int swap_idle_threshold2 = 10;
405 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold2,
406 	CTLFLAG_RW, &swap_idle_threshold2, 0, "");
407 
408 /*
409  * Swapout is driven by the pageout daemon.  Very simple, we find eligible
410  * procs and unwire their u-areas.  We try to always "swap" at least one
411  * process in case we need the room for a swapin.
412  * If any procs have been sleeping/stopped for at least maxslp seconds,
413  * they are swapped.  Else, we swap the longest-sleeping or stopped process,
414  * if any, otherwise the longest-resident process.
415  */
416 void
417 swapout_procs(action)
418 int action;
419 {
420 	register struct proc *p;
421 	struct proc *outp, *outp2;
422 	int outpri, outpri2;
423 	int didswap = 0;
424 
425 	outp = outp2 = NULL;
426 	outpri = outpri2 = INT_MIN;
427 retry:
428 	for (p = allproc.lh_first; p != 0; p = p->p_list.le_next) {
429 		struct vmspace *vm;
430 		if (!swappable(p))
431 			continue;
432 
433 		vm = p->p_vmspace;
434 
435 		switch (p->p_stat) {
436 		default:
437 			continue;
438 
439 		case SSLEEP:
440 		case SSTOP:
441 			/*
442 			 * do not swapout a realtime process
443 			 */
444 			if (RTP_PRIO_IS_REALTIME(p->p_rtprio.type))
445 				continue;
446 
447 			/*
448 			 * Do not swapout a process waiting on a critical
449 			 * event of some kind.  Also guarantee swap_idle_threshold1
450 			 * time in memory.
451 			 */
452 			if (((p->p_priority & 0x7f) < PSOCK) ||
453 				(p->p_slptime < swap_idle_threshold1))
454 				continue;
455 
456 			/*
457 			 * If the system is under memory stress, or if we are swapping
458 			 * idle processes >= swap_idle_threshold2, then swap the process
459 			 * out.
460 			 */
461 			if (((action & VM_SWAP_NORMAL) == 0) &&
462 				(((action & VM_SWAP_IDLE) == 0) ||
463 				  (p->p_slptime < swap_idle_threshold2)))
464 				continue;
465 
466 			++vm->vm_refcnt;
467 			/*
468 			 * do not swapout a process that is waiting for VM
469 			 * data structures there is a possible deadlock.
470 			 */
471 			if (lockmgr(&vm->vm_map.lock,
472 					LK_EXCLUSIVE | LK_NOWAIT,
473 					(void *)0, curproc)) {
474 				vmspace_free(vm);
475 				continue;
476 			}
477 			vm_map_unlock(&vm->vm_map);
478 			/*
479 			 * If the process has been asleep for awhile and had
480 			 * most of its pages taken away already, swap it out.
481 			 */
482 			if ((action & VM_SWAP_NORMAL) ||
483 				((action & VM_SWAP_IDLE) &&
484 				 (p->p_slptime > swap_idle_threshold2))) {
485 				swapout(p);
486 				vmspace_free(vm);
487 				didswap++;
488 				goto retry;
489 			}
490 		}
491 	}
492 	/*
493 	 * If we swapped something out, and another process needed memory,
494 	 * then wakeup the sched process.
495 	 */
496 	if (didswap)
497 		wakeup(&proc0);
498 }
499 
500 static void
501 swapout(p)
502 	register struct proc *p;
503 {
504 
505 #if defined(SWAP_DEBUG)
506 	printf("swapping out %d\n", p->p_pid);
507 #endif
508 	++p->p_stats->p_ru.ru_nswap;
509 	/*
510 	 * remember the process resident count
511 	 */
512 	p->p_vmspace->vm_swrss = vmspace_resident_count(p->p_vmspace);
513 
514 	(void) splhigh();
515 	p->p_flag &= ~P_INMEM;
516 	p->p_flag |= P_SWAPPING;
517 	if (p->p_stat == SRUN)
518 		remrunqueue(p);
519 	(void) spl0();
520 
521 	pmap_swapout_proc(p);
522 
523 	p->p_flag &= ~P_SWAPPING;
524 	p->p_swtime = 0;
525 }
526 #endif /* !NO_SWAPPING */
527