xref: /freebsd/sys/vm/vm_glue.c (revision ee2ea5ceafed78a5bd9810beb9e3ca927180c226)
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/lock.h>
70 #include <sys/mutex.h>
71 #include <sys/proc.h>
72 #include <sys/resourcevar.h>
73 #include <sys/shm.h>
74 #include <sys/vmmeter.h>
75 #include <sys/sx.h>
76 #include <sys/sysctl.h>
77 
78 #include <sys/kernel.h>
79 #include <sys/ktr.h>
80 #include <sys/unistd.h>
81 
82 #include <machine/limits.h>
83 
84 #include <vm/vm.h>
85 #include <vm/vm_param.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 extern int maxslp;
96 
97 /*
98  * System initialization
99  *
100  * Note: proc0 from proc.h
101  */
102 static void vm_init_limits(void *);
103 SYSINIT(vm_limits, SI_SUB_VM_CONF, SI_ORDER_FIRST, vm_init_limits, &proc0)
104 
105 /*
106  * THIS MUST BE THE LAST INITIALIZATION ITEM!!!
107  *
108  * Note: run scheduling should be divorced from the vm system.
109  */
110 static void scheduler(void *);
111 SYSINIT(scheduler, SI_SUB_RUN_SCHEDULER, SI_ORDER_FIRST, scheduler, NULL)
112 
113 #ifndef NO_SWAPPING
114 static void swapout(struct proc *);
115 #endif
116 
117 int
118 kernacc(addr, len, rw)
119 	caddr_t addr;
120 	int len, rw;
121 {
122 	boolean_t rv;
123 	vm_offset_t saddr, eaddr;
124 	vm_prot_t prot;
125 
126 	KASSERT((rw & ~VM_PROT_ALL) == 0,
127 	    ("illegal ``rw'' argument to kernacc (%x)\n", rw));
128 	prot = rw;
129 	saddr = trunc_page((vm_offset_t)addr);
130 	eaddr = round_page((vm_offset_t)addr + len);
131 	rv = vm_map_check_protection(kernel_map, saddr, eaddr, prot);
132 	return (rv == TRUE);
133 }
134 
135 int
136 useracc(addr, len, rw)
137 	caddr_t addr;
138 	int len, rw;
139 {
140 	boolean_t rv;
141 	vm_prot_t prot;
142 
143 	GIANT_REQUIRED;
144 
145 	KASSERT((rw & ~VM_PROT_ALL) == 0,
146 	    ("illegal ``rw'' argument to useracc (%x)\n", rw));
147 	prot = rw;
148 	/*
149 	 * XXX - check separately to disallow access to user area and user
150 	 * page tables - they are in the map.
151 	 *
152 	 * XXX - VM_MAXUSER_ADDRESS is an end address, not a max.  It was once
153 	 * only used (as an end address) in trap.c.  Use it as an end address
154 	 * here too.  This bogusness has spread.  I just fixed where it was
155 	 * used as a max in vm_mmap.c.
156 	 */
157 	if ((vm_offset_t) addr + len > /* XXX */ VM_MAXUSER_ADDRESS
158 	    || (vm_offset_t) addr + len < (vm_offset_t) addr) {
159 		return (FALSE);
160 	}
161 	rv = vm_map_check_protection(&curproc->p_vmspace->vm_map,
162 	    trunc_page((vm_offset_t)addr), round_page((vm_offset_t)addr + len),
163 	    prot);
164 	return (rv == TRUE);
165 }
166 
167 void
168 vslock(addr, len)
169 	caddr_t addr;
170 	u_int len;
171 {
172 	GIANT_REQUIRED;
173 	vm_map_pageable(&curproc->p_vmspace->vm_map,
174 	    trunc_page((vm_offset_t)addr),
175 	    round_page((vm_offset_t)addr + len), FALSE);
176 }
177 
178 void
179 vsunlock(addr, len)
180 	caddr_t addr;
181 	u_int len;
182 {
183 	GIANT_REQUIRED;
184 	vm_map_pageable(&curproc->p_vmspace->vm_map,
185 	    trunc_page((vm_offset_t)addr),
186 	    round_page((vm_offset_t)addr + len), TRUE);
187 }
188 
189 /*
190  * Implement fork's actions on an address space.
191  * Here we arrange for the address space to be copied or referenced,
192  * allocate a user struct (pcb and kernel stack), then call the
193  * machine-dependent layer to fill those in and make the new process
194  * ready to run.  The new process is set up so that it returns directly
195  * to user mode to avoid stack copying and relocation problems.
196  */
197 void
198 vm_forkproc(td, p2, td2, flags)
199 	struct thread *td;
200 	struct proc *p2;
201 	struct thread *td2;
202 	int flags;
203 {
204 	struct proc *p1 = td->td_proc;
205 	struct user *up;
206 
207 	GIANT_REQUIRED;
208 
209 	if ((flags & RFPROC) == 0) {
210 		/*
211 		 * Divorce the memory, if it is shared, essentially
212 		 * this changes shared memory amongst threads, into
213 		 * COW locally.
214 		 */
215 		if ((flags & RFMEM) == 0) {
216 			if (p1->p_vmspace->vm_refcnt > 1) {
217 				vmspace_unshare(p1);
218 			}
219 		}
220 		cpu_fork(td, p2, td2, flags);
221 		return;
222 	}
223 
224 	if (flags & RFMEM) {
225 		p2->p_vmspace = p1->p_vmspace;
226 		p1->p_vmspace->vm_refcnt++;
227 	}
228 
229 	while (vm_page_count_severe()) {
230 		VM_WAIT;
231 	}
232 
233 	if ((flags & RFMEM) == 0) {
234 		p2->p_vmspace = vmspace_fork(p1->p_vmspace);
235 
236 		pmap_pinit2(vmspace_pmap(p2->p_vmspace));
237 
238 		if (p1->p_vmspace->vm_shm)
239 			shmfork(p1, p2);
240 	}
241 
242 	pmap_new_proc(p2);
243 	pmap_new_thread(td2);		/* Initial thread */
244 
245 	/* XXXKSE this is unsatisfactory but should be adequate */
246 	up = p2->p_uarea;
247 
248 	/*
249 	 * p_stats currently points at fields in the user struct
250 	 * but not at &u, instead at p_addr. Copy parts of
251 	 * p_stats; zero the rest of p_stats (statistics).
252 	 *
253 	 * If procsig->ps_refcnt is 1 and p2->p_sigacts is NULL we dont' need
254 	 * to share sigacts, so we use the up->u_sigacts.
255 	 */
256 	p2->p_stats = &up->u_stats;
257 	if (p2->p_sigacts == NULL) {
258 		if (p2->p_procsig->ps_refcnt != 1)
259 			printf ("PID:%d NULL sigacts with refcnt not 1!\n",p2->p_pid);
260 		p2->p_sigacts = &up->u_sigacts;
261 		up->u_sigacts = *p1->p_sigacts;
262 	}
263 
264 	bzero(&up->u_stats.pstat_startzero,
265 	    (unsigned) ((caddr_t) &up->u_stats.pstat_endzero -
266 		(caddr_t) &up->u_stats.pstat_startzero));
267 	bcopy(&p1->p_stats->pstat_startcopy, &up->u_stats.pstat_startcopy,
268 	    ((caddr_t) &up->u_stats.pstat_endcopy -
269 		(caddr_t) &up->u_stats.pstat_startcopy));
270 
271 
272 	/*
273 	 * cpu_fork will copy and update the pcb, set up the kernel stack,
274 	 * and make the child ready to run.
275 	 */
276 	cpu_fork(td, p2, td2, flags);
277 }
278 
279 /*
280  * Called after process has been wait(2)'ed apon and is being reaped.
281  * The idea is to reclaim resources that we could not reclaim while
282  * the process was still executing.
283  */
284 void
285 vm_waitproc(p)
286 	struct proc *p;
287 {
288 	struct thread *td;
289 
290 	GIANT_REQUIRED;
291 	cpu_wait(p);
292 	pmap_dispose_proc(p);		/* drop per-process resources */
293 	FOREACH_THREAD_IN_PROC(p, td)
294 		pmap_dispose_thread(td);
295 	vmspace_exitfree(p);		/* and clean-out the vmspace */
296 }
297 
298 /*
299  * Set default limits for VM system.
300  * Called for proc 0, and then inherited by all others.
301  *
302  * XXX should probably act directly on proc0.
303  */
304 static void
305 vm_init_limits(udata)
306 	void *udata;
307 {
308 	struct proc *p = udata;
309 	int rss_limit;
310 
311 	/*
312 	 * Set up the initial limits on process VM. Set the maximum resident
313 	 * set size to be half of (reasonably) available memory.  Since this
314 	 * is a soft limit, it comes into effect only when the system is out
315 	 * of memory - half of main memory helps to favor smaller processes,
316 	 * and reduces thrashing of the object cache.
317 	 */
318 	p->p_rlimit[RLIMIT_STACK].rlim_cur = dflssiz;
319 	p->p_rlimit[RLIMIT_STACK].rlim_max = maxssiz;
320 	p->p_rlimit[RLIMIT_DATA].rlim_cur = dfldsiz;
321 	p->p_rlimit[RLIMIT_DATA].rlim_max = maxdsiz;
322 	/* limit the limit to no less than 2MB */
323 	rss_limit = max(cnt.v_free_count, 512);
324 	p->p_rlimit[RLIMIT_RSS].rlim_cur = ptoa(rss_limit);
325 	p->p_rlimit[RLIMIT_RSS].rlim_max = RLIM_INFINITY;
326 }
327 
328 void
329 faultin(p)
330 	struct proc *p;
331 {
332 	struct thread *td;
333 	GIANT_REQUIRED;
334 
335 	PROC_LOCK_ASSERT(p, MA_OWNED);
336 	mtx_lock_spin(&sched_lock);
337 	if ((p->p_sflag & PS_INMEM) == 0) {
338 		++p->p_lock;
339 		mtx_unlock_spin(&sched_lock);
340 		PROC_UNLOCK(p);
341 
342 		pmap_swapin_proc(p);
343 		FOREACH_THREAD_IN_PROC (p, td)
344 			pmap_swapin_thread(td);
345 
346 		PROC_LOCK(p);
347 		mtx_lock_spin(&sched_lock);
348 		FOREACH_THREAD_IN_PROC (p, td)
349 			if (td->td_proc->p_stat == SRUN)	/* XXXKSE */
350 				setrunqueue(td);
351 
352 		p->p_sflag |= PS_INMEM;
353 
354 		/* undo the effect of setting SLOCK above */
355 		--p->p_lock;
356 	}
357 	mtx_unlock_spin(&sched_lock);
358 }
359 
360 /*
361  * This swapin algorithm attempts to swap-in processes only if there
362  * is enough space for them.  Of course, if a process waits for a long
363  * time, it will be swapped in anyway.
364  *
365  *  XXXKSE - KSEGRP with highest priority counts..
366  *
367  * Giant is still held at this point, to be released in tsleep.
368  */
369 /* ARGSUSED*/
370 static void
371 scheduler(dummy)
372 	void *dummy;
373 {
374 	struct proc *p;
375 	int pri;
376 	struct proc *pp;
377 	int ppri;
378 
379 	mtx_assert(&Giant, MA_OWNED | MA_NOTRECURSED);
380 	/* GIANT_REQUIRED */
381 
382 loop:
383 	if (vm_page_count_min()) {
384 		VM_WAIT;
385 		goto loop;
386 	}
387 
388 	pp = NULL;
389 	ppri = INT_MIN;
390 	sx_slock(&allproc_lock);
391 	FOREACH_PROC_IN_SYSTEM(p) {
392 		struct ksegrp *kg;
393 		mtx_lock_spin(&sched_lock);
394 		if (p->p_stat == SRUN
395 		&& (p->p_sflag & (PS_INMEM | PS_SWAPPING)) == 0) {
396 			/* Find the minimum sleeptime for the process */
397 			FOREACH_KSEGRP_IN_PROC(p, kg) {
398 				pri = p->p_swtime + kg->kg_slptime;
399 				if ((p->p_sflag & PS_SWAPINREQ) == 0) {
400 					pri -= kg->kg_nice * 8;
401 				}
402 
403 				/*
404 				 * if this ksegrp is higher priority
405 				 * and there is enough space, then select
406 				 * this process instead of the previous
407 				 * selection.
408 				 */
409 				if (pri > ppri) {
410 					pp = p;
411 					ppri = pri;
412 				}
413 			}
414 		}
415 		mtx_unlock_spin(&sched_lock);
416 	}
417 	sx_sunlock(&allproc_lock);
418 
419 	/*
420 	 * Nothing to do, back to sleep.
421 	 */
422 	if ((p = pp) == NULL) {
423 		tsleep(&proc0, PVM, "sched", maxslp * hz / 2);
424 		goto loop;
425 	}
426 	mtx_lock_spin(&sched_lock);
427 	p->p_sflag &= ~PS_SWAPINREQ;
428 	mtx_unlock_spin(&sched_lock);
429 
430 	/*
431 	 * We would like to bring someone in. (only if there is space).
432 	 */
433 	PROC_LOCK(p);
434 	faultin(p);
435 	PROC_UNLOCK(p);
436 	mtx_lock_spin(&sched_lock);
437 	p->p_swtime = 0;
438 	mtx_unlock_spin(&sched_lock);
439 	goto loop;
440 }
441 
442 #ifndef NO_SWAPPING
443 
444 /*
445  * Swap_idle_threshold1 is the guaranteed swapped in time for a process
446  */
447 static int swap_idle_threshold1 = 2;
448 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold1,
449 	CTLFLAG_RW, &swap_idle_threshold1, 0, "");
450 
451 /*
452  * Swap_idle_threshold2 is the time that a process can be idle before
453  * it will be swapped out, if idle swapping is enabled.
454  */
455 static int swap_idle_threshold2 = 10;
456 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold2,
457 	CTLFLAG_RW, &swap_idle_threshold2, 0, "");
458 
459 /*
460  * Swapout is driven by the pageout daemon.  Very simple, we find eligible
461  * procs and unwire their u-areas.  We try to always "swap" at least one
462  * process in case we need the room for a swapin.
463  * If any procs have been sleeping/stopped for at least maxslp seconds,
464  * they are swapped.  Else, we swap the longest-sleeping or stopped process,
465  * if any, otherwise the longest-resident process.
466  */
467 void
468 swapout_procs(action)
469 int action;
470 {
471 	struct proc *p;
472 	struct ksegrp *kg;
473 	struct proc *outp, *outp2;
474 	int outpri, outpri2;
475 	int didswap = 0;
476 
477 	GIANT_REQUIRED;
478 
479 	outp = outp2 = NULL;
480 	outpri = outpri2 = INT_MIN;
481 retry:
482 	sx_slock(&allproc_lock);
483 	LIST_FOREACH(p, &allproc, p_list) {
484 		struct vmspace *vm;
485 		int minslptime = 100000;
486 
487 		PROC_LOCK(p);
488 		if (p->p_lock != 0 ||
489 		    (p->p_flag & (P_TRACED|P_SYSTEM|P_WEXIT)) != 0) {
490 			PROC_UNLOCK(p);
491 			continue;
492 		}
493 		/*
494 		 * only aiod changes vmspace, however it will be
495 		 * skipped because of the if statement above checking
496 		 * for P_SYSTEM
497 		 */
498 		vm = p->p_vmspace;
499 		mtx_lock_spin(&sched_lock);
500 		if ((p->p_sflag & (PS_INMEM|PS_SWAPPING)) != PS_INMEM) {
501 			mtx_unlock_spin(&sched_lock);
502 			PROC_UNLOCK(p);
503 			continue;
504 		}
505 
506 		switch (p->p_stat) {
507 		default:
508 			mtx_unlock_spin(&sched_lock);
509 			PROC_UNLOCK(p);
510 			continue;
511 
512 		case SSLEEP:
513 		case SSTOP:
514 			/*
515 			 * do not swapout a realtime process
516 			 * Check all the thread groups..
517 			 */
518 			FOREACH_KSEGRP_IN_PROC(p, kg) {
519 				if (PRI_IS_REALTIME(kg->kg_pri_class)) {
520 					mtx_unlock_spin(&sched_lock);
521 					PROC_UNLOCK(p);
522 					goto nextproc;
523 				}
524 
525 				/*
526 				 * Do not swapout a process waiting
527 				 * on a critical event of some kind.
528 				 * Also guarantee swap_idle_threshold1
529 				 * time in memory.
530 				 */
531 				if (((FIRST_THREAD_IN_PROC(p)->td_priority) < PSOCK) ||
532 				    (kg->kg_slptime < swap_idle_threshold1)) {
533 					mtx_unlock_spin(&sched_lock);
534 					PROC_UNLOCK(p);
535 					goto nextproc;
536 				}
537 
538 				/*
539 				 * If the system is under memory stress,
540 				 * or if we are swapping
541 				 * idle processes >= swap_idle_threshold2,
542 				 * then swap the process out.
543 				 */
544 				if (((action & VM_SWAP_NORMAL) == 0) &&
545 				    (((action & VM_SWAP_IDLE) == 0) ||
546 				    (kg->kg_slptime < swap_idle_threshold2))) {
547 					mtx_unlock_spin(&sched_lock);
548 					PROC_UNLOCK(p);
549 					goto nextproc;
550 				}
551 				if (minslptime > kg->kg_slptime)
552 					minslptime = kg->kg_slptime;
553 			}
554 
555 			mtx_unlock_spin(&sched_lock);
556 			++vm->vm_refcnt;
557 			/*
558 			 * do not swapout a process that
559 			 * is waiting for VM
560 			 * data structures there is a
561 			 * possible deadlock.
562 			 */
563 			if (!vm_map_trylock(&vm->vm_map)) {
564 				vmspace_free(vm);
565 				PROC_UNLOCK(p);
566 				goto nextproc;
567 			}
568 			vm_map_unlock(&vm->vm_map);
569 			/*
570 			 * If the process has been asleep for awhile and had
571 			 * most of its pages taken away already, swap it out.
572 			 */
573 			if ((action & VM_SWAP_NORMAL) ||
574 				((action & VM_SWAP_IDLE) &&
575 				 (minslptime > swap_idle_threshold2))) {
576 				sx_sunlock(&allproc_lock);
577 				swapout(p);
578 				vmspace_free(vm);
579 				didswap++;
580 				goto retry;
581 			}
582 			PROC_UNLOCK(p);
583 			vmspace_free(vm);
584 		}
585 nextproc:
586 		continue;
587 	}
588 	sx_sunlock(&allproc_lock);
589 	/*
590 	 * If we swapped something out, and another process needed memory,
591 	 * then wakeup the sched process.
592 	 */
593 	if (didswap)
594 		wakeup(&proc0);
595 }
596 
597 static void
598 swapout(p)
599 	struct proc *p;
600 {
601 	struct thread *td;
602 
603 	PROC_LOCK_ASSERT(p, MA_OWNED);
604 #if defined(SWAP_DEBUG)
605 	printf("swapping out %d\n", p->p_pid);
606 #endif
607 	++p->p_stats->p_ru.ru_nswap;
608 	/*
609 	 * remember the process resident count
610 	 */
611 	p->p_vmspace->vm_swrss = vmspace_resident_count(p->p_vmspace);
612 
613 	mtx_lock_spin(&sched_lock);
614 	p->p_sflag &= ~PS_INMEM;
615 	p->p_sflag |= PS_SWAPPING;
616 	PROC_UNLOCK(p);
617 	FOREACH_THREAD_IN_PROC (p, td)
618 		if (td->td_proc->p_stat == SRUN)	/* XXXKSE */
619 			remrunqueue(td);	/* XXXKSE */
620 	mtx_unlock_spin(&sched_lock);
621 
622 	pmap_swapout_proc(p);
623 	FOREACH_THREAD_IN_PROC(p, td)
624 		pmap_swapout_thread(td);
625 
626 	mtx_lock_spin(&sched_lock);
627 	p->p_sflag &= ~PS_SWAPPING;
628 	p->p_swtime = 0;
629 	mtx_unlock_spin(&sched_lock);
630 }
631 #endif /* !NO_SWAPPING */
632