xref: /dragonfly/sys/vm/vm_glue.c (revision 31524921)
1 /*
2  * (MPSAFE)
3  *
4  * Copyright (c) 1991, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  *
7  * This code is derived from software contributed to Berkeley by
8  * The Mach Operating System project at Carnegie-Mellon University.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  * 3. Neither the name of the University nor the names of its contributors
19  *    may be used to endorse or promote products derived from this software
20  *    without specific prior written permission.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  *
34  *	from: @(#)vm_glue.c	8.6 (Berkeley) 1/5/94
35  *
36  *
37  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
38  * All rights reserved.
39  *
40  * Permission to use, copy, modify and distribute this software and
41  * its documentation is hereby granted, provided that both the copyright
42  * notice and this permission notice appear in all copies of the
43  * software, derivative works or modified versions, and any portions
44  * thereof, and that both notices appear in supporting documentation.
45  *
46  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
47  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
48  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
49  *
50  * Carnegie Mellon requests users of this software to return to
51  *
52  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
53  *  School of Computer Science
54  *  Carnegie Mellon University
55  *  Pittsburgh PA 15213-3890
56  *
57  * any improvements or extensions that they make and grant Carnegie the
58  * rights to redistribute these changes.
59  *
60  * $FreeBSD: src/sys/vm/vm_glue.c,v 1.94.2.4 2003/01/13 22:51:17 dillon Exp $
61  */
62 
63 #include "opt_vm.h"
64 
65 #include <sys/param.h>
66 #include <sys/systm.h>
67 #include <sys/proc.h>
68 #include <sys/resourcevar.h>
69 #include <sys/buf.h>
70 #include <sys/shm.h>
71 #include <sys/vmmeter.h>
72 #include <sys/sysctl.h>
73 
74 #include <sys/kernel.h>
75 #include <sys/unistd.h>
76 
77 #include <machine/limits.h>
78 #include <machine/vmm.h>
79 
80 #include <vm/vm.h>
81 #include <vm/vm_param.h>
82 #include <sys/lock.h>
83 #include <vm/pmap.h>
84 #include <vm/vm_map.h>
85 #include <vm/vm_page.h>
86 #include <vm/vm_pageout.h>
87 #include <vm/vm_kern.h>
88 #include <vm/vm_extern.h>
89 
90 #include <sys/user.h>
91 #include <vm/vm_page2.h>
92 #include <sys/thread2.h>
93 #include <sys/sysref2.h>
94 
95 /*
96  * THIS MUST BE THE LAST INITIALIZATION ITEM!!!
97  *
98  * Note: run scheduling should be divorced from the vm system.
99  */
100 static void scheduler (void *);
101 SYSINIT(scheduler, SI_SUB_RUN_SCHEDULER, SI_ORDER_FIRST, scheduler, NULL);
102 
103 #ifdef INVARIANTS
104 
105 static int swap_debug = 0;
106 SYSCTL_INT(_vm, OID_AUTO, swap_debug,
107 	CTLFLAG_RW, &swap_debug, 0, "");
108 
109 #endif
110 
111 static int scheduler_notify;
112 
113 static void swapout (struct proc *);
114 
115 /*
116  * No requirements.
117  */
118 int
119 kernacc(c_caddr_t addr, int len, int rw)
120 {
121 	boolean_t rv;
122 	vm_offset_t saddr, eaddr;
123 	vm_prot_t prot;
124 
125 	KASSERT((rw & (~VM_PROT_ALL)) == 0,
126 	    ("illegal ``rw'' argument to kernacc (%x)", rw));
127 
128 	/*
129 	 * The globaldata space is not part of the kernel_map proper,
130 	 * check access separately.
131 	 */
132 	if (is_globaldata_space((vm_offset_t)addr, (vm_offset_t)(addr + len)))
133 		return (TRUE);
134 
135 	/*
136 	 * Nominal kernel memory access - check access via kernel_map.
137 	 */
138 	if ((vm_offset_t)addr + len > kernel_map.max_offset ||
139 	    (vm_offset_t)addr + len < (vm_offset_t)addr) {
140 		return (FALSE);
141 	}
142 	prot = rw;
143 	saddr = trunc_page((vm_offset_t)addr);
144 	eaddr = round_page((vm_offset_t)addr + len);
145 	rv = vm_map_check_protection(&kernel_map, saddr, eaddr, prot, FALSE);
146 
147 	return (rv == TRUE);
148 }
149 
150 /*
151  * No requirements.
152  */
153 int
154 useracc(c_caddr_t addr, int len, int rw)
155 {
156 	boolean_t rv;
157 	vm_prot_t prot;
158 	vm_map_t map;
159 	vm_map_entry_t save_hint;
160 	vm_offset_t wrap;
161 	vm_offset_t gpa;
162 
163 	KASSERT((rw & (~VM_PROT_ALL)) == 0,
164 	    ("illegal ``rw'' argument to useracc (%x)", rw));
165 	prot = rw;
166 
167 	if (curthread->td_vmm) {
168 		if (vmm_vm_get_gpa(curproc, (register_t *)&gpa, (register_t) addr))
169 			panic("%s: could not get GPA\n", __func__);
170 		addr = (c_caddr_t) gpa;
171 	}
172 
173 	/*
174 	 * XXX - check separately to disallow access to user area and user
175 	 * page tables - they are in the map.
176 	 */
177 	wrap = (vm_offset_t)addr + len;
178 	if (wrap > VM_MAX_USER_ADDRESS || wrap < (vm_offset_t)addr) {
179 		return (FALSE);
180 	}
181 	map = &curproc->p_vmspace->vm_map;
182 	vm_map_lock_read(map);
183 	/*
184 	 * We save the map hint, and restore it.  Useracc appears to distort
185 	 * the map hint unnecessarily.
186 	 */
187 	save_hint = map->hint;
188 	rv = vm_map_check_protection(map, trunc_page((vm_offset_t)addr),
189 				     round_page(wrap), prot, TRUE);
190 	map->hint = save_hint;
191 	vm_map_unlock_read(map);
192 
193 	return (rv == TRUE);
194 }
195 
196 /*
197  * No requirements.
198  */
199 void
200 vslock(caddr_t addr, u_int len)
201 {
202 	if (len) {
203 		vm_map_wire(&curproc->p_vmspace->vm_map,
204 			    trunc_page((vm_offset_t)addr),
205 			    round_page((vm_offset_t)addr + len), 0);
206 	}
207 }
208 
209 /*
210  * No requirements.
211  */
212 void
213 vsunlock(caddr_t addr, u_int len)
214 {
215 	if (len) {
216 		vm_map_wire(&curproc->p_vmspace->vm_map,
217 			    trunc_page((vm_offset_t)addr),
218 			    round_page((vm_offset_t)addr + len),
219 			    KM_PAGEABLE);
220 	}
221 }
222 
223 /*
224  * Implement fork's actions on an address space.
225  * Here we arrange for the address space to be copied or referenced,
226  * allocate a user struct (pcb and kernel stack), then call the
227  * machine-dependent layer to fill those in and make the new process
228  * ready to run.  The new process is set up so that it returns directly
229  * to user mode to avoid stack copying and relocation problems.
230  *
231  * No requirements.
232  */
233 void
234 vm_fork(struct proc *p1, struct proc *p2, int flags)
235 {
236 	if ((flags & RFPROC) == 0) {
237 		/*
238 		 * Divorce the memory, if it is shared, essentially
239 		 * this changes shared memory amongst threads, into
240 		 * COW locally.
241 		 */
242 		if ((flags & RFMEM) == 0) {
243 			if (vmspace_getrefs(p1->p_vmspace) > 1) {
244 				vmspace_unshare(p1);
245 			}
246 		}
247 		cpu_fork(ONLY_LWP_IN_PROC(p1), NULL, flags);
248 		return;
249 	}
250 
251 	if (flags & RFMEM) {
252 		vmspace_ref(p1->p_vmspace);
253 		p2->p_vmspace = p1->p_vmspace;
254 	}
255 
256 	while (vm_page_count_severe()) {
257 		vm_wait(0);
258 	}
259 
260 	if ((flags & RFMEM) == 0) {
261 		p2->p_vmspace = vmspace_fork(p1->p_vmspace);
262 
263 		pmap_pinit2(vmspace_pmap(p2->p_vmspace));
264 
265 		if (p1->p_vmspace->vm_shm)
266 			shmfork(p1, p2);
267 	}
268 
269 	pmap_init_proc(p2);
270 }
271 
272 /*
273  * Set default limits for VM system.  Call during proc0's initialization.
274  *
275  * Called from the low level boot code only.
276  */
277 void
278 vm_init_limits(struct proc *p)
279 {
280 	int rss_limit;
281 
282 	/*
283 	 * Set up the initial limits on process VM. Set the maximum resident
284 	 * set size to be half of (reasonably) available memory.  Since this
285 	 * is a soft limit, it comes into effect only when the system is out
286 	 * of memory - half of main memory helps to favor smaller processes,
287 	 * and reduces thrashing of the object cache.
288 	 */
289 	p->p_rlimit[RLIMIT_STACK].rlim_cur = dflssiz;
290 	p->p_rlimit[RLIMIT_STACK].rlim_max = maxssiz;
291 	p->p_rlimit[RLIMIT_DATA].rlim_cur = dfldsiz;
292 	p->p_rlimit[RLIMIT_DATA].rlim_max = maxdsiz;
293 	/* limit the limit to no less than 2MB */
294 	rss_limit = max(vmstats.v_free_count, 512);
295 	p->p_rlimit[RLIMIT_RSS].rlim_cur = ptoa(rss_limit);
296 	p->p_rlimit[RLIMIT_RSS].rlim_max = RLIM_INFINITY;
297 }
298 
299 /*
300  * Faultin the specified process.  Note that the process can be in any
301  * state.  Just clear P_SWAPPEDOUT and call wakeup in case the process is
302  * sleeping.
303  *
304  * No requirements.
305  */
306 void
307 faultin(struct proc *p)
308 {
309 	if (p->p_flags & P_SWAPPEDOUT) {
310 		/*
311 		 * The process is waiting in the kernel to return to user
312 		 * mode but cannot until P_SWAPPEDOUT gets cleared.
313 		 */
314 		lwkt_gettoken(&p->p_token);
315 		p->p_flags &= ~(P_SWAPPEDOUT | P_SWAPWAIT);
316 #ifdef INVARIANTS
317 		if (swap_debug)
318 			kprintf("swapping in %d (%s)\n", p->p_pid, p->p_comm);
319 #endif
320 		wakeup(p);
321 		lwkt_reltoken(&p->p_token);
322 	}
323 }
324 
325 /*
326  * Kernel initialization eventually falls through to this function,
327  * which is process 0.
328  *
329  * This swapin algorithm attempts to swap-in processes only if there
330  * is enough space for them.  Of course, if a process waits for a long
331  * time, it will be swapped in anyway.
332  */
333 struct scheduler_info {
334 	struct proc *pp;
335 	int ppri;
336 };
337 
338 static int scheduler_callback(struct proc *p, void *data);
339 
340 static void
341 scheduler(void *dummy)
342 {
343 	struct scheduler_info info;
344 	struct proc *p;
345 
346 	KKASSERT(!IN_CRITICAL_SECT(curthread));
347 loop:
348 	scheduler_notify = 0;
349 	/*
350 	 * Don't try to swap anything in if we are low on memory.
351 	 */
352 	if (vm_page_count_severe()) {
353 		vm_wait(0);
354 		goto loop;
355 	}
356 
357 	/*
358 	 * Look for a good candidate to wake up
359 	 */
360 	info.pp = NULL;
361 	info.ppri = INT_MIN;
362 	allproc_scan(scheduler_callback, &info);
363 
364 	/*
365 	 * Nothing to do, back to sleep for at least 1/10 of a second.  If
366 	 * we are woken up, immediately process the next request.  If
367 	 * multiple requests have built up the first is processed
368 	 * immediately and the rest are staggered.
369 	 */
370 	if ((p = info.pp) == NULL) {
371 		tsleep(&proc0, 0, "nowork", hz / 10);
372 		if (scheduler_notify == 0)
373 			tsleep(&scheduler_notify, 0, "nowork", 0);
374 		goto loop;
375 	}
376 
377 	/*
378 	 * Fault the selected process in, then wait for a short period of
379 	 * time and loop up.
380 	 *
381 	 * XXX we need a heuristic to get a measure of system stress and
382 	 * then adjust our stagger wakeup delay accordingly.
383 	 */
384 	lwkt_gettoken(&p->p_token);
385 	faultin(p);
386 	p->p_swtime = 0;
387 	lwkt_reltoken(&p->p_token);
388 	PRELE(p);
389 	tsleep(&proc0, 0, "swapin", hz / 10);
390 	goto loop;
391 }
392 
393 static int
394 scheduler_callback(struct proc *p, void *data)
395 {
396 	struct scheduler_info *info = data;
397 	struct lwp *lp;
398 	segsz_t pgs;
399 	int pri;
400 
401 	if (p->p_flags & P_SWAPWAIT) {
402 		pri = 0;
403 		FOREACH_LWP_IN_PROC(lp, p) {
404 			/* XXX lwp might need a different metric */
405 			pri += lp->lwp_slptime;
406 		}
407 		pri += p->p_swtime - p->p_nice * 8;
408 
409 		/*
410 		 * The more pages paged out while we were swapped,
411 		 * the more work we have to do to get up and running
412 		 * again and the lower our wakeup priority.
413 		 *
414 		 * Each second of sleep time is worth ~1MB
415 		 */
416 		lwkt_gettoken(&p->p_vmspace->vm_map.token);
417 		pgs = vmspace_resident_count(p->p_vmspace);
418 		if (pgs < p->p_vmspace->vm_swrss) {
419 			pri -= (p->p_vmspace->vm_swrss - pgs) /
420 				(1024 * 1024 / PAGE_SIZE);
421 		}
422 		lwkt_reltoken(&p->p_vmspace->vm_map.token);
423 
424 		/*
425 		 * If this process is higher priority and there is
426 		 * enough space, then select this process instead of
427 		 * the previous selection.
428 		 */
429 		if (pri > info->ppri) {
430 			if (info->pp)
431 				PRELE(info->pp);
432 			PHOLD(p);
433 			info->pp = p;
434 			info->ppri = pri;
435 		}
436 	}
437 	return(0);
438 }
439 
440 /*
441  * SMP races ok.
442  * No requirements.
443  */
444 void
445 swapin_request(void)
446 {
447 	if (scheduler_notify == 0) {
448 		scheduler_notify = 1;
449 		wakeup(&scheduler_notify);
450 	}
451 }
452 
453 #ifndef NO_SWAPPING
454 
455 #define	swappable(p) \
456 	(((p)->p_lock == 0) && \
457 	((p)->p_flags & (P_TRACED|P_SYSTEM|P_SWAPPEDOUT|P_WEXIT)) == 0)
458 
459 
460 /*
461  * Swap_idle_threshold1 is the guaranteed swapped in time for a process
462  */
463 static int swap_idle_threshold1 = 15;
464 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold1,
465 	CTLFLAG_RW, &swap_idle_threshold1, 0, "Guaranteed process resident time (sec)");
466 
467 /*
468  * Swap_idle_threshold2 is the time that a process can be idle before
469  * it will be swapped out, if idle swapping is enabled.  Default is
470  * one minute.
471  */
472 static int swap_idle_threshold2 = 60;
473 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold2,
474 	CTLFLAG_RW, &swap_idle_threshold2, 0, "Time (sec) a process can idle before being swapped");
475 
476 /*
477  * Swapout is driven by the pageout daemon.  Very simple, we find eligible
478  * procs and mark them as being swapped out.  This will cause the kernel
479  * to prefer to pageout those proc's pages first and the procs in question
480  * will not return to user mode until the swapper tells them they can.
481  *
482  * If any procs have been sleeping/stopped for at least maxslp seconds,
483  * they are swapped.  Else, we swap the longest-sleeping or stopped process,
484  * if any, otherwise the longest-resident process.
485  */
486 
487 static int swapout_procs_callback(struct proc *p, void *data);
488 
489 /*
490  * No requirements.
491  */
492 void
493 swapout_procs(int action)
494 {
495 	allproc_scan(swapout_procs_callback, &action);
496 }
497 
498 static int
499 swapout_procs_callback(struct proc *p, void *data)
500 {
501 	struct lwp *lp;
502 	int action = *(int *)data;
503 	int minslp = -1;
504 
505 	if (!swappable(p))
506 		return(0);
507 
508 	lwkt_gettoken(&p->p_token);
509 
510 	/*
511 	 * We only consider active processes.
512 	 */
513 	if (p->p_stat != SACTIVE && p->p_stat != SSTOP) {
514 		lwkt_reltoken(&p->p_token);
515 		return(0);
516 	}
517 
518 	FOREACH_LWP_IN_PROC(lp, p) {
519 		/*
520 		 * do not swap out a realtime process
521 		 */
522 		if (RTP_PRIO_IS_REALTIME(lp->lwp_rtprio.type)) {
523 			lwkt_reltoken(&p->p_token);
524 			return(0);
525 		}
526 
527 		/*
528 		 * Guarentee swap_idle_threshold time in memory
529 		 */
530 		if (lp->lwp_slptime < swap_idle_threshold1) {
531 			lwkt_reltoken(&p->p_token);
532 			return(0);
533 		}
534 
535 		/*
536 		 * If the system is under memory stress, or if we
537 		 * are swapping idle processes >= swap_idle_threshold2,
538 		 * then swap the process out.
539 		 */
540 		if (((action & VM_SWAP_NORMAL) == 0) &&
541 		    (((action & VM_SWAP_IDLE) == 0) ||
542 		     (lp->lwp_slptime < swap_idle_threshold2))) {
543 			lwkt_reltoken(&p->p_token);
544 			return(0);
545 		}
546 
547 		if (minslp == -1 || lp->lwp_slptime < minslp)
548 			minslp = lp->lwp_slptime;
549 	}
550 
551 	/*
552 	 * If the process has been asleep for awhile, swap
553 	 * it out.
554 	 */
555 	if ((action & VM_SWAP_NORMAL) ||
556 	    ((action & VM_SWAP_IDLE) &&
557 	     (minslp > swap_idle_threshold2))) {
558 		swapout(p);
559 	}
560 
561 	/*
562 	 * cleanup our reference
563 	 */
564 	lwkt_reltoken(&p->p_token);
565 
566 	return(0);
567 }
568 
569 /*
570  * The caller must hold p->p_token
571  */
572 static void
573 swapout(struct proc *p)
574 {
575 #ifdef INVARIANTS
576 	if (swap_debug)
577 		kprintf("swapping out %d (%s)\n", p->p_pid, p->p_comm);
578 #endif
579 	++p->p_ru.ru_nswap;
580 
581 	/*
582 	 * remember the process resident count
583 	 */
584 	p->p_vmspace->vm_swrss = vmspace_resident_count(p->p_vmspace);
585 	p->p_flags |= P_SWAPPEDOUT;
586 	p->p_swtime = 0;
587 }
588 
589 #endif /* !NO_SWAPPING */
590 
591