xref: /freebsd/sys/vm/vm_page.c (revision 6e8394b8baa7d5d9153ab90de6824bcd19b3b4e1)
1 /*
2  * Copyright (c) 1991 Regents of the University of California.
3  * 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_page.c	7.4 (Berkeley) 5/7/91
37  *	$Id: vm_page.c,v 1.132 1999/06/20 21:47:02 alc Exp $
38  */
39 
40 /*
41  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
42  * All rights reserved.
43  *
44  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
45  *
46  * Permission to use, copy, modify and distribute this software and
47  * its documentation is hereby granted, provided that both the copyright
48  * notice and this permission notice appear in all copies of the
49  * software, derivative works or modified versions, and any portions
50  * thereof, and that both notices appear in supporting documentation.
51  *
52  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
53  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
54  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
55  *
56  * Carnegie Mellon requests users of this software to return to
57  *
58  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
59  *  School of Computer Science
60  *  Carnegie Mellon University
61  *  Pittsburgh PA 15213-3890
62  *
63  * any improvements or extensions that they make and grant Carnegie the
64  * rights to redistribute these changes.
65  */
66 
67 /*
68  *	Resident memory management module.
69  */
70 
71 #include <sys/param.h>
72 #include <sys/systm.h>
73 #include <sys/malloc.h>
74 #include <sys/proc.h>
75 #include <sys/vmmeter.h>
76 #include <sys/vnode.h>
77 
78 #include <vm/vm.h>
79 #include <vm/vm_param.h>
80 #include <vm/vm_prot.h>
81 #include <sys/lock.h>
82 #include <vm/vm_kern.h>
83 #include <vm/vm_object.h>
84 #include <vm/vm_page.h>
85 #include <vm/vm_pageout.h>
86 #include <vm/vm_pager.h>
87 #include <vm/vm_extern.h>
88 
89 static void	vm_page_queue_init __P((void));
90 static vm_page_t vm_page_select_cache __P((vm_object_t, vm_pindex_t));
91 
92 /*
93  *	Associated with page of user-allocatable memory is a
94  *	page structure.
95  */
96 
97 static struct vm_page **vm_page_buckets; /* Array of buckets */
98 static int vm_page_bucket_count;	/* How big is array? */
99 static int vm_page_hash_mask;		/* Mask for hash function */
100 static volatile int vm_page_bucket_generation;
101 
102 struct pglist vm_page_queue_free[PQ_L2_SIZE] = {{0}};
103 struct pglist vm_page_queue_active = {0};
104 struct pglist vm_page_queue_inactive = {0};
105 struct pglist vm_page_queue_cache[PQ_L2_SIZE] = {{0}};
106 
107 static int no_queue=0;
108 
109 struct vpgqueues vm_page_queues[PQ_COUNT] = {{0}};
110 static int pqcnt[PQ_COUNT] = {0};
111 
112 static void
113 vm_page_queue_init(void) {
114 	int i;
115 
116 	vm_page_queues[PQ_NONE].pl = NULL;
117 	vm_page_queues[PQ_NONE].cnt = &no_queue;
118 	for(i=0;i<PQ_L2_SIZE;i++) {
119 		vm_page_queues[PQ_FREE+i].pl = &vm_page_queue_free[i];
120 		vm_page_queues[PQ_FREE+i].cnt = &cnt.v_free_count;
121 	}
122 	vm_page_queues[PQ_INACTIVE].pl = &vm_page_queue_inactive;
123 	vm_page_queues[PQ_INACTIVE].cnt = &cnt.v_inactive_count;
124 
125 	vm_page_queues[PQ_ACTIVE].pl = &vm_page_queue_active;
126 	vm_page_queues[PQ_ACTIVE].cnt = &cnt.v_active_count;
127 	for(i=0;i<PQ_L2_SIZE;i++) {
128 		vm_page_queues[PQ_CACHE+i].pl = &vm_page_queue_cache[i];
129 		vm_page_queues[PQ_CACHE+i].cnt = &cnt.v_cache_count;
130 	}
131 	for(i=0;i<PQ_COUNT;i++) {
132 		if (vm_page_queues[i].pl) {
133 			TAILQ_INIT(vm_page_queues[i].pl);
134 		} else if (i != 0) {
135 			panic("vm_page_queue_init: queue %d is null", i);
136 		}
137 		vm_page_queues[i].lcnt = &pqcnt[i];
138 	}
139 }
140 
141 vm_page_t vm_page_array = 0;
142 static int vm_page_array_size = 0;
143 long first_page = 0;
144 int vm_page_zero_count = 0;
145 
146 static __inline int vm_page_hash __P((vm_object_t object, vm_pindex_t pindex));
147 static void vm_page_free_wakeup __P((void));
148 
149 /*
150  *	vm_set_page_size:
151  *
152  *	Sets the page size, perhaps based upon the memory
153  *	size.  Must be called before any use of page-size
154  *	dependent functions.
155  */
156 void
157 vm_set_page_size()
158 {
159 	if (cnt.v_page_size == 0)
160 		cnt.v_page_size = PAGE_SIZE;
161 	if (((cnt.v_page_size - 1) & cnt.v_page_size) != 0)
162 		panic("vm_set_page_size: page size not a power of two");
163 }
164 
165 /*
166  *	vm_page_startup:
167  *
168  *	Initializes the resident memory module.
169  *
170  *	Allocates memory for the page cells, and
171  *	for the object/offset-to-page hash table headers.
172  *	Each page cell is initialized and placed on the free list.
173  */
174 
175 vm_offset_t
176 vm_page_startup(starta, enda, vaddr)
177 	register vm_offset_t starta;
178 	vm_offset_t enda;
179 	register vm_offset_t vaddr;
180 {
181 	register vm_offset_t mapped;
182 	register vm_page_t m;
183 	register struct vm_page **bucket;
184 	vm_size_t npages, page_range;
185 	register vm_offset_t new_start;
186 	int i;
187 	vm_offset_t pa;
188 	int nblocks;
189 	vm_offset_t first_managed_page;
190 
191 	/* the biggest memory array is the second group of pages */
192 	vm_offset_t start;
193 	vm_offset_t biggestone, biggestsize;
194 
195 	vm_offset_t total;
196 
197 	total = 0;
198 	biggestsize = 0;
199 	biggestone = 0;
200 	nblocks = 0;
201 	vaddr = round_page(vaddr);
202 
203 	for (i = 0; phys_avail[i + 1]; i += 2) {
204 		phys_avail[i] = round_page(phys_avail[i]);
205 		phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
206 	}
207 
208 	for (i = 0; phys_avail[i + 1]; i += 2) {
209 		int size = phys_avail[i + 1] - phys_avail[i];
210 
211 		if (size > biggestsize) {
212 			biggestone = i;
213 			biggestsize = size;
214 		}
215 		++nblocks;
216 		total += size;
217 	}
218 
219 	start = phys_avail[biggestone];
220 
221 	/*
222 	 * Initialize the queue headers for the free queue, the active queue
223 	 * and the inactive queue.
224 	 */
225 
226 	vm_page_queue_init();
227 
228 	/*
229 	 * Allocate (and initialize) the hash table buckets.
230 	 *
231 	 * The number of buckets MUST BE a power of 2, and the actual value is
232 	 * the next power of 2 greater than the number of physical pages in
233 	 * the system.
234 	 *
235 	 * We make the hash table approximately 2x the number of pages to
236 	 * reduce the chain length.  This is about the same size using the
237 	 * singly-linked list as the 1x hash table we were using before
238 	 * using TAILQ but the chain length will be smaller.
239 	 *
240 	 * Note: This computation can be tweaked if desired.
241 	 */
242 	vm_page_buckets = (struct vm_page **)vaddr;
243 	bucket = vm_page_buckets;
244 	if (vm_page_bucket_count == 0) {
245 		vm_page_bucket_count = 1;
246 		while (vm_page_bucket_count < atop(total))
247 			vm_page_bucket_count <<= 1;
248 	}
249 	vm_page_bucket_count <<= 1;
250 	vm_page_hash_mask = vm_page_bucket_count - 1;
251 
252 	/*
253 	 * Validate these addresses.
254 	 */
255 
256 	new_start = start + vm_page_bucket_count * sizeof(struct vm_page *);
257 	new_start = round_page(new_start);
258 	mapped = round_page(vaddr);
259 	vaddr = pmap_map(mapped, start, new_start,
260 	    VM_PROT_READ | VM_PROT_WRITE);
261 	start = new_start;
262 	vaddr = round_page(vaddr);
263 	bzero((caddr_t) mapped, vaddr - mapped);
264 
265 	for (i = 0; i < vm_page_bucket_count; i++) {
266 		*bucket = NULL;
267 		bucket++;
268 	}
269 
270 	/*
271 	 * Compute the number of pages of memory that will be available for
272 	 * use (taking into account the overhead of a page structure per
273 	 * page).
274 	 */
275 
276 	first_page = phys_avail[0] / PAGE_SIZE;
277 
278 	page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page;
279 	npages = (total - (page_range * sizeof(struct vm_page)) -
280 	    (start - phys_avail[biggestone])) / PAGE_SIZE;
281 
282 	/*
283 	 * Initialize the mem entry structures now, and put them in the free
284 	 * queue.
285 	 */
286 	vm_page_array = (vm_page_t) vaddr;
287 	mapped = vaddr;
288 
289 	/*
290 	 * Validate these addresses.
291 	 */
292 	new_start = round_page(start + page_range * sizeof(struct vm_page));
293 	mapped = pmap_map(mapped, start, new_start,
294 	    VM_PROT_READ | VM_PROT_WRITE);
295 	start = new_start;
296 
297 	first_managed_page = start / PAGE_SIZE;
298 
299 	/*
300 	 * Clear all of the page structures
301 	 */
302 	bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
303 	vm_page_array_size = page_range;
304 
305 	/*
306 	 * Construct the free queue(s) in descending order (by physical
307 	 * address) so that the first 16MB of physical memory is allocated
308 	 * last rather than first.  On large-memory machines, this avoids
309 	 * the exhaustion of low physical memory before isa_dmainit has run.
310 	 */
311 	cnt.v_page_count = 0;
312 	cnt.v_free_count = 0;
313 	for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) {
314 		if (i == biggestone)
315 			pa = ptoa(first_managed_page);
316 		else
317 			pa = phys_avail[i];
318 		while (pa < phys_avail[i + 1] && npages-- > 0) {
319 			++cnt.v_page_count;
320 			++cnt.v_free_count;
321 			m = PHYS_TO_VM_PAGE(pa);
322 			m->phys_addr = pa;
323 			m->flags = 0;
324 			m->pc = (pa >> PAGE_SHIFT) & PQ_L2_MASK;
325 			m->queue = m->pc + PQ_FREE;
326 			TAILQ_INSERT_HEAD(vm_page_queues[m->queue].pl, m, pageq);
327 			++(*vm_page_queues[m->queue].lcnt);
328 			pa += PAGE_SIZE;
329 		}
330 	}
331 	return (mapped);
332 }
333 
334 /*
335  *	vm_page_hash:
336  *
337  *	Distributes the object/offset key pair among hash buckets.
338  *
339  *	NOTE:  This macro depends on vm_page_bucket_count being a power of 2.
340  *	This routine may not block.
341  *
342  *	We try to randomize the hash based on the object to spread the pages
343  *	out in the hash table without it costing us too much.
344  */
345 static __inline int
346 vm_page_hash(object, pindex)
347 	vm_object_t object;
348 	vm_pindex_t pindex;
349 {
350 	int i = ((uintptr_t)object + pindex) ^ object->hash_rand;
351 
352 	return(i & vm_page_hash_mask);
353 }
354 
355 /*
356  *	vm_page_insert:		[ internal use only ]
357  *
358  *	Inserts the given mem entry into the object and object list.
359  *
360  *	The pagetables are not updated but will presumably fault the page
361  *	in if necessary, or if a kernel page the caller will at some point
362  *	enter the page into the kernel's pmap.  We are not allowed to block
363  *	here so we *can't* do this anyway.
364  *
365  *	The object and page must be locked, and must be splhigh.
366  *	This routine may not block.
367  */
368 
369 void
370 vm_page_insert(m, object, pindex)
371 	register vm_page_t m;
372 	register vm_object_t object;
373 	register vm_pindex_t pindex;
374 {
375 	register struct vm_page **bucket;
376 
377 	if (m->object != NULL)
378 		panic("vm_page_insert: already inserted");
379 
380 	/*
381 	 * Record the object/offset pair in this page
382 	 */
383 
384 	m->object = object;
385 	m->pindex = pindex;
386 
387 	/*
388 	 * Insert it into the object_object/offset hash table
389 	 */
390 
391 	bucket = &vm_page_buckets[vm_page_hash(object, pindex)];
392 	m->hnext = *bucket;
393 	*bucket = m;
394 	vm_page_bucket_generation++;
395 
396 	/*
397 	 * Now link into the object's list of backed pages.
398 	 */
399 
400 	TAILQ_INSERT_TAIL(&object->memq, m, listq);
401 	m->object->generation++;
402 
403 	/*
404 	 * show that the object has one more resident page.
405 	 */
406 
407 	object->resident_page_count++;
408 
409 	/*
410 	 * Since we are inserting a new and possibly dirty page,
411 	 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags.
412 	 */
413 	if (m->flags & PG_WRITEABLE)
414 	    vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
415 }
416 
417 /*
418  *	vm_page_remove:
419  *				NOTE: used by device pager as well -wfj
420  *
421  *	Removes the given mem entry from the object/offset-page
422  *	table and the object page list, but do not invalidate/terminate
423  *	the backing store.
424  *
425  *	The object and page must be locked, and at splhigh.
426  *	The underlying pmap entry (if any) is NOT removed here.
427  *	This routine may not block.
428  */
429 
430 void
431 vm_page_remove(m)
432 	vm_page_t m;
433 {
434 	vm_object_t object;
435 
436 	if (m->object == NULL)
437 		return;
438 
439 #if !defined(MAX_PERF)
440 	if ((m->flags & PG_BUSY) == 0) {
441 		panic("vm_page_remove: page not busy");
442 	}
443 #endif
444 
445 	/*
446 	 * Basically destroy the page.
447 	 */
448 
449 	vm_page_wakeup(m);
450 
451 	object = m->object;
452 
453 	/*
454 	 * Remove from the object_object/offset hash table.  The object
455 	 * must be on the hash queue, we will panic if it isn't
456 	 *
457 	 * Note: we must NULL-out m->hnext to prevent loops in detached
458 	 * buffers with vm_page_lookup().
459 	 */
460 
461 	{
462 		struct vm_page **bucket;
463 
464 		bucket = &vm_page_buckets[vm_page_hash(m->object, m->pindex)];
465 		while (*bucket != m) {
466 #if !defined(MAX_PERF)
467 			if (*bucket == NULL)
468 				panic("vm_page_remove(): page not found in hash");
469 #endif
470 			bucket = &(*bucket)->hnext;
471 		}
472 		*bucket = m->hnext;
473 		m->hnext = NULL;
474 		vm_page_bucket_generation++;
475 	}
476 
477 	/*
478 	 * Now remove from the object's list of backed pages.
479 	 */
480 
481 	TAILQ_REMOVE(&object->memq, m, listq);
482 
483 	/*
484 	 * And show that the object has one fewer resident page.
485 	 */
486 
487 	object->resident_page_count--;
488 	object->generation++;
489 
490 	m->object = NULL;
491 }
492 
493 /*
494  *	vm_page_lookup:
495  *
496  *	Returns the page associated with the object/offset
497  *	pair specified; if none is found, NULL is returned.
498  *
499  *	NOTE: the code below does not lock.  It will operate properly if
500  *	an interrupt makes a change, but the generation algorithm will not
501  *	operate properly in an SMP environment where both cpu's are able to run
502  *	kernel code simultaniously.
503  *
504  *	The object must be locked.  No side effects.
505  *	This routine may not block.
506  *	This is a critical path routine
507  */
508 
509 vm_page_t
510 vm_page_lookup(object, pindex)
511 	register vm_object_t object;
512 	register vm_pindex_t pindex;
513 {
514 	register vm_page_t m;
515 	register struct vm_page **bucket;
516 	int generation;
517 
518 	/*
519 	 * Search the hash table for this object/offset pair
520 	 */
521 
522 retry:
523 	generation = vm_page_bucket_generation;
524 	bucket = &vm_page_buckets[vm_page_hash(object, pindex)];
525 	for (m = *bucket; m != NULL; m = m->hnext) {
526 		if ((m->object == object) && (m->pindex == pindex)) {
527 			if (vm_page_bucket_generation != generation)
528 				goto retry;
529 			return (m);
530 		}
531 	}
532 	if (vm_page_bucket_generation != generation)
533 		goto retry;
534 	return (NULL);
535 }
536 
537 /*
538  *	vm_page_rename:
539  *
540  *	Move the given memory entry from its
541  *	current object to the specified target object/offset.
542  *
543  *	The object must be locked.
544  *	This routine may not block.
545  *
546  *	Note: this routine will raise itself to splvm(), the caller need not.
547  *
548  *	Note: swap associated with the page must be invalidated by the move.  We
549  *	      have to do this for several reasons:  (1) we aren't freeing the
550  *	      page, (2) we are dirtying the page, (3) the VM system is probably
551  *	      moving the page from object A to B, and will then later move
552  *	      the backing store from A to B and we can't have a conflict.
553  *
554  *	Note: we *always* dirty the page.  It is necessary both for the
555  *	      fact that we moved it, and because we may be invalidating
556  *	      swap.  If the page is on the cache, we have to deactivate it
557  *	      or vm_page_dirty() will panic.  Dirty pages are not allowed
558  *	      on the cache.
559  */
560 
561 void
562 vm_page_rename(m, new_object, new_pindex)
563 	register vm_page_t m;
564 	register vm_object_t new_object;
565 	vm_pindex_t new_pindex;
566 {
567 	int s;
568 
569 	s = splvm();
570 	vm_page_remove(m);
571 	vm_page_insert(m, new_object, new_pindex);
572 	if (m->queue - m->pc == PQ_CACHE)
573 		vm_page_deactivate(m);
574 	vm_page_dirty(m);
575 	splx(s);
576 }
577 
578 /*
579  * vm_page_unqueue_nowakeup:
580  *
581  * 	vm_page_unqueue() without any wakeup
582  *
583  *	This routine must be called at splhigh().
584  *	This routine may not block.
585  */
586 
587 void
588 vm_page_unqueue_nowakeup(m)
589 	vm_page_t m;
590 {
591 	int queue = m->queue;
592 	struct vpgqueues *pq;
593 	if (queue != PQ_NONE) {
594 		pq = &vm_page_queues[queue];
595 		m->queue = PQ_NONE;
596 		TAILQ_REMOVE(pq->pl, m, pageq);
597 		(*pq->cnt)--;
598 		(*pq->lcnt)--;
599 	}
600 }
601 
602 /*
603  * vm_page_unqueue:
604  *
605  *	Remove a page from its queue.
606  *
607  *	This routine must be called at splhigh().
608  *	This routine may not block.
609  */
610 
611 void
612 vm_page_unqueue(m)
613 	vm_page_t m;
614 {
615 	int queue = m->queue;
616 	struct vpgqueues *pq;
617 	if (queue != PQ_NONE) {
618 		m->queue = PQ_NONE;
619 		pq = &vm_page_queues[queue];
620 		TAILQ_REMOVE(pq->pl, m, pageq);
621 		(*pq->cnt)--;
622 		(*pq->lcnt)--;
623 		if ((queue - m->pc) == PQ_CACHE) {
624 			if ((cnt.v_cache_count + cnt.v_free_count) <
625 				(cnt.v_free_reserved + cnt.v_cache_min))
626 				pagedaemon_wakeup();
627 		}
628 	}
629 }
630 
631 #if PQ_L2_SIZE > 1
632 
633 /*
634  *	vm_page_list_find:
635  *
636  *	Find a page on the specified queue with color optimization.
637  *
638  *	The page coloring optimization attempts to locate a page
639  *	that does not overload other nearby pages in the object in
640  *	the cpu's L1 or L2 caches.  We need this optmization because
641  *	cpu caches tend to be physical caches, while object spaces tend
642  *	to be virtual.
643  *
644  *	This routine must be called at splvm().
645  *	This routine may not block.
646  *
647  *	This routine may only be called from the vm_page_list_find() macro
648  *	in vm_page.h
649  */
650 vm_page_t
651 _vm_page_list_find(basequeue, index)
652 	int basequeue, index;
653 {
654 	int i;
655 	vm_page_t m = NULL;
656 	struct vpgqueues *pq;
657 
658 	pq = &vm_page_queues[basequeue];
659 
660 	/*
661 	 * Note that for the first loop, index+i and index-i wind up at the
662 	 * same place.  Even though this is not totally optimal, we've already
663 	 * blown it by missing the cache case so we do not care.
664 	 */
665 
666 	for(i = PQ_L2_SIZE / 2; i > 0; --i) {
667 		if ((m = TAILQ_FIRST(pq[(index + i) & PQ_L2_MASK].pl)) != NULL)
668 			break;
669 
670 		if ((m = TAILQ_FIRST(pq[(index - i) & PQ_L2_MASK].pl)) != NULL)
671 			break;
672 	}
673 	return(m);
674 }
675 
676 #endif
677 
678 /*
679  *	vm_page_select_cache:
680  *
681  *	Find a page on the cache queue with color optimization.  As pages
682  *	might be found, but not applicable, they are deactivated.  This
683  *	keeps us from using potentially busy cached pages.
684  *
685  *	This routine must be called at splvm().
686  *	This routine may not block.
687  */
688 vm_page_t
689 vm_page_select_cache(object, pindex)
690 	vm_object_t object;
691 	vm_pindex_t pindex;
692 {
693 	vm_page_t m;
694 
695 	while (TRUE) {
696 		m = vm_page_list_find(
697 		    PQ_CACHE,
698 		    (pindex + object->pg_color) & PQ_L2_MASK,
699 		    FALSE
700 		);
701 		if (m && ((m->flags & PG_BUSY) || m->busy ||
702 			       m->hold_count || m->wire_count)) {
703 			vm_page_deactivate(m);
704 			continue;
705 		}
706 		return m;
707 	}
708 }
709 
710 /*
711  *	vm_page_select_free:
712  *
713  *	Find a free or zero page, with specified preference.  We attempt to
714  *	inline the nominal case and fall back to _vm_page_select_free()
715  *	otherwise.
716  *
717  *	This routine must be called at splvm().
718  *	This routine may not block.
719  */
720 
721 static __inline vm_page_t
722 vm_page_select_free(vm_object_t object, vm_pindex_t pindex, boolean_t prefer_zero)
723 {
724 	vm_page_t m;
725 
726 	m = vm_page_list_find(
727 		PQ_FREE,
728 		(pindex + object->pg_color) & PQ_L2_MASK,
729 		prefer_zero
730 	);
731 	return(m);
732 }
733 
734 /*
735  *	vm_page_alloc:
736  *
737  *	Allocate and return a memory cell associated
738  *	with this VM object/offset pair.
739  *
740  *	page_req classes:
741  *	VM_ALLOC_NORMAL		normal process request
742  *	VM_ALLOC_SYSTEM		system *really* needs a page
743  *	VM_ALLOC_INTERRUPT	interrupt time request
744  *	VM_ALLOC_ZERO		zero page
745  *
746  *	Object must be locked.
747  *	This routine may not block.
748  *
749  *	Additional special handling is required when called from an
750  *	interrupt (VM_ALLOC_INTERRUPT).  We are not allowed to mess with
751  *	the page cache in this case.
752  */
753 
754 vm_page_t
755 vm_page_alloc(object, pindex, page_req)
756 	vm_object_t object;
757 	vm_pindex_t pindex;
758 	int page_req;
759 {
760 	register vm_page_t m = NULL;
761 	int s;
762 
763 	KASSERT(!vm_page_lookup(object, pindex),
764 		("vm_page_alloc: page already allocated"));
765 
766 	/*
767 	 * The pager is allowed to eat deeper into the free page list.
768 	 */
769 
770 	if ((curproc == pageproc) && (page_req != VM_ALLOC_INTERRUPT)) {
771 		page_req = VM_ALLOC_SYSTEM;
772 	};
773 
774 	s = splvm();
775 
776 loop:
777 	if (cnt.v_free_count > cnt.v_free_reserved) {
778 		/*
779 		 * Allocate from the free queue if there are plenty of pages
780 		 * in it.
781 		 */
782 		if (page_req == VM_ALLOC_ZERO)
783 			m = vm_page_select_free(object, pindex, TRUE);
784 		else
785 			m = vm_page_select_free(object, pindex, FALSE);
786 	} else if (
787 	    (page_req == VM_ALLOC_SYSTEM &&
788 	     cnt.v_cache_count == 0 &&
789 	     cnt.v_free_count > cnt.v_interrupt_free_min) ||
790 	    (page_req == VM_ALLOC_INTERRUPT && cnt.v_free_count > 0)
791 	) {
792 		/*
793 		 * Interrupt or system, dig deeper into the free list.
794 		 */
795 		m = vm_page_select_free(object, pindex, FALSE);
796 	} else if (page_req != VM_ALLOC_INTERRUPT) {
797 		/*
798 		 * Allocateable from cache (non-interrupt only).  On success,
799 		 * we must free the page and try again, thus ensuring that
800 		 * cnt.v_*_free_min counters are replenished.
801 		 */
802 		m = vm_page_select_cache(object, pindex);
803 		if (m == NULL) {
804 			splx(s);
805 #if defined(DIAGNOSTIC)
806 			if (cnt.v_cache_count > 0)
807 				printf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", cnt.v_cache_count);
808 #endif
809 			vm_pageout_deficit++;
810 			pagedaemon_wakeup();
811 			return (NULL);
812 		}
813 		KASSERT(m->dirty == 0, ("Found dirty cache page %p", m));
814 		vm_page_busy(m);
815 		vm_page_protect(m, VM_PROT_NONE);
816 		vm_page_free(m);
817 		goto loop;
818 	} else {
819 		/*
820 		 * Not allocateable from cache from interrupt, give up.
821 		 */
822 		splx(s);
823 		vm_pageout_deficit++;
824 		pagedaemon_wakeup();
825 		return (NULL);
826 	}
827 
828 	/*
829 	 *  At this point we had better have found a good page.
830 	 */
831 
832 	KASSERT(
833 	    m != NULL,
834 	    ("vm_page_alloc(): missing page on free queue\n")
835 	);
836 
837 	/*
838 	 * Remove from free queue
839 	 */
840 
841 	{
842 		struct vpgqueues *pq = &vm_page_queues[m->queue];
843 
844 		TAILQ_REMOVE(pq->pl, m, pageq);
845 		(*pq->cnt)--;
846 		(*pq->lcnt)--;
847 	}
848 
849 	/*
850 	 * Initialize structure.  Only the PG_ZERO flag is inherited.
851 	 */
852 
853 	if (m->flags & PG_ZERO) {
854 		vm_page_zero_count--;
855 		m->flags = PG_ZERO | PG_BUSY;
856 	} else {
857 		m->flags = PG_BUSY;
858 	}
859 	m->wire_count = 0;
860 	m->hold_count = 0;
861 	m->act_count = 0;
862 	m->busy = 0;
863 	m->valid = 0;
864 	m->dirty = 0;
865 	m->queue = PQ_NONE;
866 
867 	/*
868 	 * vm_page_insert() is safe prior to the splx().  Note also that
869 	 * inserting a page here does not insert it into the pmap (which
870 	 * could cause us to block allocating memory).  We cannot block
871 	 * anywhere.
872 	 */
873 
874 	vm_page_insert(m, object, pindex);
875 
876 	/*
877 	 * Don't wakeup too often - wakeup the pageout daemon when
878 	 * we would be nearly out of memory.
879 	 */
880 	if (((cnt.v_free_count + cnt.v_cache_count) <
881 		(cnt.v_free_reserved + cnt.v_cache_min)) ||
882 			(cnt.v_free_count < cnt.v_pageout_free_min))
883 		pagedaemon_wakeup();
884 
885 	splx(s);
886 
887 	return (m);
888 }
889 
890 /*
891  *	vm_wait:	(also see VM_WAIT macro)
892  *
893  *	Block until free pages are available for allocation
894  */
895 
896 void
897 vm_wait()
898 {
899 	int s;
900 
901 	s = splvm();
902 	if (curproc == pageproc) {
903 		vm_pageout_pages_needed = 1;
904 		tsleep(&vm_pageout_pages_needed, PSWP, "vmwait", 0);
905 	} else {
906 		if (!vm_pages_needed) {
907 			vm_pages_needed++;
908 			wakeup(&vm_pages_needed);
909 		}
910 		tsleep(&cnt.v_free_count, PVM, "vmwait", 0);
911 	}
912 	splx(s);
913 }
914 
915 /*
916  *	vm_await:	(also see VM_AWAIT macro)
917  *
918  *	asleep on an event that will signal when free pages are available
919  *	for allocation.
920  */
921 
922 void
923 vm_await()
924 {
925 	int s;
926 
927 	s = splvm();
928 	if (curproc == pageproc) {
929 		vm_pageout_pages_needed = 1;
930 		asleep(&vm_pageout_pages_needed, PSWP, "vmwait", 0);
931 	} else {
932 		if (!vm_pages_needed) {
933 			vm_pages_needed++;
934 			wakeup(&vm_pages_needed);
935 		}
936 		asleep(&cnt.v_free_count, PVM, "vmwait", 0);
937 	}
938 	splx(s);
939 }
940 
941 #if 0
942 /*
943  *	vm_page_sleep:
944  *
945  *	Block until page is no longer busy.
946  */
947 
948 int
949 vm_page_sleep(vm_page_t m, char *msg, char *busy) {
950 	int slept = 0;
951 	if ((busy && *busy) || (m->flags & PG_BUSY)) {
952 		int s;
953 		s = splvm();
954 		if ((busy && *busy) || (m->flags & PG_BUSY)) {
955 			vm_page_flag_set(m, PG_WANTED);
956 			tsleep(m, PVM, msg, 0);
957 			slept = 1;
958 		}
959 		splx(s);
960 	}
961 	return slept;
962 }
963 
964 #endif
965 
966 #if 0
967 
968 /*
969  *	vm_page_asleep:
970  *
971  *	Similar to vm_page_sleep(), but does not block.  Returns 0 if
972  *	the page is not busy, or 1 if the page is busy.
973  *
974  *	This routine has the side effect of calling asleep() if the page
975  *	was busy (1 returned).
976  */
977 
978 int
979 vm_page_asleep(vm_page_t m, char *msg, char *busy) {
980 	int slept = 0;
981 	if ((busy && *busy) || (m->flags & PG_BUSY)) {
982 		int s;
983 		s = splvm();
984 		if ((busy && *busy) || (m->flags & PG_BUSY)) {
985 			vm_page_flag_set(m, PG_WANTED);
986 			asleep(m, PVM, msg, 0);
987 			slept = 1;
988 		}
989 		splx(s);
990 	}
991 	return slept;
992 }
993 
994 #endif
995 
996 /*
997  *	vm_page_activate:
998  *
999  *	Put the specified page on the active list (if appropriate).
1000  *
1001  *	The page queues must be locked.
1002  *	This routine may not block.
1003  */
1004 void
1005 vm_page_activate(m)
1006 	register vm_page_t m;
1007 {
1008 	int s;
1009 
1010 	s = splvm();
1011 	if (m->queue != PQ_ACTIVE) {
1012 		if ((m->queue - m->pc) == PQ_CACHE)
1013 			cnt.v_reactivated++;
1014 
1015 		vm_page_unqueue(m);
1016 
1017 		if (m->wire_count == 0) {
1018 			m->queue = PQ_ACTIVE;
1019 			++(*vm_page_queues[PQ_ACTIVE].lcnt);
1020 			TAILQ_INSERT_TAIL(&vm_page_queue_active, m, pageq);
1021 			if (m->act_count < ACT_INIT)
1022 				m->act_count = ACT_INIT;
1023 			cnt.v_active_count++;
1024 		}
1025 	} else {
1026 		if (m->act_count < ACT_INIT)
1027 			m->act_count = ACT_INIT;
1028 	}
1029 
1030 	splx(s);
1031 }
1032 
1033 /*
1034  *	vm_page_free_wakeup:
1035  *
1036  *	Helper routine for vm_page_free_toq() and vm_page_cache().  This
1037  *	routine is called when a page has been added to the cache or free
1038  *	queues.
1039  *
1040  *	This routine may not block.
1041  *	This routine must be called at splvm()
1042  */
1043 static __inline void
1044 vm_page_free_wakeup()
1045 {
1046 	/*
1047 	 * if pageout daemon needs pages, then tell it that there are
1048 	 * some free.
1049 	 */
1050 	if (vm_pageout_pages_needed) {
1051 		wakeup(&vm_pageout_pages_needed);
1052 		vm_pageout_pages_needed = 0;
1053 	}
1054 	/*
1055 	 * wakeup processes that are waiting on memory if we hit a
1056 	 * high water mark. And wakeup scheduler process if we have
1057 	 * lots of memory. this process will swapin processes.
1058 	 */
1059 	if (vm_pages_needed &&
1060 		((cnt.v_free_count + cnt.v_cache_count) >= cnt.v_free_min)) {
1061 		wakeup(&cnt.v_free_count);
1062 		vm_pages_needed = 0;
1063 	}
1064 }
1065 
1066 /*
1067  *	vm_page_free_toq:
1068  *
1069  *	Returns the given page to the PQ_FREE or PQ_ZERO list,
1070  *	disassociating it with any VM object.
1071  *
1072  *	Object and page must be locked prior to entry.
1073  *	This routine may not block.
1074  */
1075 
1076 void
1077 vm_page_free_toq(vm_page_t m)
1078 {
1079 	int s;
1080 	struct vpgqueues *pq;
1081 	vm_object_t object = m->object;
1082 
1083 	s = splvm();
1084 
1085 	cnt.v_tfree++;
1086 
1087 #if !defined(MAX_PERF)
1088 	if (m->busy || ((m->queue - m->pc) == PQ_FREE) ||
1089 		(m->hold_count != 0)) {
1090 		printf(
1091 		"vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n",
1092 		    (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0,
1093 		    m->hold_count);
1094 		if ((m->queue - m->pc) == PQ_FREE)
1095 			panic("vm_page_free: freeing free page");
1096 		else
1097 			panic("vm_page_free: freeing busy page");
1098 	}
1099 #endif
1100 
1101 	/*
1102 	 * unqueue, then remove page.  Note that we cannot destroy
1103 	 * the page here because we do not want to call the pager's
1104 	 * callback routine until after we've put the page on the
1105 	 * appropriate free queue.
1106 	 */
1107 
1108 	vm_page_unqueue_nowakeup(m);
1109 	vm_page_remove(m);
1110 
1111 	/*
1112 	 * If fictitious remove object association and
1113 	 * return, otherwise delay object association removal.
1114 	 */
1115 
1116 	if ((m->flags & PG_FICTITIOUS) != 0) {
1117 		splx(s);
1118 		return;
1119 	}
1120 
1121 	m->valid = 0;
1122 
1123 	if (m->wire_count != 0) {
1124 #if !defined(MAX_PERF)
1125 		if (m->wire_count > 1) {
1126 			panic("vm_page_free: invalid wire count (%d), pindex: 0x%x",
1127 				m->wire_count, m->pindex);
1128 		}
1129 #endif
1130 		printf("vm_page_free: freeing wired page\n");
1131 		m->wire_count = 0;
1132 		cnt.v_wire_count--;
1133 	}
1134 
1135 	/*
1136 	 * If we've exhausted the object's resident pages we want to free
1137 	 * it up.
1138 	 */
1139 
1140 	if (object &&
1141 	    (object->type == OBJT_VNODE) &&
1142 	    ((object->flags & OBJ_DEAD) == 0)
1143 	) {
1144 		struct vnode *vp = (struct vnode *)object->handle;
1145 
1146 		if (vp && VSHOULDFREE(vp)) {
1147 			if ((vp->v_flag & (VTBFREE|VDOOMED|VFREE)) == 0) {
1148 				TAILQ_INSERT_TAIL(&vnode_tobefree_list, vp, v_freelist);
1149 				vp->v_flag |= VTBFREE;
1150 			}
1151 		}
1152 	}
1153 
1154 #ifdef __alpha__
1155 	pmap_page_is_free(m);
1156 #endif
1157 
1158 	m->queue = PQ_FREE + m->pc;
1159 	pq = &vm_page_queues[m->queue];
1160 	++(*pq->lcnt);
1161 	++(*pq->cnt);
1162 
1163 	/*
1164 	 * Put zero'd pages on the end ( where we look for zero'd pages
1165 	 * first ) and non-zerod pages at the head.
1166 	 */
1167 
1168 	if (m->flags & PG_ZERO) {
1169 		TAILQ_INSERT_TAIL(pq->pl, m, pageq);
1170 		++vm_page_zero_count;
1171 	} else if (curproc == pageproc) {
1172 		/*
1173 		 * If the pageout daemon is freeing pages, the pages are
1174 		 * likely to NOT be in the L1 or L2 caches due to their age.
1175 		 * For now we do not try to do anything special with this
1176 		 * info.
1177 		 */
1178 		TAILQ_INSERT_HEAD(pq->pl, m, pageq);
1179 	} else {
1180 		TAILQ_INSERT_HEAD(pq->pl, m, pageq);
1181 	}
1182 
1183 	vm_page_free_wakeup();
1184 
1185 	splx(s);
1186 }
1187 
1188 /*
1189  *	vm_page_wire:
1190  *
1191  *	Mark this page as wired down by yet
1192  *	another map, removing it from paging queues
1193  *	as necessary.
1194  *
1195  *	The page queues must be locked.
1196  *	This routine may not block.
1197  */
1198 void
1199 vm_page_wire(m)
1200 	register vm_page_t m;
1201 {
1202 	int s;
1203 
1204 	s = splvm();
1205 	if (m->wire_count == 0) {
1206 		vm_page_unqueue(m);
1207 		cnt.v_wire_count++;
1208 	}
1209 	m->wire_count++;
1210 	splx(s);
1211 	(*vm_page_queues[PQ_NONE].lcnt)++;
1212 	vm_page_flag_set(m, PG_MAPPED);
1213 }
1214 
1215 /*
1216  *	vm_page_unwire:
1217  *
1218  *	Release one wiring of this page, potentially
1219  *	enabling it to be paged again.
1220  *
1221  *	Many pages placed on the inactive queue should actually go
1222  *	into the cache, but it is difficult to figure out which.  What
1223  *	we do instead, if the inactive target is well met, is to put
1224  *	clean pages at the head of the inactive queue instead of the tail.
1225  *	This will cause them to be moved to the cache more quickly and
1226  *	if not actively re-referenced, freed more quickly.  If we just
1227  *	stick these pages at the end of the inactive queue, heavy filesystem
1228  *	meta-data accesses can cause an unnecessary paging load on memory bound
1229  *	processes.  This optimization causes one-time-use metadata to be
1230  *	reused more quickly.
1231  *
1232  *	A number of routines use vm_page_unwire() to guarentee that the page
1233  *	will go into either the inactive or active queues, and will NEVER
1234  *	be placed in the cache - for example, just after dirtying a page.
1235  *	dirty pages in the cache are not allowed.
1236  *
1237  *	The page queues must be locked.
1238  *	This routine may not block.
1239  */
1240 void
1241 vm_page_unwire(m, activate)
1242 	register vm_page_t m;
1243 	int activate;
1244 {
1245 	int s;
1246 
1247 	s = splvm();
1248 
1249 	if (m->wire_count > 0) {
1250 		m->wire_count--;
1251 		if (m->wire_count == 0) {
1252 			cnt.v_wire_count--;
1253 			if (activate) {
1254 				TAILQ_INSERT_TAIL(&vm_page_queue_active, m, pageq);
1255 				m->queue = PQ_ACTIVE;
1256 				(*vm_page_queues[PQ_ACTIVE].lcnt)++;
1257 				cnt.v_active_count++;
1258 			} else {
1259 				TAILQ_INSERT_TAIL(&vm_page_queue_inactive, m, pageq);
1260 				m->queue = PQ_INACTIVE;
1261 				(*vm_page_queues[PQ_INACTIVE].lcnt)++;
1262 				cnt.v_inactive_count++;
1263 			}
1264 		}
1265 	} else {
1266 #if !defined(MAX_PERF)
1267 		panic("vm_page_unwire: invalid wire count: %d\n", m->wire_count);
1268 #endif
1269 	}
1270 	splx(s);
1271 }
1272 
1273 
1274 /*
1275  * Move the specified page to the inactive queue.  If the page has
1276  * any associated swap, the swap is deallocated.
1277  *
1278  * This routine may not block.
1279  */
1280 void
1281 vm_page_deactivate(m)
1282 	register vm_page_t m;
1283 {
1284 	int s;
1285 
1286 	/*
1287 	 * Ignore if already inactive.
1288 	 */
1289 	if (m->queue == PQ_INACTIVE)
1290 		return;
1291 
1292 	s = splvm();
1293 	if (m->wire_count == 0) {
1294 		if ((m->queue - m->pc) == PQ_CACHE)
1295 			cnt.v_reactivated++;
1296 		vm_page_unqueue(m);
1297 		TAILQ_INSERT_TAIL(&vm_page_queue_inactive, m, pageq);
1298 		m->queue = PQ_INACTIVE;
1299 		++(*vm_page_queues[PQ_INACTIVE].lcnt);
1300 		cnt.v_inactive_count++;
1301 	}
1302 	splx(s);
1303 }
1304 
1305 /*
1306  * vm_page_cache
1307  *
1308  * Put the specified page onto the page cache queue (if appropriate).
1309  *
1310  * This routine may not block.
1311  */
1312 void
1313 vm_page_cache(m)
1314 	register vm_page_t m;
1315 {
1316 	int s;
1317 
1318 #if !defined(MAX_PERF)
1319 	if ((m->flags & PG_BUSY) || m->busy || m->wire_count) {
1320 		printf("vm_page_cache: attempting to cache busy page\n");
1321 		return;
1322 	}
1323 #endif
1324 	if ((m->queue - m->pc) == PQ_CACHE)
1325 		return;
1326 
1327 	/*
1328 	 * Remove all pmaps and indicate that the page is not
1329 	 * writeable or mapped.
1330 	 */
1331 
1332 	vm_page_protect(m, VM_PROT_NONE);
1333 #if !defined(MAX_PERF)
1334 	if (m->dirty != 0) {
1335 		panic("vm_page_cache: caching a dirty page, pindex: %d", m->pindex);
1336 	}
1337 #endif
1338 	s = splvm();
1339 	vm_page_unqueue_nowakeup(m);
1340 	m->queue = PQ_CACHE + m->pc;
1341 	(*vm_page_queues[m->queue].lcnt)++;
1342 	TAILQ_INSERT_TAIL(vm_page_queues[m->queue].pl, m, pageq);
1343 	cnt.v_cache_count++;
1344 	vm_page_free_wakeup();
1345 	splx(s);
1346 }
1347 
1348 /*
1349  * Grab a page, waiting until we are waken up due to the page
1350  * changing state.  We keep on waiting, if the page continues
1351  * to be in the object.  If the page doesn't exist, allocate it.
1352  *
1353  * This routine may block.
1354  */
1355 vm_page_t
1356 vm_page_grab(object, pindex, allocflags)
1357 	vm_object_t object;
1358 	vm_pindex_t pindex;
1359 	int allocflags;
1360 {
1361 
1362 	vm_page_t m;
1363 	int s, generation;
1364 
1365 retrylookup:
1366 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
1367 		if (m->busy || (m->flags & PG_BUSY)) {
1368 			generation = object->generation;
1369 
1370 			s = splvm();
1371 			while ((object->generation == generation) &&
1372 					(m->busy || (m->flags & PG_BUSY))) {
1373 				vm_page_flag_set(m, PG_WANTED | PG_REFERENCED);
1374 				tsleep(m, PVM, "pgrbwt", 0);
1375 				if ((allocflags & VM_ALLOC_RETRY) == 0) {
1376 					splx(s);
1377 					return NULL;
1378 				}
1379 			}
1380 			splx(s);
1381 			goto retrylookup;
1382 		} else {
1383 			vm_page_busy(m);
1384 			return m;
1385 		}
1386 	}
1387 
1388 	m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY);
1389 	if (m == NULL) {
1390 		VM_WAIT;
1391 		if ((allocflags & VM_ALLOC_RETRY) == 0)
1392 			return NULL;
1393 		goto retrylookup;
1394 	}
1395 
1396 	return m;
1397 }
1398 
1399 /*
1400  * Mapping function for valid bits or for dirty bits in
1401  * a page.  May not block.
1402  *
1403  * Inputs are required to range within a page.
1404  */
1405 
1406 __inline int
1407 vm_page_bits(int base, int size)
1408 {
1409 	int first_bit;
1410 	int last_bit;
1411 
1412 	KASSERT(
1413 	    base + size <= PAGE_SIZE,
1414 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
1415 	);
1416 
1417 	if (size == 0)		/* handle degenerate case */
1418 		return(0);
1419 
1420 	first_bit = base >> DEV_BSHIFT;
1421 	last_bit = (base + size - 1) >> DEV_BSHIFT;
1422 
1423 	return ((2 << last_bit) - (1 << first_bit));
1424 }
1425 
1426 /*
1427  *	vm_page_set_validclean:
1428  *
1429  *	Sets portions of a page valid and clean.  The arguments are expected
1430  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
1431  *	of any partial chunks touched by the range.  The invalid portion of
1432  *	such chunks will be zero'd.
1433  *
1434  *	This routine may not block.
1435  *
1436  *	(base + size) must be less then or equal to PAGE_SIZE.
1437  */
1438 void
1439 vm_page_set_validclean(m, base, size)
1440 	vm_page_t m;
1441 	int base;
1442 	int size;
1443 {
1444 	int pagebits;
1445 	int frag;
1446 	int endoff;
1447 
1448 	if (size == 0)	/* handle degenerate case */
1449 		return;
1450 
1451 	/*
1452 	 * If the base is not DEV_BSIZE aligned and the valid
1453 	 * bit is clear, we have to zero out a portion of the
1454 	 * first block.
1455 	 */
1456 
1457 	if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
1458 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0
1459 	) {
1460 		pmap_zero_page_area(
1461 		    VM_PAGE_TO_PHYS(m),
1462 		    frag,
1463 		    base - frag
1464 		);
1465 	}
1466 
1467 	/*
1468 	 * If the ending offset is not DEV_BSIZE aligned and the
1469 	 * valid bit is clear, we have to zero out a portion of
1470 	 * the last block.
1471 	 */
1472 
1473 	endoff = base + size;
1474 
1475 	if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
1476 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0
1477 	) {
1478 		pmap_zero_page_area(
1479 		    VM_PAGE_TO_PHYS(m),
1480 		    endoff,
1481 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1))
1482 		);
1483 	}
1484 
1485 	/*
1486 	 * Set valid, clear dirty bits.  If validating the entire
1487 	 * page we can safely clear the pmap modify bit.
1488 	 */
1489 
1490 	pagebits = vm_page_bits(base, size);
1491 	m->valid |= pagebits;
1492 	m->dirty &= ~pagebits;
1493 
1494 	if (base == 0 && size == PAGE_SIZE)
1495 		pmap_clear_modify(VM_PAGE_TO_PHYS(m));
1496 }
1497 
1498 #if 0
1499 
1500 void
1501 vm_page_set_dirty(m, base, size)
1502 	vm_page_t m;
1503 	int base;
1504 	int size;
1505 {
1506 	m->dirty |= vm_page_bits(base, size);
1507 }
1508 
1509 #endif
1510 
1511 void
1512 vm_page_clear_dirty(m, base, size)
1513 	vm_page_t m;
1514 	int base;
1515 	int size;
1516 {
1517 	m->dirty &= ~vm_page_bits(base, size);
1518 }
1519 
1520 /*
1521  *	vm_page_set_invalid:
1522  *
1523  *	Invalidates DEV_BSIZE'd chunks within a page.  Both the
1524  *	valid and dirty bits for the effected areas are cleared.
1525  *
1526  *	May not block.
1527  */
1528 void
1529 vm_page_set_invalid(m, base, size)
1530 	vm_page_t m;
1531 	int base;
1532 	int size;
1533 {
1534 	int bits;
1535 
1536 	bits = vm_page_bits(base, size);
1537 	m->valid &= ~bits;
1538 	m->dirty &= ~bits;
1539 	m->object->generation++;
1540 }
1541 
1542 /*
1543  * vm_page_zero_invalid()
1544  *
1545  *	The kernel assumes that the invalid portions of a page contain
1546  *	garbage, but such pages can be mapped into memory by user code.
1547  *	When this occurs, we must zero out the non-valid portions of the
1548  *	page so user code sees what it expects.
1549  *
1550  *	Pages are most often semi-valid when the end of a file is mapped
1551  *	into memory and the file's size is not page aligned.
1552  */
1553 
1554 void
1555 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
1556 {
1557 	int b;
1558 	int i;
1559 
1560 	/*
1561 	 * Scan the valid bits looking for invalid sections that
1562 	 * must be zerod.  Invalid sub-DEV_BSIZE'd areas ( where the
1563 	 * valid bit may be set ) have already been zerod by
1564 	 * vm_page_set_validclean().
1565 	 */
1566 
1567 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
1568 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
1569 		    (m->valid & (1 << i))
1570 		) {
1571 			if (i > b) {
1572 				pmap_zero_page_area(
1573 				    VM_PAGE_TO_PHYS(m),
1574 				    b << DEV_BSHIFT,
1575 				    (i - b) << DEV_BSHIFT
1576 				);
1577 			}
1578 			b = i + 1;
1579 		}
1580 	}
1581 
1582 	/*
1583 	 * setvalid is TRUE when we can safely set the zero'd areas
1584 	 * as being valid.  We can do this if there are no cache consistancy
1585 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
1586 	 */
1587 
1588 	if (setvalid)
1589 		m->valid = VM_PAGE_BITS_ALL;
1590 }
1591 
1592 /*
1593  *	vm_page_is_valid:
1594  *
1595  *	Is (partial) page valid?  Note that the case where size == 0
1596  *	will return FALSE in the degenerate case where the page is
1597  *	entirely invalid, and TRUE otherwise.
1598  *
1599  *	May not block.
1600  */
1601 
1602 int
1603 vm_page_is_valid(m, base, size)
1604 	vm_page_t m;
1605 	int base;
1606 	int size;
1607 {
1608 	int bits = vm_page_bits(base, size);
1609 
1610 	if (m->valid && ((m->valid & bits) == bits))
1611 		return 1;
1612 	else
1613 		return 0;
1614 }
1615 
1616 /*
1617  * update dirty bits from pmap/mmu.  May not block.
1618  */
1619 
1620 void
1621 vm_page_test_dirty(m)
1622 	vm_page_t m;
1623 {
1624 	if ((m->dirty != VM_PAGE_BITS_ALL) &&
1625 	    pmap_is_modified(VM_PAGE_TO_PHYS(m))) {
1626 		vm_page_dirty(m);
1627 	}
1628 }
1629 
1630 /*
1631  * This interface is for merging with malloc() someday.
1632  * Even if we never implement compaction so that contiguous allocation
1633  * works after initialization time, malloc()'s data structures are good
1634  * for statistics and for allocations of less than a page.
1635  */
1636 void *
1637 contigmalloc1(size, type, flags, low, high, alignment, boundary, map)
1638 	unsigned long size;	/* should be size_t here and for malloc() */
1639 	struct malloc_type *type;
1640 	int flags;
1641 	unsigned long low;
1642 	unsigned long high;
1643 	unsigned long alignment;
1644 	unsigned long boundary;
1645 	vm_map_t map;
1646 {
1647 	int i, s, start;
1648 	vm_offset_t addr, phys, tmp_addr;
1649 	int pass;
1650 	vm_page_t pga = vm_page_array;
1651 
1652 	size = round_page(size);
1653 #if !defined(MAX_PERF)
1654 	if (size == 0)
1655 		panic("contigmalloc1: size must not be 0");
1656 	if ((alignment & (alignment - 1)) != 0)
1657 		panic("contigmalloc1: alignment must be a power of 2");
1658 	if ((boundary & (boundary - 1)) != 0)
1659 		panic("contigmalloc1: boundary must be a power of 2");
1660 #endif
1661 
1662 	start = 0;
1663 	for (pass = 0; pass <= 1; pass++) {
1664 		s = splvm();
1665 again:
1666 		/*
1667 		 * Find first page in array that is free, within range, aligned, and
1668 		 * such that the boundary won't be crossed.
1669 		 */
1670 		for (i = start; i < cnt.v_page_count; i++) {
1671 			int pqtype;
1672 			phys = VM_PAGE_TO_PHYS(&pga[i]);
1673 			pqtype = pga[i].queue - pga[i].pc;
1674 			if (((pqtype == PQ_FREE) || (pqtype == PQ_CACHE)) &&
1675 			    (phys >= low) && (phys < high) &&
1676 			    ((phys & (alignment - 1)) == 0) &&
1677 			    (((phys ^ (phys + size - 1)) & ~(boundary - 1)) == 0))
1678 				break;
1679 		}
1680 
1681 		/*
1682 		 * If the above failed or we will exceed the upper bound, fail.
1683 		 */
1684 		if ((i == cnt.v_page_count) ||
1685 			((VM_PAGE_TO_PHYS(&pga[i]) + size) > high)) {
1686 			vm_page_t m, next;
1687 
1688 again1:
1689 			for (m = TAILQ_FIRST(&vm_page_queue_inactive);
1690 				m != NULL;
1691 				m = next) {
1692 
1693 				if (m->queue != PQ_INACTIVE) {
1694 					break;
1695 				}
1696 
1697 				next = TAILQ_NEXT(m, pageq);
1698 				if (vm_page_sleep_busy(m, TRUE, "vpctw0"))
1699 					goto again1;
1700 				vm_page_test_dirty(m);
1701 				if (m->dirty) {
1702 					if (m->object->type == OBJT_VNODE) {
1703 						vn_lock(m->object->handle, LK_EXCLUSIVE | LK_RETRY, curproc);
1704 						vm_object_page_clean(m->object, 0, 0, OBJPC_SYNC);
1705 						VOP_UNLOCK(m->object->handle, 0, curproc);
1706 						goto again1;
1707 					} else if (m->object->type == OBJT_SWAP ||
1708 								m->object->type == OBJT_DEFAULT) {
1709 						vm_pageout_flush(&m, 1, 0);
1710 						goto again1;
1711 					}
1712 				}
1713 				if ((m->dirty == 0) && (m->busy == 0) && (m->hold_count == 0))
1714 					vm_page_cache(m);
1715 			}
1716 
1717 			for (m = TAILQ_FIRST(&vm_page_queue_active);
1718 				m != NULL;
1719 				m = next) {
1720 
1721 				if (m->queue != PQ_ACTIVE) {
1722 					break;
1723 				}
1724 
1725 				next = TAILQ_NEXT(m, pageq);
1726 				if (vm_page_sleep_busy(m, TRUE, "vpctw1"))
1727 					goto again1;
1728 				vm_page_test_dirty(m);
1729 				if (m->dirty) {
1730 					if (m->object->type == OBJT_VNODE) {
1731 						vn_lock(m->object->handle, LK_EXCLUSIVE | LK_RETRY, curproc);
1732 						vm_object_page_clean(m->object, 0, 0, OBJPC_SYNC);
1733 						VOP_UNLOCK(m->object->handle, 0, curproc);
1734 						goto again1;
1735 					} else if (m->object->type == OBJT_SWAP ||
1736 								m->object->type == OBJT_DEFAULT) {
1737 						vm_pageout_flush(&m, 1, 0);
1738 						goto again1;
1739 					}
1740 				}
1741 				if ((m->dirty == 0) && (m->busy == 0) && (m->hold_count == 0))
1742 					vm_page_cache(m);
1743 			}
1744 
1745 			splx(s);
1746 			continue;
1747 		}
1748 		start = i;
1749 
1750 		/*
1751 		 * Check successive pages for contiguous and free.
1752 		 */
1753 		for (i = start + 1; i < (start + size / PAGE_SIZE); i++) {
1754 			int pqtype;
1755 			pqtype = pga[i].queue - pga[i].pc;
1756 			if ((VM_PAGE_TO_PHYS(&pga[i]) !=
1757 			    (VM_PAGE_TO_PHYS(&pga[i - 1]) + PAGE_SIZE)) ||
1758 			    ((pqtype != PQ_FREE) && (pqtype != PQ_CACHE))) {
1759 				start++;
1760 				goto again;
1761 			}
1762 		}
1763 
1764 		for (i = start; i < (start + size / PAGE_SIZE); i++) {
1765 			int pqtype;
1766 			vm_page_t m = &pga[i];
1767 
1768 			pqtype = m->queue - m->pc;
1769 			if (pqtype == PQ_CACHE) {
1770 				vm_page_busy(m);
1771 				vm_page_free(m);
1772 			}
1773 
1774 			TAILQ_REMOVE(vm_page_queues[m->queue].pl, m, pageq);
1775 			(*vm_page_queues[m->queue].lcnt)--;
1776 			cnt.v_free_count--;
1777 			m->valid = VM_PAGE_BITS_ALL;
1778 			m->flags = 0;
1779 			m->dirty = 0;
1780 			m->wire_count = 0;
1781 			m->busy = 0;
1782 			m->queue = PQ_NONE;
1783 			m->object = NULL;
1784 			vm_page_wire(m);
1785 		}
1786 
1787 		/*
1788 		 * We've found a contiguous chunk that meets are requirements.
1789 		 * Allocate kernel VM, unfree and assign the physical pages to it and
1790 		 * return kernel VM pointer.
1791 		 */
1792 		tmp_addr = addr = kmem_alloc_pageable(map, size);
1793 		if (addr == 0) {
1794 			/*
1795 			 * XXX We almost never run out of kernel virtual
1796 			 * space, so we don't make the allocated memory
1797 			 * above available.
1798 			 */
1799 			splx(s);
1800 			return (NULL);
1801 		}
1802 
1803 		for (i = start; i < (start + size / PAGE_SIZE); i++) {
1804 			vm_page_t m = &pga[i];
1805 			vm_page_insert(m, kernel_object,
1806 				OFF_TO_IDX(tmp_addr - VM_MIN_KERNEL_ADDRESS));
1807 			pmap_kenter(tmp_addr, VM_PAGE_TO_PHYS(m));
1808 			tmp_addr += PAGE_SIZE;
1809 		}
1810 
1811 		splx(s);
1812 		return ((void *)addr);
1813 	}
1814 	return NULL;
1815 }
1816 
1817 void *
1818 contigmalloc(size, type, flags, low, high, alignment, boundary)
1819 	unsigned long size;	/* should be size_t here and for malloc() */
1820 	struct malloc_type *type;
1821 	int flags;
1822 	unsigned long low;
1823 	unsigned long high;
1824 	unsigned long alignment;
1825 	unsigned long boundary;
1826 {
1827 	return contigmalloc1(size, type, flags, low, high, alignment, boundary,
1828 			     kernel_map);
1829 }
1830 
1831 vm_offset_t
1832 vm_page_alloc_contig(size, low, high, alignment)
1833 	vm_offset_t size;
1834 	vm_offset_t low;
1835 	vm_offset_t high;
1836 	vm_offset_t alignment;
1837 {
1838 	return ((vm_offset_t)contigmalloc1(size, M_DEVBUF, M_NOWAIT, low, high,
1839 					  alignment, 0ul, kernel_map));
1840 }
1841 
1842 #include "opt_ddb.h"
1843 #ifdef DDB
1844 #include <sys/kernel.h>
1845 
1846 #include <ddb/ddb.h>
1847 
1848 DB_SHOW_COMMAND(page, vm_page_print_page_info)
1849 {
1850 	db_printf("cnt.v_free_count: %d\n", cnt.v_free_count);
1851 	db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count);
1852 	db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count);
1853 	db_printf("cnt.v_active_count: %d\n", cnt.v_active_count);
1854 	db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count);
1855 	db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved);
1856 	db_printf("cnt.v_free_min: %d\n", cnt.v_free_min);
1857 	db_printf("cnt.v_free_target: %d\n", cnt.v_free_target);
1858 	db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min);
1859 	db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target);
1860 }
1861 
1862 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
1863 {
1864 	int i;
1865 	db_printf("PQ_FREE:");
1866 	for(i=0;i<PQ_L2_SIZE;i++) {
1867 		db_printf(" %d", *vm_page_queues[PQ_FREE + i].lcnt);
1868 	}
1869 	db_printf("\n");
1870 
1871 	db_printf("PQ_CACHE:");
1872 	for(i=0;i<PQ_L2_SIZE;i++) {
1873 		db_printf(" %d", *vm_page_queues[PQ_CACHE + i].lcnt);
1874 	}
1875 	db_printf("\n");
1876 
1877 	db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n",
1878 		*vm_page_queues[PQ_ACTIVE].lcnt,
1879 		*vm_page_queues[PQ_INACTIVE].lcnt);
1880 }
1881 #endif /* DDB */
1882