xref: /dragonfly/sys/vm/vm_page.c (revision 984263bc)
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  * $FreeBSD: src/sys/vm/vm_page.c,v 1.147.2.18 2002/03/10 05:03:19 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 <sys/lock.h>
81 #include <vm/vm_kern.h>
82 #include <vm/pmap.h>
83 #include <vm/vm_map.h>
84 #include <vm/vm_object.h>
85 #include <vm/vm_page.h>
86 #include <vm/vm_pageout.h>
87 #include <vm/vm_pager.h>
88 #include <vm/vm_extern.h>
89 
90 static void	vm_page_queue_init (void);
91 static vm_page_t vm_page_select_cache (vm_object_t, vm_pindex_t);
92 
93 /*
94  *	Associated with page of user-allocatable memory is a
95  *	page structure.
96  */
97 
98 static struct vm_page **vm_page_buckets; /* Array of buckets */
99 static int vm_page_bucket_count;	/* How big is array? */
100 static int vm_page_hash_mask;		/* Mask for hash function */
101 static volatile int vm_page_bucket_generation;
102 
103 struct vpgqueues vm_page_queues[PQ_COUNT];
104 
105 static void
106 vm_page_queue_init(void) {
107 	int i;
108 
109 	for(i=0;i<PQ_L2_SIZE;i++) {
110 		vm_page_queues[PQ_FREE+i].cnt = &cnt.v_free_count;
111 	}
112 	vm_page_queues[PQ_INACTIVE].cnt = &cnt.v_inactive_count;
113 
114 	vm_page_queues[PQ_ACTIVE].cnt = &cnt.v_active_count;
115 	vm_page_queues[PQ_HOLD].cnt = &cnt.v_active_count;
116 	for(i=0;i<PQ_L2_SIZE;i++) {
117 		vm_page_queues[PQ_CACHE+i].cnt = &cnt.v_cache_count;
118 	}
119 	for(i=0;i<PQ_COUNT;i++) {
120 		TAILQ_INIT(&vm_page_queues[i].pl);
121 	}
122 }
123 
124 vm_page_t vm_page_array = 0;
125 int vm_page_array_size = 0;
126 long first_page = 0;
127 int vm_page_zero_count = 0;
128 
129 static __inline int vm_page_hash (vm_object_t object, vm_pindex_t pindex);
130 static void vm_page_free_wakeup (void);
131 
132 /*
133  *	vm_set_page_size:
134  *
135  *	Sets the page size, perhaps based upon the memory
136  *	size.  Must be called before any use of page-size
137  *	dependent functions.
138  */
139 void
140 vm_set_page_size(void)
141 {
142 	if (cnt.v_page_size == 0)
143 		cnt.v_page_size = PAGE_SIZE;
144 	if (((cnt.v_page_size - 1) & cnt.v_page_size) != 0)
145 		panic("vm_set_page_size: page size not a power of two");
146 }
147 
148 /*
149  *	vm_add_new_page:
150  *
151  *	Add a new page to the freelist for use by the system.
152  *	Must be called at splhigh().
153  */
154 vm_page_t
155 vm_add_new_page(vm_offset_t pa)
156 {
157 	vm_page_t m;
158 
159 	++cnt.v_page_count;
160 	++cnt.v_free_count;
161 	m = PHYS_TO_VM_PAGE(pa);
162 	m->phys_addr = pa;
163 	m->flags = 0;
164 	m->pc = (pa >> PAGE_SHIFT) & PQ_L2_MASK;
165 	m->queue = m->pc + PQ_FREE;
166 	TAILQ_INSERT_HEAD(&vm_page_queues[m->queue].pl, m, pageq);
167 	vm_page_queues[m->queue].lcnt++;
168 	return (m);
169 }
170 
171 /*
172  *	vm_page_startup:
173  *
174  *	Initializes the resident memory module.
175  *
176  *	Allocates memory for the page cells, and
177  *	for the object/offset-to-page hash table headers.
178  *	Each page cell is initialized and placed on the free list.
179  */
180 
181 vm_offset_t
182 vm_page_startup(vm_offset_t starta, vm_offset_t enda, vm_offset_t vaddr)
183 {
184 	vm_offset_t mapped;
185 	struct vm_page **bucket;
186 	vm_size_t npages, page_range;
187 	vm_offset_t new_end;
188 	int i;
189 	vm_offset_t pa;
190 	int nblocks;
191 	vm_offset_t last_pa;
192 
193 	/* the biggest memory array is the second group of pages */
194 	vm_offset_t end;
195 	vm_offset_t biggestone, biggestsize;
196 
197 	vm_offset_t total;
198 
199 	total = 0;
200 	biggestsize = 0;
201 	biggestone = 0;
202 	nblocks = 0;
203 	vaddr = round_page(vaddr);
204 
205 	for (i = 0; phys_avail[i + 1]; i += 2) {
206 		phys_avail[i] = round_page(phys_avail[i]);
207 		phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
208 	}
209 
210 	for (i = 0; phys_avail[i + 1]; i += 2) {
211 		int size = phys_avail[i + 1] - phys_avail[i];
212 
213 		if (size > biggestsize) {
214 			biggestone = i;
215 			biggestsize = size;
216 		}
217 		++nblocks;
218 		total += size;
219 	}
220 
221 	end = phys_avail[biggestone+1];
222 
223 	/*
224 	 * Initialize the queue headers for the free queue, the active queue
225 	 * and the inactive queue.
226 	 */
227 
228 	vm_page_queue_init();
229 
230 	/*
231 	 * Allocate (and initialize) the hash table buckets.
232 	 *
233 	 * The number of buckets MUST BE a power of 2, and the actual value is
234 	 * the next power of 2 greater than the number of physical pages in
235 	 * the system.
236 	 *
237 	 * We make the hash table approximately 2x the number of pages to
238 	 * reduce the chain length.  This is about the same size using the
239 	 * singly-linked list as the 1x hash table we were using before
240 	 * using TAILQ but the chain length will be smaller.
241 	 *
242 	 * Note: This computation can be tweaked if desired.
243 	 */
244 	vm_page_buckets = (struct vm_page **)vaddr;
245 	bucket = vm_page_buckets;
246 	if (vm_page_bucket_count == 0) {
247 		vm_page_bucket_count = 1;
248 		while (vm_page_bucket_count < atop(total))
249 			vm_page_bucket_count <<= 1;
250 	}
251 	vm_page_bucket_count <<= 1;
252 	vm_page_hash_mask = vm_page_bucket_count - 1;
253 
254 	/*
255 	 * Validate these addresses.
256 	 */
257 	new_end = end - vm_page_bucket_count * sizeof(struct vm_page *);
258 	new_end = trunc_page(new_end);
259 	mapped = round_page(vaddr);
260 	vaddr = pmap_map(mapped, new_end, end,
261 	    VM_PROT_READ | VM_PROT_WRITE);
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 	    (end - new_end)) / PAGE_SIZE;
281 
282 	end = new_end;
283 	/*
284 	 * Initialize the mem entry structures now, and put them in the free
285 	 * queue.
286 	 */
287 	vm_page_array = (vm_page_t) vaddr;
288 	mapped = vaddr;
289 
290 	/*
291 	 * Validate these addresses.
292 	 */
293 
294 	new_end = trunc_page(end - page_range * sizeof(struct vm_page));
295 	mapped = pmap_map(mapped, new_end, end,
296 	    VM_PROT_READ | VM_PROT_WRITE);
297 
298 	/*
299 	 * Clear all of the page structures
300 	 */
301 	bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
302 	vm_page_array_size = page_range;
303 
304 	/*
305 	 * Construct the free queue(s) in descending order (by physical
306 	 * address) so that the first 16MB of physical memory is allocated
307 	 * last rather than first.  On large-memory machines, this avoids
308 	 * the exhaustion of low physical memory before isa_dmainit has run.
309 	 */
310 	cnt.v_page_count = 0;
311 	cnt.v_free_count = 0;
312 	for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) {
313 		pa = phys_avail[i];
314 		if (i == biggestone)
315 			last_pa = new_end;
316 		else
317 			last_pa = phys_avail[i + 1];
318 		while (pa < last_pa && npages-- > 0) {
319 			vm_add_new_page(pa);
320 			pa += PAGE_SIZE;
321 		}
322 	}
323 	return (mapped);
324 }
325 
326 /*
327  *	vm_page_hash:
328  *
329  *	Distributes the object/offset key pair among hash buckets.
330  *
331  *	NOTE:  This macro depends on vm_page_bucket_count being a power of 2.
332  *	This routine may not block.
333  *
334  *	We try to randomize the hash based on the object to spread the pages
335  *	out in the hash table without it costing us too much.
336  */
337 static __inline int
338 vm_page_hash(vm_object_t object, vm_pindex_t pindex)
339 {
340 	int i = ((uintptr_t)object + pindex) ^ object->hash_rand;
341 
342 	return(i & vm_page_hash_mask);
343 }
344 
345 void
346 vm_page_unhold(vm_page_t mem)
347 {
348 	--mem->hold_count;
349 	KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!"));
350 	if (mem->hold_count == 0 && mem->queue == PQ_HOLD)
351 		vm_page_free_toq(mem);
352 }
353 
354 /*
355  *	vm_page_insert:		[ internal use only ]
356  *
357  *	Inserts the given mem entry into the object and object list.
358  *
359  *	The pagetables are not updated but will presumably fault the page
360  *	in if necessary, or if a kernel page the caller will at some point
361  *	enter the page into the kernel's pmap.  We are not allowed to block
362  *	here so we *can't* do this anyway.
363  *
364  *	The object and page must be locked, and must be splhigh.
365  *	This routine may not block.
366  */
367 
368 void
369 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
370 {
371 	struct vm_page **bucket;
372 
373 	if (m->object != NULL)
374 		panic("vm_page_insert: already inserted");
375 
376 	/*
377 	 * Record the object/offset pair in this page
378 	 */
379 
380 	m->object = object;
381 	m->pindex = pindex;
382 
383 	/*
384 	 * Insert it into the object_object/offset hash table
385 	 */
386 
387 	bucket = &vm_page_buckets[vm_page_hash(object, pindex)];
388 	m->hnext = *bucket;
389 	*bucket = m;
390 	vm_page_bucket_generation++;
391 
392 	/*
393 	 * Now link into the object's list of backed pages.
394 	 */
395 
396 	TAILQ_INSERT_TAIL(&object->memq, m, listq);
397 	object->generation++;
398 
399 	/*
400 	 * show that the object has one more resident page.
401 	 */
402 
403 	object->resident_page_count++;
404 
405 	/*
406 	 * Since we are inserting a new and possibly dirty page,
407 	 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags.
408 	 */
409 	if (m->flags & PG_WRITEABLE)
410 		vm_object_set_writeable_dirty(object);
411 }
412 
413 /*
414  *	vm_page_remove:
415  *				NOTE: used by device pager as well -wfj
416  *
417  *	Removes the given mem entry from the object/offset-page
418  *	table and the object page list, but do not invalidate/terminate
419  *	the backing store.
420  *
421  *	The object and page must be locked, and at splhigh.
422  *	The underlying pmap entry (if any) is NOT removed here.
423  *	This routine may not block.
424  */
425 
426 void
427 vm_page_remove(vm_page_t m)
428 {
429 	vm_object_t object;
430 
431 	if (m->object == NULL)
432 		return;
433 
434 	if ((m->flags & PG_BUSY) == 0) {
435 		panic("vm_page_remove: page not busy");
436 	}
437 
438 	/*
439 	 * Basically destroy the page.
440 	 */
441 
442 	vm_page_wakeup(m);
443 
444 	object = m->object;
445 
446 	/*
447 	 * Remove from the object_object/offset hash table.  The object
448 	 * must be on the hash queue, we will panic if it isn't
449 	 *
450 	 * Note: we must NULL-out m->hnext to prevent loops in detached
451 	 * buffers with vm_page_lookup().
452 	 */
453 
454 	{
455 		struct vm_page **bucket;
456 
457 		bucket = &vm_page_buckets[vm_page_hash(m->object, m->pindex)];
458 		while (*bucket != m) {
459 			if (*bucket == NULL)
460 				panic("vm_page_remove(): page not found in hash");
461 			bucket = &(*bucket)->hnext;
462 		}
463 		*bucket = m->hnext;
464 		m->hnext = NULL;
465 		vm_page_bucket_generation++;
466 	}
467 
468 	/*
469 	 * Now remove from the object's list of backed pages.
470 	 */
471 
472 	TAILQ_REMOVE(&object->memq, m, listq);
473 
474 	/*
475 	 * And show that the object has one fewer resident page.
476 	 */
477 
478 	object->resident_page_count--;
479 	object->generation++;
480 
481 	m->object = NULL;
482 }
483 
484 /*
485  *	vm_page_lookup:
486  *
487  *	Returns the page associated with the object/offset
488  *	pair specified; if none is found, NULL is returned.
489  *
490  *	NOTE: the code below does not lock.  It will operate properly if
491  *	an interrupt makes a change, but the generation algorithm will not
492  *	operate properly in an SMP environment where both cpu's are able to run
493  *	kernel code simultaneously.
494  *
495  *	The object must be locked.  No side effects.
496  *	This routine may not block.
497  *	This is a critical path routine
498  */
499 
500 vm_page_t
501 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
502 {
503 	vm_page_t m;
504 	struct vm_page **bucket;
505 	int generation;
506 
507 	/*
508 	 * Search the hash table for this object/offset pair
509 	 */
510 
511 retry:
512 	generation = vm_page_bucket_generation;
513 	bucket = &vm_page_buckets[vm_page_hash(object, pindex)];
514 	for (m = *bucket; m != NULL; m = m->hnext) {
515 		if ((m->object == object) && (m->pindex == pindex)) {
516 			if (vm_page_bucket_generation != generation)
517 				goto retry;
518 			return (m);
519 		}
520 	}
521 	if (vm_page_bucket_generation != generation)
522 		goto retry;
523 	return (NULL);
524 }
525 
526 /*
527  *	vm_page_rename:
528  *
529  *	Move the given memory entry from its
530  *	current object to the specified target object/offset.
531  *
532  *	The object must be locked.
533  *	This routine may not block.
534  *
535  *	Note: this routine will raise itself to splvm(), the caller need not.
536  *
537  *	Note: swap associated with the page must be invalidated by the move.  We
538  *	      have to do this for several reasons:  (1) we aren't freeing the
539  *	      page, (2) we are dirtying the page, (3) the VM system is probably
540  *	      moving the page from object A to B, and will then later move
541  *	      the backing store from A to B and we can't have a conflict.
542  *
543  *	Note: we *always* dirty the page.  It is necessary both for the
544  *	      fact that we moved it, and because we may be invalidating
545  *	      swap.  If the page is on the cache, we have to deactivate it
546  *	      or vm_page_dirty() will panic.  Dirty pages are not allowed
547  *	      on the cache.
548  */
549 
550 void
551 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
552 {
553 	int s;
554 
555 	s = splvm();
556 	vm_page_remove(m);
557 	vm_page_insert(m, new_object, new_pindex);
558 	if (m->queue - m->pc == PQ_CACHE)
559 		vm_page_deactivate(m);
560 	vm_page_dirty(m);
561 	splx(s);
562 }
563 
564 /*
565  * vm_page_unqueue_nowakeup:
566  *
567  * 	vm_page_unqueue() without any wakeup
568  *
569  *	This routine must be called at splhigh().
570  *	This routine may not block.
571  */
572 
573 void
574 vm_page_unqueue_nowakeup(vm_page_t m)
575 {
576 	int queue = m->queue;
577 	struct vpgqueues *pq;
578 	if (queue != PQ_NONE) {
579 		pq = &vm_page_queues[queue];
580 		m->queue = PQ_NONE;
581 		TAILQ_REMOVE(&pq->pl, m, pageq);
582 		(*pq->cnt)--;
583 		pq->lcnt--;
584 	}
585 }
586 
587 /*
588  * vm_page_unqueue:
589  *
590  *	Remove a page from its queue.
591  *
592  *	This routine must be called at splhigh().
593  *	This routine may not block.
594  */
595 
596 void
597 vm_page_unqueue(vm_page_t m)
598 {
599 	int queue = m->queue;
600 	struct vpgqueues *pq;
601 	if (queue != PQ_NONE) {
602 		m->queue = PQ_NONE;
603 		pq = &vm_page_queues[queue];
604 		TAILQ_REMOVE(&pq->pl, m, pageq);
605 		(*pq->cnt)--;
606 		pq->lcnt--;
607 		if ((queue - m->pc) == PQ_CACHE) {
608 			if (vm_paging_needed())
609 				pagedaemon_wakeup();
610 		}
611 	}
612 }
613 
614 #if PQ_L2_SIZE > 1
615 
616 /*
617  *	vm_page_list_find:
618  *
619  *	Find a page on the specified queue with color optimization.
620  *
621  *	The page coloring optimization attempts to locate a page
622  *	that does not overload other nearby pages in the object in
623  *	the cpu's L1 or L2 caches.  We need this optimization because
624  *	cpu caches tend to be physical caches, while object spaces tend
625  *	to be virtual.
626  *
627  *	This routine must be called at splvm().
628  *	This routine may not block.
629  *
630  *	This routine may only be called from the vm_page_list_find() macro
631  *	in vm_page.h
632  */
633 vm_page_t
634 _vm_page_list_find(int basequeue, int index)
635 {
636 	int i;
637 	vm_page_t m = NULL;
638 	struct vpgqueues *pq;
639 
640 	pq = &vm_page_queues[basequeue];
641 
642 	/*
643 	 * Note that for the first loop, index+i and index-i wind up at the
644 	 * same place.  Even though this is not totally optimal, we've already
645 	 * blown it by missing the cache case so we do not care.
646 	 */
647 
648 	for(i = PQ_L2_SIZE / 2; i > 0; --i) {
649 		if ((m = TAILQ_FIRST(&pq[(index + i) & PQ_L2_MASK].pl)) != NULL)
650 			break;
651 
652 		if ((m = TAILQ_FIRST(&pq[(index - i) & PQ_L2_MASK].pl)) != NULL)
653 			break;
654 	}
655 	return(m);
656 }
657 
658 #endif
659 
660 /*
661  *	vm_page_select_cache:
662  *
663  *	Find a page on the cache queue with color optimization.  As pages
664  *	might be found, but not applicable, they are deactivated.  This
665  *	keeps us from using potentially busy cached pages.
666  *
667  *	This routine must be called at splvm().
668  *	This routine may not block.
669  */
670 vm_page_t
671 vm_page_select_cache(vm_object_t object, vm_pindex_t pindex)
672 {
673 	vm_page_t m;
674 
675 	while (TRUE) {
676 		m = vm_page_list_find(
677 		    PQ_CACHE,
678 		    (pindex + object->pg_color) & PQ_L2_MASK,
679 		    FALSE
680 		);
681 		if (m && ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy ||
682 			       m->hold_count || m->wire_count)) {
683 			vm_page_deactivate(m);
684 			continue;
685 		}
686 		return m;
687 	}
688 }
689 
690 /*
691  *	vm_page_select_free:
692  *
693  *	Find a free or zero page, with specified preference.  We attempt to
694  *	inline the nominal case and fall back to _vm_page_select_free()
695  *	otherwise.
696  *
697  *	This routine must be called at splvm().
698  *	This routine may not block.
699  */
700 
701 static __inline vm_page_t
702 vm_page_select_free(vm_object_t object, vm_pindex_t pindex, boolean_t prefer_zero)
703 {
704 	vm_page_t m;
705 
706 	m = vm_page_list_find(
707 		PQ_FREE,
708 		(pindex + object->pg_color) & PQ_L2_MASK,
709 		prefer_zero
710 	);
711 	return(m);
712 }
713 
714 /*
715  *	vm_page_alloc:
716  *
717  *	Allocate and return a memory cell associated
718  *	with this VM object/offset pair.
719  *
720  *	page_req classes:
721  *	VM_ALLOC_NORMAL		normal process request
722  *	VM_ALLOC_SYSTEM		system *really* needs a page
723  *	VM_ALLOC_INTERRUPT	interrupt time request
724  *	VM_ALLOC_ZERO		zero page
725  *
726  *	Object must be locked.
727  *	This routine may not block.
728  *
729  *	Additional special handling is required when called from an
730  *	interrupt (VM_ALLOC_INTERRUPT).  We are not allowed to mess with
731  *	the page cache in this case.
732  */
733 
734 vm_page_t
735 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req)
736 {
737 	vm_page_t m = NULL;
738 	int s;
739 
740 	KASSERT(!vm_page_lookup(object, pindex),
741 		("vm_page_alloc: page already allocated"));
742 
743 	/*
744 	 * The pager is allowed to eat deeper into the free page list.
745 	 */
746 
747 	if ((curproc == pageproc) && (page_req != VM_ALLOC_INTERRUPT)) {
748 		page_req = VM_ALLOC_SYSTEM;
749 	};
750 
751 	s = splvm();
752 
753 loop:
754 	if (cnt.v_free_count > cnt.v_free_reserved) {
755 		/*
756 		 * Allocate from the free queue if there are plenty of pages
757 		 * in it.
758 		 */
759 		if (page_req == VM_ALLOC_ZERO)
760 			m = vm_page_select_free(object, pindex, TRUE);
761 		else
762 			m = vm_page_select_free(object, pindex, FALSE);
763 	} else if (
764 	    (page_req == VM_ALLOC_SYSTEM &&
765 	     cnt.v_cache_count == 0 &&
766 	     cnt.v_free_count > cnt.v_interrupt_free_min) ||
767 	    (page_req == VM_ALLOC_INTERRUPT && cnt.v_free_count > 0)
768 	) {
769 		/*
770 		 * Interrupt or system, dig deeper into the free list.
771 		 */
772 		m = vm_page_select_free(object, pindex, FALSE);
773 	} else if (page_req != VM_ALLOC_INTERRUPT) {
774 		/*
775 		 * Allocatable from cache (non-interrupt only).  On success,
776 		 * we must free the page and try again, thus ensuring that
777 		 * cnt.v_*_free_min counters are replenished.
778 		 */
779 		m = vm_page_select_cache(object, pindex);
780 		if (m == NULL) {
781 			splx(s);
782 #if defined(DIAGNOSTIC)
783 			if (cnt.v_cache_count > 0)
784 				printf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", cnt.v_cache_count);
785 #endif
786 			vm_pageout_deficit++;
787 			pagedaemon_wakeup();
788 			return (NULL);
789 		}
790 		KASSERT(m->dirty == 0, ("Found dirty cache page %p", m));
791 		vm_page_busy(m);
792 		vm_page_protect(m, VM_PROT_NONE);
793 		vm_page_free(m);
794 		goto loop;
795 	} else {
796 		/*
797 		 * Not allocatable from cache from interrupt, give up.
798 		 */
799 		splx(s);
800 		vm_pageout_deficit++;
801 		pagedaemon_wakeup();
802 		return (NULL);
803 	}
804 
805 	/*
806 	 *  At this point we had better have found a good page.
807 	 */
808 
809 	KASSERT(
810 	    m != NULL,
811 	    ("vm_page_alloc(): missing page on free queue\n")
812 	);
813 
814 	/*
815 	 * Remove from free queue
816 	 */
817 
818 	vm_page_unqueue_nowakeup(m);
819 
820 	/*
821 	 * Initialize structure.  Only the PG_ZERO flag is inherited.
822 	 */
823 
824 	if (m->flags & PG_ZERO) {
825 		vm_page_zero_count--;
826 		m->flags = PG_ZERO | PG_BUSY;
827 	} else {
828 		m->flags = PG_BUSY;
829 	}
830 	m->wire_count = 0;
831 	m->hold_count = 0;
832 	m->act_count = 0;
833 	m->busy = 0;
834 	m->valid = 0;
835 	KASSERT(m->dirty == 0, ("vm_page_alloc: free/cache page %p was dirty", m));
836 
837 	/*
838 	 * vm_page_insert() is safe prior to the splx().  Note also that
839 	 * inserting a page here does not insert it into the pmap (which
840 	 * could cause us to block allocating memory).  We cannot block
841 	 * anywhere.
842 	 */
843 
844 	vm_page_insert(m, object, pindex);
845 
846 	/*
847 	 * Don't wakeup too often - wakeup the pageout daemon when
848 	 * we would be nearly out of memory.
849 	 */
850 	if (vm_paging_needed())
851 		pagedaemon_wakeup();
852 
853 	splx(s);
854 
855 	return (m);
856 }
857 
858 /*
859  *	vm_wait:	(also see VM_WAIT macro)
860  *
861  *	Block until free pages are available for allocation
862  *	- Called in various places before memory allocations.
863  */
864 
865 void
866 vm_wait(void)
867 {
868 	int s;
869 
870 	s = splvm();
871 	if (curproc == pageproc) {
872 		vm_pageout_pages_needed = 1;
873 		tsleep(&vm_pageout_pages_needed, PSWP, "VMWait", 0);
874 	} else {
875 		if (!vm_pages_needed) {
876 			vm_pages_needed = 1;
877 			wakeup(&vm_pages_needed);
878 		}
879 		tsleep(&cnt.v_free_count, PVM, "vmwait", 0);
880 	}
881 	splx(s);
882 }
883 
884 /*
885  *	vm_waitpfault:	(also see VM_WAITPFAULT macro)
886  *
887  *	Block until free pages are available for allocation
888  *	- Called only in vm_fault so that processes page faulting
889  *	  can be easily tracked.
890  *	- Sleeps at a lower priority than vm_wait() so that vm_wait()ing
891  *	  processes will be able to grab memory first.  Do not change
892  *	  this balance without careful testing first.
893  */
894 
895 void
896 vm_waitpfault(void)
897 {
898 	int s;
899 
900 	s = splvm();
901 	if (!vm_pages_needed) {
902 		vm_pages_needed = 1;
903 		wakeup(&vm_pages_needed);
904 	}
905 	tsleep(&cnt.v_free_count, PUSER, "pfault", 0);
906 	splx(s);
907 }
908 
909 /*
910  *	vm_await:	(also see VM_AWAIT macro)
911  *
912  *	asleep on an event that will signal when free pages are available
913  *	for allocation.
914  */
915 
916 void
917 vm_await(void)
918 {
919 	int s;
920 
921 	s = splvm();
922 	if (curproc == pageproc) {
923 		vm_pageout_pages_needed = 1;
924 		asleep(&vm_pageout_pages_needed, PSWP, "vmwait", 0);
925 	} else {
926 		if (!vm_pages_needed) {
927 			vm_pages_needed++;
928 			wakeup(&vm_pages_needed);
929 		}
930 		asleep(&cnt.v_free_count, PVM, "vmwait", 0);
931 	}
932 	splx(s);
933 }
934 
935 #if 0
936 /*
937  *	vm_page_sleep:
938  *
939  *	Block until page is no longer busy.
940  */
941 
942 int
943 vm_page_sleep(vm_page_t m, char *msg, char *busy)
944 {
945 	int slept = 0;
946 	if ((busy && *busy) || (m->flags & PG_BUSY)) {
947 		int s;
948 		s = splvm();
949 		if ((busy && *busy) || (m->flags & PG_BUSY)) {
950 			vm_page_flag_set(m, PG_WANTED);
951 			tsleep(m, PVM, msg, 0);
952 			slept = 1;
953 		}
954 		splx(s);
955 	}
956 	return slept;
957 }
958 
959 #endif
960 
961 #if 0
962 
963 /*
964  *	vm_page_asleep:
965  *
966  *	Similar to vm_page_sleep(), but does not block.  Returns 0 if
967  *	the page is not busy, or 1 if the page is busy.
968  *
969  *	This routine has the side effect of calling asleep() if the page
970  *	was busy (1 returned).
971  */
972 
973 int
974 vm_page_asleep(vm_page_t m, char *msg, char *busy)
975 {
976 	int slept = 0;
977 	if ((busy && *busy) || (m->flags & PG_BUSY)) {
978 		int s;
979 		s = splvm();
980 		if ((busy && *busy) || (m->flags & PG_BUSY)) {
981 			vm_page_flag_set(m, PG_WANTED);
982 			asleep(m, PVM, msg, 0);
983 			slept = 1;
984 		}
985 		splx(s);
986 	}
987 	return slept;
988 }
989 
990 #endif
991 
992 /*
993  *	vm_page_activate:
994  *
995  *	Put the specified page on the active list (if appropriate).
996  *	Ensure that act_count is at least ACT_INIT but do not otherwise
997  *	mess with it.
998  *
999  *	The page queues must be locked.
1000  *	This routine may not block.
1001  */
1002 void
1003 vm_page_activate(vm_page_t m)
1004 {
1005 	int s;
1006 
1007 	s = splvm();
1008 	if (m->queue != PQ_ACTIVE) {
1009 		if ((m->queue - m->pc) == PQ_CACHE)
1010 			cnt.v_reactivated++;
1011 
1012 		vm_page_unqueue(m);
1013 
1014 		if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1015 			m->queue = PQ_ACTIVE;
1016 			vm_page_queues[PQ_ACTIVE].lcnt++;
1017 			TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1018 			if (m->act_count < ACT_INIT)
1019 				m->act_count = ACT_INIT;
1020 			cnt.v_active_count++;
1021 		}
1022 	} else {
1023 		if (m->act_count < ACT_INIT)
1024 			m->act_count = ACT_INIT;
1025 	}
1026 
1027 	splx(s);
1028 }
1029 
1030 /*
1031  *	vm_page_free_wakeup:
1032  *
1033  *	Helper routine for vm_page_free_toq() and vm_page_cache().  This
1034  *	routine is called when a page has been added to the cache or free
1035  *	queues.
1036  *
1037  *	This routine may not block.
1038  *	This routine must be called at splvm()
1039  */
1040 static __inline void
1041 vm_page_free_wakeup(void)
1042 {
1043 	/*
1044 	 * if pageout daemon needs pages, then tell it that there are
1045 	 * some free.
1046 	 */
1047 	if (vm_pageout_pages_needed &&
1048 	    cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) {
1049 		wakeup(&vm_pageout_pages_needed);
1050 		vm_pageout_pages_needed = 0;
1051 	}
1052 	/*
1053 	 * wakeup processes that are waiting on memory if we hit a
1054 	 * high water mark. And wakeup scheduler process if we have
1055 	 * lots of memory. this process will swapin processes.
1056 	 */
1057 	if (vm_pages_needed && !vm_page_count_min()) {
1058 		vm_pages_needed = 0;
1059 		wakeup(&cnt.v_free_count);
1060 	}
1061 }
1062 
1063 /*
1064  *	vm_page_free_toq:
1065  *
1066  *	Returns the given page to the PQ_FREE list,
1067  *	disassociating it with any VM object.
1068  *
1069  *	Object and page must be locked prior to entry.
1070  *	This routine may not block.
1071  */
1072 
1073 void
1074 vm_page_free_toq(vm_page_t m)
1075 {
1076 	int s;
1077 	struct vpgqueues *pq;
1078 	vm_object_t object = m->object;
1079 
1080 	s = splvm();
1081 
1082 	cnt.v_tfree++;
1083 
1084 	if (m->busy || ((m->queue - m->pc) == PQ_FREE)) {
1085 		printf(
1086 		"vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n",
1087 		    (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0,
1088 		    m->hold_count);
1089 		if ((m->queue - m->pc) == PQ_FREE)
1090 			panic("vm_page_free: freeing free page");
1091 		else
1092 			panic("vm_page_free: freeing busy page");
1093 	}
1094 
1095 	/*
1096 	 * unqueue, then remove page.  Note that we cannot destroy
1097 	 * the page here because we do not want to call the pager's
1098 	 * callback routine until after we've put the page on the
1099 	 * appropriate free queue.
1100 	 */
1101 
1102 	vm_page_unqueue_nowakeup(m);
1103 	vm_page_remove(m);
1104 
1105 	/*
1106 	 * If fictitious remove object association and
1107 	 * return, otherwise delay object association removal.
1108 	 */
1109 
1110 	if ((m->flags & PG_FICTITIOUS) != 0) {
1111 		splx(s);
1112 		return;
1113 	}
1114 
1115 	m->valid = 0;
1116 	vm_page_undirty(m);
1117 
1118 	if (m->wire_count != 0) {
1119 		if (m->wire_count > 1) {
1120 			panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx",
1121 				m->wire_count, (long)m->pindex);
1122 		}
1123 		panic("vm_page_free: freeing wired page\n");
1124 	}
1125 
1126 	/*
1127 	 * If we've exhausted the object's resident pages we want to free
1128 	 * it up.
1129 	 */
1130 
1131 	if (object &&
1132 	    (object->type == OBJT_VNODE) &&
1133 	    ((object->flags & OBJ_DEAD) == 0)
1134 	) {
1135 		struct vnode *vp = (struct vnode *)object->handle;
1136 
1137 		if (vp && VSHOULDFREE(vp))
1138 			vfree(vp);
1139 	}
1140 
1141 	/*
1142 	 * Clear the UNMANAGED flag when freeing an unmanaged page.
1143 	 */
1144 
1145 	if (m->flags & PG_UNMANAGED) {
1146 	    m->flags &= ~PG_UNMANAGED;
1147 	} else {
1148 #ifdef __alpha__
1149 	    pmap_page_is_free(m);
1150 #endif
1151 	}
1152 
1153 	if (m->hold_count != 0) {
1154 		m->flags &= ~PG_ZERO;
1155 		m->queue = PQ_HOLD;
1156 	} else
1157 		m->queue = PQ_FREE + m->pc;
1158 	pq = &vm_page_queues[m->queue];
1159 	pq->lcnt++;
1160 	++(*pq->cnt);
1161 
1162 	/*
1163 	 * Put zero'd pages on the end ( where we look for zero'd pages
1164 	 * first ) and non-zerod pages at the head.
1165 	 */
1166 
1167 	if (m->flags & PG_ZERO) {
1168 		TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
1169 		++vm_page_zero_count;
1170 	} else {
1171 		TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
1172 	}
1173 
1174 	vm_page_free_wakeup();
1175 
1176 	splx(s);
1177 }
1178 
1179 /*
1180  *	vm_page_unmanage:
1181  *
1182  * 	Prevent PV management from being done on the page.  The page is
1183  *	removed from the paging queues as if it were wired, and as a
1184  *	consequence of no longer being managed the pageout daemon will not
1185  *	touch it (since there is no way to locate the pte mappings for the
1186  *	page).  madvise() calls that mess with the pmap will also no longer
1187  *	operate on the page.
1188  *
1189  *	Beyond that the page is still reasonably 'normal'.  Freeing the page
1190  *	will clear the flag.
1191  *
1192  *	This routine is used by OBJT_PHYS objects - objects using unswappable
1193  *	physical memory as backing store rather then swap-backed memory and
1194  *	will eventually be extended to support 4MB unmanaged physical
1195  *	mappings.
1196  */
1197 
1198 void
1199 vm_page_unmanage(vm_page_t m)
1200 {
1201 	int s;
1202 
1203 	s = splvm();
1204 	if ((m->flags & PG_UNMANAGED) == 0) {
1205 		if (m->wire_count == 0)
1206 			vm_page_unqueue(m);
1207 	}
1208 	vm_page_flag_set(m, PG_UNMANAGED);
1209 	splx(s);
1210 }
1211 
1212 /*
1213  *	vm_page_wire:
1214  *
1215  *	Mark this page as wired down by yet
1216  *	another map, removing it from paging queues
1217  *	as necessary.
1218  *
1219  *	The page queues must be locked.
1220  *	This routine may not block.
1221  */
1222 void
1223 vm_page_wire(vm_page_t m)
1224 {
1225 	int s;
1226 
1227 	/*
1228 	 * Only bump the wire statistics if the page is not already wired,
1229 	 * and only unqueue the page if it is on some queue (if it is unmanaged
1230 	 * it is already off the queues).
1231 	 */
1232 	s = splvm();
1233 	if (m->wire_count == 0) {
1234 		if ((m->flags & PG_UNMANAGED) == 0)
1235 			vm_page_unqueue(m);
1236 		cnt.v_wire_count++;
1237 	}
1238 	m->wire_count++;
1239 	KASSERT(m->wire_count != 0,
1240 	    ("vm_page_wire: wire_count overflow m=%p", m));
1241 
1242 	splx(s);
1243 	vm_page_flag_set(m, PG_MAPPED);
1244 }
1245 
1246 /*
1247  *	vm_page_unwire:
1248  *
1249  *	Release one wiring of this page, potentially
1250  *	enabling it to be paged again.
1251  *
1252  *	Many pages placed on the inactive queue should actually go
1253  *	into the cache, but it is difficult to figure out which.  What
1254  *	we do instead, if the inactive target is well met, is to put
1255  *	clean pages at the head of the inactive queue instead of the tail.
1256  *	This will cause them to be moved to the cache more quickly and
1257  *	if not actively re-referenced, freed more quickly.  If we just
1258  *	stick these pages at the end of the inactive queue, heavy filesystem
1259  *	meta-data accesses can cause an unnecessary paging load on memory bound
1260  *	processes.  This optimization causes one-time-use metadata to be
1261  *	reused more quickly.
1262  *
1263  *	BUT, if we are in a low-memory situation we have no choice but to
1264  *	put clean pages on the cache queue.
1265  *
1266  *	A number of routines use vm_page_unwire() to guarantee that the page
1267  *	will go into either the inactive or active queues, and will NEVER
1268  *	be placed in the cache - for example, just after dirtying a page.
1269  *	dirty pages in the cache are not allowed.
1270  *
1271  *	The page queues must be locked.
1272  *	This routine may not block.
1273  */
1274 void
1275 vm_page_unwire(vm_page_t m, int activate)
1276 {
1277 	int s;
1278 
1279 	s = splvm();
1280 
1281 	if (m->wire_count > 0) {
1282 		m->wire_count--;
1283 		if (m->wire_count == 0) {
1284 			cnt.v_wire_count--;
1285 			if (m->flags & PG_UNMANAGED) {
1286 				;
1287 			} else if (activate) {
1288 				TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1289 				m->queue = PQ_ACTIVE;
1290 				vm_page_queues[PQ_ACTIVE].lcnt++;
1291 				cnt.v_active_count++;
1292 			} else {
1293 				vm_page_flag_clear(m, PG_WINATCFLS);
1294 				TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1295 				m->queue = PQ_INACTIVE;
1296 				vm_page_queues[PQ_INACTIVE].lcnt++;
1297 				cnt.v_inactive_count++;
1298 			}
1299 		}
1300 	} else {
1301 		panic("vm_page_unwire: invalid wire count: %d\n", m->wire_count);
1302 	}
1303 	splx(s);
1304 }
1305 
1306 
1307 /*
1308  * Move the specified page to the inactive queue.  If the page has
1309  * any associated swap, the swap is deallocated.
1310  *
1311  * Normally athead is 0 resulting in LRU operation.  athead is set
1312  * to 1 if we want this page to be 'as if it were placed in the cache',
1313  * except without unmapping it from the process address space.
1314  *
1315  * This routine may not block.
1316  */
1317 static __inline void
1318 _vm_page_deactivate(vm_page_t m, int athead)
1319 {
1320 	int s;
1321 
1322 	/*
1323 	 * Ignore if already inactive.
1324 	 */
1325 	if (m->queue == PQ_INACTIVE)
1326 		return;
1327 
1328 	s = splvm();
1329 	if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1330 		if ((m->queue - m->pc) == PQ_CACHE)
1331 			cnt.v_reactivated++;
1332 		vm_page_flag_clear(m, PG_WINATCFLS);
1333 		vm_page_unqueue(m);
1334 		if (athead)
1335 			TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1336 		else
1337 			TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1338 		m->queue = PQ_INACTIVE;
1339 		vm_page_queues[PQ_INACTIVE].lcnt++;
1340 		cnt.v_inactive_count++;
1341 	}
1342 	splx(s);
1343 }
1344 
1345 void
1346 vm_page_deactivate(vm_page_t m)
1347 {
1348     _vm_page_deactivate(m, 0);
1349 }
1350 
1351 /*
1352  * vm_page_try_to_cache:
1353  *
1354  * Returns 0 on failure, 1 on success
1355  */
1356 int
1357 vm_page_try_to_cache(vm_page_t m)
1358 {
1359 	if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1360 	    (m->flags & (PG_BUSY|PG_UNMANAGED))) {
1361 		return(0);
1362 	}
1363 	vm_page_test_dirty(m);
1364 	if (m->dirty)
1365 		return(0);
1366 	vm_page_cache(m);
1367 	return(1);
1368 }
1369 
1370 /*
1371  * vm_page_try_to_free()
1372  *
1373  *	Attempt to free the page.  If we cannot free it, we do nothing.
1374  *	1 is returned on success, 0 on failure.
1375  */
1376 
1377 int
1378 vm_page_try_to_free(vm_page_t m)
1379 {
1380 	if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1381 	    (m->flags & (PG_BUSY|PG_UNMANAGED))) {
1382 		return(0);
1383 	}
1384 	vm_page_test_dirty(m);
1385 	if (m->dirty)
1386 		return(0);
1387 	vm_page_busy(m);
1388 	vm_page_protect(m, VM_PROT_NONE);
1389 	vm_page_free(m);
1390 	return(1);
1391 }
1392 
1393 
1394 /*
1395  * vm_page_cache
1396  *
1397  * Put the specified page onto the page cache queue (if appropriate).
1398  *
1399  * This routine may not block.
1400  */
1401 void
1402 vm_page_cache(vm_page_t m)
1403 {
1404 	int s;
1405 
1406 	if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || m->wire_count) {
1407 		printf("vm_page_cache: attempting to cache busy page\n");
1408 		return;
1409 	}
1410 	if ((m->queue - m->pc) == PQ_CACHE)
1411 		return;
1412 
1413 	/*
1414 	 * Remove all pmaps and indicate that the page is not
1415 	 * writeable or mapped.
1416 	 */
1417 
1418 	vm_page_protect(m, VM_PROT_NONE);
1419 	if (m->dirty != 0) {
1420 		panic("vm_page_cache: caching a dirty page, pindex: %ld",
1421 			(long)m->pindex);
1422 	}
1423 	s = splvm();
1424 	vm_page_unqueue_nowakeup(m);
1425 	m->queue = PQ_CACHE + m->pc;
1426 	vm_page_queues[m->queue].lcnt++;
1427 	TAILQ_INSERT_TAIL(&vm_page_queues[m->queue].pl, m, pageq);
1428 	cnt.v_cache_count++;
1429 	vm_page_free_wakeup();
1430 	splx(s);
1431 }
1432 
1433 /*
1434  * vm_page_dontneed
1435  *
1436  *	Cache, deactivate, or do nothing as appropriate.  This routine
1437  *	is typically used by madvise() MADV_DONTNEED.
1438  *
1439  *	Generally speaking we want to move the page into the cache so
1440  *	it gets reused quickly.  However, this can result in a silly syndrome
1441  *	due to the page recycling too quickly.  Small objects will not be
1442  *	fully cached.  On the otherhand, if we move the page to the inactive
1443  *	queue we wind up with a problem whereby very large objects
1444  *	unnecessarily blow away our inactive and cache queues.
1445  *
1446  *	The solution is to move the pages based on a fixed weighting.  We
1447  *	either leave them alone, deactivate them, or move them to the cache,
1448  *	where moving them to the cache has the highest weighting.
1449  *	By forcing some pages into other queues we eventually force the
1450  *	system to balance the queues, potentially recovering other unrelated
1451  *	space from active.  The idea is to not force this to happen too
1452  *	often.
1453  */
1454 
1455 void
1456 vm_page_dontneed(vm_page_t m)
1457 {
1458 	static int dnweight;
1459 	int dnw;
1460 	int head;
1461 
1462 	dnw = ++dnweight;
1463 
1464 	/*
1465 	 * occassionally leave the page alone
1466 	 */
1467 
1468 	if ((dnw & 0x01F0) == 0 ||
1469 	    m->queue == PQ_INACTIVE ||
1470 	    m->queue - m->pc == PQ_CACHE
1471 	) {
1472 		if (m->act_count >= ACT_INIT)
1473 			--m->act_count;
1474 		return;
1475 	}
1476 
1477 	if (m->dirty == 0)
1478 		vm_page_test_dirty(m);
1479 
1480 	if (m->dirty || (dnw & 0x0070) == 0) {
1481 		/*
1482 		 * Deactivate the page 3 times out of 32.
1483 		 */
1484 		head = 0;
1485 	} else {
1486 		/*
1487 		 * Cache the page 28 times out of every 32.  Note that
1488 		 * the page is deactivated instead of cached, but placed
1489 		 * at the head of the queue instead of the tail.
1490 		 */
1491 		head = 1;
1492 	}
1493 	_vm_page_deactivate(m, head);
1494 }
1495 
1496 /*
1497  * Grab a page, waiting until we are waken up due to the page
1498  * changing state.  We keep on waiting, if the page continues
1499  * to be in the object.  If the page doesn't exist, allocate it.
1500  *
1501  * This routine may block.
1502  */
1503 vm_page_t
1504 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
1505 {
1506 
1507 	vm_page_t m;
1508 	int s, generation;
1509 
1510 retrylookup:
1511 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
1512 		if (m->busy || (m->flags & PG_BUSY)) {
1513 			generation = object->generation;
1514 
1515 			s = splvm();
1516 			while ((object->generation == generation) &&
1517 					(m->busy || (m->flags & PG_BUSY))) {
1518 				vm_page_flag_set(m, PG_WANTED | PG_REFERENCED);
1519 				tsleep(m, PVM, "pgrbwt", 0);
1520 				if ((allocflags & VM_ALLOC_RETRY) == 0) {
1521 					splx(s);
1522 					return NULL;
1523 				}
1524 			}
1525 			splx(s);
1526 			goto retrylookup;
1527 		} else {
1528 			vm_page_busy(m);
1529 			return m;
1530 		}
1531 	}
1532 
1533 	m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY);
1534 	if (m == NULL) {
1535 		VM_WAIT;
1536 		if ((allocflags & VM_ALLOC_RETRY) == 0)
1537 			return NULL;
1538 		goto retrylookup;
1539 	}
1540 
1541 	return m;
1542 }
1543 
1544 /*
1545  * Mapping function for valid bits or for dirty bits in
1546  * a page.  May not block.
1547  *
1548  * Inputs are required to range within a page.
1549  */
1550 
1551 __inline int
1552 vm_page_bits(int base, int size)
1553 {
1554 	int first_bit;
1555 	int last_bit;
1556 
1557 	KASSERT(
1558 	    base + size <= PAGE_SIZE,
1559 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
1560 	);
1561 
1562 	if (size == 0)		/* handle degenerate case */
1563 		return(0);
1564 
1565 	first_bit = base >> DEV_BSHIFT;
1566 	last_bit = (base + size - 1) >> DEV_BSHIFT;
1567 
1568 	return ((2 << last_bit) - (1 << first_bit));
1569 }
1570 
1571 /*
1572  *	vm_page_set_validclean:
1573  *
1574  *	Sets portions of a page valid and clean.  The arguments are expected
1575  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
1576  *	of any partial chunks touched by the range.  The invalid portion of
1577  *	such chunks will be zero'd.
1578  *
1579  *	This routine may not block.
1580  *
1581  *	(base + size) must be less then or equal to PAGE_SIZE.
1582  */
1583 void
1584 vm_page_set_validclean(vm_page_t m, int base, int size)
1585 {
1586 	int pagebits;
1587 	int frag;
1588 	int endoff;
1589 
1590 	if (size == 0)	/* handle degenerate case */
1591 		return;
1592 
1593 	/*
1594 	 * If the base is not DEV_BSIZE aligned and the valid
1595 	 * bit is clear, we have to zero out a portion of the
1596 	 * first block.
1597 	 */
1598 
1599 	if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
1600 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0
1601 	) {
1602 		pmap_zero_page_area(
1603 		    VM_PAGE_TO_PHYS(m),
1604 		    frag,
1605 		    base - frag
1606 		);
1607 	}
1608 
1609 	/*
1610 	 * If the ending offset is not DEV_BSIZE aligned and the
1611 	 * valid bit is clear, we have to zero out a portion of
1612 	 * the last block.
1613 	 */
1614 
1615 	endoff = base + size;
1616 
1617 	if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
1618 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0
1619 	) {
1620 		pmap_zero_page_area(
1621 		    VM_PAGE_TO_PHYS(m),
1622 		    endoff,
1623 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1))
1624 		);
1625 	}
1626 
1627 	/*
1628 	 * Set valid, clear dirty bits.  If validating the entire
1629 	 * page we can safely clear the pmap modify bit.  We also
1630 	 * use this opportunity to clear the PG_NOSYNC flag.  If a process
1631 	 * takes a write fault on a MAP_NOSYNC memory area the flag will
1632 	 * be set again.
1633 	 *
1634 	 * We set valid bits inclusive of any overlap, but we can only
1635 	 * clear dirty bits for DEV_BSIZE chunks that are fully within
1636 	 * the range.
1637 	 */
1638 
1639 	pagebits = vm_page_bits(base, size);
1640 	m->valid |= pagebits;
1641 #if 0	/* NOT YET */
1642 	if ((frag = base & (DEV_BSIZE - 1)) != 0) {
1643 		frag = DEV_BSIZE - frag;
1644 		base += frag;
1645 		size -= frag;
1646 		if (size < 0)
1647 		    size = 0;
1648 	}
1649 	pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
1650 #endif
1651 	m->dirty &= ~pagebits;
1652 	if (base == 0 && size == PAGE_SIZE) {
1653 		pmap_clear_modify(m);
1654 		vm_page_flag_clear(m, PG_NOSYNC);
1655 	}
1656 }
1657 
1658 #if 0
1659 
1660 void
1661 vm_page_set_dirty(vm_page_t m, int base, int size)
1662 {
1663 	m->dirty |= vm_page_bits(base, size);
1664 }
1665 
1666 #endif
1667 
1668 void
1669 vm_page_clear_dirty(vm_page_t m, int base, int size)
1670 {
1671 	m->dirty &= ~vm_page_bits(base, size);
1672 }
1673 
1674 /*
1675  *	vm_page_set_invalid:
1676  *
1677  *	Invalidates DEV_BSIZE'd chunks within a page.  Both the
1678  *	valid and dirty bits for the effected areas are cleared.
1679  *
1680  *	May not block.
1681  */
1682 void
1683 vm_page_set_invalid(vm_page_t m, int base, int size)
1684 {
1685 	int bits;
1686 
1687 	bits = vm_page_bits(base, size);
1688 	m->valid &= ~bits;
1689 	m->dirty &= ~bits;
1690 	m->object->generation++;
1691 }
1692 
1693 /*
1694  * vm_page_zero_invalid()
1695  *
1696  *	The kernel assumes that the invalid portions of a page contain
1697  *	garbage, but such pages can be mapped into memory by user code.
1698  *	When this occurs, we must zero out the non-valid portions of the
1699  *	page so user code sees what it expects.
1700  *
1701  *	Pages are most often semi-valid when the end of a file is mapped
1702  *	into memory and the file's size is not page aligned.
1703  */
1704 
1705 void
1706 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
1707 {
1708 	int b;
1709 	int i;
1710 
1711 	/*
1712 	 * Scan the valid bits looking for invalid sections that
1713 	 * must be zerod.  Invalid sub-DEV_BSIZE'd areas ( where the
1714 	 * valid bit may be set ) have already been zerod by
1715 	 * vm_page_set_validclean().
1716 	 */
1717 
1718 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
1719 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
1720 		    (m->valid & (1 << i))
1721 		) {
1722 			if (i > b) {
1723 				pmap_zero_page_area(
1724 				    VM_PAGE_TO_PHYS(m),
1725 				    b << DEV_BSHIFT,
1726 				    (i - b) << DEV_BSHIFT
1727 				);
1728 			}
1729 			b = i + 1;
1730 		}
1731 	}
1732 
1733 	/*
1734 	 * setvalid is TRUE when we can safely set the zero'd areas
1735 	 * as being valid.  We can do this if there are no cache consistency
1736 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
1737 	 */
1738 
1739 	if (setvalid)
1740 		m->valid = VM_PAGE_BITS_ALL;
1741 }
1742 
1743 /*
1744  *	vm_page_is_valid:
1745  *
1746  *	Is (partial) page valid?  Note that the case where size == 0
1747  *	will return FALSE in the degenerate case where the page is
1748  *	entirely invalid, and TRUE otherwise.
1749  *
1750  *	May not block.
1751  */
1752 
1753 int
1754 vm_page_is_valid(vm_page_t m, int base, int size)
1755 {
1756 	int bits = vm_page_bits(base, size);
1757 
1758 	if (m->valid && ((m->valid & bits) == bits))
1759 		return 1;
1760 	else
1761 		return 0;
1762 }
1763 
1764 /*
1765  * update dirty bits from pmap/mmu.  May not block.
1766  */
1767 
1768 void
1769 vm_page_test_dirty(vm_page_t m)
1770 {
1771 	if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) {
1772 		vm_page_dirty(m);
1773 	}
1774 }
1775 
1776 /*
1777  * This interface is for merging with malloc() someday.
1778  * Even if we never implement compaction so that contiguous allocation
1779  * works after initialization time, malloc()'s data structures are good
1780  * for statistics and for allocations of less than a page.
1781  */
1782 void *
1783 contigmalloc1(
1784 	unsigned long size,	/* should be size_t here and for malloc() */
1785 	struct malloc_type *type,
1786 	int flags,
1787 	unsigned long low,
1788 	unsigned long high,
1789 	unsigned long alignment,
1790 	unsigned long boundary,
1791 	vm_map_t map)
1792 {
1793 	int i, s, start;
1794 	vm_offset_t addr, phys, tmp_addr;
1795 	int pass;
1796 	vm_page_t pga = vm_page_array;
1797 
1798 	size = round_page(size);
1799 	if (size == 0)
1800 		panic("contigmalloc1: size must not be 0");
1801 	if ((alignment & (alignment - 1)) != 0)
1802 		panic("contigmalloc1: alignment must be a power of 2");
1803 	if ((boundary & (boundary - 1)) != 0)
1804 		panic("contigmalloc1: boundary must be a power of 2");
1805 
1806 	start = 0;
1807 	for (pass = 0; pass <= 1; pass++) {
1808 		s = splvm();
1809 again:
1810 		/*
1811 		 * Find first page in array that is free, within range, aligned, and
1812 		 * such that the boundary won't be crossed.
1813 		 */
1814 		for (i = start; i < cnt.v_page_count; i++) {
1815 			int pqtype;
1816 			phys = VM_PAGE_TO_PHYS(&pga[i]);
1817 			pqtype = pga[i].queue - pga[i].pc;
1818 			if (((pqtype == PQ_FREE) || (pqtype == PQ_CACHE)) &&
1819 			    (phys >= low) && (phys < high) &&
1820 			    ((phys & (alignment - 1)) == 0) &&
1821 			    (((phys ^ (phys + size - 1)) & ~(boundary - 1)) == 0))
1822 				break;
1823 		}
1824 
1825 		/*
1826 		 * If the above failed or we will exceed the upper bound, fail.
1827 		 */
1828 		if ((i == cnt.v_page_count) ||
1829 			((VM_PAGE_TO_PHYS(&pga[i]) + size) > high)) {
1830 			vm_page_t m, next;
1831 
1832 again1:
1833 			for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl);
1834 				m != NULL;
1835 				m = next) {
1836 
1837 				KASSERT(m->queue == PQ_INACTIVE,
1838 					("contigmalloc1: page %p is not PQ_INACTIVE", m));
1839 
1840 				next = TAILQ_NEXT(m, pageq);
1841 				if (vm_page_sleep_busy(m, TRUE, "vpctw0"))
1842 					goto again1;
1843 				vm_page_test_dirty(m);
1844 				if (m->dirty) {
1845 					if (m->object->type == OBJT_VNODE) {
1846 						vn_lock(m->object->handle, LK_EXCLUSIVE | LK_RETRY, curproc);
1847 						vm_object_page_clean(m->object, 0, 0, OBJPC_SYNC);
1848 						VOP_UNLOCK(m->object->handle, 0, curproc);
1849 						goto again1;
1850 					} else if (m->object->type == OBJT_SWAP ||
1851 								m->object->type == OBJT_DEFAULT) {
1852 						vm_pageout_flush(&m, 1, 0);
1853 						goto again1;
1854 					}
1855 				}
1856 				if ((m->dirty == 0) && (m->busy == 0) && (m->hold_count == 0))
1857 					vm_page_cache(m);
1858 			}
1859 
1860 			for (m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1861 				m != NULL;
1862 				m = next) {
1863 
1864 				KASSERT(m->queue == PQ_ACTIVE,
1865 					("contigmalloc1: page %p is not PQ_ACTIVE", m));
1866 
1867 				next = TAILQ_NEXT(m, pageq);
1868 				if (vm_page_sleep_busy(m, TRUE, "vpctw1"))
1869 					goto again1;
1870 				vm_page_test_dirty(m);
1871 				if (m->dirty) {
1872 					if (m->object->type == OBJT_VNODE) {
1873 						vn_lock(m->object->handle, LK_EXCLUSIVE | LK_RETRY, curproc);
1874 						vm_object_page_clean(m->object, 0, 0, OBJPC_SYNC);
1875 						VOP_UNLOCK(m->object->handle, 0, curproc);
1876 						goto again1;
1877 					} else if (m->object->type == OBJT_SWAP ||
1878 								m->object->type == OBJT_DEFAULT) {
1879 						vm_pageout_flush(&m, 1, 0);
1880 						goto again1;
1881 					}
1882 				}
1883 				if ((m->dirty == 0) && (m->busy == 0) && (m->hold_count == 0))
1884 					vm_page_cache(m);
1885 			}
1886 
1887 			splx(s);
1888 			continue;
1889 		}
1890 		start = i;
1891 
1892 		/*
1893 		 * Check successive pages for contiguous and free.
1894 		 */
1895 		for (i = start + 1; i < (start + size / PAGE_SIZE); i++) {
1896 			int pqtype;
1897 			pqtype = pga[i].queue - pga[i].pc;
1898 			if ((VM_PAGE_TO_PHYS(&pga[i]) !=
1899 			    (VM_PAGE_TO_PHYS(&pga[i - 1]) + PAGE_SIZE)) ||
1900 			    ((pqtype != PQ_FREE) && (pqtype != PQ_CACHE))) {
1901 				start++;
1902 				goto again;
1903 			}
1904 		}
1905 
1906 		for (i = start; i < (start + size / PAGE_SIZE); i++) {
1907 			int pqtype;
1908 			vm_page_t m = &pga[i];
1909 
1910 			pqtype = m->queue - m->pc;
1911 			if (pqtype == PQ_CACHE) {
1912 				vm_page_busy(m);
1913 				vm_page_free(m);
1914 			}
1915 			vm_page_unqueue_nowakeup(m);
1916 			m->valid = VM_PAGE_BITS_ALL;
1917 			if (m->flags & PG_ZERO)
1918 				vm_page_zero_count--;
1919 			m->flags = 0;
1920 			KASSERT(m->dirty == 0, ("contigmalloc1: page %p was dirty", m));
1921 			m->wire_count = 0;
1922 			m->busy = 0;
1923 			m->object = NULL;
1924 		}
1925 
1926 		/*
1927 		 * We've found a contiguous chunk that meets are requirements.
1928 		 * Allocate kernel VM, unfree and assign the physical pages to it and
1929 		 * return kernel VM pointer.
1930 		 */
1931 		vm_map_lock(map);
1932 		if (vm_map_findspace(map, vm_map_min(map), size, &addr) !=
1933 		    KERN_SUCCESS) {
1934 			/*
1935 			 * XXX We almost never run out of kernel virtual
1936 			 * space, so we don't make the allocated memory
1937 			 * above available.
1938 			 */
1939 			vm_map_unlock(map);
1940 			splx(s);
1941 			return (NULL);
1942 		}
1943 		vm_object_reference(kernel_object);
1944 		vm_map_insert(map, kernel_object, addr - VM_MIN_KERNEL_ADDRESS,
1945 		    addr, addr + size, VM_PROT_ALL, VM_PROT_ALL, 0);
1946 		vm_map_unlock(map);
1947 
1948 		tmp_addr = addr;
1949 		for (i = start; i < (start + size / PAGE_SIZE); i++) {
1950 			vm_page_t m = &pga[i];
1951 			vm_page_insert(m, kernel_object,
1952 				OFF_TO_IDX(tmp_addr - VM_MIN_KERNEL_ADDRESS));
1953 			tmp_addr += PAGE_SIZE;
1954 		}
1955 		vm_map_pageable(map, addr, addr + size, FALSE);
1956 
1957 		splx(s);
1958 		return ((void *)addr);
1959 	}
1960 	return NULL;
1961 }
1962 
1963 void *
1964 contigmalloc(
1965 	unsigned long size,	/* should be size_t here and for malloc() */
1966 	struct malloc_type *type,
1967 	int flags,
1968 	unsigned long low,
1969 	unsigned long high,
1970 	unsigned long alignment,
1971 	unsigned long boundary)
1972 {
1973 	return contigmalloc1(size, type, flags, low, high, alignment, boundary,
1974 			     kernel_map);
1975 }
1976 
1977 void
1978 contigfree(void *addr, unsigned long size, struct malloc_type *type)
1979 {
1980 	kmem_free(kernel_map, (vm_offset_t)addr, size);
1981 }
1982 
1983 vm_offset_t
1984 vm_page_alloc_contig(
1985 	vm_offset_t size,
1986 	vm_offset_t low,
1987 	vm_offset_t high,
1988 	vm_offset_t alignment)
1989 {
1990 	return ((vm_offset_t)contigmalloc1(size, M_DEVBUF, M_NOWAIT, low, high,
1991 					  alignment, 0ul, kernel_map));
1992 }
1993 
1994 #include "opt_ddb.h"
1995 #ifdef DDB
1996 #include <sys/kernel.h>
1997 
1998 #include <ddb/ddb.h>
1999 
2000 DB_SHOW_COMMAND(page, vm_page_print_page_info)
2001 {
2002 	db_printf("cnt.v_free_count: %d\n", cnt.v_free_count);
2003 	db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count);
2004 	db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count);
2005 	db_printf("cnt.v_active_count: %d\n", cnt.v_active_count);
2006 	db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count);
2007 	db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved);
2008 	db_printf("cnt.v_free_min: %d\n", cnt.v_free_min);
2009 	db_printf("cnt.v_free_target: %d\n", cnt.v_free_target);
2010 	db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min);
2011 	db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target);
2012 }
2013 
2014 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
2015 {
2016 	int i;
2017 	db_printf("PQ_FREE:");
2018 	for(i=0;i<PQ_L2_SIZE;i++) {
2019 		db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt);
2020 	}
2021 	db_printf("\n");
2022 
2023 	db_printf("PQ_CACHE:");
2024 	for(i=0;i<PQ_L2_SIZE;i++) {
2025 		db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt);
2026 	}
2027 	db_printf("\n");
2028 
2029 	db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n",
2030 		vm_page_queues[PQ_ACTIVE].lcnt,
2031 		vm_page_queues[PQ_INACTIVE].lcnt);
2032 }
2033 #endif /* DDB */
2034