xref: /dragonfly/sys/vm/vm_page.c (revision 4e7eb5cc)
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  * $DragonFly: src/sys/vm/vm_page.c,v 1.16 2004/01/20 05:04:08 dillon Exp $
39  */
40 
41 /*
42  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
43  * All rights reserved.
44  *
45  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
46  *
47  * Permission to use, copy, modify and distribute this software and
48  * its documentation is hereby granted, provided that both the copyright
49  * notice and this permission notice appear in all copies of the
50  * software, derivative works or modified versions, and any portions
51  * thereof, and that both notices appear in supporting documentation.
52  *
53  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
54  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
55  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
56  *
57  * Carnegie Mellon requests users of this software to return to
58  *
59  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
60  *  School of Computer Science
61  *  Carnegie Mellon University
62  *  Pittsburgh PA 15213-3890
63  *
64  * any improvements or extensions that they make and grant Carnegie the
65  * rights to redistribute these changes.
66  */
67 
68 /*
69  *	Resident memory management module.
70  */
71 
72 #include <sys/param.h>
73 #include <sys/systm.h>
74 #include <sys/malloc.h>
75 #include <sys/proc.h>
76 #include <sys/vmmeter.h>
77 #include <sys/vnode.h>
78 
79 #include <vm/vm.h>
80 #include <vm/vm_param.h>
81 #include <sys/lock.h>
82 #include <vm/vm_kern.h>
83 #include <vm/pmap.h>
84 #include <vm/vm_map.h>
85 #include <vm/vm_object.h>
86 #include <vm/vm_page.h>
87 #include <vm/vm_pageout.h>
88 #include <vm/vm_pager.h>
89 #include <vm/vm_extern.h>
90 #include <vm/vm_page2.h>
91 
92 static void	vm_page_queue_init (void);
93 static vm_page_t vm_page_select_cache (vm_object_t, vm_pindex_t);
94 
95 /*
96  *	Associated with page of user-allocatable memory is a
97  *	page structure.
98  */
99 
100 static struct vm_page **vm_page_buckets; /* Array of buckets */
101 static int vm_page_bucket_count;	/* How big is array? */
102 static int vm_page_hash_mask;		/* Mask for hash function */
103 static volatile int vm_page_bucket_generation;
104 
105 struct vpgqueues vm_page_queues[PQ_COUNT];
106 
107 static void
108 vm_page_queue_init(void) {
109 	int i;
110 
111 	for(i=0;i<PQ_L2_SIZE;i++) {
112 		vm_page_queues[PQ_FREE+i].cnt = &vmstats.v_free_count;
113 	}
114 	vm_page_queues[PQ_INACTIVE].cnt = &vmstats.v_inactive_count;
115 
116 	vm_page_queues[PQ_ACTIVE].cnt = &vmstats.v_active_count;
117 	vm_page_queues[PQ_HOLD].cnt = &vmstats.v_active_count;
118 	for(i=0;i<PQ_L2_SIZE;i++) {
119 		vm_page_queues[PQ_CACHE+i].cnt = &vmstats.v_cache_count;
120 	}
121 	for(i=0;i<PQ_COUNT;i++) {
122 		TAILQ_INIT(&vm_page_queues[i].pl);
123 	}
124 }
125 
126 vm_page_t vm_page_array = 0;
127 int vm_page_array_size = 0;
128 long first_page = 0;
129 int vm_page_zero_count = 0;
130 
131 static __inline int vm_page_hash (vm_object_t object, vm_pindex_t pindex);
132 static void vm_page_free_wakeup (void);
133 
134 /*
135  *	vm_set_page_size:
136  *
137  *	Sets the page size, perhaps based upon the memory
138  *	size.  Must be called before any use of page-size
139  *	dependent functions.
140  */
141 void
142 vm_set_page_size(void)
143 {
144 	if (vmstats.v_page_size == 0)
145 		vmstats.v_page_size = PAGE_SIZE;
146 	if (((vmstats.v_page_size - 1) & vmstats.v_page_size) != 0)
147 		panic("vm_set_page_size: page size not a power of two");
148 }
149 
150 /*
151  *	vm_add_new_page:
152  *
153  *	Add a new page to the freelist for use by the system.  New pages
154  *	are added to both the head and tail of the associated free page
155  *	queue in a bottom-up fashion, so both zero'd and non-zero'd page
156  *	requests pull 'recent' adds (higher physical addresses) first.
157  *
158  *	Must be called at splhigh().
159  */
160 vm_page_t
161 vm_add_new_page(vm_paddr_t pa)
162 {
163 	vm_page_t m;
164 	struct vpgqueues *vpq;
165 
166 	++vmstats.v_page_count;
167 	++vmstats.v_free_count;
168 	m = PHYS_TO_VM_PAGE(pa);
169 	m->phys_addr = pa;
170 	m->flags = 0;
171 	m->pc = (pa >> PAGE_SHIFT) & PQ_L2_MASK;
172 	m->queue = m->pc + PQ_FREE;
173 	vpq = &vm_page_queues[m->queue];
174 	if (vpq->flipflop)
175 		TAILQ_INSERT_TAIL(&vpq->pl, m, pageq);
176 	else
177 		TAILQ_INSERT_HEAD(&vpq->pl, m, pageq);
178 	vpq->flipflop = 1 - vpq->flipflop;
179 	vm_page_queues[m->queue].lcnt++;
180 	return (m);
181 }
182 
183 /*
184  *	vm_page_startup:
185  *
186  *	Initializes the resident memory module.
187  *
188  *	Allocates memory for the page cells, and
189  *	for the object/offset-to-page hash table headers.
190  *	Each page cell is initialized and placed on the free list.
191  */
192 
193 vm_offset_t
194 vm_page_startup(vm_offset_t starta, vm_offset_t enda, vm_offset_t vaddr)
195 {
196 	vm_offset_t mapped;
197 	struct vm_page **bucket;
198 	vm_size_t npages;
199 	vm_paddr_t page_range;
200 	vm_paddr_t new_end;
201 	int i;
202 	vm_paddr_t pa;
203 	int nblocks;
204 	vm_paddr_t last_pa;
205 
206 	/* the biggest memory array is the second group of pages */
207 	vm_paddr_t end;
208 	vm_paddr_t biggestone, biggestsize;
209 
210 	vm_paddr_t total;
211 
212 	total = 0;
213 	biggestsize = 0;
214 	biggestone = 0;
215 	nblocks = 0;
216 	vaddr = round_page(vaddr);
217 
218 	for (i = 0; phys_avail[i + 1]; i += 2) {
219 		phys_avail[i] = round_page(phys_avail[i]);
220 		phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
221 	}
222 
223 	for (i = 0; phys_avail[i + 1]; i += 2) {
224 		vm_paddr_t size = phys_avail[i + 1] - phys_avail[i];
225 
226 		if (size > biggestsize) {
227 			biggestone = i;
228 			biggestsize = size;
229 		}
230 		++nblocks;
231 		total += size;
232 	}
233 
234 	end = phys_avail[biggestone+1];
235 
236 	/*
237 	 * Initialize the queue headers for the free queue, the active queue
238 	 * and the inactive queue.
239 	 */
240 
241 	vm_page_queue_init();
242 
243 	/*
244 	 * Allocate (and initialize) the hash table buckets.
245 	 *
246 	 * The number of buckets MUST BE a power of 2, and the actual value is
247 	 * the next power of 2 greater than the number of physical pages in
248 	 * the system.
249 	 *
250 	 * We make the hash table approximately 2x the number of pages to
251 	 * reduce the chain length.  This is about the same size using the
252 	 * singly-linked list as the 1x hash table we were using before
253 	 * using TAILQ but the chain length will be smaller.
254 	 *
255 	 * Note: This computation can be tweaked if desired.
256 	 */
257 	vm_page_buckets = (struct vm_page **)vaddr;
258 	bucket = vm_page_buckets;
259 	if (vm_page_bucket_count == 0) {
260 		vm_page_bucket_count = 1;
261 		while (vm_page_bucket_count < atop(total))
262 			vm_page_bucket_count <<= 1;
263 	}
264 	vm_page_bucket_count <<= 1;
265 	vm_page_hash_mask = vm_page_bucket_count - 1;
266 
267 	/*
268 	 * Validate these addresses.
269 	 */
270 	new_end = end - vm_page_bucket_count * sizeof(struct vm_page *);
271 	new_end = trunc_page(new_end);
272 	mapped = round_page(vaddr);
273 	vaddr = pmap_map(mapped, new_end, end,
274 	    VM_PROT_READ | VM_PROT_WRITE);
275 	vaddr = round_page(vaddr);
276 	bzero((caddr_t) mapped, vaddr - mapped);
277 
278 	for (i = 0; i < vm_page_bucket_count; i++) {
279 		*bucket = NULL;
280 		bucket++;
281 	}
282 
283 	/*
284 	 * Compute the number of pages of memory that will be available for
285 	 * use (taking into account the overhead of a page structure per
286 	 * page).
287 	 */
288 
289 	first_page = phys_avail[0] / PAGE_SIZE;
290 
291 	page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page;
292 	npages = (total - (page_range * sizeof(struct vm_page)) -
293 	    (end - new_end)) / PAGE_SIZE;
294 
295 	end = new_end;
296 	/*
297 	 * Initialize the mem entry structures now, and put them in the free
298 	 * queue.
299 	 */
300 	vm_page_array = (vm_page_t) vaddr;
301 	mapped = vaddr;
302 
303 	/*
304 	 * Validate these addresses.
305 	 */
306 
307 	new_end = trunc_page(end - page_range * sizeof(struct vm_page));
308 	mapped = pmap_map(mapped, new_end, end,
309 	    VM_PROT_READ | VM_PROT_WRITE);
310 
311 	/*
312 	 * Clear all of the page structures
313 	 */
314 	bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
315 	vm_page_array_size = page_range;
316 
317 	/*
318 	 * Construct the free queue(s) in ascending order (by physical
319 	 * address) so that the first 16MB of physical memory is allocated
320 	 * last rather than first.  On large-memory machines, this avoids
321 	 * the exhaustion of low physical memory before isa_dmainit has run.
322 	 */
323 	vmstats.v_page_count = 0;
324 	vmstats.v_free_count = 0;
325 	for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) {
326 		pa = phys_avail[i];
327 		if (i == biggestone)
328 			last_pa = new_end;
329 		else
330 			last_pa = phys_avail[i + 1];
331 		while (pa < last_pa && npages-- > 0) {
332 			vm_add_new_page(pa);
333 			pa += PAGE_SIZE;
334 		}
335 	}
336 	return (mapped);
337 }
338 
339 /*
340  *	vm_page_hash:
341  *
342  *	Distributes the object/offset key pair among hash buckets.
343  *
344  *	NOTE:  This macro depends on vm_page_bucket_count being a power of 2.
345  *	This routine may not block.
346  *
347  *	We try to randomize the hash based on the object to spread the pages
348  *	out in the hash table without it costing us too much.
349  */
350 static __inline int
351 vm_page_hash(vm_object_t object, vm_pindex_t pindex)
352 {
353 	int i = ((uintptr_t)object + pindex) ^ object->hash_rand;
354 
355 	return(i & vm_page_hash_mask);
356 }
357 
358 void
359 vm_page_unhold(vm_page_t mem)
360 {
361 	--mem->hold_count;
362 	KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!"));
363 	if (mem->hold_count == 0 && mem->queue == PQ_HOLD)
364 		vm_page_free_toq(mem);
365 }
366 
367 /*
368  *	vm_page_insert:		[ internal use only ]
369  *
370  *	Inserts the given mem entry into the object and object list.
371  *
372  *	The pagetables are not updated but will presumably fault the page
373  *	in if necessary, or if a kernel page the caller will at some point
374  *	enter the page into the kernel's pmap.  We are not allowed to block
375  *	here so we *can't* do this anyway.
376  *
377  *	The object and page must be locked, and must be splhigh.
378  *	This routine may not block.
379  */
380 
381 void
382 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
383 {
384 	struct vm_page **bucket;
385 
386 	if (m->object != NULL)
387 		panic("vm_page_insert: already inserted");
388 
389 	/*
390 	 * Record the object/offset pair in this page
391 	 */
392 
393 	m->object = object;
394 	m->pindex = pindex;
395 
396 	/*
397 	 * Insert it into the object_object/offset hash table
398 	 */
399 
400 	bucket = &vm_page_buckets[vm_page_hash(object, pindex)];
401 	m->hnext = *bucket;
402 	*bucket = m;
403 	vm_page_bucket_generation++;
404 
405 	/*
406 	 * Now link into the object's list of backed pages.
407 	 */
408 
409 	TAILQ_INSERT_TAIL(&object->memq, m, listq);
410 	object->generation++;
411 
412 	/*
413 	 * show that the object has one more resident page.
414 	 */
415 
416 	object->resident_page_count++;
417 
418 	/*
419 	 * Since we are inserting a new and possibly dirty page,
420 	 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags.
421 	 */
422 	if (m->flags & PG_WRITEABLE)
423 		vm_object_set_writeable_dirty(object);
424 }
425 
426 /*
427  *	vm_page_remove:
428  *				NOTE: used by device pager as well -wfj
429  *
430  *	Removes the given mem entry from the object/offset-page
431  *	table and the object page list, but do not invalidate/terminate
432  *	the backing store.
433  *
434  *	The object and page must be locked, and at splhigh.
435  *	The underlying pmap entry (if any) is NOT removed here.
436  *	This routine may not block.
437  */
438 
439 void
440 vm_page_remove(vm_page_t m)
441 {
442 	vm_object_t object;
443 
444 	if (m->object == NULL)
445 		return;
446 
447 	if ((m->flags & PG_BUSY) == 0) {
448 		panic("vm_page_remove: page not busy");
449 	}
450 
451 	/*
452 	 * Basically destroy the page.
453 	 */
454 
455 	vm_page_wakeup(m);
456 
457 	object = m->object;
458 
459 	/*
460 	 * Remove from the object_object/offset hash table.  The object
461 	 * must be on the hash queue, we will panic if it isn't
462 	 *
463 	 * Note: we must NULL-out m->hnext to prevent loops in detached
464 	 * buffers with vm_page_lookup().
465 	 */
466 
467 	{
468 		struct vm_page **bucket;
469 
470 		bucket = &vm_page_buckets[vm_page_hash(m->object, m->pindex)];
471 		while (*bucket != m) {
472 			if (*bucket == NULL)
473 				panic("vm_page_remove(): page not found in hash");
474 			bucket = &(*bucket)->hnext;
475 		}
476 		*bucket = m->hnext;
477 		m->hnext = NULL;
478 		vm_page_bucket_generation++;
479 	}
480 
481 	/*
482 	 * Now remove from the object's list of backed pages.
483 	 */
484 
485 	TAILQ_REMOVE(&object->memq, m, listq);
486 
487 	/*
488 	 * And show that the object has one fewer resident page.
489 	 */
490 
491 	object->resident_page_count--;
492 	object->generation++;
493 
494 	m->object = NULL;
495 }
496 
497 /*
498  *	vm_page_lookup:
499  *
500  *	Returns the page associated with the object/offset
501  *	pair specified; if none is found, NULL is returned.
502  *
503  *	NOTE: the code below does not lock.  It will operate properly if
504  *	an interrupt makes a change, but the generation algorithm will not
505  *	operate properly in an SMP environment where both cpu's are able to run
506  *	kernel code simultaneously.
507  *
508  *	The object must be locked.  No side effects.
509  *	This routine may not block.
510  *	This is a critical path routine
511  */
512 
513 vm_page_t
514 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
515 {
516 	vm_page_t m;
517 	struct vm_page **bucket;
518 	int generation;
519 
520 	/*
521 	 * Search the hash table for this object/offset pair
522 	 */
523 
524 retry:
525 	generation = vm_page_bucket_generation;
526 	bucket = &vm_page_buckets[vm_page_hash(object, pindex)];
527 	for (m = *bucket; m != NULL; m = m->hnext) {
528 		if ((m->object == object) && (m->pindex == pindex)) {
529 			if (vm_page_bucket_generation != generation)
530 				goto retry;
531 			return (m);
532 		}
533 	}
534 	if (vm_page_bucket_generation != generation)
535 		goto retry;
536 	return (NULL);
537 }
538 
539 /*
540  *	vm_page_rename:
541  *
542  *	Move the given memory entry from its
543  *	current object to the specified target object/offset.
544  *
545  *	The object must be locked.
546  *	This routine may not block.
547  *
548  *	Note: this routine will raise itself to splvm(), the caller need not.
549  *
550  *	Note: swap associated with the page must be invalidated by the move.  We
551  *	      have to do this for several reasons:  (1) we aren't freeing the
552  *	      page, (2) we are dirtying the page, (3) the VM system is probably
553  *	      moving the page from object A to B, and will then later move
554  *	      the backing store from A to B and we can't have a conflict.
555  *
556  *	Note: we *always* dirty the page.  It is necessary both for the
557  *	      fact that we moved it, and because we may be invalidating
558  *	      swap.  If the page is on the cache, we have to deactivate it
559  *	      or vm_page_dirty() will panic.  Dirty pages are not allowed
560  *	      on the cache.
561  */
562 
563 void
564 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
565 {
566 	int s;
567 
568 	s = splvm();
569 	vm_page_remove(m);
570 	vm_page_insert(m, new_object, new_pindex);
571 	if (m->queue - m->pc == PQ_CACHE)
572 		vm_page_deactivate(m);
573 	vm_page_dirty(m);
574 	splx(s);
575 }
576 
577 /*
578  * vm_page_unqueue_nowakeup:
579  *
580  * 	vm_page_unqueue() without any wakeup
581  *
582  *	This routine must be called at splhigh().
583  *	This routine may not block.
584  */
585 
586 void
587 vm_page_unqueue_nowakeup(vm_page_t m)
588 {
589 	int queue = m->queue;
590 	struct vpgqueues *pq;
591 	if (queue != PQ_NONE) {
592 		pq = &vm_page_queues[queue];
593 		m->queue = PQ_NONE;
594 		TAILQ_REMOVE(&pq->pl, m, pageq);
595 		(*pq->cnt)--;
596 		pq->lcnt--;
597 	}
598 }
599 
600 /*
601  * vm_page_unqueue:
602  *
603  *	Remove a page from its queue.
604  *
605  *	This routine must be called at splhigh().
606  *	This routine may not block.
607  */
608 
609 void
610 vm_page_unqueue(vm_page_t m)
611 {
612 	int queue = m->queue;
613 	struct vpgqueues *pq;
614 	if (queue != PQ_NONE) {
615 		m->queue = PQ_NONE;
616 		pq = &vm_page_queues[queue];
617 		TAILQ_REMOVE(&pq->pl, m, pageq);
618 		(*pq->cnt)--;
619 		pq->lcnt--;
620 		if ((queue - m->pc) == PQ_CACHE) {
621 			if (vm_paging_needed())
622 				pagedaemon_wakeup();
623 		}
624 	}
625 }
626 
627 #if PQ_L2_SIZE > 1
628 
629 /*
630  *	vm_page_list_find:
631  *
632  *	Find a page on the specified queue with color optimization.
633  *
634  *	The page coloring optimization attempts to locate a page
635  *	that does not overload other nearby pages in the object in
636  *	the cpu's L1 or L2 caches.  We need this optimization because
637  *	cpu caches tend to be physical caches, while object spaces tend
638  *	to be virtual.
639  *
640  *	This routine must be called at splvm().
641  *	This routine may not block.
642  *
643  *	This routine may only be called from the vm_page_list_find() macro
644  *	in vm_page.h
645  */
646 vm_page_t
647 _vm_page_list_find(int basequeue, int index)
648 {
649 	int i;
650 	vm_page_t m = NULL;
651 	struct vpgqueues *pq;
652 
653 	pq = &vm_page_queues[basequeue];
654 
655 	/*
656 	 * Note that for the first loop, index+i and index-i wind up at the
657 	 * same place.  Even though this is not totally optimal, we've already
658 	 * blown it by missing the cache case so we do not care.
659 	 */
660 
661 	for(i = PQ_L2_SIZE / 2; i > 0; --i) {
662 		if ((m = TAILQ_FIRST(&pq[(index + i) & PQ_L2_MASK].pl)) != NULL)
663 			break;
664 
665 		if ((m = TAILQ_FIRST(&pq[(index - i) & PQ_L2_MASK].pl)) != NULL)
666 			break;
667 	}
668 	return(m);
669 }
670 
671 #endif
672 
673 /*
674  *	vm_page_select_cache:
675  *
676  *	Find a page on the cache queue with color optimization.  As pages
677  *	might be found, but not applicable, they are deactivated.  This
678  *	keeps us from using potentially busy cached pages.
679  *
680  *	This routine must be called at splvm().
681  *	This routine may not block.
682  */
683 vm_page_t
684 vm_page_select_cache(vm_object_t object, vm_pindex_t pindex)
685 {
686 	vm_page_t m;
687 
688 	while (TRUE) {
689 		m = vm_page_list_find(
690 		    PQ_CACHE,
691 		    (pindex + object->pg_color) & PQ_L2_MASK,
692 		    FALSE
693 		);
694 		if (m && ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy ||
695 			       m->hold_count || m->wire_count)) {
696 			vm_page_deactivate(m);
697 			continue;
698 		}
699 		return m;
700 	}
701 }
702 
703 /*
704  *	vm_page_select_free:
705  *
706  *	Find a free or zero page, with specified preference.  We attempt to
707  *	inline the nominal case and fall back to _vm_page_select_free()
708  *	otherwise.
709  *
710  *	This routine must be called at splvm().
711  *	This routine may not block.
712  */
713 
714 static __inline vm_page_t
715 vm_page_select_free(vm_object_t object, vm_pindex_t pindex, boolean_t prefer_zero)
716 {
717 	vm_page_t m;
718 
719 	m = vm_page_list_find(
720 		PQ_FREE,
721 		(pindex + object->pg_color) & PQ_L2_MASK,
722 		prefer_zero
723 	);
724 	return(m);
725 }
726 
727 /*
728  *	vm_page_alloc:
729  *
730  *	Allocate and return a memory cell associated
731  *	with this VM object/offset pair.
732  *
733  *	page_req classes:
734  *	VM_ALLOC_NORMAL		allow use of cache pages, nominal free drain
735  *	VM_ALLOC_SYSTEM		greater free drain
736  *	VM_ALLOC_INTERRUPT	allow free list to be completely drained
737  *	VM_ALLOC_ZERO		advisory request for pre-zero'd page
738  *
739  *	Object must be locked.
740  *	This routine may not block.
741  *
742  *	Additional special handling is required when called from an
743  *	interrupt (VM_ALLOC_INTERRUPT).  We are not allowed to mess with
744  *	the page cache in this case.
745  */
746 
747 vm_page_t
748 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req)
749 {
750 	vm_page_t m = NULL;
751 	int s;
752 
753 	KASSERT(!vm_page_lookup(object, pindex),
754 		("vm_page_alloc: page already allocated"));
755 	KKASSERT(page_req &
756 		(VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
757 
758 	/*
759 	 * The pager is allowed to eat deeper into the free page list.
760 	 */
761 	if (curthread == pagethread)
762 		page_req |= VM_ALLOC_SYSTEM;
763 
764 	s = splvm();
765 loop:
766 	if (vmstats.v_free_count > vmstats.v_free_reserved ||
767 	    ((page_req & VM_ALLOC_INTERRUPT) && vmstats.v_free_count > 0) ||
768 	    ((page_req & VM_ALLOC_SYSTEM) && vmstats.v_cache_count == 0 &&
769 		vmstats.v_free_count > vmstats.v_interrupt_free_min)
770 	) {
771 		/*
772 		 * The free queue has sufficient free pages to take one out.
773 		 */
774 		if (page_req & VM_ALLOC_ZERO)
775 			m = vm_page_select_free(object, pindex, TRUE);
776 		else
777 			m = vm_page_select_free(object, pindex, FALSE);
778 	} else if (page_req & VM_ALLOC_NORMAL) {
779 		/*
780 		 * Allocatable from the cache (non-interrupt only).  On
781 		 * success, we must free the page and try again, thus
782 		 * ensuring that vmstats.v_*_free_min counters are replenished.
783 		 */
784 #ifdef INVARIANTS
785 		if (curthread->td_preempted) {
786 			printf("vm_page_alloc(): warning, attempt to allocate"
787 				" cache page from preempting interrupt\n");
788 			m = NULL;
789 		} else {
790 			m = vm_page_select_cache(object, pindex);
791 		}
792 #else
793 		m = vm_page_select_cache(object, pindex);
794 #endif
795 		/*
796 		 * On succuess move the page into the free queue and loop.
797 		 */
798 		if (m != NULL) {
799 			KASSERT(m->dirty == 0,
800 			    ("Found dirty cache page %p", m));
801 			vm_page_busy(m);
802 			vm_page_protect(m, VM_PROT_NONE);
803 			vm_page_free(m);
804 			goto loop;
805 		}
806 
807 		/*
808 		 * On failure return NULL
809 		 */
810 		splx(s);
811 #if defined(DIAGNOSTIC)
812 		if (vmstats.v_cache_count > 0)
813 			printf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", vmstats.v_cache_count);
814 #endif
815 		vm_pageout_deficit++;
816 		pagedaemon_wakeup();
817 		return (NULL);
818 	} else {
819 		/*
820 		 * No pages available, wakeup the pageout daemon and give up.
821 		 */
822 		splx(s);
823 		vm_pageout_deficit++;
824 		pagedaemon_wakeup();
825 		return (NULL);
826 	}
827 
828 	/*
829 	 * Good page found.
830 	 */
831 	KASSERT(m != NULL, ("vm_page_alloc(): missing page on free queue\n"));
832 
833 	/*
834 	 * Remove from free queue
835 	 */
836 	vm_page_unqueue_nowakeup(m);
837 
838 	/*
839 	 * Initialize structure.  Only the PG_ZERO flag is inherited.
840 	 */
841 	if (m->flags & PG_ZERO) {
842 		vm_page_zero_count--;
843 		m->flags = PG_ZERO | PG_BUSY;
844 	} else {
845 		m->flags = PG_BUSY;
846 	}
847 	m->wire_count = 0;
848 	m->hold_count = 0;
849 	m->act_count = 0;
850 	m->busy = 0;
851 	m->valid = 0;
852 	KASSERT(m->dirty == 0,
853 		("vm_page_alloc: free/cache page %p was dirty", m));
854 
855 	/*
856 	 * vm_page_insert() is safe prior to the splx().  Note also that
857 	 * inserting a page here does not insert it into the pmap (which
858 	 * could cause us to block allocating memory).  We cannot block
859 	 * anywhere.
860 	 */
861 	vm_page_insert(m, object, pindex);
862 
863 	/*
864 	 * Don't wakeup too often - wakeup the pageout daemon when
865 	 * we would be nearly out of memory.
866 	 */
867 	if (vm_paging_needed())
868 		pagedaemon_wakeup();
869 
870 	splx(s);
871 	return (m);
872 }
873 
874 /*
875  *	vm_wait:	(also see VM_WAIT macro)
876  *
877  *	Block until free pages are available for allocation
878  *	- Called in various places before memory allocations.
879  */
880 
881 void
882 vm_wait(void)
883 {
884 	int s;
885 
886 	s = splvm();
887 	if (curthread == pagethread) {
888 		vm_pageout_pages_needed = 1;
889 		tsleep(&vm_pageout_pages_needed, 0, "VMWait", 0);
890 	} else {
891 		if (!vm_pages_needed) {
892 			vm_pages_needed = 1;
893 			wakeup(&vm_pages_needed);
894 		}
895 		tsleep(&vmstats.v_free_count, 0, "vmwait", 0);
896 	}
897 	splx(s);
898 }
899 
900 /*
901  *	vm_waitpfault:	(also see VM_WAITPFAULT macro)
902  *
903  *	Block until free pages are available for allocation
904  *	- Called only in vm_fault so that processes page faulting
905  *	  can be easily tracked.
906  *	- Sleeps at a lower priority than vm_wait() so that vm_wait()ing
907  *	  processes will be able to grab memory first.  Do not change
908  *	  this balance without careful testing first.
909  */
910 
911 void
912 vm_waitpfault(void)
913 {
914 	int s;
915 
916 	s = splvm();
917 	if (!vm_pages_needed) {
918 		vm_pages_needed = 1;
919 		wakeup(&vm_pages_needed);
920 	}
921 	tsleep(&vmstats.v_free_count, 0, "pfault", 0);
922 	splx(s);
923 }
924 
925 /*
926  *	vm_page_activate:
927  *
928  *	Put the specified page on the active list (if appropriate).
929  *	Ensure that act_count is at least ACT_INIT but do not otherwise
930  *	mess with it.
931  *
932  *	The page queues must be locked.
933  *	This routine may not block.
934  */
935 void
936 vm_page_activate(vm_page_t m)
937 {
938 	int s;
939 
940 	s = splvm();
941 	if (m->queue != PQ_ACTIVE) {
942 		if ((m->queue - m->pc) == PQ_CACHE)
943 			mycpu->gd_cnt.v_reactivated++;
944 
945 		vm_page_unqueue(m);
946 
947 		if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
948 			m->queue = PQ_ACTIVE;
949 			vm_page_queues[PQ_ACTIVE].lcnt++;
950 			TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
951 			if (m->act_count < ACT_INIT)
952 				m->act_count = ACT_INIT;
953 			vmstats.v_active_count++;
954 		}
955 	} else {
956 		if (m->act_count < ACT_INIT)
957 			m->act_count = ACT_INIT;
958 	}
959 
960 	splx(s);
961 }
962 
963 /*
964  *	vm_page_free_wakeup:
965  *
966  *	Helper routine for vm_page_free_toq() and vm_page_cache().  This
967  *	routine is called when a page has been added to the cache or free
968  *	queues.
969  *
970  *	This routine may not block.
971  *	This routine must be called at splvm()
972  */
973 static __inline void
974 vm_page_free_wakeup(void)
975 {
976 	/*
977 	 * if pageout daemon needs pages, then tell it that there are
978 	 * some free.
979 	 */
980 	if (vm_pageout_pages_needed &&
981 	    vmstats.v_cache_count + vmstats.v_free_count >= vmstats.v_pageout_free_min) {
982 		wakeup(&vm_pageout_pages_needed);
983 		vm_pageout_pages_needed = 0;
984 	}
985 	/*
986 	 * wakeup processes that are waiting on memory if we hit a
987 	 * high water mark. And wakeup scheduler process if we have
988 	 * lots of memory. this process will swapin processes.
989 	 */
990 	if (vm_pages_needed && !vm_page_count_min()) {
991 		vm_pages_needed = 0;
992 		wakeup(&vmstats.v_free_count);
993 	}
994 }
995 
996 /*
997  *	vm_page_free_toq:
998  *
999  *	Returns the given page to the PQ_FREE list,
1000  *	disassociating it with any VM object.
1001  *
1002  *	Object and page must be locked prior to entry.
1003  *	This routine may not block.
1004  */
1005 
1006 void
1007 vm_page_free_toq(vm_page_t m)
1008 {
1009 	int s;
1010 	struct vpgqueues *pq;
1011 	vm_object_t object = m->object;
1012 
1013 	s = splvm();
1014 
1015 	mycpu->gd_cnt.v_tfree++;
1016 
1017 	if (m->busy || ((m->queue - m->pc) == PQ_FREE)) {
1018 		printf(
1019 		"vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n",
1020 		    (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0,
1021 		    m->hold_count);
1022 		if ((m->queue - m->pc) == PQ_FREE)
1023 			panic("vm_page_free: freeing free page");
1024 		else
1025 			panic("vm_page_free: freeing busy page");
1026 	}
1027 
1028 	/*
1029 	 * unqueue, then remove page.  Note that we cannot destroy
1030 	 * the page here because we do not want to call the pager's
1031 	 * callback routine until after we've put the page on the
1032 	 * appropriate free queue.
1033 	 */
1034 
1035 	vm_page_unqueue_nowakeup(m);
1036 	vm_page_remove(m);
1037 
1038 	/*
1039 	 * If fictitious remove object association and
1040 	 * return, otherwise delay object association removal.
1041 	 */
1042 
1043 	if ((m->flags & PG_FICTITIOUS) != 0) {
1044 		splx(s);
1045 		return;
1046 	}
1047 
1048 	m->valid = 0;
1049 	vm_page_undirty(m);
1050 
1051 	if (m->wire_count != 0) {
1052 		if (m->wire_count > 1) {
1053 			panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx",
1054 				m->wire_count, (long)m->pindex);
1055 		}
1056 		panic("vm_page_free: freeing wired page\n");
1057 	}
1058 
1059 	/*
1060 	 * If we've exhausted the object's resident pages we want to free
1061 	 * it up.
1062 	 */
1063 
1064 	if (object &&
1065 	    (object->type == OBJT_VNODE) &&
1066 	    ((object->flags & OBJ_DEAD) == 0)
1067 	) {
1068 		struct vnode *vp = (struct vnode *)object->handle;
1069 
1070 		if (vp && VSHOULDFREE(vp))
1071 			vfree(vp);
1072 	}
1073 
1074 	/*
1075 	 * Clear the UNMANAGED flag when freeing an unmanaged page.
1076 	 */
1077 
1078 	if (m->flags & PG_UNMANAGED) {
1079 	    m->flags &= ~PG_UNMANAGED;
1080 	} else {
1081 #ifdef __alpha__
1082 	    pmap_page_is_free(m);
1083 #endif
1084 	}
1085 
1086 	if (m->hold_count != 0) {
1087 		m->flags &= ~PG_ZERO;
1088 		m->queue = PQ_HOLD;
1089 	} else
1090 		m->queue = PQ_FREE + m->pc;
1091 	pq = &vm_page_queues[m->queue];
1092 	pq->lcnt++;
1093 	++(*pq->cnt);
1094 
1095 	/*
1096 	 * Put zero'd pages on the end ( where we look for zero'd pages
1097 	 * first ) and non-zerod pages at the head.
1098 	 */
1099 
1100 	if (m->flags & PG_ZERO) {
1101 		TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
1102 		++vm_page_zero_count;
1103 	} else {
1104 		TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
1105 	}
1106 
1107 	vm_page_free_wakeup();
1108 
1109 	splx(s);
1110 }
1111 
1112 /*
1113  *	vm_page_unmanage:
1114  *
1115  * 	Prevent PV management from being done on the page.  The page is
1116  *	removed from the paging queues as if it were wired, and as a
1117  *	consequence of no longer being managed the pageout daemon will not
1118  *	touch it (since there is no way to locate the pte mappings for the
1119  *	page).  madvise() calls that mess with the pmap will also no longer
1120  *	operate on the page.
1121  *
1122  *	Beyond that the page is still reasonably 'normal'.  Freeing the page
1123  *	will clear the flag.
1124  *
1125  *	This routine is used by OBJT_PHYS objects - objects using unswappable
1126  *	physical memory as backing store rather then swap-backed memory and
1127  *	will eventually be extended to support 4MB unmanaged physical
1128  *	mappings.
1129  */
1130 
1131 void
1132 vm_page_unmanage(vm_page_t m)
1133 {
1134 	int s;
1135 
1136 	s = splvm();
1137 	if ((m->flags & PG_UNMANAGED) == 0) {
1138 		if (m->wire_count == 0)
1139 			vm_page_unqueue(m);
1140 	}
1141 	vm_page_flag_set(m, PG_UNMANAGED);
1142 	splx(s);
1143 }
1144 
1145 /*
1146  *	vm_page_wire:
1147  *
1148  *	Mark this page as wired down by yet
1149  *	another map, removing it from paging queues
1150  *	as necessary.
1151  *
1152  *	The page queues must be locked.
1153  *	This routine may not block.
1154  */
1155 void
1156 vm_page_wire(vm_page_t m)
1157 {
1158 	int s;
1159 
1160 	/*
1161 	 * Only bump the wire statistics if the page is not already wired,
1162 	 * and only unqueue the page if it is on some queue (if it is unmanaged
1163 	 * it is already off the queues).
1164 	 */
1165 	s = splvm();
1166 	if (m->wire_count == 0) {
1167 		if ((m->flags & PG_UNMANAGED) == 0)
1168 			vm_page_unqueue(m);
1169 		vmstats.v_wire_count++;
1170 	}
1171 	m->wire_count++;
1172 	KASSERT(m->wire_count != 0,
1173 	    ("vm_page_wire: wire_count overflow m=%p", m));
1174 
1175 	splx(s);
1176 	vm_page_flag_set(m, PG_MAPPED);
1177 }
1178 
1179 /*
1180  *	vm_page_unwire:
1181  *
1182  *	Release one wiring of this page, potentially
1183  *	enabling it to be paged again.
1184  *
1185  *	Many pages placed on the inactive queue should actually go
1186  *	into the cache, but it is difficult to figure out which.  What
1187  *	we do instead, if the inactive target is well met, is to put
1188  *	clean pages at the head of the inactive queue instead of the tail.
1189  *	This will cause them to be moved to the cache more quickly and
1190  *	if not actively re-referenced, freed more quickly.  If we just
1191  *	stick these pages at the end of the inactive queue, heavy filesystem
1192  *	meta-data accesses can cause an unnecessary paging load on memory bound
1193  *	processes.  This optimization causes one-time-use metadata to be
1194  *	reused more quickly.
1195  *
1196  *	BUT, if we are in a low-memory situation we have no choice but to
1197  *	put clean pages on the cache queue.
1198  *
1199  *	A number of routines use vm_page_unwire() to guarantee that the page
1200  *	will go into either the inactive or active queues, and will NEVER
1201  *	be placed in the cache - for example, just after dirtying a page.
1202  *	dirty pages in the cache are not allowed.
1203  *
1204  *	The page queues must be locked.
1205  *	This routine may not block.
1206  */
1207 void
1208 vm_page_unwire(vm_page_t m, int activate)
1209 {
1210 	int s;
1211 
1212 	s = splvm();
1213 
1214 	if (m->wire_count > 0) {
1215 		m->wire_count--;
1216 		if (m->wire_count == 0) {
1217 			vmstats.v_wire_count--;
1218 			if (m->flags & PG_UNMANAGED) {
1219 				;
1220 			} else if (activate) {
1221 				TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1222 				m->queue = PQ_ACTIVE;
1223 				vm_page_queues[PQ_ACTIVE].lcnt++;
1224 				vmstats.v_active_count++;
1225 			} else {
1226 				vm_page_flag_clear(m, PG_WINATCFLS);
1227 				TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1228 				m->queue = PQ_INACTIVE;
1229 				vm_page_queues[PQ_INACTIVE].lcnt++;
1230 				vmstats.v_inactive_count++;
1231 			}
1232 		}
1233 	} else {
1234 		panic("vm_page_unwire: invalid wire count: %d\n", m->wire_count);
1235 	}
1236 	splx(s);
1237 }
1238 
1239 
1240 /*
1241  * Move the specified page to the inactive queue.  If the page has
1242  * any associated swap, the swap is deallocated.
1243  *
1244  * Normally athead is 0 resulting in LRU operation.  athead is set
1245  * to 1 if we want this page to be 'as if it were placed in the cache',
1246  * except without unmapping it from the process address space.
1247  *
1248  * This routine may not block.
1249  */
1250 static __inline void
1251 _vm_page_deactivate(vm_page_t m, int athead)
1252 {
1253 	int s;
1254 
1255 	/*
1256 	 * Ignore if already inactive.
1257 	 */
1258 	if (m->queue == PQ_INACTIVE)
1259 		return;
1260 
1261 	s = splvm();
1262 	if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1263 		if ((m->queue - m->pc) == PQ_CACHE)
1264 			mycpu->gd_cnt.v_reactivated++;
1265 		vm_page_flag_clear(m, PG_WINATCFLS);
1266 		vm_page_unqueue(m);
1267 		if (athead)
1268 			TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1269 		else
1270 			TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1271 		m->queue = PQ_INACTIVE;
1272 		vm_page_queues[PQ_INACTIVE].lcnt++;
1273 		vmstats.v_inactive_count++;
1274 	}
1275 	splx(s);
1276 }
1277 
1278 void
1279 vm_page_deactivate(vm_page_t m)
1280 {
1281     _vm_page_deactivate(m, 0);
1282 }
1283 
1284 /*
1285  * vm_page_try_to_cache:
1286  *
1287  * Returns 0 on failure, 1 on success
1288  */
1289 int
1290 vm_page_try_to_cache(vm_page_t m)
1291 {
1292 	if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1293 	    (m->flags & (PG_BUSY|PG_UNMANAGED))) {
1294 		return(0);
1295 	}
1296 	vm_page_test_dirty(m);
1297 	if (m->dirty)
1298 		return(0);
1299 	vm_page_cache(m);
1300 	return(1);
1301 }
1302 
1303 /*
1304  * vm_page_try_to_free()
1305  *
1306  *	Attempt to free the page.  If we cannot free it, we do nothing.
1307  *	1 is returned on success, 0 on failure.
1308  */
1309 
1310 int
1311 vm_page_try_to_free(vm_page_t m)
1312 {
1313 	if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1314 	    (m->flags & (PG_BUSY|PG_UNMANAGED))) {
1315 		return(0);
1316 	}
1317 	vm_page_test_dirty(m);
1318 	if (m->dirty)
1319 		return(0);
1320 	vm_page_busy(m);
1321 	vm_page_protect(m, VM_PROT_NONE);
1322 	vm_page_free(m);
1323 	return(1);
1324 }
1325 
1326 
1327 /*
1328  * vm_page_cache
1329  *
1330  * Put the specified page onto the page cache queue (if appropriate).
1331  *
1332  * This routine may not block.
1333  */
1334 void
1335 vm_page_cache(vm_page_t m)
1336 {
1337 	int s;
1338 
1339 	if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || m->wire_count) {
1340 		printf("vm_page_cache: attempting to cache busy page\n");
1341 		return;
1342 	}
1343 	if ((m->queue - m->pc) == PQ_CACHE)
1344 		return;
1345 
1346 	/*
1347 	 * Remove all pmaps and indicate that the page is not
1348 	 * writeable or mapped.
1349 	 */
1350 
1351 	vm_page_protect(m, VM_PROT_NONE);
1352 	if (m->dirty != 0) {
1353 		panic("vm_page_cache: caching a dirty page, pindex: %ld",
1354 			(long)m->pindex);
1355 	}
1356 	s = splvm();
1357 	vm_page_unqueue_nowakeup(m);
1358 	m->queue = PQ_CACHE + m->pc;
1359 	vm_page_queues[m->queue].lcnt++;
1360 	TAILQ_INSERT_TAIL(&vm_page_queues[m->queue].pl, m, pageq);
1361 	vmstats.v_cache_count++;
1362 	vm_page_free_wakeup();
1363 	splx(s);
1364 }
1365 
1366 /*
1367  * vm_page_dontneed
1368  *
1369  *	Cache, deactivate, or do nothing as appropriate.  This routine
1370  *	is typically used by madvise() MADV_DONTNEED.
1371  *
1372  *	Generally speaking we want to move the page into the cache so
1373  *	it gets reused quickly.  However, this can result in a silly syndrome
1374  *	due to the page recycling too quickly.  Small objects will not be
1375  *	fully cached.  On the otherhand, if we move the page to the inactive
1376  *	queue we wind up with a problem whereby very large objects
1377  *	unnecessarily blow away our inactive and cache queues.
1378  *
1379  *	The solution is to move the pages based on a fixed weighting.  We
1380  *	either leave them alone, deactivate them, or move them to the cache,
1381  *	where moving them to the cache has the highest weighting.
1382  *	By forcing some pages into other queues we eventually force the
1383  *	system to balance the queues, potentially recovering other unrelated
1384  *	space from active.  The idea is to not force this to happen too
1385  *	often.
1386  */
1387 
1388 void
1389 vm_page_dontneed(vm_page_t m)
1390 {
1391 	static int dnweight;
1392 	int dnw;
1393 	int head;
1394 
1395 	dnw = ++dnweight;
1396 
1397 	/*
1398 	 * occassionally leave the page alone
1399 	 */
1400 
1401 	if ((dnw & 0x01F0) == 0 ||
1402 	    m->queue == PQ_INACTIVE ||
1403 	    m->queue - m->pc == PQ_CACHE
1404 	) {
1405 		if (m->act_count >= ACT_INIT)
1406 			--m->act_count;
1407 		return;
1408 	}
1409 
1410 	if (m->dirty == 0)
1411 		vm_page_test_dirty(m);
1412 
1413 	if (m->dirty || (dnw & 0x0070) == 0) {
1414 		/*
1415 		 * Deactivate the page 3 times out of 32.
1416 		 */
1417 		head = 0;
1418 	} else {
1419 		/*
1420 		 * Cache the page 28 times out of every 32.  Note that
1421 		 * the page is deactivated instead of cached, but placed
1422 		 * at the head of the queue instead of the tail.
1423 		 */
1424 		head = 1;
1425 	}
1426 	_vm_page_deactivate(m, head);
1427 }
1428 
1429 /*
1430  * Grab a page, waiting until we are waken up due to the page
1431  * changing state.  We keep on waiting, if the page continues
1432  * to be in the object.  If the page doesn't exist, allocate it.
1433  *
1434  * If VM_ALLOC_RETRY is specified VM_ALLOC_NORMAL must also be specified.
1435  *
1436  * This routine may block.
1437  */
1438 vm_page_t
1439 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
1440 {
1441 	vm_page_t m;
1442 	int s, generation;
1443 
1444 	KKASSERT(allocflags &
1445 		(VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
1446 retrylookup:
1447 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
1448 		if (m->busy || (m->flags & PG_BUSY)) {
1449 			generation = object->generation;
1450 
1451 			s = splvm();
1452 			while ((object->generation == generation) &&
1453 					(m->busy || (m->flags & PG_BUSY))) {
1454 				vm_page_flag_set(m, PG_WANTED | PG_REFERENCED);
1455 				tsleep(m, 0, "pgrbwt", 0);
1456 				if ((allocflags & VM_ALLOC_RETRY) == 0) {
1457 					splx(s);
1458 					return NULL;
1459 				}
1460 			}
1461 			splx(s);
1462 			goto retrylookup;
1463 		} else {
1464 			vm_page_busy(m);
1465 			return m;
1466 		}
1467 	}
1468 
1469 	m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY);
1470 	if (m == NULL) {
1471 		VM_WAIT;
1472 		if ((allocflags & VM_ALLOC_RETRY) == 0)
1473 			return NULL;
1474 		goto retrylookup;
1475 	}
1476 
1477 	return m;
1478 }
1479 
1480 /*
1481  * Mapping function for valid bits or for dirty bits in
1482  * a page.  May not block.
1483  *
1484  * Inputs are required to range within a page.
1485  */
1486 
1487 __inline int
1488 vm_page_bits(int base, int size)
1489 {
1490 	int first_bit;
1491 	int last_bit;
1492 
1493 	KASSERT(
1494 	    base + size <= PAGE_SIZE,
1495 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
1496 	);
1497 
1498 	if (size == 0)		/* handle degenerate case */
1499 		return(0);
1500 
1501 	first_bit = base >> DEV_BSHIFT;
1502 	last_bit = (base + size - 1) >> DEV_BSHIFT;
1503 
1504 	return ((2 << last_bit) - (1 << first_bit));
1505 }
1506 
1507 /*
1508  *	vm_page_set_validclean:
1509  *
1510  *	Sets portions of a page valid and clean.  The arguments are expected
1511  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
1512  *	of any partial chunks touched by the range.  The invalid portion of
1513  *	such chunks will be zero'd.
1514  *
1515  *	This routine may not block.
1516  *
1517  *	(base + size) must be less then or equal to PAGE_SIZE.
1518  */
1519 void
1520 vm_page_set_validclean(vm_page_t m, int base, int size)
1521 {
1522 	int pagebits;
1523 	int frag;
1524 	int endoff;
1525 
1526 	if (size == 0)	/* handle degenerate case */
1527 		return;
1528 
1529 	/*
1530 	 * If the base is not DEV_BSIZE aligned and the valid
1531 	 * bit is clear, we have to zero out a portion of the
1532 	 * first block.
1533 	 */
1534 
1535 	if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
1536 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0
1537 	) {
1538 		pmap_zero_page_area(
1539 		    VM_PAGE_TO_PHYS(m),
1540 		    frag,
1541 		    base - frag
1542 		);
1543 	}
1544 
1545 	/*
1546 	 * If the ending offset is not DEV_BSIZE aligned and the
1547 	 * valid bit is clear, we have to zero out a portion of
1548 	 * the last block.
1549 	 */
1550 
1551 	endoff = base + size;
1552 
1553 	if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
1554 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0
1555 	) {
1556 		pmap_zero_page_area(
1557 		    VM_PAGE_TO_PHYS(m),
1558 		    endoff,
1559 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1))
1560 		);
1561 	}
1562 
1563 	/*
1564 	 * Set valid, clear dirty bits.  If validating the entire
1565 	 * page we can safely clear the pmap modify bit.  We also
1566 	 * use this opportunity to clear the PG_NOSYNC flag.  If a process
1567 	 * takes a write fault on a MAP_NOSYNC memory area the flag will
1568 	 * be set again.
1569 	 *
1570 	 * We set valid bits inclusive of any overlap, but we can only
1571 	 * clear dirty bits for DEV_BSIZE chunks that are fully within
1572 	 * the range.
1573 	 */
1574 
1575 	pagebits = vm_page_bits(base, size);
1576 	m->valid |= pagebits;
1577 #if 0	/* NOT YET */
1578 	if ((frag = base & (DEV_BSIZE - 1)) != 0) {
1579 		frag = DEV_BSIZE - frag;
1580 		base += frag;
1581 		size -= frag;
1582 		if (size < 0)
1583 		    size = 0;
1584 	}
1585 	pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
1586 #endif
1587 	m->dirty &= ~pagebits;
1588 	if (base == 0 && size == PAGE_SIZE) {
1589 		pmap_clear_modify(m);
1590 		vm_page_flag_clear(m, PG_NOSYNC);
1591 	}
1592 }
1593 
1594 #if 0
1595 
1596 void
1597 vm_page_set_dirty(vm_page_t m, int base, int size)
1598 {
1599 	m->dirty |= vm_page_bits(base, size);
1600 }
1601 
1602 #endif
1603 
1604 void
1605 vm_page_clear_dirty(vm_page_t m, int base, int size)
1606 {
1607 	m->dirty &= ~vm_page_bits(base, size);
1608 }
1609 
1610 /*
1611  *	vm_page_set_invalid:
1612  *
1613  *	Invalidates DEV_BSIZE'd chunks within a page.  Both the
1614  *	valid and dirty bits for the effected areas are cleared.
1615  *
1616  *	May not block.
1617  */
1618 void
1619 vm_page_set_invalid(vm_page_t m, int base, int size)
1620 {
1621 	int bits;
1622 
1623 	bits = vm_page_bits(base, size);
1624 	m->valid &= ~bits;
1625 	m->dirty &= ~bits;
1626 	m->object->generation++;
1627 }
1628 
1629 /*
1630  * vm_page_zero_invalid()
1631  *
1632  *	The kernel assumes that the invalid portions of a page contain
1633  *	garbage, but such pages can be mapped into memory by user code.
1634  *	When this occurs, we must zero out the non-valid portions of the
1635  *	page so user code sees what it expects.
1636  *
1637  *	Pages are most often semi-valid when the end of a file is mapped
1638  *	into memory and the file's size is not page aligned.
1639  */
1640 
1641 void
1642 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
1643 {
1644 	int b;
1645 	int i;
1646 
1647 	/*
1648 	 * Scan the valid bits looking for invalid sections that
1649 	 * must be zerod.  Invalid sub-DEV_BSIZE'd areas ( where the
1650 	 * valid bit may be set ) have already been zerod by
1651 	 * vm_page_set_validclean().
1652 	 */
1653 
1654 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
1655 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
1656 		    (m->valid & (1 << i))
1657 		) {
1658 			if (i > b) {
1659 				pmap_zero_page_area(
1660 				    VM_PAGE_TO_PHYS(m),
1661 				    b << DEV_BSHIFT,
1662 				    (i - b) << DEV_BSHIFT
1663 				);
1664 			}
1665 			b = i + 1;
1666 		}
1667 	}
1668 
1669 	/*
1670 	 * setvalid is TRUE when we can safely set the zero'd areas
1671 	 * as being valid.  We can do this if there are no cache consistency
1672 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
1673 	 */
1674 
1675 	if (setvalid)
1676 		m->valid = VM_PAGE_BITS_ALL;
1677 }
1678 
1679 /*
1680  *	vm_page_is_valid:
1681  *
1682  *	Is (partial) page valid?  Note that the case where size == 0
1683  *	will return FALSE in the degenerate case where the page is
1684  *	entirely invalid, and TRUE otherwise.
1685  *
1686  *	May not block.
1687  */
1688 
1689 int
1690 vm_page_is_valid(vm_page_t m, int base, int size)
1691 {
1692 	int bits = vm_page_bits(base, size);
1693 
1694 	if (m->valid && ((m->valid & bits) == bits))
1695 		return 1;
1696 	else
1697 		return 0;
1698 }
1699 
1700 /*
1701  * update dirty bits from pmap/mmu.  May not block.
1702  */
1703 
1704 void
1705 vm_page_test_dirty(vm_page_t m)
1706 {
1707 	if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) {
1708 		vm_page_dirty(m);
1709 	}
1710 }
1711 
1712 #include "opt_ddb.h"
1713 #ifdef DDB
1714 #include <sys/kernel.h>
1715 
1716 #include <ddb/ddb.h>
1717 
1718 DB_SHOW_COMMAND(page, vm_page_print_page_info)
1719 {
1720 	db_printf("vmstats.v_free_count: %d\n", vmstats.v_free_count);
1721 	db_printf("vmstats.v_cache_count: %d\n", vmstats.v_cache_count);
1722 	db_printf("vmstats.v_inactive_count: %d\n", vmstats.v_inactive_count);
1723 	db_printf("vmstats.v_active_count: %d\n", vmstats.v_active_count);
1724 	db_printf("vmstats.v_wire_count: %d\n", vmstats.v_wire_count);
1725 	db_printf("vmstats.v_free_reserved: %d\n", vmstats.v_free_reserved);
1726 	db_printf("vmstats.v_free_min: %d\n", vmstats.v_free_min);
1727 	db_printf("vmstats.v_free_target: %d\n", vmstats.v_free_target);
1728 	db_printf("vmstats.v_cache_min: %d\n", vmstats.v_cache_min);
1729 	db_printf("vmstats.v_inactive_target: %d\n", vmstats.v_inactive_target);
1730 }
1731 
1732 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
1733 {
1734 	int i;
1735 	db_printf("PQ_FREE:");
1736 	for(i=0;i<PQ_L2_SIZE;i++) {
1737 		db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt);
1738 	}
1739 	db_printf("\n");
1740 
1741 	db_printf("PQ_CACHE:");
1742 	for(i=0;i<PQ_L2_SIZE;i++) {
1743 		db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt);
1744 	}
1745 	db_printf("\n");
1746 
1747 	db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n",
1748 		vm_page_queues[PQ_ACTIVE].lcnt,
1749 		vm_page_queues[PQ_INACTIVE].lcnt);
1750 }
1751 #endif /* DDB */
1752