xref: /freebsd/sys/vm/vm_page.c (revision 4cf49a43559ed9fdad601bdcccd2c55963008675)
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  * $FreeBSD$
38  */
39 
40 /*
41  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
42  * All rights reserved.
43  *
44  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
45  *
46  * Permission to use, copy, modify and distribute this software and
47  * its documentation is hereby granted, provided that both the copyright
48  * notice and this permission notice appear in all copies of the
49  * software, derivative works or modified versions, and any portions
50  * thereof, and that both notices appear in supporting documentation.
51  *
52  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
53  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
54  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
55  *
56  * Carnegie Mellon requests users of this software to return to
57  *
58  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
59  *  School of Computer Science
60  *  Carnegie Mellon University
61  *  Pittsburgh PA 15213-3890
62  *
63  * any improvements or extensions that they make and grant Carnegie the
64  * rights to redistribute these changes.
65  */
66 
67 /*
68  *	Resident memory management module.
69  */
70 
71 #include <sys/param.h>
72 #include <sys/systm.h>
73 #include <sys/malloc.h>
74 #include <sys/proc.h>
75 #include <sys/vmmeter.h>
76 #include <sys/vnode.h>
77 
78 #include <vm/vm.h>
79 #include <vm/vm_param.h>
80 #include <vm/vm_prot.h>
81 #include <sys/lock.h>
82 #include <vm/vm_kern.h>
83 #include <vm/vm_object.h>
84 #include <vm/vm_page.h>
85 #include <vm/vm_pageout.h>
86 #include <vm/vm_pager.h>
87 #include <vm/vm_extern.h>
88 
89 static void	vm_page_queue_init __P((void));
90 static vm_page_t vm_page_select_cache __P((vm_object_t, vm_pindex_t));
91 
92 /*
93  *	Associated with page of user-allocatable memory is a
94  *	page structure.
95  */
96 
97 static struct vm_page **vm_page_buckets; /* Array of buckets */
98 static int vm_page_bucket_count;	/* How big is array? */
99 static int vm_page_hash_mask;		/* Mask for hash function */
100 static volatile int vm_page_bucket_generation;
101 
102 struct pglist vm_page_queue_free[PQ_L2_SIZE] = {{0}};
103 struct pglist vm_page_queue_active = {0};
104 struct pglist vm_page_queue_inactive = {0};
105 struct pglist vm_page_queue_cache[PQ_L2_SIZE] = {{0}};
106 
107 struct vpgqueues vm_page_queues[PQ_COUNT] = {{0}};
108 
109 static void
110 vm_page_queue_init(void) {
111 	int i;
112 
113 	for(i=0;i<PQ_L2_SIZE;i++) {
114 		vm_page_queues[PQ_FREE+i].pl = &vm_page_queue_free[i];
115 		vm_page_queues[PQ_FREE+i].cnt = &cnt.v_free_count;
116 	}
117 	vm_page_queues[PQ_INACTIVE].pl = &vm_page_queue_inactive;
118 	vm_page_queues[PQ_INACTIVE].cnt = &cnt.v_inactive_count;
119 
120 	vm_page_queues[PQ_ACTIVE].pl = &vm_page_queue_active;
121 	vm_page_queues[PQ_ACTIVE].cnt = &cnt.v_active_count;
122 	for(i=0;i<PQ_L2_SIZE;i++) {
123 		vm_page_queues[PQ_CACHE+i].pl = &vm_page_queue_cache[i];
124 		vm_page_queues[PQ_CACHE+i].cnt = &cnt.v_cache_count;
125 	}
126 	for(i=PQ_FREE;i<PQ_COUNT;i++) {
127 		if (vm_page_queues[i].pl) {
128 			TAILQ_INIT(vm_page_queues[i].pl);
129 		} else {
130 			panic("vm_page_queue_init: queue %d is null", i);
131 		}
132 	}
133 }
134 
135 vm_page_t vm_page_array = 0;
136 static int vm_page_array_size = 0;
137 long first_page = 0;
138 int vm_page_zero_count = 0;
139 
140 static __inline int vm_page_hash __P((vm_object_t object, vm_pindex_t pindex));
141 static void vm_page_free_wakeup __P((void));
142 
143 /*
144  *	vm_set_page_size:
145  *
146  *	Sets the page size, perhaps based upon the memory
147  *	size.  Must be called before any use of page-size
148  *	dependent functions.
149  */
150 void
151 vm_set_page_size()
152 {
153 	if (cnt.v_page_size == 0)
154 		cnt.v_page_size = PAGE_SIZE;
155 	if (((cnt.v_page_size - 1) & cnt.v_page_size) != 0)
156 		panic("vm_set_page_size: page size not a power of two");
157 }
158 
159 /*
160  *	vm_page_startup:
161  *
162  *	Initializes the resident memory module.
163  *
164  *	Allocates memory for the page cells, and
165  *	for the object/offset-to-page hash table headers.
166  *	Each page cell is initialized and placed on the free list.
167  */
168 
169 vm_offset_t
170 vm_page_startup(starta, enda, vaddr)
171 	register vm_offset_t starta;
172 	vm_offset_t enda;
173 	register vm_offset_t vaddr;
174 {
175 	register vm_offset_t mapped;
176 	register vm_page_t m;
177 	register struct vm_page **bucket;
178 	vm_size_t npages, page_range;
179 	register vm_offset_t new_start;
180 	int i;
181 	vm_offset_t pa;
182 	int nblocks;
183 	vm_offset_t first_managed_page;
184 
185 	/* the biggest memory array is the second group of pages */
186 	vm_offset_t start;
187 	vm_offset_t biggestone, biggestsize;
188 
189 	vm_offset_t total;
190 
191 	total = 0;
192 	biggestsize = 0;
193 	biggestone = 0;
194 	nblocks = 0;
195 	vaddr = round_page(vaddr);
196 
197 	for (i = 0; phys_avail[i + 1]; i += 2) {
198 		phys_avail[i] = round_page(phys_avail[i]);
199 		phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
200 	}
201 
202 	for (i = 0; phys_avail[i + 1]; i += 2) {
203 		int size = phys_avail[i + 1] - phys_avail[i];
204 
205 		if (size > biggestsize) {
206 			biggestone = i;
207 			biggestsize = size;
208 		}
209 		++nblocks;
210 		total += size;
211 	}
212 
213 	start = phys_avail[biggestone];
214 
215 	/*
216 	 * Initialize the queue headers for the free queue, the active queue
217 	 * and the inactive queue.
218 	 */
219 
220 	vm_page_queue_init();
221 
222 	/*
223 	 * Allocate (and initialize) the hash table buckets.
224 	 *
225 	 * The number of buckets MUST BE a power of 2, and the actual value is
226 	 * the next power of 2 greater than the number of physical pages in
227 	 * the system.
228 	 *
229 	 * We make the hash table approximately 2x the number of pages to
230 	 * reduce the chain length.  This is about the same size using the
231 	 * singly-linked list as the 1x hash table we were using before
232 	 * using TAILQ but the chain length will be smaller.
233 	 *
234 	 * Note: This computation can be tweaked if desired.
235 	 */
236 	vm_page_buckets = (struct vm_page **)vaddr;
237 	bucket = vm_page_buckets;
238 	if (vm_page_bucket_count == 0) {
239 		vm_page_bucket_count = 1;
240 		while (vm_page_bucket_count < atop(total))
241 			vm_page_bucket_count <<= 1;
242 	}
243 	vm_page_bucket_count <<= 1;
244 	vm_page_hash_mask = vm_page_bucket_count - 1;
245 
246 	/*
247 	 * Validate these addresses.
248 	 */
249 
250 	new_start = start + vm_page_bucket_count * sizeof(struct vm_page *);
251 	new_start = round_page(new_start);
252 	mapped = round_page(vaddr);
253 	vaddr = pmap_map(mapped, start, new_start,
254 	    VM_PROT_READ | VM_PROT_WRITE);
255 	start = new_start;
256 	vaddr = round_page(vaddr);
257 	bzero((caddr_t) mapped, vaddr - mapped);
258 
259 	for (i = 0; i < vm_page_bucket_count; i++) {
260 		*bucket = NULL;
261 		bucket++;
262 	}
263 
264 	/*
265 	 * Compute the number of pages of memory that will be available for
266 	 * use (taking into account the overhead of a page structure per
267 	 * page).
268 	 */
269 
270 	first_page = phys_avail[0] / PAGE_SIZE;
271 
272 	page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page;
273 	npages = (total - (page_range * sizeof(struct vm_page)) -
274 	    (start - phys_avail[biggestone])) / PAGE_SIZE;
275 
276 	/*
277 	 * Initialize the mem entry structures now, and put them in the free
278 	 * queue.
279 	 */
280 	vm_page_array = (vm_page_t) vaddr;
281 	mapped = vaddr;
282 
283 	/*
284 	 * Validate these addresses.
285 	 */
286 	new_start = round_page(start + page_range * sizeof(struct vm_page));
287 	mapped = pmap_map(mapped, start, new_start,
288 	    VM_PROT_READ | VM_PROT_WRITE);
289 	start = new_start;
290 
291 	first_managed_page = start / PAGE_SIZE;
292 
293 	/*
294 	 * Clear all of the page structures
295 	 */
296 	bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
297 	vm_page_array_size = page_range;
298 
299 	/*
300 	 * Construct the free queue(s) in descending order (by physical
301 	 * address) so that the first 16MB of physical memory is allocated
302 	 * last rather than first.  On large-memory machines, this avoids
303 	 * the exhaustion of low physical memory before isa_dmainit has run.
304 	 */
305 	cnt.v_page_count = 0;
306 	cnt.v_free_count = 0;
307 	for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) {
308 		if (i == biggestone)
309 			pa = ptoa(first_managed_page);
310 		else
311 			pa = phys_avail[i];
312 		while (pa < phys_avail[i + 1] && npages-- > 0) {
313 			++cnt.v_page_count;
314 			++cnt.v_free_count;
315 			m = PHYS_TO_VM_PAGE(pa);
316 			m->phys_addr = pa;
317 			m->flags = 0;
318 			m->pc = (pa >> PAGE_SHIFT) & PQ_L2_MASK;
319 			m->queue = m->pc + PQ_FREE;
320 			TAILQ_INSERT_HEAD(vm_page_queues[m->queue].pl, m, pageq);
321 			vm_page_queues[m->queue].lcnt++;
322 			pa += PAGE_SIZE;
323 		}
324 	}
325 	return (mapped);
326 }
327 
328 /*
329  *	vm_page_hash:
330  *
331  *	Distributes the object/offset key pair among hash buckets.
332  *
333  *	NOTE:  This macro depends on vm_page_bucket_count being a power of 2.
334  *	This routine may not block.
335  *
336  *	We try to randomize the hash based on the object to spread the pages
337  *	out in the hash table without it costing us too much.
338  */
339 static __inline int
340 vm_page_hash(object, pindex)
341 	vm_object_t object;
342 	vm_pindex_t pindex;
343 {
344 	int i = ((uintptr_t)object + pindex) ^ object->hash_rand;
345 
346 	return(i & vm_page_hash_mask);
347 }
348 
349 /*
350  *	vm_page_insert:		[ internal use only ]
351  *
352  *	Inserts the given mem entry into the object and object list.
353  *
354  *	The pagetables are not updated but will presumably fault the page
355  *	in if necessary, or if a kernel page the caller will at some point
356  *	enter the page into the kernel's pmap.  We are not allowed to block
357  *	here so we *can't* do this anyway.
358  *
359  *	The object and page must be locked, and must be splhigh.
360  *	This routine may not block.
361  */
362 
363 void
364 vm_page_insert(m, object, pindex)
365 	register vm_page_t m;
366 	register vm_object_t object;
367 	register vm_pindex_t pindex;
368 {
369 	register struct vm_page **bucket;
370 
371 	if (m->object != NULL)
372 		panic("vm_page_insert: already inserted");
373 
374 	/*
375 	 * Record the object/offset pair in this page
376 	 */
377 
378 	m->object = object;
379 	m->pindex = pindex;
380 
381 	/*
382 	 * Insert it into the object_object/offset hash table
383 	 */
384 
385 	bucket = &vm_page_buckets[vm_page_hash(object, pindex)];
386 	m->hnext = *bucket;
387 	*bucket = m;
388 	vm_page_bucket_generation++;
389 
390 	/*
391 	 * Now link into the object's list of backed pages.
392 	 */
393 
394 	TAILQ_INSERT_TAIL(&object->memq, m, listq);
395 	object->generation++;
396 
397 	/*
398 	 * show that the object has one more resident page.
399 	 */
400 
401 	object->resident_page_count++;
402 
403 	/*
404 	 * Since we are inserting a new and possibly dirty page,
405 	 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags.
406 	 */
407 	if (m->flags & PG_WRITEABLE)
408 	    vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
409 }
410 
411 /*
412  *	vm_page_remove:
413  *				NOTE: used by device pager as well -wfj
414  *
415  *	Removes the given mem entry from the object/offset-page
416  *	table and the object page list, but do not invalidate/terminate
417  *	the backing store.
418  *
419  *	The object and page must be locked, and at splhigh.
420  *	The underlying pmap entry (if any) is NOT removed here.
421  *	This routine may not block.
422  */
423 
424 void
425 vm_page_remove(m)
426 	vm_page_t m;
427 {
428 	vm_object_t object;
429 
430 	if (m->object == NULL)
431 		return;
432 
433 #if !defined(MAX_PERF)
434 	if ((m->flags & PG_BUSY) == 0) {
435 		panic("vm_page_remove: page not busy");
436 	}
437 #endif
438 
439 	/*
440 	 * Basically destroy the page.
441 	 */
442 
443 	vm_page_wakeup(m);
444 
445 	object = m->object;
446 
447 	/*
448 	 * Remove from the object_object/offset hash table.  The object
449 	 * must be on the hash queue, we will panic if it isn't
450 	 *
451 	 * Note: we must NULL-out m->hnext to prevent loops in detached
452 	 * buffers with vm_page_lookup().
453 	 */
454 
455 	{
456 		struct vm_page **bucket;
457 
458 		bucket = &vm_page_buckets[vm_page_hash(m->object, m->pindex)];
459 		while (*bucket != m) {
460 #if !defined(MAX_PERF)
461 			if (*bucket == NULL)
462 				panic("vm_page_remove(): page not found in hash");
463 #endif
464 			bucket = &(*bucket)->hnext;
465 		}
466 		*bucket = m->hnext;
467 		m->hnext = NULL;
468 		vm_page_bucket_generation++;
469 	}
470 
471 	/*
472 	 * Now remove from the object's list of backed pages.
473 	 */
474 
475 	TAILQ_REMOVE(&object->memq, m, listq);
476 
477 	/*
478 	 * And show that the object has one fewer resident page.
479 	 */
480 
481 	object->resident_page_count--;
482 	object->generation++;
483 
484 	m->object = NULL;
485 }
486 
487 /*
488  *	vm_page_lookup:
489  *
490  *	Returns the page associated with the object/offset
491  *	pair specified; if none is found, NULL is returned.
492  *
493  *	NOTE: the code below does not lock.  It will operate properly if
494  *	an interrupt makes a change, but the generation algorithm will not
495  *	operate properly in an SMP environment where both cpu's are able to run
496  *	kernel code simultaniously.
497  *
498  *	The object must be locked.  No side effects.
499  *	This routine may not block.
500  *	This is a critical path routine
501  */
502 
503 vm_page_t
504 vm_page_lookup(object, pindex)
505 	register vm_object_t object;
506 	register vm_pindex_t pindex;
507 {
508 	register vm_page_t m;
509 	register struct vm_page **bucket;
510 	int generation;
511 
512 	/*
513 	 * Search the hash table for this object/offset pair
514 	 */
515 
516 retry:
517 	generation = vm_page_bucket_generation;
518 	bucket = &vm_page_buckets[vm_page_hash(object, pindex)];
519 	for (m = *bucket; m != NULL; m = m->hnext) {
520 		if ((m->object == object) && (m->pindex == pindex)) {
521 			if (vm_page_bucket_generation != generation)
522 				goto retry;
523 			return (m);
524 		}
525 	}
526 	if (vm_page_bucket_generation != generation)
527 		goto retry;
528 	return (NULL);
529 }
530 
531 /*
532  *	vm_page_rename:
533  *
534  *	Move the given memory entry from its
535  *	current object to the specified target object/offset.
536  *
537  *	The object must be locked.
538  *	This routine may not block.
539  *
540  *	Note: this routine will raise itself to splvm(), the caller need not.
541  *
542  *	Note: swap associated with the page must be invalidated by the move.  We
543  *	      have to do this for several reasons:  (1) we aren't freeing the
544  *	      page, (2) we are dirtying the page, (3) the VM system is probably
545  *	      moving the page from object A to B, and will then later move
546  *	      the backing store from A to B and we can't have a conflict.
547  *
548  *	Note: we *always* dirty the page.  It is necessary both for the
549  *	      fact that we moved it, and because we may be invalidating
550  *	      swap.  If the page is on the cache, we have to deactivate it
551  *	      or vm_page_dirty() will panic.  Dirty pages are not allowed
552  *	      on the cache.
553  */
554 
555 void
556 vm_page_rename(m, new_object, new_pindex)
557 	register vm_page_t m;
558 	register vm_object_t new_object;
559 	vm_pindex_t new_pindex;
560 {
561 	int s;
562 
563 	s = splvm();
564 	vm_page_remove(m);
565 	vm_page_insert(m, new_object, new_pindex);
566 	if (m->queue - m->pc == PQ_CACHE)
567 		vm_page_deactivate(m);
568 	vm_page_dirty(m);
569 	splx(s);
570 }
571 
572 /*
573  * vm_page_unqueue_nowakeup:
574  *
575  * 	vm_page_unqueue() without any wakeup
576  *
577  *	This routine must be called at splhigh().
578  *	This routine may not block.
579  */
580 
581 void
582 vm_page_unqueue_nowakeup(m)
583 	vm_page_t m;
584 {
585 	int queue = m->queue;
586 	struct vpgqueues *pq;
587 	if (queue != PQ_NONE) {
588 		pq = &vm_page_queues[queue];
589 		m->queue = PQ_NONE;
590 		TAILQ_REMOVE(pq->pl, m, pageq);
591 		(*pq->cnt)--;
592 		pq->lcnt--;
593 	}
594 }
595 
596 /*
597  * vm_page_unqueue:
598  *
599  *	Remove a page from its queue.
600  *
601  *	This routine must be called at splhigh().
602  *	This routine may not block.
603  */
604 
605 void
606 vm_page_unqueue(m)
607 	vm_page_t m;
608 {
609 	int queue = m->queue;
610 	struct vpgqueues *pq;
611 	if (queue != PQ_NONE) {
612 		m->queue = PQ_NONE;
613 		pq = &vm_page_queues[queue];
614 		TAILQ_REMOVE(pq->pl, m, pageq);
615 		(*pq->cnt)--;
616 		pq->lcnt--;
617 		if ((queue - m->pc) == PQ_CACHE) {
618 			if (vm_paging_needed())
619 				pagedaemon_wakeup();
620 		}
621 	}
622 }
623 
624 #if PQ_L2_SIZE > 1
625 
626 /*
627  *	vm_page_list_find:
628  *
629  *	Find a page on the specified queue with color optimization.
630  *
631  *	The page coloring optimization attempts to locate a page
632  *	that does not overload other nearby pages in the object in
633  *	the cpu's L1 or L2 caches.  We need this optmization because
634  *	cpu caches tend to be physical caches, while object spaces tend
635  *	to be virtual.
636  *
637  *	This routine must be called at splvm().
638  *	This routine may not block.
639  *
640  *	This routine may only be called from the vm_page_list_find() macro
641  *	in vm_page.h
642  */
643 vm_page_t
644 _vm_page_list_find(basequeue, index)
645 	int basequeue, index;
646 {
647 	int i;
648 	vm_page_t m = NULL;
649 	struct vpgqueues *pq;
650 
651 	pq = &vm_page_queues[basequeue];
652 
653 	/*
654 	 * Note that for the first loop, index+i and index-i wind up at the
655 	 * same place.  Even though this is not totally optimal, we've already
656 	 * blown it by missing the cache case so we do not care.
657 	 */
658 
659 	for(i = PQ_L2_SIZE / 2; i > 0; --i) {
660 		if ((m = TAILQ_FIRST(pq[(index + i) & PQ_L2_MASK].pl)) != NULL)
661 			break;
662 
663 		if ((m = TAILQ_FIRST(pq[(index - i) & PQ_L2_MASK].pl)) != NULL)
664 			break;
665 	}
666 	return(m);
667 }
668 
669 #endif
670 
671 /*
672  *	vm_page_select_cache:
673  *
674  *	Find a page on the cache queue with color optimization.  As pages
675  *	might be found, but not applicable, they are deactivated.  This
676  *	keeps us from using potentially busy cached pages.
677  *
678  *	This routine must be called at splvm().
679  *	This routine may not block.
680  */
681 vm_page_t
682 vm_page_select_cache(object, pindex)
683 	vm_object_t object;
684 	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) || 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		normal process request
735  *	VM_ALLOC_SYSTEM		system *really* needs a page
736  *	VM_ALLOC_INTERRUPT	interrupt time request
737  *	VM_ALLOC_ZERO		zero 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(object, pindex, page_req)
749 	vm_object_t object;
750 	vm_pindex_t pindex;
751 	int page_req;
752 {
753 	register vm_page_t m = NULL;
754 	int s;
755 
756 	KASSERT(!vm_page_lookup(object, pindex),
757 		("vm_page_alloc: page already allocated"));
758 
759 	/*
760 	 * The pager is allowed to eat deeper into the free page list.
761 	 */
762 
763 	if ((curproc == pageproc) && (page_req != VM_ALLOC_INTERRUPT)) {
764 		page_req = VM_ALLOC_SYSTEM;
765 	};
766 
767 	s = splvm();
768 
769 loop:
770 	if (cnt.v_free_count > cnt.v_free_reserved) {
771 		/*
772 		 * Allocate from the free queue if there are plenty of pages
773 		 * in it.
774 		 */
775 		if (page_req == VM_ALLOC_ZERO)
776 			m = vm_page_select_free(object, pindex, TRUE);
777 		else
778 			m = vm_page_select_free(object, pindex, FALSE);
779 	} else if (
780 	    (page_req == VM_ALLOC_SYSTEM &&
781 	     cnt.v_cache_count == 0 &&
782 	     cnt.v_free_count > cnt.v_interrupt_free_min) ||
783 	    (page_req == VM_ALLOC_INTERRUPT && cnt.v_free_count > 0)
784 	) {
785 		/*
786 		 * Interrupt or system, dig deeper into the free list.
787 		 */
788 		m = vm_page_select_free(object, pindex, FALSE);
789 	} else if (page_req != VM_ALLOC_INTERRUPT) {
790 		/*
791 		 * Allocateable from cache (non-interrupt only).  On success,
792 		 * we must free the page and try again, thus ensuring that
793 		 * cnt.v_*_free_min counters are replenished.
794 		 */
795 		m = vm_page_select_cache(object, pindex);
796 		if (m == NULL) {
797 			splx(s);
798 #if defined(DIAGNOSTIC)
799 			if (cnt.v_cache_count > 0)
800 				printf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", cnt.v_cache_count);
801 #endif
802 			vm_pageout_deficit++;
803 			pagedaemon_wakeup();
804 			return (NULL);
805 		}
806 		KASSERT(m->dirty == 0, ("Found dirty cache page %p", m));
807 		vm_page_busy(m);
808 		vm_page_protect(m, VM_PROT_NONE);
809 		vm_page_free(m);
810 		goto loop;
811 	} else {
812 		/*
813 		 * Not allocateable from cache from interrupt, give up.
814 		 */
815 		splx(s);
816 		vm_pageout_deficit++;
817 		pagedaemon_wakeup();
818 		return (NULL);
819 	}
820 
821 	/*
822 	 *  At this point we had better have found a good page.
823 	 */
824 
825 	KASSERT(
826 	    m != NULL,
827 	    ("vm_page_alloc(): missing page on free queue\n")
828 	);
829 
830 	/*
831 	 * Remove from free queue
832 	 */
833 
834 	{
835 		struct vpgqueues *pq = &vm_page_queues[m->queue];
836 
837 		TAILQ_REMOVE(pq->pl, m, pageq);
838 		(*pq->cnt)--;
839 		pq->lcnt--;
840 	}
841 
842 	/*
843 	 * Initialize structure.  Only the PG_ZERO flag is inherited.
844 	 */
845 
846 	if (m->flags & PG_ZERO) {
847 		vm_page_zero_count--;
848 		m->flags = PG_ZERO | PG_BUSY;
849 	} else {
850 		m->flags = PG_BUSY;
851 	}
852 	m->wire_count = 0;
853 	m->hold_count = 0;
854 	m->act_count = 0;
855 	m->busy = 0;
856 	m->valid = 0;
857 	KASSERT(m->dirty == 0, ("vm_page_alloc: free/cache page %p was dirty", m));
858 	m->queue = PQ_NONE;
859 
860 	/*
861 	 * vm_page_insert() is safe prior to the splx().  Note also that
862 	 * inserting a page here does not insert it into the pmap (which
863 	 * could cause us to block allocating memory).  We cannot block
864 	 * anywhere.
865 	 */
866 
867 	vm_page_insert(m, object, pindex);
868 
869 	/*
870 	 * Don't wakeup too often - wakeup the pageout daemon when
871 	 * we would be nearly out of memory.
872 	 */
873 	if (vm_paging_needed() || cnt.v_free_count < cnt.v_pageout_free_min)
874 		pagedaemon_wakeup();
875 
876 	splx(s);
877 
878 	return (m);
879 }
880 
881 /*
882  *	vm_wait:	(also see VM_WAIT macro)
883  *
884  *	Block until free pages are available for allocation
885  */
886 
887 void
888 vm_wait()
889 {
890 	int s;
891 
892 	s = splvm();
893 	if (curproc == pageproc) {
894 		vm_pageout_pages_needed = 1;
895 		tsleep(&vm_pageout_pages_needed, PSWP, "vmwait", 0);
896 	} else {
897 		if (!vm_pages_needed) {
898 			vm_pages_needed++;
899 			wakeup(&vm_pages_needed);
900 		}
901 		tsleep(&cnt.v_free_count, PVM, "vmwait", 0);
902 	}
903 	splx(s);
904 }
905 
906 /*
907  *	vm_await:	(also see VM_AWAIT macro)
908  *
909  *	asleep on an event that will signal when free pages are available
910  *	for allocation.
911  */
912 
913 void
914 vm_await()
915 {
916 	int s;
917 
918 	s = splvm();
919 	if (curproc == pageproc) {
920 		vm_pageout_pages_needed = 1;
921 		asleep(&vm_pageout_pages_needed, PSWP, "vmwait", 0);
922 	} else {
923 		if (!vm_pages_needed) {
924 			vm_pages_needed++;
925 			wakeup(&vm_pages_needed);
926 		}
927 		asleep(&cnt.v_free_count, PVM, "vmwait", 0);
928 	}
929 	splx(s);
930 }
931 
932 #if 0
933 /*
934  *	vm_page_sleep:
935  *
936  *	Block until page is no longer busy.
937  */
938 
939 int
940 vm_page_sleep(vm_page_t m, char *msg, char *busy) {
941 	int slept = 0;
942 	if ((busy && *busy) || (m->flags & PG_BUSY)) {
943 		int s;
944 		s = splvm();
945 		if ((busy && *busy) || (m->flags & PG_BUSY)) {
946 			vm_page_flag_set(m, PG_WANTED);
947 			tsleep(m, PVM, msg, 0);
948 			slept = 1;
949 		}
950 		splx(s);
951 	}
952 	return slept;
953 }
954 
955 #endif
956 
957 #if 0
958 
959 /*
960  *	vm_page_asleep:
961  *
962  *	Similar to vm_page_sleep(), but does not block.  Returns 0 if
963  *	the page is not busy, or 1 if the page is busy.
964  *
965  *	This routine has the side effect of calling asleep() if the page
966  *	was busy (1 returned).
967  */
968 
969 int
970 vm_page_asleep(vm_page_t m, char *msg, char *busy) {
971 	int slept = 0;
972 	if ((busy && *busy) || (m->flags & PG_BUSY)) {
973 		int s;
974 		s = splvm();
975 		if ((busy && *busy) || (m->flags & PG_BUSY)) {
976 			vm_page_flag_set(m, PG_WANTED);
977 			asleep(m, PVM, msg, 0);
978 			slept = 1;
979 		}
980 		splx(s);
981 	}
982 	return slept;
983 }
984 
985 #endif
986 
987 /*
988  *	vm_page_activate:
989  *
990  *	Put the specified page on the active list (if appropriate).
991  *	Ensure that act_count is at least ACT_INIT but do not otherwise
992  *	mess with it.
993  *
994  *	The page queues must be locked.
995  *	This routine may not block.
996  */
997 void
998 vm_page_activate(m)
999 	register vm_page_t m;
1000 {
1001 	int s;
1002 
1003 	s = splvm();
1004 	if (m->queue != PQ_ACTIVE) {
1005 		if ((m->queue - m->pc) == PQ_CACHE)
1006 			cnt.v_reactivated++;
1007 
1008 		vm_page_unqueue(m);
1009 
1010 		if (m->wire_count == 0) {
1011 			m->queue = PQ_ACTIVE;
1012 			vm_page_queues[PQ_ACTIVE].lcnt++;
1013 			TAILQ_INSERT_TAIL(&vm_page_queue_active, m, pageq);
1014 			if (m->act_count < ACT_INIT)
1015 				m->act_count = ACT_INIT;
1016 			cnt.v_active_count++;
1017 		}
1018 	} else {
1019 		if (m->act_count < ACT_INIT)
1020 			m->act_count = ACT_INIT;
1021 	}
1022 
1023 	splx(s);
1024 }
1025 
1026 /*
1027  *	vm_page_free_wakeup:
1028  *
1029  *	Helper routine for vm_page_free_toq() and vm_page_cache().  This
1030  *	routine is called when a page has been added to the cache or free
1031  *	queues.
1032  *
1033  *	This routine may not block.
1034  *	This routine must be called at splvm()
1035  */
1036 static __inline void
1037 vm_page_free_wakeup()
1038 {
1039 	/*
1040 	 * if pageout daemon needs pages, then tell it that there are
1041 	 * some free.
1042 	 */
1043 	if (vm_pageout_pages_needed) {
1044 		wakeup(&vm_pageout_pages_needed);
1045 		vm_pageout_pages_needed = 0;
1046 	}
1047 	/*
1048 	 * wakeup processes that are waiting on memory if we hit a
1049 	 * high water mark. And wakeup scheduler process if we have
1050 	 * lots of memory. this process will swapin processes.
1051 	 */
1052 	if (vm_pages_needed && vm_page_count_min()) {
1053 		wakeup(&cnt.v_free_count);
1054 		vm_pages_needed = 0;
1055 	}
1056 }
1057 
1058 /*
1059  *	vm_page_free_toq:
1060  *
1061  *	Returns the given page to the PQ_FREE list,
1062  *	disassociating it with any VM object.
1063  *
1064  *	Object and page must be locked prior to entry.
1065  *	This routine may not block.
1066  */
1067 
1068 void
1069 vm_page_free_toq(vm_page_t m)
1070 {
1071 	int s;
1072 	struct vpgqueues *pq;
1073 	vm_object_t object = m->object;
1074 
1075 	s = splvm();
1076 
1077 	cnt.v_tfree++;
1078 
1079 #if !defined(MAX_PERF)
1080 	if (m->busy || ((m->queue - m->pc) == PQ_FREE) ||
1081 		(m->hold_count != 0)) {
1082 		printf(
1083 		"vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n",
1084 		    (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0,
1085 		    m->hold_count);
1086 		if ((m->queue - m->pc) == PQ_FREE)
1087 			panic("vm_page_free: freeing free page");
1088 		else
1089 			panic("vm_page_free: freeing busy page");
1090 	}
1091 #endif
1092 
1093 	/*
1094 	 * unqueue, then remove page.  Note that we cannot destroy
1095 	 * the page here because we do not want to call the pager's
1096 	 * callback routine until after we've put the page on the
1097 	 * appropriate free queue.
1098 	 */
1099 
1100 	vm_page_unqueue_nowakeup(m);
1101 	vm_page_remove(m);
1102 
1103 	/*
1104 	 * If fictitious remove object association and
1105 	 * return, otherwise delay object association removal.
1106 	 */
1107 
1108 	if ((m->flags & PG_FICTITIOUS) != 0) {
1109 		splx(s);
1110 		return;
1111 	}
1112 
1113 	m->valid = 0;
1114 	vm_page_undirty(m);
1115 
1116 	if (m->wire_count != 0) {
1117 #if !defined(MAX_PERF)
1118 		if (m->wire_count > 1) {
1119 			panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx",
1120 				m->wire_count, (long)m->pindex);
1121 		}
1122 #endif
1123 		printf("vm_page_free: freeing wired page\n");
1124 		m->wire_count = 0;
1125 		cnt.v_wire_count--;
1126 	}
1127 
1128 	/*
1129 	 * If we've exhausted the object's resident pages we want to free
1130 	 * it up.
1131 	 */
1132 
1133 	if (object &&
1134 	    (object->type == OBJT_VNODE) &&
1135 	    ((object->flags & OBJ_DEAD) == 0)
1136 	) {
1137 		struct vnode *vp = (struct vnode *)object->handle;
1138 
1139 		if (vp && VSHOULDFREE(vp)) {
1140 			if ((vp->v_flag & (VTBFREE|VDOOMED|VFREE)) == 0) {
1141 				TAILQ_INSERT_TAIL(&vnode_tobefree_list, vp, v_freelist);
1142 				vp->v_flag |= VTBFREE;
1143 			}
1144 		}
1145 	}
1146 
1147 #ifdef __alpha__
1148 	pmap_page_is_free(m);
1149 #endif
1150 
1151 	m->queue = PQ_FREE + m->pc;
1152 	pq = &vm_page_queues[m->queue];
1153 	pq->lcnt++;
1154 	++(*pq->cnt);
1155 
1156 	/*
1157 	 * Put zero'd pages on the end ( where we look for zero'd pages
1158 	 * first ) and non-zerod pages at the head.
1159 	 */
1160 
1161 	if (m->flags & PG_ZERO) {
1162 		TAILQ_INSERT_TAIL(pq->pl, m, pageq);
1163 		++vm_page_zero_count;
1164 	} else {
1165 		TAILQ_INSERT_HEAD(pq->pl, m, pageq);
1166 	}
1167 
1168 	vm_page_free_wakeup();
1169 
1170 	splx(s);
1171 }
1172 
1173 /*
1174  *	vm_page_wire:
1175  *
1176  *	Mark this page as wired down by yet
1177  *	another map, removing it from paging queues
1178  *	as necessary.
1179  *
1180  *	The page queues must be locked.
1181  *	This routine may not block.
1182  */
1183 void
1184 vm_page_wire(m)
1185 	register vm_page_t m;
1186 {
1187 	int s;
1188 
1189 	s = splvm();
1190 	if (m->wire_count == 0) {
1191 		vm_page_unqueue(m);
1192 		cnt.v_wire_count++;
1193 	}
1194 	m->wire_count++;
1195 	splx(s);
1196 	vm_page_flag_set(m, PG_MAPPED);
1197 }
1198 
1199 /*
1200  *	vm_page_unwire:
1201  *
1202  *	Release one wiring of this page, potentially
1203  *	enabling it to be paged again.
1204  *
1205  *	Many pages placed on the inactive queue should actually go
1206  *	into the cache, but it is difficult to figure out which.  What
1207  *	we do instead, if the inactive target is well met, is to put
1208  *	clean pages at the head of the inactive queue instead of the tail.
1209  *	This will cause them to be moved to the cache more quickly and
1210  *	if not actively re-referenced, freed more quickly.  If we just
1211  *	stick these pages at the end of the inactive queue, heavy filesystem
1212  *	meta-data accesses can cause an unnecessary paging load on memory bound
1213  *	processes.  This optimization causes one-time-use metadata to be
1214  *	reused more quickly.
1215  *
1216  *	A number of routines use vm_page_unwire() to guarentee that the page
1217  *	will go into either the inactive or active queues, and will NEVER
1218  *	be placed in the cache - for example, just after dirtying a page.
1219  *	dirty pages in the cache are not allowed.
1220  *
1221  *	The page queues must be locked.
1222  *	This routine may not block.
1223  */
1224 void
1225 vm_page_unwire(m, activate)
1226 	register vm_page_t m;
1227 	int activate;
1228 {
1229 	int s;
1230 
1231 	s = splvm();
1232 
1233 	if (m->wire_count > 0) {
1234 		m->wire_count--;
1235 		if (m->wire_count == 0) {
1236 			cnt.v_wire_count--;
1237 			if (activate) {
1238 				TAILQ_INSERT_TAIL(&vm_page_queue_active, m, pageq);
1239 				m->queue = PQ_ACTIVE;
1240 				vm_page_queues[PQ_ACTIVE].lcnt++;
1241 				cnt.v_active_count++;
1242 			} else {
1243 				TAILQ_INSERT_TAIL(&vm_page_queue_inactive, m, pageq);
1244 				m->queue = PQ_INACTIVE;
1245 				vm_page_queues[PQ_INACTIVE].lcnt++;
1246 				cnt.v_inactive_count++;
1247 			}
1248 		}
1249 	} else {
1250 #if !defined(MAX_PERF)
1251 		panic("vm_page_unwire: invalid wire count: %d\n", m->wire_count);
1252 #endif
1253 	}
1254 	splx(s);
1255 }
1256 
1257 
1258 /*
1259  * Move the specified page to the inactive queue.  If the page has
1260  * any associated swap, the swap is deallocated.
1261  *
1262  * Normally athead is 0 resulting in LRU operation.  athead is set
1263  * to 1 if we want this page to be 'as if it were placed in the cache',
1264  * except without unmapping it from the process address space.
1265  *
1266  * This routine may not block.
1267  */
1268 static __inline void
1269 _vm_page_deactivate(vm_page_t m, int athead)
1270 {
1271 	int s;
1272 
1273 	/*
1274 	 * Ignore if already inactive.
1275 	 */
1276 	if (m->queue == PQ_INACTIVE)
1277 		return;
1278 
1279 	s = splvm();
1280 	if (m->wire_count == 0) {
1281 		if ((m->queue - m->pc) == PQ_CACHE)
1282 			cnt.v_reactivated++;
1283 		vm_page_unqueue(m);
1284 		if (athead)
1285 			TAILQ_INSERT_HEAD(&vm_page_queue_inactive, m, pageq);
1286 		else
1287 			TAILQ_INSERT_TAIL(&vm_page_queue_inactive, m, pageq);
1288 		m->queue = PQ_INACTIVE;
1289 		vm_page_queues[PQ_INACTIVE].lcnt++;
1290 		cnt.v_inactive_count++;
1291 	}
1292 	splx(s);
1293 }
1294 
1295 void
1296 vm_page_deactivate(vm_page_t m)
1297 {
1298     _vm_page_deactivate(m, 0);
1299 }
1300 
1301 /*
1302  * vm_page_cache
1303  *
1304  * Put the specified page onto the page cache queue (if appropriate).
1305  *
1306  * This routine may not block.
1307  */
1308 void
1309 vm_page_cache(m)
1310 	register vm_page_t m;
1311 {
1312 	int s;
1313 
1314 #if !defined(MAX_PERF)
1315 	if ((m->flags & PG_BUSY) || m->busy || m->wire_count) {
1316 		printf("vm_page_cache: attempting to cache busy page\n");
1317 		return;
1318 	}
1319 #endif
1320 	if ((m->queue - m->pc) == PQ_CACHE)
1321 		return;
1322 
1323 	/*
1324 	 * Remove all pmaps and indicate that the page is not
1325 	 * writeable or mapped.
1326 	 */
1327 
1328 	vm_page_protect(m, VM_PROT_NONE);
1329 #if !defined(MAX_PERF)
1330 	if (m->dirty != 0) {
1331 		panic("vm_page_cache: caching a dirty page, pindex: %ld",
1332 			(long)m->pindex);
1333 	}
1334 #endif
1335 	s = splvm();
1336 	vm_page_unqueue_nowakeup(m);
1337 	m->queue = PQ_CACHE + m->pc;
1338 	vm_page_queues[m->queue].lcnt++;
1339 	TAILQ_INSERT_TAIL(vm_page_queues[m->queue].pl, m, pageq);
1340 	cnt.v_cache_count++;
1341 	vm_page_free_wakeup();
1342 	splx(s);
1343 }
1344 
1345 /*
1346  * vm_page_dontneed
1347  *
1348  *	Cache, deactivate, or do nothing as appropriate.  This routine
1349  *	is typically used by madvise() MADV_DONTNEED.
1350  *
1351  *	Generally speaking we want to move the page into the cache so
1352  *	it gets reused quickly.  However, this can result in a silly syndrome
1353  *	due to the page recycling too quickly.  Small objects will not be
1354  *	fully cached.  On the otherhand, if we move the page to the inactive
1355  *	queue we wind up with a problem whereby very large objects
1356  *	unnecessarily blow away our inactive and cache queues.
1357  *
1358  *	The solution is to move the pages based on a fixed weighting.  We
1359  *	either leave them alone, deactivate them, or move them to the cache,
1360  *	where moving them to the cache has the highest weighting.
1361  *	By forcing some pages into other queues we eventually force the
1362  *	system to balance the queues, potentially recovering other unrelated
1363  *	space from active.  The idea is to not force this to happen too
1364  *	often.
1365  */
1366 
1367 void
1368 vm_page_dontneed(m)
1369 	vm_page_t m;
1370 {
1371 	static int dnweight;
1372 	int dnw;
1373 	int head;
1374 
1375 	dnw = ++dnweight;
1376 
1377 	/*
1378 	 * occassionally leave the page alone
1379 	 */
1380 
1381 	if ((dnw & 0x01F0) == 0 ||
1382 	    m->queue == PQ_INACTIVE ||
1383 	    m->queue - m->pc == PQ_CACHE
1384 	) {
1385 		if (m->act_count >= ACT_INIT)
1386 			--m->act_count;
1387 		return;
1388 	}
1389 
1390 	if (m->dirty == 0)
1391 		vm_page_test_dirty(m);
1392 
1393 	if (m->dirty || (dnw & 0x0070) == 0) {
1394 		/*
1395 		 * Deactivate the page 3 times out of 32.
1396 		 */
1397 		head = 0;
1398 	} else {
1399 		/*
1400 		 * Cache the page 28 times out of every 32.  Note that
1401 		 * the page is deactivated instead of cached, but placed
1402 		 * at the head of the queue instead of the tail.
1403 		 */
1404 		head = 1;
1405 	}
1406 	_vm_page_deactivate(m, head);
1407 }
1408 
1409 /*
1410  * Grab a page, waiting until we are waken up due to the page
1411  * changing state.  We keep on waiting, if the page continues
1412  * to be in the object.  If the page doesn't exist, allocate it.
1413  *
1414  * This routine may block.
1415  */
1416 vm_page_t
1417 vm_page_grab(object, pindex, allocflags)
1418 	vm_object_t object;
1419 	vm_pindex_t pindex;
1420 	int allocflags;
1421 {
1422 
1423 	vm_page_t m;
1424 	int s, generation;
1425 
1426 retrylookup:
1427 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
1428 		if (m->busy || (m->flags & PG_BUSY)) {
1429 			generation = object->generation;
1430 
1431 			s = splvm();
1432 			while ((object->generation == generation) &&
1433 					(m->busy || (m->flags & PG_BUSY))) {
1434 				vm_page_flag_set(m, PG_WANTED | PG_REFERENCED);
1435 				tsleep(m, PVM, "pgrbwt", 0);
1436 				if ((allocflags & VM_ALLOC_RETRY) == 0) {
1437 					splx(s);
1438 					return NULL;
1439 				}
1440 			}
1441 			splx(s);
1442 			goto retrylookup;
1443 		} else {
1444 			vm_page_busy(m);
1445 			return m;
1446 		}
1447 	}
1448 
1449 	m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY);
1450 	if (m == NULL) {
1451 		VM_WAIT;
1452 		if ((allocflags & VM_ALLOC_RETRY) == 0)
1453 			return NULL;
1454 		goto retrylookup;
1455 	}
1456 
1457 	return m;
1458 }
1459 
1460 /*
1461  * Mapping function for valid bits or for dirty bits in
1462  * a page.  May not block.
1463  *
1464  * Inputs are required to range within a page.
1465  */
1466 
1467 __inline int
1468 vm_page_bits(int base, int size)
1469 {
1470 	int first_bit;
1471 	int last_bit;
1472 
1473 	KASSERT(
1474 	    base + size <= PAGE_SIZE,
1475 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
1476 	);
1477 
1478 	if (size == 0)		/* handle degenerate case */
1479 		return(0);
1480 
1481 	first_bit = base >> DEV_BSHIFT;
1482 	last_bit = (base + size - 1) >> DEV_BSHIFT;
1483 
1484 	return ((2 << last_bit) - (1 << first_bit));
1485 }
1486 
1487 /*
1488  *	vm_page_set_validclean:
1489  *
1490  *	Sets portions of a page valid and clean.  The arguments are expected
1491  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
1492  *	of any partial chunks touched by the range.  The invalid portion of
1493  *	such chunks will be zero'd.
1494  *
1495  *	This routine may not block.
1496  *
1497  *	(base + size) must be less then or equal to PAGE_SIZE.
1498  */
1499 void
1500 vm_page_set_validclean(m, base, size)
1501 	vm_page_t m;
1502 	int base;
1503 	int size;
1504 {
1505 	int pagebits;
1506 	int frag;
1507 	int endoff;
1508 
1509 	if (size == 0)	/* handle degenerate case */
1510 		return;
1511 
1512 	/*
1513 	 * If the base is not DEV_BSIZE aligned and the valid
1514 	 * bit is clear, we have to zero out a portion of the
1515 	 * first block.
1516 	 */
1517 
1518 	if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
1519 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0
1520 	) {
1521 		pmap_zero_page_area(
1522 		    VM_PAGE_TO_PHYS(m),
1523 		    frag,
1524 		    base - frag
1525 		);
1526 	}
1527 
1528 	/*
1529 	 * If the ending offset is not DEV_BSIZE aligned and the
1530 	 * valid bit is clear, we have to zero out a portion of
1531 	 * the last block.
1532 	 */
1533 
1534 	endoff = base + size;
1535 
1536 	if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
1537 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0
1538 	) {
1539 		pmap_zero_page_area(
1540 		    VM_PAGE_TO_PHYS(m),
1541 		    endoff,
1542 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1))
1543 		);
1544 	}
1545 
1546 	/*
1547 	 * Set valid, clear dirty bits.  If validating the entire
1548 	 * page we can safely clear the pmap modify bit.
1549 	 */
1550 
1551 	pagebits = vm_page_bits(base, size);
1552 	m->valid |= pagebits;
1553 	m->dirty &= ~pagebits;
1554 
1555 	if (base == 0 && size == PAGE_SIZE)
1556 		pmap_clear_modify(VM_PAGE_TO_PHYS(m));
1557 }
1558 
1559 #if 0
1560 
1561 void
1562 vm_page_set_dirty(m, base, size)
1563 	vm_page_t m;
1564 	int base;
1565 	int size;
1566 {
1567 	m->dirty |= vm_page_bits(base, size);
1568 }
1569 
1570 #endif
1571 
1572 void
1573 vm_page_clear_dirty(m, base, size)
1574 	vm_page_t m;
1575 	int base;
1576 	int size;
1577 {
1578 	m->dirty &= ~vm_page_bits(base, size);
1579 }
1580 
1581 /*
1582  *	vm_page_set_invalid:
1583  *
1584  *	Invalidates DEV_BSIZE'd chunks within a page.  Both the
1585  *	valid and dirty bits for the effected areas are cleared.
1586  *
1587  *	May not block.
1588  */
1589 void
1590 vm_page_set_invalid(m, base, size)
1591 	vm_page_t m;
1592 	int base;
1593 	int size;
1594 {
1595 	int bits;
1596 
1597 	bits = vm_page_bits(base, size);
1598 	m->valid &= ~bits;
1599 	m->dirty &= ~bits;
1600 	m->object->generation++;
1601 }
1602 
1603 /*
1604  * vm_page_zero_invalid()
1605  *
1606  *	The kernel assumes that the invalid portions of a page contain
1607  *	garbage, but such pages can be mapped into memory by user code.
1608  *	When this occurs, we must zero out the non-valid portions of the
1609  *	page so user code sees what it expects.
1610  *
1611  *	Pages are most often semi-valid when the end of a file is mapped
1612  *	into memory and the file's size is not page aligned.
1613  */
1614 
1615 void
1616 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
1617 {
1618 	int b;
1619 	int i;
1620 
1621 	/*
1622 	 * Scan the valid bits looking for invalid sections that
1623 	 * must be zerod.  Invalid sub-DEV_BSIZE'd areas ( where the
1624 	 * valid bit may be set ) have already been zerod by
1625 	 * vm_page_set_validclean().
1626 	 */
1627 
1628 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
1629 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
1630 		    (m->valid & (1 << i))
1631 		) {
1632 			if (i > b) {
1633 				pmap_zero_page_area(
1634 				    VM_PAGE_TO_PHYS(m),
1635 				    b << DEV_BSHIFT,
1636 				    (i - b) << DEV_BSHIFT
1637 				);
1638 			}
1639 			b = i + 1;
1640 		}
1641 	}
1642 
1643 	/*
1644 	 * setvalid is TRUE when we can safely set the zero'd areas
1645 	 * as being valid.  We can do this if there are no cache consistancy
1646 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
1647 	 */
1648 
1649 	if (setvalid)
1650 		m->valid = VM_PAGE_BITS_ALL;
1651 }
1652 
1653 /*
1654  *	vm_page_is_valid:
1655  *
1656  *	Is (partial) page valid?  Note that the case where size == 0
1657  *	will return FALSE in the degenerate case where the page is
1658  *	entirely invalid, and TRUE otherwise.
1659  *
1660  *	May not block.
1661  */
1662 
1663 int
1664 vm_page_is_valid(m, base, size)
1665 	vm_page_t m;
1666 	int base;
1667 	int size;
1668 {
1669 	int bits = vm_page_bits(base, size);
1670 
1671 	if (m->valid && ((m->valid & bits) == bits))
1672 		return 1;
1673 	else
1674 		return 0;
1675 }
1676 
1677 /*
1678  * update dirty bits from pmap/mmu.  May not block.
1679  */
1680 
1681 void
1682 vm_page_test_dirty(m)
1683 	vm_page_t m;
1684 {
1685 	if ((m->dirty != VM_PAGE_BITS_ALL) &&
1686 	    pmap_is_modified(VM_PAGE_TO_PHYS(m))) {
1687 		vm_page_dirty(m);
1688 	}
1689 }
1690 
1691 /*
1692  * This interface is for merging with malloc() someday.
1693  * Even if we never implement compaction so that contiguous allocation
1694  * works after initialization time, malloc()'s data structures are good
1695  * for statistics and for allocations of less than a page.
1696  */
1697 void *
1698 contigmalloc1(size, type, flags, low, high, alignment, boundary, map)
1699 	unsigned long size;	/* should be size_t here and for malloc() */
1700 	struct malloc_type *type;
1701 	int flags;
1702 	unsigned long low;
1703 	unsigned long high;
1704 	unsigned long alignment;
1705 	unsigned long boundary;
1706 	vm_map_t map;
1707 {
1708 	int i, s, start;
1709 	vm_offset_t addr, phys, tmp_addr;
1710 	int pass;
1711 	vm_page_t pga = vm_page_array;
1712 
1713 	size = round_page(size);
1714 #if !defined(MAX_PERF)
1715 	if (size == 0)
1716 		panic("contigmalloc1: size must not be 0");
1717 	if ((alignment & (alignment - 1)) != 0)
1718 		panic("contigmalloc1: alignment must be a power of 2");
1719 	if ((boundary & (boundary - 1)) != 0)
1720 		panic("contigmalloc1: boundary must be a power of 2");
1721 #endif
1722 
1723 	start = 0;
1724 	for (pass = 0; pass <= 1; pass++) {
1725 		s = splvm();
1726 again:
1727 		/*
1728 		 * Find first page in array that is free, within range, aligned, and
1729 		 * such that the boundary won't be crossed.
1730 		 */
1731 		for (i = start; i < cnt.v_page_count; i++) {
1732 			int pqtype;
1733 			phys = VM_PAGE_TO_PHYS(&pga[i]);
1734 			pqtype = pga[i].queue - pga[i].pc;
1735 			if (((pqtype == PQ_FREE) || (pqtype == PQ_CACHE)) &&
1736 			    (phys >= low) && (phys < high) &&
1737 			    ((phys & (alignment - 1)) == 0) &&
1738 			    (((phys ^ (phys + size - 1)) & ~(boundary - 1)) == 0))
1739 				break;
1740 		}
1741 
1742 		/*
1743 		 * If the above failed or we will exceed the upper bound, fail.
1744 		 */
1745 		if ((i == cnt.v_page_count) ||
1746 			((VM_PAGE_TO_PHYS(&pga[i]) + size) > high)) {
1747 			vm_page_t m, next;
1748 
1749 again1:
1750 			for (m = TAILQ_FIRST(&vm_page_queue_inactive);
1751 				m != NULL;
1752 				m = next) {
1753 
1754 				KASSERT(m->queue == PQ_INACTIVE,
1755 					("contigmalloc1: page %p is not PQ_INACTIVE", m));
1756 
1757 				next = TAILQ_NEXT(m, pageq);
1758 				if (vm_page_sleep_busy(m, TRUE, "vpctw0"))
1759 					goto again1;
1760 				vm_page_test_dirty(m);
1761 				if (m->dirty) {
1762 					if (m->object->type == OBJT_VNODE) {
1763 						vn_lock(m->object->handle, LK_EXCLUSIVE | LK_RETRY, curproc);
1764 						vm_object_page_clean(m->object, 0, 0, OBJPC_SYNC);
1765 						VOP_UNLOCK(m->object->handle, 0, curproc);
1766 						goto again1;
1767 					} else if (m->object->type == OBJT_SWAP ||
1768 								m->object->type == OBJT_DEFAULT) {
1769 						vm_pageout_flush(&m, 1, 0);
1770 						goto again1;
1771 					}
1772 				}
1773 				if ((m->dirty == 0) && (m->busy == 0) && (m->hold_count == 0))
1774 					vm_page_cache(m);
1775 			}
1776 
1777 			for (m = TAILQ_FIRST(&vm_page_queue_active);
1778 				m != NULL;
1779 				m = next) {
1780 
1781 				KASSERT(m->queue == PQ_ACTIVE,
1782 					("contigmalloc1: page %p is not PQ_ACTIVE", m));
1783 
1784 				next = TAILQ_NEXT(m, pageq);
1785 				if (vm_page_sleep_busy(m, TRUE, "vpctw1"))
1786 					goto again1;
1787 				vm_page_test_dirty(m);
1788 				if (m->dirty) {
1789 					if (m->object->type == OBJT_VNODE) {
1790 						vn_lock(m->object->handle, LK_EXCLUSIVE | LK_RETRY, curproc);
1791 						vm_object_page_clean(m->object, 0, 0, OBJPC_SYNC);
1792 						VOP_UNLOCK(m->object->handle, 0, curproc);
1793 						goto again1;
1794 					} else if (m->object->type == OBJT_SWAP ||
1795 								m->object->type == OBJT_DEFAULT) {
1796 						vm_pageout_flush(&m, 1, 0);
1797 						goto again1;
1798 					}
1799 				}
1800 				if ((m->dirty == 0) && (m->busy == 0) && (m->hold_count == 0))
1801 					vm_page_cache(m);
1802 			}
1803 
1804 			splx(s);
1805 			continue;
1806 		}
1807 		start = i;
1808 
1809 		/*
1810 		 * Check successive pages for contiguous and free.
1811 		 */
1812 		for (i = start + 1; i < (start + size / PAGE_SIZE); i++) {
1813 			int pqtype;
1814 			pqtype = pga[i].queue - pga[i].pc;
1815 			if ((VM_PAGE_TO_PHYS(&pga[i]) !=
1816 			    (VM_PAGE_TO_PHYS(&pga[i - 1]) + PAGE_SIZE)) ||
1817 			    ((pqtype != PQ_FREE) && (pqtype != PQ_CACHE))) {
1818 				start++;
1819 				goto again;
1820 			}
1821 		}
1822 
1823 		for (i = start; i < (start + size / PAGE_SIZE); i++) {
1824 			int pqtype;
1825 			vm_page_t m = &pga[i];
1826 
1827 			pqtype = m->queue - m->pc;
1828 			if (pqtype == PQ_CACHE) {
1829 				vm_page_busy(m);
1830 				vm_page_free(m);
1831 			}
1832 
1833 			TAILQ_REMOVE(vm_page_queues[m->queue].pl, m, pageq);
1834 			vm_page_queues[m->queue].lcnt--;
1835 			cnt.v_free_count--;
1836 			m->valid = VM_PAGE_BITS_ALL;
1837 			m->flags = 0;
1838 			KASSERT(m->dirty == 0, ("contigmalloc1: page %p was dirty", m));
1839 			m->wire_count = 0;
1840 			m->busy = 0;
1841 			m->queue = PQ_NONE;
1842 			m->object = NULL;
1843 			vm_page_wire(m);
1844 		}
1845 
1846 		/*
1847 		 * We've found a contiguous chunk that meets are requirements.
1848 		 * Allocate kernel VM, unfree and assign the physical pages to it and
1849 		 * return kernel VM pointer.
1850 		 */
1851 		tmp_addr = addr = kmem_alloc_pageable(map, size);
1852 		if (addr == 0) {
1853 			/*
1854 			 * XXX We almost never run out of kernel virtual
1855 			 * space, so we don't make the allocated memory
1856 			 * above available.
1857 			 */
1858 			splx(s);
1859 			return (NULL);
1860 		}
1861 
1862 		for (i = start; i < (start + size / PAGE_SIZE); i++) {
1863 			vm_page_t m = &pga[i];
1864 			vm_page_insert(m, kernel_object,
1865 				OFF_TO_IDX(tmp_addr - VM_MIN_KERNEL_ADDRESS));
1866 			pmap_kenter(tmp_addr, VM_PAGE_TO_PHYS(m));
1867 			tmp_addr += PAGE_SIZE;
1868 		}
1869 
1870 		splx(s);
1871 		return ((void *)addr);
1872 	}
1873 	return NULL;
1874 }
1875 
1876 void *
1877 contigmalloc(size, type, flags, low, high, alignment, boundary)
1878 	unsigned long size;	/* should be size_t here and for malloc() */
1879 	struct malloc_type *type;
1880 	int flags;
1881 	unsigned long low;
1882 	unsigned long high;
1883 	unsigned long alignment;
1884 	unsigned long boundary;
1885 {
1886 	return contigmalloc1(size, type, flags, low, high, alignment, boundary,
1887 			     kernel_map);
1888 }
1889 
1890 void
1891 contigfree(addr, size, type)
1892 	void *addr;
1893 	unsigned long size;
1894 	struct malloc_type *type;
1895 {
1896 	kmem_free(kernel_map, (vm_offset_t)addr, size);
1897 }
1898 
1899 vm_offset_t
1900 vm_page_alloc_contig(size, low, high, alignment)
1901 	vm_offset_t size;
1902 	vm_offset_t low;
1903 	vm_offset_t high;
1904 	vm_offset_t alignment;
1905 {
1906 	return ((vm_offset_t)contigmalloc1(size, M_DEVBUF, M_NOWAIT, low, high,
1907 					  alignment, 0ul, kernel_map));
1908 }
1909 
1910 #include "opt_ddb.h"
1911 #ifdef DDB
1912 #include <sys/kernel.h>
1913 
1914 #include <ddb/ddb.h>
1915 
1916 DB_SHOW_COMMAND(page, vm_page_print_page_info)
1917 {
1918 	db_printf("cnt.v_free_count: %d\n", cnt.v_free_count);
1919 	db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count);
1920 	db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count);
1921 	db_printf("cnt.v_active_count: %d\n", cnt.v_active_count);
1922 	db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count);
1923 	db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved);
1924 	db_printf("cnt.v_free_min: %d\n", cnt.v_free_min);
1925 	db_printf("cnt.v_free_target: %d\n", cnt.v_free_target);
1926 	db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min);
1927 	db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target);
1928 }
1929 
1930 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
1931 {
1932 	int i;
1933 	db_printf("PQ_FREE:");
1934 	for(i=0;i<PQ_L2_SIZE;i++) {
1935 		db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt);
1936 	}
1937 	db_printf("\n");
1938 
1939 	db_printf("PQ_CACHE:");
1940 	for(i=0;i<PQ_L2_SIZE;i++) {
1941 		db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt);
1942 	}
1943 	db_printf("\n");
1944 
1945 	db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n",
1946 		vm_page_queues[PQ_ACTIVE].lcnt,
1947 		vm_page_queues[PQ_INACTIVE].lcnt);
1948 }
1949 #endif /* DDB */
1950