xref: /dflybsd-src/sys/vm/vm_page.c (revision ac2e3f5effc58aa364c7e5c199f35ebbae7cda81)
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 1.7 2003/07/19 21:14:53 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.
154  *	Must be called at splhigh().
155  */
156 vm_page_t
157 vm_add_new_page(vm_offset_t pa)
158 {
159 	vm_page_t m;
160 
161 	++vmstats.v_page_count;
162 	++vmstats.v_free_count;
163 	m = PHYS_TO_VM_PAGE(pa);
164 	m->phys_addr = pa;
165 	m->flags = 0;
166 	m->pc = (pa >> PAGE_SHIFT) & PQ_L2_MASK;
167 	m->queue = m->pc + PQ_FREE;
168 	TAILQ_INSERT_HEAD(&vm_page_queues[m->queue].pl, m, pageq);
169 	vm_page_queues[m->queue].lcnt++;
170 	return (m);
171 }
172 
173 /*
174  *	vm_page_startup:
175  *
176  *	Initializes the resident memory module.
177  *
178  *	Allocates memory for the page cells, and
179  *	for the object/offset-to-page hash table headers.
180  *	Each page cell is initialized and placed on the free list.
181  */
182 
183 vm_offset_t
184 vm_page_startup(vm_offset_t starta, vm_offset_t enda, vm_offset_t vaddr)
185 {
186 	vm_offset_t mapped;
187 	struct vm_page **bucket;
188 	vm_size_t npages, page_range;
189 	vm_offset_t new_end;
190 	int i;
191 	vm_offset_t pa;
192 	int nblocks;
193 	vm_offset_t last_pa;
194 
195 	/* the biggest memory array is the second group of pages */
196 	vm_offset_t end;
197 	vm_offset_t biggestone, biggestsize;
198 
199 	vm_offset_t total;
200 
201 	total = 0;
202 	biggestsize = 0;
203 	biggestone = 0;
204 	nblocks = 0;
205 	vaddr = round_page(vaddr);
206 
207 	for (i = 0; phys_avail[i + 1]; i += 2) {
208 		phys_avail[i] = round_page(phys_avail[i]);
209 		phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
210 	}
211 
212 	for (i = 0; phys_avail[i + 1]; i += 2) {
213 		int size = phys_avail[i + 1] - phys_avail[i];
214 
215 		if (size > biggestsize) {
216 			biggestone = i;
217 			biggestsize = size;
218 		}
219 		++nblocks;
220 		total += size;
221 	}
222 
223 	end = phys_avail[biggestone+1];
224 
225 	/*
226 	 * Initialize the queue headers for the free queue, the active queue
227 	 * and the inactive queue.
228 	 */
229 
230 	vm_page_queue_init();
231 
232 	/*
233 	 * Allocate (and initialize) the hash table buckets.
234 	 *
235 	 * The number of buckets MUST BE a power of 2, and the actual value is
236 	 * the next power of 2 greater than the number of physical pages in
237 	 * the system.
238 	 *
239 	 * We make the hash table approximately 2x the number of pages to
240 	 * reduce the chain length.  This is about the same size using the
241 	 * singly-linked list as the 1x hash table we were using before
242 	 * using TAILQ but the chain length will be smaller.
243 	 *
244 	 * Note: This computation can be tweaked if desired.
245 	 */
246 	vm_page_buckets = (struct vm_page **)vaddr;
247 	bucket = vm_page_buckets;
248 	if (vm_page_bucket_count == 0) {
249 		vm_page_bucket_count = 1;
250 		while (vm_page_bucket_count < atop(total))
251 			vm_page_bucket_count <<= 1;
252 	}
253 	vm_page_bucket_count <<= 1;
254 	vm_page_hash_mask = vm_page_bucket_count - 1;
255 
256 	/*
257 	 * Validate these addresses.
258 	 */
259 	new_end = end - vm_page_bucket_count * sizeof(struct vm_page *);
260 	new_end = trunc_page(new_end);
261 	mapped = round_page(vaddr);
262 	vaddr = pmap_map(mapped, new_end, end,
263 	    VM_PROT_READ | VM_PROT_WRITE);
264 	vaddr = round_page(vaddr);
265 	bzero((caddr_t) mapped, vaddr - mapped);
266 
267 	for (i = 0; i < vm_page_bucket_count; i++) {
268 		*bucket = NULL;
269 		bucket++;
270 	}
271 
272 	/*
273 	 * Compute the number of pages of memory that will be available for
274 	 * use (taking into account the overhead of a page structure per
275 	 * page).
276 	 */
277 
278 	first_page = phys_avail[0] / PAGE_SIZE;
279 
280 	page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page;
281 	npages = (total - (page_range * sizeof(struct vm_page)) -
282 	    (end - new_end)) / PAGE_SIZE;
283 
284 	end = new_end;
285 	/*
286 	 * Initialize the mem entry structures now, and put them in the free
287 	 * queue.
288 	 */
289 	vm_page_array = (vm_page_t) vaddr;
290 	mapped = vaddr;
291 
292 	/*
293 	 * Validate these addresses.
294 	 */
295 
296 	new_end = trunc_page(end - page_range * sizeof(struct vm_page));
297 	mapped = pmap_map(mapped, new_end, end,
298 	    VM_PROT_READ | VM_PROT_WRITE);
299 
300 	/*
301 	 * Clear all of the page structures
302 	 */
303 	bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
304 	vm_page_array_size = page_range;
305 
306 	/*
307 	 * Construct the free queue(s) in descending order (by physical
308 	 * address) so that the first 16MB of physical memory is allocated
309 	 * last rather than first.  On large-memory machines, this avoids
310 	 * the exhaustion of low physical memory before isa_dmainit has run.
311 	 */
312 	vmstats.v_page_count = 0;
313 	vmstats.v_free_count = 0;
314 	for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) {
315 		pa = phys_avail[i];
316 		if (i == biggestone)
317 			last_pa = new_end;
318 		else
319 			last_pa = phys_avail[i + 1];
320 		while (pa < last_pa && npages-- > 0) {
321 			vm_add_new_page(pa);
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(vm_object_t object, vm_pindex_t pindex)
341 {
342 	int i = ((uintptr_t)object + pindex) ^ object->hash_rand;
343 
344 	return(i & vm_page_hash_mask);
345 }
346 
347 void
348 vm_page_unhold(vm_page_t mem)
349 {
350 	--mem->hold_count;
351 	KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!"));
352 	if (mem->hold_count == 0 && mem->queue == PQ_HOLD)
353 		vm_page_free_toq(mem);
354 }
355 
356 /*
357  *	vm_page_insert:		[ internal use only ]
358  *
359  *	Inserts the given mem entry into the object and object list.
360  *
361  *	The pagetables are not updated but will presumably fault the page
362  *	in if necessary, or if a kernel page the caller will at some point
363  *	enter the page into the kernel's pmap.  We are not allowed to block
364  *	here so we *can't* do this anyway.
365  *
366  *	The object and page must be locked, and must be splhigh.
367  *	This routine may not block.
368  */
369 
370 void
371 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
372 {
373 	struct vm_page **bucket;
374 
375 	if (m->object != NULL)
376 		panic("vm_page_insert: already inserted");
377 
378 	/*
379 	 * Record the object/offset pair in this page
380 	 */
381 
382 	m->object = object;
383 	m->pindex = pindex;
384 
385 	/*
386 	 * Insert it into the object_object/offset hash table
387 	 */
388 
389 	bucket = &vm_page_buckets[vm_page_hash(object, pindex)];
390 	m->hnext = *bucket;
391 	*bucket = m;
392 	vm_page_bucket_generation++;
393 
394 	/*
395 	 * Now link into the object's list of backed pages.
396 	 */
397 
398 	TAILQ_INSERT_TAIL(&object->memq, m, listq);
399 	object->generation++;
400 
401 	/*
402 	 * show that the object has one more resident page.
403 	 */
404 
405 	object->resident_page_count++;
406 
407 	/*
408 	 * Since we are inserting a new and possibly dirty page,
409 	 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags.
410 	 */
411 	if (m->flags & PG_WRITEABLE)
412 		vm_object_set_writeable_dirty(object);
413 }
414 
415 /*
416  *	vm_page_remove:
417  *				NOTE: used by device pager as well -wfj
418  *
419  *	Removes the given mem entry from the object/offset-page
420  *	table and the object page list, but do not invalidate/terminate
421  *	the backing store.
422  *
423  *	The object and page must be locked, and at splhigh.
424  *	The underlying pmap entry (if any) is NOT removed here.
425  *	This routine may not block.
426  */
427 
428 void
429 vm_page_remove(vm_page_t m)
430 {
431 	vm_object_t object;
432 
433 	if (m->object == NULL)
434 		return;
435 
436 	if ((m->flags & PG_BUSY) == 0) {
437 		panic("vm_page_remove: page not busy");
438 	}
439 
440 	/*
441 	 * Basically destroy the page.
442 	 */
443 
444 	vm_page_wakeup(m);
445 
446 	object = m->object;
447 
448 	/*
449 	 * Remove from the object_object/offset hash table.  The object
450 	 * must be on the hash queue, we will panic if it isn't
451 	 *
452 	 * Note: we must NULL-out m->hnext to prevent loops in detached
453 	 * buffers with vm_page_lookup().
454 	 */
455 
456 	{
457 		struct vm_page **bucket;
458 
459 		bucket = &vm_page_buckets[vm_page_hash(m->object, m->pindex)];
460 		while (*bucket != m) {
461 			if (*bucket == NULL)
462 				panic("vm_page_remove(): page not found in hash");
463 			bucket = &(*bucket)->hnext;
464 		}
465 		*bucket = m->hnext;
466 		m->hnext = NULL;
467 		vm_page_bucket_generation++;
468 	}
469 
470 	/*
471 	 * Now remove from the object's list of backed pages.
472 	 */
473 
474 	TAILQ_REMOVE(&object->memq, m, listq);
475 
476 	/*
477 	 * And show that the object has one fewer resident page.
478 	 */
479 
480 	object->resident_page_count--;
481 	object->generation++;
482 
483 	m->object = NULL;
484 }
485 
486 /*
487  *	vm_page_lookup:
488  *
489  *	Returns the page associated with the object/offset
490  *	pair specified; if none is found, NULL is returned.
491  *
492  *	NOTE: the code below does not lock.  It will operate properly if
493  *	an interrupt makes a change, but the generation algorithm will not
494  *	operate properly in an SMP environment where both cpu's are able to run
495  *	kernel code simultaneously.
496  *
497  *	The object must be locked.  No side effects.
498  *	This routine may not block.
499  *	This is a critical path routine
500  */
501 
502 vm_page_t
503 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
504 {
505 	vm_page_t m;
506 	struct vm_page **bucket;
507 	int generation;
508 
509 	/*
510 	 * Search the hash table for this object/offset pair
511 	 */
512 
513 retry:
514 	generation = vm_page_bucket_generation;
515 	bucket = &vm_page_buckets[vm_page_hash(object, pindex)];
516 	for (m = *bucket; m != NULL; m = m->hnext) {
517 		if ((m->object == object) && (m->pindex == pindex)) {
518 			if (vm_page_bucket_generation != generation)
519 				goto retry;
520 			return (m);
521 		}
522 	}
523 	if (vm_page_bucket_generation != generation)
524 		goto retry;
525 	return (NULL);
526 }
527 
528 /*
529  *	vm_page_rename:
530  *
531  *	Move the given memory entry from its
532  *	current object to the specified target object/offset.
533  *
534  *	The object must be locked.
535  *	This routine may not block.
536  *
537  *	Note: this routine will raise itself to splvm(), the caller need not.
538  *
539  *	Note: swap associated with the page must be invalidated by the move.  We
540  *	      have to do this for several reasons:  (1) we aren't freeing the
541  *	      page, (2) we are dirtying the page, (3) the VM system is probably
542  *	      moving the page from object A to B, and will then later move
543  *	      the backing store from A to B and we can't have a conflict.
544  *
545  *	Note: we *always* dirty the page.  It is necessary both for the
546  *	      fact that we moved it, and because we may be invalidating
547  *	      swap.  If the page is on the cache, we have to deactivate it
548  *	      or vm_page_dirty() will panic.  Dirty pages are not allowed
549  *	      on the cache.
550  */
551 
552 void
553 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
554 {
555 	int s;
556 
557 	s = splvm();
558 	vm_page_remove(m);
559 	vm_page_insert(m, new_object, new_pindex);
560 	if (m->queue - m->pc == PQ_CACHE)
561 		vm_page_deactivate(m);
562 	vm_page_dirty(m);
563 	splx(s);
564 }
565 
566 /*
567  * vm_page_unqueue_nowakeup:
568  *
569  * 	vm_page_unqueue() without any wakeup
570  *
571  *	This routine must be called at splhigh().
572  *	This routine may not block.
573  */
574 
575 void
576 vm_page_unqueue_nowakeup(vm_page_t m)
577 {
578 	int queue = m->queue;
579 	struct vpgqueues *pq;
580 	if (queue != PQ_NONE) {
581 		pq = &vm_page_queues[queue];
582 		m->queue = PQ_NONE;
583 		TAILQ_REMOVE(&pq->pl, m, pageq);
584 		(*pq->cnt)--;
585 		pq->lcnt--;
586 	}
587 }
588 
589 /*
590  * vm_page_unqueue:
591  *
592  *	Remove a page from its queue.
593  *
594  *	This routine must be called at splhigh().
595  *	This routine may not block.
596  */
597 
598 void
599 vm_page_unqueue(vm_page_t m)
600 {
601 	int queue = m->queue;
602 	struct vpgqueues *pq;
603 	if (queue != PQ_NONE) {
604 		m->queue = PQ_NONE;
605 		pq = &vm_page_queues[queue];
606 		TAILQ_REMOVE(&pq->pl, m, pageq);
607 		(*pq->cnt)--;
608 		pq->lcnt--;
609 		if ((queue - m->pc) == PQ_CACHE) {
610 			if (vm_paging_needed())
611 				pagedaemon_wakeup();
612 		}
613 	}
614 }
615 
616 #if PQ_L2_SIZE > 1
617 
618 /*
619  *	vm_page_list_find:
620  *
621  *	Find a page on the specified queue with color optimization.
622  *
623  *	The page coloring optimization attempts to locate a page
624  *	that does not overload other nearby pages in the object in
625  *	the cpu's L1 or L2 caches.  We need this optimization because
626  *	cpu caches tend to be physical caches, while object spaces tend
627  *	to be virtual.
628  *
629  *	This routine must be called at splvm().
630  *	This routine may not block.
631  *
632  *	This routine may only be called from the vm_page_list_find() macro
633  *	in vm_page.h
634  */
635 vm_page_t
636 _vm_page_list_find(int basequeue, int index)
637 {
638 	int i;
639 	vm_page_t m = NULL;
640 	struct vpgqueues *pq;
641 
642 	pq = &vm_page_queues[basequeue];
643 
644 	/*
645 	 * Note that for the first loop, index+i and index-i wind up at the
646 	 * same place.  Even though this is not totally optimal, we've already
647 	 * blown it by missing the cache case so we do not care.
648 	 */
649 
650 	for(i = PQ_L2_SIZE / 2; i > 0; --i) {
651 		if ((m = TAILQ_FIRST(&pq[(index + i) & PQ_L2_MASK].pl)) != NULL)
652 			break;
653 
654 		if ((m = TAILQ_FIRST(&pq[(index - i) & PQ_L2_MASK].pl)) != NULL)
655 			break;
656 	}
657 	return(m);
658 }
659 
660 #endif
661 
662 /*
663  *	vm_page_select_cache:
664  *
665  *	Find a page on the cache queue with color optimization.  As pages
666  *	might be found, but not applicable, they are deactivated.  This
667  *	keeps us from using potentially busy cached pages.
668  *
669  *	This routine must be called at splvm().
670  *	This routine may not block.
671  */
672 vm_page_t
673 vm_page_select_cache(vm_object_t object, vm_pindex_t pindex)
674 {
675 	vm_page_t m;
676 
677 	while (TRUE) {
678 		m = vm_page_list_find(
679 		    PQ_CACHE,
680 		    (pindex + object->pg_color) & PQ_L2_MASK,
681 		    FALSE
682 		);
683 		if (m && ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy ||
684 			       m->hold_count || m->wire_count)) {
685 			vm_page_deactivate(m);
686 			continue;
687 		}
688 		return m;
689 	}
690 }
691 
692 /*
693  *	vm_page_select_free:
694  *
695  *	Find a free or zero page, with specified preference.  We attempt to
696  *	inline the nominal case and fall back to _vm_page_select_free()
697  *	otherwise.
698  *
699  *	This routine must be called at splvm().
700  *	This routine may not block.
701  */
702 
703 static __inline vm_page_t
704 vm_page_select_free(vm_object_t object, vm_pindex_t pindex, boolean_t prefer_zero)
705 {
706 	vm_page_t m;
707 
708 	m = vm_page_list_find(
709 		PQ_FREE,
710 		(pindex + object->pg_color) & PQ_L2_MASK,
711 		prefer_zero
712 	);
713 	return(m);
714 }
715 
716 /*
717  *	vm_page_alloc:
718  *
719  *	Allocate and return a memory cell associated
720  *	with this VM object/offset pair.
721  *
722  *	page_req classes:
723  *	VM_ALLOC_NORMAL		normal process request
724  *	VM_ALLOC_SYSTEM		system *really* needs a page
725  *	VM_ALLOC_INTERRUPT	interrupt time request
726  *	VM_ALLOC_ZERO		zero page
727  *
728  *	Object must be locked.
729  *	This routine may not block.
730  *
731  *	Additional special handling is required when called from an
732  *	interrupt (VM_ALLOC_INTERRUPT).  We are not allowed to mess with
733  *	the page cache in this case.
734  */
735 
736 vm_page_t
737 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req)
738 {
739 	vm_page_t m = NULL;
740 	int s;
741 
742 	KASSERT(!vm_page_lookup(object, pindex),
743 		("vm_page_alloc: page already allocated"));
744 
745 	/*
746 	 * The pager is allowed to eat deeper into the free page list.
747 	 */
748 
749 	if ((curthread == pagethread) && (page_req != VM_ALLOC_INTERRUPT)) {
750 		page_req = VM_ALLOC_SYSTEM;
751 	};
752 
753 	s = splvm();
754 
755 loop:
756 	if (vmstats.v_free_count > vmstats.v_free_reserved) {
757 		/*
758 		 * Allocate from the free queue if there are plenty of pages
759 		 * in it.
760 		 */
761 		if (page_req == VM_ALLOC_ZERO)
762 			m = vm_page_select_free(object, pindex, TRUE);
763 		else
764 			m = vm_page_select_free(object, pindex, FALSE);
765 	} else if (
766 	    (page_req == VM_ALLOC_SYSTEM &&
767 	     vmstats.v_cache_count == 0 &&
768 	     vmstats.v_free_count > vmstats.v_interrupt_free_min) ||
769 	    (page_req == VM_ALLOC_INTERRUPT && vmstats.v_free_count > 0)
770 	) {
771 		/*
772 		 * Interrupt or system, dig deeper into the free list.
773 		 */
774 		m = vm_page_select_free(object, pindex, FALSE);
775 	} else if (page_req != VM_ALLOC_INTERRUPT) {
776 		/*
777 		 * Allocatable from cache (non-interrupt only).  On success,
778 		 * we must free the page and try again, thus ensuring that
779 		 * vmstats.v_*_free_min counters are replenished.
780 		 */
781 		m = vm_page_select_cache(object, pindex);
782 		if (m == NULL) {
783 			splx(s);
784 #if defined(DIAGNOSTIC)
785 			if (vmstats.v_cache_count > 0)
786 				printf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", vmstats.v_cache_count);
787 #endif
788 			vm_pageout_deficit++;
789 			pagedaemon_wakeup();
790 			return (NULL);
791 		}
792 		KASSERT(m->dirty == 0, ("Found dirty cache page %p", m));
793 		vm_page_busy(m);
794 		vm_page_protect(m, VM_PROT_NONE);
795 		vm_page_free(m);
796 		goto loop;
797 	} else {
798 		/*
799 		 * Not allocatable from cache from interrupt, give up.
800 		 */
801 		splx(s);
802 		vm_pageout_deficit++;
803 		pagedaemon_wakeup();
804 		return (NULL);
805 	}
806 
807 	/*
808 	 *  At this point we had better have found a good page.
809 	 */
810 
811 	KASSERT(
812 	    m != NULL,
813 	    ("vm_page_alloc(): missing page on free queue\n")
814 	);
815 
816 	/*
817 	 * Remove from free queue
818 	 */
819 
820 	vm_page_unqueue_nowakeup(m);
821 
822 	/*
823 	 * Initialize structure.  Only the PG_ZERO flag is inherited.
824 	 */
825 
826 	if (m->flags & PG_ZERO) {
827 		vm_page_zero_count--;
828 		m->flags = PG_ZERO | PG_BUSY;
829 	} else {
830 		m->flags = PG_BUSY;
831 	}
832 	m->wire_count = 0;
833 	m->hold_count = 0;
834 	m->act_count = 0;
835 	m->busy = 0;
836 	m->valid = 0;
837 	KASSERT(m->dirty == 0, ("vm_page_alloc: free/cache page %p was dirty", m));
838 
839 	/*
840 	 * vm_page_insert() is safe prior to the splx().  Note also that
841 	 * inserting a page here does not insert it into the pmap (which
842 	 * could cause us to block allocating memory).  We cannot block
843 	 * anywhere.
844 	 */
845 
846 	vm_page_insert(m, object, pindex);
847 
848 	/*
849 	 * Don't wakeup too often - wakeup the pageout daemon when
850 	 * we would be nearly out of memory.
851 	 */
852 	if (vm_paging_needed())
853 		pagedaemon_wakeup();
854 
855 	splx(s);
856 
857 	return (m);
858 }
859 
860 /*
861  *	vm_wait:	(also see VM_WAIT macro)
862  *
863  *	Block until free pages are available for allocation
864  *	- Called in various places before memory allocations.
865  */
866 
867 void
868 vm_wait(void)
869 {
870 	int s;
871 
872 	s = splvm();
873 	if (curthread == pagethread) {
874 		vm_pageout_pages_needed = 1;
875 		tsleep(&vm_pageout_pages_needed, 0, "VMWait", 0);
876 	} else {
877 		if (!vm_pages_needed) {
878 			vm_pages_needed = 1;
879 			wakeup(&vm_pages_needed);
880 		}
881 		tsleep(&vmstats.v_free_count, 0, "vmwait", 0);
882 	}
883 	splx(s);
884 }
885 
886 /*
887  *	vm_waitpfault:	(also see VM_WAITPFAULT macro)
888  *
889  *	Block until free pages are available for allocation
890  *	- Called only in vm_fault so that processes page faulting
891  *	  can be easily tracked.
892  *	- Sleeps at a lower priority than vm_wait() so that vm_wait()ing
893  *	  processes will be able to grab memory first.  Do not change
894  *	  this balance without careful testing first.
895  */
896 
897 void
898 vm_waitpfault(void)
899 {
900 	int s;
901 
902 	s = splvm();
903 	if (!vm_pages_needed) {
904 		vm_pages_needed = 1;
905 		wakeup(&vm_pages_needed);
906 	}
907 	tsleep(&vmstats.v_free_count, 0, "pfault", 0);
908 	splx(s);
909 }
910 
911 /*
912  *	vm_page_activate:
913  *
914  *	Put the specified page on the active list (if appropriate).
915  *	Ensure that act_count is at least ACT_INIT but do not otherwise
916  *	mess with it.
917  *
918  *	The page queues must be locked.
919  *	This routine may not block.
920  */
921 void
922 vm_page_activate(vm_page_t m)
923 {
924 	int s;
925 
926 	s = splvm();
927 	if (m->queue != PQ_ACTIVE) {
928 		if ((m->queue - m->pc) == PQ_CACHE)
929 			mycpu->gd_cnt.v_reactivated++;
930 
931 		vm_page_unqueue(m);
932 
933 		if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
934 			m->queue = PQ_ACTIVE;
935 			vm_page_queues[PQ_ACTIVE].lcnt++;
936 			TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
937 			if (m->act_count < ACT_INIT)
938 				m->act_count = ACT_INIT;
939 			vmstats.v_active_count++;
940 		}
941 	} else {
942 		if (m->act_count < ACT_INIT)
943 			m->act_count = ACT_INIT;
944 	}
945 
946 	splx(s);
947 }
948 
949 /*
950  *	vm_page_free_wakeup:
951  *
952  *	Helper routine for vm_page_free_toq() and vm_page_cache().  This
953  *	routine is called when a page has been added to the cache or free
954  *	queues.
955  *
956  *	This routine may not block.
957  *	This routine must be called at splvm()
958  */
959 static __inline void
960 vm_page_free_wakeup(void)
961 {
962 	/*
963 	 * if pageout daemon needs pages, then tell it that there are
964 	 * some free.
965 	 */
966 	if (vm_pageout_pages_needed &&
967 	    vmstats.v_cache_count + vmstats.v_free_count >= vmstats.v_pageout_free_min) {
968 		wakeup(&vm_pageout_pages_needed);
969 		vm_pageout_pages_needed = 0;
970 	}
971 	/*
972 	 * wakeup processes that are waiting on memory if we hit a
973 	 * high water mark. And wakeup scheduler process if we have
974 	 * lots of memory. this process will swapin processes.
975 	 */
976 	if (vm_pages_needed && !vm_page_count_min()) {
977 		vm_pages_needed = 0;
978 		wakeup(&vmstats.v_free_count);
979 	}
980 }
981 
982 /*
983  *	vm_page_free_toq:
984  *
985  *	Returns the given page to the PQ_FREE list,
986  *	disassociating it with any VM object.
987  *
988  *	Object and page must be locked prior to entry.
989  *	This routine may not block.
990  */
991 
992 void
993 vm_page_free_toq(vm_page_t m)
994 {
995 	int s;
996 	struct vpgqueues *pq;
997 	vm_object_t object = m->object;
998 
999 	s = splvm();
1000 
1001 	mycpu->gd_cnt.v_tfree++;
1002 
1003 	if (m->busy || ((m->queue - m->pc) == PQ_FREE)) {
1004 		printf(
1005 		"vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n",
1006 		    (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0,
1007 		    m->hold_count);
1008 		if ((m->queue - m->pc) == PQ_FREE)
1009 			panic("vm_page_free: freeing free page");
1010 		else
1011 			panic("vm_page_free: freeing busy page");
1012 	}
1013 
1014 	/*
1015 	 * unqueue, then remove page.  Note that we cannot destroy
1016 	 * the page here because we do not want to call the pager's
1017 	 * callback routine until after we've put the page on the
1018 	 * appropriate free queue.
1019 	 */
1020 
1021 	vm_page_unqueue_nowakeup(m);
1022 	vm_page_remove(m);
1023 
1024 	/*
1025 	 * If fictitious remove object association and
1026 	 * return, otherwise delay object association removal.
1027 	 */
1028 
1029 	if ((m->flags & PG_FICTITIOUS) != 0) {
1030 		splx(s);
1031 		return;
1032 	}
1033 
1034 	m->valid = 0;
1035 	vm_page_undirty(m);
1036 
1037 	if (m->wire_count != 0) {
1038 		if (m->wire_count > 1) {
1039 			panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx",
1040 				m->wire_count, (long)m->pindex);
1041 		}
1042 		panic("vm_page_free: freeing wired page\n");
1043 	}
1044 
1045 	/*
1046 	 * If we've exhausted the object's resident pages we want to free
1047 	 * it up.
1048 	 */
1049 
1050 	if (object &&
1051 	    (object->type == OBJT_VNODE) &&
1052 	    ((object->flags & OBJ_DEAD) == 0)
1053 	) {
1054 		struct vnode *vp = (struct vnode *)object->handle;
1055 
1056 		if (vp && VSHOULDFREE(vp))
1057 			vfree(vp);
1058 	}
1059 
1060 	/*
1061 	 * Clear the UNMANAGED flag when freeing an unmanaged page.
1062 	 */
1063 
1064 	if (m->flags & PG_UNMANAGED) {
1065 	    m->flags &= ~PG_UNMANAGED;
1066 	} else {
1067 #ifdef __alpha__
1068 	    pmap_page_is_free(m);
1069 #endif
1070 	}
1071 
1072 	if (m->hold_count != 0) {
1073 		m->flags &= ~PG_ZERO;
1074 		m->queue = PQ_HOLD;
1075 	} else
1076 		m->queue = PQ_FREE + m->pc;
1077 	pq = &vm_page_queues[m->queue];
1078 	pq->lcnt++;
1079 	++(*pq->cnt);
1080 
1081 	/*
1082 	 * Put zero'd pages on the end ( where we look for zero'd pages
1083 	 * first ) and non-zerod pages at the head.
1084 	 */
1085 
1086 	if (m->flags & PG_ZERO) {
1087 		TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
1088 		++vm_page_zero_count;
1089 	} else {
1090 		TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
1091 	}
1092 
1093 	vm_page_free_wakeup();
1094 
1095 	splx(s);
1096 }
1097 
1098 /*
1099  *	vm_page_unmanage:
1100  *
1101  * 	Prevent PV management from being done on the page.  The page is
1102  *	removed from the paging queues as if it were wired, and as a
1103  *	consequence of no longer being managed the pageout daemon will not
1104  *	touch it (since there is no way to locate the pte mappings for the
1105  *	page).  madvise() calls that mess with the pmap will also no longer
1106  *	operate on the page.
1107  *
1108  *	Beyond that the page is still reasonably 'normal'.  Freeing the page
1109  *	will clear the flag.
1110  *
1111  *	This routine is used by OBJT_PHYS objects - objects using unswappable
1112  *	physical memory as backing store rather then swap-backed memory and
1113  *	will eventually be extended to support 4MB unmanaged physical
1114  *	mappings.
1115  */
1116 
1117 void
1118 vm_page_unmanage(vm_page_t m)
1119 {
1120 	int s;
1121 
1122 	s = splvm();
1123 	if ((m->flags & PG_UNMANAGED) == 0) {
1124 		if (m->wire_count == 0)
1125 			vm_page_unqueue(m);
1126 	}
1127 	vm_page_flag_set(m, PG_UNMANAGED);
1128 	splx(s);
1129 }
1130 
1131 /*
1132  *	vm_page_wire:
1133  *
1134  *	Mark this page as wired down by yet
1135  *	another map, removing it from paging queues
1136  *	as necessary.
1137  *
1138  *	The page queues must be locked.
1139  *	This routine may not block.
1140  */
1141 void
1142 vm_page_wire(vm_page_t m)
1143 {
1144 	int s;
1145 
1146 	/*
1147 	 * Only bump the wire statistics if the page is not already wired,
1148 	 * and only unqueue the page if it is on some queue (if it is unmanaged
1149 	 * it is already off the queues).
1150 	 */
1151 	s = splvm();
1152 	if (m->wire_count == 0) {
1153 		if ((m->flags & PG_UNMANAGED) == 0)
1154 			vm_page_unqueue(m);
1155 		vmstats.v_wire_count++;
1156 	}
1157 	m->wire_count++;
1158 	KASSERT(m->wire_count != 0,
1159 	    ("vm_page_wire: wire_count overflow m=%p", m));
1160 
1161 	splx(s);
1162 	vm_page_flag_set(m, PG_MAPPED);
1163 }
1164 
1165 /*
1166  *	vm_page_unwire:
1167  *
1168  *	Release one wiring of this page, potentially
1169  *	enabling it to be paged again.
1170  *
1171  *	Many pages placed on the inactive queue should actually go
1172  *	into the cache, but it is difficult to figure out which.  What
1173  *	we do instead, if the inactive target is well met, is to put
1174  *	clean pages at the head of the inactive queue instead of the tail.
1175  *	This will cause them to be moved to the cache more quickly and
1176  *	if not actively re-referenced, freed more quickly.  If we just
1177  *	stick these pages at the end of the inactive queue, heavy filesystem
1178  *	meta-data accesses can cause an unnecessary paging load on memory bound
1179  *	processes.  This optimization causes one-time-use metadata to be
1180  *	reused more quickly.
1181  *
1182  *	BUT, if we are in a low-memory situation we have no choice but to
1183  *	put clean pages on the cache queue.
1184  *
1185  *	A number of routines use vm_page_unwire() to guarantee that the page
1186  *	will go into either the inactive or active queues, and will NEVER
1187  *	be placed in the cache - for example, just after dirtying a page.
1188  *	dirty pages in the cache are not allowed.
1189  *
1190  *	The page queues must be locked.
1191  *	This routine may not block.
1192  */
1193 void
1194 vm_page_unwire(vm_page_t m, int activate)
1195 {
1196 	int s;
1197 
1198 	s = splvm();
1199 
1200 	if (m->wire_count > 0) {
1201 		m->wire_count--;
1202 		if (m->wire_count == 0) {
1203 			vmstats.v_wire_count--;
1204 			if (m->flags & PG_UNMANAGED) {
1205 				;
1206 			} else if (activate) {
1207 				TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1208 				m->queue = PQ_ACTIVE;
1209 				vm_page_queues[PQ_ACTIVE].lcnt++;
1210 				vmstats.v_active_count++;
1211 			} else {
1212 				vm_page_flag_clear(m, PG_WINATCFLS);
1213 				TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1214 				m->queue = PQ_INACTIVE;
1215 				vm_page_queues[PQ_INACTIVE].lcnt++;
1216 				vmstats.v_inactive_count++;
1217 			}
1218 		}
1219 	} else {
1220 		panic("vm_page_unwire: invalid wire count: %d\n", m->wire_count);
1221 	}
1222 	splx(s);
1223 }
1224 
1225 
1226 /*
1227  * Move the specified page to the inactive queue.  If the page has
1228  * any associated swap, the swap is deallocated.
1229  *
1230  * Normally athead is 0 resulting in LRU operation.  athead is set
1231  * to 1 if we want this page to be 'as if it were placed in the cache',
1232  * except without unmapping it from the process address space.
1233  *
1234  * This routine may not block.
1235  */
1236 static __inline void
1237 _vm_page_deactivate(vm_page_t m, int athead)
1238 {
1239 	int s;
1240 
1241 	/*
1242 	 * Ignore if already inactive.
1243 	 */
1244 	if (m->queue == PQ_INACTIVE)
1245 		return;
1246 
1247 	s = splvm();
1248 	if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1249 		if ((m->queue - m->pc) == PQ_CACHE)
1250 			mycpu->gd_cnt.v_reactivated++;
1251 		vm_page_flag_clear(m, PG_WINATCFLS);
1252 		vm_page_unqueue(m);
1253 		if (athead)
1254 			TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1255 		else
1256 			TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1257 		m->queue = PQ_INACTIVE;
1258 		vm_page_queues[PQ_INACTIVE].lcnt++;
1259 		vmstats.v_inactive_count++;
1260 	}
1261 	splx(s);
1262 }
1263 
1264 void
1265 vm_page_deactivate(vm_page_t m)
1266 {
1267     _vm_page_deactivate(m, 0);
1268 }
1269 
1270 /*
1271  * vm_page_try_to_cache:
1272  *
1273  * Returns 0 on failure, 1 on success
1274  */
1275 int
1276 vm_page_try_to_cache(vm_page_t m)
1277 {
1278 	if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1279 	    (m->flags & (PG_BUSY|PG_UNMANAGED))) {
1280 		return(0);
1281 	}
1282 	vm_page_test_dirty(m);
1283 	if (m->dirty)
1284 		return(0);
1285 	vm_page_cache(m);
1286 	return(1);
1287 }
1288 
1289 /*
1290  * vm_page_try_to_free()
1291  *
1292  *	Attempt to free the page.  If we cannot free it, we do nothing.
1293  *	1 is returned on success, 0 on failure.
1294  */
1295 
1296 int
1297 vm_page_try_to_free(vm_page_t m)
1298 {
1299 	if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1300 	    (m->flags & (PG_BUSY|PG_UNMANAGED))) {
1301 		return(0);
1302 	}
1303 	vm_page_test_dirty(m);
1304 	if (m->dirty)
1305 		return(0);
1306 	vm_page_busy(m);
1307 	vm_page_protect(m, VM_PROT_NONE);
1308 	vm_page_free(m);
1309 	return(1);
1310 }
1311 
1312 
1313 /*
1314  * vm_page_cache
1315  *
1316  * Put the specified page onto the page cache queue (if appropriate).
1317  *
1318  * This routine may not block.
1319  */
1320 void
1321 vm_page_cache(vm_page_t m)
1322 {
1323 	int s;
1324 
1325 	if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || m->wire_count) {
1326 		printf("vm_page_cache: attempting to cache busy page\n");
1327 		return;
1328 	}
1329 	if ((m->queue - m->pc) == PQ_CACHE)
1330 		return;
1331 
1332 	/*
1333 	 * Remove all pmaps and indicate that the page is not
1334 	 * writeable or mapped.
1335 	 */
1336 
1337 	vm_page_protect(m, VM_PROT_NONE);
1338 	if (m->dirty != 0) {
1339 		panic("vm_page_cache: caching a dirty page, pindex: %ld",
1340 			(long)m->pindex);
1341 	}
1342 	s = splvm();
1343 	vm_page_unqueue_nowakeup(m);
1344 	m->queue = PQ_CACHE + m->pc;
1345 	vm_page_queues[m->queue].lcnt++;
1346 	TAILQ_INSERT_TAIL(&vm_page_queues[m->queue].pl, m, pageq);
1347 	vmstats.v_cache_count++;
1348 	vm_page_free_wakeup();
1349 	splx(s);
1350 }
1351 
1352 /*
1353  * vm_page_dontneed
1354  *
1355  *	Cache, deactivate, or do nothing as appropriate.  This routine
1356  *	is typically used by madvise() MADV_DONTNEED.
1357  *
1358  *	Generally speaking we want to move the page into the cache so
1359  *	it gets reused quickly.  However, this can result in a silly syndrome
1360  *	due to the page recycling too quickly.  Small objects will not be
1361  *	fully cached.  On the otherhand, if we move the page to the inactive
1362  *	queue we wind up with a problem whereby very large objects
1363  *	unnecessarily blow away our inactive and cache queues.
1364  *
1365  *	The solution is to move the pages based on a fixed weighting.  We
1366  *	either leave them alone, deactivate them, or move them to the cache,
1367  *	where moving them to the cache has the highest weighting.
1368  *	By forcing some pages into other queues we eventually force the
1369  *	system to balance the queues, potentially recovering other unrelated
1370  *	space from active.  The idea is to not force this to happen too
1371  *	often.
1372  */
1373 
1374 void
1375 vm_page_dontneed(vm_page_t m)
1376 {
1377 	static int dnweight;
1378 	int dnw;
1379 	int head;
1380 
1381 	dnw = ++dnweight;
1382 
1383 	/*
1384 	 * occassionally leave the page alone
1385 	 */
1386 
1387 	if ((dnw & 0x01F0) == 0 ||
1388 	    m->queue == PQ_INACTIVE ||
1389 	    m->queue - m->pc == PQ_CACHE
1390 	) {
1391 		if (m->act_count >= ACT_INIT)
1392 			--m->act_count;
1393 		return;
1394 	}
1395 
1396 	if (m->dirty == 0)
1397 		vm_page_test_dirty(m);
1398 
1399 	if (m->dirty || (dnw & 0x0070) == 0) {
1400 		/*
1401 		 * Deactivate the page 3 times out of 32.
1402 		 */
1403 		head = 0;
1404 	} else {
1405 		/*
1406 		 * Cache the page 28 times out of every 32.  Note that
1407 		 * the page is deactivated instead of cached, but placed
1408 		 * at the head of the queue instead of the tail.
1409 		 */
1410 		head = 1;
1411 	}
1412 	_vm_page_deactivate(m, head);
1413 }
1414 
1415 /*
1416  * Grab a page, waiting until we are waken up due to the page
1417  * changing state.  We keep on waiting, if the page continues
1418  * to be in the object.  If the page doesn't exist, allocate it.
1419  *
1420  * This routine may block.
1421  */
1422 vm_page_t
1423 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
1424 {
1425 
1426 	vm_page_t m;
1427 	int s, generation;
1428 
1429 retrylookup:
1430 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
1431 		if (m->busy || (m->flags & PG_BUSY)) {
1432 			generation = object->generation;
1433 
1434 			s = splvm();
1435 			while ((object->generation == generation) &&
1436 					(m->busy || (m->flags & PG_BUSY))) {
1437 				vm_page_flag_set(m, PG_WANTED | PG_REFERENCED);
1438 				tsleep(m, 0, "pgrbwt", 0);
1439 				if ((allocflags & VM_ALLOC_RETRY) == 0) {
1440 					splx(s);
1441 					return NULL;
1442 				}
1443 			}
1444 			splx(s);
1445 			goto retrylookup;
1446 		} else {
1447 			vm_page_busy(m);
1448 			return m;
1449 		}
1450 	}
1451 
1452 	m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY);
1453 	if (m == NULL) {
1454 		VM_WAIT;
1455 		if ((allocflags & VM_ALLOC_RETRY) == 0)
1456 			return NULL;
1457 		goto retrylookup;
1458 	}
1459 
1460 	return m;
1461 }
1462 
1463 /*
1464  * Mapping function for valid bits or for dirty bits in
1465  * a page.  May not block.
1466  *
1467  * Inputs are required to range within a page.
1468  */
1469 
1470 __inline int
1471 vm_page_bits(int base, int size)
1472 {
1473 	int first_bit;
1474 	int last_bit;
1475 
1476 	KASSERT(
1477 	    base + size <= PAGE_SIZE,
1478 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
1479 	);
1480 
1481 	if (size == 0)		/* handle degenerate case */
1482 		return(0);
1483 
1484 	first_bit = base >> DEV_BSHIFT;
1485 	last_bit = (base + size - 1) >> DEV_BSHIFT;
1486 
1487 	return ((2 << last_bit) - (1 << first_bit));
1488 }
1489 
1490 /*
1491  *	vm_page_set_validclean:
1492  *
1493  *	Sets portions of a page valid and clean.  The arguments are expected
1494  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
1495  *	of any partial chunks touched by the range.  The invalid portion of
1496  *	such chunks will be zero'd.
1497  *
1498  *	This routine may not block.
1499  *
1500  *	(base + size) must be less then or equal to PAGE_SIZE.
1501  */
1502 void
1503 vm_page_set_validclean(vm_page_t m, int base, 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.  We also
1549 	 * use this opportunity to clear the PG_NOSYNC flag.  If a process
1550 	 * takes a write fault on a MAP_NOSYNC memory area the flag will
1551 	 * be set again.
1552 	 *
1553 	 * We set valid bits inclusive of any overlap, but we can only
1554 	 * clear dirty bits for DEV_BSIZE chunks that are fully within
1555 	 * the range.
1556 	 */
1557 
1558 	pagebits = vm_page_bits(base, size);
1559 	m->valid |= pagebits;
1560 #if 0	/* NOT YET */
1561 	if ((frag = base & (DEV_BSIZE - 1)) != 0) {
1562 		frag = DEV_BSIZE - frag;
1563 		base += frag;
1564 		size -= frag;
1565 		if (size < 0)
1566 		    size = 0;
1567 	}
1568 	pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
1569 #endif
1570 	m->dirty &= ~pagebits;
1571 	if (base == 0 && size == PAGE_SIZE) {
1572 		pmap_clear_modify(m);
1573 		vm_page_flag_clear(m, PG_NOSYNC);
1574 	}
1575 }
1576 
1577 #if 0
1578 
1579 void
1580 vm_page_set_dirty(vm_page_t m, int base, int size)
1581 {
1582 	m->dirty |= vm_page_bits(base, size);
1583 }
1584 
1585 #endif
1586 
1587 void
1588 vm_page_clear_dirty(vm_page_t m, int base, int size)
1589 {
1590 	m->dirty &= ~vm_page_bits(base, size);
1591 }
1592 
1593 /*
1594  *	vm_page_set_invalid:
1595  *
1596  *	Invalidates DEV_BSIZE'd chunks within a page.  Both the
1597  *	valid and dirty bits for the effected areas are cleared.
1598  *
1599  *	May not block.
1600  */
1601 void
1602 vm_page_set_invalid(vm_page_t m, int base, int size)
1603 {
1604 	int bits;
1605 
1606 	bits = vm_page_bits(base, size);
1607 	m->valid &= ~bits;
1608 	m->dirty &= ~bits;
1609 	m->object->generation++;
1610 }
1611 
1612 /*
1613  * vm_page_zero_invalid()
1614  *
1615  *	The kernel assumes that the invalid portions of a page contain
1616  *	garbage, but such pages can be mapped into memory by user code.
1617  *	When this occurs, we must zero out the non-valid portions of the
1618  *	page so user code sees what it expects.
1619  *
1620  *	Pages are most often semi-valid when the end of a file is mapped
1621  *	into memory and the file's size is not page aligned.
1622  */
1623 
1624 void
1625 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
1626 {
1627 	int b;
1628 	int i;
1629 
1630 	/*
1631 	 * Scan the valid bits looking for invalid sections that
1632 	 * must be zerod.  Invalid sub-DEV_BSIZE'd areas ( where the
1633 	 * valid bit may be set ) have already been zerod by
1634 	 * vm_page_set_validclean().
1635 	 */
1636 
1637 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
1638 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
1639 		    (m->valid & (1 << i))
1640 		) {
1641 			if (i > b) {
1642 				pmap_zero_page_area(
1643 				    VM_PAGE_TO_PHYS(m),
1644 				    b << DEV_BSHIFT,
1645 				    (i - b) << DEV_BSHIFT
1646 				);
1647 			}
1648 			b = i + 1;
1649 		}
1650 	}
1651 
1652 	/*
1653 	 * setvalid is TRUE when we can safely set the zero'd areas
1654 	 * as being valid.  We can do this if there are no cache consistency
1655 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
1656 	 */
1657 
1658 	if (setvalid)
1659 		m->valid = VM_PAGE_BITS_ALL;
1660 }
1661 
1662 /*
1663  *	vm_page_is_valid:
1664  *
1665  *	Is (partial) page valid?  Note that the case where size == 0
1666  *	will return FALSE in the degenerate case where the page is
1667  *	entirely invalid, and TRUE otherwise.
1668  *
1669  *	May not block.
1670  */
1671 
1672 int
1673 vm_page_is_valid(vm_page_t m, int base, int size)
1674 {
1675 	int bits = vm_page_bits(base, size);
1676 
1677 	if (m->valid && ((m->valid & bits) == bits))
1678 		return 1;
1679 	else
1680 		return 0;
1681 }
1682 
1683 /*
1684  * update dirty bits from pmap/mmu.  May not block.
1685  */
1686 
1687 void
1688 vm_page_test_dirty(vm_page_t m)
1689 {
1690 	if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) {
1691 		vm_page_dirty(m);
1692 	}
1693 }
1694 
1695 /*
1696  * This interface is for merging with malloc() someday.
1697  * Even if we never implement compaction so that contiguous allocation
1698  * works after initialization time, malloc()'s data structures are good
1699  * for statistics and for allocations of less than a page.
1700  */
1701 void *
1702 contigmalloc1(
1703 	unsigned long size,	/* should be size_t here and for malloc() */
1704 	struct malloc_type *type,
1705 	int flags,
1706 	unsigned long low,
1707 	unsigned long high,
1708 	unsigned long alignment,
1709 	unsigned long boundary,
1710 	vm_map_t map)
1711 {
1712 	int i, s, start;
1713 	vm_offset_t addr, phys, tmp_addr;
1714 	int pass;
1715 	vm_page_t pga = vm_page_array;
1716 
1717 	size = round_page(size);
1718 	if (size == 0)
1719 		panic("contigmalloc1: size must not be 0");
1720 	if ((alignment & (alignment - 1)) != 0)
1721 		panic("contigmalloc1: alignment must be a power of 2");
1722 	if ((boundary & (boundary - 1)) != 0)
1723 		panic("contigmalloc1: boundary must be a power of 2");
1724 
1725 	start = 0;
1726 	for (pass = 0; pass <= 1; pass++) {
1727 		s = splvm();
1728 again:
1729 		/*
1730 		 * Find first page in array that is free, within range, aligned, and
1731 		 * such that the boundary won't be crossed.
1732 		 */
1733 		for (i = start; i < vmstats.v_page_count; i++) {
1734 			int pqtype;
1735 			phys = VM_PAGE_TO_PHYS(&pga[i]);
1736 			pqtype = pga[i].queue - pga[i].pc;
1737 			if (((pqtype == PQ_FREE) || (pqtype == PQ_CACHE)) &&
1738 			    (phys >= low) && (phys < high) &&
1739 			    ((phys & (alignment - 1)) == 0) &&
1740 			    (((phys ^ (phys + size - 1)) & ~(boundary - 1)) == 0))
1741 				break;
1742 		}
1743 
1744 		/*
1745 		 * If the above failed or we will exceed the upper bound, fail.
1746 		 */
1747 		if ((i == vmstats.v_page_count) ||
1748 			((VM_PAGE_TO_PHYS(&pga[i]) + size) > high)) {
1749 			vm_page_t m, next;
1750 
1751 again1:
1752 			for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl);
1753 				m != NULL;
1754 				m = next) {
1755 
1756 				KASSERT(m->queue == PQ_INACTIVE,
1757 					("contigmalloc1: page %p is not PQ_INACTIVE", m));
1758 
1759 				next = TAILQ_NEXT(m, pageq);
1760 				if (vm_page_sleep_busy(m, TRUE, "vpctw0"))
1761 					goto again1;
1762 				vm_page_test_dirty(m);
1763 				if (m->dirty) {
1764 					if (m->object->type == OBJT_VNODE) {
1765 						vn_lock(m->object->handle, LK_EXCLUSIVE | LK_RETRY, curthread);
1766 						vm_object_page_clean(m->object, 0, 0, OBJPC_SYNC);
1767 						VOP_UNLOCK(m->object->handle, 0, curthread);
1768 						goto again1;
1769 					} else if (m->object->type == OBJT_SWAP ||
1770 								m->object->type == OBJT_DEFAULT) {
1771 						vm_pageout_flush(&m, 1, 0);
1772 						goto again1;
1773 					}
1774 				}
1775 				if ((m->dirty == 0) && (m->busy == 0) && (m->hold_count == 0))
1776 					vm_page_cache(m);
1777 			}
1778 
1779 			for (m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1780 				m != NULL;
1781 				m = next) {
1782 
1783 				KASSERT(m->queue == PQ_ACTIVE,
1784 					("contigmalloc1: page %p is not PQ_ACTIVE", m));
1785 
1786 				next = TAILQ_NEXT(m, pageq);
1787 				if (vm_page_sleep_busy(m, TRUE, "vpctw1"))
1788 					goto again1;
1789 				vm_page_test_dirty(m);
1790 				if (m->dirty) {
1791 					if (m->object->type == OBJT_VNODE) {
1792 						vn_lock(m->object->handle, LK_EXCLUSIVE | LK_RETRY, curthread);
1793 						vm_object_page_clean(m->object, 0, 0, OBJPC_SYNC);
1794 						VOP_UNLOCK(m->object->handle, 0, curthread);
1795 						goto again1;
1796 					} else if (m->object->type == OBJT_SWAP ||
1797 								m->object->type == OBJT_DEFAULT) {
1798 						vm_pageout_flush(&m, 1, 0);
1799 						goto again1;
1800 					}
1801 				}
1802 				if ((m->dirty == 0) && (m->busy == 0) && (m->hold_count == 0))
1803 					vm_page_cache(m);
1804 			}
1805 
1806 			splx(s);
1807 			continue;
1808 		}
1809 		start = i;
1810 
1811 		/*
1812 		 * Check successive pages for contiguous and free.
1813 		 */
1814 		for (i = start + 1; i < (start + size / PAGE_SIZE); i++) {
1815 			int pqtype;
1816 			pqtype = pga[i].queue - pga[i].pc;
1817 			if ((VM_PAGE_TO_PHYS(&pga[i]) !=
1818 			    (VM_PAGE_TO_PHYS(&pga[i - 1]) + PAGE_SIZE)) ||
1819 			    ((pqtype != PQ_FREE) && (pqtype != PQ_CACHE))) {
1820 				start++;
1821 				goto again;
1822 			}
1823 		}
1824 
1825 		for (i = start; i < (start + size / PAGE_SIZE); i++) {
1826 			int pqtype;
1827 			vm_page_t m = &pga[i];
1828 
1829 			pqtype = m->queue - m->pc;
1830 			if (pqtype == PQ_CACHE) {
1831 				vm_page_busy(m);
1832 				vm_page_free(m);
1833 			}
1834 			vm_page_unqueue_nowakeup(m);
1835 			m->valid = VM_PAGE_BITS_ALL;
1836 			if (m->flags & PG_ZERO)
1837 				vm_page_zero_count--;
1838 			m->flags = 0;
1839 			KASSERT(m->dirty == 0, ("contigmalloc1: page %p was dirty", m));
1840 			m->wire_count = 0;
1841 			m->busy = 0;
1842 			m->object = NULL;
1843 		}
1844 
1845 		/*
1846 		 * We've found a contiguous chunk that meets are requirements.
1847 		 * Allocate kernel VM, unfree and assign the physical pages to it and
1848 		 * return kernel VM pointer.
1849 		 */
1850 		vm_map_lock(map);
1851 		if (vm_map_findspace(map, vm_map_min(map), size, &addr) !=
1852 		    KERN_SUCCESS) {
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 			vm_map_unlock(map);
1859 			splx(s);
1860 			return (NULL);
1861 		}
1862 		vm_object_reference(kernel_object);
1863 		vm_map_insert(map, kernel_object, addr - VM_MIN_KERNEL_ADDRESS,
1864 		    addr, addr + size, VM_PROT_ALL, VM_PROT_ALL, 0);
1865 		vm_map_unlock(map);
1866 
1867 		tmp_addr = addr;
1868 		for (i = start; i < (start + size / PAGE_SIZE); i++) {
1869 			vm_page_t m = &pga[i];
1870 			vm_page_insert(m, kernel_object,
1871 				OFF_TO_IDX(tmp_addr - VM_MIN_KERNEL_ADDRESS));
1872 			tmp_addr += PAGE_SIZE;
1873 		}
1874 		vm_map_pageable(map, addr, addr + size, FALSE);
1875 
1876 		splx(s);
1877 		return ((void *)addr);
1878 	}
1879 	return NULL;
1880 }
1881 
1882 void *
1883 contigmalloc(
1884 	unsigned long size,	/* should be size_t here and for malloc() */
1885 	struct malloc_type *type,
1886 	int flags,
1887 	unsigned long low,
1888 	unsigned long high,
1889 	unsigned long alignment,
1890 	unsigned long boundary)
1891 {
1892 	return contigmalloc1(size, type, flags, low, high, alignment, boundary,
1893 			     kernel_map);
1894 }
1895 
1896 void
1897 contigfree(void *addr, unsigned long size, struct malloc_type *type)
1898 {
1899 	kmem_free(kernel_map, (vm_offset_t)addr, size);
1900 }
1901 
1902 vm_offset_t
1903 vm_page_alloc_contig(
1904 	vm_offset_t size,
1905 	vm_offset_t low,
1906 	vm_offset_t high,
1907 	vm_offset_t alignment)
1908 {
1909 	return ((vm_offset_t)contigmalloc1(size, M_DEVBUF, M_NOWAIT, low, high,
1910 					  alignment, 0ul, kernel_map));
1911 }
1912 
1913 #include "opt_ddb.h"
1914 #ifdef DDB
1915 #include <sys/kernel.h>
1916 
1917 #include <ddb/ddb.h>
1918 
1919 DB_SHOW_COMMAND(page, vm_page_print_page_info)
1920 {
1921 	db_printf("vmstats.v_free_count: %d\n", vmstats.v_free_count);
1922 	db_printf("vmstats.v_cache_count: %d\n", vmstats.v_cache_count);
1923 	db_printf("vmstats.v_inactive_count: %d\n", vmstats.v_inactive_count);
1924 	db_printf("vmstats.v_active_count: %d\n", vmstats.v_active_count);
1925 	db_printf("vmstats.v_wire_count: %d\n", vmstats.v_wire_count);
1926 	db_printf("vmstats.v_free_reserved: %d\n", vmstats.v_free_reserved);
1927 	db_printf("vmstats.v_free_min: %d\n", vmstats.v_free_min);
1928 	db_printf("vmstats.v_free_target: %d\n", vmstats.v_free_target);
1929 	db_printf("vmstats.v_cache_min: %d\n", vmstats.v_cache_min);
1930 	db_printf("vmstats.v_inactive_target: %d\n", vmstats.v_inactive_target);
1931 }
1932 
1933 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
1934 {
1935 	int i;
1936 	db_printf("PQ_FREE:");
1937 	for(i=0;i<PQ_L2_SIZE;i++) {
1938 		db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt);
1939 	}
1940 	db_printf("\n");
1941 
1942 	db_printf("PQ_CACHE:");
1943 	for(i=0;i<PQ_L2_SIZE;i++) {
1944 		db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt);
1945 	}
1946 	db_printf("\n");
1947 
1948 	db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n",
1949 		vm_page_queues[PQ_ACTIVE].lcnt,
1950 		vm_page_queues[PQ_INACTIVE].lcnt);
1951 }
1952 #endif /* DDB */
1953