xref: /dragonfly/sys/vm/vm_page.c (revision ab709bfb)
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 1.40 2008/08/25 17:01:42 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  * Resident memory management module.  The module manipulates 'VM pages'.
69  * A VM page is the core building block for memory management.
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 void vm_page_free_wakeup(void);
94 static vm_page_t vm_page_select_cache(vm_object_t, vm_pindex_t);
95 static vm_page_t _vm_page_list_find2(int basequeue, int index);
96 
97 struct vpgqueues vm_page_queues[PQ_COUNT]; /* Array of tailq lists */
98 
99 #define ASSERT_IN_CRIT_SECTION()	KKASSERT(crit_test(curthread));
100 
101 RB_GENERATE2(vm_page_rb_tree, vm_page, rb_entry, rb_vm_page_compare,
102 	     vm_pindex_t, pindex);
103 
104 static void
105 vm_page_queue_init(void)
106 {
107 	int i;
108 
109 	for (i = 0; i < PQ_L2_SIZE; i++)
110 		vm_page_queues[PQ_FREE+i].cnt = &vmstats.v_free_count;
111 	for (i = 0; i < PQ_L2_SIZE; i++)
112 		vm_page_queues[PQ_CACHE+i].cnt = &vmstats.v_cache_count;
113 
114 	vm_page_queues[PQ_INACTIVE].cnt = &vmstats.v_inactive_count;
115 	vm_page_queues[PQ_ACTIVE].cnt = &vmstats.v_active_count;
116 	vm_page_queues[PQ_HOLD].cnt = &vmstats.v_active_count;
117 	/* PQ_NONE has no queue */
118 
119 	for (i = 0; i < PQ_COUNT; i++)
120 		TAILQ_INIT(&vm_page_queues[i].pl);
121 }
122 
123 /*
124  * note: place in initialized data section?  Is this necessary?
125  */
126 long first_page = 0;
127 int vm_page_array_size = 0;
128 int vm_page_zero_count = 0;
129 vm_page_t vm_page_array = 0;
130 
131 /*
132  * (low level boot)
133  *
134  * Sets the page size, perhaps based upon the memory size.
135  * Must be called before any use of page-size dependent functions.
136  */
137 void
138 vm_set_page_size(void)
139 {
140 	if (vmstats.v_page_size == 0)
141 		vmstats.v_page_size = PAGE_SIZE;
142 	if (((vmstats.v_page_size - 1) & vmstats.v_page_size) != 0)
143 		panic("vm_set_page_size: page size not a power of two");
144 }
145 
146 /*
147  * (low level boot)
148  *
149  * Add a new page to the freelist for use by the system.  New pages
150  * are added to both the head and tail of the associated free page
151  * queue in a bottom-up fashion, so both zero'd and non-zero'd page
152  * requests pull 'recent' adds (higher physical addresses) first.
153  *
154  * Must be called in a critical section.
155  */
156 vm_page_t
157 vm_add_new_page(vm_paddr_t pa)
158 {
159 	struct vpgqueues *vpq;
160 	vm_page_t m;
161 
162 	++vmstats.v_page_count;
163 	++vmstats.v_free_count;
164 	m = PHYS_TO_VM_PAGE(pa);
165 	m->phys_addr = pa;
166 	m->flags = 0;
167 	m->pc = (pa >> PAGE_SHIFT) & PQ_L2_MASK;
168 	m->queue = m->pc + PQ_FREE;
169 	KKASSERT(m->dirty == 0);
170 
171 	vpq = &vm_page_queues[m->queue];
172 	if (vpq->flipflop)
173 		TAILQ_INSERT_TAIL(&vpq->pl, m, pageq);
174 	else
175 		TAILQ_INSERT_HEAD(&vpq->pl, m, pageq);
176 	vpq->flipflop = 1 - vpq->flipflop;
177 
178 	vm_page_queues[m->queue].lcnt++;
179 	return (m);
180 }
181 
182 /*
183  * (low level boot)
184  *
185  * Initializes the resident memory module.
186  *
187  * Allocates memory for the page cells, and for the object/offset-to-page
188  * hash table headers.  Each page cell is initialized and placed on the
189  * free list.
190  *
191  * starta/enda represents the range of physical memory addresses available
192  * for use (skipping memory already used by the kernel), subject to
193  * phys_avail[].  Note that phys_avail[] has already mapped out memory
194  * already in use by the kernel.
195  */
196 vm_offset_t
197 vm_page_startup(vm_offset_t vaddr)
198 {
199 	vm_offset_t mapped;
200 	vm_size_t npages;
201 	vm_paddr_t page_range;
202 	vm_paddr_t new_end;
203 	int i;
204 	vm_paddr_t pa;
205 	int nblocks;
206 	vm_paddr_t last_pa;
207 	vm_paddr_t end;
208 	vm_paddr_t biggestone, biggestsize;
209 	vm_paddr_t total;
210 
211 	total = 0;
212 	biggestsize = 0;
213 	biggestone = 0;
214 	nblocks = 0;
215 	vaddr = round_page(vaddr);
216 
217 	for (i = 0; phys_avail[i + 1]; i += 2) {
218 		phys_avail[i] = round_page(phys_avail[i]);
219 		phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
220 	}
221 
222 	for (i = 0; phys_avail[i + 1]; i += 2) {
223 		vm_paddr_t size = phys_avail[i + 1] - phys_avail[i];
224 
225 		if (size > biggestsize) {
226 			biggestone = i;
227 			biggestsize = size;
228 		}
229 		++nblocks;
230 		total += size;
231 	}
232 
233 	end = phys_avail[biggestone+1];
234 	end = trunc_page(end);
235 
236 	/*
237 	 * Initialize the queue headers for the free queue, the active queue
238 	 * and the inactive queue.
239 	 */
240 
241 	vm_page_queue_init();
242 
243 	/*
244 	 * Compute the number of pages of memory that will be available for
245 	 * use (taking into account the overhead of a page structure per
246 	 * page).
247 	 */
248 	first_page = phys_avail[0] / PAGE_SIZE;
249 	page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page;
250 	npages = (total - (page_range * sizeof(struct vm_page))) / PAGE_SIZE;
251 
252 	/*
253 	 * Initialize the mem entry structures now, and put them in the free
254 	 * queue.
255 	 */
256 	vm_page_array = (vm_page_t) vaddr;
257 	mapped = vaddr;
258 
259 	/*
260 	 * Validate these addresses.
261 	 */
262 	new_end = trunc_page(end - page_range * sizeof(struct vm_page));
263 	mapped = pmap_map(mapped, new_end, end,
264 	    VM_PROT_READ | VM_PROT_WRITE);
265 
266 	/*
267 	 * Clear all of the page structures
268 	 */
269 	bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
270 	vm_page_array_size = page_range;
271 
272 	/*
273 	 * Construct the free queue(s) in ascending order (by physical
274 	 * address) so that the first 16MB of physical memory is allocated
275 	 * last rather than first.  On large-memory machines, this avoids
276 	 * the exhaustion of low physical memory before isa_dmainit has run.
277 	 */
278 	vmstats.v_page_count = 0;
279 	vmstats.v_free_count = 0;
280 	for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) {
281 		pa = phys_avail[i];
282 		if (i == biggestone)
283 			last_pa = new_end;
284 		else
285 			last_pa = phys_avail[i + 1];
286 		while (pa < last_pa && npages-- > 0) {
287 			vm_add_new_page(pa);
288 			pa += PAGE_SIZE;
289 		}
290 	}
291 	return (mapped);
292 }
293 
294 /*
295  * Scan comparison function for Red-Black tree scans.  An inclusive
296  * (start,end) is expected.  Other fields are not used.
297  */
298 int
299 rb_vm_page_scancmp(struct vm_page *p, void *data)
300 {
301 	struct rb_vm_page_scan_info *info = data;
302 
303 	if (p->pindex < info->start_pindex)
304 		return(-1);
305 	if (p->pindex > info->end_pindex)
306 		return(1);
307 	return(0);
308 }
309 
310 int
311 rb_vm_page_compare(struct vm_page *p1, struct vm_page *p2)
312 {
313 	if (p1->pindex < p2->pindex)
314 		return(-1);
315 	if (p1->pindex > p2->pindex)
316 		return(1);
317 	return(0);
318 }
319 
320 /*
321  * The opposite of vm_page_hold().  A page can be freed while being held,
322  * which places it on the PQ_HOLD queue.  We must call vm_page_free_toq()
323  * in this case to actually free it once the hold count drops to 0.
324  *
325  * This routine must be called at splvm().
326  */
327 void
328 vm_page_unhold(vm_page_t mem)
329 {
330 	--mem->hold_count;
331 	KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!"));
332 	if (mem->hold_count == 0 && mem->queue == PQ_HOLD) {
333 		vm_page_busy(mem);
334 		vm_page_free_toq(mem);
335 	}
336 }
337 
338 /*
339  * Inserts the given mem entry into the object and object list.
340  *
341  * The pagetables are not updated but will presumably fault the page
342  * in if necessary, or if a kernel page the caller will at some point
343  * enter the page into the kernel's pmap.  We are not allowed to block
344  * here so we *can't* do this anyway.
345  *
346  * This routine may not block.
347  * This routine must be called with a critical section held.
348  */
349 void
350 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
351 {
352 	ASSERT_IN_CRIT_SECTION();
353 	if (m->object != NULL)
354 		panic("vm_page_insert: already inserted");
355 
356 	/*
357 	 * Record the object/offset pair in this page
358 	 */
359 	m->object = object;
360 	m->pindex = pindex;
361 
362 	/*
363 	 * Insert it into the object.
364 	 */
365 	vm_page_rb_tree_RB_INSERT(&object->rb_memq, m);
366 	object->generation++;
367 
368 	/*
369 	 * show that the object has one more resident page.
370 	 */
371 	object->resident_page_count++;
372 
373 	/*
374 	 * Since we are inserting a new and possibly dirty page,
375 	 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags.
376 	 */
377 	if ((m->valid & m->dirty) || (m->flags & PG_WRITEABLE))
378 		vm_object_set_writeable_dirty(object);
379 }
380 
381 /*
382  * Removes the given vm_page_t from the global (object,index) hash table
383  * and from the object's memq.
384  *
385  * The underlying pmap entry (if any) is NOT removed here.
386  * This routine may not block.
387  *
388  * The page must be BUSY and will remain BUSY on return.  No spl needs to be
389  * held on call to this routine.
390  *
391  * note: FreeBSD side effect was to unbusy the page on return.  We leave
392  * it busy.
393  */
394 void
395 vm_page_remove(vm_page_t m)
396 {
397 	vm_object_t object;
398 
399 	crit_enter();
400 	if (m->object == NULL) {
401 		crit_exit();
402 		return;
403 	}
404 
405 	if ((m->flags & PG_BUSY) == 0)
406 		panic("vm_page_remove: page not busy");
407 
408 	object = m->object;
409 
410 	/*
411 	 * Remove the page from the object and update the object.
412 	 */
413 	vm_page_rb_tree_RB_REMOVE(&object->rb_memq, m);
414 	object->resident_page_count--;
415 	object->generation++;
416 	m->object = NULL;
417 
418 	crit_exit();
419 }
420 
421 /*
422  * Locate and return the page at (object, pindex), or NULL if the
423  * page could not be found.
424  *
425  * This routine will operate properly without spl protection, but
426  * the returned page could be in flux if it is busy.  Because an
427  * interrupt can race a caller's busy check (unbusying and freeing the
428  * page we return before the caller is able to check the busy bit),
429  * the caller should generally call this routine with a critical
430  * section held.
431  *
432  * Callers may call this routine without spl protection if they know
433  * 'for sure' that the page will not be ripped out from under them
434  * by an interrupt.
435  */
436 vm_page_t
437 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
438 {
439 	vm_page_t m;
440 
441 	/*
442 	 * Search the hash table for this object/offset pair
443 	 */
444 	crit_enter();
445 	m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
446 	crit_exit();
447 	KKASSERT(m == NULL || (m->object == object && m->pindex == pindex));
448 	return(m);
449 }
450 
451 /*
452  * vm_page_rename()
453  *
454  * Move the given memory entry from its current object to the specified
455  * target object/offset.
456  *
457  * The object must be locked.
458  * This routine may not block.
459  *
460  * Note: This routine will raise itself to splvm(), the caller need not.
461  *
462  * Note: Swap associated with the page must be invalidated by the move.  We
463  *       have to do this for several reasons:  (1) we aren't freeing the
464  *       page, (2) we are dirtying the page, (3) the VM system is probably
465  *       moving the page from object A to B, and will then later move
466  *       the backing store from A to B and we can't have a conflict.
467  *
468  * Note: We *always* dirty the page.  It is necessary both for the
469  *       fact that we moved it, and because we may be invalidating
470  *	 swap.  If the page is on the cache, we have to deactivate it
471  *	 or vm_page_dirty() will panic.  Dirty pages are not allowed
472  *	 on the cache.
473  */
474 void
475 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
476 {
477 	crit_enter();
478 	vm_page_remove(m);
479 	vm_page_insert(m, new_object, new_pindex);
480 	if (m->queue - m->pc == PQ_CACHE)
481 		vm_page_deactivate(m);
482 	vm_page_dirty(m);
483 	vm_page_wakeup(m);
484 	crit_exit();
485 }
486 
487 /*
488  * vm_page_unqueue() without any wakeup.  This routine is used when a page
489  * is being moved between queues or otherwise is to remain BUSYied by the
490  * caller.
491  *
492  * This routine must be called at splhigh().
493  * This routine may not block.
494  */
495 void
496 vm_page_unqueue_nowakeup(vm_page_t m)
497 {
498 	int queue = m->queue;
499 	struct vpgqueues *pq;
500 
501 	if (queue != PQ_NONE) {
502 		pq = &vm_page_queues[queue];
503 		m->queue = PQ_NONE;
504 		TAILQ_REMOVE(&pq->pl, m, pageq);
505 		(*pq->cnt)--;
506 		pq->lcnt--;
507 	}
508 }
509 
510 /*
511  * vm_page_unqueue() - Remove a page from its queue, wakeup the pagedemon
512  * if necessary.
513  *
514  * This routine must be called at splhigh().
515  * This routine may not block.
516  */
517 void
518 vm_page_unqueue(vm_page_t m)
519 {
520 	int queue = m->queue;
521 	struct vpgqueues *pq;
522 
523 	if (queue != PQ_NONE) {
524 		m->queue = PQ_NONE;
525 		pq = &vm_page_queues[queue];
526 		TAILQ_REMOVE(&pq->pl, m, pageq);
527 		(*pq->cnt)--;
528 		pq->lcnt--;
529 		if ((queue - m->pc) == PQ_CACHE || (queue - m->pc) == PQ_FREE)
530 			pagedaemon_wakeup();
531 	}
532 }
533 
534 /*
535  * vm_page_list_find()
536  *
537  * Find a page on the specified queue with color optimization.
538  *
539  * The page coloring optimization attempts to locate a page that does
540  * not overload other nearby pages in the object in the cpu's L1 or L2
541  * caches.  We need this optimization because cpu caches tend to be
542  * physical caches, while object spaces tend to be virtual.
543  *
544  * This routine must be called at splvm().
545  * This routine may not block.
546  *
547  * Note that this routine is carefully inlined.  A non-inlined version
548  * is available for outside callers but the only critical path is
549  * from within this source file.
550  */
551 static __inline
552 vm_page_t
553 _vm_page_list_find(int basequeue, int index, boolean_t prefer_zero)
554 {
555 	vm_page_t m;
556 
557 	if (prefer_zero)
558 		m = TAILQ_LAST(&vm_page_queues[basequeue+index].pl, pglist);
559 	else
560 		m = TAILQ_FIRST(&vm_page_queues[basequeue+index].pl);
561 	if (m == NULL)
562 		m = _vm_page_list_find2(basequeue, index);
563 	return(m);
564 }
565 
566 static vm_page_t
567 _vm_page_list_find2(int basequeue, int index)
568 {
569 	int i;
570 	vm_page_t m = NULL;
571 	struct vpgqueues *pq;
572 
573 	pq = &vm_page_queues[basequeue];
574 
575 	/*
576 	 * Note that for the first loop, index+i and index-i wind up at the
577 	 * same place.  Even though this is not totally optimal, we've already
578 	 * blown it by missing the cache case so we do not care.
579 	 */
580 
581 	for(i = PQ_L2_SIZE / 2; i > 0; --i) {
582 		if ((m = TAILQ_FIRST(&pq[(index + i) & PQ_L2_MASK].pl)) != NULL)
583 			break;
584 
585 		if ((m = TAILQ_FIRST(&pq[(index - i) & PQ_L2_MASK].pl)) != NULL)
586 			break;
587 	}
588 	return(m);
589 }
590 
591 vm_page_t
592 vm_page_list_find(int basequeue, int index, boolean_t prefer_zero)
593 {
594 	return(_vm_page_list_find(basequeue, index, prefer_zero));
595 }
596 
597 /*
598  * Find a page on the cache queue with color optimization.  As pages
599  * might be found, but not applicable, they are deactivated.  This
600  * keeps us from using potentially busy cached pages.
601  *
602  * This routine must be called with a critical section held.
603  * This routine may not block.
604  */
605 vm_page_t
606 vm_page_select_cache(vm_object_t object, vm_pindex_t pindex)
607 {
608 	vm_page_t m;
609 
610 	while (TRUE) {
611 		m = _vm_page_list_find(
612 		    PQ_CACHE,
613 		    (pindex + object->pg_color) & PQ_L2_MASK,
614 		    FALSE
615 		);
616 		if (m && ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy ||
617 			       m->hold_count || m->wire_count)) {
618 			vm_page_deactivate(m);
619 			continue;
620 		}
621 		return m;
622 	}
623 	/* not reached */
624 }
625 
626 /*
627  * Find a free or zero page, with specified preference.  We attempt to
628  * inline the nominal case and fall back to _vm_page_select_free()
629  * otherwise.
630  *
631  * This routine must be called with a critical section held.
632  * This routine may not block.
633  */
634 static __inline vm_page_t
635 vm_page_select_free(vm_object_t object, vm_pindex_t pindex, boolean_t prefer_zero)
636 {
637 	vm_page_t m;
638 
639 	m = _vm_page_list_find(
640 		PQ_FREE,
641 		(pindex + object->pg_color) & PQ_L2_MASK,
642 		prefer_zero
643 	);
644 	return(m);
645 }
646 
647 /*
648  * vm_page_alloc()
649  *
650  * Allocate and return a memory cell associated with this VM object/offset
651  * pair.
652  *
653  *	page_req classes:
654  *
655  *	VM_ALLOC_NORMAL		allow use of cache pages, nominal free drain
656  *	VM_ALLOC_SYSTEM		greater free drain
657  *	VM_ALLOC_INTERRUPT	allow free list to be completely drained
658  *	VM_ALLOC_ZERO		advisory request for pre-zero'd page
659  *
660  * The object must be locked.
661  * This routine may not block.
662  * The returned page will be marked PG_BUSY
663  *
664  * Additional special handling is required when called from an interrupt
665  * (VM_ALLOC_INTERRUPT).  We are not allowed to mess with the page cache
666  * in this case.
667  */
668 vm_page_t
669 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req)
670 {
671 	vm_page_t m = NULL;
672 
673 	KKASSERT(object != NULL);
674 	KASSERT(!vm_page_lookup(object, pindex),
675 		("vm_page_alloc: page already allocated"));
676 	KKASSERT(page_req &
677 		(VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
678 
679 	/*
680 	 * Certain system threads (pageout daemon, buf_daemon's) are
681 	 * allowed to eat deeper into the free page list.
682 	 */
683 	if (curthread->td_flags & TDF_SYSTHREAD)
684 		page_req |= VM_ALLOC_SYSTEM;
685 
686 	crit_enter();
687 loop:
688 	if (vmstats.v_free_count > vmstats.v_free_reserved ||
689 	    ((page_req & VM_ALLOC_INTERRUPT) && vmstats.v_free_count > 0) ||
690 	    ((page_req & VM_ALLOC_SYSTEM) && vmstats.v_cache_count == 0 &&
691 		vmstats.v_free_count > vmstats.v_interrupt_free_min)
692 	) {
693 		/*
694 		 * The free queue has sufficient free pages to take one out.
695 		 */
696 		if (page_req & VM_ALLOC_ZERO)
697 			m = vm_page_select_free(object, pindex, TRUE);
698 		else
699 			m = vm_page_select_free(object, pindex, FALSE);
700 	} else if (page_req & VM_ALLOC_NORMAL) {
701 		/*
702 		 * Allocatable from the cache (non-interrupt only).  On
703 		 * success, we must free the page and try again, thus
704 		 * ensuring that vmstats.v_*_free_min counters are replenished.
705 		 */
706 #ifdef INVARIANTS
707 		if (curthread->td_preempted) {
708 			kprintf("vm_page_alloc(): warning, attempt to allocate"
709 				" cache page from preempting interrupt\n");
710 			m = NULL;
711 		} else {
712 			m = vm_page_select_cache(object, pindex);
713 		}
714 #else
715 		m = vm_page_select_cache(object, pindex);
716 #endif
717 		/*
718 		 * On success move the page into the free queue and loop.
719 		 */
720 		if (m != NULL) {
721 			KASSERT(m->dirty == 0,
722 			    ("Found dirty cache page %p", m));
723 			vm_page_busy(m);
724 			vm_page_protect(m, VM_PROT_NONE);
725 			vm_page_free(m);
726 			goto loop;
727 		}
728 
729 		/*
730 		 * On failure return NULL
731 		 */
732 		crit_exit();
733 #if defined(DIAGNOSTIC)
734 		if (vmstats.v_cache_count > 0)
735 			kprintf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", vmstats.v_cache_count);
736 #endif
737 		vm_pageout_deficit++;
738 		pagedaemon_wakeup();
739 		return (NULL);
740 	} else {
741 		/*
742 		 * No pages available, wakeup the pageout daemon and give up.
743 		 */
744 		crit_exit();
745 		vm_pageout_deficit++;
746 		pagedaemon_wakeup();
747 		return (NULL);
748 	}
749 
750 	/*
751 	 * Good page found.  The page has not yet been busied.  We are in
752 	 * a critical section.
753 	 */
754 	KASSERT(m != NULL, ("vm_page_alloc(): missing page on free queue\n"));
755 	KASSERT(m->dirty == 0,
756 		("vm_page_alloc: free/cache page %p was dirty", m));
757 
758 	/*
759 	 * Remove from free queue
760 	 */
761 	vm_page_unqueue_nowakeup(m);
762 
763 	/*
764 	 * Initialize structure.  Only the PG_ZERO flag is inherited.  Set
765 	 * the page PG_BUSY
766 	 */
767 	if (m->flags & PG_ZERO) {
768 		vm_page_zero_count--;
769 		m->flags = PG_ZERO | PG_BUSY;
770 	} else {
771 		m->flags = PG_BUSY;
772 	}
773 	m->wire_count = 0;
774 	m->hold_count = 0;
775 	m->act_count = 0;
776 	m->busy = 0;
777 	m->valid = 0;
778 
779 	/*
780 	 * vm_page_insert() is safe prior to the crit_exit().  Note also that
781 	 * inserting a page here does not insert it into the pmap (which
782 	 * could cause us to block allocating memory).  We cannot block
783 	 * anywhere.
784 	 */
785 	vm_page_insert(m, object, pindex);
786 
787 	/*
788 	 * Don't wakeup too often - wakeup the pageout daemon when
789 	 * we would be nearly out of memory.
790 	 */
791 	pagedaemon_wakeup();
792 
793 	crit_exit();
794 
795 	/*
796 	 * A PG_BUSY page is returned.
797 	 */
798 	return (m);
799 }
800 
801 /*
802  * Block until free pages are available for allocation, called in various
803  * places before memory allocations.
804  */
805 void
806 vm_wait(int timo)
807 {
808 	crit_enter();
809 	if (curthread == pagethread) {
810 		vm_pageout_pages_needed = 1;
811 		tsleep(&vm_pageout_pages_needed, 0, "VMWait", timo);
812 	} else {
813 		if (vm_pages_needed == 0) {
814 			vm_pages_needed = 1;
815 			wakeup(&vm_pages_needed);
816 		}
817 		tsleep(&vmstats.v_free_count, 0, "vmwait", timo);
818 	}
819 	crit_exit();
820 }
821 
822 /*
823  * Block until free pages are available for allocation
824  *
825  * Called only in vm_fault so that processes page faulting can be
826  * easily tracked.
827  */
828 void
829 vm_waitpfault(void)
830 {
831 	crit_enter();
832 	if (vm_pages_needed == 0) {
833 		vm_pages_needed = 1;
834 		wakeup(&vm_pages_needed);
835 	}
836 	tsleep(&vmstats.v_free_count, 0, "pfault", 0);
837 	crit_exit();
838 }
839 
840 /*
841  * Put the specified page on the active list (if appropriate).  Ensure
842  * that act_count is at least ACT_INIT but do not otherwise mess with it.
843  *
844  * The page queues must be locked.
845  * This routine may not block.
846  */
847 void
848 vm_page_activate(vm_page_t m)
849 {
850 	crit_enter();
851 	if (m->queue != PQ_ACTIVE) {
852 		if ((m->queue - m->pc) == PQ_CACHE)
853 			mycpu->gd_cnt.v_reactivated++;
854 
855 		vm_page_unqueue(m);
856 
857 		if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
858 			m->queue = PQ_ACTIVE;
859 			vm_page_queues[PQ_ACTIVE].lcnt++;
860 			TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl,
861 					    m, pageq);
862 			if (m->act_count < ACT_INIT)
863 				m->act_count = ACT_INIT;
864 			vmstats.v_active_count++;
865 		}
866 	} else {
867 		if (m->act_count < ACT_INIT)
868 			m->act_count = ACT_INIT;
869 	}
870 	crit_exit();
871 }
872 
873 /*
874  * Helper routine for vm_page_free_toq() and vm_page_cache().  This
875  * routine is called when a page has been added to the cache or free
876  * queues.
877  *
878  * This routine may not block.
879  * This routine must be called at splvm()
880  */
881 static __inline void
882 vm_page_free_wakeup(void)
883 {
884 	/*
885 	 * if pageout daemon needs pages, then tell it that there are
886 	 * some free.
887 	 */
888 	if (vm_pageout_pages_needed &&
889 	    vmstats.v_cache_count + vmstats.v_free_count >=
890 	    vmstats.v_pageout_free_min
891 	) {
892 		wakeup(&vm_pageout_pages_needed);
893 		vm_pageout_pages_needed = 0;
894 	}
895 
896 	/*
897 	 * wakeup processes that are waiting on memory if we hit a
898 	 * high water mark. And wakeup scheduler process if we have
899 	 * lots of memory. this process will swapin processes.
900 	 */
901 	if (vm_pages_needed && !vm_page_count_min(0)) {
902 		vm_pages_needed = 0;
903 		wakeup(&vmstats.v_free_count);
904 	}
905 }
906 
907 /*
908  *	vm_page_free_toq:
909  *
910  *	Returns the given page to the PQ_FREE list, disassociating it with
911  *	any VM object.
912  *
913  *	The vm_page must be PG_BUSY on entry.  PG_BUSY will be released on
914  *	return (the page will have been freed).  No particular spl is required
915  *	on entry.
916  *
917  *	This routine may not block.
918  */
919 void
920 vm_page_free_toq(vm_page_t m)
921 {
922 	struct vpgqueues *pq;
923 
924 	crit_enter();
925 	mycpu->gd_cnt.v_tfree++;
926 
927 	KKASSERT((m->flags & PG_MAPPED) == 0);
928 
929 	if (m->busy || ((m->queue - m->pc) == PQ_FREE)) {
930 		kprintf(
931 		"vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n",
932 		    (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0,
933 		    m->hold_count);
934 		if ((m->queue - m->pc) == PQ_FREE)
935 			panic("vm_page_free: freeing free page");
936 		else
937 			panic("vm_page_free: freeing busy page");
938 	}
939 
940 	/*
941 	 * unqueue, then remove page.  Note that we cannot destroy
942 	 * the page here because we do not want to call the pager's
943 	 * callback routine until after we've put the page on the
944 	 * appropriate free queue.
945 	 */
946 	vm_page_unqueue_nowakeup(m);
947 	vm_page_remove(m);
948 
949 	/*
950 	 * No further management of fictitious pages occurs beyond object
951 	 * and queue removal.
952 	 */
953 	if ((m->flags & PG_FICTITIOUS) != 0) {
954 		vm_page_wakeup(m);
955 		crit_exit();
956 		return;
957 	}
958 
959 	m->valid = 0;
960 	vm_page_undirty(m);
961 
962 	if (m->wire_count != 0) {
963 		if (m->wire_count > 1) {
964 		    panic(
965 			"vm_page_free: invalid wire count (%d), pindex: 0x%lx",
966 			m->wire_count, (long)m->pindex);
967 		}
968 		panic("vm_page_free: freeing wired page");
969 	}
970 
971 	/*
972 	 * Clear the UNMANAGED flag when freeing an unmanaged page.
973 	 */
974 	if (m->flags & PG_UNMANAGED) {
975 	    m->flags &= ~PG_UNMANAGED;
976 	}
977 
978 	if (m->hold_count != 0) {
979 		m->flags &= ~PG_ZERO;
980 		m->queue = PQ_HOLD;
981 	} else {
982 		m->queue = PQ_FREE + m->pc;
983 	}
984 	pq = &vm_page_queues[m->queue];
985 	pq->lcnt++;
986 	++(*pq->cnt);
987 
988 	/*
989 	 * Put zero'd pages on the end ( where we look for zero'd pages
990 	 * first ) and non-zerod pages at the head.
991 	 */
992 	if (m->flags & PG_ZERO) {
993 		TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
994 		++vm_page_zero_count;
995 	} else {
996 		TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
997 	}
998 	vm_page_wakeup(m);
999 	vm_page_free_wakeup();
1000 	crit_exit();
1001 }
1002 
1003 /*
1004  * vm_page_unmanage()
1005  *
1006  * Prevent PV management from being done on the page.  The page is
1007  * removed from the paging queues as if it were wired, and as a
1008  * consequence of no longer being managed the pageout daemon will not
1009  * touch it (since there is no way to locate the pte mappings for the
1010  * page).  madvise() calls that mess with the pmap will also no longer
1011  * operate on the page.
1012  *
1013  * Beyond that the page is still reasonably 'normal'.  Freeing the page
1014  * will clear the flag.
1015  *
1016  * This routine is used by OBJT_PHYS objects - objects using unswappable
1017  * physical memory as backing store rather then swap-backed memory and
1018  * will eventually be extended to support 4MB unmanaged physical
1019  * mappings.
1020  *
1021  * Must be called with a critical section held.
1022  */
1023 void
1024 vm_page_unmanage(vm_page_t m)
1025 {
1026 	ASSERT_IN_CRIT_SECTION();
1027 	if ((m->flags & PG_UNMANAGED) == 0) {
1028 		if (m->wire_count == 0)
1029 			vm_page_unqueue(m);
1030 	}
1031 	vm_page_flag_set(m, PG_UNMANAGED);
1032 }
1033 
1034 /*
1035  * Mark this page as wired down by yet another map, removing it from
1036  * paging queues as necessary.
1037  *
1038  * The page queues must be locked.
1039  * This routine may not block.
1040  */
1041 void
1042 vm_page_wire(vm_page_t m)
1043 {
1044 	/*
1045 	 * Only bump the wire statistics if the page is not already wired,
1046 	 * and only unqueue the page if it is on some queue (if it is unmanaged
1047 	 * it is already off the queues).  Don't do anything with fictitious
1048 	 * pages because they are always wired.
1049 	 */
1050 	crit_enter();
1051 	if ((m->flags & PG_FICTITIOUS) == 0) {
1052 		if (m->wire_count == 0) {
1053 			if ((m->flags & PG_UNMANAGED) == 0)
1054 				vm_page_unqueue(m);
1055 			vmstats.v_wire_count++;
1056 		}
1057 		m->wire_count++;
1058 		KASSERT(m->wire_count != 0,
1059 			("vm_page_wire: wire_count overflow m=%p", m));
1060 	}
1061 	crit_exit();
1062 }
1063 
1064 /*
1065  * Release one wiring of this page, potentially enabling it to be paged again.
1066  *
1067  * Many pages placed on the inactive queue should actually go
1068  * into the cache, but it is difficult to figure out which.  What
1069  * we do instead, if the inactive target is well met, is to put
1070  * clean pages at the head of the inactive queue instead of the tail.
1071  * This will cause them to be moved to the cache more quickly and
1072  * if not actively re-referenced, freed more quickly.  If we just
1073  * stick these pages at the end of the inactive queue, heavy filesystem
1074  * meta-data accesses can cause an unnecessary paging load on memory bound
1075  * processes.  This optimization causes one-time-use metadata to be
1076  * reused more quickly.
1077  *
1078  * BUT, if we are in a low-memory situation we have no choice but to
1079  * put clean pages on the cache queue.
1080  *
1081  * A number of routines use vm_page_unwire() to guarantee that the page
1082  * will go into either the inactive or active queues, and will NEVER
1083  * be placed in the cache - for example, just after dirtying a page.
1084  * dirty pages in the cache are not allowed.
1085  *
1086  * The page queues must be locked.
1087  * This routine may not block.
1088  */
1089 void
1090 vm_page_unwire(vm_page_t m, int activate)
1091 {
1092 	crit_enter();
1093 	if (m->flags & PG_FICTITIOUS) {
1094 		/* do nothing */
1095 	} else if (m->wire_count <= 0) {
1096 		panic("vm_page_unwire: invalid wire count: %d", m->wire_count);
1097 	} else {
1098 		if (--m->wire_count == 0) {
1099 			--vmstats.v_wire_count;
1100 			if (m->flags & PG_UNMANAGED) {
1101 				;
1102 			} else if (activate) {
1103 				TAILQ_INSERT_TAIL(
1104 				    &vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1105 				m->queue = PQ_ACTIVE;
1106 				vm_page_queues[PQ_ACTIVE].lcnt++;
1107 				vmstats.v_active_count++;
1108 			} else {
1109 				vm_page_flag_clear(m, PG_WINATCFLS);
1110 				TAILQ_INSERT_TAIL(
1111 				    &vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1112 				m->queue = PQ_INACTIVE;
1113 				vm_page_queues[PQ_INACTIVE].lcnt++;
1114 				vmstats.v_inactive_count++;
1115 			}
1116 		}
1117 	}
1118 	crit_exit();
1119 }
1120 
1121 
1122 /*
1123  * Move the specified page to the inactive queue.  If the page has
1124  * any associated swap, the swap is deallocated.
1125  *
1126  * Normally athead is 0 resulting in LRU operation.  athead is set
1127  * to 1 if we want this page to be 'as if it were placed in the cache',
1128  * except without unmapping it from the process address space.
1129  *
1130  * This routine may not block.
1131  */
1132 static __inline void
1133 _vm_page_deactivate(vm_page_t m, int athead)
1134 {
1135 	/*
1136 	 * Ignore if already inactive.
1137 	 */
1138 	if (m->queue == PQ_INACTIVE)
1139 		return;
1140 
1141 	if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1142 		if ((m->queue - m->pc) == PQ_CACHE)
1143 			mycpu->gd_cnt.v_reactivated++;
1144 		vm_page_flag_clear(m, PG_WINATCFLS);
1145 		vm_page_unqueue(m);
1146 		if (athead)
1147 			TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1148 		else
1149 			TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1150 		m->queue = PQ_INACTIVE;
1151 		vm_page_queues[PQ_INACTIVE].lcnt++;
1152 		vmstats.v_inactive_count++;
1153 	}
1154 }
1155 
1156 void
1157 vm_page_deactivate(vm_page_t m)
1158 {
1159     crit_enter();
1160     _vm_page_deactivate(m, 0);
1161     crit_exit();
1162 }
1163 
1164 /*
1165  * vm_page_try_to_cache:
1166  *
1167  * Returns 0 on failure, 1 on success
1168  */
1169 int
1170 vm_page_try_to_cache(vm_page_t m)
1171 {
1172 	crit_enter();
1173 	if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1174 	    (m->flags & (PG_BUSY|PG_UNMANAGED))) {
1175 		crit_exit();
1176 		return(0);
1177 	}
1178 	vm_page_test_dirty(m);
1179 	if (m->dirty) {
1180 		crit_exit();
1181 		return(0);
1182 	}
1183 	vm_page_cache(m);
1184 	crit_exit();
1185 	return(1);
1186 }
1187 
1188 /*
1189  * Attempt to free the page.  If we cannot free it, we do nothing.
1190  * 1 is returned on success, 0 on failure.
1191  */
1192 int
1193 vm_page_try_to_free(vm_page_t m)
1194 {
1195 	crit_enter();
1196 	if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1197 	    (m->flags & (PG_BUSY|PG_UNMANAGED))) {
1198 		crit_exit();
1199 		return(0);
1200 	}
1201 	vm_page_test_dirty(m);
1202 	if (m->dirty) {
1203 		crit_exit();
1204 		return(0);
1205 	}
1206 	vm_page_busy(m);
1207 	vm_page_protect(m, VM_PROT_NONE);
1208 	vm_page_free(m);
1209 	crit_exit();
1210 	return(1);
1211 }
1212 
1213 /*
1214  * vm_page_cache
1215  *
1216  * Put the specified page onto the page cache queue (if appropriate).
1217  *
1218  * This routine may not block.
1219  */
1220 void
1221 vm_page_cache(vm_page_t m)
1222 {
1223 	ASSERT_IN_CRIT_SECTION();
1224 
1225 	if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy ||
1226 			m->wire_count || m->hold_count) {
1227 		kprintf("vm_page_cache: attempting to cache busy/held page\n");
1228 		return;
1229 	}
1230 
1231 	/*
1232 	 * Already in the cache (and thus not mapped)
1233 	 */
1234 	if ((m->queue - m->pc) == PQ_CACHE) {
1235 		KKASSERT((m->flags & PG_MAPPED) == 0);
1236 		return;
1237 	}
1238 
1239 	/*
1240 	 * Caller is required to test m->dirty, but note that the act of
1241 	 * removing the page from its maps can cause it to become dirty
1242 	 * on an SMP system due to another cpu running in usermode.
1243 	 */
1244 	if (m->dirty) {
1245 		panic("vm_page_cache: caching a dirty page, pindex: %ld",
1246 			(long)m->pindex);
1247 	}
1248 
1249 	/*
1250 	 * Remove all pmaps and indicate that the page is not
1251 	 * writeable or mapped.  Our vm_page_protect() call may
1252 	 * have blocked (especially w/ VM_PROT_NONE), so recheck
1253 	 * everything.
1254 	 */
1255 	vm_page_busy(m);
1256 	vm_page_protect(m, VM_PROT_NONE);
1257 	vm_page_wakeup(m);
1258 	if ((m->flags & (PG_BUSY|PG_UNMANAGED|PG_MAPPED)) || m->busy ||
1259 			m->wire_count || m->hold_count) {
1260 		/* do nothing */
1261 	} else if (m->dirty) {
1262 		vm_page_deactivate(m);
1263 	} else {
1264 		vm_page_unqueue_nowakeup(m);
1265 		m->queue = PQ_CACHE + m->pc;
1266 		vm_page_queues[m->queue].lcnt++;
1267 		TAILQ_INSERT_TAIL(&vm_page_queues[m->queue].pl, m, pageq);
1268 		vmstats.v_cache_count++;
1269 		vm_page_free_wakeup();
1270 	}
1271 }
1272 
1273 /*
1274  * vm_page_dontneed()
1275  *
1276  * Cache, deactivate, or do nothing as appropriate.  This routine
1277  * is typically used by madvise() MADV_DONTNEED.
1278  *
1279  * Generally speaking we want to move the page into the cache so
1280  * it gets reused quickly.  However, this can result in a silly syndrome
1281  * due to the page recycling too quickly.  Small objects will not be
1282  * fully cached.  On the otherhand, if we move the page to the inactive
1283  * queue we wind up with a problem whereby very large objects
1284  * unnecessarily blow away our inactive and cache queues.
1285  *
1286  * The solution is to move the pages based on a fixed weighting.  We
1287  * either leave them alone, deactivate them, or move them to the cache,
1288  * where moving them to the cache has the highest weighting.
1289  * By forcing some pages into other queues we eventually force the
1290  * system to balance the queues, potentially recovering other unrelated
1291  * space from active.  The idea is to not force this to happen too
1292  * often.
1293  */
1294 void
1295 vm_page_dontneed(vm_page_t m)
1296 {
1297 	static int dnweight;
1298 	int dnw;
1299 	int head;
1300 
1301 	dnw = ++dnweight;
1302 
1303 	/*
1304 	 * occassionally leave the page alone
1305 	 */
1306 	crit_enter();
1307 	if ((dnw & 0x01F0) == 0 ||
1308 	    m->queue == PQ_INACTIVE ||
1309 	    m->queue - m->pc == PQ_CACHE
1310 	) {
1311 		if (m->act_count >= ACT_INIT)
1312 			--m->act_count;
1313 		crit_exit();
1314 		return;
1315 	}
1316 
1317 	if (m->dirty == 0)
1318 		vm_page_test_dirty(m);
1319 
1320 	if (m->dirty || (dnw & 0x0070) == 0) {
1321 		/*
1322 		 * Deactivate the page 3 times out of 32.
1323 		 */
1324 		head = 0;
1325 	} else {
1326 		/*
1327 		 * Cache the page 28 times out of every 32.  Note that
1328 		 * the page is deactivated instead of cached, but placed
1329 		 * at the head of the queue instead of the tail.
1330 		 */
1331 		head = 1;
1332 	}
1333 	_vm_page_deactivate(m, head);
1334 	crit_exit();
1335 }
1336 
1337 /*
1338  * Grab a page, blocking if it is busy and allocating a page if necessary.
1339  * A busy page is returned or NULL.
1340  *
1341  * If VM_ALLOC_RETRY is specified VM_ALLOC_NORMAL must also be specified.
1342  * If VM_ALLOC_RETRY is not specified
1343  *
1344  * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is
1345  * always returned if we had blocked.
1346  * This routine will never return NULL if VM_ALLOC_RETRY is set.
1347  * This routine may not be called from an interrupt.
1348  * The returned page may not be entirely valid.
1349  *
1350  * This routine may be called from mainline code without spl protection and
1351  * be guarenteed a busied page associated with the object at the specified
1352  * index.
1353  */
1354 vm_page_t
1355 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
1356 {
1357 	vm_page_t m;
1358 	int generation;
1359 
1360 	KKASSERT(allocflags &
1361 		(VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
1362 	crit_enter();
1363 retrylookup:
1364 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
1365 		if (m->busy || (m->flags & PG_BUSY)) {
1366 			generation = object->generation;
1367 
1368 			while ((object->generation == generation) &&
1369 					(m->busy || (m->flags & PG_BUSY))) {
1370 				vm_page_flag_set(m, PG_WANTED | PG_REFERENCED);
1371 				tsleep(m, 0, "pgrbwt", 0);
1372 				if ((allocflags & VM_ALLOC_RETRY) == 0) {
1373 					m = NULL;
1374 					goto done;
1375 				}
1376 			}
1377 			goto retrylookup;
1378 		} else {
1379 			vm_page_busy(m);
1380 			goto done;
1381 		}
1382 	}
1383 	m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY);
1384 	if (m == NULL) {
1385 		vm_wait(0);
1386 		if ((allocflags & VM_ALLOC_RETRY) == 0)
1387 			goto done;
1388 		goto retrylookup;
1389 	}
1390 done:
1391 	crit_exit();
1392 	return(m);
1393 }
1394 
1395 /*
1396  * Mapping function for valid bits or for dirty bits in
1397  * a page.  May not block.
1398  *
1399  * Inputs are required to range within a page.
1400  */
1401 __inline int
1402 vm_page_bits(int base, int size)
1403 {
1404 	int first_bit;
1405 	int last_bit;
1406 
1407 	KASSERT(
1408 	    base + size <= PAGE_SIZE,
1409 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
1410 	);
1411 
1412 	if (size == 0)		/* handle degenerate case */
1413 		return(0);
1414 
1415 	first_bit = base >> DEV_BSHIFT;
1416 	last_bit = (base + size - 1) >> DEV_BSHIFT;
1417 
1418 	return ((2 << last_bit) - (1 << first_bit));
1419 }
1420 
1421 /*
1422  * Sets portions of a page valid and clean.  The arguments are expected
1423  * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
1424  * of any partial chunks touched by the range.  The invalid portion of
1425  * such chunks will be zero'd.
1426  *
1427  * This routine may not block.
1428  *
1429  * (base + size) must be less then or equal to PAGE_SIZE.
1430  */
1431 void
1432 vm_page_set_validclean(vm_page_t m, int base, int size)
1433 {
1434 	int pagebits;
1435 	int frag;
1436 	int endoff;
1437 
1438 	if (size == 0)	/* handle degenerate case */
1439 		return;
1440 
1441 	/*
1442 	 * If the base is not DEV_BSIZE aligned and the valid
1443 	 * bit is clear, we have to zero out a portion of the
1444 	 * first block.
1445 	 */
1446 
1447 	if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
1448 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0
1449 	) {
1450 		pmap_zero_page_area(
1451 		    VM_PAGE_TO_PHYS(m),
1452 		    frag,
1453 		    base - frag
1454 		);
1455 	}
1456 
1457 	/*
1458 	 * If the ending offset is not DEV_BSIZE aligned and the
1459 	 * valid bit is clear, we have to zero out a portion of
1460 	 * the last block.
1461 	 */
1462 
1463 	endoff = base + size;
1464 
1465 	if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
1466 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0
1467 	) {
1468 		pmap_zero_page_area(
1469 		    VM_PAGE_TO_PHYS(m),
1470 		    endoff,
1471 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1))
1472 		);
1473 	}
1474 
1475 	/*
1476 	 * Set valid, clear dirty bits.  If validating the entire
1477 	 * page we can safely clear the pmap modify bit.  We also
1478 	 * use this opportunity to clear the PG_NOSYNC flag.  If a process
1479 	 * takes a write fault on a MAP_NOSYNC memory area the flag will
1480 	 * be set again.
1481 	 *
1482 	 * We set valid bits inclusive of any overlap, but we can only
1483 	 * clear dirty bits for DEV_BSIZE chunks that are fully within
1484 	 * the range.
1485 	 */
1486 
1487 	pagebits = vm_page_bits(base, size);
1488 	m->valid |= pagebits;
1489 #if 0	/* NOT YET */
1490 	if ((frag = base & (DEV_BSIZE - 1)) != 0) {
1491 		frag = DEV_BSIZE - frag;
1492 		base += frag;
1493 		size -= frag;
1494 		if (size < 0)
1495 		    size = 0;
1496 	}
1497 	pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
1498 #endif
1499 	m->dirty &= ~pagebits;
1500 	if (base == 0 && size == PAGE_SIZE) {
1501 		pmap_clear_modify(m);
1502 		vm_page_flag_clear(m, PG_NOSYNC);
1503 	}
1504 }
1505 
1506 void
1507 vm_page_clear_dirty(vm_page_t m, int base, int size)
1508 {
1509 	m->dirty &= ~vm_page_bits(base, size);
1510 }
1511 
1512 /*
1513  * Make the page all-dirty.
1514  *
1515  * Also make sure the related object and vnode reflect the fact that the
1516  * object may now contain a dirty page.
1517  */
1518 void
1519 vm_page_dirty(vm_page_t m)
1520 {
1521 #ifdef INVARIANTS
1522         int pqtype = m->queue - m->pc;
1523 #endif
1524         KASSERT(pqtype != PQ_CACHE && pqtype != PQ_FREE,
1525                 ("vm_page_dirty: page in free/cache queue!"));
1526 	if (m->dirty != VM_PAGE_BITS_ALL) {
1527 		m->dirty = VM_PAGE_BITS_ALL;
1528 		if (m->object)
1529 			vm_object_set_writeable_dirty(m->object);
1530 	}
1531 }
1532 
1533 /*
1534  * Invalidates DEV_BSIZE'd chunks within a page.  Both the
1535  * valid and dirty bits for the effected areas are cleared.
1536  *
1537  * May not block.
1538  */
1539 void
1540 vm_page_set_invalid(vm_page_t m, int base, int size)
1541 {
1542 	int bits;
1543 
1544 	bits = vm_page_bits(base, size);
1545 	m->valid &= ~bits;
1546 	m->dirty &= ~bits;
1547 	m->object->generation++;
1548 }
1549 
1550 /*
1551  * The kernel assumes that the invalid portions of a page contain
1552  * garbage, but such pages can be mapped into memory by user code.
1553  * When this occurs, we must zero out the non-valid portions of the
1554  * page so user code sees what it expects.
1555  *
1556  * Pages are most often semi-valid when the end of a file is mapped
1557  * into memory and the file's size is not page aligned.
1558  */
1559 void
1560 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
1561 {
1562 	int b;
1563 	int i;
1564 
1565 	/*
1566 	 * Scan the valid bits looking for invalid sections that
1567 	 * must be zerod.  Invalid sub-DEV_BSIZE'd areas ( where the
1568 	 * valid bit may be set ) have already been zerod by
1569 	 * vm_page_set_validclean().
1570 	 */
1571 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
1572 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
1573 		    (m->valid & (1 << i))
1574 		) {
1575 			if (i > b) {
1576 				pmap_zero_page_area(
1577 				    VM_PAGE_TO_PHYS(m),
1578 				    b << DEV_BSHIFT,
1579 				    (i - b) << DEV_BSHIFT
1580 				);
1581 			}
1582 			b = i + 1;
1583 		}
1584 	}
1585 
1586 	/*
1587 	 * setvalid is TRUE when we can safely set the zero'd areas
1588 	 * as being valid.  We can do this if there are no cache consistency
1589 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
1590 	 */
1591 	if (setvalid)
1592 		m->valid = VM_PAGE_BITS_ALL;
1593 }
1594 
1595 /*
1596  * Is a (partial) page valid?  Note that the case where size == 0
1597  * will return FALSE in the degenerate case where the page is entirely
1598  * invalid, and TRUE otherwise.
1599  *
1600  * May not block.
1601  */
1602 int
1603 vm_page_is_valid(vm_page_t m, int base, int size)
1604 {
1605 	int bits = vm_page_bits(base, size);
1606 
1607 	if (m->valid && ((m->valid & bits) == bits))
1608 		return 1;
1609 	else
1610 		return 0;
1611 }
1612 
1613 /*
1614  * update dirty bits from pmap/mmu.  May not block.
1615  */
1616 void
1617 vm_page_test_dirty(vm_page_t m)
1618 {
1619 	if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) {
1620 		vm_page_dirty(m);
1621 	}
1622 }
1623 
1624 /*
1625  * Issue an event on a VM page.  Corresponding action structures are
1626  * removed from the page's list and called.
1627  */
1628 void
1629 vm_page_event_internal(vm_page_t m, vm_page_event_t event)
1630 {
1631 	struct vm_page_action *scan, *next;
1632 
1633 	LIST_FOREACH_MUTABLE(scan, &m->action_list, entry, next) {
1634 		if (scan->event == event) {
1635 			scan->event = VMEVENT_NONE;
1636 			LIST_REMOVE(scan, entry);
1637 			scan->func(m, scan);
1638 		}
1639 	}
1640 }
1641 
1642 #include "opt_ddb.h"
1643 #ifdef DDB
1644 #include <sys/kernel.h>
1645 
1646 #include <ddb/ddb.h>
1647 
1648 DB_SHOW_COMMAND(page, vm_page_print_page_info)
1649 {
1650 	db_printf("vmstats.v_free_count: %d\n", vmstats.v_free_count);
1651 	db_printf("vmstats.v_cache_count: %d\n", vmstats.v_cache_count);
1652 	db_printf("vmstats.v_inactive_count: %d\n", vmstats.v_inactive_count);
1653 	db_printf("vmstats.v_active_count: %d\n", vmstats.v_active_count);
1654 	db_printf("vmstats.v_wire_count: %d\n", vmstats.v_wire_count);
1655 	db_printf("vmstats.v_free_reserved: %d\n", vmstats.v_free_reserved);
1656 	db_printf("vmstats.v_free_min: %d\n", vmstats.v_free_min);
1657 	db_printf("vmstats.v_free_target: %d\n", vmstats.v_free_target);
1658 	db_printf("vmstats.v_cache_min: %d\n", vmstats.v_cache_min);
1659 	db_printf("vmstats.v_inactive_target: %d\n", vmstats.v_inactive_target);
1660 }
1661 
1662 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
1663 {
1664 	int i;
1665 	db_printf("PQ_FREE:");
1666 	for(i=0;i<PQ_L2_SIZE;i++) {
1667 		db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt);
1668 	}
1669 	db_printf("\n");
1670 
1671 	db_printf("PQ_CACHE:");
1672 	for(i=0;i<PQ_L2_SIZE;i++) {
1673 		db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt);
1674 	}
1675 	db_printf("\n");
1676 
1677 	db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n",
1678 		vm_page_queues[PQ_ACTIVE].lcnt,
1679 		vm_page_queues[PQ_INACTIVE].lcnt);
1680 }
1681 #endif /* DDB */
1682