xref: /freebsd/sys/vm/vm_page.c (revision ebbd4fa8c8427d3dd847ba33c45c996e0500e6ff)
1 /*
2  * Copyright (c) 1991 Regents of the University of California.
3  * All rights reserved.
4  *
5  * This code is derived from software contributed to Berkeley by
6  * The Mach Operating System project at Carnegie-Mellon University.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 3. All advertising materials mentioning features or use of this software
17  *    must display the following acknowledgement:
18  *	This product includes software developed by the University of
19  *	California, Berkeley and its contributors.
20  * 4. Neither the name of the University nor the names of its contributors
21  *    may be used to endorse or promote products derived from this software
22  *    without specific prior written permission.
23  *
24  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34  * SUCH DAMAGE.
35  *
36  *	from: @(#)vm_page.c	7.4 (Berkeley) 5/7/91
37  * $FreeBSD$
38  */
39 
40 /*
41  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
42  * All rights reserved.
43  *
44  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
45  *
46  * Permission to use, copy, modify and distribute this software and
47  * its documentation is hereby granted, provided that both the copyright
48  * notice and this permission notice appear in all copies of the
49  * software, derivative works or modified versions, and any portions
50  * thereof, and that both notices appear in supporting documentation.
51  *
52  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
53  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
54  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
55  *
56  * Carnegie Mellon requests users of this software to return to
57  *
58  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
59  *  School of Computer Science
60  *  Carnegie Mellon University
61  *  Pittsburgh PA 15213-3890
62  *
63  * any improvements or extensions that they make and grant Carnegie the
64  * rights to redistribute these changes.
65  */
66 
67 /*
68  *			GENERAL RULES ON VM_PAGE MANIPULATION
69  *
70  *	- a pageq mutex is required when adding or removing a page from a
71  *	  page queue (vm_page_queue[]), regardless of other mutexes or the
72  *	  busy state of a page.
73  *
74  *	- a hash chain mutex is required when associating or disassociating
75  *	  a page from the VM PAGE CACHE hash table (vm_page_buckets),
76  *	  regardless of other mutexes or the busy state of a page.
77  *
78  *	- either a hash chain mutex OR a busied page is required in order
79  *	  to modify the page flags.  A hash chain mutex must be obtained in
80  *	  order to busy a page.  A page's flags cannot be modified by a
81  *	  hash chain mutex if the page is marked busy.
82  *
83  *	- The object memq mutex is held when inserting or removing
84  *	  pages from an object (vm_page_insert() or vm_page_remove()).  This
85  *	  is different from the object's main mutex.
86  *
87  *	Generally speaking, you have to be aware of side effects when running
88  *	vm_page ops.  A vm_page_lookup() will return with the hash chain
89  *	locked, whether it was able to lookup the page or not.  vm_page_free(),
90  *	vm_page_cache(), vm_page_activate(), and a number of other routines
91  *	will release the hash chain mutex for you.  Intermediate manipulation
92  *	routines such as vm_page_flag_set() expect the hash chain to be held
93  *	on entry and the hash chain will remain held on return.
94  *
95  *	pageq scanning can only occur with the pageq in question locked.
96  *	We have a known bottleneck with the active queue, but the cache
97  *	and free queues are actually arrays already.
98  */
99 
100 /*
101  *	Resident memory management module.
102  */
103 
104 #include <sys/param.h>
105 #include <sys/systm.h>
106 #include <sys/lock.h>
107 #include <sys/malloc.h>
108 #include <sys/mutex.h>
109 #include <sys/proc.h>
110 #include <sys/vmmeter.h>
111 #include <sys/vnode.h>
112 
113 #include <vm/vm.h>
114 #include <vm/vm_param.h>
115 #include <vm/vm_kern.h>
116 #include <vm/vm_object.h>
117 #include <vm/vm_page.h>
118 #include <vm/vm_pageout.h>
119 #include <vm/vm_pager.h>
120 #include <vm/vm_extern.h>
121 #include <vm/uma.h>
122 #include <vm/uma_int.h>
123 
124 /*
125  *	Associated with page of user-allocatable memory is a
126  *	page structure.
127  */
128 static struct mtx vm_page_buckets_mtx;
129 static struct vm_page **vm_page_buckets; /* Array of buckets */
130 static int vm_page_bucket_count;	/* How big is array? */
131 static int vm_page_hash_mask;		/* Mask for hash function */
132 
133 struct mtx vm_page_queue_mtx;
134 struct mtx vm_page_queue_free_mtx;
135 
136 vm_page_t vm_page_array = 0;
137 int vm_page_array_size = 0;
138 long first_page = 0;
139 int vm_page_zero_count = 0;
140 
141 /*
142  *	vm_set_page_size:
143  *
144  *	Sets the page size, perhaps based upon the memory
145  *	size.  Must be called before any use of page-size
146  *	dependent functions.
147  */
148 void
149 vm_set_page_size(void)
150 {
151 	if (cnt.v_page_size == 0)
152 		cnt.v_page_size = PAGE_SIZE;
153 	if (((cnt.v_page_size - 1) & cnt.v_page_size) != 0)
154 		panic("vm_set_page_size: page size not a power of two");
155 }
156 
157 /*
158  *	vm_page_startup:
159  *
160  *	Initializes the resident memory module.
161  *
162  *	Allocates memory for the page cells, and
163  *	for the object/offset-to-page hash table headers.
164  *	Each page cell is initialized and placed on the free list.
165  */
166 vm_offset_t
167 vm_page_startup(vm_offset_t starta, vm_offset_t enda, vm_offset_t vaddr)
168 {
169 	vm_offset_t mapped;
170 	struct vm_page **bucket;
171 	vm_size_t npages, page_range;
172 	vm_offset_t new_end;
173 	int i;
174 	vm_offset_t pa;
175 	int nblocks;
176 	vm_offset_t last_pa;
177 
178 	/* the biggest memory array is the second group of pages */
179 	vm_offset_t end;
180 	vm_offset_t biggestone, biggestsize;
181 
182 	vm_offset_t total;
183 	vm_size_t bootpages;
184 
185 	total = 0;
186 	biggestsize = 0;
187 	biggestone = 0;
188 	nblocks = 0;
189 	vaddr = round_page(vaddr);
190 
191 	for (i = 0; phys_avail[i + 1]; i += 2) {
192 		phys_avail[i] = round_page(phys_avail[i]);
193 		phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
194 	}
195 
196 	for (i = 0; phys_avail[i + 1]; i += 2) {
197 		vm_size_t size = phys_avail[i + 1] - phys_avail[i];
198 
199 		if (size > biggestsize) {
200 			biggestone = i;
201 			biggestsize = size;
202 		}
203 		++nblocks;
204 		total += size;
205 	}
206 
207 	end = phys_avail[biggestone+1];
208 
209 	/*
210 	 * Initialize the locks.
211 	 */
212 	mtx_init(&vm_page_queue_mtx, "vm page queue mutex", NULL, MTX_DEF);
213 	mtx_init(&vm_page_queue_free_mtx, "vm page queue free mutex", NULL,
214 	   MTX_SPIN);
215 
216 	/*
217 	 * Initialize the queue headers for the free queue, the active queue
218 	 * and the inactive queue.
219 	 */
220 	vm_pageq_init();
221 
222 	/*
223 	 * Allocate memory for use when boot strapping the kernel memory allocator
224 	 */
225 	bootpages = UMA_BOOT_PAGES * UMA_SLAB_SIZE;
226 	new_end = end - bootpages;
227 	new_end = trunc_page(new_end);
228 	mapped = pmap_map(&vaddr, new_end, end,
229 	    VM_PROT_READ | VM_PROT_WRITE);
230 	bzero((caddr_t) mapped, end - new_end);
231 	uma_startup((caddr_t)mapped);
232 
233 	end = new_end;
234 
235 	/*
236 	 * Allocate (and initialize) the hash table buckets.
237 	 *
238 	 * The number of buckets MUST BE a power of 2, and the actual value is
239 	 * the next power of 2 greater than the number of physical pages in
240 	 * the system.
241 	 *
242 	 * We make the hash table approximately 2x the number of pages to
243 	 * reduce the chain length.  This is about the same size using the
244 	 * singly-linked list as the 1x hash table we were using before
245 	 * using TAILQ but the chain length will be smaller.
246 	 *
247 	 * Note: This computation can be tweaked if desired.
248 	 */
249 	if (vm_page_bucket_count == 0) {
250 		vm_page_bucket_count = 1;
251 		while (vm_page_bucket_count < atop(total))
252 			vm_page_bucket_count <<= 1;
253 	}
254 	vm_page_bucket_count <<= 1;
255 	vm_page_hash_mask = vm_page_bucket_count - 1;
256 
257 	/*
258 	 * Validate these addresses.
259 	 */
260 	new_end = end - vm_page_bucket_count * sizeof(struct vm_page *);
261 	new_end = trunc_page(new_end);
262 	mapped = pmap_map(&vaddr, new_end, end,
263 	    VM_PROT_READ | VM_PROT_WRITE);
264 	bzero((caddr_t) mapped, end - new_end);
265 
266 	mtx_init(&vm_page_buckets_mtx, "vm page buckets mutex", NULL, MTX_SPIN);
267 	vm_page_buckets = (struct vm_page **)mapped;
268 	bucket = vm_page_buckets;
269 	for (i = 0; i < vm_page_bucket_count; i++) {
270 		*bucket = NULL;
271 		bucket++;
272 	}
273 
274 	/*
275 	 * Compute the number of pages of memory that will be available for
276 	 * use (taking into account the overhead of a page structure per
277 	 * page).
278 	 */
279 	first_page = phys_avail[0] / PAGE_SIZE;
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 	end = new_end;
284 
285 	/*
286 	 * Initialize the mem entry structures now, and put them in the free
287 	 * queue.
288 	 */
289 	new_end = trunc_page(end - page_range * sizeof(struct vm_page));
290 	mapped = pmap_map(&vaddr, new_end, end,
291 	    VM_PROT_READ | VM_PROT_WRITE);
292 	vm_page_array = (vm_page_t) mapped;
293 
294 	/*
295 	 * Clear all of the page structures
296 	 */
297 	bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
298 	vm_page_array_size = page_range;
299 
300 	/*
301 	 * Construct the free queue(s) in descending order (by physical
302 	 * address) so that the first 16MB of physical memory is allocated
303 	 * last rather than first.  On large-memory machines, this avoids
304 	 * the exhaustion of low physical memory before isa_dmainit has run.
305 	 */
306 	cnt.v_page_count = 0;
307 	cnt.v_free_count = 0;
308 	for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) {
309 		pa = phys_avail[i];
310 		if (i == biggestone)
311 			last_pa = new_end;
312 		else
313 			last_pa = phys_avail[i + 1];
314 		while (pa < last_pa && npages-- > 0) {
315 			vm_pageq_add_new_page(pa);
316 			pa += PAGE_SIZE;
317 		}
318 	}
319 	return (vaddr);
320 }
321 
322 /*
323  *	vm_page_hash:
324  *
325  *	Distributes the object/offset key pair among hash buckets.
326  *
327  *	NOTE:  This macro depends on vm_page_bucket_count being a power of 2.
328  *	This routine may not block.
329  *
330  *	We try to randomize the hash based on the object to spread the pages
331  *	out in the hash table without it costing us too much.
332  */
333 static __inline int
334 vm_page_hash(vm_object_t object, vm_pindex_t pindex)
335 {
336 	int i = ((uintptr_t)object + pindex) ^ object->hash_rand;
337 
338 	return (i & vm_page_hash_mask);
339 }
340 
341 void
342 vm_page_flag_set(vm_page_t m, unsigned short bits)
343 {
344 	GIANT_REQUIRED;
345 	m->flags |= bits;
346 }
347 
348 void
349 vm_page_flag_clear(vm_page_t m, unsigned short bits)
350 {
351 	GIANT_REQUIRED;
352 	m->flags &= ~bits;
353 }
354 
355 void
356 vm_page_busy(vm_page_t m)
357 {
358 	KASSERT((m->flags & PG_BUSY) == 0,
359 	    ("vm_page_busy: page already busy!!!"));
360 	vm_page_flag_set(m, PG_BUSY);
361 }
362 
363 /*
364  *      vm_page_flash:
365  *
366  *      wakeup anyone waiting for the page.
367  */
368 void
369 vm_page_flash(vm_page_t m)
370 {
371 	if (m->flags & PG_WANTED) {
372 		vm_page_flag_clear(m, PG_WANTED);
373 		wakeup(m);
374 	}
375 }
376 
377 /*
378  *      vm_page_wakeup:
379  *
380  *      clear the PG_BUSY flag and wakeup anyone waiting for the
381  *      page.
382  *
383  */
384 void
385 vm_page_wakeup(vm_page_t m)
386 {
387 	KASSERT(m->flags & PG_BUSY, ("vm_page_wakeup: page not busy!!!"));
388 	vm_page_flag_clear(m, PG_BUSY);
389 	vm_page_flash(m);
390 }
391 
392 /*
393  *
394  *
395  */
396 void
397 vm_page_io_start(vm_page_t m)
398 {
399 
400 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
401 	m->busy++;
402 }
403 
404 void
405 vm_page_io_finish(vm_page_t m)
406 {
407 	GIANT_REQUIRED;
408 	m->busy--;
409 	if (m->busy == 0)
410 		vm_page_flash(m);
411 }
412 
413 /*
414  * Keep page from being freed by the page daemon
415  * much of the same effect as wiring, except much lower
416  * overhead and should be used only for *very* temporary
417  * holding ("wiring").
418  */
419 void
420 vm_page_hold(vm_page_t mem)
421 {
422         GIANT_REQUIRED;
423         mem->hold_count++;
424 }
425 
426 void
427 vm_page_unhold(vm_page_t mem)
428 {
429 	GIANT_REQUIRED;
430 	--mem->hold_count;
431 	KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!"));
432 	if (mem->hold_count == 0 && mem->queue == PQ_HOLD)
433 		vm_page_free_toq(mem);
434 }
435 
436 /*
437  *	vm_page_protect:
438  *
439  *	Reduce the protection of a page.  This routine never raises the
440  *	protection and therefore can be safely called if the page is already
441  *	at VM_PROT_NONE (it will be a NOP effectively ).
442  */
443 void
444 vm_page_protect(vm_page_t mem, int prot)
445 {
446 	if (prot == VM_PROT_NONE) {
447 		if (mem->flags & (PG_WRITEABLE|PG_MAPPED)) {
448 			pmap_page_protect(mem, VM_PROT_NONE);
449 			vm_page_flag_clear(mem, PG_WRITEABLE|PG_MAPPED);
450 		}
451 	} else if ((prot == VM_PROT_READ) && (mem->flags & PG_WRITEABLE)) {
452 		pmap_page_protect(mem, VM_PROT_READ);
453 		vm_page_flag_clear(mem, PG_WRITEABLE);
454 	}
455 }
456 /*
457  *	vm_page_zero_fill:
458  *
459  *	Zero-fill the specified page.
460  *	Written as a standard pagein routine, to
461  *	be used by the zero-fill object.
462  */
463 boolean_t
464 vm_page_zero_fill(vm_page_t m)
465 {
466 	pmap_zero_page(m);
467 	return (TRUE);
468 }
469 
470 /*
471  *	vm_page_zero_fill_area:
472  *
473  *	Like vm_page_zero_fill but only fill the specified area.
474  */
475 boolean_t
476 vm_page_zero_fill_area(vm_page_t m, int off, int size)
477 {
478 	pmap_zero_page_area(m, off, size);
479 	return (TRUE);
480 }
481 
482 /*
483  *	vm_page_copy:
484  *
485  *	Copy one page to another
486  */
487 void
488 vm_page_copy(vm_page_t src_m, vm_page_t dest_m)
489 {
490 	pmap_copy_page(src_m, dest_m);
491 	dest_m->valid = VM_PAGE_BITS_ALL;
492 }
493 
494 /*
495  *	vm_page_free:
496  *
497  *	Free a page
498  *
499  *	The clearing of PG_ZERO is a temporary safety until the code can be
500  *	reviewed to determine that PG_ZERO is being properly cleared on
501  *	write faults or maps.  PG_ZERO was previously cleared in
502  *	vm_page_alloc().
503  */
504 void
505 vm_page_free(vm_page_t m)
506 {
507 	vm_page_flag_clear(m, PG_ZERO);
508 	vm_page_free_toq(m);
509 	vm_page_zero_idle_wakeup();
510 }
511 
512 /*
513  *	vm_page_free_zero:
514  *
515  *	Free a page to the zerod-pages queue
516  */
517 void
518 vm_page_free_zero(vm_page_t m)
519 {
520 	vm_page_flag_set(m, PG_ZERO);
521 	vm_page_free_toq(m);
522 }
523 
524 /*
525  *	vm_page_sleep_busy:
526  *
527  *	Wait until page is no longer PG_BUSY or (if also_m_busy is TRUE)
528  *	m->busy is zero.  Returns TRUE if it had to sleep ( including if
529  *	it almost had to sleep and made temporary spl*() mods), FALSE
530  *	otherwise.
531  *
532  *	This routine assumes that interrupts can only remove the busy
533  *	status from a page, not set the busy status or change it from
534  *	PG_BUSY to m->busy or vise versa (which would create a timing
535  *	window).
536  */
537 int
538 vm_page_sleep_busy(vm_page_t m, int also_m_busy, const char *msg)
539 {
540 	GIANT_REQUIRED;
541 	if ((m->flags & PG_BUSY) || (also_m_busy && m->busy))  {
542 		int s = splvm();
543 		if ((m->flags & PG_BUSY) || (also_m_busy && m->busy)) {
544 			/*
545 			 * Page is busy. Wait and retry.
546 			 */
547 			vm_page_flag_set(m, PG_WANTED | PG_REFERENCED);
548 			tsleep(m, PVM, msg, 0);
549 		}
550 		splx(s);
551 		return (TRUE);
552 		/* not reached */
553 	}
554 	return (FALSE);
555 }
556 
557 /*
558  *	vm_page_sleep_if_busy:
559  *
560  *	Sleep and release the page queues lock if PG_BUSY is set or,
561  *	if also_m_busy is TRUE, busy is non-zero.  Returns TRUE if the
562  *	thread slept and the page queues lock was released.
563  *	Otherwise, retains the page queues lock and returns FALSE.
564  */
565 int
566 vm_page_sleep_if_busy(vm_page_t m, int also_m_busy, const char *msg)
567 {
568 
569 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
570 	if ((m->flags & PG_BUSY) || (also_m_busy && m->busy)) {
571 		vm_page_flag_set(m, PG_WANTED | PG_REFERENCED);
572 		msleep(m, &vm_page_queue_mtx, PDROP | PVM, msg, 0);
573 		return (TRUE);
574 	}
575 	return (FALSE);
576 }
577 
578 /*
579  *	vm_page_dirty:
580  *
581  *	make page all dirty
582  */
583 void
584 vm_page_dirty(vm_page_t m)
585 {
586 	KASSERT(m->queue - m->pc != PQ_CACHE,
587 	    ("vm_page_dirty: page in cache!"));
588 	m->dirty = VM_PAGE_BITS_ALL;
589 }
590 
591 /*
592  *	vm_page_undirty:
593  *
594  *	Set page to not be dirty.  Note: does not clear pmap modify bits
595  */
596 void
597 vm_page_undirty(vm_page_t m)
598 {
599 	m->dirty = 0;
600 }
601 
602 /*
603  *	vm_page_insert:		[ internal use only ]
604  *
605  *	Inserts the given mem entry into the object and object list.
606  *
607  *	The pagetables are not updated but will presumably fault the page
608  *	in if necessary, or if a kernel page the caller will at some point
609  *	enter the page into the kernel's pmap.  We are not allowed to block
610  *	here so we *can't* do this anyway.
611  *
612  *	The object and page must be locked, and must be splhigh.
613  *	This routine may not block.
614  */
615 void
616 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
617 {
618 	struct vm_page **bucket;
619 
620 	GIANT_REQUIRED;
621 
622 	if (m->object != NULL)
623 		panic("vm_page_insert: already inserted");
624 
625 	/*
626 	 * Record the object/offset pair in this page
627 	 */
628 	m->object = object;
629 	m->pindex = pindex;
630 
631 	/*
632 	 * Insert it into the object_object/offset hash table
633 	 */
634 	bucket = &vm_page_buckets[vm_page_hash(object, pindex)];
635 	mtx_lock_spin(&vm_page_buckets_mtx);
636 	m->hnext = *bucket;
637 	*bucket = m;
638 	mtx_unlock_spin(&vm_page_buckets_mtx);
639 
640 	/*
641 	 * Now link into the object's list of backed pages.
642 	 */
643 	TAILQ_INSERT_TAIL(&object->memq, m, listq);
644 	object->generation++;
645 
646 	/*
647 	 * show that the object has one more resident page.
648 	 */
649 	object->resident_page_count++;
650 
651 	/*
652 	 * Since we are inserting a new and possibly dirty page,
653 	 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags.
654 	 */
655 	if (m->flags & PG_WRITEABLE)
656 		vm_object_set_writeable_dirty(object);
657 }
658 
659 /*
660  *	vm_page_remove:
661  *				NOTE: used by device pager as well -wfj
662  *
663  *	Removes the given mem entry from the object/offset-page
664  *	table and the object page list, but do not invalidate/terminate
665  *	the backing store.
666  *
667  *	The object and page must be locked, and at splhigh.
668  *	The underlying pmap entry (if any) is NOT removed here.
669  *	This routine may not block.
670  */
671 void
672 vm_page_remove(vm_page_t m)
673 {
674 	vm_object_t object;
675 	vm_page_t *bucket;
676 
677 	GIANT_REQUIRED;
678 
679 	if (m->object == NULL)
680 		return;
681 
682 	if ((m->flags & PG_BUSY) == 0) {
683 		panic("vm_page_remove: page not busy");
684 	}
685 
686 	/*
687 	 * Basically destroy the page.
688 	 */
689 	vm_page_wakeup(m);
690 
691 	object = m->object;
692 
693 	/*
694 	 * Remove from the object_object/offset hash table.  The object
695 	 * must be on the hash queue, we will panic if it isn't
696 	 */
697 	bucket = &vm_page_buckets[vm_page_hash(m->object, m->pindex)];
698 	mtx_lock_spin(&vm_page_buckets_mtx);
699 	while (*bucket != m) {
700 		if (*bucket == NULL)
701 			panic("vm_page_remove(): page not found in hash");
702 		bucket = &(*bucket)->hnext;
703 	}
704 	*bucket = m->hnext;
705 	m->hnext = NULL;
706 	mtx_unlock_spin(&vm_page_buckets_mtx);
707 
708 	/*
709 	 * Now remove from the object's list of backed pages.
710 	 */
711 	TAILQ_REMOVE(&object->memq, m, listq);
712 
713 	/*
714 	 * And show that the object has one fewer resident page.
715 	 */
716 	object->resident_page_count--;
717 	object->generation++;
718 
719 	m->object = NULL;
720 }
721 
722 /*
723  *	vm_page_lookup:
724  *
725  *	Returns the page associated with the object/offset
726  *	pair specified; if none is found, NULL is returned.
727  *
728  *	The object must be locked.  No side effects.
729  *	This routine may not block.
730  *	This is a critical path routine
731  */
732 vm_page_t
733 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
734 {
735 	vm_page_t m;
736 	struct vm_page **bucket;
737 
738 	/*
739 	 * Search the hash table for this object/offset pair
740 	 */
741 	bucket = &vm_page_buckets[vm_page_hash(object, pindex)];
742 	mtx_lock_spin(&vm_page_buckets_mtx);
743 	for (m = *bucket; m != NULL; m = m->hnext)
744 		if (m->object == object && m->pindex == pindex)
745 			break;
746 	mtx_unlock_spin(&vm_page_buckets_mtx);
747 	return (m);
748 }
749 
750 /*
751  *	vm_page_rename:
752  *
753  *	Move the given memory entry from its
754  *	current object to the specified target object/offset.
755  *
756  *	The object must be locked.
757  *	This routine may not block.
758  *
759  *	Note: this routine will raise itself to splvm(), the caller need not.
760  *
761  *	Note: swap associated with the page must be invalidated by the move.  We
762  *	      have to do this for several reasons:  (1) we aren't freeing the
763  *	      page, (2) we are dirtying the page, (3) the VM system is probably
764  *	      moving the page from object A to B, and will then later move
765  *	      the backing store from A to B and we can't have a conflict.
766  *
767  *	Note: we *always* dirty the page.  It is necessary both for the
768  *	      fact that we moved it, and because we may be invalidating
769  *	      swap.  If the page is on the cache, we have to deactivate it
770  *	      or vm_page_dirty() will panic.  Dirty pages are not allowed
771  *	      on the cache.
772  */
773 void
774 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
775 {
776 	int s;
777 
778 	s = splvm();
779 	vm_page_lock_queues();
780 	vm_page_remove(m);
781 	vm_page_insert(m, new_object, new_pindex);
782 	if (m->queue - m->pc == PQ_CACHE)
783 		vm_page_deactivate(m);
784 	vm_page_dirty(m);
785 	vm_page_unlock_queues();
786 	splx(s);
787 }
788 
789 /*
790  *	vm_page_select_cache:
791  *
792  *	Find a page on the cache queue with color optimization.  As pages
793  *	might be found, but not applicable, they are deactivated.  This
794  *	keeps us from using potentially busy cached pages.
795  *
796  *	This routine must be called at splvm().
797  *	This routine may not block.
798  */
799 static vm_page_t
800 vm_page_select_cache(vm_object_t object, vm_pindex_t pindex)
801 {
802 	vm_page_t m;
803 
804 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
805 	while (TRUE) {
806 		m = vm_pageq_find(
807 		    PQ_CACHE,
808 		    (pindex + object->pg_color) & PQ_L2_MASK,
809 		    FALSE
810 		);
811 		if (m && ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy ||
812 			       m->hold_count || m->wire_count)) {
813 			vm_page_deactivate(m);
814 			continue;
815 		}
816 		return m;
817 	}
818 }
819 
820 /*
821  *	vm_page_select_free:
822  *
823  *	Find a free or zero page, with specified preference.
824  *
825  *	This routine must be called at splvm().
826  *	This routine may not block.
827  */
828 static __inline vm_page_t
829 vm_page_select_free(vm_object_t object, vm_pindex_t pindex, boolean_t prefer_zero)
830 {
831 	vm_page_t m;
832 
833 	m = vm_pageq_find(
834 		PQ_FREE,
835 		(pindex + object->pg_color) & PQ_L2_MASK,
836 		prefer_zero
837 	);
838 	return (m);
839 }
840 
841 /*
842  *	vm_page_alloc:
843  *
844  *	Allocate and return a memory cell associated
845  *	with this VM object/offset pair.
846  *
847  *	page_req classes:
848  *	VM_ALLOC_NORMAL		normal process request
849  *	VM_ALLOC_SYSTEM		system *really* needs a page
850  *	VM_ALLOC_INTERRUPT	interrupt time request
851  *	VM_ALLOC_ZERO		zero page
852  *
853  *	This routine may not block.
854  *
855  *	Additional special handling is required when called from an
856  *	interrupt (VM_ALLOC_INTERRUPT).  We are not allowed to mess with
857  *	the page cache in this case.
858  */
859 vm_page_t
860 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
861 {
862 	vm_page_t m = NULL;
863 	int page_req, s;
864 
865 	GIANT_REQUIRED;
866 
867 	KASSERT(!vm_page_lookup(object, pindex),
868 		("vm_page_alloc: page already allocated"));
869 
870 	page_req = req & VM_ALLOC_CLASS_MASK;
871 
872 	/*
873 	 * The pager is allowed to eat deeper into the free page list.
874 	 */
875 	if ((curproc == pageproc) && (page_req != VM_ALLOC_INTERRUPT)) {
876 		page_req = VM_ALLOC_SYSTEM;
877 	};
878 
879 	s = splvm();
880 loop:
881 	mtx_lock_spin(&vm_page_queue_free_mtx);
882 	if (cnt.v_free_count > cnt.v_free_reserved) {
883 		/*
884 		 * Allocate from the free queue if there are plenty of pages
885 		 * in it.
886 		 */
887 		m = vm_page_select_free(object, pindex,
888 					(req & VM_ALLOC_ZERO) != 0);
889 	} else if (
890 	    (page_req == VM_ALLOC_SYSTEM &&
891 	     cnt.v_cache_count == 0 &&
892 	     cnt.v_free_count > cnt.v_interrupt_free_min) ||
893 	    (page_req == VM_ALLOC_INTERRUPT && cnt.v_free_count > 0)
894 	) {
895 		/*
896 		 * Interrupt or system, dig deeper into the free list.
897 		 */
898 		m = vm_page_select_free(object, pindex, FALSE);
899 	} else if (page_req != VM_ALLOC_INTERRUPT) {
900 		mtx_unlock_spin(&vm_page_queue_free_mtx);
901 		/*
902 		 * Allocatable from cache (non-interrupt only).  On success,
903 		 * we must free the page and try again, thus ensuring that
904 		 * cnt.v_*_free_min counters are replenished.
905 		 */
906 		vm_page_lock_queues();
907 		if ((m = vm_page_select_cache(object, pindex)) == NULL) {
908 			vm_page_unlock_queues();
909 			splx(s);
910 #if defined(DIAGNOSTIC)
911 			if (cnt.v_cache_count > 0)
912 				printf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", cnt.v_cache_count);
913 #endif
914 			vm_pageout_deficit++;
915 			pagedaemon_wakeup();
916 			return (NULL);
917 		}
918 		KASSERT(m->dirty == 0, ("Found dirty cache page %p", m));
919 		vm_page_busy(m);
920 		vm_page_protect(m, VM_PROT_NONE);
921 		vm_page_free(m);
922 		vm_page_unlock_queues();
923 		goto loop;
924 	} else {
925 		/*
926 		 * Not allocatable from cache from interrupt, give up.
927 		 */
928 		mtx_unlock_spin(&vm_page_queue_free_mtx);
929 		splx(s);
930 		vm_pageout_deficit++;
931 		pagedaemon_wakeup();
932 		return (NULL);
933 	}
934 
935 	/*
936 	 *  At this point we had better have found a good page.
937 	 */
938 
939 	KASSERT(
940 	    m != NULL,
941 	    ("vm_page_alloc(): missing page on free queue\n")
942 	);
943 
944 	/*
945 	 * Remove from free queue
946 	 */
947 
948 	vm_pageq_remove_nowakeup(m);
949 
950 	/*
951 	 * Initialize structure.  Only the PG_ZERO flag is inherited.
952 	 */
953 	if (m->flags & PG_ZERO) {
954 		vm_page_zero_count--;
955 		m->flags = PG_ZERO | PG_BUSY;
956 	} else {
957 		m->flags = PG_BUSY;
958 	}
959 	if (req & VM_ALLOC_WIRED) {
960 		cnt.v_wire_count++;
961 		m->flags |= PG_MAPPED;	/* XXX this does not belong here */
962 		m->wire_count = 1;
963 	} else
964 		m->wire_count = 0;
965 	m->hold_count = 0;
966 	m->act_count = 0;
967 	m->busy = 0;
968 	m->valid = 0;
969 	KASSERT(m->dirty == 0, ("vm_page_alloc: free/cache page %p was dirty", m));
970 	mtx_unlock_spin(&vm_page_queue_free_mtx);
971 
972 	/*
973 	 * vm_page_insert() is safe prior to the splx().  Note also that
974 	 * inserting a page here does not insert it into the pmap (which
975 	 * could cause us to block allocating memory).  We cannot block
976 	 * anywhere.
977 	 */
978 	vm_page_insert(m, object, pindex);
979 
980 	/*
981 	 * Don't wakeup too often - wakeup the pageout daemon when
982 	 * we would be nearly out of memory.
983 	 */
984 	if (vm_paging_needed())
985 		pagedaemon_wakeup();
986 
987 	splx(s);
988 	return (m);
989 }
990 
991 /*
992  *	vm_wait:	(also see VM_WAIT macro)
993  *
994  *	Block until free pages are available for allocation
995  *	- Called in various places before memory allocations.
996  */
997 void
998 vm_wait(void)
999 {
1000 	int s;
1001 
1002 	s = splvm();
1003 	if (curproc == pageproc) {
1004 		vm_pageout_pages_needed = 1;
1005 		tsleep(&vm_pageout_pages_needed, PSWP, "VMWait", 0);
1006 	} else {
1007 		if (!vm_pages_needed) {
1008 			vm_pages_needed = 1;
1009 			wakeup(&vm_pages_needed);
1010 		}
1011 		tsleep(&cnt.v_free_count, PVM, "vmwait", 0);
1012 	}
1013 	splx(s);
1014 }
1015 
1016 /*
1017  *	vm_waitpfault:	(also see VM_WAITPFAULT macro)
1018  *
1019  *	Block until free pages are available for allocation
1020  *	- Called only in vm_fault so that processes page faulting
1021  *	  can be easily tracked.
1022  *	- Sleeps at a lower priority than vm_wait() so that vm_wait()ing
1023  *	  processes will be able to grab memory first.  Do not change
1024  *	  this balance without careful testing first.
1025  */
1026 void
1027 vm_waitpfault(void)
1028 {
1029 	int s;
1030 
1031 	s = splvm();
1032 	if (!vm_pages_needed) {
1033 		vm_pages_needed = 1;
1034 		wakeup(&vm_pages_needed);
1035 	}
1036 	tsleep(&cnt.v_free_count, PUSER, "pfault", 0);
1037 	splx(s);
1038 }
1039 
1040 /*
1041  *	vm_page_activate:
1042  *
1043  *	Put the specified page on the active list (if appropriate).
1044  *	Ensure that act_count is at least ACT_INIT but do not otherwise
1045  *	mess with it.
1046  *
1047  *	The page queues must be locked.
1048  *	This routine may not block.
1049  */
1050 void
1051 vm_page_activate(vm_page_t m)
1052 {
1053 	int s;
1054 
1055 	GIANT_REQUIRED;
1056 	s = splvm();
1057 	if (m->queue != PQ_ACTIVE) {
1058 		if ((m->queue - m->pc) == PQ_CACHE)
1059 			cnt.v_reactivated++;
1060 		vm_pageq_remove(m);
1061 		if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1062 			if (m->act_count < ACT_INIT)
1063 				m->act_count = ACT_INIT;
1064 			vm_pageq_enqueue(PQ_ACTIVE, m);
1065 		}
1066 	} else {
1067 		if (m->act_count < ACT_INIT)
1068 			m->act_count = ACT_INIT;
1069 	}
1070 	splx(s);
1071 }
1072 
1073 /*
1074  *	vm_page_free_wakeup:
1075  *
1076  *	Helper routine for vm_page_free_toq() and vm_page_cache().  This
1077  *	routine is called when a page has been added to the cache or free
1078  *	queues.
1079  *
1080  *	This routine may not block.
1081  *	This routine must be called at splvm()
1082  */
1083 static __inline void
1084 vm_page_free_wakeup(void)
1085 {
1086 	/*
1087 	 * if pageout daemon needs pages, then tell it that there are
1088 	 * some free.
1089 	 */
1090 	if (vm_pageout_pages_needed &&
1091 	    cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) {
1092 		wakeup(&vm_pageout_pages_needed);
1093 		vm_pageout_pages_needed = 0;
1094 	}
1095 	/*
1096 	 * wakeup processes that are waiting on memory if we hit a
1097 	 * high water mark. And wakeup scheduler process if we have
1098 	 * lots of memory. this process will swapin processes.
1099 	 */
1100 	if (vm_pages_needed && !vm_page_count_min()) {
1101 		vm_pages_needed = 0;
1102 		wakeup(&cnt.v_free_count);
1103 	}
1104 }
1105 
1106 /*
1107  *	vm_page_free_toq:
1108  *
1109  *	Returns the given page to the PQ_FREE list,
1110  *	disassociating it with any VM object.
1111  *
1112  *	Object and page must be locked prior to entry.
1113  *	This routine may not block.
1114  */
1115 
1116 void
1117 vm_page_free_toq(vm_page_t m)
1118 {
1119 	int s;
1120 	struct vpgqueues *pq;
1121 	vm_object_t object = m->object;
1122 
1123 	GIANT_REQUIRED;
1124 	s = splvm();
1125 	cnt.v_tfree++;
1126 
1127 	if (m->busy || ((m->queue - m->pc) == PQ_FREE)) {
1128 		printf(
1129 		"vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n",
1130 		    (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0,
1131 		    m->hold_count);
1132 		if ((m->queue - m->pc) == PQ_FREE)
1133 			panic("vm_page_free: freeing free page");
1134 		else
1135 			panic("vm_page_free: freeing busy page");
1136 	}
1137 
1138 	/*
1139 	 * unqueue, then remove page.  Note that we cannot destroy
1140 	 * the page here because we do not want to call the pager's
1141 	 * callback routine until after we've put the page on the
1142 	 * appropriate free queue.
1143 	 */
1144 	vm_pageq_remove_nowakeup(m);
1145 	vm_page_remove(m);
1146 
1147 	/*
1148 	 * If fictitious remove object association and
1149 	 * return, otherwise delay object association removal.
1150 	 */
1151 	if ((m->flags & PG_FICTITIOUS) != 0) {
1152 		splx(s);
1153 		return;
1154 	}
1155 
1156 	m->valid = 0;
1157 	vm_page_undirty(m);
1158 
1159 	if (m->wire_count != 0) {
1160 		if (m->wire_count > 1) {
1161 			panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx",
1162 				m->wire_count, (long)m->pindex);
1163 		}
1164 		panic("vm_page_free: freeing wired page\n");
1165 	}
1166 
1167 	/*
1168 	 * If we've exhausted the object's resident pages we want to free
1169 	 * it up.
1170 	 */
1171 	if (object &&
1172 	    (object->type == OBJT_VNODE) &&
1173 	    ((object->flags & OBJ_DEAD) == 0)
1174 	) {
1175 		struct vnode *vp = (struct vnode *)object->handle;
1176 
1177 		if (vp && VSHOULDFREE(vp))
1178 			vfree(vp);
1179 	}
1180 
1181 	/*
1182 	 * Clear the UNMANAGED flag when freeing an unmanaged page.
1183 	 */
1184 	if (m->flags & PG_UNMANAGED) {
1185 		m->flags &= ~PG_UNMANAGED;
1186 	} else {
1187 #ifdef __alpha__
1188 		pmap_page_is_free(m);
1189 #endif
1190 	}
1191 
1192 	if (m->hold_count != 0) {
1193 		m->flags &= ~PG_ZERO;
1194 		m->queue = PQ_HOLD;
1195 	} else
1196 		m->queue = PQ_FREE + m->pc;
1197 	pq = &vm_page_queues[m->queue];
1198 	mtx_lock_spin(&vm_page_queue_free_mtx);
1199 	pq->lcnt++;
1200 	++(*pq->cnt);
1201 
1202 	/*
1203 	 * Put zero'd pages on the end ( where we look for zero'd pages
1204 	 * first ) and non-zerod pages at the head.
1205 	 */
1206 	if (m->flags & PG_ZERO) {
1207 		TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
1208 		++vm_page_zero_count;
1209 	} else {
1210 		TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
1211 	}
1212 	mtx_unlock_spin(&vm_page_queue_free_mtx);
1213 	vm_page_free_wakeup();
1214 	splx(s);
1215 }
1216 
1217 /*
1218  *	vm_page_unmanage:
1219  *
1220  * 	Prevent PV management from being done on the page.  The page is
1221  *	removed from the paging queues as if it were wired, and as a
1222  *	consequence of no longer being managed the pageout daemon will not
1223  *	touch it (since there is no way to locate the pte mappings for the
1224  *	page).  madvise() calls that mess with the pmap will also no longer
1225  *	operate on the page.
1226  *
1227  *	Beyond that the page is still reasonably 'normal'.  Freeing the page
1228  *	will clear the flag.
1229  *
1230  *	This routine is used by OBJT_PHYS objects - objects using unswappable
1231  *	physical memory as backing store rather then swap-backed memory and
1232  *	will eventually be extended to support 4MB unmanaged physical
1233  *	mappings.
1234  */
1235 void
1236 vm_page_unmanage(vm_page_t m)
1237 {
1238 	int s;
1239 
1240 	s = splvm();
1241 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1242 	if ((m->flags & PG_UNMANAGED) == 0) {
1243 		if (m->wire_count == 0)
1244 			vm_pageq_remove(m);
1245 	}
1246 	vm_page_flag_set(m, PG_UNMANAGED);
1247 	splx(s);
1248 }
1249 
1250 /*
1251  *	vm_page_wire:
1252  *
1253  *	Mark this page as wired down by yet
1254  *	another map, removing it from paging queues
1255  *	as necessary.
1256  *
1257  *	The page queues must be locked.
1258  *	This routine may not block.
1259  */
1260 void
1261 vm_page_wire(vm_page_t m)
1262 {
1263 	int s;
1264 
1265 	/*
1266 	 * Only bump the wire statistics if the page is not already wired,
1267 	 * and only unqueue the page if it is on some queue (if it is unmanaged
1268 	 * it is already off the queues).
1269 	 */
1270 	s = splvm();
1271 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1272 	if (m->wire_count == 0) {
1273 		if ((m->flags & PG_UNMANAGED) == 0)
1274 			vm_pageq_remove(m);
1275 		cnt.v_wire_count++;
1276 	}
1277 	m->wire_count++;
1278 	KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m));
1279 	splx(s);
1280 	vm_page_flag_set(m, PG_MAPPED);	/* XXX this does not belong here */
1281 }
1282 
1283 /*
1284  *	vm_page_unwire:
1285  *
1286  *	Release one wiring of this page, potentially
1287  *	enabling it to be paged again.
1288  *
1289  *	Many pages placed on the inactive queue should actually go
1290  *	into the cache, but it is difficult to figure out which.  What
1291  *	we do instead, if the inactive target is well met, is to put
1292  *	clean pages at the head of the inactive queue instead of the tail.
1293  *	This will cause them to be moved to the cache more quickly and
1294  *	if not actively re-referenced, freed more quickly.  If we just
1295  *	stick these pages at the end of the inactive queue, heavy filesystem
1296  *	meta-data accesses can cause an unnecessary paging load on memory bound
1297  *	processes.  This optimization causes one-time-use metadata to be
1298  *	reused more quickly.
1299  *
1300  *	BUT, if we are in a low-memory situation we have no choice but to
1301  *	put clean pages on the cache queue.
1302  *
1303  *	A number of routines use vm_page_unwire() to guarantee that the page
1304  *	will go into either the inactive or active queues, and will NEVER
1305  *	be placed in the cache - for example, just after dirtying a page.
1306  *	dirty pages in the cache are not allowed.
1307  *
1308  *	The page queues must be locked.
1309  *	This routine may not block.
1310  */
1311 void
1312 vm_page_unwire(vm_page_t m, int activate)
1313 {
1314 	int s;
1315 
1316 	s = splvm();
1317 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1318 	if (m->wire_count > 0) {
1319 		m->wire_count--;
1320 		if (m->wire_count == 0) {
1321 			cnt.v_wire_count--;
1322 			if (m->flags & PG_UNMANAGED) {
1323 				;
1324 			} else if (activate)
1325 				vm_pageq_enqueue(PQ_ACTIVE, m);
1326 			else {
1327 				vm_page_flag_clear(m, PG_WINATCFLS);
1328 				vm_pageq_enqueue(PQ_INACTIVE, m);
1329 			}
1330 		}
1331 	} else {
1332 		panic("vm_page_unwire: invalid wire count: %d\n", m->wire_count);
1333 	}
1334 	splx(s);
1335 }
1336 
1337 
1338 /*
1339  * Move the specified page to the inactive queue.  If the page has
1340  * any associated swap, the swap is deallocated.
1341  *
1342  * Normally athead is 0 resulting in LRU operation.  athead is set
1343  * to 1 if we want this page to be 'as if it were placed in the cache',
1344  * except without unmapping it from the process address space.
1345  *
1346  * This routine may not block.
1347  */
1348 static __inline void
1349 _vm_page_deactivate(vm_page_t m, int athead)
1350 {
1351 	int s;
1352 
1353 	GIANT_REQUIRED;
1354 	/*
1355 	 * Ignore if already inactive.
1356 	 */
1357 	if (m->queue == PQ_INACTIVE)
1358 		return;
1359 
1360 	s = splvm();
1361 	if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1362 		if ((m->queue - m->pc) == PQ_CACHE)
1363 			cnt.v_reactivated++;
1364 		vm_page_flag_clear(m, PG_WINATCFLS);
1365 		vm_pageq_remove(m);
1366 		if (athead)
1367 			TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1368 		else
1369 			TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1370 		m->queue = PQ_INACTIVE;
1371 		vm_page_queues[PQ_INACTIVE].lcnt++;
1372 		cnt.v_inactive_count++;
1373 	}
1374 	splx(s);
1375 }
1376 
1377 void
1378 vm_page_deactivate(vm_page_t m)
1379 {
1380     _vm_page_deactivate(m, 0);
1381 }
1382 
1383 /*
1384  * vm_page_try_to_cache:
1385  *
1386  * Returns 0 on failure, 1 on success
1387  */
1388 int
1389 vm_page_try_to_cache(vm_page_t m)
1390 {
1391 
1392 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1393 	if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1394 	    (m->flags & (PG_BUSY|PG_UNMANAGED))) {
1395 		return (0);
1396 	}
1397 	vm_page_test_dirty(m);
1398 	if (m->dirty)
1399 		return (0);
1400 	vm_page_cache(m);
1401 	return (1);
1402 }
1403 
1404 /*
1405  * vm_page_try_to_free()
1406  *
1407  *	Attempt to free the page.  If we cannot free it, we do nothing.
1408  *	1 is returned on success, 0 on failure.
1409  */
1410 int
1411 vm_page_try_to_free(vm_page_t m)
1412 {
1413 
1414 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1415 	if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1416 	    (m->flags & (PG_BUSY|PG_UNMANAGED))) {
1417 		return (0);
1418 	}
1419 	vm_page_test_dirty(m);
1420 	if (m->dirty)
1421 		return (0);
1422 	vm_page_busy(m);
1423 	vm_page_protect(m, VM_PROT_NONE);
1424 	vm_page_free(m);
1425 	return (1);
1426 }
1427 
1428 /*
1429  * vm_page_cache
1430  *
1431  * Put the specified page onto the page cache queue (if appropriate).
1432  *
1433  * This routine may not block.
1434  */
1435 void
1436 vm_page_cache(vm_page_t m)
1437 {
1438 	int s;
1439 
1440 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1441 	if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || m->wire_count) {
1442 		printf("vm_page_cache: attempting to cache busy page\n");
1443 		return;
1444 	}
1445 	if ((m->queue - m->pc) == PQ_CACHE)
1446 		return;
1447 
1448 	/*
1449 	 * Remove all pmaps and indicate that the page is not
1450 	 * writeable or mapped.
1451 	 */
1452 	vm_page_protect(m, VM_PROT_NONE);
1453 	if (m->dirty != 0) {
1454 		panic("vm_page_cache: caching a dirty page, pindex: %ld",
1455 			(long)m->pindex);
1456 	}
1457 	s = splvm();
1458 	vm_pageq_remove_nowakeup(m);
1459 	vm_pageq_enqueue(PQ_CACHE + m->pc, m);
1460 	vm_page_free_wakeup();
1461 	splx(s);
1462 }
1463 
1464 /*
1465  * vm_page_dontneed
1466  *
1467  *	Cache, deactivate, or do nothing as appropriate.  This routine
1468  *	is typically used by madvise() MADV_DONTNEED.
1469  *
1470  *	Generally speaking we want to move the page into the cache so
1471  *	it gets reused quickly.  However, this can result in a silly syndrome
1472  *	due to the page recycling too quickly.  Small objects will not be
1473  *	fully cached.  On the otherhand, if we move the page to the inactive
1474  *	queue we wind up with a problem whereby very large objects
1475  *	unnecessarily blow away our inactive and cache queues.
1476  *
1477  *	The solution is to move the pages based on a fixed weighting.  We
1478  *	either leave them alone, deactivate them, or move them to the cache,
1479  *	where moving them to the cache has the highest weighting.
1480  *	By forcing some pages into other queues we eventually force the
1481  *	system to balance the queues, potentially recovering other unrelated
1482  *	space from active.  The idea is to not force this to happen too
1483  *	often.
1484  */
1485 void
1486 vm_page_dontneed(vm_page_t m)
1487 {
1488 	static int dnweight;
1489 	int dnw;
1490 	int head;
1491 
1492 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1493 	dnw = ++dnweight;
1494 
1495 	/*
1496 	 * occassionally leave the page alone
1497 	 */
1498 	if ((dnw & 0x01F0) == 0 ||
1499 	    m->queue == PQ_INACTIVE ||
1500 	    m->queue - m->pc == PQ_CACHE
1501 	) {
1502 		if (m->act_count >= ACT_INIT)
1503 			--m->act_count;
1504 		return;
1505 	}
1506 
1507 	if (m->dirty == 0)
1508 		vm_page_test_dirty(m);
1509 
1510 	if (m->dirty || (dnw & 0x0070) == 0) {
1511 		/*
1512 		 * Deactivate the page 3 times out of 32.
1513 		 */
1514 		head = 0;
1515 	} else {
1516 		/*
1517 		 * Cache the page 28 times out of every 32.  Note that
1518 		 * the page is deactivated instead of cached, but placed
1519 		 * at the head of the queue instead of the tail.
1520 		 */
1521 		head = 1;
1522 	}
1523 	_vm_page_deactivate(m, head);
1524 }
1525 
1526 /*
1527  * Grab a page, waiting until we are waken up due to the page
1528  * changing state.  We keep on waiting, if the page continues
1529  * to be in the object.  If the page doesn't exist, allocate it.
1530  *
1531  * This routine may block.
1532  */
1533 vm_page_t
1534 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
1535 {
1536 	vm_page_t m;
1537 	int s, generation;
1538 
1539 	GIANT_REQUIRED;
1540 retrylookup:
1541 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
1542 		if (m->busy || (m->flags & PG_BUSY)) {
1543 			generation = object->generation;
1544 
1545 			s = splvm();
1546 			while ((object->generation == generation) &&
1547 					(m->busy || (m->flags & PG_BUSY))) {
1548 				vm_page_flag_set(m, PG_WANTED | PG_REFERENCED);
1549 				tsleep(m, PVM, "pgrbwt", 0);
1550 				if ((allocflags & VM_ALLOC_RETRY) == 0) {
1551 					splx(s);
1552 					return NULL;
1553 				}
1554 			}
1555 			splx(s);
1556 			goto retrylookup;
1557 		} else {
1558 			vm_page_lock_queues();
1559 			if (allocflags & VM_ALLOC_WIRED)
1560 				vm_page_wire(m);
1561 			vm_page_busy(m);
1562 			vm_page_unlock_queues();
1563 			return m;
1564 		}
1565 	}
1566 
1567 	m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY);
1568 	if (m == NULL) {
1569 		VM_WAIT;
1570 		if ((allocflags & VM_ALLOC_RETRY) == 0)
1571 			return NULL;
1572 		goto retrylookup;
1573 	}
1574 
1575 	return m;
1576 }
1577 
1578 /*
1579  * Mapping function for valid bits or for dirty bits in
1580  * a page.  May not block.
1581  *
1582  * Inputs are required to range within a page.
1583  */
1584 __inline int
1585 vm_page_bits(int base, int size)
1586 {
1587 	int first_bit;
1588 	int last_bit;
1589 
1590 	KASSERT(
1591 	    base + size <= PAGE_SIZE,
1592 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
1593 	);
1594 
1595 	if (size == 0)		/* handle degenerate case */
1596 		return (0);
1597 
1598 	first_bit = base >> DEV_BSHIFT;
1599 	last_bit = (base + size - 1) >> DEV_BSHIFT;
1600 
1601 	return ((2 << last_bit) - (1 << first_bit));
1602 }
1603 
1604 /*
1605  *	vm_page_set_validclean:
1606  *
1607  *	Sets portions of a page valid and clean.  The arguments are expected
1608  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
1609  *	of any partial chunks touched by the range.  The invalid portion of
1610  *	such chunks will be zero'd.
1611  *
1612  *	This routine may not block.
1613  *
1614  *	(base + size) must be less then or equal to PAGE_SIZE.
1615  */
1616 void
1617 vm_page_set_validclean(vm_page_t m, int base, int size)
1618 {
1619 	int pagebits;
1620 	int frag;
1621 	int endoff;
1622 
1623 	GIANT_REQUIRED;
1624 	if (size == 0)	/* handle degenerate case */
1625 		return;
1626 
1627 	/*
1628 	 * If the base is not DEV_BSIZE aligned and the valid
1629 	 * bit is clear, we have to zero out a portion of the
1630 	 * first block.
1631 	 */
1632 	if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
1633 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
1634 		pmap_zero_page_area(m, frag, base - frag);
1635 
1636 	/*
1637 	 * If the ending offset is not DEV_BSIZE aligned and the
1638 	 * valid bit is clear, we have to zero out a portion of
1639 	 * the last block.
1640 	 */
1641 	endoff = base + size;
1642 	if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
1643 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
1644 		pmap_zero_page_area(m, endoff,
1645 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
1646 
1647 	/*
1648 	 * Set valid, clear dirty bits.  If validating the entire
1649 	 * page we can safely clear the pmap modify bit.  We also
1650 	 * use this opportunity to clear the PG_NOSYNC flag.  If a process
1651 	 * takes a write fault on a MAP_NOSYNC memory area the flag will
1652 	 * be set again.
1653 	 *
1654 	 * We set valid bits inclusive of any overlap, but we can only
1655 	 * clear dirty bits for DEV_BSIZE chunks that are fully within
1656 	 * the range.
1657 	 */
1658 	pagebits = vm_page_bits(base, size);
1659 	m->valid |= pagebits;
1660 #if 0	/* NOT YET */
1661 	if ((frag = base & (DEV_BSIZE - 1)) != 0) {
1662 		frag = DEV_BSIZE - frag;
1663 		base += frag;
1664 		size -= frag;
1665 		if (size < 0)
1666 			size = 0;
1667 	}
1668 	pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
1669 #endif
1670 	m->dirty &= ~pagebits;
1671 	if (base == 0 && size == PAGE_SIZE) {
1672 		pmap_clear_modify(m);
1673 		vm_page_flag_clear(m, PG_NOSYNC);
1674 	}
1675 }
1676 
1677 #if 0
1678 
1679 void
1680 vm_page_set_dirty(vm_page_t m, int base, int size)
1681 {
1682 	m->dirty |= vm_page_bits(base, size);
1683 }
1684 
1685 #endif
1686 
1687 void
1688 vm_page_clear_dirty(vm_page_t m, int base, int size)
1689 {
1690 	GIANT_REQUIRED;
1691 	m->dirty &= ~vm_page_bits(base, size);
1692 }
1693 
1694 /*
1695  *	vm_page_set_invalid:
1696  *
1697  *	Invalidates DEV_BSIZE'd chunks within a page.  Both the
1698  *	valid and dirty bits for the effected areas are cleared.
1699  *
1700  *	May not block.
1701  */
1702 void
1703 vm_page_set_invalid(vm_page_t m, int base, int size)
1704 {
1705 	int bits;
1706 
1707 	GIANT_REQUIRED;
1708 	bits = vm_page_bits(base, size);
1709 	m->valid &= ~bits;
1710 	m->dirty &= ~bits;
1711 	m->object->generation++;
1712 }
1713 
1714 /*
1715  * vm_page_zero_invalid()
1716  *
1717  *	The kernel assumes that the invalid portions of a page contain
1718  *	garbage, but such pages can be mapped into memory by user code.
1719  *	When this occurs, we must zero out the non-valid portions of the
1720  *	page so user code sees what it expects.
1721  *
1722  *	Pages are most often semi-valid when the end of a file is mapped
1723  *	into memory and the file's size is not page aligned.
1724  */
1725 void
1726 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
1727 {
1728 	int b;
1729 	int i;
1730 
1731 	/*
1732 	 * Scan the valid bits looking for invalid sections that
1733 	 * must be zerod.  Invalid sub-DEV_BSIZE'd areas ( where the
1734 	 * valid bit may be set ) have already been zerod by
1735 	 * vm_page_set_validclean().
1736 	 */
1737 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
1738 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
1739 		    (m->valid & (1 << i))
1740 		) {
1741 			if (i > b) {
1742 				pmap_zero_page_area(m,
1743 				    b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
1744 			}
1745 			b = i + 1;
1746 		}
1747 	}
1748 
1749 	/*
1750 	 * setvalid is TRUE when we can safely set the zero'd areas
1751 	 * as being valid.  We can do this if there are no cache consistancy
1752 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
1753 	 */
1754 	if (setvalid)
1755 		m->valid = VM_PAGE_BITS_ALL;
1756 }
1757 
1758 /*
1759  *	vm_page_is_valid:
1760  *
1761  *	Is (partial) page valid?  Note that the case where size == 0
1762  *	will return FALSE in the degenerate case where the page is
1763  *	entirely invalid, and TRUE otherwise.
1764  *
1765  *	May not block.
1766  */
1767 int
1768 vm_page_is_valid(vm_page_t m, int base, int size)
1769 {
1770 	int bits = vm_page_bits(base, size);
1771 
1772 	if (m->valid && ((m->valid & bits) == bits))
1773 		return 1;
1774 	else
1775 		return 0;
1776 }
1777 
1778 /*
1779  * update dirty bits from pmap/mmu.  May not block.
1780  */
1781 void
1782 vm_page_test_dirty(vm_page_t m)
1783 {
1784 	if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) {
1785 		vm_page_dirty(m);
1786 	}
1787 }
1788 
1789 int so_zerocp_fullpage = 0;
1790 
1791 void
1792 vm_page_cowfault(vm_page_t m)
1793 {
1794 	vm_page_t mnew;
1795 	vm_object_t object;
1796 	vm_pindex_t pindex;
1797 
1798 	object = m->object;
1799 	pindex = m->pindex;
1800 	vm_page_busy(m);
1801 
1802  retry_alloc:
1803 	vm_page_remove(m);
1804 	mnew = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL);
1805 	if (mnew == NULL) {
1806 		vm_page_insert(m, object, pindex);
1807 		VM_WAIT;
1808 		goto retry_alloc;
1809 	}
1810 
1811 	if (m->cow == 0) {
1812 		/*
1813 		 * check to see if we raced with an xmit complete when
1814 		 * waiting to allocate a page.  If so, put things back
1815 		 * the way they were
1816 		 */
1817 		vm_page_busy(mnew);
1818 		vm_page_free(mnew);
1819 		vm_page_insert(m, object, pindex);
1820 	} else { /* clear COW & copy page */
1821 		if (so_zerocp_fullpage) {
1822 			mnew->valid = VM_PAGE_BITS_ALL;
1823 		} else {
1824 			vm_page_copy(m, mnew);
1825 		}
1826 		vm_page_dirty(mnew);
1827 		vm_page_flag_clear(mnew, PG_BUSY);
1828 	}
1829 }
1830 
1831 void
1832 vm_page_cowclear(vm_page_t m)
1833 {
1834 
1835 	/* XXX KDM find out if giant is required here. */
1836 	GIANT_REQUIRED;
1837 	if (m->cow) {
1838 		atomic_subtract_int(&m->cow, 1);
1839 		/*
1840 		 * let vm_fault add back write permission  lazily
1841 		 */
1842 	}
1843 	/*
1844 	 *  sf_buf_free() will free the page, so we needn't do it here
1845 	 */
1846 }
1847 
1848 void
1849 vm_page_cowsetup(vm_page_t m)
1850 {
1851 	/* XXX KDM find out if giant is required here */
1852 	GIANT_REQUIRED;
1853 	atomic_add_int(&m->cow, 1);
1854 	vm_page_protect(m, VM_PROT_READ);
1855 }
1856 
1857 #include "opt_ddb.h"
1858 #ifdef DDB
1859 #include <sys/kernel.h>
1860 
1861 #include <ddb/ddb.h>
1862 
1863 DB_SHOW_COMMAND(page, vm_page_print_page_info)
1864 {
1865 	db_printf("cnt.v_free_count: %d\n", cnt.v_free_count);
1866 	db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count);
1867 	db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count);
1868 	db_printf("cnt.v_active_count: %d\n", cnt.v_active_count);
1869 	db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count);
1870 	db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved);
1871 	db_printf("cnt.v_free_min: %d\n", cnt.v_free_min);
1872 	db_printf("cnt.v_free_target: %d\n", cnt.v_free_target);
1873 	db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min);
1874 	db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target);
1875 }
1876 
1877 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
1878 {
1879 	int i;
1880 	db_printf("PQ_FREE:");
1881 	for (i = 0; i < PQ_L2_SIZE; i++) {
1882 		db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt);
1883 	}
1884 	db_printf("\n");
1885 
1886 	db_printf("PQ_CACHE:");
1887 	for (i = 0; i < PQ_L2_SIZE; i++) {
1888 		db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt);
1889 	}
1890 	db_printf("\n");
1891 
1892 	db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n",
1893 		vm_page_queues[PQ_ACTIVE].lcnt,
1894 		vm_page_queues[PQ_INACTIVE].lcnt);
1895 }
1896 #endif /* DDB */
1897