xref: /freebsd/sys/vm/vm_page.c (revision 6486b015fc84e96725fef22b0e3363351399ae83)
1 /*-
2  * Copyright (c) 1991 Regents of the University of California.
3  * All rights reserved.
4  * Copyright (c) 1998 Matthew Dillon.  All Rights Reserved.
5  *
6  * This code is derived from software contributed to Berkeley by
7  * The Mach Operating System project at Carnegie-Mellon University.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  * 4. Neither the name of the University nor the names of its contributors
18  *    may be used to endorse or promote products derived from this software
19  *    without specific prior written permission.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
22  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
25  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31  * SUCH DAMAGE.
32  *
33  *	from: @(#)vm_page.c	7.4 (Berkeley) 5/7/91
34  */
35 
36 /*-
37  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
38  * All rights reserved.
39  *
40  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
41  *
42  * Permission to use, copy, modify and distribute this software and
43  * its documentation is hereby granted, provided that both the copyright
44  * notice and this permission notice appear in all copies of the
45  * software, derivative works or modified versions, and any portions
46  * thereof, and that both notices appear in supporting documentation.
47  *
48  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
49  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
50  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
51  *
52  * Carnegie Mellon requests users of this software to return to
53  *
54  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
55  *  School of Computer Science
56  *  Carnegie Mellon University
57  *  Pittsburgh PA 15213-3890
58  *
59  * any improvements or extensions that they make and grant Carnegie the
60  * rights to redistribute these changes.
61  */
62 
63 /*
64  *			GENERAL RULES ON VM_PAGE MANIPULATION
65  *
66  *	- a pageq mutex is required when adding or removing a page from a
67  *	  page queue (vm_page_queue[]), regardless of other mutexes or the
68  *	  busy state of a page.
69  *
70  *	- The object mutex is held when inserting or removing
71  *	  pages from an object (vm_page_insert() or vm_page_remove()).
72  *
73  */
74 
75 /*
76  *	Resident memory management module.
77  */
78 
79 #include <sys/cdefs.h>
80 __FBSDID("$FreeBSD$");
81 
82 #include "opt_vm.h"
83 
84 #include <sys/param.h>
85 #include <sys/systm.h>
86 #include <sys/lock.h>
87 #include <sys/kernel.h>
88 #include <sys/limits.h>
89 #include <sys/malloc.h>
90 #include <sys/msgbuf.h>
91 #include <sys/mutex.h>
92 #include <sys/proc.h>
93 #include <sys/sysctl.h>
94 #include <sys/vmmeter.h>
95 #include <sys/vnode.h>
96 
97 #include <vm/vm.h>
98 #include <vm/pmap.h>
99 #include <vm/vm_param.h>
100 #include <vm/vm_kern.h>
101 #include <vm/vm_object.h>
102 #include <vm/vm_page.h>
103 #include <vm/vm_pageout.h>
104 #include <vm/vm_pager.h>
105 #include <vm/vm_phys.h>
106 #include <vm/vm_reserv.h>
107 #include <vm/vm_extern.h>
108 #include <vm/uma.h>
109 #include <vm/uma_int.h>
110 
111 #include <machine/md_var.h>
112 
113 /*
114  *	Associated with page of user-allocatable memory is a
115  *	page structure.
116  */
117 
118 struct vpgqueues vm_page_queues[PQ_COUNT];
119 struct vpglocks vm_page_queue_lock;
120 struct vpglocks vm_page_queue_free_lock;
121 
122 struct vpglocks	pa_lock[PA_LOCK_COUNT];
123 
124 vm_page_t vm_page_array = 0;
125 int vm_page_array_size = 0;
126 long first_page = 0;
127 int vm_page_zero_count = 0;
128 
129 static int boot_pages = UMA_BOOT_PAGES;
130 TUNABLE_INT("vm.boot_pages", &boot_pages);
131 SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RD, &boot_pages, 0,
132 	"number of pages allocated for bootstrapping the VM system");
133 
134 int pa_tryrelock_restart;
135 SYSCTL_INT(_vm, OID_AUTO, tryrelock_restart, CTLFLAG_RD,
136     &pa_tryrelock_restart, 0, "Number of tryrelock restarts");
137 
138 static uma_zone_t fakepg_zone;
139 
140 static struct vnode *vm_page_alloc_init(vm_page_t m);
141 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits);
142 static void vm_page_queue_remove(int queue, vm_page_t m);
143 static void vm_page_enqueue(int queue, vm_page_t m);
144 static void vm_page_init_fakepg(void *dummy);
145 
146 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init_fakepg, NULL);
147 
148 static void
149 vm_page_init_fakepg(void *dummy)
150 {
151 
152 	fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
153 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
154 }
155 
156 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
157 #if PAGE_SIZE == 32768
158 #ifdef CTASSERT
159 CTASSERT(sizeof(u_long) >= 8);
160 #endif
161 #endif
162 
163 /*
164  * Try to acquire a physical address lock while a pmap is locked.  If we
165  * fail to trylock we unlock and lock the pmap directly and cache the
166  * locked pa in *locked.  The caller should then restart their loop in case
167  * the virtual to physical mapping has changed.
168  */
169 int
170 vm_page_pa_tryrelock(pmap_t pmap, vm_paddr_t pa, vm_paddr_t *locked)
171 {
172 	vm_paddr_t lockpa;
173 
174 	lockpa = *locked;
175 	*locked = pa;
176 	if (lockpa) {
177 		PA_LOCK_ASSERT(lockpa, MA_OWNED);
178 		if (PA_LOCKPTR(pa) == PA_LOCKPTR(lockpa))
179 			return (0);
180 		PA_UNLOCK(lockpa);
181 	}
182 	if (PA_TRYLOCK(pa))
183 		return (0);
184 	PMAP_UNLOCK(pmap);
185 	atomic_add_int(&pa_tryrelock_restart, 1);
186 	PA_LOCK(pa);
187 	PMAP_LOCK(pmap);
188 	return (EAGAIN);
189 }
190 
191 /*
192  *	vm_set_page_size:
193  *
194  *	Sets the page size, perhaps based upon the memory
195  *	size.  Must be called before any use of page-size
196  *	dependent functions.
197  */
198 void
199 vm_set_page_size(void)
200 {
201 	if (cnt.v_page_size == 0)
202 		cnt.v_page_size = PAGE_SIZE;
203 	if (((cnt.v_page_size - 1) & cnt.v_page_size) != 0)
204 		panic("vm_set_page_size: page size not a power of two");
205 }
206 
207 /*
208  *	vm_page_blacklist_lookup:
209  *
210  *	See if a physical address in this page has been listed
211  *	in the blacklist tunable.  Entries in the tunable are
212  *	separated by spaces or commas.  If an invalid integer is
213  *	encountered then the rest of the string is skipped.
214  */
215 static int
216 vm_page_blacklist_lookup(char *list, vm_paddr_t pa)
217 {
218 	vm_paddr_t bad;
219 	char *cp, *pos;
220 
221 	for (pos = list; *pos != '\0'; pos = cp) {
222 		bad = strtoq(pos, &cp, 0);
223 		if (*cp != '\0') {
224 			if (*cp == ' ' || *cp == ',') {
225 				cp++;
226 				if (cp == pos)
227 					continue;
228 			} else
229 				break;
230 		}
231 		if (pa == trunc_page(bad))
232 			return (1);
233 	}
234 	return (0);
235 }
236 
237 /*
238  *	vm_page_startup:
239  *
240  *	Initializes the resident memory module.
241  *
242  *	Allocates memory for the page cells, and
243  *	for the object/offset-to-page hash table headers.
244  *	Each page cell is initialized and placed on the free list.
245  */
246 vm_offset_t
247 vm_page_startup(vm_offset_t vaddr)
248 {
249 	vm_offset_t mapped;
250 	vm_paddr_t page_range;
251 	vm_paddr_t new_end;
252 	int i;
253 	vm_paddr_t pa;
254 	vm_paddr_t last_pa;
255 	char *list;
256 
257 	/* the biggest memory array is the second group of pages */
258 	vm_paddr_t end;
259 	vm_paddr_t biggestsize;
260 	vm_paddr_t low_water, high_water;
261 	int biggestone;
262 
263 	biggestsize = 0;
264 	biggestone = 0;
265 	vaddr = round_page(vaddr);
266 
267 	for (i = 0; phys_avail[i + 1]; i += 2) {
268 		phys_avail[i] = round_page(phys_avail[i]);
269 		phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
270 	}
271 
272 	low_water = phys_avail[0];
273 	high_water = phys_avail[1];
274 
275 	for (i = 0; phys_avail[i + 1]; i += 2) {
276 		vm_paddr_t size = phys_avail[i + 1] - phys_avail[i];
277 
278 		if (size > biggestsize) {
279 			biggestone = i;
280 			biggestsize = size;
281 		}
282 		if (phys_avail[i] < low_water)
283 			low_water = phys_avail[i];
284 		if (phys_avail[i + 1] > high_water)
285 			high_water = phys_avail[i + 1];
286 	}
287 
288 #ifdef XEN
289 	low_water = 0;
290 #endif
291 
292 	end = phys_avail[biggestone+1];
293 
294 	/*
295 	 * Initialize the locks.
296 	 */
297 	mtx_init(&vm_page_queue_mtx, "vm page queue mutex", NULL, MTX_DEF |
298 	    MTX_RECURSE);
299 	mtx_init(&vm_page_queue_free_mtx, "vm page queue free mutex", NULL,
300 	    MTX_DEF);
301 
302 	/* Setup page locks. */
303 	for (i = 0; i < PA_LOCK_COUNT; i++)
304 		mtx_init(&pa_lock[i].data, "page lock", NULL, MTX_DEF);
305 
306 	/*
307 	 * Initialize the queue headers for the hold queue, the active queue,
308 	 * and the inactive queue.
309 	 */
310 	for (i = 0; i < PQ_COUNT; i++)
311 		TAILQ_INIT(&vm_page_queues[i].pl);
312 	vm_page_queues[PQ_INACTIVE].cnt = &cnt.v_inactive_count;
313 	vm_page_queues[PQ_ACTIVE].cnt = &cnt.v_active_count;
314 	vm_page_queues[PQ_HOLD].cnt = &cnt.v_active_count;
315 
316 	/*
317 	 * Allocate memory for use when boot strapping the kernel memory
318 	 * allocator.
319 	 */
320 	new_end = end - (boot_pages * UMA_SLAB_SIZE);
321 	new_end = trunc_page(new_end);
322 	mapped = pmap_map(&vaddr, new_end, end,
323 	    VM_PROT_READ | VM_PROT_WRITE);
324 	bzero((void *)mapped, end - new_end);
325 	uma_startup((void *)mapped, boot_pages);
326 
327 #if defined(__amd64__) || defined(__i386__) || defined(__arm__) || \
328     defined(__mips__)
329 	/*
330 	 * Allocate a bitmap to indicate that a random physical page
331 	 * needs to be included in a minidump.
332 	 *
333 	 * The amd64 port needs this to indicate which direct map pages
334 	 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
335 	 *
336 	 * However, i386 still needs this workspace internally within the
337 	 * minidump code.  In theory, they are not needed on i386, but are
338 	 * included should the sf_buf code decide to use them.
339 	 */
340 	last_pa = 0;
341 	for (i = 0; dump_avail[i + 1] != 0; i += 2)
342 		if (dump_avail[i + 1] > last_pa)
343 			last_pa = dump_avail[i + 1];
344 	page_range = last_pa / PAGE_SIZE;
345 	vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
346 	new_end -= vm_page_dump_size;
347 	vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
348 	    new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
349 	bzero((void *)vm_page_dump, vm_page_dump_size);
350 #endif
351 #ifdef __amd64__
352 	/*
353 	 * Request that the physical pages underlying the message buffer be
354 	 * included in a crash dump.  Since the message buffer is accessed
355 	 * through the direct map, they are not automatically included.
356 	 */
357 	pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
358 	last_pa = pa + round_page(msgbufsize);
359 	while (pa < last_pa) {
360 		dump_add_page(pa);
361 		pa += PAGE_SIZE;
362 	}
363 #endif
364 	/*
365 	 * Compute the number of pages of memory that will be available for
366 	 * use (taking into account the overhead of a page structure per
367 	 * page).
368 	 */
369 	first_page = low_water / PAGE_SIZE;
370 #ifdef VM_PHYSSEG_SPARSE
371 	page_range = 0;
372 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
373 		page_range += atop(phys_avail[i + 1] - phys_avail[i]);
374 #elif defined(VM_PHYSSEG_DENSE)
375 	page_range = high_water / PAGE_SIZE - first_page;
376 #else
377 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
378 #endif
379 	end = new_end;
380 
381 	/*
382 	 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
383 	 */
384 	vaddr += PAGE_SIZE;
385 
386 	/*
387 	 * Initialize the mem entry structures now, and put them in the free
388 	 * queue.
389 	 */
390 	new_end = trunc_page(end - page_range * sizeof(struct vm_page));
391 	mapped = pmap_map(&vaddr, new_end, end,
392 	    VM_PROT_READ | VM_PROT_WRITE);
393 	vm_page_array = (vm_page_t) mapped;
394 #if VM_NRESERVLEVEL > 0
395 	/*
396 	 * Allocate memory for the reservation management system's data
397 	 * structures.
398 	 */
399 	new_end = vm_reserv_startup(&vaddr, new_end, high_water);
400 #endif
401 #if defined(__amd64__) || defined(__mips__)
402 	/*
403 	 * pmap_map on amd64 and mips can come out of the direct-map, not kvm
404 	 * like i386, so the pages must be tracked for a crashdump to include
405 	 * this data.  This includes the vm_page_array and the early UMA
406 	 * bootstrap pages.
407 	 */
408 	for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE)
409 		dump_add_page(pa);
410 #endif
411 	phys_avail[biggestone + 1] = new_end;
412 
413 	/*
414 	 * Clear all of the page structures
415 	 */
416 	bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
417 	for (i = 0; i < page_range; i++)
418 		vm_page_array[i].order = VM_NFREEORDER;
419 	vm_page_array_size = page_range;
420 
421 	/*
422 	 * Initialize the physical memory allocator.
423 	 */
424 	vm_phys_init();
425 
426 	/*
427 	 * Add every available physical page that is not blacklisted to
428 	 * the free lists.
429 	 */
430 	cnt.v_page_count = 0;
431 	cnt.v_free_count = 0;
432 	list = getenv("vm.blacklist");
433 	for (i = 0; phys_avail[i + 1] != 0; i += 2) {
434 		pa = phys_avail[i];
435 		last_pa = phys_avail[i + 1];
436 		while (pa < last_pa) {
437 			if (list != NULL &&
438 			    vm_page_blacklist_lookup(list, pa))
439 				printf("Skipping page with pa 0x%jx\n",
440 				    (uintmax_t)pa);
441 			else
442 				vm_phys_add_page(pa);
443 			pa += PAGE_SIZE;
444 		}
445 	}
446 	freeenv(list);
447 #if VM_NRESERVLEVEL > 0
448 	/*
449 	 * Initialize the reservation management system.
450 	 */
451 	vm_reserv_init();
452 #endif
453 	return (vaddr);
454 }
455 
456 
457 CTASSERT(offsetof(struct vm_page, aflags) % sizeof(uint32_t) == 0);
458 
459 void
460 vm_page_aflag_set(vm_page_t m, uint8_t bits)
461 {
462 	uint32_t *addr, val;
463 
464 	/*
465 	 * The PGA_WRITEABLE flag can only be set if the page is managed and
466 	 * VPO_BUSY.  Currently, this flag is only set by pmap_enter().
467 	 */
468 	KASSERT((bits & PGA_WRITEABLE) == 0 ||
469 	    (m->oflags & (VPO_UNMANAGED | VPO_BUSY)) == VPO_BUSY,
470 	    ("PGA_WRITEABLE and !VPO_BUSY"));
471 
472 	/*
473 	 * We want to use atomic updates for m->aflags, which is a
474 	 * byte wide.  Not all architectures provide atomic operations
475 	 * on the single-byte destination.  Punt and access the whole
476 	 * 4-byte word with an atomic update.  Parallel non-atomic
477 	 * updates to the fields included in the update by proximity
478 	 * are handled properly by atomics.
479 	 */
480 	addr = (void *)&m->aflags;
481 	MPASS(((uintptr_t)addr & (sizeof(uint32_t) - 1)) == 0);
482 	val = bits;
483 #if BYTE_ORDER == BIG_ENDIAN
484 	val <<= 24;
485 #endif
486 	atomic_set_32(addr, val);
487 }
488 
489 void
490 vm_page_aflag_clear(vm_page_t m, uint8_t bits)
491 {
492 	uint32_t *addr, val;
493 
494 	/*
495 	 * The PGA_REFERENCED flag can only be cleared if the object
496 	 * containing the page is locked.
497 	 */
498 	KASSERT((bits & PGA_REFERENCED) == 0 || VM_OBJECT_LOCKED(m->object),
499 	    ("PGA_REFERENCED and !VM_OBJECT_LOCKED"));
500 
501 	/*
502 	 * See the comment in vm_page_aflag_set().
503 	 */
504 	addr = (void *)&m->aflags;
505 	MPASS(((uintptr_t)addr & (sizeof(uint32_t) - 1)) == 0);
506 	val = bits;
507 #if BYTE_ORDER == BIG_ENDIAN
508 	val <<= 24;
509 #endif
510 	atomic_clear_32(addr, val);
511 }
512 
513 void
514 vm_page_reference(vm_page_t m)
515 {
516 
517 	vm_page_aflag_set(m, PGA_REFERENCED);
518 }
519 
520 void
521 vm_page_busy(vm_page_t m)
522 {
523 
524 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
525 	KASSERT((m->oflags & VPO_BUSY) == 0,
526 	    ("vm_page_busy: page already busy!!!"));
527 	m->oflags |= VPO_BUSY;
528 }
529 
530 /*
531  *      vm_page_flash:
532  *
533  *      wakeup anyone waiting for the page.
534  */
535 void
536 vm_page_flash(vm_page_t m)
537 {
538 
539 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
540 	if (m->oflags & VPO_WANTED) {
541 		m->oflags &= ~VPO_WANTED;
542 		wakeup(m);
543 	}
544 }
545 
546 /*
547  *      vm_page_wakeup:
548  *
549  *      clear the VPO_BUSY flag and wakeup anyone waiting for the
550  *      page.
551  *
552  */
553 void
554 vm_page_wakeup(vm_page_t m)
555 {
556 
557 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
558 	KASSERT(m->oflags & VPO_BUSY, ("vm_page_wakeup: page not busy!!!"));
559 	m->oflags &= ~VPO_BUSY;
560 	vm_page_flash(m);
561 }
562 
563 void
564 vm_page_io_start(vm_page_t m)
565 {
566 
567 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
568 	m->busy++;
569 }
570 
571 void
572 vm_page_io_finish(vm_page_t m)
573 {
574 
575 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
576 	KASSERT(m->busy > 0, ("vm_page_io_finish: page %p is not busy", m));
577 	m->busy--;
578 	if (m->busy == 0)
579 		vm_page_flash(m);
580 }
581 
582 /*
583  * Keep page from being freed by the page daemon
584  * much of the same effect as wiring, except much lower
585  * overhead and should be used only for *very* temporary
586  * holding ("wiring").
587  */
588 void
589 vm_page_hold(vm_page_t mem)
590 {
591 
592 	vm_page_lock_assert(mem, MA_OWNED);
593         mem->hold_count++;
594 }
595 
596 void
597 vm_page_unhold(vm_page_t mem)
598 {
599 
600 	vm_page_lock_assert(mem, MA_OWNED);
601 	--mem->hold_count;
602 	KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!"));
603 	if (mem->hold_count == 0 && mem->queue == PQ_HOLD)
604 		vm_page_free_toq(mem);
605 }
606 
607 /*
608  *	vm_page_unhold_pages:
609  *
610  *	Unhold each of the pages that is referenced by the given array.
611  */
612 void
613 vm_page_unhold_pages(vm_page_t *ma, int count)
614 {
615 	struct mtx *mtx, *new_mtx;
616 
617 	mtx = NULL;
618 	for (; count != 0; count--) {
619 		/*
620 		 * Avoid releasing and reacquiring the same page lock.
621 		 */
622 		new_mtx = vm_page_lockptr(*ma);
623 		if (mtx != new_mtx) {
624 			if (mtx != NULL)
625 				mtx_unlock(mtx);
626 			mtx = new_mtx;
627 			mtx_lock(mtx);
628 		}
629 		vm_page_unhold(*ma);
630 		ma++;
631 	}
632 	if (mtx != NULL)
633 		mtx_unlock(mtx);
634 }
635 
636 /*
637  *	vm_page_getfake:
638  *
639  *	Create a fictitious page with the specified physical address and
640  *	memory attribute.  The memory attribute is the only the machine-
641  *	dependent aspect of a fictitious page that must be initialized.
642  */
643 vm_page_t
644 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
645 {
646 	vm_page_t m;
647 
648 	m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
649 	m->phys_addr = paddr;
650 	m->queue = PQ_NONE;
651 	/* Fictitious pages don't use "segind". */
652 	m->flags = PG_FICTITIOUS;
653 	/* Fictitious pages don't use "order" or "pool". */
654 	m->oflags = VPO_BUSY | VPO_UNMANAGED;
655 	m->wire_count = 1;
656 	pmap_page_set_memattr(m, memattr);
657 	return (m);
658 }
659 
660 /*
661  *	vm_page_putfake:
662  *
663  *	Release a fictitious page.
664  */
665 void
666 vm_page_putfake(vm_page_t m)
667 {
668 
669 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
670 	    ("vm_page_putfake: bad page %p", m));
671 	uma_zfree(fakepg_zone, m);
672 }
673 
674 /*
675  *	vm_page_updatefake:
676  *
677  *	Update the given fictitious page to the specified physical address and
678  *	memory attribute.
679  */
680 void
681 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
682 {
683 
684 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
685 	    ("vm_page_updatefake: bad page %p", m));
686 	m->phys_addr = paddr;
687 	pmap_page_set_memattr(m, memattr);
688 }
689 
690 /*
691  *	vm_page_free:
692  *
693  *	Free a page.
694  */
695 void
696 vm_page_free(vm_page_t m)
697 {
698 
699 	m->flags &= ~PG_ZERO;
700 	vm_page_free_toq(m);
701 }
702 
703 /*
704  *	vm_page_free_zero:
705  *
706  *	Free a page to the zerod-pages queue
707  */
708 void
709 vm_page_free_zero(vm_page_t m)
710 {
711 
712 	m->flags |= PG_ZERO;
713 	vm_page_free_toq(m);
714 }
715 
716 /*
717  *	vm_page_sleep:
718  *
719  *	Sleep and release the page and page queues locks.
720  *
721  *	The object containing the given page must be locked.
722  */
723 void
724 vm_page_sleep(vm_page_t m, const char *msg)
725 {
726 
727 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
728 	if (mtx_owned(&vm_page_queue_mtx))
729 		vm_page_unlock_queues();
730 	if (mtx_owned(vm_page_lockptr(m)))
731 		vm_page_unlock(m);
732 
733 	/*
734 	 * It's possible that while we sleep, the page will get
735 	 * unbusied and freed.  If we are holding the object
736 	 * lock, we will assume we hold a reference to the object
737 	 * such that even if m->object changes, we can re-lock
738 	 * it.
739 	 */
740 	m->oflags |= VPO_WANTED;
741 	msleep(m, VM_OBJECT_MTX(m->object), PVM, msg, 0);
742 }
743 
744 /*
745  *	vm_page_dirty:
746  *
747  *	Set all bits in the page's dirty field.
748  *
749  *	The object containing the specified page must be locked if the
750  *	call is made from the machine-independent layer.
751  *
752  *	See vm_page_clear_dirty_mask().
753  */
754 void
755 vm_page_dirty(vm_page_t m)
756 {
757 
758 	KASSERT((m->flags & PG_CACHED) == 0,
759 	    ("vm_page_dirty: page in cache!"));
760 	KASSERT(!VM_PAGE_IS_FREE(m),
761 	    ("vm_page_dirty: page is free!"));
762 	KASSERT(m->valid == VM_PAGE_BITS_ALL,
763 	    ("vm_page_dirty: page is invalid!"));
764 	m->dirty = VM_PAGE_BITS_ALL;
765 }
766 
767 /*
768  *	vm_page_splay:
769  *
770  *	Implements Sleator and Tarjan's top-down splay algorithm.  Returns
771  *	the vm_page containing the given pindex.  If, however, that
772  *	pindex is not found in the vm_object, returns a vm_page that is
773  *	adjacent to the pindex, coming before or after it.
774  */
775 vm_page_t
776 vm_page_splay(vm_pindex_t pindex, vm_page_t root)
777 {
778 	struct vm_page dummy;
779 	vm_page_t lefttreemax, righttreemin, y;
780 
781 	if (root == NULL)
782 		return (root);
783 	lefttreemax = righttreemin = &dummy;
784 	for (;; root = y) {
785 		if (pindex < root->pindex) {
786 			if ((y = root->left) == NULL)
787 				break;
788 			if (pindex < y->pindex) {
789 				/* Rotate right. */
790 				root->left = y->right;
791 				y->right = root;
792 				root = y;
793 				if ((y = root->left) == NULL)
794 					break;
795 			}
796 			/* Link into the new root's right tree. */
797 			righttreemin->left = root;
798 			righttreemin = root;
799 		} else if (pindex > root->pindex) {
800 			if ((y = root->right) == NULL)
801 				break;
802 			if (pindex > y->pindex) {
803 				/* Rotate left. */
804 				root->right = y->left;
805 				y->left = root;
806 				root = y;
807 				if ((y = root->right) == NULL)
808 					break;
809 			}
810 			/* Link into the new root's left tree. */
811 			lefttreemax->right = root;
812 			lefttreemax = root;
813 		} else
814 			break;
815 	}
816 	/* Assemble the new root. */
817 	lefttreemax->right = root->left;
818 	righttreemin->left = root->right;
819 	root->left = dummy.right;
820 	root->right = dummy.left;
821 	return (root);
822 }
823 
824 /*
825  *	vm_page_insert:		[ internal use only ]
826  *
827  *	Inserts the given mem entry into the object and object list.
828  *
829  *	The pagetables are not updated but will presumably fault the page
830  *	in if necessary, or if a kernel page the caller will at some point
831  *	enter the page into the kernel's pmap.  We are not allowed to block
832  *	here so we *can't* do this anyway.
833  *
834  *	The object and page must be locked.
835  *	This routine may not block.
836  */
837 void
838 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
839 {
840 	vm_page_t root;
841 
842 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
843 	if (m->object != NULL)
844 		panic("vm_page_insert: page already inserted");
845 
846 	/*
847 	 * Record the object/offset pair in this page
848 	 */
849 	m->object = object;
850 	m->pindex = pindex;
851 
852 	/*
853 	 * Now link into the object's ordered list of backed pages.
854 	 */
855 	root = object->root;
856 	if (root == NULL) {
857 		m->left = NULL;
858 		m->right = NULL;
859 		TAILQ_INSERT_TAIL(&object->memq, m, listq);
860 	} else {
861 		root = vm_page_splay(pindex, root);
862 		if (pindex < root->pindex) {
863 			m->left = root->left;
864 			m->right = root;
865 			root->left = NULL;
866 			TAILQ_INSERT_BEFORE(root, m, listq);
867 		} else if (pindex == root->pindex)
868 			panic("vm_page_insert: offset already allocated");
869 		else {
870 			m->right = root->right;
871 			m->left = root;
872 			root->right = NULL;
873 			TAILQ_INSERT_AFTER(&object->memq, root, m, listq);
874 		}
875 	}
876 	object->root = m;
877 
878 	/*
879 	 * show that the object has one more resident page.
880 	 */
881 	object->resident_page_count++;
882 	/*
883 	 * Hold the vnode until the last page is released.
884 	 */
885 	if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
886 		vhold((struct vnode *)object->handle);
887 
888 	/*
889 	 * Since we are inserting a new and possibly dirty page,
890 	 * update the object's OBJ_MIGHTBEDIRTY flag.
891 	 */
892 	if (m->aflags & PGA_WRITEABLE)
893 		vm_object_set_writeable_dirty(object);
894 }
895 
896 /*
897  *	vm_page_remove:
898  *				NOTE: used by device pager as well -wfj
899  *
900  *	Removes the given mem entry from the object/offset-page
901  *	table and the object page list, but do not invalidate/terminate
902  *	the backing store.
903  *
904  *	The object and page must be locked.
905  *	The underlying pmap entry (if any) is NOT removed here.
906  *	This routine may not block.
907  */
908 void
909 vm_page_remove(vm_page_t m)
910 {
911 	vm_object_t object;
912 	vm_page_t next, prev, root;
913 
914 	if ((m->oflags & VPO_UNMANAGED) == 0)
915 		vm_page_lock_assert(m, MA_OWNED);
916 	if ((object = m->object) == NULL)
917 		return;
918 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
919 	if (m->oflags & VPO_BUSY) {
920 		m->oflags &= ~VPO_BUSY;
921 		vm_page_flash(m);
922 	}
923 
924 	/*
925 	 * Now remove from the object's list of backed pages.
926 	 */
927 	if ((next = TAILQ_NEXT(m, listq)) != NULL && next->left == m) {
928 		/*
929 		 * Since the page's successor in the list is also its parent
930 		 * in the tree, its right subtree must be empty.
931 		 */
932 		next->left = m->left;
933 		KASSERT(m->right == NULL,
934 		    ("vm_page_remove: page %p has right child", m));
935 	} else if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL &&
936 	    prev->right == m) {
937 		/*
938 		 * Since the page's predecessor in the list is also its parent
939 		 * in the tree, its left subtree must be empty.
940 		 */
941 		KASSERT(m->left == NULL,
942 		    ("vm_page_remove: page %p has left child", m));
943 		prev->right = m->right;
944 	} else {
945 		if (m != object->root)
946 			vm_page_splay(m->pindex, object->root);
947 		if (m->left == NULL)
948 			root = m->right;
949 		else if (m->right == NULL)
950 			root = m->left;
951 		else {
952 			/*
953 			 * Move the page's successor to the root, because
954 			 * pages are usually removed in ascending order.
955 			 */
956 			if (m->right != next)
957 				vm_page_splay(m->pindex, m->right);
958 			next->left = m->left;
959 			root = next;
960 		}
961 		object->root = root;
962 	}
963 	TAILQ_REMOVE(&object->memq, m, listq);
964 
965 	/*
966 	 * And show that the object has one fewer resident page.
967 	 */
968 	object->resident_page_count--;
969 	/*
970 	 * The vnode may now be recycled.
971 	 */
972 	if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
973 		vdrop((struct vnode *)object->handle);
974 
975 	m->object = NULL;
976 }
977 
978 /*
979  *	vm_page_lookup:
980  *
981  *	Returns the page associated with the object/offset
982  *	pair specified; if none is found, NULL is returned.
983  *
984  *	The object must be locked.
985  *	This routine may not block.
986  *	This is a critical path routine
987  */
988 vm_page_t
989 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
990 {
991 	vm_page_t m;
992 
993 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
994 	if ((m = object->root) != NULL && m->pindex != pindex) {
995 		m = vm_page_splay(pindex, m);
996 		if ((object->root = m)->pindex != pindex)
997 			m = NULL;
998 	}
999 	return (m);
1000 }
1001 
1002 /*
1003  *	vm_page_find_least:
1004  *
1005  *	Returns the page associated with the object with least pindex
1006  *	greater than or equal to the parameter pindex, or NULL.
1007  *
1008  *	The object must be locked.
1009  *	The routine may not block.
1010  */
1011 vm_page_t
1012 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1013 {
1014 	vm_page_t m;
1015 
1016 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1017 	if ((m = TAILQ_FIRST(&object->memq)) != NULL) {
1018 		if (m->pindex < pindex) {
1019 			m = vm_page_splay(pindex, object->root);
1020 			if ((object->root = m)->pindex < pindex)
1021 				m = TAILQ_NEXT(m, listq);
1022 		}
1023 	}
1024 	return (m);
1025 }
1026 
1027 /*
1028  * Returns the given page's successor (by pindex) within the object if it is
1029  * resident; if none is found, NULL is returned.
1030  *
1031  * The object must be locked.
1032  */
1033 vm_page_t
1034 vm_page_next(vm_page_t m)
1035 {
1036 	vm_page_t next;
1037 
1038 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1039 	if ((next = TAILQ_NEXT(m, listq)) != NULL &&
1040 	    next->pindex != m->pindex + 1)
1041 		next = NULL;
1042 	return (next);
1043 }
1044 
1045 /*
1046  * Returns the given page's predecessor (by pindex) within the object if it is
1047  * resident; if none is found, NULL is returned.
1048  *
1049  * The object must be locked.
1050  */
1051 vm_page_t
1052 vm_page_prev(vm_page_t m)
1053 {
1054 	vm_page_t prev;
1055 
1056 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1057 	if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL &&
1058 	    prev->pindex != m->pindex - 1)
1059 		prev = NULL;
1060 	return (prev);
1061 }
1062 
1063 /*
1064  *	vm_page_rename:
1065  *
1066  *	Move the given memory entry from its
1067  *	current object to the specified target object/offset.
1068  *
1069  *	The object must be locked.
1070  *	This routine may not block.
1071  *
1072  *	Note: swap associated with the page must be invalidated by the move.  We
1073  *	      have to do this for several reasons:  (1) we aren't freeing the
1074  *	      page, (2) we are dirtying the page, (3) the VM system is probably
1075  *	      moving the page from object A to B, and will then later move
1076  *	      the backing store from A to B and we can't have a conflict.
1077  *
1078  *	Note: we *always* dirty the page.  It is necessary both for the
1079  *	      fact that we moved it, and because we may be invalidating
1080  *	      swap.  If the page is on the cache, we have to deactivate it
1081  *	      or vm_page_dirty() will panic.  Dirty pages are not allowed
1082  *	      on the cache.
1083  */
1084 void
1085 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1086 {
1087 
1088 	vm_page_remove(m);
1089 	vm_page_insert(m, new_object, new_pindex);
1090 	vm_page_dirty(m);
1091 }
1092 
1093 /*
1094  *	Convert all of the given object's cached pages that have a
1095  *	pindex within the given range into free pages.  If the value
1096  *	zero is given for "end", then the range's upper bound is
1097  *	infinity.  If the given object is backed by a vnode and it
1098  *	transitions from having one or more cached pages to none, the
1099  *	vnode's hold count is reduced.
1100  */
1101 void
1102 vm_page_cache_free(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1103 {
1104 	vm_page_t m, m_next;
1105 	boolean_t empty;
1106 
1107 	mtx_lock(&vm_page_queue_free_mtx);
1108 	if (__predict_false(object->cache == NULL)) {
1109 		mtx_unlock(&vm_page_queue_free_mtx);
1110 		return;
1111 	}
1112 	m = object->cache = vm_page_splay(start, object->cache);
1113 	if (m->pindex < start) {
1114 		if (m->right == NULL)
1115 			m = NULL;
1116 		else {
1117 			m_next = vm_page_splay(start, m->right);
1118 			m_next->left = m;
1119 			m->right = NULL;
1120 			m = object->cache = m_next;
1121 		}
1122 	}
1123 
1124 	/*
1125 	 * At this point, "m" is either (1) a reference to the page
1126 	 * with the least pindex that is greater than or equal to
1127 	 * "start" or (2) NULL.
1128 	 */
1129 	for (; m != NULL && (m->pindex < end || end == 0); m = m_next) {
1130 		/*
1131 		 * Find "m"'s successor and remove "m" from the
1132 		 * object's cache.
1133 		 */
1134 		if (m->right == NULL) {
1135 			object->cache = m->left;
1136 			m_next = NULL;
1137 		} else {
1138 			m_next = vm_page_splay(start, m->right);
1139 			m_next->left = m->left;
1140 			object->cache = m_next;
1141 		}
1142 		/* Convert "m" to a free page. */
1143 		m->object = NULL;
1144 		m->valid = 0;
1145 		/* Clear PG_CACHED and set PG_FREE. */
1146 		m->flags ^= PG_CACHED | PG_FREE;
1147 		KASSERT((m->flags & (PG_CACHED | PG_FREE)) == PG_FREE,
1148 		    ("vm_page_cache_free: page %p has inconsistent flags", m));
1149 		cnt.v_cache_count--;
1150 		cnt.v_free_count++;
1151 	}
1152 	empty = object->cache == NULL;
1153 	mtx_unlock(&vm_page_queue_free_mtx);
1154 	if (object->type == OBJT_VNODE && empty)
1155 		vdrop(object->handle);
1156 }
1157 
1158 /*
1159  *	Returns the cached page that is associated with the given
1160  *	object and offset.  If, however, none exists, returns NULL.
1161  *
1162  *	The free page queue must be locked.
1163  */
1164 static inline vm_page_t
1165 vm_page_cache_lookup(vm_object_t object, vm_pindex_t pindex)
1166 {
1167 	vm_page_t m;
1168 
1169 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1170 	if ((m = object->cache) != NULL && m->pindex != pindex) {
1171 		m = vm_page_splay(pindex, m);
1172 		if ((object->cache = m)->pindex != pindex)
1173 			m = NULL;
1174 	}
1175 	return (m);
1176 }
1177 
1178 /*
1179  *	Remove the given cached page from its containing object's
1180  *	collection of cached pages.
1181  *
1182  *	The free page queue must be locked.
1183  */
1184 static void
1185 vm_page_cache_remove(vm_page_t m)
1186 {
1187 	vm_object_t object;
1188 	vm_page_t root;
1189 
1190 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1191 	KASSERT((m->flags & PG_CACHED) != 0,
1192 	    ("vm_page_cache_remove: page %p is not cached", m));
1193 	object = m->object;
1194 	if (m != object->cache) {
1195 		root = vm_page_splay(m->pindex, object->cache);
1196 		KASSERT(root == m,
1197 		    ("vm_page_cache_remove: page %p is not cached in object %p",
1198 		    m, object));
1199 	}
1200 	if (m->left == NULL)
1201 		root = m->right;
1202 	else if (m->right == NULL)
1203 		root = m->left;
1204 	else {
1205 		root = vm_page_splay(m->pindex, m->left);
1206 		root->right = m->right;
1207 	}
1208 	object->cache = root;
1209 	m->object = NULL;
1210 	cnt.v_cache_count--;
1211 }
1212 
1213 /*
1214  *	Transfer all of the cached pages with offset greater than or
1215  *	equal to 'offidxstart' from the original object's cache to the
1216  *	new object's cache.  However, any cached pages with offset
1217  *	greater than or equal to the new object's size are kept in the
1218  *	original object.  Initially, the new object's cache must be
1219  *	empty.  Offset 'offidxstart' in the original object must
1220  *	correspond to offset zero in the new object.
1221  *
1222  *	The new object must be locked.
1223  */
1224 void
1225 vm_page_cache_transfer(vm_object_t orig_object, vm_pindex_t offidxstart,
1226     vm_object_t new_object)
1227 {
1228 	vm_page_t m, m_next;
1229 
1230 	/*
1231 	 * Insertion into an object's collection of cached pages
1232 	 * requires the object to be locked.  In contrast, removal does
1233 	 * not.
1234 	 */
1235 	VM_OBJECT_LOCK_ASSERT(new_object, MA_OWNED);
1236 	KASSERT(new_object->cache == NULL,
1237 	    ("vm_page_cache_transfer: object %p has cached pages",
1238 	    new_object));
1239 	mtx_lock(&vm_page_queue_free_mtx);
1240 	if ((m = orig_object->cache) != NULL) {
1241 		/*
1242 		 * Transfer all of the pages with offset greater than or
1243 		 * equal to 'offidxstart' from the original object's
1244 		 * cache to the new object's cache.
1245 		 */
1246 		m = vm_page_splay(offidxstart, m);
1247 		if (m->pindex < offidxstart) {
1248 			orig_object->cache = m;
1249 			new_object->cache = m->right;
1250 			m->right = NULL;
1251 		} else {
1252 			orig_object->cache = m->left;
1253 			new_object->cache = m;
1254 			m->left = NULL;
1255 		}
1256 		while ((m = new_object->cache) != NULL) {
1257 			if ((m->pindex - offidxstart) >= new_object->size) {
1258 				/*
1259 				 * Return all of the cached pages with
1260 				 * offset greater than or equal to the
1261 				 * new object's size to the original
1262 				 * object's cache.
1263 				 */
1264 				new_object->cache = m->left;
1265 				m->left = orig_object->cache;
1266 				orig_object->cache = m;
1267 				break;
1268 			}
1269 			m_next = vm_page_splay(m->pindex, m->right);
1270 			/* Update the page's object and offset. */
1271 			m->object = new_object;
1272 			m->pindex -= offidxstart;
1273 			if (m_next == NULL)
1274 				break;
1275 			m->right = NULL;
1276 			m_next->left = m;
1277 			new_object->cache = m_next;
1278 		}
1279 		KASSERT(new_object->cache == NULL ||
1280 		    new_object->type == OBJT_SWAP,
1281 		    ("vm_page_cache_transfer: object %p's type is incompatible"
1282 		    " with cached pages", new_object));
1283 	}
1284 	mtx_unlock(&vm_page_queue_free_mtx);
1285 }
1286 
1287 /*
1288  *	Returns TRUE if a cached page is associated with the given object and
1289  *	offset, and FALSE otherwise.
1290  *
1291  *	The object must be locked.
1292  */
1293 boolean_t
1294 vm_page_is_cached(vm_object_t object, vm_pindex_t pindex)
1295 {
1296 	vm_page_t m;
1297 
1298 	/*
1299 	 * Insertion into an object's collection of cached pages requires the
1300 	 * object to be locked.  Therefore, if the object is locked and the
1301 	 * object's collection is empty, there is no need to acquire the free
1302 	 * page queues lock in order to prove that the specified page doesn't
1303 	 * exist.
1304 	 */
1305 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1306 	if (__predict_true(object->cache == NULL))
1307 		return (FALSE);
1308 	mtx_lock(&vm_page_queue_free_mtx);
1309 	m = vm_page_cache_lookup(object, pindex);
1310 	mtx_unlock(&vm_page_queue_free_mtx);
1311 	return (m != NULL);
1312 }
1313 
1314 /*
1315  *	vm_page_alloc:
1316  *
1317  *	Allocate and return a page that is associated with the specified
1318  *	object and offset pair.  By default, this page has the flag VPO_BUSY
1319  *	set.
1320  *
1321  *	The caller must always specify an allocation class.
1322  *
1323  *	allocation classes:
1324  *	VM_ALLOC_NORMAL		normal process request
1325  *	VM_ALLOC_SYSTEM		system *really* needs a page
1326  *	VM_ALLOC_INTERRUPT	interrupt time request
1327  *
1328  *	optional allocation flags:
1329  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
1330  *				intends to allocate
1331  *	VM_ALLOC_IFCACHED	return page only if it is cached
1332  *	VM_ALLOC_IFNOTCACHED	return NULL, do not reactivate if the page
1333  *				is cached
1334  *	VM_ALLOC_NOBUSY		do not set the flag VPO_BUSY on the page
1335  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
1336  *	VM_ALLOC_NOOBJ		page is not associated with an object and
1337  *				should not have the flag VPO_BUSY set
1338  *	VM_ALLOC_WIRED		wire the allocated page
1339  *	VM_ALLOC_ZERO		prefer a zeroed page
1340  *
1341  *	This routine may not sleep.
1342  */
1343 vm_page_t
1344 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1345 {
1346 	struct vnode *vp = NULL;
1347 	vm_object_t m_object;
1348 	vm_page_t m;
1349 	int flags, req_class;
1350 
1351 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0),
1352 	    ("vm_page_alloc: inconsistent object/req"));
1353 	if (object != NULL)
1354 		VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1355 
1356 	req_class = req & VM_ALLOC_CLASS_MASK;
1357 
1358 	/*
1359 	 * The page daemon is allowed to dig deeper into the free page list.
1360 	 */
1361 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1362 		req_class = VM_ALLOC_SYSTEM;
1363 
1364 	mtx_lock(&vm_page_queue_free_mtx);
1365 	if (cnt.v_free_count + cnt.v_cache_count > cnt.v_free_reserved ||
1366 	    (req_class == VM_ALLOC_SYSTEM &&
1367 	    cnt.v_free_count + cnt.v_cache_count > cnt.v_interrupt_free_min) ||
1368 	    (req_class == VM_ALLOC_INTERRUPT &&
1369 	    cnt.v_free_count + cnt.v_cache_count > 0)) {
1370 		/*
1371 		 * Allocate from the free queue if the number of free pages
1372 		 * exceeds the minimum for the request class.
1373 		 */
1374 		if (object != NULL &&
1375 		    (m = vm_page_cache_lookup(object, pindex)) != NULL) {
1376 			if ((req & VM_ALLOC_IFNOTCACHED) != 0) {
1377 				mtx_unlock(&vm_page_queue_free_mtx);
1378 				return (NULL);
1379 			}
1380 			if (vm_phys_unfree_page(m))
1381 				vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m, 0);
1382 #if VM_NRESERVLEVEL > 0
1383 			else if (!vm_reserv_reactivate_page(m))
1384 #else
1385 			else
1386 #endif
1387 				panic("vm_page_alloc: cache page %p is missing"
1388 				    " from the free queue", m);
1389 		} else if ((req & VM_ALLOC_IFCACHED) != 0) {
1390 			mtx_unlock(&vm_page_queue_free_mtx);
1391 			return (NULL);
1392 #if VM_NRESERVLEVEL > 0
1393 		} else if (object == NULL || object->type == OBJT_DEVICE ||
1394 		    object->type == OBJT_SG ||
1395 		    (object->flags & OBJ_COLORED) == 0 ||
1396 		    (m = vm_reserv_alloc_page(object, pindex)) == NULL) {
1397 #else
1398 		} else {
1399 #endif
1400 			m = vm_phys_alloc_pages(object != NULL ?
1401 			    VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0);
1402 #if VM_NRESERVLEVEL > 0
1403 			if (m == NULL && vm_reserv_reclaim_inactive()) {
1404 				m = vm_phys_alloc_pages(object != NULL ?
1405 				    VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT,
1406 				    0);
1407 			}
1408 #endif
1409 		}
1410 	} else {
1411 		/*
1412 		 * Not allocatable, give up.
1413 		 */
1414 		mtx_unlock(&vm_page_queue_free_mtx);
1415 		atomic_add_int(&vm_pageout_deficit,
1416 		    max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
1417 		pagedaemon_wakeup();
1418 		return (NULL);
1419 	}
1420 
1421 	/*
1422 	 *  At this point we had better have found a good page.
1423 	 */
1424 	KASSERT(m != NULL, ("vm_page_alloc: missing page"));
1425 	KASSERT(m->queue == PQ_NONE,
1426 	    ("vm_page_alloc: page %p has unexpected queue %d", m, m->queue));
1427 	KASSERT(m->wire_count == 0, ("vm_page_alloc: page %p is wired", m));
1428 	KASSERT(m->hold_count == 0, ("vm_page_alloc: page %p is held", m));
1429 	KASSERT(m->busy == 0, ("vm_page_alloc: page %p is busy", m));
1430 	KASSERT(m->dirty == 0, ("vm_page_alloc: page %p is dirty", m));
1431 	KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
1432 	    ("vm_page_alloc: page %p has unexpected memattr %d", m,
1433 	    pmap_page_get_memattr(m)));
1434 	if ((m->flags & PG_CACHED) != 0) {
1435 		KASSERT((m->flags & PG_ZERO) == 0,
1436 		    ("vm_page_alloc: cached page %p is PG_ZERO", m));
1437 		KASSERT(m->valid != 0,
1438 		    ("vm_page_alloc: cached page %p is invalid", m));
1439 		if (m->object == object && m->pindex == pindex)
1440 	  		cnt.v_reactivated++;
1441 		else
1442 			m->valid = 0;
1443 		m_object = m->object;
1444 		vm_page_cache_remove(m);
1445 		if (m_object->type == OBJT_VNODE && m_object->cache == NULL)
1446 			vp = m_object->handle;
1447 	} else {
1448 		KASSERT(VM_PAGE_IS_FREE(m),
1449 		    ("vm_page_alloc: page %p is not free", m));
1450 		KASSERT(m->valid == 0,
1451 		    ("vm_page_alloc: free page %p is valid", m));
1452 		cnt.v_free_count--;
1453 	}
1454 
1455 	/*
1456 	 * Only the PG_ZERO flag is inherited.  The PG_CACHED or PG_FREE flag
1457 	 * must be cleared before the free page queues lock is released.
1458 	 */
1459 	flags = 0;
1460 	if (req & VM_ALLOC_NODUMP)
1461 		flags |= PG_NODUMP;
1462 	if (m->flags & PG_ZERO) {
1463 		vm_page_zero_count--;
1464 		if (req & VM_ALLOC_ZERO)
1465 			flags = PG_ZERO;
1466 	}
1467 	m->flags = flags;
1468 	mtx_unlock(&vm_page_queue_free_mtx);
1469 	m->aflags = 0;
1470 	if (object == NULL || object->type == OBJT_PHYS)
1471 		m->oflags = VPO_UNMANAGED;
1472 	else
1473 		m->oflags = 0;
1474 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ)) == 0)
1475 		m->oflags |= VPO_BUSY;
1476 	if (req & VM_ALLOC_WIRED) {
1477 		/*
1478 		 * The page lock is not required for wiring a page until that
1479 		 * page is inserted into the object.
1480 		 */
1481 		atomic_add_int(&cnt.v_wire_count, 1);
1482 		m->wire_count = 1;
1483 	}
1484 	m->act_count = 0;
1485 
1486 	if (object != NULL) {
1487 		/* Ignore device objects; the pager sets "memattr" for them. */
1488 		if (object->memattr != VM_MEMATTR_DEFAULT &&
1489 		    object->type != OBJT_DEVICE && object->type != OBJT_SG)
1490 			pmap_page_set_memattr(m, object->memattr);
1491 		vm_page_insert(m, object, pindex);
1492 	} else
1493 		m->pindex = pindex;
1494 
1495 	/*
1496 	 * The following call to vdrop() must come after the above call
1497 	 * to vm_page_insert() in case both affect the same object and
1498 	 * vnode.  Otherwise, the affected vnode's hold count could
1499 	 * temporarily become zero.
1500 	 */
1501 	if (vp != NULL)
1502 		vdrop(vp);
1503 
1504 	/*
1505 	 * Don't wakeup too often - wakeup the pageout daemon when
1506 	 * we would be nearly out of memory.
1507 	 */
1508 	if (vm_paging_needed())
1509 		pagedaemon_wakeup();
1510 
1511 	return (m);
1512 }
1513 
1514 /*
1515  *	vm_page_alloc_contig:
1516  *
1517  *	Allocate a contiguous set of physical pages of the given size "npages"
1518  *	from the free lists.  All of the physical pages must be at or above
1519  *	the given physical address "low" and below the given physical address
1520  *	"high".  The given value "alignment" determines the alignment of the
1521  *	first physical page in the set.  If the given value "boundary" is
1522  *	non-zero, then the set of physical pages cannot cross any physical
1523  *	address boundary that is a multiple of that value.  Both "alignment"
1524  *	and "boundary" must be a power of two.
1525  *
1526  *	If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
1527  *	then the memory attribute setting for the physical pages is configured
1528  *	to the object's memory attribute setting.  Otherwise, the memory
1529  *	attribute setting for the physical pages is configured to "memattr",
1530  *	overriding the object's memory attribute setting.  However, if the
1531  *	object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
1532  *	memory attribute setting for the physical pages cannot be configured
1533  *	to VM_MEMATTR_DEFAULT.
1534  *
1535  *	The caller must always specify an allocation class.
1536  *
1537  *	allocation classes:
1538  *	VM_ALLOC_NORMAL		normal process request
1539  *	VM_ALLOC_SYSTEM		system *really* needs a page
1540  *	VM_ALLOC_INTERRUPT	interrupt time request
1541  *
1542  *	optional allocation flags:
1543  *	VM_ALLOC_NOBUSY		do not set the flag VPO_BUSY on the page
1544  *	VM_ALLOC_NOOBJ		page is not associated with an object and
1545  *				should not have the flag VPO_BUSY set
1546  *	VM_ALLOC_WIRED		wire the allocated page
1547  *	VM_ALLOC_ZERO		prefer a zeroed page
1548  *
1549  *	This routine may not sleep.
1550  */
1551 vm_page_t
1552 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
1553     u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
1554     vm_paddr_t boundary, vm_memattr_t memattr)
1555 {
1556 	struct vnode *drop;
1557 	vm_page_t deferred_vdrop_list, m, m_ret;
1558 	u_int flags, oflags;
1559 	int req_class;
1560 
1561 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0),
1562 	    ("vm_page_alloc_contig: inconsistent object/req"));
1563 	if (object != NULL) {
1564 		VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1565 		KASSERT(object->type == OBJT_PHYS,
1566 		    ("vm_page_alloc_contig: object %p isn't OBJT_PHYS",
1567 		    object));
1568 	}
1569 	KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
1570 	req_class = req & VM_ALLOC_CLASS_MASK;
1571 
1572 	/*
1573 	 * The page daemon is allowed to dig deeper into the free page list.
1574 	 */
1575 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1576 		req_class = VM_ALLOC_SYSTEM;
1577 
1578 	deferred_vdrop_list = NULL;
1579 	mtx_lock(&vm_page_queue_free_mtx);
1580 	if (cnt.v_free_count + cnt.v_cache_count >= npages +
1581 	    cnt.v_free_reserved || (req_class == VM_ALLOC_SYSTEM &&
1582 	    cnt.v_free_count + cnt.v_cache_count >= npages +
1583 	    cnt.v_interrupt_free_min) || (req_class == VM_ALLOC_INTERRUPT &&
1584 	    cnt.v_free_count + cnt.v_cache_count >= npages)) {
1585 #if VM_NRESERVLEVEL > 0
1586 retry:
1587 		if (object == NULL || (object->flags & OBJ_COLORED) == 0 ||
1588 		    (m_ret = vm_reserv_alloc_contig(object, pindex, npages,
1589 		    low, high, alignment, boundary)) == NULL)
1590 #endif
1591 			m_ret = vm_phys_alloc_contig(npages, low, high,
1592 			    alignment, boundary);
1593 	} else {
1594 		mtx_unlock(&vm_page_queue_free_mtx);
1595 		atomic_add_int(&vm_pageout_deficit, npages);
1596 		pagedaemon_wakeup();
1597 		return (NULL);
1598 	}
1599 	if (m_ret != NULL)
1600 		for (m = m_ret; m < &m_ret[npages]; m++) {
1601 			drop = vm_page_alloc_init(m);
1602 			if (drop != NULL) {
1603 				/*
1604 				 * Enqueue the vnode for deferred vdrop().
1605 				 *
1606 				 * Once the pages are removed from the free
1607 				 * page list, "pageq" can be safely abused to
1608 				 * construct a short-lived list of vnodes.
1609 				 */
1610 				m->pageq.tqe_prev = (void *)drop;
1611 				m->pageq.tqe_next = deferred_vdrop_list;
1612 				deferred_vdrop_list = m;
1613 			}
1614 		}
1615 	else {
1616 #if VM_NRESERVLEVEL > 0
1617 		if (vm_reserv_reclaim_contig(npages, low, high, alignment,
1618 		    boundary))
1619 			goto retry;
1620 #endif
1621 	}
1622 	mtx_unlock(&vm_page_queue_free_mtx);
1623 	if (m_ret == NULL)
1624 		return (NULL);
1625 
1626 	/*
1627 	 * Initialize the pages.  Only the PG_ZERO flag is inherited.
1628 	 */
1629 	flags = 0;
1630 	if ((req & VM_ALLOC_ZERO) != 0)
1631 		flags = PG_ZERO;
1632 	if ((req & VM_ALLOC_NODUMP) != 0)
1633 		flags |= PG_NODUMP;
1634 	if ((req & VM_ALLOC_WIRED) != 0)
1635 		atomic_add_int(&cnt.v_wire_count, npages);
1636 	oflags = VPO_UNMANAGED;
1637 	if (object != NULL) {
1638 		if ((req & VM_ALLOC_NOBUSY) == 0)
1639 			oflags |= VPO_BUSY;
1640 		if (object->memattr != VM_MEMATTR_DEFAULT &&
1641 		    memattr == VM_MEMATTR_DEFAULT)
1642 			memattr = object->memattr;
1643 	}
1644 	for (m = m_ret; m < &m_ret[npages]; m++) {
1645 		m->aflags = 0;
1646 		m->flags &= flags;
1647 		if ((req & VM_ALLOC_WIRED) != 0)
1648 			m->wire_count = 1;
1649 		/* Unmanaged pages don't use "act_count". */
1650 		m->oflags = oflags;
1651 		if (memattr != VM_MEMATTR_DEFAULT)
1652 			pmap_page_set_memattr(m, memattr);
1653 		if (object != NULL)
1654 			vm_page_insert(m, object, pindex);
1655 		else
1656 			m->pindex = pindex;
1657 		pindex++;
1658 	}
1659 	while (deferred_vdrop_list != NULL) {
1660 		vdrop((struct vnode *)deferred_vdrop_list->pageq.tqe_prev);
1661 		deferred_vdrop_list = deferred_vdrop_list->pageq.tqe_next;
1662 	}
1663 	if (vm_paging_needed())
1664 		pagedaemon_wakeup();
1665 	return (m_ret);
1666 }
1667 
1668 /*
1669  * Initialize a page that has been freshly dequeued from a freelist.
1670  * The caller has to drop the vnode returned, if it is not NULL.
1671  *
1672  * This function may only be used to initialize unmanaged pages.
1673  *
1674  * To be called with vm_page_queue_free_mtx held.
1675  */
1676 static struct vnode *
1677 vm_page_alloc_init(vm_page_t m)
1678 {
1679 	struct vnode *drop;
1680 	vm_object_t m_object;
1681 
1682 	KASSERT(m->queue == PQ_NONE,
1683 	    ("vm_page_alloc_init: page %p has unexpected queue %d",
1684 	    m, m->queue));
1685 	KASSERT(m->wire_count == 0,
1686 	    ("vm_page_alloc_init: page %p is wired", m));
1687 	KASSERT(m->hold_count == 0,
1688 	    ("vm_page_alloc_init: page %p is held", m));
1689 	KASSERT(m->busy == 0,
1690 	    ("vm_page_alloc_init: page %p is busy", m));
1691 	KASSERT(m->dirty == 0,
1692 	    ("vm_page_alloc_init: page %p is dirty", m));
1693 	KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
1694 	    ("vm_page_alloc_init: page %p has unexpected memattr %d",
1695 	    m, pmap_page_get_memattr(m)));
1696 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1697 	drop = NULL;
1698 	if ((m->flags & PG_CACHED) != 0) {
1699 		KASSERT((m->flags & PG_ZERO) == 0,
1700 		    ("vm_page_alloc_init: cached page %p is PG_ZERO", m));
1701 		m->valid = 0;
1702 		m_object = m->object;
1703 		vm_page_cache_remove(m);
1704 		if (m_object->type == OBJT_VNODE && m_object->cache == NULL)
1705 			drop = m_object->handle;
1706 	} else {
1707 		KASSERT(VM_PAGE_IS_FREE(m),
1708 		    ("vm_page_alloc_init: page %p is not free", m));
1709 		KASSERT(m->valid == 0,
1710 		    ("vm_page_alloc_init: free page %p is valid", m));
1711 		cnt.v_free_count--;
1712 		if ((m->flags & PG_ZERO) != 0)
1713 			vm_page_zero_count--;
1714 	}
1715 	/* Don't clear the PG_ZERO flag; we'll need it later. */
1716 	m->flags &= PG_ZERO;
1717 	return (drop);
1718 }
1719 
1720 /*
1721  * 	vm_page_alloc_freelist:
1722  *
1723  *	Allocate a physical page from the specified free page list.
1724  *
1725  *	The caller must always specify an allocation class.
1726  *
1727  *	allocation classes:
1728  *	VM_ALLOC_NORMAL		normal process request
1729  *	VM_ALLOC_SYSTEM		system *really* needs a page
1730  *	VM_ALLOC_INTERRUPT	interrupt time request
1731  *
1732  *	optional allocation flags:
1733  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
1734  *				intends to allocate
1735  *	VM_ALLOC_WIRED		wire the allocated page
1736  *	VM_ALLOC_ZERO		prefer a zeroed page
1737  *
1738  *	This routine may not sleep.
1739  */
1740 vm_page_t
1741 vm_page_alloc_freelist(int flind, int req)
1742 {
1743 	struct vnode *drop;
1744 	vm_page_t m;
1745 	u_int flags;
1746 	int req_class;
1747 
1748 	req_class = req & VM_ALLOC_CLASS_MASK;
1749 
1750 	/*
1751 	 * The page daemon is allowed to dig deeper into the free page list.
1752 	 */
1753 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1754 		req_class = VM_ALLOC_SYSTEM;
1755 
1756 	/*
1757 	 * Do not allocate reserved pages unless the req has asked for it.
1758 	 */
1759 	mtx_lock(&vm_page_queue_free_mtx);
1760 	if (cnt.v_free_count + cnt.v_cache_count > cnt.v_free_reserved ||
1761 	    (req_class == VM_ALLOC_SYSTEM &&
1762 	    cnt.v_free_count + cnt.v_cache_count > cnt.v_interrupt_free_min) ||
1763 	    (req_class == VM_ALLOC_INTERRUPT &&
1764 	    cnt.v_free_count + cnt.v_cache_count > 0))
1765 		m = vm_phys_alloc_freelist_pages(flind, VM_FREEPOOL_DIRECT, 0);
1766 	else {
1767 		mtx_unlock(&vm_page_queue_free_mtx);
1768 		atomic_add_int(&vm_pageout_deficit,
1769 		    max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
1770 		pagedaemon_wakeup();
1771 		return (NULL);
1772 	}
1773 	if (m == NULL) {
1774 		mtx_unlock(&vm_page_queue_free_mtx);
1775 		return (NULL);
1776 	}
1777 	drop = vm_page_alloc_init(m);
1778 	mtx_unlock(&vm_page_queue_free_mtx);
1779 
1780 	/*
1781 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
1782 	 */
1783 	m->aflags = 0;
1784 	flags = 0;
1785 	if ((req & VM_ALLOC_ZERO) != 0)
1786 		flags = PG_ZERO;
1787 	m->flags &= flags;
1788 	if ((req & VM_ALLOC_WIRED) != 0) {
1789 		/*
1790 		 * The page lock is not required for wiring a page that does
1791 		 * not belong to an object.
1792 		 */
1793 		atomic_add_int(&cnt.v_wire_count, 1);
1794 		m->wire_count = 1;
1795 	}
1796 	/* Unmanaged pages don't use "act_count". */
1797 	m->oflags = VPO_UNMANAGED;
1798 	if (drop != NULL)
1799 		vdrop(drop);
1800 	if (vm_paging_needed())
1801 		pagedaemon_wakeup();
1802 	return (m);
1803 }
1804 
1805 /*
1806  *	vm_wait:	(also see VM_WAIT macro)
1807  *
1808  *	Block until free pages are available for allocation
1809  *	- Called in various places before memory allocations.
1810  */
1811 void
1812 vm_wait(void)
1813 {
1814 
1815 	mtx_lock(&vm_page_queue_free_mtx);
1816 	if (curproc == pageproc) {
1817 		vm_pageout_pages_needed = 1;
1818 		msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx,
1819 		    PDROP | PSWP, "VMWait", 0);
1820 	} else {
1821 		if (!vm_pages_needed) {
1822 			vm_pages_needed = 1;
1823 			wakeup(&vm_pages_needed);
1824 		}
1825 		msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PVM,
1826 		    "vmwait", 0);
1827 	}
1828 }
1829 
1830 /*
1831  *	vm_waitpfault:	(also see VM_WAITPFAULT macro)
1832  *
1833  *	Block until free pages are available for allocation
1834  *	- Called only in vm_fault so that processes page faulting
1835  *	  can be easily tracked.
1836  *	- Sleeps at a lower priority than vm_wait() so that vm_wait()ing
1837  *	  processes will be able to grab memory first.  Do not change
1838  *	  this balance without careful testing first.
1839  */
1840 void
1841 vm_waitpfault(void)
1842 {
1843 
1844 	mtx_lock(&vm_page_queue_free_mtx);
1845 	if (!vm_pages_needed) {
1846 		vm_pages_needed = 1;
1847 		wakeup(&vm_pages_needed);
1848 	}
1849 	msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PUSER,
1850 	    "pfault", 0);
1851 }
1852 
1853 /*
1854  *	vm_page_requeue:
1855  *
1856  *	Move the given page to the tail of its present page queue.
1857  *
1858  *	The page queues must be locked.
1859  */
1860 void
1861 vm_page_requeue(vm_page_t m)
1862 {
1863 	struct vpgqueues *vpq;
1864 	int queue;
1865 
1866 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1867 	queue = m->queue;
1868 	KASSERT(queue != PQ_NONE,
1869 	    ("vm_page_requeue: page %p is not queued", m));
1870 	vpq = &vm_page_queues[queue];
1871 	TAILQ_REMOVE(&vpq->pl, m, pageq);
1872 	TAILQ_INSERT_TAIL(&vpq->pl, m, pageq);
1873 }
1874 
1875 /*
1876  *	vm_page_queue_remove:
1877  *
1878  *	Remove the given page from the specified queue.
1879  *
1880  *	The page and page queues must be locked.
1881  */
1882 static __inline void
1883 vm_page_queue_remove(int queue, vm_page_t m)
1884 {
1885 	struct vpgqueues *pq;
1886 
1887 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1888 	vm_page_lock_assert(m, MA_OWNED);
1889 	pq = &vm_page_queues[queue];
1890 	TAILQ_REMOVE(&pq->pl, m, pageq);
1891 	(*pq->cnt)--;
1892 }
1893 
1894 /*
1895  *	vm_pageq_remove:
1896  *
1897  *	Remove a page from its queue.
1898  *
1899  *	The given page must be locked.
1900  *	This routine may not block.
1901  */
1902 void
1903 vm_pageq_remove(vm_page_t m)
1904 {
1905 	int queue;
1906 
1907 	vm_page_lock_assert(m, MA_OWNED);
1908 	if ((queue = m->queue) != PQ_NONE) {
1909 		vm_page_lock_queues();
1910 		m->queue = PQ_NONE;
1911 		vm_page_queue_remove(queue, m);
1912 		vm_page_unlock_queues();
1913 	}
1914 }
1915 
1916 /*
1917  *	vm_page_enqueue:
1918  *
1919  *	Add the given page to the specified queue.
1920  *
1921  *	The page queues must be locked.
1922  */
1923 static void
1924 vm_page_enqueue(int queue, vm_page_t m)
1925 {
1926 	struct vpgqueues *vpq;
1927 
1928 	vpq = &vm_page_queues[queue];
1929 	m->queue = queue;
1930 	TAILQ_INSERT_TAIL(&vpq->pl, m, pageq);
1931 	++*vpq->cnt;
1932 }
1933 
1934 /*
1935  *	vm_page_activate:
1936  *
1937  *	Put the specified page on the active list (if appropriate).
1938  *	Ensure that act_count is at least ACT_INIT but do not otherwise
1939  *	mess with it.
1940  *
1941  *	The page must be locked.
1942  *	This routine may not block.
1943  */
1944 void
1945 vm_page_activate(vm_page_t m)
1946 {
1947 	int queue;
1948 
1949 	vm_page_lock_assert(m, MA_OWNED);
1950 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1951 	if ((queue = m->queue) != PQ_ACTIVE) {
1952 		if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
1953 			if (m->act_count < ACT_INIT)
1954 				m->act_count = ACT_INIT;
1955 			vm_page_lock_queues();
1956 			if (queue != PQ_NONE)
1957 				vm_page_queue_remove(queue, m);
1958 			vm_page_enqueue(PQ_ACTIVE, m);
1959 			vm_page_unlock_queues();
1960 		} else
1961 			KASSERT(queue == PQ_NONE,
1962 			    ("vm_page_activate: wired page %p is queued", m));
1963 	} else {
1964 		if (m->act_count < ACT_INIT)
1965 			m->act_count = ACT_INIT;
1966 	}
1967 }
1968 
1969 /*
1970  *	vm_page_free_wakeup:
1971  *
1972  *	Helper routine for vm_page_free_toq() and vm_page_cache().  This
1973  *	routine is called when a page has been added to the cache or free
1974  *	queues.
1975  *
1976  *	The page queues must be locked.
1977  *	This routine may not block.
1978  */
1979 static inline void
1980 vm_page_free_wakeup(void)
1981 {
1982 
1983 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1984 	/*
1985 	 * if pageout daemon needs pages, then tell it that there are
1986 	 * some free.
1987 	 */
1988 	if (vm_pageout_pages_needed &&
1989 	    cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) {
1990 		wakeup(&vm_pageout_pages_needed);
1991 		vm_pageout_pages_needed = 0;
1992 	}
1993 	/*
1994 	 * wakeup processes that are waiting on memory if we hit a
1995 	 * high water mark. And wakeup scheduler process if we have
1996 	 * lots of memory. this process will swapin processes.
1997 	 */
1998 	if (vm_pages_needed && !vm_page_count_min()) {
1999 		vm_pages_needed = 0;
2000 		wakeup(&cnt.v_free_count);
2001 	}
2002 }
2003 
2004 /*
2005  *	vm_page_free_toq:
2006  *
2007  *	Returns the given page to the free list,
2008  *	disassociating it with any VM object.
2009  *
2010  *	Object and page must be locked prior to entry.
2011  *	This routine may not block.
2012  */
2013 
2014 void
2015 vm_page_free_toq(vm_page_t m)
2016 {
2017 
2018 	if ((m->oflags & VPO_UNMANAGED) == 0) {
2019 		vm_page_lock_assert(m, MA_OWNED);
2020 		KASSERT(!pmap_page_is_mapped(m),
2021 		    ("vm_page_free_toq: freeing mapped page %p", m));
2022 	}
2023 	PCPU_INC(cnt.v_tfree);
2024 
2025 	if (VM_PAGE_IS_FREE(m))
2026 		panic("vm_page_free: freeing free page %p", m);
2027 	else if (m->busy != 0)
2028 		panic("vm_page_free: freeing busy page %p", m);
2029 
2030 	/*
2031 	 * unqueue, then remove page.  Note that we cannot destroy
2032 	 * the page here because we do not want to call the pager's
2033 	 * callback routine until after we've put the page on the
2034 	 * appropriate free queue.
2035 	 */
2036 	if ((m->oflags & VPO_UNMANAGED) == 0)
2037 		vm_pageq_remove(m);
2038 	vm_page_remove(m);
2039 
2040 	/*
2041 	 * If fictitious remove object association and
2042 	 * return, otherwise delay object association removal.
2043 	 */
2044 	if ((m->flags & PG_FICTITIOUS) != 0) {
2045 		return;
2046 	}
2047 
2048 	m->valid = 0;
2049 	vm_page_undirty(m);
2050 
2051 	if (m->wire_count != 0)
2052 		panic("vm_page_free: freeing wired page %p", m);
2053 	if (m->hold_count != 0) {
2054 		m->flags &= ~PG_ZERO;
2055 		vm_page_lock_queues();
2056 		vm_page_enqueue(PQ_HOLD, m);
2057 		vm_page_unlock_queues();
2058 	} else {
2059 		/*
2060 		 * Restore the default memory attribute to the page.
2061 		 */
2062 		if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
2063 			pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
2064 
2065 		/*
2066 		 * Insert the page into the physical memory allocator's
2067 		 * cache/free page queues.
2068 		 */
2069 		mtx_lock(&vm_page_queue_free_mtx);
2070 		m->flags |= PG_FREE;
2071 		cnt.v_free_count++;
2072 #if VM_NRESERVLEVEL > 0
2073 		if (!vm_reserv_free_page(m))
2074 #else
2075 		if (TRUE)
2076 #endif
2077 			vm_phys_free_pages(m, 0);
2078 		if ((m->flags & PG_ZERO) != 0)
2079 			++vm_page_zero_count;
2080 		else
2081 			vm_page_zero_idle_wakeup();
2082 		vm_page_free_wakeup();
2083 		mtx_unlock(&vm_page_queue_free_mtx);
2084 	}
2085 }
2086 
2087 /*
2088  *	vm_page_wire:
2089  *
2090  *	Mark this page as wired down by yet
2091  *	another map, removing it from paging queues
2092  *	as necessary.
2093  *
2094  *	If the page is fictitious, then its wire count must remain one.
2095  *
2096  *	The page must be locked.
2097  *	This routine may not block.
2098  */
2099 void
2100 vm_page_wire(vm_page_t m)
2101 {
2102 
2103 	/*
2104 	 * Only bump the wire statistics if the page is not already wired,
2105 	 * and only unqueue the page if it is on some queue (if it is unmanaged
2106 	 * it is already off the queues).
2107 	 */
2108 	vm_page_lock_assert(m, MA_OWNED);
2109 	if ((m->flags & PG_FICTITIOUS) != 0) {
2110 		KASSERT(m->wire_count == 1,
2111 		    ("vm_page_wire: fictitious page %p's wire count isn't one",
2112 		    m));
2113 		return;
2114 	}
2115 	if (m->wire_count == 0) {
2116 		if ((m->oflags & VPO_UNMANAGED) == 0)
2117 			vm_pageq_remove(m);
2118 		atomic_add_int(&cnt.v_wire_count, 1);
2119 	}
2120 	m->wire_count++;
2121 	KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m));
2122 }
2123 
2124 /*
2125  * vm_page_unwire:
2126  *
2127  * Release one wiring of the specified page, potentially enabling it to be
2128  * paged again.  If paging is enabled, then the value of the parameter
2129  * "activate" determines to which queue the page is added.  If "activate" is
2130  * non-zero, then the page is added to the active queue.  Otherwise, it is
2131  * added to the inactive queue.
2132  *
2133  * However, unless the page belongs to an object, it is not enqueued because
2134  * it cannot be paged out.
2135  *
2136  * If a page is fictitious, then its wire count must alway be one.
2137  *
2138  * A managed page must be locked.
2139  */
2140 void
2141 vm_page_unwire(vm_page_t m, int activate)
2142 {
2143 
2144 	if ((m->oflags & VPO_UNMANAGED) == 0)
2145 		vm_page_lock_assert(m, MA_OWNED);
2146 	if ((m->flags & PG_FICTITIOUS) != 0) {
2147 		KASSERT(m->wire_count == 1,
2148 	    ("vm_page_unwire: fictitious page %p's wire count isn't one", m));
2149 		return;
2150 	}
2151 	if (m->wire_count > 0) {
2152 		m->wire_count--;
2153 		if (m->wire_count == 0) {
2154 			atomic_subtract_int(&cnt.v_wire_count, 1);
2155 			if ((m->oflags & VPO_UNMANAGED) != 0 ||
2156 			    m->object == NULL)
2157 				return;
2158 			if (!activate)
2159 				m->flags &= ~PG_WINATCFLS;
2160 			vm_page_lock_queues();
2161 			vm_page_enqueue(activate ? PQ_ACTIVE : PQ_INACTIVE, m);
2162 			vm_page_unlock_queues();
2163 		}
2164 	} else
2165 		panic("vm_page_unwire: page %p's wire count is zero", m);
2166 }
2167 
2168 /*
2169  * Move the specified page to the inactive queue.
2170  *
2171  * Many pages placed on the inactive queue should actually go
2172  * into the cache, but it is difficult to figure out which.  What
2173  * we do instead, if the inactive target is well met, is to put
2174  * clean pages at the head of the inactive queue instead of the tail.
2175  * This will cause them to be moved to the cache more quickly and
2176  * if not actively re-referenced, reclaimed more quickly.  If we just
2177  * stick these pages at the end of the inactive queue, heavy filesystem
2178  * meta-data accesses can cause an unnecessary paging load on memory bound
2179  * processes.  This optimization causes one-time-use metadata to be
2180  * reused more quickly.
2181  *
2182  * Normally athead is 0 resulting in LRU operation.  athead is set
2183  * to 1 if we want this page to be 'as if it were placed in the cache',
2184  * except without unmapping it from the process address space.
2185  *
2186  * This routine may not block.
2187  */
2188 static inline void
2189 _vm_page_deactivate(vm_page_t m, int athead)
2190 {
2191 	int queue;
2192 
2193 	vm_page_lock_assert(m, MA_OWNED);
2194 
2195 	/*
2196 	 * Ignore if already inactive.
2197 	 */
2198 	if ((queue = m->queue) == PQ_INACTIVE)
2199 		return;
2200 	if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
2201 		m->flags &= ~PG_WINATCFLS;
2202 		vm_page_lock_queues();
2203 		if (queue != PQ_NONE)
2204 			vm_page_queue_remove(queue, m);
2205 		if (athead)
2206 			TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m,
2207 			    pageq);
2208 		else
2209 			TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m,
2210 			    pageq);
2211 		m->queue = PQ_INACTIVE;
2212 		cnt.v_inactive_count++;
2213 		vm_page_unlock_queues();
2214 	}
2215 }
2216 
2217 /*
2218  * Move the specified page to the inactive queue.
2219  *
2220  * The page must be locked.
2221  */
2222 void
2223 vm_page_deactivate(vm_page_t m)
2224 {
2225 
2226 	_vm_page_deactivate(m, 0);
2227 }
2228 
2229 /*
2230  * vm_page_try_to_cache:
2231  *
2232  * Returns 0 on failure, 1 on success
2233  */
2234 int
2235 vm_page_try_to_cache(vm_page_t m)
2236 {
2237 
2238 	vm_page_lock_assert(m, MA_OWNED);
2239 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2240 	if (m->dirty || m->hold_count || m->busy || m->wire_count ||
2241 	    (m->oflags & (VPO_BUSY | VPO_UNMANAGED)) != 0)
2242 		return (0);
2243 	pmap_remove_all(m);
2244 	if (m->dirty)
2245 		return (0);
2246 	vm_page_cache(m);
2247 	return (1);
2248 }
2249 
2250 /*
2251  * vm_page_try_to_free()
2252  *
2253  *	Attempt to free the page.  If we cannot free it, we do nothing.
2254  *	1 is returned on success, 0 on failure.
2255  */
2256 int
2257 vm_page_try_to_free(vm_page_t m)
2258 {
2259 
2260 	vm_page_lock_assert(m, MA_OWNED);
2261 	if (m->object != NULL)
2262 		VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2263 	if (m->dirty || m->hold_count || m->busy || m->wire_count ||
2264 	    (m->oflags & (VPO_BUSY | VPO_UNMANAGED)) != 0)
2265 		return (0);
2266 	pmap_remove_all(m);
2267 	if (m->dirty)
2268 		return (0);
2269 	vm_page_free(m);
2270 	return (1);
2271 }
2272 
2273 /*
2274  * vm_page_cache
2275  *
2276  * Put the specified page onto the page cache queue (if appropriate).
2277  *
2278  * This routine may not block.
2279  */
2280 void
2281 vm_page_cache(vm_page_t m)
2282 {
2283 	vm_object_t object;
2284 	vm_page_t next, prev, root;
2285 
2286 	vm_page_lock_assert(m, MA_OWNED);
2287 	object = m->object;
2288 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2289 	if ((m->oflags & (VPO_UNMANAGED | VPO_BUSY)) || m->busy ||
2290 	    m->hold_count || m->wire_count)
2291 		panic("vm_page_cache: attempting to cache busy page");
2292 	pmap_remove_all(m);
2293 	if (m->dirty != 0)
2294 		panic("vm_page_cache: page %p is dirty", m);
2295 	if (m->valid == 0 || object->type == OBJT_DEFAULT ||
2296 	    (object->type == OBJT_SWAP &&
2297 	    !vm_pager_has_page(object, m->pindex, NULL, NULL))) {
2298 		/*
2299 		 * Hypothesis: A cache-elgible page belonging to a
2300 		 * default object or swap object but without a backing
2301 		 * store must be zero filled.
2302 		 */
2303 		vm_page_free(m);
2304 		return;
2305 	}
2306 	KASSERT((m->flags & PG_CACHED) == 0,
2307 	    ("vm_page_cache: page %p is already cached", m));
2308 	PCPU_INC(cnt.v_tcached);
2309 
2310 	/*
2311 	 * Remove the page from the paging queues.
2312 	 */
2313 	vm_pageq_remove(m);
2314 
2315 	/*
2316 	 * Remove the page from the object's collection of resident
2317 	 * pages.
2318 	 */
2319 	if ((next = TAILQ_NEXT(m, listq)) != NULL && next->left == m) {
2320 		/*
2321 		 * Since the page's successor in the list is also its parent
2322 		 * in the tree, its right subtree must be empty.
2323 		 */
2324 		next->left = m->left;
2325 		KASSERT(m->right == NULL,
2326 		    ("vm_page_cache: page %p has right child", m));
2327 	} else if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL &&
2328 	    prev->right == m) {
2329 		/*
2330 		 * Since the page's predecessor in the list is also its parent
2331 		 * in the tree, its left subtree must be empty.
2332 		 */
2333 		KASSERT(m->left == NULL,
2334 		    ("vm_page_cache: page %p has left child", m));
2335 		prev->right = m->right;
2336 	} else {
2337 		if (m != object->root)
2338 			vm_page_splay(m->pindex, object->root);
2339 		if (m->left == NULL)
2340 			root = m->right;
2341 		else if (m->right == NULL)
2342 			root = m->left;
2343 		else {
2344 			/*
2345 			 * Move the page's successor to the root, because
2346 			 * pages are usually removed in ascending order.
2347 			 */
2348 			if (m->right != next)
2349 				vm_page_splay(m->pindex, m->right);
2350 			next->left = m->left;
2351 			root = next;
2352 		}
2353 		object->root = root;
2354 	}
2355 	TAILQ_REMOVE(&object->memq, m, listq);
2356 	object->resident_page_count--;
2357 
2358 	/*
2359 	 * Restore the default memory attribute to the page.
2360 	 */
2361 	if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
2362 		pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
2363 
2364 	/*
2365 	 * Insert the page into the object's collection of cached pages
2366 	 * and the physical memory allocator's cache/free page queues.
2367 	 */
2368 	m->flags &= ~PG_ZERO;
2369 	mtx_lock(&vm_page_queue_free_mtx);
2370 	m->flags |= PG_CACHED;
2371 	cnt.v_cache_count++;
2372 	root = object->cache;
2373 	if (root == NULL) {
2374 		m->left = NULL;
2375 		m->right = NULL;
2376 	} else {
2377 		root = vm_page_splay(m->pindex, root);
2378 		if (m->pindex < root->pindex) {
2379 			m->left = root->left;
2380 			m->right = root;
2381 			root->left = NULL;
2382 		} else if (__predict_false(m->pindex == root->pindex))
2383 			panic("vm_page_cache: offset already cached");
2384 		else {
2385 			m->right = root->right;
2386 			m->left = root;
2387 			root->right = NULL;
2388 		}
2389 	}
2390 	object->cache = m;
2391 #if VM_NRESERVLEVEL > 0
2392 	if (!vm_reserv_free_page(m)) {
2393 #else
2394 	if (TRUE) {
2395 #endif
2396 		vm_phys_set_pool(VM_FREEPOOL_CACHE, m, 0);
2397 		vm_phys_free_pages(m, 0);
2398 	}
2399 	vm_page_free_wakeup();
2400 	mtx_unlock(&vm_page_queue_free_mtx);
2401 
2402 	/*
2403 	 * Increment the vnode's hold count if this is the object's only
2404 	 * cached page.  Decrement the vnode's hold count if this was
2405 	 * the object's only resident page.
2406 	 */
2407 	if (object->type == OBJT_VNODE) {
2408 		if (root == NULL && object->resident_page_count != 0)
2409 			vhold(object->handle);
2410 		else if (root != NULL && object->resident_page_count == 0)
2411 			vdrop(object->handle);
2412 	}
2413 }
2414 
2415 /*
2416  * vm_page_dontneed
2417  *
2418  *	Cache, deactivate, or do nothing as appropriate.  This routine
2419  *	is typically used by madvise() MADV_DONTNEED.
2420  *
2421  *	Generally speaking we want to move the page into the cache so
2422  *	it gets reused quickly.  However, this can result in a silly syndrome
2423  *	due to the page recycling too quickly.  Small objects will not be
2424  *	fully cached.  On the otherhand, if we move the page to the inactive
2425  *	queue we wind up with a problem whereby very large objects
2426  *	unnecessarily blow away our inactive and cache queues.
2427  *
2428  *	The solution is to move the pages based on a fixed weighting.  We
2429  *	either leave them alone, deactivate them, or move them to the cache,
2430  *	where moving them to the cache has the highest weighting.
2431  *	By forcing some pages into other queues we eventually force the
2432  *	system to balance the queues, potentially recovering other unrelated
2433  *	space from active.  The idea is to not force this to happen too
2434  *	often.
2435  */
2436 void
2437 vm_page_dontneed(vm_page_t m)
2438 {
2439 	int dnw;
2440 	int head;
2441 
2442 	vm_page_lock_assert(m, MA_OWNED);
2443 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2444 	dnw = PCPU_GET(dnweight);
2445 	PCPU_INC(dnweight);
2446 
2447 	/*
2448 	 * Occasionally leave the page alone.
2449 	 */
2450 	if ((dnw & 0x01F0) == 0 || m->queue == PQ_INACTIVE) {
2451 		if (m->act_count >= ACT_INIT)
2452 			--m->act_count;
2453 		return;
2454 	}
2455 
2456 	/*
2457 	 * Clear any references to the page.  Otherwise, the page daemon will
2458 	 * immediately reactivate the page.
2459 	 *
2460 	 * Perform the pmap_clear_reference() first.  Otherwise, a concurrent
2461 	 * pmap operation, such as pmap_remove(), could clear a reference in
2462 	 * the pmap and set PGA_REFERENCED on the page before the
2463 	 * pmap_clear_reference() had completed.  Consequently, the page would
2464 	 * appear referenced based upon an old reference that occurred before
2465 	 * this function ran.
2466 	 */
2467 	pmap_clear_reference(m);
2468 	vm_page_aflag_clear(m, PGA_REFERENCED);
2469 
2470 	if (m->dirty == 0 && pmap_is_modified(m))
2471 		vm_page_dirty(m);
2472 
2473 	if (m->dirty || (dnw & 0x0070) == 0) {
2474 		/*
2475 		 * Deactivate the page 3 times out of 32.
2476 		 */
2477 		head = 0;
2478 	} else {
2479 		/*
2480 		 * Cache the page 28 times out of every 32.  Note that
2481 		 * the page is deactivated instead of cached, but placed
2482 		 * at the head of the queue instead of the tail.
2483 		 */
2484 		head = 1;
2485 	}
2486 	_vm_page_deactivate(m, head);
2487 }
2488 
2489 /*
2490  * Grab a page, waiting until we are waken up due to the page
2491  * changing state.  We keep on waiting, if the page continues
2492  * to be in the object.  If the page doesn't exist, first allocate it
2493  * and then conditionally zero it.
2494  *
2495  * The caller must always specify the VM_ALLOC_RETRY flag.  This is intended
2496  * to facilitate its eventual removal.
2497  *
2498  * This routine may block.
2499  */
2500 vm_page_t
2501 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
2502 {
2503 	vm_page_t m;
2504 
2505 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2506 	KASSERT((allocflags & VM_ALLOC_RETRY) != 0,
2507 	    ("vm_page_grab: VM_ALLOC_RETRY is required"));
2508 retrylookup:
2509 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
2510 		if ((m->oflags & VPO_BUSY) != 0 ||
2511 		    ((allocflags & VM_ALLOC_IGN_SBUSY) == 0 && m->busy != 0)) {
2512 			/*
2513 			 * Reference the page before unlocking and
2514 			 * sleeping so that the page daemon is less
2515 			 * likely to reclaim it.
2516 			 */
2517 			vm_page_aflag_set(m, PGA_REFERENCED);
2518 			vm_page_sleep(m, "pgrbwt");
2519 			goto retrylookup;
2520 		} else {
2521 			if ((allocflags & VM_ALLOC_WIRED) != 0) {
2522 				vm_page_lock(m);
2523 				vm_page_wire(m);
2524 				vm_page_unlock(m);
2525 			}
2526 			if ((allocflags & VM_ALLOC_NOBUSY) == 0)
2527 				vm_page_busy(m);
2528 			return (m);
2529 		}
2530 	}
2531 	m = vm_page_alloc(object, pindex, allocflags & ~(VM_ALLOC_RETRY |
2532 	    VM_ALLOC_IGN_SBUSY));
2533 	if (m == NULL) {
2534 		VM_OBJECT_UNLOCK(object);
2535 		VM_WAIT;
2536 		VM_OBJECT_LOCK(object);
2537 		goto retrylookup;
2538 	} else if (m->valid != 0)
2539 		return (m);
2540 	if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
2541 		pmap_zero_page(m);
2542 	return (m);
2543 }
2544 
2545 /*
2546  * Mapping function for valid bits or for dirty bits in
2547  * a page.  May not block.
2548  *
2549  * Inputs are required to range within a page.
2550  */
2551 vm_page_bits_t
2552 vm_page_bits(int base, int size)
2553 {
2554 	int first_bit;
2555 	int last_bit;
2556 
2557 	KASSERT(
2558 	    base + size <= PAGE_SIZE,
2559 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
2560 	);
2561 
2562 	if (size == 0)		/* handle degenerate case */
2563 		return (0);
2564 
2565 	first_bit = base >> DEV_BSHIFT;
2566 	last_bit = (base + size - 1) >> DEV_BSHIFT;
2567 
2568 	return (((vm_page_bits_t)2 << last_bit) -
2569 	    ((vm_page_bits_t)1 << first_bit));
2570 }
2571 
2572 /*
2573  *	vm_page_set_valid_range:
2574  *
2575  *	Sets portions of a page valid.  The arguments are expected
2576  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
2577  *	of any partial chunks touched by the range.  The invalid portion of
2578  *	such chunks will be zeroed.
2579  *
2580  *	(base + size) must be less then or equal to PAGE_SIZE.
2581  */
2582 void
2583 vm_page_set_valid_range(vm_page_t m, int base, int size)
2584 {
2585 	int endoff, frag;
2586 
2587 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2588 	if (size == 0)	/* handle degenerate case */
2589 		return;
2590 
2591 	/*
2592 	 * If the base is not DEV_BSIZE aligned and the valid
2593 	 * bit is clear, we have to zero out a portion of the
2594 	 * first block.
2595 	 */
2596 	if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
2597 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
2598 		pmap_zero_page_area(m, frag, base - frag);
2599 
2600 	/*
2601 	 * If the ending offset is not DEV_BSIZE aligned and the
2602 	 * valid bit is clear, we have to zero out a portion of
2603 	 * the last block.
2604 	 */
2605 	endoff = base + size;
2606 	if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
2607 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
2608 		pmap_zero_page_area(m, endoff,
2609 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
2610 
2611 	/*
2612 	 * Assert that no previously invalid block that is now being validated
2613 	 * is already dirty.
2614 	 */
2615 	KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
2616 	    ("vm_page_set_valid_range: page %p is dirty", m));
2617 
2618 	/*
2619 	 * Set valid bits inclusive of any overlap.
2620 	 */
2621 	m->valid |= vm_page_bits(base, size);
2622 }
2623 
2624 /*
2625  * Clear the given bits from the specified page's dirty field.
2626  */
2627 static __inline void
2628 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
2629 {
2630 	uintptr_t addr;
2631 #if PAGE_SIZE < 16384
2632 	int shift;
2633 #endif
2634 
2635 	/*
2636 	 * If the object is locked and the page is neither VPO_BUSY nor
2637 	 * PGA_WRITEABLE, then the page's dirty field cannot possibly be
2638 	 * set by a concurrent pmap operation.
2639 	 */
2640 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2641 	if ((m->oflags & VPO_BUSY) == 0 && (m->aflags & PGA_WRITEABLE) == 0)
2642 		m->dirty &= ~pagebits;
2643 	else {
2644 		/*
2645 		 * The pmap layer can call vm_page_dirty() without
2646 		 * holding a distinguished lock.  The combination of
2647 		 * the object's lock and an atomic operation suffice
2648 		 * to guarantee consistency of the page dirty field.
2649 		 *
2650 		 * For PAGE_SIZE == 32768 case, compiler already
2651 		 * properly aligns the dirty field, so no forcible
2652 		 * alignment is needed. Only require existence of
2653 		 * atomic_clear_64 when page size is 32768.
2654 		 */
2655 		addr = (uintptr_t)&m->dirty;
2656 #if PAGE_SIZE == 32768
2657 		atomic_clear_64((uint64_t *)addr, pagebits);
2658 #elif PAGE_SIZE == 16384
2659 		atomic_clear_32((uint32_t *)addr, pagebits);
2660 #else		/* PAGE_SIZE <= 8192 */
2661 		/*
2662 		 * Use a trick to perform a 32-bit atomic on the
2663 		 * containing aligned word, to not depend on the existence
2664 		 * of atomic_clear_{8, 16}.
2665 		 */
2666 		shift = addr & (sizeof(uint32_t) - 1);
2667 #if BYTE_ORDER == BIG_ENDIAN
2668 		shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY;
2669 #else
2670 		shift *= NBBY;
2671 #endif
2672 		addr &= ~(sizeof(uint32_t) - 1);
2673 		atomic_clear_32((uint32_t *)addr, pagebits << shift);
2674 #endif		/* PAGE_SIZE */
2675 	}
2676 }
2677 
2678 /*
2679  *	vm_page_set_validclean:
2680  *
2681  *	Sets portions of a page valid and clean.  The arguments are expected
2682  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
2683  *	of any partial chunks touched by the range.  The invalid portion of
2684  *	such chunks will be zero'd.
2685  *
2686  *	This routine may not block.
2687  *
2688  *	(base + size) must be less then or equal to PAGE_SIZE.
2689  */
2690 void
2691 vm_page_set_validclean(vm_page_t m, int base, int size)
2692 {
2693 	vm_page_bits_t oldvalid, pagebits;
2694 	int endoff, frag;
2695 
2696 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2697 	if (size == 0)	/* handle degenerate case */
2698 		return;
2699 
2700 	/*
2701 	 * If the base is not DEV_BSIZE aligned and the valid
2702 	 * bit is clear, we have to zero out a portion of the
2703 	 * first block.
2704 	 */
2705 	if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
2706 	    (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
2707 		pmap_zero_page_area(m, frag, base - frag);
2708 
2709 	/*
2710 	 * If the ending offset is not DEV_BSIZE aligned and the
2711 	 * valid bit is clear, we have to zero out a portion of
2712 	 * the last block.
2713 	 */
2714 	endoff = base + size;
2715 	if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
2716 	    (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
2717 		pmap_zero_page_area(m, endoff,
2718 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
2719 
2720 	/*
2721 	 * Set valid, clear dirty bits.  If validating the entire
2722 	 * page we can safely clear the pmap modify bit.  We also
2723 	 * use this opportunity to clear the VPO_NOSYNC flag.  If a process
2724 	 * takes a write fault on a MAP_NOSYNC memory area the flag will
2725 	 * be set again.
2726 	 *
2727 	 * We set valid bits inclusive of any overlap, but we can only
2728 	 * clear dirty bits for DEV_BSIZE chunks that are fully within
2729 	 * the range.
2730 	 */
2731 	oldvalid = m->valid;
2732 	pagebits = vm_page_bits(base, size);
2733 	m->valid |= pagebits;
2734 #if 0	/* NOT YET */
2735 	if ((frag = base & (DEV_BSIZE - 1)) != 0) {
2736 		frag = DEV_BSIZE - frag;
2737 		base += frag;
2738 		size -= frag;
2739 		if (size < 0)
2740 			size = 0;
2741 	}
2742 	pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
2743 #endif
2744 	if (base == 0 && size == PAGE_SIZE) {
2745 		/*
2746 		 * The page can only be modified within the pmap if it is
2747 		 * mapped, and it can only be mapped if it was previously
2748 		 * fully valid.
2749 		 */
2750 		if (oldvalid == VM_PAGE_BITS_ALL)
2751 			/*
2752 			 * Perform the pmap_clear_modify() first.  Otherwise,
2753 			 * a concurrent pmap operation, such as
2754 			 * pmap_protect(), could clear a modification in the
2755 			 * pmap and set the dirty field on the page before
2756 			 * pmap_clear_modify() had begun and after the dirty
2757 			 * field was cleared here.
2758 			 */
2759 			pmap_clear_modify(m);
2760 		m->dirty = 0;
2761 		m->oflags &= ~VPO_NOSYNC;
2762 	} else if (oldvalid != VM_PAGE_BITS_ALL)
2763 		m->dirty &= ~pagebits;
2764 	else
2765 		vm_page_clear_dirty_mask(m, pagebits);
2766 }
2767 
2768 void
2769 vm_page_clear_dirty(vm_page_t m, int base, int size)
2770 {
2771 
2772 	vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
2773 }
2774 
2775 /*
2776  *	vm_page_set_invalid:
2777  *
2778  *	Invalidates DEV_BSIZE'd chunks within a page.  Both the
2779  *	valid and dirty bits for the effected areas are cleared.
2780  *
2781  *	May not block.
2782  */
2783 void
2784 vm_page_set_invalid(vm_page_t m, int base, int size)
2785 {
2786 	vm_page_bits_t bits;
2787 
2788 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2789 	KASSERT((m->oflags & VPO_BUSY) == 0,
2790 	    ("vm_page_set_invalid: page %p is busy", m));
2791 	bits = vm_page_bits(base, size);
2792 	if (m->valid == VM_PAGE_BITS_ALL && bits != 0)
2793 		pmap_remove_all(m);
2794 	KASSERT(!pmap_page_is_mapped(m),
2795 	    ("vm_page_set_invalid: page %p is mapped", m));
2796 	m->valid &= ~bits;
2797 	m->dirty &= ~bits;
2798 }
2799 
2800 /*
2801  * vm_page_zero_invalid()
2802  *
2803  *	The kernel assumes that the invalid portions of a page contain
2804  *	garbage, but such pages can be mapped into memory by user code.
2805  *	When this occurs, we must zero out the non-valid portions of the
2806  *	page so user code sees what it expects.
2807  *
2808  *	Pages are most often semi-valid when the end of a file is mapped
2809  *	into memory and the file's size is not page aligned.
2810  */
2811 void
2812 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
2813 {
2814 	int b;
2815 	int i;
2816 
2817 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2818 	/*
2819 	 * Scan the valid bits looking for invalid sections that
2820 	 * must be zerod.  Invalid sub-DEV_BSIZE'd areas ( where the
2821 	 * valid bit may be set ) have already been zerod by
2822 	 * vm_page_set_validclean().
2823 	 */
2824 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
2825 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
2826 		    (m->valid & ((vm_page_bits_t)1 << i))) {
2827 			if (i > b) {
2828 				pmap_zero_page_area(m,
2829 				    b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
2830 			}
2831 			b = i + 1;
2832 		}
2833 	}
2834 
2835 	/*
2836 	 * setvalid is TRUE when we can safely set the zero'd areas
2837 	 * as being valid.  We can do this if there are no cache consistancy
2838 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
2839 	 */
2840 	if (setvalid)
2841 		m->valid = VM_PAGE_BITS_ALL;
2842 }
2843 
2844 /*
2845  *	vm_page_is_valid:
2846  *
2847  *	Is (partial) page valid?  Note that the case where size == 0
2848  *	will return FALSE in the degenerate case where the page is
2849  *	entirely invalid, and TRUE otherwise.
2850  *
2851  *	May not block.
2852  */
2853 int
2854 vm_page_is_valid(vm_page_t m, int base, int size)
2855 {
2856 	vm_page_bits_t bits;
2857 
2858 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2859 	bits = vm_page_bits(base, size);
2860 	if (m->valid && ((m->valid & bits) == bits))
2861 		return 1;
2862 	else
2863 		return 0;
2864 }
2865 
2866 /*
2867  * update dirty bits from pmap/mmu.  May not block.
2868  */
2869 void
2870 vm_page_test_dirty(vm_page_t m)
2871 {
2872 
2873 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2874 	if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
2875 		vm_page_dirty(m);
2876 }
2877 
2878 void
2879 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
2880 {
2881 
2882 	mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
2883 }
2884 
2885 void
2886 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
2887 {
2888 
2889 	mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
2890 }
2891 
2892 int
2893 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
2894 {
2895 
2896 	return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
2897 }
2898 
2899 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
2900 void
2901 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
2902 {
2903 
2904 	mtx_assert_(vm_page_lockptr(m), a, file, line);
2905 }
2906 #endif
2907 
2908 int so_zerocp_fullpage = 0;
2909 
2910 /*
2911  *	Replace the given page with a copy.  The copied page assumes
2912  *	the portion of the given page's "wire_count" that is not the
2913  *	responsibility of this copy-on-write mechanism.
2914  *
2915  *	The object containing the given page must have a non-zero
2916  *	paging-in-progress count and be locked.
2917  */
2918 void
2919 vm_page_cowfault(vm_page_t m)
2920 {
2921 	vm_page_t mnew;
2922 	vm_object_t object;
2923 	vm_pindex_t pindex;
2924 
2925 	mtx_assert(&vm_page_queue_mtx, MA_NOTOWNED);
2926 	vm_page_lock_assert(m, MA_OWNED);
2927 	object = m->object;
2928 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2929 	KASSERT(object->paging_in_progress != 0,
2930 	    ("vm_page_cowfault: object %p's paging-in-progress count is zero.",
2931 	    object));
2932 	pindex = m->pindex;
2933 
2934  retry_alloc:
2935 	pmap_remove_all(m);
2936 	vm_page_remove(m);
2937 	mnew = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL | VM_ALLOC_NOBUSY);
2938 	if (mnew == NULL) {
2939 		vm_page_insert(m, object, pindex);
2940 		vm_page_unlock(m);
2941 		VM_OBJECT_UNLOCK(object);
2942 		VM_WAIT;
2943 		VM_OBJECT_LOCK(object);
2944 		if (m == vm_page_lookup(object, pindex)) {
2945 			vm_page_lock(m);
2946 			goto retry_alloc;
2947 		} else {
2948 			/*
2949 			 * Page disappeared during the wait.
2950 			 */
2951 			return;
2952 		}
2953 	}
2954 
2955 	if (m->cow == 0) {
2956 		/*
2957 		 * check to see if we raced with an xmit complete when
2958 		 * waiting to allocate a page.  If so, put things back
2959 		 * the way they were
2960 		 */
2961 		vm_page_unlock(m);
2962 		vm_page_lock(mnew);
2963 		vm_page_free(mnew);
2964 		vm_page_unlock(mnew);
2965 		vm_page_insert(m, object, pindex);
2966 	} else { /* clear COW & copy page */
2967 		if (!so_zerocp_fullpage)
2968 			pmap_copy_page(m, mnew);
2969 		mnew->valid = VM_PAGE_BITS_ALL;
2970 		vm_page_dirty(mnew);
2971 		mnew->wire_count = m->wire_count - m->cow;
2972 		m->wire_count = m->cow;
2973 		vm_page_unlock(m);
2974 	}
2975 }
2976 
2977 void
2978 vm_page_cowclear(vm_page_t m)
2979 {
2980 
2981 	vm_page_lock_assert(m, MA_OWNED);
2982 	if (m->cow) {
2983 		m->cow--;
2984 		/*
2985 		 * let vm_fault add back write permission  lazily
2986 		 */
2987 	}
2988 	/*
2989 	 *  sf_buf_free() will free the page, so we needn't do it here
2990 	 */
2991 }
2992 
2993 int
2994 vm_page_cowsetup(vm_page_t m)
2995 {
2996 
2997 	vm_page_lock_assert(m, MA_OWNED);
2998 	if ((m->flags & PG_FICTITIOUS) != 0 ||
2999 	    (m->oflags & VPO_UNMANAGED) != 0 ||
3000 	    m->cow == USHRT_MAX - 1 || !VM_OBJECT_TRYLOCK(m->object))
3001 		return (EBUSY);
3002 	m->cow++;
3003 	pmap_remove_write(m);
3004 	VM_OBJECT_UNLOCK(m->object);
3005 	return (0);
3006 }
3007 
3008 #ifdef INVARIANTS
3009 void
3010 vm_page_object_lock_assert(vm_page_t m)
3011 {
3012 
3013 	/*
3014 	 * Certain of the page's fields may only be modified by the
3015 	 * holder of the containing object's lock or the setter of the
3016 	 * page's VPO_BUSY flag.  Unfortunately, the setter of the
3017 	 * VPO_BUSY flag is not recorded, and thus cannot be checked
3018 	 * here.
3019 	 */
3020 	if (m->object != NULL && (m->oflags & VPO_BUSY) == 0)
3021 		VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
3022 }
3023 #endif
3024 
3025 #include "opt_ddb.h"
3026 #ifdef DDB
3027 #include <sys/kernel.h>
3028 
3029 #include <ddb/ddb.h>
3030 
3031 DB_SHOW_COMMAND(page, vm_page_print_page_info)
3032 {
3033 	db_printf("cnt.v_free_count: %d\n", cnt.v_free_count);
3034 	db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count);
3035 	db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count);
3036 	db_printf("cnt.v_active_count: %d\n", cnt.v_active_count);
3037 	db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count);
3038 	db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved);
3039 	db_printf("cnt.v_free_min: %d\n", cnt.v_free_min);
3040 	db_printf("cnt.v_free_target: %d\n", cnt.v_free_target);
3041 	db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min);
3042 	db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target);
3043 }
3044 
3045 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
3046 {
3047 
3048 	db_printf("PQ_FREE:");
3049 	db_printf(" %d", cnt.v_free_count);
3050 	db_printf("\n");
3051 
3052 	db_printf("PQ_CACHE:");
3053 	db_printf(" %d", cnt.v_cache_count);
3054 	db_printf("\n");
3055 
3056 	db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n",
3057 		*vm_page_queues[PQ_ACTIVE].cnt,
3058 		*vm_page_queues[PQ_INACTIVE].cnt);
3059 }
3060 #endif /* DDB */
3061