xref: /freebsd/sys/vm/vm_page.c (revision 7cd2dcf07629713e5a3d60472cfe4701b705a167)
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 mtx_padalign vm_page_queue_mtx;
120 struct mtx_padalign vm_page_queue_free_mtx;
121 
122 struct mtx_padalign pa_lock[PA_LOCK_COUNT];
123 
124 vm_page_t vm_page_array;
125 long vm_page_array_size;
126 long first_page;
127 int vm_page_zero_count;
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 static 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 page and queue locks.
296 	 */
297 	mtx_init(&vm_page_queue_mtx, "vm page queue", NULL, MTX_DEF |
298 	    MTX_RECURSE);
299 	mtx_init(&vm_page_queue_free_mtx, "vm page free queue", NULL, MTX_DEF);
300 	for (i = 0; i < PA_LOCK_COUNT; i++)
301 		mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
302 
303 	/*
304 	 * Initialize the queue headers for the hold queue, the active queue,
305 	 * and the inactive queue.
306 	 */
307 	for (i = 0; i < PQ_COUNT; i++)
308 		TAILQ_INIT(&vm_page_queues[i].pl);
309 	vm_page_queues[PQ_INACTIVE].cnt = &cnt.v_inactive_count;
310 	vm_page_queues[PQ_ACTIVE].cnt = &cnt.v_active_count;
311 
312 	/*
313 	 * Allocate memory for use when boot strapping the kernel memory
314 	 * allocator.
315 	 */
316 	new_end = end - (boot_pages * UMA_SLAB_SIZE);
317 	new_end = trunc_page(new_end);
318 	mapped = pmap_map(&vaddr, new_end, end,
319 	    VM_PROT_READ | VM_PROT_WRITE);
320 	bzero((void *)mapped, end - new_end);
321 	uma_startup((void *)mapped, boot_pages);
322 
323 #if defined(__amd64__) || defined(__i386__) || defined(__arm__) || \
324     defined(__mips__)
325 	/*
326 	 * Allocate a bitmap to indicate that a random physical page
327 	 * needs to be included in a minidump.
328 	 *
329 	 * The amd64 port needs this to indicate which direct map pages
330 	 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
331 	 *
332 	 * However, i386 still needs this workspace internally within the
333 	 * minidump code.  In theory, they are not needed on i386, but are
334 	 * included should the sf_buf code decide to use them.
335 	 */
336 	last_pa = 0;
337 	for (i = 0; dump_avail[i + 1] != 0; i += 2)
338 		if (dump_avail[i + 1] > last_pa)
339 			last_pa = dump_avail[i + 1];
340 	page_range = last_pa / PAGE_SIZE;
341 	vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
342 	new_end -= vm_page_dump_size;
343 	vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
344 	    new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
345 	bzero((void *)vm_page_dump, vm_page_dump_size);
346 #endif
347 #ifdef __amd64__
348 	/*
349 	 * Request that the physical pages underlying the message buffer be
350 	 * included in a crash dump.  Since the message buffer is accessed
351 	 * through the direct map, they are not automatically included.
352 	 */
353 	pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
354 	last_pa = pa + round_page(msgbufsize);
355 	while (pa < last_pa) {
356 		dump_add_page(pa);
357 		pa += PAGE_SIZE;
358 	}
359 #endif
360 	/*
361 	 * Compute the number of pages of memory that will be available for
362 	 * use (taking into account the overhead of a page structure per
363 	 * page).
364 	 */
365 	first_page = low_water / PAGE_SIZE;
366 #ifdef VM_PHYSSEG_SPARSE
367 	page_range = 0;
368 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
369 		page_range += atop(phys_avail[i + 1] - phys_avail[i]);
370 #elif defined(VM_PHYSSEG_DENSE)
371 	page_range = high_water / PAGE_SIZE - first_page;
372 #else
373 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
374 #endif
375 	end = new_end;
376 
377 	/*
378 	 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
379 	 */
380 	vaddr += PAGE_SIZE;
381 
382 	/*
383 	 * Initialize the mem entry structures now, and put them in the free
384 	 * queue.
385 	 */
386 	new_end = trunc_page(end - page_range * sizeof(struct vm_page));
387 	mapped = pmap_map(&vaddr, new_end, end,
388 	    VM_PROT_READ | VM_PROT_WRITE);
389 	vm_page_array = (vm_page_t) mapped;
390 #if VM_NRESERVLEVEL > 0
391 	/*
392 	 * Allocate memory for the reservation management system's data
393 	 * structures.
394 	 */
395 	new_end = vm_reserv_startup(&vaddr, new_end, high_water);
396 #endif
397 #if defined(__amd64__) || defined(__mips__)
398 	/*
399 	 * pmap_map on amd64 and mips can come out of the direct-map, not kvm
400 	 * like i386, so the pages must be tracked for a crashdump to include
401 	 * this data.  This includes the vm_page_array and the early UMA
402 	 * bootstrap pages.
403 	 */
404 	for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE)
405 		dump_add_page(pa);
406 #endif
407 	phys_avail[biggestone + 1] = new_end;
408 
409 	/*
410 	 * Clear all of the page structures
411 	 */
412 	bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
413 	for (i = 0; i < page_range; i++)
414 		vm_page_array[i].order = VM_NFREEORDER;
415 	vm_page_array_size = page_range;
416 
417 	/*
418 	 * Initialize the physical memory allocator.
419 	 */
420 	vm_phys_init();
421 
422 	/*
423 	 * Add every available physical page that is not blacklisted to
424 	 * the free lists.
425 	 */
426 	cnt.v_page_count = 0;
427 	cnt.v_free_count = 0;
428 	list = getenv("vm.blacklist");
429 	for (i = 0; phys_avail[i + 1] != 0; i += 2) {
430 		pa = phys_avail[i];
431 		last_pa = phys_avail[i + 1];
432 		while (pa < last_pa) {
433 			if (list != NULL &&
434 			    vm_page_blacklist_lookup(list, pa))
435 				printf("Skipping page with pa 0x%jx\n",
436 				    (uintmax_t)pa);
437 			else
438 				vm_phys_add_page(pa);
439 			pa += PAGE_SIZE;
440 		}
441 	}
442 	freeenv(list);
443 #if VM_NRESERVLEVEL > 0
444 	/*
445 	 * Initialize the reservation management system.
446 	 */
447 	vm_reserv_init();
448 #endif
449 	return (vaddr);
450 }
451 
452 void
453 vm_page_reference(vm_page_t m)
454 {
455 
456 	vm_page_aflag_set(m, PGA_REFERENCED);
457 }
458 
459 void
460 vm_page_busy(vm_page_t m)
461 {
462 
463 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
464 	KASSERT((m->oflags & VPO_BUSY) == 0,
465 	    ("vm_page_busy: page already busy!!!"));
466 	m->oflags |= VPO_BUSY;
467 }
468 
469 /*
470  *      vm_page_flash:
471  *
472  *      wakeup anyone waiting for the page.
473  */
474 void
475 vm_page_flash(vm_page_t m)
476 {
477 
478 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
479 	if (m->oflags & VPO_WANTED) {
480 		m->oflags &= ~VPO_WANTED;
481 		wakeup(m);
482 	}
483 }
484 
485 /*
486  *      vm_page_wakeup:
487  *
488  *      clear the VPO_BUSY flag and wakeup anyone waiting for the
489  *      page.
490  *
491  */
492 void
493 vm_page_wakeup(vm_page_t m)
494 {
495 
496 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
497 	KASSERT(m->oflags & VPO_BUSY, ("vm_page_wakeup: page not busy!!!"));
498 	m->oflags &= ~VPO_BUSY;
499 	vm_page_flash(m);
500 }
501 
502 void
503 vm_page_io_start(vm_page_t m)
504 {
505 
506 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
507 	m->busy++;
508 }
509 
510 void
511 vm_page_io_finish(vm_page_t m)
512 {
513 
514 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
515 	KASSERT(m->busy > 0, ("vm_page_io_finish: page %p is not busy", m));
516 	m->busy--;
517 	if (m->busy == 0)
518 		vm_page_flash(m);
519 }
520 
521 /*
522  * Keep page from being freed by the page daemon
523  * much of the same effect as wiring, except much lower
524  * overhead and should be used only for *very* temporary
525  * holding ("wiring").
526  */
527 void
528 vm_page_hold(vm_page_t mem)
529 {
530 
531 	vm_page_lock_assert(mem, MA_OWNED);
532         mem->hold_count++;
533 }
534 
535 void
536 vm_page_unhold(vm_page_t mem)
537 {
538 
539 	vm_page_lock_assert(mem, MA_OWNED);
540 	--mem->hold_count;
541 	KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!"));
542 	if (mem->hold_count == 0 && (mem->flags & PG_UNHOLDFREE) != 0)
543 		vm_page_free_toq(mem);
544 }
545 
546 /*
547  *	vm_page_unhold_pages:
548  *
549  *	Unhold each of the pages that is referenced by the given array.
550  */
551 void
552 vm_page_unhold_pages(vm_page_t *ma, int count)
553 {
554 	struct mtx *mtx, *new_mtx;
555 
556 	mtx = NULL;
557 	for (; count != 0; count--) {
558 		/*
559 		 * Avoid releasing and reacquiring the same page lock.
560 		 */
561 		new_mtx = vm_page_lockptr(*ma);
562 		if (mtx != new_mtx) {
563 			if (mtx != NULL)
564 				mtx_unlock(mtx);
565 			mtx = new_mtx;
566 			mtx_lock(mtx);
567 		}
568 		vm_page_unhold(*ma);
569 		ma++;
570 	}
571 	if (mtx != NULL)
572 		mtx_unlock(mtx);
573 }
574 
575 vm_page_t
576 PHYS_TO_VM_PAGE(vm_paddr_t pa)
577 {
578 	vm_page_t m;
579 
580 #ifdef VM_PHYSSEG_SPARSE
581 	m = vm_phys_paddr_to_vm_page(pa);
582 	if (m == NULL)
583 		m = vm_phys_fictitious_to_vm_page(pa);
584 	return (m);
585 #elif defined(VM_PHYSSEG_DENSE)
586 	long pi;
587 
588 	pi = atop(pa);
589 	if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
590 		m = &vm_page_array[pi - first_page];
591 		return (m);
592 	}
593 	return (vm_phys_fictitious_to_vm_page(pa));
594 #else
595 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
596 #endif
597 }
598 
599 /*
600  *	vm_page_getfake:
601  *
602  *	Create a fictitious page with the specified physical address and
603  *	memory attribute.  The memory attribute is the only the machine-
604  *	dependent aspect of a fictitious page that must be initialized.
605  */
606 vm_page_t
607 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
608 {
609 	vm_page_t m;
610 
611 	m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
612 	vm_page_initfake(m, paddr, memattr);
613 	return (m);
614 }
615 
616 void
617 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
618 {
619 
620 	if ((m->flags & PG_FICTITIOUS) != 0) {
621 		/*
622 		 * The page's memattr might have changed since the
623 		 * previous initialization.  Update the pmap to the
624 		 * new memattr.
625 		 */
626 		goto memattr;
627 	}
628 	m->phys_addr = paddr;
629 	m->queue = PQ_NONE;
630 	/* Fictitious pages don't use "segind". */
631 	m->flags = PG_FICTITIOUS;
632 	/* Fictitious pages don't use "order" or "pool". */
633 	m->oflags = VPO_BUSY | VPO_UNMANAGED;
634 	m->wire_count = 1;
635 memattr:
636 	pmap_page_set_memattr(m, memattr);
637 }
638 
639 /*
640  *	vm_page_putfake:
641  *
642  *	Release a fictitious page.
643  */
644 void
645 vm_page_putfake(vm_page_t m)
646 {
647 
648 	KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
649 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
650 	    ("vm_page_putfake: bad page %p", m));
651 	uma_zfree(fakepg_zone, m);
652 }
653 
654 /*
655  *	vm_page_updatefake:
656  *
657  *	Update the given fictitious page to the specified physical address and
658  *	memory attribute.
659  */
660 void
661 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
662 {
663 
664 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
665 	    ("vm_page_updatefake: bad page %p", m));
666 	m->phys_addr = paddr;
667 	pmap_page_set_memattr(m, memattr);
668 }
669 
670 /*
671  *	vm_page_free:
672  *
673  *	Free a page.
674  */
675 void
676 vm_page_free(vm_page_t m)
677 {
678 
679 	m->flags &= ~PG_ZERO;
680 	vm_page_free_toq(m);
681 }
682 
683 /*
684  *	vm_page_free_zero:
685  *
686  *	Free a page to the zerod-pages queue
687  */
688 void
689 vm_page_free_zero(vm_page_t m)
690 {
691 
692 	m->flags |= PG_ZERO;
693 	vm_page_free_toq(m);
694 }
695 
696 /*
697  * Unbusy and handle the page queueing for a page from the VOP_GETPAGES()
698  * array which is not the request page.
699  */
700 void
701 vm_page_readahead_finish(vm_page_t m)
702 {
703 
704 	if (m->valid != 0) {
705 		/*
706 		 * Since the page is not the requested page, whether
707 		 * it should be activated or deactivated is not
708 		 * obvious.  Empirical results have shown that
709 		 * deactivating the page is usually the best choice,
710 		 * unless the page is wanted by another thread.
711 		 */
712 		if (m->oflags & VPO_WANTED) {
713 			vm_page_lock(m);
714 			vm_page_activate(m);
715 			vm_page_unlock(m);
716 		} else {
717 			vm_page_lock(m);
718 			vm_page_deactivate(m);
719 			vm_page_unlock(m);
720 		}
721 		vm_page_wakeup(m);
722 	} else {
723 		/*
724 		 * Free the completely invalid page.  Such page state
725 		 * occurs due to the short read operation which did
726 		 * not covered our page at all, or in case when a read
727 		 * error happens.
728 		 */
729 		vm_page_lock(m);
730 		vm_page_free(m);
731 		vm_page_unlock(m);
732 	}
733 }
734 
735 /*
736  *	vm_page_sleep:
737  *
738  *	Sleep and release the page lock.
739  *
740  *	The object containing the given page must be locked.
741  */
742 void
743 vm_page_sleep(vm_page_t m, const char *msg)
744 {
745 
746 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
747 	if (mtx_owned(vm_page_lockptr(m)))
748 		vm_page_unlock(m);
749 
750 	/*
751 	 * It's possible that while we sleep, the page will get
752 	 * unbusied and freed.  If we are holding the object
753 	 * lock, we will assume we hold a reference to the object
754 	 * such that even if m->object changes, we can re-lock
755 	 * it.
756 	 */
757 	m->oflags |= VPO_WANTED;
758 	msleep(m, VM_OBJECT_MTX(m->object), PVM, msg, 0);
759 }
760 
761 /*
762  *	vm_page_dirty_KBI:		[ internal use only ]
763  *
764  *	Set all bits in the page's dirty field.
765  *
766  *	The object containing the specified page must be locked if the
767  *	call is made from the machine-independent layer.
768  *
769  *	See vm_page_clear_dirty_mask().
770  *
771  *	This function should only be called by vm_page_dirty().
772  */
773 void
774 vm_page_dirty_KBI(vm_page_t m)
775 {
776 
777 	/* These assertions refer to this operation by its public name. */
778 	KASSERT((m->flags & PG_CACHED) == 0,
779 	    ("vm_page_dirty: page in cache!"));
780 	KASSERT(!VM_PAGE_IS_FREE(m),
781 	    ("vm_page_dirty: page is free!"));
782 	KASSERT(m->valid == VM_PAGE_BITS_ALL,
783 	    ("vm_page_dirty: page is invalid!"));
784 	m->dirty = VM_PAGE_BITS_ALL;
785 }
786 
787 /*
788  *	vm_page_splay:
789  *
790  *	Implements Sleator and Tarjan's top-down splay algorithm.  Returns
791  *	the vm_page containing the given pindex.  If, however, that
792  *	pindex is not found in the vm_object, returns a vm_page that is
793  *	adjacent to the pindex, coming before or after it.
794  */
795 vm_page_t
796 vm_page_splay(vm_pindex_t pindex, vm_page_t root)
797 {
798 	struct vm_page dummy;
799 	vm_page_t lefttreemax, righttreemin, y;
800 
801 	if (root == NULL)
802 		return (root);
803 	lefttreemax = righttreemin = &dummy;
804 	for (;; root = y) {
805 		if (pindex < root->pindex) {
806 			if ((y = root->left) == NULL)
807 				break;
808 			if (pindex < y->pindex) {
809 				/* Rotate right. */
810 				root->left = y->right;
811 				y->right = root;
812 				root = y;
813 				if ((y = root->left) == NULL)
814 					break;
815 			}
816 			/* Link into the new root's right tree. */
817 			righttreemin->left = root;
818 			righttreemin = root;
819 		} else if (pindex > root->pindex) {
820 			if ((y = root->right) == NULL)
821 				break;
822 			if (pindex > y->pindex) {
823 				/* Rotate left. */
824 				root->right = y->left;
825 				y->left = root;
826 				root = y;
827 				if ((y = root->right) == NULL)
828 					break;
829 			}
830 			/* Link into the new root's left tree. */
831 			lefttreemax->right = root;
832 			lefttreemax = root;
833 		} else
834 			break;
835 	}
836 	/* Assemble the new root. */
837 	lefttreemax->right = root->left;
838 	righttreemin->left = root->right;
839 	root->left = dummy.right;
840 	root->right = dummy.left;
841 	return (root);
842 }
843 
844 /*
845  *	vm_page_insert:		[ internal use only ]
846  *
847  *	Inserts the given mem entry into the object and object list.
848  *
849  *	The pagetables are not updated but will presumably fault the page
850  *	in if necessary, or if a kernel page the caller will at some point
851  *	enter the page into the kernel's pmap.  We are not allowed to sleep
852  *	here so we *can't* do this anyway.
853  *
854  *	The object must be locked.
855  */
856 void
857 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
858 {
859 	vm_page_t root;
860 
861 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
862 	if (m->object != NULL)
863 		panic("vm_page_insert: page already inserted");
864 
865 	/*
866 	 * Record the object/offset pair in this page
867 	 */
868 	m->object = object;
869 	m->pindex = pindex;
870 
871 	/*
872 	 * Now link into the object's ordered list of backed pages.
873 	 */
874 	root = object->root;
875 	if (root == NULL) {
876 		m->left = NULL;
877 		m->right = NULL;
878 		TAILQ_INSERT_TAIL(&object->memq, m, listq);
879 	} else {
880 		root = vm_page_splay(pindex, root);
881 		if (pindex < root->pindex) {
882 			m->left = root->left;
883 			m->right = root;
884 			root->left = NULL;
885 			TAILQ_INSERT_BEFORE(root, m, listq);
886 		} else if (pindex == root->pindex)
887 			panic("vm_page_insert: offset already allocated");
888 		else {
889 			m->right = root->right;
890 			m->left = root;
891 			root->right = NULL;
892 			TAILQ_INSERT_AFTER(&object->memq, root, m, listq);
893 		}
894 	}
895 	object->root = m;
896 
897 	/*
898 	 * Show that the object has one more resident page.
899 	 */
900 	object->resident_page_count++;
901 
902 	/*
903 	 * Hold the vnode until the last page is released.
904 	 */
905 	if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
906 		vhold(object->handle);
907 
908 	/*
909 	 * Since we are inserting a new and possibly dirty page,
910 	 * update the object's OBJ_MIGHTBEDIRTY flag.
911 	 */
912 	if (pmap_page_is_write_mapped(m))
913 		vm_object_set_writeable_dirty(object);
914 }
915 
916 /*
917  *	vm_page_remove:
918  *
919  *	Removes the given mem entry from the object/offset-page
920  *	table and the object page list, but do not invalidate/terminate
921  *	the backing store.
922  *
923  *	The underlying pmap entry (if any) is NOT removed here.
924  *
925  *	The object must be locked.  The page must be locked if it is managed.
926  */
927 void
928 vm_page_remove(vm_page_t m)
929 {
930 	vm_object_t object;
931 	vm_page_t next, prev, root;
932 
933 	if ((m->oflags & VPO_UNMANAGED) == 0)
934 		vm_page_lock_assert(m, MA_OWNED);
935 	if ((object = m->object) == NULL)
936 		return;
937 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
938 	if (m->oflags & VPO_BUSY) {
939 		m->oflags &= ~VPO_BUSY;
940 		vm_page_flash(m);
941 	}
942 
943 	/*
944 	 * Now remove from the object's list of backed pages.
945 	 */
946 	if ((next = TAILQ_NEXT(m, listq)) != NULL && next->left == m) {
947 		/*
948 		 * Since the page's successor in the list is also its parent
949 		 * in the tree, its right subtree must be empty.
950 		 */
951 		next->left = m->left;
952 		KASSERT(m->right == NULL,
953 		    ("vm_page_remove: page %p has right child", m));
954 	} else if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL &&
955 	    prev->right == m) {
956 		/*
957 		 * Since the page's predecessor in the list is also its parent
958 		 * in the tree, its left subtree must be empty.
959 		 */
960 		KASSERT(m->left == NULL,
961 		    ("vm_page_remove: page %p has left child", m));
962 		prev->right = m->right;
963 	} else {
964 		if (m != object->root)
965 			vm_page_splay(m->pindex, object->root);
966 		if (m->left == NULL)
967 			root = m->right;
968 		else if (m->right == NULL)
969 			root = m->left;
970 		else {
971 			/*
972 			 * Move the page's successor to the root, because
973 			 * pages are usually removed in ascending order.
974 			 */
975 			if (m->right != next)
976 				vm_page_splay(m->pindex, m->right);
977 			next->left = m->left;
978 			root = next;
979 		}
980 		object->root = root;
981 	}
982 	TAILQ_REMOVE(&object->memq, m, listq);
983 
984 	/*
985 	 * And show that the object has one fewer resident page.
986 	 */
987 	object->resident_page_count--;
988 
989 	/*
990 	 * The vnode may now be recycled.
991 	 */
992 	if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
993 		vdrop(object->handle);
994 
995 	m->object = NULL;
996 }
997 
998 /*
999  *	vm_page_lookup:
1000  *
1001  *	Returns the page associated with the object/offset
1002  *	pair specified; if none is found, NULL is returned.
1003  *
1004  *	The object must be locked.
1005  */
1006 vm_page_t
1007 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1008 {
1009 	vm_page_t m;
1010 
1011 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1012 	if ((m = object->root) != NULL && m->pindex != pindex) {
1013 		m = vm_page_splay(pindex, m);
1014 		if ((object->root = m)->pindex != pindex)
1015 			m = NULL;
1016 	}
1017 	return (m);
1018 }
1019 
1020 /*
1021  *	vm_page_find_least:
1022  *
1023  *	Returns the page associated with the object with least pindex
1024  *	greater than or equal to the parameter pindex, or NULL.
1025  *
1026  *	The object must be locked.
1027  */
1028 vm_page_t
1029 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1030 {
1031 	vm_page_t m;
1032 
1033 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1034 	if ((m = TAILQ_FIRST(&object->memq)) != NULL) {
1035 		if (m->pindex < pindex) {
1036 			m = vm_page_splay(pindex, object->root);
1037 			if ((object->root = m)->pindex < pindex)
1038 				m = TAILQ_NEXT(m, listq);
1039 		}
1040 	}
1041 	return (m);
1042 }
1043 
1044 /*
1045  * Returns the given page's successor (by pindex) within the object if it is
1046  * resident; if none is found, NULL is returned.
1047  *
1048  * The object must be locked.
1049  */
1050 vm_page_t
1051 vm_page_next(vm_page_t m)
1052 {
1053 	vm_page_t next;
1054 
1055 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1056 	if ((next = TAILQ_NEXT(m, listq)) != NULL &&
1057 	    next->pindex != m->pindex + 1)
1058 		next = NULL;
1059 	return (next);
1060 }
1061 
1062 /*
1063  * Returns the given page's predecessor (by pindex) within the object if it is
1064  * resident; if none is found, NULL is returned.
1065  *
1066  * The object must be locked.
1067  */
1068 vm_page_t
1069 vm_page_prev(vm_page_t m)
1070 {
1071 	vm_page_t prev;
1072 
1073 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1074 	if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL &&
1075 	    prev->pindex != m->pindex - 1)
1076 		prev = NULL;
1077 	return (prev);
1078 }
1079 
1080 /*
1081  *	vm_page_rename:
1082  *
1083  *	Move the given memory entry from its
1084  *	current object to the specified target object/offset.
1085  *
1086  *	Note: swap associated with the page must be invalidated by the move.  We
1087  *	      have to do this for several reasons:  (1) we aren't freeing the
1088  *	      page, (2) we are dirtying the page, (3) the VM system is probably
1089  *	      moving the page from object A to B, and will then later move
1090  *	      the backing store from A to B and we can't have a conflict.
1091  *
1092  *	Note: we *always* dirty the page.  It is necessary both for the
1093  *	      fact that we moved it, and because we may be invalidating
1094  *	      swap.  If the page is on the cache, we have to deactivate it
1095  *	      or vm_page_dirty() will panic.  Dirty pages are not allowed
1096  *	      on the cache.
1097  *
1098  *	The objects must be locked.  The page must be locked if it is managed.
1099  */
1100 void
1101 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1102 {
1103 
1104 	vm_page_remove(m);
1105 	vm_page_insert(m, new_object, new_pindex);
1106 	vm_page_dirty(m);
1107 }
1108 
1109 /*
1110  *	Convert all of the given object's cached pages that have a
1111  *	pindex within the given range into free pages.  If the value
1112  *	zero is given for "end", then the range's upper bound is
1113  *	infinity.  If the given object is backed by a vnode and it
1114  *	transitions from having one or more cached pages to none, the
1115  *	vnode's hold count is reduced.
1116  */
1117 void
1118 vm_page_cache_free(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1119 {
1120 	vm_page_t m, m_next;
1121 	boolean_t empty;
1122 
1123 	mtx_lock(&vm_page_queue_free_mtx);
1124 	if (__predict_false(object->cache == NULL)) {
1125 		mtx_unlock(&vm_page_queue_free_mtx);
1126 		return;
1127 	}
1128 	m = object->cache = vm_page_splay(start, object->cache);
1129 	if (m->pindex < start) {
1130 		if (m->right == NULL)
1131 			m = NULL;
1132 		else {
1133 			m_next = vm_page_splay(start, m->right);
1134 			m_next->left = m;
1135 			m->right = NULL;
1136 			m = object->cache = m_next;
1137 		}
1138 	}
1139 
1140 	/*
1141 	 * At this point, "m" is either (1) a reference to the page
1142 	 * with the least pindex that is greater than or equal to
1143 	 * "start" or (2) NULL.
1144 	 */
1145 	for (; m != NULL && (m->pindex < end || end == 0); m = m_next) {
1146 		/*
1147 		 * Find "m"'s successor and remove "m" from the
1148 		 * object's cache.
1149 		 */
1150 		if (m->right == NULL) {
1151 			object->cache = m->left;
1152 			m_next = NULL;
1153 		} else {
1154 			m_next = vm_page_splay(start, m->right);
1155 			m_next->left = m->left;
1156 			object->cache = m_next;
1157 		}
1158 		/* Convert "m" to a free page. */
1159 		m->object = NULL;
1160 		m->valid = 0;
1161 		/* Clear PG_CACHED and set PG_FREE. */
1162 		m->flags ^= PG_CACHED | PG_FREE;
1163 		KASSERT((m->flags & (PG_CACHED | PG_FREE)) == PG_FREE,
1164 		    ("vm_page_cache_free: page %p has inconsistent flags", m));
1165 		cnt.v_cache_count--;
1166 		cnt.v_free_count++;
1167 	}
1168 	empty = object->cache == NULL;
1169 	mtx_unlock(&vm_page_queue_free_mtx);
1170 	if (object->type == OBJT_VNODE && empty)
1171 		vdrop(object->handle);
1172 }
1173 
1174 /*
1175  *	Returns the cached page that is associated with the given
1176  *	object and offset.  If, however, none exists, returns NULL.
1177  *
1178  *	The free page queue must be locked.
1179  */
1180 static inline vm_page_t
1181 vm_page_cache_lookup(vm_object_t object, vm_pindex_t pindex)
1182 {
1183 	vm_page_t m;
1184 
1185 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1186 	if ((m = object->cache) != NULL && m->pindex != pindex) {
1187 		m = vm_page_splay(pindex, m);
1188 		if ((object->cache = m)->pindex != pindex)
1189 			m = NULL;
1190 	}
1191 	return (m);
1192 }
1193 
1194 /*
1195  *	Remove the given cached page from its containing object's
1196  *	collection of cached pages.
1197  *
1198  *	The free page queue must be locked.
1199  */
1200 static void
1201 vm_page_cache_remove(vm_page_t m)
1202 {
1203 	vm_object_t object;
1204 	vm_page_t root;
1205 
1206 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1207 	KASSERT((m->flags & PG_CACHED) != 0,
1208 	    ("vm_page_cache_remove: page %p is not cached", m));
1209 	object = m->object;
1210 	if (m != object->cache) {
1211 		root = vm_page_splay(m->pindex, object->cache);
1212 		KASSERT(root == m,
1213 		    ("vm_page_cache_remove: page %p is not cached in object %p",
1214 		    m, object));
1215 	}
1216 	if (m->left == NULL)
1217 		root = m->right;
1218 	else if (m->right == NULL)
1219 		root = m->left;
1220 	else {
1221 		root = vm_page_splay(m->pindex, m->left);
1222 		root->right = m->right;
1223 	}
1224 	object->cache = root;
1225 	m->object = NULL;
1226 	cnt.v_cache_count--;
1227 }
1228 
1229 /*
1230  *	Transfer all of the cached pages with offset greater than or
1231  *	equal to 'offidxstart' from the original object's cache to the
1232  *	new object's cache.  However, any cached pages with offset
1233  *	greater than or equal to the new object's size are kept in the
1234  *	original object.  Initially, the new object's cache must be
1235  *	empty.  Offset 'offidxstart' in the original object must
1236  *	correspond to offset zero in the new object.
1237  *
1238  *	The new object must be locked.
1239  */
1240 void
1241 vm_page_cache_transfer(vm_object_t orig_object, vm_pindex_t offidxstart,
1242     vm_object_t new_object)
1243 {
1244 	vm_page_t m, m_next;
1245 
1246 	/*
1247 	 * Insertion into an object's collection of cached pages
1248 	 * requires the object to be locked.  In contrast, removal does
1249 	 * not.
1250 	 */
1251 	VM_OBJECT_LOCK_ASSERT(new_object, MA_OWNED);
1252 	KASSERT(new_object->cache == NULL,
1253 	    ("vm_page_cache_transfer: object %p has cached pages",
1254 	    new_object));
1255 	mtx_lock(&vm_page_queue_free_mtx);
1256 	if ((m = orig_object->cache) != NULL) {
1257 		/*
1258 		 * Transfer all of the pages with offset greater than or
1259 		 * equal to 'offidxstart' from the original object's
1260 		 * cache to the new object's cache.
1261 		 */
1262 		m = vm_page_splay(offidxstart, m);
1263 		if (m->pindex < offidxstart) {
1264 			orig_object->cache = m;
1265 			new_object->cache = m->right;
1266 			m->right = NULL;
1267 		} else {
1268 			orig_object->cache = m->left;
1269 			new_object->cache = m;
1270 			m->left = NULL;
1271 		}
1272 		while ((m = new_object->cache) != NULL) {
1273 			if ((m->pindex - offidxstart) >= new_object->size) {
1274 				/*
1275 				 * Return all of the cached pages with
1276 				 * offset greater than or equal to the
1277 				 * new object's size to the original
1278 				 * object's cache.
1279 				 */
1280 				new_object->cache = m->left;
1281 				m->left = orig_object->cache;
1282 				orig_object->cache = m;
1283 				break;
1284 			}
1285 			m_next = vm_page_splay(m->pindex, m->right);
1286 			/* Update the page's object and offset. */
1287 			m->object = new_object;
1288 			m->pindex -= offidxstart;
1289 			if (m_next == NULL)
1290 				break;
1291 			m->right = NULL;
1292 			m_next->left = m;
1293 			new_object->cache = m_next;
1294 		}
1295 		KASSERT(new_object->cache == NULL ||
1296 		    new_object->type == OBJT_SWAP,
1297 		    ("vm_page_cache_transfer: object %p's type is incompatible"
1298 		    " with cached pages", new_object));
1299 	}
1300 	mtx_unlock(&vm_page_queue_free_mtx);
1301 }
1302 
1303 /*
1304  *	Returns TRUE if a cached page is associated with the given object and
1305  *	offset, and FALSE otherwise.
1306  *
1307  *	The object must be locked.
1308  */
1309 boolean_t
1310 vm_page_is_cached(vm_object_t object, vm_pindex_t pindex)
1311 {
1312 	vm_page_t m;
1313 
1314 	/*
1315 	 * Insertion into an object's collection of cached pages requires the
1316 	 * object to be locked.  Therefore, if the object is locked and the
1317 	 * object's collection is empty, there is no need to acquire the free
1318 	 * page queues lock in order to prove that the specified page doesn't
1319 	 * exist.
1320 	 */
1321 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1322 	if (__predict_true(object->cache == NULL))
1323 		return (FALSE);
1324 	mtx_lock(&vm_page_queue_free_mtx);
1325 	m = vm_page_cache_lookup(object, pindex);
1326 	mtx_unlock(&vm_page_queue_free_mtx);
1327 	return (m != NULL);
1328 }
1329 
1330 /*
1331  *	vm_page_alloc:
1332  *
1333  *	Allocate and return a page that is associated with the specified
1334  *	object and offset pair.  By default, this page has the flag VPO_BUSY
1335  *	set.
1336  *
1337  *	The caller must always specify an allocation class.
1338  *
1339  *	allocation classes:
1340  *	VM_ALLOC_NORMAL		normal process request
1341  *	VM_ALLOC_SYSTEM		system *really* needs a page
1342  *	VM_ALLOC_INTERRUPT	interrupt time request
1343  *
1344  *	optional allocation flags:
1345  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
1346  *				intends to allocate
1347  *	VM_ALLOC_IFCACHED	return page only if it is cached
1348  *	VM_ALLOC_IFNOTCACHED	return NULL, do not reactivate if the page
1349  *				is cached
1350  *	VM_ALLOC_NOBUSY		do not set the flag VPO_BUSY on the page
1351  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
1352  *	VM_ALLOC_NOOBJ		page is not associated with an object and
1353  *				should not have the flag VPO_BUSY set
1354  *	VM_ALLOC_WIRED		wire the allocated page
1355  *	VM_ALLOC_ZERO		prefer a zeroed page
1356  *
1357  *	This routine may not sleep.
1358  */
1359 vm_page_t
1360 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1361 {
1362 	struct vnode *vp = NULL;
1363 	vm_object_t m_object;
1364 	vm_page_t m;
1365 	int flags, req_class;
1366 
1367 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0),
1368 	    ("vm_page_alloc: inconsistent object/req"));
1369 	if (object != NULL)
1370 		VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1371 
1372 	req_class = req & VM_ALLOC_CLASS_MASK;
1373 
1374 	/*
1375 	 * The page daemon is allowed to dig deeper into the free page list.
1376 	 */
1377 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1378 		req_class = VM_ALLOC_SYSTEM;
1379 
1380 	mtx_lock(&vm_page_queue_free_mtx);
1381 	if (cnt.v_free_count + cnt.v_cache_count > cnt.v_free_reserved ||
1382 	    (req_class == VM_ALLOC_SYSTEM &&
1383 	    cnt.v_free_count + cnt.v_cache_count > cnt.v_interrupt_free_min) ||
1384 	    (req_class == VM_ALLOC_INTERRUPT &&
1385 	    cnt.v_free_count + cnt.v_cache_count > 0)) {
1386 		/*
1387 		 * Allocate from the free queue if the number of free pages
1388 		 * exceeds the minimum for the request class.
1389 		 */
1390 		if (object != NULL &&
1391 		    (m = vm_page_cache_lookup(object, pindex)) != NULL) {
1392 			if ((req & VM_ALLOC_IFNOTCACHED) != 0) {
1393 				mtx_unlock(&vm_page_queue_free_mtx);
1394 				return (NULL);
1395 			}
1396 			if (vm_phys_unfree_page(m))
1397 				vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m, 0);
1398 #if VM_NRESERVLEVEL > 0
1399 			else if (!vm_reserv_reactivate_page(m))
1400 #else
1401 			else
1402 #endif
1403 				panic("vm_page_alloc: cache page %p is missing"
1404 				    " from the free queue", m);
1405 		} else if ((req & VM_ALLOC_IFCACHED) != 0) {
1406 			mtx_unlock(&vm_page_queue_free_mtx);
1407 			return (NULL);
1408 #if VM_NRESERVLEVEL > 0
1409 		} else if (object == NULL || object->type == OBJT_DEVICE ||
1410 		    object->type == OBJT_SG ||
1411 		    (object->flags & OBJ_COLORED) == 0 ||
1412 		    (m = vm_reserv_alloc_page(object, pindex)) == NULL) {
1413 #else
1414 		} else {
1415 #endif
1416 			m = vm_phys_alloc_pages(object != NULL ?
1417 			    VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0);
1418 #if VM_NRESERVLEVEL > 0
1419 			if (m == NULL && vm_reserv_reclaim_inactive()) {
1420 				m = vm_phys_alloc_pages(object != NULL ?
1421 				    VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT,
1422 				    0);
1423 			}
1424 #endif
1425 		}
1426 	} else {
1427 		/*
1428 		 * Not allocatable, give up.
1429 		 */
1430 		mtx_unlock(&vm_page_queue_free_mtx);
1431 		atomic_add_int(&vm_pageout_deficit,
1432 		    max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
1433 		pagedaemon_wakeup();
1434 		return (NULL);
1435 	}
1436 
1437 	/*
1438 	 *  At this point we had better have found a good page.
1439 	 */
1440 	KASSERT(m != NULL, ("vm_page_alloc: missing page"));
1441 	KASSERT(m->queue == PQ_NONE,
1442 	    ("vm_page_alloc: page %p has unexpected queue %d", m, m->queue));
1443 	KASSERT(m->wire_count == 0, ("vm_page_alloc: page %p is wired", m));
1444 	KASSERT(m->hold_count == 0, ("vm_page_alloc: page %p is held", m));
1445 	KASSERT(m->busy == 0, ("vm_page_alloc: page %p is busy", m));
1446 	KASSERT(m->dirty == 0, ("vm_page_alloc: page %p is dirty", m));
1447 	KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
1448 	    ("vm_page_alloc: page %p has unexpected memattr %d", m,
1449 	    pmap_page_get_memattr(m)));
1450 	if ((m->flags & PG_CACHED) != 0) {
1451 		KASSERT((m->flags & PG_ZERO) == 0,
1452 		    ("vm_page_alloc: cached page %p is PG_ZERO", m));
1453 		KASSERT(m->valid != 0,
1454 		    ("vm_page_alloc: cached page %p is invalid", m));
1455 		if (m->object == object && m->pindex == pindex)
1456 	  		cnt.v_reactivated++;
1457 		else
1458 			m->valid = 0;
1459 		m_object = m->object;
1460 		vm_page_cache_remove(m);
1461 		if (m_object->type == OBJT_VNODE && m_object->cache == NULL)
1462 			vp = m_object->handle;
1463 	} else {
1464 		KASSERT(VM_PAGE_IS_FREE(m),
1465 		    ("vm_page_alloc: page %p is not free", m));
1466 		KASSERT(m->valid == 0,
1467 		    ("vm_page_alloc: free page %p is valid", m));
1468 		cnt.v_free_count--;
1469 	}
1470 
1471 	/*
1472 	 * Only the PG_ZERO flag is inherited.  The PG_CACHED or PG_FREE flag
1473 	 * must be cleared before the free page queues lock is released.
1474 	 */
1475 	flags = 0;
1476 	if (req & VM_ALLOC_NODUMP)
1477 		flags |= PG_NODUMP;
1478 	if (m->flags & PG_ZERO) {
1479 		vm_page_zero_count--;
1480 		if (req & VM_ALLOC_ZERO)
1481 			flags = PG_ZERO;
1482 	}
1483 	m->flags = flags;
1484 	mtx_unlock(&vm_page_queue_free_mtx);
1485 	m->aflags = 0;
1486 	if (object == NULL || object->type == OBJT_PHYS)
1487 		m->oflags = VPO_UNMANAGED;
1488 	else
1489 		m->oflags = 0;
1490 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ)) == 0)
1491 		m->oflags |= VPO_BUSY;
1492 	if (req & VM_ALLOC_WIRED) {
1493 		/*
1494 		 * The page lock is not required for wiring a page until that
1495 		 * page is inserted into the object.
1496 		 */
1497 		atomic_add_int(&cnt.v_wire_count, 1);
1498 		m->wire_count = 1;
1499 	}
1500 	m->act_count = 0;
1501 
1502 	if (object != NULL) {
1503 		/* Ignore device objects; the pager sets "memattr" for them. */
1504 		if (object->memattr != VM_MEMATTR_DEFAULT &&
1505 		    object->type != OBJT_DEVICE && object->type != OBJT_SG)
1506 			pmap_page_set_memattr(m, object->memattr);
1507 		vm_page_insert(m, object, pindex);
1508 	} else
1509 		m->pindex = pindex;
1510 
1511 	/*
1512 	 * The following call to vdrop() must come after the above call
1513 	 * to vm_page_insert() in case both affect the same object and
1514 	 * vnode.  Otherwise, the affected vnode's hold count could
1515 	 * temporarily become zero.
1516 	 */
1517 	if (vp != NULL)
1518 		vdrop(vp);
1519 
1520 	/*
1521 	 * Don't wakeup too often - wakeup the pageout daemon when
1522 	 * we would be nearly out of memory.
1523 	 */
1524 	if (vm_paging_needed())
1525 		pagedaemon_wakeup();
1526 
1527 	return (m);
1528 }
1529 
1530 /*
1531  *	vm_page_alloc_contig:
1532  *
1533  *	Allocate a contiguous set of physical pages of the given size "npages"
1534  *	from the free lists.  All of the physical pages must be at or above
1535  *	the given physical address "low" and below the given physical address
1536  *	"high".  The given value "alignment" determines the alignment of the
1537  *	first physical page in the set.  If the given value "boundary" is
1538  *	non-zero, then the set of physical pages cannot cross any physical
1539  *	address boundary that is a multiple of that value.  Both "alignment"
1540  *	and "boundary" must be a power of two.
1541  *
1542  *	If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
1543  *	then the memory attribute setting for the physical pages is configured
1544  *	to the object's memory attribute setting.  Otherwise, the memory
1545  *	attribute setting for the physical pages is configured to "memattr",
1546  *	overriding the object's memory attribute setting.  However, if the
1547  *	object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
1548  *	memory attribute setting for the physical pages cannot be configured
1549  *	to VM_MEMATTR_DEFAULT.
1550  *
1551  *	The caller must always specify an allocation class.
1552  *
1553  *	allocation classes:
1554  *	VM_ALLOC_NORMAL		normal process request
1555  *	VM_ALLOC_SYSTEM		system *really* needs a page
1556  *	VM_ALLOC_INTERRUPT	interrupt time request
1557  *
1558  *	optional allocation flags:
1559  *	VM_ALLOC_NOBUSY		do not set the flag VPO_BUSY on the page
1560  *	VM_ALLOC_NOOBJ		page is not associated with an object and
1561  *				should not have the flag VPO_BUSY set
1562  *	VM_ALLOC_WIRED		wire the allocated page
1563  *	VM_ALLOC_ZERO		prefer a zeroed page
1564  *
1565  *	This routine may not sleep.
1566  */
1567 vm_page_t
1568 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
1569     u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
1570     vm_paddr_t boundary, vm_memattr_t memattr)
1571 {
1572 	struct vnode *drop;
1573 	vm_page_t deferred_vdrop_list, m, m_ret;
1574 	u_int flags, oflags;
1575 	int req_class;
1576 
1577 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0),
1578 	    ("vm_page_alloc_contig: inconsistent object/req"));
1579 	if (object != NULL) {
1580 		VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1581 		KASSERT(object->type == OBJT_PHYS,
1582 		    ("vm_page_alloc_contig: object %p isn't OBJT_PHYS",
1583 		    object));
1584 	}
1585 	KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
1586 	req_class = req & VM_ALLOC_CLASS_MASK;
1587 
1588 	/*
1589 	 * The page daemon is allowed to dig deeper into the free page list.
1590 	 */
1591 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1592 		req_class = VM_ALLOC_SYSTEM;
1593 
1594 	deferred_vdrop_list = NULL;
1595 	mtx_lock(&vm_page_queue_free_mtx);
1596 	if (cnt.v_free_count + cnt.v_cache_count >= npages +
1597 	    cnt.v_free_reserved || (req_class == VM_ALLOC_SYSTEM &&
1598 	    cnt.v_free_count + cnt.v_cache_count >= npages +
1599 	    cnt.v_interrupt_free_min) || (req_class == VM_ALLOC_INTERRUPT &&
1600 	    cnt.v_free_count + cnt.v_cache_count >= npages)) {
1601 #if VM_NRESERVLEVEL > 0
1602 retry:
1603 		if (object == NULL || (object->flags & OBJ_COLORED) == 0 ||
1604 		    (m_ret = vm_reserv_alloc_contig(object, pindex, npages,
1605 		    low, high, alignment, boundary)) == NULL)
1606 #endif
1607 			m_ret = vm_phys_alloc_contig(npages, low, high,
1608 			    alignment, boundary);
1609 	} else {
1610 		mtx_unlock(&vm_page_queue_free_mtx);
1611 		atomic_add_int(&vm_pageout_deficit, npages);
1612 		pagedaemon_wakeup();
1613 		return (NULL);
1614 	}
1615 	if (m_ret != NULL)
1616 		for (m = m_ret; m < &m_ret[npages]; m++) {
1617 			drop = vm_page_alloc_init(m);
1618 			if (drop != NULL) {
1619 				/*
1620 				 * Enqueue the vnode for deferred vdrop().
1621 				 *
1622 				 * Once the pages are removed from the free
1623 				 * page list, "pageq" can be safely abused to
1624 				 * construct a short-lived list of vnodes.
1625 				 */
1626 				m->pageq.tqe_prev = (void *)drop;
1627 				m->pageq.tqe_next = deferred_vdrop_list;
1628 				deferred_vdrop_list = m;
1629 			}
1630 		}
1631 	else {
1632 #if VM_NRESERVLEVEL > 0
1633 		if (vm_reserv_reclaim_contig(npages, low, high, alignment,
1634 		    boundary))
1635 			goto retry;
1636 #endif
1637 	}
1638 	mtx_unlock(&vm_page_queue_free_mtx);
1639 	if (m_ret == NULL)
1640 		return (NULL);
1641 
1642 	/*
1643 	 * Initialize the pages.  Only the PG_ZERO flag is inherited.
1644 	 */
1645 	flags = 0;
1646 	if ((req & VM_ALLOC_ZERO) != 0)
1647 		flags = PG_ZERO;
1648 	if ((req & VM_ALLOC_NODUMP) != 0)
1649 		flags |= PG_NODUMP;
1650 	if ((req & VM_ALLOC_WIRED) != 0)
1651 		atomic_add_int(&cnt.v_wire_count, npages);
1652 	oflags = VPO_UNMANAGED;
1653 	if (object != NULL) {
1654 		if ((req & VM_ALLOC_NOBUSY) == 0)
1655 			oflags |= VPO_BUSY;
1656 		if (object->memattr != VM_MEMATTR_DEFAULT &&
1657 		    memattr == VM_MEMATTR_DEFAULT)
1658 			memattr = object->memattr;
1659 	}
1660 	for (m = m_ret; m < &m_ret[npages]; m++) {
1661 		m->aflags = 0;
1662 		m->flags = (m->flags | PG_NODUMP) & flags;
1663 		if ((req & VM_ALLOC_WIRED) != 0)
1664 			m->wire_count = 1;
1665 		/* Unmanaged pages don't use "act_count". */
1666 		m->oflags = oflags;
1667 		if (memattr != VM_MEMATTR_DEFAULT)
1668 			pmap_page_set_memattr(m, memattr);
1669 		if (object != NULL)
1670 			vm_page_insert(m, object, pindex);
1671 		else
1672 			m->pindex = pindex;
1673 		pindex++;
1674 	}
1675 	while (deferred_vdrop_list != NULL) {
1676 		vdrop((struct vnode *)deferred_vdrop_list->pageq.tqe_prev);
1677 		deferred_vdrop_list = deferred_vdrop_list->pageq.tqe_next;
1678 	}
1679 	if (vm_paging_needed())
1680 		pagedaemon_wakeup();
1681 	return (m_ret);
1682 }
1683 
1684 /*
1685  * Initialize a page that has been freshly dequeued from a freelist.
1686  * The caller has to drop the vnode returned, if it is not NULL.
1687  *
1688  * This function may only be used to initialize unmanaged pages.
1689  *
1690  * To be called with vm_page_queue_free_mtx held.
1691  */
1692 static struct vnode *
1693 vm_page_alloc_init(vm_page_t m)
1694 {
1695 	struct vnode *drop;
1696 	vm_object_t m_object;
1697 
1698 	KASSERT(m->queue == PQ_NONE,
1699 	    ("vm_page_alloc_init: page %p has unexpected queue %d",
1700 	    m, m->queue));
1701 	KASSERT(m->wire_count == 0,
1702 	    ("vm_page_alloc_init: page %p is wired", m));
1703 	KASSERT(m->hold_count == 0,
1704 	    ("vm_page_alloc_init: page %p is held", m));
1705 	KASSERT(m->busy == 0,
1706 	    ("vm_page_alloc_init: page %p is busy", m));
1707 	KASSERT(m->dirty == 0,
1708 	    ("vm_page_alloc_init: page %p is dirty", m));
1709 	KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
1710 	    ("vm_page_alloc_init: page %p has unexpected memattr %d",
1711 	    m, pmap_page_get_memattr(m)));
1712 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1713 	drop = NULL;
1714 	if ((m->flags & PG_CACHED) != 0) {
1715 		KASSERT((m->flags & PG_ZERO) == 0,
1716 		    ("vm_page_alloc_init: cached page %p is PG_ZERO", m));
1717 		m->valid = 0;
1718 		m_object = m->object;
1719 		vm_page_cache_remove(m);
1720 		if (m_object->type == OBJT_VNODE && m_object->cache == NULL)
1721 			drop = m_object->handle;
1722 	} else {
1723 		KASSERT(VM_PAGE_IS_FREE(m),
1724 		    ("vm_page_alloc_init: page %p is not free", m));
1725 		KASSERT(m->valid == 0,
1726 		    ("vm_page_alloc_init: free page %p is valid", m));
1727 		cnt.v_free_count--;
1728 		if ((m->flags & PG_ZERO) != 0)
1729 			vm_page_zero_count--;
1730 	}
1731 	/* Don't clear the PG_ZERO flag; we'll need it later. */
1732 	m->flags &= PG_ZERO;
1733 	return (drop);
1734 }
1735 
1736 /*
1737  * 	vm_page_alloc_freelist:
1738  *
1739  *	Allocate a physical page from the specified free page list.
1740  *
1741  *	The caller must always specify an allocation class.
1742  *
1743  *	allocation classes:
1744  *	VM_ALLOC_NORMAL		normal process request
1745  *	VM_ALLOC_SYSTEM		system *really* needs a page
1746  *	VM_ALLOC_INTERRUPT	interrupt time request
1747  *
1748  *	optional allocation flags:
1749  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
1750  *				intends to allocate
1751  *	VM_ALLOC_WIRED		wire the allocated page
1752  *	VM_ALLOC_ZERO		prefer a zeroed page
1753  *
1754  *	This routine may not sleep.
1755  */
1756 vm_page_t
1757 vm_page_alloc_freelist(int flind, int req)
1758 {
1759 	struct vnode *drop;
1760 	vm_page_t m;
1761 	u_int flags;
1762 	int req_class;
1763 
1764 	req_class = req & VM_ALLOC_CLASS_MASK;
1765 
1766 	/*
1767 	 * The page daemon is allowed to dig deeper into the free page list.
1768 	 */
1769 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1770 		req_class = VM_ALLOC_SYSTEM;
1771 
1772 	/*
1773 	 * Do not allocate reserved pages unless the req has asked for it.
1774 	 */
1775 	mtx_lock(&vm_page_queue_free_mtx);
1776 	if (cnt.v_free_count + cnt.v_cache_count > cnt.v_free_reserved ||
1777 	    (req_class == VM_ALLOC_SYSTEM &&
1778 	    cnt.v_free_count + cnt.v_cache_count > cnt.v_interrupt_free_min) ||
1779 	    (req_class == VM_ALLOC_INTERRUPT &&
1780 	    cnt.v_free_count + cnt.v_cache_count > 0))
1781 		m = vm_phys_alloc_freelist_pages(flind, VM_FREEPOOL_DIRECT, 0);
1782 	else {
1783 		mtx_unlock(&vm_page_queue_free_mtx);
1784 		atomic_add_int(&vm_pageout_deficit,
1785 		    max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
1786 		pagedaemon_wakeup();
1787 		return (NULL);
1788 	}
1789 	if (m == NULL) {
1790 		mtx_unlock(&vm_page_queue_free_mtx);
1791 		return (NULL);
1792 	}
1793 	drop = vm_page_alloc_init(m);
1794 	mtx_unlock(&vm_page_queue_free_mtx);
1795 
1796 	/*
1797 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
1798 	 */
1799 	m->aflags = 0;
1800 	flags = 0;
1801 	if ((req & VM_ALLOC_ZERO) != 0)
1802 		flags = PG_ZERO;
1803 	m->flags &= flags;
1804 	if ((req & VM_ALLOC_WIRED) != 0) {
1805 		/*
1806 		 * The page lock is not required for wiring a page that does
1807 		 * not belong to an object.
1808 		 */
1809 		atomic_add_int(&cnt.v_wire_count, 1);
1810 		m->wire_count = 1;
1811 	}
1812 	/* Unmanaged pages don't use "act_count". */
1813 	m->oflags = VPO_UNMANAGED;
1814 	if (drop != NULL)
1815 		vdrop(drop);
1816 	if (vm_paging_needed())
1817 		pagedaemon_wakeup();
1818 	return (m);
1819 }
1820 
1821 /*
1822  *	vm_wait:	(also see VM_WAIT macro)
1823  *
1824  *	Sleep until free pages are available for allocation.
1825  *	- Called in various places before memory allocations.
1826  */
1827 void
1828 vm_wait(void)
1829 {
1830 
1831 	mtx_lock(&vm_page_queue_free_mtx);
1832 	if (curproc == pageproc) {
1833 		vm_pageout_pages_needed = 1;
1834 		msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx,
1835 		    PDROP | PSWP, "VMWait", 0);
1836 	} else {
1837 		if (!vm_pages_needed) {
1838 			vm_pages_needed = 1;
1839 			wakeup(&vm_pages_needed);
1840 		}
1841 		msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PVM,
1842 		    "vmwait", 0);
1843 	}
1844 }
1845 
1846 /*
1847  *	vm_waitpfault:	(also see VM_WAITPFAULT macro)
1848  *
1849  *	Sleep until free pages are available for allocation.
1850  *	- Called only in vm_fault so that processes page faulting
1851  *	  can be easily tracked.
1852  *	- Sleeps at a lower priority than vm_wait() so that vm_wait()ing
1853  *	  processes will be able to grab memory first.  Do not change
1854  *	  this balance without careful testing first.
1855  */
1856 void
1857 vm_waitpfault(void)
1858 {
1859 
1860 	mtx_lock(&vm_page_queue_free_mtx);
1861 	if (!vm_pages_needed) {
1862 		vm_pages_needed = 1;
1863 		wakeup(&vm_pages_needed);
1864 	}
1865 	msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PUSER,
1866 	    "pfault", 0);
1867 }
1868 
1869 /*
1870  *	vm_page_queue_remove:
1871  *
1872  *	Remove the given page from the specified queue.
1873  *
1874  *	The page and page queues must be locked.
1875  */
1876 static __inline void
1877 vm_page_queue_remove(int queue, vm_page_t m)
1878 {
1879 	struct vpgqueues *pq;
1880 
1881 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1882 	vm_page_lock_assert(m, MA_OWNED);
1883 	pq = &vm_page_queues[queue];
1884 	TAILQ_REMOVE(&pq->pl, m, pageq);
1885 	(*pq->cnt)--;
1886 }
1887 
1888 /*
1889  *	vm_pageq_remove:
1890  *
1891  *	Remove a page from its queue.
1892  *
1893  *	The given page must be locked.
1894  */
1895 void
1896 vm_pageq_remove(vm_page_t m)
1897 {
1898 	int queue;
1899 
1900 	vm_page_lock_assert(m, MA_OWNED);
1901 	if ((queue = m->queue) != PQ_NONE) {
1902 		vm_page_lock_queues();
1903 		m->queue = PQ_NONE;
1904 		vm_page_queue_remove(queue, m);
1905 		vm_page_unlock_queues();
1906 	}
1907 }
1908 
1909 /*
1910  *	vm_page_enqueue:
1911  *
1912  *	Add the given page to the specified queue.
1913  *
1914  *	The page queues must be locked.
1915  */
1916 static void
1917 vm_page_enqueue(int queue, vm_page_t m)
1918 {
1919 	struct vpgqueues *vpq;
1920 
1921 	vpq = &vm_page_queues[queue];
1922 	m->queue = queue;
1923 	TAILQ_INSERT_TAIL(&vpq->pl, m, pageq);
1924 	++*vpq->cnt;
1925 }
1926 
1927 /*
1928  *	vm_page_activate:
1929  *
1930  *	Put the specified page on the active list (if appropriate).
1931  *	Ensure that act_count is at least ACT_INIT but do not otherwise
1932  *	mess with it.
1933  *
1934  *	The page must be locked.
1935  */
1936 void
1937 vm_page_activate(vm_page_t m)
1938 {
1939 	int queue;
1940 
1941 	vm_page_lock_assert(m, MA_OWNED);
1942 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1943 	if ((queue = m->queue) != PQ_ACTIVE) {
1944 		if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
1945 			if (m->act_count < ACT_INIT)
1946 				m->act_count = ACT_INIT;
1947 			vm_page_lock_queues();
1948 			if (queue != PQ_NONE)
1949 				vm_page_queue_remove(queue, m);
1950 			vm_page_enqueue(PQ_ACTIVE, m);
1951 			vm_page_unlock_queues();
1952 		} else
1953 			KASSERT(queue == PQ_NONE,
1954 			    ("vm_page_activate: wired page %p is queued", m));
1955 	} else {
1956 		if (m->act_count < ACT_INIT)
1957 			m->act_count = ACT_INIT;
1958 	}
1959 }
1960 
1961 /*
1962  *	vm_page_free_wakeup:
1963  *
1964  *	Helper routine for vm_page_free_toq() and vm_page_cache().  This
1965  *	routine is called when a page has been added to the cache or free
1966  *	queues.
1967  *
1968  *	The page queues must be locked.
1969  */
1970 static inline void
1971 vm_page_free_wakeup(void)
1972 {
1973 
1974 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1975 	/*
1976 	 * if pageout daemon needs pages, then tell it that there are
1977 	 * some free.
1978 	 */
1979 	if (vm_pageout_pages_needed &&
1980 	    cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) {
1981 		wakeup(&vm_pageout_pages_needed);
1982 		vm_pageout_pages_needed = 0;
1983 	}
1984 	/*
1985 	 * wakeup processes that are waiting on memory if we hit a
1986 	 * high water mark. And wakeup scheduler process if we have
1987 	 * lots of memory. this process will swapin processes.
1988 	 */
1989 	if (vm_pages_needed && !vm_page_count_min()) {
1990 		vm_pages_needed = 0;
1991 		wakeup(&cnt.v_free_count);
1992 	}
1993 }
1994 
1995 /*
1996  *	vm_page_free_toq:
1997  *
1998  *	Returns the given page to the free list,
1999  *	disassociating it with any VM object.
2000  *
2001  *	The object must be locked.  The page must be locked if it is managed.
2002  */
2003 void
2004 vm_page_free_toq(vm_page_t m)
2005 {
2006 
2007 	if ((m->oflags & VPO_UNMANAGED) == 0) {
2008 		vm_page_lock_assert(m, MA_OWNED);
2009 		KASSERT(!pmap_page_is_mapped(m),
2010 		    ("vm_page_free_toq: freeing mapped page %p", m));
2011 	}
2012 	PCPU_INC(cnt.v_tfree);
2013 
2014 	if (VM_PAGE_IS_FREE(m))
2015 		panic("vm_page_free: freeing free page %p", m);
2016 	else if (m->busy != 0)
2017 		panic("vm_page_free: freeing busy page %p", m);
2018 
2019 	/*
2020 	 * Unqueue, then remove page.  Note that we cannot destroy
2021 	 * the page here because we do not want to call the pager's
2022 	 * callback routine until after we've put the page on the
2023 	 * appropriate free queue.
2024 	 */
2025 	if ((m->oflags & VPO_UNMANAGED) == 0)
2026 		vm_pageq_remove(m);
2027 	vm_page_remove(m);
2028 
2029 	/*
2030 	 * If fictitious remove object association and
2031 	 * return, otherwise delay object association removal.
2032 	 */
2033 	if ((m->flags & PG_FICTITIOUS) != 0) {
2034 		return;
2035 	}
2036 
2037 	m->valid = 0;
2038 	vm_page_undirty(m);
2039 
2040 	if (m->wire_count != 0)
2041 		panic("vm_page_free: freeing wired page %p", m);
2042 	if (m->hold_count != 0) {
2043 		m->flags &= ~PG_ZERO;
2044 		KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2045 		    ("vm_page_free: freeing PG_UNHOLDFREE page %p", m));
2046 		m->flags |= PG_UNHOLDFREE;
2047 	} else {
2048 		/*
2049 		 * Restore the default memory attribute to the page.
2050 		 */
2051 		if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
2052 			pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
2053 
2054 		/*
2055 		 * Insert the page into the physical memory allocator's
2056 		 * cache/free page queues.
2057 		 */
2058 		mtx_lock(&vm_page_queue_free_mtx);
2059 		m->flags |= PG_FREE;
2060 		cnt.v_free_count++;
2061 #if VM_NRESERVLEVEL > 0
2062 		if (!vm_reserv_free_page(m))
2063 #else
2064 		if (TRUE)
2065 #endif
2066 			vm_phys_free_pages(m, 0);
2067 		if ((m->flags & PG_ZERO) != 0)
2068 			++vm_page_zero_count;
2069 		else
2070 			vm_page_zero_idle_wakeup();
2071 		vm_page_free_wakeup();
2072 		mtx_unlock(&vm_page_queue_free_mtx);
2073 	}
2074 }
2075 
2076 /*
2077  *	vm_page_wire:
2078  *
2079  *	Mark this page as wired down by yet
2080  *	another map, removing it from paging queues
2081  *	as necessary.
2082  *
2083  *	If the page is fictitious, then its wire count must remain one.
2084  *
2085  *	The page must be locked.
2086  */
2087 void
2088 vm_page_wire(vm_page_t m)
2089 {
2090 
2091 	/*
2092 	 * Only bump the wire statistics if the page is not already wired,
2093 	 * and only unqueue the page if it is on some queue (if it is unmanaged
2094 	 * it is already off the queues).
2095 	 */
2096 	vm_page_lock_assert(m, MA_OWNED);
2097 	if ((m->flags & PG_FICTITIOUS) != 0) {
2098 		KASSERT(m->wire_count == 1,
2099 		    ("vm_page_wire: fictitious page %p's wire count isn't one",
2100 		    m));
2101 		return;
2102 	}
2103 	if (m->wire_count == 0) {
2104 		if ((m->oflags & VPO_UNMANAGED) == 0)
2105 			vm_pageq_remove(m);
2106 		atomic_add_int(&cnt.v_wire_count, 1);
2107 	}
2108 	m->wire_count++;
2109 	KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m));
2110 }
2111 
2112 /*
2113  * vm_page_unwire:
2114  *
2115  * Release one wiring of the specified page, potentially enabling it to be
2116  * paged again.  If paging is enabled, then the value of the parameter
2117  * "activate" determines to which queue the page is added.  If "activate" is
2118  * non-zero, then the page is added to the active queue.  Otherwise, it is
2119  * added to the inactive queue.
2120  *
2121  * However, unless the page belongs to an object, it is not enqueued because
2122  * it cannot be paged out.
2123  *
2124  * If a page is fictitious, then its wire count must alway be one.
2125  *
2126  * A managed page must be locked.
2127  */
2128 void
2129 vm_page_unwire(vm_page_t m, int activate)
2130 {
2131 
2132 	if ((m->oflags & VPO_UNMANAGED) == 0)
2133 		vm_page_lock_assert(m, MA_OWNED);
2134 	if ((m->flags & PG_FICTITIOUS) != 0) {
2135 		KASSERT(m->wire_count == 1,
2136 	    ("vm_page_unwire: fictitious page %p's wire count isn't one", m));
2137 		return;
2138 	}
2139 	if (m->wire_count > 0) {
2140 		m->wire_count--;
2141 		if (m->wire_count == 0) {
2142 			atomic_subtract_int(&cnt.v_wire_count, 1);
2143 			if ((m->oflags & VPO_UNMANAGED) != 0 ||
2144 			    m->object == NULL)
2145 				return;
2146 			if (!activate)
2147 				m->flags &= ~PG_WINATCFLS;
2148 			vm_page_lock_queues();
2149 			vm_page_enqueue(activate ? PQ_ACTIVE : PQ_INACTIVE, m);
2150 			vm_page_unlock_queues();
2151 		}
2152 	} else
2153 		panic("vm_page_unwire: page %p's wire count is zero", m);
2154 }
2155 
2156 /*
2157  * Move the specified page to the inactive queue.
2158  *
2159  * Many pages placed on the inactive queue should actually go
2160  * into the cache, but it is difficult to figure out which.  What
2161  * we do instead, if the inactive target is well met, is to put
2162  * clean pages at the head of the inactive queue instead of the tail.
2163  * This will cause them to be moved to the cache more quickly and
2164  * if not actively re-referenced, reclaimed more quickly.  If we just
2165  * stick these pages at the end of the inactive queue, heavy filesystem
2166  * meta-data accesses can cause an unnecessary paging load on memory bound
2167  * processes.  This optimization causes one-time-use metadata to be
2168  * reused more quickly.
2169  *
2170  * Normally athead is 0 resulting in LRU operation.  athead is set
2171  * to 1 if we want this page to be 'as if it were placed in the cache',
2172  * except without unmapping it from the process address space.
2173  *
2174  * The page must be locked.
2175  */
2176 static inline void
2177 _vm_page_deactivate(vm_page_t m, int athead)
2178 {
2179 	int queue;
2180 
2181 	vm_page_lock_assert(m, MA_OWNED);
2182 
2183 	/*
2184 	 * Ignore if already inactive.
2185 	 */
2186 	if ((queue = m->queue) == PQ_INACTIVE)
2187 		return;
2188 	if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
2189 		m->flags &= ~PG_WINATCFLS;
2190 		vm_page_lock_queues();
2191 		if (queue != PQ_NONE)
2192 			vm_page_queue_remove(queue, m);
2193 		if (athead)
2194 			TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m,
2195 			    pageq);
2196 		else
2197 			TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m,
2198 			    pageq);
2199 		m->queue = PQ_INACTIVE;
2200 		cnt.v_inactive_count++;
2201 		vm_page_unlock_queues();
2202 	}
2203 }
2204 
2205 /*
2206  * Move the specified page to the inactive queue.
2207  *
2208  * The page must be locked.
2209  */
2210 void
2211 vm_page_deactivate(vm_page_t m)
2212 {
2213 
2214 	_vm_page_deactivate(m, 0);
2215 }
2216 
2217 /*
2218  * vm_page_try_to_cache:
2219  *
2220  * Returns 0 on failure, 1 on success
2221  */
2222 int
2223 vm_page_try_to_cache(vm_page_t m)
2224 {
2225 
2226 	vm_page_lock_assert(m, MA_OWNED);
2227 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2228 	if (m->dirty || m->hold_count || m->busy || m->wire_count ||
2229 	    (m->oflags & (VPO_BUSY | VPO_UNMANAGED)) != 0)
2230 		return (0);
2231 	pmap_remove_all(m);
2232 	if (m->dirty)
2233 		return (0);
2234 	vm_page_cache(m);
2235 	return (1);
2236 }
2237 
2238 /*
2239  * vm_page_try_to_free()
2240  *
2241  *	Attempt to free the page.  If we cannot free it, we do nothing.
2242  *	1 is returned on success, 0 on failure.
2243  */
2244 int
2245 vm_page_try_to_free(vm_page_t m)
2246 {
2247 
2248 	vm_page_lock_assert(m, MA_OWNED);
2249 	if (m->object != NULL)
2250 		VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2251 	if (m->dirty || m->hold_count || m->busy || m->wire_count ||
2252 	    (m->oflags & (VPO_BUSY | VPO_UNMANAGED)) != 0)
2253 		return (0);
2254 	pmap_remove_all(m);
2255 	if (m->dirty)
2256 		return (0);
2257 	vm_page_free(m);
2258 	return (1);
2259 }
2260 
2261 /*
2262  * vm_page_cache
2263  *
2264  * Put the specified page onto the page cache queue (if appropriate).
2265  *
2266  * The object and page must be locked.
2267  */
2268 void
2269 vm_page_cache(vm_page_t m)
2270 {
2271 	vm_object_t object;
2272 	vm_page_t next, prev, root;
2273 
2274 	vm_page_lock_assert(m, MA_OWNED);
2275 	object = m->object;
2276 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2277 	if ((m->oflags & (VPO_UNMANAGED | VPO_BUSY)) || m->busy ||
2278 	    m->hold_count || m->wire_count)
2279 		panic("vm_page_cache: attempting to cache busy page");
2280 	KASSERT(!pmap_page_is_mapped(m),
2281 	    ("vm_page_cache: page %p is mapped", m));
2282 	KASSERT(m->dirty == 0, ("vm_page_cache: page %p is dirty", m));
2283 	if (m->valid == 0 || object->type == OBJT_DEFAULT ||
2284 	    (object->type == OBJT_SWAP &&
2285 	    !vm_pager_has_page(object, m->pindex, NULL, NULL))) {
2286 		/*
2287 		 * Hypothesis: A cache-elgible page belonging to a
2288 		 * default object or swap object but without a backing
2289 		 * store must be zero filled.
2290 		 */
2291 		vm_page_free(m);
2292 		return;
2293 	}
2294 	KASSERT((m->flags & PG_CACHED) == 0,
2295 	    ("vm_page_cache: page %p is already cached", m));
2296 	PCPU_INC(cnt.v_tcached);
2297 
2298 	/*
2299 	 * Remove the page from the paging queues.
2300 	 */
2301 	vm_pageq_remove(m);
2302 
2303 	/*
2304 	 * Remove the page from the object's collection of resident
2305 	 * pages.
2306 	 */
2307 	if ((next = TAILQ_NEXT(m, listq)) != NULL && next->left == m) {
2308 		/*
2309 		 * Since the page's successor in the list is also its parent
2310 		 * in the tree, its right subtree must be empty.
2311 		 */
2312 		next->left = m->left;
2313 		KASSERT(m->right == NULL,
2314 		    ("vm_page_cache: page %p has right child", m));
2315 	} else if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL &&
2316 	    prev->right == m) {
2317 		/*
2318 		 * Since the page's predecessor in the list is also its parent
2319 		 * in the tree, its left subtree must be empty.
2320 		 */
2321 		KASSERT(m->left == NULL,
2322 		    ("vm_page_cache: page %p has left child", m));
2323 		prev->right = m->right;
2324 	} else {
2325 		if (m != object->root)
2326 			vm_page_splay(m->pindex, object->root);
2327 		if (m->left == NULL)
2328 			root = m->right;
2329 		else if (m->right == NULL)
2330 			root = m->left;
2331 		else {
2332 			/*
2333 			 * Move the page's successor to the root, because
2334 			 * pages are usually removed in ascending order.
2335 			 */
2336 			if (m->right != next)
2337 				vm_page_splay(m->pindex, m->right);
2338 			next->left = m->left;
2339 			root = next;
2340 		}
2341 		object->root = root;
2342 	}
2343 	TAILQ_REMOVE(&object->memq, m, listq);
2344 	object->resident_page_count--;
2345 
2346 	/*
2347 	 * Restore the default memory attribute to the page.
2348 	 */
2349 	if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
2350 		pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
2351 
2352 	/*
2353 	 * Insert the page into the object's collection of cached pages
2354 	 * and the physical memory allocator's cache/free page queues.
2355 	 */
2356 	m->flags &= ~PG_ZERO;
2357 	mtx_lock(&vm_page_queue_free_mtx);
2358 	m->flags |= PG_CACHED;
2359 	cnt.v_cache_count++;
2360 	root = object->cache;
2361 	if (root == NULL) {
2362 		m->left = NULL;
2363 		m->right = NULL;
2364 	} else {
2365 		root = vm_page_splay(m->pindex, root);
2366 		if (m->pindex < root->pindex) {
2367 			m->left = root->left;
2368 			m->right = root;
2369 			root->left = NULL;
2370 		} else if (__predict_false(m->pindex == root->pindex))
2371 			panic("vm_page_cache: offset already cached");
2372 		else {
2373 			m->right = root->right;
2374 			m->left = root;
2375 			root->right = NULL;
2376 		}
2377 	}
2378 	object->cache = m;
2379 #if VM_NRESERVLEVEL > 0
2380 	if (!vm_reserv_free_page(m)) {
2381 #else
2382 	if (TRUE) {
2383 #endif
2384 		vm_phys_set_pool(VM_FREEPOOL_CACHE, m, 0);
2385 		vm_phys_free_pages(m, 0);
2386 	}
2387 	vm_page_free_wakeup();
2388 	mtx_unlock(&vm_page_queue_free_mtx);
2389 
2390 	/*
2391 	 * Increment the vnode's hold count if this is the object's only
2392 	 * cached page.  Decrement the vnode's hold count if this was
2393 	 * the object's only resident page.
2394 	 */
2395 	if (object->type == OBJT_VNODE) {
2396 		if (root == NULL && object->resident_page_count != 0)
2397 			vhold(object->handle);
2398 		else if (root != NULL && object->resident_page_count == 0)
2399 			vdrop(object->handle);
2400 	}
2401 }
2402 
2403 /*
2404  * vm_page_dontneed
2405  *
2406  *	Cache, deactivate, or do nothing as appropriate.  This routine
2407  *	is typically used by madvise() MADV_DONTNEED.
2408  *
2409  *	Generally speaking we want to move the page into the cache so
2410  *	it gets reused quickly.  However, this can result in a silly syndrome
2411  *	due to the page recycling too quickly.  Small objects will not be
2412  *	fully cached.  On the otherhand, if we move the page to the inactive
2413  *	queue we wind up with a problem whereby very large objects
2414  *	unnecessarily blow away our inactive and cache queues.
2415  *
2416  *	The solution is to move the pages based on a fixed weighting.  We
2417  *	either leave them alone, deactivate them, or move them to the cache,
2418  *	where moving them to the cache has the highest weighting.
2419  *	By forcing some pages into other queues we eventually force the
2420  *	system to balance the queues, potentially recovering other unrelated
2421  *	space from active.  The idea is to not force this to happen too
2422  *	often.
2423  *
2424  *	The object and page must be locked.
2425  */
2426 void
2427 vm_page_dontneed(vm_page_t m)
2428 {
2429 	int dnw;
2430 	int head;
2431 
2432 	vm_page_lock_assert(m, MA_OWNED);
2433 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2434 	dnw = PCPU_GET(dnweight);
2435 	PCPU_INC(dnweight);
2436 
2437 	/*
2438 	 * Occasionally leave the page alone.
2439 	 */
2440 	if ((dnw & 0x01F0) == 0 || m->queue == PQ_INACTIVE) {
2441 		if (m->act_count >= ACT_INIT)
2442 			--m->act_count;
2443 		return;
2444 	}
2445 
2446 	/*
2447 	 * Clear any references to the page.  Otherwise, the page daemon will
2448 	 * immediately reactivate the page.
2449 	 *
2450 	 * Perform the pmap_clear_reference() first.  Otherwise, a concurrent
2451 	 * pmap operation, such as pmap_remove(), could clear a reference in
2452 	 * the pmap and set PGA_REFERENCED on the page before the
2453 	 * pmap_clear_reference() had completed.  Consequently, the page would
2454 	 * appear referenced based upon an old reference that occurred before
2455 	 * this function ran.
2456 	 */
2457 	pmap_clear_reference(m);
2458 	vm_page_aflag_clear(m, PGA_REFERENCED);
2459 
2460 	if (m->dirty == 0 && pmap_is_modified(m))
2461 		vm_page_dirty(m);
2462 
2463 	if (m->dirty || (dnw & 0x0070) == 0) {
2464 		/*
2465 		 * Deactivate the page 3 times out of 32.
2466 		 */
2467 		head = 0;
2468 	} else {
2469 		/*
2470 		 * Cache the page 28 times out of every 32.  Note that
2471 		 * the page is deactivated instead of cached, but placed
2472 		 * at the head of the queue instead of the tail.
2473 		 */
2474 		head = 1;
2475 	}
2476 	_vm_page_deactivate(m, head);
2477 }
2478 
2479 /*
2480  * Grab a page, waiting until we are waken up due to the page
2481  * changing state.  We keep on waiting, if the page continues
2482  * to be in the object.  If the page doesn't exist, first allocate it
2483  * and then conditionally zero it.
2484  *
2485  * The caller must always specify the VM_ALLOC_RETRY flag.  This is intended
2486  * to facilitate its eventual removal.
2487  *
2488  * This routine may sleep.
2489  *
2490  * The object must be locked on entry.  The lock will, however, be released
2491  * and reacquired if the routine sleeps.
2492  */
2493 vm_page_t
2494 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
2495 {
2496 	vm_page_t m;
2497 
2498 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2499 	KASSERT((allocflags & VM_ALLOC_RETRY) != 0,
2500 	    ("vm_page_grab: VM_ALLOC_RETRY is required"));
2501 retrylookup:
2502 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
2503 		if ((m->oflags & VPO_BUSY) != 0 ||
2504 		    ((allocflags & VM_ALLOC_IGN_SBUSY) == 0 && m->busy != 0)) {
2505 			/*
2506 			 * Reference the page before unlocking and
2507 			 * sleeping so that the page daemon is less
2508 			 * likely to reclaim it.
2509 			 */
2510 			vm_page_aflag_set(m, PGA_REFERENCED);
2511 			vm_page_sleep(m, "pgrbwt");
2512 			goto retrylookup;
2513 		} else {
2514 			if ((allocflags & VM_ALLOC_WIRED) != 0) {
2515 				vm_page_lock(m);
2516 				vm_page_wire(m);
2517 				vm_page_unlock(m);
2518 			}
2519 			if ((allocflags & VM_ALLOC_NOBUSY) == 0)
2520 				vm_page_busy(m);
2521 			return (m);
2522 		}
2523 	}
2524 	m = vm_page_alloc(object, pindex, allocflags & ~(VM_ALLOC_RETRY |
2525 	    VM_ALLOC_IGN_SBUSY));
2526 	if (m == NULL) {
2527 		VM_OBJECT_UNLOCK(object);
2528 		VM_WAIT;
2529 		VM_OBJECT_LOCK(object);
2530 		goto retrylookup;
2531 	} else if (m->valid != 0)
2532 		return (m);
2533 	if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
2534 		pmap_zero_page(m);
2535 	return (m);
2536 }
2537 
2538 /*
2539  * Mapping function for valid or dirty bits in a page.
2540  *
2541  * Inputs are required to range within a page.
2542  */
2543 vm_page_bits_t
2544 vm_page_bits(int base, int size)
2545 {
2546 	int first_bit;
2547 	int last_bit;
2548 
2549 	KASSERT(
2550 	    base + size <= PAGE_SIZE,
2551 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
2552 	);
2553 
2554 	if (size == 0)		/* handle degenerate case */
2555 		return (0);
2556 
2557 	first_bit = base >> DEV_BSHIFT;
2558 	last_bit = (base + size - 1) >> DEV_BSHIFT;
2559 
2560 	return (((vm_page_bits_t)2 << last_bit) -
2561 	    ((vm_page_bits_t)1 << first_bit));
2562 }
2563 
2564 /*
2565  *	vm_page_set_valid_range:
2566  *
2567  *	Sets portions of a page valid.  The arguments are expected
2568  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
2569  *	of any partial chunks touched by the range.  The invalid portion of
2570  *	such chunks will be zeroed.
2571  *
2572  *	(base + size) must be less then or equal to PAGE_SIZE.
2573  */
2574 void
2575 vm_page_set_valid_range(vm_page_t m, int base, int size)
2576 {
2577 	int endoff, frag;
2578 
2579 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2580 	if (size == 0)	/* handle degenerate case */
2581 		return;
2582 
2583 	/*
2584 	 * If the base is not DEV_BSIZE aligned and the valid
2585 	 * bit is clear, we have to zero out a portion of the
2586 	 * first block.
2587 	 */
2588 	if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
2589 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
2590 		pmap_zero_page_area(m, frag, base - frag);
2591 
2592 	/*
2593 	 * If the ending offset is not DEV_BSIZE aligned and the
2594 	 * valid bit is clear, we have to zero out a portion of
2595 	 * the last block.
2596 	 */
2597 	endoff = base + size;
2598 	if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
2599 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
2600 		pmap_zero_page_area(m, endoff,
2601 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
2602 
2603 	/*
2604 	 * Assert that no previously invalid block that is now being validated
2605 	 * is already dirty.
2606 	 */
2607 	KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
2608 	    ("vm_page_set_valid_range: page %p is dirty", m));
2609 
2610 	/*
2611 	 * Set valid bits inclusive of any overlap.
2612 	 */
2613 	m->valid |= vm_page_bits(base, size);
2614 }
2615 
2616 /*
2617  * Clear the given bits from the specified page's dirty field.
2618  */
2619 static __inline void
2620 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
2621 {
2622 	uintptr_t addr;
2623 #if PAGE_SIZE < 16384
2624 	int shift;
2625 #endif
2626 
2627 	/*
2628 	 * If the object is locked and the page is neither VPO_BUSY nor
2629 	 * write mapped, then the page's dirty field cannot possibly be
2630 	 * set by a concurrent pmap operation.
2631 	 */
2632 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2633 	if ((m->oflags & VPO_BUSY) == 0 && !pmap_page_is_write_mapped(m))
2634 		m->dirty &= ~pagebits;
2635 	else {
2636 		/*
2637 		 * The pmap layer can call vm_page_dirty() without
2638 		 * holding a distinguished lock.  The combination of
2639 		 * the object's lock and an atomic operation suffice
2640 		 * to guarantee consistency of the page dirty field.
2641 		 *
2642 		 * For PAGE_SIZE == 32768 case, compiler already
2643 		 * properly aligns the dirty field, so no forcible
2644 		 * alignment is needed. Only require existence of
2645 		 * atomic_clear_64 when page size is 32768.
2646 		 */
2647 		addr = (uintptr_t)&m->dirty;
2648 #if PAGE_SIZE == 32768
2649 		atomic_clear_64((uint64_t *)addr, pagebits);
2650 #elif PAGE_SIZE == 16384
2651 		atomic_clear_32((uint32_t *)addr, pagebits);
2652 #else		/* PAGE_SIZE <= 8192 */
2653 		/*
2654 		 * Use a trick to perform a 32-bit atomic on the
2655 		 * containing aligned word, to not depend on the existence
2656 		 * of atomic_clear_{8, 16}.
2657 		 */
2658 		shift = addr & (sizeof(uint32_t) - 1);
2659 #if BYTE_ORDER == BIG_ENDIAN
2660 		shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY;
2661 #else
2662 		shift *= NBBY;
2663 #endif
2664 		addr &= ~(sizeof(uint32_t) - 1);
2665 		atomic_clear_32((uint32_t *)addr, pagebits << shift);
2666 #endif		/* PAGE_SIZE */
2667 	}
2668 }
2669 
2670 /*
2671  *	vm_page_set_validclean:
2672  *
2673  *	Sets portions of a page valid and clean.  The arguments are expected
2674  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
2675  *	of any partial chunks touched by the range.  The invalid portion of
2676  *	such chunks will be zero'd.
2677  *
2678  *	(base + size) must be less then or equal to PAGE_SIZE.
2679  */
2680 void
2681 vm_page_set_validclean(vm_page_t m, int base, int size)
2682 {
2683 	vm_page_bits_t oldvalid, pagebits;
2684 	int endoff, frag;
2685 
2686 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2687 	if (size == 0)	/* handle degenerate case */
2688 		return;
2689 
2690 	/*
2691 	 * If the base is not DEV_BSIZE aligned and the valid
2692 	 * bit is clear, we have to zero out a portion of the
2693 	 * first block.
2694 	 */
2695 	if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
2696 	    (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
2697 		pmap_zero_page_area(m, frag, base - frag);
2698 
2699 	/*
2700 	 * If the ending offset is not DEV_BSIZE aligned and the
2701 	 * valid bit is clear, we have to zero out a portion of
2702 	 * the last block.
2703 	 */
2704 	endoff = base + size;
2705 	if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
2706 	    (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
2707 		pmap_zero_page_area(m, endoff,
2708 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
2709 
2710 	/*
2711 	 * Set valid, clear dirty bits.  If validating the entire
2712 	 * page we can safely clear the pmap modify bit.  We also
2713 	 * use this opportunity to clear the VPO_NOSYNC flag.  If a process
2714 	 * takes a write fault on a MAP_NOSYNC memory area the flag will
2715 	 * be set again.
2716 	 *
2717 	 * We set valid bits inclusive of any overlap, but we can only
2718 	 * clear dirty bits for DEV_BSIZE chunks that are fully within
2719 	 * the range.
2720 	 */
2721 	oldvalid = m->valid;
2722 	pagebits = vm_page_bits(base, size);
2723 	m->valid |= pagebits;
2724 #if 0	/* NOT YET */
2725 	if ((frag = base & (DEV_BSIZE - 1)) != 0) {
2726 		frag = DEV_BSIZE - frag;
2727 		base += frag;
2728 		size -= frag;
2729 		if (size < 0)
2730 			size = 0;
2731 	}
2732 	pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
2733 #endif
2734 	if (base == 0 && size == PAGE_SIZE) {
2735 		/*
2736 		 * The page can only be modified within the pmap if it is
2737 		 * mapped, and it can only be mapped if it was previously
2738 		 * fully valid.
2739 		 */
2740 		if (oldvalid == VM_PAGE_BITS_ALL)
2741 			/*
2742 			 * Perform the pmap_clear_modify() first.  Otherwise,
2743 			 * a concurrent pmap operation, such as
2744 			 * pmap_protect(), could clear a modification in the
2745 			 * pmap and set the dirty field on the page before
2746 			 * pmap_clear_modify() had begun and after the dirty
2747 			 * field was cleared here.
2748 			 */
2749 			pmap_clear_modify(m);
2750 		m->dirty = 0;
2751 		m->oflags &= ~VPO_NOSYNC;
2752 	} else if (oldvalid != VM_PAGE_BITS_ALL)
2753 		m->dirty &= ~pagebits;
2754 	else
2755 		vm_page_clear_dirty_mask(m, pagebits);
2756 }
2757 
2758 void
2759 vm_page_clear_dirty(vm_page_t m, int base, int size)
2760 {
2761 
2762 	vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
2763 }
2764 
2765 /*
2766  *	vm_page_set_invalid:
2767  *
2768  *	Invalidates DEV_BSIZE'd chunks within a page.  Both the
2769  *	valid and dirty bits for the effected areas are cleared.
2770  */
2771 void
2772 vm_page_set_invalid(vm_page_t m, int base, int size)
2773 {
2774 	vm_page_bits_t bits;
2775 
2776 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2777 	KASSERT((m->oflags & VPO_BUSY) == 0,
2778 	    ("vm_page_set_invalid: page %p is busy", m));
2779 	bits = vm_page_bits(base, size);
2780 	if (m->valid == VM_PAGE_BITS_ALL && bits != 0)
2781 		pmap_remove_all(m);
2782 	KASSERT(!pmap_page_is_mapped(m),
2783 	    ("vm_page_set_invalid: page %p is mapped", m));
2784 	m->valid &= ~bits;
2785 	m->dirty &= ~bits;
2786 }
2787 
2788 /*
2789  * vm_page_zero_invalid()
2790  *
2791  *	The kernel assumes that the invalid portions of a page contain
2792  *	garbage, but such pages can be mapped into memory by user code.
2793  *	When this occurs, we must zero out the non-valid portions of the
2794  *	page so user code sees what it expects.
2795  *
2796  *	Pages are most often semi-valid when the end of a file is mapped
2797  *	into memory and the file's size is not page aligned.
2798  */
2799 void
2800 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
2801 {
2802 	int b;
2803 	int i;
2804 
2805 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2806 	/*
2807 	 * Scan the valid bits looking for invalid sections that
2808 	 * must be zerod.  Invalid sub-DEV_BSIZE'd areas ( where the
2809 	 * valid bit may be set ) have already been zerod by
2810 	 * vm_page_set_validclean().
2811 	 */
2812 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
2813 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
2814 		    (m->valid & ((vm_page_bits_t)1 << i))) {
2815 			if (i > b) {
2816 				pmap_zero_page_area(m,
2817 				    b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
2818 			}
2819 			b = i + 1;
2820 		}
2821 	}
2822 
2823 	/*
2824 	 * setvalid is TRUE when we can safely set the zero'd areas
2825 	 * as being valid.  We can do this if there are no cache consistancy
2826 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
2827 	 */
2828 	if (setvalid)
2829 		m->valid = VM_PAGE_BITS_ALL;
2830 }
2831 
2832 /*
2833  *	vm_page_is_valid:
2834  *
2835  *	Is (partial) page valid?  Note that the case where size == 0
2836  *	will return FALSE in the degenerate case where the page is
2837  *	entirely invalid, and TRUE otherwise.
2838  */
2839 int
2840 vm_page_is_valid(vm_page_t m, int base, int size)
2841 {
2842 	vm_page_bits_t bits;
2843 
2844 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2845 	bits = vm_page_bits(base, size);
2846 	if (m->valid && ((m->valid & bits) == bits))
2847 		return 1;
2848 	else
2849 		return 0;
2850 }
2851 
2852 /*
2853  * Set the page's dirty bits if the page is modified.
2854  */
2855 void
2856 vm_page_test_dirty(vm_page_t m)
2857 {
2858 
2859 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2860 	if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
2861 		vm_page_dirty(m);
2862 }
2863 
2864 void
2865 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
2866 {
2867 
2868 	mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
2869 }
2870 
2871 void
2872 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
2873 {
2874 
2875 	mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
2876 }
2877 
2878 int
2879 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
2880 {
2881 
2882 	return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
2883 }
2884 
2885 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
2886 void
2887 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
2888 {
2889 
2890 	mtx_assert_(vm_page_lockptr(m), a, file, line);
2891 }
2892 #endif
2893 
2894 int so_zerocp_fullpage = 0;
2895 
2896 /*
2897  *	Replace the given page with a copy.  The copied page assumes
2898  *	the portion of the given page's "wire_count" that is not the
2899  *	responsibility of this copy-on-write mechanism.
2900  *
2901  *	The object containing the given page must have a non-zero
2902  *	paging-in-progress count and be locked.
2903  */
2904 void
2905 vm_page_cowfault(vm_page_t m)
2906 {
2907 	vm_page_t mnew;
2908 	vm_object_t object;
2909 	vm_pindex_t pindex;
2910 
2911 	vm_page_lock_assert(m, MA_OWNED);
2912 	object = m->object;
2913 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2914 	KASSERT(object->paging_in_progress != 0,
2915 	    ("vm_page_cowfault: object %p's paging-in-progress count is zero.",
2916 	    object));
2917 	pindex = m->pindex;
2918 
2919  retry_alloc:
2920 	pmap_remove_all(m);
2921 	vm_page_remove(m);
2922 	mnew = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL | VM_ALLOC_NOBUSY);
2923 	if (mnew == NULL) {
2924 		vm_page_insert(m, object, pindex);
2925 		vm_page_unlock(m);
2926 		VM_OBJECT_UNLOCK(object);
2927 		VM_WAIT;
2928 		VM_OBJECT_LOCK(object);
2929 		if (m == vm_page_lookup(object, pindex)) {
2930 			vm_page_lock(m);
2931 			goto retry_alloc;
2932 		} else {
2933 			/*
2934 			 * Page disappeared during the wait.
2935 			 */
2936 			return;
2937 		}
2938 	}
2939 
2940 	if (m->cow == 0) {
2941 		/*
2942 		 * check to see if we raced with an xmit complete when
2943 		 * waiting to allocate a page.  If so, put things back
2944 		 * the way they were
2945 		 */
2946 		vm_page_unlock(m);
2947 		vm_page_lock(mnew);
2948 		vm_page_free(mnew);
2949 		vm_page_unlock(mnew);
2950 		vm_page_insert(m, object, pindex);
2951 	} else { /* clear COW & copy page */
2952 		if (!so_zerocp_fullpage)
2953 			pmap_copy_page(m, mnew);
2954 		mnew->valid = VM_PAGE_BITS_ALL;
2955 		vm_page_dirty(mnew);
2956 		mnew->wire_count = m->wire_count - m->cow;
2957 		m->wire_count = m->cow;
2958 		vm_page_unlock(m);
2959 	}
2960 }
2961 
2962 void
2963 vm_page_cowclear(vm_page_t m)
2964 {
2965 
2966 	vm_page_lock_assert(m, MA_OWNED);
2967 	if (m->cow) {
2968 		m->cow--;
2969 		/*
2970 		 * let vm_fault add back write permission  lazily
2971 		 */
2972 	}
2973 	/*
2974 	 *  sf_buf_free() will free the page, so we needn't do it here
2975 	 */
2976 }
2977 
2978 int
2979 vm_page_cowsetup(vm_page_t m)
2980 {
2981 
2982 	vm_page_lock_assert(m, MA_OWNED);
2983 	if ((m->flags & PG_FICTITIOUS) != 0 ||
2984 	    (m->oflags & VPO_UNMANAGED) != 0 ||
2985 	    m->cow == USHRT_MAX - 1 || !VM_OBJECT_TRYLOCK(m->object))
2986 		return (EBUSY);
2987 	m->cow++;
2988 	pmap_remove_write(m);
2989 	VM_OBJECT_UNLOCK(m->object);
2990 	return (0);
2991 }
2992 
2993 #ifdef INVARIANTS
2994 void
2995 vm_page_object_lock_assert(vm_page_t m)
2996 {
2997 
2998 	/*
2999 	 * Certain of the page's fields may only be modified by the
3000 	 * holder of the containing object's lock or the setter of the
3001 	 * page's VPO_BUSY flag.  Unfortunately, the setter of the
3002 	 * VPO_BUSY flag is not recorded, and thus cannot be checked
3003 	 * here.
3004 	 */
3005 	if (m->object != NULL && (m->oflags & VPO_BUSY) == 0)
3006 		VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
3007 }
3008 #endif
3009 
3010 #include "opt_ddb.h"
3011 #ifdef DDB
3012 #include <sys/kernel.h>
3013 
3014 #include <ddb/ddb.h>
3015 
3016 DB_SHOW_COMMAND(page, vm_page_print_page_info)
3017 {
3018 	db_printf("cnt.v_free_count: %d\n", cnt.v_free_count);
3019 	db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count);
3020 	db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count);
3021 	db_printf("cnt.v_active_count: %d\n", cnt.v_active_count);
3022 	db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count);
3023 	db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved);
3024 	db_printf("cnt.v_free_min: %d\n", cnt.v_free_min);
3025 	db_printf("cnt.v_free_target: %d\n", cnt.v_free_target);
3026 	db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min);
3027 	db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target);
3028 }
3029 
3030 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
3031 {
3032 
3033 	db_printf("PQ_FREE:");
3034 	db_printf(" %d", cnt.v_free_count);
3035 	db_printf("\n");
3036 
3037 	db_printf("PQ_CACHE:");
3038 	db_printf(" %d", cnt.v_cache_count);
3039 	db_printf("\n");
3040 
3041 	db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n",
3042 		*vm_page_queues[PQ_ACTIVE].cnt,
3043 		*vm_page_queues[PQ_INACTIVE].cnt);
3044 }
3045 #endif /* DDB */
3046