xref: /freebsd/sys/vm/vm_page.c (revision 3ef51c5fb9163f2aafb1c14729e06a8bf0c4d113)
1 /*-
2  * Copyright (c) 1991 Regents of the University of California.
3  * All rights reserved.
4  * Copyright (c) 1998 Matthew Dillon.  All Rights Reserved.
5  *
6  * This code is derived from software contributed to Berkeley by
7  * The Mach Operating System project at Carnegie-Mellon University.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  * 4. Neither the name of the University nor the names of its contributors
18  *    may be used to endorse or promote products derived from this software
19  *    without specific prior written permission.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
22  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
25  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31  * SUCH DAMAGE.
32  *
33  *	from: @(#)vm_page.c	7.4 (Berkeley) 5/7/91
34  */
35 
36 /*-
37  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
38  * All rights reserved.
39  *
40  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
41  *
42  * Permission to use, copy, modify and distribute this software and
43  * its documentation is hereby granted, provided that both the copyright
44  * notice and this permission notice appear in all copies of the
45  * software, derivative works or modified versions, and any portions
46  * thereof, and that both notices appear in supporting documentation.
47  *
48  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
49  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
50  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
51  *
52  * Carnegie Mellon requests users of this software to return to
53  *
54  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
55  *  School of Computer Science
56  *  Carnegie Mellon University
57  *  Pittsburgh PA 15213-3890
58  *
59  * any improvements or extensions that they make and grant Carnegie the
60  * rights to redistribute these changes.
61  */
62 
63 /*
64  *			GENERAL RULES ON VM_PAGE MANIPULATION
65  *
66  *	- a pageq mutex is required when adding or removing a page from a
67  *	  page queue (vm_page_queue[]), regardless of other mutexes or the
68  *	  busy state of a page.
69  *
70  *	- The object mutex is held when inserting or removing
71  *	  pages from an object (vm_page_insert() or vm_page_remove()).
72  *
73  */
74 
75 /*
76  *	Resident memory management module.
77  */
78 
79 #include <sys/cdefs.h>
80 __FBSDID("$FreeBSD$");
81 
82 #include "opt_vm.h"
83 
84 #include <sys/param.h>
85 #include <sys/systm.h>
86 #include <sys/lock.h>
87 #include <sys/kernel.h>
88 #include <sys/limits.h>
89 #include <sys/malloc.h>
90 #include <sys/msgbuf.h>
91 #include <sys/mutex.h>
92 #include <sys/proc.h>
93 #include <sys/sysctl.h>
94 #include <sys/vmmeter.h>
95 #include <sys/vnode.h>
96 
97 #include <vm/vm.h>
98 #include <vm/pmap.h>
99 #include <vm/vm_param.h>
100 #include <vm/vm_kern.h>
101 #include <vm/vm_object.h>
102 #include <vm/vm_page.h>
103 #include <vm/vm_pageout.h>
104 #include <vm/vm_pager.h>
105 #include <vm/vm_phys.h>
106 #include <vm/vm_reserv.h>
107 #include <vm/vm_extern.h>
108 #include <vm/uma.h>
109 #include <vm/uma_int.h>
110 
111 #include <machine/md_var.h>
112 
113 /*
114  *	Associated with page of user-allocatable memory is a
115  *	page structure.
116  */
117 
118 struct vpgqueues vm_page_queues[PQ_COUNT];
119 struct vpglocks vm_page_queue_lock;
120 struct vpglocks vm_page_queue_free_lock;
121 
122 struct vpglocks	pa_lock[PA_LOCK_COUNT];
123 
124 vm_page_t vm_page_array = 0;
125 int vm_page_array_size = 0;
126 long first_page = 0;
127 int vm_page_zero_count = 0;
128 
129 static int boot_pages = UMA_BOOT_PAGES;
130 TUNABLE_INT("vm.boot_pages", &boot_pages);
131 SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RD, &boot_pages, 0,
132 	"number of pages allocated for bootstrapping the VM system");
133 
134 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 locks.
296 	 */
297 	mtx_init(&vm_page_queue_mtx, "vm page queue mutex", NULL, MTX_DEF |
298 	    MTX_RECURSE);
299 	mtx_init(&vm_page_queue_free_mtx, "vm page queue free mutex", NULL,
300 	    MTX_DEF);
301 
302 	/* Setup page locks. */
303 	for (i = 0; i < PA_LOCK_COUNT; i++)
304 		mtx_init(&pa_lock[i].data, "page lock", NULL, MTX_DEF);
305 
306 	/*
307 	 * Initialize the queue headers for the hold queue, the active queue,
308 	 * and the inactive queue.
309 	 */
310 	for (i = 0; i < PQ_COUNT; i++)
311 		TAILQ_INIT(&vm_page_queues[i].pl);
312 	vm_page_queues[PQ_INACTIVE].cnt = &cnt.v_inactive_count;
313 	vm_page_queues[PQ_ACTIVE].cnt = &cnt.v_active_count;
314 	vm_page_queues[PQ_HOLD].cnt = &cnt.v_active_count;
315 
316 	/*
317 	 * Allocate memory for use when boot strapping the kernel memory
318 	 * allocator.
319 	 */
320 	new_end = end - (boot_pages * UMA_SLAB_SIZE);
321 	new_end = trunc_page(new_end);
322 	mapped = pmap_map(&vaddr, new_end, end,
323 	    VM_PROT_READ | VM_PROT_WRITE);
324 	bzero((void *)mapped, end - new_end);
325 	uma_startup((void *)mapped, boot_pages);
326 
327 #if defined(__amd64__) || defined(__i386__) || defined(__arm__) || \
328     defined(__mips__)
329 	/*
330 	 * Allocate a bitmap to indicate that a random physical page
331 	 * needs to be included in a minidump.
332 	 *
333 	 * The amd64 port needs this to indicate which direct map pages
334 	 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
335 	 *
336 	 * However, i386 still needs this workspace internally within the
337 	 * minidump code.  In theory, they are not needed on i386, but are
338 	 * included should the sf_buf code decide to use them.
339 	 */
340 	last_pa = 0;
341 	for (i = 0; dump_avail[i + 1] != 0; i += 2)
342 		if (dump_avail[i + 1] > last_pa)
343 			last_pa = dump_avail[i + 1];
344 	page_range = last_pa / PAGE_SIZE;
345 	vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
346 	new_end -= vm_page_dump_size;
347 	vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
348 	    new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
349 	bzero((void *)vm_page_dump, vm_page_dump_size);
350 #endif
351 #ifdef __amd64__
352 	/*
353 	 * Request that the physical pages underlying the message buffer be
354 	 * included in a crash dump.  Since the message buffer is accessed
355 	 * through the direct map, they are not automatically included.
356 	 */
357 	pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
358 	last_pa = pa + round_page(msgbufsize);
359 	while (pa < last_pa) {
360 		dump_add_page(pa);
361 		pa += PAGE_SIZE;
362 	}
363 #endif
364 	/*
365 	 * Compute the number of pages of memory that will be available for
366 	 * use (taking into account the overhead of a page structure per
367 	 * page).
368 	 */
369 	first_page = low_water / PAGE_SIZE;
370 #ifdef VM_PHYSSEG_SPARSE
371 	page_range = 0;
372 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
373 		page_range += atop(phys_avail[i + 1] - phys_avail[i]);
374 #elif defined(VM_PHYSSEG_DENSE)
375 	page_range = high_water / PAGE_SIZE - first_page;
376 #else
377 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
378 #endif
379 	end = new_end;
380 
381 	/*
382 	 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
383 	 */
384 	vaddr += PAGE_SIZE;
385 
386 	/*
387 	 * Initialize the mem entry structures now, and put them in the free
388 	 * queue.
389 	 */
390 	new_end = trunc_page(end - page_range * sizeof(struct vm_page));
391 	mapped = pmap_map(&vaddr, new_end, end,
392 	    VM_PROT_READ | VM_PROT_WRITE);
393 	vm_page_array = (vm_page_t) mapped;
394 #if VM_NRESERVLEVEL > 0
395 	/*
396 	 * Allocate memory for the reservation management system's data
397 	 * structures.
398 	 */
399 	new_end = vm_reserv_startup(&vaddr, new_end, high_water);
400 #endif
401 #if defined(__amd64__) || defined(__mips__)
402 	/*
403 	 * pmap_map on amd64 and mips can come out of the direct-map, not kvm
404 	 * like i386, so the pages must be tracked for a crashdump to include
405 	 * this data.  This includes the vm_page_array and the early UMA
406 	 * bootstrap pages.
407 	 */
408 	for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE)
409 		dump_add_page(pa);
410 #endif
411 	phys_avail[biggestone + 1] = new_end;
412 
413 	/*
414 	 * Clear all of the page structures
415 	 */
416 	bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
417 	for (i = 0; i < page_range; i++)
418 		vm_page_array[i].order = VM_NFREEORDER;
419 	vm_page_array_size = page_range;
420 
421 	/*
422 	 * Initialize the physical memory allocator.
423 	 */
424 	vm_phys_init();
425 
426 	/*
427 	 * Add every available physical page that is not blacklisted to
428 	 * the free lists.
429 	 */
430 	cnt.v_page_count = 0;
431 	cnt.v_free_count = 0;
432 	list = getenv("vm.blacklist");
433 	for (i = 0; phys_avail[i + 1] != 0; i += 2) {
434 		pa = phys_avail[i];
435 		last_pa = phys_avail[i + 1];
436 		while (pa < last_pa) {
437 			if (list != NULL &&
438 			    vm_page_blacklist_lookup(list, pa))
439 				printf("Skipping page with pa 0x%jx\n",
440 				    (uintmax_t)pa);
441 			else
442 				vm_phys_add_page(pa);
443 			pa += PAGE_SIZE;
444 		}
445 	}
446 	freeenv(list);
447 #if VM_NRESERVLEVEL > 0
448 	/*
449 	 * Initialize the reservation management system.
450 	 */
451 	vm_reserv_init();
452 #endif
453 	return (vaddr);
454 }
455 
456 
457 CTASSERT(offsetof(struct vm_page, aflags) % sizeof(uint32_t) == 0);
458 
459 void
460 vm_page_aflag_set(vm_page_t m, uint8_t bits)
461 {
462 	uint32_t *addr, val;
463 
464 	/*
465 	 * The PGA_WRITEABLE flag can only be set if the page is managed and
466 	 * VPO_BUSY.  Currently, this flag is only set by pmap_enter().
467 	 */
468 	KASSERT((bits & PGA_WRITEABLE) == 0 ||
469 	    (m->oflags & (VPO_UNMANAGED | VPO_BUSY)) == VPO_BUSY,
470 	    ("PGA_WRITEABLE and !VPO_BUSY"));
471 
472 	/*
473 	 * We want to use atomic updates for m->aflags, which is a
474 	 * byte wide.  Not all architectures provide atomic operations
475 	 * on the single-byte destination.  Punt and access the whole
476 	 * 4-byte word with an atomic update.  Parallel non-atomic
477 	 * updates to the fields included in the update by proximity
478 	 * are handled properly by atomics.
479 	 */
480 	addr = (void *)&m->aflags;
481 	MPASS(((uintptr_t)addr & (sizeof(uint32_t) - 1)) == 0);
482 	val = bits;
483 #if BYTE_ORDER == BIG_ENDIAN
484 	val <<= 24;
485 #endif
486 	atomic_set_32(addr, val);
487 }
488 
489 void
490 vm_page_aflag_clear(vm_page_t m, uint8_t bits)
491 {
492 	uint32_t *addr, val;
493 
494 	/*
495 	 * The PGA_REFERENCED flag can only be cleared if the object
496 	 * containing the page is locked.
497 	 */
498 	KASSERT((bits & PGA_REFERENCED) == 0 || VM_OBJECT_LOCKED(m->object),
499 	    ("PGA_REFERENCED and !VM_OBJECT_LOCKED"));
500 
501 	/*
502 	 * See the comment in vm_page_aflag_set().
503 	 */
504 	addr = (void *)&m->aflags;
505 	MPASS(((uintptr_t)addr & (sizeof(uint32_t) - 1)) == 0);
506 	val = bits;
507 #if BYTE_ORDER == BIG_ENDIAN
508 	val <<= 24;
509 #endif
510 	atomic_clear_32(addr, val);
511 }
512 
513 void
514 vm_page_reference(vm_page_t m)
515 {
516 
517 	vm_page_aflag_set(m, PGA_REFERENCED);
518 }
519 
520 void
521 vm_page_busy(vm_page_t m)
522 {
523 
524 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
525 	KASSERT((m->oflags & VPO_BUSY) == 0,
526 	    ("vm_page_busy: page already busy!!!"));
527 	m->oflags |= VPO_BUSY;
528 }
529 
530 /*
531  *      vm_page_flash:
532  *
533  *      wakeup anyone waiting for the page.
534  */
535 void
536 vm_page_flash(vm_page_t m)
537 {
538 
539 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
540 	if (m->oflags & VPO_WANTED) {
541 		m->oflags &= ~VPO_WANTED;
542 		wakeup(m);
543 	}
544 }
545 
546 /*
547  *      vm_page_wakeup:
548  *
549  *      clear the VPO_BUSY flag and wakeup anyone waiting for the
550  *      page.
551  *
552  */
553 void
554 vm_page_wakeup(vm_page_t m)
555 {
556 
557 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
558 	KASSERT(m->oflags & VPO_BUSY, ("vm_page_wakeup: page not busy!!!"));
559 	m->oflags &= ~VPO_BUSY;
560 	vm_page_flash(m);
561 }
562 
563 void
564 vm_page_io_start(vm_page_t m)
565 {
566 
567 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
568 	m->busy++;
569 }
570 
571 void
572 vm_page_io_finish(vm_page_t m)
573 {
574 
575 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
576 	KASSERT(m->busy > 0, ("vm_page_io_finish: page %p is not busy", m));
577 	m->busy--;
578 	if (m->busy == 0)
579 		vm_page_flash(m);
580 }
581 
582 /*
583  * Keep page from being freed by the page daemon
584  * much of the same effect as wiring, except much lower
585  * overhead and should be used only for *very* temporary
586  * holding ("wiring").
587  */
588 void
589 vm_page_hold(vm_page_t mem)
590 {
591 
592 	vm_page_lock_assert(mem, MA_OWNED);
593         mem->hold_count++;
594 }
595 
596 void
597 vm_page_unhold(vm_page_t mem)
598 {
599 
600 	vm_page_lock_assert(mem, MA_OWNED);
601 	--mem->hold_count;
602 	KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!"));
603 	if (mem->hold_count == 0 && mem->queue == PQ_HOLD)
604 		vm_page_free_toq(mem);
605 }
606 
607 /*
608  *	vm_page_unhold_pages:
609  *
610  *	Unhold each of the pages that is referenced by the given array.
611  */
612 void
613 vm_page_unhold_pages(vm_page_t *ma, int count)
614 {
615 	struct mtx *mtx, *new_mtx;
616 
617 	mtx = NULL;
618 	for (; count != 0; count--) {
619 		/*
620 		 * Avoid releasing and reacquiring the same page lock.
621 		 */
622 		new_mtx = vm_page_lockptr(*ma);
623 		if (mtx != new_mtx) {
624 			if (mtx != NULL)
625 				mtx_unlock(mtx);
626 			mtx = new_mtx;
627 			mtx_lock(mtx);
628 		}
629 		vm_page_unhold(*ma);
630 		ma++;
631 	}
632 	if (mtx != NULL)
633 		mtx_unlock(mtx);
634 }
635 
636 /*
637  *	vm_page_getfake:
638  *
639  *	Create a fictitious page with the specified physical address and
640  *	memory attribute.  The memory attribute is the only the machine-
641  *	dependent aspect of a fictitious page that must be initialized.
642  */
643 vm_page_t
644 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
645 {
646 	vm_page_t m;
647 
648 	m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
649 	m->phys_addr = paddr;
650 	m->queue = PQ_NONE;
651 	/* Fictitious pages don't use "segind". */
652 	m->flags = PG_FICTITIOUS;
653 	/* Fictitious pages don't use "order" or "pool". */
654 	m->oflags = VPO_BUSY | VPO_UNMANAGED;
655 	m->wire_count = 1;
656 	pmap_page_set_memattr(m, memattr);
657 	return (m);
658 }
659 
660 /*
661  *	vm_page_putfake:
662  *
663  *	Release a fictitious page.
664  */
665 void
666 vm_page_putfake(vm_page_t m)
667 {
668 
669 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
670 	    ("vm_page_putfake: bad page %p", m));
671 	uma_zfree(fakepg_zone, m);
672 }
673 
674 /*
675  *	vm_page_updatefake:
676  *
677  *	Update the given fictitious page to the specified physical address and
678  *	memory attribute.
679  */
680 void
681 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
682 {
683 
684 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
685 	    ("vm_page_updatefake: bad page %p", m));
686 	m->phys_addr = paddr;
687 	pmap_page_set_memattr(m, memattr);
688 }
689 
690 /*
691  *	vm_page_free:
692  *
693  *	Free a page.
694  */
695 void
696 vm_page_free(vm_page_t m)
697 {
698 
699 	m->flags &= ~PG_ZERO;
700 	vm_page_free_toq(m);
701 }
702 
703 /*
704  *	vm_page_free_zero:
705  *
706  *	Free a page to the zerod-pages queue
707  */
708 void
709 vm_page_free_zero(vm_page_t m)
710 {
711 
712 	m->flags |= PG_ZERO;
713 	vm_page_free_toq(m);
714 }
715 
716 /*
717  *	vm_page_sleep:
718  *
719  *	Sleep and release the page and page queues locks.
720  *
721  *	The object containing the given page must be locked.
722  */
723 void
724 vm_page_sleep(vm_page_t m, const char *msg)
725 {
726 
727 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
728 	if (mtx_owned(&vm_page_queue_mtx))
729 		vm_page_unlock_queues();
730 	if (mtx_owned(vm_page_lockptr(m)))
731 		vm_page_unlock(m);
732 
733 	/*
734 	 * It's possible that while we sleep, the page will get
735 	 * unbusied and freed.  If we are holding the object
736 	 * lock, we will assume we hold a reference to the object
737 	 * such that even if m->object changes, we can re-lock
738 	 * it.
739 	 */
740 	m->oflags |= VPO_WANTED;
741 	msleep(m, VM_OBJECT_MTX(m->object), PVM, msg, 0);
742 }
743 
744 /*
745  *	vm_page_dirty:
746  *
747  *	Set all bits in the page's dirty field.
748  *
749  *	The object containing the specified page must be locked if the
750  *	call is made from the machine-independent layer.
751  *
752  *	See vm_page_clear_dirty_mask().
753  */
754 void
755 vm_page_dirty(vm_page_t m)
756 {
757 
758 	KASSERT((m->flags & PG_CACHED) == 0,
759 	    ("vm_page_dirty: page in cache!"));
760 	KASSERT(!VM_PAGE_IS_FREE(m),
761 	    ("vm_page_dirty: page is free!"));
762 	KASSERT(m->valid == VM_PAGE_BITS_ALL,
763 	    ("vm_page_dirty: page is invalid!"));
764 	m->dirty = VM_PAGE_BITS_ALL;
765 }
766 
767 /*
768  *	vm_page_splay:
769  *
770  *	Implements Sleator and Tarjan's top-down splay algorithm.  Returns
771  *	the vm_page containing the given pindex.  If, however, that
772  *	pindex is not found in the vm_object, returns a vm_page that is
773  *	adjacent to the pindex, coming before or after it.
774  */
775 vm_page_t
776 vm_page_splay(vm_pindex_t pindex, vm_page_t root)
777 {
778 	struct vm_page dummy;
779 	vm_page_t lefttreemax, righttreemin, y;
780 
781 	if (root == NULL)
782 		return (root);
783 	lefttreemax = righttreemin = &dummy;
784 	for (;; root = y) {
785 		if (pindex < root->pindex) {
786 			if ((y = root->left) == NULL)
787 				break;
788 			if (pindex < y->pindex) {
789 				/* Rotate right. */
790 				root->left = y->right;
791 				y->right = root;
792 				root = y;
793 				if ((y = root->left) == NULL)
794 					break;
795 			}
796 			/* Link into the new root's right tree. */
797 			righttreemin->left = root;
798 			righttreemin = root;
799 		} else if (pindex > root->pindex) {
800 			if ((y = root->right) == NULL)
801 				break;
802 			if (pindex > y->pindex) {
803 				/* Rotate left. */
804 				root->right = y->left;
805 				y->left = root;
806 				root = y;
807 				if ((y = root->right) == NULL)
808 					break;
809 			}
810 			/* Link into the new root's left tree. */
811 			lefttreemax->right = root;
812 			lefttreemax = root;
813 		} else
814 			break;
815 	}
816 	/* Assemble the new root. */
817 	lefttreemax->right = root->left;
818 	righttreemin->left = root->right;
819 	root->left = dummy.right;
820 	root->right = dummy.left;
821 	return (root);
822 }
823 
824 /*
825  *	vm_page_insert:		[ internal use only ]
826  *
827  *	Inserts the given mem entry into the object and object list.
828  *
829  *	The pagetables are not updated but will presumably fault the page
830  *	in if necessary, or if a kernel page the caller will at some point
831  *	enter the page into the kernel's pmap.  We are not allowed to block
832  *	here so we *can't* do this anyway.
833  *
834  *	The object and page must be locked.
835  *	This routine may not block.
836  */
837 void
838 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
839 {
840 	vm_page_t root;
841 
842 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
843 	if (m->object != NULL)
844 		panic("vm_page_insert: page already inserted");
845 
846 	/*
847 	 * Record the object/offset pair in this page
848 	 */
849 	m->object = object;
850 	m->pindex = pindex;
851 
852 	/*
853 	 * Now link into the object's ordered list of backed pages.
854 	 */
855 	root = object->root;
856 	if (root == NULL) {
857 		m->left = NULL;
858 		m->right = NULL;
859 		TAILQ_INSERT_TAIL(&object->memq, m, listq);
860 	} else {
861 		root = vm_page_splay(pindex, root);
862 		if (pindex < root->pindex) {
863 			m->left = root->left;
864 			m->right = root;
865 			root->left = NULL;
866 			TAILQ_INSERT_BEFORE(root, m, listq);
867 		} else if (pindex == root->pindex)
868 			panic("vm_page_insert: offset already allocated");
869 		else {
870 			m->right = root->right;
871 			m->left = root;
872 			root->right = NULL;
873 			TAILQ_INSERT_AFTER(&object->memq, root, m, listq);
874 		}
875 	}
876 	object->root = m;
877 
878 	/*
879 	 * show that the object has one more resident page.
880 	 */
881 	object->resident_page_count++;
882 	/*
883 	 * Hold the vnode until the last page is released.
884 	 */
885 	if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
886 		vhold((struct vnode *)object->handle);
887 
888 	/*
889 	 * Since we are inserting a new and possibly dirty page,
890 	 * update the object's OBJ_MIGHTBEDIRTY flag.
891 	 */
892 	if (m->aflags & PGA_WRITEABLE)
893 		vm_object_set_writeable_dirty(object);
894 }
895 
896 /*
897  *	vm_page_remove:
898  *				NOTE: used by device pager as well -wfj
899  *
900  *	Removes the given mem entry from the object/offset-page
901  *	table and the object page list, but do not invalidate/terminate
902  *	the backing store.
903  *
904  *	The object and page must be locked.
905  *	The underlying pmap entry (if any) is NOT removed here.
906  *	This routine may not block.
907  */
908 void
909 vm_page_remove(vm_page_t m)
910 {
911 	vm_object_t object;
912 	vm_page_t next, prev, root;
913 
914 	if ((m->oflags & VPO_UNMANAGED) == 0)
915 		vm_page_lock_assert(m, MA_OWNED);
916 	if ((object = m->object) == NULL)
917 		return;
918 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
919 	if (m->oflags & VPO_BUSY) {
920 		m->oflags &= ~VPO_BUSY;
921 		vm_page_flash(m);
922 	}
923 
924 	/*
925 	 * Now remove from the object's list of backed pages.
926 	 */
927 	if ((next = TAILQ_NEXT(m, listq)) != NULL && next->left == m) {
928 		/*
929 		 * Since the page's successor in the list is also its parent
930 		 * in the tree, its right subtree must be empty.
931 		 */
932 		next->left = m->left;
933 		KASSERT(m->right == NULL,
934 		    ("vm_page_remove: page %p has right child", m));
935 	} else if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL &&
936 	    prev->right == m) {
937 		/*
938 		 * Since the page's predecessor in the list is also its parent
939 		 * in the tree, its left subtree must be empty.
940 		 */
941 		KASSERT(m->left == NULL,
942 		    ("vm_page_remove: page %p has left child", m));
943 		prev->right = m->right;
944 	} else {
945 		if (m != object->root)
946 			vm_page_splay(m->pindex, object->root);
947 		if (m->left == NULL)
948 			root = m->right;
949 		else if (m->right == NULL)
950 			root = m->left;
951 		else {
952 			/*
953 			 * Move the page's successor to the root, because
954 			 * pages are usually removed in ascending order.
955 			 */
956 			if (m->right != next)
957 				vm_page_splay(m->pindex, m->right);
958 			next->left = m->left;
959 			root = next;
960 		}
961 		object->root = root;
962 	}
963 	TAILQ_REMOVE(&object->memq, m, listq);
964 
965 	/*
966 	 * And show that the object has one fewer resident page.
967 	 */
968 	object->resident_page_count--;
969 	/*
970 	 * The vnode may now be recycled.
971 	 */
972 	if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
973 		vdrop((struct vnode *)object->handle);
974 
975 	m->object = NULL;
976 }
977 
978 /*
979  *	vm_page_lookup:
980  *
981  *	Returns the page associated with the object/offset
982  *	pair specified; if none is found, NULL is returned.
983  *
984  *	The object must be locked.
985  *	This routine may not block.
986  *	This is a critical path routine
987  */
988 vm_page_t
989 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
990 {
991 	vm_page_t m;
992 
993 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
994 	if ((m = object->root) != NULL && m->pindex != pindex) {
995 		m = vm_page_splay(pindex, m);
996 		if ((object->root = m)->pindex != pindex)
997 			m = NULL;
998 	}
999 	return (m);
1000 }
1001 
1002 /*
1003  *	vm_page_find_least:
1004  *
1005  *	Returns the page associated with the object with least pindex
1006  *	greater than or equal to the parameter pindex, or NULL.
1007  *
1008  *	The object must be locked.
1009  *	The routine may not block.
1010  */
1011 vm_page_t
1012 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1013 {
1014 	vm_page_t m;
1015 
1016 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1017 	if ((m = TAILQ_FIRST(&object->memq)) != NULL) {
1018 		if (m->pindex < pindex) {
1019 			m = vm_page_splay(pindex, object->root);
1020 			if ((object->root = m)->pindex < pindex)
1021 				m = TAILQ_NEXT(m, listq);
1022 		}
1023 	}
1024 	return (m);
1025 }
1026 
1027 /*
1028  * Returns the given page's successor (by pindex) within the object if it is
1029  * resident; if none is found, NULL is returned.
1030  *
1031  * The object must be locked.
1032  */
1033 vm_page_t
1034 vm_page_next(vm_page_t m)
1035 {
1036 	vm_page_t next;
1037 
1038 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1039 	if ((next = TAILQ_NEXT(m, listq)) != NULL &&
1040 	    next->pindex != m->pindex + 1)
1041 		next = NULL;
1042 	return (next);
1043 }
1044 
1045 /*
1046  * Returns the given page's predecessor (by pindex) within the object if it is
1047  * resident; if none is found, NULL is returned.
1048  *
1049  * The object must be locked.
1050  */
1051 vm_page_t
1052 vm_page_prev(vm_page_t m)
1053 {
1054 	vm_page_t prev;
1055 
1056 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1057 	if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL &&
1058 	    prev->pindex != m->pindex - 1)
1059 		prev = NULL;
1060 	return (prev);
1061 }
1062 
1063 /*
1064  *	vm_page_rename:
1065  *
1066  *	Move the given memory entry from its
1067  *	current object to the specified target object/offset.
1068  *
1069  *	The object must be locked.
1070  *	This routine may not block.
1071  *
1072  *	Note: swap associated with the page must be invalidated by the move.  We
1073  *	      have to do this for several reasons:  (1) we aren't freeing the
1074  *	      page, (2) we are dirtying the page, (3) the VM system is probably
1075  *	      moving the page from object A to B, and will then later move
1076  *	      the backing store from A to B and we can't have a conflict.
1077  *
1078  *	Note: we *always* dirty the page.  It is necessary both for the
1079  *	      fact that we moved it, and because we may be invalidating
1080  *	      swap.  If the page is on the cache, we have to deactivate it
1081  *	      or vm_page_dirty() will panic.  Dirty pages are not allowed
1082  *	      on the cache.
1083  */
1084 void
1085 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1086 {
1087 
1088 	vm_page_remove(m);
1089 	vm_page_insert(m, new_object, new_pindex);
1090 	vm_page_dirty(m);
1091 }
1092 
1093 /*
1094  *	Convert all of the given object's cached pages that have a
1095  *	pindex within the given range into free pages.  If the value
1096  *	zero is given for "end", then the range's upper bound is
1097  *	infinity.  If the given object is backed by a vnode and it
1098  *	transitions from having one or more cached pages to none, the
1099  *	vnode's hold count is reduced.
1100  */
1101 void
1102 vm_page_cache_free(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1103 {
1104 	vm_page_t m, m_next;
1105 	boolean_t empty;
1106 
1107 	mtx_lock(&vm_page_queue_free_mtx);
1108 	if (__predict_false(object->cache == NULL)) {
1109 		mtx_unlock(&vm_page_queue_free_mtx);
1110 		return;
1111 	}
1112 	m = object->cache = vm_page_splay(start, object->cache);
1113 	if (m->pindex < start) {
1114 		if (m->right == NULL)
1115 			m = NULL;
1116 		else {
1117 			m_next = vm_page_splay(start, m->right);
1118 			m_next->left = m;
1119 			m->right = NULL;
1120 			m = object->cache = m_next;
1121 		}
1122 	}
1123 
1124 	/*
1125 	 * At this point, "m" is either (1) a reference to the page
1126 	 * with the least pindex that is greater than or equal to
1127 	 * "start" or (2) NULL.
1128 	 */
1129 	for (; m != NULL && (m->pindex < end || end == 0); m = m_next) {
1130 		/*
1131 		 * Find "m"'s successor and remove "m" from the
1132 		 * object's cache.
1133 		 */
1134 		if (m->right == NULL) {
1135 			object->cache = m->left;
1136 			m_next = NULL;
1137 		} else {
1138 			m_next = vm_page_splay(start, m->right);
1139 			m_next->left = m->left;
1140 			object->cache = m_next;
1141 		}
1142 		/* Convert "m" to a free page. */
1143 		m->object = NULL;
1144 		m->valid = 0;
1145 		/* Clear PG_CACHED and set PG_FREE. */
1146 		m->flags ^= PG_CACHED | PG_FREE;
1147 		KASSERT((m->flags & (PG_CACHED | PG_FREE)) == PG_FREE,
1148 		    ("vm_page_cache_free: page %p has inconsistent flags", m));
1149 		cnt.v_cache_count--;
1150 		cnt.v_free_count++;
1151 	}
1152 	empty = object->cache == NULL;
1153 	mtx_unlock(&vm_page_queue_free_mtx);
1154 	if (object->type == OBJT_VNODE && empty)
1155 		vdrop(object->handle);
1156 }
1157 
1158 /*
1159  *	Returns the cached page that is associated with the given
1160  *	object and offset.  If, however, none exists, returns NULL.
1161  *
1162  *	The free page queue must be locked.
1163  */
1164 static inline vm_page_t
1165 vm_page_cache_lookup(vm_object_t object, vm_pindex_t pindex)
1166 {
1167 	vm_page_t m;
1168 
1169 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1170 	if ((m = object->cache) != NULL && m->pindex != pindex) {
1171 		m = vm_page_splay(pindex, m);
1172 		if ((object->cache = m)->pindex != pindex)
1173 			m = NULL;
1174 	}
1175 	return (m);
1176 }
1177 
1178 /*
1179  *	Remove the given cached page from its containing object's
1180  *	collection of cached pages.
1181  *
1182  *	The free page queue must be locked.
1183  */
1184 void
1185 vm_page_cache_remove(vm_page_t m)
1186 {
1187 	vm_object_t object;
1188 	vm_page_t root;
1189 
1190 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1191 	KASSERT((m->flags & PG_CACHED) != 0,
1192 	    ("vm_page_cache_remove: page %p is not cached", m));
1193 	object = m->object;
1194 	if (m != object->cache) {
1195 		root = vm_page_splay(m->pindex, object->cache);
1196 		KASSERT(root == m,
1197 		    ("vm_page_cache_remove: page %p is not cached in object %p",
1198 		    m, object));
1199 	}
1200 	if (m->left == NULL)
1201 		root = m->right;
1202 	else if (m->right == NULL)
1203 		root = m->left;
1204 	else {
1205 		root = vm_page_splay(m->pindex, m->left);
1206 		root->right = m->right;
1207 	}
1208 	object->cache = root;
1209 	m->object = NULL;
1210 	cnt.v_cache_count--;
1211 }
1212 
1213 /*
1214  *	Transfer all of the cached pages with offset greater than or
1215  *	equal to 'offidxstart' from the original object's cache to the
1216  *	new object's cache.  However, any cached pages with offset
1217  *	greater than or equal to the new object's size are kept in the
1218  *	original object.  Initially, the new object's cache must be
1219  *	empty.  Offset 'offidxstart' in the original object must
1220  *	correspond to offset zero in the new object.
1221  *
1222  *	The new object must be locked.
1223  */
1224 void
1225 vm_page_cache_transfer(vm_object_t orig_object, vm_pindex_t offidxstart,
1226     vm_object_t new_object)
1227 {
1228 	vm_page_t m, m_next;
1229 
1230 	/*
1231 	 * Insertion into an object's collection of cached pages
1232 	 * requires the object to be locked.  In contrast, removal does
1233 	 * not.
1234 	 */
1235 	VM_OBJECT_LOCK_ASSERT(new_object, MA_OWNED);
1236 	KASSERT(new_object->cache == NULL,
1237 	    ("vm_page_cache_transfer: object %p has cached pages",
1238 	    new_object));
1239 	mtx_lock(&vm_page_queue_free_mtx);
1240 	if ((m = orig_object->cache) != NULL) {
1241 		/*
1242 		 * Transfer all of the pages with offset greater than or
1243 		 * equal to 'offidxstart' from the original object's
1244 		 * cache to the new object's cache.
1245 		 */
1246 		m = vm_page_splay(offidxstart, m);
1247 		if (m->pindex < offidxstart) {
1248 			orig_object->cache = m;
1249 			new_object->cache = m->right;
1250 			m->right = NULL;
1251 		} else {
1252 			orig_object->cache = m->left;
1253 			new_object->cache = m;
1254 			m->left = NULL;
1255 		}
1256 		while ((m = new_object->cache) != NULL) {
1257 			if ((m->pindex - offidxstart) >= new_object->size) {
1258 				/*
1259 				 * Return all of the cached pages with
1260 				 * offset greater than or equal to the
1261 				 * new object's size to the original
1262 				 * object's cache.
1263 				 */
1264 				new_object->cache = m->left;
1265 				m->left = orig_object->cache;
1266 				orig_object->cache = m;
1267 				break;
1268 			}
1269 			m_next = vm_page_splay(m->pindex, m->right);
1270 			/* Update the page's object and offset. */
1271 			m->object = new_object;
1272 			m->pindex -= offidxstart;
1273 			if (m_next == NULL)
1274 				break;
1275 			m->right = NULL;
1276 			m_next->left = m;
1277 			new_object->cache = m_next;
1278 		}
1279 		KASSERT(new_object->cache == NULL ||
1280 		    new_object->type == OBJT_SWAP,
1281 		    ("vm_page_cache_transfer: object %p's type is incompatible"
1282 		    " with cached pages", new_object));
1283 	}
1284 	mtx_unlock(&vm_page_queue_free_mtx);
1285 }
1286 
1287 /*
1288  *	vm_page_alloc:
1289  *
1290  *	Allocate and return a page that is associated with the specified
1291  *	object and offset pair.  By default, this page has the flag VPO_BUSY
1292  *	set.
1293  *
1294  *	The caller must always specify an allocation class.
1295  *
1296  *	allocation classes:
1297  *	VM_ALLOC_NORMAL		normal process request
1298  *	VM_ALLOC_SYSTEM		system *really* needs a page
1299  *	VM_ALLOC_INTERRUPT	interrupt time request
1300  *
1301  *	optional allocation flags:
1302  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
1303  *				intends to allocate
1304  *	VM_ALLOC_IFCACHED	return page only if it is cached
1305  *	VM_ALLOC_IFNOTCACHED	return NULL, do not reactivate if the page
1306  *				is cached
1307  *	VM_ALLOC_NOBUSY		do not set the flag VPO_BUSY on the page
1308  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
1309  *	VM_ALLOC_NOOBJ		page is not associated with an object and
1310  *				should not have the flag VPO_BUSY set
1311  *	VM_ALLOC_WIRED		wire the allocated page
1312  *	VM_ALLOC_ZERO		prefer a zeroed page
1313  *
1314  *	This routine may not sleep.
1315  */
1316 vm_page_t
1317 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1318 {
1319 	struct vnode *vp = NULL;
1320 	vm_object_t m_object;
1321 	vm_page_t m;
1322 	int flags, req_class;
1323 
1324 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0),
1325 	    ("vm_page_alloc: inconsistent object/req"));
1326 	if (object != NULL)
1327 		VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1328 
1329 	req_class = req & VM_ALLOC_CLASS_MASK;
1330 
1331 	/*
1332 	 * The page daemon is allowed to dig deeper into the free page list.
1333 	 */
1334 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1335 		req_class = VM_ALLOC_SYSTEM;
1336 
1337 	mtx_lock(&vm_page_queue_free_mtx);
1338 	if (cnt.v_free_count + cnt.v_cache_count > cnt.v_free_reserved ||
1339 	    (req_class == VM_ALLOC_SYSTEM &&
1340 	    cnt.v_free_count + cnt.v_cache_count > cnt.v_interrupt_free_min) ||
1341 	    (req_class == VM_ALLOC_INTERRUPT &&
1342 	    cnt.v_free_count + cnt.v_cache_count > 0)) {
1343 		/*
1344 		 * Allocate from the free queue if the number of free pages
1345 		 * exceeds the minimum for the request class.
1346 		 */
1347 		if (object != NULL &&
1348 		    (m = vm_page_cache_lookup(object, pindex)) != NULL) {
1349 			if ((req & VM_ALLOC_IFNOTCACHED) != 0) {
1350 				mtx_unlock(&vm_page_queue_free_mtx);
1351 				return (NULL);
1352 			}
1353 			if (vm_phys_unfree_page(m))
1354 				vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m, 0);
1355 #if VM_NRESERVLEVEL > 0
1356 			else if (!vm_reserv_reactivate_page(m))
1357 #else
1358 			else
1359 #endif
1360 				panic("vm_page_alloc: cache page %p is missing"
1361 				    " from the free queue", m);
1362 		} else if ((req & VM_ALLOC_IFCACHED) != 0) {
1363 			mtx_unlock(&vm_page_queue_free_mtx);
1364 			return (NULL);
1365 #if VM_NRESERVLEVEL > 0
1366 		} else if (object == NULL || object->type == OBJT_DEVICE ||
1367 		    object->type == OBJT_SG ||
1368 		    (object->flags & OBJ_COLORED) == 0 ||
1369 		    (m = vm_reserv_alloc_page(object, pindex)) == NULL) {
1370 #else
1371 		} else {
1372 #endif
1373 			m = vm_phys_alloc_pages(object != NULL ?
1374 			    VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0);
1375 #if VM_NRESERVLEVEL > 0
1376 			if (m == NULL && vm_reserv_reclaim_inactive()) {
1377 				m = vm_phys_alloc_pages(object != NULL ?
1378 				    VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT,
1379 				    0);
1380 			}
1381 #endif
1382 		}
1383 	} else {
1384 		/*
1385 		 * Not allocatable, give up.
1386 		 */
1387 		mtx_unlock(&vm_page_queue_free_mtx);
1388 		atomic_add_int(&vm_pageout_deficit,
1389 		    max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
1390 		pagedaemon_wakeup();
1391 		return (NULL);
1392 	}
1393 
1394 	/*
1395 	 *  At this point we had better have found a good page.
1396 	 */
1397 	KASSERT(m != NULL, ("vm_page_alloc: missing page"));
1398 	KASSERT(m->queue == PQ_NONE,
1399 	    ("vm_page_alloc: page %p has unexpected queue %d", m, m->queue));
1400 	KASSERT(m->wire_count == 0, ("vm_page_alloc: page %p is wired", m));
1401 	KASSERT(m->hold_count == 0, ("vm_page_alloc: page %p is held", m));
1402 	KASSERT(m->busy == 0, ("vm_page_alloc: page %p is busy", m));
1403 	KASSERT(m->dirty == 0, ("vm_page_alloc: page %p is dirty", m));
1404 	KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
1405 	    ("vm_page_alloc: page %p has unexpected memattr %d", m,
1406 	    pmap_page_get_memattr(m)));
1407 	if ((m->flags & PG_CACHED) != 0) {
1408 		KASSERT((m->flags & PG_ZERO) == 0,
1409 		    ("vm_page_alloc: cached page %p is PG_ZERO", m));
1410 		KASSERT(m->valid != 0,
1411 		    ("vm_page_alloc: cached page %p is invalid", m));
1412 		if (m->object == object && m->pindex == pindex)
1413 	  		cnt.v_reactivated++;
1414 		else
1415 			m->valid = 0;
1416 		m_object = m->object;
1417 		vm_page_cache_remove(m);
1418 		if (m_object->type == OBJT_VNODE && m_object->cache == NULL)
1419 			vp = m_object->handle;
1420 	} else {
1421 		KASSERT(VM_PAGE_IS_FREE(m),
1422 		    ("vm_page_alloc: page %p is not free", m));
1423 		KASSERT(m->valid == 0,
1424 		    ("vm_page_alloc: free page %p is valid", m));
1425 		cnt.v_free_count--;
1426 	}
1427 
1428 	/*
1429 	 * Only the PG_ZERO flag is inherited.  The PG_CACHED or PG_FREE flag
1430 	 * must be cleared before the free page queues lock is released.
1431 	 */
1432 	flags = 0;
1433 	if (req & VM_ALLOC_NODUMP)
1434 		flags |= PG_NODUMP;
1435 	if (m->flags & PG_ZERO) {
1436 		vm_page_zero_count--;
1437 		if (req & VM_ALLOC_ZERO)
1438 			flags = PG_ZERO;
1439 	}
1440 	m->flags = flags;
1441 	mtx_unlock(&vm_page_queue_free_mtx);
1442 	m->aflags = 0;
1443 	if (object == NULL || object->type == OBJT_PHYS)
1444 		m->oflags = VPO_UNMANAGED;
1445 	else
1446 		m->oflags = 0;
1447 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ)) == 0)
1448 		m->oflags |= VPO_BUSY;
1449 	if (req & VM_ALLOC_WIRED) {
1450 		/*
1451 		 * The page lock is not required for wiring a page until that
1452 		 * page is inserted into the object.
1453 		 */
1454 		atomic_add_int(&cnt.v_wire_count, 1);
1455 		m->wire_count = 1;
1456 	}
1457 	m->act_count = 0;
1458 
1459 	if (object != NULL) {
1460 		/* Ignore device objects; the pager sets "memattr" for them. */
1461 		if (object->memattr != VM_MEMATTR_DEFAULT &&
1462 		    object->type != OBJT_DEVICE && object->type != OBJT_SG)
1463 			pmap_page_set_memattr(m, object->memattr);
1464 		vm_page_insert(m, object, pindex);
1465 	} else
1466 		m->pindex = pindex;
1467 
1468 	/*
1469 	 * The following call to vdrop() must come after the above call
1470 	 * to vm_page_insert() in case both affect the same object and
1471 	 * vnode.  Otherwise, the affected vnode's hold count could
1472 	 * temporarily become zero.
1473 	 */
1474 	if (vp != NULL)
1475 		vdrop(vp);
1476 
1477 	/*
1478 	 * Don't wakeup too often - wakeup the pageout daemon when
1479 	 * we would be nearly out of memory.
1480 	 */
1481 	if (vm_paging_needed())
1482 		pagedaemon_wakeup();
1483 
1484 	return (m);
1485 }
1486 
1487 /*
1488  *	vm_page_alloc_contig:
1489  *
1490  *	Allocate a contiguous set of physical pages of the given size "npages"
1491  *	from the free lists.  All of the physical pages must be at or above
1492  *	the given physical address "low" and below the given physical address
1493  *	"high".  The given value "alignment" determines the alignment of the
1494  *	first physical page in the set.  If the given value "boundary" is
1495  *	non-zero, then the set of physical pages cannot cross any physical
1496  *	address boundary that is a multiple of that value.  Both "alignment"
1497  *	and "boundary" must be a power of two.
1498  *
1499  *	If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
1500  *	then the memory attribute setting for the physical pages is configured
1501  *	to the object's memory attribute setting.  Otherwise, the memory
1502  *	attribute setting for the physical pages is configured to "memattr",
1503  *	overriding the object's memory attribute setting.  However, if the
1504  *	object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
1505  *	memory attribute setting for the physical pages cannot be configured
1506  *	to VM_MEMATTR_DEFAULT.
1507  *
1508  *	The caller must always specify an allocation class.
1509  *
1510  *	allocation classes:
1511  *	VM_ALLOC_NORMAL		normal process request
1512  *	VM_ALLOC_SYSTEM		system *really* needs a page
1513  *	VM_ALLOC_INTERRUPT	interrupt time request
1514  *
1515  *	optional allocation flags:
1516  *	VM_ALLOC_NOBUSY		do not set the flag VPO_BUSY on the page
1517  *	VM_ALLOC_NOOBJ		page is not associated with an object and
1518  *				should not have the flag VPO_BUSY set
1519  *	VM_ALLOC_WIRED		wire the allocated page
1520  *	VM_ALLOC_ZERO		prefer a zeroed page
1521  *
1522  *	This routine may not sleep.
1523  */
1524 vm_page_t
1525 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
1526     u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
1527     vm_paddr_t boundary, vm_memattr_t memattr)
1528 {
1529 	struct vnode *drop;
1530 	vm_page_t deferred_vdrop_list, m, m_ret;
1531 	u_int flags, oflags;
1532 	int req_class;
1533 
1534 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0),
1535 	    ("vm_page_alloc_contig: inconsistent object/req"));
1536 	if (object != NULL) {
1537 		VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1538 		KASSERT(object->type == OBJT_PHYS,
1539 		    ("vm_page_alloc_contig: object %p isn't OBJT_PHYS",
1540 		    object));
1541 	}
1542 	KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
1543 	req_class = req & VM_ALLOC_CLASS_MASK;
1544 
1545 	/*
1546 	 * The page daemon is allowed to dig deeper into the free page list.
1547 	 */
1548 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1549 		req_class = VM_ALLOC_SYSTEM;
1550 
1551 	deferred_vdrop_list = NULL;
1552 	mtx_lock(&vm_page_queue_free_mtx);
1553 	if (cnt.v_free_count + cnt.v_cache_count >= npages +
1554 	    cnt.v_free_reserved || (req_class == VM_ALLOC_SYSTEM &&
1555 	    cnt.v_free_count + cnt.v_cache_count >= npages +
1556 	    cnt.v_interrupt_free_min) || (req_class == VM_ALLOC_INTERRUPT &&
1557 	    cnt.v_free_count + cnt.v_cache_count >= npages)) {
1558 #if VM_NRESERVLEVEL > 0
1559 retry:
1560 		if (object == NULL || (object->flags & OBJ_COLORED) == 0 ||
1561 		    (m_ret = vm_reserv_alloc_contig(object, pindex, npages,
1562 		    low, high, alignment, boundary)) == NULL)
1563 #endif
1564 			m_ret = vm_phys_alloc_contig(npages, low, high,
1565 			    alignment, boundary);
1566 	} else {
1567 		mtx_unlock(&vm_page_queue_free_mtx);
1568 		atomic_add_int(&vm_pageout_deficit, npages);
1569 		pagedaemon_wakeup();
1570 		return (NULL);
1571 	}
1572 	if (m_ret != NULL)
1573 		for (m = m_ret; m < &m_ret[npages]; m++) {
1574 			drop = vm_page_alloc_init(m);
1575 			if (drop != NULL) {
1576 				/*
1577 				 * Enqueue the vnode for deferred vdrop().
1578 				 *
1579 				 * Once the pages are removed from the free
1580 				 * page list, "pageq" can be safely abused to
1581 				 * construct a short-lived list of vnodes.
1582 				 */
1583 				m->pageq.tqe_prev = (void *)drop;
1584 				m->pageq.tqe_next = deferred_vdrop_list;
1585 				deferred_vdrop_list = m;
1586 			}
1587 		}
1588 	else {
1589 #if VM_NRESERVLEVEL > 0
1590 		if (vm_reserv_reclaim_contig(npages, low, high, alignment,
1591 		    boundary))
1592 			goto retry;
1593 #endif
1594 	}
1595 	mtx_unlock(&vm_page_queue_free_mtx);
1596 	if (m_ret == NULL)
1597 		return (NULL);
1598 
1599 	/*
1600 	 * Initialize the pages.  Only the PG_ZERO flag is inherited.
1601 	 */
1602 	flags = 0;
1603 	if ((req & VM_ALLOC_ZERO) != 0)
1604 		flags = PG_ZERO;
1605 	if ((req & VM_ALLOC_NODUMP) != 0)
1606 		flags |= PG_NODUMP;
1607 	if ((req & VM_ALLOC_WIRED) != 0)
1608 		atomic_add_int(&cnt.v_wire_count, npages);
1609 	oflags = VPO_UNMANAGED;
1610 	if (object != NULL) {
1611 		if ((req & VM_ALLOC_NOBUSY) == 0)
1612 			oflags |= VPO_BUSY;
1613 		if (object->memattr != VM_MEMATTR_DEFAULT &&
1614 		    memattr == VM_MEMATTR_DEFAULT)
1615 			memattr = object->memattr;
1616 	}
1617 	for (m = m_ret; m < &m_ret[npages]; m++) {
1618 		m->aflags = 0;
1619 		m->flags &= flags;
1620 		if ((req & VM_ALLOC_WIRED) != 0)
1621 			m->wire_count = 1;
1622 		/* Unmanaged pages don't use "act_count". */
1623 		m->oflags = oflags;
1624 		if (memattr != VM_MEMATTR_DEFAULT)
1625 			pmap_page_set_memattr(m, memattr);
1626 		if (object != NULL)
1627 			vm_page_insert(m, object, pindex);
1628 		else
1629 			m->pindex = pindex;
1630 		pindex++;
1631 	}
1632 	while (deferred_vdrop_list != NULL) {
1633 		vdrop((struct vnode *)deferred_vdrop_list->pageq.tqe_prev);
1634 		deferred_vdrop_list = deferred_vdrop_list->pageq.tqe_next;
1635 	}
1636 	if (vm_paging_needed())
1637 		pagedaemon_wakeup();
1638 	return (m_ret);
1639 }
1640 
1641 /*
1642  * Initialize a page that has been freshly dequeued from a freelist.
1643  * The caller has to drop the vnode returned, if it is not NULL.
1644  *
1645  * This function may only be used to initialize unmanaged pages.
1646  *
1647  * To be called with vm_page_queue_free_mtx held.
1648  */
1649 static struct vnode *
1650 vm_page_alloc_init(vm_page_t m)
1651 {
1652 	struct vnode *drop;
1653 	vm_object_t m_object;
1654 
1655 	KASSERT(m->queue == PQ_NONE,
1656 	    ("vm_page_alloc_init: page %p has unexpected queue %d",
1657 	    m, m->queue));
1658 	KASSERT(m->wire_count == 0,
1659 	    ("vm_page_alloc_init: page %p is wired", m));
1660 	KASSERT(m->hold_count == 0,
1661 	    ("vm_page_alloc_init: page %p is held", m));
1662 	KASSERT(m->busy == 0,
1663 	    ("vm_page_alloc_init: page %p is busy", m));
1664 	KASSERT(m->dirty == 0,
1665 	    ("vm_page_alloc_init: page %p is dirty", m));
1666 	KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
1667 	    ("vm_page_alloc_init: page %p has unexpected memattr %d",
1668 	    m, pmap_page_get_memattr(m)));
1669 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1670 	drop = NULL;
1671 	if ((m->flags & PG_CACHED) != 0) {
1672 		KASSERT((m->flags & PG_ZERO) == 0,
1673 		    ("vm_page_alloc_init: cached page %p is PG_ZERO", m));
1674 		m->valid = 0;
1675 		m_object = m->object;
1676 		vm_page_cache_remove(m);
1677 		if (m_object->type == OBJT_VNODE && m_object->cache == NULL)
1678 			drop = m_object->handle;
1679 	} else {
1680 		KASSERT(VM_PAGE_IS_FREE(m),
1681 		    ("vm_page_alloc_init: page %p is not free", m));
1682 		KASSERT(m->valid == 0,
1683 		    ("vm_page_alloc_init: free page %p is valid", m));
1684 		cnt.v_free_count--;
1685 		if ((m->flags & PG_ZERO) != 0)
1686 			vm_page_zero_count--;
1687 	}
1688 	/* Don't clear the PG_ZERO flag; we'll need it later. */
1689 	m->flags &= PG_ZERO;
1690 	return (drop);
1691 }
1692 
1693 /*
1694  * 	vm_page_alloc_freelist:
1695  *
1696  *	Allocate a physical page from the specified free page list.
1697  *
1698  *	The caller must always specify an allocation class.
1699  *
1700  *	allocation classes:
1701  *	VM_ALLOC_NORMAL		normal process request
1702  *	VM_ALLOC_SYSTEM		system *really* needs a page
1703  *	VM_ALLOC_INTERRUPT	interrupt time request
1704  *
1705  *	optional allocation flags:
1706  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
1707  *				intends to allocate
1708  *	VM_ALLOC_WIRED		wire the allocated page
1709  *	VM_ALLOC_ZERO		prefer a zeroed page
1710  *
1711  *	This routine may not sleep.
1712  */
1713 vm_page_t
1714 vm_page_alloc_freelist(int flind, int req)
1715 {
1716 	struct vnode *drop;
1717 	vm_page_t m;
1718 	u_int flags;
1719 	int req_class;
1720 
1721 	req_class = req & VM_ALLOC_CLASS_MASK;
1722 
1723 	/*
1724 	 * The page daemon is allowed to dig deeper into the free page list.
1725 	 */
1726 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1727 		req_class = VM_ALLOC_SYSTEM;
1728 
1729 	/*
1730 	 * Do not allocate reserved pages unless the req has asked for it.
1731 	 */
1732 	mtx_lock(&vm_page_queue_free_mtx);
1733 	if (cnt.v_free_count + cnt.v_cache_count > cnt.v_free_reserved ||
1734 	    (req_class == VM_ALLOC_SYSTEM &&
1735 	    cnt.v_free_count + cnt.v_cache_count > cnt.v_interrupt_free_min) ||
1736 	    (req_class == VM_ALLOC_INTERRUPT &&
1737 	    cnt.v_free_count + cnt.v_cache_count > 0))
1738 		m = vm_phys_alloc_freelist_pages(flind, VM_FREEPOOL_DIRECT, 0);
1739 	else {
1740 		mtx_unlock(&vm_page_queue_free_mtx);
1741 		atomic_add_int(&vm_pageout_deficit,
1742 		    max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
1743 		pagedaemon_wakeup();
1744 		return (NULL);
1745 	}
1746 	if (m == NULL) {
1747 		mtx_unlock(&vm_page_queue_free_mtx);
1748 		return (NULL);
1749 	}
1750 	drop = vm_page_alloc_init(m);
1751 	mtx_unlock(&vm_page_queue_free_mtx);
1752 
1753 	/*
1754 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
1755 	 */
1756 	m->aflags = 0;
1757 	flags = 0;
1758 	if ((req & VM_ALLOC_ZERO) != 0)
1759 		flags = PG_ZERO;
1760 	m->flags &= flags;
1761 	if ((req & VM_ALLOC_WIRED) != 0) {
1762 		/*
1763 		 * The page lock is not required for wiring a page that does
1764 		 * not belong to an object.
1765 		 */
1766 		atomic_add_int(&cnt.v_wire_count, 1);
1767 		m->wire_count = 1;
1768 	}
1769 	/* Unmanaged pages don't use "act_count". */
1770 	m->oflags = VPO_UNMANAGED;
1771 	if (drop != NULL)
1772 		vdrop(drop);
1773 	if (vm_paging_needed())
1774 		pagedaemon_wakeup();
1775 	return (m);
1776 }
1777 
1778 /*
1779  *	vm_wait:	(also see VM_WAIT macro)
1780  *
1781  *	Block until free pages are available for allocation
1782  *	- Called in various places before memory allocations.
1783  */
1784 void
1785 vm_wait(void)
1786 {
1787 
1788 	mtx_lock(&vm_page_queue_free_mtx);
1789 	if (curproc == pageproc) {
1790 		vm_pageout_pages_needed = 1;
1791 		msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx,
1792 		    PDROP | PSWP, "VMWait", 0);
1793 	} else {
1794 		if (!vm_pages_needed) {
1795 			vm_pages_needed = 1;
1796 			wakeup(&vm_pages_needed);
1797 		}
1798 		msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PVM,
1799 		    "vmwait", 0);
1800 	}
1801 }
1802 
1803 /*
1804  *	vm_waitpfault:	(also see VM_WAITPFAULT macro)
1805  *
1806  *	Block until free pages are available for allocation
1807  *	- Called only in vm_fault so that processes page faulting
1808  *	  can be easily tracked.
1809  *	- Sleeps at a lower priority than vm_wait() so that vm_wait()ing
1810  *	  processes will be able to grab memory first.  Do not change
1811  *	  this balance without careful testing first.
1812  */
1813 void
1814 vm_waitpfault(void)
1815 {
1816 
1817 	mtx_lock(&vm_page_queue_free_mtx);
1818 	if (!vm_pages_needed) {
1819 		vm_pages_needed = 1;
1820 		wakeup(&vm_pages_needed);
1821 	}
1822 	msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PUSER,
1823 	    "pfault", 0);
1824 }
1825 
1826 /*
1827  *	vm_page_requeue:
1828  *
1829  *	Move the given page to the tail of its present page queue.
1830  *
1831  *	The page queues must be locked.
1832  */
1833 void
1834 vm_page_requeue(vm_page_t m)
1835 {
1836 	struct vpgqueues *vpq;
1837 	int queue;
1838 
1839 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1840 	queue = m->queue;
1841 	KASSERT(queue != PQ_NONE,
1842 	    ("vm_page_requeue: page %p is not queued", m));
1843 	vpq = &vm_page_queues[queue];
1844 	TAILQ_REMOVE(&vpq->pl, m, pageq);
1845 	TAILQ_INSERT_TAIL(&vpq->pl, m, pageq);
1846 }
1847 
1848 /*
1849  *	vm_page_queue_remove:
1850  *
1851  *	Remove the given page from the specified queue.
1852  *
1853  *	The page and page queues must be locked.
1854  */
1855 static __inline void
1856 vm_page_queue_remove(int queue, vm_page_t m)
1857 {
1858 	struct vpgqueues *pq;
1859 
1860 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1861 	vm_page_lock_assert(m, MA_OWNED);
1862 	pq = &vm_page_queues[queue];
1863 	TAILQ_REMOVE(&pq->pl, m, pageq);
1864 	(*pq->cnt)--;
1865 }
1866 
1867 /*
1868  *	vm_pageq_remove:
1869  *
1870  *	Remove a page from its queue.
1871  *
1872  *	The given page must be locked.
1873  *	This routine may not block.
1874  */
1875 void
1876 vm_pageq_remove(vm_page_t m)
1877 {
1878 	int queue;
1879 
1880 	vm_page_lock_assert(m, MA_OWNED);
1881 	if ((queue = m->queue) != PQ_NONE) {
1882 		vm_page_lock_queues();
1883 		m->queue = PQ_NONE;
1884 		vm_page_queue_remove(queue, m);
1885 		vm_page_unlock_queues();
1886 	}
1887 }
1888 
1889 /*
1890  *	vm_page_enqueue:
1891  *
1892  *	Add the given page to the specified queue.
1893  *
1894  *	The page queues must be locked.
1895  */
1896 static void
1897 vm_page_enqueue(int queue, vm_page_t m)
1898 {
1899 	struct vpgqueues *vpq;
1900 
1901 	vpq = &vm_page_queues[queue];
1902 	m->queue = queue;
1903 	TAILQ_INSERT_TAIL(&vpq->pl, m, pageq);
1904 	++*vpq->cnt;
1905 }
1906 
1907 /*
1908  *	vm_page_activate:
1909  *
1910  *	Put the specified page on the active list (if appropriate).
1911  *	Ensure that act_count is at least ACT_INIT but do not otherwise
1912  *	mess with it.
1913  *
1914  *	The page must be locked.
1915  *	This routine may not block.
1916  */
1917 void
1918 vm_page_activate(vm_page_t m)
1919 {
1920 	int queue;
1921 
1922 	vm_page_lock_assert(m, MA_OWNED);
1923 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1924 	if ((queue = m->queue) != PQ_ACTIVE) {
1925 		if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
1926 			if (m->act_count < ACT_INIT)
1927 				m->act_count = ACT_INIT;
1928 			vm_page_lock_queues();
1929 			if (queue != PQ_NONE)
1930 				vm_page_queue_remove(queue, m);
1931 			vm_page_enqueue(PQ_ACTIVE, m);
1932 			vm_page_unlock_queues();
1933 		} else
1934 			KASSERT(queue == PQ_NONE,
1935 			    ("vm_page_activate: wired page %p is queued", m));
1936 	} else {
1937 		if (m->act_count < ACT_INIT)
1938 			m->act_count = ACT_INIT;
1939 	}
1940 }
1941 
1942 /*
1943  *	vm_page_free_wakeup:
1944  *
1945  *	Helper routine for vm_page_free_toq() and vm_page_cache().  This
1946  *	routine is called when a page has been added to the cache or free
1947  *	queues.
1948  *
1949  *	The page queues must be locked.
1950  *	This routine may not block.
1951  */
1952 static inline void
1953 vm_page_free_wakeup(void)
1954 {
1955 
1956 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1957 	/*
1958 	 * if pageout daemon needs pages, then tell it that there are
1959 	 * some free.
1960 	 */
1961 	if (vm_pageout_pages_needed &&
1962 	    cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) {
1963 		wakeup(&vm_pageout_pages_needed);
1964 		vm_pageout_pages_needed = 0;
1965 	}
1966 	/*
1967 	 * wakeup processes that are waiting on memory if we hit a
1968 	 * high water mark. And wakeup scheduler process if we have
1969 	 * lots of memory. this process will swapin processes.
1970 	 */
1971 	if (vm_pages_needed && !vm_page_count_min()) {
1972 		vm_pages_needed = 0;
1973 		wakeup(&cnt.v_free_count);
1974 	}
1975 }
1976 
1977 /*
1978  *	vm_page_free_toq:
1979  *
1980  *	Returns the given page to the free list,
1981  *	disassociating it with any VM object.
1982  *
1983  *	Object and page must be locked prior to entry.
1984  *	This routine may not block.
1985  */
1986 
1987 void
1988 vm_page_free_toq(vm_page_t m)
1989 {
1990 
1991 	if ((m->oflags & VPO_UNMANAGED) == 0) {
1992 		vm_page_lock_assert(m, MA_OWNED);
1993 		KASSERT(!pmap_page_is_mapped(m),
1994 		    ("vm_page_free_toq: freeing mapped page %p", m));
1995 	}
1996 	PCPU_INC(cnt.v_tfree);
1997 
1998 	if (VM_PAGE_IS_FREE(m))
1999 		panic("vm_page_free: freeing free page %p", m);
2000 	else if (m->busy != 0)
2001 		panic("vm_page_free: freeing busy page %p", m);
2002 
2003 	/*
2004 	 * unqueue, then remove page.  Note that we cannot destroy
2005 	 * the page here because we do not want to call the pager's
2006 	 * callback routine until after we've put the page on the
2007 	 * appropriate free queue.
2008 	 */
2009 	if ((m->oflags & VPO_UNMANAGED) == 0)
2010 		vm_pageq_remove(m);
2011 	vm_page_remove(m);
2012 
2013 	/*
2014 	 * If fictitious remove object association and
2015 	 * return, otherwise delay object association removal.
2016 	 */
2017 	if ((m->flags & PG_FICTITIOUS) != 0) {
2018 		return;
2019 	}
2020 
2021 	m->valid = 0;
2022 	vm_page_undirty(m);
2023 
2024 	if (m->wire_count != 0)
2025 		panic("vm_page_free: freeing wired page %p", m);
2026 	if (m->hold_count != 0) {
2027 		m->flags &= ~PG_ZERO;
2028 		vm_page_lock_queues();
2029 		vm_page_enqueue(PQ_HOLD, m);
2030 		vm_page_unlock_queues();
2031 	} else {
2032 		/*
2033 		 * Restore the default memory attribute to the page.
2034 		 */
2035 		if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
2036 			pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
2037 
2038 		/*
2039 		 * Insert the page into the physical memory allocator's
2040 		 * cache/free page queues.
2041 		 */
2042 		mtx_lock(&vm_page_queue_free_mtx);
2043 		m->flags |= PG_FREE;
2044 		cnt.v_free_count++;
2045 #if VM_NRESERVLEVEL > 0
2046 		if (!vm_reserv_free_page(m))
2047 #else
2048 		if (TRUE)
2049 #endif
2050 			vm_phys_free_pages(m, 0);
2051 		if ((m->flags & PG_ZERO) != 0)
2052 			++vm_page_zero_count;
2053 		else
2054 			vm_page_zero_idle_wakeup();
2055 		vm_page_free_wakeup();
2056 		mtx_unlock(&vm_page_queue_free_mtx);
2057 	}
2058 }
2059 
2060 /*
2061  *	vm_page_wire:
2062  *
2063  *	Mark this page as wired down by yet
2064  *	another map, removing it from paging queues
2065  *	as necessary.
2066  *
2067  *	If the page is fictitious, then its wire count must remain one.
2068  *
2069  *	The page must be locked.
2070  *	This routine may not block.
2071  */
2072 void
2073 vm_page_wire(vm_page_t m)
2074 {
2075 
2076 	/*
2077 	 * Only bump the wire statistics if the page is not already wired,
2078 	 * and only unqueue the page if it is on some queue (if it is unmanaged
2079 	 * it is already off the queues).
2080 	 */
2081 	vm_page_lock_assert(m, MA_OWNED);
2082 	if ((m->flags & PG_FICTITIOUS) != 0) {
2083 		KASSERT(m->wire_count == 1,
2084 		    ("vm_page_wire: fictitious page %p's wire count isn't one",
2085 		    m));
2086 		return;
2087 	}
2088 	if (m->wire_count == 0) {
2089 		if ((m->oflags & VPO_UNMANAGED) == 0)
2090 			vm_pageq_remove(m);
2091 		atomic_add_int(&cnt.v_wire_count, 1);
2092 	}
2093 	m->wire_count++;
2094 	KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m));
2095 }
2096 
2097 /*
2098  * vm_page_unwire:
2099  *
2100  * Release one wiring of the specified page, potentially enabling it to be
2101  * paged again.  If paging is enabled, then the value of the parameter
2102  * "activate" determines to which queue the page is added.  If "activate" is
2103  * non-zero, then the page is added to the active queue.  Otherwise, it is
2104  * added to the inactive queue.
2105  *
2106  * However, unless the page belongs to an object, it is not enqueued because
2107  * it cannot be paged out.
2108  *
2109  * If a page is fictitious, then its wire count must alway be one.
2110  *
2111  * A managed page must be locked.
2112  */
2113 void
2114 vm_page_unwire(vm_page_t m, int activate)
2115 {
2116 
2117 	if ((m->oflags & VPO_UNMANAGED) == 0)
2118 		vm_page_lock_assert(m, MA_OWNED);
2119 	if ((m->flags & PG_FICTITIOUS) != 0) {
2120 		KASSERT(m->wire_count == 1,
2121 	    ("vm_page_unwire: fictitious page %p's wire count isn't one", m));
2122 		return;
2123 	}
2124 	if (m->wire_count > 0) {
2125 		m->wire_count--;
2126 		if (m->wire_count == 0) {
2127 			atomic_subtract_int(&cnt.v_wire_count, 1);
2128 			if ((m->oflags & VPO_UNMANAGED) != 0 ||
2129 			    m->object == NULL)
2130 				return;
2131 			vm_page_lock_queues();
2132 			if (activate)
2133 				vm_page_enqueue(PQ_ACTIVE, m);
2134 			else {
2135 				m->flags &= ~PG_WINATCFLS;
2136 				vm_page_enqueue(PQ_INACTIVE, m);
2137 			}
2138 			vm_page_unlock_queues();
2139 		}
2140 	} else
2141 		panic("vm_page_unwire: page %p's wire count is zero", m);
2142 }
2143 
2144 /*
2145  * Move the specified page to the inactive queue.
2146  *
2147  * Many pages placed on the inactive queue should actually go
2148  * into the cache, but it is difficult to figure out which.  What
2149  * we do instead, if the inactive target is well met, is to put
2150  * clean pages at the head of the inactive queue instead of the tail.
2151  * This will cause them to be moved to the cache more quickly and
2152  * if not actively re-referenced, reclaimed more quickly.  If we just
2153  * stick these pages at the end of the inactive queue, heavy filesystem
2154  * meta-data accesses can cause an unnecessary paging load on memory bound
2155  * processes.  This optimization causes one-time-use metadata to be
2156  * reused more quickly.
2157  *
2158  * Normally athead is 0 resulting in LRU operation.  athead is set
2159  * to 1 if we want this page to be 'as if it were placed in the cache',
2160  * except without unmapping it from the process address space.
2161  *
2162  * This routine may not block.
2163  */
2164 static inline void
2165 _vm_page_deactivate(vm_page_t m, int athead)
2166 {
2167 	int queue;
2168 
2169 	vm_page_lock_assert(m, MA_OWNED);
2170 
2171 	/*
2172 	 * Ignore if already inactive.
2173 	 */
2174 	if ((queue = m->queue) == PQ_INACTIVE)
2175 		return;
2176 	if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
2177 		vm_page_lock_queues();
2178 		m->flags &= ~PG_WINATCFLS;
2179 		if (queue != PQ_NONE)
2180 			vm_page_queue_remove(queue, m);
2181 		if (athead)
2182 			TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m,
2183 			    pageq);
2184 		else
2185 			TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m,
2186 			    pageq);
2187 		m->queue = PQ_INACTIVE;
2188 		cnt.v_inactive_count++;
2189 		vm_page_unlock_queues();
2190 	}
2191 }
2192 
2193 /*
2194  * Move the specified page to the inactive queue.
2195  *
2196  * The page must be locked.
2197  */
2198 void
2199 vm_page_deactivate(vm_page_t m)
2200 {
2201 
2202 	_vm_page_deactivate(m, 0);
2203 }
2204 
2205 /*
2206  * vm_page_try_to_cache:
2207  *
2208  * Returns 0 on failure, 1 on success
2209  */
2210 int
2211 vm_page_try_to_cache(vm_page_t m)
2212 {
2213 
2214 	vm_page_lock_assert(m, MA_OWNED);
2215 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2216 	if (m->dirty || m->hold_count || m->busy || m->wire_count ||
2217 	    (m->oflags & (VPO_BUSY | VPO_UNMANAGED)) != 0)
2218 		return (0);
2219 	pmap_remove_all(m);
2220 	if (m->dirty)
2221 		return (0);
2222 	vm_page_cache(m);
2223 	return (1);
2224 }
2225 
2226 /*
2227  * vm_page_try_to_free()
2228  *
2229  *	Attempt to free the page.  If we cannot free it, we do nothing.
2230  *	1 is returned on success, 0 on failure.
2231  */
2232 int
2233 vm_page_try_to_free(vm_page_t m)
2234 {
2235 
2236 	vm_page_lock_assert(m, MA_OWNED);
2237 	if (m->object != NULL)
2238 		VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2239 	if (m->dirty || m->hold_count || m->busy || m->wire_count ||
2240 	    (m->oflags & (VPO_BUSY | VPO_UNMANAGED)) != 0)
2241 		return (0);
2242 	pmap_remove_all(m);
2243 	if (m->dirty)
2244 		return (0);
2245 	vm_page_free(m);
2246 	return (1);
2247 }
2248 
2249 /*
2250  * vm_page_cache
2251  *
2252  * Put the specified page onto the page cache queue (if appropriate).
2253  *
2254  * This routine may not block.
2255  */
2256 void
2257 vm_page_cache(vm_page_t m)
2258 {
2259 	vm_object_t object;
2260 	vm_page_t next, prev, root;
2261 
2262 	vm_page_lock_assert(m, MA_OWNED);
2263 	object = m->object;
2264 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2265 	if ((m->oflags & (VPO_UNMANAGED | VPO_BUSY)) || m->busy ||
2266 	    m->hold_count || m->wire_count)
2267 		panic("vm_page_cache: attempting to cache busy page");
2268 	pmap_remove_all(m);
2269 	if (m->dirty != 0)
2270 		panic("vm_page_cache: page %p is dirty", m);
2271 	if (m->valid == 0 || object->type == OBJT_DEFAULT ||
2272 	    (object->type == OBJT_SWAP &&
2273 	    !vm_pager_has_page(object, m->pindex, NULL, NULL))) {
2274 		/*
2275 		 * Hypothesis: A cache-elgible page belonging to a
2276 		 * default object or swap object but without a backing
2277 		 * store must be zero filled.
2278 		 */
2279 		vm_page_free(m);
2280 		return;
2281 	}
2282 	KASSERT((m->flags & PG_CACHED) == 0,
2283 	    ("vm_page_cache: page %p is already cached", m));
2284 	PCPU_INC(cnt.v_tcached);
2285 
2286 	/*
2287 	 * Remove the page from the paging queues.
2288 	 */
2289 	vm_pageq_remove(m);
2290 
2291 	/*
2292 	 * Remove the page from the object's collection of resident
2293 	 * pages.
2294 	 */
2295 	if ((next = TAILQ_NEXT(m, listq)) != NULL && next->left == m) {
2296 		/*
2297 		 * Since the page's successor in the list is also its parent
2298 		 * in the tree, its right subtree must be empty.
2299 		 */
2300 		next->left = m->left;
2301 		KASSERT(m->right == NULL,
2302 		    ("vm_page_cache: page %p has right child", m));
2303 	} else if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL &&
2304 	    prev->right == m) {
2305 		/*
2306 		 * Since the page's predecessor in the list is also its parent
2307 		 * in the tree, its left subtree must be empty.
2308 		 */
2309 		KASSERT(m->left == NULL,
2310 		    ("vm_page_cache: page %p has left child", m));
2311 		prev->right = m->right;
2312 	} else {
2313 		if (m != object->root)
2314 			vm_page_splay(m->pindex, object->root);
2315 		if (m->left == NULL)
2316 			root = m->right;
2317 		else if (m->right == NULL)
2318 			root = m->left;
2319 		else {
2320 			/*
2321 			 * Move the page's successor to the root, because
2322 			 * pages are usually removed in ascending order.
2323 			 */
2324 			if (m->right != next)
2325 				vm_page_splay(m->pindex, m->right);
2326 			next->left = m->left;
2327 			root = next;
2328 		}
2329 		object->root = root;
2330 	}
2331 	TAILQ_REMOVE(&object->memq, m, listq);
2332 	object->resident_page_count--;
2333 
2334 	/*
2335 	 * Restore the default memory attribute to the page.
2336 	 */
2337 	if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
2338 		pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
2339 
2340 	/*
2341 	 * Insert the page into the object's collection of cached pages
2342 	 * and the physical memory allocator's cache/free page queues.
2343 	 */
2344 	m->flags &= ~PG_ZERO;
2345 	mtx_lock(&vm_page_queue_free_mtx);
2346 	m->flags |= PG_CACHED;
2347 	cnt.v_cache_count++;
2348 	root = object->cache;
2349 	if (root == NULL) {
2350 		m->left = NULL;
2351 		m->right = NULL;
2352 	} else {
2353 		root = vm_page_splay(m->pindex, root);
2354 		if (m->pindex < root->pindex) {
2355 			m->left = root->left;
2356 			m->right = root;
2357 			root->left = NULL;
2358 		} else if (__predict_false(m->pindex == root->pindex))
2359 			panic("vm_page_cache: offset already cached");
2360 		else {
2361 			m->right = root->right;
2362 			m->left = root;
2363 			root->right = NULL;
2364 		}
2365 	}
2366 	object->cache = m;
2367 #if VM_NRESERVLEVEL > 0
2368 	if (!vm_reserv_free_page(m)) {
2369 #else
2370 	if (TRUE) {
2371 #endif
2372 		vm_phys_set_pool(VM_FREEPOOL_CACHE, m, 0);
2373 		vm_phys_free_pages(m, 0);
2374 	}
2375 	vm_page_free_wakeup();
2376 	mtx_unlock(&vm_page_queue_free_mtx);
2377 
2378 	/*
2379 	 * Increment the vnode's hold count if this is the object's only
2380 	 * cached page.  Decrement the vnode's hold count if this was
2381 	 * the object's only resident page.
2382 	 */
2383 	if (object->type == OBJT_VNODE) {
2384 		if (root == NULL && object->resident_page_count != 0)
2385 			vhold(object->handle);
2386 		else if (root != NULL && object->resident_page_count == 0)
2387 			vdrop(object->handle);
2388 	}
2389 }
2390 
2391 /*
2392  * vm_page_dontneed
2393  *
2394  *	Cache, deactivate, or do nothing as appropriate.  This routine
2395  *	is typically used by madvise() MADV_DONTNEED.
2396  *
2397  *	Generally speaking we want to move the page into the cache so
2398  *	it gets reused quickly.  However, this can result in a silly syndrome
2399  *	due to the page recycling too quickly.  Small objects will not be
2400  *	fully cached.  On the otherhand, if we move the page to the inactive
2401  *	queue we wind up with a problem whereby very large objects
2402  *	unnecessarily blow away our inactive and cache queues.
2403  *
2404  *	The solution is to move the pages based on a fixed weighting.  We
2405  *	either leave them alone, deactivate them, or move them to the cache,
2406  *	where moving them to the cache has the highest weighting.
2407  *	By forcing some pages into other queues we eventually force the
2408  *	system to balance the queues, potentially recovering other unrelated
2409  *	space from active.  The idea is to not force this to happen too
2410  *	often.
2411  */
2412 void
2413 vm_page_dontneed(vm_page_t m)
2414 {
2415 	int dnw;
2416 	int head;
2417 
2418 	vm_page_lock_assert(m, MA_OWNED);
2419 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2420 	dnw = PCPU_GET(dnweight);
2421 	PCPU_INC(dnweight);
2422 
2423 	/*
2424 	 * Occasionally leave the page alone.
2425 	 */
2426 	if ((dnw & 0x01F0) == 0 || m->queue == PQ_INACTIVE) {
2427 		if (m->act_count >= ACT_INIT)
2428 			--m->act_count;
2429 		return;
2430 	}
2431 
2432 	/*
2433 	 * Clear any references to the page.  Otherwise, the page daemon will
2434 	 * immediately reactivate the page.
2435 	 *
2436 	 * Perform the pmap_clear_reference() first.  Otherwise, a concurrent
2437 	 * pmap operation, such as pmap_remove(), could clear a reference in
2438 	 * the pmap and set PGA_REFERENCED on the page before the
2439 	 * pmap_clear_reference() had completed.  Consequently, the page would
2440 	 * appear referenced based upon an old reference that occurred before
2441 	 * this function ran.
2442 	 */
2443 	pmap_clear_reference(m);
2444 	vm_page_aflag_clear(m, PGA_REFERENCED);
2445 
2446 	if (m->dirty == 0 && pmap_is_modified(m))
2447 		vm_page_dirty(m);
2448 
2449 	if (m->dirty || (dnw & 0x0070) == 0) {
2450 		/*
2451 		 * Deactivate the page 3 times out of 32.
2452 		 */
2453 		head = 0;
2454 	} else {
2455 		/*
2456 		 * Cache the page 28 times out of every 32.  Note that
2457 		 * the page is deactivated instead of cached, but placed
2458 		 * at the head of the queue instead of the tail.
2459 		 */
2460 		head = 1;
2461 	}
2462 	_vm_page_deactivate(m, head);
2463 }
2464 
2465 /*
2466  * Grab a page, waiting until we are waken up due to the page
2467  * changing state.  We keep on waiting, if the page continues
2468  * to be in the object.  If the page doesn't exist, first allocate it
2469  * and then conditionally zero it.
2470  *
2471  * The caller must always specify the VM_ALLOC_RETRY flag.  This is intended
2472  * to facilitate its eventual removal.
2473  *
2474  * This routine may block.
2475  */
2476 vm_page_t
2477 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
2478 {
2479 	vm_page_t m;
2480 
2481 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2482 	KASSERT((allocflags & VM_ALLOC_RETRY) != 0,
2483 	    ("vm_page_grab: VM_ALLOC_RETRY is required"));
2484 retrylookup:
2485 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
2486 		if ((m->oflags & VPO_BUSY) != 0 ||
2487 		    ((allocflags & VM_ALLOC_IGN_SBUSY) == 0 && m->busy != 0)) {
2488 			/*
2489 			 * Reference the page before unlocking and
2490 			 * sleeping so that the page daemon is less
2491 			 * likely to reclaim it.
2492 			 */
2493 			vm_page_aflag_set(m, PGA_REFERENCED);
2494 			vm_page_sleep(m, "pgrbwt");
2495 			goto retrylookup;
2496 		} else {
2497 			if ((allocflags & VM_ALLOC_WIRED) != 0) {
2498 				vm_page_lock(m);
2499 				vm_page_wire(m);
2500 				vm_page_unlock(m);
2501 			}
2502 			if ((allocflags & VM_ALLOC_NOBUSY) == 0)
2503 				vm_page_busy(m);
2504 			return (m);
2505 		}
2506 	}
2507 	m = vm_page_alloc(object, pindex, allocflags & ~(VM_ALLOC_RETRY |
2508 	    VM_ALLOC_IGN_SBUSY));
2509 	if (m == NULL) {
2510 		VM_OBJECT_UNLOCK(object);
2511 		VM_WAIT;
2512 		VM_OBJECT_LOCK(object);
2513 		goto retrylookup;
2514 	} else if (m->valid != 0)
2515 		return (m);
2516 	if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
2517 		pmap_zero_page(m);
2518 	return (m);
2519 }
2520 
2521 /*
2522  * Mapping function for valid bits or for dirty bits in
2523  * a page.  May not block.
2524  *
2525  * Inputs are required to range within a page.
2526  */
2527 vm_page_bits_t
2528 vm_page_bits(int base, int size)
2529 {
2530 	int first_bit;
2531 	int last_bit;
2532 
2533 	KASSERT(
2534 	    base + size <= PAGE_SIZE,
2535 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
2536 	);
2537 
2538 	if (size == 0)		/* handle degenerate case */
2539 		return (0);
2540 
2541 	first_bit = base >> DEV_BSHIFT;
2542 	last_bit = (base + size - 1) >> DEV_BSHIFT;
2543 
2544 	return (((vm_page_bits_t)2 << last_bit) -
2545 	    ((vm_page_bits_t)1 << first_bit));
2546 }
2547 
2548 /*
2549  *	vm_page_set_valid_range:
2550  *
2551  *	Sets portions of a page valid.  The arguments are expected
2552  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
2553  *	of any partial chunks touched by the range.  The invalid portion of
2554  *	such chunks will be zeroed.
2555  *
2556  *	(base + size) must be less then or equal to PAGE_SIZE.
2557  */
2558 void
2559 vm_page_set_valid_range(vm_page_t m, int base, int size)
2560 {
2561 	int endoff, frag;
2562 
2563 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2564 	if (size == 0)	/* handle degenerate case */
2565 		return;
2566 
2567 	/*
2568 	 * If the base is not DEV_BSIZE aligned and the valid
2569 	 * bit is clear, we have to zero out a portion of the
2570 	 * first block.
2571 	 */
2572 	if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
2573 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
2574 		pmap_zero_page_area(m, frag, base - frag);
2575 
2576 	/*
2577 	 * If the ending offset is not DEV_BSIZE aligned and the
2578 	 * valid bit is clear, we have to zero out a portion of
2579 	 * the last block.
2580 	 */
2581 	endoff = base + size;
2582 	if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
2583 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
2584 		pmap_zero_page_area(m, endoff,
2585 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
2586 
2587 	/*
2588 	 * Assert that no previously invalid block that is now being validated
2589 	 * is already dirty.
2590 	 */
2591 	KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
2592 	    ("vm_page_set_valid_range: page %p is dirty", m));
2593 
2594 	/*
2595 	 * Set valid bits inclusive of any overlap.
2596 	 */
2597 	m->valid |= vm_page_bits(base, size);
2598 }
2599 
2600 /*
2601  * Clear the given bits from the specified page's dirty field.
2602  */
2603 static __inline void
2604 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
2605 {
2606 	uintptr_t addr;
2607 #if PAGE_SIZE < 16384
2608 	int shift;
2609 #endif
2610 
2611 	/*
2612 	 * If the object is locked and the page is neither VPO_BUSY nor
2613 	 * PGA_WRITEABLE, then the page's dirty field cannot possibly be
2614 	 * set by a concurrent pmap operation.
2615 	 */
2616 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2617 	if ((m->oflags & VPO_BUSY) == 0 && (m->aflags & PGA_WRITEABLE) == 0)
2618 		m->dirty &= ~pagebits;
2619 	else {
2620 		/*
2621 		 * The pmap layer can call vm_page_dirty() without
2622 		 * holding a distinguished lock.  The combination of
2623 		 * the object's lock and an atomic operation suffice
2624 		 * to guarantee consistency of the page dirty field.
2625 		 *
2626 		 * For PAGE_SIZE == 32768 case, compiler already
2627 		 * properly aligns the dirty field, so no forcible
2628 		 * alignment is needed. Only require existence of
2629 		 * atomic_clear_64 when page size is 32768.
2630 		 */
2631 		addr = (uintptr_t)&m->dirty;
2632 #if PAGE_SIZE == 32768
2633 		atomic_clear_64((uint64_t *)addr, pagebits);
2634 #elif PAGE_SIZE == 16384
2635 		atomic_clear_32((uint32_t *)addr, pagebits);
2636 #else		/* PAGE_SIZE <= 8192 */
2637 		/*
2638 		 * Use a trick to perform a 32-bit atomic on the
2639 		 * containing aligned word, to not depend on the existence
2640 		 * of atomic_clear_{8, 16}.
2641 		 */
2642 		shift = addr & (sizeof(uint32_t) - 1);
2643 #if BYTE_ORDER == BIG_ENDIAN
2644 		shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY;
2645 #else
2646 		shift *= NBBY;
2647 #endif
2648 		addr &= ~(sizeof(uint32_t) - 1);
2649 		atomic_clear_32((uint32_t *)addr, pagebits << shift);
2650 #endif		/* PAGE_SIZE */
2651 	}
2652 }
2653 
2654 /*
2655  *	vm_page_set_validclean:
2656  *
2657  *	Sets portions of a page valid and clean.  The arguments are expected
2658  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
2659  *	of any partial chunks touched by the range.  The invalid portion of
2660  *	such chunks will be zero'd.
2661  *
2662  *	This routine may not block.
2663  *
2664  *	(base + size) must be less then or equal to PAGE_SIZE.
2665  */
2666 void
2667 vm_page_set_validclean(vm_page_t m, int base, int size)
2668 {
2669 	vm_page_bits_t oldvalid, pagebits;
2670 	int endoff, frag;
2671 
2672 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2673 	if (size == 0)	/* handle degenerate case */
2674 		return;
2675 
2676 	/*
2677 	 * If the base is not DEV_BSIZE aligned and the valid
2678 	 * bit is clear, we have to zero out a portion of the
2679 	 * first block.
2680 	 */
2681 	if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
2682 	    (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
2683 		pmap_zero_page_area(m, frag, base - frag);
2684 
2685 	/*
2686 	 * If the ending offset is not DEV_BSIZE aligned and the
2687 	 * valid bit is clear, we have to zero out a portion of
2688 	 * the last block.
2689 	 */
2690 	endoff = base + size;
2691 	if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
2692 	    (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
2693 		pmap_zero_page_area(m, endoff,
2694 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
2695 
2696 	/*
2697 	 * Set valid, clear dirty bits.  If validating the entire
2698 	 * page we can safely clear the pmap modify bit.  We also
2699 	 * use this opportunity to clear the VPO_NOSYNC flag.  If a process
2700 	 * takes a write fault on a MAP_NOSYNC memory area the flag will
2701 	 * be set again.
2702 	 *
2703 	 * We set valid bits inclusive of any overlap, but we can only
2704 	 * clear dirty bits for DEV_BSIZE chunks that are fully within
2705 	 * the range.
2706 	 */
2707 	oldvalid = m->valid;
2708 	pagebits = vm_page_bits(base, size);
2709 	m->valid |= pagebits;
2710 #if 0	/* NOT YET */
2711 	if ((frag = base & (DEV_BSIZE - 1)) != 0) {
2712 		frag = DEV_BSIZE - frag;
2713 		base += frag;
2714 		size -= frag;
2715 		if (size < 0)
2716 			size = 0;
2717 	}
2718 	pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
2719 #endif
2720 	if (base == 0 && size == PAGE_SIZE) {
2721 		/*
2722 		 * The page can only be modified within the pmap if it is
2723 		 * mapped, and it can only be mapped if it was previously
2724 		 * fully valid.
2725 		 */
2726 		if (oldvalid == VM_PAGE_BITS_ALL)
2727 			/*
2728 			 * Perform the pmap_clear_modify() first.  Otherwise,
2729 			 * a concurrent pmap operation, such as
2730 			 * pmap_protect(), could clear a modification in the
2731 			 * pmap and set the dirty field on the page before
2732 			 * pmap_clear_modify() had begun and after the dirty
2733 			 * field was cleared here.
2734 			 */
2735 			pmap_clear_modify(m);
2736 		m->dirty = 0;
2737 		m->oflags &= ~VPO_NOSYNC;
2738 	} else if (oldvalid != VM_PAGE_BITS_ALL)
2739 		m->dirty &= ~pagebits;
2740 	else
2741 		vm_page_clear_dirty_mask(m, pagebits);
2742 }
2743 
2744 void
2745 vm_page_clear_dirty(vm_page_t m, int base, int size)
2746 {
2747 
2748 	vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
2749 }
2750 
2751 /*
2752  *	vm_page_set_invalid:
2753  *
2754  *	Invalidates DEV_BSIZE'd chunks within a page.  Both the
2755  *	valid and dirty bits for the effected areas are cleared.
2756  *
2757  *	May not block.
2758  */
2759 void
2760 vm_page_set_invalid(vm_page_t m, int base, int size)
2761 {
2762 	vm_page_bits_t bits;
2763 
2764 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2765 	KASSERT((m->oflags & VPO_BUSY) == 0,
2766 	    ("vm_page_set_invalid: page %p is busy", m));
2767 	bits = vm_page_bits(base, size);
2768 	if (m->valid == VM_PAGE_BITS_ALL && bits != 0)
2769 		pmap_remove_all(m);
2770 	KASSERT(!pmap_page_is_mapped(m),
2771 	    ("vm_page_set_invalid: page %p is mapped", m));
2772 	m->valid &= ~bits;
2773 	m->dirty &= ~bits;
2774 }
2775 
2776 /*
2777  * vm_page_zero_invalid()
2778  *
2779  *	The kernel assumes that the invalid portions of a page contain
2780  *	garbage, but such pages can be mapped into memory by user code.
2781  *	When this occurs, we must zero out the non-valid portions of the
2782  *	page so user code sees what it expects.
2783  *
2784  *	Pages are most often semi-valid when the end of a file is mapped
2785  *	into memory and the file's size is not page aligned.
2786  */
2787 void
2788 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
2789 {
2790 	int b;
2791 	int i;
2792 
2793 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2794 	/*
2795 	 * Scan the valid bits looking for invalid sections that
2796 	 * must be zerod.  Invalid sub-DEV_BSIZE'd areas ( where the
2797 	 * valid bit may be set ) have already been zerod by
2798 	 * vm_page_set_validclean().
2799 	 */
2800 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
2801 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
2802 		    (m->valid & ((vm_page_bits_t)1 << i))) {
2803 			if (i > b) {
2804 				pmap_zero_page_area(m,
2805 				    b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
2806 			}
2807 			b = i + 1;
2808 		}
2809 	}
2810 
2811 	/*
2812 	 * setvalid is TRUE when we can safely set the zero'd areas
2813 	 * as being valid.  We can do this if there are no cache consistancy
2814 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
2815 	 */
2816 	if (setvalid)
2817 		m->valid = VM_PAGE_BITS_ALL;
2818 }
2819 
2820 /*
2821  *	vm_page_is_valid:
2822  *
2823  *	Is (partial) page valid?  Note that the case where size == 0
2824  *	will return FALSE in the degenerate case where the page is
2825  *	entirely invalid, and TRUE otherwise.
2826  *
2827  *	May not block.
2828  */
2829 int
2830 vm_page_is_valid(vm_page_t m, int base, int size)
2831 {
2832 	vm_page_bits_t bits;
2833 
2834 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2835 	bits = vm_page_bits(base, size);
2836 	if (m->valid && ((m->valid & bits) == bits))
2837 		return 1;
2838 	else
2839 		return 0;
2840 }
2841 
2842 /*
2843  * update dirty bits from pmap/mmu.  May not block.
2844  */
2845 void
2846 vm_page_test_dirty(vm_page_t m)
2847 {
2848 
2849 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2850 	if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
2851 		vm_page_dirty(m);
2852 }
2853 
2854 void
2855 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
2856 {
2857 
2858 	mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
2859 }
2860 
2861 void
2862 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
2863 {
2864 
2865 	mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
2866 }
2867 
2868 int
2869 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
2870 {
2871 
2872 	return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
2873 }
2874 
2875 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
2876 void
2877 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
2878 {
2879 
2880 	mtx_assert_(vm_page_lockptr(m), a, file, line);
2881 }
2882 #endif
2883 
2884 int so_zerocp_fullpage = 0;
2885 
2886 /*
2887  *	Replace the given page with a copy.  The copied page assumes
2888  *	the portion of the given page's "wire_count" that is not the
2889  *	responsibility of this copy-on-write mechanism.
2890  *
2891  *	The object containing the given page must have a non-zero
2892  *	paging-in-progress count and be locked.
2893  */
2894 void
2895 vm_page_cowfault(vm_page_t m)
2896 {
2897 	vm_page_t mnew;
2898 	vm_object_t object;
2899 	vm_pindex_t pindex;
2900 
2901 	mtx_assert(&vm_page_queue_mtx, MA_NOTOWNED);
2902 	vm_page_lock_assert(m, MA_OWNED);
2903 	object = m->object;
2904 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2905 	KASSERT(object->paging_in_progress != 0,
2906 	    ("vm_page_cowfault: object %p's paging-in-progress count is zero.",
2907 	    object));
2908 	pindex = m->pindex;
2909 
2910  retry_alloc:
2911 	pmap_remove_all(m);
2912 	vm_page_remove(m);
2913 	mnew = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL | VM_ALLOC_NOBUSY);
2914 	if (mnew == NULL) {
2915 		vm_page_insert(m, object, pindex);
2916 		vm_page_unlock(m);
2917 		VM_OBJECT_UNLOCK(object);
2918 		VM_WAIT;
2919 		VM_OBJECT_LOCK(object);
2920 		if (m == vm_page_lookup(object, pindex)) {
2921 			vm_page_lock(m);
2922 			goto retry_alloc;
2923 		} else {
2924 			/*
2925 			 * Page disappeared during the wait.
2926 			 */
2927 			return;
2928 		}
2929 	}
2930 
2931 	if (m->cow == 0) {
2932 		/*
2933 		 * check to see if we raced with an xmit complete when
2934 		 * waiting to allocate a page.  If so, put things back
2935 		 * the way they were
2936 		 */
2937 		vm_page_unlock(m);
2938 		vm_page_lock(mnew);
2939 		vm_page_free(mnew);
2940 		vm_page_unlock(mnew);
2941 		vm_page_insert(m, object, pindex);
2942 	} else { /* clear COW & copy page */
2943 		if (!so_zerocp_fullpage)
2944 			pmap_copy_page(m, mnew);
2945 		mnew->valid = VM_PAGE_BITS_ALL;
2946 		vm_page_dirty(mnew);
2947 		mnew->wire_count = m->wire_count - m->cow;
2948 		m->wire_count = m->cow;
2949 		vm_page_unlock(m);
2950 	}
2951 }
2952 
2953 void
2954 vm_page_cowclear(vm_page_t m)
2955 {
2956 
2957 	vm_page_lock_assert(m, MA_OWNED);
2958 	if (m->cow) {
2959 		m->cow--;
2960 		/*
2961 		 * let vm_fault add back write permission  lazily
2962 		 */
2963 	}
2964 	/*
2965 	 *  sf_buf_free() will free the page, so we needn't do it here
2966 	 */
2967 }
2968 
2969 int
2970 vm_page_cowsetup(vm_page_t m)
2971 {
2972 
2973 	vm_page_lock_assert(m, MA_OWNED);
2974 	if ((m->flags & PG_FICTITIOUS) != 0 ||
2975 	    (m->oflags & VPO_UNMANAGED) != 0 ||
2976 	    m->cow == USHRT_MAX - 1 || !VM_OBJECT_TRYLOCK(m->object))
2977 		return (EBUSY);
2978 	m->cow++;
2979 	pmap_remove_write(m);
2980 	VM_OBJECT_UNLOCK(m->object);
2981 	return (0);
2982 }
2983 
2984 #ifdef INVARIANTS
2985 void
2986 vm_page_object_lock_assert(vm_page_t m)
2987 {
2988 
2989 	/*
2990 	 * Certain of the page's fields may only be modified by the
2991 	 * holder of the containing object's lock or the setter of the
2992 	 * page's VPO_BUSY flag.  Unfortunately, the setter of the
2993 	 * VPO_BUSY flag is not recorded, and thus cannot be checked
2994 	 * here.
2995 	 */
2996 	if (m->object != NULL && (m->oflags & VPO_BUSY) == 0)
2997 		VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2998 }
2999 #endif
3000 
3001 #include "opt_ddb.h"
3002 #ifdef DDB
3003 #include <sys/kernel.h>
3004 
3005 #include <ddb/ddb.h>
3006 
3007 DB_SHOW_COMMAND(page, vm_page_print_page_info)
3008 {
3009 	db_printf("cnt.v_free_count: %d\n", cnt.v_free_count);
3010 	db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count);
3011 	db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count);
3012 	db_printf("cnt.v_active_count: %d\n", cnt.v_active_count);
3013 	db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count);
3014 	db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved);
3015 	db_printf("cnt.v_free_min: %d\n", cnt.v_free_min);
3016 	db_printf("cnt.v_free_target: %d\n", cnt.v_free_target);
3017 	db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min);
3018 	db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target);
3019 }
3020 
3021 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
3022 {
3023 
3024 	db_printf("PQ_FREE:");
3025 	db_printf(" %d", cnt.v_free_count);
3026 	db_printf("\n");
3027 
3028 	db_printf("PQ_CACHE:");
3029 	db_printf(" %d", cnt.v_cache_count);
3030 	db_printf("\n");
3031 
3032 	db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n",
3033 		*vm_page_queues[PQ_ACTIVE].cnt,
3034 		*vm_page_queues[PQ_INACTIVE].cnt);
3035 }
3036 #endif /* DDB */
3037