xref: /freebsd/sys/vm/vm_page.c (revision d1d015864103b253b3fcb2f72a0da5b0cfeb31b6)
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 page queue lock is required when adding or removing a page from a
67  *	  page queue regardless of other locks or the busy state of a page.
68  *
69  *		* In general, no thread besides the page daemon can acquire or
70  *		  hold more than one page queue lock at a time.
71  *
72  *		* The page daemon can acquire and hold any pair of page queue
73  *		  locks in any order.
74  *
75  *	- The object lock is required when inserting or removing
76  *	  pages from an object (vm_page_insert() or vm_page_remove()).
77  *
78  */
79 
80 /*
81  *	Resident memory management module.
82  */
83 
84 #include <sys/cdefs.h>
85 __FBSDID("$FreeBSD$");
86 
87 #include "opt_vm.h"
88 
89 #include <sys/param.h>
90 #include <sys/systm.h>
91 #include <sys/lock.h>
92 #include <sys/kernel.h>
93 #include <sys/limits.h>
94 #include <sys/malloc.h>
95 #include <sys/mman.h>
96 #include <sys/msgbuf.h>
97 #include <sys/mutex.h>
98 #include <sys/proc.h>
99 #include <sys/rwlock.h>
100 #include <sys/sysctl.h>
101 #include <sys/vmmeter.h>
102 #include <sys/vnode.h>
103 
104 #include <vm/vm.h>
105 #include <vm/pmap.h>
106 #include <vm/vm_param.h>
107 #include <vm/vm_kern.h>
108 #include <vm/vm_object.h>
109 #include <vm/vm_page.h>
110 #include <vm/vm_pageout.h>
111 #include <vm/vm_pager.h>
112 #include <vm/vm_phys.h>
113 #include <vm/vm_radix.h>
114 #include <vm/vm_reserv.h>
115 #include <vm/vm_extern.h>
116 #include <vm/uma.h>
117 #include <vm/uma_int.h>
118 
119 #include <machine/md_var.h>
120 
121 /*
122  *	Associated with page of user-allocatable memory is a
123  *	page structure.
124  */
125 
126 struct vm_domain vm_dom[MAXMEMDOM];
127 struct mtx_padalign vm_page_queue_free_mtx;
128 
129 struct mtx_padalign pa_lock[PA_LOCK_COUNT];
130 
131 vm_page_t vm_page_array;
132 long vm_page_array_size;
133 long first_page;
134 int vm_page_zero_count;
135 
136 static int boot_pages = UMA_BOOT_PAGES;
137 TUNABLE_INT("vm.boot_pages", &boot_pages);
138 SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RD, &boot_pages, 0,
139 	"number of pages allocated for bootstrapping the VM system");
140 
141 static int pa_tryrelock_restart;
142 SYSCTL_INT(_vm, OID_AUTO, tryrelock_restart, CTLFLAG_RD,
143     &pa_tryrelock_restart, 0, "Number of tryrelock restarts");
144 
145 static uma_zone_t fakepg_zone;
146 
147 static struct vnode *vm_page_alloc_init(vm_page_t m);
148 static void vm_page_cache_turn_free(vm_page_t m);
149 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits);
150 static void vm_page_enqueue(int queue, vm_page_t m);
151 static void vm_page_init_fakepg(void *dummy);
152 static int vm_page_insert_after(vm_page_t m, vm_object_t object,
153     vm_pindex_t pindex, vm_page_t mpred);
154 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object,
155     vm_page_t mpred);
156 
157 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init_fakepg, NULL);
158 
159 static void
160 vm_page_init_fakepg(void *dummy)
161 {
162 
163 	fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
164 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
165 }
166 
167 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
168 #if PAGE_SIZE == 32768
169 #ifdef CTASSERT
170 CTASSERT(sizeof(u_long) >= 8);
171 #endif
172 #endif
173 
174 /*
175  * Try to acquire a physical address lock while a pmap is locked.  If we
176  * fail to trylock we unlock and lock the pmap directly and cache the
177  * locked pa in *locked.  The caller should then restart their loop in case
178  * the virtual to physical mapping has changed.
179  */
180 int
181 vm_page_pa_tryrelock(pmap_t pmap, vm_paddr_t pa, vm_paddr_t *locked)
182 {
183 	vm_paddr_t lockpa;
184 
185 	lockpa = *locked;
186 	*locked = pa;
187 	if (lockpa) {
188 		PA_LOCK_ASSERT(lockpa, MA_OWNED);
189 		if (PA_LOCKPTR(pa) == PA_LOCKPTR(lockpa))
190 			return (0);
191 		PA_UNLOCK(lockpa);
192 	}
193 	if (PA_TRYLOCK(pa))
194 		return (0);
195 	PMAP_UNLOCK(pmap);
196 	atomic_add_int(&pa_tryrelock_restart, 1);
197 	PA_LOCK(pa);
198 	PMAP_LOCK(pmap);
199 	return (EAGAIN);
200 }
201 
202 /*
203  *	vm_set_page_size:
204  *
205  *	Sets the page size, perhaps based upon the memory
206  *	size.  Must be called before any use of page-size
207  *	dependent functions.
208  */
209 void
210 vm_set_page_size(void)
211 {
212 	if (cnt.v_page_size == 0)
213 		cnt.v_page_size = PAGE_SIZE;
214 	if (((cnt.v_page_size - 1) & cnt.v_page_size) != 0)
215 		panic("vm_set_page_size: page size not a power of two");
216 }
217 
218 /*
219  *	vm_page_blacklist_lookup:
220  *
221  *	See if a physical address in this page has been listed
222  *	in the blacklist tunable.  Entries in the tunable are
223  *	separated by spaces or commas.  If an invalid integer is
224  *	encountered then the rest of the string is skipped.
225  */
226 static int
227 vm_page_blacklist_lookup(char *list, vm_paddr_t pa)
228 {
229 	vm_paddr_t bad;
230 	char *cp, *pos;
231 
232 	for (pos = list; *pos != '\0'; pos = cp) {
233 		bad = strtoq(pos, &cp, 0);
234 		if (*cp != '\0') {
235 			if (*cp == ' ' || *cp == ',') {
236 				cp++;
237 				if (cp == pos)
238 					continue;
239 			} else
240 				break;
241 		}
242 		if (pa == trunc_page(bad))
243 			return (1);
244 	}
245 	return (0);
246 }
247 
248 static void
249 vm_page_domain_init(struct vm_domain *vmd)
250 {
251 	struct vm_pagequeue *pq;
252 	int i;
253 
254 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) =
255 	    "vm inactive pagequeue";
256 	*__DECONST(int **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_vcnt) =
257 	    &cnt.v_inactive_count;
258 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) =
259 	    "vm active pagequeue";
260 	*__DECONST(int **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_vcnt) =
261 	    &cnt.v_active_count;
262 	vmd->vmd_page_count = 0;
263 	vmd->vmd_free_count = 0;
264 	vmd->vmd_segs = 0;
265 	vmd->vmd_oom = FALSE;
266 	vmd->vmd_pass = 0;
267 	for (i = 0; i < PQ_COUNT; i++) {
268 		pq = &vmd->vmd_pagequeues[i];
269 		TAILQ_INIT(&pq->pq_pl);
270 		mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue",
271 		    MTX_DEF | MTX_DUPOK);
272 	}
273 }
274 
275 /*
276  *	vm_page_startup:
277  *
278  *	Initializes the resident memory module.
279  *
280  *	Allocates memory for the page cells, and
281  *	for the object/offset-to-page hash table headers.
282  *	Each page cell is initialized and placed on the free list.
283  */
284 vm_offset_t
285 vm_page_startup(vm_offset_t vaddr)
286 {
287 	vm_offset_t mapped;
288 	vm_paddr_t page_range;
289 	vm_paddr_t new_end;
290 	int i;
291 	vm_paddr_t pa;
292 	vm_paddr_t last_pa;
293 	char *list;
294 
295 	/* the biggest memory array is the second group of pages */
296 	vm_paddr_t end;
297 	vm_paddr_t biggestsize;
298 	vm_paddr_t low_water, high_water;
299 	int biggestone;
300 
301 	biggestsize = 0;
302 	biggestone = 0;
303 	vaddr = round_page(vaddr);
304 
305 	for (i = 0; phys_avail[i + 1]; i += 2) {
306 		phys_avail[i] = round_page(phys_avail[i]);
307 		phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
308 	}
309 
310 	low_water = phys_avail[0];
311 	high_water = phys_avail[1];
312 
313 	for (i = 0; phys_avail[i + 1]; i += 2) {
314 		vm_paddr_t size = phys_avail[i + 1] - phys_avail[i];
315 
316 		if (size > biggestsize) {
317 			biggestone = i;
318 			biggestsize = size;
319 		}
320 		if (phys_avail[i] < low_water)
321 			low_water = phys_avail[i];
322 		if (phys_avail[i + 1] > high_water)
323 			high_water = phys_avail[i + 1];
324 	}
325 
326 #ifdef XEN
327 	low_water = 0;
328 #endif
329 
330 	end = phys_avail[biggestone+1];
331 
332 	/*
333 	 * Initialize the page and queue locks.
334 	 */
335 	mtx_init(&vm_page_queue_free_mtx, "vm page free queue", NULL, MTX_DEF);
336 	for (i = 0; i < PA_LOCK_COUNT; i++)
337 		mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
338 	for (i = 0; i < vm_ndomains; i++)
339 		vm_page_domain_init(&vm_dom[i]);
340 
341 	/*
342 	 * Allocate memory for use when boot strapping the kernel memory
343 	 * allocator.
344 	 */
345 	new_end = end - (boot_pages * UMA_SLAB_SIZE);
346 	new_end = trunc_page(new_end);
347 	mapped = pmap_map(&vaddr, new_end, end,
348 	    VM_PROT_READ | VM_PROT_WRITE);
349 	bzero((void *)mapped, end - new_end);
350 	uma_startup((void *)mapped, boot_pages);
351 
352 #if defined(__amd64__) || defined(__i386__) || defined(__arm__) || \
353     defined(__mips__)
354 	/*
355 	 * Allocate a bitmap to indicate that a random physical page
356 	 * needs to be included in a minidump.
357 	 *
358 	 * The amd64 port needs this to indicate which direct map pages
359 	 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
360 	 *
361 	 * However, i386 still needs this workspace internally within the
362 	 * minidump code.  In theory, they are not needed on i386, but are
363 	 * included should the sf_buf code decide to use them.
364 	 */
365 	last_pa = 0;
366 	for (i = 0; dump_avail[i + 1] != 0; i += 2)
367 		if (dump_avail[i + 1] > last_pa)
368 			last_pa = dump_avail[i + 1];
369 	page_range = last_pa / PAGE_SIZE;
370 	vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
371 	new_end -= vm_page_dump_size;
372 	vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
373 	    new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
374 	bzero((void *)vm_page_dump, vm_page_dump_size);
375 #endif
376 #ifdef __amd64__
377 	/*
378 	 * Request that the physical pages underlying the message buffer be
379 	 * included in a crash dump.  Since the message buffer is accessed
380 	 * through the direct map, they are not automatically included.
381 	 */
382 	pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
383 	last_pa = pa + round_page(msgbufsize);
384 	while (pa < last_pa) {
385 		dump_add_page(pa);
386 		pa += PAGE_SIZE;
387 	}
388 #endif
389 	/*
390 	 * Compute the number of pages of memory that will be available for
391 	 * use (taking into account the overhead of a page structure per
392 	 * page).
393 	 */
394 	first_page = low_water / PAGE_SIZE;
395 #ifdef VM_PHYSSEG_SPARSE
396 	page_range = 0;
397 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
398 		page_range += atop(phys_avail[i + 1] - phys_avail[i]);
399 #elif defined(VM_PHYSSEG_DENSE)
400 	page_range = high_water / PAGE_SIZE - first_page;
401 #else
402 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
403 #endif
404 	end = new_end;
405 
406 	/*
407 	 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
408 	 */
409 	vaddr += PAGE_SIZE;
410 
411 	/*
412 	 * Initialize the mem entry structures now, and put them in the free
413 	 * queue.
414 	 */
415 	new_end = trunc_page(end - page_range * sizeof(struct vm_page));
416 	mapped = pmap_map(&vaddr, new_end, end,
417 	    VM_PROT_READ | VM_PROT_WRITE);
418 	vm_page_array = (vm_page_t) mapped;
419 #if VM_NRESERVLEVEL > 0
420 	/*
421 	 * Allocate memory for the reservation management system's data
422 	 * structures.
423 	 */
424 	new_end = vm_reserv_startup(&vaddr, new_end, high_water);
425 #endif
426 #if defined(__amd64__) || defined(__mips__)
427 	/*
428 	 * pmap_map on amd64 and mips can come out of the direct-map, not kvm
429 	 * like i386, so the pages must be tracked for a crashdump to include
430 	 * this data.  This includes the vm_page_array and the early UMA
431 	 * bootstrap pages.
432 	 */
433 	for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE)
434 		dump_add_page(pa);
435 #endif
436 	phys_avail[biggestone + 1] = new_end;
437 
438 	/*
439 	 * Clear all of the page structures
440 	 */
441 	bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
442 	for (i = 0; i < page_range; i++)
443 		vm_page_array[i].order = VM_NFREEORDER;
444 	vm_page_array_size = page_range;
445 
446 	/*
447 	 * Initialize the physical memory allocator.
448 	 */
449 	vm_phys_init();
450 
451 	/*
452 	 * Add every available physical page that is not blacklisted to
453 	 * the free lists.
454 	 */
455 	cnt.v_page_count = 0;
456 	cnt.v_free_count = 0;
457 	list = getenv("vm.blacklist");
458 	for (i = 0; phys_avail[i + 1] != 0; i += 2) {
459 		pa = phys_avail[i];
460 		last_pa = phys_avail[i + 1];
461 		while (pa < last_pa) {
462 			if (list != NULL &&
463 			    vm_page_blacklist_lookup(list, pa))
464 				printf("Skipping page with pa 0x%jx\n",
465 				    (uintmax_t)pa);
466 			else
467 				vm_phys_add_page(pa);
468 			pa += PAGE_SIZE;
469 		}
470 	}
471 	freeenv(list);
472 #if VM_NRESERVLEVEL > 0
473 	/*
474 	 * Initialize the reservation management system.
475 	 */
476 	vm_reserv_init();
477 #endif
478 	return (vaddr);
479 }
480 
481 void
482 vm_page_reference(vm_page_t m)
483 {
484 
485 	vm_page_aflag_set(m, PGA_REFERENCED);
486 }
487 
488 /*
489  *	vm_page_busy_downgrade:
490  *
491  *	Downgrade an exclusive busy page into a single shared busy page.
492  */
493 void
494 vm_page_busy_downgrade(vm_page_t m)
495 {
496 	u_int x;
497 
498 	vm_page_assert_xbusied(m);
499 
500 	for (;;) {
501 		x = m->busy_lock;
502 		x &= VPB_BIT_WAITERS;
503 		if (atomic_cmpset_rel_int(&m->busy_lock,
504 		    VPB_SINGLE_EXCLUSIVER | x, VPB_SHARERS_WORD(1) | x))
505 			break;
506 	}
507 }
508 
509 /*
510  *	vm_page_sbusied:
511  *
512  *	Return a positive value if the page is shared busied, 0 otherwise.
513  */
514 int
515 vm_page_sbusied(vm_page_t m)
516 {
517 	u_int x;
518 
519 	x = m->busy_lock;
520 	return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED);
521 }
522 
523 /*
524  *	vm_page_sunbusy:
525  *
526  *	Shared unbusy a page.
527  */
528 void
529 vm_page_sunbusy(vm_page_t m)
530 {
531 	u_int x;
532 
533 	vm_page_assert_sbusied(m);
534 
535 	for (;;) {
536 		x = m->busy_lock;
537 		if (VPB_SHARERS(x) > 1) {
538 			if (atomic_cmpset_int(&m->busy_lock, x,
539 			    x - VPB_ONE_SHARER))
540 				break;
541 			continue;
542 		}
543 		if ((x & VPB_BIT_WAITERS) == 0) {
544 			KASSERT(x == VPB_SHARERS_WORD(1),
545 			    ("vm_page_sunbusy: invalid lock state"));
546 			if (atomic_cmpset_int(&m->busy_lock,
547 			    VPB_SHARERS_WORD(1), VPB_UNBUSIED))
548 				break;
549 			continue;
550 		}
551 		KASSERT(x == (VPB_SHARERS_WORD(1) | VPB_BIT_WAITERS),
552 		    ("vm_page_sunbusy: invalid lock state for waiters"));
553 
554 		vm_page_lock(m);
555 		if (!atomic_cmpset_int(&m->busy_lock, x, VPB_UNBUSIED)) {
556 			vm_page_unlock(m);
557 			continue;
558 		}
559 		wakeup(m);
560 		vm_page_unlock(m);
561 		break;
562 	}
563 }
564 
565 /*
566  *	vm_page_busy_sleep:
567  *
568  *	Sleep and release the page lock, using the page pointer as wchan.
569  *	This is used to implement the hard-path of busying mechanism.
570  *
571  *	The given page must be locked.
572  */
573 void
574 vm_page_busy_sleep(vm_page_t m, const char *wmesg)
575 {
576 	u_int x;
577 
578 	vm_page_lock_assert(m, MA_OWNED);
579 
580 	x = m->busy_lock;
581 	if (x == VPB_UNBUSIED) {
582 		vm_page_unlock(m);
583 		return;
584 	}
585 	if ((x & VPB_BIT_WAITERS) == 0 &&
586 	    !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS)) {
587 		vm_page_unlock(m);
588 		return;
589 	}
590 	msleep(m, vm_page_lockptr(m), PVM | PDROP, wmesg, 0);
591 }
592 
593 /*
594  *	vm_page_trysbusy:
595  *
596  *	Try to shared busy a page.
597  *	If the operation succeeds 1 is returned otherwise 0.
598  *	The operation never sleeps.
599  */
600 int
601 vm_page_trysbusy(vm_page_t m)
602 {
603 	u_int x;
604 
605 	for (;;) {
606 		x = m->busy_lock;
607 		if ((x & VPB_BIT_SHARED) == 0)
608 			return (0);
609 		if (atomic_cmpset_acq_int(&m->busy_lock, x, x + VPB_ONE_SHARER))
610 			return (1);
611 	}
612 }
613 
614 /*
615  *	vm_page_xunbusy_hard:
616  *
617  *	Called after the first try the exclusive unbusy of a page failed.
618  *	It is assumed that the waiters bit is on.
619  */
620 void
621 vm_page_xunbusy_hard(vm_page_t m)
622 {
623 
624 	vm_page_assert_xbusied(m);
625 
626 	vm_page_lock(m);
627 	atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
628 	wakeup(m);
629 	vm_page_unlock(m);
630 }
631 
632 /*
633  *	vm_page_flash:
634  *
635  *	Wakeup anyone waiting for the page.
636  *	The ownership bits do not change.
637  *
638  *	The given page must be locked.
639  */
640 void
641 vm_page_flash(vm_page_t m)
642 {
643 	u_int x;
644 
645 	vm_page_lock_assert(m, MA_OWNED);
646 
647 	for (;;) {
648 		x = m->busy_lock;
649 		if ((x & VPB_BIT_WAITERS) == 0)
650 			return;
651 		if (atomic_cmpset_int(&m->busy_lock, x,
652 		    x & (~VPB_BIT_WAITERS)))
653 			break;
654 	}
655 	wakeup(m);
656 }
657 
658 /*
659  * Keep page from being freed by the page daemon
660  * much of the same effect as wiring, except much lower
661  * overhead and should be used only for *very* temporary
662  * holding ("wiring").
663  */
664 void
665 vm_page_hold(vm_page_t mem)
666 {
667 
668 	vm_page_lock_assert(mem, MA_OWNED);
669         mem->hold_count++;
670 }
671 
672 void
673 vm_page_unhold(vm_page_t mem)
674 {
675 
676 	vm_page_lock_assert(mem, MA_OWNED);
677 	--mem->hold_count;
678 	KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!"));
679 	if (mem->hold_count == 0 && (mem->flags & PG_UNHOLDFREE) != 0)
680 		vm_page_free_toq(mem);
681 }
682 
683 /*
684  *	vm_page_unhold_pages:
685  *
686  *	Unhold each of the pages that is referenced by the given array.
687  */
688 void
689 vm_page_unhold_pages(vm_page_t *ma, int count)
690 {
691 	struct mtx *mtx, *new_mtx;
692 
693 	mtx = NULL;
694 	for (; count != 0; count--) {
695 		/*
696 		 * Avoid releasing and reacquiring the same page lock.
697 		 */
698 		new_mtx = vm_page_lockptr(*ma);
699 		if (mtx != new_mtx) {
700 			if (mtx != NULL)
701 				mtx_unlock(mtx);
702 			mtx = new_mtx;
703 			mtx_lock(mtx);
704 		}
705 		vm_page_unhold(*ma);
706 		ma++;
707 	}
708 	if (mtx != NULL)
709 		mtx_unlock(mtx);
710 }
711 
712 vm_page_t
713 PHYS_TO_VM_PAGE(vm_paddr_t pa)
714 {
715 	vm_page_t m;
716 
717 #ifdef VM_PHYSSEG_SPARSE
718 	m = vm_phys_paddr_to_vm_page(pa);
719 	if (m == NULL)
720 		m = vm_phys_fictitious_to_vm_page(pa);
721 	return (m);
722 #elif defined(VM_PHYSSEG_DENSE)
723 	long pi;
724 
725 	pi = atop(pa);
726 	if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
727 		m = &vm_page_array[pi - first_page];
728 		return (m);
729 	}
730 	return (vm_phys_fictitious_to_vm_page(pa));
731 #else
732 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
733 #endif
734 }
735 
736 /*
737  *	vm_page_getfake:
738  *
739  *	Create a fictitious page with the specified physical address and
740  *	memory attribute.  The memory attribute is the only the machine-
741  *	dependent aspect of a fictitious page that must be initialized.
742  */
743 vm_page_t
744 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
745 {
746 	vm_page_t m;
747 
748 	m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
749 	vm_page_initfake(m, paddr, memattr);
750 	return (m);
751 }
752 
753 void
754 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
755 {
756 
757 	if ((m->flags & PG_FICTITIOUS) != 0) {
758 		/*
759 		 * The page's memattr might have changed since the
760 		 * previous initialization.  Update the pmap to the
761 		 * new memattr.
762 		 */
763 		goto memattr;
764 	}
765 	m->phys_addr = paddr;
766 	m->queue = PQ_NONE;
767 	/* Fictitious pages don't use "segind". */
768 	m->flags = PG_FICTITIOUS;
769 	/* Fictitious pages don't use "order" or "pool". */
770 	m->oflags = VPO_UNMANAGED;
771 	m->busy_lock = VPB_SINGLE_EXCLUSIVER;
772 	m->wire_count = 1;
773 	pmap_page_init(m);
774 memattr:
775 	pmap_page_set_memattr(m, memattr);
776 }
777 
778 /*
779  *	vm_page_putfake:
780  *
781  *	Release a fictitious page.
782  */
783 void
784 vm_page_putfake(vm_page_t m)
785 {
786 
787 	KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
788 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
789 	    ("vm_page_putfake: bad page %p", m));
790 	uma_zfree(fakepg_zone, m);
791 }
792 
793 /*
794  *	vm_page_updatefake:
795  *
796  *	Update the given fictitious page to the specified physical address and
797  *	memory attribute.
798  */
799 void
800 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
801 {
802 
803 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
804 	    ("vm_page_updatefake: bad page %p", m));
805 	m->phys_addr = paddr;
806 	pmap_page_set_memattr(m, memattr);
807 }
808 
809 /*
810  *	vm_page_free:
811  *
812  *	Free a page.
813  */
814 void
815 vm_page_free(vm_page_t m)
816 {
817 
818 	m->flags &= ~PG_ZERO;
819 	vm_page_free_toq(m);
820 }
821 
822 /*
823  *	vm_page_free_zero:
824  *
825  *	Free a page to the zerod-pages queue
826  */
827 void
828 vm_page_free_zero(vm_page_t m)
829 {
830 
831 	m->flags |= PG_ZERO;
832 	vm_page_free_toq(m);
833 }
834 
835 /*
836  * Unbusy and handle the page queueing for a page from the VOP_GETPAGES()
837  * array which is not the request page.
838  */
839 void
840 vm_page_readahead_finish(vm_page_t m)
841 {
842 
843 	if (m->valid != 0) {
844 		/*
845 		 * Since the page is not the requested page, whether
846 		 * it should be activated or deactivated is not
847 		 * obvious.  Empirical results have shown that
848 		 * deactivating the page is usually the best choice,
849 		 * unless the page is wanted by another thread.
850 		 */
851 		vm_page_lock(m);
852 		if ((m->busy_lock & VPB_BIT_WAITERS) != 0)
853 			vm_page_activate(m);
854 		else
855 			vm_page_deactivate(m);
856 		vm_page_unlock(m);
857 		vm_page_xunbusy(m);
858 	} else {
859 		/*
860 		 * Free the completely invalid page.  Such page state
861 		 * occurs due to the short read operation which did
862 		 * not covered our page at all, or in case when a read
863 		 * error happens.
864 		 */
865 		vm_page_lock(m);
866 		vm_page_free(m);
867 		vm_page_unlock(m);
868 	}
869 }
870 
871 /*
872  *	vm_page_sleep_if_busy:
873  *
874  *	Sleep and release the page queues lock if the page is busied.
875  *	Returns TRUE if the thread slept.
876  *
877  *	The given page must be unlocked and object containing it must
878  *	be locked.
879  */
880 int
881 vm_page_sleep_if_busy(vm_page_t m, const char *msg)
882 {
883 	vm_object_t obj;
884 
885 	vm_page_lock_assert(m, MA_NOTOWNED);
886 	VM_OBJECT_ASSERT_WLOCKED(m->object);
887 
888 	if (vm_page_busied(m)) {
889 		/*
890 		 * The page-specific object must be cached because page
891 		 * identity can change during the sleep, causing the
892 		 * re-lock of a different object.
893 		 * It is assumed that a reference to the object is already
894 		 * held by the callers.
895 		 */
896 		obj = m->object;
897 		vm_page_lock(m);
898 		VM_OBJECT_WUNLOCK(obj);
899 		vm_page_busy_sleep(m, msg);
900 		VM_OBJECT_WLOCK(obj);
901 		return (TRUE);
902 	}
903 	return (FALSE);
904 }
905 
906 /*
907  *	vm_page_dirty_KBI:		[ internal use only ]
908  *
909  *	Set all bits in the page's dirty field.
910  *
911  *	The object containing the specified page must be locked if the
912  *	call is made from the machine-independent layer.
913  *
914  *	See vm_page_clear_dirty_mask().
915  *
916  *	This function should only be called by vm_page_dirty().
917  */
918 void
919 vm_page_dirty_KBI(vm_page_t m)
920 {
921 
922 	/* These assertions refer to this operation by its public name. */
923 	KASSERT((m->flags & PG_CACHED) == 0,
924 	    ("vm_page_dirty: page in cache!"));
925 	KASSERT(!VM_PAGE_IS_FREE(m),
926 	    ("vm_page_dirty: page is free!"));
927 	KASSERT(m->valid == VM_PAGE_BITS_ALL,
928 	    ("vm_page_dirty: page is invalid!"));
929 	m->dirty = VM_PAGE_BITS_ALL;
930 }
931 
932 /*
933  *	vm_page_insert:		[ internal use only ]
934  *
935  *	Inserts the given mem entry into the object and object list.
936  *
937  *	The object must be locked.
938  */
939 int
940 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
941 {
942 	vm_page_t mpred;
943 
944 	VM_OBJECT_ASSERT_WLOCKED(object);
945 	mpred = vm_radix_lookup_le(&object->rtree, pindex);
946 	return (vm_page_insert_after(m, object, pindex, mpred));
947 }
948 
949 /*
950  *	vm_page_insert_after:
951  *
952  *	Inserts the page "m" into the specified object at offset "pindex".
953  *
954  *	The page "mpred" must immediately precede the offset "pindex" within
955  *	the specified object.
956  *
957  *	The object must be locked.
958  */
959 static int
960 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
961     vm_page_t mpred)
962 {
963 	vm_pindex_t sidx;
964 	vm_object_t sobj;
965 	vm_page_t msucc;
966 
967 	VM_OBJECT_ASSERT_WLOCKED(object);
968 	KASSERT(m->object == NULL,
969 	    ("vm_page_insert_after: page already inserted"));
970 	if (mpred != NULL) {
971 		KASSERT(mpred->object == object ||
972 		    (mpred->flags & PG_SLAB) != 0,
973 		    ("vm_page_insert_after: object doesn't contain mpred"));
974 		KASSERT(mpred->pindex < pindex,
975 		    ("vm_page_insert_after: mpred doesn't precede pindex"));
976 		msucc = TAILQ_NEXT(mpred, listq);
977 	} else
978 		msucc = TAILQ_FIRST(&object->memq);
979 	if (msucc != NULL)
980 		KASSERT(msucc->pindex > pindex,
981 		    ("vm_page_insert_after: msucc doesn't succeed pindex"));
982 
983 	/*
984 	 * Record the object/offset pair in this page
985 	 */
986 	sobj = m->object;
987 	sidx = m->pindex;
988 	m->object = object;
989 	m->pindex = pindex;
990 
991 	/*
992 	 * Now link into the object's ordered list of backed pages.
993 	 */
994 	if (vm_radix_insert(&object->rtree, m)) {
995 		m->object = sobj;
996 		m->pindex = sidx;
997 		return (1);
998 	}
999 	vm_page_insert_radixdone(m, object, mpred);
1000 	return (0);
1001 }
1002 
1003 /*
1004  *	vm_page_insert_radixdone:
1005  *
1006  *	Complete page "m" insertion into the specified object after the
1007  *	radix trie hooking.
1008  *
1009  *	The page "mpred" must precede the offset "m->pindex" within the
1010  *	specified object.
1011  *
1012  *	The object must be locked.
1013  */
1014 static void
1015 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1016 {
1017 
1018 	VM_OBJECT_ASSERT_WLOCKED(object);
1019 	KASSERT(object != NULL && m->object == object,
1020 	    ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1021 	if (mpred != NULL) {
1022 		KASSERT(mpred->object == object ||
1023 		    (mpred->flags & PG_SLAB) != 0,
1024 		    ("vm_page_insert_after: object doesn't contain mpred"));
1025 		KASSERT(mpred->pindex < m->pindex,
1026 		    ("vm_page_insert_after: mpred doesn't precede pindex"));
1027 	}
1028 
1029 	if (mpred != NULL)
1030 		TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1031 	else
1032 		TAILQ_INSERT_HEAD(&object->memq, m, listq);
1033 
1034 	/*
1035 	 * Show that the object has one more resident page.
1036 	 */
1037 	object->resident_page_count++;
1038 
1039 	/*
1040 	 * Hold the vnode until the last page is released.
1041 	 */
1042 	if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1043 		vhold(object->handle);
1044 
1045 	/*
1046 	 * Since we are inserting a new and possibly dirty page,
1047 	 * update the object's OBJ_MIGHTBEDIRTY flag.
1048 	 */
1049 	if (pmap_page_is_write_mapped(m))
1050 		vm_object_set_writeable_dirty(object);
1051 }
1052 
1053 /*
1054  *	vm_page_remove:
1055  *
1056  *	Removes the given mem entry from the object/offset-page
1057  *	table and the object page list, but do not invalidate/terminate
1058  *	the backing store.
1059  *
1060  *	The object must be locked.  The page must be locked if it is managed.
1061  */
1062 void
1063 vm_page_remove(vm_page_t m)
1064 {
1065 	vm_object_t object;
1066 	boolean_t lockacq;
1067 
1068 	if ((m->oflags & VPO_UNMANAGED) == 0)
1069 		vm_page_lock_assert(m, MA_OWNED);
1070 	if ((object = m->object) == NULL)
1071 		return;
1072 	VM_OBJECT_ASSERT_WLOCKED(object);
1073 	if (vm_page_xbusied(m)) {
1074 		lockacq = FALSE;
1075 		if ((m->oflags & VPO_UNMANAGED) != 0 &&
1076 		    !mtx_owned(vm_page_lockptr(m))) {
1077 			lockacq = TRUE;
1078 			vm_page_lock(m);
1079 		}
1080 		vm_page_flash(m);
1081 		atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
1082 		if (lockacq)
1083 			vm_page_unlock(m);
1084 	}
1085 
1086 	/*
1087 	 * Now remove from the object's list of backed pages.
1088 	 */
1089 	vm_radix_remove(&object->rtree, m->pindex);
1090 	TAILQ_REMOVE(&object->memq, m, listq);
1091 
1092 	/*
1093 	 * And show that the object has one fewer resident page.
1094 	 */
1095 	object->resident_page_count--;
1096 
1097 	/*
1098 	 * The vnode may now be recycled.
1099 	 */
1100 	if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1101 		vdrop(object->handle);
1102 
1103 	m->object = NULL;
1104 }
1105 
1106 /*
1107  *	vm_page_lookup:
1108  *
1109  *	Returns the page associated with the object/offset
1110  *	pair specified; if none is found, NULL is returned.
1111  *
1112  *	The object must be locked.
1113  */
1114 vm_page_t
1115 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1116 {
1117 
1118 	VM_OBJECT_ASSERT_LOCKED(object);
1119 	return (vm_radix_lookup(&object->rtree, pindex));
1120 }
1121 
1122 /*
1123  *	vm_page_find_least:
1124  *
1125  *	Returns the page associated with the object with least pindex
1126  *	greater than or equal to the parameter pindex, or NULL.
1127  *
1128  *	The object must be locked.
1129  */
1130 vm_page_t
1131 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1132 {
1133 	vm_page_t m;
1134 
1135 	VM_OBJECT_ASSERT_LOCKED(object);
1136 	if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1137 		m = vm_radix_lookup_ge(&object->rtree, pindex);
1138 	return (m);
1139 }
1140 
1141 /*
1142  * Returns the given page's successor (by pindex) within the object if it is
1143  * resident; if none is found, NULL is returned.
1144  *
1145  * The object must be locked.
1146  */
1147 vm_page_t
1148 vm_page_next(vm_page_t m)
1149 {
1150 	vm_page_t next;
1151 
1152 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1153 	if ((next = TAILQ_NEXT(m, listq)) != NULL &&
1154 	    next->pindex != m->pindex + 1)
1155 		next = NULL;
1156 	return (next);
1157 }
1158 
1159 /*
1160  * Returns the given page's predecessor (by pindex) within the object if it is
1161  * resident; if none is found, NULL is returned.
1162  *
1163  * The object must be locked.
1164  */
1165 vm_page_t
1166 vm_page_prev(vm_page_t m)
1167 {
1168 	vm_page_t prev;
1169 
1170 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1171 	if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL &&
1172 	    prev->pindex != m->pindex - 1)
1173 		prev = NULL;
1174 	return (prev);
1175 }
1176 
1177 /*
1178  * Uses the page mnew as a replacement for an existing page at index
1179  * pindex which must be already present in the object.
1180  *
1181  * The existing page must not be on a paging queue.
1182  */
1183 vm_page_t
1184 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex)
1185 {
1186 	vm_page_t mold, mpred;
1187 
1188 	VM_OBJECT_ASSERT_WLOCKED(object);
1189 
1190 	/*
1191 	 * This function mostly follows vm_page_insert() and
1192 	 * vm_page_remove() without the radix, object count and vnode
1193 	 * dance.  Double check such functions for more comments.
1194 	 */
1195 	mpred = vm_radix_lookup(&object->rtree, pindex);
1196 	KASSERT(mpred != NULL,
1197 	    ("vm_page_replace: replacing page not present with pindex"));
1198 	mpred = TAILQ_PREV(mpred, respgs, listq);
1199 	if (mpred != NULL)
1200 		KASSERT(mpred->pindex < pindex,
1201 		    ("vm_page_insert_after: mpred doesn't precede pindex"));
1202 
1203 	mnew->object = object;
1204 	mnew->pindex = pindex;
1205 	mold = vm_radix_replace(&object->rtree, mnew, pindex);
1206 	KASSERT(mold->queue == PQ_NONE,
1207 	    ("vm_page_replace: mold is on a paging queue"));
1208 
1209 	/* Detach the old page from the resident tailq. */
1210 	TAILQ_REMOVE(&object->memq, mold, listq);
1211 
1212 	mold->object = NULL;
1213 	vm_page_xunbusy(mold);
1214 
1215 	/* Insert the new page in the resident tailq. */
1216 	if (mpred != NULL)
1217 		TAILQ_INSERT_AFTER(&object->memq, mpred, mnew, listq);
1218 	else
1219 		TAILQ_INSERT_HEAD(&object->memq, mnew, listq);
1220 	if (pmap_page_is_write_mapped(mnew))
1221 		vm_object_set_writeable_dirty(object);
1222 	return (mold);
1223 }
1224 
1225 /*
1226  *	vm_page_rename:
1227  *
1228  *	Move the given memory entry from its
1229  *	current object to the specified target object/offset.
1230  *
1231  *	Note: swap associated with the page must be invalidated by the move.  We
1232  *	      have to do this for several reasons:  (1) we aren't freeing the
1233  *	      page, (2) we are dirtying the page, (3) the VM system is probably
1234  *	      moving the page from object A to B, and will then later move
1235  *	      the backing store from A to B and we can't have a conflict.
1236  *
1237  *	Note: we *always* dirty the page.  It is necessary both for the
1238  *	      fact that we moved it, and because we may be invalidating
1239  *	      swap.  If the page is on the cache, we have to deactivate it
1240  *	      or vm_page_dirty() will panic.  Dirty pages are not allowed
1241  *	      on the cache.
1242  *
1243  *	The objects must be locked.
1244  */
1245 int
1246 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1247 {
1248 	vm_page_t mpred;
1249 	vm_pindex_t opidx;
1250 
1251 	VM_OBJECT_ASSERT_WLOCKED(new_object);
1252 
1253 	mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex);
1254 	KASSERT(mpred == NULL || mpred->pindex != new_pindex,
1255 	    ("vm_page_rename: pindex already renamed"));
1256 
1257 	/*
1258 	 * Create a custom version of vm_page_insert() which does not depend
1259 	 * by m_prev and can cheat on the implementation aspects of the
1260 	 * function.
1261 	 */
1262 	opidx = m->pindex;
1263 	m->pindex = new_pindex;
1264 	if (vm_radix_insert(&new_object->rtree, m)) {
1265 		m->pindex = opidx;
1266 		return (1);
1267 	}
1268 
1269 	/*
1270 	 * The operation cannot fail anymore.  The removal must happen before
1271 	 * the listq iterator is tainted.
1272 	 */
1273 	m->pindex = opidx;
1274 	vm_page_lock(m);
1275 	vm_page_remove(m);
1276 
1277 	/* Return back to the new pindex to complete vm_page_insert(). */
1278 	m->pindex = new_pindex;
1279 	m->object = new_object;
1280 	vm_page_unlock(m);
1281 	vm_page_insert_radixdone(m, new_object, mpred);
1282 	vm_page_dirty(m);
1283 	return (0);
1284 }
1285 
1286 /*
1287  *	Convert all of the given object's cached pages that have a
1288  *	pindex within the given range into free pages.  If the value
1289  *	zero is given for "end", then the range's upper bound is
1290  *	infinity.  If the given object is backed by a vnode and it
1291  *	transitions from having one or more cached pages to none, the
1292  *	vnode's hold count is reduced.
1293  */
1294 void
1295 vm_page_cache_free(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1296 {
1297 	vm_page_t m;
1298 	boolean_t empty;
1299 
1300 	mtx_lock(&vm_page_queue_free_mtx);
1301 	if (__predict_false(vm_radix_is_empty(&object->cache))) {
1302 		mtx_unlock(&vm_page_queue_free_mtx);
1303 		return;
1304 	}
1305 	while ((m = vm_radix_lookup_ge(&object->cache, start)) != NULL) {
1306 		if (end != 0 && m->pindex >= end)
1307 			break;
1308 		vm_radix_remove(&object->cache, m->pindex);
1309 		vm_page_cache_turn_free(m);
1310 	}
1311 	empty = vm_radix_is_empty(&object->cache);
1312 	mtx_unlock(&vm_page_queue_free_mtx);
1313 	if (object->type == OBJT_VNODE && empty)
1314 		vdrop(object->handle);
1315 }
1316 
1317 /*
1318  *	Returns the cached page that is associated with the given
1319  *	object and offset.  If, however, none exists, returns NULL.
1320  *
1321  *	The free page queue must be locked.
1322  */
1323 static inline vm_page_t
1324 vm_page_cache_lookup(vm_object_t object, vm_pindex_t pindex)
1325 {
1326 
1327 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1328 	return (vm_radix_lookup(&object->cache, pindex));
1329 }
1330 
1331 /*
1332  *	Remove the given cached page from its containing object's
1333  *	collection of cached pages.
1334  *
1335  *	The free page queue must be locked.
1336  */
1337 static void
1338 vm_page_cache_remove(vm_page_t m)
1339 {
1340 
1341 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1342 	KASSERT((m->flags & PG_CACHED) != 0,
1343 	    ("vm_page_cache_remove: page %p is not cached", m));
1344 	vm_radix_remove(&m->object->cache, m->pindex);
1345 	m->object = NULL;
1346 	cnt.v_cache_count--;
1347 }
1348 
1349 /*
1350  *	Transfer all of the cached pages with offset greater than or
1351  *	equal to 'offidxstart' from the original object's cache to the
1352  *	new object's cache.  However, any cached pages with offset
1353  *	greater than or equal to the new object's size are kept in the
1354  *	original object.  Initially, the new object's cache must be
1355  *	empty.  Offset 'offidxstart' in the original object must
1356  *	correspond to offset zero in the new object.
1357  *
1358  *	The new object must be locked.
1359  */
1360 void
1361 vm_page_cache_transfer(vm_object_t orig_object, vm_pindex_t offidxstart,
1362     vm_object_t new_object)
1363 {
1364 	vm_page_t m;
1365 
1366 	/*
1367 	 * Insertion into an object's collection of cached pages
1368 	 * requires the object to be locked.  In contrast, removal does
1369 	 * not.
1370 	 */
1371 	VM_OBJECT_ASSERT_WLOCKED(new_object);
1372 	KASSERT(vm_radix_is_empty(&new_object->cache),
1373 	    ("vm_page_cache_transfer: object %p has cached pages",
1374 	    new_object));
1375 	mtx_lock(&vm_page_queue_free_mtx);
1376 	while ((m = vm_radix_lookup_ge(&orig_object->cache,
1377 	    offidxstart)) != NULL) {
1378 		/*
1379 		 * Transfer all of the pages with offset greater than or
1380 		 * equal to 'offidxstart' from the original object's
1381 		 * cache to the new object's cache.
1382 		 */
1383 		if ((m->pindex - offidxstart) >= new_object->size)
1384 			break;
1385 		vm_radix_remove(&orig_object->cache, m->pindex);
1386 		/* Update the page's object and offset. */
1387 		m->object = new_object;
1388 		m->pindex -= offidxstart;
1389 		if (vm_radix_insert(&new_object->cache, m))
1390 			vm_page_cache_turn_free(m);
1391 	}
1392 	mtx_unlock(&vm_page_queue_free_mtx);
1393 }
1394 
1395 /*
1396  *	Returns TRUE if a cached page is associated with the given object and
1397  *	offset, and FALSE otherwise.
1398  *
1399  *	The object must be locked.
1400  */
1401 boolean_t
1402 vm_page_is_cached(vm_object_t object, vm_pindex_t pindex)
1403 {
1404 	vm_page_t m;
1405 
1406 	/*
1407 	 * Insertion into an object's collection of cached pages requires the
1408 	 * object to be locked.  Therefore, if the object is locked and the
1409 	 * object's collection is empty, there is no need to acquire the free
1410 	 * page queues lock in order to prove that the specified page doesn't
1411 	 * exist.
1412 	 */
1413 	VM_OBJECT_ASSERT_WLOCKED(object);
1414 	if (__predict_true(vm_object_cache_is_empty(object)))
1415 		return (FALSE);
1416 	mtx_lock(&vm_page_queue_free_mtx);
1417 	m = vm_page_cache_lookup(object, pindex);
1418 	mtx_unlock(&vm_page_queue_free_mtx);
1419 	return (m != NULL);
1420 }
1421 
1422 /*
1423  *	vm_page_alloc:
1424  *
1425  *	Allocate and return a page that is associated with the specified
1426  *	object and offset pair.  By default, this page is exclusive busied.
1427  *
1428  *	The caller must always specify an allocation class.
1429  *
1430  *	allocation classes:
1431  *	VM_ALLOC_NORMAL		normal process request
1432  *	VM_ALLOC_SYSTEM		system *really* needs a page
1433  *	VM_ALLOC_INTERRUPT	interrupt time request
1434  *
1435  *	optional allocation flags:
1436  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
1437  *				intends to allocate
1438  *	VM_ALLOC_IFCACHED	return page only if it is cached
1439  *	VM_ALLOC_IFNOTCACHED	return NULL, do not reactivate if the page
1440  *				is cached
1441  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
1442  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
1443  *	VM_ALLOC_NOOBJ		page is not associated with an object and
1444  *				should not be exclusive busy
1445  *	VM_ALLOC_SBUSY		shared busy the allocated page
1446  *	VM_ALLOC_WIRED		wire the allocated page
1447  *	VM_ALLOC_ZERO		prefer a zeroed page
1448  *
1449  *	This routine may not sleep.
1450  */
1451 vm_page_t
1452 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1453 {
1454 	struct vnode *vp = NULL;
1455 	vm_object_t m_object;
1456 	vm_page_t m, mpred;
1457 	int flags, req_class;
1458 
1459 	mpred = 0;	/* XXX: pacify gcc */
1460 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1461 	    (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1462 	    ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1463 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1464 	    ("vm_page_alloc: inconsistent object(%p)/req(%x)", (void *)object,
1465 	    req));
1466 	if (object != NULL)
1467 		VM_OBJECT_ASSERT_WLOCKED(object);
1468 
1469 	req_class = req & VM_ALLOC_CLASS_MASK;
1470 
1471 	/*
1472 	 * The page daemon is allowed to dig deeper into the free page list.
1473 	 */
1474 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1475 		req_class = VM_ALLOC_SYSTEM;
1476 
1477 	if (object != NULL) {
1478 		mpred = vm_radix_lookup_le(&object->rtree, pindex);
1479 		KASSERT(mpred == NULL || mpred->pindex != pindex,
1480 		   ("vm_page_alloc: pindex already allocated"));
1481 	}
1482 
1483 	/*
1484 	 * The page allocation request can came from consumers which already
1485 	 * hold the free page queue mutex, like vm_page_insert() in
1486 	 * vm_page_cache().
1487 	 */
1488 	mtx_lock_flags(&vm_page_queue_free_mtx, MTX_RECURSE);
1489 	if (cnt.v_free_count + cnt.v_cache_count > cnt.v_free_reserved ||
1490 	    (req_class == VM_ALLOC_SYSTEM &&
1491 	    cnt.v_free_count + cnt.v_cache_count > cnt.v_interrupt_free_min) ||
1492 	    (req_class == VM_ALLOC_INTERRUPT &&
1493 	    cnt.v_free_count + cnt.v_cache_count > 0)) {
1494 		/*
1495 		 * Allocate from the free queue if the number of free pages
1496 		 * exceeds the minimum for the request class.
1497 		 */
1498 		if (object != NULL &&
1499 		    (m = vm_page_cache_lookup(object, pindex)) != NULL) {
1500 			if ((req & VM_ALLOC_IFNOTCACHED) != 0) {
1501 				mtx_unlock(&vm_page_queue_free_mtx);
1502 				return (NULL);
1503 			}
1504 			if (vm_phys_unfree_page(m))
1505 				vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m, 0);
1506 #if VM_NRESERVLEVEL > 0
1507 			else if (!vm_reserv_reactivate_page(m))
1508 #else
1509 			else
1510 #endif
1511 				panic("vm_page_alloc: cache page %p is missing"
1512 				    " from the free queue", m);
1513 		} else if ((req & VM_ALLOC_IFCACHED) != 0) {
1514 			mtx_unlock(&vm_page_queue_free_mtx);
1515 			return (NULL);
1516 #if VM_NRESERVLEVEL > 0
1517 		} else if (object == NULL || (object->flags & (OBJ_COLORED |
1518 		    OBJ_FICTITIOUS)) != OBJ_COLORED || (m =
1519 		    vm_reserv_alloc_page(object, pindex, mpred)) == NULL) {
1520 #else
1521 		} else {
1522 #endif
1523 			m = vm_phys_alloc_pages(object != NULL ?
1524 			    VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0);
1525 #if VM_NRESERVLEVEL > 0
1526 			if (m == NULL && vm_reserv_reclaim_inactive()) {
1527 				m = vm_phys_alloc_pages(object != NULL ?
1528 				    VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT,
1529 				    0);
1530 			}
1531 #endif
1532 		}
1533 	} else {
1534 		/*
1535 		 * Not allocatable, give up.
1536 		 */
1537 		mtx_unlock(&vm_page_queue_free_mtx);
1538 		atomic_add_int(&vm_pageout_deficit,
1539 		    max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
1540 		pagedaemon_wakeup();
1541 		return (NULL);
1542 	}
1543 
1544 	/*
1545 	 *  At this point we had better have found a good page.
1546 	 */
1547 	KASSERT(m != NULL, ("vm_page_alloc: missing page"));
1548 	KASSERT(m->queue == PQ_NONE,
1549 	    ("vm_page_alloc: page %p has unexpected queue %d", m, m->queue));
1550 	KASSERT(m->wire_count == 0, ("vm_page_alloc: page %p is wired", m));
1551 	KASSERT(m->hold_count == 0, ("vm_page_alloc: page %p is held", m));
1552 	KASSERT(!vm_page_sbusied(m),
1553 	    ("vm_page_alloc: page %p is busy", m));
1554 	KASSERT(m->dirty == 0, ("vm_page_alloc: page %p is dirty", m));
1555 	KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
1556 	    ("vm_page_alloc: page %p has unexpected memattr %d", m,
1557 	    pmap_page_get_memattr(m)));
1558 	if ((m->flags & PG_CACHED) != 0) {
1559 		KASSERT((m->flags & PG_ZERO) == 0,
1560 		    ("vm_page_alloc: cached page %p is PG_ZERO", m));
1561 		KASSERT(m->valid != 0,
1562 		    ("vm_page_alloc: cached page %p is invalid", m));
1563 		if (m->object == object && m->pindex == pindex)
1564 	  		cnt.v_reactivated++;
1565 		else
1566 			m->valid = 0;
1567 		m_object = m->object;
1568 		vm_page_cache_remove(m);
1569 		if (m_object->type == OBJT_VNODE &&
1570 		    vm_object_cache_is_empty(m_object))
1571 			vp = m_object->handle;
1572 	} else {
1573 		KASSERT(VM_PAGE_IS_FREE(m),
1574 		    ("vm_page_alloc: page %p is not free", m));
1575 		KASSERT(m->valid == 0,
1576 		    ("vm_page_alloc: free page %p is valid", m));
1577 		vm_phys_freecnt_adj(m, -1);
1578 	}
1579 
1580 	/*
1581 	 * Only the PG_ZERO flag is inherited.  The PG_CACHED or PG_FREE flag
1582 	 * must be cleared before the free page queues lock is released.
1583 	 */
1584 	flags = 0;
1585 	if (m->flags & PG_ZERO) {
1586 		vm_page_zero_count--;
1587 		if (req & VM_ALLOC_ZERO)
1588 			flags = PG_ZERO;
1589 	}
1590 	if (req & VM_ALLOC_NODUMP)
1591 		flags |= PG_NODUMP;
1592 	m->flags = flags;
1593 	mtx_unlock(&vm_page_queue_free_mtx);
1594 	m->aflags = 0;
1595 	m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
1596 	    VPO_UNMANAGED : 0;
1597 	m->busy_lock = VPB_UNBUSIED;
1598 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
1599 		m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1600 	if ((req & VM_ALLOC_SBUSY) != 0)
1601 		m->busy_lock = VPB_SHARERS_WORD(1);
1602 	if (req & VM_ALLOC_WIRED) {
1603 		/*
1604 		 * The page lock is not required for wiring a page until that
1605 		 * page is inserted into the object.
1606 		 */
1607 		atomic_add_int(&cnt.v_wire_count, 1);
1608 		m->wire_count = 1;
1609 	}
1610 	m->act_count = 0;
1611 
1612 	if (object != NULL) {
1613 		if (vm_page_insert_after(m, object, pindex, mpred)) {
1614 			/* See the comment below about hold count. */
1615 			if (vp != NULL)
1616 				vdrop(vp);
1617 			pagedaemon_wakeup();
1618 			if (req & VM_ALLOC_WIRED) {
1619 				atomic_subtract_int(&cnt.v_wire_count, 1);
1620 				m->wire_count = 0;
1621 			}
1622 			m->object = NULL;
1623 			vm_page_free(m);
1624 			return (NULL);
1625 		}
1626 
1627 		/* Ignore device objects; the pager sets "memattr" for them. */
1628 		if (object->memattr != VM_MEMATTR_DEFAULT &&
1629 		    (object->flags & OBJ_FICTITIOUS) == 0)
1630 			pmap_page_set_memattr(m, object->memattr);
1631 	} else
1632 		m->pindex = pindex;
1633 
1634 	/*
1635 	 * The following call to vdrop() must come after the above call
1636 	 * to vm_page_insert() in case both affect the same object and
1637 	 * vnode.  Otherwise, the affected vnode's hold count could
1638 	 * temporarily become zero.
1639 	 */
1640 	if (vp != NULL)
1641 		vdrop(vp);
1642 
1643 	/*
1644 	 * Don't wakeup too often - wakeup the pageout daemon when
1645 	 * we would be nearly out of memory.
1646 	 */
1647 	if (vm_paging_needed())
1648 		pagedaemon_wakeup();
1649 
1650 	return (m);
1651 }
1652 
1653 static void
1654 vm_page_alloc_contig_vdrop(struct spglist *lst)
1655 {
1656 
1657 	while (!SLIST_EMPTY(lst)) {
1658 		vdrop((struct vnode *)SLIST_FIRST(lst)-> plinks.s.pv);
1659 		SLIST_REMOVE_HEAD(lst, plinks.s.ss);
1660 	}
1661 }
1662 
1663 /*
1664  *	vm_page_alloc_contig:
1665  *
1666  *	Allocate a contiguous set of physical pages of the given size "npages"
1667  *	from the free lists.  All of the physical pages must be at or above
1668  *	the given physical address "low" and below the given physical address
1669  *	"high".  The given value "alignment" determines the alignment of the
1670  *	first physical page in the set.  If the given value "boundary" is
1671  *	non-zero, then the set of physical pages cannot cross any physical
1672  *	address boundary that is a multiple of that value.  Both "alignment"
1673  *	and "boundary" must be a power of two.
1674  *
1675  *	If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
1676  *	then the memory attribute setting for the physical pages is configured
1677  *	to the object's memory attribute setting.  Otherwise, the memory
1678  *	attribute setting for the physical pages is configured to "memattr",
1679  *	overriding the object's memory attribute setting.  However, if the
1680  *	object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
1681  *	memory attribute setting for the physical pages cannot be configured
1682  *	to VM_MEMATTR_DEFAULT.
1683  *
1684  *	The caller must always specify an allocation class.
1685  *
1686  *	allocation classes:
1687  *	VM_ALLOC_NORMAL		normal process request
1688  *	VM_ALLOC_SYSTEM		system *really* needs a page
1689  *	VM_ALLOC_INTERRUPT	interrupt time request
1690  *
1691  *	optional allocation flags:
1692  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
1693  *	VM_ALLOC_NOOBJ		page is not associated with an object and
1694  *				should not be exclusive busy
1695  *	VM_ALLOC_SBUSY		shared busy the allocated page
1696  *	VM_ALLOC_WIRED		wire the allocated page
1697  *	VM_ALLOC_ZERO		prefer a zeroed page
1698  *
1699  *	This routine may not sleep.
1700  */
1701 vm_page_t
1702 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
1703     u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
1704     vm_paddr_t boundary, vm_memattr_t memattr)
1705 {
1706 	struct vnode *drop;
1707 	struct spglist deferred_vdrop_list;
1708 	vm_page_t m, m_tmp, m_ret;
1709 	u_int flags, oflags;
1710 	int req_class;
1711 
1712 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1713 	    (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1714 	    ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1715 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1716 	    ("vm_page_alloc: inconsistent object(%p)/req(%x)", (void *)object,
1717 	    req));
1718 	if (object != NULL) {
1719 		VM_OBJECT_ASSERT_WLOCKED(object);
1720 		KASSERT(object->type == OBJT_PHYS,
1721 		    ("vm_page_alloc_contig: object %p isn't OBJT_PHYS",
1722 		    object));
1723 	}
1724 	KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
1725 	req_class = req & VM_ALLOC_CLASS_MASK;
1726 
1727 	/*
1728 	 * The page daemon is allowed to dig deeper into the free page list.
1729 	 */
1730 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1731 		req_class = VM_ALLOC_SYSTEM;
1732 
1733 	SLIST_INIT(&deferred_vdrop_list);
1734 	mtx_lock(&vm_page_queue_free_mtx);
1735 	if (cnt.v_free_count + cnt.v_cache_count >= npages +
1736 	    cnt.v_free_reserved || (req_class == VM_ALLOC_SYSTEM &&
1737 	    cnt.v_free_count + cnt.v_cache_count >= npages +
1738 	    cnt.v_interrupt_free_min) || (req_class == VM_ALLOC_INTERRUPT &&
1739 	    cnt.v_free_count + cnt.v_cache_count >= npages)) {
1740 #if VM_NRESERVLEVEL > 0
1741 retry:
1742 		if (object == NULL || (object->flags & OBJ_COLORED) == 0 ||
1743 		    (m_ret = vm_reserv_alloc_contig(object, pindex, npages,
1744 		    low, high, alignment, boundary)) == NULL)
1745 #endif
1746 			m_ret = vm_phys_alloc_contig(npages, low, high,
1747 			    alignment, boundary);
1748 	} else {
1749 		mtx_unlock(&vm_page_queue_free_mtx);
1750 		atomic_add_int(&vm_pageout_deficit, npages);
1751 		pagedaemon_wakeup();
1752 		return (NULL);
1753 	}
1754 	if (m_ret != NULL)
1755 		for (m = m_ret; m < &m_ret[npages]; m++) {
1756 			drop = vm_page_alloc_init(m);
1757 			if (drop != NULL) {
1758 				/*
1759 				 * Enqueue the vnode for deferred vdrop().
1760 				 */
1761 				m->plinks.s.pv = drop;
1762 				SLIST_INSERT_HEAD(&deferred_vdrop_list, m,
1763 				    plinks.s.ss);
1764 			}
1765 		}
1766 	else {
1767 #if VM_NRESERVLEVEL > 0
1768 		if (vm_reserv_reclaim_contig(npages, low, high, alignment,
1769 		    boundary))
1770 			goto retry;
1771 #endif
1772 	}
1773 	mtx_unlock(&vm_page_queue_free_mtx);
1774 	if (m_ret == NULL)
1775 		return (NULL);
1776 
1777 	/*
1778 	 * Initialize the pages.  Only the PG_ZERO flag is inherited.
1779 	 */
1780 	flags = 0;
1781 	if ((req & VM_ALLOC_ZERO) != 0)
1782 		flags = PG_ZERO;
1783 	if ((req & VM_ALLOC_NODUMP) != 0)
1784 		flags |= PG_NODUMP;
1785 	if ((req & VM_ALLOC_WIRED) != 0)
1786 		atomic_add_int(&cnt.v_wire_count, npages);
1787 	oflags = VPO_UNMANAGED;
1788 	if (object != NULL) {
1789 		if (object->memattr != VM_MEMATTR_DEFAULT &&
1790 		    memattr == VM_MEMATTR_DEFAULT)
1791 			memattr = object->memattr;
1792 	}
1793 	for (m = m_ret; m < &m_ret[npages]; m++) {
1794 		m->aflags = 0;
1795 		m->flags = (m->flags | PG_NODUMP) & flags;
1796 		m->busy_lock = VPB_UNBUSIED;
1797 		if (object != NULL) {
1798 			if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
1799 				m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1800 			if ((req & VM_ALLOC_SBUSY) != 0)
1801 				m->busy_lock = VPB_SHARERS_WORD(1);
1802 		}
1803 		if ((req & VM_ALLOC_WIRED) != 0)
1804 			m->wire_count = 1;
1805 		/* Unmanaged pages don't use "act_count". */
1806 		m->oflags = oflags;
1807 		if (object != NULL) {
1808 			if (vm_page_insert(m, object, pindex)) {
1809 				vm_page_alloc_contig_vdrop(
1810 				    &deferred_vdrop_list);
1811 				if (vm_paging_needed())
1812 					pagedaemon_wakeup();
1813 				if ((req & VM_ALLOC_WIRED) != 0)
1814 					atomic_subtract_int(&cnt.v_wire_count,
1815 					    npages);
1816 				for (m_tmp = m, m = m_ret;
1817 				    m < &m_ret[npages]; m++) {
1818 					if ((req & VM_ALLOC_WIRED) != 0)
1819 						m->wire_count = 0;
1820 					if (m >= m_tmp)
1821 						m->object = NULL;
1822 					vm_page_free(m);
1823 				}
1824 				return (NULL);
1825 			}
1826 		} else
1827 			m->pindex = pindex;
1828 		if (memattr != VM_MEMATTR_DEFAULT)
1829 			pmap_page_set_memattr(m, memattr);
1830 		pindex++;
1831 	}
1832 	vm_page_alloc_contig_vdrop(&deferred_vdrop_list);
1833 	if (vm_paging_needed())
1834 		pagedaemon_wakeup();
1835 	return (m_ret);
1836 }
1837 
1838 /*
1839  * Initialize a page that has been freshly dequeued from a freelist.
1840  * The caller has to drop the vnode returned, if it is not NULL.
1841  *
1842  * This function may only be used to initialize unmanaged pages.
1843  *
1844  * To be called with vm_page_queue_free_mtx held.
1845  */
1846 static struct vnode *
1847 vm_page_alloc_init(vm_page_t m)
1848 {
1849 	struct vnode *drop;
1850 	vm_object_t m_object;
1851 
1852 	KASSERT(m->queue == PQ_NONE,
1853 	    ("vm_page_alloc_init: page %p has unexpected queue %d",
1854 	    m, m->queue));
1855 	KASSERT(m->wire_count == 0,
1856 	    ("vm_page_alloc_init: page %p is wired", m));
1857 	KASSERT(m->hold_count == 0,
1858 	    ("vm_page_alloc_init: page %p is held", m));
1859 	KASSERT(!vm_page_sbusied(m),
1860 	    ("vm_page_alloc_init: page %p is busy", m));
1861 	KASSERT(m->dirty == 0,
1862 	    ("vm_page_alloc_init: page %p is dirty", m));
1863 	KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
1864 	    ("vm_page_alloc_init: page %p has unexpected memattr %d",
1865 	    m, pmap_page_get_memattr(m)));
1866 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1867 	drop = NULL;
1868 	if ((m->flags & PG_CACHED) != 0) {
1869 		KASSERT((m->flags & PG_ZERO) == 0,
1870 		    ("vm_page_alloc_init: cached page %p is PG_ZERO", m));
1871 		m->valid = 0;
1872 		m_object = m->object;
1873 		vm_page_cache_remove(m);
1874 		if (m_object->type == OBJT_VNODE &&
1875 		    vm_object_cache_is_empty(m_object))
1876 			drop = m_object->handle;
1877 	} else {
1878 		KASSERT(VM_PAGE_IS_FREE(m),
1879 		    ("vm_page_alloc_init: page %p is not free", m));
1880 		KASSERT(m->valid == 0,
1881 		    ("vm_page_alloc_init: free page %p is valid", m));
1882 		vm_phys_freecnt_adj(m, -1);
1883 		if ((m->flags & PG_ZERO) != 0)
1884 			vm_page_zero_count--;
1885 	}
1886 	/* Don't clear the PG_ZERO flag; we'll need it later. */
1887 	m->flags &= PG_ZERO;
1888 	return (drop);
1889 }
1890 
1891 /*
1892  * 	vm_page_alloc_freelist:
1893  *
1894  *	Allocate a physical page from the specified free page list.
1895  *
1896  *	The caller must always specify an allocation class.
1897  *
1898  *	allocation classes:
1899  *	VM_ALLOC_NORMAL		normal process request
1900  *	VM_ALLOC_SYSTEM		system *really* needs a page
1901  *	VM_ALLOC_INTERRUPT	interrupt time request
1902  *
1903  *	optional allocation flags:
1904  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
1905  *				intends to allocate
1906  *	VM_ALLOC_WIRED		wire the allocated page
1907  *	VM_ALLOC_ZERO		prefer a zeroed page
1908  *
1909  *	This routine may not sleep.
1910  */
1911 vm_page_t
1912 vm_page_alloc_freelist(int flind, int req)
1913 {
1914 	struct vnode *drop;
1915 	vm_page_t m;
1916 	u_int flags;
1917 	int req_class;
1918 
1919 	req_class = req & VM_ALLOC_CLASS_MASK;
1920 
1921 	/*
1922 	 * The page daemon is allowed to dig deeper into the free page list.
1923 	 */
1924 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1925 		req_class = VM_ALLOC_SYSTEM;
1926 
1927 	/*
1928 	 * Do not allocate reserved pages unless the req has asked for it.
1929 	 */
1930 	mtx_lock_flags(&vm_page_queue_free_mtx, MTX_RECURSE);
1931 	if (cnt.v_free_count + cnt.v_cache_count > cnt.v_free_reserved ||
1932 	    (req_class == VM_ALLOC_SYSTEM &&
1933 	    cnt.v_free_count + cnt.v_cache_count > cnt.v_interrupt_free_min) ||
1934 	    (req_class == VM_ALLOC_INTERRUPT &&
1935 	    cnt.v_free_count + cnt.v_cache_count > 0))
1936 		m = vm_phys_alloc_freelist_pages(flind, VM_FREEPOOL_DIRECT, 0);
1937 	else {
1938 		mtx_unlock(&vm_page_queue_free_mtx);
1939 		atomic_add_int(&vm_pageout_deficit,
1940 		    max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
1941 		pagedaemon_wakeup();
1942 		return (NULL);
1943 	}
1944 	if (m == NULL) {
1945 		mtx_unlock(&vm_page_queue_free_mtx);
1946 		return (NULL);
1947 	}
1948 	drop = vm_page_alloc_init(m);
1949 	mtx_unlock(&vm_page_queue_free_mtx);
1950 
1951 	/*
1952 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
1953 	 */
1954 	m->aflags = 0;
1955 	flags = 0;
1956 	if ((req & VM_ALLOC_ZERO) != 0)
1957 		flags = PG_ZERO;
1958 	m->flags &= flags;
1959 	if ((req & VM_ALLOC_WIRED) != 0) {
1960 		/*
1961 		 * The page lock is not required for wiring a page that does
1962 		 * not belong to an object.
1963 		 */
1964 		atomic_add_int(&cnt.v_wire_count, 1);
1965 		m->wire_count = 1;
1966 	}
1967 	/* Unmanaged pages don't use "act_count". */
1968 	m->oflags = VPO_UNMANAGED;
1969 	if (drop != NULL)
1970 		vdrop(drop);
1971 	if (vm_paging_needed())
1972 		pagedaemon_wakeup();
1973 	return (m);
1974 }
1975 
1976 /*
1977  *	vm_wait:	(also see VM_WAIT macro)
1978  *
1979  *	Sleep until free pages are available for allocation.
1980  *	- Called in various places before memory allocations.
1981  */
1982 void
1983 vm_wait(void)
1984 {
1985 
1986 	mtx_lock(&vm_page_queue_free_mtx);
1987 	if (curproc == pageproc) {
1988 		vm_pageout_pages_needed = 1;
1989 		msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx,
1990 		    PDROP | PSWP, "VMWait", 0);
1991 	} else {
1992 		if (!vm_pages_needed) {
1993 			vm_pages_needed = 1;
1994 			wakeup(&vm_pages_needed);
1995 		}
1996 		msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PVM,
1997 		    "vmwait", 0);
1998 	}
1999 }
2000 
2001 /*
2002  *	vm_waitpfault:	(also see VM_WAITPFAULT macro)
2003  *
2004  *	Sleep until free pages are available for allocation.
2005  *	- Called only in vm_fault so that processes page faulting
2006  *	  can be easily tracked.
2007  *	- Sleeps at a lower priority than vm_wait() so that vm_wait()ing
2008  *	  processes will be able to grab memory first.  Do not change
2009  *	  this balance without careful testing first.
2010  */
2011 void
2012 vm_waitpfault(void)
2013 {
2014 
2015 	mtx_lock(&vm_page_queue_free_mtx);
2016 	if (!vm_pages_needed) {
2017 		vm_pages_needed = 1;
2018 		wakeup(&vm_pages_needed);
2019 	}
2020 	msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PUSER,
2021 	    "pfault", 0);
2022 }
2023 
2024 struct vm_pagequeue *
2025 vm_page_pagequeue(vm_page_t m)
2026 {
2027 
2028 	return (&vm_phys_domain(m)->vmd_pagequeues[m->queue]);
2029 }
2030 
2031 /*
2032  *	vm_page_dequeue:
2033  *
2034  *	Remove the given page from its current page queue.
2035  *
2036  *	The page must be locked.
2037  */
2038 void
2039 vm_page_dequeue(vm_page_t m)
2040 {
2041 	struct vm_pagequeue *pq;
2042 
2043 	vm_page_lock_assert(m, MA_OWNED);
2044 	KASSERT(m->queue != PQ_NONE,
2045 	    ("vm_page_dequeue: page %p is not queued", m));
2046 	pq = vm_page_pagequeue(m);
2047 	vm_pagequeue_lock(pq);
2048 	m->queue = PQ_NONE;
2049 	TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2050 	vm_pagequeue_cnt_dec(pq);
2051 	vm_pagequeue_unlock(pq);
2052 }
2053 
2054 /*
2055  *	vm_page_dequeue_locked:
2056  *
2057  *	Remove the given page from its current page queue.
2058  *
2059  *	The page and page queue must be locked.
2060  */
2061 void
2062 vm_page_dequeue_locked(vm_page_t m)
2063 {
2064 	struct vm_pagequeue *pq;
2065 
2066 	vm_page_lock_assert(m, MA_OWNED);
2067 	pq = vm_page_pagequeue(m);
2068 	vm_pagequeue_assert_locked(pq);
2069 	m->queue = PQ_NONE;
2070 	TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2071 	vm_pagequeue_cnt_dec(pq);
2072 }
2073 
2074 /*
2075  *	vm_page_enqueue:
2076  *
2077  *	Add the given page to the specified page queue.
2078  *
2079  *	The page must be locked.
2080  */
2081 static void
2082 vm_page_enqueue(int queue, vm_page_t m)
2083 {
2084 	struct vm_pagequeue *pq;
2085 
2086 	vm_page_lock_assert(m, MA_OWNED);
2087 	pq = &vm_phys_domain(m)->vmd_pagequeues[queue];
2088 	vm_pagequeue_lock(pq);
2089 	m->queue = queue;
2090 	TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2091 	vm_pagequeue_cnt_inc(pq);
2092 	vm_pagequeue_unlock(pq);
2093 }
2094 
2095 /*
2096  *	vm_page_requeue:
2097  *
2098  *	Move the given page to the tail of its current page queue.
2099  *
2100  *	The page must be locked.
2101  */
2102 void
2103 vm_page_requeue(vm_page_t m)
2104 {
2105 	struct vm_pagequeue *pq;
2106 
2107 	vm_page_lock_assert(m, MA_OWNED);
2108 	KASSERT(m->queue != PQ_NONE,
2109 	    ("vm_page_requeue: page %p is not queued", m));
2110 	pq = vm_page_pagequeue(m);
2111 	vm_pagequeue_lock(pq);
2112 	TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2113 	TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2114 	vm_pagequeue_unlock(pq);
2115 }
2116 
2117 /*
2118  *	vm_page_requeue_locked:
2119  *
2120  *	Move the given page to the tail of its current page queue.
2121  *
2122  *	The page queue must be locked.
2123  */
2124 void
2125 vm_page_requeue_locked(vm_page_t m)
2126 {
2127 	struct vm_pagequeue *pq;
2128 
2129 	KASSERT(m->queue != PQ_NONE,
2130 	    ("vm_page_requeue_locked: page %p is not queued", m));
2131 	pq = vm_page_pagequeue(m);
2132 	vm_pagequeue_assert_locked(pq);
2133 	TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2134 	TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2135 }
2136 
2137 /*
2138  *	vm_page_activate:
2139  *
2140  *	Put the specified page on the active list (if appropriate).
2141  *	Ensure that act_count is at least ACT_INIT but do not otherwise
2142  *	mess with it.
2143  *
2144  *	The page must be locked.
2145  */
2146 void
2147 vm_page_activate(vm_page_t m)
2148 {
2149 	int queue;
2150 
2151 	vm_page_lock_assert(m, MA_OWNED);
2152 	if ((queue = m->queue) != PQ_ACTIVE) {
2153 		if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
2154 			if (m->act_count < ACT_INIT)
2155 				m->act_count = ACT_INIT;
2156 			if (queue != PQ_NONE)
2157 				vm_page_dequeue(m);
2158 			vm_page_enqueue(PQ_ACTIVE, m);
2159 		} else
2160 			KASSERT(queue == PQ_NONE,
2161 			    ("vm_page_activate: wired page %p is queued", m));
2162 	} else {
2163 		if (m->act_count < ACT_INIT)
2164 			m->act_count = ACT_INIT;
2165 	}
2166 }
2167 
2168 /*
2169  *	vm_page_free_wakeup:
2170  *
2171  *	Helper routine for vm_page_free_toq() and vm_page_cache().  This
2172  *	routine is called when a page has been added to the cache or free
2173  *	queues.
2174  *
2175  *	The page queues must be locked.
2176  */
2177 static inline void
2178 vm_page_free_wakeup(void)
2179 {
2180 
2181 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
2182 	/*
2183 	 * if pageout daemon needs pages, then tell it that there are
2184 	 * some free.
2185 	 */
2186 	if (vm_pageout_pages_needed &&
2187 	    cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) {
2188 		wakeup(&vm_pageout_pages_needed);
2189 		vm_pageout_pages_needed = 0;
2190 	}
2191 	/*
2192 	 * wakeup processes that are waiting on memory if we hit a
2193 	 * high water mark. And wakeup scheduler process if we have
2194 	 * lots of memory. this process will swapin processes.
2195 	 */
2196 	if (vm_pages_needed && !vm_page_count_min()) {
2197 		vm_pages_needed = 0;
2198 		wakeup(&cnt.v_free_count);
2199 	}
2200 }
2201 
2202 /*
2203  *	Turn a cached page into a free page, by changing its attributes.
2204  *	Keep the statistics up-to-date.
2205  *
2206  *	The free page queue must be locked.
2207  */
2208 static void
2209 vm_page_cache_turn_free(vm_page_t m)
2210 {
2211 
2212 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
2213 
2214 	m->object = NULL;
2215 	m->valid = 0;
2216 	/* Clear PG_CACHED and set PG_FREE. */
2217 	m->flags ^= PG_CACHED | PG_FREE;
2218 	KASSERT((m->flags & (PG_CACHED | PG_FREE)) == PG_FREE,
2219 	    ("vm_page_cache_free: page %p has inconsistent flags", m));
2220 	cnt.v_cache_count--;
2221 	vm_phys_freecnt_adj(m, 1);
2222 }
2223 
2224 /*
2225  *	vm_page_free_toq:
2226  *
2227  *	Returns the given page to the free list,
2228  *	disassociating it with any VM object.
2229  *
2230  *	The object must be locked.  The page must be locked if it is managed.
2231  */
2232 void
2233 vm_page_free_toq(vm_page_t m)
2234 {
2235 
2236 	if ((m->oflags & VPO_UNMANAGED) == 0) {
2237 		vm_page_lock_assert(m, MA_OWNED);
2238 		KASSERT(!pmap_page_is_mapped(m),
2239 		    ("vm_page_free_toq: freeing mapped page %p", m));
2240 	} else
2241 		KASSERT(m->queue == PQ_NONE,
2242 		    ("vm_page_free_toq: unmanaged page %p is queued", m));
2243 	PCPU_INC(cnt.v_tfree);
2244 
2245 	if (VM_PAGE_IS_FREE(m))
2246 		panic("vm_page_free: freeing free page %p", m);
2247 	else if (vm_page_sbusied(m))
2248 		panic("vm_page_free: freeing busy page %p", m);
2249 
2250 	/*
2251 	 * Unqueue, then remove page.  Note that we cannot destroy
2252 	 * the page here because we do not want to call the pager's
2253 	 * callback routine until after we've put the page on the
2254 	 * appropriate free queue.
2255 	 */
2256 	vm_page_remque(m);
2257 	vm_page_remove(m);
2258 
2259 	/*
2260 	 * If fictitious remove object association and
2261 	 * return, otherwise delay object association removal.
2262 	 */
2263 	if ((m->flags & PG_FICTITIOUS) != 0) {
2264 		return;
2265 	}
2266 
2267 	m->valid = 0;
2268 	vm_page_undirty(m);
2269 
2270 	if (m->wire_count != 0)
2271 		panic("vm_page_free: freeing wired page %p", m);
2272 	if (m->hold_count != 0) {
2273 		m->flags &= ~PG_ZERO;
2274 		KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2275 		    ("vm_page_free: freeing PG_UNHOLDFREE page %p", m));
2276 		m->flags |= PG_UNHOLDFREE;
2277 	} else {
2278 		/*
2279 		 * Restore the default memory attribute to the page.
2280 		 */
2281 		if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
2282 			pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
2283 
2284 		/*
2285 		 * Insert the page into the physical memory allocator's
2286 		 * cache/free page queues.
2287 		 */
2288 		mtx_lock(&vm_page_queue_free_mtx);
2289 		m->flags |= PG_FREE;
2290 		vm_phys_freecnt_adj(m, 1);
2291 #if VM_NRESERVLEVEL > 0
2292 		if (!vm_reserv_free_page(m))
2293 #else
2294 		if (TRUE)
2295 #endif
2296 			vm_phys_free_pages(m, 0);
2297 		if ((m->flags & PG_ZERO) != 0)
2298 			++vm_page_zero_count;
2299 		else
2300 			vm_page_zero_idle_wakeup();
2301 		vm_page_free_wakeup();
2302 		mtx_unlock(&vm_page_queue_free_mtx);
2303 	}
2304 }
2305 
2306 /*
2307  *	vm_page_wire:
2308  *
2309  *	Mark this page as wired down by yet
2310  *	another map, removing it from paging queues
2311  *	as necessary.
2312  *
2313  *	If the page is fictitious, then its wire count must remain one.
2314  *
2315  *	The page must be locked.
2316  */
2317 void
2318 vm_page_wire(vm_page_t m)
2319 {
2320 
2321 	/*
2322 	 * Only bump the wire statistics if the page is not already wired,
2323 	 * and only unqueue the page if it is on some queue (if it is unmanaged
2324 	 * it is already off the queues).
2325 	 */
2326 	vm_page_lock_assert(m, MA_OWNED);
2327 	if ((m->flags & PG_FICTITIOUS) != 0) {
2328 		KASSERT(m->wire_count == 1,
2329 		    ("vm_page_wire: fictitious page %p's wire count isn't one",
2330 		    m));
2331 		return;
2332 	}
2333 	if (m->wire_count == 0) {
2334 		KASSERT((m->oflags & VPO_UNMANAGED) == 0 ||
2335 		    m->queue == PQ_NONE,
2336 		    ("vm_page_wire: unmanaged page %p is queued", m));
2337 		vm_page_remque(m);
2338 		atomic_add_int(&cnt.v_wire_count, 1);
2339 	}
2340 	m->wire_count++;
2341 	KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m));
2342 }
2343 
2344 /*
2345  * vm_page_unwire:
2346  *
2347  * Release one wiring of the specified page, potentially enabling it to be
2348  * paged again.  If paging is enabled, then the value of the parameter
2349  * "activate" determines to which queue the page is added.  If "activate" is
2350  * non-zero, then the page is added to the active queue.  Otherwise, it is
2351  * added to the inactive queue.
2352  *
2353  * However, unless the page belongs to an object, it is not enqueued because
2354  * it cannot be paged out.
2355  *
2356  * If a page is fictitious, then its wire count must always be one.
2357  *
2358  * A managed page must be locked.
2359  */
2360 void
2361 vm_page_unwire(vm_page_t m, int activate)
2362 {
2363 
2364 	if ((m->oflags & VPO_UNMANAGED) == 0)
2365 		vm_page_lock_assert(m, MA_OWNED);
2366 	if ((m->flags & PG_FICTITIOUS) != 0) {
2367 		KASSERT(m->wire_count == 1,
2368 	    ("vm_page_unwire: fictitious page %p's wire count isn't one", m));
2369 		return;
2370 	}
2371 	if (m->wire_count > 0) {
2372 		m->wire_count--;
2373 		if (m->wire_count == 0) {
2374 			atomic_subtract_int(&cnt.v_wire_count, 1);
2375 			if ((m->oflags & VPO_UNMANAGED) != 0 ||
2376 			    m->object == NULL)
2377 				return;
2378 			if (!activate)
2379 				m->flags &= ~PG_WINATCFLS;
2380 			vm_page_enqueue(activate ? PQ_ACTIVE : PQ_INACTIVE, m);
2381 		}
2382 	} else
2383 		panic("vm_page_unwire: page %p's wire count is zero", m);
2384 }
2385 
2386 /*
2387  * Move the specified page to the inactive queue.
2388  *
2389  * Many pages placed on the inactive queue should actually go
2390  * into the cache, but it is difficult to figure out which.  What
2391  * we do instead, if the inactive target is well met, is to put
2392  * clean pages at the head of the inactive queue instead of the tail.
2393  * This will cause them to be moved to the cache more quickly and
2394  * if not actively re-referenced, reclaimed more quickly.  If we just
2395  * stick these pages at the end of the inactive queue, heavy filesystem
2396  * meta-data accesses can cause an unnecessary paging load on memory bound
2397  * processes.  This optimization causes one-time-use metadata to be
2398  * reused more quickly.
2399  *
2400  * Normally athead is 0 resulting in LRU operation.  athead is set
2401  * to 1 if we want this page to be 'as if it were placed in the cache',
2402  * except without unmapping it from the process address space.
2403  *
2404  * The page must be locked.
2405  */
2406 static inline void
2407 _vm_page_deactivate(vm_page_t m, int athead)
2408 {
2409 	struct vm_pagequeue *pq;
2410 	int queue;
2411 
2412 	vm_page_lock_assert(m, MA_OWNED);
2413 
2414 	/*
2415 	 * Ignore if already inactive.
2416 	 */
2417 	if ((queue = m->queue) == PQ_INACTIVE)
2418 		return;
2419 	if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
2420 		if (queue != PQ_NONE)
2421 			vm_page_dequeue(m);
2422 		m->flags &= ~PG_WINATCFLS;
2423 		pq = &vm_phys_domain(m)->vmd_pagequeues[PQ_INACTIVE];
2424 		vm_pagequeue_lock(pq);
2425 		m->queue = PQ_INACTIVE;
2426 		if (athead)
2427 			TAILQ_INSERT_HEAD(&pq->pq_pl, m, plinks.q);
2428 		else
2429 			TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2430 		vm_pagequeue_cnt_inc(pq);
2431 		vm_pagequeue_unlock(pq);
2432 	}
2433 }
2434 
2435 /*
2436  * Move the specified page to the inactive queue.
2437  *
2438  * The page must be locked.
2439  */
2440 void
2441 vm_page_deactivate(vm_page_t m)
2442 {
2443 
2444 	_vm_page_deactivate(m, 0);
2445 }
2446 
2447 /*
2448  * vm_page_try_to_cache:
2449  *
2450  * Returns 0 on failure, 1 on success
2451  */
2452 int
2453 vm_page_try_to_cache(vm_page_t m)
2454 {
2455 
2456 	vm_page_lock_assert(m, MA_OWNED);
2457 	VM_OBJECT_ASSERT_WLOCKED(m->object);
2458 	if (m->dirty || m->hold_count || m->wire_count ||
2459 	    (m->oflags & VPO_UNMANAGED) != 0 || vm_page_busied(m))
2460 		return (0);
2461 	pmap_remove_all(m);
2462 	if (m->dirty)
2463 		return (0);
2464 	vm_page_cache(m);
2465 	return (1);
2466 }
2467 
2468 /*
2469  * vm_page_try_to_free()
2470  *
2471  *	Attempt to free the page.  If we cannot free it, we do nothing.
2472  *	1 is returned on success, 0 on failure.
2473  */
2474 int
2475 vm_page_try_to_free(vm_page_t m)
2476 {
2477 
2478 	vm_page_lock_assert(m, MA_OWNED);
2479 	if (m->object != NULL)
2480 		VM_OBJECT_ASSERT_WLOCKED(m->object);
2481 	if (m->dirty || m->hold_count || m->wire_count ||
2482 	    (m->oflags & VPO_UNMANAGED) != 0 || vm_page_busied(m))
2483 		return (0);
2484 	pmap_remove_all(m);
2485 	if (m->dirty)
2486 		return (0);
2487 	vm_page_free(m);
2488 	return (1);
2489 }
2490 
2491 /*
2492  * vm_page_cache
2493  *
2494  * Put the specified page onto the page cache queue (if appropriate).
2495  *
2496  * The object and page must be locked.
2497  */
2498 void
2499 vm_page_cache(vm_page_t m)
2500 {
2501 	vm_object_t object;
2502 	boolean_t cache_was_empty;
2503 
2504 	vm_page_lock_assert(m, MA_OWNED);
2505 	object = m->object;
2506 	VM_OBJECT_ASSERT_WLOCKED(object);
2507 	if (vm_page_busied(m) || (m->oflags & VPO_UNMANAGED) ||
2508 	    m->hold_count || m->wire_count)
2509 		panic("vm_page_cache: attempting to cache busy page");
2510 	KASSERT(!pmap_page_is_mapped(m),
2511 	    ("vm_page_cache: page %p is mapped", m));
2512 	KASSERT(m->dirty == 0, ("vm_page_cache: page %p is dirty", m));
2513 	if (m->valid == 0 || object->type == OBJT_DEFAULT ||
2514 	    (object->type == OBJT_SWAP &&
2515 	    !vm_pager_has_page(object, m->pindex, NULL, NULL))) {
2516 		/*
2517 		 * Hypothesis: A cache-elgible page belonging to a
2518 		 * default object or swap object but without a backing
2519 		 * store must be zero filled.
2520 		 */
2521 		vm_page_free(m);
2522 		return;
2523 	}
2524 	KASSERT((m->flags & PG_CACHED) == 0,
2525 	    ("vm_page_cache: page %p is already cached", m));
2526 
2527 	/*
2528 	 * Remove the page from the paging queues.
2529 	 */
2530 	vm_page_remque(m);
2531 
2532 	/*
2533 	 * Remove the page from the object's collection of resident
2534 	 * pages.
2535 	 */
2536 	vm_radix_remove(&object->rtree, m->pindex);
2537 	TAILQ_REMOVE(&object->memq, m, listq);
2538 	object->resident_page_count--;
2539 
2540 	/*
2541 	 * Restore the default memory attribute to the page.
2542 	 */
2543 	if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
2544 		pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
2545 
2546 	/*
2547 	 * Insert the page into the object's collection of cached pages
2548 	 * and the physical memory allocator's cache/free page queues.
2549 	 */
2550 	m->flags &= ~PG_ZERO;
2551 	mtx_lock(&vm_page_queue_free_mtx);
2552 	cache_was_empty = vm_radix_is_empty(&object->cache);
2553 	if (vm_radix_insert(&object->cache, m)) {
2554 		mtx_unlock(&vm_page_queue_free_mtx);
2555 		if (object->resident_page_count == 0)
2556 			vdrop(object->handle);
2557 		m->object = NULL;
2558 		vm_page_free(m);
2559 		return;
2560 	}
2561 
2562 	/*
2563 	 * The above call to vm_radix_insert() could reclaim the one pre-
2564 	 * existing cached page from this object, resulting in a call to
2565 	 * vdrop().
2566 	 */
2567 	if (!cache_was_empty)
2568 		cache_was_empty = vm_radix_is_singleton(&object->cache);
2569 
2570 	m->flags |= PG_CACHED;
2571 	cnt.v_cache_count++;
2572 	PCPU_INC(cnt.v_tcached);
2573 #if VM_NRESERVLEVEL > 0
2574 	if (!vm_reserv_free_page(m)) {
2575 #else
2576 	if (TRUE) {
2577 #endif
2578 		vm_phys_set_pool(VM_FREEPOOL_CACHE, m, 0);
2579 		vm_phys_free_pages(m, 0);
2580 	}
2581 	vm_page_free_wakeup();
2582 	mtx_unlock(&vm_page_queue_free_mtx);
2583 
2584 	/*
2585 	 * Increment the vnode's hold count if this is the object's only
2586 	 * cached page.  Decrement the vnode's hold count if this was
2587 	 * the object's only resident page.
2588 	 */
2589 	if (object->type == OBJT_VNODE) {
2590 		if (cache_was_empty && object->resident_page_count != 0)
2591 			vhold(object->handle);
2592 		else if (!cache_was_empty && object->resident_page_count == 0)
2593 			vdrop(object->handle);
2594 	}
2595 }
2596 
2597 /*
2598  * vm_page_advise
2599  *
2600  *	Cache, deactivate, or do nothing as appropriate.  This routine
2601  *	is used by madvise().
2602  *
2603  *	Generally speaking we want to move the page into the cache so
2604  *	it gets reused quickly.  However, this can result in a silly syndrome
2605  *	due to the page recycling too quickly.  Small objects will not be
2606  *	fully cached.  On the other hand, if we move the page to the inactive
2607  *	queue we wind up with a problem whereby very large objects
2608  *	unnecessarily blow away our inactive and cache queues.
2609  *
2610  *	The solution is to move the pages based on a fixed weighting.  We
2611  *	either leave them alone, deactivate them, or move them to the cache,
2612  *	where moving them to the cache has the highest weighting.
2613  *	By forcing some pages into other queues we eventually force the
2614  *	system to balance the queues, potentially recovering other unrelated
2615  *	space from active.  The idea is to not force this to happen too
2616  *	often.
2617  *
2618  *	The object and page must be locked.
2619  */
2620 void
2621 vm_page_advise(vm_page_t m, int advice)
2622 {
2623 	int dnw, head;
2624 
2625 	vm_page_assert_locked(m);
2626 	VM_OBJECT_ASSERT_WLOCKED(m->object);
2627 	if (advice == MADV_FREE) {
2628 		/*
2629 		 * Mark the page clean.  This will allow the page to be freed
2630 		 * up by the system.  However, such pages are often reused
2631 		 * quickly by malloc() so we do not do anything that would
2632 		 * cause a page fault if we can help it.
2633 		 *
2634 		 * Specifically, we do not try to actually free the page now
2635 		 * nor do we try to put it in the cache (which would cause a
2636 		 * page fault on reuse).
2637 		 *
2638 		 * But we do make the page is freeable as we can without
2639 		 * actually taking the step of unmapping it.
2640 		 */
2641 		m->dirty = 0;
2642 		m->act_count = 0;
2643 	} else if (advice != MADV_DONTNEED)
2644 		return;
2645 	dnw = PCPU_GET(dnweight);
2646 	PCPU_INC(dnweight);
2647 
2648 	/*
2649 	 * Occasionally leave the page alone.
2650 	 */
2651 	if ((dnw & 0x01F0) == 0 || m->queue == PQ_INACTIVE) {
2652 		if (m->act_count >= ACT_INIT)
2653 			--m->act_count;
2654 		return;
2655 	}
2656 
2657 	/*
2658 	 * Clear any references to the page.  Otherwise, the page daemon will
2659 	 * immediately reactivate the page.
2660 	 */
2661 	vm_page_aflag_clear(m, PGA_REFERENCED);
2662 
2663 	if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
2664 		vm_page_dirty(m);
2665 
2666 	if (m->dirty || (dnw & 0x0070) == 0) {
2667 		/*
2668 		 * Deactivate the page 3 times out of 32.
2669 		 */
2670 		head = 0;
2671 	} else {
2672 		/*
2673 		 * Cache the page 28 times out of every 32.  Note that
2674 		 * the page is deactivated instead of cached, but placed
2675 		 * at the head of the queue instead of the tail.
2676 		 */
2677 		head = 1;
2678 	}
2679 	_vm_page_deactivate(m, head);
2680 }
2681 
2682 /*
2683  * Grab a page, waiting until we are waken up due to the page
2684  * changing state.  We keep on waiting, if the page continues
2685  * to be in the object.  If the page doesn't exist, first allocate it
2686  * and then conditionally zero it.
2687  *
2688  * This routine may sleep.
2689  *
2690  * The object must be locked on entry.  The lock will, however, be released
2691  * and reacquired if the routine sleeps.
2692  */
2693 vm_page_t
2694 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
2695 {
2696 	vm_page_t m;
2697 	int sleep;
2698 
2699 	VM_OBJECT_ASSERT_WLOCKED(object);
2700 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
2701 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
2702 	    ("vm_page_grab: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
2703 retrylookup:
2704 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
2705 		sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ?
2706 		    vm_page_xbusied(m) : vm_page_busied(m);
2707 		if (sleep) {
2708 			/*
2709 			 * Reference the page before unlocking and
2710 			 * sleeping so that the page daemon is less
2711 			 * likely to reclaim it.
2712 			 */
2713 			vm_page_aflag_set(m, PGA_REFERENCED);
2714 			vm_page_lock(m);
2715 			VM_OBJECT_WUNLOCK(object);
2716 			vm_page_busy_sleep(m, "pgrbwt");
2717 			VM_OBJECT_WLOCK(object);
2718 			goto retrylookup;
2719 		} else {
2720 			if ((allocflags & VM_ALLOC_WIRED) != 0) {
2721 				vm_page_lock(m);
2722 				vm_page_wire(m);
2723 				vm_page_unlock(m);
2724 			}
2725 			if ((allocflags &
2726 			    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
2727 				vm_page_xbusy(m);
2728 			if ((allocflags & VM_ALLOC_SBUSY) != 0)
2729 				vm_page_sbusy(m);
2730 			return (m);
2731 		}
2732 	}
2733 	m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_IGN_SBUSY);
2734 	if (m == NULL) {
2735 		VM_OBJECT_WUNLOCK(object);
2736 		VM_WAIT;
2737 		VM_OBJECT_WLOCK(object);
2738 		goto retrylookup;
2739 	} else if (m->valid != 0)
2740 		return (m);
2741 	if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
2742 		pmap_zero_page(m);
2743 	return (m);
2744 }
2745 
2746 /*
2747  * Mapping function for valid or dirty bits in a page.
2748  *
2749  * Inputs are required to range within a page.
2750  */
2751 vm_page_bits_t
2752 vm_page_bits(int base, int size)
2753 {
2754 	int first_bit;
2755 	int last_bit;
2756 
2757 	KASSERT(
2758 	    base + size <= PAGE_SIZE,
2759 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
2760 	);
2761 
2762 	if (size == 0)		/* handle degenerate case */
2763 		return (0);
2764 
2765 	first_bit = base >> DEV_BSHIFT;
2766 	last_bit = (base + size - 1) >> DEV_BSHIFT;
2767 
2768 	return (((vm_page_bits_t)2 << last_bit) -
2769 	    ((vm_page_bits_t)1 << first_bit));
2770 }
2771 
2772 /*
2773  *	vm_page_set_valid_range:
2774  *
2775  *	Sets portions of a page valid.  The arguments are expected
2776  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
2777  *	of any partial chunks touched by the range.  The invalid portion of
2778  *	such chunks will be zeroed.
2779  *
2780  *	(base + size) must be less then or equal to PAGE_SIZE.
2781  */
2782 void
2783 vm_page_set_valid_range(vm_page_t m, int base, int size)
2784 {
2785 	int endoff, frag;
2786 
2787 	VM_OBJECT_ASSERT_WLOCKED(m->object);
2788 	if (size == 0)	/* handle degenerate case */
2789 		return;
2790 
2791 	/*
2792 	 * If the base is not DEV_BSIZE aligned and the valid
2793 	 * bit is clear, we have to zero out a portion of the
2794 	 * first block.
2795 	 */
2796 	if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
2797 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
2798 		pmap_zero_page_area(m, frag, base - frag);
2799 
2800 	/*
2801 	 * If the ending offset is not DEV_BSIZE aligned and the
2802 	 * valid bit is clear, we have to zero out a portion of
2803 	 * the last block.
2804 	 */
2805 	endoff = base + size;
2806 	if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
2807 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
2808 		pmap_zero_page_area(m, endoff,
2809 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
2810 
2811 	/*
2812 	 * Assert that no previously invalid block that is now being validated
2813 	 * is already dirty.
2814 	 */
2815 	KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
2816 	    ("vm_page_set_valid_range: page %p is dirty", m));
2817 
2818 	/*
2819 	 * Set valid bits inclusive of any overlap.
2820 	 */
2821 	m->valid |= vm_page_bits(base, size);
2822 }
2823 
2824 /*
2825  * Clear the given bits from the specified page's dirty field.
2826  */
2827 static __inline void
2828 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
2829 {
2830 	uintptr_t addr;
2831 #if PAGE_SIZE < 16384
2832 	int shift;
2833 #endif
2834 
2835 	/*
2836 	 * If the object is locked and the page is neither exclusive busy nor
2837 	 * write mapped, then the page's dirty field cannot possibly be
2838 	 * set by a concurrent pmap operation.
2839 	 */
2840 	VM_OBJECT_ASSERT_WLOCKED(m->object);
2841 	if (!vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
2842 		m->dirty &= ~pagebits;
2843 	else {
2844 		/*
2845 		 * The pmap layer can call vm_page_dirty() without
2846 		 * holding a distinguished lock.  The combination of
2847 		 * the object's lock and an atomic operation suffice
2848 		 * to guarantee consistency of the page dirty field.
2849 		 *
2850 		 * For PAGE_SIZE == 32768 case, compiler already
2851 		 * properly aligns the dirty field, so no forcible
2852 		 * alignment is needed. Only require existence of
2853 		 * atomic_clear_64 when page size is 32768.
2854 		 */
2855 		addr = (uintptr_t)&m->dirty;
2856 #if PAGE_SIZE == 32768
2857 		atomic_clear_64((uint64_t *)addr, pagebits);
2858 #elif PAGE_SIZE == 16384
2859 		atomic_clear_32((uint32_t *)addr, pagebits);
2860 #else		/* PAGE_SIZE <= 8192 */
2861 		/*
2862 		 * Use a trick to perform a 32-bit atomic on the
2863 		 * containing aligned word, to not depend on the existence
2864 		 * of atomic_clear_{8, 16}.
2865 		 */
2866 		shift = addr & (sizeof(uint32_t) - 1);
2867 #if BYTE_ORDER == BIG_ENDIAN
2868 		shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY;
2869 #else
2870 		shift *= NBBY;
2871 #endif
2872 		addr &= ~(sizeof(uint32_t) - 1);
2873 		atomic_clear_32((uint32_t *)addr, pagebits << shift);
2874 #endif		/* PAGE_SIZE */
2875 	}
2876 }
2877 
2878 /*
2879  *	vm_page_set_validclean:
2880  *
2881  *	Sets portions of a page valid and clean.  The arguments are expected
2882  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
2883  *	of any partial chunks touched by the range.  The invalid portion of
2884  *	such chunks will be zero'd.
2885  *
2886  *	(base + size) must be less then or equal to PAGE_SIZE.
2887  */
2888 void
2889 vm_page_set_validclean(vm_page_t m, int base, int size)
2890 {
2891 	vm_page_bits_t oldvalid, pagebits;
2892 	int endoff, frag;
2893 
2894 	VM_OBJECT_ASSERT_WLOCKED(m->object);
2895 	if (size == 0)	/* handle degenerate case */
2896 		return;
2897 
2898 	/*
2899 	 * If the base is not DEV_BSIZE aligned and the valid
2900 	 * bit is clear, we have to zero out a portion of the
2901 	 * first block.
2902 	 */
2903 	if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
2904 	    (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
2905 		pmap_zero_page_area(m, frag, base - frag);
2906 
2907 	/*
2908 	 * If the ending offset is not DEV_BSIZE aligned and the
2909 	 * valid bit is clear, we have to zero out a portion of
2910 	 * the last block.
2911 	 */
2912 	endoff = base + size;
2913 	if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
2914 	    (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
2915 		pmap_zero_page_area(m, endoff,
2916 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
2917 
2918 	/*
2919 	 * Set valid, clear dirty bits.  If validating the entire
2920 	 * page we can safely clear the pmap modify bit.  We also
2921 	 * use this opportunity to clear the VPO_NOSYNC flag.  If a process
2922 	 * takes a write fault on a MAP_NOSYNC memory area the flag will
2923 	 * be set again.
2924 	 *
2925 	 * We set valid bits inclusive of any overlap, but we can only
2926 	 * clear dirty bits for DEV_BSIZE chunks that are fully within
2927 	 * the range.
2928 	 */
2929 	oldvalid = m->valid;
2930 	pagebits = vm_page_bits(base, size);
2931 	m->valid |= pagebits;
2932 #if 0	/* NOT YET */
2933 	if ((frag = base & (DEV_BSIZE - 1)) != 0) {
2934 		frag = DEV_BSIZE - frag;
2935 		base += frag;
2936 		size -= frag;
2937 		if (size < 0)
2938 			size = 0;
2939 	}
2940 	pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
2941 #endif
2942 	if (base == 0 && size == PAGE_SIZE) {
2943 		/*
2944 		 * The page can only be modified within the pmap if it is
2945 		 * mapped, and it can only be mapped if it was previously
2946 		 * fully valid.
2947 		 */
2948 		if (oldvalid == VM_PAGE_BITS_ALL)
2949 			/*
2950 			 * Perform the pmap_clear_modify() first.  Otherwise,
2951 			 * a concurrent pmap operation, such as
2952 			 * pmap_protect(), could clear a modification in the
2953 			 * pmap and set the dirty field on the page before
2954 			 * pmap_clear_modify() had begun and after the dirty
2955 			 * field was cleared here.
2956 			 */
2957 			pmap_clear_modify(m);
2958 		m->dirty = 0;
2959 		m->oflags &= ~VPO_NOSYNC;
2960 	} else if (oldvalid != VM_PAGE_BITS_ALL)
2961 		m->dirty &= ~pagebits;
2962 	else
2963 		vm_page_clear_dirty_mask(m, pagebits);
2964 }
2965 
2966 void
2967 vm_page_clear_dirty(vm_page_t m, int base, int size)
2968 {
2969 
2970 	vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
2971 }
2972 
2973 /*
2974  *	vm_page_set_invalid:
2975  *
2976  *	Invalidates DEV_BSIZE'd chunks within a page.  Both the
2977  *	valid and dirty bits for the effected areas are cleared.
2978  */
2979 void
2980 vm_page_set_invalid(vm_page_t m, int base, int size)
2981 {
2982 	vm_page_bits_t bits;
2983 
2984 	VM_OBJECT_ASSERT_WLOCKED(m->object);
2985 	bits = vm_page_bits(base, size);
2986 	if (m->valid == VM_PAGE_BITS_ALL && bits != 0)
2987 		pmap_remove_all(m);
2988 	KASSERT(!pmap_page_is_mapped(m),
2989 	    ("vm_page_set_invalid: page %p is mapped", m));
2990 	m->valid &= ~bits;
2991 	m->dirty &= ~bits;
2992 }
2993 
2994 /*
2995  * vm_page_zero_invalid()
2996  *
2997  *	The kernel assumes that the invalid portions of a page contain
2998  *	garbage, but such pages can be mapped into memory by user code.
2999  *	When this occurs, we must zero out the non-valid portions of the
3000  *	page so user code sees what it expects.
3001  *
3002  *	Pages are most often semi-valid when the end of a file is mapped
3003  *	into memory and the file's size is not page aligned.
3004  */
3005 void
3006 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
3007 {
3008 	int b;
3009 	int i;
3010 
3011 	VM_OBJECT_ASSERT_WLOCKED(m->object);
3012 	/*
3013 	 * Scan the valid bits looking for invalid sections that
3014 	 * must be zerod.  Invalid sub-DEV_BSIZE'd areas ( where the
3015 	 * valid bit may be set ) have already been zerod by
3016 	 * vm_page_set_validclean().
3017 	 */
3018 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
3019 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
3020 		    (m->valid & ((vm_page_bits_t)1 << i))) {
3021 			if (i > b) {
3022 				pmap_zero_page_area(m,
3023 				    b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
3024 			}
3025 			b = i + 1;
3026 		}
3027 	}
3028 
3029 	/*
3030 	 * setvalid is TRUE when we can safely set the zero'd areas
3031 	 * as being valid.  We can do this if there are no cache consistancy
3032 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
3033 	 */
3034 	if (setvalid)
3035 		m->valid = VM_PAGE_BITS_ALL;
3036 }
3037 
3038 /*
3039  *	vm_page_is_valid:
3040  *
3041  *	Is (partial) page valid?  Note that the case where size == 0
3042  *	will return FALSE in the degenerate case where the page is
3043  *	entirely invalid, and TRUE otherwise.
3044  */
3045 int
3046 vm_page_is_valid(vm_page_t m, int base, int size)
3047 {
3048 	vm_page_bits_t bits;
3049 
3050 	VM_OBJECT_ASSERT_LOCKED(m->object);
3051 	bits = vm_page_bits(base, size);
3052 	return (m->valid != 0 && (m->valid & bits) == bits);
3053 }
3054 
3055 /*
3056  * Set the page's dirty bits if the page is modified.
3057  */
3058 void
3059 vm_page_test_dirty(vm_page_t m)
3060 {
3061 
3062 	VM_OBJECT_ASSERT_WLOCKED(m->object);
3063 	if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
3064 		vm_page_dirty(m);
3065 }
3066 
3067 void
3068 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
3069 {
3070 
3071 	mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
3072 }
3073 
3074 void
3075 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
3076 {
3077 
3078 	mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
3079 }
3080 
3081 int
3082 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
3083 {
3084 
3085 	return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
3086 }
3087 
3088 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
3089 void
3090 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
3091 {
3092 
3093 	vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
3094 }
3095 
3096 void
3097 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
3098 {
3099 
3100 	mtx_assert_(vm_page_lockptr(m), a, file, line);
3101 }
3102 #endif
3103 
3104 int so_zerocp_fullpage = 0;
3105 
3106 /*
3107  *	Replace the given page with a copy.  The copied page assumes
3108  *	the portion of the given page's "wire_count" that is not the
3109  *	responsibility of this copy-on-write mechanism.
3110  *
3111  *	The object containing the given page must have a non-zero
3112  *	paging-in-progress count and be locked.
3113  */
3114 void
3115 vm_page_cowfault(vm_page_t m)
3116 {
3117 	vm_page_t mnew;
3118 	vm_object_t object;
3119 	vm_pindex_t pindex;
3120 
3121 	vm_page_lock_assert(m, MA_OWNED);
3122 	object = m->object;
3123 	VM_OBJECT_ASSERT_WLOCKED(object);
3124 	KASSERT(object->paging_in_progress != 0,
3125 	    ("vm_page_cowfault: object %p's paging-in-progress count is zero.",
3126 	    object));
3127 	pindex = m->pindex;
3128 
3129  retry_alloc:
3130 	mnew = vm_page_alloc(NULL, pindex, VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ);
3131 	if (mnew == NULL) {
3132 		vm_page_unlock(m);
3133 		VM_OBJECT_WUNLOCK(object);
3134 		VM_WAIT;
3135 		VM_OBJECT_WLOCK(object);
3136 		if (m == vm_page_lookup(object, pindex)) {
3137 			vm_page_lock(m);
3138 			goto retry_alloc;
3139 		} else {
3140 			/*
3141 			 * Page disappeared during the wait.
3142 			 */
3143 			return;
3144 		}
3145 	}
3146 
3147 	if (m->cow == 0) {
3148 		/*
3149 		 * check to see if we raced with an xmit complete when
3150 		 * waiting to allocate a page.  If so, put things back
3151 		 * the way they were
3152 		 */
3153 		vm_page_unlock(m);
3154 		vm_page_lock(mnew);
3155 		vm_page_free(mnew);
3156 		vm_page_unlock(mnew);
3157 	} else { /* clear COW & copy page */
3158 		pmap_remove_all(m);
3159 		mnew->object = object;
3160 		if (object->memattr != VM_MEMATTR_DEFAULT &&
3161 		    (object->flags & OBJ_FICTITIOUS) == 0)
3162 			pmap_page_set_memattr(mnew, object->memattr);
3163 		if (vm_page_replace(mnew, object, pindex) != m)
3164 			panic("vm_page_cowfault: invalid page replacement");
3165 		if (!so_zerocp_fullpage)
3166 			pmap_copy_page(m, mnew);
3167 		mnew->valid = VM_PAGE_BITS_ALL;
3168 		vm_page_dirty(mnew);
3169 		mnew->wire_count = m->wire_count - m->cow;
3170 		m->wire_count = m->cow;
3171 		vm_page_unlock(m);
3172 	}
3173 }
3174 
3175 void
3176 vm_page_cowclear(vm_page_t m)
3177 {
3178 
3179 	vm_page_lock_assert(m, MA_OWNED);
3180 	if (m->cow) {
3181 		m->cow--;
3182 		/*
3183 		 * let vm_fault add back write permission  lazily
3184 		 */
3185 	}
3186 	/*
3187 	 *  sf_buf_free() will free the page, so we needn't do it here
3188 	 */
3189 }
3190 
3191 int
3192 vm_page_cowsetup(vm_page_t m)
3193 {
3194 
3195 	vm_page_lock_assert(m, MA_OWNED);
3196 	if ((m->flags & PG_FICTITIOUS) != 0 ||
3197 	    (m->oflags & VPO_UNMANAGED) != 0 ||
3198 	    m->cow == USHRT_MAX - 1 || !VM_OBJECT_TRYWLOCK(m->object))
3199 		return (EBUSY);
3200 	m->cow++;
3201 	pmap_remove_write(m);
3202 	VM_OBJECT_WUNLOCK(m->object);
3203 	return (0);
3204 }
3205 
3206 #ifdef INVARIANTS
3207 void
3208 vm_page_object_lock_assert(vm_page_t m)
3209 {
3210 
3211 	/*
3212 	 * Certain of the page's fields may only be modified by the
3213 	 * holder of the containing object's lock or the exclusive busy.
3214 	 * holder.  Unfortunately, the holder of the write busy is
3215 	 * not recorded, and thus cannot be checked here.
3216 	 */
3217 	if (m->object != NULL && !vm_page_xbusied(m))
3218 		VM_OBJECT_ASSERT_WLOCKED(m->object);
3219 }
3220 #endif
3221 
3222 #include "opt_ddb.h"
3223 #ifdef DDB
3224 #include <sys/kernel.h>
3225 
3226 #include <ddb/ddb.h>
3227 
3228 DB_SHOW_COMMAND(page, vm_page_print_page_info)
3229 {
3230 	db_printf("cnt.v_free_count: %d\n", cnt.v_free_count);
3231 	db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count);
3232 	db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count);
3233 	db_printf("cnt.v_active_count: %d\n", cnt.v_active_count);
3234 	db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count);
3235 	db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved);
3236 	db_printf("cnt.v_free_min: %d\n", cnt.v_free_min);
3237 	db_printf("cnt.v_free_target: %d\n", cnt.v_free_target);
3238 	db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min);
3239 	db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target);
3240 }
3241 
3242 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
3243 {
3244 	int dom;
3245 
3246 	db_printf("pq_free %d pq_cache %d\n",
3247 	    cnt.v_free_count, cnt.v_cache_count);
3248 	for (dom = 0; dom < vm_ndomains; dom++) {
3249 		db_printf(
3250 	"dom %d page_cnt %d free %d pq_act %d pq_inact %d pass %d\n",
3251 		    dom,
3252 		    vm_dom[dom].vmd_page_count,
3253 		    vm_dom[dom].vmd_free_count,
3254 		    vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
3255 		    vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
3256 		    vm_dom[dom].vmd_pass);
3257 	}
3258 }
3259 
3260 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
3261 {
3262 	vm_page_t m;
3263 	boolean_t phys;
3264 
3265 	if (!have_addr) {
3266 		db_printf("show pginfo addr\n");
3267 		return;
3268 	}
3269 
3270 	phys = strchr(modif, 'p') != NULL;
3271 	if (phys)
3272 		m = PHYS_TO_VM_PAGE(addr);
3273 	else
3274 		m = (vm_page_t)addr;
3275 	db_printf(
3276     "page %p obj %p pidx 0x%jx phys 0x%jx q %d hold %d wire %d\n"
3277     "  af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
3278 	    m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
3279 	    m->queue, m->hold_count, m->wire_count, m->aflags, m->oflags,
3280 	    m->flags, m->act_count, m->busy_lock, m->valid, m->dirty);
3281 }
3282 #endif /* DDB */
3283