xref: /freebsd/sys/vm/vm_page.c (revision ca987d4641cdcd7f27e153db17c5bf064934faf5)
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  * 3. 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/linker.h>
95 #include <sys/malloc.h>
96 #include <sys/mman.h>
97 #include <sys/msgbuf.h>
98 #include <sys/mutex.h>
99 #include <sys/proc.h>
100 #include <sys/rwlock.h>
101 #include <sys/sbuf.h>
102 #include <sys/smp.h>
103 #include <sys/sysctl.h>
104 #include <sys/vmmeter.h>
105 #include <sys/vnode.h>
106 
107 #include <vm/vm.h>
108 #include <vm/pmap.h>
109 #include <vm/vm_param.h>
110 #include <vm/vm_kern.h>
111 #include <vm/vm_object.h>
112 #include <vm/vm_page.h>
113 #include <vm/vm_pageout.h>
114 #include <vm/vm_pager.h>
115 #include <vm/vm_phys.h>
116 #include <vm/vm_radix.h>
117 #include <vm/vm_reserv.h>
118 #include <vm/vm_extern.h>
119 #include <vm/uma.h>
120 #include <vm/uma_int.h>
121 
122 #include <machine/md_var.h>
123 
124 /*
125  *	Associated with page of user-allocatable memory is a
126  *	page structure.
127  */
128 
129 struct vm_domain vm_dom[MAXMEMDOM];
130 struct mtx_padalign __exclusive_cache_line vm_page_queue_free_mtx;
131 
132 struct mtx_padalign __exclusive_cache_line pa_lock[PA_LOCK_COUNT];
133 
134 /*
135  * bogus page -- for I/O to/from partially complete buffers,
136  * or for paging into sparsely invalid regions.
137  */
138 vm_page_t bogus_page;
139 
140 vm_page_t vm_page_array;
141 long vm_page_array_size;
142 long first_page;
143 
144 static int boot_pages = UMA_BOOT_PAGES;
145 SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
146     &boot_pages, 0,
147     "number of pages allocated for bootstrapping the VM system");
148 
149 static int pa_tryrelock_restart;
150 SYSCTL_INT(_vm, OID_AUTO, tryrelock_restart, CTLFLAG_RD,
151     &pa_tryrelock_restart, 0, "Number of tryrelock restarts");
152 
153 static TAILQ_HEAD(, vm_page) blacklist_head;
154 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS);
155 SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD |
156     CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages");
157 
158 /* Is the page daemon waiting for free pages? */
159 static int vm_pageout_pages_needed;
160 
161 static uma_zone_t fakepg_zone;
162 
163 static void vm_page_alloc_check(vm_page_t m);
164 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits);
165 static void vm_page_enqueue(uint8_t queue, vm_page_t m);
166 static void vm_page_free_phys(vm_page_t m);
167 static void vm_page_free_wakeup(void);
168 static void vm_page_init(void *dummy);
169 static int vm_page_insert_after(vm_page_t m, vm_object_t object,
170     vm_pindex_t pindex, vm_page_t mpred);
171 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object,
172     vm_page_t mpred);
173 static int vm_page_reclaim_run(int req_class, u_long npages, vm_page_t m_run,
174     vm_paddr_t high);
175 static int vm_page_alloc_fail(vm_object_t object, int req);
176 
177 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL);
178 
179 static void
180 vm_page_init(void *dummy)
181 {
182 
183 	fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
184 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
185 	bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
186 	    VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
187 }
188 
189 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
190 #if PAGE_SIZE == 32768
191 #ifdef CTASSERT
192 CTASSERT(sizeof(u_long) >= 8);
193 #endif
194 #endif
195 
196 /*
197  * Try to acquire a physical address lock while a pmap is locked.  If we
198  * fail to trylock we unlock and lock the pmap directly and cache the
199  * locked pa in *locked.  The caller should then restart their loop in case
200  * the virtual to physical mapping has changed.
201  */
202 int
203 vm_page_pa_tryrelock(pmap_t pmap, vm_paddr_t pa, vm_paddr_t *locked)
204 {
205 	vm_paddr_t lockpa;
206 
207 	lockpa = *locked;
208 	*locked = pa;
209 	if (lockpa) {
210 		PA_LOCK_ASSERT(lockpa, MA_OWNED);
211 		if (PA_LOCKPTR(pa) == PA_LOCKPTR(lockpa))
212 			return (0);
213 		PA_UNLOCK(lockpa);
214 	}
215 	if (PA_TRYLOCK(pa))
216 		return (0);
217 	PMAP_UNLOCK(pmap);
218 	atomic_add_int(&pa_tryrelock_restart, 1);
219 	PA_LOCK(pa);
220 	PMAP_LOCK(pmap);
221 	return (EAGAIN);
222 }
223 
224 /*
225  *	vm_set_page_size:
226  *
227  *	Sets the page size, perhaps based upon the memory
228  *	size.  Must be called before any use of page-size
229  *	dependent functions.
230  */
231 void
232 vm_set_page_size(void)
233 {
234 	if (vm_cnt.v_page_size == 0)
235 		vm_cnt.v_page_size = PAGE_SIZE;
236 	if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0)
237 		panic("vm_set_page_size: page size not a power of two");
238 }
239 
240 /*
241  *	vm_page_blacklist_next:
242  *
243  *	Find the next entry in the provided string of blacklist
244  *	addresses.  Entries are separated by space, comma, or newline.
245  *	If an invalid integer is encountered then the rest of the
246  *	string is skipped.  Updates the list pointer to the next
247  *	character, or NULL if the string is exhausted or invalid.
248  */
249 static vm_paddr_t
250 vm_page_blacklist_next(char **list, char *end)
251 {
252 	vm_paddr_t bad;
253 	char *cp, *pos;
254 
255 	if (list == NULL || *list == NULL)
256 		return (0);
257 	if (**list =='\0') {
258 		*list = NULL;
259 		return (0);
260 	}
261 
262 	/*
263 	 * If there's no end pointer then the buffer is coming from
264 	 * the kenv and we know it's null-terminated.
265 	 */
266 	if (end == NULL)
267 		end = *list + strlen(*list);
268 
269 	/* Ensure that strtoq() won't walk off the end */
270 	if (*end != '\0') {
271 		if (*end == '\n' || *end == ' ' || *end  == ',')
272 			*end = '\0';
273 		else {
274 			printf("Blacklist not terminated, skipping\n");
275 			*list = NULL;
276 			return (0);
277 		}
278 	}
279 
280 	for (pos = *list; *pos != '\0'; pos = cp) {
281 		bad = strtoq(pos, &cp, 0);
282 		if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') {
283 			if (bad == 0) {
284 				if (++cp < end)
285 					continue;
286 				else
287 					break;
288 			}
289 		} else
290 			break;
291 		if (*cp == '\0' || ++cp >= end)
292 			*list = NULL;
293 		else
294 			*list = cp;
295 		return (trunc_page(bad));
296 	}
297 	printf("Garbage in RAM blacklist, skipping\n");
298 	*list = NULL;
299 	return (0);
300 }
301 
302 /*
303  *	vm_page_blacklist_check:
304  *
305  *	Iterate through the provided string of blacklist addresses, pulling
306  *	each entry out of the physical allocator free list and putting it
307  *	onto a list for reporting via the vm.page_blacklist sysctl.
308  */
309 static void
310 vm_page_blacklist_check(char *list, char *end)
311 {
312 	vm_paddr_t pa;
313 	vm_page_t m;
314 	char *next;
315 	int ret;
316 
317 	next = list;
318 	while (next != NULL) {
319 		if ((pa = vm_page_blacklist_next(&next, end)) == 0)
320 			continue;
321 		m = vm_phys_paddr_to_vm_page(pa);
322 		if (m == NULL)
323 			continue;
324 		mtx_lock(&vm_page_queue_free_mtx);
325 		ret = vm_phys_unfree_page(m);
326 		mtx_unlock(&vm_page_queue_free_mtx);
327 		if (ret == TRUE) {
328 			TAILQ_INSERT_TAIL(&blacklist_head, m, listq);
329 			if (bootverbose)
330 				printf("Skipping page with pa 0x%jx\n",
331 				    (uintmax_t)pa);
332 		}
333 	}
334 }
335 
336 /*
337  *	vm_page_blacklist_load:
338  *
339  *	Search for a special module named "ram_blacklist".  It'll be a
340  *	plain text file provided by the user via the loader directive
341  *	of the same name.
342  */
343 static void
344 vm_page_blacklist_load(char **list, char **end)
345 {
346 	void *mod;
347 	u_char *ptr;
348 	u_int len;
349 
350 	mod = NULL;
351 	ptr = NULL;
352 
353 	mod = preload_search_by_type("ram_blacklist");
354 	if (mod != NULL) {
355 		ptr = preload_fetch_addr(mod);
356 		len = preload_fetch_size(mod);
357         }
358 	*list = ptr;
359 	if (ptr != NULL)
360 		*end = ptr + len;
361 	else
362 		*end = NULL;
363 	return;
364 }
365 
366 static int
367 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS)
368 {
369 	vm_page_t m;
370 	struct sbuf sbuf;
371 	int error, first;
372 
373 	first = 1;
374 	error = sysctl_wire_old_buffer(req, 0);
375 	if (error != 0)
376 		return (error);
377 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
378 	TAILQ_FOREACH(m, &blacklist_head, listq) {
379 		sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",",
380 		    (uintmax_t)m->phys_addr);
381 		first = 0;
382 	}
383 	error = sbuf_finish(&sbuf);
384 	sbuf_delete(&sbuf);
385 	return (error);
386 }
387 
388 static void
389 vm_page_domain_init(struct vm_domain *vmd)
390 {
391 	struct vm_pagequeue *pq;
392 	int i;
393 
394 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) =
395 	    "vm inactive pagequeue";
396 	*__DECONST(u_int **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_vcnt) =
397 	    &vm_cnt.v_inactive_count;
398 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) =
399 	    "vm active pagequeue";
400 	*__DECONST(u_int **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_vcnt) =
401 	    &vm_cnt.v_active_count;
402 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) =
403 	    "vm laundry pagequeue";
404 	*__DECONST(int **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_vcnt) =
405 	    &vm_cnt.v_laundry_count;
406 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) =
407 	    "vm unswappable pagequeue";
408 	/* Unswappable dirty pages are counted as being in the laundry. */
409 	*__DECONST(int **, &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_vcnt) =
410 	    &vm_cnt.v_laundry_count;
411 	vmd->vmd_page_count = 0;
412 	vmd->vmd_free_count = 0;
413 	vmd->vmd_segs = 0;
414 	vmd->vmd_oom = FALSE;
415 	for (i = 0; i < PQ_COUNT; i++) {
416 		pq = &vmd->vmd_pagequeues[i];
417 		TAILQ_INIT(&pq->pq_pl);
418 		mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue",
419 		    MTX_DEF | MTX_DUPOK);
420 	}
421 }
422 
423 /*
424  *	vm_page_startup:
425  *
426  *	Initializes the resident memory module.  Allocates physical memory for
427  *	bootstrapping UMA and some data structures that are used to manage
428  *	physical pages.  Initializes these structures, and populates the free
429  *	page queues.
430  */
431 vm_offset_t
432 vm_page_startup(vm_offset_t vaddr)
433 {
434 	struct vm_domain *vmd;
435 	struct vm_phys_seg *seg;
436 	vm_page_t m;
437 	char *list, *listend;
438 	vm_offset_t mapped;
439 	vm_paddr_t end, high_avail, low_avail, new_end, page_range, size;
440 	vm_paddr_t biggestsize, last_pa, pa;
441 	u_long pagecount;
442 	int biggestone, i, pages_per_zone, segind;
443 
444 	biggestsize = 0;
445 	biggestone = 0;
446 	vaddr = round_page(vaddr);
447 
448 	for (i = 0; phys_avail[i + 1]; i += 2) {
449 		phys_avail[i] = round_page(phys_avail[i]);
450 		phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
451 	}
452 	for (i = 0; phys_avail[i + 1]; i += 2) {
453 		size = phys_avail[i + 1] - phys_avail[i];
454 		if (size > biggestsize) {
455 			biggestone = i;
456 			biggestsize = size;
457 		}
458 	}
459 
460 	end = phys_avail[biggestone+1];
461 
462 	/*
463 	 * Initialize the page and queue locks.
464 	 */
465 	mtx_init(&vm_page_queue_free_mtx, "vm page free queue", NULL, MTX_DEF);
466 	for (i = 0; i < PA_LOCK_COUNT; i++)
467 		mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
468 	for (i = 0; i < vm_ndomains; i++)
469 		vm_page_domain_init(&vm_dom[i]);
470 
471 	/*
472 	 * Almost all of the pages needed for bootstrapping UMA are used
473 	 * for zone structures, so if the number of CPUs results in those
474 	 * structures taking more than one page each, we set aside more pages
475 	 * in proportion to the zone structure size.
476 	 */
477 	pages_per_zone = howmany(sizeof(struct uma_zone) +
478 	    sizeof(struct uma_cache) * (mp_maxid + 1) +
479 	    roundup2(sizeof(struct uma_slab), sizeof(void *)), UMA_SLAB_SIZE);
480 	if (pages_per_zone > 1) {
481 		/* Reserve more pages so that we don't run out. */
482 		boot_pages = UMA_BOOT_PAGES_ZONES * pages_per_zone;
483 	}
484 
485 	/*
486 	 * Allocate memory for use when boot strapping the kernel memory
487 	 * allocator.
488 	 *
489 	 * CTFLAG_RDTUN doesn't work during the early boot process, so we must
490 	 * manually fetch the value.
491 	 */
492 	TUNABLE_INT_FETCH("vm.boot_pages", &boot_pages);
493 	new_end = end - (boot_pages * UMA_SLAB_SIZE);
494 	new_end = trunc_page(new_end);
495 	mapped = pmap_map(&vaddr, new_end, end,
496 	    VM_PROT_READ | VM_PROT_WRITE);
497 	bzero((void *)mapped, end - new_end);
498 	uma_startup((void *)mapped, boot_pages);
499 
500 #if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \
501     defined(__i386__) || defined(__mips__)
502 	/*
503 	 * Allocate a bitmap to indicate that a random physical page
504 	 * needs to be included in a minidump.
505 	 *
506 	 * The amd64 port needs this to indicate which direct map pages
507 	 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
508 	 *
509 	 * However, i386 still needs this workspace internally within the
510 	 * minidump code.  In theory, they are not needed on i386, but are
511 	 * included should the sf_buf code decide to use them.
512 	 */
513 	last_pa = 0;
514 	for (i = 0; dump_avail[i + 1] != 0; i += 2)
515 		if (dump_avail[i + 1] > last_pa)
516 			last_pa = dump_avail[i + 1];
517 	page_range = last_pa / PAGE_SIZE;
518 	vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
519 	new_end -= vm_page_dump_size;
520 	vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
521 	    new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
522 	bzero((void *)vm_page_dump, vm_page_dump_size);
523 #else
524 	(void)last_pa;
525 #endif
526 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__)
527 	/*
528 	 * Include the UMA bootstrap pages and vm_page_dump in a crash dump.
529 	 * When pmap_map() uses the direct map, they are not automatically
530 	 * included.
531 	 */
532 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
533 		dump_add_page(pa);
534 #endif
535 	phys_avail[biggestone + 1] = new_end;
536 #ifdef __amd64__
537 	/*
538 	 * Request that the physical pages underlying the message buffer be
539 	 * included in a crash dump.  Since the message buffer is accessed
540 	 * through the direct map, they are not automatically included.
541 	 */
542 	pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
543 	last_pa = pa + round_page(msgbufsize);
544 	while (pa < last_pa) {
545 		dump_add_page(pa);
546 		pa += PAGE_SIZE;
547 	}
548 #endif
549 	/*
550 	 * Compute the number of pages of memory that will be available for
551 	 * use, taking into account the overhead of a page structure per page.
552 	 * In other words, solve
553 	 *	"available physical memory" - round_page(page_range *
554 	 *	    sizeof(struct vm_page)) = page_range * PAGE_SIZE
555 	 * for page_range.
556 	 */
557 	low_avail = phys_avail[0];
558 	high_avail = phys_avail[1];
559 	for (i = 0; i < vm_phys_nsegs; i++) {
560 		if (vm_phys_segs[i].start < low_avail)
561 			low_avail = vm_phys_segs[i].start;
562 		if (vm_phys_segs[i].end > high_avail)
563 			high_avail = vm_phys_segs[i].end;
564 	}
565 	/* Skip the first chunk.  It is already accounted for. */
566 	for (i = 2; phys_avail[i + 1] != 0; i += 2) {
567 		if (phys_avail[i] < low_avail)
568 			low_avail = phys_avail[i];
569 		if (phys_avail[i + 1] > high_avail)
570 			high_avail = phys_avail[i + 1];
571 	}
572 	first_page = low_avail / PAGE_SIZE;
573 #ifdef VM_PHYSSEG_SPARSE
574 	size = 0;
575 	for (i = 0; i < vm_phys_nsegs; i++)
576 		size += vm_phys_segs[i].end - vm_phys_segs[i].start;
577 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
578 		size += phys_avail[i + 1] - phys_avail[i];
579 #elif defined(VM_PHYSSEG_DENSE)
580 	size = high_avail - low_avail;
581 #else
582 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
583 #endif
584 
585 #ifdef VM_PHYSSEG_DENSE
586 	/*
587 	 * In the VM_PHYSSEG_DENSE case, the number of pages can account for
588 	 * the overhead of a page structure per page only if vm_page_array is
589 	 * allocated from the last physical memory chunk.  Otherwise, we must
590 	 * allocate page structures representing the physical memory
591 	 * underlying vm_page_array, even though they will not be used.
592 	 */
593 	if (new_end != high_avail)
594 		page_range = size / PAGE_SIZE;
595 	else
596 #endif
597 	{
598 		page_range = size / (PAGE_SIZE + sizeof(struct vm_page));
599 
600 		/*
601 		 * If the partial bytes remaining are large enough for
602 		 * a page (PAGE_SIZE) without a corresponding
603 		 * 'struct vm_page', then new_end will contain an
604 		 * extra page after subtracting the length of the VM
605 		 * page array.  Compensate by subtracting an extra
606 		 * page from new_end.
607 		 */
608 		if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) {
609 			if (new_end == high_avail)
610 				high_avail -= PAGE_SIZE;
611 			new_end -= PAGE_SIZE;
612 		}
613 	}
614 	end = new_end;
615 
616 	/*
617 	 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
618 	 * However, because this page is allocated from KVM, out-of-bounds
619 	 * accesses using the direct map will not be trapped.
620 	 */
621 	vaddr += PAGE_SIZE;
622 
623 	/*
624 	 * Allocate physical memory for the page structures, and map it.
625 	 */
626 	new_end = trunc_page(end - page_range * sizeof(struct vm_page));
627 	mapped = pmap_map(&vaddr, new_end, end,
628 	    VM_PROT_READ | VM_PROT_WRITE);
629 	vm_page_array = (vm_page_t)mapped;
630 	vm_page_array_size = page_range;
631 
632 #if VM_NRESERVLEVEL > 0
633 	/*
634 	 * Allocate physical memory for the reservation management system's
635 	 * data structures, and map it.
636 	 */
637 	if (high_avail == end)
638 		high_avail = new_end;
639 	new_end = vm_reserv_startup(&vaddr, new_end, high_avail);
640 #endif
641 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__)
642 	/*
643 	 * Include vm_page_array and vm_reserv_array in a crash dump.
644 	 */
645 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
646 		dump_add_page(pa);
647 #endif
648 	phys_avail[biggestone + 1] = new_end;
649 
650 	/*
651 	 * Add physical memory segments corresponding to the available
652 	 * physical pages.
653 	 */
654 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
655 		vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]);
656 
657 	/*
658 	 * Initialize the physical memory allocator.
659 	 */
660 	vm_phys_init();
661 
662 	/*
663 	 * Initialize the page structures and add every available page to the
664 	 * physical memory allocator's free lists.
665 	 */
666 	vm_cnt.v_page_count = 0;
667 	vm_cnt.v_free_count = 0;
668 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
669 		seg = &vm_phys_segs[segind];
670 		for (pa = seg->start; pa < seg->end; pa += PAGE_SIZE)
671 			vm_phys_init_page(pa);
672 
673 		/*
674 		 * Add the segment to the free lists only if it is covered by
675 		 * one of the ranges in phys_avail.  Because we've added the
676 		 * ranges to the vm_phys_segs array, we can assume that each
677 		 * segment is either entirely contained in one of the ranges,
678 		 * or doesn't overlap any of them.
679 		 */
680 		for (i = 0; phys_avail[i + 1] != 0; i += 2) {
681 			if (seg->start < phys_avail[i] ||
682 			    seg->end > phys_avail[i + 1])
683 				continue;
684 
685 			m = seg->first_page;
686 			pagecount = (u_long)atop(seg->end - seg->start);
687 
688 			mtx_lock(&vm_page_queue_free_mtx);
689 			vm_phys_free_contig(m, pagecount);
690 			vm_phys_freecnt_adj(m, (int)pagecount);
691 			mtx_unlock(&vm_page_queue_free_mtx);
692 			vm_cnt.v_page_count += (u_int)pagecount;
693 
694 			vmd = &vm_dom[seg->domain];
695 			vmd->vmd_page_count += (u_int)pagecount;
696 			vmd->vmd_segs |= 1UL << m->segind;
697 			break;
698 		}
699 	}
700 
701 	/*
702 	 * Remove blacklisted pages from the physical memory allocator.
703 	 */
704 	TAILQ_INIT(&blacklist_head);
705 	vm_page_blacklist_load(&list, &listend);
706 	vm_page_blacklist_check(list, listend);
707 
708 	list = kern_getenv("vm.blacklist");
709 	vm_page_blacklist_check(list, NULL);
710 
711 	freeenv(list);
712 #if VM_NRESERVLEVEL > 0
713 	/*
714 	 * Initialize the reservation management system.
715 	 */
716 	vm_reserv_init();
717 #endif
718 	return (vaddr);
719 }
720 
721 void
722 vm_page_reference(vm_page_t m)
723 {
724 
725 	vm_page_aflag_set(m, PGA_REFERENCED);
726 }
727 
728 /*
729  *	vm_page_busy_downgrade:
730  *
731  *	Downgrade an exclusive busy page into a single shared busy page.
732  */
733 void
734 vm_page_busy_downgrade(vm_page_t m)
735 {
736 	u_int x;
737 	bool locked;
738 
739 	vm_page_assert_xbusied(m);
740 	locked = mtx_owned(vm_page_lockptr(m));
741 
742 	for (;;) {
743 		x = m->busy_lock;
744 		x &= VPB_BIT_WAITERS;
745 		if (x != 0 && !locked)
746 			vm_page_lock(m);
747 		if (atomic_cmpset_rel_int(&m->busy_lock,
748 		    VPB_SINGLE_EXCLUSIVER | x, VPB_SHARERS_WORD(1)))
749 			break;
750 		if (x != 0 && !locked)
751 			vm_page_unlock(m);
752 	}
753 	if (x != 0) {
754 		wakeup(m);
755 		if (!locked)
756 			vm_page_unlock(m);
757 	}
758 }
759 
760 /*
761  *	vm_page_sbusied:
762  *
763  *	Return a positive value if the page is shared busied, 0 otherwise.
764  */
765 int
766 vm_page_sbusied(vm_page_t m)
767 {
768 	u_int x;
769 
770 	x = m->busy_lock;
771 	return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED);
772 }
773 
774 /*
775  *	vm_page_sunbusy:
776  *
777  *	Shared unbusy a page.
778  */
779 void
780 vm_page_sunbusy(vm_page_t m)
781 {
782 	u_int x;
783 
784 	vm_page_lock_assert(m, MA_NOTOWNED);
785 	vm_page_assert_sbusied(m);
786 
787 	for (;;) {
788 		x = m->busy_lock;
789 		if (VPB_SHARERS(x) > 1) {
790 			if (atomic_cmpset_int(&m->busy_lock, x,
791 			    x - VPB_ONE_SHARER))
792 				break;
793 			continue;
794 		}
795 		if ((x & VPB_BIT_WAITERS) == 0) {
796 			KASSERT(x == VPB_SHARERS_WORD(1),
797 			    ("vm_page_sunbusy: invalid lock state"));
798 			if (atomic_cmpset_int(&m->busy_lock,
799 			    VPB_SHARERS_WORD(1), VPB_UNBUSIED))
800 				break;
801 			continue;
802 		}
803 		KASSERT(x == (VPB_SHARERS_WORD(1) | VPB_BIT_WAITERS),
804 		    ("vm_page_sunbusy: invalid lock state for waiters"));
805 
806 		vm_page_lock(m);
807 		if (!atomic_cmpset_int(&m->busy_lock, x, VPB_UNBUSIED)) {
808 			vm_page_unlock(m);
809 			continue;
810 		}
811 		wakeup(m);
812 		vm_page_unlock(m);
813 		break;
814 	}
815 }
816 
817 /*
818  *	vm_page_busy_sleep:
819  *
820  *	Sleep and release the page lock, using the page pointer as wchan.
821  *	This is used to implement the hard-path of busying mechanism.
822  *
823  *	The given page must be locked.
824  *
825  *	If nonshared is true, sleep only if the page is xbusy.
826  */
827 void
828 vm_page_busy_sleep(vm_page_t m, const char *wmesg, bool nonshared)
829 {
830 	u_int x;
831 
832 	vm_page_assert_locked(m);
833 
834 	x = m->busy_lock;
835 	if (x == VPB_UNBUSIED || (nonshared && (x & VPB_BIT_SHARED) != 0) ||
836 	    ((x & VPB_BIT_WAITERS) == 0 &&
837 	    !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS))) {
838 		vm_page_unlock(m);
839 		return;
840 	}
841 	msleep(m, vm_page_lockptr(m), PVM | PDROP, wmesg, 0);
842 }
843 
844 /*
845  *	vm_page_trysbusy:
846  *
847  *	Try to shared busy a page.
848  *	If the operation succeeds 1 is returned otherwise 0.
849  *	The operation never sleeps.
850  */
851 int
852 vm_page_trysbusy(vm_page_t m)
853 {
854 	u_int x;
855 
856 	for (;;) {
857 		x = m->busy_lock;
858 		if ((x & VPB_BIT_SHARED) == 0)
859 			return (0);
860 		if (atomic_cmpset_acq_int(&m->busy_lock, x, x + VPB_ONE_SHARER))
861 			return (1);
862 	}
863 }
864 
865 static void
866 vm_page_xunbusy_locked(vm_page_t m)
867 {
868 
869 	vm_page_assert_xbusied(m);
870 	vm_page_assert_locked(m);
871 
872 	atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
873 	/* There is a waiter, do wakeup() instead of vm_page_flash(). */
874 	wakeup(m);
875 }
876 
877 void
878 vm_page_xunbusy_maybelocked(vm_page_t m)
879 {
880 	bool lockacq;
881 
882 	vm_page_assert_xbusied(m);
883 
884 	/*
885 	 * Fast path for unbusy.  If it succeeds, we know that there
886 	 * are no waiters, so we do not need a wakeup.
887 	 */
888 	if (atomic_cmpset_rel_int(&m->busy_lock, VPB_SINGLE_EXCLUSIVER,
889 	    VPB_UNBUSIED))
890 		return;
891 
892 	lockacq = !mtx_owned(vm_page_lockptr(m));
893 	if (lockacq)
894 		vm_page_lock(m);
895 	vm_page_xunbusy_locked(m);
896 	if (lockacq)
897 		vm_page_unlock(m);
898 }
899 
900 /*
901  *	vm_page_xunbusy_hard:
902  *
903  *	Called after the first try the exclusive unbusy of a page failed.
904  *	It is assumed that the waiters bit is on.
905  */
906 void
907 vm_page_xunbusy_hard(vm_page_t m)
908 {
909 
910 	vm_page_assert_xbusied(m);
911 
912 	vm_page_lock(m);
913 	vm_page_xunbusy_locked(m);
914 	vm_page_unlock(m);
915 }
916 
917 /*
918  *	vm_page_flash:
919  *
920  *	Wakeup anyone waiting for the page.
921  *	The ownership bits do not change.
922  *
923  *	The given page must be locked.
924  */
925 void
926 vm_page_flash(vm_page_t m)
927 {
928 	u_int x;
929 
930 	vm_page_lock_assert(m, MA_OWNED);
931 
932 	for (;;) {
933 		x = m->busy_lock;
934 		if ((x & VPB_BIT_WAITERS) == 0)
935 			return;
936 		if (atomic_cmpset_int(&m->busy_lock, x,
937 		    x & (~VPB_BIT_WAITERS)))
938 			break;
939 	}
940 	wakeup(m);
941 }
942 
943 /*
944  * Avoid releasing and reacquiring the same page lock.
945  */
946 void
947 vm_page_change_lock(vm_page_t m, struct mtx **mtx)
948 {
949 	struct mtx *mtx1;
950 
951 	mtx1 = vm_page_lockptr(m);
952 	if (*mtx == mtx1)
953 		return;
954 	if (*mtx != NULL)
955 		mtx_unlock(*mtx);
956 	*mtx = mtx1;
957 	mtx_lock(mtx1);
958 }
959 
960 /*
961  * Keep page from being freed by the page daemon
962  * much of the same effect as wiring, except much lower
963  * overhead and should be used only for *very* temporary
964  * holding ("wiring").
965  */
966 void
967 vm_page_hold(vm_page_t mem)
968 {
969 
970 	vm_page_lock_assert(mem, MA_OWNED);
971         mem->hold_count++;
972 }
973 
974 void
975 vm_page_unhold(vm_page_t mem)
976 {
977 
978 	vm_page_lock_assert(mem, MA_OWNED);
979 	KASSERT(mem->hold_count >= 1, ("vm_page_unhold: hold count < 0!!!"));
980 	--mem->hold_count;
981 	if (mem->hold_count == 0 && (mem->flags & PG_UNHOLDFREE) != 0)
982 		vm_page_free_toq(mem);
983 }
984 
985 /*
986  *	vm_page_unhold_pages:
987  *
988  *	Unhold each of the pages that is referenced by the given array.
989  */
990 void
991 vm_page_unhold_pages(vm_page_t *ma, int count)
992 {
993 	struct mtx *mtx;
994 
995 	mtx = NULL;
996 	for (; count != 0; count--) {
997 		vm_page_change_lock(*ma, &mtx);
998 		vm_page_unhold(*ma);
999 		ma++;
1000 	}
1001 	if (mtx != NULL)
1002 		mtx_unlock(mtx);
1003 }
1004 
1005 vm_page_t
1006 PHYS_TO_VM_PAGE(vm_paddr_t pa)
1007 {
1008 	vm_page_t m;
1009 
1010 #ifdef VM_PHYSSEG_SPARSE
1011 	m = vm_phys_paddr_to_vm_page(pa);
1012 	if (m == NULL)
1013 		m = vm_phys_fictitious_to_vm_page(pa);
1014 	return (m);
1015 #elif defined(VM_PHYSSEG_DENSE)
1016 	long pi;
1017 
1018 	pi = atop(pa);
1019 	if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
1020 		m = &vm_page_array[pi - first_page];
1021 		return (m);
1022 	}
1023 	return (vm_phys_fictitious_to_vm_page(pa));
1024 #else
1025 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
1026 #endif
1027 }
1028 
1029 /*
1030  *	vm_page_getfake:
1031  *
1032  *	Create a fictitious page with the specified physical address and
1033  *	memory attribute.  The memory attribute is the only the machine-
1034  *	dependent aspect of a fictitious page that must be initialized.
1035  */
1036 vm_page_t
1037 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
1038 {
1039 	vm_page_t m;
1040 
1041 	m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
1042 	vm_page_initfake(m, paddr, memattr);
1043 	return (m);
1044 }
1045 
1046 void
1047 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1048 {
1049 
1050 	if ((m->flags & PG_FICTITIOUS) != 0) {
1051 		/*
1052 		 * The page's memattr might have changed since the
1053 		 * previous initialization.  Update the pmap to the
1054 		 * new memattr.
1055 		 */
1056 		goto memattr;
1057 	}
1058 	m->phys_addr = paddr;
1059 	m->queue = PQ_NONE;
1060 	/* Fictitious pages don't use "segind". */
1061 	m->flags = PG_FICTITIOUS;
1062 	/* Fictitious pages don't use "order" or "pool". */
1063 	m->oflags = VPO_UNMANAGED;
1064 	m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1065 	m->wire_count = 1;
1066 	pmap_page_init(m);
1067 memattr:
1068 	pmap_page_set_memattr(m, memattr);
1069 }
1070 
1071 /*
1072  *	vm_page_putfake:
1073  *
1074  *	Release a fictitious page.
1075  */
1076 void
1077 vm_page_putfake(vm_page_t m)
1078 {
1079 
1080 	KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
1081 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1082 	    ("vm_page_putfake: bad page %p", m));
1083 	uma_zfree(fakepg_zone, m);
1084 }
1085 
1086 /*
1087  *	vm_page_updatefake:
1088  *
1089  *	Update the given fictitious page to the specified physical address and
1090  *	memory attribute.
1091  */
1092 void
1093 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1094 {
1095 
1096 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1097 	    ("vm_page_updatefake: bad page %p", m));
1098 	m->phys_addr = paddr;
1099 	pmap_page_set_memattr(m, memattr);
1100 }
1101 
1102 /*
1103  *	vm_page_free:
1104  *
1105  *	Free a page.
1106  */
1107 void
1108 vm_page_free(vm_page_t m)
1109 {
1110 
1111 	m->flags &= ~PG_ZERO;
1112 	vm_page_free_toq(m);
1113 }
1114 
1115 /*
1116  *	vm_page_free_zero:
1117  *
1118  *	Free a page to the zerod-pages queue
1119  */
1120 void
1121 vm_page_free_zero(vm_page_t m)
1122 {
1123 
1124 	m->flags |= PG_ZERO;
1125 	vm_page_free_toq(m);
1126 }
1127 
1128 /*
1129  * Unbusy and handle the page queueing for a page from a getpages request that
1130  * was optionally read ahead or behind.
1131  */
1132 void
1133 vm_page_readahead_finish(vm_page_t m)
1134 {
1135 
1136 	/* We shouldn't put invalid pages on queues. */
1137 	KASSERT(m->valid != 0, ("%s: %p is invalid", __func__, m));
1138 
1139 	/*
1140 	 * Since the page is not the actually needed one, whether it should
1141 	 * be activated or deactivated is not obvious.  Empirical results
1142 	 * have shown that deactivating the page is usually the best choice,
1143 	 * unless the page is wanted by another thread.
1144 	 */
1145 	vm_page_lock(m);
1146 	if ((m->busy_lock & VPB_BIT_WAITERS) != 0)
1147 		vm_page_activate(m);
1148 	else
1149 		vm_page_deactivate(m);
1150 	vm_page_unlock(m);
1151 	vm_page_xunbusy(m);
1152 }
1153 
1154 /*
1155  *	vm_page_sleep_if_busy:
1156  *
1157  *	Sleep and release the page queues lock if the page is busied.
1158  *	Returns TRUE if the thread slept.
1159  *
1160  *	The given page must be unlocked and object containing it must
1161  *	be locked.
1162  */
1163 int
1164 vm_page_sleep_if_busy(vm_page_t m, const char *msg)
1165 {
1166 	vm_object_t obj;
1167 
1168 	vm_page_lock_assert(m, MA_NOTOWNED);
1169 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1170 
1171 	if (vm_page_busied(m)) {
1172 		/*
1173 		 * The page-specific object must be cached because page
1174 		 * identity can change during the sleep, causing the
1175 		 * re-lock of a different object.
1176 		 * It is assumed that a reference to the object is already
1177 		 * held by the callers.
1178 		 */
1179 		obj = m->object;
1180 		vm_page_lock(m);
1181 		VM_OBJECT_WUNLOCK(obj);
1182 		vm_page_busy_sleep(m, msg, false);
1183 		VM_OBJECT_WLOCK(obj);
1184 		return (TRUE);
1185 	}
1186 	return (FALSE);
1187 }
1188 
1189 /*
1190  *	vm_page_dirty_KBI:		[ internal use only ]
1191  *
1192  *	Set all bits in the page's dirty field.
1193  *
1194  *	The object containing the specified page must be locked if the
1195  *	call is made from the machine-independent layer.
1196  *
1197  *	See vm_page_clear_dirty_mask().
1198  *
1199  *	This function should only be called by vm_page_dirty().
1200  */
1201 void
1202 vm_page_dirty_KBI(vm_page_t m)
1203 {
1204 
1205 	/* Refer to this operation by its public name. */
1206 	KASSERT(m->valid == VM_PAGE_BITS_ALL,
1207 	    ("vm_page_dirty: page is invalid!"));
1208 	m->dirty = VM_PAGE_BITS_ALL;
1209 }
1210 
1211 /*
1212  *	vm_page_insert:		[ internal use only ]
1213  *
1214  *	Inserts the given mem entry into the object and object list.
1215  *
1216  *	The object must be locked.
1217  */
1218 int
1219 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1220 {
1221 	vm_page_t mpred;
1222 
1223 	VM_OBJECT_ASSERT_WLOCKED(object);
1224 	mpred = vm_radix_lookup_le(&object->rtree, pindex);
1225 	return (vm_page_insert_after(m, object, pindex, mpred));
1226 }
1227 
1228 /*
1229  *	vm_page_insert_after:
1230  *
1231  *	Inserts the page "m" into the specified object at offset "pindex".
1232  *
1233  *	The page "mpred" must immediately precede the offset "pindex" within
1234  *	the specified object.
1235  *
1236  *	The object must be locked.
1237  */
1238 static int
1239 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1240     vm_page_t mpred)
1241 {
1242 	vm_page_t msucc;
1243 
1244 	VM_OBJECT_ASSERT_WLOCKED(object);
1245 	KASSERT(m->object == NULL,
1246 	    ("vm_page_insert_after: page already inserted"));
1247 	if (mpred != NULL) {
1248 		KASSERT(mpred->object == object,
1249 		    ("vm_page_insert_after: object doesn't contain mpred"));
1250 		KASSERT(mpred->pindex < pindex,
1251 		    ("vm_page_insert_after: mpred doesn't precede pindex"));
1252 		msucc = TAILQ_NEXT(mpred, listq);
1253 	} else
1254 		msucc = TAILQ_FIRST(&object->memq);
1255 	if (msucc != NULL)
1256 		KASSERT(msucc->pindex > pindex,
1257 		    ("vm_page_insert_after: msucc doesn't succeed pindex"));
1258 
1259 	/*
1260 	 * Record the object/offset pair in this page
1261 	 */
1262 	m->object = object;
1263 	m->pindex = pindex;
1264 
1265 	/*
1266 	 * Now link into the object's ordered list of backed pages.
1267 	 */
1268 	if (vm_radix_insert(&object->rtree, m)) {
1269 		m->object = NULL;
1270 		m->pindex = 0;
1271 		return (1);
1272 	}
1273 	vm_page_insert_radixdone(m, object, mpred);
1274 	return (0);
1275 }
1276 
1277 /*
1278  *	vm_page_insert_radixdone:
1279  *
1280  *	Complete page "m" insertion into the specified object after the
1281  *	radix trie hooking.
1282  *
1283  *	The page "mpred" must precede the offset "m->pindex" within the
1284  *	specified object.
1285  *
1286  *	The object must be locked.
1287  */
1288 static void
1289 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1290 {
1291 
1292 	VM_OBJECT_ASSERT_WLOCKED(object);
1293 	KASSERT(object != NULL && m->object == object,
1294 	    ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1295 	if (mpred != NULL) {
1296 		KASSERT(mpred->object == object,
1297 		    ("vm_page_insert_after: object doesn't contain mpred"));
1298 		KASSERT(mpred->pindex < m->pindex,
1299 		    ("vm_page_insert_after: mpred doesn't precede pindex"));
1300 	}
1301 
1302 	if (mpred != NULL)
1303 		TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1304 	else
1305 		TAILQ_INSERT_HEAD(&object->memq, m, listq);
1306 
1307 	/*
1308 	 * Show that the object has one more resident page.
1309 	 */
1310 	object->resident_page_count++;
1311 
1312 	/*
1313 	 * Hold the vnode until the last page is released.
1314 	 */
1315 	if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1316 		vhold(object->handle);
1317 
1318 	/*
1319 	 * Since we are inserting a new and possibly dirty page,
1320 	 * update the object's OBJ_MIGHTBEDIRTY flag.
1321 	 */
1322 	if (pmap_page_is_write_mapped(m))
1323 		vm_object_set_writeable_dirty(object);
1324 }
1325 
1326 /*
1327  *	vm_page_remove:
1328  *
1329  *	Removes the specified page from its containing object, but does not
1330  *	invalidate any backing storage.
1331  *
1332  *	The object must be locked.  The page must be locked if it is managed.
1333  */
1334 void
1335 vm_page_remove(vm_page_t m)
1336 {
1337 	vm_object_t object;
1338 	vm_page_t mrem;
1339 
1340 	if ((m->oflags & VPO_UNMANAGED) == 0)
1341 		vm_page_assert_locked(m);
1342 	if ((object = m->object) == NULL)
1343 		return;
1344 	VM_OBJECT_ASSERT_WLOCKED(object);
1345 	if (vm_page_xbusied(m))
1346 		vm_page_xunbusy_maybelocked(m);
1347 	mrem = vm_radix_remove(&object->rtree, m->pindex);
1348 	KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m));
1349 
1350 	/*
1351 	 * Now remove from the object's list of backed pages.
1352 	 */
1353 	TAILQ_REMOVE(&object->memq, m, listq);
1354 
1355 	/*
1356 	 * And show that the object has one fewer resident page.
1357 	 */
1358 	object->resident_page_count--;
1359 
1360 	/*
1361 	 * The vnode may now be recycled.
1362 	 */
1363 	if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1364 		vdrop(object->handle);
1365 
1366 	m->object = NULL;
1367 }
1368 
1369 /*
1370  *	vm_page_lookup:
1371  *
1372  *	Returns the page associated with the object/offset
1373  *	pair specified; if none is found, NULL is returned.
1374  *
1375  *	The object must be locked.
1376  */
1377 vm_page_t
1378 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1379 {
1380 
1381 	VM_OBJECT_ASSERT_LOCKED(object);
1382 	return (vm_radix_lookup(&object->rtree, pindex));
1383 }
1384 
1385 /*
1386  *	vm_page_find_least:
1387  *
1388  *	Returns the page associated with the object with least pindex
1389  *	greater than or equal to the parameter pindex, or NULL.
1390  *
1391  *	The object must be locked.
1392  */
1393 vm_page_t
1394 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1395 {
1396 	vm_page_t m;
1397 
1398 	VM_OBJECT_ASSERT_LOCKED(object);
1399 	if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1400 		m = vm_radix_lookup_ge(&object->rtree, pindex);
1401 	return (m);
1402 }
1403 
1404 /*
1405  * Returns the given page's successor (by pindex) within the object if it is
1406  * resident; if none is found, NULL is returned.
1407  *
1408  * The object must be locked.
1409  */
1410 vm_page_t
1411 vm_page_next(vm_page_t m)
1412 {
1413 	vm_page_t next;
1414 
1415 	VM_OBJECT_ASSERT_LOCKED(m->object);
1416 	if ((next = TAILQ_NEXT(m, listq)) != NULL) {
1417 		MPASS(next->object == m->object);
1418 		if (next->pindex != m->pindex + 1)
1419 			next = NULL;
1420 	}
1421 	return (next);
1422 }
1423 
1424 /*
1425  * Returns the given page's predecessor (by pindex) within the object if it is
1426  * resident; if none is found, NULL is returned.
1427  *
1428  * The object must be locked.
1429  */
1430 vm_page_t
1431 vm_page_prev(vm_page_t m)
1432 {
1433 	vm_page_t prev;
1434 
1435 	VM_OBJECT_ASSERT_LOCKED(m->object);
1436 	if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) {
1437 		MPASS(prev->object == m->object);
1438 		if (prev->pindex != m->pindex - 1)
1439 			prev = NULL;
1440 	}
1441 	return (prev);
1442 }
1443 
1444 /*
1445  * Uses the page mnew as a replacement for an existing page at index
1446  * pindex which must be already present in the object.
1447  *
1448  * The existing page must not be on a paging queue.
1449  */
1450 vm_page_t
1451 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex)
1452 {
1453 	vm_page_t mold;
1454 
1455 	VM_OBJECT_ASSERT_WLOCKED(object);
1456 	KASSERT(mnew->object == NULL,
1457 	    ("vm_page_replace: page already in object"));
1458 
1459 	/*
1460 	 * This function mostly follows vm_page_insert() and
1461 	 * vm_page_remove() without the radix, object count and vnode
1462 	 * dance.  Double check such functions for more comments.
1463 	 */
1464 
1465 	mnew->object = object;
1466 	mnew->pindex = pindex;
1467 	mold = vm_radix_replace(&object->rtree, mnew);
1468 	KASSERT(mold->queue == PQ_NONE,
1469 	    ("vm_page_replace: mold is on a paging queue"));
1470 
1471 	/* Keep the resident page list in sorted order. */
1472 	TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq);
1473 	TAILQ_REMOVE(&object->memq, mold, listq);
1474 
1475 	mold->object = NULL;
1476 	vm_page_xunbusy_maybelocked(mold);
1477 
1478 	/*
1479 	 * The object's resident_page_count does not change because we have
1480 	 * swapped one page for another, but OBJ_MIGHTBEDIRTY.
1481 	 */
1482 	if (pmap_page_is_write_mapped(mnew))
1483 		vm_object_set_writeable_dirty(object);
1484 	return (mold);
1485 }
1486 
1487 /*
1488  *	vm_page_rename:
1489  *
1490  *	Move the given memory entry from its
1491  *	current object to the specified target object/offset.
1492  *
1493  *	Note: swap associated with the page must be invalidated by the move.  We
1494  *	      have to do this for several reasons:  (1) we aren't freeing the
1495  *	      page, (2) we are dirtying the page, (3) the VM system is probably
1496  *	      moving the page from object A to B, and will then later move
1497  *	      the backing store from A to B and we can't have a conflict.
1498  *
1499  *	Note: we *always* dirty the page.  It is necessary both for the
1500  *	      fact that we moved it, and because we may be invalidating
1501  *	      swap.
1502  *
1503  *	The objects must be locked.
1504  */
1505 int
1506 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1507 {
1508 	vm_page_t mpred;
1509 	vm_pindex_t opidx;
1510 
1511 	VM_OBJECT_ASSERT_WLOCKED(new_object);
1512 
1513 	mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex);
1514 	KASSERT(mpred == NULL || mpred->pindex != new_pindex,
1515 	    ("vm_page_rename: pindex already renamed"));
1516 
1517 	/*
1518 	 * Create a custom version of vm_page_insert() which does not depend
1519 	 * by m_prev and can cheat on the implementation aspects of the
1520 	 * function.
1521 	 */
1522 	opidx = m->pindex;
1523 	m->pindex = new_pindex;
1524 	if (vm_radix_insert(&new_object->rtree, m)) {
1525 		m->pindex = opidx;
1526 		return (1);
1527 	}
1528 
1529 	/*
1530 	 * The operation cannot fail anymore.  The removal must happen before
1531 	 * the listq iterator is tainted.
1532 	 */
1533 	m->pindex = opidx;
1534 	vm_page_lock(m);
1535 	vm_page_remove(m);
1536 
1537 	/* Return back to the new pindex to complete vm_page_insert(). */
1538 	m->pindex = new_pindex;
1539 	m->object = new_object;
1540 	vm_page_unlock(m);
1541 	vm_page_insert_radixdone(m, new_object, mpred);
1542 	vm_page_dirty(m);
1543 	return (0);
1544 }
1545 
1546 /*
1547  *	vm_page_alloc:
1548  *
1549  *	Allocate and return a page that is associated with the specified
1550  *	object and offset pair.  By default, this page is exclusive busied.
1551  *
1552  *	The caller must always specify an allocation class.
1553  *
1554  *	allocation classes:
1555  *	VM_ALLOC_NORMAL		normal process request
1556  *	VM_ALLOC_SYSTEM		system *really* needs a page
1557  *	VM_ALLOC_INTERRUPT	interrupt time request
1558  *
1559  *	optional allocation flags:
1560  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
1561  *				intends to allocate
1562  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
1563  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
1564  *	VM_ALLOC_NOOBJ		page is not associated with an object and
1565  *				should not be exclusive busy
1566  *	VM_ALLOC_SBUSY		shared busy the allocated page
1567  *	VM_ALLOC_WIRED		wire the allocated page
1568  *	VM_ALLOC_ZERO		prefer a zeroed page
1569  *
1570  *	This routine may not sleep.
1571  */
1572 vm_page_t
1573 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1574 {
1575 
1576 	return (vm_page_alloc_after(object, pindex, req, object != NULL ?
1577 	    vm_radix_lookup_le(&object->rtree, pindex) : NULL));
1578 }
1579 
1580 /*
1581  * Allocate a page in the specified object with the given page index.  To
1582  * optimize insertion of the page into the object, the caller must also specifiy
1583  * the resident page in the object with largest index smaller than the given
1584  * page index, or NULL if no such page exists.
1585  */
1586 vm_page_t
1587 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex, int req,
1588     vm_page_t mpred)
1589 {
1590 	vm_page_t m;
1591 	int flags, req_class;
1592 	u_int free_count;
1593 
1594 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1595 	    (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1596 	    ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1597 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1598 	    ("inconsistent object(%p)/req(%x)", object, req));
1599 	KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
1600 	    ("Can't sleep and retry object insertion."));
1601 	KASSERT(mpred == NULL || mpred->pindex < pindex,
1602 	    ("mpred %p doesn't precede pindex 0x%jx", mpred,
1603 	    (uintmax_t)pindex));
1604 	if (object != NULL)
1605 		VM_OBJECT_ASSERT_WLOCKED(object);
1606 
1607 	req_class = req & VM_ALLOC_CLASS_MASK;
1608 
1609 	/*
1610 	 * The page daemon is allowed to dig deeper into the free page list.
1611 	 */
1612 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1613 		req_class = VM_ALLOC_SYSTEM;
1614 
1615 	/*
1616 	 * Allocate a page if the number of free pages exceeds the minimum
1617 	 * for the request class.
1618 	 */
1619 again:
1620 	mtx_lock(&vm_page_queue_free_mtx);
1621 	if (vm_cnt.v_free_count > vm_cnt.v_free_reserved ||
1622 	    (req_class == VM_ALLOC_SYSTEM &&
1623 	    vm_cnt.v_free_count > vm_cnt.v_interrupt_free_min) ||
1624 	    (req_class == VM_ALLOC_INTERRUPT &&
1625 	    vm_cnt.v_free_count > 0)) {
1626 		/*
1627 		 * Can we allocate the page from a reservation?
1628 		 */
1629 #if VM_NRESERVLEVEL > 0
1630 		if (object == NULL || (object->flags & (OBJ_COLORED |
1631 		    OBJ_FICTITIOUS)) != OBJ_COLORED || (m =
1632 		    vm_reserv_alloc_page(object, pindex, mpred)) == NULL)
1633 #endif
1634 		{
1635 			/*
1636 			 * If not, allocate it from the free page queues.
1637 			 */
1638 			m = vm_phys_alloc_pages(object != NULL ?
1639 			    VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0);
1640 #if VM_NRESERVLEVEL > 0
1641 			if (m == NULL && vm_reserv_reclaim_inactive()) {
1642 				m = vm_phys_alloc_pages(object != NULL ?
1643 				    VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT,
1644 				    0);
1645 			}
1646 #endif
1647 		}
1648 	} else {
1649 		/*
1650 		 * Not allocatable, give up.
1651 		 */
1652 		if (vm_page_alloc_fail(object, req))
1653 			goto again;
1654 		return (NULL);
1655 	}
1656 
1657 	/*
1658 	 *  At this point we had better have found a good page.
1659 	 */
1660 	KASSERT(m != NULL, ("missing page"));
1661 	free_count = vm_phys_freecnt_adj(m, -1);
1662 	mtx_unlock(&vm_page_queue_free_mtx);
1663 	vm_page_alloc_check(m);
1664 
1665 	/*
1666 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
1667 	 */
1668 	flags = 0;
1669 	if ((req & VM_ALLOC_ZERO) != 0)
1670 		flags = PG_ZERO;
1671 	flags &= m->flags;
1672 	if ((req & VM_ALLOC_NODUMP) != 0)
1673 		flags |= PG_NODUMP;
1674 	m->flags = flags;
1675 	m->aflags = 0;
1676 	m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
1677 	    VPO_UNMANAGED : 0;
1678 	m->busy_lock = VPB_UNBUSIED;
1679 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
1680 		m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1681 	if ((req & VM_ALLOC_SBUSY) != 0)
1682 		m->busy_lock = VPB_SHARERS_WORD(1);
1683 	if (req & VM_ALLOC_WIRED) {
1684 		/*
1685 		 * The page lock is not required for wiring a page until that
1686 		 * page is inserted into the object.
1687 		 */
1688 		atomic_add_int(&vm_cnt.v_wire_count, 1);
1689 		m->wire_count = 1;
1690 	}
1691 	m->act_count = 0;
1692 
1693 	if (object != NULL) {
1694 		if (vm_page_insert_after(m, object, pindex, mpred)) {
1695 			pagedaemon_wakeup();
1696 			if (req & VM_ALLOC_WIRED) {
1697 				atomic_subtract_int(&vm_cnt.v_wire_count, 1);
1698 				m->wire_count = 0;
1699 			}
1700 			KASSERT(m->object == NULL, ("page %p has object", m));
1701 			m->oflags = VPO_UNMANAGED;
1702 			m->busy_lock = VPB_UNBUSIED;
1703 			/* Don't change PG_ZERO. */
1704 			vm_page_free_toq(m);
1705 			if (req & VM_ALLOC_WAITFAIL) {
1706 				VM_OBJECT_WUNLOCK(object);
1707 				vm_radix_wait();
1708 				VM_OBJECT_WLOCK(object);
1709 			}
1710 			return (NULL);
1711 		}
1712 
1713 		/* Ignore device objects; the pager sets "memattr" for them. */
1714 		if (object->memattr != VM_MEMATTR_DEFAULT &&
1715 		    (object->flags & OBJ_FICTITIOUS) == 0)
1716 			pmap_page_set_memattr(m, object->memattr);
1717 	} else
1718 		m->pindex = pindex;
1719 
1720 	/*
1721 	 * Don't wakeup too often - wakeup the pageout daemon when
1722 	 * we would be nearly out of memory.
1723 	 */
1724 	if (vm_paging_needed(free_count))
1725 		pagedaemon_wakeup();
1726 
1727 	return (m);
1728 }
1729 
1730 /*
1731  *	vm_page_alloc_contig:
1732  *
1733  *	Allocate a contiguous set of physical pages of the given size "npages"
1734  *	from the free lists.  All of the physical pages must be at or above
1735  *	the given physical address "low" and below the given physical address
1736  *	"high".  The given value "alignment" determines the alignment of the
1737  *	first physical page in the set.  If the given value "boundary" is
1738  *	non-zero, then the set of physical pages cannot cross any physical
1739  *	address boundary that is a multiple of that value.  Both "alignment"
1740  *	and "boundary" must be a power of two.
1741  *
1742  *	If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
1743  *	then the memory attribute setting for the physical pages is configured
1744  *	to the object's memory attribute setting.  Otherwise, the memory
1745  *	attribute setting for the physical pages is configured to "memattr",
1746  *	overriding the object's memory attribute setting.  However, if the
1747  *	object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
1748  *	memory attribute setting for the physical pages cannot be configured
1749  *	to VM_MEMATTR_DEFAULT.
1750  *
1751  *	The specified object may not contain fictitious pages.
1752  *
1753  *	The caller must always specify an allocation class.
1754  *
1755  *	allocation classes:
1756  *	VM_ALLOC_NORMAL		normal process request
1757  *	VM_ALLOC_SYSTEM		system *really* needs a page
1758  *	VM_ALLOC_INTERRUPT	interrupt time request
1759  *
1760  *	optional allocation flags:
1761  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
1762  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
1763  *	VM_ALLOC_NOOBJ		page is not associated with an object and
1764  *				should not be exclusive busy
1765  *	VM_ALLOC_SBUSY		shared busy the allocated page
1766  *	VM_ALLOC_WIRED		wire the allocated page
1767  *	VM_ALLOC_ZERO		prefer a zeroed page
1768  *
1769  *	This routine may not sleep.
1770  */
1771 vm_page_t
1772 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
1773     u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
1774     vm_paddr_t boundary, vm_memattr_t memattr)
1775 {
1776 	vm_page_t m, m_ret, mpred;
1777 	u_int busy_lock, flags, oflags;
1778 	int req_class;
1779 
1780 	mpred = NULL;	/* XXX: pacify gcc */
1781 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1782 	    (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1783 	    ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1784 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1785 	    ("vm_page_alloc_contig: inconsistent object(%p)/req(%x)", object,
1786 	    req));
1787 	KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
1788 	    ("Can't sleep and retry object insertion."));
1789 	if (object != NULL) {
1790 		VM_OBJECT_ASSERT_WLOCKED(object);
1791 		KASSERT((object->flags & OBJ_FICTITIOUS) == 0,
1792 		    ("vm_page_alloc_contig: object %p has fictitious pages",
1793 		    object));
1794 	}
1795 	KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
1796 	req_class = req & VM_ALLOC_CLASS_MASK;
1797 
1798 	/*
1799 	 * The page daemon is allowed to dig deeper into the free page list.
1800 	 */
1801 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1802 		req_class = VM_ALLOC_SYSTEM;
1803 
1804 	if (object != NULL) {
1805 		mpred = vm_radix_lookup_le(&object->rtree, pindex);
1806 		KASSERT(mpred == NULL || mpred->pindex != pindex,
1807 		    ("vm_page_alloc_contig: pindex already allocated"));
1808 	}
1809 
1810 	/*
1811 	 * Can we allocate the pages without the number of free pages falling
1812 	 * below the lower bound for the allocation class?
1813 	 */
1814 again:
1815 	mtx_lock(&vm_page_queue_free_mtx);
1816 	if (vm_cnt.v_free_count >= npages + vm_cnt.v_free_reserved ||
1817 	    (req_class == VM_ALLOC_SYSTEM &&
1818 	    vm_cnt.v_free_count >= npages + vm_cnt.v_interrupt_free_min) ||
1819 	    (req_class == VM_ALLOC_INTERRUPT &&
1820 	    vm_cnt.v_free_count >= npages)) {
1821 		/*
1822 		 * Can we allocate the pages from a reservation?
1823 		 */
1824 #if VM_NRESERVLEVEL > 0
1825 retry:
1826 		if (object == NULL || (object->flags & OBJ_COLORED) == 0 ||
1827 		    (m_ret = vm_reserv_alloc_contig(object, pindex, npages,
1828 		    low, high, alignment, boundary, mpred)) == NULL)
1829 #endif
1830 			/*
1831 			 * If not, allocate them from the free page queues.
1832 			 */
1833 			m_ret = vm_phys_alloc_contig(npages, low, high,
1834 			    alignment, boundary);
1835 	} else {
1836 		if (vm_page_alloc_fail(object, req))
1837 			goto again;
1838 		return (NULL);
1839 	}
1840 	if (m_ret != NULL)
1841 		vm_phys_freecnt_adj(m_ret, -npages);
1842 	else {
1843 #if VM_NRESERVLEVEL > 0
1844 		if (vm_reserv_reclaim_contig(npages, low, high, alignment,
1845 		    boundary))
1846 			goto retry;
1847 #endif
1848 	}
1849 	mtx_unlock(&vm_page_queue_free_mtx);
1850 	if (m_ret == NULL)
1851 		return (NULL);
1852 	for (m = m_ret; m < &m_ret[npages]; m++)
1853 		vm_page_alloc_check(m);
1854 
1855 	/*
1856 	 * Initialize the pages.  Only the PG_ZERO flag is inherited.
1857 	 */
1858 	flags = 0;
1859 	if ((req & VM_ALLOC_ZERO) != 0)
1860 		flags = PG_ZERO;
1861 	if ((req & VM_ALLOC_NODUMP) != 0)
1862 		flags |= PG_NODUMP;
1863 	oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
1864 	    VPO_UNMANAGED : 0;
1865 	busy_lock = VPB_UNBUSIED;
1866 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
1867 		busy_lock = VPB_SINGLE_EXCLUSIVER;
1868 	if ((req & VM_ALLOC_SBUSY) != 0)
1869 		busy_lock = VPB_SHARERS_WORD(1);
1870 	if ((req & VM_ALLOC_WIRED) != 0)
1871 		atomic_add_int(&vm_cnt.v_wire_count, npages);
1872 	if (object != NULL) {
1873 		if (object->memattr != VM_MEMATTR_DEFAULT &&
1874 		    memattr == VM_MEMATTR_DEFAULT)
1875 			memattr = object->memattr;
1876 	}
1877 	for (m = m_ret; m < &m_ret[npages]; m++) {
1878 		m->aflags = 0;
1879 		m->flags = (m->flags | PG_NODUMP) & flags;
1880 		m->busy_lock = busy_lock;
1881 		if ((req & VM_ALLOC_WIRED) != 0)
1882 			m->wire_count = 1;
1883 		m->act_count = 0;
1884 		m->oflags = oflags;
1885 		if (object != NULL) {
1886 			if (vm_page_insert_after(m, object, pindex, mpred)) {
1887 				pagedaemon_wakeup();
1888 				if ((req & VM_ALLOC_WIRED) != 0)
1889 					atomic_subtract_int(
1890 					    &vm_cnt.v_wire_count, npages);
1891 				KASSERT(m->object == NULL,
1892 				    ("page %p has object", m));
1893 				mpred = m;
1894 				for (m = m_ret; m < &m_ret[npages]; m++) {
1895 					if (m <= mpred &&
1896 					    (req & VM_ALLOC_WIRED) != 0)
1897 						m->wire_count = 0;
1898 					m->oflags = VPO_UNMANAGED;
1899 					m->busy_lock = VPB_UNBUSIED;
1900 					/* Don't change PG_ZERO. */
1901 					vm_page_free_toq(m);
1902 				}
1903 				if (req & VM_ALLOC_WAITFAIL) {
1904 					VM_OBJECT_WUNLOCK(object);
1905 					vm_radix_wait();
1906 					VM_OBJECT_WLOCK(object);
1907 				}
1908 				return (NULL);
1909 			}
1910 			mpred = m;
1911 		} else
1912 			m->pindex = pindex;
1913 		if (memattr != VM_MEMATTR_DEFAULT)
1914 			pmap_page_set_memattr(m, memattr);
1915 		pindex++;
1916 	}
1917 	if (vm_paging_needed(vm_cnt.v_free_count))
1918 		pagedaemon_wakeup();
1919 	return (m_ret);
1920 }
1921 
1922 /*
1923  * Check a page that has been freshly dequeued from a freelist.
1924  */
1925 static void
1926 vm_page_alloc_check(vm_page_t m)
1927 {
1928 
1929 	KASSERT(m->object == NULL, ("page %p has object", m));
1930 	KASSERT(m->queue == PQ_NONE,
1931 	    ("page %p has unexpected queue %d", m, m->queue));
1932 	KASSERT(m->wire_count == 0, ("page %p is wired", m));
1933 	KASSERT(m->hold_count == 0, ("page %p is held", m));
1934 	KASSERT(!vm_page_busied(m), ("page %p is busy", m));
1935 	KASSERT(m->dirty == 0, ("page %p is dirty", m));
1936 	KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
1937 	    ("page %p has unexpected memattr %d",
1938 	    m, pmap_page_get_memattr(m)));
1939 	KASSERT(m->valid == 0, ("free page %p is valid", m));
1940 }
1941 
1942 /*
1943  * 	vm_page_alloc_freelist:
1944  *
1945  *	Allocate a physical page from the specified free page list.
1946  *
1947  *	The caller must always specify an allocation class.
1948  *
1949  *	allocation classes:
1950  *	VM_ALLOC_NORMAL		normal process request
1951  *	VM_ALLOC_SYSTEM		system *really* needs a page
1952  *	VM_ALLOC_INTERRUPT	interrupt time request
1953  *
1954  *	optional allocation flags:
1955  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
1956  *				intends to allocate
1957  *	VM_ALLOC_WIRED		wire the allocated page
1958  *	VM_ALLOC_ZERO		prefer a zeroed page
1959  *
1960  *	This routine may not sleep.
1961  */
1962 vm_page_t
1963 vm_page_alloc_freelist(int flind, int req)
1964 {
1965 	vm_page_t m;
1966 	u_int flags, free_count;
1967 	int req_class;
1968 
1969 	req_class = req & VM_ALLOC_CLASS_MASK;
1970 
1971 	/*
1972 	 * The page daemon is allowed to dig deeper into the free page list.
1973 	 */
1974 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1975 		req_class = VM_ALLOC_SYSTEM;
1976 
1977 	/*
1978 	 * Do not allocate reserved pages unless the req has asked for it.
1979 	 */
1980 again:
1981 	mtx_lock(&vm_page_queue_free_mtx);
1982 	if (vm_cnt.v_free_count > vm_cnt.v_free_reserved ||
1983 	    (req_class == VM_ALLOC_SYSTEM &&
1984 	    vm_cnt.v_free_count > vm_cnt.v_interrupt_free_min) ||
1985 	    (req_class == VM_ALLOC_INTERRUPT &&
1986 	    vm_cnt.v_free_count > 0)) {
1987 		m = vm_phys_alloc_freelist_pages(flind, VM_FREEPOOL_DIRECT, 0);
1988 	} else {
1989 		if (vm_page_alloc_fail(NULL, req))
1990 			goto again;
1991 		return (NULL);
1992 	}
1993 	if (m == NULL) {
1994 		mtx_unlock(&vm_page_queue_free_mtx);
1995 		return (NULL);
1996 	}
1997 	free_count = vm_phys_freecnt_adj(m, -1);
1998 	mtx_unlock(&vm_page_queue_free_mtx);
1999 	vm_page_alloc_check(m);
2000 
2001 	/*
2002 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
2003 	 */
2004 	m->aflags = 0;
2005 	flags = 0;
2006 	if ((req & VM_ALLOC_ZERO) != 0)
2007 		flags = PG_ZERO;
2008 	m->flags &= flags;
2009 	if ((req & VM_ALLOC_WIRED) != 0) {
2010 		/*
2011 		 * The page lock is not required for wiring a page that does
2012 		 * not belong to an object.
2013 		 */
2014 		atomic_add_int(&vm_cnt.v_wire_count, 1);
2015 		m->wire_count = 1;
2016 	}
2017 	/* Unmanaged pages don't use "act_count". */
2018 	m->oflags = VPO_UNMANAGED;
2019 	if (vm_paging_needed(free_count))
2020 		pagedaemon_wakeup();
2021 	return (m);
2022 }
2023 
2024 #define	VPSC_ANY	0	/* No restrictions. */
2025 #define	VPSC_NORESERV	1	/* Skip reservations; implies VPSC_NOSUPER. */
2026 #define	VPSC_NOSUPER	2	/* Skip superpages. */
2027 
2028 /*
2029  *	vm_page_scan_contig:
2030  *
2031  *	Scan vm_page_array[] between the specified entries "m_start" and
2032  *	"m_end" for a run of contiguous physical pages that satisfy the
2033  *	specified conditions, and return the lowest page in the run.  The
2034  *	specified "alignment" determines the alignment of the lowest physical
2035  *	page in the run.  If the specified "boundary" is non-zero, then the
2036  *	run of physical pages cannot span a physical address that is a
2037  *	multiple of "boundary".
2038  *
2039  *	"m_end" is never dereferenced, so it need not point to a vm_page
2040  *	structure within vm_page_array[].
2041  *
2042  *	"npages" must be greater than zero.  "m_start" and "m_end" must not
2043  *	span a hole (or discontiguity) in the physical address space.  Both
2044  *	"alignment" and "boundary" must be a power of two.
2045  */
2046 vm_page_t
2047 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end,
2048     u_long alignment, vm_paddr_t boundary, int options)
2049 {
2050 	struct mtx *m_mtx;
2051 	vm_object_t object;
2052 	vm_paddr_t pa;
2053 	vm_page_t m, m_run;
2054 #if VM_NRESERVLEVEL > 0
2055 	int level;
2056 #endif
2057 	int m_inc, order, run_ext, run_len;
2058 
2059 	KASSERT(npages > 0, ("npages is 0"));
2060 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2061 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2062 	m_run = NULL;
2063 	run_len = 0;
2064 	m_mtx = NULL;
2065 	for (m = m_start; m < m_end && run_len < npages; m += m_inc) {
2066 		KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2067 		    ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2068 
2069 		/*
2070 		 * If the current page would be the start of a run, check its
2071 		 * physical address against the end, alignment, and boundary
2072 		 * conditions.  If it doesn't satisfy these conditions, either
2073 		 * terminate the scan or advance to the next page that
2074 		 * satisfies the failed condition.
2075 		 */
2076 		if (run_len == 0) {
2077 			KASSERT(m_run == NULL, ("m_run != NULL"));
2078 			if (m + npages > m_end)
2079 				break;
2080 			pa = VM_PAGE_TO_PHYS(m);
2081 			if ((pa & (alignment - 1)) != 0) {
2082 				m_inc = atop(roundup2(pa, alignment) - pa);
2083 				continue;
2084 			}
2085 			if (rounddown2(pa ^ (pa + ptoa(npages) - 1),
2086 			    boundary) != 0) {
2087 				m_inc = atop(roundup2(pa, boundary) - pa);
2088 				continue;
2089 			}
2090 		} else
2091 			KASSERT(m_run != NULL, ("m_run == NULL"));
2092 
2093 		vm_page_change_lock(m, &m_mtx);
2094 		m_inc = 1;
2095 retry:
2096 		if (m->wire_count != 0 || m->hold_count != 0)
2097 			run_ext = 0;
2098 #if VM_NRESERVLEVEL > 0
2099 		else if ((level = vm_reserv_level(m)) >= 0 &&
2100 		    (options & VPSC_NORESERV) != 0) {
2101 			run_ext = 0;
2102 			/* Advance to the end of the reservation. */
2103 			pa = VM_PAGE_TO_PHYS(m);
2104 			m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) -
2105 			    pa);
2106 		}
2107 #endif
2108 		else if ((object = m->object) != NULL) {
2109 			/*
2110 			 * The page is considered eligible for relocation if
2111 			 * and only if it could be laundered or reclaimed by
2112 			 * the page daemon.
2113 			 */
2114 			if (!VM_OBJECT_TRYRLOCK(object)) {
2115 				mtx_unlock(m_mtx);
2116 				VM_OBJECT_RLOCK(object);
2117 				mtx_lock(m_mtx);
2118 				if (m->object != object) {
2119 					/*
2120 					 * The page may have been freed.
2121 					 */
2122 					VM_OBJECT_RUNLOCK(object);
2123 					goto retry;
2124 				} else if (m->wire_count != 0 ||
2125 				    m->hold_count != 0) {
2126 					run_ext = 0;
2127 					goto unlock;
2128 				}
2129 			}
2130 			KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2131 			    ("page %p is PG_UNHOLDFREE", m));
2132 			/* Don't care: PG_NODUMP, PG_ZERO. */
2133 			if (object->type != OBJT_DEFAULT &&
2134 			    object->type != OBJT_SWAP &&
2135 			    object->type != OBJT_VNODE) {
2136 				run_ext = 0;
2137 #if VM_NRESERVLEVEL > 0
2138 			} else if ((options & VPSC_NOSUPER) != 0 &&
2139 			    (level = vm_reserv_level_iffullpop(m)) >= 0) {
2140 				run_ext = 0;
2141 				/* Advance to the end of the superpage. */
2142 				pa = VM_PAGE_TO_PHYS(m);
2143 				m_inc = atop(roundup2(pa + 1,
2144 				    vm_reserv_size(level)) - pa);
2145 #endif
2146 			} else if (object->memattr == VM_MEMATTR_DEFAULT &&
2147 			    m->queue != PQ_NONE && !vm_page_busied(m)) {
2148 				/*
2149 				 * The page is allocated but eligible for
2150 				 * relocation.  Extend the current run by one
2151 				 * page.
2152 				 */
2153 				KASSERT(pmap_page_get_memattr(m) ==
2154 				    VM_MEMATTR_DEFAULT,
2155 				    ("page %p has an unexpected memattr", m));
2156 				KASSERT((m->oflags & (VPO_SWAPINPROG |
2157 				    VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2158 				    ("page %p has unexpected oflags", m));
2159 				/* Don't care: VPO_NOSYNC. */
2160 				run_ext = 1;
2161 			} else
2162 				run_ext = 0;
2163 unlock:
2164 			VM_OBJECT_RUNLOCK(object);
2165 #if VM_NRESERVLEVEL > 0
2166 		} else if (level >= 0) {
2167 			/*
2168 			 * The page is reserved but not yet allocated.  In
2169 			 * other words, it is still free.  Extend the current
2170 			 * run by one page.
2171 			 */
2172 			run_ext = 1;
2173 #endif
2174 		} else if ((order = m->order) < VM_NFREEORDER) {
2175 			/*
2176 			 * The page is enqueued in the physical memory
2177 			 * allocator's free page queues.  Moreover, it is the
2178 			 * first page in a power-of-two-sized run of
2179 			 * contiguous free pages.  Add these pages to the end
2180 			 * of the current run, and jump ahead.
2181 			 */
2182 			run_ext = 1 << order;
2183 			m_inc = 1 << order;
2184 		} else {
2185 			/*
2186 			 * Skip the page for one of the following reasons: (1)
2187 			 * It is enqueued in the physical memory allocator's
2188 			 * free page queues.  However, it is not the first
2189 			 * page in a run of contiguous free pages.  (This case
2190 			 * rarely occurs because the scan is performed in
2191 			 * ascending order.) (2) It is not reserved, and it is
2192 			 * transitioning from free to allocated.  (Conversely,
2193 			 * the transition from allocated to free for managed
2194 			 * pages is blocked by the page lock.) (3) It is
2195 			 * allocated but not contained by an object and not
2196 			 * wired, e.g., allocated by Xen's balloon driver.
2197 			 */
2198 			run_ext = 0;
2199 		}
2200 
2201 		/*
2202 		 * Extend or reset the current run of pages.
2203 		 */
2204 		if (run_ext > 0) {
2205 			if (run_len == 0)
2206 				m_run = m;
2207 			run_len += run_ext;
2208 		} else {
2209 			if (run_len > 0) {
2210 				m_run = NULL;
2211 				run_len = 0;
2212 			}
2213 		}
2214 	}
2215 	if (m_mtx != NULL)
2216 		mtx_unlock(m_mtx);
2217 	if (run_len >= npages)
2218 		return (m_run);
2219 	return (NULL);
2220 }
2221 
2222 /*
2223  *	vm_page_reclaim_run:
2224  *
2225  *	Try to relocate each of the allocated virtual pages within the
2226  *	specified run of physical pages to a new physical address.  Free the
2227  *	physical pages underlying the relocated virtual pages.  A virtual page
2228  *	is relocatable if and only if it could be laundered or reclaimed by
2229  *	the page daemon.  Whenever possible, a virtual page is relocated to a
2230  *	physical address above "high".
2231  *
2232  *	Returns 0 if every physical page within the run was already free or
2233  *	just freed by a successful relocation.  Otherwise, returns a non-zero
2234  *	value indicating why the last attempt to relocate a virtual page was
2235  *	unsuccessful.
2236  *
2237  *	"req_class" must be an allocation class.
2238  */
2239 static int
2240 vm_page_reclaim_run(int req_class, u_long npages, vm_page_t m_run,
2241     vm_paddr_t high)
2242 {
2243 	struct mtx *m_mtx;
2244 	struct spglist free;
2245 	vm_object_t object;
2246 	vm_paddr_t pa;
2247 	vm_page_t m, m_end, m_new;
2248 	int error, order, req;
2249 
2250 	KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class,
2251 	    ("req_class is not an allocation class"));
2252 	SLIST_INIT(&free);
2253 	error = 0;
2254 	m = m_run;
2255 	m_end = m_run + npages;
2256 	m_mtx = NULL;
2257 	for (; error == 0 && m < m_end; m++) {
2258 		KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2259 		    ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2260 
2261 		/*
2262 		 * Avoid releasing and reacquiring the same page lock.
2263 		 */
2264 		vm_page_change_lock(m, &m_mtx);
2265 retry:
2266 		if (m->wire_count != 0 || m->hold_count != 0)
2267 			error = EBUSY;
2268 		else if ((object = m->object) != NULL) {
2269 			/*
2270 			 * The page is relocated if and only if it could be
2271 			 * laundered or reclaimed by the page daemon.
2272 			 */
2273 			if (!VM_OBJECT_TRYWLOCK(object)) {
2274 				mtx_unlock(m_mtx);
2275 				VM_OBJECT_WLOCK(object);
2276 				mtx_lock(m_mtx);
2277 				if (m->object != object) {
2278 					/*
2279 					 * The page may have been freed.
2280 					 */
2281 					VM_OBJECT_WUNLOCK(object);
2282 					goto retry;
2283 				} else if (m->wire_count != 0 ||
2284 				    m->hold_count != 0) {
2285 					error = EBUSY;
2286 					goto unlock;
2287 				}
2288 			}
2289 			KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2290 			    ("page %p is PG_UNHOLDFREE", m));
2291 			/* Don't care: PG_NODUMP, PG_ZERO. */
2292 			if (object->type != OBJT_DEFAULT &&
2293 			    object->type != OBJT_SWAP &&
2294 			    object->type != OBJT_VNODE)
2295 				error = EINVAL;
2296 			else if (object->memattr != VM_MEMATTR_DEFAULT)
2297 				error = EINVAL;
2298 			else if (m->queue != PQ_NONE && !vm_page_busied(m)) {
2299 				KASSERT(pmap_page_get_memattr(m) ==
2300 				    VM_MEMATTR_DEFAULT,
2301 				    ("page %p has an unexpected memattr", m));
2302 				KASSERT((m->oflags & (VPO_SWAPINPROG |
2303 				    VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2304 				    ("page %p has unexpected oflags", m));
2305 				/* Don't care: VPO_NOSYNC. */
2306 				if (m->valid != 0) {
2307 					/*
2308 					 * First, try to allocate a new page
2309 					 * that is above "high".  Failing
2310 					 * that, try to allocate a new page
2311 					 * that is below "m_run".  Allocate
2312 					 * the new page between the end of
2313 					 * "m_run" and "high" only as a last
2314 					 * resort.
2315 					 */
2316 					req = req_class | VM_ALLOC_NOOBJ;
2317 					if ((m->flags & PG_NODUMP) != 0)
2318 						req |= VM_ALLOC_NODUMP;
2319 					if (trunc_page(high) !=
2320 					    ~(vm_paddr_t)PAGE_MASK) {
2321 						m_new = vm_page_alloc_contig(
2322 						    NULL, 0, req, 1,
2323 						    round_page(high),
2324 						    ~(vm_paddr_t)0,
2325 						    PAGE_SIZE, 0,
2326 						    VM_MEMATTR_DEFAULT);
2327 					} else
2328 						m_new = NULL;
2329 					if (m_new == NULL) {
2330 						pa = VM_PAGE_TO_PHYS(m_run);
2331 						m_new = vm_page_alloc_contig(
2332 						    NULL, 0, req, 1,
2333 						    0, pa - 1, PAGE_SIZE, 0,
2334 						    VM_MEMATTR_DEFAULT);
2335 					}
2336 					if (m_new == NULL) {
2337 						pa += ptoa(npages);
2338 						m_new = vm_page_alloc_contig(
2339 						    NULL, 0, req, 1,
2340 						    pa, high, PAGE_SIZE, 0,
2341 						    VM_MEMATTR_DEFAULT);
2342 					}
2343 					if (m_new == NULL) {
2344 						error = ENOMEM;
2345 						goto unlock;
2346 					}
2347 					KASSERT(m_new->wire_count == 0,
2348 					    ("page %p is wired", m));
2349 
2350 					/*
2351 					 * Replace "m" with the new page.  For
2352 					 * vm_page_replace(), "m" must be busy
2353 					 * and dequeued.  Finally, change "m"
2354 					 * as if vm_page_free() was called.
2355 					 */
2356 					if (object->ref_count != 0)
2357 						pmap_remove_all(m);
2358 					m_new->aflags = m->aflags;
2359 					KASSERT(m_new->oflags == VPO_UNMANAGED,
2360 					    ("page %p is managed", m));
2361 					m_new->oflags = m->oflags & VPO_NOSYNC;
2362 					pmap_copy_page(m, m_new);
2363 					m_new->valid = m->valid;
2364 					m_new->dirty = m->dirty;
2365 					m->flags &= ~PG_ZERO;
2366 					vm_page_xbusy(m);
2367 					vm_page_remque(m);
2368 					vm_page_replace_checked(m_new, object,
2369 					    m->pindex, m);
2370 					m->valid = 0;
2371 					vm_page_undirty(m);
2372 
2373 					/*
2374 					 * The new page must be deactivated
2375 					 * before the object is unlocked.
2376 					 */
2377 					vm_page_change_lock(m_new, &m_mtx);
2378 					vm_page_deactivate(m_new);
2379 				} else {
2380 					m->flags &= ~PG_ZERO;
2381 					vm_page_remque(m);
2382 					vm_page_remove(m);
2383 					KASSERT(m->dirty == 0,
2384 					    ("page %p is dirty", m));
2385 				}
2386 				SLIST_INSERT_HEAD(&free, m, plinks.s.ss);
2387 			} else
2388 				error = EBUSY;
2389 unlock:
2390 			VM_OBJECT_WUNLOCK(object);
2391 		} else {
2392 			mtx_lock(&vm_page_queue_free_mtx);
2393 			order = m->order;
2394 			if (order < VM_NFREEORDER) {
2395 				/*
2396 				 * The page is enqueued in the physical memory
2397 				 * allocator's free page queues.  Moreover, it
2398 				 * is the first page in a power-of-two-sized
2399 				 * run of contiguous free pages.  Jump ahead
2400 				 * to the last page within that run, and
2401 				 * continue from there.
2402 				 */
2403 				m += (1 << order) - 1;
2404 			}
2405 #if VM_NRESERVLEVEL > 0
2406 			else if (vm_reserv_is_page_free(m))
2407 				order = 0;
2408 #endif
2409 			mtx_unlock(&vm_page_queue_free_mtx);
2410 			if (order == VM_NFREEORDER)
2411 				error = EINVAL;
2412 		}
2413 	}
2414 	if (m_mtx != NULL)
2415 		mtx_unlock(m_mtx);
2416 	if ((m = SLIST_FIRST(&free)) != NULL) {
2417 		mtx_lock(&vm_page_queue_free_mtx);
2418 		do {
2419 			SLIST_REMOVE_HEAD(&free, plinks.s.ss);
2420 			vm_page_free_phys(m);
2421 		} while ((m = SLIST_FIRST(&free)) != NULL);
2422 		vm_page_free_wakeup();
2423 		mtx_unlock(&vm_page_queue_free_mtx);
2424 	}
2425 	return (error);
2426 }
2427 
2428 #define	NRUNS	16
2429 
2430 CTASSERT(powerof2(NRUNS));
2431 
2432 #define	RUN_INDEX(count)	((count) & (NRUNS - 1))
2433 
2434 #define	MIN_RECLAIM	8
2435 
2436 /*
2437  *	vm_page_reclaim_contig:
2438  *
2439  *	Reclaim allocated, contiguous physical memory satisfying the specified
2440  *	conditions by relocating the virtual pages using that physical memory.
2441  *	Returns true if reclamation is successful and false otherwise.  Since
2442  *	relocation requires the allocation of physical pages, reclamation may
2443  *	fail due to a shortage of free pages.  When reclamation fails, callers
2444  *	are expected to perform VM_WAIT before retrying a failed allocation
2445  *	operation, e.g., vm_page_alloc_contig().
2446  *
2447  *	The caller must always specify an allocation class through "req".
2448  *
2449  *	allocation classes:
2450  *	VM_ALLOC_NORMAL		normal process request
2451  *	VM_ALLOC_SYSTEM		system *really* needs a page
2452  *	VM_ALLOC_INTERRUPT	interrupt time request
2453  *
2454  *	The optional allocation flags are ignored.
2455  *
2456  *	"npages" must be greater than zero.  Both "alignment" and "boundary"
2457  *	must be a power of two.
2458  */
2459 bool
2460 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high,
2461     u_long alignment, vm_paddr_t boundary)
2462 {
2463 	vm_paddr_t curr_low;
2464 	vm_page_t m_run, m_runs[NRUNS];
2465 	u_long count, reclaimed;
2466 	int error, i, options, req_class;
2467 
2468 	KASSERT(npages > 0, ("npages is 0"));
2469 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2470 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2471 	req_class = req & VM_ALLOC_CLASS_MASK;
2472 
2473 	/*
2474 	 * The page daemon is allowed to dig deeper into the free page list.
2475 	 */
2476 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2477 		req_class = VM_ALLOC_SYSTEM;
2478 
2479 	/*
2480 	 * Return if the number of free pages cannot satisfy the requested
2481 	 * allocation.
2482 	 */
2483 	count = vm_cnt.v_free_count;
2484 	if (count < npages + vm_cnt.v_free_reserved || (count < npages +
2485 	    vm_cnt.v_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) ||
2486 	    (count < npages && req_class == VM_ALLOC_INTERRUPT))
2487 		return (false);
2488 
2489 	/*
2490 	 * Scan up to three times, relaxing the restrictions ("options") on
2491 	 * the reclamation of reservations and superpages each time.
2492 	 */
2493 	for (options = VPSC_NORESERV;;) {
2494 		/*
2495 		 * Find the highest runs that satisfy the given constraints
2496 		 * and restrictions, and record them in "m_runs".
2497 		 */
2498 		curr_low = low;
2499 		count = 0;
2500 		for (;;) {
2501 			m_run = vm_phys_scan_contig(npages, curr_low, high,
2502 			    alignment, boundary, options);
2503 			if (m_run == NULL)
2504 				break;
2505 			curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages);
2506 			m_runs[RUN_INDEX(count)] = m_run;
2507 			count++;
2508 		}
2509 
2510 		/*
2511 		 * Reclaim the highest runs in LIFO (descending) order until
2512 		 * the number of reclaimed pages, "reclaimed", is at least
2513 		 * MIN_RECLAIM.  Reset "reclaimed" each time because each
2514 		 * reclamation is idempotent, and runs will (likely) recur
2515 		 * from one scan to the next as restrictions are relaxed.
2516 		 */
2517 		reclaimed = 0;
2518 		for (i = 0; count > 0 && i < NRUNS; i++) {
2519 			count--;
2520 			m_run = m_runs[RUN_INDEX(count)];
2521 			error = vm_page_reclaim_run(req_class, npages, m_run,
2522 			    high);
2523 			if (error == 0) {
2524 				reclaimed += npages;
2525 				if (reclaimed >= MIN_RECLAIM)
2526 					return (true);
2527 			}
2528 		}
2529 
2530 		/*
2531 		 * Either relax the restrictions on the next scan or return if
2532 		 * the last scan had no restrictions.
2533 		 */
2534 		if (options == VPSC_NORESERV)
2535 			options = VPSC_NOSUPER;
2536 		else if (options == VPSC_NOSUPER)
2537 			options = VPSC_ANY;
2538 		else if (options == VPSC_ANY)
2539 			return (reclaimed != 0);
2540 	}
2541 }
2542 
2543 /*
2544  *	vm_wait:	(also see VM_WAIT macro)
2545  *
2546  *	Sleep until free pages are available for allocation.
2547  *	- Called in various places before memory allocations.
2548  */
2549 static void
2550 _vm_wait(void)
2551 {
2552 
2553 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
2554 	if (curproc == pageproc) {
2555 		vm_pageout_pages_needed = 1;
2556 		msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx,
2557 		    PDROP | PSWP, "VMWait", 0);
2558 	} else {
2559 		if (__predict_false(pageproc == NULL))
2560 			panic("vm_wait in early boot");
2561 		if (!vm_pageout_wanted) {
2562 			vm_pageout_wanted = true;
2563 			wakeup(&vm_pageout_wanted);
2564 		}
2565 		vm_pages_needed = true;
2566 		msleep(&vm_cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PVM,
2567 		    "vmwait", 0);
2568 	}
2569 }
2570 
2571 void
2572 vm_wait(void)
2573 {
2574 
2575 	mtx_lock(&vm_page_queue_free_mtx);
2576 	_vm_wait();
2577 }
2578 
2579 /*
2580  *	vm_page_alloc_fail:
2581  *
2582  *	Called when a page allocation function fails.  Informs the
2583  *	pagedaemon and performs the requested wait.  Requires the
2584  *	page_queue_free and object lock on entry.  Returns with the
2585  *	object lock held and free lock released.  Returns an error when
2586  *	retry is necessary.
2587  *
2588  */
2589 static int
2590 vm_page_alloc_fail(vm_object_t object, int req)
2591 {
2592 
2593 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
2594 
2595 	atomic_add_int(&vm_pageout_deficit,
2596 	    max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
2597 	pagedaemon_wakeup();
2598 	if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) {
2599 		if (object != NULL)
2600 			VM_OBJECT_WUNLOCK(object);
2601 		_vm_wait();
2602 		if (object != NULL)
2603 			VM_OBJECT_WLOCK(object);
2604 		if (req & VM_ALLOC_WAITOK)
2605 			return (EAGAIN);
2606 	} else
2607 		mtx_unlock(&vm_page_queue_free_mtx);
2608 	return (0);
2609 }
2610 
2611 /*
2612  *	vm_waitpfault:	(also see VM_WAITPFAULT macro)
2613  *
2614  *	Sleep until free pages are available for allocation.
2615  *	- Called only in vm_fault so that processes page faulting
2616  *	  can be easily tracked.
2617  *	- Sleeps at a lower priority than vm_wait() so that vm_wait()ing
2618  *	  processes will be able to grab memory first.  Do not change
2619  *	  this balance without careful testing first.
2620  */
2621 void
2622 vm_waitpfault(void)
2623 {
2624 
2625 	mtx_lock(&vm_page_queue_free_mtx);
2626 	if (!vm_pageout_wanted) {
2627 		vm_pageout_wanted = true;
2628 		wakeup(&vm_pageout_wanted);
2629 	}
2630 	vm_pages_needed = true;
2631 	msleep(&vm_cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PUSER,
2632 	    "pfault", 0);
2633 }
2634 
2635 struct vm_pagequeue *
2636 vm_page_pagequeue(vm_page_t m)
2637 {
2638 
2639 	if (vm_page_in_laundry(m))
2640 		return (&vm_dom[0].vmd_pagequeues[m->queue]);
2641 	else
2642 		return (&vm_phys_domain(m)->vmd_pagequeues[m->queue]);
2643 }
2644 
2645 /*
2646  *	vm_page_dequeue:
2647  *
2648  *	Remove the given page from its current page queue.
2649  *
2650  *	The page must be locked.
2651  */
2652 void
2653 vm_page_dequeue(vm_page_t m)
2654 {
2655 	struct vm_pagequeue *pq;
2656 
2657 	vm_page_assert_locked(m);
2658 	KASSERT(m->queue < PQ_COUNT, ("vm_page_dequeue: page %p is not queued",
2659 	    m));
2660 	pq = vm_page_pagequeue(m);
2661 	vm_pagequeue_lock(pq);
2662 	m->queue = PQ_NONE;
2663 	TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2664 	vm_pagequeue_cnt_dec(pq);
2665 	vm_pagequeue_unlock(pq);
2666 }
2667 
2668 /*
2669  *	vm_page_dequeue_locked:
2670  *
2671  *	Remove the given page from its current page queue.
2672  *
2673  *	The page and page queue must be locked.
2674  */
2675 void
2676 vm_page_dequeue_locked(vm_page_t m)
2677 {
2678 	struct vm_pagequeue *pq;
2679 
2680 	vm_page_lock_assert(m, MA_OWNED);
2681 	pq = vm_page_pagequeue(m);
2682 	vm_pagequeue_assert_locked(pq);
2683 	m->queue = PQ_NONE;
2684 	TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2685 	vm_pagequeue_cnt_dec(pq);
2686 }
2687 
2688 /*
2689  *	vm_page_enqueue:
2690  *
2691  *	Add the given page to the specified page queue.
2692  *
2693  *	The page must be locked.
2694  */
2695 static void
2696 vm_page_enqueue(uint8_t queue, vm_page_t m)
2697 {
2698 	struct vm_pagequeue *pq;
2699 
2700 	vm_page_lock_assert(m, MA_OWNED);
2701 	KASSERT(queue < PQ_COUNT,
2702 	    ("vm_page_enqueue: invalid queue %u request for page %p",
2703 	    queue, m));
2704 	if (queue == PQ_LAUNDRY || queue == PQ_UNSWAPPABLE)
2705 		pq = &vm_dom[0].vmd_pagequeues[queue];
2706 	else
2707 		pq = &vm_phys_domain(m)->vmd_pagequeues[queue];
2708 	vm_pagequeue_lock(pq);
2709 	m->queue = queue;
2710 	TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2711 	vm_pagequeue_cnt_inc(pq);
2712 	vm_pagequeue_unlock(pq);
2713 }
2714 
2715 /*
2716  *	vm_page_requeue:
2717  *
2718  *	Move the given page to the tail of its current page queue.
2719  *
2720  *	The page must be locked.
2721  */
2722 void
2723 vm_page_requeue(vm_page_t m)
2724 {
2725 	struct vm_pagequeue *pq;
2726 
2727 	vm_page_lock_assert(m, MA_OWNED);
2728 	KASSERT(m->queue != PQ_NONE,
2729 	    ("vm_page_requeue: page %p is not queued", m));
2730 	pq = vm_page_pagequeue(m);
2731 	vm_pagequeue_lock(pq);
2732 	TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2733 	TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2734 	vm_pagequeue_unlock(pq);
2735 }
2736 
2737 /*
2738  *	vm_page_requeue_locked:
2739  *
2740  *	Move the given page to the tail of its current page queue.
2741  *
2742  *	The page queue must be locked.
2743  */
2744 void
2745 vm_page_requeue_locked(vm_page_t m)
2746 {
2747 	struct vm_pagequeue *pq;
2748 
2749 	KASSERT(m->queue != PQ_NONE,
2750 	    ("vm_page_requeue_locked: page %p is not queued", m));
2751 	pq = vm_page_pagequeue(m);
2752 	vm_pagequeue_assert_locked(pq);
2753 	TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2754 	TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2755 }
2756 
2757 /*
2758  *	vm_page_activate:
2759  *
2760  *	Put the specified page on the active list (if appropriate).
2761  *	Ensure that act_count is at least ACT_INIT but do not otherwise
2762  *	mess with it.
2763  *
2764  *	The page must be locked.
2765  */
2766 void
2767 vm_page_activate(vm_page_t m)
2768 {
2769 	int queue;
2770 
2771 	vm_page_lock_assert(m, MA_OWNED);
2772 	if ((queue = m->queue) != PQ_ACTIVE) {
2773 		if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
2774 			if (m->act_count < ACT_INIT)
2775 				m->act_count = ACT_INIT;
2776 			if (queue != PQ_NONE)
2777 				vm_page_dequeue(m);
2778 			vm_page_enqueue(PQ_ACTIVE, m);
2779 		} else
2780 			KASSERT(queue == PQ_NONE,
2781 			    ("vm_page_activate: wired page %p is queued", m));
2782 	} else {
2783 		if (m->act_count < ACT_INIT)
2784 			m->act_count = ACT_INIT;
2785 	}
2786 }
2787 
2788 /*
2789  *	vm_page_free_wakeup:
2790  *
2791  *	Helper routine for vm_page_free_toq().  This routine is called
2792  *	when a page is added to the free queues.
2793  *
2794  *	The page queues must be locked.
2795  */
2796 static void
2797 vm_page_free_wakeup(void)
2798 {
2799 
2800 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
2801 	/*
2802 	 * if pageout daemon needs pages, then tell it that there are
2803 	 * some free.
2804 	 */
2805 	if (vm_pageout_pages_needed &&
2806 	    vm_cnt.v_free_count >= vm_cnt.v_pageout_free_min) {
2807 		wakeup(&vm_pageout_pages_needed);
2808 		vm_pageout_pages_needed = 0;
2809 	}
2810 	/*
2811 	 * wakeup processes that are waiting on memory if we hit a
2812 	 * high water mark. And wakeup scheduler process if we have
2813 	 * lots of memory. this process will swapin processes.
2814 	 */
2815 	if (vm_pages_needed && !vm_page_count_min()) {
2816 		vm_pages_needed = false;
2817 		wakeup(&vm_cnt.v_free_count);
2818 	}
2819 }
2820 
2821 /*
2822  *	vm_page_free_prep:
2823  *
2824  *	Prepares the given page to be put on the free list,
2825  *	disassociating it from any VM object. The caller may return
2826  *	the page to the free list only if this function returns true.
2827  *
2828  *	The object must be locked.  The page must be locked if it is
2829  *	managed.  For a queued managed page, the pagequeue_locked
2830  *	argument specifies whether the page queue is already locked.
2831  */
2832 bool
2833 vm_page_free_prep(vm_page_t m, bool pagequeue_locked)
2834 {
2835 
2836 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP)
2837 	if ((m->flags & PG_ZERO) != 0) {
2838 		uint64_t *p;
2839 		int i;
2840 		p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
2841 		for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++)
2842 			KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx",
2843 			    m, i, (uintmax_t)*p));
2844 	}
2845 #endif
2846 	if ((m->oflags & VPO_UNMANAGED) == 0) {
2847 		vm_page_lock_assert(m, MA_OWNED);
2848 		KASSERT(!pmap_page_is_mapped(m),
2849 		    ("vm_page_free_toq: freeing mapped page %p", m));
2850 	} else
2851 		KASSERT(m->queue == PQ_NONE,
2852 		    ("vm_page_free_toq: unmanaged page %p is queued", m));
2853 	VM_CNT_INC(v_tfree);
2854 
2855 	if (vm_page_sbusied(m))
2856 		panic("vm_page_free: freeing busy page %p", m);
2857 
2858 	vm_page_remove(m);
2859 
2860 	/*
2861 	 * If fictitious remove object association and
2862 	 * return.
2863 	 */
2864 	if ((m->flags & PG_FICTITIOUS) != 0) {
2865 		KASSERT(m->wire_count == 1,
2866 		    ("fictitious page %p is not wired", m));
2867 		KASSERT(m->queue == PQ_NONE,
2868 		    ("fictitious page %p is queued", m));
2869 		return (false);
2870 	}
2871 
2872 	if (m->queue != PQ_NONE) {
2873 		if (pagequeue_locked)
2874 			vm_page_dequeue_locked(m);
2875 		else
2876 			vm_page_dequeue(m);
2877 	}
2878 	m->valid = 0;
2879 	vm_page_undirty(m);
2880 
2881 	if (m->wire_count != 0)
2882 		panic("vm_page_free: freeing wired page %p", m);
2883 	if (m->hold_count != 0) {
2884 		m->flags &= ~PG_ZERO;
2885 		KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2886 		    ("vm_page_free: freeing PG_UNHOLDFREE page %p", m));
2887 		m->flags |= PG_UNHOLDFREE;
2888 		return (false);
2889 	}
2890 
2891 	/*
2892 	 * Restore the default memory attribute to the page.
2893 	 */
2894 	if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
2895 		pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
2896 
2897 	return (true);
2898 }
2899 
2900 /*
2901  * Insert the page into the physical memory allocator's free page
2902  * queues.  This is the last step to free a page.
2903  */
2904 static void
2905 vm_page_free_phys(vm_page_t m)
2906 {
2907 
2908 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
2909 
2910 	vm_phys_freecnt_adj(m, 1);
2911 #if VM_NRESERVLEVEL > 0
2912 	if (!vm_reserv_free_page(m))
2913 #endif
2914 		vm_phys_free_pages(m, 0);
2915 }
2916 
2917 void
2918 vm_page_free_phys_pglist(struct pglist *tq)
2919 {
2920 	vm_page_t m;
2921 
2922 	if (TAILQ_EMPTY(tq))
2923 		return;
2924 	mtx_lock(&vm_page_queue_free_mtx);
2925 	TAILQ_FOREACH(m, tq, listq)
2926 		vm_page_free_phys(m);
2927 	vm_page_free_wakeup();
2928 	mtx_unlock(&vm_page_queue_free_mtx);
2929 }
2930 
2931 /*
2932  *	vm_page_free_toq:
2933  *
2934  *	Returns the given page to the free list, disassociating it
2935  *	from any VM object.
2936  *
2937  *	The object must be locked.  The page must be locked if it is
2938  *	managed.
2939  */
2940 void
2941 vm_page_free_toq(vm_page_t m)
2942 {
2943 
2944 	if (!vm_page_free_prep(m, false))
2945 		return;
2946 	mtx_lock(&vm_page_queue_free_mtx);
2947 	vm_page_free_phys(m);
2948 	vm_page_free_wakeup();
2949 	mtx_unlock(&vm_page_queue_free_mtx);
2950 }
2951 
2952 /*
2953  *	vm_page_wire:
2954  *
2955  *	Mark this page as wired down by yet
2956  *	another map, removing it from paging queues
2957  *	as necessary.
2958  *
2959  *	If the page is fictitious, then its wire count must remain one.
2960  *
2961  *	The page must be locked.
2962  */
2963 void
2964 vm_page_wire(vm_page_t m)
2965 {
2966 
2967 	/*
2968 	 * Only bump the wire statistics if the page is not already wired,
2969 	 * and only unqueue the page if it is on some queue (if it is unmanaged
2970 	 * it is already off the queues).
2971 	 */
2972 	vm_page_lock_assert(m, MA_OWNED);
2973 	if ((m->flags & PG_FICTITIOUS) != 0) {
2974 		KASSERT(m->wire_count == 1,
2975 		    ("vm_page_wire: fictitious page %p's wire count isn't one",
2976 		    m));
2977 		return;
2978 	}
2979 	if (m->wire_count == 0) {
2980 		KASSERT((m->oflags & VPO_UNMANAGED) == 0 ||
2981 		    m->queue == PQ_NONE,
2982 		    ("vm_page_wire: unmanaged page %p is queued", m));
2983 		vm_page_remque(m);
2984 		atomic_add_int(&vm_cnt.v_wire_count, 1);
2985 	}
2986 	m->wire_count++;
2987 	KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m));
2988 }
2989 
2990 /*
2991  * vm_page_unwire:
2992  *
2993  * Release one wiring of the specified page, potentially allowing it to be
2994  * paged out.  Returns TRUE if the number of wirings transitions to zero and
2995  * FALSE otherwise.
2996  *
2997  * Only managed pages belonging to an object can be paged out.  If the number
2998  * of wirings transitions to zero and the page is eligible for page out, then
2999  * the page is added to the specified paging queue (unless PQ_NONE is
3000  * specified).
3001  *
3002  * If a page is fictitious, then its wire count must always be one.
3003  *
3004  * A managed page must be locked.
3005  */
3006 boolean_t
3007 vm_page_unwire(vm_page_t m, uint8_t queue)
3008 {
3009 
3010 	KASSERT(queue < PQ_COUNT || queue == PQ_NONE,
3011 	    ("vm_page_unwire: invalid queue %u request for page %p",
3012 	    queue, m));
3013 	if ((m->oflags & VPO_UNMANAGED) == 0)
3014 		vm_page_assert_locked(m);
3015 	if ((m->flags & PG_FICTITIOUS) != 0) {
3016 		KASSERT(m->wire_count == 1,
3017 	    ("vm_page_unwire: fictitious page %p's wire count isn't one", m));
3018 		return (FALSE);
3019 	}
3020 	if (m->wire_count > 0) {
3021 		m->wire_count--;
3022 		if (m->wire_count == 0) {
3023 			atomic_subtract_int(&vm_cnt.v_wire_count, 1);
3024 			if ((m->oflags & VPO_UNMANAGED) == 0 &&
3025 			    m->object != NULL && queue != PQ_NONE)
3026 				vm_page_enqueue(queue, m);
3027 			return (TRUE);
3028 		} else
3029 			return (FALSE);
3030 	} else
3031 		panic("vm_page_unwire: page %p's wire count is zero", m);
3032 }
3033 
3034 /*
3035  * Move the specified page to the inactive queue.
3036  *
3037  * Normally, "noreuse" is FALSE, resulting in LRU ordering of the inactive
3038  * queue.  However, setting "noreuse" to TRUE will accelerate the specified
3039  * page's reclamation, but it will not unmap the page from any address space.
3040  * This is implemented by inserting the page near the head of the inactive
3041  * queue, using a marker page to guide FIFO insertion ordering.
3042  *
3043  * The page must be locked.
3044  */
3045 static inline void
3046 _vm_page_deactivate(vm_page_t m, boolean_t noreuse)
3047 {
3048 	struct vm_pagequeue *pq;
3049 	int queue;
3050 
3051 	vm_page_assert_locked(m);
3052 
3053 	/*
3054 	 * Ignore if the page is already inactive, unless it is unlikely to be
3055 	 * reactivated.
3056 	 */
3057 	if ((queue = m->queue) == PQ_INACTIVE && !noreuse)
3058 		return;
3059 	if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
3060 		pq = &vm_phys_domain(m)->vmd_pagequeues[PQ_INACTIVE];
3061 		/* Avoid multiple acquisitions of the inactive queue lock. */
3062 		if (queue == PQ_INACTIVE) {
3063 			vm_pagequeue_lock(pq);
3064 			vm_page_dequeue_locked(m);
3065 		} else {
3066 			if (queue != PQ_NONE)
3067 				vm_page_dequeue(m);
3068 			vm_pagequeue_lock(pq);
3069 		}
3070 		m->queue = PQ_INACTIVE;
3071 		if (noreuse)
3072 			TAILQ_INSERT_BEFORE(&vm_phys_domain(m)->vmd_inacthead,
3073 			    m, plinks.q);
3074 		else
3075 			TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3076 		vm_pagequeue_cnt_inc(pq);
3077 		vm_pagequeue_unlock(pq);
3078 	}
3079 }
3080 
3081 /*
3082  * Move the specified page to the inactive queue.
3083  *
3084  * The page must be locked.
3085  */
3086 void
3087 vm_page_deactivate(vm_page_t m)
3088 {
3089 
3090 	_vm_page_deactivate(m, FALSE);
3091 }
3092 
3093 /*
3094  * Move the specified page to the inactive queue with the expectation
3095  * that it is unlikely to be reused.
3096  *
3097  * The page must be locked.
3098  */
3099 void
3100 vm_page_deactivate_noreuse(vm_page_t m)
3101 {
3102 
3103 	_vm_page_deactivate(m, TRUE);
3104 }
3105 
3106 /*
3107  * vm_page_launder
3108  *
3109  * 	Put a page in the laundry.
3110  */
3111 void
3112 vm_page_launder(vm_page_t m)
3113 {
3114 	int queue;
3115 
3116 	vm_page_assert_locked(m);
3117 	if ((queue = m->queue) != PQ_LAUNDRY) {
3118 		if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
3119 			if (queue != PQ_NONE)
3120 				vm_page_dequeue(m);
3121 			vm_page_enqueue(PQ_LAUNDRY, m);
3122 		} else
3123 			KASSERT(queue == PQ_NONE,
3124 			    ("wired page %p is queued", m));
3125 	}
3126 }
3127 
3128 /*
3129  * vm_page_unswappable
3130  *
3131  *	Put a page in the PQ_UNSWAPPABLE holding queue.
3132  */
3133 void
3134 vm_page_unswappable(vm_page_t m)
3135 {
3136 
3137 	vm_page_assert_locked(m);
3138 	KASSERT(m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0,
3139 	    ("page %p already unswappable", m));
3140 	if (m->queue != PQ_NONE)
3141 		vm_page_dequeue(m);
3142 	vm_page_enqueue(PQ_UNSWAPPABLE, m);
3143 }
3144 
3145 /*
3146  * Attempt to free the page.  If it cannot be freed, do nothing.  Returns true
3147  * if the page is freed and false otherwise.
3148  *
3149  * The page must be managed.  The page and its containing object must be
3150  * locked.
3151  */
3152 bool
3153 vm_page_try_to_free(vm_page_t m)
3154 {
3155 
3156 	vm_page_assert_locked(m);
3157 	VM_OBJECT_ASSERT_WLOCKED(m->object);
3158 	KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("page %p is unmanaged", m));
3159 	if (m->dirty != 0 || m->hold_count != 0 || m->wire_count != 0 ||
3160 	    vm_page_busied(m))
3161 		return (false);
3162 	if (m->object->ref_count != 0) {
3163 		pmap_remove_all(m);
3164 		if (m->dirty != 0)
3165 			return (false);
3166 	}
3167 	vm_page_free(m);
3168 	return (true);
3169 }
3170 
3171 /*
3172  * vm_page_advise
3173  *
3174  * 	Apply the specified advice to the given page.
3175  *
3176  *	The object and page must be locked.
3177  */
3178 void
3179 vm_page_advise(vm_page_t m, int advice)
3180 {
3181 
3182 	vm_page_assert_locked(m);
3183 	VM_OBJECT_ASSERT_WLOCKED(m->object);
3184 	if (advice == MADV_FREE)
3185 		/*
3186 		 * Mark the page clean.  This will allow the page to be freed
3187 		 * without first paging it out.  MADV_FREE pages are often
3188 		 * quickly reused by malloc(3), so we do not do anything that
3189 		 * would result in a page fault on a later access.
3190 		 */
3191 		vm_page_undirty(m);
3192 	else if (advice != MADV_DONTNEED) {
3193 		if (advice == MADV_WILLNEED)
3194 			vm_page_activate(m);
3195 		return;
3196 	}
3197 
3198 	/*
3199 	 * Clear any references to the page.  Otherwise, the page daemon will
3200 	 * immediately reactivate the page.
3201 	 */
3202 	vm_page_aflag_clear(m, PGA_REFERENCED);
3203 
3204 	if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
3205 		vm_page_dirty(m);
3206 
3207 	/*
3208 	 * Place clean pages near the head of the inactive queue rather than
3209 	 * the tail, thus defeating the queue's LRU operation and ensuring that
3210 	 * the page will be reused quickly.  Dirty pages not already in the
3211 	 * laundry are moved there.
3212 	 */
3213 	if (m->dirty == 0)
3214 		vm_page_deactivate_noreuse(m);
3215 	else
3216 		vm_page_launder(m);
3217 }
3218 
3219 /*
3220  * Grab a page, waiting until we are waken up due to the page
3221  * changing state.  We keep on waiting, if the page continues
3222  * to be in the object.  If the page doesn't exist, first allocate it
3223  * and then conditionally zero it.
3224  *
3225  * This routine may sleep.
3226  *
3227  * The object must be locked on entry.  The lock will, however, be released
3228  * and reacquired if the routine sleeps.
3229  */
3230 vm_page_t
3231 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
3232 {
3233 	vm_page_t m;
3234 	int sleep;
3235 	int pflags;
3236 
3237 	VM_OBJECT_ASSERT_WLOCKED(object);
3238 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
3239 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
3240 	    ("vm_page_grab: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
3241 	pflags = allocflags &
3242 	    ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL);
3243 	if ((allocflags & VM_ALLOC_NOWAIT) == 0)
3244 		pflags |= VM_ALLOC_WAITFAIL;
3245 retrylookup:
3246 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
3247 		sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ?
3248 		    vm_page_xbusied(m) : vm_page_busied(m);
3249 		if (sleep) {
3250 			if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3251 				return (NULL);
3252 			/*
3253 			 * Reference the page before unlocking and
3254 			 * sleeping so that the page daemon is less
3255 			 * likely to reclaim it.
3256 			 */
3257 			vm_page_aflag_set(m, PGA_REFERENCED);
3258 			vm_page_lock(m);
3259 			VM_OBJECT_WUNLOCK(object);
3260 			vm_page_busy_sleep(m, "pgrbwt", (allocflags &
3261 			    VM_ALLOC_IGN_SBUSY) != 0);
3262 			VM_OBJECT_WLOCK(object);
3263 			goto retrylookup;
3264 		} else {
3265 			if ((allocflags & VM_ALLOC_WIRED) != 0) {
3266 				vm_page_lock(m);
3267 				vm_page_wire(m);
3268 				vm_page_unlock(m);
3269 			}
3270 			if ((allocflags &
3271 			    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
3272 				vm_page_xbusy(m);
3273 			if ((allocflags & VM_ALLOC_SBUSY) != 0)
3274 				vm_page_sbusy(m);
3275 			return (m);
3276 		}
3277 	}
3278 	m = vm_page_alloc(object, pindex, pflags);
3279 	if (m == NULL) {
3280 		if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3281 			return (NULL);
3282 		goto retrylookup;
3283 	}
3284 	if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
3285 		pmap_zero_page(m);
3286 	return (m);
3287 }
3288 
3289 /*
3290  * Return the specified range of pages from the given object.  For each
3291  * page offset within the range, if a page already exists within the object
3292  * at that offset and it is busy, then wait for it to change state.  If,
3293  * instead, the page doesn't exist, then allocate it.
3294  *
3295  * The caller must always specify an allocation class.
3296  *
3297  * allocation classes:
3298  *	VM_ALLOC_NORMAL		normal process request
3299  *	VM_ALLOC_SYSTEM		system *really* needs the pages
3300  *
3301  * The caller must always specify that the pages are to be busied and/or
3302  * wired.
3303  *
3304  * optional allocation flags:
3305  *	VM_ALLOC_IGN_SBUSY	do not sleep on soft busy pages
3306  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
3307  *	VM_ALLOC_NOWAIT		do not sleep
3308  *	VM_ALLOC_SBUSY		set page to sbusy state
3309  *	VM_ALLOC_WIRED		wire the pages
3310  *	VM_ALLOC_ZERO		zero and validate any invalid pages
3311  *
3312  * If VM_ALLOC_NOWAIT is not specified, this routine may sleep.  Otherwise, it
3313  * may return a partial prefix of the requested range.
3314  */
3315 int
3316 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags,
3317     vm_page_t *ma, int count)
3318 {
3319 	vm_page_t m, mpred;
3320 	int pflags;
3321 	int i;
3322 	bool sleep;
3323 
3324 	VM_OBJECT_ASSERT_WLOCKED(object);
3325 	KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0,
3326 	    ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed"));
3327 	KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 ||
3328 	    (allocflags & VM_ALLOC_WIRED) != 0,
3329 	    ("vm_page_grab_pages: the pages must be busied or wired"));
3330 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
3331 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
3332 	    ("vm_page_grab_pages: VM_ALLOC_SBUSY/IGN_SBUSY mismatch"));
3333 	if (count == 0)
3334 		return (0);
3335 	pflags = allocflags & ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK |
3336 	    VM_ALLOC_WAITFAIL | VM_ALLOC_IGN_SBUSY);
3337 	if ((allocflags & VM_ALLOC_NOWAIT) == 0)
3338 		pflags |= VM_ALLOC_WAITFAIL;
3339 	i = 0;
3340 retrylookup:
3341 	m = vm_radix_lookup_le(&object->rtree, pindex + i);
3342 	if (m == NULL || m->pindex != pindex + i) {
3343 		mpred = m;
3344 		m = NULL;
3345 	} else
3346 		mpred = TAILQ_PREV(m, pglist, listq);
3347 	for (; i < count; i++) {
3348 		if (m != NULL) {
3349 			sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ?
3350 			    vm_page_xbusied(m) : vm_page_busied(m);
3351 			if (sleep) {
3352 				if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3353 					break;
3354 				/*
3355 				 * Reference the page before unlocking and
3356 				 * sleeping so that the page daemon is less
3357 				 * likely to reclaim it.
3358 				 */
3359 				vm_page_aflag_set(m, PGA_REFERENCED);
3360 				vm_page_lock(m);
3361 				VM_OBJECT_WUNLOCK(object);
3362 				vm_page_busy_sleep(m, "grbmaw", (allocflags &
3363 				    VM_ALLOC_IGN_SBUSY) != 0);
3364 				VM_OBJECT_WLOCK(object);
3365 				goto retrylookup;
3366 			}
3367 			if ((allocflags & VM_ALLOC_WIRED) != 0) {
3368 				vm_page_lock(m);
3369 				vm_page_wire(m);
3370 				vm_page_unlock(m);
3371 			}
3372 			if ((allocflags & (VM_ALLOC_NOBUSY |
3373 			    VM_ALLOC_SBUSY)) == 0)
3374 				vm_page_xbusy(m);
3375 			if ((allocflags & VM_ALLOC_SBUSY) != 0)
3376 				vm_page_sbusy(m);
3377 		} else {
3378 			m = vm_page_alloc_after(object, pindex + i,
3379 			    pflags | VM_ALLOC_COUNT(count - i), mpred);
3380 			if (m == NULL) {
3381 				if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3382 					break;
3383 				goto retrylookup;
3384 			}
3385 		}
3386 		if (m->valid == 0 && (allocflags & VM_ALLOC_ZERO) != 0) {
3387 			if ((m->flags & PG_ZERO) == 0)
3388 				pmap_zero_page(m);
3389 			m->valid = VM_PAGE_BITS_ALL;
3390 		}
3391 		ma[i] = mpred = m;
3392 		m = vm_page_next(m);
3393 	}
3394 	return (i);
3395 }
3396 
3397 /*
3398  * Mapping function for valid or dirty bits in a page.
3399  *
3400  * Inputs are required to range within a page.
3401  */
3402 vm_page_bits_t
3403 vm_page_bits(int base, int size)
3404 {
3405 	int first_bit;
3406 	int last_bit;
3407 
3408 	KASSERT(
3409 	    base + size <= PAGE_SIZE,
3410 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
3411 	);
3412 
3413 	if (size == 0)		/* handle degenerate case */
3414 		return (0);
3415 
3416 	first_bit = base >> DEV_BSHIFT;
3417 	last_bit = (base + size - 1) >> DEV_BSHIFT;
3418 
3419 	return (((vm_page_bits_t)2 << last_bit) -
3420 	    ((vm_page_bits_t)1 << first_bit));
3421 }
3422 
3423 /*
3424  *	vm_page_set_valid_range:
3425  *
3426  *	Sets portions of a page valid.  The arguments are expected
3427  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
3428  *	of any partial chunks touched by the range.  The invalid portion of
3429  *	such chunks will be zeroed.
3430  *
3431  *	(base + size) must be less then or equal to PAGE_SIZE.
3432  */
3433 void
3434 vm_page_set_valid_range(vm_page_t m, int base, int size)
3435 {
3436 	int endoff, frag;
3437 
3438 	VM_OBJECT_ASSERT_WLOCKED(m->object);
3439 	if (size == 0)	/* handle degenerate case */
3440 		return;
3441 
3442 	/*
3443 	 * If the base is not DEV_BSIZE aligned and the valid
3444 	 * bit is clear, we have to zero out a portion of the
3445 	 * first block.
3446 	 */
3447 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
3448 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
3449 		pmap_zero_page_area(m, frag, base - frag);
3450 
3451 	/*
3452 	 * If the ending offset is not DEV_BSIZE aligned and the
3453 	 * valid bit is clear, we have to zero out a portion of
3454 	 * the last block.
3455 	 */
3456 	endoff = base + size;
3457 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
3458 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
3459 		pmap_zero_page_area(m, endoff,
3460 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
3461 
3462 	/*
3463 	 * Assert that no previously invalid block that is now being validated
3464 	 * is already dirty.
3465 	 */
3466 	KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
3467 	    ("vm_page_set_valid_range: page %p is dirty", m));
3468 
3469 	/*
3470 	 * Set valid bits inclusive of any overlap.
3471 	 */
3472 	m->valid |= vm_page_bits(base, size);
3473 }
3474 
3475 /*
3476  * Clear the given bits from the specified page's dirty field.
3477  */
3478 static __inline void
3479 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
3480 {
3481 	uintptr_t addr;
3482 #if PAGE_SIZE < 16384
3483 	int shift;
3484 #endif
3485 
3486 	/*
3487 	 * If the object is locked and the page is neither exclusive busy nor
3488 	 * write mapped, then the page's dirty field cannot possibly be
3489 	 * set by a concurrent pmap operation.
3490 	 */
3491 	VM_OBJECT_ASSERT_WLOCKED(m->object);
3492 	if (!vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
3493 		m->dirty &= ~pagebits;
3494 	else {
3495 		/*
3496 		 * The pmap layer can call vm_page_dirty() without
3497 		 * holding a distinguished lock.  The combination of
3498 		 * the object's lock and an atomic operation suffice
3499 		 * to guarantee consistency of the page dirty field.
3500 		 *
3501 		 * For PAGE_SIZE == 32768 case, compiler already
3502 		 * properly aligns the dirty field, so no forcible
3503 		 * alignment is needed. Only require existence of
3504 		 * atomic_clear_64 when page size is 32768.
3505 		 */
3506 		addr = (uintptr_t)&m->dirty;
3507 #if PAGE_SIZE == 32768
3508 		atomic_clear_64((uint64_t *)addr, pagebits);
3509 #elif PAGE_SIZE == 16384
3510 		atomic_clear_32((uint32_t *)addr, pagebits);
3511 #else		/* PAGE_SIZE <= 8192 */
3512 		/*
3513 		 * Use a trick to perform a 32-bit atomic on the
3514 		 * containing aligned word, to not depend on the existence
3515 		 * of atomic_clear_{8, 16}.
3516 		 */
3517 		shift = addr & (sizeof(uint32_t) - 1);
3518 #if BYTE_ORDER == BIG_ENDIAN
3519 		shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY;
3520 #else
3521 		shift *= NBBY;
3522 #endif
3523 		addr &= ~(sizeof(uint32_t) - 1);
3524 		atomic_clear_32((uint32_t *)addr, pagebits << shift);
3525 #endif		/* PAGE_SIZE */
3526 	}
3527 }
3528 
3529 /*
3530  *	vm_page_set_validclean:
3531  *
3532  *	Sets portions of a page valid and clean.  The arguments are expected
3533  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
3534  *	of any partial chunks touched by the range.  The invalid portion of
3535  *	such chunks will be zero'd.
3536  *
3537  *	(base + size) must be less then or equal to PAGE_SIZE.
3538  */
3539 void
3540 vm_page_set_validclean(vm_page_t m, int base, int size)
3541 {
3542 	vm_page_bits_t oldvalid, pagebits;
3543 	int endoff, frag;
3544 
3545 	VM_OBJECT_ASSERT_WLOCKED(m->object);
3546 	if (size == 0)	/* handle degenerate case */
3547 		return;
3548 
3549 	/*
3550 	 * If the base is not DEV_BSIZE aligned and the valid
3551 	 * bit is clear, we have to zero out a portion of the
3552 	 * first block.
3553 	 */
3554 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
3555 	    (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
3556 		pmap_zero_page_area(m, frag, base - frag);
3557 
3558 	/*
3559 	 * If the ending offset is not DEV_BSIZE aligned and the
3560 	 * valid bit is clear, we have to zero out a portion of
3561 	 * the last block.
3562 	 */
3563 	endoff = base + size;
3564 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
3565 	    (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
3566 		pmap_zero_page_area(m, endoff,
3567 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
3568 
3569 	/*
3570 	 * Set valid, clear dirty bits.  If validating the entire
3571 	 * page we can safely clear the pmap modify bit.  We also
3572 	 * use this opportunity to clear the VPO_NOSYNC flag.  If a process
3573 	 * takes a write fault on a MAP_NOSYNC memory area the flag will
3574 	 * be set again.
3575 	 *
3576 	 * We set valid bits inclusive of any overlap, but we can only
3577 	 * clear dirty bits for DEV_BSIZE chunks that are fully within
3578 	 * the range.
3579 	 */
3580 	oldvalid = m->valid;
3581 	pagebits = vm_page_bits(base, size);
3582 	m->valid |= pagebits;
3583 #if 0	/* NOT YET */
3584 	if ((frag = base & (DEV_BSIZE - 1)) != 0) {
3585 		frag = DEV_BSIZE - frag;
3586 		base += frag;
3587 		size -= frag;
3588 		if (size < 0)
3589 			size = 0;
3590 	}
3591 	pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
3592 #endif
3593 	if (base == 0 && size == PAGE_SIZE) {
3594 		/*
3595 		 * The page can only be modified within the pmap if it is
3596 		 * mapped, and it can only be mapped if it was previously
3597 		 * fully valid.
3598 		 */
3599 		if (oldvalid == VM_PAGE_BITS_ALL)
3600 			/*
3601 			 * Perform the pmap_clear_modify() first.  Otherwise,
3602 			 * a concurrent pmap operation, such as
3603 			 * pmap_protect(), could clear a modification in the
3604 			 * pmap and set the dirty field on the page before
3605 			 * pmap_clear_modify() had begun and after the dirty
3606 			 * field was cleared here.
3607 			 */
3608 			pmap_clear_modify(m);
3609 		m->dirty = 0;
3610 		m->oflags &= ~VPO_NOSYNC;
3611 	} else if (oldvalid != VM_PAGE_BITS_ALL)
3612 		m->dirty &= ~pagebits;
3613 	else
3614 		vm_page_clear_dirty_mask(m, pagebits);
3615 }
3616 
3617 void
3618 vm_page_clear_dirty(vm_page_t m, int base, int size)
3619 {
3620 
3621 	vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
3622 }
3623 
3624 /*
3625  *	vm_page_set_invalid:
3626  *
3627  *	Invalidates DEV_BSIZE'd chunks within a page.  Both the
3628  *	valid and dirty bits for the effected areas are cleared.
3629  */
3630 void
3631 vm_page_set_invalid(vm_page_t m, int base, int size)
3632 {
3633 	vm_page_bits_t bits;
3634 	vm_object_t object;
3635 
3636 	object = m->object;
3637 	VM_OBJECT_ASSERT_WLOCKED(object);
3638 	if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
3639 	    size >= object->un_pager.vnp.vnp_size)
3640 		bits = VM_PAGE_BITS_ALL;
3641 	else
3642 		bits = vm_page_bits(base, size);
3643 	if (object->ref_count != 0 && m->valid == VM_PAGE_BITS_ALL &&
3644 	    bits != 0)
3645 		pmap_remove_all(m);
3646 	KASSERT((bits == 0 && m->valid == VM_PAGE_BITS_ALL) ||
3647 	    !pmap_page_is_mapped(m),
3648 	    ("vm_page_set_invalid: page %p is mapped", m));
3649 	m->valid &= ~bits;
3650 	m->dirty &= ~bits;
3651 }
3652 
3653 /*
3654  * vm_page_zero_invalid()
3655  *
3656  *	The kernel assumes that the invalid portions of a page contain
3657  *	garbage, but such pages can be mapped into memory by user code.
3658  *	When this occurs, we must zero out the non-valid portions of the
3659  *	page so user code sees what it expects.
3660  *
3661  *	Pages are most often semi-valid when the end of a file is mapped
3662  *	into memory and the file's size is not page aligned.
3663  */
3664 void
3665 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
3666 {
3667 	int b;
3668 	int i;
3669 
3670 	VM_OBJECT_ASSERT_WLOCKED(m->object);
3671 	/*
3672 	 * Scan the valid bits looking for invalid sections that
3673 	 * must be zeroed.  Invalid sub-DEV_BSIZE'd areas ( where the
3674 	 * valid bit may be set ) have already been zeroed by
3675 	 * vm_page_set_validclean().
3676 	 */
3677 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
3678 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
3679 		    (m->valid & ((vm_page_bits_t)1 << i))) {
3680 			if (i > b) {
3681 				pmap_zero_page_area(m,
3682 				    b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
3683 			}
3684 			b = i + 1;
3685 		}
3686 	}
3687 
3688 	/*
3689 	 * setvalid is TRUE when we can safely set the zero'd areas
3690 	 * as being valid.  We can do this if there are no cache consistancy
3691 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
3692 	 */
3693 	if (setvalid)
3694 		m->valid = VM_PAGE_BITS_ALL;
3695 }
3696 
3697 /*
3698  *	vm_page_is_valid:
3699  *
3700  *	Is (partial) page valid?  Note that the case where size == 0
3701  *	will return FALSE in the degenerate case where the page is
3702  *	entirely invalid, and TRUE otherwise.
3703  */
3704 int
3705 vm_page_is_valid(vm_page_t m, int base, int size)
3706 {
3707 	vm_page_bits_t bits;
3708 
3709 	VM_OBJECT_ASSERT_LOCKED(m->object);
3710 	bits = vm_page_bits(base, size);
3711 	return (m->valid != 0 && (m->valid & bits) == bits);
3712 }
3713 
3714 /*
3715  * Returns true if all of the specified predicates are true for the entire
3716  * (super)page and false otherwise.
3717  */
3718 bool
3719 vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m)
3720 {
3721 	vm_object_t object;
3722 	int i, npages;
3723 
3724 	object = m->object;
3725 	VM_OBJECT_ASSERT_LOCKED(object);
3726 	npages = atop(pagesizes[m->psind]);
3727 
3728 	/*
3729 	 * The physically contiguous pages that make up a superpage, i.e., a
3730 	 * page with a page size index ("psind") greater than zero, will
3731 	 * occupy adjacent entries in vm_page_array[].
3732 	 */
3733 	for (i = 0; i < npages; i++) {
3734 		/* Always test object consistency, including "skip_m". */
3735 		if (m[i].object != object)
3736 			return (false);
3737 		if (&m[i] == skip_m)
3738 			continue;
3739 		if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i]))
3740 			return (false);
3741 		if ((flags & PS_ALL_DIRTY) != 0) {
3742 			/*
3743 			 * Calling vm_page_test_dirty() or pmap_is_modified()
3744 			 * might stop this case from spuriously returning
3745 			 * "false".  However, that would require a write lock
3746 			 * on the object containing "m[i]".
3747 			 */
3748 			if (m[i].dirty != VM_PAGE_BITS_ALL)
3749 				return (false);
3750 		}
3751 		if ((flags & PS_ALL_VALID) != 0 &&
3752 		    m[i].valid != VM_PAGE_BITS_ALL)
3753 			return (false);
3754 	}
3755 	return (true);
3756 }
3757 
3758 /*
3759  * Set the page's dirty bits if the page is modified.
3760  */
3761 void
3762 vm_page_test_dirty(vm_page_t m)
3763 {
3764 
3765 	VM_OBJECT_ASSERT_WLOCKED(m->object);
3766 	if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
3767 		vm_page_dirty(m);
3768 }
3769 
3770 void
3771 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
3772 {
3773 
3774 	mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
3775 }
3776 
3777 void
3778 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
3779 {
3780 
3781 	mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
3782 }
3783 
3784 int
3785 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
3786 {
3787 
3788 	return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
3789 }
3790 
3791 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
3792 void
3793 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
3794 {
3795 
3796 	vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
3797 }
3798 
3799 void
3800 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
3801 {
3802 
3803 	mtx_assert_(vm_page_lockptr(m), a, file, line);
3804 }
3805 #endif
3806 
3807 #ifdef INVARIANTS
3808 void
3809 vm_page_object_lock_assert(vm_page_t m)
3810 {
3811 
3812 	/*
3813 	 * Certain of the page's fields may only be modified by the
3814 	 * holder of the containing object's lock or the exclusive busy.
3815 	 * holder.  Unfortunately, the holder of the write busy is
3816 	 * not recorded, and thus cannot be checked here.
3817 	 */
3818 	if (m->object != NULL && !vm_page_xbusied(m))
3819 		VM_OBJECT_ASSERT_WLOCKED(m->object);
3820 }
3821 
3822 void
3823 vm_page_assert_pga_writeable(vm_page_t m, uint8_t bits)
3824 {
3825 
3826 	if ((bits & PGA_WRITEABLE) == 0)
3827 		return;
3828 
3829 	/*
3830 	 * The PGA_WRITEABLE flag can only be set if the page is
3831 	 * managed, is exclusively busied or the object is locked.
3832 	 * Currently, this flag is only set by pmap_enter().
3833 	 */
3834 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3835 	    ("PGA_WRITEABLE on unmanaged page"));
3836 	if (!vm_page_xbusied(m))
3837 		VM_OBJECT_ASSERT_LOCKED(m->object);
3838 }
3839 #endif
3840 
3841 #include "opt_ddb.h"
3842 #ifdef DDB
3843 #include <sys/kernel.h>
3844 
3845 #include <ddb/ddb.h>
3846 
3847 DB_SHOW_COMMAND(page, vm_page_print_page_info)
3848 {
3849 
3850 	db_printf("vm_cnt.v_free_count: %d\n", vm_cnt.v_free_count);
3851 	db_printf("vm_cnt.v_inactive_count: %d\n", vm_cnt.v_inactive_count);
3852 	db_printf("vm_cnt.v_active_count: %d\n", vm_cnt.v_active_count);
3853 	db_printf("vm_cnt.v_laundry_count: %d\n", vm_cnt.v_laundry_count);
3854 	db_printf("vm_cnt.v_wire_count: %d\n", vm_cnt.v_wire_count);
3855 	db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
3856 	db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
3857 	db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
3858 	db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
3859 }
3860 
3861 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
3862 {
3863 	int dom;
3864 
3865 	db_printf("pq_free %d\n", vm_cnt.v_free_count);
3866 	for (dom = 0; dom < vm_ndomains; dom++) {
3867 		db_printf(
3868     "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n",
3869 		    dom,
3870 		    vm_dom[dom].vmd_page_count,
3871 		    vm_dom[dom].vmd_free_count,
3872 		    vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
3873 		    vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
3874 		    vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt,
3875 		    vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt);
3876 	}
3877 }
3878 
3879 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
3880 {
3881 	vm_page_t m;
3882 	boolean_t phys;
3883 
3884 	if (!have_addr) {
3885 		db_printf("show pginfo addr\n");
3886 		return;
3887 	}
3888 
3889 	phys = strchr(modif, 'p') != NULL;
3890 	if (phys)
3891 		m = PHYS_TO_VM_PAGE(addr);
3892 	else
3893 		m = (vm_page_t)addr;
3894 	db_printf(
3895     "page %p obj %p pidx 0x%jx phys 0x%jx q %d hold %d wire %d\n"
3896     "  af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
3897 	    m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
3898 	    m->queue, m->hold_count, m->wire_count, m->aflags, m->oflags,
3899 	    m->flags, m->act_count, m->busy_lock, m->valid, m->dirty);
3900 }
3901 #endif /* DDB */
3902