xref: /freebsd/sys/vm/vm_page.c (revision 69718b786d3943ea9a99eeeb5f5f6162f11c78b7)
1 /*-
2  * Copyright (c) 1991 Regents of the University of California.
3  * All rights reserved.
4  * Copyright (c) 1998 Matthew Dillon.  All Rights Reserved.
5  *
6  * This code is derived from software contributed to Berkeley by
7  * The Mach Operating System project at Carnegie-Mellon University.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  * 4. Neither the name of the University nor the names of its contributors
18  *    may be used to endorse or promote products derived from this software
19  *    without specific prior written permission.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
22  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
25  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31  * SUCH DAMAGE.
32  *
33  *	from: @(#)vm_page.c	7.4 (Berkeley) 5/7/91
34  */
35 
36 /*-
37  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
38  * All rights reserved.
39  *
40  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
41  *
42  * Permission to use, copy, modify and distribute this software and
43  * its documentation is hereby granted, provided that both the copyright
44  * notice and this permission notice appear in all copies of the
45  * software, derivative works or modified versions, and any portions
46  * thereof, and that both notices appear in supporting documentation.
47  *
48  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
49  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
50  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
51  *
52  * Carnegie Mellon requests users of this software to return to
53  *
54  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
55  *  School of Computer Science
56  *  Carnegie Mellon University
57  *  Pittsburgh PA 15213-3890
58  *
59  * any improvements or extensions that they make and grant Carnegie the
60  * rights to redistribute these changes.
61  */
62 
63 /*
64  *			GENERAL RULES ON VM_PAGE MANIPULATION
65  *
66  *	- A page queue lock is required when adding or removing a page from a
67  *	  page queue regardless of other locks or the busy state of a page.
68  *
69  *		* In general, no thread besides the page daemon can acquire or
70  *		  hold more than one page queue lock at a time.
71  *
72  *		* The page daemon can acquire and hold any pair of page queue
73  *		  locks in any order.
74  *
75  *	- The object lock is required when inserting or removing
76  *	  pages from an object (vm_page_insert() or vm_page_remove()).
77  *
78  */
79 
80 /*
81  *	Resident memory management module.
82  */
83 
84 #include <sys/cdefs.h>
85 __FBSDID("$FreeBSD$");
86 
87 #include "opt_vm.h"
88 
89 #include <sys/param.h>
90 #include <sys/systm.h>
91 #include <sys/lock.h>
92 #include <sys/kernel.h>
93 #include <sys/limits.h>
94 #include <sys/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 vm_page_queue_free_mtx;
131 
132 struct mtx_padalign pa_lock[PA_LOCK_COUNT];
133 
134 vm_page_t vm_page_array;
135 long vm_page_array_size;
136 long first_page;
137 
138 static int boot_pages = UMA_BOOT_PAGES;
139 SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
140     &boot_pages, 0,
141     "number of pages allocated for bootstrapping the VM system");
142 
143 static int pa_tryrelock_restart;
144 SYSCTL_INT(_vm, OID_AUTO, tryrelock_restart, CTLFLAG_RD,
145     &pa_tryrelock_restart, 0, "Number of tryrelock restarts");
146 
147 static TAILQ_HEAD(, vm_page) blacklist_head;
148 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS);
149 SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD |
150     CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages");
151 
152 /* Is the page daemon waiting for free pages? */
153 static int vm_pageout_pages_needed;
154 
155 static uma_zone_t fakepg_zone;
156 
157 static void vm_page_alloc_check(vm_page_t m);
158 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits);
159 static void vm_page_enqueue(uint8_t queue, vm_page_t m);
160 static void vm_page_free_wakeup(void);
161 static void vm_page_init_fakepg(void *dummy);
162 static int vm_page_insert_after(vm_page_t m, vm_object_t object,
163     vm_pindex_t pindex, vm_page_t mpred);
164 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object,
165     vm_page_t mpred);
166 static int vm_page_reclaim_run(int req_class, u_long npages, vm_page_t m_run,
167     vm_paddr_t high);
168 
169 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init_fakepg, NULL);
170 
171 static void
172 vm_page_init_fakepg(void *dummy)
173 {
174 
175 	fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
176 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
177 }
178 
179 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
180 #if PAGE_SIZE == 32768
181 #ifdef CTASSERT
182 CTASSERT(sizeof(u_long) >= 8);
183 #endif
184 #endif
185 
186 /*
187  * Try to acquire a physical address lock while a pmap is locked.  If we
188  * fail to trylock we unlock and lock the pmap directly and cache the
189  * locked pa in *locked.  The caller should then restart their loop in case
190  * the virtual to physical mapping has changed.
191  */
192 int
193 vm_page_pa_tryrelock(pmap_t pmap, vm_paddr_t pa, vm_paddr_t *locked)
194 {
195 	vm_paddr_t lockpa;
196 
197 	lockpa = *locked;
198 	*locked = pa;
199 	if (lockpa) {
200 		PA_LOCK_ASSERT(lockpa, MA_OWNED);
201 		if (PA_LOCKPTR(pa) == PA_LOCKPTR(lockpa))
202 			return (0);
203 		PA_UNLOCK(lockpa);
204 	}
205 	if (PA_TRYLOCK(pa))
206 		return (0);
207 	PMAP_UNLOCK(pmap);
208 	atomic_add_int(&pa_tryrelock_restart, 1);
209 	PA_LOCK(pa);
210 	PMAP_LOCK(pmap);
211 	return (EAGAIN);
212 }
213 
214 /*
215  *	vm_set_page_size:
216  *
217  *	Sets the page size, perhaps based upon the memory
218  *	size.  Must be called before any use of page-size
219  *	dependent functions.
220  */
221 void
222 vm_set_page_size(void)
223 {
224 	if (vm_cnt.v_page_size == 0)
225 		vm_cnt.v_page_size = PAGE_SIZE;
226 	if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0)
227 		panic("vm_set_page_size: page size not a power of two");
228 }
229 
230 /*
231  *	vm_page_blacklist_next:
232  *
233  *	Find the next entry in the provided string of blacklist
234  *	addresses.  Entries are separated by space, comma, or newline.
235  *	If an invalid integer is encountered then the rest of the
236  *	string is skipped.  Updates the list pointer to the next
237  *	character, or NULL if the string is exhausted or invalid.
238  */
239 static vm_paddr_t
240 vm_page_blacklist_next(char **list, char *end)
241 {
242 	vm_paddr_t bad;
243 	char *cp, *pos;
244 
245 	if (list == NULL || *list == NULL)
246 		return (0);
247 	if (**list =='\0') {
248 		*list = NULL;
249 		return (0);
250 	}
251 
252 	/*
253 	 * If there's no end pointer then the buffer is coming from
254 	 * the kenv and we know it's null-terminated.
255 	 */
256 	if (end == NULL)
257 		end = *list + strlen(*list);
258 
259 	/* Ensure that strtoq() won't walk off the end */
260 	if (*end != '\0') {
261 		if (*end == '\n' || *end == ' ' || *end  == ',')
262 			*end = '\0';
263 		else {
264 			printf("Blacklist not terminated, skipping\n");
265 			*list = NULL;
266 			return (0);
267 		}
268 	}
269 
270 	for (pos = *list; *pos != '\0'; pos = cp) {
271 		bad = strtoq(pos, &cp, 0);
272 		if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') {
273 			if (bad == 0) {
274 				if (++cp < end)
275 					continue;
276 				else
277 					break;
278 			}
279 		} else
280 			break;
281 		if (*cp == '\0' || ++cp >= end)
282 			*list = NULL;
283 		else
284 			*list = cp;
285 		return (trunc_page(bad));
286 	}
287 	printf("Garbage in RAM blacklist, skipping\n");
288 	*list = NULL;
289 	return (0);
290 }
291 
292 /*
293  *	vm_page_blacklist_check:
294  *
295  *	Iterate through the provided string of blacklist addresses, pulling
296  *	each entry out of the physical allocator free list and putting it
297  *	onto a list for reporting via the vm.page_blacklist sysctl.
298  */
299 static void
300 vm_page_blacklist_check(char *list, char *end)
301 {
302 	vm_paddr_t pa;
303 	vm_page_t m;
304 	char *next;
305 	int ret;
306 
307 	next = list;
308 	while (next != NULL) {
309 		if ((pa = vm_page_blacklist_next(&next, end)) == 0)
310 			continue;
311 		m = vm_phys_paddr_to_vm_page(pa);
312 		if (m == NULL)
313 			continue;
314 		mtx_lock(&vm_page_queue_free_mtx);
315 		ret = vm_phys_unfree_page(m);
316 		mtx_unlock(&vm_page_queue_free_mtx);
317 		if (ret == TRUE) {
318 			TAILQ_INSERT_TAIL(&blacklist_head, m, listq);
319 			if (bootverbose)
320 				printf("Skipping page with pa 0x%jx\n",
321 				    (uintmax_t)pa);
322 		}
323 	}
324 }
325 
326 /*
327  *	vm_page_blacklist_load:
328  *
329  *	Search for a special module named "ram_blacklist".  It'll be a
330  *	plain text file provided by the user via the loader directive
331  *	of the same name.
332  */
333 static void
334 vm_page_blacklist_load(char **list, char **end)
335 {
336 	void *mod;
337 	u_char *ptr;
338 	u_int len;
339 
340 	mod = NULL;
341 	ptr = NULL;
342 
343 	mod = preload_search_by_type("ram_blacklist");
344 	if (mod != NULL) {
345 		ptr = preload_fetch_addr(mod);
346 		len = preload_fetch_size(mod);
347         }
348 	*list = ptr;
349 	if (ptr != NULL)
350 		*end = ptr + len;
351 	else
352 		*end = NULL;
353 	return;
354 }
355 
356 static int
357 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS)
358 {
359 	vm_page_t m;
360 	struct sbuf sbuf;
361 	int error, first;
362 
363 	first = 1;
364 	error = sysctl_wire_old_buffer(req, 0);
365 	if (error != 0)
366 		return (error);
367 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
368 	TAILQ_FOREACH(m, &blacklist_head, listq) {
369 		sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",",
370 		    (uintmax_t)m->phys_addr);
371 		first = 0;
372 	}
373 	error = sbuf_finish(&sbuf);
374 	sbuf_delete(&sbuf);
375 	return (error);
376 }
377 
378 static void
379 vm_page_domain_init(struct vm_domain *vmd)
380 {
381 	struct vm_pagequeue *pq;
382 	int i;
383 
384 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) =
385 	    "vm inactive pagequeue";
386 	*__DECONST(u_int **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_vcnt) =
387 	    &vm_cnt.v_inactive_count;
388 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) =
389 	    "vm active pagequeue";
390 	*__DECONST(u_int **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_vcnt) =
391 	    &vm_cnt.v_active_count;
392 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) =
393 	    "vm laundry pagequeue";
394 	*__DECONST(int **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_vcnt) =
395 	    &vm_cnt.v_laundry_count;
396 	vmd->vmd_page_count = 0;
397 	vmd->vmd_free_count = 0;
398 	vmd->vmd_segs = 0;
399 	vmd->vmd_oom = FALSE;
400 	for (i = 0; i < PQ_COUNT; i++) {
401 		pq = &vmd->vmd_pagequeues[i];
402 		TAILQ_INIT(&pq->pq_pl);
403 		mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue",
404 		    MTX_DEF | MTX_DUPOK);
405 	}
406 }
407 
408 /*
409  *	vm_page_startup:
410  *
411  *	Initializes the resident memory module.
412  *
413  *	Allocates memory for the page cells, and
414  *	for the object/offset-to-page hash table headers.
415  *	Each page cell is initialized and placed on the free list.
416  */
417 vm_offset_t
418 vm_page_startup(vm_offset_t vaddr)
419 {
420 	vm_offset_t mapped;
421 	vm_paddr_t page_range;
422 	vm_paddr_t new_end;
423 	int i;
424 	vm_paddr_t pa;
425 	vm_paddr_t last_pa;
426 	char *list, *listend;
427 	vm_paddr_t end;
428 	vm_paddr_t biggestsize;
429 	vm_paddr_t low_water, high_water;
430 	int biggestone;
431 	int pages_per_zone;
432 
433 	biggestsize = 0;
434 	biggestone = 0;
435 	vaddr = round_page(vaddr);
436 
437 	for (i = 0; phys_avail[i + 1]; i += 2) {
438 		phys_avail[i] = round_page(phys_avail[i]);
439 		phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
440 	}
441 
442 	low_water = phys_avail[0];
443 	high_water = phys_avail[1];
444 
445 	for (i = 0; i < vm_phys_nsegs; i++) {
446 		if (vm_phys_segs[i].start < low_water)
447 			low_water = vm_phys_segs[i].start;
448 		if (vm_phys_segs[i].end > high_water)
449 			high_water = vm_phys_segs[i].end;
450 	}
451 	for (i = 0; phys_avail[i + 1]; i += 2) {
452 		vm_paddr_t size = phys_avail[i + 1] - phys_avail[i];
453 
454 		if (size > biggestsize) {
455 			biggestone = i;
456 			biggestsize = size;
457 		}
458 		if (phys_avail[i] < low_water)
459 			low_water = phys_avail[i];
460 		if (phys_avail[i + 1] > high_water)
461 			high_water = phys_avail[i + 1];
462 	}
463 
464 	end = phys_avail[biggestone+1];
465 
466 	/*
467 	 * Initialize the page and queue locks.
468 	 */
469 	mtx_init(&vm_page_queue_free_mtx, "vm page free queue", NULL, MTX_DEF);
470 	for (i = 0; i < PA_LOCK_COUNT; i++)
471 		mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
472 	for (i = 0; i < vm_ndomains; i++)
473 		vm_page_domain_init(&vm_dom[i]);
474 
475 	/*
476 	 * Almost all of the pages needed for boot strapping UMA are used
477 	 * for zone structures, so if the number of CPUs results in those
478 	 * structures taking more than one page each, we set aside more pages
479 	 * in proportion to the zone structure size.
480 	 */
481 	pages_per_zone = howmany(sizeof(struct uma_zone) +
482 	    sizeof(struct uma_cache) * (mp_maxid + 1), UMA_SLAB_SIZE);
483 	if (pages_per_zone > 1) {
484 		/* Reserve more pages so that we don't run out. */
485 		boot_pages = UMA_BOOT_PAGES_ZONES * pages_per_zone;
486 	}
487 
488 	/*
489 	 * Allocate memory for use when boot strapping the kernel memory
490 	 * allocator.
491 	 *
492 	 * CTFLAG_RDTUN doesn't work during the early boot process, so we must
493 	 * manually fetch the value.
494 	 */
495 	TUNABLE_INT_FETCH("vm.boot_pages", &boot_pages);
496 	new_end = end - (boot_pages * UMA_SLAB_SIZE);
497 	new_end = trunc_page(new_end);
498 	mapped = pmap_map(&vaddr, new_end, end,
499 	    VM_PROT_READ | VM_PROT_WRITE);
500 	bzero((void *)mapped, end - new_end);
501 	uma_startup((void *)mapped, boot_pages);
502 
503 #if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \
504     defined(__i386__) || defined(__mips__)
505 	/*
506 	 * Allocate a bitmap to indicate that a random physical page
507 	 * needs to be included in a minidump.
508 	 *
509 	 * The amd64 port needs this to indicate which direct map pages
510 	 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
511 	 *
512 	 * However, i386 still needs this workspace internally within the
513 	 * minidump code.  In theory, they are not needed on i386, but are
514 	 * included should the sf_buf code decide to use them.
515 	 */
516 	last_pa = 0;
517 	for (i = 0; dump_avail[i + 1] != 0; i += 2)
518 		if (dump_avail[i + 1] > last_pa)
519 			last_pa = dump_avail[i + 1];
520 	page_range = last_pa / PAGE_SIZE;
521 	vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
522 	new_end -= vm_page_dump_size;
523 	vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
524 	    new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
525 	bzero((void *)vm_page_dump, vm_page_dump_size);
526 #endif
527 #ifdef __amd64__
528 	/*
529 	 * Request that the physical pages underlying the message buffer be
530 	 * included in a crash dump.  Since the message buffer is accessed
531 	 * through the direct map, they are not automatically included.
532 	 */
533 	pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
534 	last_pa = pa + round_page(msgbufsize);
535 	while (pa < last_pa) {
536 		dump_add_page(pa);
537 		pa += PAGE_SIZE;
538 	}
539 #endif
540 	/*
541 	 * Compute the number of pages of memory that will be available for
542 	 * use (taking into account the overhead of a page structure per
543 	 * page).
544 	 */
545 	first_page = low_water / PAGE_SIZE;
546 #ifdef VM_PHYSSEG_SPARSE
547 	page_range = 0;
548 	for (i = 0; i < vm_phys_nsegs; i++) {
549 		page_range += atop(vm_phys_segs[i].end -
550 		    vm_phys_segs[i].start);
551 	}
552 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
553 		page_range += atop(phys_avail[i + 1] - phys_avail[i]);
554 #elif defined(VM_PHYSSEG_DENSE)
555 	page_range = high_water / PAGE_SIZE - first_page;
556 #else
557 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
558 #endif
559 	end = new_end;
560 
561 	/*
562 	 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
563 	 */
564 	vaddr += PAGE_SIZE;
565 
566 	/*
567 	 * Initialize the mem entry structures now, and put them in the free
568 	 * queue.
569 	 */
570 	new_end = trunc_page(end - page_range * sizeof(struct vm_page));
571 	mapped = pmap_map(&vaddr, new_end, end,
572 	    VM_PROT_READ | VM_PROT_WRITE);
573 	vm_page_array = (vm_page_t) mapped;
574 #if VM_NRESERVLEVEL > 0
575 	/*
576 	 * Allocate memory for the reservation management system's data
577 	 * structures.
578 	 */
579 	new_end = vm_reserv_startup(&vaddr, new_end, high_water);
580 #endif
581 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__)
582 	/*
583 	 * pmap_map on arm64, amd64, and mips can come out of the direct-map,
584 	 * not kvm like i386, so the pages must be tracked for a crashdump to
585 	 * include this data.  This includes the vm_page_array and the early
586 	 * UMA bootstrap pages.
587 	 */
588 	for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE)
589 		dump_add_page(pa);
590 #endif
591 	phys_avail[biggestone + 1] = new_end;
592 
593 	/*
594 	 * Add physical memory segments corresponding to the available
595 	 * physical pages.
596 	 */
597 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
598 		vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]);
599 
600 	/*
601 	 * Clear all of the page structures
602 	 */
603 	bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
604 	for (i = 0; i < page_range; i++)
605 		vm_page_array[i].order = VM_NFREEORDER;
606 	vm_page_array_size = page_range;
607 
608 	/*
609 	 * Initialize the physical memory allocator.
610 	 */
611 	vm_phys_init();
612 
613 	/*
614 	 * Add every available physical page that is not blacklisted to
615 	 * the free lists.
616 	 */
617 	vm_cnt.v_page_count = 0;
618 	vm_cnt.v_free_count = 0;
619 	for (i = 0; phys_avail[i + 1] != 0; i += 2) {
620 		pa = phys_avail[i];
621 		last_pa = phys_avail[i + 1];
622 		while (pa < last_pa) {
623 			vm_phys_add_page(pa);
624 			pa += PAGE_SIZE;
625 		}
626 	}
627 
628 	TAILQ_INIT(&blacklist_head);
629 	vm_page_blacklist_load(&list, &listend);
630 	vm_page_blacklist_check(list, listend);
631 
632 	list = kern_getenv("vm.blacklist");
633 	vm_page_blacklist_check(list, NULL);
634 
635 	freeenv(list);
636 #if VM_NRESERVLEVEL > 0
637 	/*
638 	 * Initialize the reservation management system.
639 	 */
640 	vm_reserv_init();
641 #endif
642 	return (vaddr);
643 }
644 
645 void
646 vm_page_reference(vm_page_t m)
647 {
648 
649 	vm_page_aflag_set(m, PGA_REFERENCED);
650 }
651 
652 /*
653  *	vm_page_busy_downgrade:
654  *
655  *	Downgrade an exclusive busy page into a single shared busy page.
656  */
657 void
658 vm_page_busy_downgrade(vm_page_t m)
659 {
660 	u_int x;
661 	bool locked;
662 
663 	vm_page_assert_xbusied(m);
664 	locked = mtx_owned(vm_page_lockptr(m));
665 
666 	for (;;) {
667 		x = m->busy_lock;
668 		x &= VPB_BIT_WAITERS;
669 		if (x != 0 && !locked)
670 			vm_page_lock(m);
671 		if (atomic_cmpset_rel_int(&m->busy_lock,
672 		    VPB_SINGLE_EXCLUSIVER | x, VPB_SHARERS_WORD(1)))
673 			break;
674 		if (x != 0 && !locked)
675 			vm_page_unlock(m);
676 	}
677 	if (x != 0) {
678 		wakeup(m);
679 		if (!locked)
680 			vm_page_unlock(m);
681 	}
682 }
683 
684 /*
685  *	vm_page_sbusied:
686  *
687  *	Return a positive value if the page is shared busied, 0 otherwise.
688  */
689 int
690 vm_page_sbusied(vm_page_t m)
691 {
692 	u_int x;
693 
694 	x = m->busy_lock;
695 	return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED);
696 }
697 
698 /*
699  *	vm_page_sunbusy:
700  *
701  *	Shared unbusy a page.
702  */
703 void
704 vm_page_sunbusy(vm_page_t m)
705 {
706 	u_int x;
707 
708 	vm_page_assert_sbusied(m);
709 
710 	for (;;) {
711 		x = m->busy_lock;
712 		if (VPB_SHARERS(x) > 1) {
713 			if (atomic_cmpset_int(&m->busy_lock, x,
714 			    x - VPB_ONE_SHARER))
715 				break;
716 			continue;
717 		}
718 		if ((x & VPB_BIT_WAITERS) == 0) {
719 			KASSERT(x == VPB_SHARERS_WORD(1),
720 			    ("vm_page_sunbusy: invalid lock state"));
721 			if (atomic_cmpset_int(&m->busy_lock,
722 			    VPB_SHARERS_WORD(1), VPB_UNBUSIED))
723 				break;
724 			continue;
725 		}
726 		KASSERT(x == (VPB_SHARERS_WORD(1) | VPB_BIT_WAITERS),
727 		    ("vm_page_sunbusy: invalid lock state for waiters"));
728 
729 		vm_page_lock(m);
730 		if (!atomic_cmpset_int(&m->busy_lock, x, VPB_UNBUSIED)) {
731 			vm_page_unlock(m);
732 			continue;
733 		}
734 		wakeup(m);
735 		vm_page_unlock(m);
736 		break;
737 	}
738 }
739 
740 /*
741  *	vm_page_busy_sleep:
742  *
743  *	Sleep and release the page lock, using the page pointer as wchan.
744  *	This is used to implement the hard-path of busying mechanism.
745  *
746  *	The given page must be locked.
747  *
748  *	If nonshared is true, sleep only if the page is xbusy.
749  */
750 void
751 vm_page_busy_sleep(vm_page_t m, const char *wmesg, bool nonshared)
752 {
753 	u_int x;
754 
755 	vm_page_assert_locked(m);
756 
757 	x = m->busy_lock;
758 	if (x == VPB_UNBUSIED || (nonshared && (x & VPB_BIT_SHARED) != 0) ||
759 	    ((x & VPB_BIT_WAITERS) == 0 &&
760 	    !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS))) {
761 		vm_page_unlock(m);
762 		return;
763 	}
764 	msleep(m, vm_page_lockptr(m), PVM | PDROP, wmesg, 0);
765 }
766 
767 /*
768  *	vm_page_trysbusy:
769  *
770  *	Try to shared busy a page.
771  *	If the operation succeeds 1 is returned otherwise 0.
772  *	The operation never sleeps.
773  */
774 int
775 vm_page_trysbusy(vm_page_t m)
776 {
777 	u_int x;
778 
779 	for (;;) {
780 		x = m->busy_lock;
781 		if ((x & VPB_BIT_SHARED) == 0)
782 			return (0);
783 		if (atomic_cmpset_acq_int(&m->busy_lock, x, x + VPB_ONE_SHARER))
784 			return (1);
785 	}
786 }
787 
788 static void
789 vm_page_xunbusy_locked(vm_page_t m)
790 {
791 
792 	vm_page_assert_xbusied(m);
793 	vm_page_assert_locked(m);
794 
795 	atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
796 	/* There is a waiter, do wakeup() instead of vm_page_flash(). */
797 	wakeup(m);
798 }
799 
800 void
801 vm_page_xunbusy_maybelocked(vm_page_t m)
802 {
803 	bool lockacq;
804 
805 	vm_page_assert_xbusied(m);
806 
807 	/*
808 	 * Fast path for unbusy.  If it succeeds, we know that there
809 	 * are no waiters, so we do not need a wakeup.
810 	 */
811 	if (atomic_cmpset_rel_int(&m->busy_lock, VPB_SINGLE_EXCLUSIVER,
812 	    VPB_UNBUSIED))
813 		return;
814 
815 	lockacq = !mtx_owned(vm_page_lockptr(m));
816 	if (lockacq)
817 		vm_page_lock(m);
818 	vm_page_xunbusy_locked(m);
819 	if (lockacq)
820 		vm_page_unlock(m);
821 }
822 
823 /*
824  *	vm_page_xunbusy_hard:
825  *
826  *	Called after the first try the exclusive unbusy of a page failed.
827  *	It is assumed that the waiters bit is on.
828  */
829 void
830 vm_page_xunbusy_hard(vm_page_t m)
831 {
832 
833 	vm_page_assert_xbusied(m);
834 
835 	vm_page_lock(m);
836 	vm_page_xunbusy_locked(m);
837 	vm_page_unlock(m);
838 }
839 
840 /*
841  *	vm_page_flash:
842  *
843  *	Wakeup anyone waiting for the page.
844  *	The ownership bits do not change.
845  *
846  *	The given page must be locked.
847  */
848 void
849 vm_page_flash(vm_page_t m)
850 {
851 	u_int x;
852 
853 	vm_page_lock_assert(m, MA_OWNED);
854 
855 	for (;;) {
856 		x = m->busy_lock;
857 		if ((x & VPB_BIT_WAITERS) == 0)
858 			return;
859 		if (atomic_cmpset_int(&m->busy_lock, x,
860 		    x & (~VPB_BIT_WAITERS)))
861 			break;
862 	}
863 	wakeup(m);
864 }
865 
866 /*
867  * Keep page from being freed by the page daemon
868  * much of the same effect as wiring, except much lower
869  * overhead and should be used only for *very* temporary
870  * holding ("wiring").
871  */
872 void
873 vm_page_hold(vm_page_t mem)
874 {
875 
876 	vm_page_lock_assert(mem, MA_OWNED);
877         mem->hold_count++;
878 }
879 
880 void
881 vm_page_unhold(vm_page_t mem)
882 {
883 
884 	vm_page_lock_assert(mem, MA_OWNED);
885 	KASSERT(mem->hold_count >= 1, ("vm_page_unhold: hold count < 0!!!"));
886 	--mem->hold_count;
887 	if (mem->hold_count == 0 && (mem->flags & PG_UNHOLDFREE) != 0)
888 		vm_page_free_toq(mem);
889 }
890 
891 /*
892  *	vm_page_unhold_pages:
893  *
894  *	Unhold each of the pages that is referenced by the given array.
895  */
896 void
897 vm_page_unhold_pages(vm_page_t *ma, int count)
898 {
899 	struct mtx *mtx, *new_mtx;
900 
901 	mtx = NULL;
902 	for (; count != 0; count--) {
903 		/*
904 		 * Avoid releasing and reacquiring the same page lock.
905 		 */
906 		new_mtx = vm_page_lockptr(*ma);
907 		if (mtx != new_mtx) {
908 			if (mtx != NULL)
909 				mtx_unlock(mtx);
910 			mtx = new_mtx;
911 			mtx_lock(mtx);
912 		}
913 		vm_page_unhold(*ma);
914 		ma++;
915 	}
916 	if (mtx != NULL)
917 		mtx_unlock(mtx);
918 }
919 
920 vm_page_t
921 PHYS_TO_VM_PAGE(vm_paddr_t pa)
922 {
923 	vm_page_t m;
924 
925 #ifdef VM_PHYSSEG_SPARSE
926 	m = vm_phys_paddr_to_vm_page(pa);
927 	if (m == NULL)
928 		m = vm_phys_fictitious_to_vm_page(pa);
929 	return (m);
930 #elif defined(VM_PHYSSEG_DENSE)
931 	long pi;
932 
933 	pi = atop(pa);
934 	if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
935 		m = &vm_page_array[pi - first_page];
936 		return (m);
937 	}
938 	return (vm_phys_fictitious_to_vm_page(pa));
939 #else
940 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
941 #endif
942 }
943 
944 /*
945  *	vm_page_getfake:
946  *
947  *	Create a fictitious page with the specified physical address and
948  *	memory attribute.  The memory attribute is the only the machine-
949  *	dependent aspect of a fictitious page that must be initialized.
950  */
951 vm_page_t
952 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
953 {
954 	vm_page_t m;
955 
956 	m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
957 	vm_page_initfake(m, paddr, memattr);
958 	return (m);
959 }
960 
961 void
962 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
963 {
964 
965 	if ((m->flags & PG_FICTITIOUS) != 0) {
966 		/*
967 		 * The page's memattr might have changed since the
968 		 * previous initialization.  Update the pmap to the
969 		 * new memattr.
970 		 */
971 		goto memattr;
972 	}
973 	m->phys_addr = paddr;
974 	m->queue = PQ_NONE;
975 	/* Fictitious pages don't use "segind". */
976 	m->flags = PG_FICTITIOUS;
977 	/* Fictitious pages don't use "order" or "pool". */
978 	m->oflags = VPO_UNMANAGED;
979 	m->busy_lock = VPB_SINGLE_EXCLUSIVER;
980 	m->wire_count = 1;
981 	pmap_page_init(m);
982 memattr:
983 	pmap_page_set_memattr(m, memattr);
984 }
985 
986 /*
987  *	vm_page_putfake:
988  *
989  *	Release a fictitious page.
990  */
991 void
992 vm_page_putfake(vm_page_t m)
993 {
994 
995 	KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
996 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
997 	    ("vm_page_putfake: bad page %p", m));
998 	uma_zfree(fakepg_zone, m);
999 }
1000 
1001 /*
1002  *	vm_page_updatefake:
1003  *
1004  *	Update the given fictitious page to the specified physical address and
1005  *	memory attribute.
1006  */
1007 void
1008 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1009 {
1010 
1011 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1012 	    ("vm_page_updatefake: bad page %p", m));
1013 	m->phys_addr = paddr;
1014 	pmap_page_set_memattr(m, memattr);
1015 }
1016 
1017 /*
1018  *	vm_page_free:
1019  *
1020  *	Free a page.
1021  */
1022 void
1023 vm_page_free(vm_page_t m)
1024 {
1025 
1026 	m->flags &= ~PG_ZERO;
1027 	vm_page_free_toq(m);
1028 }
1029 
1030 /*
1031  *	vm_page_free_zero:
1032  *
1033  *	Free a page to the zerod-pages queue
1034  */
1035 void
1036 vm_page_free_zero(vm_page_t m)
1037 {
1038 
1039 	m->flags |= PG_ZERO;
1040 	vm_page_free_toq(m);
1041 }
1042 
1043 /*
1044  * Unbusy and handle the page queueing for a page from a getpages request that
1045  * was optionally read ahead or behind.
1046  */
1047 void
1048 vm_page_readahead_finish(vm_page_t m)
1049 {
1050 
1051 	/* We shouldn't put invalid pages on queues. */
1052 	KASSERT(m->valid != 0, ("%s: %p is invalid", __func__, m));
1053 
1054 	/*
1055 	 * Since the page is not the actually needed one, whether it should
1056 	 * be activated or deactivated is not obvious.  Empirical results
1057 	 * have shown that deactivating the page is usually the best choice,
1058 	 * unless the page is wanted by another thread.
1059 	 */
1060 	vm_page_lock(m);
1061 	if ((m->busy_lock & VPB_BIT_WAITERS) != 0)
1062 		vm_page_activate(m);
1063 	else
1064 		vm_page_deactivate(m);
1065 	vm_page_unlock(m);
1066 	vm_page_xunbusy(m);
1067 }
1068 
1069 /*
1070  *	vm_page_sleep_if_busy:
1071  *
1072  *	Sleep and release the page queues lock if the page is busied.
1073  *	Returns TRUE if the thread slept.
1074  *
1075  *	The given page must be unlocked and object containing it must
1076  *	be locked.
1077  */
1078 int
1079 vm_page_sleep_if_busy(vm_page_t m, const char *msg)
1080 {
1081 	vm_object_t obj;
1082 
1083 	vm_page_lock_assert(m, MA_NOTOWNED);
1084 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1085 
1086 	if (vm_page_busied(m)) {
1087 		/*
1088 		 * The page-specific object must be cached because page
1089 		 * identity can change during the sleep, causing the
1090 		 * re-lock of a different object.
1091 		 * It is assumed that a reference to the object is already
1092 		 * held by the callers.
1093 		 */
1094 		obj = m->object;
1095 		vm_page_lock(m);
1096 		VM_OBJECT_WUNLOCK(obj);
1097 		vm_page_busy_sleep(m, msg, false);
1098 		VM_OBJECT_WLOCK(obj);
1099 		return (TRUE);
1100 	}
1101 	return (FALSE);
1102 }
1103 
1104 /*
1105  *	vm_page_dirty_KBI:		[ internal use only ]
1106  *
1107  *	Set all bits in the page's dirty field.
1108  *
1109  *	The object containing the specified page must be locked if the
1110  *	call is made from the machine-independent layer.
1111  *
1112  *	See vm_page_clear_dirty_mask().
1113  *
1114  *	This function should only be called by vm_page_dirty().
1115  */
1116 void
1117 vm_page_dirty_KBI(vm_page_t m)
1118 {
1119 
1120 	/* Refer to this operation by its public name. */
1121 	KASSERT(m->valid == VM_PAGE_BITS_ALL,
1122 	    ("vm_page_dirty: page is invalid!"));
1123 	m->dirty = VM_PAGE_BITS_ALL;
1124 }
1125 
1126 /*
1127  *	vm_page_insert:		[ internal use only ]
1128  *
1129  *	Inserts the given mem entry into the object and object list.
1130  *
1131  *	The object must be locked.
1132  */
1133 int
1134 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1135 {
1136 	vm_page_t mpred;
1137 
1138 	VM_OBJECT_ASSERT_WLOCKED(object);
1139 	mpred = vm_radix_lookup_le(&object->rtree, pindex);
1140 	return (vm_page_insert_after(m, object, pindex, mpred));
1141 }
1142 
1143 /*
1144  *	vm_page_insert_after:
1145  *
1146  *	Inserts the page "m" into the specified object at offset "pindex".
1147  *
1148  *	The page "mpred" must immediately precede the offset "pindex" within
1149  *	the specified object.
1150  *
1151  *	The object must be locked.
1152  */
1153 static int
1154 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1155     vm_page_t mpred)
1156 {
1157 	vm_page_t msucc;
1158 
1159 	VM_OBJECT_ASSERT_WLOCKED(object);
1160 	KASSERT(m->object == NULL,
1161 	    ("vm_page_insert_after: page already inserted"));
1162 	if (mpred != NULL) {
1163 		KASSERT(mpred->object == object,
1164 		    ("vm_page_insert_after: object doesn't contain mpred"));
1165 		KASSERT(mpred->pindex < pindex,
1166 		    ("vm_page_insert_after: mpred doesn't precede pindex"));
1167 		msucc = TAILQ_NEXT(mpred, listq);
1168 	} else
1169 		msucc = TAILQ_FIRST(&object->memq);
1170 	if (msucc != NULL)
1171 		KASSERT(msucc->pindex > pindex,
1172 		    ("vm_page_insert_after: msucc doesn't succeed pindex"));
1173 
1174 	/*
1175 	 * Record the object/offset pair in this page
1176 	 */
1177 	m->object = object;
1178 	m->pindex = pindex;
1179 
1180 	/*
1181 	 * Now link into the object's ordered list of backed pages.
1182 	 */
1183 	if (vm_radix_insert(&object->rtree, m)) {
1184 		m->object = NULL;
1185 		m->pindex = 0;
1186 		return (1);
1187 	}
1188 	vm_page_insert_radixdone(m, object, mpred);
1189 	return (0);
1190 }
1191 
1192 /*
1193  *	vm_page_insert_radixdone:
1194  *
1195  *	Complete page "m" insertion into the specified object after the
1196  *	radix trie hooking.
1197  *
1198  *	The page "mpred" must precede the offset "m->pindex" within the
1199  *	specified object.
1200  *
1201  *	The object must be locked.
1202  */
1203 static void
1204 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1205 {
1206 
1207 	VM_OBJECT_ASSERT_WLOCKED(object);
1208 	KASSERT(object != NULL && m->object == object,
1209 	    ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1210 	if (mpred != NULL) {
1211 		KASSERT(mpred->object == object,
1212 		    ("vm_page_insert_after: object doesn't contain mpred"));
1213 		KASSERT(mpred->pindex < m->pindex,
1214 		    ("vm_page_insert_after: mpred doesn't precede pindex"));
1215 	}
1216 
1217 	if (mpred != NULL)
1218 		TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1219 	else
1220 		TAILQ_INSERT_HEAD(&object->memq, m, listq);
1221 
1222 	/*
1223 	 * Show that the object has one more resident page.
1224 	 */
1225 	object->resident_page_count++;
1226 
1227 	/*
1228 	 * Hold the vnode until the last page is released.
1229 	 */
1230 	if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1231 		vhold(object->handle);
1232 
1233 	/*
1234 	 * Since we are inserting a new and possibly dirty page,
1235 	 * update the object's OBJ_MIGHTBEDIRTY flag.
1236 	 */
1237 	if (pmap_page_is_write_mapped(m))
1238 		vm_object_set_writeable_dirty(object);
1239 }
1240 
1241 /*
1242  *	vm_page_remove:
1243  *
1244  *	Removes the given mem entry from the object/offset-page
1245  *	table and the object page list, but do not invalidate/terminate
1246  *	the backing store.
1247  *
1248  *	The object must be locked.  The page must be locked if it is managed.
1249  */
1250 void
1251 vm_page_remove(vm_page_t m)
1252 {
1253 	vm_object_t object;
1254 
1255 	if ((m->oflags & VPO_UNMANAGED) == 0)
1256 		vm_page_assert_locked(m);
1257 	if ((object = m->object) == NULL)
1258 		return;
1259 	VM_OBJECT_ASSERT_WLOCKED(object);
1260 	if (vm_page_xbusied(m))
1261 		vm_page_xunbusy_maybelocked(m);
1262 
1263 	/*
1264 	 * Now remove from the object's list of backed pages.
1265 	 */
1266 	vm_radix_remove(&object->rtree, m->pindex);
1267 	TAILQ_REMOVE(&object->memq, m, listq);
1268 
1269 	/*
1270 	 * And show that the object has one fewer resident page.
1271 	 */
1272 	object->resident_page_count--;
1273 
1274 	/*
1275 	 * The vnode may now be recycled.
1276 	 */
1277 	if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1278 		vdrop(object->handle);
1279 
1280 	m->object = NULL;
1281 }
1282 
1283 /*
1284  *	vm_page_lookup:
1285  *
1286  *	Returns the page associated with the object/offset
1287  *	pair specified; if none is found, NULL is returned.
1288  *
1289  *	The object must be locked.
1290  */
1291 vm_page_t
1292 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1293 {
1294 
1295 	VM_OBJECT_ASSERT_LOCKED(object);
1296 	return (vm_radix_lookup(&object->rtree, pindex));
1297 }
1298 
1299 /*
1300  *	vm_page_find_least:
1301  *
1302  *	Returns the page associated with the object with least pindex
1303  *	greater than or equal to the parameter pindex, or NULL.
1304  *
1305  *	The object must be locked.
1306  */
1307 vm_page_t
1308 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1309 {
1310 	vm_page_t m;
1311 
1312 	VM_OBJECT_ASSERT_LOCKED(object);
1313 	if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1314 		m = vm_radix_lookup_ge(&object->rtree, pindex);
1315 	return (m);
1316 }
1317 
1318 /*
1319  * Returns the given page's successor (by pindex) within the object if it is
1320  * resident; if none is found, NULL is returned.
1321  *
1322  * The object must be locked.
1323  */
1324 vm_page_t
1325 vm_page_next(vm_page_t m)
1326 {
1327 	vm_page_t next;
1328 
1329 	VM_OBJECT_ASSERT_LOCKED(m->object);
1330 	if ((next = TAILQ_NEXT(m, listq)) != NULL &&
1331 	    next->pindex != m->pindex + 1)
1332 		next = NULL;
1333 	return (next);
1334 }
1335 
1336 /*
1337  * Returns the given page's predecessor (by pindex) within the object if it is
1338  * resident; if none is found, NULL is returned.
1339  *
1340  * The object must be locked.
1341  */
1342 vm_page_t
1343 vm_page_prev(vm_page_t m)
1344 {
1345 	vm_page_t prev;
1346 
1347 	VM_OBJECT_ASSERT_LOCKED(m->object);
1348 	if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL &&
1349 	    prev->pindex != m->pindex - 1)
1350 		prev = NULL;
1351 	return (prev);
1352 }
1353 
1354 /*
1355  * Uses the page mnew as a replacement for an existing page at index
1356  * pindex which must be already present in the object.
1357  *
1358  * The existing page must not be on a paging queue.
1359  */
1360 vm_page_t
1361 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex)
1362 {
1363 	vm_page_t mold;
1364 
1365 	VM_OBJECT_ASSERT_WLOCKED(object);
1366 	KASSERT(mnew->object == NULL,
1367 	    ("vm_page_replace: page already in object"));
1368 
1369 	/*
1370 	 * This function mostly follows vm_page_insert() and
1371 	 * vm_page_remove() without the radix, object count and vnode
1372 	 * dance.  Double check such functions for more comments.
1373 	 */
1374 
1375 	mnew->object = object;
1376 	mnew->pindex = pindex;
1377 	mold = vm_radix_replace(&object->rtree, mnew);
1378 	KASSERT(mold->queue == PQ_NONE,
1379 	    ("vm_page_replace: mold is on a paging queue"));
1380 
1381 	/* Keep the resident page list in sorted order. */
1382 	TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq);
1383 	TAILQ_REMOVE(&object->memq, mold, listq);
1384 
1385 	mold->object = NULL;
1386 	vm_page_xunbusy_maybelocked(mold);
1387 
1388 	/*
1389 	 * The object's resident_page_count does not change because we have
1390 	 * swapped one page for another, but OBJ_MIGHTBEDIRTY.
1391 	 */
1392 	if (pmap_page_is_write_mapped(mnew))
1393 		vm_object_set_writeable_dirty(object);
1394 	return (mold);
1395 }
1396 
1397 /*
1398  *	vm_page_rename:
1399  *
1400  *	Move the given memory entry from its
1401  *	current object to the specified target object/offset.
1402  *
1403  *	Note: swap associated with the page must be invalidated by the move.  We
1404  *	      have to do this for several reasons:  (1) we aren't freeing the
1405  *	      page, (2) we are dirtying the page, (3) the VM system is probably
1406  *	      moving the page from object A to B, and will then later move
1407  *	      the backing store from A to B and we can't have a conflict.
1408  *
1409  *	Note: we *always* dirty the page.  It is necessary both for the
1410  *	      fact that we moved it, and because we may be invalidating
1411  *	      swap.  If the page is on the cache, we have to deactivate it
1412  *	      or vm_page_dirty() will panic.  Dirty pages are not allowed
1413  *	      on the cache.
1414  *
1415  *	The objects must be locked.
1416  */
1417 int
1418 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1419 {
1420 	vm_page_t mpred;
1421 	vm_pindex_t opidx;
1422 
1423 	VM_OBJECT_ASSERT_WLOCKED(new_object);
1424 
1425 	mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex);
1426 	KASSERT(mpred == NULL || mpred->pindex != new_pindex,
1427 	    ("vm_page_rename: pindex already renamed"));
1428 
1429 	/*
1430 	 * Create a custom version of vm_page_insert() which does not depend
1431 	 * by m_prev and can cheat on the implementation aspects of the
1432 	 * function.
1433 	 */
1434 	opidx = m->pindex;
1435 	m->pindex = new_pindex;
1436 	if (vm_radix_insert(&new_object->rtree, m)) {
1437 		m->pindex = opidx;
1438 		return (1);
1439 	}
1440 
1441 	/*
1442 	 * The operation cannot fail anymore.  The removal must happen before
1443 	 * the listq iterator is tainted.
1444 	 */
1445 	m->pindex = opidx;
1446 	vm_page_lock(m);
1447 	vm_page_remove(m);
1448 
1449 	/* Return back to the new pindex to complete vm_page_insert(). */
1450 	m->pindex = new_pindex;
1451 	m->object = new_object;
1452 	vm_page_unlock(m);
1453 	vm_page_insert_radixdone(m, new_object, mpred);
1454 	vm_page_dirty(m);
1455 	return (0);
1456 }
1457 
1458 /*
1459  *	vm_page_alloc:
1460  *
1461  *	Allocate and return a page that is associated with the specified
1462  *	object and offset pair.  By default, this page is exclusive busied.
1463  *
1464  *	The caller must always specify an allocation class.
1465  *
1466  *	allocation classes:
1467  *	VM_ALLOC_NORMAL		normal process request
1468  *	VM_ALLOC_SYSTEM		system *really* needs a page
1469  *	VM_ALLOC_INTERRUPT	interrupt time request
1470  *
1471  *	optional allocation flags:
1472  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
1473  *				intends to allocate
1474  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
1475  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
1476  *	VM_ALLOC_NOOBJ		page is not associated with an object and
1477  *				should not be exclusive busy
1478  *	VM_ALLOC_SBUSY		shared busy the allocated page
1479  *	VM_ALLOC_WIRED		wire the allocated page
1480  *	VM_ALLOC_ZERO		prefer a zeroed page
1481  *
1482  *	This routine may not sleep.
1483  */
1484 vm_page_t
1485 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1486 {
1487 	vm_page_t m, mpred;
1488 	int flags, req_class;
1489 
1490 	mpred = 0;	/* XXX: pacify gcc */
1491 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1492 	    (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1493 	    ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1494 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1495 	    ("vm_page_alloc: inconsistent object(%p)/req(%x)", (void *)object,
1496 	    req));
1497 	if (object != NULL)
1498 		VM_OBJECT_ASSERT_WLOCKED(object);
1499 
1500 	req_class = req & VM_ALLOC_CLASS_MASK;
1501 
1502 	/*
1503 	 * The page daemon is allowed to dig deeper into the free page list.
1504 	 */
1505 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1506 		req_class = VM_ALLOC_SYSTEM;
1507 
1508 	if (object != NULL) {
1509 		mpred = vm_radix_lookup_le(&object->rtree, pindex);
1510 		KASSERT(mpred == NULL || mpred->pindex != pindex,
1511 		   ("vm_page_alloc: pindex already allocated"));
1512 	}
1513 
1514 	/*
1515 	 * Allocate a page if the number of free pages exceeds the minimum
1516 	 * for the request class.
1517 	 */
1518 	mtx_lock(&vm_page_queue_free_mtx);
1519 	if (vm_cnt.v_free_count > vm_cnt.v_free_reserved ||
1520 	    (req_class == VM_ALLOC_SYSTEM &&
1521 	    vm_cnt.v_free_count > vm_cnt.v_interrupt_free_min) ||
1522 	    (req_class == VM_ALLOC_INTERRUPT &&
1523 	    vm_cnt.v_free_count > 0)) {
1524 		/*
1525 		 * Can we allocate the page from a reservation?
1526 		 */
1527 #if VM_NRESERVLEVEL > 0
1528 		if (object == NULL || (object->flags & (OBJ_COLORED |
1529 		    OBJ_FICTITIOUS)) != OBJ_COLORED || (m =
1530 		    vm_reserv_alloc_page(object, pindex, mpred)) == NULL)
1531 #endif
1532 		{
1533 			/*
1534 			 * If not, allocate it from the free page queues.
1535 			 */
1536 			m = vm_phys_alloc_pages(object != NULL ?
1537 			    VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0);
1538 #if VM_NRESERVLEVEL > 0
1539 			if (m == NULL && vm_reserv_reclaim_inactive()) {
1540 				m = vm_phys_alloc_pages(object != NULL ?
1541 				    VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT,
1542 				    0);
1543 			}
1544 #endif
1545 		}
1546 	} else {
1547 		/*
1548 		 * Not allocatable, give up.
1549 		 */
1550 		mtx_unlock(&vm_page_queue_free_mtx);
1551 		atomic_add_int(&vm_pageout_deficit,
1552 		    max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
1553 		pagedaemon_wakeup();
1554 		return (NULL);
1555 	}
1556 
1557 	/*
1558 	 *  At this point we had better have found a good page.
1559 	 */
1560 	KASSERT(m != NULL, ("vm_page_alloc: missing page"));
1561 	vm_phys_freecnt_adj(m, -1);
1562 	mtx_unlock(&vm_page_queue_free_mtx);
1563 	vm_page_alloc_check(m);
1564 
1565 	/*
1566 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
1567 	 */
1568 	flags = 0;
1569 	if ((req & VM_ALLOC_ZERO) != 0)
1570 		flags = PG_ZERO;
1571 	flags &= m->flags;
1572 	if ((req & VM_ALLOC_NODUMP) != 0)
1573 		flags |= PG_NODUMP;
1574 	m->flags = flags;
1575 	m->aflags = 0;
1576 	m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
1577 	    VPO_UNMANAGED : 0;
1578 	m->busy_lock = VPB_UNBUSIED;
1579 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
1580 		m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1581 	if ((req & VM_ALLOC_SBUSY) != 0)
1582 		m->busy_lock = VPB_SHARERS_WORD(1);
1583 	if (req & VM_ALLOC_WIRED) {
1584 		/*
1585 		 * The page lock is not required for wiring a page until that
1586 		 * page is inserted into the object.
1587 		 */
1588 		atomic_add_int(&vm_cnt.v_wire_count, 1);
1589 		m->wire_count = 1;
1590 	}
1591 	m->act_count = 0;
1592 
1593 	if (object != NULL) {
1594 		if (vm_page_insert_after(m, object, pindex, mpred)) {
1595 			pagedaemon_wakeup();
1596 			if (req & VM_ALLOC_WIRED) {
1597 				atomic_subtract_int(&vm_cnt.v_wire_count, 1);
1598 				m->wire_count = 0;
1599 			}
1600 			m->object = NULL;
1601 			m->oflags = VPO_UNMANAGED;
1602 			m->busy_lock = VPB_UNBUSIED;
1603 			vm_page_free(m);
1604 			return (NULL);
1605 		}
1606 
1607 		/* Ignore device objects; the pager sets "memattr" for them. */
1608 		if (object->memattr != VM_MEMATTR_DEFAULT &&
1609 		    (object->flags & OBJ_FICTITIOUS) == 0)
1610 			pmap_page_set_memattr(m, object->memattr);
1611 	} else
1612 		m->pindex = pindex;
1613 
1614 	/*
1615 	 * Don't wakeup too often - wakeup the pageout daemon when
1616 	 * we would be nearly out of memory.
1617 	 */
1618 	if (vm_paging_needed())
1619 		pagedaemon_wakeup();
1620 
1621 	return (m);
1622 }
1623 
1624 /*
1625  *	vm_page_alloc_contig:
1626  *
1627  *	Allocate a contiguous set of physical pages of the given size "npages"
1628  *	from the free lists.  All of the physical pages must be at or above
1629  *	the given physical address "low" and below the given physical address
1630  *	"high".  The given value "alignment" determines the alignment of the
1631  *	first physical page in the set.  If the given value "boundary" is
1632  *	non-zero, then the set of physical pages cannot cross any physical
1633  *	address boundary that is a multiple of that value.  Both "alignment"
1634  *	and "boundary" must be a power of two.
1635  *
1636  *	If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
1637  *	then the memory attribute setting for the physical pages is configured
1638  *	to the object's memory attribute setting.  Otherwise, the memory
1639  *	attribute setting for the physical pages is configured to "memattr",
1640  *	overriding the object's memory attribute setting.  However, if the
1641  *	object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
1642  *	memory attribute setting for the physical pages cannot be configured
1643  *	to VM_MEMATTR_DEFAULT.
1644  *
1645  *	The caller must always specify an allocation class.
1646  *
1647  *	allocation classes:
1648  *	VM_ALLOC_NORMAL		normal process request
1649  *	VM_ALLOC_SYSTEM		system *really* needs a page
1650  *	VM_ALLOC_INTERRUPT	interrupt time request
1651  *
1652  *	optional allocation flags:
1653  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
1654  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
1655  *	VM_ALLOC_NOOBJ		page is not associated with an object and
1656  *				should not be exclusive busy
1657  *	VM_ALLOC_SBUSY		shared busy the allocated page
1658  *	VM_ALLOC_WIRED		wire the allocated page
1659  *	VM_ALLOC_ZERO		prefer a zeroed page
1660  *
1661  *	This routine may not sleep.
1662  */
1663 vm_page_t
1664 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
1665     u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
1666     vm_paddr_t boundary, vm_memattr_t memattr)
1667 {
1668 	vm_page_t m, m_tmp, m_ret;
1669 	u_int flags;
1670 	int req_class;
1671 
1672 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1673 	    (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1674 	    ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1675 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1676 	    ("vm_page_alloc: inconsistent object(%p)/req(%x)", (void *)object,
1677 	    req));
1678 	if (object != NULL) {
1679 		VM_OBJECT_ASSERT_WLOCKED(object);
1680 		KASSERT(object->type == OBJT_PHYS,
1681 		    ("vm_page_alloc_contig: object %p isn't OBJT_PHYS",
1682 		    object));
1683 	}
1684 	KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
1685 	req_class = req & VM_ALLOC_CLASS_MASK;
1686 
1687 	/*
1688 	 * The page daemon is allowed to dig deeper into the free page list.
1689 	 */
1690 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1691 		req_class = VM_ALLOC_SYSTEM;
1692 
1693 	mtx_lock(&vm_page_queue_free_mtx);
1694 	if (vm_cnt.v_free_count >= npages + vm_cnt.v_free_reserved ||
1695 	    (req_class == VM_ALLOC_SYSTEM &&
1696 	    vm_cnt.v_free_count >= npages + vm_cnt.v_interrupt_free_min) ||
1697 	    (req_class == VM_ALLOC_INTERRUPT &&
1698 	    vm_cnt.v_free_count >= npages)) {
1699 #if VM_NRESERVLEVEL > 0
1700 retry:
1701 		if (object == NULL || (object->flags & OBJ_COLORED) == 0 ||
1702 		    (m_ret = vm_reserv_alloc_contig(object, pindex, npages,
1703 		    low, high, alignment, boundary)) == NULL)
1704 #endif
1705 			m_ret = vm_phys_alloc_contig(npages, low, high,
1706 			    alignment, boundary);
1707 	} else {
1708 		mtx_unlock(&vm_page_queue_free_mtx);
1709 		atomic_add_int(&vm_pageout_deficit, npages);
1710 		pagedaemon_wakeup();
1711 		return (NULL);
1712 	}
1713 	if (m_ret != NULL)
1714 		vm_phys_freecnt_adj(m_ret, -npages);
1715 	else {
1716 #if VM_NRESERVLEVEL > 0
1717 		if (vm_reserv_reclaim_contig(npages, low, high, alignment,
1718 		    boundary))
1719 			goto retry;
1720 #endif
1721 	}
1722 	mtx_unlock(&vm_page_queue_free_mtx);
1723 	if (m_ret == NULL)
1724 		return (NULL);
1725 	for (m = m_ret; m < &m_ret[npages]; m++)
1726 		vm_page_alloc_check(m);
1727 
1728 	/*
1729 	 * Initialize the pages.  Only the PG_ZERO flag is inherited.
1730 	 */
1731 	flags = 0;
1732 	if ((req & VM_ALLOC_ZERO) != 0)
1733 		flags = PG_ZERO;
1734 	if ((req & VM_ALLOC_NODUMP) != 0)
1735 		flags |= PG_NODUMP;
1736 	if ((req & VM_ALLOC_WIRED) != 0)
1737 		atomic_add_int(&vm_cnt.v_wire_count, npages);
1738 	if (object != NULL) {
1739 		if (object->memattr != VM_MEMATTR_DEFAULT &&
1740 		    memattr == VM_MEMATTR_DEFAULT)
1741 			memattr = object->memattr;
1742 	}
1743 	for (m = m_ret; m < &m_ret[npages]; m++) {
1744 		m->aflags = 0;
1745 		m->flags = (m->flags | PG_NODUMP) & flags;
1746 		m->busy_lock = VPB_UNBUSIED;
1747 		if (object != NULL) {
1748 			if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
1749 				m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1750 			if ((req & VM_ALLOC_SBUSY) != 0)
1751 				m->busy_lock = VPB_SHARERS_WORD(1);
1752 		}
1753 		if ((req & VM_ALLOC_WIRED) != 0)
1754 			m->wire_count = 1;
1755 		/* Unmanaged pages don't use "act_count". */
1756 		m->oflags = VPO_UNMANAGED;
1757 		if (object != NULL) {
1758 			if (vm_page_insert(m, object, pindex)) {
1759 				if (vm_paging_needed())
1760 					pagedaemon_wakeup();
1761 				if ((req & VM_ALLOC_WIRED) != 0)
1762 					atomic_subtract_int(&vm_cnt.v_wire_count,
1763 					    npages);
1764 				for (m_tmp = m, m = m_ret;
1765 				    m < &m_ret[npages]; m++) {
1766 					if ((req & VM_ALLOC_WIRED) != 0)
1767 						m->wire_count = 0;
1768 					if (m >= m_tmp) {
1769 						m->object = NULL;
1770 						m->oflags |= VPO_UNMANAGED;
1771 					}
1772 					m->busy_lock = VPB_UNBUSIED;
1773 					vm_page_free(m);
1774 				}
1775 				return (NULL);
1776 			}
1777 		} else
1778 			m->pindex = pindex;
1779 		if (memattr != VM_MEMATTR_DEFAULT)
1780 			pmap_page_set_memattr(m, memattr);
1781 		pindex++;
1782 	}
1783 	if (vm_paging_needed())
1784 		pagedaemon_wakeup();
1785 	return (m_ret);
1786 }
1787 
1788 /*
1789  * Check a page that has been freshly dequeued from a freelist.
1790  */
1791 static void
1792 vm_page_alloc_check(vm_page_t m)
1793 {
1794 
1795 	KASSERT(m->queue == PQ_NONE,
1796 	    ("page %p has unexpected queue %d", m, m->queue));
1797 	KASSERT(m->wire_count == 0, ("page %p is wired", m));
1798 	KASSERT(m->hold_count == 0, ("page %p is held", m));
1799 	KASSERT(!vm_page_busied(m), ("page %p is busy", m));
1800 	KASSERT(m->dirty == 0, ("page %p is dirty", m));
1801 	KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
1802 	    ("page %p has unexpected memattr %d",
1803 	    m, pmap_page_get_memattr(m)));
1804 	KASSERT(m->valid == 0, ("free page %p is valid", m));
1805 }
1806 
1807 /*
1808  * 	vm_page_alloc_freelist:
1809  *
1810  *	Allocate a physical page from the specified free page list.
1811  *
1812  *	The caller must always specify an allocation class.
1813  *
1814  *	allocation classes:
1815  *	VM_ALLOC_NORMAL		normal process request
1816  *	VM_ALLOC_SYSTEM		system *really* needs a page
1817  *	VM_ALLOC_INTERRUPT	interrupt time request
1818  *
1819  *	optional allocation flags:
1820  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
1821  *				intends to allocate
1822  *	VM_ALLOC_WIRED		wire the allocated page
1823  *	VM_ALLOC_ZERO		prefer a zeroed page
1824  *
1825  *	This routine may not sleep.
1826  */
1827 vm_page_t
1828 vm_page_alloc_freelist(int flind, int req)
1829 {
1830 	vm_page_t m;
1831 	u_int flags;
1832 	int req_class;
1833 
1834 	req_class = req & VM_ALLOC_CLASS_MASK;
1835 
1836 	/*
1837 	 * The page daemon is allowed to dig deeper into the free page list.
1838 	 */
1839 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1840 		req_class = VM_ALLOC_SYSTEM;
1841 
1842 	/*
1843 	 * Do not allocate reserved pages unless the req has asked for it.
1844 	 */
1845 	mtx_lock(&vm_page_queue_free_mtx);
1846 	if (vm_cnt.v_free_count > vm_cnt.v_free_reserved ||
1847 	    (req_class == VM_ALLOC_SYSTEM &&
1848 	    vm_cnt.v_free_count > vm_cnt.v_interrupt_free_min) ||
1849 	    (req_class == VM_ALLOC_INTERRUPT &&
1850 	    vm_cnt.v_free_count > 0))
1851 		m = vm_phys_alloc_freelist_pages(flind, VM_FREEPOOL_DIRECT, 0);
1852 	else {
1853 		mtx_unlock(&vm_page_queue_free_mtx);
1854 		atomic_add_int(&vm_pageout_deficit,
1855 		    max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
1856 		pagedaemon_wakeup();
1857 		return (NULL);
1858 	}
1859 	if (m == NULL) {
1860 		mtx_unlock(&vm_page_queue_free_mtx);
1861 		return (NULL);
1862 	}
1863 	vm_phys_freecnt_adj(m, -1);
1864 	mtx_unlock(&vm_page_queue_free_mtx);
1865 	vm_page_alloc_check(m);
1866 
1867 	/*
1868 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
1869 	 */
1870 	m->aflags = 0;
1871 	flags = 0;
1872 	if ((req & VM_ALLOC_ZERO) != 0)
1873 		flags = PG_ZERO;
1874 	m->flags &= flags;
1875 	if ((req & VM_ALLOC_WIRED) != 0) {
1876 		/*
1877 		 * The page lock is not required for wiring a page that does
1878 		 * not belong to an object.
1879 		 */
1880 		atomic_add_int(&vm_cnt.v_wire_count, 1);
1881 		m->wire_count = 1;
1882 	}
1883 	/* Unmanaged pages don't use "act_count". */
1884 	m->oflags = VPO_UNMANAGED;
1885 	if (vm_paging_needed())
1886 		pagedaemon_wakeup();
1887 	return (m);
1888 }
1889 
1890 #define	VPSC_ANY	0	/* No restrictions. */
1891 #define	VPSC_NORESERV	1	/* Skip reservations; implies VPSC_NOSUPER. */
1892 #define	VPSC_NOSUPER	2	/* Skip superpages. */
1893 
1894 /*
1895  *	vm_page_scan_contig:
1896  *
1897  *	Scan vm_page_array[] between the specified entries "m_start" and
1898  *	"m_end" for a run of contiguous physical pages that satisfy the
1899  *	specified conditions, and return the lowest page in the run.  The
1900  *	specified "alignment" determines the alignment of the lowest physical
1901  *	page in the run.  If the specified "boundary" is non-zero, then the
1902  *	run of physical pages cannot span a physical address that is a
1903  *	multiple of "boundary".
1904  *
1905  *	"m_end" is never dereferenced, so it need not point to a vm_page
1906  *	structure within vm_page_array[].
1907  *
1908  *	"npages" must be greater than zero.  "m_start" and "m_end" must not
1909  *	span a hole (or discontiguity) in the physical address space.  Both
1910  *	"alignment" and "boundary" must be a power of two.
1911  */
1912 vm_page_t
1913 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end,
1914     u_long alignment, vm_paddr_t boundary, int options)
1915 {
1916 	struct mtx *m_mtx, *new_mtx;
1917 	vm_object_t object;
1918 	vm_paddr_t pa;
1919 	vm_page_t m, m_run;
1920 #if VM_NRESERVLEVEL > 0
1921 	int level;
1922 #endif
1923 	int m_inc, order, run_ext, run_len;
1924 
1925 	KASSERT(npages > 0, ("npages is 0"));
1926 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
1927 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
1928 	m_run = NULL;
1929 	run_len = 0;
1930 	m_mtx = NULL;
1931 	for (m = m_start; m < m_end && run_len < npages; m += m_inc) {
1932 		KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
1933 		    ("page %p is PG_FICTITIOUS or PG_MARKER", m));
1934 
1935 		/*
1936 		 * If the current page would be the start of a run, check its
1937 		 * physical address against the end, alignment, and boundary
1938 		 * conditions.  If it doesn't satisfy these conditions, either
1939 		 * terminate the scan or advance to the next page that
1940 		 * satisfies the failed condition.
1941 		 */
1942 		if (run_len == 0) {
1943 			KASSERT(m_run == NULL, ("m_run != NULL"));
1944 			if (m + npages > m_end)
1945 				break;
1946 			pa = VM_PAGE_TO_PHYS(m);
1947 			if ((pa & (alignment - 1)) != 0) {
1948 				m_inc = atop(roundup2(pa, alignment) - pa);
1949 				continue;
1950 			}
1951 			if (rounddown2(pa ^ (pa + ptoa(npages) - 1),
1952 			    boundary) != 0) {
1953 				m_inc = atop(roundup2(pa, boundary) - pa);
1954 				continue;
1955 			}
1956 		} else
1957 			KASSERT(m_run != NULL, ("m_run == NULL"));
1958 
1959 		/*
1960 		 * Avoid releasing and reacquiring the same page lock.
1961 		 */
1962 		new_mtx = vm_page_lockptr(m);
1963 		if (m_mtx != new_mtx) {
1964 			if (m_mtx != NULL)
1965 				mtx_unlock(m_mtx);
1966 			m_mtx = new_mtx;
1967 			mtx_lock(m_mtx);
1968 		}
1969 		m_inc = 1;
1970 retry:
1971 		if (m->wire_count != 0 || m->hold_count != 0)
1972 			run_ext = 0;
1973 #if VM_NRESERVLEVEL > 0
1974 		else if ((level = vm_reserv_level(m)) >= 0 &&
1975 		    (options & VPSC_NORESERV) != 0) {
1976 			run_ext = 0;
1977 			/* Advance to the end of the reservation. */
1978 			pa = VM_PAGE_TO_PHYS(m);
1979 			m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) -
1980 			    pa);
1981 		}
1982 #endif
1983 		else if ((object = m->object) != NULL) {
1984 			/*
1985 			 * The page is considered eligible for relocation if
1986 			 * and only if it could be laundered or reclaimed by
1987 			 * the page daemon.
1988 			 */
1989 			if (!VM_OBJECT_TRYRLOCK(object)) {
1990 				mtx_unlock(m_mtx);
1991 				VM_OBJECT_RLOCK(object);
1992 				mtx_lock(m_mtx);
1993 				if (m->object != object) {
1994 					/*
1995 					 * The page may have been freed.
1996 					 */
1997 					VM_OBJECT_RUNLOCK(object);
1998 					goto retry;
1999 				} else if (m->wire_count != 0 ||
2000 				    m->hold_count != 0) {
2001 					run_ext = 0;
2002 					goto unlock;
2003 				}
2004 			}
2005 			KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2006 			    ("page %p is PG_UNHOLDFREE", m));
2007 			/* Don't care: PG_NODUMP, PG_ZERO. */
2008 			if (object->type != OBJT_DEFAULT &&
2009 			    object->type != OBJT_SWAP &&
2010 			    object->type != OBJT_VNODE) {
2011 				run_ext = 0;
2012 #if VM_NRESERVLEVEL > 0
2013 			} else if ((options & VPSC_NOSUPER) != 0 &&
2014 			    (level = vm_reserv_level_iffullpop(m)) >= 0) {
2015 				run_ext = 0;
2016 				/* Advance to the end of the superpage. */
2017 				pa = VM_PAGE_TO_PHYS(m);
2018 				m_inc = atop(roundup2(pa + 1,
2019 				    vm_reserv_size(level)) - pa);
2020 #endif
2021 			} else if (object->memattr == VM_MEMATTR_DEFAULT &&
2022 			    m->queue != PQ_NONE && !vm_page_busied(m)) {
2023 				/*
2024 				 * The page is allocated but eligible for
2025 				 * relocation.  Extend the current run by one
2026 				 * page.
2027 				 */
2028 				KASSERT(pmap_page_get_memattr(m) ==
2029 				    VM_MEMATTR_DEFAULT,
2030 				    ("page %p has an unexpected memattr", m));
2031 				KASSERT((m->oflags & (VPO_SWAPINPROG |
2032 				    VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2033 				    ("page %p has unexpected oflags", m));
2034 				/* Don't care: VPO_NOSYNC. */
2035 				run_ext = 1;
2036 			} else
2037 				run_ext = 0;
2038 unlock:
2039 			VM_OBJECT_RUNLOCK(object);
2040 #if VM_NRESERVLEVEL > 0
2041 		} else if (level >= 0) {
2042 			/*
2043 			 * The page is reserved but not yet allocated.  In
2044 			 * other words, it is still cached or free.  Extend
2045 			 * the current run by one page.
2046 			 */
2047 			run_ext = 1;
2048 #endif
2049 		} else if ((order = m->order) < VM_NFREEORDER) {
2050 			/*
2051 			 * The page is enqueued in the physical memory
2052 			 * allocator's cache/free page queues.  Moreover, it
2053 			 * is the first page in a power-of-two-sized run of
2054 			 * contiguous cache/free pages.  Add these pages to
2055 			 * the end of the current run, and jump ahead.
2056 			 */
2057 			run_ext = 1 << order;
2058 			m_inc = 1 << order;
2059 		} else {
2060 			/*
2061 			 * Skip the page for one of the following reasons: (1)
2062 			 * It is enqueued in the physical memory allocator's
2063 			 * cache/free page queues.  However, it is not the
2064 			 * first page in a run of contiguous cache/free pages.
2065 			 * (This case rarely occurs because the scan is
2066 			 * performed in ascending order.) (2) It is not
2067 			 * reserved, and it is transitioning from free to
2068 			 * allocated.  (Conversely, the transition from
2069 			 * allocated to free for managed pages is blocked by
2070 			 * the page lock.) (3) It is allocated but not
2071 			 * contained by an object and not wired, e.g.,
2072 			 * allocated by Xen's balloon driver.
2073 			 */
2074 			run_ext = 0;
2075 		}
2076 
2077 		/*
2078 		 * Extend or reset the current run of pages.
2079 		 */
2080 		if (run_ext > 0) {
2081 			if (run_len == 0)
2082 				m_run = m;
2083 			run_len += run_ext;
2084 		} else {
2085 			if (run_len > 0) {
2086 				m_run = NULL;
2087 				run_len = 0;
2088 			}
2089 		}
2090 	}
2091 	if (m_mtx != NULL)
2092 		mtx_unlock(m_mtx);
2093 	if (run_len >= npages)
2094 		return (m_run);
2095 	return (NULL);
2096 }
2097 
2098 /*
2099  *	vm_page_reclaim_run:
2100  *
2101  *	Try to relocate each of the allocated virtual pages within the
2102  *	specified run of physical pages to a new physical address.  Free the
2103  *	physical pages underlying the relocated virtual pages.  A virtual page
2104  *	is relocatable if and only if it could be laundered or reclaimed by
2105  *	the page daemon.  Whenever possible, a virtual page is relocated to a
2106  *	physical address above "high".
2107  *
2108  *	Returns 0 if every physical page within the run was already free or
2109  *	just freed by a successful relocation.  Otherwise, returns a non-zero
2110  *	value indicating why the last attempt to relocate a virtual page was
2111  *	unsuccessful.
2112  *
2113  *	"req_class" must be an allocation class.
2114  */
2115 static int
2116 vm_page_reclaim_run(int req_class, u_long npages, vm_page_t m_run,
2117     vm_paddr_t high)
2118 {
2119 	struct mtx *m_mtx, *new_mtx;
2120 	struct spglist free;
2121 	vm_object_t object;
2122 	vm_paddr_t pa;
2123 	vm_page_t m, m_end, m_new;
2124 	int error, order, req;
2125 
2126 	KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class,
2127 	    ("req_class is not an allocation class"));
2128 	SLIST_INIT(&free);
2129 	error = 0;
2130 	m = m_run;
2131 	m_end = m_run + npages;
2132 	m_mtx = NULL;
2133 	for (; error == 0 && m < m_end; m++) {
2134 		KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2135 		    ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2136 
2137 		/*
2138 		 * Avoid releasing and reacquiring the same page lock.
2139 		 */
2140 		new_mtx = vm_page_lockptr(m);
2141 		if (m_mtx != new_mtx) {
2142 			if (m_mtx != NULL)
2143 				mtx_unlock(m_mtx);
2144 			m_mtx = new_mtx;
2145 			mtx_lock(m_mtx);
2146 		}
2147 retry:
2148 		if (m->wire_count != 0 || m->hold_count != 0)
2149 			error = EBUSY;
2150 		else if ((object = m->object) != NULL) {
2151 			/*
2152 			 * The page is relocated if and only if it could be
2153 			 * laundered or reclaimed by the page daemon.
2154 			 */
2155 			if (!VM_OBJECT_TRYWLOCK(object)) {
2156 				mtx_unlock(m_mtx);
2157 				VM_OBJECT_WLOCK(object);
2158 				mtx_lock(m_mtx);
2159 				if (m->object != object) {
2160 					/*
2161 					 * The page may have been freed.
2162 					 */
2163 					VM_OBJECT_WUNLOCK(object);
2164 					goto retry;
2165 				} else if (m->wire_count != 0 ||
2166 				    m->hold_count != 0) {
2167 					error = EBUSY;
2168 					goto unlock;
2169 				}
2170 			}
2171 			KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2172 			    ("page %p is PG_UNHOLDFREE", m));
2173 			/* Don't care: PG_NODUMP, PG_ZERO. */
2174 			if (object->type != OBJT_DEFAULT &&
2175 			    object->type != OBJT_SWAP &&
2176 			    object->type != OBJT_VNODE)
2177 				error = EINVAL;
2178 			else if (object->memattr != VM_MEMATTR_DEFAULT)
2179 				error = EINVAL;
2180 			else if (m->queue != PQ_NONE && !vm_page_busied(m)) {
2181 				KASSERT(pmap_page_get_memattr(m) ==
2182 				    VM_MEMATTR_DEFAULT,
2183 				    ("page %p has an unexpected memattr", m));
2184 				KASSERT((m->oflags & (VPO_SWAPINPROG |
2185 				    VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2186 				    ("page %p has unexpected oflags", m));
2187 				/* Don't care: VPO_NOSYNC. */
2188 				if (m->valid != 0) {
2189 					/*
2190 					 * First, try to allocate a new page
2191 					 * that is above "high".  Failing
2192 					 * that, try to allocate a new page
2193 					 * that is below "m_run".  Allocate
2194 					 * the new page between the end of
2195 					 * "m_run" and "high" only as a last
2196 					 * resort.
2197 					 */
2198 					req = req_class | VM_ALLOC_NOOBJ;
2199 					if ((m->flags & PG_NODUMP) != 0)
2200 						req |= VM_ALLOC_NODUMP;
2201 					if (trunc_page(high) !=
2202 					    ~(vm_paddr_t)PAGE_MASK) {
2203 						m_new = vm_page_alloc_contig(
2204 						    NULL, 0, req, 1,
2205 						    round_page(high),
2206 						    ~(vm_paddr_t)0,
2207 						    PAGE_SIZE, 0,
2208 						    VM_MEMATTR_DEFAULT);
2209 					} else
2210 						m_new = NULL;
2211 					if (m_new == NULL) {
2212 						pa = VM_PAGE_TO_PHYS(m_run);
2213 						m_new = vm_page_alloc_contig(
2214 						    NULL, 0, req, 1,
2215 						    0, pa - 1, PAGE_SIZE, 0,
2216 						    VM_MEMATTR_DEFAULT);
2217 					}
2218 					if (m_new == NULL) {
2219 						pa += ptoa(npages);
2220 						m_new = vm_page_alloc_contig(
2221 						    NULL, 0, req, 1,
2222 						    pa, high, PAGE_SIZE, 0,
2223 						    VM_MEMATTR_DEFAULT);
2224 					}
2225 					if (m_new == NULL) {
2226 						error = ENOMEM;
2227 						goto unlock;
2228 					}
2229 					KASSERT(m_new->wire_count == 0,
2230 					    ("page %p is wired", m));
2231 
2232 					/*
2233 					 * Replace "m" with the new page.  For
2234 					 * vm_page_replace(), "m" must be busy
2235 					 * and dequeued.  Finally, change "m"
2236 					 * as if vm_page_free() was called.
2237 					 */
2238 					if (object->ref_count != 0)
2239 						pmap_remove_all(m);
2240 					m_new->aflags = m->aflags;
2241 					KASSERT(m_new->oflags == VPO_UNMANAGED,
2242 					    ("page %p is managed", m));
2243 					m_new->oflags = m->oflags & VPO_NOSYNC;
2244 					pmap_copy_page(m, m_new);
2245 					m_new->valid = m->valid;
2246 					m_new->dirty = m->dirty;
2247 					m->flags &= ~PG_ZERO;
2248 					vm_page_xbusy(m);
2249 					vm_page_remque(m);
2250 					vm_page_replace_checked(m_new, object,
2251 					    m->pindex, m);
2252 					m->valid = 0;
2253 					vm_page_undirty(m);
2254 
2255 					/*
2256 					 * The new page must be deactivated
2257 					 * before the object is unlocked.
2258 					 */
2259 					new_mtx = vm_page_lockptr(m_new);
2260 					if (m_mtx != new_mtx) {
2261 						mtx_unlock(m_mtx);
2262 						m_mtx = new_mtx;
2263 						mtx_lock(m_mtx);
2264 					}
2265 					vm_page_deactivate(m_new);
2266 				} else {
2267 					m->flags &= ~PG_ZERO;
2268 					vm_page_remque(m);
2269 					vm_page_remove(m);
2270 					KASSERT(m->dirty == 0,
2271 					    ("page %p is dirty", m));
2272 				}
2273 				SLIST_INSERT_HEAD(&free, m, plinks.s.ss);
2274 			} else
2275 				error = EBUSY;
2276 unlock:
2277 			VM_OBJECT_WUNLOCK(object);
2278 		} else {
2279 			mtx_lock(&vm_page_queue_free_mtx);
2280 			order = m->order;
2281 			if (order < VM_NFREEORDER) {
2282 				/*
2283 				 * The page is enqueued in the physical memory
2284 				 * allocator's cache/free page queues.
2285 				 * Moreover, it is the first page in a power-
2286 				 * of-two-sized run of contiguous cache/free
2287 				 * pages.  Jump ahead to the last page within
2288 				 * that run, and continue from there.
2289 				 */
2290 				m += (1 << order) - 1;
2291 			}
2292 #if VM_NRESERVLEVEL > 0
2293 			else if (vm_reserv_is_page_free(m))
2294 				order = 0;
2295 #endif
2296 			mtx_unlock(&vm_page_queue_free_mtx);
2297 			if (order == VM_NFREEORDER)
2298 				error = EINVAL;
2299 		}
2300 	}
2301 	if (m_mtx != NULL)
2302 		mtx_unlock(m_mtx);
2303 	if ((m = SLIST_FIRST(&free)) != NULL) {
2304 		mtx_lock(&vm_page_queue_free_mtx);
2305 		do {
2306 			SLIST_REMOVE_HEAD(&free, plinks.s.ss);
2307 			vm_phys_freecnt_adj(m, 1);
2308 #if VM_NRESERVLEVEL > 0
2309 			if (!vm_reserv_free_page(m))
2310 #else
2311 			if (true)
2312 #endif
2313 				vm_phys_free_pages(m, 0);
2314 		} while ((m = SLIST_FIRST(&free)) != NULL);
2315 		vm_page_free_wakeup();
2316 		mtx_unlock(&vm_page_queue_free_mtx);
2317 	}
2318 	return (error);
2319 }
2320 
2321 #define	NRUNS	16
2322 
2323 CTASSERT(powerof2(NRUNS));
2324 
2325 #define	RUN_INDEX(count)	((count) & (NRUNS - 1))
2326 
2327 #define	MIN_RECLAIM	8
2328 
2329 /*
2330  *	vm_page_reclaim_contig:
2331  *
2332  *	Reclaim allocated, contiguous physical memory satisfying the specified
2333  *	conditions by relocating the virtual pages using that physical memory.
2334  *	Returns true if reclamation is successful and false otherwise.  Since
2335  *	relocation requires the allocation of physical pages, reclamation may
2336  *	fail due to a shortage of cache/free pages.  When reclamation fails,
2337  *	callers are expected to perform VM_WAIT before retrying a failed
2338  *	allocation operation, e.g., vm_page_alloc_contig().
2339  *
2340  *	The caller must always specify an allocation class through "req".
2341  *
2342  *	allocation classes:
2343  *	VM_ALLOC_NORMAL		normal process request
2344  *	VM_ALLOC_SYSTEM		system *really* needs a page
2345  *	VM_ALLOC_INTERRUPT	interrupt time request
2346  *
2347  *	The optional allocation flags are ignored.
2348  *
2349  *	"npages" must be greater than zero.  Both "alignment" and "boundary"
2350  *	must be a power of two.
2351  */
2352 bool
2353 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high,
2354     u_long alignment, vm_paddr_t boundary)
2355 {
2356 	vm_paddr_t curr_low;
2357 	vm_page_t m_run, m_runs[NRUNS];
2358 	u_long count, reclaimed;
2359 	int error, i, options, req_class;
2360 
2361 	KASSERT(npages > 0, ("npages is 0"));
2362 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2363 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2364 	req_class = req & VM_ALLOC_CLASS_MASK;
2365 
2366 	/*
2367 	 * The page daemon is allowed to dig deeper into the free page list.
2368 	 */
2369 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2370 		req_class = VM_ALLOC_SYSTEM;
2371 
2372 	/*
2373 	 * Return if the number of cached and free pages cannot satisfy the
2374 	 * requested allocation.
2375 	 */
2376 	count = vm_cnt.v_free_count;
2377 	if (count < npages + vm_cnt.v_free_reserved || (count < npages +
2378 	    vm_cnt.v_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) ||
2379 	    (count < npages && req_class == VM_ALLOC_INTERRUPT))
2380 		return (false);
2381 
2382 	/*
2383 	 * Scan up to three times, relaxing the restrictions ("options") on
2384 	 * the reclamation of reservations and superpages each time.
2385 	 */
2386 	for (options = VPSC_NORESERV;;) {
2387 		/*
2388 		 * Find the highest runs that satisfy the given constraints
2389 		 * and restrictions, and record them in "m_runs".
2390 		 */
2391 		curr_low = low;
2392 		count = 0;
2393 		for (;;) {
2394 			m_run = vm_phys_scan_contig(npages, curr_low, high,
2395 			    alignment, boundary, options);
2396 			if (m_run == NULL)
2397 				break;
2398 			curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages);
2399 			m_runs[RUN_INDEX(count)] = m_run;
2400 			count++;
2401 		}
2402 
2403 		/*
2404 		 * Reclaim the highest runs in LIFO (descending) order until
2405 		 * the number of reclaimed pages, "reclaimed", is at least
2406 		 * MIN_RECLAIM.  Reset "reclaimed" each time because each
2407 		 * reclamation is idempotent, and runs will (likely) recur
2408 		 * from one scan to the next as restrictions are relaxed.
2409 		 */
2410 		reclaimed = 0;
2411 		for (i = 0; count > 0 && i < NRUNS; i++) {
2412 			count--;
2413 			m_run = m_runs[RUN_INDEX(count)];
2414 			error = vm_page_reclaim_run(req_class, npages, m_run,
2415 			    high);
2416 			if (error == 0) {
2417 				reclaimed += npages;
2418 				if (reclaimed >= MIN_RECLAIM)
2419 					return (true);
2420 			}
2421 		}
2422 
2423 		/*
2424 		 * Either relax the restrictions on the next scan or return if
2425 		 * the last scan had no restrictions.
2426 		 */
2427 		if (options == VPSC_NORESERV)
2428 			options = VPSC_NOSUPER;
2429 		else if (options == VPSC_NOSUPER)
2430 			options = VPSC_ANY;
2431 		else if (options == VPSC_ANY)
2432 			return (reclaimed != 0);
2433 	}
2434 }
2435 
2436 /*
2437  *	vm_wait:	(also see VM_WAIT macro)
2438  *
2439  *	Sleep until free pages are available for allocation.
2440  *	- Called in various places before memory allocations.
2441  */
2442 void
2443 vm_wait(void)
2444 {
2445 
2446 	mtx_lock(&vm_page_queue_free_mtx);
2447 	if (curproc == pageproc) {
2448 		vm_pageout_pages_needed = 1;
2449 		msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx,
2450 		    PDROP | PSWP, "VMWait", 0);
2451 	} else {
2452 		if (__predict_false(pageproc == NULL))
2453 			panic("vm_wait in early boot");
2454 		if (!vm_pageout_wanted) {
2455 			vm_pageout_wanted = true;
2456 			wakeup(&vm_pageout_wanted);
2457 		}
2458 		vm_pages_needed = true;
2459 		msleep(&vm_cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PVM,
2460 		    "vmwait", 0);
2461 	}
2462 }
2463 
2464 /*
2465  *	vm_waitpfault:	(also see VM_WAITPFAULT macro)
2466  *
2467  *	Sleep until free pages are available for allocation.
2468  *	- Called only in vm_fault so that processes page faulting
2469  *	  can be easily tracked.
2470  *	- Sleeps at a lower priority than vm_wait() so that vm_wait()ing
2471  *	  processes will be able to grab memory first.  Do not change
2472  *	  this balance without careful testing first.
2473  */
2474 void
2475 vm_waitpfault(void)
2476 {
2477 
2478 	mtx_lock(&vm_page_queue_free_mtx);
2479 	if (!vm_pageout_wanted) {
2480 		vm_pageout_wanted = true;
2481 		wakeup(&vm_pageout_wanted);
2482 	}
2483 	vm_pages_needed = true;
2484 	msleep(&vm_cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PUSER,
2485 	    "pfault", 0);
2486 }
2487 
2488 struct vm_pagequeue *
2489 vm_page_pagequeue(vm_page_t m)
2490 {
2491 
2492 	if (vm_page_in_laundry(m))
2493 		return (&vm_dom[0].vmd_pagequeues[m->queue]);
2494 	else
2495 		return (&vm_phys_domain(m)->vmd_pagequeues[m->queue]);
2496 }
2497 
2498 /*
2499  *	vm_page_dequeue:
2500  *
2501  *	Remove the given page from its current page queue.
2502  *
2503  *	The page must be locked.
2504  */
2505 void
2506 vm_page_dequeue(vm_page_t m)
2507 {
2508 	struct vm_pagequeue *pq;
2509 
2510 	vm_page_assert_locked(m);
2511 	KASSERT(m->queue < PQ_COUNT, ("vm_page_dequeue: page %p is not queued",
2512 	    m));
2513 	pq = vm_page_pagequeue(m);
2514 	vm_pagequeue_lock(pq);
2515 	m->queue = PQ_NONE;
2516 	TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2517 	vm_pagequeue_cnt_dec(pq);
2518 	vm_pagequeue_unlock(pq);
2519 }
2520 
2521 /*
2522  *	vm_page_dequeue_locked:
2523  *
2524  *	Remove the given page from its current page queue.
2525  *
2526  *	The page and page queue must be locked.
2527  */
2528 void
2529 vm_page_dequeue_locked(vm_page_t m)
2530 {
2531 	struct vm_pagequeue *pq;
2532 
2533 	vm_page_lock_assert(m, MA_OWNED);
2534 	pq = vm_page_pagequeue(m);
2535 	vm_pagequeue_assert_locked(pq);
2536 	m->queue = PQ_NONE;
2537 	TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2538 	vm_pagequeue_cnt_dec(pq);
2539 }
2540 
2541 /*
2542  *	vm_page_enqueue:
2543  *
2544  *	Add the given page to the specified page queue.
2545  *
2546  *	The page must be locked.
2547  */
2548 static void
2549 vm_page_enqueue(uint8_t queue, vm_page_t m)
2550 {
2551 	struct vm_pagequeue *pq;
2552 
2553 	vm_page_lock_assert(m, MA_OWNED);
2554 	KASSERT(queue < PQ_COUNT,
2555 	    ("vm_page_enqueue: invalid queue %u request for page %p",
2556 	    queue, m));
2557 	if (queue == PQ_LAUNDRY)
2558 		pq = &vm_dom[0].vmd_pagequeues[queue];
2559 	else
2560 		pq = &vm_phys_domain(m)->vmd_pagequeues[queue];
2561 	vm_pagequeue_lock(pq);
2562 	m->queue = queue;
2563 	TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2564 	vm_pagequeue_cnt_inc(pq);
2565 	vm_pagequeue_unlock(pq);
2566 }
2567 
2568 /*
2569  *	vm_page_requeue:
2570  *
2571  *	Move the given page to the tail of its current page queue.
2572  *
2573  *	The page must be locked.
2574  */
2575 void
2576 vm_page_requeue(vm_page_t m)
2577 {
2578 	struct vm_pagequeue *pq;
2579 
2580 	vm_page_lock_assert(m, MA_OWNED);
2581 	KASSERT(m->queue != PQ_NONE,
2582 	    ("vm_page_requeue: page %p is not queued", m));
2583 	pq = vm_page_pagequeue(m);
2584 	vm_pagequeue_lock(pq);
2585 	TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2586 	TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2587 	vm_pagequeue_unlock(pq);
2588 }
2589 
2590 /*
2591  *	vm_page_requeue_locked:
2592  *
2593  *	Move the given page to the tail of its current page queue.
2594  *
2595  *	The page queue must be locked.
2596  */
2597 void
2598 vm_page_requeue_locked(vm_page_t m)
2599 {
2600 	struct vm_pagequeue *pq;
2601 
2602 	KASSERT(m->queue != PQ_NONE,
2603 	    ("vm_page_requeue_locked: page %p is not queued", m));
2604 	pq = vm_page_pagequeue(m);
2605 	vm_pagequeue_assert_locked(pq);
2606 	TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2607 	TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2608 }
2609 
2610 /*
2611  *	vm_page_activate:
2612  *
2613  *	Put the specified page on the active list (if appropriate).
2614  *	Ensure that act_count is at least ACT_INIT but do not otherwise
2615  *	mess with it.
2616  *
2617  *	The page must be locked.
2618  */
2619 void
2620 vm_page_activate(vm_page_t m)
2621 {
2622 	int queue;
2623 
2624 	vm_page_lock_assert(m, MA_OWNED);
2625 	if ((queue = m->queue) != PQ_ACTIVE) {
2626 		if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
2627 			if (m->act_count < ACT_INIT)
2628 				m->act_count = ACT_INIT;
2629 			if (queue != PQ_NONE)
2630 				vm_page_dequeue(m);
2631 			vm_page_enqueue(PQ_ACTIVE, m);
2632 		} else
2633 			KASSERT(queue == PQ_NONE,
2634 			    ("vm_page_activate: wired page %p is queued", m));
2635 	} else {
2636 		if (m->act_count < ACT_INIT)
2637 			m->act_count = ACT_INIT;
2638 	}
2639 }
2640 
2641 /*
2642  *	vm_page_free_wakeup:
2643  *
2644  *	Helper routine for vm_page_free_toq() and vm_page_cache().  This
2645  *	routine is called when a page has been added to the cache or free
2646  *	queues.
2647  *
2648  *	The page queues must be locked.
2649  */
2650 static inline void
2651 vm_page_free_wakeup(void)
2652 {
2653 
2654 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
2655 	/*
2656 	 * if pageout daemon needs pages, then tell it that there are
2657 	 * some free.
2658 	 */
2659 	if (vm_pageout_pages_needed &&
2660 	    vm_cnt.v_free_count >= vm_cnt.v_pageout_free_min) {
2661 		wakeup(&vm_pageout_pages_needed);
2662 		vm_pageout_pages_needed = 0;
2663 	}
2664 	/*
2665 	 * wakeup processes that are waiting on memory if we hit a
2666 	 * high water mark. And wakeup scheduler process if we have
2667 	 * lots of memory. this process will swapin processes.
2668 	 */
2669 	if (vm_pages_needed && !vm_page_count_min()) {
2670 		vm_pages_needed = false;
2671 		wakeup(&vm_cnt.v_free_count);
2672 	}
2673 }
2674 
2675 /*
2676  *	vm_page_free_toq:
2677  *
2678  *	Returns the given page to the free list,
2679  *	disassociating it with any VM object.
2680  *
2681  *	The object must be locked.  The page must be locked if it is managed.
2682  */
2683 void
2684 vm_page_free_toq(vm_page_t m)
2685 {
2686 
2687 	if ((m->oflags & VPO_UNMANAGED) == 0) {
2688 		vm_page_lock_assert(m, MA_OWNED);
2689 		KASSERT(!pmap_page_is_mapped(m),
2690 		    ("vm_page_free_toq: freeing mapped page %p", m));
2691 	} else
2692 		KASSERT(m->queue == PQ_NONE,
2693 		    ("vm_page_free_toq: unmanaged page %p is queued", m));
2694 	PCPU_INC(cnt.v_tfree);
2695 
2696 	if (vm_page_sbusied(m))
2697 		panic("vm_page_free: freeing busy page %p", m);
2698 
2699 	/*
2700 	 * Unqueue, then remove page.  Note that we cannot destroy
2701 	 * the page here because we do not want to call the pager's
2702 	 * callback routine until after we've put the page on the
2703 	 * appropriate free queue.
2704 	 */
2705 	vm_page_remque(m);
2706 	vm_page_remove(m);
2707 
2708 	/*
2709 	 * If fictitious remove object association and
2710 	 * return, otherwise delay object association removal.
2711 	 */
2712 	if ((m->flags & PG_FICTITIOUS) != 0) {
2713 		return;
2714 	}
2715 
2716 	m->valid = 0;
2717 	vm_page_undirty(m);
2718 
2719 	if (m->wire_count != 0)
2720 		panic("vm_page_free: freeing wired page %p", m);
2721 	if (m->hold_count != 0) {
2722 		m->flags &= ~PG_ZERO;
2723 		KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2724 		    ("vm_page_free: freeing PG_UNHOLDFREE page %p", m));
2725 		m->flags |= PG_UNHOLDFREE;
2726 	} else {
2727 		/*
2728 		 * Restore the default memory attribute to the page.
2729 		 */
2730 		if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
2731 			pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
2732 
2733 		/*
2734 		 * Insert the page into the physical memory allocator's
2735 		 * cache/free page queues.
2736 		 */
2737 		mtx_lock(&vm_page_queue_free_mtx);
2738 		vm_phys_freecnt_adj(m, 1);
2739 #if VM_NRESERVLEVEL > 0
2740 		if (!vm_reserv_free_page(m))
2741 #else
2742 		if (TRUE)
2743 #endif
2744 			vm_phys_free_pages(m, 0);
2745 		vm_page_free_wakeup();
2746 		mtx_unlock(&vm_page_queue_free_mtx);
2747 	}
2748 }
2749 
2750 /*
2751  *	vm_page_wire:
2752  *
2753  *	Mark this page as wired down by yet
2754  *	another map, removing it from paging queues
2755  *	as necessary.
2756  *
2757  *	If the page is fictitious, then its wire count must remain one.
2758  *
2759  *	The page must be locked.
2760  */
2761 void
2762 vm_page_wire(vm_page_t m)
2763 {
2764 
2765 	/*
2766 	 * Only bump the wire statistics if the page is not already wired,
2767 	 * and only unqueue the page if it is on some queue (if it is unmanaged
2768 	 * it is already off the queues).
2769 	 */
2770 	vm_page_lock_assert(m, MA_OWNED);
2771 	if ((m->flags & PG_FICTITIOUS) != 0) {
2772 		KASSERT(m->wire_count == 1,
2773 		    ("vm_page_wire: fictitious page %p's wire count isn't one",
2774 		    m));
2775 		return;
2776 	}
2777 	if (m->wire_count == 0) {
2778 		KASSERT((m->oflags & VPO_UNMANAGED) == 0 ||
2779 		    m->queue == PQ_NONE,
2780 		    ("vm_page_wire: unmanaged page %p is queued", m));
2781 		vm_page_remque(m);
2782 		atomic_add_int(&vm_cnt.v_wire_count, 1);
2783 	}
2784 	m->wire_count++;
2785 	KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m));
2786 }
2787 
2788 /*
2789  * vm_page_unwire:
2790  *
2791  * Release one wiring of the specified page, potentially allowing it to be
2792  * paged out.  Returns TRUE if the number of wirings transitions to zero and
2793  * FALSE otherwise.
2794  *
2795  * Only managed pages belonging to an object can be paged out.  If the number
2796  * of wirings transitions to zero and the page is eligible for page out, then
2797  * the page is added to the specified paging queue (unless PQ_NONE is
2798  * specified).
2799  *
2800  * If a page is fictitious, then its wire count must always be one.
2801  *
2802  * A managed page must be locked.
2803  */
2804 boolean_t
2805 vm_page_unwire(vm_page_t m, uint8_t queue)
2806 {
2807 
2808 	KASSERT(queue < PQ_COUNT || queue == PQ_NONE,
2809 	    ("vm_page_unwire: invalid queue %u request for page %p",
2810 	    queue, m));
2811 	if ((m->oflags & VPO_UNMANAGED) == 0)
2812 		vm_page_assert_locked(m);
2813 	if ((m->flags & PG_FICTITIOUS) != 0) {
2814 		KASSERT(m->wire_count == 1,
2815 	    ("vm_page_unwire: fictitious page %p's wire count isn't one", m));
2816 		return (FALSE);
2817 	}
2818 	if (m->wire_count > 0) {
2819 		m->wire_count--;
2820 		if (m->wire_count == 0) {
2821 			atomic_subtract_int(&vm_cnt.v_wire_count, 1);
2822 			if ((m->oflags & VPO_UNMANAGED) == 0 &&
2823 			    m->object != NULL && queue != PQ_NONE)
2824 				vm_page_enqueue(queue, m);
2825 			return (TRUE);
2826 		} else
2827 			return (FALSE);
2828 	} else
2829 		panic("vm_page_unwire: page %p's wire count is zero", m);
2830 }
2831 
2832 /*
2833  * Move the specified page to the inactive queue.
2834  *
2835  * Many pages placed on the inactive queue should actually go
2836  * into the cache, but it is difficult to figure out which.  What
2837  * we do instead, if the inactive target is well met, is to put
2838  * clean pages at the head of the inactive queue instead of the tail.
2839  * This will cause them to be moved to the cache more quickly and
2840  * if not actively re-referenced, reclaimed more quickly.  If we just
2841  * stick these pages at the end of the inactive queue, heavy filesystem
2842  * meta-data accesses can cause an unnecessary paging load on memory bound
2843  * processes.  This optimization causes one-time-use metadata to be
2844  * reused more quickly.
2845  *
2846  * Normally noreuse is FALSE, resulting in LRU operation.  noreuse is set
2847  * to TRUE if we want this page to be 'as if it were placed in the cache',
2848  * except without unmapping it from the process address space.  In
2849  * practice this is implemented by inserting the page at the head of the
2850  * queue, using a marker page to guide FIFO insertion ordering.
2851  *
2852  * The page must be locked.
2853  */
2854 static inline void
2855 _vm_page_deactivate(vm_page_t m, boolean_t noreuse)
2856 {
2857 	struct vm_pagequeue *pq;
2858 	int queue;
2859 
2860 	vm_page_assert_locked(m);
2861 
2862 	/*
2863 	 * Ignore if the page is already inactive, unless it is unlikely to be
2864 	 * reactivated.
2865 	 */
2866 	if ((queue = m->queue) == PQ_INACTIVE && !noreuse)
2867 		return;
2868 	if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
2869 		pq = &vm_phys_domain(m)->vmd_pagequeues[PQ_INACTIVE];
2870 		/* Avoid multiple acquisitions of the inactive queue lock. */
2871 		if (queue == PQ_INACTIVE) {
2872 			vm_pagequeue_lock(pq);
2873 			vm_page_dequeue_locked(m);
2874 		} else {
2875 			if (queue != PQ_NONE)
2876 				vm_page_dequeue(m);
2877 			vm_pagequeue_lock(pq);
2878 		}
2879 		m->queue = PQ_INACTIVE;
2880 		if (noreuse)
2881 			TAILQ_INSERT_BEFORE(&vm_phys_domain(m)->vmd_inacthead,
2882 			    m, plinks.q);
2883 		else
2884 			TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2885 		vm_pagequeue_cnt_inc(pq);
2886 		vm_pagequeue_unlock(pq);
2887 	}
2888 }
2889 
2890 /*
2891  * Move the specified page to the inactive queue.
2892  *
2893  * The page must be locked.
2894  */
2895 void
2896 vm_page_deactivate(vm_page_t m)
2897 {
2898 
2899 	_vm_page_deactivate(m, FALSE);
2900 }
2901 
2902 /*
2903  * Move the specified page to the inactive queue with the expectation
2904  * that it is unlikely to be reused.
2905  *
2906  * The page must be locked.
2907  */
2908 void
2909 vm_page_deactivate_noreuse(vm_page_t m)
2910 {
2911 
2912 	_vm_page_deactivate(m, TRUE);
2913 }
2914 
2915 /*
2916  * vm_page_launder
2917  *
2918  * 	Put a page in the laundry.
2919  */
2920 void
2921 vm_page_launder(vm_page_t m)
2922 {
2923 	int queue;
2924 
2925 	vm_page_assert_locked(m);
2926 	if ((queue = m->queue) != PQ_LAUNDRY) {
2927 		if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
2928 			if (queue != PQ_NONE)
2929 				vm_page_dequeue(m);
2930 			vm_page_enqueue(PQ_LAUNDRY, m);
2931 		} else
2932 			KASSERT(queue == PQ_NONE,
2933 			    ("wired page %p is queued", m));
2934 	}
2935 }
2936 
2937 /*
2938  * vm_page_try_to_free()
2939  *
2940  *	Attempt to free the page.  If we cannot free it, we do nothing.
2941  *	1 is returned on success, 0 on failure.
2942  */
2943 int
2944 vm_page_try_to_free(vm_page_t m)
2945 {
2946 
2947 	vm_page_lock_assert(m, MA_OWNED);
2948 	if (m->object != NULL)
2949 		VM_OBJECT_ASSERT_WLOCKED(m->object);
2950 	if (m->dirty || m->hold_count || m->wire_count ||
2951 	    (m->oflags & VPO_UNMANAGED) != 0 || vm_page_busied(m))
2952 		return (0);
2953 	pmap_remove_all(m);
2954 	if (m->dirty)
2955 		return (0);
2956 	vm_page_free(m);
2957 	return (1);
2958 }
2959 
2960 /*
2961  * vm_page_advise
2962  *
2963  * 	Deactivate or do nothing, as appropriate.
2964  *
2965  *	The object and page must be locked.
2966  */
2967 void
2968 vm_page_advise(vm_page_t m, int advice)
2969 {
2970 
2971 	vm_page_assert_locked(m);
2972 	VM_OBJECT_ASSERT_WLOCKED(m->object);
2973 	if (advice == MADV_FREE)
2974 		/*
2975 		 * Mark the page clean.  This will allow the page to be freed
2976 		 * up by the system.  However, such pages are often reused
2977 		 * quickly by malloc() so we do not do anything that would
2978 		 * cause a page fault if we can help it.
2979 		 *
2980 		 * Specifically, we do not try to actually free the page now
2981 		 * nor do we try to put it in the cache (which would cause a
2982 		 * page fault on reuse).
2983 		 *
2984 		 * But we do make the page as freeable as we can without
2985 		 * actually taking the step of unmapping it.
2986 		 */
2987 		vm_page_undirty(m);
2988 	else if (advice != MADV_DONTNEED)
2989 		return;
2990 
2991 	/*
2992 	 * Clear any references to the page.  Otherwise, the page daemon will
2993 	 * immediately reactivate the page.
2994 	 */
2995 	vm_page_aflag_clear(m, PGA_REFERENCED);
2996 
2997 	if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
2998 		vm_page_dirty(m);
2999 
3000 	/*
3001 	 * Place clean pages near the head of the inactive queue rather than
3002 	 * the tail, thus defeating the queue's LRU operation and ensuring that
3003 	 * the page will be reused quickly.  Dirty pages not already in the
3004 	 * laundry are moved there.
3005 	 */
3006 	if (m->dirty == 0)
3007 		vm_page_deactivate_noreuse(m);
3008 	else
3009 		vm_page_launder(m);
3010 }
3011 
3012 /*
3013  * Grab a page, waiting until we are waken up due to the page
3014  * changing state.  We keep on waiting, if the page continues
3015  * to be in the object.  If the page doesn't exist, first allocate it
3016  * and then conditionally zero it.
3017  *
3018  * This routine may sleep.
3019  *
3020  * The object must be locked on entry.  The lock will, however, be released
3021  * and reacquired if the routine sleeps.
3022  */
3023 vm_page_t
3024 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
3025 {
3026 	vm_page_t m;
3027 	int sleep;
3028 
3029 	VM_OBJECT_ASSERT_WLOCKED(object);
3030 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
3031 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
3032 	    ("vm_page_grab: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
3033 retrylookup:
3034 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
3035 		sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ?
3036 		    vm_page_xbusied(m) : vm_page_busied(m);
3037 		if (sleep) {
3038 			if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3039 				return (NULL);
3040 			/*
3041 			 * Reference the page before unlocking and
3042 			 * sleeping so that the page daemon is less
3043 			 * likely to reclaim it.
3044 			 */
3045 			vm_page_aflag_set(m, PGA_REFERENCED);
3046 			vm_page_lock(m);
3047 			VM_OBJECT_WUNLOCK(object);
3048 			vm_page_busy_sleep(m, "pgrbwt", (allocflags &
3049 			    VM_ALLOC_IGN_SBUSY) != 0);
3050 			VM_OBJECT_WLOCK(object);
3051 			goto retrylookup;
3052 		} else {
3053 			if ((allocflags & VM_ALLOC_WIRED) != 0) {
3054 				vm_page_lock(m);
3055 				vm_page_wire(m);
3056 				vm_page_unlock(m);
3057 			}
3058 			if ((allocflags &
3059 			    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
3060 				vm_page_xbusy(m);
3061 			if ((allocflags & VM_ALLOC_SBUSY) != 0)
3062 				vm_page_sbusy(m);
3063 			return (m);
3064 		}
3065 	}
3066 	m = vm_page_alloc(object, pindex, allocflags);
3067 	if (m == NULL) {
3068 		if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3069 			return (NULL);
3070 		VM_OBJECT_WUNLOCK(object);
3071 		VM_WAIT;
3072 		VM_OBJECT_WLOCK(object);
3073 		goto retrylookup;
3074 	}
3075 	if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
3076 		pmap_zero_page(m);
3077 	return (m);
3078 }
3079 
3080 /*
3081  * Mapping function for valid or dirty bits in a page.
3082  *
3083  * Inputs are required to range within a page.
3084  */
3085 vm_page_bits_t
3086 vm_page_bits(int base, int size)
3087 {
3088 	int first_bit;
3089 	int last_bit;
3090 
3091 	KASSERT(
3092 	    base + size <= PAGE_SIZE,
3093 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
3094 	);
3095 
3096 	if (size == 0)		/* handle degenerate case */
3097 		return (0);
3098 
3099 	first_bit = base >> DEV_BSHIFT;
3100 	last_bit = (base + size - 1) >> DEV_BSHIFT;
3101 
3102 	return (((vm_page_bits_t)2 << last_bit) -
3103 	    ((vm_page_bits_t)1 << first_bit));
3104 }
3105 
3106 /*
3107  *	vm_page_set_valid_range:
3108  *
3109  *	Sets portions of a page valid.  The arguments are expected
3110  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
3111  *	of any partial chunks touched by the range.  The invalid portion of
3112  *	such chunks will be zeroed.
3113  *
3114  *	(base + size) must be less then or equal to PAGE_SIZE.
3115  */
3116 void
3117 vm_page_set_valid_range(vm_page_t m, int base, int size)
3118 {
3119 	int endoff, frag;
3120 
3121 	VM_OBJECT_ASSERT_WLOCKED(m->object);
3122 	if (size == 0)	/* handle degenerate case */
3123 		return;
3124 
3125 	/*
3126 	 * If the base is not DEV_BSIZE aligned and the valid
3127 	 * bit is clear, we have to zero out a portion of the
3128 	 * first block.
3129 	 */
3130 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
3131 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
3132 		pmap_zero_page_area(m, frag, base - frag);
3133 
3134 	/*
3135 	 * If the ending offset is not DEV_BSIZE aligned and the
3136 	 * valid bit is clear, we have to zero out a portion of
3137 	 * the last block.
3138 	 */
3139 	endoff = base + size;
3140 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
3141 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
3142 		pmap_zero_page_area(m, endoff,
3143 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
3144 
3145 	/*
3146 	 * Assert that no previously invalid block that is now being validated
3147 	 * is already dirty.
3148 	 */
3149 	KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
3150 	    ("vm_page_set_valid_range: page %p is dirty", m));
3151 
3152 	/*
3153 	 * Set valid bits inclusive of any overlap.
3154 	 */
3155 	m->valid |= vm_page_bits(base, size);
3156 }
3157 
3158 /*
3159  * Clear the given bits from the specified page's dirty field.
3160  */
3161 static __inline void
3162 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
3163 {
3164 	uintptr_t addr;
3165 #if PAGE_SIZE < 16384
3166 	int shift;
3167 #endif
3168 
3169 	/*
3170 	 * If the object is locked and the page is neither exclusive busy nor
3171 	 * write mapped, then the page's dirty field cannot possibly be
3172 	 * set by a concurrent pmap operation.
3173 	 */
3174 	VM_OBJECT_ASSERT_WLOCKED(m->object);
3175 	if (!vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
3176 		m->dirty &= ~pagebits;
3177 	else {
3178 		/*
3179 		 * The pmap layer can call vm_page_dirty() without
3180 		 * holding a distinguished lock.  The combination of
3181 		 * the object's lock and an atomic operation suffice
3182 		 * to guarantee consistency of the page dirty field.
3183 		 *
3184 		 * For PAGE_SIZE == 32768 case, compiler already
3185 		 * properly aligns the dirty field, so no forcible
3186 		 * alignment is needed. Only require existence of
3187 		 * atomic_clear_64 when page size is 32768.
3188 		 */
3189 		addr = (uintptr_t)&m->dirty;
3190 #if PAGE_SIZE == 32768
3191 		atomic_clear_64((uint64_t *)addr, pagebits);
3192 #elif PAGE_SIZE == 16384
3193 		atomic_clear_32((uint32_t *)addr, pagebits);
3194 #else		/* PAGE_SIZE <= 8192 */
3195 		/*
3196 		 * Use a trick to perform a 32-bit atomic on the
3197 		 * containing aligned word, to not depend on the existence
3198 		 * of atomic_clear_{8, 16}.
3199 		 */
3200 		shift = addr & (sizeof(uint32_t) - 1);
3201 #if BYTE_ORDER == BIG_ENDIAN
3202 		shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY;
3203 #else
3204 		shift *= NBBY;
3205 #endif
3206 		addr &= ~(sizeof(uint32_t) - 1);
3207 		atomic_clear_32((uint32_t *)addr, pagebits << shift);
3208 #endif		/* PAGE_SIZE */
3209 	}
3210 }
3211 
3212 /*
3213  *	vm_page_set_validclean:
3214  *
3215  *	Sets portions of a page valid and clean.  The arguments are expected
3216  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
3217  *	of any partial chunks touched by the range.  The invalid portion of
3218  *	such chunks will be zero'd.
3219  *
3220  *	(base + size) must be less then or equal to PAGE_SIZE.
3221  */
3222 void
3223 vm_page_set_validclean(vm_page_t m, int base, int size)
3224 {
3225 	vm_page_bits_t oldvalid, pagebits;
3226 	int endoff, frag;
3227 
3228 	VM_OBJECT_ASSERT_WLOCKED(m->object);
3229 	if (size == 0)	/* handle degenerate case */
3230 		return;
3231 
3232 	/*
3233 	 * If the base is not DEV_BSIZE aligned and the valid
3234 	 * bit is clear, we have to zero out a portion of the
3235 	 * first block.
3236 	 */
3237 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
3238 	    (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
3239 		pmap_zero_page_area(m, frag, base - frag);
3240 
3241 	/*
3242 	 * If the ending offset is not DEV_BSIZE aligned and the
3243 	 * valid bit is clear, we have to zero out a portion of
3244 	 * the last block.
3245 	 */
3246 	endoff = base + size;
3247 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
3248 	    (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
3249 		pmap_zero_page_area(m, endoff,
3250 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
3251 
3252 	/*
3253 	 * Set valid, clear dirty bits.  If validating the entire
3254 	 * page we can safely clear the pmap modify bit.  We also
3255 	 * use this opportunity to clear the VPO_NOSYNC flag.  If a process
3256 	 * takes a write fault on a MAP_NOSYNC memory area the flag will
3257 	 * be set again.
3258 	 *
3259 	 * We set valid bits inclusive of any overlap, but we can only
3260 	 * clear dirty bits for DEV_BSIZE chunks that are fully within
3261 	 * the range.
3262 	 */
3263 	oldvalid = m->valid;
3264 	pagebits = vm_page_bits(base, size);
3265 	m->valid |= pagebits;
3266 #if 0	/* NOT YET */
3267 	if ((frag = base & (DEV_BSIZE - 1)) != 0) {
3268 		frag = DEV_BSIZE - frag;
3269 		base += frag;
3270 		size -= frag;
3271 		if (size < 0)
3272 			size = 0;
3273 	}
3274 	pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
3275 #endif
3276 	if (base == 0 && size == PAGE_SIZE) {
3277 		/*
3278 		 * The page can only be modified within the pmap if it is
3279 		 * mapped, and it can only be mapped if it was previously
3280 		 * fully valid.
3281 		 */
3282 		if (oldvalid == VM_PAGE_BITS_ALL)
3283 			/*
3284 			 * Perform the pmap_clear_modify() first.  Otherwise,
3285 			 * a concurrent pmap operation, such as
3286 			 * pmap_protect(), could clear a modification in the
3287 			 * pmap and set the dirty field on the page before
3288 			 * pmap_clear_modify() had begun and after the dirty
3289 			 * field was cleared here.
3290 			 */
3291 			pmap_clear_modify(m);
3292 		m->dirty = 0;
3293 		m->oflags &= ~VPO_NOSYNC;
3294 	} else if (oldvalid != VM_PAGE_BITS_ALL)
3295 		m->dirty &= ~pagebits;
3296 	else
3297 		vm_page_clear_dirty_mask(m, pagebits);
3298 }
3299 
3300 void
3301 vm_page_clear_dirty(vm_page_t m, int base, int size)
3302 {
3303 
3304 	vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
3305 }
3306 
3307 /*
3308  *	vm_page_set_invalid:
3309  *
3310  *	Invalidates DEV_BSIZE'd chunks within a page.  Both the
3311  *	valid and dirty bits for the effected areas are cleared.
3312  */
3313 void
3314 vm_page_set_invalid(vm_page_t m, int base, int size)
3315 {
3316 	vm_page_bits_t bits;
3317 	vm_object_t object;
3318 
3319 	object = m->object;
3320 	VM_OBJECT_ASSERT_WLOCKED(object);
3321 	if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
3322 	    size >= object->un_pager.vnp.vnp_size)
3323 		bits = VM_PAGE_BITS_ALL;
3324 	else
3325 		bits = vm_page_bits(base, size);
3326 	if (object->ref_count != 0 && m->valid == VM_PAGE_BITS_ALL &&
3327 	    bits != 0)
3328 		pmap_remove_all(m);
3329 	KASSERT((bits == 0 && m->valid == VM_PAGE_BITS_ALL) ||
3330 	    !pmap_page_is_mapped(m),
3331 	    ("vm_page_set_invalid: page %p is mapped", m));
3332 	m->valid &= ~bits;
3333 	m->dirty &= ~bits;
3334 }
3335 
3336 /*
3337  * vm_page_zero_invalid()
3338  *
3339  *	The kernel assumes that the invalid portions of a page contain
3340  *	garbage, but such pages can be mapped into memory by user code.
3341  *	When this occurs, we must zero out the non-valid portions of the
3342  *	page so user code sees what it expects.
3343  *
3344  *	Pages are most often semi-valid when the end of a file is mapped
3345  *	into memory and the file's size is not page aligned.
3346  */
3347 void
3348 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
3349 {
3350 	int b;
3351 	int i;
3352 
3353 	VM_OBJECT_ASSERT_WLOCKED(m->object);
3354 	/*
3355 	 * Scan the valid bits looking for invalid sections that
3356 	 * must be zeroed.  Invalid sub-DEV_BSIZE'd areas ( where the
3357 	 * valid bit may be set ) have already been zeroed by
3358 	 * vm_page_set_validclean().
3359 	 */
3360 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
3361 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
3362 		    (m->valid & ((vm_page_bits_t)1 << i))) {
3363 			if (i > b) {
3364 				pmap_zero_page_area(m,
3365 				    b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
3366 			}
3367 			b = i + 1;
3368 		}
3369 	}
3370 
3371 	/*
3372 	 * setvalid is TRUE when we can safely set the zero'd areas
3373 	 * as being valid.  We can do this if there are no cache consistancy
3374 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
3375 	 */
3376 	if (setvalid)
3377 		m->valid = VM_PAGE_BITS_ALL;
3378 }
3379 
3380 /*
3381  *	vm_page_is_valid:
3382  *
3383  *	Is (partial) page valid?  Note that the case where size == 0
3384  *	will return FALSE in the degenerate case where the page is
3385  *	entirely invalid, and TRUE otherwise.
3386  */
3387 int
3388 vm_page_is_valid(vm_page_t m, int base, int size)
3389 {
3390 	vm_page_bits_t bits;
3391 
3392 	VM_OBJECT_ASSERT_LOCKED(m->object);
3393 	bits = vm_page_bits(base, size);
3394 	return (m->valid != 0 && (m->valid & bits) == bits);
3395 }
3396 
3397 /*
3398  *	vm_page_ps_is_valid:
3399  *
3400  *	Returns TRUE if the entire (super)page is valid and FALSE otherwise.
3401  */
3402 boolean_t
3403 vm_page_ps_is_valid(vm_page_t m)
3404 {
3405 	int i, npages;
3406 
3407 	VM_OBJECT_ASSERT_LOCKED(m->object);
3408 	npages = atop(pagesizes[m->psind]);
3409 
3410 	/*
3411 	 * The physically contiguous pages that make up a superpage, i.e., a
3412 	 * page with a page size index ("psind") greater than zero, will
3413 	 * occupy adjacent entries in vm_page_array[].
3414 	 */
3415 	for (i = 0; i < npages; i++) {
3416 		if (m[i].valid != VM_PAGE_BITS_ALL)
3417 			return (FALSE);
3418 	}
3419 	return (TRUE);
3420 }
3421 
3422 /*
3423  * Set the page's dirty bits if the page is modified.
3424  */
3425 void
3426 vm_page_test_dirty(vm_page_t m)
3427 {
3428 
3429 	VM_OBJECT_ASSERT_WLOCKED(m->object);
3430 	if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
3431 		vm_page_dirty(m);
3432 }
3433 
3434 void
3435 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
3436 {
3437 
3438 	mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
3439 }
3440 
3441 void
3442 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
3443 {
3444 
3445 	mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
3446 }
3447 
3448 int
3449 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
3450 {
3451 
3452 	return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
3453 }
3454 
3455 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
3456 void
3457 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
3458 {
3459 
3460 	vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
3461 }
3462 
3463 void
3464 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
3465 {
3466 
3467 	mtx_assert_(vm_page_lockptr(m), a, file, line);
3468 }
3469 #endif
3470 
3471 #ifdef INVARIANTS
3472 void
3473 vm_page_object_lock_assert(vm_page_t m)
3474 {
3475 
3476 	/*
3477 	 * Certain of the page's fields may only be modified by the
3478 	 * holder of the containing object's lock or the exclusive busy.
3479 	 * holder.  Unfortunately, the holder of the write busy is
3480 	 * not recorded, and thus cannot be checked here.
3481 	 */
3482 	if (m->object != NULL && !vm_page_xbusied(m))
3483 		VM_OBJECT_ASSERT_WLOCKED(m->object);
3484 }
3485 
3486 void
3487 vm_page_assert_pga_writeable(vm_page_t m, uint8_t bits)
3488 {
3489 
3490 	if ((bits & PGA_WRITEABLE) == 0)
3491 		return;
3492 
3493 	/*
3494 	 * The PGA_WRITEABLE flag can only be set if the page is
3495 	 * managed, is exclusively busied or the object is locked.
3496 	 * Currently, this flag is only set by pmap_enter().
3497 	 */
3498 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3499 	    ("PGA_WRITEABLE on unmanaged page"));
3500 	if (!vm_page_xbusied(m))
3501 		VM_OBJECT_ASSERT_LOCKED(m->object);
3502 }
3503 #endif
3504 
3505 #include "opt_ddb.h"
3506 #ifdef DDB
3507 #include <sys/kernel.h>
3508 
3509 #include <ddb/ddb.h>
3510 
3511 DB_SHOW_COMMAND(page, vm_page_print_page_info)
3512 {
3513 
3514 	db_printf("vm_cnt.v_free_count: %d\n", vm_cnt.v_free_count);
3515 	db_printf("vm_cnt.v_inactive_count: %d\n", vm_cnt.v_inactive_count);
3516 	db_printf("vm_cnt.v_active_count: %d\n", vm_cnt.v_active_count);
3517 	db_printf("vm_cnt.v_laundry_count: %d\n", vm_cnt.v_laundry_count);
3518 	db_printf("vm_cnt.v_wire_count: %d\n", vm_cnt.v_wire_count);
3519 	db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
3520 	db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
3521 	db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
3522 	db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
3523 }
3524 
3525 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
3526 {
3527 	int dom;
3528 
3529 	db_printf("pq_free %d\n", vm_cnt.v_free_count);
3530 	for (dom = 0; dom < vm_ndomains; dom++) {
3531 		db_printf(
3532 	    "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d\n",
3533 		    dom,
3534 		    vm_dom[dom].vmd_page_count,
3535 		    vm_dom[dom].vmd_free_count,
3536 		    vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
3537 		    vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
3538 		    vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt);
3539 	}
3540 }
3541 
3542 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
3543 {
3544 	vm_page_t m;
3545 	boolean_t phys;
3546 
3547 	if (!have_addr) {
3548 		db_printf("show pginfo addr\n");
3549 		return;
3550 	}
3551 
3552 	phys = strchr(modif, 'p') != NULL;
3553 	if (phys)
3554 		m = PHYS_TO_VM_PAGE(addr);
3555 	else
3556 		m = (vm_page_t)addr;
3557 	db_printf(
3558     "page %p obj %p pidx 0x%jx phys 0x%jx q %d hold %d wire %d\n"
3559     "  af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
3560 	    m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
3561 	    m->queue, m->hold_count, m->wire_count, m->aflags, m->oflags,
3562 	    m->flags, m->act_count, m->busy_lock, m->valid, m->dirty);
3563 }
3564 #endif /* DDB */
3565