xref: /freebsd/sys/vm/vm_page.c (revision 74fe6c29fb7eef3418d7919dcd41dc1a04a982a1)
1 /*-
2  * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU)
3  *
4  * Copyright (c) 1991 Regents of the University of California.
5  * All rights reserved.
6  * Copyright (c) 1998 Matthew Dillon.  All Rights Reserved.
7  *
8  * This code is derived from software contributed to Berkeley by
9  * The Mach Operating System project at Carnegie-Mellon University.
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  * 1. Redistributions of source code must retain the above copyright
15  *    notice, this list of conditions and the following disclaimer.
16  * 2. Redistributions in binary form must reproduce the above copyright
17  *    notice, this list of conditions and the following disclaimer in the
18  *    documentation and/or other materials provided with the distribution.
19  * 3. Neither the name of the University nor the names of its contributors
20  *    may be used to endorse or promote products derived from this software
21  *    without specific prior written permission.
22  *
23  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
24  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
25  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
26  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
27  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
28  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
29  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
30  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
31  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
32  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33  * SUCH DAMAGE.
34  *
35  *	from: @(#)vm_page.c	7.4 (Berkeley) 5/7/91
36  */
37 
38 /*-
39  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
40  * All rights reserved.
41  *
42  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
43  *
44  * Permission to use, copy, modify and distribute this software and
45  * its documentation is hereby granted, provided that both the copyright
46  * notice and this permission notice appear in all copies of the
47  * software, derivative works or modified versions, and any portions
48  * thereof, and that both notices appear in supporting documentation.
49  *
50  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
51  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
52  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
53  *
54  * Carnegie Mellon requests users of this software to return to
55  *
56  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
57  *  School of Computer Science
58  *  Carnegie Mellon University
59  *  Pittsburgh PA 15213-3890
60  *
61  * any improvements or extensions that they make and grant Carnegie the
62  * rights to redistribute these changes.
63  */
64 
65 /*
66  *			GENERAL RULES ON VM_PAGE MANIPULATION
67  *
68  *	- A page queue lock is required when adding or removing a page from a
69  *	  page queue regardless of other locks or the busy state of a page.
70  *
71  *		* In general, no thread besides the page daemon can acquire or
72  *		  hold more than one page queue lock at a time.
73  *
74  *		* The page daemon can acquire and hold any pair of page queue
75  *		  locks in any order.
76  *
77  *	- The object lock is required when inserting or removing
78  *	  pages from an object (vm_page_insert() or vm_page_remove()).
79  *
80  */
81 
82 /*
83  *	Resident memory management module.
84  */
85 
86 #include <sys/cdefs.h>
87 __FBSDID("$FreeBSD$");
88 
89 #include "opt_vm.h"
90 
91 #include <sys/param.h>
92 #include <sys/systm.h>
93 #include <sys/lock.h>
94 #include <sys/domainset.h>
95 #include <sys/kernel.h>
96 #include <sys/limits.h>
97 #include <sys/linker.h>
98 #include <sys/malloc.h>
99 #include <sys/mman.h>
100 #include <sys/msgbuf.h>
101 #include <sys/mutex.h>
102 #include <sys/proc.h>
103 #include <sys/rwlock.h>
104 #include <sys/sbuf.h>
105 #include <sys/smp.h>
106 #include <sys/sysctl.h>
107 #include <sys/vmmeter.h>
108 #include <sys/vnode.h>
109 
110 #include <vm/vm.h>
111 #include <vm/pmap.h>
112 #include <vm/vm_param.h>
113 #include <vm/vm_domainset.h>
114 #include <vm/vm_kern.h>
115 #include <vm/vm_map.h>
116 #include <vm/vm_object.h>
117 #include <vm/vm_page.h>
118 #include <vm/vm_pageout.h>
119 #include <vm/vm_phys.h>
120 #include <vm/vm_pagequeue.h>
121 #include <vm/vm_pager.h>
122 #include <vm/vm_radix.h>
123 #include <vm/vm_reserv.h>
124 #include <vm/vm_extern.h>
125 #include <vm/uma.h>
126 #include <vm/uma_int.h>
127 
128 #include <machine/md_var.h>
129 
130 extern int	uma_startup_count(int);
131 extern void	uma_startup(void *, int);
132 extern int	vmem_startup_count(void);
133 
134 /*
135  *	Associated with page of user-allocatable memory is a
136  *	page structure.
137  */
138 
139 struct vm_domain vm_dom[MAXMEMDOM];
140 
141 struct mtx_padalign __exclusive_cache_line pa_lock[PA_LOCK_COUNT];
142 
143 struct mtx_padalign __exclusive_cache_line vm_domainset_lock;
144 /* The following fields are protected by the domainset lock. */
145 domainset_t __exclusive_cache_line vm_min_domains;
146 domainset_t __exclusive_cache_line vm_severe_domains;
147 static int vm_min_waiters;
148 static int vm_severe_waiters;
149 static int vm_pageproc_waiters;
150 
151 /*
152  * bogus page -- for I/O to/from partially complete buffers,
153  * or for paging into sparsely invalid regions.
154  */
155 vm_page_t bogus_page;
156 
157 vm_page_t vm_page_array;
158 long vm_page_array_size;
159 long first_page;
160 
161 static int boot_pages;
162 SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
163     &boot_pages, 0,
164     "number of pages allocated for bootstrapping the VM system");
165 
166 static int pa_tryrelock_restart;
167 SYSCTL_INT(_vm, OID_AUTO, tryrelock_restart, CTLFLAG_RD,
168     &pa_tryrelock_restart, 0, "Number of tryrelock restarts");
169 
170 static TAILQ_HEAD(, vm_page) blacklist_head;
171 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS);
172 SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD |
173     CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages");
174 
175 static uma_zone_t fakepg_zone;
176 
177 static void vm_page_alloc_check(vm_page_t m);
178 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits);
179 static void vm_page_enqueue(uint8_t queue, vm_page_t m);
180 static void vm_page_free_phys(struct vm_domain *vmd, vm_page_t m);
181 static void vm_page_init(void *dummy);
182 static int vm_page_insert_after(vm_page_t m, vm_object_t object,
183     vm_pindex_t pindex, vm_page_t mpred);
184 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object,
185     vm_page_t mpred);
186 static int vm_page_reclaim_run(int req_class, int domain, u_long npages,
187     vm_page_t m_run, vm_paddr_t high);
188 static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object,
189     int req);
190 
191 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL);
192 
193 static void
194 vm_page_init(void *dummy)
195 {
196 
197 	fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
198 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
199 	bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
200 	    VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
201 }
202 
203 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
204 #if PAGE_SIZE == 32768
205 #ifdef CTASSERT
206 CTASSERT(sizeof(u_long) >= 8);
207 #endif
208 #endif
209 
210 /*
211  * Try to acquire a physical address lock while a pmap is locked.  If we
212  * fail to trylock we unlock and lock the pmap directly and cache the
213  * locked pa in *locked.  The caller should then restart their loop in case
214  * the virtual to physical mapping has changed.
215  */
216 int
217 vm_page_pa_tryrelock(pmap_t pmap, vm_paddr_t pa, vm_paddr_t *locked)
218 {
219 	vm_paddr_t lockpa;
220 
221 	lockpa = *locked;
222 	*locked = pa;
223 	if (lockpa) {
224 		PA_LOCK_ASSERT(lockpa, MA_OWNED);
225 		if (PA_LOCKPTR(pa) == PA_LOCKPTR(lockpa))
226 			return (0);
227 		PA_UNLOCK(lockpa);
228 	}
229 	if (PA_TRYLOCK(pa))
230 		return (0);
231 	PMAP_UNLOCK(pmap);
232 	atomic_add_int(&pa_tryrelock_restart, 1);
233 	PA_LOCK(pa);
234 	PMAP_LOCK(pmap);
235 	return (EAGAIN);
236 }
237 
238 /*
239  *	vm_set_page_size:
240  *
241  *	Sets the page size, perhaps based upon the memory
242  *	size.  Must be called before any use of page-size
243  *	dependent functions.
244  */
245 void
246 vm_set_page_size(void)
247 {
248 	if (vm_cnt.v_page_size == 0)
249 		vm_cnt.v_page_size = PAGE_SIZE;
250 	if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0)
251 		panic("vm_set_page_size: page size not a power of two");
252 }
253 
254 /*
255  *	vm_page_blacklist_next:
256  *
257  *	Find the next entry in the provided string of blacklist
258  *	addresses.  Entries are separated by space, comma, or newline.
259  *	If an invalid integer is encountered then the rest of the
260  *	string is skipped.  Updates the list pointer to the next
261  *	character, or NULL if the string is exhausted or invalid.
262  */
263 static vm_paddr_t
264 vm_page_blacklist_next(char **list, char *end)
265 {
266 	vm_paddr_t bad;
267 	char *cp, *pos;
268 
269 	if (list == NULL || *list == NULL)
270 		return (0);
271 	if (**list =='\0') {
272 		*list = NULL;
273 		return (0);
274 	}
275 
276 	/*
277 	 * If there's no end pointer then the buffer is coming from
278 	 * the kenv and we know it's null-terminated.
279 	 */
280 	if (end == NULL)
281 		end = *list + strlen(*list);
282 
283 	/* Ensure that strtoq() won't walk off the end */
284 	if (*end != '\0') {
285 		if (*end == '\n' || *end == ' ' || *end  == ',')
286 			*end = '\0';
287 		else {
288 			printf("Blacklist not terminated, skipping\n");
289 			*list = NULL;
290 			return (0);
291 		}
292 	}
293 
294 	for (pos = *list; *pos != '\0'; pos = cp) {
295 		bad = strtoq(pos, &cp, 0);
296 		if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') {
297 			if (bad == 0) {
298 				if (++cp < end)
299 					continue;
300 				else
301 					break;
302 			}
303 		} else
304 			break;
305 		if (*cp == '\0' || ++cp >= end)
306 			*list = NULL;
307 		else
308 			*list = cp;
309 		return (trunc_page(bad));
310 	}
311 	printf("Garbage in RAM blacklist, skipping\n");
312 	*list = NULL;
313 	return (0);
314 }
315 
316 /*
317  *	vm_page_blacklist_check:
318  *
319  *	Iterate through the provided string of blacklist addresses, pulling
320  *	each entry out of the physical allocator free list and putting it
321  *	onto a list for reporting via the vm.page_blacklist sysctl.
322  */
323 static void
324 vm_page_blacklist_check(char *list, char *end)
325 {
326 	struct vm_domain *vmd;
327 	vm_paddr_t pa;
328 	vm_page_t m;
329 	char *next;
330 	int ret;
331 
332 	next = list;
333 	while (next != NULL) {
334 		if ((pa = vm_page_blacklist_next(&next, end)) == 0)
335 			continue;
336 		m = vm_phys_paddr_to_vm_page(pa);
337 		if (m == NULL)
338 			continue;
339 		vmd = vm_pagequeue_domain(m);
340 		vm_domain_free_lock(vmd);
341 		ret = vm_phys_unfree_page(m);
342 		vm_domain_free_unlock(vmd);
343 		if (ret == TRUE) {
344 			TAILQ_INSERT_TAIL(&blacklist_head, m, listq);
345 			if (bootverbose)
346 				printf("Skipping page with pa 0x%jx\n",
347 				    (uintmax_t)pa);
348 		}
349 	}
350 }
351 
352 /*
353  *	vm_page_blacklist_load:
354  *
355  *	Search for a special module named "ram_blacklist".  It'll be a
356  *	plain text file provided by the user via the loader directive
357  *	of the same name.
358  */
359 static void
360 vm_page_blacklist_load(char **list, char **end)
361 {
362 	void *mod;
363 	u_char *ptr;
364 	u_int len;
365 
366 	mod = NULL;
367 	ptr = NULL;
368 
369 	mod = preload_search_by_type("ram_blacklist");
370 	if (mod != NULL) {
371 		ptr = preload_fetch_addr(mod);
372 		len = preload_fetch_size(mod);
373         }
374 	*list = ptr;
375 	if (ptr != NULL)
376 		*end = ptr + len;
377 	else
378 		*end = NULL;
379 	return;
380 }
381 
382 static int
383 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS)
384 {
385 	vm_page_t m;
386 	struct sbuf sbuf;
387 	int error, first;
388 
389 	first = 1;
390 	error = sysctl_wire_old_buffer(req, 0);
391 	if (error != 0)
392 		return (error);
393 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
394 	TAILQ_FOREACH(m, &blacklist_head, listq) {
395 		sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",",
396 		    (uintmax_t)m->phys_addr);
397 		first = 0;
398 	}
399 	error = sbuf_finish(&sbuf);
400 	sbuf_delete(&sbuf);
401 	return (error);
402 }
403 
404 static void
405 vm_page_domain_init(int domain)
406 {
407 	struct vm_domain *vmd;
408 	struct vm_pagequeue *pq;
409 	int i;
410 
411 	vmd = VM_DOMAIN(domain);
412 	bzero(vmd, sizeof(*vmd));
413 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) =
414 	    "vm inactive pagequeue";
415 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) =
416 	    "vm active pagequeue";
417 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) =
418 	    "vm laundry pagequeue";
419 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) =
420 	    "vm unswappable pagequeue";
421 	vmd->vmd_domain = domain;
422 	vmd->vmd_page_count = 0;
423 	vmd->vmd_free_count = 0;
424 	vmd->vmd_segs = 0;
425 	vmd->vmd_oom = FALSE;
426 	for (i = 0; i < PQ_COUNT; i++) {
427 		pq = &vmd->vmd_pagequeues[i];
428 		TAILQ_INIT(&pq->pq_pl);
429 		mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue",
430 		    MTX_DEF | MTX_DUPOK);
431 	}
432 	mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF);
433 	mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF);
434 	snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain);
435 }
436 
437 /*
438  * Initialize a physical page in preparation for adding it to the free
439  * lists.
440  */
441 static void
442 vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind)
443 {
444 
445 	m->object = NULL;
446 	m->wire_count = 0;
447 	m->busy_lock = VPB_UNBUSIED;
448 	m->hold_count = 0;
449 	m->flags = 0;
450 	m->phys_addr = pa;
451 	m->queue = PQ_NONE;
452 	m->psind = 0;
453 	m->segind = segind;
454 	m->order = VM_NFREEORDER;
455 	m->pool = VM_FREEPOOL_DEFAULT;
456 	m->valid = m->dirty = 0;
457 	pmap_page_init(m);
458 }
459 
460 /*
461  *	vm_page_startup:
462  *
463  *	Initializes the resident memory module.  Allocates physical memory for
464  *	bootstrapping UMA and some data structures that are used to manage
465  *	physical pages.  Initializes these structures, and populates the free
466  *	page queues.
467  */
468 vm_offset_t
469 vm_page_startup(vm_offset_t vaddr)
470 {
471 	struct vm_phys_seg *seg;
472 	vm_page_t m;
473 	char *list, *listend;
474 	vm_offset_t mapped;
475 	vm_paddr_t end, high_avail, low_avail, new_end, page_range, size;
476 	vm_paddr_t biggestsize, last_pa, pa;
477 	u_long pagecount;
478 	int biggestone, i, segind;
479 
480 	biggestsize = 0;
481 	biggestone = 0;
482 	vaddr = round_page(vaddr);
483 
484 	for (i = 0; phys_avail[i + 1]; i += 2) {
485 		phys_avail[i] = round_page(phys_avail[i]);
486 		phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
487 	}
488 	for (i = 0; phys_avail[i + 1]; i += 2) {
489 		size = phys_avail[i + 1] - phys_avail[i];
490 		if (size > biggestsize) {
491 			biggestone = i;
492 			biggestsize = size;
493 		}
494 	}
495 
496 	end = phys_avail[biggestone+1];
497 
498 	/*
499 	 * Initialize the page and queue locks.
500 	 */
501 	mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF);
502 	for (i = 0; i < PA_LOCK_COUNT; i++)
503 		mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
504 	for (i = 0; i < vm_ndomains; i++)
505 		vm_page_domain_init(i);
506 
507 	/*
508 	 * Allocate memory for use when boot strapping the kernel memory
509 	 * allocator.  Tell UMA how many zones we are going to create
510 	 * before going fully functional.  UMA will add its zones.
511 	 *
512 	 * VM startup zones: vmem, vmem_btag, VM OBJECT, RADIX NODE, MAP,
513 	 * KMAP ENTRY, MAP ENTRY, VMSPACE.
514 	 */
515 	boot_pages = uma_startup_count(8);
516 
517 #ifndef UMA_MD_SMALL_ALLOC
518 	/* vmem_startup() calls uma_prealloc(). */
519 	boot_pages += vmem_startup_count();
520 	/* vm_map_startup() calls uma_prealloc(). */
521 	boot_pages += howmany(MAX_KMAP,
522 	    UMA_SLAB_SPACE / sizeof(struct vm_map));
523 
524 	/*
525 	 * Before going fully functional kmem_init() does allocation
526 	 * from "KMAP ENTRY" and vmem_create() does allocation from "vmem".
527 	 */
528 	boot_pages += 2;
529 #endif
530 	/*
531 	 * CTFLAG_RDTUN doesn't work during the early boot process, so we must
532 	 * manually fetch the value.
533 	 */
534 	TUNABLE_INT_FETCH("vm.boot_pages", &boot_pages);
535 	new_end = end - (boot_pages * UMA_SLAB_SIZE);
536 	new_end = trunc_page(new_end);
537 	mapped = pmap_map(&vaddr, new_end, end,
538 	    VM_PROT_READ | VM_PROT_WRITE);
539 	bzero((void *)mapped, end - new_end);
540 	uma_startup((void *)mapped, boot_pages);
541 
542 #if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \
543     defined(__i386__) || defined(__mips__)
544 	/*
545 	 * Allocate a bitmap to indicate that a random physical page
546 	 * needs to be included in a minidump.
547 	 *
548 	 * The amd64 port needs this to indicate which direct map pages
549 	 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
550 	 *
551 	 * However, i386 still needs this workspace internally within the
552 	 * minidump code.  In theory, they are not needed on i386, but are
553 	 * included should the sf_buf code decide to use them.
554 	 */
555 	last_pa = 0;
556 	for (i = 0; dump_avail[i + 1] != 0; i += 2)
557 		if (dump_avail[i + 1] > last_pa)
558 			last_pa = dump_avail[i + 1];
559 	page_range = last_pa / PAGE_SIZE;
560 	vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
561 	new_end -= vm_page_dump_size;
562 	vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
563 	    new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
564 	bzero((void *)vm_page_dump, vm_page_dump_size);
565 #else
566 	(void)last_pa;
567 #endif
568 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__)
569 	/*
570 	 * Include the UMA bootstrap pages and vm_page_dump in a crash dump.
571 	 * When pmap_map() uses the direct map, they are not automatically
572 	 * included.
573 	 */
574 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
575 		dump_add_page(pa);
576 #endif
577 	phys_avail[biggestone + 1] = new_end;
578 #ifdef __amd64__
579 	/*
580 	 * Request that the physical pages underlying the message buffer be
581 	 * included in a crash dump.  Since the message buffer is accessed
582 	 * through the direct map, they are not automatically included.
583 	 */
584 	pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
585 	last_pa = pa + round_page(msgbufsize);
586 	while (pa < last_pa) {
587 		dump_add_page(pa);
588 		pa += PAGE_SIZE;
589 	}
590 #endif
591 	/*
592 	 * Compute the number of pages of memory that will be available for
593 	 * use, taking into account the overhead of a page structure per page.
594 	 * In other words, solve
595 	 *	"available physical memory" - round_page(page_range *
596 	 *	    sizeof(struct vm_page)) = page_range * PAGE_SIZE
597 	 * for page_range.
598 	 */
599 	low_avail = phys_avail[0];
600 	high_avail = phys_avail[1];
601 	for (i = 0; i < vm_phys_nsegs; i++) {
602 		if (vm_phys_segs[i].start < low_avail)
603 			low_avail = vm_phys_segs[i].start;
604 		if (vm_phys_segs[i].end > high_avail)
605 			high_avail = vm_phys_segs[i].end;
606 	}
607 	/* Skip the first chunk.  It is already accounted for. */
608 	for (i = 2; phys_avail[i + 1] != 0; i += 2) {
609 		if (phys_avail[i] < low_avail)
610 			low_avail = phys_avail[i];
611 		if (phys_avail[i + 1] > high_avail)
612 			high_avail = phys_avail[i + 1];
613 	}
614 	first_page = low_avail / PAGE_SIZE;
615 #ifdef VM_PHYSSEG_SPARSE
616 	size = 0;
617 	for (i = 0; i < vm_phys_nsegs; i++)
618 		size += vm_phys_segs[i].end - vm_phys_segs[i].start;
619 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
620 		size += phys_avail[i + 1] - phys_avail[i];
621 #elif defined(VM_PHYSSEG_DENSE)
622 	size = high_avail - low_avail;
623 #else
624 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
625 #endif
626 
627 #ifdef VM_PHYSSEG_DENSE
628 	/*
629 	 * In the VM_PHYSSEG_DENSE case, the number of pages can account for
630 	 * the overhead of a page structure per page only if vm_page_array is
631 	 * allocated from the last physical memory chunk.  Otherwise, we must
632 	 * allocate page structures representing the physical memory
633 	 * underlying vm_page_array, even though they will not be used.
634 	 */
635 	if (new_end != high_avail)
636 		page_range = size / PAGE_SIZE;
637 	else
638 #endif
639 	{
640 		page_range = size / (PAGE_SIZE + sizeof(struct vm_page));
641 
642 		/*
643 		 * If the partial bytes remaining are large enough for
644 		 * a page (PAGE_SIZE) without a corresponding
645 		 * 'struct vm_page', then new_end will contain an
646 		 * extra page after subtracting the length of the VM
647 		 * page array.  Compensate by subtracting an extra
648 		 * page from new_end.
649 		 */
650 		if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) {
651 			if (new_end == high_avail)
652 				high_avail -= PAGE_SIZE;
653 			new_end -= PAGE_SIZE;
654 		}
655 	}
656 	end = new_end;
657 
658 	/*
659 	 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
660 	 * However, because this page is allocated from KVM, out-of-bounds
661 	 * accesses using the direct map will not be trapped.
662 	 */
663 	vaddr += PAGE_SIZE;
664 
665 	/*
666 	 * Allocate physical memory for the page structures, and map it.
667 	 */
668 	new_end = trunc_page(end - page_range * sizeof(struct vm_page));
669 	mapped = pmap_map(&vaddr, new_end, end,
670 	    VM_PROT_READ | VM_PROT_WRITE);
671 	vm_page_array = (vm_page_t)mapped;
672 	vm_page_array_size = page_range;
673 
674 #if VM_NRESERVLEVEL > 0
675 	/*
676 	 * Allocate physical memory for the reservation management system's
677 	 * data structures, and map it.
678 	 */
679 	if (high_avail == end)
680 		high_avail = new_end;
681 	new_end = vm_reserv_startup(&vaddr, new_end, high_avail);
682 #endif
683 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__)
684 	/*
685 	 * Include vm_page_array and vm_reserv_array in a crash dump.
686 	 */
687 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
688 		dump_add_page(pa);
689 #endif
690 	phys_avail[biggestone + 1] = new_end;
691 
692 	/*
693 	 * Add physical memory segments corresponding to the available
694 	 * physical pages.
695 	 */
696 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
697 		vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]);
698 
699 	/*
700 	 * Initialize the physical memory allocator.
701 	 */
702 	vm_phys_init();
703 
704 	/*
705 	 * Initialize the page structures and add every available page to the
706 	 * physical memory allocator's free lists.
707 	 */
708 	vm_cnt.v_page_count = 0;
709 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
710 		seg = &vm_phys_segs[segind];
711 		for (m = seg->first_page, pa = seg->start; pa < seg->end;
712 		    m++, pa += PAGE_SIZE)
713 			vm_page_init_page(m, pa, segind);
714 
715 		/*
716 		 * Add the segment to the free lists only if it is covered by
717 		 * one of the ranges in phys_avail.  Because we've added the
718 		 * ranges to the vm_phys_segs array, we can assume that each
719 		 * segment is either entirely contained in one of the ranges,
720 		 * or doesn't overlap any of them.
721 		 */
722 		for (i = 0; phys_avail[i + 1] != 0; i += 2) {
723 			struct vm_domain *vmd;
724 
725 			if (seg->start < phys_avail[i] ||
726 			    seg->end > phys_avail[i + 1])
727 				continue;
728 
729 			m = seg->first_page;
730 			pagecount = (u_long)atop(seg->end - seg->start);
731 
732 			vmd = VM_DOMAIN(seg->domain);
733 			vm_domain_free_lock(vmd);
734 			vm_phys_free_contig(m, pagecount);
735 			vm_domain_free_unlock(vmd);
736 			vm_domain_freecnt_inc(vmd, pagecount);
737 			vm_cnt.v_page_count += (u_int)pagecount;
738 
739 			vmd = VM_DOMAIN(seg->domain);
740 			vmd->vmd_page_count += (u_int)pagecount;
741 			vmd->vmd_segs |= 1UL << m->segind;
742 			break;
743 		}
744 	}
745 
746 	/*
747 	 * Remove blacklisted pages from the physical memory allocator.
748 	 */
749 	TAILQ_INIT(&blacklist_head);
750 	vm_page_blacklist_load(&list, &listend);
751 	vm_page_blacklist_check(list, listend);
752 
753 	list = kern_getenv("vm.blacklist");
754 	vm_page_blacklist_check(list, NULL);
755 
756 	freeenv(list);
757 #if VM_NRESERVLEVEL > 0
758 	/*
759 	 * Initialize the reservation management system.
760 	 */
761 	vm_reserv_init();
762 #endif
763 	/*
764 	 * Set an initial domain policy for thread0 so that allocations
765 	 * can work.
766 	 */
767 	domainset_zero();
768 
769 	return (vaddr);
770 }
771 
772 void
773 vm_page_reference(vm_page_t m)
774 {
775 
776 	vm_page_aflag_set(m, PGA_REFERENCED);
777 }
778 
779 /*
780  *	vm_page_busy_downgrade:
781  *
782  *	Downgrade an exclusive busy page into a single shared busy page.
783  */
784 void
785 vm_page_busy_downgrade(vm_page_t m)
786 {
787 	u_int x;
788 	bool locked;
789 
790 	vm_page_assert_xbusied(m);
791 	locked = mtx_owned(vm_page_lockptr(m));
792 
793 	for (;;) {
794 		x = m->busy_lock;
795 		x &= VPB_BIT_WAITERS;
796 		if (x != 0 && !locked)
797 			vm_page_lock(m);
798 		if (atomic_cmpset_rel_int(&m->busy_lock,
799 		    VPB_SINGLE_EXCLUSIVER | x, VPB_SHARERS_WORD(1)))
800 			break;
801 		if (x != 0 && !locked)
802 			vm_page_unlock(m);
803 	}
804 	if (x != 0) {
805 		wakeup(m);
806 		if (!locked)
807 			vm_page_unlock(m);
808 	}
809 }
810 
811 /*
812  *	vm_page_sbusied:
813  *
814  *	Return a positive value if the page is shared busied, 0 otherwise.
815  */
816 int
817 vm_page_sbusied(vm_page_t m)
818 {
819 	u_int x;
820 
821 	x = m->busy_lock;
822 	return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED);
823 }
824 
825 /*
826  *	vm_page_sunbusy:
827  *
828  *	Shared unbusy a page.
829  */
830 void
831 vm_page_sunbusy(vm_page_t m)
832 {
833 	u_int x;
834 
835 	vm_page_lock_assert(m, MA_NOTOWNED);
836 	vm_page_assert_sbusied(m);
837 
838 	for (;;) {
839 		x = m->busy_lock;
840 		if (VPB_SHARERS(x) > 1) {
841 			if (atomic_cmpset_int(&m->busy_lock, x,
842 			    x - VPB_ONE_SHARER))
843 				break;
844 			continue;
845 		}
846 		if ((x & VPB_BIT_WAITERS) == 0) {
847 			KASSERT(x == VPB_SHARERS_WORD(1),
848 			    ("vm_page_sunbusy: invalid lock state"));
849 			if (atomic_cmpset_int(&m->busy_lock,
850 			    VPB_SHARERS_WORD(1), VPB_UNBUSIED))
851 				break;
852 			continue;
853 		}
854 		KASSERT(x == (VPB_SHARERS_WORD(1) | VPB_BIT_WAITERS),
855 		    ("vm_page_sunbusy: invalid lock state for waiters"));
856 
857 		vm_page_lock(m);
858 		if (!atomic_cmpset_int(&m->busy_lock, x, VPB_UNBUSIED)) {
859 			vm_page_unlock(m);
860 			continue;
861 		}
862 		wakeup(m);
863 		vm_page_unlock(m);
864 		break;
865 	}
866 }
867 
868 /*
869  *	vm_page_busy_sleep:
870  *
871  *	Sleep and release the page lock, using the page pointer as wchan.
872  *	This is used to implement the hard-path of busying mechanism.
873  *
874  *	The given page must be locked.
875  *
876  *	If nonshared is true, sleep only if the page is xbusy.
877  */
878 void
879 vm_page_busy_sleep(vm_page_t m, const char *wmesg, bool nonshared)
880 {
881 	u_int x;
882 
883 	vm_page_assert_locked(m);
884 
885 	x = m->busy_lock;
886 	if (x == VPB_UNBUSIED || (nonshared && (x & VPB_BIT_SHARED) != 0) ||
887 	    ((x & VPB_BIT_WAITERS) == 0 &&
888 	    !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS))) {
889 		vm_page_unlock(m);
890 		return;
891 	}
892 	msleep(m, vm_page_lockptr(m), PVM | PDROP, wmesg, 0);
893 }
894 
895 /*
896  *	vm_page_trysbusy:
897  *
898  *	Try to shared busy a page.
899  *	If the operation succeeds 1 is returned otherwise 0.
900  *	The operation never sleeps.
901  */
902 int
903 vm_page_trysbusy(vm_page_t m)
904 {
905 	u_int x;
906 
907 	for (;;) {
908 		x = m->busy_lock;
909 		if ((x & VPB_BIT_SHARED) == 0)
910 			return (0);
911 		if (atomic_cmpset_acq_int(&m->busy_lock, x, x + VPB_ONE_SHARER))
912 			return (1);
913 	}
914 }
915 
916 static void
917 vm_page_xunbusy_locked(vm_page_t m)
918 {
919 
920 	vm_page_assert_xbusied(m);
921 	vm_page_assert_locked(m);
922 
923 	atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
924 	/* There is a waiter, do wakeup() instead of vm_page_flash(). */
925 	wakeup(m);
926 }
927 
928 void
929 vm_page_xunbusy_maybelocked(vm_page_t m)
930 {
931 	bool lockacq;
932 
933 	vm_page_assert_xbusied(m);
934 
935 	/*
936 	 * Fast path for unbusy.  If it succeeds, we know that there
937 	 * are no waiters, so we do not need a wakeup.
938 	 */
939 	if (atomic_cmpset_rel_int(&m->busy_lock, VPB_SINGLE_EXCLUSIVER,
940 	    VPB_UNBUSIED))
941 		return;
942 
943 	lockacq = !mtx_owned(vm_page_lockptr(m));
944 	if (lockacq)
945 		vm_page_lock(m);
946 	vm_page_xunbusy_locked(m);
947 	if (lockacq)
948 		vm_page_unlock(m);
949 }
950 
951 /*
952  *	vm_page_xunbusy_hard:
953  *
954  *	Called after the first try the exclusive unbusy of a page failed.
955  *	It is assumed that the waiters bit is on.
956  */
957 void
958 vm_page_xunbusy_hard(vm_page_t m)
959 {
960 
961 	vm_page_assert_xbusied(m);
962 
963 	vm_page_lock(m);
964 	vm_page_xunbusy_locked(m);
965 	vm_page_unlock(m);
966 }
967 
968 /*
969  *	vm_page_flash:
970  *
971  *	Wakeup anyone waiting for the page.
972  *	The ownership bits do not change.
973  *
974  *	The given page must be locked.
975  */
976 void
977 vm_page_flash(vm_page_t m)
978 {
979 	u_int x;
980 
981 	vm_page_lock_assert(m, MA_OWNED);
982 
983 	for (;;) {
984 		x = m->busy_lock;
985 		if ((x & VPB_BIT_WAITERS) == 0)
986 			return;
987 		if (atomic_cmpset_int(&m->busy_lock, x,
988 		    x & (~VPB_BIT_WAITERS)))
989 			break;
990 	}
991 	wakeup(m);
992 }
993 
994 /*
995  * Avoid releasing and reacquiring the same page lock.
996  */
997 void
998 vm_page_change_lock(vm_page_t m, struct mtx **mtx)
999 {
1000 	struct mtx *mtx1;
1001 
1002 	mtx1 = vm_page_lockptr(m);
1003 	if (*mtx == mtx1)
1004 		return;
1005 	if (*mtx != NULL)
1006 		mtx_unlock(*mtx);
1007 	*mtx = mtx1;
1008 	mtx_lock(mtx1);
1009 }
1010 
1011 /*
1012  * Keep page from being freed by the page daemon
1013  * much of the same effect as wiring, except much lower
1014  * overhead and should be used only for *very* temporary
1015  * holding ("wiring").
1016  */
1017 void
1018 vm_page_hold(vm_page_t mem)
1019 {
1020 
1021 	vm_page_lock_assert(mem, MA_OWNED);
1022         mem->hold_count++;
1023 }
1024 
1025 void
1026 vm_page_unhold(vm_page_t mem)
1027 {
1028 
1029 	vm_page_lock_assert(mem, MA_OWNED);
1030 	KASSERT(mem->hold_count >= 1, ("vm_page_unhold: hold count < 0!!!"));
1031 	--mem->hold_count;
1032 	if (mem->hold_count == 0 && (mem->flags & PG_UNHOLDFREE) != 0)
1033 		vm_page_free_toq(mem);
1034 }
1035 
1036 /*
1037  *	vm_page_unhold_pages:
1038  *
1039  *	Unhold each of the pages that is referenced by the given array.
1040  */
1041 void
1042 vm_page_unhold_pages(vm_page_t *ma, int count)
1043 {
1044 	struct mtx *mtx;
1045 
1046 	mtx = NULL;
1047 	for (; count != 0; count--) {
1048 		vm_page_change_lock(*ma, &mtx);
1049 		vm_page_unhold(*ma);
1050 		ma++;
1051 	}
1052 	if (mtx != NULL)
1053 		mtx_unlock(mtx);
1054 }
1055 
1056 vm_page_t
1057 PHYS_TO_VM_PAGE(vm_paddr_t pa)
1058 {
1059 	vm_page_t m;
1060 
1061 #ifdef VM_PHYSSEG_SPARSE
1062 	m = vm_phys_paddr_to_vm_page(pa);
1063 	if (m == NULL)
1064 		m = vm_phys_fictitious_to_vm_page(pa);
1065 	return (m);
1066 #elif defined(VM_PHYSSEG_DENSE)
1067 	long pi;
1068 
1069 	pi = atop(pa);
1070 	if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
1071 		m = &vm_page_array[pi - first_page];
1072 		return (m);
1073 	}
1074 	return (vm_phys_fictitious_to_vm_page(pa));
1075 #else
1076 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
1077 #endif
1078 }
1079 
1080 /*
1081  *	vm_page_getfake:
1082  *
1083  *	Create a fictitious page with the specified physical address and
1084  *	memory attribute.  The memory attribute is the only the machine-
1085  *	dependent aspect of a fictitious page that must be initialized.
1086  */
1087 vm_page_t
1088 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
1089 {
1090 	vm_page_t m;
1091 
1092 	m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
1093 	vm_page_initfake(m, paddr, memattr);
1094 	return (m);
1095 }
1096 
1097 void
1098 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1099 {
1100 
1101 	if ((m->flags & PG_FICTITIOUS) != 0) {
1102 		/*
1103 		 * The page's memattr might have changed since the
1104 		 * previous initialization.  Update the pmap to the
1105 		 * new memattr.
1106 		 */
1107 		goto memattr;
1108 	}
1109 	m->phys_addr = paddr;
1110 	m->queue = PQ_NONE;
1111 	/* Fictitious pages don't use "segind". */
1112 	m->flags = PG_FICTITIOUS;
1113 	/* Fictitious pages don't use "order" or "pool". */
1114 	m->oflags = VPO_UNMANAGED;
1115 	m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1116 	m->wire_count = 1;
1117 	pmap_page_init(m);
1118 memattr:
1119 	pmap_page_set_memattr(m, memattr);
1120 }
1121 
1122 /*
1123  *	vm_page_putfake:
1124  *
1125  *	Release a fictitious page.
1126  */
1127 void
1128 vm_page_putfake(vm_page_t m)
1129 {
1130 
1131 	KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
1132 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1133 	    ("vm_page_putfake: bad page %p", m));
1134 	uma_zfree(fakepg_zone, m);
1135 }
1136 
1137 /*
1138  *	vm_page_updatefake:
1139  *
1140  *	Update the given fictitious page to the specified physical address and
1141  *	memory attribute.
1142  */
1143 void
1144 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1145 {
1146 
1147 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1148 	    ("vm_page_updatefake: bad page %p", m));
1149 	m->phys_addr = paddr;
1150 	pmap_page_set_memattr(m, memattr);
1151 }
1152 
1153 /*
1154  *	vm_page_free:
1155  *
1156  *	Free a page.
1157  */
1158 void
1159 vm_page_free(vm_page_t m)
1160 {
1161 
1162 	m->flags &= ~PG_ZERO;
1163 	vm_page_free_toq(m);
1164 }
1165 
1166 /*
1167  *	vm_page_free_zero:
1168  *
1169  *	Free a page to the zerod-pages queue
1170  */
1171 void
1172 vm_page_free_zero(vm_page_t m)
1173 {
1174 
1175 	m->flags |= PG_ZERO;
1176 	vm_page_free_toq(m);
1177 }
1178 
1179 /*
1180  * Unbusy and handle the page queueing for a page from a getpages request that
1181  * was optionally read ahead or behind.
1182  */
1183 void
1184 vm_page_readahead_finish(vm_page_t m)
1185 {
1186 
1187 	/* We shouldn't put invalid pages on queues. */
1188 	KASSERT(m->valid != 0, ("%s: %p is invalid", __func__, m));
1189 
1190 	/*
1191 	 * Since the page is not the actually needed one, whether it should
1192 	 * be activated or deactivated is not obvious.  Empirical results
1193 	 * have shown that deactivating the page is usually the best choice,
1194 	 * unless the page is wanted by another thread.
1195 	 */
1196 	vm_page_lock(m);
1197 	if ((m->busy_lock & VPB_BIT_WAITERS) != 0)
1198 		vm_page_activate(m);
1199 	else
1200 		vm_page_deactivate(m);
1201 	vm_page_unlock(m);
1202 	vm_page_xunbusy(m);
1203 }
1204 
1205 /*
1206  *	vm_page_sleep_if_busy:
1207  *
1208  *	Sleep and release the page queues lock if the page is busied.
1209  *	Returns TRUE if the thread slept.
1210  *
1211  *	The given page must be unlocked and object containing it must
1212  *	be locked.
1213  */
1214 int
1215 vm_page_sleep_if_busy(vm_page_t m, const char *msg)
1216 {
1217 	vm_object_t obj;
1218 
1219 	vm_page_lock_assert(m, MA_NOTOWNED);
1220 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1221 
1222 	if (vm_page_busied(m)) {
1223 		/*
1224 		 * The page-specific object must be cached because page
1225 		 * identity can change during the sleep, causing the
1226 		 * re-lock of a different object.
1227 		 * It is assumed that a reference to the object is already
1228 		 * held by the callers.
1229 		 */
1230 		obj = m->object;
1231 		vm_page_lock(m);
1232 		VM_OBJECT_WUNLOCK(obj);
1233 		vm_page_busy_sleep(m, msg, false);
1234 		VM_OBJECT_WLOCK(obj);
1235 		return (TRUE);
1236 	}
1237 	return (FALSE);
1238 }
1239 
1240 /*
1241  *	vm_page_dirty_KBI:		[ internal use only ]
1242  *
1243  *	Set all bits in the page's dirty field.
1244  *
1245  *	The object containing the specified page must be locked if the
1246  *	call is made from the machine-independent layer.
1247  *
1248  *	See vm_page_clear_dirty_mask().
1249  *
1250  *	This function should only be called by vm_page_dirty().
1251  */
1252 void
1253 vm_page_dirty_KBI(vm_page_t m)
1254 {
1255 
1256 	/* Refer to this operation by its public name. */
1257 	KASSERT(m->valid == VM_PAGE_BITS_ALL,
1258 	    ("vm_page_dirty: page is invalid!"));
1259 	m->dirty = VM_PAGE_BITS_ALL;
1260 }
1261 
1262 /*
1263  *	vm_page_insert:		[ internal use only ]
1264  *
1265  *	Inserts the given mem entry into the object and object list.
1266  *
1267  *	The object must be locked.
1268  */
1269 int
1270 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1271 {
1272 	vm_page_t mpred;
1273 
1274 	VM_OBJECT_ASSERT_WLOCKED(object);
1275 	mpred = vm_radix_lookup_le(&object->rtree, pindex);
1276 	return (vm_page_insert_after(m, object, pindex, mpred));
1277 }
1278 
1279 /*
1280  *	vm_page_insert_after:
1281  *
1282  *	Inserts the page "m" into the specified object at offset "pindex".
1283  *
1284  *	The page "mpred" must immediately precede the offset "pindex" within
1285  *	the specified object.
1286  *
1287  *	The object must be locked.
1288  */
1289 static int
1290 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1291     vm_page_t mpred)
1292 {
1293 	vm_page_t msucc;
1294 
1295 	VM_OBJECT_ASSERT_WLOCKED(object);
1296 	KASSERT(m->object == NULL,
1297 	    ("vm_page_insert_after: page already inserted"));
1298 	if (mpred != NULL) {
1299 		KASSERT(mpred->object == object,
1300 		    ("vm_page_insert_after: object doesn't contain mpred"));
1301 		KASSERT(mpred->pindex < pindex,
1302 		    ("vm_page_insert_after: mpred doesn't precede pindex"));
1303 		msucc = TAILQ_NEXT(mpred, listq);
1304 	} else
1305 		msucc = TAILQ_FIRST(&object->memq);
1306 	if (msucc != NULL)
1307 		KASSERT(msucc->pindex > pindex,
1308 		    ("vm_page_insert_after: msucc doesn't succeed pindex"));
1309 
1310 	/*
1311 	 * Record the object/offset pair in this page
1312 	 */
1313 	m->object = object;
1314 	m->pindex = pindex;
1315 
1316 	/*
1317 	 * Now link into the object's ordered list of backed pages.
1318 	 */
1319 	if (vm_radix_insert(&object->rtree, m)) {
1320 		m->object = NULL;
1321 		m->pindex = 0;
1322 		return (1);
1323 	}
1324 	vm_page_insert_radixdone(m, object, mpred);
1325 	return (0);
1326 }
1327 
1328 /*
1329  *	vm_page_insert_radixdone:
1330  *
1331  *	Complete page "m" insertion into the specified object after the
1332  *	radix trie hooking.
1333  *
1334  *	The page "mpred" must precede the offset "m->pindex" within the
1335  *	specified object.
1336  *
1337  *	The object must be locked.
1338  */
1339 static void
1340 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1341 {
1342 
1343 	VM_OBJECT_ASSERT_WLOCKED(object);
1344 	KASSERT(object != NULL && m->object == object,
1345 	    ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1346 	if (mpred != NULL) {
1347 		KASSERT(mpred->object == object,
1348 		    ("vm_page_insert_after: object doesn't contain mpred"));
1349 		KASSERT(mpred->pindex < m->pindex,
1350 		    ("vm_page_insert_after: mpred doesn't precede pindex"));
1351 	}
1352 
1353 	if (mpred != NULL)
1354 		TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1355 	else
1356 		TAILQ_INSERT_HEAD(&object->memq, m, listq);
1357 
1358 	/*
1359 	 * Show that the object has one more resident page.
1360 	 */
1361 	object->resident_page_count++;
1362 
1363 	/*
1364 	 * Hold the vnode until the last page is released.
1365 	 */
1366 	if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1367 		vhold(object->handle);
1368 
1369 	/*
1370 	 * Since we are inserting a new and possibly dirty page,
1371 	 * update the object's OBJ_MIGHTBEDIRTY flag.
1372 	 */
1373 	if (pmap_page_is_write_mapped(m))
1374 		vm_object_set_writeable_dirty(object);
1375 }
1376 
1377 /*
1378  *	vm_page_remove:
1379  *
1380  *	Removes the specified page from its containing object, but does not
1381  *	invalidate any backing storage.
1382  *
1383  *	The object must be locked.  The page must be locked if it is managed.
1384  */
1385 void
1386 vm_page_remove(vm_page_t m)
1387 {
1388 	vm_object_t object;
1389 	vm_page_t mrem;
1390 
1391 	if ((m->oflags & VPO_UNMANAGED) == 0)
1392 		vm_page_assert_locked(m);
1393 	if ((object = m->object) == NULL)
1394 		return;
1395 	VM_OBJECT_ASSERT_WLOCKED(object);
1396 	if (vm_page_xbusied(m))
1397 		vm_page_xunbusy_maybelocked(m);
1398 	mrem = vm_radix_remove(&object->rtree, m->pindex);
1399 	KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m));
1400 
1401 	/*
1402 	 * Now remove from the object's list of backed pages.
1403 	 */
1404 	TAILQ_REMOVE(&object->memq, m, listq);
1405 
1406 	/*
1407 	 * And show that the object has one fewer resident page.
1408 	 */
1409 	object->resident_page_count--;
1410 
1411 	/*
1412 	 * The vnode may now be recycled.
1413 	 */
1414 	if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1415 		vdrop(object->handle);
1416 
1417 	m->object = NULL;
1418 }
1419 
1420 /*
1421  *	vm_page_lookup:
1422  *
1423  *	Returns the page associated with the object/offset
1424  *	pair specified; if none is found, NULL is returned.
1425  *
1426  *	The object must be locked.
1427  */
1428 vm_page_t
1429 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1430 {
1431 
1432 	VM_OBJECT_ASSERT_LOCKED(object);
1433 	return (vm_radix_lookup(&object->rtree, pindex));
1434 }
1435 
1436 /*
1437  *	vm_page_find_least:
1438  *
1439  *	Returns the page associated with the object with least pindex
1440  *	greater than or equal to the parameter pindex, or NULL.
1441  *
1442  *	The object must be locked.
1443  */
1444 vm_page_t
1445 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1446 {
1447 	vm_page_t m;
1448 
1449 	VM_OBJECT_ASSERT_LOCKED(object);
1450 	if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1451 		m = vm_radix_lookup_ge(&object->rtree, pindex);
1452 	return (m);
1453 }
1454 
1455 /*
1456  * Returns the given page's successor (by pindex) within the object if it is
1457  * resident; if none is found, NULL is returned.
1458  *
1459  * The object must be locked.
1460  */
1461 vm_page_t
1462 vm_page_next(vm_page_t m)
1463 {
1464 	vm_page_t next;
1465 
1466 	VM_OBJECT_ASSERT_LOCKED(m->object);
1467 	if ((next = TAILQ_NEXT(m, listq)) != NULL) {
1468 		MPASS(next->object == m->object);
1469 		if (next->pindex != m->pindex + 1)
1470 			next = NULL;
1471 	}
1472 	return (next);
1473 }
1474 
1475 /*
1476  * Returns the given page's predecessor (by pindex) within the object if it is
1477  * resident; if none is found, NULL is returned.
1478  *
1479  * The object must be locked.
1480  */
1481 vm_page_t
1482 vm_page_prev(vm_page_t m)
1483 {
1484 	vm_page_t prev;
1485 
1486 	VM_OBJECT_ASSERT_LOCKED(m->object);
1487 	if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) {
1488 		MPASS(prev->object == m->object);
1489 		if (prev->pindex != m->pindex - 1)
1490 			prev = NULL;
1491 	}
1492 	return (prev);
1493 }
1494 
1495 /*
1496  * Uses the page mnew as a replacement for an existing page at index
1497  * pindex which must be already present in the object.
1498  *
1499  * The existing page must not be on a paging queue.
1500  */
1501 vm_page_t
1502 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex)
1503 {
1504 	vm_page_t mold;
1505 
1506 	VM_OBJECT_ASSERT_WLOCKED(object);
1507 	KASSERT(mnew->object == NULL,
1508 	    ("vm_page_replace: page %p already in object", mnew));
1509 	KASSERT(mnew->queue == PQ_NONE,
1510 	    ("vm_page_replace: new page %p is on a paging queue", mnew));
1511 
1512 	/*
1513 	 * This function mostly follows vm_page_insert() and
1514 	 * vm_page_remove() without the radix, object count and vnode
1515 	 * dance.  Double check such functions for more comments.
1516 	 */
1517 
1518 	mnew->object = object;
1519 	mnew->pindex = pindex;
1520 	mold = vm_radix_replace(&object->rtree, mnew);
1521 	KASSERT(mold->queue == PQ_NONE,
1522 	    ("vm_page_replace: old page %p is on a paging queue", mold));
1523 
1524 	/* Keep the resident page list in sorted order. */
1525 	TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq);
1526 	TAILQ_REMOVE(&object->memq, mold, listq);
1527 
1528 	mold->object = NULL;
1529 	vm_page_xunbusy_maybelocked(mold);
1530 
1531 	/*
1532 	 * The object's resident_page_count does not change because we have
1533 	 * swapped one page for another, but OBJ_MIGHTBEDIRTY.
1534 	 */
1535 	if (pmap_page_is_write_mapped(mnew))
1536 		vm_object_set_writeable_dirty(object);
1537 	return (mold);
1538 }
1539 
1540 /*
1541  *	vm_page_rename:
1542  *
1543  *	Move the given memory entry from its
1544  *	current object to the specified target object/offset.
1545  *
1546  *	Note: swap associated with the page must be invalidated by the move.  We
1547  *	      have to do this for several reasons:  (1) we aren't freeing the
1548  *	      page, (2) we are dirtying the page, (3) the VM system is probably
1549  *	      moving the page from object A to B, and will then later move
1550  *	      the backing store from A to B and we can't have a conflict.
1551  *
1552  *	Note: we *always* dirty the page.  It is necessary both for the
1553  *	      fact that we moved it, and because we may be invalidating
1554  *	      swap.
1555  *
1556  *	The objects must be locked.
1557  */
1558 int
1559 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1560 {
1561 	vm_page_t mpred;
1562 	vm_pindex_t opidx;
1563 
1564 	VM_OBJECT_ASSERT_WLOCKED(new_object);
1565 
1566 	mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex);
1567 	KASSERT(mpred == NULL || mpred->pindex != new_pindex,
1568 	    ("vm_page_rename: pindex already renamed"));
1569 
1570 	/*
1571 	 * Create a custom version of vm_page_insert() which does not depend
1572 	 * by m_prev and can cheat on the implementation aspects of the
1573 	 * function.
1574 	 */
1575 	opidx = m->pindex;
1576 	m->pindex = new_pindex;
1577 	if (vm_radix_insert(&new_object->rtree, m)) {
1578 		m->pindex = opidx;
1579 		return (1);
1580 	}
1581 
1582 	/*
1583 	 * The operation cannot fail anymore.  The removal must happen before
1584 	 * the listq iterator is tainted.
1585 	 */
1586 	m->pindex = opidx;
1587 	vm_page_lock(m);
1588 	vm_page_remove(m);
1589 
1590 	/* Return back to the new pindex to complete vm_page_insert(). */
1591 	m->pindex = new_pindex;
1592 	m->object = new_object;
1593 	vm_page_unlock(m);
1594 	vm_page_insert_radixdone(m, new_object, mpred);
1595 	vm_page_dirty(m);
1596 	return (0);
1597 }
1598 
1599 /*
1600  *	vm_page_alloc:
1601  *
1602  *	Allocate and return a page that is associated with the specified
1603  *	object and offset pair.  By default, this page is exclusive busied.
1604  *
1605  *	The caller must always specify an allocation class.
1606  *
1607  *	allocation classes:
1608  *	VM_ALLOC_NORMAL		normal process request
1609  *	VM_ALLOC_SYSTEM		system *really* needs a page
1610  *	VM_ALLOC_INTERRUPT	interrupt time request
1611  *
1612  *	optional allocation flags:
1613  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
1614  *				intends to allocate
1615  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
1616  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
1617  *	VM_ALLOC_NOOBJ		page is not associated with an object and
1618  *				should not be exclusive busy
1619  *	VM_ALLOC_SBUSY		shared busy the allocated page
1620  *	VM_ALLOC_WIRED		wire the allocated page
1621  *	VM_ALLOC_ZERO		prefer a zeroed page
1622  */
1623 vm_page_t
1624 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1625 {
1626 
1627 	return (vm_page_alloc_after(object, pindex, req, object != NULL ?
1628 	    vm_radix_lookup_le(&object->rtree, pindex) : NULL));
1629 }
1630 
1631 vm_page_t
1632 vm_page_alloc_domain(vm_object_t object, vm_pindex_t pindex, int domain,
1633     int req)
1634 {
1635 
1636 	return (vm_page_alloc_domain_after(object, pindex, domain, req,
1637 	    object != NULL ? vm_radix_lookup_le(&object->rtree, pindex) :
1638 	    NULL));
1639 }
1640 
1641 /*
1642  * Allocate a page in the specified object with the given page index.  To
1643  * optimize insertion of the page into the object, the caller must also specifiy
1644  * the resident page in the object with largest index smaller than the given
1645  * page index, or NULL if no such page exists.
1646  */
1647 vm_page_t
1648 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex,
1649     int req, vm_page_t mpred)
1650 {
1651 	struct vm_domainset_iter di;
1652 	vm_page_t m;
1653 	int domain;
1654 
1655 	vm_domainset_iter_page_init(&di, object, &domain, &req);
1656 	do {
1657 		m = vm_page_alloc_domain_after(object, pindex, domain, req,
1658 		    mpred);
1659 		if (m != NULL)
1660 			break;
1661 	} while (vm_domainset_iter_page(&di, &domain, &req) == 0);
1662 
1663 	return (m);
1664 }
1665 
1666 /*
1667  * Returns true if the number of free pages exceeds the minimum
1668  * for the request class and false otherwise.
1669  */
1670 int
1671 vm_domain_available(struct vm_domain *vmd, int req, int npages)
1672 {
1673 
1674 	vm_domain_free_assert_locked(vmd);
1675 	req = req & VM_ALLOC_CLASS_MASK;
1676 
1677 	/*
1678 	 * The page daemon is allowed to dig deeper into the free page list.
1679 	 */
1680 	if (curproc == pageproc && req != VM_ALLOC_INTERRUPT)
1681 		req = VM_ALLOC_SYSTEM;
1682 
1683 	if (vmd->vmd_free_count >= npages + vmd->vmd_free_reserved ||
1684 	    (req == VM_ALLOC_SYSTEM &&
1685 	    vmd->vmd_free_count >= npages + vmd->vmd_interrupt_free_min) ||
1686 	    (req == VM_ALLOC_INTERRUPT &&
1687 	    vmd->vmd_free_count >= npages))
1688 		return (1);
1689 
1690 	return (0);
1691 }
1692 
1693 vm_page_t
1694 vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain,
1695     int req, vm_page_t mpred)
1696 {
1697 	struct vm_domain *vmd;
1698 	vm_page_t m;
1699 	int flags;
1700 
1701 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1702 	    (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1703 	    ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1704 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1705 	    ("inconsistent object(%p)/req(%x)", object, req));
1706 	KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
1707 	    ("Can't sleep and retry object insertion."));
1708 	KASSERT(mpred == NULL || mpred->pindex < pindex,
1709 	    ("mpred %p doesn't precede pindex 0x%jx", mpred,
1710 	    (uintmax_t)pindex));
1711 	if (object != NULL)
1712 		VM_OBJECT_ASSERT_WLOCKED(object);
1713 
1714 again:
1715 	m = NULL;
1716 #if VM_NRESERVLEVEL > 0
1717 	if (vm_object_reserv(object) &&
1718 	    (m = vm_reserv_extend(req, object, pindex, domain, mpred))
1719 	    != NULL) {
1720 		domain = vm_phys_domain(m);
1721 		vmd = VM_DOMAIN(domain);
1722 		goto found;
1723 	}
1724 #endif
1725 	vmd = VM_DOMAIN(domain);
1726 	vm_domain_free_lock(vmd);
1727 	if (vm_domain_available(vmd, req, 1)) {
1728 		/*
1729 		 * Can we allocate the page from a reservation?
1730 		 */
1731 #if VM_NRESERVLEVEL > 0
1732 		if (!vm_object_reserv(object) ||
1733 		    (m = vm_reserv_alloc_page(object, pindex,
1734 		    domain, mpred)) == NULL)
1735 #endif
1736 		{
1737 			/*
1738 			 * If not, allocate it from the free page queues.
1739 			 */
1740 			m = vm_phys_alloc_pages(domain, object != NULL ?
1741 			    VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0);
1742 #if VM_NRESERVLEVEL > 0
1743 			if (m == NULL && vm_reserv_reclaim_inactive(domain)) {
1744 				m = vm_phys_alloc_pages(domain,
1745 				    object != NULL ?
1746 				    VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT,
1747 				    0);
1748 			}
1749 #endif
1750 		}
1751 	}
1752 	if (m != NULL)
1753 		vm_domain_freecnt_dec(vmd, 1);
1754 	vm_domain_free_unlock(vmd);
1755 	if (m == NULL) {
1756 		/*
1757 		 * Not allocatable, give up.
1758 		 */
1759 		if (vm_domain_alloc_fail(vmd, object, req))
1760 			goto again;
1761 		return (NULL);
1762 	}
1763 
1764 	/*
1765 	 *  At this point we had better have found a good page.
1766 	 */
1767 	KASSERT(m != NULL, ("missing page"));
1768 
1769 #if VM_NRESERVLEVEL > 0
1770 found:
1771 #endif
1772 	vm_page_alloc_check(m);
1773 
1774 	/*
1775 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
1776 	 */
1777 	flags = 0;
1778 	if ((req & VM_ALLOC_ZERO) != 0)
1779 		flags = PG_ZERO;
1780 	flags &= m->flags;
1781 	if ((req & VM_ALLOC_NODUMP) != 0)
1782 		flags |= PG_NODUMP;
1783 	m->flags = flags;
1784 	m->aflags = 0;
1785 	m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
1786 	    VPO_UNMANAGED : 0;
1787 	m->busy_lock = VPB_UNBUSIED;
1788 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
1789 		m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1790 	if ((req & VM_ALLOC_SBUSY) != 0)
1791 		m->busy_lock = VPB_SHARERS_WORD(1);
1792 	if (req & VM_ALLOC_WIRED) {
1793 		/*
1794 		 * The page lock is not required for wiring a page until that
1795 		 * page is inserted into the object.
1796 		 */
1797 		vm_wire_add(1);
1798 		m->wire_count = 1;
1799 	}
1800 	m->act_count = 0;
1801 
1802 	if (object != NULL) {
1803 		if (vm_page_insert_after(m, object, pindex, mpred)) {
1804 			if (req & VM_ALLOC_WIRED) {
1805 				vm_wire_sub(1);
1806 				m->wire_count = 0;
1807 			}
1808 			KASSERT(m->object == NULL, ("page %p has object", m));
1809 			m->oflags = VPO_UNMANAGED;
1810 			m->busy_lock = VPB_UNBUSIED;
1811 			/* Don't change PG_ZERO. */
1812 			vm_page_free_toq(m);
1813 			if (req & VM_ALLOC_WAITFAIL) {
1814 				VM_OBJECT_WUNLOCK(object);
1815 				vm_radix_wait();
1816 				VM_OBJECT_WLOCK(object);
1817 			}
1818 			return (NULL);
1819 		}
1820 
1821 		/* Ignore device objects; the pager sets "memattr" for them. */
1822 		if (object->memattr != VM_MEMATTR_DEFAULT &&
1823 		    (object->flags & OBJ_FICTITIOUS) == 0)
1824 			pmap_page_set_memattr(m, object->memattr);
1825 	} else
1826 		m->pindex = pindex;
1827 
1828 	return (m);
1829 }
1830 
1831 /*
1832  *	vm_page_alloc_contig:
1833  *
1834  *	Allocate a contiguous set of physical pages of the given size "npages"
1835  *	from the free lists.  All of the physical pages must be at or above
1836  *	the given physical address "low" and below the given physical address
1837  *	"high".  The given value "alignment" determines the alignment of the
1838  *	first physical page in the set.  If the given value "boundary" is
1839  *	non-zero, then the set of physical pages cannot cross any physical
1840  *	address boundary that is a multiple of that value.  Both "alignment"
1841  *	and "boundary" must be a power of two.
1842  *
1843  *	If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
1844  *	then the memory attribute setting for the physical pages is configured
1845  *	to the object's memory attribute setting.  Otherwise, the memory
1846  *	attribute setting for the physical pages is configured to "memattr",
1847  *	overriding the object's memory attribute setting.  However, if the
1848  *	object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
1849  *	memory attribute setting for the physical pages cannot be configured
1850  *	to VM_MEMATTR_DEFAULT.
1851  *
1852  *	The specified object may not contain fictitious pages.
1853  *
1854  *	The caller must always specify an allocation class.
1855  *
1856  *	allocation classes:
1857  *	VM_ALLOC_NORMAL		normal process request
1858  *	VM_ALLOC_SYSTEM		system *really* needs a page
1859  *	VM_ALLOC_INTERRUPT	interrupt time request
1860  *
1861  *	optional allocation flags:
1862  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
1863  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
1864  *	VM_ALLOC_NOOBJ		page is not associated with an object and
1865  *				should not be exclusive busy
1866  *	VM_ALLOC_SBUSY		shared busy the allocated page
1867  *	VM_ALLOC_WIRED		wire the allocated page
1868  *	VM_ALLOC_ZERO		prefer a zeroed page
1869  */
1870 vm_page_t
1871 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
1872     u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
1873     vm_paddr_t boundary, vm_memattr_t memattr)
1874 {
1875 	struct vm_domainset_iter di;
1876 	vm_page_t m;
1877 	int domain;
1878 
1879 	vm_domainset_iter_page_init(&di, object, &domain, &req);
1880 	do {
1881 		m = vm_page_alloc_contig_domain(object, pindex, domain, req,
1882 		    npages, low, high, alignment, boundary, memattr);
1883 		if (m != NULL)
1884 			break;
1885 	} while (vm_domainset_iter_page(&di, &domain, &req) == 0);
1886 
1887 	return (m);
1888 }
1889 
1890 vm_page_t
1891 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain,
1892     int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
1893     vm_paddr_t boundary, vm_memattr_t memattr)
1894 {
1895 	struct vm_domain *vmd;
1896 	vm_page_t m, m_ret, mpred;
1897 	u_int busy_lock, flags, oflags;
1898 
1899 	mpred = NULL;	/* XXX: pacify gcc */
1900 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1901 	    (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1902 	    ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1903 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1904 	    ("vm_page_alloc_contig: inconsistent object(%p)/req(%x)", object,
1905 	    req));
1906 	KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
1907 	    ("Can't sleep and retry object insertion."));
1908 	if (object != NULL) {
1909 		VM_OBJECT_ASSERT_WLOCKED(object);
1910 		KASSERT((object->flags & OBJ_FICTITIOUS) == 0,
1911 		    ("vm_page_alloc_contig: object %p has fictitious pages",
1912 		    object));
1913 	}
1914 	KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
1915 
1916 	if (object != NULL) {
1917 		mpred = vm_radix_lookup_le(&object->rtree, pindex);
1918 		KASSERT(mpred == NULL || mpred->pindex != pindex,
1919 		    ("vm_page_alloc_contig: pindex already allocated"));
1920 	}
1921 
1922 	/*
1923 	 * Can we allocate the pages without the number of free pages falling
1924 	 * below the lower bound for the allocation class?
1925 	 */
1926 again:
1927 #if VM_NRESERVLEVEL > 0
1928 	if (vm_object_reserv(object) &&
1929 	    (m_ret = vm_reserv_extend_contig(req, object, pindex, domain,
1930 	    npages, low, high, alignment, boundary, mpred)) != NULL) {
1931 		domain = vm_phys_domain(m_ret);
1932 		vmd = VM_DOMAIN(domain);
1933 		goto found;
1934 	}
1935 #endif
1936 	m_ret = NULL;
1937 	vmd = VM_DOMAIN(domain);
1938 	vm_domain_free_lock(vmd);
1939 	if (vm_domain_available(vmd, req, npages)) {
1940 		/*
1941 		 * Can we allocate the pages from a reservation?
1942 		 */
1943 #if VM_NRESERVLEVEL > 0
1944 retry:
1945 		if (!vm_object_reserv(object) ||
1946 		    (m_ret = vm_reserv_alloc_contig(object, pindex, domain,
1947 		    npages, low, high, alignment, boundary, mpred)) == NULL)
1948 #endif
1949 			/*
1950 			 * If not, allocate them from the free page queues.
1951 			 */
1952 			m_ret = vm_phys_alloc_contig(domain, npages, low, high,
1953 			    alignment, boundary);
1954 #if VM_NRESERVLEVEL > 0
1955 		if (m_ret == NULL && vm_reserv_reclaim_contig(
1956 		    domain, npages, low, high, alignment, boundary))
1957 			goto retry;
1958 #endif
1959 	}
1960 	if (m_ret != NULL)
1961 		vm_domain_freecnt_dec(vmd, npages);
1962 	vm_domain_free_unlock(vmd);
1963 	if (m_ret == NULL) {
1964 		if (vm_domain_alloc_fail(vmd, object, req))
1965 			goto again;
1966 		return (NULL);
1967 	}
1968 #if VM_NRESERVLEVEL > 0
1969 found:
1970 #endif
1971 	for (m = m_ret; m < &m_ret[npages]; m++)
1972 		vm_page_alloc_check(m);
1973 
1974 	/*
1975 	 * Initialize the pages.  Only the PG_ZERO flag is inherited.
1976 	 */
1977 	flags = 0;
1978 	if ((req & VM_ALLOC_ZERO) != 0)
1979 		flags = PG_ZERO;
1980 	if ((req & VM_ALLOC_NODUMP) != 0)
1981 		flags |= PG_NODUMP;
1982 	oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
1983 	    VPO_UNMANAGED : 0;
1984 	busy_lock = VPB_UNBUSIED;
1985 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
1986 		busy_lock = VPB_SINGLE_EXCLUSIVER;
1987 	if ((req & VM_ALLOC_SBUSY) != 0)
1988 		busy_lock = VPB_SHARERS_WORD(1);
1989 	if ((req & VM_ALLOC_WIRED) != 0)
1990 		vm_wire_add(npages);
1991 	if (object != NULL) {
1992 		if (object->memattr != VM_MEMATTR_DEFAULT &&
1993 		    memattr == VM_MEMATTR_DEFAULT)
1994 			memattr = object->memattr;
1995 	}
1996 	for (m = m_ret; m < &m_ret[npages]; m++) {
1997 		m->aflags = 0;
1998 		m->flags = (m->flags | PG_NODUMP) & flags;
1999 		m->busy_lock = busy_lock;
2000 		if ((req & VM_ALLOC_WIRED) != 0)
2001 			m->wire_count = 1;
2002 		m->act_count = 0;
2003 		m->oflags = oflags;
2004 		if (object != NULL) {
2005 			if (vm_page_insert_after(m, object, pindex, mpred)) {
2006 				if ((req & VM_ALLOC_WIRED) != 0)
2007 					vm_wire_sub(npages);
2008 				KASSERT(m->object == NULL,
2009 				    ("page %p has object", m));
2010 				mpred = m;
2011 				for (m = m_ret; m < &m_ret[npages]; m++) {
2012 					if (m <= mpred &&
2013 					    (req & VM_ALLOC_WIRED) != 0)
2014 						m->wire_count = 0;
2015 					m->oflags = VPO_UNMANAGED;
2016 					m->busy_lock = VPB_UNBUSIED;
2017 					/* Don't change PG_ZERO. */
2018 					vm_page_free_toq(m);
2019 				}
2020 				if (req & VM_ALLOC_WAITFAIL) {
2021 					VM_OBJECT_WUNLOCK(object);
2022 					vm_radix_wait();
2023 					VM_OBJECT_WLOCK(object);
2024 				}
2025 				return (NULL);
2026 			}
2027 			mpred = m;
2028 		} else
2029 			m->pindex = pindex;
2030 		if (memattr != VM_MEMATTR_DEFAULT)
2031 			pmap_page_set_memattr(m, memattr);
2032 		pindex++;
2033 	}
2034 	return (m_ret);
2035 }
2036 
2037 /*
2038  * Check a page that has been freshly dequeued from a freelist.
2039  */
2040 static void
2041 vm_page_alloc_check(vm_page_t m)
2042 {
2043 
2044 	KASSERT(m->object == NULL, ("page %p has object", m));
2045 	KASSERT(m->queue == PQ_NONE,
2046 	    ("page %p has unexpected queue %d", m, m->queue));
2047 	KASSERT(!vm_page_held(m), ("page %p is held", m));
2048 	KASSERT(!vm_page_busied(m), ("page %p is busy", m));
2049 	KASSERT(m->dirty == 0, ("page %p is dirty", m));
2050 	KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
2051 	    ("page %p has unexpected memattr %d",
2052 	    m, pmap_page_get_memattr(m)));
2053 	KASSERT(m->valid == 0, ("free page %p is valid", m));
2054 }
2055 
2056 /*
2057  * 	vm_page_alloc_freelist:
2058  *
2059  *	Allocate a physical page from the specified free page list.
2060  *
2061  *	The caller must always specify an allocation class.
2062  *
2063  *	allocation classes:
2064  *	VM_ALLOC_NORMAL		normal process request
2065  *	VM_ALLOC_SYSTEM		system *really* needs a page
2066  *	VM_ALLOC_INTERRUPT	interrupt time request
2067  *
2068  *	optional allocation flags:
2069  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
2070  *				intends to allocate
2071  *	VM_ALLOC_WIRED		wire the allocated page
2072  *	VM_ALLOC_ZERO		prefer a zeroed page
2073  */
2074 vm_page_t
2075 vm_page_alloc_freelist(int freelist, int req)
2076 {
2077 	struct vm_domainset_iter di;
2078 	vm_page_t m;
2079 	int domain;
2080 
2081 	vm_domainset_iter_page_init(&di, kernel_object, &domain, &req);
2082 	do {
2083 		m = vm_page_alloc_freelist_domain(domain, freelist, req);
2084 		if (m != NULL)
2085 			break;
2086 	} while (vm_domainset_iter_page(&di, &domain, &req) == 0);
2087 
2088 	return (m);
2089 }
2090 
2091 vm_page_t
2092 vm_page_alloc_freelist_domain(int domain, int freelist, int req)
2093 {
2094 	struct vm_domain *vmd;
2095 	vm_page_t m;
2096 	u_int flags;
2097 
2098 	/*
2099 	 * Do not allocate reserved pages unless the req has asked for it.
2100 	 */
2101 	vmd = VM_DOMAIN(domain);
2102 again:
2103 	vm_domain_free_lock(vmd);
2104 	if (vm_domain_available(vmd, req, 1))
2105 		m = vm_phys_alloc_freelist_pages(domain, freelist,
2106 		    VM_FREEPOOL_DIRECT, 0);
2107 	if (m != NULL)
2108 		vm_domain_freecnt_dec(vmd, 1);
2109 	vm_domain_free_unlock(vmd);
2110 	if (m == NULL) {
2111 		if (vm_domain_alloc_fail(vmd, NULL, req))
2112 			goto again;
2113 		return (NULL);
2114 	}
2115 	vm_page_alloc_check(m);
2116 
2117 	/*
2118 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
2119 	 */
2120 	m->aflags = 0;
2121 	flags = 0;
2122 	if ((req & VM_ALLOC_ZERO) != 0)
2123 		flags = PG_ZERO;
2124 	m->flags &= flags;
2125 	if ((req & VM_ALLOC_WIRED) != 0) {
2126 		/*
2127 		 * The page lock is not required for wiring a page that does
2128 		 * not belong to an object.
2129 		 */
2130 		vm_wire_add(1);
2131 		m->wire_count = 1;
2132 	}
2133 	/* Unmanaged pages don't use "act_count". */
2134 	m->oflags = VPO_UNMANAGED;
2135 	return (m);
2136 }
2137 
2138 #define	VPSC_ANY	0	/* No restrictions. */
2139 #define	VPSC_NORESERV	1	/* Skip reservations; implies VPSC_NOSUPER. */
2140 #define	VPSC_NOSUPER	2	/* Skip superpages. */
2141 
2142 /*
2143  *	vm_page_scan_contig:
2144  *
2145  *	Scan vm_page_array[] between the specified entries "m_start" and
2146  *	"m_end" for a run of contiguous physical pages that satisfy the
2147  *	specified conditions, and return the lowest page in the run.  The
2148  *	specified "alignment" determines the alignment of the lowest physical
2149  *	page in the run.  If the specified "boundary" is non-zero, then the
2150  *	run of physical pages cannot span a physical address that is a
2151  *	multiple of "boundary".
2152  *
2153  *	"m_end" is never dereferenced, so it need not point to a vm_page
2154  *	structure within vm_page_array[].
2155  *
2156  *	"npages" must be greater than zero.  "m_start" and "m_end" must not
2157  *	span a hole (or discontiguity) in the physical address space.  Both
2158  *	"alignment" and "boundary" must be a power of two.
2159  */
2160 vm_page_t
2161 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end,
2162     u_long alignment, vm_paddr_t boundary, int options)
2163 {
2164 	struct mtx *m_mtx;
2165 	vm_object_t object;
2166 	vm_paddr_t pa;
2167 	vm_page_t m, m_run;
2168 #if VM_NRESERVLEVEL > 0
2169 	int level;
2170 #endif
2171 	int m_inc, order, run_ext, run_len;
2172 
2173 	KASSERT(npages > 0, ("npages is 0"));
2174 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2175 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2176 	m_run = NULL;
2177 	run_len = 0;
2178 	m_mtx = NULL;
2179 	for (m = m_start; m < m_end && run_len < npages; m += m_inc) {
2180 		KASSERT((m->flags & PG_MARKER) == 0,
2181 		    ("page %p is PG_MARKER", m));
2182 		KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->wire_count == 1,
2183 		    ("fictitious page %p has invalid wire count", m));
2184 
2185 		/*
2186 		 * If the current page would be the start of a run, check its
2187 		 * physical address against the end, alignment, and boundary
2188 		 * conditions.  If it doesn't satisfy these conditions, either
2189 		 * terminate the scan or advance to the next page that
2190 		 * satisfies the failed condition.
2191 		 */
2192 		if (run_len == 0) {
2193 			KASSERT(m_run == NULL, ("m_run != NULL"));
2194 			if (m + npages > m_end)
2195 				break;
2196 			pa = VM_PAGE_TO_PHYS(m);
2197 			if ((pa & (alignment - 1)) != 0) {
2198 				m_inc = atop(roundup2(pa, alignment) - pa);
2199 				continue;
2200 			}
2201 			if (rounddown2(pa ^ (pa + ptoa(npages) - 1),
2202 			    boundary) != 0) {
2203 				m_inc = atop(roundup2(pa, boundary) - pa);
2204 				continue;
2205 			}
2206 		} else
2207 			KASSERT(m_run != NULL, ("m_run == NULL"));
2208 
2209 		vm_page_change_lock(m, &m_mtx);
2210 		m_inc = 1;
2211 retry:
2212 		if (vm_page_held(m))
2213 			run_ext = 0;
2214 #if VM_NRESERVLEVEL > 0
2215 		else if ((level = vm_reserv_level(m)) >= 0 &&
2216 		    (options & VPSC_NORESERV) != 0) {
2217 			run_ext = 0;
2218 			/* Advance to the end of the reservation. */
2219 			pa = VM_PAGE_TO_PHYS(m);
2220 			m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) -
2221 			    pa);
2222 		}
2223 #endif
2224 		else if ((object = m->object) != NULL) {
2225 			/*
2226 			 * The page is considered eligible for relocation if
2227 			 * and only if it could be laundered or reclaimed by
2228 			 * the page daemon.
2229 			 */
2230 			if (!VM_OBJECT_TRYRLOCK(object)) {
2231 				mtx_unlock(m_mtx);
2232 				VM_OBJECT_RLOCK(object);
2233 				mtx_lock(m_mtx);
2234 				if (m->object != object) {
2235 					/*
2236 					 * The page may have been freed.
2237 					 */
2238 					VM_OBJECT_RUNLOCK(object);
2239 					goto retry;
2240 				} else if (vm_page_held(m)) {
2241 					run_ext = 0;
2242 					goto unlock;
2243 				}
2244 			}
2245 			KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2246 			    ("page %p is PG_UNHOLDFREE", m));
2247 			/* Don't care: PG_NODUMP, PG_ZERO. */
2248 			if (object->type != OBJT_DEFAULT &&
2249 			    object->type != OBJT_SWAP &&
2250 			    object->type != OBJT_VNODE) {
2251 				run_ext = 0;
2252 #if VM_NRESERVLEVEL > 0
2253 			} else if ((options & VPSC_NOSUPER) != 0 &&
2254 			    (level = vm_reserv_level_iffullpop(m)) >= 0) {
2255 				run_ext = 0;
2256 				/* Advance to the end of the superpage. */
2257 				pa = VM_PAGE_TO_PHYS(m);
2258 				m_inc = atop(roundup2(pa + 1,
2259 				    vm_reserv_size(level)) - pa);
2260 #endif
2261 			} else if (object->memattr == VM_MEMATTR_DEFAULT &&
2262 			    m->queue != PQ_NONE && !vm_page_busied(m)) {
2263 				/*
2264 				 * The page is allocated but eligible for
2265 				 * relocation.  Extend the current run by one
2266 				 * page.
2267 				 */
2268 				KASSERT(pmap_page_get_memattr(m) ==
2269 				    VM_MEMATTR_DEFAULT,
2270 				    ("page %p has an unexpected memattr", m));
2271 				KASSERT((m->oflags & (VPO_SWAPINPROG |
2272 				    VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2273 				    ("page %p has unexpected oflags", m));
2274 				/* Don't care: VPO_NOSYNC. */
2275 				run_ext = 1;
2276 			} else
2277 				run_ext = 0;
2278 unlock:
2279 			VM_OBJECT_RUNLOCK(object);
2280 #if VM_NRESERVLEVEL > 0
2281 		} else if (level >= 0) {
2282 			/*
2283 			 * The page is reserved but not yet allocated.  In
2284 			 * other words, it is still free.  Extend the current
2285 			 * run by one page.
2286 			 */
2287 			run_ext = 1;
2288 #endif
2289 		} else if ((order = m->order) < VM_NFREEORDER) {
2290 			/*
2291 			 * The page is enqueued in the physical memory
2292 			 * allocator's free page queues.  Moreover, it is the
2293 			 * first page in a power-of-two-sized run of
2294 			 * contiguous free pages.  Add these pages to the end
2295 			 * of the current run, and jump ahead.
2296 			 */
2297 			run_ext = 1 << order;
2298 			m_inc = 1 << order;
2299 		} else {
2300 			/*
2301 			 * Skip the page for one of the following reasons: (1)
2302 			 * It is enqueued in the physical memory allocator's
2303 			 * free page queues.  However, it is not the first
2304 			 * page in a run of contiguous free pages.  (This case
2305 			 * rarely occurs because the scan is performed in
2306 			 * ascending order.) (2) It is not reserved, and it is
2307 			 * transitioning from free to allocated.  (Conversely,
2308 			 * the transition from allocated to free for managed
2309 			 * pages is blocked by the page lock.) (3) It is
2310 			 * allocated but not contained by an object and not
2311 			 * wired, e.g., allocated by Xen's balloon driver.
2312 			 */
2313 			run_ext = 0;
2314 		}
2315 
2316 		/*
2317 		 * Extend or reset the current run of pages.
2318 		 */
2319 		if (run_ext > 0) {
2320 			if (run_len == 0)
2321 				m_run = m;
2322 			run_len += run_ext;
2323 		} else {
2324 			if (run_len > 0) {
2325 				m_run = NULL;
2326 				run_len = 0;
2327 			}
2328 		}
2329 	}
2330 	if (m_mtx != NULL)
2331 		mtx_unlock(m_mtx);
2332 	if (run_len >= npages)
2333 		return (m_run);
2334 	return (NULL);
2335 }
2336 
2337 /*
2338  *	vm_page_reclaim_run:
2339  *
2340  *	Try to relocate each of the allocated virtual pages within the
2341  *	specified run of physical pages to a new physical address.  Free the
2342  *	physical pages underlying the relocated virtual pages.  A virtual page
2343  *	is relocatable if and only if it could be laundered or reclaimed by
2344  *	the page daemon.  Whenever possible, a virtual page is relocated to a
2345  *	physical address above "high".
2346  *
2347  *	Returns 0 if every physical page within the run was already free or
2348  *	just freed by a successful relocation.  Otherwise, returns a non-zero
2349  *	value indicating why the last attempt to relocate a virtual page was
2350  *	unsuccessful.
2351  *
2352  *	"req_class" must be an allocation class.
2353  */
2354 static int
2355 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run,
2356     vm_paddr_t high)
2357 {
2358 	struct vm_domain *vmd;
2359 	struct mtx *m_mtx;
2360 	struct spglist free;
2361 	vm_object_t object;
2362 	vm_paddr_t pa;
2363 	vm_page_t m, m_end, m_new;
2364 	int error, order, req;
2365 
2366 	KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class,
2367 	    ("req_class is not an allocation class"));
2368 	SLIST_INIT(&free);
2369 	error = 0;
2370 	m = m_run;
2371 	m_end = m_run + npages;
2372 	m_mtx = NULL;
2373 	for (; error == 0 && m < m_end; m++) {
2374 		KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2375 		    ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2376 
2377 		/*
2378 		 * Avoid releasing and reacquiring the same page lock.
2379 		 */
2380 		vm_page_change_lock(m, &m_mtx);
2381 retry:
2382 		if (vm_page_held(m))
2383 			error = EBUSY;
2384 		else if ((object = m->object) != NULL) {
2385 			/*
2386 			 * The page is relocated if and only if it could be
2387 			 * laundered or reclaimed by the page daemon.
2388 			 */
2389 			if (!VM_OBJECT_TRYWLOCK(object)) {
2390 				mtx_unlock(m_mtx);
2391 				VM_OBJECT_WLOCK(object);
2392 				mtx_lock(m_mtx);
2393 				if (m->object != object) {
2394 					/*
2395 					 * The page may have been freed.
2396 					 */
2397 					VM_OBJECT_WUNLOCK(object);
2398 					goto retry;
2399 				} else if (vm_page_held(m)) {
2400 					error = EBUSY;
2401 					goto unlock;
2402 				}
2403 			}
2404 			KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2405 			    ("page %p is PG_UNHOLDFREE", m));
2406 			/* Don't care: PG_NODUMP, PG_ZERO. */
2407 			if (object->type != OBJT_DEFAULT &&
2408 			    object->type != OBJT_SWAP &&
2409 			    object->type != OBJT_VNODE)
2410 				error = EINVAL;
2411 			else if (object->memattr != VM_MEMATTR_DEFAULT)
2412 				error = EINVAL;
2413 			else if (m->queue != PQ_NONE && !vm_page_busied(m)) {
2414 				KASSERT(pmap_page_get_memattr(m) ==
2415 				    VM_MEMATTR_DEFAULT,
2416 				    ("page %p has an unexpected memattr", m));
2417 				KASSERT((m->oflags & (VPO_SWAPINPROG |
2418 				    VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2419 				    ("page %p has unexpected oflags", m));
2420 				/* Don't care: VPO_NOSYNC. */
2421 				if (m->valid != 0) {
2422 					/*
2423 					 * First, try to allocate a new page
2424 					 * that is above "high".  Failing
2425 					 * that, try to allocate a new page
2426 					 * that is below "m_run".  Allocate
2427 					 * the new page between the end of
2428 					 * "m_run" and "high" only as a last
2429 					 * resort.
2430 					 */
2431 					req = req_class | VM_ALLOC_NOOBJ;
2432 					if ((m->flags & PG_NODUMP) != 0)
2433 						req |= VM_ALLOC_NODUMP;
2434 					if (trunc_page(high) !=
2435 					    ~(vm_paddr_t)PAGE_MASK) {
2436 						m_new = vm_page_alloc_contig(
2437 						    NULL, 0, req, 1,
2438 						    round_page(high),
2439 						    ~(vm_paddr_t)0,
2440 						    PAGE_SIZE, 0,
2441 						    VM_MEMATTR_DEFAULT);
2442 					} else
2443 						m_new = NULL;
2444 					if (m_new == NULL) {
2445 						pa = VM_PAGE_TO_PHYS(m_run);
2446 						m_new = vm_page_alloc_contig(
2447 						    NULL, 0, req, 1,
2448 						    0, pa - 1, PAGE_SIZE, 0,
2449 						    VM_MEMATTR_DEFAULT);
2450 					}
2451 					if (m_new == NULL) {
2452 						pa += ptoa(npages);
2453 						m_new = vm_page_alloc_contig(
2454 						    NULL, 0, req, 1,
2455 						    pa, high, PAGE_SIZE, 0,
2456 						    VM_MEMATTR_DEFAULT);
2457 					}
2458 					if (m_new == NULL) {
2459 						error = ENOMEM;
2460 						goto unlock;
2461 					}
2462 					KASSERT(m_new->wire_count == 0,
2463 					    ("page %p is wired", m));
2464 
2465 					/*
2466 					 * Replace "m" with the new page.  For
2467 					 * vm_page_replace(), "m" must be busy
2468 					 * and dequeued.  Finally, change "m"
2469 					 * as if vm_page_free() was called.
2470 					 */
2471 					if (object->ref_count != 0)
2472 						pmap_remove_all(m);
2473 					m_new->aflags = m->aflags;
2474 					KASSERT(m_new->oflags == VPO_UNMANAGED,
2475 					    ("page %p is managed", m));
2476 					m_new->oflags = m->oflags & VPO_NOSYNC;
2477 					pmap_copy_page(m, m_new);
2478 					m_new->valid = m->valid;
2479 					m_new->dirty = m->dirty;
2480 					m->flags &= ~PG_ZERO;
2481 					vm_page_xbusy(m);
2482 					vm_page_remque(m);
2483 					vm_page_replace_checked(m_new, object,
2484 					    m->pindex, m);
2485 					m->valid = 0;
2486 					vm_page_undirty(m);
2487 
2488 					/*
2489 					 * The new page must be deactivated
2490 					 * before the object is unlocked.
2491 					 */
2492 					vm_page_change_lock(m_new, &m_mtx);
2493 					vm_page_deactivate(m_new);
2494 				} else {
2495 					m->flags &= ~PG_ZERO;
2496 					vm_page_remque(m);
2497 					vm_page_remove(m);
2498 					KASSERT(m->dirty == 0,
2499 					    ("page %p is dirty", m));
2500 				}
2501 				SLIST_INSERT_HEAD(&free, m, plinks.s.ss);
2502 			} else
2503 				error = EBUSY;
2504 unlock:
2505 			VM_OBJECT_WUNLOCK(object);
2506 		} else {
2507 			MPASS(vm_phys_domain(m) == domain);
2508 			vmd = VM_DOMAIN(domain);
2509 			vm_domain_free_lock(vmd);
2510 			order = m->order;
2511 			if (order < VM_NFREEORDER) {
2512 				/*
2513 				 * The page is enqueued in the physical memory
2514 				 * allocator's free page queues.  Moreover, it
2515 				 * is the first page in a power-of-two-sized
2516 				 * run of contiguous free pages.  Jump ahead
2517 				 * to the last page within that run, and
2518 				 * continue from there.
2519 				 */
2520 				m += (1 << order) - 1;
2521 			}
2522 #if VM_NRESERVLEVEL > 0
2523 			else if (vm_reserv_is_page_free(m))
2524 				order = 0;
2525 #endif
2526 			vm_domain_free_unlock(vmd);
2527 			if (order == VM_NFREEORDER)
2528 				error = EINVAL;
2529 		}
2530 	}
2531 	if (m_mtx != NULL)
2532 		mtx_unlock(m_mtx);
2533 	if ((m = SLIST_FIRST(&free)) != NULL) {
2534 		int cnt;
2535 
2536 		vmd = VM_DOMAIN(domain);
2537 		cnt = 0;
2538 		vm_domain_free_lock(vmd);
2539 		do {
2540 			MPASS(vm_phys_domain(m) == domain);
2541 			SLIST_REMOVE_HEAD(&free, plinks.s.ss);
2542 			vm_page_free_phys(vmd, m);
2543 			cnt++;
2544 		} while ((m = SLIST_FIRST(&free)) != NULL);
2545 		vm_domain_free_unlock(vmd);
2546 		vm_domain_freecnt_inc(vmd, cnt);
2547 	}
2548 	return (error);
2549 }
2550 
2551 #define	NRUNS	16
2552 
2553 CTASSERT(powerof2(NRUNS));
2554 
2555 #define	RUN_INDEX(count)	((count) & (NRUNS - 1))
2556 
2557 #define	MIN_RECLAIM	8
2558 
2559 /*
2560  *	vm_page_reclaim_contig:
2561  *
2562  *	Reclaim allocated, contiguous physical memory satisfying the specified
2563  *	conditions by relocating the virtual pages using that physical memory.
2564  *	Returns true if reclamation is successful and false otherwise.  Since
2565  *	relocation requires the allocation of physical pages, reclamation may
2566  *	fail due to a shortage of free pages.  When reclamation fails, callers
2567  *	are expected to perform vm_wait() before retrying a failed allocation
2568  *	operation, e.g., vm_page_alloc_contig().
2569  *
2570  *	The caller must always specify an allocation class through "req".
2571  *
2572  *	allocation classes:
2573  *	VM_ALLOC_NORMAL		normal process request
2574  *	VM_ALLOC_SYSTEM		system *really* needs a page
2575  *	VM_ALLOC_INTERRUPT	interrupt time request
2576  *
2577  *	The optional allocation flags are ignored.
2578  *
2579  *	"npages" must be greater than zero.  Both "alignment" and "boundary"
2580  *	must be a power of two.
2581  */
2582 bool
2583 vm_page_reclaim_contig_domain(int domain, int req, u_long npages,
2584     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
2585 {
2586 	struct vm_domain *vmd;
2587 	vm_paddr_t curr_low;
2588 	vm_page_t m_run, m_runs[NRUNS];
2589 	u_long count, reclaimed;
2590 	int error, i, options, req_class;
2591 
2592 	KASSERT(npages > 0, ("npages is 0"));
2593 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2594 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2595 	req_class = req & VM_ALLOC_CLASS_MASK;
2596 
2597 	/*
2598 	 * The page daemon is allowed to dig deeper into the free page list.
2599 	 */
2600 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2601 		req_class = VM_ALLOC_SYSTEM;
2602 
2603 	/*
2604 	 * Return if the number of free pages cannot satisfy the requested
2605 	 * allocation.
2606 	 */
2607 	vmd = VM_DOMAIN(domain);
2608 	count = vmd->vmd_free_count;
2609 	if (count < npages + vmd->vmd_free_reserved || (count < npages +
2610 	    vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) ||
2611 	    (count < npages && req_class == VM_ALLOC_INTERRUPT))
2612 		return (false);
2613 
2614 	/*
2615 	 * Scan up to three times, relaxing the restrictions ("options") on
2616 	 * the reclamation of reservations and superpages each time.
2617 	 */
2618 	for (options = VPSC_NORESERV;;) {
2619 		/*
2620 		 * Find the highest runs that satisfy the given constraints
2621 		 * and restrictions, and record them in "m_runs".
2622 		 */
2623 		curr_low = low;
2624 		count = 0;
2625 		for (;;) {
2626 			m_run = vm_phys_scan_contig(domain, npages, curr_low,
2627 			    high, alignment, boundary, options);
2628 			if (m_run == NULL)
2629 				break;
2630 			curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages);
2631 			m_runs[RUN_INDEX(count)] = m_run;
2632 			count++;
2633 		}
2634 
2635 		/*
2636 		 * Reclaim the highest runs in LIFO (descending) order until
2637 		 * the number of reclaimed pages, "reclaimed", is at least
2638 		 * MIN_RECLAIM.  Reset "reclaimed" each time because each
2639 		 * reclamation is idempotent, and runs will (likely) recur
2640 		 * from one scan to the next as restrictions are relaxed.
2641 		 */
2642 		reclaimed = 0;
2643 		for (i = 0; count > 0 && i < NRUNS; i++) {
2644 			count--;
2645 			m_run = m_runs[RUN_INDEX(count)];
2646 			error = vm_page_reclaim_run(req_class, domain, npages,
2647 			    m_run, high);
2648 			if (error == 0) {
2649 				reclaimed += npages;
2650 				if (reclaimed >= MIN_RECLAIM)
2651 					return (true);
2652 			}
2653 		}
2654 
2655 		/*
2656 		 * Either relax the restrictions on the next scan or return if
2657 		 * the last scan had no restrictions.
2658 		 */
2659 		if (options == VPSC_NORESERV)
2660 			options = VPSC_NOSUPER;
2661 		else if (options == VPSC_NOSUPER)
2662 			options = VPSC_ANY;
2663 		else if (options == VPSC_ANY)
2664 			return (reclaimed != 0);
2665 	}
2666 }
2667 
2668 bool
2669 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high,
2670     u_long alignment, vm_paddr_t boundary)
2671 {
2672 	struct vm_domainset_iter di;
2673 	int domain;
2674 	bool ret;
2675 
2676 	vm_domainset_iter_page_init(&di, kernel_object, &domain, &req);
2677 	do {
2678 		ret = vm_page_reclaim_contig_domain(domain, req, npages, low,
2679 		    high, alignment, boundary);
2680 		if (ret)
2681 			break;
2682 	} while (vm_domainset_iter_page(&di, &domain, &req) == 0);
2683 
2684 	return (ret);
2685 }
2686 
2687 /*
2688  * Set the domain in the appropriate page level domainset.
2689  */
2690 void
2691 vm_domain_set(struct vm_domain *vmd)
2692 {
2693 
2694 	mtx_lock(&vm_domainset_lock);
2695 	if (!vmd->vmd_minset && vm_paging_min(vmd)) {
2696 		vmd->vmd_minset = 1;
2697 		DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains);
2698 	}
2699 	if (!vmd->vmd_severeset && vm_paging_severe(vmd)) {
2700 		vmd->vmd_severeset = 1;
2701 		DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains);
2702 	}
2703 	mtx_unlock(&vm_domainset_lock);
2704 }
2705 
2706 /*
2707  * Clear the domain from the appropriate page level domainset.
2708  */
2709 void
2710 vm_domain_clear(struct vm_domain *vmd)
2711 {
2712 
2713 	mtx_lock(&vm_domainset_lock);
2714 	if (vmd->vmd_minset && !vm_paging_min(vmd)) {
2715 		vmd->vmd_minset = 0;
2716 		DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains);
2717 		if (vm_min_waiters != 0) {
2718 			vm_min_waiters = 0;
2719 			wakeup(&vm_min_domains);
2720 		}
2721 	}
2722 	if (vmd->vmd_severeset && !vm_paging_severe(vmd)) {
2723 		vmd->vmd_severeset = 0;
2724 		DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains);
2725 		if (vm_severe_waiters != 0) {
2726 			vm_severe_waiters = 0;
2727 			wakeup(&vm_severe_domains);
2728 		}
2729 	}
2730 
2731 	/*
2732 	 * If pageout daemon needs pages, then tell it that there are
2733 	 * some free.
2734 	 */
2735 	if (vmd->vmd_pageout_pages_needed &&
2736 	    vmd->vmd_free_count >= vmd->vmd_pageout_free_min) {
2737 		wakeup(&vmd->vmd_pageout_pages_needed);
2738 		vmd->vmd_pageout_pages_needed = 0;
2739 	}
2740 
2741 	/* See comments in vm_wait_doms(). */
2742 	if (vm_pageproc_waiters) {
2743 		vm_pageproc_waiters = 0;
2744 		wakeup(&vm_pageproc_waiters);
2745 	}
2746 	mtx_unlock(&vm_domainset_lock);
2747 }
2748 
2749 /*
2750  * Wait for free pages to exceed the min threshold globally.
2751  */
2752 void
2753 vm_wait_min(void)
2754 {
2755 
2756 	mtx_lock(&vm_domainset_lock);
2757 	while (vm_page_count_min()) {
2758 		vm_min_waiters++;
2759 		msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0);
2760 	}
2761 	mtx_unlock(&vm_domainset_lock);
2762 }
2763 
2764 /*
2765  * Wait for free pages to exceed the severe threshold globally.
2766  */
2767 void
2768 vm_wait_severe(void)
2769 {
2770 
2771 	mtx_lock(&vm_domainset_lock);
2772 	while (vm_page_count_severe()) {
2773 		vm_severe_waiters++;
2774 		msleep(&vm_severe_domains, &vm_domainset_lock, PVM,
2775 		    "vmwait", 0);
2776 	}
2777 	mtx_unlock(&vm_domainset_lock);
2778 }
2779 
2780 u_int
2781 vm_wait_count(void)
2782 {
2783 
2784 	return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters);
2785 }
2786 
2787 static void
2788 vm_wait_doms(const domainset_t *wdoms)
2789 {
2790 
2791 	/*
2792 	 * We use racey wakeup synchronization to avoid expensive global
2793 	 * locking for the pageproc when sleeping with a non-specific vm_wait.
2794 	 * To handle this, we only sleep for one tick in this instance.  It
2795 	 * is expected that most allocations for the pageproc will come from
2796 	 * kmem or vm_page_grab* which will use the more specific and
2797 	 * race-free vm_wait_domain().
2798 	 */
2799 	if (curproc == pageproc) {
2800 		mtx_lock(&vm_domainset_lock);
2801 		vm_pageproc_waiters++;
2802 		msleep(&vm_pageproc_waiters, &vm_domainset_lock, PVM | PDROP,
2803 		    "pageprocwait", 1);
2804 	} else {
2805 		/*
2806 		 * XXX Ideally we would wait only until the allocation could
2807 		 * be satisfied.  This condition can cause new allocators to
2808 		 * consume all freed pages while old allocators wait.
2809 		 */
2810 		mtx_lock(&vm_domainset_lock);
2811 		if (DOMAINSET_SUBSET(&vm_min_domains, wdoms)) {
2812 			vm_min_waiters++;
2813 			msleep(&vm_min_domains, &vm_domainset_lock, PVM,
2814 			    "vmwait", 0);
2815 		}
2816 		mtx_unlock(&vm_domainset_lock);
2817 	}
2818 }
2819 
2820 /*
2821  *	vm_wait_domain:
2822  *
2823  *	Sleep until free pages are available for allocation.
2824  *	- Called in various places after failed memory allocations.
2825  */
2826 void
2827 vm_wait_domain(int domain)
2828 {
2829 	struct vm_domain *vmd;
2830 	domainset_t wdom;
2831 
2832 	vmd = VM_DOMAIN(domain);
2833 	vm_domain_free_assert_unlocked(vmd);
2834 
2835 	if (curproc == pageproc) {
2836 		mtx_lock(&vm_domainset_lock);
2837 		if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) {
2838 			vmd->vmd_pageout_pages_needed = 1;
2839 			msleep(&vmd->vmd_pageout_pages_needed,
2840 			    &vm_domainset_lock, PDROP | PSWP, "VMWait", 0);
2841 		} else
2842 			mtx_unlock(&vm_domainset_lock);
2843 	} else {
2844 		if (pageproc == NULL)
2845 			panic("vm_wait in early boot");
2846 		DOMAINSET_ZERO(&wdom);
2847 		DOMAINSET_SET(vmd->vmd_domain, &wdom);
2848 		vm_wait_doms(&wdom);
2849 	}
2850 }
2851 
2852 /*
2853  *	vm_wait:
2854  *
2855  *	Sleep until free pages are available for allocation in the
2856  *	affinity domains of the obj.  If obj is NULL, the domain set
2857  *	for the calling thread is used.
2858  *	Called in various places after failed memory allocations.
2859  */
2860 void
2861 vm_wait(vm_object_t obj)
2862 {
2863 	struct domainset *d;
2864 
2865 	d = NULL;
2866 
2867 	/*
2868 	 * Carefully fetch pointers only once: the struct domainset
2869 	 * itself is ummutable but the pointer might change.
2870 	 */
2871 	if (obj != NULL)
2872 		d = obj->domain.dr_policy;
2873 	if (d == NULL)
2874 		d = curthread->td_domain.dr_policy;
2875 
2876 	vm_wait_doms(&d->ds_mask);
2877 }
2878 
2879 /*
2880  *	vm_domain_alloc_fail:
2881  *
2882  *	Called when a page allocation function fails.  Informs the
2883  *	pagedaemon and performs the requested wait.  Requires the
2884  *	domain_free and object lock on entry.  Returns with the
2885  *	object lock held and free lock released.  Returns an error when
2886  *	retry is necessary.
2887  *
2888  */
2889 static int
2890 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req)
2891 {
2892 
2893 	vm_domain_free_assert_unlocked(vmd);
2894 
2895 	atomic_add_int(&vmd->vmd_pageout_deficit,
2896 	    max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
2897 	if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) {
2898 		if (object != NULL)
2899 			VM_OBJECT_WUNLOCK(object);
2900 		vm_wait_domain(vmd->vmd_domain);
2901 		if (object != NULL)
2902 			VM_OBJECT_WLOCK(object);
2903 		if (req & VM_ALLOC_WAITOK)
2904 			return (EAGAIN);
2905 	}
2906 
2907 	return (0);
2908 }
2909 
2910 /*
2911  *	vm_waitpfault:
2912  *
2913  *	Sleep until free pages are available for allocation.
2914  *	- Called only in vm_fault so that processes page faulting
2915  *	  can be easily tracked.
2916  *	- Sleeps at a lower priority than vm_wait() so that vm_wait()ing
2917  *	  processes will be able to grab memory first.  Do not change
2918  *	  this balance without careful testing first.
2919  */
2920 void
2921 vm_waitpfault(void)
2922 {
2923 
2924 	mtx_lock(&vm_domainset_lock);
2925 	if (vm_page_count_min()) {
2926 		vm_min_waiters++;
2927 		msleep(&vm_min_domains, &vm_domainset_lock, PUSER, "pfault", 0);
2928 	}
2929 	mtx_unlock(&vm_domainset_lock);
2930 }
2931 
2932 struct vm_pagequeue *
2933 vm_page_pagequeue(vm_page_t m)
2934 {
2935 
2936 	return (&vm_pagequeue_domain(m)->vmd_pagequeues[m->queue]);
2937 }
2938 
2939 /*
2940  *	vm_page_dequeue:
2941  *
2942  *	Remove the given page from its current page queue.
2943  *
2944  *	The page must be locked.
2945  */
2946 void
2947 vm_page_dequeue(vm_page_t m)
2948 {
2949 	struct vm_pagequeue *pq;
2950 
2951 	vm_page_assert_locked(m);
2952 	KASSERT(m->queue < PQ_COUNT, ("vm_page_dequeue: page %p is not queued",
2953 	    m));
2954 	pq = vm_page_pagequeue(m);
2955 	vm_pagequeue_lock(pq);
2956 	m->queue = PQ_NONE;
2957 	TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2958 	vm_pagequeue_cnt_dec(pq);
2959 	vm_pagequeue_unlock(pq);
2960 }
2961 
2962 /*
2963  *	vm_page_dequeue_locked:
2964  *
2965  *	Remove the given page from its current page queue.
2966  *
2967  *	The page and page queue must be locked.
2968  */
2969 void
2970 vm_page_dequeue_locked(vm_page_t m)
2971 {
2972 	struct vm_pagequeue *pq;
2973 
2974 	vm_page_lock_assert(m, MA_OWNED);
2975 	pq = vm_page_pagequeue(m);
2976 	vm_pagequeue_assert_locked(pq);
2977 	m->queue = PQ_NONE;
2978 	TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2979 	vm_pagequeue_cnt_dec(pq);
2980 }
2981 
2982 /*
2983  *	vm_page_enqueue:
2984  *
2985  *	Add the given page to the specified page queue.
2986  *
2987  *	The page must be locked.
2988  */
2989 static void
2990 vm_page_enqueue(uint8_t queue, vm_page_t m)
2991 {
2992 	struct vm_pagequeue *pq;
2993 
2994 	vm_page_lock_assert(m, MA_OWNED);
2995 	KASSERT(queue < PQ_COUNT,
2996 	    ("vm_page_enqueue: invalid queue %u request for page %p",
2997 	    queue, m));
2998 	pq = &vm_pagequeue_domain(m)->vmd_pagequeues[queue];
2999 	vm_pagequeue_lock(pq);
3000 	m->queue = queue;
3001 	TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3002 	vm_pagequeue_cnt_inc(pq);
3003 	vm_pagequeue_unlock(pq);
3004 }
3005 
3006 /*
3007  *	vm_page_requeue:
3008  *
3009  *	Move the given page to the tail of its current page queue.
3010  *
3011  *	The page must be locked.
3012  */
3013 void
3014 vm_page_requeue(vm_page_t m)
3015 {
3016 	struct vm_pagequeue *pq;
3017 
3018 	vm_page_lock_assert(m, MA_OWNED);
3019 	KASSERT(m->queue != PQ_NONE,
3020 	    ("vm_page_requeue: page %p is not queued", m));
3021 	pq = vm_page_pagequeue(m);
3022 	vm_pagequeue_lock(pq);
3023 	TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3024 	TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3025 	vm_pagequeue_unlock(pq);
3026 }
3027 
3028 /*
3029  *	vm_page_requeue_locked:
3030  *
3031  *	Move the given page to the tail of its current page queue.
3032  *
3033  *	The page queue must be locked.
3034  */
3035 void
3036 vm_page_requeue_locked(vm_page_t m)
3037 {
3038 	struct vm_pagequeue *pq;
3039 
3040 	KASSERT(m->queue != PQ_NONE,
3041 	    ("vm_page_requeue_locked: page %p is not queued", m));
3042 	pq = vm_page_pagequeue(m);
3043 	vm_pagequeue_assert_locked(pq);
3044 	TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3045 	TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3046 }
3047 
3048 /*
3049  *	vm_page_activate:
3050  *
3051  *	Put the specified page on the active list (if appropriate).
3052  *	Ensure that act_count is at least ACT_INIT but do not otherwise
3053  *	mess with it.
3054  *
3055  *	The page must be locked.
3056  */
3057 void
3058 vm_page_activate(vm_page_t m)
3059 {
3060 	int queue;
3061 
3062 	vm_page_lock_assert(m, MA_OWNED);
3063 	if ((queue = m->queue) != PQ_ACTIVE) {
3064 		if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
3065 			if (m->act_count < ACT_INIT)
3066 				m->act_count = ACT_INIT;
3067 			if (queue != PQ_NONE)
3068 				vm_page_dequeue(m);
3069 			vm_page_enqueue(PQ_ACTIVE, m);
3070 		}
3071 	} else {
3072 		if (m->act_count < ACT_INIT)
3073 			m->act_count = ACT_INIT;
3074 	}
3075 }
3076 
3077 /*
3078  *	vm_page_free_prep:
3079  *
3080  *	Prepares the given page to be put on the free list,
3081  *	disassociating it from any VM object. The caller may return
3082  *	the page to the free list only if this function returns true.
3083  *
3084  *	The object must be locked.  The page must be locked if it is
3085  *	managed.  For a queued managed page, the pagequeue_locked
3086  *	argument specifies whether the page queue is already locked.
3087  */
3088 bool
3089 vm_page_free_prep(vm_page_t m, bool pagequeue_locked)
3090 {
3091 
3092 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP)
3093 	if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) {
3094 		uint64_t *p;
3095 		int i;
3096 		p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
3097 		for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++)
3098 			KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx",
3099 			    m, i, (uintmax_t)*p));
3100 	}
3101 #endif
3102 	if ((m->oflags & VPO_UNMANAGED) == 0) {
3103 		vm_page_lock_assert(m, MA_OWNED);
3104 		KASSERT(!pmap_page_is_mapped(m),
3105 		    ("vm_page_free_toq: freeing mapped page %p", m));
3106 	} else
3107 		KASSERT(m->queue == PQ_NONE,
3108 		    ("vm_page_free_toq: unmanaged page %p is queued", m));
3109 	VM_CNT_INC(v_tfree);
3110 
3111 	if (vm_page_sbusied(m))
3112 		panic("vm_page_free: freeing busy page %p", m);
3113 
3114 	vm_page_remove(m);
3115 
3116 	/*
3117 	 * If fictitious remove object association and
3118 	 * return.
3119 	 */
3120 	if ((m->flags & PG_FICTITIOUS) != 0) {
3121 		KASSERT(m->wire_count == 1,
3122 		    ("fictitious page %p is not wired", m));
3123 		KASSERT(m->queue == PQ_NONE,
3124 		    ("fictitious page %p is queued", m));
3125 		return (false);
3126 	}
3127 
3128 	if (m->queue != PQ_NONE) {
3129 		if (pagequeue_locked)
3130 			vm_page_dequeue_locked(m);
3131 		else
3132 			vm_page_dequeue(m);
3133 	}
3134 	m->valid = 0;
3135 	vm_page_undirty(m);
3136 
3137 	if (m->wire_count != 0)
3138 		panic("vm_page_free: freeing wired page %p", m);
3139 	if (m->hold_count != 0) {
3140 		m->flags &= ~PG_ZERO;
3141 		KASSERT((m->flags & PG_UNHOLDFREE) == 0,
3142 		    ("vm_page_free: freeing PG_UNHOLDFREE page %p", m));
3143 		m->flags |= PG_UNHOLDFREE;
3144 		return (false);
3145 	}
3146 
3147 	/*
3148 	 * Restore the default memory attribute to the page.
3149 	 */
3150 	if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
3151 		pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
3152 
3153 	return (true);
3154 }
3155 
3156 /*
3157  * Insert the page into the physical memory allocator's free page
3158  * queues.  This is the last step to free a page.  The caller is
3159  * responsible for adjusting the free page count.
3160  */
3161 static void
3162 vm_page_free_phys(struct vm_domain *vmd, vm_page_t m)
3163 {
3164 
3165 	vm_domain_free_assert_locked(vmd);
3166 
3167 #if VM_NRESERVLEVEL > 0
3168 	if (!vm_reserv_free_page(m))
3169 #endif
3170 		vm_phys_free_pages(m, 0);
3171 }
3172 
3173 void
3174 vm_page_free_phys_pglist(struct pglist *tq)
3175 {
3176 	struct vm_domain *vmd;
3177 	vm_page_t m;
3178 	int cnt;
3179 
3180 	if (TAILQ_EMPTY(tq))
3181 		return;
3182 	vmd = NULL;
3183 	cnt = 0;
3184 	TAILQ_FOREACH(m, tq, listq) {
3185 		if (vmd != vm_pagequeue_domain(m)) {
3186 			if (vmd != NULL) {
3187 				vm_domain_free_unlock(vmd);
3188 				vm_domain_freecnt_inc(vmd, cnt);
3189 				cnt = 0;
3190 			}
3191 			vmd = vm_pagequeue_domain(m);
3192 			vm_domain_free_lock(vmd);
3193 		}
3194 		vm_page_free_phys(vmd, m);
3195 		cnt++;
3196 	}
3197 	if (vmd != NULL) {
3198 		vm_domain_free_unlock(vmd);
3199 		vm_domain_freecnt_inc(vmd, cnt);
3200 	}
3201 }
3202 
3203 /*
3204  *	vm_page_free_toq:
3205  *
3206  *	Returns the given page to the free list, disassociating it
3207  *	from any VM object.
3208  *
3209  *	The object must be locked.  The page must be locked if it is
3210  *	managed.
3211  */
3212 void
3213 vm_page_free_toq(vm_page_t m)
3214 {
3215 	struct vm_domain *vmd;
3216 
3217 	if (!vm_page_free_prep(m, false))
3218 		return;
3219 	vmd = vm_pagequeue_domain(m);
3220 	vm_domain_free_lock(vmd);
3221 	vm_page_free_phys(vmd, m);
3222 	vm_domain_free_unlock(vmd);
3223 	vm_domain_freecnt_inc(vmd, 1);
3224 }
3225 
3226 /*
3227  *	vm_page_free_pages_toq:
3228  *
3229  *	Returns a list of pages to the free list, disassociating it
3230  *	from any VM object.  In other words, this is equivalent to
3231  *	calling vm_page_free_toq() for each page of a list of VM objects.
3232  *
3233  *	The objects must be locked.  The pages must be locked if it is
3234  *	managed.
3235  */
3236 void
3237 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count)
3238 {
3239 	vm_page_t m;
3240 	struct pglist pgl;
3241 	int count;
3242 
3243 	if (SLIST_EMPTY(free))
3244 		return;
3245 
3246 	count = 0;
3247 	TAILQ_INIT(&pgl);
3248 	while ((m = SLIST_FIRST(free)) != NULL) {
3249 		count++;
3250 		SLIST_REMOVE_HEAD(free, plinks.s.ss);
3251 		if (vm_page_free_prep(m, false))
3252 			TAILQ_INSERT_TAIL(&pgl, m, listq);
3253 	}
3254 
3255 	vm_page_free_phys_pglist(&pgl);
3256 
3257 	if (update_wire_count)
3258 		vm_wire_sub(count);
3259 }
3260 
3261 /*
3262  *	vm_page_wire:
3263  *
3264  * Mark this page as wired down.  If the page is fictitious, then
3265  * its wire count must remain one.
3266  *
3267  * The page must be locked.
3268  */
3269 void
3270 vm_page_wire(vm_page_t m)
3271 {
3272 
3273 	vm_page_assert_locked(m);
3274 	if ((m->flags & PG_FICTITIOUS) != 0) {
3275 		KASSERT(m->wire_count == 1,
3276 		    ("vm_page_wire: fictitious page %p's wire count isn't one",
3277 		    m));
3278 		return;
3279 	}
3280 	if (m->wire_count == 0) {
3281 		KASSERT((m->oflags & VPO_UNMANAGED) == 0 ||
3282 		    m->queue == PQ_NONE,
3283 		    ("vm_page_wire: unmanaged page %p is queued", m));
3284 		vm_wire_add(1);
3285 	}
3286 	m->wire_count++;
3287 	KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m));
3288 }
3289 
3290 /*
3291  * vm_page_unwire:
3292  *
3293  * Release one wiring of the specified page, potentially allowing it to be
3294  * paged out.  Returns TRUE if the number of wirings transitions to zero and
3295  * FALSE otherwise.
3296  *
3297  * Only managed pages belonging to an object can be paged out.  If the number
3298  * of wirings transitions to zero and the page is eligible for page out, then
3299  * the page is added to the specified paging queue (unless PQ_NONE is
3300  * specified, in which case the page is dequeued if it belongs to a paging
3301  * queue).
3302  *
3303  * If a page is fictitious, then its wire count must always be one.
3304  *
3305  * A managed page must be locked.
3306  */
3307 bool
3308 vm_page_unwire(vm_page_t m, uint8_t queue)
3309 {
3310 	bool unwired;
3311 
3312 	KASSERT(queue < PQ_COUNT || queue == PQ_NONE,
3313 	    ("vm_page_unwire: invalid queue %u request for page %p",
3314 	    queue, m));
3315 
3316 	unwired = vm_page_unwire_noq(m);
3317 	if (unwired && (m->oflags & VPO_UNMANAGED) == 0 && m->object != NULL) {
3318 		if (m->queue == queue) {
3319 			if (queue == PQ_ACTIVE)
3320 				vm_page_reference(m);
3321 			else if (queue != PQ_NONE)
3322 				vm_page_requeue(m);
3323 		} else {
3324 			vm_page_remque(m);
3325 			if (queue != PQ_NONE) {
3326 				vm_page_enqueue(queue, m);
3327 				if (queue == PQ_ACTIVE)
3328 					/* Initialize act_count. */
3329 					vm_page_activate(m);
3330 			}
3331 		}
3332 	}
3333 	return (unwired);
3334 }
3335 
3336 /*
3337  *
3338  * vm_page_unwire_noq:
3339  *
3340  * Unwire a page without (re-)inserting it into a page queue.  It is up
3341  * to the caller to enqueue, requeue, or free the page as appropriate.
3342  * In most cases, vm_page_unwire() should be used instead.
3343  */
3344 bool
3345 vm_page_unwire_noq(vm_page_t m)
3346 {
3347 
3348 	if ((m->oflags & VPO_UNMANAGED) == 0)
3349 		vm_page_assert_locked(m);
3350 	if ((m->flags & PG_FICTITIOUS) != 0) {
3351 		KASSERT(m->wire_count == 1,
3352 	    ("vm_page_unwire: fictitious page %p's wire count isn't one", m));
3353 		return (false);
3354 	}
3355 	if (m->wire_count == 0)
3356 		panic("vm_page_unwire: page %p's wire count is zero", m);
3357 	m->wire_count--;
3358 	if (m->wire_count == 0) {
3359 		vm_wire_sub(1);
3360 		return (true);
3361 	} else
3362 		return (false);
3363 }
3364 
3365 /*
3366  * Move the specified page to the inactive queue, or requeue the page if it is
3367  * already in the inactive queue.
3368  *
3369  * Normally, "noreuse" is FALSE, resulting in LRU ordering of the inactive
3370  * queue.  However, setting "noreuse" to TRUE will accelerate the specified
3371  * page's reclamation, but it will not unmap the page from any address space.
3372  * This is implemented by inserting the page near the head of the inactive
3373  * queue, using a marker page to guide FIFO insertion ordering.
3374  *
3375  * The page must be locked.
3376  */
3377 static inline void
3378 _vm_page_deactivate(vm_page_t m, boolean_t noreuse)
3379 {
3380 	struct vm_pagequeue *pq;
3381 	int queue;
3382 
3383 	vm_page_assert_locked(m);
3384 
3385 	if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
3386 		pq = &vm_pagequeue_domain(m)->vmd_pagequeues[PQ_INACTIVE];
3387 		/* Avoid multiple acquisitions of the inactive queue lock. */
3388 		queue = m->queue;
3389 		if (queue == PQ_INACTIVE) {
3390 			vm_pagequeue_lock(pq);
3391 			vm_page_dequeue_locked(m);
3392 		} else {
3393 			if (queue != PQ_NONE)
3394 				vm_page_dequeue(m);
3395 			vm_pagequeue_lock(pq);
3396 		}
3397 		m->queue = PQ_INACTIVE;
3398 		if (noreuse)
3399 			TAILQ_INSERT_BEFORE(
3400 			    &vm_pagequeue_domain(m)->vmd_inacthead, m,
3401 			    plinks.q);
3402 		else
3403 			TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3404 		vm_pagequeue_cnt_inc(pq);
3405 		vm_pagequeue_unlock(pq);
3406 	}
3407 }
3408 
3409 /*
3410  * Move the specified page to the inactive queue, or requeue the page if it is
3411  * already in the inactive queue.
3412  *
3413  * The page must be locked.
3414  */
3415 void
3416 vm_page_deactivate(vm_page_t m)
3417 {
3418 
3419 	_vm_page_deactivate(m, FALSE);
3420 }
3421 
3422 /*
3423  * Move the specified page to the inactive queue with the expectation
3424  * that it is unlikely to be reused.
3425  *
3426  * The page must be locked.
3427  */
3428 void
3429 vm_page_deactivate_noreuse(vm_page_t m)
3430 {
3431 
3432 	_vm_page_deactivate(m, TRUE);
3433 }
3434 
3435 /*
3436  * vm_page_launder
3437  *
3438  * 	Put a page in the laundry, or requeue it if it is already there.
3439  */
3440 void
3441 vm_page_launder(vm_page_t m)
3442 {
3443 
3444 	vm_page_assert_locked(m);
3445 	if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
3446 		if (m->queue == PQ_LAUNDRY)
3447 			vm_page_requeue(m);
3448 		else {
3449 			vm_page_remque(m);
3450 			vm_page_enqueue(PQ_LAUNDRY, m);
3451 		}
3452 	}
3453 }
3454 
3455 /*
3456  * vm_page_unswappable
3457  *
3458  *	Put a page in the PQ_UNSWAPPABLE holding queue.
3459  */
3460 void
3461 vm_page_unswappable(vm_page_t m)
3462 {
3463 
3464 	vm_page_assert_locked(m);
3465 	KASSERT(m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0,
3466 	    ("page %p already unswappable", m));
3467 	if (m->queue != PQ_NONE)
3468 		vm_page_dequeue(m);
3469 	vm_page_enqueue(PQ_UNSWAPPABLE, m);
3470 }
3471 
3472 /*
3473  * Attempt to free the page.  If it cannot be freed, do nothing.  Returns true
3474  * if the page is freed and false otherwise.
3475  *
3476  * The page must be managed.  The page and its containing object must be
3477  * locked.
3478  */
3479 bool
3480 vm_page_try_to_free(vm_page_t m)
3481 {
3482 
3483 	vm_page_assert_locked(m);
3484 	VM_OBJECT_ASSERT_WLOCKED(m->object);
3485 	KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("page %p is unmanaged", m));
3486 	if (m->dirty != 0 || vm_page_held(m) || vm_page_busied(m))
3487 		return (false);
3488 	if (m->object->ref_count != 0) {
3489 		pmap_remove_all(m);
3490 		if (m->dirty != 0)
3491 			return (false);
3492 	}
3493 	vm_page_free(m);
3494 	return (true);
3495 }
3496 
3497 /*
3498  * vm_page_advise
3499  *
3500  * 	Apply the specified advice to the given page.
3501  *
3502  *	The object and page must be locked.
3503  */
3504 void
3505 vm_page_advise(vm_page_t m, int advice)
3506 {
3507 
3508 	vm_page_assert_locked(m);
3509 	VM_OBJECT_ASSERT_WLOCKED(m->object);
3510 	if (advice == MADV_FREE)
3511 		/*
3512 		 * Mark the page clean.  This will allow the page to be freed
3513 		 * without first paging it out.  MADV_FREE pages are often
3514 		 * quickly reused by malloc(3), so we do not do anything that
3515 		 * would result in a page fault on a later access.
3516 		 */
3517 		vm_page_undirty(m);
3518 	else if (advice != MADV_DONTNEED) {
3519 		if (advice == MADV_WILLNEED)
3520 			vm_page_activate(m);
3521 		return;
3522 	}
3523 
3524 	/*
3525 	 * Clear any references to the page.  Otherwise, the page daemon will
3526 	 * immediately reactivate the page.
3527 	 */
3528 	vm_page_aflag_clear(m, PGA_REFERENCED);
3529 
3530 	if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
3531 		vm_page_dirty(m);
3532 
3533 	/*
3534 	 * Place clean pages near the head of the inactive queue rather than
3535 	 * the tail, thus defeating the queue's LRU operation and ensuring that
3536 	 * the page will be reused quickly.  Dirty pages not already in the
3537 	 * laundry are moved there.
3538 	 */
3539 	if (m->dirty == 0)
3540 		vm_page_deactivate_noreuse(m);
3541 	else if (!vm_page_in_laundry(m))
3542 		vm_page_launder(m);
3543 }
3544 
3545 /*
3546  * Grab a page, waiting until we are waken up due to the page
3547  * changing state.  We keep on waiting, if the page continues
3548  * to be in the object.  If the page doesn't exist, first allocate it
3549  * and then conditionally zero it.
3550  *
3551  * This routine may sleep.
3552  *
3553  * The object must be locked on entry.  The lock will, however, be released
3554  * and reacquired if the routine sleeps.
3555  */
3556 vm_page_t
3557 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
3558 {
3559 	vm_page_t m;
3560 	int sleep;
3561 	int pflags;
3562 
3563 	VM_OBJECT_ASSERT_WLOCKED(object);
3564 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
3565 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
3566 	    ("vm_page_grab: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
3567 	pflags = allocflags &
3568 	    ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL);
3569 	if ((allocflags & VM_ALLOC_NOWAIT) == 0)
3570 		pflags |= VM_ALLOC_WAITFAIL;
3571 retrylookup:
3572 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
3573 		sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ?
3574 		    vm_page_xbusied(m) : vm_page_busied(m);
3575 		if (sleep) {
3576 			if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3577 				return (NULL);
3578 			/*
3579 			 * Reference the page before unlocking and
3580 			 * sleeping so that the page daemon is less
3581 			 * likely to reclaim it.
3582 			 */
3583 			vm_page_aflag_set(m, PGA_REFERENCED);
3584 			vm_page_lock(m);
3585 			VM_OBJECT_WUNLOCK(object);
3586 			vm_page_busy_sleep(m, "pgrbwt", (allocflags &
3587 			    VM_ALLOC_IGN_SBUSY) != 0);
3588 			VM_OBJECT_WLOCK(object);
3589 			goto retrylookup;
3590 		} else {
3591 			if ((allocflags & VM_ALLOC_WIRED) != 0) {
3592 				vm_page_lock(m);
3593 				vm_page_wire(m);
3594 				vm_page_unlock(m);
3595 			}
3596 			if ((allocflags &
3597 			    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
3598 				vm_page_xbusy(m);
3599 			if ((allocflags & VM_ALLOC_SBUSY) != 0)
3600 				vm_page_sbusy(m);
3601 			return (m);
3602 		}
3603 	}
3604 	m = vm_page_alloc(object, pindex, pflags);
3605 	if (m == NULL) {
3606 		if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3607 			return (NULL);
3608 		goto retrylookup;
3609 	}
3610 	if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
3611 		pmap_zero_page(m);
3612 	return (m);
3613 }
3614 
3615 /*
3616  * Return the specified range of pages from the given object.  For each
3617  * page offset within the range, if a page already exists within the object
3618  * at that offset and it is busy, then wait for it to change state.  If,
3619  * instead, the page doesn't exist, then allocate it.
3620  *
3621  * The caller must always specify an allocation class.
3622  *
3623  * allocation classes:
3624  *	VM_ALLOC_NORMAL		normal process request
3625  *	VM_ALLOC_SYSTEM		system *really* needs the pages
3626  *
3627  * The caller must always specify that the pages are to be busied and/or
3628  * wired.
3629  *
3630  * optional allocation flags:
3631  *	VM_ALLOC_IGN_SBUSY	do not sleep on soft busy pages
3632  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
3633  *	VM_ALLOC_NOWAIT		do not sleep
3634  *	VM_ALLOC_SBUSY		set page to sbusy state
3635  *	VM_ALLOC_WIRED		wire the pages
3636  *	VM_ALLOC_ZERO		zero and validate any invalid pages
3637  *
3638  * If VM_ALLOC_NOWAIT is not specified, this routine may sleep.  Otherwise, it
3639  * may return a partial prefix of the requested range.
3640  */
3641 int
3642 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags,
3643     vm_page_t *ma, int count)
3644 {
3645 	vm_page_t m, mpred;
3646 	int pflags;
3647 	int i;
3648 	bool sleep;
3649 
3650 	VM_OBJECT_ASSERT_WLOCKED(object);
3651 	KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0,
3652 	    ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed"));
3653 	KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 ||
3654 	    (allocflags & VM_ALLOC_WIRED) != 0,
3655 	    ("vm_page_grab_pages: the pages must be busied or wired"));
3656 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
3657 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
3658 	    ("vm_page_grab_pages: VM_ALLOC_SBUSY/IGN_SBUSY mismatch"));
3659 	if (count == 0)
3660 		return (0);
3661 	pflags = allocflags & ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK |
3662 	    VM_ALLOC_WAITFAIL | VM_ALLOC_IGN_SBUSY);
3663 	if ((allocflags & VM_ALLOC_NOWAIT) == 0)
3664 		pflags |= VM_ALLOC_WAITFAIL;
3665 	i = 0;
3666 retrylookup:
3667 	m = vm_radix_lookup_le(&object->rtree, pindex + i);
3668 	if (m == NULL || m->pindex != pindex + i) {
3669 		mpred = m;
3670 		m = NULL;
3671 	} else
3672 		mpred = TAILQ_PREV(m, pglist, listq);
3673 	for (; i < count; i++) {
3674 		if (m != NULL) {
3675 			sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ?
3676 			    vm_page_xbusied(m) : vm_page_busied(m);
3677 			if (sleep) {
3678 				if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3679 					break;
3680 				/*
3681 				 * Reference the page before unlocking and
3682 				 * sleeping so that the page daemon is less
3683 				 * likely to reclaim it.
3684 				 */
3685 				vm_page_aflag_set(m, PGA_REFERENCED);
3686 				vm_page_lock(m);
3687 				VM_OBJECT_WUNLOCK(object);
3688 				vm_page_busy_sleep(m, "grbmaw", (allocflags &
3689 				    VM_ALLOC_IGN_SBUSY) != 0);
3690 				VM_OBJECT_WLOCK(object);
3691 				goto retrylookup;
3692 			}
3693 			if ((allocflags & VM_ALLOC_WIRED) != 0) {
3694 				vm_page_lock(m);
3695 				vm_page_wire(m);
3696 				vm_page_unlock(m);
3697 			}
3698 			if ((allocflags & (VM_ALLOC_NOBUSY |
3699 			    VM_ALLOC_SBUSY)) == 0)
3700 				vm_page_xbusy(m);
3701 			if ((allocflags & VM_ALLOC_SBUSY) != 0)
3702 				vm_page_sbusy(m);
3703 		} else {
3704 			m = vm_page_alloc_after(object, pindex + i,
3705 			    pflags | VM_ALLOC_COUNT(count - i), mpred);
3706 			if (m == NULL) {
3707 				if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3708 					break;
3709 				goto retrylookup;
3710 			}
3711 		}
3712 		if (m->valid == 0 && (allocflags & VM_ALLOC_ZERO) != 0) {
3713 			if ((m->flags & PG_ZERO) == 0)
3714 				pmap_zero_page(m);
3715 			m->valid = VM_PAGE_BITS_ALL;
3716 		}
3717 		ma[i] = mpred = m;
3718 		m = vm_page_next(m);
3719 	}
3720 	return (i);
3721 }
3722 
3723 /*
3724  * Mapping function for valid or dirty bits in a page.
3725  *
3726  * Inputs are required to range within a page.
3727  */
3728 vm_page_bits_t
3729 vm_page_bits(int base, int size)
3730 {
3731 	int first_bit;
3732 	int last_bit;
3733 
3734 	KASSERT(
3735 	    base + size <= PAGE_SIZE,
3736 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
3737 	);
3738 
3739 	if (size == 0)		/* handle degenerate case */
3740 		return (0);
3741 
3742 	first_bit = base >> DEV_BSHIFT;
3743 	last_bit = (base + size - 1) >> DEV_BSHIFT;
3744 
3745 	return (((vm_page_bits_t)2 << last_bit) -
3746 	    ((vm_page_bits_t)1 << first_bit));
3747 }
3748 
3749 /*
3750  *	vm_page_set_valid_range:
3751  *
3752  *	Sets portions of a page valid.  The arguments are expected
3753  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
3754  *	of any partial chunks touched by the range.  The invalid portion of
3755  *	such chunks will be zeroed.
3756  *
3757  *	(base + size) must be less then or equal to PAGE_SIZE.
3758  */
3759 void
3760 vm_page_set_valid_range(vm_page_t m, int base, int size)
3761 {
3762 	int endoff, frag;
3763 
3764 	VM_OBJECT_ASSERT_WLOCKED(m->object);
3765 	if (size == 0)	/* handle degenerate case */
3766 		return;
3767 
3768 	/*
3769 	 * If the base is not DEV_BSIZE aligned and the valid
3770 	 * bit is clear, we have to zero out a portion of the
3771 	 * first block.
3772 	 */
3773 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
3774 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
3775 		pmap_zero_page_area(m, frag, base - frag);
3776 
3777 	/*
3778 	 * If the ending offset is not DEV_BSIZE aligned and the
3779 	 * valid bit is clear, we have to zero out a portion of
3780 	 * the last block.
3781 	 */
3782 	endoff = base + size;
3783 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
3784 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
3785 		pmap_zero_page_area(m, endoff,
3786 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
3787 
3788 	/*
3789 	 * Assert that no previously invalid block that is now being validated
3790 	 * is already dirty.
3791 	 */
3792 	KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
3793 	    ("vm_page_set_valid_range: page %p is dirty", m));
3794 
3795 	/*
3796 	 * Set valid bits inclusive of any overlap.
3797 	 */
3798 	m->valid |= vm_page_bits(base, size);
3799 }
3800 
3801 /*
3802  * Clear the given bits from the specified page's dirty field.
3803  */
3804 static __inline void
3805 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
3806 {
3807 	uintptr_t addr;
3808 #if PAGE_SIZE < 16384
3809 	int shift;
3810 #endif
3811 
3812 	/*
3813 	 * If the object is locked and the page is neither exclusive busy nor
3814 	 * write mapped, then the page's dirty field cannot possibly be
3815 	 * set by a concurrent pmap operation.
3816 	 */
3817 	VM_OBJECT_ASSERT_WLOCKED(m->object);
3818 	if (!vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
3819 		m->dirty &= ~pagebits;
3820 	else {
3821 		/*
3822 		 * The pmap layer can call vm_page_dirty() without
3823 		 * holding a distinguished lock.  The combination of
3824 		 * the object's lock and an atomic operation suffice
3825 		 * to guarantee consistency of the page dirty field.
3826 		 *
3827 		 * For PAGE_SIZE == 32768 case, compiler already
3828 		 * properly aligns the dirty field, so no forcible
3829 		 * alignment is needed. Only require existence of
3830 		 * atomic_clear_64 when page size is 32768.
3831 		 */
3832 		addr = (uintptr_t)&m->dirty;
3833 #if PAGE_SIZE == 32768
3834 		atomic_clear_64((uint64_t *)addr, pagebits);
3835 #elif PAGE_SIZE == 16384
3836 		atomic_clear_32((uint32_t *)addr, pagebits);
3837 #else		/* PAGE_SIZE <= 8192 */
3838 		/*
3839 		 * Use a trick to perform a 32-bit atomic on the
3840 		 * containing aligned word, to not depend on the existence
3841 		 * of atomic_clear_{8, 16}.
3842 		 */
3843 		shift = addr & (sizeof(uint32_t) - 1);
3844 #if BYTE_ORDER == BIG_ENDIAN
3845 		shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY;
3846 #else
3847 		shift *= NBBY;
3848 #endif
3849 		addr &= ~(sizeof(uint32_t) - 1);
3850 		atomic_clear_32((uint32_t *)addr, pagebits << shift);
3851 #endif		/* PAGE_SIZE */
3852 	}
3853 }
3854 
3855 /*
3856  *	vm_page_set_validclean:
3857  *
3858  *	Sets portions of a page valid and clean.  The arguments are expected
3859  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
3860  *	of any partial chunks touched by the range.  The invalid portion of
3861  *	such chunks will be zero'd.
3862  *
3863  *	(base + size) must be less then or equal to PAGE_SIZE.
3864  */
3865 void
3866 vm_page_set_validclean(vm_page_t m, int base, int size)
3867 {
3868 	vm_page_bits_t oldvalid, pagebits;
3869 	int endoff, frag;
3870 
3871 	VM_OBJECT_ASSERT_WLOCKED(m->object);
3872 	if (size == 0)	/* handle degenerate case */
3873 		return;
3874 
3875 	/*
3876 	 * If the base is not DEV_BSIZE aligned and the valid
3877 	 * bit is clear, we have to zero out a portion of the
3878 	 * first block.
3879 	 */
3880 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
3881 	    (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
3882 		pmap_zero_page_area(m, frag, base - frag);
3883 
3884 	/*
3885 	 * If the ending offset is not DEV_BSIZE aligned and the
3886 	 * valid bit is clear, we have to zero out a portion of
3887 	 * the last block.
3888 	 */
3889 	endoff = base + size;
3890 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
3891 	    (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
3892 		pmap_zero_page_area(m, endoff,
3893 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
3894 
3895 	/*
3896 	 * Set valid, clear dirty bits.  If validating the entire
3897 	 * page we can safely clear the pmap modify bit.  We also
3898 	 * use this opportunity to clear the VPO_NOSYNC flag.  If a process
3899 	 * takes a write fault on a MAP_NOSYNC memory area the flag will
3900 	 * be set again.
3901 	 *
3902 	 * We set valid bits inclusive of any overlap, but we can only
3903 	 * clear dirty bits for DEV_BSIZE chunks that are fully within
3904 	 * the range.
3905 	 */
3906 	oldvalid = m->valid;
3907 	pagebits = vm_page_bits(base, size);
3908 	m->valid |= pagebits;
3909 #if 0	/* NOT YET */
3910 	if ((frag = base & (DEV_BSIZE - 1)) != 0) {
3911 		frag = DEV_BSIZE - frag;
3912 		base += frag;
3913 		size -= frag;
3914 		if (size < 0)
3915 			size = 0;
3916 	}
3917 	pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
3918 #endif
3919 	if (base == 0 && size == PAGE_SIZE) {
3920 		/*
3921 		 * The page can only be modified within the pmap if it is
3922 		 * mapped, and it can only be mapped if it was previously
3923 		 * fully valid.
3924 		 */
3925 		if (oldvalid == VM_PAGE_BITS_ALL)
3926 			/*
3927 			 * Perform the pmap_clear_modify() first.  Otherwise,
3928 			 * a concurrent pmap operation, such as
3929 			 * pmap_protect(), could clear a modification in the
3930 			 * pmap and set the dirty field on the page before
3931 			 * pmap_clear_modify() had begun and after the dirty
3932 			 * field was cleared here.
3933 			 */
3934 			pmap_clear_modify(m);
3935 		m->dirty = 0;
3936 		m->oflags &= ~VPO_NOSYNC;
3937 	} else if (oldvalid != VM_PAGE_BITS_ALL)
3938 		m->dirty &= ~pagebits;
3939 	else
3940 		vm_page_clear_dirty_mask(m, pagebits);
3941 }
3942 
3943 void
3944 vm_page_clear_dirty(vm_page_t m, int base, int size)
3945 {
3946 
3947 	vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
3948 }
3949 
3950 /*
3951  *	vm_page_set_invalid:
3952  *
3953  *	Invalidates DEV_BSIZE'd chunks within a page.  Both the
3954  *	valid and dirty bits for the effected areas are cleared.
3955  */
3956 void
3957 vm_page_set_invalid(vm_page_t m, int base, int size)
3958 {
3959 	vm_page_bits_t bits;
3960 	vm_object_t object;
3961 
3962 	object = m->object;
3963 	VM_OBJECT_ASSERT_WLOCKED(object);
3964 	if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
3965 	    size >= object->un_pager.vnp.vnp_size)
3966 		bits = VM_PAGE_BITS_ALL;
3967 	else
3968 		bits = vm_page_bits(base, size);
3969 	if (object->ref_count != 0 && m->valid == VM_PAGE_BITS_ALL &&
3970 	    bits != 0)
3971 		pmap_remove_all(m);
3972 	KASSERT((bits == 0 && m->valid == VM_PAGE_BITS_ALL) ||
3973 	    !pmap_page_is_mapped(m),
3974 	    ("vm_page_set_invalid: page %p is mapped", m));
3975 	m->valid &= ~bits;
3976 	m->dirty &= ~bits;
3977 }
3978 
3979 /*
3980  * vm_page_zero_invalid()
3981  *
3982  *	The kernel assumes that the invalid portions of a page contain
3983  *	garbage, but such pages can be mapped into memory by user code.
3984  *	When this occurs, we must zero out the non-valid portions of the
3985  *	page so user code sees what it expects.
3986  *
3987  *	Pages are most often semi-valid when the end of a file is mapped
3988  *	into memory and the file's size is not page aligned.
3989  */
3990 void
3991 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
3992 {
3993 	int b;
3994 	int i;
3995 
3996 	VM_OBJECT_ASSERT_WLOCKED(m->object);
3997 	/*
3998 	 * Scan the valid bits looking for invalid sections that
3999 	 * must be zeroed.  Invalid sub-DEV_BSIZE'd areas ( where the
4000 	 * valid bit may be set ) have already been zeroed by
4001 	 * vm_page_set_validclean().
4002 	 */
4003 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
4004 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
4005 		    (m->valid & ((vm_page_bits_t)1 << i))) {
4006 			if (i > b) {
4007 				pmap_zero_page_area(m,
4008 				    b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
4009 			}
4010 			b = i + 1;
4011 		}
4012 	}
4013 
4014 	/*
4015 	 * setvalid is TRUE when we can safely set the zero'd areas
4016 	 * as being valid.  We can do this if there are no cache consistancy
4017 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
4018 	 */
4019 	if (setvalid)
4020 		m->valid = VM_PAGE_BITS_ALL;
4021 }
4022 
4023 /*
4024  *	vm_page_is_valid:
4025  *
4026  *	Is (partial) page valid?  Note that the case where size == 0
4027  *	will return FALSE in the degenerate case where the page is
4028  *	entirely invalid, and TRUE otherwise.
4029  */
4030 int
4031 vm_page_is_valid(vm_page_t m, int base, int size)
4032 {
4033 	vm_page_bits_t bits;
4034 
4035 	VM_OBJECT_ASSERT_LOCKED(m->object);
4036 	bits = vm_page_bits(base, size);
4037 	return (m->valid != 0 && (m->valid & bits) == bits);
4038 }
4039 
4040 /*
4041  * Returns true if all of the specified predicates are true for the entire
4042  * (super)page and false otherwise.
4043  */
4044 bool
4045 vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m)
4046 {
4047 	vm_object_t object;
4048 	int i, npages;
4049 
4050 	object = m->object;
4051 	VM_OBJECT_ASSERT_LOCKED(object);
4052 	npages = atop(pagesizes[m->psind]);
4053 
4054 	/*
4055 	 * The physically contiguous pages that make up a superpage, i.e., a
4056 	 * page with a page size index ("psind") greater than zero, will
4057 	 * occupy adjacent entries in vm_page_array[].
4058 	 */
4059 	for (i = 0; i < npages; i++) {
4060 		/* Always test object consistency, including "skip_m". */
4061 		if (m[i].object != object)
4062 			return (false);
4063 		if (&m[i] == skip_m)
4064 			continue;
4065 		if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i]))
4066 			return (false);
4067 		if ((flags & PS_ALL_DIRTY) != 0) {
4068 			/*
4069 			 * Calling vm_page_test_dirty() or pmap_is_modified()
4070 			 * might stop this case from spuriously returning
4071 			 * "false".  However, that would require a write lock
4072 			 * on the object containing "m[i]".
4073 			 */
4074 			if (m[i].dirty != VM_PAGE_BITS_ALL)
4075 				return (false);
4076 		}
4077 		if ((flags & PS_ALL_VALID) != 0 &&
4078 		    m[i].valid != VM_PAGE_BITS_ALL)
4079 			return (false);
4080 	}
4081 	return (true);
4082 }
4083 
4084 /*
4085  * Set the page's dirty bits if the page is modified.
4086  */
4087 void
4088 vm_page_test_dirty(vm_page_t m)
4089 {
4090 
4091 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4092 	if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
4093 		vm_page_dirty(m);
4094 }
4095 
4096 void
4097 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
4098 {
4099 
4100 	mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
4101 }
4102 
4103 void
4104 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
4105 {
4106 
4107 	mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
4108 }
4109 
4110 int
4111 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
4112 {
4113 
4114 	return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
4115 }
4116 
4117 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
4118 void
4119 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
4120 {
4121 
4122 	vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
4123 }
4124 
4125 void
4126 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
4127 {
4128 
4129 	mtx_assert_(vm_page_lockptr(m), a, file, line);
4130 }
4131 #endif
4132 
4133 #ifdef INVARIANTS
4134 void
4135 vm_page_object_lock_assert(vm_page_t m)
4136 {
4137 
4138 	/*
4139 	 * Certain of the page's fields may only be modified by the
4140 	 * holder of the containing object's lock or the exclusive busy.
4141 	 * holder.  Unfortunately, the holder of the write busy is
4142 	 * not recorded, and thus cannot be checked here.
4143 	 */
4144 	if (m->object != NULL && !vm_page_xbusied(m))
4145 		VM_OBJECT_ASSERT_WLOCKED(m->object);
4146 }
4147 
4148 void
4149 vm_page_assert_pga_writeable(vm_page_t m, uint8_t bits)
4150 {
4151 
4152 	if ((bits & PGA_WRITEABLE) == 0)
4153 		return;
4154 
4155 	/*
4156 	 * The PGA_WRITEABLE flag can only be set if the page is
4157 	 * managed, is exclusively busied or the object is locked.
4158 	 * Currently, this flag is only set by pmap_enter().
4159 	 */
4160 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4161 	    ("PGA_WRITEABLE on unmanaged page"));
4162 	if (!vm_page_xbusied(m))
4163 		VM_OBJECT_ASSERT_LOCKED(m->object);
4164 }
4165 #endif
4166 
4167 #include "opt_ddb.h"
4168 #ifdef DDB
4169 #include <sys/kernel.h>
4170 
4171 #include <ddb/ddb.h>
4172 
4173 DB_SHOW_COMMAND(page, vm_page_print_page_info)
4174 {
4175 
4176 	db_printf("vm_cnt.v_free_count: %d\n", vm_free_count());
4177 	db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count());
4178 	db_printf("vm_cnt.v_active_count: %d\n", vm_active_count());
4179 	db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count());
4180 	db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count());
4181 	db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
4182 	db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
4183 	db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
4184 	db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
4185 }
4186 
4187 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
4188 {
4189 	int dom;
4190 
4191 	db_printf("pq_free %d\n", vm_free_count());
4192 	for (dom = 0; dom < vm_ndomains; dom++) {
4193 		db_printf(
4194     "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n",
4195 		    dom,
4196 		    vm_dom[dom].vmd_page_count,
4197 		    vm_dom[dom].vmd_free_count,
4198 		    vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
4199 		    vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
4200 		    vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt,
4201 		    vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt);
4202 	}
4203 }
4204 
4205 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
4206 {
4207 	vm_page_t m;
4208 	boolean_t phys;
4209 
4210 	if (!have_addr) {
4211 		db_printf("show pginfo addr\n");
4212 		return;
4213 	}
4214 
4215 	phys = strchr(modif, 'p') != NULL;
4216 	if (phys)
4217 		m = PHYS_TO_VM_PAGE(addr);
4218 	else
4219 		m = (vm_page_t)addr;
4220 	db_printf(
4221     "page %p obj %p pidx 0x%jx phys 0x%jx q %d hold %d wire %d\n"
4222     "  af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
4223 	    m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
4224 	    m->queue, m->hold_count, m->wire_count, m->aflags, m->oflags,
4225 	    m->flags, m->act_count, m->busy_lock, m->valid, m->dirty);
4226 }
4227 #endif /* DDB */
4228