xref: /freebsd/sys/vm/vm_page.c (revision 5ca85583098f2506a34ef7ddec764f841cedadb4)
1 /*-
2  * Copyright (c) 1991 Regents of the University of California.
3  * All rights reserved.
4  * Copyright (c) 1998 Matthew Dillon.  All Rights Reserved.
5  *
6  * This code is derived from software contributed to Berkeley by
7  * The Mach Operating System project at Carnegie-Mellon University.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  * 4. Neither the name of the University nor the names of its contributors
18  *    may be used to endorse or promote products derived from this software
19  *    without specific prior written permission.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
22  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
25  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31  * SUCH DAMAGE.
32  *
33  *	from: @(#)vm_page.c	7.4 (Berkeley) 5/7/91
34  */
35 
36 /*-
37  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
38  * All rights reserved.
39  *
40  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
41  *
42  * Permission to use, copy, modify and distribute this software and
43  * its documentation is hereby granted, provided that both the copyright
44  * notice and this permission notice appear in all copies of the
45  * software, derivative works or modified versions, and any portions
46  * thereof, and that both notices appear in supporting documentation.
47  *
48  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
49  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
50  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
51  *
52  * Carnegie Mellon requests users of this software to return to
53  *
54  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
55  *  School of Computer Science
56  *  Carnegie Mellon University
57  *  Pittsburgh PA 15213-3890
58  *
59  * any improvements or extensions that they make and grant Carnegie the
60  * rights to redistribute these changes.
61  */
62 
63 /*
64  *			GENERAL RULES ON VM_PAGE MANIPULATION
65  *
66  *	- A page queue lock is required when adding or removing a page from a
67  *	  page queue regardless of other locks or the busy state of a page.
68  *
69  *		* In general, no thread besides the page daemon can acquire or
70  *		  hold more than one page queue lock at a time.
71  *
72  *		* The page daemon can acquire and hold any pair of page queue
73  *		  locks in any order.
74  *
75  *	- The object lock is required when inserting or removing
76  *	  pages from an object (vm_page_insert() or vm_page_remove()).
77  *
78  */
79 
80 /*
81  *	Resident memory management module.
82  */
83 
84 #include <sys/cdefs.h>
85 __FBSDID("$FreeBSD$");
86 
87 #include "opt_vm.h"
88 
89 #include <sys/param.h>
90 #include <sys/systm.h>
91 #include <sys/lock.h>
92 #include <sys/kernel.h>
93 #include <sys/limits.h>
94 #include <sys/linker.h>
95 #include <sys/malloc.h>
96 #include <sys/mman.h>
97 #include <sys/msgbuf.h>
98 #include <sys/mutex.h>
99 #include <sys/proc.h>
100 #include <sys/rwlock.h>
101 #include <sys/sbuf.h>
102 #include <sys/smp.h>
103 #include <sys/sysctl.h>
104 #include <sys/vmmeter.h>
105 #include <sys/vnode.h>
106 
107 #include <vm/vm.h>
108 #include <vm/pmap.h>
109 #include <vm/vm_param.h>
110 #include <vm/vm_kern.h>
111 #include <vm/vm_object.h>
112 #include <vm/vm_page.h>
113 #include <vm/vm_pageout.h>
114 #include <vm/vm_pager.h>
115 #include <vm/vm_phys.h>
116 #include <vm/vm_radix.h>
117 #include <vm/vm_reserv.h>
118 #include <vm/vm_extern.h>
119 #include <vm/uma.h>
120 #include <vm/uma_int.h>
121 
122 #include <machine/md_var.h>
123 
124 /*
125  *	Associated with page of user-allocatable memory is a
126  *	page structure.
127  */
128 
129 struct vm_domain vm_dom[MAXMEMDOM];
130 struct mtx_padalign vm_page_queue_free_mtx;
131 
132 struct mtx_padalign pa_lock[PA_LOCK_COUNT];
133 
134 /*
135  * bogus page -- for I/O to/from partially complete buffers,
136  * or for paging into sparsely invalid regions.
137  */
138 vm_page_t bogus_page;
139 
140 vm_page_t vm_page_array;
141 long vm_page_array_size;
142 long first_page;
143 
144 static int boot_pages = UMA_BOOT_PAGES;
145 SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
146     &boot_pages, 0,
147     "number of pages allocated for bootstrapping the VM system");
148 
149 static int pa_tryrelock_restart;
150 SYSCTL_INT(_vm, OID_AUTO, tryrelock_restart, CTLFLAG_RD,
151     &pa_tryrelock_restart, 0, "Number of tryrelock restarts");
152 
153 static TAILQ_HEAD(, vm_page) blacklist_head;
154 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS);
155 SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD |
156     CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages");
157 
158 /* Is the page daemon waiting for free pages? */
159 static int vm_pageout_pages_needed;
160 
161 static uma_zone_t fakepg_zone;
162 
163 static void vm_page_alloc_check(vm_page_t m);
164 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits);
165 static void vm_page_enqueue(uint8_t queue, vm_page_t m);
166 static void vm_page_free_wakeup(void);
167 static void vm_page_init(void *dummy);
168 static int vm_page_insert_after(vm_page_t m, vm_object_t object,
169     vm_pindex_t pindex, vm_page_t mpred);
170 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object,
171     vm_page_t mpred);
172 static int vm_page_reclaim_run(int req_class, u_long npages, vm_page_t m_run,
173     vm_paddr_t high);
174 
175 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL);
176 
177 static void
178 vm_page_init(void *dummy)
179 {
180 
181 	fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
182 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
183 	bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
184 	    VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
185 }
186 
187 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
188 #if PAGE_SIZE == 32768
189 #ifdef CTASSERT
190 CTASSERT(sizeof(u_long) >= 8);
191 #endif
192 #endif
193 
194 /*
195  * Try to acquire a physical address lock while a pmap is locked.  If we
196  * fail to trylock we unlock and lock the pmap directly and cache the
197  * locked pa in *locked.  The caller should then restart their loop in case
198  * the virtual to physical mapping has changed.
199  */
200 int
201 vm_page_pa_tryrelock(pmap_t pmap, vm_paddr_t pa, vm_paddr_t *locked)
202 {
203 	vm_paddr_t lockpa;
204 
205 	lockpa = *locked;
206 	*locked = pa;
207 	if (lockpa) {
208 		PA_LOCK_ASSERT(lockpa, MA_OWNED);
209 		if (PA_LOCKPTR(pa) == PA_LOCKPTR(lockpa))
210 			return (0);
211 		PA_UNLOCK(lockpa);
212 	}
213 	if (PA_TRYLOCK(pa))
214 		return (0);
215 	PMAP_UNLOCK(pmap);
216 	atomic_add_int(&pa_tryrelock_restart, 1);
217 	PA_LOCK(pa);
218 	PMAP_LOCK(pmap);
219 	return (EAGAIN);
220 }
221 
222 /*
223  *	vm_set_page_size:
224  *
225  *	Sets the page size, perhaps based upon the memory
226  *	size.  Must be called before any use of page-size
227  *	dependent functions.
228  */
229 void
230 vm_set_page_size(void)
231 {
232 	if (vm_cnt.v_page_size == 0)
233 		vm_cnt.v_page_size = PAGE_SIZE;
234 	if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0)
235 		panic("vm_set_page_size: page size not a power of two");
236 }
237 
238 /*
239  *	vm_page_blacklist_next:
240  *
241  *	Find the next entry in the provided string of blacklist
242  *	addresses.  Entries are separated by space, comma, or newline.
243  *	If an invalid integer is encountered then the rest of the
244  *	string is skipped.  Updates the list pointer to the next
245  *	character, or NULL if the string is exhausted or invalid.
246  */
247 static vm_paddr_t
248 vm_page_blacklist_next(char **list, char *end)
249 {
250 	vm_paddr_t bad;
251 	char *cp, *pos;
252 
253 	if (list == NULL || *list == NULL)
254 		return (0);
255 	if (**list =='\0') {
256 		*list = NULL;
257 		return (0);
258 	}
259 
260 	/*
261 	 * If there's no end pointer then the buffer is coming from
262 	 * the kenv and we know it's null-terminated.
263 	 */
264 	if (end == NULL)
265 		end = *list + strlen(*list);
266 
267 	/* Ensure that strtoq() won't walk off the end */
268 	if (*end != '\0') {
269 		if (*end == '\n' || *end == ' ' || *end  == ',')
270 			*end = '\0';
271 		else {
272 			printf("Blacklist not terminated, skipping\n");
273 			*list = NULL;
274 			return (0);
275 		}
276 	}
277 
278 	for (pos = *list; *pos != '\0'; pos = cp) {
279 		bad = strtoq(pos, &cp, 0);
280 		if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') {
281 			if (bad == 0) {
282 				if (++cp < end)
283 					continue;
284 				else
285 					break;
286 			}
287 		} else
288 			break;
289 		if (*cp == '\0' || ++cp >= end)
290 			*list = NULL;
291 		else
292 			*list = cp;
293 		return (trunc_page(bad));
294 	}
295 	printf("Garbage in RAM blacklist, skipping\n");
296 	*list = NULL;
297 	return (0);
298 }
299 
300 /*
301  *	vm_page_blacklist_check:
302  *
303  *	Iterate through the provided string of blacklist addresses, pulling
304  *	each entry out of the physical allocator free list and putting it
305  *	onto a list for reporting via the vm.page_blacklist sysctl.
306  */
307 static void
308 vm_page_blacklist_check(char *list, char *end)
309 {
310 	vm_paddr_t pa;
311 	vm_page_t m;
312 	char *next;
313 	int ret;
314 
315 	next = list;
316 	while (next != NULL) {
317 		if ((pa = vm_page_blacklist_next(&next, end)) == 0)
318 			continue;
319 		m = vm_phys_paddr_to_vm_page(pa);
320 		if (m == NULL)
321 			continue;
322 		mtx_lock(&vm_page_queue_free_mtx);
323 		ret = vm_phys_unfree_page(m);
324 		mtx_unlock(&vm_page_queue_free_mtx);
325 		if (ret == TRUE) {
326 			TAILQ_INSERT_TAIL(&blacklist_head, m, listq);
327 			if (bootverbose)
328 				printf("Skipping page with pa 0x%jx\n",
329 				    (uintmax_t)pa);
330 		}
331 	}
332 }
333 
334 /*
335  *	vm_page_blacklist_load:
336  *
337  *	Search for a special module named "ram_blacklist".  It'll be a
338  *	plain text file provided by the user via the loader directive
339  *	of the same name.
340  */
341 static void
342 vm_page_blacklist_load(char **list, char **end)
343 {
344 	void *mod;
345 	u_char *ptr;
346 	u_int len;
347 
348 	mod = NULL;
349 	ptr = NULL;
350 
351 	mod = preload_search_by_type("ram_blacklist");
352 	if (mod != NULL) {
353 		ptr = preload_fetch_addr(mod);
354 		len = preload_fetch_size(mod);
355         }
356 	*list = ptr;
357 	if (ptr != NULL)
358 		*end = ptr + len;
359 	else
360 		*end = NULL;
361 	return;
362 }
363 
364 static int
365 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS)
366 {
367 	vm_page_t m;
368 	struct sbuf sbuf;
369 	int error, first;
370 
371 	first = 1;
372 	error = sysctl_wire_old_buffer(req, 0);
373 	if (error != 0)
374 		return (error);
375 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
376 	TAILQ_FOREACH(m, &blacklist_head, listq) {
377 		sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",",
378 		    (uintmax_t)m->phys_addr);
379 		first = 0;
380 	}
381 	error = sbuf_finish(&sbuf);
382 	sbuf_delete(&sbuf);
383 	return (error);
384 }
385 
386 static void
387 vm_page_domain_init(struct vm_domain *vmd)
388 {
389 	struct vm_pagequeue *pq;
390 	int i;
391 
392 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) =
393 	    "vm inactive pagequeue";
394 	*__DECONST(u_int **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_vcnt) =
395 	    &vm_cnt.v_inactive_count;
396 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) =
397 	    "vm active pagequeue";
398 	*__DECONST(u_int **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_vcnt) =
399 	    &vm_cnt.v_active_count;
400 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) =
401 	    "vm laundry pagequeue";
402 	*__DECONST(int **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_vcnt) =
403 	    &vm_cnt.v_laundry_count;
404 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) =
405 	    "vm unswappable pagequeue";
406 	/* Unswappable dirty pages are counted as being in the laundry. */
407 	*__DECONST(int **, &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_vcnt) =
408 	    &vm_cnt.v_laundry_count;
409 	vmd->vmd_page_count = 0;
410 	vmd->vmd_free_count = 0;
411 	vmd->vmd_segs = 0;
412 	vmd->vmd_oom = FALSE;
413 	for (i = 0; i < PQ_COUNT; i++) {
414 		pq = &vmd->vmd_pagequeues[i];
415 		TAILQ_INIT(&pq->pq_pl);
416 		mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue",
417 		    MTX_DEF | MTX_DUPOK);
418 	}
419 }
420 
421 /*
422  *	vm_page_startup:
423  *
424  *	Initializes the resident memory module.
425  *
426  *	Allocates memory for the page cells, and
427  *	for the object/offset-to-page hash table headers.
428  *	Each page cell is initialized and placed on the free list.
429  */
430 vm_offset_t
431 vm_page_startup(vm_offset_t vaddr)
432 {
433 	vm_offset_t mapped;
434 	vm_paddr_t page_range;
435 	vm_paddr_t new_end;
436 	int i;
437 	vm_paddr_t pa;
438 	vm_paddr_t last_pa;
439 	char *list, *listend;
440 	vm_paddr_t end;
441 	vm_paddr_t biggestsize;
442 	vm_paddr_t low_water, high_water;
443 	int biggestone;
444 	int pages_per_zone;
445 
446 	biggestsize = 0;
447 	biggestone = 0;
448 	vaddr = round_page(vaddr);
449 
450 	for (i = 0; phys_avail[i + 1]; i += 2) {
451 		phys_avail[i] = round_page(phys_avail[i]);
452 		phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
453 	}
454 
455 	low_water = phys_avail[0];
456 	high_water = phys_avail[1];
457 
458 	for (i = 0; i < vm_phys_nsegs; i++) {
459 		if (vm_phys_segs[i].start < low_water)
460 			low_water = vm_phys_segs[i].start;
461 		if (vm_phys_segs[i].end > high_water)
462 			high_water = vm_phys_segs[i].end;
463 	}
464 	for (i = 0; phys_avail[i + 1]; i += 2) {
465 		vm_paddr_t size = phys_avail[i + 1] - phys_avail[i];
466 
467 		if (size > biggestsize) {
468 			biggestone = i;
469 			biggestsize = size;
470 		}
471 		if (phys_avail[i] < low_water)
472 			low_water = phys_avail[i];
473 		if (phys_avail[i + 1] > high_water)
474 			high_water = phys_avail[i + 1];
475 	}
476 
477 	end = phys_avail[biggestone+1];
478 
479 	/*
480 	 * Initialize the page and queue locks.
481 	 */
482 	mtx_init(&vm_page_queue_free_mtx, "vm page free queue", NULL, MTX_DEF);
483 	for (i = 0; i < PA_LOCK_COUNT; i++)
484 		mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
485 	for (i = 0; i < vm_ndomains; i++)
486 		vm_page_domain_init(&vm_dom[i]);
487 
488 	/*
489 	 * Almost all of the pages needed for boot strapping UMA are used
490 	 * for zone structures, so if the number of CPUs results in those
491 	 * structures taking more than one page each, we set aside more pages
492 	 * in proportion to the zone structure size.
493 	 */
494 	pages_per_zone = howmany(sizeof(struct uma_zone) +
495 	    sizeof(struct uma_cache) * (mp_maxid + 1), UMA_SLAB_SIZE);
496 	if (pages_per_zone > 1) {
497 		/* Reserve more pages so that we don't run out. */
498 		boot_pages = UMA_BOOT_PAGES_ZONES * pages_per_zone;
499 	}
500 
501 	/*
502 	 * Allocate memory for use when boot strapping the kernel memory
503 	 * allocator.
504 	 *
505 	 * CTFLAG_RDTUN doesn't work during the early boot process, so we must
506 	 * manually fetch the value.
507 	 */
508 	TUNABLE_INT_FETCH("vm.boot_pages", &boot_pages);
509 	new_end = end - (boot_pages * UMA_SLAB_SIZE);
510 	new_end = trunc_page(new_end);
511 	mapped = pmap_map(&vaddr, new_end, end,
512 	    VM_PROT_READ | VM_PROT_WRITE);
513 	bzero((void *)mapped, end - new_end);
514 	uma_startup((void *)mapped, boot_pages);
515 
516 #if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \
517     defined(__i386__) || defined(__mips__)
518 	/*
519 	 * Allocate a bitmap to indicate that a random physical page
520 	 * needs to be included in a minidump.
521 	 *
522 	 * The amd64 port needs this to indicate which direct map pages
523 	 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
524 	 *
525 	 * However, i386 still needs this workspace internally within the
526 	 * minidump code.  In theory, they are not needed on i386, but are
527 	 * included should the sf_buf code decide to use them.
528 	 */
529 	last_pa = 0;
530 	for (i = 0; dump_avail[i + 1] != 0; i += 2)
531 		if (dump_avail[i + 1] > last_pa)
532 			last_pa = dump_avail[i + 1];
533 	page_range = last_pa / PAGE_SIZE;
534 	vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
535 	new_end -= vm_page_dump_size;
536 	vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
537 	    new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
538 	bzero((void *)vm_page_dump, vm_page_dump_size);
539 #endif
540 #ifdef __amd64__
541 	/*
542 	 * Request that the physical pages underlying the message buffer be
543 	 * included in a crash dump.  Since the message buffer is accessed
544 	 * through the direct map, they are not automatically included.
545 	 */
546 	pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
547 	last_pa = pa + round_page(msgbufsize);
548 	while (pa < last_pa) {
549 		dump_add_page(pa);
550 		pa += PAGE_SIZE;
551 	}
552 #endif
553 	/*
554 	 * Compute the number of pages of memory that will be available for
555 	 * use (taking into account the overhead of a page structure per
556 	 * page).
557 	 */
558 	first_page = low_water / PAGE_SIZE;
559 #ifdef VM_PHYSSEG_SPARSE
560 	page_range = 0;
561 	for (i = 0; i < vm_phys_nsegs; i++) {
562 		page_range += atop(vm_phys_segs[i].end -
563 		    vm_phys_segs[i].start);
564 	}
565 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
566 		page_range += atop(phys_avail[i + 1] - phys_avail[i]);
567 #elif defined(VM_PHYSSEG_DENSE)
568 	page_range = high_water / PAGE_SIZE - first_page;
569 #else
570 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
571 #endif
572 	end = new_end;
573 
574 	/*
575 	 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
576 	 */
577 	vaddr += PAGE_SIZE;
578 
579 	/*
580 	 * Initialize the mem entry structures now, and put them in the free
581 	 * queue.
582 	 */
583 	new_end = trunc_page(end - page_range * sizeof(struct vm_page));
584 	mapped = pmap_map(&vaddr, new_end, end,
585 	    VM_PROT_READ | VM_PROT_WRITE);
586 	vm_page_array = (vm_page_t) mapped;
587 #if VM_NRESERVLEVEL > 0
588 	/*
589 	 * Allocate memory for the reservation management system's data
590 	 * structures.
591 	 */
592 	new_end = vm_reserv_startup(&vaddr, new_end, high_water);
593 #endif
594 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__)
595 	/*
596 	 * pmap_map on arm64, amd64, and mips can come out of the direct-map,
597 	 * not kvm like i386, so the pages must be tracked for a crashdump to
598 	 * include this data.  This includes the vm_page_array and the early
599 	 * UMA bootstrap pages.
600 	 */
601 	for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE)
602 		dump_add_page(pa);
603 #endif
604 	phys_avail[biggestone + 1] = new_end;
605 
606 	/*
607 	 * Add physical memory segments corresponding to the available
608 	 * physical pages.
609 	 */
610 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
611 		vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]);
612 
613 	/*
614 	 * Clear all of the page structures
615 	 */
616 	bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
617 	for (i = 0; i < page_range; i++)
618 		vm_page_array[i].order = VM_NFREEORDER;
619 	vm_page_array_size = page_range;
620 
621 	/*
622 	 * Initialize the physical memory allocator.
623 	 */
624 	vm_phys_init();
625 
626 	/*
627 	 * Add every available physical page that is not blacklisted to
628 	 * the free lists.
629 	 */
630 	vm_cnt.v_page_count = 0;
631 	vm_cnt.v_free_count = 0;
632 	for (i = 0; phys_avail[i + 1] != 0; i += 2) {
633 		pa = phys_avail[i];
634 		last_pa = phys_avail[i + 1];
635 		while (pa < last_pa) {
636 			vm_phys_add_page(pa);
637 			pa += PAGE_SIZE;
638 		}
639 	}
640 
641 	TAILQ_INIT(&blacklist_head);
642 	vm_page_blacklist_load(&list, &listend);
643 	vm_page_blacklist_check(list, listend);
644 
645 	list = kern_getenv("vm.blacklist");
646 	vm_page_blacklist_check(list, NULL);
647 
648 	freeenv(list);
649 #if VM_NRESERVLEVEL > 0
650 	/*
651 	 * Initialize the reservation management system.
652 	 */
653 	vm_reserv_init();
654 #endif
655 	return (vaddr);
656 }
657 
658 void
659 vm_page_reference(vm_page_t m)
660 {
661 
662 	vm_page_aflag_set(m, PGA_REFERENCED);
663 }
664 
665 /*
666  *	vm_page_busy_downgrade:
667  *
668  *	Downgrade an exclusive busy page into a single shared busy page.
669  */
670 void
671 vm_page_busy_downgrade(vm_page_t m)
672 {
673 	u_int x;
674 	bool locked;
675 
676 	vm_page_assert_xbusied(m);
677 	locked = mtx_owned(vm_page_lockptr(m));
678 
679 	for (;;) {
680 		x = m->busy_lock;
681 		x &= VPB_BIT_WAITERS;
682 		if (x != 0 && !locked)
683 			vm_page_lock(m);
684 		if (atomic_cmpset_rel_int(&m->busy_lock,
685 		    VPB_SINGLE_EXCLUSIVER | x, VPB_SHARERS_WORD(1)))
686 			break;
687 		if (x != 0 && !locked)
688 			vm_page_unlock(m);
689 	}
690 	if (x != 0) {
691 		wakeup(m);
692 		if (!locked)
693 			vm_page_unlock(m);
694 	}
695 }
696 
697 /*
698  *	vm_page_sbusied:
699  *
700  *	Return a positive value if the page is shared busied, 0 otherwise.
701  */
702 int
703 vm_page_sbusied(vm_page_t m)
704 {
705 	u_int x;
706 
707 	x = m->busy_lock;
708 	return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED);
709 }
710 
711 /*
712  *	vm_page_sunbusy:
713  *
714  *	Shared unbusy a page.
715  */
716 void
717 vm_page_sunbusy(vm_page_t m)
718 {
719 	u_int x;
720 
721 	vm_page_assert_sbusied(m);
722 
723 	for (;;) {
724 		x = m->busy_lock;
725 		if (VPB_SHARERS(x) > 1) {
726 			if (atomic_cmpset_int(&m->busy_lock, x,
727 			    x - VPB_ONE_SHARER))
728 				break;
729 			continue;
730 		}
731 		if ((x & VPB_BIT_WAITERS) == 0) {
732 			KASSERT(x == VPB_SHARERS_WORD(1),
733 			    ("vm_page_sunbusy: invalid lock state"));
734 			if (atomic_cmpset_int(&m->busy_lock,
735 			    VPB_SHARERS_WORD(1), VPB_UNBUSIED))
736 				break;
737 			continue;
738 		}
739 		KASSERT(x == (VPB_SHARERS_WORD(1) | VPB_BIT_WAITERS),
740 		    ("vm_page_sunbusy: invalid lock state for waiters"));
741 
742 		vm_page_lock(m);
743 		if (!atomic_cmpset_int(&m->busy_lock, x, VPB_UNBUSIED)) {
744 			vm_page_unlock(m);
745 			continue;
746 		}
747 		wakeup(m);
748 		vm_page_unlock(m);
749 		break;
750 	}
751 }
752 
753 /*
754  *	vm_page_busy_sleep:
755  *
756  *	Sleep and release the page lock, using the page pointer as wchan.
757  *	This is used to implement the hard-path of busying mechanism.
758  *
759  *	The given page must be locked.
760  *
761  *	If nonshared is true, sleep only if the page is xbusy.
762  */
763 void
764 vm_page_busy_sleep(vm_page_t m, const char *wmesg, bool nonshared)
765 {
766 	u_int x;
767 
768 	vm_page_assert_locked(m);
769 
770 	x = m->busy_lock;
771 	if (x == VPB_UNBUSIED || (nonshared && (x & VPB_BIT_SHARED) != 0) ||
772 	    ((x & VPB_BIT_WAITERS) == 0 &&
773 	    !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS))) {
774 		vm_page_unlock(m);
775 		return;
776 	}
777 	msleep(m, vm_page_lockptr(m), PVM | PDROP, wmesg, 0);
778 }
779 
780 /*
781  *	vm_page_trysbusy:
782  *
783  *	Try to shared busy a page.
784  *	If the operation succeeds 1 is returned otherwise 0.
785  *	The operation never sleeps.
786  */
787 int
788 vm_page_trysbusy(vm_page_t m)
789 {
790 	u_int x;
791 
792 	for (;;) {
793 		x = m->busy_lock;
794 		if ((x & VPB_BIT_SHARED) == 0)
795 			return (0);
796 		if (atomic_cmpset_acq_int(&m->busy_lock, x, x + VPB_ONE_SHARER))
797 			return (1);
798 	}
799 }
800 
801 static void
802 vm_page_xunbusy_locked(vm_page_t m)
803 {
804 
805 	vm_page_assert_xbusied(m);
806 	vm_page_assert_locked(m);
807 
808 	atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
809 	/* There is a waiter, do wakeup() instead of vm_page_flash(). */
810 	wakeup(m);
811 }
812 
813 void
814 vm_page_xunbusy_maybelocked(vm_page_t m)
815 {
816 	bool lockacq;
817 
818 	vm_page_assert_xbusied(m);
819 
820 	/*
821 	 * Fast path for unbusy.  If it succeeds, we know that there
822 	 * are no waiters, so we do not need a wakeup.
823 	 */
824 	if (atomic_cmpset_rel_int(&m->busy_lock, VPB_SINGLE_EXCLUSIVER,
825 	    VPB_UNBUSIED))
826 		return;
827 
828 	lockacq = !mtx_owned(vm_page_lockptr(m));
829 	if (lockacq)
830 		vm_page_lock(m);
831 	vm_page_xunbusy_locked(m);
832 	if (lockacq)
833 		vm_page_unlock(m);
834 }
835 
836 /*
837  *	vm_page_xunbusy_hard:
838  *
839  *	Called after the first try the exclusive unbusy of a page failed.
840  *	It is assumed that the waiters bit is on.
841  */
842 void
843 vm_page_xunbusy_hard(vm_page_t m)
844 {
845 
846 	vm_page_assert_xbusied(m);
847 
848 	vm_page_lock(m);
849 	vm_page_xunbusy_locked(m);
850 	vm_page_unlock(m);
851 }
852 
853 /*
854  *	vm_page_flash:
855  *
856  *	Wakeup anyone waiting for the page.
857  *	The ownership bits do not change.
858  *
859  *	The given page must be locked.
860  */
861 void
862 vm_page_flash(vm_page_t m)
863 {
864 	u_int x;
865 
866 	vm_page_lock_assert(m, MA_OWNED);
867 
868 	for (;;) {
869 		x = m->busy_lock;
870 		if ((x & VPB_BIT_WAITERS) == 0)
871 			return;
872 		if (atomic_cmpset_int(&m->busy_lock, x,
873 		    x & (~VPB_BIT_WAITERS)))
874 			break;
875 	}
876 	wakeup(m);
877 }
878 
879 /*
880  * Keep page from being freed by the page daemon
881  * much of the same effect as wiring, except much lower
882  * overhead and should be used only for *very* temporary
883  * holding ("wiring").
884  */
885 void
886 vm_page_hold(vm_page_t mem)
887 {
888 
889 	vm_page_lock_assert(mem, MA_OWNED);
890         mem->hold_count++;
891 }
892 
893 void
894 vm_page_unhold(vm_page_t mem)
895 {
896 
897 	vm_page_lock_assert(mem, MA_OWNED);
898 	KASSERT(mem->hold_count >= 1, ("vm_page_unhold: hold count < 0!!!"));
899 	--mem->hold_count;
900 	if (mem->hold_count == 0 && (mem->flags & PG_UNHOLDFREE) != 0)
901 		vm_page_free_toq(mem);
902 }
903 
904 /*
905  *	vm_page_unhold_pages:
906  *
907  *	Unhold each of the pages that is referenced by the given array.
908  */
909 void
910 vm_page_unhold_pages(vm_page_t *ma, int count)
911 {
912 	struct mtx *mtx, *new_mtx;
913 
914 	mtx = NULL;
915 	for (; count != 0; count--) {
916 		/*
917 		 * Avoid releasing and reacquiring the same page lock.
918 		 */
919 		new_mtx = vm_page_lockptr(*ma);
920 		if (mtx != new_mtx) {
921 			if (mtx != NULL)
922 				mtx_unlock(mtx);
923 			mtx = new_mtx;
924 			mtx_lock(mtx);
925 		}
926 		vm_page_unhold(*ma);
927 		ma++;
928 	}
929 	if (mtx != NULL)
930 		mtx_unlock(mtx);
931 }
932 
933 vm_page_t
934 PHYS_TO_VM_PAGE(vm_paddr_t pa)
935 {
936 	vm_page_t m;
937 
938 #ifdef VM_PHYSSEG_SPARSE
939 	m = vm_phys_paddr_to_vm_page(pa);
940 	if (m == NULL)
941 		m = vm_phys_fictitious_to_vm_page(pa);
942 	return (m);
943 #elif defined(VM_PHYSSEG_DENSE)
944 	long pi;
945 
946 	pi = atop(pa);
947 	if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
948 		m = &vm_page_array[pi - first_page];
949 		return (m);
950 	}
951 	return (vm_phys_fictitious_to_vm_page(pa));
952 #else
953 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
954 #endif
955 }
956 
957 /*
958  *	vm_page_getfake:
959  *
960  *	Create a fictitious page with the specified physical address and
961  *	memory attribute.  The memory attribute is the only the machine-
962  *	dependent aspect of a fictitious page that must be initialized.
963  */
964 vm_page_t
965 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
966 {
967 	vm_page_t m;
968 
969 	m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
970 	vm_page_initfake(m, paddr, memattr);
971 	return (m);
972 }
973 
974 void
975 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
976 {
977 
978 	if ((m->flags & PG_FICTITIOUS) != 0) {
979 		/*
980 		 * The page's memattr might have changed since the
981 		 * previous initialization.  Update the pmap to the
982 		 * new memattr.
983 		 */
984 		goto memattr;
985 	}
986 	m->phys_addr = paddr;
987 	m->queue = PQ_NONE;
988 	/* Fictitious pages don't use "segind". */
989 	m->flags = PG_FICTITIOUS;
990 	/* Fictitious pages don't use "order" or "pool". */
991 	m->oflags = VPO_UNMANAGED;
992 	m->busy_lock = VPB_SINGLE_EXCLUSIVER;
993 	m->wire_count = 1;
994 	pmap_page_init(m);
995 memattr:
996 	pmap_page_set_memattr(m, memattr);
997 }
998 
999 /*
1000  *	vm_page_putfake:
1001  *
1002  *	Release a fictitious page.
1003  */
1004 void
1005 vm_page_putfake(vm_page_t m)
1006 {
1007 
1008 	KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
1009 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1010 	    ("vm_page_putfake: bad page %p", m));
1011 	uma_zfree(fakepg_zone, m);
1012 }
1013 
1014 /*
1015  *	vm_page_updatefake:
1016  *
1017  *	Update the given fictitious page to the specified physical address and
1018  *	memory attribute.
1019  */
1020 void
1021 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1022 {
1023 
1024 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1025 	    ("vm_page_updatefake: bad page %p", m));
1026 	m->phys_addr = paddr;
1027 	pmap_page_set_memattr(m, memattr);
1028 }
1029 
1030 /*
1031  *	vm_page_free:
1032  *
1033  *	Free a page.
1034  */
1035 void
1036 vm_page_free(vm_page_t m)
1037 {
1038 
1039 	m->flags &= ~PG_ZERO;
1040 	vm_page_free_toq(m);
1041 }
1042 
1043 /*
1044  *	vm_page_free_zero:
1045  *
1046  *	Free a page to the zerod-pages queue
1047  */
1048 void
1049 vm_page_free_zero(vm_page_t m)
1050 {
1051 
1052 	m->flags |= PG_ZERO;
1053 	vm_page_free_toq(m);
1054 }
1055 
1056 /*
1057  * Unbusy and handle the page queueing for a page from a getpages request that
1058  * was optionally read ahead or behind.
1059  */
1060 void
1061 vm_page_readahead_finish(vm_page_t m)
1062 {
1063 
1064 	/* We shouldn't put invalid pages on queues. */
1065 	KASSERT(m->valid != 0, ("%s: %p is invalid", __func__, m));
1066 
1067 	/*
1068 	 * Since the page is not the actually needed one, whether it should
1069 	 * be activated or deactivated is not obvious.  Empirical results
1070 	 * have shown that deactivating the page is usually the best choice,
1071 	 * unless the page is wanted by another thread.
1072 	 */
1073 	vm_page_lock(m);
1074 	if ((m->busy_lock & VPB_BIT_WAITERS) != 0)
1075 		vm_page_activate(m);
1076 	else
1077 		vm_page_deactivate(m);
1078 	vm_page_unlock(m);
1079 	vm_page_xunbusy(m);
1080 }
1081 
1082 /*
1083  *	vm_page_sleep_if_busy:
1084  *
1085  *	Sleep and release the page queues lock if the page is busied.
1086  *	Returns TRUE if the thread slept.
1087  *
1088  *	The given page must be unlocked and object containing it must
1089  *	be locked.
1090  */
1091 int
1092 vm_page_sleep_if_busy(vm_page_t m, const char *msg)
1093 {
1094 	vm_object_t obj;
1095 
1096 	vm_page_lock_assert(m, MA_NOTOWNED);
1097 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1098 
1099 	if (vm_page_busied(m)) {
1100 		/*
1101 		 * The page-specific object must be cached because page
1102 		 * identity can change during the sleep, causing the
1103 		 * re-lock of a different object.
1104 		 * It is assumed that a reference to the object is already
1105 		 * held by the callers.
1106 		 */
1107 		obj = m->object;
1108 		vm_page_lock(m);
1109 		VM_OBJECT_WUNLOCK(obj);
1110 		vm_page_busy_sleep(m, msg, false);
1111 		VM_OBJECT_WLOCK(obj);
1112 		return (TRUE);
1113 	}
1114 	return (FALSE);
1115 }
1116 
1117 /*
1118  *	vm_page_dirty_KBI:		[ internal use only ]
1119  *
1120  *	Set all bits in the page's dirty field.
1121  *
1122  *	The object containing the specified page must be locked if the
1123  *	call is made from the machine-independent layer.
1124  *
1125  *	See vm_page_clear_dirty_mask().
1126  *
1127  *	This function should only be called by vm_page_dirty().
1128  */
1129 void
1130 vm_page_dirty_KBI(vm_page_t m)
1131 {
1132 
1133 	/* Refer to this operation by its public name. */
1134 	KASSERT(m->valid == VM_PAGE_BITS_ALL,
1135 	    ("vm_page_dirty: page is invalid!"));
1136 	m->dirty = VM_PAGE_BITS_ALL;
1137 }
1138 
1139 /*
1140  *	vm_page_insert:		[ internal use only ]
1141  *
1142  *	Inserts the given mem entry into the object and object list.
1143  *
1144  *	The object must be locked.
1145  */
1146 int
1147 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1148 {
1149 	vm_page_t mpred;
1150 
1151 	VM_OBJECT_ASSERT_WLOCKED(object);
1152 	mpred = vm_radix_lookup_le(&object->rtree, pindex);
1153 	return (vm_page_insert_after(m, object, pindex, mpred));
1154 }
1155 
1156 /*
1157  *	vm_page_insert_after:
1158  *
1159  *	Inserts the page "m" into the specified object at offset "pindex".
1160  *
1161  *	The page "mpred" must immediately precede the offset "pindex" within
1162  *	the specified object.
1163  *
1164  *	The object must be locked.
1165  */
1166 static int
1167 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1168     vm_page_t mpred)
1169 {
1170 	vm_page_t msucc;
1171 
1172 	VM_OBJECT_ASSERT_WLOCKED(object);
1173 	KASSERT(m->object == NULL,
1174 	    ("vm_page_insert_after: page already inserted"));
1175 	if (mpred != NULL) {
1176 		KASSERT(mpred->object == object,
1177 		    ("vm_page_insert_after: object doesn't contain mpred"));
1178 		KASSERT(mpred->pindex < pindex,
1179 		    ("vm_page_insert_after: mpred doesn't precede pindex"));
1180 		msucc = TAILQ_NEXT(mpred, listq);
1181 	} else
1182 		msucc = TAILQ_FIRST(&object->memq);
1183 	if (msucc != NULL)
1184 		KASSERT(msucc->pindex > pindex,
1185 		    ("vm_page_insert_after: msucc doesn't succeed pindex"));
1186 
1187 	/*
1188 	 * Record the object/offset pair in this page
1189 	 */
1190 	m->object = object;
1191 	m->pindex = pindex;
1192 
1193 	/*
1194 	 * Now link into the object's ordered list of backed pages.
1195 	 */
1196 	if (vm_radix_insert(&object->rtree, m)) {
1197 		m->object = NULL;
1198 		m->pindex = 0;
1199 		return (1);
1200 	}
1201 	vm_page_insert_radixdone(m, object, mpred);
1202 	return (0);
1203 }
1204 
1205 /*
1206  *	vm_page_insert_radixdone:
1207  *
1208  *	Complete page "m" insertion into the specified object after the
1209  *	radix trie hooking.
1210  *
1211  *	The page "mpred" must precede the offset "m->pindex" within the
1212  *	specified object.
1213  *
1214  *	The object must be locked.
1215  */
1216 static void
1217 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1218 {
1219 
1220 	VM_OBJECT_ASSERT_WLOCKED(object);
1221 	KASSERT(object != NULL && m->object == object,
1222 	    ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1223 	if (mpred != NULL) {
1224 		KASSERT(mpred->object == object,
1225 		    ("vm_page_insert_after: object doesn't contain mpred"));
1226 		KASSERT(mpred->pindex < m->pindex,
1227 		    ("vm_page_insert_after: mpred doesn't precede pindex"));
1228 	}
1229 
1230 	if (mpred != NULL)
1231 		TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1232 	else
1233 		TAILQ_INSERT_HEAD(&object->memq, m, listq);
1234 
1235 	/*
1236 	 * Show that the object has one more resident page.
1237 	 */
1238 	object->resident_page_count++;
1239 
1240 	/*
1241 	 * Hold the vnode until the last page is released.
1242 	 */
1243 	if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1244 		vhold(object->handle);
1245 
1246 	/*
1247 	 * Since we are inserting a new and possibly dirty page,
1248 	 * update the object's OBJ_MIGHTBEDIRTY flag.
1249 	 */
1250 	if (pmap_page_is_write_mapped(m))
1251 		vm_object_set_writeable_dirty(object);
1252 }
1253 
1254 /*
1255  *	vm_page_remove:
1256  *
1257  *	Removes the specified page from its containing object, but does not
1258  *	invalidate any backing storage.
1259  *
1260  *	The object must be locked.  The page must be locked if it is managed.
1261  */
1262 void
1263 vm_page_remove(vm_page_t m)
1264 {
1265 	vm_object_t object;
1266 	vm_page_t mrem;
1267 
1268 	if ((m->oflags & VPO_UNMANAGED) == 0)
1269 		vm_page_assert_locked(m);
1270 	if ((object = m->object) == NULL)
1271 		return;
1272 	VM_OBJECT_ASSERT_WLOCKED(object);
1273 	if (vm_page_xbusied(m))
1274 		vm_page_xunbusy_maybelocked(m);
1275 	mrem = vm_radix_remove(&object->rtree, m->pindex);
1276 	KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m));
1277 
1278 	/*
1279 	 * Now remove from the object's list of backed pages.
1280 	 */
1281 	TAILQ_REMOVE(&object->memq, m, listq);
1282 
1283 	/*
1284 	 * And show that the object has one fewer resident page.
1285 	 */
1286 	object->resident_page_count--;
1287 
1288 	/*
1289 	 * The vnode may now be recycled.
1290 	 */
1291 	if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1292 		vdrop(object->handle);
1293 
1294 	m->object = NULL;
1295 }
1296 
1297 /*
1298  *	vm_page_lookup:
1299  *
1300  *	Returns the page associated with the object/offset
1301  *	pair specified; if none is found, NULL is returned.
1302  *
1303  *	The object must be locked.
1304  */
1305 vm_page_t
1306 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1307 {
1308 
1309 	VM_OBJECT_ASSERT_LOCKED(object);
1310 	return (vm_radix_lookup(&object->rtree, pindex));
1311 }
1312 
1313 /*
1314  *	vm_page_find_least:
1315  *
1316  *	Returns the page associated with the object with least pindex
1317  *	greater than or equal to the parameter pindex, or NULL.
1318  *
1319  *	The object must be locked.
1320  */
1321 vm_page_t
1322 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1323 {
1324 	vm_page_t m;
1325 
1326 	VM_OBJECT_ASSERT_LOCKED(object);
1327 	if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1328 		m = vm_radix_lookup_ge(&object->rtree, pindex);
1329 	return (m);
1330 }
1331 
1332 /*
1333  * Returns the given page's successor (by pindex) within the object if it is
1334  * resident; if none is found, NULL is returned.
1335  *
1336  * The object must be locked.
1337  */
1338 vm_page_t
1339 vm_page_next(vm_page_t m)
1340 {
1341 	vm_page_t next;
1342 
1343 	VM_OBJECT_ASSERT_LOCKED(m->object);
1344 	if ((next = TAILQ_NEXT(m, listq)) != NULL) {
1345 		MPASS(next->object == m->object);
1346 		if (next->pindex != m->pindex + 1)
1347 			next = NULL;
1348 	}
1349 	return (next);
1350 }
1351 
1352 /*
1353  * Returns the given page's predecessor (by pindex) within the object if it is
1354  * resident; if none is found, NULL is returned.
1355  *
1356  * The object must be locked.
1357  */
1358 vm_page_t
1359 vm_page_prev(vm_page_t m)
1360 {
1361 	vm_page_t prev;
1362 
1363 	VM_OBJECT_ASSERT_LOCKED(m->object);
1364 	if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) {
1365 		MPASS(prev->object == m->object);
1366 		if (prev->pindex != m->pindex - 1)
1367 			prev = NULL;
1368 	}
1369 	return (prev);
1370 }
1371 
1372 /*
1373  * Uses the page mnew as a replacement for an existing page at index
1374  * pindex which must be already present in the object.
1375  *
1376  * The existing page must not be on a paging queue.
1377  */
1378 vm_page_t
1379 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex)
1380 {
1381 	vm_page_t mold;
1382 
1383 	VM_OBJECT_ASSERT_WLOCKED(object);
1384 	KASSERT(mnew->object == NULL,
1385 	    ("vm_page_replace: page already in object"));
1386 
1387 	/*
1388 	 * This function mostly follows vm_page_insert() and
1389 	 * vm_page_remove() without the radix, object count and vnode
1390 	 * dance.  Double check such functions for more comments.
1391 	 */
1392 
1393 	mnew->object = object;
1394 	mnew->pindex = pindex;
1395 	mold = vm_radix_replace(&object->rtree, mnew);
1396 	KASSERT(mold->queue == PQ_NONE,
1397 	    ("vm_page_replace: mold is on a paging queue"));
1398 
1399 	/* Keep the resident page list in sorted order. */
1400 	TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq);
1401 	TAILQ_REMOVE(&object->memq, mold, listq);
1402 
1403 	mold->object = NULL;
1404 	vm_page_xunbusy_maybelocked(mold);
1405 
1406 	/*
1407 	 * The object's resident_page_count does not change because we have
1408 	 * swapped one page for another, but OBJ_MIGHTBEDIRTY.
1409 	 */
1410 	if (pmap_page_is_write_mapped(mnew))
1411 		vm_object_set_writeable_dirty(object);
1412 	return (mold);
1413 }
1414 
1415 /*
1416  *	vm_page_rename:
1417  *
1418  *	Move the given memory entry from its
1419  *	current object to the specified target object/offset.
1420  *
1421  *	Note: swap associated with the page must be invalidated by the move.  We
1422  *	      have to do this for several reasons:  (1) we aren't freeing the
1423  *	      page, (2) we are dirtying the page, (3) the VM system is probably
1424  *	      moving the page from object A to B, and will then later move
1425  *	      the backing store from A to B and we can't have a conflict.
1426  *
1427  *	Note: we *always* dirty the page.  It is necessary both for the
1428  *	      fact that we moved it, and because we may be invalidating
1429  *	      swap.
1430  *
1431  *	The objects must be locked.
1432  */
1433 int
1434 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1435 {
1436 	vm_page_t mpred;
1437 	vm_pindex_t opidx;
1438 
1439 	VM_OBJECT_ASSERT_WLOCKED(new_object);
1440 
1441 	mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex);
1442 	KASSERT(mpred == NULL || mpred->pindex != new_pindex,
1443 	    ("vm_page_rename: pindex already renamed"));
1444 
1445 	/*
1446 	 * Create a custom version of vm_page_insert() which does not depend
1447 	 * by m_prev and can cheat on the implementation aspects of the
1448 	 * function.
1449 	 */
1450 	opidx = m->pindex;
1451 	m->pindex = new_pindex;
1452 	if (vm_radix_insert(&new_object->rtree, m)) {
1453 		m->pindex = opidx;
1454 		return (1);
1455 	}
1456 
1457 	/*
1458 	 * The operation cannot fail anymore.  The removal must happen before
1459 	 * the listq iterator is tainted.
1460 	 */
1461 	m->pindex = opidx;
1462 	vm_page_lock(m);
1463 	vm_page_remove(m);
1464 
1465 	/* Return back to the new pindex to complete vm_page_insert(). */
1466 	m->pindex = new_pindex;
1467 	m->object = new_object;
1468 	vm_page_unlock(m);
1469 	vm_page_insert_radixdone(m, new_object, mpred);
1470 	vm_page_dirty(m);
1471 	return (0);
1472 }
1473 
1474 /*
1475  *	vm_page_alloc:
1476  *
1477  *	Allocate and return a page that is associated with the specified
1478  *	object and offset pair.  By default, this page is exclusive busied.
1479  *
1480  *	The caller must always specify an allocation class.
1481  *
1482  *	allocation classes:
1483  *	VM_ALLOC_NORMAL		normal process request
1484  *	VM_ALLOC_SYSTEM		system *really* needs a page
1485  *	VM_ALLOC_INTERRUPT	interrupt time request
1486  *
1487  *	optional allocation flags:
1488  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
1489  *				intends to allocate
1490  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
1491  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
1492  *	VM_ALLOC_NOOBJ		page is not associated with an object and
1493  *				should not be exclusive busy
1494  *	VM_ALLOC_SBUSY		shared busy the allocated page
1495  *	VM_ALLOC_WIRED		wire the allocated page
1496  *	VM_ALLOC_ZERO		prefer a zeroed page
1497  *
1498  *	This routine may not sleep.
1499  */
1500 vm_page_t
1501 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1502 {
1503 	vm_page_t m, mpred;
1504 	int flags, req_class;
1505 
1506 	mpred = NULL;	/* XXX: pacify gcc */
1507 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1508 	    (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1509 	    ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1510 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1511 	    ("vm_page_alloc: inconsistent object(%p)/req(%x)", object, req));
1512 	if (object != NULL)
1513 		VM_OBJECT_ASSERT_WLOCKED(object);
1514 
1515 	req_class = req & VM_ALLOC_CLASS_MASK;
1516 
1517 	/*
1518 	 * The page daemon is allowed to dig deeper into the free page list.
1519 	 */
1520 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1521 		req_class = VM_ALLOC_SYSTEM;
1522 
1523 	if (object != NULL) {
1524 		mpred = vm_radix_lookup_le(&object->rtree, pindex);
1525 		KASSERT(mpred == NULL || mpred->pindex != pindex,
1526 		   ("vm_page_alloc: pindex already allocated"));
1527 	}
1528 
1529 	/*
1530 	 * Allocate a page if the number of free pages exceeds the minimum
1531 	 * for the request class.
1532 	 */
1533 	mtx_lock(&vm_page_queue_free_mtx);
1534 	if (vm_cnt.v_free_count > vm_cnt.v_free_reserved ||
1535 	    (req_class == VM_ALLOC_SYSTEM &&
1536 	    vm_cnt.v_free_count > vm_cnt.v_interrupt_free_min) ||
1537 	    (req_class == VM_ALLOC_INTERRUPT &&
1538 	    vm_cnt.v_free_count > 0)) {
1539 		/*
1540 		 * Can we allocate the page from a reservation?
1541 		 */
1542 #if VM_NRESERVLEVEL > 0
1543 		if (object == NULL || (object->flags & (OBJ_COLORED |
1544 		    OBJ_FICTITIOUS)) != OBJ_COLORED || (m =
1545 		    vm_reserv_alloc_page(object, pindex, mpred)) == NULL)
1546 #endif
1547 		{
1548 			/*
1549 			 * If not, allocate it from the free page queues.
1550 			 */
1551 			m = vm_phys_alloc_pages(object != NULL ?
1552 			    VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0);
1553 #if VM_NRESERVLEVEL > 0
1554 			if (m == NULL && vm_reserv_reclaim_inactive()) {
1555 				m = vm_phys_alloc_pages(object != NULL ?
1556 				    VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT,
1557 				    0);
1558 			}
1559 #endif
1560 		}
1561 	} else {
1562 		/*
1563 		 * Not allocatable, give up.
1564 		 */
1565 		mtx_unlock(&vm_page_queue_free_mtx);
1566 		atomic_add_int(&vm_pageout_deficit,
1567 		    max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
1568 		pagedaemon_wakeup();
1569 		return (NULL);
1570 	}
1571 
1572 	/*
1573 	 *  At this point we had better have found a good page.
1574 	 */
1575 	KASSERT(m != NULL, ("vm_page_alloc: missing page"));
1576 	vm_phys_freecnt_adj(m, -1);
1577 	mtx_unlock(&vm_page_queue_free_mtx);
1578 	vm_page_alloc_check(m);
1579 
1580 	/*
1581 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
1582 	 */
1583 	flags = 0;
1584 	if ((req & VM_ALLOC_ZERO) != 0)
1585 		flags = PG_ZERO;
1586 	flags &= m->flags;
1587 	if ((req & VM_ALLOC_NODUMP) != 0)
1588 		flags |= PG_NODUMP;
1589 	m->flags = flags;
1590 	m->aflags = 0;
1591 	m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
1592 	    VPO_UNMANAGED : 0;
1593 	m->busy_lock = VPB_UNBUSIED;
1594 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
1595 		m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1596 	if ((req & VM_ALLOC_SBUSY) != 0)
1597 		m->busy_lock = VPB_SHARERS_WORD(1);
1598 	if (req & VM_ALLOC_WIRED) {
1599 		/*
1600 		 * The page lock is not required for wiring a page until that
1601 		 * page is inserted into the object.
1602 		 */
1603 		atomic_add_int(&vm_cnt.v_wire_count, 1);
1604 		m->wire_count = 1;
1605 	}
1606 	m->act_count = 0;
1607 
1608 	if (object != NULL) {
1609 		if (vm_page_insert_after(m, object, pindex, mpred)) {
1610 			pagedaemon_wakeup();
1611 			if (req & VM_ALLOC_WIRED) {
1612 				atomic_subtract_int(&vm_cnt.v_wire_count, 1);
1613 				m->wire_count = 0;
1614 			}
1615 			KASSERT(m->object == NULL, ("page %p has object", m));
1616 			m->oflags = VPO_UNMANAGED;
1617 			m->busy_lock = VPB_UNBUSIED;
1618 			/* Don't change PG_ZERO. */
1619 			vm_page_free_toq(m);
1620 			return (NULL);
1621 		}
1622 
1623 		/* Ignore device objects; the pager sets "memattr" for them. */
1624 		if (object->memattr != VM_MEMATTR_DEFAULT &&
1625 		    (object->flags & OBJ_FICTITIOUS) == 0)
1626 			pmap_page_set_memattr(m, object->memattr);
1627 	} else
1628 		m->pindex = pindex;
1629 
1630 	/*
1631 	 * Don't wakeup too often - wakeup the pageout daemon when
1632 	 * we would be nearly out of memory.
1633 	 */
1634 	if (vm_paging_needed())
1635 		pagedaemon_wakeup();
1636 
1637 	return (m);
1638 }
1639 
1640 /*
1641  *	vm_page_alloc_contig:
1642  *
1643  *	Allocate a contiguous set of physical pages of the given size "npages"
1644  *	from the free lists.  All of the physical pages must be at or above
1645  *	the given physical address "low" and below the given physical address
1646  *	"high".  The given value "alignment" determines the alignment of the
1647  *	first physical page in the set.  If the given value "boundary" is
1648  *	non-zero, then the set of physical pages cannot cross any physical
1649  *	address boundary that is a multiple of that value.  Both "alignment"
1650  *	and "boundary" must be a power of two.
1651  *
1652  *	If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
1653  *	then the memory attribute setting for the physical pages is configured
1654  *	to the object's memory attribute setting.  Otherwise, the memory
1655  *	attribute setting for the physical pages is configured to "memattr",
1656  *	overriding the object's memory attribute setting.  However, if the
1657  *	object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
1658  *	memory attribute setting for the physical pages cannot be configured
1659  *	to VM_MEMATTR_DEFAULT.
1660  *
1661  *	The specified object may not contain fictitious pages.
1662  *
1663  *	The caller must always specify an allocation class.
1664  *
1665  *	allocation classes:
1666  *	VM_ALLOC_NORMAL		normal process request
1667  *	VM_ALLOC_SYSTEM		system *really* needs a page
1668  *	VM_ALLOC_INTERRUPT	interrupt time request
1669  *
1670  *	optional allocation flags:
1671  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
1672  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
1673  *	VM_ALLOC_NOOBJ		page is not associated with an object and
1674  *				should not be exclusive busy
1675  *	VM_ALLOC_SBUSY		shared busy the allocated page
1676  *	VM_ALLOC_WIRED		wire the allocated page
1677  *	VM_ALLOC_ZERO		prefer a zeroed page
1678  *
1679  *	This routine may not sleep.
1680  */
1681 vm_page_t
1682 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
1683     u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
1684     vm_paddr_t boundary, vm_memattr_t memattr)
1685 {
1686 	vm_page_t m, m_ret, mpred;
1687 	u_int busy_lock, flags, oflags;
1688 	int req_class;
1689 
1690 	mpred = NULL;	/* XXX: pacify gcc */
1691 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1692 	    (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1693 	    ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1694 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1695 	    ("vm_page_alloc_contig: inconsistent object(%p)/req(%x)", object,
1696 	    req));
1697 	if (object != NULL) {
1698 		VM_OBJECT_ASSERT_WLOCKED(object);
1699 		KASSERT((object->flags & OBJ_FICTITIOUS) == 0,
1700 		    ("vm_page_alloc_contig: object %p has fictitious pages",
1701 		    object));
1702 	}
1703 	KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
1704 	req_class = req & VM_ALLOC_CLASS_MASK;
1705 
1706 	/*
1707 	 * The page daemon is allowed to dig deeper into the free page list.
1708 	 */
1709 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1710 		req_class = VM_ALLOC_SYSTEM;
1711 
1712 	if (object != NULL) {
1713 		mpred = vm_radix_lookup_le(&object->rtree, pindex);
1714 		KASSERT(mpred == NULL || mpred->pindex != pindex,
1715 		    ("vm_page_alloc_contig: pindex already allocated"));
1716 	}
1717 
1718 	/*
1719 	 * Can we allocate the pages without the number of free pages falling
1720 	 * below the lower bound for the allocation class?
1721 	 */
1722 	mtx_lock(&vm_page_queue_free_mtx);
1723 	if (vm_cnt.v_free_count >= npages + vm_cnt.v_free_reserved ||
1724 	    (req_class == VM_ALLOC_SYSTEM &&
1725 	    vm_cnt.v_free_count >= npages + vm_cnt.v_interrupt_free_min) ||
1726 	    (req_class == VM_ALLOC_INTERRUPT &&
1727 	    vm_cnt.v_free_count >= npages)) {
1728 		/*
1729 		 * Can we allocate the pages from a reservation?
1730 		 */
1731 #if VM_NRESERVLEVEL > 0
1732 retry:
1733 		if (object == NULL || (object->flags & OBJ_COLORED) == 0 ||
1734 		    (m_ret = vm_reserv_alloc_contig(object, pindex, npages,
1735 		    low, high, alignment, boundary, mpred)) == NULL)
1736 #endif
1737 			/*
1738 			 * If not, allocate them from the free page queues.
1739 			 */
1740 			m_ret = vm_phys_alloc_contig(npages, low, high,
1741 			    alignment, boundary);
1742 	} else {
1743 		mtx_unlock(&vm_page_queue_free_mtx);
1744 		atomic_add_int(&vm_pageout_deficit, npages);
1745 		pagedaemon_wakeup();
1746 		return (NULL);
1747 	}
1748 	if (m_ret != NULL)
1749 		vm_phys_freecnt_adj(m_ret, -npages);
1750 	else {
1751 #if VM_NRESERVLEVEL > 0
1752 		if (vm_reserv_reclaim_contig(npages, low, high, alignment,
1753 		    boundary))
1754 			goto retry;
1755 #endif
1756 	}
1757 	mtx_unlock(&vm_page_queue_free_mtx);
1758 	if (m_ret == NULL)
1759 		return (NULL);
1760 	for (m = m_ret; m < &m_ret[npages]; m++)
1761 		vm_page_alloc_check(m);
1762 
1763 	/*
1764 	 * Initialize the pages.  Only the PG_ZERO flag is inherited.
1765 	 */
1766 	flags = 0;
1767 	if ((req & VM_ALLOC_ZERO) != 0)
1768 		flags = PG_ZERO;
1769 	if ((req & VM_ALLOC_NODUMP) != 0)
1770 		flags |= PG_NODUMP;
1771 	oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
1772 	    VPO_UNMANAGED : 0;
1773 	busy_lock = VPB_UNBUSIED;
1774 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
1775 		busy_lock = VPB_SINGLE_EXCLUSIVER;
1776 	if ((req & VM_ALLOC_SBUSY) != 0)
1777 		busy_lock = VPB_SHARERS_WORD(1);
1778 	if ((req & VM_ALLOC_WIRED) != 0)
1779 		atomic_add_int(&vm_cnt.v_wire_count, npages);
1780 	if (object != NULL) {
1781 		if (object->memattr != VM_MEMATTR_DEFAULT &&
1782 		    memattr == VM_MEMATTR_DEFAULT)
1783 			memattr = object->memattr;
1784 	}
1785 	for (m = m_ret; m < &m_ret[npages]; m++) {
1786 		m->aflags = 0;
1787 		m->flags = (m->flags | PG_NODUMP) & flags;
1788 		m->busy_lock = busy_lock;
1789 		if ((req & VM_ALLOC_WIRED) != 0)
1790 			m->wire_count = 1;
1791 		m->act_count = 0;
1792 		m->oflags = oflags;
1793 		if (object != NULL) {
1794 			if (vm_page_insert_after(m, object, pindex, mpred)) {
1795 				pagedaemon_wakeup();
1796 				if ((req & VM_ALLOC_WIRED) != 0)
1797 					atomic_subtract_int(
1798 					    &vm_cnt.v_wire_count, npages);
1799 				KASSERT(m->object == NULL,
1800 				    ("page %p has object", m));
1801 				mpred = m;
1802 				for (m = m_ret; m < &m_ret[npages]; m++) {
1803 					if (m <= mpred &&
1804 					    (req & VM_ALLOC_WIRED) != 0)
1805 						m->wire_count = 0;
1806 					m->oflags = VPO_UNMANAGED;
1807 					m->busy_lock = VPB_UNBUSIED;
1808 					/* Don't change PG_ZERO. */
1809 					vm_page_free_toq(m);
1810 				}
1811 				return (NULL);
1812 			}
1813 			mpred = m;
1814 		} else
1815 			m->pindex = pindex;
1816 		if (memattr != VM_MEMATTR_DEFAULT)
1817 			pmap_page_set_memattr(m, memattr);
1818 		pindex++;
1819 	}
1820 	if (vm_paging_needed())
1821 		pagedaemon_wakeup();
1822 	return (m_ret);
1823 }
1824 
1825 /*
1826  * Check a page that has been freshly dequeued from a freelist.
1827  */
1828 static void
1829 vm_page_alloc_check(vm_page_t m)
1830 {
1831 
1832 	KASSERT(m->object == NULL, ("page %p has object", m));
1833 	KASSERT(m->queue == PQ_NONE,
1834 	    ("page %p has unexpected queue %d", m, m->queue));
1835 	KASSERT(m->wire_count == 0, ("page %p is wired", m));
1836 	KASSERT(m->hold_count == 0, ("page %p is held", m));
1837 	KASSERT(!vm_page_busied(m), ("page %p is busy", m));
1838 	KASSERT(m->dirty == 0, ("page %p is dirty", m));
1839 	KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
1840 	    ("page %p has unexpected memattr %d",
1841 	    m, pmap_page_get_memattr(m)));
1842 	KASSERT(m->valid == 0, ("free page %p is valid", m));
1843 }
1844 
1845 /*
1846  * 	vm_page_alloc_freelist:
1847  *
1848  *	Allocate a physical page from the specified free page list.
1849  *
1850  *	The caller must always specify an allocation class.
1851  *
1852  *	allocation classes:
1853  *	VM_ALLOC_NORMAL		normal process request
1854  *	VM_ALLOC_SYSTEM		system *really* needs a page
1855  *	VM_ALLOC_INTERRUPT	interrupt time request
1856  *
1857  *	optional allocation flags:
1858  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
1859  *				intends to allocate
1860  *	VM_ALLOC_WIRED		wire the allocated page
1861  *	VM_ALLOC_ZERO		prefer a zeroed page
1862  *
1863  *	This routine may not sleep.
1864  */
1865 vm_page_t
1866 vm_page_alloc_freelist(int flind, int req)
1867 {
1868 	vm_page_t m;
1869 	u_int flags;
1870 	int req_class;
1871 
1872 	req_class = req & VM_ALLOC_CLASS_MASK;
1873 
1874 	/*
1875 	 * The page daemon is allowed to dig deeper into the free page list.
1876 	 */
1877 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1878 		req_class = VM_ALLOC_SYSTEM;
1879 
1880 	/*
1881 	 * Do not allocate reserved pages unless the req has asked for it.
1882 	 */
1883 	mtx_lock(&vm_page_queue_free_mtx);
1884 	if (vm_cnt.v_free_count > vm_cnt.v_free_reserved ||
1885 	    (req_class == VM_ALLOC_SYSTEM &&
1886 	    vm_cnt.v_free_count > vm_cnt.v_interrupt_free_min) ||
1887 	    (req_class == VM_ALLOC_INTERRUPT &&
1888 	    vm_cnt.v_free_count > 0))
1889 		m = vm_phys_alloc_freelist_pages(flind, VM_FREEPOOL_DIRECT, 0);
1890 	else {
1891 		mtx_unlock(&vm_page_queue_free_mtx);
1892 		atomic_add_int(&vm_pageout_deficit,
1893 		    max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
1894 		pagedaemon_wakeup();
1895 		return (NULL);
1896 	}
1897 	if (m == NULL) {
1898 		mtx_unlock(&vm_page_queue_free_mtx);
1899 		return (NULL);
1900 	}
1901 	vm_phys_freecnt_adj(m, -1);
1902 	mtx_unlock(&vm_page_queue_free_mtx);
1903 	vm_page_alloc_check(m);
1904 
1905 	/*
1906 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
1907 	 */
1908 	m->aflags = 0;
1909 	flags = 0;
1910 	if ((req & VM_ALLOC_ZERO) != 0)
1911 		flags = PG_ZERO;
1912 	m->flags &= flags;
1913 	if ((req & VM_ALLOC_WIRED) != 0) {
1914 		/*
1915 		 * The page lock is not required for wiring a page that does
1916 		 * not belong to an object.
1917 		 */
1918 		atomic_add_int(&vm_cnt.v_wire_count, 1);
1919 		m->wire_count = 1;
1920 	}
1921 	/* Unmanaged pages don't use "act_count". */
1922 	m->oflags = VPO_UNMANAGED;
1923 	if (vm_paging_needed())
1924 		pagedaemon_wakeup();
1925 	return (m);
1926 }
1927 
1928 #define	VPSC_ANY	0	/* No restrictions. */
1929 #define	VPSC_NORESERV	1	/* Skip reservations; implies VPSC_NOSUPER. */
1930 #define	VPSC_NOSUPER	2	/* Skip superpages. */
1931 
1932 /*
1933  *	vm_page_scan_contig:
1934  *
1935  *	Scan vm_page_array[] between the specified entries "m_start" and
1936  *	"m_end" for a run of contiguous physical pages that satisfy the
1937  *	specified conditions, and return the lowest page in the run.  The
1938  *	specified "alignment" determines the alignment of the lowest physical
1939  *	page in the run.  If the specified "boundary" is non-zero, then the
1940  *	run of physical pages cannot span a physical address that is a
1941  *	multiple of "boundary".
1942  *
1943  *	"m_end" is never dereferenced, so it need not point to a vm_page
1944  *	structure within vm_page_array[].
1945  *
1946  *	"npages" must be greater than zero.  "m_start" and "m_end" must not
1947  *	span a hole (or discontiguity) in the physical address space.  Both
1948  *	"alignment" and "boundary" must be a power of two.
1949  */
1950 vm_page_t
1951 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end,
1952     u_long alignment, vm_paddr_t boundary, int options)
1953 {
1954 	struct mtx *m_mtx, *new_mtx;
1955 	vm_object_t object;
1956 	vm_paddr_t pa;
1957 	vm_page_t m, m_run;
1958 #if VM_NRESERVLEVEL > 0
1959 	int level;
1960 #endif
1961 	int m_inc, order, run_ext, run_len;
1962 
1963 	KASSERT(npages > 0, ("npages is 0"));
1964 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
1965 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
1966 	m_run = NULL;
1967 	run_len = 0;
1968 	m_mtx = NULL;
1969 	for (m = m_start; m < m_end && run_len < npages; m += m_inc) {
1970 		KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
1971 		    ("page %p is PG_FICTITIOUS or PG_MARKER", m));
1972 
1973 		/*
1974 		 * If the current page would be the start of a run, check its
1975 		 * physical address against the end, alignment, and boundary
1976 		 * conditions.  If it doesn't satisfy these conditions, either
1977 		 * terminate the scan or advance to the next page that
1978 		 * satisfies the failed condition.
1979 		 */
1980 		if (run_len == 0) {
1981 			KASSERT(m_run == NULL, ("m_run != NULL"));
1982 			if (m + npages > m_end)
1983 				break;
1984 			pa = VM_PAGE_TO_PHYS(m);
1985 			if ((pa & (alignment - 1)) != 0) {
1986 				m_inc = atop(roundup2(pa, alignment) - pa);
1987 				continue;
1988 			}
1989 			if (rounddown2(pa ^ (pa + ptoa(npages) - 1),
1990 			    boundary) != 0) {
1991 				m_inc = atop(roundup2(pa, boundary) - pa);
1992 				continue;
1993 			}
1994 		} else
1995 			KASSERT(m_run != NULL, ("m_run == NULL"));
1996 
1997 		/*
1998 		 * Avoid releasing and reacquiring the same page lock.
1999 		 */
2000 		new_mtx = vm_page_lockptr(m);
2001 		if (m_mtx != new_mtx) {
2002 			if (m_mtx != NULL)
2003 				mtx_unlock(m_mtx);
2004 			m_mtx = new_mtx;
2005 			mtx_lock(m_mtx);
2006 		}
2007 		m_inc = 1;
2008 retry:
2009 		if (m->wire_count != 0 || m->hold_count != 0)
2010 			run_ext = 0;
2011 #if VM_NRESERVLEVEL > 0
2012 		else if ((level = vm_reserv_level(m)) >= 0 &&
2013 		    (options & VPSC_NORESERV) != 0) {
2014 			run_ext = 0;
2015 			/* Advance to the end of the reservation. */
2016 			pa = VM_PAGE_TO_PHYS(m);
2017 			m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) -
2018 			    pa);
2019 		}
2020 #endif
2021 		else if ((object = m->object) != NULL) {
2022 			/*
2023 			 * The page is considered eligible for relocation if
2024 			 * and only if it could be laundered or reclaimed by
2025 			 * the page daemon.
2026 			 */
2027 			if (!VM_OBJECT_TRYRLOCK(object)) {
2028 				mtx_unlock(m_mtx);
2029 				VM_OBJECT_RLOCK(object);
2030 				mtx_lock(m_mtx);
2031 				if (m->object != object) {
2032 					/*
2033 					 * The page may have been freed.
2034 					 */
2035 					VM_OBJECT_RUNLOCK(object);
2036 					goto retry;
2037 				} else if (m->wire_count != 0 ||
2038 				    m->hold_count != 0) {
2039 					run_ext = 0;
2040 					goto unlock;
2041 				}
2042 			}
2043 			KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2044 			    ("page %p is PG_UNHOLDFREE", m));
2045 			/* Don't care: PG_NODUMP, PG_ZERO. */
2046 			if (object->type != OBJT_DEFAULT &&
2047 			    object->type != OBJT_SWAP &&
2048 			    object->type != OBJT_VNODE) {
2049 				run_ext = 0;
2050 #if VM_NRESERVLEVEL > 0
2051 			} else if ((options & VPSC_NOSUPER) != 0 &&
2052 			    (level = vm_reserv_level_iffullpop(m)) >= 0) {
2053 				run_ext = 0;
2054 				/* Advance to the end of the superpage. */
2055 				pa = VM_PAGE_TO_PHYS(m);
2056 				m_inc = atop(roundup2(pa + 1,
2057 				    vm_reserv_size(level)) - pa);
2058 #endif
2059 			} else if (object->memattr == VM_MEMATTR_DEFAULT &&
2060 			    m->queue != PQ_NONE && !vm_page_busied(m)) {
2061 				/*
2062 				 * The page is allocated but eligible for
2063 				 * relocation.  Extend the current run by one
2064 				 * page.
2065 				 */
2066 				KASSERT(pmap_page_get_memattr(m) ==
2067 				    VM_MEMATTR_DEFAULT,
2068 				    ("page %p has an unexpected memattr", m));
2069 				KASSERT((m->oflags & (VPO_SWAPINPROG |
2070 				    VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2071 				    ("page %p has unexpected oflags", m));
2072 				/* Don't care: VPO_NOSYNC. */
2073 				run_ext = 1;
2074 			} else
2075 				run_ext = 0;
2076 unlock:
2077 			VM_OBJECT_RUNLOCK(object);
2078 #if VM_NRESERVLEVEL > 0
2079 		} else if (level >= 0) {
2080 			/*
2081 			 * The page is reserved but not yet allocated.  In
2082 			 * other words, it is still free.  Extend the current
2083 			 * run by one page.
2084 			 */
2085 			run_ext = 1;
2086 #endif
2087 		} else if ((order = m->order) < VM_NFREEORDER) {
2088 			/*
2089 			 * The page is enqueued in the physical memory
2090 			 * allocator's free page queues.  Moreover, it is the
2091 			 * first page in a power-of-two-sized run of
2092 			 * contiguous free pages.  Add these pages to the end
2093 			 * of the current run, and jump ahead.
2094 			 */
2095 			run_ext = 1 << order;
2096 			m_inc = 1 << order;
2097 		} else {
2098 			/*
2099 			 * Skip the page for one of the following reasons: (1)
2100 			 * It is enqueued in the physical memory allocator's
2101 			 * free page queues.  However, it is not the first
2102 			 * page in a run of contiguous free pages.  (This case
2103 			 * rarely occurs because the scan is performed in
2104 			 * ascending order.) (2) It is not reserved, and it is
2105 			 * transitioning from free to allocated.  (Conversely,
2106 			 * the transition from allocated to free for managed
2107 			 * pages is blocked by the page lock.) (3) It is
2108 			 * allocated but not contained by an object and not
2109 			 * wired, e.g., allocated by Xen's balloon driver.
2110 			 */
2111 			run_ext = 0;
2112 		}
2113 
2114 		/*
2115 		 * Extend or reset the current run of pages.
2116 		 */
2117 		if (run_ext > 0) {
2118 			if (run_len == 0)
2119 				m_run = m;
2120 			run_len += run_ext;
2121 		} else {
2122 			if (run_len > 0) {
2123 				m_run = NULL;
2124 				run_len = 0;
2125 			}
2126 		}
2127 	}
2128 	if (m_mtx != NULL)
2129 		mtx_unlock(m_mtx);
2130 	if (run_len >= npages)
2131 		return (m_run);
2132 	return (NULL);
2133 }
2134 
2135 /*
2136  *	vm_page_reclaim_run:
2137  *
2138  *	Try to relocate each of the allocated virtual pages within the
2139  *	specified run of physical pages to a new physical address.  Free the
2140  *	physical pages underlying the relocated virtual pages.  A virtual page
2141  *	is relocatable if and only if it could be laundered or reclaimed by
2142  *	the page daemon.  Whenever possible, a virtual page is relocated to a
2143  *	physical address above "high".
2144  *
2145  *	Returns 0 if every physical page within the run was already free or
2146  *	just freed by a successful relocation.  Otherwise, returns a non-zero
2147  *	value indicating why the last attempt to relocate a virtual page was
2148  *	unsuccessful.
2149  *
2150  *	"req_class" must be an allocation class.
2151  */
2152 static int
2153 vm_page_reclaim_run(int req_class, u_long npages, vm_page_t m_run,
2154     vm_paddr_t high)
2155 {
2156 	struct mtx *m_mtx, *new_mtx;
2157 	struct spglist free;
2158 	vm_object_t object;
2159 	vm_paddr_t pa;
2160 	vm_page_t m, m_end, m_new;
2161 	int error, order, req;
2162 
2163 	KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class,
2164 	    ("req_class is not an allocation class"));
2165 	SLIST_INIT(&free);
2166 	error = 0;
2167 	m = m_run;
2168 	m_end = m_run + npages;
2169 	m_mtx = NULL;
2170 	for (; error == 0 && m < m_end; m++) {
2171 		KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2172 		    ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2173 
2174 		/*
2175 		 * Avoid releasing and reacquiring the same page lock.
2176 		 */
2177 		new_mtx = vm_page_lockptr(m);
2178 		if (m_mtx != new_mtx) {
2179 			if (m_mtx != NULL)
2180 				mtx_unlock(m_mtx);
2181 			m_mtx = new_mtx;
2182 			mtx_lock(m_mtx);
2183 		}
2184 retry:
2185 		if (m->wire_count != 0 || m->hold_count != 0)
2186 			error = EBUSY;
2187 		else if ((object = m->object) != NULL) {
2188 			/*
2189 			 * The page is relocated if and only if it could be
2190 			 * laundered or reclaimed by the page daemon.
2191 			 */
2192 			if (!VM_OBJECT_TRYWLOCK(object)) {
2193 				mtx_unlock(m_mtx);
2194 				VM_OBJECT_WLOCK(object);
2195 				mtx_lock(m_mtx);
2196 				if (m->object != object) {
2197 					/*
2198 					 * The page may have been freed.
2199 					 */
2200 					VM_OBJECT_WUNLOCK(object);
2201 					goto retry;
2202 				} else if (m->wire_count != 0 ||
2203 				    m->hold_count != 0) {
2204 					error = EBUSY;
2205 					goto unlock;
2206 				}
2207 			}
2208 			KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2209 			    ("page %p is PG_UNHOLDFREE", m));
2210 			/* Don't care: PG_NODUMP, PG_ZERO. */
2211 			if (object->type != OBJT_DEFAULT &&
2212 			    object->type != OBJT_SWAP &&
2213 			    object->type != OBJT_VNODE)
2214 				error = EINVAL;
2215 			else if (object->memattr != VM_MEMATTR_DEFAULT)
2216 				error = EINVAL;
2217 			else if (m->queue != PQ_NONE && !vm_page_busied(m)) {
2218 				KASSERT(pmap_page_get_memattr(m) ==
2219 				    VM_MEMATTR_DEFAULT,
2220 				    ("page %p has an unexpected memattr", m));
2221 				KASSERT((m->oflags & (VPO_SWAPINPROG |
2222 				    VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2223 				    ("page %p has unexpected oflags", m));
2224 				/* Don't care: VPO_NOSYNC. */
2225 				if (m->valid != 0) {
2226 					/*
2227 					 * First, try to allocate a new page
2228 					 * that is above "high".  Failing
2229 					 * that, try to allocate a new page
2230 					 * that is below "m_run".  Allocate
2231 					 * the new page between the end of
2232 					 * "m_run" and "high" only as a last
2233 					 * resort.
2234 					 */
2235 					req = req_class | VM_ALLOC_NOOBJ;
2236 					if ((m->flags & PG_NODUMP) != 0)
2237 						req |= VM_ALLOC_NODUMP;
2238 					if (trunc_page(high) !=
2239 					    ~(vm_paddr_t)PAGE_MASK) {
2240 						m_new = vm_page_alloc_contig(
2241 						    NULL, 0, req, 1,
2242 						    round_page(high),
2243 						    ~(vm_paddr_t)0,
2244 						    PAGE_SIZE, 0,
2245 						    VM_MEMATTR_DEFAULT);
2246 					} else
2247 						m_new = NULL;
2248 					if (m_new == NULL) {
2249 						pa = VM_PAGE_TO_PHYS(m_run);
2250 						m_new = vm_page_alloc_contig(
2251 						    NULL, 0, req, 1,
2252 						    0, pa - 1, PAGE_SIZE, 0,
2253 						    VM_MEMATTR_DEFAULT);
2254 					}
2255 					if (m_new == NULL) {
2256 						pa += ptoa(npages);
2257 						m_new = vm_page_alloc_contig(
2258 						    NULL, 0, req, 1,
2259 						    pa, high, PAGE_SIZE, 0,
2260 						    VM_MEMATTR_DEFAULT);
2261 					}
2262 					if (m_new == NULL) {
2263 						error = ENOMEM;
2264 						goto unlock;
2265 					}
2266 					KASSERT(m_new->wire_count == 0,
2267 					    ("page %p is wired", m));
2268 
2269 					/*
2270 					 * Replace "m" with the new page.  For
2271 					 * vm_page_replace(), "m" must be busy
2272 					 * and dequeued.  Finally, change "m"
2273 					 * as if vm_page_free() was called.
2274 					 */
2275 					if (object->ref_count != 0)
2276 						pmap_remove_all(m);
2277 					m_new->aflags = m->aflags;
2278 					KASSERT(m_new->oflags == VPO_UNMANAGED,
2279 					    ("page %p is managed", m));
2280 					m_new->oflags = m->oflags & VPO_NOSYNC;
2281 					pmap_copy_page(m, m_new);
2282 					m_new->valid = m->valid;
2283 					m_new->dirty = m->dirty;
2284 					m->flags &= ~PG_ZERO;
2285 					vm_page_xbusy(m);
2286 					vm_page_remque(m);
2287 					vm_page_replace_checked(m_new, object,
2288 					    m->pindex, m);
2289 					m->valid = 0;
2290 					vm_page_undirty(m);
2291 
2292 					/*
2293 					 * The new page must be deactivated
2294 					 * before the object is unlocked.
2295 					 */
2296 					new_mtx = vm_page_lockptr(m_new);
2297 					if (m_mtx != new_mtx) {
2298 						mtx_unlock(m_mtx);
2299 						m_mtx = new_mtx;
2300 						mtx_lock(m_mtx);
2301 					}
2302 					vm_page_deactivate(m_new);
2303 				} else {
2304 					m->flags &= ~PG_ZERO;
2305 					vm_page_remque(m);
2306 					vm_page_remove(m);
2307 					KASSERT(m->dirty == 0,
2308 					    ("page %p is dirty", m));
2309 				}
2310 				SLIST_INSERT_HEAD(&free, m, plinks.s.ss);
2311 			} else
2312 				error = EBUSY;
2313 unlock:
2314 			VM_OBJECT_WUNLOCK(object);
2315 		} else {
2316 			mtx_lock(&vm_page_queue_free_mtx);
2317 			order = m->order;
2318 			if (order < VM_NFREEORDER) {
2319 				/*
2320 				 * The page is enqueued in the physical memory
2321 				 * allocator's free page queues.  Moreover, it
2322 				 * is the first page in a power-of-two-sized
2323 				 * run of contiguous free pages.  Jump ahead
2324 				 * to the last page within that run, and
2325 				 * continue from there.
2326 				 */
2327 				m += (1 << order) - 1;
2328 			}
2329 #if VM_NRESERVLEVEL > 0
2330 			else if (vm_reserv_is_page_free(m))
2331 				order = 0;
2332 #endif
2333 			mtx_unlock(&vm_page_queue_free_mtx);
2334 			if (order == VM_NFREEORDER)
2335 				error = EINVAL;
2336 		}
2337 	}
2338 	if (m_mtx != NULL)
2339 		mtx_unlock(m_mtx);
2340 	if ((m = SLIST_FIRST(&free)) != NULL) {
2341 		mtx_lock(&vm_page_queue_free_mtx);
2342 		do {
2343 			SLIST_REMOVE_HEAD(&free, plinks.s.ss);
2344 			vm_phys_freecnt_adj(m, 1);
2345 #if VM_NRESERVLEVEL > 0
2346 			if (!vm_reserv_free_page(m))
2347 #else
2348 			if (true)
2349 #endif
2350 				vm_phys_free_pages(m, 0);
2351 		} while ((m = SLIST_FIRST(&free)) != NULL);
2352 		vm_page_free_wakeup();
2353 		mtx_unlock(&vm_page_queue_free_mtx);
2354 	}
2355 	return (error);
2356 }
2357 
2358 #define	NRUNS	16
2359 
2360 CTASSERT(powerof2(NRUNS));
2361 
2362 #define	RUN_INDEX(count)	((count) & (NRUNS - 1))
2363 
2364 #define	MIN_RECLAIM	8
2365 
2366 /*
2367  *	vm_page_reclaim_contig:
2368  *
2369  *	Reclaim allocated, contiguous physical memory satisfying the specified
2370  *	conditions by relocating the virtual pages using that physical memory.
2371  *	Returns true if reclamation is successful and false otherwise.  Since
2372  *	relocation requires the allocation of physical pages, reclamation may
2373  *	fail due to a shortage of free pages.  When reclamation fails, callers
2374  *	are expected to perform VM_WAIT before retrying a failed allocation
2375  *	operation, e.g., vm_page_alloc_contig().
2376  *
2377  *	The caller must always specify an allocation class through "req".
2378  *
2379  *	allocation classes:
2380  *	VM_ALLOC_NORMAL		normal process request
2381  *	VM_ALLOC_SYSTEM		system *really* needs a page
2382  *	VM_ALLOC_INTERRUPT	interrupt time request
2383  *
2384  *	The optional allocation flags are ignored.
2385  *
2386  *	"npages" must be greater than zero.  Both "alignment" and "boundary"
2387  *	must be a power of two.
2388  */
2389 bool
2390 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high,
2391     u_long alignment, vm_paddr_t boundary)
2392 {
2393 	vm_paddr_t curr_low;
2394 	vm_page_t m_run, m_runs[NRUNS];
2395 	u_long count, reclaimed;
2396 	int error, i, options, req_class;
2397 
2398 	KASSERT(npages > 0, ("npages is 0"));
2399 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2400 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2401 	req_class = req & VM_ALLOC_CLASS_MASK;
2402 
2403 	/*
2404 	 * The page daemon is allowed to dig deeper into the free page list.
2405 	 */
2406 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2407 		req_class = VM_ALLOC_SYSTEM;
2408 
2409 	/*
2410 	 * Return if the number of free pages cannot satisfy the requested
2411 	 * allocation.
2412 	 */
2413 	count = vm_cnt.v_free_count;
2414 	if (count < npages + vm_cnt.v_free_reserved || (count < npages +
2415 	    vm_cnt.v_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) ||
2416 	    (count < npages && req_class == VM_ALLOC_INTERRUPT))
2417 		return (false);
2418 
2419 	/*
2420 	 * Scan up to three times, relaxing the restrictions ("options") on
2421 	 * the reclamation of reservations and superpages each time.
2422 	 */
2423 	for (options = VPSC_NORESERV;;) {
2424 		/*
2425 		 * Find the highest runs that satisfy the given constraints
2426 		 * and restrictions, and record them in "m_runs".
2427 		 */
2428 		curr_low = low;
2429 		count = 0;
2430 		for (;;) {
2431 			m_run = vm_phys_scan_contig(npages, curr_low, high,
2432 			    alignment, boundary, options);
2433 			if (m_run == NULL)
2434 				break;
2435 			curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages);
2436 			m_runs[RUN_INDEX(count)] = m_run;
2437 			count++;
2438 		}
2439 
2440 		/*
2441 		 * Reclaim the highest runs in LIFO (descending) order until
2442 		 * the number of reclaimed pages, "reclaimed", is at least
2443 		 * MIN_RECLAIM.  Reset "reclaimed" each time because each
2444 		 * reclamation is idempotent, and runs will (likely) recur
2445 		 * from one scan to the next as restrictions are relaxed.
2446 		 */
2447 		reclaimed = 0;
2448 		for (i = 0; count > 0 && i < NRUNS; i++) {
2449 			count--;
2450 			m_run = m_runs[RUN_INDEX(count)];
2451 			error = vm_page_reclaim_run(req_class, npages, m_run,
2452 			    high);
2453 			if (error == 0) {
2454 				reclaimed += npages;
2455 				if (reclaimed >= MIN_RECLAIM)
2456 					return (true);
2457 			}
2458 		}
2459 
2460 		/*
2461 		 * Either relax the restrictions on the next scan or return if
2462 		 * the last scan had no restrictions.
2463 		 */
2464 		if (options == VPSC_NORESERV)
2465 			options = VPSC_NOSUPER;
2466 		else if (options == VPSC_NOSUPER)
2467 			options = VPSC_ANY;
2468 		else if (options == VPSC_ANY)
2469 			return (reclaimed != 0);
2470 	}
2471 }
2472 
2473 /*
2474  *	vm_wait:	(also see VM_WAIT macro)
2475  *
2476  *	Sleep until free pages are available for allocation.
2477  *	- Called in various places before memory allocations.
2478  */
2479 void
2480 vm_wait(void)
2481 {
2482 
2483 	mtx_lock(&vm_page_queue_free_mtx);
2484 	if (curproc == pageproc) {
2485 		vm_pageout_pages_needed = 1;
2486 		msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx,
2487 		    PDROP | PSWP, "VMWait", 0);
2488 	} else {
2489 		if (__predict_false(pageproc == NULL))
2490 			panic("vm_wait in early boot");
2491 		if (!vm_pageout_wanted) {
2492 			vm_pageout_wanted = true;
2493 			wakeup(&vm_pageout_wanted);
2494 		}
2495 		vm_pages_needed = true;
2496 		msleep(&vm_cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PVM,
2497 		    "vmwait", 0);
2498 	}
2499 }
2500 
2501 /*
2502  *	vm_waitpfault:	(also see VM_WAITPFAULT macro)
2503  *
2504  *	Sleep until free pages are available for allocation.
2505  *	- Called only in vm_fault so that processes page faulting
2506  *	  can be easily tracked.
2507  *	- Sleeps at a lower priority than vm_wait() so that vm_wait()ing
2508  *	  processes will be able to grab memory first.  Do not change
2509  *	  this balance without careful testing first.
2510  */
2511 void
2512 vm_waitpfault(void)
2513 {
2514 
2515 	mtx_lock(&vm_page_queue_free_mtx);
2516 	if (!vm_pageout_wanted) {
2517 		vm_pageout_wanted = true;
2518 		wakeup(&vm_pageout_wanted);
2519 	}
2520 	vm_pages_needed = true;
2521 	msleep(&vm_cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PUSER,
2522 	    "pfault", 0);
2523 }
2524 
2525 struct vm_pagequeue *
2526 vm_page_pagequeue(vm_page_t m)
2527 {
2528 
2529 	if (vm_page_in_laundry(m))
2530 		return (&vm_dom[0].vmd_pagequeues[m->queue]);
2531 	else
2532 		return (&vm_phys_domain(m)->vmd_pagequeues[m->queue]);
2533 }
2534 
2535 /*
2536  *	vm_page_dequeue:
2537  *
2538  *	Remove the given page from its current page queue.
2539  *
2540  *	The page must be locked.
2541  */
2542 void
2543 vm_page_dequeue(vm_page_t m)
2544 {
2545 	struct vm_pagequeue *pq;
2546 
2547 	vm_page_assert_locked(m);
2548 	KASSERT(m->queue < PQ_COUNT, ("vm_page_dequeue: page %p is not queued",
2549 	    m));
2550 	pq = vm_page_pagequeue(m);
2551 	vm_pagequeue_lock(pq);
2552 	m->queue = PQ_NONE;
2553 	TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2554 	vm_pagequeue_cnt_dec(pq);
2555 	vm_pagequeue_unlock(pq);
2556 }
2557 
2558 /*
2559  *	vm_page_dequeue_locked:
2560  *
2561  *	Remove the given page from its current page queue.
2562  *
2563  *	The page and page queue must be locked.
2564  */
2565 void
2566 vm_page_dequeue_locked(vm_page_t m)
2567 {
2568 	struct vm_pagequeue *pq;
2569 
2570 	vm_page_lock_assert(m, MA_OWNED);
2571 	pq = vm_page_pagequeue(m);
2572 	vm_pagequeue_assert_locked(pq);
2573 	m->queue = PQ_NONE;
2574 	TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2575 	vm_pagequeue_cnt_dec(pq);
2576 }
2577 
2578 /*
2579  *	vm_page_enqueue:
2580  *
2581  *	Add the given page to the specified page queue.
2582  *
2583  *	The page must be locked.
2584  */
2585 static void
2586 vm_page_enqueue(uint8_t queue, vm_page_t m)
2587 {
2588 	struct vm_pagequeue *pq;
2589 
2590 	vm_page_lock_assert(m, MA_OWNED);
2591 	KASSERT(queue < PQ_COUNT,
2592 	    ("vm_page_enqueue: invalid queue %u request for page %p",
2593 	    queue, m));
2594 	if (queue == PQ_LAUNDRY || queue == PQ_UNSWAPPABLE)
2595 		pq = &vm_dom[0].vmd_pagequeues[queue];
2596 	else
2597 		pq = &vm_phys_domain(m)->vmd_pagequeues[queue];
2598 	vm_pagequeue_lock(pq);
2599 	m->queue = queue;
2600 	TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2601 	vm_pagequeue_cnt_inc(pq);
2602 	vm_pagequeue_unlock(pq);
2603 }
2604 
2605 /*
2606  *	vm_page_requeue:
2607  *
2608  *	Move the given page to the tail of its current page queue.
2609  *
2610  *	The page must be locked.
2611  */
2612 void
2613 vm_page_requeue(vm_page_t m)
2614 {
2615 	struct vm_pagequeue *pq;
2616 
2617 	vm_page_lock_assert(m, MA_OWNED);
2618 	KASSERT(m->queue != PQ_NONE,
2619 	    ("vm_page_requeue: page %p is not queued", m));
2620 	pq = vm_page_pagequeue(m);
2621 	vm_pagequeue_lock(pq);
2622 	TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2623 	TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2624 	vm_pagequeue_unlock(pq);
2625 }
2626 
2627 /*
2628  *	vm_page_requeue_locked:
2629  *
2630  *	Move the given page to the tail of its current page queue.
2631  *
2632  *	The page queue must be locked.
2633  */
2634 void
2635 vm_page_requeue_locked(vm_page_t m)
2636 {
2637 	struct vm_pagequeue *pq;
2638 
2639 	KASSERT(m->queue != PQ_NONE,
2640 	    ("vm_page_requeue_locked: page %p is not queued", m));
2641 	pq = vm_page_pagequeue(m);
2642 	vm_pagequeue_assert_locked(pq);
2643 	TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2644 	TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2645 }
2646 
2647 /*
2648  *	vm_page_activate:
2649  *
2650  *	Put the specified page on the active list (if appropriate).
2651  *	Ensure that act_count is at least ACT_INIT but do not otherwise
2652  *	mess with it.
2653  *
2654  *	The page must be locked.
2655  */
2656 void
2657 vm_page_activate(vm_page_t m)
2658 {
2659 	int queue;
2660 
2661 	vm_page_lock_assert(m, MA_OWNED);
2662 	if ((queue = m->queue) != PQ_ACTIVE) {
2663 		if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
2664 			if (m->act_count < ACT_INIT)
2665 				m->act_count = ACT_INIT;
2666 			if (queue != PQ_NONE)
2667 				vm_page_dequeue(m);
2668 			vm_page_enqueue(PQ_ACTIVE, m);
2669 		} else
2670 			KASSERT(queue == PQ_NONE,
2671 			    ("vm_page_activate: wired page %p is queued", m));
2672 	} else {
2673 		if (m->act_count < ACT_INIT)
2674 			m->act_count = ACT_INIT;
2675 	}
2676 }
2677 
2678 /*
2679  *	vm_page_free_wakeup:
2680  *
2681  *	Helper routine for vm_page_free_toq().  This routine is called
2682  *	when a page is added to the free queues.
2683  *
2684  *	The page queues must be locked.
2685  */
2686 static inline void
2687 vm_page_free_wakeup(void)
2688 {
2689 
2690 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
2691 	/*
2692 	 * if pageout daemon needs pages, then tell it that there are
2693 	 * some free.
2694 	 */
2695 	if (vm_pageout_pages_needed &&
2696 	    vm_cnt.v_free_count >= vm_cnt.v_pageout_free_min) {
2697 		wakeup(&vm_pageout_pages_needed);
2698 		vm_pageout_pages_needed = 0;
2699 	}
2700 	/*
2701 	 * wakeup processes that are waiting on memory if we hit a
2702 	 * high water mark. And wakeup scheduler process if we have
2703 	 * lots of memory. this process will swapin processes.
2704 	 */
2705 	if (vm_pages_needed && !vm_page_count_min()) {
2706 		vm_pages_needed = false;
2707 		wakeup(&vm_cnt.v_free_count);
2708 	}
2709 }
2710 
2711 /*
2712  *	vm_page_free_toq:
2713  *
2714  *	Returns the given page to the free list,
2715  *	disassociating it with any VM object.
2716  *
2717  *	The object must be locked.  The page must be locked if it is managed.
2718  */
2719 void
2720 vm_page_free_toq(vm_page_t m)
2721 {
2722 
2723 	if ((m->oflags & VPO_UNMANAGED) == 0) {
2724 		vm_page_lock_assert(m, MA_OWNED);
2725 		KASSERT(!pmap_page_is_mapped(m),
2726 		    ("vm_page_free_toq: freeing mapped page %p", m));
2727 	} else
2728 		KASSERT(m->queue == PQ_NONE,
2729 		    ("vm_page_free_toq: unmanaged page %p is queued", m));
2730 	PCPU_INC(cnt.v_tfree);
2731 
2732 	if (vm_page_sbusied(m))
2733 		panic("vm_page_free: freeing busy page %p", m);
2734 
2735 	/*
2736 	 * Unqueue, then remove page.  Note that we cannot destroy
2737 	 * the page here because we do not want to call the pager's
2738 	 * callback routine until after we've put the page on the
2739 	 * appropriate free queue.
2740 	 */
2741 	vm_page_remque(m);
2742 	vm_page_remove(m);
2743 
2744 	/*
2745 	 * If fictitious remove object association and
2746 	 * return, otherwise delay object association removal.
2747 	 */
2748 	if ((m->flags & PG_FICTITIOUS) != 0) {
2749 		return;
2750 	}
2751 
2752 	m->valid = 0;
2753 	vm_page_undirty(m);
2754 
2755 	if (m->wire_count != 0)
2756 		panic("vm_page_free: freeing wired page %p", m);
2757 	if (m->hold_count != 0) {
2758 		m->flags &= ~PG_ZERO;
2759 		KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2760 		    ("vm_page_free: freeing PG_UNHOLDFREE page %p", m));
2761 		m->flags |= PG_UNHOLDFREE;
2762 	} else {
2763 		/*
2764 		 * Restore the default memory attribute to the page.
2765 		 */
2766 		if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
2767 			pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
2768 
2769 		/*
2770 		 * Insert the page into the physical memory allocator's free
2771 		 * page queues.
2772 		 */
2773 		mtx_lock(&vm_page_queue_free_mtx);
2774 		vm_phys_freecnt_adj(m, 1);
2775 #if VM_NRESERVLEVEL > 0
2776 		if (!vm_reserv_free_page(m))
2777 #else
2778 		if (TRUE)
2779 #endif
2780 			vm_phys_free_pages(m, 0);
2781 		vm_page_free_wakeup();
2782 		mtx_unlock(&vm_page_queue_free_mtx);
2783 	}
2784 }
2785 
2786 /*
2787  *	vm_page_wire:
2788  *
2789  *	Mark this page as wired down by yet
2790  *	another map, removing it from paging queues
2791  *	as necessary.
2792  *
2793  *	If the page is fictitious, then its wire count must remain one.
2794  *
2795  *	The page must be locked.
2796  */
2797 void
2798 vm_page_wire(vm_page_t m)
2799 {
2800 
2801 	/*
2802 	 * Only bump the wire statistics if the page is not already wired,
2803 	 * and only unqueue the page if it is on some queue (if it is unmanaged
2804 	 * it is already off the queues).
2805 	 */
2806 	vm_page_lock_assert(m, MA_OWNED);
2807 	if ((m->flags & PG_FICTITIOUS) != 0) {
2808 		KASSERT(m->wire_count == 1,
2809 		    ("vm_page_wire: fictitious page %p's wire count isn't one",
2810 		    m));
2811 		return;
2812 	}
2813 	if (m->wire_count == 0) {
2814 		KASSERT((m->oflags & VPO_UNMANAGED) == 0 ||
2815 		    m->queue == PQ_NONE,
2816 		    ("vm_page_wire: unmanaged page %p is queued", m));
2817 		vm_page_remque(m);
2818 		atomic_add_int(&vm_cnt.v_wire_count, 1);
2819 	}
2820 	m->wire_count++;
2821 	KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m));
2822 }
2823 
2824 /*
2825  * vm_page_unwire:
2826  *
2827  * Release one wiring of the specified page, potentially allowing it to be
2828  * paged out.  Returns TRUE if the number of wirings transitions to zero and
2829  * FALSE otherwise.
2830  *
2831  * Only managed pages belonging to an object can be paged out.  If the number
2832  * of wirings transitions to zero and the page is eligible for page out, then
2833  * the page is added to the specified paging queue (unless PQ_NONE is
2834  * specified).
2835  *
2836  * If a page is fictitious, then its wire count must always be one.
2837  *
2838  * A managed page must be locked.
2839  */
2840 boolean_t
2841 vm_page_unwire(vm_page_t m, uint8_t queue)
2842 {
2843 
2844 	KASSERT(queue < PQ_COUNT || queue == PQ_NONE,
2845 	    ("vm_page_unwire: invalid queue %u request for page %p",
2846 	    queue, m));
2847 	if ((m->oflags & VPO_UNMANAGED) == 0)
2848 		vm_page_assert_locked(m);
2849 	if ((m->flags & PG_FICTITIOUS) != 0) {
2850 		KASSERT(m->wire_count == 1,
2851 	    ("vm_page_unwire: fictitious page %p's wire count isn't one", m));
2852 		return (FALSE);
2853 	}
2854 	if (m->wire_count > 0) {
2855 		m->wire_count--;
2856 		if (m->wire_count == 0) {
2857 			atomic_subtract_int(&vm_cnt.v_wire_count, 1);
2858 			if ((m->oflags & VPO_UNMANAGED) == 0 &&
2859 			    m->object != NULL && queue != PQ_NONE)
2860 				vm_page_enqueue(queue, m);
2861 			return (TRUE);
2862 		} else
2863 			return (FALSE);
2864 	} else
2865 		panic("vm_page_unwire: page %p's wire count is zero", m);
2866 }
2867 
2868 /*
2869  * Move the specified page to the inactive queue.
2870  *
2871  * Normally, "noreuse" is FALSE, resulting in LRU ordering of the inactive
2872  * queue.  However, setting "noreuse" to TRUE will accelerate the specified
2873  * page's reclamation, but it will not unmap the page from any address space.
2874  * This is implemented by inserting the page near the head of the inactive
2875  * queue, using a marker page to guide FIFO insertion ordering.
2876  *
2877  * The page must be locked.
2878  */
2879 static inline void
2880 _vm_page_deactivate(vm_page_t m, boolean_t noreuse)
2881 {
2882 	struct vm_pagequeue *pq;
2883 	int queue;
2884 
2885 	vm_page_assert_locked(m);
2886 
2887 	/*
2888 	 * Ignore if the page is already inactive, unless it is unlikely to be
2889 	 * reactivated.
2890 	 */
2891 	if ((queue = m->queue) == PQ_INACTIVE && !noreuse)
2892 		return;
2893 	if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
2894 		pq = &vm_phys_domain(m)->vmd_pagequeues[PQ_INACTIVE];
2895 		/* Avoid multiple acquisitions of the inactive queue lock. */
2896 		if (queue == PQ_INACTIVE) {
2897 			vm_pagequeue_lock(pq);
2898 			vm_page_dequeue_locked(m);
2899 		} else {
2900 			if (queue != PQ_NONE)
2901 				vm_page_dequeue(m);
2902 			vm_pagequeue_lock(pq);
2903 		}
2904 		m->queue = PQ_INACTIVE;
2905 		if (noreuse)
2906 			TAILQ_INSERT_BEFORE(&vm_phys_domain(m)->vmd_inacthead,
2907 			    m, plinks.q);
2908 		else
2909 			TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2910 		vm_pagequeue_cnt_inc(pq);
2911 		vm_pagequeue_unlock(pq);
2912 	}
2913 }
2914 
2915 /*
2916  * Move the specified page to the inactive queue.
2917  *
2918  * The page must be locked.
2919  */
2920 void
2921 vm_page_deactivate(vm_page_t m)
2922 {
2923 
2924 	_vm_page_deactivate(m, FALSE);
2925 }
2926 
2927 /*
2928  * Move the specified page to the inactive queue with the expectation
2929  * that it is unlikely to be reused.
2930  *
2931  * The page must be locked.
2932  */
2933 void
2934 vm_page_deactivate_noreuse(vm_page_t m)
2935 {
2936 
2937 	_vm_page_deactivate(m, TRUE);
2938 }
2939 
2940 /*
2941  * vm_page_launder
2942  *
2943  * 	Put a page in the laundry.
2944  */
2945 void
2946 vm_page_launder(vm_page_t m)
2947 {
2948 	int queue;
2949 
2950 	vm_page_assert_locked(m);
2951 	if ((queue = m->queue) != PQ_LAUNDRY) {
2952 		if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
2953 			if (queue != PQ_NONE)
2954 				vm_page_dequeue(m);
2955 			vm_page_enqueue(PQ_LAUNDRY, m);
2956 		} else
2957 			KASSERT(queue == PQ_NONE,
2958 			    ("wired page %p is queued", m));
2959 	}
2960 }
2961 
2962 /*
2963  * vm_page_unswappable
2964  *
2965  *	Put a page in the PQ_UNSWAPPABLE holding queue.
2966  */
2967 void
2968 vm_page_unswappable(vm_page_t m)
2969 {
2970 
2971 	vm_page_assert_locked(m);
2972 	KASSERT(m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0,
2973 	    ("page %p already unswappable", m));
2974 	if (m->queue != PQ_NONE)
2975 		vm_page_dequeue(m);
2976 	vm_page_enqueue(PQ_UNSWAPPABLE, m);
2977 }
2978 
2979 /*
2980  * vm_page_try_to_free()
2981  *
2982  *	Attempt to free the page.  If we cannot free it, we do nothing.
2983  *	1 is returned on success, 0 on failure.
2984  */
2985 int
2986 vm_page_try_to_free(vm_page_t m)
2987 {
2988 
2989 	vm_page_lock_assert(m, MA_OWNED);
2990 	if (m->object != NULL)
2991 		VM_OBJECT_ASSERT_WLOCKED(m->object);
2992 	if (m->dirty || m->hold_count || m->wire_count ||
2993 	    (m->oflags & VPO_UNMANAGED) != 0 || vm_page_busied(m))
2994 		return (0);
2995 	pmap_remove_all(m);
2996 	if (m->dirty)
2997 		return (0);
2998 	vm_page_free(m);
2999 	return (1);
3000 }
3001 
3002 /*
3003  * vm_page_advise
3004  *
3005  * 	Apply the specified advice to the given page.
3006  *
3007  *	The object and page must be locked.
3008  */
3009 void
3010 vm_page_advise(vm_page_t m, int advice)
3011 {
3012 
3013 	vm_page_assert_locked(m);
3014 	VM_OBJECT_ASSERT_WLOCKED(m->object);
3015 	if (advice == MADV_FREE)
3016 		/*
3017 		 * Mark the page clean.  This will allow the page to be freed
3018 		 * without first paging it out.  MADV_FREE pages are often
3019 		 * quickly reused by malloc(3), so we do not do anything that
3020 		 * would result in a page fault on a later access.
3021 		 */
3022 		vm_page_undirty(m);
3023 	else if (advice != MADV_DONTNEED) {
3024 		if (advice == MADV_WILLNEED)
3025 			vm_page_activate(m);
3026 		return;
3027 	}
3028 
3029 	/*
3030 	 * Clear any references to the page.  Otherwise, the page daemon will
3031 	 * immediately reactivate the page.
3032 	 */
3033 	vm_page_aflag_clear(m, PGA_REFERENCED);
3034 
3035 	if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
3036 		vm_page_dirty(m);
3037 
3038 	/*
3039 	 * Place clean pages near the head of the inactive queue rather than
3040 	 * the tail, thus defeating the queue's LRU operation and ensuring that
3041 	 * the page will be reused quickly.  Dirty pages not already in the
3042 	 * laundry are moved there.
3043 	 */
3044 	if (m->dirty == 0)
3045 		vm_page_deactivate_noreuse(m);
3046 	else
3047 		vm_page_launder(m);
3048 }
3049 
3050 /*
3051  * Grab a page, waiting until we are waken up due to the page
3052  * changing state.  We keep on waiting, if the page continues
3053  * to be in the object.  If the page doesn't exist, first allocate it
3054  * and then conditionally zero it.
3055  *
3056  * This routine may sleep.
3057  *
3058  * The object must be locked on entry.  The lock will, however, be released
3059  * and reacquired if the routine sleeps.
3060  */
3061 vm_page_t
3062 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
3063 {
3064 	vm_page_t m;
3065 	int sleep;
3066 
3067 	VM_OBJECT_ASSERT_WLOCKED(object);
3068 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
3069 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
3070 	    ("vm_page_grab: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
3071 retrylookup:
3072 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
3073 		sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ?
3074 		    vm_page_xbusied(m) : vm_page_busied(m);
3075 		if (sleep) {
3076 			if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3077 				return (NULL);
3078 			/*
3079 			 * Reference the page before unlocking and
3080 			 * sleeping so that the page daemon is less
3081 			 * likely to reclaim it.
3082 			 */
3083 			vm_page_aflag_set(m, PGA_REFERENCED);
3084 			vm_page_lock(m);
3085 			VM_OBJECT_WUNLOCK(object);
3086 			vm_page_busy_sleep(m, "pgrbwt", (allocflags &
3087 			    VM_ALLOC_IGN_SBUSY) != 0);
3088 			VM_OBJECT_WLOCK(object);
3089 			goto retrylookup;
3090 		} else {
3091 			if ((allocflags & VM_ALLOC_WIRED) != 0) {
3092 				vm_page_lock(m);
3093 				vm_page_wire(m);
3094 				vm_page_unlock(m);
3095 			}
3096 			if ((allocflags &
3097 			    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
3098 				vm_page_xbusy(m);
3099 			if ((allocflags & VM_ALLOC_SBUSY) != 0)
3100 				vm_page_sbusy(m);
3101 			return (m);
3102 		}
3103 	}
3104 	m = vm_page_alloc(object, pindex, allocflags);
3105 	if (m == NULL) {
3106 		if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3107 			return (NULL);
3108 		VM_OBJECT_WUNLOCK(object);
3109 		VM_WAIT;
3110 		VM_OBJECT_WLOCK(object);
3111 		goto retrylookup;
3112 	}
3113 	if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
3114 		pmap_zero_page(m);
3115 	return (m);
3116 }
3117 
3118 /*
3119  * Mapping function for valid or dirty bits in a page.
3120  *
3121  * Inputs are required to range within a page.
3122  */
3123 vm_page_bits_t
3124 vm_page_bits(int base, int size)
3125 {
3126 	int first_bit;
3127 	int last_bit;
3128 
3129 	KASSERT(
3130 	    base + size <= PAGE_SIZE,
3131 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
3132 	);
3133 
3134 	if (size == 0)		/* handle degenerate case */
3135 		return (0);
3136 
3137 	first_bit = base >> DEV_BSHIFT;
3138 	last_bit = (base + size - 1) >> DEV_BSHIFT;
3139 
3140 	return (((vm_page_bits_t)2 << last_bit) -
3141 	    ((vm_page_bits_t)1 << first_bit));
3142 }
3143 
3144 /*
3145  *	vm_page_set_valid_range:
3146  *
3147  *	Sets portions of a page valid.  The arguments are expected
3148  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
3149  *	of any partial chunks touched by the range.  The invalid portion of
3150  *	such chunks will be zeroed.
3151  *
3152  *	(base + size) must be less then or equal to PAGE_SIZE.
3153  */
3154 void
3155 vm_page_set_valid_range(vm_page_t m, int base, int size)
3156 {
3157 	int endoff, frag;
3158 
3159 	VM_OBJECT_ASSERT_WLOCKED(m->object);
3160 	if (size == 0)	/* handle degenerate case */
3161 		return;
3162 
3163 	/*
3164 	 * If the base is not DEV_BSIZE aligned and the valid
3165 	 * bit is clear, we have to zero out a portion of the
3166 	 * first block.
3167 	 */
3168 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
3169 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
3170 		pmap_zero_page_area(m, frag, base - frag);
3171 
3172 	/*
3173 	 * If the ending offset is not DEV_BSIZE aligned and the
3174 	 * valid bit is clear, we have to zero out a portion of
3175 	 * the last block.
3176 	 */
3177 	endoff = base + size;
3178 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
3179 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
3180 		pmap_zero_page_area(m, endoff,
3181 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
3182 
3183 	/*
3184 	 * Assert that no previously invalid block that is now being validated
3185 	 * is already dirty.
3186 	 */
3187 	KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
3188 	    ("vm_page_set_valid_range: page %p is dirty", m));
3189 
3190 	/*
3191 	 * Set valid bits inclusive of any overlap.
3192 	 */
3193 	m->valid |= vm_page_bits(base, size);
3194 }
3195 
3196 /*
3197  * Clear the given bits from the specified page's dirty field.
3198  */
3199 static __inline void
3200 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
3201 {
3202 	uintptr_t addr;
3203 #if PAGE_SIZE < 16384
3204 	int shift;
3205 #endif
3206 
3207 	/*
3208 	 * If the object is locked and the page is neither exclusive busy nor
3209 	 * write mapped, then the page's dirty field cannot possibly be
3210 	 * set by a concurrent pmap operation.
3211 	 */
3212 	VM_OBJECT_ASSERT_WLOCKED(m->object);
3213 	if (!vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
3214 		m->dirty &= ~pagebits;
3215 	else {
3216 		/*
3217 		 * The pmap layer can call vm_page_dirty() without
3218 		 * holding a distinguished lock.  The combination of
3219 		 * the object's lock and an atomic operation suffice
3220 		 * to guarantee consistency of the page dirty field.
3221 		 *
3222 		 * For PAGE_SIZE == 32768 case, compiler already
3223 		 * properly aligns the dirty field, so no forcible
3224 		 * alignment is needed. Only require existence of
3225 		 * atomic_clear_64 when page size is 32768.
3226 		 */
3227 		addr = (uintptr_t)&m->dirty;
3228 #if PAGE_SIZE == 32768
3229 		atomic_clear_64((uint64_t *)addr, pagebits);
3230 #elif PAGE_SIZE == 16384
3231 		atomic_clear_32((uint32_t *)addr, pagebits);
3232 #else		/* PAGE_SIZE <= 8192 */
3233 		/*
3234 		 * Use a trick to perform a 32-bit atomic on the
3235 		 * containing aligned word, to not depend on the existence
3236 		 * of atomic_clear_{8, 16}.
3237 		 */
3238 		shift = addr & (sizeof(uint32_t) - 1);
3239 #if BYTE_ORDER == BIG_ENDIAN
3240 		shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY;
3241 #else
3242 		shift *= NBBY;
3243 #endif
3244 		addr &= ~(sizeof(uint32_t) - 1);
3245 		atomic_clear_32((uint32_t *)addr, pagebits << shift);
3246 #endif		/* PAGE_SIZE */
3247 	}
3248 }
3249 
3250 /*
3251  *	vm_page_set_validclean:
3252  *
3253  *	Sets portions of a page valid and clean.  The arguments are expected
3254  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
3255  *	of any partial chunks touched by the range.  The invalid portion of
3256  *	such chunks will be zero'd.
3257  *
3258  *	(base + size) must be less then or equal to PAGE_SIZE.
3259  */
3260 void
3261 vm_page_set_validclean(vm_page_t m, int base, int size)
3262 {
3263 	vm_page_bits_t oldvalid, pagebits;
3264 	int endoff, frag;
3265 
3266 	VM_OBJECT_ASSERT_WLOCKED(m->object);
3267 	if (size == 0)	/* handle degenerate case */
3268 		return;
3269 
3270 	/*
3271 	 * If the base is not DEV_BSIZE aligned and the valid
3272 	 * bit is clear, we have to zero out a portion of the
3273 	 * first block.
3274 	 */
3275 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
3276 	    (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
3277 		pmap_zero_page_area(m, frag, base - frag);
3278 
3279 	/*
3280 	 * If the ending offset is not DEV_BSIZE aligned and the
3281 	 * valid bit is clear, we have to zero out a portion of
3282 	 * the last block.
3283 	 */
3284 	endoff = base + size;
3285 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
3286 	    (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
3287 		pmap_zero_page_area(m, endoff,
3288 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
3289 
3290 	/*
3291 	 * Set valid, clear dirty bits.  If validating the entire
3292 	 * page we can safely clear the pmap modify bit.  We also
3293 	 * use this opportunity to clear the VPO_NOSYNC flag.  If a process
3294 	 * takes a write fault on a MAP_NOSYNC memory area the flag will
3295 	 * be set again.
3296 	 *
3297 	 * We set valid bits inclusive of any overlap, but we can only
3298 	 * clear dirty bits for DEV_BSIZE chunks that are fully within
3299 	 * the range.
3300 	 */
3301 	oldvalid = m->valid;
3302 	pagebits = vm_page_bits(base, size);
3303 	m->valid |= pagebits;
3304 #if 0	/* NOT YET */
3305 	if ((frag = base & (DEV_BSIZE - 1)) != 0) {
3306 		frag = DEV_BSIZE - frag;
3307 		base += frag;
3308 		size -= frag;
3309 		if (size < 0)
3310 			size = 0;
3311 	}
3312 	pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
3313 #endif
3314 	if (base == 0 && size == PAGE_SIZE) {
3315 		/*
3316 		 * The page can only be modified within the pmap if it is
3317 		 * mapped, and it can only be mapped if it was previously
3318 		 * fully valid.
3319 		 */
3320 		if (oldvalid == VM_PAGE_BITS_ALL)
3321 			/*
3322 			 * Perform the pmap_clear_modify() first.  Otherwise,
3323 			 * a concurrent pmap operation, such as
3324 			 * pmap_protect(), could clear a modification in the
3325 			 * pmap and set the dirty field on the page before
3326 			 * pmap_clear_modify() had begun and after the dirty
3327 			 * field was cleared here.
3328 			 */
3329 			pmap_clear_modify(m);
3330 		m->dirty = 0;
3331 		m->oflags &= ~VPO_NOSYNC;
3332 	} else if (oldvalid != VM_PAGE_BITS_ALL)
3333 		m->dirty &= ~pagebits;
3334 	else
3335 		vm_page_clear_dirty_mask(m, pagebits);
3336 }
3337 
3338 void
3339 vm_page_clear_dirty(vm_page_t m, int base, int size)
3340 {
3341 
3342 	vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
3343 }
3344 
3345 /*
3346  *	vm_page_set_invalid:
3347  *
3348  *	Invalidates DEV_BSIZE'd chunks within a page.  Both the
3349  *	valid and dirty bits for the effected areas are cleared.
3350  */
3351 void
3352 vm_page_set_invalid(vm_page_t m, int base, int size)
3353 {
3354 	vm_page_bits_t bits;
3355 	vm_object_t object;
3356 
3357 	object = m->object;
3358 	VM_OBJECT_ASSERT_WLOCKED(object);
3359 	if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
3360 	    size >= object->un_pager.vnp.vnp_size)
3361 		bits = VM_PAGE_BITS_ALL;
3362 	else
3363 		bits = vm_page_bits(base, size);
3364 	if (object->ref_count != 0 && m->valid == VM_PAGE_BITS_ALL &&
3365 	    bits != 0)
3366 		pmap_remove_all(m);
3367 	KASSERT((bits == 0 && m->valid == VM_PAGE_BITS_ALL) ||
3368 	    !pmap_page_is_mapped(m),
3369 	    ("vm_page_set_invalid: page %p is mapped", m));
3370 	m->valid &= ~bits;
3371 	m->dirty &= ~bits;
3372 }
3373 
3374 /*
3375  * vm_page_zero_invalid()
3376  *
3377  *	The kernel assumes that the invalid portions of a page contain
3378  *	garbage, but such pages can be mapped into memory by user code.
3379  *	When this occurs, we must zero out the non-valid portions of the
3380  *	page so user code sees what it expects.
3381  *
3382  *	Pages are most often semi-valid when the end of a file is mapped
3383  *	into memory and the file's size is not page aligned.
3384  */
3385 void
3386 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
3387 {
3388 	int b;
3389 	int i;
3390 
3391 	VM_OBJECT_ASSERT_WLOCKED(m->object);
3392 	/*
3393 	 * Scan the valid bits looking for invalid sections that
3394 	 * must be zeroed.  Invalid sub-DEV_BSIZE'd areas ( where the
3395 	 * valid bit may be set ) have already been zeroed by
3396 	 * vm_page_set_validclean().
3397 	 */
3398 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
3399 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
3400 		    (m->valid & ((vm_page_bits_t)1 << i))) {
3401 			if (i > b) {
3402 				pmap_zero_page_area(m,
3403 				    b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
3404 			}
3405 			b = i + 1;
3406 		}
3407 	}
3408 
3409 	/*
3410 	 * setvalid is TRUE when we can safely set the zero'd areas
3411 	 * as being valid.  We can do this if there are no cache consistancy
3412 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
3413 	 */
3414 	if (setvalid)
3415 		m->valid = VM_PAGE_BITS_ALL;
3416 }
3417 
3418 /*
3419  *	vm_page_is_valid:
3420  *
3421  *	Is (partial) page valid?  Note that the case where size == 0
3422  *	will return FALSE in the degenerate case where the page is
3423  *	entirely invalid, and TRUE otherwise.
3424  */
3425 int
3426 vm_page_is_valid(vm_page_t m, int base, int size)
3427 {
3428 	vm_page_bits_t bits;
3429 
3430 	VM_OBJECT_ASSERT_LOCKED(m->object);
3431 	bits = vm_page_bits(base, size);
3432 	return (m->valid != 0 && (m->valid & bits) == bits);
3433 }
3434 
3435 /*
3436  *	vm_page_ps_is_valid:
3437  *
3438  *	Returns TRUE if the entire (super)page is valid and FALSE otherwise.
3439  */
3440 boolean_t
3441 vm_page_ps_is_valid(vm_page_t m)
3442 {
3443 	int i, npages;
3444 
3445 	VM_OBJECT_ASSERT_LOCKED(m->object);
3446 	npages = atop(pagesizes[m->psind]);
3447 
3448 	/*
3449 	 * The physically contiguous pages that make up a superpage, i.e., a
3450 	 * page with a page size index ("psind") greater than zero, will
3451 	 * occupy adjacent entries in vm_page_array[].
3452 	 */
3453 	for (i = 0; i < npages; i++) {
3454 		if (m[i].valid != VM_PAGE_BITS_ALL)
3455 			return (FALSE);
3456 	}
3457 	return (TRUE);
3458 }
3459 
3460 /*
3461  * Set the page's dirty bits if the page is modified.
3462  */
3463 void
3464 vm_page_test_dirty(vm_page_t m)
3465 {
3466 
3467 	VM_OBJECT_ASSERT_WLOCKED(m->object);
3468 	if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
3469 		vm_page_dirty(m);
3470 }
3471 
3472 void
3473 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
3474 {
3475 
3476 	mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
3477 }
3478 
3479 void
3480 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
3481 {
3482 
3483 	mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
3484 }
3485 
3486 int
3487 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
3488 {
3489 
3490 	return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
3491 }
3492 
3493 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
3494 void
3495 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
3496 {
3497 
3498 	vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
3499 }
3500 
3501 void
3502 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
3503 {
3504 
3505 	mtx_assert_(vm_page_lockptr(m), a, file, line);
3506 }
3507 #endif
3508 
3509 #ifdef INVARIANTS
3510 void
3511 vm_page_object_lock_assert(vm_page_t m)
3512 {
3513 
3514 	/*
3515 	 * Certain of the page's fields may only be modified by the
3516 	 * holder of the containing object's lock or the exclusive busy.
3517 	 * holder.  Unfortunately, the holder of the write busy is
3518 	 * not recorded, and thus cannot be checked here.
3519 	 */
3520 	if (m->object != NULL && !vm_page_xbusied(m))
3521 		VM_OBJECT_ASSERT_WLOCKED(m->object);
3522 }
3523 
3524 void
3525 vm_page_assert_pga_writeable(vm_page_t m, uint8_t bits)
3526 {
3527 
3528 	if ((bits & PGA_WRITEABLE) == 0)
3529 		return;
3530 
3531 	/*
3532 	 * The PGA_WRITEABLE flag can only be set if the page is
3533 	 * managed, is exclusively busied or the object is locked.
3534 	 * Currently, this flag is only set by pmap_enter().
3535 	 */
3536 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3537 	    ("PGA_WRITEABLE on unmanaged page"));
3538 	if (!vm_page_xbusied(m))
3539 		VM_OBJECT_ASSERT_LOCKED(m->object);
3540 }
3541 #endif
3542 
3543 #include "opt_ddb.h"
3544 #ifdef DDB
3545 #include <sys/kernel.h>
3546 
3547 #include <ddb/ddb.h>
3548 
3549 DB_SHOW_COMMAND(page, vm_page_print_page_info)
3550 {
3551 
3552 	db_printf("vm_cnt.v_free_count: %d\n", vm_cnt.v_free_count);
3553 	db_printf("vm_cnt.v_inactive_count: %d\n", vm_cnt.v_inactive_count);
3554 	db_printf("vm_cnt.v_active_count: %d\n", vm_cnt.v_active_count);
3555 	db_printf("vm_cnt.v_laundry_count: %d\n", vm_cnt.v_laundry_count);
3556 	db_printf("vm_cnt.v_wire_count: %d\n", vm_cnt.v_wire_count);
3557 	db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
3558 	db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
3559 	db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
3560 	db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
3561 }
3562 
3563 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
3564 {
3565 	int dom;
3566 
3567 	db_printf("pq_free %d\n", vm_cnt.v_free_count);
3568 	for (dom = 0; dom < vm_ndomains; dom++) {
3569 		db_printf(
3570     "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n",
3571 		    dom,
3572 		    vm_dom[dom].vmd_page_count,
3573 		    vm_dom[dom].vmd_free_count,
3574 		    vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
3575 		    vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
3576 		    vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt,
3577 		    vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt);
3578 	}
3579 }
3580 
3581 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
3582 {
3583 	vm_page_t m;
3584 	boolean_t phys;
3585 
3586 	if (!have_addr) {
3587 		db_printf("show pginfo addr\n");
3588 		return;
3589 	}
3590 
3591 	phys = strchr(modif, 'p') != NULL;
3592 	if (phys)
3593 		m = PHYS_TO_VM_PAGE(addr);
3594 	else
3595 		m = (vm_page_t)addr;
3596 	db_printf(
3597     "page %p obj %p pidx 0x%jx phys 0x%jx q %d hold %d wire %d\n"
3598     "  af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
3599 	    m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
3600 	    m->queue, m->hold_count, m->wire_count, m->aflags, m->oflags,
3601 	    m->flags, m->act_count, m->busy_lock, m->valid, m->dirty);
3602 }
3603 #endif /* DDB */
3604