xref: /freebsd/sys/vm/vm_page.c (revision 145992504973bd16cf3518af9ba5ce185fefa82a)
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 pageq mutex is required when adding or removing a page from a
67  *	  page queue (vm_page_queue[]), regardless of other mutexes or the
68  *	  busy state of a page.
69  *
70  *	- The object mutex is held when inserting or removing
71  *	  pages from an object (vm_page_insert() or vm_page_remove()).
72  *
73  */
74 
75 /*
76  *	Resident memory management module.
77  */
78 
79 #include <sys/cdefs.h>
80 __FBSDID("$FreeBSD$");
81 
82 #include "opt_vm.h"
83 
84 #include <sys/param.h>
85 #include <sys/systm.h>
86 #include <sys/lock.h>
87 #include <sys/kernel.h>
88 #include <sys/limits.h>
89 #include <sys/malloc.h>
90 #include <sys/msgbuf.h>
91 #include <sys/mutex.h>
92 #include <sys/proc.h>
93 #include <sys/sysctl.h>
94 #include <sys/vmmeter.h>
95 #include <sys/vnode.h>
96 
97 #include <vm/vm.h>
98 #include <vm/pmap.h>
99 #include <vm/vm_param.h>
100 #include <vm/vm_kern.h>
101 #include <vm/vm_object.h>
102 #include <vm/vm_page.h>
103 #include <vm/vm_pageout.h>
104 #include <vm/vm_pager.h>
105 #include <vm/vm_phys.h>
106 #include <vm/vm_reserv.h>
107 #include <vm/vm_extern.h>
108 #include <vm/uma.h>
109 #include <vm/uma_int.h>
110 
111 #include <machine/md_var.h>
112 
113 /*
114  *	Associated with page of user-allocatable memory is a
115  *	page structure.
116  */
117 
118 struct vpgqueues vm_page_queues[PQ_COUNT];
119 struct vpglocks vm_page_queue_lock;
120 struct vpglocks vm_page_queue_free_lock;
121 
122 struct vpglocks	pa_lock[PA_LOCK_COUNT];
123 
124 vm_page_t vm_page_array;
125 long vm_page_array_size;
126 long first_page;
127 int vm_page_zero_count;
128 
129 static int boot_pages = UMA_BOOT_PAGES;
130 TUNABLE_INT("vm.boot_pages", &boot_pages);
131 SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RD, &boot_pages, 0,
132 	"number of pages allocated for bootstrapping the VM system");
133 
134 static int pa_tryrelock_restart;
135 SYSCTL_INT(_vm, OID_AUTO, tryrelock_restart, CTLFLAG_RD,
136     &pa_tryrelock_restart, 0, "Number of tryrelock restarts");
137 
138 static uma_zone_t fakepg_zone;
139 
140 static struct vnode *vm_page_alloc_init(vm_page_t m);
141 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits);
142 static void vm_page_queue_remove(int queue, vm_page_t m);
143 static void vm_page_enqueue(int queue, vm_page_t m);
144 static void vm_page_init_fakepg(void *dummy);
145 
146 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init_fakepg, NULL);
147 
148 static void
149 vm_page_init_fakepg(void *dummy)
150 {
151 
152 	fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
153 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
154 }
155 
156 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
157 #if PAGE_SIZE == 32768
158 #ifdef CTASSERT
159 CTASSERT(sizeof(u_long) >= 8);
160 #endif
161 #endif
162 
163 /*
164  * Try to acquire a physical address lock while a pmap is locked.  If we
165  * fail to trylock we unlock and lock the pmap directly and cache the
166  * locked pa in *locked.  The caller should then restart their loop in case
167  * the virtual to physical mapping has changed.
168  */
169 int
170 vm_page_pa_tryrelock(pmap_t pmap, vm_paddr_t pa, vm_paddr_t *locked)
171 {
172 	vm_paddr_t lockpa;
173 
174 	lockpa = *locked;
175 	*locked = pa;
176 	if (lockpa) {
177 		PA_LOCK_ASSERT(lockpa, MA_OWNED);
178 		if (PA_LOCKPTR(pa) == PA_LOCKPTR(lockpa))
179 			return (0);
180 		PA_UNLOCK(lockpa);
181 	}
182 	if (PA_TRYLOCK(pa))
183 		return (0);
184 	PMAP_UNLOCK(pmap);
185 	atomic_add_int(&pa_tryrelock_restart, 1);
186 	PA_LOCK(pa);
187 	PMAP_LOCK(pmap);
188 	return (EAGAIN);
189 }
190 
191 /*
192  *	vm_set_page_size:
193  *
194  *	Sets the page size, perhaps based upon the memory
195  *	size.  Must be called before any use of page-size
196  *	dependent functions.
197  */
198 void
199 vm_set_page_size(void)
200 {
201 	if (cnt.v_page_size == 0)
202 		cnt.v_page_size = PAGE_SIZE;
203 	if (((cnt.v_page_size - 1) & cnt.v_page_size) != 0)
204 		panic("vm_set_page_size: page size not a power of two");
205 }
206 
207 /*
208  *	vm_page_blacklist_lookup:
209  *
210  *	See if a physical address in this page has been listed
211  *	in the blacklist tunable.  Entries in the tunable are
212  *	separated by spaces or commas.  If an invalid integer is
213  *	encountered then the rest of the string is skipped.
214  */
215 static int
216 vm_page_blacklist_lookup(char *list, vm_paddr_t pa)
217 {
218 	vm_paddr_t bad;
219 	char *cp, *pos;
220 
221 	for (pos = list; *pos != '\0'; pos = cp) {
222 		bad = strtoq(pos, &cp, 0);
223 		if (*cp != '\0') {
224 			if (*cp == ' ' || *cp == ',') {
225 				cp++;
226 				if (cp == pos)
227 					continue;
228 			} else
229 				break;
230 		}
231 		if (pa == trunc_page(bad))
232 			return (1);
233 	}
234 	return (0);
235 }
236 
237 /*
238  *	vm_page_startup:
239  *
240  *	Initializes the resident memory module.
241  *
242  *	Allocates memory for the page cells, and
243  *	for the object/offset-to-page hash table headers.
244  *	Each page cell is initialized and placed on the free list.
245  */
246 vm_offset_t
247 vm_page_startup(vm_offset_t vaddr)
248 {
249 	vm_offset_t mapped;
250 	vm_paddr_t page_range;
251 	vm_paddr_t new_end;
252 	int i;
253 	vm_paddr_t pa;
254 	vm_paddr_t last_pa;
255 	char *list;
256 
257 	/* the biggest memory array is the second group of pages */
258 	vm_paddr_t end;
259 	vm_paddr_t biggestsize;
260 	vm_paddr_t low_water, high_water;
261 	int biggestone;
262 
263 	biggestsize = 0;
264 	biggestone = 0;
265 	vaddr = round_page(vaddr);
266 
267 	for (i = 0; phys_avail[i + 1]; i += 2) {
268 		phys_avail[i] = round_page(phys_avail[i]);
269 		phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
270 	}
271 
272 	low_water = phys_avail[0];
273 	high_water = phys_avail[1];
274 
275 	for (i = 0; phys_avail[i + 1]; i += 2) {
276 		vm_paddr_t size = phys_avail[i + 1] - phys_avail[i];
277 
278 		if (size > biggestsize) {
279 			biggestone = i;
280 			biggestsize = size;
281 		}
282 		if (phys_avail[i] < low_water)
283 			low_water = phys_avail[i];
284 		if (phys_avail[i + 1] > high_water)
285 			high_water = phys_avail[i + 1];
286 	}
287 
288 #ifdef XEN
289 	low_water = 0;
290 #endif
291 
292 	end = phys_avail[biggestone+1];
293 
294 	/*
295 	 * Initialize the page and queue locks.
296 	 */
297 	mtx_init(&vm_page_queue_mtx, "vm page queue", NULL, MTX_DEF |
298 	    MTX_RECURSE);
299 	mtx_init(&vm_page_queue_free_mtx, "vm page free queue", NULL, MTX_DEF);
300 	for (i = 0; i < PA_LOCK_COUNT; i++)
301 		mtx_init(&pa_lock[i].data, "vm page", NULL, MTX_DEF);
302 
303 	/*
304 	 * Initialize the queue headers for the hold queue, the active queue,
305 	 * and the inactive queue.
306 	 */
307 	for (i = 0; i < PQ_COUNT; i++)
308 		TAILQ_INIT(&vm_page_queues[i].pl);
309 	vm_page_queues[PQ_INACTIVE].cnt = &cnt.v_inactive_count;
310 	vm_page_queues[PQ_ACTIVE].cnt = &cnt.v_active_count;
311 	vm_page_queues[PQ_HOLD].cnt = &cnt.v_active_count;
312 
313 	/*
314 	 * Allocate memory for use when boot strapping the kernel memory
315 	 * allocator.
316 	 */
317 	new_end = end - (boot_pages * UMA_SLAB_SIZE);
318 	new_end = trunc_page(new_end);
319 	mapped = pmap_map(&vaddr, new_end, end,
320 	    VM_PROT_READ | VM_PROT_WRITE);
321 	bzero((void *)mapped, end - new_end);
322 	uma_startup((void *)mapped, boot_pages);
323 
324 #if defined(__amd64__) || defined(__i386__) || defined(__arm__) || \
325     defined(__mips__)
326 	/*
327 	 * Allocate a bitmap to indicate that a random physical page
328 	 * needs to be included in a minidump.
329 	 *
330 	 * The amd64 port needs this to indicate which direct map pages
331 	 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
332 	 *
333 	 * However, i386 still needs this workspace internally within the
334 	 * minidump code.  In theory, they are not needed on i386, but are
335 	 * included should the sf_buf code decide to use them.
336 	 */
337 	last_pa = 0;
338 	for (i = 0; dump_avail[i + 1] != 0; i += 2)
339 		if (dump_avail[i + 1] > last_pa)
340 			last_pa = dump_avail[i + 1];
341 	page_range = last_pa / PAGE_SIZE;
342 	vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
343 	new_end -= vm_page_dump_size;
344 	vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
345 	    new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
346 	bzero((void *)vm_page_dump, vm_page_dump_size);
347 #endif
348 #ifdef __amd64__
349 	/*
350 	 * Request that the physical pages underlying the message buffer be
351 	 * included in a crash dump.  Since the message buffer is accessed
352 	 * through the direct map, they are not automatically included.
353 	 */
354 	pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
355 	last_pa = pa + round_page(msgbufsize);
356 	while (pa < last_pa) {
357 		dump_add_page(pa);
358 		pa += PAGE_SIZE;
359 	}
360 #endif
361 	/*
362 	 * Compute the number of pages of memory that will be available for
363 	 * use (taking into account the overhead of a page structure per
364 	 * page).
365 	 */
366 	first_page = low_water / PAGE_SIZE;
367 #ifdef VM_PHYSSEG_SPARSE
368 	page_range = 0;
369 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
370 		page_range += atop(phys_avail[i + 1] - phys_avail[i]);
371 #elif defined(VM_PHYSSEG_DENSE)
372 	page_range = high_water / PAGE_SIZE - first_page;
373 #else
374 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
375 #endif
376 	end = new_end;
377 
378 	/*
379 	 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
380 	 */
381 	vaddr += PAGE_SIZE;
382 
383 	/*
384 	 * Initialize the mem entry structures now, and put them in the free
385 	 * queue.
386 	 */
387 	new_end = trunc_page(end - page_range * sizeof(struct vm_page));
388 	mapped = pmap_map(&vaddr, new_end, end,
389 	    VM_PROT_READ | VM_PROT_WRITE);
390 	vm_page_array = (vm_page_t) mapped;
391 #if VM_NRESERVLEVEL > 0
392 	/*
393 	 * Allocate memory for the reservation management system's data
394 	 * structures.
395 	 */
396 	new_end = vm_reserv_startup(&vaddr, new_end, high_water);
397 #endif
398 #if defined(__amd64__) || defined(__mips__)
399 	/*
400 	 * pmap_map on amd64 and mips can come out of the direct-map, not kvm
401 	 * like i386, so the pages must be tracked for a crashdump to include
402 	 * this data.  This includes the vm_page_array and the early UMA
403 	 * bootstrap pages.
404 	 */
405 	for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE)
406 		dump_add_page(pa);
407 #endif
408 	phys_avail[biggestone + 1] = new_end;
409 
410 	/*
411 	 * Clear all of the page structures
412 	 */
413 	bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
414 	for (i = 0; i < page_range; i++)
415 		vm_page_array[i].order = VM_NFREEORDER;
416 	vm_page_array_size = page_range;
417 
418 	/*
419 	 * Initialize the physical memory allocator.
420 	 */
421 	vm_phys_init();
422 
423 	/*
424 	 * Add every available physical page that is not blacklisted to
425 	 * the free lists.
426 	 */
427 	cnt.v_page_count = 0;
428 	cnt.v_free_count = 0;
429 	list = getenv("vm.blacklist");
430 	for (i = 0; phys_avail[i + 1] != 0; i += 2) {
431 		pa = phys_avail[i];
432 		last_pa = phys_avail[i + 1];
433 		while (pa < last_pa) {
434 			if (list != NULL &&
435 			    vm_page_blacklist_lookup(list, pa))
436 				printf("Skipping page with pa 0x%jx\n",
437 				    (uintmax_t)pa);
438 			else
439 				vm_phys_add_page(pa);
440 			pa += PAGE_SIZE;
441 		}
442 	}
443 	freeenv(list);
444 #if VM_NRESERVLEVEL > 0
445 	/*
446 	 * Initialize the reservation management system.
447 	 */
448 	vm_reserv_init();
449 #endif
450 	return (vaddr);
451 }
452 
453 void
454 vm_page_reference(vm_page_t m)
455 {
456 
457 	vm_page_aflag_set(m, PGA_REFERENCED);
458 }
459 
460 void
461 vm_page_busy(vm_page_t m)
462 {
463 
464 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
465 	KASSERT((m->oflags & VPO_BUSY) == 0,
466 	    ("vm_page_busy: page already busy!!!"));
467 	m->oflags |= VPO_BUSY;
468 }
469 
470 /*
471  *      vm_page_flash:
472  *
473  *      wakeup anyone waiting for the page.
474  */
475 void
476 vm_page_flash(vm_page_t m)
477 {
478 
479 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
480 	if (m->oflags & VPO_WANTED) {
481 		m->oflags &= ~VPO_WANTED;
482 		wakeup(m);
483 	}
484 }
485 
486 /*
487  *      vm_page_wakeup:
488  *
489  *      clear the VPO_BUSY flag and wakeup anyone waiting for the
490  *      page.
491  *
492  */
493 void
494 vm_page_wakeup(vm_page_t m)
495 {
496 
497 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
498 	KASSERT(m->oflags & VPO_BUSY, ("vm_page_wakeup: page not busy!!!"));
499 	m->oflags &= ~VPO_BUSY;
500 	vm_page_flash(m);
501 }
502 
503 void
504 vm_page_io_start(vm_page_t m)
505 {
506 
507 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
508 	m->busy++;
509 }
510 
511 void
512 vm_page_io_finish(vm_page_t m)
513 {
514 
515 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
516 	KASSERT(m->busy > 0, ("vm_page_io_finish: page %p is not busy", m));
517 	m->busy--;
518 	if (m->busy == 0)
519 		vm_page_flash(m);
520 }
521 
522 /*
523  * Keep page from being freed by the page daemon
524  * much of the same effect as wiring, except much lower
525  * overhead and should be used only for *very* temporary
526  * holding ("wiring").
527  */
528 void
529 vm_page_hold(vm_page_t mem)
530 {
531 
532 	vm_page_lock_assert(mem, MA_OWNED);
533         mem->hold_count++;
534 }
535 
536 void
537 vm_page_unhold(vm_page_t mem)
538 {
539 
540 	vm_page_lock_assert(mem, MA_OWNED);
541 	--mem->hold_count;
542 	KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!"));
543 	if (mem->hold_count == 0 && mem->queue == PQ_HOLD)
544 		vm_page_free_toq(mem);
545 }
546 
547 /*
548  *	vm_page_unhold_pages:
549  *
550  *	Unhold each of the pages that is referenced by the given array.
551  */
552 void
553 vm_page_unhold_pages(vm_page_t *ma, int count)
554 {
555 	struct mtx *mtx, *new_mtx;
556 
557 	mtx = NULL;
558 	for (; count != 0; count--) {
559 		/*
560 		 * Avoid releasing and reacquiring the same page lock.
561 		 */
562 		new_mtx = vm_page_lockptr(*ma);
563 		if (mtx != new_mtx) {
564 			if (mtx != NULL)
565 				mtx_unlock(mtx);
566 			mtx = new_mtx;
567 			mtx_lock(mtx);
568 		}
569 		vm_page_unhold(*ma);
570 		ma++;
571 	}
572 	if (mtx != NULL)
573 		mtx_unlock(mtx);
574 }
575 
576 vm_page_t
577 PHYS_TO_VM_PAGE(vm_paddr_t pa)
578 {
579 	vm_page_t m;
580 
581 #ifdef VM_PHYSSEG_SPARSE
582 	m = vm_phys_paddr_to_vm_page(pa);
583 	if (m == NULL)
584 		m = vm_phys_fictitious_to_vm_page(pa);
585 	return (m);
586 #elif defined(VM_PHYSSEG_DENSE)
587 	long pi;
588 
589 	pi = atop(pa);
590 	if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
591 		m = &vm_page_array[pi - first_page];
592 		return (m);
593 	}
594 	return (vm_phys_fictitious_to_vm_page(pa));
595 #else
596 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
597 #endif
598 }
599 
600 /*
601  *	vm_page_getfake:
602  *
603  *	Create a fictitious page with the specified physical address and
604  *	memory attribute.  The memory attribute is the only the machine-
605  *	dependent aspect of a fictitious page that must be initialized.
606  */
607 vm_page_t
608 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
609 {
610 	vm_page_t m;
611 
612 	m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
613 	vm_page_initfake(m, paddr, memattr);
614 	return (m);
615 }
616 
617 void
618 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
619 {
620 
621 	if ((m->flags & PG_FICTITIOUS) != 0) {
622 		/*
623 		 * The page's memattr might have changed since the
624 		 * previous initialization.  Update the pmap to the
625 		 * new memattr.
626 		 */
627 		goto memattr;
628 	}
629 	m->phys_addr = paddr;
630 	m->queue = PQ_NONE;
631 	/* Fictitious pages don't use "segind". */
632 	m->flags = PG_FICTITIOUS;
633 	/* Fictitious pages don't use "order" or "pool". */
634 	m->oflags = VPO_BUSY | VPO_UNMANAGED;
635 	m->wire_count = 1;
636 memattr:
637 	pmap_page_set_memattr(m, memattr);
638 }
639 
640 /*
641  *	vm_page_putfake:
642  *
643  *	Release a fictitious page.
644  */
645 void
646 vm_page_putfake(vm_page_t m)
647 {
648 
649 	KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
650 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
651 	    ("vm_page_putfake: bad page %p", m));
652 	uma_zfree(fakepg_zone, m);
653 }
654 
655 /*
656  *	vm_page_updatefake:
657  *
658  *	Update the given fictitious page to the specified physical address and
659  *	memory attribute.
660  */
661 void
662 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
663 {
664 
665 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
666 	    ("vm_page_updatefake: bad page %p", m));
667 	m->phys_addr = paddr;
668 	pmap_page_set_memattr(m, memattr);
669 }
670 
671 /*
672  *	vm_page_free:
673  *
674  *	Free a page.
675  */
676 void
677 vm_page_free(vm_page_t m)
678 {
679 
680 	m->flags &= ~PG_ZERO;
681 	vm_page_free_toq(m);
682 }
683 
684 /*
685  *	vm_page_free_zero:
686  *
687  *	Free a page to the zerod-pages queue
688  */
689 void
690 vm_page_free_zero(vm_page_t m)
691 {
692 
693 	m->flags |= PG_ZERO;
694 	vm_page_free_toq(m);
695 }
696 
697 /*
698  * Unbusy and handle the page queueing for a page from the VOP_GETPAGES()
699  * array which is not the request page.
700  */
701 void
702 vm_page_readahead_finish(vm_page_t m)
703 {
704 
705 	if (m->valid != 0) {
706 		/*
707 		 * Since the page is not the requested page, whether
708 		 * it should be activated or deactivated is not
709 		 * obvious.  Empirical results have shown that
710 		 * deactivating the page is usually the best choice,
711 		 * unless the page is wanted by another thread.
712 		 */
713 		if (m->oflags & VPO_WANTED) {
714 			vm_page_lock(m);
715 			vm_page_activate(m);
716 			vm_page_unlock(m);
717 		} else {
718 			vm_page_lock(m);
719 			vm_page_deactivate(m);
720 			vm_page_unlock(m);
721 		}
722 		vm_page_wakeup(m);
723 	} else {
724 		/*
725 		 * Free the completely invalid page.  Such page state
726 		 * occurs due to the short read operation which did
727 		 * not covered our page at all, or in case when a read
728 		 * error happens.
729 		 */
730 		vm_page_lock(m);
731 		vm_page_free(m);
732 		vm_page_unlock(m);
733 	}
734 }
735 
736 /*
737  *	vm_page_sleep:
738  *
739  *	Sleep and release the page lock.
740  *
741  *	The object containing the given page must be locked.
742  */
743 void
744 vm_page_sleep(vm_page_t m, const char *msg)
745 {
746 
747 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
748 	if (mtx_owned(vm_page_lockptr(m)))
749 		vm_page_unlock(m);
750 
751 	/*
752 	 * It's possible that while we sleep, the page will get
753 	 * unbusied and freed.  If we are holding the object
754 	 * lock, we will assume we hold a reference to the object
755 	 * such that even if m->object changes, we can re-lock
756 	 * it.
757 	 */
758 	m->oflags |= VPO_WANTED;
759 	msleep(m, VM_OBJECT_MTX(m->object), PVM, msg, 0);
760 }
761 
762 /*
763  *	vm_page_dirty_KBI:		[ internal use only ]
764  *
765  *	Set all bits in the page's dirty field.
766  *
767  *	The object containing the specified page must be locked if the
768  *	call is made from the machine-independent layer.
769  *
770  *	See vm_page_clear_dirty_mask().
771  *
772  *	This function should only be called by vm_page_dirty().
773  */
774 void
775 vm_page_dirty_KBI(vm_page_t m)
776 {
777 
778 	/* These assertions refer to this operation by its public name. */
779 	KASSERT((m->flags & PG_CACHED) == 0,
780 	    ("vm_page_dirty: page in cache!"));
781 	KASSERT(!VM_PAGE_IS_FREE(m),
782 	    ("vm_page_dirty: page is free!"));
783 	KASSERT(m->valid == VM_PAGE_BITS_ALL,
784 	    ("vm_page_dirty: page is invalid!"));
785 	m->dirty = VM_PAGE_BITS_ALL;
786 }
787 
788 /*
789  *	vm_page_splay:
790  *
791  *	Implements Sleator and Tarjan's top-down splay algorithm.  Returns
792  *	the vm_page containing the given pindex.  If, however, that
793  *	pindex is not found in the vm_object, returns a vm_page that is
794  *	adjacent to the pindex, coming before or after it.
795  */
796 vm_page_t
797 vm_page_splay(vm_pindex_t pindex, vm_page_t root)
798 {
799 	struct vm_page dummy;
800 	vm_page_t lefttreemax, righttreemin, y;
801 
802 	if (root == NULL)
803 		return (root);
804 	lefttreemax = righttreemin = &dummy;
805 	for (;; root = y) {
806 		if (pindex < root->pindex) {
807 			if ((y = root->left) == NULL)
808 				break;
809 			if (pindex < y->pindex) {
810 				/* Rotate right. */
811 				root->left = y->right;
812 				y->right = root;
813 				root = y;
814 				if ((y = root->left) == NULL)
815 					break;
816 			}
817 			/* Link into the new root's right tree. */
818 			righttreemin->left = root;
819 			righttreemin = root;
820 		} else if (pindex > root->pindex) {
821 			if ((y = root->right) == NULL)
822 				break;
823 			if (pindex > y->pindex) {
824 				/* Rotate left. */
825 				root->right = y->left;
826 				y->left = root;
827 				root = y;
828 				if ((y = root->right) == NULL)
829 					break;
830 			}
831 			/* Link into the new root's left tree. */
832 			lefttreemax->right = root;
833 			lefttreemax = root;
834 		} else
835 			break;
836 	}
837 	/* Assemble the new root. */
838 	lefttreemax->right = root->left;
839 	righttreemin->left = root->right;
840 	root->left = dummy.right;
841 	root->right = dummy.left;
842 	return (root);
843 }
844 
845 /*
846  *	vm_page_insert:		[ internal use only ]
847  *
848  *	Inserts the given mem entry into the object and object list.
849  *
850  *	The pagetables are not updated but will presumably fault the page
851  *	in if necessary, or if a kernel page the caller will at some point
852  *	enter the page into the kernel's pmap.  We are not allowed to sleep
853  *	here so we *can't* do this anyway.
854  *
855  *	The object must be locked.
856  */
857 void
858 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
859 {
860 	vm_page_t root;
861 
862 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
863 	if (m->object != NULL)
864 		panic("vm_page_insert: page already inserted");
865 
866 	/*
867 	 * Record the object/offset pair in this page
868 	 */
869 	m->object = object;
870 	m->pindex = pindex;
871 
872 	/*
873 	 * Now link into the object's ordered list of backed pages.
874 	 */
875 	root = object->root;
876 	if (root == NULL) {
877 		m->left = NULL;
878 		m->right = NULL;
879 		TAILQ_INSERT_TAIL(&object->memq, m, listq);
880 	} else {
881 		root = vm_page_splay(pindex, root);
882 		if (pindex < root->pindex) {
883 			m->left = root->left;
884 			m->right = root;
885 			root->left = NULL;
886 			TAILQ_INSERT_BEFORE(root, m, listq);
887 		} else if (pindex == root->pindex)
888 			panic("vm_page_insert: offset already allocated");
889 		else {
890 			m->right = root->right;
891 			m->left = root;
892 			root->right = NULL;
893 			TAILQ_INSERT_AFTER(&object->memq, root, m, listq);
894 		}
895 	}
896 	object->root = m;
897 
898 	/*
899 	 * Show that the object has one more resident page.
900 	 */
901 	object->resident_page_count++;
902 
903 	/*
904 	 * Hold the vnode until the last page is released.
905 	 */
906 	if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
907 		vhold(object->handle);
908 
909 	/*
910 	 * Since we are inserting a new and possibly dirty page,
911 	 * update the object's OBJ_MIGHTBEDIRTY flag.
912 	 */
913 	if (pmap_page_is_write_mapped(m))
914 		vm_object_set_writeable_dirty(object);
915 }
916 
917 /*
918  *	vm_page_remove:
919  *
920  *	Removes the given mem entry from the object/offset-page
921  *	table and the object page list, but do not invalidate/terminate
922  *	the backing store.
923  *
924  *	The underlying pmap entry (if any) is NOT removed here.
925  *
926  *	The object must be locked.  The page must be locked if it is managed.
927  */
928 void
929 vm_page_remove(vm_page_t m)
930 {
931 	vm_object_t object;
932 	vm_page_t next, prev, root;
933 
934 	if ((m->oflags & VPO_UNMANAGED) == 0)
935 		vm_page_lock_assert(m, MA_OWNED);
936 	if ((object = m->object) == NULL)
937 		return;
938 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
939 	if (m->oflags & VPO_BUSY) {
940 		m->oflags &= ~VPO_BUSY;
941 		vm_page_flash(m);
942 	}
943 
944 	/*
945 	 * Now remove from the object's list of backed pages.
946 	 */
947 	if ((next = TAILQ_NEXT(m, listq)) != NULL && next->left == m) {
948 		/*
949 		 * Since the page's successor in the list is also its parent
950 		 * in the tree, its right subtree must be empty.
951 		 */
952 		next->left = m->left;
953 		KASSERT(m->right == NULL,
954 		    ("vm_page_remove: page %p has right child", m));
955 	} else if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL &&
956 	    prev->right == m) {
957 		/*
958 		 * Since the page's predecessor in the list is also its parent
959 		 * in the tree, its left subtree must be empty.
960 		 */
961 		KASSERT(m->left == NULL,
962 		    ("vm_page_remove: page %p has left child", m));
963 		prev->right = m->right;
964 	} else {
965 		if (m != object->root)
966 			vm_page_splay(m->pindex, object->root);
967 		if (m->left == NULL)
968 			root = m->right;
969 		else if (m->right == NULL)
970 			root = m->left;
971 		else {
972 			/*
973 			 * Move the page's successor to the root, because
974 			 * pages are usually removed in ascending order.
975 			 */
976 			if (m->right != next)
977 				vm_page_splay(m->pindex, m->right);
978 			next->left = m->left;
979 			root = next;
980 		}
981 		object->root = root;
982 	}
983 	TAILQ_REMOVE(&object->memq, m, listq);
984 
985 	/*
986 	 * And show that the object has one fewer resident page.
987 	 */
988 	object->resident_page_count--;
989 
990 	/*
991 	 * The vnode may now be recycled.
992 	 */
993 	if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
994 		vdrop(object->handle);
995 
996 	m->object = NULL;
997 }
998 
999 /*
1000  *	vm_page_lookup:
1001  *
1002  *	Returns the page associated with the object/offset
1003  *	pair specified; if none is found, NULL is returned.
1004  *
1005  *	The object must be locked.
1006  */
1007 vm_page_t
1008 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1009 {
1010 	vm_page_t m;
1011 
1012 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1013 	if ((m = object->root) != NULL && m->pindex != pindex) {
1014 		m = vm_page_splay(pindex, m);
1015 		if ((object->root = m)->pindex != pindex)
1016 			m = NULL;
1017 	}
1018 	return (m);
1019 }
1020 
1021 /*
1022  *	vm_page_find_least:
1023  *
1024  *	Returns the page associated with the object with least pindex
1025  *	greater than or equal to the parameter pindex, or NULL.
1026  *
1027  *	The object must be locked.
1028  */
1029 vm_page_t
1030 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1031 {
1032 	vm_page_t m;
1033 
1034 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1035 	if ((m = TAILQ_FIRST(&object->memq)) != NULL) {
1036 		if (m->pindex < pindex) {
1037 			m = vm_page_splay(pindex, object->root);
1038 			if ((object->root = m)->pindex < pindex)
1039 				m = TAILQ_NEXT(m, listq);
1040 		}
1041 	}
1042 	return (m);
1043 }
1044 
1045 /*
1046  * Returns the given page's successor (by pindex) within the object if it is
1047  * resident; if none is found, NULL is returned.
1048  *
1049  * The object must be locked.
1050  */
1051 vm_page_t
1052 vm_page_next(vm_page_t m)
1053 {
1054 	vm_page_t next;
1055 
1056 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1057 	if ((next = TAILQ_NEXT(m, listq)) != NULL &&
1058 	    next->pindex != m->pindex + 1)
1059 		next = NULL;
1060 	return (next);
1061 }
1062 
1063 /*
1064  * Returns the given page's predecessor (by pindex) within the object if it is
1065  * resident; if none is found, NULL is returned.
1066  *
1067  * The object must be locked.
1068  */
1069 vm_page_t
1070 vm_page_prev(vm_page_t m)
1071 {
1072 	vm_page_t prev;
1073 
1074 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1075 	if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL &&
1076 	    prev->pindex != m->pindex - 1)
1077 		prev = NULL;
1078 	return (prev);
1079 }
1080 
1081 /*
1082  *	vm_page_rename:
1083  *
1084  *	Move the given memory entry from its
1085  *	current object to the specified target object/offset.
1086  *
1087  *	Note: swap associated with the page must be invalidated by the move.  We
1088  *	      have to do this for several reasons:  (1) we aren't freeing the
1089  *	      page, (2) we are dirtying the page, (3) the VM system is probably
1090  *	      moving the page from object A to B, and will then later move
1091  *	      the backing store from A to B and we can't have a conflict.
1092  *
1093  *	Note: we *always* dirty the page.  It is necessary both for the
1094  *	      fact that we moved it, and because we may be invalidating
1095  *	      swap.  If the page is on the cache, we have to deactivate it
1096  *	      or vm_page_dirty() will panic.  Dirty pages are not allowed
1097  *	      on the cache.
1098  *
1099  *	The objects must be locked.  The page must be locked if it is managed.
1100  */
1101 void
1102 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1103 {
1104 
1105 	vm_page_remove(m);
1106 	vm_page_insert(m, new_object, new_pindex);
1107 	vm_page_dirty(m);
1108 }
1109 
1110 /*
1111  *	Convert all of the given object's cached pages that have a
1112  *	pindex within the given range into free pages.  If the value
1113  *	zero is given for "end", then the range's upper bound is
1114  *	infinity.  If the given object is backed by a vnode and it
1115  *	transitions from having one or more cached pages to none, the
1116  *	vnode's hold count is reduced.
1117  */
1118 void
1119 vm_page_cache_free(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1120 {
1121 	vm_page_t m, m_next;
1122 	boolean_t empty;
1123 
1124 	mtx_lock(&vm_page_queue_free_mtx);
1125 	if (__predict_false(object->cache == NULL)) {
1126 		mtx_unlock(&vm_page_queue_free_mtx);
1127 		return;
1128 	}
1129 	m = object->cache = vm_page_splay(start, object->cache);
1130 	if (m->pindex < start) {
1131 		if (m->right == NULL)
1132 			m = NULL;
1133 		else {
1134 			m_next = vm_page_splay(start, m->right);
1135 			m_next->left = m;
1136 			m->right = NULL;
1137 			m = object->cache = m_next;
1138 		}
1139 	}
1140 
1141 	/*
1142 	 * At this point, "m" is either (1) a reference to the page
1143 	 * with the least pindex that is greater than or equal to
1144 	 * "start" or (2) NULL.
1145 	 */
1146 	for (; m != NULL && (m->pindex < end || end == 0); m = m_next) {
1147 		/*
1148 		 * Find "m"'s successor and remove "m" from the
1149 		 * object's cache.
1150 		 */
1151 		if (m->right == NULL) {
1152 			object->cache = m->left;
1153 			m_next = NULL;
1154 		} else {
1155 			m_next = vm_page_splay(start, m->right);
1156 			m_next->left = m->left;
1157 			object->cache = m_next;
1158 		}
1159 		/* Convert "m" to a free page. */
1160 		m->object = NULL;
1161 		m->valid = 0;
1162 		/* Clear PG_CACHED and set PG_FREE. */
1163 		m->flags ^= PG_CACHED | PG_FREE;
1164 		KASSERT((m->flags & (PG_CACHED | PG_FREE)) == PG_FREE,
1165 		    ("vm_page_cache_free: page %p has inconsistent flags", m));
1166 		cnt.v_cache_count--;
1167 		cnt.v_free_count++;
1168 	}
1169 	empty = object->cache == NULL;
1170 	mtx_unlock(&vm_page_queue_free_mtx);
1171 	if (object->type == OBJT_VNODE && empty)
1172 		vdrop(object->handle);
1173 }
1174 
1175 /*
1176  *	Returns the cached page that is associated with the given
1177  *	object and offset.  If, however, none exists, returns NULL.
1178  *
1179  *	The free page queue must be locked.
1180  */
1181 static inline vm_page_t
1182 vm_page_cache_lookup(vm_object_t object, vm_pindex_t pindex)
1183 {
1184 	vm_page_t m;
1185 
1186 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1187 	if ((m = object->cache) != NULL && m->pindex != pindex) {
1188 		m = vm_page_splay(pindex, m);
1189 		if ((object->cache = m)->pindex != pindex)
1190 			m = NULL;
1191 	}
1192 	return (m);
1193 }
1194 
1195 /*
1196  *	Remove the given cached page from its containing object's
1197  *	collection of cached pages.
1198  *
1199  *	The free page queue must be locked.
1200  */
1201 static void
1202 vm_page_cache_remove(vm_page_t m)
1203 {
1204 	vm_object_t object;
1205 	vm_page_t root;
1206 
1207 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1208 	KASSERT((m->flags & PG_CACHED) != 0,
1209 	    ("vm_page_cache_remove: page %p is not cached", m));
1210 	object = m->object;
1211 	if (m != object->cache) {
1212 		root = vm_page_splay(m->pindex, object->cache);
1213 		KASSERT(root == m,
1214 		    ("vm_page_cache_remove: page %p is not cached in object %p",
1215 		    m, object));
1216 	}
1217 	if (m->left == NULL)
1218 		root = m->right;
1219 	else if (m->right == NULL)
1220 		root = m->left;
1221 	else {
1222 		root = vm_page_splay(m->pindex, m->left);
1223 		root->right = m->right;
1224 	}
1225 	object->cache = root;
1226 	m->object = NULL;
1227 	cnt.v_cache_count--;
1228 }
1229 
1230 /*
1231  *	Transfer all of the cached pages with offset greater than or
1232  *	equal to 'offidxstart' from the original object's cache to the
1233  *	new object's cache.  However, any cached pages with offset
1234  *	greater than or equal to the new object's size are kept in the
1235  *	original object.  Initially, the new object's cache must be
1236  *	empty.  Offset 'offidxstart' in the original object must
1237  *	correspond to offset zero in the new object.
1238  *
1239  *	The new object must be locked.
1240  */
1241 void
1242 vm_page_cache_transfer(vm_object_t orig_object, vm_pindex_t offidxstart,
1243     vm_object_t new_object)
1244 {
1245 	vm_page_t m, m_next;
1246 
1247 	/*
1248 	 * Insertion into an object's collection of cached pages
1249 	 * requires the object to be locked.  In contrast, removal does
1250 	 * not.
1251 	 */
1252 	VM_OBJECT_LOCK_ASSERT(new_object, MA_OWNED);
1253 	KASSERT(new_object->cache == NULL,
1254 	    ("vm_page_cache_transfer: object %p has cached pages",
1255 	    new_object));
1256 	mtx_lock(&vm_page_queue_free_mtx);
1257 	if ((m = orig_object->cache) != NULL) {
1258 		/*
1259 		 * Transfer all of the pages with offset greater than or
1260 		 * equal to 'offidxstart' from the original object's
1261 		 * cache to the new object's cache.
1262 		 */
1263 		m = vm_page_splay(offidxstart, m);
1264 		if (m->pindex < offidxstart) {
1265 			orig_object->cache = m;
1266 			new_object->cache = m->right;
1267 			m->right = NULL;
1268 		} else {
1269 			orig_object->cache = m->left;
1270 			new_object->cache = m;
1271 			m->left = NULL;
1272 		}
1273 		while ((m = new_object->cache) != NULL) {
1274 			if ((m->pindex - offidxstart) >= new_object->size) {
1275 				/*
1276 				 * Return all of the cached pages with
1277 				 * offset greater than or equal to the
1278 				 * new object's size to the original
1279 				 * object's cache.
1280 				 */
1281 				new_object->cache = m->left;
1282 				m->left = orig_object->cache;
1283 				orig_object->cache = m;
1284 				break;
1285 			}
1286 			m_next = vm_page_splay(m->pindex, m->right);
1287 			/* Update the page's object and offset. */
1288 			m->object = new_object;
1289 			m->pindex -= offidxstart;
1290 			if (m_next == NULL)
1291 				break;
1292 			m->right = NULL;
1293 			m_next->left = m;
1294 			new_object->cache = m_next;
1295 		}
1296 		KASSERT(new_object->cache == NULL ||
1297 		    new_object->type == OBJT_SWAP,
1298 		    ("vm_page_cache_transfer: object %p's type is incompatible"
1299 		    " with cached pages", new_object));
1300 	}
1301 	mtx_unlock(&vm_page_queue_free_mtx);
1302 }
1303 
1304 /*
1305  *	Returns TRUE if a cached page is associated with the given object and
1306  *	offset, and FALSE otherwise.
1307  *
1308  *	The object must be locked.
1309  */
1310 boolean_t
1311 vm_page_is_cached(vm_object_t object, vm_pindex_t pindex)
1312 {
1313 	vm_page_t m;
1314 
1315 	/*
1316 	 * Insertion into an object's collection of cached pages requires the
1317 	 * object to be locked.  Therefore, if the object is locked and the
1318 	 * object's collection is empty, there is no need to acquire the free
1319 	 * page queues lock in order to prove that the specified page doesn't
1320 	 * exist.
1321 	 */
1322 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1323 	if (__predict_true(object->cache == NULL))
1324 		return (FALSE);
1325 	mtx_lock(&vm_page_queue_free_mtx);
1326 	m = vm_page_cache_lookup(object, pindex);
1327 	mtx_unlock(&vm_page_queue_free_mtx);
1328 	return (m != NULL);
1329 }
1330 
1331 /*
1332  *	vm_page_alloc:
1333  *
1334  *	Allocate and return a page that is associated with the specified
1335  *	object and offset pair.  By default, this page has the flag VPO_BUSY
1336  *	set.
1337  *
1338  *	The caller must always specify an allocation class.
1339  *
1340  *	allocation classes:
1341  *	VM_ALLOC_NORMAL		normal process request
1342  *	VM_ALLOC_SYSTEM		system *really* needs a page
1343  *	VM_ALLOC_INTERRUPT	interrupt time request
1344  *
1345  *	optional allocation flags:
1346  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
1347  *				intends to allocate
1348  *	VM_ALLOC_IFCACHED	return page only if it is cached
1349  *	VM_ALLOC_IFNOTCACHED	return NULL, do not reactivate if the page
1350  *				is cached
1351  *	VM_ALLOC_NOBUSY		do not set the flag VPO_BUSY on the page
1352  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
1353  *	VM_ALLOC_NOOBJ		page is not associated with an object and
1354  *				should not have the flag VPO_BUSY set
1355  *	VM_ALLOC_WIRED		wire the allocated page
1356  *	VM_ALLOC_ZERO		prefer a zeroed page
1357  *
1358  *	This routine may not sleep.
1359  */
1360 vm_page_t
1361 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1362 {
1363 	struct vnode *vp = NULL;
1364 	vm_object_t m_object;
1365 	vm_page_t m;
1366 	int flags, req_class;
1367 
1368 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0),
1369 	    ("vm_page_alloc: inconsistent object/req"));
1370 	if (object != NULL)
1371 		VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1372 
1373 	req_class = req & VM_ALLOC_CLASS_MASK;
1374 
1375 	/*
1376 	 * The page daemon is allowed to dig deeper into the free page list.
1377 	 */
1378 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1379 		req_class = VM_ALLOC_SYSTEM;
1380 
1381 	mtx_lock(&vm_page_queue_free_mtx);
1382 	if (cnt.v_free_count + cnt.v_cache_count > cnt.v_free_reserved ||
1383 	    (req_class == VM_ALLOC_SYSTEM &&
1384 	    cnt.v_free_count + cnt.v_cache_count > cnt.v_interrupt_free_min) ||
1385 	    (req_class == VM_ALLOC_INTERRUPT &&
1386 	    cnt.v_free_count + cnt.v_cache_count > 0)) {
1387 		/*
1388 		 * Allocate from the free queue if the number of free pages
1389 		 * exceeds the minimum for the request class.
1390 		 */
1391 		if (object != NULL &&
1392 		    (m = vm_page_cache_lookup(object, pindex)) != NULL) {
1393 			if ((req & VM_ALLOC_IFNOTCACHED) != 0) {
1394 				mtx_unlock(&vm_page_queue_free_mtx);
1395 				return (NULL);
1396 			}
1397 			if (vm_phys_unfree_page(m))
1398 				vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m, 0);
1399 #if VM_NRESERVLEVEL > 0
1400 			else if (!vm_reserv_reactivate_page(m))
1401 #else
1402 			else
1403 #endif
1404 				panic("vm_page_alloc: cache page %p is missing"
1405 				    " from the free queue", m);
1406 		} else if ((req & VM_ALLOC_IFCACHED) != 0) {
1407 			mtx_unlock(&vm_page_queue_free_mtx);
1408 			return (NULL);
1409 #if VM_NRESERVLEVEL > 0
1410 		} else if (object == NULL || object->type == OBJT_DEVICE ||
1411 		    object->type == OBJT_SG ||
1412 		    (object->flags & OBJ_COLORED) == 0 ||
1413 		    (m = vm_reserv_alloc_page(object, pindex)) == NULL) {
1414 #else
1415 		} else {
1416 #endif
1417 			m = vm_phys_alloc_pages(object != NULL ?
1418 			    VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0);
1419 #if VM_NRESERVLEVEL > 0
1420 			if (m == NULL && vm_reserv_reclaim_inactive()) {
1421 				m = vm_phys_alloc_pages(object != NULL ?
1422 				    VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT,
1423 				    0);
1424 			}
1425 #endif
1426 		}
1427 	} else {
1428 		/*
1429 		 * Not allocatable, give up.
1430 		 */
1431 		mtx_unlock(&vm_page_queue_free_mtx);
1432 		atomic_add_int(&vm_pageout_deficit,
1433 		    max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
1434 		pagedaemon_wakeup();
1435 		return (NULL);
1436 	}
1437 
1438 	/*
1439 	 *  At this point we had better have found a good page.
1440 	 */
1441 	KASSERT(m != NULL, ("vm_page_alloc: missing page"));
1442 	KASSERT(m->queue == PQ_NONE,
1443 	    ("vm_page_alloc: page %p has unexpected queue %d", m, m->queue));
1444 	KASSERT(m->wire_count == 0, ("vm_page_alloc: page %p is wired", m));
1445 	KASSERT(m->hold_count == 0, ("vm_page_alloc: page %p is held", m));
1446 	KASSERT(m->busy == 0, ("vm_page_alloc: page %p is busy", m));
1447 	KASSERT(m->dirty == 0, ("vm_page_alloc: page %p is dirty", m));
1448 	KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
1449 	    ("vm_page_alloc: page %p has unexpected memattr %d", m,
1450 	    pmap_page_get_memattr(m)));
1451 	if ((m->flags & PG_CACHED) != 0) {
1452 		KASSERT((m->flags & PG_ZERO) == 0,
1453 		    ("vm_page_alloc: cached page %p is PG_ZERO", m));
1454 		KASSERT(m->valid != 0,
1455 		    ("vm_page_alloc: cached page %p is invalid", m));
1456 		if (m->object == object && m->pindex == pindex)
1457 	  		cnt.v_reactivated++;
1458 		else
1459 			m->valid = 0;
1460 		m_object = m->object;
1461 		vm_page_cache_remove(m);
1462 		if (m_object->type == OBJT_VNODE && m_object->cache == NULL)
1463 			vp = m_object->handle;
1464 	} else {
1465 		KASSERT(VM_PAGE_IS_FREE(m),
1466 		    ("vm_page_alloc: page %p is not free", m));
1467 		KASSERT(m->valid == 0,
1468 		    ("vm_page_alloc: free page %p is valid", m));
1469 		cnt.v_free_count--;
1470 	}
1471 
1472 	/*
1473 	 * Only the PG_ZERO flag is inherited.  The PG_CACHED or PG_FREE flag
1474 	 * must be cleared before the free page queues lock is released.
1475 	 */
1476 	flags = 0;
1477 	if (req & VM_ALLOC_NODUMP)
1478 		flags |= PG_NODUMP;
1479 	if (m->flags & PG_ZERO) {
1480 		vm_page_zero_count--;
1481 		if (req & VM_ALLOC_ZERO)
1482 			flags = PG_ZERO;
1483 	}
1484 	m->flags = flags;
1485 	mtx_unlock(&vm_page_queue_free_mtx);
1486 	m->aflags = 0;
1487 	if (object == NULL || object->type == OBJT_PHYS)
1488 		m->oflags = VPO_UNMANAGED;
1489 	else
1490 		m->oflags = 0;
1491 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ)) == 0)
1492 		m->oflags |= VPO_BUSY;
1493 	if (req & VM_ALLOC_WIRED) {
1494 		/*
1495 		 * The page lock is not required for wiring a page until that
1496 		 * page is inserted into the object.
1497 		 */
1498 		atomic_add_int(&cnt.v_wire_count, 1);
1499 		m->wire_count = 1;
1500 	}
1501 	m->act_count = 0;
1502 
1503 	if (object != NULL) {
1504 		/* Ignore device objects; the pager sets "memattr" for them. */
1505 		if (object->memattr != VM_MEMATTR_DEFAULT &&
1506 		    object->type != OBJT_DEVICE && object->type != OBJT_SG)
1507 			pmap_page_set_memattr(m, object->memattr);
1508 		vm_page_insert(m, object, pindex);
1509 	} else
1510 		m->pindex = pindex;
1511 
1512 	/*
1513 	 * The following call to vdrop() must come after the above call
1514 	 * to vm_page_insert() in case both affect the same object and
1515 	 * vnode.  Otherwise, the affected vnode's hold count could
1516 	 * temporarily become zero.
1517 	 */
1518 	if (vp != NULL)
1519 		vdrop(vp);
1520 
1521 	/*
1522 	 * Don't wakeup too often - wakeup the pageout daemon when
1523 	 * we would be nearly out of memory.
1524 	 */
1525 	if (vm_paging_needed())
1526 		pagedaemon_wakeup();
1527 
1528 	return (m);
1529 }
1530 
1531 /*
1532  *	vm_page_alloc_contig:
1533  *
1534  *	Allocate a contiguous set of physical pages of the given size "npages"
1535  *	from the free lists.  All of the physical pages must be at or above
1536  *	the given physical address "low" and below the given physical address
1537  *	"high".  The given value "alignment" determines the alignment of the
1538  *	first physical page in the set.  If the given value "boundary" is
1539  *	non-zero, then the set of physical pages cannot cross any physical
1540  *	address boundary that is a multiple of that value.  Both "alignment"
1541  *	and "boundary" must be a power of two.
1542  *
1543  *	If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
1544  *	then the memory attribute setting for the physical pages is configured
1545  *	to the object's memory attribute setting.  Otherwise, the memory
1546  *	attribute setting for the physical pages is configured to "memattr",
1547  *	overriding the object's memory attribute setting.  However, if the
1548  *	object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
1549  *	memory attribute setting for the physical pages cannot be configured
1550  *	to VM_MEMATTR_DEFAULT.
1551  *
1552  *	The caller must always specify an allocation class.
1553  *
1554  *	allocation classes:
1555  *	VM_ALLOC_NORMAL		normal process request
1556  *	VM_ALLOC_SYSTEM		system *really* needs a page
1557  *	VM_ALLOC_INTERRUPT	interrupt time request
1558  *
1559  *	optional allocation flags:
1560  *	VM_ALLOC_NOBUSY		do not set the flag VPO_BUSY on the page
1561  *	VM_ALLOC_NOOBJ		page is not associated with an object and
1562  *				should not have the flag VPO_BUSY set
1563  *	VM_ALLOC_WIRED		wire the allocated page
1564  *	VM_ALLOC_ZERO		prefer a zeroed page
1565  *
1566  *	This routine may not sleep.
1567  */
1568 vm_page_t
1569 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
1570     u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
1571     vm_paddr_t boundary, vm_memattr_t memattr)
1572 {
1573 	struct vnode *drop;
1574 	vm_page_t deferred_vdrop_list, m, m_ret;
1575 	u_int flags, oflags;
1576 	int req_class;
1577 
1578 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0),
1579 	    ("vm_page_alloc_contig: inconsistent object/req"));
1580 	if (object != NULL) {
1581 		VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1582 		KASSERT(object->type == OBJT_PHYS,
1583 		    ("vm_page_alloc_contig: object %p isn't OBJT_PHYS",
1584 		    object));
1585 	}
1586 	KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
1587 	req_class = req & VM_ALLOC_CLASS_MASK;
1588 
1589 	/*
1590 	 * The page daemon is allowed to dig deeper into the free page list.
1591 	 */
1592 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1593 		req_class = VM_ALLOC_SYSTEM;
1594 
1595 	deferred_vdrop_list = NULL;
1596 	mtx_lock(&vm_page_queue_free_mtx);
1597 	if (cnt.v_free_count + cnt.v_cache_count >= npages +
1598 	    cnt.v_free_reserved || (req_class == VM_ALLOC_SYSTEM &&
1599 	    cnt.v_free_count + cnt.v_cache_count >= npages +
1600 	    cnt.v_interrupt_free_min) || (req_class == VM_ALLOC_INTERRUPT &&
1601 	    cnt.v_free_count + cnt.v_cache_count >= npages)) {
1602 #if VM_NRESERVLEVEL > 0
1603 retry:
1604 		if (object == NULL || (object->flags & OBJ_COLORED) == 0 ||
1605 		    (m_ret = vm_reserv_alloc_contig(object, pindex, npages,
1606 		    low, high, alignment, boundary)) == NULL)
1607 #endif
1608 			m_ret = vm_phys_alloc_contig(npages, low, high,
1609 			    alignment, boundary);
1610 	} else {
1611 		mtx_unlock(&vm_page_queue_free_mtx);
1612 		atomic_add_int(&vm_pageout_deficit, npages);
1613 		pagedaemon_wakeup();
1614 		return (NULL);
1615 	}
1616 	if (m_ret != NULL)
1617 		for (m = m_ret; m < &m_ret[npages]; m++) {
1618 			drop = vm_page_alloc_init(m);
1619 			if (drop != NULL) {
1620 				/*
1621 				 * Enqueue the vnode for deferred vdrop().
1622 				 *
1623 				 * Once the pages are removed from the free
1624 				 * page list, "pageq" can be safely abused to
1625 				 * construct a short-lived list of vnodes.
1626 				 */
1627 				m->pageq.tqe_prev = (void *)drop;
1628 				m->pageq.tqe_next = deferred_vdrop_list;
1629 				deferred_vdrop_list = m;
1630 			}
1631 		}
1632 	else {
1633 #if VM_NRESERVLEVEL > 0
1634 		if (vm_reserv_reclaim_contig(npages, low, high, alignment,
1635 		    boundary))
1636 			goto retry;
1637 #endif
1638 	}
1639 	mtx_unlock(&vm_page_queue_free_mtx);
1640 	if (m_ret == NULL)
1641 		return (NULL);
1642 
1643 	/*
1644 	 * Initialize the pages.  Only the PG_ZERO flag is inherited.
1645 	 */
1646 	flags = 0;
1647 	if ((req & VM_ALLOC_ZERO) != 0)
1648 		flags = PG_ZERO;
1649 	if ((req & VM_ALLOC_NODUMP) != 0)
1650 		flags |= PG_NODUMP;
1651 	if ((req & VM_ALLOC_WIRED) != 0)
1652 		atomic_add_int(&cnt.v_wire_count, npages);
1653 	oflags = VPO_UNMANAGED;
1654 	if (object != NULL) {
1655 		if ((req & VM_ALLOC_NOBUSY) == 0)
1656 			oflags |= VPO_BUSY;
1657 		if (object->memattr != VM_MEMATTR_DEFAULT &&
1658 		    memattr == VM_MEMATTR_DEFAULT)
1659 			memattr = object->memattr;
1660 	}
1661 	for (m = m_ret; m < &m_ret[npages]; m++) {
1662 		m->aflags = 0;
1663 		m->flags = (m->flags | PG_NODUMP) & flags;
1664 		if ((req & VM_ALLOC_WIRED) != 0)
1665 			m->wire_count = 1;
1666 		/* Unmanaged pages don't use "act_count". */
1667 		m->oflags = oflags;
1668 		if (memattr != VM_MEMATTR_DEFAULT)
1669 			pmap_page_set_memattr(m, memattr);
1670 		if (object != NULL)
1671 			vm_page_insert(m, object, pindex);
1672 		else
1673 			m->pindex = pindex;
1674 		pindex++;
1675 	}
1676 	while (deferred_vdrop_list != NULL) {
1677 		vdrop((struct vnode *)deferred_vdrop_list->pageq.tqe_prev);
1678 		deferred_vdrop_list = deferred_vdrop_list->pageq.tqe_next;
1679 	}
1680 	if (vm_paging_needed())
1681 		pagedaemon_wakeup();
1682 	return (m_ret);
1683 }
1684 
1685 /*
1686  * Initialize a page that has been freshly dequeued from a freelist.
1687  * The caller has to drop the vnode returned, if it is not NULL.
1688  *
1689  * This function may only be used to initialize unmanaged pages.
1690  *
1691  * To be called with vm_page_queue_free_mtx held.
1692  */
1693 static struct vnode *
1694 vm_page_alloc_init(vm_page_t m)
1695 {
1696 	struct vnode *drop;
1697 	vm_object_t m_object;
1698 
1699 	KASSERT(m->queue == PQ_NONE,
1700 	    ("vm_page_alloc_init: page %p has unexpected queue %d",
1701 	    m, m->queue));
1702 	KASSERT(m->wire_count == 0,
1703 	    ("vm_page_alloc_init: page %p is wired", m));
1704 	KASSERT(m->hold_count == 0,
1705 	    ("vm_page_alloc_init: page %p is held", m));
1706 	KASSERT(m->busy == 0,
1707 	    ("vm_page_alloc_init: page %p is busy", m));
1708 	KASSERT(m->dirty == 0,
1709 	    ("vm_page_alloc_init: page %p is dirty", m));
1710 	KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
1711 	    ("vm_page_alloc_init: page %p has unexpected memattr %d",
1712 	    m, pmap_page_get_memattr(m)));
1713 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1714 	drop = NULL;
1715 	if ((m->flags & PG_CACHED) != 0) {
1716 		KASSERT((m->flags & PG_ZERO) == 0,
1717 		    ("vm_page_alloc_init: cached page %p is PG_ZERO", m));
1718 		m->valid = 0;
1719 		m_object = m->object;
1720 		vm_page_cache_remove(m);
1721 		if (m_object->type == OBJT_VNODE && m_object->cache == NULL)
1722 			drop = m_object->handle;
1723 	} else {
1724 		KASSERT(VM_PAGE_IS_FREE(m),
1725 		    ("vm_page_alloc_init: page %p is not free", m));
1726 		KASSERT(m->valid == 0,
1727 		    ("vm_page_alloc_init: free page %p is valid", m));
1728 		cnt.v_free_count--;
1729 		if ((m->flags & PG_ZERO) != 0)
1730 			vm_page_zero_count--;
1731 	}
1732 	/* Don't clear the PG_ZERO flag; we'll need it later. */
1733 	m->flags &= PG_ZERO;
1734 	return (drop);
1735 }
1736 
1737 /*
1738  * 	vm_page_alloc_freelist:
1739  *
1740  *	Allocate a physical page from the specified free page list.
1741  *
1742  *	The caller must always specify an allocation class.
1743  *
1744  *	allocation classes:
1745  *	VM_ALLOC_NORMAL		normal process request
1746  *	VM_ALLOC_SYSTEM		system *really* needs a page
1747  *	VM_ALLOC_INTERRUPT	interrupt time request
1748  *
1749  *	optional allocation flags:
1750  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
1751  *				intends to allocate
1752  *	VM_ALLOC_WIRED		wire the allocated page
1753  *	VM_ALLOC_ZERO		prefer a zeroed page
1754  *
1755  *	This routine may not sleep.
1756  */
1757 vm_page_t
1758 vm_page_alloc_freelist(int flind, int req)
1759 {
1760 	struct vnode *drop;
1761 	vm_page_t m;
1762 	u_int flags;
1763 	int req_class;
1764 
1765 	req_class = req & VM_ALLOC_CLASS_MASK;
1766 
1767 	/*
1768 	 * The page daemon is allowed to dig deeper into the free page list.
1769 	 */
1770 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1771 		req_class = VM_ALLOC_SYSTEM;
1772 
1773 	/*
1774 	 * Do not allocate reserved pages unless the req has asked for it.
1775 	 */
1776 	mtx_lock(&vm_page_queue_free_mtx);
1777 	if (cnt.v_free_count + cnt.v_cache_count > cnt.v_free_reserved ||
1778 	    (req_class == VM_ALLOC_SYSTEM &&
1779 	    cnt.v_free_count + cnt.v_cache_count > cnt.v_interrupt_free_min) ||
1780 	    (req_class == VM_ALLOC_INTERRUPT &&
1781 	    cnt.v_free_count + cnt.v_cache_count > 0))
1782 		m = vm_phys_alloc_freelist_pages(flind, VM_FREEPOOL_DIRECT, 0);
1783 	else {
1784 		mtx_unlock(&vm_page_queue_free_mtx);
1785 		atomic_add_int(&vm_pageout_deficit,
1786 		    max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
1787 		pagedaemon_wakeup();
1788 		return (NULL);
1789 	}
1790 	if (m == NULL) {
1791 		mtx_unlock(&vm_page_queue_free_mtx);
1792 		return (NULL);
1793 	}
1794 	drop = vm_page_alloc_init(m);
1795 	mtx_unlock(&vm_page_queue_free_mtx);
1796 
1797 	/*
1798 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
1799 	 */
1800 	m->aflags = 0;
1801 	flags = 0;
1802 	if ((req & VM_ALLOC_ZERO) != 0)
1803 		flags = PG_ZERO;
1804 	m->flags &= flags;
1805 	if ((req & VM_ALLOC_WIRED) != 0) {
1806 		/*
1807 		 * The page lock is not required for wiring a page that does
1808 		 * not belong to an object.
1809 		 */
1810 		atomic_add_int(&cnt.v_wire_count, 1);
1811 		m->wire_count = 1;
1812 	}
1813 	/* Unmanaged pages don't use "act_count". */
1814 	m->oflags = VPO_UNMANAGED;
1815 	if (drop != NULL)
1816 		vdrop(drop);
1817 	if (vm_paging_needed())
1818 		pagedaemon_wakeup();
1819 	return (m);
1820 }
1821 
1822 /*
1823  *	vm_wait:	(also see VM_WAIT macro)
1824  *
1825  *	Sleep until free pages are available for allocation.
1826  *	- Called in various places before memory allocations.
1827  */
1828 void
1829 vm_wait(void)
1830 {
1831 
1832 	mtx_lock(&vm_page_queue_free_mtx);
1833 	if (curproc == pageproc) {
1834 		vm_pageout_pages_needed = 1;
1835 		msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx,
1836 		    PDROP | PSWP, "VMWait", 0);
1837 	} else {
1838 		if (!vm_pages_needed) {
1839 			vm_pages_needed = 1;
1840 			wakeup(&vm_pages_needed);
1841 		}
1842 		msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PVM,
1843 		    "vmwait", 0);
1844 	}
1845 }
1846 
1847 /*
1848  *	vm_waitpfault:	(also see VM_WAITPFAULT macro)
1849  *
1850  *	Sleep until free pages are available for allocation.
1851  *	- Called only in vm_fault so that processes page faulting
1852  *	  can be easily tracked.
1853  *	- Sleeps at a lower priority than vm_wait() so that vm_wait()ing
1854  *	  processes will be able to grab memory first.  Do not change
1855  *	  this balance without careful testing first.
1856  */
1857 void
1858 vm_waitpfault(void)
1859 {
1860 
1861 	mtx_lock(&vm_page_queue_free_mtx);
1862 	if (!vm_pages_needed) {
1863 		vm_pages_needed = 1;
1864 		wakeup(&vm_pages_needed);
1865 	}
1866 	msleep(&cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PUSER,
1867 	    "pfault", 0);
1868 }
1869 
1870 /*
1871  *	vm_page_queue_remove:
1872  *
1873  *	Remove the given page from the specified queue.
1874  *
1875  *	The page and page queues must be locked.
1876  */
1877 static __inline void
1878 vm_page_queue_remove(int queue, vm_page_t m)
1879 {
1880 	struct vpgqueues *pq;
1881 
1882 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1883 	vm_page_lock_assert(m, MA_OWNED);
1884 	pq = &vm_page_queues[queue];
1885 	TAILQ_REMOVE(&pq->pl, m, pageq);
1886 	(*pq->cnt)--;
1887 }
1888 
1889 /*
1890  *	vm_pageq_remove:
1891  *
1892  *	Remove a page from its queue.
1893  *
1894  *	The given page must be locked.
1895  */
1896 void
1897 vm_pageq_remove(vm_page_t m)
1898 {
1899 	int queue;
1900 
1901 	vm_page_lock_assert(m, MA_OWNED);
1902 	if ((queue = m->queue) != PQ_NONE) {
1903 		vm_page_lock_queues();
1904 		m->queue = PQ_NONE;
1905 		vm_page_queue_remove(queue, m);
1906 		vm_page_unlock_queues();
1907 	}
1908 }
1909 
1910 /*
1911  *	vm_page_enqueue:
1912  *
1913  *	Add the given page to the specified queue.
1914  *
1915  *	The page queues must be locked.
1916  */
1917 static void
1918 vm_page_enqueue(int queue, vm_page_t m)
1919 {
1920 	struct vpgqueues *vpq;
1921 
1922 	vpq = &vm_page_queues[queue];
1923 	m->queue = queue;
1924 	TAILQ_INSERT_TAIL(&vpq->pl, m, pageq);
1925 	++*vpq->cnt;
1926 }
1927 
1928 /*
1929  *	vm_page_activate:
1930  *
1931  *	Put the specified page on the active list (if appropriate).
1932  *	Ensure that act_count is at least ACT_INIT but do not otherwise
1933  *	mess with it.
1934  *
1935  *	The page must be locked.
1936  */
1937 void
1938 vm_page_activate(vm_page_t m)
1939 {
1940 	int queue;
1941 
1942 	vm_page_lock_assert(m, MA_OWNED);
1943 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1944 	if ((queue = m->queue) != PQ_ACTIVE) {
1945 		if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
1946 			if (m->act_count < ACT_INIT)
1947 				m->act_count = ACT_INIT;
1948 			vm_page_lock_queues();
1949 			if (queue != PQ_NONE)
1950 				vm_page_queue_remove(queue, m);
1951 			vm_page_enqueue(PQ_ACTIVE, m);
1952 			vm_page_unlock_queues();
1953 		} else
1954 			KASSERT(queue == PQ_NONE,
1955 			    ("vm_page_activate: wired page %p is queued", m));
1956 	} else {
1957 		if (m->act_count < ACT_INIT)
1958 			m->act_count = ACT_INIT;
1959 	}
1960 }
1961 
1962 /*
1963  *	vm_page_free_wakeup:
1964  *
1965  *	Helper routine for vm_page_free_toq() and vm_page_cache().  This
1966  *	routine is called when a page has been added to the cache or free
1967  *	queues.
1968  *
1969  *	The page queues must be locked.
1970  */
1971 static inline void
1972 vm_page_free_wakeup(void)
1973 {
1974 
1975 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1976 	/*
1977 	 * if pageout daemon needs pages, then tell it that there are
1978 	 * some free.
1979 	 */
1980 	if (vm_pageout_pages_needed &&
1981 	    cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) {
1982 		wakeup(&vm_pageout_pages_needed);
1983 		vm_pageout_pages_needed = 0;
1984 	}
1985 	/*
1986 	 * wakeup processes that are waiting on memory if we hit a
1987 	 * high water mark. And wakeup scheduler process if we have
1988 	 * lots of memory. this process will swapin processes.
1989 	 */
1990 	if (vm_pages_needed && !vm_page_count_min()) {
1991 		vm_pages_needed = 0;
1992 		wakeup(&cnt.v_free_count);
1993 	}
1994 }
1995 
1996 /*
1997  *	vm_page_free_toq:
1998  *
1999  *	Returns the given page to the free list,
2000  *	disassociating it with any VM object.
2001  *
2002  *	The object must be locked.  The page must be locked if it is managed.
2003  */
2004 void
2005 vm_page_free_toq(vm_page_t m)
2006 {
2007 
2008 	if ((m->oflags & VPO_UNMANAGED) == 0) {
2009 		vm_page_lock_assert(m, MA_OWNED);
2010 		KASSERT(!pmap_page_is_mapped(m),
2011 		    ("vm_page_free_toq: freeing mapped page %p", m));
2012 	}
2013 	PCPU_INC(cnt.v_tfree);
2014 
2015 	if (VM_PAGE_IS_FREE(m))
2016 		panic("vm_page_free: freeing free page %p", m);
2017 	else if (m->busy != 0)
2018 		panic("vm_page_free: freeing busy page %p", m);
2019 
2020 	/*
2021 	 * Unqueue, then remove page.  Note that we cannot destroy
2022 	 * the page here because we do not want to call the pager's
2023 	 * callback routine until after we've put the page on the
2024 	 * appropriate free queue.
2025 	 */
2026 	if ((m->oflags & VPO_UNMANAGED) == 0)
2027 		vm_pageq_remove(m);
2028 	vm_page_remove(m);
2029 
2030 	/*
2031 	 * If fictitious remove object association and
2032 	 * return, otherwise delay object association removal.
2033 	 */
2034 	if ((m->flags & PG_FICTITIOUS) != 0) {
2035 		return;
2036 	}
2037 
2038 	m->valid = 0;
2039 	vm_page_undirty(m);
2040 
2041 	if (m->wire_count != 0)
2042 		panic("vm_page_free: freeing wired page %p", m);
2043 	if (m->hold_count != 0) {
2044 		m->flags &= ~PG_ZERO;
2045 		vm_page_lock_queues();
2046 		vm_page_enqueue(PQ_HOLD, m);
2047 		vm_page_unlock_queues();
2048 	} else {
2049 		/*
2050 		 * Restore the default memory attribute to the page.
2051 		 */
2052 		if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
2053 			pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
2054 
2055 		/*
2056 		 * Insert the page into the physical memory allocator's
2057 		 * cache/free page queues.
2058 		 */
2059 		mtx_lock(&vm_page_queue_free_mtx);
2060 		m->flags |= PG_FREE;
2061 		cnt.v_free_count++;
2062 #if VM_NRESERVLEVEL > 0
2063 		if (!vm_reserv_free_page(m))
2064 #else
2065 		if (TRUE)
2066 #endif
2067 			vm_phys_free_pages(m, 0);
2068 		if ((m->flags & PG_ZERO) != 0)
2069 			++vm_page_zero_count;
2070 		else
2071 			vm_page_zero_idle_wakeup();
2072 		vm_page_free_wakeup();
2073 		mtx_unlock(&vm_page_queue_free_mtx);
2074 	}
2075 }
2076 
2077 /*
2078  *	vm_page_wire:
2079  *
2080  *	Mark this page as wired down by yet
2081  *	another map, removing it from paging queues
2082  *	as necessary.
2083  *
2084  *	If the page is fictitious, then its wire count must remain one.
2085  *
2086  *	The page must be locked.
2087  */
2088 void
2089 vm_page_wire(vm_page_t m)
2090 {
2091 
2092 	/*
2093 	 * Only bump the wire statistics if the page is not already wired,
2094 	 * and only unqueue the page if it is on some queue (if it is unmanaged
2095 	 * it is already off the queues).
2096 	 */
2097 	vm_page_lock_assert(m, MA_OWNED);
2098 	if ((m->flags & PG_FICTITIOUS) != 0) {
2099 		KASSERT(m->wire_count == 1,
2100 		    ("vm_page_wire: fictitious page %p's wire count isn't one",
2101 		    m));
2102 		return;
2103 	}
2104 	if (m->wire_count == 0) {
2105 		if ((m->oflags & VPO_UNMANAGED) == 0)
2106 			vm_pageq_remove(m);
2107 		atomic_add_int(&cnt.v_wire_count, 1);
2108 	}
2109 	m->wire_count++;
2110 	KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m));
2111 }
2112 
2113 /*
2114  * vm_page_unwire:
2115  *
2116  * Release one wiring of the specified page, potentially enabling it to be
2117  * paged again.  If paging is enabled, then the value of the parameter
2118  * "activate" determines to which queue the page is added.  If "activate" is
2119  * non-zero, then the page is added to the active queue.  Otherwise, it is
2120  * added to the inactive queue.
2121  *
2122  * However, unless the page belongs to an object, it is not enqueued because
2123  * it cannot be paged out.
2124  *
2125  * If a page is fictitious, then its wire count must alway be one.
2126  *
2127  * A managed page must be locked.
2128  */
2129 void
2130 vm_page_unwire(vm_page_t m, int activate)
2131 {
2132 
2133 	if ((m->oflags & VPO_UNMANAGED) == 0)
2134 		vm_page_lock_assert(m, MA_OWNED);
2135 	if ((m->flags & PG_FICTITIOUS) != 0) {
2136 		KASSERT(m->wire_count == 1,
2137 	    ("vm_page_unwire: fictitious page %p's wire count isn't one", m));
2138 		return;
2139 	}
2140 	if (m->wire_count > 0) {
2141 		m->wire_count--;
2142 		if (m->wire_count == 0) {
2143 			atomic_subtract_int(&cnt.v_wire_count, 1);
2144 			if ((m->oflags & VPO_UNMANAGED) != 0 ||
2145 			    m->object == NULL)
2146 				return;
2147 			if (!activate)
2148 				m->flags &= ~PG_WINATCFLS;
2149 			vm_page_lock_queues();
2150 			vm_page_enqueue(activate ? PQ_ACTIVE : PQ_INACTIVE, m);
2151 			vm_page_unlock_queues();
2152 		}
2153 	} else
2154 		panic("vm_page_unwire: page %p's wire count is zero", m);
2155 }
2156 
2157 /*
2158  * Move the specified page to the inactive queue.
2159  *
2160  * Many pages placed on the inactive queue should actually go
2161  * into the cache, but it is difficult to figure out which.  What
2162  * we do instead, if the inactive target is well met, is to put
2163  * clean pages at the head of the inactive queue instead of the tail.
2164  * This will cause them to be moved to the cache more quickly and
2165  * if not actively re-referenced, reclaimed more quickly.  If we just
2166  * stick these pages at the end of the inactive queue, heavy filesystem
2167  * meta-data accesses can cause an unnecessary paging load on memory bound
2168  * processes.  This optimization causes one-time-use metadata to be
2169  * reused more quickly.
2170  *
2171  * Normally athead is 0 resulting in LRU operation.  athead is set
2172  * to 1 if we want this page to be 'as if it were placed in the cache',
2173  * except without unmapping it from the process address space.
2174  *
2175  * The page must be locked.
2176  */
2177 static inline void
2178 _vm_page_deactivate(vm_page_t m, int athead)
2179 {
2180 	int queue;
2181 
2182 	vm_page_lock_assert(m, MA_OWNED);
2183 
2184 	/*
2185 	 * Ignore if already inactive.
2186 	 */
2187 	if ((queue = m->queue) == PQ_INACTIVE)
2188 		return;
2189 	if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
2190 		m->flags &= ~PG_WINATCFLS;
2191 		vm_page_lock_queues();
2192 		if (queue != PQ_NONE)
2193 			vm_page_queue_remove(queue, m);
2194 		if (athead)
2195 			TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m,
2196 			    pageq);
2197 		else
2198 			TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m,
2199 			    pageq);
2200 		m->queue = PQ_INACTIVE;
2201 		cnt.v_inactive_count++;
2202 		vm_page_unlock_queues();
2203 	}
2204 }
2205 
2206 /*
2207  * Move the specified page to the inactive queue.
2208  *
2209  * The page must be locked.
2210  */
2211 void
2212 vm_page_deactivate(vm_page_t m)
2213 {
2214 
2215 	_vm_page_deactivate(m, 0);
2216 }
2217 
2218 /*
2219  * vm_page_try_to_cache:
2220  *
2221  * Returns 0 on failure, 1 on success
2222  */
2223 int
2224 vm_page_try_to_cache(vm_page_t m)
2225 {
2226 
2227 	vm_page_lock_assert(m, MA_OWNED);
2228 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2229 	if (m->dirty || m->hold_count || m->busy || m->wire_count ||
2230 	    (m->oflags & (VPO_BUSY | VPO_UNMANAGED)) != 0)
2231 		return (0);
2232 	pmap_remove_all(m);
2233 	if (m->dirty)
2234 		return (0);
2235 	vm_page_cache(m);
2236 	return (1);
2237 }
2238 
2239 /*
2240  * vm_page_try_to_free()
2241  *
2242  *	Attempt to free the page.  If we cannot free it, we do nothing.
2243  *	1 is returned on success, 0 on failure.
2244  */
2245 int
2246 vm_page_try_to_free(vm_page_t m)
2247 {
2248 
2249 	vm_page_lock_assert(m, MA_OWNED);
2250 	if (m->object != NULL)
2251 		VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2252 	if (m->dirty || m->hold_count || m->busy || m->wire_count ||
2253 	    (m->oflags & (VPO_BUSY | VPO_UNMANAGED)) != 0)
2254 		return (0);
2255 	pmap_remove_all(m);
2256 	if (m->dirty)
2257 		return (0);
2258 	vm_page_free(m);
2259 	return (1);
2260 }
2261 
2262 /*
2263  * vm_page_cache
2264  *
2265  * Put the specified page onto the page cache queue (if appropriate).
2266  *
2267  * The object and page must be locked.
2268  */
2269 void
2270 vm_page_cache(vm_page_t m)
2271 {
2272 	vm_object_t object;
2273 	vm_page_t next, prev, root;
2274 
2275 	vm_page_lock_assert(m, MA_OWNED);
2276 	object = m->object;
2277 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2278 	if ((m->oflags & (VPO_UNMANAGED | VPO_BUSY)) || m->busy ||
2279 	    m->hold_count || m->wire_count)
2280 		panic("vm_page_cache: attempting to cache busy page");
2281 	pmap_remove_all(m);
2282 	if (m->dirty != 0)
2283 		panic("vm_page_cache: page %p is dirty", m);
2284 	if (m->valid == 0 || object->type == OBJT_DEFAULT ||
2285 	    (object->type == OBJT_SWAP &&
2286 	    !vm_pager_has_page(object, m->pindex, NULL, NULL))) {
2287 		/*
2288 		 * Hypothesis: A cache-elgible page belonging to a
2289 		 * default object or swap object but without a backing
2290 		 * store must be zero filled.
2291 		 */
2292 		vm_page_free(m);
2293 		return;
2294 	}
2295 	KASSERT((m->flags & PG_CACHED) == 0,
2296 	    ("vm_page_cache: page %p is already cached", m));
2297 	PCPU_INC(cnt.v_tcached);
2298 
2299 	/*
2300 	 * Remove the page from the paging queues.
2301 	 */
2302 	vm_pageq_remove(m);
2303 
2304 	/*
2305 	 * Remove the page from the object's collection of resident
2306 	 * pages.
2307 	 */
2308 	if ((next = TAILQ_NEXT(m, listq)) != NULL && next->left == m) {
2309 		/*
2310 		 * Since the page's successor in the list is also its parent
2311 		 * in the tree, its right subtree must be empty.
2312 		 */
2313 		next->left = m->left;
2314 		KASSERT(m->right == NULL,
2315 		    ("vm_page_cache: page %p has right child", m));
2316 	} else if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL &&
2317 	    prev->right == m) {
2318 		/*
2319 		 * Since the page's predecessor in the list is also its parent
2320 		 * in the tree, its left subtree must be empty.
2321 		 */
2322 		KASSERT(m->left == NULL,
2323 		    ("vm_page_cache: page %p has left child", m));
2324 		prev->right = m->right;
2325 	} else {
2326 		if (m != object->root)
2327 			vm_page_splay(m->pindex, object->root);
2328 		if (m->left == NULL)
2329 			root = m->right;
2330 		else if (m->right == NULL)
2331 			root = m->left;
2332 		else {
2333 			/*
2334 			 * Move the page's successor to the root, because
2335 			 * pages are usually removed in ascending order.
2336 			 */
2337 			if (m->right != next)
2338 				vm_page_splay(m->pindex, m->right);
2339 			next->left = m->left;
2340 			root = next;
2341 		}
2342 		object->root = root;
2343 	}
2344 	TAILQ_REMOVE(&object->memq, m, listq);
2345 	object->resident_page_count--;
2346 
2347 	/*
2348 	 * Restore the default memory attribute to the page.
2349 	 */
2350 	if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
2351 		pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
2352 
2353 	/*
2354 	 * Insert the page into the object's collection of cached pages
2355 	 * and the physical memory allocator's cache/free page queues.
2356 	 */
2357 	m->flags &= ~PG_ZERO;
2358 	mtx_lock(&vm_page_queue_free_mtx);
2359 	m->flags |= PG_CACHED;
2360 	cnt.v_cache_count++;
2361 	root = object->cache;
2362 	if (root == NULL) {
2363 		m->left = NULL;
2364 		m->right = NULL;
2365 	} else {
2366 		root = vm_page_splay(m->pindex, root);
2367 		if (m->pindex < root->pindex) {
2368 			m->left = root->left;
2369 			m->right = root;
2370 			root->left = NULL;
2371 		} else if (__predict_false(m->pindex == root->pindex))
2372 			panic("vm_page_cache: offset already cached");
2373 		else {
2374 			m->right = root->right;
2375 			m->left = root;
2376 			root->right = NULL;
2377 		}
2378 	}
2379 	object->cache = m;
2380 #if VM_NRESERVLEVEL > 0
2381 	if (!vm_reserv_free_page(m)) {
2382 #else
2383 	if (TRUE) {
2384 #endif
2385 		vm_phys_set_pool(VM_FREEPOOL_CACHE, m, 0);
2386 		vm_phys_free_pages(m, 0);
2387 	}
2388 	vm_page_free_wakeup();
2389 	mtx_unlock(&vm_page_queue_free_mtx);
2390 
2391 	/*
2392 	 * Increment the vnode's hold count if this is the object's only
2393 	 * cached page.  Decrement the vnode's hold count if this was
2394 	 * the object's only resident page.
2395 	 */
2396 	if (object->type == OBJT_VNODE) {
2397 		if (root == NULL && object->resident_page_count != 0)
2398 			vhold(object->handle);
2399 		else if (root != NULL && object->resident_page_count == 0)
2400 			vdrop(object->handle);
2401 	}
2402 }
2403 
2404 /*
2405  * vm_page_dontneed
2406  *
2407  *	Cache, deactivate, or do nothing as appropriate.  This routine
2408  *	is typically used by madvise() MADV_DONTNEED.
2409  *
2410  *	Generally speaking we want to move the page into the cache so
2411  *	it gets reused quickly.  However, this can result in a silly syndrome
2412  *	due to the page recycling too quickly.  Small objects will not be
2413  *	fully cached.  On the otherhand, if we move the page to the inactive
2414  *	queue we wind up with a problem whereby very large objects
2415  *	unnecessarily blow away our inactive and cache queues.
2416  *
2417  *	The solution is to move the pages based on a fixed weighting.  We
2418  *	either leave them alone, deactivate them, or move them to the cache,
2419  *	where moving them to the cache has the highest weighting.
2420  *	By forcing some pages into other queues we eventually force the
2421  *	system to balance the queues, potentially recovering other unrelated
2422  *	space from active.  The idea is to not force this to happen too
2423  *	often.
2424  *
2425  *	The object and page must be locked.
2426  */
2427 void
2428 vm_page_dontneed(vm_page_t m)
2429 {
2430 	int dnw;
2431 	int head;
2432 
2433 	vm_page_lock_assert(m, MA_OWNED);
2434 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2435 	dnw = PCPU_GET(dnweight);
2436 	PCPU_INC(dnweight);
2437 
2438 	/*
2439 	 * Occasionally leave the page alone.
2440 	 */
2441 	if ((dnw & 0x01F0) == 0 || m->queue == PQ_INACTIVE) {
2442 		if (m->act_count >= ACT_INIT)
2443 			--m->act_count;
2444 		return;
2445 	}
2446 
2447 	/*
2448 	 * Clear any references to the page.  Otherwise, the page daemon will
2449 	 * immediately reactivate the page.
2450 	 *
2451 	 * Perform the pmap_clear_reference() first.  Otherwise, a concurrent
2452 	 * pmap operation, such as pmap_remove(), could clear a reference in
2453 	 * the pmap and set PGA_REFERENCED on the page before the
2454 	 * pmap_clear_reference() had completed.  Consequently, the page would
2455 	 * appear referenced based upon an old reference that occurred before
2456 	 * this function ran.
2457 	 */
2458 	pmap_clear_reference(m);
2459 	vm_page_aflag_clear(m, PGA_REFERENCED);
2460 
2461 	if (m->dirty == 0 && pmap_is_modified(m))
2462 		vm_page_dirty(m);
2463 
2464 	if (m->dirty || (dnw & 0x0070) == 0) {
2465 		/*
2466 		 * Deactivate the page 3 times out of 32.
2467 		 */
2468 		head = 0;
2469 	} else {
2470 		/*
2471 		 * Cache the page 28 times out of every 32.  Note that
2472 		 * the page is deactivated instead of cached, but placed
2473 		 * at the head of the queue instead of the tail.
2474 		 */
2475 		head = 1;
2476 	}
2477 	_vm_page_deactivate(m, head);
2478 }
2479 
2480 /*
2481  * Grab a page, waiting until we are waken up due to the page
2482  * changing state.  We keep on waiting, if the page continues
2483  * to be in the object.  If the page doesn't exist, first allocate it
2484  * and then conditionally zero it.
2485  *
2486  * The caller must always specify the VM_ALLOC_RETRY flag.  This is intended
2487  * to facilitate its eventual removal.
2488  *
2489  * This routine may sleep.
2490  *
2491  * The object must be locked on entry.  The lock will, however, be released
2492  * and reacquired if the routine sleeps.
2493  */
2494 vm_page_t
2495 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
2496 {
2497 	vm_page_t m;
2498 
2499 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2500 	KASSERT((allocflags & VM_ALLOC_RETRY) != 0,
2501 	    ("vm_page_grab: VM_ALLOC_RETRY is required"));
2502 retrylookup:
2503 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
2504 		if ((m->oflags & VPO_BUSY) != 0 ||
2505 		    ((allocflags & VM_ALLOC_IGN_SBUSY) == 0 && m->busy != 0)) {
2506 			/*
2507 			 * Reference the page before unlocking and
2508 			 * sleeping so that the page daemon is less
2509 			 * likely to reclaim it.
2510 			 */
2511 			vm_page_aflag_set(m, PGA_REFERENCED);
2512 			vm_page_sleep(m, "pgrbwt");
2513 			goto retrylookup;
2514 		} else {
2515 			if ((allocflags & VM_ALLOC_WIRED) != 0) {
2516 				vm_page_lock(m);
2517 				vm_page_wire(m);
2518 				vm_page_unlock(m);
2519 			}
2520 			if ((allocflags & VM_ALLOC_NOBUSY) == 0)
2521 				vm_page_busy(m);
2522 			return (m);
2523 		}
2524 	}
2525 	m = vm_page_alloc(object, pindex, allocflags & ~(VM_ALLOC_RETRY |
2526 	    VM_ALLOC_IGN_SBUSY));
2527 	if (m == NULL) {
2528 		VM_OBJECT_UNLOCK(object);
2529 		VM_WAIT;
2530 		VM_OBJECT_LOCK(object);
2531 		goto retrylookup;
2532 	} else if (m->valid != 0)
2533 		return (m);
2534 	if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
2535 		pmap_zero_page(m);
2536 	return (m);
2537 }
2538 
2539 /*
2540  * Mapping function for valid or dirty bits in a page.
2541  *
2542  * Inputs are required to range within a page.
2543  */
2544 vm_page_bits_t
2545 vm_page_bits(int base, int size)
2546 {
2547 	int first_bit;
2548 	int last_bit;
2549 
2550 	KASSERT(
2551 	    base + size <= PAGE_SIZE,
2552 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
2553 	);
2554 
2555 	if (size == 0)		/* handle degenerate case */
2556 		return (0);
2557 
2558 	first_bit = base >> DEV_BSHIFT;
2559 	last_bit = (base + size - 1) >> DEV_BSHIFT;
2560 
2561 	return (((vm_page_bits_t)2 << last_bit) -
2562 	    ((vm_page_bits_t)1 << first_bit));
2563 }
2564 
2565 /*
2566  *	vm_page_set_valid_range:
2567  *
2568  *	Sets portions of a page valid.  The arguments are expected
2569  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
2570  *	of any partial chunks touched by the range.  The invalid portion of
2571  *	such chunks will be zeroed.
2572  *
2573  *	(base + size) must be less then or equal to PAGE_SIZE.
2574  */
2575 void
2576 vm_page_set_valid_range(vm_page_t m, int base, int size)
2577 {
2578 	int endoff, frag;
2579 
2580 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2581 	if (size == 0)	/* handle degenerate case */
2582 		return;
2583 
2584 	/*
2585 	 * If the base is not DEV_BSIZE aligned and the valid
2586 	 * bit is clear, we have to zero out a portion of the
2587 	 * first block.
2588 	 */
2589 	if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
2590 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
2591 		pmap_zero_page_area(m, frag, base - frag);
2592 
2593 	/*
2594 	 * If the ending offset is not DEV_BSIZE aligned and the
2595 	 * valid bit is clear, we have to zero out a portion of
2596 	 * the last block.
2597 	 */
2598 	endoff = base + size;
2599 	if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
2600 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
2601 		pmap_zero_page_area(m, endoff,
2602 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
2603 
2604 	/*
2605 	 * Assert that no previously invalid block that is now being validated
2606 	 * is already dirty.
2607 	 */
2608 	KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
2609 	    ("vm_page_set_valid_range: page %p is dirty", m));
2610 
2611 	/*
2612 	 * Set valid bits inclusive of any overlap.
2613 	 */
2614 	m->valid |= vm_page_bits(base, size);
2615 }
2616 
2617 /*
2618  * Clear the given bits from the specified page's dirty field.
2619  */
2620 static __inline void
2621 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
2622 {
2623 	uintptr_t addr;
2624 #if PAGE_SIZE < 16384
2625 	int shift;
2626 #endif
2627 
2628 	/*
2629 	 * If the object is locked and the page is neither VPO_BUSY nor
2630 	 * write mapped, then the page's dirty field cannot possibly be
2631 	 * set by a concurrent pmap operation.
2632 	 */
2633 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2634 	if ((m->oflags & VPO_BUSY) == 0 && !pmap_page_is_write_mapped(m))
2635 		m->dirty &= ~pagebits;
2636 	else {
2637 		/*
2638 		 * The pmap layer can call vm_page_dirty() without
2639 		 * holding a distinguished lock.  The combination of
2640 		 * the object's lock and an atomic operation suffice
2641 		 * to guarantee consistency of the page dirty field.
2642 		 *
2643 		 * For PAGE_SIZE == 32768 case, compiler already
2644 		 * properly aligns the dirty field, so no forcible
2645 		 * alignment is needed. Only require existence of
2646 		 * atomic_clear_64 when page size is 32768.
2647 		 */
2648 		addr = (uintptr_t)&m->dirty;
2649 #if PAGE_SIZE == 32768
2650 		atomic_clear_64((uint64_t *)addr, pagebits);
2651 #elif PAGE_SIZE == 16384
2652 		atomic_clear_32((uint32_t *)addr, pagebits);
2653 #else		/* PAGE_SIZE <= 8192 */
2654 		/*
2655 		 * Use a trick to perform a 32-bit atomic on the
2656 		 * containing aligned word, to not depend on the existence
2657 		 * of atomic_clear_{8, 16}.
2658 		 */
2659 		shift = addr & (sizeof(uint32_t) - 1);
2660 #if BYTE_ORDER == BIG_ENDIAN
2661 		shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY;
2662 #else
2663 		shift *= NBBY;
2664 #endif
2665 		addr &= ~(sizeof(uint32_t) - 1);
2666 		atomic_clear_32((uint32_t *)addr, pagebits << shift);
2667 #endif		/* PAGE_SIZE */
2668 	}
2669 }
2670 
2671 /*
2672  *	vm_page_set_validclean:
2673  *
2674  *	Sets portions of a page valid and clean.  The arguments are expected
2675  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
2676  *	of any partial chunks touched by the range.  The invalid portion of
2677  *	such chunks will be zero'd.
2678  *
2679  *	(base + size) must be less then or equal to PAGE_SIZE.
2680  */
2681 void
2682 vm_page_set_validclean(vm_page_t m, int base, int size)
2683 {
2684 	vm_page_bits_t oldvalid, pagebits;
2685 	int endoff, frag;
2686 
2687 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2688 	if (size == 0)	/* handle degenerate case */
2689 		return;
2690 
2691 	/*
2692 	 * If the base is not DEV_BSIZE aligned and the valid
2693 	 * bit is clear, we have to zero out a portion of the
2694 	 * first block.
2695 	 */
2696 	if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
2697 	    (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
2698 		pmap_zero_page_area(m, frag, base - frag);
2699 
2700 	/*
2701 	 * If the ending offset is not DEV_BSIZE aligned and the
2702 	 * valid bit is clear, we have to zero out a portion of
2703 	 * the last block.
2704 	 */
2705 	endoff = base + size;
2706 	if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
2707 	    (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
2708 		pmap_zero_page_area(m, endoff,
2709 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
2710 
2711 	/*
2712 	 * Set valid, clear dirty bits.  If validating the entire
2713 	 * page we can safely clear the pmap modify bit.  We also
2714 	 * use this opportunity to clear the VPO_NOSYNC flag.  If a process
2715 	 * takes a write fault on a MAP_NOSYNC memory area the flag will
2716 	 * be set again.
2717 	 *
2718 	 * We set valid bits inclusive of any overlap, but we can only
2719 	 * clear dirty bits for DEV_BSIZE chunks that are fully within
2720 	 * the range.
2721 	 */
2722 	oldvalid = m->valid;
2723 	pagebits = vm_page_bits(base, size);
2724 	m->valid |= pagebits;
2725 #if 0	/* NOT YET */
2726 	if ((frag = base & (DEV_BSIZE - 1)) != 0) {
2727 		frag = DEV_BSIZE - frag;
2728 		base += frag;
2729 		size -= frag;
2730 		if (size < 0)
2731 			size = 0;
2732 	}
2733 	pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
2734 #endif
2735 	if (base == 0 && size == PAGE_SIZE) {
2736 		/*
2737 		 * The page can only be modified within the pmap if it is
2738 		 * mapped, and it can only be mapped if it was previously
2739 		 * fully valid.
2740 		 */
2741 		if (oldvalid == VM_PAGE_BITS_ALL)
2742 			/*
2743 			 * Perform the pmap_clear_modify() first.  Otherwise,
2744 			 * a concurrent pmap operation, such as
2745 			 * pmap_protect(), could clear a modification in the
2746 			 * pmap and set the dirty field on the page before
2747 			 * pmap_clear_modify() had begun and after the dirty
2748 			 * field was cleared here.
2749 			 */
2750 			pmap_clear_modify(m);
2751 		m->dirty = 0;
2752 		m->oflags &= ~VPO_NOSYNC;
2753 	} else if (oldvalid != VM_PAGE_BITS_ALL)
2754 		m->dirty &= ~pagebits;
2755 	else
2756 		vm_page_clear_dirty_mask(m, pagebits);
2757 }
2758 
2759 void
2760 vm_page_clear_dirty(vm_page_t m, int base, int size)
2761 {
2762 
2763 	vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
2764 }
2765 
2766 /*
2767  *	vm_page_set_invalid:
2768  *
2769  *	Invalidates DEV_BSIZE'd chunks within a page.  Both the
2770  *	valid and dirty bits for the effected areas are cleared.
2771  */
2772 void
2773 vm_page_set_invalid(vm_page_t m, int base, int size)
2774 {
2775 	vm_page_bits_t bits;
2776 
2777 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2778 	KASSERT((m->oflags & VPO_BUSY) == 0,
2779 	    ("vm_page_set_invalid: page %p is busy", m));
2780 	bits = vm_page_bits(base, size);
2781 	if (m->valid == VM_PAGE_BITS_ALL && bits != 0)
2782 		pmap_remove_all(m);
2783 	KASSERT(!pmap_page_is_mapped(m),
2784 	    ("vm_page_set_invalid: page %p is mapped", m));
2785 	m->valid &= ~bits;
2786 	m->dirty &= ~bits;
2787 }
2788 
2789 /*
2790  * vm_page_zero_invalid()
2791  *
2792  *	The kernel assumes that the invalid portions of a page contain
2793  *	garbage, but such pages can be mapped into memory by user code.
2794  *	When this occurs, we must zero out the non-valid portions of the
2795  *	page so user code sees what it expects.
2796  *
2797  *	Pages are most often semi-valid when the end of a file is mapped
2798  *	into memory and the file's size is not page aligned.
2799  */
2800 void
2801 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
2802 {
2803 	int b;
2804 	int i;
2805 
2806 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2807 	/*
2808 	 * Scan the valid bits looking for invalid sections that
2809 	 * must be zerod.  Invalid sub-DEV_BSIZE'd areas ( where the
2810 	 * valid bit may be set ) have already been zerod by
2811 	 * vm_page_set_validclean().
2812 	 */
2813 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
2814 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
2815 		    (m->valid & ((vm_page_bits_t)1 << i))) {
2816 			if (i > b) {
2817 				pmap_zero_page_area(m,
2818 				    b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
2819 			}
2820 			b = i + 1;
2821 		}
2822 	}
2823 
2824 	/*
2825 	 * setvalid is TRUE when we can safely set the zero'd areas
2826 	 * as being valid.  We can do this if there are no cache consistancy
2827 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
2828 	 */
2829 	if (setvalid)
2830 		m->valid = VM_PAGE_BITS_ALL;
2831 }
2832 
2833 /*
2834  *	vm_page_is_valid:
2835  *
2836  *	Is (partial) page valid?  Note that the case where size == 0
2837  *	will return FALSE in the degenerate case where the page is
2838  *	entirely invalid, and TRUE otherwise.
2839  */
2840 int
2841 vm_page_is_valid(vm_page_t m, int base, int size)
2842 {
2843 	vm_page_bits_t bits;
2844 
2845 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2846 	bits = vm_page_bits(base, size);
2847 	if (m->valid && ((m->valid & bits) == bits))
2848 		return 1;
2849 	else
2850 		return 0;
2851 }
2852 
2853 /*
2854  * Set the page's dirty bits if the page is modified.
2855  */
2856 void
2857 vm_page_test_dirty(vm_page_t m)
2858 {
2859 
2860 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2861 	if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
2862 		vm_page_dirty(m);
2863 }
2864 
2865 void
2866 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
2867 {
2868 
2869 	mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
2870 }
2871 
2872 void
2873 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
2874 {
2875 
2876 	mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
2877 }
2878 
2879 int
2880 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
2881 {
2882 
2883 	return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
2884 }
2885 
2886 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
2887 void
2888 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
2889 {
2890 
2891 	mtx_assert_(vm_page_lockptr(m), a, file, line);
2892 }
2893 #endif
2894 
2895 int so_zerocp_fullpage = 0;
2896 
2897 /*
2898  *	Replace the given page with a copy.  The copied page assumes
2899  *	the portion of the given page's "wire_count" that is not the
2900  *	responsibility of this copy-on-write mechanism.
2901  *
2902  *	The object containing the given page must have a non-zero
2903  *	paging-in-progress count and be locked.
2904  */
2905 void
2906 vm_page_cowfault(vm_page_t m)
2907 {
2908 	vm_page_t mnew;
2909 	vm_object_t object;
2910 	vm_pindex_t pindex;
2911 
2912 	vm_page_lock_assert(m, MA_OWNED);
2913 	object = m->object;
2914 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2915 	KASSERT(object->paging_in_progress != 0,
2916 	    ("vm_page_cowfault: object %p's paging-in-progress count is zero.",
2917 	    object));
2918 	pindex = m->pindex;
2919 
2920  retry_alloc:
2921 	pmap_remove_all(m);
2922 	vm_page_remove(m);
2923 	mnew = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL | VM_ALLOC_NOBUSY);
2924 	if (mnew == NULL) {
2925 		vm_page_insert(m, object, pindex);
2926 		vm_page_unlock(m);
2927 		VM_OBJECT_UNLOCK(object);
2928 		VM_WAIT;
2929 		VM_OBJECT_LOCK(object);
2930 		if (m == vm_page_lookup(object, pindex)) {
2931 			vm_page_lock(m);
2932 			goto retry_alloc;
2933 		} else {
2934 			/*
2935 			 * Page disappeared during the wait.
2936 			 */
2937 			return;
2938 		}
2939 	}
2940 
2941 	if (m->cow == 0) {
2942 		/*
2943 		 * check to see if we raced with an xmit complete when
2944 		 * waiting to allocate a page.  If so, put things back
2945 		 * the way they were
2946 		 */
2947 		vm_page_unlock(m);
2948 		vm_page_lock(mnew);
2949 		vm_page_free(mnew);
2950 		vm_page_unlock(mnew);
2951 		vm_page_insert(m, object, pindex);
2952 	} else { /* clear COW & copy page */
2953 		if (!so_zerocp_fullpage)
2954 			pmap_copy_page(m, mnew);
2955 		mnew->valid = VM_PAGE_BITS_ALL;
2956 		vm_page_dirty(mnew);
2957 		mnew->wire_count = m->wire_count - m->cow;
2958 		m->wire_count = m->cow;
2959 		vm_page_unlock(m);
2960 	}
2961 }
2962 
2963 void
2964 vm_page_cowclear(vm_page_t m)
2965 {
2966 
2967 	vm_page_lock_assert(m, MA_OWNED);
2968 	if (m->cow) {
2969 		m->cow--;
2970 		/*
2971 		 * let vm_fault add back write permission  lazily
2972 		 */
2973 	}
2974 	/*
2975 	 *  sf_buf_free() will free the page, so we needn't do it here
2976 	 */
2977 }
2978 
2979 int
2980 vm_page_cowsetup(vm_page_t m)
2981 {
2982 
2983 	vm_page_lock_assert(m, MA_OWNED);
2984 	if ((m->flags & PG_FICTITIOUS) != 0 ||
2985 	    (m->oflags & VPO_UNMANAGED) != 0 ||
2986 	    m->cow == USHRT_MAX - 1 || !VM_OBJECT_TRYLOCK(m->object))
2987 		return (EBUSY);
2988 	m->cow++;
2989 	pmap_remove_write(m);
2990 	VM_OBJECT_UNLOCK(m->object);
2991 	return (0);
2992 }
2993 
2994 #ifdef INVARIANTS
2995 void
2996 vm_page_object_lock_assert(vm_page_t m)
2997 {
2998 
2999 	/*
3000 	 * Certain of the page's fields may only be modified by the
3001 	 * holder of the containing object's lock or the setter of the
3002 	 * page's VPO_BUSY flag.  Unfortunately, the setter of the
3003 	 * VPO_BUSY flag is not recorded, and thus cannot be checked
3004 	 * here.
3005 	 */
3006 	if (m->object != NULL && (m->oflags & VPO_BUSY) == 0)
3007 		VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
3008 }
3009 #endif
3010 
3011 #include "opt_ddb.h"
3012 #ifdef DDB
3013 #include <sys/kernel.h>
3014 
3015 #include <ddb/ddb.h>
3016 
3017 DB_SHOW_COMMAND(page, vm_page_print_page_info)
3018 {
3019 	db_printf("cnt.v_free_count: %d\n", cnt.v_free_count);
3020 	db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count);
3021 	db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count);
3022 	db_printf("cnt.v_active_count: %d\n", cnt.v_active_count);
3023 	db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count);
3024 	db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved);
3025 	db_printf("cnt.v_free_min: %d\n", cnt.v_free_min);
3026 	db_printf("cnt.v_free_target: %d\n", cnt.v_free_target);
3027 	db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min);
3028 	db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target);
3029 }
3030 
3031 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
3032 {
3033 
3034 	db_printf("PQ_FREE:");
3035 	db_printf(" %d", cnt.v_free_count);
3036 	db_printf("\n");
3037 
3038 	db_printf("PQ_CACHE:");
3039 	db_printf(" %d", cnt.v_cache_count);
3040 	db_printf("\n");
3041 
3042 	db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n",
3043 		*vm_page_queues[PQ_ACTIVE].cnt,
3044 		*vm_page_queues[PQ_INACTIVE].cnt);
3045 }
3046 #endif /* DDB */
3047