xref: /dflybsd-src/sys/vm/vm_page.c (revision 4d9022e3888d071c1d026d8aa01aecd099e7bd9b)
1 /*
2  * (MPSAFE)
3  *
4  * Copyright (c) 1991 Regents of the University of California.
5  * All rights reserved.
6  *
7  * This code is derived from software contributed to Berkeley by
8  * The Mach Operating System project at Carnegie-Mellon University.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  * 3. All advertising materials mentioning features or use of this software
19  *    must display the following acknowledgement:
20  *	This product includes software developed by the University of
21  *	California, Berkeley and its contributors.
22  * 4. Neither the name of the University nor the names of its contributors
23  *    may be used to endorse or promote products derived from this software
24  *    without specific prior written permission.
25  *
26  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
27  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
28  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
29  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
30  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
31  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
32  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
33  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
34  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
35  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36  * SUCH DAMAGE.
37  *
38  *	from: @(#)vm_page.c	7.4 (Berkeley) 5/7/91
39  * $FreeBSD: src/sys/vm/vm_page.c,v 1.147.2.18 2002/03/10 05:03:19 alc Exp $
40  * $DragonFly: src/sys/vm/vm_page.c,v 1.40 2008/08/25 17:01:42 dillon Exp $
41  */
42 
43 /*
44  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
45  * All rights reserved.
46  *
47  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
48  *
49  * Permission to use, copy, modify and distribute this software and
50  * its documentation is hereby granted, provided that both the copyright
51  * notice and this permission notice appear in all copies of the
52  * software, derivative works or modified versions, and any portions
53  * thereof, and that both notices appear in supporting documentation.
54  *
55  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
56  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
57  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
58  *
59  * Carnegie Mellon requests users of this software to return to
60  *
61  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
62  *  School of Computer Science
63  *  Carnegie Mellon University
64  *  Pittsburgh PA 15213-3890
65  *
66  * any improvements or extensions that they make and grant Carnegie the
67  * rights to redistribute these changes.
68  */
69 /*
70  * Resident memory management module.  The module manipulates 'VM pages'.
71  * A VM page is the core building block for memory management.
72  */
73 
74 #include <sys/param.h>
75 #include <sys/systm.h>
76 #include <sys/malloc.h>
77 #include <sys/proc.h>
78 #include <sys/vmmeter.h>
79 #include <sys/vnode.h>
80 
81 #include <vm/vm.h>
82 #include <vm/vm_param.h>
83 #include <sys/lock.h>
84 #include <vm/vm_kern.h>
85 #include <vm/pmap.h>
86 #include <vm/vm_map.h>
87 #include <vm/vm_object.h>
88 #include <vm/vm_page.h>
89 #include <vm/vm_pageout.h>
90 #include <vm/vm_pager.h>
91 #include <vm/vm_extern.h>
92 #include <vm/swap_pager.h>
93 
94 #include <machine/md_var.h>
95 
96 #include <vm/vm_page2.h>
97 #include <sys/mplock2.h>
98 
99 static void vm_page_queue_init(void);
100 static void vm_page_free_wakeup(void);
101 static vm_page_t vm_page_select_cache(vm_object_t, vm_pindex_t);
102 static vm_page_t _vm_page_list_find2(int basequeue, int index);
103 
104 struct vpgqueues vm_page_queues[PQ_COUNT]; /* Array of tailq lists */
105 
106 #define ASSERT_IN_CRIT_SECTION()	KKASSERT(crit_test(curthread));
107 
108 RB_GENERATE2(vm_page_rb_tree, vm_page, rb_entry, rb_vm_page_compare,
109 	     vm_pindex_t, pindex);
110 
111 static void
112 vm_page_queue_init(void)
113 {
114 	int i;
115 
116 	for (i = 0; i < PQ_L2_SIZE; i++)
117 		vm_page_queues[PQ_FREE+i].cnt = &vmstats.v_free_count;
118 	for (i = 0; i < PQ_L2_SIZE; i++)
119 		vm_page_queues[PQ_CACHE+i].cnt = &vmstats.v_cache_count;
120 
121 	vm_page_queues[PQ_INACTIVE].cnt = &vmstats.v_inactive_count;
122 	vm_page_queues[PQ_ACTIVE].cnt = &vmstats.v_active_count;
123 	vm_page_queues[PQ_HOLD].cnt = &vmstats.v_active_count;
124 	/* PQ_NONE has no queue */
125 
126 	for (i = 0; i < PQ_COUNT; i++)
127 		TAILQ_INIT(&vm_page_queues[i].pl);
128 }
129 
130 /*
131  * note: place in initialized data section?  Is this necessary?
132  */
133 long first_page = 0;
134 int vm_page_array_size = 0;
135 int vm_page_zero_count = 0;
136 vm_page_t vm_page_array = 0;
137 
138 /*
139  * (low level boot)
140  *
141  * Sets the page size, perhaps based upon the memory size.
142  * Must be called before any use of page-size dependent functions.
143  */
144 void
145 vm_set_page_size(void)
146 {
147 	if (vmstats.v_page_size == 0)
148 		vmstats.v_page_size = PAGE_SIZE;
149 	if (((vmstats.v_page_size - 1) & vmstats.v_page_size) != 0)
150 		panic("vm_set_page_size: page size not a power of two");
151 }
152 
153 /*
154  * (low level boot)
155  *
156  * Add a new page to the freelist for use by the system.  New pages
157  * are added to both the head and tail of the associated free page
158  * queue in a bottom-up fashion, so both zero'd and non-zero'd page
159  * requests pull 'recent' adds (higher physical addresses) first.
160  *
161  * Must be called in a critical section.
162  */
163 vm_page_t
164 vm_add_new_page(vm_paddr_t pa)
165 {
166 	struct vpgqueues *vpq;
167 	vm_page_t m;
168 
169 	++vmstats.v_page_count;
170 	++vmstats.v_free_count;
171 	m = PHYS_TO_VM_PAGE(pa);
172 	m->phys_addr = pa;
173 	m->flags = 0;
174 	m->pc = (pa >> PAGE_SHIFT) & PQ_L2_MASK;
175 	m->queue = m->pc + PQ_FREE;
176 	KKASSERT(m->dirty == 0);
177 
178 	vpq = &vm_page_queues[m->queue];
179 	if (vpq->flipflop)
180 		TAILQ_INSERT_TAIL(&vpq->pl, m, pageq);
181 	else
182 		TAILQ_INSERT_HEAD(&vpq->pl, m, pageq);
183 	vpq->flipflop = 1 - vpq->flipflop;
184 
185 	vm_page_queues[m->queue].lcnt++;
186 	return (m);
187 }
188 
189 /*
190  * (low level boot)
191  *
192  * Initializes the resident memory module.
193  *
194  * Allocates memory for the page cells, and for the object/offset-to-page
195  * hash table headers.  Each page cell is initialized and placed on the
196  * free list.
197  *
198  * starta/enda represents the range of physical memory addresses available
199  * for use (skipping memory already used by the kernel), subject to
200  * phys_avail[].  Note that phys_avail[] has already mapped out memory
201  * already in use by the kernel.
202  */
203 vm_offset_t
204 vm_page_startup(vm_offset_t vaddr)
205 {
206 	vm_offset_t mapped;
207 	vm_size_t npages;
208 	vm_paddr_t page_range;
209 	vm_paddr_t new_end;
210 	int i;
211 	vm_paddr_t pa;
212 	int nblocks;
213 	vm_paddr_t last_pa;
214 	vm_paddr_t end;
215 	vm_paddr_t biggestone, biggestsize;
216 	vm_paddr_t total;
217 
218 	total = 0;
219 	biggestsize = 0;
220 	biggestone = 0;
221 	nblocks = 0;
222 	vaddr = round_page(vaddr);
223 
224 	for (i = 0; phys_avail[i + 1]; i += 2) {
225 		phys_avail[i] = round_page64(phys_avail[i]);
226 		phys_avail[i + 1] = trunc_page64(phys_avail[i + 1]);
227 	}
228 
229 	for (i = 0; phys_avail[i + 1]; i += 2) {
230 		vm_paddr_t size = phys_avail[i + 1] - phys_avail[i];
231 
232 		if (size > biggestsize) {
233 			biggestone = i;
234 			biggestsize = size;
235 		}
236 		++nblocks;
237 		total += size;
238 	}
239 
240 	end = phys_avail[biggestone+1];
241 	end = trunc_page(end);
242 
243 	/*
244 	 * Initialize the queue headers for the free queue, the active queue
245 	 * and the inactive queue.
246 	 */
247 
248 	vm_page_queue_init();
249 
250 	/* VKERNELs don't support minidumps and as such don't need vm_page_dump */
251 #if !defined(_KERNEL_VIRTUAL)
252 	/*
253 	 * Allocate a bitmap to indicate that a random physical page
254 	 * needs to be included in a minidump.
255 	 *
256 	 * The amd64 port needs this to indicate which direct map pages
257 	 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
258 	 *
259 	 * However, i386 still needs this workspace internally within the
260 	 * minidump code.  In theory, they are not needed on i386, but are
261 	 * included should the sf_buf code decide to use them.
262 	 */
263 	page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE;
264 	vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
265 	end -= vm_page_dump_size;
266 	vm_page_dump = (void *)pmap_map(&vaddr, end, end + vm_page_dump_size,
267 	    VM_PROT_READ | VM_PROT_WRITE);
268 	bzero((void *)vm_page_dump, vm_page_dump_size);
269 #endif
270 
271 	/*
272 	 * Compute the number of pages of memory that will be available for
273 	 * use (taking into account the overhead of a page structure per
274 	 * page).
275 	 */
276 	first_page = phys_avail[0] / PAGE_SIZE;
277 	page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page;
278 	npages = (total - (page_range * sizeof(struct vm_page))) / PAGE_SIZE;
279 
280 	/*
281 	 * Initialize the mem entry structures now, and put them in the free
282 	 * queue.
283 	 */
284 	new_end = trunc_page(end - page_range * sizeof(struct vm_page));
285 	mapped = pmap_map(&vaddr, new_end, end,
286 	    VM_PROT_READ | VM_PROT_WRITE);
287 	vm_page_array = (vm_page_t)mapped;
288 
289 #if defined(__x86_64__) && !defined(_KERNEL_VIRTUAL)
290 	/*
291 	 * since pmap_map on amd64 returns stuff out of a direct-map region,
292 	 * we have to manually add these pages to the minidump tracking so
293 	 * that they can be dumped, including the vm_page_array.
294 	 */
295 	for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE)
296 		dump_add_page(pa);
297 #endif
298 
299 	/*
300 	 * Clear all of the page structures
301 	 */
302 	bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
303 	vm_page_array_size = page_range;
304 
305 	/*
306 	 * Construct the free queue(s) in ascending order (by physical
307 	 * address) so that the first 16MB of physical memory is allocated
308 	 * last rather than first.  On large-memory machines, this avoids
309 	 * the exhaustion of low physical memory before isa_dmainit has run.
310 	 */
311 	vmstats.v_page_count = 0;
312 	vmstats.v_free_count = 0;
313 	for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) {
314 		pa = phys_avail[i];
315 		if (i == biggestone)
316 			last_pa = new_end;
317 		else
318 			last_pa = phys_avail[i + 1];
319 		while (pa < last_pa && npages-- > 0) {
320 			vm_add_new_page(pa);
321 			pa += PAGE_SIZE;
322 		}
323 	}
324 	return (vaddr);
325 }
326 
327 /*
328  * Scan comparison function for Red-Black tree scans.  An inclusive
329  * (start,end) is expected.  Other fields are not used.
330  */
331 int
332 rb_vm_page_scancmp(struct vm_page *p, void *data)
333 {
334 	struct rb_vm_page_scan_info *info = data;
335 
336 	if (p->pindex < info->start_pindex)
337 		return(-1);
338 	if (p->pindex > info->end_pindex)
339 		return(1);
340 	return(0);
341 }
342 
343 int
344 rb_vm_page_compare(struct vm_page *p1, struct vm_page *p2)
345 {
346 	if (p1->pindex < p2->pindex)
347 		return(-1);
348 	if (p1->pindex > p2->pindex)
349 		return(1);
350 	return(0);
351 }
352 
353 /*
354  * The opposite of vm_page_hold().  A page can be freed while being held,
355  * which places it on the PQ_HOLD queue.  We must call vm_page_free_toq()
356  * in this case to actually free it once the hold count drops to 0.
357  *
358  * This routine must be called at splvm().
359  */
360 void
361 vm_page_unhold(vm_page_t mem)
362 {
363 	--mem->hold_count;
364 	KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!"));
365 	if (mem->hold_count == 0 && mem->queue == PQ_HOLD) {
366 		vm_page_busy(mem);
367 		vm_page_free_toq(mem);
368 	}
369 }
370 
371 /*
372  * Inserts the given mem entry into the object and object list.
373  *
374  * The pagetables are not updated but will presumably fault the page
375  * in if necessary, or if a kernel page the caller will at some point
376  * enter the page into the kernel's pmap.  We are not allowed to block
377  * here so we *can't* do this anyway.
378  *
379  * This routine may not block.
380  * This routine must be called with a critical section held.
381  */
382 void
383 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
384 {
385 	ASSERT_IN_CRIT_SECTION();
386 	if (m->object != NULL)
387 		panic("vm_page_insert: already inserted");
388 
389 	/*
390 	 * Record the object/offset pair in this page
391 	 */
392 	m->object = object;
393 	m->pindex = pindex;
394 
395 	/*
396 	 * Insert it into the object.
397 	 */
398 	ASSERT_MP_LOCK_HELD(curthread);
399 	vm_page_rb_tree_RB_INSERT(&object->rb_memq, m);
400 	object->generation++;
401 
402 	/*
403 	 * show that the object has one more resident page.
404 	 */
405 	object->resident_page_count++;
406 
407 	/*
408 	 * Since we are inserting a new and possibly dirty page,
409 	 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags.
410 	 */
411 	if ((m->valid & m->dirty) || (m->flags & PG_WRITEABLE))
412 		vm_object_set_writeable_dirty(object);
413 
414 	/*
415 	 * Checks for a swap assignment and sets PG_SWAPPED if appropriate.
416 	 */
417 	swap_pager_page_inserted(m);
418 }
419 
420 /*
421  * Removes the given vm_page_t from the global (object,index) hash table
422  * and from the object's memq.
423  *
424  * The underlying pmap entry (if any) is NOT removed here.
425  * This routine may not block.
426  *
427  * The page must be BUSY and will remain BUSY on return.
428  * No other requirements.
429  *
430  * NOTE: FreeBSD side effect was to unbusy the page on return.  We leave
431  *	 it busy.
432  */
433 void
434 vm_page_remove(vm_page_t m)
435 {
436 	vm_object_t object;
437 
438 	crit_enter();
439 	lwkt_gettoken(&vm_token);
440 	if (m->object == NULL) {
441 		lwkt_reltoken(&vm_token);
442 		crit_exit();
443 		return;
444 	}
445 
446 	if ((m->flags & PG_BUSY) == 0)
447 		panic("vm_page_remove: page not busy");
448 
449 	object = m->object;
450 
451 	/*
452 	 * Remove the page from the object and update the object.
453 	 */
454 	ASSERT_MP_LOCK_HELD(curthread);
455 	vm_page_rb_tree_RB_REMOVE(&object->rb_memq, m);
456 	object->resident_page_count--;
457 	object->generation++;
458 	m->object = NULL;
459 
460 	lwkt_reltoken(&vm_token);
461 	crit_exit();
462 }
463 
464 /*
465  * Locate and return the page at (object, pindex), or NULL if the
466  * page could not be found.
467  *
468  * This routine will operate properly without spl protection, but
469  * the returned page could be in flux if it is busy.  Because an
470  * interrupt can race a caller's busy check (unbusying and freeing the
471  * page we return before the caller is able to check the busy bit),
472  * the caller should generally call this routine with a critical
473  * section held.
474  *
475  * Callers may call this routine without spl protection if they know
476  * 'for sure' that the page will not be ripped out from under them
477  * by an interrupt.
478  */
479 vm_page_t
480 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
481 {
482 	vm_page_t m;
483 
484 	/*
485 	 * Search the hash table for this object/offset pair
486 	 */
487 	ASSERT_MP_LOCK_HELD(curthread);
488 	crit_enter();
489 	lwkt_gettoken(&vm_token);
490 	m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
491 	lwkt_reltoken(&vm_token);
492 	crit_exit();
493 	KKASSERT(m == NULL || (m->object == object && m->pindex == pindex));
494 	return(m);
495 }
496 
497 /*
498  * vm_page_rename()
499  *
500  * Move the given memory entry from its current object to the specified
501  * target object/offset.
502  *
503  * The object must be locked.
504  * This routine may not block.
505  *
506  * Note: This routine will raise itself to splvm(), the caller need not.
507  *
508  * Note: Swap associated with the page must be invalidated by the move.  We
509  *       have to do this for several reasons:  (1) we aren't freeing the
510  *       page, (2) we are dirtying the page, (3) the VM system is probably
511  *       moving the page from object A to B, and will then later move
512  *       the backing store from A to B and we can't have a conflict.
513  *
514  * Note: We *always* dirty the page.  It is necessary both for the
515  *       fact that we moved it, and because we may be invalidating
516  *	 swap.  If the page is on the cache, we have to deactivate it
517  *	 or vm_page_dirty() will panic.  Dirty pages are not allowed
518  *	 on the cache.
519  */
520 void
521 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
522 {
523 	crit_enter();
524 	lwkt_gettoken(&vm_token);
525 	vm_page_remove(m);
526 	vm_page_insert(m, new_object, new_pindex);
527 	if (m->queue - m->pc == PQ_CACHE)
528 		vm_page_deactivate(m);
529 	vm_page_dirty(m);
530 	vm_page_wakeup(m);
531 	lwkt_reltoken(&vm_token);
532 	crit_exit();
533 }
534 
535 /*
536  * vm_page_unqueue() without any wakeup.  This routine is used when a page
537  * is being moved between queues or otherwise is to remain BUSYied by the
538  * caller.
539  *
540  * This routine must be called at splhigh().
541  * This routine may not block.
542  */
543 void
544 vm_page_unqueue_nowakeup(vm_page_t m)
545 {
546 	int queue = m->queue;
547 	struct vpgqueues *pq;
548 
549 	if (queue != PQ_NONE) {
550 		pq = &vm_page_queues[queue];
551 		m->queue = PQ_NONE;
552 		TAILQ_REMOVE(&pq->pl, m, pageq);
553 		(*pq->cnt)--;
554 		pq->lcnt--;
555 	}
556 }
557 
558 /*
559  * vm_page_unqueue() - Remove a page from its queue, wakeup the pagedemon
560  * if necessary.
561  *
562  * This routine must be called at splhigh().
563  * This routine may not block.
564  */
565 void
566 vm_page_unqueue(vm_page_t m)
567 {
568 	int queue = m->queue;
569 	struct vpgqueues *pq;
570 
571 	if (queue != PQ_NONE) {
572 		m->queue = PQ_NONE;
573 		pq = &vm_page_queues[queue];
574 		TAILQ_REMOVE(&pq->pl, m, pageq);
575 		(*pq->cnt)--;
576 		pq->lcnt--;
577 		if ((queue - m->pc) == PQ_CACHE || (queue - m->pc) == PQ_FREE)
578 			pagedaemon_wakeup();
579 	}
580 }
581 
582 /*
583  * vm_page_list_find()
584  *
585  * Find a page on the specified queue with color optimization.
586  *
587  * The page coloring optimization attempts to locate a page that does
588  * not overload other nearby pages in the object in the cpu's L1 or L2
589  * caches.  We need this optimization because cpu caches tend to be
590  * physical caches, while object spaces tend to be virtual.
591  *
592  * This routine must be called at splvm().
593  * This routine may not block.
594  *
595  * Note that this routine is carefully inlined.  A non-inlined version
596  * is available for outside callers but the only critical path is
597  * from within this source file.
598  */
599 static __inline
600 vm_page_t
601 _vm_page_list_find(int basequeue, int index, boolean_t prefer_zero)
602 {
603 	vm_page_t m;
604 
605 	if (prefer_zero)
606 		m = TAILQ_LAST(&vm_page_queues[basequeue+index].pl, pglist);
607 	else
608 		m = TAILQ_FIRST(&vm_page_queues[basequeue+index].pl);
609 	if (m == NULL)
610 		m = _vm_page_list_find2(basequeue, index);
611 	return(m);
612 }
613 
614 static vm_page_t
615 _vm_page_list_find2(int basequeue, int index)
616 {
617 	int i;
618 	vm_page_t m = NULL;
619 	struct vpgqueues *pq;
620 
621 	pq = &vm_page_queues[basequeue];
622 
623 	/*
624 	 * Note that for the first loop, index+i and index-i wind up at the
625 	 * same place.  Even though this is not totally optimal, we've already
626 	 * blown it by missing the cache case so we do not care.
627 	 */
628 
629 	for(i = PQ_L2_SIZE / 2; i > 0; --i) {
630 		if ((m = TAILQ_FIRST(&pq[(index + i) & PQ_L2_MASK].pl)) != NULL)
631 			break;
632 
633 		if ((m = TAILQ_FIRST(&pq[(index - i) & PQ_L2_MASK].pl)) != NULL)
634 			break;
635 	}
636 	return(m);
637 }
638 
639 vm_page_t
640 vm_page_list_find(int basequeue, int index, boolean_t prefer_zero)
641 {
642 	return(_vm_page_list_find(basequeue, index, prefer_zero));
643 }
644 
645 /*
646  * Find a page on the cache queue with color optimization.  As pages
647  * might be found, but not applicable, they are deactivated.  This
648  * keeps us from using potentially busy cached pages.
649  *
650  * This routine must be called with a critical section held.
651  * This routine may not block.
652  */
653 vm_page_t
654 vm_page_select_cache(vm_object_t object, vm_pindex_t pindex)
655 {
656 	vm_page_t m;
657 
658 	while (TRUE) {
659 		m = _vm_page_list_find(
660 		    PQ_CACHE,
661 		    (pindex + object->pg_color) & PQ_L2_MASK,
662 		    FALSE
663 		);
664 		if (m && ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy ||
665 			       m->hold_count || m->wire_count)) {
666 			vm_page_deactivate(m);
667 			continue;
668 		}
669 		return m;
670 	}
671 	/* not reached */
672 }
673 
674 /*
675  * Find a free or zero page, with specified preference.  We attempt to
676  * inline the nominal case and fall back to _vm_page_select_free()
677  * otherwise.
678  *
679  * This routine must be called with a critical section held.
680  * This routine may not block.
681  */
682 static __inline vm_page_t
683 vm_page_select_free(vm_object_t object, vm_pindex_t pindex, boolean_t prefer_zero)
684 {
685 	vm_page_t m;
686 
687 	m = _vm_page_list_find(
688 		PQ_FREE,
689 		(pindex + object->pg_color) & PQ_L2_MASK,
690 		prefer_zero
691 	);
692 	return(m);
693 }
694 
695 /*
696  * vm_page_alloc()
697  *
698  * Allocate and return a memory cell associated with this VM object/offset
699  * pair.
700  *
701  *	page_req classes:
702  *
703  *	VM_ALLOC_NORMAL		allow use of cache pages, nominal free drain
704  *	VM_ALLOC_QUICK		like normal but cannot use cache
705  *	VM_ALLOC_SYSTEM		greater free drain
706  *	VM_ALLOC_INTERRUPT	allow free list to be completely drained
707  *	VM_ALLOC_ZERO		advisory request for pre-zero'd page
708  *
709  * The object must be locked.
710  * This routine may not block.
711  * The returned page will be marked PG_BUSY
712  *
713  * Additional special handling is required when called from an interrupt
714  * (VM_ALLOC_INTERRUPT).  We are not allowed to mess with the page cache
715  * in this case.
716  */
717 vm_page_t
718 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req)
719 {
720 	vm_page_t m = NULL;
721 
722 	KKASSERT(object != NULL);
723 	KASSERT(!vm_page_lookup(object, pindex),
724 		("vm_page_alloc: page already allocated"));
725 	KKASSERT(page_req &
726 		(VM_ALLOC_NORMAL|VM_ALLOC_QUICK|
727 		 VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
728 
729 	/*
730 	 * Certain system threads (pageout daemon, buf_daemon's) are
731 	 * allowed to eat deeper into the free page list.
732 	 */
733 	if (curthread->td_flags & TDF_SYSTHREAD)
734 		page_req |= VM_ALLOC_SYSTEM;
735 
736 	crit_enter();
737 	lwkt_gettoken(&vm_token);
738 loop:
739 	if (vmstats.v_free_count > vmstats.v_free_reserved ||
740 	    ((page_req & VM_ALLOC_INTERRUPT) && vmstats.v_free_count > 0) ||
741 	    ((page_req & VM_ALLOC_SYSTEM) && vmstats.v_cache_count == 0 &&
742 		vmstats.v_free_count > vmstats.v_interrupt_free_min)
743 	) {
744 		/*
745 		 * The free queue has sufficient free pages to take one out.
746 		 */
747 		if (page_req & VM_ALLOC_ZERO)
748 			m = vm_page_select_free(object, pindex, TRUE);
749 		else
750 			m = vm_page_select_free(object, pindex, FALSE);
751 	} else if (page_req & VM_ALLOC_NORMAL) {
752 		/*
753 		 * Allocatable from the cache (non-interrupt only).  On
754 		 * success, we must free the page and try again, thus
755 		 * ensuring that vmstats.v_*_free_min counters are replenished.
756 		 */
757 #ifdef INVARIANTS
758 		if (curthread->td_preempted) {
759 			kprintf("vm_page_alloc(): warning, attempt to allocate"
760 				" cache page from preempting interrupt\n");
761 			m = NULL;
762 		} else {
763 			m = vm_page_select_cache(object, pindex);
764 		}
765 #else
766 		m = vm_page_select_cache(object, pindex);
767 #endif
768 		/*
769 		 * On success move the page into the free queue and loop.
770 		 */
771 		if (m != NULL) {
772 			KASSERT(m->dirty == 0,
773 			    ("Found dirty cache page %p", m));
774 			vm_page_busy(m);
775 			vm_page_protect(m, VM_PROT_NONE);
776 			vm_page_free(m);
777 			goto loop;
778 		}
779 
780 		/*
781 		 * On failure return NULL
782 		 */
783 		lwkt_reltoken(&vm_token);
784 		crit_exit();
785 #if defined(DIAGNOSTIC)
786 		if (vmstats.v_cache_count > 0)
787 			kprintf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", vmstats.v_cache_count);
788 #endif
789 		vm_pageout_deficit++;
790 		pagedaemon_wakeup();
791 		return (NULL);
792 	} else {
793 		/*
794 		 * No pages available, wakeup the pageout daemon and give up.
795 		 */
796 		lwkt_reltoken(&vm_token);
797 		crit_exit();
798 		vm_pageout_deficit++;
799 		pagedaemon_wakeup();
800 		return (NULL);
801 	}
802 
803 	/*
804 	 * Good page found.  The page has not yet been busied.  We are in
805 	 * a critical section.
806 	 */
807 	KASSERT(m != NULL, ("vm_page_alloc(): missing page on free queue\n"));
808 	KASSERT(m->dirty == 0,
809 		("vm_page_alloc: free/cache page %p was dirty", m));
810 
811 	/*
812 	 * Remove from free queue
813 	 */
814 	vm_page_unqueue_nowakeup(m);
815 
816 	/*
817 	 * Initialize structure.  Only the PG_ZERO flag is inherited.  Set
818 	 * the page PG_BUSY
819 	 */
820 	if (m->flags & PG_ZERO) {
821 		vm_page_zero_count--;
822 		m->flags = PG_ZERO | PG_BUSY;
823 	} else {
824 		m->flags = PG_BUSY;
825 	}
826 	m->wire_count = 0;
827 	m->hold_count = 0;
828 	m->act_count = 0;
829 	m->busy = 0;
830 	m->valid = 0;
831 
832 	/*
833 	 * vm_page_insert() is safe prior to the crit_exit().  Note also that
834 	 * inserting a page here does not insert it into the pmap (which
835 	 * could cause us to block allocating memory).  We cannot block
836 	 * anywhere.
837 	 */
838 	vm_page_insert(m, object, pindex);
839 
840 	/*
841 	 * Don't wakeup too often - wakeup the pageout daemon when
842 	 * we would be nearly out of memory.
843 	 */
844 	pagedaemon_wakeup();
845 
846 	lwkt_reltoken(&vm_token);
847 	crit_exit();
848 
849 	/*
850 	 * A PG_BUSY page is returned.
851 	 */
852 	return (m);
853 }
854 
855 /*
856  * Wait for sufficient free memory for nominal heavy memory use kernel
857  * operations.
858  */
859 void
860 vm_wait_nominal(void)
861 {
862 	while (vm_page_count_min(0))
863 		vm_wait(0);
864 }
865 
866 /*
867  * Test if vm_wait_nominal() would block.
868  */
869 int
870 vm_test_nominal(void)
871 {
872 	if (vm_page_count_min(0))
873 		return(1);
874 	return(0);
875 }
876 
877 /*
878  * Block until free pages are available for allocation, called in various
879  * places before memory allocations.
880  */
881 void
882 vm_wait(int timo)
883 {
884 	crit_enter();
885 	lwkt_gettoken(&vm_token);
886 	if (curthread == pagethread) {
887 		vm_pageout_pages_needed = 1;
888 		tsleep(&vm_pageout_pages_needed, 0, "VMWait", timo);
889 	} else {
890 		if (vm_pages_needed == 0) {
891 			vm_pages_needed = 1;
892 			wakeup(&vm_pages_needed);
893 		}
894 		tsleep(&vmstats.v_free_count, 0, "vmwait", timo);
895 	}
896 	lwkt_reltoken(&vm_token);
897 	crit_exit();
898 }
899 
900 /*
901  * Block until free pages are available for allocation
902  *
903  * Called only in vm_fault so that processes page faulting can be
904  * easily tracked.
905  */
906 void
907 vm_waitpfault(void)
908 {
909 	crit_enter();
910 	lwkt_gettoken(&vm_token);
911 	if (vm_pages_needed == 0) {
912 		vm_pages_needed = 1;
913 		wakeup(&vm_pages_needed);
914 	}
915 	tsleep(&vmstats.v_free_count, 0, "pfault", 0);
916 	lwkt_reltoken(&vm_token);
917 	crit_exit();
918 }
919 
920 /*
921  * Put the specified page on the active list (if appropriate).  Ensure
922  * that act_count is at least ACT_INIT but do not otherwise mess with it.
923  *
924  * The page queues must be locked.
925  * This routine may not block.
926  */
927 void
928 vm_page_activate(vm_page_t m)
929 {
930 	crit_enter();
931 	lwkt_gettoken(&vm_token);
932 	if (m->queue != PQ_ACTIVE) {
933 		if ((m->queue - m->pc) == PQ_CACHE)
934 			mycpu->gd_cnt.v_reactivated++;
935 
936 		vm_page_unqueue(m);
937 
938 		if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
939 			m->queue = PQ_ACTIVE;
940 			vm_page_queues[PQ_ACTIVE].lcnt++;
941 			TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl,
942 					    m, pageq);
943 			if (m->act_count < ACT_INIT)
944 				m->act_count = ACT_INIT;
945 			vmstats.v_active_count++;
946 		}
947 	} else {
948 		if (m->act_count < ACT_INIT)
949 			m->act_count = ACT_INIT;
950 	}
951 	lwkt_reltoken(&vm_token);
952 	crit_exit();
953 }
954 
955 /*
956  * Helper routine for vm_page_free_toq() and vm_page_cache().  This
957  * routine is called when a page has been added to the cache or free
958  * queues.
959  *
960  * This routine may not block.
961  * This routine must be called at splvm()
962  */
963 static __inline void
964 vm_page_free_wakeup(void)
965 {
966 	/*
967 	 * if pageout daemon needs pages, then tell it that there are
968 	 * some free.
969 	 */
970 	if (vm_pageout_pages_needed &&
971 	    vmstats.v_cache_count + vmstats.v_free_count >=
972 	    vmstats.v_pageout_free_min
973 	) {
974 		wakeup(&vm_pageout_pages_needed);
975 		vm_pageout_pages_needed = 0;
976 	}
977 
978 	/*
979 	 * wakeup processes that are waiting on memory if we hit a
980 	 * high water mark. And wakeup scheduler process if we have
981 	 * lots of memory. this process will swapin processes.
982 	 */
983 	if (vm_pages_needed && !vm_page_count_min(0)) {
984 		vm_pages_needed = 0;
985 		wakeup(&vmstats.v_free_count);
986 	}
987 }
988 
989 /*
990  *	vm_page_free_toq:
991  *
992  *	Returns the given page to the PQ_FREE list, disassociating it with
993  *	any VM object.
994  *
995  *	The vm_page must be PG_BUSY on entry.  PG_BUSY will be released on
996  *	return (the page will have been freed).  No particular spl is required
997  *	on entry.
998  *
999  *	This routine may not block.
1000  */
1001 void
1002 vm_page_free_toq(vm_page_t m)
1003 {
1004 	struct vpgqueues *pq;
1005 
1006 	crit_enter();
1007 	lwkt_gettoken(&vm_token);
1008 	mycpu->gd_cnt.v_tfree++;
1009 
1010 	KKASSERT((m->flags & PG_MAPPED) == 0);
1011 
1012 	if (m->busy || ((m->queue - m->pc) == PQ_FREE)) {
1013 		kprintf(
1014 		"vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n",
1015 		    (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0,
1016 		    m->hold_count);
1017 		if ((m->queue - m->pc) == PQ_FREE)
1018 			panic("vm_page_free: freeing free page");
1019 		else
1020 			panic("vm_page_free: freeing busy page");
1021 	}
1022 
1023 	/*
1024 	 * unqueue, then remove page.  Note that we cannot destroy
1025 	 * the page here because we do not want to call the pager's
1026 	 * callback routine until after we've put the page on the
1027 	 * appropriate free queue.
1028 	 */
1029 	vm_page_unqueue_nowakeup(m);
1030 	vm_page_remove(m);
1031 
1032 	/*
1033 	 * No further management of fictitious pages occurs beyond object
1034 	 * and queue removal.
1035 	 */
1036 	if ((m->flags & PG_FICTITIOUS) != 0) {
1037 		vm_page_wakeup(m);
1038 		lwkt_reltoken(&vm_token);
1039 		crit_exit();
1040 		return;
1041 	}
1042 
1043 	m->valid = 0;
1044 	vm_page_undirty(m);
1045 
1046 	if (m->wire_count != 0) {
1047 		if (m->wire_count > 1) {
1048 		    panic(
1049 			"vm_page_free: invalid wire count (%d), pindex: 0x%lx",
1050 			m->wire_count, (long)m->pindex);
1051 		}
1052 		panic("vm_page_free: freeing wired page");
1053 	}
1054 
1055 	/*
1056 	 * Clear the UNMANAGED flag when freeing an unmanaged page.
1057 	 */
1058 	if (m->flags & PG_UNMANAGED) {
1059 	    m->flags &= ~PG_UNMANAGED;
1060 	}
1061 
1062 	if (m->hold_count != 0) {
1063 		m->flags &= ~PG_ZERO;
1064 		m->queue = PQ_HOLD;
1065 	} else {
1066 		m->queue = PQ_FREE + m->pc;
1067 	}
1068 	pq = &vm_page_queues[m->queue];
1069 	pq->lcnt++;
1070 	++(*pq->cnt);
1071 
1072 	/*
1073 	 * Put zero'd pages on the end ( where we look for zero'd pages
1074 	 * first ) and non-zerod pages at the head.
1075 	 */
1076 	if (m->flags & PG_ZERO) {
1077 		TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
1078 		++vm_page_zero_count;
1079 	} else {
1080 		TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
1081 	}
1082 	vm_page_wakeup(m);
1083 	vm_page_free_wakeup();
1084 	lwkt_reltoken(&vm_token);
1085 	crit_exit();
1086 }
1087 
1088 /*
1089  * vm_page_free_fromq_fast()
1090  *
1091  * Remove a non-zero page from one of the free queues; the page is removed for
1092  * zeroing, so do not issue a wakeup.
1093  *
1094  * MPUNSAFE
1095  */
1096 vm_page_t
1097 vm_page_free_fromq_fast(void)
1098 {
1099 	static int qi;
1100 	vm_page_t m;
1101 	int i;
1102 
1103 	crit_enter();
1104 	lwkt_gettoken(&vm_token);
1105 	for (i = 0; i < PQ_L2_SIZE; ++i) {
1106 		m = vm_page_list_find(PQ_FREE, qi, FALSE);
1107 		qi = (qi + PQ_PRIME2) & PQ_L2_MASK;
1108 		if (m && (m->flags & PG_ZERO) == 0) {
1109 			vm_page_unqueue_nowakeup(m);
1110 			vm_page_busy(m);
1111 			break;
1112 		}
1113 		m = NULL;
1114 	}
1115 	lwkt_reltoken(&vm_token);
1116 	crit_exit();
1117 	return (m);
1118 }
1119 
1120 /*
1121  * vm_page_unmanage()
1122  *
1123  * Prevent PV management from being done on the page.  The page is
1124  * removed from the paging queues as if it were wired, and as a
1125  * consequence of no longer being managed the pageout daemon will not
1126  * touch it (since there is no way to locate the pte mappings for the
1127  * page).  madvise() calls that mess with the pmap will also no longer
1128  * operate on the page.
1129  *
1130  * Beyond that the page is still reasonably 'normal'.  Freeing the page
1131  * will clear the flag.
1132  *
1133  * This routine is used by OBJT_PHYS objects - objects using unswappable
1134  * physical memory as backing store rather then swap-backed memory and
1135  * will eventually be extended to support 4MB unmanaged physical
1136  * mappings.
1137  *
1138  * Must be called with a critical section held.
1139  */
1140 void
1141 vm_page_unmanage(vm_page_t m)
1142 {
1143 	ASSERT_IN_CRIT_SECTION();
1144 	if ((m->flags & PG_UNMANAGED) == 0) {
1145 		if (m->wire_count == 0)
1146 			vm_page_unqueue(m);
1147 	}
1148 	vm_page_flag_set(m, PG_UNMANAGED);
1149 }
1150 
1151 /*
1152  * Mark this page as wired down by yet another map, removing it from
1153  * paging queues as necessary.
1154  *
1155  * The page queues must be locked.
1156  * This routine may not block.
1157  */
1158 void
1159 vm_page_wire(vm_page_t m)
1160 {
1161 	/*
1162 	 * Only bump the wire statistics if the page is not already wired,
1163 	 * and only unqueue the page if it is on some queue (if it is unmanaged
1164 	 * it is already off the queues).  Don't do anything with fictitious
1165 	 * pages because they are always wired.
1166 	 */
1167 	crit_enter();
1168 	lwkt_gettoken(&vm_token);
1169 	if ((m->flags & PG_FICTITIOUS) == 0) {
1170 		if (m->wire_count == 0) {
1171 			if ((m->flags & PG_UNMANAGED) == 0)
1172 				vm_page_unqueue(m);
1173 			vmstats.v_wire_count++;
1174 		}
1175 		m->wire_count++;
1176 		KASSERT(m->wire_count != 0,
1177 			("vm_page_wire: wire_count overflow m=%p", m));
1178 	}
1179 	lwkt_reltoken(&vm_token);
1180 	crit_exit();
1181 }
1182 
1183 /*
1184  * Release one wiring of this page, potentially enabling it to be paged again.
1185  *
1186  * Many pages placed on the inactive queue should actually go
1187  * into the cache, but it is difficult to figure out which.  What
1188  * we do instead, if the inactive target is well met, is to put
1189  * clean pages at the head of the inactive queue instead of the tail.
1190  * This will cause them to be moved to the cache more quickly and
1191  * if not actively re-referenced, freed more quickly.  If we just
1192  * stick these pages at the end of the inactive queue, heavy filesystem
1193  * meta-data accesses can cause an unnecessary paging load on memory bound
1194  * processes.  This optimization causes one-time-use metadata to be
1195  * reused more quickly.
1196  *
1197  * BUT, if we are in a low-memory situation we have no choice but to
1198  * put clean pages on the cache queue.
1199  *
1200  * A number of routines use vm_page_unwire() to guarantee that the page
1201  * will go into either the inactive or active queues, and will NEVER
1202  * be placed in the cache - for example, just after dirtying a page.
1203  * dirty pages in the cache are not allowed.
1204  *
1205  * The page queues must be locked.
1206  * This routine may not block.
1207  */
1208 void
1209 vm_page_unwire(vm_page_t m, int activate)
1210 {
1211 	crit_enter();
1212 	lwkt_gettoken(&vm_token);
1213 	if (m->flags & PG_FICTITIOUS) {
1214 		/* do nothing */
1215 	} else if (m->wire_count <= 0) {
1216 		panic("vm_page_unwire: invalid wire count: %d", m->wire_count);
1217 	} else {
1218 		if (--m->wire_count == 0) {
1219 			--vmstats.v_wire_count;
1220 			if (m->flags & PG_UNMANAGED) {
1221 				;
1222 			} else if (activate) {
1223 				TAILQ_INSERT_TAIL(
1224 				    &vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1225 				m->queue = PQ_ACTIVE;
1226 				vm_page_queues[PQ_ACTIVE].lcnt++;
1227 				vmstats.v_active_count++;
1228 			} else {
1229 				vm_page_flag_clear(m, PG_WINATCFLS);
1230 				TAILQ_INSERT_TAIL(
1231 				    &vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1232 				m->queue = PQ_INACTIVE;
1233 				vm_page_queues[PQ_INACTIVE].lcnt++;
1234 				vmstats.v_inactive_count++;
1235 				++vm_swapcache_inactive_heuristic;
1236 			}
1237 		}
1238 	}
1239 	lwkt_reltoken(&vm_token);
1240 	crit_exit();
1241 }
1242 
1243 
1244 /*
1245  * Move the specified page to the inactive queue.  If the page has
1246  * any associated swap, the swap is deallocated.
1247  *
1248  * Normally athead is 0 resulting in LRU operation.  athead is set
1249  * to 1 if we want this page to be 'as if it were placed in the cache',
1250  * except without unmapping it from the process address space.
1251  *
1252  * This routine may not block.
1253  */
1254 static __inline void
1255 _vm_page_deactivate(vm_page_t m, int athead)
1256 {
1257 	/*
1258 	 * Ignore if already inactive.
1259 	 */
1260 	if (m->queue == PQ_INACTIVE)
1261 		return;
1262 
1263 	if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1264 		if ((m->queue - m->pc) == PQ_CACHE)
1265 			mycpu->gd_cnt.v_reactivated++;
1266 		vm_page_flag_clear(m, PG_WINATCFLS);
1267 		vm_page_unqueue(m);
1268 		if (athead) {
1269 			TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl,
1270 					  m, pageq);
1271 		} else {
1272 			TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl,
1273 					  m, pageq);
1274 			++vm_swapcache_inactive_heuristic;
1275 		}
1276 		m->queue = PQ_INACTIVE;
1277 		vm_page_queues[PQ_INACTIVE].lcnt++;
1278 		vmstats.v_inactive_count++;
1279 	}
1280 }
1281 
1282 void
1283 vm_page_deactivate(vm_page_t m)
1284 {
1285     crit_enter();
1286     lwkt_gettoken(&vm_token);
1287     _vm_page_deactivate(m, 0);
1288     lwkt_reltoken(&vm_token);
1289     crit_exit();
1290 }
1291 
1292 /*
1293  * vm_page_try_to_cache:
1294  *
1295  * Returns 0 on failure, 1 on success
1296  */
1297 int
1298 vm_page_try_to_cache(vm_page_t m)
1299 {
1300 	crit_enter();
1301 	lwkt_gettoken(&vm_token);
1302 	if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1303 	    (m->flags & (PG_BUSY|PG_UNMANAGED))) {
1304 		lwkt_reltoken(&vm_token);
1305 		crit_exit();
1306 		return(0);
1307 	}
1308 	vm_page_test_dirty(m);
1309 	if (m->dirty) {
1310 		lwkt_reltoken(&vm_token);
1311 		crit_exit();
1312 		return(0);
1313 	}
1314 	vm_page_cache(m);
1315 	lwkt_reltoken(&vm_token);
1316 	crit_exit();
1317 	return(1);
1318 }
1319 
1320 /*
1321  * Attempt to free the page.  If we cannot free it, we do nothing.
1322  * 1 is returned on success, 0 on failure.
1323  */
1324 int
1325 vm_page_try_to_free(vm_page_t m)
1326 {
1327 	crit_enter();
1328 	lwkt_gettoken(&vm_token);
1329 	if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1330 	    (m->flags & (PG_BUSY|PG_UNMANAGED))) {
1331 		lwkt_reltoken(&vm_token);
1332 		crit_exit();
1333 		return(0);
1334 	}
1335 	vm_page_test_dirty(m);
1336 	if (m->dirty) {
1337 		lwkt_reltoken(&vm_token);
1338 		crit_exit();
1339 		return(0);
1340 	}
1341 	vm_page_busy(m);
1342 	vm_page_protect(m, VM_PROT_NONE);
1343 	vm_page_free(m);
1344 	lwkt_reltoken(&vm_token);
1345 	crit_exit();
1346 	return(1);
1347 }
1348 
1349 /*
1350  * vm_page_cache
1351  *
1352  * Put the specified page onto the page cache queue (if appropriate).
1353  *
1354  * This routine may not block.
1355  */
1356 void
1357 vm_page_cache(vm_page_t m)
1358 {
1359 	ASSERT_IN_CRIT_SECTION();
1360 
1361 	if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy ||
1362 			m->wire_count || m->hold_count) {
1363 		kprintf("vm_page_cache: attempting to cache busy/held page\n");
1364 		return;
1365 	}
1366 
1367 	/*
1368 	 * Already in the cache (and thus not mapped)
1369 	 */
1370 	if ((m->queue - m->pc) == PQ_CACHE) {
1371 		KKASSERT((m->flags & PG_MAPPED) == 0);
1372 		return;
1373 	}
1374 
1375 	/*
1376 	 * Caller is required to test m->dirty, but note that the act of
1377 	 * removing the page from its maps can cause it to become dirty
1378 	 * on an SMP system due to another cpu running in usermode.
1379 	 */
1380 	if (m->dirty) {
1381 		panic("vm_page_cache: caching a dirty page, pindex: %ld",
1382 			(long)m->pindex);
1383 	}
1384 
1385 	/*
1386 	 * Remove all pmaps and indicate that the page is not
1387 	 * writeable or mapped.  Our vm_page_protect() call may
1388 	 * have blocked (especially w/ VM_PROT_NONE), so recheck
1389 	 * everything.
1390 	 */
1391 	vm_page_busy(m);
1392 	vm_page_protect(m, VM_PROT_NONE);
1393 	vm_page_wakeup(m);
1394 	if ((m->flags & (PG_BUSY|PG_UNMANAGED|PG_MAPPED)) || m->busy ||
1395 			m->wire_count || m->hold_count) {
1396 		/* do nothing */
1397 	} else if (m->dirty) {
1398 		vm_page_deactivate(m);
1399 	} else {
1400 		vm_page_unqueue_nowakeup(m);
1401 		m->queue = PQ_CACHE + m->pc;
1402 		vm_page_queues[m->queue].lcnt++;
1403 		TAILQ_INSERT_TAIL(&vm_page_queues[m->queue].pl, m, pageq);
1404 		vmstats.v_cache_count++;
1405 		vm_page_free_wakeup();
1406 	}
1407 }
1408 
1409 /*
1410  * vm_page_dontneed()
1411  *
1412  * Cache, deactivate, or do nothing as appropriate.  This routine
1413  * is typically used by madvise() MADV_DONTNEED.
1414  *
1415  * Generally speaking we want to move the page into the cache so
1416  * it gets reused quickly.  However, this can result in a silly syndrome
1417  * due to the page recycling too quickly.  Small objects will not be
1418  * fully cached.  On the otherhand, if we move the page to the inactive
1419  * queue we wind up with a problem whereby very large objects
1420  * unnecessarily blow away our inactive and cache queues.
1421  *
1422  * The solution is to move the pages based on a fixed weighting.  We
1423  * either leave them alone, deactivate them, or move them to the cache,
1424  * where moving them to the cache has the highest weighting.
1425  * By forcing some pages into other queues we eventually force the
1426  * system to balance the queues, potentially recovering other unrelated
1427  * space from active.  The idea is to not force this to happen too
1428  * often.
1429  */
1430 void
1431 vm_page_dontneed(vm_page_t m)
1432 {
1433 	static int dnweight;
1434 	int dnw;
1435 	int head;
1436 
1437 	dnw = ++dnweight;
1438 
1439 	/*
1440 	 * occassionally leave the page alone
1441 	 */
1442 	crit_enter();
1443 	lwkt_gettoken(&vm_token);
1444 	if ((dnw & 0x01F0) == 0 ||
1445 	    m->queue == PQ_INACTIVE ||
1446 	    m->queue - m->pc == PQ_CACHE
1447 	) {
1448 		if (m->act_count >= ACT_INIT)
1449 			--m->act_count;
1450 		lwkt_reltoken(&vm_token);
1451 		crit_exit();
1452 		return;
1453 	}
1454 
1455 	if (m->dirty == 0)
1456 		vm_page_test_dirty(m);
1457 
1458 	if (m->dirty || (dnw & 0x0070) == 0) {
1459 		/*
1460 		 * Deactivate the page 3 times out of 32.
1461 		 */
1462 		head = 0;
1463 	} else {
1464 		/*
1465 		 * Cache the page 28 times out of every 32.  Note that
1466 		 * the page is deactivated instead of cached, but placed
1467 		 * at the head of the queue instead of the tail.
1468 		 */
1469 		head = 1;
1470 	}
1471 	_vm_page_deactivate(m, head);
1472 	lwkt_reltoken(&vm_token);
1473 	crit_exit();
1474 }
1475 
1476 /*
1477  * Grab a page, blocking if it is busy and allocating a page if necessary.
1478  * A busy page is returned or NULL.
1479  *
1480  * If VM_ALLOC_RETRY is specified VM_ALLOC_NORMAL must also be specified.
1481  * If VM_ALLOC_RETRY is not specified
1482  *
1483  * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is
1484  * always returned if we had blocked.
1485  * This routine will never return NULL if VM_ALLOC_RETRY is set.
1486  * This routine may not be called from an interrupt.
1487  * The returned page may not be entirely valid.
1488  *
1489  * This routine may be called from mainline code without spl protection and
1490  * be guarenteed a busied page associated with the object at the specified
1491  * index.
1492  */
1493 vm_page_t
1494 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
1495 {
1496 	vm_page_t m;
1497 	int generation;
1498 
1499 	KKASSERT(allocflags &
1500 		(VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
1501 	crit_enter();
1502 	lwkt_gettoken(&vm_token);
1503 retrylookup:
1504 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
1505 		if (m->busy || (m->flags & PG_BUSY)) {
1506 			generation = object->generation;
1507 
1508 			while ((object->generation == generation) &&
1509 					(m->busy || (m->flags & PG_BUSY))) {
1510 				vm_page_flag_set(m, PG_WANTED | PG_REFERENCED);
1511 				tsleep(m, 0, "pgrbwt", 0);
1512 				if ((allocflags & VM_ALLOC_RETRY) == 0) {
1513 					m = NULL;
1514 					goto done;
1515 				}
1516 			}
1517 			goto retrylookup;
1518 		} else {
1519 			vm_page_busy(m);
1520 			goto done;
1521 		}
1522 	}
1523 	m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY);
1524 	if (m == NULL) {
1525 		vm_wait(0);
1526 		if ((allocflags & VM_ALLOC_RETRY) == 0)
1527 			goto done;
1528 		goto retrylookup;
1529 	}
1530 done:
1531 	lwkt_reltoken(&vm_token);
1532 	crit_exit();
1533 	return(m);
1534 }
1535 
1536 /*
1537  * Mapping function for valid bits or for dirty bits in
1538  * a page.  May not block.
1539  *
1540  * Inputs are required to range within a page.
1541  */
1542 __inline int
1543 vm_page_bits(int base, int size)
1544 {
1545 	int first_bit;
1546 	int last_bit;
1547 
1548 	KASSERT(
1549 	    base + size <= PAGE_SIZE,
1550 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
1551 	);
1552 
1553 	if (size == 0)		/* handle degenerate case */
1554 		return(0);
1555 
1556 	first_bit = base >> DEV_BSHIFT;
1557 	last_bit = (base + size - 1) >> DEV_BSHIFT;
1558 
1559 	return ((2 << last_bit) - (1 << first_bit));
1560 }
1561 
1562 /*
1563  * Sets portions of a page valid and clean.  The arguments are expected
1564  * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
1565  * of any partial chunks touched by the range.  The invalid portion of
1566  * such chunks will be zero'd.
1567  *
1568  * NOTE: When truncating a buffer vnode_pager_setsize() will automatically
1569  *	 align base to DEV_BSIZE so as not to mark clean a partially
1570  *	 truncated device block.  Otherwise the dirty page status might be
1571  *	 lost.
1572  *
1573  * This routine may not block.
1574  *
1575  * (base + size) must be less then or equal to PAGE_SIZE.
1576  */
1577 static void
1578 _vm_page_zero_valid(vm_page_t m, int base, int size)
1579 {
1580 	int frag;
1581 	int endoff;
1582 
1583 	if (size == 0)	/* handle degenerate case */
1584 		return;
1585 
1586 	/*
1587 	 * If the base is not DEV_BSIZE aligned and the valid
1588 	 * bit is clear, we have to zero out a portion of the
1589 	 * first block.
1590 	 */
1591 
1592 	if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
1593 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0
1594 	) {
1595 		pmap_zero_page_area(
1596 		    VM_PAGE_TO_PHYS(m),
1597 		    frag,
1598 		    base - frag
1599 		);
1600 	}
1601 
1602 	/*
1603 	 * If the ending offset is not DEV_BSIZE aligned and the
1604 	 * valid bit is clear, we have to zero out a portion of
1605 	 * the last block.
1606 	 */
1607 
1608 	endoff = base + size;
1609 
1610 	if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
1611 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0
1612 	) {
1613 		pmap_zero_page_area(
1614 		    VM_PAGE_TO_PHYS(m),
1615 		    endoff,
1616 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1))
1617 		);
1618 	}
1619 }
1620 
1621 /*
1622  * Set valid, clear dirty bits.  If validating the entire
1623  * page we can safely clear the pmap modify bit.  We also
1624  * use this opportunity to clear the PG_NOSYNC flag.  If a process
1625  * takes a write fault on a MAP_NOSYNC memory area the flag will
1626  * be set again.
1627  *
1628  * We set valid bits inclusive of any overlap, but we can only
1629  * clear dirty bits for DEV_BSIZE chunks that are fully within
1630  * the range.
1631  */
1632 void
1633 vm_page_set_valid(vm_page_t m, int base, int size)
1634 {
1635 	_vm_page_zero_valid(m, base, size);
1636 	m->valid |= vm_page_bits(base, size);
1637 }
1638 
1639 
1640 /*
1641  * Set valid bits and clear dirty bits.
1642  *
1643  * NOTE: This function does not clear the pmap modified bit.
1644  *	 Also note that e.g. NFS may use a byte-granular base
1645  *	 and size.
1646  */
1647 void
1648 vm_page_set_validclean(vm_page_t m, int base, int size)
1649 {
1650 	int pagebits;
1651 
1652 	_vm_page_zero_valid(m, base, size);
1653 	pagebits = vm_page_bits(base, size);
1654 	m->valid |= pagebits;
1655 	m->dirty &= ~pagebits;
1656 	if (base == 0 && size == PAGE_SIZE) {
1657 		/*pmap_clear_modify(m);*/
1658 		vm_page_flag_clear(m, PG_NOSYNC);
1659 	}
1660 }
1661 
1662 /*
1663  * Set valid & dirty.  Used by buwrite()
1664  */
1665 void
1666 vm_page_set_validdirty(vm_page_t m, int base, int size)
1667 {
1668 	int pagebits;
1669 
1670 	pagebits = vm_page_bits(base, size);
1671 	m->valid |= pagebits;
1672 	m->dirty |= pagebits;
1673 	if (m->object)
1674 		vm_object_set_writeable_dirty(m->object);
1675 }
1676 
1677 /*
1678  * Clear dirty bits.
1679  *
1680  * NOTE: This function does not clear the pmap modified bit.
1681  *	 Also note that e.g. NFS may use a byte-granular base
1682  *	 and size.
1683  */
1684 void
1685 vm_page_clear_dirty(vm_page_t m, int base, int size)
1686 {
1687 	m->dirty &= ~vm_page_bits(base, size);
1688 	if (base == 0 && size == PAGE_SIZE) {
1689 		/*pmap_clear_modify(m);*/
1690 		vm_page_flag_clear(m, PG_NOSYNC);
1691 	}
1692 }
1693 
1694 /*
1695  * Make the page all-dirty.
1696  *
1697  * Also make sure the related object and vnode reflect the fact that the
1698  * object may now contain a dirty page.
1699  */
1700 void
1701 vm_page_dirty(vm_page_t m)
1702 {
1703 #ifdef INVARIANTS
1704         int pqtype = m->queue - m->pc;
1705 #endif
1706         KASSERT(pqtype != PQ_CACHE && pqtype != PQ_FREE,
1707                 ("vm_page_dirty: page in free/cache queue!"));
1708 	if (m->dirty != VM_PAGE_BITS_ALL) {
1709 		m->dirty = VM_PAGE_BITS_ALL;
1710 		if (m->object)
1711 			vm_object_set_writeable_dirty(m->object);
1712 	}
1713 }
1714 
1715 /*
1716  * Invalidates DEV_BSIZE'd chunks within a page.  Both the
1717  * valid and dirty bits for the effected areas are cleared.
1718  *
1719  * May not block.
1720  */
1721 void
1722 vm_page_set_invalid(vm_page_t m, int base, int size)
1723 {
1724 	int bits;
1725 
1726 	bits = vm_page_bits(base, size);
1727 	m->valid &= ~bits;
1728 	m->dirty &= ~bits;
1729 	m->object->generation++;
1730 }
1731 
1732 /*
1733  * The kernel assumes that the invalid portions of a page contain
1734  * garbage, but such pages can be mapped into memory by user code.
1735  * When this occurs, we must zero out the non-valid portions of the
1736  * page so user code sees what it expects.
1737  *
1738  * Pages are most often semi-valid when the end of a file is mapped
1739  * into memory and the file's size is not page aligned.
1740  */
1741 void
1742 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
1743 {
1744 	int b;
1745 	int i;
1746 
1747 	/*
1748 	 * Scan the valid bits looking for invalid sections that
1749 	 * must be zerod.  Invalid sub-DEV_BSIZE'd areas ( where the
1750 	 * valid bit may be set ) have already been zerod by
1751 	 * vm_page_set_validclean().
1752 	 */
1753 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
1754 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
1755 		    (m->valid & (1 << i))
1756 		) {
1757 			if (i > b) {
1758 				pmap_zero_page_area(
1759 				    VM_PAGE_TO_PHYS(m),
1760 				    b << DEV_BSHIFT,
1761 				    (i - b) << DEV_BSHIFT
1762 				);
1763 			}
1764 			b = i + 1;
1765 		}
1766 	}
1767 
1768 	/*
1769 	 * setvalid is TRUE when we can safely set the zero'd areas
1770 	 * as being valid.  We can do this if there are no cache consistency
1771 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
1772 	 */
1773 	if (setvalid)
1774 		m->valid = VM_PAGE_BITS_ALL;
1775 }
1776 
1777 /*
1778  * Is a (partial) page valid?  Note that the case where size == 0
1779  * will return FALSE in the degenerate case where the page is entirely
1780  * invalid, and TRUE otherwise.
1781  *
1782  * May not block.
1783  */
1784 int
1785 vm_page_is_valid(vm_page_t m, int base, int size)
1786 {
1787 	int bits = vm_page_bits(base, size);
1788 
1789 	if (m->valid && ((m->valid & bits) == bits))
1790 		return 1;
1791 	else
1792 		return 0;
1793 }
1794 
1795 /*
1796  * update dirty bits from pmap/mmu.  May not block.
1797  */
1798 void
1799 vm_page_test_dirty(vm_page_t m)
1800 {
1801 	if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) {
1802 		vm_page_dirty(m);
1803 	}
1804 }
1805 
1806 /*
1807  * Issue an event on a VM page.  Corresponding action structures are
1808  * removed from the page's list and called.
1809  */
1810 void
1811 vm_page_event_internal(vm_page_t m, vm_page_event_t event)
1812 {
1813 	struct vm_page_action *scan, *next;
1814 
1815 	LIST_FOREACH_MUTABLE(scan, &m->action_list, entry, next) {
1816 		if (scan->event == event) {
1817 			scan->event = VMEVENT_NONE;
1818 			LIST_REMOVE(scan, entry);
1819 			scan->func(m, scan);
1820 		}
1821 	}
1822 }
1823 
1824 
1825 #include "opt_ddb.h"
1826 #ifdef DDB
1827 #include <sys/kernel.h>
1828 
1829 #include <ddb/ddb.h>
1830 
1831 DB_SHOW_COMMAND(page, vm_page_print_page_info)
1832 {
1833 	db_printf("vmstats.v_free_count: %d\n", vmstats.v_free_count);
1834 	db_printf("vmstats.v_cache_count: %d\n", vmstats.v_cache_count);
1835 	db_printf("vmstats.v_inactive_count: %d\n", vmstats.v_inactive_count);
1836 	db_printf("vmstats.v_active_count: %d\n", vmstats.v_active_count);
1837 	db_printf("vmstats.v_wire_count: %d\n", vmstats.v_wire_count);
1838 	db_printf("vmstats.v_free_reserved: %d\n", vmstats.v_free_reserved);
1839 	db_printf("vmstats.v_free_min: %d\n", vmstats.v_free_min);
1840 	db_printf("vmstats.v_free_target: %d\n", vmstats.v_free_target);
1841 	db_printf("vmstats.v_cache_min: %d\n", vmstats.v_cache_min);
1842 	db_printf("vmstats.v_inactive_target: %d\n", vmstats.v_inactive_target);
1843 }
1844 
1845 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
1846 {
1847 	int i;
1848 	db_printf("PQ_FREE:");
1849 	for(i=0;i<PQ_L2_SIZE;i++) {
1850 		db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt);
1851 	}
1852 	db_printf("\n");
1853 
1854 	db_printf("PQ_CACHE:");
1855 	for(i=0;i<PQ_L2_SIZE;i++) {
1856 		db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt);
1857 	}
1858 	db_printf("\n");
1859 
1860 	db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n",
1861 		vm_page_queues[PQ_ACTIVE].lcnt,
1862 		vm_page_queues[PQ_INACTIVE].lcnt);
1863 }
1864 #endif /* DDB */
1865