xref: /dflybsd-src/sys/vm/vm_page.c (revision b14d39bea25ac6b89cc943b26215359832fdb51a)
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 #define VMACTION_HSIZE	256
100 #define VMACTION_HMASK	(VMACTION_HSIZE - 1)
101 
102 static void vm_page_queue_init(void);
103 static void vm_page_free_wakeup(void);
104 static vm_page_t vm_page_select_cache(vm_object_t, vm_pindex_t);
105 static vm_page_t _vm_page_list_find2(int basequeue, int index);
106 
107 struct vpgqueues vm_page_queues[PQ_COUNT]; /* Array of tailq lists */
108 
109 LIST_HEAD(vm_page_action_list, vm_page_action);
110 struct vm_page_action_list	action_list[VMACTION_HSIZE];
111 
112 
113 #define ASSERT_IN_CRIT_SECTION()	KKASSERT(crit_test(curthread));
114 
115 RB_GENERATE2(vm_page_rb_tree, vm_page, rb_entry, rb_vm_page_compare,
116 	     vm_pindex_t, pindex);
117 
118 static void
119 vm_page_queue_init(void)
120 {
121 	int i;
122 
123 	for (i = 0; i < PQ_L2_SIZE; i++)
124 		vm_page_queues[PQ_FREE+i].cnt = &vmstats.v_free_count;
125 	for (i = 0; i < PQ_L2_SIZE; i++)
126 		vm_page_queues[PQ_CACHE+i].cnt = &vmstats.v_cache_count;
127 
128 	vm_page_queues[PQ_INACTIVE].cnt = &vmstats.v_inactive_count;
129 	vm_page_queues[PQ_ACTIVE].cnt = &vmstats.v_active_count;
130 	vm_page_queues[PQ_HOLD].cnt = &vmstats.v_active_count;
131 	/* PQ_NONE has no queue */
132 
133 	for (i = 0; i < PQ_COUNT; i++)
134 		TAILQ_INIT(&vm_page_queues[i].pl);
135 
136 	for (i = 0; i < VMACTION_HSIZE; i++)
137 		LIST_INIT(&action_list[i]);
138 }
139 
140 /*
141  * note: place in initialized data section?  Is this necessary?
142  */
143 long first_page = 0;
144 int vm_page_array_size = 0;
145 int vm_page_zero_count = 0;
146 vm_page_t vm_page_array = 0;
147 
148 /*
149  * (low level boot)
150  *
151  * Sets the page size, perhaps based upon the memory size.
152  * Must be called before any use of page-size dependent functions.
153  */
154 void
155 vm_set_page_size(void)
156 {
157 	if (vmstats.v_page_size == 0)
158 		vmstats.v_page_size = PAGE_SIZE;
159 	if (((vmstats.v_page_size - 1) & vmstats.v_page_size) != 0)
160 		panic("vm_set_page_size: page size not a power of two");
161 }
162 
163 /*
164  * (low level boot)
165  *
166  * Add a new page to the freelist for use by the system.  New pages
167  * are added to both the head and tail of the associated free page
168  * queue in a bottom-up fashion, so both zero'd and non-zero'd page
169  * requests pull 'recent' adds (higher physical addresses) first.
170  *
171  * Must be called in a critical section.
172  */
173 vm_page_t
174 vm_add_new_page(vm_paddr_t pa)
175 {
176 	struct vpgqueues *vpq;
177 	vm_page_t m;
178 
179 	++vmstats.v_page_count;
180 	++vmstats.v_free_count;
181 	m = PHYS_TO_VM_PAGE(pa);
182 	m->phys_addr = pa;
183 	m->flags = 0;
184 	m->pc = (pa >> PAGE_SHIFT) & PQ_L2_MASK;
185 	m->queue = m->pc + PQ_FREE;
186 	KKASSERT(m->dirty == 0);
187 
188 	vpq = &vm_page_queues[m->queue];
189 	if (vpq->flipflop)
190 		TAILQ_INSERT_TAIL(&vpq->pl, m, pageq);
191 	else
192 		TAILQ_INSERT_HEAD(&vpq->pl, m, pageq);
193 	vpq->flipflop = 1 - vpq->flipflop;
194 
195 	vm_page_queues[m->queue].lcnt++;
196 	return (m);
197 }
198 
199 /*
200  * (low level boot)
201  *
202  * Initializes the resident memory module.
203  *
204  * Allocates memory for the page cells, and for the object/offset-to-page
205  * hash table headers.  Each page cell is initialized and placed on the
206  * free list.
207  *
208  * starta/enda represents the range of physical memory addresses available
209  * for use (skipping memory already used by the kernel), subject to
210  * phys_avail[].  Note that phys_avail[] has already mapped out memory
211  * already in use by the kernel.
212  */
213 vm_offset_t
214 vm_page_startup(vm_offset_t vaddr)
215 {
216 	vm_offset_t mapped;
217 	vm_size_t npages;
218 	vm_paddr_t page_range;
219 	vm_paddr_t new_end;
220 	int i;
221 	vm_paddr_t pa;
222 	int nblocks;
223 	vm_paddr_t last_pa;
224 	vm_paddr_t end;
225 	vm_paddr_t biggestone, biggestsize;
226 	vm_paddr_t total;
227 
228 	total = 0;
229 	biggestsize = 0;
230 	biggestone = 0;
231 	nblocks = 0;
232 	vaddr = round_page(vaddr);
233 
234 	for (i = 0; phys_avail[i + 1]; i += 2) {
235 		phys_avail[i] = round_page64(phys_avail[i]);
236 		phys_avail[i + 1] = trunc_page64(phys_avail[i + 1]);
237 	}
238 
239 	for (i = 0; phys_avail[i + 1]; i += 2) {
240 		vm_paddr_t size = phys_avail[i + 1] - phys_avail[i];
241 
242 		if (size > biggestsize) {
243 			biggestone = i;
244 			biggestsize = size;
245 		}
246 		++nblocks;
247 		total += size;
248 	}
249 
250 	end = phys_avail[biggestone+1];
251 	end = trunc_page(end);
252 
253 	/*
254 	 * Initialize the queue headers for the free queue, the active queue
255 	 * and the inactive queue.
256 	 */
257 
258 	vm_page_queue_init();
259 
260 	/* VKERNELs don't support minidumps and as such don't need vm_page_dump */
261 #if !defined(_KERNEL_VIRTUAL)
262 	/*
263 	 * Allocate a bitmap to indicate that a random physical page
264 	 * needs to be included in a minidump.
265 	 *
266 	 * The amd64 port needs this to indicate which direct map pages
267 	 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
268 	 *
269 	 * However, i386 still needs this workspace internally within the
270 	 * minidump code.  In theory, they are not needed on i386, but are
271 	 * included should the sf_buf code decide to use them.
272 	 */
273 	page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE;
274 	vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
275 	end -= vm_page_dump_size;
276 	vm_page_dump = (void *)pmap_map(&vaddr, end, end + vm_page_dump_size,
277 	    VM_PROT_READ | VM_PROT_WRITE);
278 	bzero((void *)vm_page_dump, vm_page_dump_size);
279 #endif
280 
281 	/*
282 	 * Compute the number of pages of memory that will be available for
283 	 * use (taking into account the overhead of a page structure per
284 	 * page).
285 	 */
286 	first_page = phys_avail[0] / PAGE_SIZE;
287 	page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page;
288 	npages = (total - (page_range * sizeof(struct vm_page))) / PAGE_SIZE;
289 
290 	/*
291 	 * Initialize the mem entry structures now, and put them in the free
292 	 * queue.
293 	 */
294 	new_end = trunc_page(end - page_range * sizeof(struct vm_page));
295 	mapped = pmap_map(&vaddr, new_end, end,
296 	    VM_PROT_READ | VM_PROT_WRITE);
297 	vm_page_array = (vm_page_t)mapped;
298 
299 #if defined(__x86_64__) && !defined(_KERNEL_VIRTUAL)
300 	/*
301 	 * since pmap_map on amd64 returns stuff out of a direct-map region,
302 	 * we have to manually add these pages to the minidump tracking so
303 	 * that they can be dumped, including the vm_page_array.
304 	 */
305 	for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE)
306 		dump_add_page(pa);
307 #endif
308 
309 	/*
310 	 * Clear all of the page structures
311 	 */
312 	bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
313 	vm_page_array_size = page_range;
314 
315 	/*
316 	 * Construct the free queue(s) in ascending order (by physical
317 	 * address) so that the first 16MB of physical memory is allocated
318 	 * last rather than first.  On large-memory machines, this avoids
319 	 * the exhaustion of low physical memory before isa_dmainit has run.
320 	 */
321 	vmstats.v_page_count = 0;
322 	vmstats.v_free_count = 0;
323 	for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) {
324 		pa = phys_avail[i];
325 		if (i == biggestone)
326 			last_pa = new_end;
327 		else
328 			last_pa = phys_avail[i + 1];
329 		while (pa < last_pa && npages-- > 0) {
330 			vm_add_new_page(pa);
331 			pa += PAGE_SIZE;
332 		}
333 	}
334 	return (vaddr);
335 }
336 
337 /*
338  * Scan comparison function for Red-Black tree scans.  An inclusive
339  * (start,end) is expected.  Other fields are not used.
340  */
341 int
342 rb_vm_page_scancmp(struct vm_page *p, void *data)
343 {
344 	struct rb_vm_page_scan_info *info = data;
345 
346 	if (p->pindex < info->start_pindex)
347 		return(-1);
348 	if (p->pindex > info->end_pindex)
349 		return(1);
350 	return(0);
351 }
352 
353 int
354 rb_vm_page_compare(struct vm_page *p1, struct vm_page *p2)
355 {
356 	if (p1->pindex < p2->pindex)
357 		return(-1);
358 	if (p1->pindex > p2->pindex)
359 		return(1);
360 	return(0);
361 }
362 
363 /*
364  * Holding a page keeps it from being reused.  Other parts of the system
365  * can still disassociate the page from its current object and free it, or
366  * perform read or write I/O on it and/or otherwise manipulate the page,
367  * but if the page is held the VM system will leave the page and its data
368  * intact and not reuse the page for other purposes until the last hold
369  * reference is released.  (see vm_page_wire() if you want to prevent the
370  * page from being disassociated from its object too).
371  *
372  * The caller must hold vm_token.
373  *
374  * The caller must still validate the contents of the page and, if necessary,
375  * wait for any pending I/O (e.g. vm_page_sleep_busy() loop) to complete
376  * before manipulating the page.
377  */
378 void
379 vm_page_hold(vm_page_t m)
380 {
381 	ASSERT_LWKT_TOKEN_HELD(&vm_token);
382 	++m->hold_count;
383 }
384 
385 /*
386  * The opposite of vm_page_hold().  A page can be freed while being held,
387  * which places it on the PQ_HOLD queue.  We must call vm_page_free_toq()
388  * in this case to actually free it once the hold count drops to 0.
389  *
390  * The caller must hold vm_token if non-blocking operation is desired,
391  * but otherwise does not need to.
392  */
393 void
394 vm_page_unhold(vm_page_t m)
395 {
396 	lwkt_gettoken(&vm_token);
397 	--m->hold_count;
398 	KASSERT(m->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!"));
399 	if (m->hold_count == 0 && m->queue == PQ_HOLD) {
400 		vm_page_busy(m);
401 		vm_page_free_toq(m);
402 	}
403 	lwkt_reltoken(&vm_token);
404 }
405 
406 /*
407  * Inserts the given vm_page into the object and object list.
408  *
409  * The pagetables are not updated but will presumably fault the page
410  * in if necessary, or if a kernel page the caller will at some point
411  * enter the page into the kernel's pmap.  We are not allowed to block
412  * here so we *can't* do this anyway.
413  *
414  * This routine may not block.
415  * This routine must be called with the vm_token held.
416  * This routine must be called with a critical section held.
417  */
418 void
419 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
420 {
421 	ASSERT_IN_CRIT_SECTION();
422 	ASSERT_LWKT_TOKEN_HELD(&vm_token);
423 	if (m->object != NULL)
424 		panic("vm_page_insert: already inserted");
425 
426 	/*
427 	 * Record the object/offset pair in this page
428 	 */
429 	m->object = object;
430 	m->pindex = pindex;
431 
432 	/*
433 	 * Insert it into the object.
434 	 */
435 	vm_page_rb_tree_RB_INSERT(&object->rb_memq, m);
436 	object->generation++;
437 
438 	/*
439 	 * show that the object has one more resident page.
440 	 */
441 	object->resident_page_count++;
442 
443 	/*
444 	 * Add the pv_list_cout of the page when its inserted in
445 	 * the object
446 	*/
447 	object->agg_pv_list_count = object->agg_pv_list_count + m->md.pv_list_count;
448 
449 	/*
450 	 * Since we are inserting a new and possibly dirty page,
451 	 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags.
452 	 */
453 	if ((m->valid & m->dirty) || (m->flags & PG_WRITEABLE))
454 		vm_object_set_writeable_dirty(object);
455 
456 	/*
457 	 * Checks for a swap assignment and sets PG_SWAPPED if appropriate.
458 	 */
459 	swap_pager_page_inserted(m);
460 }
461 
462 /*
463  * Removes the given vm_page_t from the global (object,index) hash table
464  * and from the object's memq.
465  *
466  * The underlying pmap entry (if any) is NOT removed here.
467  * This routine may not block.
468  *
469  * The page must be BUSY and will remain BUSY on return.
470  * No other requirements.
471  *
472  * NOTE: FreeBSD side effect was to unbusy the page on return.  We leave
473  *	 it busy.
474  */
475 void
476 vm_page_remove(vm_page_t m)
477 {
478 	vm_object_t object;
479 
480 	crit_enter();
481 	lwkt_gettoken(&vm_token);
482 	if (m->object == NULL) {
483 		lwkt_reltoken(&vm_token);
484 		crit_exit();
485 		return;
486 	}
487 
488 	if ((m->flags & PG_BUSY) == 0)
489 		panic("vm_page_remove: page not busy");
490 
491 	object = m->object;
492 
493 	/*
494 	 * Remove the page from the object and update the object.
495 	 */
496 	vm_page_rb_tree_RB_REMOVE(&object->rb_memq, m);
497 	object->resident_page_count--;
498 	object->agg_pv_list_count = object->agg_pv_list_count - m->md.pv_list_count;
499 	object->generation++;
500 	m->object = NULL;
501 
502 	lwkt_reltoken(&vm_token);
503 	crit_exit();
504 }
505 
506 /*
507  * Locate and return the page at (object, pindex), or NULL if the
508  * page could not be found.
509  *
510  * The caller must hold vm_token.
511  */
512 vm_page_t
513 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
514 {
515 	vm_page_t m;
516 
517 	/*
518 	 * Search the hash table for this object/offset pair
519 	 */
520 	ASSERT_LWKT_TOKEN_HELD(&vm_token);
521 	crit_enter();
522 	m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
523 	crit_exit();
524 	KKASSERT(m == NULL || (m->object == object && m->pindex == pindex));
525 	return(m);
526 }
527 
528 /*
529  * vm_page_rename()
530  *
531  * Move the given memory entry from its current object to the specified
532  * target object/offset.
533  *
534  * The object must be locked.
535  * This routine may not block.
536  *
537  * Note: This routine will raise itself to splvm(), the caller need not.
538  *
539  * Note: Swap associated with the page must be invalidated by the move.  We
540  *       have to do this for several reasons:  (1) we aren't freeing the
541  *       page, (2) we are dirtying the page, (3) the VM system is probably
542  *       moving the page from object A to B, and will then later move
543  *       the backing store from A to B and we can't have a conflict.
544  *
545  * Note: We *always* dirty the page.  It is necessary both for the
546  *       fact that we moved it, and because we may be invalidating
547  *	 swap.  If the page is on the cache, we have to deactivate it
548  *	 or vm_page_dirty() will panic.  Dirty pages are not allowed
549  *	 on the cache.
550  */
551 void
552 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
553 {
554 	crit_enter();
555 	lwkt_gettoken(&vm_token);
556 	vm_page_remove(m);
557 	vm_page_insert(m, new_object, new_pindex);
558 	if (m->queue - m->pc == PQ_CACHE)
559 		vm_page_deactivate(m);
560 	vm_page_dirty(m);
561 	vm_page_wakeup(m);
562 	lwkt_reltoken(&vm_token);
563 	crit_exit();
564 }
565 
566 /*
567  * vm_page_unqueue() without any wakeup.  This routine is used when a page
568  * is being moved between queues or otherwise is to remain BUSYied by the
569  * caller.
570  *
571  * The caller must hold vm_token
572  * This routine may not block.
573  */
574 void
575 vm_page_unqueue_nowakeup(vm_page_t m)
576 {
577 	int queue = m->queue;
578 	struct vpgqueues *pq;
579 
580 	ASSERT_LWKT_TOKEN_HELD(&vm_token);
581 	if (queue != PQ_NONE) {
582 		pq = &vm_page_queues[queue];
583 		m->queue = PQ_NONE;
584 		TAILQ_REMOVE(&pq->pl, m, pageq);
585 		(*pq->cnt)--;
586 		pq->lcnt--;
587 	}
588 }
589 
590 /*
591  * vm_page_unqueue() - Remove a page from its queue, wakeup the pagedemon
592  * if necessary.
593  *
594  * The caller must hold vm_token
595  * This routine may not block.
596  */
597 void
598 vm_page_unqueue(vm_page_t m)
599 {
600 	int queue = m->queue;
601 	struct vpgqueues *pq;
602 
603 	ASSERT_LWKT_TOKEN_HELD(&vm_token);
604 	if (queue != PQ_NONE) {
605 		m->queue = PQ_NONE;
606 		pq = &vm_page_queues[queue];
607 		TAILQ_REMOVE(&pq->pl, m, pageq);
608 		(*pq->cnt)--;
609 		pq->lcnt--;
610 		if ((queue - m->pc) == PQ_CACHE || (queue - m->pc) == PQ_FREE)
611 			pagedaemon_wakeup();
612 	}
613 }
614 
615 /*
616  * vm_page_list_find()
617  *
618  * Find a page on the specified queue with color optimization.
619  *
620  * The page coloring optimization attempts to locate a page that does
621  * not overload other nearby pages in the object in the cpu's L1 or L2
622  * caches.  We need this optimization because cpu caches tend to be
623  * physical caches, while object spaces tend to be virtual.
624  *
625  * Must be called with vm_token held.
626  * This routine may not block.
627  *
628  * Note that this routine is carefully inlined.  A non-inlined version
629  * is available for outside callers but the only critical path is
630  * from within this source file.
631  */
632 static __inline
633 vm_page_t
634 _vm_page_list_find(int basequeue, int index, boolean_t prefer_zero)
635 {
636 	vm_page_t m;
637 
638 	if (prefer_zero)
639 		m = TAILQ_LAST(&vm_page_queues[basequeue+index].pl, pglist);
640 	else
641 		m = TAILQ_FIRST(&vm_page_queues[basequeue+index].pl);
642 	if (m == NULL)
643 		m = _vm_page_list_find2(basequeue, index);
644 	return(m);
645 }
646 
647 static vm_page_t
648 _vm_page_list_find2(int basequeue, int index)
649 {
650 	int i;
651 	vm_page_t m = NULL;
652 	struct vpgqueues *pq;
653 
654 	pq = &vm_page_queues[basequeue];
655 
656 	/*
657 	 * Note that for the first loop, index+i and index-i wind up at the
658 	 * same place.  Even though this is not totally optimal, we've already
659 	 * blown it by missing the cache case so we do not care.
660 	 */
661 
662 	for(i = PQ_L2_SIZE / 2; i > 0; --i) {
663 		if ((m = TAILQ_FIRST(&pq[(index + i) & PQ_L2_MASK].pl)) != NULL)
664 			break;
665 
666 		if ((m = TAILQ_FIRST(&pq[(index - i) & PQ_L2_MASK].pl)) != NULL)
667 			break;
668 	}
669 	return(m);
670 }
671 
672 /*
673  * Must be called with vm_token held if the caller desired non-blocking
674  * operation and a stable result.
675  */
676 vm_page_t
677 vm_page_list_find(int basequeue, int index, boolean_t prefer_zero)
678 {
679 	return(_vm_page_list_find(basequeue, index, prefer_zero));
680 }
681 
682 /*
683  * Find a page on the cache queue with color optimization.  As pages
684  * might be found, but not applicable, they are deactivated.  This
685  * keeps us from using potentially busy cached pages.
686  *
687  * This routine may not block.
688  * Must be called with vm_token held.
689  */
690 vm_page_t
691 vm_page_select_cache(vm_object_t object, vm_pindex_t pindex)
692 {
693 	vm_page_t m;
694 
695 	ASSERT_LWKT_TOKEN_HELD(&vm_token);
696 	while (TRUE) {
697 		m = _vm_page_list_find(
698 		    PQ_CACHE,
699 		    (pindex + object->pg_color) & PQ_L2_MASK,
700 		    FALSE
701 		);
702 		if (m && ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy ||
703 			       m->hold_count || m->wire_count)) {
704 			vm_page_deactivate(m);
705 			continue;
706 		}
707 		return m;
708 	}
709 	/* not reached */
710 }
711 
712 /*
713  * Find a free or zero page, with specified preference.  We attempt to
714  * inline the nominal case and fall back to _vm_page_select_free()
715  * otherwise.
716  *
717  * This routine must be called with a critical section held.
718  * This routine may not block.
719  */
720 static __inline vm_page_t
721 vm_page_select_free(vm_object_t object, vm_pindex_t pindex, boolean_t prefer_zero)
722 {
723 	vm_page_t m;
724 
725 	m = _vm_page_list_find(
726 		PQ_FREE,
727 		(pindex + object->pg_color) & PQ_L2_MASK,
728 		prefer_zero
729 	);
730 	return(m);
731 }
732 
733 /*
734  * vm_page_alloc()
735  *
736  * Allocate and return a memory cell associated with this VM object/offset
737  * pair.
738  *
739  *	page_req classes:
740  *
741  *	VM_ALLOC_NORMAL		allow use of cache pages, nominal free drain
742  *	VM_ALLOC_QUICK		like normal but cannot use cache
743  *	VM_ALLOC_SYSTEM		greater free drain
744  *	VM_ALLOC_INTERRUPT	allow free list to be completely drained
745  *	VM_ALLOC_ZERO		advisory request for pre-zero'd page
746  *
747  * The object must be locked.
748  * This routine may not block.
749  * The returned page will be marked PG_BUSY
750  *
751  * Additional special handling is required when called from an interrupt
752  * (VM_ALLOC_INTERRUPT).  We are not allowed to mess with the page cache
753  * in this case.
754  */
755 vm_page_t
756 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req)
757 {
758 	vm_page_t m = NULL;
759 
760 	crit_enter();
761 	lwkt_gettoken(&vm_token);
762 
763 	KKASSERT(object != NULL);
764 	KASSERT(!vm_page_lookup(object, pindex),
765 		("vm_page_alloc: page already allocated"));
766 	KKASSERT(page_req &
767 		(VM_ALLOC_NORMAL|VM_ALLOC_QUICK|
768 		 VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
769 
770 	/*
771 	 * Certain system threads (pageout daemon, buf_daemon's) are
772 	 * allowed to eat deeper into the free page list.
773 	 */
774 	if (curthread->td_flags & TDF_SYSTHREAD)
775 		page_req |= VM_ALLOC_SYSTEM;
776 
777 loop:
778 	if (vmstats.v_free_count > vmstats.v_free_reserved ||
779 	    ((page_req & VM_ALLOC_INTERRUPT) && vmstats.v_free_count > 0) ||
780 	    ((page_req & VM_ALLOC_SYSTEM) && vmstats.v_cache_count == 0 &&
781 		vmstats.v_free_count > vmstats.v_interrupt_free_min)
782 	) {
783 		/*
784 		 * The free queue has sufficient free pages to take one out.
785 		 */
786 		if (page_req & VM_ALLOC_ZERO)
787 			m = vm_page_select_free(object, pindex, TRUE);
788 		else
789 			m = vm_page_select_free(object, pindex, FALSE);
790 	} else if (page_req & VM_ALLOC_NORMAL) {
791 		/*
792 		 * Allocatable from the cache (non-interrupt only).  On
793 		 * success, we must free the page and try again, thus
794 		 * ensuring that vmstats.v_*_free_min counters are replenished.
795 		 */
796 #ifdef INVARIANTS
797 		if (curthread->td_preempted) {
798 			kprintf("vm_page_alloc(): warning, attempt to allocate"
799 				" cache page from preempting interrupt\n");
800 			m = NULL;
801 		} else {
802 			m = vm_page_select_cache(object, pindex);
803 		}
804 #else
805 		m = vm_page_select_cache(object, pindex);
806 #endif
807 		/*
808 		 * On success move the page into the free queue and loop.
809 		 */
810 		if (m != NULL) {
811 			KASSERT(m->dirty == 0,
812 			    ("Found dirty cache page %p", m));
813 			vm_page_busy(m);
814 			vm_page_protect(m, VM_PROT_NONE);
815 			vm_page_free(m);
816 			goto loop;
817 		}
818 
819 		/*
820 		 * On failure return NULL
821 		 */
822 		lwkt_reltoken(&vm_token);
823 		crit_exit();
824 #if defined(DIAGNOSTIC)
825 		if (vmstats.v_cache_count > 0)
826 			kprintf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", vmstats.v_cache_count);
827 #endif
828 		vm_pageout_deficit++;
829 		pagedaemon_wakeup();
830 		return (NULL);
831 	} else {
832 		/*
833 		 * No pages available, wakeup the pageout daemon and give up.
834 		 */
835 		lwkt_reltoken(&vm_token);
836 		crit_exit();
837 		vm_pageout_deficit++;
838 		pagedaemon_wakeup();
839 		return (NULL);
840 	}
841 
842 	/*
843 	 * Good page found.  The page has not yet been busied.  We are in
844 	 * a critical section.
845 	 */
846 	KASSERT(m != NULL, ("vm_page_alloc(): missing page on free queue\n"));
847 	KASSERT(m->dirty == 0,
848 		("vm_page_alloc: free/cache page %p was dirty", m));
849 
850 	/*
851 	 * Remove from free queue
852 	 */
853 	vm_page_unqueue_nowakeup(m);
854 
855 	/*
856 	 * Initialize structure.  Only the PG_ZERO flag is inherited.  Set
857 	 * the page PG_BUSY
858 	 */
859 	if (m->flags & PG_ZERO) {
860 		vm_page_zero_count--;
861 		m->flags = PG_ZERO | PG_BUSY;
862 	} else {
863 		m->flags = PG_BUSY;
864 	}
865 	m->wire_count = 0;
866 	m->hold_count = 0;
867 	m->act_count = 0;
868 	m->busy = 0;
869 	m->valid = 0;
870 
871 	/*
872 	 * vm_page_insert() is safe prior to the crit_exit().  Note also that
873 	 * inserting a page here does not insert it into the pmap (which
874 	 * could cause us to block allocating memory).  We cannot block
875 	 * anywhere.
876 	 */
877 	vm_page_insert(m, object, pindex);
878 
879 	/*
880 	 * Don't wakeup too often - wakeup the pageout daemon when
881 	 * we would be nearly out of memory.
882 	 */
883 	pagedaemon_wakeup();
884 
885 	lwkt_reltoken(&vm_token);
886 	crit_exit();
887 
888 	/*
889 	 * A PG_BUSY page is returned.
890 	 */
891 	return (m);
892 }
893 
894 /*
895  * Wait for sufficient free memory for nominal heavy memory use kernel
896  * operations.
897  */
898 void
899 vm_wait_nominal(void)
900 {
901 	while (vm_page_count_min(0))
902 		vm_wait(0);
903 }
904 
905 /*
906  * Test if vm_wait_nominal() would block.
907  */
908 int
909 vm_test_nominal(void)
910 {
911 	if (vm_page_count_min(0))
912 		return(1);
913 	return(0);
914 }
915 
916 /*
917  * Block until free pages are available for allocation, called in various
918  * places before memory allocations.
919  */
920 void
921 vm_wait(int timo)
922 {
923 	crit_enter();
924 	lwkt_gettoken(&vm_token);
925 	if (curthread == pagethread) {
926 		vm_pageout_pages_needed = 1;
927 		tsleep(&vm_pageout_pages_needed, 0, "VMWait", timo);
928 	} else {
929 		if (vm_pages_needed == 0) {
930 			vm_pages_needed = 1;
931 			wakeup(&vm_pages_needed);
932 		}
933 		tsleep(&vmstats.v_free_count, 0, "vmwait", timo);
934 	}
935 	lwkt_reltoken(&vm_token);
936 	crit_exit();
937 }
938 
939 /*
940  * Block until free pages are available for allocation
941  *
942  * Called only in vm_fault so that processes page faulting can be
943  * easily tracked.
944  */
945 void
946 vm_waitpfault(void)
947 {
948 	crit_enter();
949 	lwkt_gettoken(&vm_token);
950 	if (vm_pages_needed == 0) {
951 		vm_pages_needed = 1;
952 		wakeup(&vm_pages_needed);
953 	}
954 	tsleep(&vmstats.v_free_count, 0, "pfault", 0);
955 	lwkt_reltoken(&vm_token);
956 	crit_exit();
957 }
958 
959 /*
960  * Put the specified page on the active list (if appropriate).  Ensure
961  * that act_count is at least ACT_INIT but do not otherwise mess with it.
962  *
963  * The page queues must be locked.
964  * This routine may not block.
965  */
966 void
967 vm_page_activate(vm_page_t m)
968 {
969 	crit_enter();
970 	lwkt_gettoken(&vm_token);
971 	if (m->queue != PQ_ACTIVE) {
972 		if ((m->queue - m->pc) == PQ_CACHE)
973 			mycpu->gd_cnt.v_reactivated++;
974 
975 		vm_page_unqueue(m);
976 
977 		if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
978 			m->queue = PQ_ACTIVE;
979 			vm_page_queues[PQ_ACTIVE].lcnt++;
980 			TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl,
981 					    m, pageq);
982 			if (m->act_count < ACT_INIT)
983 				m->act_count = ACT_INIT;
984 			vmstats.v_active_count++;
985 		}
986 	} else {
987 		if (m->act_count < ACT_INIT)
988 			m->act_count = ACT_INIT;
989 	}
990 	lwkt_reltoken(&vm_token);
991 	crit_exit();
992 }
993 
994 /*
995  * Helper routine for vm_page_free_toq() and vm_page_cache().  This
996  * routine is called when a page has been added to the cache or free
997  * queues.
998  *
999  * This routine may not block.
1000  * This routine must be called at splvm()
1001  */
1002 static __inline void
1003 vm_page_free_wakeup(void)
1004 {
1005 	/*
1006 	 * if pageout daemon needs pages, then tell it that there are
1007 	 * some free.
1008 	 */
1009 	if (vm_pageout_pages_needed &&
1010 	    vmstats.v_cache_count + vmstats.v_free_count >=
1011 	    vmstats.v_pageout_free_min
1012 	) {
1013 		wakeup(&vm_pageout_pages_needed);
1014 		vm_pageout_pages_needed = 0;
1015 	}
1016 
1017 	/*
1018 	 * wakeup processes that are waiting on memory if we hit a
1019 	 * high water mark. And wakeup scheduler process if we have
1020 	 * lots of memory. this process will swapin processes.
1021 	 */
1022 	if (vm_pages_needed && !vm_page_count_min(0)) {
1023 		vm_pages_needed = 0;
1024 		wakeup(&vmstats.v_free_count);
1025 	}
1026 }
1027 
1028 /*
1029  *	vm_page_free_toq:
1030  *
1031  *	Returns the given page to the PQ_FREE list, disassociating it with
1032  *	any VM object.
1033  *
1034  *	The vm_page must be PG_BUSY on entry.  PG_BUSY will be released on
1035  *	return (the page will have been freed).  No particular spl is required
1036  *	on entry.
1037  *
1038  *	This routine may not block.
1039  */
1040 void
1041 vm_page_free_toq(vm_page_t m)
1042 {
1043 	struct vpgqueues *pq;
1044 
1045 	crit_enter();
1046 	lwkt_gettoken(&vm_token);
1047 	mycpu->gd_cnt.v_tfree++;
1048 
1049 	KKASSERT((m->flags & PG_MAPPED) == 0);
1050 
1051 	if (m->busy || ((m->queue - m->pc) == PQ_FREE)) {
1052 		kprintf(
1053 		"vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n",
1054 		    (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0,
1055 		    m->hold_count);
1056 		if ((m->queue - m->pc) == PQ_FREE)
1057 			panic("vm_page_free: freeing free page");
1058 		else
1059 			panic("vm_page_free: freeing busy page");
1060 	}
1061 
1062 	/*
1063 	 * unqueue, then remove page.  Note that we cannot destroy
1064 	 * the page here because we do not want to call the pager's
1065 	 * callback routine until after we've put the page on the
1066 	 * appropriate free queue.
1067 	 */
1068 	vm_page_unqueue_nowakeup(m);
1069 	vm_page_remove(m);
1070 
1071 	/*
1072 	 * No further management of fictitious pages occurs beyond object
1073 	 * and queue removal.
1074 	 */
1075 	if ((m->flags & PG_FICTITIOUS) != 0) {
1076 		vm_page_wakeup(m);
1077 		lwkt_reltoken(&vm_token);
1078 		crit_exit();
1079 		return;
1080 	}
1081 
1082 	m->valid = 0;
1083 	vm_page_undirty(m);
1084 
1085 	if (m->wire_count != 0) {
1086 		if (m->wire_count > 1) {
1087 		    panic(
1088 			"vm_page_free: invalid wire count (%d), pindex: 0x%lx",
1089 			m->wire_count, (long)m->pindex);
1090 		}
1091 		panic("vm_page_free: freeing wired page");
1092 	}
1093 
1094 	/*
1095 	 * Clear the UNMANAGED flag when freeing an unmanaged page.
1096 	 */
1097 	if (m->flags & PG_UNMANAGED) {
1098 	    m->flags &= ~PG_UNMANAGED;
1099 	}
1100 
1101 	if (m->hold_count != 0) {
1102 		m->flags &= ~PG_ZERO;
1103 		m->queue = PQ_HOLD;
1104 	} else {
1105 		m->queue = PQ_FREE + m->pc;
1106 	}
1107 	pq = &vm_page_queues[m->queue];
1108 	pq->lcnt++;
1109 	++(*pq->cnt);
1110 
1111 	/*
1112 	 * Put zero'd pages on the end ( where we look for zero'd pages
1113 	 * first ) and non-zerod pages at the head.
1114 	 */
1115 	if (m->flags & PG_ZERO) {
1116 		TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
1117 		++vm_page_zero_count;
1118 	} else {
1119 		TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
1120 	}
1121 	vm_page_wakeup(m);
1122 	vm_page_free_wakeup();
1123 	lwkt_reltoken(&vm_token);
1124 	crit_exit();
1125 }
1126 
1127 /*
1128  * vm_page_free_fromq_fast()
1129  *
1130  * Remove a non-zero page from one of the free queues; the page is removed for
1131  * zeroing, so do not issue a wakeup.
1132  *
1133  * MPUNSAFE
1134  */
1135 vm_page_t
1136 vm_page_free_fromq_fast(void)
1137 {
1138 	static int qi;
1139 	vm_page_t m;
1140 	int i;
1141 
1142 	crit_enter();
1143 	lwkt_gettoken(&vm_token);
1144 	for (i = 0; i < PQ_L2_SIZE; ++i) {
1145 		m = vm_page_list_find(PQ_FREE, qi, FALSE);
1146 		qi = (qi + PQ_PRIME2) & PQ_L2_MASK;
1147 		if (m && (m->flags & PG_ZERO) == 0) {
1148 			vm_page_unqueue_nowakeup(m);
1149 			vm_page_busy(m);
1150 			break;
1151 		}
1152 		m = NULL;
1153 	}
1154 	lwkt_reltoken(&vm_token);
1155 	crit_exit();
1156 	return (m);
1157 }
1158 
1159 /*
1160  * vm_page_unmanage()
1161  *
1162  * Prevent PV management from being done on the page.  The page is
1163  * removed from the paging queues as if it were wired, and as a
1164  * consequence of no longer being managed the pageout daemon will not
1165  * touch it (since there is no way to locate the pte mappings for the
1166  * page).  madvise() calls that mess with the pmap will also no longer
1167  * operate on the page.
1168  *
1169  * Beyond that the page is still reasonably 'normal'.  Freeing the page
1170  * will clear the flag.
1171  *
1172  * This routine is used by OBJT_PHYS objects - objects using unswappable
1173  * physical memory as backing store rather then swap-backed memory and
1174  * will eventually be extended to support 4MB unmanaged physical
1175  * mappings.
1176  *
1177  * Must be called with a critical section held.
1178  * Must be called with vm_token held.
1179  */
1180 void
1181 vm_page_unmanage(vm_page_t m)
1182 {
1183 	ASSERT_IN_CRIT_SECTION();
1184 	ASSERT_LWKT_TOKEN_HELD(&vm_token);
1185 	if ((m->flags & PG_UNMANAGED) == 0) {
1186 		if (m->wire_count == 0)
1187 			vm_page_unqueue(m);
1188 	}
1189 	vm_page_flag_set(m, PG_UNMANAGED);
1190 }
1191 
1192 /*
1193  * Mark this page as wired down by yet another map, removing it from
1194  * paging queues as necessary.
1195  *
1196  * The page queues must be locked.
1197  * This routine may not block.
1198  */
1199 void
1200 vm_page_wire(vm_page_t m)
1201 {
1202 	/*
1203 	 * Only bump the wire statistics if the page is not already wired,
1204 	 * and only unqueue the page if it is on some queue (if it is unmanaged
1205 	 * it is already off the queues).  Don't do anything with fictitious
1206 	 * pages because they are always wired.
1207 	 */
1208 	crit_enter();
1209 	lwkt_gettoken(&vm_token);
1210 	if ((m->flags & PG_FICTITIOUS) == 0) {
1211 		if (m->wire_count == 0) {
1212 			if ((m->flags & PG_UNMANAGED) == 0)
1213 				vm_page_unqueue(m);
1214 			vmstats.v_wire_count++;
1215 		}
1216 		m->wire_count++;
1217 		KASSERT(m->wire_count != 0,
1218 			("vm_page_wire: wire_count overflow m=%p", m));
1219 	}
1220 	lwkt_reltoken(&vm_token);
1221 	crit_exit();
1222 }
1223 
1224 /*
1225  * Release one wiring of this page, potentially enabling it to be paged again.
1226  *
1227  * Many pages placed on the inactive queue should actually go
1228  * into the cache, but it is difficult to figure out which.  What
1229  * we do instead, if the inactive target is well met, is to put
1230  * clean pages at the head of the inactive queue instead of the tail.
1231  * This will cause them to be moved to the cache more quickly and
1232  * if not actively re-referenced, freed more quickly.  If we just
1233  * stick these pages at the end of the inactive queue, heavy filesystem
1234  * meta-data accesses can cause an unnecessary paging load on memory bound
1235  * processes.  This optimization causes one-time-use metadata to be
1236  * reused more quickly.
1237  *
1238  * BUT, if we are in a low-memory situation we have no choice but to
1239  * put clean pages on the cache queue.
1240  *
1241  * A number of routines use vm_page_unwire() to guarantee that the page
1242  * will go into either the inactive or active queues, and will NEVER
1243  * be placed in the cache - for example, just after dirtying a page.
1244  * dirty pages in the cache are not allowed.
1245  *
1246  * The page queues must be locked.
1247  * This routine may not block.
1248  */
1249 void
1250 vm_page_unwire(vm_page_t m, int activate)
1251 {
1252 	crit_enter();
1253 	lwkt_gettoken(&vm_token);
1254 	if (m->flags & PG_FICTITIOUS) {
1255 		/* do nothing */
1256 	} else if (m->wire_count <= 0) {
1257 		panic("vm_page_unwire: invalid wire count: %d", m->wire_count);
1258 	} else {
1259 		if (--m->wire_count == 0) {
1260 			--vmstats.v_wire_count;
1261 			if (m->flags & PG_UNMANAGED) {
1262 				;
1263 			} else if (activate) {
1264 				TAILQ_INSERT_TAIL(
1265 				    &vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1266 				m->queue = PQ_ACTIVE;
1267 				vm_page_queues[PQ_ACTIVE].lcnt++;
1268 				vmstats.v_active_count++;
1269 			} else {
1270 				vm_page_flag_clear(m, PG_WINATCFLS);
1271 				TAILQ_INSERT_TAIL(
1272 				    &vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1273 				m->queue = PQ_INACTIVE;
1274 				vm_page_queues[PQ_INACTIVE].lcnt++;
1275 				vmstats.v_inactive_count++;
1276 				++vm_swapcache_inactive_heuristic;
1277 			}
1278 		}
1279 	}
1280 	lwkt_reltoken(&vm_token);
1281 	crit_exit();
1282 }
1283 
1284 
1285 /*
1286  * Move the specified page to the inactive queue.  If the page has
1287  * any associated swap, the swap is deallocated.
1288  *
1289  * Normally athead is 0 resulting in LRU operation.  athead is set
1290  * to 1 if we want this page to be 'as if it were placed in the cache',
1291  * except without unmapping it from the process address space.
1292  *
1293  * This routine may not block.
1294  * The caller must hold vm_token.
1295  */
1296 static __inline void
1297 _vm_page_deactivate(vm_page_t m, int athead)
1298 {
1299 	/*
1300 	 * Ignore if already inactive.
1301 	 */
1302 	if (m->queue == PQ_INACTIVE)
1303 		return;
1304 
1305 	if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1306 		if ((m->queue - m->pc) == PQ_CACHE)
1307 			mycpu->gd_cnt.v_reactivated++;
1308 		vm_page_flag_clear(m, PG_WINATCFLS);
1309 		vm_page_unqueue(m);
1310 		if (athead) {
1311 			TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl,
1312 					  m, pageq);
1313 		} else {
1314 			TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl,
1315 					  m, pageq);
1316 			++vm_swapcache_inactive_heuristic;
1317 		}
1318 		m->queue = PQ_INACTIVE;
1319 		vm_page_queues[PQ_INACTIVE].lcnt++;
1320 		vmstats.v_inactive_count++;
1321 	}
1322 }
1323 
1324 /*
1325  * Attempt to deactivate a page.
1326  *
1327  * No requirements.
1328  */
1329 void
1330 vm_page_deactivate(vm_page_t m)
1331 {
1332 	crit_enter();
1333 	lwkt_gettoken(&vm_token);
1334 	_vm_page_deactivate(m, 0);
1335 	lwkt_reltoken(&vm_token);
1336 	crit_exit();
1337 }
1338 
1339 /*
1340  * Attempt to move a page to PQ_CACHE.
1341  * Returns 0 on failure, 1 on success
1342  *
1343  * No requirements.
1344  */
1345 int
1346 vm_page_try_to_cache(vm_page_t m)
1347 {
1348 	crit_enter();
1349 	lwkt_gettoken(&vm_token);
1350 	if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1351 	    (m->flags & (PG_BUSY|PG_UNMANAGED))) {
1352 		lwkt_reltoken(&vm_token);
1353 		crit_exit();
1354 		return(0);
1355 	}
1356 	vm_page_test_dirty(m);
1357 	if (m->dirty) {
1358 		lwkt_reltoken(&vm_token);
1359 		crit_exit();
1360 		return(0);
1361 	}
1362 	vm_page_cache(m);
1363 	lwkt_reltoken(&vm_token);
1364 	crit_exit();
1365 	return(1);
1366 }
1367 
1368 /*
1369  * Attempt to free the page.  If we cannot free it, we do nothing.
1370  * 1 is returned on success, 0 on failure.
1371  *
1372  * No requirements.
1373  */
1374 int
1375 vm_page_try_to_free(vm_page_t m)
1376 {
1377 	crit_enter();
1378 	lwkt_gettoken(&vm_token);
1379 	if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1380 	    (m->flags & (PG_BUSY|PG_UNMANAGED))) {
1381 		lwkt_reltoken(&vm_token);
1382 		crit_exit();
1383 		return(0);
1384 	}
1385 	vm_page_test_dirty(m);
1386 	if (m->dirty) {
1387 		lwkt_reltoken(&vm_token);
1388 		crit_exit();
1389 		return(0);
1390 	}
1391 	vm_page_busy(m);
1392 	vm_page_protect(m, VM_PROT_NONE);
1393 	vm_page_free(m);
1394 	lwkt_reltoken(&vm_token);
1395 	crit_exit();
1396 	return(1);
1397 }
1398 
1399 /*
1400  * vm_page_cache
1401  *
1402  * Put the specified page onto the page cache queue (if appropriate).
1403  *
1404  * The caller must hold vm_token.
1405  * This routine may not block.
1406  */
1407 void
1408 vm_page_cache(vm_page_t m)
1409 {
1410 	ASSERT_IN_CRIT_SECTION();
1411 	ASSERT_LWKT_TOKEN_HELD(&vm_token);
1412 
1413 	if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy ||
1414 			m->wire_count || m->hold_count) {
1415 		kprintf("vm_page_cache: attempting to cache busy/held page\n");
1416 		return;
1417 	}
1418 
1419 	/*
1420 	 * Already in the cache (and thus not mapped)
1421 	 */
1422 	if ((m->queue - m->pc) == PQ_CACHE) {
1423 		KKASSERT((m->flags & PG_MAPPED) == 0);
1424 		return;
1425 	}
1426 
1427 	/*
1428 	 * Caller is required to test m->dirty, but note that the act of
1429 	 * removing the page from its maps can cause it to become dirty
1430 	 * on an SMP system due to another cpu running in usermode.
1431 	 */
1432 	if (m->dirty) {
1433 		panic("vm_page_cache: caching a dirty page, pindex: %ld",
1434 			(long)m->pindex);
1435 	}
1436 
1437 	/*
1438 	 * Remove all pmaps and indicate that the page is not
1439 	 * writeable or mapped.  Our vm_page_protect() call may
1440 	 * have blocked (especially w/ VM_PROT_NONE), so recheck
1441 	 * everything.
1442 	 */
1443 	vm_page_busy(m);
1444 	vm_page_protect(m, VM_PROT_NONE);
1445 	vm_page_wakeup(m);
1446 	if ((m->flags & (PG_BUSY|PG_UNMANAGED|PG_MAPPED)) || m->busy ||
1447 			m->wire_count || m->hold_count) {
1448 		/* do nothing */
1449 	} else if (m->dirty) {
1450 		vm_page_deactivate(m);
1451 	} else {
1452 		vm_page_unqueue_nowakeup(m);
1453 		m->queue = PQ_CACHE + m->pc;
1454 		vm_page_queues[m->queue].lcnt++;
1455 		TAILQ_INSERT_TAIL(&vm_page_queues[m->queue].pl, m, pageq);
1456 		vmstats.v_cache_count++;
1457 		vm_page_free_wakeup();
1458 	}
1459 }
1460 
1461 /*
1462  * vm_page_dontneed()
1463  *
1464  * Cache, deactivate, or do nothing as appropriate.  This routine
1465  * is typically used by madvise() MADV_DONTNEED.
1466  *
1467  * Generally speaking we want to move the page into the cache so
1468  * it gets reused quickly.  However, this can result in a silly syndrome
1469  * due to the page recycling too quickly.  Small objects will not be
1470  * fully cached.  On the otherhand, if we move the page to the inactive
1471  * queue we wind up with a problem whereby very large objects
1472  * unnecessarily blow away our inactive and cache queues.
1473  *
1474  * The solution is to move the pages based on a fixed weighting.  We
1475  * either leave them alone, deactivate them, or move them to the cache,
1476  * where moving them to the cache has the highest weighting.
1477  * By forcing some pages into other queues we eventually force the
1478  * system to balance the queues, potentially recovering other unrelated
1479  * space from active.  The idea is to not force this to happen too
1480  * often.
1481  *
1482  * No requirements.
1483  */
1484 void
1485 vm_page_dontneed(vm_page_t m)
1486 {
1487 	static int dnweight;
1488 	int dnw;
1489 	int head;
1490 
1491 	dnw = ++dnweight;
1492 
1493 	/*
1494 	 * occassionally leave the page alone
1495 	 */
1496 	crit_enter();
1497 	lwkt_gettoken(&vm_token);
1498 	if ((dnw & 0x01F0) == 0 ||
1499 	    m->queue == PQ_INACTIVE ||
1500 	    m->queue - m->pc == PQ_CACHE
1501 	) {
1502 		if (m->act_count >= ACT_INIT)
1503 			--m->act_count;
1504 		lwkt_reltoken(&vm_token);
1505 		crit_exit();
1506 		return;
1507 	}
1508 
1509 	if (m->dirty == 0)
1510 		vm_page_test_dirty(m);
1511 
1512 	if (m->dirty || (dnw & 0x0070) == 0) {
1513 		/*
1514 		 * Deactivate the page 3 times out of 32.
1515 		 */
1516 		head = 0;
1517 	} else {
1518 		/*
1519 		 * Cache the page 28 times out of every 32.  Note that
1520 		 * the page is deactivated instead of cached, but placed
1521 		 * at the head of the queue instead of the tail.
1522 		 */
1523 		head = 1;
1524 	}
1525 	_vm_page_deactivate(m, head);
1526 	lwkt_reltoken(&vm_token);
1527 	crit_exit();
1528 }
1529 
1530 /*
1531  * Grab a page, blocking if it is busy and allocating a page if necessary.
1532  * A busy page is returned or NULL.
1533  *
1534  * If VM_ALLOC_RETRY is specified VM_ALLOC_NORMAL must also be specified.
1535  * If VM_ALLOC_RETRY is not specified
1536  *
1537  * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is
1538  * always returned if we had blocked.
1539  * This routine will never return NULL if VM_ALLOC_RETRY is set.
1540  * This routine may not be called from an interrupt.
1541  * The returned page may not be entirely valid.
1542  *
1543  * This routine may be called from mainline code without spl protection and
1544  * be guarenteed a busied page associated with the object at the specified
1545  * index.
1546  *
1547  * No requirements.
1548  */
1549 vm_page_t
1550 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
1551 {
1552 	vm_page_t m;
1553 	int generation;
1554 
1555 	KKASSERT(allocflags &
1556 		(VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
1557 	crit_enter();
1558 	lwkt_gettoken(&vm_token);
1559 retrylookup:
1560 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
1561 		if (m->busy || (m->flags & PG_BUSY)) {
1562 			generation = object->generation;
1563 
1564 			while ((object->generation == generation) &&
1565 					(m->busy || (m->flags & PG_BUSY))) {
1566 				vm_page_flag_set(m, PG_WANTED | PG_REFERENCED);
1567 				tsleep(m, 0, "pgrbwt", 0);
1568 				if ((allocflags & VM_ALLOC_RETRY) == 0) {
1569 					m = NULL;
1570 					goto done;
1571 				}
1572 			}
1573 			goto retrylookup;
1574 		} else {
1575 			vm_page_busy(m);
1576 			goto done;
1577 		}
1578 	}
1579 	m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY);
1580 	if (m == NULL) {
1581 		vm_wait(0);
1582 		if ((allocflags & VM_ALLOC_RETRY) == 0)
1583 			goto done;
1584 		goto retrylookup;
1585 	}
1586 done:
1587 	lwkt_reltoken(&vm_token);
1588 	crit_exit();
1589 	return(m);
1590 }
1591 
1592 /*
1593  * Mapping function for valid bits or for dirty bits in
1594  * a page.  May not block.
1595  *
1596  * Inputs are required to range within a page.
1597  *
1598  * No requirements.
1599  * Non blocking.
1600  */
1601 int
1602 vm_page_bits(int base, int size)
1603 {
1604 	int first_bit;
1605 	int last_bit;
1606 
1607 	KASSERT(
1608 	    base + size <= PAGE_SIZE,
1609 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
1610 	);
1611 
1612 	if (size == 0)		/* handle degenerate case */
1613 		return(0);
1614 
1615 	first_bit = base >> DEV_BSHIFT;
1616 	last_bit = (base + size - 1) >> DEV_BSHIFT;
1617 
1618 	return ((2 << last_bit) - (1 << first_bit));
1619 }
1620 
1621 /*
1622  * Sets portions of a page valid and clean.  The arguments are expected
1623  * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
1624  * of any partial chunks touched by the range.  The invalid portion of
1625  * such chunks will be zero'd.
1626  *
1627  * NOTE: When truncating a buffer vnode_pager_setsize() will automatically
1628  *	 align base to DEV_BSIZE so as not to mark clean a partially
1629  *	 truncated device block.  Otherwise the dirty page status might be
1630  *	 lost.
1631  *
1632  * This routine may not block.
1633  *
1634  * (base + size) must be less then or equal to PAGE_SIZE.
1635  */
1636 static void
1637 _vm_page_zero_valid(vm_page_t m, int base, int size)
1638 {
1639 	int frag;
1640 	int endoff;
1641 
1642 	if (size == 0)	/* handle degenerate case */
1643 		return;
1644 
1645 	/*
1646 	 * If the base is not DEV_BSIZE aligned and the valid
1647 	 * bit is clear, we have to zero out a portion of the
1648 	 * first block.
1649 	 */
1650 
1651 	if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
1652 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0
1653 	) {
1654 		pmap_zero_page_area(
1655 		    VM_PAGE_TO_PHYS(m),
1656 		    frag,
1657 		    base - frag
1658 		);
1659 	}
1660 
1661 	/*
1662 	 * If the ending offset is not DEV_BSIZE aligned and the
1663 	 * valid bit is clear, we have to zero out a portion of
1664 	 * the last block.
1665 	 */
1666 
1667 	endoff = base + size;
1668 
1669 	if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
1670 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0
1671 	) {
1672 		pmap_zero_page_area(
1673 		    VM_PAGE_TO_PHYS(m),
1674 		    endoff,
1675 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1))
1676 		);
1677 	}
1678 }
1679 
1680 /*
1681  * Set valid, clear dirty bits.  If validating the entire
1682  * page we can safely clear the pmap modify bit.  We also
1683  * use this opportunity to clear the PG_NOSYNC flag.  If a process
1684  * takes a write fault on a MAP_NOSYNC memory area the flag will
1685  * be set again.
1686  *
1687  * We set valid bits inclusive of any overlap, but we can only
1688  * clear dirty bits for DEV_BSIZE chunks that are fully within
1689  * the range.
1690  *
1691  * Page must be busied?
1692  * No other requirements.
1693  */
1694 void
1695 vm_page_set_valid(vm_page_t m, int base, int size)
1696 {
1697 	_vm_page_zero_valid(m, base, size);
1698 	m->valid |= vm_page_bits(base, size);
1699 }
1700 
1701 
1702 /*
1703  * Set valid bits and clear dirty bits.
1704  *
1705  * NOTE: This function does not clear the pmap modified bit.
1706  *	 Also note that e.g. NFS may use a byte-granular base
1707  *	 and size.
1708  *
1709  * Page must be busied?
1710  * No other requirements.
1711  */
1712 void
1713 vm_page_set_validclean(vm_page_t m, int base, int size)
1714 {
1715 	int pagebits;
1716 
1717 	_vm_page_zero_valid(m, base, size);
1718 	pagebits = vm_page_bits(base, size);
1719 	m->valid |= pagebits;
1720 	m->dirty &= ~pagebits;
1721 	if (base == 0 && size == PAGE_SIZE) {
1722 		/*pmap_clear_modify(m);*/
1723 		vm_page_flag_clear(m, PG_NOSYNC);
1724 	}
1725 }
1726 
1727 /*
1728  * Set valid & dirty.  Used by buwrite()
1729  *
1730  * Page must be busied?
1731  * No other requirements.
1732  */
1733 void
1734 vm_page_set_validdirty(vm_page_t m, int base, int size)
1735 {
1736 	int pagebits;
1737 
1738 	pagebits = vm_page_bits(base, size);
1739 	m->valid |= pagebits;
1740 	m->dirty |= pagebits;
1741 	if (m->object)
1742 		vm_object_set_writeable_dirty(m->object);
1743 }
1744 
1745 /*
1746  * Clear dirty bits.
1747  *
1748  * NOTE: This function does not clear the pmap modified bit.
1749  *	 Also note that e.g. NFS may use a byte-granular base
1750  *	 and size.
1751  *
1752  * Page must be busied?
1753  * No other requirements.
1754  */
1755 void
1756 vm_page_clear_dirty(vm_page_t m, int base, int size)
1757 {
1758 	m->dirty &= ~vm_page_bits(base, size);
1759 	if (base == 0 && size == PAGE_SIZE) {
1760 		/*pmap_clear_modify(m);*/
1761 		vm_page_flag_clear(m, PG_NOSYNC);
1762 	}
1763 }
1764 
1765 /*
1766  * Make the page all-dirty.
1767  *
1768  * Also make sure the related object and vnode reflect the fact that the
1769  * object may now contain a dirty page.
1770  *
1771  * Page must be busied?
1772  * No other requirements.
1773  */
1774 void
1775 vm_page_dirty(vm_page_t m)
1776 {
1777 #ifdef INVARIANTS
1778         int pqtype = m->queue - m->pc;
1779 #endif
1780         KASSERT(pqtype != PQ_CACHE && pqtype != PQ_FREE,
1781                 ("vm_page_dirty: page in free/cache queue!"));
1782 	if (m->dirty != VM_PAGE_BITS_ALL) {
1783 		m->dirty = VM_PAGE_BITS_ALL;
1784 		if (m->object)
1785 			vm_object_set_writeable_dirty(m->object);
1786 	}
1787 }
1788 
1789 /*
1790  * Invalidates DEV_BSIZE'd chunks within a page.  Both the
1791  * valid and dirty bits for the effected areas are cleared.
1792  *
1793  * Page must be busied?
1794  * Does not block.
1795  * No other requirements.
1796  */
1797 void
1798 vm_page_set_invalid(vm_page_t m, int base, int size)
1799 {
1800 	int bits;
1801 
1802 	bits = vm_page_bits(base, size);
1803 	m->valid &= ~bits;
1804 	m->dirty &= ~bits;
1805 	m->object->generation++;
1806 }
1807 
1808 /*
1809  * The kernel assumes that the invalid portions of a page contain
1810  * garbage, but such pages can be mapped into memory by user code.
1811  * When this occurs, we must zero out the non-valid portions of the
1812  * page so user code sees what it expects.
1813  *
1814  * Pages are most often semi-valid when the end of a file is mapped
1815  * into memory and the file's size is not page aligned.
1816  *
1817  * Page must be busied?
1818  * No other requirements.
1819  */
1820 void
1821 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
1822 {
1823 	int b;
1824 	int i;
1825 
1826 	/*
1827 	 * Scan the valid bits looking for invalid sections that
1828 	 * must be zerod.  Invalid sub-DEV_BSIZE'd areas ( where the
1829 	 * valid bit may be set ) have already been zerod by
1830 	 * vm_page_set_validclean().
1831 	 */
1832 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
1833 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
1834 		    (m->valid & (1 << i))
1835 		) {
1836 			if (i > b) {
1837 				pmap_zero_page_area(
1838 				    VM_PAGE_TO_PHYS(m),
1839 				    b << DEV_BSHIFT,
1840 				    (i - b) << DEV_BSHIFT
1841 				);
1842 			}
1843 			b = i + 1;
1844 		}
1845 	}
1846 
1847 	/*
1848 	 * setvalid is TRUE when we can safely set the zero'd areas
1849 	 * as being valid.  We can do this if there are no cache consistency
1850 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
1851 	 */
1852 	if (setvalid)
1853 		m->valid = VM_PAGE_BITS_ALL;
1854 }
1855 
1856 /*
1857  * Is a (partial) page valid?  Note that the case where size == 0
1858  * will return FALSE in the degenerate case where the page is entirely
1859  * invalid, and TRUE otherwise.
1860  *
1861  * Does not block.
1862  * No other requirements.
1863  */
1864 int
1865 vm_page_is_valid(vm_page_t m, int base, int size)
1866 {
1867 	int bits = vm_page_bits(base, size);
1868 
1869 	if (m->valid && ((m->valid & bits) == bits))
1870 		return 1;
1871 	else
1872 		return 0;
1873 }
1874 
1875 /*
1876  * update dirty bits from pmap/mmu.  May not block.
1877  *
1878  * Caller must hold vm_token if non-blocking operation desired.
1879  * No other requirements.
1880  */
1881 void
1882 vm_page_test_dirty(vm_page_t m)
1883 {
1884 	if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) {
1885 		vm_page_dirty(m);
1886 	}
1887 }
1888 
1889 /*
1890  * Register an action, associating it with its vm_page
1891  */
1892 void
1893 vm_page_register_action(vm_page_action_t action, vm_page_event_t event)
1894 {
1895 	struct vm_page_action_list *list;
1896 	int hv;
1897 
1898 	hv = (int)((intptr_t)action->m >> 8) & VMACTION_HMASK;
1899 	list = &action_list[hv];
1900 
1901 	lwkt_gettoken(&vm_token);
1902 	vm_page_flag_set(action->m, PG_ACTIONLIST);
1903 	action->event = event;
1904 	LIST_INSERT_HEAD(list, action, entry);
1905 	lwkt_reltoken(&vm_token);
1906 }
1907 
1908 /*
1909  * Unregister an action, disassociating it from its related vm_page
1910  */
1911 void
1912 vm_page_unregister_action(vm_page_action_t action)
1913 {
1914 	struct vm_page_action_list *list;
1915 	int hv;
1916 
1917 	lwkt_gettoken(&vm_token);
1918 	if (action->event != VMEVENT_NONE) {
1919 		action->event = VMEVENT_NONE;
1920 		LIST_REMOVE(action, entry);
1921 
1922 		hv = (int)((intptr_t)action->m >> 8) & VMACTION_HMASK;
1923 		list = &action_list[hv];
1924 		if (LIST_EMPTY(list))
1925 			vm_page_flag_clear(action->m, PG_ACTIONLIST);
1926 	}
1927 	lwkt_reltoken(&vm_token);
1928 }
1929 
1930 /*
1931  * Issue an event on a VM page.  Corresponding action structures are
1932  * removed from the page's list and called.
1933  *
1934  * If the vm_page has no more pending action events we clear its
1935  * PG_ACTIONLIST flag.
1936  */
1937 void
1938 vm_page_event_internal(vm_page_t m, vm_page_event_t event)
1939 {
1940 	struct vm_page_action_list *list;
1941 	struct vm_page_action *scan;
1942 	struct vm_page_action *next;
1943 	int hv;
1944 	int all;
1945 
1946 	hv = (int)((intptr_t)m >> 8) & VMACTION_HMASK;
1947 	list = &action_list[hv];
1948 	all = 1;
1949 
1950 	lwkt_gettoken(&vm_token);
1951 	LIST_FOREACH_MUTABLE(scan, list, entry, next) {
1952 		if (scan->m == m) {
1953 			if (scan->event == event) {
1954 				scan->event = VMEVENT_NONE;
1955 				LIST_REMOVE(scan, entry);
1956 				scan->func(m, scan);
1957 				/* XXX */
1958 			} else {
1959 				all = 0;
1960 			}
1961 		}
1962 	}
1963 	if (all)
1964 		vm_page_flag_clear(m, PG_ACTIONLIST);
1965 	lwkt_reltoken(&vm_token);
1966 }
1967 
1968 
1969 #include "opt_ddb.h"
1970 #ifdef DDB
1971 #include <sys/kernel.h>
1972 
1973 #include <ddb/ddb.h>
1974 
1975 DB_SHOW_COMMAND(page, vm_page_print_page_info)
1976 {
1977 	db_printf("vmstats.v_free_count: %d\n", vmstats.v_free_count);
1978 	db_printf("vmstats.v_cache_count: %d\n", vmstats.v_cache_count);
1979 	db_printf("vmstats.v_inactive_count: %d\n", vmstats.v_inactive_count);
1980 	db_printf("vmstats.v_active_count: %d\n", vmstats.v_active_count);
1981 	db_printf("vmstats.v_wire_count: %d\n", vmstats.v_wire_count);
1982 	db_printf("vmstats.v_free_reserved: %d\n", vmstats.v_free_reserved);
1983 	db_printf("vmstats.v_free_min: %d\n", vmstats.v_free_min);
1984 	db_printf("vmstats.v_free_target: %d\n", vmstats.v_free_target);
1985 	db_printf("vmstats.v_cache_min: %d\n", vmstats.v_cache_min);
1986 	db_printf("vmstats.v_inactive_target: %d\n", vmstats.v_inactive_target);
1987 }
1988 
1989 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
1990 {
1991 	int i;
1992 	db_printf("PQ_FREE:");
1993 	for(i=0;i<PQ_L2_SIZE;i++) {
1994 		db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt);
1995 	}
1996 	db_printf("\n");
1997 
1998 	db_printf("PQ_CACHE:");
1999 	for(i=0;i<PQ_L2_SIZE;i++) {
2000 		db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt);
2001 	}
2002 	db_printf("\n");
2003 
2004 	db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n",
2005 		vm_page_queues[PQ_ACTIVE].lcnt,
2006 		vm_page_queues[PQ_INACTIVE].lcnt);
2007 }
2008 #endif /* DDB */
2009