xref: /freebsd/sys/vm/vm_page.c (revision 729362425c09cf6b362366aabc6fb547eee8035a)
1 /*
2  * Copyright (c) 1991 Regents of the University of California.
3  * All rights reserved.
4  *
5  * This code is derived from software contributed to Berkeley by
6  * The Mach Operating System project at Carnegie-Mellon University.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 3. All advertising materials mentioning features or use of this software
17  *    must display the following acknowledgement:
18  *	This product includes software developed by the University of
19  *	California, Berkeley and its contributors.
20  * 4. Neither the name of the University nor the names of its contributors
21  *    may be used to endorse or promote products derived from this software
22  *    without specific prior written permission.
23  *
24  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34  * SUCH DAMAGE.
35  *
36  *	from: @(#)vm_page.c	7.4 (Berkeley) 5/7/91
37  * $FreeBSD$
38  */
39 
40 /*
41  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
42  * All rights reserved.
43  *
44  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
45  *
46  * Permission to use, copy, modify and distribute this software and
47  * its documentation is hereby granted, provided that both the copyright
48  * notice and this permission notice appear in all copies of the
49  * software, derivative works or modified versions, and any portions
50  * thereof, and that both notices appear in supporting documentation.
51  *
52  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
53  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
54  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
55  *
56  * Carnegie Mellon requests users of this software to return to
57  *
58  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
59  *  School of Computer Science
60  *  Carnegie Mellon University
61  *  Pittsburgh PA 15213-3890
62  *
63  * any improvements or extensions that they make and grant Carnegie the
64  * rights to redistribute these changes.
65  */
66 
67 /*
68  *			GENERAL RULES ON VM_PAGE MANIPULATION
69  *
70  *	- a pageq mutex is required when adding or removing a page from a
71  *	  page queue (vm_page_queue[]), regardless of other mutexes or the
72  *	  busy state of a page.
73  *
74  *	- a hash chain mutex is required when associating or disassociating
75  *	  a page from the VM PAGE CACHE hash table (vm_page_buckets),
76  *	  regardless of other mutexes or the busy state of a page.
77  *
78  *	- either a hash chain mutex OR a busied page is required in order
79  *	  to modify the page flags.  A hash chain mutex must be obtained in
80  *	  order to busy a page.  A page's flags cannot be modified by a
81  *	  hash chain mutex if the page is marked busy.
82  *
83  *	- The object memq mutex is held when inserting or removing
84  *	  pages from an object (vm_page_insert() or vm_page_remove()).  This
85  *	  is different from the object's main mutex.
86  *
87  *	Generally speaking, you have to be aware of side effects when running
88  *	vm_page ops.  A vm_page_lookup() will return with the hash chain
89  *	locked, whether it was able to lookup the page or not.  vm_page_free(),
90  *	vm_page_cache(), vm_page_activate(), and a number of other routines
91  *	will release the hash chain mutex for you.  Intermediate manipulation
92  *	routines such as vm_page_flag_set() expect the hash chain to be held
93  *	on entry and the hash chain will remain held on return.
94  *
95  *	pageq scanning can only occur with the pageq in question locked.
96  *	We have a known bottleneck with the active queue, but the cache
97  *	and free queues are actually arrays already.
98  */
99 
100 /*
101  *	Resident memory management module.
102  */
103 
104 #include <sys/param.h>
105 #include <sys/systm.h>
106 #include <sys/lock.h>
107 #include <sys/malloc.h>
108 #include <sys/mutex.h>
109 #include <sys/proc.h>
110 #include <sys/vmmeter.h>
111 #include <sys/vnode.h>
112 
113 #include <vm/vm.h>
114 #include <vm/vm_param.h>
115 #include <vm/vm_kern.h>
116 #include <vm/vm_object.h>
117 #include <vm/vm_page.h>
118 #include <vm/vm_pageout.h>
119 #include <vm/vm_pager.h>
120 #include <vm/vm_extern.h>
121 #include <vm/uma.h>
122 #include <vm/uma_int.h>
123 
124 /*
125  *	Associated with page of user-allocatable memory is a
126  *	page structure.
127  */
128 
129 struct mtx vm_page_queue_mtx;
130 struct mtx vm_page_queue_free_mtx;
131 
132 vm_page_t vm_page_array = 0;
133 int vm_page_array_size = 0;
134 long first_page = 0;
135 int vm_page_zero_count = 0;
136 
137 /*
138  *	vm_set_page_size:
139  *
140  *	Sets the page size, perhaps based upon the memory
141  *	size.  Must be called before any use of page-size
142  *	dependent functions.
143  */
144 void
145 vm_set_page_size(void)
146 {
147 	if (cnt.v_page_size == 0)
148 		cnt.v_page_size = PAGE_SIZE;
149 	if (((cnt.v_page_size - 1) & cnt.v_page_size) != 0)
150 		panic("vm_set_page_size: page size not a power of two");
151 }
152 
153 /*
154  *	vm_page_startup:
155  *
156  *	Initializes the resident memory module.
157  *
158  *	Allocates memory for the page cells, and
159  *	for the object/offset-to-page hash table headers.
160  *	Each page cell is initialized and placed on the free list.
161  */
162 vm_offset_t
163 vm_page_startup(vm_offset_t starta, vm_offset_t enda, vm_offset_t vaddr)
164 {
165 	vm_offset_t mapped;
166 	vm_size_t npages;
167 	vm_paddr_t page_range;
168 	vm_paddr_t new_end;
169 	int i;
170 	vm_paddr_t pa;
171 	int nblocks;
172 	vm_paddr_t last_pa;
173 
174 	/* the biggest memory array is the second group of pages */
175 	vm_paddr_t end;
176 	vm_paddr_t biggestsize;
177 	int biggestone;
178 
179 	vm_paddr_t total;
180 	vm_size_t bootpages;
181 
182 	total = 0;
183 	biggestsize = 0;
184 	biggestone = 0;
185 	nblocks = 0;
186 	vaddr = round_page(vaddr);
187 
188 	for (i = 0; phys_avail[i + 1]; i += 2) {
189 		phys_avail[i] = round_page(phys_avail[i]);
190 		phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
191 	}
192 
193 	for (i = 0; phys_avail[i + 1]; i += 2) {
194 		vm_paddr_t size = phys_avail[i + 1] - phys_avail[i];
195 
196 		if (size > biggestsize) {
197 			biggestone = i;
198 			biggestsize = size;
199 		}
200 		++nblocks;
201 		total += size;
202 	}
203 
204 	end = phys_avail[biggestone+1];
205 
206 	/*
207 	 * Initialize the locks.
208 	 */
209 	mtx_init(&vm_page_queue_mtx, "vm page queue mutex", NULL, MTX_DEF);
210 	mtx_init(&vm_page_queue_free_mtx, "vm page queue free mutex", NULL,
211 	   MTX_SPIN);
212 
213 	/*
214 	 * Initialize the queue headers for the free queue, the active queue
215 	 * and the inactive queue.
216 	 */
217 	vm_pageq_init();
218 
219 	/*
220 	 * Allocate memory for use when boot strapping the kernel memory
221 	 * allocator.
222 	 */
223 	bootpages = UMA_BOOT_PAGES * UMA_SLAB_SIZE;
224 	new_end = end - bootpages;
225 	new_end = trunc_page(new_end);
226 	mapped = pmap_map(&vaddr, new_end, end,
227 	    VM_PROT_READ | VM_PROT_WRITE);
228 	bzero((caddr_t) mapped, end - new_end);
229 	uma_startup((caddr_t)mapped);
230 
231 	/*
232 	 * Compute the number of pages of memory that will be available for
233 	 * use (taking into account the overhead of a page structure per
234 	 * page).
235 	 */
236 	first_page = phys_avail[0] / PAGE_SIZE;
237 	page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page;
238 	npages = (total - (page_range * sizeof(struct vm_page)) -
239 	    (end - new_end)) / PAGE_SIZE;
240 	end = new_end;
241 
242 	/*
243 	 * Initialize the mem entry structures now, and put them in the free
244 	 * queue.
245 	 */
246 	new_end = trunc_page(end - page_range * sizeof(struct vm_page));
247 	mapped = pmap_map(&vaddr, new_end, end,
248 	    VM_PROT_READ | VM_PROT_WRITE);
249 	vm_page_array = (vm_page_t) mapped;
250 	phys_avail[biggestone + 1] = new_end;
251 
252 	/*
253 	 * Clear all of the page structures
254 	 */
255 	bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
256 	vm_page_array_size = page_range;
257 
258 	/*
259 	 * Construct the free queue(s) in descending order (by physical
260 	 * address) so that the first 16MB of physical memory is allocated
261 	 * last rather than first.  On large-memory machines, this avoids
262 	 * the exhaustion of low physical memory before isa_dmainit has run.
263 	 */
264 	cnt.v_page_count = 0;
265 	cnt.v_free_count = 0;
266 	for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) {
267 		pa = phys_avail[i];
268 		last_pa = phys_avail[i + 1];
269 		while (pa < last_pa && npages-- > 0) {
270 			vm_pageq_add_new_page(pa);
271 			pa += PAGE_SIZE;
272 		}
273 	}
274 	return (vaddr);
275 }
276 
277 void
278 vm_page_flag_set(vm_page_t m, unsigned short bits)
279 {
280 
281 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
282 	m->flags |= bits;
283 }
284 
285 void
286 vm_page_flag_clear(vm_page_t m, unsigned short bits)
287 {
288 
289 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
290 	m->flags &= ~bits;
291 }
292 
293 void
294 vm_page_busy(vm_page_t m)
295 {
296 	KASSERT((m->flags & PG_BUSY) == 0,
297 	    ("vm_page_busy: page already busy!!!"));
298 	vm_page_flag_set(m, PG_BUSY);
299 }
300 
301 /*
302  *      vm_page_flash:
303  *
304  *      wakeup anyone waiting for the page.
305  */
306 void
307 vm_page_flash(vm_page_t m)
308 {
309 	if (m->flags & PG_WANTED) {
310 		vm_page_flag_clear(m, PG_WANTED);
311 		wakeup(m);
312 	}
313 }
314 
315 /*
316  *      vm_page_wakeup:
317  *
318  *      clear the PG_BUSY flag and wakeup anyone waiting for the
319  *      page.
320  *
321  */
322 void
323 vm_page_wakeup(vm_page_t m)
324 {
325 	KASSERT(m->flags & PG_BUSY, ("vm_page_wakeup: page not busy!!!"));
326 	vm_page_flag_clear(m, PG_BUSY);
327 	vm_page_flash(m);
328 }
329 
330 void
331 vm_page_io_start(vm_page_t m)
332 {
333 
334 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
335 	m->busy++;
336 }
337 
338 void
339 vm_page_io_finish(vm_page_t m)
340 {
341 
342 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
343 	m->busy--;
344 	if (m->busy == 0)
345 		vm_page_flash(m);
346 }
347 
348 /*
349  * Keep page from being freed by the page daemon
350  * much of the same effect as wiring, except much lower
351  * overhead and should be used only for *very* temporary
352  * holding ("wiring").
353  */
354 void
355 vm_page_hold(vm_page_t mem)
356 {
357 
358 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
359         mem->hold_count++;
360 }
361 
362 void
363 vm_page_unhold(vm_page_t mem)
364 {
365 
366 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
367 	--mem->hold_count;
368 	KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!"));
369 	if (mem->hold_count == 0 && mem->queue == PQ_HOLD)
370 		vm_page_free_toq(mem);
371 }
372 
373 /*
374  *	vm_page_copy:
375  *
376  *	Copy one page to another
377  */
378 void
379 vm_page_copy(vm_page_t src_m, vm_page_t dest_m)
380 {
381 	pmap_copy_page(src_m, dest_m);
382 	dest_m->valid = VM_PAGE_BITS_ALL;
383 }
384 
385 /*
386  *	vm_page_free:
387  *
388  *	Free a page
389  *
390  *	The clearing of PG_ZERO is a temporary safety until the code can be
391  *	reviewed to determine that PG_ZERO is being properly cleared on
392  *	write faults or maps.  PG_ZERO was previously cleared in
393  *	vm_page_alloc().
394  */
395 void
396 vm_page_free(vm_page_t m)
397 {
398 	vm_page_flag_clear(m, PG_ZERO);
399 	vm_page_free_toq(m);
400 	vm_page_zero_idle_wakeup();
401 }
402 
403 /*
404  *	vm_page_free_zero:
405  *
406  *	Free a page to the zerod-pages queue
407  */
408 void
409 vm_page_free_zero(vm_page_t m)
410 {
411 	vm_page_flag_set(m, PG_ZERO);
412 	vm_page_free_toq(m);
413 }
414 
415 /*
416  *	vm_page_sleep_if_busy:
417  *
418  *	Sleep and release the page queues lock if PG_BUSY is set or,
419  *	if also_m_busy is TRUE, busy is non-zero.  Returns TRUE if the
420  *	thread slept and the page queues lock was released.
421  *	Otherwise, retains the page queues lock and returns FALSE.
422  */
423 int
424 vm_page_sleep_if_busy(vm_page_t m, int also_m_busy, const char *msg)
425 {
426 	int is_object_locked;
427 
428 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
429 	if ((m->flags & PG_BUSY) || (also_m_busy && m->busy)) {
430 		vm_page_flag_set(m, PG_WANTED | PG_REFERENCED);
431 		/*
432 		 * Remove mtx_owned() after vm_object locking is finished.
433 		 */
434 		if ((is_object_locked = m->object != NULL &&
435 		     mtx_owned(&m->object->mtx)))
436 			mtx_unlock(&m->object->mtx);
437 		msleep(m, &vm_page_queue_mtx, PDROP | PVM, msg, 0);
438 		if (is_object_locked)
439 			mtx_lock(&m->object->mtx);
440 		return (TRUE);
441 	}
442 	return (FALSE);
443 }
444 
445 /*
446  *	vm_page_dirty:
447  *
448  *	make page all dirty
449  */
450 void
451 vm_page_dirty(vm_page_t m)
452 {
453 	KASSERT(m->queue - m->pc != PQ_CACHE,
454 	    ("vm_page_dirty: page in cache!"));
455 	KASSERT(m->queue - m->pc != PQ_FREE,
456 	    ("vm_page_dirty: page is free!"));
457 	m->dirty = VM_PAGE_BITS_ALL;
458 }
459 
460 /*
461  *	vm_page_splay:
462  *
463  *	Implements Sleator and Tarjan's top-down splay algorithm.  Returns
464  *	the vm_page containing the given pindex.  If, however, that
465  *	pindex is not found in the vm_object, returns a vm_page that is
466  *	adjacent to the pindex, coming before or after it.
467  */
468 vm_page_t
469 vm_page_splay(vm_pindex_t pindex, vm_page_t root)
470 {
471 	struct vm_page dummy;
472 	vm_page_t lefttreemax, righttreemin, y;
473 
474 	if (root == NULL)
475 		return (root);
476 	lefttreemax = righttreemin = &dummy;
477 	for (;; root = y) {
478 		if (pindex < root->pindex) {
479 			if ((y = root->left) == NULL)
480 				break;
481 			if (pindex < y->pindex) {
482 				/* Rotate right. */
483 				root->left = y->right;
484 				y->right = root;
485 				root = y;
486 				if ((y = root->left) == NULL)
487 					break;
488 			}
489 			/* Link into the new root's right tree. */
490 			righttreemin->left = root;
491 			righttreemin = root;
492 		} else if (pindex > root->pindex) {
493 			if ((y = root->right) == NULL)
494 				break;
495 			if (pindex > y->pindex) {
496 				/* Rotate left. */
497 				root->right = y->left;
498 				y->left = root;
499 				root = y;
500 				if ((y = root->right) == NULL)
501 					break;
502 			}
503 			/* Link into the new root's left tree. */
504 			lefttreemax->right = root;
505 			lefttreemax = root;
506 		} else
507 			break;
508 	}
509 	/* Assemble the new root. */
510 	lefttreemax->right = root->left;
511 	righttreemin->left = root->right;
512 	root->left = dummy.right;
513 	root->right = dummy.left;
514 	return (root);
515 }
516 
517 /*
518  *	vm_page_insert:		[ internal use only ]
519  *
520  *	Inserts the given mem entry into the object and object list.
521  *
522  *	The pagetables are not updated but will presumably fault the page
523  *	in if necessary, or if a kernel page the caller will at some point
524  *	enter the page into the kernel's pmap.  We are not allowed to block
525  *	here so we *can't* do this anyway.
526  *
527  *	The object and page must be locked, and must be splhigh.
528  *	This routine may not block.
529  */
530 void
531 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
532 {
533 	vm_page_t root;
534 
535 	if (m->object != NULL)
536 		panic("vm_page_insert: already inserted");
537 
538 	/*
539 	 * Record the object/offset pair in this page
540 	 */
541 	m->object = object;
542 	m->pindex = pindex;
543 
544 	mtx_assert(object == kmem_object ? &object->mtx : &Giant, MA_OWNED);
545 	/*
546 	 * Now link into the object's ordered list of backed pages.
547 	 */
548 	root = object->root;
549 	if (root == NULL) {
550 		m->left = NULL;
551 		m->right = NULL;
552 		TAILQ_INSERT_TAIL(&object->memq, m, listq);
553 	} else {
554 		root = vm_page_splay(pindex, root);
555 		if (pindex < root->pindex) {
556 			m->left = root->left;
557 			m->right = root;
558 			root->left = NULL;
559 			TAILQ_INSERT_BEFORE(root, m, listq);
560 		} else {
561 			m->right = root->right;
562 			m->left = root;
563 			root->right = NULL;
564 			TAILQ_INSERT_AFTER(&object->memq, root, m, listq);
565 		}
566 	}
567 	object->root = m;
568 	object->generation++;
569 
570 	/*
571 	 * show that the object has one more resident page.
572 	 */
573 	object->resident_page_count++;
574 
575 	/*
576 	 * Since we are inserting a new and possibly dirty page,
577 	 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags.
578 	 */
579 	if (m->flags & PG_WRITEABLE)
580 		vm_object_set_writeable_dirty(object);
581 }
582 
583 /*
584  *	vm_page_remove:
585  *				NOTE: used by device pager as well -wfj
586  *
587  *	Removes the given mem entry from the object/offset-page
588  *	table and the object page list, but do not invalidate/terminate
589  *	the backing store.
590  *
591  *	The object and page must be locked, and at splhigh.
592  *	The underlying pmap entry (if any) is NOT removed here.
593  *	This routine may not block.
594  */
595 void
596 vm_page_remove(vm_page_t m)
597 {
598 	vm_object_t object;
599 	vm_page_t root;
600 
601 	GIANT_REQUIRED;
602 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
603 	if (m->object == NULL)
604 		return;
605 
606 	if ((m->flags & PG_BUSY) == 0) {
607 		panic("vm_page_remove: page not busy");
608 	}
609 
610 	/*
611 	 * Basically destroy the page.
612 	 */
613 	vm_page_wakeup(m);
614 
615 	object = m->object;
616 
617 	/*
618 	 * Now remove from the object's list of backed pages.
619 	 */
620 	if (m != object->root)
621 		vm_page_splay(m->pindex, object->root);
622 	if (m->left == NULL)
623 		root = m->right;
624 	else {
625 		root = vm_page_splay(m->pindex, m->left);
626 		root->right = m->right;
627 	}
628 	object->root = root;
629 	TAILQ_REMOVE(&object->memq, m, listq);
630 
631 	/*
632 	 * And show that the object has one fewer resident page.
633 	 */
634 	object->resident_page_count--;
635 	object->generation++;
636 
637 	m->object = NULL;
638 }
639 
640 /*
641  *	vm_page_lookup:
642  *
643  *	Returns the page associated with the object/offset
644  *	pair specified; if none is found, NULL is returned.
645  *
646  *	The object must be locked.
647  *	This routine may not block.
648  *	This is a critical path routine
649  */
650 vm_page_t
651 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
652 {
653 	vm_page_t m;
654 
655 	mtx_assert(object == kmem_object ? &object->mtx : &Giant, MA_OWNED);
656 	m = vm_page_splay(pindex, object->root);
657 	if ((object->root = m) != NULL && m->pindex != pindex)
658 		m = NULL;
659 	return (m);
660 }
661 
662 /*
663  *	vm_page_rename:
664  *
665  *	Move the given memory entry from its
666  *	current object to the specified target object/offset.
667  *
668  *	The object must be locked.
669  *	This routine may not block.
670  *
671  *	Note: this routine will raise itself to splvm(), the caller need not.
672  *
673  *	Note: swap associated with the page must be invalidated by the move.  We
674  *	      have to do this for several reasons:  (1) we aren't freeing the
675  *	      page, (2) we are dirtying the page, (3) the VM system is probably
676  *	      moving the page from object A to B, and will then later move
677  *	      the backing store from A to B and we can't have a conflict.
678  *
679  *	Note: we *always* dirty the page.  It is necessary both for the
680  *	      fact that we moved it, and because we may be invalidating
681  *	      swap.  If the page is on the cache, we have to deactivate it
682  *	      or vm_page_dirty() will panic.  Dirty pages are not allowed
683  *	      on the cache.
684  */
685 void
686 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
687 {
688 	int s;
689 
690 	s = splvm();
691 	vm_page_remove(m);
692 	vm_page_insert(m, new_object, new_pindex);
693 	if (m->queue - m->pc == PQ_CACHE)
694 		vm_page_deactivate(m);
695 	vm_page_dirty(m);
696 	splx(s);
697 }
698 
699 /*
700  *	vm_page_select_cache:
701  *
702  *	Find a page on the cache queue with color optimization.  As pages
703  *	might be found, but not applicable, they are deactivated.  This
704  *	keeps us from using potentially busy cached pages.
705  *
706  *	This routine must be called at splvm().
707  *	This routine may not block.
708  */
709 static vm_page_t
710 vm_page_select_cache(vm_pindex_t color)
711 {
712 	vm_page_t m;
713 
714 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
715 	while (TRUE) {
716 		m = vm_pageq_find(PQ_CACHE, color & PQ_L2_MASK, FALSE);
717 		if (m && ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy ||
718 			       m->hold_count || m->wire_count)) {
719 			vm_page_deactivate(m);
720 			continue;
721 		}
722 		return m;
723 	}
724 }
725 
726 /*
727  *	vm_page_select_free:
728  *
729  *	Find a free or zero page, with specified preference.
730  *
731  *	This routine must be called at splvm().
732  *	This routine may not block.
733  */
734 static __inline vm_page_t
735 vm_page_select_free(vm_pindex_t color, boolean_t prefer_zero)
736 {
737 	vm_page_t m;
738 
739 	m = vm_pageq_find(PQ_FREE, color & PQ_L2_MASK, prefer_zero);
740 	return (m);
741 }
742 
743 /*
744  *	vm_page_alloc:
745  *
746  *	Allocate and return a memory cell associated
747  *	with this VM object/offset pair.
748  *
749  *	page_req classes:
750  *	VM_ALLOC_NORMAL		normal process request
751  *	VM_ALLOC_SYSTEM		system *really* needs a page
752  *	VM_ALLOC_INTERRUPT	interrupt time request
753  *	VM_ALLOC_ZERO		zero page
754  *
755  *	This routine may not block.
756  *
757  *	Additional special handling is required when called from an
758  *	interrupt (VM_ALLOC_INTERRUPT).  We are not allowed to mess with
759  *	the page cache in this case.
760  */
761 vm_page_t
762 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
763 {
764 	vm_page_t m = NULL;
765 	vm_pindex_t color;
766 	int flags, page_req, s;
767 
768 	page_req = req & VM_ALLOC_CLASS_MASK;
769 
770 	if ((req & VM_ALLOC_NOOBJ) == 0) {
771 		KASSERT(object != NULL,
772 		    ("vm_page_alloc: NULL object."));
773 		mtx_assert(object == kmem_object ? &object->mtx : &Giant,
774 		    MA_OWNED);
775 		KASSERT(!vm_page_lookup(object, pindex),
776 		    ("vm_page_alloc: page already allocated"));
777 		color = pindex + object->pg_color;
778 	} else
779 		color = pindex;
780 
781 	/*
782 	 * The pager is allowed to eat deeper into the free page list.
783 	 */
784 	if ((curproc == pageproc) && (page_req != VM_ALLOC_INTERRUPT)) {
785 		page_req = VM_ALLOC_SYSTEM;
786 	};
787 
788 	s = splvm();
789 loop:
790 	mtx_lock_spin(&vm_page_queue_free_mtx);
791 	if (cnt.v_free_count > cnt.v_free_reserved ||
792 	    (page_req == VM_ALLOC_SYSTEM &&
793 	     cnt.v_cache_count == 0 &&
794 	     cnt.v_free_count > cnt.v_interrupt_free_min) ||
795 	    (page_req == VM_ALLOC_INTERRUPT && cnt.v_free_count > 0)) {
796 		/*
797 		 * Allocate from the free queue if the number of free pages
798 		 * exceeds the minimum for the request class.
799 		 */
800 		m = vm_page_select_free(color, (req & VM_ALLOC_ZERO) != 0);
801 	} else if (page_req != VM_ALLOC_INTERRUPT) {
802 		mtx_unlock_spin(&vm_page_queue_free_mtx);
803 		/*
804 		 * Allocatable from cache (non-interrupt only).  On success,
805 		 * we must free the page and try again, thus ensuring that
806 		 * cnt.v_*_free_min counters are replenished.
807 		 */
808 		vm_page_lock_queues();
809 		if ((m = vm_page_select_cache(color)) == NULL) {
810 			vm_page_unlock_queues();
811 			splx(s);
812 #if defined(DIAGNOSTIC)
813 			if (cnt.v_cache_count > 0)
814 				printf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", cnt.v_cache_count);
815 #endif
816 			atomic_add_int(&vm_pageout_deficit, 1);
817 			pagedaemon_wakeup();
818 			return (NULL);
819 		}
820 		KASSERT(m->dirty == 0, ("Found dirty cache page %p", m));
821 		vm_page_busy(m);
822 		pmap_remove_all(m);
823 		vm_page_free(m);
824 		vm_page_unlock_queues();
825 		goto loop;
826 	} else {
827 		/*
828 		 * Not allocatable from cache from interrupt, give up.
829 		 */
830 		mtx_unlock_spin(&vm_page_queue_free_mtx);
831 		splx(s);
832 		atomic_add_int(&vm_pageout_deficit, 1);
833 		pagedaemon_wakeup();
834 		return (NULL);
835 	}
836 
837 	/*
838 	 *  At this point we had better have found a good page.
839 	 */
840 
841 	KASSERT(
842 	    m != NULL,
843 	    ("vm_page_alloc(): missing page on free queue\n")
844 	);
845 
846 	/*
847 	 * Remove from free queue
848 	 */
849 
850 	vm_pageq_remove_nowakeup(m);
851 
852 	/*
853 	 * Initialize structure.  Only the PG_ZERO flag is inherited.
854 	 */
855 	flags = PG_BUSY;
856 	if (m->flags & PG_ZERO) {
857 		vm_page_zero_count--;
858 		if (req & VM_ALLOC_ZERO)
859 			flags = PG_ZERO | PG_BUSY;
860 	}
861 	m->flags = flags;
862 	if (req & VM_ALLOC_WIRED) {
863 		atomic_add_int(&cnt.v_wire_count, 1);
864 		m->wire_count = 1;
865 	} else
866 		m->wire_count = 0;
867 	m->hold_count = 0;
868 	m->act_count = 0;
869 	m->busy = 0;
870 	m->valid = 0;
871 	KASSERT(m->dirty == 0, ("vm_page_alloc: free/cache page %p was dirty", m));
872 	mtx_unlock_spin(&vm_page_queue_free_mtx);
873 
874 	/*
875 	 * vm_page_insert() is safe prior to the splx().  Note also that
876 	 * inserting a page here does not insert it into the pmap (which
877 	 * could cause us to block allocating memory).  We cannot block
878 	 * anywhere.
879 	 */
880 	if ((req & VM_ALLOC_NOOBJ) == 0)
881 		vm_page_insert(m, object, pindex);
882 
883 	/*
884 	 * Don't wakeup too often - wakeup the pageout daemon when
885 	 * we would be nearly out of memory.
886 	 */
887 	if (vm_paging_needed())
888 		pagedaemon_wakeup();
889 
890 	splx(s);
891 	return (m);
892 }
893 
894 /*
895  *	vm_wait:	(also see VM_WAIT macro)
896  *
897  *	Block until free pages are available for allocation
898  *	- Called in various places before memory allocations.
899  */
900 void
901 vm_wait(void)
902 {
903 	int s;
904 
905 	s = splvm();
906 	vm_page_lock_queues();
907 	if (curproc == pageproc) {
908 		vm_pageout_pages_needed = 1;
909 		msleep(&vm_pageout_pages_needed, &vm_page_queue_mtx,
910 		    PDROP | PSWP, "VMWait", 0);
911 	} else {
912 		if (!vm_pages_needed) {
913 			vm_pages_needed = 1;
914 			wakeup(&vm_pages_needed);
915 		}
916 		msleep(&cnt.v_free_count, &vm_page_queue_mtx, PDROP | PVM,
917 		    "vmwait", 0);
918 	}
919 	splx(s);
920 }
921 
922 /*
923  *	vm_waitpfault:	(also see VM_WAITPFAULT macro)
924  *
925  *	Block until free pages are available for allocation
926  *	- Called only in vm_fault so that processes page faulting
927  *	  can be easily tracked.
928  *	- Sleeps at a lower priority than vm_wait() so that vm_wait()ing
929  *	  processes will be able to grab memory first.  Do not change
930  *	  this balance without careful testing first.
931  */
932 void
933 vm_waitpfault(void)
934 {
935 	int s;
936 
937 	s = splvm();
938 	vm_page_lock_queues();
939 	if (!vm_pages_needed) {
940 		vm_pages_needed = 1;
941 		wakeup(&vm_pages_needed);
942 	}
943 	msleep(&cnt.v_free_count, &vm_page_queue_mtx, PDROP | PUSER,
944 	    "pfault", 0);
945 	splx(s);
946 }
947 
948 /*
949  *	vm_page_activate:
950  *
951  *	Put the specified page on the active list (if appropriate).
952  *	Ensure that act_count is at least ACT_INIT but do not otherwise
953  *	mess with it.
954  *
955  *	The page queues must be locked.
956  *	This routine may not block.
957  */
958 void
959 vm_page_activate(vm_page_t m)
960 {
961 	int s;
962 
963 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
964 	s = splvm();
965 	if (m->queue != PQ_ACTIVE) {
966 		if ((m->queue - m->pc) == PQ_CACHE)
967 			cnt.v_reactivated++;
968 		vm_pageq_remove(m);
969 		if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
970 			if (m->act_count < ACT_INIT)
971 				m->act_count = ACT_INIT;
972 			vm_pageq_enqueue(PQ_ACTIVE, m);
973 		}
974 	} else {
975 		if (m->act_count < ACT_INIT)
976 			m->act_count = ACT_INIT;
977 	}
978 	splx(s);
979 }
980 
981 /*
982  *	vm_page_free_wakeup:
983  *
984  *	Helper routine for vm_page_free_toq() and vm_page_cache().  This
985  *	routine is called when a page has been added to the cache or free
986  *	queues.
987  *
988  *	This routine may not block.
989  *	This routine must be called at splvm()
990  */
991 static __inline void
992 vm_page_free_wakeup(void)
993 {
994 
995 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
996 	/*
997 	 * if pageout daemon needs pages, then tell it that there are
998 	 * some free.
999 	 */
1000 	if (vm_pageout_pages_needed &&
1001 	    cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) {
1002 		wakeup(&vm_pageout_pages_needed);
1003 		vm_pageout_pages_needed = 0;
1004 	}
1005 	/*
1006 	 * wakeup processes that are waiting on memory if we hit a
1007 	 * high water mark. And wakeup scheduler process if we have
1008 	 * lots of memory. this process will swapin processes.
1009 	 */
1010 	if (vm_pages_needed && !vm_page_count_min()) {
1011 		vm_pages_needed = 0;
1012 		wakeup(&cnt.v_free_count);
1013 	}
1014 }
1015 
1016 /*
1017  *	vm_page_free_toq:
1018  *
1019  *	Returns the given page to the PQ_FREE list,
1020  *	disassociating it with any VM object.
1021  *
1022  *	Object and page must be locked prior to entry.
1023  *	This routine may not block.
1024  */
1025 
1026 void
1027 vm_page_free_toq(vm_page_t m)
1028 {
1029 	int s;
1030 	struct vpgqueues *pq;
1031 	vm_object_t object = m->object;
1032 
1033 	GIANT_REQUIRED;
1034 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1035 	s = splvm();
1036 	cnt.v_tfree++;
1037 
1038 	if (m->busy || ((m->queue - m->pc) == PQ_FREE)) {
1039 		printf(
1040 		"vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n",
1041 		    (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0,
1042 		    m->hold_count);
1043 		if ((m->queue - m->pc) == PQ_FREE)
1044 			panic("vm_page_free: freeing free page");
1045 		else
1046 			panic("vm_page_free: freeing busy page");
1047 	}
1048 
1049 	/*
1050 	 * unqueue, then remove page.  Note that we cannot destroy
1051 	 * the page here because we do not want to call the pager's
1052 	 * callback routine until after we've put the page on the
1053 	 * appropriate free queue.
1054 	 */
1055 	vm_pageq_remove_nowakeup(m);
1056 	vm_page_remove(m);
1057 
1058 	/*
1059 	 * If fictitious remove object association and
1060 	 * return, otherwise delay object association removal.
1061 	 */
1062 	if ((m->flags & PG_FICTITIOUS) != 0) {
1063 		splx(s);
1064 		return;
1065 	}
1066 
1067 	m->valid = 0;
1068 	vm_page_undirty(m);
1069 
1070 	if (m->wire_count != 0) {
1071 		if (m->wire_count > 1) {
1072 			panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx",
1073 				m->wire_count, (long)m->pindex);
1074 		}
1075 		panic("vm_page_free: freeing wired page\n");
1076 	}
1077 
1078 	/*
1079 	 * If we've exhausted the object's resident pages we want to free
1080 	 * it up.
1081 	 */
1082 	if (object &&
1083 	    (object->type == OBJT_VNODE) &&
1084 	    ((object->flags & OBJ_DEAD) == 0)
1085 	) {
1086 		struct vnode *vp = (struct vnode *)object->handle;
1087 
1088 		if (vp) {
1089 			VI_LOCK(vp);
1090 			if (VSHOULDFREE(vp))
1091 				vfree(vp);
1092 			VI_UNLOCK(vp);
1093 		}
1094 	}
1095 
1096 	/*
1097 	 * Clear the UNMANAGED flag when freeing an unmanaged page.
1098 	 */
1099 	if (m->flags & PG_UNMANAGED) {
1100 		m->flags &= ~PG_UNMANAGED;
1101 	} else {
1102 #ifdef __alpha__
1103 		pmap_page_is_free(m);
1104 #endif
1105 	}
1106 
1107 	if (m->hold_count != 0) {
1108 		m->flags &= ~PG_ZERO;
1109 		m->queue = PQ_HOLD;
1110 	} else
1111 		m->queue = PQ_FREE + m->pc;
1112 	pq = &vm_page_queues[m->queue];
1113 	mtx_lock_spin(&vm_page_queue_free_mtx);
1114 	pq->lcnt++;
1115 	++(*pq->cnt);
1116 
1117 	/*
1118 	 * Put zero'd pages on the end ( where we look for zero'd pages
1119 	 * first ) and non-zerod pages at the head.
1120 	 */
1121 	if (m->flags & PG_ZERO) {
1122 		TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
1123 		++vm_page_zero_count;
1124 	} else {
1125 		TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
1126 	}
1127 	mtx_unlock_spin(&vm_page_queue_free_mtx);
1128 	vm_page_free_wakeup();
1129 	splx(s);
1130 }
1131 
1132 /*
1133  *	vm_page_unmanage:
1134  *
1135  * 	Prevent PV management from being done on the page.  The page is
1136  *	removed from the paging queues as if it were wired, and as a
1137  *	consequence of no longer being managed the pageout daemon will not
1138  *	touch it (since there is no way to locate the pte mappings for the
1139  *	page).  madvise() calls that mess with the pmap will also no longer
1140  *	operate on the page.
1141  *
1142  *	Beyond that the page is still reasonably 'normal'.  Freeing the page
1143  *	will clear the flag.
1144  *
1145  *	This routine is used by OBJT_PHYS objects - objects using unswappable
1146  *	physical memory as backing store rather then swap-backed memory and
1147  *	will eventually be extended to support 4MB unmanaged physical
1148  *	mappings.
1149  */
1150 void
1151 vm_page_unmanage(vm_page_t m)
1152 {
1153 	int s;
1154 
1155 	s = splvm();
1156 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1157 	if ((m->flags & PG_UNMANAGED) == 0) {
1158 		if (m->wire_count == 0)
1159 			vm_pageq_remove(m);
1160 	}
1161 	vm_page_flag_set(m, PG_UNMANAGED);
1162 	splx(s);
1163 }
1164 
1165 /*
1166  *	vm_page_wire:
1167  *
1168  *	Mark this page as wired down by yet
1169  *	another map, removing it from paging queues
1170  *	as necessary.
1171  *
1172  *	The page queues must be locked.
1173  *	This routine may not block.
1174  */
1175 void
1176 vm_page_wire(vm_page_t m)
1177 {
1178 	int s;
1179 
1180 	/*
1181 	 * Only bump the wire statistics if the page is not already wired,
1182 	 * and only unqueue the page if it is on some queue (if it is unmanaged
1183 	 * it is already off the queues).
1184 	 */
1185 	s = splvm();
1186 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1187 	if (m->wire_count == 0) {
1188 		if ((m->flags & PG_UNMANAGED) == 0)
1189 			vm_pageq_remove(m);
1190 		atomic_add_int(&cnt.v_wire_count, 1);
1191 	}
1192 	m->wire_count++;
1193 	KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m));
1194 	splx(s);
1195 }
1196 
1197 /*
1198  *	vm_page_unwire:
1199  *
1200  *	Release one wiring of this page, potentially
1201  *	enabling it to be paged again.
1202  *
1203  *	Many pages placed on the inactive queue should actually go
1204  *	into the cache, but it is difficult to figure out which.  What
1205  *	we do instead, if the inactive target is well met, is to put
1206  *	clean pages at the head of the inactive queue instead of the tail.
1207  *	This will cause them to be moved to the cache more quickly and
1208  *	if not actively re-referenced, freed more quickly.  If we just
1209  *	stick these pages at the end of the inactive queue, heavy filesystem
1210  *	meta-data accesses can cause an unnecessary paging load on memory bound
1211  *	processes.  This optimization causes one-time-use metadata to be
1212  *	reused more quickly.
1213  *
1214  *	BUT, if we are in a low-memory situation we have no choice but to
1215  *	put clean pages on the cache queue.
1216  *
1217  *	A number of routines use vm_page_unwire() to guarantee that the page
1218  *	will go into either the inactive or active queues, and will NEVER
1219  *	be placed in the cache - for example, just after dirtying a page.
1220  *	dirty pages in the cache are not allowed.
1221  *
1222  *	The page queues must be locked.
1223  *	This routine may not block.
1224  */
1225 void
1226 vm_page_unwire(vm_page_t m, int activate)
1227 {
1228 	int s;
1229 
1230 	s = splvm();
1231 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1232 	if (m->wire_count > 0) {
1233 		m->wire_count--;
1234 		if (m->wire_count == 0) {
1235 			atomic_subtract_int(&cnt.v_wire_count, 1);
1236 			if (m->flags & PG_UNMANAGED) {
1237 				;
1238 			} else if (activate)
1239 				vm_pageq_enqueue(PQ_ACTIVE, m);
1240 			else {
1241 				vm_page_flag_clear(m, PG_WINATCFLS);
1242 				vm_pageq_enqueue(PQ_INACTIVE, m);
1243 			}
1244 		}
1245 	} else {
1246 		panic("vm_page_unwire: invalid wire count: %d\n", m->wire_count);
1247 	}
1248 	splx(s);
1249 }
1250 
1251 
1252 /*
1253  * Move the specified page to the inactive queue.  If the page has
1254  * any associated swap, the swap is deallocated.
1255  *
1256  * Normally athead is 0 resulting in LRU operation.  athead is set
1257  * to 1 if we want this page to be 'as if it were placed in the cache',
1258  * except without unmapping it from the process address space.
1259  *
1260  * This routine may not block.
1261  */
1262 static __inline void
1263 _vm_page_deactivate(vm_page_t m, int athead)
1264 {
1265 	int s;
1266 
1267 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1268 	/*
1269 	 * Ignore if already inactive.
1270 	 */
1271 	if (m->queue == PQ_INACTIVE)
1272 		return;
1273 
1274 	s = splvm();
1275 	if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
1276 		if ((m->queue - m->pc) == PQ_CACHE)
1277 			cnt.v_reactivated++;
1278 		vm_page_flag_clear(m, PG_WINATCFLS);
1279 		vm_pageq_remove(m);
1280 		if (athead)
1281 			TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1282 		else
1283 			TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1284 		m->queue = PQ_INACTIVE;
1285 		vm_page_queues[PQ_INACTIVE].lcnt++;
1286 		cnt.v_inactive_count++;
1287 	}
1288 	splx(s);
1289 }
1290 
1291 void
1292 vm_page_deactivate(vm_page_t m)
1293 {
1294     _vm_page_deactivate(m, 0);
1295 }
1296 
1297 /*
1298  * vm_page_try_to_cache:
1299  *
1300  * Returns 0 on failure, 1 on success
1301  */
1302 int
1303 vm_page_try_to_cache(vm_page_t m)
1304 {
1305 
1306 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1307 	if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1308 	    (m->flags & (PG_BUSY|PG_UNMANAGED))) {
1309 		return (0);
1310 	}
1311 	vm_page_test_dirty(m);
1312 	if (m->dirty)
1313 		return (0);
1314 	vm_page_cache(m);
1315 	return (1);
1316 }
1317 
1318 /*
1319  * vm_page_try_to_free()
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 
1328 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1329 	if (m->dirty || m->hold_count || m->busy || m->wire_count ||
1330 	    (m->flags & (PG_BUSY|PG_UNMANAGED))) {
1331 		return (0);
1332 	}
1333 	vm_page_test_dirty(m);
1334 	if (m->dirty)
1335 		return (0);
1336 	vm_page_busy(m);
1337 	pmap_remove_all(m);
1338 	vm_page_free(m);
1339 	return (1);
1340 }
1341 
1342 /*
1343  * vm_page_cache
1344  *
1345  * Put the specified page onto the page cache queue (if appropriate).
1346  *
1347  * This routine may not block.
1348  */
1349 void
1350 vm_page_cache(vm_page_t m)
1351 {
1352 	int s;
1353 
1354 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1355 	if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || m->wire_count) {
1356 		printf("vm_page_cache: attempting to cache busy page\n");
1357 		return;
1358 	}
1359 	if ((m->queue - m->pc) == PQ_CACHE)
1360 		return;
1361 
1362 	/*
1363 	 * Remove all pmaps and indicate that the page is not
1364 	 * writeable or mapped.
1365 	 */
1366 	pmap_remove_all(m);
1367 	if (m->dirty != 0) {
1368 		panic("vm_page_cache: caching a dirty page, pindex: %ld",
1369 			(long)m->pindex);
1370 	}
1371 	s = splvm();
1372 	vm_pageq_remove_nowakeup(m);
1373 	vm_pageq_enqueue(PQ_CACHE + m->pc, m);
1374 	vm_page_free_wakeup();
1375 	splx(s);
1376 }
1377 
1378 /*
1379  * vm_page_dontneed
1380  *
1381  *	Cache, deactivate, or do nothing as appropriate.  This routine
1382  *	is typically used by madvise() MADV_DONTNEED.
1383  *
1384  *	Generally speaking we want to move the page into the cache so
1385  *	it gets reused quickly.  However, this can result in a silly syndrome
1386  *	due to the page recycling too quickly.  Small objects will not be
1387  *	fully cached.  On the otherhand, if we move the page to the inactive
1388  *	queue we wind up with a problem whereby very large objects
1389  *	unnecessarily blow away our inactive and cache queues.
1390  *
1391  *	The solution is to move the pages based on a fixed weighting.  We
1392  *	either leave them alone, deactivate them, or move them to the cache,
1393  *	where moving them to the cache has the highest weighting.
1394  *	By forcing some pages into other queues we eventually force the
1395  *	system to balance the queues, potentially recovering other unrelated
1396  *	space from active.  The idea is to not force this to happen too
1397  *	often.
1398  */
1399 void
1400 vm_page_dontneed(vm_page_t m)
1401 {
1402 	static int dnweight;
1403 	int dnw;
1404 	int head;
1405 
1406 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1407 	dnw = ++dnweight;
1408 
1409 	/*
1410 	 * occassionally leave the page alone
1411 	 */
1412 	if ((dnw & 0x01F0) == 0 ||
1413 	    m->queue == PQ_INACTIVE ||
1414 	    m->queue - m->pc == PQ_CACHE
1415 	) {
1416 		if (m->act_count >= ACT_INIT)
1417 			--m->act_count;
1418 		return;
1419 	}
1420 
1421 	if (m->dirty == 0)
1422 		vm_page_test_dirty(m);
1423 
1424 	if (m->dirty || (dnw & 0x0070) == 0) {
1425 		/*
1426 		 * Deactivate the page 3 times out of 32.
1427 		 */
1428 		head = 0;
1429 	} else {
1430 		/*
1431 		 * Cache the page 28 times out of every 32.  Note that
1432 		 * the page is deactivated instead of cached, but placed
1433 		 * at the head of the queue instead of the tail.
1434 		 */
1435 		head = 1;
1436 	}
1437 	_vm_page_deactivate(m, head);
1438 }
1439 
1440 /*
1441  * Grab a page, waiting until we are waken up due to the page
1442  * changing state.  We keep on waiting, if the page continues
1443  * to be in the object.  If the page doesn't exist, allocate it.
1444  *
1445  * This routine may block.
1446  */
1447 vm_page_t
1448 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
1449 {
1450 	vm_page_t m;
1451 	int s, generation;
1452 
1453 	GIANT_REQUIRED;
1454 retrylookup:
1455 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
1456 		vm_page_lock_queues();
1457 		if (m->busy || (m->flags & PG_BUSY)) {
1458 			generation = object->generation;
1459 
1460 			s = splvm();
1461 			while ((object->generation == generation) &&
1462 					(m->busy || (m->flags & PG_BUSY))) {
1463 				vm_page_flag_set(m, PG_WANTED | PG_REFERENCED);
1464 				msleep(m, &vm_page_queue_mtx, PVM, "pgrbwt", 0);
1465 				if ((allocflags & VM_ALLOC_RETRY) == 0) {
1466 					vm_page_unlock_queues();
1467 					splx(s);
1468 					return NULL;
1469 				}
1470 			}
1471 			vm_page_unlock_queues();
1472 			splx(s);
1473 			goto retrylookup;
1474 		} else {
1475 			if (allocflags & VM_ALLOC_WIRED)
1476 				vm_page_wire(m);
1477 			vm_page_busy(m);
1478 			vm_page_unlock_queues();
1479 			return m;
1480 		}
1481 	}
1482 
1483 	m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY);
1484 	if (m == NULL) {
1485 		VM_WAIT;
1486 		if ((allocflags & VM_ALLOC_RETRY) == 0)
1487 			return NULL;
1488 		goto retrylookup;
1489 	}
1490 
1491 	return m;
1492 }
1493 
1494 /*
1495  * Mapping function for valid bits or for dirty bits in
1496  * a page.  May not block.
1497  *
1498  * Inputs are required to range within a page.
1499  */
1500 __inline int
1501 vm_page_bits(int base, int size)
1502 {
1503 	int first_bit;
1504 	int last_bit;
1505 
1506 	KASSERT(
1507 	    base + size <= PAGE_SIZE,
1508 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
1509 	);
1510 
1511 	if (size == 0)		/* handle degenerate case */
1512 		return (0);
1513 
1514 	first_bit = base >> DEV_BSHIFT;
1515 	last_bit = (base + size - 1) >> DEV_BSHIFT;
1516 
1517 	return ((2 << last_bit) - (1 << first_bit));
1518 }
1519 
1520 /*
1521  *	vm_page_set_validclean:
1522  *
1523  *	Sets portions of a page valid and clean.  The arguments are expected
1524  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
1525  *	of any partial chunks touched by the range.  The invalid portion of
1526  *	such chunks will be zero'd.
1527  *
1528  *	This routine may not block.
1529  *
1530  *	(base + size) must be less then or equal to PAGE_SIZE.
1531  */
1532 void
1533 vm_page_set_validclean(vm_page_t m, int base, int size)
1534 {
1535 	int pagebits;
1536 	int frag;
1537 	int endoff;
1538 
1539 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1540 	if (size == 0)	/* handle degenerate case */
1541 		return;
1542 
1543 	/*
1544 	 * If the base is not DEV_BSIZE aligned and the valid
1545 	 * bit is clear, we have to zero out a portion of the
1546 	 * first block.
1547 	 */
1548 	if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
1549 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
1550 		pmap_zero_page_area(m, frag, base - frag);
1551 
1552 	/*
1553 	 * If the ending offset is not DEV_BSIZE aligned and the
1554 	 * valid bit is clear, we have to zero out a portion of
1555 	 * the last block.
1556 	 */
1557 	endoff = base + size;
1558 	if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
1559 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
1560 		pmap_zero_page_area(m, endoff,
1561 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
1562 
1563 	/*
1564 	 * Set valid, clear dirty bits.  If validating the entire
1565 	 * page we can safely clear the pmap modify bit.  We also
1566 	 * use this opportunity to clear the PG_NOSYNC flag.  If a process
1567 	 * takes a write fault on a MAP_NOSYNC memory area the flag will
1568 	 * be set again.
1569 	 *
1570 	 * We set valid bits inclusive of any overlap, but we can only
1571 	 * clear dirty bits for DEV_BSIZE chunks that are fully within
1572 	 * the range.
1573 	 */
1574 	pagebits = vm_page_bits(base, size);
1575 	m->valid |= pagebits;
1576 #if 0	/* NOT YET */
1577 	if ((frag = base & (DEV_BSIZE - 1)) != 0) {
1578 		frag = DEV_BSIZE - frag;
1579 		base += frag;
1580 		size -= frag;
1581 		if (size < 0)
1582 			size = 0;
1583 	}
1584 	pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
1585 #endif
1586 	m->dirty &= ~pagebits;
1587 	if (base == 0 && size == PAGE_SIZE) {
1588 		pmap_clear_modify(m);
1589 		vm_page_flag_clear(m, PG_NOSYNC);
1590 	}
1591 }
1592 
1593 #if 0
1594 
1595 void
1596 vm_page_set_dirty(vm_page_t m, int base, int size)
1597 {
1598 	m->dirty |= vm_page_bits(base, size);
1599 }
1600 
1601 #endif
1602 
1603 void
1604 vm_page_clear_dirty(vm_page_t m, int base, int size)
1605 {
1606 	GIANT_REQUIRED;
1607 	m->dirty &= ~vm_page_bits(base, size);
1608 }
1609 
1610 /*
1611  *	vm_page_set_invalid:
1612  *
1613  *	Invalidates DEV_BSIZE'd chunks within a page.  Both the
1614  *	valid and dirty bits for the effected areas are cleared.
1615  *
1616  *	May not block.
1617  */
1618 void
1619 vm_page_set_invalid(vm_page_t m, int base, int size)
1620 {
1621 	int bits;
1622 
1623 	GIANT_REQUIRED;
1624 	bits = vm_page_bits(base, size);
1625 	m->valid &= ~bits;
1626 	m->dirty &= ~bits;
1627 	m->object->generation++;
1628 }
1629 
1630 /*
1631  * vm_page_zero_invalid()
1632  *
1633  *	The kernel assumes that the invalid portions of a page contain
1634  *	garbage, but such pages can be mapped into memory by user code.
1635  *	When this occurs, we must zero out the non-valid portions of the
1636  *	page so user code sees what it expects.
1637  *
1638  *	Pages are most often semi-valid when the end of a file is mapped
1639  *	into memory and the file's size is not page aligned.
1640  */
1641 void
1642 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
1643 {
1644 	int b;
1645 	int i;
1646 
1647 	/*
1648 	 * Scan the valid bits looking for invalid sections that
1649 	 * must be zerod.  Invalid sub-DEV_BSIZE'd areas ( where the
1650 	 * valid bit may be set ) have already been zerod by
1651 	 * vm_page_set_validclean().
1652 	 */
1653 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
1654 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
1655 		    (m->valid & (1 << i))
1656 		) {
1657 			if (i > b) {
1658 				pmap_zero_page_area(m,
1659 				    b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
1660 			}
1661 			b = i + 1;
1662 		}
1663 	}
1664 
1665 	/*
1666 	 * setvalid is TRUE when we can safely set the zero'd areas
1667 	 * as being valid.  We can do this if there are no cache consistancy
1668 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
1669 	 */
1670 	if (setvalid)
1671 		m->valid = VM_PAGE_BITS_ALL;
1672 }
1673 
1674 /*
1675  *	vm_page_is_valid:
1676  *
1677  *	Is (partial) page valid?  Note that the case where size == 0
1678  *	will return FALSE in the degenerate case where the page is
1679  *	entirely invalid, and TRUE otherwise.
1680  *
1681  *	May not block.
1682  */
1683 int
1684 vm_page_is_valid(vm_page_t m, int base, int size)
1685 {
1686 	int bits = vm_page_bits(base, size);
1687 
1688 	if (m->valid && ((m->valid & bits) == bits))
1689 		return 1;
1690 	else
1691 		return 0;
1692 }
1693 
1694 /*
1695  * update dirty bits from pmap/mmu.  May not block.
1696  */
1697 void
1698 vm_page_test_dirty(vm_page_t m)
1699 {
1700 	if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) {
1701 		vm_page_dirty(m);
1702 	}
1703 }
1704 
1705 int so_zerocp_fullpage = 0;
1706 
1707 void
1708 vm_page_cowfault(vm_page_t m)
1709 {
1710 	vm_page_t mnew;
1711 	vm_object_t object;
1712 	vm_pindex_t pindex;
1713 
1714 	object = m->object;
1715 	pindex = m->pindex;
1716 	vm_page_busy(m);
1717 
1718  retry_alloc:
1719 	vm_page_remove(m);
1720 	/*
1721 	 * An interrupt allocation is requested because the page
1722 	 * queues lock is held.
1723 	 */
1724 	mnew = vm_page_alloc(object, pindex, VM_ALLOC_INTERRUPT);
1725 	if (mnew == NULL) {
1726 		vm_page_insert(m, object, pindex);
1727 		vm_page_unlock_queues();
1728 		VM_WAIT;
1729 		vm_page_lock_queues();
1730 		goto retry_alloc;
1731 	}
1732 
1733 	if (m->cow == 0) {
1734 		/*
1735 		 * check to see if we raced with an xmit complete when
1736 		 * waiting to allocate a page.  If so, put things back
1737 		 * the way they were
1738 		 */
1739 		vm_page_busy(mnew);
1740 		vm_page_free(mnew);
1741 		vm_page_insert(m, object, pindex);
1742 	} else { /* clear COW & copy page */
1743 		if (so_zerocp_fullpage) {
1744 			mnew->valid = VM_PAGE_BITS_ALL;
1745 		} else {
1746 			vm_page_copy(m, mnew);
1747 		}
1748 		vm_page_dirty(mnew);
1749 		vm_page_flag_clear(mnew, PG_BUSY);
1750 	}
1751 }
1752 
1753 void
1754 vm_page_cowclear(vm_page_t m)
1755 {
1756 
1757 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1758 	if (m->cow) {
1759 		m->cow--;
1760 		/*
1761 		 * let vm_fault add back write permission  lazily
1762 		 */
1763 	}
1764 	/*
1765 	 *  sf_buf_free() will free the page, so we needn't do it here
1766 	 */
1767 }
1768 
1769 void
1770 vm_page_cowsetup(vm_page_t m)
1771 {
1772 
1773 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1774 	m->cow++;
1775 	pmap_page_protect(m, VM_PROT_READ);
1776 }
1777 
1778 #include "opt_ddb.h"
1779 #ifdef DDB
1780 #include <sys/kernel.h>
1781 
1782 #include <ddb/ddb.h>
1783 
1784 DB_SHOW_COMMAND(page, vm_page_print_page_info)
1785 {
1786 	db_printf("cnt.v_free_count: %d\n", cnt.v_free_count);
1787 	db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count);
1788 	db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count);
1789 	db_printf("cnt.v_active_count: %d\n", cnt.v_active_count);
1790 	db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count);
1791 	db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved);
1792 	db_printf("cnt.v_free_min: %d\n", cnt.v_free_min);
1793 	db_printf("cnt.v_free_target: %d\n", cnt.v_free_target);
1794 	db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min);
1795 	db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target);
1796 }
1797 
1798 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
1799 {
1800 	int i;
1801 	db_printf("PQ_FREE:");
1802 	for (i = 0; i < PQ_L2_SIZE; i++) {
1803 		db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt);
1804 	}
1805 	db_printf("\n");
1806 
1807 	db_printf("PQ_CACHE:");
1808 	for (i = 0; i < PQ_L2_SIZE; i++) {
1809 		db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt);
1810 	}
1811 	db_printf("\n");
1812 
1813 	db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n",
1814 		vm_page_queues[PQ_ACTIVE].lcnt,
1815 		vm_page_queues[PQ_INACTIVE].lcnt);
1816 }
1817 #endif /* DDB */
1818