xref: /freebsd/sys/vm/vm_fault.c (revision 4c8945a06b01a5c8122cdeb402af36bb46a06acc)
1 /*-
2  * Copyright (c) 1991, 1993
3  *	The Regents of the University of California.  All rights reserved.
4  * Copyright (c) 1994 John S. Dyson
5  * All rights reserved.
6  * Copyright (c) 1994 David Greenman
7  * All rights reserved.
8  *
9  *
10  * This code is derived from software contributed to Berkeley by
11  * The Mach Operating System project at Carnegie-Mellon University.
12  *
13  * Redistribution and use in source and binary forms, with or without
14  * modification, are permitted provided that the following conditions
15  * are met:
16  * 1. Redistributions of source code must retain the above copyright
17  *    notice, this list of conditions and the following disclaimer.
18  * 2. Redistributions in binary form must reproduce the above copyright
19  *    notice, this list of conditions and the following disclaimer in the
20  *    documentation and/or other materials provided with the distribution.
21  * 3. All advertising materials mentioning features or use of this software
22  *    must display the following acknowledgement:
23  *	This product includes software developed by the University of
24  *	California, Berkeley and its contributors.
25  * 4. Neither the name of the University nor the names of its contributors
26  *    may be used to endorse or promote products derived from this software
27  *    without specific prior written permission.
28  *
29  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
30  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
31  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
32  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
33  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
34  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
35  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
36  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
37  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
38  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
39  * SUCH DAMAGE.
40  *
41  *	from: @(#)vm_fault.c	8.4 (Berkeley) 1/12/94
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 /*
71  *	Page fault handling module.
72  */
73 
74 #include <sys/cdefs.h>
75 __FBSDID("$FreeBSD$");
76 
77 #include "opt_vm.h"
78 
79 #include <sys/param.h>
80 #include <sys/systm.h>
81 #include <sys/kernel.h>
82 #include <sys/lock.h>
83 #include <sys/mutex.h>
84 #include <sys/proc.h>
85 #include <sys/resourcevar.h>
86 #include <sys/sysctl.h>
87 #include <sys/vmmeter.h>
88 #include <sys/vnode.h>
89 
90 #include <vm/vm.h>
91 #include <vm/vm_param.h>
92 #include <vm/pmap.h>
93 #include <vm/vm_map.h>
94 #include <vm/vm_object.h>
95 #include <vm/vm_page.h>
96 #include <vm/vm_pageout.h>
97 #include <vm/vm_kern.h>
98 #include <vm/vm_pager.h>
99 #include <vm/vm_extern.h>
100 
101 #include <sys/mount.h>	/* XXX Temporary for VFS_LOCK_GIANT() */
102 
103 #define PFBAK 4
104 #define PFFOR 4
105 #define PAGEORDER_SIZE (PFBAK+PFFOR)
106 
107 static int prefault_pageorder[] = {
108 	-1 * PAGE_SIZE, 1 * PAGE_SIZE,
109 	-2 * PAGE_SIZE, 2 * PAGE_SIZE,
110 	-3 * PAGE_SIZE, 3 * PAGE_SIZE,
111 	-4 * PAGE_SIZE, 4 * PAGE_SIZE
112 };
113 
114 static int vm_fault_additional_pages(vm_page_t, int, int, vm_page_t *, int *);
115 static void vm_fault_prefault(pmap_t, vm_offset_t, vm_map_entry_t);
116 
117 #define VM_FAULT_READ_AHEAD 8
118 #define VM_FAULT_READ_BEHIND 7
119 #define VM_FAULT_READ (VM_FAULT_READ_AHEAD+VM_FAULT_READ_BEHIND+1)
120 
121 struct faultstate {
122 	vm_page_t m;
123 	vm_object_t object;
124 	vm_pindex_t pindex;
125 	vm_page_t first_m;
126 	vm_object_t	first_object;
127 	vm_pindex_t first_pindex;
128 	vm_map_t map;
129 	vm_map_entry_t entry;
130 	int lookup_still_valid;
131 	struct vnode *vp;
132 	int vfslocked;
133 };
134 
135 static inline void
136 release_page(struct faultstate *fs)
137 {
138 
139 	vm_page_wakeup(fs->m);
140 	vm_page_lock(fs->m);
141 	vm_page_deactivate(fs->m);
142 	vm_page_unlock(fs->m);
143 	fs->m = NULL;
144 }
145 
146 static inline void
147 unlock_map(struct faultstate *fs)
148 {
149 
150 	if (fs->lookup_still_valid) {
151 		vm_map_lookup_done(fs->map, fs->entry);
152 		fs->lookup_still_valid = FALSE;
153 	}
154 }
155 
156 static void
157 unlock_and_deallocate(struct faultstate *fs)
158 {
159 
160 	vm_object_pip_wakeup(fs->object);
161 	VM_OBJECT_UNLOCK(fs->object);
162 	if (fs->object != fs->first_object) {
163 		VM_OBJECT_LOCK(fs->first_object);
164 		vm_page_lock(fs->first_m);
165 		vm_page_free(fs->first_m);
166 		vm_page_unlock(fs->first_m);
167 		vm_object_pip_wakeup(fs->first_object);
168 		VM_OBJECT_UNLOCK(fs->first_object);
169 		fs->first_m = NULL;
170 	}
171 	vm_object_deallocate(fs->first_object);
172 	unlock_map(fs);
173 	if (fs->vp != NULL) {
174 		vput(fs->vp);
175 		fs->vp = NULL;
176 	}
177 	VFS_UNLOCK_GIANT(fs->vfslocked);
178 	fs->vfslocked = 0;
179 }
180 
181 /*
182  * TRYPAGER - used by vm_fault to calculate whether the pager for the
183  *	      current object *might* contain the page.
184  *
185  *	      default objects are zero-fill, there is no real pager.
186  */
187 #define TRYPAGER	(fs.object->type != OBJT_DEFAULT && \
188 			((fault_flags & VM_FAULT_CHANGE_WIRING) == 0 || wired))
189 
190 /*
191  *	vm_fault:
192  *
193  *	Handle a page fault occurring at the given address,
194  *	requiring the given permissions, in the map specified.
195  *	If successful, the page is inserted into the
196  *	associated physical map.
197  *
198  *	NOTE: the given address should be truncated to the
199  *	proper page address.
200  *
201  *	KERN_SUCCESS is returned if the page fault is handled; otherwise,
202  *	a standard error specifying why the fault is fatal is returned.
203  *
204  *	The map in question must be referenced, and remains so.
205  *	Caller may hold no locks.
206  */
207 int
208 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
209     int fault_flags)
210 {
211 
212 	return (vm_fault_hold(map, vaddr, fault_type, fault_flags, NULL));
213 }
214 
215 int
216 vm_fault_hold(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
217     int fault_flags, vm_page_t *m_hold)
218 {
219 	vm_prot_t prot;
220 	int is_first_object_locked, result;
221 	boolean_t growstack, wired;
222 	int map_generation;
223 	vm_object_t next_object;
224 	vm_page_t marray[VM_FAULT_READ], mt, mt_prev;
225 	int hardfault;
226 	int faultcount, ahead, behind, alloc_req;
227 	struct faultstate fs;
228 	struct vnode *vp;
229 	int locked, error;
230 
231 	hardfault = 0;
232 	growstack = TRUE;
233 	PCPU_INC(cnt.v_vm_faults);
234 	fs.vp = NULL;
235 	fs.vfslocked = 0;
236 	faultcount = behind = 0;
237 
238 RetryFault:;
239 
240 	/*
241 	 * Find the backing store object and offset into it to begin the
242 	 * search.
243 	 */
244 	fs.map = map;
245 	result = vm_map_lookup(&fs.map, vaddr, fault_type, &fs.entry,
246 	    &fs.first_object, &fs.first_pindex, &prot, &wired);
247 	if (result != KERN_SUCCESS) {
248 		if (growstack && result == KERN_INVALID_ADDRESS &&
249 		    map != kernel_map) {
250 			result = vm_map_growstack(curproc, vaddr);
251 			if (result != KERN_SUCCESS)
252 				return (KERN_FAILURE);
253 			growstack = FALSE;
254 			goto RetryFault;
255 		}
256 		return (result);
257 	}
258 
259 	map_generation = fs.map->timestamp;
260 
261 	if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
262 		panic("vm_fault: fault on nofault entry, addr: %lx",
263 		    (u_long)vaddr);
264 	}
265 
266 	/*
267 	 * Make a reference to this object to prevent its disposal while we
268 	 * are messing with it.  Once we have the reference, the map is free
269 	 * to be diddled.  Since objects reference their shadows (and copies),
270 	 * they will stay around as well.
271 	 *
272 	 * Bump the paging-in-progress count to prevent size changes (e.g.
273 	 * truncation operations) during I/O.  This must be done after
274 	 * obtaining the vnode lock in order to avoid possible deadlocks.
275 	 */
276 	VM_OBJECT_LOCK(fs.first_object);
277 	vm_object_reference_locked(fs.first_object);
278 	vm_object_pip_add(fs.first_object, 1);
279 
280 	fs.lookup_still_valid = TRUE;
281 
282 	if (wired)
283 		fault_type = prot | (fault_type & VM_PROT_COPY);
284 
285 	fs.first_m = NULL;
286 
287 	/*
288 	 * Search for the page at object/offset.
289 	 */
290 	fs.object = fs.first_object;
291 	fs.pindex = fs.first_pindex;
292 	while (TRUE) {
293 		/*
294 		 * If the object is dead, we stop here
295 		 */
296 		if (fs.object->flags & OBJ_DEAD) {
297 			unlock_and_deallocate(&fs);
298 			return (KERN_PROTECTION_FAILURE);
299 		}
300 
301 		/*
302 		 * See if page is resident
303 		 */
304 		fs.m = vm_page_lookup(fs.object, fs.pindex);
305 		if (fs.m != NULL) {
306 			/*
307 			 * check for page-based copy on write.
308 			 * We check fs.object == fs.first_object so
309 			 * as to ensure the legacy COW mechanism is
310 			 * used when the page in question is part of
311 			 * a shadow object.  Otherwise, vm_page_cowfault()
312 			 * removes the page from the backing object,
313 			 * which is not what we want.
314 			 */
315 			vm_page_lock(fs.m);
316 			if ((fs.m->cow) &&
317 			    (fault_type & VM_PROT_WRITE) &&
318 			    (fs.object == fs.first_object)) {
319 				vm_page_cowfault(fs.m);
320 				unlock_and_deallocate(&fs);
321 				goto RetryFault;
322 			}
323 
324 			/*
325 			 * Wait/Retry if the page is busy.  We have to do this
326 			 * if the page is busy via either VPO_BUSY or
327 			 * vm_page_t->busy because the vm_pager may be using
328 			 * vm_page_t->busy for pageouts ( and even pageins if
329 			 * it is the vnode pager ), and we could end up trying
330 			 * to pagein and pageout the same page simultaneously.
331 			 *
332 			 * We can theoretically allow the busy case on a read
333 			 * fault if the page is marked valid, but since such
334 			 * pages are typically already pmap'd, putting that
335 			 * special case in might be more effort then it is
336 			 * worth.  We cannot under any circumstances mess
337 			 * around with a vm_page_t->busy page except, perhaps,
338 			 * to pmap it.
339 			 */
340 			if ((fs.m->oflags & VPO_BUSY) || fs.m->busy) {
341 				/*
342 				 * Reference the page before unlocking and
343 				 * sleeping so that the page daemon is less
344 				 * likely to reclaim it.
345 				 */
346 				vm_page_lock_queues();
347 				vm_page_flag_set(fs.m, PG_REFERENCED);
348 				vm_page_unlock_queues();
349 				vm_page_unlock(fs.m);
350 				if (fs.object != fs.first_object) {
351 					if (!VM_OBJECT_TRYLOCK(
352 					    fs.first_object)) {
353 						VM_OBJECT_UNLOCK(fs.object);
354 						VM_OBJECT_LOCK(fs.first_object);
355 						VM_OBJECT_LOCK(fs.object);
356 					}
357 					vm_page_lock(fs.first_m);
358 					vm_page_free(fs.first_m);
359 					vm_page_unlock(fs.first_m);
360 					vm_object_pip_wakeup(fs.first_object);
361 					VM_OBJECT_UNLOCK(fs.first_object);
362 					fs.first_m = NULL;
363 				}
364 				unlock_map(&fs);
365 				if (fs.m == vm_page_lookup(fs.object,
366 				    fs.pindex)) {
367 					vm_page_sleep_if_busy(fs.m, TRUE,
368 					    "vmpfw");
369 				}
370 				vm_object_pip_wakeup(fs.object);
371 				VM_OBJECT_UNLOCK(fs.object);
372 				PCPU_INC(cnt.v_intrans);
373 				vm_object_deallocate(fs.first_object);
374 				goto RetryFault;
375 			}
376 			vm_pageq_remove(fs.m);
377 			vm_page_unlock(fs.m);
378 
379 			/*
380 			 * Mark page busy for other processes, and the
381 			 * pagedaemon.  If it still isn't completely valid
382 			 * (readable), jump to readrest, else break-out ( we
383 			 * found the page ).
384 			 */
385 			vm_page_busy(fs.m);
386 			if (fs.m->valid != VM_PAGE_BITS_ALL &&
387 				fs.m->object != kernel_object && fs.m->object != kmem_object) {
388 				goto readrest;
389 			}
390 
391 			break;
392 		}
393 
394 		/*
395 		 * Page is not resident, If this is the search termination
396 		 * or the pager might contain the page, allocate a new page.
397 		 */
398 		if (TRYPAGER || fs.object == fs.first_object) {
399 			if (fs.pindex >= fs.object->size) {
400 				unlock_and_deallocate(&fs);
401 				return (KERN_PROTECTION_FAILURE);
402 			}
403 
404 			/*
405 			 * Allocate a new page for this object/offset pair.
406 			 *
407 			 * Unlocked read of the p_flag is harmless. At
408 			 * worst, the P_KILLED might be not observed
409 			 * there, and allocation can fail, causing
410 			 * restart and new reading of the p_flag.
411 			 */
412 			fs.m = NULL;
413 			if (!vm_page_count_severe() || P_KILLED(curproc)) {
414 #if VM_NRESERVLEVEL > 0
415 				if ((fs.object->flags & OBJ_COLORED) == 0) {
416 					fs.object->flags |= OBJ_COLORED;
417 					fs.object->pg_color = atop(vaddr) -
418 					    fs.pindex;
419 				}
420 #endif
421 				alloc_req = P_KILLED(curproc) ?
422 				    VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL;
423 				if (fs.object->type != OBJT_VNODE &&
424 				    fs.object->backing_object == NULL)
425 					alloc_req |= VM_ALLOC_ZERO;
426 				fs.m = vm_page_alloc(fs.object, fs.pindex,
427 				    alloc_req);
428 			}
429 			if (fs.m == NULL) {
430 				unlock_and_deallocate(&fs);
431 				VM_WAITPFAULT;
432 				goto RetryFault;
433 			} else if (fs.m->valid == VM_PAGE_BITS_ALL)
434 				break;
435 		}
436 
437 readrest:
438 		/*
439 		 * We have found a valid page or we have allocated a new page.
440 		 * The page thus may not be valid or may not be entirely
441 		 * valid.
442 		 *
443 		 * Attempt to fault-in the page if there is a chance that the
444 		 * pager has it, and potentially fault in additional pages
445 		 * at the same time.
446 		 */
447 		if (TRYPAGER) {
448 			int rv;
449 			int reqpage = 0;
450 			u_char behavior = vm_map_entry_behavior(fs.entry);
451 
452 			if (behavior == MAP_ENTRY_BEHAV_RANDOM ||
453 			    P_KILLED(curproc)) {
454 				ahead = 0;
455 				behind = 0;
456 			} else {
457 				behind = (vaddr - fs.entry->start) >> PAGE_SHIFT;
458 				if (behind > VM_FAULT_READ_BEHIND)
459 					behind = VM_FAULT_READ_BEHIND;
460 
461 				ahead = ((fs.entry->end - vaddr) >> PAGE_SHIFT) - 1;
462 				if (ahead > VM_FAULT_READ_AHEAD)
463 					ahead = VM_FAULT_READ_AHEAD;
464 			}
465 			is_first_object_locked = FALSE;
466 			if ((behavior == MAP_ENTRY_BEHAV_SEQUENTIAL ||
467 			     (behavior != MAP_ENTRY_BEHAV_RANDOM &&
468 			      fs.pindex >= fs.entry->lastr &&
469 			      fs.pindex < fs.entry->lastr + VM_FAULT_READ)) &&
470 			    (fs.first_object == fs.object ||
471 			     (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object))) &&
472 			    fs.first_object->type != OBJT_DEVICE &&
473 			    fs.first_object->type != OBJT_PHYS &&
474 			    fs.first_object->type != OBJT_SG) {
475 				vm_pindex_t firstpindex;
476 
477 				if (fs.first_pindex < 2 * VM_FAULT_READ)
478 					firstpindex = 0;
479 				else
480 					firstpindex = fs.first_pindex - 2 * VM_FAULT_READ;
481 				mt = fs.first_object != fs.object ?
482 				    fs.first_m : fs.m;
483 				KASSERT(mt != NULL, ("vm_fault: missing mt"));
484 				KASSERT((mt->oflags & VPO_BUSY) != 0,
485 				    ("vm_fault: mt %p not busy", mt));
486 				mt_prev = vm_page_prev(mt);
487 
488 				/*
489 				 * note: partially valid pages cannot be
490 				 * included in the lookahead - NFS piecemeal
491 				 * writes will barf on it badly.
492 				 */
493 				while ((mt = mt_prev) != NULL &&
494 				    mt->pindex >= firstpindex &&
495 				    mt->valid == VM_PAGE_BITS_ALL) {
496 					mt_prev = vm_page_prev(mt);
497 					if (mt->busy ||
498 					    (mt->oflags & VPO_BUSY))
499 						continue;
500 					vm_page_lock(mt);
501 					if (mt->hold_count ||
502 					    mt->wire_count) {
503 						vm_page_unlock(mt);
504 						continue;
505 					}
506 					pmap_remove_all(mt);
507 					if (mt->dirty != 0)
508 						vm_page_deactivate(mt);
509 					else
510 						vm_page_cache(mt);
511 					vm_page_unlock(mt);
512 				}
513 				ahead += behind;
514 				behind = 0;
515 			}
516 			if (is_first_object_locked)
517 				VM_OBJECT_UNLOCK(fs.first_object);
518 
519 			/*
520 			 * Call the pager to retrieve the data, if any, after
521 			 * releasing the lock on the map.  We hold a ref on
522 			 * fs.object and the pages are VPO_BUSY'd.
523 			 */
524 			unlock_map(&fs);
525 
526 vnode_lock:
527 			if (fs.object->type == OBJT_VNODE) {
528 				vp = fs.object->handle;
529 				if (vp == fs.vp)
530 					goto vnode_locked;
531 				else if (fs.vp != NULL) {
532 					vput(fs.vp);
533 					fs.vp = NULL;
534 				}
535 				locked = VOP_ISLOCKED(vp);
536 
537 				if (VFS_NEEDSGIANT(vp->v_mount) && !fs.vfslocked) {
538 					fs.vfslocked = 1;
539 					if (!mtx_trylock(&Giant)) {
540 						VM_OBJECT_UNLOCK(fs.object);
541 						mtx_lock(&Giant);
542 						VM_OBJECT_LOCK(fs.object);
543 						goto vnode_lock;
544 					}
545 				}
546 				if (locked != LK_EXCLUSIVE)
547 					locked = LK_SHARED;
548 				/* Do not sleep for vnode lock while fs.m is busy */
549 				error = vget(vp, locked | LK_CANRECURSE |
550 				    LK_NOWAIT, curthread);
551 				if (error != 0) {
552 					int vfslocked;
553 
554 					vfslocked = fs.vfslocked;
555 					fs.vfslocked = 0; /* Keep Giant */
556 					vhold(vp);
557 					release_page(&fs);
558 					unlock_and_deallocate(&fs);
559 					error = vget(vp, locked | LK_RETRY |
560 					    LK_CANRECURSE, curthread);
561 					vdrop(vp);
562 					fs.vp = vp;
563 					fs.vfslocked = vfslocked;
564 					KASSERT(error == 0,
565 					    ("vm_fault: vget failed"));
566 					goto RetryFault;
567 				}
568 				fs.vp = vp;
569 			}
570 vnode_locked:
571 			KASSERT(fs.vp == NULL || !fs.map->system_map,
572 			    ("vm_fault: vnode-backed object mapped by system map"));
573 
574 			/*
575 			 * now we find out if any other pages should be paged
576 			 * in at this time this routine checks to see if the
577 			 * pages surrounding this fault reside in the same
578 			 * object as the page for this fault.  If they do,
579 			 * then they are faulted in also into the object.  The
580 			 * array "marray" returned contains an array of
581 			 * vm_page_t structs where one of them is the
582 			 * vm_page_t passed to the routine.  The reqpage
583 			 * return value is the index into the marray for the
584 			 * vm_page_t passed to the routine.
585 			 *
586 			 * fs.m plus the additional pages are VPO_BUSY'd.
587 			 */
588 			faultcount = vm_fault_additional_pages(
589 			    fs.m, behind, ahead, marray, &reqpage);
590 
591 			rv = faultcount ?
592 			    vm_pager_get_pages(fs.object, marray, faultcount,
593 				reqpage) : VM_PAGER_FAIL;
594 
595 			if (rv == VM_PAGER_OK) {
596 				/*
597 				 * Found the page. Leave it busy while we play
598 				 * with it.
599 				 */
600 
601 				/*
602 				 * Relookup in case pager changed page. Pager
603 				 * is responsible for disposition of old page
604 				 * if moved.
605 				 */
606 				fs.m = vm_page_lookup(fs.object, fs.pindex);
607 				if (!fs.m) {
608 					unlock_and_deallocate(&fs);
609 					goto RetryFault;
610 				}
611 
612 				hardfault++;
613 				break; /* break to PAGE HAS BEEN FOUND */
614 			}
615 			/*
616 			 * Remove the bogus page (which does not exist at this
617 			 * object/offset); before doing so, we must get back
618 			 * our object lock to preserve our invariant.
619 			 *
620 			 * Also wake up any other process that may want to bring
621 			 * in this page.
622 			 *
623 			 * If this is the top-level object, we must leave the
624 			 * busy page to prevent another process from rushing
625 			 * past us, and inserting the page in that object at
626 			 * the same time that we are.
627 			 */
628 			if (rv == VM_PAGER_ERROR)
629 				printf("vm_fault: pager read error, pid %d (%s)\n",
630 				    curproc->p_pid, curproc->p_comm);
631 			/*
632 			 * Data outside the range of the pager or an I/O error
633 			 */
634 			/*
635 			 * XXX - the check for kernel_map is a kludge to work
636 			 * around having the machine panic on a kernel space
637 			 * fault w/ I/O error.
638 			 */
639 			if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) ||
640 				(rv == VM_PAGER_BAD)) {
641 				vm_page_lock(fs.m);
642 				vm_page_free(fs.m);
643 				vm_page_unlock(fs.m);
644 				fs.m = NULL;
645 				unlock_and_deallocate(&fs);
646 				return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
647 			}
648 			if (fs.object != fs.first_object) {
649 				vm_page_lock(fs.m);
650 				vm_page_free(fs.m);
651 				vm_page_unlock(fs.m);
652 				fs.m = NULL;
653 				/*
654 				 * XXX - we cannot just fall out at this
655 				 * point, m has been freed and is invalid!
656 				 */
657 			}
658 		}
659 
660 		/*
661 		 * We get here if the object has default pager (or unwiring)
662 		 * or the pager doesn't have the page.
663 		 */
664 		if (fs.object == fs.first_object)
665 			fs.first_m = fs.m;
666 
667 		/*
668 		 * Move on to the next object.  Lock the next object before
669 		 * unlocking the current one.
670 		 */
671 		fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset);
672 		next_object = fs.object->backing_object;
673 		if (next_object == NULL) {
674 			/*
675 			 * If there's no object left, fill the page in the top
676 			 * object with zeros.
677 			 */
678 			if (fs.object != fs.first_object) {
679 				vm_object_pip_wakeup(fs.object);
680 				VM_OBJECT_UNLOCK(fs.object);
681 
682 				fs.object = fs.first_object;
683 				fs.pindex = fs.first_pindex;
684 				fs.m = fs.first_m;
685 				VM_OBJECT_LOCK(fs.object);
686 			}
687 			fs.first_m = NULL;
688 
689 			/*
690 			 * Zero the page if necessary and mark it valid.
691 			 */
692 			if ((fs.m->flags & PG_ZERO) == 0) {
693 				pmap_zero_page(fs.m);
694 			} else {
695 				PCPU_INC(cnt.v_ozfod);
696 			}
697 			PCPU_INC(cnt.v_zfod);
698 			fs.m->valid = VM_PAGE_BITS_ALL;
699 			break;	/* break to PAGE HAS BEEN FOUND */
700 		} else {
701 			KASSERT(fs.object != next_object,
702 			    ("object loop %p", next_object));
703 			VM_OBJECT_LOCK(next_object);
704 			vm_object_pip_add(next_object, 1);
705 			if (fs.object != fs.first_object)
706 				vm_object_pip_wakeup(fs.object);
707 			VM_OBJECT_UNLOCK(fs.object);
708 			fs.object = next_object;
709 		}
710 	}
711 
712 	KASSERT((fs.m->oflags & VPO_BUSY) != 0,
713 	    ("vm_fault: not busy after main loop"));
714 
715 	/*
716 	 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
717 	 * is held.]
718 	 */
719 
720 	/*
721 	 * If the page is being written, but isn't already owned by the
722 	 * top-level object, we have to copy it into a new page owned by the
723 	 * top-level object.
724 	 */
725 	if (fs.object != fs.first_object) {
726 		/*
727 		 * We only really need to copy if we want to write it.
728 		 */
729 		if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
730 			/*
731 			 * This allows pages to be virtually copied from a
732 			 * backing_object into the first_object, where the
733 			 * backing object has no other refs to it, and cannot
734 			 * gain any more refs.  Instead of a bcopy, we just
735 			 * move the page from the backing object to the
736 			 * first object.  Note that we must mark the page
737 			 * dirty in the first object so that it will go out
738 			 * to swap when needed.
739 			 */
740 			is_first_object_locked = FALSE;
741 			if (
742 				/*
743 				 * Only one shadow object
744 				 */
745 				(fs.object->shadow_count == 1) &&
746 				/*
747 				 * No COW refs, except us
748 				 */
749 				(fs.object->ref_count == 1) &&
750 				/*
751 				 * No one else can look this object up
752 				 */
753 				(fs.object->handle == NULL) &&
754 				/*
755 				 * No other ways to look the object up
756 				 */
757 				((fs.object->type == OBJT_DEFAULT) ||
758 				 (fs.object->type == OBJT_SWAP)) &&
759 			    (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object)) &&
760 				/*
761 				 * We don't chase down the shadow chain
762 				 */
763 			    fs.object == fs.first_object->backing_object) {
764 				/*
765 				 * get rid of the unnecessary page
766 				 */
767 				vm_page_lock(fs.first_m);
768 				vm_page_free(fs.first_m);
769 				vm_page_unlock(fs.first_m);
770 				/*
771 				 * grab the page and put it into the
772 				 * process'es object.  The page is
773 				 * automatically made dirty.
774 				 */
775 				vm_page_lock(fs.m);
776 				vm_page_rename(fs.m, fs.first_object, fs.first_pindex);
777 				vm_page_unlock(fs.m);
778 				vm_page_busy(fs.m);
779 				fs.first_m = fs.m;
780 				fs.m = NULL;
781 				PCPU_INC(cnt.v_cow_optim);
782 			} else {
783 				/*
784 				 * Oh, well, lets copy it.
785 				 */
786 				pmap_copy_page(fs.m, fs.first_m);
787 				fs.first_m->valid = VM_PAGE_BITS_ALL;
788 				if (wired && (fault_flags &
789 				    VM_FAULT_CHANGE_WIRING) == 0) {
790 					vm_page_lock(fs.first_m);
791 					vm_page_wire(fs.first_m);
792 					vm_page_unlock(fs.first_m);
793 
794 					vm_page_lock(fs.m);
795 					vm_page_unwire(fs.m, FALSE);
796 					vm_page_unlock(fs.m);
797 				}
798 				/*
799 				 * We no longer need the old page or object.
800 				 */
801 				release_page(&fs);
802 			}
803 			/*
804 			 * fs.object != fs.first_object due to above
805 			 * conditional
806 			 */
807 			vm_object_pip_wakeup(fs.object);
808 			VM_OBJECT_UNLOCK(fs.object);
809 			/*
810 			 * Only use the new page below...
811 			 */
812 			fs.object = fs.first_object;
813 			fs.pindex = fs.first_pindex;
814 			fs.m = fs.first_m;
815 			if (!is_first_object_locked)
816 				VM_OBJECT_LOCK(fs.object);
817 			PCPU_INC(cnt.v_cow_faults);
818 		} else {
819 			prot &= ~VM_PROT_WRITE;
820 		}
821 	}
822 
823 	/*
824 	 * We must verify that the maps have not changed since our last
825 	 * lookup.
826 	 */
827 	if (!fs.lookup_still_valid) {
828 		vm_object_t retry_object;
829 		vm_pindex_t retry_pindex;
830 		vm_prot_t retry_prot;
831 
832 		if (!vm_map_trylock_read(fs.map)) {
833 			release_page(&fs);
834 			unlock_and_deallocate(&fs);
835 			goto RetryFault;
836 		}
837 		fs.lookup_still_valid = TRUE;
838 		if (fs.map->timestamp != map_generation) {
839 			result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
840 			    &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
841 
842 			/*
843 			 * If we don't need the page any longer, put it on the inactive
844 			 * list (the easiest thing to do here).  If no one needs it,
845 			 * pageout will grab it eventually.
846 			 */
847 			if (result != KERN_SUCCESS) {
848 				release_page(&fs);
849 				unlock_and_deallocate(&fs);
850 
851 				/*
852 				 * If retry of map lookup would have blocked then
853 				 * retry fault from start.
854 				 */
855 				if (result == KERN_FAILURE)
856 					goto RetryFault;
857 				return (result);
858 			}
859 			if ((retry_object != fs.first_object) ||
860 			    (retry_pindex != fs.first_pindex)) {
861 				release_page(&fs);
862 				unlock_and_deallocate(&fs);
863 				goto RetryFault;
864 			}
865 
866 			/*
867 			 * Check whether the protection has changed or the object has
868 			 * been copied while we left the map unlocked. Changing from
869 			 * read to write permission is OK - we leave the page
870 			 * write-protected, and catch the write fault. Changing from
871 			 * write to read permission means that we can't mark the page
872 			 * write-enabled after all.
873 			 */
874 			prot &= retry_prot;
875 		}
876 	}
877 	/*
878 	 * If the page was filled by a pager, update the map entry's
879 	 * last read offset.  Since the pager does not return the
880 	 * actual set of pages that it read, this update is based on
881 	 * the requested set.  Typically, the requested and actual
882 	 * sets are the same.
883 	 *
884 	 * XXX The following assignment modifies the map
885 	 * without holding a write lock on it.
886 	 */
887 	if (hardfault)
888 		fs.entry->lastr = fs.pindex + faultcount - behind;
889 
890 	if ((prot & VM_PROT_WRITE) != 0 ||
891 	    (fault_flags & VM_FAULT_DIRTY) != 0) {
892 		vm_object_set_writeable_dirty(fs.object);
893 
894 		/*
895 		 * If this is a NOSYNC mmap we do not want to set VPO_NOSYNC
896 		 * if the page is already dirty to prevent data written with
897 		 * the expectation of being synced from not being synced.
898 		 * Likewise if this entry does not request NOSYNC then make
899 		 * sure the page isn't marked NOSYNC.  Applications sharing
900 		 * data should use the same flags to avoid ping ponging.
901 		 */
902 		if (fs.entry->eflags & MAP_ENTRY_NOSYNC) {
903 			if (fs.m->dirty == 0)
904 				fs.m->oflags |= VPO_NOSYNC;
905 		} else {
906 			fs.m->oflags &= ~VPO_NOSYNC;
907 		}
908 
909 		/*
910 		 * If the fault is a write, we know that this page is being
911 		 * written NOW so dirty it explicitly to save on
912 		 * pmap_is_modified() calls later.
913 		 *
914 		 * Also tell the backing pager, if any, that it should remove
915 		 * any swap backing since the page is now dirty.
916 		 */
917 		if (((fault_type & VM_PROT_WRITE) != 0 &&
918 		    (fault_flags & VM_FAULT_CHANGE_WIRING) == 0) ||
919 		    (fault_flags & VM_FAULT_DIRTY) != 0) {
920 			vm_page_dirty(fs.m);
921 			vm_pager_page_unswapped(fs.m);
922 		}
923 	}
924 
925 	/*
926 	 * Page had better still be busy
927 	 */
928 	KASSERT(fs.m->oflags & VPO_BUSY,
929 		("vm_fault: page %p not busy!", fs.m));
930 	/*
931 	 * Page must be completely valid or it is not fit to
932 	 * map into user space.  vm_pager_get_pages() ensures this.
933 	 */
934 	KASSERT(fs.m->valid == VM_PAGE_BITS_ALL,
935 	    ("vm_fault: page %p partially invalid", fs.m));
936 	VM_OBJECT_UNLOCK(fs.object);
937 
938 	/*
939 	 * Put this page into the physical map.  We had to do the unlock above
940 	 * because pmap_enter() may sleep.  We don't put the page
941 	 * back on the active queue until later so that the pageout daemon
942 	 * won't find it (yet).
943 	 */
944 	pmap_enter(fs.map->pmap, vaddr, fault_type, fs.m, prot, wired);
945 	if ((fault_flags & VM_FAULT_CHANGE_WIRING) == 0 && wired == 0)
946 		vm_fault_prefault(fs.map->pmap, vaddr, fs.entry);
947 	VM_OBJECT_LOCK(fs.object);
948 	vm_page_lock(fs.m);
949 
950 	/*
951 	 * If the page is not wired down, then put it where the pageout daemon
952 	 * can find it.
953 	 */
954 	if (fault_flags & VM_FAULT_CHANGE_WIRING) {
955 		if (wired)
956 			vm_page_wire(fs.m);
957 		else
958 			vm_page_unwire(fs.m, 1);
959 	} else
960 		vm_page_activate(fs.m);
961 	if (m_hold != NULL) {
962 		*m_hold = fs.m;
963 		vm_page_hold(fs.m);
964 	}
965 	vm_page_unlock(fs.m);
966 	vm_page_wakeup(fs.m);
967 
968 	/*
969 	 * Unlock everything, and return
970 	 */
971 	unlock_and_deallocate(&fs);
972 	if (hardfault)
973 		curthread->td_ru.ru_majflt++;
974 	else
975 		curthread->td_ru.ru_minflt++;
976 
977 	return (KERN_SUCCESS);
978 }
979 
980 /*
981  * vm_fault_prefault provides a quick way of clustering
982  * pagefaults into a processes address space.  It is a "cousin"
983  * of vm_map_pmap_enter, except it runs at page fault time instead
984  * of mmap time.
985  */
986 static void
987 vm_fault_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry)
988 {
989 	int i;
990 	vm_offset_t addr, starta;
991 	vm_pindex_t pindex;
992 	vm_page_t m;
993 	vm_object_t object;
994 
995 	if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
996 		return;
997 
998 	object = entry->object.vm_object;
999 
1000 	starta = addra - PFBAK * PAGE_SIZE;
1001 	if (starta < entry->start) {
1002 		starta = entry->start;
1003 	} else if (starta > addra) {
1004 		starta = 0;
1005 	}
1006 
1007 	for (i = 0; i < PAGEORDER_SIZE; i++) {
1008 		vm_object_t backing_object, lobject;
1009 
1010 		addr = addra + prefault_pageorder[i];
1011 		if (addr > addra + (PFFOR * PAGE_SIZE))
1012 			addr = 0;
1013 
1014 		if (addr < starta || addr >= entry->end)
1015 			continue;
1016 
1017 		if (!pmap_is_prefaultable(pmap, addr))
1018 			continue;
1019 
1020 		pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1021 		lobject = object;
1022 		VM_OBJECT_LOCK(lobject);
1023 		while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1024 		    lobject->type == OBJT_DEFAULT &&
1025 		    (backing_object = lobject->backing_object) != NULL) {
1026 			KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1027 			    0, ("vm_fault_prefault: unaligned object offset"));
1028 			pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1029 			VM_OBJECT_LOCK(backing_object);
1030 			VM_OBJECT_UNLOCK(lobject);
1031 			lobject = backing_object;
1032 		}
1033 		/*
1034 		 * give-up when a page is not in memory
1035 		 */
1036 		if (m == NULL) {
1037 			VM_OBJECT_UNLOCK(lobject);
1038 			break;
1039 		}
1040 		if (m->valid == VM_PAGE_BITS_ALL &&
1041 		    (m->flags & PG_FICTITIOUS) == 0)
1042 			pmap_enter_quick(pmap, addr, m, entry->protection);
1043 		VM_OBJECT_UNLOCK(lobject);
1044 	}
1045 }
1046 
1047 /*
1048  * Hold each of the physical pages that are mapped by the specified range of
1049  * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
1050  * and allow the specified types of access, "prot".  If all of the implied
1051  * pages are successfully held, then the number of held pages is returned
1052  * together with pointers to those pages in the array "ma".  However, if any
1053  * of the pages cannot be held, -1 is returned.
1054  */
1055 int
1056 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
1057     vm_prot_t prot, vm_page_t *ma, int max_count)
1058 {
1059 	vm_offset_t end, va;
1060 	vm_page_t *mp;
1061 	int count;
1062 	boolean_t pmap_failed;
1063 
1064 	end = round_page(addr + len);
1065 	addr = trunc_page(addr);
1066 
1067 	/*
1068 	 * Check for illegal addresses.
1069 	 */
1070 	if (addr < vm_map_min(map) || addr > end || end > vm_map_max(map))
1071 		return (-1);
1072 
1073 	count = howmany(end - addr, PAGE_SIZE);
1074 	if (count > max_count)
1075 		panic("vm_fault_quick_hold_pages: count > max_count");
1076 
1077 	/*
1078 	 * Most likely, the physical pages are resident in the pmap, so it is
1079 	 * faster to try pmap_extract_and_hold() first.
1080 	 */
1081 	pmap_failed = FALSE;
1082 	for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) {
1083 		*mp = pmap_extract_and_hold(map->pmap, va, prot);
1084 		if (*mp == NULL)
1085 			pmap_failed = TRUE;
1086 		else if ((prot & VM_PROT_WRITE) != 0 &&
1087 		    (*mp)->dirty != VM_PAGE_BITS_ALL) {
1088 			/*
1089 			 * Explicitly dirty the physical page.  Otherwise, the
1090 			 * caller's changes may go unnoticed because they are
1091 			 * performed through an unmanaged mapping or by a DMA
1092 			 * operation.
1093 			 */
1094 			vm_page_lock_queues();
1095 			vm_page_dirty(*mp);
1096 			vm_page_unlock_queues();
1097 		}
1098 	}
1099 	if (pmap_failed) {
1100 		/*
1101 		 * One or more pages could not be held by the pmap.  Either no
1102 		 * page was mapped at the specified virtual address or that
1103 		 * mapping had insufficient permissions.  Attempt to fault in
1104 		 * and hold these pages.
1105 		 */
1106 		for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE)
1107 			if (*mp == NULL && vm_fault_hold(map, va, prot,
1108 			    VM_FAULT_NORMAL, mp) != KERN_SUCCESS)
1109 				goto error;
1110 	}
1111 	return (count);
1112 error:
1113 	for (mp = ma; mp < ma + count; mp++)
1114 		if (*mp != NULL) {
1115 			vm_page_lock(*mp);
1116 			vm_page_unhold(*mp);
1117 			vm_page_unlock(*mp);
1118 		}
1119 	return (-1);
1120 }
1121 
1122 /*
1123  *	vm_fault_wire:
1124  *
1125  *	Wire down a range of virtual addresses in a map.
1126  */
1127 int
1128 vm_fault_wire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1129     boolean_t fictitious)
1130 {
1131 	vm_offset_t va;
1132 	int rv;
1133 
1134 	/*
1135 	 * We simulate a fault to get the page and enter it in the physical
1136 	 * map.  For user wiring, we only ask for read access on currently
1137 	 * read-only sections.
1138 	 */
1139 	for (va = start; va < end; va += PAGE_SIZE) {
1140 		rv = vm_fault(map, va, VM_PROT_NONE, VM_FAULT_CHANGE_WIRING);
1141 		if (rv) {
1142 			if (va != start)
1143 				vm_fault_unwire(map, start, va, fictitious);
1144 			return (rv);
1145 		}
1146 	}
1147 	return (KERN_SUCCESS);
1148 }
1149 
1150 /*
1151  *	vm_fault_unwire:
1152  *
1153  *	Unwire a range of virtual addresses in a map.
1154  */
1155 void
1156 vm_fault_unwire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1157     boolean_t fictitious)
1158 {
1159 	vm_paddr_t pa;
1160 	vm_offset_t va;
1161 	vm_page_t m;
1162 	pmap_t pmap;
1163 
1164 	pmap = vm_map_pmap(map);
1165 
1166 	/*
1167 	 * Since the pages are wired down, we must be able to get their
1168 	 * mappings from the physical map system.
1169 	 */
1170 	for (va = start; va < end; va += PAGE_SIZE) {
1171 		pa = pmap_extract(pmap, va);
1172 		if (pa != 0) {
1173 			pmap_change_wiring(pmap, va, FALSE);
1174 			if (!fictitious) {
1175 				m = PHYS_TO_VM_PAGE(pa);
1176 				vm_page_lock(m);
1177 				vm_page_unwire(m, TRUE);
1178 				vm_page_unlock(m);
1179 			}
1180 		}
1181 	}
1182 }
1183 
1184 /*
1185  *	Routine:
1186  *		vm_fault_copy_entry
1187  *	Function:
1188  *		Create new shadow object backing dst_entry with private copy of
1189  *		all underlying pages. When src_entry is equal to dst_entry,
1190  *		function implements COW for wired-down map entry. Otherwise,
1191  *		it forks wired entry into dst_map.
1192  *
1193  *	In/out conditions:
1194  *		The source and destination maps must be locked for write.
1195  *		The source map entry must be wired down (or be a sharing map
1196  *		entry corresponding to a main map entry that is wired down).
1197  */
1198 void
1199 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1200     vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1201     vm_ooffset_t *fork_charge)
1202 {
1203 	vm_object_t backing_object, dst_object, object, src_object;
1204 	vm_pindex_t dst_pindex, pindex, src_pindex;
1205 	vm_prot_t access, prot;
1206 	vm_offset_t vaddr;
1207 	vm_page_t dst_m;
1208 	vm_page_t src_m;
1209 	boolean_t src_readonly, upgrade;
1210 
1211 #ifdef	lint
1212 	src_map++;
1213 #endif	/* lint */
1214 
1215 	upgrade = src_entry == dst_entry;
1216 
1217 	src_object = src_entry->object.vm_object;
1218 	src_pindex = OFF_TO_IDX(src_entry->offset);
1219 	src_readonly = (src_entry->protection & VM_PROT_WRITE) == 0;
1220 
1221 	/*
1222 	 * Create the top-level object for the destination entry. (Doesn't
1223 	 * actually shadow anything - we copy the pages directly.)
1224 	 */
1225 	dst_object = vm_object_allocate(OBJT_DEFAULT,
1226 	    OFF_TO_IDX(dst_entry->end - dst_entry->start));
1227 #if VM_NRESERVLEVEL > 0
1228 	dst_object->flags |= OBJ_COLORED;
1229 	dst_object->pg_color = atop(dst_entry->start);
1230 #endif
1231 
1232 	VM_OBJECT_LOCK(dst_object);
1233 	KASSERT(upgrade || dst_entry->object.vm_object == NULL,
1234 	    ("vm_fault_copy_entry: vm_object not NULL"));
1235 	dst_entry->object.vm_object = dst_object;
1236 	dst_entry->offset = 0;
1237 	dst_object->charge = dst_entry->end - dst_entry->start;
1238 	if (fork_charge != NULL) {
1239 		KASSERT(dst_entry->cred == NULL,
1240 		    ("vm_fault_copy_entry: leaked swp charge"));
1241 		dst_object->cred = curthread->td_ucred;
1242 		crhold(dst_object->cred);
1243 		*fork_charge += dst_object->charge;
1244 	} else {
1245 		dst_object->cred = dst_entry->cred;
1246 		dst_entry->cred = NULL;
1247 	}
1248 	access = prot = dst_entry->protection;
1249 	/*
1250 	 * If not an upgrade, then enter the mappings in the pmap as
1251 	 * read and/or execute accesses.  Otherwise, enter them as
1252 	 * write accesses.
1253 	 *
1254 	 * A writeable large page mapping is only created if all of
1255 	 * the constituent small page mappings are modified. Marking
1256 	 * PTEs as modified on inception allows promotion to happen
1257 	 * without taking potentially large number of soft faults.
1258 	 */
1259 	if (!upgrade)
1260 		access &= ~VM_PROT_WRITE;
1261 
1262 	/*
1263 	 * Loop through all of the pages in the entry's range, copying each
1264 	 * one from the source object (it should be there) to the destination
1265 	 * object.
1266 	 */
1267 	for (vaddr = dst_entry->start, dst_pindex = 0;
1268 	    vaddr < dst_entry->end;
1269 	    vaddr += PAGE_SIZE, dst_pindex++) {
1270 
1271 		/*
1272 		 * Allocate a page in the destination object.
1273 		 */
1274 		do {
1275 			dst_m = vm_page_alloc(dst_object, dst_pindex,
1276 			    VM_ALLOC_NORMAL);
1277 			if (dst_m == NULL) {
1278 				VM_OBJECT_UNLOCK(dst_object);
1279 				VM_WAIT;
1280 				VM_OBJECT_LOCK(dst_object);
1281 			}
1282 		} while (dst_m == NULL);
1283 
1284 		/*
1285 		 * Find the page in the source object, and copy it in.
1286 		 * (Because the source is wired down, the page will be in
1287 		 * memory.)
1288 		 */
1289 		VM_OBJECT_LOCK(src_object);
1290 		object = src_object;
1291 		pindex = src_pindex + dst_pindex;
1292 		while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
1293 		    src_readonly &&
1294 		    (backing_object = object->backing_object) != NULL) {
1295 			/*
1296 			 * Allow fallback to backing objects if we are reading.
1297 			 */
1298 			VM_OBJECT_LOCK(backing_object);
1299 			pindex += OFF_TO_IDX(object->backing_object_offset);
1300 			VM_OBJECT_UNLOCK(object);
1301 			object = backing_object;
1302 		}
1303 		if (src_m == NULL)
1304 			panic("vm_fault_copy_wired: page missing");
1305 		pmap_copy_page(src_m, dst_m);
1306 		VM_OBJECT_UNLOCK(object);
1307 		dst_m->valid = VM_PAGE_BITS_ALL;
1308 		VM_OBJECT_UNLOCK(dst_object);
1309 
1310 		/*
1311 		 * Enter it in the pmap. If a wired, copy-on-write
1312 		 * mapping is being replaced by a write-enabled
1313 		 * mapping, then wire that new mapping.
1314 		 */
1315 		pmap_enter(dst_map->pmap, vaddr, access, dst_m, prot, upgrade);
1316 
1317 		/*
1318 		 * Mark it no longer busy, and put it on the active list.
1319 		 */
1320 		VM_OBJECT_LOCK(dst_object);
1321 
1322 		if (upgrade) {
1323 			vm_page_lock(src_m);
1324 			vm_page_unwire(src_m, 0);
1325 			vm_page_unlock(src_m);
1326 
1327 			vm_page_lock(dst_m);
1328 			vm_page_wire(dst_m);
1329 			vm_page_unlock(dst_m);
1330 		} else {
1331 			vm_page_lock(dst_m);
1332 			vm_page_activate(dst_m);
1333 			vm_page_unlock(dst_m);
1334 		}
1335 		vm_page_wakeup(dst_m);
1336 	}
1337 	VM_OBJECT_UNLOCK(dst_object);
1338 	if (upgrade) {
1339 		dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
1340 		vm_object_deallocate(src_object);
1341 	}
1342 }
1343 
1344 
1345 /*
1346  * This routine checks around the requested page for other pages that
1347  * might be able to be faulted in.  This routine brackets the viable
1348  * pages for the pages to be paged in.
1349  *
1350  * Inputs:
1351  *	m, rbehind, rahead
1352  *
1353  * Outputs:
1354  *  marray (array of vm_page_t), reqpage (index of requested page)
1355  *
1356  * Return value:
1357  *  number of pages in marray
1358  */
1359 static int
1360 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage)
1361 	vm_page_t m;
1362 	int rbehind;
1363 	int rahead;
1364 	vm_page_t *marray;
1365 	int *reqpage;
1366 {
1367 	int i,j;
1368 	vm_object_t object;
1369 	vm_pindex_t pindex, startpindex, endpindex, tpindex;
1370 	vm_page_t rtm;
1371 	int cbehind, cahead;
1372 
1373 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1374 
1375 	object = m->object;
1376 	pindex = m->pindex;
1377 	cbehind = cahead = 0;
1378 
1379 	/*
1380 	 * if the requested page is not available, then give up now
1381 	 */
1382 	if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1383 		return 0;
1384 	}
1385 
1386 	if ((cbehind == 0) && (cahead == 0)) {
1387 		*reqpage = 0;
1388 		marray[0] = m;
1389 		return 1;
1390 	}
1391 
1392 	if (rahead > cahead) {
1393 		rahead = cahead;
1394 	}
1395 
1396 	if (rbehind > cbehind) {
1397 		rbehind = cbehind;
1398 	}
1399 
1400 	/*
1401 	 * scan backward for the read behind pages -- in memory
1402 	 */
1403 	if (pindex > 0) {
1404 		if (rbehind > pindex) {
1405 			rbehind = pindex;
1406 			startpindex = 0;
1407 		} else {
1408 			startpindex = pindex - rbehind;
1409 		}
1410 
1411 		if ((rtm = TAILQ_PREV(m, pglist, listq)) != NULL &&
1412 		    rtm->pindex >= startpindex)
1413 			startpindex = rtm->pindex + 1;
1414 
1415 		/* tpindex is unsigned; beware of numeric underflow. */
1416 		for (i = 0, tpindex = pindex - 1; tpindex >= startpindex &&
1417 		    tpindex < pindex; i++, tpindex--) {
1418 
1419 			rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1420 			    VM_ALLOC_IFNOTCACHED);
1421 			if (rtm == NULL) {
1422 				/*
1423 				 * Shift the allocated pages to the
1424 				 * beginning of the array.
1425 				 */
1426 				for (j = 0; j < i; j++) {
1427 					marray[j] = marray[j + tpindex + 1 -
1428 					    startpindex];
1429 				}
1430 				break;
1431 			}
1432 
1433 			marray[tpindex - startpindex] = rtm;
1434 		}
1435 	} else {
1436 		startpindex = 0;
1437 		i = 0;
1438 	}
1439 
1440 	marray[i] = m;
1441 	/* page offset of the required page */
1442 	*reqpage = i;
1443 
1444 	tpindex = pindex + 1;
1445 	i++;
1446 
1447 	/*
1448 	 * scan forward for the read ahead pages
1449 	 */
1450 	endpindex = tpindex + rahead;
1451 	if ((rtm = TAILQ_NEXT(m, listq)) != NULL && rtm->pindex < endpindex)
1452 		endpindex = rtm->pindex;
1453 	if (endpindex > object->size)
1454 		endpindex = object->size;
1455 
1456 	for (; tpindex < endpindex; i++, tpindex++) {
1457 
1458 		rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1459 		    VM_ALLOC_IFNOTCACHED);
1460 		if (rtm == NULL) {
1461 			break;
1462 		}
1463 
1464 		marray[i] = rtm;
1465 	}
1466 
1467 	/* return number of pages */
1468 	return i;
1469 }
1470