xref: /freebsd/sys/vm/vm_fault.c (revision 682c9e0fed0115eb6f283e755901c0aac90e86e8)
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 	if ((curthread->td_pflags & TDP_NOFAULTING) != 0)
213 		return (KERN_PROTECTION_FAILURE);
214 	return (vm_fault_hold(map, vaddr, fault_type, fault_flags, NULL));
215 }
216 
217 int
218 vm_fault_hold(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
219     int fault_flags, vm_page_t *m_hold)
220 {
221 	vm_prot_t prot;
222 	int is_first_object_locked, result;
223 	boolean_t growstack, wired;
224 	int map_generation;
225 	vm_object_t next_object;
226 	vm_page_t marray[VM_FAULT_READ], mt, mt_prev;
227 	int hardfault;
228 	int faultcount, ahead, behind, alloc_req;
229 	struct faultstate fs;
230 	struct vnode *vp;
231 	int locked, error;
232 
233 	hardfault = 0;
234 	growstack = TRUE;
235 	PCPU_INC(cnt.v_vm_faults);
236 	fs.vp = NULL;
237 	fs.vfslocked = 0;
238 	faultcount = behind = 0;
239 
240 RetryFault:;
241 
242 	/*
243 	 * Find the backing store object and offset into it to begin the
244 	 * search.
245 	 */
246 	fs.map = map;
247 	result = vm_map_lookup(&fs.map, vaddr, fault_type, &fs.entry,
248 	    &fs.first_object, &fs.first_pindex, &prot, &wired);
249 	if (result != KERN_SUCCESS) {
250 		if (growstack && result == KERN_INVALID_ADDRESS &&
251 		    map != kernel_map) {
252 			result = vm_map_growstack(curproc, vaddr);
253 			if (result != KERN_SUCCESS)
254 				return (KERN_FAILURE);
255 			growstack = FALSE;
256 			goto RetryFault;
257 		}
258 		return (result);
259 	}
260 
261 	map_generation = fs.map->timestamp;
262 
263 	if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
264 		panic("vm_fault: fault on nofault entry, addr: %lx",
265 		    (u_long)vaddr);
266 	}
267 
268 	/*
269 	 * Make a reference to this object to prevent its disposal while we
270 	 * are messing with it.  Once we have the reference, the map is free
271 	 * to be diddled.  Since objects reference their shadows (and copies),
272 	 * they will stay around as well.
273 	 *
274 	 * Bump the paging-in-progress count to prevent size changes (e.g.
275 	 * truncation operations) during I/O.  This must be done after
276 	 * obtaining the vnode lock in order to avoid possible deadlocks.
277 	 */
278 	VM_OBJECT_LOCK(fs.first_object);
279 	vm_object_reference_locked(fs.first_object);
280 	vm_object_pip_add(fs.first_object, 1);
281 
282 	fs.lookup_still_valid = TRUE;
283 
284 	if (wired)
285 		fault_type = prot | (fault_type & VM_PROT_COPY);
286 
287 	fs.first_m = NULL;
288 
289 	/*
290 	 * Search for the page at object/offset.
291 	 */
292 	fs.object = fs.first_object;
293 	fs.pindex = fs.first_pindex;
294 	while (TRUE) {
295 		/*
296 		 * If the object is dead, we stop here
297 		 */
298 		if (fs.object->flags & OBJ_DEAD) {
299 			unlock_and_deallocate(&fs);
300 			return (KERN_PROTECTION_FAILURE);
301 		}
302 
303 		/*
304 		 * See if page is resident
305 		 */
306 		fs.m = vm_page_lookup(fs.object, fs.pindex);
307 		if (fs.m != NULL) {
308 			/*
309 			 * check for page-based copy on write.
310 			 * We check fs.object == fs.first_object so
311 			 * as to ensure the legacy COW mechanism is
312 			 * used when the page in question is part of
313 			 * a shadow object.  Otherwise, vm_page_cowfault()
314 			 * removes the page from the backing object,
315 			 * which is not what we want.
316 			 */
317 			vm_page_lock(fs.m);
318 			if ((fs.m->cow) &&
319 			    (fault_type & VM_PROT_WRITE) &&
320 			    (fs.object == fs.first_object)) {
321 				vm_page_cowfault(fs.m);
322 				unlock_and_deallocate(&fs);
323 				goto RetryFault;
324 			}
325 
326 			/*
327 			 * Wait/Retry if the page is busy.  We have to do this
328 			 * if the page is busy via either VPO_BUSY or
329 			 * vm_page_t->busy because the vm_pager may be using
330 			 * vm_page_t->busy for pageouts ( and even pageins if
331 			 * it is the vnode pager ), and we could end up trying
332 			 * to pagein and pageout the same page simultaneously.
333 			 *
334 			 * We can theoretically allow the busy case on a read
335 			 * fault if the page is marked valid, but since such
336 			 * pages are typically already pmap'd, putting that
337 			 * special case in might be more effort then it is
338 			 * worth.  We cannot under any circumstances mess
339 			 * around with a vm_page_t->busy page except, perhaps,
340 			 * to pmap it.
341 			 */
342 			if ((fs.m->oflags & VPO_BUSY) || fs.m->busy) {
343 				/*
344 				 * Reference the page before unlocking and
345 				 * sleeping so that the page daemon is less
346 				 * likely to reclaim it.
347 				 */
348 				vm_page_lock_queues();
349 				vm_page_flag_set(fs.m, PG_REFERENCED);
350 				vm_page_unlock_queues();
351 				vm_page_unlock(fs.m);
352 				if (fs.object != fs.first_object) {
353 					if (!VM_OBJECT_TRYLOCK(
354 					    fs.first_object)) {
355 						VM_OBJECT_UNLOCK(fs.object);
356 						VM_OBJECT_LOCK(fs.first_object);
357 						VM_OBJECT_LOCK(fs.object);
358 					}
359 					vm_page_lock(fs.first_m);
360 					vm_page_free(fs.first_m);
361 					vm_page_unlock(fs.first_m);
362 					vm_object_pip_wakeup(fs.first_object);
363 					VM_OBJECT_UNLOCK(fs.first_object);
364 					fs.first_m = NULL;
365 				}
366 				unlock_map(&fs);
367 				if (fs.m == vm_page_lookup(fs.object,
368 				    fs.pindex)) {
369 					vm_page_sleep_if_busy(fs.m, TRUE,
370 					    "vmpfw");
371 				}
372 				vm_object_pip_wakeup(fs.object);
373 				VM_OBJECT_UNLOCK(fs.object);
374 				PCPU_INC(cnt.v_intrans);
375 				vm_object_deallocate(fs.first_object);
376 				goto RetryFault;
377 			}
378 			vm_pageq_remove(fs.m);
379 			vm_page_unlock(fs.m);
380 
381 			/*
382 			 * Mark page busy for other processes, and the
383 			 * pagedaemon.  If it still isn't completely valid
384 			 * (readable), jump to readrest, else break-out ( we
385 			 * found the page ).
386 			 */
387 			vm_page_busy(fs.m);
388 			if (fs.m->valid != VM_PAGE_BITS_ALL)
389 				goto readrest;
390 			break;
391 		}
392 
393 		/*
394 		 * Page is not resident, If this is the search termination
395 		 * or the pager might contain the page, allocate a new page.
396 		 */
397 		if (TRYPAGER || fs.object == fs.first_object) {
398 			if (fs.pindex >= fs.object->size) {
399 				unlock_and_deallocate(&fs);
400 				return (KERN_PROTECTION_FAILURE);
401 			}
402 
403 			/*
404 			 * Allocate a new page for this object/offset pair.
405 			 *
406 			 * Unlocked read of the p_flag is harmless. At
407 			 * worst, the P_KILLED might be not observed
408 			 * there, and allocation can fail, causing
409 			 * restart and new reading of the p_flag.
410 			 */
411 			fs.m = NULL;
412 			if (!vm_page_count_severe() || P_KILLED(curproc)) {
413 #if VM_NRESERVLEVEL > 0
414 				if ((fs.object->flags & OBJ_COLORED) == 0) {
415 					fs.object->flags |= OBJ_COLORED;
416 					fs.object->pg_color = atop(vaddr) -
417 					    fs.pindex;
418 				}
419 #endif
420 				alloc_req = P_KILLED(curproc) ?
421 				    VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL;
422 				if (fs.object->type != OBJT_VNODE &&
423 				    fs.object->backing_object == NULL)
424 					alloc_req |= VM_ALLOC_ZERO;
425 				fs.m = vm_page_alloc(fs.object, fs.pindex,
426 				    alloc_req);
427 			}
428 			if (fs.m == NULL) {
429 				unlock_and_deallocate(&fs);
430 				VM_WAITPFAULT;
431 				goto RetryFault;
432 			} else if (fs.m->valid == VM_PAGE_BITS_ALL)
433 				break;
434 		}
435 
436 readrest:
437 		/*
438 		 * We have found a valid page or we have allocated a new page.
439 		 * The page thus may not be valid or may not be entirely
440 		 * valid.
441 		 *
442 		 * Attempt to fault-in the page if there is a chance that the
443 		 * pager has it, and potentially fault in additional pages
444 		 * at the same time.
445 		 */
446 		if (TRYPAGER) {
447 			int rv;
448 			int reqpage = 0;
449 			u_char behavior = vm_map_entry_behavior(fs.entry);
450 
451 			if (behavior == MAP_ENTRY_BEHAV_RANDOM ||
452 			    P_KILLED(curproc)) {
453 				ahead = 0;
454 				behind = 0;
455 			} else {
456 				behind = (vaddr - fs.entry->start) >> PAGE_SHIFT;
457 				if (behind > VM_FAULT_READ_BEHIND)
458 					behind = VM_FAULT_READ_BEHIND;
459 
460 				ahead = ((fs.entry->end - vaddr) >> PAGE_SHIFT) - 1;
461 				if (ahead > VM_FAULT_READ_AHEAD)
462 					ahead = VM_FAULT_READ_AHEAD;
463 			}
464 			is_first_object_locked = FALSE;
465 			if ((behavior == MAP_ENTRY_BEHAV_SEQUENTIAL ||
466 			     (behavior != MAP_ENTRY_BEHAV_RANDOM &&
467 			      fs.pindex >= fs.entry->lastr &&
468 			      fs.pindex < fs.entry->lastr + VM_FAULT_READ)) &&
469 			    (fs.first_object == fs.object ||
470 			     (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object))) &&
471 			    fs.first_object->type != OBJT_DEVICE &&
472 			    fs.first_object->type != OBJT_PHYS &&
473 			    fs.first_object->type != OBJT_SG) {
474 				vm_pindex_t firstpindex;
475 
476 				if (fs.first_pindex < 2 * VM_FAULT_READ)
477 					firstpindex = 0;
478 				else
479 					firstpindex = fs.first_pindex - 2 * VM_FAULT_READ;
480 				mt = fs.first_object != fs.object ?
481 				    fs.first_m : fs.m;
482 				KASSERT(mt != NULL, ("vm_fault: missing mt"));
483 				KASSERT((mt->oflags & VPO_BUSY) != 0,
484 				    ("vm_fault: mt %p not busy", mt));
485 				mt_prev = vm_page_prev(mt);
486 
487 				/*
488 				 * note: partially valid pages cannot be
489 				 * included in the lookahead - NFS piecemeal
490 				 * writes will barf on it badly.
491 				 */
492 				while ((mt = mt_prev) != NULL &&
493 				    mt->pindex >= firstpindex &&
494 				    mt->valid == VM_PAGE_BITS_ALL) {
495 					mt_prev = vm_page_prev(mt);
496 					if (mt->busy ||
497 					    (mt->oflags & VPO_BUSY))
498 						continue;
499 					vm_page_lock(mt);
500 					if (mt->hold_count ||
501 					    mt->wire_count) {
502 						vm_page_unlock(mt);
503 						continue;
504 					}
505 					pmap_remove_all(mt);
506 					if (mt->dirty != 0)
507 						vm_page_deactivate(mt);
508 					else
509 						vm_page_cache(mt);
510 					vm_page_unlock(mt);
511 				}
512 				ahead += behind;
513 				behind = 0;
514 			}
515 			if (is_first_object_locked)
516 				VM_OBJECT_UNLOCK(fs.first_object);
517 
518 			/*
519 			 * Call the pager to retrieve the data, if any, after
520 			 * releasing the lock on the map.  We hold a ref on
521 			 * fs.object and the pages are VPO_BUSY'd.
522 			 */
523 			unlock_map(&fs);
524 
525 vnode_lock:
526 			if (fs.object->type == OBJT_VNODE) {
527 				vp = fs.object->handle;
528 				if (vp == fs.vp)
529 					goto vnode_locked;
530 				else if (fs.vp != NULL) {
531 					vput(fs.vp);
532 					fs.vp = NULL;
533 				}
534 				locked = VOP_ISLOCKED(vp);
535 
536 				if (VFS_NEEDSGIANT(vp->v_mount) && !fs.vfslocked) {
537 					fs.vfslocked = 1;
538 					if (!mtx_trylock(&Giant)) {
539 						VM_OBJECT_UNLOCK(fs.object);
540 						mtx_lock(&Giant);
541 						VM_OBJECT_LOCK(fs.object);
542 						goto vnode_lock;
543 					}
544 				}
545 				if (locked != LK_EXCLUSIVE)
546 					locked = LK_SHARED;
547 				/* Do not sleep for vnode lock while fs.m is busy */
548 				error = vget(vp, locked | LK_CANRECURSE |
549 				    LK_NOWAIT, curthread);
550 				if (error != 0) {
551 					int vfslocked;
552 
553 					vfslocked = fs.vfslocked;
554 					fs.vfslocked = 0; /* Keep Giant */
555 					vhold(vp);
556 					release_page(&fs);
557 					unlock_and_deallocate(&fs);
558 					error = vget(vp, locked | LK_RETRY |
559 					    LK_CANRECURSE, curthread);
560 					vdrop(vp);
561 					fs.vp = vp;
562 					fs.vfslocked = vfslocked;
563 					KASSERT(error == 0,
564 					    ("vm_fault: vget failed"));
565 					goto RetryFault;
566 				}
567 				fs.vp = vp;
568 			}
569 vnode_locked:
570 			KASSERT(fs.vp == NULL || !fs.map->system_map,
571 			    ("vm_fault: vnode-backed object mapped by system map"));
572 
573 			/*
574 			 * now we find out if any other pages should be paged
575 			 * in at this time this routine checks to see if the
576 			 * pages surrounding this fault reside in the same
577 			 * object as the page for this fault.  If they do,
578 			 * then they are faulted in also into the object.  The
579 			 * array "marray" returned contains an array of
580 			 * vm_page_t structs where one of them is the
581 			 * vm_page_t passed to the routine.  The reqpage
582 			 * return value is the index into the marray for the
583 			 * vm_page_t passed to the routine.
584 			 *
585 			 * fs.m plus the additional pages are VPO_BUSY'd.
586 			 */
587 			faultcount = vm_fault_additional_pages(
588 			    fs.m, behind, ahead, marray, &reqpage);
589 
590 			rv = faultcount ?
591 			    vm_pager_get_pages(fs.object, marray, faultcount,
592 				reqpage) : VM_PAGER_FAIL;
593 
594 			if (rv == VM_PAGER_OK) {
595 				/*
596 				 * Found the page. Leave it busy while we play
597 				 * with it.
598 				 */
599 
600 				/*
601 				 * Relookup in case pager changed page. Pager
602 				 * is responsible for disposition of old page
603 				 * if moved.
604 				 */
605 				fs.m = vm_page_lookup(fs.object, fs.pindex);
606 				if (!fs.m) {
607 					unlock_and_deallocate(&fs);
608 					goto RetryFault;
609 				}
610 
611 				hardfault++;
612 				break; /* break to PAGE HAS BEEN FOUND */
613 			}
614 			/*
615 			 * Remove the bogus page (which does not exist at this
616 			 * object/offset); before doing so, we must get back
617 			 * our object lock to preserve our invariant.
618 			 *
619 			 * Also wake up any other process that may want to bring
620 			 * in this page.
621 			 *
622 			 * If this is the top-level object, we must leave the
623 			 * busy page to prevent another process from rushing
624 			 * past us, and inserting the page in that object at
625 			 * the same time that we are.
626 			 */
627 			if (rv == VM_PAGER_ERROR)
628 				printf("vm_fault: pager read error, pid %d (%s)\n",
629 				    curproc->p_pid, curproc->p_comm);
630 			/*
631 			 * Data outside the range of the pager or an I/O error
632 			 */
633 			/*
634 			 * XXX - the check for kernel_map is a kludge to work
635 			 * around having the machine panic on a kernel space
636 			 * fault w/ I/O error.
637 			 */
638 			if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) ||
639 				(rv == VM_PAGER_BAD)) {
640 				vm_page_lock(fs.m);
641 				vm_page_free(fs.m);
642 				vm_page_unlock(fs.m);
643 				fs.m = NULL;
644 				unlock_and_deallocate(&fs);
645 				return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
646 			}
647 			if (fs.object != fs.first_object) {
648 				vm_page_lock(fs.m);
649 				vm_page_free(fs.m);
650 				vm_page_unlock(fs.m);
651 				fs.m = NULL;
652 				/*
653 				 * XXX - we cannot just fall out at this
654 				 * point, m has been freed and is invalid!
655 				 */
656 			}
657 		}
658 
659 		/*
660 		 * We get here if the object has default pager (or unwiring)
661 		 * or the pager doesn't have the page.
662 		 */
663 		if (fs.object == fs.first_object)
664 			fs.first_m = fs.m;
665 
666 		/*
667 		 * Move on to the next object.  Lock the next object before
668 		 * unlocking the current one.
669 		 */
670 		fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset);
671 		next_object = fs.object->backing_object;
672 		if (next_object == NULL) {
673 			/*
674 			 * If there's no object left, fill the page in the top
675 			 * object with zeros.
676 			 */
677 			if (fs.object != fs.first_object) {
678 				vm_object_pip_wakeup(fs.object);
679 				VM_OBJECT_UNLOCK(fs.object);
680 
681 				fs.object = fs.first_object;
682 				fs.pindex = fs.first_pindex;
683 				fs.m = fs.first_m;
684 				VM_OBJECT_LOCK(fs.object);
685 			}
686 			fs.first_m = NULL;
687 
688 			/*
689 			 * Zero the page if necessary and mark it valid.
690 			 */
691 			if ((fs.m->flags & PG_ZERO) == 0) {
692 				pmap_zero_page(fs.m);
693 			} else {
694 				PCPU_INC(cnt.v_ozfod);
695 			}
696 			PCPU_INC(cnt.v_zfod);
697 			fs.m->valid = VM_PAGE_BITS_ALL;
698 			break;	/* break to PAGE HAS BEEN FOUND */
699 		} else {
700 			KASSERT(fs.object != next_object,
701 			    ("object loop %p", next_object));
702 			VM_OBJECT_LOCK(next_object);
703 			vm_object_pip_add(next_object, 1);
704 			if (fs.object != fs.first_object)
705 				vm_object_pip_wakeup(fs.object);
706 			VM_OBJECT_UNLOCK(fs.object);
707 			fs.object = next_object;
708 		}
709 	}
710 
711 	KASSERT((fs.m->oflags & VPO_BUSY) != 0,
712 	    ("vm_fault: not busy after main loop"));
713 
714 	/*
715 	 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
716 	 * is held.]
717 	 */
718 
719 	/*
720 	 * If the page is being written, but isn't already owned by the
721 	 * top-level object, we have to copy it into a new page owned by the
722 	 * top-level object.
723 	 */
724 	if (fs.object != fs.first_object) {
725 		/*
726 		 * We only really need to copy if we want to write it.
727 		 */
728 		if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
729 			/*
730 			 * This allows pages to be virtually copied from a
731 			 * backing_object into the first_object, where the
732 			 * backing object has no other refs to it, and cannot
733 			 * gain any more refs.  Instead of a bcopy, we just
734 			 * move the page from the backing object to the
735 			 * first object.  Note that we must mark the page
736 			 * dirty in the first object so that it will go out
737 			 * to swap when needed.
738 			 */
739 			is_first_object_locked = FALSE;
740 			if (
741 				/*
742 				 * Only one shadow object
743 				 */
744 				(fs.object->shadow_count == 1) &&
745 				/*
746 				 * No COW refs, except us
747 				 */
748 				(fs.object->ref_count == 1) &&
749 				/*
750 				 * No one else can look this object up
751 				 */
752 				(fs.object->handle == NULL) &&
753 				/*
754 				 * No other ways to look the object up
755 				 */
756 				((fs.object->type == OBJT_DEFAULT) ||
757 				 (fs.object->type == OBJT_SWAP)) &&
758 			    (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object)) &&
759 				/*
760 				 * We don't chase down the shadow chain
761 				 */
762 			    fs.object == fs.first_object->backing_object) {
763 				/*
764 				 * get rid of the unnecessary page
765 				 */
766 				vm_page_lock(fs.first_m);
767 				vm_page_free(fs.first_m);
768 				vm_page_unlock(fs.first_m);
769 				/*
770 				 * grab the page and put it into the
771 				 * process'es object.  The page is
772 				 * automatically made dirty.
773 				 */
774 				vm_page_lock(fs.m);
775 				vm_page_rename(fs.m, fs.first_object, fs.first_pindex);
776 				vm_page_unlock(fs.m);
777 				vm_page_busy(fs.m);
778 				fs.first_m = fs.m;
779 				fs.m = NULL;
780 				PCPU_INC(cnt.v_cow_optim);
781 			} else {
782 				/*
783 				 * Oh, well, lets copy it.
784 				 */
785 				pmap_copy_page(fs.m, fs.first_m);
786 				fs.first_m->valid = VM_PAGE_BITS_ALL;
787 				if (wired && (fault_flags &
788 				    VM_FAULT_CHANGE_WIRING) == 0) {
789 					vm_page_lock(fs.first_m);
790 					vm_page_wire(fs.first_m);
791 					vm_page_unlock(fs.first_m);
792 
793 					vm_page_lock(fs.m);
794 					vm_page_unwire(fs.m, FALSE);
795 					vm_page_unlock(fs.m);
796 				}
797 				/*
798 				 * We no longer need the old page or object.
799 				 */
800 				release_page(&fs);
801 			}
802 			/*
803 			 * fs.object != fs.first_object due to above
804 			 * conditional
805 			 */
806 			vm_object_pip_wakeup(fs.object);
807 			VM_OBJECT_UNLOCK(fs.object);
808 			/*
809 			 * Only use the new page below...
810 			 */
811 			fs.object = fs.first_object;
812 			fs.pindex = fs.first_pindex;
813 			fs.m = fs.first_m;
814 			if (!is_first_object_locked)
815 				VM_OBJECT_LOCK(fs.object);
816 			PCPU_INC(cnt.v_cow_faults);
817 		} else {
818 			prot &= ~VM_PROT_WRITE;
819 		}
820 	}
821 
822 	/*
823 	 * We must verify that the maps have not changed since our last
824 	 * lookup.
825 	 */
826 	if (!fs.lookup_still_valid) {
827 		vm_object_t retry_object;
828 		vm_pindex_t retry_pindex;
829 		vm_prot_t retry_prot;
830 
831 		if (!vm_map_trylock_read(fs.map)) {
832 			release_page(&fs);
833 			unlock_and_deallocate(&fs);
834 			goto RetryFault;
835 		}
836 		fs.lookup_still_valid = TRUE;
837 		if (fs.map->timestamp != map_generation) {
838 			result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
839 			    &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
840 
841 			/*
842 			 * If we don't need the page any longer, put it on the inactive
843 			 * list (the easiest thing to do here).  If no one needs it,
844 			 * pageout will grab it eventually.
845 			 */
846 			if (result != KERN_SUCCESS) {
847 				release_page(&fs);
848 				unlock_and_deallocate(&fs);
849 
850 				/*
851 				 * If retry of map lookup would have blocked then
852 				 * retry fault from start.
853 				 */
854 				if (result == KERN_FAILURE)
855 					goto RetryFault;
856 				return (result);
857 			}
858 			if ((retry_object != fs.first_object) ||
859 			    (retry_pindex != fs.first_pindex)) {
860 				release_page(&fs);
861 				unlock_and_deallocate(&fs);
862 				goto RetryFault;
863 			}
864 
865 			/*
866 			 * Check whether the protection has changed or the object has
867 			 * been copied while we left the map unlocked. Changing from
868 			 * read to write permission is OK - we leave the page
869 			 * write-protected, and catch the write fault. Changing from
870 			 * write to read permission means that we can't mark the page
871 			 * write-enabled after all.
872 			 */
873 			prot &= retry_prot;
874 		}
875 	}
876 	/*
877 	 * If the page was filled by a pager, update the map entry's
878 	 * last read offset.  Since the pager does not return the
879 	 * actual set of pages that it read, this update is based on
880 	 * the requested set.  Typically, the requested and actual
881 	 * sets are the same.
882 	 *
883 	 * XXX The following assignment modifies the map
884 	 * without holding a write lock on it.
885 	 */
886 	if (hardfault)
887 		fs.entry->lastr = fs.pindex + faultcount - behind;
888 
889 	if ((prot & VM_PROT_WRITE) != 0 ||
890 	    (fault_flags & VM_FAULT_DIRTY) != 0) {
891 		vm_object_set_writeable_dirty(fs.object);
892 
893 		/*
894 		 * If this is a NOSYNC mmap we do not want to set VPO_NOSYNC
895 		 * if the page is already dirty to prevent data written with
896 		 * the expectation of being synced from not being synced.
897 		 * Likewise if this entry does not request NOSYNC then make
898 		 * sure the page isn't marked NOSYNC.  Applications sharing
899 		 * data should use the same flags to avoid ping ponging.
900 		 */
901 		if (fs.entry->eflags & MAP_ENTRY_NOSYNC) {
902 			if (fs.m->dirty == 0)
903 				fs.m->oflags |= VPO_NOSYNC;
904 		} else {
905 			fs.m->oflags &= ~VPO_NOSYNC;
906 		}
907 
908 		/*
909 		 * If the fault is a write, we know that this page is being
910 		 * written NOW so dirty it explicitly to save on
911 		 * pmap_is_modified() calls later.
912 		 *
913 		 * Also tell the backing pager, if any, that it should remove
914 		 * any swap backing since the page is now dirty.
915 		 */
916 		if (((fault_type & VM_PROT_WRITE) != 0 &&
917 		    (fault_flags & VM_FAULT_CHANGE_WIRING) == 0) ||
918 		    (fault_flags & VM_FAULT_DIRTY) != 0) {
919 			vm_page_dirty(fs.m);
920 			vm_pager_page_unswapped(fs.m);
921 		}
922 	}
923 
924 	/*
925 	 * Page had better still be busy
926 	 */
927 	KASSERT(fs.m->oflags & VPO_BUSY,
928 		("vm_fault: page %p not busy!", fs.m));
929 	/*
930 	 * Page must be completely valid or it is not fit to
931 	 * map into user space.  vm_pager_get_pages() ensures this.
932 	 */
933 	KASSERT(fs.m->valid == VM_PAGE_BITS_ALL,
934 	    ("vm_fault: page %p partially invalid", fs.m));
935 	VM_OBJECT_UNLOCK(fs.object);
936 
937 	/*
938 	 * Put this page into the physical map.  We had to do the unlock above
939 	 * because pmap_enter() may sleep.  We don't put the page
940 	 * back on the active queue until later so that the pageout daemon
941 	 * won't find it (yet).
942 	 */
943 	pmap_enter(fs.map->pmap, vaddr, fault_type, fs.m, prot, wired);
944 	if ((fault_flags & VM_FAULT_CHANGE_WIRING) == 0 && wired == 0)
945 		vm_fault_prefault(fs.map->pmap, vaddr, fs.entry);
946 	VM_OBJECT_LOCK(fs.object);
947 	vm_page_lock(fs.m);
948 
949 	/*
950 	 * If the page is not wired down, then put it where the pageout daemon
951 	 * can find it.
952 	 */
953 	if (fault_flags & VM_FAULT_CHANGE_WIRING) {
954 		if (wired)
955 			vm_page_wire(fs.m);
956 		else
957 			vm_page_unwire(fs.m, 1);
958 	} else
959 		vm_page_activate(fs.m);
960 	if (m_hold != NULL) {
961 		*m_hold = fs.m;
962 		vm_page_hold(fs.m);
963 	}
964 	vm_page_unlock(fs.m);
965 	vm_page_wakeup(fs.m);
966 
967 	/*
968 	 * Unlock everything, and return
969 	 */
970 	unlock_and_deallocate(&fs);
971 	if (hardfault)
972 		curthread->td_ru.ru_majflt++;
973 	else
974 		curthread->td_ru.ru_minflt++;
975 
976 	return (KERN_SUCCESS);
977 }
978 
979 /*
980  * vm_fault_prefault provides a quick way of clustering
981  * pagefaults into a processes address space.  It is a "cousin"
982  * of vm_map_pmap_enter, except it runs at page fault time instead
983  * of mmap time.
984  */
985 static void
986 vm_fault_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry)
987 {
988 	int i;
989 	vm_offset_t addr, starta;
990 	vm_pindex_t pindex;
991 	vm_page_t m;
992 	vm_object_t object;
993 
994 	if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
995 		return;
996 
997 	object = entry->object.vm_object;
998 
999 	starta = addra - PFBAK * PAGE_SIZE;
1000 	if (starta < entry->start) {
1001 		starta = entry->start;
1002 	} else if (starta > addra) {
1003 		starta = 0;
1004 	}
1005 
1006 	for (i = 0; i < PAGEORDER_SIZE; i++) {
1007 		vm_object_t backing_object, lobject;
1008 
1009 		addr = addra + prefault_pageorder[i];
1010 		if (addr > addra + (PFFOR * PAGE_SIZE))
1011 			addr = 0;
1012 
1013 		if (addr < starta || addr >= entry->end)
1014 			continue;
1015 
1016 		if (!pmap_is_prefaultable(pmap, addr))
1017 			continue;
1018 
1019 		pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1020 		lobject = object;
1021 		VM_OBJECT_LOCK(lobject);
1022 		while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1023 		    lobject->type == OBJT_DEFAULT &&
1024 		    (backing_object = lobject->backing_object) != NULL) {
1025 			KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1026 			    0, ("vm_fault_prefault: unaligned object offset"));
1027 			pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1028 			VM_OBJECT_LOCK(backing_object);
1029 			VM_OBJECT_UNLOCK(lobject);
1030 			lobject = backing_object;
1031 		}
1032 		/*
1033 		 * give-up when a page is not in memory
1034 		 */
1035 		if (m == NULL) {
1036 			VM_OBJECT_UNLOCK(lobject);
1037 			break;
1038 		}
1039 		if (m->valid == VM_PAGE_BITS_ALL &&
1040 		    (m->flags & PG_FICTITIOUS) == 0)
1041 			pmap_enter_quick(pmap, addr, m, entry->protection);
1042 		VM_OBJECT_UNLOCK(lobject);
1043 	}
1044 }
1045 
1046 /*
1047  * Hold each of the physical pages that are mapped by the specified range of
1048  * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
1049  * and allow the specified types of access, "prot".  If all of the implied
1050  * pages are successfully held, then the number of held pages is returned
1051  * together with pointers to those pages in the array "ma".  However, if any
1052  * of the pages cannot be held, -1 is returned.
1053  */
1054 int
1055 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
1056     vm_prot_t prot, vm_page_t *ma, int max_count)
1057 {
1058 	vm_offset_t end, va;
1059 	vm_page_t *mp;
1060 	int count;
1061 	boolean_t pmap_failed;
1062 
1063 	if (len == 0)
1064 		return (0);
1065 	end = round_page(addr + len);
1066 	addr = trunc_page(addr);
1067 
1068 	/*
1069 	 * Check for illegal addresses.
1070 	 */
1071 	if (addr < vm_map_min(map) || addr > end || end > vm_map_max(map))
1072 		return (-1);
1073 
1074 	count = howmany(end - addr, PAGE_SIZE);
1075 	if (count > max_count)
1076 		panic("vm_fault_quick_hold_pages: count > max_count");
1077 
1078 	/*
1079 	 * Most likely, the physical pages are resident in the pmap, so it is
1080 	 * faster to try pmap_extract_and_hold() first.
1081 	 */
1082 	pmap_failed = FALSE;
1083 	for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) {
1084 		*mp = pmap_extract_and_hold(map->pmap, va, prot);
1085 		if (*mp == NULL)
1086 			pmap_failed = TRUE;
1087 		else if ((prot & VM_PROT_WRITE) != 0 &&
1088 		    (*mp)->dirty != VM_PAGE_BITS_ALL) {
1089 			/*
1090 			 * Explicitly dirty the physical page.  Otherwise, the
1091 			 * caller's changes may go unnoticed because they are
1092 			 * performed through an unmanaged mapping or by a DMA
1093 			 * operation.
1094 			 *
1095 			 * The object lock is not held here.  Therefore, like
1096 			 * a pmap operation, the page queues lock may be
1097 			 * required in order to call vm_page_dirty().  See
1098 			 * vm_page_clear_dirty_mask().
1099 			 */
1100 #if defined(__amd64__) || defined(__i386__) || defined(__ia64__)
1101 			vm_page_dirty(*mp);
1102 #else
1103 			vm_page_lock_queues();
1104 			vm_page_dirty(*mp);
1105 			vm_page_unlock_queues();
1106 #endif
1107 		}
1108 	}
1109 	if (pmap_failed) {
1110 		/*
1111 		 * One or more pages could not be held by the pmap.  Either no
1112 		 * page was mapped at the specified virtual address or that
1113 		 * mapping had insufficient permissions.  Attempt to fault in
1114 		 * and hold these pages.
1115 		 */
1116 		for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE)
1117 			if (*mp == NULL && vm_fault_hold(map, va, prot,
1118 			    VM_FAULT_NORMAL, mp) != KERN_SUCCESS)
1119 				goto error;
1120 	}
1121 	return (count);
1122 error:
1123 	for (mp = ma; mp < ma + count; mp++)
1124 		if (*mp != NULL) {
1125 			vm_page_lock(*mp);
1126 			vm_page_unhold(*mp);
1127 			vm_page_unlock(*mp);
1128 		}
1129 	return (-1);
1130 }
1131 
1132 /*
1133  *	vm_fault_wire:
1134  *
1135  *	Wire down a range of virtual addresses in a map.
1136  */
1137 int
1138 vm_fault_wire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1139     boolean_t fictitious)
1140 {
1141 	vm_offset_t va;
1142 	int rv;
1143 
1144 	/*
1145 	 * We simulate a fault to get the page and enter it in the physical
1146 	 * map.  For user wiring, we only ask for read access on currently
1147 	 * read-only sections.
1148 	 */
1149 	for (va = start; va < end; va += PAGE_SIZE) {
1150 		rv = vm_fault(map, va, VM_PROT_NONE, VM_FAULT_CHANGE_WIRING);
1151 		if (rv) {
1152 			if (va != start)
1153 				vm_fault_unwire(map, start, va, fictitious);
1154 			return (rv);
1155 		}
1156 	}
1157 	return (KERN_SUCCESS);
1158 }
1159 
1160 /*
1161  *	vm_fault_unwire:
1162  *
1163  *	Unwire a range of virtual addresses in a map.
1164  */
1165 void
1166 vm_fault_unwire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1167     boolean_t fictitious)
1168 {
1169 	vm_paddr_t pa;
1170 	vm_offset_t va;
1171 	vm_page_t m;
1172 	pmap_t pmap;
1173 
1174 	pmap = vm_map_pmap(map);
1175 
1176 	/*
1177 	 * Since the pages are wired down, we must be able to get their
1178 	 * mappings from the physical map system.
1179 	 */
1180 	for (va = start; va < end; va += PAGE_SIZE) {
1181 		pa = pmap_extract(pmap, va);
1182 		if (pa != 0) {
1183 			pmap_change_wiring(pmap, va, FALSE);
1184 			if (!fictitious) {
1185 				m = PHYS_TO_VM_PAGE(pa);
1186 				vm_page_lock(m);
1187 				vm_page_unwire(m, TRUE);
1188 				vm_page_unlock(m);
1189 			}
1190 		}
1191 	}
1192 }
1193 
1194 /*
1195  *	Routine:
1196  *		vm_fault_copy_entry
1197  *	Function:
1198  *		Create new shadow object backing dst_entry with private copy of
1199  *		all underlying pages. When src_entry is equal to dst_entry,
1200  *		function implements COW for wired-down map entry. Otherwise,
1201  *		it forks wired entry into dst_map.
1202  *
1203  *	In/out conditions:
1204  *		The source and destination maps must be locked for write.
1205  *		The source map entry must be wired down (or be a sharing map
1206  *		entry corresponding to a main map entry that is wired down).
1207  */
1208 void
1209 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1210     vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1211     vm_ooffset_t *fork_charge)
1212 {
1213 	vm_object_t backing_object, dst_object, object, src_object;
1214 	vm_pindex_t dst_pindex, pindex, src_pindex;
1215 	vm_prot_t access, prot;
1216 	vm_offset_t vaddr;
1217 	vm_page_t dst_m;
1218 	vm_page_t src_m;
1219 	boolean_t src_readonly, upgrade;
1220 
1221 #ifdef	lint
1222 	src_map++;
1223 #endif	/* lint */
1224 
1225 	upgrade = src_entry == dst_entry;
1226 
1227 	src_object = src_entry->object.vm_object;
1228 	src_pindex = OFF_TO_IDX(src_entry->offset);
1229 	src_readonly = (src_entry->protection & VM_PROT_WRITE) == 0;
1230 
1231 	/*
1232 	 * Create the top-level object for the destination entry. (Doesn't
1233 	 * actually shadow anything - we copy the pages directly.)
1234 	 */
1235 	dst_object = vm_object_allocate(OBJT_DEFAULT,
1236 	    OFF_TO_IDX(dst_entry->end - dst_entry->start));
1237 #if VM_NRESERVLEVEL > 0
1238 	dst_object->flags |= OBJ_COLORED;
1239 	dst_object->pg_color = atop(dst_entry->start);
1240 #endif
1241 
1242 	VM_OBJECT_LOCK(dst_object);
1243 	KASSERT(upgrade || dst_entry->object.vm_object == NULL,
1244 	    ("vm_fault_copy_entry: vm_object not NULL"));
1245 	dst_entry->object.vm_object = dst_object;
1246 	dst_entry->offset = 0;
1247 	dst_object->charge = dst_entry->end - dst_entry->start;
1248 	if (fork_charge != NULL) {
1249 		KASSERT(dst_entry->cred == NULL,
1250 		    ("vm_fault_copy_entry: leaked swp charge"));
1251 		dst_object->cred = curthread->td_ucred;
1252 		crhold(dst_object->cred);
1253 		*fork_charge += dst_object->charge;
1254 	} else {
1255 		dst_object->cred = dst_entry->cred;
1256 		dst_entry->cred = NULL;
1257 	}
1258 	access = prot = dst_entry->protection;
1259 	/*
1260 	 * If not an upgrade, then enter the mappings in the pmap as
1261 	 * read and/or execute accesses.  Otherwise, enter them as
1262 	 * write accesses.
1263 	 *
1264 	 * A writeable large page mapping is only created if all of
1265 	 * the constituent small page mappings are modified. Marking
1266 	 * PTEs as modified on inception allows promotion to happen
1267 	 * without taking potentially large number of soft faults.
1268 	 */
1269 	if (!upgrade)
1270 		access &= ~VM_PROT_WRITE;
1271 
1272 	/*
1273 	 * Loop through all of the pages in the entry's range, copying each
1274 	 * one from the source object (it should be there) to the destination
1275 	 * object.
1276 	 */
1277 	for (vaddr = dst_entry->start, dst_pindex = 0;
1278 	    vaddr < dst_entry->end;
1279 	    vaddr += PAGE_SIZE, dst_pindex++) {
1280 
1281 		/*
1282 		 * Allocate a page in the destination object.
1283 		 */
1284 		do {
1285 			dst_m = vm_page_alloc(dst_object, dst_pindex,
1286 			    VM_ALLOC_NORMAL);
1287 			if (dst_m == NULL) {
1288 				VM_OBJECT_UNLOCK(dst_object);
1289 				VM_WAIT;
1290 				VM_OBJECT_LOCK(dst_object);
1291 			}
1292 		} while (dst_m == NULL);
1293 
1294 		/*
1295 		 * Find the page in the source object, and copy it in.
1296 		 * (Because the source is wired down, the page will be in
1297 		 * memory.)
1298 		 */
1299 		VM_OBJECT_LOCK(src_object);
1300 		object = src_object;
1301 		pindex = src_pindex + dst_pindex;
1302 		while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
1303 		    src_readonly &&
1304 		    (backing_object = object->backing_object) != NULL) {
1305 			/*
1306 			 * Allow fallback to backing objects if we are reading.
1307 			 */
1308 			VM_OBJECT_LOCK(backing_object);
1309 			pindex += OFF_TO_IDX(object->backing_object_offset);
1310 			VM_OBJECT_UNLOCK(object);
1311 			object = backing_object;
1312 		}
1313 		if (src_m == NULL)
1314 			panic("vm_fault_copy_wired: page missing");
1315 		pmap_copy_page(src_m, dst_m);
1316 		VM_OBJECT_UNLOCK(object);
1317 		dst_m->valid = VM_PAGE_BITS_ALL;
1318 		VM_OBJECT_UNLOCK(dst_object);
1319 
1320 		/*
1321 		 * Enter it in the pmap. If a wired, copy-on-write
1322 		 * mapping is being replaced by a write-enabled
1323 		 * mapping, then wire that new mapping.
1324 		 */
1325 		pmap_enter(dst_map->pmap, vaddr, access, dst_m, prot, upgrade);
1326 
1327 		/*
1328 		 * Mark it no longer busy, and put it on the active list.
1329 		 */
1330 		VM_OBJECT_LOCK(dst_object);
1331 
1332 		if (upgrade) {
1333 			vm_page_lock(src_m);
1334 			vm_page_unwire(src_m, 0);
1335 			vm_page_unlock(src_m);
1336 
1337 			vm_page_lock(dst_m);
1338 			vm_page_wire(dst_m);
1339 			vm_page_unlock(dst_m);
1340 		} else {
1341 			vm_page_lock(dst_m);
1342 			vm_page_activate(dst_m);
1343 			vm_page_unlock(dst_m);
1344 		}
1345 		vm_page_wakeup(dst_m);
1346 	}
1347 	VM_OBJECT_UNLOCK(dst_object);
1348 	if (upgrade) {
1349 		dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
1350 		vm_object_deallocate(src_object);
1351 	}
1352 }
1353 
1354 
1355 /*
1356  * This routine checks around the requested page for other pages that
1357  * might be able to be faulted in.  This routine brackets the viable
1358  * pages for the pages to be paged in.
1359  *
1360  * Inputs:
1361  *	m, rbehind, rahead
1362  *
1363  * Outputs:
1364  *  marray (array of vm_page_t), reqpage (index of requested page)
1365  *
1366  * Return value:
1367  *  number of pages in marray
1368  */
1369 static int
1370 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage)
1371 	vm_page_t m;
1372 	int rbehind;
1373 	int rahead;
1374 	vm_page_t *marray;
1375 	int *reqpage;
1376 {
1377 	int i,j;
1378 	vm_object_t object;
1379 	vm_pindex_t pindex, startpindex, endpindex, tpindex;
1380 	vm_page_t rtm;
1381 	int cbehind, cahead;
1382 
1383 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1384 
1385 	object = m->object;
1386 	pindex = m->pindex;
1387 	cbehind = cahead = 0;
1388 
1389 	/*
1390 	 * if the requested page is not available, then give up now
1391 	 */
1392 	if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1393 		return 0;
1394 	}
1395 
1396 	if ((cbehind == 0) && (cahead == 0)) {
1397 		*reqpage = 0;
1398 		marray[0] = m;
1399 		return 1;
1400 	}
1401 
1402 	if (rahead > cahead) {
1403 		rahead = cahead;
1404 	}
1405 
1406 	if (rbehind > cbehind) {
1407 		rbehind = cbehind;
1408 	}
1409 
1410 	/*
1411 	 * scan backward for the read behind pages -- in memory
1412 	 */
1413 	if (pindex > 0) {
1414 		if (rbehind > pindex) {
1415 			rbehind = pindex;
1416 			startpindex = 0;
1417 		} else {
1418 			startpindex = pindex - rbehind;
1419 		}
1420 
1421 		if ((rtm = TAILQ_PREV(m, pglist, listq)) != NULL &&
1422 		    rtm->pindex >= startpindex)
1423 			startpindex = rtm->pindex + 1;
1424 
1425 		/* tpindex is unsigned; beware of numeric underflow. */
1426 		for (i = 0, tpindex = pindex - 1; tpindex >= startpindex &&
1427 		    tpindex < pindex; i++, tpindex--) {
1428 
1429 			rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1430 			    VM_ALLOC_IFNOTCACHED);
1431 			if (rtm == NULL) {
1432 				/*
1433 				 * Shift the allocated pages to the
1434 				 * beginning of the array.
1435 				 */
1436 				for (j = 0; j < i; j++) {
1437 					marray[j] = marray[j + tpindex + 1 -
1438 					    startpindex];
1439 				}
1440 				break;
1441 			}
1442 
1443 			marray[tpindex - startpindex] = rtm;
1444 		}
1445 	} else {
1446 		startpindex = 0;
1447 		i = 0;
1448 	}
1449 
1450 	marray[i] = m;
1451 	/* page offset of the required page */
1452 	*reqpage = i;
1453 
1454 	tpindex = pindex + 1;
1455 	i++;
1456 
1457 	/*
1458 	 * scan forward for the read ahead pages
1459 	 */
1460 	endpindex = tpindex + rahead;
1461 	if ((rtm = TAILQ_NEXT(m, listq)) != NULL && rtm->pindex < endpindex)
1462 		endpindex = rtm->pindex;
1463 	if (endpindex > object->size)
1464 		endpindex = object->size;
1465 
1466 	for (; tpindex < endpindex; i++, tpindex++) {
1467 
1468 		rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1469 		    VM_ALLOC_IFNOTCACHED);
1470 		if (rtm == NULL) {
1471 			break;
1472 		}
1473 
1474 		marray[i] = rtm;
1475 	}
1476 
1477 	/* return number of pages */
1478 	return i;
1479 }
1480 
1481 int
1482 vm_fault_disable_pagefaults(void)
1483 {
1484 
1485 	return (curthread_pflags_set(TDP_NOFAULTING));
1486 }
1487 
1488 void
1489 vm_fault_enable_pagefaults(int save)
1490 {
1491 
1492 	curthread_pflags_restore(save);
1493 }
1494