xref: /freebsd/sys/vm/vm_fault.c (revision a3cf0ef5a295c885c895fabfd56470c0d1db322d)
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  *
205  *	The map in question must be referenced, and remains so.
206  *	Caller may hold no locks.
207  */
208 int
209 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
210 	 int fault_flags)
211 {
212 	vm_prot_t prot;
213 	int is_first_object_locked, result;
214 	boolean_t growstack, wired;
215 	int map_generation;
216 	vm_object_t next_object;
217 	vm_page_t marray[VM_FAULT_READ], mt, mt_prev;
218 	int hardfault;
219 	int faultcount, ahead, behind, alloc_req;
220 	struct faultstate fs;
221 	struct vnode *vp;
222 	int locked, error;
223 
224 	hardfault = 0;
225 	growstack = TRUE;
226 	PCPU_INC(cnt.v_vm_faults);
227 	fs.vp = NULL;
228 	fs.vfslocked = 0;
229 	faultcount = behind = 0;
230 
231 RetryFault:;
232 
233 	/*
234 	 * Find the backing store object and offset into it to begin the
235 	 * search.
236 	 */
237 	fs.map = map;
238 	result = vm_map_lookup(&fs.map, vaddr, fault_type, &fs.entry,
239 	    &fs.first_object, &fs.first_pindex, &prot, &wired);
240 	if (result != KERN_SUCCESS) {
241 		if (growstack && result == KERN_INVALID_ADDRESS &&
242 		    map != kernel_map) {
243 			result = vm_map_growstack(curproc, vaddr);
244 			if (result != KERN_SUCCESS)
245 				return (KERN_FAILURE);
246 			growstack = FALSE;
247 			goto RetryFault;
248 		}
249 		return (result);
250 	}
251 
252 	map_generation = fs.map->timestamp;
253 
254 	if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
255 		panic("vm_fault: fault on nofault entry, addr: %lx",
256 		    (u_long)vaddr);
257 	}
258 
259 	/*
260 	 * Make a reference to this object to prevent its disposal while we
261 	 * are messing with it.  Once we have the reference, the map is free
262 	 * to be diddled.  Since objects reference their shadows (and copies),
263 	 * they will stay around as well.
264 	 *
265 	 * Bump the paging-in-progress count to prevent size changes (e.g.
266 	 * truncation operations) during I/O.  This must be done after
267 	 * obtaining the vnode lock in order to avoid possible deadlocks.
268 	 */
269 	VM_OBJECT_LOCK(fs.first_object);
270 	vm_object_reference_locked(fs.first_object);
271 	vm_object_pip_add(fs.first_object, 1);
272 
273 	fs.lookup_still_valid = TRUE;
274 
275 	if (wired)
276 		fault_type = prot | (fault_type & VM_PROT_COPY);
277 
278 	fs.first_m = NULL;
279 
280 	/*
281 	 * Search for the page at object/offset.
282 	 */
283 	fs.object = fs.first_object;
284 	fs.pindex = fs.first_pindex;
285 	while (TRUE) {
286 		/*
287 		 * If the object is dead, we stop here
288 		 */
289 		if (fs.object->flags & OBJ_DEAD) {
290 			unlock_and_deallocate(&fs);
291 			return (KERN_PROTECTION_FAILURE);
292 		}
293 
294 		/*
295 		 * See if page is resident
296 		 */
297 		fs.m = vm_page_lookup(fs.object, fs.pindex);
298 		if (fs.m != NULL) {
299 			/*
300 			 * check for page-based copy on write.
301 			 * We check fs.object == fs.first_object so
302 			 * as to ensure the legacy COW mechanism is
303 			 * used when the page in question is part of
304 			 * a shadow object.  Otherwise, vm_page_cowfault()
305 			 * removes the page from the backing object,
306 			 * which is not what we want.
307 			 */
308 			vm_page_lock(fs.m);
309 			if ((fs.m->cow) &&
310 			    (fault_type & VM_PROT_WRITE) &&
311 			    (fs.object == fs.first_object)) {
312 				vm_page_cowfault(fs.m);
313 				unlock_and_deallocate(&fs);
314 				goto RetryFault;
315 			}
316 
317 			/*
318 			 * Wait/Retry if the page is busy.  We have to do this
319 			 * if the page is busy via either VPO_BUSY or
320 			 * vm_page_t->busy because the vm_pager may be using
321 			 * vm_page_t->busy for pageouts ( and even pageins if
322 			 * it is the vnode pager ), and we could end up trying
323 			 * to pagein and pageout the same page simultaneously.
324 			 *
325 			 * We can theoretically allow the busy case on a read
326 			 * fault if the page is marked valid, but since such
327 			 * pages are typically already pmap'd, putting that
328 			 * special case in might be more effort then it is
329 			 * worth.  We cannot under any circumstances mess
330 			 * around with a vm_page_t->busy page except, perhaps,
331 			 * to pmap it.
332 			 */
333 			if ((fs.m->oflags & VPO_BUSY) || fs.m->busy) {
334 				/*
335 				 * Reference the page before unlocking and
336 				 * sleeping so that the page daemon is less
337 				 * likely to reclaim it.
338 				 */
339 				vm_page_lock_queues();
340 				vm_page_flag_set(fs.m, PG_REFERENCED);
341 				vm_page_unlock_queues();
342 				vm_page_unlock(fs.m);
343 				if (fs.object != fs.first_object) {
344 					if (!VM_OBJECT_TRYLOCK(
345 					    fs.first_object)) {
346 						VM_OBJECT_UNLOCK(fs.object);
347 						VM_OBJECT_LOCK(fs.first_object);
348 						VM_OBJECT_LOCK(fs.object);
349 					}
350 					vm_page_lock(fs.first_m);
351 					vm_page_free(fs.first_m);
352 					vm_page_unlock(fs.first_m);
353 					vm_object_pip_wakeup(fs.first_object);
354 					VM_OBJECT_UNLOCK(fs.first_object);
355 					fs.first_m = NULL;
356 				}
357 				unlock_map(&fs);
358 				if (fs.m == vm_page_lookup(fs.object,
359 				    fs.pindex)) {
360 					vm_page_sleep_if_busy(fs.m, TRUE,
361 					    "vmpfw");
362 				}
363 				vm_object_pip_wakeup(fs.object);
364 				VM_OBJECT_UNLOCK(fs.object);
365 				PCPU_INC(cnt.v_intrans);
366 				vm_object_deallocate(fs.first_object);
367 				goto RetryFault;
368 			}
369 			vm_pageq_remove(fs.m);
370 			vm_page_unlock(fs.m);
371 
372 			/*
373 			 * Mark page busy for other processes, and the
374 			 * pagedaemon.  If it still isn't completely valid
375 			 * (readable), jump to readrest, else break-out ( we
376 			 * found the page ).
377 			 */
378 			vm_page_busy(fs.m);
379 			if (fs.m->valid != VM_PAGE_BITS_ALL &&
380 				fs.m->object != kernel_object && fs.m->object != kmem_object) {
381 				goto readrest;
382 			}
383 
384 			break;
385 		}
386 
387 		/*
388 		 * Page is not resident, If this is the search termination
389 		 * or the pager might contain the page, allocate a new page.
390 		 */
391 		if (TRYPAGER || fs.object == fs.first_object) {
392 			if (fs.pindex >= fs.object->size) {
393 				unlock_and_deallocate(&fs);
394 				return (KERN_PROTECTION_FAILURE);
395 			}
396 
397 			/*
398 			 * Allocate a new page for this object/offset pair.
399 			 *
400 			 * Unlocked read of the p_flag is harmless. At
401 			 * worst, the P_KILLED might be not observed
402 			 * there, and allocation can fail, causing
403 			 * restart and new reading of the p_flag.
404 			 */
405 			fs.m = NULL;
406 			if (!vm_page_count_severe() || P_KILLED(curproc)) {
407 #if VM_NRESERVLEVEL > 0
408 				if ((fs.object->flags & OBJ_COLORED) == 0) {
409 					fs.object->flags |= OBJ_COLORED;
410 					fs.object->pg_color = atop(vaddr) -
411 					    fs.pindex;
412 				}
413 #endif
414 				alloc_req = P_KILLED(curproc) ?
415 				    VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL;
416 				if (fs.object->type != OBJT_VNODE &&
417 				    fs.object->backing_object == NULL)
418 					alloc_req |= VM_ALLOC_ZERO;
419 				fs.m = vm_page_alloc(fs.object, fs.pindex,
420 				    alloc_req);
421 			}
422 			if (fs.m == NULL) {
423 				unlock_and_deallocate(&fs);
424 				VM_WAITPFAULT;
425 				goto RetryFault;
426 			} else if (fs.m->valid == VM_PAGE_BITS_ALL)
427 				break;
428 		}
429 
430 readrest:
431 		/*
432 		 * We have found a valid page or we have allocated a new page.
433 		 * The page thus may not be valid or may not be entirely
434 		 * valid.
435 		 *
436 		 * Attempt to fault-in the page if there is a chance that the
437 		 * pager has it, and potentially fault in additional pages
438 		 * at the same time.
439 		 */
440 		if (TRYPAGER) {
441 			int rv;
442 			int reqpage = 0;
443 			u_char behavior = vm_map_entry_behavior(fs.entry);
444 
445 			if (behavior == MAP_ENTRY_BEHAV_RANDOM ||
446 			    P_KILLED(curproc)) {
447 				ahead = 0;
448 				behind = 0;
449 			} else {
450 				behind = (vaddr - fs.entry->start) >> PAGE_SHIFT;
451 				if (behind > VM_FAULT_READ_BEHIND)
452 					behind = VM_FAULT_READ_BEHIND;
453 
454 				ahead = ((fs.entry->end - vaddr) >> PAGE_SHIFT) - 1;
455 				if (ahead > VM_FAULT_READ_AHEAD)
456 					ahead = VM_FAULT_READ_AHEAD;
457 			}
458 			is_first_object_locked = FALSE;
459 			if ((behavior == MAP_ENTRY_BEHAV_SEQUENTIAL ||
460 			     (behavior != MAP_ENTRY_BEHAV_RANDOM &&
461 			      fs.pindex >= fs.entry->lastr &&
462 			      fs.pindex < fs.entry->lastr + VM_FAULT_READ)) &&
463 			    (fs.first_object == fs.object ||
464 			     (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object))) &&
465 			    fs.first_object->type != OBJT_DEVICE &&
466 			    fs.first_object->type != OBJT_PHYS &&
467 			    fs.first_object->type != OBJT_SG) {
468 				vm_pindex_t firstpindex;
469 
470 				if (fs.first_pindex < 2 * VM_FAULT_READ)
471 					firstpindex = 0;
472 				else
473 					firstpindex = fs.first_pindex - 2 * VM_FAULT_READ;
474 				mt = fs.first_object != fs.object ?
475 				    fs.first_m : fs.m;
476 				KASSERT(mt != NULL, ("vm_fault: missing mt"));
477 				KASSERT((mt->oflags & VPO_BUSY) != 0,
478 				    ("vm_fault: mt %p not busy", mt));
479 				mt_prev = vm_page_prev(mt);
480 
481 				/*
482 				 * note: partially valid pages cannot be
483 				 * included in the lookahead - NFS piecemeal
484 				 * writes will barf on it badly.
485 				 */
486 				while ((mt = mt_prev) != NULL &&
487 				    mt->pindex >= firstpindex &&
488 				    mt->valid == VM_PAGE_BITS_ALL) {
489 					mt_prev = vm_page_prev(mt);
490 					if (mt->busy ||
491 					    (mt->oflags & VPO_BUSY))
492 						continue;
493 					vm_page_lock(mt);
494 					if (mt->hold_count ||
495 					    mt->wire_count) {
496 						vm_page_unlock(mt);
497 						continue;
498 					}
499 					pmap_remove_all(mt);
500 					if (mt->dirty != 0)
501 						vm_page_deactivate(mt);
502 					else
503 						vm_page_cache(mt);
504 					vm_page_unlock(mt);
505 				}
506 				ahead += behind;
507 				behind = 0;
508 			}
509 			if (is_first_object_locked)
510 				VM_OBJECT_UNLOCK(fs.first_object);
511 
512 			/*
513 			 * Call the pager to retrieve the data, if any, after
514 			 * releasing the lock on the map.  We hold a ref on
515 			 * fs.object and the pages are VPO_BUSY'd.
516 			 */
517 			unlock_map(&fs);
518 
519 vnode_lock:
520 			if (fs.object->type == OBJT_VNODE) {
521 				vp = fs.object->handle;
522 				if (vp == fs.vp)
523 					goto vnode_locked;
524 				else if (fs.vp != NULL) {
525 					vput(fs.vp);
526 					fs.vp = NULL;
527 				}
528 				locked = VOP_ISLOCKED(vp);
529 
530 				if (VFS_NEEDSGIANT(vp->v_mount) && !fs.vfslocked) {
531 					fs.vfslocked = 1;
532 					if (!mtx_trylock(&Giant)) {
533 						VM_OBJECT_UNLOCK(fs.object);
534 						mtx_lock(&Giant);
535 						VM_OBJECT_LOCK(fs.object);
536 						goto vnode_lock;
537 					}
538 				}
539 				if (locked != LK_EXCLUSIVE)
540 					locked = LK_SHARED;
541 				/* Do not sleep for vnode lock while fs.m is busy */
542 				error = vget(vp, locked | LK_CANRECURSE |
543 				    LK_NOWAIT, curthread);
544 				if (error != 0) {
545 					int vfslocked;
546 
547 					vfslocked = fs.vfslocked;
548 					fs.vfslocked = 0; /* Keep Giant */
549 					vhold(vp);
550 					release_page(&fs);
551 					unlock_and_deallocate(&fs);
552 					error = vget(vp, locked | LK_RETRY |
553 					    LK_CANRECURSE, curthread);
554 					vdrop(vp);
555 					fs.vp = vp;
556 					fs.vfslocked = vfslocked;
557 					KASSERT(error == 0,
558 					    ("vm_fault: vget failed"));
559 					goto RetryFault;
560 				}
561 				fs.vp = vp;
562 			}
563 vnode_locked:
564 			KASSERT(fs.vp == NULL || !fs.map->system_map,
565 			    ("vm_fault: vnode-backed object mapped by system map"));
566 
567 			/*
568 			 * now we find out if any other pages should be paged
569 			 * in at this time this routine checks to see if the
570 			 * pages surrounding this fault reside in the same
571 			 * object as the page for this fault.  If they do,
572 			 * then they are faulted in also into the object.  The
573 			 * array "marray" returned contains an array of
574 			 * vm_page_t structs where one of them is the
575 			 * vm_page_t passed to the routine.  The reqpage
576 			 * return value is the index into the marray for the
577 			 * vm_page_t passed to the routine.
578 			 *
579 			 * fs.m plus the additional pages are VPO_BUSY'd.
580 			 */
581 			faultcount = vm_fault_additional_pages(
582 			    fs.m, behind, ahead, marray, &reqpage);
583 
584 			rv = faultcount ?
585 			    vm_pager_get_pages(fs.object, marray, faultcount,
586 				reqpage) : VM_PAGER_FAIL;
587 
588 			if (rv == VM_PAGER_OK) {
589 				/*
590 				 * Found the page. Leave it busy while we play
591 				 * with it.
592 				 */
593 
594 				/*
595 				 * Relookup in case pager changed page. Pager
596 				 * is responsible for disposition of old page
597 				 * if moved.
598 				 */
599 				fs.m = vm_page_lookup(fs.object, fs.pindex);
600 				if (!fs.m) {
601 					unlock_and_deallocate(&fs);
602 					goto RetryFault;
603 				}
604 
605 				hardfault++;
606 				break; /* break to PAGE HAS BEEN FOUND */
607 			}
608 			/*
609 			 * Remove the bogus page (which does not exist at this
610 			 * object/offset); before doing so, we must get back
611 			 * our object lock to preserve our invariant.
612 			 *
613 			 * Also wake up any other process that may want to bring
614 			 * in this page.
615 			 *
616 			 * If this is the top-level object, we must leave the
617 			 * busy page to prevent another process from rushing
618 			 * past us, and inserting the page in that object at
619 			 * the same time that we are.
620 			 */
621 			if (rv == VM_PAGER_ERROR)
622 				printf("vm_fault: pager read error, pid %d (%s)\n",
623 				    curproc->p_pid, curproc->p_comm);
624 			/*
625 			 * Data outside the range of the pager or an I/O error
626 			 */
627 			/*
628 			 * XXX - the check for kernel_map is a kludge to work
629 			 * around having the machine panic on a kernel space
630 			 * fault w/ I/O error.
631 			 */
632 			if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) ||
633 				(rv == VM_PAGER_BAD)) {
634 				vm_page_lock(fs.m);
635 				vm_page_free(fs.m);
636 				vm_page_unlock(fs.m);
637 				fs.m = NULL;
638 				unlock_and_deallocate(&fs);
639 				return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
640 			}
641 			if (fs.object != fs.first_object) {
642 				vm_page_lock(fs.m);
643 				vm_page_free(fs.m);
644 				vm_page_unlock(fs.m);
645 				fs.m = NULL;
646 				/*
647 				 * XXX - we cannot just fall out at this
648 				 * point, m has been freed and is invalid!
649 				 */
650 			}
651 		}
652 
653 		/*
654 		 * We get here if the object has default pager (or unwiring)
655 		 * or the pager doesn't have the page.
656 		 */
657 		if (fs.object == fs.first_object)
658 			fs.first_m = fs.m;
659 
660 		/*
661 		 * Move on to the next object.  Lock the next object before
662 		 * unlocking the current one.
663 		 */
664 		fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset);
665 		next_object = fs.object->backing_object;
666 		if (next_object == NULL) {
667 			/*
668 			 * If there's no object left, fill the page in the top
669 			 * object with zeros.
670 			 */
671 			if (fs.object != fs.first_object) {
672 				vm_object_pip_wakeup(fs.object);
673 				VM_OBJECT_UNLOCK(fs.object);
674 
675 				fs.object = fs.first_object;
676 				fs.pindex = fs.first_pindex;
677 				fs.m = fs.first_m;
678 				VM_OBJECT_LOCK(fs.object);
679 			}
680 			fs.first_m = NULL;
681 
682 			/*
683 			 * Zero the page if necessary and mark it valid.
684 			 */
685 			if ((fs.m->flags & PG_ZERO) == 0) {
686 				pmap_zero_page(fs.m);
687 			} else {
688 				PCPU_INC(cnt.v_ozfod);
689 			}
690 			PCPU_INC(cnt.v_zfod);
691 			fs.m->valid = VM_PAGE_BITS_ALL;
692 			break;	/* break to PAGE HAS BEEN FOUND */
693 		} else {
694 			KASSERT(fs.object != next_object,
695 			    ("object loop %p", next_object));
696 			VM_OBJECT_LOCK(next_object);
697 			vm_object_pip_add(next_object, 1);
698 			if (fs.object != fs.first_object)
699 				vm_object_pip_wakeup(fs.object);
700 			VM_OBJECT_UNLOCK(fs.object);
701 			fs.object = next_object;
702 		}
703 	}
704 
705 	KASSERT((fs.m->oflags & VPO_BUSY) != 0,
706 	    ("vm_fault: not busy after main loop"));
707 
708 	/*
709 	 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
710 	 * is held.]
711 	 */
712 
713 	/*
714 	 * If the page is being written, but isn't already owned by the
715 	 * top-level object, we have to copy it into a new page owned by the
716 	 * top-level object.
717 	 */
718 	if (fs.object != fs.first_object) {
719 		/*
720 		 * We only really need to copy if we want to write it.
721 		 */
722 		if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
723 			/*
724 			 * This allows pages to be virtually copied from a
725 			 * backing_object into the first_object, where the
726 			 * backing object has no other refs to it, and cannot
727 			 * gain any more refs.  Instead of a bcopy, we just
728 			 * move the page from the backing object to the
729 			 * first object.  Note that we must mark the page
730 			 * dirty in the first object so that it will go out
731 			 * to swap when needed.
732 			 */
733 			is_first_object_locked = FALSE;
734 			if (
735 				/*
736 				 * Only one shadow object
737 				 */
738 				(fs.object->shadow_count == 1) &&
739 				/*
740 				 * No COW refs, except us
741 				 */
742 				(fs.object->ref_count == 1) &&
743 				/*
744 				 * No one else can look this object up
745 				 */
746 				(fs.object->handle == NULL) &&
747 				/*
748 				 * No other ways to look the object up
749 				 */
750 				((fs.object->type == OBJT_DEFAULT) ||
751 				 (fs.object->type == OBJT_SWAP)) &&
752 			    (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object)) &&
753 				/*
754 				 * We don't chase down the shadow chain
755 				 */
756 			    fs.object == fs.first_object->backing_object) {
757 				/*
758 				 * get rid of the unnecessary page
759 				 */
760 				vm_page_lock(fs.first_m);
761 				vm_page_free(fs.first_m);
762 				vm_page_unlock(fs.first_m);
763 				/*
764 				 * grab the page and put it into the
765 				 * process'es object.  The page is
766 				 * automatically made dirty.
767 				 */
768 				vm_page_lock(fs.m);
769 				vm_page_rename(fs.m, fs.first_object, fs.first_pindex);
770 				vm_page_unlock(fs.m);
771 				vm_page_busy(fs.m);
772 				fs.first_m = fs.m;
773 				fs.m = NULL;
774 				PCPU_INC(cnt.v_cow_optim);
775 			} else {
776 				/*
777 				 * Oh, well, lets copy it.
778 				 */
779 				pmap_copy_page(fs.m, fs.first_m);
780 				fs.first_m->valid = VM_PAGE_BITS_ALL;
781 				if (wired && (fault_flags &
782 				    VM_FAULT_CHANGE_WIRING) == 0) {
783 					vm_page_lock(fs.first_m);
784 					vm_page_wire(fs.first_m);
785 					vm_page_unlock(fs.first_m);
786 
787 					vm_page_lock(fs.m);
788 					vm_page_unwire(fs.m, FALSE);
789 					vm_page_unlock(fs.m);
790 				}
791 				/*
792 				 * We no longer need the old page or object.
793 				 */
794 				release_page(&fs);
795 			}
796 			/*
797 			 * fs.object != fs.first_object due to above
798 			 * conditional
799 			 */
800 			vm_object_pip_wakeup(fs.object);
801 			VM_OBJECT_UNLOCK(fs.object);
802 			/*
803 			 * Only use the new page below...
804 			 */
805 			fs.object = fs.first_object;
806 			fs.pindex = fs.first_pindex;
807 			fs.m = fs.first_m;
808 			if (!is_first_object_locked)
809 				VM_OBJECT_LOCK(fs.object);
810 			PCPU_INC(cnt.v_cow_faults);
811 		} else {
812 			prot &= ~VM_PROT_WRITE;
813 		}
814 	}
815 
816 	/*
817 	 * We must verify that the maps have not changed since our last
818 	 * lookup.
819 	 */
820 	if (!fs.lookup_still_valid) {
821 		vm_object_t retry_object;
822 		vm_pindex_t retry_pindex;
823 		vm_prot_t retry_prot;
824 
825 		if (!vm_map_trylock_read(fs.map)) {
826 			release_page(&fs);
827 			unlock_and_deallocate(&fs);
828 			goto RetryFault;
829 		}
830 		fs.lookup_still_valid = TRUE;
831 		if (fs.map->timestamp != map_generation) {
832 			result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
833 			    &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
834 
835 			/*
836 			 * If we don't need the page any longer, put it on the inactive
837 			 * list (the easiest thing to do here).  If no one needs it,
838 			 * pageout will grab it eventually.
839 			 */
840 			if (result != KERN_SUCCESS) {
841 				release_page(&fs);
842 				unlock_and_deallocate(&fs);
843 
844 				/*
845 				 * If retry of map lookup would have blocked then
846 				 * retry fault from start.
847 				 */
848 				if (result == KERN_FAILURE)
849 					goto RetryFault;
850 				return (result);
851 			}
852 			if ((retry_object != fs.first_object) ||
853 			    (retry_pindex != fs.first_pindex)) {
854 				release_page(&fs);
855 				unlock_and_deallocate(&fs);
856 				goto RetryFault;
857 			}
858 
859 			/*
860 			 * Check whether the protection has changed or the object has
861 			 * been copied while we left the map unlocked. Changing from
862 			 * read to write permission is OK - we leave the page
863 			 * write-protected, and catch the write fault. Changing from
864 			 * write to read permission means that we can't mark the page
865 			 * write-enabled after all.
866 			 */
867 			prot &= retry_prot;
868 		}
869 	}
870 	/*
871 	 * If the page was filled by a pager, update the map entry's
872 	 * last read offset.  Since the pager does not return the
873 	 * actual set of pages that it read, this update is based on
874 	 * the requested set.  Typically, the requested and actual
875 	 * sets are the same.
876 	 *
877 	 * XXX The following assignment modifies the map
878 	 * without holding a write lock on it.
879 	 */
880 	if (hardfault)
881 		fs.entry->lastr = fs.pindex + faultcount - behind;
882 
883 	if (prot & VM_PROT_WRITE) {
884 		vm_object_set_writeable_dirty(fs.object);
885 
886 		/*
887 		 * If this is a NOSYNC mmap we do not want to set VPO_NOSYNC
888 		 * if the page is already dirty to prevent data written with
889 		 * the expectation of being synced from not being synced.
890 		 * Likewise if this entry does not request NOSYNC then make
891 		 * sure the page isn't marked NOSYNC.  Applications sharing
892 		 * data should use the same flags to avoid ping ponging.
893 		 */
894 		if (fs.entry->eflags & MAP_ENTRY_NOSYNC) {
895 			if (fs.m->dirty == 0)
896 				fs.m->oflags |= VPO_NOSYNC;
897 		} else {
898 			fs.m->oflags &= ~VPO_NOSYNC;
899 		}
900 
901 		/*
902 		 * If the fault is a write, we know that this page is being
903 		 * written NOW so dirty it explicitly to save on
904 		 * pmap_is_modified() calls later.
905 		 *
906 		 * Also tell the backing pager, if any, that it should remove
907 		 * any swap backing since the page is now dirty.
908 		 */
909 		if ((fault_type & VM_PROT_WRITE) != 0 &&
910 		    (fault_flags & VM_FAULT_CHANGE_WIRING) == 0) {
911 			vm_page_dirty(fs.m);
912 			vm_pager_page_unswapped(fs.m);
913 		}
914 	}
915 
916 	/*
917 	 * Page had better still be busy
918 	 */
919 	KASSERT(fs.m->oflags & VPO_BUSY,
920 		("vm_fault: page %p not busy!", fs.m));
921 	/*
922 	 * Page must be completely valid or it is not fit to
923 	 * map into user space.  vm_pager_get_pages() ensures this.
924 	 */
925 	KASSERT(fs.m->valid == VM_PAGE_BITS_ALL,
926 	    ("vm_fault: page %p partially invalid", fs.m));
927 	VM_OBJECT_UNLOCK(fs.object);
928 
929 	/*
930 	 * Put this page into the physical map.  We had to do the unlock above
931 	 * because pmap_enter() may sleep.  We don't put the page
932 	 * back on the active queue until later so that the pageout daemon
933 	 * won't find it (yet).
934 	 */
935 	pmap_enter(fs.map->pmap, vaddr, fault_type, fs.m, prot, wired);
936 	if ((fault_flags & VM_FAULT_CHANGE_WIRING) == 0 && wired == 0)
937 		vm_fault_prefault(fs.map->pmap, vaddr, fs.entry);
938 	VM_OBJECT_LOCK(fs.object);
939 	vm_page_lock(fs.m);
940 
941 	/*
942 	 * If the page is not wired down, then put it where the pageout daemon
943 	 * can find it.
944 	 */
945 	if (fault_flags & VM_FAULT_CHANGE_WIRING) {
946 		if (wired)
947 			vm_page_wire(fs.m);
948 		else
949 			vm_page_unwire(fs.m, 1);
950 	} else
951 		vm_page_activate(fs.m);
952 	vm_page_unlock(fs.m);
953 	vm_page_wakeup(fs.m);
954 
955 	/*
956 	 * Unlock everything, and return
957 	 */
958 	unlock_and_deallocate(&fs);
959 	if (hardfault)
960 		curthread->td_ru.ru_majflt++;
961 	else
962 		curthread->td_ru.ru_minflt++;
963 
964 	return (KERN_SUCCESS);
965 }
966 
967 /*
968  * vm_fault_prefault provides a quick way of clustering
969  * pagefaults into a processes address space.  It is a "cousin"
970  * of vm_map_pmap_enter, except it runs at page fault time instead
971  * of mmap time.
972  */
973 static void
974 vm_fault_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry)
975 {
976 	int i;
977 	vm_offset_t addr, starta;
978 	vm_pindex_t pindex;
979 	vm_page_t m;
980 	vm_object_t object;
981 
982 	if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
983 		return;
984 
985 	object = entry->object.vm_object;
986 
987 	starta = addra - PFBAK * PAGE_SIZE;
988 	if (starta < entry->start) {
989 		starta = entry->start;
990 	} else if (starta > addra) {
991 		starta = 0;
992 	}
993 
994 	for (i = 0; i < PAGEORDER_SIZE; i++) {
995 		vm_object_t backing_object, lobject;
996 
997 		addr = addra + prefault_pageorder[i];
998 		if (addr > addra + (PFFOR * PAGE_SIZE))
999 			addr = 0;
1000 
1001 		if (addr < starta || addr >= entry->end)
1002 			continue;
1003 
1004 		if (!pmap_is_prefaultable(pmap, addr))
1005 			continue;
1006 
1007 		pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1008 		lobject = object;
1009 		VM_OBJECT_LOCK(lobject);
1010 		while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1011 		    lobject->type == OBJT_DEFAULT &&
1012 		    (backing_object = lobject->backing_object) != NULL) {
1013 			KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1014 			    0, ("vm_fault_prefault: unaligned object offset"));
1015 			pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1016 			VM_OBJECT_LOCK(backing_object);
1017 			VM_OBJECT_UNLOCK(lobject);
1018 			lobject = backing_object;
1019 		}
1020 		/*
1021 		 * give-up when a page is not in memory
1022 		 */
1023 		if (m == NULL) {
1024 			VM_OBJECT_UNLOCK(lobject);
1025 			break;
1026 		}
1027 		if (m->valid == VM_PAGE_BITS_ALL &&
1028 		    (m->flags & PG_FICTITIOUS) == 0)
1029 			pmap_enter_quick(pmap, addr, m, entry->protection);
1030 		VM_OBJECT_UNLOCK(lobject);
1031 	}
1032 }
1033 
1034 /*
1035  *	vm_fault_quick:
1036  *
1037  *	Ensure that the requested virtual address, which may be in userland,
1038  *	is valid.  Fault-in the page if necessary.  Return -1 on failure.
1039  */
1040 int
1041 vm_fault_quick(caddr_t v, int prot)
1042 {
1043 	int r;
1044 
1045 	if (prot & VM_PROT_WRITE)
1046 		r = subyte(v, fubyte(v));
1047 	else
1048 		r = fubyte(v);
1049 	return(r);
1050 }
1051 
1052 /*
1053  *	vm_fault_wire:
1054  *
1055  *	Wire down a range of virtual addresses in a map.
1056  */
1057 int
1058 vm_fault_wire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1059     boolean_t fictitious)
1060 {
1061 	vm_offset_t va;
1062 	int rv;
1063 
1064 	/*
1065 	 * We simulate a fault to get the page and enter it in the physical
1066 	 * map.  For user wiring, we only ask for read access on currently
1067 	 * read-only sections.
1068 	 */
1069 	for (va = start; va < end; va += PAGE_SIZE) {
1070 		rv = vm_fault(map, va, VM_PROT_NONE, VM_FAULT_CHANGE_WIRING);
1071 		if (rv) {
1072 			if (va != start)
1073 				vm_fault_unwire(map, start, va, fictitious);
1074 			return (rv);
1075 		}
1076 	}
1077 	return (KERN_SUCCESS);
1078 }
1079 
1080 /*
1081  *	vm_fault_unwire:
1082  *
1083  *	Unwire a range of virtual addresses in a map.
1084  */
1085 void
1086 vm_fault_unwire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1087     boolean_t fictitious)
1088 {
1089 	vm_paddr_t pa;
1090 	vm_offset_t va;
1091 	vm_page_t m;
1092 	pmap_t pmap;
1093 
1094 	pmap = vm_map_pmap(map);
1095 
1096 	/*
1097 	 * Since the pages are wired down, we must be able to get their
1098 	 * mappings from the physical map system.
1099 	 */
1100 	for (va = start; va < end; va += PAGE_SIZE) {
1101 		pa = pmap_extract(pmap, va);
1102 		if (pa != 0) {
1103 			pmap_change_wiring(pmap, va, FALSE);
1104 			if (!fictitious) {
1105 				m = PHYS_TO_VM_PAGE(pa);
1106 				vm_page_lock(m);
1107 				vm_page_unwire(m, TRUE);
1108 				vm_page_unlock(m);
1109 			}
1110 		}
1111 	}
1112 }
1113 
1114 /*
1115  *	Routine:
1116  *		vm_fault_copy_entry
1117  *	Function:
1118  *		Create new shadow object backing dst_entry with private copy of
1119  *		all underlying pages. When src_entry is equal to dst_entry,
1120  *		function implements COW for wired-down map entry. Otherwise,
1121  *		it forks wired entry into dst_map.
1122  *
1123  *	In/out conditions:
1124  *		The source and destination maps must be locked for write.
1125  *		The source map entry must be wired down (or be a sharing map
1126  *		entry corresponding to a main map entry that is wired down).
1127  */
1128 void
1129 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1130     vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1131     vm_ooffset_t *fork_charge)
1132 {
1133 	vm_object_t backing_object, dst_object, object, src_object;
1134 	vm_pindex_t dst_pindex, pindex, src_pindex;
1135 	vm_prot_t access, prot;
1136 	vm_offset_t vaddr;
1137 	vm_page_t dst_m;
1138 	vm_page_t src_m;
1139 	boolean_t src_readonly, upgrade;
1140 
1141 #ifdef	lint
1142 	src_map++;
1143 #endif	/* lint */
1144 
1145 	upgrade = src_entry == dst_entry;
1146 
1147 	src_object = src_entry->object.vm_object;
1148 	src_pindex = OFF_TO_IDX(src_entry->offset);
1149 	src_readonly = (src_entry->protection & VM_PROT_WRITE) == 0;
1150 
1151 	/*
1152 	 * Create the top-level object for the destination entry. (Doesn't
1153 	 * actually shadow anything - we copy the pages directly.)
1154 	 */
1155 	dst_object = vm_object_allocate(OBJT_DEFAULT,
1156 	    OFF_TO_IDX(dst_entry->end - dst_entry->start));
1157 #if VM_NRESERVLEVEL > 0
1158 	dst_object->flags |= OBJ_COLORED;
1159 	dst_object->pg_color = atop(dst_entry->start);
1160 #endif
1161 
1162 	VM_OBJECT_LOCK(dst_object);
1163 	KASSERT(upgrade || dst_entry->object.vm_object == NULL,
1164 	    ("vm_fault_copy_entry: vm_object not NULL"));
1165 	dst_entry->object.vm_object = dst_object;
1166 	dst_entry->offset = 0;
1167 	dst_object->charge = dst_entry->end - dst_entry->start;
1168 	if (fork_charge != NULL) {
1169 		KASSERT(dst_entry->uip == NULL,
1170 		    ("vm_fault_copy_entry: leaked swp charge"));
1171 		dst_object->uip = curthread->td_ucred->cr_ruidinfo;
1172 		uihold(dst_object->uip);
1173 		*fork_charge += dst_object->charge;
1174 	} else {
1175 		dst_object->uip = dst_entry->uip;
1176 		dst_entry->uip = NULL;
1177 	}
1178 	access = prot = dst_entry->protection;
1179 	/*
1180 	 * If not an upgrade, then enter the mappings in the pmap as
1181 	 * read and/or execute accesses.  Otherwise, enter them as
1182 	 * write accesses.
1183 	 *
1184 	 * A writeable large page mapping is only created if all of
1185 	 * the constituent small page mappings are modified. Marking
1186 	 * PTEs as modified on inception allows promotion to happen
1187 	 * without taking potentially large number of soft faults.
1188 	 */
1189 	if (!upgrade)
1190 		access &= ~VM_PROT_WRITE;
1191 
1192 	/*
1193 	 * Loop through all of the pages in the entry's range, copying each
1194 	 * one from the source object (it should be there) to the destination
1195 	 * object.
1196 	 */
1197 	for (vaddr = dst_entry->start, dst_pindex = 0;
1198 	    vaddr < dst_entry->end;
1199 	    vaddr += PAGE_SIZE, dst_pindex++) {
1200 
1201 		/*
1202 		 * Allocate a page in the destination object.
1203 		 */
1204 		do {
1205 			dst_m = vm_page_alloc(dst_object, dst_pindex,
1206 			    VM_ALLOC_NORMAL);
1207 			if (dst_m == NULL) {
1208 				VM_OBJECT_UNLOCK(dst_object);
1209 				VM_WAIT;
1210 				VM_OBJECT_LOCK(dst_object);
1211 			}
1212 		} while (dst_m == NULL);
1213 
1214 		/*
1215 		 * Find the page in the source object, and copy it in.
1216 		 * (Because the source is wired down, the page will be in
1217 		 * memory.)
1218 		 */
1219 		VM_OBJECT_LOCK(src_object);
1220 		object = src_object;
1221 		pindex = src_pindex + dst_pindex;
1222 		while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
1223 		    src_readonly &&
1224 		    (backing_object = object->backing_object) != NULL) {
1225 			/*
1226 			 * Allow fallback to backing objects if we are reading.
1227 			 */
1228 			VM_OBJECT_LOCK(backing_object);
1229 			pindex += OFF_TO_IDX(object->backing_object_offset);
1230 			VM_OBJECT_UNLOCK(object);
1231 			object = backing_object;
1232 		}
1233 		if (src_m == NULL)
1234 			panic("vm_fault_copy_wired: page missing");
1235 		pmap_copy_page(src_m, dst_m);
1236 		VM_OBJECT_UNLOCK(object);
1237 		dst_m->valid = VM_PAGE_BITS_ALL;
1238 		VM_OBJECT_UNLOCK(dst_object);
1239 
1240 		/*
1241 		 * Enter it in the pmap. If a wired, copy-on-write
1242 		 * mapping is being replaced by a write-enabled
1243 		 * mapping, then wire that new mapping.
1244 		 */
1245 		pmap_enter(dst_map->pmap, vaddr, access, dst_m, prot, upgrade);
1246 
1247 		/*
1248 		 * Mark it no longer busy, and put it on the active list.
1249 		 */
1250 		VM_OBJECT_LOCK(dst_object);
1251 
1252 		if (upgrade) {
1253 			vm_page_lock(src_m);
1254 			vm_page_unwire(src_m, 0);
1255 			vm_page_unlock(src_m);
1256 
1257 			vm_page_lock(dst_m);
1258 			vm_page_wire(dst_m);
1259 			vm_page_unlock(dst_m);
1260 		} else {
1261 			vm_page_lock(dst_m);
1262 			vm_page_activate(dst_m);
1263 			vm_page_unlock(dst_m);
1264 		}
1265 		vm_page_wakeup(dst_m);
1266 	}
1267 	VM_OBJECT_UNLOCK(dst_object);
1268 	if (upgrade) {
1269 		dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
1270 		vm_object_deallocate(src_object);
1271 	}
1272 }
1273 
1274 
1275 /*
1276  * This routine checks around the requested page for other pages that
1277  * might be able to be faulted in.  This routine brackets the viable
1278  * pages for the pages to be paged in.
1279  *
1280  * Inputs:
1281  *	m, rbehind, rahead
1282  *
1283  * Outputs:
1284  *  marray (array of vm_page_t), reqpage (index of requested page)
1285  *
1286  * Return value:
1287  *  number of pages in marray
1288  */
1289 static int
1290 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage)
1291 	vm_page_t m;
1292 	int rbehind;
1293 	int rahead;
1294 	vm_page_t *marray;
1295 	int *reqpage;
1296 {
1297 	int i,j;
1298 	vm_object_t object;
1299 	vm_pindex_t pindex, startpindex, endpindex, tpindex;
1300 	vm_page_t rtm;
1301 	int cbehind, cahead;
1302 
1303 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1304 
1305 	object = m->object;
1306 	pindex = m->pindex;
1307 	cbehind = cahead = 0;
1308 
1309 	/*
1310 	 * if the requested page is not available, then give up now
1311 	 */
1312 	if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1313 		return 0;
1314 	}
1315 
1316 	if ((cbehind == 0) && (cahead == 0)) {
1317 		*reqpage = 0;
1318 		marray[0] = m;
1319 		return 1;
1320 	}
1321 
1322 	if (rahead > cahead) {
1323 		rahead = cahead;
1324 	}
1325 
1326 	if (rbehind > cbehind) {
1327 		rbehind = cbehind;
1328 	}
1329 
1330 	/*
1331 	 * scan backward for the read behind pages -- in memory
1332 	 */
1333 	if (pindex > 0) {
1334 		if (rbehind > pindex) {
1335 			rbehind = pindex;
1336 			startpindex = 0;
1337 		} else {
1338 			startpindex = pindex - rbehind;
1339 		}
1340 
1341 		if ((rtm = TAILQ_PREV(m, pglist, listq)) != NULL &&
1342 		    rtm->pindex >= startpindex)
1343 			startpindex = rtm->pindex + 1;
1344 
1345 		/* tpindex is unsigned; beware of numeric underflow. */
1346 		for (i = 0, tpindex = pindex - 1; tpindex >= startpindex &&
1347 		    tpindex < pindex; i++, tpindex--) {
1348 
1349 			rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1350 			    VM_ALLOC_IFNOTCACHED);
1351 			if (rtm == NULL) {
1352 				/*
1353 				 * Shift the allocated pages to the
1354 				 * beginning of the array.
1355 				 */
1356 				for (j = 0; j < i; j++) {
1357 					marray[j] = marray[j + tpindex + 1 -
1358 					    startpindex];
1359 				}
1360 				break;
1361 			}
1362 
1363 			marray[tpindex - startpindex] = rtm;
1364 		}
1365 	} else {
1366 		startpindex = 0;
1367 		i = 0;
1368 	}
1369 
1370 	marray[i] = m;
1371 	/* page offset of the required page */
1372 	*reqpage = i;
1373 
1374 	tpindex = pindex + 1;
1375 	i++;
1376 
1377 	/*
1378 	 * scan forward for the read ahead pages
1379 	 */
1380 	endpindex = tpindex + rahead;
1381 	if ((rtm = TAILQ_NEXT(m, listq)) != NULL && rtm->pindex < endpindex)
1382 		endpindex = rtm->pindex;
1383 	if (endpindex > object->size)
1384 		endpindex = object->size;
1385 
1386 	for (; tpindex < endpindex; i++, tpindex++) {
1387 
1388 		rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1389 		    VM_ALLOC_IFNOTCACHED);
1390 		if (rtm == NULL) {
1391 			break;
1392 		}
1393 
1394 		marray[i] = rtm;
1395 	}
1396 
1397 	/* return number of pages */
1398 	return i;
1399 }
1400