xref: /freebsd/sys/vm/vm_fault.c (revision 84ee9401a3fc8d3c22424266f421a928989cd692)
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 <sys/param.h>
78 #include <sys/systm.h>
79 #include <sys/kernel.h>
80 #include <sys/lock.h>
81 #include <sys/mutex.h>
82 #include <sys/proc.h>
83 #include <sys/resourcevar.h>
84 #include <sys/sysctl.h>
85 #include <sys/vmmeter.h>
86 #include <sys/vnode.h>
87 
88 #include <vm/vm.h>
89 #include <vm/vm_param.h>
90 #include <vm/pmap.h>
91 #include <vm/vm_map.h>
92 #include <vm/vm_object.h>
93 #include <vm/vm_page.h>
94 #include <vm/vm_pageout.h>
95 #include <vm/vm_kern.h>
96 #include <vm/vm_pager.h>
97 #include <vm/vnode_pager.h>
98 #include <vm/vm_extern.h>
99 
100 #include <sys/mount.h>	/* XXX Temporary for VFS_LOCK_GIANT() */
101 
102 #define PFBAK 4
103 #define PFFOR 4
104 #define PAGEORDER_SIZE (PFBAK+PFFOR)
105 
106 static int prefault_pageorder[] = {
107 	-1 * PAGE_SIZE, 1 * PAGE_SIZE,
108 	-2 * PAGE_SIZE, 2 * PAGE_SIZE,
109 	-3 * PAGE_SIZE, 3 * PAGE_SIZE,
110 	-4 * PAGE_SIZE, 4 * PAGE_SIZE
111 };
112 
113 static int vm_fault_additional_pages(vm_page_t, int, int, vm_page_t *, int *);
114 static void vm_fault_prefault(pmap_t, vm_offset_t, vm_map_entry_t);
115 
116 #define VM_FAULT_READ_AHEAD 8
117 #define VM_FAULT_READ_BEHIND 7
118 #define VM_FAULT_READ (VM_FAULT_READ_AHEAD+VM_FAULT_READ_BEHIND+1)
119 
120 struct faultstate {
121 	vm_page_t m;
122 	vm_object_t object;
123 	vm_pindex_t pindex;
124 	vm_page_t first_m;
125 	vm_object_t	first_object;
126 	vm_pindex_t first_pindex;
127 	vm_map_t map;
128 	vm_map_entry_t entry;
129 	int lookup_still_valid;
130 	struct vnode *vp;
131 };
132 
133 static inline void
134 release_page(struct faultstate *fs)
135 {
136 	vm_page_lock_queues();
137 	vm_page_wakeup(fs->m);
138 	vm_page_deactivate(fs->m);
139 	vm_page_unlock_queues();
140 	fs->m = NULL;
141 }
142 
143 static inline void
144 unlock_map(struct faultstate *fs)
145 {
146 	if (fs->lookup_still_valid) {
147 		vm_map_lookup_done(fs->map, fs->entry);
148 		fs->lookup_still_valid = FALSE;
149 	}
150 }
151 
152 static void
153 unlock_and_deallocate(struct faultstate *fs)
154 {
155 
156 	vm_object_pip_wakeup(fs->object);
157 	VM_OBJECT_UNLOCK(fs->object);
158 	if (fs->object != fs->first_object) {
159 		VM_OBJECT_LOCK(fs->first_object);
160 		vm_page_lock_queues();
161 		vm_page_free(fs->first_m);
162 		vm_page_unlock_queues();
163 		vm_object_pip_wakeup(fs->first_object);
164 		VM_OBJECT_UNLOCK(fs->first_object);
165 		fs->first_m = NULL;
166 	}
167 	vm_object_deallocate(fs->first_object);
168 	unlock_map(fs);
169 	if (fs->vp != NULL) {
170 		int vfslocked;
171 
172 		vfslocked = VFS_LOCK_GIANT(fs->vp->v_mount);
173 		vput(fs->vp);
174 		fs->vp = NULL;
175 		VFS_UNLOCK_GIANT(vfslocked);
176 	}
177 }
178 
179 /*
180  * TRYPAGER - used by vm_fault to calculate whether the pager for the
181  *	      current object *might* contain the page.
182  *
183  *	      default objects are zero-fill, there is no real pager.
184  */
185 #define TRYPAGER	(fs.object->type != OBJT_DEFAULT && \
186 			(((fault_flags & VM_FAULT_WIRE_MASK) == 0) || wired))
187 
188 /*
189  *	vm_fault:
190  *
191  *	Handle a page fault occurring at the given address,
192  *	requiring the given permissions, in the map specified.
193  *	If successful, the page is inserted into the
194  *	associated physical map.
195  *
196  *	NOTE: the given address should be truncated to the
197  *	proper page address.
198  *
199  *	KERN_SUCCESS is returned if the page fault is handled; otherwise,
200  *	a standard error specifying why the fault is fatal is returned.
201  *
202  *
203  *	The map in question must be referenced, and remains so.
204  *	Caller may hold no locks.
205  */
206 int
207 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
208 	 int fault_flags)
209 {
210 	vm_prot_t prot;
211 	int is_first_object_locked, result;
212 	boolean_t growstack, wired;
213 	int map_generation;
214 	vm_object_t next_object;
215 	vm_page_t marray[VM_FAULT_READ];
216 	int hardfault;
217 	int faultcount;
218 	struct faultstate fs;
219 
220 	hardfault = 0;
221 	growstack = TRUE;
222 	atomic_add_int(&cnt.v_vm_faults, 1);
223 
224 RetryFault:;
225 
226 	/*
227 	 * Find the backing store object and offset into it to begin the
228 	 * search.
229 	 */
230 	fs.map = map;
231 	result = vm_map_lookup(&fs.map, vaddr, fault_type, &fs.entry,
232 	    &fs.first_object, &fs.first_pindex, &prot, &wired);
233 	if (result != KERN_SUCCESS) {
234 		if (result != KERN_PROTECTION_FAILURE ||
235 		    (fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE) {
236 			if (growstack && result == KERN_INVALID_ADDRESS &&
237 			    map != kernel_map && curproc != NULL) {
238 				result = vm_map_growstack(curproc, vaddr);
239 				if (result != KERN_SUCCESS)
240 					return (KERN_FAILURE);
241 				growstack = FALSE;
242 				goto RetryFault;
243 			}
244 			return (result);
245 		}
246 
247 		/*
248    		 * If we are user-wiring a r/w segment, and it is COW, then
249    		 * we need to do the COW operation.  Note that we don't COW
250    		 * currently RO sections now, because it is NOT desirable
251    		 * to COW .text.  We simply keep .text from ever being COW'ed
252    		 * and take the heat that one cannot debug wired .text sections.
253    		 */
254 		result = vm_map_lookup(&fs.map, vaddr,
255 			VM_PROT_READ|VM_PROT_WRITE|VM_PROT_OVERRIDE_WRITE,
256 			&fs.entry, &fs.first_object, &fs.first_pindex, &prot, &wired);
257 		if (result != KERN_SUCCESS)
258 			return (result);
259 
260 		/*
261 		 * If we don't COW now, on a user wire, the user will never
262 		 * be able to write to the mapping.  If we don't make this
263 		 * restriction, the bookkeeping would be nearly impossible.
264 		 *
265 		 * XXX The following assignment modifies the map without
266 		 * holding a write lock on it.
267 		 */
268 		if ((fs.entry->protection & VM_PROT_WRITE) == 0)
269 			fs.entry->max_protection &= ~VM_PROT_WRITE;
270 	}
271 
272 	map_generation = fs.map->timestamp;
273 
274 	if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
275 		panic("vm_fault: fault on nofault entry, addr: %lx",
276 		    (u_long)vaddr);
277 	}
278 
279 	/*
280 	 * Make a reference to this object to prevent its disposal while we
281 	 * are messing with it.  Once we have the reference, the map is free
282 	 * to be diddled.  Since objects reference their shadows (and copies),
283 	 * they will stay around as well.
284 	 *
285 	 * Bump the paging-in-progress count to prevent size changes (e.g.
286 	 * truncation operations) during I/O.  This must be done after
287 	 * obtaining the vnode lock in order to avoid possible deadlocks.
288 	 *
289 	 * XXX vnode_pager_lock() can block without releasing the map lock.
290 	 */
291 	if (fs.first_object->flags & OBJ_NEEDGIANT)
292 		mtx_lock(&Giant);
293 	VM_OBJECT_LOCK(fs.first_object);
294 	vm_object_reference_locked(fs.first_object);
295 	fs.vp = vnode_pager_lock(fs.first_object);
296 	KASSERT(fs.vp == NULL || !fs.map->system_map,
297 	    ("vm_fault: vnode-backed object mapped by system map"));
298 	KASSERT((fs.first_object->flags & OBJ_NEEDGIANT) == 0 ||
299 	    !fs.map->system_map,
300 	    ("vm_fault: Object requiring giant mapped by system map"));
301 	if (fs.first_object->flags & OBJ_NEEDGIANT)
302 		mtx_unlock(&Giant);
303 	vm_object_pip_add(fs.first_object, 1);
304 
305 	fs.lookup_still_valid = TRUE;
306 
307 	if (wired)
308 		fault_type = prot;
309 
310 	fs.first_m = NULL;
311 
312 	/*
313 	 * Search for the page at object/offset.
314 	 */
315 	fs.object = fs.first_object;
316 	fs.pindex = fs.first_pindex;
317 	while (TRUE) {
318 		/*
319 		 * If the object is dead, we stop here
320 		 */
321 		if (fs.object->flags & OBJ_DEAD) {
322 			unlock_and_deallocate(&fs);
323 			return (KERN_PROTECTION_FAILURE);
324 		}
325 
326 		/*
327 		 * See if page is resident
328 		 */
329 		fs.m = vm_page_lookup(fs.object, fs.pindex);
330 		if (fs.m != NULL) {
331 			int queue;
332 
333 			/*
334 			 * check for page-based copy on write.
335 			 * We check fs.object == fs.first_object so
336 			 * as to ensure the legacy COW mechanism is
337 			 * used when the page in question is part of
338 			 * a shadow object.  Otherwise, vm_page_cowfault()
339 			 * removes the page from the backing object,
340 			 * which is not what we want.
341 			 */
342 			vm_page_lock_queues();
343 			if ((fs.m->cow) &&
344 			    (fault_type & VM_PROT_WRITE) &&
345 			    (fs.object == fs.first_object)) {
346 				vm_page_cowfault(fs.m);
347 				vm_page_unlock_queues();
348 				unlock_and_deallocate(&fs);
349 				goto RetryFault;
350 			}
351 
352 			/*
353 			 * Wait/Retry if the page is busy.  We have to do this
354 			 * if the page is busy via either PG_BUSY or
355 			 * vm_page_t->busy because the vm_pager may be using
356 			 * vm_page_t->busy for pageouts ( and even pageins if
357 			 * it is the vnode pager ), and we could end up trying
358 			 * to pagein and pageout the same page simultaneously.
359 			 *
360 			 * We can theoretically allow the busy case on a read
361 			 * fault if the page is marked valid, but since such
362 			 * pages are typically already pmap'd, putting that
363 			 * special case in might be more effort then it is
364 			 * worth.  We cannot under any circumstances mess
365 			 * around with a vm_page_t->busy page except, perhaps,
366 			 * to pmap it.
367 			 */
368 			if ((fs.m->flags & PG_BUSY) || fs.m->busy) {
369 				vm_page_unlock_queues();
370 				VM_OBJECT_UNLOCK(fs.object);
371 				if (fs.object != fs.first_object) {
372 					VM_OBJECT_LOCK(fs.first_object);
373 					vm_page_lock_queues();
374 					vm_page_free(fs.first_m);
375 					vm_page_unlock_queues();
376 					vm_object_pip_wakeup(fs.first_object);
377 					VM_OBJECT_UNLOCK(fs.first_object);
378 					fs.first_m = NULL;
379 				}
380 				unlock_map(&fs);
381 				if (fs.vp != NULL) {
382 					int vfslck;
383 
384 					vfslck = VFS_LOCK_GIANT(fs.vp->v_mount);
385 					vput(fs.vp);
386 					fs.vp = NULL;
387 					VFS_UNLOCK_GIANT(vfslck);
388 				}
389 				VM_OBJECT_LOCK(fs.object);
390 				if (fs.m == vm_page_lookup(fs.object,
391 				    fs.pindex)) {
392 					vm_page_sleep_if_busy(fs.m, TRUE,
393 					    "vmpfw");
394 				}
395 				vm_object_pip_wakeup(fs.object);
396 				VM_OBJECT_UNLOCK(fs.object);
397 				atomic_add_int(&cnt.v_intrans, 1);
398 				vm_object_deallocate(fs.first_object);
399 				goto RetryFault;
400 			}
401 			queue = fs.m->queue;
402 
403 			vm_pageq_remove_nowakeup(fs.m);
404 
405 			if (VM_PAGE_RESOLVEQUEUE(fs.m, queue) == PQ_CACHE &&
406 			    vm_page_count_severe()) {
407 				vm_page_activate(fs.m);
408 				vm_page_unlock_queues();
409 				unlock_and_deallocate(&fs);
410 				VM_WAITPFAULT;
411 				goto RetryFault;
412 			}
413 
414 			/*
415 			 * Mark page busy for other processes, and the
416 			 * pagedaemon.  If it still isn't completely valid
417 			 * (readable), jump to readrest, else break-out ( we
418 			 * found the page ).
419 			 */
420 			vm_page_busy(fs.m);
421 			vm_page_unlock_queues();
422 			if (((fs.m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) &&
423 				fs.m->object != kernel_object && fs.m->object != kmem_object) {
424 				goto readrest;
425 			}
426 
427 			break;
428 		}
429 
430 		/*
431 		 * Page is not resident, If this is the search termination
432 		 * or the pager might contain the page, allocate a new page.
433 		 */
434 		if (TRYPAGER || fs.object == fs.first_object) {
435 			if (fs.pindex >= fs.object->size) {
436 				unlock_and_deallocate(&fs);
437 				return (KERN_PROTECTION_FAILURE);
438 			}
439 
440 			/*
441 			 * Allocate a new page for this object/offset pair.
442 			 */
443 			fs.m = NULL;
444 			if (!vm_page_count_severe()) {
445 				fs.m = vm_page_alloc(fs.object, fs.pindex,
446 				    (fs.vp || fs.object->backing_object)? VM_ALLOC_NORMAL: VM_ALLOC_ZERO);
447 			}
448 			if (fs.m == NULL) {
449 				unlock_and_deallocate(&fs);
450 				VM_WAITPFAULT;
451 				goto RetryFault;
452 			}
453 		}
454 
455 readrest:
456 		/*
457 		 * We have found a valid page or we have allocated a new page.
458 		 * The page thus may not be valid or may not be entirely
459 		 * valid.
460 		 *
461 		 * Attempt to fault-in the page if there is a chance that the
462 		 * pager has it, and potentially fault in additional pages
463 		 * at the same time.
464 		 */
465 		if (TRYPAGER) {
466 			int rv;
467 			int reqpage;
468 			int ahead, behind;
469 			u_char behavior = vm_map_entry_behavior(fs.entry);
470 
471 			if (behavior == MAP_ENTRY_BEHAV_RANDOM) {
472 				ahead = 0;
473 				behind = 0;
474 			} else {
475 				behind = (vaddr - fs.entry->start) >> PAGE_SHIFT;
476 				if (behind > VM_FAULT_READ_BEHIND)
477 					behind = VM_FAULT_READ_BEHIND;
478 
479 				ahead = ((fs.entry->end - vaddr) >> PAGE_SHIFT) - 1;
480 				if (ahead > VM_FAULT_READ_AHEAD)
481 					ahead = VM_FAULT_READ_AHEAD;
482 			}
483 			is_first_object_locked = FALSE;
484 			if ((behavior == MAP_ENTRY_BEHAV_SEQUENTIAL ||
485 			     (behavior != MAP_ENTRY_BEHAV_RANDOM &&
486 			      fs.pindex >= fs.entry->lastr &&
487 			      fs.pindex < fs.entry->lastr + VM_FAULT_READ)) &&
488 			    (fs.first_object == fs.object ||
489 			     (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object))) &&
490 			    fs.first_object->type != OBJT_DEVICE) {
491 				vm_pindex_t firstpindex, tmppindex;
492 
493 				if (fs.first_pindex < 2 * VM_FAULT_READ)
494 					firstpindex = 0;
495 				else
496 					firstpindex = fs.first_pindex - 2 * VM_FAULT_READ;
497 
498 				vm_page_lock_queues();
499 				/*
500 				 * note: partially valid pages cannot be
501 				 * included in the lookahead - NFS piecemeal
502 				 * writes will barf on it badly.
503 				 */
504 				for (tmppindex = fs.first_pindex - 1;
505 					tmppindex >= firstpindex;
506 					--tmppindex) {
507 					vm_page_t mt;
508 
509 					mt = vm_page_lookup(fs.first_object, tmppindex);
510 					if (mt == NULL || (mt->valid != VM_PAGE_BITS_ALL))
511 						break;
512 					if (mt->busy ||
513 						(mt->flags & (PG_BUSY | PG_FICTITIOUS | PG_UNMANAGED)) ||
514 						mt->hold_count ||
515 						mt->wire_count)
516 						continue;
517 					pmap_remove_all(mt);
518 					if (mt->dirty) {
519 						vm_page_deactivate(mt);
520 					} else {
521 						vm_page_cache(mt);
522 					}
523 				}
524 				vm_page_unlock_queues();
525 				ahead += behind;
526 				behind = 0;
527 			}
528 			if (is_first_object_locked)
529 				VM_OBJECT_UNLOCK(fs.first_object);
530 			/*
531 			 * now we find out if any other pages should be paged
532 			 * in at this time this routine checks to see if the
533 			 * pages surrounding this fault reside in the same
534 			 * object as the page for this fault.  If they do,
535 			 * then they are faulted in also into the object.  The
536 			 * array "marray" returned contains an array of
537 			 * vm_page_t structs where one of them is the
538 			 * vm_page_t passed to the routine.  The reqpage
539 			 * return value is the index into the marray for the
540 			 * vm_page_t passed to the routine.
541 			 *
542 			 * fs.m plus the additional pages are PG_BUSY'd.
543 			 *
544 			 * XXX vm_fault_additional_pages() can block
545 			 * without releasing the map lock.
546 			 */
547 			faultcount = vm_fault_additional_pages(
548 			    fs.m, behind, ahead, marray, &reqpage);
549 
550 			/*
551 			 * update lastr imperfectly (we do not know how much
552 			 * getpages will actually read), but good enough.
553 			 *
554 			 * XXX The following assignment modifies the map
555 			 * without holding a write lock on it.
556 			 */
557 			fs.entry->lastr = fs.pindex + faultcount - behind;
558 
559 			/*
560 			 * Call the pager to retrieve the data, if any, after
561 			 * releasing the lock on the map.  We hold a ref on
562 			 * fs.object and the pages are PG_BUSY'd.
563 			 */
564 			unlock_map(&fs);
565 
566 			rv = faultcount ?
567 			    vm_pager_get_pages(fs.object, marray, faultcount,
568 				reqpage) : VM_PAGER_FAIL;
569 
570 			if (rv == VM_PAGER_OK) {
571 				/*
572 				 * Found the page. Leave it busy while we play
573 				 * with it.
574 				 */
575 
576 				/*
577 				 * Relookup in case pager changed page. Pager
578 				 * is responsible for disposition of old page
579 				 * if moved.
580 				 */
581 				fs.m = vm_page_lookup(fs.object, fs.pindex);
582 				if (!fs.m) {
583 					unlock_and_deallocate(&fs);
584 					goto RetryFault;
585 				}
586 
587 				hardfault++;
588 				break; /* break to PAGE HAS BEEN FOUND */
589 			}
590 			/*
591 			 * Remove the bogus page (which does not exist at this
592 			 * object/offset); before doing so, we must get back
593 			 * our object lock to preserve our invariant.
594 			 *
595 			 * Also wake up any other process that may want to bring
596 			 * in this page.
597 			 *
598 			 * If this is the top-level object, we must leave the
599 			 * busy page to prevent another process from rushing
600 			 * past us, and inserting the page in that object at
601 			 * the same time that we are.
602 			 */
603 			if (rv == VM_PAGER_ERROR)
604 				printf("vm_fault: pager read error, pid %d (%s)\n",
605 				    curproc->p_pid, curproc->p_comm);
606 			/*
607 			 * Data outside the range of the pager or an I/O error
608 			 */
609 			/*
610 			 * XXX - the check for kernel_map is a kludge to work
611 			 * around having the machine panic on a kernel space
612 			 * fault w/ I/O error.
613 			 */
614 			if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) ||
615 				(rv == VM_PAGER_BAD)) {
616 				vm_page_lock_queues();
617 				vm_page_free(fs.m);
618 				vm_page_unlock_queues();
619 				fs.m = NULL;
620 				unlock_and_deallocate(&fs);
621 				return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
622 			}
623 			if (fs.object != fs.first_object) {
624 				vm_page_lock_queues();
625 				vm_page_free(fs.m);
626 				vm_page_unlock_queues();
627 				fs.m = NULL;
628 				/*
629 				 * XXX - we cannot just fall out at this
630 				 * point, m has been freed and is invalid!
631 				 */
632 			}
633 		}
634 
635 		/*
636 		 * We get here if the object has default pager (or unwiring)
637 		 * or the pager doesn't have the page.
638 		 */
639 		if (fs.object == fs.first_object)
640 			fs.first_m = fs.m;
641 
642 		/*
643 		 * Move on to the next object.  Lock the next object before
644 		 * unlocking the current one.
645 		 */
646 		fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset);
647 		next_object = fs.object->backing_object;
648 		if (next_object == NULL) {
649 			/*
650 			 * If there's no object left, fill the page in the top
651 			 * object with zeros.
652 			 */
653 			if (fs.object != fs.first_object) {
654 				vm_object_pip_wakeup(fs.object);
655 				VM_OBJECT_UNLOCK(fs.object);
656 
657 				fs.object = fs.first_object;
658 				fs.pindex = fs.first_pindex;
659 				fs.m = fs.first_m;
660 				VM_OBJECT_LOCK(fs.object);
661 			}
662 			fs.first_m = NULL;
663 
664 			/*
665 			 * Zero the page if necessary and mark it valid.
666 			 */
667 			if ((fs.m->flags & PG_ZERO) == 0) {
668 				pmap_zero_page(fs.m);
669 			} else {
670 				atomic_add_int(&cnt.v_ozfod, 1);
671 			}
672 			atomic_add_int(&cnt.v_zfod, 1);
673 			fs.m->valid = VM_PAGE_BITS_ALL;
674 			break;	/* break to PAGE HAS BEEN FOUND */
675 		} else {
676 			KASSERT(fs.object != next_object,
677 			    ("object loop %p", next_object));
678 			VM_OBJECT_LOCK(next_object);
679 			vm_object_pip_add(next_object, 1);
680 			if (fs.object != fs.first_object)
681 				vm_object_pip_wakeup(fs.object);
682 			VM_OBJECT_UNLOCK(fs.object);
683 			fs.object = next_object;
684 		}
685 	}
686 
687 	KASSERT((fs.m->flags & PG_BUSY) != 0,
688 	    ("vm_fault: not busy after main loop"));
689 
690 	/*
691 	 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
692 	 * is held.]
693 	 */
694 
695 	/*
696 	 * If the page is being written, but isn't already owned by the
697 	 * top-level object, we have to copy it into a new page owned by the
698 	 * top-level object.
699 	 */
700 	if (fs.object != fs.first_object) {
701 		/*
702 		 * We only really need to copy if we want to write it.
703 		 */
704 		if (fault_type & VM_PROT_WRITE) {
705 			/*
706 			 * This allows pages to be virtually copied from a
707 			 * backing_object into the first_object, where the
708 			 * backing object has no other refs to it, and cannot
709 			 * gain any more refs.  Instead of a bcopy, we just
710 			 * move the page from the backing object to the
711 			 * first object.  Note that we must mark the page
712 			 * dirty in the first object so that it will go out
713 			 * to swap when needed.
714 			 */
715 			is_first_object_locked = FALSE;
716 			if (
717 				/*
718 				 * Only one shadow object
719 				 */
720 				(fs.object->shadow_count == 1) &&
721 				/*
722 				 * No COW refs, except us
723 				 */
724 				(fs.object->ref_count == 1) &&
725 				/*
726 				 * No one else can look this object up
727 				 */
728 				(fs.object->handle == NULL) &&
729 				/*
730 				 * No other ways to look the object up
731 				 */
732 				((fs.object->type == OBJT_DEFAULT) ||
733 				 (fs.object->type == OBJT_SWAP)) &&
734 			    (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object)) &&
735 				/*
736 				 * We don't chase down the shadow chain
737 				 */
738 			    fs.object == fs.first_object->backing_object) {
739 				vm_page_lock_queues();
740 				/*
741 				 * get rid of the unnecessary page
742 				 */
743 				vm_page_free(fs.first_m);
744 				/*
745 				 * grab the page and put it into the
746 				 * process'es object.  The page is
747 				 * automatically made dirty.
748 				 */
749 				vm_page_rename(fs.m, fs.first_object, fs.first_pindex);
750 				vm_page_busy(fs.m);
751 				vm_page_unlock_queues();
752 				fs.first_m = fs.m;
753 				fs.m = NULL;
754 				atomic_add_int(&cnt.v_cow_optim, 1);
755 			} else {
756 				/*
757 				 * Oh, well, lets copy it.
758 				 */
759 				pmap_copy_page(fs.m, fs.first_m);
760 				fs.first_m->valid = VM_PAGE_BITS_ALL;
761 			}
762 			if (fs.m) {
763 				/*
764 				 * We no longer need the old page or object.
765 				 */
766 				release_page(&fs);
767 			}
768 			/*
769 			 * fs.object != fs.first_object due to above
770 			 * conditional
771 			 */
772 			vm_object_pip_wakeup(fs.object);
773 			VM_OBJECT_UNLOCK(fs.object);
774 			/*
775 			 * Only use the new page below...
776 			 */
777 			fs.object = fs.first_object;
778 			fs.pindex = fs.first_pindex;
779 			fs.m = fs.first_m;
780 			if (!is_first_object_locked)
781 				VM_OBJECT_LOCK(fs.object);
782 			atomic_add_int(&cnt.v_cow_faults, 1);
783 		} else {
784 			prot &= ~VM_PROT_WRITE;
785 		}
786 	}
787 
788 	/*
789 	 * We must verify that the maps have not changed since our last
790 	 * lookup.
791 	 */
792 	if (!fs.lookup_still_valid) {
793 		vm_object_t retry_object;
794 		vm_pindex_t retry_pindex;
795 		vm_prot_t retry_prot;
796 
797 		if (!vm_map_trylock_read(fs.map)) {
798 			release_page(&fs);
799 			unlock_and_deallocate(&fs);
800 			goto RetryFault;
801 		}
802 		fs.lookup_still_valid = TRUE;
803 		if (fs.map->timestamp != map_generation) {
804 			result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
805 			    &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
806 
807 			/*
808 			 * If we don't need the page any longer, put it on the inactive
809 			 * list (the easiest thing to do here).  If no one needs it,
810 			 * pageout will grab it eventually.
811 			 */
812 			if (result != KERN_SUCCESS) {
813 				release_page(&fs);
814 				unlock_and_deallocate(&fs);
815 
816 				/*
817 				 * If retry of map lookup would have blocked then
818 				 * retry fault from start.
819 				 */
820 				if (result == KERN_FAILURE)
821 					goto RetryFault;
822 				return (result);
823 			}
824 			if ((retry_object != fs.first_object) ||
825 			    (retry_pindex != fs.first_pindex)) {
826 				release_page(&fs);
827 				unlock_and_deallocate(&fs);
828 				goto RetryFault;
829 			}
830 
831 			/*
832 			 * Check whether the protection has changed or the object has
833 			 * been copied while we left the map unlocked. Changing from
834 			 * read to write permission is OK - we leave the page
835 			 * write-protected, and catch the write fault. Changing from
836 			 * write to read permission means that we can't mark the page
837 			 * write-enabled after all.
838 			 */
839 			prot &= retry_prot;
840 		}
841 	}
842 	if (prot & VM_PROT_WRITE) {
843 		vm_page_lock_queues();
844 		vm_page_flag_set(fs.m, PG_WRITEABLE);
845 		vm_page_unlock_queues();
846 		vm_object_set_writeable_dirty(fs.object);
847 
848 		/*
849 		 * If the fault is a write, we know that this page is being
850 		 * written NOW so dirty it explicitly to save on
851 		 * pmap_is_modified() calls later.
852 		 *
853 		 * If this is a NOSYNC mmap we do not want to set VPO_NOSYNC
854 		 * if the page is already dirty to prevent data written with
855 		 * the expectation of being synced from not being synced.
856 		 * Likewise if this entry does not request NOSYNC then make
857 		 * sure the page isn't marked NOSYNC.  Applications sharing
858 		 * data should use the same flags to avoid ping ponging.
859 		 *
860 		 * Also tell the backing pager, if any, that it should remove
861 		 * any swap backing since the page is now dirty.
862 		 */
863 		if (fs.entry->eflags & MAP_ENTRY_NOSYNC) {
864 			if (fs.m->dirty == 0)
865 				fs.m->oflags |= VPO_NOSYNC;
866 		} else {
867 			fs.m->oflags &= ~VPO_NOSYNC;
868 		}
869 		if (fault_flags & VM_FAULT_DIRTY) {
870 			vm_page_dirty(fs.m);
871 			vm_pager_page_unswapped(fs.m);
872 		}
873 	}
874 
875 	/*
876 	 * Page had better still be busy
877 	 */
878 	KASSERT(fs.m->flags & PG_BUSY,
879 		("vm_fault: page %p not busy!", fs.m));
880 	/*
881 	 * Sanity check: page must be completely valid or it is not fit to
882 	 * map into user space.  vm_pager_get_pages() ensures this.
883 	 */
884 	if (fs.m->valid != VM_PAGE_BITS_ALL) {
885 		vm_page_zero_invalid(fs.m, TRUE);
886 		printf("Warning: page %p partially invalid on fault\n", fs.m);
887 	}
888 	VM_OBJECT_UNLOCK(fs.object);
889 
890 	/*
891 	 * Put this page into the physical map.  We had to do the unlock above
892 	 * because pmap_enter() may sleep.  We don't put the page
893 	 * back on the active queue until later so that the pageout daemon
894 	 * won't find it (yet).
895 	 */
896 	pmap_enter(fs.map->pmap, vaddr, fs.m, prot, wired);
897 	if (((fault_flags & VM_FAULT_WIRE_MASK) == 0) && (wired == 0)) {
898 		vm_fault_prefault(fs.map->pmap, vaddr, fs.entry);
899 	}
900 	VM_OBJECT_LOCK(fs.object);
901 	vm_page_lock_queues();
902 	vm_page_flag_set(fs.m, PG_REFERENCED);
903 
904 	/*
905 	 * If the page is not wired down, then put it where the pageout daemon
906 	 * can find it.
907 	 */
908 	if (fault_flags & VM_FAULT_WIRE_MASK) {
909 		if (wired)
910 			vm_page_wire(fs.m);
911 		else
912 			vm_page_unwire(fs.m, 1);
913 	} else {
914 		vm_page_activate(fs.m);
915 	}
916 	vm_page_wakeup(fs.m);
917 	vm_page_unlock_queues();
918 
919 	/*
920 	 * Unlock everything, and return
921 	 */
922 	unlock_and_deallocate(&fs);
923 	PROC_LOCK(curproc);
924 	if ((curproc->p_sflag & PS_INMEM) && curproc->p_stats) {
925 		if (hardfault) {
926 			curproc->p_stats->p_ru.ru_majflt++;
927 		} else {
928 			curproc->p_stats->p_ru.ru_minflt++;
929 		}
930 	}
931 	PROC_UNLOCK(curproc);
932 
933 	return (KERN_SUCCESS);
934 }
935 
936 /*
937  * vm_fault_prefault provides a quick way of clustering
938  * pagefaults into a processes address space.  It is a "cousin"
939  * of vm_map_pmap_enter, except it runs at page fault time instead
940  * of mmap time.
941  */
942 static void
943 vm_fault_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry)
944 {
945 	int i;
946 	vm_offset_t addr, starta;
947 	vm_pindex_t pindex;
948 	vm_page_t m;
949 	vm_object_t object;
950 
951 	if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
952 		return;
953 
954 	object = entry->object.vm_object;
955 
956 	starta = addra - PFBAK * PAGE_SIZE;
957 	if (starta < entry->start) {
958 		starta = entry->start;
959 	} else if (starta > addra) {
960 		starta = 0;
961 	}
962 
963 	for (i = 0; i < PAGEORDER_SIZE; i++) {
964 		vm_object_t backing_object, lobject;
965 
966 		addr = addra + prefault_pageorder[i];
967 		if (addr > addra + (PFFOR * PAGE_SIZE))
968 			addr = 0;
969 
970 		if (addr < starta || addr >= entry->end)
971 			continue;
972 
973 		if (!pmap_is_prefaultable(pmap, addr))
974 			continue;
975 
976 		pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
977 		lobject = object;
978 		VM_OBJECT_LOCK(lobject);
979 		while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
980 		    lobject->type == OBJT_DEFAULT &&
981 		    (backing_object = lobject->backing_object) != NULL) {
982 			if (lobject->backing_object_offset & PAGE_MASK)
983 				break;
984 			pindex += lobject->backing_object_offset >> PAGE_SHIFT;
985 			VM_OBJECT_LOCK(backing_object);
986 			VM_OBJECT_UNLOCK(lobject);
987 			lobject = backing_object;
988 		}
989 		/*
990 		 * give-up when a page is not in memory
991 		 */
992 		if (m == NULL) {
993 			VM_OBJECT_UNLOCK(lobject);
994 			break;
995 		}
996 		if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
997 			(m->busy == 0) &&
998 		    (m->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
999 
1000 			vm_page_lock_queues();
1001 			if (VM_PAGE_INQUEUE1(m, PQ_CACHE))
1002 				vm_page_deactivate(m);
1003 			pmap_enter_quick(pmap, addr, m, entry->protection);
1004 			vm_page_unlock_queues();
1005 		}
1006 		VM_OBJECT_UNLOCK(lobject);
1007 	}
1008 }
1009 
1010 /*
1011  *	vm_fault_quick:
1012  *
1013  *	Ensure that the requested virtual address, which may be in userland,
1014  *	is valid.  Fault-in the page if necessary.  Return -1 on failure.
1015  */
1016 int
1017 vm_fault_quick(caddr_t v, int prot)
1018 {
1019 	int r;
1020 
1021 	if (prot & VM_PROT_WRITE)
1022 		r = subyte(v, fubyte(v));
1023 	else
1024 		r = fubyte(v);
1025 	return(r);
1026 }
1027 
1028 /*
1029  *	vm_fault_wire:
1030  *
1031  *	Wire down a range of virtual addresses in a map.
1032  */
1033 int
1034 vm_fault_wire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1035     boolean_t user_wire, boolean_t fictitious)
1036 {
1037 	vm_offset_t va;
1038 	int rv;
1039 
1040 	/*
1041 	 * We simulate a fault to get the page and enter it in the physical
1042 	 * map.  For user wiring, we only ask for read access on currently
1043 	 * read-only sections.
1044 	 */
1045 	for (va = start; va < end; va += PAGE_SIZE) {
1046 		rv = vm_fault(map, va,
1047 		    user_wire ? VM_PROT_READ : VM_PROT_READ | VM_PROT_WRITE,
1048 		    user_wire ? VM_FAULT_USER_WIRE : VM_FAULT_CHANGE_WIRING);
1049 		if (rv) {
1050 			if (va != start)
1051 				vm_fault_unwire(map, start, va, fictitious);
1052 			return (rv);
1053 		}
1054 	}
1055 	return (KERN_SUCCESS);
1056 }
1057 
1058 /*
1059  *	vm_fault_unwire:
1060  *
1061  *	Unwire a range of virtual addresses in a map.
1062  */
1063 void
1064 vm_fault_unwire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1065     boolean_t fictitious)
1066 {
1067 	vm_paddr_t pa;
1068 	vm_offset_t va;
1069 	pmap_t pmap;
1070 
1071 	pmap = vm_map_pmap(map);
1072 
1073 	/*
1074 	 * Since the pages are wired down, we must be able to get their
1075 	 * mappings from the physical map system.
1076 	 */
1077 	for (va = start; va < end; va += PAGE_SIZE) {
1078 		pa = pmap_extract(pmap, va);
1079 		if (pa != 0) {
1080 			pmap_change_wiring(pmap, va, FALSE);
1081 			if (!fictitious) {
1082 				vm_page_lock_queues();
1083 				vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1);
1084 				vm_page_unlock_queues();
1085 			}
1086 		}
1087 	}
1088 }
1089 
1090 /*
1091  *	Routine:
1092  *		vm_fault_copy_entry
1093  *	Function:
1094  *		Copy all of the pages from a wired-down map entry to another.
1095  *
1096  *	In/out conditions:
1097  *		The source and destination maps must be locked for write.
1098  *		The source map entry must be wired down (or be a sharing map
1099  *		entry corresponding to a main map entry that is wired down).
1100  */
1101 void
1102 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry)
1103 	vm_map_t dst_map;
1104 	vm_map_t src_map;
1105 	vm_map_entry_t dst_entry;
1106 	vm_map_entry_t src_entry;
1107 {
1108 	vm_object_t backing_object, dst_object, object;
1109 	vm_object_t src_object;
1110 	vm_ooffset_t dst_offset;
1111 	vm_ooffset_t src_offset;
1112 	vm_pindex_t pindex;
1113 	vm_prot_t prot;
1114 	vm_offset_t vaddr;
1115 	vm_page_t dst_m;
1116 	vm_page_t src_m;
1117 
1118 #ifdef	lint
1119 	src_map++;
1120 #endif	/* lint */
1121 
1122 	src_object = src_entry->object.vm_object;
1123 	src_offset = src_entry->offset;
1124 
1125 	/*
1126 	 * Create the top-level object for the destination entry. (Doesn't
1127 	 * actually shadow anything - we copy the pages directly.)
1128 	 */
1129 	dst_object = vm_object_allocate(OBJT_DEFAULT,
1130 	    OFF_TO_IDX(dst_entry->end - dst_entry->start));
1131 
1132 	VM_OBJECT_LOCK(dst_object);
1133 	dst_entry->object.vm_object = dst_object;
1134 	dst_entry->offset = 0;
1135 
1136 	prot = dst_entry->max_protection;
1137 
1138 	/*
1139 	 * Loop through all of the pages in the entry's range, copying each
1140 	 * one from the source object (it should be there) to the destination
1141 	 * object.
1142 	 */
1143 	for (vaddr = dst_entry->start, dst_offset = 0;
1144 	    vaddr < dst_entry->end;
1145 	    vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) {
1146 
1147 		/*
1148 		 * Allocate a page in the destination object
1149 		 */
1150 		do {
1151 			dst_m = vm_page_alloc(dst_object,
1152 				OFF_TO_IDX(dst_offset), VM_ALLOC_NORMAL);
1153 			if (dst_m == NULL) {
1154 				VM_OBJECT_UNLOCK(dst_object);
1155 				VM_WAIT;
1156 				VM_OBJECT_LOCK(dst_object);
1157 			}
1158 		} while (dst_m == NULL);
1159 
1160 		/*
1161 		 * Find the page in the source object, and copy it in.
1162 		 * (Because the source is wired down, the page will be in
1163 		 * memory.)
1164 		 */
1165 		VM_OBJECT_LOCK(src_object);
1166 		object = src_object;
1167 		pindex = 0;
1168 		while ((src_m = vm_page_lookup(object, pindex +
1169 		    OFF_TO_IDX(dst_offset + src_offset))) == NULL &&
1170 		    (src_entry->protection & VM_PROT_WRITE) == 0 &&
1171 		    (backing_object = object->backing_object) != NULL) {
1172 			/*
1173 			 * Allow fallback to backing objects if we are reading.
1174 			 */
1175 			VM_OBJECT_LOCK(backing_object);
1176 			pindex += OFF_TO_IDX(object->backing_object_offset);
1177 			VM_OBJECT_UNLOCK(object);
1178 			object = backing_object;
1179 		}
1180 		if (src_m == NULL)
1181 			panic("vm_fault_copy_wired: page missing");
1182 		pmap_copy_page(src_m, dst_m);
1183 		VM_OBJECT_UNLOCK(object);
1184 		dst_m->valid = VM_PAGE_BITS_ALL;
1185 		VM_OBJECT_UNLOCK(dst_object);
1186 
1187 		/*
1188 		 * Enter it in the pmap...
1189 		 */
1190 		pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE);
1191 		VM_OBJECT_LOCK(dst_object);
1192 		vm_page_lock_queues();
1193 		if ((prot & VM_PROT_WRITE) != 0)
1194 			vm_page_flag_set(dst_m, PG_WRITEABLE);
1195 
1196 		/*
1197 		 * Mark it no longer busy, and put it on the active list.
1198 		 */
1199 		vm_page_activate(dst_m);
1200 		vm_page_wakeup(dst_m);
1201 		vm_page_unlock_queues();
1202 	}
1203 	VM_OBJECT_UNLOCK(dst_object);
1204 }
1205 
1206 
1207 /*
1208  * This routine checks around the requested page for other pages that
1209  * might be able to be faulted in.  This routine brackets the viable
1210  * pages for the pages to be paged in.
1211  *
1212  * Inputs:
1213  *	m, rbehind, rahead
1214  *
1215  * Outputs:
1216  *  marray (array of vm_page_t), reqpage (index of requested page)
1217  *
1218  * Return value:
1219  *  number of pages in marray
1220  *
1221  * This routine can't block.
1222  */
1223 static int
1224 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage)
1225 	vm_page_t m;
1226 	int rbehind;
1227 	int rahead;
1228 	vm_page_t *marray;
1229 	int *reqpage;
1230 {
1231 	int i,j;
1232 	vm_object_t object;
1233 	vm_pindex_t pindex, startpindex, endpindex, tpindex;
1234 	vm_page_t rtm;
1235 	int cbehind, cahead;
1236 
1237 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1238 
1239 	object = m->object;
1240 	pindex = m->pindex;
1241 
1242 	/*
1243 	 * we don't fault-ahead for device pager
1244 	 */
1245 	if (object->type == OBJT_DEVICE) {
1246 		*reqpage = 0;
1247 		marray[0] = m;
1248 		return 1;
1249 	}
1250 
1251 	/*
1252 	 * if the requested page is not available, then give up now
1253 	 */
1254 	if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1255 		return 0;
1256 	}
1257 
1258 	if ((cbehind == 0) && (cahead == 0)) {
1259 		*reqpage = 0;
1260 		marray[0] = m;
1261 		return 1;
1262 	}
1263 
1264 	if (rahead > cahead) {
1265 		rahead = cahead;
1266 	}
1267 
1268 	if (rbehind > cbehind) {
1269 		rbehind = cbehind;
1270 	}
1271 
1272 	/*
1273 	 * try to do any readahead that we might have free pages for.
1274 	 */
1275 	if ((rahead + rbehind) >
1276 		((cnt.v_free_count + cnt.v_cache_count) - cnt.v_free_reserved)) {
1277 		pagedaemon_wakeup();
1278 		marray[0] = m;
1279 		*reqpage = 0;
1280 		return 1;
1281 	}
1282 
1283 	/*
1284 	 * scan backward for the read behind pages -- in memory
1285 	 */
1286 	if (pindex > 0) {
1287 		if (rbehind > pindex) {
1288 			rbehind = pindex;
1289 			startpindex = 0;
1290 		} else {
1291 			startpindex = pindex - rbehind;
1292 		}
1293 
1294 		if ((rtm = TAILQ_PREV(m, pglist, listq)) != NULL &&
1295 		    rtm->pindex >= startpindex)
1296 			startpindex = rtm->pindex + 1;
1297 
1298 		for (i = 0, tpindex = startpindex; tpindex < pindex; i++, tpindex++) {
1299 
1300 			rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
1301 			if (rtm == NULL) {
1302 				vm_page_lock_queues();
1303 				for (j = 0; j < i; j++) {
1304 					vm_page_free(marray[j]);
1305 				}
1306 				vm_page_unlock_queues();
1307 				marray[0] = m;
1308 				*reqpage = 0;
1309 				return 1;
1310 			}
1311 
1312 			marray[i] = rtm;
1313 		}
1314 	} else {
1315 		startpindex = 0;
1316 		i = 0;
1317 	}
1318 
1319 	marray[i] = m;
1320 	/* page offset of the required page */
1321 	*reqpage = i;
1322 
1323 	tpindex = pindex + 1;
1324 	i++;
1325 
1326 	/*
1327 	 * scan forward for the read ahead pages
1328 	 */
1329 	endpindex = tpindex + rahead;
1330 	if ((rtm = TAILQ_NEXT(m, listq)) != NULL && rtm->pindex < endpindex)
1331 		endpindex = rtm->pindex;
1332 	if (endpindex > object->size)
1333 		endpindex = object->size;
1334 
1335 	for (; tpindex < endpindex; i++, tpindex++) {
1336 
1337 		rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
1338 		if (rtm == NULL) {
1339 			break;
1340 		}
1341 
1342 		marray[i] = rtm;
1343 	}
1344 
1345 	/* return number of bytes of pages */
1346 	return i;
1347 }
1348