xref: /freebsd/sys/vm/vm_fault.c (revision acd3428b7d3e94cef0e1881c868cb4b131d4ff41)
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_wakeup(fs->m);
137 	vm_page_lock_queues();
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 VPO_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->oflags & VPO_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->oflags & VPO_BUSY) ||
514 					    (mt->flags & (PG_FICTITIOUS | PG_UNMANAGED)) ||
515 						mt->hold_count ||
516 						mt->wire_count)
517 						continue;
518 					pmap_remove_all(mt);
519 					if (mt->dirty) {
520 						vm_page_deactivate(mt);
521 					} else {
522 						vm_page_cache(mt);
523 					}
524 				}
525 				vm_page_unlock_queues();
526 				ahead += behind;
527 				behind = 0;
528 			}
529 			if (is_first_object_locked)
530 				VM_OBJECT_UNLOCK(fs.first_object);
531 			/*
532 			 * now we find out if any other pages should be paged
533 			 * in at this time this routine checks to see if the
534 			 * pages surrounding this fault reside in the same
535 			 * object as the page for this fault.  If they do,
536 			 * then they are faulted in also into the object.  The
537 			 * array "marray" returned contains an array of
538 			 * vm_page_t structs where one of them is the
539 			 * vm_page_t passed to the routine.  The reqpage
540 			 * return value is the index into the marray for the
541 			 * vm_page_t passed to the routine.
542 			 *
543 			 * fs.m plus the additional pages are VPO_BUSY'd.
544 			 *
545 			 * XXX vm_fault_additional_pages() can block
546 			 * without releasing the map lock.
547 			 */
548 			faultcount = vm_fault_additional_pages(
549 			    fs.m, behind, ahead, marray, &reqpage);
550 
551 			/*
552 			 * update lastr imperfectly (we do not know how much
553 			 * getpages will actually read), but good enough.
554 			 *
555 			 * XXX The following assignment modifies the map
556 			 * without holding a write lock on it.
557 			 */
558 			fs.entry->lastr = fs.pindex + faultcount - behind;
559 
560 			/*
561 			 * Call the pager to retrieve the data, if any, after
562 			 * releasing the lock on the map.  We hold a ref on
563 			 * fs.object and the pages are VPO_BUSY'd.
564 			 */
565 			unlock_map(&fs);
566 
567 			rv = faultcount ?
568 			    vm_pager_get_pages(fs.object, marray, faultcount,
569 				reqpage) : VM_PAGER_FAIL;
570 
571 			if (rv == VM_PAGER_OK) {
572 				/*
573 				 * Found the page. Leave it busy while we play
574 				 * with it.
575 				 */
576 
577 				/*
578 				 * Relookup in case pager changed page. Pager
579 				 * is responsible for disposition of old page
580 				 * if moved.
581 				 */
582 				fs.m = vm_page_lookup(fs.object, fs.pindex);
583 				if (!fs.m) {
584 					unlock_and_deallocate(&fs);
585 					goto RetryFault;
586 				}
587 
588 				hardfault++;
589 				break; /* break to PAGE HAS BEEN FOUND */
590 			}
591 			/*
592 			 * Remove the bogus page (which does not exist at this
593 			 * object/offset); before doing so, we must get back
594 			 * our object lock to preserve our invariant.
595 			 *
596 			 * Also wake up any other process that may want to bring
597 			 * in this page.
598 			 *
599 			 * If this is the top-level object, we must leave the
600 			 * busy page to prevent another process from rushing
601 			 * past us, and inserting the page in that object at
602 			 * the same time that we are.
603 			 */
604 			if (rv == VM_PAGER_ERROR)
605 				printf("vm_fault: pager read error, pid %d (%s)\n",
606 				    curproc->p_pid, curproc->p_comm);
607 			/*
608 			 * Data outside the range of the pager or an I/O error
609 			 */
610 			/*
611 			 * XXX - the check for kernel_map is a kludge to work
612 			 * around having the machine panic on a kernel space
613 			 * fault w/ I/O error.
614 			 */
615 			if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) ||
616 				(rv == VM_PAGER_BAD)) {
617 				vm_page_lock_queues();
618 				vm_page_free(fs.m);
619 				vm_page_unlock_queues();
620 				fs.m = NULL;
621 				unlock_and_deallocate(&fs);
622 				return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
623 			}
624 			if (fs.object != fs.first_object) {
625 				vm_page_lock_queues();
626 				vm_page_free(fs.m);
627 				vm_page_unlock_queues();
628 				fs.m = NULL;
629 				/*
630 				 * XXX - we cannot just fall out at this
631 				 * point, m has been freed and is invalid!
632 				 */
633 			}
634 		}
635 
636 		/*
637 		 * We get here if the object has default pager (or unwiring)
638 		 * or the pager doesn't have the page.
639 		 */
640 		if (fs.object == fs.first_object)
641 			fs.first_m = fs.m;
642 
643 		/*
644 		 * Move on to the next object.  Lock the next object before
645 		 * unlocking the current one.
646 		 */
647 		fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset);
648 		next_object = fs.object->backing_object;
649 		if (next_object == NULL) {
650 			/*
651 			 * If there's no object left, fill the page in the top
652 			 * object with zeros.
653 			 */
654 			if (fs.object != fs.first_object) {
655 				vm_object_pip_wakeup(fs.object);
656 				VM_OBJECT_UNLOCK(fs.object);
657 
658 				fs.object = fs.first_object;
659 				fs.pindex = fs.first_pindex;
660 				fs.m = fs.first_m;
661 				VM_OBJECT_LOCK(fs.object);
662 			}
663 			fs.first_m = NULL;
664 
665 			/*
666 			 * Zero the page if necessary and mark it valid.
667 			 */
668 			if ((fs.m->flags & PG_ZERO) == 0) {
669 				pmap_zero_page(fs.m);
670 			} else {
671 				atomic_add_int(&cnt.v_ozfod, 1);
672 			}
673 			atomic_add_int(&cnt.v_zfod, 1);
674 			fs.m->valid = VM_PAGE_BITS_ALL;
675 			break;	/* break to PAGE HAS BEEN FOUND */
676 		} else {
677 			KASSERT(fs.object != next_object,
678 			    ("object loop %p", next_object));
679 			VM_OBJECT_LOCK(next_object);
680 			vm_object_pip_add(next_object, 1);
681 			if (fs.object != fs.first_object)
682 				vm_object_pip_wakeup(fs.object);
683 			VM_OBJECT_UNLOCK(fs.object);
684 			fs.object = next_object;
685 		}
686 	}
687 
688 	KASSERT((fs.m->oflags & VPO_BUSY) != 0,
689 	    ("vm_fault: not busy after main loop"));
690 
691 	/*
692 	 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
693 	 * is held.]
694 	 */
695 
696 	/*
697 	 * If the page is being written, but isn't already owned by the
698 	 * top-level object, we have to copy it into a new page owned by the
699 	 * top-level object.
700 	 */
701 	if (fs.object != fs.first_object) {
702 		/*
703 		 * We only really need to copy if we want to write it.
704 		 */
705 		if (fault_type & VM_PROT_WRITE) {
706 			/*
707 			 * This allows pages to be virtually copied from a
708 			 * backing_object into the first_object, where the
709 			 * backing object has no other refs to it, and cannot
710 			 * gain any more refs.  Instead of a bcopy, we just
711 			 * move the page from the backing object to the
712 			 * first object.  Note that we must mark the page
713 			 * dirty in the first object so that it will go out
714 			 * to swap when needed.
715 			 */
716 			is_first_object_locked = FALSE;
717 			if (
718 				/*
719 				 * Only one shadow object
720 				 */
721 				(fs.object->shadow_count == 1) &&
722 				/*
723 				 * No COW refs, except us
724 				 */
725 				(fs.object->ref_count == 1) &&
726 				/*
727 				 * No one else can look this object up
728 				 */
729 				(fs.object->handle == NULL) &&
730 				/*
731 				 * No other ways to look the object up
732 				 */
733 				((fs.object->type == OBJT_DEFAULT) ||
734 				 (fs.object->type == OBJT_SWAP)) &&
735 			    (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object)) &&
736 				/*
737 				 * We don't chase down the shadow chain
738 				 */
739 			    fs.object == fs.first_object->backing_object) {
740 				vm_page_lock_queues();
741 				/*
742 				 * get rid of the unnecessary page
743 				 */
744 				vm_page_free(fs.first_m);
745 				/*
746 				 * grab the page and put it into the
747 				 * process'es object.  The page is
748 				 * automatically made dirty.
749 				 */
750 				vm_page_rename(fs.m, fs.first_object, fs.first_pindex);
751 				vm_page_busy(fs.m);
752 				vm_page_unlock_queues();
753 				fs.first_m = fs.m;
754 				fs.m = NULL;
755 				atomic_add_int(&cnt.v_cow_optim, 1);
756 			} else {
757 				/*
758 				 * Oh, well, lets copy it.
759 				 */
760 				pmap_copy_page(fs.m, fs.first_m);
761 				fs.first_m->valid = VM_PAGE_BITS_ALL;
762 			}
763 			if (fs.m) {
764 				/*
765 				 * We no longer need the old page or object.
766 				 */
767 				release_page(&fs);
768 			}
769 			/*
770 			 * fs.object != fs.first_object due to above
771 			 * conditional
772 			 */
773 			vm_object_pip_wakeup(fs.object);
774 			VM_OBJECT_UNLOCK(fs.object);
775 			/*
776 			 * Only use the new page below...
777 			 */
778 			fs.object = fs.first_object;
779 			fs.pindex = fs.first_pindex;
780 			fs.m = fs.first_m;
781 			if (!is_first_object_locked)
782 				VM_OBJECT_LOCK(fs.object);
783 			atomic_add_int(&cnt.v_cow_faults, 1);
784 		} else {
785 			prot &= ~VM_PROT_WRITE;
786 		}
787 	}
788 
789 	/*
790 	 * We must verify that the maps have not changed since our last
791 	 * lookup.
792 	 */
793 	if (!fs.lookup_still_valid) {
794 		vm_object_t retry_object;
795 		vm_pindex_t retry_pindex;
796 		vm_prot_t retry_prot;
797 
798 		if (!vm_map_trylock_read(fs.map)) {
799 			release_page(&fs);
800 			unlock_and_deallocate(&fs);
801 			goto RetryFault;
802 		}
803 		fs.lookup_still_valid = TRUE;
804 		if (fs.map->timestamp != map_generation) {
805 			result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
806 			    &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
807 
808 			/*
809 			 * If we don't need the page any longer, put it on the inactive
810 			 * list (the easiest thing to do here).  If no one needs it,
811 			 * pageout will grab it eventually.
812 			 */
813 			if (result != KERN_SUCCESS) {
814 				release_page(&fs);
815 				unlock_and_deallocate(&fs);
816 
817 				/*
818 				 * If retry of map lookup would have blocked then
819 				 * retry fault from start.
820 				 */
821 				if (result == KERN_FAILURE)
822 					goto RetryFault;
823 				return (result);
824 			}
825 			if ((retry_object != fs.first_object) ||
826 			    (retry_pindex != fs.first_pindex)) {
827 				release_page(&fs);
828 				unlock_and_deallocate(&fs);
829 				goto RetryFault;
830 			}
831 
832 			/*
833 			 * Check whether the protection has changed or the object has
834 			 * been copied while we left the map unlocked. Changing from
835 			 * read to write permission is OK - we leave the page
836 			 * write-protected, and catch the write fault. Changing from
837 			 * write to read permission means that we can't mark the page
838 			 * write-enabled after all.
839 			 */
840 			prot &= retry_prot;
841 		}
842 	}
843 	if (prot & VM_PROT_WRITE) {
844 		vm_page_lock_queues();
845 		vm_page_flag_set(fs.m, PG_WRITEABLE);
846 		vm_page_unlock_queues();
847 		vm_object_set_writeable_dirty(fs.object);
848 
849 		/*
850 		 * If the fault is a write, we know that this page is being
851 		 * written NOW so dirty it explicitly to save on
852 		 * pmap_is_modified() calls later.
853 		 *
854 		 * If this is a NOSYNC mmap we do not want to set VPO_NOSYNC
855 		 * if the page is already dirty to prevent data written with
856 		 * the expectation of being synced from not being synced.
857 		 * Likewise if this entry does not request NOSYNC then make
858 		 * sure the page isn't marked NOSYNC.  Applications sharing
859 		 * data should use the same flags to avoid ping ponging.
860 		 *
861 		 * Also tell the backing pager, if any, that it should remove
862 		 * any swap backing since the page is now dirty.
863 		 */
864 		if (fs.entry->eflags & MAP_ENTRY_NOSYNC) {
865 			if (fs.m->dirty == 0)
866 				fs.m->oflags |= VPO_NOSYNC;
867 		} else {
868 			fs.m->oflags &= ~VPO_NOSYNC;
869 		}
870 		if (fault_flags & VM_FAULT_DIRTY) {
871 			vm_page_dirty(fs.m);
872 			vm_pager_page_unswapped(fs.m);
873 		}
874 	}
875 
876 	/*
877 	 * Page had better still be busy
878 	 */
879 	KASSERT(fs.m->oflags & VPO_BUSY,
880 		("vm_fault: page %p not busy!", fs.m));
881 	/*
882 	 * Sanity check: page must be completely valid or it is not fit to
883 	 * map into user space.  vm_pager_get_pages() ensures this.
884 	 */
885 	if (fs.m->valid != VM_PAGE_BITS_ALL) {
886 		vm_page_zero_invalid(fs.m, TRUE);
887 		printf("Warning: page %p partially invalid on fault\n", fs.m);
888 	}
889 	VM_OBJECT_UNLOCK(fs.object);
890 
891 	/*
892 	 * Put this page into the physical map.  We had to do the unlock above
893 	 * because pmap_enter() may sleep.  We don't put the page
894 	 * back on the active queue until later so that the pageout daemon
895 	 * won't find it (yet).
896 	 */
897 	pmap_enter(fs.map->pmap, vaddr, fs.m, prot, wired);
898 	if (((fault_flags & VM_FAULT_WIRE_MASK) == 0) && (wired == 0)) {
899 		vm_fault_prefault(fs.map->pmap, vaddr, fs.entry);
900 	}
901 	VM_OBJECT_LOCK(fs.object);
902 	vm_page_lock_queues();
903 	vm_page_flag_set(fs.m, PG_REFERENCED);
904 
905 	/*
906 	 * If the page is not wired down, then put it where the pageout daemon
907 	 * can find it.
908 	 */
909 	if (fault_flags & VM_FAULT_WIRE_MASK) {
910 		if (wired)
911 			vm_page_wire(fs.m);
912 		else
913 			vm_page_unwire(fs.m, 1);
914 	} else {
915 		vm_page_activate(fs.m);
916 	}
917 	vm_page_unlock_queues();
918 	vm_page_wakeup(fs.m);
919 
920 	/*
921 	 * Unlock everything, and return
922 	 */
923 	unlock_and_deallocate(&fs);
924 	PROC_LOCK(curproc);
925 	if ((curproc->p_sflag & PS_INMEM) && curproc->p_stats) {
926 		if (hardfault) {
927 			curproc->p_stats->p_ru.ru_majflt++;
928 		} else {
929 			curproc->p_stats->p_ru.ru_minflt++;
930 		}
931 	}
932 	PROC_UNLOCK(curproc);
933 
934 	return (KERN_SUCCESS);
935 }
936 
937 /*
938  * vm_fault_prefault provides a quick way of clustering
939  * pagefaults into a processes address space.  It is a "cousin"
940  * of vm_map_pmap_enter, except it runs at page fault time instead
941  * of mmap time.
942  */
943 static void
944 vm_fault_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry)
945 {
946 	int i;
947 	vm_offset_t addr, starta;
948 	vm_pindex_t pindex;
949 	vm_page_t m;
950 	vm_object_t object;
951 
952 	if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
953 		return;
954 
955 	object = entry->object.vm_object;
956 
957 	starta = addra - PFBAK * PAGE_SIZE;
958 	if (starta < entry->start) {
959 		starta = entry->start;
960 	} else if (starta > addra) {
961 		starta = 0;
962 	}
963 
964 	for (i = 0; i < PAGEORDER_SIZE; i++) {
965 		vm_object_t backing_object, lobject;
966 
967 		addr = addra + prefault_pageorder[i];
968 		if (addr > addra + (PFFOR * PAGE_SIZE))
969 			addr = 0;
970 
971 		if (addr < starta || addr >= entry->end)
972 			continue;
973 
974 		if (!pmap_is_prefaultable(pmap, addr))
975 			continue;
976 
977 		pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
978 		lobject = object;
979 		VM_OBJECT_LOCK(lobject);
980 		while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
981 		    lobject->type == OBJT_DEFAULT &&
982 		    (backing_object = lobject->backing_object) != NULL) {
983 			if (lobject->backing_object_offset & PAGE_MASK)
984 				break;
985 			pindex += lobject->backing_object_offset >> PAGE_SHIFT;
986 			VM_OBJECT_LOCK(backing_object);
987 			VM_OBJECT_UNLOCK(lobject);
988 			lobject = backing_object;
989 		}
990 		/*
991 		 * give-up when a page is not in memory
992 		 */
993 		if (m == NULL) {
994 			VM_OBJECT_UNLOCK(lobject);
995 			break;
996 		}
997 		if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
998 			(m->busy == 0) &&
999 		    (m->flags & PG_FICTITIOUS) == 0) {
1000 
1001 			vm_page_lock_queues();
1002 			if (VM_PAGE_INQUEUE1(m, PQ_CACHE))
1003 				vm_page_deactivate(m);
1004 			pmap_enter_quick(pmap, addr, m, entry->protection);
1005 			vm_page_unlock_queues();
1006 		}
1007 		VM_OBJECT_UNLOCK(lobject);
1008 	}
1009 }
1010 
1011 /*
1012  *	vm_fault_quick:
1013  *
1014  *	Ensure that the requested virtual address, which may be in userland,
1015  *	is valid.  Fault-in the page if necessary.  Return -1 on failure.
1016  */
1017 int
1018 vm_fault_quick(caddr_t v, int prot)
1019 {
1020 	int r;
1021 
1022 	if (prot & VM_PROT_WRITE)
1023 		r = subyte(v, fubyte(v));
1024 	else
1025 		r = fubyte(v);
1026 	return(r);
1027 }
1028 
1029 /*
1030  *	vm_fault_wire:
1031  *
1032  *	Wire down a range of virtual addresses in a map.
1033  */
1034 int
1035 vm_fault_wire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1036     boolean_t user_wire, boolean_t fictitious)
1037 {
1038 	vm_offset_t va;
1039 	int rv;
1040 
1041 	/*
1042 	 * We simulate a fault to get the page and enter it in the physical
1043 	 * map.  For user wiring, we only ask for read access on currently
1044 	 * read-only sections.
1045 	 */
1046 	for (va = start; va < end; va += PAGE_SIZE) {
1047 		rv = vm_fault(map, va,
1048 		    user_wire ? VM_PROT_READ : VM_PROT_READ | VM_PROT_WRITE,
1049 		    user_wire ? VM_FAULT_USER_WIRE : VM_FAULT_CHANGE_WIRING);
1050 		if (rv) {
1051 			if (va != start)
1052 				vm_fault_unwire(map, start, va, fictitious);
1053 			return (rv);
1054 		}
1055 	}
1056 	return (KERN_SUCCESS);
1057 }
1058 
1059 /*
1060  *	vm_fault_unwire:
1061  *
1062  *	Unwire a range of virtual addresses in a map.
1063  */
1064 void
1065 vm_fault_unwire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1066     boolean_t fictitious)
1067 {
1068 	vm_paddr_t pa;
1069 	vm_offset_t va;
1070 	pmap_t pmap;
1071 
1072 	pmap = vm_map_pmap(map);
1073 
1074 	/*
1075 	 * Since the pages are wired down, we must be able to get their
1076 	 * mappings from the physical map system.
1077 	 */
1078 	for (va = start; va < end; va += PAGE_SIZE) {
1079 		pa = pmap_extract(pmap, va);
1080 		if (pa != 0) {
1081 			pmap_change_wiring(pmap, va, FALSE);
1082 			if (!fictitious) {
1083 				vm_page_lock_queues();
1084 				vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1);
1085 				vm_page_unlock_queues();
1086 			}
1087 		}
1088 	}
1089 }
1090 
1091 /*
1092  *	Routine:
1093  *		vm_fault_copy_entry
1094  *	Function:
1095  *		Copy all of the pages from a wired-down map entry to another.
1096  *
1097  *	In/out conditions:
1098  *		The source and destination maps must be locked for write.
1099  *		The source map entry must be wired down (or be a sharing map
1100  *		entry corresponding to a main map entry that is wired down).
1101  */
1102 void
1103 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry)
1104 	vm_map_t dst_map;
1105 	vm_map_t src_map;
1106 	vm_map_entry_t dst_entry;
1107 	vm_map_entry_t src_entry;
1108 {
1109 	vm_object_t backing_object, dst_object, object;
1110 	vm_object_t src_object;
1111 	vm_ooffset_t dst_offset;
1112 	vm_ooffset_t src_offset;
1113 	vm_pindex_t pindex;
1114 	vm_prot_t prot;
1115 	vm_offset_t vaddr;
1116 	vm_page_t dst_m;
1117 	vm_page_t src_m;
1118 
1119 #ifdef	lint
1120 	src_map++;
1121 #endif	/* lint */
1122 
1123 	src_object = src_entry->object.vm_object;
1124 	src_offset = src_entry->offset;
1125 
1126 	/*
1127 	 * Create the top-level object for the destination entry. (Doesn't
1128 	 * actually shadow anything - we copy the pages directly.)
1129 	 */
1130 	dst_object = vm_object_allocate(OBJT_DEFAULT,
1131 	    OFF_TO_IDX(dst_entry->end - dst_entry->start));
1132 
1133 	VM_OBJECT_LOCK(dst_object);
1134 	dst_entry->object.vm_object = dst_object;
1135 	dst_entry->offset = 0;
1136 
1137 	prot = dst_entry->max_protection;
1138 
1139 	/*
1140 	 * Loop through all of the pages in the entry's range, copying each
1141 	 * one from the source object (it should be there) to the destination
1142 	 * object.
1143 	 */
1144 	for (vaddr = dst_entry->start, dst_offset = 0;
1145 	    vaddr < dst_entry->end;
1146 	    vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) {
1147 
1148 		/*
1149 		 * Allocate a page in the destination object
1150 		 */
1151 		do {
1152 			dst_m = vm_page_alloc(dst_object,
1153 				OFF_TO_IDX(dst_offset), VM_ALLOC_NORMAL);
1154 			if (dst_m == NULL) {
1155 				VM_OBJECT_UNLOCK(dst_object);
1156 				VM_WAIT;
1157 				VM_OBJECT_LOCK(dst_object);
1158 			}
1159 		} while (dst_m == NULL);
1160 
1161 		/*
1162 		 * Find the page in the source object, and copy it in.
1163 		 * (Because the source is wired down, the page will be in
1164 		 * memory.)
1165 		 */
1166 		VM_OBJECT_LOCK(src_object);
1167 		object = src_object;
1168 		pindex = 0;
1169 		while ((src_m = vm_page_lookup(object, pindex +
1170 		    OFF_TO_IDX(dst_offset + src_offset))) == NULL &&
1171 		    (src_entry->protection & VM_PROT_WRITE) == 0 &&
1172 		    (backing_object = object->backing_object) != NULL) {
1173 			/*
1174 			 * Allow fallback to backing objects if we are reading.
1175 			 */
1176 			VM_OBJECT_LOCK(backing_object);
1177 			pindex += OFF_TO_IDX(object->backing_object_offset);
1178 			VM_OBJECT_UNLOCK(object);
1179 			object = backing_object;
1180 		}
1181 		if (src_m == NULL)
1182 			panic("vm_fault_copy_wired: page missing");
1183 		pmap_copy_page(src_m, dst_m);
1184 		VM_OBJECT_UNLOCK(object);
1185 		dst_m->valid = VM_PAGE_BITS_ALL;
1186 		VM_OBJECT_UNLOCK(dst_object);
1187 
1188 		/*
1189 		 * Enter it in the pmap...
1190 		 */
1191 		pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE);
1192 		VM_OBJECT_LOCK(dst_object);
1193 		vm_page_lock_queues();
1194 		if ((prot & VM_PROT_WRITE) != 0)
1195 			vm_page_flag_set(dst_m, PG_WRITEABLE);
1196 
1197 		/*
1198 		 * Mark it no longer busy, and put it on the active list.
1199 		 */
1200 		vm_page_activate(dst_m);
1201 		vm_page_unlock_queues();
1202 		vm_page_wakeup(dst_m);
1203 	}
1204 	VM_OBJECT_UNLOCK(dst_object);
1205 }
1206 
1207 
1208 /*
1209  * This routine checks around the requested page for other pages that
1210  * might be able to be faulted in.  This routine brackets the viable
1211  * pages for the pages to be paged in.
1212  *
1213  * Inputs:
1214  *	m, rbehind, rahead
1215  *
1216  * Outputs:
1217  *  marray (array of vm_page_t), reqpage (index of requested page)
1218  *
1219  * Return value:
1220  *  number of pages in marray
1221  *
1222  * This routine can't block.
1223  */
1224 static int
1225 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage)
1226 	vm_page_t m;
1227 	int rbehind;
1228 	int rahead;
1229 	vm_page_t *marray;
1230 	int *reqpage;
1231 {
1232 	int i,j;
1233 	vm_object_t object;
1234 	vm_pindex_t pindex, startpindex, endpindex, tpindex;
1235 	vm_page_t rtm;
1236 	int cbehind, cahead;
1237 
1238 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1239 
1240 	object = m->object;
1241 	pindex = m->pindex;
1242 
1243 	/*
1244 	 * we don't fault-ahead for device pager
1245 	 */
1246 	if (object->type == OBJT_DEVICE) {
1247 		*reqpage = 0;
1248 		marray[0] = m;
1249 		return 1;
1250 	}
1251 
1252 	/*
1253 	 * if the requested page is not available, then give up now
1254 	 */
1255 	if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1256 		return 0;
1257 	}
1258 
1259 	if ((cbehind == 0) && (cahead == 0)) {
1260 		*reqpage = 0;
1261 		marray[0] = m;
1262 		return 1;
1263 	}
1264 
1265 	if (rahead > cahead) {
1266 		rahead = cahead;
1267 	}
1268 
1269 	if (rbehind > cbehind) {
1270 		rbehind = cbehind;
1271 	}
1272 
1273 	/*
1274 	 * try to do any readahead that we might have free pages for.
1275 	 */
1276 	if ((rahead + rbehind) >
1277 		((cnt.v_free_count + cnt.v_cache_count) - cnt.v_free_reserved)) {
1278 		pagedaemon_wakeup();
1279 		marray[0] = m;
1280 		*reqpage = 0;
1281 		return 1;
1282 	}
1283 
1284 	/*
1285 	 * scan backward for the read behind pages -- in memory
1286 	 */
1287 	if (pindex > 0) {
1288 		if (rbehind > pindex) {
1289 			rbehind = pindex;
1290 			startpindex = 0;
1291 		} else {
1292 			startpindex = pindex - rbehind;
1293 		}
1294 
1295 		if ((rtm = TAILQ_PREV(m, pglist, listq)) != NULL &&
1296 		    rtm->pindex >= startpindex)
1297 			startpindex = rtm->pindex + 1;
1298 
1299 		for (i = 0, tpindex = startpindex; tpindex < pindex; i++, tpindex++) {
1300 
1301 			rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
1302 			if (rtm == NULL) {
1303 				vm_page_lock_queues();
1304 				for (j = 0; j < i; j++) {
1305 					vm_page_free(marray[j]);
1306 				}
1307 				vm_page_unlock_queues();
1308 				marray[0] = m;
1309 				*reqpage = 0;
1310 				return 1;
1311 			}
1312 
1313 			marray[i] = rtm;
1314 		}
1315 	} else {
1316 		startpindex = 0;
1317 		i = 0;
1318 	}
1319 
1320 	marray[i] = m;
1321 	/* page offset of the required page */
1322 	*reqpage = i;
1323 
1324 	tpindex = pindex + 1;
1325 	i++;
1326 
1327 	/*
1328 	 * scan forward for the read ahead pages
1329 	 */
1330 	endpindex = tpindex + rahead;
1331 	if ((rtm = TAILQ_NEXT(m, listq)) != NULL && rtm->pindex < endpindex)
1332 		endpindex = rtm->pindex;
1333 	if (endpindex > object->size)
1334 		endpindex = object->size;
1335 
1336 	for (; tpindex < endpindex; i++, tpindex++) {
1337 
1338 		rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
1339 		if (rtm == NULL) {
1340 			break;
1341 		}
1342 
1343 		marray[i] = rtm;
1344 	}
1345 
1346 	/* return number of bytes of pages */
1347 	return i;
1348 }
1349