xref: /freebsd/sys/vm/vm_fault.c (revision 5d9359578951e7d657021353ab09a71cdbad0cbc)
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_ktrace.h"
78 #include "opt_vm.h"
79 
80 #include <sys/param.h>
81 #include <sys/systm.h>
82 #include <sys/kernel.h>
83 #include <sys/lock.h>
84 #include <sys/mman.h>
85 #include <sys/proc.h>
86 #include <sys/resourcevar.h>
87 #include <sys/rwlock.h>
88 #include <sys/sysctl.h>
89 #include <sys/vmmeter.h>
90 #include <sys/vnode.h>
91 #ifdef KTRACE
92 #include <sys/ktrace.h>
93 #endif
94 
95 #include <vm/vm.h>
96 #include <vm/vm_param.h>
97 #include <vm/pmap.h>
98 #include <vm/vm_map.h>
99 #include <vm/vm_object.h>
100 #include <vm/vm_page.h>
101 #include <vm/vm_pageout.h>
102 #include <vm/vm_kern.h>
103 #include <vm/vm_pager.h>
104 #include <vm/vm_extern.h>
105 #include <vm/vm_reserv.h>
106 
107 #define PFBAK 4
108 #define PFFOR 4
109 
110 static int vm_fault_additional_pages(vm_page_t, int, int, vm_page_t *, int *);
111 
112 #define	VM_FAULT_READ_BEHIND	8
113 #define	VM_FAULT_READ_DEFAULT	(1 + VM_FAULT_READ_AHEAD_INIT)
114 #define	VM_FAULT_READ_MAX	(1 + VM_FAULT_READ_AHEAD_MAX)
115 #define	VM_FAULT_NINCR		(VM_FAULT_READ_MAX / VM_FAULT_READ_BEHIND)
116 #define	VM_FAULT_SUM		(VM_FAULT_NINCR * (VM_FAULT_NINCR + 1) / 2)
117 
118 #define	VM_FAULT_DONTNEED_MIN	1048576
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 void vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr,
134 	    int ahead);
135 static void vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
136 	    int faultcount, int reqpage);
137 
138 static inline void
139 release_page(struct faultstate *fs)
140 {
141 
142 	vm_page_xunbusy(fs->m);
143 	vm_page_lock(fs->m);
144 	vm_page_deactivate(fs->m);
145 	vm_page_unlock(fs->m);
146 	fs->m = NULL;
147 }
148 
149 static inline void
150 unlock_map(struct faultstate *fs)
151 {
152 
153 	if (fs->lookup_still_valid) {
154 		vm_map_lookup_done(fs->map, fs->entry);
155 		fs->lookup_still_valid = FALSE;
156 	}
157 }
158 
159 static void
160 unlock_and_deallocate(struct faultstate *fs)
161 {
162 
163 	vm_object_pip_wakeup(fs->object);
164 	VM_OBJECT_WUNLOCK(fs->object);
165 	if (fs->object != fs->first_object) {
166 		VM_OBJECT_WLOCK(fs->first_object);
167 		vm_page_lock(fs->first_m);
168 		vm_page_free(fs->first_m);
169 		vm_page_unlock(fs->first_m);
170 		vm_object_pip_wakeup(fs->first_object);
171 		VM_OBJECT_WUNLOCK(fs->first_object);
172 		fs->first_m = NULL;
173 	}
174 	vm_object_deallocate(fs->first_object);
175 	unlock_map(fs);
176 	if (fs->vp != NULL) {
177 		vput(fs->vp);
178 		fs->vp = NULL;
179 	}
180 }
181 
182 static void
183 vm_fault_dirty(vm_map_entry_t entry, vm_page_t m, vm_prot_t prot,
184     vm_prot_t fault_type, int fault_flags, boolean_t set_wd)
185 {
186 	boolean_t need_dirty;
187 
188 	if (((prot & VM_PROT_WRITE) == 0 &&
189 	    (fault_flags & VM_FAULT_DIRTY) == 0) ||
190 	    (m->oflags & VPO_UNMANAGED) != 0)
191 		return;
192 
193 	VM_OBJECT_ASSERT_LOCKED(m->object);
194 
195 	need_dirty = ((fault_type & VM_PROT_WRITE) != 0 &&
196 	    (fault_flags & VM_FAULT_CHANGE_WIRING) == 0) ||
197 	    (fault_flags & VM_FAULT_DIRTY) != 0;
198 
199 	if (set_wd)
200 		vm_object_set_writeable_dirty(m->object);
201 	else
202 		/*
203 		 * If two callers of vm_fault_dirty() with set_wd ==
204 		 * FALSE, one for the map entry with MAP_ENTRY_NOSYNC
205 		 * flag set, other with flag clear, race, it is
206 		 * possible for the no-NOSYNC thread to see m->dirty
207 		 * != 0 and not clear VPO_NOSYNC.  Take vm_page lock
208 		 * around manipulation of VPO_NOSYNC and
209 		 * vm_page_dirty() call, to avoid the race and keep
210 		 * m->oflags consistent.
211 		 */
212 		vm_page_lock(m);
213 
214 	/*
215 	 * If this is a NOSYNC mmap we do not want to set VPO_NOSYNC
216 	 * if the page is already dirty to prevent data written with
217 	 * the expectation of being synced from not being synced.
218 	 * Likewise if this entry does not request NOSYNC then make
219 	 * sure the page isn't marked NOSYNC.  Applications sharing
220 	 * data should use the same flags to avoid ping ponging.
221 	 */
222 	if ((entry->eflags & MAP_ENTRY_NOSYNC) != 0) {
223 		if (m->dirty == 0) {
224 			m->oflags |= VPO_NOSYNC;
225 		}
226 	} else {
227 		m->oflags &= ~VPO_NOSYNC;
228 	}
229 
230 	/*
231 	 * If the fault is a write, we know that this page is being
232 	 * written NOW so dirty it explicitly to save on
233 	 * pmap_is_modified() calls later.
234 	 *
235 	 * Also tell the backing pager, if any, that it should remove
236 	 * any swap backing since the page is now dirty.
237 	 */
238 	if (need_dirty)
239 		vm_page_dirty(m);
240 	if (!set_wd)
241 		vm_page_unlock(m);
242 	if (need_dirty)
243 		vm_pager_page_unswapped(m);
244 }
245 
246 /*
247  * TRYPAGER - used by vm_fault to calculate whether the pager for the
248  *	      current object *might* contain the page.
249  *
250  *	      default objects are zero-fill, there is no real pager.
251  */
252 #define TRYPAGER	(fs.object->type != OBJT_DEFAULT && \
253 			((fault_flags & VM_FAULT_CHANGE_WIRING) == 0 || wired))
254 
255 /*
256  *	vm_fault:
257  *
258  *	Handle a page fault occurring at the given address,
259  *	requiring the given permissions, in the map specified.
260  *	If successful, the page is inserted into the
261  *	associated physical map.
262  *
263  *	NOTE: the given address should be truncated to the
264  *	proper page address.
265  *
266  *	KERN_SUCCESS is returned if the page fault is handled; otherwise,
267  *	a standard error specifying why the fault is fatal is returned.
268  *
269  *	The map in question must be referenced, and remains so.
270  *	Caller may hold no locks.
271  */
272 int
273 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
274     int fault_flags)
275 {
276 	struct thread *td;
277 	int result;
278 
279 	td = curthread;
280 	if ((td->td_pflags & TDP_NOFAULTING) != 0)
281 		return (KERN_PROTECTION_FAILURE);
282 #ifdef KTRACE
283 	if (map != kernel_map && KTRPOINT(td, KTR_FAULT))
284 		ktrfault(vaddr, fault_type);
285 #endif
286 	result = vm_fault_hold(map, trunc_page(vaddr), fault_type, fault_flags,
287 	    NULL);
288 #ifdef KTRACE
289 	if (map != kernel_map && KTRPOINT(td, KTR_FAULTEND))
290 		ktrfaultend(result);
291 #endif
292 	return (result);
293 }
294 
295 int
296 vm_fault_hold(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
297     int fault_flags, vm_page_t *m_hold)
298 {
299 	vm_prot_t prot;
300 	int alloc_req, era, faultcount, nera, reqpage, result;
301 	boolean_t growstack, is_first_object_locked, wired;
302 	int map_generation;
303 	vm_object_t next_object;
304 	vm_page_t marray[VM_FAULT_READ_MAX];
305 	int hardfault;
306 	struct faultstate fs;
307 	struct vnode *vp;
308 	vm_page_t m;
309 	int ahead, behind, cluster_offset, error, locked;
310 
311 	hardfault = 0;
312 	growstack = TRUE;
313 	PCPU_INC(cnt.v_vm_faults);
314 	fs.vp = NULL;
315 	faultcount = reqpage = 0;
316 
317 RetryFault:;
318 
319 	/*
320 	 * Find the backing store object and offset into it to begin the
321 	 * search.
322 	 */
323 	fs.map = map;
324 	result = vm_map_lookup(&fs.map, vaddr, fault_type, &fs.entry,
325 	    &fs.first_object, &fs.first_pindex, &prot, &wired);
326 	if (result != KERN_SUCCESS) {
327 		if (growstack && result == KERN_INVALID_ADDRESS &&
328 		    map != kernel_map) {
329 			result = vm_map_growstack(curproc, vaddr);
330 			if (result != KERN_SUCCESS)
331 				return (KERN_FAILURE);
332 			growstack = FALSE;
333 			goto RetryFault;
334 		}
335 		return (result);
336 	}
337 
338 	map_generation = fs.map->timestamp;
339 
340 	if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
341 		panic("vm_fault: fault on nofault entry, addr: %lx",
342 		    (u_long)vaddr);
343 	}
344 
345 	if (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION &&
346 	    fs.entry->wiring_thread != curthread) {
347 		vm_map_unlock_read(fs.map);
348 		vm_map_lock(fs.map);
349 		if (vm_map_lookup_entry(fs.map, vaddr, &fs.entry) &&
350 		    (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION)) {
351 			fs.entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
352 			vm_map_unlock_and_wait(fs.map, 0);
353 		} else
354 			vm_map_unlock(fs.map);
355 		goto RetryFault;
356 	}
357 
358 	if (wired)
359 		fault_type = prot | (fault_type & VM_PROT_COPY);
360 
361 	if (fs.vp == NULL /* avoid locked vnode leak */ &&
362 	    (fault_flags & (VM_FAULT_CHANGE_WIRING | VM_FAULT_DIRTY)) == 0 &&
363 	    /* avoid calling vm_object_set_writeable_dirty() */
364 	    ((prot & VM_PROT_WRITE) == 0 ||
365 	    (fs.first_object->type != OBJT_VNODE &&
366 	    (fs.first_object->flags & OBJ_TMPFS_NODE) == 0) ||
367 	    (fs.first_object->flags & OBJ_MIGHTBEDIRTY) != 0)) {
368 		VM_OBJECT_RLOCK(fs.first_object);
369 		if ((prot & VM_PROT_WRITE) != 0 &&
370 		    (fs.first_object->type == OBJT_VNODE ||
371 		    (fs.first_object->flags & OBJ_TMPFS_NODE) != 0) &&
372 		    (fs.first_object->flags & OBJ_MIGHTBEDIRTY) == 0)
373 			goto fast_failed;
374 		m = vm_page_lookup(fs.first_object, fs.first_pindex);
375 		/* A busy page can be mapped for read|execute access. */
376 		if (m == NULL || ((prot & VM_PROT_WRITE) != 0 &&
377 		    vm_page_busied(m)) || m->valid != VM_PAGE_BITS_ALL)
378 			goto fast_failed;
379 		result = pmap_enter(fs.map->pmap, vaddr, m, prot,
380 		   fault_type | PMAP_ENTER_NOSLEEP | (wired ? PMAP_ENTER_WIRED :
381 		   0), 0);
382 		if (result != KERN_SUCCESS)
383 			goto fast_failed;
384 		if (m_hold != NULL) {
385 			*m_hold = m;
386 			vm_page_lock(m);
387 			vm_page_hold(m);
388 			vm_page_unlock(m);
389 		}
390 		vm_fault_dirty(fs.entry, m, prot, fault_type, fault_flags,
391 		    FALSE);
392 		VM_OBJECT_RUNLOCK(fs.first_object);
393 		if (!wired)
394 			vm_fault_prefault(&fs, vaddr, 0, 0);
395 		vm_map_lookup_done(fs.map, fs.entry);
396 		curthread->td_ru.ru_minflt++;
397 		return (KERN_SUCCESS);
398 fast_failed:
399 		if (!VM_OBJECT_TRYUPGRADE(fs.first_object)) {
400 			VM_OBJECT_RUNLOCK(fs.first_object);
401 			VM_OBJECT_WLOCK(fs.first_object);
402 		}
403 	} else {
404 		VM_OBJECT_WLOCK(fs.first_object);
405 	}
406 
407 	/*
408 	 * Make a reference to this object to prevent its disposal while we
409 	 * are messing with it.  Once we have the reference, the map is free
410 	 * to be diddled.  Since objects reference their shadows (and copies),
411 	 * they will stay around as well.
412 	 *
413 	 * Bump the paging-in-progress count to prevent size changes (e.g.
414 	 * truncation operations) during I/O.  This must be done after
415 	 * obtaining the vnode lock in order to avoid possible deadlocks.
416 	 */
417 	vm_object_reference_locked(fs.first_object);
418 	vm_object_pip_add(fs.first_object, 1);
419 
420 	fs.lookup_still_valid = TRUE;
421 
422 	fs.first_m = NULL;
423 
424 	/*
425 	 * Search for the page at object/offset.
426 	 */
427 	fs.object = fs.first_object;
428 	fs.pindex = fs.first_pindex;
429 	while (TRUE) {
430 		/*
431 		 * If the object is dead, we stop here
432 		 */
433 		if (fs.object->flags & OBJ_DEAD) {
434 			unlock_and_deallocate(&fs);
435 			return (KERN_PROTECTION_FAILURE);
436 		}
437 
438 		/*
439 		 * See if page is resident
440 		 */
441 		fs.m = vm_page_lookup(fs.object, fs.pindex);
442 		if (fs.m != NULL) {
443 			/*
444 			 * Wait/Retry if the page is busy.  We have to do this
445 			 * if the page is either exclusive or shared busy
446 			 * because the vm_pager may be using read busy for
447 			 * pageouts (and even pageins if it is the vnode
448 			 * pager), and we could end up trying to pagein and
449 			 * pageout the same page simultaneously.
450 			 *
451 			 * We can theoretically allow the busy case on a read
452 			 * fault if the page is marked valid, but since such
453 			 * pages are typically already pmap'd, putting that
454 			 * special case in might be more effort then it is
455 			 * worth.  We cannot under any circumstances mess
456 			 * around with a shared busied page except, perhaps,
457 			 * to pmap it.
458 			 */
459 			if (vm_page_busied(fs.m)) {
460 				/*
461 				 * Reference the page before unlocking and
462 				 * sleeping so that the page daemon is less
463 				 * likely to reclaim it.
464 				 */
465 				vm_page_aflag_set(fs.m, PGA_REFERENCED);
466 				if (fs.object != fs.first_object) {
467 					if (!VM_OBJECT_TRYWLOCK(
468 					    fs.first_object)) {
469 						VM_OBJECT_WUNLOCK(fs.object);
470 						VM_OBJECT_WLOCK(fs.first_object);
471 						VM_OBJECT_WLOCK(fs.object);
472 					}
473 					vm_page_lock(fs.first_m);
474 					vm_page_free(fs.first_m);
475 					vm_page_unlock(fs.first_m);
476 					vm_object_pip_wakeup(fs.first_object);
477 					VM_OBJECT_WUNLOCK(fs.first_object);
478 					fs.first_m = NULL;
479 				}
480 				unlock_map(&fs);
481 				if (fs.m == vm_page_lookup(fs.object,
482 				    fs.pindex)) {
483 					vm_page_sleep_if_busy(fs.m, "vmpfw");
484 				}
485 				vm_object_pip_wakeup(fs.object);
486 				VM_OBJECT_WUNLOCK(fs.object);
487 				PCPU_INC(cnt.v_intrans);
488 				vm_object_deallocate(fs.first_object);
489 				goto RetryFault;
490 			}
491 			vm_page_lock(fs.m);
492 			vm_page_remque(fs.m);
493 			vm_page_unlock(fs.m);
494 
495 			/*
496 			 * Mark page busy for other processes, and the
497 			 * pagedaemon.  If it still isn't completely valid
498 			 * (readable), jump to readrest, else break-out ( we
499 			 * found the page ).
500 			 */
501 			vm_page_xbusy(fs.m);
502 			if (fs.m->valid != VM_PAGE_BITS_ALL)
503 				goto readrest;
504 			break;
505 		}
506 
507 		/*
508 		 * Page is not resident, If this is the search termination
509 		 * or the pager might contain the page, allocate a new page.
510 		 */
511 		if (TRYPAGER || fs.object == fs.first_object) {
512 			if (fs.pindex >= fs.object->size) {
513 				unlock_and_deallocate(&fs);
514 				return (KERN_PROTECTION_FAILURE);
515 			}
516 
517 			/*
518 			 * Allocate a new page for this object/offset pair.
519 			 *
520 			 * Unlocked read of the p_flag is harmless. At
521 			 * worst, the P_KILLED might be not observed
522 			 * there, and allocation can fail, causing
523 			 * restart and new reading of the p_flag.
524 			 */
525 			fs.m = NULL;
526 			if (!vm_page_count_severe() || P_KILLED(curproc)) {
527 #if VM_NRESERVLEVEL > 0
528 				vm_object_color(fs.object, atop(vaddr) -
529 				    fs.pindex);
530 #endif
531 				alloc_req = P_KILLED(curproc) ?
532 				    VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL;
533 				if (fs.object->type != OBJT_VNODE &&
534 				    fs.object->backing_object == NULL)
535 					alloc_req |= VM_ALLOC_ZERO;
536 				fs.m = vm_page_alloc(fs.object, fs.pindex,
537 				    alloc_req);
538 			}
539 			if (fs.m == NULL) {
540 				unlock_and_deallocate(&fs);
541 				VM_WAITPFAULT;
542 				goto RetryFault;
543 			} else if (fs.m->valid == VM_PAGE_BITS_ALL)
544 				break;
545 		}
546 
547 readrest:
548 		/*
549 		 * We have found a valid page or we have allocated a new page.
550 		 * The page thus may not be valid or may not be entirely
551 		 * valid.
552 		 *
553 		 * Attempt to fault-in the page if there is a chance that the
554 		 * pager has it, and potentially fault in additional pages
555 		 * at the same time.
556 		 */
557 		if (TRYPAGER) {
558 			int rv;
559 			u_char behavior = vm_map_entry_behavior(fs.entry);
560 
561 			era = fs.entry->read_ahead;
562 			if (behavior == MAP_ENTRY_BEHAV_RANDOM ||
563 			    P_KILLED(curproc)) {
564 				behind = 0;
565 				nera = 0;
566 				ahead = 0;
567 			} else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) {
568 				behind = 0;
569 				nera = VM_FAULT_READ_AHEAD_MAX;
570 				ahead = nera;
571 				if (fs.pindex == fs.entry->next_read)
572 					vm_fault_dontneed(&fs, vaddr, ahead);
573 			} else if (fs.pindex == fs.entry->next_read) {
574 				/*
575 				 * This is a sequential fault.  Arithmetically
576 				 * increase the requested number of pages in
577 				 * the read-ahead window.  The requested
578 				 * number of pages is "# of sequential faults
579 				 * x (read ahead min + 1) + read ahead min"
580 				 */
581 				behind = 0;
582 				nera = VM_FAULT_READ_AHEAD_MIN;
583 				if (era > 0) {
584 					nera += era + 1;
585 					if (nera > VM_FAULT_READ_AHEAD_MAX)
586 						nera = VM_FAULT_READ_AHEAD_MAX;
587 				}
588 				ahead = nera;
589 				if (era == VM_FAULT_READ_AHEAD_MAX)
590 					vm_fault_dontneed(&fs, vaddr, ahead);
591 			} else {
592 				/*
593 				 * This is a non-sequential fault.  Request a
594 				 * cluster of pages that is aligned to a
595 				 * VM_FAULT_READ_DEFAULT page offset boundary
596 				 * within the object.  Alignment to a page
597 				 * offset boundary is more likely to coincide
598 				 * with the underlying file system block than
599 				 * alignment to a virtual address boundary.
600 				 */
601 				cluster_offset = fs.pindex %
602 				    VM_FAULT_READ_DEFAULT;
603 				behind = ulmin(cluster_offset,
604 				    atop(vaddr - fs.entry->start));
605 				nera = 0;
606 				ahead = VM_FAULT_READ_DEFAULT - 1 -
607 				    cluster_offset;
608 			}
609 			ahead = ulmin(ahead, atop(fs.entry->end - vaddr) - 1);
610 			if (era != nera)
611 				fs.entry->read_ahead = nera;
612 
613 			/*
614 			 * Call the pager to retrieve the data, if any, after
615 			 * releasing the lock on the map.  We hold a ref on
616 			 * fs.object and the pages are exclusive busied.
617 			 */
618 			unlock_map(&fs);
619 
620 			if (fs.object->type == OBJT_VNODE) {
621 				vp = fs.object->handle;
622 				if (vp == fs.vp)
623 					goto vnode_locked;
624 				else if (fs.vp != NULL) {
625 					vput(fs.vp);
626 					fs.vp = NULL;
627 				}
628 				locked = VOP_ISLOCKED(vp);
629 
630 				if (locked != LK_EXCLUSIVE)
631 					locked = LK_SHARED;
632 				/* Do not sleep for vnode lock while fs.m is busy */
633 				error = vget(vp, locked | LK_CANRECURSE |
634 				    LK_NOWAIT, curthread);
635 				if (error != 0) {
636 					vhold(vp);
637 					release_page(&fs);
638 					unlock_and_deallocate(&fs);
639 					error = vget(vp, locked | LK_RETRY |
640 					    LK_CANRECURSE, curthread);
641 					vdrop(vp);
642 					fs.vp = vp;
643 					KASSERT(error == 0,
644 					    ("vm_fault: vget failed"));
645 					goto RetryFault;
646 				}
647 				fs.vp = vp;
648 			}
649 vnode_locked:
650 			KASSERT(fs.vp == NULL || !fs.map->system_map,
651 			    ("vm_fault: vnode-backed object mapped by system map"));
652 
653 			/*
654 			 * now we find out if any other pages should be paged
655 			 * in at this time this routine checks to see if the
656 			 * pages surrounding this fault reside in the same
657 			 * object as the page for this fault.  If they do,
658 			 * then they are faulted in also into the object.  The
659 			 * array "marray" returned contains an array of
660 			 * vm_page_t structs where one of them is the
661 			 * vm_page_t passed to the routine.  The reqpage
662 			 * return value is the index into the marray for the
663 			 * vm_page_t passed to the routine.
664 			 *
665 			 * fs.m plus the additional pages are exclusive busied.
666 			 */
667 			faultcount = vm_fault_additional_pages(
668 			    fs.m, behind, ahead, marray, &reqpage);
669 
670 			rv = faultcount ?
671 			    vm_pager_get_pages(fs.object, marray, faultcount,
672 				reqpage) : VM_PAGER_FAIL;
673 
674 			if (rv == VM_PAGER_OK) {
675 				/*
676 				 * Found the page. Leave it busy while we play
677 				 * with it.
678 				 */
679 
680 				/*
681 				 * Relookup in case pager changed page. Pager
682 				 * is responsible for disposition of old page
683 				 * if moved.
684 				 */
685 				fs.m = vm_page_lookup(fs.object, fs.pindex);
686 				if (!fs.m) {
687 					unlock_and_deallocate(&fs);
688 					goto RetryFault;
689 				}
690 
691 				hardfault++;
692 				break; /* break to PAGE HAS BEEN FOUND */
693 			}
694 			/*
695 			 * Remove the bogus page (which does not exist at this
696 			 * object/offset); before doing so, we must get back
697 			 * our object lock to preserve our invariant.
698 			 *
699 			 * Also wake up any other process that may want to bring
700 			 * in this page.
701 			 *
702 			 * If this is the top-level object, we must leave the
703 			 * busy page to prevent another process from rushing
704 			 * past us, and inserting the page in that object at
705 			 * the same time that we are.
706 			 */
707 			if (rv == VM_PAGER_ERROR)
708 				printf("vm_fault: pager read error, pid %d (%s)\n",
709 				    curproc->p_pid, curproc->p_comm);
710 			/*
711 			 * Data outside the range of the pager or an I/O error
712 			 */
713 			/*
714 			 * XXX - the check for kernel_map is a kludge to work
715 			 * around having the machine panic on a kernel space
716 			 * fault w/ I/O error.
717 			 */
718 			if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) ||
719 				(rv == VM_PAGER_BAD)) {
720 				vm_page_lock(fs.m);
721 				vm_page_free(fs.m);
722 				vm_page_unlock(fs.m);
723 				fs.m = NULL;
724 				unlock_and_deallocate(&fs);
725 				return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
726 			}
727 			if (fs.object != fs.first_object) {
728 				vm_page_lock(fs.m);
729 				vm_page_free(fs.m);
730 				vm_page_unlock(fs.m);
731 				fs.m = NULL;
732 				/*
733 				 * XXX - we cannot just fall out at this
734 				 * point, m has been freed and is invalid!
735 				 */
736 			}
737 		}
738 
739 		/*
740 		 * We get here if the object has default pager (or unwiring)
741 		 * or the pager doesn't have the page.
742 		 */
743 		if (fs.object == fs.first_object)
744 			fs.first_m = fs.m;
745 
746 		/*
747 		 * Move on to the next object.  Lock the next object before
748 		 * unlocking the current one.
749 		 */
750 		fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset);
751 		next_object = fs.object->backing_object;
752 		if (next_object == NULL) {
753 			/*
754 			 * If there's no object left, fill the page in the top
755 			 * object with zeros.
756 			 */
757 			if (fs.object != fs.first_object) {
758 				vm_object_pip_wakeup(fs.object);
759 				VM_OBJECT_WUNLOCK(fs.object);
760 
761 				fs.object = fs.first_object;
762 				fs.pindex = fs.first_pindex;
763 				fs.m = fs.first_m;
764 				VM_OBJECT_WLOCK(fs.object);
765 			}
766 			fs.first_m = NULL;
767 
768 			/*
769 			 * Zero the page if necessary and mark it valid.
770 			 */
771 			if ((fs.m->flags & PG_ZERO) == 0) {
772 				pmap_zero_page(fs.m);
773 			} else {
774 				PCPU_INC(cnt.v_ozfod);
775 			}
776 			PCPU_INC(cnt.v_zfod);
777 			fs.m->valid = VM_PAGE_BITS_ALL;
778 			/* Don't try to prefault neighboring pages. */
779 			faultcount = 1;
780 			break;	/* break to PAGE HAS BEEN FOUND */
781 		} else {
782 			KASSERT(fs.object != next_object,
783 			    ("object loop %p", next_object));
784 			VM_OBJECT_WLOCK(next_object);
785 			vm_object_pip_add(next_object, 1);
786 			if (fs.object != fs.first_object)
787 				vm_object_pip_wakeup(fs.object);
788 			VM_OBJECT_WUNLOCK(fs.object);
789 			fs.object = next_object;
790 		}
791 	}
792 
793 	vm_page_assert_xbusied(fs.m);
794 
795 	/*
796 	 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
797 	 * is held.]
798 	 */
799 
800 	/*
801 	 * If the page is being written, but isn't already owned by the
802 	 * top-level object, we have to copy it into a new page owned by the
803 	 * top-level object.
804 	 */
805 	if (fs.object != fs.first_object) {
806 		/*
807 		 * We only really need to copy if we want to write it.
808 		 */
809 		if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
810 			/*
811 			 * This allows pages to be virtually copied from a
812 			 * backing_object into the first_object, where the
813 			 * backing object has no other refs to it, and cannot
814 			 * gain any more refs.  Instead of a bcopy, we just
815 			 * move the page from the backing object to the
816 			 * first object.  Note that we must mark the page
817 			 * dirty in the first object so that it will go out
818 			 * to swap when needed.
819 			 */
820 			is_first_object_locked = FALSE;
821 			if (
822 				/*
823 				 * Only one shadow object
824 				 */
825 				(fs.object->shadow_count == 1) &&
826 				/*
827 				 * No COW refs, except us
828 				 */
829 				(fs.object->ref_count == 1) &&
830 				/*
831 				 * No one else can look this object up
832 				 */
833 				(fs.object->handle == NULL) &&
834 				/*
835 				 * No other ways to look the object up
836 				 */
837 				((fs.object->type == OBJT_DEFAULT) ||
838 				 (fs.object->type == OBJT_SWAP)) &&
839 			    (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs.first_object)) &&
840 				/*
841 				 * We don't chase down the shadow chain
842 				 */
843 			    fs.object == fs.first_object->backing_object) {
844 				/*
845 				 * get rid of the unnecessary page
846 				 */
847 				vm_page_lock(fs.first_m);
848 				vm_page_free(fs.first_m);
849 				vm_page_unlock(fs.first_m);
850 				/*
851 				 * grab the page and put it into the
852 				 * process'es object.  The page is
853 				 * automatically made dirty.
854 				 */
855 				if (vm_page_rename(fs.m, fs.first_object,
856 				    fs.first_pindex)) {
857 					unlock_and_deallocate(&fs);
858 					goto RetryFault;
859 				}
860 #if VM_NRESERVLEVEL > 0
861 				/*
862 				 * Rename the reservation.
863 				 */
864 				vm_reserv_rename(fs.m, fs.first_object,
865 				    fs.object, OFF_TO_IDX(
866 				    fs.first_object->backing_object_offset));
867 #endif
868 				vm_page_xbusy(fs.m);
869 				fs.first_m = fs.m;
870 				fs.m = NULL;
871 				PCPU_INC(cnt.v_cow_optim);
872 			} else {
873 				/*
874 				 * Oh, well, lets copy it.
875 				 */
876 				pmap_copy_page(fs.m, fs.first_m);
877 				fs.first_m->valid = VM_PAGE_BITS_ALL;
878 				if (wired && (fault_flags &
879 				    VM_FAULT_CHANGE_WIRING) == 0) {
880 					vm_page_lock(fs.first_m);
881 					vm_page_wire(fs.first_m);
882 					vm_page_unlock(fs.first_m);
883 
884 					vm_page_lock(fs.m);
885 					vm_page_unwire(fs.m, PQ_INACTIVE);
886 					vm_page_unlock(fs.m);
887 				}
888 				/*
889 				 * We no longer need the old page or object.
890 				 */
891 				release_page(&fs);
892 			}
893 			/*
894 			 * fs.object != fs.first_object due to above
895 			 * conditional
896 			 */
897 			vm_object_pip_wakeup(fs.object);
898 			VM_OBJECT_WUNLOCK(fs.object);
899 			/*
900 			 * Only use the new page below...
901 			 */
902 			fs.object = fs.first_object;
903 			fs.pindex = fs.first_pindex;
904 			fs.m = fs.first_m;
905 			if (!is_first_object_locked)
906 				VM_OBJECT_WLOCK(fs.object);
907 			PCPU_INC(cnt.v_cow_faults);
908 			curthread->td_cow++;
909 		} else {
910 			prot &= ~VM_PROT_WRITE;
911 		}
912 	}
913 
914 	/*
915 	 * We must verify that the maps have not changed since our last
916 	 * lookup.
917 	 */
918 	if (!fs.lookup_still_valid) {
919 		vm_object_t retry_object;
920 		vm_pindex_t retry_pindex;
921 		vm_prot_t retry_prot;
922 
923 		if (!vm_map_trylock_read(fs.map)) {
924 			release_page(&fs);
925 			unlock_and_deallocate(&fs);
926 			goto RetryFault;
927 		}
928 		fs.lookup_still_valid = TRUE;
929 		if (fs.map->timestamp != map_generation) {
930 			result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
931 			    &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
932 
933 			/*
934 			 * If we don't need the page any longer, put it on the inactive
935 			 * list (the easiest thing to do here).  If no one needs it,
936 			 * pageout will grab it eventually.
937 			 */
938 			if (result != KERN_SUCCESS) {
939 				release_page(&fs);
940 				unlock_and_deallocate(&fs);
941 
942 				/*
943 				 * If retry of map lookup would have blocked then
944 				 * retry fault from start.
945 				 */
946 				if (result == KERN_FAILURE)
947 					goto RetryFault;
948 				return (result);
949 			}
950 			if ((retry_object != fs.first_object) ||
951 			    (retry_pindex != fs.first_pindex)) {
952 				release_page(&fs);
953 				unlock_and_deallocate(&fs);
954 				goto RetryFault;
955 			}
956 
957 			/*
958 			 * Check whether the protection has changed or the object has
959 			 * been copied while we left the map unlocked. Changing from
960 			 * read to write permission is OK - we leave the page
961 			 * write-protected, and catch the write fault. Changing from
962 			 * write to read permission means that we can't mark the page
963 			 * write-enabled after all.
964 			 */
965 			prot &= retry_prot;
966 		}
967 	}
968 	/*
969 	 * If the page was filled by a pager, update the map entry's
970 	 * last read offset.  Since the pager does not return the
971 	 * actual set of pages that it read, this update is based on
972 	 * the requested set.  Typically, the requested and actual
973 	 * sets are the same.
974 	 *
975 	 * XXX The following assignment modifies the map
976 	 * without holding a write lock on it.
977 	 */
978 	if (hardfault)
979 		fs.entry->next_read = fs.pindex + faultcount - reqpage;
980 
981 	vm_fault_dirty(fs.entry, fs.m, prot, fault_type, fault_flags, TRUE);
982 	vm_page_assert_xbusied(fs.m);
983 
984 	/*
985 	 * Page must be completely valid or it is not fit to
986 	 * map into user space.  vm_pager_get_pages() ensures this.
987 	 */
988 	KASSERT(fs.m->valid == VM_PAGE_BITS_ALL,
989 	    ("vm_fault: page %p partially invalid", fs.m));
990 	VM_OBJECT_WUNLOCK(fs.object);
991 
992 	/*
993 	 * Put this page into the physical map.  We had to do the unlock above
994 	 * because pmap_enter() may sleep.  We don't put the page
995 	 * back on the active queue until later so that the pageout daemon
996 	 * won't find it (yet).
997 	 */
998 	pmap_enter(fs.map->pmap, vaddr, fs.m, prot,
999 	    fault_type | (wired ? PMAP_ENTER_WIRED : 0), 0);
1000 	if (faultcount != 1 && (fault_flags & VM_FAULT_CHANGE_WIRING) == 0 &&
1001 	    wired == 0)
1002 		vm_fault_prefault(&fs, vaddr, faultcount, reqpage);
1003 	VM_OBJECT_WLOCK(fs.object);
1004 	vm_page_lock(fs.m);
1005 
1006 	/*
1007 	 * If the page is not wired down, then put it where the pageout daemon
1008 	 * can find it.
1009 	 */
1010 	if (fault_flags & VM_FAULT_CHANGE_WIRING) {
1011 		if (wired)
1012 			vm_page_wire(fs.m);
1013 		else
1014 			vm_page_unwire(fs.m, PQ_ACTIVE);
1015 	} else
1016 		vm_page_activate(fs.m);
1017 	if (m_hold != NULL) {
1018 		*m_hold = fs.m;
1019 		vm_page_hold(fs.m);
1020 	}
1021 	vm_page_unlock(fs.m);
1022 	vm_page_xunbusy(fs.m);
1023 
1024 	/*
1025 	 * Unlock everything, and return
1026 	 */
1027 	unlock_and_deallocate(&fs);
1028 	if (hardfault) {
1029 		PCPU_INC(cnt.v_io_faults);
1030 		curthread->td_ru.ru_majflt++;
1031 	} else
1032 		curthread->td_ru.ru_minflt++;
1033 
1034 	return (KERN_SUCCESS);
1035 }
1036 
1037 /*
1038  * Speed up the reclamation of pages that precede the faulting pindex within
1039  * the first object of the shadow chain.  Essentially, perform the equivalent
1040  * to madvise(..., MADV_DONTNEED) on a large cluster of pages that precedes
1041  * the faulting pindex by the cluster size when the pages read by vm_fault()
1042  * cross a cluster-size boundary.  The cluster size is the greater of the
1043  * smallest superpage size and VM_FAULT_DONTNEED_MIN.
1044  *
1045  * When "fs->first_object" is a shadow object, the pages in the backing object
1046  * that precede the faulting pindex are deactivated by vm_fault().  So, this
1047  * function must only be concerned with pages in the first object.
1048  */
1049 static void
1050 vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr, int ahead)
1051 {
1052 	vm_map_entry_t entry;
1053 	vm_object_t first_object, object;
1054 	vm_offset_t end, start;
1055 	vm_page_t m, m_next;
1056 	vm_pindex_t pend, pstart;
1057 	vm_size_t size;
1058 
1059 	object = fs->object;
1060 	VM_OBJECT_ASSERT_WLOCKED(object);
1061 	first_object = fs->first_object;
1062 	if (first_object != object) {
1063 		if (!VM_OBJECT_TRYWLOCK(first_object)) {
1064 			VM_OBJECT_WUNLOCK(object);
1065 			VM_OBJECT_WLOCK(first_object);
1066 			VM_OBJECT_WLOCK(object);
1067 		}
1068 	}
1069 	/* Neither fictitious nor unmanaged pages can be reclaimed. */
1070 	if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) {
1071 		size = VM_FAULT_DONTNEED_MIN;
1072 		if (MAXPAGESIZES > 1 && size < pagesizes[1])
1073 			size = pagesizes[1];
1074 		end = rounddown2(vaddr, size);
1075 		if (vaddr - end >= size - PAGE_SIZE - ptoa(ahead) &&
1076 		    (entry = fs->entry)->start < end) {
1077 			if (end - entry->start < size)
1078 				start = entry->start;
1079 			else
1080 				start = end - size;
1081 			pmap_advise(fs->map->pmap, start, end, MADV_DONTNEED);
1082 			pstart = OFF_TO_IDX(entry->offset) + atop(start -
1083 			    entry->start);
1084 			m_next = vm_page_find_least(first_object, pstart);
1085 			pend = OFF_TO_IDX(entry->offset) + atop(end -
1086 			    entry->start);
1087 			while ((m = m_next) != NULL && m->pindex < pend) {
1088 				m_next = TAILQ_NEXT(m, listq);
1089 				if (m->valid != VM_PAGE_BITS_ALL ||
1090 				    vm_page_busied(m))
1091 					continue;
1092 				vm_page_lock(m);
1093 				if (m->hold_count == 0 && m->wire_count == 0)
1094 					vm_page_advise(m, MADV_DONTNEED);
1095 				vm_page_unlock(m);
1096 			}
1097 		}
1098 	}
1099 	if (first_object != object)
1100 		VM_OBJECT_WUNLOCK(first_object);
1101 }
1102 
1103 /*
1104  * vm_fault_prefault provides a quick way of clustering
1105  * pagefaults into a processes address space.  It is a "cousin"
1106  * of vm_map_pmap_enter, except it runs at page fault time instead
1107  * of mmap time.
1108  */
1109 static void
1110 vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
1111     int faultcount, int reqpage)
1112 {
1113 	pmap_t pmap;
1114 	vm_map_entry_t entry;
1115 	vm_object_t backing_object, lobject;
1116 	vm_offset_t addr, starta;
1117 	vm_pindex_t pindex;
1118 	vm_page_t m;
1119 	int backward, forward, i;
1120 
1121 	pmap = fs->map->pmap;
1122 	if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
1123 		return;
1124 
1125 	if (faultcount > 0) {
1126 		backward = reqpage;
1127 		forward = faultcount - reqpage - 1;
1128 	} else {
1129 		backward = PFBAK;
1130 		forward = PFFOR;
1131 	}
1132 	entry = fs->entry;
1133 
1134 	starta = addra - backward * PAGE_SIZE;
1135 	if (starta < entry->start) {
1136 		starta = entry->start;
1137 	} else if (starta > addra) {
1138 		starta = 0;
1139 	}
1140 
1141 	/*
1142 	 * Generate the sequence of virtual addresses that are candidates for
1143 	 * prefaulting in an outward spiral from the faulting virtual address,
1144 	 * "addra".  Specifically, the sequence is "addra - PAGE_SIZE", "addra
1145 	 * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ...
1146 	 * If the candidate address doesn't have a backing physical page, then
1147 	 * the loop immediately terminates.
1148 	 */
1149 	for (i = 0; i < 2 * imax(backward, forward); i++) {
1150 		addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE :
1151 		    PAGE_SIZE);
1152 		if (addr > addra + forward * PAGE_SIZE)
1153 			addr = 0;
1154 
1155 		if (addr < starta || addr >= entry->end)
1156 			continue;
1157 
1158 		if (!pmap_is_prefaultable(pmap, addr))
1159 			continue;
1160 
1161 		pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1162 		lobject = entry->object.vm_object;
1163 		VM_OBJECT_RLOCK(lobject);
1164 		while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1165 		    lobject->type == OBJT_DEFAULT &&
1166 		    (backing_object = lobject->backing_object) != NULL) {
1167 			KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1168 			    0, ("vm_fault_prefault: unaligned object offset"));
1169 			pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1170 			VM_OBJECT_RLOCK(backing_object);
1171 			VM_OBJECT_RUNLOCK(lobject);
1172 			lobject = backing_object;
1173 		}
1174 		if (m == NULL) {
1175 			VM_OBJECT_RUNLOCK(lobject);
1176 			break;
1177 		}
1178 		if (m->valid == VM_PAGE_BITS_ALL &&
1179 		    (m->flags & PG_FICTITIOUS) == 0)
1180 			pmap_enter_quick(pmap, addr, m, entry->protection);
1181 		VM_OBJECT_RUNLOCK(lobject);
1182 	}
1183 }
1184 
1185 /*
1186  * Hold each of the physical pages that are mapped by the specified range of
1187  * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
1188  * and allow the specified types of access, "prot".  If all of the implied
1189  * pages are successfully held, then the number of held pages is returned
1190  * together with pointers to those pages in the array "ma".  However, if any
1191  * of the pages cannot be held, -1 is returned.
1192  */
1193 int
1194 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
1195     vm_prot_t prot, vm_page_t *ma, int max_count)
1196 {
1197 	vm_offset_t end, va;
1198 	vm_page_t *mp;
1199 	int count;
1200 	boolean_t pmap_failed;
1201 
1202 	if (len == 0)
1203 		return (0);
1204 	end = round_page(addr + len);
1205 	addr = trunc_page(addr);
1206 
1207 	/*
1208 	 * Check for illegal addresses.
1209 	 */
1210 	if (addr < vm_map_min(map) || addr > end || end > vm_map_max(map))
1211 		return (-1);
1212 
1213 	if (atop(end - addr) > max_count)
1214 		panic("vm_fault_quick_hold_pages: count > max_count");
1215 	count = atop(end - addr);
1216 
1217 	/*
1218 	 * Most likely, the physical pages are resident in the pmap, so it is
1219 	 * faster to try pmap_extract_and_hold() first.
1220 	 */
1221 	pmap_failed = FALSE;
1222 	for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) {
1223 		*mp = pmap_extract_and_hold(map->pmap, va, prot);
1224 		if (*mp == NULL)
1225 			pmap_failed = TRUE;
1226 		else if ((prot & VM_PROT_WRITE) != 0 &&
1227 		    (*mp)->dirty != VM_PAGE_BITS_ALL) {
1228 			/*
1229 			 * Explicitly dirty the physical page.  Otherwise, the
1230 			 * caller's changes may go unnoticed because they are
1231 			 * performed through an unmanaged mapping or by a DMA
1232 			 * operation.
1233 			 *
1234 			 * The object lock is not held here.
1235 			 * See vm_page_clear_dirty_mask().
1236 			 */
1237 			vm_page_dirty(*mp);
1238 		}
1239 	}
1240 	if (pmap_failed) {
1241 		/*
1242 		 * One or more pages could not be held by the pmap.  Either no
1243 		 * page was mapped at the specified virtual address or that
1244 		 * mapping had insufficient permissions.  Attempt to fault in
1245 		 * and hold these pages.
1246 		 */
1247 		for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE)
1248 			if (*mp == NULL && vm_fault_hold(map, va, prot,
1249 			    VM_FAULT_NORMAL, mp) != KERN_SUCCESS)
1250 				goto error;
1251 	}
1252 	return (count);
1253 error:
1254 	for (mp = ma; mp < ma + count; mp++)
1255 		if (*mp != NULL) {
1256 			vm_page_lock(*mp);
1257 			vm_page_unhold(*mp);
1258 			vm_page_unlock(*mp);
1259 		}
1260 	return (-1);
1261 }
1262 
1263 /*
1264  *	Routine:
1265  *		vm_fault_copy_entry
1266  *	Function:
1267  *		Create new shadow object backing dst_entry with private copy of
1268  *		all underlying pages. When src_entry is equal to dst_entry,
1269  *		function implements COW for wired-down map entry. Otherwise,
1270  *		it forks wired entry into dst_map.
1271  *
1272  *	In/out conditions:
1273  *		The source and destination maps must be locked for write.
1274  *		The source map entry must be wired down (or be a sharing map
1275  *		entry corresponding to a main map entry that is wired down).
1276  */
1277 void
1278 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1279     vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1280     vm_ooffset_t *fork_charge)
1281 {
1282 	vm_object_t backing_object, dst_object, object, src_object;
1283 	vm_pindex_t dst_pindex, pindex, src_pindex;
1284 	vm_prot_t access, prot;
1285 	vm_offset_t vaddr;
1286 	vm_page_t dst_m;
1287 	vm_page_t src_m;
1288 	boolean_t upgrade;
1289 
1290 #ifdef	lint
1291 	src_map++;
1292 #endif	/* lint */
1293 
1294 	upgrade = src_entry == dst_entry;
1295 	access = prot = dst_entry->protection;
1296 
1297 	src_object = src_entry->object.vm_object;
1298 	src_pindex = OFF_TO_IDX(src_entry->offset);
1299 
1300 	if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
1301 		dst_object = src_object;
1302 		vm_object_reference(dst_object);
1303 	} else {
1304 		/*
1305 		 * Create the top-level object for the destination entry. (Doesn't
1306 		 * actually shadow anything - we copy the pages directly.)
1307 		 */
1308 		dst_object = vm_object_allocate(OBJT_DEFAULT,
1309 		    OFF_TO_IDX(dst_entry->end - dst_entry->start));
1310 #if VM_NRESERVLEVEL > 0
1311 		dst_object->flags |= OBJ_COLORED;
1312 		dst_object->pg_color = atop(dst_entry->start);
1313 #endif
1314 	}
1315 
1316 	VM_OBJECT_WLOCK(dst_object);
1317 	KASSERT(upgrade || dst_entry->object.vm_object == NULL,
1318 	    ("vm_fault_copy_entry: vm_object not NULL"));
1319 	if (src_object != dst_object) {
1320 		dst_entry->object.vm_object = dst_object;
1321 		dst_entry->offset = 0;
1322 		dst_object->charge = dst_entry->end - dst_entry->start;
1323 	}
1324 	if (fork_charge != NULL) {
1325 		KASSERT(dst_entry->cred == NULL,
1326 		    ("vm_fault_copy_entry: leaked swp charge"));
1327 		dst_object->cred = curthread->td_ucred;
1328 		crhold(dst_object->cred);
1329 		*fork_charge += dst_object->charge;
1330 	} else if (dst_object->cred == NULL) {
1331 		KASSERT(dst_entry->cred != NULL, ("no cred for entry %p",
1332 		    dst_entry));
1333 		dst_object->cred = dst_entry->cred;
1334 		dst_entry->cred = NULL;
1335 	}
1336 
1337 	/*
1338 	 * If not an upgrade, then enter the mappings in the pmap as
1339 	 * read and/or execute accesses.  Otherwise, enter them as
1340 	 * write accesses.
1341 	 *
1342 	 * A writeable large page mapping is only created if all of
1343 	 * the constituent small page mappings are modified. Marking
1344 	 * PTEs as modified on inception allows promotion to happen
1345 	 * without taking potentially large number of soft faults.
1346 	 */
1347 	if (!upgrade)
1348 		access &= ~VM_PROT_WRITE;
1349 
1350 	/*
1351 	 * Loop through all of the virtual pages within the entry's
1352 	 * range, copying each page from the source object to the
1353 	 * destination object.  Since the source is wired, those pages
1354 	 * must exist.  In contrast, the destination is pageable.
1355 	 * Since the destination object does share any backing storage
1356 	 * with the source object, all of its pages must be dirtied,
1357 	 * regardless of whether they can be written.
1358 	 */
1359 	for (vaddr = dst_entry->start, dst_pindex = 0;
1360 	    vaddr < dst_entry->end;
1361 	    vaddr += PAGE_SIZE, dst_pindex++) {
1362 again:
1363 		/*
1364 		 * Find the page in the source object, and copy it in.
1365 		 * Because the source is wired down, the page will be
1366 		 * in memory.
1367 		 */
1368 		if (src_object != dst_object)
1369 			VM_OBJECT_RLOCK(src_object);
1370 		object = src_object;
1371 		pindex = src_pindex + dst_pindex;
1372 		while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
1373 		    (backing_object = object->backing_object) != NULL) {
1374 			/*
1375 			 * Unless the source mapping is read-only or
1376 			 * it is presently being upgraded from
1377 			 * read-only, the first object in the shadow
1378 			 * chain should provide all of the pages.  In
1379 			 * other words, this loop body should never be
1380 			 * executed when the source mapping is already
1381 			 * read/write.
1382 			 */
1383 			KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 ||
1384 			    upgrade,
1385 			    ("vm_fault_copy_entry: main object missing page"));
1386 
1387 			VM_OBJECT_RLOCK(backing_object);
1388 			pindex += OFF_TO_IDX(object->backing_object_offset);
1389 			if (object != dst_object)
1390 				VM_OBJECT_RUNLOCK(object);
1391 			object = backing_object;
1392 		}
1393 		KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing"));
1394 
1395 		if (object != dst_object) {
1396 			/*
1397 			 * Allocate a page in the destination object.
1398 			 */
1399 			dst_m = vm_page_alloc(dst_object, (src_object ==
1400 			    dst_object ? src_pindex : 0) + dst_pindex,
1401 			    VM_ALLOC_NORMAL);
1402 			if (dst_m == NULL) {
1403 				VM_OBJECT_WUNLOCK(dst_object);
1404 				VM_OBJECT_RUNLOCK(object);
1405 				VM_WAIT;
1406 				VM_OBJECT_WLOCK(dst_object);
1407 				goto again;
1408 			}
1409 			pmap_copy_page(src_m, dst_m);
1410 			VM_OBJECT_RUNLOCK(object);
1411 			dst_m->valid = VM_PAGE_BITS_ALL;
1412 			dst_m->dirty = VM_PAGE_BITS_ALL;
1413 		} else {
1414 			dst_m = src_m;
1415 			if (vm_page_sleep_if_busy(dst_m, "fltupg"))
1416 				goto again;
1417 			vm_page_xbusy(dst_m);
1418 			KASSERT(dst_m->valid == VM_PAGE_BITS_ALL,
1419 			    ("invalid dst page %p", dst_m));
1420 		}
1421 		VM_OBJECT_WUNLOCK(dst_object);
1422 
1423 		/*
1424 		 * Enter it in the pmap. If a wired, copy-on-write
1425 		 * mapping is being replaced by a write-enabled
1426 		 * mapping, then wire that new mapping.
1427 		 */
1428 		pmap_enter(dst_map->pmap, vaddr, dst_m, prot,
1429 		    access | (upgrade ? PMAP_ENTER_WIRED : 0), 0);
1430 
1431 		/*
1432 		 * Mark it no longer busy, and put it on the active list.
1433 		 */
1434 		VM_OBJECT_WLOCK(dst_object);
1435 
1436 		if (upgrade) {
1437 			if (src_m != dst_m) {
1438 				vm_page_lock(src_m);
1439 				vm_page_unwire(src_m, PQ_INACTIVE);
1440 				vm_page_unlock(src_m);
1441 				vm_page_lock(dst_m);
1442 				vm_page_wire(dst_m);
1443 				vm_page_unlock(dst_m);
1444 			} else {
1445 				KASSERT(dst_m->wire_count > 0,
1446 				    ("dst_m %p is not wired", dst_m));
1447 			}
1448 		} else {
1449 			vm_page_lock(dst_m);
1450 			vm_page_activate(dst_m);
1451 			vm_page_unlock(dst_m);
1452 		}
1453 		vm_page_xunbusy(dst_m);
1454 	}
1455 	VM_OBJECT_WUNLOCK(dst_object);
1456 	if (upgrade) {
1457 		dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
1458 		vm_object_deallocate(src_object);
1459 	}
1460 }
1461 
1462 
1463 /*
1464  * This routine checks around the requested page for other pages that
1465  * might be able to be faulted in.  This routine brackets the viable
1466  * pages for the pages to be paged in.
1467  *
1468  * Inputs:
1469  *	m, rbehind, rahead
1470  *
1471  * Outputs:
1472  *  marray (array of vm_page_t), reqpage (index of requested page)
1473  *
1474  * Return value:
1475  *  number of pages in marray
1476  */
1477 static int
1478 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage)
1479 	vm_page_t m;
1480 	int rbehind;
1481 	int rahead;
1482 	vm_page_t *marray;
1483 	int *reqpage;
1484 {
1485 	int i,j;
1486 	vm_object_t object;
1487 	vm_pindex_t pindex, startpindex, endpindex, tpindex;
1488 	vm_page_t rtm;
1489 	int cbehind, cahead;
1490 
1491 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1492 
1493 	object = m->object;
1494 	pindex = m->pindex;
1495 	cbehind = cahead = 0;
1496 
1497 	/*
1498 	 * if the requested page is not available, then give up now
1499 	 */
1500 	if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1501 		return 0;
1502 	}
1503 
1504 	if ((cbehind == 0) && (cahead == 0)) {
1505 		*reqpage = 0;
1506 		marray[0] = m;
1507 		return 1;
1508 	}
1509 
1510 	if (rahead > cahead) {
1511 		rahead = cahead;
1512 	}
1513 
1514 	if (rbehind > cbehind) {
1515 		rbehind = cbehind;
1516 	}
1517 
1518 	/*
1519 	 * scan backward for the read behind pages -- in memory
1520 	 */
1521 	if (pindex > 0) {
1522 		if (rbehind > pindex) {
1523 			rbehind = pindex;
1524 			startpindex = 0;
1525 		} else {
1526 			startpindex = pindex - rbehind;
1527 		}
1528 
1529 		if ((rtm = TAILQ_PREV(m, pglist, listq)) != NULL &&
1530 		    rtm->pindex >= startpindex)
1531 			startpindex = rtm->pindex + 1;
1532 
1533 		/* tpindex is unsigned; beware of numeric underflow. */
1534 		for (i = 0, tpindex = pindex - 1; tpindex >= startpindex &&
1535 		    tpindex < pindex; i++, tpindex--) {
1536 
1537 			rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1538 			    VM_ALLOC_IFNOTCACHED);
1539 			if (rtm == NULL) {
1540 				/*
1541 				 * Shift the allocated pages to the
1542 				 * beginning of the array.
1543 				 */
1544 				for (j = 0; j < i; j++) {
1545 					marray[j] = marray[j + tpindex + 1 -
1546 					    startpindex];
1547 				}
1548 				break;
1549 			}
1550 
1551 			marray[tpindex - startpindex] = rtm;
1552 		}
1553 	} else {
1554 		startpindex = 0;
1555 		i = 0;
1556 	}
1557 
1558 	marray[i] = m;
1559 	/* page offset of the required page */
1560 	*reqpage = i;
1561 
1562 	tpindex = pindex + 1;
1563 	i++;
1564 
1565 	/*
1566 	 * scan forward for the read ahead pages
1567 	 */
1568 	endpindex = tpindex + rahead;
1569 	if ((rtm = TAILQ_NEXT(m, listq)) != NULL && rtm->pindex < endpindex)
1570 		endpindex = rtm->pindex;
1571 	if (endpindex > object->size)
1572 		endpindex = object->size;
1573 
1574 	for (; tpindex < endpindex; i++, tpindex++) {
1575 
1576 		rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1577 		    VM_ALLOC_IFNOTCACHED);
1578 		if (rtm == NULL) {
1579 			break;
1580 		}
1581 
1582 		marray[i] = rtm;
1583 	}
1584 
1585 	/* return number of pages */
1586 	return i;
1587 }
1588 
1589 /*
1590  * Block entry into the machine-independent layer's page fault handler by
1591  * the calling thread.  Subsequent calls to vm_fault() by that thread will
1592  * return KERN_PROTECTION_FAILURE.  Enable machine-dependent handling of
1593  * spurious page faults.
1594  */
1595 int
1596 vm_fault_disable_pagefaults(void)
1597 {
1598 
1599 	return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR));
1600 }
1601 
1602 void
1603 vm_fault_enable_pagefaults(int save)
1604 {
1605 
1606 	curthread_pflags_restore(save);
1607 }
1608