xref: /freebsd/sys/vm/vm_fault.c (revision ec0e626bafb335b30c499d06066997f54b10c092)
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/proc.h>
85 #include <sys/resourcevar.h>
86 #include <sys/rwlock.h>
87 #include <sys/sysctl.h>
88 #include <sys/vmmeter.h>
89 #include <sys/vnode.h>
90 #ifdef KTRACE
91 #include <sys/ktrace.h>
92 #endif
93 
94 #include <vm/vm.h>
95 #include <vm/vm_param.h>
96 #include <vm/pmap.h>
97 #include <vm/vm_map.h>
98 #include <vm/vm_object.h>
99 #include <vm/vm_page.h>
100 #include <vm/vm_pageout.h>
101 #include <vm/vm_kern.h>
102 #include <vm/vm_pager.h>
103 #include <vm/vm_extern.h>
104 #include <vm/vm_reserv.h>
105 
106 #define PFBAK 4
107 #define PFFOR 4
108 
109 static int vm_fault_additional_pages(vm_page_t, int, int, vm_page_t *, int *);
110 
111 #define	VM_FAULT_READ_BEHIND	8
112 #define	VM_FAULT_READ_DEFAULT	(1 + VM_FAULT_READ_AHEAD_INIT)
113 #define	VM_FAULT_READ_MAX	(1 + VM_FAULT_READ_AHEAD_MAX)
114 #define	VM_FAULT_NINCR		(VM_FAULT_READ_MAX / VM_FAULT_READ_BEHIND)
115 #define	VM_FAULT_SUM		(VM_FAULT_NINCR * (VM_FAULT_NINCR + 1) / 2)
116 #define	VM_FAULT_CACHE_BEHIND	(VM_FAULT_READ_BEHIND * VM_FAULT_SUM)
117 
118 struct faultstate {
119 	vm_page_t m;
120 	vm_object_t object;
121 	vm_pindex_t pindex;
122 	vm_page_t first_m;
123 	vm_object_t	first_object;
124 	vm_pindex_t first_pindex;
125 	vm_map_t map;
126 	vm_map_entry_t entry;
127 	int lookup_still_valid;
128 	struct vnode *vp;
129 };
130 
131 static void vm_fault_cache_behind(const struct faultstate *fs, int distance);
132 static void vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
133 	    int faultcount, int reqpage);
134 
135 static inline void
136 release_page(struct faultstate *fs)
137 {
138 
139 	vm_page_xunbusy(fs->m);
140 	vm_page_lock(fs->m);
141 	vm_page_deactivate(fs->m);
142 	vm_page_unlock(fs->m);
143 	fs->m = NULL;
144 }
145 
146 static inline void
147 unlock_map(struct faultstate *fs)
148 {
149 
150 	if (fs->lookup_still_valid) {
151 		vm_map_lookup_done(fs->map, fs->entry);
152 		fs->lookup_still_valid = FALSE;
153 	}
154 }
155 
156 static void
157 unlock_and_deallocate(struct faultstate *fs)
158 {
159 
160 	vm_object_pip_wakeup(fs->object);
161 	VM_OBJECT_WUNLOCK(fs->object);
162 	if (fs->object != fs->first_object) {
163 		VM_OBJECT_WLOCK(fs->first_object);
164 		vm_page_lock(fs->first_m);
165 		vm_page_free(fs->first_m);
166 		vm_page_unlock(fs->first_m);
167 		vm_object_pip_wakeup(fs->first_object);
168 		VM_OBJECT_WUNLOCK(fs->first_object);
169 		fs->first_m = NULL;
170 	}
171 	vm_object_deallocate(fs->first_object);
172 	unlock_map(fs);
173 	if (fs->vp != NULL) {
174 		vput(fs->vp);
175 		fs->vp = NULL;
176 	}
177 }
178 
179 static void
180 vm_fault_dirty(vm_map_entry_t entry, vm_page_t m, vm_prot_t prot,
181     vm_prot_t fault_type, int fault_flags, boolean_t set_wd)
182 {
183 	boolean_t need_dirty;
184 
185 	if (((prot & VM_PROT_WRITE) == 0 &&
186 	    (fault_flags & VM_FAULT_DIRTY) == 0) ||
187 	    (m->oflags & VPO_UNMANAGED) != 0)
188 		return;
189 
190 	VM_OBJECT_ASSERT_LOCKED(m->object);
191 
192 	need_dirty = ((fault_type & VM_PROT_WRITE) != 0 &&
193 	    (fault_flags & VM_FAULT_CHANGE_WIRING) == 0) ||
194 	    (fault_flags & VM_FAULT_DIRTY) != 0;
195 
196 	if (set_wd)
197 		vm_object_set_writeable_dirty(m->object);
198 	else
199 		/*
200 		 * If two callers of vm_fault_dirty() with set_wd ==
201 		 * FALSE, one for the map entry with MAP_ENTRY_NOSYNC
202 		 * flag set, other with flag clear, race, it is
203 		 * possible for the no-NOSYNC thread to see m->dirty
204 		 * != 0 and not clear VPO_NOSYNC.  Take vm_page lock
205 		 * around manipulation of VPO_NOSYNC and
206 		 * vm_page_dirty() call, to avoid the race and keep
207 		 * m->oflags consistent.
208 		 */
209 		vm_page_lock(m);
210 
211 	/*
212 	 * If this is a NOSYNC mmap we do not want to set VPO_NOSYNC
213 	 * if the page is already dirty to prevent data written with
214 	 * the expectation of being synced from not being synced.
215 	 * Likewise if this entry does not request NOSYNC then make
216 	 * sure the page isn't marked NOSYNC.  Applications sharing
217 	 * data should use the same flags to avoid ping ponging.
218 	 */
219 	if ((entry->eflags & MAP_ENTRY_NOSYNC) != 0) {
220 		if (m->dirty == 0) {
221 			m->oflags |= VPO_NOSYNC;
222 		}
223 	} else {
224 		m->oflags &= ~VPO_NOSYNC;
225 	}
226 
227 	/*
228 	 * If the fault is a write, we know that this page is being
229 	 * written NOW so dirty it explicitly to save on
230 	 * pmap_is_modified() calls later.
231 	 *
232 	 * Also tell the backing pager, if any, that it should remove
233 	 * any swap backing since the page is now dirty.
234 	 */
235 	if (need_dirty)
236 		vm_page_dirty(m);
237 	if (!set_wd)
238 		vm_page_unlock(m);
239 	if (need_dirty)
240 		vm_pager_page_unswapped(m);
241 }
242 
243 /*
244  * TRYPAGER - used by vm_fault to calculate whether the pager for the
245  *	      current object *might* contain the page.
246  *
247  *	      default objects are zero-fill, there is no real pager.
248  */
249 #define TRYPAGER	(fs.object->type != OBJT_DEFAULT && \
250 			((fault_flags & VM_FAULT_CHANGE_WIRING) == 0 || wired))
251 
252 /*
253  *	vm_fault:
254  *
255  *	Handle a page fault occurring at the given address,
256  *	requiring the given permissions, in the map specified.
257  *	If successful, the page is inserted into the
258  *	associated physical map.
259  *
260  *	NOTE: the given address should be truncated to the
261  *	proper page address.
262  *
263  *	KERN_SUCCESS is returned if the page fault is handled; otherwise,
264  *	a standard error specifying why the fault is fatal is returned.
265  *
266  *	The map in question must be referenced, and remains so.
267  *	Caller may hold no locks.
268  */
269 int
270 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
271     int fault_flags)
272 {
273 	struct thread *td;
274 	int result;
275 
276 	td = curthread;
277 	if ((td->td_pflags & TDP_NOFAULTING) != 0)
278 		return (KERN_PROTECTION_FAILURE);
279 #ifdef KTRACE
280 	if (map != kernel_map && KTRPOINT(td, KTR_FAULT))
281 		ktrfault(vaddr, fault_type);
282 #endif
283 	result = vm_fault_hold(map, trunc_page(vaddr), fault_type, fault_flags,
284 	    NULL);
285 #ifdef KTRACE
286 	if (map != kernel_map && KTRPOINT(td, KTR_FAULTEND))
287 		ktrfaultend(result);
288 #endif
289 	return (result);
290 }
291 
292 int
293 vm_fault_hold(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
294     int fault_flags, vm_page_t *m_hold)
295 {
296 	vm_prot_t prot;
297 	int alloc_req, era, faultcount, nera, reqpage, result;
298 	boolean_t growstack, is_first_object_locked, wired;
299 	int map_generation;
300 	vm_object_t next_object;
301 	vm_page_t marray[VM_FAULT_READ_MAX];
302 	int hardfault;
303 	struct faultstate fs;
304 	struct vnode *vp;
305 	vm_page_t m;
306 	int ahead, behind, cluster_offset, error, locked;
307 
308 	hardfault = 0;
309 	growstack = TRUE;
310 	PCPU_INC(cnt.v_vm_faults);
311 	fs.vp = NULL;
312 	faultcount = reqpage = 0;
313 
314 RetryFault:;
315 
316 	/*
317 	 * Find the backing store object and offset into it to begin the
318 	 * search.
319 	 */
320 	fs.map = map;
321 	result = vm_map_lookup(&fs.map, vaddr, fault_type, &fs.entry,
322 	    &fs.first_object, &fs.first_pindex, &prot, &wired);
323 	if (result != KERN_SUCCESS) {
324 		if (growstack && result == KERN_INVALID_ADDRESS &&
325 		    map != kernel_map) {
326 			result = vm_map_growstack(curproc, vaddr);
327 			if (result != KERN_SUCCESS)
328 				return (KERN_FAILURE);
329 			growstack = FALSE;
330 			goto RetryFault;
331 		}
332 		return (result);
333 	}
334 
335 	map_generation = fs.map->timestamp;
336 
337 	if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
338 		panic("vm_fault: fault on nofault entry, addr: %lx",
339 		    (u_long)vaddr);
340 	}
341 
342 	if (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION &&
343 	    fs.entry->wiring_thread != curthread) {
344 		vm_map_unlock_read(fs.map);
345 		vm_map_lock(fs.map);
346 		if (vm_map_lookup_entry(fs.map, vaddr, &fs.entry) &&
347 		    (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION)) {
348 			fs.entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
349 			vm_map_unlock_and_wait(fs.map, 0);
350 		} else
351 			vm_map_unlock(fs.map);
352 		goto RetryFault;
353 	}
354 
355 	if (wired)
356 		fault_type = prot | (fault_type & VM_PROT_COPY);
357 
358 	if (fs.vp == NULL /* avoid locked vnode leak */ &&
359 	    (fault_flags & (VM_FAULT_CHANGE_WIRING | VM_FAULT_DIRTY)) == 0 &&
360 	    /* avoid calling vm_object_set_writeable_dirty() */
361 	    ((prot & VM_PROT_WRITE) == 0 ||
362 	    (fs.first_object->type != OBJT_VNODE &&
363 	    (fs.first_object->flags & OBJ_TMPFS_NODE) == 0) ||
364 	    (fs.first_object->flags & OBJ_MIGHTBEDIRTY) != 0)) {
365 		VM_OBJECT_RLOCK(fs.first_object);
366 		if ((prot & VM_PROT_WRITE) != 0 &&
367 		    (fs.first_object->type == OBJT_VNODE ||
368 		    (fs.first_object->flags & OBJ_TMPFS_NODE) != 0) &&
369 		    (fs.first_object->flags & OBJ_MIGHTBEDIRTY) == 0)
370 			goto fast_failed;
371 		m = vm_page_lookup(fs.first_object, fs.first_pindex);
372 		/* A busy page can be mapped for read|execute access. */
373 		if (m == NULL || ((prot & VM_PROT_WRITE) != 0 &&
374 		    vm_page_busied(m)) || m->valid != VM_PAGE_BITS_ALL)
375 			goto fast_failed;
376 		result = pmap_enter(fs.map->pmap, vaddr, m, prot,
377 		   fault_type | PMAP_ENTER_NOSLEEP | (wired ? PMAP_ENTER_WIRED :
378 		   0), 0);
379 		if (result != KERN_SUCCESS)
380 			goto fast_failed;
381 		if (m_hold != NULL) {
382 			*m_hold = m;
383 			vm_page_lock(m);
384 			vm_page_hold(m);
385 			vm_page_unlock(m);
386 		}
387 		vm_fault_dirty(fs.entry, m, prot, fault_type, fault_flags,
388 		    FALSE);
389 		VM_OBJECT_RUNLOCK(fs.first_object);
390 		if (!wired)
391 			vm_fault_prefault(&fs, vaddr, 0, 0);
392 		vm_map_lookup_done(fs.map, fs.entry);
393 		curthread->td_ru.ru_minflt++;
394 		return (KERN_SUCCESS);
395 fast_failed:
396 		if (!VM_OBJECT_TRYUPGRADE(fs.first_object)) {
397 			VM_OBJECT_RUNLOCK(fs.first_object);
398 			VM_OBJECT_WLOCK(fs.first_object);
399 		}
400 	} else {
401 		VM_OBJECT_WLOCK(fs.first_object);
402 	}
403 
404 	/*
405 	 * Make a reference to this object to prevent its disposal while we
406 	 * are messing with it.  Once we have the reference, the map is free
407 	 * to be diddled.  Since objects reference their shadows (and copies),
408 	 * they will stay around as well.
409 	 *
410 	 * Bump the paging-in-progress count to prevent size changes (e.g.
411 	 * truncation operations) during I/O.  This must be done after
412 	 * obtaining the vnode lock in order to avoid possible deadlocks.
413 	 */
414 	vm_object_reference_locked(fs.first_object);
415 	vm_object_pip_add(fs.first_object, 1);
416 
417 	fs.lookup_still_valid = TRUE;
418 
419 	fs.first_m = NULL;
420 
421 	/*
422 	 * Search for the page at object/offset.
423 	 */
424 	fs.object = fs.first_object;
425 	fs.pindex = fs.first_pindex;
426 	while (TRUE) {
427 		/*
428 		 * If the object is dead, we stop here
429 		 */
430 		if (fs.object->flags & OBJ_DEAD) {
431 			unlock_and_deallocate(&fs);
432 			return (KERN_PROTECTION_FAILURE);
433 		}
434 
435 		/*
436 		 * See if page is resident
437 		 */
438 		fs.m = vm_page_lookup(fs.object, fs.pindex);
439 		if (fs.m != NULL) {
440 			/*
441 			 * Wait/Retry if the page is busy.  We have to do this
442 			 * if the page is either exclusive or shared busy
443 			 * because the vm_pager may be using read busy for
444 			 * pageouts (and even pageins if it is the vnode
445 			 * pager), and we could end up trying to pagein and
446 			 * pageout the same page simultaneously.
447 			 *
448 			 * We can theoretically allow the busy case on a read
449 			 * fault if the page is marked valid, but since such
450 			 * pages are typically already pmap'd, putting that
451 			 * special case in might be more effort then it is
452 			 * worth.  We cannot under any circumstances mess
453 			 * around with a shared busied page except, perhaps,
454 			 * to pmap it.
455 			 */
456 			if (vm_page_busied(fs.m)) {
457 				/*
458 				 * Reference the page before unlocking and
459 				 * sleeping so that the page daemon is less
460 				 * likely to reclaim it.
461 				 */
462 				vm_page_aflag_set(fs.m, PGA_REFERENCED);
463 				if (fs.object != fs.first_object) {
464 					if (!VM_OBJECT_TRYWLOCK(
465 					    fs.first_object)) {
466 						VM_OBJECT_WUNLOCK(fs.object);
467 						VM_OBJECT_WLOCK(fs.first_object);
468 						VM_OBJECT_WLOCK(fs.object);
469 					}
470 					vm_page_lock(fs.first_m);
471 					vm_page_free(fs.first_m);
472 					vm_page_unlock(fs.first_m);
473 					vm_object_pip_wakeup(fs.first_object);
474 					VM_OBJECT_WUNLOCK(fs.first_object);
475 					fs.first_m = NULL;
476 				}
477 				unlock_map(&fs);
478 				if (fs.m == vm_page_lookup(fs.object,
479 				    fs.pindex)) {
480 					vm_page_sleep_if_busy(fs.m, "vmpfw");
481 				}
482 				vm_object_pip_wakeup(fs.object);
483 				VM_OBJECT_WUNLOCK(fs.object);
484 				PCPU_INC(cnt.v_intrans);
485 				vm_object_deallocate(fs.first_object);
486 				goto RetryFault;
487 			}
488 			vm_page_lock(fs.m);
489 			vm_page_remque(fs.m);
490 			vm_page_unlock(fs.m);
491 
492 			/*
493 			 * Mark page busy for other processes, and the
494 			 * pagedaemon.  If it still isn't completely valid
495 			 * (readable), jump to readrest, else break-out ( we
496 			 * found the page ).
497 			 */
498 			vm_page_xbusy(fs.m);
499 			if (fs.m->valid != VM_PAGE_BITS_ALL)
500 				goto readrest;
501 			break;
502 		}
503 
504 		/*
505 		 * Page is not resident, If this is the search termination
506 		 * or the pager might contain the page, allocate a new page.
507 		 */
508 		if (TRYPAGER || fs.object == fs.first_object) {
509 			if (fs.pindex >= fs.object->size) {
510 				unlock_and_deallocate(&fs);
511 				return (KERN_PROTECTION_FAILURE);
512 			}
513 
514 			/*
515 			 * Allocate a new page for this object/offset pair.
516 			 *
517 			 * Unlocked read of the p_flag is harmless. At
518 			 * worst, the P_KILLED might be not observed
519 			 * there, and allocation can fail, causing
520 			 * restart and new reading of the p_flag.
521 			 */
522 			fs.m = NULL;
523 			if (!vm_page_count_severe() || P_KILLED(curproc)) {
524 #if VM_NRESERVLEVEL > 0
525 				if ((fs.object->flags & OBJ_COLORED) == 0) {
526 					fs.object->flags |= OBJ_COLORED;
527 					fs.object->pg_color = atop(vaddr) -
528 					    fs.pindex;
529 				}
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_cache_behind(&fs,
573 					    VM_FAULT_READ_MAX);
574 			} else if (fs.pindex == fs.entry->next_read) {
575 				/*
576 				 * This is a sequential fault.  Arithmetically
577 				 * increase the requested number of pages in
578 				 * the read-ahead window.  The requested
579 				 * number of pages is "# of sequential faults
580 				 * x (read ahead min + 1) + read ahead min"
581 				 */
582 				behind = 0;
583 				nera = VM_FAULT_READ_AHEAD_MIN;
584 				if (era > 0) {
585 					nera += era + 1;
586 					if (nera > VM_FAULT_READ_AHEAD_MAX)
587 						nera = VM_FAULT_READ_AHEAD_MAX;
588 				}
589 				ahead = nera;
590 				if (era == VM_FAULT_READ_AHEAD_MAX)
591 					vm_fault_cache_behind(&fs,
592 					    VM_FAULT_CACHE_BEHIND);
593 			} else {
594 				/*
595 				 * This is a non-sequential fault.  Request a
596 				 * cluster of pages that is aligned to a
597 				 * VM_FAULT_READ_DEFAULT page offset boundary
598 				 * within the object.  Alignment to a page
599 				 * offset boundary is more likely to coincide
600 				 * with the underlying file system block than
601 				 * alignment to a virtual address boundary.
602 				 */
603 				cluster_offset = fs.pindex %
604 				    VM_FAULT_READ_DEFAULT;
605 				behind = ulmin(cluster_offset,
606 				    atop(vaddr - fs.entry->start));
607 				nera = 0;
608 				ahead = VM_FAULT_READ_DEFAULT - 1 -
609 				    cluster_offset;
610 			}
611 			ahead = ulmin(ahead, atop(fs.entry->end - vaddr) - 1);
612 			if (era != nera)
613 				fs.entry->read_ahead = nera;
614 
615 			/*
616 			 * Call the pager to retrieve the data, if any, after
617 			 * releasing the lock on the map.  We hold a ref on
618 			 * fs.object and the pages are exclusive busied.
619 			 */
620 			unlock_map(&fs);
621 
622 			if (fs.object->type == OBJT_VNODE) {
623 				vp = fs.object->handle;
624 				if (vp == fs.vp)
625 					goto vnode_locked;
626 				else if (fs.vp != NULL) {
627 					vput(fs.vp);
628 					fs.vp = NULL;
629 				}
630 				locked = VOP_ISLOCKED(vp);
631 
632 				if (locked != LK_EXCLUSIVE)
633 					locked = LK_SHARED;
634 				/* Do not sleep for vnode lock while fs.m is busy */
635 				error = vget(vp, locked | LK_CANRECURSE |
636 				    LK_NOWAIT, curthread);
637 				if (error != 0) {
638 					vhold(vp);
639 					release_page(&fs);
640 					unlock_and_deallocate(&fs);
641 					error = vget(vp, locked | LK_RETRY |
642 					    LK_CANRECURSE, curthread);
643 					vdrop(vp);
644 					fs.vp = vp;
645 					KASSERT(error == 0,
646 					    ("vm_fault: vget failed"));
647 					goto RetryFault;
648 				}
649 				fs.vp = vp;
650 			}
651 vnode_locked:
652 			KASSERT(fs.vp == NULL || !fs.map->system_map,
653 			    ("vm_fault: vnode-backed object mapped by system map"));
654 
655 			/*
656 			 * now we find out if any other pages should be paged
657 			 * in at this time this routine checks to see if the
658 			 * pages surrounding this fault reside in the same
659 			 * object as the page for this fault.  If they do,
660 			 * then they are faulted in also into the object.  The
661 			 * array "marray" returned contains an array of
662 			 * vm_page_t structs where one of them is the
663 			 * vm_page_t passed to the routine.  The reqpage
664 			 * return value is the index into the marray for the
665 			 * vm_page_t passed to the routine.
666 			 *
667 			 * fs.m plus the additional pages are exclusive busied.
668 			 */
669 			faultcount = vm_fault_additional_pages(
670 			    fs.m, behind, ahead, marray, &reqpage);
671 
672 			rv = faultcount ?
673 			    vm_pager_get_pages(fs.object, marray, faultcount,
674 				reqpage) : VM_PAGER_FAIL;
675 
676 			if (rv == VM_PAGER_OK) {
677 				/*
678 				 * Found the page. Leave it busy while we play
679 				 * with it.
680 				 */
681 
682 				/*
683 				 * Relookup in case pager changed page. Pager
684 				 * is responsible for disposition of old page
685 				 * if moved.
686 				 */
687 				fs.m = vm_page_lookup(fs.object, fs.pindex);
688 				if (!fs.m) {
689 					unlock_and_deallocate(&fs);
690 					goto RetryFault;
691 				}
692 
693 				hardfault++;
694 				break; /* break to PAGE HAS BEEN FOUND */
695 			}
696 			/*
697 			 * Remove the bogus page (which does not exist at this
698 			 * object/offset); before doing so, we must get back
699 			 * our object lock to preserve our invariant.
700 			 *
701 			 * Also wake up any other process that may want to bring
702 			 * in this page.
703 			 *
704 			 * If this is the top-level object, we must leave the
705 			 * busy page to prevent another process from rushing
706 			 * past us, and inserting the page in that object at
707 			 * the same time that we are.
708 			 */
709 			if (rv == VM_PAGER_ERROR)
710 				printf("vm_fault: pager read error, pid %d (%s)\n",
711 				    curproc->p_pid, curproc->p_comm);
712 			/*
713 			 * Data outside the range of the pager or an I/O error
714 			 */
715 			/*
716 			 * XXX - the check for kernel_map is a kludge to work
717 			 * around having the machine panic on a kernel space
718 			 * fault w/ I/O error.
719 			 */
720 			if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) ||
721 				(rv == VM_PAGER_BAD)) {
722 				vm_page_lock(fs.m);
723 				vm_page_free(fs.m);
724 				vm_page_unlock(fs.m);
725 				fs.m = NULL;
726 				unlock_and_deallocate(&fs);
727 				return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
728 			}
729 			if (fs.object != fs.first_object) {
730 				vm_page_lock(fs.m);
731 				vm_page_free(fs.m);
732 				vm_page_unlock(fs.m);
733 				fs.m = NULL;
734 				/*
735 				 * XXX - we cannot just fall out at this
736 				 * point, m has been freed and is invalid!
737 				 */
738 			}
739 		}
740 
741 		/*
742 		 * We get here if the object has default pager (or unwiring)
743 		 * or the pager doesn't have the page.
744 		 */
745 		if (fs.object == fs.first_object)
746 			fs.first_m = fs.m;
747 
748 		/*
749 		 * Move on to the next object.  Lock the next object before
750 		 * unlocking the current one.
751 		 */
752 		fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset);
753 		next_object = fs.object->backing_object;
754 		if (next_object == NULL) {
755 			/*
756 			 * If there's no object left, fill the page in the top
757 			 * object with zeros.
758 			 */
759 			if (fs.object != fs.first_object) {
760 				vm_object_pip_wakeup(fs.object);
761 				VM_OBJECT_WUNLOCK(fs.object);
762 
763 				fs.object = fs.first_object;
764 				fs.pindex = fs.first_pindex;
765 				fs.m = fs.first_m;
766 				VM_OBJECT_WLOCK(fs.object);
767 			}
768 			fs.first_m = NULL;
769 
770 			/*
771 			 * Zero the page if necessary and mark it valid.
772 			 */
773 			if ((fs.m->flags & PG_ZERO) == 0) {
774 				pmap_zero_page(fs.m);
775 			} else {
776 				PCPU_INC(cnt.v_ozfod);
777 			}
778 			PCPU_INC(cnt.v_zfod);
779 			fs.m->valid = VM_PAGE_BITS_ALL;
780 			/* Don't try to prefault neighboring pages. */
781 			faultcount = 1;
782 			break;	/* break to PAGE HAS BEEN FOUND */
783 		} else {
784 			KASSERT(fs.object != next_object,
785 			    ("object loop %p", next_object));
786 			VM_OBJECT_WLOCK(next_object);
787 			vm_object_pip_add(next_object, 1);
788 			if (fs.object != fs.first_object)
789 				vm_object_pip_wakeup(fs.object);
790 			VM_OBJECT_WUNLOCK(fs.object);
791 			fs.object = next_object;
792 		}
793 	}
794 
795 	vm_page_assert_xbusied(fs.m);
796 
797 	/*
798 	 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
799 	 * is held.]
800 	 */
801 
802 	/*
803 	 * If the page is being written, but isn't already owned by the
804 	 * top-level object, we have to copy it into a new page owned by the
805 	 * top-level object.
806 	 */
807 	if (fs.object != fs.first_object) {
808 		/*
809 		 * We only really need to copy if we want to write it.
810 		 */
811 		if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
812 			/*
813 			 * This allows pages to be virtually copied from a
814 			 * backing_object into the first_object, where the
815 			 * backing object has no other refs to it, and cannot
816 			 * gain any more refs.  Instead of a bcopy, we just
817 			 * move the page from the backing object to the
818 			 * first object.  Note that we must mark the page
819 			 * dirty in the first object so that it will go out
820 			 * to swap when needed.
821 			 */
822 			is_first_object_locked = FALSE;
823 			if (
824 				/*
825 				 * Only one shadow object
826 				 */
827 				(fs.object->shadow_count == 1) &&
828 				/*
829 				 * No COW refs, except us
830 				 */
831 				(fs.object->ref_count == 1) &&
832 				/*
833 				 * No one else can look this object up
834 				 */
835 				(fs.object->handle == NULL) &&
836 				/*
837 				 * No other ways to look the object up
838 				 */
839 				((fs.object->type == OBJT_DEFAULT) ||
840 				 (fs.object->type == OBJT_SWAP)) &&
841 			    (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs.first_object)) &&
842 				/*
843 				 * We don't chase down the shadow chain
844 				 */
845 			    fs.object == fs.first_object->backing_object) {
846 				/*
847 				 * get rid of the unnecessary page
848 				 */
849 				vm_page_lock(fs.first_m);
850 				vm_page_free(fs.first_m);
851 				vm_page_unlock(fs.first_m);
852 				/*
853 				 * grab the page and put it into the
854 				 * process'es object.  The page is
855 				 * automatically made dirty.
856 				 */
857 				if (vm_page_rename(fs.m, fs.first_object,
858 				    fs.first_pindex)) {
859 					unlock_and_deallocate(&fs);
860 					goto RetryFault;
861 				}
862 #if VM_NRESERVLEVEL > 0
863 				/*
864 				 * Rename the reservation.
865 				 */
866 				vm_reserv_rename(fs.m, fs.first_object,
867 				    fs.object, OFF_TO_IDX(
868 				    fs.first_object->backing_object_offset));
869 #endif
870 				vm_page_xbusy(fs.m);
871 				fs.first_m = fs.m;
872 				fs.m = NULL;
873 				PCPU_INC(cnt.v_cow_optim);
874 			} else {
875 				/*
876 				 * Oh, well, lets copy it.
877 				 */
878 				pmap_copy_page(fs.m, fs.first_m);
879 				fs.first_m->valid = VM_PAGE_BITS_ALL;
880 				if (wired && (fault_flags &
881 				    VM_FAULT_CHANGE_WIRING) == 0) {
882 					vm_page_lock(fs.first_m);
883 					vm_page_wire(fs.first_m);
884 					vm_page_unlock(fs.first_m);
885 
886 					vm_page_lock(fs.m);
887 					vm_page_unwire(fs.m, PQ_INACTIVE);
888 					vm_page_unlock(fs.m);
889 				}
890 				/*
891 				 * We no longer need the old page or object.
892 				 */
893 				release_page(&fs);
894 			}
895 			/*
896 			 * fs.object != fs.first_object due to above
897 			 * conditional
898 			 */
899 			vm_object_pip_wakeup(fs.object);
900 			VM_OBJECT_WUNLOCK(fs.object);
901 			/*
902 			 * Only use the new page below...
903 			 */
904 			fs.object = fs.first_object;
905 			fs.pindex = fs.first_pindex;
906 			fs.m = fs.first_m;
907 			if (!is_first_object_locked)
908 				VM_OBJECT_WLOCK(fs.object);
909 			PCPU_INC(cnt.v_cow_faults);
910 			curthread->td_cow++;
911 		} else {
912 			prot &= ~VM_PROT_WRITE;
913 		}
914 	}
915 
916 	/*
917 	 * We must verify that the maps have not changed since our last
918 	 * lookup.
919 	 */
920 	if (!fs.lookup_still_valid) {
921 		vm_object_t retry_object;
922 		vm_pindex_t retry_pindex;
923 		vm_prot_t retry_prot;
924 
925 		if (!vm_map_trylock_read(fs.map)) {
926 			release_page(&fs);
927 			unlock_and_deallocate(&fs);
928 			goto RetryFault;
929 		}
930 		fs.lookup_still_valid = TRUE;
931 		if (fs.map->timestamp != map_generation) {
932 			result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
933 			    &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
934 
935 			/*
936 			 * If we don't need the page any longer, put it on the inactive
937 			 * list (the easiest thing to do here).  If no one needs it,
938 			 * pageout will grab it eventually.
939 			 */
940 			if (result != KERN_SUCCESS) {
941 				release_page(&fs);
942 				unlock_and_deallocate(&fs);
943 
944 				/*
945 				 * If retry of map lookup would have blocked then
946 				 * retry fault from start.
947 				 */
948 				if (result == KERN_FAILURE)
949 					goto RetryFault;
950 				return (result);
951 			}
952 			if ((retry_object != fs.first_object) ||
953 			    (retry_pindex != fs.first_pindex)) {
954 				release_page(&fs);
955 				unlock_and_deallocate(&fs);
956 				goto RetryFault;
957 			}
958 
959 			/*
960 			 * Check whether the protection has changed or the object has
961 			 * been copied while we left the map unlocked. Changing from
962 			 * read to write permission is OK - we leave the page
963 			 * write-protected, and catch the write fault. Changing from
964 			 * write to read permission means that we can't mark the page
965 			 * write-enabled after all.
966 			 */
967 			prot &= retry_prot;
968 		}
969 	}
970 	/*
971 	 * If the page was filled by a pager, update the map entry's
972 	 * last read offset.  Since the pager does not return the
973 	 * actual set of pages that it read, this update is based on
974 	 * the requested set.  Typically, the requested and actual
975 	 * sets are the same.
976 	 *
977 	 * XXX The following assignment modifies the map
978 	 * without holding a write lock on it.
979 	 */
980 	if (hardfault)
981 		fs.entry->next_read = fs.pindex + faultcount - reqpage;
982 
983 	vm_fault_dirty(fs.entry, fs.m, prot, fault_type, fault_flags, TRUE);
984 	vm_page_assert_xbusied(fs.m);
985 
986 	/*
987 	 * Page must be completely valid or it is not fit to
988 	 * map into user space.  vm_pager_get_pages() ensures this.
989 	 */
990 	KASSERT(fs.m->valid == VM_PAGE_BITS_ALL,
991 	    ("vm_fault: page %p partially invalid", fs.m));
992 	VM_OBJECT_WUNLOCK(fs.object);
993 
994 	/*
995 	 * Put this page into the physical map.  We had to do the unlock above
996 	 * because pmap_enter() may sleep.  We don't put the page
997 	 * back on the active queue until later so that the pageout daemon
998 	 * won't find it (yet).
999 	 */
1000 	pmap_enter(fs.map->pmap, vaddr, fs.m, prot,
1001 	    fault_type | (wired ? PMAP_ENTER_WIRED : 0), 0);
1002 	if (faultcount != 1 && (fault_flags & VM_FAULT_CHANGE_WIRING) == 0 &&
1003 	    wired == 0)
1004 		vm_fault_prefault(&fs, vaddr, faultcount, reqpage);
1005 	VM_OBJECT_WLOCK(fs.object);
1006 	vm_page_lock(fs.m);
1007 
1008 	/*
1009 	 * If the page is not wired down, then put it where the pageout daemon
1010 	 * can find it.
1011 	 */
1012 	if (fault_flags & VM_FAULT_CHANGE_WIRING) {
1013 		if (wired)
1014 			vm_page_wire(fs.m);
1015 		else
1016 			vm_page_unwire(fs.m, PQ_ACTIVE);
1017 	} else
1018 		vm_page_activate(fs.m);
1019 	if (m_hold != NULL) {
1020 		*m_hold = fs.m;
1021 		vm_page_hold(fs.m);
1022 	}
1023 	vm_page_unlock(fs.m);
1024 	vm_page_xunbusy(fs.m);
1025 
1026 	/*
1027 	 * Unlock everything, and return
1028 	 */
1029 	unlock_and_deallocate(&fs);
1030 	if (hardfault) {
1031 		PCPU_INC(cnt.v_io_faults);
1032 		curthread->td_ru.ru_majflt++;
1033 	} else
1034 		curthread->td_ru.ru_minflt++;
1035 
1036 	return (KERN_SUCCESS);
1037 }
1038 
1039 /*
1040  * Speed up the reclamation of up to "distance" pages that precede the
1041  * faulting pindex within the first object of the shadow chain.
1042  */
1043 static void
1044 vm_fault_cache_behind(const struct faultstate *fs, int distance)
1045 {
1046 	vm_object_t first_object, object;
1047 	vm_page_t m, m_prev;
1048 	vm_pindex_t pindex;
1049 
1050 	object = fs->object;
1051 	VM_OBJECT_ASSERT_WLOCKED(object);
1052 	first_object = fs->first_object;
1053 	if (first_object != object) {
1054 		if (!VM_OBJECT_TRYWLOCK(first_object)) {
1055 			VM_OBJECT_WUNLOCK(object);
1056 			VM_OBJECT_WLOCK(first_object);
1057 			VM_OBJECT_WLOCK(object);
1058 		}
1059 	}
1060 	/* Neither fictitious nor unmanaged pages can be cached. */
1061 	if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) {
1062 		if (fs->first_pindex < distance)
1063 			pindex = 0;
1064 		else
1065 			pindex = fs->first_pindex - distance;
1066 		if (pindex < OFF_TO_IDX(fs->entry->offset))
1067 			pindex = OFF_TO_IDX(fs->entry->offset);
1068 		m = first_object != object ? fs->first_m : fs->m;
1069 		vm_page_assert_xbusied(m);
1070 		m_prev = vm_page_prev(m);
1071 		while ((m = m_prev) != NULL && m->pindex >= pindex &&
1072 		    m->valid == VM_PAGE_BITS_ALL) {
1073 			m_prev = vm_page_prev(m);
1074 			if (vm_page_busied(m))
1075 				continue;
1076 			vm_page_lock(m);
1077 			if (m->hold_count == 0 && m->wire_count == 0) {
1078 				pmap_remove_all(m);
1079 				vm_page_aflag_clear(m, PGA_REFERENCED);
1080 				if (m->dirty != 0)
1081 					vm_page_deactivate(m);
1082 				else
1083 					vm_page_cache(m);
1084 			}
1085 			vm_page_unlock(m);
1086 		}
1087 	}
1088 	if (first_object != object)
1089 		VM_OBJECT_WUNLOCK(first_object);
1090 }
1091 
1092 /*
1093  * vm_fault_prefault provides a quick way of clustering
1094  * pagefaults into a processes address space.  It is a "cousin"
1095  * of vm_map_pmap_enter, except it runs at page fault time instead
1096  * of mmap time.
1097  */
1098 static void
1099 vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
1100     int faultcount, int reqpage)
1101 {
1102 	pmap_t pmap;
1103 	vm_map_entry_t entry;
1104 	vm_object_t backing_object, lobject;
1105 	vm_offset_t addr, starta;
1106 	vm_pindex_t pindex;
1107 	vm_page_t m;
1108 	int backward, forward, i;
1109 
1110 	pmap = fs->map->pmap;
1111 	if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
1112 		return;
1113 
1114 	if (faultcount > 0) {
1115 		backward = reqpage;
1116 		forward = faultcount - reqpage - 1;
1117 	} else {
1118 		backward = PFBAK;
1119 		forward = PFFOR;
1120 	}
1121 	entry = fs->entry;
1122 
1123 	starta = addra - backward * PAGE_SIZE;
1124 	if (starta < entry->start) {
1125 		starta = entry->start;
1126 	} else if (starta > addra) {
1127 		starta = 0;
1128 	}
1129 
1130 	/*
1131 	 * Generate the sequence of virtual addresses that are candidates for
1132 	 * prefaulting in an outward spiral from the faulting virtual address,
1133 	 * "addra".  Specifically, the sequence is "addra - PAGE_SIZE", "addra
1134 	 * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ...
1135 	 * If the candidate address doesn't have a backing physical page, then
1136 	 * the loop immediately terminates.
1137 	 */
1138 	for (i = 0; i < 2 * imax(backward, forward); i++) {
1139 		addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE :
1140 		    PAGE_SIZE);
1141 		if (addr > addra + forward * PAGE_SIZE)
1142 			addr = 0;
1143 
1144 		if (addr < starta || addr >= entry->end)
1145 			continue;
1146 
1147 		if (!pmap_is_prefaultable(pmap, addr))
1148 			continue;
1149 
1150 		pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1151 		lobject = entry->object.vm_object;
1152 		VM_OBJECT_RLOCK(lobject);
1153 		while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1154 		    lobject->type == OBJT_DEFAULT &&
1155 		    (backing_object = lobject->backing_object) != NULL) {
1156 			KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1157 			    0, ("vm_fault_prefault: unaligned object offset"));
1158 			pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1159 			VM_OBJECT_RLOCK(backing_object);
1160 			VM_OBJECT_RUNLOCK(lobject);
1161 			lobject = backing_object;
1162 		}
1163 		if (m == NULL) {
1164 			VM_OBJECT_RUNLOCK(lobject);
1165 			break;
1166 		}
1167 		if (m->valid == VM_PAGE_BITS_ALL &&
1168 		    (m->flags & PG_FICTITIOUS) == 0)
1169 			pmap_enter_quick(pmap, addr, m, entry->protection);
1170 		VM_OBJECT_RUNLOCK(lobject);
1171 	}
1172 }
1173 
1174 /*
1175  * Hold each of the physical pages that are mapped by the specified range of
1176  * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
1177  * and allow the specified types of access, "prot".  If all of the implied
1178  * pages are successfully held, then the number of held pages is returned
1179  * together with pointers to those pages in the array "ma".  However, if any
1180  * of the pages cannot be held, -1 is returned.
1181  */
1182 int
1183 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
1184     vm_prot_t prot, vm_page_t *ma, int max_count)
1185 {
1186 	vm_offset_t end, va;
1187 	vm_page_t *mp;
1188 	int count;
1189 	boolean_t pmap_failed;
1190 
1191 	if (len == 0)
1192 		return (0);
1193 	end = round_page(addr + len);
1194 	addr = trunc_page(addr);
1195 
1196 	/*
1197 	 * Check for illegal addresses.
1198 	 */
1199 	if (addr < vm_map_min(map) || addr > end || end > vm_map_max(map))
1200 		return (-1);
1201 
1202 	if (atop(end - addr) > max_count)
1203 		panic("vm_fault_quick_hold_pages: count > max_count");
1204 	count = atop(end - addr);
1205 
1206 	/*
1207 	 * Most likely, the physical pages are resident in the pmap, so it is
1208 	 * faster to try pmap_extract_and_hold() first.
1209 	 */
1210 	pmap_failed = FALSE;
1211 	for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) {
1212 		*mp = pmap_extract_and_hold(map->pmap, va, prot);
1213 		if (*mp == NULL)
1214 			pmap_failed = TRUE;
1215 		else if ((prot & VM_PROT_WRITE) != 0 &&
1216 		    (*mp)->dirty != VM_PAGE_BITS_ALL) {
1217 			/*
1218 			 * Explicitly dirty the physical page.  Otherwise, the
1219 			 * caller's changes may go unnoticed because they are
1220 			 * performed through an unmanaged mapping or by a DMA
1221 			 * operation.
1222 			 *
1223 			 * The object lock is not held here.
1224 			 * See vm_page_clear_dirty_mask().
1225 			 */
1226 			vm_page_dirty(*mp);
1227 		}
1228 	}
1229 	if (pmap_failed) {
1230 		/*
1231 		 * One or more pages could not be held by the pmap.  Either no
1232 		 * page was mapped at the specified virtual address or that
1233 		 * mapping had insufficient permissions.  Attempt to fault in
1234 		 * and hold these pages.
1235 		 */
1236 		for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE)
1237 			if (*mp == NULL && vm_fault_hold(map, va, prot,
1238 			    VM_FAULT_NORMAL, mp) != KERN_SUCCESS)
1239 				goto error;
1240 	}
1241 	return (count);
1242 error:
1243 	for (mp = ma; mp < ma + count; mp++)
1244 		if (*mp != NULL) {
1245 			vm_page_lock(*mp);
1246 			vm_page_unhold(*mp);
1247 			vm_page_unlock(*mp);
1248 		}
1249 	return (-1);
1250 }
1251 
1252 /*
1253  *	Routine:
1254  *		vm_fault_copy_entry
1255  *	Function:
1256  *		Create new shadow object backing dst_entry with private copy of
1257  *		all underlying pages. When src_entry is equal to dst_entry,
1258  *		function implements COW for wired-down map entry. Otherwise,
1259  *		it forks wired entry into dst_map.
1260  *
1261  *	In/out conditions:
1262  *		The source and destination maps must be locked for write.
1263  *		The source map entry must be wired down (or be a sharing map
1264  *		entry corresponding to a main map entry that is wired down).
1265  */
1266 void
1267 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1268     vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1269     vm_ooffset_t *fork_charge)
1270 {
1271 	vm_object_t backing_object, dst_object, object, src_object;
1272 	vm_pindex_t dst_pindex, pindex, src_pindex;
1273 	vm_prot_t access, prot;
1274 	vm_offset_t vaddr;
1275 	vm_page_t dst_m;
1276 	vm_page_t src_m;
1277 	boolean_t upgrade;
1278 
1279 #ifdef	lint
1280 	src_map++;
1281 #endif	/* lint */
1282 
1283 	upgrade = src_entry == dst_entry;
1284 	access = prot = dst_entry->protection;
1285 
1286 	src_object = src_entry->object.vm_object;
1287 	src_pindex = OFF_TO_IDX(src_entry->offset);
1288 
1289 	if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
1290 		dst_object = src_object;
1291 		vm_object_reference(dst_object);
1292 	} else {
1293 		/*
1294 		 * Create the top-level object for the destination entry. (Doesn't
1295 		 * actually shadow anything - we copy the pages directly.)
1296 		 */
1297 		dst_object = vm_object_allocate(OBJT_DEFAULT,
1298 		    OFF_TO_IDX(dst_entry->end - dst_entry->start));
1299 #if VM_NRESERVLEVEL > 0
1300 		dst_object->flags |= OBJ_COLORED;
1301 		dst_object->pg_color = atop(dst_entry->start);
1302 #endif
1303 	}
1304 
1305 	VM_OBJECT_WLOCK(dst_object);
1306 	KASSERT(upgrade || dst_entry->object.vm_object == NULL,
1307 	    ("vm_fault_copy_entry: vm_object not NULL"));
1308 	if (src_object != dst_object) {
1309 		dst_entry->object.vm_object = dst_object;
1310 		dst_entry->offset = 0;
1311 		dst_object->charge = dst_entry->end - dst_entry->start;
1312 	}
1313 	if (fork_charge != NULL) {
1314 		KASSERT(dst_entry->cred == NULL,
1315 		    ("vm_fault_copy_entry: leaked swp charge"));
1316 		dst_object->cred = curthread->td_ucred;
1317 		crhold(dst_object->cred);
1318 		*fork_charge += dst_object->charge;
1319 	} else if (dst_object->cred == NULL) {
1320 		KASSERT(dst_entry->cred != NULL, ("no cred for entry %p",
1321 		    dst_entry));
1322 		dst_object->cred = dst_entry->cred;
1323 		dst_entry->cred = NULL;
1324 	}
1325 
1326 	/*
1327 	 * If not an upgrade, then enter the mappings in the pmap as
1328 	 * read and/or execute accesses.  Otherwise, enter them as
1329 	 * write accesses.
1330 	 *
1331 	 * A writeable large page mapping is only created if all of
1332 	 * the constituent small page mappings are modified. Marking
1333 	 * PTEs as modified on inception allows promotion to happen
1334 	 * without taking potentially large number of soft faults.
1335 	 */
1336 	if (!upgrade)
1337 		access &= ~VM_PROT_WRITE;
1338 
1339 	/*
1340 	 * Loop through all of the virtual pages within the entry's
1341 	 * range, copying each page from the source object to the
1342 	 * destination object.  Since the source is wired, those pages
1343 	 * must exist.  In contrast, the destination is pageable.
1344 	 * Since the destination object does share any backing storage
1345 	 * with the source object, all of its pages must be dirtied,
1346 	 * regardless of whether they can be written.
1347 	 */
1348 	for (vaddr = dst_entry->start, dst_pindex = 0;
1349 	    vaddr < dst_entry->end;
1350 	    vaddr += PAGE_SIZE, dst_pindex++) {
1351 again:
1352 		/*
1353 		 * Find the page in the source object, and copy it in.
1354 		 * Because the source is wired down, the page will be
1355 		 * in memory.
1356 		 */
1357 		if (src_object != dst_object)
1358 			VM_OBJECT_RLOCK(src_object);
1359 		object = src_object;
1360 		pindex = src_pindex + dst_pindex;
1361 		while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
1362 		    (backing_object = object->backing_object) != NULL) {
1363 			/*
1364 			 * Unless the source mapping is read-only or
1365 			 * it is presently being upgraded from
1366 			 * read-only, the first object in the shadow
1367 			 * chain should provide all of the pages.  In
1368 			 * other words, this loop body should never be
1369 			 * executed when the source mapping is already
1370 			 * read/write.
1371 			 */
1372 			KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 ||
1373 			    upgrade,
1374 			    ("vm_fault_copy_entry: main object missing page"));
1375 
1376 			VM_OBJECT_RLOCK(backing_object);
1377 			pindex += OFF_TO_IDX(object->backing_object_offset);
1378 			if (object != dst_object)
1379 				VM_OBJECT_RUNLOCK(object);
1380 			object = backing_object;
1381 		}
1382 		KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing"));
1383 
1384 		if (object != dst_object) {
1385 			/*
1386 			 * Allocate a page in the destination object.
1387 			 */
1388 			dst_m = vm_page_alloc(dst_object, (src_object ==
1389 			    dst_object ? src_pindex : 0) + dst_pindex,
1390 			    VM_ALLOC_NORMAL);
1391 			if (dst_m == NULL) {
1392 				VM_OBJECT_WUNLOCK(dst_object);
1393 				VM_OBJECT_RUNLOCK(object);
1394 				VM_WAIT;
1395 				VM_OBJECT_WLOCK(dst_object);
1396 				goto again;
1397 			}
1398 			pmap_copy_page(src_m, dst_m);
1399 			VM_OBJECT_RUNLOCK(object);
1400 			dst_m->valid = VM_PAGE_BITS_ALL;
1401 			dst_m->dirty = VM_PAGE_BITS_ALL;
1402 		} else {
1403 			dst_m = src_m;
1404 			if (vm_page_sleep_if_busy(dst_m, "fltupg"))
1405 				goto again;
1406 			vm_page_xbusy(dst_m);
1407 			KASSERT(dst_m->valid == VM_PAGE_BITS_ALL,
1408 			    ("invalid dst page %p", dst_m));
1409 		}
1410 		VM_OBJECT_WUNLOCK(dst_object);
1411 
1412 		/*
1413 		 * Enter it in the pmap. If a wired, copy-on-write
1414 		 * mapping is being replaced by a write-enabled
1415 		 * mapping, then wire that new mapping.
1416 		 */
1417 		pmap_enter(dst_map->pmap, vaddr, dst_m, prot,
1418 		    access | (upgrade ? PMAP_ENTER_WIRED : 0), 0);
1419 
1420 		/*
1421 		 * Mark it no longer busy, and put it on the active list.
1422 		 */
1423 		VM_OBJECT_WLOCK(dst_object);
1424 
1425 		if (upgrade) {
1426 			if (src_m != dst_m) {
1427 				vm_page_lock(src_m);
1428 				vm_page_unwire(src_m, PQ_INACTIVE);
1429 				vm_page_unlock(src_m);
1430 				vm_page_lock(dst_m);
1431 				vm_page_wire(dst_m);
1432 				vm_page_unlock(dst_m);
1433 			} else {
1434 				KASSERT(dst_m->wire_count > 0,
1435 				    ("dst_m %p is not wired", dst_m));
1436 			}
1437 		} else {
1438 			vm_page_lock(dst_m);
1439 			vm_page_activate(dst_m);
1440 			vm_page_unlock(dst_m);
1441 		}
1442 		vm_page_xunbusy(dst_m);
1443 	}
1444 	VM_OBJECT_WUNLOCK(dst_object);
1445 	if (upgrade) {
1446 		dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
1447 		vm_object_deallocate(src_object);
1448 	}
1449 }
1450 
1451 
1452 /*
1453  * This routine checks around the requested page for other pages that
1454  * might be able to be faulted in.  This routine brackets the viable
1455  * pages for the pages to be paged in.
1456  *
1457  * Inputs:
1458  *	m, rbehind, rahead
1459  *
1460  * Outputs:
1461  *  marray (array of vm_page_t), reqpage (index of requested page)
1462  *
1463  * Return value:
1464  *  number of pages in marray
1465  */
1466 static int
1467 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage)
1468 	vm_page_t m;
1469 	int rbehind;
1470 	int rahead;
1471 	vm_page_t *marray;
1472 	int *reqpage;
1473 {
1474 	int i,j;
1475 	vm_object_t object;
1476 	vm_pindex_t pindex, startpindex, endpindex, tpindex;
1477 	vm_page_t rtm;
1478 	int cbehind, cahead;
1479 
1480 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1481 
1482 	object = m->object;
1483 	pindex = m->pindex;
1484 	cbehind = cahead = 0;
1485 
1486 	/*
1487 	 * if the requested page is not available, then give up now
1488 	 */
1489 	if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1490 		return 0;
1491 	}
1492 
1493 	if ((cbehind == 0) && (cahead == 0)) {
1494 		*reqpage = 0;
1495 		marray[0] = m;
1496 		return 1;
1497 	}
1498 
1499 	if (rahead > cahead) {
1500 		rahead = cahead;
1501 	}
1502 
1503 	if (rbehind > cbehind) {
1504 		rbehind = cbehind;
1505 	}
1506 
1507 	/*
1508 	 * scan backward for the read behind pages -- in memory
1509 	 */
1510 	if (pindex > 0) {
1511 		if (rbehind > pindex) {
1512 			rbehind = pindex;
1513 			startpindex = 0;
1514 		} else {
1515 			startpindex = pindex - rbehind;
1516 		}
1517 
1518 		if ((rtm = TAILQ_PREV(m, pglist, listq)) != NULL &&
1519 		    rtm->pindex >= startpindex)
1520 			startpindex = rtm->pindex + 1;
1521 
1522 		/* tpindex is unsigned; beware of numeric underflow. */
1523 		for (i = 0, tpindex = pindex - 1; tpindex >= startpindex &&
1524 		    tpindex < pindex; i++, tpindex--) {
1525 
1526 			rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1527 			    VM_ALLOC_IFNOTCACHED);
1528 			if (rtm == NULL) {
1529 				/*
1530 				 * Shift the allocated pages to the
1531 				 * beginning of the array.
1532 				 */
1533 				for (j = 0; j < i; j++) {
1534 					marray[j] = marray[j + tpindex + 1 -
1535 					    startpindex];
1536 				}
1537 				break;
1538 			}
1539 
1540 			marray[tpindex - startpindex] = rtm;
1541 		}
1542 	} else {
1543 		startpindex = 0;
1544 		i = 0;
1545 	}
1546 
1547 	marray[i] = m;
1548 	/* page offset of the required page */
1549 	*reqpage = i;
1550 
1551 	tpindex = pindex + 1;
1552 	i++;
1553 
1554 	/*
1555 	 * scan forward for the read ahead pages
1556 	 */
1557 	endpindex = tpindex + rahead;
1558 	if ((rtm = TAILQ_NEXT(m, listq)) != NULL && rtm->pindex < endpindex)
1559 		endpindex = rtm->pindex;
1560 	if (endpindex > object->size)
1561 		endpindex = object->size;
1562 
1563 	for (; tpindex < endpindex; i++, tpindex++) {
1564 
1565 		rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1566 		    VM_ALLOC_IFNOTCACHED);
1567 		if (rtm == NULL) {
1568 			break;
1569 		}
1570 
1571 		marray[i] = rtm;
1572 	}
1573 
1574 	/* return number of pages */
1575 	return i;
1576 }
1577 
1578 /*
1579  * Block entry into the machine-independent layer's page fault handler by
1580  * the calling thread.  Subsequent calls to vm_fault() by that thread will
1581  * return KERN_PROTECTION_FAILURE.  Enable machine-dependent handling of
1582  * spurious page faults.
1583  */
1584 int
1585 vm_fault_disable_pagefaults(void)
1586 {
1587 
1588 	return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR));
1589 }
1590 
1591 void
1592 vm_fault_enable_pagefaults(int save)
1593 {
1594 
1595 	curthread_pflags_restore(save);
1596 }
1597