xref: /freebsd/sys/vm/vm_fault.c (revision ba8d15d3a8b10be9f3a0cb86a60246f225a36736)
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 
105 #define PFBAK 4
106 #define PFFOR 4
107 
108 static int vm_fault_additional_pages(vm_page_t, int, int, vm_page_t *, int *);
109 
110 #define	VM_FAULT_READ_BEHIND	8
111 #define	VM_FAULT_READ_DEFAULT	(1 + VM_FAULT_READ_AHEAD_INIT)
112 #define	VM_FAULT_READ_MAX	(1 + VM_FAULT_READ_AHEAD_MAX)
113 #define	VM_FAULT_NINCR		(VM_FAULT_READ_MAX / VM_FAULT_READ_BEHIND)
114 #define	VM_FAULT_SUM		(VM_FAULT_NINCR * (VM_FAULT_NINCR + 1) / 2)
115 #define	VM_FAULT_CACHE_BEHIND	(VM_FAULT_READ_BEHIND * VM_FAULT_SUM)
116 
117 struct faultstate {
118 	vm_page_t m;
119 	vm_object_t object;
120 	vm_pindex_t pindex;
121 	vm_page_t first_m;
122 	vm_object_t	first_object;
123 	vm_pindex_t first_pindex;
124 	vm_map_t map;
125 	vm_map_entry_t entry;
126 	int lookup_still_valid;
127 	struct vnode *vp;
128 };
129 
130 static void vm_fault_cache_behind(const struct faultstate *fs, int distance);
131 static void vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
132 	    int faultcount, int reqpage);
133 
134 static inline void
135 release_page(struct faultstate *fs)
136 {
137 
138 	vm_page_xunbusy(fs->m);
139 	vm_page_lock(fs->m);
140 	vm_page_deactivate(fs->m);
141 	vm_page_unlock(fs->m);
142 	fs->m = NULL;
143 }
144 
145 static inline void
146 unlock_map(struct faultstate *fs)
147 {
148 
149 	if (fs->lookup_still_valid) {
150 		vm_map_lookup_done(fs->map, fs->entry);
151 		fs->lookup_still_valid = FALSE;
152 	}
153 }
154 
155 static void
156 unlock_and_deallocate(struct faultstate *fs)
157 {
158 
159 	vm_object_pip_wakeup(fs->object);
160 	VM_OBJECT_WUNLOCK(fs->object);
161 	if (fs->object != fs->first_object) {
162 		VM_OBJECT_WLOCK(fs->first_object);
163 		vm_page_lock(fs->first_m);
164 		vm_page_free(fs->first_m);
165 		vm_page_unlock(fs->first_m);
166 		vm_object_pip_wakeup(fs->first_object);
167 		VM_OBJECT_WUNLOCK(fs->first_object);
168 		fs->first_m = NULL;
169 	}
170 	vm_object_deallocate(fs->first_object);
171 	unlock_map(fs);
172 	if (fs->vp != NULL) {
173 		vput(fs->vp);
174 		fs->vp = NULL;
175 	}
176 }
177 
178 static void
179 vm_fault_dirty(vm_map_entry_t entry, vm_page_t m, vm_prot_t prot,
180     vm_prot_t fault_type, int fault_flags, boolean_t set_wd)
181 {
182 	boolean_t need_dirty;
183 
184 	if (((prot & VM_PROT_WRITE) == 0 &&
185 	    (fault_flags & VM_FAULT_DIRTY) == 0) ||
186 	    (m->oflags & VPO_UNMANAGED) != 0)
187 		return;
188 
189 	VM_OBJECT_ASSERT_LOCKED(m->object);
190 
191 	need_dirty = ((fault_type & VM_PROT_WRITE) != 0 &&
192 	    (fault_flags & VM_FAULT_CHANGE_WIRING) == 0) ||
193 	    (fault_flags & VM_FAULT_DIRTY) != 0;
194 
195 	if (set_wd)
196 		vm_object_set_writeable_dirty(m->object);
197 	else
198 		/*
199 		 * If two callers of vm_fault_dirty() with set_wd ==
200 		 * FALSE, one for the map entry with MAP_ENTRY_NOSYNC
201 		 * flag set, other with flag clear, race, it is
202 		 * possible for the no-NOSYNC thread to see m->dirty
203 		 * != 0 and not clear VPO_NOSYNC.  Take vm_page lock
204 		 * around manipulation of VPO_NOSYNC and
205 		 * vm_page_dirty() call, to avoid the race and keep
206 		 * m->oflags consistent.
207 		 */
208 		vm_page_lock(m);
209 
210 	/*
211 	 * If this is a NOSYNC mmap we do not want to set VPO_NOSYNC
212 	 * if the page is already dirty to prevent data written with
213 	 * the expectation of being synced from not being synced.
214 	 * Likewise if this entry does not request NOSYNC then make
215 	 * sure the page isn't marked NOSYNC.  Applications sharing
216 	 * data should use the same flags to avoid ping ponging.
217 	 */
218 	if ((entry->eflags & MAP_ENTRY_NOSYNC) != 0) {
219 		if (m->dirty == 0) {
220 			m->oflags |= VPO_NOSYNC;
221 		}
222 	} else {
223 		m->oflags &= ~VPO_NOSYNC;
224 	}
225 
226 	/*
227 	 * If the fault is a write, we know that this page is being
228 	 * written NOW so dirty it explicitly to save on
229 	 * pmap_is_modified() calls later.
230 	 *
231 	 * Also tell the backing pager, if any, that it should remove
232 	 * any swap backing since the page is now dirty.
233 	 */
234 	if (need_dirty)
235 		vm_page_dirty(m);
236 	if (!set_wd)
237 		vm_page_unlock(m);
238 	if (need_dirty)
239 		vm_pager_page_unswapped(m);
240 }
241 
242 /*
243  * TRYPAGER - used by vm_fault to calculate whether the pager for the
244  *	      current object *might* contain the page.
245  *
246  *	      default objects are zero-fill, there is no real pager.
247  */
248 #define TRYPAGER	(fs.object->type != OBJT_DEFAULT && \
249 			((fault_flags & VM_FAULT_CHANGE_WIRING) == 0 || wired))
250 
251 /*
252  *	vm_fault:
253  *
254  *	Handle a page fault occurring at the given address,
255  *	requiring the given permissions, in the map specified.
256  *	If successful, the page is inserted into the
257  *	associated physical map.
258  *
259  *	NOTE: the given address should be truncated to the
260  *	proper page address.
261  *
262  *	KERN_SUCCESS is returned if the page fault is handled; otherwise,
263  *	a standard error specifying why the fault is fatal is returned.
264  *
265  *	The map in question must be referenced, and remains so.
266  *	Caller may hold no locks.
267  */
268 int
269 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
270     int fault_flags)
271 {
272 	struct thread *td;
273 	int result;
274 
275 	td = curthread;
276 	if ((td->td_pflags & TDP_NOFAULTING) != 0)
277 		return (KERN_PROTECTION_FAILURE);
278 #ifdef KTRACE
279 	if (map != kernel_map && KTRPOINT(td, KTR_FAULT))
280 		ktrfault(vaddr, fault_type);
281 #endif
282 	result = vm_fault_hold(map, trunc_page(vaddr), fault_type, fault_flags,
283 	    NULL);
284 #ifdef KTRACE
285 	if (map != kernel_map && KTRPOINT(td, KTR_FAULTEND))
286 		ktrfaultend(result);
287 #endif
288 	return (result);
289 }
290 
291 int
292 vm_fault_hold(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
293     int fault_flags, vm_page_t *m_hold)
294 {
295 	vm_prot_t prot;
296 	int alloc_req, era, faultcount, nera, reqpage, result;
297 	boolean_t growstack, is_first_object_locked, wired;
298 	int map_generation;
299 	vm_object_t next_object;
300 	vm_page_t marray[VM_FAULT_READ_MAX];
301 	int hardfault;
302 	struct faultstate fs;
303 	struct vnode *vp;
304 	vm_page_t m;
305 	int ahead, behind, cluster_offset, error, locked;
306 
307 	hardfault = 0;
308 	growstack = TRUE;
309 	PCPU_INC(cnt.v_vm_faults);
310 	fs.vp = NULL;
311 	faultcount = reqpage = 0;
312 
313 RetryFault:;
314 
315 	/*
316 	 * Find the backing store object and offset into it to begin the
317 	 * search.
318 	 */
319 	fs.map = map;
320 	result = vm_map_lookup(&fs.map, vaddr, fault_type, &fs.entry,
321 	    &fs.first_object, &fs.first_pindex, &prot, &wired);
322 	if (result != KERN_SUCCESS) {
323 		if (growstack && result == KERN_INVALID_ADDRESS &&
324 		    map != kernel_map) {
325 			result = vm_map_growstack(curproc, vaddr);
326 			if (result != KERN_SUCCESS)
327 				return (KERN_FAILURE);
328 			growstack = FALSE;
329 			goto RetryFault;
330 		}
331 		return (result);
332 	}
333 
334 	map_generation = fs.map->timestamp;
335 
336 	if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
337 		panic("vm_fault: fault on nofault entry, addr: %lx",
338 		    (u_long)vaddr);
339 	}
340 
341 	if (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION &&
342 	    fs.entry->wiring_thread != curthread) {
343 		vm_map_unlock_read(fs.map);
344 		vm_map_lock(fs.map);
345 		if (vm_map_lookup_entry(fs.map, vaddr, &fs.entry) &&
346 		    (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION)) {
347 			fs.entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
348 			vm_map_unlock_and_wait(fs.map, 0);
349 		} else
350 			vm_map_unlock(fs.map);
351 		goto RetryFault;
352 	}
353 
354 	if (wired)
355 		fault_type = prot | (fault_type & VM_PROT_COPY);
356 
357 	if (fs.vp == NULL /* avoid locked vnode leak */ &&
358 	    (fault_flags & (VM_FAULT_CHANGE_WIRING | VM_FAULT_DIRTY)) == 0 &&
359 	    /* avoid calling vm_object_set_writeable_dirty() */
360 	    ((prot & VM_PROT_WRITE) == 0 ||
361 	    (fs.first_object->type != OBJT_VNODE &&
362 	    (fs.first_object->flags & OBJ_TMPFS_NODE) == 0) ||
363 	    (fs.first_object->flags & OBJ_MIGHTBEDIRTY) != 0)) {
364 		VM_OBJECT_RLOCK(fs.first_object);
365 		if ((prot & VM_PROT_WRITE) != 0 &&
366 		    (fs.first_object->type == OBJT_VNODE ||
367 		    (fs.first_object->flags & OBJ_TMPFS_NODE) != 0) &&
368 		    (fs.first_object->flags & OBJ_MIGHTBEDIRTY) == 0)
369 			goto fast_failed;
370 		m = vm_page_lookup(fs.first_object, fs.first_pindex);
371 		/* A busy page can be mapped for read|execute access. */
372 		if (m == NULL || ((prot & VM_PROT_WRITE) != 0 &&
373 		    vm_page_busied(m)) || m->valid != VM_PAGE_BITS_ALL)
374 			goto fast_failed;
375 		result = pmap_enter(fs.map->pmap, vaddr, m, prot,
376 		   fault_type | PMAP_ENTER_NOSLEEP | (wired ? PMAP_ENTER_WIRED :
377 		   0), 0);
378 		if (result != KERN_SUCCESS)
379 			goto fast_failed;
380 		if (m_hold != NULL) {
381 			*m_hold = m;
382 			vm_page_lock(m);
383 			vm_page_hold(m);
384 			vm_page_unlock(m);
385 		}
386 		vm_fault_dirty(fs.entry, m, prot, fault_type, fault_flags,
387 		    FALSE);
388 		VM_OBJECT_RUNLOCK(fs.first_object);
389 		if (!wired)
390 			vm_fault_prefault(&fs, vaddr, 0, 0);
391 		vm_map_lookup_done(fs.map, fs.entry);
392 		curthread->td_ru.ru_minflt++;
393 		return (KERN_SUCCESS);
394 fast_failed:
395 		if (!VM_OBJECT_TRYUPGRADE(fs.first_object)) {
396 			VM_OBJECT_RUNLOCK(fs.first_object);
397 			VM_OBJECT_WLOCK(fs.first_object);
398 		}
399 	} else {
400 		VM_OBJECT_WLOCK(fs.first_object);
401 	}
402 
403 	/*
404 	 * Make a reference to this object to prevent its disposal while we
405 	 * are messing with it.  Once we have the reference, the map is free
406 	 * to be diddled.  Since objects reference their shadows (and copies),
407 	 * they will stay around as well.
408 	 *
409 	 * Bump the paging-in-progress count to prevent size changes (e.g.
410 	 * truncation operations) during I/O.  This must be done after
411 	 * obtaining the vnode lock in order to avoid possible deadlocks.
412 	 */
413 	vm_object_reference_locked(fs.first_object);
414 	vm_object_pip_add(fs.first_object, 1);
415 
416 	fs.lookup_still_valid = TRUE;
417 
418 	fs.first_m = NULL;
419 
420 	/*
421 	 * Search for the page at object/offset.
422 	 */
423 	fs.object = fs.first_object;
424 	fs.pindex = fs.first_pindex;
425 	while (TRUE) {
426 		/*
427 		 * If the object is dead, we stop here
428 		 */
429 		if (fs.object->flags & OBJ_DEAD) {
430 			unlock_and_deallocate(&fs);
431 			return (KERN_PROTECTION_FAILURE);
432 		}
433 
434 		/*
435 		 * See if page is resident
436 		 */
437 		fs.m = vm_page_lookup(fs.object, fs.pindex);
438 		if (fs.m != NULL) {
439 			/*
440 			 * Wait/Retry if the page is busy.  We have to do this
441 			 * if the page is either exclusive or shared busy
442 			 * because the vm_pager may be using read busy for
443 			 * pageouts (and even pageins if it is the vnode
444 			 * pager), and we could end up trying to pagein and
445 			 * pageout the same page simultaneously.
446 			 *
447 			 * We can theoretically allow the busy case on a read
448 			 * fault if the page is marked valid, but since such
449 			 * pages are typically already pmap'd, putting that
450 			 * special case in might be more effort then it is
451 			 * worth.  We cannot under any circumstances mess
452 			 * around with a shared busied page except, perhaps,
453 			 * to pmap it.
454 			 */
455 			if (vm_page_busied(fs.m)) {
456 				/*
457 				 * Reference the page before unlocking and
458 				 * sleeping so that the page daemon is less
459 				 * likely to reclaim it.
460 				 */
461 				vm_page_aflag_set(fs.m, PGA_REFERENCED);
462 				if (fs.object != fs.first_object) {
463 					if (!VM_OBJECT_TRYWLOCK(
464 					    fs.first_object)) {
465 						VM_OBJECT_WUNLOCK(fs.object);
466 						VM_OBJECT_WLOCK(fs.first_object);
467 						VM_OBJECT_WLOCK(fs.object);
468 					}
469 					vm_page_lock(fs.first_m);
470 					vm_page_free(fs.first_m);
471 					vm_page_unlock(fs.first_m);
472 					vm_object_pip_wakeup(fs.first_object);
473 					VM_OBJECT_WUNLOCK(fs.first_object);
474 					fs.first_m = NULL;
475 				}
476 				unlock_map(&fs);
477 				if (fs.m == vm_page_lookup(fs.object,
478 				    fs.pindex)) {
479 					vm_page_sleep_if_busy(fs.m, "vmpfw");
480 				}
481 				vm_object_pip_wakeup(fs.object);
482 				VM_OBJECT_WUNLOCK(fs.object);
483 				PCPU_INC(cnt.v_intrans);
484 				vm_object_deallocate(fs.first_object);
485 				goto RetryFault;
486 			}
487 			vm_page_lock(fs.m);
488 			vm_page_remque(fs.m);
489 			vm_page_unlock(fs.m);
490 
491 			/*
492 			 * Mark page busy for other processes, and the
493 			 * pagedaemon.  If it still isn't completely valid
494 			 * (readable), jump to readrest, else break-out ( we
495 			 * found the page ).
496 			 */
497 			vm_page_xbusy(fs.m);
498 			if (fs.m->valid != VM_PAGE_BITS_ALL)
499 				goto readrest;
500 			break;
501 		}
502 
503 		/*
504 		 * Page is not resident, If this is the search termination
505 		 * or the pager might contain the page, allocate a new page.
506 		 */
507 		if (TRYPAGER || fs.object == fs.first_object) {
508 			if (fs.pindex >= fs.object->size) {
509 				unlock_and_deallocate(&fs);
510 				return (KERN_PROTECTION_FAILURE);
511 			}
512 
513 			/*
514 			 * Allocate a new page for this object/offset pair.
515 			 *
516 			 * Unlocked read of the p_flag is harmless. At
517 			 * worst, the P_KILLED might be not observed
518 			 * there, and allocation can fail, causing
519 			 * restart and new reading of the p_flag.
520 			 */
521 			fs.m = NULL;
522 			if (!vm_page_count_severe() || P_KILLED(curproc)) {
523 #if VM_NRESERVLEVEL > 0
524 				if ((fs.object->flags & OBJ_COLORED) == 0) {
525 					fs.object->flags |= OBJ_COLORED;
526 					fs.object->pg_color = atop(vaddr) -
527 					    fs.pindex;
528 				}
529 #endif
530 				alloc_req = P_KILLED(curproc) ?
531 				    VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL;
532 				if (fs.object->type != OBJT_VNODE &&
533 				    fs.object->backing_object == NULL)
534 					alloc_req |= VM_ALLOC_ZERO;
535 				fs.m = vm_page_alloc(fs.object, fs.pindex,
536 				    alloc_req);
537 			}
538 			if (fs.m == NULL) {
539 				unlock_and_deallocate(&fs);
540 				VM_WAITPFAULT;
541 				goto RetryFault;
542 			} else if (fs.m->valid == VM_PAGE_BITS_ALL)
543 				break;
544 		}
545 
546 readrest:
547 		/*
548 		 * We have found a valid page or we have allocated a new page.
549 		 * The page thus may not be valid or may not be entirely
550 		 * valid.
551 		 *
552 		 * Attempt to fault-in the page if there is a chance that the
553 		 * pager has it, and potentially fault in additional pages
554 		 * at the same time.
555 		 */
556 		if (TRYPAGER) {
557 			int rv;
558 			u_char behavior = vm_map_entry_behavior(fs.entry);
559 
560 			era = fs.entry->read_ahead;
561 			if (behavior == MAP_ENTRY_BEHAV_RANDOM ||
562 			    P_KILLED(curproc)) {
563 				behind = 0;
564 				nera = 0;
565 				ahead = 0;
566 			} else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) {
567 				behind = 0;
568 				nera = VM_FAULT_READ_AHEAD_MAX;
569 				ahead = nera;
570 				if (fs.pindex == fs.entry->next_read)
571 					vm_fault_cache_behind(&fs,
572 					    VM_FAULT_READ_MAX);
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_cache_behind(&fs,
591 					    VM_FAULT_CACHE_BEHIND);
592 			} else {
593 				/*
594 				 * This is a non-sequential fault.  Request a
595 				 * cluster of pages that is aligned to a
596 				 * VM_FAULT_READ_DEFAULT page offset boundary
597 				 * within the object.  Alignment to a page
598 				 * offset boundary is more likely to coincide
599 				 * with the underlying file system block than
600 				 * alignment to a virtual address boundary.
601 				 */
602 				cluster_offset = fs.pindex %
603 				    VM_FAULT_READ_DEFAULT;
604 				behind = ulmin(cluster_offset,
605 				    atop(vaddr - fs.entry->start));
606 				nera = 0;
607 				ahead = VM_FAULT_READ_DEFAULT - 1 -
608 				    cluster_offset;
609 			}
610 			ahead = ulmin(ahead, atop(fs.entry->end - vaddr) - 1);
611 			if (era != nera)
612 				fs.entry->read_ahead = nera;
613 
614 			/*
615 			 * Call the pager to retrieve the data, if any, after
616 			 * releasing the lock on the map.  We hold a ref on
617 			 * fs.object and the pages are exclusive busied.
618 			 */
619 			unlock_map(&fs);
620 
621 			if (fs.object->type == OBJT_VNODE) {
622 				vp = fs.object->handle;
623 				if (vp == fs.vp)
624 					goto vnode_locked;
625 				else if (fs.vp != NULL) {
626 					vput(fs.vp);
627 					fs.vp = NULL;
628 				}
629 				locked = VOP_ISLOCKED(vp);
630 
631 				if (locked != LK_EXCLUSIVE)
632 					locked = LK_SHARED;
633 				/* Do not sleep for vnode lock while fs.m is busy */
634 				error = vget(vp, locked | LK_CANRECURSE |
635 				    LK_NOWAIT, curthread);
636 				if (error != 0) {
637 					vhold(vp);
638 					release_page(&fs);
639 					unlock_and_deallocate(&fs);
640 					error = vget(vp, locked | LK_RETRY |
641 					    LK_CANRECURSE, curthread);
642 					vdrop(vp);
643 					fs.vp = vp;
644 					KASSERT(error == 0,
645 					    ("vm_fault: vget failed"));
646 					goto RetryFault;
647 				}
648 				fs.vp = vp;
649 			}
650 vnode_locked:
651 			KASSERT(fs.vp == NULL || !fs.map->system_map,
652 			    ("vm_fault: vnode-backed object mapped by system map"));
653 
654 			/*
655 			 * now we find out if any other pages should be paged
656 			 * in at this time this routine checks to see if the
657 			 * pages surrounding this fault reside in the same
658 			 * object as the page for this fault.  If they do,
659 			 * then they are faulted in also into the object.  The
660 			 * array "marray" returned contains an array of
661 			 * vm_page_t structs where one of them is the
662 			 * vm_page_t passed to the routine.  The reqpage
663 			 * return value is the index into the marray for the
664 			 * vm_page_t passed to the routine.
665 			 *
666 			 * fs.m plus the additional pages are exclusive busied.
667 			 */
668 			faultcount = vm_fault_additional_pages(
669 			    fs.m, behind, ahead, marray, &reqpage);
670 
671 			rv = faultcount ?
672 			    vm_pager_get_pages(fs.object, marray, faultcount,
673 				reqpage) : VM_PAGER_FAIL;
674 
675 			if (rv == VM_PAGER_OK) {
676 				/*
677 				 * Found the page. Leave it busy while we play
678 				 * with it.
679 				 */
680 
681 				/*
682 				 * Relookup in case pager changed page. Pager
683 				 * is responsible for disposition of old page
684 				 * if moved.
685 				 */
686 				fs.m = vm_page_lookup(fs.object, fs.pindex);
687 				if (!fs.m) {
688 					unlock_and_deallocate(&fs);
689 					goto RetryFault;
690 				}
691 
692 				hardfault++;
693 				break; /* break to PAGE HAS BEEN FOUND */
694 			}
695 			/*
696 			 * Remove the bogus page (which does not exist at this
697 			 * object/offset); before doing so, we must get back
698 			 * our object lock to preserve our invariant.
699 			 *
700 			 * Also wake up any other process that may want to bring
701 			 * in this page.
702 			 *
703 			 * If this is the top-level object, we must leave the
704 			 * busy page to prevent another process from rushing
705 			 * past us, and inserting the page in that object at
706 			 * the same time that we are.
707 			 */
708 			if (rv == VM_PAGER_ERROR)
709 				printf("vm_fault: pager read error, pid %d (%s)\n",
710 				    curproc->p_pid, curproc->p_comm);
711 			/*
712 			 * Data outside the range of the pager or an I/O error
713 			 */
714 			/*
715 			 * XXX - the check for kernel_map is a kludge to work
716 			 * around having the machine panic on a kernel space
717 			 * fault w/ I/O error.
718 			 */
719 			if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) ||
720 				(rv == VM_PAGER_BAD)) {
721 				vm_page_lock(fs.m);
722 				vm_page_free(fs.m);
723 				vm_page_unlock(fs.m);
724 				fs.m = NULL;
725 				unlock_and_deallocate(&fs);
726 				return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
727 			}
728 			if (fs.object != fs.first_object) {
729 				vm_page_lock(fs.m);
730 				vm_page_free(fs.m);
731 				vm_page_unlock(fs.m);
732 				fs.m = NULL;
733 				/*
734 				 * XXX - we cannot just fall out at this
735 				 * point, m has been freed and is invalid!
736 				 */
737 			}
738 		}
739 
740 		/*
741 		 * We get here if the object has default pager (or unwiring)
742 		 * or the pager doesn't have the page.
743 		 */
744 		if (fs.object == fs.first_object)
745 			fs.first_m = fs.m;
746 
747 		/*
748 		 * Move on to the next object.  Lock the next object before
749 		 * unlocking the current one.
750 		 */
751 		fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset);
752 		next_object = fs.object->backing_object;
753 		if (next_object == NULL) {
754 			/*
755 			 * If there's no object left, fill the page in the top
756 			 * object with zeros.
757 			 */
758 			if (fs.object != fs.first_object) {
759 				vm_object_pip_wakeup(fs.object);
760 				VM_OBJECT_WUNLOCK(fs.object);
761 
762 				fs.object = fs.first_object;
763 				fs.pindex = fs.first_pindex;
764 				fs.m = fs.first_m;
765 				VM_OBJECT_WLOCK(fs.object);
766 			}
767 			fs.first_m = NULL;
768 
769 			/*
770 			 * Zero the page if necessary and mark it valid.
771 			 */
772 			if ((fs.m->flags & PG_ZERO) == 0) {
773 				pmap_zero_page(fs.m);
774 			} else {
775 				PCPU_INC(cnt.v_ozfod);
776 			}
777 			PCPU_INC(cnt.v_zfod);
778 			fs.m->valid = VM_PAGE_BITS_ALL;
779 			/* Don't try to prefault neighboring pages. */
780 			faultcount = 1;
781 			break;	/* break to PAGE HAS BEEN FOUND */
782 		} else {
783 			KASSERT(fs.object != next_object,
784 			    ("object loop %p", next_object));
785 			VM_OBJECT_WLOCK(next_object);
786 			vm_object_pip_add(next_object, 1);
787 			if (fs.object != fs.first_object)
788 				vm_object_pip_wakeup(fs.object);
789 			VM_OBJECT_WUNLOCK(fs.object);
790 			fs.object = next_object;
791 		}
792 	}
793 
794 	vm_page_assert_xbusied(fs.m);
795 
796 	/*
797 	 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
798 	 * is held.]
799 	 */
800 
801 	/*
802 	 * If the page is being written, but isn't already owned by the
803 	 * top-level object, we have to copy it into a new page owned by the
804 	 * top-level object.
805 	 */
806 	if (fs.object != fs.first_object) {
807 		/*
808 		 * We only really need to copy if we want to write it.
809 		 */
810 		if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
811 			/*
812 			 * This allows pages to be virtually copied from a
813 			 * backing_object into the first_object, where the
814 			 * backing object has no other refs to it, and cannot
815 			 * gain any more refs.  Instead of a bcopy, we just
816 			 * move the page from the backing object to the
817 			 * first object.  Note that we must mark the page
818 			 * dirty in the first object so that it will go out
819 			 * to swap when needed.
820 			 */
821 			is_first_object_locked = FALSE;
822 			if (
823 				/*
824 				 * Only one shadow object
825 				 */
826 				(fs.object->shadow_count == 1) &&
827 				/*
828 				 * No COW refs, except us
829 				 */
830 				(fs.object->ref_count == 1) &&
831 				/*
832 				 * No one else can look this object up
833 				 */
834 				(fs.object->handle == NULL) &&
835 				/*
836 				 * No other ways to look the object up
837 				 */
838 				((fs.object->type == OBJT_DEFAULT) ||
839 				 (fs.object->type == OBJT_SWAP)) &&
840 			    (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs.first_object)) &&
841 				/*
842 				 * We don't chase down the shadow chain
843 				 */
844 			    fs.object == fs.first_object->backing_object) {
845 				/*
846 				 * get rid of the unnecessary page
847 				 */
848 				vm_page_lock(fs.first_m);
849 				vm_page_free(fs.first_m);
850 				vm_page_unlock(fs.first_m);
851 				/*
852 				 * grab the page and put it into the
853 				 * process'es object.  The page is
854 				 * automatically made dirty.
855 				 */
856 				if (vm_page_rename(fs.m, fs.first_object,
857 				    fs.first_pindex)) {
858 					unlock_and_deallocate(&fs);
859 					goto RetryFault;
860 				}
861 				vm_page_xbusy(fs.m);
862 				fs.first_m = fs.m;
863 				fs.m = NULL;
864 				PCPU_INC(cnt.v_cow_optim);
865 			} else {
866 				/*
867 				 * Oh, well, lets copy it.
868 				 */
869 				pmap_copy_page(fs.m, fs.first_m);
870 				fs.first_m->valid = VM_PAGE_BITS_ALL;
871 				if (wired && (fault_flags &
872 				    VM_FAULT_CHANGE_WIRING) == 0) {
873 					vm_page_lock(fs.first_m);
874 					vm_page_wire(fs.first_m);
875 					vm_page_unlock(fs.first_m);
876 
877 					vm_page_lock(fs.m);
878 					vm_page_unwire(fs.m, PQ_INACTIVE);
879 					vm_page_unlock(fs.m);
880 				}
881 				/*
882 				 * We no longer need the old page or object.
883 				 */
884 				release_page(&fs);
885 			}
886 			/*
887 			 * fs.object != fs.first_object due to above
888 			 * conditional
889 			 */
890 			vm_object_pip_wakeup(fs.object);
891 			VM_OBJECT_WUNLOCK(fs.object);
892 			/*
893 			 * Only use the new page below...
894 			 */
895 			fs.object = fs.first_object;
896 			fs.pindex = fs.first_pindex;
897 			fs.m = fs.first_m;
898 			if (!is_first_object_locked)
899 				VM_OBJECT_WLOCK(fs.object);
900 			PCPU_INC(cnt.v_cow_faults);
901 			curthread->td_cow++;
902 		} else {
903 			prot &= ~VM_PROT_WRITE;
904 		}
905 	}
906 
907 	/*
908 	 * We must verify that the maps have not changed since our last
909 	 * lookup.
910 	 */
911 	if (!fs.lookup_still_valid) {
912 		vm_object_t retry_object;
913 		vm_pindex_t retry_pindex;
914 		vm_prot_t retry_prot;
915 
916 		if (!vm_map_trylock_read(fs.map)) {
917 			release_page(&fs);
918 			unlock_and_deallocate(&fs);
919 			goto RetryFault;
920 		}
921 		fs.lookup_still_valid = TRUE;
922 		if (fs.map->timestamp != map_generation) {
923 			result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
924 			    &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
925 
926 			/*
927 			 * If we don't need the page any longer, put it on the inactive
928 			 * list (the easiest thing to do here).  If no one needs it,
929 			 * pageout will grab it eventually.
930 			 */
931 			if (result != KERN_SUCCESS) {
932 				release_page(&fs);
933 				unlock_and_deallocate(&fs);
934 
935 				/*
936 				 * If retry of map lookup would have blocked then
937 				 * retry fault from start.
938 				 */
939 				if (result == KERN_FAILURE)
940 					goto RetryFault;
941 				return (result);
942 			}
943 			if ((retry_object != fs.first_object) ||
944 			    (retry_pindex != fs.first_pindex)) {
945 				release_page(&fs);
946 				unlock_and_deallocate(&fs);
947 				goto RetryFault;
948 			}
949 
950 			/*
951 			 * Check whether the protection has changed or the object has
952 			 * been copied while we left the map unlocked. Changing from
953 			 * read to write permission is OK - we leave the page
954 			 * write-protected, and catch the write fault. Changing from
955 			 * write to read permission means that we can't mark the page
956 			 * write-enabled after all.
957 			 */
958 			prot &= retry_prot;
959 		}
960 	}
961 	/*
962 	 * If the page was filled by a pager, update the map entry's
963 	 * last read offset.  Since the pager does not return the
964 	 * actual set of pages that it read, this update is based on
965 	 * the requested set.  Typically, the requested and actual
966 	 * sets are the same.
967 	 *
968 	 * XXX The following assignment modifies the map
969 	 * without holding a write lock on it.
970 	 */
971 	if (hardfault)
972 		fs.entry->next_read = fs.pindex + faultcount - reqpage;
973 
974 	vm_fault_dirty(fs.entry, fs.m, prot, fault_type, fault_flags, TRUE);
975 	vm_page_assert_xbusied(fs.m);
976 
977 	/*
978 	 * Page must be completely valid or it is not fit to
979 	 * map into user space.  vm_pager_get_pages() ensures this.
980 	 */
981 	KASSERT(fs.m->valid == VM_PAGE_BITS_ALL,
982 	    ("vm_fault: page %p partially invalid", fs.m));
983 	VM_OBJECT_WUNLOCK(fs.object);
984 
985 	/*
986 	 * Put this page into the physical map.  We had to do the unlock above
987 	 * because pmap_enter() may sleep.  We don't put the page
988 	 * back on the active queue until later so that the pageout daemon
989 	 * won't find it (yet).
990 	 */
991 	pmap_enter(fs.map->pmap, vaddr, fs.m, prot,
992 	    fault_type | (wired ? PMAP_ENTER_WIRED : 0), 0);
993 	if (faultcount != 1 && (fault_flags & VM_FAULT_CHANGE_WIRING) == 0 &&
994 	    wired == 0)
995 		vm_fault_prefault(&fs, vaddr, faultcount, reqpage);
996 	VM_OBJECT_WLOCK(fs.object);
997 	vm_page_lock(fs.m);
998 
999 	/*
1000 	 * If the page is not wired down, then put it where the pageout daemon
1001 	 * can find it.
1002 	 */
1003 	if (fault_flags & VM_FAULT_CHANGE_WIRING) {
1004 		if (wired)
1005 			vm_page_wire(fs.m);
1006 		else
1007 			vm_page_unwire(fs.m, PQ_ACTIVE);
1008 	} else
1009 		vm_page_activate(fs.m);
1010 	if (m_hold != NULL) {
1011 		*m_hold = fs.m;
1012 		vm_page_hold(fs.m);
1013 	}
1014 	vm_page_unlock(fs.m);
1015 	vm_page_xunbusy(fs.m);
1016 
1017 	/*
1018 	 * Unlock everything, and return
1019 	 */
1020 	unlock_and_deallocate(&fs);
1021 	if (hardfault) {
1022 		PCPU_INC(cnt.v_io_faults);
1023 		curthread->td_ru.ru_majflt++;
1024 	} else
1025 		curthread->td_ru.ru_minflt++;
1026 
1027 	return (KERN_SUCCESS);
1028 }
1029 
1030 /*
1031  * Speed up the reclamation of up to "distance" pages that precede the
1032  * faulting pindex within the first object of the shadow chain.
1033  */
1034 static void
1035 vm_fault_cache_behind(const struct faultstate *fs, int distance)
1036 {
1037 	vm_object_t first_object, object;
1038 	vm_page_t m, m_prev;
1039 	vm_pindex_t pindex;
1040 
1041 	object = fs->object;
1042 	VM_OBJECT_ASSERT_WLOCKED(object);
1043 	first_object = fs->first_object;
1044 	if (first_object != object) {
1045 		if (!VM_OBJECT_TRYWLOCK(first_object)) {
1046 			VM_OBJECT_WUNLOCK(object);
1047 			VM_OBJECT_WLOCK(first_object);
1048 			VM_OBJECT_WLOCK(object);
1049 		}
1050 	}
1051 	/* Neither fictitious nor unmanaged pages can be cached. */
1052 	if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) {
1053 		if (fs->first_pindex < distance)
1054 			pindex = 0;
1055 		else
1056 			pindex = fs->first_pindex - distance;
1057 		if (pindex < OFF_TO_IDX(fs->entry->offset))
1058 			pindex = OFF_TO_IDX(fs->entry->offset);
1059 		m = first_object != object ? fs->first_m : fs->m;
1060 		vm_page_assert_xbusied(m);
1061 		m_prev = vm_page_prev(m);
1062 		while ((m = m_prev) != NULL && m->pindex >= pindex &&
1063 		    m->valid == VM_PAGE_BITS_ALL) {
1064 			m_prev = vm_page_prev(m);
1065 			if (vm_page_busied(m))
1066 				continue;
1067 			vm_page_lock(m);
1068 			if (m->hold_count == 0 && m->wire_count == 0) {
1069 				pmap_remove_all(m);
1070 				vm_page_aflag_clear(m, PGA_REFERENCED);
1071 				if (m->dirty != 0)
1072 					vm_page_deactivate(m);
1073 				else
1074 					vm_page_cache(m);
1075 			}
1076 			vm_page_unlock(m);
1077 		}
1078 	}
1079 	if (first_object != object)
1080 		VM_OBJECT_WUNLOCK(first_object);
1081 }
1082 
1083 /*
1084  * vm_fault_prefault provides a quick way of clustering
1085  * pagefaults into a processes address space.  It is a "cousin"
1086  * of vm_map_pmap_enter, except it runs at page fault time instead
1087  * of mmap time.
1088  */
1089 static void
1090 vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
1091     int faultcount, int reqpage)
1092 {
1093 	pmap_t pmap;
1094 	vm_map_entry_t entry;
1095 	vm_object_t backing_object, lobject;
1096 	vm_offset_t addr, starta;
1097 	vm_pindex_t pindex;
1098 	vm_page_t m;
1099 	int backward, forward, i;
1100 
1101 	pmap = fs->map->pmap;
1102 	if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
1103 		return;
1104 
1105 	if (faultcount > 0) {
1106 		backward = reqpage;
1107 		forward = faultcount - reqpage - 1;
1108 	} else {
1109 		backward = PFBAK;
1110 		forward = PFFOR;
1111 	}
1112 	entry = fs->entry;
1113 
1114 	starta = addra - backward * PAGE_SIZE;
1115 	if (starta < entry->start) {
1116 		starta = entry->start;
1117 	} else if (starta > addra) {
1118 		starta = 0;
1119 	}
1120 
1121 	/*
1122 	 * Generate the sequence of virtual addresses that are candidates for
1123 	 * prefaulting in an outward spiral from the faulting virtual address,
1124 	 * "addra".  Specifically, the sequence is "addra - PAGE_SIZE", "addra
1125 	 * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ...
1126 	 * If the candidate address doesn't have a backing physical page, then
1127 	 * the loop immediately terminates.
1128 	 */
1129 	for (i = 0; i < 2 * imax(backward, forward); i++) {
1130 		addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE :
1131 		    PAGE_SIZE);
1132 		if (addr > addra + forward * PAGE_SIZE)
1133 			addr = 0;
1134 
1135 		if (addr < starta || addr >= entry->end)
1136 			continue;
1137 
1138 		if (!pmap_is_prefaultable(pmap, addr))
1139 			continue;
1140 
1141 		pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1142 		lobject = entry->object.vm_object;
1143 		VM_OBJECT_RLOCK(lobject);
1144 		while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1145 		    lobject->type == OBJT_DEFAULT &&
1146 		    (backing_object = lobject->backing_object) != NULL) {
1147 			KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1148 			    0, ("vm_fault_prefault: unaligned object offset"));
1149 			pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1150 			VM_OBJECT_RLOCK(backing_object);
1151 			VM_OBJECT_RUNLOCK(lobject);
1152 			lobject = backing_object;
1153 		}
1154 		if (m == NULL) {
1155 			VM_OBJECT_RUNLOCK(lobject);
1156 			break;
1157 		}
1158 		if (m->valid == VM_PAGE_BITS_ALL &&
1159 		    (m->flags & PG_FICTITIOUS) == 0)
1160 			pmap_enter_quick(pmap, addr, m, entry->protection);
1161 		VM_OBJECT_RUNLOCK(lobject);
1162 	}
1163 }
1164 
1165 /*
1166  * Hold each of the physical pages that are mapped by the specified range of
1167  * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
1168  * and allow the specified types of access, "prot".  If all of the implied
1169  * pages are successfully held, then the number of held pages is returned
1170  * together with pointers to those pages in the array "ma".  However, if any
1171  * of the pages cannot be held, -1 is returned.
1172  */
1173 int
1174 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
1175     vm_prot_t prot, vm_page_t *ma, int max_count)
1176 {
1177 	vm_offset_t end, va;
1178 	vm_page_t *mp;
1179 	int count;
1180 	boolean_t pmap_failed;
1181 
1182 	if (len == 0)
1183 		return (0);
1184 	end = round_page(addr + len);
1185 	addr = trunc_page(addr);
1186 
1187 	/*
1188 	 * Check for illegal addresses.
1189 	 */
1190 	if (addr < vm_map_min(map) || addr > end || end > vm_map_max(map))
1191 		return (-1);
1192 
1193 	if (atop(end - addr) > max_count)
1194 		panic("vm_fault_quick_hold_pages: count > max_count");
1195 	count = atop(end - addr);
1196 
1197 	/*
1198 	 * Most likely, the physical pages are resident in the pmap, so it is
1199 	 * faster to try pmap_extract_and_hold() first.
1200 	 */
1201 	pmap_failed = FALSE;
1202 	for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) {
1203 		*mp = pmap_extract_and_hold(map->pmap, va, prot);
1204 		if (*mp == NULL)
1205 			pmap_failed = TRUE;
1206 		else if ((prot & VM_PROT_WRITE) != 0 &&
1207 		    (*mp)->dirty != VM_PAGE_BITS_ALL) {
1208 			/*
1209 			 * Explicitly dirty the physical page.  Otherwise, the
1210 			 * caller's changes may go unnoticed because they are
1211 			 * performed through an unmanaged mapping or by a DMA
1212 			 * operation.
1213 			 *
1214 			 * The object lock is not held here.
1215 			 * See vm_page_clear_dirty_mask().
1216 			 */
1217 			vm_page_dirty(*mp);
1218 		}
1219 	}
1220 	if (pmap_failed) {
1221 		/*
1222 		 * One or more pages could not be held by the pmap.  Either no
1223 		 * page was mapped at the specified virtual address or that
1224 		 * mapping had insufficient permissions.  Attempt to fault in
1225 		 * and hold these pages.
1226 		 */
1227 		for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE)
1228 			if (*mp == NULL && vm_fault_hold(map, va, prot,
1229 			    VM_FAULT_NORMAL, mp) != KERN_SUCCESS)
1230 				goto error;
1231 	}
1232 	return (count);
1233 error:
1234 	for (mp = ma; mp < ma + count; mp++)
1235 		if (*mp != NULL) {
1236 			vm_page_lock(*mp);
1237 			vm_page_unhold(*mp);
1238 			vm_page_unlock(*mp);
1239 		}
1240 	return (-1);
1241 }
1242 
1243 /*
1244  *	Routine:
1245  *		vm_fault_copy_entry
1246  *	Function:
1247  *		Create new shadow object backing dst_entry with private copy of
1248  *		all underlying pages. When src_entry is equal to dst_entry,
1249  *		function implements COW for wired-down map entry. Otherwise,
1250  *		it forks wired entry into dst_map.
1251  *
1252  *	In/out conditions:
1253  *		The source and destination maps must be locked for write.
1254  *		The source map entry must be wired down (or be a sharing map
1255  *		entry corresponding to a main map entry that is wired down).
1256  */
1257 void
1258 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1259     vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1260     vm_ooffset_t *fork_charge)
1261 {
1262 	vm_object_t backing_object, dst_object, object, src_object;
1263 	vm_pindex_t dst_pindex, pindex, src_pindex;
1264 	vm_prot_t access, prot;
1265 	vm_offset_t vaddr;
1266 	vm_page_t dst_m;
1267 	vm_page_t src_m;
1268 	boolean_t upgrade;
1269 
1270 #ifdef	lint
1271 	src_map++;
1272 #endif	/* lint */
1273 
1274 	upgrade = src_entry == dst_entry;
1275 	access = prot = dst_entry->protection;
1276 
1277 	src_object = src_entry->object.vm_object;
1278 	src_pindex = OFF_TO_IDX(src_entry->offset);
1279 
1280 	if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
1281 		dst_object = src_object;
1282 		vm_object_reference(dst_object);
1283 	} else {
1284 		/*
1285 		 * Create the top-level object for the destination entry. (Doesn't
1286 		 * actually shadow anything - we copy the pages directly.)
1287 		 */
1288 		dst_object = vm_object_allocate(OBJT_DEFAULT,
1289 		    OFF_TO_IDX(dst_entry->end - dst_entry->start));
1290 #if VM_NRESERVLEVEL > 0
1291 		dst_object->flags |= OBJ_COLORED;
1292 		dst_object->pg_color = atop(dst_entry->start);
1293 #endif
1294 	}
1295 
1296 	VM_OBJECT_WLOCK(dst_object);
1297 	KASSERT(upgrade || dst_entry->object.vm_object == NULL,
1298 	    ("vm_fault_copy_entry: vm_object not NULL"));
1299 	if (src_object != dst_object) {
1300 		dst_entry->object.vm_object = dst_object;
1301 		dst_entry->offset = 0;
1302 		dst_object->charge = dst_entry->end - dst_entry->start;
1303 	}
1304 	if (fork_charge != NULL) {
1305 		KASSERT(dst_entry->cred == NULL,
1306 		    ("vm_fault_copy_entry: leaked swp charge"));
1307 		dst_object->cred = curthread->td_ucred;
1308 		crhold(dst_object->cred);
1309 		*fork_charge += dst_object->charge;
1310 	} else if (dst_object->cred == NULL) {
1311 		KASSERT(dst_entry->cred != NULL, ("no cred for entry %p",
1312 		    dst_entry));
1313 		dst_object->cred = dst_entry->cred;
1314 		dst_entry->cred = NULL;
1315 	}
1316 
1317 	/*
1318 	 * If not an upgrade, then enter the mappings in the pmap as
1319 	 * read and/or execute accesses.  Otherwise, enter them as
1320 	 * write accesses.
1321 	 *
1322 	 * A writeable large page mapping is only created if all of
1323 	 * the constituent small page mappings are modified. Marking
1324 	 * PTEs as modified on inception allows promotion to happen
1325 	 * without taking potentially large number of soft faults.
1326 	 */
1327 	if (!upgrade)
1328 		access &= ~VM_PROT_WRITE;
1329 
1330 	/*
1331 	 * Loop through all of the virtual pages within the entry's
1332 	 * range, copying each page from the source object to the
1333 	 * destination object.  Since the source is wired, those pages
1334 	 * must exist.  In contrast, the destination is pageable.
1335 	 * Since the destination object does share any backing storage
1336 	 * with the source object, all of its pages must be dirtied,
1337 	 * regardless of whether they can be written.
1338 	 */
1339 	for (vaddr = dst_entry->start, dst_pindex = 0;
1340 	    vaddr < dst_entry->end;
1341 	    vaddr += PAGE_SIZE, dst_pindex++) {
1342 again:
1343 		/*
1344 		 * Find the page in the source object, and copy it in.
1345 		 * Because the source is wired down, the page will be
1346 		 * in memory.
1347 		 */
1348 		if (src_object != dst_object)
1349 			VM_OBJECT_RLOCK(src_object);
1350 		object = src_object;
1351 		pindex = src_pindex + dst_pindex;
1352 		while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
1353 		    (backing_object = object->backing_object) != NULL) {
1354 			/*
1355 			 * Unless the source mapping is read-only or
1356 			 * it is presently being upgraded from
1357 			 * read-only, the first object in the shadow
1358 			 * chain should provide all of the pages.  In
1359 			 * other words, this loop body should never be
1360 			 * executed when the source mapping is already
1361 			 * read/write.
1362 			 */
1363 			KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 ||
1364 			    upgrade,
1365 			    ("vm_fault_copy_entry: main object missing page"));
1366 
1367 			VM_OBJECT_RLOCK(backing_object);
1368 			pindex += OFF_TO_IDX(object->backing_object_offset);
1369 			if (object != dst_object)
1370 				VM_OBJECT_RUNLOCK(object);
1371 			object = backing_object;
1372 		}
1373 		KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing"));
1374 
1375 		if (object != dst_object) {
1376 			/*
1377 			 * Allocate a page in the destination object.
1378 			 */
1379 			dst_m = vm_page_alloc(dst_object, (src_object ==
1380 			    dst_object ? src_pindex : 0) + dst_pindex,
1381 			    VM_ALLOC_NORMAL);
1382 			if (dst_m == NULL) {
1383 				VM_OBJECT_WUNLOCK(dst_object);
1384 				VM_OBJECT_RUNLOCK(object);
1385 				VM_WAIT;
1386 				VM_OBJECT_WLOCK(dst_object);
1387 				goto again;
1388 			}
1389 			pmap_copy_page(src_m, dst_m);
1390 			VM_OBJECT_RUNLOCK(object);
1391 			dst_m->valid = VM_PAGE_BITS_ALL;
1392 			dst_m->dirty = VM_PAGE_BITS_ALL;
1393 		} else {
1394 			dst_m = src_m;
1395 			if (vm_page_sleep_if_busy(dst_m, "fltupg"))
1396 				goto again;
1397 			vm_page_xbusy(dst_m);
1398 			KASSERT(dst_m->valid == VM_PAGE_BITS_ALL,
1399 			    ("invalid dst page %p", dst_m));
1400 		}
1401 		VM_OBJECT_WUNLOCK(dst_object);
1402 
1403 		/*
1404 		 * Enter it in the pmap. If a wired, copy-on-write
1405 		 * mapping is being replaced by a write-enabled
1406 		 * mapping, then wire that new mapping.
1407 		 */
1408 		pmap_enter(dst_map->pmap, vaddr, dst_m, prot,
1409 		    access | (upgrade ? PMAP_ENTER_WIRED : 0), 0);
1410 
1411 		/*
1412 		 * Mark it no longer busy, and put it on the active list.
1413 		 */
1414 		VM_OBJECT_WLOCK(dst_object);
1415 
1416 		if (upgrade) {
1417 			if (src_m != dst_m) {
1418 				vm_page_lock(src_m);
1419 				vm_page_unwire(src_m, PQ_INACTIVE);
1420 				vm_page_unlock(src_m);
1421 				vm_page_lock(dst_m);
1422 				vm_page_wire(dst_m);
1423 				vm_page_unlock(dst_m);
1424 			} else {
1425 				KASSERT(dst_m->wire_count > 0,
1426 				    ("dst_m %p is not wired", dst_m));
1427 			}
1428 		} else {
1429 			vm_page_lock(dst_m);
1430 			vm_page_activate(dst_m);
1431 			vm_page_unlock(dst_m);
1432 		}
1433 		vm_page_xunbusy(dst_m);
1434 	}
1435 	VM_OBJECT_WUNLOCK(dst_object);
1436 	if (upgrade) {
1437 		dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
1438 		vm_object_deallocate(src_object);
1439 	}
1440 }
1441 
1442 
1443 /*
1444  * This routine checks around the requested page for other pages that
1445  * might be able to be faulted in.  This routine brackets the viable
1446  * pages for the pages to be paged in.
1447  *
1448  * Inputs:
1449  *	m, rbehind, rahead
1450  *
1451  * Outputs:
1452  *  marray (array of vm_page_t), reqpage (index of requested page)
1453  *
1454  * Return value:
1455  *  number of pages in marray
1456  */
1457 static int
1458 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage)
1459 	vm_page_t m;
1460 	int rbehind;
1461 	int rahead;
1462 	vm_page_t *marray;
1463 	int *reqpage;
1464 {
1465 	int i,j;
1466 	vm_object_t object;
1467 	vm_pindex_t pindex, startpindex, endpindex, tpindex;
1468 	vm_page_t rtm;
1469 	int cbehind, cahead;
1470 
1471 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1472 
1473 	object = m->object;
1474 	pindex = m->pindex;
1475 	cbehind = cahead = 0;
1476 
1477 	/*
1478 	 * if the requested page is not available, then give up now
1479 	 */
1480 	if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1481 		return 0;
1482 	}
1483 
1484 	if ((cbehind == 0) && (cahead == 0)) {
1485 		*reqpage = 0;
1486 		marray[0] = m;
1487 		return 1;
1488 	}
1489 
1490 	if (rahead > cahead) {
1491 		rahead = cahead;
1492 	}
1493 
1494 	if (rbehind > cbehind) {
1495 		rbehind = cbehind;
1496 	}
1497 
1498 	/*
1499 	 * scan backward for the read behind pages -- in memory
1500 	 */
1501 	if (pindex > 0) {
1502 		if (rbehind > pindex) {
1503 			rbehind = pindex;
1504 			startpindex = 0;
1505 		} else {
1506 			startpindex = pindex - rbehind;
1507 		}
1508 
1509 		if ((rtm = TAILQ_PREV(m, pglist, listq)) != NULL &&
1510 		    rtm->pindex >= startpindex)
1511 			startpindex = rtm->pindex + 1;
1512 
1513 		/* tpindex is unsigned; beware of numeric underflow. */
1514 		for (i = 0, tpindex = pindex - 1; tpindex >= startpindex &&
1515 		    tpindex < pindex; i++, tpindex--) {
1516 
1517 			rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1518 			    VM_ALLOC_IFNOTCACHED);
1519 			if (rtm == NULL) {
1520 				/*
1521 				 * Shift the allocated pages to the
1522 				 * beginning of the array.
1523 				 */
1524 				for (j = 0; j < i; j++) {
1525 					marray[j] = marray[j + tpindex + 1 -
1526 					    startpindex];
1527 				}
1528 				break;
1529 			}
1530 
1531 			marray[tpindex - startpindex] = rtm;
1532 		}
1533 	} else {
1534 		startpindex = 0;
1535 		i = 0;
1536 	}
1537 
1538 	marray[i] = m;
1539 	/* page offset of the required page */
1540 	*reqpage = i;
1541 
1542 	tpindex = pindex + 1;
1543 	i++;
1544 
1545 	/*
1546 	 * scan forward for the read ahead pages
1547 	 */
1548 	endpindex = tpindex + rahead;
1549 	if ((rtm = TAILQ_NEXT(m, listq)) != NULL && rtm->pindex < endpindex)
1550 		endpindex = rtm->pindex;
1551 	if (endpindex > object->size)
1552 		endpindex = object->size;
1553 
1554 	for (; tpindex < endpindex; i++, tpindex++) {
1555 
1556 		rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1557 		    VM_ALLOC_IFNOTCACHED);
1558 		if (rtm == NULL) {
1559 			break;
1560 		}
1561 
1562 		marray[i] = rtm;
1563 	}
1564 
1565 	/* return number of pages */
1566 	return i;
1567 }
1568 
1569 /*
1570  * Block entry into the machine-independent layer's page fault handler by
1571  * the calling thread.  Subsequent calls to vm_fault() by that thread will
1572  * return KERN_PROTECTION_FAILURE.  Enable machine-dependent handling of
1573  * spurious page faults.
1574  */
1575 int
1576 vm_fault_disable_pagefaults(void)
1577 {
1578 
1579 	return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR));
1580 }
1581 
1582 void
1583 vm_fault_enable_pagefaults(int save)
1584 {
1585 
1586 	curthread_pflags_restore(save);
1587 }
1588