xref: /freebsd/sys/vm/vm_fault.c (revision ece7a5e9849032f4a31e725714c2db89d055b706)
1 /*-
2  * SPDX-License-Identifier: (BSD-4-Clause AND MIT-CMU)
3  *
4  * Copyright (c) 1991, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  * Copyright (c) 1994 John S. Dyson
7  * All rights reserved.
8  * Copyright (c) 1994 David Greenman
9  * All rights reserved.
10  *
11  *
12  * This code is derived from software contributed to Berkeley by
13  * The Mach Operating System project at Carnegie-Mellon University.
14  *
15  * Redistribution and use in source and binary forms, with or without
16  * modification, are permitted provided that the following conditions
17  * are met:
18  * 1. Redistributions of source code must retain the above copyright
19  *    notice, this list of conditions and the following disclaimer.
20  * 2. Redistributions in binary form must reproduce the above copyright
21  *    notice, this list of conditions and the following disclaimer in the
22  *    documentation and/or other materials provided with the distribution.
23  * 3. All advertising materials mentioning features or use of this software
24  *    must display the following acknowledgement:
25  *	This product includes software developed by the University of
26  *	California, Berkeley and its contributors.
27  * 4. Neither the name of the University nor the names of its contributors
28  *    may be used to endorse or promote products derived from this software
29  *    without specific prior written permission.
30  *
31  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
32  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
33  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
34  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
35  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
36  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
37  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
38  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
39  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
40  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
41  * SUCH DAMAGE.
42  *
43  *	from: @(#)vm_fault.c	8.4 (Berkeley) 1/12/94
44  *
45  *
46  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
47  * All rights reserved.
48  *
49  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
50  *
51  * Permission to use, copy, modify and distribute this software and
52  * its documentation is hereby granted, provided that both the copyright
53  * notice and this permission notice appear in all copies of the
54  * software, derivative works or modified versions, and any portions
55  * thereof, and that both notices appear in supporting documentation.
56  *
57  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
58  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
59  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
60  *
61  * Carnegie Mellon requests users of this software to return to
62  *
63  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
64  *  School of Computer Science
65  *  Carnegie Mellon University
66  *  Pittsburgh PA 15213-3890
67  *
68  * any improvements or extensions that they make and grant Carnegie the
69  * rights to redistribute these changes.
70  */
71 
72 /*
73  *	Page fault handling module.
74  */
75 
76 #include <sys/cdefs.h>
77 __FBSDID("$FreeBSD$");
78 
79 #include "opt_ktrace.h"
80 #include "opt_vm.h"
81 
82 #include <sys/param.h>
83 #include <sys/systm.h>
84 #include <sys/kernel.h>
85 #include <sys/lock.h>
86 #include <sys/mman.h>
87 #include <sys/mutex.h>
88 #include <sys/proc.h>
89 #include <sys/racct.h>
90 #include <sys/refcount.h>
91 #include <sys/resourcevar.h>
92 #include <sys/rwlock.h>
93 #include <sys/signalvar.h>
94 #include <sys/sysctl.h>
95 #include <sys/sysent.h>
96 #include <sys/vmmeter.h>
97 #include <sys/vnode.h>
98 #ifdef KTRACE
99 #include <sys/ktrace.h>
100 #endif
101 
102 #include <vm/vm.h>
103 #include <vm/vm_param.h>
104 #include <vm/pmap.h>
105 #include <vm/vm_map.h>
106 #include <vm/vm_object.h>
107 #include <vm/vm_page.h>
108 #include <vm/vm_pageout.h>
109 #include <vm/vm_kern.h>
110 #include <vm/vm_pager.h>
111 #include <vm/vm_extern.h>
112 #include <vm/vm_reserv.h>
113 
114 #define PFBAK 4
115 #define PFFOR 4
116 
117 #define	VM_FAULT_READ_DEFAULT	(1 + VM_FAULT_READ_AHEAD_INIT)
118 #define	VM_FAULT_READ_MAX	(1 + VM_FAULT_READ_AHEAD_MAX)
119 
120 #define	VM_FAULT_DONTNEED_MIN	1048576
121 
122 struct faultstate {
123 	vm_page_t m;
124 	vm_object_t object;
125 	vm_pindex_t pindex;
126 	vm_page_t first_m;
127 	vm_object_t	first_object;
128 	vm_pindex_t first_pindex;
129 	vm_map_t map;
130 	vm_map_entry_t entry;
131 	int map_generation;
132 	bool lookup_still_valid;
133 	struct vnode *vp;
134 };
135 
136 static void vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr,
137 	    int ahead);
138 static void vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
139 	    int backward, int forward, bool obj_locked);
140 
141 static int vm_pfault_oom_attempts = 3;
142 SYSCTL_INT(_vm, OID_AUTO, pfault_oom_attempts, CTLFLAG_RWTUN,
143     &vm_pfault_oom_attempts, 0,
144     "Number of page allocation attempts in page fault handler before it "
145     "triggers OOM handling");
146 
147 static int vm_pfault_oom_wait = 10;
148 SYSCTL_INT(_vm, OID_AUTO, pfault_oom_wait, CTLFLAG_RWTUN,
149     &vm_pfault_oom_wait, 0,
150     "Number of seconds to wait for free pages before retrying "
151     "the page fault handler");
152 
153 static inline void
154 fault_page_release(vm_page_t *mp)
155 {
156 	vm_page_t m;
157 
158 	m = *mp;
159 	if (m != NULL) {
160 		/*
161 		 * We are likely to loop around again and attempt to busy
162 		 * this page.  Deactivating it leaves it available for
163 		 * pageout while optimizing fault restarts.
164 		 */
165 		vm_page_lock(m);
166 		vm_page_deactivate(m);
167 		vm_page_unlock(m);
168 		vm_page_xunbusy(m);
169 		*mp = NULL;
170 	}
171 }
172 
173 static inline void
174 fault_page_free(vm_page_t *mp)
175 {
176 	vm_page_t m;
177 
178 	m = *mp;
179 	if (m != NULL) {
180 		VM_OBJECT_ASSERT_WLOCKED(m->object);
181 		if (!vm_page_wired(m))
182 			vm_page_free(m);
183 		else
184 			vm_page_xunbusy(m);
185 		*mp = NULL;
186 	}
187 }
188 
189 static inline void
190 unlock_map(struct faultstate *fs)
191 {
192 
193 	if (fs->lookup_still_valid) {
194 		vm_map_lookup_done(fs->map, fs->entry);
195 		fs->lookup_still_valid = false;
196 	}
197 }
198 
199 static void
200 unlock_vp(struct faultstate *fs)
201 {
202 
203 	if (fs->vp != NULL) {
204 		vput(fs->vp);
205 		fs->vp = NULL;
206 	}
207 }
208 
209 static void
210 fault_deallocate(struct faultstate *fs)
211 {
212 
213 	fault_page_release(&fs->m);
214 	vm_object_pip_wakeup(fs->object);
215 	if (fs->object != fs->first_object) {
216 		VM_OBJECT_WLOCK(fs->first_object);
217 		fault_page_free(&fs->first_m);
218 		VM_OBJECT_WUNLOCK(fs->first_object);
219 		vm_object_pip_wakeup(fs->first_object);
220 	}
221 	vm_object_deallocate(fs->first_object);
222 	unlock_map(fs);
223 	unlock_vp(fs);
224 }
225 
226 static void
227 unlock_and_deallocate(struct faultstate *fs)
228 {
229 
230 	VM_OBJECT_WUNLOCK(fs->object);
231 	fault_deallocate(fs);
232 }
233 
234 static void
235 vm_fault_dirty(vm_map_entry_t entry, vm_page_t m, vm_prot_t prot,
236     vm_prot_t fault_type, int fault_flags)
237 {
238 	bool need_dirty;
239 
240 	if (((prot & VM_PROT_WRITE) == 0 &&
241 	    (fault_flags & VM_FAULT_DIRTY) == 0) ||
242 	    (m->oflags & VPO_UNMANAGED) != 0)
243 		return;
244 
245 	VM_PAGE_OBJECT_BUSY_ASSERT(m);
246 
247 	need_dirty = ((fault_type & VM_PROT_WRITE) != 0 &&
248 	    (fault_flags & VM_FAULT_WIRE) == 0) ||
249 	    (fault_flags & VM_FAULT_DIRTY) != 0;
250 
251 	vm_object_set_writeable_dirty(m->object);
252 
253 	/*
254 	 * If the fault is a write, we know that this page is being
255 	 * written NOW so dirty it explicitly to save on
256 	 * pmap_is_modified() calls later.
257 	 *
258 	 * Also, since the page is now dirty, we can possibly tell
259 	 * the pager to release any swap backing the page.
260 	 */
261 	if (need_dirty && vm_page_set_dirty(m) == 0) {
262 		/*
263 		 * If this is a NOSYNC mmap we do not want to set PGA_NOSYNC
264 		 * if the page is already dirty to prevent data written with
265 		 * the expectation of being synced from not being synced.
266 		 * Likewise if this entry does not request NOSYNC then make
267 		 * sure the page isn't marked NOSYNC.  Applications sharing
268 		 * data should use the same flags to avoid ping ponging.
269 		 */
270 		if ((entry->eflags & MAP_ENTRY_NOSYNC) != 0)
271 			vm_page_aflag_set(m, PGA_NOSYNC);
272 		else
273 			vm_page_aflag_clear(m, PGA_NOSYNC);
274 	}
275 
276 }
277 
278 /*
279  * Unlocks fs.first_object and fs.map on success.
280  */
281 static int
282 vm_fault_soft_fast(struct faultstate *fs, vm_offset_t vaddr, vm_prot_t prot,
283     int fault_type, int fault_flags, boolean_t wired, vm_page_t *m_hold)
284 {
285 	vm_page_t m, m_map;
286 #if (defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \
287     __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv)) && \
288     VM_NRESERVLEVEL > 0
289 	vm_page_t m_super;
290 	int flags;
291 #endif
292 	int psind, rv;
293 
294 	MPASS(fs->vp == NULL);
295 	vm_object_busy(fs->first_object);
296 	m = vm_page_lookup(fs->first_object, fs->first_pindex);
297 	/* A busy page can be mapped for read|execute access. */
298 	if (m == NULL || ((prot & VM_PROT_WRITE) != 0 &&
299 	    vm_page_busied(m)) || !vm_page_all_valid(m)) {
300 		rv = KERN_FAILURE;
301 		goto out;
302 	}
303 	m_map = m;
304 	psind = 0;
305 #if (defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \
306     __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv)) && \
307     VM_NRESERVLEVEL > 0
308 	if ((m->flags & PG_FICTITIOUS) == 0 &&
309 	    (m_super = vm_reserv_to_superpage(m)) != NULL &&
310 	    rounddown2(vaddr, pagesizes[m_super->psind]) >= fs->entry->start &&
311 	    roundup2(vaddr + 1, pagesizes[m_super->psind]) <= fs->entry->end &&
312 	    (vaddr & (pagesizes[m_super->psind] - 1)) == (VM_PAGE_TO_PHYS(m) &
313 	    (pagesizes[m_super->psind] - 1)) && !wired &&
314 	    pmap_ps_enabled(fs->map->pmap)) {
315 		flags = PS_ALL_VALID;
316 		if ((prot & VM_PROT_WRITE) != 0) {
317 			/*
318 			 * Create a superpage mapping allowing write access
319 			 * only if none of the constituent pages are busy and
320 			 * all of them are already dirty (except possibly for
321 			 * the page that was faulted on).
322 			 */
323 			flags |= PS_NONE_BUSY;
324 			if ((fs->first_object->flags & OBJ_UNMANAGED) == 0)
325 				flags |= PS_ALL_DIRTY;
326 		}
327 		if (vm_page_ps_test(m_super, flags, m)) {
328 			m_map = m_super;
329 			psind = m_super->psind;
330 			vaddr = rounddown2(vaddr, pagesizes[psind]);
331 			/* Preset the modified bit for dirty superpages. */
332 			if ((flags & PS_ALL_DIRTY) != 0)
333 				fault_type |= VM_PROT_WRITE;
334 		}
335 	}
336 #endif
337 	rv = pmap_enter(fs->map->pmap, vaddr, m_map, prot, fault_type |
338 	    PMAP_ENTER_NOSLEEP | (wired ? PMAP_ENTER_WIRED : 0), psind);
339 	if (rv != KERN_SUCCESS)
340 		goto out;
341 	if (m_hold != NULL) {
342 		*m_hold = m;
343 		vm_page_wire(m);
344 	}
345 	vm_fault_dirty(fs->entry, m, prot, fault_type, fault_flags);
346 	if (psind == 0 && !wired)
347 		vm_fault_prefault(fs, vaddr, PFBAK, PFFOR, true);
348 	VM_OBJECT_RUNLOCK(fs->first_object);
349 	vm_map_lookup_done(fs->map, fs->entry);
350 	curthread->td_ru.ru_minflt++;
351 
352 out:
353 	vm_object_unbusy(fs->first_object);
354 	return (rv);
355 }
356 
357 static void
358 vm_fault_restore_map_lock(struct faultstate *fs)
359 {
360 
361 	VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
362 	MPASS(REFCOUNT_COUNT(fs->first_object->paging_in_progress) > 0);
363 
364 	if (!vm_map_trylock_read(fs->map)) {
365 		VM_OBJECT_WUNLOCK(fs->first_object);
366 		vm_map_lock_read(fs->map);
367 		VM_OBJECT_WLOCK(fs->first_object);
368 	}
369 	fs->lookup_still_valid = true;
370 }
371 
372 static void
373 vm_fault_populate_check_page(vm_page_t m)
374 {
375 
376 	/*
377 	 * Check each page to ensure that the pager is obeying the
378 	 * interface: the page must be installed in the object, fully
379 	 * valid, and exclusively busied.
380 	 */
381 	MPASS(m != NULL);
382 	MPASS(vm_page_all_valid(m));
383 	MPASS(vm_page_xbusied(m));
384 }
385 
386 static void
387 vm_fault_populate_cleanup(vm_object_t object, vm_pindex_t first,
388     vm_pindex_t last)
389 {
390 	vm_page_t m;
391 	vm_pindex_t pidx;
392 
393 	VM_OBJECT_ASSERT_WLOCKED(object);
394 	MPASS(first <= last);
395 	for (pidx = first, m = vm_page_lookup(object, pidx);
396 	    pidx <= last; pidx++, m = vm_page_next(m)) {
397 		vm_fault_populate_check_page(m);
398 		vm_page_lock(m);
399 		vm_page_deactivate(m);
400 		vm_page_unlock(m);
401 		vm_page_xunbusy(m);
402 	}
403 }
404 
405 static int
406 vm_fault_populate(struct faultstate *fs, vm_prot_t prot, int fault_type,
407     int fault_flags, boolean_t wired, vm_page_t *m_hold)
408 {
409 	struct mtx *m_mtx;
410 	vm_offset_t vaddr;
411 	vm_page_t m;
412 	vm_pindex_t map_first, map_last, pager_first, pager_last, pidx;
413 	int i, npages, psind, rv;
414 
415 	MPASS(fs->object == fs->first_object);
416 	VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
417 	MPASS(REFCOUNT_COUNT(fs->first_object->paging_in_progress) > 0);
418 	MPASS(fs->first_object->backing_object == NULL);
419 	MPASS(fs->lookup_still_valid);
420 
421 	pager_first = OFF_TO_IDX(fs->entry->offset);
422 	pager_last = pager_first + atop(fs->entry->end - fs->entry->start) - 1;
423 	unlock_map(fs);
424 	unlock_vp(fs);
425 
426 	/*
427 	 * Call the pager (driver) populate() method.
428 	 *
429 	 * There is no guarantee that the method will be called again
430 	 * if the current fault is for read, and a future fault is
431 	 * for write.  Report the entry's maximum allowed protection
432 	 * to the driver.
433 	 */
434 	rv = vm_pager_populate(fs->first_object, fs->first_pindex,
435 	    fault_type, fs->entry->max_protection, &pager_first, &pager_last);
436 
437 	VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
438 	if (rv == VM_PAGER_BAD) {
439 		/*
440 		 * VM_PAGER_BAD is the backdoor for a pager to request
441 		 * normal fault handling.
442 		 */
443 		vm_fault_restore_map_lock(fs);
444 		if (fs->map->timestamp != fs->map_generation)
445 			return (KERN_RESOURCE_SHORTAGE); /* RetryFault */
446 		return (KERN_NOT_RECEIVER);
447 	}
448 	if (rv != VM_PAGER_OK)
449 		return (KERN_FAILURE); /* AKA SIGSEGV */
450 
451 	/* Ensure that the driver is obeying the interface. */
452 	MPASS(pager_first <= pager_last);
453 	MPASS(fs->first_pindex <= pager_last);
454 	MPASS(fs->first_pindex >= pager_first);
455 	MPASS(pager_last < fs->first_object->size);
456 
457 	vm_fault_restore_map_lock(fs);
458 	if (fs->map->timestamp != fs->map_generation) {
459 		vm_fault_populate_cleanup(fs->first_object, pager_first,
460 		    pager_last);
461 		return (KERN_RESOURCE_SHORTAGE); /* RetryFault */
462 	}
463 
464 	/*
465 	 * The map is unchanged after our last unlock.  Process the fault.
466 	 *
467 	 * The range [pager_first, pager_last] that is given to the
468 	 * pager is only a hint.  The pager may populate any range
469 	 * within the object that includes the requested page index.
470 	 * In case the pager expanded the range, clip it to fit into
471 	 * the map entry.
472 	 */
473 	map_first = OFF_TO_IDX(fs->entry->offset);
474 	if (map_first > pager_first) {
475 		vm_fault_populate_cleanup(fs->first_object, pager_first,
476 		    map_first - 1);
477 		pager_first = map_first;
478 	}
479 	map_last = map_first + atop(fs->entry->end - fs->entry->start) - 1;
480 	if (map_last < pager_last) {
481 		vm_fault_populate_cleanup(fs->first_object, map_last + 1,
482 		    pager_last);
483 		pager_last = map_last;
484 	}
485 	for (pidx = pager_first, m = vm_page_lookup(fs->first_object, pidx);
486 	    pidx <= pager_last;
487 	    pidx += npages, m = vm_page_next(&m[npages - 1])) {
488 		vaddr = fs->entry->start + IDX_TO_OFF(pidx) - fs->entry->offset;
489 #if defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \
490     __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv)
491 		psind = m->psind;
492 		if (psind > 0 && ((vaddr & (pagesizes[psind] - 1)) != 0 ||
493 		    pidx + OFF_TO_IDX(pagesizes[psind]) - 1 > pager_last ||
494 		    !pmap_ps_enabled(fs->map->pmap) || wired))
495 			psind = 0;
496 #else
497 		psind = 0;
498 #endif
499 		npages = atop(pagesizes[psind]);
500 		for (i = 0; i < npages; i++) {
501 			vm_fault_populate_check_page(&m[i]);
502 			vm_fault_dirty(fs->entry, &m[i], prot, fault_type,
503 			    fault_flags);
504 		}
505 		VM_OBJECT_WUNLOCK(fs->first_object);
506 		rv = pmap_enter(fs->map->pmap, vaddr, m, prot, fault_type |
507 		    (wired ? PMAP_ENTER_WIRED : 0), psind);
508 #if defined(__amd64__)
509 		if (psind > 0 && rv == KERN_FAILURE) {
510 			for (i = 0; i < npages; i++) {
511 				rv = pmap_enter(fs->map->pmap, vaddr + ptoa(i),
512 				    &m[i], prot, fault_type |
513 				    (wired ? PMAP_ENTER_WIRED : 0), 0);
514 				MPASS(rv == KERN_SUCCESS);
515 			}
516 		}
517 #else
518 		MPASS(rv == KERN_SUCCESS);
519 #endif
520 		VM_OBJECT_WLOCK(fs->first_object);
521 		m_mtx = NULL;
522 		for (i = 0; i < npages; i++) {
523 			if ((fault_flags & VM_FAULT_WIRE) != 0) {
524 				vm_page_wire(&m[i]);
525 			} else {
526 				vm_page_change_lock(&m[i], &m_mtx);
527 				vm_page_activate(&m[i]);
528 			}
529 			if (m_hold != NULL && m[i].pindex == fs->first_pindex) {
530 				*m_hold = &m[i];
531 				vm_page_wire(&m[i]);
532 			}
533 			vm_page_xunbusy(&m[i]);
534 		}
535 		if (m_mtx != NULL)
536 			mtx_unlock(m_mtx);
537 	}
538 	curthread->td_ru.ru_majflt++;
539 	return (KERN_SUCCESS);
540 }
541 
542 static int prot_fault_translation;
543 SYSCTL_INT(_machdep, OID_AUTO, prot_fault_translation, CTLFLAG_RWTUN,
544     &prot_fault_translation, 0,
545     "Control signal to deliver on protection fault");
546 
547 /* compat definition to keep common code for signal translation */
548 #define	UCODE_PAGEFLT	12
549 #ifdef T_PAGEFLT
550 _Static_assert(UCODE_PAGEFLT == T_PAGEFLT, "T_PAGEFLT");
551 #endif
552 
553 /*
554  *	vm_fault_trap:
555  *
556  *	Handle a page fault occurring at the given address,
557  *	requiring the given permissions, in the map specified.
558  *	If successful, the page is inserted into the
559  *	associated physical map.
560  *
561  *	NOTE: the given address should be truncated to the
562  *	proper page address.
563  *
564  *	KERN_SUCCESS is returned if the page fault is handled; otherwise,
565  *	a standard error specifying why the fault is fatal is returned.
566  *
567  *	The map in question must be referenced, and remains so.
568  *	Caller may hold no locks.
569  */
570 int
571 vm_fault_trap(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
572     int fault_flags, int *signo, int *ucode)
573 {
574 	int result;
575 
576 	MPASS(signo == NULL || ucode != NULL);
577 #ifdef KTRACE
578 	if (map != kernel_map && KTRPOINT(curthread, KTR_FAULT))
579 		ktrfault(vaddr, fault_type);
580 #endif
581 	result = vm_fault(map, trunc_page(vaddr), fault_type, fault_flags,
582 	    NULL);
583 	KASSERT(result == KERN_SUCCESS || result == KERN_FAILURE ||
584 	    result == KERN_INVALID_ADDRESS ||
585 	    result == KERN_RESOURCE_SHORTAGE ||
586 	    result == KERN_PROTECTION_FAILURE ||
587 	    result == KERN_OUT_OF_BOUNDS,
588 	    ("Unexpected Mach error %d from vm_fault()", result));
589 #ifdef KTRACE
590 	if (map != kernel_map && KTRPOINT(curthread, KTR_FAULTEND))
591 		ktrfaultend(result);
592 #endif
593 	if (result != KERN_SUCCESS && signo != NULL) {
594 		switch (result) {
595 		case KERN_FAILURE:
596 		case KERN_INVALID_ADDRESS:
597 			*signo = SIGSEGV;
598 			*ucode = SEGV_MAPERR;
599 			break;
600 		case KERN_RESOURCE_SHORTAGE:
601 			*signo = SIGBUS;
602 			*ucode = BUS_OOMERR;
603 			break;
604 		case KERN_OUT_OF_BOUNDS:
605 			*signo = SIGBUS;
606 			*ucode = BUS_OBJERR;
607 			break;
608 		case KERN_PROTECTION_FAILURE:
609 			if (prot_fault_translation == 0) {
610 				/*
611 				 * Autodetect.  This check also covers
612 				 * the images without the ABI-tag ELF
613 				 * note.
614 				 */
615 				if (SV_CURPROC_ABI() == SV_ABI_FREEBSD &&
616 				    curproc->p_osrel >= P_OSREL_SIGSEGV) {
617 					*signo = SIGSEGV;
618 					*ucode = SEGV_ACCERR;
619 				} else {
620 					*signo = SIGBUS;
621 					*ucode = UCODE_PAGEFLT;
622 				}
623 			} else if (prot_fault_translation == 1) {
624 				/* Always compat mode. */
625 				*signo = SIGBUS;
626 				*ucode = UCODE_PAGEFLT;
627 			} else {
628 				/* Always SIGSEGV mode. */
629 				*signo = SIGSEGV;
630 				*ucode = SEGV_ACCERR;
631 			}
632 			break;
633 		default:
634 			KASSERT(0, ("Unexpected Mach error %d from vm_fault()",
635 			    result));
636 			break;
637 		}
638 	}
639 	return (result);
640 }
641 
642 static int
643 vm_fault_lock_vnode(struct faultstate *fs)
644 {
645 	struct vnode *vp;
646 	int error, locked;
647 
648 	if (fs->object->type != OBJT_VNODE)
649 		return (KERN_SUCCESS);
650 	vp = fs->object->handle;
651 	if (vp == fs->vp) {
652 		ASSERT_VOP_LOCKED(vp, "saved vnode is not locked");
653 		return (KERN_SUCCESS);
654 	}
655 
656 	/*
657 	 * Perform an unlock in case the desired vnode changed while
658 	 * the map was unlocked during a retry.
659 	 */
660 	unlock_vp(fs);
661 
662 	locked = VOP_ISLOCKED(vp);
663 	if (locked != LK_EXCLUSIVE)
664 		locked = LK_SHARED;
665 
666 	/*
667 	 * We must not sleep acquiring the vnode lock while we have
668 	 * the page exclusive busied or the object's
669 	 * paging-in-progress count incremented.  Otherwise, we could
670 	 * deadlock.
671 	 */
672 	error = vget(vp, locked | LK_CANRECURSE | LK_NOWAIT, curthread);
673 	if (error == 0) {
674 		fs->vp = vp;
675 		return (KERN_SUCCESS);
676 	}
677 
678 	vhold(vp);
679 	unlock_and_deallocate(fs);
680 	error = vget(vp, locked | LK_RETRY | LK_CANRECURSE, curthread);
681 	vdrop(vp);
682 	fs->vp = vp;
683 	KASSERT(error == 0, ("vm_fault: vget failed %d", error));
684 	return (KERN_RESOURCE_SHORTAGE);
685 }
686 
687 /*
688  * Wait/Retry if the page is busy.  We have to do this if the page is
689  * either exclusive or shared busy because the vm_pager may be using
690  * read busy for pageouts (and even pageins if it is the vnode pager),
691  * and we could end up trying to pagein and pageout the same page
692  * simultaneously.
693  *
694  * We can theoretically allow the busy case on a read fault if the page
695  * is marked valid, but since such pages are typically already pmap'd,
696  * putting that special case in might be more effort then it is worth.
697  * We cannot under any circumstances mess around with a shared busied
698  * page except, perhaps, to pmap it.
699  */
700 static void
701 vm_fault_busy_sleep(struct faultstate *fs)
702 {
703 	/*
704 	 * Reference the page before unlocking and
705 	 * sleeping so that the page daemon is less
706 	 * likely to reclaim it.
707 	 */
708 	vm_page_aflag_set(fs->m, PGA_REFERENCED);
709 	if (fs->object != fs->first_object) {
710 		fault_page_release(&fs->first_m);
711 		vm_object_pip_wakeup(fs->first_object);
712 	}
713 	vm_object_pip_wakeup(fs->object);
714 	unlock_map(fs);
715 	if (fs->m == vm_page_lookup(fs->object, fs->pindex))
716 		vm_page_busy_sleep(fs->m, "vmpfw", false);
717 	else
718 		VM_OBJECT_WUNLOCK(fs->object);
719 	VM_CNT_INC(v_intrans);
720 	vm_object_deallocate(fs->first_object);
721 }
722 
723 int
724 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
725     int fault_flags, vm_page_t *m_hold)
726 {
727 	struct faultstate fs;
728 	struct domainset *dset;
729 	vm_object_t next_object, retry_object;
730 	vm_offset_t e_end, e_start;
731 	vm_pindex_t retry_pindex;
732 	vm_prot_t prot, retry_prot;
733 	int ahead, alloc_req, behind, cluster_offset, era, faultcount;
734 	int nera, oom, result, rv;
735 	u_char behavior;
736 	boolean_t wired;	/* Passed by reference. */
737 	bool dead, hardfault, is_first_object_locked;
738 
739 	VM_CNT_INC(v_vm_faults);
740 
741 	if ((curthread->td_pflags & TDP_NOFAULTING) != 0)
742 		return (KERN_PROTECTION_FAILURE);
743 
744 	fs.vp = NULL;
745 	faultcount = 0;
746 	nera = -1;
747 	hardfault = false;
748 
749 RetryFault:
750 	oom = 0;
751 RetryFault_oom:
752 
753 	/*
754 	 * Find the backing store object and offset into it to begin the
755 	 * search.
756 	 */
757 	fs.map = map;
758 	result = vm_map_lookup(&fs.map, vaddr, fault_type |
759 	    VM_PROT_FAULT_LOOKUP, &fs.entry, &fs.first_object,
760 	    &fs.first_pindex, &prot, &wired);
761 	if (result != KERN_SUCCESS) {
762 		unlock_vp(&fs);
763 		return (result);
764 	}
765 
766 	fs.map_generation = fs.map->timestamp;
767 
768 	if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
769 		panic("%s: fault on nofault entry, addr: %#lx",
770 		    __func__, (u_long)vaddr);
771 	}
772 
773 	if (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION &&
774 	    fs.entry->wiring_thread != curthread) {
775 		vm_map_unlock_read(fs.map);
776 		vm_map_lock(fs.map);
777 		if (vm_map_lookup_entry(fs.map, vaddr, &fs.entry) &&
778 		    (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION)) {
779 			unlock_vp(&fs);
780 			fs.entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
781 			vm_map_unlock_and_wait(fs.map, 0);
782 		} else
783 			vm_map_unlock(fs.map);
784 		goto RetryFault;
785 	}
786 
787 	MPASS((fs.entry->eflags & MAP_ENTRY_GUARD) == 0);
788 
789 	if (wired)
790 		fault_type = prot | (fault_type & VM_PROT_COPY);
791 	else
792 		KASSERT((fault_flags & VM_FAULT_WIRE) == 0,
793 		    ("!wired && VM_FAULT_WIRE"));
794 
795 	/*
796 	 * Try to avoid lock contention on the top-level object through
797 	 * special-case handling of some types of page faults, specifically,
798 	 * those that are mapping an existing page from the top-level object.
799 	 * Under this condition, a read lock on the object suffices, allowing
800 	 * multiple page faults of a similar type to run in parallel.
801 	 */
802 	if (fs.vp == NULL /* avoid locked vnode leak */ &&
803 	    (fault_flags & (VM_FAULT_WIRE | VM_FAULT_DIRTY)) == 0) {
804 		VM_OBJECT_RLOCK(fs.first_object);
805 		rv = vm_fault_soft_fast(&fs, vaddr, prot, fault_type,
806 		    fault_flags, wired, m_hold);
807 		if (rv == KERN_SUCCESS)
808 			return (rv);
809 		if (!VM_OBJECT_TRYUPGRADE(fs.first_object)) {
810 			VM_OBJECT_RUNLOCK(fs.first_object);
811 			VM_OBJECT_WLOCK(fs.first_object);
812 		}
813 	} else {
814 		VM_OBJECT_WLOCK(fs.first_object);
815 	}
816 
817 	/*
818 	 * Make a reference to this object to prevent its disposal while we
819 	 * are messing with it.  Once we have the reference, the map is free
820 	 * to be diddled.  Since objects reference their shadows (and copies),
821 	 * they will stay around as well.
822 	 *
823 	 * Bump the paging-in-progress count to prevent size changes (e.g.
824 	 * truncation operations) during I/O.
825 	 */
826 	vm_object_reference_locked(fs.first_object);
827 	vm_object_pip_add(fs.first_object, 1);
828 
829 	fs.lookup_still_valid = true;
830 
831 	fs.m = fs.first_m = NULL;
832 
833 	/*
834 	 * Search for the page at object/offset.
835 	 */
836 	fs.object = fs.first_object;
837 	fs.pindex = fs.first_pindex;
838 	while (TRUE) {
839 		KASSERT(fs.m == NULL,
840 		    ("page still set %p at loop start", fs.m));
841 		/*
842 		 * If the object is marked for imminent termination,
843 		 * we retry here, since the collapse pass has raced
844 		 * with us.  Otherwise, if we see terminally dead
845 		 * object, return fail.
846 		 */
847 		if ((fs.object->flags & OBJ_DEAD) != 0) {
848 			dead = fs.object->type == OBJT_DEAD;
849 			unlock_and_deallocate(&fs);
850 			if (dead)
851 				return (KERN_PROTECTION_FAILURE);
852 			pause("vmf_de", 1);
853 			goto RetryFault;
854 		}
855 
856 		/*
857 		 * See if page is resident
858 		 */
859 		fs.m = vm_page_lookup(fs.object, fs.pindex);
860 		if (fs.m != NULL) {
861 			if (vm_page_tryxbusy(fs.m) == 0) {
862 				vm_fault_busy_sleep(&fs);
863 				goto RetryFault;
864 			}
865 
866 			/*
867 			 * The page is marked busy for other processes and the
868 			 * pagedaemon.  If it still isn't completely valid
869 			 * (readable), jump to readrest, else break-out ( we
870 			 * found the page ).
871 			 */
872 			if (!vm_page_all_valid(fs.m))
873 				goto readrest;
874 			break; /* break to PAGE HAS BEEN FOUND */
875 		}
876 		KASSERT(fs.m == NULL, ("fs.m should be NULL, not %p", fs.m));
877 
878 		/*
879 		 * Page is not resident.  If the pager might contain the page
880 		 * or this is the beginning of the search, allocate a new
881 		 * page.  (Default objects are zero-fill, so there is no real
882 		 * pager for them.)
883 		 */
884 		if (fs.object->type != OBJT_DEFAULT ||
885 		    fs.object == fs.first_object) {
886 			if ((fs.object->flags & OBJ_SIZEVNLOCK) != 0) {
887 				rv = vm_fault_lock_vnode(&fs);
888 				MPASS(rv == KERN_SUCCESS ||
889 				    rv == KERN_RESOURCE_SHORTAGE);
890 				if (rv == KERN_RESOURCE_SHORTAGE)
891 					goto RetryFault;
892 			}
893 			if (fs.pindex >= fs.object->size) {
894 				unlock_and_deallocate(&fs);
895 				return (KERN_OUT_OF_BOUNDS);
896 			}
897 
898 			if (fs.object == fs.first_object &&
899 			    (fs.first_object->flags & OBJ_POPULATE) != 0 &&
900 			    fs.first_object->shadow_count == 0) {
901 				rv = vm_fault_populate(&fs, prot, fault_type,
902 				    fault_flags, wired, m_hold);
903 				switch (rv) {
904 				case KERN_SUCCESS:
905 				case KERN_FAILURE:
906 					unlock_and_deallocate(&fs);
907 					return (rv);
908 				case KERN_RESOURCE_SHORTAGE:
909 					unlock_and_deallocate(&fs);
910 					goto RetryFault;
911 				case KERN_NOT_RECEIVER:
912 					/*
913 					 * Pager's populate() method
914 					 * returned VM_PAGER_BAD.
915 					 */
916 					break;
917 				default:
918 					panic("inconsistent return codes");
919 				}
920 			}
921 
922 			/*
923 			 * Allocate a new page for this object/offset pair.
924 			 *
925 			 * Unlocked read of the p_flag is harmless. At
926 			 * worst, the P_KILLED might be not observed
927 			 * there, and allocation can fail, causing
928 			 * restart and new reading of the p_flag.
929 			 */
930 			dset = fs.object->domain.dr_policy;
931 			if (dset == NULL)
932 				dset = curthread->td_domain.dr_policy;
933 			if (!vm_page_count_severe_set(&dset->ds_mask) ||
934 			    P_KILLED(curproc)) {
935 #if VM_NRESERVLEVEL > 0
936 				vm_object_color(fs.object, atop(vaddr) -
937 				    fs.pindex);
938 #endif
939 				alloc_req = P_KILLED(curproc) ?
940 				    VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL;
941 				if (fs.object->type != OBJT_VNODE &&
942 				    fs.object->backing_object == NULL)
943 					alloc_req |= VM_ALLOC_ZERO;
944 				fs.m = vm_page_alloc(fs.object, fs.pindex,
945 				    alloc_req);
946 			}
947 			if (fs.m == NULL) {
948 				unlock_and_deallocate(&fs);
949 				if (vm_pfault_oom_attempts < 0 ||
950 				    oom < vm_pfault_oom_attempts) {
951 					oom++;
952 					vm_waitpfault(dset,
953 					    vm_pfault_oom_wait * hz);
954 					goto RetryFault_oom;
955 				}
956 				if (bootverbose)
957 					printf(
958 	"proc %d (%s) failed to alloc page on fault, starting OOM\n",
959 					    curproc->p_pid, curproc->p_comm);
960 				vm_pageout_oom(VM_OOM_MEM_PF);
961 				goto RetryFault;
962 			}
963 		}
964 
965 readrest:
966 		/*
967 		 * At this point, we have either allocated a new page or found
968 		 * an existing page that is only partially valid.
969 		 *
970 		 * We hold a reference on the current object and the page is
971 		 * exclusive busied.
972 		 */
973 
974 		/*
975 		 * If the pager for the current object might have the page,
976 		 * then determine the number of additional pages to read and
977 		 * potentially reprioritize previously read pages for earlier
978 		 * reclamation.  These operations should only be performed
979 		 * once per page fault.  Even if the current pager doesn't
980 		 * have the page, the number of additional pages to read will
981 		 * apply to subsequent objects in the shadow chain.
982 		 */
983 		if (fs.object->type != OBJT_DEFAULT && nera == -1 &&
984 		    !P_KILLED(curproc)) {
985 			KASSERT(fs.lookup_still_valid, ("map unlocked"));
986 			era = fs.entry->read_ahead;
987 			behavior = vm_map_entry_behavior(fs.entry);
988 			if (behavior == MAP_ENTRY_BEHAV_RANDOM) {
989 				nera = 0;
990 			} else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) {
991 				nera = VM_FAULT_READ_AHEAD_MAX;
992 				if (vaddr == fs.entry->next_read)
993 					vm_fault_dontneed(&fs, vaddr, nera);
994 			} else if (vaddr == fs.entry->next_read) {
995 				/*
996 				 * This is a sequential fault.  Arithmetically
997 				 * increase the requested number of pages in
998 				 * the read-ahead window.  The requested
999 				 * number of pages is "# of sequential faults
1000 				 * x (read ahead min + 1) + read ahead min"
1001 				 */
1002 				nera = VM_FAULT_READ_AHEAD_MIN;
1003 				if (era > 0) {
1004 					nera += era + 1;
1005 					if (nera > VM_FAULT_READ_AHEAD_MAX)
1006 						nera = VM_FAULT_READ_AHEAD_MAX;
1007 				}
1008 				if (era == VM_FAULT_READ_AHEAD_MAX)
1009 					vm_fault_dontneed(&fs, vaddr, nera);
1010 			} else {
1011 				/*
1012 				 * This is a non-sequential fault.
1013 				 */
1014 				nera = 0;
1015 			}
1016 			if (era != nera) {
1017 				/*
1018 				 * A read lock on the map suffices to update
1019 				 * the read ahead count safely.
1020 				 */
1021 				fs.entry->read_ahead = nera;
1022 			}
1023 
1024 			/*
1025 			 * Prepare for unlocking the map.  Save the map
1026 			 * entry's start and end addresses, which are used to
1027 			 * optimize the size of the pager operation below.
1028 			 * Even if the map entry's addresses change after
1029 			 * unlocking the map, using the saved addresses is
1030 			 * safe.
1031 			 */
1032 			e_start = fs.entry->start;
1033 			e_end = fs.entry->end;
1034 		}
1035 
1036 		/*
1037 		 * Call the pager to retrieve the page if there is a chance
1038 		 * that the pager has it, and potentially retrieve additional
1039 		 * pages at the same time.
1040 		 */
1041 		if (fs.object->type != OBJT_DEFAULT) {
1042 			/*
1043 			 * Release the map lock before locking the vnode or
1044 			 * sleeping in the pager.  (If the current object has
1045 			 * a shadow, then an earlier iteration of this loop
1046 			 * may have already unlocked the map.)
1047 			 */
1048 			unlock_map(&fs);
1049 
1050 			rv = vm_fault_lock_vnode(&fs);
1051 			MPASS(rv == KERN_SUCCESS ||
1052 			    rv == KERN_RESOURCE_SHORTAGE);
1053 			if (rv == KERN_RESOURCE_SHORTAGE)
1054 				goto RetryFault;
1055 			KASSERT(fs.vp == NULL || !fs.map->system_map,
1056 			    ("vm_fault: vnode-backed object mapped by system map"));
1057 
1058 			/*
1059 			 * Page in the requested page and hint the pager,
1060 			 * that it may bring up surrounding pages.
1061 			 */
1062 			if (nera == -1 || behavior == MAP_ENTRY_BEHAV_RANDOM ||
1063 			    P_KILLED(curproc)) {
1064 				behind = 0;
1065 				ahead = 0;
1066 			} else {
1067 				/* Is this a sequential fault? */
1068 				if (nera > 0) {
1069 					behind = 0;
1070 					ahead = nera;
1071 				} else {
1072 					/*
1073 					 * Request a cluster of pages that is
1074 					 * aligned to a VM_FAULT_READ_DEFAULT
1075 					 * page offset boundary within the
1076 					 * object.  Alignment to a page offset
1077 					 * boundary is more likely to coincide
1078 					 * with the underlying file system
1079 					 * block than alignment to a virtual
1080 					 * address boundary.
1081 					 */
1082 					cluster_offset = fs.pindex %
1083 					    VM_FAULT_READ_DEFAULT;
1084 					behind = ulmin(cluster_offset,
1085 					    atop(vaddr - e_start));
1086 					ahead = VM_FAULT_READ_DEFAULT - 1 -
1087 					    cluster_offset;
1088 				}
1089 				ahead = ulmin(ahead, atop(e_end - vaddr) - 1);
1090 			}
1091 			rv = vm_pager_get_pages(fs.object, &fs.m, 1,
1092 			    &behind, &ahead);
1093 			if (rv == VM_PAGER_OK) {
1094 				faultcount = behind + 1 + ahead;
1095 				hardfault = true;
1096 				break; /* break to PAGE HAS BEEN FOUND */
1097 			}
1098 			if (rv == VM_PAGER_ERROR)
1099 				printf("vm_fault: pager read error, pid %d (%s)\n",
1100 				    curproc->p_pid, curproc->p_comm);
1101 
1102 			/*
1103 			 * If an I/O error occurred or the requested page was
1104 			 * outside the range of the pager, clean up and return
1105 			 * an error.
1106 			 */
1107 			if (rv == VM_PAGER_ERROR || rv == VM_PAGER_BAD) {
1108 				fault_page_free(&fs.m);
1109 				unlock_and_deallocate(&fs);
1110 				return (KERN_OUT_OF_BOUNDS);
1111 			}
1112 
1113 		}
1114 
1115 		/*
1116 		 * The requested page does not exist at this object/
1117 		 * offset.  Remove the invalid page from the object,
1118 		 * waking up anyone waiting for it, and continue on to
1119 		 * the next object.  However, if this is the top-level
1120 		 * object, we must leave the busy page in place to
1121 		 * prevent another process from rushing past us, and
1122 		 * inserting the page in that object at the same time
1123 		 * that we are.
1124 		 */
1125 		if (fs.object == fs.first_object) {
1126 			fs.first_m = fs.m;
1127 			fs.m = NULL;
1128 		} else
1129 			fault_page_free(&fs.m);
1130 
1131 		/*
1132 		 * Move on to the next object.  Lock the next object before
1133 		 * unlocking the current one.
1134 		 */
1135 		next_object = fs.object->backing_object;
1136 		if (next_object == NULL) {
1137 			/*
1138 			 * If there's no object left, fill the page in the top
1139 			 * object with zeros.
1140 			 */
1141 			if (fs.object != fs.first_object) {
1142 				vm_object_pip_wakeup(fs.object);
1143 				VM_OBJECT_WUNLOCK(fs.object);
1144 
1145 				fs.object = fs.first_object;
1146 				fs.pindex = fs.first_pindex;
1147 				VM_OBJECT_WLOCK(fs.object);
1148 			}
1149 			MPASS(fs.first_m != NULL);
1150 			MPASS(fs.m == NULL);
1151 			fs.m = fs.first_m;
1152 			fs.first_m = NULL;
1153 
1154 			/*
1155 			 * Zero the page if necessary and mark it valid.
1156 			 */
1157 			if ((fs.m->flags & PG_ZERO) == 0) {
1158 				pmap_zero_page(fs.m);
1159 			} else {
1160 				VM_CNT_INC(v_ozfod);
1161 			}
1162 			VM_CNT_INC(v_zfod);
1163 			vm_page_valid(fs.m);
1164 			/* Don't try to prefault neighboring pages. */
1165 			faultcount = 1;
1166 			break;	/* break to PAGE HAS BEEN FOUND */
1167 		} else {
1168 			MPASS(fs.first_m != NULL);
1169 			KASSERT(fs.object != next_object,
1170 			    ("object loop %p", next_object));
1171 			VM_OBJECT_WLOCK(next_object);
1172 			vm_object_pip_add(next_object, 1);
1173 			if (fs.object != fs.first_object)
1174 				vm_object_pip_wakeup(fs.object);
1175 			fs.pindex +=
1176 			    OFF_TO_IDX(fs.object->backing_object_offset);
1177 			VM_OBJECT_WUNLOCK(fs.object);
1178 			fs.object = next_object;
1179 		}
1180 	}
1181 
1182 	vm_page_assert_xbusied(fs.m);
1183 
1184 	/*
1185 	 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
1186 	 * is held.]
1187 	 */
1188 
1189 	/*
1190 	 * If the page is being written, but isn't already owned by the
1191 	 * top-level object, we have to copy it into a new page owned by the
1192 	 * top-level object.
1193 	 */
1194 	if (fs.object != fs.first_object) {
1195 		/*
1196 		 * We only really need to copy if we want to write it.
1197 		 */
1198 		if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1199 			/*
1200 			 * This allows pages to be virtually copied from a
1201 			 * backing_object into the first_object, where the
1202 			 * backing object has no other refs to it, and cannot
1203 			 * gain any more refs.  Instead of a bcopy, we just
1204 			 * move the page from the backing object to the
1205 			 * first object.  Note that we must mark the page
1206 			 * dirty in the first object so that it will go out
1207 			 * to swap when needed.
1208 			 */
1209 			is_first_object_locked = false;
1210 			if (
1211 				/*
1212 				 * Only one shadow object
1213 				 */
1214 				(fs.object->shadow_count == 1) &&
1215 				/*
1216 				 * No COW refs, except us
1217 				 */
1218 				(fs.object->ref_count == 1) &&
1219 				/*
1220 				 * No one else can look this object up
1221 				 */
1222 				(fs.object->handle == NULL) &&
1223 				/*
1224 				 * No other ways to look the object up
1225 				 */
1226 				((fs.object->flags & OBJ_ANON) != 0) &&
1227 			    (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs.first_object)) &&
1228 				/*
1229 				 * We don't chase down the shadow chain
1230 				 */
1231 			    fs.object == fs.first_object->backing_object) {
1232 
1233 				/*
1234 				 * Remove but keep xbusy for replace.  fs.m is
1235 				 * moved into fs.first_object and left busy
1236 				 * while fs.first_m is conditionally freed.
1237 				 */
1238 				vm_page_remove_xbusy(fs.m);
1239 				vm_page_replace(fs.m, fs.first_object,
1240 				    fs.first_pindex, fs.first_m);
1241 				vm_page_dirty(fs.m);
1242 #if VM_NRESERVLEVEL > 0
1243 				/*
1244 				 * Rename the reservation.
1245 				 */
1246 				vm_reserv_rename(fs.m, fs.first_object,
1247 				    fs.object, OFF_TO_IDX(
1248 				    fs.first_object->backing_object_offset));
1249 #endif
1250 				VM_OBJECT_WUNLOCK(fs.object);
1251 				fs.first_m = fs.m;
1252 				fs.m = NULL;
1253 				VM_CNT_INC(v_cow_optim);
1254 			} else {
1255 				VM_OBJECT_WUNLOCK(fs.object);
1256 				/*
1257 				 * Oh, well, lets copy it.
1258 				 */
1259 				pmap_copy_page(fs.m, fs.first_m);
1260 				vm_page_valid(fs.first_m);
1261 				if (wired && (fault_flags &
1262 				    VM_FAULT_WIRE) == 0) {
1263 					vm_page_wire(fs.first_m);
1264 					vm_page_unwire(fs.m, PQ_INACTIVE);
1265 				}
1266 				/*
1267 				 * We no longer need the old page or object.
1268 				 */
1269 				fault_page_release(&fs.m);
1270 			}
1271 			/*
1272 			 * fs.object != fs.first_object due to above
1273 			 * conditional
1274 			 */
1275 			vm_object_pip_wakeup(fs.object);
1276 
1277 			/*
1278 			 * We only try to prefault read-only mappings to the
1279 			 * neighboring pages when this copy-on-write fault is
1280 			 * a hard fault.  In other cases, trying to prefault
1281 			 * is typically wasted effort.
1282 			 */
1283 			if (faultcount == 0)
1284 				faultcount = 1;
1285 
1286 			/*
1287 			 * Only use the new page below...
1288 			 */
1289 			fs.object = fs.first_object;
1290 			fs.pindex = fs.first_pindex;
1291 			fs.m = fs.first_m;
1292 			if (!is_first_object_locked)
1293 				VM_OBJECT_WLOCK(fs.object);
1294 			VM_CNT_INC(v_cow_faults);
1295 			curthread->td_cow++;
1296 		} else {
1297 			prot &= ~VM_PROT_WRITE;
1298 		}
1299 	}
1300 
1301 	/*
1302 	 * We must verify that the maps have not changed since our last
1303 	 * lookup.
1304 	 */
1305 	if (!fs.lookup_still_valid) {
1306 		if (!vm_map_trylock_read(fs.map)) {
1307 			unlock_and_deallocate(&fs);
1308 			goto RetryFault;
1309 		}
1310 		fs.lookup_still_valid = true;
1311 		if (fs.map->timestamp != fs.map_generation) {
1312 			result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
1313 			    &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
1314 
1315 			/*
1316 			 * If we don't need the page any longer, put it on the inactive
1317 			 * list (the easiest thing to do here).  If no one needs it,
1318 			 * pageout will grab it eventually.
1319 			 */
1320 			if (result != KERN_SUCCESS) {
1321 				unlock_and_deallocate(&fs);
1322 
1323 				/*
1324 				 * If retry of map lookup would have blocked then
1325 				 * retry fault from start.
1326 				 */
1327 				if (result == KERN_FAILURE)
1328 					goto RetryFault;
1329 				return (result);
1330 			}
1331 			if ((retry_object != fs.first_object) ||
1332 			    (retry_pindex != fs.first_pindex)) {
1333 				unlock_and_deallocate(&fs);
1334 				goto RetryFault;
1335 			}
1336 
1337 			/*
1338 			 * Check whether the protection has changed or the object has
1339 			 * been copied while we left the map unlocked. Changing from
1340 			 * read to write permission is OK - we leave the page
1341 			 * write-protected, and catch the write fault. Changing from
1342 			 * write to read permission means that we can't mark the page
1343 			 * write-enabled after all.
1344 			 */
1345 			prot &= retry_prot;
1346 			fault_type &= retry_prot;
1347 			if (prot == 0) {
1348 				unlock_and_deallocate(&fs);
1349 				goto RetryFault;
1350 			}
1351 
1352 			/* Reassert because wired may have changed. */
1353 			KASSERT(wired || (fault_flags & VM_FAULT_WIRE) == 0,
1354 			    ("!wired && VM_FAULT_WIRE"));
1355 		}
1356 	}
1357 
1358 	/*
1359 	 * If the page was filled by a pager, save the virtual address that
1360 	 * should be faulted on next under a sequential access pattern to the
1361 	 * map entry.  A read lock on the map suffices to update this address
1362 	 * safely.
1363 	 */
1364 	if (hardfault)
1365 		fs.entry->next_read = vaddr + ptoa(ahead) + PAGE_SIZE;
1366 
1367 	vm_page_assert_xbusied(fs.m);
1368 	vm_fault_dirty(fs.entry, fs.m, prot, fault_type, fault_flags);
1369 
1370 	/*
1371 	 * Page must be completely valid or it is not fit to
1372 	 * map into user space.  vm_pager_get_pages() ensures this.
1373 	 */
1374 	KASSERT(vm_page_all_valid(fs.m),
1375 	    ("vm_fault: page %p partially invalid", fs.m));
1376 	VM_OBJECT_WUNLOCK(fs.object);
1377 
1378 	/*
1379 	 * Put this page into the physical map.  We had to do the unlock above
1380 	 * because pmap_enter() may sleep.  We don't put the page
1381 	 * back on the active queue until later so that the pageout daemon
1382 	 * won't find it (yet).
1383 	 */
1384 	pmap_enter(fs.map->pmap, vaddr, fs.m, prot,
1385 	    fault_type | (wired ? PMAP_ENTER_WIRED : 0), 0);
1386 	if (faultcount != 1 && (fault_flags & VM_FAULT_WIRE) == 0 &&
1387 	    wired == 0)
1388 		vm_fault_prefault(&fs, vaddr,
1389 		    faultcount > 0 ? behind : PFBAK,
1390 		    faultcount > 0 ? ahead : PFFOR, false);
1391 
1392 	/*
1393 	 * If the page is not wired down, then put it where the pageout daemon
1394 	 * can find it.
1395 	 */
1396 	if ((fault_flags & VM_FAULT_WIRE) != 0) {
1397 		vm_page_wire(fs.m);
1398 	} else {
1399 		vm_page_lock(fs.m);
1400 		vm_page_activate(fs.m);
1401 		vm_page_unlock(fs.m);
1402 	}
1403 	if (m_hold != NULL) {
1404 		*m_hold = fs.m;
1405 		vm_page_wire(fs.m);
1406 	}
1407 	vm_page_xunbusy(fs.m);
1408 	fs.m = NULL;
1409 
1410 	/*
1411 	 * Unlock everything, and return
1412 	 */
1413 	fault_deallocate(&fs);
1414 	if (hardfault) {
1415 		VM_CNT_INC(v_io_faults);
1416 		curthread->td_ru.ru_majflt++;
1417 #ifdef RACCT
1418 		if (racct_enable && fs.object->type == OBJT_VNODE) {
1419 			PROC_LOCK(curproc);
1420 			if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1421 				racct_add_force(curproc, RACCT_WRITEBPS,
1422 				    PAGE_SIZE + behind * PAGE_SIZE);
1423 				racct_add_force(curproc, RACCT_WRITEIOPS, 1);
1424 			} else {
1425 				racct_add_force(curproc, RACCT_READBPS,
1426 				    PAGE_SIZE + ahead * PAGE_SIZE);
1427 				racct_add_force(curproc, RACCT_READIOPS, 1);
1428 			}
1429 			PROC_UNLOCK(curproc);
1430 		}
1431 #endif
1432 	} else
1433 		curthread->td_ru.ru_minflt++;
1434 
1435 	return (KERN_SUCCESS);
1436 }
1437 
1438 /*
1439  * Speed up the reclamation of pages that precede the faulting pindex within
1440  * the first object of the shadow chain.  Essentially, perform the equivalent
1441  * to madvise(..., MADV_DONTNEED) on a large cluster of pages that precedes
1442  * the faulting pindex by the cluster size when the pages read by vm_fault()
1443  * cross a cluster-size boundary.  The cluster size is the greater of the
1444  * smallest superpage size and VM_FAULT_DONTNEED_MIN.
1445  *
1446  * When "fs->first_object" is a shadow object, the pages in the backing object
1447  * that precede the faulting pindex are deactivated by vm_fault().  So, this
1448  * function must only be concerned with pages in the first object.
1449  */
1450 static void
1451 vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr, int ahead)
1452 {
1453 	vm_map_entry_t entry;
1454 	vm_object_t first_object, object;
1455 	vm_offset_t end, start;
1456 	vm_page_t m, m_next;
1457 	vm_pindex_t pend, pstart;
1458 	vm_size_t size;
1459 
1460 	object = fs->object;
1461 	VM_OBJECT_ASSERT_WLOCKED(object);
1462 	first_object = fs->first_object;
1463 	if (first_object != object) {
1464 		if (!VM_OBJECT_TRYWLOCK(first_object)) {
1465 			VM_OBJECT_WUNLOCK(object);
1466 			VM_OBJECT_WLOCK(first_object);
1467 			VM_OBJECT_WLOCK(object);
1468 		}
1469 	}
1470 	/* Neither fictitious nor unmanaged pages can be reclaimed. */
1471 	if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) {
1472 		size = VM_FAULT_DONTNEED_MIN;
1473 		if (MAXPAGESIZES > 1 && size < pagesizes[1])
1474 			size = pagesizes[1];
1475 		end = rounddown2(vaddr, size);
1476 		if (vaddr - end >= size - PAGE_SIZE - ptoa(ahead) &&
1477 		    (entry = fs->entry)->start < end) {
1478 			if (end - entry->start < size)
1479 				start = entry->start;
1480 			else
1481 				start = end - size;
1482 			pmap_advise(fs->map->pmap, start, end, MADV_DONTNEED);
1483 			pstart = OFF_TO_IDX(entry->offset) + atop(start -
1484 			    entry->start);
1485 			m_next = vm_page_find_least(first_object, pstart);
1486 			pend = OFF_TO_IDX(entry->offset) + atop(end -
1487 			    entry->start);
1488 			while ((m = m_next) != NULL && m->pindex < pend) {
1489 				m_next = TAILQ_NEXT(m, listq);
1490 				if (!vm_page_all_valid(m) ||
1491 				    vm_page_busied(m))
1492 					continue;
1493 
1494 				/*
1495 				 * Don't clear PGA_REFERENCED, since it would
1496 				 * likely represent a reference by a different
1497 				 * process.
1498 				 *
1499 				 * Typically, at this point, prefetched pages
1500 				 * are still in the inactive queue.  Only
1501 				 * pages that triggered page faults are in the
1502 				 * active queue.
1503 				 */
1504 				vm_page_lock(m);
1505 				if (!vm_page_inactive(m))
1506 					vm_page_deactivate(m);
1507 				vm_page_unlock(m);
1508 			}
1509 		}
1510 	}
1511 	if (first_object != object)
1512 		VM_OBJECT_WUNLOCK(first_object);
1513 }
1514 
1515 /*
1516  * vm_fault_prefault provides a quick way of clustering
1517  * pagefaults into a processes address space.  It is a "cousin"
1518  * of vm_map_pmap_enter, except it runs at page fault time instead
1519  * of mmap time.
1520  */
1521 static void
1522 vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
1523     int backward, int forward, bool obj_locked)
1524 {
1525 	pmap_t pmap;
1526 	vm_map_entry_t entry;
1527 	vm_object_t backing_object, lobject;
1528 	vm_offset_t addr, starta;
1529 	vm_pindex_t pindex;
1530 	vm_page_t m;
1531 	int i;
1532 
1533 	pmap = fs->map->pmap;
1534 	if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
1535 		return;
1536 
1537 	entry = fs->entry;
1538 
1539 	if (addra < backward * PAGE_SIZE) {
1540 		starta = entry->start;
1541 	} else {
1542 		starta = addra - backward * PAGE_SIZE;
1543 		if (starta < entry->start)
1544 			starta = entry->start;
1545 	}
1546 
1547 	/*
1548 	 * Generate the sequence of virtual addresses that are candidates for
1549 	 * prefaulting in an outward spiral from the faulting virtual address,
1550 	 * "addra".  Specifically, the sequence is "addra - PAGE_SIZE", "addra
1551 	 * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ...
1552 	 * If the candidate address doesn't have a backing physical page, then
1553 	 * the loop immediately terminates.
1554 	 */
1555 	for (i = 0; i < 2 * imax(backward, forward); i++) {
1556 		addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE :
1557 		    PAGE_SIZE);
1558 		if (addr > addra + forward * PAGE_SIZE)
1559 			addr = 0;
1560 
1561 		if (addr < starta || addr >= entry->end)
1562 			continue;
1563 
1564 		if (!pmap_is_prefaultable(pmap, addr))
1565 			continue;
1566 
1567 		pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1568 		lobject = entry->object.vm_object;
1569 		if (!obj_locked)
1570 			VM_OBJECT_RLOCK(lobject);
1571 		while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1572 		    lobject->type == OBJT_DEFAULT &&
1573 		    (backing_object = lobject->backing_object) != NULL) {
1574 			KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1575 			    0, ("vm_fault_prefault: unaligned object offset"));
1576 			pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1577 			VM_OBJECT_RLOCK(backing_object);
1578 			if (!obj_locked || lobject != entry->object.vm_object)
1579 				VM_OBJECT_RUNLOCK(lobject);
1580 			lobject = backing_object;
1581 		}
1582 		if (m == NULL) {
1583 			if (!obj_locked || lobject != entry->object.vm_object)
1584 				VM_OBJECT_RUNLOCK(lobject);
1585 			break;
1586 		}
1587 		if (vm_page_all_valid(m) &&
1588 		    (m->flags & PG_FICTITIOUS) == 0)
1589 			pmap_enter_quick(pmap, addr, m, entry->protection);
1590 		if (!obj_locked || lobject != entry->object.vm_object)
1591 			VM_OBJECT_RUNLOCK(lobject);
1592 	}
1593 }
1594 
1595 /*
1596  * Hold each of the physical pages that are mapped by the specified range of
1597  * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
1598  * and allow the specified types of access, "prot".  If all of the implied
1599  * pages are successfully held, then the number of held pages is returned
1600  * together with pointers to those pages in the array "ma".  However, if any
1601  * of the pages cannot be held, -1 is returned.
1602  */
1603 int
1604 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
1605     vm_prot_t prot, vm_page_t *ma, int max_count)
1606 {
1607 	vm_offset_t end, va;
1608 	vm_page_t *mp;
1609 	int count;
1610 	boolean_t pmap_failed;
1611 
1612 	if (len == 0)
1613 		return (0);
1614 	end = round_page(addr + len);
1615 	addr = trunc_page(addr);
1616 
1617 	/*
1618 	 * Check for illegal addresses.
1619 	 */
1620 	if (addr < vm_map_min(map) || addr > end || end > vm_map_max(map))
1621 		return (-1);
1622 
1623 	if (atop(end - addr) > max_count)
1624 		panic("vm_fault_quick_hold_pages: count > max_count");
1625 	count = atop(end - addr);
1626 
1627 	/*
1628 	 * Most likely, the physical pages are resident in the pmap, so it is
1629 	 * faster to try pmap_extract_and_hold() first.
1630 	 */
1631 	pmap_failed = FALSE;
1632 	for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) {
1633 		*mp = pmap_extract_and_hold(map->pmap, va, prot);
1634 		if (*mp == NULL)
1635 			pmap_failed = TRUE;
1636 		else if ((prot & VM_PROT_WRITE) != 0 &&
1637 		    (*mp)->dirty != VM_PAGE_BITS_ALL) {
1638 			/*
1639 			 * Explicitly dirty the physical page.  Otherwise, the
1640 			 * caller's changes may go unnoticed because they are
1641 			 * performed through an unmanaged mapping or by a DMA
1642 			 * operation.
1643 			 *
1644 			 * The object lock is not held here.
1645 			 * See vm_page_clear_dirty_mask().
1646 			 */
1647 			vm_page_dirty(*mp);
1648 		}
1649 	}
1650 	if (pmap_failed) {
1651 		/*
1652 		 * One or more pages could not be held by the pmap.  Either no
1653 		 * page was mapped at the specified virtual address or that
1654 		 * mapping had insufficient permissions.  Attempt to fault in
1655 		 * and hold these pages.
1656 		 *
1657 		 * If vm_fault_disable_pagefaults() was called,
1658 		 * i.e., TDP_NOFAULTING is set, we must not sleep nor
1659 		 * acquire MD VM locks, which means we must not call
1660 		 * vm_fault().  Some (out of tree) callers mark
1661 		 * too wide a code area with vm_fault_disable_pagefaults()
1662 		 * already, use the VM_PROT_QUICK_NOFAULT flag to request
1663 		 * the proper behaviour explicitly.
1664 		 */
1665 		if ((prot & VM_PROT_QUICK_NOFAULT) != 0 &&
1666 		    (curthread->td_pflags & TDP_NOFAULTING) != 0)
1667 			goto error;
1668 		for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE)
1669 			if (*mp == NULL && vm_fault(map, va, prot,
1670 			    VM_FAULT_NORMAL, mp) != KERN_SUCCESS)
1671 				goto error;
1672 	}
1673 	return (count);
1674 error:
1675 	for (mp = ma; mp < ma + count; mp++)
1676 		if (*mp != NULL)
1677 			vm_page_unwire(*mp, PQ_INACTIVE);
1678 	return (-1);
1679 }
1680 
1681 /*
1682  *	Routine:
1683  *		vm_fault_copy_entry
1684  *	Function:
1685  *		Create new shadow object backing dst_entry with private copy of
1686  *		all underlying pages. When src_entry is equal to dst_entry,
1687  *		function implements COW for wired-down map entry. Otherwise,
1688  *		it forks wired entry into dst_map.
1689  *
1690  *	In/out conditions:
1691  *		The source and destination maps must be locked for write.
1692  *		The source map entry must be wired down (or be a sharing map
1693  *		entry corresponding to a main map entry that is wired down).
1694  */
1695 void
1696 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1697     vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1698     vm_ooffset_t *fork_charge)
1699 {
1700 	vm_object_t backing_object, dst_object, object, src_object;
1701 	vm_pindex_t dst_pindex, pindex, src_pindex;
1702 	vm_prot_t access, prot;
1703 	vm_offset_t vaddr;
1704 	vm_page_t dst_m;
1705 	vm_page_t src_m;
1706 	boolean_t upgrade;
1707 
1708 #ifdef	lint
1709 	src_map++;
1710 #endif	/* lint */
1711 
1712 	upgrade = src_entry == dst_entry;
1713 	access = prot = dst_entry->protection;
1714 
1715 	src_object = src_entry->object.vm_object;
1716 	src_pindex = OFF_TO_IDX(src_entry->offset);
1717 
1718 	if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
1719 		dst_object = src_object;
1720 		vm_object_reference(dst_object);
1721 	} else {
1722 		/*
1723 		 * Create the top-level object for the destination entry.
1724 		 * Doesn't actually shadow anything - we copy the pages
1725 		 * directly.
1726 		 */
1727 		dst_object = vm_object_allocate_anon(atop(dst_entry->end -
1728 		    dst_entry->start), NULL, NULL, 0);
1729 #if VM_NRESERVLEVEL > 0
1730 		dst_object->flags |= OBJ_COLORED;
1731 		dst_object->pg_color = atop(dst_entry->start);
1732 #endif
1733 		dst_object->domain = src_object->domain;
1734 		dst_object->charge = dst_entry->end - dst_entry->start;
1735 	}
1736 
1737 	VM_OBJECT_WLOCK(dst_object);
1738 	KASSERT(upgrade || dst_entry->object.vm_object == NULL,
1739 	    ("vm_fault_copy_entry: vm_object not NULL"));
1740 	if (src_object != dst_object) {
1741 		dst_entry->object.vm_object = dst_object;
1742 		dst_entry->offset = 0;
1743 		dst_entry->eflags &= ~MAP_ENTRY_VN_EXEC;
1744 	}
1745 	if (fork_charge != NULL) {
1746 		KASSERT(dst_entry->cred == NULL,
1747 		    ("vm_fault_copy_entry: leaked swp charge"));
1748 		dst_object->cred = curthread->td_ucred;
1749 		crhold(dst_object->cred);
1750 		*fork_charge += dst_object->charge;
1751 	} else if ((dst_object->type == OBJT_DEFAULT ||
1752 	    dst_object->type == OBJT_SWAP) &&
1753 	    dst_object->cred == NULL) {
1754 		KASSERT(dst_entry->cred != NULL, ("no cred for entry %p",
1755 		    dst_entry));
1756 		dst_object->cred = dst_entry->cred;
1757 		dst_entry->cred = NULL;
1758 	}
1759 
1760 	/*
1761 	 * If not an upgrade, then enter the mappings in the pmap as
1762 	 * read and/or execute accesses.  Otherwise, enter them as
1763 	 * write accesses.
1764 	 *
1765 	 * A writeable large page mapping is only created if all of
1766 	 * the constituent small page mappings are modified. Marking
1767 	 * PTEs as modified on inception allows promotion to happen
1768 	 * without taking potentially large number of soft faults.
1769 	 */
1770 	if (!upgrade)
1771 		access &= ~VM_PROT_WRITE;
1772 
1773 	/*
1774 	 * Loop through all of the virtual pages within the entry's
1775 	 * range, copying each page from the source object to the
1776 	 * destination object.  Since the source is wired, those pages
1777 	 * must exist.  In contrast, the destination is pageable.
1778 	 * Since the destination object doesn't share any backing storage
1779 	 * with the source object, all of its pages must be dirtied,
1780 	 * regardless of whether they can be written.
1781 	 */
1782 	for (vaddr = dst_entry->start, dst_pindex = 0;
1783 	    vaddr < dst_entry->end;
1784 	    vaddr += PAGE_SIZE, dst_pindex++) {
1785 again:
1786 		/*
1787 		 * Find the page in the source object, and copy it in.
1788 		 * Because the source is wired down, the page will be
1789 		 * in memory.
1790 		 */
1791 		if (src_object != dst_object)
1792 			VM_OBJECT_RLOCK(src_object);
1793 		object = src_object;
1794 		pindex = src_pindex + dst_pindex;
1795 		while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
1796 		    (backing_object = object->backing_object) != NULL) {
1797 			/*
1798 			 * Unless the source mapping is read-only or
1799 			 * it is presently being upgraded from
1800 			 * read-only, the first object in the shadow
1801 			 * chain should provide all of the pages.  In
1802 			 * other words, this loop body should never be
1803 			 * executed when the source mapping is already
1804 			 * read/write.
1805 			 */
1806 			KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 ||
1807 			    upgrade,
1808 			    ("vm_fault_copy_entry: main object missing page"));
1809 
1810 			VM_OBJECT_RLOCK(backing_object);
1811 			pindex += OFF_TO_IDX(object->backing_object_offset);
1812 			if (object != dst_object)
1813 				VM_OBJECT_RUNLOCK(object);
1814 			object = backing_object;
1815 		}
1816 		KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing"));
1817 
1818 		if (object != dst_object) {
1819 			/*
1820 			 * Allocate a page in the destination object.
1821 			 */
1822 			dst_m = vm_page_alloc(dst_object, (src_object ==
1823 			    dst_object ? src_pindex : 0) + dst_pindex,
1824 			    VM_ALLOC_NORMAL);
1825 			if (dst_m == NULL) {
1826 				VM_OBJECT_WUNLOCK(dst_object);
1827 				VM_OBJECT_RUNLOCK(object);
1828 				vm_wait(dst_object);
1829 				VM_OBJECT_WLOCK(dst_object);
1830 				goto again;
1831 			}
1832 			pmap_copy_page(src_m, dst_m);
1833 			VM_OBJECT_RUNLOCK(object);
1834 			dst_m->dirty = dst_m->valid = src_m->valid;
1835 		} else {
1836 			dst_m = src_m;
1837 			if (vm_page_busy_acquire(dst_m, VM_ALLOC_WAITFAIL) == 0)
1838 				goto again;
1839 			if (dst_m->pindex >= dst_object->size) {
1840 				/*
1841 				 * We are upgrading.  Index can occur
1842 				 * out of bounds if the object type is
1843 				 * vnode and the file was truncated.
1844 				 */
1845 				vm_page_xunbusy(dst_m);
1846 				break;
1847 			}
1848 		}
1849 		VM_OBJECT_WUNLOCK(dst_object);
1850 
1851 		/*
1852 		 * Enter it in the pmap. If a wired, copy-on-write
1853 		 * mapping is being replaced by a write-enabled
1854 		 * mapping, then wire that new mapping.
1855 		 *
1856 		 * The page can be invalid if the user called
1857 		 * msync(MS_INVALIDATE) or truncated the backing vnode
1858 		 * or shared memory object.  In this case, do not
1859 		 * insert it into pmap, but still do the copy so that
1860 		 * all copies of the wired map entry have similar
1861 		 * backing pages.
1862 		 */
1863 		if (vm_page_all_valid(dst_m)) {
1864 			pmap_enter(dst_map->pmap, vaddr, dst_m, prot,
1865 			    access | (upgrade ? PMAP_ENTER_WIRED : 0), 0);
1866 		}
1867 
1868 		/*
1869 		 * Mark it no longer busy, and put it on the active list.
1870 		 */
1871 		VM_OBJECT_WLOCK(dst_object);
1872 
1873 		if (upgrade) {
1874 			if (src_m != dst_m) {
1875 				vm_page_unwire(src_m, PQ_INACTIVE);
1876 				vm_page_wire(dst_m);
1877 			} else {
1878 				KASSERT(vm_page_wired(dst_m),
1879 				    ("dst_m %p is not wired", dst_m));
1880 			}
1881 		} else {
1882 			vm_page_lock(dst_m);
1883 			vm_page_activate(dst_m);
1884 			vm_page_unlock(dst_m);
1885 		}
1886 		vm_page_xunbusy(dst_m);
1887 	}
1888 	VM_OBJECT_WUNLOCK(dst_object);
1889 	if (upgrade) {
1890 		dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
1891 		vm_object_deallocate(src_object);
1892 	}
1893 }
1894 
1895 /*
1896  * Block entry into the machine-independent layer's page fault handler by
1897  * the calling thread.  Subsequent calls to vm_fault() by that thread will
1898  * return KERN_PROTECTION_FAILURE.  Enable machine-dependent handling of
1899  * spurious page faults.
1900  */
1901 int
1902 vm_fault_disable_pagefaults(void)
1903 {
1904 
1905 	return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR));
1906 }
1907 
1908 void
1909 vm_fault_enable_pagefaults(int save)
1910 {
1911 
1912 	curthread_pflags_restore(save);
1913 }
1914