xref: /freebsd/sys/vm/vm_fault.c (revision eac7052fdebb90caf2f653e06187bdbca837b9c7)
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 	/* Fault parameters. */
124 	vm_offset_t	vaddr;
125 	vm_page_t	*m_hold;
126 	vm_prot_t	fault_type;
127 	vm_prot_t	prot;
128 	int		fault_flags;
129 	int		oom;
130 	boolean_t	wired;
131 
132 	/* Page reference for cow. */
133 	vm_page_t m_cow;
134 
135 	/* Current object. */
136 	vm_object_t	object;
137 	vm_pindex_t	pindex;
138 	vm_page_t	m;
139 
140 	/* Top-level map object. */
141 	vm_object_t	first_object;
142 	vm_pindex_t	first_pindex;
143 	vm_page_t	first_m;
144 
145 	/* Map state. */
146 	vm_map_t	map;
147 	vm_map_entry_t	entry;
148 	int		map_generation;
149 	bool		lookup_still_valid;
150 
151 	/* Vnode if locked. */
152 	struct vnode	*vp;
153 };
154 
155 static void vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr,
156 	    int ahead);
157 static void vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
158 	    int backward, int forward, bool obj_locked);
159 
160 static int vm_pfault_oom_attempts = 3;
161 SYSCTL_INT(_vm, OID_AUTO, pfault_oom_attempts, CTLFLAG_RWTUN,
162     &vm_pfault_oom_attempts, 0,
163     "Number of page allocation attempts in page fault handler before it "
164     "triggers OOM handling");
165 
166 static int vm_pfault_oom_wait = 10;
167 SYSCTL_INT(_vm, OID_AUTO, pfault_oom_wait, CTLFLAG_RWTUN,
168     &vm_pfault_oom_wait, 0,
169     "Number of seconds to wait for free pages before retrying "
170     "the page fault handler");
171 
172 static inline void
173 fault_page_release(vm_page_t *mp)
174 {
175 	vm_page_t m;
176 
177 	m = *mp;
178 	if (m != NULL) {
179 		/*
180 		 * We are likely to loop around again and attempt to busy
181 		 * this page.  Deactivating it leaves it available for
182 		 * pageout while optimizing fault restarts.
183 		 */
184 		vm_page_deactivate(m);
185 		vm_page_xunbusy(m);
186 		*mp = NULL;
187 	}
188 }
189 
190 static inline void
191 fault_page_free(vm_page_t *mp)
192 {
193 	vm_page_t m;
194 
195 	m = *mp;
196 	if (m != NULL) {
197 		VM_OBJECT_ASSERT_WLOCKED(m->object);
198 		if (!vm_page_wired(m))
199 			vm_page_free(m);
200 		else
201 			vm_page_xunbusy(m);
202 		*mp = NULL;
203 	}
204 }
205 
206 static inline void
207 unlock_map(struct faultstate *fs)
208 {
209 
210 	if (fs->lookup_still_valid) {
211 		vm_map_lookup_done(fs->map, fs->entry);
212 		fs->lookup_still_valid = false;
213 	}
214 }
215 
216 static void
217 unlock_vp(struct faultstate *fs)
218 {
219 
220 	if (fs->vp != NULL) {
221 		vput(fs->vp);
222 		fs->vp = NULL;
223 	}
224 }
225 
226 static void
227 fault_deallocate(struct faultstate *fs)
228 {
229 
230 	fault_page_release(&fs->m_cow);
231 	fault_page_release(&fs->m);
232 	vm_object_pip_wakeup(fs->object);
233 	if (fs->object != fs->first_object) {
234 		VM_OBJECT_WLOCK(fs->first_object);
235 		fault_page_free(&fs->first_m);
236 		VM_OBJECT_WUNLOCK(fs->first_object);
237 		vm_object_pip_wakeup(fs->first_object);
238 	}
239 	vm_object_deallocate(fs->first_object);
240 	unlock_map(fs);
241 	unlock_vp(fs);
242 }
243 
244 static void
245 unlock_and_deallocate(struct faultstate *fs)
246 {
247 
248 	VM_OBJECT_WUNLOCK(fs->object);
249 	fault_deallocate(fs);
250 }
251 
252 static void
253 vm_fault_dirty(struct faultstate *fs, vm_page_t m)
254 {
255 	bool need_dirty;
256 
257 	if (((fs->prot & VM_PROT_WRITE) == 0 &&
258 	    (fs->fault_flags & VM_FAULT_DIRTY) == 0) ||
259 	    (m->oflags & VPO_UNMANAGED) != 0)
260 		return;
261 
262 	VM_PAGE_OBJECT_BUSY_ASSERT(m);
263 
264 	need_dirty = ((fs->fault_type & VM_PROT_WRITE) != 0 &&
265 	    (fs->fault_flags & VM_FAULT_WIRE) == 0) ||
266 	    (fs->fault_flags & VM_FAULT_DIRTY) != 0;
267 
268 	vm_object_set_writeable_dirty(m->object);
269 
270 	/*
271 	 * If the fault is a write, we know that this page is being
272 	 * written NOW so dirty it explicitly to save on
273 	 * pmap_is_modified() calls later.
274 	 *
275 	 * Also, since the page is now dirty, we can possibly tell
276 	 * the pager to release any swap backing the page.
277 	 */
278 	if (need_dirty && vm_page_set_dirty(m) == 0) {
279 		/*
280 		 * If this is a NOSYNC mmap we do not want to set PGA_NOSYNC
281 		 * if the page is already dirty to prevent data written with
282 		 * the expectation of being synced from not being synced.
283 		 * Likewise if this entry does not request NOSYNC then make
284 		 * sure the page isn't marked NOSYNC.  Applications sharing
285 		 * data should use the same flags to avoid ping ponging.
286 		 */
287 		if ((fs->entry->eflags & MAP_ENTRY_NOSYNC) != 0)
288 			vm_page_aflag_set(m, PGA_NOSYNC);
289 		else
290 			vm_page_aflag_clear(m, PGA_NOSYNC);
291 	}
292 
293 }
294 
295 /*
296  * Unlocks fs.first_object and fs.map on success.
297  */
298 static int
299 vm_fault_soft_fast(struct faultstate *fs)
300 {
301 	vm_page_t m, m_map;
302 #if VM_NRESERVLEVEL > 0
303 	vm_page_t m_super;
304 	int flags;
305 #endif
306 	int psind, rv;
307 	vm_offset_t vaddr;
308 
309 	MPASS(fs->vp == NULL);
310 	vaddr = fs->vaddr;
311 	vm_object_busy(fs->first_object);
312 	m = vm_page_lookup(fs->first_object, fs->first_pindex);
313 	/* A busy page can be mapped for read|execute access. */
314 	if (m == NULL || ((fs->prot & VM_PROT_WRITE) != 0 &&
315 	    vm_page_busied(m)) || !vm_page_all_valid(m)) {
316 		rv = KERN_FAILURE;
317 		goto out;
318 	}
319 	m_map = m;
320 	psind = 0;
321 #if VM_NRESERVLEVEL > 0
322 	if ((m->flags & PG_FICTITIOUS) == 0 &&
323 	    (m_super = vm_reserv_to_superpage(m)) != NULL &&
324 	    rounddown2(vaddr, pagesizes[m_super->psind]) >= fs->entry->start &&
325 	    roundup2(vaddr + 1, pagesizes[m_super->psind]) <= fs->entry->end &&
326 	    (vaddr & (pagesizes[m_super->psind] - 1)) == (VM_PAGE_TO_PHYS(m) &
327 	    (pagesizes[m_super->psind] - 1)) && !fs->wired &&
328 	    pmap_ps_enabled(fs->map->pmap)) {
329 		flags = PS_ALL_VALID;
330 		if ((fs->prot & VM_PROT_WRITE) != 0) {
331 			/*
332 			 * Create a superpage mapping allowing write access
333 			 * only if none of the constituent pages are busy and
334 			 * all of them are already dirty (except possibly for
335 			 * the page that was faulted on).
336 			 */
337 			flags |= PS_NONE_BUSY;
338 			if ((fs->first_object->flags & OBJ_UNMANAGED) == 0)
339 				flags |= PS_ALL_DIRTY;
340 		}
341 		if (vm_page_ps_test(m_super, flags, m)) {
342 			m_map = m_super;
343 			psind = m_super->psind;
344 			vaddr = rounddown2(vaddr, pagesizes[psind]);
345 			/* Preset the modified bit for dirty superpages. */
346 			if ((flags & PS_ALL_DIRTY) != 0)
347 				fs->fault_type |= VM_PROT_WRITE;
348 		}
349 	}
350 #endif
351 	rv = pmap_enter(fs->map->pmap, vaddr, m_map, fs->prot, fs->fault_type |
352 	    PMAP_ENTER_NOSLEEP | (fs->wired ? PMAP_ENTER_WIRED : 0), psind);
353 	if (rv != KERN_SUCCESS)
354 		goto out;
355 	if (fs->m_hold != NULL) {
356 		(*fs->m_hold) = m;
357 		vm_page_wire(m);
358 	}
359 	if (psind == 0 && !fs->wired)
360 		vm_fault_prefault(fs, vaddr, PFBAK, PFFOR, true);
361 	VM_OBJECT_RUNLOCK(fs->first_object);
362 	vm_fault_dirty(fs, m);
363 	vm_map_lookup_done(fs->map, fs->entry);
364 	curthread->td_ru.ru_minflt++;
365 
366 out:
367 	vm_object_unbusy(fs->first_object);
368 	return (rv);
369 }
370 
371 static void
372 vm_fault_restore_map_lock(struct faultstate *fs)
373 {
374 
375 	VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
376 	MPASS(blockcount_read(&fs->first_object->paging_in_progress) > 0);
377 
378 	if (!vm_map_trylock_read(fs->map)) {
379 		VM_OBJECT_WUNLOCK(fs->first_object);
380 		vm_map_lock_read(fs->map);
381 		VM_OBJECT_WLOCK(fs->first_object);
382 	}
383 	fs->lookup_still_valid = true;
384 }
385 
386 static void
387 vm_fault_populate_check_page(vm_page_t m)
388 {
389 
390 	/*
391 	 * Check each page to ensure that the pager is obeying the
392 	 * interface: the page must be installed in the object, fully
393 	 * valid, and exclusively busied.
394 	 */
395 	MPASS(m != NULL);
396 	MPASS(vm_page_all_valid(m));
397 	MPASS(vm_page_xbusied(m));
398 }
399 
400 static void
401 vm_fault_populate_cleanup(vm_object_t object, vm_pindex_t first,
402     vm_pindex_t last)
403 {
404 	vm_page_t m;
405 	vm_pindex_t pidx;
406 
407 	VM_OBJECT_ASSERT_WLOCKED(object);
408 	MPASS(first <= last);
409 	for (pidx = first, m = vm_page_lookup(object, pidx);
410 	    pidx <= last; pidx++, m = vm_page_next(m)) {
411 		vm_fault_populate_check_page(m);
412 		vm_page_deactivate(m);
413 		vm_page_xunbusy(m);
414 	}
415 }
416 
417 static int
418 vm_fault_populate(struct faultstate *fs)
419 {
420 	vm_offset_t vaddr;
421 	vm_page_t m;
422 	vm_pindex_t map_first, map_last, pager_first, pager_last, pidx;
423 	int i, npages, psind, rv;
424 
425 	MPASS(fs->object == fs->first_object);
426 	VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
427 	MPASS(blockcount_read(&fs->first_object->paging_in_progress) > 0);
428 	MPASS(fs->first_object->backing_object == NULL);
429 	MPASS(fs->lookup_still_valid);
430 
431 	pager_first = OFF_TO_IDX(fs->entry->offset);
432 	pager_last = pager_first + atop(fs->entry->end - fs->entry->start) - 1;
433 	unlock_map(fs);
434 	unlock_vp(fs);
435 
436 	/*
437 	 * Call the pager (driver) populate() method.
438 	 *
439 	 * There is no guarantee that the method will be called again
440 	 * if the current fault is for read, and a future fault is
441 	 * for write.  Report the entry's maximum allowed protection
442 	 * to the driver.
443 	 */
444 	rv = vm_pager_populate(fs->first_object, fs->first_pindex,
445 	    fs->fault_type, fs->entry->max_protection, &pager_first, &pager_last);
446 
447 	VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
448 	if (rv == VM_PAGER_BAD) {
449 		/*
450 		 * VM_PAGER_BAD is the backdoor for a pager to request
451 		 * normal fault handling.
452 		 */
453 		vm_fault_restore_map_lock(fs);
454 		if (fs->map->timestamp != fs->map_generation)
455 			return (KERN_RESTART);
456 		return (KERN_NOT_RECEIVER);
457 	}
458 	if (rv != VM_PAGER_OK)
459 		return (KERN_FAILURE); /* AKA SIGSEGV */
460 
461 	/* Ensure that the driver is obeying the interface. */
462 	MPASS(pager_first <= pager_last);
463 	MPASS(fs->first_pindex <= pager_last);
464 	MPASS(fs->first_pindex >= pager_first);
465 	MPASS(pager_last < fs->first_object->size);
466 
467 	vm_fault_restore_map_lock(fs);
468 	if (fs->map->timestamp != fs->map_generation) {
469 		vm_fault_populate_cleanup(fs->first_object, pager_first,
470 		    pager_last);
471 		return (KERN_RESTART);
472 	}
473 
474 	/*
475 	 * The map is unchanged after our last unlock.  Process the fault.
476 	 *
477 	 * The range [pager_first, pager_last] that is given to the
478 	 * pager is only a hint.  The pager may populate any range
479 	 * within the object that includes the requested page index.
480 	 * In case the pager expanded the range, clip it to fit into
481 	 * the map entry.
482 	 */
483 	map_first = OFF_TO_IDX(fs->entry->offset);
484 	if (map_first > pager_first) {
485 		vm_fault_populate_cleanup(fs->first_object, pager_first,
486 		    map_first - 1);
487 		pager_first = map_first;
488 	}
489 	map_last = map_first + atop(fs->entry->end - fs->entry->start) - 1;
490 	if (map_last < pager_last) {
491 		vm_fault_populate_cleanup(fs->first_object, map_last + 1,
492 		    pager_last);
493 		pager_last = map_last;
494 	}
495 	for (pidx = pager_first, m = vm_page_lookup(fs->first_object, pidx);
496 	    pidx <= pager_last;
497 	    pidx += npages, m = vm_page_next(&m[npages - 1])) {
498 		vaddr = fs->entry->start + IDX_TO_OFF(pidx) - fs->entry->offset;
499 #if defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \
500     __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv)
501 		psind = m->psind;
502 		if (psind > 0 && ((vaddr & (pagesizes[psind] - 1)) != 0 ||
503 		    pidx + OFF_TO_IDX(pagesizes[psind]) - 1 > pager_last ||
504 		    !pmap_ps_enabled(fs->map->pmap) || fs->wired))
505 			psind = 0;
506 #else
507 		psind = 0;
508 #endif
509 		npages = atop(pagesizes[psind]);
510 		for (i = 0; i < npages; i++) {
511 			vm_fault_populate_check_page(&m[i]);
512 			vm_fault_dirty(fs, &m[i]);
513 		}
514 		VM_OBJECT_WUNLOCK(fs->first_object);
515 		rv = pmap_enter(fs->map->pmap, vaddr, m, fs->prot, fs->fault_type |
516 		    (fs->wired ? PMAP_ENTER_WIRED : 0), psind);
517 #if defined(__amd64__)
518 		if (psind > 0 && rv == KERN_FAILURE) {
519 			for (i = 0; i < npages; i++) {
520 				rv = pmap_enter(fs->map->pmap, vaddr + ptoa(i),
521 				    &m[i], fs->prot, fs->fault_type |
522 				    (fs->wired ? PMAP_ENTER_WIRED : 0), 0);
523 				MPASS(rv == KERN_SUCCESS);
524 			}
525 		}
526 #else
527 		MPASS(rv == KERN_SUCCESS);
528 #endif
529 		VM_OBJECT_WLOCK(fs->first_object);
530 		for (i = 0; i < npages; i++) {
531 			if ((fs->fault_flags & VM_FAULT_WIRE) != 0)
532 				vm_page_wire(&m[i]);
533 			else
534 				vm_page_activate(&m[i]);
535 			if (fs->m_hold != NULL && m[i].pindex == fs->first_pindex) {
536 				(*fs->m_hold) = &m[i];
537 				vm_page_wire(&m[i]);
538 			}
539 			vm_page_xunbusy(&m[i]);
540 		}
541 	}
542 	curthread->td_ru.ru_majflt++;
543 	return (KERN_SUCCESS);
544 }
545 
546 static int prot_fault_translation;
547 SYSCTL_INT(_machdep, OID_AUTO, prot_fault_translation, CTLFLAG_RWTUN,
548     &prot_fault_translation, 0,
549     "Control signal to deliver on protection fault");
550 
551 /* compat definition to keep common code for signal translation */
552 #define	UCODE_PAGEFLT	12
553 #ifdef T_PAGEFLT
554 _Static_assert(UCODE_PAGEFLT == T_PAGEFLT, "T_PAGEFLT");
555 #endif
556 
557 /*
558  *	vm_fault_trap:
559  *
560  *	Handle a page fault occurring at the given address,
561  *	requiring the given permissions, in the map specified.
562  *	If successful, the page is inserted into the
563  *	associated physical map.
564  *
565  *	NOTE: the given address should be truncated to the
566  *	proper page address.
567  *
568  *	KERN_SUCCESS is returned if the page fault is handled; otherwise,
569  *	a standard error specifying why the fault is fatal is returned.
570  *
571  *	The map in question must be referenced, and remains so.
572  *	Caller may hold no locks.
573  */
574 int
575 vm_fault_trap(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
576     int fault_flags, int *signo, int *ucode)
577 {
578 	int result;
579 
580 	MPASS(signo == NULL || ucode != NULL);
581 #ifdef KTRACE
582 	if (map != kernel_map && KTRPOINT(curthread, KTR_FAULT))
583 		ktrfault(vaddr, fault_type);
584 #endif
585 	result = vm_fault(map, trunc_page(vaddr), fault_type, fault_flags,
586 	    NULL);
587 	KASSERT(result == KERN_SUCCESS || result == KERN_FAILURE ||
588 	    result == KERN_INVALID_ADDRESS ||
589 	    result == KERN_RESOURCE_SHORTAGE ||
590 	    result == KERN_PROTECTION_FAILURE ||
591 	    result == KERN_OUT_OF_BOUNDS,
592 	    ("Unexpected Mach error %d from vm_fault()", result));
593 #ifdef KTRACE
594 	if (map != kernel_map && KTRPOINT(curthread, KTR_FAULTEND))
595 		ktrfaultend(result);
596 #endif
597 	if (result != KERN_SUCCESS && signo != NULL) {
598 		switch (result) {
599 		case KERN_FAILURE:
600 		case KERN_INVALID_ADDRESS:
601 			*signo = SIGSEGV;
602 			*ucode = SEGV_MAPERR;
603 			break;
604 		case KERN_RESOURCE_SHORTAGE:
605 			*signo = SIGBUS;
606 			*ucode = BUS_OOMERR;
607 			break;
608 		case KERN_OUT_OF_BOUNDS:
609 			*signo = SIGBUS;
610 			*ucode = BUS_OBJERR;
611 			break;
612 		case KERN_PROTECTION_FAILURE:
613 			if (prot_fault_translation == 0) {
614 				/*
615 				 * Autodetect.  This check also covers
616 				 * the images without the ABI-tag ELF
617 				 * note.
618 				 */
619 				if (SV_CURPROC_ABI() == SV_ABI_FREEBSD &&
620 				    curproc->p_osrel >= P_OSREL_SIGSEGV) {
621 					*signo = SIGSEGV;
622 					*ucode = SEGV_ACCERR;
623 				} else {
624 					*signo = SIGBUS;
625 					*ucode = UCODE_PAGEFLT;
626 				}
627 			} else if (prot_fault_translation == 1) {
628 				/* Always compat mode. */
629 				*signo = SIGBUS;
630 				*ucode = UCODE_PAGEFLT;
631 			} else {
632 				/* Always SIGSEGV mode. */
633 				*signo = SIGSEGV;
634 				*ucode = SEGV_ACCERR;
635 			}
636 			break;
637 		default:
638 			KASSERT(0, ("Unexpected Mach error %d from vm_fault()",
639 			    result));
640 			break;
641 		}
642 	}
643 	return (result);
644 }
645 
646 static int
647 vm_fault_lock_vnode(struct faultstate *fs, bool objlocked)
648 {
649 	struct vnode *vp;
650 	int error, locked;
651 
652 	if (fs->object->type != OBJT_VNODE)
653 		return (KERN_SUCCESS);
654 	vp = fs->object->handle;
655 	if (vp == fs->vp) {
656 		ASSERT_VOP_LOCKED(vp, "saved vnode is not locked");
657 		return (KERN_SUCCESS);
658 	}
659 
660 	/*
661 	 * Perform an unlock in case the desired vnode changed while
662 	 * the map was unlocked during a retry.
663 	 */
664 	unlock_vp(fs);
665 
666 	locked = VOP_ISLOCKED(vp);
667 	if (locked != LK_EXCLUSIVE)
668 		locked = LK_SHARED;
669 
670 	/*
671 	 * We must not sleep acquiring the vnode lock while we have
672 	 * the page exclusive busied or the object's
673 	 * paging-in-progress count incremented.  Otherwise, we could
674 	 * deadlock.
675 	 */
676 	error = vget(vp, locked | LK_CANRECURSE | LK_NOWAIT);
677 	if (error == 0) {
678 		fs->vp = vp;
679 		return (KERN_SUCCESS);
680 	}
681 
682 	vhold(vp);
683 	if (objlocked)
684 		unlock_and_deallocate(fs);
685 	else
686 		fault_deallocate(fs);
687 	error = vget(vp, locked | LK_RETRY | LK_CANRECURSE);
688 	vdrop(vp);
689 	fs->vp = vp;
690 	KASSERT(error == 0, ("vm_fault: vget failed %d", error));
691 	return (KERN_RESOURCE_SHORTAGE);
692 }
693 
694 /*
695  * Calculate the desired readahead.  Handle drop-behind.
696  *
697  * Returns the number of readahead blocks to pass to the pager.
698  */
699 static int
700 vm_fault_readahead(struct faultstate *fs)
701 {
702 	int era, nera;
703 	u_char behavior;
704 
705 	KASSERT(fs->lookup_still_valid, ("map unlocked"));
706 	era = fs->entry->read_ahead;
707 	behavior = vm_map_entry_behavior(fs->entry);
708 	if (behavior == MAP_ENTRY_BEHAV_RANDOM) {
709 		nera = 0;
710 	} else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) {
711 		nera = VM_FAULT_READ_AHEAD_MAX;
712 		if (fs->vaddr == fs->entry->next_read)
713 			vm_fault_dontneed(fs, fs->vaddr, nera);
714 	} else if (fs->vaddr == fs->entry->next_read) {
715 		/*
716 		 * This is a sequential fault.  Arithmetically
717 		 * increase the requested number of pages in
718 		 * the read-ahead window.  The requested
719 		 * number of pages is "# of sequential faults
720 		 * x (read ahead min + 1) + read ahead min"
721 		 */
722 		nera = VM_FAULT_READ_AHEAD_MIN;
723 		if (era > 0) {
724 			nera += era + 1;
725 			if (nera > VM_FAULT_READ_AHEAD_MAX)
726 				nera = VM_FAULT_READ_AHEAD_MAX;
727 		}
728 		if (era == VM_FAULT_READ_AHEAD_MAX)
729 			vm_fault_dontneed(fs, fs->vaddr, nera);
730 	} else {
731 		/*
732 		 * This is a non-sequential fault.
733 		 */
734 		nera = 0;
735 	}
736 	if (era != nera) {
737 		/*
738 		 * A read lock on the map suffices to update
739 		 * the read ahead count safely.
740 		 */
741 		fs->entry->read_ahead = nera;
742 	}
743 
744 	return (nera);
745 }
746 
747 static int
748 vm_fault_lookup(struct faultstate *fs)
749 {
750 	int result;
751 
752 	KASSERT(!fs->lookup_still_valid,
753 	   ("vm_fault_lookup: Map already locked."));
754 	result = vm_map_lookup(&fs->map, fs->vaddr, fs->fault_type |
755 	    VM_PROT_FAULT_LOOKUP, &fs->entry, &fs->first_object,
756 	    &fs->first_pindex, &fs->prot, &fs->wired);
757 	if (result != KERN_SUCCESS) {
758 		unlock_vp(fs);
759 		return (result);
760 	}
761 
762 	fs->map_generation = fs->map->timestamp;
763 
764 	if (fs->entry->eflags & MAP_ENTRY_NOFAULT) {
765 		panic("%s: fault on nofault entry, addr: %#lx",
766 		    __func__, (u_long)fs->vaddr);
767 	}
768 
769 	if (fs->entry->eflags & MAP_ENTRY_IN_TRANSITION &&
770 	    fs->entry->wiring_thread != curthread) {
771 		vm_map_unlock_read(fs->map);
772 		vm_map_lock(fs->map);
773 		if (vm_map_lookup_entry(fs->map, fs->vaddr, &fs->entry) &&
774 		    (fs->entry->eflags & MAP_ENTRY_IN_TRANSITION)) {
775 			unlock_vp(fs);
776 			fs->entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
777 			vm_map_unlock_and_wait(fs->map, 0);
778 		} else
779 			vm_map_unlock(fs->map);
780 		return (KERN_RESOURCE_SHORTAGE);
781 	}
782 
783 	MPASS((fs->entry->eflags & MAP_ENTRY_GUARD) == 0);
784 
785 	if (fs->wired)
786 		fs->fault_type = fs->prot | (fs->fault_type & VM_PROT_COPY);
787 	else
788 		KASSERT((fs->fault_flags & VM_FAULT_WIRE) == 0,
789 		    ("!fs->wired && VM_FAULT_WIRE"));
790 	fs->lookup_still_valid = true;
791 
792 	return (KERN_SUCCESS);
793 }
794 
795 static int
796 vm_fault_relookup(struct faultstate *fs)
797 {
798 	vm_object_t retry_object;
799 	vm_pindex_t retry_pindex;
800 	vm_prot_t retry_prot;
801 	int result;
802 
803 	if (!vm_map_trylock_read(fs->map))
804 		return (KERN_RESTART);
805 
806 	fs->lookup_still_valid = true;
807 	if (fs->map->timestamp == fs->map_generation)
808 		return (KERN_SUCCESS);
809 
810 	result = vm_map_lookup_locked(&fs->map, fs->vaddr, fs->fault_type,
811 	    &fs->entry, &retry_object, &retry_pindex, &retry_prot,
812 	    &fs->wired);
813 	if (result != KERN_SUCCESS) {
814 		/*
815 		 * If retry of map lookup would have blocked then
816 		 * retry fault from start.
817 		 */
818 		if (result == KERN_FAILURE)
819 			return (KERN_RESTART);
820 		return (result);
821 	}
822 	if (retry_object != fs->first_object ||
823 	    retry_pindex != fs->first_pindex)
824 		return (KERN_RESTART);
825 
826 	/*
827 	 * Check whether the protection has changed or the object has
828 	 * been copied while we left the map unlocked. Changing from
829 	 * read to write permission is OK - we leave the page
830 	 * write-protected, and catch the write fault. Changing from
831 	 * write to read permission means that we can't mark the page
832 	 * write-enabled after all.
833 	 */
834 	fs->prot &= retry_prot;
835 	fs->fault_type &= retry_prot;
836 	if (fs->prot == 0)
837 		return (KERN_RESTART);
838 
839 	/* Reassert because wired may have changed. */
840 	KASSERT(fs->wired || (fs->fault_flags & VM_FAULT_WIRE) == 0,
841 	    ("!wired && VM_FAULT_WIRE"));
842 
843 	return (KERN_SUCCESS);
844 }
845 
846 static void
847 vm_fault_cow(struct faultstate *fs)
848 {
849 	bool is_first_object_locked;
850 
851 	/*
852 	 * This allows pages to be virtually copied from a backing_object
853 	 * into the first_object, where the backing object has no other
854 	 * refs to it, and cannot gain any more refs.  Instead of a bcopy,
855 	 * we just move the page from the backing object to the first
856 	 * object.  Note that we must mark the page dirty in the first
857 	 * object so that it will go out to swap when needed.
858 	 */
859 	is_first_object_locked = false;
860 	if (
861 	    /*
862 	     * Only one shadow object and no other refs.
863 	     */
864 	    fs->object->shadow_count == 1 && fs->object->ref_count == 1 &&
865 	    /*
866 	     * No other ways to look the object up
867 	     */
868 	    fs->object->handle == NULL && (fs->object->flags & OBJ_ANON) != 0 &&
869 	    /*
870 	     * We don't chase down the shadow chain and we can acquire locks.
871 	     */
872 	    (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs->first_object)) &&
873 	    fs->object == fs->first_object->backing_object &&
874 	    VM_OBJECT_TRYWLOCK(fs->object)) {
875 		/*
876 		 * Remove but keep xbusy for replace.  fs->m is moved into
877 		 * fs->first_object and left busy while fs->first_m is
878 		 * conditionally freed.
879 		 */
880 		vm_page_remove_xbusy(fs->m);
881 		vm_page_replace(fs->m, fs->first_object, fs->first_pindex,
882 		    fs->first_m);
883 		vm_page_dirty(fs->m);
884 #if VM_NRESERVLEVEL > 0
885 		/*
886 		 * Rename the reservation.
887 		 */
888 		vm_reserv_rename(fs->m, fs->first_object, fs->object,
889 		    OFF_TO_IDX(fs->first_object->backing_object_offset));
890 #endif
891 		VM_OBJECT_WUNLOCK(fs->object);
892 		VM_OBJECT_WUNLOCK(fs->first_object);
893 		fs->first_m = fs->m;
894 		fs->m = NULL;
895 		VM_CNT_INC(v_cow_optim);
896 	} else {
897 		if (is_first_object_locked)
898 			VM_OBJECT_WUNLOCK(fs->first_object);
899 		/*
900 		 * Oh, well, lets copy it.
901 		 */
902 		pmap_copy_page(fs->m, fs->first_m);
903 		vm_page_valid(fs->first_m);
904 		if (fs->wired && (fs->fault_flags & VM_FAULT_WIRE) == 0) {
905 			vm_page_wire(fs->first_m);
906 			vm_page_unwire(fs->m, PQ_INACTIVE);
907 		}
908 		/*
909 		 * Save the cow page to be released after
910 		 * pmap_enter is complete.
911 		 */
912 		fs->m_cow = fs->m;
913 		fs->m = NULL;
914 	}
915 	/*
916 	 * fs->object != fs->first_object due to above
917 	 * conditional
918 	 */
919 	vm_object_pip_wakeup(fs->object);
920 
921 	/*
922 	 * Only use the new page below...
923 	 */
924 	fs->object = fs->first_object;
925 	fs->pindex = fs->first_pindex;
926 	fs->m = fs->first_m;
927 	VM_CNT_INC(v_cow_faults);
928 	curthread->td_cow++;
929 }
930 
931 static bool
932 vm_fault_next(struct faultstate *fs)
933 {
934 	vm_object_t next_object;
935 
936 	/*
937 	 * The requested page does not exist at this object/
938 	 * offset.  Remove the invalid page from the object,
939 	 * waking up anyone waiting for it, and continue on to
940 	 * the next object.  However, if this is the top-level
941 	 * object, we must leave the busy page in place to
942 	 * prevent another process from rushing past us, and
943 	 * inserting the page in that object at the same time
944 	 * that we are.
945 	 */
946 	if (fs->object == fs->first_object) {
947 		fs->first_m = fs->m;
948 		fs->m = NULL;
949 	} else
950 		fault_page_free(&fs->m);
951 
952 	/*
953 	 * Move on to the next object.  Lock the next object before
954 	 * unlocking the current one.
955 	 */
956 	VM_OBJECT_ASSERT_WLOCKED(fs->object);
957 	next_object = fs->object->backing_object;
958 	if (next_object == NULL)
959 		return (false);
960 	MPASS(fs->first_m != NULL);
961 	KASSERT(fs->object != next_object, ("object loop %p", next_object));
962 	VM_OBJECT_WLOCK(next_object);
963 	vm_object_pip_add(next_object, 1);
964 	if (fs->object != fs->first_object)
965 		vm_object_pip_wakeup(fs->object);
966 	fs->pindex += OFF_TO_IDX(fs->object->backing_object_offset);
967 	VM_OBJECT_WUNLOCK(fs->object);
968 	fs->object = next_object;
969 
970 	return (true);
971 }
972 
973 static void
974 vm_fault_zerofill(struct faultstate *fs)
975 {
976 
977 	/*
978 	 * If there's no object left, fill the page in the top
979 	 * object with zeros.
980 	 */
981 	if (fs->object != fs->first_object) {
982 		vm_object_pip_wakeup(fs->object);
983 		fs->object = fs->first_object;
984 		fs->pindex = fs->first_pindex;
985 	}
986 	MPASS(fs->first_m != NULL);
987 	MPASS(fs->m == NULL);
988 	fs->m = fs->first_m;
989 	fs->first_m = NULL;
990 
991 	/*
992 	 * Zero the page if necessary and mark it valid.
993 	 */
994 	if ((fs->m->flags & PG_ZERO) == 0) {
995 		pmap_zero_page(fs->m);
996 	} else {
997 		VM_CNT_INC(v_ozfod);
998 	}
999 	VM_CNT_INC(v_zfod);
1000 	vm_page_valid(fs->m);
1001 }
1002 
1003 /*
1004  * Allocate a page directly or via the object populate method.
1005  */
1006 static int
1007 vm_fault_allocate(struct faultstate *fs)
1008 {
1009 	struct domainset *dset;
1010 	int alloc_req;
1011 	int rv;
1012 
1013 	if ((fs->object->flags & OBJ_SIZEVNLOCK) != 0) {
1014 		rv = vm_fault_lock_vnode(fs, true);
1015 		MPASS(rv == KERN_SUCCESS || rv == KERN_RESOURCE_SHORTAGE);
1016 		if (rv == KERN_RESOURCE_SHORTAGE)
1017 			return (rv);
1018 	}
1019 
1020 	if (fs->pindex >= fs->object->size)
1021 		return (KERN_OUT_OF_BOUNDS);
1022 
1023 	if (fs->object == fs->first_object &&
1024 	    (fs->first_object->flags & OBJ_POPULATE) != 0 &&
1025 	    fs->first_object->shadow_count == 0) {
1026 		rv = vm_fault_populate(fs);
1027 		switch (rv) {
1028 		case KERN_SUCCESS:
1029 		case KERN_FAILURE:
1030 		case KERN_RESTART:
1031 			return (rv);
1032 		case KERN_NOT_RECEIVER:
1033 			/*
1034 			 * Pager's populate() method
1035 			 * returned VM_PAGER_BAD.
1036 			 */
1037 			break;
1038 		default:
1039 			panic("inconsistent return codes");
1040 		}
1041 	}
1042 
1043 	/*
1044 	 * Allocate a new page for this object/offset pair.
1045 	 *
1046 	 * Unlocked read of the p_flag is harmless. At worst, the P_KILLED
1047 	 * might be not observed there, and allocation can fail, causing
1048 	 * restart and new reading of the p_flag.
1049 	 */
1050 	dset = fs->object->domain.dr_policy;
1051 	if (dset == NULL)
1052 		dset = curthread->td_domain.dr_policy;
1053 	if (!vm_page_count_severe_set(&dset->ds_mask) || P_KILLED(curproc)) {
1054 #if VM_NRESERVLEVEL > 0
1055 		vm_object_color(fs->object, atop(fs->vaddr) - fs->pindex);
1056 #endif
1057 		alloc_req = P_KILLED(curproc) ?
1058 		    VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL;
1059 		if (fs->object->type != OBJT_VNODE &&
1060 		    fs->object->backing_object == NULL)
1061 			alloc_req |= VM_ALLOC_ZERO;
1062 		fs->m = vm_page_alloc(fs->object, fs->pindex, alloc_req);
1063 	}
1064 	if (fs->m == NULL) {
1065 		unlock_and_deallocate(fs);
1066 		if (vm_pfault_oom_attempts < 0 ||
1067 		    fs->oom < vm_pfault_oom_attempts) {
1068 			fs->oom++;
1069 			vm_waitpfault(dset, vm_pfault_oom_wait * hz);
1070 		} else 	{
1071 			if (bootverbose)
1072 				printf(
1073 		"proc %d (%s) failed to alloc page on fault, starting OOM\n",
1074 				    curproc->p_pid, curproc->p_comm);
1075 			vm_pageout_oom(VM_OOM_MEM_PF);
1076 			fs->oom = 0;
1077 		}
1078 		return (KERN_RESOURCE_SHORTAGE);
1079 	}
1080 	fs->oom = 0;
1081 
1082 	return (KERN_NOT_RECEIVER);
1083 }
1084 
1085 /*
1086  * Call the pager to retrieve the page if there is a chance
1087  * that the pager has it, and potentially retrieve additional
1088  * pages at the same time.
1089  */
1090 static int
1091 vm_fault_getpages(struct faultstate *fs, int nera, int *behindp, int *aheadp)
1092 {
1093 	vm_offset_t e_end, e_start;
1094 	int ahead, behind, cluster_offset, rv;
1095 	u_char behavior;
1096 
1097 	/*
1098 	 * Prepare for unlocking the map.  Save the map
1099 	 * entry's start and end addresses, which are used to
1100 	 * optimize the size of the pager operation below.
1101 	 * Even if the map entry's addresses change after
1102 	 * unlocking the map, using the saved addresses is
1103 	 * safe.
1104 	 */
1105 	e_start = fs->entry->start;
1106 	e_end = fs->entry->end;
1107 	behavior = vm_map_entry_behavior(fs->entry);
1108 
1109 	/*
1110 	 * Release the map lock before locking the vnode or
1111 	 * sleeping in the pager.  (If the current object has
1112 	 * a shadow, then an earlier iteration of this loop
1113 	 * may have already unlocked the map.)
1114 	 */
1115 	unlock_map(fs);
1116 
1117 	rv = vm_fault_lock_vnode(fs, false);
1118 	MPASS(rv == KERN_SUCCESS || rv == KERN_RESOURCE_SHORTAGE);
1119 	if (rv == KERN_RESOURCE_SHORTAGE)
1120 		return (rv);
1121 	KASSERT(fs->vp == NULL || !fs->map->system_map,
1122 	    ("vm_fault: vnode-backed object mapped by system map"));
1123 
1124 	/*
1125 	 * Page in the requested page and hint the pager,
1126 	 * that it may bring up surrounding pages.
1127 	 */
1128 	if (nera == -1 || behavior == MAP_ENTRY_BEHAV_RANDOM ||
1129 	    P_KILLED(curproc)) {
1130 		behind = 0;
1131 		ahead = 0;
1132 	} else {
1133 		/* Is this a sequential fault? */
1134 		if (nera > 0) {
1135 			behind = 0;
1136 			ahead = nera;
1137 		} else {
1138 			/*
1139 			 * Request a cluster of pages that is
1140 			 * aligned to a VM_FAULT_READ_DEFAULT
1141 			 * page offset boundary within the
1142 			 * object.  Alignment to a page offset
1143 			 * boundary is more likely to coincide
1144 			 * with the underlying file system
1145 			 * block than alignment to a virtual
1146 			 * address boundary.
1147 			 */
1148 			cluster_offset = fs->pindex % VM_FAULT_READ_DEFAULT;
1149 			behind = ulmin(cluster_offset,
1150 			    atop(fs->vaddr - e_start));
1151 			ahead = VM_FAULT_READ_DEFAULT - 1 - cluster_offset;
1152 		}
1153 		ahead = ulmin(ahead, atop(e_end - fs->vaddr) - 1);
1154 	}
1155 	*behindp = behind;
1156 	*aheadp = ahead;
1157 	rv = vm_pager_get_pages(fs->object, &fs->m, 1, behindp, aheadp);
1158 	if (rv == VM_PAGER_OK)
1159 		return (KERN_SUCCESS);
1160 	if (rv == VM_PAGER_ERROR)
1161 		printf("vm_fault: pager read error, pid %d (%s)\n",
1162 		    curproc->p_pid, curproc->p_comm);
1163 	/*
1164 	 * If an I/O error occurred or the requested page was
1165 	 * outside the range of the pager, clean up and return
1166 	 * an error.
1167 	 */
1168 	if (rv == VM_PAGER_ERROR || rv == VM_PAGER_BAD)
1169 		return (KERN_OUT_OF_BOUNDS);
1170 	return (KERN_NOT_RECEIVER);
1171 }
1172 
1173 /*
1174  * Wait/Retry if the page is busy.  We have to do this if the page is
1175  * either exclusive or shared busy because the vm_pager may be using
1176  * read busy for pageouts (and even pageins if it is the vnode pager),
1177  * and we could end up trying to pagein and pageout the same page
1178  * simultaneously.
1179  *
1180  * We can theoretically allow the busy case on a read fault if the page
1181  * is marked valid, but since such pages are typically already pmap'd,
1182  * putting that special case in might be more effort then it is worth.
1183  * We cannot under any circumstances mess around with a shared busied
1184  * page except, perhaps, to pmap it.
1185  */
1186 static void
1187 vm_fault_busy_sleep(struct faultstate *fs)
1188 {
1189 	/*
1190 	 * Reference the page before unlocking and
1191 	 * sleeping so that the page daemon is less
1192 	 * likely to reclaim it.
1193 	 */
1194 	vm_page_aflag_set(fs->m, PGA_REFERENCED);
1195 	if (fs->object != fs->first_object) {
1196 		fault_page_release(&fs->first_m);
1197 		vm_object_pip_wakeup(fs->first_object);
1198 	}
1199 	vm_object_pip_wakeup(fs->object);
1200 	unlock_map(fs);
1201 	if (fs->m == vm_page_lookup(fs->object, fs->pindex))
1202 		vm_page_busy_sleep(fs->m, "vmpfw", false);
1203 	else
1204 		VM_OBJECT_WUNLOCK(fs->object);
1205 	VM_CNT_INC(v_intrans);
1206 	vm_object_deallocate(fs->first_object);
1207 }
1208 
1209 int
1210 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
1211     int fault_flags, vm_page_t *m_hold)
1212 {
1213 	struct faultstate fs;
1214 	int ahead, behind, faultcount;
1215 	int nera, result, rv;
1216 	bool dead, hardfault;
1217 
1218 	VM_CNT_INC(v_vm_faults);
1219 
1220 	if ((curthread->td_pflags & TDP_NOFAULTING) != 0)
1221 		return (KERN_PROTECTION_FAILURE);
1222 
1223 	fs.vp = NULL;
1224 	fs.vaddr = vaddr;
1225 	fs.m_hold = m_hold;
1226 	fs.fault_flags = fault_flags;
1227 	fs.map = map;
1228 	fs.lookup_still_valid = false;
1229 	fs.oom = 0;
1230 	faultcount = 0;
1231 	nera = -1;
1232 	hardfault = false;
1233 
1234 RetryFault:
1235 	fs.fault_type = fault_type;
1236 
1237 	/*
1238 	 * Find the backing store object and offset into it to begin the
1239 	 * search.
1240 	 */
1241 	result = vm_fault_lookup(&fs);
1242 	if (result != KERN_SUCCESS) {
1243 		if (result == KERN_RESOURCE_SHORTAGE)
1244 			goto RetryFault;
1245 		return (result);
1246 	}
1247 
1248 	/*
1249 	 * Try to avoid lock contention on the top-level object through
1250 	 * special-case handling of some types of page faults, specifically,
1251 	 * those that are mapping an existing page from the top-level object.
1252 	 * Under this condition, a read lock on the object suffices, allowing
1253 	 * multiple page faults of a similar type to run in parallel.
1254 	 */
1255 	if (fs.vp == NULL /* avoid locked vnode leak */ &&
1256 	    (fs.fault_flags & (VM_FAULT_WIRE | VM_FAULT_DIRTY)) == 0) {
1257 		VM_OBJECT_RLOCK(fs.first_object);
1258 		rv = vm_fault_soft_fast(&fs);
1259 		if (rv == KERN_SUCCESS)
1260 			return (rv);
1261 		if (!VM_OBJECT_TRYUPGRADE(fs.first_object)) {
1262 			VM_OBJECT_RUNLOCK(fs.first_object);
1263 			VM_OBJECT_WLOCK(fs.first_object);
1264 		}
1265 	} else {
1266 		VM_OBJECT_WLOCK(fs.first_object);
1267 	}
1268 
1269 	/*
1270 	 * Make a reference to this object to prevent its disposal while we
1271 	 * are messing with it.  Once we have the reference, the map is free
1272 	 * to be diddled.  Since objects reference their shadows (and copies),
1273 	 * they will stay around as well.
1274 	 *
1275 	 * Bump the paging-in-progress count to prevent size changes (e.g.
1276 	 * truncation operations) during I/O.
1277 	 */
1278 	vm_object_reference_locked(fs.first_object);
1279 	vm_object_pip_add(fs.first_object, 1);
1280 
1281 	fs.m_cow = fs.m = fs.first_m = NULL;
1282 
1283 	/*
1284 	 * Search for the page at object/offset.
1285 	 */
1286 	fs.object = fs.first_object;
1287 	fs.pindex = fs.first_pindex;
1288 	while (TRUE) {
1289 		KASSERT(fs.m == NULL,
1290 		    ("page still set %p at loop start", fs.m));
1291 		/*
1292 		 * If the object is marked for imminent termination,
1293 		 * we retry here, since the collapse pass has raced
1294 		 * with us.  Otherwise, if we see terminally dead
1295 		 * object, return fail.
1296 		 */
1297 		if ((fs.object->flags & OBJ_DEAD) != 0) {
1298 			dead = fs.object->type == OBJT_DEAD;
1299 			unlock_and_deallocate(&fs);
1300 			if (dead)
1301 				return (KERN_PROTECTION_FAILURE);
1302 			pause("vmf_de", 1);
1303 			goto RetryFault;
1304 		}
1305 
1306 		/*
1307 		 * See if page is resident
1308 		 */
1309 		fs.m = vm_page_lookup(fs.object, fs.pindex);
1310 		if (fs.m != NULL) {
1311 			if (vm_page_tryxbusy(fs.m) == 0) {
1312 				vm_fault_busy_sleep(&fs);
1313 				goto RetryFault;
1314 			}
1315 
1316 			/*
1317 			 * The page is marked busy for other processes and the
1318 			 * pagedaemon.  If it still is completely valid we
1319 			 * are done.
1320 			 */
1321 			if (vm_page_all_valid(fs.m)) {
1322 				VM_OBJECT_WUNLOCK(fs.object);
1323 				break; /* break to PAGE HAS BEEN FOUND. */
1324 			}
1325 		}
1326 		VM_OBJECT_ASSERT_WLOCKED(fs.object);
1327 
1328 		/*
1329 		 * Page is not resident.  If the pager might contain the page
1330 		 * or this is the beginning of the search, allocate a new
1331 		 * page.  (Default objects are zero-fill, so there is no real
1332 		 * pager for them.)
1333 		 */
1334 		if (fs.m == NULL && (fs.object->type != OBJT_DEFAULT ||
1335 		    fs.object == fs.first_object)) {
1336 			rv = vm_fault_allocate(&fs);
1337 			switch (rv) {
1338 			case KERN_RESTART:
1339 				unlock_and_deallocate(&fs);
1340 				/* FALLTHROUGH */
1341 			case KERN_RESOURCE_SHORTAGE:
1342 				goto RetryFault;
1343 			case KERN_SUCCESS:
1344 			case KERN_FAILURE:
1345 			case KERN_OUT_OF_BOUNDS:
1346 				unlock_and_deallocate(&fs);
1347 				return (rv);
1348 			case KERN_NOT_RECEIVER:
1349 				break;
1350 			default:
1351 				panic("vm_fault: Unhandled rv %d", rv);
1352 			}
1353 		}
1354 
1355 		/*
1356 		 * Default objects have no pager so no exclusive busy exists
1357 		 * to protect this page in the chain.  Skip to the next
1358 		 * object without dropping the lock to preserve atomicity of
1359 		 * shadow faults.
1360 		 */
1361 		if (fs.object->type != OBJT_DEFAULT) {
1362 			/*
1363 			 * At this point, we have either allocated a new page
1364 			 * or found an existing page that is only partially
1365 			 * valid.
1366 			 *
1367 			 * We hold a reference on the current object and the
1368 			 * page is exclusive busied.  The exclusive busy
1369 			 * prevents simultaneous faults and collapses while
1370 			 * the object lock is dropped.
1371 		 	 */
1372 			VM_OBJECT_WUNLOCK(fs.object);
1373 
1374 			/*
1375 			 * If the pager for the current object might have
1376 			 * the page, then determine the number of additional
1377 			 * pages to read and potentially reprioritize
1378 			 * previously read pages for earlier reclamation.
1379 			 * These operations should only be performed once per
1380 			 * page fault.  Even if the current pager doesn't
1381 			 * have the page, the number of additional pages to
1382 			 * read will apply to subsequent objects in the
1383 			 * shadow chain.
1384 			 */
1385 			if (nera == -1 && !P_KILLED(curproc))
1386 				nera = vm_fault_readahead(&fs);
1387 
1388 			rv = vm_fault_getpages(&fs, nera, &behind, &ahead);
1389 			if (rv == KERN_SUCCESS) {
1390 				faultcount = behind + 1 + ahead;
1391 				hardfault = true;
1392 				break; /* break to PAGE HAS BEEN FOUND. */
1393 			}
1394 			if (rv == KERN_RESOURCE_SHORTAGE)
1395 				goto RetryFault;
1396 			VM_OBJECT_WLOCK(fs.object);
1397 			if (rv == KERN_OUT_OF_BOUNDS) {
1398 				fault_page_free(&fs.m);
1399 				unlock_and_deallocate(&fs);
1400 				return (rv);
1401 			}
1402 		}
1403 
1404 		/*
1405 		 * The page was not found in the current object.  Try to
1406 		 * traverse into a backing object or zero fill if none is
1407 		 * found.
1408 		 */
1409 		if (vm_fault_next(&fs))
1410 			continue;
1411 		VM_OBJECT_WUNLOCK(fs.object);
1412 		vm_fault_zerofill(&fs);
1413 		/* Don't try to prefault neighboring pages. */
1414 		faultcount = 1;
1415 		break;	/* break to PAGE HAS BEEN FOUND. */
1416 	}
1417 
1418 	/*
1419 	 * PAGE HAS BEEN FOUND.  A valid page has been found and exclusively
1420 	 * busied.  The object lock must no longer be held.
1421 	 */
1422 	vm_page_assert_xbusied(fs.m);
1423 	VM_OBJECT_ASSERT_UNLOCKED(fs.object);
1424 
1425 	/*
1426 	 * If the page is being written, but isn't already owned by the
1427 	 * top-level object, we have to copy it into a new page owned by the
1428 	 * top-level object.
1429 	 */
1430 	if (fs.object != fs.first_object) {
1431 		/*
1432 		 * We only really need to copy if we want to write it.
1433 		 */
1434 		if ((fs.fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1435 			vm_fault_cow(&fs);
1436 			/*
1437 			 * We only try to prefault read-only mappings to the
1438 			 * neighboring pages when this copy-on-write fault is
1439 			 * a hard fault.  In other cases, trying to prefault
1440 			 * is typically wasted effort.
1441 			 */
1442 			if (faultcount == 0)
1443 				faultcount = 1;
1444 
1445 		} else {
1446 			fs.prot &= ~VM_PROT_WRITE;
1447 		}
1448 	}
1449 
1450 	/*
1451 	 * We must verify that the maps have not changed since our last
1452 	 * lookup.
1453 	 */
1454 	if (!fs.lookup_still_valid) {
1455 		result = vm_fault_relookup(&fs);
1456 		if (result != KERN_SUCCESS) {
1457 			fault_deallocate(&fs);
1458 			if (result == KERN_RESTART)
1459 				goto RetryFault;
1460 			return (result);
1461 		}
1462 	}
1463 	VM_OBJECT_ASSERT_UNLOCKED(fs.object);
1464 
1465 	/*
1466 	 * If the page was filled by a pager, save the virtual address that
1467 	 * should be faulted on next under a sequential access pattern to the
1468 	 * map entry.  A read lock on the map suffices to update this address
1469 	 * safely.
1470 	 */
1471 	if (hardfault)
1472 		fs.entry->next_read = vaddr + ptoa(ahead) + PAGE_SIZE;
1473 
1474 	/*
1475 	 * Page must be completely valid or it is not fit to
1476 	 * map into user space.  vm_pager_get_pages() ensures this.
1477 	 */
1478 	vm_page_assert_xbusied(fs.m);
1479 	KASSERT(vm_page_all_valid(fs.m),
1480 	    ("vm_fault: page %p partially invalid", fs.m));
1481 
1482 	vm_fault_dirty(&fs, fs.m);
1483 
1484 	/*
1485 	 * Put this page into the physical map.  We had to do the unlock above
1486 	 * because pmap_enter() may sleep.  We don't put the page
1487 	 * back on the active queue until later so that the pageout daemon
1488 	 * won't find it (yet).
1489 	 */
1490 	pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot,
1491 	    fs.fault_type | (fs.wired ? PMAP_ENTER_WIRED : 0), 0);
1492 	if (faultcount != 1 && (fs.fault_flags & VM_FAULT_WIRE) == 0 &&
1493 	    fs.wired == 0)
1494 		vm_fault_prefault(&fs, vaddr,
1495 		    faultcount > 0 ? behind : PFBAK,
1496 		    faultcount > 0 ? ahead : PFFOR, false);
1497 
1498 	/*
1499 	 * If the page is not wired down, then put it where the pageout daemon
1500 	 * can find it.
1501 	 */
1502 	if ((fs.fault_flags & VM_FAULT_WIRE) != 0)
1503 		vm_page_wire(fs.m);
1504 	else
1505 		vm_page_activate(fs.m);
1506 	if (fs.m_hold != NULL) {
1507 		(*fs.m_hold) = fs.m;
1508 		vm_page_wire(fs.m);
1509 	}
1510 	vm_page_xunbusy(fs.m);
1511 	fs.m = NULL;
1512 
1513 	/*
1514 	 * Unlock everything, and return
1515 	 */
1516 	fault_deallocate(&fs);
1517 	if (hardfault) {
1518 		VM_CNT_INC(v_io_faults);
1519 		curthread->td_ru.ru_majflt++;
1520 #ifdef RACCT
1521 		if (racct_enable && fs.object->type == OBJT_VNODE) {
1522 			PROC_LOCK(curproc);
1523 			if ((fs.fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1524 				racct_add_force(curproc, RACCT_WRITEBPS,
1525 				    PAGE_SIZE + behind * PAGE_SIZE);
1526 				racct_add_force(curproc, RACCT_WRITEIOPS, 1);
1527 			} else {
1528 				racct_add_force(curproc, RACCT_READBPS,
1529 				    PAGE_SIZE + ahead * PAGE_SIZE);
1530 				racct_add_force(curproc, RACCT_READIOPS, 1);
1531 			}
1532 			PROC_UNLOCK(curproc);
1533 		}
1534 #endif
1535 	} else
1536 		curthread->td_ru.ru_minflt++;
1537 
1538 	return (KERN_SUCCESS);
1539 }
1540 
1541 /*
1542  * Speed up the reclamation of pages that precede the faulting pindex within
1543  * the first object of the shadow chain.  Essentially, perform the equivalent
1544  * to madvise(..., MADV_DONTNEED) on a large cluster of pages that precedes
1545  * the faulting pindex by the cluster size when the pages read by vm_fault()
1546  * cross a cluster-size boundary.  The cluster size is the greater of the
1547  * smallest superpage size and VM_FAULT_DONTNEED_MIN.
1548  *
1549  * When "fs->first_object" is a shadow object, the pages in the backing object
1550  * that precede the faulting pindex are deactivated by vm_fault().  So, this
1551  * function must only be concerned with pages in the first object.
1552  */
1553 static void
1554 vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr, int ahead)
1555 {
1556 	vm_map_entry_t entry;
1557 	vm_object_t first_object, object;
1558 	vm_offset_t end, start;
1559 	vm_page_t m, m_next;
1560 	vm_pindex_t pend, pstart;
1561 	vm_size_t size;
1562 
1563 	object = fs->object;
1564 	VM_OBJECT_ASSERT_UNLOCKED(object);
1565 	first_object = fs->first_object;
1566 	/* Neither fictitious nor unmanaged pages can be reclaimed. */
1567 	if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) {
1568 		VM_OBJECT_RLOCK(first_object);
1569 		size = VM_FAULT_DONTNEED_MIN;
1570 		if (MAXPAGESIZES > 1 && size < pagesizes[1])
1571 			size = pagesizes[1];
1572 		end = rounddown2(vaddr, size);
1573 		if (vaddr - end >= size - PAGE_SIZE - ptoa(ahead) &&
1574 		    (entry = fs->entry)->start < end) {
1575 			if (end - entry->start < size)
1576 				start = entry->start;
1577 			else
1578 				start = end - size;
1579 			pmap_advise(fs->map->pmap, start, end, MADV_DONTNEED);
1580 			pstart = OFF_TO_IDX(entry->offset) + atop(start -
1581 			    entry->start);
1582 			m_next = vm_page_find_least(first_object, pstart);
1583 			pend = OFF_TO_IDX(entry->offset) + atop(end -
1584 			    entry->start);
1585 			while ((m = m_next) != NULL && m->pindex < pend) {
1586 				m_next = TAILQ_NEXT(m, listq);
1587 				if (!vm_page_all_valid(m) ||
1588 				    vm_page_busied(m))
1589 					continue;
1590 
1591 				/*
1592 				 * Don't clear PGA_REFERENCED, since it would
1593 				 * likely represent a reference by a different
1594 				 * process.
1595 				 *
1596 				 * Typically, at this point, prefetched pages
1597 				 * are still in the inactive queue.  Only
1598 				 * pages that triggered page faults are in the
1599 				 * active queue.  The test for whether the page
1600 				 * is in the inactive queue is racy; in the
1601 				 * worst case we will requeue the page
1602 				 * unnecessarily.
1603 				 */
1604 				if (!vm_page_inactive(m))
1605 					vm_page_deactivate(m);
1606 			}
1607 		}
1608 		VM_OBJECT_RUNLOCK(first_object);
1609 	}
1610 }
1611 
1612 /*
1613  * vm_fault_prefault provides a quick way of clustering
1614  * pagefaults into a processes address space.  It is a "cousin"
1615  * of vm_map_pmap_enter, except it runs at page fault time instead
1616  * of mmap time.
1617  */
1618 static void
1619 vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
1620     int backward, int forward, bool obj_locked)
1621 {
1622 	pmap_t pmap;
1623 	vm_map_entry_t entry;
1624 	vm_object_t backing_object, lobject;
1625 	vm_offset_t addr, starta;
1626 	vm_pindex_t pindex;
1627 	vm_page_t m;
1628 	int i;
1629 
1630 	pmap = fs->map->pmap;
1631 	if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
1632 		return;
1633 
1634 	entry = fs->entry;
1635 
1636 	if (addra < backward * PAGE_SIZE) {
1637 		starta = entry->start;
1638 	} else {
1639 		starta = addra - backward * PAGE_SIZE;
1640 		if (starta < entry->start)
1641 			starta = entry->start;
1642 	}
1643 
1644 	/*
1645 	 * Generate the sequence of virtual addresses that are candidates for
1646 	 * prefaulting in an outward spiral from the faulting virtual address,
1647 	 * "addra".  Specifically, the sequence is "addra - PAGE_SIZE", "addra
1648 	 * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ...
1649 	 * If the candidate address doesn't have a backing physical page, then
1650 	 * the loop immediately terminates.
1651 	 */
1652 	for (i = 0; i < 2 * imax(backward, forward); i++) {
1653 		addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE :
1654 		    PAGE_SIZE);
1655 		if (addr > addra + forward * PAGE_SIZE)
1656 			addr = 0;
1657 
1658 		if (addr < starta || addr >= entry->end)
1659 			continue;
1660 
1661 		if (!pmap_is_prefaultable(pmap, addr))
1662 			continue;
1663 
1664 		pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1665 		lobject = entry->object.vm_object;
1666 		if (!obj_locked)
1667 			VM_OBJECT_RLOCK(lobject);
1668 		while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1669 		    lobject->type == OBJT_DEFAULT &&
1670 		    (backing_object = lobject->backing_object) != NULL) {
1671 			KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1672 			    0, ("vm_fault_prefault: unaligned object offset"));
1673 			pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1674 			VM_OBJECT_RLOCK(backing_object);
1675 			if (!obj_locked || lobject != entry->object.vm_object)
1676 				VM_OBJECT_RUNLOCK(lobject);
1677 			lobject = backing_object;
1678 		}
1679 		if (m == NULL) {
1680 			if (!obj_locked || lobject != entry->object.vm_object)
1681 				VM_OBJECT_RUNLOCK(lobject);
1682 			break;
1683 		}
1684 		if (vm_page_all_valid(m) &&
1685 		    (m->flags & PG_FICTITIOUS) == 0)
1686 			pmap_enter_quick(pmap, addr, m, entry->protection);
1687 		if (!obj_locked || lobject != entry->object.vm_object)
1688 			VM_OBJECT_RUNLOCK(lobject);
1689 	}
1690 }
1691 
1692 /*
1693  * Hold each of the physical pages that are mapped by the specified range of
1694  * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
1695  * and allow the specified types of access, "prot".  If all of the implied
1696  * pages are successfully held, then the number of held pages is returned
1697  * together with pointers to those pages in the array "ma".  However, if any
1698  * of the pages cannot be held, -1 is returned.
1699  */
1700 int
1701 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
1702     vm_prot_t prot, vm_page_t *ma, int max_count)
1703 {
1704 	vm_offset_t end, va;
1705 	vm_page_t *mp;
1706 	int count;
1707 	boolean_t pmap_failed;
1708 
1709 	if (len == 0)
1710 		return (0);
1711 	end = round_page(addr + len);
1712 	addr = trunc_page(addr);
1713 
1714 	if (!vm_map_range_valid(map, addr, end))
1715 		return (-1);
1716 
1717 	if (atop(end - addr) > max_count)
1718 		panic("vm_fault_quick_hold_pages: count > max_count");
1719 	count = atop(end - addr);
1720 
1721 	/*
1722 	 * Most likely, the physical pages are resident in the pmap, so it is
1723 	 * faster to try pmap_extract_and_hold() first.
1724 	 */
1725 	pmap_failed = FALSE;
1726 	for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) {
1727 		*mp = pmap_extract_and_hold(map->pmap, va, prot);
1728 		if (*mp == NULL)
1729 			pmap_failed = TRUE;
1730 		else if ((prot & VM_PROT_WRITE) != 0 &&
1731 		    (*mp)->dirty != VM_PAGE_BITS_ALL) {
1732 			/*
1733 			 * Explicitly dirty the physical page.  Otherwise, the
1734 			 * caller's changes may go unnoticed because they are
1735 			 * performed through an unmanaged mapping or by a DMA
1736 			 * operation.
1737 			 *
1738 			 * The object lock is not held here.
1739 			 * See vm_page_clear_dirty_mask().
1740 			 */
1741 			vm_page_dirty(*mp);
1742 		}
1743 	}
1744 	if (pmap_failed) {
1745 		/*
1746 		 * One or more pages could not be held by the pmap.  Either no
1747 		 * page was mapped at the specified virtual address or that
1748 		 * mapping had insufficient permissions.  Attempt to fault in
1749 		 * and hold these pages.
1750 		 *
1751 		 * If vm_fault_disable_pagefaults() was called,
1752 		 * i.e., TDP_NOFAULTING is set, we must not sleep nor
1753 		 * acquire MD VM locks, which means we must not call
1754 		 * vm_fault().  Some (out of tree) callers mark
1755 		 * too wide a code area with vm_fault_disable_pagefaults()
1756 		 * already, use the VM_PROT_QUICK_NOFAULT flag to request
1757 		 * the proper behaviour explicitly.
1758 		 */
1759 		if ((prot & VM_PROT_QUICK_NOFAULT) != 0 &&
1760 		    (curthread->td_pflags & TDP_NOFAULTING) != 0)
1761 			goto error;
1762 		for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE)
1763 			if (*mp == NULL && vm_fault(map, va, prot,
1764 			    VM_FAULT_NORMAL, mp) != KERN_SUCCESS)
1765 				goto error;
1766 	}
1767 	return (count);
1768 error:
1769 	for (mp = ma; mp < ma + count; mp++)
1770 		if (*mp != NULL)
1771 			vm_page_unwire(*mp, PQ_INACTIVE);
1772 	return (-1);
1773 }
1774 
1775 /*
1776  *	Routine:
1777  *		vm_fault_copy_entry
1778  *	Function:
1779  *		Create new shadow object backing dst_entry with private copy of
1780  *		all underlying pages. When src_entry is equal to dst_entry,
1781  *		function implements COW for wired-down map entry. Otherwise,
1782  *		it forks wired entry into dst_map.
1783  *
1784  *	In/out conditions:
1785  *		The source and destination maps must be locked for write.
1786  *		The source map entry must be wired down (or be a sharing map
1787  *		entry corresponding to a main map entry that is wired down).
1788  */
1789 void
1790 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1791     vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1792     vm_ooffset_t *fork_charge)
1793 {
1794 	vm_object_t backing_object, dst_object, object, src_object;
1795 	vm_pindex_t dst_pindex, pindex, src_pindex;
1796 	vm_prot_t access, prot;
1797 	vm_offset_t vaddr;
1798 	vm_page_t dst_m;
1799 	vm_page_t src_m;
1800 	boolean_t upgrade;
1801 
1802 #ifdef	lint
1803 	src_map++;
1804 #endif	/* lint */
1805 
1806 	upgrade = src_entry == dst_entry;
1807 	access = prot = dst_entry->protection;
1808 
1809 	src_object = src_entry->object.vm_object;
1810 	src_pindex = OFF_TO_IDX(src_entry->offset);
1811 
1812 	if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
1813 		dst_object = src_object;
1814 		vm_object_reference(dst_object);
1815 	} else {
1816 		/*
1817 		 * Create the top-level object for the destination entry.
1818 		 * Doesn't actually shadow anything - we copy the pages
1819 		 * directly.
1820 		 */
1821 		dst_object = vm_object_allocate_anon(atop(dst_entry->end -
1822 		    dst_entry->start), NULL, NULL, 0);
1823 #if VM_NRESERVLEVEL > 0
1824 		dst_object->flags |= OBJ_COLORED;
1825 		dst_object->pg_color = atop(dst_entry->start);
1826 #endif
1827 		dst_object->domain = src_object->domain;
1828 		dst_object->charge = dst_entry->end - dst_entry->start;
1829 	}
1830 
1831 	VM_OBJECT_WLOCK(dst_object);
1832 	KASSERT(upgrade || dst_entry->object.vm_object == NULL,
1833 	    ("vm_fault_copy_entry: vm_object not NULL"));
1834 	if (src_object != dst_object) {
1835 		dst_entry->object.vm_object = dst_object;
1836 		dst_entry->offset = 0;
1837 		dst_entry->eflags &= ~MAP_ENTRY_VN_EXEC;
1838 	}
1839 	if (fork_charge != NULL) {
1840 		KASSERT(dst_entry->cred == NULL,
1841 		    ("vm_fault_copy_entry: leaked swp charge"));
1842 		dst_object->cred = curthread->td_ucred;
1843 		crhold(dst_object->cred);
1844 		*fork_charge += dst_object->charge;
1845 	} else if ((dst_object->type == OBJT_DEFAULT ||
1846 	    dst_object->type == OBJT_SWAP) &&
1847 	    dst_object->cred == NULL) {
1848 		KASSERT(dst_entry->cred != NULL, ("no cred for entry %p",
1849 		    dst_entry));
1850 		dst_object->cred = dst_entry->cred;
1851 		dst_entry->cred = NULL;
1852 	}
1853 
1854 	/*
1855 	 * If not an upgrade, then enter the mappings in the pmap as
1856 	 * read and/or execute accesses.  Otherwise, enter them as
1857 	 * write accesses.
1858 	 *
1859 	 * A writeable large page mapping is only created if all of
1860 	 * the constituent small page mappings are modified. Marking
1861 	 * PTEs as modified on inception allows promotion to happen
1862 	 * without taking potentially large number of soft faults.
1863 	 */
1864 	if (!upgrade)
1865 		access &= ~VM_PROT_WRITE;
1866 
1867 	/*
1868 	 * Loop through all of the virtual pages within the entry's
1869 	 * range, copying each page from the source object to the
1870 	 * destination object.  Since the source is wired, those pages
1871 	 * must exist.  In contrast, the destination is pageable.
1872 	 * Since the destination object doesn't share any backing storage
1873 	 * with the source object, all of its pages must be dirtied,
1874 	 * regardless of whether they can be written.
1875 	 */
1876 	for (vaddr = dst_entry->start, dst_pindex = 0;
1877 	    vaddr < dst_entry->end;
1878 	    vaddr += PAGE_SIZE, dst_pindex++) {
1879 again:
1880 		/*
1881 		 * Find the page in the source object, and copy it in.
1882 		 * Because the source is wired down, the page will be
1883 		 * in memory.
1884 		 */
1885 		if (src_object != dst_object)
1886 			VM_OBJECT_RLOCK(src_object);
1887 		object = src_object;
1888 		pindex = src_pindex + dst_pindex;
1889 		while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
1890 		    (backing_object = object->backing_object) != NULL) {
1891 			/*
1892 			 * Unless the source mapping is read-only or
1893 			 * it is presently being upgraded from
1894 			 * read-only, the first object in the shadow
1895 			 * chain should provide all of the pages.  In
1896 			 * other words, this loop body should never be
1897 			 * executed when the source mapping is already
1898 			 * read/write.
1899 			 */
1900 			KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 ||
1901 			    upgrade,
1902 			    ("vm_fault_copy_entry: main object missing page"));
1903 
1904 			VM_OBJECT_RLOCK(backing_object);
1905 			pindex += OFF_TO_IDX(object->backing_object_offset);
1906 			if (object != dst_object)
1907 				VM_OBJECT_RUNLOCK(object);
1908 			object = backing_object;
1909 		}
1910 		KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing"));
1911 
1912 		if (object != dst_object) {
1913 			/*
1914 			 * Allocate a page in the destination object.
1915 			 */
1916 			dst_m = vm_page_alloc(dst_object, (src_object ==
1917 			    dst_object ? src_pindex : 0) + dst_pindex,
1918 			    VM_ALLOC_NORMAL);
1919 			if (dst_m == NULL) {
1920 				VM_OBJECT_WUNLOCK(dst_object);
1921 				VM_OBJECT_RUNLOCK(object);
1922 				vm_wait(dst_object);
1923 				VM_OBJECT_WLOCK(dst_object);
1924 				goto again;
1925 			}
1926 			pmap_copy_page(src_m, dst_m);
1927 			VM_OBJECT_RUNLOCK(object);
1928 			dst_m->dirty = dst_m->valid = src_m->valid;
1929 		} else {
1930 			dst_m = src_m;
1931 			if (vm_page_busy_acquire(dst_m, VM_ALLOC_WAITFAIL) == 0)
1932 				goto again;
1933 			if (dst_m->pindex >= dst_object->size) {
1934 				/*
1935 				 * We are upgrading.  Index can occur
1936 				 * out of bounds if the object type is
1937 				 * vnode and the file was truncated.
1938 				 */
1939 				vm_page_xunbusy(dst_m);
1940 				break;
1941 			}
1942 		}
1943 		VM_OBJECT_WUNLOCK(dst_object);
1944 
1945 		/*
1946 		 * Enter it in the pmap. If a wired, copy-on-write
1947 		 * mapping is being replaced by a write-enabled
1948 		 * mapping, then wire that new mapping.
1949 		 *
1950 		 * The page can be invalid if the user called
1951 		 * msync(MS_INVALIDATE) or truncated the backing vnode
1952 		 * or shared memory object.  In this case, do not
1953 		 * insert it into pmap, but still do the copy so that
1954 		 * all copies of the wired map entry have similar
1955 		 * backing pages.
1956 		 */
1957 		if (vm_page_all_valid(dst_m)) {
1958 			pmap_enter(dst_map->pmap, vaddr, dst_m, prot,
1959 			    access | (upgrade ? PMAP_ENTER_WIRED : 0), 0);
1960 		}
1961 
1962 		/*
1963 		 * Mark it no longer busy, and put it on the active list.
1964 		 */
1965 		VM_OBJECT_WLOCK(dst_object);
1966 
1967 		if (upgrade) {
1968 			if (src_m != dst_m) {
1969 				vm_page_unwire(src_m, PQ_INACTIVE);
1970 				vm_page_wire(dst_m);
1971 			} else {
1972 				KASSERT(vm_page_wired(dst_m),
1973 				    ("dst_m %p is not wired", dst_m));
1974 			}
1975 		} else {
1976 			vm_page_activate(dst_m);
1977 		}
1978 		vm_page_xunbusy(dst_m);
1979 	}
1980 	VM_OBJECT_WUNLOCK(dst_object);
1981 	if (upgrade) {
1982 		dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
1983 		vm_object_deallocate(src_object);
1984 	}
1985 }
1986 
1987 /*
1988  * Block entry into the machine-independent layer's page fault handler by
1989  * the calling thread.  Subsequent calls to vm_fault() by that thread will
1990  * return KERN_PROTECTION_FAILURE.  Enable machine-dependent handling of
1991  * spurious page faults.
1992  */
1993 int
1994 vm_fault_disable_pagefaults(void)
1995 {
1996 
1997 	return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR));
1998 }
1999 
2000 void
2001 vm_fault_enable_pagefaults(int save)
2002 {
2003 
2004 	curthread_pflags_restore(save);
2005 }
2006