xref: /freebsd/sys/vm/vm_fault.c (revision 3332f1b444d4a73238e9f59cca27bfc95fe936bd)
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 
119 #define	VM_FAULT_DONTNEED_MIN	1048576
120 
121 struct faultstate {
122 	/* Fault parameters. */
123 	vm_offset_t	vaddr;
124 	vm_page_t	*m_hold;
125 	vm_prot_t	fault_type;
126 	vm_prot_t	prot;
127 	int		fault_flags;
128 	struct timeval	oom_start_time;
129 	bool		oom_started;
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 bdry_idx, 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,
446 	    &pager_last);
447 
448 	VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
449 	if (rv == VM_PAGER_BAD) {
450 		/*
451 		 * VM_PAGER_BAD is the backdoor for a pager to request
452 		 * normal fault handling.
453 		 */
454 		vm_fault_restore_map_lock(fs);
455 		if (fs->map->timestamp != fs->map_generation)
456 			return (KERN_RESTART);
457 		return (KERN_NOT_RECEIVER);
458 	}
459 	if (rv != VM_PAGER_OK)
460 		return (KERN_FAILURE); /* AKA SIGSEGV */
461 
462 	/* Ensure that the driver is obeying the interface. */
463 	MPASS(pager_first <= pager_last);
464 	MPASS(fs->first_pindex <= pager_last);
465 	MPASS(fs->first_pindex >= pager_first);
466 	MPASS(pager_last < fs->first_object->size);
467 
468 	vm_fault_restore_map_lock(fs);
469 	bdry_idx = (fs->entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) >>
470 	    MAP_ENTRY_SPLIT_BOUNDARY_SHIFT;
471 	if (fs->map->timestamp != fs->map_generation) {
472 		if (bdry_idx == 0) {
473 			vm_fault_populate_cleanup(fs->first_object, pager_first,
474 			    pager_last);
475 		} else {
476 			m = vm_page_lookup(fs->first_object, pager_first);
477 			if (m != fs->m)
478 				vm_page_xunbusy(m);
479 		}
480 		return (KERN_RESTART);
481 	}
482 
483 	/*
484 	 * The map is unchanged after our last unlock.  Process the fault.
485 	 *
486 	 * First, the special case of largepage mappings, where
487 	 * populate only busies the first page in superpage run.
488 	 */
489 	if (bdry_idx != 0) {
490 		KASSERT(PMAP_HAS_LARGEPAGES,
491 		    ("missing pmap support for large pages"));
492 		m = vm_page_lookup(fs->first_object, pager_first);
493 		vm_fault_populate_check_page(m);
494 		VM_OBJECT_WUNLOCK(fs->first_object);
495 		vaddr = fs->entry->start + IDX_TO_OFF(pager_first) -
496 		    fs->entry->offset;
497 		/* assert alignment for entry */
498 		KASSERT((vaddr & (pagesizes[bdry_idx] - 1)) == 0,
499     ("unaligned superpage start %#jx pager_first %#jx offset %#jx vaddr %#jx",
500 		    (uintmax_t)fs->entry->start, (uintmax_t)pager_first,
501 		    (uintmax_t)fs->entry->offset, (uintmax_t)vaddr));
502 		KASSERT((VM_PAGE_TO_PHYS(m) & (pagesizes[bdry_idx] - 1)) == 0,
503 		    ("unaligned superpage m %p %#jx", m,
504 		    (uintmax_t)VM_PAGE_TO_PHYS(m)));
505 		rv = pmap_enter(fs->map->pmap, vaddr, m, fs->prot,
506 		    fs->fault_type | (fs->wired ? PMAP_ENTER_WIRED : 0) |
507 		    PMAP_ENTER_LARGEPAGE, bdry_idx);
508 		VM_OBJECT_WLOCK(fs->first_object);
509 		vm_page_xunbusy(m);
510 		if (rv != KERN_SUCCESS)
511 			goto out;
512 		if ((fs->fault_flags & VM_FAULT_WIRE) != 0) {
513 			for (i = 0; i < atop(pagesizes[bdry_idx]); i++)
514 				vm_page_wire(m + i);
515 		}
516 		if (fs->m_hold != NULL) {
517 			*fs->m_hold = m + (fs->first_pindex - pager_first);
518 			vm_page_wire(*fs->m_hold);
519 		}
520 		goto out;
521 	}
522 
523 	/*
524 	 * The range [pager_first, pager_last] that is given to the
525 	 * pager is only a hint.  The pager may populate any range
526 	 * within the object that includes the requested page index.
527 	 * In case the pager expanded the range, clip it to fit into
528 	 * the map entry.
529 	 */
530 	map_first = OFF_TO_IDX(fs->entry->offset);
531 	if (map_first > pager_first) {
532 		vm_fault_populate_cleanup(fs->first_object, pager_first,
533 		    map_first - 1);
534 		pager_first = map_first;
535 	}
536 	map_last = map_first + atop(fs->entry->end - fs->entry->start) - 1;
537 	if (map_last < pager_last) {
538 		vm_fault_populate_cleanup(fs->first_object, map_last + 1,
539 		    pager_last);
540 		pager_last = map_last;
541 	}
542 	for (pidx = pager_first, m = vm_page_lookup(fs->first_object, pidx);
543 	    pidx <= pager_last;
544 	    pidx += npages, m = vm_page_next(&m[npages - 1])) {
545 		vaddr = fs->entry->start + IDX_TO_OFF(pidx) - fs->entry->offset;
546 
547 		psind = m->psind;
548 		if (psind > 0 && ((vaddr & (pagesizes[psind] - 1)) != 0 ||
549 		    pidx + OFF_TO_IDX(pagesizes[psind]) - 1 > pager_last ||
550 		    !pmap_ps_enabled(fs->map->pmap) || fs->wired))
551 			psind = 0;
552 
553 		npages = atop(pagesizes[psind]);
554 		for (i = 0; i < npages; i++) {
555 			vm_fault_populate_check_page(&m[i]);
556 			vm_fault_dirty(fs, &m[i]);
557 		}
558 		VM_OBJECT_WUNLOCK(fs->first_object);
559 		rv = pmap_enter(fs->map->pmap, vaddr, m, fs->prot, fs->fault_type |
560 		    (fs->wired ? PMAP_ENTER_WIRED : 0), psind);
561 
562 		/*
563 		 * pmap_enter() may fail for a superpage mapping if additional
564 		 * protection policies prevent the full mapping.
565 		 * For example, this will happen on amd64 if the entire
566 		 * address range does not share the same userspace protection
567 		 * key.  Revert to single-page mappings if this happens.
568 		 */
569 		MPASS(rv == KERN_SUCCESS ||
570 		    (psind > 0 && rv == KERN_PROTECTION_FAILURE));
571 		if (__predict_false(psind > 0 &&
572 		    rv == KERN_PROTECTION_FAILURE)) {
573 			for (i = 0; i < npages; i++) {
574 				rv = pmap_enter(fs->map->pmap, vaddr + ptoa(i),
575 				    &m[i], fs->prot, fs->fault_type |
576 				    (fs->wired ? PMAP_ENTER_WIRED : 0), 0);
577 				MPASS(rv == KERN_SUCCESS);
578 			}
579 		}
580 
581 		VM_OBJECT_WLOCK(fs->first_object);
582 		for (i = 0; i < npages; i++) {
583 			if ((fs->fault_flags & VM_FAULT_WIRE) != 0)
584 				vm_page_wire(&m[i]);
585 			else
586 				vm_page_activate(&m[i]);
587 			if (fs->m_hold != NULL && m[i].pindex == fs->first_pindex) {
588 				(*fs->m_hold) = &m[i];
589 				vm_page_wire(&m[i]);
590 			}
591 			vm_page_xunbusy(&m[i]);
592 		}
593 	}
594 out:
595 	curthread->td_ru.ru_majflt++;
596 	return (rv);
597 }
598 
599 static int prot_fault_translation;
600 SYSCTL_INT(_machdep, OID_AUTO, prot_fault_translation, CTLFLAG_RWTUN,
601     &prot_fault_translation, 0,
602     "Control signal to deliver on protection fault");
603 
604 /* compat definition to keep common code for signal translation */
605 #define	UCODE_PAGEFLT	12
606 #ifdef T_PAGEFLT
607 _Static_assert(UCODE_PAGEFLT == T_PAGEFLT, "T_PAGEFLT");
608 #endif
609 
610 /*
611  *	vm_fault_trap:
612  *
613  *	Handle a page fault occurring at the given address,
614  *	requiring the given permissions, in the map specified.
615  *	If successful, the page is inserted into the
616  *	associated physical map.
617  *
618  *	NOTE: the given address should be truncated to the
619  *	proper page address.
620  *
621  *	KERN_SUCCESS is returned if the page fault is handled; otherwise,
622  *	a standard error specifying why the fault is fatal is returned.
623  *
624  *	The map in question must be referenced, and remains so.
625  *	Caller may hold no locks.
626  */
627 int
628 vm_fault_trap(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
629     int fault_flags, int *signo, int *ucode)
630 {
631 	int result;
632 
633 	MPASS(signo == NULL || ucode != NULL);
634 #ifdef KTRACE
635 	if (map != kernel_map && KTRPOINT(curthread, KTR_FAULT))
636 		ktrfault(vaddr, fault_type);
637 #endif
638 	result = vm_fault(map, trunc_page(vaddr), fault_type, fault_flags,
639 	    NULL);
640 	KASSERT(result == KERN_SUCCESS || result == KERN_FAILURE ||
641 	    result == KERN_INVALID_ADDRESS ||
642 	    result == KERN_RESOURCE_SHORTAGE ||
643 	    result == KERN_PROTECTION_FAILURE ||
644 	    result == KERN_OUT_OF_BOUNDS,
645 	    ("Unexpected Mach error %d from vm_fault()", result));
646 #ifdef KTRACE
647 	if (map != kernel_map && KTRPOINT(curthread, KTR_FAULTEND))
648 		ktrfaultend(result);
649 #endif
650 	if (result != KERN_SUCCESS && signo != NULL) {
651 		switch (result) {
652 		case KERN_FAILURE:
653 		case KERN_INVALID_ADDRESS:
654 			*signo = SIGSEGV;
655 			*ucode = SEGV_MAPERR;
656 			break;
657 		case KERN_RESOURCE_SHORTAGE:
658 			*signo = SIGBUS;
659 			*ucode = BUS_OOMERR;
660 			break;
661 		case KERN_OUT_OF_BOUNDS:
662 			*signo = SIGBUS;
663 			*ucode = BUS_OBJERR;
664 			break;
665 		case KERN_PROTECTION_FAILURE:
666 			if (prot_fault_translation == 0) {
667 				/*
668 				 * Autodetect.  This check also covers
669 				 * the images without the ABI-tag ELF
670 				 * note.
671 				 */
672 				if (SV_CURPROC_ABI() == SV_ABI_FREEBSD &&
673 				    curproc->p_osrel >= P_OSREL_SIGSEGV) {
674 					*signo = SIGSEGV;
675 					*ucode = SEGV_ACCERR;
676 				} else {
677 					*signo = SIGBUS;
678 					*ucode = UCODE_PAGEFLT;
679 				}
680 			} else if (prot_fault_translation == 1) {
681 				/* Always compat mode. */
682 				*signo = SIGBUS;
683 				*ucode = UCODE_PAGEFLT;
684 			} else {
685 				/* Always SIGSEGV mode. */
686 				*signo = SIGSEGV;
687 				*ucode = SEGV_ACCERR;
688 			}
689 			break;
690 		default:
691 			KASSERT(0, ("Unexpected Mach error %d from vm_fault()",
692 			    result));
693 			break;
694 		}
695 	}
696 	return (result);
697 }
698 
699 static int
700 vm_fault_lock_vnode(struct faultstate *fs, bool objlocked)
701 {
702 	struct vnode *vp;
703 	int error, locked;
704 
705 	if (fs->object->type != OBJT_VNODE)
706 		return (KERN_SUCCESS);
707 	vp = fs->object->handle;
708 	if (vp == fs->vp) {
709 		ASSERT_VOP_LOCKED(vp, "saved vnode is not locked");
710 		return (KERN_SUCCESS);
711 	}
712 
713 	/*
714 	 * Perform an unlock in case the desired vnode changed while
715 	 * the map was unlocked during a retry.
716 	 */
717 	unlock_vp(fs);
718 
719 	locked = VOP_ISLOCKED(vp);
720 	if (locked != LK_EXCLUSIVE)
721 		locked = LK_SHARED;
722 
723 	/*
724 	 * We must not sleep acquiring the vnode lock while we have
725 	 * the page exclusive busied or the object's
726 	 * paging-in-progress count incremented.  Otherwise, we could
727 	 * deadlock.
728 	 */
729 	error = vget(vp, locked | LK_CANRECURSE | LK_NOWAIT);
730 	if (error == 0) {
731 		fs->vp = vp;
732 		return (KERN_SUCCESS);
733 	}
734 
735 	vhold(vp);
736 	if (objlocked)
737 		unlock_and_deallocate(fs);
738 	else
739 		fault_deallocate(fs);
740 	error = vget(vp, locked | LK_RETRY | LK_CANRECURSE);
741 	vdrop(vp);
742 	fs->vp = vp;
743 	KASSERT(error == 0, ("vm_fault: vget failed %d", error));
744 	return (KERN_RESOURCE_SHORTAGE);
745 }
746 
747 /*
748  * Calculate the desired readahead.  Handle drop-behind.
749  *
750  * Returns the number of readahead blocks to pass to the pager.
751  */
752 static int
753 vm_fault_readahead(struct faultstate *fs)
754 {
755 	int era, nera;
756 	u_char behavior;
757 
758 	KASSERT(fs->lookup_still_valid, ("map unlocked"));
759 	era = fs->entry->read_ahead;
760 	behavior = vm_map_entry_behavior(fs->entry);
761 	if (behavior == MAP_ENTRY_BEHAV_RANDOM) {
762 		nera = 0;
763 	} else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) {
764 		nera = VM_FAULT_READ_AHEAD_MAX;
765 		if (fs->vaddr == fs->entry->next_read)
766 			vm_fault_dontneed(fs, fs->vaddr, nera);
767 	} else if (fs->vaddr == fs->entry->next_read) {
768 		/*
769 		 * This is a sequential fault.  Arithmetically
770 		 * increase the requested number of pages in
771 		 * the read-ahead window.  The requested
772 		 * number of pages is "# of sequential faults
773 		 * x (read ahead min + 1) + read ahead min"
774 		 */
775 		nera = VM_FAULT_READ_AHEAD_MIN;
776 		if (era > 0) {
777 			nera += era + 1;
778 			if (nera > VM_FAULT_READ_AHEAD_MAX)
779 				nera = VM_FAULT_READ_AHEAD_MAX;
780 		}
781 		if (era == VM_FAULT_READ_AHEAD_MAX)
782 			vm_fault_dontneed(fs, fs->vaddr, nera);
783 	} else {
784 		/*
785 		 * This is a non-sequential fault.
786 		 */
787 		nera = 0;
788 	}
789 	if (era != nera) {
790 		/*
791 		 * A read lock on the map suffices to update
792 		 * the read ahead count safely.
793 		 */
794 		fs->entry->read_ahead = nera;
795 	}
796 
797 	return (nera);
798 }
799 
800 static int
801 vm_fault_lookup(struct faultstate *fs)
802 {
803 	int result;
804 
805 	KASSERT(!fs->lookup_still_valid,
806 	   ("vm_fault_lookup: Map already locked."));
807 	result = vm_map_lookup(&fs->map, fs->vaddr, fs->fault_type |
808 	    VM_PROT_FAULT_LOOKUP, &fs->entry, &fs->first_object,
809 	    &fs->first_pindex, &fs->prot, &fs->wired);
810 	if (result != KERN_SUCCESS) {
811 		unlock_vp(fs);
812 		return (result);
813 	}
814 
815 	fs->map_generation = fs->map->timestamp;
816 
817 	if (fs->entry->eflags & MAP_ENTRY_NOFAULT) {
818 		panic("%s: fault on nofault entry, addr: %#lx",
819 		    __func__, (u_long)fs->vaddr);
820 	}
821 
822 	if (fs->entry->eflags & MAP_ENTRY_IN_TRANSITION &&
823 	    fs->entry->wiring_thread != curthread) {
824 		vm_map_unlock_read(fs->map);
825 		vm_map_lock(fs->map);
826 		if (vm_map_lookup_entry(fs->map, fs->vaddr, &fs->entry) &&
827 		    (fs->entry->eflags & MAP_ENTRY_IN_TRANSITION)) {
828 			unlock_vp(fs);
829 			fs->entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
830 			vm_map_unlock_and_wait(fs->map, 0);
831 		} else
832 			vm_map_unlock(fs->map);
833 		return (KERN_RESOURCE_SHORTAGE);
834 	}
835 
836 	MPASS((fs->entry->eflags & MAP_ENTRY_GUARD) == 0);
837 
838 	if (fs->wired)
839 		fs->fault_type = fs->prot | (fs->fault_type & VM_PROT_COPY);
840 	else
841 		KASSERT((fs->fault_flags & VM_FAULT_WIRE) == 0,
842 		    ("!fs->wired && VM_FAULT_WIRE"));
843 	fs->lookup_still_valid = true;
844 
845 	return (KERN_SUCCESS);
846 }
847 
848 static int
849 vm_fault_relookup(struct faultstate *fs)
850 {
851 	vm_object_t retry_object;
852 	vm_pindex_t retry_pindex;
853 	vm_prot_t retry_prot;
854 	int result;
855 
856 	if (!vm_map_trylock_read(fs->map))
857 		return (KERN_RESTART);
858 
859 	fs->lookup_still_valid = true;
860 	if (fs->map->timestamp == fs->map_generation)
861 		return (KERN_SUCCESS);
862 
863 	result = vm_map_lookup_locked(&fs->map, fs->vaddr, fs->fault_type,
864 	    &fs->entry, &retry_object, &retry_pindex, &retry_prot,
865 	    &fs->wired);
866 	if (result != KERN_SUCCESS) {
867 		/*
868 		 * If retry of map lookup would have blocked then
869 		 * retry fault from start.
870 		 */
871 		if (result == KERN_FAILURE)
872 			return (KERN_RESTART);
873 		return (result);
874 	}
875 	if (retry_object != fs->first_object ||
876 	    retry_pindex != fs->first_pindex)
877 		return (KERN_RESTART);
878 
879 	/*
880 	 * Check whether the protection has changed or the object has
881 	 * been copied while we left the map unlocked. Changing from
882 	 * read to write permission is OK - we leave the page
883 	 * write-protected, and catch the write fault. Changing from
884 	 * write to read permission means that we can't mark the page
885 	 * write-enabled after all.
886 	 */
887 	fs->prot &= retry_prot;
888 	fs->fault_type &= retry_prot;
889 	if (fs->prot == 0)
890 		return (KERN_RESTART);
891 
892 	/* Reassert because wired may have changed. */
893 	KASSERT(fs->wired || (fs->fault_flags & VM_FAULT_WIRE) == 0,
894 	    ("!wired && VM_FAULT_WIRE"));
895 
896 	return (KERN_SUCCESS);
897 }
898 
899 static void
900 vm_fault_cow(struct faultstate *fs)
901 {
902 	bool is_first_object_locked;
903 
904 	KASSERT(fs->object != fs->first_object,
905 	    ("source and target COW objects are identical"));
906 
907 	/*
908 	 * This allows pages to be virtually copied from a backing_object
909 	 * into the first_object, where the backing object has no other
910 	 * refs to it, and cannot gain any more refs.  Instead of a bcopy,
911 	 * we just move the page from the backing object to the first
912 	 * object.  Note that we must mark the page dirty in the first
913 	 * object so that it will go out to swap when needed.
914 	 */
915 	is_first_object_locked = false;
916 	if (
917 	    /*
918 	     * Only one shadow object and no other refs.
919 	     */
920 	    fs->object->shadow_count == 1 && fs->object->ref_count == 1 &&
921 	    /*
922 	     * No other ways to look the object up
923 	     */
924 	    fs->object->handle == NULL && (fs->object->flags & OBJ_ANON) != 0 &&
925 	    /*
926 	     * We don't chase down the shadow chain and we can acquire locks.
927 	     */
928 	    (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs->first_object)) &&
929 	    fs->object == fs->first_object->backing_object &&
930 	    VM_OBJECT_TRYWLOCK(fs->object)) {
931 		/*
932 		 * Remove but keep xbusy for replace.  fs->m is moved into
933 		 * fs->first_object and left busy while fs->first_m is
934 		 * conditionally freed.
935 		 */
936 		vm_page_remove_xbusy(fs->m);
937 		vm_page_replace(fs->m, fs->first_object, fs->first_pindex,
938 		    fs->first_m);
939 		vm_page_dirty(fs->m);
940 #if VM_NRESERVLEVEL > 0
941 		/*
942 		 * Rename the reservation.
943 		 */
944 		vm_reserv_rename(fs->m, fs->first_object, fs->object,
945 		    OFF_TO_IDX(fs->first_object->backing_object_offset));
946 #endif
947 		VM_OBJECT_WUNLOCK(fs->object);
948 		VM_OBJECT_WUNLOCK(fs->first_object);
949 		fs->first_m = fs->m;
950 		fs->m = NULL;
951 		VM_CNT_INC(v_cow_optim);
952 	} else {
953 		if (is_first_object_locked)
954 			VM_OBJECT_WUNLOCK(fs->first_object);
955 		/*
956 		 * Oh, well, lets copy it.
957 		 */
958 		pmap_copy_page(fs->m, fs->first_m);
959 		vm_page_valid(fs->first_m);
960 		if (fs->wired && (fs->fault_flags & VM_FAULT_WIRE) == 0) {
961 			vm_page_wire(fs->first_m);
962 			vm_page_unwire(fs->m, PQ_INACTIVE);
963 		}
964 		/*
965 		 * Save the cow page to be released after
966 		 * pmap_enter is complete.
967 		 */
968 		fs->m_cow = fs->m;
969 		fs->m = NULL;
970 
971 		/*
972 		 * Typically, the shadow object is either private to this
973 		 * address space (OBJ_ONEMAPPING) or its pages are read only.
974 		 * In the highly unusual case where the pages of a shadow object
975 		 * are read/write shared between this and other address spaces,
976 		 * we need to ensure that any pmap-level mappings to the
977 		 * original, copy-on-write page from the backing object are
978 		 * removed from those other address spaces.
979 		 *
980 		 * The flag check is racy, but this is tolerable: if
981 		 * OBJ_ONEMAPPING is cleared after the check, the busy state
982 		 * ensures that new mappings of m_cow can't be created.
983 		 * pmap_enter() will replace an existing mapping in the current
984 		 * address space.  If OBJ_ONEMAPPING is set after the check,
985 		 * removing mappings will at worse trigger some unnecessary page
986 		 * faults.
987 		 */
988 		vm_page_assert_xbusied(fs->m_cow);
989 		if ((fs->first_object->flags & OBJ_ONEMAPPING) == 0)
990 			pmap_remove_all(fs->m_cow);
991 	}
992 
993 	vm_object_pip_wakeup(fs->object);
994 
995 	/*
996 	 * Only use the new page below...
997 	 */
998 	fs->object = fs->first_object;
999 	fs->pindex = fs->first_pindex;
1000 	fs->m = fs->first_m;
1001 	VM_CNT_INC(v_cow_faults);
1002 	curthread->td_cow++;
1003 }
1004 
1005 static bool
1006 vm_fault_next(struct faultstate *fs)
1007 {
1008 	vm_object_t next_object;
1009 
1010 	/*
1011 	 * The requested page does not exist at this object/
1012 	 * offset.  Remove the invalid page from the object,
1013 	 * waking up anyone waiting for it, and continue on to
1014 	 * the next object.  However, if this is the top-level
1015 	 * object, we must leave the busy page in place to
1016 	 * prevent another process from rushing past us, and
1017 	 * inserting the page in that object at the same time
1018 	 * that we are.
1019 	 */
1020 	if (fs->object == fs->first_object) {
1021 		fs->first_m = fs->m;
1022 		fs->m = NULL;
1023 	} else
1024 		fault_page_free(&fs->m);
1025 
1026 	/*
1027 	 * Move on to the next object.  Lock the next object before
1028 	 * unlocking the current one.
1029 	 */
1030 	VM_OBJECT_ASSERT_WLOCKED(fs->object);
1031 	next_object = fs->object->backing_object;
1032 	if (next_object == NULL)
1033 		return (false);
1034 	MPASS(fs->first_m != NULL);
1035 	KASSERT(fs->object != next_object, ("object loop %p", next_object));
1036 	VM_OBJECT_WLOCK(next_object);
1037 	vm_object_pip_add(next_object, 1);
1038 	if (fs->object != fs->first_object)
1039 		vm_object_pip_wakeup(fs->object);
1040 	fs->pindex += OFF_TO_IDX(fs->object->backing_object_offset);
1041 	VM_OBJECT_WUNLOCK(fs->object);
1042 	fs->object = next_object;
1043 
1044 	return (true);
1045 }
1046 
1047 static void
1048 vm_fault_zerofill(struct faultstate *fs)
1049 {
1050 
1051 	/*
1052 	 * If there's no object left, fill the page in the top
1053 	 * object with zeros.
1054 	 */
1055 	if (fs->object != fs->first_object) {
1056 		vm_object_pip_wakeup(fs->object);
1057 		fs->object = fs->first_object;
1058 		fs->pindex = fs->first_pindex;
1059 	}
1060 	MPASS(fs->first_m != NULL);
1061 	MPASS(fs->m == NULL);
1062 	fs->m = fs->first_m;
1063 	fs->first_m = NULL;
1064 
1065 	/*
1066 	 * Zero the page if necessary and mark it valid.
1067 	 */
1068 	if ((fs->m->flags & PG_ZERO) == 0) {
1069 		pmap_zero_page(fs->m);
1070 	} else {
1071 		VM_CNT_INC(v_ozfod);
1072 	}
1073 	VM_CNT_INC(v_zfod);
1074 	vm_page_valid(fs->m);
1075 }
1076 
1077 /*
1078  * Initiate page fault after timeout.  Returns true if caller should
1079  * do vm_waitpfault() after the call.
1080  */
1081 static bool
1082 vm_fault_allocate_oom(struct faultstate *fs)
1083 {
1084 	struct timeval now;
1085 
1086 	unlock_and_deallocate(fs);
1087 	if (vm_pfault_oom_attempts < 0)
1088 		return (true);
1089 	if (!fs->oom_started) {
1090 		fs->oom_started = true;
1091 		getmicrotime(&fs->oom_start_time);
1092 		return (true);
1093 	}
1094 
1095 	getmicrotime(&now);
1096 	timevalsub(&now, &fs->oom_start_time);
1097 	if (now.tv_sec < vm_pfault_oom_attempts * vm_pfault_oom_wait)
1098 		return (true);
1099 
1100 	if (bootverbose)
1101 		printf(
1102 	    "proc %d (%s) failed to alloc page on fault, starting OOM\n",
1103 		    curproc->p_pid, curproc->p_comm);
1104 	vm_pageout_oom(VM_OOM_MEM_PF);
1105 	fs->oom_started = false;
1106 	return (false);
1107 }
1108 
1109 /*
1110  * Allocate a page directly or via the object populate method.
1111  */
1112 static int
1113 vm_fault_allocate(struct faultstate *fs)
1114 {
1115 	struct domainset *dset;
1116 	int rv;
1117 
1118 	if ((fs->object->flags & OBJ_SIZEVNLOCK) != 0) {
1119 		rv = vm_fault_lock_vnode(fs, true);
1120 		MPASS(rv == KERN_SUCCESS || rv == KERN_RESOURCE_SHORTAGE);
1121 		if (rv == KERN_RESOURCE_SHORTAGE)
1122 			return (rv);
1123 	}
1124 
1125 	if (fs->pindex >= fs->object->size)
1126 		return (KERN_OUT_OF_BOUNDS);
1127 
1128 	if (fs->object == fs->first_object &&
1129 	    (fs->first_object->flags & OBJ_POPULATE) != 0 &&
1130 	    fs->first_object->shadow_count == 0) {
1131 		rv = vm_fault_populate(fs);
1132 		switch (rv) {
1133 		case KERN_SUCCESS:
1134 		case KERN_FAILURE:
1135 		case KERN_PROTECTION_FAILURE:
1136 		case KERN_RESTART:
1137 			return (rv);
1138 		case KERN_NOT_RECEIVER:
1139 			/*
1140 			 * Pager's populate() method
1141 			 * returned VM_PAGER_BAD.
1142 			 */
1143 			break;
1144 		default:
1145 			panic("inconsistent return codes");
1146 		}
1147 	}
1148 
1149 	/*
1150 	 * Allocate a new page for this object/offset pair.
1151 	 *
1152 	 * If the process has a fatal signal pending, prioritize the allocation
1153 	 * with the expectation that the process will exit shortly and free some
1154 	 * pages.  In particular, the signal may have been posted by the page
1155 	 * daemon in an attempt to resolve an out-of-memory condition.
1156 	 *
1157 	 * The unlocked read of the p_flag is harmless.  At worst, the P_KILLED
1158 	 * might be not observed here, and allocation fails, causing a restart
1159 	 * and new reading of the p_flag.
1160 	 */
1161 	dset = fs->object->domain.dr_policy;
1162 	if (dset == NULL)
1163 		dset = curthread->td_domain.dr_policy;
1164 	if (!vm_page_count_severe_set(&dset->ds_mask) || P_KILLED(curproc)) {
1165 #if VM_NRESERVLEVEL > 0
1166 		vm_object_color(fs->object, atop(fs->vaddr) - fs->pindex);
1167 #endif
1168 		fs->m = vm_page_alloc(fs->object, fs->pindex,
1169 		    P_KILLED(curproc) ? VM_ALLOC_SYSTEM : 0);
1170 	}
1171 	if (fs->m == NULL) {
1172 		if (vm_fault_allocate_oom(fs))
1173 			vm_waitpfault(dset, vm_pfault_oom_wait * hz);
1174 		return (KERN_RESOURCE_SHORTAGE);
1175 	}
1176 	fs->oom_started = false;
1177 
1178 	return (KERN_NOT_RECEIVER);
1179 }
1180 
1181 /*
1182  * Call the pager to retrieve the page if there is a chance
1183  * that the pager has it, and potentially retrieve additional
1184  * pages at the same time.
1185  */
1186 static int
1187 vm_fault_getpages(struct faultstate *fs, int nera, int *behindp, int *aheadp)
1188 {
1189 	vm_offset_t e_end, e_start;
1190 	int ahead, behind, cluster_offset, rv;
1191 	u_char behavior;
1192 
1193 	/*
1194 	 * Prepare for unlocking the map.  Save the map
1195 	 * entry's start and end addresses, which are used to
1196 	 * optimize the size of the pager operation below.
1197 	 * Even if the map entry's addresses change after
1198 	 * unlocking the map, using the saved addresses is
1199 	 * safe.
1200 	 */
1201 	e_start = fs->entry->start;
1202 	e_end = fs->entry->end;
1203 	behavior = vm_map_entry_behavior(fs->entry);
1204 
1205 	/*
1206 	 * Release the map lock before locking the vnode or
1207 	 * sleeping in the pager.  (If the current object has
1208 	 * a shadow, then an earlier iteration of this loop
1209 	 * may have already unlocked the map.)
1210 	 */
1211 	unlock_map(fs);
1212 
1213 	rv = vm_fault_lock_vnode(fs, false);
1214 	MPASS(rv == KERN_SUCCESS || rv == KERN_RESOURCE_SHORTAGE);
1215 	if (rv == KERN_RESOURCE_SHORTAGE)
1216 		return (rv);
1217 	KASSERT(fs->vp == NULL || !fs->map->system_map,
1218 	    ("vm_fault: vnode-backed object mapped by system map"));
1219 
1220 	/*
1221 	 * Page in the requested page and hint the pager,
1222 	 * that it may bring up surrounding pages.
1223 	 */
1224 	if (nera == -1 || behavior == MAP_ENTRY_BEHAV_RANDOM ||
1225 	    P_KILLED(curproc)) {
1226 		behind = 0;
1227 		ahead = 0;
1228 	} else {
1229 		/* Is this a sequential fault? */
1230 		if (nera > 0) {
1231 			behind = 0;
1232 			ahead = nera;
1233 		} else {
1234 			/*
1235 			 * Request a cluster of pages that is
1236 			 * aligned to a VM_FAULT_READ_DEFAULT
1237 			 * page offset boundary within the
1238 			 * object.  Alignment to a page offset
1239 			 * boundary is more likely to coincide
1240 			 * with the underlying file system
1241 			 * block than alignment to a virtual
1242 			 * address boundary.
1243 			 */
1244 			cluster_offset = fs->pindex % VM_FAULT_READ_DEFAULT;
1245 			behind = ulmin(cluster_offset,
1246 			    atop(fs->vaddr - e_start));
1247 			ahead = VM_FAULT_READ_DEFAULT - 1 - cluster_offset;
1248 		}
1249 		ahead = ulmin(ahead, atop(e_end - fs->vaddr) - 1);
1250 	}
1251 	*behindp = behind;
1252 	*aheadp = ahead;
1253 	rv = vm_pager_get_pages(fs->object, &fs->m, 1, behindp, aheadp);
1254 	if (rv == VM_PAGER_OK)
1255 		return (KERN_SUCCESS);
1256 	if (rv == VM_PAGER_ERROR)
1257 		printf("vm_fault: pager read error, pid %d (%s)\n",
1258 		    curproc->p_pid, curproc->p_comm);
1259 	/*
1260 	 * If an I/O error occurred or the requested page was
1261 	 * outside the range of the pager, clean up and return
1262 	 * an error.
1263 	 */
1264 	if (rv == VM_PAGER_ERROR || rv == VM_PAGER_BAD)
1265 		return (KERN_OUT_OF_BOUNDS);
1266 	return (KERN_NOT_RECEIVER);
1267 }
1268 
1269 /*
1270  * Wait/Retry if the page is busy.  We have to do this if the page is
1271  * either exclusive or shared busy because the vm_pager may be using
1272  * read busy for pageouts (and even pageins if it is the vnode pager),
1273  * and we could end up trying to pagein and pageout the same page
1274  * simultaneously.
1275  *
1276  * We can theoretically allow the busy case on a read fault if the page
1277  * is marked valid, but since such pages are typically already pmap'd,
1278  * putting that special case in might be more effort then it is worth.
1279  * We cannot under any circumstances mess around with a shared busied
1280  * page except, perhaps, to pmap it.
1281  */
1282 static void
1283 vm_fault_busy_sleep(struct faultstate *fs)
1284 {
1285 	/*
1286 	 * Reference the page before unlocking and
1287 	 * sleeping so that the page daemon is less
1288 	 * likely to reclaim it.
1289 	 */
1290 	vm_page_aflag_set(fs->m, PGA_REFERENCED);
1291 	if (fs->object != fs->first_object) {
1292 		fault_page_release(&fs->first_m);
1293 		vm_object_pip_wakeup(fs->first_object);
1294 	}
1295 	vm_object_pip_wakeup(fs->object);
1296 	unlock_map(fs);
1297 	if (fs->m == vm_page_lookup(fs->object, fs->pindex))
1298 		vm_page_busy_sleep(fs->m, "vmpfw", false);
1299 	else
1300 		VM_OBJECT_WUNLOCK(fs->object);
1301 	VM_CNT_INC(v_intrans);
1302 	vm_object_deallocate(fs->first_object);
1303 }
1304 
1305 int
1306 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
1307     int fault_flags, vm_page_t *m_hold)
1308 {
1309 	struct faultstate fs;
1310 	int ahead, behind, faultcount;
1311 	int nera, result, rv;
1312 	bool dead, hardfault;
1313 
1314 	VM_CNT_INC(v_vm_faults);
1315 
1316 	if ((curthread->td_pflags & TDP_NOFAULTING) != 0)
1317 		return (KERN_PROTECTION_FAILURE);
1318 
1319 	fs.vp = NULL;
1320 	fs.vaddr = vaddr;
1321 	fs.m_hold = m_hold;
1322 	fs.fault_flags = fault_flags;
1323 	fs.map = map;
1324 	fs.lookup_still_valid = false;
1325 	fs.oom_started = false;
1326 	faultcount = 0;
1327 	nera = -1;
1328 	hardfault = false;
1329 
1330 RetryFault:
1331 	fs.fault_type = fault_type;
1332 
1333 	/*
1334 	 * Find the backing store object and offset into it to begin the
1335 	 * search.
1336 	 */
1337 	result = vm_fault_lookup(&fs);
1338 	if (result != KERN_SUCCESS) {
1339 		if (result == KERN_RESOURCE_SHORTAGE)
1340 			goto RetryFault;
1341 		return (result);
1342 	}
1343 
1344 	/*
1345 	 * Try to avoid lock contention on the top-level object through
1346 	 * special-case handling of some types of page faults, specifically,
1347 	 * those that are mapping an existing page from the top-level object.
1348 	 * Under this condition, a read lock on the object suffices, allowing
1349 	 * multiple page faults of a similar type to run in parallel.
1350 	 */
1351 	if (fs.vp == NULL /* avoid locked vnode leak */ &&
1352 	    (fs.entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) == 0 &&
1353 	    (fs.fault_flags & (VM_FAULT_WIRE | VM_FAULT_DIRTY)) == 0) {
1354 		VM_OBJECT_RLOCK(fs.first_object);
1355 		rv = vm_fault_soft_fast(&fs);
1356 		if (rv == KERN_SUCCESS)
1357 			return (rv);
1358 		if (!VM_OBJECT_TRYUPGRADE(fs.first_object)) {
1359 			VM_OBJECT_RUNLOCK(fs.first_object);
1360 			VM_OBJECT_WLOCK(fs.first_object);
1361 		}
1362 	} else {
1363 		VM_OBJECT_WLOCK(fs.first_object);
1364 	}
1365 
1366 	/*
1367 	 * Make a reference to this object to prevent its disposal while we
1368 	 * are messing with it.  Once we have the reference, the map is free
1369 	 * to be diddled.  Since objects reference their shadows (and copies),
1370 	 * they will stay around as well.
1371 	 *
1372 	 * Bump the paging-in-progress count to prevent size changes (e.g.
1373 	 * truncation operations) during I/O.
1374 	 */
1375 	vm_object_reference_locked(fs.first_object);
1376 	vm_object_pip_add(fs.first_object, 1);
1377 
1378 	fs.m_cow = fs.m = fs.first_m = NULL;
1379 
1380 	/*
1381 	 * Search for the page at object/offset.
1382 	 */
1383 	fs.object = fs.first_object;
1384 	fs.pindex = fs.first_pindex;
1385 
1386 	if ((fs.entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) != 0) {
1387 		rv = vm_fault_allocate(&fs);
1388 		switch (rv) {
1389 		case KERN_RESTART:
1390 			unlock_and_deallocate(&fs);
1391 			/* FALLTHROUGH */
1392 		case KERN_RESOURCE_SHORTAGE:
1393 			goto RetryFault;
1394 		case KERN_SUCCESS:
1395 		case KERN_FAILURE:
1396 		case KERN_PROTECTION_FAILURE:
1397 		case KERN_OUT_OF_BOUNDS:
1398 			unlock_and_deallocate(&fs);
1399 			return (rv);
1400 		case KERN_NOT_RECEIVER:
1401 			break;
1402 		default:
1403 			panic("vm_fault: Unhandled rv %d", rv);
1404 		}
1405 	}
1406 
1407 	while (TRUE) {
1408 		KASSERT(fs.m == NULL,
1409 		    ("page still set %p at loop start", fs.m));
1410 		/*
1411 		 * If the object is marked for imminent termination,
1412 		 * we retry here, since the collapse pass has raced
1413 		 * with us.  Otherwise, if we see terminally dead
1414 		 * object, return fail.
1415 		 */
1416 		if ((fs.object->flags & OBJ_DEAD) != 0) {
1417 			dead = fs.object->type == OBJT_DEAD;
1418 			unlock_and_deallocate(&fs);
1419 			if (dead)
1420 				return (KERN_PROTECTION_FAILURE);
1421 			pause("vmf_de", 1);
1422 			goto RetryFault;
1423 		}
1424 
1425 		/*
1426 		 * See if page is resident
1427 		 */
1428 		fs.m = vm_page_lookup(fs.object, fs.pindex);
1429 		if (fs.m != NULL) {
1430 			if (vm_page_tryxbusy(fs.m) == 0) {
1431 				vm_fault_busy_sleep(&fs);
1432 				goto RetryFault;
1433 			}
1434 
1435 			/*
1436 			 * The page is marked busy for other processes and the
1437 			 * pagedaemon.  If it still is completely valid we
1438 			 * are done.
1439 			 */
1440 			if (vm_page_all_valid(fs.m)) {
1441 				VM_OBJECT_WUNLOCK(fs.object);
1442 				break; /* break to PAGE HAS BEEN FOUND. */
1443 			}
1444 		}
1445 		VM_OBJECT_ASSERT_WLOCKED(fs.object);
1446 
1447 		/*
1448 		 * Page is not resident.  If the pager might contain the page
1449 		 * or this is the beginning of the search, allocate a new
1450 		 * page.  (Default objects are zero-fill, so there is no real
1451 		 * pager for them.)
1452 		 */
1453 		if (fs.m == NULL && (fs.object->type != OBJT_DEFAULT ||
1454 		    fs.object == fs.first_object)) {
1455 			rv = vm_fault_allocate(&fs);
1456 			switch (rv) {
1457 			case KERN_RESTART:
1458 				unlock_and_deallocate(&fs);
1459 				/* FALLTHROUGH */
1460 			case KERN_RESOURCE_SHORTAGE:
1461 				goto RetryFault;
1462 			case KERN_SUCCESS:
1463 			case KERN_FAILURE:
1464 			case KERN_PROTECTION_FAILURE:
1465 			case KERN_OUT_OF_BOUNDS:
1466 				unlock_and_deallocate(&fs);
1467 				return (rv);
1468 			case KERN_NOT_RECEIVER:
1469 				break;
1470 			default:
1471 				panic("vm_fault: Unhandled rv %d", rv);
1472 			}
1473 		}
1474 
1475 		/*
1476 		 * Default objects have no pager so no exclusive busy exists
1477 		 * to protect this page in the chain.  Skip to the next
1478 		 * object without dropping the lock to preserve atomicity of
1479 		 * shadow faults.
1480 		 */
1481 		if (fs.object->type != OBJT_DEFAULT) {
1482 			/*
1483 			 * At this point, we have either allocated a new page
1484 			 * or found an existing page that is only partially
1485 			 * valid.
1486 			 *
1487 			 * We hold a reference on the current object and the
1488 			 * page is exclusive busied.  The exclusive busy
1489 			 * prevents simultaneous faults and collapses while
1490 			 * the object lock is dropped.
1491 		 	 */
1492 			VM_OBJECT_WUNLOCK(fs.object);
1493 
1494 			/*
1495 			 * If the pager for the current object might have
1496 			 * the page, then determine the number of additional
1497 			 * pages to read and potentially reprioritize
1498 			 * previously read pages for earlier reclamation.
1499 			 * These operations should only be performed once per
1500 			 * page fault.  Even if the current pager doesn't
1501 			 * have the page, the number of additional pages to
1502 			 * read will apply to subsequent objects in the
1503 			 * shadow chain.
1504 			 */
1505 			if (nera == -1 && !P_KILLED(curproc))
1506 				nera = vm_fault_readahead(&fs);
1507 
1508 			rv = vm_fault_getpages(&fs, nera, &behind, &ahead);
1509 			if (rv == KERN_SUCCESS) {
1510 				faultcount = behind + 1 + ahead;
1511 				hardfault = true;
1512 				break; /* break to PAGE HAS BEEN FOUND. */
1513 			}
1514 			if (rv == KERN_RESOURCE_SHORTAGE)
1515 				goto RetryFault;
1516 			VM_OBJECT_WLOCK(fs.object);
1517 			if (rv == KERN_OUT_OF_BOUNDS) {
1518 				fault_page_free(&fs.m);
1519 				unlock_and_deallocate(&fs);
1520 				return (rv);
1521 			}
1522 		}
1523 
1524 		/*
1525 		 * The page was not found in the current object.  Try to
1526 		 * traverse into a backing object or zero fill if none is
1527 		 * found.
1528 		 */
1529 		if (vm_fault_next(&fs))
1530 			continue;
1531 		if ((fs.fault_flags & VM_FAULT_NOFILL) != 0) {
1532 			if (fs.first_object == fs.object)
1533 				fault_page_free(&fs.first_m);
1534 			unlock_and_deallocate(&fs);
1535 			return (KERN_OUT_OF_BOUNDS);
1536 		}
1537 		VM_OBJECT_WUNLOCK(fs.object);
1538 		vm_fault_zerofill(&fs);
1539 		/* Don't try to prefault neighboring pages. */
1540 		faultcount = 1;
1541 		break;	/* break to PAGE HAS BEEN FOUND. */
1542 	}
1543 
1544 	/*
1545 	 * PAGE HAS BEEN FOUND.  A valid page has been found and exclusively
1546 	 * busied.  The object lock must no longer be held.
1547 	 */
1548 	vm_page_assert_xbusied(fs.m);
1549 	VM_OBJECT_ASSERT_UNLOCKED(fs.object);
1550 
1551 	/*
1552 	 * If the page is being written, but isn't already owned by the
1553 	 * top-level object, we have to copy it into a new page owned by the
1554 	 * top-level object.
1555 	 */
1556 	if (fs.object != fs.first_object) {
1557 		/*
1558 		 * We only really need to copy if we want to write it.
1559 		 */
1560 		if ((fs.fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1561 			vm_fault_cow(&fs);
1562 			/*
1563 			 * We only try to prefault read-only mappings to the
1564 			 * neighboring pages when this copy-on-write fault is
1565 			 * a hard fault.  In other cases, trying to prefault
1566 			 * is typically wasted effort.
1567 			 */
1568 			if (faultcount == 0)
1569 				faultcount = 1;
1570 
1571 		} else {
1572 			fs.prot &= ~VM_PROT_WRITE;
1573 		}
1574 	}
1575 
1576 	/*
1577 	 * We must verify that the maps have not changed since our last
1578 	 * lookup.
1579 	 */
1580 	if (!fs.lookup_still_valid) {
1581 		result = vm_fault_relookup(&fs);
1582 		if (result != KERN_SUCCESS) {
1583 			fault_deallocate(&fs);
1584 			if (result == KERN_RESTART)
1585 				goto RetryFault;
1586 			return (result);
1587 		}
1588 	}
1589 	VM_OBJECT_ASSERT_UNLOCKED(fs.object);
1590 
1591 	/*
1592 	 * If the page was filled by a pager, save the virtual address that
1593 	 * should be faulted on next under a sequential access pattern to the
1594 	 * map entry.  A read lock on the map suffices to update this address
1595 	 * safely.
1596 	 */
1597 	if (hardfault)
1598 		fs.entry->next_read = vaddr + ptoa(ahead) + PAGE_SIZE;
1599 
1600 	/*
1601 	 * Page must be completely valid or it is not fit to
1602 	 * map into user space.  vm_pager_get_pages() ensures this.
1603 	 */
1604 	vm_page_assert_xbusied(fs.m);
1605 	KASSERT(vm_page_all_valid(fs.m),
1606 	    ("vm_fault: page %p partially invalid", fs.m));
1607 
1608 	vm_fault_dirty(&fs, fs.m);
1609 
1610 	/*
1611 	 * Put this page into the physical map.  We had to do the unlock above
1612 	 * because pmap_enter() may sleep.  We don't put the page
1613 	 * back on the active queue until later so that the pageout daemon
1614 	 * won't find it (yet).
1615 	 */
1616 	pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot,
1617 	    fs.fault_type | (fs.wired ? PMAP_ENTER_WIRED : 0), 0);
1618 	if (faultcount != 1 && (fs.fault_flags & VM_FAULT_WIRE) == 0 &&
1619 	    fs.wired == 0)
1620 		vm_fault_prefault(&fs, vaddr,
1621 		    faultcount > 0 ? behind : PFBAK,
1622 		    faultcount > 0 ? ahead : PFFOR, false);
1623 
1624 	/*
1625 	 * If the page is not wired down, then put it where the pageout daemon
1626 	 * can find it.
1627 	 */
1628 	if ((fs.fault_flags & VM_FAULT_WIRE) != 0)
1629 		vm_page_wire(fs.m);
1630 	else
1631 		vm_page_activate(fs.m);
1632 	if (fs.m_hold != NULL) {
1633 		(*fs.m_hold) = fs.m;
1634 		vm_page_wire(fs.m);
1635 	}
1636 	vm_page_xunbusy(fs.m);
1637 	fs.m = NULL;
1638 
1639 	/*
1640 	 * Unlock everything, and return
1641 	 */
1642 	fault_deallocate(&fs);
1643 	if (hardfault) {
1644 		VM_CNT_INC(v_io_faults);
1645 		curthread->td_ru.ru_majflt++;
1646 #ifdef RACCT
1647 		if (racct_enable && fs.object->type == OBJT_VNODE) {
1648 			PROC_LOCK(curproc);
1649 			if ((fs.fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1650 				racct_add_force(curproc, RACCT_WRITEBPS,
1651 				    PAGE_SIZE + behind * PAGE_SIZE);
1652 				racct_add_force(curproc, RACCT_WRITEIOPS, 1);
1653 			} else {
1654 				racct_add_force(curproc, RACCT_READBPS,
1655 				    PAGE_SIZE + ahead * PAGE_SIZE);
1656 				racct_add_force(curproc, RACCT_READIOPS, 1);
1657 			}
1658 			PROC_UNLOCK(curproc);
1659 		}
1660 #endif
1661 	} else
1662 		curthread->td_ru.ru_minflt++;
1663 
1664 	return (KERN_SUCCESS);
1665 }
1666 
1667 /*
1668  * Speed up the reclamation of pages that precede the faulting pindex within
1669  * the first object of the shadow chain.  Essentially, perform the equivalent
1670  * to madvise(..., MADV_DONTNEED) on a large cluster of pages that precedes
1671  * the faulting pindex by the cluster size when the pages read by vm_fault()
1672  * cross a cluster-size boundary.  The cluster size is the greater of the
1673  * smallest superpage size and VM_FAULT_DONTNEED_MIN.
1674  *
1675  * When "fs->first_object" is a shadow object, the pages in the backing object
1676  * that precede the faulting pindex are deactivated by vm_fault().  So, this
1677  * function must only be concerned with pages in the first object.
1678  */
1679 static void
1680 vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr, int ahead)
1681 {
1682 	vm_map_entry_t entry;
1683 	vm_object_t first_object, object;
1684 	vm_offset_t end, start;
1685 	vm_page_t m, m_next;
1686 	vm_pindex_t pend, pstart;
1687 	vm_size_t size;
1688 
1689 	object = fs->object;
1690 	VM_OBJECT_ASSERT_UNLOCKED(object);
1691 	first_object = fs->first_object;
1692 	/* Neither fictitious nor unmanaged pages can be reclaimed. */
1693 	if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) {
1694 		VM_OBJECT_RLOCK(first_object);
1695 		size = VM_FAULT_DONTNEED_MIN;
1696 		if (MAXPAGESIZES > 1 && size < pagesizes[1])
1697 			size = pagesizes[1];
1698 		end = rounddown2(vaddr, size);
1699 		if (vaddr - end >= size - PAGE_SIZE - ptoa(ahead) &&
1700 		    (entry = fs->entry)->start < end) {
1701 			if (end - entry->start < size)
1702 				start = entry->start;
1703 			else
1704 				start = end - size;
1705 			pmap_advise(fs->map->pmap, start, end, MADV_DONTNEED);
1706 			pstart = OFF_TO_IDX(entry->offset) + atop(start -
1707 			    entry->start);
1708 			m_next = vm_page_find_least(first_object, pstart);
1709 			pend = OFF_TO_IDX(entry->offset) + atop(end -
1710 			    entry->start);
1711 			while ((m = m_next) != NULL && m->pindex < pend) {
1712 				m_next = TAILQ_NEXT(m, listq);
1713 				if (!vm_page_all_valid(m) ||
1714 				    vm_page_busied(m))
1715 					continue;
1716 
1717 				/*
1718 				 * Don't clear PGA_REFERENCED, since it would
1719 				 * likely represent a reference by a different
1720 				 * process.
1721 				 *
1722 				 * Typically, at this point, prefetched pages
1723 				 * are still in the inactive queue.  Only
1724 				 * pages that triggered page faults are in the
1725 				 * active queue.  The test for whether the page
1726 				 * is in the inactive queue is racy; in the
1727 				 * worst case we will requeue the page
1728 				 * unnecessarily.
1729 				 */
1730 				if (!vm_page_inactive(m))
1731 					vm_page_deactivate(m);
1732 			}
1733 		}
1734 		VM_OBJECT_RUNLOCK(first_object);
1735 	}
1736 }
1737 
1738 /*
1739  * vm_fault_prefault provides a quick way of clustering
1740  * pagefaults into a processes address space.  It is a "cousin"
1741  * of vm_map_pmap_enter, except it runs at page fault time instead
1742  * of mmap time.
1743  */
1744 static void
1745 vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
1746     int backward, int forward, bool obj_locked)
1747 {
1748 	pmap_t pmap;
1749 	vm_map_entry_t entry;
1750 	vm_object_t backing_object, lobject;
1751 	vm_offset_t addr, starta;
1752 	vm_pindex_t pindex;
1753 	vm_page_t m;
1754 	int i;
1755 
1756 	pmap = fs->map->pmap;
1757 	if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
1758 		return;
1759 
1760 	entry = fs->entry;
1761 
1762 	if (addra < backward * PAGE_SIZE) {
1763 		starta = entry->start;
1764 	} else {
1765 		starta = addra - backward * PAGE_SIZE;
1766 		if (starta < entry->start)
1767 			starta = entry->start;
1768 	}
1769 
1770 	/*
1771 	 * Generate the sequence of virtual addresses that are candidates for
1772 	 * prefaulting in an outward spiral from the faulting virtual address,
1773 	 * "addra".  Specifically, the sequence is "addra - PAGE_SIZE", "addra
1774 	 * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ...
1775 	 * If the candidate address doesn't have a backing physical page, then
1776 	 * the loop immediately terminates.
1777 	 */
1778 	for (i = 0; i < 2 * imax(backward, forward); i++) {
1779 		addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE :
1780 		    PAGE_SIZE);
1781 		if (addr > addra + forward * PAGE_SIZE)
1782 			addr = 0;
1783 
1784 		if (addr < starta || addr >= entry->end)
1785 			continue;
1786 
1787 		if (!pmap_is_prefaultable(pmap, addr))
1788 			continue;
1789 
1790 		pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1791 		lobject = entry->object.vm_object;
1792 		if (!obj_locked)
1793 			VM_OBJECT_RLOCK(lobject);
1794 		while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1795 		    lobject->type == OBJT_DEFAULT &&
1796 		    (backing_object = lobject->backing_object) != NULL) {
1797 			KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1798 			    0, ("vm_fault_prefault: unaligned object offset"));
1799 			pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1800 			VM_OBJECT_RLOCK(backing_object);
1801 			if (!obj_locked || lobject != entry->object.vm_object)
1802 				VM_OBJECT_RUNLOCK(lobject);
1803 			lobject = backing_object;
1804 		}
1805 		if (m == NULL) {
1806 			if (!obj_locked || lobject != entry->object.vm_object)
1807 				VM_OBJECT_RUNLOCK(lobject);
1808 			break;
1809 		}
1810 		if (vm_page_all_valid(m) &&
1811 		    (m->flags & PG_FICTITIOUS) == 0)
1812 			pmap_enter_quick(pmap, addr, m, entry->protection);
1813 		if (!obj_locked || lobject != entry->object.vm_object)
1814 			VM_OBJECT_RUNLOCK(lobject);
1815 	}
1816 }
1817 
1818 /*
1819  * Hold each of the physical pages that are mapped by the specified range of
1820  * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
1821  * and allow the specified types of access, "prot".  If all of the implied
1822  * pages are successfully held, then the number of held pages is returned
1823  * together with pointers to those pages in the array "ma".  However, if any
1824  * of the pages cannot be held, -1 is returned.
1825  */
1826 int
1827 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
1828     vm_prot_t prot, vm_page_t *ma, int max_count)
1829 {
1830 	vm_offset_t end, va;
1831 	vm_page_t *mp;
1832 	int count;
1833 	boolean_t pmap_failed;
1834 
1835 	if (len == 0)
1836 		return (0);
1837 	end = round_page(addr + len);
1838 	addr = trunc_page(addr);
1839 
1840 	if (!vm_map_range_valid(map, addr, end))
1841 		return (-1);
1842 
1843 	if (atop(end - addr) > max_count)
1844 		panic("vm_fault_quick_hold_pages: count > max_count");
1845 	count = atop(end - addr);
1846 
1847 	/*
1848 	 * Most likely, the physical pages are resident in the pmap, so it is
1849 	 * faster to try pmap_extract_and_hold() first.
1850 	 */
1851 	pmap_failed = FALSE;
1852 	for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) {
1853 		*mp = pmap_extract_and_hold(map->pmap, va, prot);
1854 		if (*mp == NULL)
1855 			pmap_failed = TRUE;
1856 		else if ((prot & VM_PROT_WRITE) != 0 &&
1857 		    (*mp)->dirty != VM_PAGE_BITS_ALL) {
1858 			/*
1859 			 * Explicitly dirty the physical page.  Otherwise, the
1860 			 * caller's changes may go unnoticed because they are
1861 			 * performed through an unmanaged mapping or by a DMA
1862 			 * operation.
1863 			 *
1864 			 * The object lock is not held here.
1865 			 * See vm_page_clear_dirty_mask().
1866 			 */
1867 			vm_page_dirty(*mp);
1868 		}
1869 	}
1870 	if (pmap_failed) {
1871 		/*
1872 		 * One or more pages could not be held by the pmap.  Either no
1873 		 * page was mapped at the specified virtual address or that
1874 		 * mapping had insufficient permissions.  Attempt to fault in
1875 		 * and hold these pages.
1876 		 *
1877 		 * If vm_fault_disable_pagefaults() was called,
1878 		 * i.e., TDP_NOFAULTING is set, we must not sleep nor
1879 		 * acquire MD VM locks, which means we must not call
1880 		 * vm_fault().  Some (out of tree) callers mark
1881 		 * too wide a code area with vm_fault_disable_pagefaults()
1882 		 * already, use the VM_PROT_QUICK_NOFAULT flag to request
1883 		 * the proper behaviour explicitly.
1884 		 */
1885 		if ((prot & VM_PROT_QUICK_NOFAULT) != 0 &&
1886 		    (curthread->td_pflags & TDP_NOFAULTING) != 0)
1887 			goto error;
1888 		for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE)
1889 			if (*mp == NULL && vm_fault(map, va, prot,
1890 			    VM_FAULT_NORMAL, mp) != KERN_SUCCESS)
1891 				goto error;
1892 	}
1893 	return (count);
1894 error:
1895 	for (mp = ma; mp < ma + count; mp++)
1896 		if (*mp != NULL)
1897 			vm_page_unwire(*mp, PQ_INACTIVE);
1898 	return (-1);
1899 }
1900 
1901 /*
1902  *	Routine:
1903  *		vm_fault_copy_entry
1904  *	Function:
1905  *		Create new shadow object backing dst_entry with private copy of
1906  *		all underlying pages. When src_entry is equal to dst_entry,
1907  *		function implements COW for wired-down map entry. Otherwise,
1908  *		it forks wired entry into dst_map.
1909  *
1910  *	In/out conditions:
1911  *		The source and destination maps must be locked for write.
1912  *		The source map entry must be wired down (or be a sharing map
1913  *		entry corresponding to a main map entry that is wired down).
1914  */
1915 void
1916 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1917     vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1918     vm_ooffset_t *fork_charge)
1919 {
1920 	vm_object_t backing_object, dst_object, object, src_object;
1921 	vm_pindex_t dst_pindex, pindex, src_pindex;
1922 	vm_prot_t access, prot;
1923 	vm_offset_t vaddr;
1924 	vm_page_t dst_m;
1925 	vm_page_t src_m;
1926 	boolean_t upgrade;
1927 
1928 #ifdef	lint
1929 	src_map++;
1930 #endif	/* lint */
1931 
1932 	upgrade = src_entry == dst_entry;
1933 	access = prot = dst_entry->protection;
1934 
1935 	src_object = src_entry->object.vm_object;
1936 	src_pindex = OFF_TO_IDX(src_entry->offset);
1937 
1938 	if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
1939 		dst_object = src_object;
1940 		vm_object_reference(dst_object);
1941 	} else {
1942 		/*
1943 		 * Create the top-level object for the destination entry.
1944 		 * Doesn't actually shadow anything - we copy the pages
1945 		 * directly.
1946 		 */
1947 		dst_object = vm_object_allocate_anon(atop(dst_entry->end -
1948 		    dst_entry->start), NULL, NULL, 0);
1949 #if VM_NRESERVLEVEL > 0
1950 		dst_object->flags |= OBJ_COLORED;
1951 		dst_object->pg_color = atop(dst_entry->start);
1952 #endif
1953 		dst_object->domain = src_object->domain;
1954 		dst_object->charge = dst_entry->end - dst_entry->start;
1955 	}
1956 
1957 	VM_OBJECT_WLOCK(dst_object);
1958 	KASSERT(upgrade || dst_entry->object.vm_object == NULL,
1959 	    ("vm_fault_copy_entry: vm_object not NULL"));
1960 	if (src_object != dst_object) {
1961 		dst_entry->object.vm_object = dst_object;
1962 		dst_entry->offset = 0;
1963 		dst_entry->eflags &= ~MAP_ENTRY_VN_EXEC;
1964 	}
1965 	if (fork_charge != NULL) {
1966 		KASSERT(dst_entry->cred == NULL,
1967 		    ("vm_fault_copy_entry: leaked swp charge"));
1968 		dst_object->cred = curthread->td_ucred;
1969 		crhold(dst_object->cred);
1970 		*fork_charge += dst_object->charge;
1971 	} else if ((dst_object->type == OBJT_DEFAULT ||
1972 	    (dst_object->flags & OBJ_SWAP) != 0) &&
1973 	    dst_object->cred == NULL) {
1974 		KASSERT(dst_entry->cred != NULL, ("no cred for entry %p",
1975 		    dst_entry));
1976 		dst_object->cred = dst_entry->cred;
1977 		dst_entry->cred = NULL;
1978 	}
1979 
1980 	/*
1981 	 * If not an upgrade, then enter the mappings in the pmap as
1982 	 * read and/or execute accesses.  Otherwise, enter them as
1983 	 * write accesses.
1984 	 *
1985 	 * A writeable large page mapping is only created if all of
1986 	 * the constituent small page mappings are modified. Marking
1987 	 * PTEs as modified on inception allows promotion to happen
1988 	 * without taking potentially large number of soft faults.
1989 	 */
1990 	if (!upgrade)
1991 		access &= ~VM_PROT_WRITE;
1992 
1993 	/*
1994 	 * Loop through all of the virtual pages within the entry's
1995 	 * range, copying each page from the source object to the
1996 	 * destination object.  Since the source is wired, those pages
1997 	 * must exist.  In contrast, the destination is pageable.
1998 	 * Since the destination object doesn't share any backing storage
1999 	 * with the source object, all of its pages must be dirtied,
2000 	 * regardless of whether they can be written.
2001 	 */
2002 	for (vaddr = dst_entry->start, dst_pindex = 0;
2003 	    vaddr < dst_entry->end;
2004 	    vaddr += PAGE_SIZE, dst_pindex++) {
2005 again:
2006 		/*
2007 		 * Find the page in the source object, and copy it in.
2008 		 * Because the source is wired down, the page will be
2009 		 * in memory.
2010 		 */
2011 		if (src_object != dst_object)
2012 			VM_OBJECT_RLOCK(src_object);
2013 		object = src_object;
2014 		pindex = src_pindex + dst_pindex;
2015 		while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
2016 		    (backing_object = object->backing_object) != NULL) {
2017 			/*
2018 			 * Unless the source mapping is read-only or
2019 			 * it is presently being upgraded from
2020 			 * read-only, the first object in the shadow
2021 			 * chain should provide all of the pages.  In
2022 			 * other words, this loop body should never be
2023 			 * executed when the source mapping is already
2024 			 * read/write.
2025 			 */
2026 			KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 ||
2027 			    upgrade,
2028 			    ("vm_fault_copy_entry: main object missing page"));
2029 
2030 			VM_OBJECT_RLOCK(backing_object);
2031 			pindex += OFF_TO_IDX(object->backing_object_offset);
2032 			if (object != dst_object)
2033 				VM_OBJECT_RUNLOCK(object);
2034 			object = backing_object;
2035 		}
2036 		KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing"));
2037 
2038 		if (object != dst_object) {
2039 			/*
2040 			 * Allocate a page in the destination object.
2041 			 */
2042 			dst_m = vm_page_alloc(dst_object, (src_object ==
2043 			    dst_object ? src_pindex : 0) + dst_pindex,
2044 			    VM_ALLOC_NORMAL);
2045 			if (dst_m == NULL) {
2046 				VM_OBJECT_WUNLOCK(dst_object);
2047 				VM_OBJECT_RUNLOCK(object);
2048 				vm_wait(dst_object);
2049 				VM_OBJECT_WLOCK(dst_object);
2050 				goto again;
2051 			}
2052 			pmap_copy_page(src_m, dst_m);
2053 			VM_OBJECT_RUNLOCK(object);
2054 			dst_m->dirty = dst_m->valid = src_m->valid;
2055 		} else {
2056 			dst_m = src_m;
2057 			if (vm_page_busy_acquire(dst_m, VM_ALLOC_WAITFAIL) == 0)
2058 				goto again;
2059 			if (dst_m->pindex >= dst_object->size) {
2060 				/*
2061 				 * We are upgrading.  Index can occur
2062 				 * out of bounds if the object type is
2063 				 * vnode and the file was truncated.
2064 				 */
2065 				vm_page_xunbusy(dst_m);
2066 				break;
2067 			}
2068 		}
2069 		VM_OBJECT_WUNLOCK(dst_object);
2070 
2071 		/*
2072 		 * Enter it in the pmap. If a wired, copy-on-write
2073 		 * mapping is being replaced by a write-enabled
2074 		 * mapping, then wire that new mapping.
2075 		 *
2076 		 * The page can be invalid if the user called
2077 		 * msync(MS_INVALIDATE) or truncated the backing vnode
2078 		 * or shared memory object.  In this case, do not
2079 		 * insert it into pmap, but still do the copy so that
2080 		 * all copies of the wired map entry have similar
2081 		 * backing pages.
2082 		 */
2083 		if (vm_page_all_valid(dst_m)) {
2084 			pmap_enter(dst_map->pmap, vaddr, dst_m, prot,
2085 			    access | (upgrade ? PMAP_ENTER_WIRED : 0), 0);
2086 		}
2087 
2088 		/*
2089 		 * Mark it no longer busy, and put it on the active list.
2090 		 */
2091 		VM_OBJECT_WLOCK(dst_object);
2092 
2093 		if (upgrade) {
2094 			if (src_m != dst_m) {
2095 				vm_page_unwire(src_m, PQ_INACTIVE);
2096 				vm_page_wire(dst_m);
2097 			} else {
2098 				KASSERT(vm_page_wired(dst_m),
2099 				    ("dst_m %p is not wired", dst_m));
2100 			}
2101 		} else {
2102 			vm_page_activate(dst_m);
2103 		}
2104 		vm_page_xunbusy(dst_m);
2105 	}
2106 	VM_OBJECT_WUNLOCK(dst_object);
2107 	if (upgrade) {
2108 		dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
2109 		vm_object_deallocate(src_object);
2110 	}
2111 }
2112 
2113 /*
2114  * Block entry into the machine-independent layer's page fault handler by
2115  * the calling thread.  Subsequent calls to vm_fault() by that thread will
2116  * return KERN_PROTECTION_FAILURE.  Enable machine-dependent handling of
2117  * spurious page faults.
2118  */
2119 int
2120 vm_fault_disable_pagefaults(void)
2121 {
2122 
2123 	return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR));
2124 }
2125 
2126 void
2127 vm_fault_enable_pagefaults(int save)
2128 {
2129 
2130 	curthread_pflags_restore(save);
2131 }
2132