xref: /freebsd/sys/vm/vm_fault.c (revision 5956d97f4b3204318ceb6aa9c77bd0bc6ea87a41)
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/pctrie.h>
89 #include <sys/proc.h>
90 #include <sys/racct.h>
91 #include <sys/refcount.h>
92 #include <sys/resourcevar.h>
93 #include <sys/rwlock.h>
94 #include <sys/signalvar.h>
95 #include <sys/sysctl.h>
96 #include <sys/sysent.h>
97 #include <sys/vmmeter.h>
98 #include <sys/vnode.h>
99 #ifdef KTRACE
100 #include <sys/ktrace.h>
101 #endif
102 
103 #include <vm/vm.h>
104 #include <vm/vm_param.h>
105 #include <vm/pmap.h>
106 #include <vm/vm_map.h>
107 #include <vm/vm_object.h>
108 #include <vm/vm_page.h>
109 #include <vm/vm_pageout.h>
110 #include <vm/vm_kern.h>
111 #include <vm/vm_pager.h>
112 #include <vm/vm_extern.h>
113 #include <vm/vm_reserv.h>
114 
115 #define PFBAK 4
116 #define PFFOR 4
117 
118 #define	VM_FAULT_READ_DEFAULT	(1 + VM_FAULT_READ_AHEAD_INIT)
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 	boolean_t	wired;
130 
131 	/* Control state. */
132 	struct timeval	oom_start_time;
133 	bool		oom_started;
134 	int		nera;
135 
136 	/* Page reference for cow. */
137 	vm_page_t m_cow;
138 
139 	/* Current object. */
140 	vm_object_t	object;
141 	vm_pindex_t	pindex;
142 	vm_page_t	m;
143 
144 	/* Top-level map object. */
145 	vm_object_t	first_object;
146 	vm_pindex_t	first_pindex;
147 	vm_page_t	first_m;
148 
149 	/* Map state. */
150 	vm_map_t	map;
151 	vm_map_entry_t	entry;
152 	int		map_generation;
153 	bool		lookup_still_valid;
154 
155 	/* Vnode if locked. */
156 	struct vnode	*vp;
157 };
158 
159 /*
160  * Return codes for internal fault routines.
161  */
162 enum fault_status {
163 	FAULT_SUCCESS = 1,	/* Return success to user. */
164 	FAULT_FAILURE,		/* Return failure to user. */
165 	FAULT_CONTINUE,		/* Continue faulting. */
166 	FAULT_RESTART,		/* Restart fault. */
167 	FAULT_OUT_OF_BOUNDS,	/* Invalid address for pager. */
168 	FAULT_HARD,		/* Performed I/O. */
169 	FAULT_SOFT,		/* Found valid page. */
170 	FAULT_PROTECTION_FAILURE, /* Invalid access. */
171 };
172 
173 static void vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr,
174 	    int ahead);
175 static void vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
176 	    int backward, int forward, bool obj_locked);
177 
178 static int vm_pfault_oom_attempts = 3;
179 SYSCTL_INT(_vm, OID_AUTO, pfault_oom_attempts, CTLFLAG_RWTUN,
180     &vm_pfault_oom_attempts, 0,
181     "Number of page allocation attempts in page fault handler before it "
182     "triggers OOM handling");
183 
184 static int vm_pfault_oom_wait = 10;
185 SYSCTL_INT(_vm, OID_AUTO, pfault_oom_wait, CTLFLAG_RWTUN,
186     &vm_pfault_oom_wait, 0,
187     "Number of seconds to wait for free pages before retrying "
188     "the page fault handler");
189 
190 static inline void
191 fault_page_release(vm_page_t *mp)
192 {
193 	vm_page_t m;
194 
195 	m = *mp;
196 	if (m != NULL) {
197 		/*
198 		 * We are likely to loop around again and attempt to busy
199 		 * this page.  Deactivating it leaves it available for
200 		 * pageout while optimizing fault restarts.
201 		 */
202 		vm_page_deactivate(m);
203 		vm_page_xunbusy(m);
204 		*mp = NULL;
205 	}
206 }
207 
208 static inline void
209 fault_page_free(vm_page_t *mp)
210 {
211 	vm_page_t m;
212 
213 	m = *mp;
214 	if (m != NULL) {
215 		VM_OBJECT_ASSERT_WLOCKED(m->object);
216 		if (!vm_page_wired(m))
217 			vm_page_free(m);
218 		else
219 			vm_page_xunbusy(m);
220 		*mp = NULL;
221 	}
222 }
223 
224 /*
225  * Return true if a vm_pager_get_pages() call is needed in order to check
226  * whether the pager might have a particular page, false if it can be determined
227  * immediately that the pager can not have a copy.  For swap objects, this can
228  * be checked quickly.
229  */
230 static inline bool
231 fault_object_needs_getpages(vm_object_t object)
232 {
233 	VM_OBJECT_ASSERT_LOCKED(object);
234 
235 	return ((object->flags & OBJ_SWAP) == 0 ||
236 	    !pctrie_is_empty(&object->un_pager.swp.swp_blks));
237 }
238 
239 static inline void
240 unlock_map(struct faultstate *fs)
241 {
242 
243 	if (fs->lookup_still_valid) {
244 		vm_map_lookup_done(fs->map, fs->entry);
245 		fs->lookup_still_valid = false;
246 	}
247 }
248 
249 static void
250 unlock_vp(struct faultstate *fs)
251 {
252 
253 	if (fs->vp != NULL) {
254 		vput(fs->vp);
255 		fs->vp = NULL;
256 	}
257 }
258 
259 static void
260 fault_deallocate(struct faultstate *fs)
261 {
262 
263 	fault_page_release(&fs->m_cow);
264 	fault_page_release(&fs->m);
265 	vm_object_pip_wakeup(fs->object);
266 	if (fs->object != fs->first_object) {
267 		VM_OBJECT_WLOCK(fs->first_object);
268 		fault_page_free(&fs->first_m);
269 		VM_OBJECT_WUNLOCK(fs->first_object);
270 		vm_object_pip_wakeup(fs->first_object);
271 	}
272 	vm_object_deallocate(fs->first_object);
273 	unlock_map(fs);
274 	unlock_vp(fs);
275 }
276 
277 static void
278 unlock_and_deallocate(struct faultstate *fs)
279 {
280 
281 	VM_OBJECT_WUNLOCK(fs->object);
282 	fault_deallocate(fs);
283 }
284 
285 static void
286 vm_fault_dirty(struct faultstate *fs, vm_page_t m)
287 {
288 	bool need_dirty;
289 
290 	if (((fs->prot & VM_PROT_WRITE) == 0 &&
291 	    (fs->fault_flags & VM_FAULT_DIRTY) == 0) ||
292 	    (m->oflags & VPO_UNMANAGED) != 0)
293 		return;
294 
295 	VM_PAGE_OBJECT_BUSY_ASSERT(m);
296 
297 	need_dirty = ((fs->fault_type & VM_PROT_WRITE) != 0 &&
298 	    (fs->fault_flags & VM_FAULT_WIRE) == 0) ||
299 	    (fs->fault_flags & VM_FAULT_DIRTY) != 0;
300 
301 	vm_object_set_writeable_dirty(m->object);
302 
303 	/*
304 	 * If the fault is a write, we know that this page is being
305 	 * written NOW so dirty it explicitly to save on
306 	 * pmap_is_modified() calls later.
307 	 *
308 	 * Also, since the page is now dirty, we can possibly tell
309 	 * the pager to release any swap backing the page.
310 	 */
311 	if (need_dirty && vm_page_set_dirty(m) == 0) {
312 		/*
313 		 * If this is a NOSYNC mmap we do not want to set PGA_NOSYNC
314 		 * if the page is already dirty to prevent data written with
315 		 * the expectation of being synced from not being synced.
316 		 * Likewise if this entry does not request NOSYNC then make
317 		 * sure the page isn't marked NOSYNC.  Applications sharing
318 		 * data should use the same flags to avoid ping ponging.
319 		 */
320 		if ((fs->entry->eflags & MAP_ENTRY_NOSYNC) != 0)
321 			vm_page_aflag_set(m, PGA_NOSYNC);
322 		else
323 			vm_page_aflag_clear(m, PGA_NOSYNC);
324 	}
325 
326 }
327 
328 /*
329  * Unlocks fs.first_object and fs.map on success.
330  */
331 static enum fault_status
332 vm_fault_soft_fast(struct faultstate *fs)
333 {
334 	vm_page_t m, m_map;
335 #if VM_NRESERVLEVEL > 0
336 	vm_page_t m_super;
337 	int flags;
338 #endif
339 	int psind;
340 	vm_offset_t vaddr;
341 	enum fault_status res;
342 
343 	MPASS(fs->vp == NULL);
344 
345 	res = FAULT_SUCCESS;
346 	vaddr = fs->vaddr;
347 	vm_object_busy(fs->first_object);
348 	m = vm_page_lookup(fs->first_object, fs->first_pindex);
349 	/* A busy page can be mapped for read|execute access. */
350 	if (m == NULL || ((fs->prot & VM_PROT_WRITE) != 0 &&
351 	    vm_page_busied(m)) || !vm_page_all_valid(m)) {
352 		res = FAULT_FAILURE;
353 		goto out;
354 	}
355 	m_map = m;
356 	psind = 0;
357 #if VM_NRESERVLEVEL > 0
358 	if ((m->flags & PG_FICTITIOUS) == 0 &&
359 	    (m_super = vm_reserv_to_superpage(m)) != NULL &&
360 	    rounddown2(vaddr, pagesizes[m_super->psind]) >= fs->entry->start &&
361 	    roundup2(vaddr + 1, pagesizes[m_super->psind]) <= fs->entry->end &&
362 	    (vaddr & (pagesizes[m_super->psind] - 1)) == (VM_PAGE_TO_PHYS(m) &
363 	    (pagesizes[m_super->psind] - 1)) && !fs->wired &&
364 	    pmap_ps_enabled(fs->map->pmap)) {
365 		flags = PS_ALL_VALID;
366 		if ((fs->prot & VM_PROT_WRITE) != 0) {
367 			/*
368 			 * Create a superpage mapping allowing write access
369 			 * only if none of the constituent pages are busy and
370 			 * all of them are already dirty (except possibly for
371 			 * the page that was faulted on).
372 			 */
373 			flags |= PS_NONE_BUSY;
374 			if ((fs->first_object->flags & OBJ_UNMANAGED) == 0)
375 				flags |= PS_ALL_DIRTY;
376 		}
377 		if (vm_page_ps_test(m_super, flags, m)) {
378 			m_map = m_super;
379 			psind = m_super->psind;
380 			vaddr = rounddown2(vaddr, pagesizes[psind]);
381 			/* Preset the modified bit for dirty superpages. */
382 			if ((flags & PS_ALL_DIRTY) != 0)
383 				fs->fault_type |= VM_PROT_WRITE;
384 		}
385 	}
386 #endif
387 	if (pmap_enter(fs->map->pmap, vaddr, m_map, fs->prot, fs->fault_type |
388 	    PMAP_ENTER_NOSLEEP | (fs->wired ? PMAP_ENTER_WIRED : 0), psind) !=
389 	    KERN_SUCCESS) {
390 		res = FAULT_FAILURE;
391 		goto out;
392 	}
393 	if (fs->m_hold != NULL) {
394 		(*fs->m_hold) = m;
395 		vm_page_wire(m);
396 	}
397 	if (psind == 0 && !fs->wired)
398 		vm_fault_prefault(fs, vaddr, PFBAK, PFFOR, true);
399 	VM_OBJECT_RUNLOCK(fs->first_object);
400 	vm_fault_dirty(fs, m);
401 	vm_map_lookup_done(fs->map, fs->entry);
402 	curthread->td_ru.ru_minflt++;
403 
404 out:
405 	vm_object_unbusy(fs->first_object);
406 	return (res);
407 }
408 
409 static void
410 vm_fault_restore_map_lock(struct faultstate *fs)
411 {
412 
413 	VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
414 	MPASS(blockcount_read(&fs->first_object->paging_in_progress) > 0);
415 
416 	if (!vm_map_trylock_read(fs->map)) {
417 		VM_OBJECT_WUNLOCK(fs->first_object);
418 		vm_map_lock_read(fs->map);
419 		VM_OBJECT_WLOCK(fs->first_object);
420 	}
421 	fs->lookup_still_valid = true;
422 }
423 
424 static void
425 vm_fault_populate_check_page(vm_page_t m)
426 {
427 
428 	/*
429 	 * Check each page to ensure that the pager is obeying the
430 	 * interface: the page must be installed in the object, fully
431 	 * valid, and exclusively busied.
432 	 */
433 	MPASS(m != NULL);
434 	MPASS(vm_page_all_valid(m));
435 	MPASS(vm_page_xbusied(m));
436 }
437 
438 static void
439 vm_fault_populate_cleanup(vm_object_t object, vm_pindex_t first,
440     vm_pindex_t last)
441 {
442 	vm_page_t m;
443 	vm_pindex_t pidx;
444 
445 	VM_OBJECT_ASSERT_WLOCKED(object);
446 	MPASS(first <= last);
447 	for (pidx = first, m = vm_page_lookup(object, pidx);
448 	    pidx <= last; pidx++, m = vm_page_next(m)) {
449 		vm_fault_populate_check_page(m);
450 		vm_page_deactivate(m);
451 		vm_page_xunbusy(m);
452 	}
453 }
454 
455 static enum fault_status
456 vm_fault_populate(struct faultstate *fs)
457 {
458 	vm_offset_t vaddr;
459 	vm_page_t m;
460 	vm_pindex_t map_first, map_last, pager_first, pager_last, pidx;
461 	int bdry_idx, i, npages, psind, rv;
462 	enum fault_status res;
463 
464 	MPASS(fs->object == fs->first_object);
465 	VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
466 	MPASS(blockcount_read(&fs->first_object->paging_in_progress) > 0);
467 	MPASS(fs->first_object->backing_object == NULL);
468 	MPASS(fs->lookup_still_valid);
469 
470 	pager_first = OFF_TO_IDX(fs->entry->offset);
471 	pager_last = pager_first + atop(fs->entry->end - fs->entry->start) - 1;
472 	unlock_map(fs);
473 	unlock_vp(fs);
474 
475 	res = FAULT_SUCCESS;
476 
477 	/*
478 	 * Call the pager (driver) populate() method.
479 	 *
480 	 * There is no guarantee that the method will be called again
481 	 * if the current fault is for read, and a future fault is
482 	 * for write.  Report the entry's maximum allowed protection
483 	 * to the driver.
484 	 */
485 	rv = vm_pager_populate(fs->first_object, fs->first_pindex,
486 	    fs->fault_type, fs->entry->max_protection, &pager_first,
487 	    &pager_last);
488 
489 	VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
490 	if (rv == VM_PAGER_BAD) {
491 		/*
492 		 * VM_PAGER_BAD is the backdoor for a pager to request
493 		 * normal fault handling.
494 		 */
495 		vm_fault_restore_map_lock(fs);
496 		if (fs->map->timestamp != fs->map_generation)
497 			return (FAULT_RESTART);
498 		return (FAULT_CONTINUE);
499 	}
500 	if (rv != VM_PAGER_OK)
501 		return (FAULT_FAILURE); /* AKA SIGSEGV */
502 
503 	/* Ensure that the driver is obeying the interface. */
504 	MPASS(pager_first <= pager_last);
505 	MPASS(fs->first_pindex <= pager_last);
506 	MPASS(fs->first_pindex >= pager_first);
507 	MPASS(pager_last < fs->first_object->size);
508 
509 	vm_fault_restore_map_lock(fs);
510 	bdry_idx = (fs->entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) >>
511 	    MAP_ENTRY_SPLIT_BOUNDARY_SHIFT;
512 	if (fs->map->timestamp != fs->map_generation) {
513 		if (bdry_idx == 0) {
514 			vm_fault_populate_cleanup(fs->first_object, pager_first,
515 			    pager_last);
516 		} else {
517 			m = vm_page_lookup(fs->first_object, pager_first);
518 			if (m != fs->m)
519 				vm_page_xunbusy(m);
520 		}
521 		return (FAULT_RESTART);
522 	}
523 
524 	/*
525 	 * The map is unchanged after our last unlock.  Process the fault.
526 	 *
527 	 * First, the special case of largepage mappings, where
528 	 * populate only busies the first page in superpage run.
529 	 */
530 	if (bdry_idx != 0) {
531 		KASSERT(PMAP_HAS_LARGEPAGES,
532 		    ("missing pmap support for large pages"));
533 		m = vm_page_lookup(fs->first_object, pager_first);
534 		vm_fault_populate_check_page(m);
535 		VM_OBJECT_WUNLOCK(fs->first_object);
536 		vaddr = fs->entry->start + IDX_TO_OFF(pager_first) -
537 		    fs->entry->offset;
538 		/* assert alignment for entry */
539 		KASSERT((vaddr & (pagesizes[bdry_idx] - 1)) == 0,
540     ("unaligned superpage start %#jx pager_first %#jx offset %#jx vaddr %#jx",
541 		    (uintmax_t)fs->entry->start, (uintmax_t)pager_first,
542 		    (uintmax_t)fs->entry->offset, (uintmax_t)vaddr));
543 		KASSERT((VM_PAGE_TO_PHYS(m) & (pagesizes[bdry_idx] - 1)) == 0,
544 		    ("unaligned superpage m %p %#jx", m,
545 		    (uintmax_t)VM_PAGE_TO_PHYS(m)));
546 		rv = pmap_enter(fs->map->pmap, vaddr, m, fs->prot,
547 		    fs->fault_type | (fs->wired ? PMAP_ENTER_WIRED : 0) |
548 		    PMAP_ENTER_LARGEPAGE, bdry_idx);
549 		VM_OBJECT_WLOCK(fs->first_object);
550 		vm_page_xunbusy(m);
551 		if (rv != KERN_SUCCESS) {
552 			res = FAULT_FAILURE;
553 			goto out;
554 		}
555 		if ((fs->fault_flags & VM_FAULT_WIRE) != 0) {
556 			for (i = 0; i < atop(pagesizes[bdry_idx]); i++)
557 				vm_page_wire(m + i);
558 		}
559 		if (fs->m_hold != NULL) {
560 			*fs->m_hold = m + (fs->first_pindex - pager_first);
561 			vm_page_wire(*fs->m_hold);
562 		}
563 		goto out;
564 	}
565 
566 	/*
567 	 * The range [pager_first, pager_last] that is given to the
568 	 * pager is only a hint.  The pager may populate any range
569 	 * within the object that includes the requested page index.
570 	 * In case the pager expanded the range, clip it to fit into
571 	 * the map entry.
572 	 */
573 	map_first = OFF_TO_IDX(fs->entry->offset);
574 	if (map_first > pager_first) {
575 		vm_fault_populate_cleanup(fs->first_object, pager_first,
576 		    map_first - 1);
577 		pager_first = map_first;
578 	}
579 	map_last = map_first + atop(fs->entry->end - fs->entry->start) - 1;
580 	if (map_last < pager_last) {
581 		vm_fault_populate_cleanup(fs->first_object, map_last + 1,
582 		    pager_last);
583 		pager_last = map_last;
584 	}
585 	for (pidx = pager_first, m = vm_page_lookup(fs->first_object, pidx);
586 	    pidx <= pager_last;
587 	    pidx += npages, m = vm_page_next(&m[npages - 1])) {
588 		vaddr = fs->entry->start + IDX_TO_OFF(pidx) - fs->entry->offset;
589 
590 		psind = m->psind;
591 		if (psind > 0 && ((vaddr & (pagesizes[psind] - 1)) != 0 ||
592 		    pidx + OFF_TO_IDX(pagesizes[psind]) - 1 > pager_last ||
593 		    !pmap_ps_enabled(fs->map->pmap) || fs->wired))
594 			psind = 0;
595 
596 		npages = atop(pagesizes[psind]);
597 		for (i = 0; i < npages; i++) {
598 			vm_fault_populate_check_page(&m[i]);
599 			vm_fault_dirty(fs, &m[i]);
600 		}
601 		VM_OBJECT_WUNLOCK(fs->first_object);
602 		rv = pmap_enter(fs->map->pmap, vaddr, m, fs->prot, fs->fault_type |
603 		    (fs->wired ? PMAP_ENTER_WIRED : 0), psind);
604 
605 		/*
606 		 * pmap_enter() may fail for a superpage mapping if additional
607 		 * protection policies prevent the full mapping.
608 		 * For example, this will happen on amd64 if the entire
609 		 * address range does not share the same userspace protection
610 		 * key.  Revert to single-page mappings if this happens.
611 		 */
612 		MPASS(rv == KERN_SUCCESS ||
613 		    (psind > 0 && rv == KERN_PROTECTION_FAILURE));
614 		if (__predict_false(psind > 0 &&
615 		    rv == KERN_PROTECTION_FAILURE)) {
616 			MPASS(!fs->wired);
617 			for (i = 0; i < npages; i++) {
618 				rv = pmap_enter(fs->map->pmap, vaddr + ptoa(i),
619 				    &m[i], fs->prot, fs->fault_type, 0);
620 				MPASS(rv == KERN_SUCCESS);
621 			}
622 		}
623 
624 		VM_OBJECT_WLOCK(fs->first_object);
625 		for (i = 0; i < npages; i++) {
626 			if ((fs->fault_flags & VM_FAULT_WIRE) != 0 &&
627 			    m[i].pindex == fs->first_pindex)
628 				vm_page_wire(&m[i]);
629 			else
630 				vm_page_activate(&m[i]);
631 			if (fs->m_hold != NULL &&
632 			    m[i].pindex == fs->first_pindex) {
633 				(*fs->m_hold) = &m[i];
634 				vm_page_wire(&m[i]);
635 			}
636 			vm_page_xunbusy(&m[i]);
637 		}
638 	}
639 out:
640 	curthread->td_ru.ru_majflt++;
641 	return (res);
642 }
643 
644 static int prot_fault_translation;
645 SYSCTL_INT(_machdep, OID_AUTO, prot_fault_translation, CTLFLAG_RWTUN,
646     &prot_fault_translation, 0,
647     "Control signal to deliver on protection fault");
648 
649 /* compat definition to keep common code for signal translation */
650 #define	UCODE_PAGEFLT	12
651 #ifdef T_PAGEFLT
652 _Static_assert(UCODE_PAGEFLT == T_PAGEFLT, "T_PAGEFLT");
653 #endif
654 
655 /*
656  *	vm_fault_trap:
657  *
658  *	Handle a page fault occurring at the given address,
659  *	requiring the given permissions, in the map specified.
660  *	If successful, the page is inserted into the
661  *	associated physical map.
662  *
663  *	NOTE: the given address should be truncated to the
664  *	proper page address.
665  *
666  *	KERN_SUCCESS is returned if the page fault is handled; otherwise,
667  *	a standard error specifying why the fault is fatal is returned.
668  *
669  *	The map in question must be referenced, and remains so.
670  *	Caller may hold no locks.
671  */
672 int
673 vm_fault_trap(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
674     int fault_flags, int *signo, int *ucode)
675 {
676 	int result;
677 
678 	MPASS(signo == NULL || ucode != NULL);
679 #ifdef KTRACE
680 	if (map != kernel_map && KTRPOINT(curthread, KTR_FAULT))
681 		ktrfault(vaddr, fault_type);
682 #endif
683 	result = vm_fault(map, trunc_page(vaddr), fault_type, fault_flags,
684 	    NULL);
685 	KASSERT(result == KERN_SUCCESS || result == KERN_FAILURE ||
686 	    result == KERN_INVALID_ADDRESS ||
687 	    result == KERN_RESOURCE_SHORTAGE ||
688 	    result == KERN_PROTECTION_FAILURE ||
689 	    result == KERN_OUT_OF_BOUNDS,
690 	    ("Unexpected Mach error %d from vm_fault()", result));
691 #ifdef KTRACE
692 	if (map != kernel_map && KTRPOINT(curthread, KTR_FAULTEND))
693 		ktrfaultend(result);
694 #endif
695 	if (result != KERN_SUCCESS && signo != NULL) {
696 		switch (result) {
697 		case KERN_FAILURE:
698 		case KERN_INVALID_ADDRESS:
699 			*signo = SIGSEGV;
700 			*ucode = SEGV_MAPERR;
701 			break;
702 		case KERN_RESOURCE_SHORTAGE:
703 			*signo = SIGBUS;
704 			*ucode = BUS_OOMERR;
705 			break;
706 		case KERN_OUT_OF_BOUNDS:
707 			*signo = SIGBUS;
708 			*ucode = BUS_OBJERR;
709 			break;
710 		case KERN_PROTECTION_FAILURE:
711 			if (prot_fault_translation == 0) {
712 				/*
713 				 * Autodetect.  This check also covers
714 				 * the images without the ABI-tag ELF
715 				 * note.
716 				 */
717 				if (SV_CURPROC_ABI() == SV_ABI_FREEBSD &&
718 				    curproc->p_osrel >= P_OSREL_SIGSEGV) {
719 					*signo = SIGSEGV;
720 					*ucode = SEGV_ACCERR;
721 				} else {
722 					*signo = SIGBUS;
723 					*ucode = UCODE_PAGEFLT;
724 				}
725 			} else if (prot_fault_translation == 1) {
726 				/* Always compat mode. */
727 				*signo = SIGBUS;
728 				*ucode = UCODE_PAGEFLT;
729 			} else {
730 				/* Always SIGSEGV mode. */
731 				*signo = SIGSEGV;
732 				*ucode = SEGV_ACCERR;
733 			}
734 			break;
735 		default:
736 			KASSERT(0, ("Unexpected Mach error %d from vm_fault()",
737 			    result));
738 			break;
739 		}
740 	}
741 	return (result);
742 }
743 
744 static enum fault_status
745 vm_fault_lock_vnode(struct faultstate *fs, bool objlocked)
746 {
747 	struct vnode *vp;
748 	int error, locked;
749 
750 	if (fs->object->type != OBJT_VNODE)
751 		return (FAULT_CONTINUE);
752 	vp = fs->object->handle;
753 	if (vp == fs->vp) {
754 		ASSERT_VOP_LOCKED(vp, "saved vnode is not locked");
755 		return (FAULT_CONTINUE);
756 	}
757 
758 	/*
759 	 * Perform an unlock in case the desired vnode changed while
760 	 * the map was unlocked during a retry.
761 	 */
762 	unlock_vp(fs);
763 
764 	locked = VOP_ISLOCKED(vp);
765 	if (locked != LK_EXCLUSIVE)
766 		locked = LK_SHARED;
767 
768 	/*
769 	 * We must not sleep acquiring the vnode lock while we have
770 	 * the page exclusive busied or the object's
771 	 * paging-in-progress count incremented.  Otherwise, we could
772 	 * deadlock.
773 	 */
774 	error = vget(vp, locked | LK_CANRECURSE | LK_NOWAIT);
775 	if (error == 0) {
776 		fs->vp = vp;
777 		return (FAULT_CONTINUE);
778 	}
779 
780 	vhold(vp);
781 	if (objlocked)
782 		unlock_and_deallocate(fs);
783 	else
784 		fault_deallocate(fs);
785 	error = vget(vp, locked | LK_RETRY | LK_CANRECURSE);
786 	vdrop(vp);
787 	fs->vp = vp;
788 	KASSERT(error == 0, ("vm_fault: vget failed %d", error));
789 	return (FAULT_RESTART);
790 }
791 
792 /*
793  * Calculate the desired readahead.  Handle drop-behind.
794  *
795  * Returns the number of readahead blocks to pass to the pager.
796  */
797 static int
798 vm_fault_readahead(struct faultstate *fs)
799 {
800 	int era, nera;
801 	u_char behavior;
802 
803 	KASSERT(fs->lookup_still_valid, ("map unlocked"));
804 	era = fs->entry->read_ahead;
805 	behavior = vm_map_entry_behavior(fs->entry);
806 	if (behavior == MAP_ENTRY_BEHAV_RANDOM) {
807 		nera = 0;
808 	} else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) {
809 		nera = VM_FAULT_READ_AHEAD_MAX;
810 		if (fs->vaddr == fs->entry->next_read)
811 			vm_fault_dontneed(fs, fs->vaddr, nera);
812 	} else if (fs->vaddr == fs->entry->next_read) {
813 		/*
814 		 * This is a sequential fault.  Arithmetically
815 		 * increase the requested number of pages in
816 		 * the read-ahead window.  The requested
817 		 * number of pages is "# of sequential faults
818 		 * x (read ahead min + 1) + read ahead min"
819 		 */
820 		nera = VM_FAULT_READ_AHEAD_MIN;
821 		if (era > 0) {
822 			nera += era + 1;
823 			if (nera > VM_FAULT_READ_AHEAD_MAX)
824 				nera = VM_FAULT_READ_AHEAD_MAX;
825 		}
826 		if (era == VM_FAULT_READ_AHEAD_MAX)
827 			vm_fault_dontneed(fs, fs->vaddr, nera);
828 	} else {
829 		/*
830 		 * This is a non-sequential fault.
831 		 */
832 		nera = 0;
833 	}
834 	if (era != nera) {
835 		/*
836 		 * A read lock on the map suffices to update
837 		 * the read ahead count safely.
838 		 */
839 		fs->entry->read_ahead = nera;
840 	}
841 
842 	return (nera);
843 }
844 
845 static int
846 vm_fault_lookup(struct faultstate *fs)
847 {
848 	int result;
849 
850 	KASSERT(!fs->lookup_still_valid,
851 	   ("vm_fault_lookup: Map already locked."));
852 	result = vm_map_lookup(&fs->map, fs->vaddr, fs->fault_type |
853 	    VM_PROT_FAULT_LOOKUP, &fs->entry, &fs->first_object,
854 	    &fs->first_pindex, &fs->prot, &fs->wired);
855 	if (result != KERN_SUCCESS) {
856 		unlock_vp(fs);
857 		return (result);
858 	}
859 
860 	fs->map_generation = fs->map->timestamp;
861 
862 	if (fs->entry->eflags & MAP_ENTRY_NOFAULT) {
863 		panic("%s: fault on nofault entry, addr: %#lx",
864 		    __func__, (u_long)fs->vaddr);
865 	}
866 
867 	if (fs->entry->eflags & MAP_ENTRY_IN_TRANSITION &&
868 	    fs->entry->wiring_thread != curthread) {
869 		vm_map_unlock_read(fs->map);
870 		vm_map_lock(fs->map);
871 		if (vm_map_lookup_entry(fs->map, fs->vaddr, &fs->entry) &&
872 		    (fs->entry->eflags & MAP_ENTRY_IN_TRANSITION)) {
873 			unlock_vp(fs);
874 			fs->entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
875 			vm_map_unlock_and_wait(fs->map, 0);
876 		} else
877 			vm_map_unlock(fs->map);
878 		return (KERN_RESOURCE_SHORTAGE);
879 	}
880 
881 	MPASS((fs->entry->eflags & MAP_ENTRY_GUARD) == 0);
882 
883 	if (fs->wired)
884 		fs->fault_type = fs->prot | (fs->fault_type & VM_PROT_COPY);
885 	else
886 		KASSERT((fs->fault_flags & VM_FAULT_WIRE) == 0,
887 		    ("!fs->wired && VM_FAULT_WIRE"));
888 	fs->lookup_still_valid = true;
889 
890 	return (KERN_SUCCESS);
891 }
892 
893 static int
894 vm_fault_relookup(struct faultstate *fs)
895 {
896 	vm_object_t retry_object;
897 	vm_pindex_t retry_pindex;
898 	vm_prot_t retry_prot;
899 	int result;
900 
901 	if (!vm_map_trylock_read(fs->map))
902 		return (KERN_RESTART);
903 
904 	fs->lookup_still_valid = true;
905 	if (fs->map->timestamp == fs->map_generation)
906 		return (KERN_SUCCESS);
907 
908 	result = vm_map_lookup_locked(&fs->map, fs->vaddr, fs->fault_type,
909 	    &fs->entry, &retry_object, &retry_pindex, &retry_prot,
910 	    &fs->wired);
911 	if (result != KERN_SUCCESS) {
912 		/*
913 		 * If retry of map lookup would have blocked then
914 		 * retry fault from start.
915 		 */
916 		if (result == KERN_FAILURE)
917 			return (KERN_RESTART);
918 		return (result);
919 	}
920 	if (retry_object != fs->first_object ||
921 	    retry_pindex != fs->first_pindex)
922 		return (KERN_RESTART);
923 
924 	/*
925 	 * Check whether the protection has changed or the object has
926 	 * been copied while we left the map unlocked. Changing from
927 	 * read to write permission is OK - we leave the page
928 	 * write-protected, and catch the write fault. Changing from
929 	 * write to read permission means that we can't mark the page
930 	 * write-enabled after all.
931 	 */
932 	fs->prot &= retry_prot;
933 	fs->fault_type &= retry_prot;
934 	if (fs->prot == 0)
935 		return (KERN_RESTART);
936 
937 	/* Reassert because wired may have changed. */
938 	KASSERT(fs->wired || (fs->fault_flags & VM_FAULT_WIRE) == 0,
939 	    ("!wired && VM_FAULT_WIRE"));
940 
941 	return (KERN_SUCCESS);
942 }
943 
944 static void
945 vm_fault_cow(struct faultstate *fs)
946 {
947 	bool is_first_object_locked;
948 
949 	KASSERT(fs->object != fs->first_object,
950 	    ("source and target COW objects are identical"));
951 
952 	/*
953 	 * This allows pages to be virtually copied from a backing_object
954 	 * into the first_object, where the backing object has no other
955 	 * refs to it, and cannot gain any more refs.  Instead of a bcopy,
956 	 * we just move the page from the backing object to the first
957 	 * object.  Note that we must mark the page dirty in the first
958 	 * object so that it will go out to swap when needed.
959 	 */
960 	is_first_object_locked = false;
961 	if (
962 	    /*
963 	     * Only one shadow object and no other refs.
964 	     */
965 	    fs->object->shadow_count == 1 && fs->object->ref_count == 1 &&
966 	    /*
967 	     * No other ways to look the object up
968 	     */
969 	    fs->object->handle == NULL && (fs->object->flags & OBJ_ANON) != 0 &&
970 	    /*
971 	     * We don't chase down the shadow chain and we can acquire locks.
972 	     */
973 	    (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs->first_object)) &&
974 	    fs->object == fs->first_object->backing_object &&
975 	    VM_OBJECT_TRYWLOCK(fs->object)) {
976 		/*
977 		 * Remove but keep xbusy for replace.  fs->m is moved into
978 		 * fs->first_object and left busy while fs->first_m is
979 		 * conditionally freed.
980 		 */
981 		vm_page_remove_xbusy(fs->m);
982 		vm_page_replace(fs->m, fs->first_object, fs->first_pindex,
983 		    fs->first_m);
984 		vm_page_dirty(fs->m);
985 #if VM_NRESERVLEVEL > 0
986 		/*
987 		 * Rename the reservation.
988 		 */
989 		vm_reserv_rename(fs->m, fs->first_object, fs->object,
990 		    OFF_TO_IDX(fs->first_object->backing_object_offset));
991 #endif
992 		VM_OBJECT_WUNLOCK(fs->object);
993 		VM_OBJECT_WUNLOCK(fs->first_object);
994 		fs->first_m = fs->m;
995 		fs->m = NULL;
996 		VM_CNT_INC(v_cow_optim);
997 	} else {
998 		if (is_first_object_locked)
999 			VM_OBJECT_WUNLOCK(fs->first_object);
1000 		/*
1001 		 * Oh, well, lets copy it.
1002 		 */
1003 		pmap_copy_page(fs->m, fs->first_m);
1004 		vm_page_valid(fs->first_m);
1005 		if (fs->wired && (fs->fault_flags & VM_FAULT_WIRE) == 0) {
1006 			vm_page_wire(fs->first_m);
1007 			vm_page_unwire(fs->m, PQ_INACTIVE);
1008 		}
1009 		/*
1010 		 * Save the cow page to be released after
1011 		 * pmap_enter is complete.
1012 		 */
1013 		fs->m_cow = fs->m;
1014 		fs->m = NULL;
1015 
1016 		/*
1017 		 * Typically, the shadow object is either private to this
1018 		 * address space (OBJ_ONEMAPPING) or its pages are read only.
1019 		 * In the highly unusual case where the pages of a shadow object
1020 		 * are read/write shared between this and other address spaces,
1021 		 * we need to ensure that any pmap-level mappings to the
1022 		 * original, copy-on-write page from the backing object are
1023 		 * removed from those other address spaces.
1024 		 *
1025 		 * The flag check is racy, but this is tolerable: if
1026 		 * OBJ_ONEMAPPING is cleared after the check, the busy state
1027 		 * ensures that new mappings of m_cow can't be created.
1028 		 * pmap_enter() will replace an existing mapping in the current
1029 		 * address space.  If OBJ_ONEMAPPING is set after the check,
1030 		 * removing mappings will at worse trigger some unnecessary page
1031 		 * faults.
1032 		 */
1033 		vm_page_assert_xbusied(fs->m_cow);
1034 		if ((fs->first_object->flags & OBJ_ONEMAPPING) == 0)
1035 			pmap_remove_all(fs->m_cow);
1036 	}
1037 
1038 	vm_object_pip_wakeup(fs->object);
1039 
1040 	/*
1041 	 * Only use the new page below...
1042 	 */
1043 	fs->object = fs->first_object;
1044 	fs->pindex = fs->first_pindex;
1045 	fs->m = fs->first_m;
1046 	VM_CNT_INC(v_cow_faults);
1047 	curthread->td_cow++;
1048 }
1049 
1050 static bool
1051 vm_fault_next(struct faultstate *fs)
1052 {
1053 	vm_object_t next_object;
1054 
1055 	/*
1056 	 * The requested page does not exist at this object/
1057 	 * offset.  Remove the invalid page from the object,
1058 	 * waking up anyone waiting for it, and continue on to
1059 	 * the next object.  However, if this is the top-level
1060 	 * object, we must leave the busy page in place to
1061 	 * prevent another process from rushing past us, and
1062 	 * inserting the page in that object at the same time
1063 	 * that we are.
1064 	 */
1065 	if (fs->object == fs->first_object) {
1066 		fs->first_m = fs->m;
1067 		fs->m = NULL;
1068 	} else
1069 		fault_page_free(&fs->m);
1070 
1071 	/*
1072 	 * Move on to the next object.  Lock the next object before
1073 	 * unlocking the current one.
1074 	 */
1075 	VM_OBJECT_ASSERT_WLOCKED(fs->object);
1076 	next_object = fs->object->backing_object;
1077 	if (next_object == NULL)
1078 		return (false);
1079 	MPASS(fs->first_m != NULL);
1080 	KASSERT(fs->object != next_object, ("object loop %p", next_object));
1081 	VM_OBJECT_WLOCK(next_object);
1082 	vm_object_pip_add(next_object, 1);
1083 	if (fs->object != fs->first_object)
1084 		vm_object_pip_wakeup(fs->object);
1085 	fs->pindex += OFF_TO_IDX(fs->object->backing_object_offset);
1086 	VM_OBJECT_WUNLOCK(fs->object);
1087 	fs->object = next_object;
1088 
1089 	return (true);
1090 }
1091 
1092 static void
1093 vm_fault_zerofill(struct faultstate *fs)
1094 {
1095 
1096 	/*
1097 	 * If there's no object left, fill the page in the top
1098 	 * object with zeros.
1099 	 */
1100 	if (fs->object != fs->first_object) {
1101 		vm_object_pip_wakeup(fs->object);
1102 		fs->object = fs->first_object;
1103 		fs->pindex = fs->first_pindex;
1104 	}
1105 	MPASS(fs->first_m != NULL);
1106 	MPASS(fs->m == NULL);
1107 	fs->m = fs->first_m;
1108 	fs->first_m = NULL;
1109 
1110 	/*
1111 	 * Zero the page if necessary and mark it valid.
1112 	 */
1113 	if ((fs->m->flags & PG_ZERO) == 0) {
1114 		pmap_zero_page(fs->m);
1115 	} else {
1116 		VM_CNT_INC(v_ozfod);
1117 	}
1118 	VM_CNT_INC(v_zfod);
1119 	vm_page_valid(fs->m);
1120 }
1121 
1122 /*
1123  * Initiate page fault after timeout.  Returns true if caller should
1124  * do vm_waitpfault() after the call.
1125  */
1126 static bool
1127 vm_fault_allocate_oom(struct faultstate *fs)
1128 {
1129 	struct timeval now;
1130 
1131 	unlock_and_deallocate(fs);
1132 	if (vm_pfault_oom_attempts < 0)
1133 		return (true);
1134 	if (!fs->oom_started) {
1135 		fs->oom_started = true;
1136 		getmicrotime(&fs->oom_start_time);
1137 		return (true);
1138 	}
1139 
1140 	getmicrotime(&now);
1141 	timevalsub(&now, &fs->oom_start_time);
1142 	if (now.tv_sec < vm_pfault_oom_attempts * vm_pfault_oom_wait)
1143 		return (true);
1144 
1145 	if (bootverbose)
1146 		printf(
1147 	    "proc %d (%s) failed to alloc page on fault, starting OOM\n",
1148 		    curproc->p_pid, curproc->p_comm);
1149 	vm_pageout_oom(VM_OOM_MEM_PF);
1150 	fs->oom_started = false;
1151 	return (false);
1152 }
1153 
1154 /*
1155  * Allocate a page directly or via the object populate method.
1156  */
1157 static enum fault_status
1158 vm_fault_allocate(struct faultstate *fs)
1159 {
1160 	struct domainset *dset;
1161 	enum fault_status res;
1162 
1163 	if ((fs->object->flags & OBJ_SIZEVNLOCK) != 0) {
1164 		res = vm_fault_lock_vnode(fs, true);
1165 		MPASS(res == FAULT_CONTINUE || res == FAULT_RESTART);
1166 		if (res == FAULT_RESTART)
1167 			return (res);
1168 	}
1169 
1170 	if (fs->pindex >= fs->object->size) {
1171 		unlock_and_deallocate(fs);
1172 		return (FAULT_OUT_OF_BOUNDS);
1173 	}
1174 
1175 	if (fs->object == fs->first_object &&
1176 	    (fs->first_object->flags & OBJ_POPULATE) != 0 &&
1177 	    fs->first_object->shadow_count == 0) {
1178 		res = vm_fault_populate(fs);
1179 		switch (res) {
1180 		case FAULT_SUCCESS:
1181 		case FAULT_FAILURE:
1182 		case FAULT_RESTART:
1183 			unlock_and_deallocate(fs);
1184 			return (res);
1185 		case FAULT_CONTINUE:
1186 			/*
1187 			 * Pager's populate() method
1188 			 * returned VM_PAGER_BAD.
1189 			 */
1190 			break;
1191 		default:
1192 			panic("inconsistent return codes");
1193 		}
1194 	}
1195 
1196 	/*
1197 	 * Allocate a new page for this object/offset pair.
1198 	 *
1199 	 * If the process has a fatal signal pending, prioritize the allocation
1200 	 * with the expectation that the process will exit shortly and free some
1201 	 * pages.  In particular, the signal may have been posted by the page
1202 	 * daemon in an attempt to resolve an out-of-memory condition.
1203 	 *
1204 	 * The unlocked read of the p_flag is harmless.  At worst, the P_KILLED
1205 	 * might be not observed here, and allocation fails, causing a restart
1206 	 * and new reading of the p_flag.
1207 	 */
1208 	dset = fs->object->domain.dr_policy;
1209 	if (dset == NULL)
1210 		dset = curthread->td_domain.dr_policy;
1211 	if (!vm_page_count_severe_set(&dset->ds_mask) || P_KILLED(curproc)) {
1212 #if VM_NRESERVLEVEL > 0
1213 		vm_object_color(fs->object, atop(fs->vaddr) - fs->pindex);
1214 #endif
1215 		fs->m = vm_page_alloc(fs->object, fs->pindex,
1216 		    P_KILLED(curproc) ? VM_ALLOC_SYSTEM : 0);
1217 	}
1218 	if (fs->m == NULL) {
1219 		if (vm_fault_allocate_oom(fs))
1220 			vm_waitpfault(dset, vm_pfault_oom_wait * hz);
1221 		return (FAULT_RESTART);
1222 	}
1223 	fs->oom_started = false;
1224 
1225 	return (FAULT_CONTINUE);
1226 }
1227 
1228 /*
1229  * Call the pager to retrieve the page if there is a chance
1230  * that the pager has it, and potentially retrieve additional
1231  * pages at the same time.
1232  */
1233 static enum fault_status
1234 vm_fault_getpages(struct faultstate *fs, int *behindp, int *aheadp)
1235 {
1236 	vm_offset_t e_end, e_start;
1237 	int ahead, behind, cluster_offset, rv;
1238 	enum fault_status status;
1239 	u_char behavior;
1240 
1241 	/*
1242 	 * Prepare for unlocking the map.  Save the map
1243 	 * entry's start and end addresses, which are used to
1244 	 * optimize the size of the pager operation below.
1245 	 * Even if the map entry's addresses change after
1246 	 * unlocking the map, using the saved addresses is
1247 	 * safe.
1248 	 */
1249 	e_start = fs->entry->start;
1250 	e_end = fs->entry->end;
1251 	behavior = vm_map_entry_behavior(fs->entry);
1252 
1253 	/*
1254 	 * If the pager for the current object might have
1255 	 * the page, then determine the number of additional
1256 	 * pages to read and potentially reprioritize
1257 	 * previously read pages for earlier reclamation.
1258 	 * These operations should only be performed once per
1259 	 * page fault.  Even if the current pager doesn't
1260 	 * have the page, the number of additional pages to
1261 	 * read will apply to subsequent objects in the
1262 	 * shadow chain.
1263 	 */
1264 	if (fs->nera == -1 && !P_KILLED(curproc))
1265 		fs->nera = vm_fault_readahead(fs);
1266 
1267 	/*
1268 	 * Release the map lock before locking the vnode or
1269 	 * sleeping in the pager.  (If the current object has
1270 	 * a shadow, then an earlier iteration of this loop
1271 	 * may have already unlocked the map.)
1272 	 */
1273 	unlock_map(fs);
1274 
1275 	status = vm_fault_lock_vnode(fs, false);
1276 	MPASS(status == FAULT_CONTINUE || status == FAULT_RESTART);
1277 	if (status == FAULT_RESTART)
1278 		return (status);
1279 	KASSERT(fs->vp == NULL || !fs->map->system_map,
1280 	    ("vm_fault: vnode-backed object mapped by system map"));
1281 
1282 	/*
1283 	 * Page in the requested page and hint the pager,
1284 	 * that it may bring up surrounding pages.
1285 	 */
1286 	if (fs->nera == -1 || behavior == MAP_ENTRY_BEHAV_RANDOM ||
1287 	    P_KILLED(curproc)) {
1288 		behind = 0;
1289 		ahead = 0;
1290 	} else {
1291 		/* Is this a sequential fault? */
1292 		if (fs->nera > 0) {
1293 			behind = 0;
1294 			ahead = fs->nera;
1295 		} else {
1296 			/*
1297 			 * Request a cluster of pages that is
1298 			 * aligned to a VM_FAULT_READ_DEFAULT
1299 			 * page offset boundary within the
1300 			 * object.  Alignment to a page offset
1301 			 * boundary is more likely to coincide
1302 			 * with the underlying file system
1303 			 * block than alignment to a virtual
1304 			 * address boundary.
1305 			 */
1306 			cluster_offset = fs->pindex % VM_FAULT_READ_DEFAULT;
1307 			behind = ulmin(cluster_offset,
1308 			    atop(fs->vaddr - e_start));
1309 			ahead = VM_FAULT_READ_DEFAULT - 1 - cluster_offset;
1310 		}
1311 		ahead = ulmin(ahead, atop(e_end - fs->vaddr) - 1);
1312 	}
1313 	*behindp = behind;
1314 	*aheadp = ahead;
1315 	rv = vm_pager_get_pages(fs->object, &fs->m, 1, behindp, aheadp);
1316 	if (rv == VM_PAGER_OK)
1317 		return (FAULT_HARD);
1318 	if (rv == VM_PAGER_ERROR)
1319 		printf("vm_fault: pager read error, pid %d (%s)\n",
1320 		    curproc->p_pid, curproc->p_comm);
1321 	/*
1322 	 * If an I/O error occurred or the requested page was
1323 	 * outside the range of the pager, clean up and return
1324 	 * an error.
1325 	 */
1326 	if (rv == VM_PAGER_ERROR || rv == VM_PAGER_BAD) {
1327 		VM_OBJECT_WLOCK(fs->object);
1328 		fault_page_free(&fs->m);
1329 		unlock_and_deallocate(fs);
1330 		return (FAULT_OUT_OF_BOUNDS);
1331 	}
1332 	KASSERT(rv == VM_PAGER_FAIL,
1333 	    ("%s: unexpected pager error %d", __func__, rv));
1334 	return (FAULT_CONTINUE);
1335 }
1336 
1337 /*
1338  * Wait/Retry if the page is busy.  We have to do this if the page is
1339  * either exclusive or shared busy because the vm_pager may be using
1340  * read busy for pageouts (and even pageins if it is the vnode pager),
1341  * and we could end up trying to pagein and pageout the same page
1342  * simultaneously.
1343  *
1344  * We can theoretically allow the busy case on a read fault if the page
1345  * is marked valid, but since such pages are typically already pmap'd,
1346  * putting that special case in might be more effort then it is worth.
1347  * We cannot under any circumstances mess around with a shared busied
1348  * page except, perhaps, to pmap it.
1349  */
1350 static void
1351 vm_fault_busy_sleep(struct faultstate *fs)
1352 {
1353 	/*
1354 	 * Reference the page before unlocking and
1355 	 * sleeping so that the page daemon is less
1356 	 * likely to reclaim it.
1357 	 */
1358 	vm_page_aflag_set(fs->m, PGA_REFERENCED);
1359 	if (fs->object != fs->first_object) {
1360 		fault_page_release(&fs->first_m);
1361 		vm_object_pip_wakeup(fs->first_object);
1362 	}
1363 	vm_object_pip_wakeup(fs->object);
1364 	unlock_map(fs);
1365 	if (fs->m != vm_page_lookup(fs->object, fs->pindex) ||
1366 	    !vm_page_busy_sleep(fs->m, "vmpfw", 0))
1367 		VM_OBJECT_WUNLOCK(fs->object);
1368 	VM_CNT_INC(v_intrans);
1369 	vm_object_deallocate(fs->first_object);
1370 }
1371 
1372 /*
1373  * Handle page lookup, populate, allocate, page-in for the current
1374  * object.
1375  *
1376  * The object is locked on entry and will remain locked with a return
1377  * code of FAULT_CONTINUE so that fault may follow the shadow chain.
1378  * Otherwise, the object will be unlocked upon return.
1379  */
1380 static enum fault_status
1381 vm_fault_object(struct faultstate *fs, int *behindp, int *aheadp)
1382 {
1383 	enum fault_status res;
1384 	bool dead;
1385 
1386 	/*
1387 	 * If the object is marked for imminent termination, we retry
1388 	 * here, since the collapse pass has raced with us.  Otherwise,
1389 	 * if we see terminally dead object, return fail.
1390 	 */
1391 	if ((fs->object->flags & OBJ_DEAD) != 0) {
1392 		dead = fs->object->type == OBJT_DEAD;
1393 		unlock_and_deallocate(fs);
1394 		if (dead)
1395 			return (FAULT_PROTECTION_FAILURE);
1396 		pause("vmf_de", 1);
1397 		return (FAULT_RESTART);
1398 	}
1399 
1400 	/*
1401 	 * See if the page is resident.
1402 	 */
1403 	fs->m = vm_page_lookup(fs->object, fs->pindex);
1404 	if (fs->m != NULL) {
1405 		if (!vm_page_tryxbusy(fs->m)) {
1406 			vm_fault_busy_sleep(fs);
1407 			return (FAULT_RESTART);
1408 		}
1409 
1410 		/*
1411 		 * The page is marked busy for other processes and the
1412 		 * pagedaemon.  If it is still completely valid we are
1413 		 * done.
1414 		 */
1415 		if (vm_page_all_valid(fs->m)) {
1416 			VM_OBJECT_WUNLOCK(fs->object);
1417 			return (FAULT_SOFT);
1418 		}
1419 	}
1420 	VM_OBJECT_ASSERT_WLOCKED(fs->object);
1421 
1422 	/*
1423 	 * Page is not resident.  If the pager might contain the page
1424 	 * or this is the beginning of the search, allocate a new
1425 	 * page.
1426 	 */
1427 	if (fs->m == NULL && (fault_object_needs_getpages(fs->object) ||
1428 	    fs->object == fs->first_object)) {
1429 		res = vm_fault_allocate(fs);
1430 		if (res != FAULT_CONTINUE)
1431 			return (res);
1432 	}
1433 
1434 	/*
1435 	 * Default objects have no pager so no exclusive busy exists
1436 	 * to protect this page in the chain.  Skip to the next
1437 	 * object without dropping the lock to preserve atomicity of
1438 	 * shadow faults.
1439 	 */
1440 	if (fault_object_needs_getpages(fs->object)) {
1441 		/*
1442 		 * At this point, we have either allocated a new page
1443 		 * or found an existing page that is only partially
1444 		 * valid.
1445 		 *
1446 		 * We hold a reference on the current object and the
1447 		 * page is exclusive busied.  The exclusive busy
1448 		 * prevents simultaneous faults and collapses while
1449 		 * the object lock is dropped.
1450 		 */
1451 		VM_OBJECT_WUNLOCK(fs->object);
1452 		res = vm_fault_getpages(fs, behindp, aheadp);
1453 		if (res == FAULT_CONTINUE)
1454 			VM_OBJECT_WLOCK(fs->object);
1455 	} else {
1456 		res = FAULT_CONTINUE;
1457 	}
1458 	return (res);
1459 }
1460 
1461 int
1462 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
1463     int fault_flags, vm_page_t *m_hold)
1464 {
1465 	struct faultstate fs;
1466 	int ahead, behind, faultcount, rv;
1467 	enum fault_status res;
1468 	bool hardfault;
1469 
1470 	VM_CNT_INC(v_vm_faults);
1471 
1472 	if ((curthread->td_pflags & TDP_NOFAULTING) != 0)
1473 		return (KERN_PROTECTION_FAILURE);
1474 
1475 	fs.vp = NULL;
1476 	fs.vaddr = vaddr;
1477 	fs.m_hold = m_hold;
1478 	fs.fault_flags = fault_flags;
1479 	fs.map = map;
1480 	fs.lookup_still_valid = false;
1481 	fs.oom_started = false;
1482 	fs.nera = -1;
1483 	faultcount = 0;
1484 	hardfault = false;
1485 
1486 RetryFault:
1487 	fs.fault_type = fault_type;
1488 
1489 	/*
1490 	 * Find the backing store object and offset into it to begin the
1491 	 * search.
1492 	 */
1493 	rv = vm_fault_lookup(&fs);
1494 	if (rv != KERN_SUCCESS) {
1495 		if (rv == KERN_RESOURCE_SHORTAGE)
1496 			goto RetryFault;
1497 		return (rv);
1498 	}
1499 
1500 	/*
1501 	 * Try to avoid lock contention on the top-level object through
1502 	 * special-case handling of some types of page faults, specifically,
1503 	 * those that are mapping an existing page from the top-level object.
1504 	 * Under this condition, a read lock on the object suffices, allowing
1505 	 * multiple page faults of a similar type to run in parallel.
1506 	 */
1507 	if (fs.vp == NULL /* avoid locked vnode leak */ &&
1508 	    (fs.entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) == 0 &&
1509 	    (fs.fault_flags & (VM_FAULT_WIRE | VM_FAULT_DIRTY)) == 0) {
1510 		VM_OBJECT_RLOCK(fs.first_object);
1511 		res = vm_fault_soft_fast(&fs);
1512 		if (res == FAULT_SUCCESS)
1513 			return (KERN_SUCCESS);
1514 		if (!VM_OBJECT_TRYUPGRADE(fs.first_object)) {
1515 			VM_OBJECT_RUNLOCK(fs.first_object);
1516 			VM_OBJECT_WLOCK(fs.first_object);
1517 		}
1518 	} else {
1519 		VM_OBJECT_WLOCK(fs.first_object);
1520 	}
1521 
1522 	/*
1523 	 * Make a reference to this object to prevent its disposal while we
1524 	 * are messing with it.  Once we have the reference, the map is free
1525 	 * to be diddled.  Since objects reference their shadows (and copies),
1526 	 * they will stay around as well.
1527 	 *
1528 	 * Bump the paging-in-progress count to prevent size changes (e.g.
1529 	 * truncation operations) during I/O.
1530 	 */
1531 	vm_object_reference_locked(fs.first_object);
1532 	vm_object_pip_add(fs.first_object, 1);
1533 
1534 	fs.m_cow = fs.m = fs.first_m = NULL;
1535 
1536 	/*
1537 	 * Search for the page at object/offset.
1538 	 */
1539 	fs.object = fs.first_object;
1540 	fs.pindex = fs.first_pindex;
1541 
1542 	if ((fs.entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) != 0) {
1543 		res = vm_fault_allocate(&fs);
1544 		switch (res) {
1545 		case FAULT_RESTART:
1546 			goto RetryFault;
1547 		case FAULT_SUCCESS:
1548 			return (KERN_SUCCESS);
1549 		case FAULT_FAILURE:
1550 			return (KERN_FAILURE);
1551 		case FAULT_OUT_OF_BOUNDS:
1552 			return (KERN_OUT_OF_BOUNDS);
1553 		case FAULT_CONTINUE:
1554 			break;
1555 		default:
1556 			panic("vm_fault: Unhandled status %d", res);
1557 		}
1558 	}
1559 
1560 	while (TRUE) {
1561 		KASSERT(fs.m == NULL,
1562 		    ("page still set %p at loop start", fs.m));
1563 
1564 		res = vm_fault_object(&fs, &behind, &ahead);
1565 		switch (res) {
1566 		case FAULT_SOFT:
1567 			goto found;
1568 		case FAULT_HARD:
1569 			faultcount = behind + 1 + ahead;
1570 			hardfault = true;
1571 			goto found;
1572 		case FAULT_RESTART:
1573 			goto RetryFault;
1574 		case FAULT_SUCCESS:
1575 			return (KERN_SUCCESS);
1576 		case FAULT_FAILURE:
1577 			return (KERN_FAILURE);
1578 		case FAULT_OUT_OF_BOUNDS:
1579 			return (KERN_OUT_OF_BOUNDS);
1580 		case FAULT_PROTECTION_FAILURE:
1581 			return (KERN_PROTECTION_FAILURE);
1582 		case FAULT_CONTINUE:
1583 			break;
1584 		default:
1585 			panic("vm_fault: Unhandled status %d", res);
1586 		}
1587 
1588 		/*
1589 		 * The page was not found in the current object.  Try to
1590 		 * traverse into a backing object or zero fill if none is
1591 		 * found.
1592 		 */
1593 		if (vm_fault_next(&fs))
1594 			continue;
1595 		if ((fs.fault_flags & VM_FAULT_NOFILL) != 0) {
1596 			if (fs.first_object == fs.object)
1597 				fault_page_free(&fs.first_m);
1598 			unlock_and_deallocate(&fs);
1599 			return (KERN_OUT_OF_BOUNDS);
1600 		}
1601 		VM_OBJECT_WUNLOCK(fs.object);
1602 		vm_fault_zerofill(&fs);
1603 		/* Don't try to prefault neighboring pages. */
1604 		faultcount = 1;
1605 		break;
1606 	}
1607 
1608 found:
1609 	/*
1610 	 * A valid page has been found and exclusively busied.  The
1611 	 * object lock must no longer be held.
1612 	 */
1613 	vm_page_assert_xbusied(fs.m);
1614 	VM_OBJECT_ASSERT_UNLOCKED(fs.object);
1615 
1616 	/*
1617 	 * If the page is being written, but isn't already owned by the
1618 	 * top-level object, we have to copy it into a new page owned by the
1619 	 * top-level object.
1620 	 */
1621 	if (fs.object != fs.first_object) {
1622 		/*
1623 		 * We only really need to copy if we want to write it.
1624 		 */
1625 		if ((fs.fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1626 			vm_fault_cow(&fs);
1627 			/*
1628 			 * We only try to prefault read-only mappings to the
1629 			 * neighboring pages when this copy-on-write fault is
1630 			 * a hard fault.  In other cases, trying to prefault
1631 			 * is typically wasted effort.
1632 			 */
1633 			if (faultcount == 0)
1634 				faultcount = 1;
1635 
1636 		} else {
1637 			fs.prot &= ~VM_PROT_WRITE;
1638 		}
1639 	}
1640 
1641 	/*
1642 	 * We must verify that the maps have not changed since our last
1643 	 * lookup.
1644 	 */
1645 	if (!fs.lookup_still_valid) {
1646 		rv = vm_fault_relookup(&fs);
1647 		if (rv != KERN_SUCCESS) {
1648 			fault_deallocate(&fs);
1649 			if (rv == KERN_RESTART)
1650 				goto RetryFault;
1651 			return (rv);
1652 		}
1653 	}
1654 	VM_OBJECT_ASSERT_UNLOCKED(fs.object);
1655 
1656 	/*
1657 	 * If the page was filled by a pager, save the virtual address that
1658 	 * should be faulted on next under a sequential access pattern to the
1659 	 * map entry.  A read lock on the map suffices to update this address
1660 	 * safely.
1661 	 */
1662 	if (hardfault)
1663 		fs.entry->next_read = vaddr + ptoa(ahead) + PAGE_SIZE;
1664 
1665 	/*
1666 	 * Page must be completely valid or it is not fit to
1667 	 * map into user space.  vm_pager_get_pages() ensures this.
1668 	 */
1669 	vm_page_assert_xbusied(fs.m);
1670 	KASSERT(vm_page_all_valid(fs.m),
1671 	    ("vm_fault: page %p partially invalid", fs.m));
1672 
1673 	vm_fault_dirty(&fs, fs.m);
1674 
1675 	/*
1676 	 * Put this page into the physical map.  We had to do the unlock above
1677 	 * because pmap_enter() may sleep.  We don't put the page
1678 	 * back on the active queue until later so that the pageout daemon
1679 	 * won't find it (yet).
1680 	 */
1681 	pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot,
1682 	    fs.fault_type | (fs.wired ? PMAP_ENTER_WIRED : 0), 0);
1683 	if (faultcount != 1 && (fs.fault_flags & VM_FAULT_WIRE) == 0 &&
1684 	    fs.wired == 0)
1685 		vm_fault_prefault(&fs, vaddr,
1686 		    faultcount > 0 ? behind : PFBAK,
1687 		    faultcount > 0 ? ahead : PFFOR, false);
1688 
1689 	/*
1690 	 * If the page is not wired down, then put it where the pageout daemon
1691 	 * can find it.
1692 	 */
1693 	if ((fs.fault_flags & VM_FAULT_WIRE) != 0)
1694 		vm_page_wire(fs.m);
1695 	else
1696 		vm_page_activate(fs.m);
1697 	if (fs.m_hold != NULL) {
1698 		(*fs.m_hold) = fs.m;
1699 		vm_page_wire(fs.m);
1700 	}
1701 	vm_page_xunbusy(fs.m);
1702 	fs.m = NULL;
1703 
1704 	/*
1705 	 * Unlock everything, and return
1706 	 */
1707 	fault_deallocate(&fs);
1708 	if (hardfault) {
1709 		VM_CNT_INC(v_io_faults);
1710 		curthread->td_ru.ru_majflt++;
1711 #ifdef RACCT
1712 		if (racct_enable && fs.object->type == OBJT_VNODE) {
1713 			PROC_LOCK(curproc);
1714 			if ((fs.fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1715 				racct_add_force(curproc, RACCT_WRITEBPS,
1716 				    PAGE_SIZE + behind * PAGE_SIZE);
1717 				racct_add_force(curproc, RACCT_WRITEIOPS, 1);
1718 			} else {
1719 				racct_add_force(curproc, RACCT_READBPS,
1720 				    PAGE_SIZE + ahead * PAGE_SIZE);
1721 				racct_add_force(curproc, RACCT_READIOPS, 1);
1722 			}
1723 			PROC_UNLOCK(curproc);
1724 		}
1725 #endif
1726 	} else
1727 		curthread->td_ru.ru_minflt++;
1728 
1729 	return (KERN_SUCCESS);
1730 }
1731 
1732 /*
1733  * Speed up the reclamation of pages that precede the faulting pindex within
1734  * the first object of the shadow chain.  Essentially, perform the equivalent
1735  * to madvise(..., MADV_DONTNEED) on a large cluster of pages that precedes
1736  * the faulting pindex by the cluster size when the pages read by vm_fault()
1737  * cross a cluster-size boundary.  The cluster size is the greater of the
1738  * smallest superpage size and VM_FAULT_DONTNEED_MIN.
1739  *
1740  * When "fs->first_object" is a shadow object, the pages in the backing object
1741  * that precede the faulting pindex are deactivated by vm_fault().  So, this
1742  * function must only be concerned with pages in the first object.
1743  */
1744 static void
1745 vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr, int ahead)
1746 {
1747 	vm_map_entry_t entry;
1748 	vm_object_t first_object;
1749 	vm_offset_t end, start;
1750 	vm_page_t m, m_next;
1751 	vm_pindex_t pend, pstart;
1752 	vm_size_t size;
1753 
1754 	VM_OBJECT_ASSERT_UNLOCKED(fs->object);
1755 	first_object = fs->first_object;
1756 	/* Neither fictitious nor unmanaged pages can be reclaimed. */
1757 	if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) {
1758 		VM_OBJECT_RLOCK(first_object);
1759 		size = VM_FAULT_DONTNEED_MIN;
1760 		if (MAXPAGESIZES > 1 && size < pagesizes[1])
1761 			size = pagesizes[1];
1762 		end = rounddown2(vaddr, size);
1763 		if (vaddr - end >= size - PAGE_SIZE - ptoa(ahead) &&
1764 		    (entry = fs->entry)->start < end) {
1765 			if (end - entry->start < size)
1766 				start = entry->start;
1767 			else
1768 				start = end - size;
1769 			pmap_advise(fs->map->pmap, start, end, MADV_DONTNEED);
1770 			pstart = OFF_TO_IDX(entry->offset) + atop(start -
1771 			    entry->start);
1772 			m_next = vm_page_find_least(first_object, pstart);
1773 			pend = OFF_TO_IDX(entry->offset) + atop(end -
1774 			    entry->start);
1775 			while ((m = m_next) != NULL && m->pindex < pend) {
1776 				m_next = TAILQ_NEXT(m, listq);
1777 				if (!vm_page_all_valid(m) ||
1778 				    vm_page_busied(m))
1779 					continue;
1780 
1781 				/*
1782 				 * Don't clear PGA_REFERENCED, since it would
1783 				 * likely represent a reference by a different
1784 				 * process.
1785 				 *
1786 				 * Typically, at this point, prefetched pages
1787 				 * are still in the inactive queue.  Only
1788 				 * pages that triggered page faults are in the
1789 				 * active queue.  The test for whether the page
1790 				 * is in the inactive queue is racy; in the
1791 				 * worst case we will requeue the page
1792 				 * unnecessarily.
1793 				 */
1794 				if (!vm_page_inactive(m))
1795 					vm_page_deactivate(m);
1796 			}
1797 		}
1798 		VM_OBJECT_RUNLOCK(first_object);
1799 	}
1800 }
1801 
1802 /*
1803  * vm_fault_prefault provides a quick way of clustering
1804  * pagefaults into a processes address space.  It is a "cousin"
1805  * of vm_map_pmap_enter, except it runs at page fault time instead
1806  * of mmap time.
1807  */
1808 static void
1809 vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
1810     int backward, int forward, bool obj_locked)
1811 {
1812 	pmap_t pmap;
1813 	vm_map_entry_t entry;
1814 	vm_object_t backing_object, lobject;
1815 	vm_offset_t addr, starta;
1816 	vm_pindex_t pindex;
1817 	vm_page_t m;
1818 	int i;
1819 
1820 	pmap = fs->map->pmap;
1821 	if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
1822 		return;
1823 
1824 	entry = fs->entry;
1825 
1826 	if (addra < backward * PAGE_SIZE) {
1827 		starta = entry->start;
1828 	} else {
1829 		starta = addra - backward * PAGE_SIZE;
1830 		if (starta < entry->start)
1831 			starta = entry->start;
1832 	}
1833 
1834 	/*
1835 	 * Generate the sequence of virtual addresses that are candidates for
1836 	 * prefaulting in an outward spiral from the faulting virtual address,
1837 	 * "addra".  Specifically, the sequence is "addra - PAGE_SIZE", "addra
1838 	 * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ...
1839 	 * If the candidate address doesn't have a backing physical page, then
1840 	 * the loop immediately terminates.
1841 	 */
1842 	for (i = 0; i < 2 * imax(backward, forward); i++) {
1843 		addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE :
1844 		    PAGE_SIZE);
1845 		if (addr > addra + forward * PAGE_SIZE)
1846 			addr = 0;
1847 
1848 		if (addr < starta || addr >= entry->end)
1849 			continue;
1850 
1851 		if (!pmap_is_prefaultable(pmap, addr))
1852 			continue;
1853 
1854 		pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1855 		lobject = entry->object.vm_object;
1856 		if (!obj_locked)
1857 			VM_OBJECT_RLOCK(lobject);
1858 		while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1859 		    !fault_object_needs_getpages(lobject) &&
1860 		    (backing_object = lobject->backing_object) != NULL) {
1861 			KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1862 			    0, ("vm_fault_prefault: unaligned object offset"));
1863 			pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1864 			VM_OBJECT_RLOCK(backing_object);
1865 			if (!obj_locked || lobject != entry->object.vm_object)
1866 				VM_OBJECT_RUNLOCK(lobject);
1867 			lobject = backing_object;
1868 		}
1869 		if (m == NULL) {
1870 			if (!obj_locked || lobject != entry->object.vm_object)
1871 				VM_OBJECT_RUNLOCK(lobject);
1872 			break;
1873 		}
1874 		if (vm_page_all_valid(m) &&
1875 		    (m->flags & PG_FICTITIOUS) == 0)
1876 			pmap_enter_quick(pmap, addr, m, entry->protection);
1877 		if (!obj_locked || lobject != entry->object.vm_object)
1878 			VM_OBJECT_RUNLOCK(lobject);
1879 	}
1880 }
1881 
1882 /*
1883  * Hold each of the physical pages that are mapped by the specified range of
1884  * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
1885  * and allow the specified types of access, "prot".  If all of the implied
1886  * pages are successfully held, then the number of held pages is returned
1887  * together with pointers to those pages in the array "ma".  However, if any
1888  * of the pages cannot be held, -1 is returned.
1889  */
1890 int
1891 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
1892     vm_prot_t prot, vm_page_t *ma, int max_count)
1893 {
1894 	vm_offset_t end, va;
1895 	vm_page_t *mp;
1896 	int count;
1897 	boolean_t pmap_failed;
1898 
1899 	if (len == 0)
1900 		return (0);
1901 	end = round_page(addr + len);
1902 	addr = trunc_page(addr);
1903 
1904 	if (!vm_map_range_valid(map, addr, end))
1905 		return (-1);
1906 
1907 	if (atop(end - addr) > max_count)
1908 		panic("vm_fault_quick_hold_pages: count > max_count");
1909 	count = atop(end - addr);
1910 
1911 	/*
1912 	 * Most likely, the physical pages are resident in the pmap, so it is
1913 	 * faster to try pmap_extract_and_hold() first.
1914 	 */
1915 	pmap_failed = FALSE;
1916 	for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) {
1917 		*mp = pmap_extract_and_hold(map->pmap, va, prot);
1918 		if (*mp == NULL)
1919 			pmap_failed = TRUE;
1920 		else if ((prot & VM_PROT_WRITE) != 0 &&
1921 		    (*mp)->dirty != VM_PAGE_BITS_ALL) {
1922 			/*
1923 			 * Explicitly dirty the physical page.  Otherwise, the
1924 			 * caller's changes may go unnoticed because they are
1925 			 * performed through an unmanaged mapping or by a DMA
1926 			 * operation.
1927 			 *
1928 			 * The object lock is not held here.
1929 			 * See vm_page_clear_dirty_mask().
1930 			 */
1931 			vm_page_dirty(*mp);
1932 		}
1933 	}
1934 	if (pmap_failed) {
1935 		/*
1936 		 * One or more pages could not be held by the pmap.  Either no
1937 		 * page was mapped at the specified virtual address or that
1938 		 * mapping had insufficient permissions.  Attempt to fault in
1939 		 * and hold these pages.
1940 		 *
1941 		 * If vm_fault_disable_pagefaults() was called,
1942 		 * i.e., TDP_NOFAULTING is set, we must not sleep nor
1943 		 * acquire MD VM locks, which means we must not call
1944 		 * vm_fault().  Some (out of tree) callers mark
1945 		 * too wide a code area with vm_fault_disable_pagefaults()
1946 		 * already, use the VM_PROT_QUICK_NOFAULT flag to request
1947 		 * the proper behaviour explicitly.
1948 		 */
1949 		if ((prot & VM_PROT_QUICK_NOFAULT) != 0 &&
1950 		    (curthread->td_pflags & TDP_NOFAULTING) != 0)
1951 			goto error;
1952 		for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE)
1953 			if (*mp == NULL && vm_fault(map, va, prot,
1954 			    VM_FAULT_NORMAL, mp) != KERN_SUCCESS)
1955 				goto error;
1956 	}
1957 	return (count);
1958 error:
1959 	for (mp = ma; mp < ma + count; mp++)
1960 		if (*mp != NULL)
1961 			vm_page_unwire(*mp, PQ_INACTIVE);
1962 	return (-1);
1963 }
1964 
1965 /*
1966  *	Routine:
1967  *		vm_fault_copy_entry
1968  *	Function:
1969  *		Create new object backing dst_entry with private copy of all
1970  *		underlying pages. When src_entry is equal to dst_entry, function
1971  *		implements COW for wired-down map entry. Otherwise, it forks
1972  *		wired entry into dst_map.
1973  *
1974  *	In/out conditions:
1975  *		The source and destination maps must be locked for write.
1976  *		The source map entry must be wired down (or be a sharing map
1977  *		entry corresponding to a main map entry that is wired down).
1978  */
1979 void
1980 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map __unused,
1981     vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1982     vm_ooffset_t *fork_charge)
1983 {
1984 	vm_object_t backing_object, dst_object, object, src_object;
1985 	vm_pindex_t dst_pindex, pindex, src_pindex;
1986 	vm_prot_t access, prot;
1987 	vm_offset_t vaddr;
1988 	vm_page_t dst_m;
1989 	vm_page_t src_m;
1990 	bool upgrade;
1991 
1992 	upgrade = src_entry == dst_entry;
1993 	KASSERT(upgrade || dst_entry->object.vm_object == NULL,
1994 	    ("vm_fault_copy_entry: vm_object not NULL"));
1995 
1996 	/*
1997 	 * If not an upgrade, then enter the mappings in the pmap as
1998 	 * read and/or execute accesses.  Otherwise, enter them as
1999 	 * write accesses.
2000 	 *
2001 	 * A writeable large page mapping is only created if all of
2002 	 * the constituent small page mappings are modified. Marking
2003 	 * PTEs as modified on inception allows promotion to happen
2004 	 * without taking potentially large number of soft faults.
2005 	 */
2006 	access = prot = dst_entry->protection;
2007 	if (!upgrade)
2008 		access &= ~VM_PROT_WRITE;
2009 
2010 	src_object = src_entry->object.vm_object;
2011 	src_pindex = OFF_TO_IDX(src_entry->offset);
2012 
2013 	if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
2014 		dst_object = src_object;
2015 		vm_object_reference(dst_object);
2016 	} else {
2017 		/*
2018 		 * Create the top-level object for the destination entry.
2019 		 * Doesn't actually shadow anything - we copy the pages
2020 		 * directly.
2021 		 */
2022 		dst_object = vm_object_allocate_anon(atop(dst_entry->end -
2023 		    dst_entry->start), NULL, NULL, 0);
2024 #if VM_NRESERVLEVEL > 0
2025 		dst_object->flags |= OBJ_COLORED;
2026 		dst_object->pg_color = atop(dst_entry->start);
2027 #endif
2028 		dst_object->domain = src_object->domain;
2029 		dst_object->charge = dst_entry->end - dst_entry->start;
2030 
2031 		dst_entry->object.vm_object = dst_object;
2032 		dst_entry->offset = 0;
2033 		dst_entry->eflags &= ~MAP_ENTRY_VN_EXEC;
2034 	}
2035 
2036 	VM_OBJECT_WLOCK(dst_object);
2037 	if (fork_charge != NULL) {
2038 		KASSERT(dst_entry->cred == NULL,
2039 		    ("vm_fault_copy_entry: leaked swp charge"));
2040 		dst_object->cred = curthread->td_ucred;
2041 		crhold(dst_object->cred);
2042 		*fork_charge += dst_object->charge;
2043 	} else if ((dst_object->flags & OBJ_SWAP) != 0 &&
2044 	    dst_object->cred == NULL) {
2045 		KASSERT(dst_entry->cred != NULL, ("no cred for entry %p",
2046 		    dst_entry));
2047 		dst_object->cred = dst_entry->cred;
2048 		dst_entry->cred = NULL;
2049 	}
2050 
2051 	/*
2052 	 * Loop through all of the virtual pages within the entry's
2053 	 * range, copying each page from the source object to the
2054 	 * destination object.  Since the source is wired, those pages
2055 	 * must exist.  In contrast, the destination is pageable.
2056 	 * Since the destination object doesn't share any backing storage
2057 	 * with the source object, all of its pages must be dirtied,
2058 	 * regardless of whether they can be written.
2059 	 */
2060 	for (vaddr = dst_entry->start, dst_pindex = 0;
2061 	    vaddr < dst_entry->end;
2062 	    vaddr += PAGE_SIZE, dst_pindex++) {
2063 again:
2064 		/*
2065 		 * Find the page in the source object, and copy it in.
2066 		 * Because the source is wired down, the page will be
2067 		 * in memory.
2068 		 */
2069 		if (src_object != dst_object)
2070 			VM_OBJECT_RLOCK(src_object);
2071 		object = src_object;
2072 		pindex = src_pindex + dst_pindex;
2073 		while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
2074 		    (backing_object = object->backing_object) != NULL) {
2075 			/*
2076 			 * Unless the source mapping is read-only or
2077 			 * it is presently being upgraded from
2078 			 * read-only, the first object in the shadow
2079 			 * chain should provide all of the pages.  In
2080 			 * other words, this loop body should never be
2081 			 * executed when the source mapping is already
2082 			 * read/write.
2083 			 */
2084 			KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 ||
2085 			    upgrade,
2086 			    ("vm_fault_copy_entry: main object missing page"));
2087 
2088 			VM_OBJECT_RLOCK(backing_object);
2089 			pindex += OFF_TO_IDX(object->backing_object_offset);
2090 			if (object != dst_object)
2091 				VM_OBJECT_RUNLOCK(object);
2092 			object = backing_object;
2093 		}
2094 		KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing"));
2095 
2096 		if (object != dst_object) {
2097 			/*
2098 			 * Allocate a page in the destination object.
2099 			 */
2100 			dst_m = vm_page_alloc(dst_object, (src_object ==
2101 			    dst_object ? src_pindex : 0) + dst_pindex,
2102 			    VM_ALLOC_NORMAL);
2103 			if (dst_m == NULL) {
2104 				VM_OBJECT_WUNLOCK(dst_object);
2105 				VM_OBJECT_RUNLOCK(object);
2106 				vm_wait(dst_object);
2107 				VM_OBJECT_WLOCK(dst_object);
2108 				goto again;
2109 			}
2110 
2111 			/*
2112 			 * See the comment in vm_fault_cow().
2113 			 */
2114 			if (src_object == dst_object &&
2115 			    (object->flags & OBJ_ONEMAPPING) == 0)
2116 				pmap_remove_all(src_m);
2117 			pmap_copy_page(src_m, dst_m);
2118 
2119 			/*
2120 			 * The object lock does not guarantee that "src_m" will
2121 			 * transition from invalid to valid, but it does ensure
2122 			 * that "src_m" will not transition from valid to
2123 			 * invalid.
2124 			 */
2125 			dst_m->dirty = dst_m->valid = src_m->valid;
2126 			VM_OBJECT_RUNLOCK(object);
2127 		} else {
2128 			dst_m = src_m;
2129 			if (vm_page_busy_acquire(dst_m, VM_ALLOC_WAITFAIL) == 0)
2130 				goto again;
2131 			if (dst_m->pindex >= dst_object->size) {
2132 				/*
2133 				 * We are upgrading.  Index can occur
2134 				 * out of bounds if the object type is
2135 				 * vnode and the file was truncated.
2136 				 */
2137 				vm_page_xunbusy(dst_m);
2138 				break;
2139 			}
2140 		}
2141 
2142 		/*
2143 		 * Enter it in the pmap. If a wired, copy-on-write
2144 		 * mapping is being replaced by a write-enabled
2145 		 * mapping, then wire that new mapping.
2146 		 *
2147 		 * The page can be invalid if the user called
2148 		 * msync(MS_INVALIDATE) or truncated the backing vnode
2149 		 * or shared memory object.  In this case, do not
2150 		 * insert it into pmap, but still do the copy so that
2151 		 * all copies of the wired map entry have similar
2152 		 * backing pages.
2153 		 */
2154 		if (vm_page_all_valid(dst_m)) {
2155 			VM_OBJECT_WUNLOCK(dst_object);
2156 			pmap_enter(dst_map->pmap, vaddr, dst_m, prot,
2157 			    access | (upgrade ? PMAP_ENTER_WIRED : 0), 0);
2158 			VM_OBJECT_WLOCK(dst_object);
2159 		}
2160 
2161 		/*
2162 		 * Mark it no longer busy, and put it on the active list.
2163 		 */
2164 		if (upgrade) {
2165 			if (src_m != dst_m) {
2166 				vm_page_unwire(src_m, PQ_INACTIVE);
2167 				vm_page_wire(dst_m);
2168 			} else {
2169 				KASSERT(vm_page_wired(dst_m),
2170 				    ("dst_m %p is not wired", dst_m));
2171 			}
2172 		} else {
2173 			vm_page_activate(dst_m);
2174 		}
2175 		vm_page_xunbusy(dst_m);
2176 	}
2177 	VM_OBJECT_WUNLOCK(dst_object);
2178 	if (upgrade) {
2179 		dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
2180 		vm_object_deallocate(src_object);
2181 	}
2182 }
2183 
2184 /*
2185  * Block entry into the machine-independent layer's page fault handler by
2186  * the calling thread.  Subsequent calls to vm_fault() by that thread will
2187  * return KERN_PROTECTION_FAILURE.  Enable machine-dependent handling of
2188  * spurious page faults.
2189  */
2190 int
2191 vm_fault_disable_pagefaults(void)
2192 {
2193 
2194 	return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR));
2195 }
2196 
2197 void
2198 vm_fault_enable_pagefaults(int save)
2199 {
2200 
2201 	curthread_pflags_restore(save);
2202 }
2203