xref: /freebsd/sys/vm/vm_fault.c (revision a0409676120c1e558d0ade943019934e0f15118d)
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 	int		oom;
129 	boolean_t	wired;
130 
131 	/* Page reference for cow. */
132 	vm_page_t m_cow;
133 
134 	/* Current object. */
135 	vm_object_t	object;
136 	vm_pindex_t	pindex;
137 	vm_page_t	m;
138 
139 	/* Top-level map object. */
140 	vm_object_t	first_object;
141 	vm_pindex_t	first_pindex;
142 	vm_page_t	first_m;
143 
144 	/* Map state. */
145 	vm_map_t	map;
146 	vm_map_entry_t	entry;
147 	int		map_generation;
148 	bool		lookup_still_valid;
149 
150 	/* Vnode if locked. */
151 	struct vnode	*vp;
152 };
153 
154 static void vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr,
155 	    int ahead);
156 static void vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
157 	    int backward, int forward, bool obj_locked);
158 
159 static int vm_pfault_oom_attempts = 3;
160 SYSCTL_INT(_vm, OID_AUTO, pfault_oom_attempts, CTLFLAG_RWTUN,
161     &vm_pfault_oom_attempts, 0,
162     "Number of page allocation attempts in page fault handler before it "
163     "triggers OOM handling");
164 
165 static int vm_pfault_oom_wait = 10;
166 SYSCTL_INT(_vm, OID_AUTO, pfault_oom_wait, CTLFLAG_RWTUN,
167     &vm_pfault_oom_wait, 0,
168     "Number of seconds to wait for free pages before retrying "
169     "the page fault handler");
170 
171 static inline void
172 fault_page_release(vm_page_t *mp)
173 {
174 	vm_page_t m;
175 
176 	m = *mp;
177 	if (m != NULL) {
178 		/*
179 		 * We are likely to loop around again and attempt to busy
180 		 * this page.  Deactivating it leaves it available for
181 		 * pageout while optimizing fault restarts.
182 		 */
183 		vm_page_deactivate(m);
184 		vm_page_xunbusy(m);
185 		*mp = NULL;
186 	}
187 }
188 
189 static inline void
190 fault_page_free(vm_page_t *mp)
191 {
192 	vm_page_t m;
193 
194 	m = *mp;
195 	if (m != NULL) {
196 		VM_OBJECT_ASSERT_WLOCKED(m->object);
197 		if (!vm_page_wired(m))
198 			vm_page_free(m);
199 		else
200 			vm_page_xunbusy(m);
201 		*mp = NULL;
202 	}
203 }
204 
205 static inline void
206 unlock_map(struct faultstate *fs)
207 {
208 
209 	if (fs->lookup_still_valid) {
210 		vm_map_lookup_done(fs->map, fs->entry);
211 		fs->lookup_still_valid = false;
212 	}
213 }
214 
215 static void
216 unlock_vp(struct faultstate *fs)
217 {
218 
219 	if (fs->vp != NULL) {
220 		vput(fs->vp);
221 		fs->vp = NULL;
222 	}
223 }
224 
225 static void
226 fault_deallocate(struct faultstate *fs)
227 {
228 
229 	fault_page_release(&fs->m_cow);
230 	fault_page_release(&fs->m);
231 	vm_object_pip_wakeup(fs->object);
232 	if (fs->object != fs->first_object) {
233 		VM_OBJECT_WLOCK(fs->first_object);
234 		fault_page_free(&fs->first_m);
235 		VM_OBJECT_WUNLOCK(fs->first_object);
236 		vm_object_pip_wakeup(fs->first_object);
237 	}
238 	vm_object_deallocate(fs->first_object);
239 	unlock_map(fs);
240 	unlock_vp(fs);
241 }
242 
243 static void
244 unlock_and_deallocate(struct faultstate *fs)
245 {
246 
247 	VM_OBJECT_WUNLOCK(fs->object);
248 	fault_deallocate(fs);
249 }
250 
251 static void
252 vm_fault_dirty(struct faultstate *fs, vm_page_t m)
253 {
254 	bool need_dirty;
255 
256 	if (((fs->prot & VM_PROT_WRITE) == 0 &&
257 	    (fs->fault_flags & VM_FAULT_DIRTY) == 0) ||
258 	    (m->oflags & VPO_UNMANAGED) != 0)
259 		return;
260 
261 	VM_PAGE_OBJECT_BUSY_ASSERT(m);
262 
263 	need_dirty = ((fs->fault_type & VM_PROT_WRITE) != 0 &&
264 	    (fs->fault_flags & VM_FAULT_WIRE) == 0) ||
265 	    (fs->fault_flags & VM_FAULT_DIRTY) != 0;
266 
267 	vm_object_set_writeable_dirty(m->object);
268 
269 	/*
270 	 * If the fault is a write, we know that this page is being
271 	 * written NOW so dirty it explicitly to save on
272 	 * pmap_is_modified() calls later.
273 	 *
274 	 * Also, since the page is now dirty, we can possibly tell
275 	 * the pager to release any swap backing the page.
276 	 */
277 	if (need_dirty && vm_page_set_dirty(m) == 0) {
278 		/*
279 		 * If this is a NOSYNC mmap we do not want to set PGA_NOSYNC
280 		 * if the page is already dirty to prevent data written with
281 		 * the expectation of being synced from not being synced.
282 		 * Likewise if this entry does not request NOSYNC then make
283 		 * sure the page isn't marked NOSYNC.  Applications sharing
284 		 * data should use the same flags to avoid ping ponging.
285 		 */
286 		if ((fs->entry->eflags & MAP_ENTRY_NOSYNC) != 0)
287 			vm_page_aflag_set(m, PGA_NOSYNC);
288 		else
289 			vm_page_aflag_clear(m, PGA_NOSYNC);
290 	}
291 
292 }
293 
294 /*
295  * Unlocks fs.first_object and fs.map on success.
296  */
297 static int
298 vm_fault_soft_fast(struct faultstate *fs)
299 {
300 	vm_page_t m, m_map;
301 #if VM_NRESERVLEVEL > 0
302 	vm_page_t m_super;
303 	int flags;
304 #endif
305 	int psind, rv;
306 	vm_offset_t vaddr;
307 
308 	MPASS(fs->vp == NULL);
309 	vaddr = fs->vaddr;
310 	vm_object_busy(fs->first_object);
311 	m = vm_page_lookup(fs->first_object, fs->first_pindex);
312 	/* A busy page can be mapped for read|execute access. */
313 	if (m == NULL || ((fs->prot & VM_PROT_WRITE) != 0 &&
314 	    vm_page_busied(m)) || !vm_page_all_valid(m)) {
315 		rv = KERN_FAILURE;
316 		goto out;
317 	}
318 	m_map = m;
319 	psind = 0;
320 #if VM_NRESERVLEVEL > 0
321 	if ((m->flags & PG_FICTITIOUS) == 0 &&
322 	    (m_super = vm_reserv_to_superpage(m)) != NULL &&
323 	    rounddown2(vaddr, pagesizes[m_super->psind]) >= fs->entry->start &&
324 	    roundup2(vaddr + 1, pagesizes[m_super->psind]) <= fs->entry->end &&
325 	    (vaddr & (pagesizes[m_super->psind] - 1)) == (VM_PAGE_TO_PHYS(m) &
326 	    (pagesizes[m_super->psind] - 1)) && !fs->wired &&
327 	    pmap_ps_enabled(fs->map->pmap)) {
328 		flags = PS_ALL_VALID;
329 		if ((fs->prot & VM_PROT_WRITE) != 0) {
330 			/*
331 			 * Create a superpage mapping allowing write access
332 			 * only if none of the constituent pages are busy and
333 			 * all of them are already dirty (except possibly for
334 			 * the page that was faulted on).
335 			 */
336 			flags |= PS_NONE_BUSY;
337 			if ((fs->first_object->flags & OBJ_UNMANAGED) == 0)
338 				flags |= PS_ALL_DIRTY;
339 		}
340 		if (vm_page_ps_test(m_super, flags, m)) {
341 			m_map = m_super;
342 			psind = m_super->psind;
343 			vaddr = rounddown2(vaddr, pagesizes[psind]);
344 			/* Preset the modified bit for dirty superpages. */
345 			if ((flags & PS_ALL_DIRTY) != 0)
346 				fs->fault_type |= VM_PROT_WRITE;
347 		}
348 	}
349 #endif
350 	rv = pmap_enter(fs->map->pmap, vaddr, m_map, fs->prot, fs->fault_type |
351 	    PMAP_ENTER_NOSLEEP | (fs->wired ? PMAP_ENTER_WIRED : 0), psind);
352 	if (rv != KERN_SUCCESS)
353 		goto out;
354 	if (fs->m_hold != NULL) {
355 		(*fs->m_hold) = m;
356 		vm_page_wire(m);
357 	}
358 	if (psind == 0 && !fs->wired)
359 		vm_fault_prefault(fs, vaddr, PFBAK, PFFOR, true);
360 	VM_OBJECT_RUNLOCK(fs->first_object);
361 	vm_fault_dirty(fs, m);
362 	vm_map_lookup_done(fs->map, fs->entry);
363 	curthread->td_ru.ru_minflt++;
364 
365 out:
366 	vm_object_unbusy(fs->first_object);
367 	return (rv);
368 }
369 
370 static void
371 vm_fault_restore_map_lock(struct faultstate *fs)
372 {
373 
374 	VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
375 	MPASS(blockcount_read(&fs->first_object->paging_in_progress) > 0);
376 
377 	if (!vm_map_trylock_read(fs->map)) {
378 		VM_OBJECT_WUNLOCK(fs->first_object);
379 		vm_map_lock_read(fs->map);
380 		VM_OBJECT_WLOCK(fs->first_object);
381 	}
382 	fs->lookup_still_valid = true;
383 }
384 
385 static void
386 vm_fault_populate_check_page(vm_page_t m)
387 {
388 
389 	/*
390 	 * Check each page to ensure that the pager is obeying the
391 	 * interface: the page must be installed in the object, fully
392 	 * valid, and exclusively busied.
393 	 */
394 	MPASS(m != NULL);
395 	MPASS(vm_page_all_valid(m));
396 	MPASS(vm_page_xbusied(m));
397 }
398 
399 static void
400 vm_fault_populate_cleanup(vm_object_t object, vm_pindex_t first,
401     vm_pindex_t last)
402 {
403 	vm_page_t m;
404 	vm_pindex_t pidx;
405 
406 	VM_OBJECT_ASSERT_WLOCKED(object);
407 	MPASS(first <= last);
408 	for (pidx = first, m = vm_page_lookup(object, pidx);
409 	    pidx <= last; pidx++, m = vm_page_next(m)) {
410 		vm_fault_populate_check_page(m);
411 		vm_page_deactivate(m);
412 		vm_page_xunbusy(m);
413 	}
414 }
415 
416 static int
417 vm_fault_populate(struct faultstate *fs)
418 {
419 	vm_offset_t vaddr;
420 	vm_page_t m;
421 	vm_pindex_t map_first, map_last, pager_first, pager_last, pidx;
422 	int bdry_idx, i, npages, psind, rv;
423 
424 	MPASS(fs->object == fs->first_object);
425 	VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
426 	MPASS(blockcount_read(&fs->first_object->paging_in_progress) > 0);
427 	MPASS(fs->first_object->backing_object == NULL);
428 	MPASS(fs->lookup_still_valid);
429 
430 	pager_first = OFF_TO_IDX(fs->entry->offset);
431 	pager_last = pager_first + atop(fs->entry->end - fs->entry->start) - 1;
432 	unlock_map(fs);
433 	unlock_vp(fs);
434 
435 	/*
436 	 * Call the pager (driver) populate() method.
437 	 *
438 	 * There is no guarantee that the method will be called again
439 	 * if the current fault is for read, and a future fault is
440 	 * for write.  Report the entry's maximum allowed protection
441 	 * to the driver.
442 	 */
443 	rv = vm_pager_populate(fs->first_object, fs->first_pindex,
444 	    fs->fault_type, fs->entry->max_protection, &pager_first,
445 	    &pager_last);
446 
447 	VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
448 	if (rv == VM_PAGER_BAD) {
449 		/*
450 		 * VM_PAGER_BAD is the backdoor for a pager to request
451 		 * normal fault handling.
452 		 */
453 		vm_fault_restore_map_lock(fs);
454 		if (fs->map->timestamp != fs->map_generation)
455 			return (KERN_RESTART);
456 		return (KERN_NOT_RECEIVER);
457 	}
458 	if (rv != VM_PAGER_OK)
459 		return (KERN_FAILURE); /* AKA SIGSEGV */
460 
461 	/* Ensure that the driver is obeying the interface. */
462 	MPASS(pager_first <= pager_last);
463 	MPASS(fs->first_pindex <= pager_last);
464 	MPASS(fs->first_pindex >= pager_first);
465 	MPASS(pager_last < fs->first_object->size);
466 
467 	vm_fault_restore_map_lock(fs);
468 	bdry_idx = (fs->entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) >>
469 	    MAP_ENTRY_SPLIT_BOUNDARY_SHIFT;
470 	if (fs->map->timestamp != fs->map_generation) {
471 		if (bdry_idx == 0) {
472 			vm_fault_populate_cleanup(fs->first_object, pager_first,
473 			    pager_last);
474 		} else {
475 			m = vm_page_lookup(fs->first_object, pager_first);
476 			if (m != fs->m)
477 				vm_page_xunbusy(m);
478 		}
479 		return (KERN_RESTART);
480 	}
481 
482 	/*
483 	 * The map is unchanged after our last unlock.  Process the fault.
484 	 *
485 	 * First, the special case of largepage mappings, where
486 	 * populate only busies the first page in superpage run.
487 	 */
488 	if (bdry_idx != 0) {
489 		KASSERT(PMAP_HAS_LARGEPAGES,
490 		    ("missing pmap support for large pages"));
491 		m = vm_page_lookup(fs->first_object, pager_first);
492 		vm_fault_populate_check_page(m);
493 		VM_OBJECT_WUNLOCK(fs->first_object);
494 		vaddr = fs->entry->start + IDX_TO_OFF(pager_first) -
495 		    fs->entry->offset;
496 		/* assert alignment for entry */
497 		KASSERT((vaddr & (pagesizes[bdry_idx] - 1)) == 0,
498     ("unaligned superpage start %#jx pager_first %#jx offset %#jx vaddr %#jx",
499 		    (uintmax_t)fs->entry->start, (uintmax_t)pager_first,
500 		    (uintmax_t)fs->entry->offset, (uintmax_t)vaddr));
501 		KASSERT((VM_PAGE_TO_PHYS(m) & (pagesizes[bdry_idx] - 1)) == 0,
502 		    ("unaligned superpage m %p %#jx", m,
503 		    (uintmax_t)VM_PAGE_TO_PHYS(m)));
504 		rv = pmap_enter(fs->map->pmap, vaddr, m, fs->prot,
505 		    fs->fault_type | (fs->wired ? PMAP_ENTER_WIRED : 0) |
506 		    PMAP_ENTER_LARGEPAGE, bdry_idx);
507 		VM_OBJECT_WLOCK(fs->first_object);
508 		vm_page_xunbusy(m);
509 		if ((fs->fault_flags & VM_FAULT_WIRE) != 0) {
510 			for (i = 0; i < atop(pagesizes[bdry_idx]); i++)
511 				vm_page_wire(m + i);
512 		}
513 		if (fs->m_hold != NULL) {
514 			*fs->m_hold = m + (fs->first_pindex - pager_first);
515 			vm_page_wire(*fs->m_hold);
516 		}
517 		goto out;
518 	}
519 
520 	/*
521 	 * The range [pager_first, pager_last] that is given to the
522 	 * pager is only a hint.  The pager may populate any range
523 	 * within the object that includes the requested page index.
524 	 * In case the pager expanded the range, clip it to fit into
525 	 * the map entry.
526 	 */
527 	map_first = OFF_TO_IDX(fs->entry->offset);
528 	if (map_first > pager_first) {
529 		vm_fault_populate_cleanup(fs->first_object, pager_first,
530 		    map_first - 1);
531 		pager_first = map_first;
532 	}
533 	map_last = map_first + atop(fs->entry->end - fs->entry->start) - 1;
534 	if (map_last < pager_last) {
535 		vm_fault_populate_cleanup(fs->first_object, map_last + 1,
536 		    pager_last);
537 		pager_last = map_last;
538 	}
539 	for (pidx = pager_first, m = vm_page_lookup(fs->first_object, pidx);
540 	    pidx <= pager_last;
541 	    pidx += npages, m = vm_page_next(&m[npages - 1])) {
542 		vaddr = fs->entry->start + IDX_TO_OFF(pidx) - fs->entry->offset;
543 #if defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \
544     __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv) || \
545     defined(__powerpc64__)
546 		psind = m->psind;
547 		if (psind > 0 && ((vaddr & (pagesizes[psind] - 1)) != 0 ||
548 		    pidx + OFF_TO_IDX(pagesizes[psind]) - 1 > pager_last ||
549 		    !pmap_ps_enabled(fs->map->pmap) || fs->wired))
550 			psind = 0;
551 #else
552 		psind = 0;
553 #endif
554 		npages = atop(pagesizes[psind]);
555 		for (i = 0; i < npages; i++) {
556 			vm_fault_populate_check_page(&m[i]);
557 			vm_fault_dirty(fs, &m[i]);
558 		}
559 		VM_OBJECT_WUNLOCK(fs->first_object);
560 		rv = pmap_enter(fs->map->pmap, vaddr, m, fs->prot, fs->fault_type |
561 		    (fs->wired ? PMAP_ENTER_WIRED : 0), psind);
562 #if defined(__amd64__)
563 		if (psind > 0 && rv == KERN_FAILURE) {
564 			for (i = 0; i < npages; i++) {
565 				rv = pmap_enter(fs->map->pmap, vaddr + ptoa(i),
566 				    &m[i], fs->prot, fs->fault_type |
567 				    (fs->wired ? PMAP_ENTER_WIRED : 0), 0);
568 				MPASS(rv == KERN_SUCCESS);
569 			}
570 		}
571 #else
572 		MPASS(rv == KERN_SUCCESS);
573 #endif
574 		VM_OBJECT_WLOCK(fs->first_object);
575 		for (i = 0; i < npages; i++) {
576 			if ((fs->fault_flags & VM_FAULT_WIRE) != 0)
577 				vm_page_wire(&m[i]);
578 			else
579 				vm_page_activate(&m[i]);
580 			if (fs->m_hold != NULL && m[i].pindex == fs->first_pindex) {
581 				(*fs->m_hold) = &m[i];
582 				vm_page_wire(&m[i]);
583 			}
584 			vm_page_xunbusy(&m[i]);
585 		}
586 	}
587 out:
588 	curthread->td_ru.ru_majflt++;
589 	return (KERN_SUCCESS);
590 }
591 
592 static int prot_fault_translation;
593 SYSCTL_INT(_machdep, OID_AUTO, prot_fault_translation, CTLFLAG_RWTUN,
594     &prot_fault_translation, 0,
595     "Control signal to deliver on protection fault");
596 
597 /* compat definition to keep common code for signal translation */
598 #define	UCODE_PAGEFLT	12
599 #ifdef T_PAGEFLT
600 _Static_assert(UCODE_PAGEFLT == T_PAGEFLT, "T_PAGEFLT");
601 #endif
602 
603 /*
604  *	vm_fault_trap:
605  *
606  *	Handle a page fault occurring at the given address,
607  *	requiring the given permissions, in the map specified.
608  *	If successful, the page is inserted into the
609  *	associated physical map.
610  *
611  *	NOTE: the given address should be truncated to the
612  *	proper page address.
613  *
614  *	KERN_SUCCESS is returned if the page fault is handled; otherwise,
615  *	a standard error specifying why the fault is fatal is returned.
616  *
617  *	The map in question must be referenced, and remains so.
618  *	Caller may hold no locks.
619  */
620 int
621 vm_fault_trap(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
622     int fault_flags, int *signo, int *ucode)
623 {
624 	int result;
625 
626 	MPASS(signo == NULL || ucode != NULL);
627 #ifdef KTRACE
628 	if (map != kernel_map && KTRPOINT(curthread, KTR_FAULT))
629 		ktrfault(vaddr, fault_type);
630 #endif
631 	result = vm_fault(map, trunc_page(vaddr), fault_type, fault_flags,
632 	    NULL);
633 	KASSERT(result == KERN_SUCCESS || result == KERN_FAILURE ||
634 	    result == KERN_INVALID_ADDRESS ||
635 	    result == KERN_RESOURCE_SHORTAGE ||
636 	    result == KERN_PROTECTION_FAILURE ||
637 	    result == KERN_OUT_OF_BOUNDS,
638 	    ("Unexpected Mach error %d from vm_fault()", result));
639 #ifdef KTRACE
640 	if (map != kernel_map && KTRPOINT(curthread, KTR_FAULTEND))
641 		ktrfaultend(result);
642 #endif
643 	if (result != KERN_SUCCESS && signo != NULL) {
644 		switch (result) {
645 		case KERN_FAILURE:
646 		case KERN_INVALID_ADDRESS:
647 			*signo = SIGSEGV;
648 			*ucode = SEGV_MAPERR;
649 			break;
650 		case KERN_RESOURCE_SHORTAGE:
651 			*signo = SIGBUS;
652 			*ucode = BUS_OOMERR;
653 			break;
654 		case KERN_OUT_OF_BOUNDS:
655 			*signo = SIGBUS;
656 			*ucode = BUS_OBJERR;
657 			break;
658 		case KERN_PROTECTION_FAILURE:
659 			if (prot_fault_translation == 0) {
660 				/*
661 				 * Autodetect.  This check also covers
662 				 * the images without the ABI-tag ELF
663 				 * note.
664 				 */
665 				if (SV_CURPROC_ABI() == SV_ABI_FREEBSD &&
666 				    curproc->p_osrel >= P_OSREL_SIGSEGV) {
667 					*signo = SIGSEGV;
668 					*ucode = SEGV_ACCERR;
669 				} else {
670 					*signo = SIGBUS;
671 					*ucode = UCODE_PAGEFLT;
672 				}
673 			} else if (prot_fault_translation == 1) {
674 				/* Always compat mode. */
675 				*signo = SIGBUS;
676 				*ucode = UCODE_PAGEFLT;
677 			} else {
678 				/* Always SIGSEGV mode. */
679 				*signo = SIGSEGV;
680 				*ucode = SEGV_ACCERR;
681 			}
682 			break;
683 		default:
684 			KASSERT(0, ("Unexpected Mach error %d from vm_fault()",
685 			    result));
686 			break;
687 		}
688 	}
689 	return (result);
690 }
691 
692 static int
693 vm_fault_lock_vnode(struct faultstate *fs, bool objlocked)
694 {
695 	struct vnode *vp;
696 	int error, locked;
697 
698 	if (fs->object->type != OBJT_VNODE)
699 		return (KERN_SUCCESS);
700 	vp = fs->object->handle;
701 	if (vp == fs->vp) {
702 		ASSERT_VOP_LOCKED(vp, "saved vnode is not locked");
703 		return (KERN_SUCCESS);
704 	}
705 
706 	/*
707 	 * Perform an unlock in case the desired vnode changed while
708 	 * the map was unlocked during a retry.
709 	 */
710 	unlock_vp(fs);
711 
712 	locked = VOP_ISLOCKED(vp);
713 	if (locked != LK_EXCLUSIVE)
714 		locked = LK_SHARED;
715 
716 	/*
717 	 * We must not sleep acquiring the vnode lock while we have
718 	 * the page exclusive busied or the object's
719 	 * paging-in-progress count incremented.  Otherwise, we could
720 	 * deadlock.
721 	 */
722 	error = vget(vp, locked | LK_CANRECURSE | LK_NOWAIT);
723 	if (error == 0) {
724 		fs->vp = vp;
725 		return (KERN_SUCCESS);
726 	}
727 
728 	vhold(vp);
729 	if (objlocked)
730 		unlock_and_deallocate(fs);
731 	else
732 		fault_deallocate(fs);
733 	error = vget(vp, locked | LK_RETRY | LK_CANRECURSE);
734 	vdrop(vp);
735 	fs->vp = vp;
736 	KASSERT(error == 0, ("vm_fault: vget failed %d", error));
737 	return (KERN_RESOURCE_SHORTAGE);
738 }
739 
740 /*
741  * Calculate the desired readahead.  Handle drop-behind.
742  *
743  * Returns the number of readahead blocks to pass to the pager.
744  */
745 static int
746 vm_fault_readahead(struct faultstate *fs)
747 {
748 	int era, nera;
749 	u_char behavior;
750 
751 	KASSERT(fs->lookup_still_valid, ("map unlocked"));
752 	era = fs->entry->read_ahead;
753 	behavior = vm_map_entry_behavior(fs->entry);
754 	if (behavior == MAP_ENTRY_BEHAV_RANDOM) {
755 		nera = 0;
756 	} else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) {
757 		nera = VM_FAULT_READ_AHEAD_MAX;
758 		if (fs->vaddr == fs->entry->next_read)
759 			vm_fault_dontneed(fs, fs->vaddr, nera);
760 	} else if (fs->vaddr == fs->entry->next_read) {
761 		/*
762 		 * This is a sequential fault.  Arithmetically
763 		 * increase the requested number of pages in
764 		 * the read-ahead window.  The requested
765 		 * number of pages is "# of sequential faults
766 		 * x (read ahead min + 1) + read ahead min"
767 		 */
768 		nera = VM_FAULT_READ_AHEAD_MIN;
769 		if (era > 0) {
770 			nera += era + 1;
771 			if (nera > VM_FAULT_READ_AHEAD_MAX)
772 				nera = VM_FAULT_READ_AHEAD_MAX;
773 		}
774 		if (era == VM_FAULT_READ_AHEAD_MAX)
775 			vm_fault_dontneed(fs, fs->vaddr, nera);
776 	} else {
777 		/*
778 		 * This is a non-sequential fault.
779 		 */
780 		nera = 0;
781 	}
782 	if (era != nera) {
783 		/*
784 		 * A read lock on the map suffices to update
785 		 * the read ahead count safely.
786 		 */
787 		fs->entry->read_ahead = nera;
788 	}
789 
790 	return (nera);
791 }
792 
793 static int
794 vm_fault_lookup(struct faultstate *fs)
795 {
796 	int result;
797 
798 	KASSERT(!fs->lookup_still_valid,
799 	   ("vm_fault_lookup: Map already locked."));
800 	result = vm_map_lookup(&fs->map, fs->vaddr, fs->fault_type |
801 	    VM_PROT_FAULT_LOOKUP, &fs->entry, &fs->first_object,
802 	    &fs->first_pindex, &fs->prot, &fs->wired);
803 	if (result != KERN_SUCCESS) {
804 		unlock_vp(fs);
805 		return (result);
806 	}
807 
808 	fs->map_generation = fs->map->timestamp;
809 
810 	if (fs->entry->eflags & MAP_ENTRY_NOFAULT) {
811 		panic("%s: fault on nofault entry, addr: %#lx",
812 		    __func__, (u_long)fs->vaddr);
813 	}
814 
815 	if (fs->entry->eflags & MAP_ENTRY_IN_TRANSITION &&
816 	    fs->entry->wiring_thread != curthread) {
817 		vm_map_unlock_read(fs->map);
818 		vm_map_lock(fs->map);
819 		if (vm_map_lookup_entry(fs->map, fs->vaddr, &fs->entry) &&
820 		    (fs->entry->eflags & MAP_ENTRY_IN_TRANSITION)) {
821 			unlock_vp(fs);
822 			fs->entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
823 			vm_map_unlock_and_wait(fs->map, 0);
824 		} else
825 			vm_map_unlock(fs->map);
826 		return (KERN_RESOURCE_SHORTAGE);
827 	}
828 
829 	MPASS((fs->entry->eflags & MAP_ENTRY_GUARD) == 0);
830 
831 	if (fs->wired)
832 		fs->fault_type = fs->prot | (fs->fault_type & VM_PROT_COPY);
833 	else
834 		KASSERT((fs->fault_flags & VM_FAULT_WIRE) == 0,
835 		    ("!fs->wired && VM_FAULT_WIRE"));
836 	fs->lookup_still_valid = true;
837 
838 	return (KERN_SUCCESS);
839 }
840 
841 static int
842 vm_fault_relookup(struct faultstate *fs)
843 {
844 	vm_object_t retry_object;
845 	vm_pindex_t retry_pindex;
846 	vm_prot_t retry_prot;
847 	int result;
848 
849 	if (!vm_map_trylock_read(fs->map))
850 		return (KERN_RESTART);
851 
852 	fs->lookup_still_valid = true;
853 	if (fs->map->timestamp == fs->map_generation)
854 		return (KERN_SUCCESS);
855 
856 	result = vm_map_lookup_locked(&fs->map, fs->vaddr, fs->fault_type,
857 	    &fs->entry, &retry_object, &retry_pindex, &retry_prot,
858 	    &fs->wired);
859 	if (result != KERN_SUCCESS) {
860 		/*
861 		 * If retry of map lookup would have blocked then
862 		 * retry fault from start.
863 		 */
864 		if (result == KERN_FAILURE)
865 			return (KERN_RESTART);
866 		return (result);
867 	}
868 	if (retry_object != fs->first_object ||
869 	    retry_pindex != fs->first_pindex)
870 		return (KERN_RESTART);
871 
872 	/*
873 	 * Check whether the protection has changed or the object has
874 	 * been copied while we left the map unlocked. Changing from
875 	 * read to write permission is OK - we leave the page
876 	 * write-protected, and catch the write fault. Changing from
877 	 * write to read permission means that we can't mark the page
878 	 * write-enabled after all.
879 	 */
880 	fs->prot &= retry_prot;
881 	fs->fault_type &= retry_prot;
882 	if (fs->prot == 0)
883 		return (KERN_RESTART);
884 
885 	/* Reassert because wired may have changed. */
886 	KASSERT(fs->wired || (fs->fault_flags & VM_FAULT_WIRE) == 0,
887 	    ("!wired && VM_FAULT_WIRE"));
888 
889 	return (KERN_SUCCESS);
890 }
891 
892 static void
893 vm_fault_cow(struct faultstate *fs)
894 {
895 	bool is_first_object_locked;
896 
897 	/*
898 	 * This allows pages to be virtually copied from a backing_object
899 	 * into the first_object, where the backing object has no other
900 	 * refs to it, and cannot gain any more refs.  Instead of a bcopy,
901 	 * we just move the page from the backing object to the first
902 	 * object.  Note that we must mark the page dirty in the first
903 	 * object so that it will go out to swap when needed.
904 	 */
905 	is_first_object_locked = false;
906 	if (
907 	    /*
908 	     * Only one shadow object and no other refs.
909 	     */
910 	    fs->object->shadow_count == 1 && fs->object->ref_count == 1 &&
911 	    /*
912 	     * No other ways to look the object up
913 	     */
914 	    fs->object->handle == NULL && (fs->object->flags & OBJ_ANON) != 0 &&
915 	    /*
916 	     * We don't chase down the shadow chain and we can acquire locks.
917 	     */
918 	    (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs->first_object)) &&
919 	    fs->object == fs->first_object->backing_object &&
920 	    VM_OBJECT_TRYWLOCK(fs->object)) {
921 		/*
922 		 * Remove but keep xbusy for replace.  fs->m is moved into
923 		 * fs->first_object and left busy while fs->first_m is
924 		 * conditionally freed.
925 		 */
926 		vm_page_remove_xbusy(fs->m);
927 		vm_page_replace(fs->m, fs->first_object, fs->first_pindex,
928 		    fs->first_m);
929 		vm_page_dirty(fs->m);
930 #if VM_NRESERVLEVEL > 0
931 		/*
932 		 * Rename the reservation.
933 		 */
934 		vm_reserv_rename(fs->m, fs->first_object, fs->object,
935 		    OFF_TO_IDX(fs->first_object->backing_object_offset));
936 #endif
937 		VM_OBJECT_WUNLOCK(fs->object);
938 		VM_OBJECT_WUNLOCK(fs->first_object);
939 		fs->first_m = fs->m;
940 		fs->m = NULL;
941 		VM_CNT_INC(v_cow_optim);
942 	} else {
943 		if (is_first_object_locked)
944 			VM_OBJECT_WUNLOCK(fs->first_object);
945 		/*
946 		 * Oh, well, lets copy it.
947 		 */
948 		pmap_copy_page(fs->m, fs->first_m);
949 		vm_page_valid(fs->first_m);
950 		if (fs->wired && (fs->fault_flags & VM_FAULT_WIRE) == 0) {
951 			vm_page_wire(fs->first_m);
952 			vm_page_unwire(fs->m, PQ_INACTIVE);
953 		}
954 		/*
955 		 * Save the cow page to be released after
956 		 * pmap_enter is complete.
957 		 */
958 		fs->m_cow = fs->m;
959 		fs->m = NULL;
960 	}
961 	/*
962 	 * fs->object != fs->first_object due to above
963 	 * conditional
964 	 */
965 	vm_object_pip_wakeup(fs->object);
966 
967 	/*
968 	 * Only use the new page below...
969 	 */
970 	fs->object = fs->first_object;
971 	fs->pindex = fs->first_pindex;
972 	fs->m = fs->first_m;
973 	VM_CNT_INC(v_cow_faults);
974 	curthread->td_cow++;
975 }
976 
977 static bool
978 vm_fault_next(struct faultstate *fs)
979 {
980 	vm_object_t next_object;
981 
982 	/*
983 	 * The requested page does not exist at this object/
984 	 * offset.  Remove the invalid page from the object,
985 	 * waking up anyone waiting for it, and continue on to
986 	 * the next object.  However, if this is the top-level
987 	 * object, we must leave the busy page in place to
988 	 * prevent another process from rushing past us, and
989 	 * inserting the page in that object at the same time
990 	 * that we are.
991 	 */
992 	if (fs->object == fs->first_object) {
993 		fs->first_m = fs->m;
994 		fs->m = NULL;
995 	} else
996 		fault_page_free(&fs->m);
997 
998 	/*
999 	 * Move on to the next object.  Lock the next object before
1000 	 * unlocking the current one.
1001 	 */
1002 	VM_OBJECT_ASSERT_WLOCKED(fs->object);
1003 	next_object = fs->object->backing_object;
1004 	if (next_object == NULL)
1005 		return (false);
1006 	MPASS(fs->first_m != NULL);
1007 	KASSERT(fs->object != next_object, ("object loop %p", next_object));
1008 	VM_OBJECT_WLOCK(next_object);
1009 	vm_object_pip_add(next_object, 1);
1010 	if (fs->object != fs->first_object)
1011 		vm_object_pip_wakeup(fs->object);
1012 	fs->pindex += OFF_TO_IDX(fs->object->backing_object_offset);
1013 	VM_OBJECT_WUNLOCK(fs->object);
1014 	fs->object = next_object;
1015 
1016 	return (true);
1017 }
1018 
1019 static void
1020 vm_fault_zerofill(struct faultstate *fs)
1021 {
1022 
1023 	/*
1024 	 * If there's no object left, fill the page in the top
1025 	 * object with zeros.
1026 	 */
1027 	if (fs->object != fs->first_object) {
1028 		vm_object_pip_wakeup(fs->object);
1029 		fs->object = fs->first_object;
1030 		fs->pindex = fs->first_pindex;
1031 	}
1032 	MPASS(fs->first_m != NULL);
1033 	MPASS(fs->m == NULL);
1034 	fs->m = fs->first_m;
1035 	fs->first_m = NULL;
1036 
1037 	/*
1038 	 * Zero the page if necessary and mark it valid.
1039 	 */
1040 	if ((fs->m->flags & PG_ZERO) == 0) {
1041 		pmap_zero_page(fs->m);
1042 	} else {
1043 		VM_CNT_INC(v_ozfod);
1044 	}
1045 	VM_CNT_INC(v_zfod);
1046 	vm_page_valid(fs->m);
1047 }
1048 
1049 /*
1050  * Allocate a page directly or via the object populate method.
1051  */
1052 static int
1053 vm_fault_allocate(struct faultstate *fs)
1054 {
1055 	struct domainset *dset;
1056 	int alloc_req;
1057 	int rv;
1058 
1059 	if ((fs->object->flags & OBJ_SIZEVNLOCK) != 0) {
1060 		rv = vm_fault_lock_vnode(fs, true);
1061 		MPASS(rv == KERN_SUCCESS || rv == KERN_RESOURCE_SHORTAGE);
1062 		if (rv == KERN_RESOURCE_SHORTAGE)
1063 			return (rv);
1064 	}
1065 
1066 	if (fs->pindex >= fs->object->size)
1067 		return (KERN_OUT_OF_BOUNDS);
1068 
1069 	if (fs->object == fs->first_object &&
1070 	    (fs->first_object->flags & OBJ_POPULATE) != 0 &&
1071 	    fs->first_object->shadow_count == 0) {
1072 		rv = vm_fault_populate(fs);
1073 		switch (rv) {
1074 		case KERN_SUCCESS:
1075 		case KERN_FAILURE:
1076 		case KERN_RESTART:
1077 			return (rv);
1078 		case KERN_NOT_RECEIVER:
1079 			/*
1080 			 * Pager's populate() method
1081 			 * returned VM_PAGER_BAD.
1082 			 */
1083 			break;
1084 		default:
1085 			panic("inconsistent return codes");
1086 		}
1087 	}
1088 
1089 	/*
1090 	 * Allocate a new page for this object/offset pair.
1091 	 *
1092 	 * Unlocked read of the p_flag is harmless. At worst, the P_KILLED
1093 	 * might be not observed there, and allocation can fail, causing
1094 	 * restart and new reading of the p_flag.
1095 	 */
1096 	dset = fs->object->domain.dr_policy;
1097 	if (dset == NULL)
1098 		dset = curthread->td_domain.dr_policy;
1099 	if (!vm_page_count_severe_set(&dset->ds_mask) || P_KILLED(curproc)) {
1100 #if VM_NRESERVLEVEL > 0
1101 		vm_object_color(fs->object, atop(fs->vaddr) - fs->pindex);
1102 #endif
1103 		alloc_req = P_KILLED(curproc) ?
1104 		    VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL;
1105 		if (fs->object->type != OBJT_VNODE &&
1106 		    fs->object->backing_object == NULL)
1107 			alloc_req |= VM_ALLOC_ZERO;
1108 		fs->m = vm_page_alloc(fs->object, fs->pindex, alloc_req);
1109 	}
1110 	if (fs->m == NULL) {
1111 		unlock_and_deallocate(fs);
1112 		if (vm_pfault_oom_attempts < 0 ||
1113 		    fs->oom < vm_pfault_oom_attempts) {
1114 			fs->oom++;
1115 			vm_waitpfault(dset, vm_pfault_oom_wait * hz);
1116 		} else 	{
1117 			if (bootverbose)
1118 				printf(
1119 		"proc %d (%s) failed to alloc page on fault, starting OOM\n",
1120 				    curproc->p_pid, curproc->p_comm);
1121 			vm_pageout_oom(VM_OOM_MEM_PF);
1122 			fs->oom = 0;
1123 		}
1124 		return (KERN_RESOURCE_SHORTAGE);
1125 	}
1126 	fs->oom = 0;
1127 
1128 	return (KERN_NOT_RECEIVER);
1129 }
1130 
1131 /*
1132  * Call the pager to retrieve the page if there is a chance
1133  * that the pager has it, and potentially retrieve additional
1134  * pages at the same time.
1135  */
1136 static int
1137 vm_fault_getpages(struct faultstate *fs, int nera, int *behindp, int *aheadp)
1138 {
1139 	vm_offset_t e_end, e_start;
1140 	int ahead, behind, cluster_offset, rv;
1141 	u_char behavior;
1142 
1143 	/*
1144 	 * Prepare for unlocking the map.  Save the map
1145 	 * entry's start and end addresses, which are used to
1146 	 * optimize the size of the pager operation below.
1147 	 * Even if the map entry's addresses change after
1148 	 * unlocking the map, using the saved addresses is
1149 	 * safe.
1150 	 */
1151 	e_start = fs->entry->start;
1152 	e_end = fs->entry->end;
1153 	behavior = vm_map_entry_behavior(fs->entry);
1154 
1155 	/*
1156 	 * Release the map lock before locking the vnode or
1157 	 * sleeping in the pager.  (If the current object has
1158 	 * a shadow, then an earlier iteration of this loop
1159 	 * may have already unlocked the map.)
1160 	 */
1161 	unlock_map(fs);
1162 
1163 	rv = vm_fault_lock_vnode(fs, false);
1164 	MPASS(rv == KERN_SUCCESS || rv == KERN_RESOURCE_SHORTAGE);
1165 	if (rv == KERN_RESOURCE_SHORTAGE)
1166 		return (rv);
1167 	KASSERT(fs->vp == NULL || !fs->map->system_map,
1168 	    ("vm_fault: vnode-backed object mapped by system map"));
1169 
1170 	/*
1171 	 * Page in the requested page and hint the pager,
1172 	 * that it may bring up surrounding pages.
1173 	 */
1174 	if (nera == -1 || behavior == MAP_ENTRY_BEHAV_RANDOM ||
1175 	    P_KILLED(curproc)) {
1176 		behind = 0;
1177 		ahead = 0;
1178 	} else {
1179 		/* Is this a sequential fault? */
1180 		if (nera > 0) {
1181 			behind = 0;
1182 			ahead = nera;
1183 		} else {
1184 			/*
1185 			 * Request a cluster of pages that is
1186 			 * aligned to a VM_FAULT_READ_DEFAULT
1187 			 * page offset boundary within the
1188 			 * object.  Alignment to a page offset
1189 			 * boundary is more likely to coincide
1190 			 * with the underlying file system
1191 			 * block than alignment to a virtual
1192 			 * address boundary.
1193 			 */
1194 			cluster_offset = fs->pindex % VM_FAULT_READ_DEFAULT;
1195 			behind = ulmin(cluster_offset,
1196 			    atop(fs->vaddr - e_start));
1197 			ahead = VM_FAULT_READ_DEFAULT - 1 - cluster_offset;
1198 		}
1199 		ahead = ulmin(ahead, atop(e_end - fs->vaddr) - 1);
1200 	}
1201 	*behindp = behind;
1202 	*aheadp = ahead;
1203 	rv = vm_pager_get_pages(fs->object, &fs->m, 1, behindp, aheadp);
1204 	if (rv == VM_PAGER_OK)
1205 		return (KERN_SUCCESS);
1206 	if (rv == VM_PAGER_ERROR)
1207 		printf("vm_fault: pager read error, pid %d (%s)\n",
1208 		    curproc->p_pid, curproc->p_comm);
1209 	/*
1210 	 * If an I/O error occurred or the requested page was
1211 	 * outside the range of the pager, clean up and return
1212 	 * an error.
1213 	 */
1214 	if (rv == VM_PAGER_ERROR || rv == VM_PAGER_BAD)
1215 		return (KERN_OUT_OF_BOUNDS);
1216 	return (KERN_NOT_RECEIVER);
1217 }
1218 
1219 /*
1220  * Wait/Retry if the page is busy.  We have to do this if the page is
1221  * either exclusive or shared busy because the vm_pager may be using
1222  * read busy for pageouts (and even pageins if it is the vnode pager),
1223  * and we could end up trying to pagein and pageout the same page
1224  * simultaneously.
1225  *
1226  * We can theoretically allow the busy case on a read fault if the page
1227  * is marked valid, but since such pages are typically already pmap'd,
1228  * putting that special case in might be more effort then it is worth.
1229  * We cannot under any circumstances mess around with a shared busied
1230  * page except, perhaps, to pmap it.
1231  */
1232 static void
1233 vm_fault_busy_sleep(struct faultstate *fs)
1234 {
1235 	/*
1236 	 * Reference the page before unlocking and
1237 	 * sleeping so that the page daemon is less
1238 	 * likely to reclaim it.
1239 	 */
1240 	vm_page_aflag_set(fs->m, PGA_REFERENCED);
1241 	if (fs->object != fs->first_object) {
1242 		fault_page_release(&fs->first_m);
1243 		vm_object_pip_wakeup(fs->first_object);
1244 	}
1245 	vm_object_pip_wakeup(fs->object);
1246 	unlock_map(fs);
1247 	if (fs->m == vm_page_lookup(fs->object, fs->pindex))
1248 		vm_page_busy_sleep(fs->m, "vmpfw", false);
1249 	else
1250 		VM_OBJECT_WUNLOCK(fs->object);
1251 	VM_CNT_INC(v_intrans);
1252 	vm_object_deallocate(fs->first_object);
1253 }
1254 
1255 int
1256 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
1257     int fault_flags, vm_page_t *m_hold)
1258 {
1259 	struct faultstate fs;
1260 	int ahead, behind, faultcount;
1261 	int nera, result, rv;
1262 	bool dead, hardfault;
1263 
1264 	VM_CNT_INC(v_vm_faults);
1265 
1266 	if ((curthread->td_pflags & TDP_NOFAULTING) != 0)
1267 		return (KERN_PROTECTION_FAILURE);
1268 
1269 	fs.vp = NULL;
1270 	fs.vaddr = vaddr;
1271 	fs.m_hold = m_hold;
1272 	fs.fault_flags = fault_flags;
1273 	fs.map = map;
1274 	fs.lookup_still_valid = false;
1275 	fs.oom = 0;
1276 	faultcount = 0;
1277 	nera = -1;
1278 	hardfault = false;
1279 
1280 RetryFault:
1281 	fs.fault_type = fault_type;
1282 
1283 	/*
1284 	 * Find the backing store object and offset into it to begin the
1285 	 * search.
1286 	 */
1287 	result = vm_fault_lookup(&fs);
1288 	if (result != KERN_SUCCESS) {
1289 		if (result == KERN_RESOURCE_SHORTAGE)
1290 			goto RetryFault;
1291 		return (result);
1292 	}
1293 
1294 	/*
1295 	 * Try to avoid lock contention on the top-level object through
1296 	 * special-case handling of some types of page faults, specifically,
1297 	 * those that are mapping an existing page from the top-level object.
1298 	 * Under this condition, a read lock on the object suffices, allowing
1299 	 * multiple page faults of a similar type to run in parallel.
1300 	 */
1301 	if (fs.vp == NULL /* avoid locked vnode leak */ &&
1302 	    (fs.entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) == 0 &&
1303 	    (fs.fault_flags & (VM_FAULT_WIRE | VM_FAULT_DIRTY)) == 0) {
1304 		VM_OBJECT_RLOCK(fs.first_object);
1305 		rv = vm_fault_soft_fast(&fs);
1306 		if (rv == KERN_SUCCESS)
1307 			return (rv);
1308 		if (!VM_OBJECT_TRYUPGRADE(fs.first_object)) {
1309 			VM_OBJECT_RUNLOCK(fs.first_object);
1310 			VM_OBJECT_WLOCK(fs.first_object);
1311 		}
1312 	} else {
1313 		VM_OBJECT_WLOCK(fs.first_object);
1314 	}
1315 
1316 	/*
1317 	 * Make a reference to this object to prevent its disposal while we
1318 	 * are messing with it.  Once we have the reference, the map is free
1319 	 * to be diddled.  Since objects reference their shadows (and copies),
1320 	 * they will stay around as well.
1321 	 *
1322 	 * Bump the paging-in-progress count to prevent size changes (e.g.
1323 	 * truncation operations) during I/O.
1324 	 */
1325 	vm_object_reference_locked(fs.first_object);
1326 	vm_object_pip_add(fs.first_object, 1);
1327 
1328 	fs.m_cow = fs.m = fs.first_m = NULL;
1329 
1330 	/*
1331 	 * Search for the page at object/offset.
1332 	 */
1333 	fs.object = fs.first_object;
1334 	fs.pindex = fs.first_pindex;
1335 
1336 	if ((fs.entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) != 0) {
1337 		rv = vm_fault_allocate(&fs);
1338 		switch (rv) {
1339 		case KERN_RESTART:
1340 			unlock_and_deallocate(&fs);
1341 			/* FALLTHROUGH */
1342 		case KERN_RESOURCE_SHORTAGE:
1343 			goto RetryFault;
1344 		case KERN_SUCCESS:
1345 		case KERN_FAILURE:
1346 		case KERN_OUT_OF_BOUNDS:
1347 			unlock_and_deallocate(&fs);
1348 			return (rv);
1349 		case KERN_NOT_RECEIVER:
1350 			break;
1351 		default:
1352 			panic("vm_fault: Unhandled rv %d", rv);
1353 		}
1354 	}
1355 
1356 	while (TRUE) {
1357 		KASSERT(fs.m == NULL,
1358 		    ("page still set %p at loop start", fs.m));
1359 		/*
1360 		 * If the object is marked for imminent termination,
1361 		 * we retry here, since the collapse pass has raced
1362 		 * with us.  Otherwise, if we see terminally dead
1363 		 * object, return fail.
1364 		 */
1365 		if ((fs.object->flags & OBJ_DEAD) != 0) {
1366 			dead = fs.object->type == OBJT_DEAD;
1367 			unlock_and_deallocate(&fs);
1368 			if (dead)
1369 				return (KERN_PROTECTION_FAILURE);
1370 			pause("vmf_de", 1);
1371 			goto RetryFault;
1372 		}
1373 
1374 		/*
1375 		 * See if page is resident
1376 		 */
1377 		fs.m = vm_page_lookup(fs.object, fs.pindex);
1378 		if (fs.m != NULL) {
1379 			if (vm_page_tryxbusy(fs.m) == 0) {
1380 				vm_fault_busy_sleep(&fs);
1381 				goto RetryFault;
1382 			}
1383 
1384 			/*
1385 			 * The page is marked busy for other processes and the
1386 			 * pagedaemon.  If it still is completely valid we
1387 			 * are done.
1388 			 */
1389 			if (vm_page_all_valid(fs.m)) {
1390 				VM_OBJECT_WUNLOCK(fs.object);
1391 				break; /* break to PAGE HAS BEEN FOUND. */
1392 			}
1393 		}
1394 		VM_OBJECT_ASSERT_WLOCKED(fs.object);
1395 
1396 		/*
1397 		 * Page is not resident.  If the pager might contain the page
1398 		 * or this is the beginning of the search, allocate a new
1399 		 * page.  (Default objects are zero-fill, so there is no real
1400 		 * pager for them.)
1401 		 */
1402 		if (fs.m == NULL && (fs.object->type != OBJT_DEFAULT ||
1403 		    fs.object == fs.first_object)) {
1404 			rv = vm_fault_allocate(&fs);
1405 			switch (rv) {
1406 			case KERN_RESTART:
1407 				unlock_and_deallocate(&fs);
1408 				/* FALLTHROUGH */
1409 			case KERN_RESOURCE_SHORTAGE:
1410 				goto RetryFault;
1411 			case KERN_SUCCESS:
1412 			case KERN_FAILURE:
1413 			case KERN_OUT_OF_BOUNDS:
1414 				unlock_and_deallocate(&fs);
1415 				return (rv);
1416 			case KERN_NOT_RECEIVER:
1417 				break;
1418 			default:
1419 				panic("vm_fault: Unhandled rv %d", rv);
1420 			}
1421 		}
1422 
1423 		/*
1424 		 * Default objects have no pager so no exclusive busy exists
1425 		 * to protect this page in the chain.  Skip to the next
1426 		 * object without dropping the lock to preserve atomicity of
1427 		 * shadow faults.
1428 		 */
1429 		if (fs.object->type != OBJT_DEFAULT) {
1430 			/*
1431 			 * At this point, we have either allocated a new page
1432 			 * or found an existing page that is only partially
1433 			 * valid.
1434 			 *
1435 			 * We hold a reference on the current object and the
1436 			 * page is exclusive busied.  The exclusive busy
1437 			 * prevents simultaneous faults and collapses while
1438 			 * the object lock is dropped.
1439 		 	 */
1440 			VM_OBJECT_WUNLOCK(fs.object);
1441 
1442 			/*
1443 			 * If the pager for the current object might have
1444 			 * the page, then determine the number of additional
1445 			 * pages to read and potentially reprioritize
1446 			 * previously read pages for earlier reclamation.
1447 			 * These operations should only be performed once per
1448 			 * page fault.  Even if the current pager doesn't
1449 			 * have the page, the number of additional pages to
1450 			 * read will apply to subsequent objects in the
1451 			 * shadow chain.
1452 			 */
1453 			if (nera == -1 && !P_KILLED(curproc))
1454 				nera = vm_fault_readahead(&fs);
1455 
1456 			rv = vm_fault_getpages(&fs, nera, &behind, &ahead);
1457 			if (rv == KERN_SUCCESS) {
1458 				faultcount = behind + 1 + ahead;
1459 				hardfault = true;
1460 				break; /* break to PAGE HAS BEEN FOUND. */
1461 			}
1462 			if (rv == KERN_RESOURCE_SHORTAGE)
1463 				goto RetryFault;
1464 			VM_OBJECT_WLOCK(fs.object);
1465 			if (rv == KERN_OUT_OF_BOUNDS) {
1466 				fault_page_free(&fs.m);
1467 				unlock_and_deallocate(&fs);
1468 				return (rv);
1469 			}
1470 		}
1471 
1472 		/*
1473 		 * The page was not found in the current object.  Try to
1474 		 * traverse into a backing object or zero fill if none is
1475 		 * found.
1476 		 */
1477 		if (vm_fault_next(&fs))
1478 			continue;
1479 		if ((fs.fault_flags & VM_FAULT_NOFILL) != 0) {
1480 			if (fs.first_object == fs.object)
1481 				fault_page_free(&fs.first_m);
1482 			unlock_and_deallocate(&fs);
1483 			return (KERN_OUT_OF_BOUNDS);
1484 		}
1485 		VM_OBJECT_WUNLOCK(fs.object);
1486 		vm_fault_zerofill(&fs);
1487 		/* Don't try to prefault neighboring pages. */
1488 		faultcount = 1;
1489 		break;	/* break to PAGE HAS BEEN FOUND. */
1490 	}
1491 
1492 	/*
1493 	 * PAGE HAS BEEN FOUND.  A valid page has been found and exclusively
1494 	 * busied.  The object lock must no longer be held.
1495 	 */
1496 	vm_page_assert_xbusied(fs.m);
1497 	VM_OBJECT_ASSERT_UNLOCKED(fs.object);
1498 
1499 	/*
1500 	 * If the page is being written, but isn't already owned by the
1501 	 * top-level object, we have to copy it into a new page owned by the
1502 	 * top-level object.
1503 	 */
1504 	if (fs.object != fs.first_object) {
1505 		/*
1506 		 * We only really need to copy if we want to write it.
1507 		 */
1508 		if ((fs.fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1509 			vm_fault_cow(&fs);
1510 			/*
1511 			 * We only try to prefault read-only mappings to the
1512 			 * neighboring pages when this copy-on-write fault is
1513 			 * a hard fault.  In other cases, trying to prefault
1514 			 * is typically wasted effort.
1515 			 */
1516 			if (faultcount == 0)
1517 				faultcount = 1;
1518 
1519 		} else {
1520 			fs.prot &= ~VM_PROT_WRITE;
1521 		}
1522 	}
1523 
1524 	/*
1525 	 * We must verify that the maps have not changed since our last
1526 	 * lookup.
1527 	 */
1528 	if (!fs.lookup_still_valid) {
1529 		result = vm_fault_relookup(&fs);
1530 		if (result != KERN_SUCCESS) {
1531 			fault_deallocate(&fs);
1532 			if (result == KERN_RESTART)
1533 				goto RetryFault;
1534 			return (result);
1535 		}
1536 	}
1537 	VM_OBJECT_ASSERT_UNLOCKED(fs.object);
1538 
1539 	/*
1540 	 * If the page was filled by a pager, save the virtual address that
1541 	 * should be faulted on next under a sequential access pattern to the
1542 	 * map entry.  A read lock on the map suffices to update this address
1543 	 * safely.
1544 	 */
1545 	if (hardfault)
1546 		fs.entry->next_read = vaddr + ptoa(ahead) + PAGE_SIZE;
1547 
1548 	/*
1549 	 * Page must be completely valid or it is not fit to
1550 	 * map into user space.  vm_pager_get_pages() ensures this.
1551 	 */
1552 	vm_page_assert_xbusied(fs.m);
1553 	KASSERT(vm_page_all_valid(fs.m),
1554 	    ("vm_fault: page %p partially invalid", fs.m));
1555 
1556 	vm_fault_dirty(&fs, fs.m);
1557 
1558 	/*
1559 	 * Put this page into the physical map.  We had to do the unlock above
1560 	 * because pmap_enter() may sleep.  We don't put the page
1561 	 * back on the active queue until later so that the pageout daemon
1562 	 * won't find it (yet).
1563 	 */
1564 	pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot,
1565 	    fs.fault_type | (fs.wired ? PMAP_ENTER_WIRED : 0), 0);
1566 	if (faultcount != 1 && (fs.fault_flags & VM_FAULT_WIRE) == 0 &&
1567 	    fs.wired == 0)
1568 		vm_fault_prefault(&fs, vaddr,
1569 		    faultcount > 0 ? behind : PFBAK,
1570 		    faultcount > 0 ? ahead : PFFOR, false);
1571 
1572 	/*
1573 	 * If the page is not wired down, then put it where the pageout daemon
1574 	 * can find it.
1575 	 */
1576 	if ((fs.fault_flags & VM_FAULT_WIRE) != 0)
1577 		vm_page_wire(fs.m);
1578 	else
1579 		vm_page_activate(fs.m);
1580 	if (fs.m_hold != NULL) {
1581 		(*fs.m_hold) = fs.m;
1582 		vm_page_wire(fs.m);
1583 	}
1584 	vm_page_xunbusy(fs.m);
1585 	fs.m = NULL;
1586 
1587 	/*
1588 	 * Unlock everything, and return
1589 	 */
1590 	fault_deallocate(&fs);
1591 	if (hardfault) {
1592 		VM_CNT_INC(v_io_faults);
1593 		curthread->td_ru.ru_majflt++;
1594 #ifdef RACCT
1595 		if (racct_enable && fs.object->type == OBJT_VNODE) {
1596 			PROC_LOCK(curproc);
1597 			if ((fs.fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1598 				racct_add_force(curproc, RACCT_WRITEBPS,
1599 				    PAGE_SIZE + behind * PAGE_SIZE);
1600 				racct_add_force(curproc, RACCT_WRITEIOPS, 1);
1601 			} else {
1602 				racct_add_force(curproc, RACCT_READBPS,
1603 				    PAGE_SIZE + ahead * PAGE_SIZE);
1604 				racct_add_force(curproc, RACCT_READIOPS, 1);
1605 			}
1606 			PROC_UNLOCK(curproc);
1607 		}
1608 #endif
1609 	} else
1610 		curthread->td_ru.ru_minflt++;
1611 
1612 	return (KERN_SUCCESS);
1613 }
1614 
1615 /*
1616  * Speed up the reclamation of pages that precede the faulting pindex within
1617  * the first object of the shadow chain.  Essentially, perform the equivalent
1618  * to madvise(..., MADV_DONTNEED) on a large cluster of pages that precedes
1619  * the faulting pindex by the cluster size when the pages read by vm_fault()
1620  * cross a cluster-size boundary.  The cluster size is the greater of the
1621  * smallest superpage size and VM_FAULT_DONTNEED_MIN.
1622  *
1623  * When "fs->first_object" is a shadow object, the pages in the backing object
1624  * that precede the faulting pindex are deactivated by vm_fault().  So, this
1625  * function must only be concerned with pages in the first object.
1626  */
1627 static void
1628 vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr, int ahead)
1629 {
1630 	vm_map_entry_t entry;
1631 	vm_object_t first_object, object;
1632 	vm_offset_t end, start;
1633 	vm_page_t m, m_next;
1634 	vm_pindex_t pend, pstart;
1635 	vm_size_t size;
1636 
1637 	object = fs->object;
1638 	VM_OBJECT_ASSERT_UNLOCKED(object);
1639 	first_object = fs->first_object;
1640 	/* Neither fictitious nor unmanaged pages can be reclaimed. */
1641 	if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) {
1642 		VM_OBJECT_RLOCK(first_object);
1643 		size = VM_FAULT_DONTNEED_MIN;
1644 		if (MAXPAGESIZES > 1 && size < pagesizes[1])
1645 			size = pagesizes[1];
1646 		end = rounddown2(vaddr, size);
1647 		if (vaddr - end >= size - PAGE_SIZE - ptoa(ahead) &&
1648 		    (entry = fs->entry)->start < end) {
1649 			if (end - entry->start < size)
1650 				start = entry->start;
1651 			else
1652 				start = end - size;
1653 			pmap_advise(fs->map->pmap, start, end, MADV_DONTNEED);
1654 			pstart = OFF_TO_IDX(entry->offset) + atop(start -
1655 			    entry->start);
1656 			m_next = vm_page_find_least(first_object, pstart);
1657 			pend = OFF_TO_IDX(entry->offset) + atop(end -
1658 			    entry->start);
1659 			while ((m = m_next) != NULL && m->pindex < pend) {
1660 				m_next = TAILQ_NEXT(m, listq);
1661 				if (!vm_page_all_valid(m) ||
1662 				    vm_page_busied(m))
1663 					continue;
1664 
1665 				/*
1666 				 * Don't clear PGA_REFERENCED, since it would
1667 				 * likely represent a reference by a different
1668 				 * process.
1669 				 *
1670 				 * Typically, at this point, prefetched pages
1671 				 * are still in the inactive queue.  Only
1672 				 * pages that triggered page faults are in the
1673 				 * active queue.  The test for whether the page
1674 				 * is in the inactive queue is racy; in the
1675 				 * worst case we will requeue the page
1676 				 * unnecessarily.
1677 				 */
1678 				if (!vm_page_inactive(m))
1679 					vm_page_deactivate(m);
1680 			}
1681 		}
1682 		VM_OBJECT_RUNLOCK(first_object);
1683 	}
1684 }
1685 
1686 /*
1687  * vm_fault_prefault provides a quick way of clustering
1688  * pagefaults into a processes address space.  It is a "cousin"
1689  * of vm_map_pmap_enter, except it runs at page fault time instead
1690  * of mmap time.
1691  */
1692 static void
1693 vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
1694     int backward, int forward, bool obj_locked)
1695 {
1696 	pmap_t pmap;
1697 	vm_map_entry_t entry;
1698 	vm_object_t backing_object, lobject;
1699 	vm_offset_t addr, starta;
1700 	vm_pindex_t pindex;
1701 	vm_page_t m;
1702 	int i;
1703 
1704 	pmap = fs->map->pmap;
1705 	if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
1706 		return;
1707 
1708 	entry = fs->entry;
1709 
1710 	if (addra < backward * PAGE_SIZE) {
1711 		starta = entry->start;
1712 	} else {
1713 		starta = addra - backward * PAGE_SIZE;
1714 		if (starta < entry->start)
1715 			starta = entry->start;
1716 	}
1717 
1718 	/*
1719 	 * Generate the sequence of virtual addresses that are candidates for
1720 	 * prefaulting in an outward spiral from the faulting virtual address,
1721 	 * "addra".  Specifically, the sequence is "addra - PAGE_SIZE", "addra
1722 	 * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ...
1723 	 * If the candidate address doesn't have a backing physical page, then
1724 	 * the loop immediately terminates.
1725 	 */
1726 	for (i = 0; i < 2 * imax(backward, forward); i++) {
1727 		addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE :
1728 		    PAGE_SIZE);
1729 		if (addr > addra + forward * PAGE_SIZE)
1730 			addr = 0;
1731 
1732 		if (addr < starta || addr >= entry->end)
1733 			continue;
1734 
1735 		if (!pmap_is_prefaultable(pmap, addr))
1736 			continue;
1737 
1738 		pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1739 		lobject = entry->object.vm_object;
1740 		if (!obj_locked)
1741 			VM_OBJECT_RLOCK(lobject);
1742 		while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1743 		    lobject->type == OBJT_DEFAULT &&
1744 		    (backing_object = lobject->backing_object) != NULL) {
1745 			KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1746 			    0, ("vm_fault_prefault: unaligned object offset"));
1747 			pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1748 			VM_OBJECT_RLOCK(backing_object);
1749 			if (!obj_locked || lobject != entry->object.vm_object)
1750 				VM_OBJECT_RUNLOCK(lobject);
1751 			lobject = backing_object;
1752 		}
1753 		if (m == NULL) {
1754 			if (!obj_locked || lobject != entry->object.vm_object)
1755 				VM_OBJECT_RUNLOCK(lobject);
1756 			break;
1757 		}
1758 		if (vm_page_all_valid(m) &&
1759 		    (m->flags & PG_FICTITIOUS) == 0)
1760 			pmap_enter_quick(pmap, addr, m, entry->protection);
1761 		if (!obj_locked || lobject != entry->object.vm_object)
1762 			VM_OBJECT_RUNLOCK(lobject);
1763 	}
1764 }
1765 
1766 /*
1767  * Hold each of the physical pages that are mapped by the specified range of
1768  * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
1769  * and allow the specified types of access, "prot".  If all of the implied
1770  * pages are successfully held, then the number of held pages is returned
1771  * together with pointers to those pages in the array "ma".  However, if any
1772  * of the pages cannot be held, -1 is returned.
1773  */
1774 int
1775 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
1776     vm_prot_t prot, vm_page_t *ma, int max_count)
1777 {
1778 	vm_offset_t end, va;
1779 	vm_page_t *mp;
1780 	int count;
1781 	boolean_t pmap_failed;
1782 
1783 	if (len == 0)
1784 		return (0);
1785 	end = round_page(addr + len);
1786 	addr = trunc_page(addr);
1787 
1788 	if (!vm_map_range_valid(map, addr, end))
1789 		return (-1);
1790 
1791 	if (atop(end - addr) > max_count)
1792 		panic("vm_fault_quick_hold_pages: count > max_count");
1793 	count = atop(end - addr);
1794 
1795 	/*
1796 	 * Most likely, the physical pages are resident in the pmap, so it is
1797 	 * faster to try pmap_extract_and_hold() first.
1798 	 */
1799 	pmap_failed = FALSE;
1800 	for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) {
1801 		*mp = pmap_extract_and_hold(map->pmap, va, prot);
1802 		if (*mp == NULL)
1803 			pmap_failed = TRUE;
1804 		else if ((prot & VM_PROT_WRITE) != 0 &&
1805 		    (*mp)->dirty != VM_PAGE_BITS_ALL) {
1806 			/*
1807 			 * Explicitly dirty the physical page.  Otherwise, the
1808 			 * caller's changes may go unnoticed because they are
1809 			 * performed through an unmanaged mapping or by a DMA
1810 			 * operation.
1811 			 *
1812 			 * The object lock is not held here.
1813 			 * See vm_page_clear_dirty_mask().
1814 			 */
1815 			vm_page_dirty(*mp);
1816 		}
1817 	}
1818 	if (pmap_failed) {
1819 		/*
1820 		 * One or more pages could not be held by the pmap.  Either no
1821 		 * page was mapped at the specified virtual address or that
1822 		 * mapping had insufficient permissions.  Attempt to fault in
1823 		 * and hold these pages.
1824 		 *
1825 		 * If vm_fault_disable_pagefaults() was called,
1826 		 * i.e., TDP_NOFAULTING is set, we must not sleep nor
1827 		 * acquire MD VM locks, which means we must not call
1828 		 * vm_fault().  Some (out of tree) callers mark
1829 		 * too wide a code area with vm_fault_disable_pagefaults()
1830 		 * already, use the VM_PROT_QUICK_NOFAULT flag to request
1831 		 * the proper behaviour explicitly.
1832 		 */
1833 		if ((prot & VM_PROT_QUICK_NOFAULT) != 0 &&
1834 		    (curthread->td_pflags & TDP_NOFAULTING) != 0)
1835 			goto error;
1836 		for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE)
1837 			if (*mp == NULL && vm_fault(map, va, prot,
1838 			    VM_FAULT_NORMAL, mp) != KERN_SUCCESS)
1839 				goto error;
1840 	}
1841 	return (count);
1842 error:
1843 	for (mp = ma; mp < ma + count; mp++)
1844 		if (*mp != NULL)
1845 			vm_page_unwire(*mp, PQ_INACTIVE);
1846 	return (-1);
1847 }
1848 
1849 /*
1850  *	Routine:
1851  *		vm_fault_copy_entry
1852  *	Function:
1853  *		Create new shadow object backing dst_entry with private copy of
1854  *		all underlying pages. When src_entry is equal to dst_entry,
1855  *		function implements COW for wired-down map entry. Otherwise,
1856  *		it forks wired entry into dst_map.
1857  *
1858  *	In/out conditions:
1859  *		The source and destination maps must be locked for write.
1860  *		The source map entry must be wired down (or be a sharing map
1861  *		entry corresponding to a main map entry that is wired down).
1862  */
1863 void
1864 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1865     vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1866     vm_ooffset_t *fork_charge)
1867 {
1868 	vm_object_t backing_object, dst_object, object, src_object;
1869 	vm_pindex_t dst_pindex, pindex, src_pindex;
1870 	vm_prot_t access, prot;
1871 	vm_offset_t vaddr;
1872 	vm_page_t dst_m;
1873 	vm_page_t src_m;
1874 	boolean_t upgrade;
1875 
1876 #ifdef	lint
1877 	src_map++;
1878 #endif	/* lint */
1879 
1880 	upgrade = src_entry == dst_entry;
1881 	access = prot = dst_entry->protection;
1882 
1883 	src_object = src_entry->object.vm_object;
1884 	src_pindex = OFF_TO_IDX(src_entry->offset);
1885 
1886 	if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
1887 		dst_object = src_object;
1888 		vm_object_reference(dst_object);
1889 	} else {
1890 		/*
1891 		 * Create the top-level object for the destination entry.
1892 		 * Doesn't actually shadow anything - we copy the pages
1893 		 * directly.
1894 		 */
1895 		dst_object = vm_object_allocate_anon(atop(dst_entry->end -
1896 		    dst_entry->start), NULL, NULL, 0);
1897 #if VM_NRESERVLEVEL > 0
1898 		dst_object->flags |= OBJ_COLORED;
1899 		dst_object->pg_color = atop(dst_entry->start);
1900 #endif
1901 		dst_object->domain = src_object->domain;
1902 		dst_object->charge = dst_entry->end - dst_entry->start;
1903 	}
1904 
1905 	VM_OBJECT_WLOCK(dst_object);
1906 	KASSERT(upgrade || dst_entry->object.vm_object == NULL,
1907 	    ("vm_fault_copy_entry: vm_object not NULL"));
1908 	if (src_object != dst_object) {
1909 		dst_entry->object.vm_object = dst_object;
1910 		dst_entry->offset = 0;
1911 		dst_entry->eflags &= ~MAP_ENTRY_VN_EXEC;
1912 	}
1913 	if (fork_charge != NULL) {
1914 		KASSERT(dst_entry->cred == NULL,
1915 		    ("vm_fault_copy_entry: leaked swp charge"));
1916 		dst_object->cred = curthread->td_ucred;
1917 		crhold(dst_object->cred);
1918 		*fork_charge += dst_object->charge;
1919 	} else if ((dst_object->type == OBJT_DEFAULT ||
1920 	    dst_object->type == OBJT_SWAP) &&
1921 	    dst_object->cred == NULL) {
1922 		KASSERT(dst_entry->cred != NULL, ("no cred for entry %p",
1923 		    dst_entry));
1924 		dst_object->cred = dst_entry->cred;
1925 		dst_entry->cred = NULL;
1926 	}
1927 
1928 	/*
1929 	 * If not an upgrade, then enter the mappings in the pmap as
1930 	 * read and/or execute accesses.  Otherwise, enter them as
1931 	 * write accesses.
1932 	 *
1933 	 * A writeable large page mapping is only created if all of
1934 	 * the constituent small page mappings are modified. Marking
1935 	 * PTEs as modified on inception allows promotion to happen
1936 	 * without taking potentially large number of soft faults.
1937 	 */
1938 	if (!upgrade)
1939 		access &= ~VM_PROT_WRITE;
1940 
1941 	/*
1942 	 * Loop through all of the virtual pages within the entry's
1943 	 * range, copying each page from the source object to the
1944 	 * destination object.  Since the source is wired, those pages
1945 	 * must exist.  In contrast, the destination is pageable.
1946 	 * Since the destination object doesn't share any backing storage
1947 	 * with the source object, all of its pages must be dirtied,
1948 	 * regardless of whether they can be written.
1949 	 */
1950 	for (vaddr = dst_entry->start, dst_pindex = 0;
1951 	    vaddr < dst_entry->end;
1952 	    vaddr += PAGE_SIZE, dst_pindex++) {
1953 again:
1954 		/*
1955 		 * Find the page in the source object, and copy it in.
1956 		 * Because the source is wired down, the page will be
1957 		 * in memory.
1958 		 */
1959 		if (src_object != dst_object)
1960 			VM_OBJECT_RLOCK(src_object);
1961 		object = src_object;
1962 		pindex = src_pindex + dst_pindex;
1963 		while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
1964 		    (backing_object = object->backing_object) != NULL) {
1965 			/*
1966 			 * Unless the source mapping is read-only or
1967 			 * it is presently being upgraded from
1968 			 * read-only, the first object in the shadow
1969 			 * chain should provide all of the pages.  In
1970 			 * other words, this loop body should never be
1971 			 * executed when the source mapping is already
1972 			 * read/write.
1973 			 */
1974 			KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 ||
1975 			    upgrade,
1976 			    ("vm_fault_copy_entry: main object missing page"));
1977 
1978 			VM_OBJECT_RLOCK(backing_object);
1979 			pindex += OFF_TO_IDX(object->backing_object_offset);
1980 			if (object != dst_object)
1981 				VM_OBJECT_RUNLOCK(object);
1982 			object = backing_object;
1983 		}
1984 		KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing"));
1985 
1986 		if (object != dst_object) {
1987 			/*
1988 			 * Allocate a page in the destination object.
1989 			 */
1990 			dst_m = vm_page_alloc(dst_object, (src_object ==
1991 			    dst_object ? src_pindex : 0) + dst_pindex,
1992 			    VM_ALLOC_NORMAL);
1993 			if (dst_m == NULL) {
1994 				VM_OBJECT_WUNLOCK(dst_object);
1995 				VM_OBJECT_RUNLOCK(object);
1996 				vm_wait(dst_object);
1997 				VM_OBJECT_WLOCK(dst_object);
1998 				goto again;
1999 			}
2000 			pmap_copy_page(src_m, dst_m);
2001 			VM_OBJECT_RUNLOCK(object);
2002 			dst_m->dirty = dst_m->valid = src_m->valid;
2003 		} else {
2004 			dst_m = src_m;
2005 			if (vm_page_busy_acquire(dst_m, VM_ALLOC_WAITFAIL) == 0)
2006 				goto again;
2007 			if (dst_m->pindex >= dst_object->size) {
2008 				/*
2009 				 * We are upgrading.  Index can occur
2010 				 * out of bounds if the object type is
2011 				 * vnode and the file was truncated.
2012 				 */
2013 				vm_page_xunbusy(dst_m);
2014 				break;
2015 			}
2016 		}
2017 		VM_OBJECT_WUNLOCK(dst_object);
2018 
2019 		/*
2020 		 * Enter it in the pmap. If a wired, copy-on-write
2021 		 * mapping is being replaced by a write-enabled
2022 		 * mapping, then wire that new mapping.
2023 		 *
2024 		 * The page can be invalid if the user called
2025 		 * msync(MS_INVALIDATE) or truncated the backing vnode
2026 		 * or shared memory object.  In this case, do not
2027 		 * insert it into pmap, but still do the copy so that
2028 		 * all copies of the wired map entry have similar
2029 		 * backing pages.
2030 		 */
2031 		if (vm_page_all_valid(dst_m)) {
2032 			pmap_enter(dst_map->pmap, vaddr, dst_m, prot,
2033 			    access | (upgrade ? PMAP_ENTER_WIRED : 0), 0);
2034 		}
2035 
2036 		/*
2037 		 * Mark it no longer busy, and put it on the active list.
2038 		 */
2039 		VM_OBJECT_WLOCK(dst_object);
2040 
2041 		if (upgrade) {
2042 			if (src_m != dst_m) {
2043 				vm_page_unwire(src_m, PQ_INACTIVE);
2044 				vm_page_wire(dst_m);
2045 			} else {
2046 				KASSERT(vm_page_wired(dst_m),
2047 				    ("dst_m %p is not wired", dst_m));
2048 			}
2049 		} else {
2050 			vm_page_activate(dst_m);
2051 		}
2052 		vm_page_xunbusy(dst_m);
2053 	}
2054 	VM_OBJECT_WUNLOCK(dst_object);
2055 	if (upgrade) {
2056 		dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
2057 		vm_object_deallocate(src_object);
2058 	}
2059 }
2060 
2061 /*
2062  * Block entry into the machine-independent layer's page fault handler by
2063  * the calling thread.  Subsequent calls to vm_fault() by that thread will
2064  * return KERN_PROTECTION_FAILURE.  Enable machine-dependent handling of
2065  * spurious page faults.
2066  */
2067 int
2068 vm_fault_disable_pagefaults(void)
2069 {
2070 
2071 	return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR));
2072 }
2073 
2074 void
2075 vm_fault_enable_pagefaults(int save)
2076 {
2077 
2078 	curthread_pflags_restore(save);
2079 }
2080