xref: /freebsd/sys/vm/vm_fault.c (revision 729362425c09cf6b362366aabc6fb547eee8035a)
1 /*
2  * Copyright (c) 1991, 1993
3  *	The Regents of the University of California.  All rights reserved.
4  * Copyright (c) 1994 John S. Dyson
5  * All rights reserved.
6  * Copyright (c) 1994 David Greenman
7  * All rights reserved.
8  *
9  *
10  * This code is derived from software contributed to Berkeley by
11  * The Mach Operating System project at Carnegie-Mellon University.
12  *
13  * Redistribution and use in source and binary forms, with or without
14  * modification, are permitted provided that the following conditions
15  * are met:
16  * 1. Redistributions of source code must retain the above copyright
17  *    notice, this list of conditions and the following disclaimer.
18  * 2. Redistributions in binary form must reproduce the above copyright
19  *    notice, this list of conditions and the following disclaimer in the
20  *    documentation and/or other materials provided with the distribution.
21  * 3. All advertising materials mentioning features or use of this software
22  *    must display the following acknowledgement:
23  *	This product includes software developed by the University of
24  *	California, Berkeley and its contributors.
25  * 4. Neither the name of the University nor the names of its contributors
26  *    may be used to endorse or promote products derived from this software
27  *    without specific prior written permission.
28  *
29  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
30  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
31  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
32  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
33  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
34  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
35  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
36  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
37  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
38  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
39  * SUCH DAMAGE.
40  *
41  *	from: @(#)vm_fault.c	8.4 (Berkeley) 1/12/94
42  *
43  *
44  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
45  * All rights reserved.
46  *
47  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
48  *
49  * Permission to use, copy, modify and distribute this software and
50  * its documentation is hereby granted, provided that both the copyright
51  * notice and this permission notice appear in all copies of the
52  * software, derivative works or modified versions, and any portions
53  * thereof, and that both notices appear in supporting documentation.
54  *
55  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
56  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
57  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
58  *
59  * Carnegie Mellon requests users of this software to return to
60  *
61  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
62  *  School of Computer Science
63  *  Carnegie Mellon University
64  *  Pittsburgh PA 15213-3890
65  *
66  * any improvements or extensions that they make and grant Carnegie the
67  * rights to redistribute these changes.
68  *
69  * $FreeBSD$
70  */
71 
72 /*
73  *	Page fault handling module.
74  */
75 #include <sys/param.h>
76 #include <sys/systm.h>
77 #include <sys/kernel.h>
78 #include <sys/lock.h>
79 #include <sys/mutex.h>
80 #include <sys/proc.h>
81 #include <sys/resourcevar.h>
82 #include <sys/sysctl.h>
83 #include <sys/vmmeter.h>
84 #include <sys/vnode.h>
85 
86 #include <vm/vm.h>
87 #include <vm/vm_param.h>
88 #include <vm/pmap.h>
89 #include <vm/vm_map.h>
90 #include <vm/vm_object.h>
91 #include <vm/vm_page.h>
92 #include <vm/vm_pageout.h>
93 #include <vm/vm_kern.h>
94 #include <vm/vm_pager.h>
95 #include <vm/vnode_pager.h>
96 #include <vm/vm_extern.h>
97 
98 static int vm_fault_additional_pages(vm_page_t, int, int, vm_page_t *, int *);
99 
100 #define VM_FAULT_READ_AHEAD 8
101 #define VM_FAULT_READ_BEHIND 7
102 #define VM_FAULT_READ (VM_FAULT_READ_AHEAD+VM_FAULT_READ_BEHIND+1)
103 
104 struct faultstate {
105 	vm_page_t m;
106 	vm_object_t object;
107 	vm_pindex_t pindex;
108 	vm_page_t first_m;
109 	vm_object_t	first_object;
110 	vm_pindex_t first_pindex;
111 	vm_map_t map;
112 	vm_map_entry_t entry;
113 	int lookup_still_valid;
114 	struct vnode *vp;
115 };
116 
117 static __inline void
118 release_page(struct faultstate *fs)
119 {
120 	vm_page_lock_queues();
121 	vm_page_wakeup(fs->m);
122 	vm_page_deactivate(fs->m);
123 	vm_page_unlock_queues();
124 	fs->m = NULL;
125 }
126 
127 static __inline void
128 unlock_map(struct faultstate *fs)
129 {
130 	if (fs->lookup_still_valid) {
131 		vm_map_lookup_done(fs->map, fs->entry);
132 		fs->lookup_still_valid = FALSE;
133 	}
134 }
135 
136 static void
137 _unlock_things(struct faultstate *fs, int dealloc)
138 {
139 	GIANT_REQUIRED;
140 	vm_object_pip_wakeup(fs->object);
141 	if (fs->object != fs->first_object) {
142 		vm_page_lock_queues();
143 		vm_page_free(fs->first_m);
144 		vm_page_unlock_queues();
145 		vm_object_pip_wakeup(fs->first_object);
146 		fs->first_m = NULL;
147 	}
148 	if (dealloc) {
149 		vm_object_deallocate(fs->first_object);
150 	}
151 	unlock_map(fs);
152 	if (fs->vp != NULL) {
153 		vput(fs->vp);
154 		fs->vp = NULL;
155 	}
156 }
157 
158 #define unlock_things(fs) _unlock_things(fs, 0)
159 #define unlock_and_deallocate(fs) _unlock_things(fs, 1)
160 
161 /*
162  * TRYPAGER - used by vm_fault to calculate whether the pager for the
163  *	      current object *might* contain the page.
164  *
165  *	      default objects are zero-fill, there is no real pager.
166  */
167 #define TRYPAGER	(fs.object->type != OBJT_DEFAULT && \
168 			(((fault_flags & VM_FAULT_WIRE_MASK) == 0) || wired))
169 
170 /*
171  *	vm_fault:
172  *
173  *	Handle a page fault occurring at the given address,
174  *	requiring the given permissions, in the map specified.
175  *	If successful, the page is inserted into the
176  *	associated physical map.
177  *
178  *	NOTE: the given address should be truncated to the
179  *	proper page address.
180  *
181  *	KERN_SUCCESS is returned if the page fault is handled; otherwise,
182  *	a standard error specifying why the fault is fatal is returned.
183  *
184  *
185  *	The map in question must be referenced, and remains so.
186  *	Caller may hold no locks.
187  */
188 int
189 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
190 	 int fault_flags)
191 {
192 	vm_prot_t prot;
193 	int result;
194 	boolean_t growstack, wired;
195 	int map_generation;
196 	vm_object_t next_object;
197 	vm_page_t marray[VM_FAULT_READ];
198 	int hardfault;
199 	int faultcount;
200 	struct faultstate fs;
201 
202 	hardfault = 0;
203 	growstack = TRUE;
204 	atomic_add_int(&cnt.v_vm_faults, 1);
205 
206 	mtx_lock(&Giant);
207 RetryFault:;
208 
209 	/*
210 	 * Find the backing store object and offset into it to begin the
211 	 * search.
212 	 */
213 	fs.map = map;
214 	result = vm_map_lookup(&fs.map, vaddr, fault_type, &fs.entry,
215 	    &fs.first_object, &fs.first_pindex, &prot, &wired);
216 	if (result != KERN_SUCCESS) {
217 		if (result != KERN_PROTECTION_FAILURE ||
218 		    (fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE) {
219 			if (growstack && result == KERN_INVALID_ADDRESS &&
220 			    map != kernel_map && curproc != NULL) {
221 				result = vm_map_growstack(curproc, vaddr);
222 				if (result != KERN_SUCCESS) {
223 					mtx_unlock(&Giant);
224 					return (KERN_FAILURE);
225 				}
226 				growstack = FALSE;
227 				goto RetryFault;
228 			}
229 			mtx_unlock(&Giant);
230 			return (result);
231 		}
232 
233 		/*
234    		 * If we are user-wiring a r/w segment, and it is COW, then
235    		 * we need to do the COW operation.  Note that we don't COW
236    		 * currently RO sections now, because it is NOT desirable
237    		 * to COW .text.  We simply keep .text from ever being COW'ed
238    		 * and take the heat that one cannot debug wired .text sections.
239    		 */
240 		result = vm_map_lookup(&fs.map, vaddr,
241 			VM_PROT_READ|VM_PROT_WRITE|VM_PROT_OVERRIDE_WRITE,
242 			&fs.entry, &fs.first_object, &fs.first_pindex, &prot, &wired);
243 		if (result != KERN_SUCCESS) {
244 			mtx_unlock(&Giant);
245 			return (result);
246 		}
247 
248 		/*
249 		 * If we don't COW now, on a user wire, the user will never
250 		 * be able to write to the mapping.  If we don't make this
251 		 * restriction, the bookkeeping would be nearly impossible.
252 		 *
253 		 * XXX The following assignment modifies the map without
254 		 * holding a write lock on it.
255 		 */
256 		if ((fs.entry->protection & VM_PROT_WRITE) == 0)
257 			fs.entry->max_protection &= ~VM_PROT_WRITE;
258 	}
259 
260 	map_generation = fs.map->timestamp;
261 
262 	if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
263 		panic("vm_fault: fault on nofault entry, addr: %lx",
264 		    (u_long)vaddr);
265 	}
266 
267 	/*
268 	 * Make a reference to this object to prevent its disposal while we
269 	 * are messing with it.  Once we have the reference, the map is free
270 	 * to be diddled.  Since objects reference their shadows (and copies),
271 	 * they will stay around as well.
272 	 *
273 	 * Bump the paging-in-progress count to prevent size changes (e.g.
274 	 * truncation operations) during I/O.  This must be done after
275 	 * obtaining the vnode lock in order to avoid possible deadlocks.
276 	 *
277 	 * XXX vnode_pager_lock() can block without releasing the map lock.
278 	 */
279 	vm_object_reference(fs.first_object);
280 	fs.vp = vnode_pager_lock(fs.first_object);
281 	vm_object_pip_add(fs.first_object, 1);
282 
283 	fs.lookup_still_valid = TRUE;
284 
285 	if (wired)
286 		fault_type = prot;
287 
288 	fs.first_m = NULL;
289 
290 	/*
291 	 * Search for the page at object/offset.
292 	 */
293 	fs.object = fs.first_object;
294 	fs.pindex = fs.first_pindex;
295 	while (TRUE) {
296 		/*
297 		 * If the object is dead, we stop here
298 		 */
299 		if (fs.object->flags & OBJ_DEAD) {
300 			unlock_and_deallocate(&fs);
301 			mtx_unlock(&Giant);
302 			return (KERN_PROTECTION_FAILURE);
303 		}
304 
305 		/*
306 		 * See if page is resident
307 		 */
308 		fs.m = vm_page_lookup(fs.object, fs.pindex);
309 		if (fs.m != NULL) {
310 			int queue, s;
311 
312 			/*
313 			 * check for page-based copy on write.
314 			 * We check fs.object == fs.first_object so
315 			 * as to ensure the legacy COW mechanism is
316 			 * used when the page in question is part of
317 			 * a shadow object.  Otherwise, vm_page_cowfault()
318 			 * removes the page from the backing object,
319 			 * which is not what we want.
320 			 */
321 			vm_page_lock_queues();
322 			if ((fs.m->cow) &&
323 			    (fault_type & VM_PROT_WRITE) &&
324 			    (fs.object == fs.first_object)) {
325 				s = splvm();
326 				vm_page_cowfault(fs.m);
327 				splx(s);
328 				vm_page_unlock_queues();
329 				unlock_things(&fs);
330 				goto RetryFault;
331 			}
332 
333 			/*
334 			 * Wait/Retry if the page is busy.  We have to do this
335 			 * if the page is busy via either PG_BUSY or
336 			 * vm_page_t->busy because the vm_pager may be using
337 			 * vm_page_t->busy for pageouts ( and even pageins if
338 			 * it is the vnode pager ), and we could end up trying
339 			 * to pagein and pageout the same page simultaneously.
340 			 *
341 			 * We can theoretically allow the busy case on a read
342 			 * fault if the page is marked valid, but since such
343 			 * pages are typically already pmap'd, putting that
344 			 * special case in might be more effort then it is
345 			 * worth.  We cannot under any circumstances mess
346 			 * around with a vm_page_t->busy page except, perhaps,
347 			 * to pmap it.
348 			 */
349 			if ((fs.m->flags & PG_BUSY) || fs.m->busy) {
350 				vm_page_unlock_queues();
351 				unlock_things(&fs);
352 				vm_page_lock_queues();
353 				if (!vm_page_sleep_if_busy(fs.m, TRUE, "vmpfw"))
354 					vm_page_unlock_queues();
355 				cnt.v_intrans++;
356 				vm_object_deallocate(fs.first_object);
357 				goto RetryFault;
358 			}
359 			queue = fs.m->queue;
360 
361 			s = splvm();
362 			vm_pageq_remove_nowakeup(fs.m);
363 			splx(s);
364 
365 			if ((queue - fs.m->pc) == PQ_CACHE && vm_page_count_severe()) {
366 				vm_page_activate(fs.m);
367 				vm_page_unlock_queues();
368 				unlock_and_deallocate(&fs);
369 				VM_WAITPFAULT;
370 				goto RetryFault;
371 			}
372 
373 			/*
374 			 * Mark page busy for other processes, and the
375 			 * pagedaemon.  If it still isn't completely valid
376 			 * (readable), jump to readrest, else break-out ( we
377 			 * found the page ).
378 			 */
379 			vm_page_busy(fs.m);
380 			vm_page_unlock_queues();
381 			if (((fs.m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) &&
382 				fs.m->object != kernel_object && fs.m->object != kmem_object) {
383 				goto readrest;
384 			}
385 
386 			break;
387 		}
388 
389 		/*
390 		 * Page is not resident, If this is the search termination
391 		 * or the pager might contain the page, allocate a new page.
392 		 */
393 		if (TRYPAGER || fs.object == fs.first_object) {
394 			if (fs.pindex >= fs.object->size) {
395 				unlock_and_deallocate(&fs);
396 				mtx_unlock(&Giant);
397 				return (KERN_PROTECTION_FAILURE);
398 			}
399 
400 			/*
401 			 * Allocate a new page for this object/offset pair.
402 			 */
403 			fs.m = NULL;
404 			if (!vm_page_count_severe()) {
405 				fs.m = vm_page_alloc(fs.object, fs.pindex,
406 				    (fs.vp || fs.object->backing_object)? VM_ALLOC_NORMAL: VM_ALLOC_ZERO);
407 			}
408 			if (fs.m == NULL) {
409 				unlock_and_deallocate(&fs);
410 				VM_WAITPFAULT;
411 				goto RetryFault;
412 			}
413 		}
414 
415 readrest:
416 		/*
417 		 * We have found a valid page or we have allocated a new page.
418 		 * The page thus may not be valid or may not be entirely
419 		 * valid.
420 		 *
421 		 * Attempt to fault-in the page if there is a chance that the
422 		 * pager has it, and potentially fault in additional pages
423 		 * at the same time.
424 		 */
425 		if (TRYPAGER) {
426 			int rv;
427 			int reqpage;
428 			int ahead, behind;
429 			u_char behavior = vm_map_entry_behavior(fs.entry);
430 
431 			if (behavior == MAP_ENTRY_BEHAV_RANDOM) {
432 				ahead = 0;
433 				behind = 0;
434 			} else {
435 				behind = (vaddr - fs.entry->start) >> PAGE_SHIFT;
436 				if (behind > VM_FAULT_READ_BEHIND)
437 					behind = VM_FAULT_READ_BEHIND;
438 
439 				ahead = ((fs.entry->end - vaddr) >> PAGE_SHIFT) - 1;
440 				if (ahead > VM_FAULT_READ_AHEAD)
441 					ahead = VM_FAULT_READ_AHEAD;
442 			}
443 
444 			if ((fs.first_object->type != OBJT_DEVICE) &&
445 			    (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL ||
446                                 (behavior != MAP_ENTRY_BEHAV_RANDOM &&
447                                 fs.pindex >= fs.entry->lastr &&
448                                 fs.pindex < fs.entry->lastr + VM_FAULT_READ))
449 			) {
450 				vm_pindex_t firstpindex, tmppindex;
451 
452 				if (fs.first_pindex < 2 * VM_FAULT_READ)
453 					firstpindex = 0;
454 				else
455 					firstpindex = fs.first_pindex - 2 * VM_FAULT_READ;
456 
457 				vm_page_lock_queues();
458 				/*
459 				 * note: partially valid pages cannot be
460 				 * included in the lookahead - NFS piecemeal
461 				 * writes will barf on it badly.
462 				 */
463 				for (tmppindex = fs.first_pindex - 1;
464 					tmppindex >= firstpindex;
465 					--tmppindex) {
466 					vm_page_t mt;
467 
468 					mt = vm_page_lookup(fs.first_object, tmppindex);
469 					if (mt == NULL || (mt->valid != VM_PAGE_BITS_ALL))
470 						break;
471 					if (mt->busy ||
472 						(mt->flags & (PG_BUSY | PG_FICTITIOUS | PG_UNMANAGED)) ||
473 						mt->hold_count ||
474 						mt->wire_count)
475 						continue;
476 					if (mt->dirty == 0)
477 						vm_page_test_dirty(mt);
478 					if (mt->dirty) {
479 						pmap_remove_all(mt);
480 						vm_page_deactivate(mt);
481 					} else {
482 						vm_page_cache(mt);
483 					}
484 				}
485 				vm_page_unlock_queues();
486 				ahead += behind;
487 				behind = 0;
488 			}
489 
490 			/*
491 			 * now we find out if any other pages should be paged
492 			 * in at this time this routine checks to see if the
493 			 * pages surrounding this fault reside in the same
494 			 * object as the page for this fault.  If they do,
495 			 * then they are faulted in also into the object.  The
496 			 * array "marray" returned contains an array of
497 			 * vm_page_t structs where one of them is the
498 			 * vm_page_t passed to the routine.  The reqpage
499 			 * return value is the index into the marray for the
500 			 * vm_page_t passed to the routine.
501 			 *
502 			 * fs.m plus the additional pages are PG_BUSY'd.
503 			 *
504 			 * XXX vm_fault_additional_pages() can block
505 			 * without releasing the map lock.
506 			 */
507 			faultcount = vm_fault_additional_pages(
508 			    fs.m, behind, ahead, marray, &reqpage);
509 
510 			/*
511 			 * update lastr imperfectly (we do not know how much
512 			 * getpages will actually read), but good enough.
513 			 *
514 			 * XXX The following assignment modifies the map
515 			 * without holding a write lock on it.
516 			 */
517 			fs.entry->lastr = fs.pindex + faultcount - behind;
518 
519 			/*
520 			 * Call the pager to retrieve the data, if any, after
521 			 * releasing the lock on the map.  We hold a ref on
522 			 * fs.object and the pages are PG_BUSY'd.
523 			 */
524 			unlock_map(&fs);
525 
526 			rv = faultcount ?
527 			    vm_pager_get_pages(fs.object, marray, faultcount,
528 				reqpage) : VM_PAGER_FAIL;
529 
530 			if (rv == VM_PAGER_OK) {
531 				/*
532 				 * Found the page. Leave it busy while we play
533 				 * with it.
534 				 */
535 
536 				/*
537 				 * Relookup in case pager changed page. Pager
538 				 * is responsible for disposition of old page
539 				 * if moved.
540 				 */
541 				fs.m = vm_page_lookup(fs.object, fs.pindex);
542 				if (!fs.m) {
543 					unlock_and_deallocate(&fs);
544 					goto RetryFault;
545 				}
546 
547 				hardfault++;
548 				break; /* break to PAGE HAS BEEN FOUND */
549 			}
550 			/*
551 			 * Remove the bogus page (which does not exist at this
552 			 * object/offset); before doing so, we must get back
553 			 * our object lock to preserve our invariant.
554 			 *
555 			 * Also wake up any other process that may want to bring
556 			 * in this page.
557 			 *
558 			 * If this is the top-level object, we must leave the
559 			 * busy page to prevent another process from rushing
560 			 * past us, and inserting the page in that object at
561 			 * the same time that we are.
562 			 */
563 			if (rv == VM_PAGER_ERROR)
564 				printf("vm_fault: pager read error, pid %d (%s)\n",
565 				    curproc->p_pid, curproc->p_comm);
566 			/*
567 			 * Data outside the range of the pager or an I/O error
568 			 */
569 			/*
570 			 * XXX - the check for kernel_map is a kludge to work
571 			 * around having the machine panic on a kernel space
572 			 * fault w/ I/O error.
573 			 */
574 			if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) ||
575 				(rv == VM_PAGER_BAD)) {
576 				vm_page_lock_queues();
577 				vm_page_free(fs.m);
578 				vm_page_unlock_queues();
579 				fs.m = NULL;
580 				unlock_and_deallocate(&fs);
581 				mtx_unlock(&Giant);
582 				return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
583 			}
584 			if (fs.object != fs.first_object) {
585 				vm_page_lock_queues();
586 				vm_page_free(fs.m);
587 				vm_page_unlock_queues();
588 				fs.m = NULL;
589 				/*
590 				 * XXX - we cannot just fall out at this
591 				 * point, m has been freed and is invalid!
592 				 */
593 			}
594 		}
595 
596 		/*
597 		 * We get here if the object has default pager (or unwiring)
598 		 * or the pager doesn't have the page.
599 		 */
600 		if (fs.object == fs.first_object)
601 			fs.first_m = fs.m;
602 
603 		/*
604 		 * Move on to the next object.  Lock the next object before
605 		 * unlocking the current one.
606 		 */
607 		fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset);
608 		next_object = fs.object->backing_object;
609 		if (next_object == NULL) {
610 			/*
611 			 * If there's no object left, fill the page in the top
612 			 * object with zeros.
613 			 */
614 			if (fs.object != fs.first_object) {
615 				vm_object_pip_wakeup(fs.object);
616 
617 				fs.object = fs.first_object;
618 				fs.pindex = fs.first_pindex;
619 				fs.m = fs.first_m;
620 			}
621 			fs.first_m = NULL;
622 
623 			/*
624 			 * Zero the page if necessary and mark it valid.
625 			 */
626 			if ((fs.m->flags & PG_ZERO) == 0) {
627 				pmap_zero_page(fs.m);
628 			} else {
629 				cnt.v_ozfod++;
630 			}
631 			cnt.v_zfod++;
632 			fs.m->valid = VM_PAGE_BITS_ALL;
633 			break;	/* break to PAGE HAS BEEN FOUND */
634 		} else {
635 			if (fs.object != fs.first_object) {
636 				vm_object_pip_wakeup(fs.object);
637 			}
638 			KASSERT(fs.object != next_object, ("object loop %p", next_object));
639 			fs.object = next_object;
640 			vm_object_pip_add(fs.object, 1);
641 		}
642 	}
643 
644 	KASSERT((fs.m->flags & PG_BUSY) != 0,
645 	    ("vm_fault: not busy after main loop"));
646 
647 	/*
648 	 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
649 	 * is held.]
650 	 */
651 
652 	/*
653 	 * If the page is being written, but isn't already owned by the
654 	 * top-level object, we have to copy it into a new page owned by the
655 	 * top-level object.
656 	 */
657 	if (fs.object != fs.first_object) {
658 		/*
659 		 * We only really need to copy if we want to write it.
660 		 */
661 		if (fault_type & VM_PROT_WRITE) {
662 			/*
663 			 * This allows pages to be virtually copied from a
664 			 * backing_object into the first_object, where the
665 			 * backing object has no other refs to it, and cannot
666 			 * gain any more refs.  Instead of a bcopy, we just
667 			 * move the page from the backing object to the
668 			 * first object.  Note that we must mark the page
669 			 * dirty in the first object so that it will go out
670 			 * to swap when needed.
671 			 */
672 			if (map_generation == fs.map->timestamp &&
673 				/*
674 				 * Only one shadow object
675 				 */
676 				(fs.object->shadow_count == 1) &&
677 				/*
678 				 * No COW refs, except us
679 				 */
680 				(fs.object->ref_count == 1) &&
681 				/*
682 				 * No one else can look this object up
683 				 */
684 				(fs.object->handle == NULL) &&
685 				/*
686 				 * No other ways to look the object up
687 				 */
688 				((fs.object->type == OBJT_DEFAULT) ||
689 				 (fs.object->type == OBJT_SWAP)) &&
690 				/*
691 				 * We don't chase down the shadow chain
692 				 */
693 				(fs.object == fs.first_object->backing_object) &&
694 
695 				/*
696 				 * grab the lock if we need to
697 				 */
698 			    (fs.lookup_still_valid || vm_map_trylock(fs.map))) {
699 
700 				fs.lookup_still_valid = 1;
701 				/*
702 				 * get rid of the unnecessary page
703 				 */
704 				vm_page_lock_queues();
705 				pmap_remove_all(fs.first_m);
706 				vm_page_free(fs.first_m);
707 				fs.first_m = NULL;
708 
709 				/*
710 				 * grab the page and put it into the
711 				 * process'es object.  The page is
712 				 * automatically made dirty.
713 				 */
714 				vm_page_rename(fs.m, fs.first_object, fs.first_pindex);
715 				fs.first_m = fs.m;
716 				vm_page_busy(fs.first_m);
717 				vm_page_unlock_queues();
718 				fs.m = NULL;
719 				cnt.v_cow_optim++;
720 			} else {
721 				/*
722 				 * Oh, well, lets copy it.
723 				 */
724 				vm_page_copy(fs.m, fs.first_m);
725 			}
726 
727 			if (fs.m) {
728 				/*
729 				 * We no longer need the old page or object.
730 				 */
731 				release_page(&fs);
732 			}
733 
734 			/*
735 			 * fs.object != fs.first_object due to above
736 			 * conditional
737 			 */
738 			vm_object_pip_wakeup(fs.object);
739 
740 			/*
741 			 * Only use the new page below...
742 			 */
743 			cnt.v_cow_faults++;
744 			fs.m = fs.first_m;
745 			fs.object = fs.first_object;
746 			fs.pindex = fs.first_pindex;
747 
748 		} else {
749 			prot &= ~VM_PROT_WRITE;
750 		}
751 	}
752 
753 	/*
754 	 * We must verify that the maps have not changed since our last
755 	 * lookup.
756 	 */
757 	if (!fs.lookup_still_valid &&
758 		(fs.map->timestamp != map_generation)) {
759 		vm_object_t retry_object;
760 		vm_pindex_t retry_pindex;
761 		vm_prot_t retry_prot;
762 
763 		/*
764 		 * Since map entries may be pageable, make sure we can take a
765 		 * page fault on them.
766 		 */
767 
768 		/*
769 		 * Unlock vnode before the lookup to avoid deadlock.   E.G.
770 		 * avoid a deadlock between the inode and exec_map that can
771 		 * occur due to locks being obtained in different orders.
772 		 */
773 		if (fs.vp != NULL) {
774 			vput(fs.vp);
775 			fs.vp = NULL;
776 		}
777 
778 		if (fs.map->infork) {
779 			release_page(&fs);
780 			unlock_and_deallocate(&fs);
781 			goto RetryFault;
782 		}
783 
784 		/*
785 		 * To avoid trying to write_lock the map while another process
786 		 * has it read_locked (in vm_map_pageable), we do not try for
787 		 * write permission.  If the page is still writable, we will
788 		 * get write permission.  If it is not, or has been marked
789 		 * needs_copy, we enter the mapping without write permission,
790 		 * and will merely take another fault.
791 		 */
792 		result = vm_map_lookup(&fs.map, vaddr, fault_type & ~VM_PROT_WRITE,
793 		    &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
794 		map_generation = fs.map->timestamp;
795 
796 		/*
797 		 * If we don't need the page any longer, put it on the active
798 		 * list (the easiest thing to do here).  If no one needs it,
799 		 * pageout will grab it eventually.
800 		 */
801 		if (result != KERN_SUCCESS) {
802 			release_page(&fs);
803 			unlock_and_deallocate(&fs);
804 			mtx_unlock(&Giant);
805 			return (result);
806 		}
807 		fs.lookup_still_valid = TRUE;
808 
809 		if ((retry_object != fs.first_object) ||
810 		    (retry_pindex != fs.first_pindex)) {
811 			release_page(&fs);
812 			unlock_and_deallocate(&fs);
813 			goto RetryFault;
814 		}
815 		/*
816 		 * Check whether the protection has changed or the object has
817 		 * been copied while we left the map unlocked. Changing from
818 		 * read to write permission is OK - we leave the page
819 		 * write-protected, and catch the write fault. Changing from
820 		 * write to read permission means that we can't mark the page
821 		 * write-enabled after all.
822 		 */
823 		prot &= retry_prot;
824 	}
825 
826 	/*
827 	 * Put this page into the physical map. We had to do the unlock above
828 	 * because pmap_enter may cause other faults.   We don't put the page
829 	 * back on the active queue until later so that the page-out daemon
830 	 * won't find us (yet).
831 	 */
832 
833 	if (prot & VM_PROT_WRITE) {
834 		vm_page_lock_queues();
835 		vm_page_flag_set(fs.m, PG_WRITEABLE);
836 		vm_object_set_writeable_dirty(fs.m->object);
837 
838 		/*
839 		 * If the fault is a write, we know that this page is being
840 		 * written NOW so dirty it explicitly to save on
841 		 * pmap_is_modified() calls later.
842 		 *
843 		 * If this is a NOSYNC mmap we do not want to set PG_NOSYNC
844 		 * if the page is already dirty to prevent data written with
845 		 * the expectation of being synced from not being synced.
846 		 * Likewise if this entry does not request NOSYNC then make
847 		 * sure the page isn't marked NOSYNC.  Applications sharing
848 		 * data should use the same flags to avoid ping ponging.
849 		 *
850 		 * Also tell the backing pager, if any, that it should remove
851 		 * any swap backing since the page is now dirty.
852 		 */
853 		if (fs.entry->eflags & MAP_ENTRY_NOSYNC) {
854 			if (fs.m->dirty == 0)
855 				vm_page_flag_set(fs.m, PG_NOSYNC);
856 		} else {
857 			vm_page_flag_clear(fs.m, PG_NOSYNC);
858 		}
859 		vm_page_unlock_queues();
860 		if (fault_flags & VM_FAULT_DIRTY) {
861 			int s;
862 			vm_page_dirty(fs.m);
863 			s = splvm();
864 			vm_pager_page_unswapped(fs.m);
865 			splx(s);
866 		}
867 	}
868 
869 	/*
870 	 * Page had better still be busy
871 	 */
872 	KASSERT(fs.m->flags & PG_BUSY,
873 		("vm_fault: page %p not busy!", fs.m));
874 	unlock_things(&fs);
875 
876 	/*
877 	 * Sanity check: page must be completely valid or it is not fit to
878 	 * map into user space.  vm_pager_get_pages() ensures this.
879 	 */
880 	if (fs.m->valid != VM_PAGE_BITS_ALL) {
881 		vm_page_zero_invalid(fs.m, TRUE);
882 		printf("Warning: page %p partially invalid on fault\n", fs.m);
883 	}
884 	pmap_enter(fs.map->pmap, vaddr, fs.m, prot, wired);
885 	if (((fault_flags & VM_FAULT_WIRE_MASK) == 0) && (wired == 0)) {
886 		pmap_prefault(fs.map->pmap, vaddr, fs.entry);
887 	}
888 	vm_page_lock_queues();
889 	vm_page_flag_clear(fs.m, PG_ZERO);
890 	vm_page_flag_set(fs.m, PG_REFERENCED);
891 
892 	/*
893 	 * If the page is not wired down, then put it where the pageout daemon
894 	 * can find it.
895 	 */
896 	if (fault_flags & VM_FAULT_WIRE_MASK) {
897 		if (wired)
898 			vm_page_wire(fs.m);
899 		else
900 			vm_page_unwire(fs.m, 1);
901 	} else {
902 		vm_page_activate(fs.m);
903 	}
904 	vm_page_wakeup(fs.m);
905 	vm_page_unlock_queues();
906 	mtx_lock_spin(&sched_lock);
907 	if (curproc && (curproc->p_sflag & PS_INMEM) && curproc->p_stats) {
908 		if (hardfault) {
909 			curproc->p_stats->p_ru.ru_majflt++;
910 		} else {
911 			curproc->p_stats->p_ru.ru_minflt++;
912 		}
913 	}
914 	mtx_unlock_spin(&sched_lock);
915 
916 	/*
917 	 * Unlock everything, and return
918 	 */
919 	vm_object_deallocate(fs.first_object);
920 	mtx_unlock(&Giant);
921 	return (KERN_SUCCESS);
922 }
923 
924 /*
925  *	vm_fault_quick:
926  *
927  *	Ensure that the requested virtual address, which may be in userland,
928  *	is valid.  Fault-in the page if necessary.  Return -1 on failure.
929  */
930 int
931 vm_fault_quick(caddr_t v, int prot)
932 {
933 	int r;
934 
935 	if (prot & VM_PROT_WRITE)
936 		r = subyte(v, fubyte(v));
937 	else
938 		r = fubyte(v);
939 	return(r);
940 }
941 
942 /*
943  *	vm_fault_wire:
944  *
945  *	Wire down a range of virtual addresses in a map.
946  */
947 int
948 vm_fault_wire(map, start, end, user_wire)
949 	vm_map_t map;
950 	vm_offset_t start, end;
951 	boolean_t user_wire;
952 {
953 	vm_offset_t va;
954 	int rv;
955 
956 	/*
957 	 * We simulate a fault to get the page and enter it in the physical
958 	 * map.  For user wiring, we only ask for read access on currently
959 	 * read-only sections.
960 	 */
961 	for (va = start; va < end; va += PAGE_SIZE) {
962 		rv = vm_fault(map, va,
963 		    user_wire ? VM_PROT_READ : VM_PROT_READ | VM_PROT_WRITE,
964 		    user_wire ? VM_FAULT_USER_WIRE : VM_FAULT_CHANGE_WIRING);
965 		if (rv) {
966 			if (va != start)
967 				vm_fault_unwire(map, start, va);
968 			return (rv);
969 		}
970 	}
971 	return (KERN_SUCCESS);
972 }
973 
974 /*
975  *	vm_fault_unwire:
976  *
977  *	Unwire a range of virtual addresses in a map.
978  */
979 void
980 vm_fault_unwire(map, start, end)
981 	vm_map_t map;
982 	vm_offset_t start, end;
983 {
984 	vm_paddr_t pa;
985 	vm_offset_t va;
986 	pmap_t pmap;
987 
988 	pmap = vm_map_pmap(map);
989 
990 	mtx_lock(&Giant);
991 	/*
992 	 * Since the pages are wired down, we must be able to get their
993 	 * mappings from the physical map system.
994 	 */
995 	for (va = start; va < end; va += PAGE_SIZE) {
996 		pa = pmap_extract(pmap, va);
997 		if (pa != 0) {
998 			pmap_change_wiring(pmap, va, FALSE);
999 			vm_page_lock_queues();
1000 			vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1);
1001 			vm_page_unlock_queues();
1002 		}
1003 	}
1004 	mtx_unlock(&Giant);
1005 }
1006 
1007 /*
1008  *	Routine:
1009  *		vm_fault_copy_entry
1010  *	Function:
1011  *		Copy all of the pages from a wired-down map entry to another.
1012  *
1013  *	In/out conditions:
1014  *		The source and destination maps must be locked for write.
1015  *		The source map entry must be wired down (or be a sharing map
1016  *		entry corresponding to a main map entry that is wired down).
1017  */
1018 void
1019 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry)
1020 	vm_map_t dst_map;
1021 	vm_map_t src_map;
1022 	vm_map_entry_t dst_entry;
1023 	vm_map_entry_t src_entry;
1024 {
1025 	vm_object_t dst_object;
1026 	vm_object_t src_object;
1027 	vm_ooffset_t dst_offset;
1028 	vm_ooffset_t src_offset;
1029 	vm_prot_t prot;
1030 	vm_offset_t vaddr;
1031 	vm_page_t dst_m;
1032 	vm_page_t src_m;
1033 
1034 #ifdef	lint
1035 	src_map++;
1036 #endif	/* lint */
1037 
1038 	src_object = src_entry->object.vm_object;
1039 	src_offset = src_entry->offset;
1040 
1041 	/*
1042 	 * Create the top-level object for the destination entry. (Doesn't
1043 	 * actually shadow anything - we copy the pages directly.)
1044 	 */
1045 	dst_object = vm_object_allocate(OBJT_DEFAULT,
1046 	    (vm_size_t) OFF_TO_IDX(dst_entry->end - dst_entry->start));
1047 
1048 	dst_entry->object.vm_object = dst_object;
1049 	dst_entry->offset = 0;
1050 
1051 	prot = dst_entry->max_protection;
1052 
1053 	/*
1054 	 * Loop through all of the pages in the entry's range, copying each
1055 	 * one from the source object (it should be there) to the destination
1056 	 * object.
1057 	 */
1058 	for (vaddr = dst_entry->start, dst_offset = 0;
1059 	    vaddr < dst_entry->end;
1060 	    vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) {
1061 
1062 		/*
1063 		 * Allocate a page in the destination object
1064 		 */
1065 		do {
1066 			dst_m = vm_page_alloc(dst_object,
1067 				OFF_TO_IDX(dst_offset), VM_ALLOC_NORMAL);
1068 			if (dst_m == NULL) {
1069 				VM_WAIT;
1070 			}
1071 		} while (dst_m == NULL);
1072 
1073 		/*
1074 		 * Find the page in the source object, and copy it in.
1075 		 * (Because the source is wired down, the page will be in
1076 		 * memory.)
1077 		 */
1078 		src_m = vm_page_lookup(src_object,
1079 			OFF_TO_IDX(dst_offset + src_offset));
1080 		if (src_m == NULL)
1081 			panic("vm_fault_copy_wired: page missing");
1082 
1083 		vm_page_copy(src_m, dst_m);
1084 
1085 		/*
1086 		 * Enter it in the pmap...
1087 		 */
1088 		pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE);
1089 		vm_page_lock_queues();
1090 		vm_page_flag_set(dst_m, PG_WRITEABLE);
1091 
1092 		/*
1093 		 * Mark it no longer busy, and put it on the active list.
1094 		 */
1095 		vm_page_activate(dst_m);
1096 		vm_page_wakeup(dst_m);
1097 		vm_page_unlock_queues();
1098 	}
1099 }
1100 
1101 
1102 /*
1103  * This routine checks around the requested page for other pages that
1104  * might be able to be faulted in.  This routine brackets the viable
1105  * pages for the pages to be paged in.
1106  *
1107  * Inputs:
1108  *	m, rbehind, rahead
1109  *
1110  * Outputs:
1111  *  marray (array of vm_page_t), reqpage (index of requested page)
1112  *
1113  * Return value:
1114  *  number of pages in marray
1115  *
1116  * This routine can't block.
1117  */
1118 static int
1119 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage)
1120 	vm_page_t m;
1121 	int rbehind;
1122 	int rahead;
1123 	vm_page_t *marray;
1124 	int *reqpage;
1125 {
1126 	int i,j;
1127 	vm_object_t object;
1128 	vm_pindex_t pindex, startpindex, endpindex, tpindex;
1129 	vm_page_t rtm;
1130 	int cbehind, cahead;
1131 
1132 	GIANT_REQUIRED;
1133 
1134 	object = m->object;
1135 	pindex = m->pindex;
1136 
1137 	/*
1138 	 * we don't fault-ahead for device pager
1139 	 */
1140 	if (object->type == OBJT_DEVICE) {
1141 		*reqpage = 0;
1142 		marray[0] = m;
1143 		return 1;
1144 	}
1145 
1146 	/*
1147 	 * if the requested page is not available, then give up now
1148 	 */
1149 	if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1150 		return 0;
1151 	}
1152 
1153 	if ((cbehind == 0) && (cahead == 0)) {
1154 		*reqpage = 0;
1155 		marray[0] = m;
1156 		return 1;
1157 	}
1158 
1159 	if (rahead > cahead) {
1160 		rahead = cahead;
1161 	}
1162 
1163 	if (rbehind > cbehind) {
1164 		rbehind = cbehind;
1165 	}
1166 
1167 	/*
1168 	 * try to do any readahead that we might have free pages for.
1169 	 */
1170 	if ((rahead + rbehind) >
1171 		((cnt.v_free_count + cnt.v_cache_count) - cnt.v_free_reserved)) {
1172 		pagedaemon_wakeup();
1173 		marray[0] = m;
1174 		*reqpage = 0;
1175 		return 1;
1176 	}
1177 
1178 	/*
1179 	 * scan backward for the read behind pages -- in memory
1180 	 */
1181 	if (pindex > 0) {
1182 		if (rbehind > pindex) {
1183 			rbehind = pindex;
1184 			startpindex = 0;
1185 		} else {
1186 			startpindex = pindex - rbehind;
1187 		}
1188 
1189 		for (tpindex = pindex - 1; tpindex >= startpindex; tpindex -= 1) {
1190 			if (vm_page_lookup(object, tpindex)) {
1191 				startpindex = tpindex + 1;
1192 				break;
1193 			}
1194 			if (tpindex == 0)
1195 				break;
1196 		}
1197 
1198 		for (i = 0, tpindex = startpindex; tpindex < pindex; i++, tpindex++) {
1199 
1200 			rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
1201 			if (rtm == NULL) {
1202 				vm_page_lock_queues();
1203 				for (j = 0; j < i; j++) {
1204 					vm_page_free(marray[j]);
1205 				}
1206 				vm_page_unlock_queues();
1207 				marray[0] = m;
1208 				*reqpage = 0;
1209 				return 1;
1210 			}
1211 
1212 			marray[i] = rtm;
1213 		}
1214 	} else {
1215 		startpindex = 0;
1216 		i = 0;
1217 	}
1218 
1219 	marray[i] = m;
1220 	/* page offset of the required page */
1221 	*reqpage = i;
1222 
1223 	tpindex = pindex + 1;
1224 	i++;
1225 
1226 	/*
1227 	 * scan forward for the read ahead pages
1228 	 */
1229 	endpindex = tpindex + rahead;
1230 	if (endpindex > object->size)
1231 		endpindex = object->size;
1232 
1233 	for (; tpindex < endpindex; i++, tpindex++) {
1234 
1235 		if (vm_page_lookup(object, tpindex)) {
1236 			break;
1237 		}
1238 
1239 		rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
1240 		if (rtm == NULL) {
1241 			break;
1242 		}
1243 
1244 		marray[i] = rtm;
1245 	}
1246 
1247 	/* return number of bytes of pages */
1248 	return i;
1249 }
1250