xref: /freebsd/sys/vm/vm_fault.c (revision ce4946daa5ce852d28008dac492029500ab2ee95)
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 
76 #include <sys/param.h>
77 #include <sys/systm.h>
78 #include <sys/lock.h>
79 #include <sys/mutex.h>
80 #include <sys/proc.h>
81 #include <sys/vnode.h>
82 #include <sys/resourcevar.h>
83 #include <sys/vmmeter.h>
84 
85 #include <vm/vm.h>
86 #include <vm/vm_param.h>
87 #include <vm/pmap.h>
88 #include <vm/vm_map.h>
89 #include <vm/vm_object.h>
90 #include <vm/vm_page.h>
91 #include <vm/vm_pageout.h>
92 #include <vm/vm_kern.h>
93 #include <vm/vm_pager.h>
94 #include <vm/vnode_pager.h>
95 #include <vm/vm_extern.h>
96 
97 static int vm_fault_additional_pages __P((vm_page_t, int,
98 					  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_wakeup(fs->m);
121 	vm_page_deactivate(fs->m);
122 	fs->m = NULL;
123 }
124 
125 static __inline void
126 unlock_map(struct faultstate *fs)
127 {
128 	if (fs->lookup_still_valid) {
129 		vm_map_lookup_done(fs->map, fs->entry);
130 		fs->lookup_still_valid = FALSE;
131 	}
132 }
133 
134 static void
135 _unlock_things(struct faultstate *fs, int dealloc)
136 {
137 	vm_object_pip_wakeup(fs->object);
138 	if (fs->object != fs->first_object) {
139 		vm_page_free(fs->first_m);
140 		vm_object_pip_wakeup(fs->first_object);
141 		fs->first_m = NULL;
142 	}
143 	if (dealloc) {
144 		vm_object_deallocate(fs->first_object);
145 	}
146 	unlock_map(fs);
147 	if (fs->vp != NULL) {
148 		vput(fs->vp);
149 		fs->vp = NULL;
150 	}
151 }
152 
153 #define unlock_things(fs) _unlock_things(fs, 0)
154 #define unlock_and_deallocate(fs) _unlock_things(fs, 1)
155 
156 /*
157  * TRYPAGER - used by vm_fault to calculate whether the pager for the
158  *	      current object *might* contain the page.
159  *
160  *	      default objects are zero-fill, there is no real pager.
161  */
162 
163 #define TRYPAGER	(fs.object->type != OBJT_DEFAULT && \
164 			(((fault_flags & VM_FAULT_WIRE_MASK) == 0) || wired))
165 
166 /*
167  *	vm_fault:
168  *
169  *	Handle a page fault occurring at the given address,
170  *	requiring the given permissions, in the map specified.
171  *	If successful, the page is inserted into the
172  *	associated physical map.
173  *
174  *	NOTE: the given address should be truncated to the
175  *	proper page address.
176  *
177  *	KERN_SUCCESS is returned if the page fault is handled; otherwise,
178  *	a standard error specifying why the fault is fatal is returned.
179  *
180  *
181  *	The map in question must be referenced, and remains so.
182  *	Caller may hold no locks.
183  */
184 int
185 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, int fault_flags)
186 {
187 	vm_prot_t prot;
188 	int result;
189 	boolean_t wired;
190 	int map_generation;
191 	vm_object_t next_object;
192 	vm_page_t marray[VM_FAULT_READ];
193 	int hardfault;
194 	int faultcount;
195 	struct faultstate fs;
196 
197 	cnt.v_vm_faults++;	/* needs lock XXX */
198 	hardfault = 0;
199 
200 RetryFault:;
201 
202 	/*
203 	 * Find the backing store object and offset into it to begin the
204 	 * search.
205 	 */
206 	fs.map = map;
207 	if ((result = vm_map_lookup(&fs.map, vaddr,
208 		fault_type, &fs.entry, &fs.first_object,
209 		&fs.first_pindex, &prot, &wired)) != KERN_SUCCESS) {
210 		if ((result != KERN_PROTECTION_FAILURE) ||
211 			((fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE)) {
212 			return result;
213 		}
214 
215 		/*
216    		 * If we are user-wiring a r/w segment, and it is COW, then
217    		 * we need to do the COW operation.  Note that we don't COW
218    		 * currently RO sections now, because it is NOT desirable
219    		 * to COW .text.  We simply keep .text from ever being COW'ed
220    		 * and take the heat that one cannot debug wired .text sections.
221    		 */
222 		result = vm_map_lookup(&fs.map, vaddr,
223 			VM_PROT_READ|VM_PROT_WRITE|VM_PROT_OVERRIDE_WRITE,
224 			&fs.entry, &fs.first_object, &fs.first_pindex, &prot, &wired);
225 		if (result != KERN_SUCCESS) {
226 			return result;
227 		}
228 
229 		/*
230 		 * If we don't COW now, on a user wire, the user will never
231 		 * be able to write to the mapping.  If we don't make this
232 		 * restriction, the bookkeeping would be nearly impossible.
233 		 */
234 		if ((fs.entry->protection & VM_PROT_WRITE) == 0)
235 			fs.entry->max_protection &= ~VM_PROT_WRITE;
236 	}
237 
238 	map_generation = fs.map->timestamp;
239 
240 	if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
241 		panic("vm_fault: fault on nofault entry, addr: %lx",
242 		    (u_long)vaddr);
243 	}
244 
245 	/*
246 	 * Make a reference to this object to prevent its disposal while we
247 	 * are messing with it.  Once we have the reference, the map is free
248 	 * to be diddled.  Since objects reference their shadows (and copies),
249 	 * they will stay around as well.
250 	 */
251 	vm_object_reference(fs.first_object);
252 	vm_object_pip_add(fs.first_object, 1);
253 
254 	fs.vp = vnode_pager_lock(fs.first_object);
255 	if ((fault_type & VM_PROT_WRITE) &&
256 		(fs.first_object->type == OBJT_VNODE)) {
257 		vm_freeze_copyopts(fs.first_object,
258 			fs.first_pindex, fs.first_pindex + 1);
259 	}
260 
261 	fs.lookup_still_valid = TRUE;
262 
263 	if (wired)
264 		fault_type = prot;
265 
266 	fs.first_m = NULL;
267 
268 	/*
269 	 * Search for the page at object/offset.
270 	 */
271 
272 	fs.object = fs.first_object;
273 	fs.pindex = fs.first_pindex;
274 
275 	while (TRUE) {
276 		/*
277 		 * If the object is dead, we stop here
278 		 */
279 
280 		if (fs.object->flags & OBJ_DEAD) {
281 			unlock_and_deallocate(&fs);
282 			return (KERN_PROTECTION_FAILURE);
283 		}
284 
285 		/*
286 		 * See if page is resident
287 		 */
288 
289 		fs.m = vm_page_lookup(fs.object, fs.pindex);
290 		if (fs.m != NULL) {
291 			int queue, s;
292 			/*
293 			 * Wait/Retry if the page is busy.  We have to do this
294 			 * if the page is busy via either PG_BUSY or
295 			 * vm_page_t->busy because the vm_pager may be using
296 			 * vm_page_t->busy for pageouts ( and even pageins if
297 			 * it is the vnode pager ), and we could end up trying
298 			 * to pagein and pageout the same page simultaneously.
299 			 *
300 			 * We can theoretically allow the busy case on a read
301 			 * fault if the page is marked valid, but since such
302 			 * pages are typically already pmap'd, putting that
303 			 * special case in might be more effort then it is
304 			 * worth.  We cannot under any circumstances mess
305 			 * around with a vm_page_t->busy page except, perhaps,
306 			 * to pmap it.
307 			 */
308 			if ((fs.m->flags & PG_BUSY) || fs.m->busy) {
309 				unlock_things(&fs);
310 				(void)vm_page_sleep_busy(fs.m, TRUE, "vmpfw");
311 				cnt.v_intrans++;
312 				vm_object_deallocate(fs.first_object);
313 				goto RetryFault;
314 			}
315 
316 			queue = fs.m->queue;
317 			s = splvm();
318 			vm_page_unqueue_nowakeup(fs.m);
319 			splx(s);
320 
321 			if ((queue - fs.m->pc) == PQ_CACHE && vm_page_count_severe()) {
322 				vm_page_activate(fs.m);
323 				unlock_and_deallocate(&fs);
324 				VM_WAIT;
325 				goto RetryFault;
326 			}
327 
328 			/*
329 			 * Mark page busy for other processes, and the
330 			 * pagedaemon.  If it still isn't completely valid
331 			 * (readable), jump to readrest, else break-out ( we
332 			 * found the page ).
333 			 */
334 
335 			vm_page_busy(fs.m);
336 			if (((fs.m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) &&
337 				fs.m->object != kernel_object && fs.m->object != kmem_object) {
338 				goto readrest;
339 			}
340 
341 			break;
342 		}
343 
344 		/*
345 		 * Page is not resident, If this is the search termination
346 		 * or the pager might contain the page, allocate a new page.
347 		 */
348 
349 		if (TRYPAGER || fs.object == fs.first_object) {
350 			if (fs.pindex >= fs.object->size) {
351 				unlock_and_deallocate(&fs);
352 				return (KERN_PROTECTION_FAILURE);
353 			}
354 
355 			/*
356 			 * Allocate a new page for this object/offset pair.
357 			 */
358 			fs.m = NULL;
359 			if (!vm_page_count_severe()) {
360 				fs.m = vm_page_alloc(fs.object, fs.pindex,
361 				    (fs.vp || fs.object->backing_object)? VM_ALLOC_NORMAL: VM_ALLOC_ZERO);
362 			}
363 			if (fs.m == NULL) {
364 				unlock_and_deallocate(&fs);
365 				VM_WAIT;
366 				goto RetryFault;
367 			}
368 		}
369 
370 readrest:
371 		/*
372 		 * We have found a valid page or we have allocated a new page.
373 		 * The page thus may not be valid or may not be entirely
374 		 * valid.
375 		 *
376 		 * Attempt to fault-in the page if there is a chance that the
377 		 * pager has it, and potentially fault in additional pages
378 		 * at the same time.
379 		 */
380 
381 		if (TRYPAGER) {
382 			int rv;
383 			int reqpage;
384 			int ahead, behind;
385 			u_char behavior = vm_map_entry_behavior(fs.entry);
386 
387 			if (behavior == MAP_ENTRY_BEHAV_RANDOM) {
388 				ahead = 0;
389 				behind = 0;
390 			} else {
391 				behind = (vaddr - fs.entry->start) >> PAGE_SHIFT;
392 				if (behind > VM_FAULT_READ_BEHIND)
393 					behind = VM_FAULT_READ_BEHIND;
394 
395 				ahead = ((fs.entry->end - vaddr) >> PAGE_SHIFT) - 1;
396 				if (ahead > VM_FAULT_READ_AHEAD)
397 					ahead = VM_FAULT_READ_AHEAD;
398 			}
399 
400 			if ((fs.first_object->type != OBJT_DEVICE) &&
401 			    (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL ||
402                                 (behavior != MAP_ENTRY_BEHAV_RANDOM &&
403                                 fs.pindex >= fs.entry->lastr &&
404                                 fs.pindex < fs.entry->lastr + VM_FAULT_READ))
405 			) {
406 				vm_pindex_t firstpindex, tmppindex;
407 
408 				if (fs.first_pindex < 2 * VM_FAULT_READ)
409 					firstpindex = 0;
410 				else
411 					firstpindex = fs.first_pindex - 2 * VM_FAULT_READ;
412 
413 				/*
414 				 * note: partially valid pages cannot be
415 				 * included in the lookahead - NFS piecemeal
416 				 * writes will barf on it badly.
417 				 */
418 
419 				for(tmppindex = fs.first_pindex - 1;
420 					tmppindex >= firstpindex;
421 					--tmppindex) {
422 					vm_page_t mt;
423 					mt = vm_page_lookup( fs.first_object, tmppindex);
424 					if (mt == NULL || (mt->valid != VM_PAGE_BITS_ALL))
425 						break;
426 					if (mt->busy ||
427 						(mt->flags & (PG_BUSY | PG_FICTITIOUS | PG_UNMANAGED)) ||
428 						mt->hold_count ||
429 						mt->wire_count)
430 						continue;
431 					if (mt->dirty == 0)
432 						vm_page_test_dirty(mt);
433 					if (mt->dirty) {
434 						vm_page_protect(mt, VM_PROT_NONE);
435 						vm_page_deactivate(mt);
436 					} else {
437 						vm_page_cache(mt);
438 					}
439 				}
440 
441 				ahead += behind;
442 				behind = 0;
443 			}
444 
445 			/*
446 			 * now we find out if any other pages should be paged
447 			 * in at this time this routine checks to see if the
448 			 * pages surrounding this fault reside in the same
449 			 * object as the page for this fault.  If they do,
450 			 * then they are faulted in also into the object.  The
451 			 * array "marray" returned contains an array of
452 			 * vm_page_t structs where one of them is the
453 			 * vm_page_t passed to the routine.  The reqpage
454 			 * return value is the index into the marray for the
455 			 * vm_page_t passed to the routine.
456 			 *
457 			 * fs.m plus the additional pages are PG_BUSY'd.
458 			 */
459 			faultcount = vm_fault_additional_pages(
460 			    fs.m, behind, ahead, marray, &reqpage);
461 
462 			/*
463 			 * update lastr imperfectly (we do not know how much
464 			 * getpages will actually read), but good enough.
465 			 */
466 			fs.entry->lastr = fs.pindex + faultcount - behind;
467 
468 			/*
469 			 * Call the pager to retrieve the data, if any, after
470 			 * releasing the lock on the map.  We hold a ref on
471 			 * fs.object and the pages are PG_BUSY'd.
472 			 */
473 			unlock_map(&fs);
474 
475 			rv = faultcount ?
476 			    vm_pager_get_pages(fs.object, marray, faultcount,
477 				reqpage) : VM_PAGER_FAIL;
478 
479 			if (rv == VM_PAGER_OK) {
480 				/*
481 				 * Found the page. Leave it busy while we play
482 				 * with it.
483 				 */
484 
485 				/*
486 				 * Relookup in case pager changed page. Pager
487 				 * is responsible for disposition of old page
488 				 * if moved.
489 				 */
490 				fs.m = vm_page_lookup(fs.object, fs.pindex);
491 				if(!fs.m) {
492 					unlock_and_deallocate(&fs);
493 					goto RetryFault;
494 				}
495 
496 				hardfault++;
497 				break; /* break to PAGE HAS BEEN FOUND */
498 			}
499 			/*
500 			 * Remove the bogus page (which does not exist at this
501 			 * object/offset); before doing so, we must get back
502 			 * our object lock to preserve our invariant.
503 			 *
504 			 * Also wake up any other process that may want to bring
505 			 * in this page.
506 			 *
507 			 * If this is the top-level object, we must leave the
508 			 * busy page to prevent another process from rushing
509 			 * past us, and inserting the page in that object at
510 			 * the same time that we are.
511 			 */
512 
513 			if (rv == VM_PAGER_ERROR)
514 				printf("vm_fault: pager read error, pid %d (%s)\n",
515 				    curproc->p_pid, curproc->p_comm);
516 			/*
517 			 * Data outside the range of the pager or an I/O error
518 			 */
519 			/*
520 			 * XXX - the check for kernel_map is a kludge to work
521 			 * around having the machine panic on a kernel space
522 			 * fault w/ I/O error.
523 			 */
524 			if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) ||
525 				(rv == VM_PAGER_BAD)) {
526 				vm_page_free(fs.m);
527 				fs.m = NULL;
528 				unlock_and_deallocate(&fs);
529 				return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
530 			}
531 			if (fs.object != fs.first_object) {
532 				vm_page_free(fs.m);
533 				fs.m = NULL;
534 				/*
535 				 * XXX - we cannot just fall out at this
536 				 * point, m has been freed and is invalid!
537 				 */
538 			}
539 		}
540 
541 		/*
542 		 * We get here if the object has default pager (or unwiring)
543 		 * or the pager doesn't have the page.
544 		 */
545 		if (fs.object == fs.first_object)
546 			fs.first_m = fs.m;
547 
548 		/*
549 		 * Move on to the next object.  Lock the next object before
550 		 * unlocking the current one.
551 		 */
552 
553 		fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset);
554 		next_object = fs.object->backing_object;
555 		if (next_object == NULL) {
556 			/*
557 			 * If there's no object left, fill the page in the top
558 			 * object with zeros.
559 			 */
560 			if (fs.object != fs.first_object) {
561 				vm_object_pip_wakeup(fs.object);
562 
563 				fs.object = fs.first_object;
564 				fs.pindex = fs.first_pindex;
565 				fs.m = fs.first_m;
566 			}
567 			fs.first_m = NULL;
568 
569 			/*
570 			 * Zero the page if necessary and mark it valid.
571 			 */
572 			if ((fs.m->flags & PG_ZERO) == 0) {
573 				vm_page_zero_fill(fs.m);
574 			} else {
575 				cnt.v_ozfod++;
576 			}
577 			cnt.v_zfod++;
578 			fs.m->valid = VM_PAGE_BITS_ALL;
579 			break;	/* break to PAGE HAS BEEN FOUND */
580 		} else {
581 			if (fs.object != fs.first_object) {
582 				vm_object_pip_wakeup(fs.object);
583 			}
584 			KASSERT(fs.object != next_object, ("object loop %p", next_object));
585 			fs.object = next_object;
586 			vm_object_pip_add(fs.object, 1);
587 		}
588 	}
589 
590 	KASSERT((fs.m->flags & PG_BUSY) != 0,
591 	    ("vm_fault: not busy after main loop"));
592 
593 	/*
594 	 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
595 	 * is held.]
596 	 */
597 
598 	/*
599 	 * If the page is being written, but isn't already owned by the
600 	 * top-level object, we have to copy it into a new page owned by the
601 	 * top-level object.
602 	 */
603 
604 	if (fs.object != fs.first_object) {
605 		/*
606 		 * We only really need to copy if we want to write it.
607 		 */
608 
609 		if (fault_type & VM_PROT_WRITE) {
610 			/*
611 			 * This allows pages to be virtually copied from a
612 			 * backing_object into the first_object, where the
613 			 * backing object has no other refs to it, and cannot
614 			 * gain any more refs.  Instead of a bcopy, we just
615 			 * move the page from the backing object to the
616 			 * first object.  Note that we must mark the page
617 			 * dirty in the first object so that it will go out
618 			 * to swap when needed.
619 			 */
620 			if (map_generation == fs.map->timestamp &&
621 				/*
622 				 * Only one shadow object
623 				 */
624 				(fs.object->shadow_count == 1) &&
625 				/*
626 				 * No COW refs, except us
627 				 */
628 				(fs.object->ref_count == 1) &&
629 				/*
630 				 * No one else can look this object up
631 				 */
632 				(fs.object->handle == NULL) &&
633 				/*
634 				 * No other ways to look the object up
635 				 */
636 				((fs.object->type == OBJT_DEFAULT) ||
637 				 (fs.object->type == OBJT_SWAP)) &&
638 				/*
639 				 * We don't chase down the shadow chain
640 				 */
641 				(fs.object == fs.first_object->backing_object) &&
642 
643 				/*
644 				 * grab the lock if we need to
645 				 */
646 				(fs.lookup_still_valid ||
647 				 lockmgr(&fs.map->lock, LK_EXCLUSIVE|LK_NOWAIT, (void *)0, curproc) == 0)
648 			    ) {
649 
650 				fs.lookup_still_valid = 1;
651 				/*
652 				 * get rid of the unnecessary page
653 				 */
654 				vm_page_protect(fs.first_m, VM_PROT_NONE);
655 				vm_page_free(fs.first_m);
656 				fs.first_m = NULL;
657 
658 				/*
659 				 * grab the page and put it into the
660 				 * process'es object.  The page is
661 				 * automatically made dirty.
662 				 */
663 				vm_page_rename(fs.m, fs.first_object, fs.first_pindex);
664 				fs.first_m = fs.m;
665 				vm_page_busy(fs.first_m);
666 				fs.m = NULL;
667 				cnt.v_cow_optim++;
668 			} else {
669 				/*
670 				 * Oh, well, lets copy it.
671 				 */
672 				vm_page_copy(fs.m, fs.first_m);
673 			}
674 
675 			if (fs.m) {
676 				/*
677 				 * We no longer need the old page or object.
678 				 */
679 				release_page(&fs);
680 			}
681 
682 			/*
683 			 * fs.object != fs.first_object due to above
684 			 * conditional
685 			 */
686 
687 			vm_object_pip_wakeup(fs.object);
688 
689 			/*
690 			 * Only use the new page below...
691 			 */
692 
693 			cnt.v_cow_faults++;
694 			fs.m = fs.first_m;
695 			fs.object = fs.first_object;
696 			fs.pindex = fs.first_pindex;
697 
698 		} else {
699 			prot &= ~VM_PROT_WRITE;
700 		}
701 	}
702 
703 	/*
704 	 * We must verify that the maps have not changed since our last
705 	 * lookup.
706 	 */
707 
708 	if (!fs.lookup_still_valid &&
709 		(fs.map->timestamp != map_generation)) {
710 		vm_object_t retry_object;
711 		vm_pindex_t retry_pindex;
712 		vm_prot_t retry_prot;
713 
714 		/*
715 		 * Since map entries may be pageable, make sure we can take a
716 		 * page fault on them.
717 		 */
718 
719 		/*
720 		 * Unlock vnode before the lookup to avoid deadlock.   E.G.
721 		 * avoid a deadlock between the inode and exec_map that can
722 		 * occur due to locks being obtained in different orders.
723 		 */
724 
725 		if (fs.vp != NULL) {
726 			vput(fs.vp);
727 			fs.vp = NULL;
728 		}
729 
730 		if (fs.map->infork) {
731 			release_page(&fs);
732 			unlock_and_deallocate(&fs);
733 			goto RetryFault;
734 		}
735 
736 		/*
737 		 * To avoid trying to write_lock the map while another process
738 		 * has it read_locked (in vm_map_pageable), we do not try for
739 		 * write permission.  If the page is still writable, we will
740 		 * get write permission.  If it is not, or has been marked
741 		 * needs_copy, we enter the mapping without write permission,
742 		 * and will merely take another fault.
743 		 */
744 		result = vm_map_lookup(&fs.map, vaddr, fault_type & ~VM_PROT_WRITE,
745 		    &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
746 		map_generation = fs.map->timestamp;
747 
748 		/*
749 		 * If we don't need the page any longer, put it on the active
750 		 * list (the easiest thing to do here).  If no one needs it,
751 		 * pageout will grab it eventually.
752 		 */
753 
754 		if (result != KERN_SUCCESS) {
755 			release_page(&fs);
756 			unlock_and_deallocate(&fs);
757 			return (result);
758 		}
759 		fs.lookup_still_valid = TRUE;
760 
761 		if ((retry_object != fs.first_object) ||
762 		    (retry_pindex != fs.first_pindex)) {
763 			release_page(&fs);
764 			unlock_and_deallocate(&fs);
765 			goto RetryFault;
766 		}
767 		/*
768 		 * Check whether the protection has changed or the object has
769 		 * been copied while we left the map unlocked. Changing from
770 		 * read to write permission is OK - we leave the page
771 		 * write-protected, and catch the write fault. Changing from
772 		 * write to read permission means that we can't mark the page
773 		 * write-enabled after all.
774 		 */
775 		prot &= retry_prot;
776 	}
777 
778 	/*
779 	 * Put this page into the physical map. We had to do the unlock above
780 	 * because pmap_enter may cause other faults.   We don't put the page
781 	 * back on the active queue until later so that the page-out daemon
782 	 * won't find us (yet).
783 	 */
784 
785 	if (prot & VM_PROT_WRITE) {
786 		vm_page_flag_set(fs.m, PG_WRITEABLE);
787 		vm_object_set_flag(fs.m->object,
788 				   OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
789 
790 		/*
791 		 * If the fault is a write, we know that this page is being
792 		 * written NOW so dirty it explicitly to save on
793 		 * pmap_is_modified() calls later.
794 		 *
795 		 * If this is a NOSYNC mmap we do not want to set PG_NOSYNC
796 		 * if the page is already dirty to prevent data written with
797 		 * the expectation of being synced from not being synced.
798 		 * Likewise if this entry does not request NOSYNC then make
799 		 * sure the page isn't marked NOSYNC.  Applications sharing
800 		 * data should use the same flags to avoid ping ponging.
801 		 *
802 		 * Also tell the backing pager, if any, that it should remove
803 		 * any swap backing since the page is now dirty.
804 		 */
805 		if (fs.entry->eflags & MAP_ENTRY_NOSYNC) {
806 			if (fs.m->dirty == 0)
807 				vm_page_flag_set(fs.m, PG_NOSYNC);
808 		} else {
809 			vm_page_flag_clear(fs.m, PG_NOSYNC);
810 		}
811 		if (fault_flags & VM_FAULT_DIRTY) {
812 			int s;
813 			vm_page_dirty(fs.m);
814 			s = splvm();
815 			vm_pager_page_unswapped(fs.m);
816 			splx(s);
817 		}
818 	}
819 
820 	/*
821 	 * Page had better still be busy
822 	 */
823 
824 	KASSERT(fs.m->flags & PG_BUSY,
825 		("vm_fault: page %p not busy!", fs.m));
826 
827 	unlock_things(&fs);
828 
829 	/*
830 	 * Sanity check: page must be completely valid or it is not fit to
831 	 * map into user space.  vm_pager_get_pages() ensures this.
832 	 */
833 
834 	if (fs.m->valid != VM_PAGE_BITS_ALL) {
835 		vm_page_zero_invalid(fs.m, TRUE);
836 		printf("Warning: page %p partially invalid on fault\n", fs.m);
837 	}
838 
839 	pmap_enter(fs.map->pmap, vaddr, fs.m, prot, wired);
840 
841 	if (((fault_flags & VM_FAULT_WIRE_MASK) == 0) && (wired == 0)) {
842 		pmap_prefault(fs.map->pmap, vaddr, fs.entry);
843 	}
844 
845 	vm_page_flag_clear(fs.m, PG_ZERO);
846 	vm_page_flag_set(fs.m, PG_MAPPED|PG_REFERENCED);
847 	if (fault_flags & VM_FAULT_HOLD)
848 		vm_page_hold(fs.m);
849 
850 	/*
851 	 * If the page is not wired down, then put it where the pageout daemon
852 	 * can find it.
853 	 */
854 
855 	if (fault_flags & VM_FAULT_WIRE_MASK) {
856 		if (wired)
857 			vm_page_wire(fs.m);
858 		else
859 			vm_page_unwire(fs.m, 1);
860 	} else {
861 		vm_page_activate(fs.m);
862 	}
863 
864 	mtx_lock_spin(&sched_lock);
865 	if (curproc && (curproc->p_sflag & PS_INMEM) && curproc->p_stats) {
866 		if (hardfault) {
867 			curproc->p_stats->p_ru.ru_majflt++;
868 		} else {
869 			curproc->p_stats->p_ru.ru_minflt++;
870 		}
871 	}
872 	mtx_unlock_spin(&sched_lock);
873 
874 	/*
875 	 * Unlock everything, and return
876 	 */
877 
878 	vm_page_wakeup(fs.m);
879 	vm_object_deallocate(fs.first_object);
880 
881 	return (KERN_SUCCESS);
882 
883 }
884 
885 /*
886  *	vm_fault_wire:
887  *
888  *	Wire down a range of virtual addresses in a map.
889  */
890 int
891 vm_fault_wire(map, start, end)
892 	vm_map_t map;
893 	vm_offset_t start, end;
894 {
895 
896 	register vm_offset_t va;
897 	register pmap_t pmap;
898 	int rv;
899 
900 	pmap = vm_map_pmap(map);
901 
902 	/*
903 	 * Inform the physical mapping system that the range of addresses may
904 	 * not fault, so that page tables and such can be locked down as well.
905 	 */
906 
907 	pmap_pageable(pmap, start, end, FALSE);
908 
909 	/*
910 	 * We simulate a fault to get the page and enter it in the physical
911 	 * map.
912 	 */
913 
914 	for (va = start; va < end; va += PAGE_SIZE) {
915 		rv = vm_fault(map, va, VM_PROT_READ|VM_PROT_WRITE,
916 			VM_FAULT_CHANGE_WIRING);
917 		if (rv) {
918 			if (va != start)
919 				vm_fault_unwire(map, start, va);
920 			return (rv);
921 		}
922 	}
923 	return (KERN_SUCCESS);
924 }
925 
926 /*
927  *	vm_fault_user_wire:
928  *
929  *	Wire down a range of virtual addresses in a map.  This
930  *	is for user mode though, so we only ask for read access
931  *	on currently read only sections.
932  */
933 int
934 vm_fault_user_wire(map, start, end)
935 	vm_map_t map;
936 	vm_offset_t start, end;
937 {
938 
939 	register vm_offset_t va;
940 	register pmap_t pmap;
941 	int rv;
942 
943 	pmap = vm_map_pmap(map);
944 
945 	/*
946 	 * Inform the physical mapping system that the range of addresses may
947 	 * not fault, so that page tables and such can be locked down as well.
948 	 */
949 
950 	pmap_pageable(pmap, start, end, FALSE);
951 
952 	/*
953 	 * We simulate a fault to get the page and enter it in the physical
954 	 * map.
955 	 */
956 	for (va = start; va < end; va += PAGE_SIZE) {
957 		rv = vm_fault(map, va, VM_PROT_READ, VM_FAULT_USER_WIRE);
958 		if (rv) {
959 			if (va != start)
960 				vm_fault_unwire(map, start, va);
961 			return (rv);
962 		}
963 	}
964 	return (KERN_SUCCESS);
965 }
966 
967 
968 /*
969  *	vm_fault_unwire:
970  *
971  *	Unwire a range of virtual addresses in a map.
972  */
973 void
974 vm_fault_unwire(map, start, end)
975 	vm_map_t map;
976 	vm_offset_t start, end;
977 {
978 
979 	register vm_offset_t va, pa;
980 	register pmap_t pmap;
981 
982 	pmap = vm_map_pmap(map);
983 
984 	/*
985 	 * Since the pages are wired down, we must be able to get their
986 	 * mappings from the physical map system.
987 	 */
988 
989 	for (va = start; va < end; va += PAGE_SIZE) {
990 		pa = pmap_extract(pmap, va);
991 		if (pa != (vm_offset_t) 0) {
992 			pmap_change_wiring(pmap, va, FALSE);
993 			vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1);
994 		}
995 	}
996 
997 	/*
998 	 * Inform the physical mapping system that the range of addresses may
999 	 * fault, so that page tables and such may be unwired themselves.
1000 	 */
1001 
1002 	pmap_pageable(pmap, start, end, TRUE);
1003 
1004 }
1005 
1006 /*
1007  *	Routine:
1008  *		vm_fault_copy_entry
1009  *	Function:
1010  *		Copy all of the pages from a wired-down map entry to another.
1011  *
1012  *	In/out conditions:
1013  *		The source and destination maps must be locked for write.
1014  *		The source map entry must be wired down (or be a sharing map
1015  *		entry corresponding to a main map entry that is wired down).
1016  */
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 
1089 		vm_page_flag_clear(dst_m, PG_ZERO);
1090 		pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE);
1091 		vm_page_flag_set(dst_m, PG_WRITEABLE|PG_MAPPED);
1092 
1093 		/*
1094 		 * Mark it no longer busy, and put it on the active list.
1095 		 */
1096 		vm_page_activate(dst_m);
1097 		vm_page_wakeup(dst_m);
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 static int
1117 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage)
1118 	vm_page_t m;
1119 	int rbehind;
1120 	int rahead;
1121 	vm_page_t *marray;
1122 	int *reqpage;
1123 {
1124 	int i,j;
1125 	vm_object_t object;
1126 	vm_pindex_t pindex, startpindex, endpindex, tpindex;
1127 	vm_page_t rtm;
1128 	int cbehind, cahead;
1129 
1130 	object = m->object;
1131 	pindex = m->pindex;
1132 
1133 	/*
1134 	 * we don't fault-ahead for device pager
1135 	 */
1136 	if (object->type == OBJT_DEVICE) {
1137 		*reqpage = 0;
1138 		marray[0] = m;
1139 		return 1;
1140 	}
1141 
1142 	/*
1143 	 * if the requested page is not available, then give up now
1144 	 */
1145 
1146 	if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1147 		return 0;
1148 	}
1149 
1150 	if ((cbehind == 0) && (cahead == 0)) {
1151 		*reqpage = 0;
1152 		marray[0] = m;
1153 		return 1;
1154 	}
1155 
1156 	if (rahead > cahead) {
1157 		rahead = cahead;
1158 	}
1159 
1160 	if (rbehind > cbehind) {
1161 		rbehind = cbehind;
1162 	}
1163 
1164 	/*
1165 	 * try to do any readahead that we might have free pages for.
1166 	 */
1167 	if ((rahead + rbehind) >
1168 		((cnt.v_free_count + cnt.v_cache_count) - cnt.v_free_reserved)) {
1169 		pagedaemon_wakeup();
1170 		marray[0] = m;
1171 		*reqpage = 0;
1172 		return 1;
1173 	}
1174 
1175 	/*
1176 	 * scan backward for the read behind pages -- in memory
1177 	 */
1178 	if (pindex > 0) {
1179 		if (rbehind > pindex) {
1180 			rbehind = pindex;
1181 			startpindex = 0;
1182 		} else {
1183 			startpindex = pindex - rbehind;
1184 		}
1185 
1186 		for ( tpindex = pindex - 1; tpindex >= startpindex; tpindex -= 1) {
1187 			if (vm_page_lookup( object, tpindex)) {
1188 				startpindex = tpindex + 1;
1189 				break;
1190 			}
1191 			if (tpindex == 0)
1192 				break;
1193 		}
1194 
1195 		for(i = 0, tpindex = startpindex; tpindex < pindex; i++, tpindex++) {
1196 
1197 			rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
1198 			if (rtm == NULL) {
1199 				for (j = 0; j < i; j++) {
1200 					vm_page_free(marray[j]);
1201 				}
1202 				marray[0] = m;
1203 				*reqpage = 0;
1204 				return 1;
1205 			}
1206 
1207 			marray[i] = rtm;
1208 		}
1209 	} else {
1210 		startpindex = 0;
1211 		i = 0;
1212 	}
1213 
1214 	marray[i] = m;
1215 	/* page offset of the required page */
1216 	*reqpage = i;
1217 
1218 	tpindex = pindex + 1;
1219 	i++;
1220 
1221 	/*
1222 	 * scan forward for the read ahead pages
1223 	 */
1224 	endpindex = tpindex + rahead;
1225 	if (endpindex > object->size)
1226 		endpindex = object->size;
1227 
1228 	for( ; tpindex < endpindex; i++, tpindex++) {
1229 
1230 		if (vm_page_lookup(object, tpindex)) {
1231 			break;
1232 		}
1233 
1234 		rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
1235 		if (rtm == NULL) {
1236 			break;
1237 		}
1238 
1239 		marray[i] = rtm;
1240 	}
1241 
1242 	/* return number of bytes of pages */
1243 	return i;
1244 }
1245