xref: /freebsd/sys/vm/vm_fault.c (revision b601c69bdbe8755d26570261d7fd4c02ee4eff74)
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/proc.h>
79 #include <sys/vnode.h>
80 #include <sys/resourcevar.h>
81 #include <sys/vmmeter.h>
82 
83 #include <vm/vm.h>
84 #include <vm/vm_param.h>
85 #include <sys/lock.h>
86 #include <vm/pmap.h>
87 #include <vm/vm_map.h>
88 #include <vm/vm_object.h>
89 #include <vm/vm_page.h>
90 #include <vm/vm_pageout.h>
91 #include <vm/vm_kern.h>
92 #include <vm/vm_pager.h>
93 #include <vm/vnode_pager.h>
94 #include <vm/vm_extern.h>
95 
96 static int vm_fault_additional_pages __P((vm_page_t, int,
97 					  int, vm_page_t *, int *));
98 
99 #define VM_FAULT_READ_AHEAD 8
100 #define VM_FAULT_READ_BEHIND 7
101 #define VM_FAULT_READ (VM_FAULT_READ_AHEAD+VM_FAULT_READ_BEHIND+1)
102 
103 struct faultstate {
104 	vm_page_t m;
105 	vm_object_t object;
106 	vm_pindex_t pindex;
107 	vm_page_t first_m;
108 	vm_object_t	first_object;
109 	vm_pindex_t first_pindex;
110 	vm_map_t map;
111 	vm_map_entry_t entry;
112 	int lookup_still_valid;
113 	struct vnode *vp;
114 };
115 
116 static __inline void
117 release_page(struct faultstate *fs)
118 {
119 	vm_page_wakeup(fs->m);
120 	vm_page_deactivate(fs->m);
121 	fs->m = NULL;
122 }
123 
124 static __inline void
125 unlock_map(struct faultstate *fs)
126 {
127 	if (fs->lookup_still_valid) {
128 		vm_map_lookup_done(fs->map, fs->entry);
129 		fs->lookup_still_valid = FALSE;
130 	}
131 }
132 
133 static void
134 _unlock_things(struct faultstate *fs, int dealloc)
135 {
136 	vm_object_pip_wakeup(fs->object);
137 	if (fs->object != fs->first_object) {
138 		vm_page_free(fs->first_m);
139 		vm_object_pip_wakeup(fs->first_object);
140 		fs->first_m = NULL;
141 	}
142 	if (dealloc) {
143 		vm_object_deallocate(fs->first_object);
144 	}
145 	unlock_map(fs);
146 	if (fs->vp != NULL) {
147 		vput(fs->vp);
148 		fs->vp = NULL;
149 	}
150 }
151 
152 #define unlock_things(fs) _unlock_things(fs, 0)
153 #define unlock_and_deallocate(fs) _unlock_things(fs, 1)
154 
155 /*
156  * TRYPAGER - used by vm_fault to calculate whether the pager for the
157  *	      current object *might* contain the page.
158  *
159  *	      default objects are zero-fill, there is no real pager.
160  */
161 
162 #define TRYPAGER	(fs.object->type != OBJT_DEFAULT && \
163 			(((fault_flags & VM_FAULT_WIRE_MASK) == 0) || wired))
164 
165 /*
166  *	vm_fault:
167  *
168  *	Handle a page fault occurring at the given address,
169  *	requiring the given permissions, in the map specified.
170  *	If successful, the page is inserted into the
171  *	associated physical map.
172  *
173  *	NOTE: the given address should be truncated to the
174  *	proper page address.
175  *
176  *	KERN_SUCCESS is returned if the page fault is handled; otherwise,
177  *	a standard error specifying why the fault is fatal is returned.
178  *
179  *
180  *	The map in question must be referenced, and remains so.
181  *	Caller may hold no locks.
182  */
183 int
184 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, int fault_flags)
185 {
186 	vm_prot_t prot;
187 	int result;
188 	boolean_t wired;
189 	int map_generation;
190 	vm_object_t next_object;
191 	vm_page_t marray[VM_FAULT_READ];
192 	int hardfault;
193 	int faultcount;
194 	struct faultstate fs;
195 
196 	cnt.v_vm_faults++;	/* needs lock XXX */
197 	hardfault = 0;
198 
199 RetryFault:;
200 
201 	/*
202 	 * Find the backing store object and offset into it to begin the
203 	 * search.
204 	 */
205 	fs.map = map;
206 	if ((result = vm_map_lookup(&fs.map, vaddr,
207 		fault_type, &fs.entry, &fs.first_object,
208 		&fs.first_pindex, &prot, &wired)) != KERN_SUCCESS) {
209 		if ((result != KERN_PROTECTION_FAILURE) ||
210 			((fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE)) {
211 			return result;
212 		}
213 
214 		/*
215    		 * If we are user-wiring a r/w segment, and it is COW, then
216    		 * we need to do the COW operation.  Note that we don't COW
217    		 * currently RO sections now, because it is NOT desirable
218    		 * to COW .text.  We simply keep .text from ever being COW'ed
219    		 * and take the heat that one cannot debug wired .text sections.
220    		 */
221 		result = vm_map_lookup(&fs.map, vaddr,
222 			VM_PROT_READ|VM_PROT_WRITE|VM_PROT_OVERRIDE_WRITE,
223 			&fs.entry, &fs.first_object, &fs.first_pindex, &prot, &wired);
224 		if (result != KERN_SUCCESS) {
225 			return result;
226 		}
227 
228 		/*
229 		 * If we don't COW now, on a user wire, the user will never
230 		 * be able to write to the mapping.  If we don't make this
231 		 * restriction, the bookkeeping would be nearly impossible.
232 		 */
233 		if ((fs.entry->protection & VM_PROT_WRITE) == 0)
234 			fs.entry->max_protection &= ~VM_PROT_WRITE;
235 	}
236 
237 	map_generation = fs.map->timestamp;
238 
239 	if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
240 		panic("vm_fault: fault on nofault entry, addr: %lx",
241 		    (u_long)vaddr);
242 	}
243 
244 	/*
245 	 * Make a reference to this object to prevent its disposal while we
246 	 * are messing with it.  Once we have the reference, the map is free
247 	 * to be diddled.  Since objects reference their shadows (and copies),
248 	 * they will stay around as well.
249 	 */
250 	vm_object_reference(fs.first_object);
251 	vm_object_pip_add(fs.first_object, 1);
252 
253 	fs.vp = vnode_pager_lock(fs.first_object);
254 	if ((fault_type & VM_PROT_WRITE) &&
255 		(fs.first_object->type == OBJT_VNODE)) {
256 		vm_freeze_copyopts(fs.first_object,
257 			fs.first_pindex, fs.first_pindex + 1);
258 	}
259 
260 	fs.lookup_still_valid = TRUE;
261 
262 	if (wired)
263 		fault_type = prot;
264 
265 	fs.first_m = NULL;
266 
267 	/*
268 	 * Search for the page at object/offset.
269 	 */
270 
271 	fs.object = fs.first_object;
272 	fs.pindex = fs.first_pindex;
273 
274 	while (TRUE) {
275 		/*
276 		 * If the object is dead, we stop here
277 		 */
278 
279 		if (fs.object->flags & OBJ_DEAD) {
280 			unlock_and_deallocate(&fs);
281 			return (KERN_PROTECTION_FAILURE);
282 		}
283 
284 		/*
285 		 * See if page is resident
286 		 */
287 
288 		fs.m = vm_page_lookup(fs.object, fs.pindex);
289 		if (fs.m != NULL) {
290 			int queue, s;
291 			/*
292 			 * Wait/Retry if the page is busy.  We have to do this
293 			 * if the page is busy via either PG_BUSY or
294 			 * vm_page_t->busy because the vm_pager may be using
295 			 * vm_page_t->busy for pageouts ( and even pageins if
296 			 * it is the vnode pager ), and we could end up trying
297 			 * to pagein and pageout the same page simultaneously.
298 			 *
299 			 * We can theoretically allow the busy case on a read
300 			 * fault if the page is marked valid, but since such
301 			 * pages are typically already pmap'd, putting that
302 			 * special case in might be more effort then it is
303 			 * worth.  We cannot under any circumstances mess
304 			 * around with a vm_page_t->busy page except, perhaps,
305 			 * to pmap it.
306 			 */
307 			if ((fs.m->flags & PG_BUSY) || fs.m->busy) {
308 				unlock_things(&fs);
309 				(void)vm_page_sleep_busy(fs.m, TRUE, "vmpfw");
310 				cnt.v_intrans++;
311 				vm_object_deallocate(fs.first_object);
312 				goto RetryFault;
313 			}
314 
315 			queue = fs.m->queue;
316 			s = splvm();
317 			vm_page_unqueue_nowakeup(fs.m);
318 			splx(s);
319 
320 			if ((queue - fs.m->pc) == PQ_CACHE && vm_page_count_severe()) {
321 				vm_page_activate(fs.m);
322 				unlock_and_deallocate(&fs);
323 				VM_WAIT;
324 				goto RetryFault;
325 			}
326 
327 			/*
328 			 * Mark page busy for other processes, and the
329 			 * pagedaemon.  If it still isn't completely valid
330 			 * (readable), jump to readrest, else break-out ( we
331 			 * found the page ).
332 			 */
333 
334 			vm_page_busy(fs.m);
335 			if (((fs.m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) &&
336 				fs.m->object != kernel_object && fs.m->object != kmem_object) {
337 				goto readrest;
338 			}
339 
340 			break;
341 		}
342 
343 		/*
344 		 * Page is not resident, If this is the search termination
345 		 * or the pager might contain the page, allocate a new page.
346 		 */
347 
348 		if (TRYPAGER || fs.object == fs.first_object) {
349 			if (fs.pindex >= fs.object->size) {
350 				unlock_and_deallocate(&fs);
351 				return (KERN_PROTECTION_FAILURE);
352 			}
353 
354 			/*
355 			 * Allocate a new page for this object/offset pair.
356 			 */
357 			fs.m = NULL;
358 			if (!vm_page_count_severe()) {
359 				fs.m = vm_page_alloc(fs.object, fs.pindex,
360 				    (fs.vp || fs.object->backing_object)? VM_ALLOC_NORMAL: VM_ALLOC_ZERO);
361 			}
362 			if (fs.m == NULL) {
363 				unlock_and_deallocate(&fs);
364 				VM_WAIT;
365 				goto RetryFault;
366 			}
367 		}
368 
369 readrest:
370 		/*
371 		 * We have found a valid page or we have allocated a new page.
372 		 * The page thus may not be valid or may not be entirely
373 		 * valid.
374 		 *
375 		 * Attempt to fault-in the page if there is a chance that the
376 		 * pager has it, and potentially fault in additional pages
377 		 * at the same time.
378 		 */
379 
380 		if (TRYPAGER) {
381 			int rv;
382 			int reqpage;
383 			int ahead, behind;
384 			u_char behavior = vm_map_entry_behavior(fs.entry);
385 
386 			if (behavior == MAP_ENTRY_BEHAV_RANDOM) {
387 				ahead = 0;
388 				behind = 0;
389 			} else {
390 				behind = (vaddr - fs.entry->start) >> PAGE_SHIFT;
391 				if (behind > VM_FAULT_READ_BEHIND)
392 					behind = VM_FAULT_READ_BEHIND;
393 
394 				ahead = ((fs.entry->end - vaddr) >> PAGE_SHIFT) - 1;
395 				if (ahead > VM_FAULT_READ_AHEAD)
396 					ahead = VM_FAULT_READ_AHEAD;
397 			}
398 
399 			if ((fs.first_object->type != OBJT_DEVICE) &&
400 			    (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL ||
401                                 (behavior != MAP_ENTRY_BEHAV_RANDOM &&
402                                 fs.pindex >= fs.entry->lastr &&
403                                 fs.pindex < fs.entry->lastr + VM_FAULT_READ))
404 			) {
405 				vm_pindex_t firstpindex, tmppindex;
406 
407 				if (fs.first_pindex < 2 * VM_FAULT_READ)
408 					firstpindex = 0;
409 				else
410 					firstpindex = fs.first_pindex - 2 * VM_FAULT_READ;
411 
412 				/*
413 				 * note: partially valid pages cannot be
414 				 * included in the lookahead - NFS piecemeal
415 				 * writes will barf on it badly.
416 				 */
417 
418 				for(tmppindex = fs.first_pindex - 1;
419 					tmppindex >= firstpindex;
420 					--tmppindex) {
421 					vm_page_t mt;
422 					mt = vm_page_lookup( fs.first_object, tmppindex);
423 					if (mt == NULL || (mt->valid != VM_PAGE_BITS_ALL))
424 						break;
425 					if (mt->busy ||
426 						(mt->flags & (PG_BUSY | PG_FICTITIOUS | PG_UNMANAGED)) ||
427 						mt->hold_count ||
428 						mt->wire_count)
429 						continue;
430 					if (mt->dirty == 0)
431 						vm_page_test_dirty(mt);
432 					if (mt->dirty) {
433 						vm_page_protect(mt, VM_PROT_NONE);
434 						vm_page_deactivate(mt);
435 					} else {
436 						vm_page_cache(mt);
437 					}
438 				}
439 
440 				ahead += behind;
441 				behind = 0;
442 			}
443 
444 			/*
445 			 * now we find out if any other pages should be paged
446 			 * in at this time this routine checks to see if the
447 			 * pages surrounding this fault reside in the same
448 			 * object as the page for this fault.  If they do,
449 			 * then they are faulted in also into the object.  The
450 			 * array "marray" returned contains an array of
451 			 * vm_page_t structs where one of them is the
452 			 * vm_page_t passed to the routine.  The reqpage
453 			 * return value is the index into the marray for the
454 			 * vm_page_t passed to the routine.
455 			 *
456 			 * fs.m plus the additional pages are PG_BUSY'd.
457 			 */
458 			faultcount = vm_fault_additional_pages(
459 			    fs.m, behind, ahead, marray, &reqpage);
460 
461 			/*
462 			 * update lastr imperfectly (we do not know how much
463 			 * getpages will actually read), but good enough.
464 			 */
465 			fs.entry->lastr = fs.pindex + faultcount - behind;
466 
467 			/*
468 			 * Call the pager to retrieve the data, if any, after
469 			 * releasing the lock on the map.  We hold a ref on
470 			 * fs.object and the pages are PG_BUSY'd.
471 			 */
472 			unlock_map(&fs);
473 
474 			rv = faultcount ?
475 			    vm_pager_get_pages(fs.object, marray, faultcount,
476 				reqpage) : VM_PAGER_FAIL;
477 
478 			if (rv == VM_PAGER_OK) {
479 				/*
480 				 * Found the page. Leave it busy while we play
481 				 * with it.
482 				 */
483 
484 				/*
485 				 * Relookup in case pager changed page. Pager
486 				 * is responsible for disposition of old page
487 				 * if moved.
488 				 */
489 				fs.m = vm_page_lookup(fs.object, fs.pindex);
490 				if(!fs.m) {
491 					unlock_and_deallocate(&fs);
492 					goto RetryFault;
493 				}
494 
495 				hardfault++;
496 				break; /* break to PAGE HAS BEEN FOUND */
497 			}
498 			/*
499 			 * Remove the bogus page (which does not exist at this
500 			 * object/offset); before doing so, we must get back
501 			 * our object lock to preserve our invariant.
502 			 *
503 			 * Also wake up any other process that may want to bring
504 			 * in this page.
505 			 *
506 			 * If this is the top-level object, we must leave the
507 			 * busy page to prevent another process from rushing
508 			 * past us, and inserting the page in that object at
509 			 * the same time that we are.
510 			 */
511 
512 			if (rv == VM_PAGER_ERROR)
513 				printf("vm_fault: pager read error, pid %d (%s)\n",
514 				    curproc->p_pid, curproc->p_comm);
515 			/*
516 			 * Data outside the range of the pager or an I/O error
517 			 */
518 			/*
519 			 * XXX - the check for kernel_map is a kludge to work
520 			 * around having the machine panic on a kernel space
521 			 * fault w/ I/O error.
522 			 */
523 			if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) ||
524 				(rv == VM_PAGER_BAD)) {
525 				vm_page_free(fs.m);
526 				fs.m = NULL;
527 				unlock_and_deallocate(&fs);
528 				return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
529 			}
530 			if (fs.object != fs.first_object) {
531 				vm_page_free(fs.m);
532 				fs.m = NULL;
533 				/*
534 				 * XXX - we cannot just fall out at this
535 				 * point, m has been freed and is invalid!
536 				 */
537 			}
538 		}
539 
540 		/*
541 		 * We get here if the object has default pager (or unwiring)
542 		 * or the pager doesn't have the page.
543 		 */
544 		if (fs.object == fs.first_object)
545 			fs.first_m = fs.m;
546 
547 		/*
548 		 * Move on to the next object.  Lock the next object before
549 		 * unlocking the current one.
550 		 */
551 
552 		fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset);
553 		next_object = fs.object->backing_object;
554 		if (next_object == NULL) {
555 			/*
556 			 * If there's no object left, fill the page in the top
557 			 * object with zeros.
558 			 */
559 			if (fs.object != fs.first_object) {
560 				vm_object_pip_wakeup(fs.object);
561 
562 				fs.object = fs.first_object;
563 				fs.pindex = fs.first_pindex;
564 				fs.m = fs.first_m;
565 			}
566 			fs.first_m = NULL;
567 
568 			/*
569 			 * Zero the page if necessary and mark it valid.
570 			 */
571 			if ((fs.m->flags & PG_ZERO) == 0) {
572 				vm_page_zero_fill(fs.m);
573 			} else {
574 				cnt.v_ozfod++;
575 			}
576 			cnt.v_zfod++;
577 			fs.m->valid = VM_PAGE_BITS_ALL;
578 			break;	/* break to PAGE HAS BEEN FOUND */
579 		} else {
580 			if (fs.object != fs.first_object) {
581 				vm_object_pip_wakeup(fs.object);
582 			}
583 			KASSERT(fs.object != next_object, ("object loop %p", next_object));
584 			fs.object = next_object;
585 			vm_object_pip_add(fs.object, 1);
586 		}
587 	}
588 
589 	KASSERT((fs.m->flags & PG_BUSY) != 0,
590 	    ("vm_fault: not busy after main loop"));
591 
592 	/*
593 	 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
594 	 * is held.]
595 	 */
596 
597 	/*
598 	 * If the page is being written, but isn't already owned by the
599 	 * top-level object, we have to copy it into a new page owned by the
600 	 * top-level object.
601 	 */
602 
603 	if (fs.object != fs.first_object) {
604 		/*
605 		 * We only really need to copy if we want to write it.
606 		 */
607 
608 		if (fault_type & VM_PROT_WRITE) {
609 			/*
610 			 * This allows pages to be virtually copied from a
611 			 * backing_object into the first_object, where the
612 			 * backing object has no other refs to it, and cannot
613 			 * gain any more refs.  Instead of a bcopy, we just
614 			 * move the page from the backing object to the
615 			 * first object.  Note that we must mark the page
616 			 * dirty in the first object so that it will go out
617 			 * to swap when needed.
618 			 */
619 			if (map_generation == fs.map->timestamp &&
620 				/*
621 				 * Only one shadow object
622 				 */
623 				(fs.object->shadow_count == 1) &&
624 				/*
625 				 * No COW refs, except us
626 				 */
627 				(fs.object->ref_count == 1) &&
628 				/*
629 				 * No one else can look this object up
630 				 */
631 				(fs.object->handle == NULL) &&
632 				/*
633 				 * No other ways to look the object up
634 				 */
635 				((fs.object->type == OBJT_DEFAULT) ||
636 				 (fs.object->type == OBJT_SWAP)) &&
637 				/*
638 				 * We don't chase down the shadow chain
639 				 */
640 				(fs.object == fs.first_object->backing_object) &&
641 
642 				/*
643 				 * grab the lock if we need to
644 				 */
645 				(fs.lookup_still_valid ||
646 				 lockmgr(&fs.map->lock, LK_EXCLUSIVE|LK_NOWAIT, (void *)0, curproc) == 0)
647 			    ) {
648 
649 				fs.lookup_still_valid = 1;
650 				/*
651 				 * get rid of the unnecessary page
652 				 */
653 				vm_page_protect(fs.first_m, VM_PROT_NONE);
654 				vm_page_free(fs.first_m);
655 				fs.first_m = NULL;
656 
657 				/*
658 				 * grab the page and put it into the
659 				 * process'es object.  The page is
660 				 * automatically made dirty.
661 				 */
662 				vm_page_rename(fs.m, fs.first_object, fs.first_pindex);
663 				fs.first_m = fs.m;
664 				vm_page_busy(fs.first_m);
665 				fs.m = NULL;
666 				cnt.v_cow_optim++;
667 			} else {
668 				/*
669 				 * Oh, well, lets copy it.
670 				 */
671 				vm_page_copy(fs.m, fs.first_m);
672 			}
673 
674 			if (fs.m) {
675 				/*
676 				 * We no longer need the old page or object.
677 				 */
678 				release_page(&fs);
679 			}
680 
681 			/*
682 			 * fs.object != fs.first_object due to above
683 			 * conditional
684 			 */
685 
686 			vm_object_pip_wakeup(fs.object);
687 
688 			/*
689 			 * Only use the new page below...
690 			 */
691 
692 			cnt.v_cow_faults++;
693 			fs.m = fs.first_m;
694 			fs.object = fs.first_object;
695 			fs.pindex = fs.first_pindex;
696 
697 		} else {
698 			prot &= ~VM_PROT_WRITE;
699 		}
700 	}
701 
702 	/*
703 	 * We must verify that the maps have not changed since our last
704 	 * lookup.
705 	 */
706 
707 	if (!fs.lookup_still_valid &&
708 		(fs.map->timestamp != map_generation)) {
709 		vm_object_t retry_object;
710 		vm_pindex_t retry_pindex;
711 		vm_prot_t retry_prot;
712 
713 		/*
714 		 * Since map entries may be pageable, make sure we can take a
715 		 * page fault on them.
716 		 */
717 
718 		/*
719 		 * Unlock vnode before the lookup to avoid deadlock.   E.G.
720 		 * avoid a deadlock between the inode and exec_map that can
721 		 * occur due to locks being obtained in different orders.
722 		 */
723 
724 		if (fs.vp != NULL) {
725 			vput(fs.vp);
726 			fs.vp = NULL;
727 		}
728 
729 		/*
730 		 * To avoid trying to write_lock the map while another process
731 		 * has it read_locked (in vm_map_pageable), we do not try for
732 		 * write permission.  If the page is still writable, we will
733 		 * get write permission.  If it is not, or has been marked
734 		 * needs_copy, we enter the mapping without write permission,
735 		 * and will merely take another fault.
736 		 */
737 		result = vm_map_lookup(&fs.map, vaddr, fault_type & ~VM_PROT_WRITE,
738 		    &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
739 		map_generation = fs.map->timestamp;
740 
741 		/*
742 		 * If we don't need the page any longer, put it on the active
743 		 * list (the easiest thing to do here).  If no one needs it,
744 		 * pageout will grab it eventually.
745 		 */
746 
747 		if (result != KERN_SUCCESS) {
748 			release_page(&fs);
749 			unlock_and_deallocate(&fs);
750 			return (result);
751 		}
752 		fs.lookup_still_valid = TRUE;
753 
754 		if ((retry_object != fs.first_object) ||
755 		    (retry_pindex != fs.first_pindex)) {
756 			release_page(&fs);
757 			unlock_and_deallocate(&fs);
758 			goto RetryFault;
759 		}
760 		/*
761 		 * Check whether the protection has changed or the object has
762 		 * been copied while we left the map unlocked. Changing from
763 		 * read to write permission is OK - we leave the page
764 		 * write-protected, and catch the write fault. Changing from
765 		 * write to read permission means that we can't mark the page
766 		 * write-enabled after all.
767 		 */
768 		prot &= retry_prot;
769 	}
770 
771 	/*
772 	 * Put this page into the physical map. We had to do the unlock above
773 	 * because pmap_enter may cause other faults.   We don't put the page
774 	 * back on the active queue until later so that the page-out daemon
775 	 * won't find us (yet).
776 	 */
777 
778 	if (prot & VM_PROT_WRITE) {
779 		vm_page_flag_set(fs.m, PG_WRITEABLE);
780 		vm_object_set_flag(fs.m->object,
781 				   OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
782 
783 		/*
784 		 * If the fault is a write, we know that this page is being
785 		 * written NOW so dirty it explicitly to save on
786 		 * pmap_is_modified() calls later.
787 		 *
788 		 * If this is a NOSYNC mmap we do not want to set PG_NOSYNC
789 		 * if the page is already dirty to prevent data written with
790 		 * the expectation of being synced from not being synced.
791 		 * Likewise if this entry does not request NOSYNC then make
792 		 * sure the page isn't marked NOSYNC.  Applications sharing
793 		 * data should use the same flags to avoid ping ponging.
794 		 *
795 		 * Also tell the backing pager, if any, that it should remove
796 		 * any swap backing since the page is now dirty.
797 		 */
798 		if (fault_flags & VM_FAULT_DIRTY) {
799 			if (fs.entry->eflags & MAP_ENTRY_NOSYNC) {
800 				if (fs.m->dirty == 0)
801 					vm_page_flag_set(fs.m, PG_NOSYNC);
802 			} else {
803 				vm_page_flag_clear(fs.m, PG_NOSYNC);
804 			}
805 			vm_page_dirty(fs.m);
806 			vm_pager_page_unswapped(fs.m);
807 		}
808 	}
809 
810 	/*
811 	 * Page had better still be busy
812 	 */
813 
814 	KASSERT(fs.m->flags & PG_BUSY,
815 		("vm_fault: page %p not busy!", fs.m));
816 
817 	unlock_things(&fs);
818 
819 	/*
820 	 * Sanity check: page must be completely valid or it is not fit to
821 	 * map into user space.  vm_pager_get_pages() ensures this.
822 	 */
823 
824 	if (fs.m->valid != VM_PAGE_BITS_ALL) {
825 		vm_page_zero_invalid(fs.m, TRUE);
826 		printf("Warning: page %p partially invalid on fault\n", fs.m);
827 	}
828 
829 	pmap_enter(fs.map->pmap, vaddr, fs.m, prot, wired);
830 
831 	if (((fault_flags & VM_FAULT_WIRE_MASK) == 0) && (wired == 0)) {
832 		pmap_prefault(fs.map->pmap, vaddr, fs.entry);
833 	}
834 
835 	vm_page_flag_clear(fs.m, PG_ZERO);
836 	vm_page_flag_set(fs.m, PG_MAPPED|PG_REFERENCED);
837 	if (fault_flags & VM_FAULT_HOLD)
838 		vm_page_hold(fs.m);
839 
840 	/*
841 	 * If the page is not wired down, then put it where the pageout daemon
842 	 * can find it.
843 	 */
844 
845 	if (fault_flags & VM_FAULT_WIRE_MASK) {
846 		if (wired)
847 			vm_page_wire(fs.m);
848 		else
849 			vm_page_unwire(fs.m, 1);
850 	} else {
851 		vm_page_activate(fs.m);
852 	}
853 
854 	if (curproc && (curproc->p_flag & P_INMEM) && curproc->p_stats) {
855 		if (hardfault) {
856 			curproc->p_stats->p_ru.ru_majflt++;
857 		} else {
858 			curproc->p_stats->p_ru.ru_minflt++;
859 		}
860 	}
861 
862 	/*
863 	 * Unlock everything, and return
864 	 */
865 
866 	vm_page_wakeup(fs.m);
867 	vm_object_deallocate(fs.first_object);
868 
869 	return (KERN_SUCCESS);
870 
871 }
872 
873 /*
874  *	vm_fault_wire:
875  *
876  *	Wire down a range of virtual addresses in a map.
877  */
878 int
879 vm_fault_wire(map, start, end)
880 	vm_map_t map;
881 	vm_offset_t start, end;
882 {
883 
884 	register vm_offset_t va;
885 	register pmap_t pmap;
886 	int rv;
887 
888 	pmap = vm_map_pmap(map);
889 
890 	/*
891 	 * Inform the physical mapping system that the range of addresses may
892 	 * not fault, so that page tables and such can be locked down as well.
893 	 */
894 
895 	pmap_pageable(pmap, start, end, FALSE);
896 
897 	/*
898 	 * We simulate a fault to get the page and enter it in the physical
899 	 * map.
900 	 */
901 
902 	for (va = start; va < end; va += PAGE_SIZE) {
903 		rv = vm_fault(map, va, VM_PROT_READ|VM_PROT_WRITE,
904 			VM_FAULT_CHANGE_WIRING);
905 		if (rv) {
906 			if (va != start)
907 				vm_fault_unwire(map, start, va);
908 			return (rv);
909 		}
910 	}
911 	return (KERN_SUCCESS);
912 }
913 
914 /*
915  *	vm_fault_user_wire:
916  *
917  *	Wire down a range of virtual addresses in a map.  This
918  *	is for user mode though, so we only ask for read access
919  *	on currently read only sections.
920  */
921 int
922 vm_fault_user_wire(map, start, end)
923 	vm_map_t map;
924 	vm_offset_t start, end;
925 {
926 
927 	register vm_offset_t va;
928 	register pmap_t pmap;
929 	int rv;
930 
931 	pmap = vm_map_pmap(map);
932 
933 	/*
934 	 * Inform the physical mapping system that the range of addresses may
935 	 * not fault, so that page tables and such can be locked down as well.
936 	 */
937 
938 	pmap_pageable(pmap, start, end, FALSE);
939 
940 	/*
941 	 * We simulate a fault to get the page and enter it in the physical
942 	 * map.
943 	 */
944 	for (va = start; va < end; va += PAGE_SIZE) {
945 		rv = vm_fault(map, va, VM_PROT_READ, VM_FAULT_USER_WIRE);
946 		if (rv) {
947 			if (va != start)
948 				vm_fault_unwire(map, start, va);
949 			return (rv);
950 		}
951 	}
952 	return (KERN_SUCCESS);
953 }
954 
955 
956 /*
957  *	vm_fault_unwire:
958  *
959  *	Unwire a range of virtual addresses in a map.
960  */
961 void
962 vm_fault_unwire(map, start, end)
963 	vm_map_t map;
964 	vm_offset_t start, end;
965 {
966 
967 	register vm_offset_t va, pa;
968 	register pmap_t pmap;
969 
970 	pmap = vm_map_pmap(map);
971 
972 	/*
973 	 * Since the pages are wired down, we must be able to get their
974 	 * mappings from the physical map system.
975 	 */
976 
977 	for (va = start; va < end; va += PAGE_SIZE) {
978 		pa = pmap_extract(pmap, va);
979 		if (pa != (vm_offset_t) 0) {
980 			pmap_change_wiring(pmap, va, FALSE);
981 			vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1);
982 		}
983 	}
984 
985 	/*
986 	 * Inform the physical mapping system that the range of addresses may
987 	 * fault, so that page tables and such may be unwired themselves.
988 	 */
989 
990 	pmap_pageable(pmap, start, end, TRUE);
991 
992 }
993 
994 /*
995  *	Routine:
996  *		vm_fault_copy_entry
997  *	Function:
998  *		Copy all of the pages from a wired-down map entry to another.
999  *
1000  *	In/out conditions:
1001  *		The source and destination maps must be locked for write.
1002  *		The source map entry must be wired down (or be a sharing map
1003  *		entry corresponding to a main map entry that is wired down).
1004  */
1005 
1006 void
1007 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry)
1008 	vm_map_t dst_map;
1009 	vm_map_t src_map;
1010 	vm_map_entry_t dst_entry;
1011 	vm_map_entry_t src_entry;
1012 {
1013 	vm_object_t dst_object;
1014 	vm_object_t src_object;
1015 	vm_ooffset_t dst_offset;
1016 	vm_ooffset_t src_offset;
1017 	vm_prot_t prot;
1018 	vm_offset_t vaddr;
1019 	vm_page_t dst_m;
1020 	vm_page_t src_m;
1021 
1022 #ifdef	lint
1023 	src_map++;
1024 #endif	/* lint */
1025 
1026 	src_object = src_entry->object.vm_object;
1027 	src_offset = src_entry->offset;
1028 
1029 	/*
1030 	 * Create the top-level object for the destination entry. (Doesn't
1031 	 * actually shadow anything - we copy the pages directly.)
1032 	 */
1033 	dst_object = vm_object_allocate(OBJT_DEFAULT,
1034 	    (vm_size_t) OFF_TO_IDX(dst_entry->end - dst_entry->start));
1035 
1036 	dst_entry->object.vm_object = dst_object;
1037 	dst_entry->offset = 0;
1038 
1039 	prot = dst_entry->max_protection;
1040 
1041 	/*
1042 	 * Loop through all of the pages in the entry's range, copying each
1043 	 * one from the source object (it should be there) to the destination
1044 	 * object.
1045 	 */
1046 	for (vaddr = dst_entry->start, dst_offset = 0;
1047 	    vaddr < dst_entry->end;
1048 	    vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) {
1049 
1050 		/*
1051 		 * Allocate a page in the destination object
1052 		 */
1053 		do {
1054 			dst_m = vm_page_alloc(dst_object,
1055 				OFF_TO_IDX(dst_offset), VM_ALLOC_NORMAL);
1056 			if (dst_m == NULL) {
1057 				VM_WAIT;
1058 			}
1059 		} while (dst_m == NULL);
1060 
1061 		/*
1062 		 * Find the page in the source object, and copy it in.
1063 		 * (Because the source is wired down, the page will be in
1064 		 * memory.)
1065 		 */
1066 		src_m = vm_page_lookup(src_object,
1067 			OFF_TO_IDX(dst_offset + src_offset));
1068 		if (src_m == NULL)
1069 			panic("vm_fault_copy_wired: page missing");
1070 
1071 		vm_page_copy(src_m, dst_m);
1072 
1073 		/*
1074 		 * Enter it in the pmap...
1075 		 */
1076 
1077 		vm_page_flag_clear(dst_m, PG_ZERO);
1078 		pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE);
1079 		vm_page_flag_set(dst_m, PG_WRITEABLE|PG_MAPPED);
1080 
1081 		/*
1082 		 * Mark it no longer busy, and put it on the active list.
1083 		 */
1084 		vm_page_activate(dst_m);
1085 		vm_page_wakeup(dst_m);
1086 	}
1087 }
1088 
1089 
1090 /*
1091  * This routine checks around the requested page for other pages that
1092  * might be able to be faulted in.  This routine brackets the viable
1093  * pages for the pages to be paged in.
1094  *
1095  * Inputs:
1096  *	m, rbehind, rahead
1097  *
1098  * Outputs:
1099  *  marray (array of vm_page_t), reqpage (index of requested page)
1100  *
1101  * Return value:
1102  *  number of pages in marray
1103  */
1104 static int
1105 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage)
1106 	vm_page_t m;
1107 	int rbehind;
1108 	int rahead;
1109 	vm_page_t *marray;
1110 	int *reqpage;
1111 {
1112 	int i,j;
1113 	vm_object_t object;
1114 	vm_pindex_t pindex, startpindex, endpindex, tpindex;
1115 	vm_page_t rtm;
1116 	int cbehind, cahead;
1117 
1118 	object = m->object;
1119 	pindex = m->pindex;
1120 
1121 	/*
1122 	 * we don't fault-ahead for device pager
1123 	 */
1124 	if (object->type == OBJT_DEVICE) {
1125 		*reqpage = 0;
1126 		marray[0] = m;
1127 		return 1;
1128 	}
1129 
1130 	/*
1131 	 * if the requested page is not available, then give up now
1132 	 */
1133 
1134 	if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1135 		return 0;
1136 	}
1137 
1138 	if ((cbehind == 0) && (cahead == 0)) {
1139 		*reqpage = 0;
1140 		marray[0] = m;
1141 		return 1;
1142 	}
1143 
1144 	if (rahead > cahead) {
1145 		rahead = cahead;
1146 	}
1147 
1148 	if (rbehind > cbehind) {
1149 		rbehind = cbehind;
1150 	}
1151 
1152 	/*
1153 	 * try to do any readahead that we might have free pages for.
1154 	 */
1155 	if ((rahead + rbehind) >
1156 		((cnt.v_free_count + cnt.v_cache_count) - cnt.v_free_reserved)) {
1157 		pagedaemon_wakeup();
1158 		marray[0] = m;
1159 		*reqpage = 0;
1160 		return 1;
1161 	}
1162 
1163 	/*
1164 	 * scan backward for the read behind pages -- in memory
1165 	 */
1166 	if (pindex > 0) {
1167 		if (rbehind > pindex) {
1168 			rbehind = pindex;
1169 			startpindex = 0;
1170 		} else {
1171 			startpindex = pindex - rbehind;
1172 		}
1173 
1174 		for ( tpindex = pindex - 1; tpindex >= startpindex; tpindex -= 1) {
1175 			if (vm_page_lookup( object, tpindex)) {
1176 				startpindex = tpindex + 1;
1177 				break;
1178 			}
1179 			if (tpindex == 0)
1180 				break;
1181 		}
1182 
1183 		for(i = 0, tpindex = startpindex; tpindex < pindex; i++, tpindex++) {
1184 
1185 			rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
1186 			if (rtm == NULL) {
1187 				for (j = 0; j < i; j++) {
1188 					vm_page_free(marray[j]);
1189 				}
1190 				marray[0] = m;
1191 				*reqpage = 0;
1192 				return 1;
1193 			}
1194 
1195 			marray[i] = rtm;
1196 		}
1197 	} else {
1198 		startpindex = 0;
1199 		i = 0;
1200 	}
1201 
1202 	marray[i] = m;
1203 	/* page offset of the required page */
1204 	*reqpage = i;
1205 
1206 	tpindex = pindex + 1;
1207 	i++;
1208 
1209 	/*
1210 	 * scan forward for the read ahead pages
1211 	 */
1212 	endpindex = tpindex + rahead;
1213 	if (endpindex > object->size)
1214 		endpindex = object->size;
1215 
1216 	for( ; tpindex < endpindex; i++, tpindex++) {
1217 
1218 		if (vm_page_lookup(object, tpindex)) {
1219 			break;
1220 		}
1221 
1222 		rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
1223 		if (rtm == NULL) {
1224 			break;
1225 		}
1226 
1227 		marray[i] = rtm;
1228 	}
1229 
1230 	/* return number of bytes of pages */
1231 	return i;
1232 }
1233