xref: /freebsd/sys/vm/vm_fault.c (revision 6990ffd8a95caaba6858ad44ff1b3157d1efba8f)
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/kernel.h>
79 #include <sys/lock.h>
80 #include <sys/mutex.h>
81 #include <sys/proc.h>
82 #include <sys/resourcevar.h>
83 #include <sys/sysctl.h>
84 #include <sys/vmmeter.h>
85 #include <sys/vnode.h>
86 
87 #include <vm/vm.h>
88 #include <vm/vm_param.h>
89 #include <vm/pmap.h>
90 #include <vm/vm_map.h>
91 #include <vm/vm_object.h>
92 #include <vm/vm_page.h>
93 #include <vm/vm_pageout.h>
94 #include <vm/vm_kern.h>
95 #include <vm/vm_pager.h>
96 #include <vm/vnode_pager.h>
97 #include <vm/vm_extern.h>
98 
99 static int vm_fault_additional_pages __P((vm_page_t, int,
100 					  int, vm_page_t *, int *));
101 
102 #define VM_FAULT_READ_AHEAD 8
103 #define VM_FAULT_READ_BEHIND 7
104 #define VM_FAULT_READ (VM_FAULT_READ_AHEAD+VM_FAULT_READ_BEHIND+1)
105 
106 struct faultstate {
107 	vm_page_t m;
108 	vm_object_t object;
109 	vm_pindex_t pindex;
110 	vm_page_t first_m;
111 	vm_object_t	first_object;
112 	vm_pindex_t first_pindex;
113 	vm_map_t map;
114 	vm_map_entry_t entry;
115 	int lookup_still_valid;
116 	struct vnode *vp;
117 };
118 
119 static __inline void
120 release_page(struct faultstate *fs)
121 {
122 	vm_page_wakeup(fs->m);
123 	vm_page_deactivate(fs->m);
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_free(fs->first_m);
143 		vm_object_pip_wakeup(fs->first_object);
144 		fs->first_m = NULL;
145 	}
146 	if (dealloc) {
147 		vm_object_deallocate(fs->first_object);
148 	}
149 	unlock_map(fs);
150 	if (fs->vp != NULL) {
151 		vput(fs->vp);
152 		fs->vp = NULL;
153 	}
154 }
155 
156 #define unlock_things(fs) _unlock_things(fs, 0)
157 #define unlock_and_deallocate(fs) _unlock_things(fs, 1)
158 
159 /*
160  * TRYPAGER - used by vm_fault to calculate whether the pager for the
161  *	      current object *might* contain the page.
162  *
163  *	      default objects are zero-fill, there is no real pager.
164  */
165 
166 #define TRYPAGER	(fs.object->type != OBJT_DEFAULT && \
167 			(((fault_flags & VM_FAULT_WIRE_MASK) == 0) || wired))
168 
169 /*
170  *	vm_fault:
171  *
172  *	Handle a page fault occurring at the given address,
173  *	requiring the given permissions, in the map specified.
174  *	If successful, the page is inserted into the
175  *	associated physical map.
176  *
177  *	NOTE: the given address should be truncated to the
178  *	proper page address.
179  *
180  *	KERN_SUCCESS is returned if the page fault is handled; otherwise,
181  *	a standard error specifying why the fault is fatal is returned.
182  *
183  *
184  *	The map in question must be referenced, and remains so.
185  *	Caller may hold no locks.
186  */
187 static int vm_fault1 __P((vm_map_t, vm_offset_t, vm_prot_t, int));
188 
189 int
190 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
191 	 int fault_flags)
192 {
193 	int ret;
194 
195 	mtx_lock(&Giant);
196 	/* GIANT_REQUIRED */
197 
198 	ret = vm_fault1(map, vaddr, fault_type, fault_flags);
199 	mtx_unlock(&Giant);
200 	return (ret);
201 }
202 
203 static int
204 vm_fault1(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
205 	  int fault_flags)
206 {
207 	vm_prot_t prot;
208 	int result;
209 	boolean_t wired;
210 	int map_generation;
211 	vm_object_t next_object;
212 	vm_page_t marray[VM_FAULT_READ];
213 	int hardfault;
214 	int faultcount;
215 	struct faultstate fs;
216 
217 	GIANT_REQUIRED;
218 
219 	cnt.v_vm_faults++;
220 	hardfault = 0;
221 
222 RetryFault:;
223 
224 	/*
225 	 * Find the backing store object and offset into it to begin the
226 	 * search.
227 	 */
228 	fs.map = map;
229 	if ((result = vm_map_lookup(&fs.map, vaddr,
230 		fault_type, &fs.entry, &fs.first_object,
231 		&fs.first_pindex, &prot, &wired)) != KERN_SUCCESS) {
232 		if ((result != KERN_PROTECTION_FAILURE) ||
233 			((fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE)) {
234 			return result;
235 		}
236 
237 		/*
238    		 * If we are user-wiring a r/w segment, and it is COW, then
239    		 * we need to do the COW operation.  Note that we don't COW
240    		 * currently RO sections now, because it is NOT desirable
241    		 * to COW .text.  We simply keep .text from ever being COW'ed
242    		 * and take the heat that one cannot debug wired .text sections.
243    		 */
244 		result = vm_map_lookup(&fs.map, vaddr,
245 			VM_PROT_READ|VM_PROT_WRITE|VM_PROT_OVERRIDE_WRITE,
246 			&fs.entry, &fs.first_object, &fs.first_pindex, &prot, &wired);
247 		if (result != KERN_SUCCESS) {
248 			return result;
249 		}
250 
251 		/*
252 		 * If we don't COW now, on a user wire, the user will never
253 		 * be able to write to the mapping.  If we don't make this
254 		 * restriction, the bookkeeping would be nearly impossible.
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 	vm_object_reference(fs.first_object);
274 	vm_object_pip_add(fs.first_object, 1);
275 
276 	fs.vp = vnode_pager_lock(fs.first_object);
277 	if ((fault_type & VM_PROT_WRITE) &&
278 		(fs.first_object->type == OBJT_VNODE)) {
279 		vm_freeze_copyopts(fs.first_object,
280 			fs.first_pindex, fs.first_pindex + 1);
281 	}
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 
294 	fs.object = fs.first_object;
295 	fs.pindex = fs.first_pindex;
296 
297 	while (TRUE) {
298 		/*
299 		 * If the object is dead, we stop here
300 		 */
301 
302 		if (fs.object->flags & OBJ_DEAD) {
303 			unlock_and_deallocate(&fs);
304 			return (KERN_PROTECTION_FAILURE);
305 		}
306 
307 		/*
308 		 * See if page is resident
309 		 */
310 
311 		fs.m = vm_page_lookup(fs.object, fs.pindex);
312 		if (fs.m != NULL) {
313 			int queue, s;
314 			/*
315 			 * Wait/Retry if the page is busy.  We have to do this
316 			 * if the page is busy via either PG_BUSY or
317 			 * vm_page_t->busy because the vm_pager may be using
318 			 * vm_page_t->busy for pageouts ( and even pageins if
319 			 * it is the vnode pager ), and we could end up trying
320 			 * to pagein and pageout the same page simultaneously.
321 			 *
322 			 * We can theoretically allow the busy case on a read
323 			 * fault if the page is marked valid, but since such
324 			 * pages are typically already pmap'd, putting that
325 			 * special case in might be more effort then it is
326 			 * worth.  We cannot under any circumstances mess
327 			 * around with a vm_page_t->busy page except, perhaps,
328 			 * to pmap it.
329 			 */
330 			if ((fs.m->flags & PG_BUSY) || fs.m->busy) {
331 				unlock_things(&fs);
332 				(void)vm_page_sleep_busy(fs.m, TRUE, "vmpfw");
333 				cnt.v_intrans++;
334 				vm_object_deallocate(fs.first_object);
335 				goto RetryFault;
336 			}
337 
338 			queue = fs.m->queue;
339 			s = splvm();
340 			vm_pageq_remove_nowakeup(fs.m);
341 			splx(s);
342 
343 			if ((queue - fs.m->pc) == PQ_CACHE && vm_page_count_severe()) {
344 				vm_page_activate(fs.m);
345 				unlock_and_deallocate(&fs);
346 				VM_WAIT;
347 				goto RetryFault;
348 			}
349 
350 			/*
351 			 * Mark page busy for other processes, and the
352 			 * pagedaemon.  If it still isn't completely valid
353 			 * (readable), jump to readrest, else break-out ( we
354 			 * found the page ).
355 			 */
356 
357 			vm_page_busy(fs.m);
358 			if (((fs.m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) &&
359 				fs.m->object != kernel_object && fs.m->object != kmem_object) {
360 				goto readrest;
361 			}
362 
363 			break;
364 		}
365 
366 		/*
367 		 * Page is not resident, If this is the search termination
368 		 * or the pager might contain the page, allocate a new page.
369 		 */
370 
371 		if (TRYPAGER || fs.object == fs.first_object) {
372 			if (fs.pindex >= fs.object->size) {
373 				unlock_and_deallocate(&fs);
374 				return (KERN_PROTECTION_FAILURE);
375 			}
376 
377 			/*
378 			 * Allocate a new page for this object/offset pair.
379 			 */
380 			fs.m = NULL;
381 			if (!vm_page_count_severe()) {
382 				fs.m = vm_page_alloc(fs.object, fs.pindex,
383 				    (fs.vp || fs.object->backing_object)? VM_ALLOC_NORMAL: VM_ALLOC_ZERO);
384 			}
385 			if (fs.m == NULL) {
386 				unlock_and_deallocate(&fs);
387 				VM_WAIT;
388 				goto RetryFault;
389 			}
390 		}
391 
392 readrest:
393 		/*
394 		 * We have found a valid page or we have allocated a new page.
395 		 * The page thus may not be valid or may not be entirely
396 		 * valid.
397 		 *
398 		 * Attempt to fault-in the page if there is a chance that the
399 		 * pager has it, and potentially fault in additional pages
400 		 * at the same time.
401 		 */
402 
403 		if (TRYPAGER) {
404 			int rv;
405 			int reqpage;
406 			int ahead, behind;
407 			u_char behavior = vm_map_entry_behavior(fs.entry);
408 
409 			if (behavior == MAP_ENTRY_BEHAV_RANDOM) {
410 				ahead = 0;
411 				behind = 0;
412 			} else {
413 				behind = (vaddr - fs.entry->start) >> PAGE_SHIFT;
414 				if (behind > VM_FAULT_READ_BEHIND)
415 					behind = VM_FAULT_READ_BEHIND;
416 
417 				ahead = ((fs.entry->end - vaddr) >> PAGE_SHIFT) - 1;
418 				if (ahead > VM_FAULT_READ_AHEAD)
419 					ahead = VM_FAULT_READ_AHEAD;
420 			}
421 
422 			if ((fs.first_object->type != OBJT_DEVICE) &&
423 			    (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL ||
424                                 (behavior != MAP_ENTRY_BEHAV_RANDOM &&
425                                 fs.pindex >= fs.entry->lastr &&
426                                 fs.pindex < fs.entry->lastr + VM_FAULT_READ))
427 			) {
428 				vm_pindex_t firstpindex, tmppindex;
429 
430 				if (fs.first_pindex < 2 * VM_FAULT_READ)
431 					firstpindex = 0;
432 				else
433 					firstpindex = fs.first_pindex - 2 * VM_FAULT_READ;
434 
435 				/*
436 				 * note: partially valid pages cannot be
437 				 * included in the lookahead - NFS piecemeal
438 				 * writes will barf on it badly.
439 				 */
440 
441 				for(tmppindex = fs.first_pindex - 1;
442 					tmppindex >= firstpindex;
443 					--tmppindex) {
444 					vm_page_t mt;
445 					mt = vm_page_lookup( fs.first_object, tmppindex);
446 					if (mt == NULL || (mt->valid != VM_PAGE_BITS_ALL))
447 						break;
448 					if (mt->busy ||
449 						(mt->flags & (PG_BUSY | PG_FICTITIOUS | PG_UNMANAGED)) ||
450 						mt->hold_count ||
451 						mt->wire_count)
452 						continue;
453 					if (mt->dirty == 0)
454 						vm_page_test_dirty(mt);
455 					if (mt->dirty) {
456 						vm_page_protect(mt, VM_PROT_NONE);
457 						vm_page_deactivate(mt);
458 					} else {
459 						vm_page_cache(mt);
460 					}
461 				}
462 
463 				ahead += behind;
464 				behind = 0;
465 			}
466 
467 			/*
468 			 * now we find out if any other pages should be paged
469 			 * in at this time this routine checks to see if the
470 			 * pages surrounding this fault reside in the same
471 			 * object as the page for this fault.  If they do,
472 			 * then they are faulted in also into the object.  The
473 			 * array "marray" returned contains an array of
474 			 * vm_page_t structs where one of them is the
475 			 * vm_page_t passed to the routine.  The reqpage
476 			 * return value is the index into the marray for the
477 			 * vm_page_t passed to the routine.
478 			 *
479 			 * fs.m plus the additional pages are PG_BUSY'd.
480 			 */
481 			faultcount = vm_fault_additional_pages(
482 			    fs.m, behind, ahead, marray, &reqpage);
483 
484 			/*
485 			 * update lastr imperfectly (we do not know how much
486 			 * getpages will actually read), but good enough.
487 			 */
488 			fs.entry->lastr = fs.pindex + faultcount - behind;
489 
490 			/*
491 			 * Call the pager to retrieve the data, if any, after
492 			 * releasing the lock on the map.  We hold a ref on
493 			 * fs.object and the pages are PG_BUSY'd.
494 			 */
495 			unlock_map(&fs);
496 
497 			rv = faultcount ?
498 			    vm_pager_get_pages(fs.object, marray, faultcount,
499 				reqpage) : VM_PAGER_FAIL;
500 
501 			if (rv == VM_PAGER_OK) {
502 				/*
503 				 * Found the page. Leave it busy while we play
504 				 * with it.
505 				 */
506 
507 				/*
508 				 * Relookup in case pager changed page. Pager
509 				 * is responsible for disposition of old page
510 				 * if moved.
511 				 */
512 				fs.m = vm_page_lookup(fs.object, fs.pindex);
513 				if(!fs.m) {
514 					unlock_and_deallocate(&fs);
515 					goto RetryFault;
516 				}
517 
518 				hardfault++;
519 				break; /* break to PAGE HAS BEEN FOUND */
520 			}
521 			/*
522 			 * Remove the bogus page (which does not exist at this
523 			 * object/offset); before doing so, we must get back
524 			 * our object lock to preserve our invariant.
525 			 *
526 			 * Also wake up any other process that may want to bring
527 			 * in this page.
528 			 *
529 			 * If this is the top-level object, we must leave the
530 			 * busy page to prevent another process from rushing
531 			 * past us, and inserting the page in that object at
532 			 * the same time that we are.
533 			 */
534 
535 			if (rv == VM_PAGER_ERROR)
536 				printf("vm_fault: pager read error, pid %d (%s)\n",
537 				    curproc->p_pid, curproc->p_comm);
538 			/*
539 			 * Data outside the range of the pager or an I/O error
540 			 */
541 			/*
542 			 * XXX - the check for kernel_map is a kludge to work
543 			 * around having the machine panic on a kernel space
544 			 * fault w/ I/O error.
545 			 */
546 			if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) ||
547 				(rv == VM_PAGER_BAD)) {
548 				vm_page_free(fs.m);
549 				fs.m = NULL;
550 				unlock_and_deallocate(&fs);
551 				return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
552 			}
553 			if (fs.object != fs.first_object) {
554 				vm_page_free(fs.m);
555 				fs.m = NULL;
556 				/*
557 				 * XXX - we cannot just fall out at this
558 				 * point, m has been freed and is invalid!
559 				 */
560 			}
561 		}
562 
563 		/*
564 		 * We get here if the object has default pager (or unwiring)
565 		 * or the pager doesn't have the page.
566 		 */
567 		if (fs.object == fs.first_object)
568 			fs.first_m = fs.m;
569 
570 		/*
571 		 * Move on to the next object.  Lock the next object before
572 		 * unlocking the current one.
573 		 */
574 
575 		fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset);
576 		next_object = fs.object->backing_object;
577 		if (next_object == NULL) {
578 			/*
579 			 * If there's no object left, fill the page in the top
580 			 * object with zeros.
581 			 */
582 			if (fs.object != fs.first_object) {
583 				vm_object_pip_wakeup(fs.object);
584 
585 				fs.object = fs.first_object;
586 				fs.pindex = fs.first_pindex;
587 				fs.m = fs.first_m;
588 			}
589 			fs.first_m = NULL;
590 
591 			/*
592 			 * Zero the page if necessary and mark it valid.
593 			 */
594 			if ((fs.m->flags & PG_ZERO) == 0) {
595 				vm_page_zero_fill(fs.m);
596 			} else {
597 				cnt.v_ozfod++;
598 			}
599 			cnt.v_zfod++;
600 			fs.m->valid = VM_PAGE_BITS_ALL;
601 			break;	/* break to PAGE HAS BEEN FOUND */
602 		} else {
603 			if (fs.object != fs.first_object) {
604 				vm_object_pip_wakeup(fs.object);
605 			}
606 			KASSERT(fs.object != next_object, ("object loop %p", next_object));
607 			fs.object = next_object;
608 			vm_object_pip_add(fs.object, 1);
609 		}
610 	}
611 
612 	KASSERT((fs.m->flags & PG_BUSY) != 0,
613 	    ("vm_fault: not busy after main loop"));
614 
615 	/*
616 	 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
617 	 * is held.]
618 	 */
619 
620 	/*
621 	 * If the page is being written, but isn't already owned by the
622 	 * top-level object, we have to copy it into a new page owned by the
623 	 * top-level object.
624 	 */
625 
626 	if (fs.object != fs.first_object) {
627 		/*
628 		 * We only really need to copy if we want to write it.
629 		 */
630 
631 		if (fault_type & VM_PROT_WRITE) {
632 			/*
633 			 * This allows pages to be virtually copied from a
634 			 * backing_object into the first_object, where the
635 			 * backing object has no other refs to it, and cannot
636 			 * gain any more refs.  Instead of a bcopy, we just
637 			 * move the page from the backing object to the
638 			 * first object.  Note that we must mark the page
639 			 * dirty in the first object so that it will go out
640 			 * to swap when needed.
641 			 */
642 			if (map_generation == fs.map->timestamp &&
643 				/*
644 				 * Only one shadow object
645 				 */
646 				(fs.object->shadow_count == 1) &&
647 				/*
648 				 * No COW refs, except us
649 				 */
650 				(fs.object->ref_count == 1) &&
651 				/*
652 				 * No one else can look this object up
653 				 */
654 				(fs.object->handle == NULL) &&
655 				/*
656 				 * No other ways to look the object up
657 				 */
658 				((fs.object->type == OBJT_DEFAULT) ||
659 				 (fs.object->type == OBJT_SWAP)) &&
660 				/*
661 				 * We don't chase down the shadow chain
662 				 */
663 				(fs.object == fs.first_object->backing_object) &&
664 
665 				/*
666 				 * grab the lock if we need to
667 				 */
668 				(fs.lookup_still_valid ||
669 				 lockmgr(&fs.map->lock, LK_EXCLUSIVE|LK_NOWAIT, (void *)0, curthread) == 0)
670 			    ) {
671 
672 				fs.lookup_still_valid = 1;
673 				/*
674 				 * get rid of the unnecessary page
675 				 */
676 				vm_page_protect(fs.first_m, VM_PROT_NONE);
677 				vm_page_free(fs.first_m);
678 				fs.first_m = NULL;
679 
680 				/*
681 				 * grab the page and put it into the
682 				 * process'es object.  The page is
683 				 * automatically made dirty.
684 				 */
685 				vm_page_rename(fs.m, fs.first_object, fs.first_pindex);
686 				fs.first_m = fs.m;
687 				vm_page_busy(fs.first_m);
688 				fs.m = NULL;
689 				cnt.v_cow_optim++;
690 			} else {
691 				/*
692 				 * Oh, well, lets copy it.
693 				 */
694 				vm_page_copy(fs.m, fs.first_m);
695 			}
696 
697 			if (fs.m) {
698 				/*
699 				 * We no longer need the old page or object.
700 				 */
701 				release_page(&fs);
702 			}
703 
704 			/*
705 			 * fs.object != fs.first_object due to above
706 			 * conditional
707 			 */
708 
709 			vm_object_pip_wakeup(fs.object);
710 
711 			/*
712 			 * Only use the new page below...
713 			 */
714 
715 			cnt.v_cow_faults++;
716 			fs.m = fs.first_m;
717 			fs.object = fs.first_object;
718 			fs.pindex = fs.first_pindex;
719 
720 		} else {
721 			prot &= ~VM_PROT_WRITE;
722 		}
723 	}
724 
725 	/*
726 	 * We must verify that the maps have not changed since our last
727 	 * lookup.
728 	 */
729 
730 	if (!fs.lookup_still_valid &&
731 		(fs.map->timestamp != map_generation)) {
732 		vm_object_t retry_object;
733 		vm_pindex_t retry_pindex;
734 		vm_prot_t retry_prot;
735 
736 		/*
737 		 * Since map entries may be pageable, make sure we can take a
738 		 * page fault on them.
739 		 */
740 
741 		/*
742 		 * Unlock vnode before the lookup to avoid deadlock.   E.G.
743 		 * avoid a deadlock between the inode and exec_map that can
744 		 * occur due to locks being obtained in different orders.
745 		 */
746 
747 		if (fs.vp != NULL) {
748 			vput(fs.vp);
749 			fs.vp = NULL;
750 		}
751 
752 		if (fs.map->infork) {
753 			release_page(&fs);
754 			unlock_and_deallocate(&fs);
755 			goto RetryFault;
756 		}
757 
758 		/*
759 		 * To avoid trying to write_lock the map while another process
760 		 * has it read_locked (in vm_map_pageable), we do not try for
761 		 * write permission.  If the page is still writable, we will
762 		 * get write permission.  If it is not, or has been marked
763 		 * needs_copy, we enter the mapping without write permission,
764 		 * and will merely take another fault.
765 		 */
766 		result = vm_map_lookup(&fs.map, vaddr, fault_type & ~VM_PROT_WRITE,
767 		    &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
768 		map_generation = fs.map->timestamp;
769 
770 		/*
771 		 * If we don't need the page any longer, put it on the active
772 		 * list (the easiest thing to do here).  If no one needs it,
773 		 * pageout will grab it eventually.
774 		 */
775 
776 		if (result != KERN_SUCCESS) {
777 			release_page(&fs);
778 			unlock_and_deallocate(&fs);
779 			return (result);
780 		}
781 		fs.lookup_still_valid = TRUE;
782 
783 		if ((retry_object != fs.first_object) ||
784 		    (retry_pindex != fs.first_pindex)) {
785 			release_page(&fs);
786 			unlock_and_deallocate(&fs);
787 			goto RetryFault;
788 		}
789 		/*
790 		 * Check whether the protection has changed or the object has
791 		 * been copied while we left the map unlocked. Changing from
792 		 * read to write permission is OK - we leave the page
793 		 * write-protected, and catch the write fault. Changing from
794 		 * write to read permission means that we can't mark the page
795 		 * write-enabled after all.
796 		 */
797 		prot &= retry_prot;
798 	}
799 
800 	/*
801 	 * Put this page into the physical map. We had to do the unlock above
802 	 * because pmap_enter may cause other faults.   We don't put the page
803 	 * back on the active queue until later so that the page-out daemon
804 	 * won't find us (yet).
805 	 */
806 
807 	if (prot & VM_PROT_WRITE) {
808 		vm_page_flag_set(fs.m, PG_WRITEABLE);
809 		vm_object_set_flag(fs.m->object,
810 				   OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
811 
812 		/*
813 		 * If the fault is a write, we know that this page is being
814 		 * written NOW so dirty it explicitly to save on
815 		 * pmap_is_modified() calls later.
816 		 *
817 		 * If this is a NOSYNC mmap we do not want to set PG_NOSYNC
818 		 * if the page is already dirty to prevent data written with
819 		 * the expectation of being synced from not being synced.
820 		 * Likewise if this entry does not request NOSYNC then make
821 		 * sure the page isn't marked NOSYNC.  Applications sharing
822 		 * data should use the same flags to avoid ping ponging.
823 		 *
824 		 * Also tell the backing pager, if any, that it should remove
825 		 * any swap backing since the page is now dirty.
826 		 */
827 		if (fs.entry->eflags & MAP_ENTRY_NOSYNC) {
828 			if (fs.m->dirty == 0)
829 				vm_page_flag_set(fs.m, PG_NOSYNC);
830 		} else {
831 			vm_page_flag_clear(fs.m, PG_NOSYNC);
832 		}
833 		if (fault_flags & VM_FAULT_DIRTY) {
834 			int s;
835 			vm_page_dirty(fs.m);
836 			s = splvm();
837 			vm_pager_page_unswapped(fs.m);
838 			splx(s);
839 		}
840 	}
841 
842 	/*
843 	 * Page had better still be busy
844 	 */
845 
846 	KASSERT(fs.m->flags & PG_BUSY,
847 		("vm_fault: page %p not busy!", fs.m));
848 
849 	unlock_things(&fs);
850 
851 	/*
852 	 * Sanity check: page must be completely valid or it is not fit to
853 	 * map into user space.  vm_pager_get_pages() ensures this.
854 	 */
855 
856 	if (fs.m->valid != VM_PAGE_BITS_ALL) {
857 		vm_page_zero_invalid(fs.m, TRUE);
858 		printf("Warning: page %p partially invalid on fault\n", fs.m);
859 	}
860 
861 	pmap_enter(fs.map->pmap, vaddr, fs.m, prot, wired);
862 
863 	if (((fault_flags & VM_FAULT_WIRE_MASK) == 0) && (wired == 0)) {
864 		pmap_prefault(fs.map->pmap, vaddr, fs.entry);
865 	}
866 
867 	vm_page_flag_clear(fs.m, PG_ZERO);
868 	vm_page_flag_set(fs.m, PG_MAPPED|PG_REFERENCED);
869 	if (fault_flags & VM_FAULT_HOLD)
870 		vm_page_hold(fs.m);
871 
872 	/*
873 	 * If the page is not wired down, then put it where the pageout daemon
874 	 * can find it.
875 	 */
876 
877 	if (fault_flags & VM_FAULT_WIRE_MASK) {
878 		if (wired)
879 			vm_page_wire(fs.m);
880 		else
881 			vm_page_unwire(fs.m, 1);
882 	} else {
883 		vm_page_activate(fs.m);
884 	}
885 
886 	mtx_lock_spin(&sched_lock);
887 	if (curproc && (curproc->p_sflag & PS_INMEM) && curproc->p_stats) {
888 		if (hardfault) {
889 			curproc->p_stats->p_ru.ru_majflt++;
890 		} else {
891 			curproc->p_stats->p_ru.ru_minflt++;
892 		}
893 	}
894 	mtx_unlock_spin(&sched_lock);
895 
896 	/*
897 	 * Unlock everything, and return
898 	 */
899 
900 	vm_page_wakeup(fs.m);
901 	vm_object_deallocate(fs.first_object);
902 
903 	return (KERN_SUCCESS);
904 
905 }
906 
907 /*
908  *	vm_fault_wire:
909  *
910  *	Wire down a range of virtual addresses in a map.
911  */
912 int
913 vm_fault_wire(map, start, end)
914 	vm_map_t map;
915 	vm_offset_t start, end;
916 {
917 
918 	vm_offset_t va;
919 	pmap_t pmap;
920 	int rv;
921 
922 	pmap = vm_map_pmap(map);
923 
924 	/*
925 	 * Inform the physical mapping system that the range of addresses may
926 	 * not fault, so that page tables and such can be locked down as well.
927 	 */
928 
929 	pmap_pageable(pmap, start, end, FALSE);
930 
931 	/*
932 	 * We simulate a fault to get the page and enter it in the physical
933 	 * map.
934 	 */
935 
936 	for (va = start; va < end; va += PAGE_SIZE) {
937 		rv = vm_fault(map, va, VM_PROT_READ|VM_PROT_WRITE,
938 			VM_FAULT_CHANGE_WIRING);
939 		if (rv) {
940 			if (va != start)
941 				vm_fault_unwire(map, start, va);
942 			return (rv);
943 		}
944 	}
945 	return (KERN_SUCCESS);
946 }
947 
948 /*
949  *	vm_fault_user_wire:
950  *
951  *	Wire down a range of virtual addresses in a map.  This
952  *	is for user mode though, so we only ask for read access
953  *	on currently read only sections.
954  */
955 int
956 vm_fault_user_wire(map, start, end)
957 	vm_map_t map;
958 	vm_offset_t start, end;
959 {
960 
961 	vm_offset_t va;
962 	pmap_t pmap;
963 	int rv;
964 
965 	GIANT_REQUIRED;
966 
967 	pmap = vm_map_pmap(map);
968 
969 	/*
970 	 * Inform the physical mapping system that the range of addresses may
971 	 * not fault, so that page tables and such can be locked down as well.
972 	 */
973 
974 	pmap_pageable(pmap, start, end, FALSE);
975 
976 	/*
977 	 * We simulate a fault to get the page and enter it in the physical
978 	 * map.
979 	 */
980 	for (va = start; va < end; va += PAGE_SIZE) {
981 		rv = vm_fault(map, va, VM_PROT_READ, VM_FAULT_USER_WIRE);
982 		if (rv) {
983 			if (va != start)
984 				vm_fault_unwire(map, start, va);
985 			return (rv);
986 		}
987 	}
988 	return (KERN_SUCCESS);
989 }
990 
991 
992 /*
993  *	vm_fault_unwire:
994  *
995  *	Unwire a range of virtual addresses in a map.
996  */
997 void
998 vm_fault_unwire(map, start, end)
999 	vm_map_t map;
1000 	vm_offset_t start, end;
1001 {
1002 
1003 	vm_offset_t va, pa;
1004 	pmap_t pmap;
1005 
1006 	pmap = vm_map_pmap(map);
1007 
1008 	/*
1009 	 * Since the pages are wired down, we must be able to get their
1010 	 * mappings from the physical map system.
1011 	 */
1012 
1013 	for (va = start; va < end; va += PAGE_SIZE) {
1014 		pa = pmap_extract(pmap, va);
1015 		if (pa != (vm_offset_t) 0) {
1016 			pmap_change_wiring(pmap, va, FALSE);
1017 			vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1);
1018 		}
1019 	}
1020 
1021 	/*
1022 	 * Inform the physical mapping system that the range of addresses may
1023 	 * fault, so that page tables and such may be unwired themselves.
1024 	 */
1025 
1026 	pmap_pageable(pmap, start, end, TRUE);
1027 
1028 }
1029 
1030 /*
1031  *	Routine:
1032  *		vm_fault_copy_entry
1033  *	Function:
1034  *		Copy all of the pages from a wired-down map entry to another.
1035  *
1036  *	In/out conditions:
1037  *		The source and destination maps must be locked for write.
1038  *		The source map entry must be wired down (or be a sharing map
1039  *		entry corresponding to a main map entry that is wired down).
1040  */
1041 
1042 void
1043 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry)
1044 	vm_map_t dst_map;
1045 	vm_map_t src_map;
1046 	vm_map_entry_t dst_entry;
1047 	vm_map_entry_t src_entry;
1048 {
1049 	vm_object_t dst_object;
1050 	vm_object_t src_object;
1051 	vm_ooffset_t dst_offset;
1052 	vm_ooffset_t src_offset;
1053 	vm_prot_t prot;
1054 	vm_offset_t vaddr;
1055 	vm_page_t dst_m;
1056 	vm_page_t src_m;
1057 
1058 #ifdef	lint
1059 	src_map++;
1060 #endif	/* lint */
1061 
1062 	src_object = src_entry->object.vm_object;
1063 	src_offset = src_entry->offset;
1064 
1065 	/*
1066 	 * Create the top-level object for the destination entry. (Doesn't
1067 	 * actually shadow anything - we copy the pages directly.)
1068 	 */
1069 	dst_object = vm_object_allocate(OBJT_DEFAULT,
1070 	    (vm_size_t) OFF_TO_IDX(dst_entry->end - dst_entry->start));
1071 
1072 	dst_entry->object.vm_object = dst_object;
1073 	dst_entry->offset = 0;
1074 
1075 	prot = dst_entry->max_protection;
1076 
1077 	/*
1078 	 * Loop through all of the pages in the entry's range, copying each
1079 	 * one from the source object (it should be there) to the destination
1080 	 * object.
1081 	 */
1082 	for (vaddr = dst_entry->start, dst_offset = 0;
1083 	    vaddr < dst_entry->end;
1084 	    vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) {
1085 
1086 		/*
1087 		 * Allocate a page in the destination object
1088 		 */
1089 		do {
1090 			dst_m = vm_page_alloc(dst_object,
1091 				OFF_TO_IDX(dst_offset), VM_ALLOC_NORMAL);
1092 			if (dst_m == NULL) {
1093 				VM_WAIT;
1094 			}
1095 		} while (dst_m == NULL);
1096 
1097 		/*
1098 		 * Find the page in the source object, and copy it in.
1099 		 * (Because the source is wired down, the page will be in
1100 		 * memory.)
1101 		 */
1102 		src_m = vm_page_lookup(src_object,
1103 			OFF_TO_IDX(dst_offset + src_offset));
1104 		if (src_m == NULL)
1105 			panic("vm_fault_copy_wired: page missing");
1106 
1107 		vm_page_copy(src_m, dst_m);
1108 
1109 		/*
1110 		 * Enter it in the pmap...
1111 		 */
1112 
1113 		vm_page_flag_clear(dst_m, PG_ZERO);
1114 		pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE);
1115 		vm_page_flag_set(dst_m, PG_WRITEABLE|PG_MAPPED);
1116 
1117 		/*
1118 		 * Mark it no longer busy, and put it on the active list.
1119 		 */
1120 		vm_page_activate(dst_m);
1121 		vm_page_wakeup(dst_m);
1122 	}
1123 }
1124 
1125 
1126 /*
1127  * This routine checks around the requested page for other pages that
1128  * might be able to be faulted in.  This routine brackets the viable
1129  * pages for the pages to be paged in.
1130  *
1131  * Inputs:
1132  *	m, rbehind, rahead
1133  *
1134  * Outputs:
1135  *  marray (array of vm_page_t), reqpage (index of requested page)
1136  *
1137  * Return value:
1138  *  number of pages in marray
1139  *
1140  * This routine can't block.
1141  */
1142 static int
1143 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage)
1144 	vm_page_t m;
1145 	int rbehind;
1146 	int rahead;
1147 	vm_page_t *marray;
1148 	int *reqpage;
1149 {
1150 	int i,j;
1151 	vm_object_t object;
1152 	vm_pindex_t pindex, startpindex, endpindex, tpindex;
1153 	vm_page_t rtm;
1154 	int cbehind, cahead;
1155 
1156 	GIANT_REQUIRED;
1157 
1158 	object = m->object;
1159 	pindex = m->pindex;
1160 
1161 	/*
1162 	 * we don't fault-ahead for device pager
1163 	 */
1164 	if (object->type == OBJT_DEVICE) {
1165 		*reqpage = 0;
1166 		marray[0] = m;
1167 		return 1;
1168 	}
1169 
1170 	/*
1171 	 * if the requested page is not available, then give up now
1172 	 */
1173 
1174 	if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1175 		return 0;
1176 	}
1177 
1178 	if ((cbehind == 0) && (cahead == 0)) {
1179 		*reqpage = 0;
1180 		marray[0] = m;
1181 		return 1;
1182 	}
1183 
1184 	if (rahead > cahead) {
1185 		rahead = cahead;
1186 	}
1187 
1188 	if (rbehind > cbehind) {
1189 		rbehind = cbehind;
1190 	}
1191 
1192 	/*
1193 	 * try to do any readahead that we might have free pages for.
1194 	 */
1195 	if ((rahead + rbehind) >
1196 		((cnt.v_free_count + cnt.v_cache_count) - cnt.v_free_reserved)) {
1197 		pagedaemon_wakeup();
1198 		marray[0] = m;
1199 		*reqpage = 0;
1200 		return 1;
1201 	}
1202 
1203 	/*
1204 	 * scan backward for the read behind pages -- in memory
1205 	 */
1206 	if (pindex > 0) {
1207 		if (rbehind > pindex) {
1208 			rbehind = pindex;
1209 			startpindex = 0;
1210 		} else {
1211 			startpindex = pindex - rbehind;
1212 		}
1213 
1214 		for ( tpindex = pindex - 1; tpindex >= startpindex; tpindex -= 1) {
1215 			if (vm_page_lookup( object, tpindex)) {
1216 				startpindex = tpindex + 1;
1217 				break;
1218 			}
1219 			if (tpindex == 0)
1220 				break;
1221 		}
1222 
1223 		for(i = 0, tpindex = startpindex; tpindex < pindex; i++, tpindex++) {
1224 
1225 			rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
1226 			if (rtm == NULL) {
1227 				for (j = 0; j < i; j++) {
1228 					vm_page_free(marray[j]);
1229 				}
1230 				marray[0] = m;
1231 				*reqpage = 0;
1232 				return 1;
1233 			}
1234 
1235 			marray[i] = rtm;
1236 		}
1237 	} else {
1238 		startpindex = 0;
1239 		i = 0;
1240 	}
1241 
1242 	marray[i] = m;
1243 	/* page offset of the required page */
1244 	*reqpage = i;
1245 
1246 	tpindex = pindex + 1;
1247 	i++;
1248 
1249 	/*
1250 	 * scan forward for the read ahead pages
1251 	 */
1252 	endpindex = tpindex + rahead;
1253 	if (endpindex > object->size)
1254 		endpindex = object->size;
1255 
1256 	for( ; tpindex < endpindex; i++, tpindex++) {
1257 
1258 		if (vm_page_lookup(object, tpindex)) {
1259 			break;
1260 		}
1261 
1262 		rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
1263 		if (rtm == NULL) {
1264 			break;
1265 		}
1266 
1267 		marray[i] = rtm;
1268 	}
1269 
1270 	/* return number of bytes of pages */
1271 	return i;
1272 }
1273