xref: /freebsd/sys/vm/vm_fault.c (revision 3ff369fed2a08f32dda232c10470b949bef9489f)
1 /*
2  * Copyright (c) 1991, 1993
3  *	The Regents of the University of California.  All rights reserved.
4  * Copyright (c) 1994 John S. Dyson
5  * All rights reserved.
6  * Copyright (c) 1994 David Greenman
7  * All rights reserved.
8  *
9  *
10  * This code is derived from software contributed to Berkeley by
11  * The Mach Operating System project at Carnegie-Mellon University.
12  *
13  * Redistribution and use in source and binary forms, with or without
14  * modification, are permitted provided that the following conditions
15  * are met:
16  * 1. Redistributions of source code must retain the above copyright
17  *    notice, this list of conditions and the following disclaimer.
18  * 2. Redistributions in binary form must reproduce the above copyright
19  *    notice, this list of conditions and the following disclaimer in the
20  *    documentation and/or other materials provided with the distribution.
21  * 3. All advertising materials mentioning features or use of this software
22  *    must display the following acknowledgement:
23  *	This product includes software developed by the University of
24  *	California, Berkeley and its contributors.
25  * 4. Neither the name of the University nor the names of its contributors
26  *    may be used to endorse or promote products derived from this software
27  *    without specific prior written permission.
28  *
29  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
30  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
31  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
32  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
33  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
34  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
35  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
36  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
37  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
38  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
39  * SUCH DAMAGE.
40  *
41  *	from: @(#)vm_fault.c	8.4 (Berkeley) 1/12/94
42  *
43  *
44  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
45  * All rights reserved.
46  *
47  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
48  *
49  * Permission to use, copy, modify and distribute this software and
50  * its documentation is hereby granted, provided that both the copyright
51  * notice and this permission notice appear in all copies of the
52  * software, derivative works or modified versions, and any portions
53  * thereof, and that both notices appear in supporting documentation.
54  *
55  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
56  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
57  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
58  *
59  * Carnegie Mellon requests users of this software to return to
60  *
61  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
62  *  School of Computer Science
63  *  Carnegie Mellon University
64  *  Pittsburgh PA 15213-3890
65  *
66  * any improvements or extensions that they make and grant Carnegie the
67  * rights to redistribute these changes.
68  *
69  * $FreeBSD$
70  */
71 
72 /*
73  *	Page fault handling module.
74  */
75 #include <sys/param.h>
76 #include <sys/systm.h>
77 #include <sys/kernel.h>
78 #include <sys/lock.h>
79 #include <sys/mutex.h>
80 #include <sys/proc.h>
81 #include <sys/resourcevar.h>
82 #include <sys/sysctl.h>
83 #include <sys/vmmeter.h>
84 #include <sys/vnode.h>
85 
86 #include <vm/vm.h>
87 #include <vm/vm_param.h>
88 #include <vm/pmap.h>
89 #include <vm/vm_map.h>
90 #include <vm/vm_object.h>
91 #include <vm/vm_page.h>
92 #include <vm/vm_pageout.h>
93 #include <vm/vm_kern.h>
94 #include <vm/vm_pager.h>
95 #include <vm/vnode_pager.h>
96 #include <vm/vm_extern.h>
97 
98 static int vm_fault_additional_pages(vm_page_t, int, int, vm_page_t *, int *);
99 
100 #define VM_FAULT_READ_AHEAD 8
101 #define VM_FAULT_READ_BEHIND 7
102 #define VM_FAULT_READ (VM_FAULT_READ_AHEAD+VM_FAULT_READ_BEHIND+1)
103 
104 struct faultstate {
105 	vm_page_t m;
106 	vm_object_t object;
107 	vm_pindex_t pindex;
108 	vm_page_t first_m;
109 	vm_object_t	first_object;
110 	vm_pindex_t first_pindex;
111 	vm_map_t map;
112 	vm_map_entry_t entry;
113 	int lookup_still_valid;
114 	struct vnode *vp;
115 };
116 
117 static __inline void
118 release_page(struct faultstate *fs)
119 {
120 	vm_page_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 	GIANT_REQUIRED;
138 	vm_object_pip_wakeup(fs->object);
139 	if (fs->object != fs->first_object) {
140 		vm_page_free(fs->first_m);
141 		vm_object_pip_wakeup(fs->first_object);
142 		fs->first_m = NULL;
143 	}
144 	if (dealloc) {
145 		vm_object_deallocate(fs->first_object);
146 	}
147 	unlock_map(fs);
148 	if (fs->vp != NULL) {
149 		vput(fs->vp);
150 		fs->vp = NULL;
151 	}
152 }
153 
154 #define unlock_things(fs) _unlock_things(fs, 0)
155 #define unlock_and_deallocate(fs) _unlock_things(fs, 1)
156 
157 /*
158  * TRYPAGER - used by vm_fault to calculate whether the pager for the
159  *	      current object *might* contain the page.
160  *
161  *	      default objects are zero-fill, there is no real pager.
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,
186 	 int fault_flags)
187 {
188 	vm_prot_t prot;
189 	int result;
190 	boolean_t growstack, wired;
191 	int map_generation;
192 	vm_object_t next_object;
193 	vm_page_t marray[VM_FAULT_READ];
194 	int hardfault;
195 	int faultcount;
196 	struct faultstate fs;
197 
198 	hardfault = 0;
199 	growstack = TRUE;
200 	atomic_add_int(&cnt.v_vm_faults, 1);
201 
202 	mtx_lock(&Giant);
203 RetryFault:;
204 
205 	/*
206 	 * Find the backing store object and offset into it to begin the
207 	 * search.
208 	 */
209 	fs.map = map;
210 	result = vm_map_lookup(&fs.map, vaddr, fault_type, &fs.entry,
211 	    &fs.first_object, &fs.first_pindex, &prot, &wired);
212 	if (result != KERN_SUCCESS) {
213 		if (result != KERN_PROTECTION_FAILURE ||
214 		    (fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE) {
215 			if (growstack && result == KERN_INVALID_ADDRESS &&
216 			    map != kernel_map && curproc != NULL) {
217 				result = vm_map_growstack(curproc, vaddr);
218 				if (result != KERN_SUCCESS) {
219 					mtx_unlock(&Giant);
220 					return (KERN_FAILURE);
221 				}
222 				growstack = FALSE;
223 				goto RetryFault;
224 			}
225 			mtx_unlock(&Giant);
226 			return (result);
227 		}
228 
229 		/*
230    		 * If we are user-wiring a r/w segment, and it is COW, then
231    		 * we need to do the COW operation.  Note that we don't COW
232    		 * currently RO sections now, because it is NOT desirable
233    		 * to COW .text.  We simply keep .text from ever being COW'ed
234    		 * and take the heat that one cannot debug wired .text sections.
235    		 */
236 		result = vm_map_lookup(&fs.map, vaddr,
237 			VM_PROT_READ|VM_PROT_WRITE|VM_PROT_OVERRIDE_WRITE,
238 			&fs.entry, &fs.first_object, &fs.first_pindex, &prot, &wired);
239 		if (result != KERN_SUCCESS) {
240 			mtx_unlock(&Giant);
241 			return (result);
242 		}
243 
244 		/*
245 		 * If we don't COW now, on a user wire, the user will never
246 		 * be able to write to the mapping.  If we don't make this
247 		 * restriction, the bookkeeping would be nearly impossible.
248 		 *
249 		 * XXX The following assignment modifies the map without
250 		 * holding a write lock on it.
251 		 */
252 		if ((fs.entry->protection & VM_PROT_WRITE) == 0)
253 			fs.entry->max_protection &= ~VM_PROT_WRITE;
254 	}
255 
256 	map_generation = fs.map->timestamp;
257 
258 	if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
259 		panic("vm_fault: fault on nofault entry, addr: %lx",
260 		    (u_long)vaddr);
261 	}
262 
263 	/*
264 	 * Make a reference to this object to prevent its disposal while we
265 	 * are messing with it.  Once we have the reference, the map is free
266 	 * to be diddled.  Since objects reference their shadows (and copies),
267 	 * they will stay around as well.
268 	 *
269 	 * Bump the paging-in-progress count to prevent size changes (e.g.
270 	 * truncation operations) during I/O.  This must be done after
271 	 * obtaining the vnode lock in order to avoid possible deadlocks.
272 	 *
273 	 * XXX vnode_pager_lock() can block without releasing the map lock.
274 	 */
275 	vm_object_reference(fs.first_object);
276 	fs.vp = vnode_pager_lock(fs.first_object);
277 	vm_object_pip_add(fs.first_object, 1);
278 
279 #ifdef ENABLE_VFS_IOOPT
280 	if ((fault_type & VM_PROT_WRITE) &&
281 		(fs.first_object->type == OBJT_VNODE)) {
282 		vm_freeze_copyopts(fs.first_object,
283 			fs.first_pindex, fs.first_pindex + 1);
284 	}
285 #endif
286 	fs.lookup_still_valid = TRUE;
287 
288 	if (wired)
289 		fault_type = prot;
290 
291 	fs.first_m = NULL;
292 
293 	/*
294 	 * Search for the page at object/offset.
295 	 */
296 	fs.object = fs.first_object;
297 	fs.pindex = fs.first_pindex;
298 	while (TRUE) {
299 		/*
300 		 * If the object is dead, we stop here
301 		 */
302 		if (fs.object->flags & OBJ_DEAD) {
303 			unlock_and_deallocate(&fs);
304 			mtx_unlock(&Giant);
305 			return (KERN_PROTECTION_FAILURE);
306 		}
307 
308 		/*
309 		 * See if page is resident
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 			queue = fs.m->queue;
338 
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_WAITPFAULT;
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 			vm_page_busy(fs.m);
357 			if (((fs.m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) &&
358 				fs.m->object != kernel_object && fs.m->object != kmem_object) {
359 				goto readrest;
360 			}
361 
362 			break;
363 		}
364 
365 		/*
366 		 * Page is not resident, If this is the search termination
367 		 * or the pager might contain the page, allocate a new page.
368 		 */
369 		if (TRYPAGER || fs.object == fs.first_object) {
370 			if (fs.pindex >= fs.object->size) {
371 				unlock_and_deallocate(&fs);
372 				mtx_unlock(&Giant);
373 				return (KERN_PROTECTION_FAILURE);
374 			}
375 
376 			/*
377 			 * Allocate a new page for this object/offset pair.
378 			 */
379 			fs.m = NULL;
380 			if (!vm_page_count_severe()) {
381 				fs.m = vm_page_alloc(fs.object, fs.pindex,
382 				    (fs.vp || fs.object->backing_object)? VM_ALLOC_NORMAL: VM_ALLOC_ZERO);
383 			}
384 			if (fs.m == NULL) {
385 				unlock_and_deallocate(&fs);
386 				VM_WAITPFAULT;
387 				goto RetryFault;
388 			}
389 		}
390 
391 readrest:
392 		/*
393 		 * We have found a valid page or we have allocated a new page.
394 		 * The page thus may not be valid or may not be entirely
395 		 * valid.
396 		 *
397 		 * Attempt to fault-in the page if there is a chance that the
398 		 * pager has it, and potentially fault in additional pages
399 		 * at the same time.
400 		 */
401 		if (TRYPAGER) {
402 			int rv;
403 			int reqpage;
404 			int ahead, behind;
405 			u_char behavior = vm_map_entry_behavior(fs.entry);
406 
407 			if (behavior == MAP_ENTRY_BEHAV_RANDOM) {
408 				ahead = 0;
409 				behind = 0;
410 			} else {
411 				behind = (vaddr - fs.entry->start) >> PAGE_SHIFT;
412 				if (behind > VM_FAULT_READ_BEHIND)
413 					behind = VM_FAULT_READ_BEHIND;
414 
415 				ahead = ((fs.entry->end - vaddr) >> PAGE_SHIFT) - 1;
416 				if (ahead > VM_FAULT_READ_AHEAD)
417 					ahead = VM_FAULT_READ_AHEAD;
418 			}
419 
420 			if ((fs.first_object->type != OBJT_DEVICE) &&
421 			    (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL ||
422                                 (behavior != MAP_ENTRY_BEHAV_RANDOM &&
423                                 fs.pindex >= fs.entry->lastr &&
424                                 fs.pindex < fs.entry->lastr + VM_FAULT_READ))
425 			) {
426 				vm_pindex_t firstpindex, tmppindex;
427 
428 				if (fs.first_pindex < 2 * VM_FAULT_READ)
429 					firstpindex = 0;
430 				else
431 					firstpindex = fs.first_pindex - 2 * VM_FAULT_READ;
432 
433 				/*
434 				 * note: partially valid pages cannot be
435 				 * included in the lookahead - NFS piecemeal
436 				 * writes will barf on it badly.
437 				 */
438 				for (tmppindex = fs.first_pindex - 1;
439 					tmppindex >= firstpindex;
440 					--tmppindex) {
441 					vm_page_t mt;
442 
443 					mt = vm_page_lookup(fs.first_object, tmppindex);
444 					if (mt == NULL || (mt->valid != VM_PAGE_BITS_ALL))
445 						break;
446 					if (mt->busy ||
447 						(mt->flags & (PG_BUSY | PG_FICTITIOUS | PG_UNMANAGED)) ||
448 						mt->hold_count ||
449 						mt->wire_count)
450 						continue;
451 					if (mt->dirty == 0)
452 						vm_page_test_dirty(mt);
453 					if (mt->dirty) {
454 						vm_page_protect(mt, VM_PROT_NONE);
455 						vm_page_deactivate(mt);
456 					} else {
457 						vm_page_cache(mt);
458 					}
459 				}
460 
461 				ahead += behind;
462 				behind = 0;
463 			}
464 
465 			/*
466 			 * now we find out if any other pages should be paged
467 			 * in at this time this routine checks to see if the
468 			 * pages surrounding this fault reside in the same
469 			 * object as the page for this fault.  If they do,
470 			 * then they are faulted in also into the object.  The
471 			 * array "marray" returned contains an array of
472 			 * vm_page_t structs where one of them is the
473 			 * vm_page_t passed to the routine.  The reqpage
474 			 * return value is the index into the marray for the
475 			 * vm_page_t passed to the routine.
476 			 *
477 			 * fs.m plus the additional pages are PG_BUSY'd.
478 			 *
479 			 * XXX vm_fault_additional_pages() can block
480 			 * without releasing the map lock.
481 			 */
482 			faultcount = vm_fault_additional_pages(
483 			    fs.m, behind, ahead, marray, &reqpage);
484 
485 			/*
486 			 * update lastr imperfectly (we do not know how much
487 			 * getpages will actually read), but good enough.
488 			 *
489 			 * XXX The following assignment modifies the map
490 			 * without holding a write lock on it.
491 			 */
492 			fs.entry->lastr = fs.pindex + faultcount - behind;
493 
494 			/*
495 			 * Call the pager to retrieve the data, if any, after
496 			 * releasing the lock on the map.  We hold a ref on
497 			 * fs.object and the pages are PG_BUSY'd.
498 			 */
499 			unlock_map(&fs);
500 
501 			rv = faultcount ?
502 			    vm_pager_get_pages(fs.object, marray, faultcount,
503 				reqpage) : VM_PAGER_FAIL;
504 
505 			if (rv == VM_PAGER_OK) {
506 				/*
507 				 * Found the page. Leave it busy while we play
508 				 * with it.
509 				 */
510 
511 				/*
512 				 * Relookup in case pager changed page. Pager
513 				 * is responsible for disposition of old page
514 				 * if moved.
515 				 */
516 				fs.m = vm_page_lookup(fs.object, fs.pindex);
517 				if (!fs.m) {
518 					unlock_and_deallocate(&fs);
519 					goto RetryFault;
520 				}
521 
522 				hardfault++;
523 				break; /* break to PAGE HAS BEEN FOUND */
524 			}
525 			/*
526 			 * Remove the bogus page (which does not exist at this
527 			 * object/offset); before doing so, we must get back
528 			 * our object lock to preserve our invariant.
529 			 *
530 			 * Also wake up any other process that may want to bring
531 			 * in this page.
532 			 *
533 			 * If this is the top-level object, we must leave the
534 			 * busy page to prevent another process from rushing
535 			 * past us, and inserting the page in that object at
536 			 * the same time that we are.
537 			 */
538 			if (rv == VM_PAGER_ERROR)
539 				printf("vm_fault: pager read error, pid %d (%s)\n",
540 				    curproc->p_pid, curproc->p_comm);
541 			/*
542 			 * Data outside the range of the pager or an I/O error
543 			 */
544 			/*
545 			 * XXX - the check for kernel_map is a kludge to work
546 			 * around having the machine panic on a kernel space
547 			 * fault w/ I/O error.
548 			 */
549 			if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) ||
550 				(rv == VM_PAGER_BAD)) {
551 				vm_page_free(fs.m);
552 				fs.m = NULL;
553 				unlock_and_deallocate(&fs);
554 				mtx_unlock(&Giant);
555 				return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
556 			}
557 			if (fs.object != fs.first_object) {
558 				vm_page_free(fs.m);
559 				fs.m = NULL;
560 				/*
561 				 * XXX - we cannot just fall out at this
562 				 * point, m has been freed and is invalid!
563 				 */
564 			}
565 		}
566 
567 		/*
568 		 * We get here if the object has default pager (or unwiring)
569 		 * or the pager doesn't have the page.
570 		 */
571 		if (fs.object == fs.first_object)
572 			fs.first_m = fs.m;
573 
574 		/*
575 		 * Move on to the next object.  Lock the next object before
576 		 * unlocking the current one.
577 		 */
578 		fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset);
579 		next_object = fs.object->backing_object;
580 		if (next_object == NULL) {
581 			/*
582 			 * If there's no object left, fill the page in the top
583 			 * object with zeros.
584 			 */
585 			if (fs.object != fs.first_object) {
586 				vm_object_pip_wakeup(fs.object);
587 
588 				fs.object = fs.first_object;
589 				fs.pindex = fs.first_pindex;
590 				fs.m = fs.first_m;
591 			}
592 			fs.first_m = NULL;
593 
594 			/*
595 			 * Zero the page if necessary and mark it valid.
596 			 */
597 			if ((fs.m->flags & PG_ZERO) == 0) {
598 				vm_page_zero_fill(fs.m);
599 			} else {
600 				cnt.v_ozfod++;
601 			}
602 			cnt.v_zfod++;
603 			fs.m->valid = VM_PAGE_BITS_ALL;
604 			break;	/* break to PAGE HAS BEEN FOUND */
605 		} else {
606 			if (fs.object != fs.first_object) {
607 				vm_object_pip_wakeup(fs.object);
608 			}
609 			KASSERT(fs.object != next_object, ("object loop %p", next_object));
610 			fs.object = next_object;
611 			vm_object_pip_add(fs.object, 1);
612 		}
613 	}
614 
615 	KASSERT((fs.m->flags & PG_BUSY) != 0,
616 	    ("vm_fault: not busy after main loop"));
617 
618 	/*
619 	 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
620 	 * is held.]
621 	 */
622 
623 	/*
624 	 * If the page is being written, but isn't already owned by the
625 	 * top-level object, we have to copy it into a new page owned by the
626 	 * top-level object.
627 	 */
628 	if (fs.object != fs.first_object) {
629 		/*
630 		 * We only really need to copy if we want to write it.
631 		 */
632 		if (fault_type & VM_PROT_WRITE) {
633 			/*
634 			 * This allows pages to be virtually copied from a
635 			 * backing_object into the first_object, where the
636 			 * backing object has no other refs to it, and cannot
637 			 * gain any more refs.  Instead of a bcopy, we just
638 			 * move the page from the backing object to the
639 			 * first object.  Note that we must mark the page
640 			 * dirty in the first object so that it will go out
641 			 * to swap when needed.
642 			 */
643 			if (map_generation == fs.map->timestamp &&
644 				/*
645 				 * Only one shadow object
646 				 */
647 				(fs.object->shadow_count == 1) &&
648 				/*
649 				 * No COW refs, except us
650 				 */
651 				(fs.object->ref_count == 1) &&
652 				/*
653 				 * No one else can look this object up
654 				 */
655 				(fs.object->handle == NULL) &&
656 				/*
657 				 * No other ways to look the object up
658 				 */
659 				((fs.object->type == OBJT_DEFAULT) ||
660 				 (fs.object->type == OBJT_SWAP)) &&
661 				/*
662 				 * We don't chase down the shadow chain
663 				 */
664 				(fs.object == fs.first_object->backing_object) &&
665 
666 				/*
667 				 * grab the lock if we need to
668 				 */
669 			    (fs.lookup_still_valid || vm_map_trylock(fs.map))) {
670 
671 				fs.lookup_still_valid = 1;
672 				/*
673 				 * get rid of the unnecessary page
674 				 */
675 				vm_page_protect(fs.first_m, VM_PROT_NONE);
676 				vm_page_free(fs.first_m);
677 				fs.first_m = NULL;
678 
679 				/*
680 				 * grab the page and put it into the
681 				 * process'es object.  The page is
682 				 * automatically made dirty.
683 				 */
684 				vm_page_rename(fs.m, fs.first_object, fs.first_pindex);
685 				fs.first_m = fs.m;
686 				vm_page_busy(fs.first_m);
687 				fs.m = NULL;
688 				cnt.v_cow_optim++;
689 			} else {
690 				/*
691 				 * Oh, well, lets copy it.
692 				 */
693 				vm_page_copy(fs.m, fs.first_m);
694 			}
695 
696 			if (fs.m) {
697 				/*
698 				 * We no longer need the old page or object.
699 				 */
700 				release_page(&fs);
701 			}
702 
703 			/*
704 			 * fs.object != fs.first_object due to above
705 			 * conditional
706 			 */
707 			vm_object_pip_wakeup(fs.object);
708 
709 			/*
710 			 * Only use the new page below...
711 			 */
712 			cnt.v_cow_faults++;
713 			fs.m = fs.first_m;
714 			fs.object = fs.first_object;
715 			fs.pindex = fs.first_pindex;
716 
717 		} else {
718 			prot &= ~VM_PROT_WRITE;
719 		}
720 	}
721 
722 	/*
723 	 * We must verify that the maps have not changed since our last
724 	 * lookup.
725 	 */
726 	if (!fs.lookup_still_valid &&
727 		(fs.map->timestamp != map_generation)) {
728 		vm_object_t retry_object;
729 		vm_pindex_t retry_pindex;
730 		vm_prot_t retry_prot;
731 
732 		/*
733 		 * Since map entries may be pageable, make sure we can take a
734 		 * page fault on them.
735 		 */
736 
737 		/*
738 		 * Unlock vnode before the lookup to avoid deadlock.   E.G.
739 		 * avoid a deadlock between the inode and exec_map that can
740 		 * occur due to locks being obtained in different orders.
741 		 */
742 		if (fs.vp != NULL) {
743 			vput(fs.vp);
744 			fs.vp = NULL;
745 		}
746 
747 		if (fs.map->infork) {
748 			release_page(&fs);
749 			unlock_and_deallocate(&fs);
750 			goto RetryFault;
751 		}
752 
753 		/*
754 		 * To avoid trying to write_lock the map while another process
755 		 * has it read_locked (in vm_map_pageable), we do not try for
756 		 * write permission.  If the page is still writable, we will
757 		 * get write permission.  If it is not, or has been marked
758 		 * needs_copy, we enter the mapping without write permission,
759 		 * and will merely take another fault.
760 		 */
761 		result = vm_map_lookup(&fs.map, vaddr, fault_type & ~VM_PROT_WRITE,
762 		    &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
763 		map_generation = fs.map->timestamp;
764 
765 		/*
766 		 * If we don't need the page any longer, put it on the active
767 		 * list (the easiest thing to do here).  If no one needs it,
768 		 * pageout will grab it eventually.
769 		 */
770 		if (result != KERN_SUCCESS) {
771 			release_page(&fs);
772 			unlock_and_deallocate(&fs);
773 			mtx_unlock(&Giant);
774 			return (result);
775 		}
776 		fs.lookup_still_valid = TRUE;
777 
778 		if ((retry_object != fs.first_object) ||
779 		    (retry_pindex != fs.first_pindex)) {
780 			release_page(&fs);
781 			unlock_and_deallocate(&fs);
782 			goto RetryFault;
783 		}
784 		/*
785 		 * Check whether the protection has changed or the object has
786 		 * been copied while we left the map unlocked. Changing from
787 		 * read to write permission is OK - we leave the page
788 		 * write-protected, and catch the write fault. Changing from
789 		 * write to read permission means that we can't mark the page
790 		 * write-enabled after all.
791 		 */
792 		prot &= retry_prot;
793 	}
794 
795 	/*
796 	 * Put this page into the physical map. We had to do the unlock above
797 	 * because pmap_enter may cause other faults.   We don't put the page
798 	 * back on the active queue until later so that the page-out daemon
799 	 * won't find us (yet).
800 	 */
801 
802 	if (prot & VM_PROT_WRITE) {
803 		vm_page_flag_set(fs.m, PG_WRITEABLE);
804 		vm_object_set_writeable_dirty(fs.m->object);
805 
806 		/*
807 		 * If the fault is a write, we know that this page is being
808 		 * written NOW so dirty it explicitly to save on
809 		 * pmap_is_modified() calls later.
810 		 *
811 		 * If this is a NOSYNC mmap we do not want to set PG_NOSYNC
812 		 * if the page is already dirty to prevent data written with
813 		 * the expectation of being synced from not being synced.
814 		 * Likewise if this entry does not request NOSYNC then make
815 		 * sure the page isn't marked NOSYNC.  Applications sharing
816 		 * data should use the same flags to avoid ping ponging.
817 		 *
818 		 * Also tell the backing pager, if any, that it should remove
819 		 * any swap backing since the page is now dirty.
820 		 */
821 		if (fs.entry->eflags & MAP_ENTRY_NOSYNC) {
822 			if (fs.m->dirty == 0)
823 				vm_page_flag_set(fs.m, PG_NOSYNC);
824 		} else {
825 			vm_page_flag_clear(fs.m, PG_NOSYNC);
826 		}
827 		if (fault_flags & VM_FAULT_DIRTY) {
828 			int s;
829 			vm_page_dirty(fs.m);
830 			s = splvm();
831 			vm_pager_page_unswapped(fs.m);
832 			splx(s);
833 		}
834 	}
835 
836 	/*
837 	 * Page had better still be busy
838 	 */
839 	KASSERT(fs.m->flags & PG_BUSY,
840 		("vm_fault: page %p not busy!", fs.m));
841 	unlock_things(&fs);
842 
843 	/*
844 	 * Sanity check: page must be completely valid or it is not fit to
845 	 * map into user space.  vm_pager_get_pages() ensures this.
846 	 */
847 	if (fs.m->valid != VM_PAGE_BITS_ALL) {
848 		vm_page_zero_invalid(fs.m, TRUE);
849 		printf("Warning: page %p partially invalid on fault\n", fs.m);
850 	}
851 	pmap_enter(fs.map->pmap, vaddr, fs.m, prot, wired);
852 	if (((fault_flags & VM_FAULT_WIRE_MASK) == 0) && (wired == 0)) {
853 		pmap_prefault(fs.map->pmap, vaddr, fs.entry);
854 	}
855 	vm_page_flag_clear(fs.m, PG_ZERO);
856 	vm_page_flag_set(fs.m, PG_MAPPED|PG_REFERENCED);
857 
858 	/*
859 	 * If the page is not wired down, then put it where the pageout daemon
860 	 * can find it.
861 	 */
862 	if (fault_flags & VM_FAULT_WIRE_MASK) {
863 		if (wired)
864 			vm_page_wire(fs.m);
865 		else
866 			vm_page_unwire(fs.m, 1);
867 	} else {
868 		vm_page_activate(fs.m);
869 	}
870 
871 	mtx_lock_spin(&sched_lock);
872 	if (curproc && (curproc->p_sflag & PS_INMEM) && curproc->p_stats) {
873 		if (hardfault) {
874 			curproc->p_stats->p_ru.ru_majflt++;
875 		} else {
876 			curproc->p_stats->p_ru.ru_minflt++;
877 		}
878 	}
879 	mtx_unlock_spin(&sched_lock);
880 
881 	/*
882 	 * Unlock everything, and return
883 	 */
884 	vm_page_wakeup(fs.m);
885 	vm_object_deallocate(fs.first_object);
886 	mtx_unlock(&Giant);
887 	return (KERN_SUCCESS);
888 }
889 
890 /*
891  *	vm_fault_wire:
892  *
893  *	Wire down a range of virtual addresses in a map.
894  */
895 int
896 vm_fault_wire(map, start, end)
897 	vm_map_t map;
898 	vm_offset_t start, end;
899 {
900 
901 	vm_offset_t va;
902 	pmap_t pmap;
903 	int rv;
904 
905 	pmap = vm_map_pmap(map);
906 
907 	/*
908 	 * Inform the physical mapping system that the range of addresses may
909 	 * not fault, so that page tables and such can be locked down as well.
910 	 */
911 	pmap_pageable(pmap, start, end, FALSE);
912 
913 	/*
914 	 * We simulate a fault to get the page and enter it in the physical
915 	 * map.
916 	 */
917 	for (va = start; va < end; va += PAGE_SIZE) {
918 		rv = vm_fault(map, va, VM_PROT_READ|VM_PROT_WRITE,
919 			VM_FAULT_CHANGE_WIRING);
920 		if (rv) {
921 			if (va != start)
922 				vm_fault_unwire(map, start, va);
923 			return (rv);
924 		}
925 	}
926 	return (KERN_SUCCESS);
927 }
928 
929 /*
930  *	vm_fault_user_wire:
931  *
932  *	Wire down a range of virtual addresses in a map.  This
933  *	is for user mode though, so we only ask for read access
934  *	on currently read only sections.
935  */
936 int
937 vm_fault_user_wire(map, start, end)
938 	vm_map_t map;
939 	vm_offset_t start, end;
940 {
941 	vm_offset_t va;
942 	pmap_t pmap;
943 	int rv;
944 
945 	pmap = vm_map_pmap(map);
946 
947 	/*
948 	 * Inform the physical mapping system that the range of addresses may
949 	 * not fault, so that page tables and such can be locked down as well.
950 	 */
951 	pmap_pageable(pmap, start, end, FALSE);
952 
953 	/*
954 	 * We simulate a fault to get the page and enter it in the physical
955 	 * map.
956 	 */
957 	for (va = start; va < end; va += PAGE_SIZE) {
958 		rv = vm_fault(map, va, VM_PROT_READ, VM_FAULT_USER_WIRE);
959 		if (rv) {
960 			if (va != start)
961 				vm_fault_unwire(map, start, va);
962 			return (rv);
963 		}
964 	}
965 	return (KERN_SUCCESS);
966 }
967 
968 
969 /*
970  *	vm_fault_unwire:
971  *
972  *	Unwire a range of virtual addresses in a map.
973  */
974 void
975 vm_fault_unwire(map, start, end)
976 	vm_map_t map;
977 	vm_offset_t start, end;
978 {
979 	vm_offset_t va, pa;
980 	pmap_t pmap;
981 
982 	pmap = vm_map_pmap(map);
983 
984 	mtx_lock(&Giant);
985 	/*
986 	 * Since the pages are wired down, we must be able to get their
987 	 * mappings from the physical map system.
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 	mtx_unlock(&Giant);
997 
998 	/*
999 	 * Inform the physical mapping system that the range of addresses may
1000 	 * fault, so that page tables and such may be unwired themselves.
1001 	 */
1002 	pmap_pageable(pmap, start, end, TRUE);
1003 }
1004 
1005 /*
1006  *	Routine:
1007  *		vm_fault_copy_entry
1008  *	Function:
1009  *		Copy all of the pages from a wired-down map entry to another.
1010  *
1011  *	In/out conditions:
1012  *		The source and destination maps must be locked for write.
1013  *		The source map entry must be wired down (or be a sharing map
1014  *		entry corresponding to a main map entry that is wired down).
1015  */
1016 void
1017 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry)
1018 	vm_map_t dst_map;
1019 	vm_map_t src_map;
1020 	vm_map_entry_t dst_entry;
1021 	vm_map_entry_t src_entry;
1022 {
1023 	vm_object_t dst_object;
1024 	vm_object_t src_object;
1025 	vm_ooffset_t dst_offset;
1026 	vm_ooffset_t src_offset;
1027 	vm_prot_t prot;
1028 	vm_offset_t vaddr;
1029 	vm_page_t dst_m;
1030 	vm_page_t src_m;
1031 
1032 #ifdef	lint
1033 	src_map++;
1034 #endif	/* lint */
1035 
1036 	src_object = src_entry->object.vm_object;
1037 	src_offset = src_entry->offset;
1038 
1039 	/*
1040 	 * Create the top-level object for the destination entry. (Doesn't
1041 	 * actually shadow anything - we copy the pages directly.)
1042 	 */
1043 	dst_object = vm_object_allocate(OBJT_DEFAULT,
1044 	    (vm_size_t) OFF_TO_IDX(dst_entry->end - dst_entry->start));
1045 
1046 	dst_entry->object.vm_object = dst_object;
1047 	dst_entry->offset = 0;
1048 
1049 	prot = dst_entry->max_protection;
1050 
1051 	/*
1052 	 * Loop through all of the pages in the entry's range, copying each
1053 	 * one from the source object (it should be there) to the destination
1054 	 * object.
1055 	 */
1056 	for (vaddr = dst_entry->start, dst_offset = 0;
1057 	    vaddr < dst_entry->end;
1058 	    vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) {
1059 
1060 		/*
1061 		 * Allocate a page in the destination object
1062 		 */
1063 		do {
1064 			dst_m = vm_page_alloc(dst_object,
1065 				OFF_TO_IDX(dst_offset), VM_ALLOC_NORMAL);
1066 			if (dst_m == NULL) {
1067 				VM_WAIT;
1068 			}
1069 		} while (dst_m == NULL);
1070 
1071 		/*
1072 		 * Find the page in the source object, and copy it in.
1073 		 * (Because the source is wired down, the page will be in
1074 		 * memory.)
1075 		 */
1076 		src_m = vm_page_lookup(src_object,
1077 			OFF_TO_IDX(dst_offset + src_offset));
1078 		if (src_m == NULL)
1079 			panic("vm_fault_copy_wired: page missing");
1080 
1081 		vm_page_copy(src_m, dst_m);
1082 
1083 		/*
1084 		 * Enter it in the pmap...
1085 		 */
1086 		vm_page_flag_clear(dst_m, PG_ZERO);
1087 		pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE);
1088 		vm_page_flag_set(dst_m, PG_WRITEABLE|PG_MAPPED);
1089 
1090 		/*
1091 		 * Mark it no longer busy, and put it on the active list.
1092 		 */
1093 		vm_page_activate(dst_m);
1094 		vm_page_wakeup(dst_m);
1095 	}
1096 }
1097 
1098 
1099 /*
1100  * This routine checks around the requested page for other pages that
1101  * might be able to be faulted in.  This routine brackets the viable
1102  * pages for the pages to be paged in.
1103  *
1104  * Inputs:
1105  *	m, rbehind, rahead
1106  *
1107  * Outputs:
1108  *  marray (array of vm_page_t), reqpage (index of requested page)
1109  *
1110  * Return value:
1111  *  number of pages in marray
1112  *
1113  * This routine can't block.
1114  */
1115 static int
1116 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage)
1117 	vm_page_t m;
1118 	int rbehind;
1119 	int rahead;
1120 	vm_page_t *marray;
1121 	int *reqpage;
1122 {
1123 	int i,j;
1124 	vm_object_t object;
1125 	vm_pindex_t pindex, startpindex, endpindex, tpindex;
1126 	vm_page_t rtm;
1127 	int cbehind, cahead;
1128 
1129 	GIANT_REQUIRED;
1130 
1131 	object = m->object;
1132 	pindex = m->pindex;
1133 
1134 	/*
1135 	 * we don't fault-ahead for device pager
1136 	 */
1137 	if (object->type == OBJT_DEVICE) {
1138 		*reqpage = 0;
1139 		marray[0] = m;
1140 		return 1;
1141 	}
1142 
1143 	/*
1144 	 * if the requested page is not available, then give up now
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