xref: /freebsd/sys/vm/vm_fault.c (revision 0b87f79976047c8f4332bbf7dc03146f6b0de79f)
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 			/*
316 			 * check for page-based copy on write
317 			 */
318 
319 			if ((fs.m->cow) &&
320 			    (fault_type & VM_PROT_WRITE)) {
321 				s = splvm();
322 				vm_page_cowfault(fs.m);
323 				splx(s);
324 				unlock_things(&fs);
325 				goto RetryFault;
326 			}
327 
328 			/*
329 			 * Wait/Retry if the page is busy.  We have to do this
330 			 * if the page is busy via either PG_BUSY or
331 			 * vm_page_t->busy because the vm_pager may be using
332 			 * vm_page_t->busy for pageouts ( and even pageins if
333 			 * it is the vnode pager ), and we could end up trying
334 			 * to pagein and pageout the same page simultaneously.
335 			 *
336 			 * We can theoretically allow the busy case on a read
337 			 * fault if the page is marked valid, but since such
338 			 * pages are typically already pmap'd, putting that
339 			 * special case in might be more effort then it is
340 			 * worth.  We cannot under any circumstances mess
341 			 * around with a vm_page_t->busy page except, perhaps,
342 			 * to pmap it.
343 			 */
344 			if ((fs.m->flags & PG_BUSY) || fs.m->busy) {
345 				unlock_things(&fs);
346 				(void)vm_page_sleep_busy(fs.m, TRUE, "vmpfw");
347 				cnt.v_intrans++;
348 				vm_object_deallocate(fs.first_object);
349 				goto RetryFault;
350 			}
351 			queue = fs.m->queue;
352 
353 			s = splvm();
354 			vm_pageq_remove_nowakeup(fs.m);
355 			splx(s);
356 
357 			if ((queue - fs.m->pc) == PQ_CACHE && vm_page_count_severe()) {
358 				vm_page_activate(fs.m);
359 				unlock_and_deallocate(&fs);
360 				VM_WAITPFAULT;
361 				goto RetryFault;
362 			}
363 
364 			/*
365 			 * Mark page busy for other processes, and the
366 			 * pagedaemon.  If it still isn't completely valid
367 			 * (readable), jump to readrest, else break-out ( we
368 			 * found the page ).
369 			 */
370 			vm_page_busy(fs.m);
371 			if (((fs.m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) &&
372 				fs.m->object != kernel_object && fs.m->object != kmem_object) {
373 				goto readrest;
374 			}
375 
376 			break;
377 		}
378 
379 		/*
380 		 * Page is not resident, If this is the search termination
381 		 * or the pager might contain the page, allocate a new page.
382 		 */
383 		if (TRYPAGER || fs.object == fs.first_object) {
384 			if (fs.pindex >= fs.object->size) {
385 				unlock_and_deallocate(&fs);
386 				mtx_unlock(&Giant);
387 				return (KERN_PROTECTION_FAILURE);
388 			}
389 
390 			/*
391 			 * Allocate a new page for this object/offset pair.
392 			 */
393 			fs.m = NULL;
394 			if (!vm_page_count_severe()) {
395 				fs.m = vm_page_alloc(fs.object, fs.pindex,
396 				    (fs.vp || fs.object->backing_object)? VM_ALLOC_NORMAL: VM_ALLOC_ZERO);
397 			}
398 			if (fs.m == NULL) {
399 				unlock_and_deallocate(&fs);
400 				VM_WAITPFAULT;
401 				goto RetryFault;
402 			}
403 		}
404 
405 readrest:
406 		/*
407 		 * We have found a valid page or we have allocated a new page.
408 		 * The page thus may not be valid or may not be entirely
409 		 * valid.
410 		 *
411 		 * Attempt to fault-in the page if there is a chance that the
412 		 * pager has it, and potentially fault in additional pages
413 		 * at the same time.
414 		 */
415 		if (TRYPAGER) {
416 			int rv;
417 			int reqpage;
418 			int ahead, behind;
419 			u_char behavior = vm_map_entry_behavior(fs.entry);
420 
421 			if (behavior == MAP_ENTRY_BEHAV_RANDOM) {
422 				ahead = 0;
423 				behind = 0;
424 			} else {
425 				behind = (vaddr - fs.entry->start) >> PAGE_SHIFT;
426 				if (behind > VM_FAULT_READ_BEHIND)
427 					behind = VM_FAULT_READ_BEHIND;
428 
429 				ahead = ((fs.entry->end - vaddr) >> PAGE_SHIFT) - 1;
430 				if (ahead > VM_FAULT_READ_AHEAD)
431 					ahead = VM_FAULT_READ_AHEAD;
432 			}
433 
434 			if ((fs.first_object->type != OBJT_DEVICE) &&
435 			    (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL ||
436                                 (behavior != MAP_ENTRY_BEHAV_RANDOM &&
437                                 fs.pindex >= fs.entry->lastr &&
438                                 fs.pindex < fs.entry->lastr + VM_FAULT_READ))
439 			) {
440 				vm_pindex_t firstpindex, tmppindex;
441 
442 				if (fs.first_pindex < 2 * VM_FAULT_READ)
443 					firstpindex = 0;
444 				else
445 					firstpindex = fs.first_pindex - 2 * VM_FAULT_READ;
446 
447 				/*
448 				 * note: partially valid pages cannot be
449 				 * included in the lookahead - NFS piecemeal
450 				 * writes will barf on it badly.
451 				 */
452 				for (tmppindex = fs.first_pindex - 1;
453 					tmppindex >= firstpindex;
454 					--tmppindex) {
455 					vm_page_t mt;
456 
457 					mt = vm_page_lookup(fs.first_object, tmppindex);
458 					if (mt == NULL || (mt->valid != VM_PAGE_BITS_ALL))
459 						break;
460 					if (mt->busy ||
461 						(mt->flags & (PG_BUSY | PG_FICTITIOUS | PG_UNMANAGED)) ||
462 						mt->hold_count ||
463 						mt->wire_count)
464 						continue;
465 					if (mt->dirty == 0)
466 						vm_page_test_dirty(mt);
467 					if (mt->dirty) {
468 						vm_page_protect(mt, VM_PROT_NONE);
469 						vm_page_deactivate(mt);
470 					} else {
471 						vm_page_cache(mt);
472 					}
473 				}
474 
475 				ahead += behind;
476 				behind = 0;
477 			}
478 
479 			/*
480 			 * now we find out if any other pages should be paged
481 			 * in at this time this routine checks to see if the
482 			 * pages surrounding this fault reside in the same
483 			 * object as the page for this fault.  If they do,
484 			 * then they are faulted in also into the object.  The
485 			 * array "marray" returned contains an array of
486 			 * vm_page_t structs where one of them is the
487 			 * vm_page_t passed to the routine.  The reqpage
488 			 * return value is the index into the marray for the
489 			 * vm_page_t passed to the routine.
490 			 *
491 			 * fs.m plus the additional pages are PG_BUSY'd.
492 			 *
493 			 * XXX vm_fault_additional_pages() can block
494 			 * without releasing the map lock.
495 			 */
496 			faultcount = vm_fault_additional_pages(
497 			    fs.m, behind, ahead, marray, &reqpage);
498 
499 			/*
500 			 * update lastr imperfectly (we do not know how much
501 			 * getpages will actually read), but good enough.
502 			 *
503 			 * XXX The following assignment modifies the map
504 			 * without holding a write lock on it.
505 			 */
506 			fs.entry->lastr = fs.pindex + faultcount - behind;
507 
508 			/*
509 			 * Call the pager to retrieve the data, if any, after
510 			 * releasing the lock on the map.  We hold a ref on
511 			 * fs.object and the pages are PG_BUSY'd.
512 			 */
513 			unlock_map(&fs);
514 
515 			rv = faultcount ?
516 			    vm_pager_get_pages(fs.object, marray, faultcount,
517 				reqpage) : VM_PAGER_FAIL;
518 
519 			if (rv == VM_PAGER_OK) {
520 				/*
521 				 * Found the page. Leave it busy while we play
522 				 * with it.
523 				 */
524 
525 				/*
526 				 * Relookup in case pager changed page. Pager
527 				 * is responsible for disposition of old page
528 				 * if moved.
529 				 */
530 				fs.m = vm_page_lookup(fs.object, fs.pindex);
531 				if (!fs.m) {
532 					unlock_and_deallocate(&fs);
533 					goto RetryFault;
534 				}
535 
536 				hardfault++;
537 				break; /* break to PAGE HAS BEEN FOUND */
538 			}
539 			/*
540 			 * Remove the bogus page (which does not exist at this
541 			 * object/offset); before doing so, we must get back
542 			 * our object lock to preserve our invariant.
543 			 *
544 			 * Also wake up any other process that may want to bring
545 			 * in this page.
546 			 *
547 			 * If this is the top-level object, we must leave the
548 			 * busy page to prevent another process from rushing
549 			 * past us, and inserting the page in that object at
550 			 * the same time that we are.
551 			 */
552 			if (rv == VM_PAGER_ERROR)
553 				printf("vm_fault: pager read error, pid %d (%s)\n",
554 				    curproc->p_pid, curproc->p_comm);
555 			/*
556 			 * Data outside the range of the pager or an I/O error
557 			 */
558 			/*
559 			 * XXX - the check for kernel_map is a kludge to work
560 			 * around having the machine panic on a kernel space
561 			 * fault w/ I/O error.
562 			 */
563 			if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) ||
564 				(rv == VM_PAGER_BAD)) {
565 				vm_page_free(fs.m);
566 				fs.m = NULL;
567 				unlock_and_deallocate(&fs);
568 				mtx_unlock(&Giant);
569 				return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
570 			}
571 			if (fs.object != fs.first_object) {
572 				vm_page_free(fs.m);
573 				fs.m = NULL;
574 				/*
575 				 * XXX - we cannot just fall out at this
576 				 * point, m has been freed and is invalid!
577 				 */
578 			}
579 		}
580 
581 		/*
582 		 * We get here if the object has default pager (or unwiring)
583 		 * or the pager doesn't have the page.
584 		 */
585 		if (fs.object == fs.first_object)
586 			fs.first_m = fs.m;
587 
588 		/*
589 		 * Move on to the next object.  Lock the next object before
590 		 * unlocking the current one.
591 		 */
592 		fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset);
593 		next_object = fs.object->backing_object;
594 		if (next_object == NULL) {
595 			/*
596 			 * If there's no object left, fill the page in the top
597 			 * object with zeros.
598 			 */
599 			if (fs.object != fs.first_object) {
600 				vm_object_pip_wakeup(fs.object);
601 
602 				fs.object = fs.first_object;
603 				fs.pindex = fs.first_pindex;
604 				fs.m = fs.first_m;
605 			}
606 			fs.first_m = NULL;
607 
608 			/*
609 			 * Zero the page if necessary and mark it valid.
610 			 */
611 			if ((fs.m->flags & PG_ZERO) == 0) {
612 				vm_page_zero_fill(fs.m);
613 			} else {
614 				cnt.v_ozfod++;
615 			}
616 			cnt.v_zfod++;
617 			fs.m->valid = VM_PAGE_BITS_ALL;
618 			break;	/* break to PAGE HAS BEEN FOUND */
619 		} else {
620 			if (fs.object != fs.first_object) {
621 				vm_object_pip_wakeup(fs.object);
622 			}
623 			KASSERT(fs.object != next_object, ("object loop %p", next_object));
624 			fs.object = next_object;
625 			vm_object_pip_add(fs.object, 1);
626 		}
627 	}
628 
629 	KASSERT((fs.m->flags & PG_BUSY) != 0,
630 	    ("vm_fault: not busy after main loop"));
631 
632 	/*
633 	 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
634 	 * is held.]
635 	 */
636 
637 	/*
638 	 * If the page is being written, but isn't already owned by the
639 	 * top-level object, we have to copy it into a new page owned by the
640 	 * top-level object.
641 	 */
642 	if (fs.object != fs.first_object) {
643 		/*
644 		 * We only really need to copy if we want to write it.
645 		 */
646 		if (fault_type & VM_PROT_WRITE) {
647 			/*
648 			 * This allows pages to be virtually copied from a
649 			 * backing_object into the first_object, where the
650 			 * backing object has no other refs to it, and cannot
651 			 * gain any more refs.  Instead of a bcopy, we just
652 			 * move the page from the backing object to the
653 			 * first object.  Note that we must mark the page
654 			 * dirty in the first object so that it will go out
655 			 * to swap when needed.
656 			 */
657 			if (map_generation == fs.map->timestamp &&
658 				/*
659 				 * Only one shadow object
660 				 */
661 				(fs.object->shadow_count == 1) &&
662 				/*
663 				 * No COW refs, except us
664 				 */
665 				(fs.object->ref_count == 1) &&
666 				/*
667 				 * No one else can look this object up
668 				 */
669 				(fs.object->handle == NULL) &&
670 				/*
671 				 * No other ways to look the object up
672 				 */
673 				((fs.object->type == OBJT_DEFAULT) ||
674 				 (fs.object->type == OBJT_SWAP)) &&
675 				/*
676 				 * We don't chase down the shadow chain
677 				 */
678 				(fs.object == fs.first_object->backing_object) &&
679 
680 				/*
681 				 * grab the lock if we need to
682 				 */
683 			    (fs.lookup_still_valid || vm_map_trylock(fs.map))) {
684 
685 				fs.lookup_still_valid = 1;
686 				/*
687 				 * get rid of the unnecessary page
688 				 */
689 				vm_page_protect(fs.first_m, VM_PROT_NONE);
690 				vm_page_free(fs.first_m);
691 				fs.first_m = NULL;
692 
693 				/*
694 				 * grab the page and put it into the
695 				 * process'es object.  The page is
696 				 * automatically made dirty.
697 				 */
698 				vm_page_rename(fs.m, fs.first_object, fs.first_pindex);
699 				fs.first_m = fs.m;
700 				vm_page_busy(fs.first_m);
701 				fs.m = NULL;
702 				cnt.v_cow_optim++;
703 			} else {
704 				/*
705 				 * Oh, well, lets copy it.
706 				 */
707 				vm_page_copy(fs.m, fs.first_m);
708 			}
709 
710 			if (fs.m) {
711 				/*
712 				 * We no longer need the old page or object.
713 				 */
714 				release_page(&fs);
715 			}
716 
717 			/*
718 			 * fs.object != fs.first_object due to above
719 			 * conditional
720 			 */
721 			vm_object_pip_wakeup(fs.object);
722 
723 			/*
724 			 * Only use the new page below...
725 			 */
726 			cnt.v_cow_faults++;
727 			fs.m = fs.first_m;
728 			fs.object = fs.first_object;
729 			fs.pindex = fs.first_pindex;
730 
731 		} else {
732 			prot &= ~VM_PROT_WRITE;
733 		}
734 	}
735 
736 	/*
737 	 * We must verify that the maps have not changed since our last
738 	 * lookup.
739 	 */
740 	if (!fs.lookup_still_valid &&
741 		(fs.map->timestamp != map_generation)) {
742 		vm_object_t retry_object;
743 		vm_pindex_t retry_pindex;
744 		vm_prot_t retry_prot;
745 
746 		/*
747 		 * Since map entries may be pageable, make sure we can take a
748 		 * page fault on them.
749 		 */
750 
751 		/*
752 		 * Unlock vnode before the lookup to avoid deadlock.   E.G.
753 		 * avoid a deadlock between the inode and exec_map that can
754 		 * occur due to locks being obtained in different orders.
755 		 */
756 		if (fs.vp != NULL) {
757 			vput(fs.vp);
758 			fs.vp = NULL;
759 		}
760 
761 		if (fs.map->infork) {
762 			release_page(&fs);
763 			unlock_and_deallocate(&fs);
764 			goto RetryFault;
765 		}
766 
767 		/*
768 		 * To avoid trying to write_lock the map while another process
769 		 * has it read_locked (in vm_map_pageable), we do not try for
770 		 * write permission.  If the page is still writable, we will
771 		 * get write permission.  If it is not, or has been marked
772 		 * needs_copy, we enter the mapping without write permission,
773 		 * and will merely take another fault.
774 		 */
775 		result = vm_map_lookup(&fs.map, vaddr, fault_type & ~VM_PROT_WRITE,
776 		    &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
777 		map_generation = fs.map->timestamp;
778 
779 		/*
780 		 * If we don't need the page any longer, put it on the active
781 		 * list (the easiest thing to do here).  If no one needs it,
782 		 * pageout will grab it eventually.
783 		 */
784 		if (result != KERN_SUCCESS) {
785 			release_page(&fs);
786 			unlock_and_deallocate(&fs);
787 			mtx_unlock(&Giant);
788 			return (result);
789 		}
790 		fs.lookup_still_valid = TRUE;
791 
792 		if ((retry_object != fs.first_object) ||
793 		    (retry_pindex != fs.first_pindex)) {
794 			release_page(&fs);
795 			unlock_and_deallocate(&fs);
796 			goto RetryFault;
797 		}
798 		/*
799 		 * Check whether the protection has changed or the object has
800 		 * been copied while we left the map unlocked. Changing from
801 		 * read to write permission is OK - we leave the page
802 		 * write-protected, and catch the write fault. Changing from
803 		 * write to read permission means that we can't mark the page
804 		 * write-enabled after all.
805 		 */
806 		prot &= retry_prot;
807 	}
808 
809 	/*
810 	 * Put this page into the physical map. We had to do the unlock above
811 	 * because pmap_enter may cause other faults.   We don't put the page
812 	 * back on the active queue until later so that the page-out daemon
813 	 * won't find us (yet).
814 	 */
815 
816 	if (prot & VM_PROT_WRITE) {
817 		vm_page_flag_set(fs.m, PG_WRITEABLE);
818 		vm_object_set_writeable_dirty(fs.m->object);
819 
820 		/*
821 		 * If the fault is a write, we know that this page is being
822 		 * written NOW so dirty it explicitly to save on
823 		 * pmap_is_modified() calls later.
824 		 *
825 		 * If this is a NOSYNC mmap we do not want to set PG_NOSYNC
826 		 * if the page is already dirty to prevent data written with
827 		 * the expectation of being synced from not being synced.
828 		 * Likewise if this entry does not request NOSYNC then make
829 		 * sure the page isn't marked NOSYNC.  Applications sharing
830 		 * data should use the same flags to avoid ping ponging.
831 		 *
832 		 * Also tell the backing pager, if any, that it should remove
833 		 * any swap backing since the page is now dirty.
834 		 */
835 		if (fs.entry->eflags & MAP_ENTRY_NOSYNC) {
836 			if (fs.m->dirty == 0)
837 				vm_page_flag_set(fs.m, PG_NOSYNC);
838 		} else {
839 			vm_page_flag_clear(fs.m, PG_NOSYNC);
840 		}
841 		if (fault_flags & VM_FAULT_DIRTY) {
842 			int s;
843 			vm_page_dirty(fs.m);
844 			s = splvm();
845 			vm_pager_page_unswapped(fs.m);
846 			splx(s);
847 		}
848 	}
849 
850 	/*
851 	 * Page had better still be busy
852 	 */
853 	KASSERT(fs.m->flags & PG_BUSY,
854 		("vm_fault: page %p not busy!", fs.m));
855 	unlock_things(&fs);
856 
857 	/*
858 	 * Sanity check: page must be completely valid or it is not fit to
859 	 * map into user space.  vm_pager_get_pages() ensures this.
860 	 */
861 	if (fs.m->valid != VM_PAGE_BITS_ALL) {
862 		vm_page_zero_invalid(fs.m, TRUE);
863 		printf("Warning: page %p partially invalid on fault\n", fs.m);
864 	}
865 	pmap_enter(fs.map->pmap, vaddr, fs.m, prot, wired);
866 	if (((fault_flags & VM_FAULT_WIRE_MASK) == 0) && (wired == 0)) {
867 		pmap_prefault(fs.map->pmap, vaddr, fs.entry);
868 	}
869 	vm_page_lock_queues();
870 	vm_page_flag_clear(fs.m, PG_ZERO);
871 	vm_page_flag_set(fs.m, PG_MAPPED|PG_REFERENCED);
872 
873 	/*
874 	 * If the page is not wired down, then put it where the pageout daemon
875 	 * can find it.
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 	vm_page_unlock_queues();
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 	vm_page_wakeup(fs.m);
900 	vm_object_deallocate(fs.first_object);
901 	mtx_unlock(&Giant);
902 	return (KERN_SUCCESS);
903 }
904 
905 /*
906  *	vm_fault_wire:
907  *
908  *	Wire down a range of virtual addresses in a map.
909  */
910 int
911 vm_fault_wire(map, start, end)
912 	vm_map_t map;
913 	vm_offset_t start, end;
914 {
915 
916 	vm_offset_t va;
917 	pmap_t pmap;
918 	int rv;
919 
920 	pmap = vm_map_pmap(map);
921 
922 	/*
923 	 * Inform the physical mapping system that the range of addresses may
924 	 * not fault, so that page tables and such can be locked down as well.
925 	 */
926 	pmap_pageable(pmap, start, end, FALSE);
927 
928 	/*
929 	 * We simulate a fault to get the page and enter it in the physical
930 	 * map.
931 	 */
932 	for (va = start; va < end; va += PAGE_SIZE) {
933 		rv = vm_fault(map, va, VM_PROT_READ|VM_PROT_WRITE,
934 			VM_FAULT_CHANGE_WIRING);
935 		if (rv) {
936 			if (va != start)
937 				vm_fault_unwire(map, start, va);
938 			return (rv);
939 		}
940 	}
941 	return (KERN_SUCCESS);
942 }
943 
944 /*
945  *	vm_fault_user_wire:
946  *
947  *	Wire down a range of virtual addresses in a map.  This
948  *	is for user mode though, so we only ask for read access
949  *	on currently read only sections.
950  */
951 int
952 vm_fault_user_wire(map, start, end)
953 	vm_map_t map;
954 	vm_offset_t start, end;
955 {
956 	vm_offset_t va;
957 	pmap_t pmap;
958 	int rv;
959 
960 	pmap = vm_map_pmap(map);
961 
962 	/*
963 	 * Inform the physical mapping system that the range of addresses may
964 	 * not fault, so that page tables and such can be locked down as well.
965 	 */
966 	pmap_pageable(pmap, start, end, FALSE);
967 
968 	/*
969 	 * We simulate a fault to get the page and enter it in the physical
970 	 * map.
971 	 */
972 	for (va = start; va < end; va += PAGE_SIZE) {
973 		rv = vm_fault(map, va, VM_PROT_READ, VM_FAULT_USER_WIRE);
974 		if (rv) {
975 			if (va != start)
976 				vm_fault_unwire(map, start, va);
977 			return (rv);
978 		}
979 	}
980 	return (KERN_SUCCESS);
981 }
982 
983 
984 /*
985  *	vm_fault_unwire:
986  *
987  *	Unwire a range of virtual addresses in a map.
988  */
989 void
990 vm_fault_unwire(map, start, end)
991 	vm_map_t map;
992 	vm_offset_t start, end;
993 {
994 	vm_offset_t va, pa;
995 	pmap_t pmap;
996 
997 	pmap = vm_map_pmap(map);
998 
999 	mtx_lock(&Giant);
1000 	/*
1001 	 * Since the pages are wired down, we must be able to get their
1002 	 * mappings from the physical map system.
1003 	 */
1004 	for (va = start; va < end; va += PAGE_SIZE) {
1005 		pa = pmap_extract(pmap, va);
1006 		if (pa != (vm_offset_t) 0) {
1007 			pmap_change_wiring(pmap, va, FALSE);
1008 			vm_page_lock_queues();
1009 			vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1);
1010 			vm_page_unlock_queues();
1011 		}
1012 	}
1013 	mtx_unlock(&Giant);
1014 
1015 	/*
1016 	 * Inform the physical mapping system that the range of addresses may
1017 	 * fault, so that page tables and such may be unwired themselves.
1018 	 */
1019 	pmap_pageable(pmap, start, end, TRUE);
1020 }
1021 
1022 /*
1023  *	Routine:
1024  *		vm_fault_copy_entry
1025  *	Function:
1026  *		Copy all of the pages from a wired-down map entry to another.
1027  *
1028  *	In/out conditions:
1029  *		The source and destination maps must be locked for write.
1030  *		The source map entry must be wired down (or be a sharing map
1031  *		entry corresponding to a main map entry that is wired down).
1032  */
1033 void
1034 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry)
1035 	vm_map_t dst_map;
1036 	vm_map_t src_map;
1037 	vm_map_entry_t dst_entry;
1038 	vm_map_entry_t src_entry;
1039 {
1040 	vm_object_t dst_object;
1041 	vm_object_t src_object;
1042 	vm_ooffset_t dst_offset;
1043 	vm_ooffset_t src_offset;
1044 	vm_prot_t prot;
1045 	vm_offset_t vaddr;
1046 	vm_page_t dst_m;
1047 	vm_page_t src_m;
1048 
1049 #ifdef	lint
1050 	src_map++;
1051 #endif	/* lint */
1052 
1053 	src_object = src_entry->object.vm_object;
1054 	src_offset = src_entry->offset;
1055 
1056 	/*
1057 	 * Create the top-level object for the destination entry. (Doesn't
1058 	 * actually shadow anything - we copy the pages directly.)
1059 	 */
1060 	dst_object = vm_object_allocate(OBJT_DEFAULT,
1061 	    (vm_size_t) OFF_TO_IDX(dst_entry->end - dst_entry->start));
1062 
1063 	dst_entry->object.vm_object = dst_object;
1064 	dst_entry->offset = 0;
1065 
1066 	prot = dst_entry->max_protection;
1067 
1068 	/*
1069 	 * Loop through all of the pages in the entry's range, copying each
1070 	 * one from the source object (it should be there) to the destination
1071 	 * object.
1072 	 */
1073 	for (vaddr = dst_entry->start, dst_offset = 0;
1074 	    vaddr < dst_entry->end;
1075 	    vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) {
1076 
1077 		/*
1078 		 * Allocate a page in the destination object
1079 		 */
1080 		do {
1081 			dst_m = vm_page_alloc(dst_object,
1082 				OFF_TO_IDX(dst_offset), VM_ALLOC_NORMAL);
1083 			if (dst_m == NULL) {
1084 				VM_WAIT;
1085 			}
1086 		} while (dst_m == NULL);
1087 
1088 		/*
1089 		 * Find the page in the source object, and copy it in.
1090 		 * (Because the source is wired down, the page will be in
1091 		 * memory.)
1092 		 */
1093 		src_m = vm_page_lookup(src_object,
1094 			OFF_TO_IDX(dst_offset + src_offset));
1095 		if (src_m == NULL)
1096 			panic("vm_fault_copy_wired: page missing");
1097 
1098 		vm_page_copy(src_m, dst_m);
1099 
1100 		/*
1101 		 * Enter it in the pmap...
1102 		 */
1103 		vm_page_flag_clear(dst_m, PG_ZERO);
1104 		pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE);
1105 		vm_page_flag_set(dst_m, PG_WRITEABLE|PG_MAPPED);
1106 
1107 		/*
1108 		 * Mark it no longer busy, and put it on the active list.
1109 		 */
1110 		vm_page_activate(dst_m);
1111 		vm_page_wakeup(dst_m);
1112 	}
1113 }
1114 
1115 
1116 /*
1117  * This routine checks around the requested page for other pages that
1118  * might be able to be faulted in.  This routine brackets the viable
1119  * pages for the pages to be paged in.
1120  *
1121  * Inputs:
1122  *	m, rbehind, rahead
1123  *
1124  * Outputs:
1125  *  marray (array of vm_page_t), reqpage (index of requested page)
1126  *
1127  * Return value:
1128  *  number of pages in marray
1129  *
1130  * This routine can't block.
1131  */
1132 static int
1133 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage)
1134 	vm_page_t m;
1135 	int rbehind;
1136 	int rahead;
1137 	vm_page_t *marray;
1138 	int *reqpage;
1139 {
1140 	int i,j;
1141 	vm_object_t object;
1142 	vm_pindex_t pindex, startpindex, endpindex, tpindex;
1143 	vm_page_t rtm;
1144 	int cbehind, cahead;
1145 
1146 	GIANT_REQUIRED;
1147 
1148 	object = m->object;
1149 	pindex = m->pindex;
1150 
1151 	/*
1152 	 * we don't fault-ahead for device pager
1153 	 */
1154 	if (object->type == OBJT_DEVICE) {
1155 		*reqpage = 0;
1156 		marray[0] = m;
1157 		return 1;
1158 	}
1159 
1160 	/*
1161 	 * if the requested page is not available, then give up now
1162 	 */
1163 	if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1164 		return 0;
1165 	}
1166 
1167 	if ((cbehind == 0) && (cahead == 0)) {
1168 		*reqpage = 0;
1169 		marray[0] = m;
1170 		return 1;
1171 	}
1172 
1173 	if (rahead > cahead) {
1174 		rahead = cahead;
1175 	}
1176 
1177 	if (rbehind > cbehind) {
1178 		rbehind = cbehind;
1179 	}
1180 
1181 	/*
1182 	 * try to do any readahead that we might have free pages for.
1183 	 */
1184 	if ((rahead + rbehind) >
1185 		((cnt.v_free_count + cnt.v_cache_count) - cnt.v_free_reserved)) {
1186 		pagedaemon_wakeup();
1187 		marray[0] = m;
1188 		*reqpage = 0;
1189 		return 1;
1190 	}
1191 
1192 	/*
1193 	 * scan backward for the read behind pages -- in memory
1194 	 */
1195 	if (pindex > 0) {
1196 		if (rbehind > pindex) {
1197 			rbehind = pindex;
1198 			startpindex = 0;
1199 		} else {
1200 			startpindex = pindex - rbehind;
1201 		}
1202 
1203 		for (tpindex = pindex - 1; tpindex >= startpindex; tpindex -= 1) {
1204 			if (vm_page_lookup(object, tpindex)) {
1205 				startpindex = tpindex + 1;
1206 				break;
1207 			}
1208 			if (tpindex == 0)
1209 				break;
1210 		}
1211 
1212 		for (i = 0, tpindex = startpindex; tpindex < pindex; i++, tpindex++) {
1213 
1214 			rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
1215 			if (rtm == NULL) {
1216 				for (j = 0; j < i; j++) {
1217 					vm_page_free(marray[j]);
1218 				}
1219 				marray[0] = m;
1220 				*reqpage = 0;
1221 				return 1;
1222 			}
1223 
1224 			marray[i] = rtm;
1225 		}
1226 	} else {
1227 		startpindex = 0;
1228 		i = 0;
1229 	}
1230 
1231 	marray[i] = m;
1232 	/* page offset of the required page */
1233 	*reqpage = i;
1234 
1235 	tpindex = pindex + 1;
1236 	i++;
1237 
1238 	/*
1239 	 * scan forward for the read ahead pages
1240 	 */
1241 	endpindex = tpindex + rahead;
1242 	if (endpindex > object->size)
1243 		endpindex = object->size;
1244 
1245 	for (; tpindex < endpindex; i++, tpindex++) {
1246 
1247 		if (vm_page_lookup(object, tpindex)) {
1248 			break;
1249 		}
1250 
1251 		rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
1252 		if (rtm == NULL) {
1253 			break;
1254 		}
1255 
1256 		marray[i] = rtm;
1257 	}
1258 
1259 	/* return number of bytes of pages */
1260 	return i;
1261 }
1262