xref: /freebsd/sys/vm/vm_fault.c (revision 51a9219f5780e61e1437d25220bf8750d9df7f8b)
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_lock_queues();
121 	vm_page_wakeup(fs->m);
122 	vm_page_deactivate(fs->m);
123 	vm_page_unlock_queues();
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_lock_queues();
143 		vm_page_free(fs->first_m);
144 		vm_page_unlock_queues();
145 		vm_object_pip_wakeup(fs->first_object);
146 		fs->first_m = NULL;
147 	}
148 	if (dealloc) {
149 		vm_object_deallocate(fs->first_object);
150 	}
151 	unlock_map(fs);
152 	if (fs->vp != NULL) {
153 		vput(fs->vp);
154 		fs->vp = NULL;
155 	}
156 }
157 
158 #define unlock_things(fs) _unlock_things(fs, 0)
159 #define unlock_and_deallocate(fs) _unlock_things(fs, 1)
160 
161 /*
162  * TRYPAGER - used by vm_fault to calculate whether the pager for the
163  *	      current object *might* contain the page.
164  *
165  *	      default objects are zero-fill, there is no real pager.
166  */
167 #define TRYPAGER	(fs.object->type != OBJT_DEFAULT && \
168 			(((fault_flags & VM_FAULT_WIRE_MASK) == 0) || wired))
169 
170 /*
171  *	vm_fault:
172  *
173  *	Handle a page fault occurring at the given address,
174  *	requiring the given permissions, in the map specified.
175  *	If successful, the page is inserted into the
176  *	associated physical map.
177  *
178  *	NOTE: the given address should be truncated to the
179  *	proper page address.
180  *
181  *	KERN_SUCCESS is returned if the page fault is handled; otherwise,
182  *	a standard error specifying why the fault is fatal is returned.
183  *
184  *
185  *	The map in question must be referenced, and remains so.
186  *	Caller may hold no locks.
187  */
188 int
189 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
190 	 int fault_flags)
191 {
192 	vm_prot_t prot;
193 	int result;
194 	boolean_t growstack, wired;
195 	int map_generation;
196 	vm_object_t next_object;
197 	vm_page_t marray[VM_FAULT_READ];
198 	int hardfault;
199 	int faultcount;
200 	struct faultstate fs;
201 
202 	hardfault = 0;
203 	growstack = TRUE;
204 	atomic_add_int(&cnt.v_vm_faults, 1);
205 
206 	mtx_lock(&Giant);
207 RetryFault:;
208 
209 	/*
210 	 * Find the backing store object and offset into it to begin the
211 	 * search.
212 	 */
213 	fs.map = map;
214 	result = vm_map_lookup(&fs.map, vaddr, fault_type, &fs.entry,
215 	    &fs.first_object, &fs.first_pindex, &prot, &wired);
216 	if (result != KERN_SUCCESS) {
217 		if (result != KERN_PROTECTION_FAILURE ||
218 		    (fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE) {
219 			if (growstack && result == KERN_INVALID_ADDRESS &&
220 			    map != kernel_map && curproc != NULL) {
221 				result = vm_map_growstack(curproc, vaddr);
222 				if (result != KERN_SUCCESS) {
223 					mtx_unlock(&Giant);
224 					return (KERN_FAILURE);
225 				}
226 				growstack = FALSE;
227 				goto RetryFault;
228 			}
229 			mtx_unlock(&Giant);
230 			return (result);
231 		}
232 
233 		/*
234    		 * If we are user-wiring a r/w segment, and it is COW, then
235    		 * we need to do the COW operation.  Note that we don't COW
236    		 * currently RO sections now, because it is NOT desirable
237    		 * to COW .text.  We simply keep .text from ever being COW'ed
238    		 * and take the heat that one cannot debug wired .text sections.
239    		 */
240 		result = vm_map_lookup(&fs.map, vaddr,
241 			VM_PROT_READ|VM_PROT_WRITE|VM_PROT_OVERRIDE_WRITE,
242 			&fs.entry, &fs.first_object, &fs.first_pindex, &prot, &wired);
243 		if (result != KERN_SUCCESS) {
244 			mtx_unlock(&Giant);
245 			return (result);
246 		}
247 
248 		/*
249 		 * If we don't COW now, on a user wire, the user will never
250 		 * be able to write to the mapping.  If we don't make this
251 		 * restriction, the bookkeeping would be nearly impossible.
252 		 *
253 		 * XXX The following assignment modifies the map without
254 		 * holding a write lock on it.
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 	 * Bump the paging-in-progress count to prevent size changes (e.g.
274 	 * truncation operations) during I/O.  This must be done after
275 	 * obtaining the vnode lock in order to avoid possible deadlocks.
276 	 *
277 	 * XXX vnode_pager_lock() can block without releasing the map lock.
278 	 */
279 	vm_object_reference(fs.first_object);
280 	fs.vp = vnode_pager_lock(fs.first_object);
281 	vm_object_pip_add(fs.first_object, 1);
282 
283 #ifdef ENABLE_VFS_IOOPT
284 	if ((fault_type & VM_PROT_WRITE) &&
285 		(fs.first_object->type == OBJT_VNODE)) {
286 		vm_freeze_copyopts(fs.first_object,
287 			fs.first_pindex, fs.first_pindex + 1);
288 	}
289 #endif
290 	fs.lookup_still_valid = TRUE;
291 
292 	if (wired)
293 		fault_type = prot;
294 
295 	fs.first_m = NULL;
296 
297 	/*
298 	 * Search for the page at object/offset.
299 	 */
300 	fs.object = fs.first_object;
301 	fs.pindex = fs.first_pindex;
302 	while (TRUE) {
303 		/*
304 		 * If the object is dead, we stop here
305 		 */
306 		if (fs.object->flags & OBJ_DEAD) {
307 			unlock_and_deallocate(&fs);
308 			mtx_unlock(&Giant);
309 			return (KERN_PROTECTION_FAILURE);
310 		}
311 
312 		/*
313 		 * See if page is resident
314 		 */
315 		fs.m = vm_page_lookup(fs.object, fs.pindex);
316 		if (fs.m != NULL) {
317 			int queue, s;
318 
319 			/*
320 			 * check for page-based copy on write
321 			 */
322 			vm_page_lock_queues();
323 			if ((fs.m->cow) &&
324 			    (fault_type & VM_PROT_WRITE)) {
325 				s = splvm();
326 				vm_page_cowfault(fs.m);
327 				splx(s);
328 				vm_page_unlock_queues();
329 				unlock_things(&fs);
330 				goto RetryFault;
331 			}
332 
333 			/*
334 			 * Wait/Retry if the page is busy.  We have to do this
335 			 * if the page is busy via either PG_BUSY or
336 			 * vm_page_t->busy because the vm_pager may be using
337 			 * vm_page_t->busy for pageouts ( and even pageins if
338 			 * it is the vnode pager ), and we could end up trying
339 			 * to pagein and pageout the same page simultaneously.
340 			 *
341 			 * We can theoretically allow the busy case on a read
342 			 * fault if the page is marked valid, but since such
343 			 * pages are typically already pmap'd, putting that
344 			 * special case in might be more effort then it is
345 			 * worth.  We cannot under any circumstances mess
346 			 * around with a vm_page_t->busy page except, perhaps,
347 			 * to pmap it.
348 			 */
349 			if ((fs.m->flags & PG_BUSY) || fs.m->busy) {
350 				vm_page_unlock_queues();
351 				unlock_things(&fs);
352 				(void)vm_page_sleep_busy(fs.m, TRUE, "vmpfw");
353 				cnt.v_intrans++;
354 				vm_object_deallocate(fs.first_object);
355 				goto RetryFault;
356 			}
357 			queue = fs.m->queue;
358 
359 			s = splvm();
360 			vm_pageq_remove_nowakeup(fs.m);
361 			splx(s);
362 
363 			if ((queue - fs.m->pc) == PQ_CACHE && vm_page_count_severe()) {
364 				vm_page_activate(fs.m);
365 				vm_page_unlock_queues();
366 				unlock_and_deallocate(&fs);
367 				VM_WAITPFAULT;
368 				goto RetryFault;
369 			}
370 
371 			/*
372 			 * Mark page busy for other processes, and the
373 			 * pagedaemon.  If it still isn't completely valid
374 			 * (readable), jump to readrest, else break-out ( we
375 			 * found the page ).
376 			 */
377 			vm_page_busy(fs.m);
378 			vm_page_unlock_queues();
379 			if (((fs.m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) &&
380 				fs.m->object != kernel_object && fs.m->object != kmem_object) {
381 				goto readrest;
382 			}
383 
384 			break;
385 		}
386 
387 		/*
388 		 * Page is not resident, If this is the search termination
389 		 * or the pager might contain the page, allocate a new page.
390 		 */
391 		if (TRYPAGER || fs.object == fs.first_object) {
392 			if (fs.pindex >= fs.object->size) {
393 				unlock_and_deallocate(&fs);
394 				mtx_unlock(&Giant);
395 				return (KERN_PROTECTION_FAILURE);
396 			}
397 
398 			/*
399 			 * Allocate a new page for this object/offset pair.
400 			 */
401 			fs.m = NULL;
402 			if (!vm_page_count_severe()) {
403 				fs.m = vm_page_alloc(fs.object, fs.pindex,
404 				    (fs.vp || fs.object->backing_object)? VM_ALLOC_NORMAL: VM_ALLOC_ZERO);
405 			}
406 			if (fs.m == NULL) {
407 				unlock_and_deallocate(&fs);
408 				VM_WAITPFAULT;
409 				goto RetryFault;
410 			}
411 		}
412 
413 readrest:
414 		/*
415 		 * We have found a valid page or we have allocated a new page.
416 		 * The page thus may not be valid or may not be entirely
417 		 * valid.
418 		 *
419 		 * Attempt to fault-in the page if there is a chance that the
420 		 * pager has it, and potentially fault in additional pages
421 		 * at the same time.
422 		 */
423 		if (TRYPAGER) {
424 			int rv;
425 			int reqpage;
426 			int ahead, behind;
427 			u_char behavior = vm_map_entry_behavior(fs.entry);
428 
429 			if (behavior == MAP_ENTRY_BEHAV_RANDOM) {
430 				ahead = 0;
431 				behind = 0;
432 			} else {
433 				behind = (vaddr - fs.entry->start) >> PAGE_SHIFT;
434 				if (behind > VM_FAULT_READ_BEHIND)
435 					behind = VM_FAULT_READ_BEHIND;
436 
437 				ahead = ((fs.entry->end - vaddr) >> PAGE_SHIFT) - 1;
438 				if (ahead > VM_FAULT_READ_AHEAD)
439 					ahead = VM_FAULT_READ_AHEAD;
440 			}
441 
442 			if ((fs.first_object->type != OBJT_DEVICE) &&
443 			    (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL ||
444                                 (behavior != MAP_ENTRY_BEHAV_RANDOM &&
445                                 fs.pindex >= fs.entry->lastr &&
446                                 fs.pindex < fs.entry->lastr + VM_FAULT_READ))
447 			) {
448 				vm_pindex_t firstpindex, tmppindex;
449 
450 				if (fs.first_pindex < 2 * VM_FAULT_READ)
451 					firstpindex = 0;
452 				else
453 					firstpindex = fs.first_pindex - 2 * VM_FAULT_READ;
454 
455 				vm_page_lock_queues();
456 				/*
457 				 * note: partially valid pages cannot be
458 				 * included in the lookahead - NFS piecemeal
459 				 * writes will barf on it badly.
460 				 */
461 				for (tmppindex = fs.first_pindex - 1;
462 					tmppindex >= firstpindex;
463 					--tmppindex) {
464 					vm_page_t mt;
465 
466 					mt = vm_page_lookup(fs.first_object, tmppindex);
467 					if (mt == NULL || (mt->valid != VM_PAGE_BITS_ALL))
468 						break;
469 					if (mt->busy ||
470 						(mt->flags & (PG_BUSY | PG_FICTITIOUS | PG_UNMANAGED)) ||
471 						mt->hold_count ||
472 						mt->wire_count)
473 						continue;
474 					if (mt->dirty == 0)
475 						vm_page_test_dirty(mt);
476 					if (mt->dirty) {
477 						vm_page_protect(mt, VM_PROT_NONE);
478 						vm_page_deactivate(mt);
479 					} else {
480 						vm_page_cache(mt);
481 					}
482 				}
483 				vm_page_unlock_queues();
484 				ahead += behind;
485 				behind = 0;
486 			}
487 
488 			/*
489 			 * now we find out if any other pages should be paged
490 			 * in at this time this routine checks to see if the
491 			 * pages surrounding this fault reside in the same
492 			 * object as the page for this fault.  If they do,
493 			 * then they are faulted in also into the object.  The
494 			 * array "marray" returned contains an array of
495 			 * vm_page_t structs where one of them is the
496 			 * vm_page_t passed to the routine.  The reqpage
497 			 * return value is the index into the marray for the
498 			 * vm_page_t passed to the routine.
499 			 *
500 			 * fs.m plus the additional pages are PG_BUSY'd.
501 			 *
502 			 * XXX vm_fault_additional_pages() can block
503 			 * without releasing the map lock.
504 			 */
505 			faultcount = vm_fault_additional_pages(
506 			    fs.m, behind, ahead, marray, &reqpage);
507 
508 			/*
509 			 * update lastr imperfectly (we do not know how much
510 			 * getpages will actually read), but good enough.
511 			 *
512 			 * XXX The following assignment modifies the map
513 			 * without holding a write lock on it.
514 			 */
515 			fs.entry->lastr = fs.pindex + faultcount - behind;
516 
517 			/*
518 			 * Call the pager to retrieve the data, if any, after
519 			 * releasing the lock on the map.  We hold a ref on
520 			 * fs.object and the pages are PG_BUSY'd.
521 			 */
522 			unlock_map(&fs);
523 
524 			rv = faultcount ?
525 			    vm_pager_get_pages(fs.object, marray, faultcount,
526 				reqpage) : VM_PAGER_FAIL;
527 
528 			if (rv == VM_PAGER_OK) {
529 				/*
530 				 * Found the page. Leave it busy while we play
531 				 * with it.
532 				 */
533 
534 				/*
535 				 * Relookup in case pager changed page. Pager
536 				 * is responsible for disposition of old page
537 				 * if moved.
538 				 */
539 				fs.m = vm_page_lookup(fs.object, fs.pindex);
540 				if (!fs.m) {
541 					unlock_and_deallocate(&fs);
542 					goto RetryFault;
543 				}
544 
545 				hardfault++;
546 				break; /* break to PAGE HAS BEEN FOUND */
547 			}
548 			/*
549 			 * Remove the bogus page (which does not exist at this
550 			 * object/offset); before doing so, we must get back
551 			 * our object lock to preserve our invariant.
552 			 *
553 			 * Also wake up any other process that may want to bring
554 			 * in this page.
555 			 *
556 			 * If this is the top-level object, we must leave the
557 			 * busy page to prevent another process from rushing
558 			 * past us, and inserting the page in that object at
559 			 * the same time that we are.
560 			 */
561 			if (rv == VM_PAGER_ERROR)
562 				printf("vm_fault: pager read error, pid %d (%s)\n",
563 				    curproc->p_pid, curproc->p_comm);
564 			/*
565 			 * Data outside the range of the pager or an I/O error
566 			 */
567 			/*
568 			 * XXX - the check for kernel_map is a kludge to work
569 			 * around having the machine panic on a kernel space
570 			 * fault w/ I/O error.
571 			 */
572 			if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) ||
573 				(rv == VM_PAGER_BAD)) {
574 				vm_page_lock_queues();
575 				vm_page_free(fs.m);
576 				vm_page_unlock_queues();
577 				fs.m = NULL;
578 				unlock_and_deallocate(&fs);
579 				mtx_unlock(&Giant);
580 				return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
581 			}
582 			if (fs.object != fs.first_object) {
583 				vm_page_lock_queues();
584 				vm_page_free(fs.m);
585 				vm_page_unlock_queues();
586 				fs.m = NULL;
587 				/*
588 				 * XXX - we cannot just fall out at this
589 				 * point, m has been freed and is invalid!
590 				 */
591 			}
592 		}
593 
594 		/*
595 		 * We get here if the object has default pager (or unwiring)
596 		 * or the pager doesn't have the page.
597 		 */
598 		if (fs.object == fs.first_object)
599 			fs.first_m = fs.m;
600 
601 		/*
602 		 * Move on to the next object.  Lock the next object before
603 		 * unlocking the current one.
604 		 */
605 		fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset);
606 		next_object = fs.object->backing_object;
607 		if (next_object == NULL) {
608 			/*
609 			 * If there's no object left, fill the page in the top
610 			 * object with zeros.
611 			 */
612 			if (fs.object != fs.first_object) {
613 				vm_object_pip_wakeup(fs.object);
614 
615 				fs.object = fs.first_object;
616 				fs.pindex = fs.first_pindex;
617 				fs.m = fs.first_m;
618 			}
619 			fs.first_m = NULL;
620 
621 			/*
622 			 * Zero the page if necessary and mark it valid.
623 			 */
624 			if ((fs.m->flags & PG_ZERO) == 0) {
625 				pmap_zero_page(fs.m);
626 			} else {
627 				cnt.v_ozfod++;
628 			}
629 			cnt.v_zfod++;
630 			fs.m->valid = VM_PAGE_BITS_ALL;
631 			break;	/* break to PAGE HAS BEEN FOUND */
632 		} else {
633 			if (fs.object != fs.first_object) {
634 				vm_object_pip_wakeup(fs.object);
635 			}
636 			KASSERT(fs.object != next_object, ("object loop %p", next_object));
637 			fs.object = next_object;
638 			vm_object_pip_add(fs.object, 1);
639 		}
640 	}
641 
642 	KASSERT((fs.m->flags & PG_BUSY) != 0,
643 	    ("vm_fault: not busy after main loop"));
644 
645 	/*
646 	 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
647 	 * is held.]
648 	 */
649 
650 	/*
651 	 * If the page is being written, but isn't already owned by the
652 	 * top-level object, we have to copy it into a new page owned by the
653 	 * top-level object.
654 	 */
655 	if (fs.object != fs.first_object) {
656 		/*
657 		 * We only really need to copy if we want to write it.
658 		 */
659 		if (fault_type & VM_PROT_WRITE) {
660 			/*
661 			 * This allows pages to be virtually copied from a
662 			 * backing_object into the first_object, where the
663 			 * backing object has no other refs to it, and cannot
664 			 * gain any more refs.  Instead of a bcopy, we just
665 			 * move the page from the backing object to the
666 			 * first object.  Note that we must mark the page
667 			 * dirty in the first object so that it will go out
668 			 * to swap when needed.
669 			 */
670 			if (map_generation == fs.map->timestamp &&
671 				/*
672 				 * Only one shadow object
673 				 */
674 				(fs.object->shadow_count == 1) &&
675 				/*
676 				 * No COW refs, except us
677 				 */
678 				(fs.object->ref_count == 1) &&
679 				/*
680 				 * No one else can look this object up
681 				 */
682 				(fs.object->handle == NULL) &&
683 				/*
684 				 * No other ways to look the object up
685 				 */
686 				((fs.object->type == OBJT_DEFAULT) ||
687 				 (fs.object->type == OBJT_SWAP)) &&
688 				/*
689 				 * We don't chase down the shadow chain
690 				 */
691 				(fs.object == fs.first_object->backing_object) &&
692 
693 				/*
694 				 * grab the lock if we need to
695 				 */
696 			    (fs.lookup_still_valid || vm_map_trylock(fs.map))) {
697 
698 				fs.lookup_still_valid = 1;
699 				/*
700 				 * get rid of the unnecessary page
701 				 */
702 				vm_page_lock_queues();
703 				vm_page_protect(fs.first_m, VM_PROT_NONE);
704 				vm_page_free(fs.first_m);
705 				vm_page_unlock_queues();
706 				fs.first_m = NULL;
707 
708 				/*
709 				 * grab the page and put it into the
710 				 * process'es object.  The page is
711 				 * automatically made dirty.
712 				 */
713 				vm_page_rename(fs.m, fs.first_object, fs.first_pindex);
714 				fs.first_m = fs.m;
715 				vm_page_busy(fs.first_m);
716 				fs.m = NULL;
717 				cnt.v_cow_optim++;
718 			} else {
719 				/*
720 				 * Oh, well, lets copy it.
721 				 */
722 				vm_page_copy(fs.m, fs.first_m);
723 			}
724 
725 			if (fs.m) {
726 				/*
727 				 * We no longer need the old page or object.
728 				 */
729 				release_page(&fs);
730 			}
731 
732 			/*
733 			 * fs.object != fs.first_object due to above
734 			 * conditional
735 			 */
736 			vm_object_pip_wakeup(fs.object);
737 
738 			/*
739 			 * Only use the new page below...
740 			 */
741 			cnt.v_cow_faults++;
742 			fs.m = fs.first_m;
743 			fs.object = fs.first_object;
744 			fs.pindex = fs.first_pindex;
745 
746 		} else {
747 			prot &= ~VM_PROT_WRITE;
748 		}
749 	}
750 
751 	/*
752 	 * We must verify that the maps have not changed since our last
753 	 * lookup.
754 	 */
755 	if (!fs.lookup_still_valid &&
756 		(fs.map->timestamp != map_generation)) {
757 		vm_object_t retry_object;
758 		vm_pindex_t retry_pindex;
759 		vm_prot_t retry_prot;
760 
761 		/*
762 		 * Since map entries may be pageable, make sure we can take a
763 		 * page fault on them.
764 		 */
765 
766 		/*
767 		 * Unlock vnode before the lookup to avoid deadlock.   E.G.
768 		 * avoid a deadlock between the inode and exec_map that can
769 		 * occur due to locks being obtained in different orders.
770 		 */
771 		if (fs.vp != NULL) {
772 			vput(fs.vp);
773 			fs.vp = NULL;
774 		}
775 
776 		if (fs.map->infork) {
777 			release_page(&fs);
778 			unlock_and_deallocate(&fs);
779 			goto RetryFault;
780 		}
781 
782 		/*
783 		 * To avoid trying to write_lock the map while another process
784 		 * has it read_locked (in vm_map_pageable), we do not try for
785 		 * write permission.  If the page is still writable, we will
786 		 * get write permission.  If it is not, or has been marked
787 		 * needs_copy, we enter the mapping without write permission,
788 		 * and will merely take another fault.
789 		 */
790 		result = vm_map_lookup(&fs.map, vaddr, fault_type & ~VM_PROT_WRITE,
791 		    &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
792 		map_generation = fs.map->timestamp;
793 
794 		/*
795 		 * If we don't need the page any longer, put it on the active
796 		 * list (the easiest thing to do here).  If no one needs it,
797 		 * pageout will grab it eventually.
798 		 */
799 		if (result != KERN_SUCCESS) {
800 			release_page(&fs);
801 			unlock_and_deallocate(&fs);
802 			mtx_unlock(&Giant);
803 			return (result);
804 		}
805 		fs.lookup_still_valid = TRUE;
806 
807 		if ((retry_object != fs.first_object) ||
808 		    (retry_pindex != fs.first_pindex)) {
809 			release_page(&fs);
810 			unlock_and_deallocate(&fs);
811 			goto RetryFault;
812 		}
813 		/*
814 		 * Check whether the protection has changed or the object has
815 		 * been copied while we left the map unlocked. Changing from
816 		 * read to write permission is OK - we leave the page
817 		 * write-protected, and catch the write fault. Changing from
818 		 * write to read permission means that we can't mark the page
819 		 * write-enabled after all.
820 		 */
821 		prot &= retry_prot;
822 	}
823 
824 	/*
825 	 * Put this page into the physical map. We had to do the unlock above
826 	 * because pmap_enter may cause other faults.   We don't put the page
827 	 * back on the active queue until later so that the page-out daemon
828 	 * won't find us (yet).
829 	 */
830 
831 	if (prot & VM_PROT_WRITE) {
832 		vm_page_flag_set(fs.m, PG_WRITEABLE);
833 		vm_object_set_writeable_dirty(fs.m->object);
834 
835 		/*
836 		 * If the fault is a write, we know that this page is being
837 		 * written NOW so dirty it explicitly to save on
838 		 * pmap_is_modified() calls later.
839 		 *
840 		 * If this is a NOSYNC mmap we do not want to set PG_NOSYNC
841 		 * if the page is already dirty to prevent data written with
842 		 * the expectation of being synced from not being synced.
843 		 * Likewise if this entry does not request NOSYNC then make
844 		 * sure the page isn't marked NOSYNC.  Applications sharing
845 		 * data should use the same flags to avoid ping ponging.
846 		 *
847 		 * Also tell the backing pager, if any, that it should remove
848 		 * any swap backing since the page is now dirty.
849 		 */
850 		if (fs.entry->eflags & MAP_ENTRY_NOSYNC) {
851 			if (fs.m->dirty == 0)
852 				vm_page_flag_set(fs.m, PG_NOSYNC);
853 		} else {
854 			vm_page_flag_clear(fs.m, PG_NOSYNC);
855 		}
856 		if (fault_flags & VM_FAULT_DIRTY) {
857 			int s;
858 			vm_page_dirty(fs.m);
859 			s = splvm();
860 			vm_pager_page_unswapped(fs.m);
861 			splx(s);
862 		}
863 	}
864 
865 	/*
866 	 * Page had better still be busy
867 	 */
868 	KASSERT(fs.m->flags & PG_BUSY,
869 		("vm_fault: page %p not busy!", fs.m));
870 	unlock_things(&fs);
871 
872 	/*
873 	 * Sanity check: page must be completely valid or it is not fit to
874 	 * map into user space.  vm_pager_get_pages() ensures this.
875 	 */
876 	if (fs.m->valid != VM_PAGE_BITS_ALL) {
877 		vm_page_zero_invalid(fs.m, TRUE);
878 		printf("Warning: page %p partially invalid on fault\n", fs.m);
879 	}
880 	pmap_enter(fs.map->pmap, vaddr, fs.m, prot, wired);
881 	if (((fault_flags & VM_FAULT_WIRE_MASK) == 0) && (wired == 0)) {
882 		pmap_prefault(fs.map->pmap, vaddr, fs.entry);
883 	}
884 	vm_page_lock_queues();
885 	vm_page_flag_clear(fs.m, PG_ZERO);
886 	vm_page_flag_set(fs.m, PG_REFERENCED);
887 
888 	/*
889 	 * If the page is not wired down, then put it where the pageout daemon
890 	 * can find it.
891 	 */
892 	if (fault_flags & VM_FAULT_WIRE_MASK) {
893 		if (wired)
894 			vm_page_wire(fs.m);
895 		else
896 			vm_page_unwire(fs.m, 1);
897 	} else {
898 		vm_page_activate(fs.m);
899 	}
900 	vm_page_wakeup(fs.m);
901 	vm_page_unlock_queues();
902 	mtx_lock_spin(&sched_lock);
903 	if (curproc && (curproc->p_sflag & PS_INMEM) && curproc->p_stats) {
904 		if (hardfault) {
905 			curproc->p_stats->p_ru.ru_majflt++;
906 		} else {
907 			curproc->p_stats->p_ru.ru_minflt++;
908 		}
909 	}
910 	mtx_unlock_spin(&sched_lock);
911 
912 	/*
913 	 * Unlock everything, and return
914 	 */
915 	vm_object_deallocate(fs.first_object);
916 	mtx_unlock(&Giant);
917 	return (KERN_SUCCESS);
918 }
919 
920 /*
921  *	vm_fault_wire:
922  *
923  *	Wire down a range of virtual addresses in a map.
924  */
925 int
926 vm_fault_wire(map, start, end, user_wire)
927 	vm_map_t map;
928 	vm_offset_t start, end;
929 	boolean_t user_wire;
930 {
931 	vm_offset_t va;
932 	int rv;
933 
934 	/*
935 	 * We simulate a fault to get the page and enter it in the physical
936 	 * map.  For user wiring, we only ask for read access on currently
937 	 * read-only sections.
938 	 */
939 	for (va = start; va < end; va += PAGE_SIZE) {
940 		rv = vm_fault(map, va,
941 		    user_wire ? VM_PROT_READ : VM_PROT_READ | VM_PROT_WRITE,
942 		    user_wire ? VM_FAULT_USER_WIRE : VM_FAULT_CHANGE_WIRING);
943 		if (rv) {
944 			if (va != start)
945 				vm_fault_unwire(map, start, va);
946 			return (rv);
947 		}
948 	}
949 	return (KERN_SUCCESS);
950 }
951 
952 /*
953  *	vm_fault_unwire:
954  *
955  *	Unwire a range of virtual addresses in a map.
956  */
957 void
958 vm_fault_unwire(map, start, end)
959 	vm_map_t map;
960 	vm_offset_t start, end;
961 {
962 	vm_offset_t va, pa;
963 	pmap_t pmap;
964 
965 	pmap = vm_map_pmap(map);
966 
967 	mtx_lock(&Giant);
968 	/*
969 	 * Since the pages are wired down, we must be able to get their
970 	 * mappings from the physical map system.
971 	 */
972 	for (va = start; va < end; va += PAGE_SIZE) {
973 		pa = pmap_extract(pmap, va);
974 		if (pa != (vm_offset_t) 0) {
975 			pmap_change_wiring(pmap, va, FALSE);
976 			vm_page_lock_queues();
977 			vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1);
978 			vm_page_unlock_queues();
979 		}
980 	}
981 	mtx_unlock(&Giant);
982 }
983 
984 /*
985  *	Routine:
986  *		vm_fault_copy_entry
987  *	Function:
988  *		Copy all of the pages from a wired-down map entry to another.
989  *
990  *	In/out conditions:
991  *		The source and destination maps must be locked for write.
992  *		The source map entry must be wired down (or be a sharing map
993  *		entry corresponding to a main map entry that is wired down).
994  */
995 void
996 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry)
997 	vm_map_t dst_map;
998 	vm_map_t src_map;
999 	vm_map_entry_t dst_entry;
1000 	vm_map_entry_t src_entry;
1001 {
1002 	vm_object_t dst_object;
1003 	vm_object_t src_object;
1004 	vm_ooffset_t dst_offset;
1005 	vm_ooffset_t src_offset;
1006 	vm_prot_t prot;
1007 	vm_offset_t vaddr;
1008 	vm_page_t dst_m;
1009 	vm_page_t src_m;
1010 
1011 #ifdef	lint
1012 	src_map++;
1013 #endif	/* lint */
1014 
1015 	src_object = src_entry->object.vm_object;
1016 	src_offset = src_entry->offset;
1017 
1018 	/*
1019 	 * Create the top-level object for the destination entry. (Doesn't
1020 	 * actually shadow anything - we copy the pages directly.)
1021 	 */
1022 	dst_object = vm_object_allocate(OBJT_DEFAULT,
1023 	    (vm_size_t) OFF_TO_IDX(dst_entry->end - dst_entry->start));
1024 
1025 	dst_entry->object.vm_object = dst_object;
1026 	dst_entry->offset = 0;
1027 
1028 	prot = dst_entry->max_protection;
1029 
1030 	/*
1031 	 * Loop through all of the pages in the entry's range, copying each
1032 	 * one from the source object (it should be there) to the destination
1033 	 * object.
1034 	 */
1035 	for (vaddr = dst_entry->start, dst_offset = 0;
1036 	    vaddr < dst_entry->end;
1037 	    vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) {
1038 
1039 		/*
1040 		 * Allocate a page in the destination object
1041 		 */
1042 		do {
1043 			dst_m = vm_page_alloc(dst_object,
1044 				OFF_TO_IDX(dst_offset), VM_ALLOC_NORMAL);
1045 			if (dst_m == NULL) {
1046 				VM_WAIT;
1047 			}
1048 		} while (dst_m == NULL);
1049 
1050 		/*
1051 		 * Find the page in the source object, and copy it in.
1052 		 * (Because the source is wired down, the page will be in
1053 		 * memory.)
1054 		 */
1055 		src_m = vm_page_lookup(src_object,
1056 			OFF_TO_IDX(dst_offset + src_offset));
1057 		if (src_m == NULL)
1058 			panic("vm_fault_copy_wired: page missing");
1059 
1060 		vm_page_copy(src_m, dst_m);
1061 
1062 		/*
1063 		 * Enter it in the pmap...
1064 		 */
1065 		vm_page_flag_clear(dst_m, PG_ZERO);
1066 		pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE);
1067 		vm_page_lock_queues();
1068 		vm_page_flag_set(dst_m, PG_WRITEABLE);
1069 
1070 		/*
1071 		 * Mark it no longer busy, and put it on the active list.
1072 		 */
1073 		vm_page_activate(dst_m);
1074 		vm_page_wakeup(dst_m);
1075 		vm_page_unlock_queues();
1076 	}
1077 }
1078 
1079 
1080 /*
1081  * This routine checks around the requested page for other pages that
1082  * might be able to be faulted in.  This routine brackets the viable
1083  * pages for the pages to be paged in.
1084  *
1085  * Inputs:
1086  *	m, rbehind, rahead
1087  *
1088  * Outputs:
1089  *  marray (array of vm_page_t), reqpage (index of requested page)
1090  *
1091  * Return value:
1092  *  number of pages in marray
1093  *
1094  * This routine can't block.
1095  */
1096 static int
1097 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage)
1098 	vm_page_t m;
1099 	int rbehind;
1100 	int rahead;
1101 	vm_page_t *marray;
1102 	int *reqpage;
1103 {
1104 	int i,j;
1105 	vm_object_t object;
1106 	vm_pindex_t pindex, startpindex, endpindex, tpindex;
1107 	vm_page_t rtm;
1108 	int cbehind, cahead;
1109 
1110 	GIANT_REQUIRED;
1111 
1112 	object = m->object;
1113 	pindex = m->pindex;
1114 
1115 	/*
1116 	 * we don't fault-ahead for device pager
1117 	 */
1118 	if (object->type == OBJT_DEVICE) {
1119 		*reqpage = 0;
1120 		marray[0] = m;
1121 		return 1;
1122 	}
1123 
1124 	/*
1125 	 * if the requested page is not available, then give up now
1126 	 */
1127 	if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1128 		return 0;
1129 	}
1130 
1131 	if ((cbehind == 0) && (cahead == 0)) {
1132 		*reqpage = 0;
1133 		marray[0] = m;
1134 		return 1;
1135 	}
1136 
1137 	if (rahead > cahead) {
1138 		rahead = cahead;
1139 	}
1140 
1141 	if (rbehind > cbehind) {
1142 		rbehind = cbehind;
1143 	}
1144 
1145 	/*
1146 	 * try to do any readahead that we might have free pages for.
1147 	 */
1148 	if ((rahead + rbehind) >
1149 		((cnt.v_free_count + cnt.v_cache_count) - cnt.v_free_reserved)) {
1150 		pagedaemon_wakeup();
1151 		marray[0] = m;
1152 		*reqpage = 0;
1153 		return 1;
1154 	}
1155 
1156 	/*
1157 	 * scan backward for the read behind pages -- in memory
1158 	 */
1159 	if (pindex > 0) {
1160 		if (rbehind > pindex) {
1161 			rbehind = pindex;
1162 			startpindex = 0;
1163 		} else {
1164 			startpindex = pindex - rbehind;
1165 		}
1166 
1167 		for (tpindex = pindex - 1; tpindex >= startpindex; tpindex -= 1) {
1168 			if (vm_page_lookup(object, tpindex)) {
1169 				startpindex = tpindex + 1;
1170 				break;
1171 			}
1172 			if (tpindex == 0)
1173 				break;
1174 		}
1175 
1176 		for (i = 0, tpindex = startpindex; tpindex < pindex; i++, tpindex++) {
1177 
1178 			rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
1179 			if (rtm == NULL) {
1180 				vm_page_lock_queues();
1181 				for (j = 0; j < i; j++) {
1182 					vm_page_free(marray[j]);
1183 				}
1184 				vm_page_unlock_queues();
1185 				marray[0] = m;
1186 				*reqpage = 0;
1187 				return 1;
1188 			}
1189 
1190 			marray[i] = rtm;
1191 		}
1192 	} else {
1193 		startpindex = 0;
1194 		i = 0;
1195 	}
1196 
1197 	marray[i] = m;
1198 	/* page offset of the required page */
1199 	*reqpage = i;
1200 
1201 	tpindex = pindex + 1;
1202 	i++;
1203 
1204 	/*
1205 	 * scan forward for the read ahead pages
1206 	 */
1207 	endpindex = tpindex + rahead;
1208 	if (endpindex > object->size)
1209 		endpindex = object->size;
1210 
1211 	for (; tpindex < endpindex; i++, tpindex++) {
1212 
1213 		if (vm_page_lookup(object, tpindex)) {
1214 			break;
1215 		}
1216 
1217 		rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
1218 		if (rtm == NULL) {
1219 			break;
1220 		}
1221 
1222 		marray[i] = rtm;
1223 	}
1224 
1225 	/* return number of bytes of pages */
1226 	return i;
1227 }
1228