xref: /dflybsd-src/sys/vm/vm_fault.c (revision d4b8aec4bb44a374c3e91969c1a7df6569da7be3)
1 /*
2  * (MPSAFE)
3  *
4  * Copyright (c) 1991, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  * Copyright (c) 1994 John S. Dyson
7  * All rights reserved.
8  * Copyright (c) 1994 David Greenman
9  * All rights reserved.
10  *
11  *
12  * This code is derived from software contributed to Berkeley by
13  * The Mach Operating System project at Carnegie-Mellon University.
14  *
15  * Redistribution and use in source and binary forms, with or without
16  * modification, are permitted provided that the following conditions
17  * are met:
18  * 1. Redistributions of source code must retain the above copyright
19  *    notice, this list of conditions and the following disclaimer.
20  * 2. Redistributions in binary form must reproduce the above copyright
21  *    notice, this list of conditions and the following disclaimer in the
22  *    documentation and/or other materials provided with the distribution.
23  * 3. All advertising materials mentioning features or use of this software
24  *    must display the following acknowledgement:
25  *	This product includes software developed by the University of
26  *	California, Berkeley and its contributors.
27  * 4. Neither the name of the University nor the names of its contributors
28  *    may be used to endorse or promote products derived from this software
29  *    without specific prior written permission.
30  *
31  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
32  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
33  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
34  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
35  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
36  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
37  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
38  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
39  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
40  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
41  * SUCH DAMAGE.
42  *
43  *	from: @(#)vm_fault.c	8.4 (Berkeley) 1/12/94
44  *
45  *
46  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
47  * All rights reserved.
48  *
49  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
50  *
51  * Permission to use, copy, modify and distribute this software and
52  * its documentation is hereby granted, provided that both the copyright
53  * notice and this permission notice appear in all copies of the
54  * software, derivative works or modified versions, and any portions
55  * thereof, and that both notices appear in supporting documentation.
56  *
57  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
58  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
59  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
60  *
61  * Carnegie Mellon requests users of this software to return to
62  *
63  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
64  *  School of Computer Science
65  *  Carnegie Mellon University
66  *  Pittsburgh PA 15213-3890
67  *
68  * any improvements or extensions that they make and grant Carnegie the
69  * rights to redistribute these changes.
70  *
71  * $FreeBSD: src/sys/vm/vm_fault.c,v 1.108.2.8 2002/02/26 05:49:27 silby Exp $
72  * $DragonFly: src/sys/vm/vm_fault.c,v 1.47 2008/07/01 02:02:56 dillon Exp $
73  */
74 
75 /*
76  *	Page fault handling module.
77  */
78 
79 #include <sys/param.h>
80 #include <sys/systm.h>
81 #include <sys/kernel.h>
82 #include <sys/proc.h>
83 #include <sys/vnode.h>
84 #include <sys/resourcevar.h>
85 #include <sys/vmmeter.h>
86 #include <sys/vkernel.h>
87 #include <sys/lock.h>
88 #include <sys/sysctl.h>
89 
90 #include <cpu/lwbuf.h>
91 
92 #include <vm/vm.h>
93 #include <vm/vm_param.h>
94 #include <vm/pmap.h>
95 #include <vm/vm_map.h>
96 #include <vm/vm_object.h>
97 #include <vm/vm_page.h>
98 #include <vm/vm_pageout.h>
99 #include <vm/vm_kern.h>
100 #include <vm/vm_pager.h>
101 #include <vm/vnode_pager.h>
102 #include <vm/vm_extern.h>
103 
104 #include <sys/thread2.h>
105 #include <vm/vm_page2.h>
106 
107 struct faultstate {
108 	vm_page_t m;
109 	vm_object_t object;
110 	vm_pindex_t pindex;
111 	vm_prot_t prot;
112 	vm_page_t first_m;
113 	vm_object_t first_object;
114 	vm_prot_t first_prot;
115 	vm_map_t map;
116 	vm_map_entry_t entry;
117 	int lookup_still_valid;
118 	int didlimit;
119 	int hardfault;
120 	int fault_flags;
121 	int map_generation;
122 	boolean_t wired;
123 	struct vnode *vp;
124 };
125 
126 static int vm_fast_fault = 1;
127 SYSCTL_INT(_vm, OID_AUTO, fast_fault, CTLFLAG_RW, &vm_fast_fault, 0,
128 	   "Burst fault zero-fill regions");
129 static int debug_cluster = 0;
130 SYSCTL_INT(_vm, OID_AUTO, debug_cluster, CTLFLAG_RW, &debug_cluster, 0, "");
131 
132 static int vm_fault_object(struct faultstate *, vm_pindex_t, vm_prot_t);
133 static int vm_fault_vpagetable(struct faultstate *, vm_pindex_t *, vpte_t, int);
134 #if 0
135 static int vm_fault_additional_pages (vm_page_t, int, int, vm_page_t *, int *);
136 #endif
137 static int vm_fault_ratelimit(struct vmspace *);
138 static void vm_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry,
139 			int prot);
140 
141 /*
142  * The caller must hold vm_token.
143  */
144 static __inline void
145 release_page(struct faultstate *fs)
146 {
147 	vm_page_deactivate(fs->m);
148 	vm_page_wakeup(fs->m);
149 	fs->m = NULL;
150 }
151 
152 /*
153  * The caller must hold vm_token.
154  */
155 static __inline void
156 unlock_map(struct faultstate *fs)
157 {
158 	if (fs->lookup_still_valid && fs->map) {
159 		vm_map_lookup_done(fs->map, fs->entry, 0);
160 		fs->lookup_still_valid = FALSE;
161 	}
162 }
163 
164 /*
165  * Clean up after a successful call to vm_fault_object() so another call
166  * to vm_fault_object() can be made.
167  *
168  * The caller must hold vm_token.
169  */
170 static void
171 _cleanup_successful_fault(struct faultstate *fs, int relock)
172 {
173 	if (fs->object != fs->first_object) {
174 		vm_page_free(fs->first_m);
175 		vm_object_pip_wakeup(fs->object);
176 		fs->first_m = NULL;
177 	}
178 	fs->object = fs->first_object;
179 	if (relock && fs->lookup_still_valid == FALSE) {
180 		if (fs->map)
181 			vm_map_lock_read(fs->map);
182 		fs->lookup_still_valid = TRUE;
183 	}
184 }
185 
186 /*
187  * The caller must hold vm_token.
188  */
189 static void
190 _unlock_things(struct faultstate *fs, int dealloc)
191 {
192 	vm_object_pip_wakeup(fs->first_object);
193 	_cleanup_successful_fault(fs, 0);
194 	if (dealloc) {
195 		vm_object_deallocate(fs->first_object);
196 		fs->first_object = NULL;
197 	}
198 	unlock_map(fs);
199 	if (fs->vp != NULL) {
200 		vput(fs->vp);
201 		fs->vp = NULL;
202 	}
203 }
204 
205 #define unlock_things(fs) _unlock_things(fs, 0)
206 #define unlock_and_deallocate(fs) _unlock_things(fs, 1)
207 #define cleanup_successful_fault(fs) _cleanup_successful_fault(fs, 1)
208 
209 /*
210  * TRYPAGER
211  *
212  * Determine if the pager for the current object *might* contain the page.
213  *
214  * We only need to try the pager if this is not a default object (default
215  * objects are zero-fill and have no real pager), and if we are not taking
216  * a wiring fault or if the FS entry is wired.
217  */
218 #define TRYPAGER(fs)	\
219 		(fs->object->type != OBJT_DEFAULT && \
220 		(((fs->fault_flags & VM_FAULT_WIRE_MASK) == 0) || fs->wired))
221 
222 /*
223  * vm_fault:
224  *
225  * Handle a page fault occuring at the given address, requiring the given
226  * permissions, in the map specified.  If successful, the page is inserted
227  * into the associated physical map.
228  *
229  * NOTE: The given address should be truncated to the proper page address.
230  *
231  * KERN_SUCCESS is returned if the page fault is handled; otherwise,
232  * a standard error specifying why the fault is fatal is returned.
233  *
234  * The map in question must be referenced, and remains so.
235  * The caller may hold no locks.
236  * No other requirements.
237  */
238 int
239 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, int fault_flags)
240 {
241 	int result;
242 	vm_pindex_t first_pindex;
243 	struct faultstate fs;
244 	int growstack;
245 
246 	mycpu->gd_cnt.v_vm_faults++;
247 
248 	fs.didlimit = 0;
249 	fs.hardfault = 0;
250 	fs.fault_flags = fault_flags;
251 	growstack = 1;
252 
253 RetryFault:
254 	/*
255 	 * Find the vm_map_entry representing the backing store and resolve
256 	 * the top level object and page index.  This may have the side
257 	 * effect of executing a copy-on-write on the map entry and/or
258 	 * creating a shadow object, but will not COW any actual VM pages.
259 	 *
260 	 * On success fs.map is left read-locked and various other fields
261 	 * are initialized but not otherwise referenced or locked.
262 	 *
263 	 * NOTE!  vm_map_lookup will try to upgrade the fault_type to
264 	 * VM_FAULT_WRITE if the map entry is a virtual page table and also
265 	 * writable, so we can set the 'A'accessed bit in the virtual page
266 	 * table entry.
267 	 */
268 	fs.map = map;
269 	result = vm_map_lookup(&fs.map, vaddr, fault_type,
270 			       &fs.entry, &fs.first_object,
271 			       &first_pindex, &fs.first_prot, &fs.wired);
272 
273 	/*
274 	 * If the lookup failed or the map protections are incompatible,
275 	 * the fault generally fails.  However, if the caller is trying
276 	 * to do a user wiring we have more work to do.
277 	 */
278 	if (result != KERN_SUCCESS) {
279 		if (result != KERN_PROTECTION_FAILURE ||
280 		    (fs.fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE)
281 		{
282 			if (result == KERN_INVALID_ADDRESS && growstack &&
283 			    map != &kernel_map && curproc != NULL) {
284 				result = vm_map_growstack(curproc, vaddr);
285 				if (result != KERN_SUCCESS)
286 					return (KERN_FAILURE);
287 				growstack = 0;
288 				goto RetryFault;
289 			}
290 			return (result);
291 		}
292 
293 		/*
294    		 * If we are user-wiring a r/w segment, and it is COW, then
295    		 * we need to do the COW operation.  Note that we don't
296 		 * currently COW RO sections now, because it is NOT desirable
297    		 * to COW .text.  We simply keep .text from ever being COW'ed
298    		 * and take the heat that one cannot debug wired .text sections.
299    		 */
300 		result = vm_map_lookup(&fs.map, vaddr,
301 				       VM_PROT_READ|VM_PROT_WRITE|
302 				        VM_PROT_OVERRIDE_WRITE,
303 				       &fs.entry, &fs.first_object,
304 				       &first_pindex, &fs.first_prot,
305 				       &fs.wired);
306 		if (result != KERN_SUCCESS)
307 			return result;
308 
309 		/*
310 		 * If we don't COW now, on a user wire, the user will never
311 		 * be able to write to the mapping.  If we don't make this
312 		 * restriction, the bookkeeping would be nearly impossible.
313 		 */
314 		if ((fs.entry->protection & VM_PROT_WRITE) == 0)
315 			fs.entry->max_protection &= ~VM_PROT_WRITE;
316 	}
317 
318 	/*
319 	 * fs.map is read-locked
320 	 *
321 	 * Misc checks.  Save the map generation number to detect races.
322 	 */
323 	fs.map_generation = fs.map->timestamp;
324 
325 	if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
326 		panic("vm_fault: fault on nofault entry, addr: %lx",
327 		    (u_long)vaddr);
328 	}
329 
330 	/*
331 	 * A system map entry may return a NULL object.  No object means
332 	 * no pager means an unrecoverable kernel fault.
333 	 */
334 	if (fs.first_object == NULL) {
335 		panic("vm_fault: unrecoverable fault at %p in entry %p",
336 			(void *)vaddr, fs.entry);
337 	}
338 
339 	/*
340 	 * Make a reference to this object to prevent its disposal while we
341 	 * are messing with it.  Once we have the reference, the map is free
342 	 * to be diddled.  Since objects reference their shadows (and copies),
343 	 * they will stay around as well.
344 	 *
345 	 * Bump the paging-in-progress count to prevent size changes (e.g.
346 	 * truncation operations) during I/O.  This must be done after
347 	 * obtaining the vnode lock in order to avoid possible deadlocks.
348 	 *
349 	 * The vm_token is needed to manipulate the vm_object
350 	 */
351 	lwkt_gettoken(&vm_token);
352 	vm_object_reference(fs.first_object);
353 	fs.vp = vnode_pager_lock(fs.first_object);
354 	vm_object_pip_add(fs.first_object, 1);
355 	lwkt_reltoken(&vm_token);
356 
357 	fs.lookup_still_valid = TRUE;
358 	fs.first_m = NULL;
359 	fs.object = fs.first_object;	/* so unlock_and_deallocate works */
360 
361 	/*
362 	 * If the entry is wired we cannot change the page protection.
363 	 */
364 	if (fs.wired)
365 		fault_type = fs.first_prot;
366 
367 	/*
368 	 * The page we want is at (first_object, first_pindex), but if the
369 	 * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the
370 	 * page table to figure out the actual pindex.
371 	 *
372 	 * NOTE!  DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION
373 	 * ONLY
374 	 */
375 	if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) {
376 		result = vm_fault_vpagetable(&fs, &first_pindex,
377 					     fs.entry->aux.master_pde,
378 					     fault_type);
379 		if (result == KERN_TRY_AGAIN)
380 			goto RetryFault;
381 		if (result != KERN_SUCCESS)
382 			return (result);
383 	}
384 
385 	/*
386 	 * Now we have the actual (object, pindex), fault in the page.  If
387 	 * vm_fault_object() fails it will unlock and deallocate the FS
388 	 * data.   If it succeeds everything remains locked and fs->object
389 	 * will have an additional PIP count if it is not equal to
390 	 * fs->first_object
391 	 *
392 	 * vm_fault_object will set fs->prot for the pmap operation.  It is
393 	 * allowed to set VM_PROT_WRITE if fault_type == VM_PROT_READ if the
394 	 * page can be safely written.  However, it will force a read-only
395 	 * mapping for a read fault if the memory is managed by a virtual
396 	 * page table.
397 	 */
398 	result = vm_fault_object(&fs, first_pindex, fault_type);
399 
400 	if (result == KERN_TRY_AGAIN)
401 		goto RetryFault;
402 	if (result != KERN_SUCCESS)
403 		return (result);
404 
405 	/*
406 	 * On success vm_fault_object() does not unlock or deallocate, and fs.m
407 	 * will contain a busied page.
408 	 *
409 	 * Enter the page into the pmap and do pmap-related adjustments.
410 	 */
411 	pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot, fs.wired);
412 
413 	/*
414 	 * Burst in a few more pages if possible.  The fs.map should still
415 	 * be locked.
416 	 */
417 	if (fault_flags & VM_FAULT_BURST) {
418 		if ((fs.fault_flags & VM_FAULT_WIRE_MASK) == 0 &&
419 		    fs.wired == 0) {
420 			vm_prefault(fs.map->pmap, vaddr, fs.entry, fs.prot);
421 		}
422 	}
423 	unlock_things(&fs);
424 
425 	vm_page_flag_clear(fs.m, PG_ZERO);
426 	vm_page_flag_set(fs.m, PG_REFERENCED);
427 
428 	/*
429 	 * If the page is not wired down, then put it where the pageout daemon
430 	 * can find it.
431 	 *
432 	 * We do not really need to get vm_token here but since all the
433 	 * vm_*() calls have to doing it here improves efficiency.
434 	 */
435 	lwkt_gettoken(&vm_token);
436 	if (fs.fault_flags & VM_FAULT_WIRE_MASK) {
437 		if (fs.wired)
438 			vm_page_wire(fs.m);
439 		else
440 			vm_page_unwire(fs.m, 1);
441 	} else {
442 		vm_page_activate(fs.m);
443 	}
444 
445 	if (curthread->td_lwp) {
446 		if (fs.hardfault) {
447 			curthread->td_lwp->lwp_ru.ru_majflt++;
448 		} else {
449 			curthread->td_lwp->lwp_ru.ru_minflt++;
450 		}
451 	}
452 
453 	/*
454 	 * Unlock everything, and return
455 	 */
456 	vm_page_wakeup(fs.m);
457 	vm_object_deallocate(fs.first_object);
458 	lwkt_reltoken(&vm_token);
459 
460 	return (KERN_SUCCESS);
461 }
462 
463 /*
464  * Fault in the specified virtual address in the current process map,
465  * returning a held VM page or NULL.  See vm_fault_page() for more
466  * information.
467  *
468  * No requirements.
469  */
470 vm_page_t
471 vm_fault_page_quick(vm_offset_t va, vm_prot_t fault_type, int *errorp)
472 {
473 	struct lwp *lp = curthread->td_lwp;
474 	vm_page_t m;
475 
476 	m = vm_fault_page(&lp->lwp_vmspace->vm_map, va,
477 			  fault_type, VM_FAULT_NORMAL, errorp);
478 	return(m);
479 }
480 
481 /*
482  * Fault in the specified virtual address in the specified map, doing all
483  * necessary manipulation of the object store and all necessary I/O.  Return
484  * a held VM page or NULL, and set *errorp.  The related pmap is not
485  * updated.
486  *
487  * The returned page will be properly dirtied if VM_PROT_WRITE was specified,
488  * and marked PG_REFERENCED as well.
489  *
490  * If the page cannot be faulted writable and VM_PROT_WRITE was specified, an
491  * error will be returned.
492  *
493  * No requirements.
494  */
495 vm_page_t
496 vm_fault_page(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
497 	      int fault_flags, int *errorp)
498 {
499 	vm_pindex_t first_pindex;
500 	struct faultstate fs;
501 	int result;
502 	vm_prot_t orig_fault_type = fault_type;
503 
504 	mycpu->gd_cnt.v_vm_faults++;
505 
506 	fs.didlimit = 0;
507 	fs.hardfault = 0;
508 	fs.fault_flags = fault_flags;
509 	KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0);
510 
511 RetryFault:
512 	/*
513 	 * Find the vm_map_entry representing the backing store and resolve
514 	 * the top level object and page index.  This may have the side
515 	 * effect of executing a copy-on-write on the map entry and/or
516 	 * creating a shadow object, but will not COW any actual VM pages.
517 	 *
518 	 * On success fs.map is left read-locked and various other fields
519 	 * are initialized but not otherwise referenced or locked.
520 	 *
521 	 * NOTE!  vm_map_lookup will upgrade the fault_type to VM_FAULT_WRITE
522 	 * if the map entry is a virtual page table and also writable,
523 	 * so we can set the 'A'accessed bit in the virtual page table entry.
524 	 */
525 	fs.map = map;
526 	result = vm_map_lookup(&fs.map, vaddr, fault_type,
527 			       &fs.entry, &fs.first_object,
528 			       &first_pindex, &fs.first_prot, &fs.wired);
529 
530 	if (result != KERN_SUCCESS) {
531 		*errorp = result;
532 		return (NULL);
533 	}
534 
535 	/*
536 	 * fs.map is read-locked
537 	 *
538 	 * Misc checks.  Save the map generation number to detect races.
539 	 */
540 	fs.map_generation = fs.map->timestamp;
541 
542 	if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
543 		panic("vm_fault: fault on nofault entry, addr: %lx",
544 		    (u_long)vaddr);
545 	}
546 
547 	/*
548 	 * A system map entry may return a NULL object.  No object means
549 	 * no pager means an unrecoverable kernel fault.
550 	 */
551 	if (fs.first_object == NULL) {
552 		panic("vm_fault: unrecoverable fault at %p in entry %p",
553 			(void *)vaddr, fs.entry);
554 	}
555 
556 	/*
557 	 * Make a reference to this object to prevent its disposal while we
558 	 * are messing with it.  Once we have the reference, the map is free
559 	 * to be diddled.  Since objects reference their shadows (and copies),
560 	 * they will stay around as well.
561 	 *
562 	 * Bump the paging-in-progress count to prevent size changes (e.g.
563 	 * truncation operations) during I/O.  This must be done after
564 	 * obtaining the vnode lock in order to avoid possible deadlocks.
565 	 *
566 	 * The vm_token is needed to manipulate the vm_object
567 	 */
568 	lwkt_gettoken(&vm_token);
569 	vm_object_reference(fs.first_object);
570 	fs.vp = vnode_pager_lock(fs.first_object);
571 	vm_object_pip_add(fs.first_object, 1);
572 	lwkt_reltoken(&vm_token);
573 
574 	fs.lookup_still_valid = TRUE;
575 	fs.first_m = NULL;
576 	fs.object = fs.first_object;	/* so unlock_and_deallocate works */
577 
578 	/*
579 	 * If the entry is wired we cannot change the page protection.
580 	 */
581 	if (fs.wired)
582 		fault_type = fs.first_prot;
583 
584 	/*
585 	 * The page we want is at (first_object, first_pindex), but if the
586 	 * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the
587 	 * page table to figure out the actual pindex.
588 	 *
589 	 * NOTE!  DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION
590 	 * ONLY
591 	 */
592 	if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) {
593 		result = vm_fault_vpagetable(&fs, &first_pindex,
594 					     fs.entry->aux.master_pde,
595 					     fault_type);
596 		if (result == KERN_TRY_AGAIN)
597 			goto RetryFault;
598 		if (result != KERN_SUCCESS) {
599 			*errorp = result;
600 			return (NULL);
601 		}
602 	}
603 
604 	/*
605 	 * Now we have the actual (object, pindex), fault in the page.  If
606 	 * vm_fault_object() fails it will unlock and deallocate the FS
607 	 * data.   If it succeeds everything remains locked and fs->object
608 	 * will have an additinal PIP count if it is not equal to
609 	 * fs->first_object
610 	 */
611 	result = vm_fault_object(&fs, first_pindex, fault_type);
612 
613 	if (result == KERN_TRY_AGAIN)
614 		goto RetryFault;
615 	if (result != KERN_SUCCESS) {
616 		*errorp = result;
617 		return(NULL);
618 	}
619 
620 	if ((orig_fault_type & VM_PROT_WRITE) &&
621 	    (fs.prot & VM_PROT_WRITE) == 0) {
622 		*errorp = KERN_PROTECTION_FAILURE;
623 		unlock_and_deallocate(&fs);
624 		return(NULL);
625 	}
626 
627 	/*
628 	 * On success vm_fault_object() does not unlock or deallocate, and fs.m
629 	 * will contain a busied page.
630 	 */
631 	unlock_things(&fs);
632 
633 	/*
634 	 * Return a held page.  We are not doing any pmap manipulation so do
635 	 * not set PG_MAPPED.  However, adjust the page flags according to
636 	 * the fault type because the caller may not use a managed pmapping
637 	 * (so we don't want to lose the fact that the page will be dirtied
638 	 * if a write fault was specified).
639 	 */
640 	lwkt_gettoken(&vm_token);
641 	vm_page_hold(fs.m);
642 	vm_page_flag_clear(fs.m, PG_ZERO);
643 	if (fault_type & VM_PROT_WRITE)
644 		vm_page_dirty(fs.m);
645 
646 	/*
647 	 * Update the pmap.  We really only have to do this if a COW
648 	 * occured to replace the read-only page with the new page.  For
649 	 * now just do it unconditionally. XXX
650 	 */
651 	pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot, fs.wired);
652 	vm_page_flag_set(fs.m, PG_REFERENCED);
653 
654 	/*
655 	 * Unbusy the page by activating it.  It remains held and will not
656 	 * be reclaimed.
657 	 */
658 	vm_page_activate(fs.m);
659 
660 	if (curthread->td_lwp) {
661 		if (fs.hardfault) {
662 			curthread->td_lwp->lwp_ru.ru_majflt++;
663 		} else {
664 			curthread->td_lwp->lwp_ru.ru_minflt++;
665 		}
666 	}
667 
668 	/*
669 	 * Unlock everything, and return the held page.
670 	 */
671 	vm_page_wakeup(fs.m);
672 	vm_object_deallocate(fs.first_object);
673 	lwkt_reltoken(&vm_token);
674 
675 	*errorp = 0;
676 	return(fs.m);
677 }
678 
679 /*
680  * Fault in the specified (object,offset), dirty the returned page as
681  * needed.  If the requested fault_type cannot be done NULL and an
682  * error is returned.
683  *
684  * A held (but not busied) page is returned.
685  *
686  * No requirements.
687  */
688 vm_page_t
689 vm_fault_object_page(vm_object_t object, vm_ooffset_t offset,
690 		     vm_prot_t fault_type, int fault_flags, int *errorp)
691 {
692 	int result;
693 	vm_pindex_t first_pindex;
694 	struct faultstate fs;
695 	struct vm_map_entry entry;
696 
697 	bzero(&entry, sizeof(entry));
698 	entry.object.vm_object = object;
699 	entry.maptype = VM_MAPTYPE_NORMAL;
700 	entry.protection = entry.max_protection = fault_type;
701 
702 	fs.didlimit = 0;
703 	fs.hardfault = 0;
704 	fs.fault_flags = fault_flags;
705 	fs.map = NULL;
706 	KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0);
707 
708 RetryFault:
709 
710 	fs.first_object = object;
711 	first_pindex = OFF_TO_IDX(offset);
712 	fs.entry = &entry;
713 	fs.first_prot = fault_type;
714 	fs.wired = 0;
715 	/*fs.map_generation = 0; unused */
716 
717 	/*
718 	 * Make a reference to this object to prevent its disposal while we
719 	 * are messing with it.  Once we have the reference, the map is free
720 	 * to be diddled.  Since objects reference their shadows (and copies),
721 	 * they will stay around as well.
722 	 *
723 	 * Bump the paging-in-progress count to prevent size changes (e.g.
724 	 * truncation operations) during I/O.  This must be done after
725 	 * obtaining the vnode lock in order to avoid possible deadlocks.
726 	 */
727 	lwkt_gettoken(&vm_token);
728 	vm_object_reference(fs.first_object);
729 	fs.vp = vnode_pager_lock(fs.first_object);
730 	vm_object_pip_add(fs.first_object, 1);
731 	lwkt_reltoken(&vm_token);
732 
733 	fs.lookup_still_valid = TRUE;
734 	fs.first_m = NULL;
735 	fs.object = fs.first_object;	/* so unlock_and_deallocate works */
736 
737 #if 0
738 	/* XXX future - ability to operate on VM object using vpagetable */
739 	if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) {
740 		result = vm_fault_vpagetable(&fs, &first_pindex,
741 					     fs.entry->aux.master_pde,
742 					     fault_type);
743 		if (result == KERN_TRY_AGAIN)
744 			goto RetryFault;
745 		if (result != KERN_SUCCESS) {
746 			*errorp = result;
747 			return (NULL);
748 		}
749 	}
750 #endif
751 
752 	/*
753 	 * Now we have the actual (object, pindex), fault in the page.  If
754 	 * vm_fault_object() fails it will unlock and deallocate the FS
755 	 * data.   If it succeeds everything remains locked and fs->object
756 	 * will have an additinal PIP count if it is not equal to
757 	 * fs->first_object
758 	 */
759 	result = vm_fault_object(&fs, first_pindex, fault_type);
760 
761 	if (result == KERN_TRY_AGAIN)
762 		goto RetryFault;
763 	if (result != KERN_SUCCESS) {
764 		*errorp = result;
765 		return(NULL);
766 	}
767 
768 	if ((fault_type & VM_PROT_WRITE) && (fs.prot & VM_PROT_WRITE) == 0) {
769 		*errorp = KERN_PROTECTION_FAILURE;
770 		unlock_and_deallocate(&fs);
771 		return(NULL);
772 	}
773 
774 	/*
775 	 * On success vm_fault_object() does not unlock or deallocate, and fs.m
776 	 * will contain a busied page.
777 	 */
778 	unlock_things(&fs);
779 
780 	/*
781 	 * Return a held page.  We are not doing any pmap manipulation so do
782 	 * not set PG_MAPPED.  However, adjust the page flags according to
783 	 * the fault type because the caller may not use a managed pmapping
784 	 * (so we don't want to lose the fact that the page will be dirtied
785 	 * if a write fault was specified).
786 	 */
787 	lwkt_gettoken(&vm_token);
788 	vm_page_hold(fs.m);
789 	vm_page_flag_clear(fs.m, PG_ZERO);
790 	if (fault_type & VM_PROT_WRITE)
791 		vm_page_dirty(fs.m);
792 
793 	/*
794 	 * Indicate that the page was accessed.
795 	 */
796 	vm_page_flag_set(fs.m, PG_REFERENCED);
797 
798 	/*
799 	 * Unbusy the page by activating it.  It remains held and will not
800 	 * be reclaimed.
801 	 */
802 	vm_page_activate(fs.m);
803 
804 	if (curthread->td_lwp) {
805 		if (fs.hardfault) {
806 			mycpu->gd_cnt.v_vm_faults++;
807 			curthread->td_lwp->lwp_ru.ru_majflt++;
808 		} else {
809 			curthread->td_lwp->lwp_ru.ru_minflt++;
810 		}
811 	}
812 
813 	/*
814 	 * Unlock everything, and return the held page.
815 	 */
816 	vm_page_wakeup(fs.m);
817 	vm_object_deallocate(fs.first_object);
818 	lwkt_reltoken(&vm_token);
819 
820 	*errorp = 0;
821 	return(fs.m);
822 }
823 
824 /*
825  * Translate the virtual page number (first_pindex) that is relative
826  * to the address space into a logical page number that is relative to the
827  * backing object.  Use the virtual page table pointed to by (vpte).
828  *
829  * This implements an N-level page table.  Any level can terminate the
830  * scan by setting VPTE_PS.   A linear mapping is accomplished by setting
831  * VPTE_PS in the master page directory entry set via mcontrol(MADV_SETMAP).
832  *
833  * No requirements (vm_token need not be held).
834  */
835 static
836 int
837 vm_fault_vpagetable(struct faultstate *fs, vm_pindex_t *pindex,
838 		    vpte_t vpte, int fault_type)
839 {
840 	struct lwbuf *lwb;
841 	int vshift = VPTE_FRAME_END - PAGE_SHIFT; /* index bits remaining */
842 	int result = KERN_SUCCESS;
843 	vpte_t *ptep;
844 
845 	for (;;) {
846 		/*
847 		 * We cannot proceed if the vpte is not valid, not readable
848 		 * for a read fault, or not writable for a write fault.
849 		 */
850 		if ((vpte & VPTE_V) == 0) {
851 			unlock_and_deallocate(fs);
852 			return (KERN_FAILURE);
853 		}
854 		if ((fault_type & VM_PROT_READ) && (vpte & VPTE_R) == 0) {
855 			unlock_and_deallocate(fs);
856 			return (KERN_FAILURE);
857 		}
858 		if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_W) == 0) {
859 			unlock_and_deallocate(fs);
860 			return (KERN_FAILURE);
861 		}
862 		if ((vpte & VPTE_PS) || vshift == 0)
863 			break;
864 		KKASSERT(vshift >= VPTE_PAGE_BITS);
865 
866 		/*
867 		 * Get the page table page.  Nominally we only read the page
868 		 * table, but since we are actively setting VPTE_M and VPTE_A,
869 		 * tell vm_fault_object() that we are writing it.
870 		 *
871 		 * There is currently no real need to optimize this.
872 		 */
873 		result = vm_fault_object(fs, (vpte & VPTE_FRAME) >> PAGE_SHIFT,
874 					 VM_PROT_READ|VM_PROT_WRITE);
875 		if (result != KERN_SUCCESS)
876 			return (result);
877 
878 		/*
879 		 * Process the returned fs.m and look up the page table
880 		 * entry in the page table page.
881 		 */
882 		vshift -= VPTE_PAGE_BITS;
883 		lwb = lwbuf_alloc(fs->m);
884 		ptep = ((vpte_t *)lwbuf_kva(lwb) +
885 		        ((*pindex >> vshift) & VPTE_PAGE_MASK));
886 		vpte = *ptep;
887 
888 		/*
889 		 * Page table write-back.  If the vpte is valid for the
890 		 * requested operation, do a write-back to the page table.
891 		 *
892 		 * XXX VPTE_M is not set properly for page directory pages.
893 		 * It doesn't get set in the page directory if the page table
894 		 * is modified during a read access.
895 		 */
896 		if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_V) &&
897 		    (vpte & VPTE_W)) {
898 			if ((vpte & (VPTE_M|VPTE_A)) != (VPTE_M|VPTE_A)) {
899 				atomic_set_long(ptep, VPTE_M | VPTE_A);
900 				vm_page_dirty(fs->m);
901 			}
902 		}
903 		if ((fault_type & VM_PROT_READ) && (vpte & VPTE_V) &&
904 		    (vpte & VPTE_R)) {
905 			if ((vpte & VPTE_A) == 0) {
906 				atomic_set_long(ptep, VPTE_A);
907 				vm_page_dirty(fs->m);
908 			}
909 		}
910 		lwbuf_free(lwb);
911 		vm_page_flag_set(fs->m, PG_REFERENCED);
912 		vm_page_activate(fs->m);
913 		vm_page_wakeup(fs->m);
914 		cleanup_successful_fault(fs);
915 	}
916 	/*
917 	 * Combine remaining address bits with the vpte.
918 	 */
919 	/* JG how many bits from each? */
920 	*pindex = ((vpte & VPTE_FRAME) >> PAGE_SHIFT) +
921 		  (*pindex & ((1L << vshift) - 1));
922 	return (KERN_SUCCESS);
923 }
924 
925 
926 /*
927  * This is the core of the vm_fault code.
928  *
929  * Do all operations required to fault-in (fs.first_object, pindex).  Run
930  * through the shadow chain as necessary and do required COW or virtual
931  * copy operations.  The caller has already fully resolved the vm_map_entry
932  * and, if appropriate, has created a copy-on-write layer.  All we need to
933  * do is iterate the object chain.
934  *
935  * On failure (fs) is unlocked and deallocated and the caller may return or
936  * retry depending on the failure code.  On success (fs) is NOT unlocked or
937  * deallocated, fs.m will contained a resolved, busied page, and fs.object
938  * will have an additional PIP count if it is not equal to fs.first_object.
939  *
940  * No requirements.
941  */
942 static
943 int
944 vm_fault_object(struct faultstate *fs,
945 		vm_pindex_t first_pindex, vm_prot_t fault_type)
946 {
947 	vm_object_t next_object;
948 	vm_pindex_t pindex;
949 
950 	fs->prot = fs->first_prot;
951 	fs->object = fs->first_object;
952 	pindex = first_pindex;
953 
954 	/*
955 	 * If a read fault occurs we try to make the page writable if
956 	 * possible.  There are three cases where we cannot make the
957 	 * page mapping writable:
958 	 *
959 	 * (1) The mapping is read-only or the VM object is read-only,
960 	 *     fs->prot above will simply not have VM_PROT_WRITE set.
961 	 *
962 	 * (2) If the mapping is a virtual page table we need to be able
963 	 *     to detect writes so we can set VPTE_M in the virtual page
964 	 *     table.
965 	 *
966 	 * (3) If the VM page is read-only or copy-on-write, upgrading would
967 	 *     just result in an unnecessary COW fault.
968 	 *
969 	 * VM_PROT_VPAGED is set if faulting via a virtual page table and
970 	 * causes adjustments to the 'M'odify bit to also turn off write
971 	 * access to force a re-fault.
972 	 */
973 	if (fs->entry->maptype == VM_MAPTYPE_VPAGETABLE) {
974 		if ((fault_type & VM_PROT_WRITE) == 0)
975 			fs->prot &= ~VM_PROT_WRITE;
976 	}
977 
978 	lwkt_gettoken(&vm_token);
979 
980 	for (;;) {
981 		/*
982 		 * If the object is dead, we stop here
983 		 */
984 		if (fs->object->flags & OBJ_DEAD) {
985 			unlock_and_deallocate(fs);
986 			lwkt_reltoken(&vm_token);
987 			return (KERN_PROTECTION_FAILURE);
988 		}
989 
990 		/*
991 		 * See if page is resident.  spl protection is required
992 		 * to avoid an interrupt unbusy/free race against our
993 		 * lookup.  We must hold the protection through a page
994 		 * allocation or busy.
995 		 */
996 		crit_enter();
997 		fs->m = vm_page_lookup(fs->object, pindex);
998 		if (fs->m != NULL) {
999 			int queue;
1000 			/*
1001 			 * Wait/Retry if the page is busy.  We have to do this
1002 			 * if the page is busy via either PG_BUSY or
1003 			 * vm_page_t->busy because the vm_pager may be using
1004 			 * vm_page_t->busy for pageouts ( and even pageins if
1005 			 * it is the vnode pager ), and we could end up trying
1006 			 * to pagein and pageout the same page simultaneously.
1007 			 *
1008 			 * We can theoretically allow the busy case on a read
1009 			 * fault if the page is marked valid, but since such
1010 			 * pages are typically already pmap'd, putting that
1011 			 * special case in might be more effort then it is
1012 			 * worth.  We cannot under any circumstances mess
1013 			 * around with a vm_page_t->busy page except, perhaps,
1014 			 * to pmap it.
1015 			 */
1016 			if ((fs->m->flags & PG_BUSY) || fs->m->busy) {
1017 				unlock_things(fs);
1018 				vm_page_sleep_busy(fs->m, TRUE, "vmpfw");
1019 				mycpu->gd_cnt.v_intrans++;
1020 				vm_object_deallocate(fs->first_object);
1021 				fs->first_object = NULL;
1022 				lwkt_reltoken(&vm_token);
1023 				crit_exit();
1024 				return (KERN_TRY_AGAIN);
1025 			}
1026 
1027 			/*
1028 			 * If reactivating a page from PQ_CACHE we may have
1029 			 * to rate-limit.
1030 			 */
1031 			queue = fs->m->queue;
1032 			vm_page_unqueue_nowakeup(fs->m);
1033 
1034 			if ((queue - fs->m->pc) == PQ_CACHE &&
1035 			    vm_page_count_severe()) {
1036 				vm_page_activate(fs->m);
1037 				unlock_and_deallocate(fs);
1038 				vm_waitpfault();
1039 				lwkt_reltoken(&vm_token);
1040 				crit_exit();
1041 				return (KERN_TRY_AGAIN);
1042 			}
1043 
1044 			/*
1045 			 * Mark page busy for other processes, and the
1046 			 * pagedaemon.  If it still isn't completely valid
1047 			 * (readable), or if a read-ahead-mark is set on
1048 			 * the VM page, jump to readrest, else we found the
1049 			 * page and can return.
1050 			 *
1051 			 * We can release the spl once we have marked the
1052 			 * page busy.
1053 			 */
1054 			vm_page_busy(fs->m);
1055 			crit_exit();
1056 
1057 			if (fs->m->object != &kernel_object) {
1058 				if ((fs->m->valid & VM_PAGE_BITS_ALL) !=
1059 				    VM_PAGE_BITS_ALL) {
1060 					goto readrest;
1061 				}
1062 				if (fs->m->flags & PG_RAM) {
1063 					if (debug_cluster)
1064 						kprintf("R");
1065 					vm_page_flag_clear(fs->m, PG_RAM);
1066 					goto readrest;
1067 				}
1068 			}
1069 			break; /* break to PAGE HAS BEEN FOUND */
1070 		}
1071 
1072 		/*
1073 		 * Page is not resident, If this is the search termination
1074 		 * or the pager might contain the page, allocate a new page.
1075 		 *
1076 		 * NOTE: We are still in a critical section.
1077 		 */
1078 		if (TRYPAGER(fs) || fs->object == fs->first_object) {
1079 			/*
1080 			 * If the page is beyond the object size we fail
1081 			 */
1082 			if (pindex >= fs->object->size) {
1083 				lwkt_reltoken(&vm_token);
1084 				crit_exit();
1085 				unlock_and_deallocate(fs);
1086 				return (KERN_PROTECTION_FAILURE);
1087 			}
1088 
1089 			/*
1090 			 * Ratelimit.
1091 			 */
1092 			if (fs->didlimit == 0 && curproc != NULL) {
1093 				int limticks;
1094 
1095 				limticks = vm_fault_ratelimit(curproc->p_vmspace);
1096 				if (limticks) {
1097 					lwkt_reltoken(&vm_token);
1098 					crit_exit();
1099 					unlock_and_deallocate(fs);
1100 					tsleep(curproc, 0, "vmrate", limticks);
1101 					fs->didlimit = 1;
1102 					return (KERN_TRY_AGAIN);
1103 				}
1104 			}
1105 
1106 			/*
1107 			 * Allocate a new page for this object/offset pair.
1108 			 */
1109 			fs->m = NULL;
1110 			if (!vm_page_count_severe()) {
1111 				fs->m = vm_page_alloc(fs->object, pindex,
1112 				    (fs->vp || fs->object->backing_object) ? VM_ALLOC_NORMAL : VM_ALLOC_NORMAL | VM_ALLOC_ZERO);
1113 			}
1114 			if (fs->m == NULL) {
1115 				lwkt_reltoken(&vm_token);
1116 				crit_exit();
1117 				unlock_and_deallocate(fs);
1118 				vm_waitpfault();
1119 				return (KERN_TRY_AGAIN);
1120 			}
1121 		}
1122 		crit_exit();
1123 
1124 readrest:
1125 		/*
1126 		 * We have found an invalid or partially valid page, a
1127 		 * page with a read-ahead mark which might be partially or
1128 		 * fully valid (and maybe dirty too), or we have allocated
1129 		 * a new page.
1130 		 *
1131 		 * Attempt to fault-in the page if there is a chance that the
1132 		 * pager has it, and potentially fault in additional pages
1133 		 * at the same time.
1134 		 *
1135 		 * We are NOT in splvm here and if TRYPAGER is true then
1136 		 * fs.m will be non-NULL and will be PG_BUSY for us.
1137 		 */
1138 		if (TRYPAGER(fs)) {
1139 			int rv;
1140 			int seqaccess;
1141 			u_char behavior = vm_map_entry_behavior(fs->entry);
1142 
1143 			if (behavior == MAP_ENTRY_BEHAV_RANDOM)
1144 				seqaccess = 0;
1145 			else
1146 				seqaccess = -1;
1147 
1148 			/*
1149 			 * If sequential access is detected then attempt
1150 			 * to deactivate/cache pages behind the scan to
1151 			 * prevent resource hogging.
1152 			 *
1153 			 * Use of PG_RAM to detect sequential access
1154 			 * also simulates multi-zone sequential access
1155 			 * detection for free.
1156 			 *
1157 			 * NOTE: Partially valid dirty pages cannot be
1158 			 *	 deactivated without causing NFS picemeal
1159 			 *	 writes to barf.
1160 			 */
1161 			if ((fs->first_object->type != OBJT_DEVICE) &&
1162 			    (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL ||
1163                                 (behavior != MAP_ENTRY_BEHAV_RANDOM &&
1164 				 (fs->m->flags & PG_RAM)))
1165 			) {
1166 				vm_pindex_t scan_pindex;
1167 				int scan_count = 16;
1168 
1169 				if (first_pindex < 16) {
1170 					scan_pindex = 0;
1171 					scan_count = 0;
1172 				} else {
1173 					scan_pindex = first_pindex - 16;
1174 					if (scan_pindex < 16)
1175 						scan_count = scan_pindex;
1176 					else
1177 						scan_count = 16;
1178 				}
1179 
1180 				crit_enter();
1181 				while (scan_count) {
1182 					vm_page_t mt;
1183 
1184 					mt = vm_page_lookup(fs->first_object,
1185 							    scan_pindex);
1186 					if (mt == NULL ||
1187 					    (mt->valid != VM_PAGE_BITS_ALL)) {
1188 						break;
1189 					}
1190 					if (mt->busy ||
1191 					    (mt->flags & (PG_BUSY | PG_FICTITIOUS | PG_UNMANAGED)) ||
1192 					    mt->hold_count ||
1193 					    mt->wire_count)  {
1194 						goto skip;
1195 					}
1196 					if (mt->dirty == 0)
1197 						vm_page_test_dirty(mt);
1198 					if (mt->dirty) {
1199 						vm_page_busy(mt);
1200 						vm_page_protect(mt,
1201 								VM_PROT_NONE);
1202 						vm_page_deactivate(mt);
1203 						vm_page_wakeup(mt);
1204 					} else {
1205 						vm_page_cache(mt);
1206 					}
1207 skip:
1208 					--scan_count;
1209 					--scan_pindex;
1210 				}
1211 				crit_exit();
1212 
1213 				seqaccess = 1;
1214 			}
1215 
1216 			/*
1217 			 * Avoid deadlocking against the map when doing I/O.
1218 			 * fs.object and the page is PG_BUSY'd.
1219 			 */
1220 			unlock_map(fs);
1221 
1222 			/*
1223 			 * Acquire the page data.  We still hold a ref on
1224 			 * fs.object and the page has been PG_BUSY's.
1225 			 *
1226 			 * The pager may replace the page (for example, in
1227 			 * order to enter a fictitious page into the
1228 			 * object).  If it does so it is responsible for
1229 			 * cleaning up the passed page and properly setting
1230 			 * the new page PG_BUSY.
1231 			 *
1232 			 * If we got here through a PG_RAM read-ahead
1233 			 * mark the page may be partially dirty and thus
1234 			 * not freeable.  Don't bother checking to see
1235 			 * if the pager has the page because we can't free
1236 			 * it anyway.  We have to depend on the get_page
1237 			 * operation filling in any gaps whether there is
1238 			 * backing store or not.
1239 			 */
1240 			rv = vm_pager_get_page(fs->object, &fs->m, seqaccess);
1241 
1242 			if (rv == VM_PAGER_OK) {
1243 				/*
1244 				 * Relookup in case pager changed page. Pager
1245 				 * is responsible for disposition of old page
1246 				 * if moved.
1247 				 *
1248 				 * XXX other code segments do relookups too.
1249 				 * It's a bad abstraction that needs to be
1250 				 * fixed/removed.
1251 				 */
1252 				fs->m = vm_page_lookup(fs->object, pindex);
1253 				if (fs->m == NULL) {
1254 					lwkt_reltoken(&vm_token);
1255 					unlock_and_deallocate(fs);
1256 					return (KERN_TRY_AGAIN);
1257 				}
1258 
1259 				++fs->hardfault;
1260 				break; /* break to PAGE HAS BEEN FOUND */
1261 			}
1262 
1263 			/*
1264 			 * Remove the bogus page (which does not exist at this
1265 			 * object/offset); before doing so, we must get back
1266 			 * our object lock to preserve our invariant.
1267 			 *
1268 			 * Also wake up any other process that may want to bring
1269 			 * in this page.
1270 			 *
1271 			 * If this is the top-level object, we must leave the
1272 			 * busy page to prevent another process from rushing
1273 			 * past us, and inserting the page in that object at
1274 			 * the same time that we are.
1275 			 */
1276 			if (rv == VM_PAGER_ERROR) {
1277 				if (curproc)
1278 					kprintf("vm_fault: pager read error, pid %d (%s)\n", curproc->p_pid, curproc->p_comm);
1279 				else
1280 					kprintf("vm_fault: pager read error, thread %p (%s)\n", curthread, curproc->p_comm);
1281 			}
1282 
1283 			/*
1284 			 * Data outside the range of the pager or an I/O error
1285 			 *
1286 			 * The page may have been wired during the pagein,
1287 			 * e.g. by the buffer cache, and cannot simply be
1288 			 * freed.  Call vnode_pager_freepage() to deal with it.
1289 			 */
1290 			/*
1291 			 * XXX - the check for kernel_map is a kludge to work
1292 			 * around having the machine panic on a kernel space
1293 			 * fault w/ I/O error.
1294 			 */
1295 			if (((fs->map != &kernel_map) &&
1296 			    (rv == VM_PAGER_ERROR)) || (rv == VM_PAGER_BAD)) {
1297 				vnode_pager_freepage(fs->m);
1298 				lwkt_reltoken(&vm_token);
1299 				fs->m = NULL;
1300 				unlock_and_deallocate(fs);
1301 				if (rv == VM_PAGER_ERROR)
1302 					return (KERN_FAILURE);
1303 				else
1304 					return (KERN_PROTECTION_FAILURE);
1305 				/* NOT REACHED */
1306 			}
1307 			if (fs->object != fs->first_object) {
1308 				vnode_pager_freepage(fs->m);
1309 				fs->m = NULL;
1310 				/*
1311 				 * XXX - we cannot just fall out at this
1312 				 * point, m has been freed and is invalid!
1313 				 */
1314 			}
1315 		}
1316 
1317 		/*
1318 		 * We get here if the object has a default pager (or unwiring)
1319 		 * or the pager doesn't have the page.
1320 		 */
1321 		if (fs->object == fs->first_object)
1322 			fs->first_m = fs->m;
1323 
1324 		/*
1325 		 * Move on to the next object.  Lock the next object before
1326 		 * unlocking the current one.
1327 		 */
1328 		pindex += OFF_TO_IDX(fs->object->backing_object_offset);
1329 		next_object = fs->object->backing_object;
1330 		if (next_object == NULL) {
1331 			/*
1332 			 * If there's no object left, fill the page in the top
1333 			 * object with zeros.
1334 			 */
1335 			if (fs->object != fs->first_object) {
1336 				vm_object_pip_wakeup(fs->object);
1337 
1338 				fs->object = fs->first_object;
1339 				pindex = first_pindex;
1340 				fs->m = fs->first_m;
1341 			}
1342 			fs->first_m = NULL;
1343 
1344 			/*
1345 			 * Zero the page if necessary and mark it valid.
1346 			 */
1347 			if ((fs->m->flags & PG_ZERO) == 0) {
1348 				vm_page_zero_fill(fs->m);
1349 			} else {
1350 				mycpu->gd_cnt.v_ozfod++;
1351 			}
1352 			mycpu->gd_cnt.v_zfod++;
1353 			fs->m->valid = VM_PAGE_BITS_ALL;
1354 			break;	/* break to PAGE HAS BEEN FOUND */
1355 		}
1356 		if (fs->object != fs->first_object) {
1357 			vm_object_pip_wakeup(fs->object);
1358 		}
1359 		KASSERT(fs->object != next_object,
1360 			("object loop %p", next_object));
1361 		fs->object = next_object;
1362 		vm_object_pip_add(fs->object, 1);
1363 	}
1364 
1365 	/*
1366 	 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
1367 	 * is held.]
1368 	 *
1369 	 * vm_token is still held
1370 	 *
1371 	 * If the page is being written, but isn't already owned by the
1372 	 * top-level object, we have to copy it into a new page owned by the
1373 	 * top-level object.
1374 	 */
1375 	KASSERT((fs->m->flags & PG_BUSY) != 0,
1376 		("vm_fault: not busy after main loop"));
1377 
1378 	if (fs->object != fs->first_object) {
1379 		/*
1380 		 * We only really need to copy if we want to write it.
1381 		 */
1382 		if (fault_type & VM_PROT_WRITE) {
1383 			/*
1384 			 * This allows pages to be virtually copied from a
1385 			 * backing_object into the first_object, where the
1386 			 * backing object has no other refs to it, and cannot
1387 			 * gain any more refs.  Instead of a bcopy, we just
1388 			 * move the page from the backing object to the
1389 			 * first object.  Note that we must mark the page
1390 			 * dirty in the first object so that it will go out
1391 			 * to swap when needed.
1392 			 */
1393 			if (
1394 				/*
1395 				 * Map, if present, has not changed
1396 				 */
1397 				(fs->map == NULL ||
1398 				fs->map_generation == fs->map->timestamp) &&
1399 				/*
1400 				 * Only one shadow object
1401 				 */
1402 				(fs->object->shadow_count == 1) &&
1403 				/*
1404 				 * No COW refs, except us
1405 				 */
1406 				(fs->object->ref_count == 1) &&
1407 				/*
1408 				 * No one else can look this object up
1409 				 */
1410 				(fs->object->handle == NULL) &&
1411 				/*
1412 				 * No other ways to look the object up
1413 				 */
1414 				((fs->object->type == OBJT_DEFAULT) ||
1415 				 (fs->object->type == OBJT_SWAP)) &&
1416 				/*
1417 				 * We don't chase down the shadow chain
1418 				 */
1419 				(fs->object == fs->first_object->backing_object) &&
1420 
1421 				/*
1422 				 * grab the lock if we need to
1423 				 */
1424 				(fs->lookup_still_valid ||
1425 				 fs->map == NULL ||
1426 				 lockmgr(&fs->map->lock, LK_EXCLUSIVE|LK_NOWAIT) == 0)
1427 			    ) {
1428 
1429 				fs->lookup_still_valid = 1;
1430 				/*
1431 				 * get rid of the unnecessary page
1432 				 */
1433 				vm_page_protect(fs->first_m, VM_PROT_NONE);
1434 				vm_page_free(fs->first_m);
1435 				fs->first_m = NULL;
1436 
1437 				/*
1438 				 * grab the page and put it into the
1439 				 * process'es object.  The page is
1440 				 * automatically made dirty.
1441 				 */
1442 				vm_page_rename(fs->m, fs->first_object, first_pindex);
1443 				fs->first_m = fs->m;
1444 				vm_page_busy(fs->first_m);
1445 				fs->m = NULL;
1446 				mycpu->gd_cnt.v_cow_optim++;
1447 			} else {
1448 				/*
1449 				 * Oh, well, lets copy it.
1450 				 */
1451 				vm_page_copy(fs->m, fs->first_m);
1452 				vm_page_event(fs->m, VMEVENT_COW);
1453 			}
1454 
1455 			if (fs->m) {
1456 				/*
1457 				 * We no longer need the old page or object.
1458 				 */
1459 				release_page(fs);
1460 			}
1461 
1462 			/*
1463 			 * fs->object != fs->first_object due to above
1464 			 * conditional
1465 			 */
1466 			vm_object_pip_wakeup(fs->object);
1467 
1468 			/*
1469 			 * Only use the new page below...
1470 			 */
1471 
1472 			mycpu->gd_cnt.v_cow_faults++;
1473 			fs->m = fs->first_m;
1474 			fs->object = fs->first_object;
1475 			pindex = first_pindex;
1476 		} else {
1477 			/*
1478 			 * If it wasn't a write fault avoid having to copy
1479 			 * the page by mapping it read-only.
1480 			 */
1481 			fs->prot &= ~VM_PROT_WRITE;
1482 		}
1483 	}
1484 
1485 	/*
1486 	 * We may have had to unlock a map to do I/O.  If we did then
1487 	 * lookup_still_valid will be FALSE.  If the map generation count
1488 	 * also changed then all sorts of things could have happened while
1489 	 * we were doing the I/O and we need to retry.
1490 	 */
1491 
1492 	if (!fs->lookup_still_valid &&
1493 	    fs->map != NULL &&
1494 	    (fs->map->timestamp != fs->map_generation)) {
1495 		release_page(fs);
1496 		lwkt_reltoken(&vm_token);
1497 		unlock_and_deallocate(fs);
1498 		return (KERN_TRY_AGAIN);
1499 	}
1500 
1501 	/*
1502 	 * If the fault is a write, we know that this page is being
1503 	 * written NOW so dirty it explicitly to save on pmap_is_modified()
1504 	 * calls later.
1505 	 *
1506 	 * If this is a NOSYNC mmap we do not want to set PG_NOSYNC
1507 	 * if the page is already dirty to prevent data written with
1508 	 * the expectation of being synced from not being synced.
1509 	 * Likewise if this entry does not request NOSYNC then make
1510 	 * sure the page isn't marked NOSYNC.  Applications sharing
1511 	 * data should use the same flags to avoid ping ponging.
1512 	 *
1513 	 * Also tell the backing pager, if any, that it should remove
1514 	 * any swap backing since the page is now dirty.
1515 	 */
1516 	if (fs->prot & VM_PROT_WRITE) {
1517 		vm_object_set_writeable_dirty(fs->m->object);
1518 		if (fs->entry->eflags & MAP_ENTRY_NOSYNC) {
1519 			if (fs->m->dirty == 0)
1520 				vm_page_flag_set(fs->m, PG_NOSYNC);
1521 		} else {
1522 			vm_page_flag_clear(fs->m, PG_NOSYNC);
1523 		}
1524 		if (fs->fault_flags & VM_FAULT_DIRTY) {
1525 			crit_enter();
1526 			vm_page_dirty(fs->m);
1527 			swap_pager_unswapped(fs->m);
1528 			crit_exit();
1529 		}
1530 	}
1531 
1532 	lwkt_reltoken(&vm_token);
1533 
1534 	/*
1535 	 * Page had better still be busy.  We are still locked up and
1536 	 * fs->object will have another PIP reference if it is not equal
1537 	 * to fs->first_object.
1538 	 */
1539 	KASSERT(fs->m->flags & PG_BUSY,
1540 		("vm_fault: page %p not busy!", fs->m));
1541 
1542 	/*
1543 	 * Sanity check: page must be completely valid or it is not fit to
1544 	 * map into user space.  vm_pager_get_pages() ensures this.
1545 	 */
1546 	if (fs->m->valid != VM_PAGE_BITS_ALL) {
1547 		vm_page_zero_invalid(fs->m, TRUE);
1548 		kprintf("Warning: page %p partially invalid on fault\n", fs->m);
1549 	}
1550 
1551 	return (KERN_SUCCESS);
1552 }
1553 
1554 /*
1555  * Wire down a range of virtual addresses in a map.  The entry in question
1556  * should be marked in-transition and the map must be locked.  We must
1557  * release the map temporarily while faulting-in the page to avoid a
1558  * deadlock.  Note that the entry may be clipped while we are blocked but
1559  * will never be freed.
1560  *
1561  * No requirements.
1562  */
1563 int
1564 vm_fault_wire(vm_map_t map, vm_map_entry_t entry, boolean_t user_wire)
1565 {
1566 	boolean_t fictitious;
1567 	vm_offset_t start;
1568 	vm_offset_t end;
1569 	vm_offset_t va;
1570 	vm_paddr_t pa;
1571 	pmap_t pmap;
1572 	int rv;
1573 
1574 	pmap = vm_map_pmap(map);
1575 	start = entry->start;
1576 	end = entry->end;
1577 	fictitious = entry->object.vm_object &&
1578 			(entry->object.vm_object->type == OBJT_DEVICE);
1579 
1580 	lwkt_gettoken(&vm_token);
1581 	vm_map_unlock(map);
1582 	map->timestamp++;
1583 
1584 	/*
1585 	 * We simulate a fault to get the page and enter it in the physical
1586 	 * map.
1587 	 */
1588 	for (va = start; va < end; va += PAGE_SIZE) {
1589 		if (user_wire) {
1590 			rv = vm_fault(map, va, VM_PROT_READ,
1591 					VM_FAULT_USER_WIRE);
1592 		} else {
1593 			rv = vm_fault(map, va, VM_PROT_READ|VM_PROT_WRITE,
1594 					VM_FAULT_CHANGE_WIRING);
1595 		}
1596 		if (rv) {
1597 			while (va > start) {
1598 				va -= PAGE_SIZE;
1599 				if ((pa = pmap_extract(pmap, va)) == 0)
1600 					continue;
1601 				pmap_change_wiring(pmap, va, FALSE);
1602 				if (!fictitious)
1603 					vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1);
1604 			}
1605 			vm_map_lock(map);
1606 			lwkt_reltoken(&vm_token);
1607 			return (rv);
1608 		}
1609 	}
1610 	vm_map_lock(map);
1611 	lwkt_reltoken(&vm_token);
1612 	return (KERN_SUCCESS);
1613 }
1614 
1615 /*
1616  * Unwire a range of virtual addresses in a map.  The map should be
1617  * locked.
1618  */
1619 void
1620 vm_fault_unwire(vm_map_t map, vm_map_entry_t entry)
1621 {
1622 	boolean_t fictitious;
1623 	vm_offset_t start;
1624 	vm_offset_t end;
1625 	vm_offset_t va;
1626 	vm_paddr_t pa;
1627 	pmap_t pmap;
1628 
1629 	pmap = vm_map_pmap(map);
1630 	start = entry->start;
1631 	end = entry->end;
1632 	fictitious = entry->object.vm_object &&
1633 			(entry->object.vm_object->type == OBJT_DEVICE);
1634 
1635 	/*
1636 	 * Since the pages are wired down, we must be able to get their
1637 	 * mappings from the physical map system.
1638 	 */
1639 	lwkt_gettoken(&vm_token);
1640 	for (va = start; va < end; va += PAGE_SIZE) {
1641 		pa = pmap_extract(pmap, va);
1642 		if (pa != 0) {
1643 			pmap_change_wiring(pmap, va, FALSE);
1644 			if (!fictitious)
1645 				vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1);
1646 		}
1647 	}
1648 	lwkt_reltoken(&vm_token);
1649 }
1650 
1651 /*
1652  * Reduce the rate at which memory is allocated to a process based
1653  * on the perceived load on the VM system. As the load increases
1654  * the allocation burst rate goes down and the delay increases.
1655  *
1656  * Rate limiting does not apply when faulting active or inactive
1657  * pages.  When faulting 'cache' pages, rate limiting only applies
1658  * if the system currently has a severe page deficit.
1659  *
1660  * XXX vm_pagesupply should be increased when a page is freed.
1661  *
1662  * We sleep up to 1/10 of a second.
1663  */
1664 static int
1665 vm_fault_ratelimit(struct vmspace *vmspace)
1666 {
1667 	if (vm_load_enable == 0)
1668 		return(0);
1669 	if (vmspace->vm_pagesupply > 0) {
1670 		--vmspace->vm_pagesupply;	/* SMP race ok */
1671 		return(0);
1672 	}
1673 #ifdef INVARIANTS
1674 	if (vm_load_debug) {
1675 		kprintf("load %-4d give %d pgs, wait %d, pid %-5d (%s)\n",
1676 			vm_load,
1677 			(1000 - vm_load ) / 10, vm_load * hz / 10000,
1678 			curproc->p_pid, curproc->p_comm);
1679 	}
1680 #endif
1681 	vmspace->vm_pagesupply = (1000 - vm_load) / 10;
1682 	return(vm_load * hz / 10000);
1683 }
1684 
1685 /*
1686  * Copy all of the pages from a wired-down map entry to another.
1687  *
1688  * The source and destination maps must be locked for write.
1689  * The source map entry must be wired down (or be a sharing map
1690  * entry corresponding to a main map entry that is wired down).
1691  *
1692  * No other requirements.
1693  */
1694 void
1695 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1696 		    vm_map_entry_t dst_entry, vm_map_entry_t src_entry)
1697 {
1698 	vm_object_t dst_object;
1699 	vm_object_t src_object;
1700 	vm_ooffset_t dst_offset;
1701 	vm_ooffset_t src_offset;
1702 	vm_prot_t prot;
1703 	vm_offset_t vaddr;
1704 	vm_page_t dst_m;
1705 	vm_page_t src_m;
1706 
1707 #ifdef	lint
1708 	src_map++;
1709 #endif	/* lint */
1710 
1711 	src_object = src_entry->object.vm_object;
1712 	src_offset = src_entry->offset;
1713 
1714 	/*
1715 	 * Create the top-level object for the destination entry. (Doesn't
1716 	 * actually shadow anything - we copy the pages directly.)
1717 	 */
1718 	vm_map_entry_allocate_object(dst_entry);
1719 	dst_object = dst_entry->object.vm_object;
1720 
1721 	prot = dst_entry->max_protection;
1722 
1723 	/*
1724 	 * Loop through all of the pages in the entry's range, copying each
1725 	 * one from the source object (it should be there) to the destination
1726 	 * object.
1727 	 */
1728 	for (vaddr = dst_entry->start, dst_offset = 0;
1729 	    vaddr < dst_entry->end;
1730 	    vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) {
1731 
1732 		/*
1733 		 * Allocate a page in the destination object
1734 		 */
1735 		do {
1736 			dst_m = vm_page_alloc(dst_object,
1737 				OFF_TO_IDX(dst_offset), VM_ALLOC_NORMAL);
1738 			if (dst_m == NULL) {
1739 				vm_wait(0);
1740 			}
1741 		} while (dst_m == NULL);
1742 
1743 		/*
1744 		 * Find the page in the source object, and copy it in.
1745 		 * (Because the source is wired down, the page will be in
1746 		 * memory.)
1747 		 */
1748 		src_m = vm_page_lookup(src_object,
1749 			OFF_TO_IDX(dst_offset + src_offset));
1750 		if (src_m == NULL)
1751 			panic("vm_fault_copy_wired: page missing");
1752 
1753 		vm_page_copy(src_m, dst_m);
1754 		vm_page_event(src_m, VMEVENT_COW);
1755 
1756 		/*
1757 		 * Enter it in the pmap...
1758 		 */
1759 
1760 		vm_page_flag_clear(dst_m, PG_ZERO);
1761 		pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE);
1762 
1763 		/*
1764 		 * Mark it no longer busy, and put it on the active list.
1765 		 */
1766 		vm_page_activate(dst_m);
1767 		vm_page_wakeup(dst_m);
1768 	}
1769 }
1770 
1771 #if 0
1772 
1773 /*
1774  * This routine checks around the requested page for other pages that
1775  * might be able to be faulted in.  This routine brackets the viable
1776  * pages for the pages to be paged in.
1777  *
1778  * Inputs:
1779  *	m, rbehind, rahead
1780  *
1781  * Outputs:
1782  *  marray (array of vm_page_t), reqpage (index of requested page)
1783  *
1784  * Return value:
1785  *  number of pages in marray
1786  */
1787 static int
1788 vm_fault_additional_pages(vm_page_t m, int rbehind, int rahead,
1789 			  vm_page_t *marray, int *reqpage)
1790 {
1791 	int i,j;
1792 	vm_object_t object;
1793 	vm_pindex_t pindex, startpindex, endpindex, tpindex;
1794 	vm_page_t rtm;
1795 	int cbehind, cahead;
1796 
1797 	object = m->object;
1798 	pindex = m->pindex;
1799 
1800 	/*
1801 	 * we don't fault-ahead for device pager
1802 	 */
1803 	if (object->type == OBJT_DEVICE) {
1804 		*reqpage = 0;
1805 		marray[0] = m;
1806 		return 1;
1807 	}
1808 
1809 	/*
1810 	 * if the requested page is not available, then give up now
1811 	 */
1812 	if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1813 		*reqpage = 0;	/* not used by caller, fix compiler warn */
1814 		return 0;
1815 	}
1816 
1817 	if ((cbehind == 0) && (cahead == 0)) {
1818 		*reqpage = 0;
1819 		marray[0] = m;
1820 		return 1;
1821 	}
1822 
1823 	if (rahead > cahead) {
1824 		rahead = cahead;
1825 	}
1826 
1827 	if (rbehind > cbehind) {
1828 		rbehind = cbehind;
1829 	}
1830 
1831 	/*
1832 	 * Do not do any readahead if we have insufficient free memory.
1833 	 *
1834 	 * XXX code was broken disabled before and has instability
1835 	 * with this conditonal fixed, so shortcut for now.
1836 	 */
1837 	if (burst_fault == 0 || vm_page_count_severe()) {
1838 		marray[0] = m;
1839 		*reqpage = 0;
1840 		return 1;
1841 	}
1842 
1843 	/*
1844 	 * scan backward for the read behind pages -- in memory
1845 	 *
1846 	 * Assume that if the page is not found an interrupt will not
1847 	 * create it.  Theoretically interrupts can only remove (busy)
1848 	 * pages, not create new associations.
1849 	 */
1850 	if (pindex > 0) {
1851 		if (rbehind > pindex) {
1852 			rbehind = pindex;
1853 			startpindex = 0;
1854 		} else {
1855 			startpindex = pindex - rbehind;
1856 		}
1857 
1858 		crit_enter();
1859 		lwkt_gettoken(&vm_token);
1860 		for (tpindex = pindex; tpindex > startpindex; --tpindex) {
1861 			if (vm_page_lookup(object, tpindex - 1))
1862 				break;
1863 		}
1864 
1865 		i = 0;
1866 		while (tpindex < pindex) {
1867 			rtm = vm_page_alloc(object, tpindex, VM_ALLOC_SYSTEM);
1868 			if (rtm == NULL) {
1869 				lwkt_reltoken(&vm_token);
1870 				crit_exit();
1871 				for (j = 0; j < i; j++) {
1872 					vm_page_free(marray[j]);
1873 				}
1874 				marray[0] = m;
1875 				*reqpage = 0;
1876 				return 1;
1877 			}
1878 			marray[i] = rtm;
1879 			++i;
1880 			++tpindex;
1881 		}
1882 		lwkt_reltoken(&vm_token);
1883 		crit_exit();
1884 	} else {
1885 		i = 0;
1886 	}
1887 
1888 	/*
1889 	 * Assign requested page
1890 	 */
1891 	marray[i] = m;
1892 	*reqpage = i;
1893 	++i;
1894 
1895 	/*
1896 	 * Scan forwards for read-ahead pages
1897 	 */
1898 	tpindex = pindex + 1;
1899 	endpindex = tpindex + rahead;
1900 	if (endpindex > object->size)
1901 		endpindex = object->size;
1902 
1903 	crit_enter();
1904 	lwkt_gettoken(&vm_token);
1905 	while (tpindex < endpindex) {
1906 		if (vm_page_lookup(object, tpindex))
1907 			break;
1908 		rtm = vm_page_alloc(object, tpindex, VM_ALLOC_SYSTEM);
1909 		if (rtm == NULL)
1910 			break;
1911 		marray[i] = rtm;
1912 		++i;
1913 		++tpindex;
1914 	}
1915 	lwkt_reltoken(&vm_token);
1916 	crit_exit();
1917 
1918 	return (i);
1919 }
1920 
1921 #endif
1922 
1923 /*
1924  * vm_prefault() provides a quick way of clustering pagefaults into a
1925  * processes address space.  It is a "cousin" of pmap_object_init_pt,
1926  * except it runs at page fault time instead of mmap time.
1927  *
1928  * This code used to be per-platform pmap_prefault().  It is now
1929  * machine-independent and enhanced to also pre-fault zero-fill pages
1930  * (see vm.fast_fault) as well as make them writable, which greatly
1931  * reduces the number of page faults programs incur.
1932  *
1933  * Application performance when pre-faulting zero-fill pages is heavily
1934  * dependent on the application.  Very tiny applications like /bin/echo
1935  * lose a little performance while applications of any appreciable size
1936  * gain performance.  Prefaulting multiple pages also reduces SMP
1937  * congestion and can improve SMP performance significantly.
1938  *
1939  * NOTE!  prot may allow writing but this only applies to the top level
1940  *	  object.  If we wind up mapping a page extracted from a backing
1941  *	  object we have to make sure it is read-only.
1942  *
1943  * NOTE!  The caller has already handled any COW operations on the
1944  *	  vm_map_entry via the normal fault code.  Do NOT call this
1945  *	  shortcut unless the normal fault code has run on this entry.
1946  *
1947  * No other requirements.
1948  */
1949 #define PFBAK 4
1950 #define PFFOR 4
1951 #define PAGEORDER_SIZE (PFBAK+PFFOR)
1952 
1953 static int vm_prefault_pageorder[] = {
1954 	-PAGE_SIZE, PAGE_SIZE,
1955 	-2 * PAGE_SIZE, 2 * PAGE_SIZE,
1956 	-3 * PAGE_SIZE, 3 * PAGE_SIZE,
1957 	-4 * PAGE_SIZE, 4 * PAGE_SIZE
1958 };
1959 
1960 static void
1961 vm_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry, int prot)
1962 {
1963 	struct lwp *lp;
1964 	vm_page_t m;
1965 	vm_offset_t starta;
1966 	vm_offset_t addr;
1967 	vm_pindex_t index;
1968 	vm_pindex_t pindex;
1969 	vm_object_t object;
1970 	int pprot;
1971 	int i;
1972 
1973 	/*
1974 	 * We do not currently prefault mappings that use virtual page
1975 	 * tables.  We do not prefault foreign pmaps.
1976 	 */
1977 	if (entry->maptype == VM_MAPTYPE_VPAGETABLE)
1978 		return;
1979 	lp = curthread->td_lwp;
1980 	if (lp == NULL || (pmap != vmspace_pmap(lp->lwp_vmspace)))
1981 		return;
1982 
1983 	object = entry->object.vm_object;
1984 
1985 	starta = addra - PFBAK * PAGE_SIZE;
1986 	if (starta < entry->start)
1987 		starta = entry->start;
1988 	else if (starta > addra)
1989 		starta = 0;
1990 
1991 	/*
1992 	 * critical section protection is required to maintain the
1993 	 * page/object association, interrupts can free pages and remove
1994 	 * them from their objects.
1995 	 */
1996 	crit_enter();
1997 	lwkt_gettoken(&vm_token);
1998 	for (i = 0; i < PAGEORDER_SIZE; i++) {
1999 		vm_object_t lobject;
2000 		int allocated = 0;
2001 
2002 		addr = addra + vm_prefault_pageorder[i];
2003 		if (addr > addra + (PFFOR * PAGE_SIZE))
2004 			addr = 0;
2005 
2006 		if (addr < starta || addr >= entry->end)
2007 			continue;
2008 
2009 		if (pmap_prefault_ok(pmap, addr) == 0)
2010 			continue;
2011 
2012 		/*
2013 		 * Follow the VM object chain to obtain the page to be mapped
2014 		 * into the pmap.
2015 		 *
2016 		 * If we reach the terminal object without finding a page
2017 		 * and we determine it would be advantageous, then allocate
2018 		 * a zero-fill page for the base object.  The base object
2019 		 * is guaranteed to be OBJT_DEFAULT for this case.
2020 		 *
2021 		 * In order to not have to check the pager via *haspage*()
2022 		 * we stop if any non-default object is encountered.  e.g.
2023 		 * a vnode or swap object would stop the loop.
2024 		 */
2025 		index = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
2026 		lobject = object;
2027 		pindex = index;
2028 		pprot = prot;
2029 
2030 		while ((m = vm_page_lookup(lobject, pindex)) == NULL) {
2031 			if (lobject->type != OBJT_DEFAULT)
2032 				break;
2033 			if (lobject->backing_object == NULL) {
2034 				if (vm_fast_fault == 0)
2035 					break;
2036 				if (vm_prefault_pageorder[i] < 0 ||
2037 				    (prot & VM_PROT_WRITE) == 0 ||
2038 				    vm_page_count_min(0)) {
2039 					break;
2040 				}
2041 				/* note: allocate from base object */
2042 				m = vm_page_alloc(object, index,
2043 					      VM_ALLOC_NORMAL | VM_ALLOC_ZERO);
2044 
2045 				if ((m->flags & PG_ZERO) == 0) {
2046 					vm_page_zero_fill(m);
2047 				} else {
2048 					vm_page_flag_clear(m, PG_ZERO);
2049 					mycpu->gd_cnt.v_ozfod++;
2050 				}
2051 				mycpu->gd_cnt.v_zfod++;
2052 				m->valid = VM_PAGE_BITS_ALL;
2053 				allocated = 1;
2054 				pprot = prot;
2055 				/* lobject = object .. not needed */
2056 				break;
2057 			}
2058 			if (lobject->backing_object_offset & PAGE_MASK)
2059 				break;
2060 			pindex += lobject->backing_object_offset >> PAGE_SHIFT;
2061 			lobject = lobject->backing_object;
2062 			pprot &= ~VM_PROT_WRITE;
2063 		}
2064 		/*
2065 		 * NOTE: lobject now invalid (if we did a zero-fill we didn't
2066 		 *	 bother assigning lobject = object).
2067 		 *
2068 		 * Give-up if the page is not available.
2069 		 */
2070 		if (m == NULL)
2071 			break;
2072 
2073 		/*
2074 		 * Do not conditionalize on PG_RAM.  If pages are present in
2075 		 * the VM system we assume optimal caching.  If caching is
2076 		 * not optimal the I/O gravy train will be restarted when we
2077 		 * hit an unavailable page.  We do not want to try to restart
2078 		 * the gravy train now because we really don't know how much
2079 		 * of the object has been cached.  The cost for restarting
2080 		 * the gravy train should be low (since accesses will likely
2081 		 * be I/O bound anyway).
2082 		 *
2083 		 * The object must be marked dirty if we are mapping a
2084 		 * writable page.
2085 		 */
2086 		if (pprot & VM_PROT_WRITE)
2087 			vm_object_set_writeable_dirty(m->object);
2088 
2089 		/*
2090 		 * Enter the page into the pmap if appropriate.  If we had
2091 		 * allocated the page we have to place it on a queue.  If not
2092 		 * we just have to make sure it isn't on the cache queue
2093 		 * (pages on the cache queue are not allowed to be mapped).
2094 		 */
2095 		if (allocated) {
2096 			pmap_enter(pmap, addr, m, pprot, 0);
2097 			vm_page_deactivate(m);
2098 			vm_page_wakeup(m);
2099 		} else if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2100 		    (m->busy == 0) &&
2101 		    (m->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2102 
2103 			if ((m->queue - m->pc) == PQ_CACHE) {
2104 				vm_page_deactivate(m);
2105 			}
2106 			vm_page_busy(m);
2107 			pmap_enter(pmap, addr, m, pprot, 0);
2108 			vm_page_wakeup(m);
2109 		}
2110 	}
2111 	lwkt_reltoken(&vm_token);
2112 	crit_exit();
2113 }
2114