xref: /dflybsd-src/sys/vm/vm_fault.c (revision 747e2961e929a839fd1f570e8b2c2fce73f0fe5c)
1 /*
2  * Copyright (c) 2003-2019 The DragonFly Project.  All rights reserved.
3  *
4  * This code is derived from software contributed to The DragonFly Project
5  * by Matthew Dillon <dillon@backplane.com>
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  *
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in
15  *    the documentation and/or other materials provided with the
16  *    distribution.
17  * 3. Neither the name of The DragonFly Project nor the names of its
18  *    contributors may be used to endorse or promote products derived
19  *    from this software without specific, prior written permission.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
25  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  *
34  * ---
35  *
36  * Copyright (c) 1991, 1993
37  *	The Regents of the University of California.  All rights reserved.
38  * Copyright (c) 1994 John S. Dyson
39  * All rights reserved.
40  * Copyright (c) 1994 David Greenman
41  * All rights reserved.
42  *
43  *
44  * This code is derived from software contributed to Berkeley by
45  * The Mach Operating System project at Carnegie-Mellon University.
46  *
47  * Redistribution and use in source and binary forms, with or without
48  * modification, are permitted provided that the following conditions
49  * are met:
50  * 1. Redistributions of source code must retain the above copyright
51  *    notice, this list of conditions and the following disclaimer.
52  * 2. Redistributions in binary form must reproduce the above copyright
53  *    notice, this list of conditions and the following disclaimer in the
54  *    documentation and/or other materials provided with the distribution.
55  * 3. Neither the name of the University nor the names of its contributors
56  *    may be used to endorse or promote products derived from this software
57  *    without specific prior written permission.
58  *
59  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
60  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
61  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
62  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
63  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
64  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
65  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
66  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
67  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
68  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
69  * SUCH DAMAGE.
70  *
71  * ---
72  *
73  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
74  * All rights reserved.
75  *
76  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
77  *
78  * Permission to use, copy, modify and distribute this software and
79  * its documentation is hereby granted, provided that both the copyright
80  * notice and this permission notice appear in all copies of the
81  * software, derivative works or modified versions, and any portions
82  * thereof, and that both notices appear in supporting documentation.
83  *
84  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
85  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
86  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
87  *
88  * Carnegie Mellon requests users of this software to return to
89  *
90  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
91  *  School of Computer Science
92  *  Carnegie Mellon University
93  *  Pittsburgh PA 15213-3890
94  *
95  * any improvements or extensions that they make and grant Carnegie the
96  * rights to redistribute these changes.
97  */
98 
99 /*
100  *	Page fault handling module.
101  */
102 
103 #include <sys/param.h>
104 #include <sys/systm.h>
105 #include <sys/kernel.h>
106 #include <sys/proc.h>
107 #include <sys/vnode.h>
108 #include <sys/resourcevar.h>
109 #include <sys/vmmeter.h>
110 #include <sys/vkernel.h>
111 #include <sys/lock.h>
112 #include <sys/sysctl.h>
113 
114 #include <cpu/lwbuf.h>
115 
116 #include <vm/vm.h>
117 #include <vm/vm_param.h>
118 #include <vm/pmap.h>
119 #include <vm/vm_map.h>
120 #include <vm/vm_object.h>
121 #include <vm/vm_page.h>
122 #include <vm/vm_pageout.h>
123 #include <vm/vm_kern.h>
124 #include <vm/vm_pager.h>
125 #include <vm/vnode_pager.h>
126 #include <vm/swap_pager.h>
127 #include <vm/vm_extern.h>
128 
129 #include <vm/vm_page2.h>
130 
131 struct faultstate {
132 	vm_page_t m;
133 	vm_map_backing_t ba;
134 	vm_prot_t prot;
135 	vm_page_t first_m;
136 	vm_map_backing_t first_ba;
137 	vm_prot_t first_prot;
138 	vm_map_t map;
139 	vm_map_entry_t entry;
140 	int lookup_still_valid;	/* 0=inv 1=valid/rel -1=valid/atomic */
141 	int hardfault;
142 	int fault_flags;
143 	int shared;
144 	int msoftonly;
145 	int first_shared;
146 	int wflags;
147 	int first_ba_held;	/* 0=unlocked 1=locked/rel -1=lock/atomic */
148 	struct vnode *vp;
149 };
150 
151 __read_mostly static int debug_fault = 0;
152 SYSCTL_INT(_vm, OID_AUTO, debug_fault, CTLFLAG_RW, &debug_fault, 0, "");
153 __read_mostly static int debug_cluster = 0;
154 SYSCTL_INT(_vm, OID_AUTO, debug_cluster, CTLFLAG_RW, &debug_cluster, 0, "");
155 #if 0
156 static int virtual_copy_enable = 1;
157 SYSCTL_INT(_vm, OID_AUTO, virtual_copy_enable, CTLFLAG_RW,
158 		&virtual_copy_enable, 0, "");
159 #endif
160 __read_mostly int vm_shared_fault = 1;
161 TUNABLE_INT("vm.shared_fault", &vm_shared_fault);
162 SYSCTL_INT(_vm, OID_AUTO, shared_fault, CTLFLAG_RW,
163 		&vm_shared_fault, 0, "Allow shared token on vm_object");
164 __read_mostly static int vm_fault_quick_enable = 0;
165 TUNABLE_INT("vm.fault_quick", &vm_fault_quick_enable);
166 SYSCTL_INT(_vm, OID_AUTO, fault_quick, CTLFLAG_RW,
167 		&vm_fault_quick_enable, 0, "Allow fast vm_fault shortcut");
168 #ifdef VM_FAULT_QUICK_DEBUG
169 static long vm_fault_quick_success_count = 0;
170 SYSCTL_LONG(_vm, OID_AUTO, fault_quick_success_count, CTLFLAG_RW,
171 		&vm_fault_quick_success_count, 0, "");
172 static long vm_fault_quick_failure_count1 = 0;
173 SYSCTL_LONG(_vm, OID_AUTO, fault_quick_failure_count1, CTLFLAG_RW,
174 		&vm_fault_quick_failure_count1, 0, "");
175 static long vm_fault_quick_failure_count2 = 0;
176 SYSCTL_LONG(_vm, OID_AUTO, fault_quick_failure_count2, CTLFLAG_RW,
177 		&vm_fault_quick_failure_count2, 0, "");
178 static long vm_fault_quick_failure_count3 = 0;
179 SYSCTL_LONG(_vm, OID_AUTO, fault_quick_failure_count3, CTLFLAG_RW,
180 		&vm_fault_quick_failure_count3, 0, "");
181 static long vm_fault_quick_failure_count4 = 0;
182 SYSCTL_LONG(_vm, OID_AUTO, fault_quick_failure_count4, CTLFLAG_RW,
183 		&vm_fault_quick_failure_count4, 0, "");
184 #endif
185 
186 static int vm_fault_quick(struct faultstate *fs, vm_pindex_t first_pindex,
187 			vm_prot_t fault_type);
188 static int vm_fault_object(struct faultstate *, vm_pindex_t, vm_prot_t, int);
189 static int vm_fault_vpagetable(struct faultstate *, vm_pindex_t *,
190 			vpte_t, int, int);
191 #if 0
192 static int vm_fault_additional_pages (vm_page_t, int, int, vm_page_t *, int *);
193 #endif
194 static void vm_set_nosync(vm_page_t m, vm_map_entry_t entry);
195 static void vm_prefault(pmap_t pmap, vm_offset_t addra,
196 			vm_map_entry_t entry, int prot, int fault_flags);
197 static void vm_prefault_quick(pmap_t pmap, vm_offset_t addra,
198 			vm_map_entry_t entry, int prot, int fault_flags);
199 
200 static __inline void
201 release_page(struct faultstate *fs)
202 {
203 	vm_page_deactivate(fs->m);
204 	vm_page_wakeup(fs->m);
205 	fs->m = NULL;
206 }
207 
208 static __inline void
209 unlock_map(struct faultstate *fs)
210 {
211 	if (fs->ba != fs->first_ba)
212 		vm_object_drop(fs->ba->object);
213 	if (fs->first_ba && fs->first_ba_held == 1) {
214 		vm_object_drop(fs->first_ba->object);
215 		fs->first_ba_held = 0;
216 		fs->first_ba = NULL;
217 	}
218 	fs->ba = NULL;
219 
220 	/*
221 	 * NOTE: If lookup_still_valid == -1 the map is assumed to be locked
222 	 *	 and caller expects it to remain locked atomically.
223 	 */
224 	if (fs->lookup_still_valid == 1 && fs->map) {
225 		vm_map_lookup_done(fs->map, fs->entry, 0);
226 		fs->lookup_still_valid = 0;
227 		fs->entry = NULL;
228 	}
229 }
230 
231 /*
232  * Clean up after a successful call to vm_fault_object() so another call
233  * to vm_fault_object() can be made.
234  */
235 static void
236 cleanup_fault(struct faultstate *fs)
237 {
238 	/*
239 	 * We allocated a junk page for a COW operation that did
240 	 * not occur, the page must be freed.
241 	 */
242 	if (fs->ba != fs->first_ba) {
243 		KKASSERT(fs->first_shared == 0);
244 
245 		/*
246 		 * first_m could be completely valid and we got here
247 		 * because of a PG_RAM, don't mistakenly free it!
248 		 */
249 		if ((fs->first_m->valid & VM_PAGE_BITS_ALL) ==
250 		    VM_PAGE_BITS_ALL) {
251 			vm_page_wakeup(fs->first_m);
252 		} else {
253 			vm_page_free(fs->first_m);
254 		}
255 		vm_object_pip_wakeup(fs->ba->object);
256 		fs->first_m = NULL;
257 
258 		/*
259 		 * Reset fs->ba (used by vm_fault_vpagetahble() without
260 		 * calling unlock_map(), so we need a little duplication.
261 		 */
262 		vm_object_drop(fs->ba->object);
263 		fs->ba = fs->first_ba;
264 	}
265 }
266 
267 static void
268 unlock_things(struct faultstate *fs)
269 {
270 	cleanup_fault(fs);
271 	unlock_map(fs);
272 	if (fs->vp != NULL) {
273 		vput(fs->vp);
274 		fs->vp = NULL;
275 	}
276 }
277 
278 #if 0
279 /*
280  * Virtual copy tests.   Used by the fault code to determine if a
281  * page can be moved from an orphan vm_object into its shadow
282  * instead of copying its contents.
283  */
284 static __inline int
285 virtual_copy_test(struct faultstate *fs)
286 {
287 	/*
288 	 * Must be holding exclusive locks
289 	 */
290 	if (fs->first_shared || fs->shared || virtual_copy_enable == 0)
291 		return 0;
292 
293 	/*
294 	 * Map, if present, has not changed
295 	 */
296 	if (fs->map && fs->map_generation != fs->map->timestamp)
297 		return 0;
298 
299 	/*
300 	 * No refs, except us
301 	 */
302 	if (fs->ba->object->ref_count != 1)
303 		return 0;
304 
305 	/*
306 	 * No one else can look this object up
307 	 */
308 	if (fs->ba->object->handle != NULL)
309 		return 0;
310 
311 	/*
312 	 * No other ways to look the object up
313 	 */
314 	if (fs->ba->object->type != OBJT_DEFAULT &&
315 	    fs->ba->object->type != OBJT_SWAP)
316 		return 0;
317 
318 	/*
319 	 * We don't chase down the shadow chain
320 	 */
321 	if (fs->ba != fs->first_ba->backing_ba)
322 		return 0;
323 
324 	return 1;
325 }
326 
327 static __inline int
328 virtual_copy_ok(struct faultstate *fs)
329 {
330 	if (virtual_copy_test(fs)) {
331 		/*
332 		 * Grab the lock and re-test changeable items.
333 		 */
334 		if (fs->lookup_still_valid == 0 && fs->map) {
335 			if (lockmgr(&fs->map->lock, LK_EXCLUSIVE|LK_NOWAIT))
336 				return 0;
337 			fs->lookup_still_valid = 1;
338 			if (virtual_copy_test(fs)) {
339 				fs->map_generation = ++fs->map->timestamp;
340 				return 1;
341 			}
342 			fs->lookup_still_valid = 0;
343 			lockmgr(&fs->map->lock, LK_RELEASE);
344 		}
345 	}
346 	return 0;
347 }
348 #endif
349 
350 /*
351  * TRYPAGER
352  *
353  * Determine if the pager for the current object *might* contain the page.
354  *
355  * We only need to try the pager if this is not a default object (default
356  * objects are zero-fill and have no real pager), and if we are not taking
357  * a wiring fault or if the FS entry is wired.
358  */
359 #define TRYPAGER(fs)	\
360 		(fs->ba->object->type != OBJT_DEFAULT &&		\
361 		(((fs->fault_flags & VM_FAULT_WIRE_MASK) == 0) ||	\
362 		 (fs->wflags & FW_WIRED)))
363 
364 /*
365  * vm_fault:
366  *
367  * Handle a page fault occuring at the given address, requiring the given
368  * permissions, in the map specified.  If successful, the page is inserted
369  * into the associated physical map.
370  *
371  * NOTE: The given address should be truncated to the proper page address.
372  *
373  * KERN_SUCCESS is returned if the page fault is handled; otherwise,
374  * a standard error specifying why the fault is fatal is returned.
375  *
376  * The map in question must be referenced, and remains so.
377  * The caller may hold no locks.
378  * No other requirements.
379  */
380 int
381 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, int fault_flags)
382 {
383 	int result;
384 	vm_pindex_t first_pindex;
385 	struct faultstate fs;
386 	struct lwp *lp;
387 	struct proc *p;
388 	thread_t td;
389 	struct vm_map_ilock ilock;
390 	int didilock;
391 	int growstack;
392 	int retry = 0;
393 	int inherit_prot;
394 
395 	inherit_prot = fault_type & VM_PROT_NOSYNC;
396 	fs.hardfault = 0;
397 	fs.fault_flags = fault_flags;
398 	fs.vp = NULL;
399 	fs.shared = vm_shared_fault;
400 	fs.first_shared = vm_shared_fault;
401 	growstack = 1;
402 
403 	/*
404 	 * vm_map interactions
405 	 */
406 	td = curthread;
407 	if ((lp = td->td_lwp) != NULL)
408 		lp->lwp_flags |= LWP_PAGING;
409 
410 RetryFault:
411 	/*
412 	 * vm_fault_quick() can shortcut us.
413 	 */
414 	fs.msoftonly = 0;
415 	fs.first_ba_held = 0;
416 
417 	/*
418 	 * Find the vm_map_entry representing the backing store and resolve
419 	 * the top level object and page index.  This may have the side
420 	 * effect of executing a copy-on-write on the map entry,
421 	 * creating a shadow object, or splitting an anonymous entry for
422 	 * performance, but will not COW any actual VM pages.
423 	 *
424 	 * On success fs.map is left read-locked and various other fields
425 	 * are initialized but not otherwise referenced or locked.
426 	 *
427 	 * NOTE!  vm_map_lookup will try to upgrade the fault_type to
428 	 *	  VM_FAULT_WRITE if the map entry is a virtual page table
429 	 *	  and also writable, so we can set the 'A'accessed bit in
430 	 *	  the virtual page table entry.
431 	 */
432 	fs.map = map;
433 	result = vm_map_lookup(&fs.map, vaddr, fault_type,
434 			       &fs.entry, &fs.first_ba,
435 			       &first_pindex, &fs.first_prot, &fs.wflags);
436 
437 	/*
438 	 * If the lookup failed or the map protections are incompatible,
439 	 * the fault generally fails.
440 	 *
441 	 * The failure could be due to TDF_NOFAULT if vm_map_lookup()
442 	 * tried to do a COW fault.
443 	 *
444 	 * If the caller is trying to do a user wiring we have more work
445 	 * to do.
446 	 */
447 	if (result != KERN_SUCCESS) {
448 		if (result == KERN_FAILURE_NOFAULT) {
449 			result = KERN_FAILURE;
450 			goto done;
451 		}
452 		if (result != KERN_PROTECTION_FAILURE ||
453 		    (fs.fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE)
454 		{
455 			if (result == KERN_INVALID_ADDRESS && growstack &&
456 			    map != &kernel_map && curproc != NULL) {
457 				result = vm_map_growstack(map, vaddr);
458 				if (result == KERN_SUCCESS) {
459 					growstack = 0;
460 					++retry;
461 					goto RetryFault;
462 				}
463 				result = KERN_FAILURE;
464 			}
465 			goto done;
466 		}
467 
468 		/*
469 		 * If we are user-wiring a r/w segment, and it is COW, then
470 		 * we need to do the COW operation.  Note that we don't
471 		 * currently COW RO sections now, because it is NOT desirable
472 		 * to COW .text.  We simply keep .text from ever being COW'ed
473 		 * and take the heat that one cannot debug wired .text sections.
474 		 *
475 		 * XXX Try to allow the above by specifying OVERRIDE_WRITE.
476 		 */
477 		result = vm_map_lookup(&fs.map, vaddr,
478 				       VM_PROT_READ|VM_PROT_WRITE|
479 				        VM_PROT_OVERRIDE_WRITE,
480 				       &fs.entry, &fs.first_ba,
481 				       &first_pindex, &fs.first_prot,
482 				       &fs.wflags);
483 		if (result != KERN_SUCCESS) {
484 			/* could also be KERN_FAILURE_NOFAULT */
485 			result = KERN_FAILURE;
486 			goto done;
487 		}
488 
489 		/*
490 		 * If we don't COW now, on a user wire, the user will never
491 		 * be able to write to the mapping.  If we don't make this
492 		 * restriction, the bookkeeping would be nearly impossible.
493 		 *
494 		 * XXX We have a shared lock, this will have a MP race but
495 		 * I don't see how it can hurt anything.
496 		 */
497 		if ((fs.entry->protection & VM_PROT_WRITE) == 0) {
498 			atomic_clear_char(&fs.entry->max_protection,
499 					  VM_PROT_WRITE);
500 		}
501 	}
502 
503 	/*
504 	 * fs.map is read-locked
505 	 *
506 	 * Misc checks.  Save the map generation number to detect races.
507 	 */
508 	fs.lookup_still_valid = 1;
509 	fs.first_m = NULL;
510 	fs.ba = fs.first_ba;		/* so unlock_things() works */
511 	fs.prot = fs.first_prot;	/* default (used by uksmap) */
512 
513 	if (fs.entry->eflags & (MAP_ENTRY_NOFAULT | MAP_ENTRY_KSTACK)) {
514 		if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
515 			panic("vm_fault: fault on nofault entry, addr: %p",
516 			      (void *)vaddr);
517 		}
518 		if ((fs.entry->eflags & MAP_ENTRY_KSTACK) &&
519 		    vaddr >= fs.entry->start &&
520 		    vaddr < fs.entry->start + PAGE_SIZE) {
521 			panic("vm_fault: fault on stack guard, addr: %p",
522 			      (void *)vaddr);
523 		}
524 	}
525 
526 	/*
527 	 * A user-kernel shared map has no VM object and bypasses
528 	 * everything.  We execute the uksmap function with a temporary
529 	 * fictitious vm_page.  The address is directly mapped with no
530 	 * management.
531 	 */
532 	if (fs.entry->maptype == VM_MAPTYPE_UKSMAP) {
533 		struct vm_page fakem;
534 
535 		bzero(&fakem, sizeof(fakem));
536 		fakem.pindex = first_pindex;
537 		fakem.flags = PG_FICTITIOUS | PG_UNMANAGED;
538 		fakem.busy_count = PBUSY_LOCKED;
539 		fakem.valid = VM_PAGE_BITS_ALL;
540 		fakem.pat_mode = VM_MEMATTR_DEFAULT;
541 		if (fs.entry->ba.uksmap(fs.entry->aux.dev, &fakem)) {
542 			result = KERN_FAILURE;
543 			unlock_things(&fs);
544 			goto done2;
545 		}
546 		pmap_enter(fs.map->pmap, vaddr, &fakem, fs.prot | inherit_prot,
547 			   (fs.wflags & FW_WIRED), fs.entry);
548 		goto done_success;
549 	}
550 
551 	/*
552 	 * A system map entry may return a NULL object.  No object means
553 	 * no pager means an unrecoverable kernel fault.
554 	 */
555 	if (fs.first_ba == NULL) {
556 		panic("vm_fault: unrecoverable fault at %p in entry %p",
557 			(void *)vaddr, fs.entry);
558 	}
559 
560 	/*
561 	 * Fail here if not a trivial anonymous page fault and TDF_NOFAULT
562 	 * is set.
563 	 *
564 	 * Unfortunately a deadlock can occur if we are forced to page-in
565 	 * from swap, but diving all the way into the vm_pager_get_page()
566 	 * function to find out is too much.  Just check the object type.
567 	 *
568 	 * The deadlock is a CAM deadlock on a busy VM page when trying
569 	 * to finish an I/O if another process gets stuck in
570 	 * vop_helper_read_shortcut() due to a swap fault.
571 	 */
572 	if ((td->td_flags & TDF_NOFAULT) &&
573 	    (retry ||
574 	     fs.first_ba->object->type == OBJT_VNODE ||
575 	     fs.first_ba->object->type == OBJT_SWAP ||
576 	     fs.first_ba->backing_ba)) {
577 		result = KERN_FAILURE;
578 		unlock_things(&fs);
579 		goto done2;
580 	}
581 
582 	/*
583 	 * If the entry is wired we cannot change the page protection.
584 	 */
585 	if (fs.wflags & FW_WIRED)
586 		fault_type = fs.first_prot;
587 
588 	/*
589 	 * We generally want to avoid unnecessary exclusive modes on backing
590 	 * and terminal objects because this can seriously interfere with
591 	 * heavily fork()'d processes (particularly /bin/sh scripts).
592 	 *
593 	 * However, we also want to avoid unnecessary retries due to needed
594 	 * shared->exclusive promotion for common faults.  Exclusive mode is
595 	 * always needed if any page insertion, rename, or free occurs in an
596 	 * object (and also indirectly if any I/O is done).
597 	 *
598 	 * The main issue here is going to be fs.first_shared.  If the
599 	 * first_object has a backing object which isn't shadowed and the
600 	 * process is single-threaded we might as well use an exclusive
601 	 * lock/chain right off the bat.
602 	 */
603 #if 0
604 	/* WORK IN PROGRESS, CODE REMOVED */
605 	if (fs.first_shared && fs.first_object->backing_object &&
606 	    LIST_EMPTY(&fs.first_object->shadow_head) &&
607 	    td->td_proc && td->td_proc->p_nthreads == 1) {
608 		fs.first_shared = 0;
609 	}
610 #endif
611 
612 	/*
613 	 * VM_FAULT_UNSWAP - swap_pager_unswapped() needs an exclusive object
614 	 * VM_FAULT_DIRTY  - may require swap_pager_unswapped() later, but
615 	 *		     we can try shared first.
616 	 */
617 	if (fault_flags & VM_FAULT_UNSWAP)
618 		fs.first_shared = 0;
619 
620 	/*
621 	 * Try to shortcut the entire mess and run the fault lockless.
622 	 */
623 	if (vm_fault_quick_enable &&
624 	    vm_fault_quick(&fs, first_pindex, fault_type) == KERN_SUCCESS) {
625 		didilock = 0;
626 		fault_flags &= ~VM_FAULT_BURST;
627 		goto success;
628 	}
629 
630 	/*
631 	 * Exclusive heuristic (alloc page vs page exists)
632 	 */
633 	if (fs.first_ba->flags & VM_MAP_BACK_EXCL_HEUR)
634 		fs.first_shared = 0;
635 
636 	/*
637 	 * Obtain a top-level object lock, shared or exclusive depending
638 	 * on fs.first_shared.  If a shared lock winds up being insufficient
639 	 * we will retry with an exclusive lock.
640 	 *
641 	 * The vnode pager lock is always shared.
642 	 */
643 	if (fs.first_shared)
644 		vm_object_hold_shared(fs.first_ba->object);
645 	else
646 		vm_object_hold(fs.first_ba->object);
647 	if (fs.vp == NULL)
648 		fs.vp = vnode_pager_lock(fs.first_ba);
649 	fs.first_ba_held = 1;
650 
651 	/*
652 	 * The page we want is at (first_object, first_pindex), but if the
653 	 * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the
654 	 * page table to figure out the actual pindex.
655 	 *
656 	 * NOTE!  DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION
657 	 * ONLY
658 	 */
659 	didilock = 0;
660 	if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) {
661 		vm_map_interlock(fs.map, &ilock, vaddr, vaddr + PAGE_SIZE);
662 		didilock = 1;
663 		result = vm_fault_vpagetable(&fs, &first_pindex,
664 					     fs.entry->aux.master_pde,
665 					     fault_type, 1);
666 		if (result == KERN_TRY_AGAIN) {
667 			vm_map_deinterlock(fs.map, &ilock);
668 			++retry;
669 			goto RetryFault;
670 		}
671 		if (result != KERN_SUCCESS) {
672 			vm_map_deinterlock(fs.map, &ilock);
673 			goto done;
674 		}
675 	}
676 
677 	/*
678 	 * Now we have the actual (object, pindex), fault in the page.  If
679 	 * vm_fault_object() fails it will unlock and deallocate the FS
680 	 * data.   If it succeeds everything remains locked and fs->ba->object
681 	 * will have an additional PIP count if fs->ba != fs->first_ba.
682 	 *
683 	 * vm_fault_object will set fs->prot for the pmap operation.  It is
684 	 * allowed to set VM_PROT_WRITE if fault_type == VM_PROT_READ if the
685 	 * page can be safely written.  However, it will force a read-only
686 	 * mapping for a read fault if the memory is managed by a virtual
687 	 * page table.
688 	 *
689 	 * If the fault code uses the shared object lock shortcut
690 	 * we must not try to burst (we can't allocate VM pages).
691 	 */
692 	result = vm_fault_object(&fs, first_pindex, fault_type, 1);
693 
694 	if (debug_fault > 0) {
695 		--debug_fault;
696 		kprintf("VM_FAULT result %d addr=%jx type=%02x flags=%02x "
697 			"fs.m=%p fs.prot=%02x fs.wflags=%02x fs.entry=%p\n",
698 			result, (intmax_t)vaddr, fault_type, fault_flags,
699 			fs.m, fs.prot, fs.wflags, fs.entry);
700 	}
701 
702 	if (result == KERN_TRY_AGAIN) {
703 		if (didilock)
704 			vm_map_deinterlock(fs.map, &ilock);
705 		++retry;
706 		goto RetryFault;
707 	}
708 	if (result != KERN_SUCCESS) {
709 		if (didilock)
710 			vm_map_deinterlock(fs.map, &ilock);
711 		goto done;
712 	}
713 
714 success:
715 	/*
716 	 * On success vm_fault_object() does not unlock or deallocate, and fs.m
717 	 * will contain a busied page.  It does drop fs->ba if appropriate.
718 	 *
719 	 * Enter the page into the pmap and do pmap-related adjustments.
720 	 *
721 	 * WARNING! Soft-busied fs.m's can only be manipulated in limited
722 	 *	    ways.
723 	 */
724 	KKASSERT(fs.lookup_still_valid != 0);
725 	vm_page_flag_set(fs.m, PG_REFERENCED);
726 	pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot | inherit_prot,
727 		   fs.wflags & FW_WIRED, fs.entry);
728 
729 	if (didilock)
730 		vm_map_deinterlock(fs.map, &ilock);
731 
732 	/*
733 	 * If the page is not wired down, then put it where the pageout daemon
734 	 * can find it.
735 	 *
736 	 * NOTE: We cannot safely wire, unwire, or adjust queues for a
737 	 *	 soft-busied page.
738 	 */
739 	if (fs.msoftonly) {
740 		KKASSERT(fs.m->busy_count & PBUSY_MASK);
741 		KKASSERT((fs.fault_flags & VM_FAULT_WIRE_MASK) == 0);
742 		vm_page_sbusy_drop(fs.m);
743 	} else {
744 		if (fs.fault_flags & VM_FAULT_WIRE_MASK) {
745 			if (fs.wflags & FW_WIRED)
746 				vm_page_wire(fs.m);
747 			else
748 				vm_page_unwire(fs.m, 1);
749 		} else {
750 			vm_page_activate(fs.m);
751 		}
752 		KKASSERT(fs.m->busy_count & PBUSY_LOCKED);
753 		vm_page_wakeup(fs.m);
754 	}
755 
756 	/*
757 	 * Burst in a few more pages if possible.  The fs.map should still
758 	 * be locked.  To avoid interlocking against a vnode->getblk
759 	 * operation we had to be sure to unbusy our primary vm_page above
760 	 * first.
761 	 *
762 	 * A normal burst can continue down backing store, only execute
763 	 * if we are holding an exclusive lock, otherwise the exclusive
764 	 * locks the burst code gets might cause excessive SMP collisions.
765 	 *
766 	 * A quick burst can be utilized when there is no backing object
767 	 * (i.e. a shared file mmap).
768 	 */
769 	if ((fault_flags & VM_FAULT_BURST) &&
770 	    (fs.fault_flags & VM_FAULT_WIRE_MASK) == 0 &&
771 	    (fs.wflags & FW_WIRED) == 0) {
772 		if (fs.first_shared == 0 && fs.shared == 0) {
773 			vm_prefault(fs.map->pmap, vaddr,
774 				    fs.entry, fs.prot, fault_flags);
775 		} else {
776 			vm_prefault_quick(fs.map->pmap, vaddr,
777 					  fs.entry, fs.prot, fault_flags);
778 		}
779 	}
780 
781 done_success:
782 	mycpu->gd_cnt.v_vm_faults++;
783 	if (td->td_lwp)
784 		++td->td_lwp->lwp_ru.ru_minflt;
785 
786 	/*
787 	 * Unlock everything, and return
788 	 */
789 	unlock_things(&fs);
790 
791 	if (td->td_lwp) {
792 		if (fs.hardfault) {
793 			td->td_lwp->lwp_ru.ru_majflt++;
794 		} else {
795 			td->td_lwp->lwp_ru.ru_minflt++;
796 		}
797 	}
798 
799 	/*vm_object_deallocate(fs.first_ba->object);*/
800 	/*fs.m = NULL; */
801 
802 	result = KERN_SUCCESS;
803 done:
804 	if (fs.first_ba && fs.first_ba->object && fs.first_ba_held == 1) {
805 		vm_object_drop(fs.first_ba->object);
806 		fs.first_ba_held = 0;
807 	}
808 done2:
809 	if (lp)
810 		lp->lwp_flags &= ~LWP_PAGING;
811 
812 #if !defined(NO_SWAPPING)
813 	/*
814 	 * Check the process RSS limit and force deactivation and
815 	 * (asynchronous) paging if necessary.  This is a complex operation,
816 	 * only do it for direct user-mode faults, for now.
817 	 *
818 	 * To reduce overhead implement approximately a ~16MB hysteresis.
819 	 */
820 	p = td->td_proc;
821 	if ((fault_flags & VM_FAULT_USERMODE) && lp &&
822 	    p->p_limit && map->pmap && vm_pageout_memuse_mode >= 1 &&
823 	    map != &kernel_map) {
824 		vm_pindex_t limit;
825 		vm_pindex_t size;
826 
827 		limit = OFF_TO_IDX(qmin(p->p_rlimit[RLIMIT_RSS].rlim_cur,
828 					p->p_rlimit[RLIMIT_RSS].rlim_max));
829 		size = pmap_resident_tlnw_count(map->pmap);
830 		if (limit >= 0 && size > 4096 && size - 4096 >= limit) {
831 			vm_pageout_map_deactivate_pages(map, limit);
832 		}
833 	}
834 #endif
835 
836 	if (result != KERN_SUCCESS && debug_fault < 0) {
837 		kprintf("VM_FAULT %d:%d (%s) result %d "
838 			"addr=%jx type=%02x flags=%02x "
839 			"fs.m=%p fs.prot=%02x fs.wflags=%02x fs.entry=%p\n",
840 			(curthread->td_proc ? curthread->td_proc->p_pid : -1),
841 			(curthread->td_lwp ? curthread->td_lwp->lwp_tid : -1),
842 			curthread->td_comm,
843 			result,
844 			(intmax_t)vaddr, fault_type, fault_flags,
845 			fs.m, fs.prot, fs.wflags, fs.entry);
846 		while (debug_fault < 0 && (debug_fault & 1))
847 			tsleep(&debug_fault, 0, "DEBUG", hz);
848 	}
849 
850 	return (result);
851 }
852 
853 /*
854  * Attempt a lockless vm_fault() shortcut.  The stars have to align for this
855  * to work.  But if it does we can get our page only soft-busied and not
856  * have to touch the vm_object or vnode locks at all.
857  */
858 static
859 int
860 vm_fault_quick(struct faultstate *fs, vm_pindex_t first_pindex,
861 	       vm_prot_t fault_type)
862 {
863 	vm_page_t m;
864 	vm_object_t obj;	/* NOT LOCKED */
865 
866 	/*
867 	 * Don't waste time if the object is only being used by one vm_map.
868 	 *
869 	 * WARNING! We can't rely on obj->ref_count here because it might
870 	 *	    be part of a shared ba chain, and we can't rely on
871 	 *	    ba->refs for the same reason.  The combination of it
872 	 *	    being the ba embedded in the entry (aka first_ba) AND
873 	 *	    ref_count == 1 would work, but OBJ_ONEMAPPING is better
874 	 *	    because it will remain flagged even when ref_count > 1
875 	 *	    for situations where entries are clipped.
876 	 */
877 	obj = fs->first_ba->object;
878 	if (obj->flags & OBJ_ONEMAPPING)
879 		return KERN_FAILURE;
880 
881 	/*
882 	 * This will try to wire/unwire a page, which can't be done with
883 	 * a soft-busied page.
884 	 */
885 	if (fs->fault_flags & VM_FAULT_WIRE_MASK)
886 		return KERN_FAILURE;
887 
888 	/*
889 	 * Ick, can't handle this
890 	 */
891 	if (fs->entry->maptype == VM_MAPTYPE_VPAGETABLE) {
892 #ifdef VM_FAULT_QUICK_DEBUG
893 		++vm_fault_quick_failure_count1;
894 #endif
895 		return KERN_FAILURE;
896 	}
897 
898 	/*
899 	 * Ok, try to get the vm_page quickly via the hash table.  The
900 	 * page will be soft-busied on success (NOT hard-busied).
901 	 */
902 	m = vm_page_hash_get(obj, first_pindex);
903 	if (m == NULL) {
904 #ifdef VM_FAULT_QUICK_DEBUG
905 		++vm_fault_quick_failure_count2;
906 #endif
907 		return KERN_FAILURE;
908 	}
909 	if ((obj->flags & OBJ_DEAD) ||
910 	    m->valid != VM_PAGE_BITS_ALL ||
911 	    m->queue - m->pc == PQ_CACHE ||
912 	    (m->flags & PG_SWAPPED)) {
913 		vm_page_sbusy_drop(m);
914 #ifdef VM_FAULT_QUICK_DEBUG
915 		++vm_fault_quick_failure_count3;
916 #endif
917 		return KERN_FAILURE;
918 	}
919 
920 	/*
921 	 * The page is already fully valid, ACTIVE, and is not PG_SWAPPED.
922 	 *
923 	 * Don't map the page writable when emulating the dirty bit, a
924 	 * fault must be taken for proper emulation (vkernel).
925 	 */
926 	if (curthread->td_lwp && curthread->td_lwp->lwp_vmspace &&
927 	    pmap_emulate_ad_bits(&curthread->td_lwp->lwp_vmspace->vm_pmap)) {
928 		if ((fault_type & VM_PROT_WRITE) == 0)
929 			fs->prot &= ~VM_PROT_WRITE;
930 	}
931 
932 	/*
933 	 * If this is a write fault the object and the page must already
934 	 * be writable.  Since we don't hold an object lock and only a
935 	 * soft-busy on the page, we cannot manipulate the object or
936 	 * the page state (other than the page queue).
937 	 */
938 	if (fs->prot & VM_PROT_WRITE) {
939 		if ((obj->flags & (OBJ_WRITEABLE | OBJ_MIGHTBEDIRTY)) !=
940 		    (OBJ_WRITEABLE | OBJ_MIGHTBEDIRTY) ||
941 		    m->dirty != VM_PAGE_BITS_ALL) {
942 			vm_page_sbusy_drop(m);
943 #ifdef VM_FAULT_QUICK_DEBUG
944 			++vm_fault_quick_failure_count4;
945 #endif
946 			return KERN_FAILURE;
947 		}
948 		vm_set_nosync(m, fs->entry);
949 	}
950 
951 	/*
952 	 * Even though we are only soft-busied we can still move pages
953 	 * around in the normal queue(s).  The soft-busy prevents the
954 	 * page from being removed from the object, etc (normal operation).
955 	 */
956 	vm_page_activate(m);
957 	fs->m = m;
958 	fs->msoftonly = 1;
959 #ifdef VM_FAULT_QUICK_DEBUG
960 	++vm_fault_quick_success_count;
961 #endif
962 
963 	return KERN_SUCCESS;
964 }
965 
966 /*
967  * Fault in the specified virtual address in the current process map,
968  * returning a held VM page or NULL.  See vm_fault_page() for more
969  * information.
970  *
971  * No requirements.
972  */
973 vm_page_t
974 vm_fault_page_quick(vm_offset_t va, vm_prot_t fault_type,
975 		    int *errorp, int *busyp)
976 {
977 	struct lwp *lp = curthread->td_lwp;
978 	vm_page_t m;
979 
980 	m = vm_fault_page(&lp->lwp_vmspace->vm_map, va,
981 			  fault_type, VM_FAULT_NORMAL,
982 			  errorp, busyp);
983 	return(m);
984 }
985 
986 /*
987  * Fault in the specified virtual address in the specified map, doing all
988  * necessary manipulation of the object store and all necessary I/O.  Return
989  * a held VM page or NULL, and set *errorp.  The related pmap is not
990  * updated.
991  *
992  * If busyp is not NULL then *busyp will be set to TRUE if this routine
993  * decides to return a busied page (aka VM_PROT_WRITE), or FALSE if it
994  * does not (VM_PROT_WRITE not specified or busyp is NULL).  If busyp is
995  * NULL the returned page is only held.
996  *
997  * If the caller has no intention of writing to the page's contents, busyp
998  * can be passed as NULL along with VM_PROT_WRITE to force a COW operation
999  * without busying the page.
1000  *
1001  * The returned page will also be marked PG_REFERENCED.
1002  *
1003  * If the page cannot be faulted writable and VM_PROT_WRITE was specified, an
1004  * error will be returned.
1005  *
1006  * No requirements.
1007  */
1008 vm_page_t
1009 vm_fault_page(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
1010 	      int fault_flags, int *errorp, int *busyp)
1011 {
1012 	vm_pindex_t first_pindex;
1013 	struct faultstate fs;
1014 	int result;
1015 	int retry;
1016 	int growstack;
1017 	int didcow;
1018 	vm_prot_t orig_fault_type = fault_type;
1019 
1020 	retry = 0;
1021 	didcow = 0;
1022 	fs.hardfault = 0;
1023 	fs.fault_flags = fault_flags;
1024 	KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0);
1025 
1026 	/*
1027 	 * Dive the pmap (concurrency possible).  If we find the
1028 	 * appropriate page we can terminate early and quickly.
1029 	 *
1030 	 * This works great for normal programs but will always return
1031 	 * NULL for host lookups of vkernel maps in VMM mode.
1032 	 *
1033 	 * NOTE: pmap_fault_page_quick() might not busy the page.  If
1034 	 *	 VM_PROT_WRITE is set in fault_type and pmap_fault_page_quick()
1035 	 *	 returns non-NULL, it will safely dirty the returned vm_page_t
1036 	 *	 for us.  We cannot safely dirty it here (it might not be
1037 	 *	 busy).
1038 	 */
1039 	fs.m = pmap_fault_page_quick(map->pmap, vaddr, fault_type, busyp);
1040 	if (fs.m) {
1041 		*errorp = 0;
1042 		return(fs.m);
1043 	}
1044 
1045 	/*
1046 	 * Otherwise take a concurrency hit and do a formal page
1047 	 * fault.
1048 	 */
1049 	fs.vp = NULL;
1050 	fs.shared = vm_shared_fault;
1051 	fs.first_shared = vm_shared_fault;
1052 	fs.msoftonly = 0;
1053 	growstack = 1;
1054 
1055 	/*
1056 	 * VM_FAULT_UNSWAP - swap_pager_unswapped() needs an exclusive object
1057 	 * VM_FAULT_DIRTY  - may require swap_pager_unswapped() later, but
1058 	 *		     we can try shared first.
1059 	 */
1060 	if (fault_flags & VM_FAULT_UNSWAP) {
1061 		fs.first_shared = 0;
1062 	}
1063 
1064 RetryFault:
1065 	/*
1066 	 * Find the vm_map_entry representing the backing store and resolve
1067 	 * the top level object and page index.  This may have the side
1068 	 * effect of executing a copy-on-write on the map entry and/or
1069 	 * creating a shadow object, but will not COW any actual VM pages.
1070 	 *
1071 	 * On success fs.map is left read-locked and various other fields
1072 	 * are initialized but not otherwise referenced or locked.
1073 	 *
1074 	 * NOTE!  vm_map_lookup will upgrade the fault_type to VM_FAULT_WRITE
1075 	 *	  if the map entry is a virtual page table and also writable,
1076 	 *	  so we can set the 'A'accessed bit in the virtual page table
1077 	 *	  entry.
1078 	 */
1079 	fs.map = map;
1080 	fs.first_ba_held = 0;
1081 	result = vm_map_lookup(&fs.map, vaddr, fault_type,
1082 			       &fs.entry, &fs.first_ba,
1083 			       &first_pindex, &fs.first_prot, &fs.wflags);
1084 
1085 	if (result != KERN_SUCCESS) {
1086 		if (result == KERN_FAILURE_NOFAULT) {
1087 			*errorp = KERN_FAILURE;
1088 			fs.m = NULL;
1089 			goto done;
1090 		}
1091 		if (result != KERN_PROTECTION_FAILURE ||
1092 		    (fs.fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE)
1093 		{
1094 			if (result == KERN_INVALID_ADDRESS && growstack &&
1095 			    map != &kernel_map && curproc != NULL) {
1096 				result = vm_map_growstack(map, vaddr);
1097 				if (result == KERN_SUCCESS) {
1098 					growstack = 0;
1099 					++retry;
1100 					goto RetryFault;
1101 				}
1102 				result = KERN_FAILURE;
1103 			}
1104 			fs.m = NULL;
1105 			*errorp = result;
1106 			goto done;
1107 		}
1108 
1109 		/*
1110 		 * If we are user-wiring a r/w segment, and it is COW, then
1111 		 * we need to do the COW operation.  Note that we don't
1112 		 * currently COW RO sections now, because it is NOT desirable
1113 		 * to COW .text.  We simply keep .text from ever being COW'ed
1114 		 * and take the heat that one cannot debug wired .text sections.
1115 		 */
1116 		result = vm_map_lookup(&fs.map, vaddr,
1117 				       VM_PROT_READ|VM_PROT_WRITE|
1118 				        VM_PROT_OVERRIDE_WRITE,
1119 				       &fs.entry, &fs.first_ba,
1120 				       &first_pindex, &fs.first_prot,
1121 				       &fs.wflags);
1122 		if (result != KERN_SUCCESS) {
1123 			/* could also be KERN_FAILURE_NOFAULT */
1124 			*errorp = KERN_FAILURE;
1125 			fs.m = NULL;
1126 			goto done;
1127 		}
1128 
1129 		/*
1130 		 * If we don't COW now, on a user wire, the user will never
1131 		 * be able to write to the mapping.  If we don't make this
1132 		 * restriction, the bookkeeping would be nearly impossible.
1133 		 *
1134 		 * XXX We have a shared lock, this will have a MP race but
1135 		 * I don't see how it can hurt anything.
1136 		 */
1137 		if ((fs.entry->protection & VM_PROT_WRITE) == 0) {
1138 			atomic_clear_char(&fs.entry->max_protection,
1139 					  VM_PROT_WRITE);
1140 		}
1141 	}
1142 
1143 	/*
1144 	 * fs.map is read-locked
1145 	 *
1146 	 * Misc checks.  Save the map generation number to detect races.
1147 	 */
1148 	fs.lookup_still_valid = 1;
1149 	fs.first_m = NULL;
1150 	fs.ba = fs.first_ba;
1151 
1152 	if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
1153 		panic("vm_fault: fault on nofault entry, addr: %lx",
1154 		    (u_long)vaddr);
1155 	}
1156 
1157 	/*
1158 	 * A user-kernel shared map has no VM object and bypasses
1159 	 * everything.  We execute the uksmap function with a temporary
1160 	 * fictitious vm_page.  The address is directly mapped with no
1161 	 * management.
1162 	 */
1163 	if (fs.entry->maptype == VM_MAPTYPE_UKSMAP) {
1164 		struct vm_page fakem;
1165 
1166 		bzero(&fakem, sizeof(fakem));
1167 		fakem.pindex = first_pindex;
1168 		fakem.flags = PG_FICTITIOUS | PG_UNMANAGED;
1169 		fakem.busy_count = PBUSY_LOCKED;
1170 		fakem.valid = VM_PAGE_BITS_ALL;
1171 		fakem.pat_mode = VM_MEMATTR_DEFAULT;
1172 		if (fs.entry->ba.uksmap(fs.entry->aux.dev, &fakem)) {
1173 			*errorp = KERN_FAILURE;
1174 			fs.m = NULL;
1175 			unlock_things(&fs);
1176 			goto done2;
1177 		}
1178 		fs.m = PHYS_TO_VM_PAGE(fakem.phys_addr);
1179 		vm_page_hold(fs.m);
1180 		if (busyp)
1181 			*busyp = 0;	/* don't need to busy R or W */
1182 		unlock_things(&fs);
1183 		*errorp = 0;
1184 		goto done;
1185 	}
1186 
1187 
1188 	/*
1189 	 * A system map entry may return a NULL object.  No object means
1190 	 * no pager means an unrecoverable kernel fault.
1191 	 */
1192 	if (fs.first_ba == NULL) {
1193 		panic("vm_fault: unrecoverable fault at %p in entry %p",
1194 			(void *)vaddr, fs.entry);
1195 	}
1196 
1197 	/*
1198 	 * Fail here if not a trivial anonymous page fault and TDF_NOFAULT
1199 	 * is set.
1200 	 *
1201 	 * Unfortunately a deadlock can occur if we are forced to page-in
1202 	 * from swap, but diving all the way into the vm_pager_get_page()
1203 	 * function to find out is too much.  Just check the object type.
1204 	 */
1205 	if ((curthread->td_flags & TDF_NOFAULT) &&
1206 	    (retry ||
1207 	     fs.first_ba->object->type == OBJT_VNODE ||
1208 	     fs.first_ba->object->type == OBJT_SWAP ||
1209 	     fs.first_ba->backing_ba)) {
1210 		*errorp = KERN_FAILURE;
1211 		unlock_things(&fs);
1212 		fs.m = NULL;
1213 		goto done2;
1214 	}
1215 
1216 	/*
1217 	 * If the entry is wired we cannot change the page protection.
1218 	 */
1219 	if (fs.wflags & FW_WIRED)
1220 		fault_type = fs.first_prot;
1221 
1222 	/*
1223 	 * Make a reference to this object to prevent its disposal while we
1224 	 * are messing with it.  Once we have the reference, the map is free
1225 	 * to be diddled.  Since objects reference their shadows (and copies),
1226 	 * they will stay around as well.
1227 	 *
1228 	 * The reference should also prevent an unexpected collapse of the
1229 	 * parent that might move pages from the current object into the
1230 	 * parent unexpectedly, resulting in corruption.
1231 	 *
1232 	 * Bump the paging-in-progress count to prevent size changes (e.g.
1233 	 * truncation operations) during I/O.  This must be done after
1234 	 * obtaining the vnode lock in order to avoid possible deadlocks.
1235 	 */
1236 	if (fs.first_ba->flags & VM_MAP_BACK_EXCL_HEUR)
1237 		fs.first_shared = 0;
1238 
1239 	if (fs.first_shared)
1240 		vm_object_hold_shared(fs.first_ba->object);
1241 	else
1242 		vm_object_hold(fs.first_ba->object);
1243 	fs.first_ba_held = 1;
1244 	if (fs.vp == NULL)
1245 		fs.vp = vnode_pager_lock(fs.first_ba);	/* shared */
1246 
1247 	/*
1248 	 * The page we want is at (first_object, first_pindex), but if the
1249 	 * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the
1250 	 * page table to figure out the actual pindex.
1251 	 *
1252 	 * NOTE!  DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION
1253 	 * ONLY
1254 	 */
1255 	if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) {
1256 		result = vm_fault_vpagetable(&fs, &first_pindex,
1257 					     fs.entry->aux.master_pde,
1258 					     fault_type, 1);
1259 		if (result == KERN_TRY_AGAIN) {
1260 			++retry;
1261 			goto RetryFault;
1262 		}
1263 		if (result != KERN_SUCCESS) {
1264 			*errorp = result;
1265 			fs.m = NULL;
1266 			goto done;
1267 		}
1268 	}
1269 
1270 	/*
1271 	 * Now we have the actual (object, pindex), fault in the page.  If
1272 	 * vm_fault_object() fails it will unlock and deallocate the FS
1273 	 * data.   If it succeeds everything remains locked and fs->ba->object
1274 	 * will have an additinal PIP count if fs->ba != fs->first_ba.
1275 	 */
1276 	fs.m = NULL;
1277 	result = vm_fault_object(&fs, first_pindex, fault_type, 1);
1278 
1279 	if (result == KERN_TRY_AGAIN) {
1280 		KKASSERT(fs.first_ba_held == 0);
1281 		++retry;
1282 		didcow |= fs.wflags & FW_DIDCOW;
1283 		goto RetryFault;
1284 	}
1285 	if (result != KERN_SUCCESS) {
1286 		*errorp = result;
1287 		fs.m = NULL;
1288 		goto done;
1289 	}
1290 
1291 	if ((orig_fault_type & VM_PROT_WRITE) &&
1292 	    (fs.prot & VM_PROT_WRITE) == 0) {
1293 		*errorp = KERN_PROTECTION_FAILURE;
1294 		unlock_things(&fs);
1295 		fs.m = NULL;
1296 		goto done;
1297 	}
1298 
1299 	/*
1300 	 * Generally speaking we don't want to update the pmap because
1301 	 * this routine can be called many times for situations that do
1302 	 * not require updating the pmap, not to mention the page might
1303 	 * already be in the pmap.
1304 	 *
1305 	 * However, if our vm_map_lookup() results in a COW, we need to
1306 	 * at least remove the pte from the pmap to guarantee proper
1307 	 * visibility of modifications made to the process.  For example,
1308 	 * modifications made by vkernel uiocopy/related routines and
1309 	 * modifications made by ptrace().
1310 	 */
1311 	vm_page_flag_set(fs.m, PG_REFERENCED);
1312 #if 0
1313 	pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot,
1314 		   fs.wflags & FW_WIRED, NULL);
1315 	mycpu->gd_cnt.v_vm_faults++;
1316 	if (curthread->td_lwp)
1317 		++curthread->td_lwp->lwp_ru.ru_minflt;
1318 #endif
1319 	if ((fs.wflags | didcow) | FW_DIDCOW) {
1320 		pmap_remove(fs.map->pmap,
1321 			    vaddr & ~PAGE_MASK,
1322 			    (vaddr & ~PAGE_MASK) + PAGE_SIZE);
1323 	}
1324 
1325 	/*
1326 	 * On success vm_fault_object() does not unlock or deallocate, and fs.m
1327 	 * will contain a busied page.  So we must unlock here after having
1328 	 * messed with the pmap.
1329 	 */
1330 	unlock_things(&fs);
1331 
1332 	/*
1333 	 * Return a held page.  We are not doing any pmap manipulation so do
1334 	 * not set PG_MAPPED.  However, adjust the page flags according to
1335 	 * the fault type because the caller may not use a managed pmapping
1336 	 * (so we don't want to lose the fact that the page will be dirtied
1337 	 * if a write fault was specified).
1338 	 */
1339 	if (fault_type & VM_PROT_WRITE)
1340 		vm_page_dirty(fs.m);
1341 	vm_page_activate(fs.m);
1342 
1343 	if (curthread->td_lwp) {
1344 		if (fs.hardfault) {
1345 			curthread->td_lwp->lwp_ru.ru_majflt++;
1346 		} else {
1347 			curthread->td_lwp->lwp_ru.ru_minflt++;
1348 		}
1349 	}
1350 
1351 	/*
1352 	 * Unlock everything, and return the held or busied page.
1353 	 */
1354 	if (busyp) {
1355 		if (fault_type & VM_PROT_WRITE) {
1356 			vm_page_dirty(fs.m);
1357 			*busyp = 1;
1358 		} else {
1359 			*busyp = 0;
1360 			vm_page_hold(fs.m);
1361 			vm_page_wakeup(fs.m);
1362 		}
1363 	} else {
1364 		vm_page_hold(fs.m);
1365 		vm_page_wakeup(fs.m);
1366 	}
1367 	/*vm_object_deallocate(fs.first_ba->object);*/
1368 	*errorp = 0;
1369 
1370 done:
1371 	KKASSERT(fs.first_ba_held == 0);
1372 done2:
1373 	return(fs.m);
1374 }
1375 
1376 /*
1377  * Fault in the specified (object,offset), dirty the returned page as
1378  * needed.  If the requested fault_type cannot be done NULL and an
1379  * error is returned.
1380  *
1381  * A held (but not busied) page is returned.
1382  *
1383  * The passed in object must be held as specified by the shared
1384  * argument.
1385  */
1386 vm_page_t
1387 vm_fault_object_page(vm_object_t object, vm_ooffset_t offset,
1388 		     vm_prot_t fault_type, int fault_flags,
1389 		     int *sharedp, int *errorp)
1390 {
1391 	int result;
1392 	vm_pindex_t first_pindex;
1393 	struct faultstate fs;
1394 	struct vm_map_entry entry;
1395 
1396 	ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1397 	bzero(&entry, sizeof(entry));
1398 	entry.maptype = VM_MAPTYPE_NORMAL;
1399 	entry.protection = entry.max_protection = fault_type;
1400 	entry.ba.backing_ba = NULL;
1401 	entry.ba.object = object;
1402 	entry.ba.offset = 0;
1403 
1404 	fs.hardfault = 0;
1405 	fs.fault_flags = fault_flags;
1406 	fs.map = NULL;
1407 	fs.shared = vm_shared_fault;
1408 	fs.first_shared = *sharedp;
1409 	fs.msoftonly = 0;
1410 	fs.vp = NULL;
1411 	fs.first_ba_held = -1;	/* object held across call, prevent drop */
1412 	KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0);
1413 
1414 	/*
1415 	 * VM_FAULT_UNSWAP - swap_pager_unswapped() needs an exclusive object
1416 	 * VM_FAULT_DIRTY  - may require swap_pager_unswapped() later, but
1417 	 *		     we can try shared first.
1418 	 */
1419 	if (fs.first_shared && (fault_flags & VM_FAULT_UNSWAP)) {
1420 		fs.first_shared = 0;
1421 		vm_object_upgrade(object);
1422 	}
1423 
1424 	/*
1425 	 * Retry loop as needed (typically for shared->exclusive transitions)
1426 	 */
1427 RetryFault:
1428 	*sharedp = fs.first_shared;
1429 	first_pindex = OFF_TO_IDX(offset);
1430 	fs.first_ba = &entry.ba;
1431 	fs.ba = fs.first_ba;
1432 	fs.entry = &entry;
1433 	fs.first_prot = fault_type;
1434 	fs.wflags = 0;
1435 
1436 	/*
1437 	 * Make a reference to this object to prevent its disposal while we
1438 	 * are messing with it.  Once we have the reference, the map is free
1439 	 * to be diddled.  Since objects reference their shadows (and copies),
1440 	 * they will stay around as well.
1441 	 *
1442 	 * The reference should also prevent an unexpected collapse of the
1443 	 * parent that might move pages from the current object into the
1444 	 * parent unexpectedly, resulting in corruption.
1445 	 *
1446 	 * Bump the paging-in-progress count to prevent size changes (e.g.
1447 	 * truncation operations) during I/O.  This must be done after
1448 	 * obtaining the vnode lock in order to avoid possible deadlocks.
1449 	 */
1450 	if (fs.vp == NULL)
1451 		fs.vp = vnode_pager_lock(fs.first_ba);
1452 
1453 	fs.lookup_still_valid = 1;
1454 	fs.first_m = NULL;
1455 
1456 #if 0
1457 	/* XXX future - ability to operate on VM object using vpagetable */
1458 	if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) {
1459 		result = vm_fault_vpagetable(&fs, &first_pindex,
1460 					     fs.entry->aux.master_pde,
1461 					     fault_type, 0);
1462 		if (result == KERN_TRY_AGAIN) {
1463 			if (fs.first_shared == 0 && *sharedp)
1464 				vm_object_upgrade(object);
1465 			goto RetryFault;
1466 		}
1467 		if (result != KERN_SUCCESS) {
1468 			*errorp = result;
1469 			return (NULL);
1470 		}
1471 	}
1472 #endif
1473 
1474 	/*
1475 	 * Now we have the actual (object, pindex), fault in the page.  If
1476 	 * vm_fault_object() fails it will unlock and deallocate the FS
1477 	 * data.   If it succeeds everything remains locked and fs->ba->object
1478 	 * will have an additinal PIP count if fs->ba != fs->first_ba.
1479 	 *
1480 	 * On KERN_TRY_AGAIN vm_fault_object() leaves fs.first_ba intact.
1481 	 * We may have to upgrade its lock to handle the requested fault.
1482 	 */
1483 	result = vm_fault_object(&fs, first_pindex, fault_type, 0);
1484 
1485 	if (result == KERN_TRY_AGAIN) {
1486 		if (fs.first_shared == 0 && *sharedp)
1487 			vm_object_upgrade(object);
1488 		goto RetryFault;
1489 	}
1490 	if (result != KERN_SUCCESS) {
1491 		*errorp = result;
1492 		return(NULL);
1493 	}
1494 
1495 	if ((fault_type & VM_PROT_WRITE) && (fs.prot & VM_PROT_WRITE) == 0) {
1496 		*errorp = KERN_PROTECTION_FAILURE;
1497 		unlock_things(&fs);
1498 		return(NULL);
1499 	}
1500 
1501 	/*
1502 	 * On success vm_fault_object() does not unlock or deallocate, so we
1503 	 * do it here.  Note that the returned fs.m will be busied.
1504 	 */
1505 	unlock_things(&fs);
1506 
1507 	/*
1508 	 * Return a held page.  We are not doing any pmap manipulation so do
1509 	 * not set PG_MAPPED.  However, adjust the page flags according to
1510 	 * the fault type because the caller may not use a managed pmapping
1511 	 * (so we don't want to lose the fact that the page will be dirtied
1512 	 * if a write fault was specified).
1513 	 */
1514 	vm_page_hold(fs.m);
1515 	vm_page_activate(fs.m);
1516 	if ((fault_type & VM_PROT_WRITE) || (fault_flags & VM_FAULT_DIRTY))
1517 		vm_page_dirty(fs.m);
1518 	if (fault_flags & VM_FAULT_UNSWAP)
1519 		swap_pager_unswapped(fs.m);
1520 
1521 	/*
1522 	 * Indicate that the page was accessed.
1523 	 */
1524 	vm_page_flag_set(fs.m, PG_REFERENCED);
1525 
1526 	if (curthread->td_lwp) {
1527 		if (fs.hardfault) {
1528 			curthread->td_lwp->lwp_ru.ru_majflt++;
1529 		} else {
1530 			curthread->td_lwp->lwp_ru.ru_minflt++;
1531 		}
1532 	}
1533 
1534 	/*
1535 	 * Unlock everything, and return the held page.
1536 	 */
1537 	vm_page_wakeup(fs.m);
1538 	/*vm_object_deallocate(fs.first_ba->object);*/
1539 
1540 	*errorp = 0;
1541 	return(fs.m);
1542 }
1543 
1544 /*
1545  * Translate the virtual page number (first_pindex) that is relative
1546  * to the address space into a logical page number that is relative to the
1547  * backing object.  Use the virtual page table pointed to by (vpte).
1548  *
1549  * Possibly downgrade the protection based on the vpte bits.
1550  *
1551  * This implements an N-level page table.  Any level can terminate the
1552  * scan by setting VPTE_PS.   A linear mapping is accomplished by setting
1553  * VPTE_PS in the master page directory entry set via mcontrol(MADV_SETMAP).
1554  */
1555 static
1556 int
1557 vm_fault_vpagetable(struct faultstate *fs, vm_pindex_t *pindex,
1558 		    vpte_t vpte, int fault_type, int allow_nofault)
1559 {
1560 	struct lwbuf *lwb;
1561 	struct lwbuf lwb_cache;
1562 	int vshift = VPTE_FRAME_END - PAGE_SHIFT; /* index bits remaining */
1563 	int result;
1564 	vpte_t *ptep;
1565 
1566 	ASSERT_LWKT_TOKEN_HELD(vm_object_token(fs->first_ba->object));
1567 	for (;;) {
1568 		/*
1569 		 * We cannot proceed if the vpte is not valid, not readable
1570 		 * for a read fault, not writable for a write fault, or
1571 		 * not executable for an instruction execution fault.
1572 		 */
1573 		if ((vpte & VPTE_V) == 0) {
1574 			unlock_things(fs);
1575 			return (KERN_FAILURE);
1576 		}
1577 		if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_RW) == 0) {
1578 			unlock_things(fs);
1579 			return (KERN_FAILURE);
1580 		}
1581 		if ((fault_type & VM_PROT_EXECUTE) && (vpte & VPTE_NX)) {
1582 			unlock_things(fs);
1583 			return (KERN_FAILURE);
1584 		}
1585 		if ((vpte & VPTE_PS) || vshift == 0)
1586 			break;
1587 
1588 		/*
1589 		 * Get the page table page.  Nominally we only read the page
1590 		 * table, but since we are actively setting VPTE_M and VPTE_A,
1591 		 * tell vm_fault_object() that we are writing it.
1592 		 *
1593 		 * There is currently no real need to optimize this.
1594 		 */
1595 		result = vm_fault_object(fs, (vpte & VPTE_FRAME) >> PAGE_SHIFT,
1596 					 VM_PROT_READ|VM_PROT_WRITE,
1597 					 allow_nofault);
1598 		if (result != KERN_SUCCESS)
1599 			return (result);
1600 
1601 		/*
1602 		 * Process the returned fs.m and look up the page table
1603 		 * entry in the page table page.
1604 		 */
1605 		vshift -= VPTE_PAGE_BITS;
1606 		lwb = lwbuf_alloc(fs->m, &lwb_cache);
1607 		ptep = ((vpte_t *)lwbuf_kva(lwb) +
1608 		        ((*pindex >> vshift) & VPTE_PAGE_MASK));
1609 		vm_page_activate(fs->m);
1610 
1611 		/*
1612 		 * Page table write-back - entire operation including
1613 		 * validation of the pte must be atomic to avoid races
1614 		 * against the vkernel changing the pte.
1615 		 *
1616 		 * If the vpte is valid for the* requested operation, do
1617 		 * a write-back to the page table.
1618 		 *
1619 		 * XXX VPTE_M is not set properly for page directory pages.
1620 		 * It doesn't get set in the page directory if the page table
1621 		 * is modified during a read access.
1622 		 */
1623 		for (;;) {
1624 			vpte_t nvpte;
1625 
1626 			/*
1627 			 * Reload for the cmpset, but make sure the pte is
1628 			 * still valid.
1629 			 */
1630 			vpte = *ptep;
1631 			cpu_ccfence();
1632 			nvpte = vpte;
1633 
1634 			if ((vpte & VPTE_V) == 0)
1635 				break;
1636 
1637 			if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_RW))
1638 				nvpte |= VPTE_M | VPTE_A;
1639 			if (fault_type & (VM_PROT_READ | VM_PROT_EXECUTE))
1640 				nvpte |= VPTE_A;
1641 			if (vpte == nvpte)
1642 				break;
1643 			if (atomic_cmpset_long(ptep, vpte, nvpte)) {
1644 				vm_page_dirty(fs->m);
1645 				break;
1646 			}
1647 		}
1648 		lwbuf_free(lwb);
1649 		vm_page_flag_set(fs->m, PG_REFERENCED);
1650 		vm_page_wakeup(fs->m);
1651 		fs->m = NULL;
1652 		cleanup_fault(fs);
1653 	}
1654 
1655 	/*
1656 	 * When the vkernel sets VPTE_RW it expects the real kernel to
1657 	 * reflect VPTE_M back when the page is modified via the mapping.
1658 	 * In order to accomplish this the real kernel must map the page
1659 	 * read-only for read faults and use write faults to reflect VPTE_M
1660 	 * back.
1661 	 *
1662 	 * Once VPTE_M has been set, the real kernel's pte allows writing.
1663 	 * If the vkernel clears VPTE_M the vkernel must be sure to
1664 	 * MADV_INVAL the real kernel's mappings to force the real kernel
1665 	 * to re-fault on the next write so oit can set VPTE_M again.
1666 	 */
1667 	if ((fault_type & VM_PROT_WRITE) == 0 &&
1668 	    (vpte & (VPTE_RW | VPTE_M)) != (VPTE_RW | VPTE_M)) {
1669 		fs->first_prot &= ~VM_PROT_WRITE;
1670 	}
1671 
1672 	/*
1673 	 * Disable EXECUTE perms if NX bit is set.
1674 	 */
1675 	if (vpte & VPTE_NX)
1676 		fs->first_prot &= ~VM_PROT_EXECUTE;
1677 
1678 	/*
1679 	 * Combine remaining address bits with the vpte.
1680 	 */
1681 	*pindex = ((vpte & VPTE_FRAME) >> PAGE_SHIFT) +
1682 		  (*pindex & ((1L << vshift) - 1));
1683 	return (KERN_SUCCESS);
1684 }
1685 
1686 
1687 /*
1688  * This is the core of the vm_fault code.
1689  *
1690  * Do all operations required to fault-in (fs.first_ba->object, pindex).
1691  * Run through the backing store as necessary and do required COW or virtual
1692  * copy operations.  The caller has already fully resolved the vm_map_entry
1693  * and, if appropriate, has created a copy-on-write layer.  All we need to
1694  * do is iterate the object chain.
1695  *
1696  * On failure (fs) is unlocked and deallocated and the caller may return or
1697  * retry depending on the failure code.  On success (fs) is NOT unlocked or
1698  * deallocated, fs.m will contained a resolved, busied page, and fs.ba's
1699  * object will have an additional PIP count if it is not equal to
1700  * fs.first_ba.
1701  *
1702  * If locks based on fs->first_shared or fs->shared are insufficient,
1703  * clear the appropriate field(s) and return RETRY.  COWs require that
1704  * first_shared be 0, while page allocations (or frees) require that
1705  * shared be 0.  Renames require that both be 0.
1706  *
1707  * NOTE! fs->[first_]shared might be set with VM_FAULT_DIRTY also set.
1708  *	 we will have to retry with it exclusive if the vm_page is
1709  *	 PG_SWAPPED.
1710  *
1711  * fs->first_ba->object must be held on call.
1712  */
1713 static
1714 int
1715 vm_fault_object(struct faultstate *fs, vm_pindex_t first_pindex,
1716 		vm_prot_t fault_type, int allow_nofault)
1717 {
1718 	vm_map_backing_t next_ba;
1719 	vm_pindex_t pindex;
1720 	int error;
1721 
1722 	ASSERT_LWKT_TOKEN_HELD(vm_object_token(fs->first_ba->object));
1723 	fs->prot = fs->first_prot;
1724 	pindex = first_pindex;
1725 	KKASSERT(fs->ba == fs->first_ba);
1726 
1727 	vm_object_pip_add(fs->first_ba->object, 1);
1728 
1729 	/*
1730 	 * If a read fault occurs we try to upgrade the page protection
1731 	 * and make it also writable if possible.  There are three cases
1732 	 * where we cannot make the page mapping writable:
1733 	 *
1734 	 * (1) The mapping is read-only or the VM object is read-only,
1735 	 *     fs->prot above will simply not have VM_PROT_WRITE set.
1736 	 *
1737 	 * (2) If the mapping is a virtual page table fs->first_prot will
1738 	 *     have already been properly adjusted by vm_fault_vpagetable().
1739 	 *     to detect writes so we can set VPTE_M in the virtual page
1740 	 *     table.  Used by vkernels.
1741 	 *
1742 	 * (3) If the VM page is read-only or copy-on-write, upgrading would
1743 	 *     just result in an unnecessary COW fault.
1744 	 *
1745 	 * (4) If the pmap specifically requests A/M bit emulation, downgrade
1746 	 *     here.
1747 	 */
1748 #if 0
1749 	/* see vpagetable code */
1750 	if (fs->entry->maptype == VM_MAPTYPE_VPAGETABLE) {
1751 		if ((fault_type & VM_PROT_WRITE) == 0)
1752 			fs->prot &= ~VM_PROT_WRITE;
1753 	}
1754 #endif
1755 
1756 	if (curthread->td_lwp && curthread->td_lwp->lwp_vmspace &&
1757 	    pmap_emulate_ad_bits(&curthread->td_lwp->lwp_vmspace->vm_pmap)) {
1758 		if ((fault_type & VM_PROT_WRITE) == 0)
1759 			fs->prot &= ~VM_PROT_WRITE;
1760 	}
1761 
1762 	/* vm_object_hold(fs->ba->object); implied b/c ba == first_ba */
1763 
1764 	for (;;) {
1765 		/*
1766 		 * If the object is dead, we stop here
1767 		 */
1768 		if (fs->ba->object->flags & OBJ_DEAD) {
1769 			vm_object_pip_wakeup(fs->first_ba->object);
1770 			unlock_things(fs);
1771 			return (KERN_PROTECTION_FAILURE);
1772 		}
1773 
1774 		/*
1775 		 * See if the page is resident.  Wait/Retry if the page is
1776 		 * busy (lots of stuff may have changed so we can't continue
1777 		 * in that case).
1778 		 *
1779 		 * We can theoretically allow the soft-busy case on a read
1780 		 * fault if the page is marked valid, but since such
1781 		 * pages are typically already pmap'd, putting that
1782 		 * special case in might be more effort then it is
1783 		 * worth.  We cannot under any circumstances mess
1784 		 * around with a vm_page_t->busy page except, perhaps,
1785 		 * to pmap it.
1786 		 */
1787 		fs->m = vm_page_lookup_busy_try(fs->ba->object, pindex,
1788 						TRUE, &error);
1789 		if (error) {
1790 			vm_object_pip_wakeup(fs->first_ba->object);
1791 			unlock_things(fs);
1792 			vm_page_sleep_busy(fs->m, TRUE, "vmpfw");
1793 			mycpu->gd_cnt.v_intrans++;
1794 			fs->m = NULL;
1795 			return (KERN_TRY_AGAIN);
1796 		}
1797 		if (fs->m) {
1798 			/*
1799 			 * The page is busied for us.
1800 			 *
1801 			 * If reactivating a page from PQ_CACHE we may have
1802 			 * to rate-limit.
1803 			 */
1804 			int queue = fs->m->queue;
1805 			vm_page_unqueue_nowakeup(fs->m);
1806 
1807 			if ((queue - fs->m->pc) == PQ_CACHE &&
1808 			    vm_page_count_severe()) {
1809 				vm_page_activate(fs->m);
1810 				vm_page_wakeup(fs->m);
1811 				fs->m = NULL;
1812 				vm_object_pip_wakeup(fs->first_ba->object);
1813 				unlock_things(fs);
1814 				if (allow_nofault == 0 ||
1815 				    (curthread->td_flags & TDF_NOFAULT) == 0) {
1816 					thread_t td;
1817 
1818 					vm_wait_pfault();
1819 					td = curthread;
1820 					if (td->td_proc && (td->td_proc->p_flags & P_LOWMEMKILL))
1821 						return (KERN_PROTECTION_FAILURE);
1822 				}
1823 				return (KERN_TRY_AGAIN);
1824 			}
1825 
1826 			/*
1827 			 * If it still isn't completely valid (readable),
1828 			 * or if a read-ahead-mark is set on the VM page,
1829 			 * jump to readrest, else we found the page and
1830 			 * can return.
1831 			 *
1832 			 * We can release the spl once we have marked the
1833 			 * page busy.
1834 			 */
1835 			if (fs->m->object != &kernel_object) {
1836 				if ((fs->m->valid & VM_PAGE_BITS_ALL) !=
1837 				    VM_PAGE_BITS_ALL) {
1838 					goto readrest;
1839 				}
1840 				if (fs->m->flags & PG_RAM) {
1841 					if (debug_cluster)
1842 						kprintf("R");
1843 					vm_page_flag_clear(fs->m, PG_RAM);
1844 					goto readrest;
1845 				}
1846 			}
1847 			fs->first_ba->flags &= ~VM_MAP_BACK_EXCL_HEUR;
1848 			break; /* break to PAGE HAS BEEN FOUND */
1849 		}
1850 
1851 		/*
1852 		 * Page is not resident, If this is the search termination
1853 		 * or the pager might contain the page, allocate a new page.
1854 		 */
1855 		if (TRYPAGER(fs) || fs->ba == fs->first_ba) {
1856 			/*
1857 			 * If this is a SWAP object we can use the shared
1858 			 * lock to check existence of a swap block.  If
1859 			 * there isn't one we can skip to the next object.
1860 			 *
1861 			 * However, if this is the first object we allocate
1862 			 * a page now just in case we need to copy to it
1863 			 * later.
1864 			 */
1865 			if (fs->ba != fs->first_ba &&
1866 			    fs->ba->object->type == OBJT_SWAP) {
1867 				if (swap_pager_haspage_locked(fs->ba->object,
1868 							      pindex) == 0) {
1869 					goto next;
1870 				}
1871 			}
1872 
1873 			/*
1874 			 * Allocating, must be exclusive.
1875 			 */
1876 			fs->first_ba->flags |= VM_MAP_BACK_EXCL_HEUR;
1877 			if (fs->ba == fs->first_ba && fs->first_shared) {
1878 				fs->first_shared = 0;
1879 				vm_object_pip_wakeup(fs->first_ba->object);
1880 				unlock_things(fs);
1881 				return (KERN_TRY_AGAIN);
1882 			}
1883 			if (fs->ba != fs->first_ba && fs->shared) {
1884 				fs->first_shared = 0;
1885 				fs->shared = 0;
1886 				vm_object_pip_wakeup(fs->first_ba->object);
1887 				unlock_things(fs);
1888 				return (KERN_TRY_AGAIN);
1889 			}
1890 
1891 			/*
1892 			 * If the page is beyond the object size we fail
1893 			 */
1894 			if (pindex >= fs->ba->object->size) {
1895 				vm_object_pip_wakeup(fs->first_ba->object);
1896 				unlock_things(fs);
1897 				return (KERN_PROTECTION_FAILURE);
1898 			}
1899 
1900 			/*
1901 			 * Allocate a new page for this object/offset pair.
1902 			 *
1903 			 * It is possible for the allocation to race, so
1904 			 * handle the case.
1905 			 */
1906 			fs->m = NULL;
1907 			if (!vm_page_count_severe()) {
1908 				fs->m = vm_page_alloc(fs->ba->object, pindex,
1909 				    ((fs->vp || fs->ba->backing_ba) ?
1910 					VM_ALLOC_NULL_OK | VM_ALLOC_NORMAL :
1911 					VM_ALLOC_NULL_OK | VM_ALLOC_NORMAL |
1912 					VM_ALLOC_USE_GD | VM_ALLOC_ZERO));
1913 			}
1914 			if (fs->m == NULL) {
1915 				vm_object_pip_wakeup(fs->first_ba->object);
1916 				unlock_things(fs);
1917 				if (allow_nofault == 0 ||
1918 				    (curthread->td_flags & TDF_NOFAULT) == 0) {
1919 					thread_t td;
1920 
1921 					vm_wait_pfault();
1922 					td = curthread;
1923 					if (td->td_proc && (td->td_proc->p_flags & P_LOWMEMKILL))
1924 						return (KERN_PROTECTION_FAILURE);
1925 				}
1926 				return (KERN_TRY_AGAIN);
1927 			}
1928 
1929 			/*
1930 			 * Fall through to readrest.  We have a new page which
1931 			 * will have to be paged (since m->valid will be 0).
1932 			 */
1933 		}
1934 
1935 readrest:
1936 		/*
1937 		 * We have found an invalid or partially valid page, a
1938 		 * page with a read-ahead mark which might be partially or
1939 		 * fully valid (and maybe dirty too), or we have allocated
1940 		 * a new page.
1941 		 *
1942 		 * Attempt to fault-in the page if there is a chance that the
1943 		 * pager has it, and potentially fault in additional pages
1944 		 * at the same time.
1945 		 *
1946 		 * If TRYPAGER is true then fs.m will be non-NULL and busied
1947 		 * for us.
1948 		 */
1949 		if (TRYPAGER(fs)) {
1950 			u_char behavior = vm_map_entry_behavior(fs->entry);
1951 			vm_object_t object;
1952 			vm_page_t first_m;
1953 			int seqaccess;
1954 			int rv;
1955 
1956 			if (behavior == MAP_ENTRY_BEHAV_RANDOM)
1957 				seqaccess = 0;
1958 			else
1959 				seqaccess = -1;
1960 
1961 			/*
1962 			 * Doing I/O may synchronously insert additional
1963 			 * pages so we can't be shared at this point either.
1964 			 *
1965 			 * NOTE: We can't free fs->m here in the allocated
1966 			 *	 case (fs->ba != fs->first_ba) as this
1967 			 *	 would require an exclusively locked
1968 			 *	 VM object.
1969 			 */
1970 			if (fs->ba == fs->first_ba && fs->first_shared) {
1971 				vm_page_deactivate(fs->m);
1972 				vm_page_wakeup(fs->m);
1973 				fs->m = NULL;
1974 				fs->first_shared = 0;
1975 				vm_object_pip_wakeup(fs->first_ba->object);
1976 				unlock_things(fs);
1977 				return (KERN_TRY_AGAIN);
1978 			}
1979 			if (fs->ba != fs->first_ba && fs->shared) {
1980 				vm_page_deactivate(fs->m);
1981 				vm_page_wakeup(fs->m);
1982 				fs->m = NULL;
1983 				fs->first_shared = 0;
1984 				fs->shared = 0;
1985 				vm_object_pip_wakeup(fs->first_ba->object);
1986 				unlock_things(fs);
1987 				return (KERN_TRY_AGAIN);
1988 			}
1989 
1990 			object = fs->ba->object;
1991 			first_m = NULL;
1992 
1993 			/* object is held, no more access to entry or ba's */
1994 
1995 			/*
1996 			 * Acquire the page data.  We still hold object
1997 			 * and the page has been BUSY's.
1998 			 *
1999 			 * We own the page, but we must re-issue the lookup
2000 			 * because the pager may have replaced it (for example,
2001 			 * in order to enter a fictitious page into the
2002 			 * object).  In this situation the pager will have
2003 			 * cleaned up the old page and left the new one
2004 			 * busy for us.
2005 			 *
2006 			 * If we got here through a PG_RAM read-ahead
2007 			 * mark the page may be partially dirty and thus
2008 			 * not freeable.  Don't bother checking to see
2009 			 * if the pager has the page because we can't free
2010 			 * it anyway.  We have to depend on the get_page
2011 			 * operation filling in any gaps whether there is
2012 			 * backing store or not.
2013 			 *
2014 			 * We must dispose of the page (fs->m) and also
2015 			 * possibly first_m (the fronting layer).  If
2016 			 * this is a write fault leave the page intact
2017 			 * because we will probably have to copy fs->m
2018 			 * to fs->first_m on the retry.  If this is a
2019 			 * read fault we probably won't need the page.
2020 			 */
2021 			rv = vm_pager_get_page(object, &fs->m, seqaccess);
2022 
2023 			if (rv == VM_PAGER_OK) {
2024 				++fs->hardfault;
2025 				fs->m = vm_page_lookup(object, pindex);
2026 				if (fs->m) {
2027 					vm_page_activate(fs->m);
2028 					vm_page_wakeup(fs->m);
2029 					fs->m = NULL;
2030 				}
2031 
2032 				if (fs->m) {
2033 					/* have page */
2034 					break;
2035 				}
2036 				vm_object_pip_wakeup(fs->first_ba->object);
2037 				unlock_things(fs);
2038 				return (KERN_TRY_AGAIN);
2039 			}
2040 
2041 			/*
2042 			 * If the pager doesn't have the page, continue on
2043 			 * to the next object.  Retain the vm_page if this
2044 			 * is the first object, we may need to copy into
2045 			 * it later.
2046 			 */
2047 			if (rv == VM_PAGER_FAIL) {
2048 				if (fs->ba != fs->first_ba) {
2049 					vm_page_free(fs->m);
2050 					fs->m = NULL;
2051 				}
2052 				goto next;
2053 			}
2054 
2055 			/*
2056 			 * Remove the bogus page (which does not exist at this
2057 			 * object/offset).
2058 			 *
2059 			 * Also wake up any other process that may want to bring
2060 			 * in this page.
2061 			 *
2062 			 * If this is the top-level object, we must leave the
2063 			 * busy page to prevent another process from rushing
2064 			 * past us, and inserting the page in that object at
2065 			 * the same time that we are.
2066 			 */
2067 			if (rv == VM_PAGER_ERROR) {
2068 				if (curproc) {
2069 					kprintf("vm_fault: pager read error, "
2070 						"pid %d (%s)\n",
2071 						curproc->p_pid,
2072 						curproc->p_comm);
2073 				} else {
2074 					kprintf("vm_fault: pager read error, "
2075 						"thread %p (%s)\n",
2076 						curthread,
2077 						curthread->td_comm);
2078 				}
2079 			}
2080 
2081 			/*
2082 			 * I/O error or data outside pager's range.
2083 			 */
2084 			if (fs->m) {
2085 				vnode_pager_freepage(fs->m);
2086 				fs->m = NULL;
2087 			}
2088 			if (first_m) {
2089 				vm_page_free(first_m);
2090 				first_m = NULL;		/* safety */
2091 			}
2092 			vm_object_pip_wakeup(object);
2093 			unlock_things(fs);
2094 
2095 			switch(rv) {
2096 			case VM_PAGER_ERROR:
2097 				return (KERN_FAILURE);
2098 			case VM_PAGER_BAD:
2099 				return (KERN_PROTECTION_FAILURE);
2100 			default:
2101 				return (KERN_PROTECTION_FAILURE);
2102 			}
2103 
2104 #if 0
2105 			/*
2106 			 * Data outside the range of the pager or an I/O error
2107 			 *
2108 			 * The page may have been wired during the pagein,
2109 			 * e.g. by the buffer cache, and cannot simply be
2110 			 * freed.  Call vnode_pager_freepage() to deal with it.
2111 			 *
2112 			 * The object is not held shared so we can safely
2113 			 * free the page.
2114 			 */
2115 			if (fs->ba != fs->first_ba) {
2116 
2117 				/*
2118 				 * XXX - we cannot just fall out at this
2119 				 * point, m has been freed and is invalid!
2120 				 */
2121 			}
2122 
2123 			/*
2124 			 * XXX - the check for kernel_map is a kludge to work
2125 			 * around having the machine panic on a kernel space
2126 			 * fault w/ I/O error.
2127 			 */
2128 			if (((fs->map != &kernel_map) &&
2129 			    (rv == VM_PAGER_ERROR)) || (rv == VM_PAGER_BAD)) {
2130 				if (fs->m) {
2131 					/* from just above */
2132 					KKASSERT(fs->first_shared == 0);
2133 					vnode_pager_freepage(fs->m);
2134 					fs->m = NULL;
2135 				}
2136 				/* NOT REACHED */
2137 			}
2138 #endif
2139 		}
2140 
2141 next:
2142 		/*
2143 		 * We get here if the object has a default pager (or unwiring)
2144 		 * or the pager doesn't have the page.
2145 		 *
2146 		 * fs->first_m will be used for the COW unless we find a
2147 		 * deeper page to be mapped read-only, in which case the
2148 		 * unlock*(fs) will free first_m.
2149 		 */
2150 		if (fs->ba == fs->first_ba)
2151 			fs->first_m = fs->m;
2152 
2153 		/*
2154 		 * Move on to the next object.  The chain lock should prevent
2155 		 * the backing_object from getting ripped out from under us.
2156 		 *
2157 		 * The object lock for the next object is governed by
2158 		 * fs->shared.
2159 		 */
2160 		if ((next_ba = fs->ba->backing_ba) != NULL) {
2161 			if (fs->shared)
2162 				vm_object_hold_shared(next_ba->object);
2163 			else
2164 				vm_object_hold(next_ba->object);
2165 			KKASSERT(next_ba == fs->ba->backing_ba);
2166 			pindex += OFF_TO_IDX(next_ba->offset);
2167 		}
2168 
2169 		if (next_ba == NULL) {
2170 			/*
2171 			 * If there's no object left, fill the page in the top
2172 			 * object with zeros.
2173 			 */
2174 			if (fs->ba != fs->first_ba) {
2175 				vm_object_pip_wakeup(fs->ba->object);
2176 				vm_object_drop(fs->ba->object);
2177 				fs->ba = fs->first_ba;
2178 				pindex = first_pindex;
2179 				fs->m = fs->first_m;
2180 			}
2181 			fs->first_m = NULL;
2182 
2183 			/*
2184 			 * Zero the page and mark it valid.
2185 			 */
2186 			vm_page_zero_fill(fs->m);
2187 			mycpu->gd_cnt.v_zfod++;
2188 			fs->m->valid = VM_PAGE_BITS_ALL;
2189 			break;	/* break to PAGE HAS BEEN FOUND */
2190 		}
2191 		if (fs->ba != fs->first_ba) {
2192 			vm_object_pip_wakeup(fs->ba->object);
2193 			vm_object_lock_swap();	/* flip ba/next_ba */
2194 			vm_object_drop(fs->ba->object);
2195 		}
2196 		fs->ba = next_ba;
2197 		vm_object_pip_add(next_ba->object, 1);
2198 	}
2199 
2200 	/*
2201 	 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
2202 	 * is held.]
2203 	 *
2204 	 * object still held.
2205 	 * vm_map may not be locked (determined by fs->lookup_still_valid)
2206 	 *
2207 	 * local shared variable may be different from fs->shared.
2208 	 *
2209 	 * If the page is being written, but isn't already owned by the
2210 	 * top-level object, we have to copy it into a new page owned by the
2211 	 * top-level object.
2212 	 */
2213 	KASSERT((fs->m->busy_count & PBUSY_LOCKED) != 0,
2214 		("vm_fault: not busy after main loop"));
2215 
2216 	if (fs->ba != fs->first_ba) {
2217 		/*
2218 		 * We only really need to copy if we want to write it.
2219 		 */
2220 		if (fault_type & VM_PROT_WRITE) {
2221 #if 0
2222 			/* CODE REFACTOR IN PROGRESS, REMOVE OPTIMIZATION */
2223 			/*
2224 			 * This allows pages to be virtually copied from a
2225 			 * backing_object into the first_object, where the
2226 			 * backing object has no other refs to it, and cannot
2227 			 * gain any more refs.  Instead of a bcopy, we just
2228 			 * move the page from the backing object to the
2229 			 * first object.  Note that we must mark the page
2230 			 * dirty in the first object so that it will go out
2231 			 * to swap when needed.
2232 			 */
2233 			if (virtual_copy_ok(fs)) {
2234 				/*
2235 				 * (first_m) and (m) are both busied.  We have
2236 				 * move (m) into (first_m)'s object/pindex
2237 				 * in an atomic fashion, then free (first_m).
2238 				 *
2239 				 * first_object is held so second remove
2240 				 * followed by the rename should wind
2241 				 * up being atomic.  vm_page_free() might
2242 				 * block so we don't do it until after the
2243 				 * rename.
2244 				 */
2245 				vm_page_protect(fs->first_m, VM_PROT_NONE);
2246 				vm_page_remove(fs->first_m);
2247 				vm_page_rename(fs->m,
2248 					       fs->first_ba->object,
2249 					       first_pindex);
2250 				vm_page_free(fs->first_m);
2251 				fs->first_m = fs->m;
2252 				fs->m = NULL;
2253 				mycpu->gd_cnt.v_cow_optim++;
2254 			} else
2255 #endif
2256 			{
2257 				/*
2258 				 * Oh, well, lets copy it.
2259 				 *
2260 				 * Why are we unmapping the original page
2261 				 * here?  Well, in short, not all accessors
2262 				 * of user memory go through the pmap.  The
2263 				 * procfs code doesn't have access user memory
2264 				 * via a local pmap, so vm_fault_page*()
2265 				 * can't call pmap_enter().  And the umtx*()
2266 				 * code may modify the COW'd page via a DMAP
2267 				 * or kernel mapping and not via the pmap,
2268 				 * leaving the original page still mapped
2269 				 * read-only into the pmap.
2270 				 *
2271 				 * So we have to remove the page from at
2272 				 * least the current pmap if it is in it.
2273 				 *
2274 				 * We used to just remove it from all pmaps
2275 				 * but that creates inefficiencies on SMP,
2276 				 * particularly for COW program & library
2277 				 * mappings that are concurrently exec'd.
2278 				 * Only remove the page from the current
2279 				 * pmap.
2280 				 */
2281 				KKASSERT(fs->first_shared == 0);
2282 				vm_page_copy(fs->m, fs->first_m);
2283 				/*vm_page_protect(fs->m, VM_PROT_NONE);*/
2284 				pmap_remove_specific(
2285 				    &curthread->td_lwp->lwp_vmspace->vm_pmap,
2286 				    fs->m);
2287 			}
2288 
2289 			/*
2290 			 * We no longer need the old page or object.
2291 			 */
2292 			if (fs->m)
2293 				release_page(fs);
2294 
2295 			/*
2296 			 * fs->ba != fs->first_ba due to above conditional
2297 			 */
2298 			vm_object_pip_wakeup(fs->ba->object);
2299 			vm_object_drop(fs->ba->object);
2300 			fs->ba = fs->first_ba;
2301 
2302 			/*
2303 			 * Only use the new page below...
2304 			 */
2305 			mycpu->gd_cnt.v_cow_faults++;
2306 			fs->m = fs->first_m;
2307 			pindex = first_pindex;
2308 		} else {
2309 			/*
2310 			 * If it wasn't a write fault avoid having to copy
2311 			 * the page by mapping it read-only from backing
2312 			 * store.  The process is not allowed to modify
2313 			 * backing pages.
2314 			 */
2315 			fs->prot &= ~VM_PROT_WRITE;
2316 		}
2317 	}
2318 
2319 	/*
2320 	 * Relock the map if necessary, then check the generation count.
2321 	 * relock_map() will update fs->timestamp to account for the
2322 	 * relocking if necessary.
2323 	 *
2324 	 * If the count has changed after relocking then all sorts of
2325 	 * crap may have happened and we have to retry.
2326 	 *
2327 	 * NOTE: The relock_map() can fail due to a deadlock against
2328 	 *	 the vm_page we are holding BUSY.
2329 	 */
2330 	KKASSERT(fs->lookup_still_valid != 0);
2331 #if 0
2332 	if (fs->lookup_still_valid == 0 && fs->map) {
2333 		if (relock_map(fs) ||
2334 		    fs->map->timestamp != fs->map_generation) {
2335 			release_page(fs);
2336 			vm_object_pip_wakeup(fs->first_ba->object);
2337 			unlock_things(fs);
2338 			return (KERN_TRY_AGAIN);
2339 		}
2340 	}
2341 #endif
2342 
2343 	/*
2344 	 * If the fault is a write, we know that this page is being
2345 	 * written NOW so dirty it explicitly to save on pmap_is_modified()
2346 	 * calls later.
2347 	 *
2348 	 * If this is a NOSYNC mmap we do not want to set PG_NOSYNC
2349 	 * if the page is already dirty to prevent data written with
2350 	 * the expectation of being synced from not being synced.
2351 	 * Likewise if this entry does not request NOSYNC then make
2352 	 * sure the page isn't marked NOSYNC.  Applications sharing
2353 	 * data should use the same flags to avoid ping ponging.
2354 	 *
2355 	 * Also tell the backing pager, if any, that it should remove
2356 	 * any swap backing since the page is now dirty.
2357 	 */
2358 	vm_page_activate(fs->m);
2359 	if (fs->prot & VM_PROT_WRITE) {
2360 		vm_object_set_writeable_dirty(fs->m->object);
2361 		vm_set_nosync(fs->m, fs->entry);
2362 		if (fs->fault_flags & VM_FAULT_DIRTY) {
2363 			vm_page_dirty(fs->m);
2364 			if (fs->m->flags & PG_SWAPPED) {
2365 				/*
2366 				 * If the page is swapped out we have to call
2367 				 * swap_pager_unswapped() which requires an
2368 				 * exclusive object lock.  If we are shared,
2369 				 * we must clear the shared flag and retry.
2370 				 */
2371 				if ((fs->ba == fs->first_ba &&
2372 				     fs->first_shared) ||
2373 				    (fs->ba != fs->first_ba && fs->shared)) {
2374 					vm_page_wakeup(fs->m);
2375 					fs->m = NULL;
2376 					if (fs->ba == fs->first_ba)
2377 						fs->first_shared = 0;
2378 					else
2379 						fs->shared = 0;
2380 					vm_object_pip_wakeup(
2381 							fs->first_ba->object);
2382 					unlock_things(fs);
2383 					return (KERN_TRY_AGAIN);
2384 				}
2385 				swap_pager_unswapped(fs->m);
2386 			}
2387 		}
2388 	}
2389 
2390 	/*
2391 	 * We found our page at backing layer ba.  Leave the layer state
2392 	 * intact.
2393 	 */
2394 
2395 	vm_object_pip_wakeup(fs->first_ba->object);
2396 #if 0
2397 	if (fs->ba != fs->first_ba)
2398 		vm_object_drop(fs->ba->object);
2399 #endif
2400 
2401 	/*
2402 	 * Page had better still be busy.  We are still locked up and
2403 	 * fs->ba->object will have another PIP reference for the case
2404 	 * where fs->ba != fs->first_ba.
2405 	 */
2406 	KASSERT(fs->m->busy_count & PBUSY_LOCKED,
2407 		("vm_fault: page %p not busy!", fs->m));
2408 
2409 	/*
2410 	 * Sanity check: page must be completely valid or it is not fit to
2411 	 * map into user space.  vm_pager_get_pages() ensures this.
2412 	 */
2413 	if (fs->m->valid != VM_PAGE_BITS_ALL) {
2414 		vm_page_zero_invalid(fs->m, TRUE);
2415 		kprintf("Warning: page %p partially invalid on fault\n", fs->m);
2416 	}
2417 
2418 	return (KERN_SUCCESS);
2419 }
2420 
2421 /*
2422  * Wire down a range of virtual addresses in a map.  The entry in question
2423  * should be marked in-transition and the map must be locked.  We must
2424  * release the map temporarily while faulting-in the page to avoid a
2425  * deadlock.  Note that the entry may be clipped while we are blocked but
2426  * will never be freed.
2427  *
2428  * map must be locked on entry.
2429  */
2430 int
2431 vm_fault_wire(vm_map_t map, vm_map_entry_t entry,
2432 	      boolean_t user_wire, int kmflags)
2433 {
2434 	boolean_t fictitious;
2435 	vm_offset_t start;
2436 	vm_offset_t end;
2437 	vm_offset_t va;
2438 	pmap_t pmap;
2439 	int rv;
2440 	int wire_prot;
2441 	int fault_flags;
2442 	vm_page_t m;
2443 
2444 	if (user_wire) {
2445 		wire_prot = VM_PROT_READ;
2446 		fault_flags = VM_FAULT_USER_WIRE;
2447 	} else {
2448 		wire_prot = VM_PROT_READ | VM_PROT_WRITE;
2449 		fault_flags = VM_FAULT_CHANGE_WIRING;
2450 	}
2451 	if (kmflags & KM_NOTLBSYNC)
2452 		wire_prot |= VM_PROT_NOSYNC;
2453 
2454 	pmap = vm_map_pmap(map);
2455 	start = entry->start;
2456 	end = entry->end;
2457 
2458 	switch(entry->maptype) {
2459 	case VM_MAPTYPE_NORMAL:
2460 	case VM_MAPTYPE_VPAGETABLE:
2461 		fictitious = entry->ba.object &&
2462 			    ((entry->ba.object->type == OBJT_DEVICE) ||
2463 			     (entry->ba.object->type == OBJT_MGTDEVICE));
2464 		break;
2465 	case VM_MAPTYPE_UKSMAP:
2466 		fictitious = TRUE;
2467 		break;
2468 	default:
2469 		fictitious = FALSE;
2470 		break;
2471 	}
2472 
2473 	if (entry->eflags & MAP_ENTRY_KSTACK)
2474 		start += PAGE_SIZE;
2475 	map->timestamp++;
2476 	vm_map_unlock(map);
2477 
2478 	/*
2479 	 * We simulate a fault to get the page and enter it in the physical
2480 	 * map.
2481 	 */
2482 	for (va = start; va < end; va += PAGE_SIZE) {
2483 		rv = vm_fault(map, va, wire_prot, fault_flags);
2484 		if (rv) {
2485 			while (va > start) {
2486 				va -= PAGE_SIZE;
2487 				m = pmap_unwire(pmap, va);
2488 				if (m && !fictitious) {
2489 					vm_page_busy_wait(m, FALSE, "vmwrpg");
2490 					vm_page_unwire(m, 1);
2491 					vm_page_wakeup(m);
2492 				}
2493 			}
2494 			goto done;
2495 		}
2496 	}
2497 	rv = KERN_SUCCESS;
2498 done:
2499 	vm_map_lock(map);
2500 
2501 	return (rv);
2502 }
2503 
2504 /*
2505  * Unwire a range of virtual addresses in a map.  The map should be
2506  * locked.
2507  */
2508 void
2509 vm_fault_unwire(vm_map_t map, vm_map_entry_t entry)
2510 {
2511 	boolean_t fictitious;
2512 	vm_offset_t start;
2513 	vm_offset_t end;
2514 	vm_offset_t va;
2515 	pmap_t pmap;
2516 	vm_page_t m;
2517 
2518 	pmap = vm_map_pmap(map);
2519 	start = entry->start;
2520 	end = entry->end;
2521 	fictitious = entry->ba.object &&
2522 			((entry->ba.object->type == OBJT_DEVICE) ||
2523 			 (entry->ba.object->type == OBJT_MGTDEVICE));
2524 	if (entry->eflags & MAP_ENTRY_KSTACK)
2525 		start += PAGE_SIZE;
2526 
2527 	/*
2528 	 * Since the pages are wired down, we must be able to get their
2529 	 * mappings from the physical map system.
2530 	 */
2531 	for (va = start; va < end; va += PAGE_SIZE) {
2532 		m = pmap_unwire(pmap, va);
2533 		if (m && !fictitious) {
2534 			vm_page_busy_wait(m, FALSE, "vmwrpg");
2535 			vm_page_unwire(m, 1);
2536 			vm_page_wakeup(m);
2537 		}
2538 	}
2539 }
2540 
2541 /*
2542  * Simulate write faults to bring all data into the head object, return
2543  * KERN_SUCCESS on success (which should be always unless the system runs
2544  * out of memory).
2545  *
2546  * The caller will handle destroying the backing_ba's.
2547  */
2548 int
2549 vm_fault_collapse(vm_map_t map, vm_map_entry_t entry)
2550 {
2551 	struct faultstate fs;
2552 	vm_ooffset_t scan;
2553 	vm_pindex_t pindex;
2554 	vm_object_t object;
2555 	int rv;
2556 	int all_shadowed;
2557 
2558 	bzero(&fs, sizeof(fs));
2559 	object = entry->ba.object;
2560 
2561 	fs.first_prot = entry->max_protection |	/* optional VM_PROT_EXECUTE */
2562 			VM_PROT_READ | VM_PROT_WRITE | VM_PROT_OVERRIDE_WRITE;
2563 	fs.fault_flags = VM_FAULT_NORMAL;
2564 	fs.map = map;
2565 	fs.entry = entry;
2566 	fs.lookup_still_valid = -1;	/* leave map atomically locked */
2567 	fs.first_ba = &entry->ba;
2568 	fs.first_ba_held = -1;		/* leave object held */
2569 
2570 	/* fs.hardfault */
2571 
2572 	vm_object_hold(object);
2573 	rv = KERN_SUCCESS;
2574 
2575 	scan = entry->start;
2576 	all_shadowed = 1;
2577 
2578 	while (scan < entry->end) {
2579 		pindex = OFF_TO_IDX(entry->ba.offset + (scan - entry->start));
2580 
2581 		if (vm_page_lookup(object, pindex)) {
2582 			scan += PAGE_SIZE;
2583 			continue;
2584 		}
2585 
2586 		all_shadowed = 0;
2587 		fs.ba = fs.first_ba;
2588 		fs.prot = fs.first_prot;
2589 
2590 		rv = vm_fault_object(&fs, pindex, fs.first_prot, 1);
2591 		if (rv == KERN_TRY_AGAIN)
2592 			continue;
2593 		if (rv != KERN_SUCCESS)
2594 			break;
2595 		vm_page_flag_set(fs.m, PG_REFERENCED);
2596 		vm_page_activate(fs.m);
2597 		vm_page_wakeup(fs.m);
2598 		scan += PAGE_SIZE;
2599 	}
2600 	KKASSERT(entry->ba.object == object);
2601 	vm_object_drop(object);
2602 
2603 	/*
2604 	 * If the fronting object did not have every page we have to clear
2605 	 * the pmap range due to the pages being changed so we can fault-in
2606 	 * the proper pages.
2607 	 */
2608 	if (all_shadowed == 0)
2609 		pmap_remove(map->pmap, entry->start, entry->end);
2610 
2611 	return rv;
2612 }
2613 
2614 /*
2615  * Copy all of the pages from one map entry to another.  If the source
2616  * is wired down we just use vm_page_lookup().  If not we use
2617  * vm_fault_object().
2618  *
2619  * The source and destination maps must be locked for write.
2620  * The source and destination maps token must be held
2621  *
2622  * No other requirements.
2623  *
2624  * XXX do segment optimization
2625  */
2626 void
2627 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
2628 		    vm_map_entry_t dst_entry, vm_map_entry_t src_entry)
2629 {
2630 	vm_object_t dst_object;
2631 	vm_object_t src_object;
2632 	vm_ooffset_t dst_offset;
2633 	vm_ooffset_t src_offset;
2634 	vm_prot_t prot;
2635 	vm_offset_t vaddr;
2636 	vm_page_t dst_m;
2637 	vm_page_t src_m;
2638 
2639 	src_object = src_entry->ba.object;
2640 	src_offset = src_entry->ba.offset;
2641 
2642 	/*
2643 	 * Create the top-level object for the destination entry. (Doesn't
2644 	 * actually shadow anything - we copy the pages directly.)
2645 	 */
2646 	vm_map_entry_allocate_object(dst_entry);
2647 	dst_object = dst_entry->ba.object;
2648 
2649 	prot = dst_entry->max_protection;
2650 
2651 	/*
2652 	 * Loop through all of the pages in the entry's range, copying each
2653 	 * one from the source object (it should be there) to the destination
2654 	 * object.
2655 	 */
2656 	vm_object_hold(src_object);
2657 	vm_object_hold(dst_object);
2658 
2659 	for (vaddr = dst_entry->start, dst_offset = 0;
2660 	     vaddr < dst_entry->end;
2661 	     vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) {
2662 
2663 		/*
2664 		 * Allocate a page in the destination object
2665 		 */
2666 		do {
2667 			dst_m = vm_page_alloc(dst_object,
2668 					      OFF_TO_IDX(dst_offset),
2669 					      VM_ALLOC_NORMAL);
2670 			if (dst_m == NULL) {
2671 				vm_wait(0);
2672 			}
2673 		} while (dst_m == NULL);
2674 
2675 		/*
2676 		 * Find the page in the source object, and copy it in.
2677 		 * (Because the source is wired down, the page will be in
2678 		 * memory.)
2679 		 */
2680 		src_m = vm_page_lookup(src_object,
2681 				       OFF_TO_IDX(dst_offset + src_offset));
2682 		if (src_m == NULL)
2683 			panic("vm_fault_copy_wired: page missing");
2684 
2685 		vm_page_copy(src_m, dst_m);
2686 
2687 		/*
2688 		 * Enter it in the pmap...
2689 		 */
2690 		pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE, dst_entry);
2691 
2692 		/*
2693 		 * Mark it no longer busy, and put it on the active list.
2694 		 */
2695 		vm_page_activate(dst_m);
2696 		vm_page_wakeup(dst_m);
2697 	}
2698 	vm_object_drop(dst_object);
2699 	vm_object_drop(src_object);
2700 }
2701 
2702 #if 0
2703 
2704 /*
2705  * This routine checks around the requested page for other pages that
2706  * might be able to be faulted in.  This routine brackets the viable
2707  * pages for the pages to be paged in.
2708  *
2709  * Inputs:
2710  *	m, rbehind, rahead
2711  *
2712  * Outputs:
2713  *  marray (array of vm_page_t), reqpage (index of requested page)
2714  *
2715  * Return value:
2716  *  number of pages in marray
2717  */
2718 static int
2719 vm_fault_additional_pages(vm_page_t m, int rbehind, int rahead,
2720 			  vm_page_t *marray, int *reqpage)
2721 {
2722 	int i,j;
2723 	vm_object_t object;
2724 	vm_pindex_t pindex, startpindex, endpindex, tpindex;
2725 	vm_page_t rtm;
2726 	int cbehind, cahead;
2727 
2728 	object = m->object;
2729 	pindex = m->pindex;
2730 
2731 	/*
2732 	 * we don't fault-ahead for device pager
2733 	 */
2734 	if ((object->type == OBJT_DEVICE) ||
2735 	    (object->type == OBJT_MGTDEVICE)) {
2736 		*reqpage = 0;
2737 		marray[0] = m;
2738 		return 1;
2739 	}
2740 
2741 	/*
2742 	 * if the requested page is not available, then give up now
2743 	 */
2744 	if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
2745 		*reqpage = 0;	/* not used by caller, fix compiler warn */
2746 		return 0;
2747 	}
2748 
2749 	if ((cbehind == 0) && (cahead == 0)) {
2750 		*reqpage = 0;
2751 		marray[0] = m;
2752 		return 1;
2753 	}
2754 
2755 	if (rahead > cahead) {
2756 		rahead = cahead;
2757 	}
2758 
2759 	if (rbehind > cbehind) {
2760 		rbehind = cbehind;
2761 	}
2762 
2763 	/*
2764 	 * Do not do any readahead if we have insufficient free memory.
2765 	 *
2766 	 * XXX code was broken disabled before and has instability
2767 	 * with this conditonal fixed, so shortcut for now.
2768 	 */
2769 	if (burst_fault == 0 || vm_page_count_severe()) {
2770 		marray[0] = m;
2771 		*reqpage = 0;
2772 		return 1;
2773 	}
2774 
2775 	/*
2776 	 * scan backward for the read behind pages -- in memory
2777 	 *
2778 	 * Assume that if the page is not found an interrupt will not
2779 	 * create it.  Theoretically interrupts can only remove (busy)
2780 	 * pages, not create new associations.
2781 	 */
2782 	if (pindex > 0) {
2783 		if (rbehind > pindex) {
2784 			rbehind = pindex;
2785 			startpindex = 0;
2786 		} else {
2787 			startpindex = pindex - rbehind;
2788 		}
2789 
2790 		vm_object_hold(object);
2791 		for (tpindex = pindex; tpindex > startpindex; --tpindex) {
2792 			if (vm_page_lookup(object, tpindex - 1))
2793 				break;
2794 		}
2795 
2796 		i = 0;
2797 		while (tpindex < pindex) {
2798 			rtm = vm_page_alloc(object, tpindex, VM_ALLOC_SYSTEM |
2799 							     VM_ALLOC_NULL_OK);
2800 			if (rtm == NULL) {
2801 				for (j = 0; j < i; j++) {
2802 					vm_page_free(marray[j]);
2803 				}
2804 				vm_object_drop(object);
2805 				marray[0] = m;
2806 				*reqpage = 0;
2807 				return 1;
2808 			}
2809 			marray[i] = rtm;
2810 			++i;
2811 			++tpindex;
2812 		}
2813 		vm_object_drop(object);
2814 	} else {
2815 		i = 0;
2816 	}
2817 
2818 	/*
2819 	 * Assign requested page
2820 	 */
2821 	marray[i] = m;
2822 	*reqpage = i;
2823 	++i;
2824 
2825 	/*
2826 	 * Scan forwards for read-ahead pages
2827 	 */
2828 	tpindex = pindex + 1;
2829 	endpindex = tpindex + rahead;
2830 	if (endpindex > object->size)
2831 		endpindex = object->size;
2832 
2833 	vm_object_hold(object);
2834 	while (tpindex < endpindex) {
2835 		if (vm_page_lookup(object, tpindex))
2836 			break;
2837 		rtm = vm_page_alloc(object, tpindex, VM_ALLOC_SYSTEM |
2838 						     VM_ALLOC_NULL_OK);
2839 		if (rtm == NULL)
2840 			break;
2841 		marray[i] = rtm;
2842 		++i;
2843 		++tpindex;
2844 	}
2845 	vm_object_drop(object);
2846 
2847 	return (i);
2848 }
2849 
2850 #endif
2851 
2852 /*
2853  * vm_prefault() provides a quick way of clustering pagefaults into a
2854  * processes address space.  It is a "cousin" of pmap_object_init_pt,
2855  * except it runs at page fault time instead of mmap time.
2856  *
2857  * vm.fast_fault	Enables pre-faulting zero-fill pages
2858  *
2859  * vm.prefault_pages	Number of pages (1/2 negative, 1/2 positive) to
2860  *			prefault.  Scan stops in either direction when
2861  *			a page is found to already exist.
2862  *
2863  * This code used to be per-platform pmap_prefault().  It is now
2864  * machine-independent and enhanced to also pre-fault zero-fill pages
2865  * (see vm.fast_fault) as well as make them writable, which greatly
2866  * reduces the number of page faults programs incur.
2867  *
2868  * Application performance when pre-faulting zero-fill pages is heavily
2869  * dependent on the application.  Very tiny applications like /bin/echo
2870  * lose a little performance while applications of any appreciable size
2871  * gain performance.  Prefaulting multiple pages also reduces SMP
2872  * congestion and can improve SMP performance significantly.
2873  *
2874  * NOTE!  prot may allow writing but this only applies to the top level
2875  *	  object.  If we wind up mapping a page extracted from a backing
2876  *	  object we have to make sure it is read-only.
2877  *
2878  * NOTE!  The caller has already handled any COW operations on the
2879  *	  vm_map_entry via the normal fault code.  Do NOT call this
2880  *	  shortcut unless the normal fault code has run on this entry.
2881  *
2882  * The related map must be locked.
2883  * No other requirements.
2884  */
2885 __read_mostly static int vm_prefault_pages = 8;
2886 SYSCTL_INT(_vm, OID_AUTO, prefault_pages, CTLFLAG_RW, &vm_prefault_pages, 0,
2887 	   "Maximum number of pages to pre-fault");
2888 __read_mostly static int vm_fast_fault = 1;
2889 SYSCTL_INT(_vm, OID_AUTO, fast_fault, CTLFLAG_RW, &vm_fast_fault, 0,
2890 	   "Burst fault zero-fill regions");
2891 
2892 /*
2893  * Set PG_NOSYNC if the map entry indicates so, but only if the page
2894  * is not already dirty by other means.  This will prevent passive
2895  * filesystem syncing as well as 'sync' from writing out the page.
2896  */
2897 static void
2898 vm_set_nosync(vm_page_t m, vm_map_entry_t entry)
2899 {
2900 	if (entry->eflags & MAP_ENTRY_NOSYNC) {
2901 		if (m->dirty == 0)
2902 			vm_page_flag_set(m, PG_NOSYNC);
2903 	} else {
2904 		vm_page_flag_clear(m, PG_NOSYNC);
2905 	}
2906 }
2907 
2908 static void
2909 vm_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry, int prot,
2910 	    int fault_flags)
2911 {
2912 	vm_map_backing_t ba;	/* first ba */
2913 	struct lwp *lp;
2914 	vm_page_t m;
2915 	vm_offset_t addr;
2916 	vm_pindex_t index;
2917 	vm_pindex_t pindex;
2918 	vm_object_t object;
2919 	int pprot;
2920 	int i;
2921 	int noneg;
2922 	int nopos;
2923 	int maxpages;
2924 
2925 	/*
2926 	 * Get stable max count value, disabled if set to 0
2927 	 */
2928 	maxpages = vm_prefault_pages;
2929 	cpu_ccfence();
2930 	if (maxpages <= 0)
2931 		return;
2932 
2933 	/*
2934 	 * We do not currently prefault mappings that use virtual page
2935 	 * tables.  We do not prefault foreign pmaps.
2936 	 */
2937 	if (entry->maptype != VM_MAPTYPE_NORMAL)
2938 		return;
2939 	lp = curthread->td_lwp;
2940 	if (lp == NULL || (pmap != vmspace_pmap(lp->lwp_vmspace)))
2941 		return;
2942 
2943 	/*
2944 	 * Limit pre-fault count to 1024 pages.
2945 	 */
2946 	if (maxpages > 1024)
2947 		maxpages = 1024;
2948 
2949 	ba = &entry->ba;
2950 	object = entry->ba.object;
2951 	KKASSERT(object != NULL);
2952 
2953 	/*
2954 	 * NOTE: VM_FAULT_DIRTY allowed later so must hold object exclusively
2955 	 *	 now (or do something more complex XXX).
2956 	 */
2957 	vm_object_hold(object);
2958 
2959 	noneg = 0;
2960 	nopos = 0;
2961 	for (i = 0; i < maxpages; ++i) {
2962 		vm_object_t lobject;
2963 		vm_object_t nobject;
2964 		vm_map_backing_t last_ba;	/* last ba */
2965 		vm_map_backing_t next_ba;	/* last ba */
2966 		int allocated = 0;
2967 		int error;
2968 
2969 		/*
2970 		 * This can eat a lot of time on a heavily contended
2971 		 * machine so yield on the tick if needed.
2972 		 */
2973 		if ((i & 7) == 7)
2974 			lwkt_yield();
2975 
2976 		/*
2977 		 * Calculate the page to pre-fault, stopping the scan in
2978 		 * each direction separately if the limit is reached.
2979 		 */
2980 		if (i & 1) {
2981 			if (noneg)
2982 				continue;
2983 			addr = addra - ((i + 1) >> 1) * PAGE_SIZE;
2984 		} else {
2985 			if (nopos)
2986 				continue;
2987 			addr = addra + ((i + 2) >> 1) * PAGE_SIZE;
2988 		}
2989 		if (addr < entry->start) {
2990 			noneg = 1;
2991 			if (noneg && nopos)
2992 				break;
2993 			continue;
2994 		}
2995 		if (addr >= entry->end) {
2996 			nopos = 1;
2997 			if (noneg && nopos)
2998 				break;
2999 			continue;
3000 		}
3001 
3002 		/*
3003 		 * Skip pages already mapped, and stop scanning in that
3004 		 * direction.  When the scan terminates in both directions
3005 		 * we are done.
3006 		 */
3007 		if (pmap_prefault_ok(pmap, addr) == 0) {
3008 			if (i & 1)
3009 				noneg = 1;
3010 			else
3011 				nopos = 1;
3012 			if (noneg && nopos)
3013 				break;
3014 			continue;
3015 		}
3016 
3017 		/*
3018 		 * Follow the backing layers to obtain the page to be mapped
3019 		 * into the pmap.
3020 		 *
3021 		 * If we reach the terminal object without finding a page
3022 		 * and we determine it would be advantageous, then allocate
3023 		 * a zero-fill page for the base object.  The base object
3024 		 * is guaranteed to be OBJT_DEFAULT for this case.
3025 		 *
3026 		 * In order to not have to check the pager via *haspage*()
3027 		 * we stop if any non-default object is encountered.  e.g.
3028 		 * a vnode or swap object would stop the loop.
3029 		 */
3030 		index = ((addr - entry->start) + entry->ba.offset) >>
3031 			PAGE_SHIFT;
3032 		last_ba = ba;
3033 		lobject = object;
3034 		pindex = index;
3035 		pprot = prot;
3036 
3037 		/*vm_object_hold(lobject); implied */
3038 
3039 		while ((m = vm_page_lookup_busy_try(lobject, pindex,
3040 						    TRUE, &error)) == NULL) {
3041 			if (lobject->type != OBJT_DEFAULT)
3042 				break;
3043 			if ((next_ba = last_ba->backing_ba) == NULL) {
3044 				if (vm_fast_fault == 0)
3045 					break;
3046 				if ((prot & VM_PROT_WRITE) == 0 ||
3047 				    vm_page_count_min(0)) {
3048 					break;
3049 				}
3050 
3051 				/*
3052 				 * NOTE: Allocated from base object
3053 				 */
3054 				m = vm_page_alloc(object, index,
3055 						  VM_ALLOC_NORMAL |
3056 						  VM_ALLOC_ZERO |
3057 						  VM_ALLOC_USE_GD |
3058 						  VM_ALLOC_NULL_OK);
3059 				if (m == NULL)
3060 					break;
3061 				allocated = 1;
3062 				pprot = prot;
3063 				/* lobject = object .. not needed */
3064 				break;
3065 			}
3066 			if (next_ba->offset & PAGE_MASK)
3067 				break;
3068 			nobject = next_ba->object;
3069 			vm_object_hold(nobject);
3070 			pindex += next_ba->offset >> PAGE_SHIFT;
3071 			if (last_ba != ba) {
3072 				vm_object_lock_swap();
3073 				vm_object_drop(lobject);
3074 			}
3075 			lobject = nobject;
3076 			last_ba = next_ba;
3077 			pprot &= ~VM_PROT_WRITE;
3078 		}
3079 
3080 		/*
3081 		 * NOTE: A non-NULL (m) will be associated with lobject if
3082 		 *	 it was found there, otherwise it is probably a
3083 		 *	 zero-fill page associated with the base object.
3084 		 *
3085 		 * Give-up if no page is available.
3086 		 */
3087 		if (m == NULL) {
3088 			if (last_ba != ba)
3089 				vm_object_drop(lobject);
3090 			break;
3091 		}
3092 
3093 		/*
3094 		 * The object must be marked dirty if we are mapping a
3095 		 * writable page.  m->object is either lobject or object,
3096 		 * both of which are still held.  Do this before we
3097 		 * potentially drop the object.
3098 		 */
3099 		if (pprot & VM_PROT_WRITE)
3100 			vm_object_set_writeable_dirty(m->object);
3101 
3102 		/*
3103 		 * Do not conditionalize on PG_RAM.  If pages are present in
3104 		 * the VM system we assume optimal caching.  If caching is
3105 		 * not optimal the I/O gravy train will be restarted when we
3106 		 * hit an unavailable page.  We do not want to try to restart
3107 		 * the gravy train now because we really don't know how much
3108 		 * of the object has been cached.  The cost for restarting
3109 		 * the gravy train should be low (since accesses will likely
3110 		 * be I/O bound anyway).
3111 		 */
3112 		if (last_ba != ba)
3113 			vm_object_drop(lobject);
3114 
3115 		/*
3116 		 * Enter the page into the pmap if appropriate.  If we had
3117 		 * allocated the page we have to place it on a queue.  If not
3118 		 * we just have to make sure it isn't on the cache queue
3119 		 * (pages on the cache queue are not allowed to be mapped).
3120 		 */
3121 		if (allocated) {
3122 			/*
3123 			 * Page must be zerod.
3124 			 */
3125 			vm_page_zero_fill(m);
3126 			mycpu->gd_cnt.v_zfod++;
3127 			m->valid = VM_PAGE_BITS_ALL;
3128 
3129 			/*
3130 			 * Handle dirty page case
3131 			 */
3132 			if (pprot & VM_PROT_WRITE)
3133 				vm_set_nosync(m, entry);
3134 			pmap_enter(pmap, addr, m, pprot, 0, entry);
3135 			mycpu->gd_cnt.v_vm_faults++;
3136 			if (curthread->td_lwp)
3137 				++curthread->td_lwp->lwp_ru.ru_minflt;
3138 			vm_page_deactivate(m);
3139 			if (pprot & VM_PROT_WRITE) {
3140 				/*vm_object_set_writeable_dirty(m->object);*/
3141 				vm_set_nosync(m, entry);
3142 				if (fault_flags & VM_FAULT_DIRTY) {
3143 					vm_page_dirty(m);
3144 					/*XXX*/
3145 					swap_pager_unswapped(m);
3146 				}
3147 			}
3148 			vm_page_wakeup(m);
3149 		} else if (error) {
3150 			/* couldn't busy page, no wakeup */
3151 		} else if (
3152 		    ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
3153 		    (m->flags & PG_FICTITIOUS) == 0) {
3154 			/*
3155 			 * A fully valid page not undergoing soft I/O can
3156 			 * be immediately entered into the pmap.
3157 			 */
3158 			if ((m->queue - m->pc) == PQ_CACHE)
3159 				vm_page_deactivate(m);
3160 			if (pprot & VM_PROT_WRITE) {
3161 				/*vm_object_set_writeable_dirty(m->object);*/
3162 				vm_set_nosync(m, entry);
3163 				if (fault_flags & VM_FAULT_DIRTY) {
3164 					vm_page_dirty(m);
3165 					/*XXX*/
3166 					swap_pager_unswapped(m);
3167 				}
3168 			}
3169 			if (pprot & VM_PROT_WRITE)
3170 				vm_set_nosync(m, entry);
3171 			pmap_enter(pmap, addr, m, pprot, 0, entry);
3172 			mycpu->gd_cnt.v_vm_faults++;
3173 			if (curthread->td_lwp)
3174 				++curthread->td_lwp->lwp_ru.ru_minflt;
3175 			vm_page_wakeup(m);
3176 		} else {
3177 			vm_page_wakeup(m);
3178 		}
3179 	}
3180 	vm_object_drop(object);
3181 }
3182 
3183 /*
3184  * Object can be held shared
3185  */
3186 static void
3187 vm_prefault_quick(pmap_t pmap, vm_offset_t addra,
3188 		  vm_map_entry_t entry, int prot, int fault_flags)
3189 {
3190 	struct lwp *lp;
3191 	vm_page_t m;
3192 	vm_offset_t addr;
3193 	vm_pindex_t pindex;
3194 	vm_object_t object;
3195 	int i;
3196 	int noneg;
3197 	int nopos;
3198 	int maxpages;
3199 
3200 	/*
3201 	 * Get stable max count value, disabled if set to 0
3202 	 */
3203 	maxpages = vm_prefault_pages;
3204 	cpu_ccfence();
3205 	if (maxpages <= 0)
3206 		return;
3207 
3208 	/*
3209 	 * We do not currently prefault mappings that use virtual page
3210 	 * tables.  We do not prefault foreign pmaps.
3211 	 */
3212 	if (entry->maptype != VM_MAPTYPE_NORMAL)
3213 		return;
3214 	lp = curthread->td_lwp;
3215 	if (lp == NULL || (pmap != vmspace_pmap(lp->lwp_vmspace)))
3216 		return;
3217 	object = entry->ba.object;
3218 	if (entry->ba.backing_ba != NULL)
3219 		return;
3220 	ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
3221 
3222 	/*
3223 	 * Limit pre-fault count to 1024 pages.
3224 	 */
3225 	if (maxpages > 1024)
3226 		maxpages = 1024;
3227 
3228 	noneg = 0;
3229 	nopos = 0;
3230 	for (i = 0; i < maxpages; ++i) {
3231 		int error;
3232 
3233 		/*
3234 		 * Calculate the page to pre-fault, stopping the scan in
3235 		 * each direction separately if the limit is reached.
3236 		 */
3237 		if (i & 1) {
3238 			if (noneg)
3239 				continue;
3240 			addr = addra - ((i + 1) >> 1) * PAGE_SIZE;
3241 		} else {
3242 			if (nopos)
3243 				continue;
3244 			addr = addra + ((i + 2) >> 1) * PAGE_SIZE;
3245 		}
3246 		if (addr < entry->start) {
3247 			noneg = 1;
3248 			if (noneg && nopos)
3249 				break;
3250 			continue;
3251 		}
3252 		if (addr >= entry->end) {
3253 			nopos = 1;
3254 			if (noneg && nopos)
3255 				break;
3256 			continue;
3257 		}
3258 
3259 		/*
3260 		 * Follow the VM object chain to obtain the page to be mapped
3261 		 * into the pmap.  This version of the prefault code only
3262 		 * works with terminal objects.
3263 		 *
3264 		 * The page must already exist.  If we encounter a problem
3265 		 * we stop here.
3266 		 *
3267 		 * WARNING!  We cannot call swap_pager_unswapped() or insert
3268 		 *	     a new vm_page with a shared token.
3269 		 */
3270 		pindex = ((addr - entry->start) + entry->ba.offset) >>
3271 			 PAGE_SHIFT;
3272 
3273 		/*
3274 		 * Skip pages already mapped, and stop scanning in that
3275 		 * direction.  When the scan terminates in both directions
3276 		 * we are done.
3277 		 */
3278 		if (pmap_prefault_ok(pmap, addr) == 0) {
3279 			if (i & 1)
3280 				noneg = 1;
3281 			else
3282 				nopos = 1;
3283 			if (noneg && nopos)
3284 				break;
3285 			continue;
3286 		}
3287 
3288 		/*
3289 		 * Shortcut the read-only mapping case using the far more
3290 		 * efficient vm_page_lookup_sbusy_try() function.  This
3291 		 * allows us to acquire the page soft-busied only which
3292 		 * is especially nice for concurrent execs of the same
3293 		 * program.
3294 		 *
3295 		 * The lookup function also validates page suitability
3296 		 * (all valid bits set, and not fictitious).
3297 		 *
3298 		 * If the page is in PQ_CACHE we have to fall-through
3299 		 * and hard-busy it so we can move it out of PQ_CACHE.
3300 		 */
3301 		if ((prot & VM_PROT_WRITE) == 0) {
3302 			m = vm_page_lookup_sbusy_try(object, pindex,
3303 						     0, PAGE_SIZE);
3304 			if (m == NULL)
3305 				break;
3306 			if ((m->queue - m->pc) != PQ_CACHE) {
3307 				pmap_enter(pmap, addr, m, prot, 0, entry);
3308 				mycpu->gd_cnt.v_vm_faults++;
3309 				if (curthread->td_lwp)
3310 					++curthread->td_lwp->lwp_ru.ru_minflt;
3311 				vm_page_sbusy_drop(m);
3312 				continue;
3313 			}
3314 			vm_page_sbusy_drop(m);
3315 		}
3316 
3317 		/*
3318 		 * Fallback to normal vm_page lookup code.  This code
3319 		 * hard-busies the page.  Not only that, but the page
3320 		 * can remain in that state for a significant period
3321 		 * time due to pmap_enter()'s overhead.
3322 		 */
3323 		m = vm_page_lookup_busy_try(object, pindex, TRUE, &error);
3324 		if (m == NULL || error)
3325 			break;
3326 
3327 		/*
3328 		 * Stop if the page cannot be trivially entered into the
3329 		 * pmap.
3330 		 */
3331 		if (((m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) ||
3332 		    (m->flags & PG_FICTITIOUS) ||
3333 		    ((m->flags & PG_SWAPPED) &&
3334 		     (prot & VM_PROT_WRITE) &&
3335 		     (fault_flags & VM_FAULT_DIRTY))) {
3336 			vm_page_wakeup(m);
3337 			break;
3338 		}
3339 
3340 		/*
3341 		 * Enter the page into the pmap.  The object might be held
3342 		 * shared so we can't do any (serious) modifying operation
3343 		 * on it.
3344 		 */
3345 		if ((m->queue - m->pc) == PQ_CACHE)
3346 			vm_page_deactivate(m);
3347 		if (prot & VM_PROT_WRITE) {
3348 			vm_object_set_writeable_dirty(m->object);
3349 			vm_set_nosync(m, entry);
3350 			if (fault_flags & VM_FAULT_DIRTY) {
3351 				vm_page_dirty(m);
3352 				/* can't happeen due to conditional above */
3353 				/* swap_pager_unswapped(m); */
3354 			}
3355 		}
3356 		pmap_enter(pmap, addr, m, prot, 0, entry);
3357 		mycpu->gd_cnt.v_vm_faults++;
3358 		if (curthread->td_lwp)
3359 			++curthread->td_lwp->lwp_ru.ru_minflt;
3360 		vm_page_wakeup(m);
3361 	}
3362 }
3363