xref: /dflybsd-src/sys/vm/vm_fault.c (revision d4b0d6715086a6d05c21e8e4dd1cfd5da335cf79)
1 /*
2  * Copyright (c) 2003-2014 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/vm_extern.h>
127 
128 #include <sys/thread2.h>
129 #include <vm/vm_page2.h>
130 
131 struct faultstate {
132 	vm_page_t m;
133 	vm_object_t object;
134 	vm_pindex_t pindex;
135 	vm_prot_t prot;
136 	vm_page_t first_m;
137 	vm_object_t first_object;
138 	vm_prot_t first_prot;
139 	vm_map_t map;
140 	vm_map_entry_t entry;
141 	int lookup_still_valid;
142 	int hardfault;
143 	int fault_flags;
144 	int map_generation;
145 	int shared;
146 	int first_shared;
147 	boolean_t wired;
148 	struct vnode *vp;
149 };
150 
151 static int debug_fault = 0;
152 SYSCTL_INT(_vm, OID_AUTO, debug_fault, CTLFLAG_RW, &debug_fault, 0, "");
153 static int debug_cluster = 0;
154 SYSCTL_INT(_vm, OID_AUTO, debug_cluster, CTLFLAG_RW, &debug_cluster, 0, "");
155 int vm_shared_fault = 1;
156 TUNABLE_INT("vm.shared_fault", &vm_shared_fault);
157 SYSCTL_INT(_vm, OID_AUTO, shared_fault, CTLFLAG_RW, &vm_shared_fault, 0,
158 	   "Allow shared token on vm_object");
159 static long vm_shared_hit = 0;
160 SYSCTL_LONG(_vm, OID_AUTO, shared_hit, CTLFLAG_RW, &vm_shared_hit, 0,
161 	   "Successful shared faults");
162 static long vm_shared_count = 0;
163 SYSCTL_LONG(_vm, OID_AUTO, shared_count, CTLFLAG_RW, &vm_shared_count, 0,
164 	   "Shared fault attempts");
165 static long vm_shared_miss = 0;
166 SYSCTL_LONG(_vm, OID_AUTO, shared_miss, CTLFLAG_RW, &vm_shared_miss, 0,
167 	   "Unsuccessful shared faults");
168 
169 static int vm_fault_object(struct faultstate *, vm_pindex_t, vm_prot_t, int);
170 static int vm_fault_vpagetable(struct faultstate *, vm_pindex_t *,
171 			vpte_t, int, int);
172 #if 0
173 static int vm_fault_additional_pages (vm_page_t, int, int, vm_page_t *, int *);
174 #endif
175 static void vm_set_nosync(vm_page_t m, vm_map_entry_t entry);
176 static void vm_prefault(pmap_t pmap, vm_offset_t addra,
177 			vm_map_entry_t entry, int prot, int fault_flags);
178 static void vm_prefault_quick(pmap_t pmap, vm_offset_t addra,
179 			vm_map_entry_t entry, int prot, int fault_flags);
180 
181 static __inline void
182 release_page(struct faultstate *fs)
183 {
184 	vm_page_deactivate(fs->m);
185 	vm_page_wakeup(fs->m);
186 	fs->m = NULL;
187 }
188 
189 /*
190  * NOTE: Once unlocked any cached fs->entry becomes invalid, any reuse
191  *	 requires relocking and then checking the timestamp.
192  *
193  * NOTE: vm_map_lock_read() does not bump fs->map->timestamp so we do
194  *	 not have to update fs->map_generation here.
195  *
196  * NOTE: This function can fail due to a deadlock against the caller's
197  *	 holding of a vm_page BUSY.
198  */
199 static __inline int
200 relock_map(struct faultstate *fs)
201 {
202 	int error;
203 
204 	if (fs->lookup_still_valid == FALSE && fs->map) {
205 		error = vm_map_lock_read_to(fs->map);
206 		if (error == 0)
207 			fs->lookup_still_valid = TRUE;
208 	} else {
209 		error = 0;
210 	}
211 	return error;
212 }
213 
214 static __inline void
215 unlock_map(struct faultstate *fs)
216 {
217 	if (fs->lookup_still_valid && fs->map) {
218 		vm_map_lookup_done(fs->map, fs->entry, 0);
219 		fs->lookup_still_valid = FALSE;
220 	}
221 }
222 
223 /*
224  * Clean up after a successful call to vm_fault_object() so another call
225  * to vm_fault_object() can be made.
226  */
227 static void
228 _cleanup_successful_fault(struct faultstate *fs, int relock)
229 {
230 	/*
231 	 * We allocated a junk page for a COW operation that did
232 	 * not occur, the page must be freed.
233 	 */
234 	if (fs->object != fs->first_object) {
235 		KKASSERT(fs->first_shared == 0);
236 		vm_page_free(fs->first_m);
237 		vm_object_pip_wakeup(fs->object);
238 		fs->first_m = NULL;
239 	}
240 
241 	/*
242 	 * Reset fs->object.
243 	 */
244 	fs->object = fs->first_object;
245 	if (relock && fs->lookup_still_valid == FALSE) {
246 		if (fs->map)
247 			vm_map_lock_read(fs->map);
248 		fs->lookup_still_valid = TRUE;
249 	}
250 }
251 
252 static void
253 _unlock_things(struct faultstate *fs, int dealloc)
254 {
255 	_cleanup_successful_fault(fs, 0);
256 	if (dealloc) {
257 		/*vm_object_deallocate(fs->first_object);*/
258 		/*fs->first_object = NULL; drop used later on */
259 	}
260 	unlock_map(fs);
261 	if (fs->vp != NULL) {
262 		vput(fs->vp);
263 		fs->vp = NULL;
264 	}
265 }
266 
267 #define unlock_things(fs) _unlock_things(fs, 0)
268 #define unlock_and_deallocate(fs) _unlock_things(fs, 1)
269 #define cleanup_successful_fault(fs) _cleanup_successful_fault(fs, 1)
270 
271 /*
272  * TRYPAGER
273  *
274  * Determine if the pager for the current object *might* contain the page.
275  *
276  * We only need to try the pager if this is not a default object (default
277  * objects are zero-fill and have no real pager), and if we are not taking
278  * a wiring fault or if the FS entry is wired.
279  */
280 #define TRYPAGER(fs)	\
281 		(fs->object->type != OBJT_DEFAULT && \
282 		(((fs->fault_flags & VM_FAULT_WIRE_MASK) == 0) || fs->wired))
283 
284 /*
285  * vm_fault:
286  *
287  * Handle a page fault occuring at the given address, requiring the given
288  * permissions, in the map specified.  If successful, the page is inserted
289  * into the associated physical map.
290  *
291  * NOTE: The given address should be truncated to the proper page address.
292  *
293  * KERN_SUCCESS is returned if the page fault is handled; otherwise,
294  * a standard error specifying why the fault is fatal is returned.
295  *
296  * The map in question must be referenced, and remains so.
297  * The caller may hold no locks.
298  * No other requirements.
299  */
300 int
301 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, int fault_flags)
302 {
303 	int result;
304 	vm_pindex_t first_pindex;
305 	struct faultstate fs;
306 	struct lwp *lp;
307 	int growstack;
308 	int retry = 0;
309 	int inherit_prot;
310 
311 	inherit_prot = fault_type & VM_PROT_NOSYNC;
312 	fs.hardfault = 0;
313 	fs.fault_flags = fault_flags;
314 	fs.vp = NULL;
315 	fs.shared = vm_shared_fault;
316 	fs.first_shared = vm_shared_fault;
317 	growstack = 1;
318 	if (vm_shared_fault)
319 		++vm_shared_count;
320 
321 	/*
322 	 * vm_map interactions
323 	 */
324 	if ((lp = curthread->td_lwp) != NULL)
325 		lp->lwp_flags |= LWP_PAGING;
326 	lwkt_gettoken(&map->token);
327 
328 RetryFault:
329 	/*
330 	 * Find the vm_map_entry representing the backing store and resolve
331 	 * the top level object and page index.  This may have the side
332 	 * effect of executing a copy-on-write on the map entry and/or
333 	 * creating a shadow object, but will not COW any actual VM pages.
334 	 *
335 	 * On success fs.map is left read-locked and various other fields
336 	 * are initialized but not otherwise referenced or locked.
337 	 *
338 	 * NOTE!  vm_map_lookup will try to upgrade the fault_type to
339 	 * VM_FAULT_WRITE if the map entry is a virtual page table and also
340 	 * writable, so we can set the 'A'accessed bit in the virtual page
341 	 * table entry.
342 	 */
343 	fs.map = map;
344 	result = vm_map_lookup(&fs.map, vaddr, fault_type,
345 			       &fs.entry, &fs.first_object,
346 			       &first_pindex, &fs.first_prot, &fs.wired);
347 
348 	/*
349 	 * If the lookup failed or the map protections are incompatible,
350 	 * the fault generally fails.
351 	 *
352 	 * The failure could be due to TDF_NOFAULT if vm_map_lookup()
353 	 * tried to do a COW fault.
354 	 *
355 	 * If the caller is trying to do a user wiring we have more work
356 	 * to do.
357 	 */
358 	if (result != KERN_SUCCESS) {
359 		if (result == KERN_FAILURE_NOFAULT) {
360 			result = KERN_FAILURE;
361 			goto done;
362 		}
363 		if (result != KERN_PROTECTION_FAILURE ||
364 		    (fs.fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE)
365 		{
366 			if (result == KERN_INVALID_ADDRESS && growstack &&
367 			    map != &kernel_map && curproc != NULL) {
368 				result = vm_map_growstack(curproc, vaddr);
369 				if (result == KERN_SUCCESS) {
370 					growstack = 0;
371 					++retry;
372 					goto RetryFault;
373 				}
374 				result = KERN_FAILURE;
375 			}
376 			goto done;
377 		}
378 
379 		/*
380    		 * If we are user-wiring a r/w segment, and it is COW, then
381    		 * we need to do the COW operation.  Note that we don't
382 		 * currently COW RO sections now, because it is NOT desirable
383    		 * to COW .text.  We simply keep .text from ever being COW'ed
384    		 * and take the heat that one cannot debug wired .text sections.
385    		 */
386 		result = vm_map_lookup(&fs.map, vaddr,
387 				       VM_PROT_READ|VM_PROT_WRITE|
388 				        VM_PROT_OVERRIDE_WRITE,
389 				       &fs.entry, &fs.first_object,
390 				       &first_pindex, &fs.first_prot,
391 				       &fs.wired);
392 		if (result != KERN_SUCCESS) {
393 			/* could also be KERN_FAILURE_NOFAULT */
394 			result = KERN_FAILURE;
395 			goto done;
396 		}
397 
398 		/*
399 		 * If we don't COW now, on a user wire, the user will never
400 		 * be able to write to the mapping.  If we don't make this
401 		 * restriction, the bookkeeping would be nearly impossible.
402 		 *
403 		 * XXX We have a shared lock, this will have a MP race but
404 		 * I don't see how it can hurt anything.
405 		 */
406 		if ((fs.entry->protection & VM_PROT_WRITE) == 0)
407 			fs.entry->max_protection &= ~VM_PROT_WRITE;
408 	}
409 
410 	/*
411 	 * fs.map is read-locked
412 	 *
413 	 * Misc checks.  Save the map generation number to detect races.
414 	 */
415 	fs.map_generation = fs.map->timestamp;
416 	fs.lookup_still_valid = TRUE;
417 	fs.first_m = NULL;
418 	fs.object = fs.first_object;	/* so unlock_and_deallocate works */
419 	fs.prot = fs.first_prot;	/* default (used by uksmap) */
420 
421 	if (fs.entry->eflags & (MAP_ENTRY_NOFAULT | MAP_ENTRY_KSTACK)) {
422 		if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
423 			panic("vm_fault: fault on nofault entry, addr: %p",
424 			      (void *)vaddr);
425 		}
426 		if ((fs.entry->eflags & MAP_ENTRY_KSTACK) &&
427 		    vaddr >= fs.entry->start &&
428 		    vaddr < fs.entry->start + PAGE_SIZE) {
429 			panic("vm_fault: fault on stack guard, addr: %p",
430 			      (void *)vaddr);
431 		}
432 	}
433 
434 	/*
435 	 * A user-kernel shared map has no VM object and bypasses
436 	 * everything.  We execute the uksmap function with a temporary
437 	 * fictitious vm_page.  The address is directly mapped with no
438 	 * management.
439 	 */
440 	if (fs.entry->maptype == VM_MAPTYPE_UKSMAP) {
441 		struct vm_page fakem;
442 
443 		bzero(&fakem, sizeof(fakem));
444 		fakem.pindex = first_pindex;
445 		fakem.flags = PG_BUSY | PG_FICTITIOUS | PG_UNMANAGED;
446 		fakem.valid = VM_PAGE_BITS_ALL;
447 		fakem.pat_mode = VM_MEMATTR_DEFAULT;
448 		if (fs.entry->object.uksmap(fs.entry->aux.dev, &fakem)) {
449 			result = KERN_FAILURE;
450 			unlock_things(&fs);
451 			goto done2;
452 		}
453 		pmap_enter(fs.map->pmap, vaddr, &fakem, fs.prot | inherit_prot,
454 			   fs.wired, fs.entry);
455 		goto done_success;
456 	}
457 
458 	/*
459 	 * A system map entry may return a NULL object.  No object means
460 	 * no pager means an unrecoverable kernel fault.
461 	 */
462 	if (fs.first_object == NULL) {
463 		panic("vm_fault: unrecoverable fault at %p in entry %p",
464 			(void *)vaddr, fs.entry);
465 	}
466 
467 	/*
468 	 * Fail here if not a trivial anonymous page fault and TDF_NOFAULT
469 	 * is set.
470 	 *
471 	 * Unfortunately a deadlock can occur if we are forced to page-in
472 	 * from swap, but diving all the way into the vm_pager_get_page()
473 	 * function to find out is too much.  Just check the object type.
474 	 *
475 	 * The deadlock is a CAM deadlock on a busy VM page when trying
476 	 * to finish an I/O if another process gets stuck in
477 	 * vop_helper_read_shortcut() due to a swap fault.
478 	 */
479 	if ((curthread->td_flags & TDF_NOFAULT) &&
480 	    (retry ||
481 	     fs.first_object->type == OBJT_VNODE ||
482 	     fs.first_object->type == OBJT_SWAP ||
483 	     fs.first_object->backing_object)) {
484 		result = KERN_FAILURE;
485 		unlock_things(&fs);
486 		goto done2;
487 	}
488 
489 	/*
490 	 * If the entry is wired we cannot change the page protection.
491 	 */
492 	if (fs.wired)
493 		fault_type = fs.first_prot;
494 
495 	/*
496 	 * We generally want to avoid unnecessary exclusive modes on backing
497 	 * and terminal objects because this can seriously interfere with
498 	 * heavily fork()'d processes (particularly /bin/sh scripts).
499 	 *
500 	 * However, we also want to avoid unnecessary retries due to needed
501 	 * shared->exclusive promotion for common faults.  Exclusive mode is
502 	 * always needed if any page insertion, rename, or free occurs in an
503 	 * object (and also indirectly if any I/O is done).
504 	 *
505 	 * The main issue here is going to be fs.first_shared.  If the
506 	 * first_object has a backing object which isn't shadowed and the
507 	 * process is single-threaded we might as well use an exclusive
508 	 * lock/chain right off the bat.
509 	 */
510 	if (fs.first_shared && fs.first_object->backing_object &&
511 	    LIST_EMPTY(&fs.first_object->shadow_head) &&
512 	    curthread->td_proc && curthread->td_proc->p_nthreads == 1) {
513 		fs.first_shared = 0;
514 	}
515 
516 	/*
517 	 * swap_pager_unswapped() needs an exclusive object
518 	 */
519 	if (fault_flags & (VM_FAULT_UNSWAP | VM_FAULT_DIRTY)) {
520 		fs.first_shared = 0;
521 	}
522 
523 	/*
524 	 * Obtain a top-level object lock, shared or exclusive depending
525 	 * on fs.first_shared.  If a shared lock winds up being insufficient
526 	 * we will retry with an exclusive lock.
527 	 *
528 	 * The vnode pager lock is always shared.
529 	 */
530 	if (fs.first_shared)
531 		vm_object_hold_shared(fs.first_object);
532 	else
533 		vm_object_hold(fs.first_object);
534 	if (fs.vp == NULL)
535 		fs.vp = vnode_pager_lock(fs.first_object);
536 
537 	/*
538 	 * The page we want is at (first_object, first_pindex), but if the
539 	 * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the
540 	 * page table to figure out the actual pindex.
541 	 *
542 	 * NOTE!  DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION
543 	 * ONLY
544 	 */
545 	if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) {
546 		result = vm_fault_vpagetable(&fs, &first_pindex,
547 					     fs.entry->aux.master_pde,
548 					     fault_type, 1);
549 		if (result == KERN_TRY_AGAIN) {
550 			vm_object_drop(fs.first_object);
551 			++retry;
552 			goto RetryFault;
553 		}
554 		if (result != KERN_SUCCESS)
555 			goto done;
556 	}
557 
558 	/*
559 	 * Now we have the actual (object, pindex), fault in the page.  If
560 	 * vm_fault_object() fails it will unlock and deallocate the FS
561 	 * data.   If it succeeds everything remains locked and fs->object
562 	 * will have an additional PIP count if it is not equal to
563 	 * fs->first_object
564 	 *
565 	 * vm_fault_object will set fs->prot for the pmap operation.  It is
566 	 * allowed to set VM_PROT_WRITE if fault_type == VM_PROT_READ if the
567 	 * page can be safely written.  However, it will force a read-only
568 	 * mapping for a read fault if the memory is managed by a virtual
569 	 * page table.
570 	 *
571 	 * If the fault code uses the shared object lock shortcut
572 	 * we must not try to burst (we can't allocate VM pages).
573 	 */
574 	result = vm_fault_object(&fs, first_pindex, fault_type, 1);
575 
576 	if (debug_fault > 0) {
577 		--debug_fault;
578 		kprintf("VM_FAULT result %d addr=%jx type=%02x flags=%02x "
579 			"fs.m=%p fs.prot=%02x fs.wired=%02x fs.entry=%p\n",
580 			result, (intmax_t)vaddr, fault_type, fault_flags,
581 			fs.m, fs.prot, fs.wired, fs.entry);
582 	}
583 
584 	if (result == KERN_TRY_AGAIN) {
585 		vm_object_drop(fs.first_object);
586 		++retry;
587 		goto RetryFault;
588 	}
589 	if (result != KERN_SUCCESS)
590 		goto done;
591 
592 	/*
593 	 * On success vm_fault_object() does not unlock or deallocate, and fs.m
594 	 * will contain a busied page.
595 	 *
596 	 * Enter the page into the pmap and do pmap-related adjustments.
597 	 */
598 	KKASSERT(fs.lookup_still_valid == TRUE);
599 	vm_page_flag_set(fs.m, PG_REFERENCED);
600 	pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot | inherit_prot,
601 		   fs.wired, fs.entry);
602 
603 	/*KKASSERT(fs.m->queue == PQ_NONE); page-in op may deactivate page */
604 	KKASSERT(fs.m->flags & PG_BUSY);
605 
606 	/*
607 	 * If the page is not wired down, then put it where the pageout daemon
608 	 * can find it.
609 	 */
610 	if (fs.fault_flags & VM_FAULT_WIRE_MASK) {
611 		if (fs.wired)
612 			vm_page_wire(fs.m);
613 		else
614 			vm_page_unwire(fs.m, 1);
615 	} else {
616 		vm_page_activate(fs.m);
617 	}
618 	vm_page_wakeup(fs.m);
619 
620 	/*
621 	 * Burst in a few more pages if possible.  The fs.map should still
622 	 * be locked.  To avoid interlocking against a vnode->getblk
623 	 * operation we had to be sure to unbusy our primary vm_page above
624 	 * first.
625 	 *
626 	 * A normal burst can continue down backing store, only execute
627 	 * if we are holding an exclusive lock, otherwise the exclusive
628 	 * locks the burst code gets might cause excessive SMP collisions.
629 	 *
630 	 * A quick burst can be utilized when there is no backing object
631 	 * (i.e. a shared file mmap).
632 	 */
633 	if ((fault_flags & VM_FAULT_BURST) &&
634 	    (fs.fault_flags & VM_FAULT_WIRE_MASK) == 0 &&
635 	    fs.wired == 0) {
636 		if (fs.first_shared == 0 && fs.shared == 0) {
637 			vm_prefault(fs.map->pmap, vaddr,
638 				    fs.entry, fs.prot, fault_flags);
639 		} else {
640 			vm_prefault_quick(fs.map->pmap, vaddr,
641 					  fs.entry, fs.prot, fault_flags);
642 		}
643 	}
644 
645 done_success:
646 	mycpu->gd_cnt.v_vm_faults++;
647 	if (curthread->td_lwp)
648 		++curthread->td_lwp->lwp_ru.ru_minflt;
649 
650 	/*
651 	 * Unlock everything, and return
652 	 */
653 	unlock_things(&fs);
654 
655 	if (curthread->td_lwp) {
656 		if (fs.hardfault) {
657 			curthread->td_lwp->lwp_ru.ru_majflt++;
658 		} else {
659 			curthread->td_lwp->lwp_ru.ru_minflt++;
660 		}
661 	}
662 
663 	/*vm_object_deallocate(fs.first_object);*/
664 	/*fs.m = NULL; */
665 	/*fs.first_object = NULL; must still drop later */
666 
667 	result = KERN_SUCCESS;
668 done:
669 	if (fs.first_object)
670 		vm_object_drop(fs.first_object);
671 done2:
672 	lwkt_reltoken(&map->token);
673 	if (lp)
674 		lp->lwp_flags &= ~LWP_PAGING;
675 	if (vm_shared_fault && fs.shared == 0)
676 		++vm_shared_miss;
677 	return (result);
678 }
679 
680 /*
681  * Fault in the specified virtual address in the current process map,
682  * returning a held VM page or NULL.  See vm_fault_page() for more
683  * information.
684  *
685  * No requirements.
686  */
687 vm_page_t
688 vm_fault_page_quick(vm_offset_t va, vm_prot_t fault_type, int *errorp)
689 {
690 	struct lwp *lp = curthread->td_lwp;
691 	vm_page_t m;
692 
693 	m = vm_fault_page(&lp->lwp_vmspace->vm_map, va,
694 			  fault_type, VM_FAULT_NORMAL, errorp);
695 	return(m);
696 }
697 
698 /*
699  * Fault in the specified virtual address in the specified map, doing all
700  * necessary manipulation of the object store and all necessary I/O.  Return
701  * a held VM page or NULL, and set *errorp.  The related pmap is not
702  * updated.
703  *
704  * The returned page will be properly dirtied if VM_PROT_WRITE was specified,
705  * and marked PG_REFERENCED as well.
706  *
707  * If the page cannot be faulted writable and VM_PROT_WRITE was specified, an
708  * error will be returned.
709  *
710  * No requirements.
711  */
712 vm_page_t
713 vm_fault_page(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
714 	      int fault_flags, int *errorp)
715 {
716 	vm_pindex_t first_pindex;
717 	struct faultstate fs;
718 	int result;
719 	int retry = 0;
720 	vm_prot_t orig_fault_type = fault_type;
721 
722 	fs.hardfault = 0;
723 	fs.fault_flags = fault_flags;
724 	KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0);
725 
726 	/*
727 	 * Dive the pmap (concurrency possible).  If we find the
728 	 * appropriate page we can terminate early and quickly.
729 	 */
730 	fs.m = pmap_fault_page_quick(map->pmap, vaddr, fault_type);
731 	if (fs.m) {
732 		*errorp = 0;
733 		return(fs.m);
734 	}
735 
736 	/*
737 	 * Otherwise take a concurrency hit and do a formal page
738 	 * fault.
739 	 */
740 	fs.shared = vm_shared_fault;
741 	fs.first_shared = vm_shared_fault;
742 	fs.vp = NULL;
743 	lwkt_gettoken(&map->token);
744 
745 	/*
746 	 * swap_pager_unswapped() needs an exclusive object
747 	 */
748 	if (fault_flags & (VM_FAULT_UNSWAP | VM_FAULT_DIRTY)) {
749 		fs.first_shared = 0;
750 	}
751 
752 RetryFault:
753 	/*
754 	 * Find the vm_map_entry representing the backing store and resolve
755 	 * the top level object and page index.  This may have the side
756 	 * effect of executing a copy-on-write on the map entry and/or
757 	 * creating a shadow object, but will not COW any actual VM pages.
758 	 *
759 	 * On success fs.map is left read-locked and various other fields
760 	 * are initialized but not otherwise referenced or locked.
761 	 *
762 	 * NOTE!  vm_map_lookup will upgrade the fault_type to VM_FAULT_WRITE
763 	 * if the map entry is a virtual page table and also writable,
764 	 * so we can set the 'A'accessed bit in the virtual page table entry.
765 	 */
766 	fs.map = map;
767 	result = vm_map_lookup(&fs.map, vaddr, fault_type,
768 			       &fs.entry, &fs.first_object,
769 			       &first_pindex, &fs.first_prot, &fs.wired);
770 
771 	if (result != KERN_SUCCESS) {
772 		*errorp = result;
773 		fs.m = NULL;
774 		goto done;
775 	}
776 
777 	/*
778 	 * fs.map is read-locked
779 	 *
780 	 * Misc checks.  Save the map generation number to detect races.
781 	 */
782 	fs.map_generation = fs.map->timestamp;
783 	fs.lookup_still_valid = TRUE;
784 	fs.first_m = NULL;
785 	fs.object = fs.first_object;	/* so unlock_and_deallocate works */
786 
787 	if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
788 		panic("vm_fault: fault on nofault entry, addr: %lx",
789 		    (u_long)vaddr);
790 	}
791 
792 	/*
793 	 * A user-kernel shared map has no VM object and bypasses
794 	 * everything.  We execute the uksmap function with a temporary
795 	 * fictitious vm_page.  The address is directly mapped with no
796 	 * management.
797 	 */
798 	if (fs.entry->maptype == VM_MAPTYPE_UKSMAP) {
799 		struct vm_page fakem;
800 
801 		bzero(&fakem, sizeof(fakem));
802 		fakem.pindex = first_pindex;
803 		fakem.flags = PG_BUSY | PG_FICTITIOUS | PG_UNMANAGED;
804 		fakem.valid = VM_PAGE_BITS_ALL;
805 		fakem.pat_mode = VM_MEMATTR_DEFAULT;
806 		if (fs.entry->object.uksmap(fs.entry->aux.dev, &fakem)) {
807 			*errorp = KERN_FAILURE;
808 			fs.m = NULL;
809 			unlock_things(&fs);
810 			goto done2;
811 		}
812 		fs.m = PHYS_TO_VM_PAGE(fakem.phys_addr);
813 		vm_page_hold(fs.m);
814 
815 		unlock_things(&fs);
816 		*errorp = 0;
817 		goto done;
818 	}
819 
820 
821 	/*
822 	 * A system map entry may return a NULL object.  No object means
823 	 * no pager means an unrecoverable kernel fault.
824 	 */
825 	if (fs.first_object == NULL) {
826 		panic("vm_fault: unrecoverable fault at %p in entry %p",
827 			(void *)vaddr, fs.entry);
828 	}
829 
830 	/*
831 	 * Fail here if not a trivial anonymous page fault and TDF_NOFAULT
832 	 * is set.
833 	 *
834 	 * Unfortunately a deadlock can occur if we are forced to page-in
835 	 * from swap, but diving all the way into the vm_pager_get_page()
836 	 * function to find out is too much.  Just check the object type.
837 	 */
838 	if ((curthread->td_flags & TDF_NOFAULT) &&
839 	    (retry ||
840 	     fs.first_object->type == OBJT_VNODE ||
841 	     fs.first_object->type == OBJT_SWAP ||
842 	     fs.first_object->backing_object)) {
843 		*errorp = KERN_FAILURE;
844 		unlock_things(&fs);
845 		goto done2;
846 	}
847 
848 	/*
849 	 * If the entry is wired we cannot change the page protection.
850 	 */
851 	if (fs.wired)
852 		fault_type = fs.first_prot;
853 
854 	/*
855 	 * Make a reference to this object to prevent its disposal while we
856 	 * are messing with it.  Once we have the reference, the map is free
857 	 * to be diddled.  Since objects reference their shadows (and copies),
858 	 * they will stay around as well.
859 	 *
860 	 * The reference should also prevent an unexpected collapse of the
861 	 * parent that might move pages from the current object into the
862 	 * parent unexpectedly, resulting in corruption.
863 	 *
864 	 * Bump the paging-in-progress count to prevent size changes (e.g.
865 	 * truncation operations) during I/O.  This must be done after
866 	 * obtaining the vnode lock in order to avoid possible deadlocks.
867 	 */
868 	if (fs.first_shared)
869 		vm_object_hold_shared(fs.first_object);
870 	else
871 		vm_object_hold(fs.first_object);
872 	if (fs.vp == NULL)
873 		fs.vp = vnode_pager_lock(fs.first_object);	/* shared */
874 
875 	/*
876 	 * The page we want is at (first_object, first_pindex), but if the
877 	 * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the
878 	 * page table to figure out the actual pindex.
879 	 *
880 	 * NOTE!  DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION
881 	 * ONLY
882 	 */
883 	if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) {
884 		result = vm_fault_vpagetable(&fs, &first_pindex,
885 					     fs.entry->aux.master_pde,
886 					     fault_type, 1);
887 		if (result == KERN_TRY_AGAIN) {
888 			vm_object_drop(fs.first_object);
889 			++retry;
890 			goto RetryFault;
891 		}
892 		if (result != KERN_SUCCESS) {
893 			*errorp = result;
894 			fs.m = NULL;
895 			goto done;
896 		}
897 	}
898 
899 	/*
900 	 * Now we have the actual (object, pindex), fault in the page.  If
901 	 * vm_fault_object() fails it will unlock and deallocate the FS
902 	 * data.   If it succeeds everything remains locked and fs->object
903 	 * will have an additinal PIP count if it is not equal to
904 	 * fs->first_object
905 	 */
906 	fs.m = NULL;
907 	result = vm_fault_object(&fs, first_pindex, fault_type, 1);
908 
909 	if (result == KERN_TRY_AGAIN) {
910 		vm_object_drop(fs.first_object);
911 		++retry;
912 		goto RetryFault;
913 	}
914 	if (result != KERN_SUCCESS) {
915 		*errorp = result;
916 		fs.m = NULL;
917 		goto done;
918 	}
919 
920 	if ((orig_fault_type & VM_PROT_WRITE) &&
921 	    (fs.prot & VM_PROT_WRITE) == 0) {
922 		*errorp = KERN_PROTECTION_FAILURE;
923 		unlock_and_deallocate(&fs);
924 		fs.m = NULL;
925 		goto done;
926 	}
927 
928 	/*
929 	 * DO NOT UPDATE THE PMAP!!!  This function may be called for
930 	 * a pmap unrelated to the current process pmap, in which case
931 	 * the current cpu core will not be listed in the pmap's pm_active
932 	 * mask.  Thus invalidation interlocks will fail to work properly.
933 	 *
934 	 * (for example, 'ps' uses procfs to read program arguments from
935 	 * each process's stack).
936 	 *
937 	 * In addition to the above this function will be called to acquire
938 	 * a page that might already be faulted in, re-faulting it
939 	 * continuously is a waste of time.
940 	 *
941 	 * XXX could this have been the cause of our random seg-fault
942 	 *     issues?  procfs accesses user stacks.
943 	 */
944 	vm_page_flag_set(fs.m, PG_REFERENCED);
945 #if 0
946 	pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot, fs.wired, NULL);
947 	mycpu->gd_cnt.v_vm_faults++;
948 	if (curthread->td_lwp)
949 		++curthread->td_lwp->lwp_ru.ru_minflt;
950 #endif
951 
952 	/*
953 	 * On success vm_fault_object() does not unlock or deallocate, and fs.m
954 	 * will contain a busied page.  So we must unlock here after having
955 	 * messed with the pmap.
956 	 */
957 	unlock_things(&fs);
958 
959 	/*
960 	 * Return a held page.  We are not doing any pmap manipulation so do
961 	 * not set PG_MAPPED.  However, adjust the page flags according to
962 	 * the fault type because the caller may not use a managed pmapping
963 	 * (so we don't want to lose the fact that the page will be dirtied
964 	 * if a write fault was specified).
965 	 */
966 	vm_page_hold(fs.m);
967 	vm_page_activate(fs.m);
968 	if (fault_type & VM_PROT_WRITE)
969 		vm_page_dirty(fs.m);
970 
971 	if (curthread->td_lwp) {
972 		if (fs.hardfault) {
973 			curthread->td_lwp->lwp_ru.ru_majflt++;
974 		} else {
975 			curthread->td_lwp->lwp_ru.ru_minflt++;
976 		}
977 	}
978 
979 	/*
980 	 * Unlock everything, and return the held page.
981 	 */
982 	vm_page_wakeup(fs.m);
983 	/*vm_object_deallocate(fs.first_object);*/
984 	/*fs.first_object = NULL; */
985 	*errorp = 0;
986 
987 done:
988 	if (fs.first_object)
989 		vm_object_drop(fs.first_object);
990 done2:
991 	lwkt_reltoken(&map->token);
992 	return(fs.m);
993 }
994 
995 /*
996  * Fault in the specified (object,offset), dirty the returned page as
997  * needed.  If the requested fault_type cannot be done NULL and an
998  * error is returned.
999  *
1000  * A held (but not busied) page is returned.
1001  *
1002  * The passed in object must be held as specified by the shared
1003  * argument.
1004  */
1005 vm_page_t
1006 vm_fault_object_page(vm_object_t object, vm_ooffset_t offset,
1007 		     vm_prot_t fault_type, int fault_flags,
1008 		     int *sharedp, int *errorp)
1009 {
1010 	int result;
1011 	vm_pindex_t first_pindex;
1012 	struct faultstate fs;
1013 	struct vm_map_entry entry;
1014 
1015 	ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1016 	bzero(&entry, sizeof(entry));
1017 	entry.object.vm_object = object;
1018 	entry.maptype = VM_MAPTYPE_NORMAL;
1019 	entry.protection = entry.max_protection = fault_type;
1020 
1021 	fs.hardfault = 0;
1022 	fs.fault_flags = fault_flags;
1023 	fs.map = NULL;
1024 	fs.shared = vm_shared_fault;
1025 	fs.first_shared = *sharedp;
1026 	fs.vp = NULL;
1027 	KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0);
1028 
1029 	/*
1030 	 * Might require swap block adjustments
1031 	 */
1032 	if (fs.first_shared && (fault_flags & (VM_FAULT_UNSWAP | VM_FAULT_DIRTY))) {
1033 		fs.first_shared = 0;
1034 		vm_object_upgrade(object);
1035 	}
1036 
1037 	/*
1038 	 * Retry loop as needed (typically for shared->exclusive transitions)
1039 	 */
1040 RetryFault:
1041 	*sharedp = fs.first_shared;
1042 	first_pindex = OFF_TO_IDX(offset);
1043 	fs.first_object = object;
1044 	fs.entry = &entry;
1045 	fs.first_prot = fault_type;
1046 	fs.wired = 0;
1047 	/*fs.map_generation = 0; unused */
1048 
1049 	/*
1050 	 * Make a reference to this object to prevent its disposal while we
1051 	 * are messing with it.  Once we have the reference, the map is free
1052 	 * to be diddled.  Since objects reference their shadows (and copies),
1053 	 * they will stay around as well.
1054 	 *
1055 	 * The reference should also prevent an unexpected collapse of the
1056 	 * parent that might move pages from the current object into the
1057 	 * parent unexpectedly, resulting in corruption.
1058 	 *
1059 	 * Bump the paging-in-progress count to prevent size changes (e.g.
1060 	 * truncation operations) during I/O.  This must be done after
1061 	 * obtaining the vnode lock in order to avoid possible deadlocks.
1062 	 */
1063 	if (fs.vp == NULL)
1064 		fs.vp = vnode_pager_lock(fs.first_object);
1065 
1066 	fs.lookup_still_valid = TRUE;
1067 	fs.first_m = NULL;
1068 	fs.object = fs.first_object;	/* so unlock_and_deallocate works */
1069 
1070 #if 0
1071 	/* XXX future - ability to operate on VM object using vpagetable */
1072 	if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) {
1073 		result = vm_fault_vpagetable(&fs, &first_pindex,
1074 					     fs.entry->aux.master_pde,
1075 					     fault_type, 0);
1076 		if (result == KERN_TRY_AGAIN) {
1077 			if (fs.first_shared == 0 && *sharedp)
1078 				vm_object_upgrade(object);
1079 			goto RetryFault;
1080 		}
1081 		if (result != KERN_SUCCESS) {
1082 			*errorp = result;
1083 			return (NULL);
1084 		}
1085 	}
1086 #endif
1087 
1088 	/*
1089 	 * Now we have the actual (object, pindex), fault in the page.  If
1090 	 * vm_fault_object() fails it will unlock and deallocate the FS
1091 	 * data.   If it succeeds everything remains locked and fs->object
1092 	 * will have an additinal PIP count if it is not equal to
1093 	 * fs->first_object
1094 	 *
1095 	 * On KERN_TRY_AGAIN vm_fault_object() leaves fs.first_object intact.
1096 	 * We may have to upgrade its lock to handle the requested fault.
1097 	 */
1098 	result = vm_fault_object(&fs, first_pindex, fault_type, 0);
1099 
1100 	if (result == KERN_TRY_AGAIN) {
1101 		if (fs.first_shared == 0 && *sharedp)
1102 			vm_object_upgrade(object);
1103 		goto RetryFault;
1104 	}
1105 	if (result != KERN_SUCCESS) {
1106 		*errorp = result;
1107 		return(NULL);
1108 	}
1109 
1110 	if ((fault_type & VM_PROT_WRITE) && (fs.prot & VM_PROT_WRITE) == 0) {
1111 		*errorp = KERN_PROTECTION_FAILURE;
1112 		unlock_and_deallocate(&fs);
1113 		return(NULL);
1114 	}
1115 
1116 	/*
1117 	 * On success vm_fault_object() does not unlock or deallocate, so we
1118 	 * do it here.  Note that the returned fs.m will be busied.
1119 	 */
1120 	unlock_things(&fs);
1121 
1122 	/*
1123 	 * Return a held page.  We are not doing any pmap manipulation so do
1124 	 * not set PG_MAPPED.  However, adjust the page flags according to
1125 	 * the fault type because the caller may not use a managed pmapping
1126 	 * (so we don't want to lose the fact that the page will be dirtied
1127 	 * if a write fault was specified).
1128 	 */
1129 	vm_page_hold(fs.m);
1130 	vm_page_activate(fs.m);
1131 	if ((fault_type & VM_PROT_WRITE) || (fault_flags & VM_FAULT_DIRTY))
1132 		vm_page_dirty(fs.m);
1133 	if (fault_flags & VM_FAULT_UNSWAP)
1134 		swap_pager_unswapped(fs.m);
1135 
1136 	/*
1137 	 * Indicate that the page was accessed.
1138 	 */
1139 	vm_page_flag_set(fs.m, PG_REFERENCED);
1140 
1141 	if (curthread->td_lwp) {
1142 		if (fs.hardfault) {
1143 			curthread->td_lwp->lwp_ru.ru_majflt++;
1144 		} else {
1145 			curthread->td_lwp->lwp_ru.ru_minflt++;
1146 		}
1147 	}
1148 
1149 	/*
1150 	 * Unlock everything, and return the held page.
1151 	 */
1152 	vm_page_wakeup(fs.m);
1153 	/*vm_object_deallocate(fs.first_object);*/
1154 	/*fs.first_object = NULL; */
1155 
1156 	*errorp = 0;
1157 	return(fs.m);
1158 }
1159 
1160 /*
1161  * Translate the virtual page number (first_pindex) that is relative
1162  * to the address space into a logical page number that is relative to the
1163  * backing object.  Use the virtual page table pointed to by (vpte).
1164  *
1165  * This implements an N-level page table.  Any level can terminate the
1166  * scan by setting VPTE_PS.   A linear mapping is accomplished by setting
1167  * VPTE_PS in the master page directory entry set via mcontrol(MADV_SETMAP).
1168  */
1169 static
1170 int
1171 vm_fault_vpagetable(struct faultstate *fs, vm_pindex_t *pindex,
1172 		    vpte_t vpte, int fault_type, int allow_nofault)
1173 {
1174 	struct lwbuf *lwb;
1175 	struct lwbuf lwb_cache;
1176 	int vshift = VPTE_FRAME_END - PAGE_SHIFT; /* index bits remaining */
1177 	int result = KERN_SUCCESS;
1178 	vpte_t *ptep;
1179 
1180 	ASSERT_LWKT_TOKEN_HELD(vm_object_token(fs->first_object));
1181 	for (;;) {
1182 		/*
1183 		 * We cannot proceed if the vpte is not valid, not readable
1184 		 * for a read fault, or not writable for a write fault.
1185 		 */
1186 		if ((vpte & VPTE_V) == 0) {
1187 			unlock_and_deallocate(fs);
1188 			return (KERN_FAILURE);
1189 		}
1190 		if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_RW) == 0) {
1191 			unlock_and_deallocate(fs);
1192 			return (KERN_FAILURE);
1193 		}
1194 		if ((vpte & VPTE_PS) || vshift == 0)
1195 			break;
1196 		KKASSERT(vshift >= VPTE_PAGE_BITS);
1197 
1198 		/*
1199 		 * Get the page table page.  Nominally we only read the page
1200 		 * table, but since we are actively setting VPTE_M and VPTE_A,
1201 		 * tell vm_fault_object() that we are writing it.
1202 		 *
1203 		 * There is currently no real need to optimize this.
1204 		 */
1205 		result = vm_fault_object(fs, (vpte & VPTE_FRAME) >> PAGE_SHIFT,
1206 					 VM_PROT_READ|VM_PROT_WRITE,
1207 					 allow_nofault);
1208 		if (result != KERN_SUCCESS)
1209 			return (result);
1210 
1211 		/*
1212 		 * Process the returned fs.m and look up the page table
1213 		 * entry in the page table page.
1214 		 */
1215 		vshift -= VPTE_PAGE_BITS;
1216 		lwb = lwbuf_alloc(fs->m, &lwb_cache);
1217 		ptep = ((vpte_t *)lwbuf_kva(lwb) +
1218 		        ((*pindex >> vshift) & VPTE_PAGE_MASK));
1219 		vpte = *ptep;
1220 
1221 		/*
1222 		 * Page table write-back.  If the vpte is valid for the
1223 		 * requested operation, do a write-back to the page table.
1224 		 *
1225 		 * XXX VPTE_M is not set properly for page directory pages.
1226 		 * It doesn't get set in the page directory if the page table
1227 		 * is modified during a read access.
1228 		 */
1229 		vm_page_activate(fs->m);
1230 		if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_V) &&
1231 		    (vpte & VPTE_RW)) {
1232 			if ((vpte & (VPTE_M|VPTE_A)) != (VPTE_M|VPTE_A)) {
1233 				atomic_set_long(ptep, VPTE_M | VPTE_A);
1234 				vm_page_dirty(fs->m);
1235 			}
1236 		}
1237 		if ((fault_type & VM_PROT_READ) && (vpte & VPTE_V)) {
1238 			if ((vpte & VPTE_A) == 0) {
1239 				atomic_set_long(ptep, VPTE_A);
1240 				vm_page_dirty(fs->m);
1241 			}
1242 		}
1243 		lwbuf_free(lwb);
1244 		vm_page_flag_set(fs->m, PG_REFERENCED);
1245 		vm_page_wakeup(fs->m);
1246 		fs->m = NULL;
1247 		cleanup_successful_fault(fs);
1248 	}
1249 	/*
1250 	 * Combine remaining address bits with the vpte.
1251 	 */
1252 	/* JG how many bits from each? */
1253 	*pindex = ((vpte & VPTE_FRAME) >> PAGE_SHIFT) +
1254 		  (*pindex & ((1L << vshift) - 1));
1255 	return (KERN_SUCCESS);
1256 }
1257 
1258 
1259 /*
1260  * This is the core of the vm_fault code.
1261  *
1262  * Do all operations required to fault-in (fs.first_object, pindex).  Run
1263  * through the shadow chain as necessary and do required COW or virtual
1264  * copy operations.  The caller has already fully resolved the vm_map_entry
1265  * and, if appropriate, has created a copy-on-write layer.  All we need to
1266  * do is iterate the object chain.
1267  *
1268  * On failure (fs) is unlocked and deallocated and the caller may return or
1269  * retry depending on the failure code.  On success (fs) is NOT unlocked or
1270  * deallocated, fs.m will contained a resolved, busied page, and fs.object
1271  * will have an additional PIP count if it is not equal to fs.first_object.
1272  *
1273  * If locks based on fs->first_shared or fs->shared are insufficient,
1274  * clear the appropriate field(s) and return RETRY.  COWs require that
1275  * first_shared be 0, while page allocations (or frees) require that
1276  * shared be 0.  Renames require that both be 0.
1277  *
1278  * fs->first_object must be held on call.
1279  */
1280 static
1281 int
1282 vm_fault_object(struct faultstate *fs, vm_pindex_t first_pindex,
1283 		vm_prot_t fault_type, int allow_nofault)
1284 {
1285 	vm_object_t next_object;
1286 	vm_pindex_t pindex;
1287 	int error;
1288 
1289 	ASSERT_LWKT_TOKEN_HELD(vm_object_token(fs->first_object));
1290 	fs->prot = fs->first_prot;
1291 	fs->object = fs->first_object;
1292 	pindex = first_pindex;
1293 
1294 	vm_object_chain_acquire(fs->first_object, fs->shared);
1295 	vm_object_pip_add(fs->first_object, 1);
1296 
1297 	/*
1298 	 * If a read fault occurs we try to make the page writable if
1299 	 * possible.  There are three cases where we cannot make the
1300 	 * page mapping writable:
1301 	 *
1302 	 * (1) The mapping is read-only or the VM object is read-only,
1303 	 *     fs->prot above will simply not have VM_PROT_WRITE set.
1304 	 *
1305 	 * (2) If the mapping is a virtual page table we need to be able
1306 	 *     to detect writes so we can set VPTE_M in the virtual page
1307 	 *     table.
1308 	 *
1309 	 * (3) If the VM page is read-only or copy-on-write, upgrading would
1310 	 *     just result in an unnecessary COW fault.
1311 	 *
1312 	 * VM_PROT_VPAGED is set if faulting via a virtual page table and
1313 	 * causes adjustments to the 'M'odify bit to also turn off write
1314 	 * access to force a re-fault.
1315 	 */
1316 	if (fs->entry->maptype == VM_MAPTYPE_VPAGETABLE) {
1317 		if ((fault_type & VM_PROT_WRITE) == 0)
1318 			fs->prot &= ~VM_PROT_WRITE;
1319 	}
1320 
1321 	if (curthread->td_lwp && curthread->td_lwp->lwp_vmspace &&
1322 	    pmap_emulate_ad_bits(&curthread->td_lwp->lwp_vmspace->vm_pmap)) {
1323 		if ((fault_type & VM_PROT_WRITE) == 0)
1324 			fs->prot &= ~VM_PROT_WRITE;
1325 	}
1326 
1327 	/* vm_object_hold(fs->object); implied b/c object == first_object */
1328 
1329 	for (;;) {
1330 		/*
1331 		 * The entire backing chain from first_object to object
1332 		 * inclusive is chainlocked.
1333 		 *
1334 		 * If the object is dead, we stop here
1335 		 */
1336 		if (fs->object->flags & OBJ_DEAD) {
1337 			vm_object_pip_wakeup(fs->first_object);
1338 			vm_object_chain_release_all(fs->first_object,
1339 						    fs->object);
1340 			if (fs->object != fs->first_object)
1341 				vm_object_drop(fs->object);
1342 			unlock_and_deallocate(fs);
1343 			return (KERN_PROTECTION_FAILURE);
1344 		}
1345 
1346 		/*
1347 		 * See if the page is resident.  Wait/Retry if the page is
1348 		 * busy (lots of stuff may have changed so we can't continue
1349 		 * in that case).
1350 		 *
1351 		 * We can theoretically allow the soft-busy case on a read
1352 		 * fault if the page is marked valid, but since such
1353 		 * pages are typically already pmap'd, putting that
1354 		 * special case in might be more effort then it is
1355 		 * worth.  We cannot under any circumstances mess
1356 		 * around with a vm_page_t->busy page except, perhaps,
1357 		 * to pmap it.
1358 		 */
1359 		fs->m = vm_page_lookup_busy_try(fs->object, pindex,
1360 						TRUE, &error);
1361 		if (error) {
1362 			vm_object_pip_wakeup(fs->first_object);
1363 			vm_object_chain_release_all(fs->first_object,
1364 						    fs->object);
1365 			if (fs->object != fs->first_object)
1366 				vm_object_drop(fs->object);
1367 			unlock_things(fs);
1368 			vm_page_sleep_busy(fs->m, TRUE, "vmpfw");
1369 			mycpu->gd_cnt.v_intrans++;
1370 			/*vm_object_deallocate(fs->first_object);*/
1371 			/*fs->first_object = NULL;*/
1372 			fs->m = NULL;
1373 			return (KERN_TRY_AGAIN);
1374 		}
1375 		if (fs->m) {
1376 			/*
1377 			 * The page is busied for us.
1378 			 *
1379 			 * If reactivating a page from PQ_CACHE we may have
1380 			 * to rate-limit.
1381 			 */
1382 			int queue = fs->m->queue;
1383 			vm_page_unqueue_nowakeup(fs->m);
1384 
1385 			if ((queue - fs->m->pc) == PQ_CACHE &&
1386 			    vm_page_count_severe()) {
1387 				vm_page_activate(fs->m);
1388 				vm_page_wakeup(fs->m);
1389 				fs->m = NULL;
1390 				vm_object_pip_wakeup(fs->first_object);
1391 				vm_object_chain_release_all(fs->first_object,
1392 							    fs->object);
1393 				if (fs->object != fs->first_object)
1394 					vm_object_drop(fs->object);
1395 				unlock_and_deallocate(fs);
1396 				if (allow_nofault == 0 ||
1397 				    (curthread->td_flags & TDF_NOFAULT) == 0) {
1398 					vm_wait_pfault();
1399 				}
1400 				return (KERN_TRY_AGAIN);
1401 			}
1402 
1403 			/*
1404 			 * If it still isn't completely valid (readable),
1405 			 * or if a read-ahead-mark is set on the VM page,
1406 			 * jump to readrest, else we found the page and
1407 			 * can return.
1408 			 *
1409 			 * We can release the spl once we have marked the
1410 			 * page busy.
1411 			 */
1412 			if (fs->m->object != &kernel_object) {
1413 				if ((fs->m->valid & VM_PAGE_BITS_ALL) !=
1414 				    VM_PAGE_BITS_ALL) {
1415 					goto readrest;
1416 				}
1417 				if (fs->m->flags & PG_RAM) {
1418 					if (debug_cluster)
1419 						kprintf("R");
1420 					vm_page_flag_clear(fs->m, PG_RAM);
1421 					goto readrest;
1422 				}
1423 			}
1424 			break; /* break to PAGE HAS BEEN FOUND */
1425 		}
1426 
1427 		/*
1428 		 * Page is not resident, If this is the search termination
1429 		 * or the pager might contain the page, allocate a new page.
1430 		 */
1431 		if (TRYPAGER(fs) || fs->object == fs->first_object) {
1432 			/*
1433 			 * Allocating, must be exclusive.
1434 			 */
1435 			if (fs->object == fs->first_object &&
1436 			    fs->first_shared) {
1437 				fs->first_shared = 0;
1438 				vm_object_pip_wakeup(fs->first_object);
1439 				vm_object_chain_release_all(fs->first_object,
1440 							    fs->object);
1441 				if (fs->object != fs->first_object)
1442 					vm_object_drop(fs->object);
1443 				unlock_and_deallocate(fs);
1444 				return (KERN_TRY_AGAIN);
1445 			}
1446 			if (fs->object != fs->first_object &&
1447 			    fs->shared) {
1448 				fs->first_shared = 0;
1449 				fs->shared = 0;
1450 				vm_object_pip_wakeup(fs->first_object);
1451 				vm_object_chain_release_all(fs->first_object,
1452 							    fs->object);
1453 				if (fs->object != fs->first_object)
1454 					vm_object_drop(fs->object);
1455 				unlock_and_deallocate(fs);
1456 				return (KERN_TRY_AGAIN);
1457 			}
1458 
1459 			/*
1460 			 * If the page is beyond the object size we fail
1461 			 */
1462 			if (pindex >= fs->object->size) {
1463 				vm_object_pip_wakeup(fs->first_object);
1464 				vm_object_chain_release_all(fs->first_object,
1465 							    fs->object);
1466 				if (fs->object != fs->first_object)
1467 					vm_object_drop(fs->object);
1468 				unlock_and_deallocate(fs);
1469 				return (KERN_PROTECTION_FAILURE);
1470 			}
1471 
1472 			/*
1473 			 * Allocate a new page for this object/offset pair.
1474 			 *
1475 			 * It is possible for the allocation to race, so
1476 			 * handle the case.
1477 			 */
1478 			fs->m = NULL;
1479 			if (!vm_page_count_severe()) {
1480 				fs->m = vm_page_alloc(fs->object, pindex,
1481 				    ((fs->vp || fs->object->backing_object) ?
1482 					VM_ALLOC_NULL_OK | VM_ALLOC_NORMAL :
1483 					VM_ALLOC_NULL_OK | VM_ALLOC_NORMAL |
1484 					VM_ALLOC_USE_GD | VM_ALLOC_ZERO));
1485 			}
1486 			if (fs->m == NULL) {
1487 				vm_object_pip_wakeup(fs->first_object);
1488 				vm_object_chain_release_all(fs->first_object,
1489 							    fs->object);
1490 				if (fs->object != fs->first_object)
1491 					vm_object_drop(fs->object);
1492 				unlock_and_deallocate(fs);
1493 				if (allow_nofault == 0 ||
1494 				    (curthread->td_flags & TDF_NOFAULT) == 0) {
1495 					vm_wait_pfault();
1496 				}
1497 				return (KERN_TRY_AGAIN);
1498 			}
1499 
1500 			/*
1501 			 * Fall through to readrest.  We have a new page which
1502 			 * will have to be paged (since m->valid will be 0).
1503 			 */
1504 		}
1505 
1506 readrest:
1507 		/*
1508 		 * We have found an invalid or partially valid page, a
1509 		 * page with a read-ahead mark which might be partially or
1510 		 * fully valid (and maybe dirty too), or we have allocated
1511 		 * a new page.
1512 		 *
1513 		 * Attempt to fault-in the page if there is a chance that the
1514 		 * pager has it, and potentially fault in additional pages
1515 		 * at the same time.
1516 		 *
1517 		 * If TRYPAGER is true then fs.m will be non-NULL and busied
1518 		 * for us.
1519 		 */
1520 		if (TRYPAGER(fs)) {
1521 			int rv;
1522 			int seqaccess;
1523 			u_char behavior = vm_map_entry_behavior(fs->entry);
1524 
1525 			if (behavior == MAP_ENTRY_BEHAV_RANDOM)
1526 				seqaccess = 0;
1527 			else
1528 				seqaccess = -1;
1529 
1530 			/*
1531 			 * Doing I/O may synchronously insert additional
1532 			 * pages so we can't be shared at this point either.
1533 			 *
1534 			 * NOTE: We can't free fs->m here in the allocated
1535 			 *	 case (fs->object != fs->first_object) as
1536 			 *	 this would require an exclusively locked
1537 			 *	 VM object.
1538 			 */
1539 			if (fs->object == fs->first_object &&
1540 			    fs->first_shared) {
1541 				vm_page_deactivate(fs->m);
1542 				vm_page_wakeup(fs->m);
1543 				fs->m = NULL;
1544 				fs->first_shared = 0;
1545 				vm_object_pip_wakeup(fs->first_object);
1546 				vm_object_chain_release_all(fs->first_object,
1547 							    fs->object);
1548 				if (fs->object != fs->first_object)
1549 					vm_object_drop(fs->object);
1550 				unlock_and_deallocate(fs);
1551 				return (KERN_TRY_AGAIN);
1552 			}
1553 			if (fs->object != fs->first_object &&
1554 			    fs->shared) {
1555 				vm_page_deactivate(fs->m);
1556 				vm_page_wakeup(fs->m);
1557 				fs->m = NULL;
1558 				fs->first_shared = 0;
1559 				fs->shared = 0;
1560 				vm_object_pip_wakeup(fs->first_object);
1561 				vm_object_chain_release_all(fs->first_object,
1562 							    fs->object);
1563 				if (fs->object != fs->first_object)
1564 					vm_object_drop(fs->object);
1565 				unlock_and_deallocate(fs);
1566 				return (KERN_TRY_AGAIN);
1567 			}
1568 
1569 			/*
1570 			 * Avoid deadlocking against the map when doing I/O.
1571 			 * fs.object and the page is PG_BUSY'd.
1572 			 *
1573 			 * NOTE: Once unlocked, fs->entry can become stale
1574 			 *	 so this will NULL it out.
1575 			 *
1576 			 * NOTE: fs->entry is invalid until we relock the
1577 			 *	 map and verify that the timestamp has not
1578 			 *	 changed.
1579 			 */
1580 			unlock_map(fs);
1581 
1582 			/*
1583 			 * Acquire the page data.  We still hold a ref on
1584 			 * fs.object and the page has been PG_BUSY's.
1585 			 *
1586 			 * The pager may replace the page (for example, in
1587 			 * order to enter a fictitious page into the
1588 			 * object).  If it does so it is responsible for
1589 			 * cleaning up the passed page and properly setting
1590 			 * the new page PG_BUSY.
1591 			 *
1592 			 * If we got here through a PG_RAM read-ahead
1593 			 * mark the page may be partially dirty and thus
1594 			 * not freeable.  Don't bother checking to see
1595 			 * if the pager has the page because we can't free
1596 			 * it anyway.  We have to depend on the get_page
1597 			 * operation filling in any gaps whether there is
1598 			 * backing store or not.
1599 			 */
1600 			rv = vm_pager_get_page(fs->object, &fs->m, seqaccess);
1601 
1602 			if (rv == VM_PAGER_OK) {
1603 				/*
1604 				 * Relookup in case pager changed page. Pager
1605 				 * is responsible for disposition of old page
1606 				 * if moved.
1607 				 *
1608 				 * XXX other code segments do relookups too.
1609 				 * It's a bad abstraction that needs to be
1610 				 * fixed/removed.
1611 				 */
1612 				fs->m = vm_page_lookup(fs->object, pindex);
1613 				if (fs->m == NULL) {
1614 					vm_object_pip_wakeup(fs->first_object);
1615 					vm_object_chain_release_all(
1616 						fs->first_object, fs->object);
1617 					if (fs->object != fs->first_object)
1618 						vm_object_drop(fs->object);
1619 					unlock_and_deallocate(fs);
1620 					return (KERN_TRY_AGAIN);
1621 				}
1622 				++fs->hardfault;
1623 				break; /* break to PAGE HAS BEEN FOUND */
1624 			}
1625 
1626 			/*
1627 			 * Remove the bogus page (which does not exist at this
1628 			 * object/offset); before doing so, we must get back
1629 			 * our object lock to preserve our invariant.
1630 			 *
1631 			 * Also wake up any other process that may want to bring
1632 			 * in this page.
1633 			 *
1634 			 * If this is the top-level object, we must leave the
1635 			 * busy page to prevent another process from rushing
1636 			 * past us, and inserting the page in that object at
1637 			 * the same time that we are.
1638 			 */
1639 			if (rv == VM_PAGER_ERROR) {
1640 				if (curproc) {
1641 					kprintf("vm_fault: pager read error, "
1642 						"pid %d (%s)\n",
1643 						curproc->p_pid,
1644 						curproc->p_comm);
1645 				} else {
1646 					kprintf("vm_fault: pager read error, "
1647 						"thread %p (%s)\n",
1648 						curthread,
1649 						curproc->p_comm);
1650 				}
1651 			}
1652 
1653 			/*
1654 			 * Data outside the range of the pager or an I/O error
1655 			 *
1656 			 * The page may have been wired during the pagein,
1657 			 * e.g. by the buffer cache, and cannot simply be
1658 			 * freed.  Call vnode_pager_freepage() to deal with it.
1659 			 *
1660 			 * Also note that we cannot free the page if we are
1661 			 * holding the related object shared. XXX not sure
1662 			 * what to do in that case.
1663 			 */
1664 			if (fs->object != fs->first_object) {
1665 				vnode_pager_freepage(fs->m);
1666 				fs->m = NULL;
1667 				/*
1668 				 * XXX - we cannot just fall out at this
1669 				 * point, m has been freed and is invalid!
1670 				 */
1671 			}
1672 			/*
1673 			 * XXX - the check for kernel_map is a kludge to work
1674 			 * around having the machine panic on a kernel space
1675 			 * fault w/ I/O error.
1676 			 */
1677 			if (((fs->map != &kernel_map) &&
1678 			    (rv == VM_PAGER_ERROR)) || (rv == VM_PAGER_BAD)) {
1679 				if (fs->m) {
1680 					if (fs->first_shared) {
1681 						vm_page_deactivate(fs->m);
1682 						vm_page_wakeup(fs->m);
1683 					} else {
1684 						vnode_pager_freepage(fs->m);
1685 					}
1686 					fs->m = NULL;
1687 				}
1688 				vm_object_pip_wakeup(fs->first_object);
1689 				vm_object_chain_release_all(fs->first_object,
1690 							    fs->object);
1691 				if (fs->object != fs->first_object)
1692 					vm_object_drop(fs->object);
1693 				unlock_and_deallocate(fs);
1694 				if (rv == VM_PAGER_ERROR)
1695 					return (KERN_FAILURE);
1696 				else
1697 					return (KERN_PROTECTION_FAILURE);
1698 				/* NOT REACHED */
1699 			}
1700 		}
1701 
1702 		/*
1703 		 * We get here if the object has a default pager (or unwiring)
1704 		 * or the pager doesn't have the page.
1705 		 *
1706 		 * fs->first_m will be used for the COW unless we find a
1707 		 * deeper page to be mapped read-only, in which case the
1708 		 * unlock*(fs) will free first_m.
1709 		 */
1710 		if (fs->object == fs->first_object)
1711 			fs->first_m = fs->m;
1712 
1713 		/*
1714 		 * Move on to the next object.  The chain lock should prevent
1715 		 * the backing_object from getting ripped out from under us.
1716 		 *
1717 		 * The object lock for the next object is governed by
1718 		 * fs->shared.
1719 		 */
1720 		if ((next_object = fs->object->backing_object) != NULL) {
1721 			if (fs->shared)
1722 				vm_object_hold_shared(next_object);
1723 			else
1724 				vm_object_hold(next_object);
1725 			vm_object_chain_acquire(next_object, fs->shared);
1726 			KKASSERT(next_object == fs->object->backing_object);
1727 			pindex += OFF_TO_IDX(fs->object->backing_object_offset);
1728 		}
1729 
1730 		if (next_object == NULL) {
1731 			/*
1732 			 * If there's no object left, fill the page in the top
1733 			 * object with zeros.
1734 			 */
1735 			if (fs->object != fs->first_object) {
1736 #if 0
1737 				if (fs->first_object->backing_object !=
1738 				    fs->object) {
1739 					vm_object_hold(fs->first_object->backing_object);
1740 				}
1741 #endif
1742 				vm_object_chain_release_all(
1743 					fs->first_object->backing_object,
1744 					fs->object);
1745 #if 0
1746 				if (fs->first_object->backing_object !=
1747 				    fs->object) {
1748 					vm_object_drop(fs->first_object->backing_object);
1749 				}
1750 #endif
1751 				vm_object_pip_wakeup(fs->object);
1752 				vm_object_drop(fs->object);
1753 				fs->object = fs->first_object;
1754 				pindex = first_pindex;
1755 				fs->m = fs->first_m;
1756 			}
1757 			fs->first_m = NULL;
1758 
1759 			/*
1760 			 * Zero the page and mark it valid.
1761 			 */
1762 			vm_page_zero_fill(fs->m);
1763 			mycpu->gd_cnt.v_zfod++;
1764 			fs->m->valid = VM_PAGE_BITS_ALL;
1765 			break;	/* break to PAGE HAS BEEN FOUND */
1766 		}
1767 		if (fs->object != fs->first_object) {
1768 			vm_object_pip_wakeup(fs->object);
1769 			vm_object_lock_swap();
1770 			vm_object_drop(fs->object);
1771 		}
1772 		KASSERT(fs->object != next_object,
1773 			("object loop %p", next_object));
1774 		fs->object = next_object;
1775 		vm_object_pip_add(fs->object, 1);
1776 	}
1777 
1778 	/*
1779 	 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
1780 	 * is held.]
1781 	 *
1782 	 * object still held.
1783 	 *
1784 	 * local shared variable may be different from fs->shared.
1785 	 *
1786 	 * If the page is being written, but isn't already owned by the
1787 	 * top-level object, we have to copy it into a new page owned by the
1788 	 * top-level object.
1789 	 */
1790 	KASSERT((fs->m->flags & PG_BUSY) != 0,
1791 		("vm_fault: not busy after main loop"));
1792 
1793 	if (fs->object != fs->first_object) {
1794 		/*
1795 		 * We only really need to copy if we want to write it.
1796 		 */
1797 		if (fault_type & VM_PROT_WRITE) {
1798 			/*
1799 			 * This allows pages to be virtually copied from a
1800 			 * backing_object into the first_object, where the
1801 			 * backing object has no other refs to it, and cannot
1802 			 * gain any more refs.  Instead of a bcopy, we just
1803 			 * move the page from the backing object to the
1804 			 * first object.  Note that we must mark the page
1805 			 * dirty in the first object so that it will go out
1806 			 * to swap when needed.
1807 			 */
1808 			if (
1809 				/*
1810 				 * Must be holding exclusive locks
1811 				 */
1812 				fs->first_shared == 0 &&
1813 				fs->shared == 0 &&
1814 				/*
1815 				 * Map, if present, has not changed
1816 				 */
1817 				(fs->map == NULL ||
1818 				fs->map_generation == fs->map->timestamp) &&
1819 				/*
1820 				 * Only one shadow object
1821 				 */
1822 				(fs->object->shadow_count == 1) &&
1823 				/*
1824 				 * No COW refs, except us
1825 				 */
1826 				(fs->object->ref_count == 1) &&
1827 				/*
1828 				 * No one else can look this object up
1829 				 */
1830 				(fs->object->handle == NULL) &&
1831 				/*
1832 				 * No other ways to look the object up
1833 				 */
1834 				((fs->object->type == OBJT_DEFAULT) ||
1835 				 (fs->object->type == OBJT_SWAP)) &&
1836 				/*
1837 				 * We don't chase down the shadow chain
1838 				 */
1839 				(fs->object == fs->first_object->backing_object) &&
1840 
1841 				/*
1842 				 * grab the lock if we need to
1843 				 */
1844 				(fs->lookup_still_valid ||
1845 				 fs->map == NULL ||
1846 				 lockmgr(&fs->map->lock, LK_EXCLUSIVE|LK_NOWAIT) == 0)
1847 			    ) {
1848 				/*
1849 				 * (first_m) and (m) are both busied.  We have
1850 				 * move (m) into (first_m)'s object/pindex
1851 				 * in an atomic fashion, then free (first_m).
1852 				 *
1853 				 * first_object is held so second remove
1854 				 * followed by the rename should wind
1855 				 * up being atomic.  vm_page_free() might
1856 				 * block so we don't do it until after the
1857 				 * rename.
1858 				 */
1859 				fs->lookup_still_valid = 1;
1860 				vm_page_protect(fs->first_m, VM_PROT_NONE);
1861 				vm_page_remove(fs->first_m);
1862 				vm_page_rename(fs->m, fs->first_object,
1863 					       first_pindex);
1864 				vm_page_free(fs->first_m);
1865 				fs->first_m = fs->m;
1866 				fs->m = NULL;
1867 				mycpu->gd_cnt.v_cow_optim++;
1868 			} else {
1869 				/*
1870 				 * Oh, well, lets copy it.
1871 				 *
1872 				 * Why are we unmapping the original page
1873 				 * here?  Well, in short, not all accessors
1874 				 * of user memory go through the pmap.  The
1875 				 * procfs code doesn't have access user memory
1876 				 * via a local pmap, so vm_fault_page*()
1877 				 * can't call pmap_enter().  And the umtx*()
1878 				 * code may modify the COW'd page via a DMAP
1879 				 * or kernel mapping and not via the pmap,
1880 				 * leaving the original page still mapped
1881 				 * read-only into the pmap.
1882 				 *
1883 				 * So we have to remove the page from at
1884 				 * least the current pmap if it is in it.
1885 				 * Just remove it from all pmaps.
1886 				 */
1887 				KKASSERT(fs->first_shared == 0);
1888 				vm_page_copy(fs->m, fs->first_m);
1889 				vm_page_protect(fs->m, VM_PROT_NONE);
1890 				vm_page_event(fs->m, VMEVENT_COW);
1891 			}
1892 
1893 			/*
1894 			 * We no longer need the old page or object.
1895 			 */
1896 			if (fs->m)
1897 				release_page(fs);
1898 
1899 			/*
1900 			 * We intend to revert to first_object, undo the
1901 			 * chain lock through to that.
1902 			 */
1903 #if 0
1904 			if (fs->first_object->backing_object != fs->object)
1905 				vm_object_hold(fs->first_object->backing_object);
1906 #endif
1907 			vm_object_chain_release_all(
1908 					fs->first_object->backing_object,
1909 					fs->object);
1910 #if 0
1911 			if (fs->first_object->backing_object != fs->object)
1912 				vm_object_drop(fs->first_object->backing_object);
1913 #endif
1914 
1915 			/*
1916 			 * fs->object != fs->first_object due to above
1917 			 * conditional
1918 			 */
1919 			vm_object_pip_wakeup(fs->object);
1920 			vm_object_drop(fs->object);
1921 
1922 			/*
1923 			 * Only use the new page below...
1924 			 */
1925 			mycpu->gd_cnt.v_cow_faults++;
1926 			fs->m = fs->first_m;
1927 			fs->object = fs->first_object;
1928 			pindex = first_pindex;
1929 		} else {
1930 			/*
1931 			 * If it wasn't a write fault avoid having to copy
1932 			 * the page by mapping it read-only.
1933 			 */
1934 			fs->prot &= ~VM_PROT_WRITE;
1935 		}
1936 	}
1937 
1938 	/*
1939 	 * Relock the map if necessary, then check the generation count.
1940 	 * relock_map() will update fs->timestamp to account for the
1941 	 * relocking if necessary.
1942 	 *
1943 	 * If the count has changed after relocking then all sorts of
1944 	 * crap may have happened and we have to retry.
1945 	 *
1946 	 * NOTE: The relock_map() can fail due to a deadlock against
1947 	 *	 the vm_page we are holding BUSY.
1948 	 */
1949 	if (fs->lookup_still_valid == FALSE && fs->map) {
1950 		if (relock_map(fs) ||
1951 		    fs->map->timestamp != fs->map_generation) {
1952 			release_page(fs);
1953 			vm_object_pip_wakeup(fs->first_object);
1954 			vm_object_chain_release_all(fs->first_object,
1955 						    fs->object);
1956 			if (fs->object != fs->first_object)
1957 				vm_object_drop(fs->object);
1958 			unlock_and_deallocate(fs);
1959 			return (KERN_TRY_AGAIN);
1960 		}
1961 	}
1962 
1963 	/*
1964 	 * If the fault is a write, we know that this page is being
1965 	 * written NOW so dirty it explicitly to save on pmap_is_modified()
1966 	 * calls later.
1967 	 *
1968 	 * If this is a NOSYNC mmap we do not want to set PG_NOSYNC
1969 	 * if the page is already dirty to prevent data written with
1970 	 * the expectation of being synced from not being synced.
1971 	 * Likewise if this entry does not request NOSYNC then make
1972 	 * sure the page isn't marked NOSYNC.  Applications sharing
1973 	 * data should use the same flags to avoid ping ponging.
1974 	 *
1975 	 * Also tell the backing pager, if any, that it should remove
1976 	 * any swap backing since the page is now dirty.
1977 	 */
1978 	vm_page_activate(fs->m);
1979 	if (fs->prot & VM_PROT_WRITE) {
1980 		vm_object_set_writeable_dirty(fs->m->object);
1981 		vm_set_nosync(fs->m, fs->entry);
1982 		if (fs->fault_flags & VM_FAULT_DIRTY) {
1983 			vm_page_dirty(fs->m);
1984 			swap_pager_unswapped(fs->m);
1985 		}
1986 	}
1987 
1988 	vm_object_pip_wakeup(fs->first_object);
1989 	vm_object_chain_release_all(fs->first_object, fs->object);
1990 	if (fs->object != fs->first_object)
1991 		vm_object_drop(fs->object);
1992 
1993 	/*
1994 	 * Page had better still be busy.  We are still locked up and
1995 	 * fs->object will have another PIP reference if it is not equal
1996 	 * to fs->first_object.
1997 	 */
1998 	KASSERT(fs->m->flags & PG_BUSY,
1999 		("vm_fault: page %p not busy!", fs->m));
2000 
2001 	/*
2002 	 * Sanity check: page must be completely valid or it is not fit to
2003 	 * map into user space.  vm_pager_get_pages() ensures this.
2004 	 */
2005 	if (fs->m->valid != VM_PAGE_BITS_ALL) {
2006 		vm_page_zero_invalid(fs->m, TRUE);
2007 		kprintf("Warning: page %p partially invalid on fault\n", fs->m);
2008 	}
2009 
2010 	return (KERN_SUCCESS);
2011 }
2012 
2013 /*
2014  * Hold each of the physical pages that are mapped by the specified range of
2015  * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
2016  * and allow the specified types of access, "prot".  If all of the implied
2017  * pages are successfully held, then the number of held pages is returned
2018  * together with pointers to those pages in the array "ma".  However, if any
2019  * of the pages cannot be held, -1 is returned.
2020  */
2021 int
2022 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
2023     vm_prot_t prot, vm_page_t *ma, int max_count)
2024 {
2025 	vm_offset_t start, end;
2026 	int i, npages, error;
2027 
2028 	start = trunc_page(addr);
2029 	end = round_page(addr + len);
2030 
2031 	npages = howmany(end - start, PAGE_SIZE);
2032 
2033 	if (npages > max_count)
2034 		return -1;
2035 
2036 	for (i = 0; i < npages; i++) {
2037 		// XXX error handling
2038 		ma[i] = vm_fault_page_quick(start + (i * PAGE_SIZE),
2039 			prot,
2040 			&error);
2041 	}
2042 
2043 	return npages;
2044 }
2045 
2046 /*
2047  * Wire down a range of virtual addresses in a map.  The entry in question
2048  * should be marked in-transition and the map must be locked.  We must
2049  * release the map temporarily while faulting-in the page to avoid a
2050  * deadlock.  Note that the entry may be clipped while we are blocked but
2051  * will never be freed.
2052  *
2053  * No requirements.
2054  */
2055 int
2056 vm_fault_wire(vm_map_t map, vm_map_entry_t entry,
2057 	      boolean_t user_wire, int kmflags)
2058 {
2059 	boolean_t fictitious;
2060 	vm_offset_t start;
2061 	vm_offset_t end;
2062 	vm_offset_t va;
2063 	vm_paddr_t pa;
2064 	vm_page_t m;
2065 	pmap_t pmap;
2066 	int rv;
2067 	int wire_prot;
2068 	int fault_flags;
2069 
2070 	lwkt_gettoken(&map->token);
2071 
2072 	if (user_wire) {
2073 		wire_prot = VM_PROT_READ;
2074 		fault_flags = VM_FAULT_USER_WIRE;
2075 	} else {
2076 		wire_prot = VM_PROT_READ | VM_PROT_WRITE;
2077 		fault_flags = VM_FAULT_CHANGE_WIRING;
2078 	}
2079 	if (kmflags & KM_NOTLBSYNC)
2080 		wire_prot |= VM_PROT_NOSYNC;
2081 
2082 	pmap = vm_map_pmap(map);
2083 	start = entry->start;
2084 	end = entry->end;
2085 	switch(entry->maptype) {
2086 	case VM_MAPTYPE_NORMAL:
2087 	case VM_MAPTYPE_VPAGETABLE:
2088 		fictitious = entry->object.vm_object &&
2089 			    ((entry->object.vm_object->type == OBJT_DEVICE) ||
2090 			     (entry->object.vm_object->type == OBJT_MGTDEVICE));
2091 		break;
2092 	case VM_MAPTYPE_UKSMAP:
2093 		fictitious = TRUE;
2094 		break;
2095 	default:
2096 		fictitious = FALSE;
2097 		break;
2098 	}
2099 
2100 	if (entry->eflags & MAP_ENTRY_KSTACK)
2101 		start += PAGE_SIZE;
2102 	map->timestamp++;
2103 	vm_map_unlock(map);
2104 
2105 	/*
2106 	 * We simulate a fault to get the page and enter it in the physical
2107 	 * map.
2108 	 */
2109 	for (va = start; va < end; va += PAGE_SIZE) {
2110 		rv = vm_fault(map, va, wire_prot, fault_flags);
2111 		if (rv) {
2112 			while (va > start) {
2113 				va -= PAGE_SIZE;
2114 				if ((pa = pmap_extract(pmap, va)) == 0)
2115 					continue;
2116 				pmap_change_wiring(pmap, va, FALSE, entry);
2117 				if (!fictitious) {
2118 					m = PHYS_TO_VM_PAGE(pa);
2119 					vm_page_busy_wait(m, FALSE, "vmwrpg");
2120 					vm_page_unwire(m, 1);
2121 					vm_page_wakeup(m);
2122 				}
2123 			}
2124 			goto done;
2125 		}
2126 	}
2127 	rv = KERN_SUCCESS;
2128 done:
2129 	vm_map_lock(map);
2130 	lwkt_reltoken(&map->token);
2131 	return (rv);
2132 }
2133 
2134 /*
2135  * Unwire a range of virtual addresses in a map.  The map should be
2136  * locked.
2137  */
2138 void
2139 vm_fault_unwire(vm_map_t map, vm_map_entry_t entry)
2140 {
2141 	boolean_t fictitious;
2142 	vm_offset_t start;
2143 	vm_offset_t end;
2144 	vm_offset_t va;
2145 	vm_paddr_t pa;
2146 	vm_page_t m;
2147 	pmap_t pmap;
2148 
2149 	lwkt_gettoken(&map->token);
2150 
2151 	pmap = vm_map_pmap(map);
2152 	start = entry->start;
2153 	end = entry->end;
2154 	fictitious = entry->object.vm_object &&
2155 			((entry->object.vm_object->type == OBJT_DEVICE) ||
2156 			 (entry->object.vm_object->type == OBJT_MGTDEVICE));
2157 	if (entry->eflags & MAP_ENTRY_KSTACK)
2158 		start += PAGE_SIZE;
2159 
2160 	/*
2161 	 * Since the pages are wired down, we must be able to get their
2162 	 * mappings from the physical map system.
2163 	 */
2164 	for (va = start; va < end; va += PAGE_SIZE) {
2165 		pa = pmap_extract(pmap, va);
2166 		if (pa != 0) {
2167 			pmap_change_wiring(pmap, va, FALSE, entry);
2168 			if (!fictitious) {
2169 				m = PHYS_TO_VM_PAGE(pa);
2170 				vm_page_busy_wait(m, FALSE, "vmwupg");
2171 				vm_page_unwire(m, 1);
2172 				vm_page_wakeup(m);
2173 			}
2174 		}
2175 	}
2176 	lwkt_reltoken(&map->token);
2177 }
2178 
2179 /*
2180  * Copy all of the pages from a wired-down map entry to another.
2181  *
2182  * The source and destination maps must be locked for write.
2183  * The source and destination maps token must be held
2184  * The source map entry must be wired down (or be a sharing map
2185  * entry corresponding to a main map entry that is wired down).
2186  *
2187  * No other requirements.
2188  *
2189  * XXX do segment optimization
2190  */
2191 void
2192 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
2193 		    vm_map_entry_t dst_entry, vm_map_entry_t src_entry)
2194 {
2195 	vm_object_t dst_object;
2196 	vm_object_t src_object;
2197 	vm_ooffset_t dst_offset;
2198 	vm_ooffset_t src_offset;
2199 	vm_prot_t prot;
2200 	vm_offset_t vaddr;
2201 	vm_page_t dst_m;
2202 	vm_page_t src_m;
2203 
2204 	src_object = src_entry->object.vm_object;
2205 	src_offset = src_entry->offset;
2206 
2207 	/*
2208 	 * Create the top-level object for the destination entry. (Doesn't
2209 	 * actually shadow anything - we copy the pages directly.)
2210 	 */
2211 	vm_map_entry_allocate_object(dst_entry);
2212 	dst_object = dst_entry->object.vm_object;
2213 
2214 	prot = dst_entry->max_protection;
2215 
2216 	/*
2217 	 * Loop through all of the pages in the entry's range, copying each
2218 	 * one from the source object (it should be there) to the destination
2219 	 * object.
2220 	 */
2221 	vm_object_hold(src_object);
2222 	vm_object_hold(dst_object);
2223 	for (vaddr = dst_entry->start, dst_offset = 0;
2224 	    vaddr < dst_entry->end;
2225 	    vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) {
2226 
2227 		/*
2228 		 * Allocate a page in the destination object
2229 		 */
2230 		do {
2231 			dst_m = vm_page_alloc(dst_object,
2232 					      OFF_TO_IDX(dst_offset),
2233 					      VM_ALLOC_NORMAL);
2234 			if (dst_m == NULL) {
2235 				vm_wait(0);
2236 			}
2237 		} while (dst_m == NULL);
2238 
2239 		/*
2240 		 * Find the page in the source object, and copy it in.
2241 		 * (Because the source is wired down, the page will be in
2242 		 * memory.)
2243 		 */
2244 		src_m = vm_page_lookup(src_object,
2245 				       OFF_TO_IDX(dst_offset + src_offset));
2246 		if (src_m == NULL)
2247 			panic("vm_fault_copy_wired: page missing");
2248 
2249 		vm_page_copy(src_m, dst_m);
2250 		vm_page_event(src_m, VMEVENT_COW);
2251 
2252 		/*
2253 		 * Enter it in the pmap...
2254 		 */
2255 		pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE, dst_entry);
2256 
2257 		/*
2258 		 * Mark it no longer busy, and put it on the active list.
2259 		 */
2260 		vm_page_activate(dst_m);
2261 		vm_page_wakeup(dst_m);
2262 	}
2263 	vm_object_drop(dst_object);
2264 	vm_object_drop(src_object);
2265 }
2266 
2267 #if 0
2268 
2269 /*
2270  * This routine checks around the requested page for other pages that
2271  * might be able to be faulted in.  This routine brackets the viable
2272  * pages for the pages to be paged in.
2273  *
2274  * Inputs:
2275  *	m, rbehind, rahead
2276  *
2277  * Outputs:
2278  *  marray (array of vm_page_t), reqpage (index of requested page)
2279  *
2280  * Return value:
2281  *  number of pages in marray
2282  */
2283 static int
2284 vm_fault_additional_pages(vm_page_t m, int rbehind, int rahead,
2285 			  vm_page_t *marray, int *reqpage)
2286 {
2287 	int i,j;
2288 	vm_object_t object;
2289 	vm_pindex_t pindex, startpindex, endpindex, tpindex;
2290 	vm_page_t rtm;
2291 	int cbehind, cahead;
2292 
2293 	object = m->object;
2294 	pindex = m->pindex;
2295 
2296 	/*
2297 	 * we don't fault-ahead for device pager
2298 	 */
2299 	if ((object->type == OBJT_DEVICE) ||
2300 	    (object->type == OBJT_MGTDEVICE)) {
2301 		*reqpage = 0;
2302 		marray[0] = m;
2303 		return 1;
2304 	}
2305 
2306 	/*
2307 	 * if the requested page is not available, then give up now
2308 	 */
2309 	if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
2310 		*reqpage = 0;	/* not used by caller, fix compiler warn */
2311 		return 0;
2312 	}
2313 
2314 	if ((cbehind == 0) && (cahead == 0)) {
2315 		*reqpage = 0;
2316 		marray[0] = m;
2317 		return 1;
2318 	}
2319 
2320 	if (rahead > cahead) {
2321 		rahead = cahead;
2322 	}
2323 
2324 	if (rbehind > cbehind) {
2325 		rbehind = cbehind;
2326 	}
2327 
2328 	/*
2329 	 * Do not do any readahead if we have insufficient free memory.
2330 	 *
2331 	 * XXX code was broken disabled before and has instability
2332 	 * with this conditonal fixed, so shortcut for now.
2333 	 */
2334 	if (burst_fault == 0 || vm_page_count_severe()) {
2335 		marray[0] = m;
2336 		*reqpage = 0;
2337 		return 1;
2338 	}
2339 
2340 	/*
2341 	 * scan backward for the read behind pages -- in memory
2342 	 *
2343 	 * Assume that if the page is not found an interrupt will not
2344 	 * create it.  Theoretically interrupts can only remove (busy)
2345 	 * pages, not create new associations.
2346 	 */
2347 	if (pindex > 0) {
2348 		if (rbehind > pindex) {
2349 			rbehind = pindex;
2350 			startpindex = 0;
2351 		} else {
2352 			startpindex = pindex - rbehind;
2353 		}
2354 
2355 		vm_object_hold(object);
2356 		for (tpindex = pindex; tpindex > startpindex; --tpindex) {
2357 			if (vm_page_lookup(object, tpindex - 1))
2358 				break;
2359 		}
2360 
2361 		i = 0;
2362 		while (tpindex < pindex) {
2363 			rtm = vm_page_alloc(object, tpindex, VM_ALLOC_SYSTEM |
2364 							     VM_ALLOC_NULL_OK);
2365 			if (rtm == NULL) {
2366 				for (j = 0; j < i; j++) {
2367 					vm_page_free(marray[j]);
2368 				}
2369 				vm_object_drop(object);
2370 				marray[0] = m;
2371 				*reqpage = 0;
2372 				return 1;
2373 			}
2374 			marray[i] = rtm;
2375 			++i;
2376 			++tpindex;
2377 		}
2378 		vm_object_drop(object);
2379 	} else {
2380 		i = 0;
2381 	}
2382 
2383 	/*
2384 	 * Assign requested page
2385 	 */
2386 	marray[i] = m;
2387 	*reqpage = i;
2388 	++i;
2389 
2390 	/*
2391 	 * Scan forwards for read-ahead pages
2392 	 */
2393 	tpindex = pindex + 1;
2394 	endpindex = tpindex + rahead;
2395 	if (endpindex > object->size)
2396 		endpindex = object->size;
2397 
2398 	vm_object_hold(object);
2399 	while (tpindex < endpindex) {
2400 		if (vm_page_lookup(object, tpindex))
2401 			break;
2402 		rtm = vm_page_alloc(object, tpindex, VM_ALLOC_SYSTEM |
2403 						     VM_ALLOC_NULL_OK);
2404 		if (rtm == NULL)
2405 			break;
2406 		marray[i] = rtm;
2407 		++i;
2408 		++tpindex;
2409 	}
2410 	vm_object_drop(object);
2411 
2412 	return (i);
2413 }
2414 
2415 #endif
2416 
2417 /*
2418  * vm_prefault() provides a quick way of clustering pagefaults into a
2419  * processes address space.  It is a "cousin" of pmap_object_init_pt,
2420  * except it runs at page fault time instead of mmap time.
2421  *
2422  * vm.fast_fault	Enables pre-faulting zero-fill pages
2423  *
2424  * vm.prefault_pages	Number of pages (1/2 negative, 1/2 positive) to
2425  *			prefault.  Scan stops in either direction when
2426  *			a page is found to already exist.
2427  *
2428  * This code used to be per-platform pmap_prefault().  It is now
2429  * machine-independent and enhanced to also pre-fault zero-fill pages
2430  * (see vm.fast_fault) as well as make them writable, which greatly
2431  * reduces the number of page faults programs incur.
2432  *
2433  * Application performance when pre-faulting zero-fill pages is heavily
2434  * dependent on the application.  Very tiny applications like /bin/echo
2435  * lose a little performance while applications of any appreciable size
2436  * gain performance.  Prefaulting multiple pages also reduces SMP
2437  * congestion and can improve SMP performance significantly.
2438  *
2439  * NOTE!  prot may allow writing but this only applies to the top level
2440  *	  object.  If we wind up mapping a page extracted from a backing
2441  *	  object we have to make sure it is read-only.
2442  *
2443  * NOTE!  The caller has already handled any COW operations on the
2444  *	  vm_map_entry via the normal fault code.  Do NOT call this
2445  *	  shortcut unless the normal fault code has run on this entry.
2446  *
2447  * The related map must be locked.
2448  * No other requirements.
2449  */
2450 static int vm_prefault_pages = 8;
2451 SYSCTL_INT(_vm, OID_AUTO, prefault_pages, CTLFLAG_RW, &vm_prefault_pages, 0,
2452 	   "Maximum number of pages to pre-fault");
2453 static int vm_fast_fault = 1;
2454 SYSCTL_INT(_vm, OID_AUTO, fast_fault, CTLFLAG_RW, &vm_fast_fault, 0,
2455 	   "Burst fault zero-fill regions");
2456 
2457 /*
2458  * Set PG_NOSYNC if the map entry indicates so, but only if the page
2459  * is not already dirty by other means.  This will prevent passive
2460  * filesystem syncing as well as 'sync' from writing out the page.
2461  */
2462 static void
2463 vm_set_nosync(vm_page_t m, vm_map_entry_t entry)
2464 {
2465 	if (entry->eflags & MAP_ENTRY_NOSYNC) {
2466 		if (m->dirty == 0)
2467 			vm_page_flag_set(m, PG_NOSYNC);
2468 	} else {
2469 		vm_page_flag_clear(m, PG_NOSYNC);
2470 	}
2471 }
2472 
2473 static void
2474 vm_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry, int prot,
2475 	    int fault_flags)
2476 {
2477 	struct lwp *lp;
2478 	vm_page_t m;
2479 	vm_offset_t addr;
2480 	vm_pindex_t index;
2481 	vm_pindex_t pindex;
2482 	vm_object_t object;
2483 	int pprot;
2484 	int i;
2485 	int noneg;
2486 	int nopos;
2487 	int maxpages;
2488 
2489 	/*
2490 	 * Get stable max count value, disabled if set to 0
2491 	 */
2492 	maxpages = vm_prefault_pages;
2493 	cpu_ccfence();
2494 	if (maxpages <= 0)
2495 		return;
2496 
2497 	/*
2498 	 * We do not currently prefault mappings that use virtual page
2499 	 * tables.  We do not prefault foreign pmaps.
2500 	 */
2501 	if (entry->maptype != VM_MAPTYPE_NORMAL)
2502 		return;
2503 	lp = curthread->td_lwp;
2504 	if (lp == NULL || (pmap != vmspace_pmap(lp->lwp_vmspace)))
2505 		return;
2506 
2507 	/*
2508 	 * Limit pre-fault count to 1024 pages.
2509 	 */
2510 	if (maxpages > 1024)
2511 		maxpages = 1024;
2512 
2513 	object = entry->object.vm_object;
2514 	KKASSERT(object != NULL);
2515 	KKASSERT(object == entry->object.vm_object);
2516 	vm_object_hold(object);
2517 	vm_object_chain_acquire(object, 0);
2518 
2519 	noneg = 0;
2520 	nopos = 0;
2521 	for (i = 0; i < maxpages; ++i) {
2522 		vm_object_t lobject;
2523 		vm_object_t nobject;
2524 		int allocated = 0;
2525 		int error;
2526 
2527 		/*
2528 		 * This can eat a lot of time on a heavily contended
2529 		 * machine so yield on the tick if needed.
2530 		 */
2531 		if ((i & 7) == 7)
2532 			lwkt_yield();
2533 
2534 		/*
2535 		 * Calculate the page to pre-fault, stopping the scan in
2536 		 * each direction separately if the limit is reached.
2537 		 */
2538 		if (i & 1) {
2539 			if (noneg)
2540 				continue;
2541 			addr = addra - ((i + 1) >> 1) * PAGE_SIZE;
2542 		} else {
2543 			if (nopos)
2544 				continue;
2545 			addr = addra + ((i + 2) >> 1) * PAGE_SIZE;
2546 		}
2547 		if (addr < entry->start) {
2548 			noneg = 1;
2549 			if (noneg && nopos)
2550 				break;
2551 			continue;
2552 		}
2553 		if (addr >= entry->end) {
2554 			nopos = 1;
2555 			if (noneg && nopos)
2556 				break;
2557 			continue;
2558 		}
2559 
2560 		/*
2561 		 * Skip pages already mapped, and stop scanning in that
2562 		 * direction.  When the scan terminates in both directions
2563 		 * we are done.
2564 		 */
2565 		if (pmap_prefault_ok(pmap, addr) == 0) {
2566 			if (i & 1)
2567 				noneg = 1;
2568 			else
2569 				nopos = 1;
2570 			if (noneg && nopos)
2571 				break;
2572 			continue;
2573 		}
2574 
2575 		/*
2576 		 * Follow the VM object chain to obtain the page to be mapped
2577 		 * into the pmap.
2578 		 *
2579 		 * If we reach the terminal object without finding a page
2580 		 * and we determine it would be advantageous, then allocate
2581 		 * a zero-fill page for the base object.  The base object
2582 		 * is guaranteed to be OBJT_DEFAULT for this case.
2583 		 *
2584 		 * In order to not have to check the pager via *haspage*()
2585 		 * we stop if any non-default object is encountered.  e.g.
2586 		 * a vnode or swap object would stop the loop.
2587 		 */
2588 		index = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
2589 		lobject = object;
2590 		pindex = index;
2591 		pprot = prot;
2592 
2593 		KKASSERT(lobject == entry->object.vm_object);
2594 		/*vm_object_hold(lobject); implied */
2595 
2596 		while ((m = vm_page_lookup_busy_try(lobject, pindex,
2597 						    TRUE, &error)) == NULL) {
2598 			if (lobject->type != OBJT_DEFAULT)
2599 				break;
2600 			if (lobject->backing_object == NULL) {
2601 				if (vm_fast_fault == 0)
2602 					break;
2603 				if ((prot & VM_PROT_WRITE) == 0 ||
2604 				    vm_page_count_min(0)) {
2605 					break;
2606 				}
2607 
2608 				/*
2609 				 * NOTE: Allocated from base object
2610 				 */
2611 				m = vm_page_alloc(object, index,
2612 						  VM_ALLOC_NORMAL |
2613 						  VM_ALLOC_ZERO |
2614 						  VM_ALLOC_USE_GD |
2615 						  VM_ALLOC_NULL_OK);
2616 				if (m == NULL)
2617 					break;
2618 				allocated = 1;
2619 				pprot = prot;
2620 				/* lobject = object .. not needed */
2621 				break;
2622 			}
2623 			if (lobject->backing_object_offset & PAGE_MASK)
2624 				break;
2625 			nobject = lobject->backing_object;
2626 			vm_object_hold(nobject);
2627 			KKASSERT(nobject == lobject->backing_object);
2628 			pindex += lobject->backing_object_offset >> PAGE_SHIFT;
2629 			if (lobject != object) {
2630 				vm_object_lock_swap();
2631 				vm_object_drop(lobject);
2632 			}
2633 			lobject = nobject;
2634 			pprot &= ~VM_PROT_WRITE;
2635 			vm_object_chain_acquire(lobject, 0);
2636 		}
2637 
2638 		/*
2639 		 * NOTE: A non-NULL (m) will be associated with lobject if
2640 		 *	 it was found there, otherwise it is probably a
2641 		 *	 zero-fill page associated with the base object.
2642 		 *
2643 		 * Give-up if no page is available.
2644 		 */
2645 		if (m == NULL) {
2646 			if (lobject != object) {
2647 #if 0
2648 				if (object->backing_object != lobject)
2649 					vm_object_hold(object->backing_object);
2650 #endif
2651 				vm_object_chain_release_all(
2652 					object->backing_object, lobject);
2653 #if 0
2654 				if (object->backing_object != lobject)
2655 					vm_object_drop(object->backing_object);
2656 #endif
2657 				vm_object_drop(lobject);
2658 			}
2659 			break;
2660 		}
2661 
2662 		/*
2663 		 * The object must be marked dirty if we are mapping a
2664 		 * writable page.  m->object is either lobject or object,
2665 		 * both of which are still held.  Do this before we
2666 		 * potentially drop the object.
2667 		 */
2668 		if (pprot & VM_PROT_WRITE)
2669 			vm_object_set_writeable_dirty(m->object);
2670 
2671 		/*
2672 		 * Do not conditionalize on PG_RAM.  If pages are present in
2673 		 * the VM system we assume optimal caching.  If caching is
2674 		 * not optimal the I/O gravy train will be restarted when we
2675 		 * hit an unavailable page.  We do not want to try to restart
2676 		 * the gravy train now because we really don't know how much
2677 		 * of the object has been cached.  The cost for restarting
2678 		 * the gravy train should be low (since accesses will likely
2679 		 * be I/O bound anyway).
2680 		 */
2681 		if (lobject != object) {
2682 #if 0
2683 			if (object->backing_object != lobject)
2684 				vm_object_hold(object->backing_object);
2685 #endif
2686 			vm_object_chain_release_all(object->backing_object,
2687 						    lobject);
2688 #if 0
2689 			if (object->backing_object != lobject)
2690 				vm_object_drop(object->backing_object);
2691 #endif
2692 			vm_object_drop(lobject);
2693 		}
2694 
2695 		/*
2696 		 * Enter the page into the pmap if appropriate.  If we had
2697 		 * allocated the page we have to place it on a queue.  If not
2698 		 * we just have to make sure it isn't on the cache queue
2699 		 * (pages on the cache queue are not allowed to be mapped).
2700 		 */
2701 		if (allocated) {
2702 			/*
2703 			 * Page must be zerod.
2704 			 */
2705 			vm_page_zero_fill(m);
2706 			mycpu->gd_cnt.v_zfod++;
2707 			m->valid = VM_PAGE_BITS_ALL;
2708 
2709 			/*
2710 			 * Handle dirty page case
2711 			 */
2712 			if (pprot & VM_PROT_WRITE)
2713 				vm_set_nosync(m, entry);
2714 			pmap_enter(pmap, addr, m, pprot, 0, entry);
2715 			mycpu->gd_cnt.v_vm_faults++;
2716 			if (curthread->td_lwp)
2717 				++curthread->td_lwp->lwp_ru.ru_minflt;
2718 			vm_page_deactivate(m);
2719 			if (pprot & VM_PROT_WRITE) {
2720 				/*vm_object_set_writeable_dirty(m->object);*/
2721 				vm_set_nosync(m, entry);
2722 				if (fault_flags & VM_FAULT_DIRTY) {
2723 					vm_page_dirty(m);
2724 					/*XXX*/
2725 					swap_pager_unswapped(m);
2726 				}
2727 			}
2728 			vm_page_wakeup(m);
2729 		} else if (error) {
2730 			/* couldn't busy page, no wakeup */
2731 		} else if (
2732 		    ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2733 		    (m->flags & PG_FICTITIOUS) == 0) {
2734 			/*
2735 			 * A fully valid page not undergoing soft I/O can
2736 			 * be immediately entered into the pmap.
2737 			 */
2738 			if ((m->queue - m->pc) == PQ_CACHE)
2739 				vm_page_deactivate(m);
2740 			if (pprot & VM_PROT_WRITE) {
2741 				/*vm_object_set_writeable_dirty(m->object);*/
2742 				vm_set_nosync(m, entry);
2743 				if (fault_flags & VM_FAULT_DIRTY) {
2744 					vm_page_dirty(m);
2745 					/*XXX*/
2746 					swap_pager_unswapped(m);
2747 				}
2748 			}
2749 			if (pprot & VM_PROT_WRITE)
2750 				vm_set_nosync(m, entry);
2751 			pmap_enter(pmap, addr, m, pprot, 0, entry);
2752 			mycpu->gd_cnt.v_vm_faults++;
2753 			if (curthread->td_lwp)
2754 				++curthread->td_lwp->lwp_ru.ru_minflt;
2755 			vm_page_wakeup(m);
2756 		} else {
2757 			vm_page_wakeup(m);
2758 		}
2759 	}
2760 	vm_object_chain_release(object);
2761 	vm_object_drop(object);
2762 }
2763 
2764 /*
2765  * Object can be held shared
2766  */
2767 static void
2768 vm_prefault_quick(pmap_t pmap, vm_offset_t addra,
2769 		  vm_map_entry_t entry, int prot, int fault_flags)
2770 {
2771 	struct lwp *lp;
2772 	vm_page_t m;
2773 	vm_offset_t addr;
2774 	vm_pindex_t pindex;
2775 	vm_object_t object;
2776 	int i;
2777 	int noneg;
2778 	int nopos;
2779 	int maxpages;
2780 
2781 	/*
2782 	 * Get stable max count value, disabled if set to 0
2783 	 */
2784 	maxpages = vm_prefault_pages;
2785 	cpu_ccfence();
2786 	if (maxpages <= 0)
2787 		return;
2788 
2789 	/*
2790 	 * We do not currently prefault mappings that use virtual page
2791 	 * tables.  We do not prefault foreign pmaps.
2792 	 */
2793 	if (entry->maptype != VM_MAPTYPE_NORMAL)
2794 		return;
2795 	lp = curthread->td_lwp;
2796 	if (lp == NULL || (pmap != vmspace_pmap(lp->lwp_vmspace)))
2797 		return;
2798 	object = entry->object.vm_object;
2799 	if (object->backing_object != NULL)
2800 		return;
2801 	ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2802 
2803 	/*
2804 	 * Limit pre-fault count to 1024 pages.
2805 	 */
2806 	if (maxpages > 1024)
2807 		maxpages = 1024;
2808 
2809 	noneg = 0;
2810 	nopos = 0;
2811 	for (i = 0; i < maxpages; ++i) {
2812 		int error;
2813 
2814 		/*
2815 		 * Calculate the page to pre-fault, stopping the scan in
2816 		 * each direction separately if the limit is reached.
2817 		 */
2818 		if (i & 1) {
2819 			if (noneg)
2820 				continue;
2821 			addr = addra - ((i + 1) >> 1) * PAGE_SIZE;
2822 		} else {
2823 			if (nopos)
2824 				continue;
2825 			addr = addra + ((i + 2) >> 1) * PAGE_SIZE;
2826 		}
2827 		if (addr < entry->start) {
2828 			noneg = 1;
2829 			if (noneg && nopos)
2830 				break;
2831 			continue;
2832 		}
2833 		if (addr >= entry->end) {
2834 			nopos = 1;
2835 			if (noneg && nopos)
2836 				break;
2837 			continue;
2838 		}
2839 
2840 		/*
2841 		 * Follow the VM object chain to obtain the page to be mapped
2842 		 * into the pmap.  This version of the prefault code only
2843 		 * works with terminal objects.
2844 		 *
2845 		 * The page must already exist.  If we encounter a problem
2846 		 * we stop here.
2847 		 *
2848 		 * WARNING!  We cannot call swap_pager_unswapped() or insert
2849 		 *	     a new vm_page with a shared token.
2850 		 */
2851 		pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
2852 
2853 		m = vm_page_lookup_busy_try(object, pindex, TRUE, &error);
2854 		if (m == NULL || error)
2855 			break;
2856 
2857 		/*
2858 		 * Skip pages already mapped, and stop scanning in that
2859 		 * direction.  When the scan terminates in both directions
2860 		 * we are done.
2861 		 */
2862 		if (pmap_prefault_ok(pmap, addr) == 0) {
2863 			vm_page_wakeup(m);
2864 			if (i & 1)
2865 				noneg = 1;
2866 			else
2867 				nopos = 1;
2868 			if (noneg && nopos)
2869 				break;
2870 			continue;
2871 		}
2872 
2873 		/*
2874 		 * Stop if the page cannot be trivially entered into the
2875 		 * pmap.
2876 		 */
2877 		if (((m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) ||
2878 		    (m->flags & PG_FICTITIOUS) ||
2879 		    ((m->flags & PG_SWAPPED) &&
2880 		     (prot & VM_PROT_WRITE) &&
2881 		     (fault_flags & VM_FAULT_DIRTY))) {
2882 			vm_page_wakeup(m);
2883 			break;
2884 		}
2885 
2886 		/*
2887 		 * Enter the page into the pmap.  The object might be held
2888 		 * shared so we can't do any (serious) modifying operation
2889 		 * on it.
2890 		 */
2891 		if ((m->queue - m->pc) == PQ_CACHE)
2892 			vm_page_deactivate(m);
2893 		if (prot & VM_PROT_WRITE) {
2894 			vm_object_set_writeable_dirty(m->object);
2895 			vm_set_nosync(m, entry);
2896 			if (fault_flags & VM_FAULT_DIRTY) {
2897 				vm_page_dirty(m);
2898 				/* can't happeen due to conditional above */
2899 				/* swap_pager_unswapped(m); */
2900 			}
2901 		}
2902 		pmap_enter(pmap, addr, m, prot, 0, entry);
2903 		mycpu->gd_cnt.v_vm_faults++;
2904 		if (curthread->td_lwp)
2905 			++curthread->td_lwp->lwp_ru.ru_minflt;
2906 		vm_page_wakeup(m);
2907 	}
2908 }
2909